fs/ceph/mds_client.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/ceph/ceph_debug.h>
   3
   4 #include <linux/fs.h>
   5 #include <linux/wait.h>
   6 #include <linux/slab.h>
   7 #include <linux/gfp.h>
   8 #include <linux/sched.h>
   9 #include <linux/debugfs.h>
  10 #include <linux/seq_file.h>
  11 #include <linux/ratelimit.h>
  12 #include <linux/bits.h>
  13 #include <linux/ktime.h>
  14 #include <linux/bitmap.h>
  15 #include <linux/mnt_idmapping.h>
  16
  17 #include "super.h"
  18 #include "mds_client.h"
  19 #include "crypto.h"
  20
  21 #include <linux/ceph/ceph_features.h>
  22 #include <linux/ceph/messenger.h>
  23 #include <linux/ceph/decode.h>
  24 #include <linux/ceph/pagelist.h>
  25 #include <linux/ceph/auth.h>
  26 #include <linux/ceph/debugfs.h>
  27
  28 #define RECONNECT_MAX_SIZE (INT_MAX - PAGE_SIZE)
  29
  30 /*
  31  * A cluster of MDS (metadata server) daemons is responsible for
  32  * managing the file system namespace (the directory hierarchy and
  33  * inodes) and for coordinating shared access to storage.  Metadata is
  34  * partitioning hierarchically across a number of servers, and that
  35  * partition varies over time as the cluster adjusts the distribution
  36  * in order to balance load.
  37  *
  38  * The MDS client is primarily responsible to managing synchronous
  39  * metadata requests for operations like open, unlink, and so forth.
  40  * If there is a MDS failure, we find out about it when we (possibly
  41  * request and) receive a new MDS map, and can resubmit affected
  42  * requests.
  43  *
  44  * For the most part, though, we take advantage of a lossless
  45  * communications channel to the MDS, and do not need to worry about
  46  * timing out or resubmitting requests.
  47  *
  48  * We maintain a stateful "session" with each MDS we interact with.
  49  * Within each session, we sent periodic heartbeat messages to ensure
  50  * any capabilities or leases we have been issues remain valid.  If
  51  * the session times out and goes stale, our leases and capabilities
  52  * are no longer valid.
  53  */
  54
  55 struct ceph_reconnect_state {
  56         struct ceph_mds_session *session;
  57         int nr_caps, nr_realms;
  58         struct ceph_pagelist *pagelist;
  59         unsigned msg_version;
  60         bool allow_multi;
  61 };
  62
  63 static void __wake_requests(struct ceph_mds_client *mdsc,
  64                             struct list_head *head);
  65 static void ceph_cap_release_work(struct work_struct *work);
  66 static void ceph_cap_reclaim_work(struct work_struct *work);
  67
  68 static const struct ceph_connection_operations mds_con_ops;
  69
  70
  71 /*
  72  * mds reply parsing
  73  */
  74
  75 static int parse_reply_info_quota(void **p, void *end,
  76                                   struct ceph_mds_reply_info_in *info)
  77 {
  78         u8 struct_v, struct_compat;
  79         u32 struct_len;
  80
  81         ceph_decode_8_safe(p, end, struct_v, bad);
  82         ceph_decode_8_safe(p, end, struct_compat, bad);
  83         /* struct_v is expected to be >= 1. we only
  84          * understand encoding with struct_compat == 1. */
  85         if (!struct_v || struct_compat != 1)
  86                 goto bad;
  87         ceph_decode_32_safe(p, end, struct_len, bad);
  88         ceph_decode_need(p, end, struct_len, bad);
  89         end = *p + struct_len;
  90         ceph_decode_64_safe(p, end, info->max_bytes, bad);
  91         ceph_decode_64_safe(p, end, info->max_files, bad);
  92         *p = end;
  93         return 0;
  94 bad:
  95         return -EIO;
  96 }
  97
  98 /*
  99  * parse individual inode info
 100  */
 101 static int parse_reply_info_in(void **p, void *end,
 102                                struct ceph_mds_reply_info_in *info,
 103                                u64 features)
 104 {
 105         int err = 0;
 106         u8 struct_v = 0;
 107
 108         if (features == (u64)-1) {
 109                 u32 struct_len;
 110                 u8 struct_compat;
 111                 ceph_decode_8_safe(p, end, struct_v, bad);
 112                 ceph_decode_8_safe(p, end, struct_compat, bad);
 113                 /* struct_v is expected to be >= 1. we only understand
 114                  * encoding with struct_compat == 1. */
 115                 if (!struct_v || struct_compat != 1)
 116                         goto bad;
 117                 ceph_decode_32_safe(p, end, struct_len, bad);
 118                 ceph_decode_need(p, end, struct_len, bad);
 119                 end = *p + struct_len;
 120         }
 121
 122         ceph_decode_need(p, end, sizeof(struct ceph_mds_reply_inode), bad);
 123         info->in = *p;
 124         *p += sizeof(struct ceph_mds_reply_inode) +
 125                 sizeof(*info->in->fragtree.splits) *
 126                 le32_to_cpu(info->in->fragtree.nsplits);
 127
 128         ceph_decode_32_safe(p, end, info->symlink_len, bad);
 129         ceph_decode_need(p, end, info->symlink_len, bad);
 130         info->symlink = *p;
 131         *p += info->symlink_len;
 132
 133         ceph_decode_copy_safe(p, end, &info->dir_layout,
 134                               sizeof(info->dir_layout), bad);
 135         ceph_decode_32_safe(p, end, info->xattr_len, bad);
 136         ceph_decode_need(p, end, info->xattr_len, bad);
 137         info->xattr_data = *p;
 138         *p += info->xattr_len;
 139
 140         if (features == (u64)-1) {
 141                 /* inline data */
 142                 ceph_decode_64_safe(p, end, info->inline_version, bad);
 143                 ceph_decode_32_safe(p, end, info->inline_len, bad);
 144                 ceph_decode_need(p, end, info->inline_len, bad);
 145                 info->inline_data = *p;
 146                 *p += info->inline_len;
 147                 /* quota */
 148                 err = parse_reply_info_quota(p, end, info);
 149                 if (err < 0)
 150                         goto out_bad;
 151                 /* pool namespace */
 152                 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
 153                 if (info->pool_ns_len > 0) {
 154                         ceph_decode_need(p, end, info->pool_ns_len, bad);
 155                         info->pool_ns_data = *p;
 156                         *p += info->pool_ns_len;
 157                 }
 158
 159                 /* btime */
 160                 ceph_decode_need(p, end, sizeof(info->btime), bad);
 161                 ceph_decode_copy(p, &info->btime, sizeof(info->btime));
 162
 163                 /* change attribute */
 164                 ceph_decode_64_safe(p, end, info->change_attr, bad);
 165
 166                 /* dir pin */
 167                 if (struct_v >= 2) {
 168                         ceph_decode_32_safe(p, end, info->dir_pin, bad);
 169                 } else {
 170                         info->dir_pin = -ENODATA;
 171                 }
 172
 173                 /* snapshot birth time, remains zero for v<=2 */
 174                 if (struct_v >= 3) {
 175                         ceph_decode_need(p, end, sizeof(info->snap_btime), bad);
 176                         ceph_decode_copy(p, &info->snap_btime,
 177                                          sizeof(info->snap_btime));
 178                 } else {
 179                         memset(&info->snap_btime, 0, sizeof(info->snap_btime));
 180                 }
 181
 182                 /* snapshot count, remains zero for v<=3 */
 183                 if (struct_v >= 4) {
 184                         ceph_decode_64_safe(p, end, info->rsnaps, bad);
 185                 } else {
 186                         info->rsnaps = 0;
 187                 }
 188
 189                 if (struct_v >= 5) {
 190                         u32 alen;
 191
 192                         ceph_decode_32_safe(p, end, alen, bad);
 193
 194                         while (alen--) {
 195                                 u32 len;
 196
 197                                 /* key */
 198                                 ceph_decode_32_safe(p, end, len, bad);
 199                                 ceph_decode_skip_n(p, end, len, bad);
 200                                 /* value */
 201                                 ceph_decode_32_safe(p, end, len, bad);
 202                                 ceph_decode_skip_n(p, end, len, bad);
 203                         }
 204                 }
 205
 206                 /* fscrypt flag -- ignore */
 207                 if (struct_v >= 6)
 208                         ceph_decode_skip_8(p, end, bad);
 209
 210                 info->fscrypt_auth = NULL;
 211                 info->fscrypt_auth_len = 0;
 212                 info->fscrypt_file = NULL;
 213                 info->fscrypt_file_len = 0;
 214                 if (struct_v >= 7) {
 215                         ceph_decode_32_safe(p, end, info->fscrypt_auth_len, bad);
 216                         if (info->fscrypt_auth_len) {
 217                                 info->fscrypt_auth = kmalloc(info->fscrypt_auth_len,
 218                                                              GFP_KERNEL);
 219                                 if (!info->fscrypt_auth)
 220                                         return -ENOMEM;
 221                                 ceph_decode_copy_safe(p, end, info->fscrypt_auth,
 222                                                       info->fscrypt_auth_len, bad);
 223                         }
 224                         ceph_decode_32_safe(p, end, info->fscrypt_file_len, bad);
 225                         if (info->fscrypt_file_len) {
 226                                 info->fscrypt_file = kmalloc(info->fscrypt_file_len,
 227                                                              GFP_KERNEL);
 228                                 if (!info->fscrypt_file)
 229                                         return -ENOMEM;
 230                                 ceph_decode_copy_safe(p, end, info->fscrypt_file,
 231                                                       info->fscrypt_file_len, bad);
 232                         }
 233                 }
 234                 *p = end;
 235         } else {
 236                 /* legacy (unversioned) struct */
 237                 if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
 238                         ceph_decode_64_safe(p, end, info->inline_version, bad);
 239                         ceph_decode_32_safe(p, end, info->inline_len, bad);
 240                         ceph_decode_need(p, end, info->inline_len, bad);
 241                         info->inline_data = *p;
 242                         *p += info->inline_len;
 243                 } else
 244                         info->inline_version = CEPH_INLINE_NONE;
 245
 246                 if (features & CEPH_FEATURE_MDS_QUOTA) {
 247                         err = parse_reply_info_quota(p, end, info);
 248                         if (err < 0)
 249                                 goto out_bad;
 250                 } else {
 251                         info->max_bytes = 0;
 252                         info->max_files = 0;
 253                 }
 254
 255                 info->pool_ns_len = 0;
 256                 info->pool_ns_data = NULL;
 257                 if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
 258                         ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
 259                         if (info->pool_ns_len > 0) {
 260                                 ceph_decode_need(p, end, info->pool_ns_len, bad);
 261                                 info->pool_ns_data = *p;
 262                                 *p += info->pool_ns_len;
 263                         }
 264                 }
 265
 266                 if (features & CEPH_FEATURE_FS_BTIME) {
 267                         ceph_decode_need(p, end, sizeof(info->btime), bad);
 268                         ceph_decode_copy(p, &info->btime, sizeof(info->btime));
 269                         ceph_decode_64_safe(p, end, info->change_attr, bad);
 270                 }
 271
 272                 info->dir_pin = -ENODATA;
 273                 /* info->snap_btime and info->rsnaps remain zero */
 274         }
 275         return 0;
 276 bad:
 277         err = -EIO;
 278 out_bad:
 279         return err;
 280 }
 281
 282 static int parse_reply_info_dir(void **p, void *end,
 283                                 struct ceph_mds_reply_dirfrag **dirfrag,
 284                                 u64 features)
 285 {
 286         if (features == (u64)-1) {
 287                 u8 struct_v, struct_compat;
 288                 u32 struct_len;
 289                 ceph_decode_8_safe(p, end, struct_v, bad);
 290                 ceph_decode_8_safe(p, end, struct_compat, bad);
 291                 /* struct_v is expected to be >= 1. we only understand
 292                  * encoding whose struct_compat == 1. */
 293                 if (!struct_v || struct_compat != 1)
 294                         goto bad;
 295                 ceph_decode_32_safe(p, end, struct_len, bad);
 296                 ceph_decode_need(p, end, struct_len, bad);
 297                 end = *p + struct_len;
 298         }
 299
 300         ceph_decode_need(p, end, sizeof(**dirfrag), bad);
 301         *dirfrag = *p;
 302         *p += sizeof(**dirfrag) + sizeof(u32) * le32_to_cpu((*dirfrag)->ndist);
 303         if (unlikely(*p > end))
 304                 goto bad;
 305         if (features == (u64)-1)
 306                 *p = end;
 307         return 0;
 308 bad:
 309         return -EIO;
 310 }
 311
 312 static int parse_reply_info_lease(void **p, void *end,
 313                                   struct ceph_mds_reply_lease **lease,
 314                                   u64 features, u32 *altname_len, u8 **altname)
 315 {
 316         u8 struct_v;
 317         u32 struct_len;
 318         void *lend;
 319
 320         if (features == (u64)-1) {
 321                 u8 struct_compat;
 322
 323                 ceph_decode_8_safe(p, end, struct_v, bad);
 324                 ceph_decode_8_safe(p, end, struct_compat, bad);
 325
 326                 /* struct_v is expected to be >= 1. we only understand
 327                  * encoding whose struct_compat == 1. */
 328                 if (!struct_v || struct_compat != 1)
 329                         goto bad;
 330
 331                 ceph_decode_32_safe(p, end, struct_len, bad);
 332         } else {
 333                 struct_len = sizeof(**lease);
 334                 *altname_len = 0;
 335                 *altname = NULL;
 336         }
 337
 338         lend = *p + struct_len;
 339         ceph_decode_need(p, end, struct_len, bad);
 340         *lease = *p;
 341         *p += sizeof(**lease);
 342
 343         if (features == (u64)-1) {
 344                 if (struct_v >= 2) {
 345                         ceph_decode_32_safe(p, end, *altname_len, bad);
 346                         ceph_decode_need(p, end, *altname_len, bad);
 347                         *altname = *p;
 348                         *p += *altname_len;
 349                 } else {
 350                         *altname = NULL;
 351                         *altname_len = 0;
 352                 }
 353         }
 354         *p = lend;
 355         return 0;
 356 bad:
 357         return -EIO;
 358 }
 359
 360 /*
 361  * parse a normal reply, which may contain a (dir+)dentry and/or a
 362  * target inode.
 363  */
 364 static int parse_reply_info_trace(void **p, void *end,
 365                                   struct ceph_mds_reply_info_parsed *info,
 366                                   u64 features)
 367 {
 368         int err;
 369
 370         if (info->head->is_dentry) {
 371                 err = parse_reply_info_in(p, end, &info->diri, features);
 372                 if (err < 0)
 373                         goto out_bad;
 374
 375                 err = parse_reply_info_dir(p, end, &info->dirfrag, features);
 376                 if (err < 0)
 377                         goto out_bad;
 378
 379                 ceph_decode_32_safe(p, end, info->dname_len, bad);
 380                 ceph_decode_need(p, end, info->dname_len, bad);
 381                 info->dname = *p;
 382                 *p += info->dname_len;
 383
 384                 err = parse_reply_info_lease(p, end, &info->dlease, features,
 385                                              &info->altname_len, &info->altname);
 386                 if (err < 0)
 387                         goto out_bad;
 388         }
 389
 390         if (info->head->is_target) {
 391                 err = parse_reply_info_in(p, end, &info->targeti, features);
 392                 if (err < 0)
 393                         goto out_bad;
 394         }
 395
 396         if (unlikely(*p != end))
 397                 goto bad;
 398         return 0;
 399
 400 bad:
 401         err = -EIO;
 402 out_bad:
 403         pr_err("problem parsing mds trace %d\n", err);
 404         return err;
 405 }
 406
 407 /*
 408  * parse readdir results
 409  */
 410 static int parse_reply_info_readdir(void **p, void *end,
 411                                     struct ceph_mds_request *req,
 412                                     u64 features)
 413 {
 414         struct ceph_mds_reply_info_parsed *info = &req->r_reply_info;
 415         struct ceph_client *cl = req->r_mdsc->fsc->client;
 416         u32 num, i = 0;
 417         int err;
 418
 419         err = parse_reply_info_dir(p, end, &info->dir_dir, features);
 420         if (err < 0)
 421                 goto out_bad;
 422
 423         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 424         num = ceph_decode_32(p);
 425         {
 426                 u16 flags = ceph_decode_16(p);
 427                 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
 428                 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
 429                 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
 430                 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
 431         }
 432         if (num == 0)
 433                 goto done;
 434
 435         BUG_ON(!info->dir_entries);
 436         if ((unsigned long)(info->dir_entries + num) >
 437             (unsigned long)info->dir_entries + info->dir_buf_size) {
 438                 pr_err_client(cl, "dir contents are larger than expected\n");
 439                 WARN_ON(1);
 440                 goto bad;
 441         }
 442
 443         info->dir_nr = num;
 444         while (num) {
 445                 struct inode *inode = d_inode(req->r_dentry);
 446                 struct ceph_inode_info *ci = ceph_inode(inode);
 447                 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
 448                 struct fscrypt_str tname = FSTR_INIT(NULL, 0);
 449                 struct fscrypt_str oname = FSTR_INIT(NULL, 0);
 450                 struct ceph_fname fname;
 451                 u32 altname_len, _name_len;
 452                 u8 *altname, *_name;
 453
 454                 /* dentry */
 455                 ceph_decode_32_safe(p, end, _name_len, bad);
 456                 ceph_decode_need(p, end, _name_len, bad);
 457                 _name = *p;
 458                 *p += _name_len;
 459                 doutc(cl, "parsed dir dname '%.*s'\n", _name_len, _name);
 460
 461                 if (info->hash_order)
 462                         rde->raw_hash = ceph_str_hash(ci->i_dir_layout.dl_dir_hash,
 463                                                       _name, _name_len);
 464
 465                 /* dentry lease */
 466                 err = parse_reply_info_lease(p, end, &rde->lease, features,
 467                                              &altname_len, &altname);
 468                 if (err)
 469                         goto out_bad;
 470
 471                 /*
 472                  * Try to dencrypt the dentry names and update them
 473                  * in the ceph_mds_reply_dir_entry struct.
 474                  */
 475                 fname.dir = inode;
 476                 fname.name = _name;
 477                 fname.name_len = _name_len;
 478                 fname.ctext = altname;
 479                 fname.ctext_len = altname_len;
 480                 /*
 481                  * The _name_len maybe larger than altname_len, such as
 482                  * when the human readable name length is in range of
 483                  * (CEPH_NOHASH_NAME_MAX, CEPH_NOHASH_NAME_MAX + SHA256_DIGEST_SIZE),
 484                  * then the copy in ceph_fname_to_usr will corrupt the
 485                  * data if there has no encryption key.
 486                  *
 487                  * Just set the no_copy flag and then if there has no
 488                  * encryption key the oname.name will be assigned to
 489                  * _name always.
 490                  */
 491                 fname.no_copy = true;
 492                 if (altname_len == 0) {
 493                         /*
 494                          * Set tname to _name, and this will be used
 495                          * to do the base64_decode in-place. It's
 496                          * safe because the decoded string should
 497                          * always be shorter, which is 3/4 of origin
 498                          * string.
 499                          */
 500                         tname.name = _name;
 501
 502                         /*
 503                          * Set oname to _name too, and this will be
 504                          * used to do the dencryption in-place.
 505                          */
 506                         oname.name = _name;
 507                         oname.len = _name_len;
 508                 } else {
 509                         /*
 510                          * This will do the decryption only in-place
 511                          * from altname cryptext directly.
 512                          */
 513                         oname.name = altname;
 514                         oname.len = altname_len;
 515                 }
 516                 rde->is_nokey = false;
 517                 err = ceph_fname_to_usr(&fname, &tname, &oname, &rde->is_nokey);
 518                 if (err) {
 519                         pr_err_client(cl, "unable to decode %.*s, got %d\n",
 520                                       _name_len, _name, err);
 521                         goto out_bad;
 522                 }
 523                 rde->name = oname.name;
 524                 rde->name_len = oname.len;
 525
 526                 /* inode */
 527                 err = parse_reply_info_in(p, end, &rde->inode, features);
 528                 if (err < 0)
 529                         goto out_bad;
 530                 /* ceph_readdir_prepopulate() will update it */
 531                 rde->offset = 0;
 532                 i++;
 533                 num--;
 534         }
 535
 536 done:
 537         /* Skip over any unrecognized fields */
 538         *p = end;
 539         return 0;
 540
 541 bad:
 542         err = -EIO;
 543 out_bad:
 544         pr_err_client(cl, "problem parsing dir contents %d\n", err);
 545         return err;
 546 }
 547
 548 /*
 549  * parse fcntl F_GETLK results
 550  */
 551 static int parse_reply_info_filelock(void **p, void *end,
 552                                      struct ceph_mds_reply_info_parsed *info,
 553                                      u64 features)
 554 {
 555         if (*p + sizeof(*info->filelock_reply) > end)
 556                 goto bad;
 557
 558         info->filelock_reply = *p;
 559
 560         /* Skip over any unrecognized fields */
 561         *p = end;
 562         return 0;
 563 bad:
 564         return -EIO;
 565 }
 566
 567
 568 #if BITS_PER_LONG == 64
 569
 570 #define DELEGATED_INO_AVAILABLE         xa_mk_value(1)
 571
 572 static int ceph_parse_deleg_inos(void **p, void *end,
 573                                  struct ceph_mds_session *s)
 574 {
 575         struct ceph_client *cl = s->s_mdsc->fsc->client;
 576         u32 sets;
 577
 578         ceph_decode_32_safe(p, end, sets, bad);
 579         doutc(cl, "got %u sets of delegated inodes\n", sets);
 580         while (sets--) {
 581                 u64 start, len;
 582
 583                 ceph_decode_64_safe(p, end, start, bad);
 584                 ceph_decode_64_safe(p, end, len, bad);
 585
 586                 /* Don't accept a delegation of system inodes */
 587                 if (start < CEPH_INO_SYSTEM_BASE) {
 588                         pr_warn_ratelimited_client(cl,
 589                                 "ignoring reserved inode range delegation (start=0x%llx len=0x%llx)\n",
 590                                 start, len);
 591                         continue;
 592                 }
 593                 while (len--) {
 594                         int err = xa_insert(&s->s_delegated_inos, start++,
 595                                             DELEGATED_INO_AVAILABLE,
 596                                             GFP_KERNEL);
 597                         if (!err) {
 598                                 doutc(cl, "added delegated inode 0x%llx\n", start - 1);
 599                         } else if (err == -EBUSY) {
 600                                 pr_warn_client(cl,
 601                                         "MDS delegated inode 0x%llx more than once.\n",
 602                                         start - 1);
 603                         } else {
 604                                 return err;
 605                         }
 606                 }
 607         }
 608         return 0;
 609 bad:
 610         return -EIO;
 611 }
 612
 613 u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
 614 {
 615         unsigned long ino;
 616         void *val;
 617
 618         xa_for_each(&s->s_delegated_inos, ino, val) {
 619                 val = xa_erase(&s->s_delegated_inos, ino);
 620                 if (val == DELEGATED_INO_AVAILABLE)
 621                         return ino;
 622         }
 623         return 0;
 624 }
 625
 626 int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
 627 {
 628         return xa_insert(&s->s_delegated_inos, ino, DELEGATED_INO_AVAILABLE,
 629                          GFP_KERNEL);
 630 }
 631 #else /* BITS_PER_LONG == 64 */
 632 /*
 633  * FIXME: xarrays can't handle 64-bit indexes on a 32-bit arch. For now, just
 634  * ignore delegated_inos on 32 bit arch. Maybe eventually add xarrays for top
 635  * and bottom words?
 636  */
 637 static int ceph_parse_deleg_inos(void **p, void *end,
 638                                  struct ceph_mds_session *s)
 639 {
 640         u32 sets;
 641
 642         ceph_decode_32_safe(p, end, sets, bad);
 643         if (sets)
 644                 ceph_decode_skip_n(p, end, sets * 2 * sizeof(__le64), bad);
 645         return 0;
 646 bad:
 647         return -EIO;
 648 }
 649
 650 u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
 651 {
 652         return 0;
 653 }
 654
 655 int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
 656 {
 657         return 0;
 658 }
 659 #endif /* BITS_PER_LONG == 64 */
 660
 661 /*
 662  * parse create results
 663  */
 664 static int parse_reply_info_create(void **p, void *end,
 665                                   struct ceph_mds_reply_info_parsed *info,
 666                                   u64 features, struct ceph_mds_session *s)
 667 {
 668         int ret;
 669
 670         if (features == (u64)-1 ||
 671             (features & CEPH_FEATURE_REPLY_CREATE_INODE)) {
 672                 if (*p == end) {
 673                         /* Malformed reply? */
 674                         info->has_create_ino = false;
 675                 } else if (test_bit(CEPHFS_FEATURE_DELEG_INO, &s->s_features)) {
 676                         info->has_create_ino = true;
 677                         /* struct_v, struct_compat, and len */
 678                         ceph_decode_skip_n(p, end, 2 + sizeof(u32), bad);
 679                         ceph_decode_64_safe(p, end, info->ino, bad);
 680                         ret = ceph_parse_deleg_inos(p, end, s);
 681                         if (ret)
 682                                 return ret;
 683                 } else {
 684                         /* legacy */
 685                         ceph_decode_64_safe(p, end, info->ino, bad);
 686                         info->has_create_ino = true;
 687                 }
 688         } else {
 689                 if (*p != end)
 690                         goto bad;
 691         }
 692
 693         /* Skip over any unrecognized fields */
 694         *p = end;
 695         return 0;
 696 bad:
 697         return -EIO;
 698 }
 699
 700 static int parse_reply_info_getvxattr(void **p, void *end,
 701                                       struct ceph_mds_reply_info_parsed *info,
 702                                       u64 features)
 703 {
 704         u32 value_len;
 705
 706         ceph_decode_skip_8(p, end, bad); /* skip current version: 1 */
 707         ceph_decode_skip_8(p, end, bad); /* skip first version: 1 */
 708         ceph_decode_skip_32(p, end, bad); /* skip payload length */
 709
 710         ceph_decode_32_safe(p, end, value_len, bad);
 711
 712         if (value_len == end - *p) {
 713           info->xattr_info.xattr_value = *p;
 714           info->xattr_info.xattr_value_len = value_len;
 715           *p = end;
 716           return value_len;
 717         }
 718 bad:
 719         return -EIO;
 720 }
 721
 722 /*
 723  * parse extra results
 724  */
 725 static int parse_reply_info_extra(void **p, void *end,
 726                                   struct ceph_mds_request *req,
 727                                   u64 features, struct ceph_mds_session *s)
 728 {
 729         struct ceph_mds_reply_info_parsed *info = &req->r_reply_info;
 730         u32 op = le32_to_cpu(info->head->op);
 731
 732         if (op == CEPH_MDS_OP_GETFILELOCK)
 733                 return parse_reply_info_filelock(p, end, info, features);
 734         else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
 735                 return parse_reply_info_readdir(p, end, req, features);
 736         else if (op == CEPH_MDS_OP_CREATE)
 737                 return parse_reply_info_create(p, end, info, features, s);
 738         else if (op == CEPH_MDS_OP_GETVXATTR)
 739                 return parse_reply_info_getvxattr(p, end, info, features);
 740         else
 741                 return -EIO;
 742 }
 743
 744 /*
 745  * parse entire mds reply
 746  */
 747 static int parse_reply_info(struct ceph_mds_session *s, struct ceph_msg *msg,
 748                             struct ceph_mds_request *req, u64 features)
 749 {
 750         struct ceph_mds_reply_info_parsed *info = &req->r_reply_info;
 751         struct ceph_client *cl = s->s_mdsc->fsc->client;
 752         void *p, *end;
 753         u32 len;
 754         int err;
 755
 756         info->head = msg->front.iov_base;
 757         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 758         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 759
 760         /* trace */
 761         ceph_decode_32_safe(&p, end, len, bad);
 762         if (len > 0) {
 763                 ceph_decode_need(&p, end, len, bad);
 764                 err = parse_reply_info_trace(&p, p+len, info, features);
 765                 if (err < 0)
 766                         goto out_bad;
 767         }
 768
 769         /* extra */
 770         ceph_decode_32_safe(&p, end, len, bad);
 771         if (len > 0) {
 772                 ceph_decode_need(&p, end, len, bad);
 773                 err = parse_reply_info_extra(&p, p+len, req, features, s);
 774                 if (err < 0)
 775                         goto out_bad;
 776         }
 777
 778         /* snap blob */
 779         ceph_decode_32_safe(&p, end, len, bad);
 780         info->snapblob_len = len;
 781         info->snapblob = p;
 782         p += len;
 783
 784         if (p != end)
 785                 goto bad;
 786         return 0;
 787
 788 bad:
 789         err = -EIO;
 790 out_bad:
 791         pr_err_client(cl, "mds parse_reply err %d\n", err);
 792         ceph_msg_dump(msg);
 793         return err;
 794 }
 795
 796 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 797 {
 798         int i;
 799
 800         kfree(info->diri.fscrypt_auth);
 801         kfree(info->diri.fscrypt_file);
 802         kfree(info->targeti.fscrypt_auth);
 803         kfree(info->targeti.fscrypt_file);
 804         if (!info->dir_entries)
 805                 return;
 806
 807         for (i = 0; i < info->dir_nr; i++) {
 808                 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
 809
 810                 kfree(rde->inode.fscrypt_auth);
 811                 kfree(rde->inode.fscrypt_file);
 812         }
 813         free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
 814 }
 815
 816 /*
 817  * In async unlink case the kclient won't wait for the first reply
 818  * from MDS and just drop all the links and unhash the dentry and then
 819  * succeeds immediately.
 820  *
 821  * For any new create/link/rename,etc requests followed by using the
 822  * same file names we must wait for the first reply of the inflight
 823  * unlink request, or the MDS possibly will fail these following
 824  * requests with -EEXIST if the inflight async unlink request was
 825  * delayed for some reasons.
 826  *
 827  * And the worst case is that for the none async openc request it will
 828  * successfully open the file if the CDentry hasn't been unlinked yet,
 829  * but later the previous delayed async unlink request will remove the
 830  * CDentry. That means the just created file is possibly deleted later
 831  * by accident.
 832  *
 833  * We need to wait for the inflight async unlink requests to finish
 834  * when creating new files/directories by using the same file names.
 835  */
 836 int ceph_wait_on_conflict_unlink(struct dentry *dentry)
 837 {
 838         struct ceph_fs_client *fsc = ceph_sb_to_fs_client(dentry->d_sb);
 839         struct ceph_client *cl = fsc->client;
 840         struct dentry *pdentry = dentry->d_parent;
 841         struct dentry *udentry, *found = NULL;
 842         struct ceph_dentry_info *di;
 843         struct qstr dname;
 844         u32 hash = dentry->d_name.hash;
 845         int err;
 846
 847         dname.name = dentry->d_name.name;
 848         dname.len = dentry->d_name.len;
 849
 850         rcu_read_lock();
 851         hash_for_each_possible_rcu(fsc->async_unlink_conflict, di,
 852                                    hnode, hash) {
 853                 udentry = di->dentry;
 854
 855                 spin_lock(&udentry->d_lock);
 856                 if (udentry->d_name.hash != hash)
 857                         goto next;
 858                 if (unlikely(udentry->d_parent != pdentry))
 859                         goto next;
 860                 if (!hash_hashed(&di->hnode))
 861                         goto next;
 862
 863                 if (!test_bit(CEPH_DENTRY_ASYNC_UNLINK_BIT, &di->flags))
 864                         pr_warn_client(cl, "dentry %p:%pd async unlink bit is not set\n",
 865                                        dentry, dentry);
 866
 867                 if (!d_same_name(udentry, pdentry, &dname))
 868                         goto next;
 869
 870                 found = dget_dlock(udentry);
 871                 spin_unlock(&udentry->d_lock);
 872                 break;
 873 next:
 874                 spin_unlock(&udentry->d_lock);
 875         }
 876         rcu_read_unlock();
 877
 878         if (likely(!found))
 879                 return 0;
 880
 881         doutc(cl, "dentry %p:%pd conflict with old %p:%pd\n", dentry, dentry,
 882               found, found);
 883
 884         err = wait_on_bit(&di->flags, CEPH_DENTRY_ASYNC_UNLINK_BIT,
 885                           TASK_KILLABLE);
 886         dput(found);
 887         return err;
 888 }
 889
 890
 891 /*
 892  * sessions
 893  */
 894 const char *ceph_session_state_name(int s)
 895 {
 896         switch (s) {
 897         case CEPH_MDS_SESSION_NEW: return "new";
 898         case CEPH_MDS_SESSION_OPENING: return "opening";
 899         case CEPH_MDS_SESSION_OPEN: return "open";
 900         case CEPH_MDS_SESSION_HUNG: return "hung";
 901         case CEPH_MDS_SESSION_CLOSING: return "closing";
 902         case CEPH_MDS_SESSION_CLOSED: return "closed";
 903         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 904         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 905         case CEPH_MDS_SESSION_REJECTED: return "rejected";
 906         default: return "???";
 907         }
 908 }
 909
 910 struct ceph_mds_session *ceph_get_mds_session(struct ceph_mds_session *s)
 911 {
 912         if (refcount_inc_not_zero(&s->s_ref))
 913                 return s;
 914         return NULL;
 915 }
 916
 917 void ceph_put_mds_session(struct ceph_mds_session *s)
 918 {
 919         if (IS_ERR_OR_NULL(s))
 920                 return;
 921
 922         if (refcount_dec_and_test(&s->s_ref)) {
 923                 if (s->s_auth.authorizer)
 924                         ceph_auth_destroy_authorizer(s->s_auth.authorizer);
 925                 WARN_ON(mutex_is_locked(&s->s_mutex));
 926                 xa_destroy(&s->s_delegated_inos);
 927                 kfree(s);
 928         }
 929 }
 930
 931 /*
 932  * called under mdsc->mutex
 933  */
 934 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 935                                                    int mds)
 936 {
 937         if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
 938                 return NULL;
 939         return ceph_get_mds_session(mdsc->sessions[mds]);
 940 }
 941
 942 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 943 {
 944         if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
 945                 return false;
 946         else
 947                 return true;
 948 }
 949
 950 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 951                                        struct ceph_mds_session *s)
 952 {
 953         if (s->s_mds >= mdsc->max_sessions ||
 954             mdsc->sessions[s->s_mds] != s)
 955                 return -ENOENT;
 956         return 0;
 957 }
 958
 959 /*
 960  * create+register a new session for given mds.
 961  * called under mdsc->mutex.
 962  */
 963 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 964                                                  int mds)
 965 {
 966         struct ceph_client *cl = mdsc->fsc->client;
 967         struct ceph_mds_session *s;
 968
 969         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_FENCE_IO)
 970                 return ERR_PTR(-EIO);
 971
 972         if (mds >= mdsc->mdsmap->possible_max_rank)
 973                 return ERR_PTR(-EINVAL);
 974
 975         s = kzalloc(sizeof(*s), GFP_NOFS);
 976         if (!s)
 977                 return ERR_PTR(-ENOMEM);
 978
 979         if (mds >= mdsc->max_sessions) {
 980                 int newmax = 1 << get_count_order(mds + 1);
 981                 struct ceph_mds_session **sa;
 982
 983                 doutc(cl, "realloc to %d\n", newmax);
 984                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 985                 if (!sa)
 986                         goto fail_realloc;
 987                 if (mdsc->sessions) {
 988                         memcpy(sa, mdsc->sessions,
 989                                mdsc->max_sessions * sizeof(void *));
 990                         kfree(mdsc->sessions);
 991                 }
 992                 mdsc->sessions = sa;
 993                 mdsc->max_sessions = newmax;
 994         }
 995
 996         doutc(cl, "mds%d\n", mds);
 997         s->s_mdsc = mdsc;
 998         s->s_mds = mds;
 999         s->s_state = CEPH_MDS_SESSION_NEW;
1000         mutex_init(&s->s_mutex);
1001
1002         ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
1003
1004         atomic_set(&s->s_cap_gen, 1);
1005         s->s_cap_ttl = jiffies - 1;
1006
1007         spin_lock_init(&s->s_cap_lock);
1008         INIT_LIST_HEAD(&s->s_caps);
1009         refcount_set(&s->s_ref, 1);
1010         INIT_LIST_HEAD(&s->s_waiting);
1011         INIT_LIST_HEAD(&s->s_unsafe);
1012         xa_init(&s->s_delegated_inos);
1013         INIT_LIST_HEAD(&s->s_cap_releases);
1014         INIT_WORK(&s->s_cap_release_work, ceph_cap_release_work);
1015
1016         INIT_LIST_HEAD(&s->s_cap_dirty);
1017         INIT_LIST_HEAD(&s->s_cap_flushing);
1018
1019         mdsc->sessions[mds] = s;
1020         atomic_inc(&mdsc->num_sessions);
1021         refcount_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
1022
1023         ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
1024                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
1025
1026         return s;
1027
1028 fail_realloc:
1029         kfree(s);
1030         return ERR_PTR(-ENOMEM);
1031 }
1032
1033 /*
1034  * called under mdsc->mutex
1035  */
1036 static void __unregister_session(struct ceph_mds_client *mdsc,
1037                                struct ceph_mds_session *s)
1038 {
1039         doutc(mdsc->fsc->client, "mds%d %p\n", s->s_mds, s);
1040         BUG_ON(mdsc->sessions[s->s_mds] != s);
1041         mdsc->sessions[s->s_mds] = NULL;
1042         ceph_con_close(&s->s_con);
1043         ceph_put_mds_session(s);
1044         atomic_dec(&mdsc->num_sessions);
1045 }
1046
1047 /*
1048  * drop session refs in request.
1049  *
1050  * should be last request ref, or hold mdsc->mutex
1051  */
1052 static void put_request_session(struct ceph_mds_request *req)
1053 {
1054         if (req->r_session) {
1055                 ceph_put_mds_session(req->r_session);
1056                 req->r_session = NULL;
1057         }
1058 }
1059
1060 void ceph_mdsc_iterate_sessions(struct ceph_mds_client *mdsc,
1061                                 void (*cb)(struct ceph_mds_session *),
1062                                 bool check_state)
1063 {
1064         int mds;
1065
1066         mutex_lock(&mdsc->mutex);
1067         for (mds = 0; mds < mdsc->max_sessions; ++mds) {
1068                 struct ceph_mds_session *s;
1069
1070                 s = __ceph_lookup_mds_session(mdsc, mds);
1071                 if (!s)
1072                         continue;
1073
1074                 if (check_state && !check_session_state(s)) {
1075                         ceph_put_mds_session(s);
1076                         continue;
1077                 }
1078
1079                 mutex_unlock(&mdsc->mutex);
1080                 cb(s);
1081                 ceph_put_mds_session(s);
1082                 mutex_lock(&mdsc->mutex);
1083         }
1084         mutex_unlock(&mdsc->mutex);
1085 }
1086
1087 void ceph_mdsc_release_request(struct kref *kref)
1088 {
1089         struct ceph_mds_request *req = container_of(kref,
1090                                                     struct ceph_mds_request,
1091                                                     r_kref);
1092         ceph_mdsc_release_dir_caps_async(req);
1093         destroy_reply_info(&req->r_reply_info);
1094         if (req->r_request)
1095                 ceph_msg_put(req->r_request);
1096         if (req->r_reply)
1097                 ceph_msg_put(req->r_reply);
1098         if (req->r_inode) {
1099                 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1100                 iput(req->r_inode);
1101         }
1102         if (req->r_parent) {
1103                 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
1104                 iput(req->r_parent);
1105         }
1106         iput(req->r_target_inode);
1107         iput(req->r_new_inode);
1108         if (req->r_dentry)
1109                 dput(req->r_dentry);
1110         if (req->r_old_dentry)
1111                 dput(req->r_old_dentry);
1112         if (req->r_old_dentry_dir) {
1113                 /*
1114                  * track (and drop pins for) r_old_dentry_dir
1115                  * separately, since r_old_dentry's d_parent may have
1116                  * changed between the dir mutex being dropped and
1117                  * this request being freed.
1118                  */
1119                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
1120                                   CEPH_CAP_PIN);
1121                 iput(req->r_old_dentry_dir);
1122         }
1123         kfree(req->r_path1);
1124         kfree(req->r_path2);
1125         put_cred(req->r_cred);
1126         if (req->r_mnt_idmap)
1127                 mnt_idmap_put(req->r_mnt_idmap);
1128         if (req->r_pagelist)
1129                 ceph_pagelist_release(req->r_pagelist);
1130         kfree(req->r_fscrypt_auth);
1131         kfree(req->r_altname);
1132         put_request_session(req);
1133         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
1134         WARN_ON_ONCE(!list_empty(&req->r_wait));
1135         kmem_cache_free(ceph_mds_request_cachep, req);
1136 }
1137
1138 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
1139
1140 /*
1141  * lookup session, bump ref if found.
1142  *
1143  * called under mdsc->mutex.
1144  */
1145 static struct ceph_mds_request *
1146 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
1147 {
1148         struct ceph_mds_request *req;
1149
1150         req = lookup_request(&mdsc->request_tree, tid);
1151         if (req)
1152                 ceph_mdsc_get_request(req);
1153
1154         return req;
1155 }
1156
1157 /*
1158  * Register an in-flight request, and assign a tid.  Link to directory
1159  * are modifying (if any).
1160  *
1161  * Called under mdsc->mutex.
1162  */
1163 static void __register_request(struct ceph_mds_client *mdsc,
1164                                struct ceph_mds_request *req,
1165                                struct inode *dir)
1166 {
1167         struct ceph_client *cl = mdsc->fsc->client;
1168         int ret = 0;
1169
1170         req->r_tid = ++mdsc->last_tid;
1171         if (req->r_num_caps) {
1172                 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
1173                                         req->r_num_caps);
1174                 if (ret < 0) {
1175                         pr_err_client(cl, "%p failed to reserve caps: %d\n",
1176                                       req, ret);
1177                         /* set req->r_err to fail early from __do_request */
1178                         req->r_err = ret;
1179                         return;
1180                 }
1181         }
1182         doutc(cl, "%p tid %lld\n", req, req->r_tid);
1183         ceph_mdsc_get_request(req);
1184         insert_request(&mdsc->request_tree, req);
1185
1186         req->r_cred = get_current_cred();
1187         if (!req->r_mnt_idmap)
1188                 req->r_mnt_idmap = &nop_mnt_idmap;
1189
1190         if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
1191                 mdsc->oldest_tid = req->r_tid;
1192
1193         if (dir) {
1194                 struct ceph_inode_info *ci = ceph_inode(dir);
1195
1196                 ihold(dir);
1197                 req->r_unsafe_dir = dir;
1198                 spin_lock(&ci->i_unsafe_lock);
1199                 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
1200                 spin_unlock(&ci->i_unsafe_lock);
1201         }
1202 }
1203
1204 static void __unregister_request(struct ceph_mds_client *mdsc,
1205                                  struct ceph_mds_request *req)
1206 {
1207         doutc(mdsc->fsc->client, "%p tid %lld\n", req, req->r_tid);
1208
1209         /* Never leave an unregistered request on an unsafe list! */
1210         list_del_init(&req->r_unsafe_item);
1211
1212         if (req->r_tid == mdsc->oldest_tid) {
1213                 struct rb_node *p = rb_next(&req->r_node);
1214                 mdsc->oldest_tid = 0;
1215                 while (p) {
1216                         struct ceph_mds_request *next_req =
1217                                 rb_entry(p, struct ceph_mds_request, r_node);
1218                         if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
1219                                 mdsc->oldest_tid = next_req->r_tid;
1220                                 break;
1221                         }
1222                         p = rb_next(p);
1223                 }
1224         }
1225
1226         erase_request(&mdsc->request_tree, req);
1227
1228         if (req->r_unsafe_dir) {
1229                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
1230                 spin_lock(&ci->i_unsafe_lock);
1231                 list_del_init(&req->r_unsafe_dir_item);
1232                 spin_unlock(&ci->i_unsafe_lock);
1233         }
1234         if (req->r_target_inode &&
1235             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
1236                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
1237                 spin_lock(&ci->i_unsafe_lock);
1238                 list_del_init(&req->r_unsafe_target_item);
1239                 spin_unlock(&ci->i_unsafe_lock);
1240         }
1241
1242         if (req->r_unsafe_dir) {
1243                 iput(req->r_unsafe_dir);
1244                 req->r_unsafe_dir = NULL;
1245         }
1246
1247         complete_all(&req->r_safe_completion);
1248
1249         ceph_mdsc_put_request(req);
1250 }
1251
1252 /*
1253  * Walk back up the dentry tree until we hit a dentry representing a
1254  * non-snapshot inode. We do this using the rcu_read_lock (which must be held
1255  * when calling this) to ensure that the objects won't disappear while we're
1256  * working with them. Once we hit a candidate dentry, we attempt to take a
1257  * reference to it, and return that as the result.
1258  */
1259 static struct inode *get_nonsnap_parent(struct dentry *dentry)
1260 {
1261         struct inode *inode = NULL;
1262
1263         while (dentry && !IS_ROOT(dentry)) {
1264                 inode = d_inode_rcu(dentry);
1265                 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
1266                         break;
1267                 dentry = dentry->d_parent;
1268         }
1269         if (inode)
1270                 inode = igrab(inode);
1271         return inode;
1272 }
1273
1274 /*
1275  * Choose mds to send request to next.  If there is a hint set in the
1276  * request (e.g., due to a prior forward hint from the mds), use that.
1277  * Otherwise, consult frag tree and/or caps to identify the
1278  * appropriate mds.  If all else fails, choose randomly.
1279  *
1280  * Called under mdsc->mutex.
1281  */
1282 static int __choose_mds(struct ceph_mds_client *mdsc,
1283                         struct ceph_mds_request *req,
1284                         bool *random)
1285 {
1286         struct inode *inode;
1287         struct ceph_inode_info *ci;
1288         struct ceph_cap *cap;
1289         int mode = req->r_direct_mode;
1290         int mds = -1;
1291         u32 hash = req->r_direct_hash;
1292         bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
1293         struct ceph_client *cl = mdsc->fsc->client;
1294
1295         if (random)
1296                 *random = false;
1297
1298         /*
1299          * is there a specific mds we should try?  ignore hint if we have
1300          * no session and the mds is not up (active or recovering).
1301          */
1302         if (req->r_resend_mds >= 0 &&
1303             (__have_session(mdsc, req->r_resend_mds) ||
1304              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
1305                 doutc(cl, "using resend_mds mds%d\n", req->r_resend_mds);
1306                 return req->r_resend_mds;
1307         }
1308
1309         if (mode == USE_RANDOM_MDS)
1310                 goto random;
1311
1312         inode = NULL;
1313         if (req->r_inode) {
1314                 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
1315                         inode = req->r_inode;
1316                         ihold(inode);
1317                 } else {
1318                         /* req->r_dentry is non-null for LSSNAP request */
1319                         rcu_read_lock();
1320                         inode = get_nonsnap_parent(req->r_dentry);
1321                         rcu_read_unlock();
1322                         doutc(cl, "using snapdir's parent %p %llx.%llx\n",
1323                               inode, ceph_vinop(inode));
1324                 }
1325         } else if (req->r_dentry) {
1326                 /* ignore race with rename; old or new d_parent is okay */
1327                 struct dentry *parent;
1328                 struct inode *dir;
1329
1330                 rcu_read_lock();
1331                 parent = READ_ONCE(req->r_dentry->d_parent);
1332                 dir = req->r_parent ? : d_inode_rcu(parent);
1333
1334                 if (!dir || dir->i_sb != mdsc->fsc->sb) {
1335                         /*  not this fs or parent went negative */
1336                         inode = d_inode(req->r_dentry);
1337                         if (inode)
1338                                 ihold(inode);
1339                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
1340                         /* direct snapped/virtual snapdir requests
1341                          * based on parent dir inode */
1342                         inode = get_nonsnap_parent(parent);
1343                         doutc(cl, "using nonsnap parent %p %llx.%llx\n",
1344                               inode, ceph_vinop(inode));
1345                 } else {
1346                         /* dentry target */
1347                         inode = d_inode(req->r_dentry);
1348                         if (!inode || mode == USE_AUTH_MDS) {
1349                                 /* dir + name */
1350                                 inode = igrab(dir);
1351                                 hash = ceph_dentry_hash(dir, req->r_dentry);
1352                                 is_hash = true;
1353                         } else {
1354                                 ihold(inode);
1355                         }
1356                 }
1357                 rcu_read_unlock();
1358         }
1359
1360         if (!inode)
1361                 goto random;
1362
1363         doutc(cl, "%p %llx.%llx is_hash=%d (0x%x) mode %d\n", inode,
1364               ceph_vinop(inode), (int)is_hash, hash, mode);
1365         ci = ceph_inode(inode);
1366
1367         if (is_hash && S_ISDIR(inode->i_mode)) {
1368                 struct ceph_inode_frag frag;
1369                 int found;
1370
1371                 ceph_choose_frag(ci, hash, &frag, &found);
1372                 if (found) {
1373                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
1374                                 u8 r;
1375
1376                                 /* choose a random replica */
1377                                 get_random_bytes(&r, 1);
1378                                 r %= frag.ndist;
1379                                 mds = frag.dist[r];
1380                                 doutc(cl, "%p %llx.%llx frag %u mds%d (%d/%d)\n",
1381                                       inode, ceph_vinop(inode), frag.frag,
1382                                       mds, (int)r, frag.ndist);
1383                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1384                                     CEPH_MDS_STATE_ACTIVE &&
1385                                     !ceph_mdsmap_is_laggy(mdsc->mdsmap, mds))
1386                                         goto out;
1387                         }
1388
1389                         /* since this file/dir wasn't known to be
1390                          * replicated, then we want to look for the
1391                          * authoritative mds. */
1392                         if (frag.mds >= 0) {
1393                                 /* choose auth mds */
1394                                 mds = frag.mds;
1395                                 doutc(cl, "%p %llx.%llx frag %u mds%d (auth)\n",
1396                                       inode, ceph_vinop(inode), frag.frag, mds);
1397                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1398                                     CEPH_MDS_STATE_ACTIVE) {
1399                                         if (!ceph_mdsmap_is_laggy(mdsc->mdsmap,
1400                                                                   mds))
1401                                                 goto out;
1402                                 }
1403                         }
1404                         mode = USE_AUTH_MDS;
1405                 }
1406         }
1407
1408         spin_lock(&ci->i_ceph_lock);
1409         cap = NULL;
1410         if (mode == USE_AUTH_MDS)
1411                 cap = ci->i_auth_cap;
1412         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
1413                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
1414         if (!cap) {
1415                 spin_unlock(&ci->i_ceph_lock);
1416                 iput(inode);
1417                 goto random;
1418         }
1419         mds = cap->session->s_mds;
1420         doutc(cl, "%p %llx.%llx mds%d (%scap %p)\n", inode,
1421               ceph_vinop(inode), mds,
1422               cap == ci->i_auth_cap ? "auth " : "", cap);
1423         spin_unlock(&ci->i_ceph_lock);
1424 out:
1425         iput(inode);
1426         return mds;
1427
1428 random:
1429         if (random)
1430                 *random = true;
1431
1432         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
1433         doutc(cl, "chose random mds%d\n", mds);
1434         return mds;
1435 }
1436
1437
1438 /*
1439  * session messages
1440  */
1441 struct ceph_msg *ceph_create_session_msg(u32 op, u64 seq)
1442 {
1443         struct ceph_msg *msg;
1444         struct ceph_mds_session_head *h;
1445
1446         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
1447                            false);
1448         if (!msg) {
1449                 pr_err("ENOMEM creating session %s msg\n",
1450                        ceph_session_op_name(op));
1451                 return NULL;
1452         }
1453         h = msg->front.iov_base;
1454         h->op = cpu_to_le32(op);
1455         h->seq = cpu_to_le64(seq);
1456
1457         return msg;
1458 }
1459
1460 static const unsigned char feature_bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
1461 #define FEATURE_BYTES(c) (DIV_ROUND_UP((size_t)feature_bits[c - 1] + 1, 64) * 8)
1462 static int encode_supported_features(void **p, void *end)
1463 {
1464         static const size_t count = ARRAY_SIZE(feature_bits);
1465
1466         if (count > 0) {
1467                 size_t i;
1468                 size_t size = FEATURE_BYTES(count);
1469                 unsigned long bit;
1470
1471                 if (WARN_ON_ONCE(*p + 4 + size > end))
1472                         return -ERANGE;
1473
1474                 ceph_encode_32(p, size);
1475                 memset(*p, 0, size);
1476                 for (i = 0; i < count; i++) {
1477                         bit = feature_bits[i];
1478                         ((unsigned char *)(*p))[bit / 8] |= BIT(bit % 8);
1479                 }
1480                 *p += size;
1481         } else {
1482                 if (WARN_ON_ONCE(*p + 4 > end))
1483                         return -ERANGE;
1484
1485                 ceph_encode_32(p, 0);
1486         }
1487
1488         return 0;
1489 }
1490
1491 static const unsigned char metric_bits[] = CEPHFS_METRIC_SPEC_CLIENT_SUPPORTED;
1492 #define METRIC_BYTES(cnt) (DIV_ROUND_UP((size_t)metric_bits[cnt - 1] + 1, 64) * 8)
1493 static int encode_metric_spec(void **p, void *end)
1494 {
1495         static const size_t count = ARRAY_SIZE(metric_bits);
1496
1497         /* header */
1498         if (WARN_ON_ONCE(*p + 2 > end))
1499                 return -ERANGE;
1500
1501         ceph_encode_8(p, 1); /* version */
1502         ceph_encode_8(p, 1); /* compat */
1503
1504         if (count > 0) {
1505                 size_t i;
1506                 size_t size = METRIC_BYTES(count);
1507
1508                 if (WARN_ON_ONCE(*p + 4 + 4 + size > end))
1509                         return -ERANGE;
1510
1511                 /* metric spec info length */
1512                 ceph_encode_32(p, 4 + size);
1513
1514                 /* metric spec */
1515                 ceph_encode_32(p, size);
1516                 memset(*p, 0, size);
1517                 for (i = 0; i < count; i++)
1518                         ((unsigned char *)(*p))[i / 8] |= BIT(metric_bits[i] % 8);
1519                 *p += size;
1520         } else {
1521                 if (WARN_ON_ONCE(*p + 4 + 4 > end))
1522                         return -ERANGE;
1523
1524                 /* metric spec info length */
1525                 ceph_encode_32(p, 4);
1526                 /* metric spec */
1527                 ceph_encode_32(p, 0);
1528         }
1529
1530         return 0;
1531 }
1532
1533 /*
1534  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
1535  * to include additional client metadata fields.
1536  */
1537 static struct ceph_msg *
1538 create_session_full_msg(struct ceph_mds_client *mdsc, int op, u64 seq)
1539 {
1540         struct ceph_msg *msg;
1541         struct ceph_mds_session_head *h;
1542         int i;
1543         int extra_bytes = 0;
1544         int metadata_key_count = 0;
1545         struct ceph_options *opt = mdsc->fsc->client->options;
1546         struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1547         struct ceph_client *cl = mdsc->fsc->client;
1548         size_t size, count;
1549         void *p, *end;
1550         int ret;
1551
1552         const char* metadata[][2] = {
1553                 {"hostname", mdsc->nodename},
1554                 {"kernel_version", init_utsname()->release},
1555                 {"entity_id", opt->name ? : ""},
1556                 {"root", fsopt->server_path ? : "/"},
1557                 {NULL, NULL}
1558         };
1559
1560         /* Calculate serialized length of metadata */
1561         extra_bytes = 4;  /* map length */
1562         for (i = 0; metadata[i][0]; ++i) {
1563                 extra_bytes += 8 + strlen(metadata[i][0]) +
1564                         strlen(metadata[i][1]);
1565                 metadata_key_count++;
1566         }
1567
1568         /* supported feature */
1569         size = 0;
1570         count = ARRAY_SIZE(feature_bits);
1571         if (count > 0)
1572                 size = FEATURE_BYTES(count);
1573         extra_bytes += 4 + size;
1574
1575         /* metric spec */
1576         size = 0;
1577         count = ARRAY_SIZE(metric_bits);
1578         if (count > 0)
1579                 size = METRIC_BYTES(count);
1580         extra_bytes += 2 + 4 + 4 + size;
1581
1582         /* flags, mds auth caps and oldest_client_tid */
1583         extra_bytes += 4 + 4 + 8;
1584
1585         /* Allocate the message */
1586         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1587                            GFP_NOFS, false);
1588         if (!msg) {
1589                 pr_err_client(cl, "ENOMEM creating session open msg\n");
1590                 return ERR_PTR(-ENOMEM);
1591         }
1592         p = msg->front.iov_base;
1593         end = p + msg->front.iov_len;
1594
1595         h = p;
1596         h->op = cpu_to_le32(op);
1597         h->seq = cpu_to_le64(seq);
1598
1599         /*
1600          * Serialize client metadata into waiting buffer space, using
1601          * the format that userspace expects for map<string, string>
1602          *
1603          * ClientSession messages with metadata are v7
1604          */
1605         msg->hdr.version = cpu_to_le16(7);
1606         msg->hdr.compat_version = cpu_to_le16(1);
1607
1608         /* The write pointer, following the session_head structure */
1609         p += sizeof(*h);
1610
1611         /* Number of entries in the map */
1612         ceph_encode_32(&p, metadata_key_count);
1613
1614         /* Two length-prefixed strings for each entry in the map */
1615         for (i = 0; metadata[i][0]; ++i) {
1616                 size_t const key_len = strlen(metadata[i][0]);
1617                 size_t const val_len = strlen(metadata[i][1]);
1618
1619                 ceph_encode_32(&p, key_len);
1620                 memcpy(p, metadata[i][0], key_len);
1621                 p += key_len;
1622                 ceph_encode_32(&p, val_len);
1623                 memcpy(p, metadata[i][1], val_len);
1624                 p += val_len;
1625         }
1626
1627         ret = encode_supported_features(&p, end);
1628         if (ret) {
1629                 pr_err_client(cl, "encode_supported_features failed!\n");
1630                 ceph_msg_put(msg);
1631                 return ERR_PTR(ret);
1632         }
1633
1634         ret = encode_metric_spec(&p, end);
1635         if (ret) {
1636                 pr_err_client(cl, "encode_metric_spec failed!\n");
1637                 ceph_msg_put(msg);
1638                 return ERR_PTR(ret);
1639         }
1640
1641         /* version == 5, flags */
1642         ceph_encode_32(&p, 0);
1643
1644         /* version == 6, mds auth caps */
1645         ceph_encode_32(&p, 0);
1646
1647         /* version == 7, oldest_client_tid */
1648         ceph_encode_64(&p, mdsc->oldest_tid);
1649
1650         msg->front.iov_len = p - msg->front.iov_base;
1651         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1652
1653         return msg;
1654 }
1655
1656 /*
1657  * send session open request.
1658  *
1659  * called under mdsc->mutex
1660  */
1661 static int __open_session(struct ceph_mds_client *mdsc,
1662                           struct ceph_mds_session *session)
1663 {
1664         struct ceph_msg *msg;
1665         int mstate;
1666         int mds = session->s_mds;
1667
1668         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_FENCE_IO)
1669                 return -EIO;
1670
1671         /* wait for mds to go active? */
1672         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1673         doutc(mdsc->fsc->client, "open_session to mds%d (%s)\n", mds,
1674               ceph_mds_state_name(mstate));
1675         session->s_state = CEPH_MDS_SESSION_OPENING;
1676         session->s_renew_requested = jiffies;
1677
1678         /* send connect message */
1679         msg = create_session_full_msg(mdsc, CEPH_SESSION_REQUEST_OPEN,
1680                                       session->s_seq);
1681         if (IS_ERR(msg))
1682                 return PTR_ERR(msg);
1683         ceph_con_send(&session->s_con, msg);
1684         return 0;
1685 }
1686
1687 /*
1688  * open sessions for any export targets for the given mds
1689  *
1690  * called under mdsc->mutex
1691  */
1692 static struct ceph_mds_session *
1693 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1694 {
1695         struct ceph_mds_session *session;
1696         int ret;
1697
1698         session = __ceph_lookup_mds_session(mdsc, target);
1699         if (!session) {
1700                 session = register_session(mdsc, target);
1701                 if (IS_ERR(session))
1702                         return session;
1703         }
1704         if (session->s_state == CEPH_MDS_SESSION_NEW ||
1705             session->s_state == CEPH_MDS_SESSION_CLOSING) {
1706                 ret = __open_session(mdsc, session);
1707                 if (ret)
1708                         return ERR_PTR(ret);
1709         }
1710
1711         return session;
1712 }
1713
1714 struct ceph_mds_session *
1715 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1716 {
1717         struct ceph_mds_session *session;
1718         struct ceph_client *cl = mdsc->fsc->client;
1719
1720         doutc(cl, "to mds%d\n", target);
1721
1722         mutex_lock(&mdsc->mutex);
1723         session = __open_export_target_session(mdsc, target);
1724         mutex_unlock(&mdsc->mutex);
1725
1726         return session;
1727 }
1728
1729 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1730                                           struct ceph_mds_session *session)
1731 {
1732         struct ceph_mds_info *mi;
1733         struct ceph_mds_session *ts;
1734         int i, mds = session->s_mds;
1735         struct ceph_client *cl = mdsc->fsc->client;
1736
1737         if (mds >= mdsc->mdsmap->possible_max_rank)
1738                 return;
1739
1740         mi = &mdsc->mdsmap->m_info[mds];
1741         doutc(cl, "for mds%d (%d targets)\n", session->s_mds,
1742               mi->num_export_targets);
1743
1744         for (i = 0; i < mi->num_export_targets; i++) {
1745                 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1746                 ceph_put_mds_session(ts);
1747         }
1748 }
1749
1750 /*
1751  * session caps
1752  */
1753
1754 static void detach_cap_releases(struct ceph_mds_session *session,
1755                                 struct list_head *target)
1756 {
1757         struct ceph_client *cl = session->s_mdsc->fsc->client;
1758
1759         lockdep_assert_held(&session->s_cap_lock);
1760
1761         list_splice_init(&session->s_cap_releases, target);
1762         session->s_num_cap_releases = 0;
1763         doutc(cl, "mds%d\n", session->s_mds);
1764 }
1765
1766 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1767                                  struct list_head *dispose)
1768 {
1769         while (!list_empty(dispose)) {
1770                 struct ceph_cap *cap;
1771                 /* zero out the in-progress message */
1772                 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1773                 list_del(&cap->session_caps);
1774                 ceph_put_cap(mdsc, cap);
1775         }
1776 }
1777
1778 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1779                                      struct ceph_mds_session *session)
1780 {
1781         struct ceph_client *cl = mdsc->fsc->client;
1782         struct ceph_mds_request *req;
1783         struct rb_node *p;
1784
1785         doutc(cl, "mds%d\n", session->s_mds);
1786         mutex_lock(&mdsc->mutex);
1787         while (!list_empty(&session->s_unsafe)) {
1788                 req = list_first_entry(&session->s_unsafe,
1789                                        struct ceph_mds_request, r_unsafe_item);
1790                 pr_warn_ratelimited_client(cl, " dropping unsafe request %llu\n",
1791                                            req->r_tid);
1792                 if (req->r_target_inode)
1793                         mapping_set_error(req->r_target_inode->i_mapping, -EIO);
1794                 if (req->r_unsafe_dir)
1795                         mapping_set_error(req->r_unsafe_dir->i_mapping, -EIO);
1796                 __unregister_request(mdsc, req);
1797         }
1798         /* zero r_attempts, so kick_requests() will re-send requests */
1799         p = rb_first(&mdsc->request_tree);
1800         while (p) {
1801                 req = rb_entry(p, struct ceph_mds_request, r_node);
1802                 p = rb_next(p);
1803                 if (req->r_session &&
1804                     req->r_session->s_mds == session->s_mds)
1805                         req->r_attempts = 0;
1806         }
1807         mutex_unlock(&mdsc->mutex);
1808 }
1809
1810 /*
1811  * Helper to safely iterate over all caps associated with a session, with
1812  * special care taken to handle a racing __ceph_remove_cap().
1813  *
1814  * Caller must hold session s_mutex.
1815  */
1816 int ceph_iterate_session_caps(struct ceph_mds_session *session,
1817                               int (*cb)(struct inode *, int mds, void *),
1818                               void *arg)
1819 {
1820         struct ceph_client *cl = session->s_mdsc->fsc->client;
1821         struct list_head *p;
1822         struct ceph_cap *cap;
1823         struct inode *inode, *last_inode = NULL;
1824         struct ceph_cap *old_cap = NULL;
1825         int ret;
1826
1827         doutc(cl, "%p mds%d\n", session, session->s_mds);
1828         spin_lock(&session->s_cap_lock);
1829         p = session->s_caps.next;
1830         while (p != &session->s_caps) {
1831                 int mds;
1832
1833                 cap = list_entry(p, struct ceph_cap, session_caps);
1834                 inode = igrab(&cap->ci->netfs.inode);
1835                 if (!inode) {
1836                         p = p->next;
1837                         continue;
1838                 }
1839                 session->s_cap_iterator = cap;
1840                 mds = cap->mds;
1841                 spin_unlock(&session->s_cap_lock);
1842
1843                 if (last_inode) {
1844                         iput(last_inode);
1845                         last_inode = NULL;
1846                 }
1847                 if (old_cap) {
1848                         ceph_put_cap(session->s_mdsc, old_cap);
1849                         old_cap = NULL;
1850                 }
1851
1852                 ret = cb(inode, mds, arg);
1853                 last_inode = inode;
1854
1855                 spin_lock(&session->s_cap_lock);
1856                 p = p->next;
1857                 if (!cap->ci) {
1858                         doutc(cl, "finishing cap %p removal\n", cap);
1859                         BUG_ON(cap->session != session);
1860                         cap->session = NULL;
1861                         list_del_init(&cap->session_caps);
1862                         session->s_nr_caps--;
1863                         atomic64_dec(&session->s_mdsc->metric.total_caps);
1864                         if (cap->queue_release)
1865                                 __ceph_queue_cap_release(session, cap);
1866                         else
1867                                 old_cap = cap;  /* put_cap it w/o locks held */
1868                 }
1869                 if (ret < 0)
1870                         goto out;
1871         }
1872         ret = 0;
1873 out:
1874         session->s_cap_iterator = NULL;
1875         spin_unlock(&session->s_cap_lock);
1876
1877         iput(last_inode);
1878         if (old_cap)
1879                 ceph_put_cap(session->s_mdsc, old_cap);
1880
1881         return ret;
1882 }
1883
1884 static int remove_session_caps_cb(struct inode *inode, int mds, void *arg)
1885 {
1886         struct ceph_inode_info *ci = ceph_inode(inode);
1887         struct ceph_client *cl = ceph_inode_to_client(inode);
1888         bool invalidate = false;
1889         struct ceph_cap *cap;
1890         int iputs = 0;
1891
1892         spin_lock(&ci->i_ceph_lock);
1893         cap = __get_cap_for_mds(ci, mds);
1894         if (cap) {
1895                 doutc(cl, " removing cap %p, ci is %p, inode is %p\n",
1896                       cap, ci, &ci->netfs.inode);
1897
1898                 iputs = ceph_purge_inode_cap(inode, cap, &invalidate);
1899         }
1900         spin_unlock(&ci->i_ceph_lock);
1901
1902         if (cap)
1903                 wake_up_all(&ci->i_cap_wq);
1904         if (invalidate)
1905                 ceph_queue_invalidate(inode);
1906         while (iputs--)
1907                 iput(inode);
1908         return 0;
1909 }
1910
1911 /*
1912  * caller must hold session s_mutex
1913  */
1914 static void remove_session_caps(struct ceph_mds_session *session)
1915 {
1916         struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1917         struct super_block *sb = fsc->sb;
1918         LIST_HEAD(dispose);
1919
1920         doutc(fsc->client, "on %p\n", session);
1921         ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1922
1923         wake_up_all(&fsc->mdsc->cap_flushing_wq);
1924
1925         spin_lock(&session->s_cap_lock);
1926         if (session->s_nr_caps > 0) {
1927                 struct inode *inode;
1928                 struct ceph_cap *cap, *prev = NULL;
1929                 struct ceph_vino vino;
1930                 /*
1931                  * iterate_session_caps() skips inodes that are being
1932                  * deleted, we need to wait until deletions are complete.
1933                  * __wait_on_freeing_inode() is designed for the job,
1934                  * but it is not exported, so use lookup inode function
1935                  * to access it.
1936                  */
1937                 while (!list_empty(&session->s_caps)) {
1938                         cap = list_entry(session->s_caps.next,
1939                                          struct ceph_cap, session_caps);
1940                         if (cap == prev)
1941                                 break;
1942                         prev = cap;
1943                         vino = cap->ci->i_vino;
1944                         spin_unlock(&session->s_cap_lock);
1945
1946                         inode = ceph_find_inode(sb, vino);
1947                         iput(inode);
1948
1949                         spin_lock(&session->s_cap_lock);
1950                 }
1951         }
1952
1953         // drop cap expires and unlock s_cap_lock
1954         detach_cap_releases(session, &dispose);
1955
1956         BUG_ON(session->s_nr_caps > 0);
1957         BUG_ON(!list_empty(&session->s_cap_flushing));
1958         spin_unlock(&session->s_cap_lock);
1959         dispose_cap_releases(session->s_mdsc, &dispose);
1960 }
1961
1962 enum {
1963         RECONNECT,
1964         RENEWCAPS,
1965         FORCE_RO,
1966 };
1967
1968 /*
1969  * wake up any threads waiting on this session's caps.  if the cap is
1970  * old (didn't get renewed on the client reconnect), remove it now.
1971  *
1972  * caller must hold s_mutex.
1973  */
1974 static int wake_up_session_cb(struct inode *inode, int mds, void *arg)
1975 {
1976         struct ceph_inode_info *ci = ceph_inode(inode);
1977         unsigned long ev = (unsigned long)arg;
1978
1979         if (ev == RECONNECT) {
1980                 spin_lock(&ci->i_ceph_lock);
1981                 ci->i_wanted_max_size = 0;
1982                 ci->i_requested_max_size = 0;
1983                 spin_unlock(&ci->i_ceph_lock);
1984         } else if (ev == RENEWCAPS) {
1985                 struct ceph_cap *cap;
1986
1987                 spin_lock(&ci->i_ceph_lock);
1988                 cap = __get_cap_for_mds(ci, mds);
1989                 /* mds did not re-issue stale cap */
1990                 if (cap && cap->cap_gen < atomic_read(&cap->session->s_cap_gen))
1991                         cap->issued = cap->implemented = CEPH_CAP_PIN;
1992                 spin_unlock(&ci->i_ceph_lock);
1993         } else if (ev == FORCE_RO) {
1994         }
1995         wake_up_all(&ci->i_cap_wq);
1996         return 0;
1997 }
1998
1999 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
2000 {
2001         struct ceph_client *cl = session->s_mdsc->fsc->client;
2002
2003         doutc(cl, "session %p mds%d\n", session, session->s_mds);
2004         ceph_iterate_session_caps(session, wake_up_session_cb,
2005                                   (void *)(unsigned long)ev);
2006 }
2007
2008 /*
2009  * Send periodic message to MDS renewing all currently held caps.  The
2010  * ack will reset the expiration for all caps from this session.
2011  *
2012  * caller holds s_mutex
2013  */
2014 static int send_renew_caps(struct ceph_mds_client *mdsc,
2015                            struct ceph_mds_session *session)
2016 {
2017         struct ceph_client *cl = mdsc->fsc->client;
2018         struct ceph_msg *msg;
2019         int state;
2020
2021         if (time_after_eq(jiffies, session->s_cap_ttl) &&
2022             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
2023                 pr_info_client(cl, "mds%d caps stale\n", session->s_mds);
2024         session->s_renew_requested = jiffies;
2025
2026         /* do not try to renew caps until a recovering mds has reconnected
2027          * with its clients. */
2028         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
2029         if (state < CEPH_MDS_STATE_RECONNECT) {
2030                 doutc(cl, "ignoring mds%d (%s)\n", session->s_mds,
2031                       ceph_mds_state_name(state));
2032                 return 0;
2033         }
2034
2035         doutc(cl, "to mds%d (%s)\n", session->s_mds,
2036               ceph_mds_state_name(state));
2037         msg = create_session_full_msg(mdsc, CEPH_SESSION_REQUEST_RENEWCAPS,
2038                                       ++session->s_renew_seq);
2039         if (IS_ERR(msg))
2040                 return PTR_ERR(msg);
2041         ceph_con_send(&session->s_con, msg);
2042         return 0;
2043 }
2044
2045 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
2046                              struct ceph_mds_session *session, u64 seq)
2047 {
2048         struct ceph_client *cl = mdsc->fsc->client;
2049         struct ceph_msg *msg;
2050
2051         doutc(cl, "to mds%d (%s)s seq %lld\n", session->s_mds,
2052               ceph_session_state_name(session->s_state), seq);
2053         msg = ceph_create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
2054         if (!msg)
2055                 return -ENOMEM;
2056         ceph_con_send(&session->s_con, msg);
2057         return 0;
2058 }
2059
2060
2061 /*
2062  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
2063  *
2064  * Called under session->s_mutex
2065  */
2066 static void renewed_caps(struct ceph_mds_client *mdsc,
2067                          struct ceph_mds_session *session, int is_renew)
2068 {
2069         struct ceph_client *cl = mdsc->fsc->client;
2070         int was_stale;
2071         int wake = 0;
2072
2073         spin_lock(&session->s_cap_lock);
2074         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
2075
2076         session->s_cap_ttl = session->s_renew_requested +
2077                 mdsc->mdsmap->m_session_timeout*HZ;
2078
2079         if (was_stale) {
2080                 if (time_before(jiffies, session->s_cap_ttl)) {
2081                         pr_info_client(cl, "mds%d caps renewed\n",
2082                                        session->s_mds);
2083                         wake = 1;
2084                 } else {
2085                         pr_info_client(cl, "mds%d caps still stale\n",
2086                                        session->s_mds);
2087                 }
2088         }
2089         doutc(cl, "mds%d ttl now %lu, was %s, now %s\n", session->s_mds,
2090               session->s_cap_ttl, was_stale ? "stale" : "fresh",
2091               time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
2092         spin_unlock(&session->s_cap_lock);
2093
2094         if (wake)
2095                 wake_up_session_caps(session, RENEWCAPS);
2096 }
2097
2098 /*
2099  * send a session close request
2100  */
2101 static int request_close_session(struct ceph_mds_session *session)
2102 {
2103         struct ceph_client *cl = session->s_mdsc->fsc->client;
2104         struct ceph_msg *msg;
2105
2106         doutc(cl, "mds%d state %s seq %lld\n", session->s_mds,
2107               ceph_session_state_name(session->s_state), session->s_seq);
2108         msg = ceph_create_session_msg(CEPH_SESSION_REQUEST_CLOSE,
2109                                       session->s_seq);
2110         if (!msg)
2111                 return -ENOMEM;
2112         ceph_con_send(&session->s_con, msg);
2113         return 1;
2114 }
2115
2116 /*
2117  * Called with s_mutex held.
2118  */
2119 static int __close_session(struct ceph_mds_client *mdsc,
2120                          struct ceph_mds_session *session)
2121 {
2122         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
2123                 return 0;
2124         session->s_state = CEPH_MDS_SESSION_CLOSING;
2125         return request_close_session(session);
2126 }
2127
2128 static bool drop_negative_children(struct dentry *dentry)
2129 {
2130         struct dentry *child;
2131         bool all_negative = true;
2132
2133         if (!d_is_dir(dentry))
2134                 goto out;
2135
2136         spin_lock(&dentry->d_lock);
2137         hlist_for_each_entry(child, &dentry->d_children, d_sib) {
2138                 if (d_really_is_positive(child)) {
2139                         all_negative = false;
2140                         break;
2141                 }
2142         }
2143         spin_unlock(&dentry->d_lock);
2144
2145         if (all_negative)
2146                 shrink_dcache_parent(dentry);
2147 out:
2148         return all_negative;
2149 }
2150
2151 /*
2152  * Trim old(er) caps.
2153  *
2154  * Because we can't cache an inode without one or more caps, we do
2155  * this indirectly: if a cap is unused, we prune its aliases, at which
2156  * point the inode will hopefully get dropped to.
2157  *
2158  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
2159  * memory pressure from the MDS, though, so it needn't be perfect.
2160  */
2161 static int trim_caps_cb(struct inode *inode, int mds, void *arg)
2162 {
2163         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
2164         struct ceph_client *cl = mdsc->fsc->client;
2165         int *remaining = arg;
2166         struct ceph_inode_info *ci = ceph_inode(inode);
2167         int used, wanted, oissued, mine;
2168         struct ceph_cap *cap;
2169
2170         if (*remaining <= 0)
2171                 return -1;
2172
2173         spin_lock(&ci->i_ceph_lock);
2174         cap = __get_cap_for_mds(ci, mds);
2175         if (!cap) {
2176                 spin_unlock(&ci->i_ceph_lock);
2177                 return 0;
2178         }
2179         mine = cap->issued | cap->implemented;
2180         used = __ceph_caps_used(ci);
2181         wanted = __ceph_caps_file_wanted(ci);
2182         oissued = __ceph_caps_issued_other(ci, cap);
2183
2184         doutc(cl, "%p %llx.%llx cap %p mine %s oissued %s used %s wanted %s\n",
2185               inode, ceph_vinop(inode), cap, ceph_cap_string(mine),
2186               ceph_cap_string(oissued), ceph_cap_string(used),
2187               ceph_cap_string(wanted));
2188         if (cap == ci->i_auth_cap) {
2189                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
2190                     !list_empty(&ci->i_cap_snaps))
2191                         goto out;
2192                 if ((used | wanted) & CEPH_CAP_ANY_WR)
2193                         goto out;
2194                 /* Note: it's possible that i_filelock_ref becomes non-zero
2195                  * after dropping auth caps. It doesn't hurt because reply
2196                  * of lock mds request will re-add auth caps. */
2197                 if (atomic_read(&ci->i_filelock_ref) > 0)
2198                         goto out;
2199         }
2200         /* The inode has cached pages, but it's no longer used.
2201          * we can safely drop it */
2202         if (S_ISREG(inode->i_mode) &&
2203             wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
2204             !(oissued & CEPH_CAP_FILE_CACHE)) {
2205           used = 0;
2206           oissued = 0;
2207         }
2208         if ((used | wanted) & ~oissued & mine)
2209                 goto out;   /* we need these caps */
2210
2211         if (oissued) {
2212                 /* we aren't the only cap.. just remove us */
2213                 ceph_remove_cap(mdsc, cap, true);
2214                 (*remaining)--;
2215         } else {
2216                 struct dentry *dentry;
2217                 /* try dropping referring dentries */
2218                 spin_unlock(&ci->i_ceph_lock);
2219                 dentry = d_find_any_alias(inode);
2220                 if (dentry && drop_negative_children(dentry)) {
2221                         int count;
2222                         dput(dentry);
2223                         d_prune_aliases(inode);
2224                         count = atomic_read(&inode->i_count);
2225                         if (count == 1)
2226                                 (*remaining)--;
2227                         doutc(cl, "%p %llx.%llx cap %p pruned, count now %d\n",
2228                               inode, ceph_vinop(inode), cap, count);
2229                 } else {
2230                         dput(dentry);
2231                 }
2232                 return 0;
2233         }
2234
2235 out:
2236         spin_unlock(&ci->i_ceph_lock);
2237         return 0;
2238 }
2239
2240 /*
2241  * Trim session cap count down to some max number.
2242  */
2243 int ceph_trim_caps(struct ceph_mds_client *mdsc,
2244                    struct ceph_mds_session *session,
2245                    int max_caps)
2246 {
2247         struct ceph_client *cl = mdsc->fsc->client;
2248         int trim_caps = session->s_nr_caps - max_caps;
2249
2250         doutc(cl, "mds%d start: %d / %d, trim %d\n", session->s_mds,
2251               session->s_nr_caps, max_caps, trim_caps);
2252         if (trim_caps > 0) {
2253                 int remaining = trim_caps;
2254
2255                 ceph_iterate_session_caps(session, trim_caps_cb, &remaining);
2256                 doutc(cl, "mds%d done: %d / %d, trimmed %d\n",
2257                       session->s_mds, session->s_nr_caps, max_caps,
2258                       trim_caps - remaining);
2259         }
2260
2261         ceph_flush_session_cap_releases(mdsc, session);
2262         return 0;
2263 }
2264
2265 static int check_caps_flush(struct ceph_mds_client *mdsc,
2266                             u64 want_flush_tid)
2267 {
2268         struct ceph_client *cl = mdsc->fsc->client;
2269         int ret = 1;
2270
2271         spin_lock(&mdsc->cap_dirty_lock);
2272         if (!list_empty(&mdsc->cap_flush_list)) {
2273                 struct ceph_cap_flush *cf =
2274                         list_first_entry(&mdsc->cap_flush_list,
2275                                          struct ceph_cap_flush, g_list);
2276                 if (cf->tid <= want_flush_tid) {
2277                         doutc(cl, "still flushing tid %llu <= %llu\n",
2278                               cf->tid, want_flush_tid);
2279                         ret = 0;
2280                 }
2281         }
2282         spin_unlock(&mdsc->cap_dirty_lock);
2283         return ret;
2284 }
2285
2286 /*
2287  * flush all dirty inode data to disk.
2288  *
2289  * returns true if we've flushed through want_flush_tid
2290  */
2291 static void wait_caps_flush(struct ceph_mds_client *mdsc,
2292                             u64 want_flush_tid)
2293 {
2294         struct ceph_client *cl = mdsc->fsc->client;
2295
2296         doutc(cl, "want %llu\n", want_flush_tid);
2297
2298         wait_event(mdsc->cap_flushing_wq,
2299                    check_caps_flush(mdsc, want_flush_tid));
2300
2301         doutc(cl, "ok, flushed thru %llu\n", want_flush_tid);
2302 }
2303
2304 /*
2305  * called under s_mutex
2306  */
2307 static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
2308                                    struct ceph_mds_session *session)
2309 {
2310         struct ceph_client *cl = mdsc->fsc->client;
2311         struct ceph_msg *msg = NULL;
2312         struct ceph_mds_cap_release *head;
2313         struct ceph_mds_cap_item *item;
2314         struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
2315         struct ceph_cap *cap;
2316         LIST_HEAD(tmp_list);
2317         int num_cap_releases;
2318         __le32  barrier, *cap_barrier;
2319
2320         down_read(&osdc->lock);
2321         barrier = cpu_to_le32(osdc->epoch_barrier);
2322         up_read(&osdc->lock);
2323
2324         spin_lock(&session->s_cap_lock);
2325 again:
2326         list_splice_init(&session->s_cap_releases, &tmp_list);
2327         num_cap_releases = session->s_num_cap_releases;
2328         session->s_num_cap_releases = 0;
2329         spin_unlock(&session->s_cap_lock);
2330
2331         while (!list_empty(&tmp_list)) {
2332                 if (!msg) {
2333                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
2334                                         PAGE_SIZE, GFP_NOFS, false);
2335                         if (!msg)
2336                                 goto out_err;
2337                         head = msg->front.iov_base;
2338                         head->num = cpu_to_le32(0);
2339                         msg->front.iov_len = sizeof(*head);
2340
2341                         msg->hdr.version = cpu_to_le16(2);
2342                         msg->hdr.compat_version = cpu_to_le16(1);
2343                 }
2344
2345                 cap = list_first_entry(&tmp_list, struct ceph_cap,
2346                                         session_caps);
2347                 list_del(&cap->session_caps);
2348                 num_cap_releases--;
2349
2350                 head = msg->front.iov_base;
2351                 put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
2352                                    &head->num);
2353                 item = msg->front.iov_base + msg->front.iov_len;
2354                 item->ino = cpu_to_le64(cap->cap_ino);
2355                 item->cap_id = cpu_to_le64(cap->cap_id);
2356                 item->migrate_seq = cpu_to_le32(cap->mseq);
2357                 item->issue_seq = cpu_to_le32(cap->issue_seq);
2358                 msg->front.iov_len += sizeof(*item);
2359
2360                 ceph_put_cap(mdsc, cap);
2361
2362                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
2363                         // Append cap_barrier field
2364                         cap_barrier = msg->front.iov_base + msg->front.iov_len;
2365                         *cap_barrier = barrier;
2366                         msg->front.iov_len += sizeof(*cap_barrier);
2367
2368                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2369                         doutc(cl, "mds%d %p\n", session->s_mds, msg);
2370                         ceph_con_send(&session->s_con, msg);
2371                         msg = NULL;
2372                 }
2373         }
2374
2375         BUG_ON(num_cap_releases != 0);
2376
2377         spin_lock(&session->s_cap_lock);
2378         if (!list_empty(&session->s_cap_releases))
2379                 goto again;
2380         spin_unlock(&session->s_cap_lock);
2381
2382         if (msg) {
2383                 // Append cap_barrier field
2384                 cap_barrier = msg->front.iov_base + msg->front.iov_len;
2385                 *cap_barrier = barrier;
2386                 msg->front.iov_len += sizeof(*cap_barrier);
2387
2388                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2389                 doutc(cl, "mds%d %p\n", session->s_mds, msg);
2390                 ceph_con_send(&session->s_con, msg);
2391         }
2392         return;
2393 out_err:
2394         pr_err_client(cl, "mds%d, failed to allocate message\n",
2395                       session->s_mds);
2396         spin_lock(&session->s_cap_lock);
2397         list_splice(&tmp_list, &session->s_cap_releases);
2398         session->s_num_cap_releases += num_cap_releases;
2399         spin_unlock(&session->s_cap_lock);
2400 }
2401
2402 static void ceph_cap_release_work(struct work_struct *work)
2403 {
2404         struct ceph_mds_session *session =
2405                 container_of(work, struct ceph_mds_session, s_cap_release_work);
2406
2407         mutex_lock(&session->s_mutex);
2408         if (session->s_state == CEPH_MDS_SESSION_OPEN ||
2409             session->s_state == CEPH_MDS_SESSION_HUNG)
2410                 ceph_send_cap_releases(session->s_mdsc, session);
2411         mutex_unlock(&session->s_mutex);
2412         ceph_put_mds_session(session);
2413 }
2414
2415 void ceph_flush_session_cap_releases(struct ceph_mds_client *mdsc,
2416                              struct ceph_mds_session *session)
2417 {
2418         struct ceph_client *cl = mdsc->fsc->client;
2419         if (mdsc->stopping)
2420                 return;
2421
2422         ceph_get_mds_session(session);
2423         if (queue_work(mdsc->fsc->cap_wq,
2424                        &session->s_cap_release_work)) {
2425                 doutc(cl, "cap release work queued\n");
2426         } else {
2427                 ceph_put_mds_session(session);
2428                 doutc(cl, "failed to queue cap release work\n");
2429         }
2430 }
2431
2432 /*
2433  * caller holds session->s_cap_lock
2434  */
2435 void __ceph_queue_cap_release(struct ceph_mds_session *session,
2436                               struct ceph_cap *cap)
2437 {
2438         list_add_tail(&cap->session_caps, &session->s_cap_releases);
2439         session->s_num_cap_releases++;
2440
2441         if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
2442                 ceph_flush_session_cap_releases(session->s_mdsc, session);
2443 }
2444
2445 static void ceph_cap_reclaim_work(struct work_struct *work)
2446 {
2447         struct ceph_mds_client *mdsc =
2448                 container_of(work, struct ceph_mds_client, cap_reclaim_work);
2449         int ret = ceph_trim_dentries(mdsc);
2450         if (ret == -EAGAIN)
2451                 ceph_queue_cap_reclaim_work(mdsc);
2452 }
2453
2454 void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
2455 {
2456         struct ceph_client *cl = mdsc->fsc->client;
2457         if (mdsc->stopping)
2458                 return;
2459
2460         if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
2461                 doutc(cl, "caps reclaim work queued\n");
2462         } else {
2463                 doutc(cl, "failed to queue caps release work\n");
2464         }
2465 }
2466
2467 void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
2468 {
2469         int val;
2470         if (!nr)
2471                 return;
2472         val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
2473         if ((val % CEPH_CAPS_PER_RELEASE) < nr) {
2474                 atomic_set(&mdsc->cap_reclaim_pending, 0);
2475                 ceph_queue_cap_reclaim_work(mdsc);
2476         }
2477 }
2478
2479 void ceph_queue_cap_unlink_work(struct ceph_mds_client *mdsc)
2480 {
2481         struct ceph_client *cl = mdsc->fsc->client;
2482         if (mdsc->stopping)
2483                 return;
2484
2485         if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_unlink_work)) {
2486                 doutc(cl, "caps unlink work queued\n");
2487         } else {
2488                 doutc(cl, "failed to queue caps unlink work\n");
2489         }
2490 }
2491
2492 static void ceph_cap_unlink_work(struct work_struct *work)
2493 {
2494         struct ceph_mds_client *mdsc =
2495                 container_of(work, struct ceph_mds_client, cap_unlink_work);
2496         struct ceph_client *cl = mdsc->fsc->client;
2497
2498         doutc(cl, "begin\n");
2499         spin_lock(&mdsc->cap_delay_lock);
2500         while (!list_empty(&mdsc->cap_unlink_delay_list)) {
2501                 struct ceph_inode_info *ci;
2502                 struct inode *inode;
2503
2504                 ci = list_first_entry(&mdsc->cap_unlink_delay_list,
2505                                       struct ceph_inode_info,
2506                                       i_cap_delay_list);
2507                 list_del_init(&ci->i_cap_delay_list);
2508
2509                 inode = igrab(&ci->netfs.inode);
2510                 if (inode) {
2511                         spin_unlock(&mdsc->cap_delay_lock);
2512                         doutc(cl, "on %p %llx.%llx\n", inode,
2513                               ceph_vinop(inode));
2514                         ceph_check_caps(ci, CHECK_CAPS_FLUSH);
2515                         iput(inode);
2516                         spin_lock(&mdsc->cap_delay_lock);
2517                 }
2518         }
2519         spin_unlock(&mdsc->cap_delay_lock);
2520         doutc(cl, "done\n");
2521 }
2522
2523 /*
2524  * requests
2525  */
2526
2527 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
2528                                     struct inode *dir)
2529 {
2530         struct ceph_inode_info *ci = ceph_inode(dir);
2531         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2532         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2533         size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2534         unsigned int num_entries;
2535         int order;
2536
2537         spin_lock(&ci->i_ceph_lock);
2538         num_entries = ci->i_files + ci->i_subdirs;
2539         spin_unlock(&ci->i_ceph_lock);
2540         num_entries = max(num_entries, 1U);
2541         num_entries = min(num_entries, opt->max_readdir);
2542
2543         order = get_order(size * num_entries);
2544         while (order >= 0) {
2545                 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2546                                                              __GFP_NOWARN |
2547                                                              __GFP_ZERO,
2548                                                              order);
2549                 if (rinfo->dir_entries)
2550                         break;
2551                 order--;
2552         }
2553         if (!rinfo->dir_entries)
2554                 return -ENOMEM;
2555
2556         num_entries = (PAGE_SIZE << order) / size;
2557         num_entries = min(num_entries, opt->max_readdir);
2558
2559         rinfo->dir_buf_size = PAGE_SIZE << order;
2560         req->r_num_caps = num_entries + 1;
2561         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2562         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2563         return 0;
2564 }
2565
2566 /*
2567  * Create an mds request.
2568  */
2569 struct ceph_mds_request *
2570 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2571 {
2572         struct ceph_mds_request *req;
2573
2574         req = kmem_cache_zalloc(ceph_mds_request_cachep, GFP_NOFS);
2575         if (!req)
2576                 return ERR_PTR(-ENOMEM);
2577
2578         mutex_init(&req->r_fill_mutex);
2579         req->r_mdsc = mdsc;
2580         req->r_started = jiffies;
2581         req->r_start_latency = ktime_get();
2582         req->r_resend_mds = -1;
2583         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2584         INIT_LIST_HEAD(&req->r_unsafe_target_item);
2585         req->r_fmode = -1;
2586         req->r_feature_needed = -1;
2587         kref_init(&req->r_kref);
2588         RB_CLEAR_NODE(&req->r_node);
2589         INIT_LIST_HEAD(&req->r_wait);
2590         init_completion(&req->r_completion);
2591         init_completion(&req->r_safe_completion);
2592         INIT_LIST_HEAD(&req->r_unsafe_item);
2593
2594         ktime_get_coarse_real_ts64(&req->r_stamp);
2595
2596         req->r_op = op;
2597         req->r_direct_mode = mode;
2598         return req;
2599 }
2600
2601 /*
2602  * return oldest (lowest) request, tid in request tree, 0 if none.
2603  *
2604  * called under mdsc->mutex.
2605  */
2606 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2607 {
2608         if (RB_EMPTY_ROOT(&mdsc->request_tree))
2609                 return NULL;
2610         return rb_entry(rb_first(&mdsc->request_tree),
2611                         struct ceph_mds_request, r_node);
2612 }
2613
2614 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2615 {
2616         return mdsc->oldest_tid;
2617 }
2618
2619 #if IS_ENABLED(CONFIG_FS_ENCRYPTION)
2620 static u8 *get_fscrypt_altname(const struct ceph_mds_request *req, u32 *plen)
2621 {
2622         struct inode *dir = req->r_parent;
2623         struct dentry *dentry = req->r_dentry;
2624         const struct qstr *name = req->r_dname;
2625         u8 *cryptbuf = NULL;
2626         u32 len = 0;
2627         int ret = 0;
2628
2629         /* only encode if we have parent and dentry */
2630         if (!dir || !dentry)
2631                 goto success;
2632
2633         /* No-op unless this is encrypted */
2634         if (!IS_ENCRYPTED(dir))
2635                 goto success;
2636
2637         ret = ceph_fscrypt_prepare_readdir(dir);
2638         if (ret < 0)
2639                 return ERR_PTR(ret);
2640
2641         /* No key? Just ignore it. */
2642         if (!fscrypt_has_encryption_key(dir))
2643                 goto success;
2644
2645         if (!name)
2646                 name = &dentry->d_name;
2647
2648         if (!fscrypt_fname_encrypted_size(dir, name->len, NAME_MAX, &len)) {
2649                 WARN_ON_ONCE(1);
2650                 return ERR_PTR(-ENAMETOOLONG);
2651         }
2652
2653         /* No need to append altname if name is short enough */
2654         if (len <= CEPH_NOHASH_NAME_MAX) {
2655                 len = 0;
2656                 goto success;
2657         }
2658
2659         cryptbuf = kmalloc(len, GFP_KERNEL);
2660         if (!cryptbuf)
2661                 return ERR_PTR(-ENOMEM);
2662
2663         ret = fscrypt_fname_encrypt(dir, name, cryptbuf, len);
2664         if (ret) {
2665                 kfree(cryptbuf);
2666                 return ERR_PTR(ret);
2667         }
2668 success:
2669         *plen = len;
2670         return cryptbuf;
2671 }
2672 #else
2673 static u8 *get_fscrypt_altname(const struct ceph_mds_request *req, u32 *plen)
2674 {
2675         *plen = 0;
2676         return NULL;
2677 }
2678 #endif
2679
2680 /**
2681  * ceph_mdsc_build_path - build a path string to a given dentry
2682  * @mdsc: mds client
2683  * @dentry: dentry to which path should be built
2684  * @plen: returned length of string
2685  * @pbase: returned base inode number
2686  * @for_wire: is this path going to be sent to the MDS?
2687  *
2688  * Build a string that represents the path to the dentry. This is mostly called
2689  * for two different purposes:
2690  *
2691  * 1) we need to build a path string to send to the MDS (for_wire == true)
2692  * 2) we need a path string for local presentation (e.g. debugfs)
2693  *    (for_wire == false)
2694  *
2695  * The path is built in reverse, starting with the dentry. Walk back up toward
2696  * the root, building the path until the first non-snapped inode is reached
2697  * (for_wire) or the root inode is reached (!for_wire).
2698  *
2699  * Encode hidden .snap dirs as a double /, i.e.
2700  *   foo/.snap/bar -> foo//bar
2701  */
2702 char *ceph_mdsc_build_path(struct ceph_mds_client *mdsc, struct dentry *dentry,
2703                            int *plen, u64 *pbase, int for_wire)
2704 {
2705         struct ceph_client *cl = mdsc->fsc->client;
2706         struct dentry *cur;
2707         struct inode *inode;
2708         char *path;
2709         int pos;
2710         unsigned seq;
2711         u64 base;
2712
2713         if (!dentry)
2714                 return ERR_PTR(-EINVAL);
2715
2716         path = __getname();
2717         if (!path)
2718                 return ERR_PTR(-ENOMEM);
2719 retry:
2720         pos = PATH_MAX - 1;
2721         path[pos] = '\0';
2722
2723         seq = read_seqbegin(&rename_lock);
2724         cur = dget(dentry);
2725         for (;;) {
2726                 struct dentry *parent;
2727
2728                 spin_lock(&cur->d_lock);
2729                 inode = d_inode(cur);
2730                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2731                         doutc(cl, "path+%d: %p SNAPDIR\n", pos, cur);
2732                         spin_unlock(&cur->d_lock);
2733                         parent = dget_parent(cur);
2734                 } else if (for_wire && inode && dentry != cur &&
2735                            ceph_snap(inode) == CEPH_NOSNAP) {
2736                         spin_unlock(&cur->d_lock);
2737                         pos++; /* get rid of any prepended '/' */
2738                         break;
2739                 } else if (!for_wire || !IS_ENCRYPTED(d_inode(cur->d_parent))) {
2740                         pos -= cur->d_name.len;
2741                         if (pos < 0) {
2742                                 spin_unlock(&cur->d_lock);
2743                                 break;
2744                         }
2745                         memcpy(path + pos, cur->d_name.name, cur->d_name.len);
2746                         spin_unlock(&cur->d_lock);
2747                         parent = dget_parent(cur);
2748                 } else {
2749                         int len, ret;
2750                         char buf[NAME_MAX];
2751
2752                         /*
2753                          * Proactively copy name into buf, in case we need to
2754                          * present it as-is.
2755                          */
2756                         memcpy(buf, cur->d_name.name, cur->d_name.len);
2757                         len = cur->d_name.len;
2758                         spin_unlock(&cur->d_lock);
2759                         parent = dget_parent(cur);
2760
2761                         ret = ceph_fscrypt_prepare_readdir(d_inode(parent));
2762                         if (ret < 0) {
2763                                 dput(parent);
2764                                 dput(cur);
2765                                 return ERR_PTR(ret);
2766                         }
2767
2768                         if (fscrypt_has_encryption_key(d_inode(parent))) {
2769                                 len = ceph_encode_encrypted_fname(d_inode(parent),
2770                                                                   cur, buf);
2771                                 if (len < 0) {
2772                                         dput(parent);
2773                                         dput(cur);
2774                                         return ERR_PTR(len);
2775                                 }
2776                         }
2777                         pos -= len;
2778                         if (pos < 0) {
2779                                 dput(parent);
2780                                 break;
2781                         }
2782                         memcpy(path + pos, buf, len);
2783                 }
2784                 dput(cur);
2785                 cur = parent;
2786
2787                 /* Are we at the root? */
2788                 if (IS_ROOT(cur))
2789                         break;
2790
2791                 /* Are we out of buffer? */
2792                 if (--pos < 0)
2793                         break;
2794
2795                 path[pos] = '/';
2796         }
2797         inode = d_inode(cur);
2798         base = inode ? ceph_ino(inode) : 0;
2799         dput(cur);
2800
2801         if (read_seqretry(&rename_lock, seq))
2802                 goto retry;
2803
2804         if (pos < 0) {
2805                 /*
2806                  * The path is longer than PATH_MAX and this function
2807                  * cannot ever succeed.  Creating paths that long is
2808                  * possible with Ceph, but Linux cannot use them.
2809                  */
2810                 return ERR_PTR(-ENAMETOOLONG);
2811         }
2812
2813         *pbase = base;
2814         *plen = PATH_MAX - 1 - pos;
2815         doutc(cl, "on %p %d built %llx '%.*s'\n", dentry, d_count(dentry),
2816               base, *plen, path + pos);
2817         return path + pos;
2818 }
2819
2820 static int build_dentry_path(struct ceph_mds_client *mdsc, struct dentry *dentry,
2821                              struct inode *dir, const char **ppath, int *ppathlen,
2822                              u64 *pino, bool *pfreepath, bool parent_locked)
2823 {
2824         char *path;
2825
2826         rcu_read_lock();
2827         if (!dir)
2828                 dir = d_inode_rcu(dentry->d_parent);
2829         if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP &&
2830             !IS_ENCRYPTED(dir)) {
2831                 *pino = ceph_ino(dir);
2832                 rcu_read_unlock();
2833                 *ppath = dentry->d_name.name;
2834                 *ppathlen = dentry->d_name.len;
2835                 return 0;
2836         }
2837         rcu_read_unlock();
2838         path = ceph_mdsc_build_path(mdsc, dentry, ppathlen, pino, 1);
2839         if (IS_ERR(path))
2840                 return PTR_ERR(path);
2841         *ppath = path;
2842         *pfreepath = true;
2843         return 0;
2844 }
2845
2846 static int build_inode_path(struct inode *inode,
2847                             const char **ppath, int *ppathlen, u64 *pino,
2848                             bool *pfreepath)
2849 {
2850         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
2851         struct dentry *dentry;
2852         char *path;
2853
2854         if (ceph_snap(inode) == CEPH_NOSNAP) {
2855                 *pino = ceph_ino(inode);
2856                 *ppathlen = 0;
2857                 return 0;
2858         }
2859         dentry = d_find_alias(inode);
2860         path = ceph_mdsc_build_path(mdsc, dentry, ppathlen, pino, 1);
2861         dput(dentry);
2862         if (IS_ERR(path))
2863                 return PTR_ERR(path);
2864         *ppath = path;
2865         *pfreepath = true;
2866         return 0;
2867 }
2868
2869 /*
2870  * request arguments may be specified via an inode *, a dentry *, or
2871  * an explicit ino+path.
2872  */
2873 static int set_request_path_attr(struct ceph_mds_client *mdsc, struct inode *rinode,
2874                                  struct dentry *rdentry, struct inode *rdiri,
2875                                  const char *rpath, u64 rino, const char **ppath,
2876                                  int *pathlen, u64 *ino, bool *freepath,
2877                                  bool parent_locked)
2878 {
2879         struct ceph_client *cl = mdsc->fsc->client;
2880         int r = 0;
2881
2882         if (rinode) {
2883                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2884                 doutc(cl, " inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2885                       ceph_snap(rinode));
2886         } else if (rdentry) {
2887                 r = build_dentry_path(mdsc, rdentry, rdiri, ppath, pathlen, ino,
2888                                         freepath, parent_locked);
2889                 doutc(cl, " dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen, *ppath);
2890         } else if (rpath || rino) {
2891                 *ino = rino;
2892                 *ppath = rpath;
2893                 *pathlen = rpath ? strlen(rpath) : 0;
2894                 doutc(cl, " path %.*s\n", *pathlen, rpath);
2895         }
2896
2897         return r;
2898 }
2899
2900 static void encode_mclientrequest_tail(void **p,
2901                                        const struct ceph_mds_request *req)
2902 {
2903         struct ceph_timespec ts;
2904         int i;
2905
2906         ceph_encode_timespec64(&ts, &req->r_stamp);
2907         ceph_encode_copy(p, &ts, sizeof(ts));
2908
2909         /* v4: gid_list */
2910         ceph_encode_32(p, req->r_cred->group_info->ngroups);
2911         for (i = 0; i < req->r_cred->group_info->ngroups; i++)
2912                 ceph_encode_64(p, from_kgid(&init_user_ns,
2913                                             req->r_cred->group_info->gid[i]));
2914
2915         /* v5: altname */
2916         ceph_encode_32(p, req->r_altname_len);
2917         ceph_encode_copy(p, req->r_altname, req->r_altname_len);
2918
2919         /* v6: fscrypt_auth and fscrypt_file */
2920         if (req->r_fscrypt_auth) {
2921                 u32 authlen = ceph_fscrypt_auth_len(req->r_fscrypt_auth);
2922
2923                 ceph_encode_32(p, authlen);
2924                 ceph_encode_copy(p, req->r_fscrypt_auth, authlen);
2925         } else {
2926                 ceph_encode_32(p, 0);
2927         }
2928         if (test_bit(CEPH_MDS_R_FSCRYPT_FILE, &req->r_req_flags)) {
2929                 ceph_encode_32(p, sizeof(__le64));
2930                 ceph_encode_64(p, req->r_fscrypt_file);
2931         } else {
2932                 ceph_encode_32(p, 0);
2933         }
2934 }
2935
2936 static inline u16 mds_supported_head_version(struct ceph_mds_session *session)
2937 {
2938         if (!test_bit(CEPHFS_FEATURE_32BITS_RETRY_FWD, &session->s_features))
2939                 return 1;
2940
2941         if (!test_bit(CEPHFS_FEATURE_HAS_OWNER_UIDGID, &session->s_features))
2942                 return 2;
2943
2944         return CEPH_MDS_REQUEST_HEAD_VERSION;
2945 }
2946
2947 static struct ceph_mds_request_head_legacy *
2948 find_legacy_request_head(void *p, u64 features)
2949 {
2950         bool legacy = !(features & CEPH_FEATURE_FS_BTIME);
2951         struct ceph_mds_request_head *head;
2952
2953         if (legacy)
2954                 return (struct ceph_mds_request_head_legacy *)p;
2955         head = (struct ceph_mds_request_head *)p;
2956         return (struct ceph_mds_request_head_legacy *)&head->oldest_client_tid;
2957 }
2958
2959 /*
2960  * called under mdsc->mutex
2961  */
2962 static struct ceph_msg *create_request_message(struct ceph_mds_session *session,
2963                                                struct ceph_mds_request *req,
2964                                                bool drop_cap_releases)
2965 {
2966         int mds = session->s_mds;
2967         struct ceph_mds_client *mdsc = session->s_mdsc;
2968         struct ceph_client *cl = mdsc->fsc->client;
2969         struct ceph_msg *msg;
2970         struct ceph_mds_request_head_legacy *lhead;
2971         const char *path1 = NULL;
2972         const char *path2 = NULL;
2973         u64 ino1 = 0, ino2 = 0;
2974         int pathlen1 = 0, pathlen2 = 0;
2975         bool freepath1 = false, freepath2 = false;
2976         struct dentry *old_dentry = NULL;
2977         int len;
2978         u16 releases;
2979         void *p, *end;
2980         int ret;
2981         bool legacy = !(session->s_con.peer_features & CEPH_FEATURE_FS_BTIME);
2982         u16 request_head_version = mds_supported_head_version(session);
2983         kuid_t caller_fsuid = req->r_cred->fsuid;
2984         kgid_t caller_fsgid = req->r_cred->fsgid;
2985
2986         ret = set_request_path_attr(mdsc, req->r_inode, req->r_dentry,
2987                               req->r_parent, req->r_path1, req->r_ino1.ino,
2988                               &path1, &pathlen1, &ino1, &freepath1,
2989                               test_bit(CEPH_MDS_R_PARENT_LOCKED,
2990                                         &req->r_req_flags));
2991         if (ret < 0) {
2992                 msg = ERR_PTR(ret);
2993                 goto out;
2994         }
2995
2996         /* If r_old_dentry is set, then assume that its parent is locked */
2997         if (req->r_old_dentry &&
2998             !(req->r_old_dentry->d_flags & DCACHE_DISCONNECTED))
2999                 old_dentry = req->r_old_dentry;
3000         ret = set_request_path_attr(mdsc, NULL, old_dentry,
3001                               req->r_old_dentry_dir,
3002                               req->r_path2, req->r_ino2.ino,
3003                               &path2, &pathlen2, &ino2, &freepath2, true);
3004         if (ret < 0) {
3005                 msg = ERR_PTR(ret);
3006                 goto out_free1;
3007         }
3008
3009         req->r_altname = get_fscrypt_altname(req, &req->r_altname_len);
3010         if (IS_ERR(req->r_altname)) {
3011                 msg = ERR_CAST(req->r_altname);
3012                 req->r_altname = NULL;
3013                 goto out_free2;
3014         }
3015
3016         /*
3017          * For old cephs without supporting the 32bit retry/fwd feature
3018          * it will copy the raw memories directly when decoding the
3019          * requests. While new cephs will decode the head depending the
3020          * version member, so we need to make sure it will be compatible
3021          * with them both.
3022          */
3023         if (legacy)
3024                 len = sizeof(struct ceph_mds_request_head_legacy);
3025         else if (request_head_version == 1)
3026                 len = offsetofend(struct ceph_mds_request_head, args);
3027         else if (request_head_version == 2)
3028                 len = offsetofend(struct ceph_mds_request_head, ext_num_fwd);
3029         else
3030                 len = sizeof(struct ceph_mds_request_head);
3031
3032         /* filepaths */
3033         len += 2 * (1 + sizeof(u32) + sizeof(u64));
3034         len += pathlen1 + pathlen2;
3035
3036         /* cap releases */
3037         len += sizeof(struct ceph_mds_request_release) *
3038                 (!!req->r_inode_drop + !!req->r_dentry_drop +
3039                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
3040
3041         if (req->r_dentry_drop)
3042                 len += pathlen1;
3043         if (req->r_old_dentry_drop)
3044                 len += pathlen2;
3045
3046         /* MClientRequest tail */
3047
3048         /* req->r_stamp */
3049         len += sizeof(struct ceph_timespec);
3050
3051         /* gid list */
3052         len += sizeof(u32) + (sizeof(u64) * req->r_cred->group_info->ngroups);
3053
3054         /* alternate name */
3055         len += sizeof(u32) + req->r_altname_len;
3056
3057         /* fscrypt_auth */
3058         len += sizeof(u32); // fscrypt_auth
3059         if (req->r_fscrypt_auth)
3060                 len += ceph_fscrypt_auth_len(req->r_fscrypt_auth);
3061
3062         /* fscrypt_file */
3063         len += sizeof(u32);
3064         if (test_bit(CEPH_MDS_R_FSCRYPT_FILE, &req->r_req_flags))
3065                 len += sizeof(__le64);
3066
3067         msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
3068         if (!msg) {
3069                 msg = ERR_PTR(-ENOMEM);
3070                 goto out_free2;
3071         }
3072
3073         msg->hdr.tid = cpu_to_le64(req->r_tid);
3074
3075         lhead = find_legacy_request_head(msg->front.iov_base,
3076                                          session->s_con.peer_features);
3077
3078         if ((req->r_mnt_idmap != &nop_mnt_idmap) &&
3079             !test_bit(CEPHFS_FEATURE_HAS_OWNER_UIDGID, &session->s_features)) {
3080                 WARN_ON_ONCE(!IS_CEPH_MDS_OP_NEWINODE(req->r_op));
3081
3082                 if (enable_unsafe_idmap) {
3083                         pr_warn_once_client(cl,
3084                                 "idmapped mount is used and CEPHFS_FEATURE_HAS_OWNER_UIDGID"
3085                                 " is not supported by MDS. UID/GID-based restrictions may"
3086                                 " not work properly.\n");
3087
3088                         caller_fsuid = from_vfsuid(req->r_mnt_idmap, &init_user_ns,
3089                                                    VFSUIDT_INIT(req->r_cred->fsuid));
3090                         caller_fsgid = from_vfsgid(req->r_mnt_idmap, &init_user_ns,
3091                                                    VFSGIDT_INIT(req->r_cred->fsgid));
3092                 } else {
3093                         pr_err_ratelimited_client(cl,
3094                                 "idmapped mount is used and CEPHFS_FEATURE_HAS_OWNER_UIDGID"
3095                                 " is not supported by MDS. Fail request with -EIO.\n");
3096
3097                         ret = -EIO;
3098                         goto out_err;
3099                 }
3100         }
3101
3102         /*
3103          * The ceph_mds_request_head_legacy didn't contain a version field, and
3104          * one was added when we moved the message version from 3->4.
3105          */
3106         if (legacy) {
3107                 msg->hdr.version = cpu_to_le16(3);
3108                 p = msg->front.iov_base + sizeof(*lhead);
3109         } else if (request_head_version == 1) {
3110                 struct ceph_mds_request_head *nhead = msg->front.iov_base;
3111
3112                 msg->hdr.version = cpu_to_le16(4);
3113                 nhead->version = cpu_to_le16(1);
3114                 p = msg->front.iov_base + offsetofend(struct ceph_mds_request_head, args);
3115         } else if (request_head_version == 2) {
3116                 struct ceph_mds_request_head *nhead = msg->front.iov_base;
3117
3118                 msg->hdr.version = cpu_to_le16(6);
3119                 nhead->version = cpu_to_le16(2);
3120
3121                 p = msg->front.iov_base + offsetofend(struct ceph_mds_request_head, ext_num_fwd);
3122         } else {
3123                 struct ceph_mds_request_head *nhead = msg->front.iov_base;
3124                 kuid_t owner_fsuid;
3125                 kgid_t owner_fsgid;
3126
3127                 msg->hdr.version = cpu_to_le16(6);
3128                 nhead->version = cpu_to_le16(CEPH_MDS_REQUEST_HEAD_VERSION);
3129                 nhead->struct_len = cpu_to_le32(sizeof(struct ceph_mds_request_head));
3130
3131                 if (IS_CEPH_MDS_OP_NEWINODE(req->r_op)) {
3132                         owner_fsuid = from_vfsuid(req->r_mnt_idmap, &init_user_ns,
3133                                                 VFSUIDT_INIT(req->r_cred->fsuid));
3134                         owner_fsgid = from_vfsgid(req->r_mnt_idmap, &init_user_ns,
3135                                                 VFSGIDT_INIT(req->r_cred->fsgid));
3136                         nhead->owner_uid = cpu_to_le32(from_kuid(&init_user_ns, owner_fsuid));
3137                         nhead->owner_gid = cpu_to_le32(from_kgid(&init_user_ns, owner_fsgid));
3138                 } else {
3139                         nhead->owner_uid = cpu_to_le32(-1);
3140                         nhead->owner_gid = cpu_to_le32(-1);
3141                 }
3142
3143                 p = msg->front.iov_base + sizeof(*nhead);
3144         }
3145
3146         end = msg->front.iov_base + msg->front.iov_len;
3147
3148         lhead->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
3149         lhead->op = cpu_to_le32(req->r_op);
3150         lhead->caller_uid = cpu_to_le32(from_kuid(&init_user_ns,
3151                                                   caller_fsuid));
3152         lhead->caller_gid = cpu_to_le32(from_kgid(&init_user_ns,
3153                                                   caller_fsgid));
3154         lhead->ino = cpu_to_le64(req->r_deleg_ino);
3155         lhead->args = req->r_args;
3156
3157         ceph_encode_filepath(&p, end, ino1, path1);
3158         ceph_encode_filepath(&p, end, ino2, path2);
3159
3160         /* make note of release offset, in case we need to replay */
3161         req->r_request_release_offset = p - msg->front.iov_base;
3162
3163         /* cap releases */
3164         releases = 0;
3165         if (req->r_inode_drop)
3166                 releases += ceph_encode_inode_release(&p,
3167                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
3168                       mds, req->r_inode_drop, req->r_inode_unless,
3169                       req->r_op == CEPH_MDS_OP_READDIR);
3170         if (req->r_dentry_drop) {
3171                 ret = ceph_encode_dentry_release(&p, req->r_dentry,
3172                                 req->r_parent, mds, req->r_dentry_drop,
3173                                 req->r_dentry_unless);
3174                 if (ret < 0)
3175                         goto out_err;
3176                 releases += ret;
3177         }
3178         if (req->r_old_dentry_drop) {
3179                 ret = ceph_encode_dentry_release(&p, req->r_old_dentry,
3180                                 req->r_old_dentry_dir, mds,
3181                                 req->r_old_dentry_drop,
3182                                 req->r_old_dentry_unless);
3183                 if (ret < 0)
3184                         goto out_err;
3185                 releases += ret;
3186         }
3187         if (req->r_old_inode_drop)
3188                 releases += ceph_encode_inode_release(&p,
3189                       d_inode(req->r_old_dentry),
3190                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
3191
3192         if (drop_cap_releases) {
3193                 releases = 0;
3194                 p = msg->front.iov_base + req->r_request_release_offset;
3195         }
3196
3197         lhead->num_releases = cpu_to_le16(releases);
3198
3199         encode_mclientrequest_tail(&p, req);
3200
3201         if (WARN_ON_ONCE(p > end)) {
3202                 ceph_msg_put(msg);
3203                 msg = ERR_PTR(-ERANGE);
3204                 goto out_free2;
3205         }
3206
3207         msg->front.iov_len = p - msg->front.iov_base;
3208         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
3209
3210         if (req->r_pagelist) {
3211                 struct ceph_pagelist *pagelist = req->r_pagelist;
3212                 ceph_msg_data_add_pagelist(msg, pagelist);
3213                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
3214         } else {
3215                 msg->hdr.data_len = 0;
3216         }
3217
3218         msg->hdr.data_off = cpu_to_le16(0);
3219
3220 out_free2:
3221         if (freepath2)
3222                 ceph_mdsc_free_path((char *)path2, pathlen2);
3223 out_free1:
3224         if (freepath1)
3225                 ceph_mdsc_free_path((char *)path1, pathlen1);
3226 out:
3227         return msg;
3228 out_err:
3229         ceph_msg_put(msg);
3230         msg = ERR_PTR(ret);
3231         goto out_free2;
3232 }
3233
3234 /*
3235  * called under mdsc->mutex if error, under no mutex if
3236  * success.
3237  */
3238 static void complete_request(struct ceph_mds_client *mdsc,
3239                              struct ceph_mds_request *req)
3240 {
3241         req->r_end_latency = ktime_get();
3242
3243         if (req->r_callback)
3244                 req->r_callback(mdsc, req);
3245         complete_all(&req->r_completion);
3246 }
3247
3248 /*
3249  * called under mdsc->mutex
3250  */
3251 static int __prepare_send_request(struct ceph_mds_session *session,
3252                                   struct ceph_mds_request *req,
3253                                   bool drop_cap_releases)
3254 {
3255         int mds = session->s_mds;
3256         struct ceph_mds_client *mdsc = session->s_mdsc;
3257         struct ceph_client *cl = mdsc->fsc->client;
3258         struct ceph_mds_request_head_legacy *lhead;
3259         struct ceph_mds_request_head *nhead;
3260         struct ceph_msg *msg;
3261         int flags = 0, old_max_retry;
3262         bool old_version = !test_bit(CEPHFS_FEATURE_32BITS_RETRY_FWD,
3263                                      &session->s_features);
3264
3265         /*
3266          * Avoid infinite retrying after overflow. The client will
3267          * increase the retry count and if the MDS is old version,
3268          * so we limit to retry at most 256 times.
3269          */
3270         if (req->r_attempts) {
3271                old_max_retry = sizeof_field(struct ceph_mds_request_head,
3272                                             num_retry);
3273                old_max_retry = 1 << (old_max_retry * BITS_PER_BYTE);
3274                if ((old_version && req->r_attempts >= old_max_retry) ||
3275                    ((uint32_t)req->r_attempts >= U32_MAX)) {
3276                         pr_warn_ratelimited_client(cl, "request tid %llu seq overflow\n",
3277                                                    req->r_tid);
3278                         return -EMULTIHOP;
3279                }
3280         }
3281
3282         req->r_attempts++;
3283         if (req->r_inode) {
3284                 struct ceph_cap *cap =
3285                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
3286
3287                 if (cap)
3288                         req->r_sent_on_mseq = cap->mseq;
3289                 else
3290                         req->r_sent_on_mseq = -1;
3291         }
3292         doutc(cl, "%p tid %lld %s (attempt %d)\n", req, req->r_tid,
3293               ceph_mds_op_name(req->r_op), req->r_attempts);
3294
3295         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3296                 void *p;
3297
3298                 /*
3299                  * Replay.  Do not regenerate message (and rebuild
3300                  * paths, etc.); just use the original message.
3301                  * Rebuilding paths will break for renames because
3302                  * d_move mangles the src name.
3303                  */
3304                 msg = req->r_request;
3305                 lhead = find_legacy_request_head(msg->front.iov_base,
3306                                                  session->s_con.peer_features);
3307
3308                 flags = le32_to_cpu(lhead->flags);
3309                 flags |= CEPH_MDS_FLAG_REPLAY;
3310                 lhead->flags = cpu_to_le32(flags);
3311
3312                 if (req->r_target_inode)
3313                         lhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
3314
3315                 lhead->num_retry = req->r_attempts - 1;
3316                 if (!old_version) {
3317                         nhead = (struct ceph_mds_request_head*)msg->front.iov_base;
3318                         nhead->ext_num_retry = cpu_to_le32(req->r_attempts - 1);
3319                 }
3320
3321                 /* remove cap/dentry releases from message */
3322                 lhead->num_releases = 0;
3323
3324                 p = msg->front.iov_base + req->r_request_release_offset;
3325                 encode_mclientrequest_tail(&p, req);
3326
3327                 msg->front.iov_len = p - msg->front.iov_base;
3328                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
3329                 return 0;
3330         }
3331
3332         if (req->r_request) {
3333                 ceph_msg_put(req->r_request);
3334                 req->r_request = NULL;
3335         }
3336         msg = create_request_message(session, req, drop_cap_releases);
3337         if (IS_ERR(msg)) {
3338                 req->r_err = PTR_ERR(msg);
3339                 return PTR_ERR(msg);
3340         }
3341         req->r_request = msg;
3342
3343         lhead = find_legacy_request_head(msg->front.iov_base,
3344                                          session->s_con.peer_features);
3345         lhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
3346         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3347                 flags |= CEPH_MDS_FLAG_REPLAY;
3348         if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags))
3349                 flags |= CEPH_MDS_FLAG_ASYNC;
3350         if (req->r_parent)
3351                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
3352         lhead->flags = cpu_to_le32(flags);
3353         lhead->num_fwd = req->r_num_fwd;
3354         lhead->num_retry = req->r_attempts - 1;
3355         if (!old_version) {
3356                 nhead = (struct ceph_mds_request_head*)msg->front.iov_base;
3357                 nhead->ext_num_fwd = cpu_to_le32(req->r_num_fwd);
3358                 nhead->ext_num_retry = cpu_to_le32(req->r_attempts - 1);
3359         }
3360
3361         doutc(cl, " r_parent = %p\n", req->r_parent);
3362         return 0;
3363 }
3364
3365 /*
3366  * called under mdsc->mutex
3367  */
3368 static int __send_request(struct ceph_mds_session *session,
3369                           struct ceph_mds_request *req,
3370                           bool drop_cap_releases)
3371 {
3372         int err;
3373
3374         err = __prepare_send_request(session, req, drop_cap_releases);
3375         if (!err) {
3376                 ceph_msg_get(req->r_request);
3377                 ceph_con_send(&session->s_con, req->r_request);
3378         }
3379
3380         return err;
3381 }
3382
3383 /*
3384  * send request, or put it on the appropriate wait list.
3385  */
3386 static void __do_request(struct ceph_mds_client *mdsc,
3387                         struct ceph_mds_request *req)
3388 {
3389         struct ceph_client *cl = mdsc->fsc->client;
3390         struct ceph_mds_session *session = NULL;
3391         int mds = -1;
3392         int err = 0;
3393         bool random;
3394
3395         if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
3396                 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
3397                         __unregister_request(mdsc, req);
3398                 return;
3399         }
3400
3401         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_FENCE_IO) {
3402                 doutc(cl, "metadata corrupted\n");
3403                 err = -EIO;
3404                 goto finish;
3405         }
3406         if (req->r_timeout &&
3407             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
3408                 doutc(cl, "timed out\n");
3409                 err = -ETIMEDOUT;
3410                 goto finish;
3411         }
3412         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
3413                 doutc(cl, "forced umount\n");
3414                 err = -EIO;
3415                 goto finish;
3416         }
3417         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
3418                 if (mdsc->mdsmap_err) {
3419                         err = mdsc->mdsmap_err;
3420                         doutc(cl, "mdsmap err %d\n", err);
3421                         goto finish;
3422                 }
3423                 if (mdsc->mdsmap->m_epoch == 0) {
3424                         doutc(cl, "no mdsmap, waiting for map\n");
3425                         list_add(&req->r_wait, &mdsc->waiting_for_map);
3426                         return;
3427                 }
3428                 if (!(mdsc->fsc->mount_options->flags &
3429                       CEPH_MOUNT_OPT_MOUNTWAIT) &&
3430                     !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
3431                         err = -EHOSTUNREACH;
3432                         goto finish;
3433                 }
3434         }
3435
3436         put_request_session(req);
3437
3438         mds = __choose_mds(mdsc, req, &random);
3439         if (mds < 0 ||
3440             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
3441                 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
3442                         err = -EJUKEBOX;
3443                         goto finish;
3444                 }
3445                 doutc(cl, "no mds or not active, waiting for map\n");
3446                 list_add(&req->r_wait, &mdsc->waiting_for_map);
3447                 return;
3448         }
3449
3450         /* get, open session */
3451         session = __ceph_lookup_mds_session(mdsc, mds);
3452         if (!session) {
3453                 session = register_session(mdsc, mds);
3454                 if (IS_ERR(session)) {
3455                         err = PTR_ERR(session);
3456                         goto finish;
3457                 }
3458         }
3459         req->r_session = ceph_get_mds_session(session);
3460
3461         doutc(cl, "mds%d session %p state %s\n", mds, session,
3462               ceph_session_state_name(session->s_state));
3463
3464         /*
3465          * The old ceph will crash the MDSs when see unknown OPs
3466          */
3467         if (req->r_feature_needed > 0 &&
3468             !test_bit(req->r_feature_needed, &session->s_features)) {
3469                 err = -EOPNOTSUPP;
3470                 goto out_session;
3471         }
3472
3473         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
3474             session->s_state != CEPH_MDS_SESSION_HUNG) {
3475                 /*
3476                  * We cannot queue async requests since the caps and delegated
3477                  * inodes are bound to the session. Just return -EJUKEBOX and
3478                  * let the caller retry a sync request in that case.
3479                  */
3480                 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
3481                         err = -EJUKEBOX;
3482                         goto out_session;
3483                 }
3484
3485                 /*
3486                  * If the session has been REJECTED, then return a hard error,
3487                  * unless it's a CLEANRECOVER mount, in which case we'll queue
3488                  * it to the mdsc queue.
3489                  */
3490                 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
3491                         if (ceph_test_mount_opt(mdsc->fsc, CLEANRECOVER))
3492                                 list_add(&req->r_wait, &mdsc->waiting_for_map);
3493                         else
3494                                 err = -EACCES;
3495                         goto out_session;
3496                 }
3497
3498                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
3499                     session->s_state == CEPH_MDS_SESSION_CLOSING) {
3500                         err = __open_session(mdsc, session);
3501                         if (err)
3502                                 goto out_session;
3503                         /* retry the same mds later */
3504                         if (random)
3505                                 req->r_resend_mds = mds;
3506                 }
3507                 list_add(&req->r_wait, &session->s_waiting);
3508                 goto out_session;
3509         }
3510
3511         /* send request */
3512         req->r_resend_mds = -1;   /* forget any previous mds hint */
3513
3514         if (req->r_request_started == 0)   /* note request start time */
3515                 req->r_request_started = jiffies;
3516
3517         /*
3518          * For async create we will choose the auth MDS of frag in parent
3519          * directory to send the request and usually this works fine, but
3520          * if the migrated the dirtory to another MDS before it could handle
3521          * it the request will be forwarded.
3522          *
3523          * And then the auth cap will be changed.
3524          */
3525         if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags) && req->r_num_fwd) {
3526                 struct ceph_dentry_info *di = ceph_dentry(req->r_dentry);
3527                 struct ceph_inode_info *ci;
3528                 struct ceph_cap *cap;
3529
3530                 /*
3531                  * The request maybe handled very fast and the new inode
3532                  * hasn't been linked to the dentry yet. We need to wait
3533                  * for the ceph_finish_async_create(), which shouldn't be
3534                  * stuck too long or fail in thoery, to finish when forwarding
3535                  * the request.
3536                  */
3537                 if (!d_inode(req->r_dentry)) {
3538                         err = wait_on_bit(&di->flags, CEPH_DENTRY_ASYNC_CREATE_BIT,
3539                                           TASK_KILLABLE);
3540                         if (err) {
3541                                 mutex_lock(&req->r_fill_mutex);
3542                                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
3543                                 mutex_unlock(&req->r_fill_mutex);
3544                                 goto out_session;
3545                         }
3546                 }
3547
3548                 ci = ceph_inode(d_inode(req->r_dentry));
3549
3550                 spin_lock(&ci->i_ceph_lock);
3551                 cap = ci->i_auth_cap;
3552                 if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE && mds != cap->mds) {
3553                         doutc(cl, "session changed for auth cap %d -> %d\n",
3554                               cap->session->s_mds, session->s_mds);
3555
3556                         /* Remove the auth cap from old session */
3557                         spin_lock(&cap->session->s_cap_lock);
3558                         cap->session->s_nr_caps--;
3559                         list_del_init(&cap->session_caps);
3560                         spin_unlock(&cap->session->s_cap_lock);
3561
3562                         /* Add the auth cap to the new session */
3563                         cap->mds = mds;
3564                         cap->session = session;
3565                         spin_lock(&session->s_cap_lock);
3566                         session->s_nr_caps++;
3567                         list_add_tail(&cap->session_caps, &session->s_caps);
3568                         spin_unlock(&session->s_cap_lock);
3569
3570                         change_auth_cap_ses(ci, session);
3571                 }
3572                 spin_unlock(&ci->i_ceph_lock);
3573         }
3574
3575         err = __send_request(session, req, false);
3576
3577 out_session:
3578         ceph_put_mds_session(session);
3579 finish:
3580         if (err) {
3581                 doutc(cl, "early error %d\n", err);
3582                 req->r_err = err;
3583                 complete_request(mdsc, req);
3584                 __unregister_request(mdsc, req);
3585         }
3586         return;
3587 }
3588
3589 /*
3590  * called under mdsc->mutex
3591  */
3592 static void __wake_requests(struct ceph_mds_client *mdsc,
3593                             struct list_head *head)
3594 {
3595         struct ceph_client *cl = mdsc->fsc->client;
3596         struct ceph_mds_request *req;
3597         LIST_HEAD(tmp_list);
3598
3599         list_splice_init(head, &tmp_list);
3600
3601         while (!list_empty(&tmp_list)) {
3602                 req = list_entry(tmp_list.next,
3603                                  struct ceph_mds_request, r_wait);
3604                 list_del_init(&req->r_wait);
3605                 doutc(cl, " wake request %p tid %llu\n", req,
3606                       req->r_tid);
3607                 __do_request(mdsc, req);
3608         }
3609 }
3610
3611 /*
3612  * Wake up threads with requests pending for @mds, so that they can
3613  * resubmit their requests to a possibly different mds.
3614  */
3615 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
3616 {
3617         struct ceph_client *cl = mdsc->fsc->client;
3618         struct ceph_mds_request *req;
3619         struct rb_node *p = rb_first(&mdsc->request_tree);
3620
3621         doutc(cl, "kick_requests mds%d\n", mds);
3622         while (p) {
3623                 req = rb_entry(p, struct ceph_mds_request, r_node);
3624                 p = rb_next(p);
3625                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3626                         continue;
3627                 if (req->r_attempts > 0)
3628                         continue; /* only new requests */
3629                 if (req->r_session &&
3630                     req->r_session->s_mds == mds) {
3631                         doutc(cl, " kicking tid %llu\n", req->r_tid);
3632                         list_del_init(&req->r_wait);
3633                         __do_request(mdsc, req);
3634                 }
3635         }
3636 }
3637
3638 int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
3639                               struct ceph_mds_request *req)
3640 {
3641         struct ceph_client *cl = mdsc->fsc->client;
3642         int err = 0;
3643
3644         /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
3645         if (req->r_inode)
3646                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
3647         if (req->r_parent) {
3648                 struct ceph_inode_info *ci = ceph_inode(req->r_parent);
3649                 int fmode = (req->r_op & CEPH_MDS_OP_WRITE) ?
3650                             CEPH_FILE_MODE_WR : CEPH_FILE_MODE_RD;
3651                 spin_lock(&ci->i_ceph_lock);
3652                 ceph_take_cap_refs(ci, CEPH_CAP_PIN, false);
3653                 __ceph_touch_fmode(ci, mdsc, fmode);
3654                 spin_unlock(&ci->i_ceph_lock);
3655         }
3656         if (req->r_old_dentry_dir)
3657                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
3658                                   CEPH_CAP_PIN);
3659
3660         if (req->r_inode) {
3661                 err = ceph_wait_on_async_create(req->r_inode);
3662                 if (err) {
3663                         doutc(cl, "wait for async create returned: %d\n", err);
3664                         return err;
3665                 }
3666         }
3667
3668         if (!err && req->r_old_inode) {
3669                 err = ceph_wait_on_async_create(req->r_old_inode);
3670                 if (err) {
3671                         doutc(cl, "wait for async create returned: %d\n", err);
3672                         return err;
3673                 }
3674         }
3675
3676         doutc(cl, "submit_request on %p for inode %p\n", req, dir);
3677         mutex_lock(&mdsc->mutex);
3678         __register_request(mdsc, req, dir);
3679         __do_request(mdsc, req);
3680         err = req->r_err;
3681         mutex_unlock(&mdsc->mutex);
3682         return err;
3683 }
3684
3685 int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
3686                            struct ceph_mds_request *req,
3687                            ceph_mds_request_wait_callback_t wait_func)
3688 {
3689         struct ceph_client *cl = mdsc->fsc->client;
3690         int err;
3691
3692         /* wait */
3693         doutc(cl, "do_request waiting\n");
3694         if (wait_func) {
3695                 err = wait_func(mdsc, req);
3696         } else {
3697                 long timeleft = wait_for_completion_killable_timeout(
3698                                         &req->r_completion,
3699                                         ceph_timeout_jiffies(req->r_timeout));
3700                 if (timeleft > 0)
3701                         err = 0;
3702                 else if (!timeleft)
3703                         err = -ETIMEDOUT;  /* timed out */
3704                 else
3705                         err = timeleft;  /* killed */
3706         }
3707         doutc(cl, "do_request waited, got %d\n", err);
3708         mutex_lock(&mdsc->mutex);
3709
3710         /* only abort if we didn't race with a real reply */
3711         if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
3712                 err = le32_to_cpu(req->r_reply_info.head->result);
3713         } else if (err < 0) {
3714                 doutc(cl, "aborted request %lld with %d\n", req->r_tid, err);
3715
3716                 /*
3717                  * ensure we aren't running concurrently with
3718                  * ceph_fill_trace or ceph_readdir_prepopulate, which
3719                  * rely on locks (dir mutex) held by our caller.
3720                  */
3721                 mutex_lock(&req->r_fill_mutex);
3722                 req->r_err = err;
3723                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
3724                 mutex_unlock(&req->r_fill_mutex);
3725
3726                 if (req->r_parent &&
3727                     (req->r_op & CEPH_MDS_OP_WRITE))
3728                         ceph_invalidate_dir_request(req);
3729         } else {
3730                 err = req->r_err;
3731         }
3732
3733         mutex_unlock(&mdsc->mutex);
3734         return err;
3735 }
3736
3737 /*
3738  * Synchrously perform an mds request.  Take care of all of the
3739  * session setup, forwarding, retry details.
3740  */
3741 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
3742                          struct inode *dir,
3743                          struct ceph_mds_request *req)
3744 {
3745         struct ceph_client *cl = mdsc->fsc->client;
3746         int err;
3747
3748         doutc(cl, "do_request on %p\n", req);
3749
3750         /* issue */
3751         err = ceph_mdsc_submit_request(mdsc, dir, req);
3752         if (!err)
3753                 err = ceph_mdsc_wait_request(mdsc, req, NULL);
3754         doutc(cl, "do_request %p done, result %d\n", req, err);
3755         return err;
3756 }
3757
3758 /*
3759  * Invalidate dir's completeness, dentry lease state on an aborted MDS
3760  * namespace request.
3761  */
3762 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
3763 {
3764         struct inode *dir = req->r_parent;
3765         struct inode *old_dir = req->r_old_dentry_dir;
3766         struct ceph_client *cl = req->r_mdsc->fsc->client;
3767
3768         doutc(cl, "invalidate_dir_request %p %p (complete, lease(s))\n",
3769               dir, old_dir);
3770
3771         ceph_dir_clear_complete(dir);
3772         if (old_dir)
3773                 ceph_dir_clear_complete(old_dir);
3774         if (req->r_dentry)
3775                 ceph_invalidate_dentry_lease(req->r_dentry);
3776         if (req->r_old_dentry)
3777                 ceph_invalidate_dentry_lease(req->r_old_dentry);
3778 }
3779
3780 /*
3781  * Handle mds reply.
3782  *
3783  * We take the session mutex and parse and process the reply immediately.
3784  * This preserves the logical ordering of replies, capabilities, etc., sent
3785  * by the MDS as they are applied to our local cache.
3786  */
3787 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
3788 {
3789         struct ceph_mds_client *mdsc = session->s_mdsc;
3790         struct ceph_client *cl = mdsc->fsc->client;
3791         struct ceph_mds_request *req;
3792         struct ceph_mds_reply_head *head = msg->front.iov_base;
3793         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
3794         struct ceph_snap_realm *realm;
3795         u64 tid;
3796         int err, result;
3797         int mds = session->s_mds;
3798         bool close_sessions = false;
3799
3800         if (msg->front.iov_len < sizeof(*head)) {
3801                 pr_err_client(cl, "got corrupt (short) reply\n");
3802                 ceph_msg_dump(msg);
3803                 return;
3804         }
3805
3806         /* get request, session */
3807         tid = le64_to_cpu(msg->hdr.tid);
3808         mutex_lock(&mdsc->mutex);
3809         req = lookup_get_request(mdsc, tid);
3810         if (!req) {
3811                 doutc(cl, "on unknown tid %llu\n", tid);
3812                 mutex_unlock(&mdsc->mutex);
3813                 return;
3814         }
3815         doutc(cl, "handle_reply %p\n", req);
3816
3817         /* correct session? */
3818         if (req->r_session != session) {
3819                 pr_err_client(cl, "got %llu on session mds%d not mds%d\n",
3820                               tid, session->s_mds,
3821                               req->r_session ? req->r_session->s_mds : -1);
3822                 mutex_unlock(&mdsc->mutex);
3823                 goto out;
3824         }
3825
3826         /* dup? */
3827         if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
3828             (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
3829                 pr_warn_client(cl, "got a dup %s reply on %llu from mds%d\n",
3830                                head->safe ? "safe" : "unsafe", tid, mds);
3831                 mutex_unlock(&mdsc->mutex);
3832                 goto out;
3833         }
3834         if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
3835                 pr_warn_client(cl, "got unsafe after safe on %llu from mds%d\n",
3836                                tid, mds);
3837                 mutex_unlock(&mdsc->mutex);
3838                 goto out;
3839         }
3840
3841         result = le32_to_cpu(head->result);
3842
3843         if (head->safe) {
3844                 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
3845                 __unregister_request(mdsc, req);
3846
3847                 /* last request during umount? */
3848                 if (mdsc->stopping && !__get_oldest_req(mdsc))
3849                         complete_all(&mdsc->safe_umount_waiters);
3850
3851                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3852                         /*
3853                          * We already handled the unsafe response, now do the
3854                          * cleanup.  No need to examine the response; the MDS
3855                          * doesn't include any result info in the safe
3856                          * response.  And even if it did, there is nothing
3857                          * useful we could do with a revised return value.
3858                          */
3859                         doutc(cl, "got safe reply %llu, mds%d\n", tid, mds);
3860
3861                         mutex_unlock(&mdsc->mutex);
3862                         goto out;
3863                 }
3864         } else {
3865                 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
3866                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
3867         }
3868
3869         doutc(cl, "tid %lld result %d\n", tid, result);
3870         if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
3871                 err = parse_reply_info(session, msg, req, (u64)-1);
3872         else
3873                 err = parse_reply_info(session, msg, req,
3874                                        session->s_con.peer_features);
3875         mutex_unlock(&mdsc->mutex);
3876
3877         /* Must find target inode outside of mutexes to avoid deadlocks */
3878         rinfo = &req->r_reply_info;
3879         if ((err >= 0) && rinfo->head->is_target) {
3880                 struct inode *in = xchg(&req->r_new_inode, NULL);
3881                 struct ceph_vino tvino = {
3882                         .ino  = le64_to_cpu(rinfo->targeti.in->ino),
3883                         .snap = le64_to_cpu(rinfo->targeti.in->snapid)
3884                 };
3885
3886                 /*
3887                  * If we ended up opening an existing inode, discard
3888                  * r_new_inode
3889                  */
3890                 if (req->r_op == CEPH_MDS_OP_CREATE &&
3891                     !req->r_reply_info.has_create_ino) {
3892                         /* This should never happen on an async create */
3893                         WARN_ON_ONCE(req->r_deleg_ino);
3894                         iput(in);
3895                         in = NULL;
3896                 }
3897
3898                 in = ceph_get_inode(mdsc->fsc->sb, tvino, in);
3899                 if (IS_ERR(in)) {
3900                         err = PTR_ERR(in);
3901                         mutex_lock(&session->s_mutex);
3902                         goto out_err;
3903                 }
3904                 req->r_target_inode = in;
3905         }
3906
3907         mutex_lock(&session->s_mutex);
3908         if (err < 0) {
3909                 pr_err_client(cl, "got corrupt reply mds%d(tid:%lld)\n",
3910                               mds, tid);
3911                 ceph_msg_dump(msg);
3912                 goto out_err;
3913         }
3914
3915         /* snap trace */
3916         realm = NULL;
3917         if (rinfo->snapblob_len) {
3918                 down_write(&mdsc->snap_rwsem);
3919                 err = ceph_update_snap_trace(mdsc, rinfo->snapblob,
3920                                 rinfo->snapblob + rinfo->snapblob_len,
3921                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
3922                                 &realm);
3923                 if (err) {
3924                         up_write(&mdsc->snap_rwsem);
3925                         close_sessions = true;
3926                         if (err == -EIO)
3927                                 ceph_msg_dump(msg);
3928                         goto out_err;
3929                 }
3930                 downgrade_write(&mdsc->snap_rwsem);
3931         } else {
3932                 down_read(&mdsc->snap_rwsem);
3933         }
3934
3935         /* insert trace into our cache */
3936         mutex_lock(&req->r_fill_mutex);
3937         current->journal_info = req;
3938         err = ceph_fill_trace(mdsc->fsc->sb, req);
3939         if (err == 0) {
3940                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
3941                                     req->r_op == CEPH_MDS_OP_LSSNAP))
3942                         err = ceph_readdir_prepopulate(req, req->r_session);
3943         }
3944         current->journal_info = NULL;
3945         mutex_unlock(&req->r_fill_mutex);
3946
3947         up_read(&mdsc->snap_rwsem);
3948         if (realm)
3949                 ceph_put_snap_realm(mdsc, realm);
3950
3951         if (err == 0) {
3952                 if (req->r_target_inode &&
3953                     test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3954                         struct ceph_inode_info *ci =
3955                                 ceph_inode(req->r_target_inode);
3956                         spin_lock(&ci->i_unsafe_lock);
3957                         list_add_tail(&req->r_unsafe_target_item,
3958                                       &ci->i_unsafe_iops);
3959                         spin_unlock(&ci->i_unsafe_lock);
3960                 }
3961
3962                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
3963         }
3964 out_err:
3965         mutex_lock(&mdsc->mutex);
3966         if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3967                 if (err) {
3968                         req->r_err = err;
3969                 } else {
3970                         req->r_reply =  ceph_msg_get(msg);
3971                         set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
3972                 }
3973         } else {
3974                 doutc(cl, "reply arrived after request %lld was aborted\n", tid);
3975         }
3976         mutex_unlock(&mdsc->mutex);
3977
3978         mutex_unlock(&session->s_mutex);
3979
3980         /* kick calling process */
3981         complete_request(mdsc, req);
3982
3983         ceph_update_metadata_metrics(&mdsc->metric, req->r_start_latency,
3984                                      req->r_end_latency, err);
3985 out:
3986         ceph_mdsc_put_request(req);
3987
3988         /* Defer closing the sessions after s_mutex lock being released */
3989         if (close_sessions)
3990                 ceph_mdsc_close_sessions(mdsc);
3991         return;
3992 }
3993
3994
3995
3996 /*
3997  * handle mds notification that our request has been forwarded.
3998  */
3999 static void handle_forward(struct ceph_mds_client *mdsc,
4000                            struct ceph_mds_session *session,
4001                            struct ceph_msg *msg)
4002 {
4003         struct ceph_client *cl = mdsc->fsc->client;
4004         struct ceph_mds_request *req;
4005         u64 tid = le64_to_cpu(msg->hdr.tid);
4006         u32 next_mds;
4007         u32 fwd_seq;
4008         int err = -EINVAL;
4009         void *p = msg->front.iov_base;
4010         void *end = p + msg->front.iov_len;
4011         bool aborted = false;
4012
4013         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
4014         next_mds = ceph_decode_32(&p);
4015         fwd_seq = ceph_decode_32(&p);
4016
4017         mutex_lock(&mdsc->mutex);
4018         req = lookup_get_request(mdsc, tid);
4019         if (!req) {
4020                 mutex_unlock(&mdsc->mutex);
4021                 doutc(cl, "forward tid %llu to mds%d - req dne\n", tid, next_mds);
4022                 return;  /* dup reply? */
4023         }
4024
4025         if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
4026                 doutc(cl, "forward tid %llu aborted, unregistering\n", tid);
4027                 __unregister_request(mdsc, req);
4028         } else if (fwd_seq <= req->r_num_fwd || (uint32_t)fwd_seq >= U32_MAX) {
4029                 /*
4030                  * Avoid infinite retrying after overflow.
4031                  *
4032                  * The MDS will increase the fwd count and in client side
4033                  * if the num_fwd is less than the one saved in request
4034                  * that means the MDS is an old version and overflowed of
4035                  * 8 bits.
4036                  */
4037                 mutex_lock(&req->r_fill_mutex);
4038                 req->r_err = -EMULTIHOP;
4039                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
4040                 mutex_unlock(&req->r_fill_mutex);
4041                 aborted = true;
4042                 pr_warn_ratelimited_client(cl, "forward tid %llu seq overflow\n",
4043                                            tid);
4044         } else {
4045                 /* resend. forward race not possible; mds would drop */
4046                 doutc(cl, "forward tid %llu to mds%d (we resend)\n", tid, next_mds);
4047                 BUG_ON(req->r_err);
4048                 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
4049                 req->r_attempts = 0;
4050                 req->r_num_fwd = fwd_seq;
4051                 req->r_resend_mds = next_mds;
4052                 put_request_session(req);
4053                 __do_request(mdsc, req);
4054         }
4055         mutex_unlock(&mdsc->mutex);
4056
4057         /* kick calling process */
4058         if (aborted)
4059                 complete_request(mdsc, req);
4060         ceph_mdsc_put_request(req);
4061         return;
4062
4063 bad:
4064         pr_err_client(cl, "decode error err=%d\n", err);
4065         ceph_msg_dump(msg);
4066 }
4067
4068 static int __decode_session_metadata(void **p, void *end,
4069                                      bool *blocklisted)
4070 {
4071         /* map<string,string> */
4072         u32 n;
4073         bool err_str;
4074         ceph_decode_32_safe(p, end, n, bad);
4075         while (n-- > 0) {
4076                 u32 len;
4077                 ceph_decode_32_safe(p, end, len, bad);
4078                 ceph_decode_need(p, end, len, bad);
4079                 err_str = !strncmp(*p, "error_string", len);
4080                 *p += len;
4081                 ceph_decode_32_safe(p, end, len, bad);
4082                 ceph_decode_need(p, end, len, bad);
4083                 /*
4084                  * Match "blocklisted (blacklisted)" from newer MDSes,
4085                  * or "blacklisted" from older MDSes.
4086                  */
4087                 if (err_str && strnstr(*p, "blacklisted", len))
4088                         *blocklisted = true;
4089                 *p += len;
4090         }
4091         return 0;
4092 bad:
4093         return -1;
4094 }
4095
4096 /*
4097  * handle a mds session control message
4098  */
4099 static void handle_session(struct ceph_mds_session *session,
4100                            struct ceph_msg *msg)
4101 {
4102         struct ceph_mds_client *mdsc = session->s_mdsc;
4103         struct ceph_client *cl = mdsc->fsc->client;
4104         int mds = session->s_mds;
4105         int msg_version = le16_to_cpu(msg->hdr.version);
4106         void *p = msg->front.iov_base;
4107         void *end = p + msg->front.iov_len;
4108         struct ceph_mds_session_head *h;
4109         struct ceph_mds_cap_auth *cap_auths = NULL;
4110         u32 op, cap_auths_num = 0;
4111         u64 seq, features = 0;
4112         int wake = 0;
4113         bool blocklisted = false;
4114         u32 i;
4115
4116
4117         /* decode */
4118         ceph_decode_need(&p, end, sizeof(*h), bad);
4119         h = p;
4120         p += sizeof(*h);
4121
4122         op = le32_to_cpu(h->op);
4123         seq = le64_to_cpu(h->seq);
4124
4125         if (msg_version >= 3) {
4126                 u32 len;
4127                 /* version >= 2 and < 5, decode metadata, skip otherwise
4128                  * as it's handled via flags.
4129                  */
4130                 if (msg_version >= 5)
4131                         ceph_decode_skip_map(&p, end, string, string, bad);
4132                 else if (__decode_session_metadata(&p, end, &blocklisted) < 0)
4133                         goto bad;
4134
4135                 /* version >= 3, feature bits */
4136                 ceph_decode_32_safe(&p, end, len, bad);
4137                 if (len) {
4138                         ceph_decode_64_safe(&p, end, features, bad);
4139                         p += len - sizeof(features);
4140                 }
4141         }
4142
4143         if (msg_version >= 5) {
4144                 u32 flags, len;
4145
4146                 /* version >= 4 */
4147                 ceph_decode_skip_16(&p, end, bad); /* struct_v, struct_cv */
4148                 ceph_decode_32_safe(&p, end, len, bad); /* len */
4149                 ceph_decode_skip_n(&p, end, len, bad); /* metric_spec */
4150
4151                 /* version >= 5, flags   */
4152                 ceph_decode_32_safe(&p, end, flags, bad);
4153                 if (flags & CEPH_SESSION_BLOCKLISTED) {
4154                         pr_warn_client(cl, "mds%d session blocklisted\n",
4155                                        session->s_mds);
4156                         blocklisted = true;
4157                 }
4158         }
4159
4160         if (msg_version >= 6) {
4161                 ceph_decode_32_safe(&p, end, cap_auths_num, bad);
4162                 doutc(cl, "cap_auths_num %d\n", cap_auths_num);
4163
4164                 if (cap_auths_num && op != CEPH_SESSION_OPEN) {
4165                         WARN_ON_ONCE(op != CEPH_SESSION_OPEN);
4166                         goto skip_cap_auths;
4167                 }
4168
4169                 cap_auths = kcalloc(cap_auths_num,
4170                                     sizeof(struct ceph_mds_cap_auth),
4171                                     GFP_KERNEL);
4172                 if (!cap_auths) {
4173                         pr_err_client(cl, "No memory for cap_auths\n");
4174                         return;
4175                 }
4176
4177                 for (i = 0; i < cap_auths_num; i++) {
4178                         u32 _len, j;
4179
4180                         /* struct_v, struct_compat, and struct_len in MDSCapAuth */
4181                         ceph_decode_skip_n(&p, end, 2 + sizeof(u32), bad);
4182
4183                         /* struct_v, struct_compat, and struct_len in MDSCapMatch */
4184                         ceph_decode_skip_n(&p, end, 2 + sizeof(u32), bad);
4185                         ceph_decode_64_safe(&p, end, cap_auths[i].match.uid, bad);
4186                         ceph_decode_32_safe(&p, end, _len, bad);
4187                         if (_len) {
4188                                 cap_auths[i].match.gids = kcalloc(_len, sizeof(u32),
4189                                                                   GFP_KERNEL);
4190                                 if (!cap_auths[i].match.gids) {
4191                                         pr_err_client(cl, "No memory for gids\n");
4192                                         goto fail;
4193                                 }
4194
4195                                 cap_auths[i].match.num_gids = _len;
4196                                 for (j = 0; j < _len; j++)
4197                                         ceph_decode_32_safe(&p, end,
4198                                                             cap_auths[i].match.gids[j],
4199                                                             bad);
4200                         }
4201
4202                         ceph_decode_32_safe(&p, end, _len, bad);
4203                         if (_len) {
4204                                 cap_auths[i].match.path = kcalloc(_len + 1, sizeof(char),
4205                                                                   GFP_KERNEL);
4206                                 if (!cap_auths[i].match.path) {
4207                                         pr_err_client(cl, "No memory for path\n");
4208                                         goto fail;
4209                                 }
4210                                 ceph_decode_copy(&p, cap_auths[i].match.path, _len);
4211
4212                                 /* Remove the tailing '/' */
4213                                 while (_len && cap_auths[i].match.path[_len - 1] == '/') {
4214                                         cap_auths[i].match.path[_len - 1] = '\0';
4215                                         _len -= 1;
4216                                 }
4217                         }
4218
4219                         ceph_decode_32_safe(&p, end, _len, bad);
4220                         if (_len) {
4221                                 cap_auths[i].match.fs_name = kcalloc(_len + 1, sizeof(char),
4222                                                                      GFP_KERNEL);
4223                                 if (!cap_auths[i].match.fs_name) {
4224                                         pr_err_client(cl, "No memory for fs_name\n");
4225                                         goto fail;
4226                                 }
4227                                 ceph_decode_copy(&p, cap_auths[i].match.fs_name, _len);
4228                         }
4229
4230                         ceph_decode_8_safe(&p, end, cap_auths[i].match.root_squash, bad);
4231                         ceph_decode_8_safe(&p, end, cap_auths[i].readable, bad);
4232                         ceph_decode_8_safe(&p, end, cap_auths[i].writeable, bad);
4233                         doutc(cl, "uid %lld, num_gids %u, path %s, fs_name %s, root_squash %d, readable %d, writeable %d\n",
4234                               cap_auths[i].match.uid, cap_auths[i].match.num_gids,
4235                               cap_auths[i].match.path, cap_auths[i].match.fs_name,
4236                               cap_auths[i].match.root_squash,
4237                               cap_auths[i].readable, cap_auths[i].writeable);
4238                 }
4239         }
4240
4241 skip_cap_auths:
4242         mutex_lock(&mdsc->mutex);
4243         if (op == CEPH_SESSION_OPEN) {
4244                 if (mdsc->s_cap_auths) {
4245                         for (i = 0; i < mdsc->s_cap_auths_num; i++) {
4246                                 kfree(mdsc->s_cap_auths[i].match.gids);
4247                                 kfree(mdsc->s_cap_auths[i].match.path);
4248                                 kfree(mdsc->s_cap_auths[i].match.fs_name);
4249                         }
4250                         kfree(mdsc->s_cap_auths);
4251                 }
4252                 mdsc->s_cap_auths_num = cap_auths_num;
4253                 mdsc->s_cap_auths = cap_auths;
4254         }
4255         if (op == CEPH_SESSION_CLOSE) {
4256                 ceph_get_mds_session(session);
4257                 __unregister_session(mdsc, session);
4258         }
4259         /* FIXME: this ttl calculation is generous */
4260         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
4261         mutex_unlock(&mdsc->mutex);
4262
4263         mutex_lock(&session->s_mutex);
4264
4265         doutc(cl, "mds%d %s %p state %s seq %llu\n", mds,
4266               ceph_session_op_name(op), session,
4267               ceph_session_state_name(session->s_state), seq);
4268
4269         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
4270                 session->s_state = CEPH_MDS_SESSION_OPEN;
4271                 pr_info_client(cl, "mds%d came back\n", session->s_mds);
4272         }
4273
4274         switch (op) {
4275         case CEPH_SESSION_OPEN:
4276                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
4277                         pr_info_client(cl, "mds%d reconnect success\n",
4278                                        session->s_mds);
4279
4280                 session->s_features = features;
4281                 if (session->s_state == CEPH_MDS_SESSION_OPEN) {
4282                         pr_notice_client(cl, "mds%d is already opened\n",
4283                                          session->s_mds);
4284                 } else {
4285                         session->s_state = CEPH_MDS_SESSION_OPEN;
4286                         renewed_caps(mdsc, session, 0);
4287                         if (test_bit(CEPHFS_FEATURE_METRIC_COLLECT,
4288                                      &session->s_features))
4289                                 metric_schedule_delayed(&mdsc->metric);
4290                 }
4291
4292                 /*
4293                  * The connection maybe broken and the session in client
4294                  * side has been reinitialized, need to update the seq
4295                  * anyway.
4296                  */
4297                 if (!session->s_seq && seq)
4298                         session->s_seq = seq;
4299
4300                 wake = 1;
4301                 if (mdsc->stopping)
4302                         __close_session(mdsc, session);
4303                 break;
4304
4305         case CEPH_SESSION_RENEWCAPS:
4306                 if (session->s_renew_seq == seq)
4307                         renewed_caps(mdsc, session, 1);
4308                 break;
4309
4310         case CEPH_SESSION_CLOSE:
4311                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
4312                         pr_info_client(cl, "mds%d reconnect denied\n",
4313                                        session->s_mds);
4314                 session->s_state = CEPH_MDS_SESSION_CLOSED;
4315                 cleanup_session_requests(mdsc, session);
4316                 remove_session_caps(session);
4317                 wake = 2; /* for good measure */
4318                 wake_up_all(&mdsc->session_close_wq);
4319                 break;
4320
4321         case CEPH_SESSION_STALE:
4322                 pr_info_client(cl, "mds%d caps went stale, renewing\n",
4323                                session->s_mds);
4324                 atomic_inc(&session->s_cap_gen);
4325                 session->s_cap_ttl = jiffies - 1;
4326                 send_renew_caps(mdsc, session);
4327                 break;
4328
4329         case CEPH_SESSION_RECALL_STATE:
4330                 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
4331                 break;
4332
4333         case CEPH_SESSION_FLUSHMSG:
4334                 /* flush cap releases */
4335                 spin_lock(&session->s_cap_lock);
4336                 if (session->s_num_cap_releases)
4337                         ceph_flush_session_cap_releases(mdsc, session);
4338                 spin_unlock(&session->s_cap_lock);
4339
4340                 send_flushmsg_ack(mdsc, session, seq);
4341                 break;
4342
4343         case CEPH_SESSION_FORCE_RO:
4344                 doutc(cl, "force_session_readonly %p\n", session);
4345                 spin_lock(&session->s_cap_lock);
4346                 session->s_readonly = true;
4347                 spin_unlock(&session->s_cap_lock);
4348                 wake_up_session_caps(session, FORCE_RO);
4349                 break;
4350
4351         case CEPH_SESSION_REJECT:
4352                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
4353                 pr_info_client(cl, "mds%d rejected session\n",
4354                                session->s_mds);
4355                 session->s_state = CEPH_MDS_SESSION_REJECTED;
4356                 cleanup_session_requests(mdsc, session);
4357                 remove_session_caps(session);
4358                 if (blocklisted)
4359                         mdsc->fsc->blocklisted = true;
4360                 wake = 2; /* for good measure */
4361                 break;
4362
4363         default:
4364                 pr_err_client(cl, "bad op %d mds%d\n", op, mds);
4365                 WARN_ON(1);
4366         }
4367
4368         mutex_unlock(&session->s_mutex);
4369         if (wake) {
4370                 mutex_lock(&mdsc->mutex);
4371                 __wake_requests(mdsc, &session->s_waiting);
4372                 if (wake == 2)
4373                         kick_requests(mdsc, mds);
4374                 mutex_unlock(&mdsc->mutex);
4375         }
4376         if (op == CEPH_SESSION_CLOSE)
4377                 ceph_put_mds_session(session);
4378         return;
4379
4380 bad:
4381         pr_err_client(cl, "corrupt message mds%d len %d\n", mds,
4382                       (int)msg->front.iov_len);
4383         ceph_msg_dump(msg);
4384 fail:
4385         for (i = 0; i < cap_auths_num; i++) {
4386                 kfree(cap_auths[i].match.gids);
4387                 kfree(cap_auths[i].match.path);
4388                 kfree(cap_auths[i].match.fs_name);
4389         }
4390         kfree(cap_auths);
4391         return;
4392 }
4393
4394 void ceph_mdsc_release_dir_caps(struct ceph_mds_request *req)
4395 {
4396         struct ceph_client *cl = req->r_mdsc->fsc->client;
4397         int dcaps;
4398
4399         dcaps = xchg(&req->r_dir_caps, 0);
4400         if (dcaps) {
4401                 doutc(cl, "releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
4402                 ceph_put_cap_refs(ceph_inode(req->r_parent), dcaps);
4403         }
4404 }
4405
4406 void ceph_mdsc_release_dir_caps_async(struct ceph_mds_request *req)
4407 {
4408         struct ceph_client *cl = req->r_mdsc->fsc->client;
4409         int dcaps;
4410
4411         dcaps = xchg(&req->r_dir_caps, 0);
4412         if (dcaps) {
4413                 doutc(cl, "releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
4414                 ceph_put_cap_refs_async(ceph_inode(req->r_parent), dcaps);
4415         }
4416 }
4417
4418 /*
4419  * called under session->mutex.
4420  */
4421 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
4422                                    struct ceph_mds_session *session)
4423 {
4424         struct ceph_mds_request *req, *nreq;
4425         struct rb_node *p;
4426
4427         doutc(mdsc->fsc->client, "mds%d\n", session->s_mds);
4428
4429         mutex_lock(&mdsc->mutex);
4430         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item)
4431                 __send_request(session, req, true);
4432
4433         /*
4434          * also re-send old requests when MDS enters reconnect stage. So that MDS
4435          * can process completed request in clientreplay stage.
4436          */
4437         p = rb_first(&mdsc->request_tree);
4438         while (p) {
4439                 req = rb_entry(p, struct ceph_mds_request, r_node);
4440                 p = rb_next(p);
4441                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
4442                         continue;
4443                 if (req->r_attempts == 0)
4444                         continue; /* only old requests */
4445                 if (!req->r_session)
4446                         continue;
4447                 if (req->r_session->s_mds != session->s_mds)
4448                         continue;
4449
4450                 ceph_mdsc_release_dir_caps_async(req);
4451
4452                 __send_request(session, req, true);
4453         }
4454         mutex_unlock(&mdsc->mutex);
4455 }
4456
4457 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
4458 {
4459         struct ceph_msg *reply;
4460         struct ceph_pagelist *_pagelist;
4461         struct page *page;
4462         __le32 *addr;
4463         int err = -ENOMEM;
4464
4465         if (!recon_state->allow_multi)
4466                 return -ENOSPC;
4467
4468         /* can't handle message that contains both caps and realm */
4469         BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
4470
4471         /* pre-allocate new pagelist */
4472         _pagelist = ceph_pagelist_alloc(GFP_NOFS);
4473         if (!_pagelist)
4474                 return -ENOMEM;
4475
4476         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
4477         if (!reply)
4478                 goto fail_msg;
4479
4480         /* placeholder for nr_caps */
4481         err = ceph_pagelist_encode_32(_pagelist, 0);
4482         if (err < 0)
4483                 goto fail;
4484
4485         if (recon_state->nr_caps) {
4486                 /* currently encoding caps */
4487                 err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
4488                 if (err)
4489                         goto fail;
4490         } else {
4491                 /* placeholder for nr_realms (currently encoding relams) */
4492                 err = ceph_pagelist_encode_32(_pagelist, 0);
4493                 if (err < 0)
4494                         goto fail;
4495         }
4496
4497         err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
4498         if (err)
4499                 goto fail;
4500
4501         page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
4502         addr = kmap_atomic(page);
4503         if (recon_state->nr_caps) {
4504                 /* currently encoding caps */
4505                 *addr = cpu_to_le32(recon_state->nr_caps);
4506         } else {
4507                 /* currently encoding relams */
4508                 *(addr + 1) = cpu_to_le32(recon_state->nr_realms);
4509         }
4510         kunmap_atomic(addr);
4511
4512         reply->hdr.version = cpu_to_le16(5);
4513         reply->hdr.compat_version = cpu_to_le16(4);
4514
4515         reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
4516         ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
4517
4518         ceph_con_send(&recon_state->session->s_con, reply);
4519         ceph_pagelist_release(recon_state->pagelist);
4520
4521         recon_state->pagelist = _pagelist;
4522         recon_state->nr_caps = 0;
4523         recon_state->nr_realms = 0;
4524         recon_state->msg_version = 5;
4525         return 0;
4526 fail:
4527         ceph_msg_put(reply);
4528 fail_msg:
4529         ceph_pagelist_release(_pagelist);
4530         return err;
4531 }
4532
4533 static struct dentry* d_find_primary(struct inode *inode)
4534 {
4535         struct dentry *alias, *dn = NULL;
4536
4537         if (hlist_empty(&inode->i_dentry))
4538                 return NULL;
4539
4540         spin_lock(&inode->i_lock);
4541         if (hlist_empty(&inode->i_dentry))
4542                 goto out_unlock;
4543
4544         if (S_ISDIR(inode->i_mode)) {
4545                 alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
4546                 if (!IS_ROOT(alias))
4547                         dn = dget(alias);
4548                 goto out_unlock;
4549         }
4550
4551         hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
4552                 spin_lock(&alias->d_lock);
4553                 if (!d_unhashed(alias) &&
4554                     (ceph_dentry(alias)->flags & CEPH_DENTRY_PRIMARY_LINK)) {
4555                         dn = dget_dlock(alias);
4556                 }
4557                 spin_unlock(&alias->d_lock);
4558                 if (dn)
4559                         break;
4560         }
4561 out_unlock:
4562         spin_unlock(&inode->i_lock);
4563         return dn;
4564 }
4565
4566 /*
4567  * Encode information about a cap for a reconnect with the MDS.
4568  */
4569 static int reconnect_caps_cb(struct inode *inode, int mds, void *arg)
4570 {
4571         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
4572         struct ceph_client *cl = ceph_inode_to_client(inode);
4573         union {
4574                 struct ceph_mds_cap_reconnect v2;
4575                 struct ceph_mds_cap_reconnect_v1 v1;
4576         } rec;
4577         struct ceph_inode_info *ci = ceph_inode(inode);
4578         struct ceph_reconnect_state *recon_state = arg;
4579         struct ceph_pagelist *pagelist = recon_state->pagelist;
4580         struct dentry *dentry;
4581         struct ceph_cap *cap;
4582         char *path;
4583         int pathlen = 0, err;
4584         u64 pathbase;
4585         u64 snap_follows;
4586
4587         dentry = d_find_primary(inode);
4588         if (dentry) {
4589                 /* set pathbase to parent dir when msg_version >= 2 */
4590                 path = ceph_mdsc_build_path(mdsc, dentry, &pathlen, &pathbase,
4591                                             recon_state->msg_version >= 2);
4592                 dput(dentry);
4593                 if (IS_ERR(path)) {
4594                         err = PTR_ERR(path);
4595                         goto out_err;
4596                 }
4597         } else {
4598                 path = NULL;
4599                 pathbase = 0;
4600         }
4601
4602         spin_lock(&ci->i_ceph_lock);
4603         cap = __get_cap_for_mds(ci, mds);
4604         if (!cap) {
4605                 spin_unlock(&ci->i_ceph_lock);
4606                 err = 0;
4607                 goto out_err;
4608         }
4609         doutc(cl, " adding %p ino %llx.%llx cap %p %lld %s\n", inode,
4610               ceph_vinop(inode), cap, cap->cap_id,
4611               ceph_cap_string(cap->issued));
4612
4613         cap->seq = 0;        /* reset cap seq */
4614         cap->issue_seq = 0;  /* and issue_seq */
4615         cap->mseq = 0;       /* and migrate_seq */
4616         cap->cap_gen = atomic_read(&cap->session->s_cap_gen);
4617
4618         /* These are lost when the session goes away */
4619         if (S_ISDIR(inode->i_mode)) {
4620                 if (cap->issued & CEPH_CAP_DIR_CREATE) {
4621                         ceph_put_string(rcu_dereference_raw(ci->i_cached_layout.pool_ns));
4622                         memset(&ci->i_cached_layout, 0, sizeof(ci->i_cached_layout));
4623                 }
4624                 cap->issued &= ~CEPH_CAP_ANY_DIR_OPS;
4625         }
4626
4627         if (recon_state->msg_version >= 2) {
4628                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
4629                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
4630                 rec.v2.issued = cpu_to_le32(cap->issued);
4631                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
4632                 rec.v2.pathbase = cpu_to_le64(pathbase);
4633                 rec.v2.flock_len = (__force __le32)
4634                         ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
4635         } else {
4636                 struct timespec64 ts;
4637
4638                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
4639                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
4640                 rec.v1.issued = cpu_to_le32(cap->issued);
4641                 rec.v1.size = cpu_to_le64(i_size_read(inode));
4642                 ts = inode_get_mtime(inode);
4643                 ceph_encode_timespec64(&rec.v1.mtime, &ts);
4644                 ts = inode_get_atime(inode);
4645                 ceph_encode_timespec64(&rec.v1.atime, &ts);
4646                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
4647                 rec.v1.pathbase = cpu_to_le64(pathbase);
4648         }
4649
4650         if (list_empty(&ci->i_cap_snaps)) {
4651                 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
4652         } else {
4653                 struct ceph_cap_snap *capsnap =
4654                         list_first_entry(&ci->i_cap_snaps,
4655                                          struct ceph_cap_snap, ci_item);
4656                 snap_follows = capsnap->follows;
4657         }
4658         spin_unlock(&ci->i_ceph_lock);
4659
4660         if (recon_state->msg_version >= 2) {
4661                 int num_fcntl_locks, num_flock_locks;
4662                 struct ceph_filelock *flocks = NULL;
4663                 size_t struct_len, total_len = sizeof(u64);
4664                 u8 struct_v = 0;
4665
4666 encode_again:
4667                 if (rec.v2.flock_len) {
4668                         ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
4669                 } else {
4670                         num_fcntl_locks = 0;
4671                         num_flock_locks = 0;
4672                 }
4673                 if (num_fcntl_locks + num_flock_locks > 0) {
4674                         flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
4675                                                sizeof(struct ceph_filelock),
4676                                                GFP_NOFS);
4677                         if (!flocks) {
4678                                 err = -ENOMEM;
4679                                 goto out_err;
4680                         }
4681                         err = ceph_encode_locks_to_buffer(inode, flocks,
4682                                                           num_fcntl_locks,
4683                                                           num_flock_locks);
4684                         if (err) {
4685                                 kfree(flocks);
4686                                 flocks = NULL;
4687                                 if (err == -ENOSPC)
4688                                         goto encode_again;
4689                                 goto out_err;
4690                         }
4691                 } else {
4692                         kfree(flocks);
4693                         flocks = NULL;
4694                 }
4695
4696                 if (recon_state->msg_version >= 3) {
4697                         /* version, compat_version and struct_len */
4698                         total_len += 2 * sizeof(u8) + sizeof(u32);
4699                         struct_v = 2;
4700                 }
4701                 /*
4702                  * number of encoded locks is stable, so copy to pagelist
4703                  */
4704                 struct_len = 2 * sizeof(u32) +
4705                             (num_fcntl_locks + num_flock_locks) *
4706                             sizeof(struct ceph_filelock);
4707                 rec.v2.flock_len = cpu_to_le32(struct_len);
4708
4709                 struct_len += sizeof(u32) + pathlen + sizeof(rec.v2);
4710
4711                 if (struct_v >= 2)
4712                         struct_len += sizeof(u64); /* snap_follows */
4713
4714                 total_len += struct_len;
4715
4716                 if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
4717                         err = send_reconnect_partial(recon_state);
4718                         if (err)
4719                                 goto out_freeflocks;
4720                         pagelist = recon_state->pagelist;
4721                 }
4722
4723                 err = ceph_pagelist_reserve(pagelist, total_len);
4724                 if (err)
4725                         goto out_freeflocks;
4726
4727                 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
4728                 if (recon_state->msg_version >= 3) {
4729                         ceph_pagelist_encode_8(pagelist, struct_v);
4730                         ceph_pagelist_encode_8(pagelist, 1);
4731                         ceph_pagelist_encode_32(pagelist, struct_len);
4732                 }
4733                 ceph_pagelist_encode_string(pagelist, path, pathlen);
4734                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
4735                 ceph_locks_to_pagelist(flocks, pagelist,
4736                                        num_fcntl_locks, num_flock_locks);
4737                 if (struct_v >= 2)
4738                         ceph_pagelist_encode_64(pagelist, snap_follows);
4739 out_freeflocks:
4740                 kfree(flocks);
4741         } else {
4742                 err = ceph_pagelist_reserve(pagelist,
4743                                             sizeof(u64) + sizeof(u32) +
4744                                             pathlen + sizeof(rec.v1));
4745                 if (err)
4746                         goto out_err;
4747
4748                 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
4749                 ceph_pagelist_encode_string(pagelist, path, pathlen);
4750                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
4751         }
4752
4753 out_err:
4754         ceph_mdsc_free_path(path, pathlen);
4755         if (!err)
4756                 recon_state->nr_caps++;
4757         return err;
4758 }
4759
4760 static int encode_snap_realms(struct ceph_mds_client *mdsc,
4761                               struct ceph_reconnect_state *recon_state)
4762 {
4763         struct rb_node *p;
4764         struct ceph_pagelist *pagelist = recon_state->pagelist;
4765         struct ceph_client *cl = mdsc->fsc->client;
4766         int err = 0;
4767
4768         if (recon_state->msg_version >= 4) {
4769                 err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
4770                 if (err < 0)
4771                         goto fail;
4772         }
4773
4774         /*
4775          * snaprealms.  we provide mds with the ino, seq (version), and
4776          * parent for all of our realms.  If the mds has any newer info,
4777          * it will tell us.
4778          */
4779         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
4780                 struct ceph_snap_realm *realm =
4781                        rb_entry(p, struct ceph_snap_realm, node);
4782                 struct ceph_mds_snaprealm_reconnect sr_rec;
4783
4784                 if (recon_state->msg_version >= 4) {
4785                         size_t need = sizeof(u8) * 2 + sizeof(u32) +
4786                                       sizeof(sr_rec);
4787
4788                         if (pagelist->length + need > RECONNECT_MAX_SIZE) {
4789                                 err = send_reconnect_partial(recon_state);
4790                                 if (err)
4791                                         goto fail;
4792                                 pagelist = recon_state->pagelist;
4793                         }
4794
4795                         err = ceph_pagelist_reserve(pagelist, need);
4796                         if (err)
4797                                 goto fail;
4798
4799                         ceph_pagelist_encode_8(pagelist, 1);
4800                         ceph_pagelist_encode_8(pagelist, 1);
4801                         ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
4802                 }
4803
4804                 doutc(cl, " adding snap realm %llx seq %lld parent %llx\n",
4805                       realm->ino, realm->seq, realm->parent_ino);
4806                 sr_rec.ino = cpu_to_le64(realm->ino);
4807                 sr_rec.seq = cpu_to_le64(realm->seq);
4808                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
4809
4810                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
4811                 if (err)
4812                         goto fail;
4813
4814                 recon_state->nr_realms++;
4815         }
4816 fail:
4817         return err;
4818 }
4819
4820
4821 /*
4822  * If an MDS fails and recovers, clients need to reconnect in order to
4823  * reestablish shared state.  This includes all caps issued through
4824  * this session _and_ the snap_realm hierarchy.  Because it's not
4825  * clear which snap realms the mds cares about, we send everything we
4826  * know about.. that ensures we'll then get any new info the
4827  * recovering MDS might have.
4828  *
4829  * This is a relatively heavyweight operation, but it's rare.
4830  */
4831 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
4832                                struct ceph_mds_session *session)
4833 {
4834         struct ceph_client *cl = mdsc->fsc->client;
4835         struct ceph_msg *reply;
4836         int mds = session->s_mds;
4837         int err = -ENOMEM;
4838         struct ceph_reconnect_state recon_state = {
4839                 .session = session,
4840         };
4841         LIST_HEAD(dispose);
4842
4843         pr_info_client(cl, "mds%d reconnect start\n", mds);
4844
4845         recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
4846         if (!recon_state.pagelist)
4847                 goto fail_nopagelist;
4848
4849         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
4850         if (!reply)
4851                 goto fail_nomsg;
4852
4853         xa_destroy(&session->s_delegated_inos);
4854
4855         mutex_lock(&session->s_mutex);
4856         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
4857         session->s_seq = 0;
4858
4859         doutc(cl, "session %p state %s\n", session,
4860               ceph_session_state_name(session->s_state));
4861
4862         atomic_inc(&session->s_cap_gen);
4863
4864         spin_lock(&session->s_cap_lock);
4865         /* don't know if session is readonly */
4866         session->s_readonly = 0;
4867         /*
4868          * notify __ceph_remove_cap() that we are composing cap reconnect.
4869          * If a cap get released before being added to the cap reconnect,
4870          * __ceph_remove_cap() should skip queuing cap release.
4871          */
4872         session->s_cap_reconnect = 1;
4873         /* drop old cap expires; we're about to reestablish that state */
4874         detach_cap_releases(session, &dispose);
4875         spin_unlock(&session->s_cap_lock);
4876         dispose_cap_releases(mdsc, &dispose);
4877
4878         /* trim unused caps to reduce MDS's cache rejoin time */
4879         if (mdsc->fsc->sb->s_root)
4880                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
4881
4882         ceph_con_close(&session->s_con);
4883         ceph_con_open(&session->s_con,
4884                       CEPH_ENTITY_TYPE_MDS, mds,
4885                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
4886
4887         /* replay unsafe requests */
4888         replay_unsafe_requests(mdsc, session);
4889
4890         ceph_early_kick_flushing_caps(mdsc, session);
4891
4892         down_read(&mdsc->snap_rwsem);
4893
4894         /* placeholder for nr_caps */
4895         err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
4896         if (err)
4897                 goto fail;
4898
4899         if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
4900                 recon_state.msg_version = 3;
4901                 recon_state.allow_multi = true;
4902         } else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
4903                 recon_state.msg_version = 3;
4904         } else {
4905                 recon_state.msg_version = 2;
4906         }
4907         /* traverse this session's caps */
4908         err = ceph_iterate_session_caps(session, reconnect_caps_cb, &recon_state);
4909
4910         spin_lock(&session->s_cap_lock);
4911         session->s_cap_reconnect = 0;
4912         spin_unlock(&session->s_cap_lock);
4913
4914         if (err < 0)
4915                 goto fail;
4916
4917         /* check if all realms can be encoded into current message */
4918         if (mdsc->num_snap_realms) {
4919                 size_t total_len =
4920                         recon_state.pagelist->length +
4921                         mdsc->num_snap_realms *
4922                         sizeof(struct ceph_mds_snaprealm_reconnect);
4923                 if (recon_state.msg_version >= 4) {
4924                         /* number of realms */
4925                         total_len += sizeof(u32);
4926                         /* version, compat_version and struct_len */
4927                         total_len += mdsc->num_snap_realms *
4928                                      (2 * sizeof(u8) + sizeof(u32));
4929                 }
4930                 if (total_len > RECONNECT_MAX_SIZE) {
4931                         if (!recon_state.allow_multi) {
4932                                 err = -ENOSPC;
4933                                 goto fail;
4934                         }
4935                         if (recon_state.nr_caps) {
4936                                 err = send_reconnect_partial(&recon_state);
4937                                 if (err)
4938                                         goto fail;
4939                         }
4940                         recon_state.msg_version = 5;
4941                 }
4942         }
4943
4944         err = encode_snap_realms(mdsc, &recon_state);
4945         if (err < 0)
4946                 goto fail;
4947
4948         if (recon_state.msg_version >= 5) {
4949                 err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
4950                 if (err < 0)
4951                         goto fail;
4952         }
4953
4954         if (recon_state.nr_caps || recon_state.nr_realms) {
4955                 struct page *page =
4956                         list_first_entry(&recon_state.pagelist->head,
4957                                         struct page, lru);
4958                 __le32 *addr = kmap_atomic(page);
4959                 if (recon_state.nr_caps) {
4960                         WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
4961                         *addr = cpu_to_le32(recon_state.nr_caps);
4962                 } else if (recon_state.msg_version >= 4) {
4963                         *(addr + 1) = cpu_to_le32(recon_state.nr_realms);
4964                 }
4965                 kunmap_atomic(addr);
4966         }
4967
4968         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
4969         if (recon_state.msg_version >= 4)
4970                 reply->hdr.compat_version = cpu_to_le16(4);
4971
4972         reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
4973         ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
4974
4975         ceph_con_send(&session->s_con, reply);
4976
4977         mutex_unlock(&session->s_mutex);
4978
4979         mutex_lock(&mdsc->mutex);
4980         __wake_requests(mdsc, &session->s_waiting);
4981         mutex_unlock(&mdsc->mutex);
4982
4983         up_read(&mdsc->snap_rwsem);
4984         ceph_pagelist_release(recon_state.pagelist);
4985         return;
4986
4987 fail:
4988         ceph_msg_put(reply);
4989         up_read(&mdsc->snap_rwsem);
4990         mutex_unlock(&session->s_mutex);
4991 fail_nomsg:
4992         ceph_pagelist_release(recon_state.pagelist);
4993 fail_nopagelist:
4994         pr_err_client(cl, "error %d preparing reconnect for mds%d\n",
4995                       err, mds);
4996         return;
4997 }
4998
4999
5000 /*
5001  * compare old and new mdsmaps, kicking requests
5002  * and closing out old connections as necessary
5003  *
5004  * called under mdsc->mutex.
5005  */
5006 static void check_new_map(struct ceph_mds_client *mdsc,
5007                           struct ceph_mdsmap *newmap,
5008                           struct ceph_mdsmap *oldmap)
5009 {
5010         int i, j, err;
5011         int oldstate, newstate;
5012         struct ceph_mds_session *s;
5013         unsigned long targets[DIV_ROUND_UP(CEPH_MAX_MDS, sizeof(unsigned long))] = {0};
5014         struct ceph_client *cl = mdsc->fsc->client;
5015
5016         doutc(cl, "new %u old %u\n", newmap->m_epoch, oldmap->m_epoch);
5017
5018         if (newmap->m_info) {
5019                 for (i = 0; i < newmap->possible_max_rank; i++) {
5020                         for (j = 0; j < newmap->m_info[i].num_export_targets; j++)
5021                                 set_bit(newmap->m_info[i].export_targets[j], targets);
5022                 }
5023         }
5024
5025         for (i = 0; i < oldmap->possible_max_rank && i < mdsc->max_sessions; i++) {
5026                 if (!mdsc->sessions[i])
5027                         continue;
5028                 s = mdsc->sessions[i];
5029                 oldstate = ceph_mdsmap_get_state(oldmap, i);
5030                 newstate = ceph_mdsmap_get_state(newmap, i);
5031
5032                 doutc(cl, "mds%d state %s%s -> %s%s (session %s)\n",
5033                       i, ceph_mds_state_name(oldstate),
5034                       ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
5035                       ceph_mds_state_name(newstate),
5036                       ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
5037                       ceph_session_state_name(s->s_state));
5038
5039                 if (i >= newmap->possible_max_rank) {
5040                         /* force close session for stopped mds */
5041                         ceph_get_mds_session(s);
5042                         __unregister_session(mdsc, s);
5043                         __wake_requests(mdsc, &s->s_waiting);
5044                         mutex_unlock(&mdsc->mutex);
5045
5046                         mutex_lock(&s->s_mutex);
5047                         cleanup_session_requests(mdsc, s);
5048                         remove_session_caps(s);
5049                         mutex_unlock(&s->s_mutex);
5050
5051                         ceph_put_mds_session(s);
5052
5053                         mutex_lock(&mdsc->mutex);
5054                         kick_requests(mdsc, i);
5055                         continue;
5056                 }
5057
5058                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
5059                            ceph_mdsmap_get_addr(newmap, i),
5060                            sizeof(struct ceph_entity_addr))) {
5061                         /* just close it */
5062                         mutex_unlock(&mdsc->mutex);
5063                         mutex_lock(&s->s_mutex);
5064                         mutex_lock(&mdsc->mutex);
5065                         ceph_con_close(&s->s_con);
5066                         mutex_unlock(&s->s_mutex);
5067                         s->s_state = CEPH_MDS_SESSION_RESTARTING;
5068                 } else if (oldstate == newstate) {
5069                         continue;  /* nothing new with this mds */
5070                 }
5071
5072                 /*
5073                  * send reconnect?
5074                  */
5075                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
5076                     newstate >= CEPH_MDS_STATE_RECONNECT) {
5077                         mutex_unlock(&mdsc->mutex);
5078                         clear_bit(i, targets);
5079                         send_mds_reconnect(mdsc, s);
5080                         mutex_lock(&mdsc->mutex);
5081                 }
5082
5083                 /*
5084                  * kick request on any mds that has gone active.
5085                  */
5086                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
5087                     newstate >= CEPH_MDS_STATE_ACTIVE) {
5088                         if (oldstate != CEPH_MDS_STATE_CREATING &&
5089                             oldstate != CEPH_MDS_STATE_STARTING)
5090                                 pr_info_client(cl, "mds%d recovery completed\n",
5091                                                s->s_mds);
5092                         kick_requests(mdsc, i);
5093                         mutex_unlock(&mdsc->mutex);
5094                         mutex_lock(&s->s_mutex);
5095                         mutex_lock(&mdsc->mutex);
5096                         ceph_kick_flushing_caps(mdsc, s);
5097                         mutex_unlock(&s->s_mutex);
5098                         wake_up_session_caps(s, RECONNECT);
5099                 }
5100         }
5101
5102         /*
5103          * Only open and reconnect sessions that don't exist yet.
5104          */
5105         for (i = 0; i < newmap->possible_max_rank; i++) {
5106                 /*
5107                  * In case the import MDS is crashed just after
5108                  * the EImportStart journal is flushed, so when
5109                  * a standby MDS takes over it and is replaying
5110                  * the EImportStart journal the new MDS daemon
5111                  * will wait the client to reconnect it, but the
5112                  * client may never register/open the session yet.
5113                  *
5114                  * Will try to reconnect that MDS daemon if the
5115                  * rank number is in the export targets array and
5116                  * is the up:reconnect state.
5117                  */
5118                 newstate = ceph_mdsmap_get_state(newmap, i);
5119                 if (!test_bit(i, targets) || newstate != CEPH_MDS_STATE_RECONNECT)
5120                         continue;
5121
5122                 /*
5123                  * The session maybe registered and opened by some
5124                  * requests which were choosing random MDSes during
5125                  * the mdsc->mutex's unlock/lock gap below in rare
5126                  * case. But the related MDS daemon will just queue
5127                  * that requests and be still waiting for the client's
5128                  * reconnection request in up:reconnect state.
5129                  */
5130                 s = __ceph_lookup_mds_session(mdsc, i);
5131                 if (likely(!s)) {
5132                         s = __open_export_target_session(mdsc, i);
5133                         if (IS_ERR(s)) {
5134                                 err = PTR_ERR(s);
5135                                 pr_err_client(cl,
5136                                               "failed to open export target session, err %d\n",
5137                                               err);
5138                                 continue;
5139                         }
5140                 }
5141                 doutc(cl, "send reconnect to export target mds.%d\n", i);
5142                 mutex_unlock(&mdsc->mutex);
5143                 send_mds_reconnect(mdsc, s);
5144                 ceph_put_mds_session(s);
5145                 mutex_lock(&mdsc->mutex);
5146         }
5147
5148         for (i = 0; i < newmap->possible_max_rank && i < mdsc->max_sessions; i++) {
5149                 s = mdsc->sessions[i];
5150                 if (!s)
5151                         continue;
5152                 if (!ceph_mdsmap_is_laggy(newmap, i))
5153                         continue;
5154                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
5155                     s->s_state == CEPH_MDS_SESSION_HUNG ||
5156                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
5157                         doutc(cl, " connecting to export targets of laggy mds%d\n", i);
5158                         __open_export_target_sessions(mdsc, s);
5159                 }
5160         }
5161 }
5162
5163
5164
5165 /*
5166  * leases
5167  */
5168
5169 /*
5170  * caller must hold session s_mutex, dentry->d_lock
5171  */
5172 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
5173 {
5174         struct ceph_dentry_info *di = ceph_dentry(dentry);
5175
5176         ceph_put_mds_session(di->lease_session);
5177         di->lease_session = NULL;
5178 }
5179
5180 static void handle_lease(struct ceph_mds_client *mdsc,
5181                          struct ceph_mds_session *session,
5182                          struct ceph_msg *msg)
5183 {
5184         struct ceph_client *cl = mdsc->fsc->client;
5185         struct super_block *sb = mdsc->fsc->sb;
5186         struct inode *inode;
5187         struct dentry *parent, *dentry;
5188         struct ceph_dentry_info *di;
5189         int mds = session->s_mds;
5190         struct ceph_mds_lease *h = msg->front.iov_base;
5191         u32 seq;
5192         struct ceph_vino vino;
5193         struct qstr dname;
5194         int release = 0;
5195
5196         doutc(cl, "from mds%d\n", mds);
5197
5198         if (!ceph_inc_mds_stopping_blocker(mdsc, session))
5199                 return;
5200
5201         /* decode */
5202         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
5203                 goto bad;
5204         vino.ino = le64_to_cpu(h->ino);
5205         vino.snap = CEPH_NOSNAP;
5206         seq = le32_to_cpu(h->seq);
5207         dname.len = get_unaligned_le32(h + 1);
5208         if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
5209                 goto bad;
5210         dname.name = (void *)(h + 1) + sizeof(u32);
5211
5212         /* lookup inode */
5213         inode = ceph_find_inode(sb, vino);
5214         doutc(cl, "%s, ino %llx %p %.*s\n", ceph_lease_op_name(h->action),
5215               vino.ino, inode, dname.len, dname.name);
5216
5217         mutex_lock(&session->s_mutex);
5218         if (!inode) {
5219                 doutc(cl, "no inode %llx\n", vino.ino);
5220                 goto release;
5221         }
5222
5223         /* dentry */
5224         parent = d_find_alias(inode);
5225         if (!parent) {
5226                 doutc(cl, "no parent dentry on inode %p\n", inode);
5227                 WARN_ON(1);
5228                 goto release;  /* hrm... */
5229         }
5230         dname.hash = full_name_hash(parent, dname.name, dname.len);
5231         dentry = d_lookup(parent, &dname);
5232         dput(parent);
5233         if (!dentry)
5234                 goto release;
5235
5236         spin_lock(&dentry->d_lock);
5237         di = ceph_dentry(dentry);
5238         switch (h->action) {
5239         case CEPH_MDS_LEASE_REVOKE:
5240                 if (di->lease_session == session) {
5241                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
5242                                 h->seq = cpu_to_le32(di->lease_seq);
5243                         __ceph_mdsc_drop_dentry_lease(dentry);
5244                 }
5245                 release = 1;
5246                 break;
5247
5248         case CEPH_MDS_LEASE_RENEW:
5249                 if (di->lease_session == session &&
5250                     di->lease_gen == atomic_read(&session->s_cap_gen) &&
5251                     di->lease_renew_from &&
5252                     di->lease_renew_after == 0) {
5253                         unsigned long duration =
5254                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
5255
5256                         di->lease_seq = seq;
5257                         di->time = di->lease_renew_from + duration;
5258                         di->lease_renew_after = di->lease_renew_from +
5259                                 (duration >> 1);
5260                         di->lease_renew_from = 0;
5261                 }
5262                 break;
5263         }
5264         spin_unlock(&dentry->d_lock);
5265         dput(dentry);
5266
5267         if (!release)
5268                 goto out;
5269
5270 release:
5271         /* let's just reuse the same message */
5272         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
5273         ceph_msg_get(msg);
5274         ceph_con_send(&session->s_con, msg);
5275
5276 out:
5277         mutex_unlock(&session->s_mutex);
5278         iput(inode);
5279
5280         ceph_dec_mds_stopping_blocker(mdsc);
5281         return;
5282
5283 bad:
5284         ceph_dec_mds_stopping_blocker(mdsc);
5285
5286         pr_err_client(cl, "corrupt lease message\n");
5287         ceph_msg_dump(msg);
5288 }
5289
5290 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
5291                               struct dentry *dentry, char action,
5292                               u32 seq)
5293 {
5294         struct ceph_client *cl = session->s_mdsc->fsc->client;
5295         struct ceph_msg *msg;
5296         struct ceph_mds_lease *lease;
5297         struct inode *dir;
5298         int len = sizeof(*lease) + sizeof(u32) + NAME_MAX;
5299
5300         doutc(cl, "identry %p %s to mds%d\n", dentry, ceph_lease_op_name(action),
5301               session->s_mds);
5302
5303         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
5304         if (!msg)
5305                 return;
5306         lease = msg->front.iov_base;
5307         lease->action = action;
5308         lease->seq = cpu_to_le32(seq);
5309
5310         spin_lock(&dentry->d_lock);
5311         dir = d_inode(dentry->d_parent);
5312         lease->ino = cpu_to_le64(ceph_ino(dir));
5313         lease->first = lease->last = cpu_to_le64(ceph_snap(dir));
5314
5315         put_unaligned_le32(dentry->d_name.len, lease + 1);
5316         memcpy((void *)(lease + 1) + 4,
5317                dentry->d_name.name, dentry->d_name.len);
5318         spin_unlock(&dentry->d_lock);
5319
5320         ceph_con_send(&session->s_con, msg);
5321 }
5322
5323 /*
5324  * lock unlock the session, to wait ongoing session activities
5325  */
5326 static void lock_unlock_session(struct ceph_mds_session *s)
5327 {
5328         mutex_lock(&s->s_mutex);
5329         mutex_unlock(&s->s_mutex);
5330 }
5331
5332 static void maybe_recover_session(struct ceph_mds_client *mdsc)
5333 {
5334         struct ceph_client *cl = mdsc->fsc->client;
5335         struct ceph_fs_client *fsc = mdsc->fsc;
5336
5337         if (!ceph_test_mount_opt(fsc, CLEANRECOVER))
5338                 return;
5339
5340         if (READ_ONCE(fsc->mount_state) != CEPH_MOUNT_MOUNTED)
5341                 return;
5342
5343         if (!READ_ONCE(fsc->blocklisted))
5344                 return;
5345
5346         pr_info_client(cl, "auto reconnect after blocklisted\n");
5347         ceph_force_reconnect(fsc->sb);
5348 }
5349
5350 bool check_session_state(struct ceph_mds_session *s)
5351 {
5352         struct ceph_client *cl = s->s_mdsc->fsc->client;
5353
5354         switch (s->s_state) {
5355         case CEPH_MDS_SESSION_OPEN:
5356                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
5357                         s->s_state = CEPH_MDS_SESSION_HUNG;
5358                         pr_info_client(cl, "mds%d hung\n", s->s_mds);
5359                 }
5360                 break;
5361         case CEPH_MDS_SESSION_CLOSING:
5362         case CEPH_MDS_SESSION_NEW:
5363         case CEPH_MDS_SESSION_RESTARTING:
5364         case CEPH_MDS_SESSION_CLOSED:
5365         case CEPH_MDS_SESSION_REJECTED:
5366                 return false;
5367         }
5368
5369         return true;
5370 }
5371
5372 /*
5373  * If the sequence is incremented while we're waiting on a REQUEST_CLOSE reply,
5374  * then we need to retransmit that request.
5375  */
5376 void inc_session_sequence(struct ceph_mds_session *s)
5377 {
5378         struct ceph_client *cl = s->s_mdsc->fsc->client;
5379
5380         lockdep_assert_held(&s->s_mutex);
5381
5382         s->s_seq++;
5383
5384         if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
5385                 int ret;
5386
5387                 doutc(cl, "resending session close request for mds%d\n", s->s_mds);
5388                 ret = request_close_session(s);
5389                 if (ret < 0)
5390                         pr_err_client(cl, "unable to close session to mds%d: %d\n",
5391                                       s->s_mds, ret);
5392         }
5393 }
5394
5395 /*
5396  * delayed work -- periodically trim expired leases, renew caps with mds.  If
5397  * the @delay parameter is set to 0 or if it's more than 5 secs, the default
5398  * workqueue delay value of 5 secs will be used.
5399  */
5400 static void schedule_delayed(struct ceph_mds_client *mdsc, unsigned long delay)
5401 {
5402         unsigned long max_delay = HZ * 5;
5403
5404         /* 5 secs default delay */
5405         if (!delay || (delay > max_delay))
5406                 delay = max_delay;
5407         schedule_delayed_work(&mdsc->delayed_work,
5408                               round_jiffies_relative(delay));
5409 }
5410
5411 static void delayed_work(struct work_struct *work)
5412 {
5413         struct ceph_mds_client *mdsc =
5414                 container_of(work, struct ceph_mds_client, delayed_work.work);
5415         unsigned long delay;
5416         int renew_interval;
5417         int renew_caps;
5418         int i;
5419
5420         doutc(mdsc->fsc->client, "mdsc delayed_work\n");
5421
5422         if (mdsc->stopping >= CEPH_MDSC_STOPPING_FLUSHED)
5423                 return;
5424
5425         mutex_lock(&mdsc->mutex);
5426         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
5427         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
5428                                    mdsc->last_renew_caps);
5429         if (renew_caps)
5430                 mdsc->last_renew_caps = jiffies;
5431
5432         for (i = 0; i < mdsc->max_sessions; i++) {
5433                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
5434                 if (!s)
5435                         continue;
5436
5437                 if (!check_session_state(s)) {
5438                         ceph_put_mds_session(s);
5439                         continue;
5440                 }
5441                 mutex_unlock(&mdsc->mutex);
5442
5443                 ceph_flush_session_cap_releases(mdsc, s);
5444
5445                 mutex_lock(&s->s_mutex);
5446                 if (renew_caps)
5447                         send_renew_caps(mdsc, s);
5448                 else
5449                         ceph_con_keepalive(&s->s_con);
5450                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
5451                     s->s_state == CEPH_MDS_SESSION_HUNG)
5452                         ceph_send_cap_releases(mdsc, s);
5453                 mutex_unlock(&s->s_mutex);
5454                 ceph_put_mds_session(s);
5455
5456                 mutex_lock(&mdsc->mutex);
5457         }
5458         mutex_unlock(&mdsc->mutex);
5459
5460         delay = ceph_check_delayed_caps(mdsc);
5461
5462         ceph_queue_cap_reclaim_work(mdsc);
5463
5464         ceph_trim_snapid_map(mdsc);
5465
5466         maybe_recover_session(mdsc);
5467
5468         schedule_delayed(mdsc, delay);
5469 }
5470
5471 int ceph_mdsc_init(struct ceph_fs_client *fsc)
5472
5473 {
5474         struct ceph_mds_client *mdsc;
5475         int err;
5476
5477         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
5478         if (!mdsc)
5479                 return -ENOMEM;
5480         mdsc->fsc = fsc;
5481         mutex_init(&mdsc->mutex);
5482         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
5483         if (!mdsc->mdsmap) {
5484                 err = -ENOMEM;
5485                 goto err_mdsc;
5486         }
5487
5488         init_completion(&mdsc->safe_umount_waiters);
5489         spin_lock_init(&mdsc->stopping_lock);
5490         atomic_set(&mdsc->stopping_blockers, 0);
5491         init_completion(&mdsc->stopping_waiter);
5492         atomic64_set(&mdsc->dirty_folios, 0);
5493         init_waitqueue_head(&mdsc->flush_end_wq);
5494         init_waitqueue_head(&mdsc->session_close_wq);
5495         INIT_LIST_HEAD(&mdsc->waiting_for_map);
5496         mdsc->quotarealms_inodes = RB_ROOT;
5497         mutex_init(&mdsc->quotarealms_inodes_mutex);
5498         init_rwsem(&mdsc->snap_rwsem);
5499         mdsc->snap_realms = RB_ROOT;
5500         INIT_LIST_HEAD(&mdsc->snap_empty);
5501         spin_lock_init(&mdsc->snap_empty_lock);
5502         mdsc->request_tree = RB_ROOT;
5503         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
5504         mdsc->last_renew_caps = jiffies;
5505         INIT_LIST_HEAD(&mdsc->cap_delay_list);
5506 #ifdef CONFIG_DEBUG_FS
5507         INIT_LIST_HEAD(&mdsc->cap_wait_list);
5508 #endif
5509         spin_lock_init(&mdsc->cap_delay_lock);
5510         INIT_LIST_HEAD(&mdsc->cap_unlink_delay_list);
5511         INIT_LIST_HEAD(&mdsc->snap_flush_list);
5512         spin_lock_init(&mdsc->snap_flush_lock);
5513         mdsc->last_cap_flush_tid = 1;
5514         INIT_LIST_HEAD(&mdsc->cap_flush_list);
5515         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
5516         spin_lock_init(&mdsc->cap_dirty_lock);
5517         init_waitqueue_head(&mdsc->cap_flushing_wq);
5518         INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
5519         INIT_WORK(&mdsc->cap_unlink_work, ceph_cap_unlink_work);
5520         err = ceph_metric_init(&mdsc->metric);
5521         if (err)
5522                 goto err_mdsmap;
5523
5524         spin_lock_init(&mdsc->dentry_list_lock);
5525         INIT_LIST_HEAD(&mdsc->dentry_leases);
5526         INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
5527
5528         ceph_caps_init(mdsc);
5529         ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
5530
5531         spin_lock_init(&mdsc->snapid_map_lock);
5532         mdsc->snapid_map_tree = RB_ROOT;
5533         INIT_LIST_HEAD(&mdsc->snapid_map_lru);
5534
5535         init_rwsem(&mdsc->pool_perm_rwsem);
5536         mdsc->pool_perm_tree = RB_ROOT;
5537
5538         strscpy(mdsc->nodename, utsname()->nodename,
5539                 sizeof(mdsc->nodename));
5540
5541         fsc->mdsc = mdsc;
5542         return 0;
5543
5544 err_mdsmap:
5545         kfree(mdsc->mdsmap);
5546 err_mdsc:
5547         kfree(mdsc);
5548         return err;
5549 }
5550
5551 /*
5552  * Wait for safe replies on open mds requests.  If we time out, drop
5553  * all requests from the tree to avoid dangling dentry refs.
5554  */
5555 static void wait_requests(struct ceph_mds_client *mdsc)
5556 {
5557         struct ceph_client *cl = mdsc->fsc->client;
5558         struct ceph_options *opts = mdsc->fsc->client->options;
5559         struct ceph_mds_request *req;
5560
5561         mutex_lock(&mdsc->mutex);
5562         if (__get_oldest_req(mdsc)) {
5563                 mutex_unlock(&mdsc->mutex);
5564
5565                 doutc(cl, "waiting for requests\n");
5566                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
5567                                     ceph_timeout_jiffies(opts->mount_timeout));
5568
5569                 /* tear down remaining requests */
5570                 mutex_lock(&mdsc->mutex);
5571                 while ((req = __get_oldest_req(mdsc))) {
5572                         doutc(cl, "timed out on tid %llu\n", req->r_tid);
5573                         list_del_init(&req->r_wait);
5574                         __unregister_request(mdsc, req);
5575                 }
5576         }
5577         mutex_unlock(&mdsc->mutex);
5578         doutc(cl, "done\n");
5579 }
5580
5581 void send_flush_mdlog(struct ceph_mds_session *s)
5582 {
5583         struct ceph_client *cl = s->s_mdsc->fsc->client;
5584         struct ceph_msg *msg;
5585
5586         /*
5587          * Pre-luminous MDS crashes when it sees an unknown session request
5588          */
5589         if (!CEPH_HAVE_FEATURE(s->s_con.peer_features, SERVER_LUMINOUS))
5590                 return;
5591
5592         mutex_lock(&s->s_mutex);
5593         doutc(cl, "request mdlog flush to mds%d (%s)s seq %lld\n",
5594               s->s_mds, ceph_session_state_name(s->s_state), s->s_seq);
5595         msg = ceph_create_session_msg(CEPH_SESSION_REQUEST_FLUSH_MDLOG,
5596                                       s->s_seq);
5597         if (!msg) {
5598                 pr_err_client(cl, "failed to request mdlog flush to mds%d (%s) seq %lld\n",
5599                               s->s_mds, ceph_session_state_name(s->s_state), s->s_seq);
5600         } else {
5601                 ceph_con_send(&s->s_con, msg);
5602         }
5603         mutex_unlock(&s->s_mutex);
5604 }
5605
5606 static int ceph_mds_auth_match(struct ceph_mds_client *mdsc,
5607                                struct ceph_mds_cap_auth *auth,
5608                                const struct cred *cred,
5609                                char *tpath)
5610 {
5611         u32 caller_uid = from_kuid(&init_user_ns, cred->fsuid);
5612         u32 caller_gid = from_kgid(&init_user_ns, cred->fsgid);
5613         struct ceph_client *cl = mdsc->fsc->client;
5614         const char *spath = mdsc->fsc->mount_options->server_path;
5615         bool gid_matched = false;
5616         u32 gid, tlen, len;
5617         int i, j;
5618
5619         doutc(cl, "match.uid %lld\n", auth->match.uid);
5620         if (auth->match.uid != MDS_AUTH_UID_ANY) {
5621                 if (auth->match.uid != caller_uid)
5622                         return 0;
5623                 if (auth->match.num_gids) {
5624                         for (i = 0; i < auth->match.num_gids; i++) {
5625                                 if (caller_gid == auth->match.gids[i])
5626                                         gid_matched = true;
5627                         }
5628                         if (!gid_matched && cred->group_info->ngroups) {
5629                                 for (i = 0; i < cred->group_info->ngroups; i++) {
5630                                         gid = from_kgid(&init_user_ns,
5631                                                         cred->group_info->gid[i]);
5632                                         for (j = 0; j < auth->match.num_gids; j++) {
5633                                                 if (gid == auth->match.gids[j]) {
5634                                                         gid_matched = true;
5635                                                         break;
5636                                                 }
5637                                         }
5638                                         if (gid_matched)
5639                                                 break;
5640                                 }
5641                         }
5642                         if (!gid_matched)
5643                                 return 0;
5644                 }
5645         }
5646
5647         /* path match */
5648         if (auth->match.path) {
5649                 if (!tpath)
5650                         return 0;
5651
5652                 tlen = strlen(tpath);
5653                 len = strlen(auth->match.path);
5654                 if (len) {
5655                         char *_tpath = tpath;
5656                         bool free_tpath = false;
5657                         int m, n;
5658
5659                         doutc(cl, "server path %s, tpath %s, match.path %s\n",
5660                               spath, tpath, auth->match.path);
5661                         if (spath && (m = strlen(spath)) != 1) {
5662                                 /* mount path + '/' + tpath + an extra space */
5663                                 n = m + 1 + tlen + 1;
5664                                 _tpath = kmalloc(n, GFP_NOFS);
5665                                 if (!_tpath)
5666                                         return -ENOMEM;
5667                                 /* remove the leading '/' */
5668                                 snprintf(_tpath, n, "%s/%s", spath + 1, tpath);
5669                                 free_tpath = true;
5670                                 tlen = strlen(_tpath);
5671                         }
5672
5673                         /*
5674                          * Please note the tailing '/' for match.path has already
5675                          * been removed when parsing.
5676                          *
5677                          * Remove the tailing '/' for the target path.
5678                          */
5679                         while (tlen && _tpath[tlen - 1] == '/') {
5680                                 _tpath[tlen - 1] = '\0';
5681                                 tlen -= 1;
5682                         }
5683                         doutc(cl, "_tpath %s\n", _tpath);
5684
5685                         /*
5686                          * In case first == _tpath && tlen == len:
5687                          *  match.path=/foo  --> /foo _path=/foo     --> match
5688                          *  match.path=/foo/ --> /foo _path=/foo     --> match
5689                          *
5690                          * In case first == _tmatch.path && tlen > len:
5691                          *  match.path=/foo/ --> /foo _path=/foo/    --> match
5692                          *  match.path=/foo  --> /foo _path=/foo/    --> match
5693                          *  match.path=/foo/ --> /foo _path=/foo/d   --> match
5694                          *  match.path=/foo  --> /foo _path=/food    --> mismatch
5695                          *
5696                          * All the other cases                       --> mismatch
5697                          */
5698                         bool path_matched = true;
5699                         char *first = strstr(_tpath, auth->match.path);
5700                         if (first != _tpath ||
5701                             (tlen > len && _tpath[len] != '/')) {
5702                                 path_matched = false;
5703                         }
5704
5705                         if (free_tpath)
5706                                 kfree(_tpath);
5707
5708                         if (!path_matched)
5709                                 return 0;
5710                 }
5711         }
5712
5713         doutc(cl, "matched\n");
5714         return 1;
5715 }
5716
5717 int ceph_mds_check_access(struct ceph_mds_client *mdsc, char *tpath, int mask)
5718 {
5719         const struct cred *cred = get_current_cred();
5720         u32 caller_uid = from_kuid(&init_user_ns, cred->fsuid);
5721         u32 caller_gid = from_kgid(&init_user_ns, cred->fsgid);
5722         struct ceph_mds_cap_auth *rw_perms_s = NULL;
5723         struct ceph_client *cl = mdsc->fsc->client;
5724         bool root_squash_perms = true;
5725         int i, err;
5726
5727         doutc(cl, "tpath '%s', mask %d, caller_uid %d, caller_gid %d\n",
5728               tpath, mask, caller_uid, caller_gid);
5729
5730         for (i = 0; i < mdsc->s_cap_auths_num; i++) {
5731                 struct ceph_mds_cap_auth *s = &mdsc->s_cap_auths[i];
5732
5733                 err = ceph_mds_auth_match(mdsc, s, cred, tpath);
5734                 if (err < 0) {
5735                         put_cred(cred);
5736                         return err;
5737                 } else if (err > 0) {
5738                         /* always follow the last auth caps' permission */
5739                         root_squash_perms = true;
5740                         rw_perms_s = NULL;
5741                         if ((mask & MAY_WRITE) && s->writeable &&
5742                             s->match.root_squash && (!caller_uid || !caller_gid))
5743                                 root_squash_perms = false;
5744
5745                         if (((mask & MAY_WRITE) && !s->writeable) ||
5746                             ((mask & MAY_READ) && !s->readable))
5747                                 rw_perms_s = s;
5748                 }
5749         }
5750
5751         put_cred(cred);
5752
5753         doutc(cl, "root_squash_perms %d, rw_perms_s %p\n", root_squash_perms,
5754               rw_perms_s);
5755         if (root_squash_perms && rw_perms_s == NULL) {
5756                 doutc(cl, "access allowed\n");
5757                 return 0;
5758         }
5759
5760         if (!root_squash_perms) {
5761                 doutc(cl, "root_squash is enabled and user(%d %d) isn't allowed to write",
5762                       caller_uid, caller_gid);
5763         }
5764         if (rw_perms_s) {
5765                 doutc(cl, "mds auth caps readable/writeable %d/%d while request r/w %d/%d",
5766                       rw_perms_s->readable, rw_perms_s->writeable,
5767                       !!(mask & MAY_READ), !!(mask & MAY_WRITE));
5768         }
5769         doutc(cl, "access denied\n");
5770         return -EACCES;
5771 }
5772
5773 /*
5774  * called before mount is ro, and before dentries are torn down.
5775  * (hmm, does this still race with new lookups?)
5776  */
5777 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
5778 {
5779         doutc(mdsc->fsc->client, "begin\n");
5780         mdsc->stopping = CEPH_MDSC_STOPPING_BEGIN;
5781
5782         ceph_mdsc_iterate_sessions(mdsc, send_flush_mdlog, true);
5783         ceph_mdsc_iterate_sessions(mdsc, lock_unlock_session, false);
5784         ceph_flush_dirty_caps(mdsc);
5785         wait_requests(mdsc);
5786
5787         /*
5788          * wait for reply handlers to drop their request refs and
5789          * their inode/dcache refs
5790          */
5791         ceph_msgr_flush();
5792
5793         ceph_cleanup_quotarealms_inodes(mdsc);
5794         doutc(mdsc->fsc->client, "done\n");
5795 }
5796
5797 /*
5798  * flush the mdlog and wait for all write mds requests to flush.
5799  */
5800 static void flush_mdlog_and_wait_mdsc_unsafe_requests(struct ceph_mds_client *mdsc,
5801                                                  u64 want_tid)
5802 {
5803         struct ceph_client *cl = mdsc->fsc->client;
5804         struct ceph_mds_request *req = NULL, *nextreq;
5805         struct ceph_mds_session *last_session = NULL;
5806         struct rb_node *n;
5807
5808         mutex_lock(&mdsc->mutex);
5809         doutc(cl, "want %lld\n", want_tid);
5810 restart:
5811         req = __get_oldest_req(mdsc);
5812         while (req && req->r_tid <= want_tid) {
5813                 /* find next request */
5814                 n = rb_next(&req->r_node);
5815                 if (n)
5816                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
5817                 else
5818                         nextreq = NULL;
5819                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
5820                     (req->r_op & CEPH_MDS_OP_WRITE)) {
5821                         struct ceph_mds_session *s = req->r_session;
5822
5823                         if (!s) {
5824                                 req = nextreq;
5825                                 continue;
5826                         }
5827
5828                         /* write op */
5829                         ceph_mdsc_get_request(req);
5830                         if (nextreq)
5831                                 ceph_mdsc_get_request(nextreq);
5832                         s = ceph_get_mds_session(s);
5833                         mutex_unlock(&mdsc->mutex);
5834
5835                         /* send flush mdlog request to MDS */
5836                         if (last_session != s) {
5837                                 send_flush_mdlog(s);
5838                                 ceph_put_mds_session(last_session);
5839                                 last_session = s;
5840                         } else {
5841                                 ceph_put_mds_session(s);
5842                         }
5843                         doutc(cl, "wait on %llu (want %llu)\n",
5844                               req->r_tid, want_tid);
5845                         wait_for_completion(&req->r_safe_completion);
5846
5847                         mutex_lock(&mdsc->mutex);
5848                         ceph_mdsc_put_request(req);
5849                         if (!nextreq)
5850                                 break;  /* next dne before, so we're done! */
5851                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
5852                                 /* next request was removed from tree */
5853                                 ceph_mdsc_put_request(nextreq);
5854                                 goto restart;
5855                         }
5856                         ceph_mdsc_put_request(nextreq);  /* won't go away */
5857                 }
5858                 req = nextreq;
5859         }
5860         mutex_unlock(&mdsc->mutex);
5861         ceph_put_mds_session(last_session);
5862         doutc(cl, "done\n");
5863 }
5864
5865 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
5866 {
5867         struct ceph_client *cl = mdsc->fsc->client;
5868         u64 want_tid, want_flush;
5869
5870         if (READ_ONCE(mdsc->fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN)
5871                 return;
5872
5873         doutc(cl, "sync\n");
5874         mutex_lock(&mdsc->mutex);
5875         want_tid = mdsc->last_tid;
5876         mutex_unlock(&mdsc->mutex);
5877
5878         ceph_flush_dirty_caps(mdsc);
5879         ceph_flush_cap_releases(mdsc);
5880         spin_lock(&mdsc->cap_dirty_lock);
5881         want_flush = mdsc->last_cap_flush_tid;
5882         if (!list_empty(&mdsc->cap_flush_list)) {
5883                 struct ceph_cap_flush *cf =
5884                         list_last_entry(&mdsc->cap_flush_list,
5885                                         struct ceph_cap_flush, g_list);
5886                 cf->wake = true;
5887         }
5888         spin_unlock(&mdsc->cap_dirty_lock);
5889
5890         doutc(cl, "sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
5891
5892         flush_mdlog_and_wait_mdsc_unsafe_requests(mdsc, want_tid);
5893         wait_caps_flush(mdsc, want_flush);
5894 }
5895
5896 /*
5897  * true if all sessions are closed, or we force unmount
5898  */
5899 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
5900 {
5901         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
5902                 return true;
5903         return atomic_read(&mdsc->num_sessions) <= skipped;
5904 }
5905
5906 /*
5907  * called after sb is ro or when metadata corrupted.
5908  */
5909 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
5910 {
5911         struct ceph_options *opts = mdsc->fsc->client->options;
5912         struct ceph_client *cl = mdsc->fsc->client;
5913         struct ceph_mds_session *session;
5914         int i;
5915         int skipped = 0;
5916
5917         doutc(cl, "begin\n");
5918
5919         /* close sessions */
5920         mutex_lock(&mdsc->mutex);
5921         for (i = 0; i < mdsc->max_sessions; i++) {
5922                 session = __ceph_lookup_mds_session(mdsc, i);
5923                 if (!session)
5924                         continue;
5925                 mutex_unlock(&mdsc->mutex);
5926                 mutex_lock(&session->s_mutex);
5927                 if (__close_session(mdsc, session) <= 0)
5928                         skipped++;
5929                 mutex_unlock(&session->s_mutex);
5930                 ceph_put_mds_session(session);
5931                 mutex_lock(&mdsc->mutex);
5932         }
5933         mutex_unlock(&mdsc->mutex);
5934
5935         doutc(cl, "waiting for sessions to close\n");
5936         wait_event_timeout(mdsc->session_close_wq,
5937                            done_closing_sessions(mdsc, skipped),
5938                            ceph_timeout_jiffies(opts->mount_timeout));
5939
5940         /* tear down remaining sessions */
5941         mutex_lock(&mdsc->mutex);
5942         for (i = 0; i < mdsc->max_sessions; i++) {
5943                 if (mdsc->sessions[i]) {
5944                         session = ceph_get_mds_session(mdsc->sessions[i]);
5945                         __unregister_session(mdsc, session);
5946                         mutex_unlock(&mdsc->mutex);
5947                         mutex_lock(&session->s_mutex);
5948                         remove_session_caps(session);
5949                         mutex_unlock(&session->s_mutex);
5950                         ceph_put_mds_session(session);
5951                         mutex_lock(&mdsc->mutex);
5952                 }
5953         }
5954         WARN_ON(!list_empty(&mdsc->cap_delay_list));
5955         mutex_unlock(&mdsc->mutex);
5956
5957         ceph_cleanup_snapid_map(mdsc);
5958         ceph_cleanup_global_and_empty_realms(mdsc);
5959
5960         cancel_work_sync(&mdsc->cap_reclaim_work);
5961         cancel_work_sync(&mdsc->cap_unlink_work);
5962         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
5963
5964         doutc(cl, "done\n");
5965 }
5966
5967 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
5968 {
5969         struct ceph_mds_session *session;
5970         int mds;
5971
5972         doutc(mdsc->fsc->client, "force umount\n");
5973
5974         mutex_lock(&mdsc->mutex);
5975         for (mds = 0; mds < mdsc->max_sessions; mds++) {
5976                 session = __ceph_lookup_mds_session(mdsc, mds);
5977                 if (!session)
5978                         continue;
5979
5980                 if (session->s_state == CEPH_MDS_SESSION_REJECTED)
5981                         __unregister_session(mdsc, session);
5982                 __wake_requests(mdsc, &session->s_waiting);
5983                 mutex_unlock(&mdsc->mutex);
5984
5985                 mutex_lock(&session->s_mutex);
5986                 __close_session(mdsc, session);
5987                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
5988                         cleanup_session_requests(mdsc, session);
5989                         remove_session_caps(session);
5990                 }
5991                 mutex_unlock(&session->s_mutex);
5992                 ceph_put_mds_session(session);
5993
5994                 mutex_lock(&mdsc->mutex);
5995                 kick_requests(mdsc, mds);
5996         }
5997         __wake_requests(mdsc, &mdsc->waiting_for_map);
5998         mutex_unlock(&mdsc->mutex);
5999 }
6000
6001 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
6002 {
6003         doutc(mdsc->fsc->client, "stop\n");
6004         /*
6005          * Make sure the delayed work stopped before releasing
6006          * the resources.
6007          *
6008          * Because the cancel_delayed_work_sync() will only
6009          * guarantee that the work finishes executing. But the
6010          * delayed work will re-arm itself again after that.
6011          */
6012         flush_delayed_work(&mdsc->delayed_work);
6013
6014         if (mdsc->mdsmap)
6015                 ceph_mdsmap_destroy(mdsc->mdsmap);
6016         kfree(mdsc->sessions);
6017         ceph_caps_finalize(mdsc);
6018
6019         if (mdsc->s_cap_auths) {
6020                 int i;
6021
6022                 for (i = 0; i < mdsc->s_cap_auths_num; i++) {
6023                         kfree(mdsc->s_cap_auths[i].match.gids);
6024                         kfree(mdsc->s_cap_auths[i].match.path);
6025                         kfree(mdsc->s_cap_auths[i].match.fs_name);
6026                 }
6027                 kfree(mdsc->s_cap_auths);
6028         }
6029
6030         ceph_pool_perm_destroy(mdsc);
6031 }
6032
6033 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
6034 {
6035         struct ceph_mds_client *mdsc = fsc->mdsc;
6036         doutc(fsc->client, "%p\n", mdsc);
6037
6038         if (!mdsc)
6039                 return;
6040
6041         /* flush out any connection work with references to us */
6042         ceph_msgr_flush();
6043
6044         ceph_mdsc_stop(mdsc);
6045
6046         ceph_metric_destroy(&mdsc->metric);
6047
6048         fsc->mdsc = NULL;
6049         kfree(mdsc);
6050         doutc(fsc->client, "%p done\n", mdsc);
6051 }
6052
6053 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
6054 {
6055         struct ceph_fs_client *fsc = mdsc->fsc;
6056         struct ceph_client *cl = fsc->client;
6057         const char *mds_namespace = fsc->mount_options->mds_namespace;
6058         void *p = msg->front.iov_base;
6059         void *end = p + msg->front.iov_len;
6060         u32 epoch;
6061         u32 num_fs;
6062         u32 mount_fscid = (u32)-1;
6063         int err = -EINVAL;
6064
6065         ceph_decode_need(&p, end, sizeof(u32), bad);
6066         epoch = ceph_decode_32(&p);
6067
6068         doutc(cl, "epoch %u\n", epoch);
6069
6070         /* struct_v, struct_cv, map_len, epoch, legacy_client_fscid */
6071         ceph_decode_skip_n(&p, end, 2 + sizeof(u32) * 3, bad);
6072
6073         ceph_decode_32_safe(&p, end, num_fs, bad);
6074         while (num_fs-- > 0) {
6075                 void *info_p, *info_end;
6076                 u32 info_len;
6077                 u32 fscid, namelen;
6078
6079                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
6080                 p += 2;         // info_v, info_cv
6081                 info_len = ceph_decode_32(&p);
6082                 ceph_decode_need(&p, end, info_len, bad);
6083                 info_p = p;
6084                 info_end = p + info_len;
6085                 p = info_end;
6086
6087                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
6088                 fscid = ceph_decode_32(&info_p);
6089                 namelen = ceph_decode_32(&info_p);
6090                 ceph_decode_need(&info_p, info_end, namelen, bad);
6091
6092                 if (mds_namespace &&
6093                     strlen(mds_namespace) == namelen &&
6094                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
6095                         mount_fscid = fscid;
6096                         break;
6097                 }
6098         }
6099
6100         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
6101         if (mount_fscid != (u32)-1) {
6102                 fsc->client->monc.fs_cluster_id = mount_fscid;
6103                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
6104                                    0, true);
6105                 ceph_monc_renew_subs(&fsc->client->monc);
6106         } else {
6107                 err = -ENOENT;
6108                 goto err_out;
6109         }
6110         return;
6111
6112 bad:
6113         pr_err_client(cl, "error decoding fsmap %d. Shutting down mount.\n",
6114                       err);
6115         ceph_umount_begin(mdsc->fsc->sb);
6116         ceph_msg_dump(msg);
6117 err_out:
6118         mutex_lock(&mdsc->mutex);
6119         mdsc->mdsmap_err = err;
6120         __wake_requests(mdsc, &mdsc->waiting_for_map);
6121         mutex_unlock(&mdsc->mutex);
6122 }
6123
6124 /*
6125  * handle mds map update.
6126  */
6127 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
6128 {
6129         struct ceph_client *cl = mdsc->fsc->client;
6130         u32 epoch;
6131         u32 maplen;
6132         void *p = msg->front.iov_base;
6133         void *end = p + msg->front.iov_len;
6134         struct ceph_mdsmap *newmap, *oldmap;
6135         struct ceph_fsid fsid;
6136         int err = -EINVAL;
6137
6138         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
6139         ceph_decode_copy(&p, &fsid, sizeof(fsid));
6140         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
6141                 return;
6142         epoch = ceph_decode_32(&p);
6143         maplen = ceph_decode_32(&p);
6144         doutc(cl, "epoch %u len %d\n", epoch, (int)maplen);
6145
6146         /* do we need it? */
6147         mutex_lock(&mdsc->mutex);
6148         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
6149                 doutc(cl, "epoch %u <= our %u\n", epoch, mdsc->mdsmap->m_epoch);
6150                 mutex_unlock(&mdsc->mutex);
6151                 return;
6152         }
6153
6154         newmap = ceph_mdsmap_decode(mdsc, &p, end, ceph_msgr2(mdsc->fsc->client));
6155         if (IS_ERR(newmap)) {
6156                 err = PTR_ERR(newmap);
6157                 goto bad_unlock;
6158         }
6159
6160         /* swap into place */
6161         if (mdsc->mdsmap) {
6162                 oldmap = mdsc->mdsmap;
6163                 mdsc->mdsmap = newmap;
6164                 check_new_map(mdsc, newmap, oldmap);
6165                 ceph_mdsmap_destroy(oldmap);
6166         } else {
6167                 mdsc->mdsmap = newmap;  /* first mds map */
6168         }
6169         mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
6170                                         MAX_LFS_FILESIZE);
6171
6172         __wake_requests(mdsc, &mdsc->waiting_for_map);
6173         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
6174                           mdsc->mdsmap->m_epoch);
6175
6176         mutex_unlock(&mdsc->mutex);
6177         schedule_delayed(mdsc, 0);
6178         return;
6179
6180 bad_unlock:
6181         mutex_unlock(&mdsc->mutex);
6182 bad:
6183         pr_err_client(cl, "error decoding mdsmap %d. Shutting down mount.\n",
6184                       err);
6185         ceph_umount_begin(mdsc->fsc->sb);
6186         ceph_msg_dump(msg);
6187         return;
6188 }
6189
6190 static struct ceph_connection *mds_get_con(struct ceph_connection *con)
6191 {
6192         struct ceph_mds_session *s = con->private;
6193
6194         if (ceph_get_mds_session(s))
6195                 return con;
6196         return NULL;
6197 }
6198
6199 static void mds_put_con(struct ceph_connection *con)
6200 {
6201         struct ceph_mds_session *s = con->private;
6202
6203         ceph_put_mds_session(s);
6204 }
6205
6206 /*
6207  * if the client is unresponsive for long enough, the mds will kill
6208  * the session entirely.
6209  */
6210 static void mds_peer_reset(struct ceph_connection *con)
6211 {
6212         struct ceph_mds_session *s = con->private;
6213         struct ceph_mds_client *mdsc = s->s_mdsc;
6214
6215         pr_warn_client(mdsc->fsc->client, "mds%d closed our session\n",
6216                        s->s_mds);
6217         if (READ_ONCE(mdsc->fsc->mount_state) != CEPH_MOUNT_FENCE_IO &&
6218             ceph_mdsmap_get_state(mdsc->mdsmap, s->s_mds) >= CEPH_MDS_STATE_RECONNECT)
6219                 send_mds_reconnect(mdsc, s);
6220 }
6221
6222 static void mds_dispatch(struct ceph_connection *con, struct ceph_msg *msg)
6223 {
6224         struct ceph_mds_session *s = con->private;
6225         struct ceph_mds_client *mdsc = s->s_mdsc;
6226         struct ceph_client *cl = mdsc->fsc->client;
6227         int type = le16_to_cpu(msg->hdr.type);
6228
6229         mutex_lock(&mdsc->mutex);
6230         if (__verify_registered_session(mdsc, s) < 0) {
6231                 mutex_unlock(&mdsc->mutex);
6232                 goto out;
6233         }
6234         mutex_unlock(&mdsc->mutex);
6235
6236         switch (type) {
6237         case CEPH_MSG_MDS_MAP:
6238                 ceph_mdsc_handle_mdsmap(mdsc, msg);
6239                 break;
6240         case CEPH_MSG_FS_MAP_USER:
6241                 ceph_mdsc_handle_fsmap(mdsc, msg);
6242                 break;
6243         case CEPH_MSG_CLIENT_SESSION:
6244                 handle_session(s, msg);
6245                 break;
6246         case CEPH_MSG_CLIENT_REPLY:
6247                 handle_reply(s, msg);
6248                 break;
6249         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
6250                 handle_forward(mdsc, s, msg);
6251                 break;
6252         case CEPH_MSG_CLIENT_CAPS:
6253                 ceph_handle_caps(s, msg);
6254                 break;
6255         case CEPH_MSG_CLIENT_SNAP:
6256                 ceph_handle_snap(mdsc, s, msg);
6257                 break;
6258         case CEPH_MSG_CLIENT_LEASE:
6259                 handle_lease(mdsc, s, msg);
6260                 break;
6261         case CEPH_MSG_CLIENT_QUOTA:
6262                 ceph_handle_quota(mdsc, s, msg);
6263                 break;
6264
6265         default:
6266                 pr_err_client(cl, "received unknown message type %d %s\n",
6267                               type, ceph_msg_type_name(type));
6268         }
6269 out:
6270         ceph_msg_put(msg);
6271 }
6272
6273 /*
6274  * authentication
6275  */
6276
6277 /*
6278  * Note: returned pointer is the address of a structure that's
6279  * managed separately.  Caller must *not* attempt to free it.
6280  */
6281 static struct ceph_auth_handshake *
6282 mds_get_authorizer(struct ceph_connection *con, int *proto, int force_new)
6283 {
6284         struct ceph_mds_session *s = con->private;
6285         struct ceph_mds_client *mdsc = s->s_mdsc;
6286         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
6287         struct ceph_auth_handshake *auth = &s->s_auth;
6288         int ret;
6289
6290         ret = __ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
6291                                          force_new, proto, NULL, NULL);
6292         if (ret)
6293                 return ERR_PTR(ret);
6294
6295         return auth;
6296 }
6297
6298 static int mds_add_authorizer_challenge(struct ceph_connection *con,
6299                                     void *challenge_buf, int challenge_buf_len)
6300 {
6301         struct ceph_mds_session *s = con->private;
6302         struct ceph_mds_client *mdsc = s->s_mdsc;
6303         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
6304
6305         return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
6306                                             challenge_buf, challenge_buf_len);
6307 }
6308
6309 static int mds_verify_authorizer_reply(struct ceph_connection *con)
6310 {
6311         struct ceph_mds_session *s = con->private;
6312         struct ceph_mds_client *mdsc = s->s_mdsc;
6313         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
6314         struct ceph_auth_handshake *auth = &s->s_auth;
6315
6316         return ceph_auth_verify_authorizer_reply(ac, auth->authorizer,
6317                 auth->authorizer_reply_buf, auth->authorizer_reply_buf_len,
6318                 NULL, NULL, NULL, NULL);
6319 }
6320
6321 static int mds_invalidate_authorizer(struct ceph_connection *con)
6322 {
6323         struct ceph_mds_session *s = con->private;
6324         struct ceph_mds_client *mdsc = s->s_mdsc;
6325         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
6326
6327         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
6328
6329         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
6330 }
6331
6332 static int mds_get_auth_request(struct ceph_connection *con,
6333                                 void *buf, int *buf_len,
6334                                 void **authorizer, int *authorizer_len)
6335 {
6336         struct ceph_mds_session *s = con->private;
6337         struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
6338         struct ceph_auth_handshake *auth = &s->s_auth;
6339         int ret;
6340
6341         ret = ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
6342                                        buf, buf_len);
6343         if (ret)
6344                 return ret;
6345
6346         *authorizer = auth->authorizer_buf;
6347         *authorizer_len = auth->authorizer_buf_len;
6348         return 0;
6349 }
6350
6351 static int mds_handle_auth_reply_more(struct ceph_connection *con,
6352                                       void *reply, int reply_len,
6353                                       void *buf, int *buf_len,
6354                                       void **authorizer, int *authorizer_len)
6355 {
6356         struct ceph_mds_session *s = con->private;
6357         struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
6358         struct ceph_auth_handshake *auth = &s->s_auth;
6359         int ret;
6360
6361         ret = ceph_auth_handle_svc_reply_more(ac, auth, reply, reply_len,
6362                                               buf, buf_len);
6363         if (ret)
6364                 return ret;
6365
6366         *authorizer = auth->authorizer_buf;
6367         *authorizer_len = auth->authorizer_buf_len;
6368         return 0;
6369 }
6370
6371 static int mds_handle_auth_done(struct ceph_connection *con,
6372                                 u64 global_id, void *reply, int reply_len,
6373                                 u8 *session_key, int *session_key_len,
6374                                 u8 *con_secret, int *con_secret_len)
6375 {
6376         struct ceph_mds_session *s = con->private;
6377         struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
6378         struct ceph_auth_handshake *auth = &s->s_auth;
6379
6380         return ceph_auth_handle_svc_reply_done(ac, auth, reply, reply_len,
6381                                                session_key, session_key_len,
6382                                                con_secret, con_secret_len);
6383 }
6384
6385 static int mds_handle_auth_bad_method(struct ceph_connection *con,
6386                                       int used_proto, int result,
6387                                       const int *allowed_protos, int proto_cnt,
6388                                       const int *allowed_modes, int mode_cnt)
6389 {
6390         struct ceph_mds_session *s = con->private;
6391         struct ceph_mon_client *monc = &s->s_mdsc->fsc->client->monc;
6392         int ret;
6393
6394         if (ceph_auth_handle_bad_authorizer(monc->auth, CEPH_ENTITY_TYPE_MDS,
6395                                             used_proto, result,
6396                                             allowed_protos, proto_cnt,
6397                                             allowed_modes, mode_cnt)) {
6398                 ret = ceph_monc_validate_auth(monc);
6399                 if (ret)
6400                         return ret;
6401         }
6402
6403         return -EACCES;
6404 }
6405
6406 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
6407                                 struct ceph_msg_header *hdr, int *skip)
6408 {
6409         struct ceph_msg *msg;
6410         int type = (int) le16_to_cpu(hdr->type);
6411         int front_len = (int) le32_to_cpu(hdr->front_len);
6412
6413         if (con->in_msg)
6414                 return con->in_msg;
6415
6416         *skip = 0;
6417         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
6418         if (!msg) {
6419                 pr_err("unable to allocate msg type %d len %d\n",
6420                        type, front_len);
6421                 return NULL;
6422         }
6423
6424         return msg;
6425 }
6426
6427 static int mds_sign_message(struct ceph_msg *msg)
6428 {
6429        struct ceph_mds_session *s = msg->con->private;
6430        struct ceph_auth_handshake *auth = &s->s_auth;
6431
6432        return ceph_auth_sign_message(auth, msg);
6433 }
6434
6435 static int mds_check_message_signature(struct ceph_msg *msg)
6436 {
6437        struct ceph_mds_session *s = msg->con->private;
6438        struct ceph_auth_handshake *auth = &s->s_auth;
6439
6440        return ceph_auth_check_message_signature(auth, msg);
6441 }
6442
6443 static const struct ceph_connection_operations mds_con_ops = {
6444         .get = mds_get_con,
6445         .put = mds_put_con,
6446         .alloc_msg = mds_alloc_msg,
6447         .dispatch = mds_dispatch,
6448         .peer_reset = mds_peer_reset,
6449         .get_authorizer = mds_get_authorizer,
6450         .add_authorizer_challenge = mds_add_authorizer_challenge,
6451         .verify_authorizer_reply = mds_verify_authorizer_reply,
6452         .invalidate_authorizer = mds_invalidate_authorizer,
6453         .sign_message = mds_sign_message,
6454         .check_message_signature = mds_check_message_signature,
6455         .get_auth_request = mds_get_auth_request,
6456         .handle_auth_reply_more = mds_handle_auth_reply_more,
6457         .handle_auth_done = mds_handle_auth_done,
6458         .handle_auth_bad_method = mds_handle_auth_bad_method,
6459 };
6460
6461 /* eof */