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
  13 #include "super.h"
  14 #include "mds_client.h"
  15
  16 #include <linux/ceph/ceph_features.h>
  17 #include <linux/ceph/messenger.h>
  18 #include <linux/ceph/decode.h>
  19 #include <linux/ceph/pagelist.h>
  20 #include <linux/ceph/auth.h>
  21 #include <linux/ceph/debugfs.h>
  22
  23 /*
  24  * A cluster of MDS (metadata server) daemons is responsible for
  25  * managing the file system namespace (the directory hierarchy and
  26  * inodes) and for coordinating shared access to storage.  Metadata is
  27  * partitioning hierarchically across a number of servers, and that
  28  * partition varies over time as the cluster adjusts the distribution
  29  * in order to balance load.
  30  *
  31  * The MDS client is primarily responsible to managing synchronous
  32  * metadata requests for operations like open, unlink, and so forth.
  33  * If there is a MDS failure, we find out about it when we (possibly
  34  * request and) receive a new MDS map, and can resubmit affected
  35  * requests.
  36  *
  37  * For the most part, though, we take advantage of a lossless
  38  * communications channel to the MDS, and do not need to worry about
  39  * timing out or resubmitting requests.
  40  *
  41  * We maintain a stateful "session" with each MDS we interact with.
  42  * Within each session, we sent periodic heartbeat messages to ensure
  43  * any capabilities or leases we have been issues remain valid.  If
  44  * the session times out and goes stale, our leases and capabilities
  45  * are no longer valid.
  46  */
  47
  48 struct ceph_reconnect_state {
  49         int nr_caps;
  50         struct ceph_pagelist *pagelist;
  51         unsigned msg_version;
  52 };
  53
  54 static void __wake_requests(struct ceph_mds_client *mdsc,
  55                             struct list_head *head);
  56
  57 static const struct ceph_connection_operations mds_con_ops;
  58
  59
  60 /*
  61  * mds reply parsing
  62  */
  63
  64 /*
  65  * parse individual inode info
  66  */
  67 static int parse_reply_info_in(void **p, void *end,
  68                                struct ceph_mds_reply_info_in *info,
  69                                u64 features)
  70 {
  71         int err = -EIO;
  72
  73         info->in = *p;
  74         *p += sizeof(struct ceph_mds_reply_inode) +
  75                 sizeof(*info->in->fragtree.splits) *
  76                 le32_to_cpu(info->in->fragtree.nsplits);
  77
  78         ceph_decode_32_safe(p, end, info->symlink_len, bad);
  79         ceph_decode_need(p, end, info->symlink_len, bad);
  80         info->symlink = *p;
  81         *p += info->symlink_len;
  82
  83         if (features & CEPH_FEATURE_DIRLAYOUTHASH)
  84                 ceph_decode_copy_safe(p, end, &info->dir_layout,
  85                                       sizeof(info->dir_layout), bad);
  86         else
  87                 memset(&info->dir_layout, 0, sizeof(info->dir_layout));
  88
  89         ceph_decode_32_safe(p, end, info->xattr_len, bad);
  90         ceph_decode_need(p, end, info->xattr_len, bad);
  91         info->xattr_data = *p;
  92         *p += info->xattr_len;
  93
  94         if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
  95                 ceph_decode_64_safe(p, end, info->inline_version, bad);
  96                 ceph_decode_32_safe(p, end, info->inline_len, bad);
  97                 ceph_decode_need(p, end, info->inline_len, bad);
  98                 info->inline_data = *p;
  99                 *p += info->inline_len;
 100         } else
 101                 info->inline_version = CEPH_INLINE_NONE;
 102
 103         info->pool_ns_len = 0;
 104         info->pool_ns_data = NULL;
 105         if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
 106                 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
 107                 if (info->pool_ns_len > 0) {
 108                         ceph_decode_need(p, end, info->pool_ns_len, bad);
 109                         info->pool_ns_data = *p;
 110                         *p += info->pool_ns_len;
 111                 }
 112         }
 113
 114         return 0;
 115 bad:
 116         return err;
 117 }
 118
 119 /*
 120  * parse a normal reply, which may contain a (dir+)dentry and/or a
 121  * target inode.
 122  */
 123 static int parse_reply_info_trace(void **p, void *end,
 124                                   struct ceph_mds_reply_info_parsed *info,
 125                                   u64 features)
 126 {
 127         int err;
 128
 129         if (info->head->is_dentry) {
 130                 err = parse_reply_info_in(p, end, &info->diri, features);
 131                 if (err < 0)
 132                         goto out_bad;
 133
 134                 if (unlikely(*p + sizeof(*info->dirfrag) > end))
 135                         goto bad;
 136                 info->dirfrag = *p;
 137                 *p += sizeof(*info->dirfrag) +
 138                         sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
 139                 if (unlikely(*p > end))
 140                         goto bad;
 141
 142                 ceph_decode_32_safe(p, end, info->dname_len, bad);
 143                 ceph_decode_need(p, end, info->dname_len, bad);
 144                 info->dname = *p;
 145                 *p += info->dname_len;
 146                 info->dlease = *p;
 147                 *p += sizeof(*info->dlease);
 148         }
 149
 150         if (info->head->is_target) {
 151                 err = parse_reply_info_in(p, end, &info->targeti, features);
 152                 if (err < 0)
 153                         goto out_bad;
 154         }
 155
 156         if (unlikely(*p != end))
 157                 goto bad;
 158         return 0;
 159
 160 bad:
 161         err = -EIO;
 162 out_bad:
 163         pr_err("problem parsing mds trace %d\n", err);
 164         return err;
 165 }
 166
 167 /*
 168  * parse readdir results
 169  */
 170 static int parse_reply_info_dir(void **p, void *end,
 171                                 struct ceph_mds_reply_info_parsed *info,
 172                                 u64 features)
 173 {
 174         u32 num, i = 0;
 175         int err;
 176
 177         info->dir_dir = *p;
 178         if (*p + sizeof(*info->dir_dir) > end)
 179                 goto bad;
 180         *p += sizeof(*info->dir_dir) +
 181                 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
 182         if (*p > end)
 183                 goto bad;
 184
 185         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 186         num = ceph_decode_32(p);
 187         {
 188                 u16 flags = ceph_decode_16(p);
 189                 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
 190                 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
 191                 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
 192                 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
 193         }
 194         if (num == 0)
 195                 goto done;
 196
 197         BUG_ON(!info->dir_entries);
 198         if ((unsigned long)(info->dir_entries + num) >
 199             (unsigned long)info->dir_entries + info->dir_buf_size) {
 200                 pr_err("dir contents are larger than expected\n");
 201                 WARN_ON(1);
 202                 goto bad;
 203         }
 204
 205         info->dir_nr = num;
 206         while (num) {
 207                 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
 208                 /* dentry */
 209                 ceph_decode_need(p, end, sizeof(u32)*2, bad);
 210                 rde->name_len = ceph_decode_32(p);
 211                 ceph_decode_need(p, end, rde->name_len, bad);
 212                 rde->name = *p;
 213                 *p += rde->name_len;
 214                 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
 215                 rde->lease = *p;
 216                 *p += sizeof(struct ceph_mds_reply_lease);
 217
 218                 /* inode */
 219                 err = parse_reply_info_in(p, end, &rde->inode, features);
 220                 if (err < 0)
 221                         goto out_bad;
 222                 /* ceph_readdir_prepopulate() will update it */
 223                 rde->offset = 0;
 224                 i++;
 225                 num--;
 226         }
 227
 228 done:
 229         if (*p != end)
 230                 goto bad;
 231         return 0;
 232
 233 bad:
 234         err = -EIO;
 235 out_bad:
 236         pr_err("problem parsing dir contents %d\n", err);
 237         return err;
 238 }
 239
 240 /*
 241  * parse fcntl F_GETLK results
 242  */
 243 static int parse_reply_info_filelock(void **p, void *end,
 244                                      struct ceph_mds_reply_info_parsed *info,
 245                                      u64 features)
 246 {
 247         if (*p + sizeof(*info->filelock_reply) > end)
 248                 goto bad;
 249
 250         info->filelock_reply = *p;
 251         *p += sizeof(*info->filelock_reply);
 252
 253         if (unlikely(*p != end))
 254                 goto bad;
 255         return 0;
 256
 257 bad:
 258         return -EIO;
 259 }
 260
 261 /*
 262  * parse create results
 263  */
 264 static int parse_reply_info_create(void **p, void *end,
 265                                   struct ceph_mds_reply_info_parsed *info,
 266                                   u64 features)
 267 {
 268         if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
 269                 if (*p == end) {
 270                         info->has_create_ino = false;
 271                 } else {
 272                         info->has_create_ino = true;
 273                         info->ino = ceph_decode_64(p);
 274                 }
 275         }
 276
 277         if (unlikely(*p != end))
 278                 goto bad;
 279         return 0;
 280
 281 bad:
 282         return -EIO;
 283 }
 284
 285 /*
 286  * parse extra results
 287  */
 288 static int parse_reply_info_extra(void **p, void *end,
 289                                   struct ceph_mds_reply_info_parsed *info,
 290                                   u64 features)
 291 {
 292         u32 op = le32_to_cpu(info->head->op);
 293
 294         if (op == CEPH_MDS_OP_GETFILELOCK)
 295                 return parse_reply_info_filelock(p, end, info, features);
 296         else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
 297                 return parse_reply_info_dir(p, end, info, features);
 298         else if (op == CEPH_MDS_OP_CREATE)
 299                 return parse_reply_info_create(p, end, info, features);
 300         else
 301                 return -EIO;
 302 }
 303
 304 /*
 305  * parse entire mds reply
 306  */
 307 static int parse_reply_info(struct ceph_msg *msg,
 308                             struct ceph_mds_reply_info_parsed *info,
 309                             u64 features)
 310 {
 311         void *p, *end;
 312         u32 len;
 313         int err;
 314
 315         info->head = msg->front.iov_base;
 316         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 317         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 318
 319         /* trace */
 320         ceph_decode_32_safe(&p, end, len, bad);
 321         if (len > 0) {
 322                 ceph_decode_need(&p, end, len, bad);
 323                 err = parse_reply_info_trace(&p, p+len, info, features);
 324                 if (err < 0)
 325                         goto out_bad;
 326         }
 327
 328         /* extra */
 329         ceph_decode_32_safe(&p, end, len, bad);
 330         if (len > 0) {
 331                 ceph_decode_need(&p, end, len, bad);
 332                 err = parse_reply_info_extra(&p, p+len, info, features);
 333                 if (err < 0)
 334                         goto out_bad;
 335         }
 336
 337         /* snap blob */
 338         ceph_decode_32_safe(&p, end, len, bad);
 339         info->snapblob_len = len;
 340         info->snapblob = p;
 341         p += len;
 342
 343         if (p != end)
 344                 goto bad;
 345         return 0;
 346
 347 bad:
 348         err = -EIO;
 349 out_bad:
 350         pr_err("mds parse_reply err %d\n", err);
 351         return err;
 352 }
 353
 354 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 355 {
 356         if (!info->dir_entries)
 357                 return;
 358         free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
 359 }
 360
 361
 362 /*
 363  * sessions
 364  */
 365 const char *ceph_session_state_name(int s)
 366 {
 367         switch (s) {
 368         case CEPH_MDS_SESSION_NEW: return "new";
 369         case CEPH_MDS_SESSION_OPENING: return "opening";
 370         case CEPH_MDS_SESSION_OPEN: return "open";
 371         case CEPH_MDS_SESSION_HUNG: return "hung";
 372         case CEPH_MDS_SESSION_CLOSING: return "closing";
 373         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 374         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 375         case CEPH_MDS_SESSION_REJECTED: return "rejected";
 376         default: return "???";
 377         }
 378 }
 379
 380 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
 381 {
 382         if (refcount_inc_not_zero(&s->s_ref)) {
 383                 dout("mdsc get_session %p %d -> %d\n", s,
 384                      refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
 385                 return s;
 386         } else {
 387                 dout("mdsc get_session %p 0 -- FAIL", s);
 388                 return NULL;
 389         }
 390 }
 391
 392 void ceph_put_mds_session(struct ceph_mds_session *s)
 393 {
 394         dout("mdsc put_session %p %d -> %d\n", s,
 395              refcount_read(&s->s_ref), refcount_read(&s->s_ref)-1);
 396         if (refcount_dec_and_test(&s->s_ref)) {
 397                 if (s->s_auth.authorizer)
 398                         ceph_auth_destroy_authorizer(s->s_auth.authorizer);
 399                 kfree(s);
 400         }
 401 }
 402
 403 /*
 404  * called under mdsc->mutex
 405  */
 406 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 407                                                    int mds)
 408 {
 409         struct ceph_mds_session *session;
 410
 411         if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
 412                 return NULL;
 413         session = mdsc->sessions[mds];
 414         dout("lookup_mds_session %p %d\n", session,
 415              refcount_read(&session->s_ref));
 416         get_session(session);
 417         return session;
 418 }
 419
 420 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 421 {
 422         if (mds >= mdsc->max_sessions)
 423                 return false;
 424         return mdsc->sessions[mds];
 425 }
 426
 427 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 428                                        struct ceph_mds_session *s)
 429 {
 430         if (s->s_mds >= mdsc->max_sessions ||
 431             mdsc->sessions[s->s_mds] != s)
 432                 return -ENOENT;
 433         return 0;
 434 }
 435
 436 /*
 437  * create+register a new session for given mds.
 438  * called under mdsc->mutex.
 439  */
 440 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 441                                                  int mds)
 442 {
 443         struct ceph_mds_session *s;
 444
 445         if (mds >= mdsc->mdsmap->m_num_mds)
 446                 return ERR_PTR(-EINVAL);
 447
 448         s = kzalloc(sizeof(*s), GFP_NOFS);
 449         if (!s)
 450                 return ERR_PTR(-ENOMEM);
 451         s->s_mdsc = mdsc;
 452         s->s_mds = mds;
 453         s->s_state = CEPH_MDS_SESSION_NEW;
 454         s->s_ttl = 0;
 455         s->s_seq = 0;
 456         mutex_init(&s->s_mutex);
 457
 458         ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
 459
 460         spin_lock_init(&s->s_gen_ttl_lock);
 461         s->s_cap_gen = 0;
 462         s->s_cap_ttl = jiffies - 1;
 463
 464         spin_lock_init(&s->s_cap_lock);
 465         s->s_renew_requested = 0;
 466         s->s_renew_seq = 0;
 467         INIT_LIST_HEAD(&s->s_caps);
 468         s->s_nr_caps = 0;
 469         s->s_trim_caps = 0;
 470         refcount_set(&s->s_ref, 1);
 471         INIT_LIST_HEAD(&s->s_waiting);
 472         INIT_LIST_HEAD(&s->s_unsafe);
 473         s->s_num_cap_releases = 0;
 474         s->s_cap_reconnect = 0;
 475         s->s_cap_iterator = NULL;
 476         INIT_LIST_HEAD(&s->s_cap_releases);
 477         INIT_LIST_HEAD(&s->s_cap_flushing);
 478
 479         dout("register_session mds%d\n", mds);
 480         if (mds >= mdsc->max_sessions) {
 481                 int newmax = 1 << get_count_order(mds+1);
 482                 struct ceph_mds_session **sa;
 483
 484                 dout("register_session realloc to %d\n", newmax);
 485                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 486                 if (!sa)
 487                         goto fail_realloc;
 488                 if (mdsc->sessions) {
 489                         memcpy(sa, mdsc->sessions,
 490                                mdsc->max_sessions * sizeof(void *));
 491                         kfree(mdsc->sessions);
 492                 }
 493                 mdsc->sessions = sa;
 494                 mdsc->max_sessions = newmax;
 495         }
 496         mdsc->sessions[mds] = s;
 497         atomic_inc(&mdsc->num_sessions);
 498         refcount_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
 499
 500         ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
 501                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
 502
 503         return s;
 504
 505 fail_realloc:
 506         kfree(s);
 507         return ERR_PTR(-ENOMEM);
 508 }
 509
 510 /*
 511  * called under mdsc->mutex
 512  */
 513 static void __unregister_session(struct ceph_mds_client *mdsc,
 514                                struct ceph_mds_session *s)
 515 {
 516         dout("__unregister_session mds%d %p\n", s->s_mds, s);
 517         BUG_ON(mdsc->sessions[s->s_mds] != s);
 518         mdsc->sessions[s->s_mds] = NULL;
 519         ceph_con_close(&s->s_con);
 520         ceph_put_mds_session(s);
 521         atomic_dec(&mdsc->num_sessions);
 522 }
 523
 524 /*
 525  * drop session refs in request.
 526  *
 527  * should be last request ref, or hold mdsc->mutex
 528  */
 529 static void put_request_session(struct ceph_mds_request *req)
 530 {
 531         if (req->r_session) {
 532                 ceph_put_mds_session(req->r_session);
 533                 req->r_session = NULL;
 534         }
 535 }
 536
 537 void ceph_mdsc_release_request(struct kref *kref)
 538 {
 539         struct ceph_mds_request *req = container_of(kref,
 540                                                     struct ceph_mds_request,
 541                                                     r_kref);
 542         destroy_reply_info(&req->r_reply_info);
 543         if (req->r_request)
 544                 ceph_msg_put(req->r_request);
 545         if (req->r_reply)
 546                 ceph_msg_put(req->r_reply);
 547         if (req->r_inode) {
 548                 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
 549                 iput(req->r_inode);
 550         }
 551         if (req->r_parent)
 552                 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
 553         iput(req->r_target_inode);
 554         if (req->r_dentry)
 555                 dput(req->r_dentry);
 556         if (req->r_old_dentry)
 557                 dput(req->r_old_dentry);
 558         if (req->r_old_dentry_dir) {
 559                 /*
 560                  * track (and drop pins for) r_old_dentry_dir
 561                  * separately, since r_old_dentry's d_parent may have
 562                  * changed between the dir mutex being dropped and
 563                  * this request being freed.
 564                  */
 565                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
 566                                   CEPH_CAP_PIN);
 567                 iput(req->r_old_dentry_dir);
 568         }
 569         kfree(req->r_path1);
 570         kfree(req->r_path2);
 571         if (req->r_pagelist)
 572                 ceph_pagelist_release(req->r_pagelist);
 573         put_request_session(req);
 574         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
 575         kfree(req);
 576 }
 577
 578 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
 579
 580 /*
 581  * lookup session, bump ref if found.
 582  *
 583  * called under mdsc->mutex.
 584  */
 585 static struct ceph_mds_request *
 586 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
 587 {
 588         struct ceph_mds_request *req;
 589
 590         req = lookup_request(&mdsc->request_tree, tid);
 591         if (req)
 592                 ceph_mdsc_get_request(req);
 593
 594         return req;
 595 }
 596
 597 /*
 598  * Register an in-flight request, and assign a tid.  Link to directory
 599  * are modifying (if any).
 600  *
 601  * Called under mdsc->mutex.
 602  */
 603 static void __register_request(struct ceph_mds_client *mdsc,
 604                                struct ceph_mds_request *req,
 605                                struct inode *dir)
 606 {
 607         req->r_tid = ++mdsc->last_tid;
 608         if (req->r_num_caps)
 609                 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
 610                                   req->r_num_caps);
 611         dout("__register_request %p tid %lld\n", req, req->r_tid);
 612         ceph_mdsc_get_request(req);
 613         insert_request(&mdsc->request_tree, req);
 614
 615         req->r_uid = current_fsuid();
 616         req->r_gid = current_fsgid();
 617
 618         if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
 619                 mdsc->oldest_tid = req->r_tid;
 620
 621         if (dir) {
 622                 ihold(dir);
 623                 req->r_unsafe_dir = dir;
 624         }
 625 }
 626
 627 static void __unregister_request(struct ceph_mds_client *mdsc,
 628                                  struct ceph_mds_request *req)
 629 {
 630         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 631
 632         /* Never leave an unregistered request on an unsafe list! */
 633         list_del_init(&req->r_unsafe_item);
 634
 635         if (req->r_tid == mdsc->oldest_tid) {
 636                 struct rb_node *p = rb_next(&req->r_node);
 637                 mdsc->oldest_tid = 0;
 638                 while (p) {
 639                         struct ceph_mds_request *next_req =
 640                                 rb_entry(p, struct ceph_mds_request, r_node);
 641                         if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
 642                                 mdsc->oldest_tid = next_req->r_tid;
 643                                 break;
 644                         }
 645                         p = rb_next(p);
 646                 }
 647         }
 648
 649         erase_request(&mdsc->request_tree, req);
 650
 651         if (req->r_unsafe_dir  &&
 652             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 653                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
 654                 spin_lock(&ci->i_unsafe_lock);
 655                 list_del_init(&req->r_unsafe_dir_item);
 656                 spin_unlock(&ci->i_unsafe_lock);
 657         }
 658         if (req->r_target_inode &&
 659             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
 660                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
 661                 spin_lock(&ci->i_unsafe_lock);
 662                 list_del_init(&req->r_unsafe_target_item);
 663                 spin_unlock(&ci->i_unsafe_lock);
 664         }
 665
 666         if (req->r_unsafe_dir) {
 667                 iput(req->r_unsafe_dir);
 668                 req->r_unsafe_dir = NULL;
 669         }
 670
 671         complete_all(&req->r_safe_completion);
 672
 673         ceph_mdsc_put_request(req);
 674 }
 675
 676 /*
 677  * Walk back up the dentry tree until we hit a dentry representing a
 678  * non-snapshot inode. We do this using the rcu_read_lock (which must be held
 679  * when calling this) to ensure that the objects won't disappear while we're
 680  * working with them. Once we hit a candidate dentry, we attempt to take a
 681  * reference to it, and return that as the result.
 682  */
 683 static struct inode *get_nonsnap_parent(struct dentry *dentry)
 684 {
 685         struct inode *inode = NULL;
 686
 687         while (dentry && !IS_ROOT(dentry)) {
 688                 inode = d_inode_rcu(dentry);
 689                 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
 690                         break;
 691                 dentry = dentry->d_parent;
 692         }
 693         if (inode)
 694                 inode = igrab(inode);
 695         return inode;
 696 }
 697
 698 /*
 699  * Choose mds to send request to next.  If there is a hint set in the
 700  * request (e.g., due to a prior forward hint from the mds), use that.
 701  * Otherwise, consult frag tree and/or caps to identify the
 702  * appropriate mds.  If all else fails, choose randomly.
 703  *
 704  * Called under mdsc->mutex.
 705  */
 706 static int __choose_mds(struct ceph_mds_client *mdsc,
 707                         struct ceph_mds_request *req)
 708 {
 709         struct inode *inode;
 710         struct ceph_inode_info *ci;
 711         struct ceph_cap *cap;
 712         int mode = req->r_direct_mode;
 713         int mds = -1;
 714         u32 hash = req->r_direct_hash;
 715         bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
 716
 717         /*
 718          * is there a specific mds we should try?  ignore hint if we have
 719          * no session and the mds is not up (active or recovering).
 720          */
 721         if (req->r_resend_mds >= 0 &&
 722             (__have_session(mdsc, req->r_resend_mds) ||
 723              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
 724                 dout("choose_mds using resend_mds mds%d\n",
 725                      req->r_resend_mds);
 726                 return req->r_resend_mds;
 727         }
 728
 729         if (mode == USE_RANDOM_MDS)
 730                 goto random;
 731
 732         inode = NULL;
 733         if (req->r_inode) {
 734                 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
 735                         inode = req->r_inode;
 736                         ihold(inode);
 737                 } else {
 738                         /* req->r_dentry is non-null for LSSNAP request */
 739                         rcu_read_lock();
 740                         inode = get_nonsnap_parent(req->r_dentry);
 741                         rcu_read_unlock();
 742                         dout("__choose_mds using snapdir's parent %p\n", inode);
 743                 }
 744         } else if (req->r_dentry) {
 745                 /* ignore race with rename; old or new d_parent is okay */
 746                 struct dentry *parent;
 747                 struct inode *dir;
 748
 749                 rcu_read_lock();
 750                 parent = req->r_dentry->d_parent;
 751                 dir = req->r_parent ? : d_inode_rcu(parent);
 752
 753                 if (!dir || dir->i_sb != mdsc->fsc->sb) {
 754                         /*  not this fs or parent went negative */
 755                         inode = d_inode(req->r_dentry);
 756                         if (inode)
 757                                 ihold(inode);
 758                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
 759                         /* direct snapped/virtual snapdir requests
 760                          * based on parent dir inode */
 761                         inode = get_nonsnap_parent(parent);
 762                         dout("__choose_mds using nonsnap parent %p\n", inode);
 763                 } else {
 764                         /* dentry target */
 765                         inode = d_inode(req->r_dentry);
 766                         if (!inode || mode == USE_AUTH_MDS) {
 767                                 /* dir + name */
 768                                 inode = igrab(dir);
 769                                 hash = ceph_dentry_hash(dir, req->r_dentry);
 770                                 is_hash = true;
 771                         } else {
 772                                 ihold(inode);
 773                         }
 774                 }
 775                 rcu_read_unlock();
 776         }
 777
 778         dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
 779              (int)hash, mode);
 780         if (!inode)
 781                 goto random;
 782         ci = ceph_inode(inode);
 783
 784         if (is_hash && S_ISDIR(inode->i_mode)) {
 785                 struct ceph_inode_frag frag;
 786                 int found;
 787
 788                 ceph_choose_frag(ci, hash, &frag, &found);
 789                 if (found) {
 790                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
 791                                 u8 r;
 792
 793                                 /* choose a random replica */
 794                                 get_random_bytes(&r, 1);
 795                                 r %= frag.ndist;
 796                                 mds = frag.dist[r];
 797                                 dout("choose_mds %p %llx.%llx "
 798                                      "frag %u mds%d (%d/%d)\n",
 799                                      inode, ceph_vinop(inode),
 800                                      frag.frag, mds,
 801                                      (int)r, frag.ndist);
 802                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 803                                     CEPH_MDS_STATE_ACTIVE)
 804                                         goto out;
 805                         }
 806
 807                         /* since this file/dir wasn't known to be
 808                          * replicated, then we want to look for the
 809                          * authoritative mds. */
 810                         mode = USE_AUTH_MDS;
 811                         if (frag.mds >= 0) {
 812                                 /* choose auth mds */
 813                                 mds = frag.mds;
 814                                 dout("choose_mds %p %llx.%llx "
 815                                      "frag %u mds%d (auth)\n",
 816                                      inode, ceph_vinop(inode), frag.frag, mds);
 817                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
 818                                     CEPH_MDS_STATE_ACTIVE)
 819                                         goto out;
 820                         }
 821                 }
 822         }
 823
 824         spin_lock(&ci->i_ceph_lock);
 825         cap = NULL;
 826         if (mode == USE_AUTH_MDS)
 827                 cap = ci->i_auth_cap;
 828         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
 829                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
 830         if (!cap) {
 831                 spin_unlock(&ci->i_ceph_lock);
 832                 iput(inode);
 833                 goto random;
 834         }
 835         mds = cap->session->s_mds;
 836         dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
 837              inode, ceph_vinop(inode), mds,
 838              cap == ci->i_auth_cap ? "auth " : "", cap);
 839         spin_unlock(&ci->i_ceph_lock);
 840 out:
 841         iput(inode);
 842         return mds;
 843
 844 random:
 845         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
 846         dout("choose_mds chose random mds%d\n", mds);
 847         return mds;
 848 }
 849
 850
 851 /*
 852  * session messages
 853  */
 854 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
 855 {
 856         struct ceph_msg *msg;
 857         struct ceph_mds_session_head *h;
 858
 859         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
 860                            false);
 861         if (!msg) {
 862                 pr_err("create_session_msg ENOMEM creating msg\n");
 863                 return NULL;
 864         }
 865         h = msg->front.iov_base;
 866         h->op = cpu_to_le32(op);
 867         h->seq = cpu_to_le64(seq);
 868
 869         return msg;
 870 }
 871
 872 /*
 873  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
 874  * to include additional client metadata fields.
 875  */
 876 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
 877 {
 878         struct ceph_msg *msg;
 879         struct ceph_mds_session_head *h;
 880         int i = -1;
 881         int metadata_bytes = 0;
 882         int metadata_key_count = 0;
 883         struct ceph_options *opt = mdsc->fsc->client->options;
 884         struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
 885         void *p;
 886
 887         const char* metadata[][2] = {
 888                 {"hostname", mdsc->nodename},
 889                 {"kernel_version", init_utsname()->release},
 890                 {"entity_id", opt->name ? : ""},
 891                 {"root", fsopt->server_path ? : "/"},
 892                 {NULL, NULL}
 893         };
 894
 895         /* Calculate serialized length of metadata */
 896         metadata_bytes = 4;  /* map length */
 897         for (i = 0; metadata[i][0]; ++i) {
 898                 metadata_bytes += 8 + strlen(metadata[i][0]) +
 899                         strlen(metadata[i][1]);
 900                 metadata_key_count++;
 901         }
 902
 903         /* Allocate the message */
 904         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
 905                            GFP_NOFS, false);
 906         if (!msg) {
 907                 pr_err("create_session_msg ENOMEM creating msg\n");
 908                 return NULL;
 909         }
 910         h = msg->front.iov_base;
 911         h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
 912         h->seq = cpu_to_le64(seq);
 913
 914         /*
 915          * Serialize client metadata into waiting buffer space, using
 916          * the format that userspace expects for map<string, string>
 917          *
 918          * ClientSession messages with metadata are v2
 919          */
 920         msg->hdr.version = cpu_to_le16(2);
 921         msg->hdr.compat_version = cpu_to_le16(1);
 922
 923         /* The write pointer, following the session_head structure */
 924         p = msg->front.iov_base + sizeof(*h);
 925
 926         /* Number of entries in the map */
 927         ceph_encode_32(&p, metadata_key_count);
 928
 929         /* Two length-prefixed strings for each entry in the map */
 930         for (i = 0; metadata[i][0]; ++i) {
 931                 size_t const key_len = strlen(metadata[i][0]);
 932                 size_t const val_len = strlen(metadata[i][1]);
 933
 934                 ceph_encode_32(&p, key_len);
 935                 memcpy(p, metadata[i][0], key_len);
 936                 p += key_len;
 937                 ceph_encode_32(&p, val_len);
 938                 memcpy(p, metadata[i][1], val_len);
 939                 p += val_len;
 940         }
 941
 942         return msg;
 943 }
 944
 945 /*
 946  * send session open request.
 947  *
 948  * called under mdsc->mutex
 949  */
 950 static int __open_session(struct ceph_mds_client *mdsc,
 951                           struct ceph_mds_session *session)
 952 {
 953         struct ceph_msg *msg;
 954         int mstate;
 955         int mds = session->s_mds;
 956
 957         /* wait for mds to go active? */
 958         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
 959         dout("open_session to mds%d (%s)\n", mds,
 960              ceph_mds_state_name(mstate));
 961         session->s_state = CEPH_MDS_SESSION_OPENING;
 962         session->s_renew_requested = jiffies;
 963
 964         /* send connect message */
 965         msg = create_session_open_msg(mdsc, session->s_seq);
 966         if (!msg)
 967                 return -ENOMEM;
 968         ceph_con_send(&session->s_con, msg);
 969         return 0;
 970 }
 971
 972 /*
 973  * open sessions for any export targets for the given mds
 974  *
 975  * called under mdsc->mutex
 976  */
 977 static struct ceph_mds_session *
 978 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
 979 {
 980         struct ceph_mds_session *session;
 981
 982         session = __ceph_lookup_mds_session(mdsc, target);
 983         if (!session) {
 984                 session = register_session(mdsc, target);
 985                 if (IS_ERR(session))
 986                         return session;
 987         }
 988         if (session->s_state == CEPH_MDS_SESSION_NEW ||
 989             session->s_state == CEPH_MDS_SESSION_CLOSING)
 990                 __open_session(mdsc, session);
 991
 992         return session;
 993 }
 994
 995 struct ceph_mds_session *
 996 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
 997 {
 998         struct ceph_mds_session *session;
 999
1000         dout("open_export_target_session to mds%d\n", target);
1001
1002         mutex_lock(&mdsc->mutex);
1003         session = __open_export_target_session(mdsc, target);
1004         mutex_unlock(&mdsc->mutex);
1005
1006         return session;
1007 }
1008
1009 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1010                                           struct ceph_mds_session *session)
1011 {
1012         struct ceph_mds_info *mi;
1013         struct ceph_mds_session *ts;
1014         int i, mds = session->s_mds;
1015
1016         if (mds >= mdsc->mdsmap->m_num_mds)
1017                 return;
1018
1019         mi = &mdsc->mdsmap->m_info[mds];
1020         dout("open_export_target_sessions for mds%d (%d targets)\n",
1021              session->s_mds, mi->num_export_targets);
1022
1023         for (i = 0; i < mi->num_export_targets; i++) {
1024                 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1025                 if (!IS_ERR(ts))
1026                         ceph_put_mds_session(ts);
1027         }
1028 }
1029
1030 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1031                                            struct ceph_mds_session *session)
1032 {
1033         mutex_lock(&mdsc->mutex);
1034         __open_export_target_sessions(mdsc, session);
1035         mutex_unlock(&mdsc->mutex);
1036 }
1037
1038 /*
1039  * session caps
1040  */
1041
1042 /* caller holds s_cap_lock, we drop it */
1043 static void cleanup_cap_releases(struct ceph_mds_client *mdsc,
1044                                  struct ceph_mds_session *session)
1045         __releases(session->s_cap_lock)
1046 {
1047         LIST_HEAD(tmp_list);
1048         list_splice_init(&session->s_cap_releases, &tmp_list);
1049         session->s_num_cap_releases = 0;
1050         spin_unlock(&session->s_cap_lock);
1051
1052         dout("cleanup_cap_releases mds%d\n", session->s_mds);
1053         while (!list_empty(&tmp_list)) {
1054                 struct ceph_cap *cap;
1055                 /* zero out the in-progress message */
1056                 cap = list_first_entry(&tmp_list,
1057                                         struct ceph_cap, session_caps);
1058                 list_del(&cap->session_caps);
1059                 ceph_put_cap(mdsc, cap);
1060         }
1061 }
1062
1063 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1064                                      struct ceph_mds_session *session)
1065 {
1066         struct ceph_mds_request *req;
1067         struct rb_node *p;
1068
1069         dout("cleanup_session_requests mds%d\n", session->s_mds);
1070         mutex_lock(&mdsc->mutex);
1071         while (!list_empty(&session->s_unsafe)) {
1072                 req = list_first_entry(&session->s_unsafe,
1073                                        struct ceph_mds_request, r_unsafe_item);
1074                 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1075                                     req->r_tid);
1076                 __unregister_request(mdsc, req);
1077         }
1078         /* zero r_attempts, so kick_requests() will re-send requests */
1079         p = rb_first(&mdsc->request_tree);
1080         while (p) {
1081                 req = rb_entry(p, struct ceph_mds_request, r_node);
1082                 p = rb_next(p);
1083                 if (req->r_session &&
1084                     req->r_session->s_mds == session->s_mds)
1085                         req->r_attempts = 0;
1086         }
1087         mutex_unlock(&mdsc->mutex);
1088 }
1089
1090 /*
1091  * Helper to safely iterate over all caps associated with a session, with
1092  * special care taken to handle a racing __ceph_remove_cap().
1093  *
1094  * Caller must hold session s_mutex.
1095  */
1096 static int iterate_session_caps(struct ceph_mds_session *session,
1097                                  int (*cb)(struct inode *, struct ceph_cap *,
1098                                             void *), void *arg)
1099 {
1100         struct list_head *p;
1101         struct ceph_cap *cap;
1102         struct inode *inode, *last_inode = NULL;
1103         struct ceph_cap *old_cap = NULL;
1104         int ret;
1105
1106         dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1107         spin_lock(&session->s_cap_lock);
1108         p = session->s_caps.next;
1109         while (p != &session->s_caps) {
1110                 cap = list_entry(p, struct ceph_cap, session_caps);
1111                 inode = igrab(&cap->ci->vfs_inode);
1112                 if (!inode) {
1113                         p = p->next;
1114                         continue;
1115                 }
1116                 session->s_cap_iterator = cap;
1117                 spin_unlock(&session->s_cap_lock);
1118
1119                 if (last_inode) {
1120                         iput(last_inode);
1121                         last_inode = NULL;
1122                 }
1123                 if (old_cap) {
1124                         ceph_put_cap(session->s_mdsc, old_cap);
1125                         old_cap = NULL;
1126                 }
1127
1128                 ret = cb(inode, cap, arg);
1129                 last_inode = inode;
1130
1131                 spin_lock(&session->s_cap_lock);
1132                 p = p->next;
1133                 if (!cap->ci) {
1134                         dout("iterate_session_caps  finishing cap %p removal\n",
1135                              cap);
1136                         BUG_ON(cap->session != session);
1137                         cap->session = NULL;
1138                         list_del_init(&cap->session_caps);
1139                         session->s_nr_caps--;
1140                         if (cap->queue_release) {
1141                                 list_add_tail(&cap->session_caps,
1142                                               &session->s_cap_releases);
1143                                 session->s_num_cap_releases++;
1144                         } else {
1145                                 old_cap = cap;  /* put_cap it w/o locks held */
1146                         }
1147                 }
1148                 if (ret < 0)
1149                         goto out;
1150         }
1151         ret = 0;
1152 out:
1153         session->s_cap_iterator = NULL;
1154         spin_unlock(&session->s_cap_lock);
1155
1156         iput(last_inode);
1157         if (old_cap)
1158                 ceph_put_cap(session->s_mdsc, old_cap);
1159
1160         return ret;
1161 }
1162
1163 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1164                                   void *arg)
1165 {
1166         struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1167         struct ceph_inode_info *ci = ceph_inode(inode);
1168         LIST_HEAD(to_remove);
1169         bool drop = false;
1170         bool invalidate = false;
1171
1172         dout("removing cap %p, ci is %p, inode is %p\n",
1173              cap, ci, &ci->vfs_inode);
1174         spin_lock(&ci->i_ceph_lock);
1175         __ceph_remove_cap(cap, false);
1176         if (!ci->i_auth_cap) {
1177                 struct ceph_cap_flush *cf;
1178                 struct ceph_mds_client *mdsc = fsc->mdsc;
1179
1180                 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1181
1182                 if (ci->i_wrbuffer_ref > 0 &&
1183                     READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
1184                         invalidate = true;
1185
1186                 while (!list_empty(&ci->i_cap_flush_list)) {
1187                         cf = list_first_entry(&ci->i_cap_flush_list,
1188                                               struct ceph_cap_flush, i_list);
1189                         list_move(&cf->i_list, &to_remove);
1190                 }
1191
1192                 spin_lock(&mdsc->cap_dirty_lock);
1193
1194                 list_for_each_entry(cf, &to_remove, i_list)
1195                         list_del(&cf->g_list);
1196
1197                 if (!list_empty(&ci->i_dirty_item)) {
1198                         pr_warn_ratelimited(
1199                                 " dropping dirty %s state for %p %lld\n",
1200                                 ceph_cap_string(ci->i_dirty_caps),
1201                                 inode, ceph_ino(inode));
1202                         ci->i_dirty_caps = 0;
1203                         list_del_init(&ci->i_dirty_item);
1204                         drop = true;
1205                 }
1206                 if (!list_empty(&ci->i_flushing_item)) {
1207                         pr_warn_ratelimited(
1208                                 " dropping dirty+flushing %s state for %p %lld\n",
1209                                 ceph_cap_string(ci->i_flushing_caps),
1210                                 inode, ceph_ino(inode));
1211                         ci->i_flushing_caps = 0;
1212                         list_del_init(&ci->i_flushing_item);
1213                         mdsc->num_cap_flushing--;
1214                         drop = true;
1215                 }
1216                 spin_unlock(&mdsc->cap_dirty_lock);
1217
1218                 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1219                         list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1220                         ci->i_prealloc_cap_flush = NULL;
1221                 }
1222         }
1223         spin_unlock(&ci->i_ceph_lock);
1224         while (!list_empty(&to_remove)) {
1225                 struct ceph_cap_flush *cf;
1226                 cf = list_first_entry(&to_remove,
1227                                       struct ceph_cap_flush, i_list);
1228                 list_del(&cf->i_list);
1229                 ceph_free_cap_flush(cf);
1230         }
1231
1232         wake_up_all(&ci->i_cap_wq);
1233         if (invalidate)
1234                 ceph_queue_invalidate(inode);
1235         if (drop)
1236                 iput(inode);
1237         return 0;
1238 }
1239
1240 /*
1241  * caller must hold session s_mutex
1242  */
1243 static void remove_session_caps(struct ceph_mds_session *session)
1244 {
1245         struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1246         struct super_block *sb = fsc->sb;
1247         dout("remove_session_caps on %p\n", session);
1248         iterate_session_caps(session, remove_session_caps_cb, fsc);
1249
1250         wake_up_all(&fsc->mdsc->cap_flushing_wq);
1251
1252         spin_lock(&session->s_cap_lock);
1253         if (session->s_nr_caps > 0) {
1254                 struct inode *inode;
1255                 struct ceph_cap *cap, *prev = NULL;
1256                 struct ceph_vino vino;
1257                 /*
1258                  * iterate_session_caps() skips inodes that are being
1259                  * deleted, we need to wait until deletions are complete.
1260                  * __wait_on_freeing_inode() is designed for the job,
1261                  * but it is not exported, so use lookup inode function
1262                  * to access it.
1263                  */
1264                 while (!list_empty(&session->s_caps)) {
1265                         cap = list_entry(session->s_caps.next,
1266                                          struct ceph_cap, session_caps);
1267                         if (cap == prev)
1268                                 break;
1269                         prev = cap;
1270                         vino = cap->ci->i_vino;
1271                         spin_unlock(&session->s_cap_lock);
1272
1273                         inode = ceph_find_inode(sb, vino);
1274                         iput(inode);
1275
1276                         spin_lock(&session->s_cap_lock);
1277                 }
1278         }
1279
1280         // drop cap expires and unlock s_cap_lock
1281         cleanup_cap_releases(session->s_mdsc, session);
1282
1283         BUG_ON(session->s_nr_caps > 0);
1284         BUG_ON(!list_empty(&session->s_cap_flushing));
1285 }
1286
1287 /*
1288  * wake up any threads waiting on this session's caps.  if the cap is
1289  * old (didn't get renewed on the client reconnect), remove it now.
1290  *
1291  * caller must hold s_mutex.
1292  */
1293 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1294                               void *arg)
1295 {
1296         struct ceph_inode_info *ci = ceph_inode(inode);
1297
1298         if (arg) {
1299                 spin_lock(&ci->i_ceph_lock);
1300                 ci->i_wanted_max_size = 0;
1301                 ci->i_requested_max_size = 0;
1302                 spin_unlock(&ci->i_ceph_lock);
1303         }
1304         wake_up_all(&ci->i_cap_wq);
1305         return 0;
1306 }
1307
1308 static void wake_up_session_caps(struct ceph_mds_session *session,
1309                                  int reconnect)
1310 {
1311         dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1312         iterate_session_caps(session, wake_up_session_cb,
1313                              (void *)(unsigned long)reconnect);
1314 }
1315
1316 /*
1317  * Send periodic message to MDS renewing all currently held caps.  The
1318  * ack will reset the expiration for all caps from this session.
1319  *
1320  * caller holds s_mutex
1321  */
1322 static int send_renew_caps(struct ceph_mds_client *mdsc,
1323                            struct ceph_mds_session *session)
1324 {
1325         struct ceph_msg *msg;
1326         int state;
1327
1328         if (time_after_eq(jiffies, session->s_cap_ttl) &&
1329             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1330                 pr_info("mds%d caps stale\n", session->s_mds);
1331         session->s_renew_requested = jiffies;
1332
1333         /* do not try to renew caps until a recovering mds has reconnected
1334          * with its clients. */
1335         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1336         if (state < CEPH_MDS_STATE_RECONNECT) {
1337                 dout("send_renew_caps ignoring mds%d (%s)\n",
1338                      session->s_mds, ceph_mds_state_name(state));
1339                 return 0;
1340         }
1341
1342         dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1343                 ceph_mds_state_name(state));
1344         msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1345                                  ++session->s_renew_seq);
1346         if (!msg)
1347                 return -ENOMEM;
1348         ceph_con_send(&session->s_con, msg);
1349         return 0;
1350 }
1351
1352 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1353                              struct ceph_mds_session *session, u64 seq)
1354 {
1355         struct ceph_msg *msg;
1356
1357         dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1358              session->s_mds, ceph_session_state_name(session->s_state), seq);
1359         msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1360         if (!msg)
1361                 return -ENOMEM;
1362         ceph_con_send(&session->s_con, msg);
1363         return 0;
1364 }
1365
1366
1367 /*
1368  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1369  *
1370  * Called under session->s_mutex
1371  */
1372 static void renewed_caps(struct ceph_mds_client *mdsc,
1373                          struct ceph_mds_session *session, int is_renew)
1374 {
1375         int was_stale;
1376         int wake = 0;
1377
1378         spin_lock(&session->s_cap_lock);
1379         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1380
1381         session->s_cap_ttl = session->s_renew_requested +
1382                 mdsc->mdsmap->m_session_timeout*HZ;
1383
1384         if (was_stale) {
1385                 if (time_before(jiffies, session->s_cap_ttl)) {
1386                         pr_info("mds%d caps renewed\n", session->s_mds);
1387                         wake = 1;
1388                 } else {
1389                         pr_info("mds%d caps still stale\n", session->s_mds);
1390                 }
1391         }
1392         dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1393              session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1394              time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1395         spin_unlock(&session->s_cap_lock);
1396
1397         if (wake)
1398                 wake_up_session_caps(session, 0);
1399 }
1400
1401 /*
1402  * send a session close request
1403  */
1404 static int request_close_session(struct ceph_mds_client *mdsc,
1405                                  struct ceph_mds_session *session)
1406 {
1407         struct ceph_msg *msg;
1408
1409         dout("request_close_session mds%d state %s seq %lld\n",
1410              session->s_mds, ceph_session_state_name(session->s_state),
1411              session->s_seq);
1412         msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1413         if (!msg)
1414                 return -ENOMEM;
1415         ceph_con_send(&session->s_con, msg);
1416         return 1;
1417 }
1418
1419 /*
1420  * Called with s_mutex held.
1421  */
1422 static int __close_session(struct ceph_mds_client *mdsc,
1423                          struct ceph_mds_session *session)
1424 {
1425         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1426                 return 0;
1427         session->s_state = CEPH_MDS_SESSION_CLOSING;
1428         return request_close_session(mdsc, session);
1429 }
1430
1431 /*
1432  * Trim old(er) caps.
1433  *
1434  * Because we can't cache an inode without one or more caps, we do
1435  * this indirectly: if a cap is unused, we prune its aliases, at which
1436  * point the inode will hopefully get dropped to.
1437  *
1438  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
1439  * memory pressure from the MDS, though, so it needn't be perfect.
1440  */
1441 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1442 {
1443         struct ceph_mds_session *session = arg;
1444         struct ceph_inode_info *ci = ceph_inode(inode);
1445         int used, wanted, oissued, mine;
1446
1447         if (session->s_trim_caps <= 0)
1448                 return -1;
1449
1450         spin_lock(&ci->i_ceph_lock);
1451         mine = cap->issued | cap->implemented;
1452         used = __ceph_caps_used(ci);
1453         wanted = __ceph_caps_file_wanted(ci);
1454         oissued = __ceph_caps_issued_other(ci, cap);
1455
1456         dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1457              inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1458              ceph_cap_string(used), ceph_cap_string(wanted));
1459         if (cap == ci->i_auth_cap) {
1460                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1461                     !list_empty(&ci->i_cap_snaps))
1462                         goto out;
1463                 if ((used | wanted) & CEPH_CAP_ANY_WR)
1464                         goto out;
1465         }
1466         /* The inode has cached pages, but it's no longer used.
1467          * we can safely drop it */
1468         if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1469             !(oissued & CEPH_CAP_FILE_CACHE)) {
1470           used = 0;
1471           oissued = 0;
1472         }
1473         if ((used | wanted) & ~oissued & mine)
1474                 goto out;   /* we need these caps */
1475
1476         session->s_trim_caps--;
1477         if (oissued) {
1478                 /* we aren't the only cap.. just remove us */
1479                 __ceph_remove_cap(cap, true);
1480         } else {
1481                 /* try dropping referring dentries */
1482                 spin_unlock(&ci->i_ceph_lock);
1483                 d_prune_aliases(inode);
1484                 dout("trim_caps_cb %p cap %p  pruned, count now %d\n",
1485                      inode, cap, atomic_read(&inode->i_count));
1486                 return 0;
1487         }
1488
1489 out:
1490         spin_unlock(&ci->i_ceph_lock);
1491         return 0;
1492 }
1493
1494 /*
1495  * Trim session cap count down to some max number.
1496  */
1497 static int trim_caps(struct ceph_mds_client *mdsc,
1498                      struct ceph_mds_session *session,
1499                      int max_caps)
1500 {
1501         int trim_caps = session->s_nr_caps - max_caps;
1502
1503         dout("trim_caps mds%d start: %d / %d, trim %d\n",
1504              session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1505         if (trim_caps > 0) {
1506                 session->s_trim_caps = trim_caps;
1507                 iterate_session_caps(session, trim_caps_cb, session);
1508                 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1509                      session->s_mds, session->s_nr_caps, max_caps,
1510                         trim_caps - session->s_trim_caps);
1511                 session->s_trim_caps = 0;
1512         }
1513
1514         ceph_send_cap_releases(mdsc, session);
1515         return 0;
1516 }
1517
1518 static int check_caps_flush(struct ceph_mds_client *mdsc,
1519                             u64 want_flush_tid)
1520 {
1521         int ret = 1;
1522
1523         spin_lock(&mdsc->cap_dirty_lock);
1524         if (!list_empty(&mdsc->cap_flush_list)) {
1525                 struct ceph_cap_flush *cf =
1526                         list_first_entry(&mdsc->cap_flush_list,
1527                                          struct ceph_cap_flush, g_list);
1528                 if (cf->tid <= want_flush_tid) {
1529                         dout("check_caps_flush still flushing tid "
1530                              "%llu <= %llu\n", cf->tid, want_flush_tid);
1531                         ret = 0;
1532                 }
1533         }
1534         spin_unlock(&mdsc->cap_dirty_lock);
1535         return ret;
1536 }
1537
1538 /*
1539  * flush all dirty inode data to disk.
1540  *
1541  * returns true if we've flushed through want_flush_tid
1542  */
1543 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1544                             u64 want_flush_tid)
1545 {
1546         dout("check_caps_flush want %llu\n", want_flush_tid);
1547
1548         wait_event(mdsc->cap_flushing_wq,
1549                    check_caps_flush(mdsc, want_flush_tid));
1550
1551         dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1552 }
1553
1554 /*
1555  * called under s_mutex
1556  */
1557 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1558                             struct ceph_mds_session *session)
1559 {
1560         struct ceph_msg *msg = NULL;
1561         struct ceph_mds_cap_release *head;
1562         struct ceph_mds_cap_item *item;
1563         struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1564         struct ceph_cap *cap;
1565         LIST_HEAD(tmp_list);
1566         int num_cap_releases;
1567         __le32  barrier, *cap_barrier;
1568
1569         down_read(&osdc->lock);
1570         barrier = cpu_to_le32(osdc->epoch_barrier);
1571         up_read(&osdc->lock);
1572
1573         spin_lock(&session->s_cap_lock);
1574 again:
1575         list_splice_init(&session->s_cap_releases, &tmp_list);
1576         num_cap_releases = session->s_num_cap_releases;
1577         session->s_num_cap_releases = 0;
1578         spin_unlock(&session->s_cap_lock);
1579
1580         while (!list_empty(&tmp_list)) {
1581                 if (!msg) {
1582                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1583                                         PAGE_SIZE, GFP_NOFS, false);
1584                         if (!msg)
1585                                 goto out_err;
1586                         head = msg->front.iov_base;
1587                         head->num = cpu_to_le32(0);
1588                         msg->front.iov_len = sizeof(*head);
1589
1590                         msg->hdr.version = cpu_to_le16(2);
1591                         msg->hdr.compat_version = cpu_to_le16(1);
1592                 }
1593
1594                 cap = list_first_entry(&tmp_list, struct ceph_cap,
1595                                         session_caps);
1596                 list_del(&cap->session_caps);
1597                 num_cap_releases--;
1598
1599                 head = msg->front.iov_base;
1600                 le32_add_cpu(&head->num, 1);
1601                 item = msg->front.iov_base + msg->front.iov_len;
1602                 item->ino = cpu_to_le64(cap->cap_ino);
1603                 item->cap_id = cpu_to_le64(cap->cap_id);
1604                 item->migrate_seq = cpu_to_le32(cap->mseq);
1605                 item->seq = cpu_to_le32(cap->issue_seq);
1606                 msg->front.iov_len += sizeof(*item);
1607
1608                 ceph_put_cap(mdsc, cap);
1609
1610                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1611                         // Append cap_barrier field
1612                         cap_barrier = msg->front.iov_base + msg->front.iov_len;
1613                         *cap_barrier = barrier;
1614                         msg->front.iov_len += sizeof(*cap_barrier);
1615
1616                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1617                         dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1618                         ceph_con_send(&session->s_con, msg);
1619                         msg = NULL;
1620                 }
1621         }
1622
1623         BUG_ON(num_cap_releases != 0);
1624
1625         spin_lock(&session->s_cap_lock);
1626         if (!list_empty(&session->s_cap_releases))
1627                 goto again;
1628         spin_unlock(&session->s_cap_lock);
1629
1630         if (msg) {
1631                 // Append cap_barrier field
1632                 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1633                 *cap_barrier = barrier;
1634                 msg->front.iov_len += sizeof(*cap_barrier);
1635
1636                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1637                 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1638                 ceph_con_send(&session->s_con, msg);
1639         }
1640         return;
1641 out_err:
1642         pr_err("send_cap_releases mds%d, failed to allocate message\n",
1643                 session->s_mds);
1644         spin_lock(&session->s_cap_lock);
1645         list_splice(&tmp_list, &session->s_cap_releases);
1646         session->s_num_cap_releases += num_cap_releases;
1647         spin_unlock(&session->s_cap_lock);
1648 }
1649
1650 /*
1651  * requests
1652  */
1653
1654 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1655                                     struct inode *dir)
1656 {
1657         struct ceph_inode_info *ci = ceph_inode(dir);
1658         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1659         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1660         size_t size = sizeof(struct ceph_mds_reply_dir_entry);
1661         int order, num_entries;
1662
1663         spin_lock(&ci->i_ceph_lock);
1664         num_entries = ci->i_files + ci->i_subdirs;
1665         spin_unlock(&ci->i_ceph_lock);
1666         num_entries = max(num_entries, 1);
1667         num_entries = min(num_entries, opt->max_readdir);
1668
1669         order = get_order(size * num_entries);
1670         while (order >= 0) {
1671                 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
1672                                                              __GFP_NOWARN,
1673                                                              order);
1674                 if (rinfo->dir_entries)
1675                         break;
1676                 order--;
1677         }
1678         if (!rinfo->dir_entries)
1679                 return -ENOMEM;
1680
1681         num_entries = (PAGE_SIZE << order) / size;
1682         num_entries = min(num_entries, opt->max_readdir);
1683
1684         rinfo->dir_buf_size = PAGE_SIZE << order;
1685         req->r_num_caps = num_entries + 1;
1686         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1687         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1688         return 0;
1689 }
1690
1691 /*
1692  * Create an mds request.
1693  */
1694 struct ceph_mds_request *
1695 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1696 {
1697         struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1698
1699         if (!req)
1700                 return ERR_PTR(-ENOMEM);
1701
1702         mutex_init(&req->r_fill_mutex);
1703         req->r_mdsc = mdsc;
1704         req->r_started = jiffies;
1705         req->r_resend_mds = -1;
1706         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1707         INIT_LIST_HEAD(&req->r_unsafe_target_item);
1708         req->r_fmode = -1;
1709         kref_init(&req->r_kref);
1710         RB_CLEAR_NODE(&req->r_node);
1711         INIT_LIST_HEAD(&req->r_wait);
1712         init_completion(&req->r_completion);
1713         init_completion(&req->r_safe_completion);
1714         INIT_LIST_HEAD(&req->r_unsafe_item);
1715
1716         req->r_stamp = timespec_trunc(current_kernel_time(), mdsc->fsc->sb->s_time_gran);
1717
1718         req->r_op = op;
1719         req->r_direct_mode = mode;
1720         return req;
1721 }
1722
1723 /*
1724  * return oldest (lowest) request, tid in request tree, 0 if none.
1725  *
1726  * called under mdsc->mutex.
1727  */
1728 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1729 {
1730         if (RB_EMPTY_ROOT(&mdsc->request_tree))
1731                 return NULL;
1732         return rb_entry(rb_first(&mdsc->request_tree),
1733                         struct ceph_mds_request, r_node);
1734 }
1735
1736 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1737 {
1738         return mdsc->oldest_tid;
1739 }
1740
1741 /*
1742  * Build a dentry's path.  Allocate on heap; caller must kfree.  Based
1743  * on build_path_from_dentry in fs/cifs/dir.c.
1744  *
1745  * If @stop_on_nosnap, generate path relative to the first non-snapped
1746  * inode.
1747  *
1748  * Encode hidden .snap dirs as a double /, i.e.
1749  *   foo/.snap/bar -> foo//bar
1750  */
1751 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1752                            int stop_on_nosnap)
1753 {
1754         struct dentry *temp;
1755         char *path;
1756         int len, pos;
1757         unsigned seq;
1758
1759         if (!dentry)
1760                 return ERR_PTR(-EINVAL);
1761
1762 retry:
1763         len = 0;
1764         seq = read_seqbegin(&rename_lock);
1765         rcu_read_lock();
1766         for (temp = dentry; !IS_ROOT(temp);) {
1767                 struct inode *inode = d_inode(temp);
1768                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1769                         len++;  /* slash only */
1770                 else if (stop_on_nosnap && inode &&
1771                          ceph_snap(inode) == CEPH_NOSNAP)
1772                         break;
1773                 else
1774                         len += 1 + temp->d_name.len;
1775                 temp = temp->d_parent;
1776         }
1777         rcu_read_unlock();
1778         if (len)
1779                 len--;  /* no leading '/' */
1780
1781         path = kmalloc(len+1, GFP_NOFS);
1782         if (!path)
1783                 return ERR_PTR(-ENOMEM);
1784         pos = len;
1785         path[pos] = 0;  /* trailing null */
1786         rcu_read_lock();
1787         for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1788                 struct inode *inode;
1789
1790                 spin_lock(&temp->d_lock);
1791                 inode = d_inode(temp);
1792                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1793                         dout("build_path path+%d: %p SNAPDIR\n",
1794                              pos, temp);
1795                 } else if (stop_on_nosnap && inode &&
1796                            ceph_snap(inode) == CEPH_NOSNAP) {
1797                         spin_unlock(&temp->d_lock);
1798                         break;
1799                 } else {
1800                         pos -= temp->d_name.len;
1801                         if (pos < 0) {
1802                                 spin_unlock(&temp->d_lock);
1803                                 break;
1804                         }
1805                         strncpy(path + pos, temp->d_name.name,
1806                                 temp->d_name.len);
1807                 }
1808                 spin_unlock(&temp->d_lock);
1809                 if (pos)
1810                         path[--pos] = '/';
1811                 temp = temp->d_parent;
1812         }
1813         rcu_read_unlock();
1814         if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1815                 pr_err("build_path did not end path lookup where "
1816                        "expected, namelen is %d, pos is %d\n", len, pos);
1817                 /* presumably this is only possible if racing with a
1818                    rename of one of the parent directories (we can not
1819                    lock the dentries above us to prevent this, but
1820                    retrying should be harmless) */
1821                 kfree(path);
1822                 goto retry;
1823         }
1824
1825         *base = ceph_ino(d_inode(temp));
1826         *plen = len;
1827         dout("build_path on %p %d built %llx '%.*s'\n",
1828              dentry, d_count(dentry), *base, len, path);
1829         return path;
1830 }
1831
1832 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
1833                              const char **ppath, int *ppathlen, u64 *pino,
1834                              int *pfreepath)
1835 {
1836         char *path;
1837
1838         rcu_read_lock();
1839         if (!dir)
1840                 dir = d_inode_rcu(dentry->d_parent);
1841         if (dir && ceph_snap(dir) == CEPH_NOSNAP) {
1842                 *pino = ceph_ino(dir);
1843                 rcu_read_unlock();
1844                 *ppath = dentry->d_name.name;
1845                 *ppathlen = dentry->d_name.len;
1846                 return 0;
1847         }
1848         rcu_read_unlock();
1849         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1850         if (IS_ERR(path))
1851                 return PTR_ERR(path);
1852         *ppath = path;
1853         *pfreepath = 1;
1854         return 0;
1855 }
1856
1857 static int build_inode_path(struct inode *inode,
1858                             const char **ppath, int *ppathlen, u64 *pino,
1859                             int *pfreepath)
1860 {
1861         struct dentry *dentry;
1862         char *path;
1863
1864         if (ceph_snap(inode) == CEPH_NOSNAP) {
1865                 *pino = ceph_ino(inode);
1866                 *ppathlen = 0;
1867                 return 0;
1868         }
1869         dentry = d_find_alias(inode);
1870         path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1871         dput(dentry);
1872         if (IS_ERR(path))
1873                 return PTR_ERR(path);
1874         *ppath = path;
1875         *pfreepath = 1;
1876         return 0;
1877 }
1878
1879 /*
1880  * request arguments may be specified via an inode *, a dentry *, or
1881  * an explicit ino+path.
1882  */
1883 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1884                                   struct inode *rdiri, const char *rpath,
1885                                   u64 rino, const char **ppath, int *pathlen,
1886                                   u64 *ino, int *freepath)
1887 {
1888         int r = 0;
1889
1890         if (rinode) {
1891                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1892                 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1893                      ceph_snap(rinode));
1894         } else if (rdentry) {
1895                 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
1896                                         freepath);
1897                 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1898                      *ppath);
1899         } else if (rpath || rino) {
1900                 *ino = rino;
1901                 *ppath = rpath;
1902                 *pathlen = rpath ? strlen(rpath) : 0;
1903                 dout(" path %.*s\n", *pathlen, rpath);
1904         }
1905
1906         return r;
1907 }
1908
1909 /*
1910  * called under mdsc->mutex
1911  */
1912 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1913                                                struct ceph_mds_request *req,
1914                                                int mds, bool drop_cap_releases)
1915 {
1916         struct ceph_msg *msg;
1917         struct ceph_mds_request_head *head;
1918         const char *path1 = NULL;
1919         const char *path2 = NULL;
1920         u64 ino1 = 0, ino2 = 0;
1921         int pathlen1 = 0, pathlen2 = 0;
1922         int freepath1 = 0, freepath2 = 0;
1923         int len;
1924         u16 releases;
1925         void *p, *end;
1926         int ret;
1927
1928         ret = set_request_path_attr(req->r_inode, req->r_dentry,
1929                               req->r_parent, req->r_path1, req->r_ino1.ino,
1930                               &path1, &pathlen1, &ino1, &freepath1);
1931         if (ret < 0) {
1932                 msg = ERR_PTR(ret);
1933                 goto out;
1934         }
1935
1936         ret = set_request_path_attr(NULL, req->r_old_dentry,
1937                               req->r_old_dentry_dir,
1938                               req->r_path2, req->r_ino2.ino,
1939                               &path2, &pathlen2, &ino2, &freepath2);
1940         if (ret < 0) {
1941                 msg = ERR_PTR(ret);
1942                 goto out_free1;
1943         }
1944
1945         len = sizeof(*head) +
1946                 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1947                 sizeof(struct ceph_timespec);
1948
1949         /* calculate (max) length for cap releases */
1950         len += sizeof(struct ceph_mds_request_release) *
1951                 (!!req->r_inode_drop + !!req->r_dentry_drop +
1952                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1953         if (req->r_dentry_drop)
1954                 len += req->r_dentry->d_name.len;
1955         if (req->r_old_dentry_drop)
1956                 len += req->r_old_dentry->d_name.len;
1957
1958         msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1959         if (!msg) {
1960                 msg = ERR_PTR(-ENOMEM);
1961                 goto out_free2;
1962         }
1963
1964         msg->hdr.version = cpu_to_le16(2);
1965         msg->hdr.tid = cpu_to_le64(req->r_tid);
1966
1967         head = msg->front.iov_base;
1968         p = msg->front.iov_base + sizeof(*head);
1969         end = msg->front.iov_base + msg->front.iov_len;
1970
1971         head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1972         head->op = cpu_to_le32(req->r_op);
1973         head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1974         head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1975         head->args = req->r_args;
1976
1977         ceph_encode_filepath(&p, end, ino1, path1);
1978         ceph_encode_filepath(&p, end, ino2, path2);
1979
1980         /* make note of release offset, in case we need to replay */
1981         req->r_request_release_offset = p - msg->front.iov_base;
1982
1983         /* cap releases */
1984         releases = 0;
1985         if (req->r_inode_drop)
1986                 releases += ceph_encode_inode_release(&p,
1987                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
1988                       mds, req->r_inode_drop, req->r_inode_unless, 0);
1989         if (req->r_dentry_drop)
1990                 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1991                                 req->r_parent, mds, req->r_dentry_drop,
1992                                 req->r_dentry_unless);
1993         if (req->r_old_dentry_drop)
1994                 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1995                                 req->r_old_dentry_dir, mds,
1996                                 req->r_old_dentry_drop,
1997                                 req->r_old_dentry_unless);
1998         if (req->r_old_inode_drop)
1999                 releases += ceph_encode_inode_release(&p,
2000                       d_inode(req->r_old_dentry),
2001                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2002
2003         if (drop_cap_releases) {
2004                 releases = 0;
2005                 p = msg->front.iov_base + req->r_request_release_offset;
2006         }
2007
2008         head->num_releases = cpu_to_le16(releases);
2009
2010         /* time stamp */
2011         {
2012                 struct ceph_timespec ts;
2013                 ceph_encode_timespec(&ts, &req->r_stamp);
2014                 ceph_encode_copy(&p, &ts, sizeof(ts));
2015         }
2016
2017         BUG_ON(p > end);
2018         msg->front.iov_len = p - msg->front.iov_base;
2019         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2020
2021         if (req->r_pagelist) {
2022                 struct ceph_pagelist *pagelist = req->r_pagelist;
2023                 refcount_inc(&pagelist->refcnt);
2024                 ceph_msg_data_add_pagelist(msg, pagelist);
2025                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2026         } else {
2027                 msg->hdr.data_len = 0;
2028         }
2029
2030         msg->hdr.data_off = cpu_to_le16(0);
2031
2032 out_free2:
2033         if (freepath2)
2034                 kfree((char *)path2);
2035 out_free1:
2036         if (freepath1)
2037                 kfree((char *)path1);
2038 out:
2039         return msg;
2040 }
2041
2042 /*
2043  * called under mdsc->mutex if error, under no mutex if
2044  * success.
2045  */
2046 static void complete_request(struct ceph_mds_client *mdsc,
2047                              struct ceph_mds_request *req)
2048 {
2049         if (req->r_callback)
2050                 req->r_callback(mdsc, req);
2051         else
2052                 complete_all(&req->r_completion);
2053 }
2054
2055 /*
2056  * called under mdsc->mutex
2057  */
2058 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2059                                   struct ceph_mds_request *req,
2060                                   int mds, bool drop_cap_releases)
2061 {
2062         struct ceph_mds_request_head *rhead;
2063         struct ceph_msg *msg;
2064         int flags = 0;
2065
2066         req->r_attempts++;
2067         if (req->r_inode) {
2068                 struct ceph_cap *cap =
2069                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2070
2071                 if (cap)
2072                         req->r_sent_on_mseq = cap->mseq;
2073                 else
2074                         req->r_sent_on_mseq = -1;
2075         }
2076         dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2077              req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2078
2079         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2080                 void *p;
2081                 /*
2082                  * Replay.  Do not regenerate message (and rebuild
2083                  * paths, etc.); just use the original message.
2084                  * Rebuilding paths will break for renames because
2085                  * d_move mangles the src name.
2086                  */
2087                 msg = req->r_request;
2088                 rhead = msg->front.iov_base;
2089
2090                 flags = le32_to_cpu(rhead->flags);
2091                 flags |= CEPH_MDS_FLAG_REPLAY;
2092                 rhead->flags = cpu_to_le32(flags);
2093
2094                 if (req->r_target_inode)
2095                         rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2096
2097                 rhead->num_retry = req->r_attempts - 1;
2098
2099                 /* remove cap/dentry releases from message */
2100                 rhead->num_releases = 0;
2101
2102                 /* time stamp */
2103                 p = msg->front.iov_base + req->r_request_release_offset;
2104                 {
2105                         struct ceph_timespec ts;
2106                         ceph_encode_timespec(&ts, &req->r_stamp);
2107                         ceph_encode_copy(&p, &ts, sizeof(ts));
2108                 }
2109
2110                 msg->front.iov_len = p - msg->front.iov_base;
2111                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2112                 return 0;
2113         }
2114
2115         if (req->r_request) {
2116                 ceph_msg_put(req->r_request);
2117                 req->r_request = NULL;
2118         }
2119         msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2120         if (IS_ERR(msg)) {
2121                 req->r_err = PTR_ERR(msg);
2122                 return PTR_ERR(msg);
2123         }
2124         req->r_request = msg;
2125
2126         rhead = msg->front.iov_base;
2127         rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2128         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2129                 flags |= CEPH_MDS_FLAG_REPLAY;
2130         if (req->r_parent)
2131                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2132         rhead->flags = cpu_to_le32(flags);
2133         rhead->num_fwd = req->r_num_fwd;
2134         rhead->num_retry = req->r_attempts - 1;
2135         rhead->ino = 0;
2136
2137         dout(" r_parent = %p\n", req->r_parent);
2138         return 0;
2139 }
2140
2141 /*
2142  * send request, or put it on the appropriate wait list.
2143  */
2144 static int __do_request(struct ceph_mds_client *mdsc,
2145                         struct ceph_mds_request *req)
2146 {
2147         struct ceph_mds_session *session = NULL;
2148         int mds = -1;
2149         int err = 0;
2150
2151         if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2152                 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2153                         __unregister_request(mdsc, req);
2154                 goto out;
2155         }
2156
2157         if (req->r_timeout &&
2158             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2159                 dout("do_request timed out\n");
2160                 err = -EIO;
2161                 goto finish;
2162         }
2163         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2164                 dout("do_request forced umount\n");
2165                 err = -EIO;
2166                 goto finish;
2167         }
2168         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2169                 if (mdsc->mdsmap_err) {
2170                         err = mdsc->mdsmap_err;
2171                         dout("do_request mdsmap err %d\n", err);
2172                         goto finish;
2173                 }
2174                 if (mdsc->mdsmap->m_epoch == 0) {
2175                         dout("do_request no mdsmap, waiting for map\n");
2176                         list_add(&req->r_wait, &mdsc->waiting_for_map);
2177                         goto finish;
2178                 }
2179                 if (!(mdsc->fsc->mount_options->flags &
2180                       CEPH_MOUNT_OPT_MOUNTWAIT) &&
2181                     !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2182                         err = -ENOENT;
2183                         pr_info("probably no mds server is up\n");
2184                         goto finish;
2185                 }
2186         }
2187
2188         put_request_session(req);
2189
2190         mds = __choose_mds(mdsc, req);
2191         if (mds < 0 ||
2192             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2193                 dout("do_request no mds or not active, waiting for map\n");
2194                 list_add(&req->r_wait, &mdsc->waiting_for_map);
2195                 goto out;
2196         }
2197
2198         /* get, open session */
2199         session = __ceph_lookup_mds_session(mdsc, mds);
2200         if (!session) {
2201                 session = register_session(mdsc, mds);
2202                 if (IS_ERR(session)) {
2203                         err = PTR_ERR(session);
2204                         goto finish;
2205                 }
2206         }
2207         req->r_session = get_session(session);
2208
2209         dout("do_request mds%d session %p state %s\n", mds, session,
2210              ceph_session_state_name(session->s_state));
2211         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2212             session->s_state != CEPH_MDS_SESSION_HUNG) {
2213                 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2214                         err = -EACCES;
2215                         goto out_session;
2216                 }
2217                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2218                     session->s_state == CEPH_MDS_SESSION_CLOSING)
2219                         __open_session(mdsc, session);
2220                 list_add(&req->r_wait, &session->s_waiting);
2221                 goto out_session;
2222         }
2223
2224         /* send request */
2225         req->r_resend_mds = -1;   /* forget any previous mds hint */
2226
2227         if (req->r_request_started == 0)   /* note request start time */
2228                 req->r_request_started = jiffies;
2229
2230         err = __prepare_send_request(mdsc, req, mds, false);
2231         if (!err) {
2232                 ceph_msg_get(req->r_request);
2233                 ceph_con_send(&session->s_con, req->r_request);
2234         }
2235
2236 out_session:
2237         ceph_put_mds_session(session);
2238 finish:
2239         if (err) {
2240                 dout("__do_request early error %d\n", err);
2241                 req->r_err = err;
2242                 complete_request(mdsc, req);
2243                 __unregister_request(mdsc, req);
2244         }
2245 out:
2246         return err;
2247 }
2248
2249 /*
2250  * called under mdsc->mutex
2251  */
2252 static void __wake_requests(struct ceph_mds_client *mdsc,
2253                             struct list_head *head)
2254 {
2255         struct ceph_mds_request *req;
2256         LIST_HEAD(tmp_list);
2257
2258         list_splice_init(head, &tmp_list);
2259
2260         while (!list_empty(&tmp_list)) {
2261                 req = list_entry(tmp_list.next,
2262                                  struct ceph_mds_request, r_wait);
2263                 list_del_init(&req->r_wait);
2264                 dout(" wake request %p tid %llu\n", req, req->r_tid);
2265                 __do_request(mdsc, req);
2266         }
2267 }
2268
2269 /*
2270  * Wake up threads with requests pending for @mds, so that they can
2271  * resubmit their requests to a possibly different mds.
2272  */
2273 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2274 {
2275         struct ceph_mds_request *req;
2276         struct rb_node *p = rb_first(&mdsc->request_tree);
2277
2278         dout("kick_requests mds%d\n", mds);
2279         while (p) {
2280                 req = rb_entry(p, struct ceph_mds_request, r_node);
2281                 p = rb_next(p);
2282                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2283                         continue;
2284                 if (req->r_attempts > 0)
2285                         continue; /* only new requests */
2286                 if (req->r_session &&
2287                     req->r_session->s_mds == mds) {
2288                         dout(" kicking tid %llu\n", req->r_tid);
2289                         list_del_init(&req->r_wait);
2290                         __do_request(mdsc, req);
2291                 }
2292         }
2293 }
2294
2295 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2296                               struct ceph_mds_request *req)
2297 {
2298         dout("submit_request on %p\n", req);
2299         mutex_lock(&mdsc->mutex);
2300         __register_request(mdsc, req, NULL);
2301         __do_request(mdsc, req);
2302         mutex_unlock(&mdsc->mutex);
2303 }
2304
2305 /*
2306  * Synchrously perform an mds request.  Take care of all of the
2307  * session setup, forwarding, retry details.
2308  */
2309 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2310                          struct inode *dir,
2311                          struct ceph_mds_request *req)
2312 {
2313         int err;
2314
2315         dout("do_request on %p\n", req);
2316
2317         /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2318         if (req->r_inode)
2319                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2320         if (req->r_parent)
2321                 ceph_get_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
2322         if (req->r_old_dentry_dir)
2323                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2324                                   CEPH_CAP_PIN);
2325
2326         /* issue */
2327         mutex_lock(&mdsc->mutex);
2328         __register_request(mdsc, req, dir);
2329         __do_request(mdsc, req);
2330
2331         if (req->r_err) {
2332                 err = req->r_err;
2333                 goto out;
2334         }
2335
2336         /* wait */
2337         mutex_unlock(&mdsc->mutex);
2338         dout("do_request waiting\n");
2339         if (!req->r_timeout && req->r_wait_for_completion) {
2340                 err = req->r_wait_for_completion(mdsc, req);
2341         } else {
2342                 long timeleft = wait_for_completion_killable_timeout(
2343                                         &req->r_completion,
2344                                         ceph_timeout_jiffies(req->r_timeout));
2345                 if (timeleft > 0)
2346                         err = 0;
2347                 else if (!timeleft)
2348                         err = -EIO;  /* timed out */
2349                 else
2350                         err = timeleft;  /* killed */
2351         }
2352         dout("do_request waited, got %d\n", err);
2353         mutex_lock(&mdsc->mutex);
2354
2355         /* only abort if we didn't race with a real reply */
2356         if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2357                 err = le32_to_cpu(req->r_reply_info.head->result);
2358         } else if (err < 0) {
2359                 dout("aborted request %lld with %d\n", req->r_tid, err);
2360
2361                 /*
2362                  * ensure we aren't running concurrently with
2363                  * ceph_fill_trace or ceph_readdir_prepopulate, which
2364                  * rely on locks (dir mutex) held by our caller.
2365                  */
2366                 mutex_lock(&req->r_fill_mutex);
2367                 req->r_err = err;
2368                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2369                 mutex_unlock(&req->r_fill_mutex);
2370
2371                 if (req->r_parent &&
2372                     (req->r_op & CEPH_MDS_OP_WRITE))
2373                         ceph_invalidate_dir_request(req);
2374         } else {
2375                 err = req->r_err;
2376         }
2377
2378 out:
2379         mutex_unlock(&mdsc->mutex);
2380         dout("do_request %p done, result %d\n", req, err);
2381         return err;
2382 }
2383
2384 /*
2385  * Invalidate dir's completeness, dentry lease state on an aborted MDS
2386  * namespace request.
2387  */
2388 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2389 {
2390         struct inode *inode = req->r_parent;
2391
2392         dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2393
2394         ceph_dir_clear_complete(inode);
2395         if (req->r_dentry)
2396                 ceph_invalidate_dentry_lease(req->r_dentry);
2397         if (req->r_old_dentry)
2398                 ceph_invalidate_dentry_lease(req->r_old_dentry);
2399 }
2400
2401 /*
2402  * Handle mds reply.
2403  *
2404  * We take the session mutex and parse and process the reply immediately.
2405  * This preserves the logical ordering of replies, capabilities, etc., sent
2406  * by the MDS as they are applied to our local cache.
2407  */
2408 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2409 {
2410         struct ceph_mds_client *mdsc = session->s_mdsc;
2411         struct ceph_mds_request *req;
2412         struct ceph_mds_reply_head *head = msg->front.iov_base;
2413         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
2414         struct ceph_snap_realm *realm;
2415         u64 tid;
2416         int err, result;
2417         int mds = session->s_mds;
2418
2419         if (msg->front.iov_len < sizeof(*head)) {
2420                 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2421                 ceph_msg_dump(msg);
2422                 return;
2423         }
2424
2425         /* get request, session */
2426         tid = le64_to_cpu(msg->hdr.tid);
2427         mutex_lock(&mdsc->mutex);
2428         req = lookup_get_request(mdsc, tid);
2429         if (!req) {
2430                 dout("handle_reply on unknown tid %llu\n", tid);
2431                 mutex_unlock(&mdsc->mutex);
2432                 return;
2433         }
2434         dout("handle_reply %p\n", req);
2435
2436         /* correct session? */
2437         if (req->r_session != session) {
2438                 pr_err("mdsc_handle_reply got %llu on session mds%d"
2439                        " not mds%d\n", tid, session->s_mds,
2440                        req->r_session ? req->r_session->s_mds : -1);
2441                 mutex_unlock(&mdsc->mutex);
2442                 goto out;
2443         }
2444
2445         /* dup? */
2446         if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
2447             (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
2448                 pr_warn("got a dup %s reply on %llu from mds%d\n",
2449                            head->safe ? "safe" : "unsafe", tid, mds);
2450                 mutex_unlock(&mdsc->mutex);
2451                 goto out;
2452         }
2453         if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
2454                 pr_warn("got unsafe after safe on %llu from mds%d\n",
2455                            tid, mds);
2456                 mutex_unlock(&mdsc->mutex);
2457                 goto out;
2458         }
2459
2460         result = le32_to_cpu(head->result);
2461
2462         /*
2463          * Handle an ESTALE
2464          * if we're not talking to the authority, send to them
2465          * if the authority has changed while we weren't looking,
2466          * send to new authority
2467          * Otherwise we just have to return an ESTALE
2468          */
2469         if (result == -ESTALE) {
2470                 dout("got ESTALE on request %llu", req->r_tid);
2471                 req->r_resend_mds = -1;
2472                 if (req->r_direct_mode != USE_AUTH_MDS) {
2473                         dout("not using auth, setting for that now");
2474                         req->r_direct_mode = USE_AUTH_MDS;
2475                         __do_request(mdsc, req);
2476                         mutex_unlock(&mdsc->mutex);
2477                         goto out;
2478                 } else  {
2479                         int mds = __choose_mds(mdsc, req);
2480                         if (mds >= 0 && mds != req->r_session->s_mds) {
2481                                 dout("but auth changed, so resending");
2482                                 __do_request(mdsc, req);
2483                                 mutex_unlock(&mdsc->mutex);
2484                                 goto out;
2485                         }
2486                 }
2487                 dout("have to return ESTALE on request %llu", req->r_tid);
2488         }
2489
2490
2491         if (head->safe) {
2492                 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
2493                 __unregister_request(mdsc, req);
2494
2495                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2496                         /*
2497                          * We already handled the unsafe response, now do the
2498                          * cleanup.  No need to examine the response; the MDS
2499                          * doesn't include any result info in the safe
2500                          * response.  And even if it did, there is nothing
2501                          * useful we could do with a revised return value.
2502                          */
2503                         dout("got safe reply %llu, mds%d\n", tid, mds);
2504
2505                         /* last unsafe request during umount? */
2506                         if (mdsc->stopping && !__get_oldest_req(mdsc))
2507                                 complete_all(&mdsc->safe_umount_waiters);
2508                         mutex_unlock(&mdsc->mutex);
2509                         goto out;
2510                 }
2511         } else {
2512                 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
2513                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2514                 if (req->r_unsafe_dir) {
2515                         struct ceph_inode_info *ci =
2516                                         ceph_inode(req->r_unsafe_dir);
2517                         spin_lock(&ci->i_unsafe_lock);
2518                         list_add_tail(&req->r_unsafe_dir_item,
2519                                       &ci->i_unsafe_dirops);
2520                         spin_unlock(&ci->i_unsafe_lock);
2521                 }
2522         }
2523
2524         dout("handle_reply tid %lld result %d\n", tid, result);
2525         rinfo = &req->r_reply_info;
2526         err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2527         mutex_unlock(&mdsc->mutex);
2528
2529         mutex_lock(&session->s_mutex);
2530         if (err < 0) {
2531                 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2532                 ceph_msg_dump(msg);
2533                 goto out_err;
2534         }
2535
2536         /* snap trace */
2537         realm = NULL;
2538         if (rinfo->snapblob_len) {
2539                 down_write(&mdsc->snap_rwsem);
2540                 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2541                                 rinfo->snapblob + rinfo->snapblob_len,
2542                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2543                                 &realm);
2544                 downgrade_write(&mdsc->snap_rwsem);
2545         } else {
2546                 down_read(&mdsc->snap_rwsem);
2547         }
2548
2549         /* insert trace into our cache */
2550         mutex_lock(&req->r_fill_mutex);
2551         current->journal_info = req;
2552         err = ceph_fill_trace(mdsc->fsc->sb, req);
2553         if (err == 0) {
2554                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2555                                     req->r_op == CEPH_MDS_OP_LSSNAP))
2556                         ceph_readdir_prepopulate(req, req->r_session);
2557                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2558         }
2559         current->journal_info = NULL;
2560         mutex_unlock(&req->r_fill_mutex);
2561
2562         up_read(&mdsc->snap_rwsem);
2563         if (realm)
2564                 ceph_put_snap_realm(mdsc, realm);
2565
2566         if (err == 0 && req->r_target_inode &&
2567             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2568                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2569                 spin_lock(&ci->i_unsafe_lock);
2570                 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2571                 spin_unlock(&ci->i_unsafe_lock);
2572         }
2573 out_err:
2574         mutex_lock(&mdsc->mutex);
2575         if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2576                 if (err) {
2577                         req->r_err = err;
2578                 } else {
2579                         req->r_reply =  ceph_msg_get(msg);
2580                         set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
2581                 }
2582         } else {
2583                 dout("reply arrived after request %lld was aborted\n", tid);
2584         }
2585         mutex_unlock(&mdsc->mutex);
2586
2587         mutex_unlock(&session->s_mutex);
2588
2589         /* kick calling process */
2590         complete_request(mdsc, req);
2591 out:
2592         ceph_mdsc_put_request(req);
2593         return;
2594 }
2595
2596
2597
2598 /*
2599  * handle mds notification that our request has been forwarded.
2600  */
2601 static void handle_forward(struct ceph_mds_client *mdsc,
2602                            struct ceph_mds_session *session,
2603                            struct ceph_msg *msg)
2604 {
2605         struct ceph_mds_request *req;
2606         u64 tid = le64_to_cpu(msg->hdr.tid);
2607         u32 next_mds;
2608         u32 fwd_seq;
2609         int err = -EINVAL;
2610         void *p = msg->front.iov_base;
2611         void *end = p + msg->front.iov_len;
2612
2613         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2614         next_mds = ceph_decode_32(&p);
2615         fwd_seq = ceph_decode_32(&p);
2616
2617         mutex_lock(&mdsc->mutex);
2618         req = lookup_get_request(mdsc, tid);
2619         if (!req) {
2620                 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2621                 goto out;  /* dup reply? */
2622         }
2623
2624         if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2625                 dout("forward tid %llu aborted, unregistering\n", tid);
2626                 __unregister_request(mdsc, req);
2627         } else if (fwd_seq <= req->r_num_fwd) {
2628                 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2629                      tid, next_mds, req->r_num_fwd, fwd_seq);
2630         } else {
2631                 /* resend. forward race not possible; mds would drop */
2632                 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2633                 BUG_ON(req->r_err);
2634                 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
2635                 req->r_attempts = 0;
2636                 req->r_num_fwd = fwd_seq;
2637                 req->r_resend_mds = next_mds;
2638                 put_request_session(req);
2639                 __do_request(mdsc, req);
2640         }
2641         ceph_mdsc_put_request(req);
2642 out:
2643         mutex_unlock(&mdsc->mutex);
2644         return;
2645
2646 bad:
2647         pr_err("mdsc_handle_forward decode error err=%d\n", err);
2648 }
2649
2650 /*
2651  * handle a mds session control message
2652  */
2653 static void handle_session(struct ceph_mds_session *session,
2654                            struct ceph_msg *msg)
2655 {
2656         struct ceph_mds_client *mdsc = session->s_mdsc;
2657         u32 op;
2658         u64 seq;
2659         int mds = session->s_mds;
2660         struct ceph_mds_session_head *h = msg->front.iov_base;
2661         int wake = 0;
2662
2663         /* decode */
2664         if (msg->front.iov_len != sizeof(*h))
2665                 goto bad;
2666         op = le32_to_cpu(h->op);
2667         seq = le64_to_cpu(h->seq);
2668
2669         mutex_lock(&mdsc->mutex);
2670         if (op == CEPH_SESSION_CLOSE) {
2671                 get_session(session);
2672                 __unregister_session(mdsc, session);
2673         }
2674         /* FIXME: this ttl calculation is generous */
2675         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2676         mutex_unlock(&mdsc->mutex);
2677
2678         mutex_lock(&session->s_mutex);
2679
2680         dout("handle_session mds%d %s %p state %s seq %llu\n",
2681              mds, ceph_session_op_name(op), session,
2682              ceph_session_state_name(session->s_state), seq);
2683
2684         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2685                 session->s_state = CEPH_MDS_SESSION_OPEN;
2686                 pr_info("mds%d came back\n", session->s_mds);
2687         }
2688
2689         switch (op) {
2690         case CEPH_SESSION_OPEN:
2691                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2692                         pr_info("mds%d reconnect success\n", session->s_mds);
2693                 session->s_state = CEPH_MDS_SESSION_OPEN;
2694                 renewed_caps(mdsc, session, 0);
2695                 wake = 1;
2696                 if (mdsc->stopping)
2697                         __close_session(mdsc, session);
2698                 break;
2699
2700         case CEPH_SESSION_RENEWCAPS:
2701                 if (session->s_renew_seq == seq)
2702                         renewed_caps(mdsc, session, 1);
2703                 break;
2704
2705         case CEPH_SESSION_CLOSE:
2706                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2707                         pr_info("mds%d reconnect denied\n", session->s_mds);
2708                 cleanup_session_requests(mdsc, session);
2709                 remove_session_caps(session);
2710                 wake = 2; /* for good measure */
2711                 wake_up_all(&mdsc->session_close_wq);
2712                 break;
2713
2714         case CEPH_SESSION_STALE:
2715                 pr_info("mds%d caps went stale, renewing\n",
2716                         session->s_mds);
2717                 spin_lock(&session->s_gen_ttl_lock);
2718                 session->s_cap_gen++;
2719                 session->s_cap_ttl = jiffies - 1;
2720                 spin_unlock(&session->s_gen_ttl_lock);
2721                 send_renew_caps(mdsc, session);
2722                 break;
2723
2724         case CEPH_SESSION_RECALL_STATE:
2725                 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2726                 break;
2727
2728         case CEPH_SESSION_FLUSHMSG:
2729                 send_flushmsg_ack(mdsc, session, seq);
2730                 break;
2731
2732         case CEPH_SESSION_FORCE_RO:
2733                 dout("force_session_readonly %p\n", session);
2734                 spin_lock(&session->s_cap_lock);
2735                 session->s_readonly = true;
2736                 spin_unlock(&session->s_cap_lock);
2737                 wake_up_session_caps(session, 0);
2738                 break;
2739
2740         case CEPH_SESSION_REJECT:
2741                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
2742                 pr_info("mds%d rejected session\n", session->s_mds);
2743                 session->s_state = CEPH_MDS_SESSION_REJECTED;
2744                 cleanup_session_requests(mdsc, session);
2745                 remove_session_caps(session);
2746                 wake = 2; /* for good measure */
2747                 break;
2748
2749         default:
2750                 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2751                 WARN_ON(1);
2752         }
2753
2754         mutex_unlock(&session->s_mutex);
2755         if (wake) {
2756                 mutex_lock(&mdsc->mutex);
2757                 __wake_requests(mdsc, &session->s_waiting);
2758                 if (wake == 2)
2759                         kick_requests(mdsc, mds);
2760                 mutex_unlock(&mdsc->mutex);
2761         }
2762         if (op == CEPH_SESSION_CLOSE)
2763                 ceph_put_mds_session(session);
2764         return;
2765
2766 bad:
2767         pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2768                (int)msg->front.iov_len);
2769         ceph_msg_dump(msg);
2770         return;
2771 }
2772
2773
2774 /*
2775  * called under session->mutex.
2776  */
2777 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2778                                    struct ceph_mds_session *session)
2779 {
2780         struct ceph_mds_request *req, *nreq;
2781         struct rb_node *p;
2782         int err;
2783
2784         dout("replay_unsafe_requests mds%d\n", session->s_mds);
2785
2786         mutex_lock(&mdsc->mutex);
2787         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2788                 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2789                 if (!err) {
2790                         ceph_msg_get(req->r_request);
2791                         ceph_con_send(&session->s_con, req->r_request);
2792                 }
2793         }
2794
2795         /*
2796          * also re-send old requests when MDS enters reconnect stage. So that MDS
2797          * can process completed request in clientreplay stage.
2798          */
2799         p = rb_first(&mdsc->request_tree);
2800         while (p) {
2801                 req = rb_entry(p, struct ceph_mds_request, r_node);
2802                 p = rb_next(p);
2803                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2804                         continue;
2805                 if (req->r_attempts == 0)
2806                         continue; /* only old requests */
2807                 if (req->r_session &&
2808                     req->r_session->s_mds == session->s_mds) {
2809                         err = __prepare_send_request(mdsc, req,
2810                                                      session->s_mds, true);
2811                         if (!err) {
2812                                 ceph_msg_get(req->r_request);
2813                                 ceph_con_send(&session->s_con, req->r_request);
2814                         }
2815                 }
2816         }
2817         mutex_unlock(&mdsc->mutex);
2818 }
2819
2820 /*
2821  * Encode information about a cap for a reconnect with the MDS.
2822  */
2823 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2824                           void *arg)
2825 {
2826         union {
2827                 struct ceph_mds_cap_reconnect v2;
2828                 struct ceph_mds_cap_reconnect_v1 v1;
2829         } rec;
2830         struct ceph_inode_info *ci;
2831         struct ceph_reconnect_state *recon_state = arg;
2832         struct ceph_pagelist *pagelist = recon_state->pagelist;
2833         char *path;
2834         int pathlen, err;
2835         u64 pathbase;
2836         u64 snap_follows;
2837         struct dentry *dentry;
2838
2839         ci = cap->ci;
2840
2841         dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2842              inode, ceph_vinop(inode), cap, cap->cap_id,
2843              ceph_cap_string(cap->issued));
2844         err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2845         if (err)
2846                 return err;
2847
2848         dentry = d_find_alias(inode);
2849         if (dentry) {
2850                 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2851                 if (IS_ERR(path)) {
2852                         err = PTR_ERR(path);
2853                         goto out_dput;
2854                 }
2855         } else {
2856                 path = NULL;
2857                 pathlen = 0;
2858                 pathbase = 0;
2859         }
2860
2861         spin_lock(&ci->i_ceph_lock);
2862         cap->seq = 0;        /* reset cap seq */
2863         cap->issue_seq = 0;  /* and issue_seq */
2864         cap->mseq = 0;       /* and migrate_seq */
2865         cap->cap_gen = cap->session->s_cap_gen;
2866
2867         if (recon_state->msg_version >= 2) {
2868                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2869                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2870                 rec.v2.issued = cpu_to_le32(cap->issued);
2871                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2872                 rec.v2.pathbase = cpu_to_le64(pathbase);
2873                 rec.v2.flock_len = 0;
2874         } else {
2875                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2876                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2877                 rec.v1.issued = cpu_to_le32(cap->issued);
2878                 rec.v1.size = cpu_to_le64(inode->i_size);
2879                 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2880                 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2881                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2882                 rec.v1.pathbase = cpu_to_le64(pathbase);
2883         }
2884
2885         if (list_empty(&ci->i_cap_snaps)) {
2886                 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
2887         } else {
2888                 struct ceph_cap_snap *capsnap =
2889                         list_first_entry(&ci->i_cap_snaps,
2890                                          struct ceph_cap_snap, ci_item);
2891                 snap_follows = capsnap->follows;
2892         }
2893         spin_unlock(&ci->i_ceph_lock);
2894
2895         if (recon_state->msg_version >= 2) {
2896                 int num_fcntl_locks, num_flock_locks;
2897                 struct ceph_filelock *flocks;
2898                 size_t struct_len, total_len = 0;
2899                 u8 struct_v = 0;
2900
2901 encode_again:
2902                 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2903                 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2904                                  sizeof(struct ceph_filelock), GFP_NOFS);
2905                 if (!flocks) {
2906                         err = -ENOMEM;
2907                         goto out_free;
2908                 }
2909                 err = ceph_encode_locks_to_buffer(inode, flocks,
2910                                                   num_fcntl_locks,
2911                                                   num_flock_locks);
2912                 if (err) {
2913                         kfree(flocks);
2914                         if (err == -ENOSPC)
2915                                 goto encode_again;
2916                         goto out_free;
2917                 }
2918
2919                 if (recon_state->msg_version >= 3) {
2920                         /* version, compat_version and struct_len */
2921                         total_len = 2 * sizeof(u8) + sizeof(u32);
2922                         struct_v = 2;
2923                 }
2924                 /*
2925                  * number of encoded locks is stable, so copy to pagelist
2926                  */
2927                 struct_len = 2 * sizeof(u32) +
2928                             (num_fcntl_locks + num_flock_locks) *
2929                             sizeof(struct ceph_filelock);
2930                 rec.v2.flock_len = cpu_to_le32(struct_len);
2931
2932                 struct_len += sizeof(rec.v2);
2933                 struct_len += sizeof(u32) + pathlen;
2934
2935                 if (struct_v >= 2)
2936                         struct_len += sizeof(u64); /* snap_follows */
2937
2938                 total_len += struct_len;
2939                 err = ceph_pagelist_reserve(pagelist, total_len);
2940
2941                 if (!err) {
2942                         if (recon_state->msg_version >= 3) {
2943                                 ceph_pagelist_encode_8(pagelist, struct_v);
2944                                 ceph_pagelist_encode_8(pagelist, 1);
2945                                 ceph_pagelist_encode_32(pagelist, struct_len);
2946                         }
2947                         ceph_pagelist_encode_string(pagelist, path, pathlen);
2948                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
2949                         ceph_locks_to_pagelist(flocks, pagelist,
2950                                                num_fcntl_locks,
2951                                                num_flock_locks);
2952                         if (struct_v >= 2)
2953                                 ceph_pagelist_encode_64(pagelist, snap_follows);
2954                 }
2955                 kfree(flocks);
2956         } else {
2957                 size_t size = sizeof(u32) + pathlen + sizeof(rec.v1);
2958                 err = ceph_pagelist_reserve(pagelist, size);
2959                 if (!err) {
2960                         ceph_pagelist_encode_string(pagelist, path, pathlen);
2961                         ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
2962                 }
2963         }
2964
2965         recon_state->nr_caps++;
2966 out_free:
2967         kfree(path);
2968 out_dput:
2969         dput(dentry);
2970         return err;
2971 }
2972
2973
2974 /*
2975  * If an MDS fails and recovers, clients need to reconnect in order to
2976  * reestablish shared state.  This includes all caps issued through
2977  * this session _and_ the snap_realm hierarchy.  Because it's not
2978  * clear which snap realms the mds cares about, we send everything we
2979  * know about.. that ensures we'll then get any new info the
2980  * recovering MDS might have.
2981  *
2982  * This is a relatively heavyweight operation, but it's rare.
2983  *
2984  * called with mdsc->mutex held.
2985  */
2986 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2987                                struct ceph_mds_session *session)
2988 {
2989         struct ceph_msg *reply;
2990         struct rb_node *p;
2991         int mds = session->s_mds;
2992         int err = -ENOMEM;
2993         int s_nr_caps;
2994         struct ceph_pagelist *pagelist;
2995         struct ceph_reconnect_state recon_state;
2996
2997         pr_info("mds%d reconnect start\n", mds);
2998
2999         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
3000         if (!pagelist)
3001                 goto fail_nopagelist;
3002         ceph_pagelist_init(pagelist);
3003
3004         reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
3005         if (!reply)
3006                 goto fail_nomsg;
3007
3008         mutex_lock(&session->s_mutex);
3009         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3010         session->s_seq = 0;
3011
3012         dout("session %p state %s\n", session,
3013              ceph_session_state_name(session->s_state));
3014
3015         spin_lock(&session->s_gen_ttl_lock);
3016         session->s_cap_gen++;
3017         spin_unlock(&session->s_gen_ttl_lock);
3018
3019         spin_lock(&session->s_cap_lock);
3020         /* don't know if session is readonly */
3021         session->s_readonly = 0;
3022         /*
3023          * notify __ceph_remove_cap() that we are composing cap reconnect.
3024          * If a cap get released before being added to the cap reconnect,
3025          * __ceph_remove_cap() should skip queuing cap release.
3026          */
3027         session->s_cap_reconnect = 1;
3028         /* drop old cap expires; we're about to reestablish that state */
3029         cleanup_cap_releases(mdsc, session);
3030
3031         /* trim unused caps to reduce MDS's cache rejoin time */
3032         if (mdsc->fsc->sb->s_root)
3033                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3034
3035         ceph_con_close(&session->s_con);
3036         ceph_con_open(&session->s_con,
3037                       CEPH_ENTITY_TYPE_MDS, mds,
3038                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3039
3040         /* replay unsafe requests */
3041         replay_unsafe_requests(mdsc, session);
3042
3043         down_read(&mdsc->snap_rwsem);
3044
3045         /* traverse this session's caps */
3046         s_nr_caps = session->s_nr_caps;
3047         err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
3048         if (err)
3049                 goto fail;
3050
3051         recon_state.nr_caps = 0;
3052         recon_state.pagelist = pagelist;
3053         if (session->s_con.peer_features & CEPH_FEATURE_MDSENC)
3054                 recon_state.msg_version = 3;
3055         else if (session->s_con.peer_features & CEPH_FEATURE_FLOCK)
3056                 recon_state.msg_version = 2;
3057         else
3058                 recon_state.msg_version = 1;
3059         err = iterate_session_caps(session, encode_caps_cb, &recon_state);
3060         if (err < 0)
3061                 goto fail;
3062
3063         spin_lock(&session->s_cap_lock);
3064         session->s_cap_reconnect = 0;
3065         spin_unlock(&session->s_cap_lock);
3066
3067         /*
3068          * snaprealms.  we provide mds with the ino, seq (version), and
3069          * parent for all of our realms.  If the mds has any newer info,
3070          * it will tell us.
3071          */
3072         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3073                 struct ceph_snap_realm *realm =
3074                         rb_entry(p, struct ceph_snap_realm, node);
3075                 struct ceph_mds_snaprealm_reconnect sr_rec;
3076
3077                 dout(" adding snap realm %llx seq %lld parent %llx\n",
3078                      realm->ino, realm->seq, realm->parent_ino);
3079                 sr_rec.ino = cpu_to_le64(realm->ino);
3080                 sr_rec.seq = cpu_to_le64(realm->seq);
3081                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3082                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3083                 if (err)
3084                         goto fail;
3085         }
3086
3087         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3088
3089         /* raced with cap release? */
3090         if (s_nr_caps != recon_state.nr_caps) {
3091                 struct page *page = list_first_entry(&pagelist->head,
3092                                                      struct page, lru);
3093                 __le32 *addr = kmap_atomic(page);
3094                 *addr = cpu_to_le32(recon_state.nr_caps);
3095                 kunmap_atomic(addr);
3096         }
3097
3098         reply->hdr.data_len = cpu_to_le32(pagelist->length);
3099         ceph_msg_data_add_pagelist(reply, pagelist);
3100
3101         ceph_early_kick_flushing_caps(mdsc, session);
3102
3103         ceph_con_send(&session->s_con, reply);
3104
3105         mutex_unlock(&session->s_mutex);
3106
3107         mutex_lock(&mdsc->mutex);
3108         __wake_requests(mdsc, &session->s_waiting);
3109         mutex_unlock(&mdsc->mutex);
3110
3111         up_read(&mdsc->snap_rwsem);
3112         return;
3113
3114 fail:
3115         ceph_msg_put(reply);
3116         up_read(&mdsc->snap_rwsem);
3117         mutex_unlock(&session->s_mutex);
3118 fail_nomsg:
3119         ceph_pagelist_release(pagelist);
3120 fail_nopagelist:
3121         pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3122         return;
3123 }
3124
3125
3126 /*
3127  * compare old and new mdsmaps, kicking requests
3128  * and closing out old connections as necessary
3129  *
3130  * called under mdsc->mutex.
3131  */
3132 static void check_new_map(struct ceph_mds_client *mdsc,
3133                           struct ceph_mdsmap *newmap,
3134                           struct ceph_mdsmap *oldmap)
3135 {
3136         int i;
3137         int oldstate, newstate;
3138         struct ceph_mds_session *s;
3139
3140         dout("check_new_map new %u old %u\n",
3141              newmap->m_epoch, oldmap->m_epoch);
3142
3143         for (i = 0; i < oldmap->m_num_mds && i < mdsc->max_sessions; i++) {
3144                 if (!mdsc->sessions[i])
3145                         continue;
3146                 s = mdsc->sessions[i];
3147                 oldstate = ceph_mdsmap_get_state(oldmap, i);
3148                 newstate = ceph_mdsmap_get_state(newmap, i);
3149
3150                 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3151                      i, ceph_mds_state_name(oldstate),
3152                      ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3153                      ceph_mds_state_name(newstate),
3154                      ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3155                      ceph_session_state_name(s->s_state));
3156
3157                 if (i >= newmap->m_num_mds ||
3158                     memcmp(ceph_mdsmap_get_addr(oldmap, i),
3159                            ceph_mdsmap_get_addr(newmap, i),
3160                            sizeof(struct ceph_entity_addr))) {
3161                         if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3162                                 /* the session never opened, just close it
3163                                  * out now */
3164                                 get_session(s);
3165                                 __unregister_session(mdsc, s);
3166                                 __wake_requests(mdsc, &s->s_waiting);
3167                                 ceph_put_mds_session(s);
3168                         } else if (i >= newmap->m_num_mds) {
3169                                 /* force close session for stopped mds */
3170                                 get_session(s);
3171                                 __unregister_session(mdsc, s);
3172                                 __wake_requests(mdsc, &s->s_waiting);
3173                                 kick_requests(mdsc, i);
3174                                 mutex_unlock(&mdsc->mutex);
3175
3176                                 mutex_lock(&s->s_mutex);
3177                                 cleanup_session_requests(mdsc, s);
3178                                 remove_session_caps(s);
3179                                 mutex_unlock(&s->s_mutex);
3180
3181                                 ceph_put_mds_session(s);
3182
3183                                 mutex_lock(&mdsc->mutex);
3184                         } else {
3185                                 /* just close it */
3186                                 mutex_unlock(&mdsc->mutex);
3187                                 mutex_lock(&s->s_mutex);
3188                                 mutex_lock(&mdsc->mutex);
3189                                 ceph_con_close(&s->s_con);
3190                                 mutex_unlock(&s->s_mutex);
3191                                 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3192                         }
3193                 } else if (oldstate == newstate) {
3194                         continue;  /* nothing new with this mds */
3195                 }
3196
3197                 /*
3198                  * send reconnect?
3199                  */
3200                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3201                     newstate >= CEPH_MDS_STATE_RECONNECT) {
3202                         mutex_unlock(&mdsc->mutex);
3203                         send_mds_reconnect(mdsc, s);
3204                         mutex_lock(&mdsc->mutex);
3205                 }
3206
3207                 /*
3208                  * kick request on any mds that has gone active.
3209                  */
3210                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3211                     newstate >= CEPH_MDS_STATE_ACTIVE) {
3212                         if (oldstate != CEPH_MDS_STATE_CREATING &&
3213                             oldstate != CEPH_MDS_STATE_STARTING)
3214                                 pr_info("mds%d recovery completed\n", s->s_mds);
3215                         kick_requests(mdsc, i);
3216                         ceph_kick_flushing_caps(mdsc, s);
3217                         wake_up_session_caps(s, 1);
3218                 }
3219         }
3220
3221         for (i = 0; i < newmap->m_num_mds && i < mdsc->max_sessions; i++) {
3222                 s = mdsc->sessions[i];
3223                 if (!s)
3224                         continue;
3225                 if (!ceph_mdsmap_is_laggy(newmap, i))
3226                         continue;
3227                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3228                     s->s_state == CEPH_MDS_SESSION_HUNG ||
3229                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
3230                         dout(" connecting to export targets of laggy mds%d\n",
3231                              i);
3232                         __open_export_target_sessions(mdsc, s);
3233                 }
3234         }
3235 }
3236
3237
3238
3239 /*
3240  * leases
3241  */
3242
3243 /*
3244  * caller must hold session s_mutex, dentry->d_lock
3245  */
3246 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3247 {
3248         struct ceph_dentry_info *di = ceph_dentry(dentry);
3249
3250         ceph_put_mds_session(di->lease_session);
3251         di->lease_session = NULL;
3252 }
3253
3254 static void handle_lease(struct ceph_mds_client *mdsc,
3255                          struct ceph_mds_session *session,
3256                          struct ceph_msg *msg)
3257 {
3258         struct super_block *sb = mdsc->fsc->sb;
3259         struct inode *inode;
3260         struct dentry *parent, *dentry;
3261         struct ceph_dentry_info *di;
3262         int mds = session->s_mds;
3263         struct ceph_mds_lease *h = msg->front.iov_base;
3264         u32 seq;
3265         struct ceph_vino vino;
3266         struct qstr dname;
3267         int release = 0;
3268
3269         dout("handle_lease from mds%d\n", mds);
3270
3271         /* decode */
3272         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3273                 goto bad;
3274         vino.ino = le64_to_cpu(h->ino);
3275         vino.snap = CEPH_NOSNAP;
3276         seq = le32_to_cpu(h->seq);
3277         dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3278         dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3279         if (dname.len != get_unaligned_le32(h+1))
3280                 goto bad;
3281
3282         /* lookup inode */
3283         inode = ceph_find_inode(sb, vino);
3284         dout("handle_lease %s, ino %llx %p %.*s\n",
3285              ceph_lease_op_name(h->action), vino.ino, inode,
3286              dname.len, dname.name);
3287
3288         mutex_lock(&session->s_mutex);
3289         session->s_seq++;
3290
3291         if (!inode) {
3292                 dout("handle_lease no inode %llx\n", vino.ino);
3293                 goto release;
3294         }
3295
3296         /* dentry */
3297         parent = d_find_alias(inode);
3298         if (!parent) {
3299                 dout("no parent dentry on inode %p\n", inode);
3300                 WARN_ON(1);
3301                 goto release;  /* hrm... */
3302         }
3303         dname.hash = full_name_hash(parent, dname.name, dname.len);
3304         dentry = d_lookup(parent, &dname);
3305         dput(parent);
3306         if (!dentry)
3307                 goto release;
3308
3309         spin_lock(&dentry->d_lock);
3310         di = ceph_dentry(dentry);
3311         switch (h->action) {
3312         case CEPH_MDS_LEASE_REVOKE:
3313                 if (di->lease_session == session) {
3314                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3315                                 h->seq = cpu_to_le32(di->lease_seq);
3316                         __ceph_mdsc_drop_dentry_lease(dentry);
3317                 }
3318                 release = 1;
3319                 break;
3320
3321         case CEPH_MDS_LEASE_RENEW:
3322                 if (di->lease_session == session &&
3323                     di->lease_gen == session->s_cap_gen &&
3324                     di->lease_renew_from &&
3325                     di->lease_renew_after == 0) {
3326                         unsigned long duration =
3327                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3328
3329                         di->lease_seq = seq;
3330                         di->time = di->lease_renew_from + duration;
3331                         di->lease_renew_after = di->lease_renew_from +
3332                                 (duration >> 1);
3333                         di->lease_renew_from = 0;
3334                 }
3335                 break;
3336         }
3337         spin_unlock(&dentry->d_lock);
3338         dput(dentry);
3339
3340         if (!release)
3341                 goto out;
3342
3343 release:
3344         /* let's just reuse the same message */
3345         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3346         ceph_msg_get(msg);
3347         ceph_con_send(&session->s_con, msg);
3348
3349 out:
3350         iput(inode);
3351         mutex_unlock(&session->s_mutex);
3352         return;
3353
3354 bad:
3355         pr_err("corrupt lease message\n");
3356         ceph_msg_dump(msg);
3357 }
3358
3359 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3360                               struct inode *inode,
3361                               struct dentry *dentry, char action,
3362                               u32 seq)
3363 {
3364         struct ceph_msg *msg;
3365         struct ceph_mds_lease *lease;
3366         int len = sizeof(*lease) + sizeof(u32);
3367         int dnamelen = 0;
3368
3369         dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3370              inode, dentry, ceph_lease_op_name(action), session->s_mds);
3371         dnamelen = dentry->d_name.len;
3372         len += dnamelen;
3373
3374         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3375         if (!msg)
3376                 return;
3377         lease = msg->front.iov_base;
3378         lease->action = action;
3379         lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3380         lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3381         lease->seq = cpu_to_le32(seq);
3382         put_unaligned_le32(dnamelen, lease + 1);
3383         memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3384
3385         /*
3386          * if this is a preemptive lease RELEASE, no need to
3387          * flush request stream, since the actual request will
3388          * soon follow.
3389          */
3390         msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3391
3392         ceph_con_send(&session->s_con, msg);
3393 }
3394
3395 /*
3396  * drop all leases (and dentry refs) in preparation for umount
3397  */
3398 static void drop_leases(struct ceph_mds_client *mdsc)
3399 {
3400         int i;
3401
3402         dout("drop_leases\n");
3403         mutex_lock(&mdsc->mutex);
3404         for (i = 0; i < mdsc->max_sessions; i++) {
3405                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3406                 if (!s)
3407                         continue;
3408                 mutex_unlock(&mdsc->mutex);
3409                 mutex_lock(&s->s_mutex);
3410                 mutex_unlock(&s->s_mutex);
3411                 ceph_put_mds_session(s);
3412                 mutex_lock(&mdsc->mutex);
3413         }
3414         mutex_unlock(&mdsc->mutex);
3415 }
3416
3417
3418
3419 /*
3420  * delayed work -- periodically trim expired leases, renew caps with mds
3421  */
3422 static void schedule_delayed(struct ceph_mds_client *mdsc)
3423 {
3424         int delay = 5;
3425         unsigned hz = round_jiffies_relative(HZ * delay);
3426         schedule_delayed_work(&mdsc->delayed_work, hz);
3427 }
3428
3429 static void delayed_work(struct work_struct *work)
3430 {
3431         int i;
3432         struct ceph_mds_client *mdsc =
3433                 container_of(work, struct ceph_mds_client, delayed_work.work);
3434         int renew_interval;
3435         int renew_caps;
3436
3437         dout("mdsc delayed_work\n");
3438         ceph_check_delayed_caps(mdsc);
3439
3440         mutex_lock(&mdsc->mutex);
3441         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3442         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3443                                    mdsc->last_renew_caps);
3444         if (renew_caps)
3445                 mdsc->last_renew_caps = jiffies;
3446
3447         for (i = 0; i < mdsc->max_sessions; i++) {
3448                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3449                 if (!s)
3450                         continue;
3451                 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3452                         dout("resending session close request for mds%d\n",
3453                              s->s_mds);
3454                         request_close_session(mdsc, s);
3455                         ceph_put_mds_session(s);
3456                         continue;
3457                 }
3458                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3459                         if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3460                                 s->s_state = CEPH_MDS_SESSION_HUNG;
3461                                 pr_info("mds%d hung\n", s->s_mds);
3462                         }
3463                 }
3464                 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3465                         /* this mds is failed or recovering, just wait */
3466                         ceph_put_mds_session(s);
3467                         continue;
3468                 }
3469                 mutex_unlock(&mdsc->mutex);
3470
3471                 mutex_lock(&s->s_mutex);
3472                 if (renew_caps)
3473                         send_renew_caps(mdsc, s);
3474                 else
3475                         ceph_con_keepalive(&s->s_con);
3476                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3477                     s->s_state == CEPH_MDS_SESSION_HUNG)
3478                         ceph_send_cap_releases(mdsc, s);
3479                 mutex_unlock(&s->s_mutex);
3480                 ceph_put_mds_session(s);
3481
3482                 mutex_lock(&mdsc->mutex);
3483         }
3484         mutex_unlock(&mdsc->mutex);
3485
3486         schedule_delayed(mdsc);
3487 }
3488
3489 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3490
3491 {
3492         struct ceph_mds_client *mdsc;
3493
3494         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3495         if (!mdsc)
3496                 return -ENOMEM;
3497         mdsc->fsc = fsc;
3498         fsc->mdsc = mdsc;
3499         mutex_init(&mdsc->mutex);
3500         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3501         if (!mdsc->mdsmap) {
3502                 kfree(mdsc);
3503                 return -ENOMEM;
3504         }
3505
3506         init_completion(&mdsc->safe_umount_waiters);
3507         init_waitqueue_head(&mdsc->session_close_wq);
3508         INIT_LIST_HEAD(&mdsc->waiting_for_map);
3509         mdsc->sessions = NULL;
3510         atomic_set(&mdsc->num_sessions, 0);
3511         mdsc->max_sessions = 0;
3512         mdsc->stopping = 0;
3513         mdsc->last_snap_seq = 0;
3514         init_rwsem(&mdsc->snap_rwsem);
3515         mdsc->snap_realms = RB_ROOT;
3516         INIT_LIST_HEAD(&mdsc->snap_empty);
3517         spin_lock_init(&mdsc->snap_empty_lock);
3518         mdsc->last_tid = 0;
3519         mdsc->oldest_tid = 0;
3520         mdsc->request_tree = RB_ROOT;
3521         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3522         mdsc->last_renew_caps = jiffies;
3523         INIT_LIST_HEAD(&mdsc->cap_delay_list);
3524         spin_lock_init(&mdsc->cap_delay_lock);
3525         INIT_LIST_HEAD(&mdsc->snap_flush_list);
3526         spin_lock_init(&mdsc->snap_flush_lock);
3527         mdsc->last_cap_flush_tid = 1;
3528         INIT_LIST_HEAD(&mdsc->cap_flush_list);
3529         INIT_LIST_HEAD(&mdsc->cap_dirty);
3530         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3531         mdsc->num_cap_flushing = 0;
3532         spin_lock_init(&mdsc->cap_dirty_lock);
3533         init_waitqueue_head(&mdsc->cap_flushing_wq);
3534         spin_lock_init(&mdsc->dentry_lru_lock);
3535         INIT_LIST_HEAD(&mdsc->dentry_lru);
3536
3537         ceph_caps_init(mdsc);
3538         ceph_adjust_min_caps(mdsc, fsc->min_caps);
3539
3540         init_rwsem(&mdsc->pool_perm_rwsem);
3541         mdsc->pool_perm_tree = RB_ROOT;
3542
3543         strncpy(mdsc->nodename, utsname()->nodename,
3544                 sizeof(mdsc->nodename) - 1);
3545         return 0;
3546 }
3547
3548 /*
3549  * Wait for safe replies on open mds requests.  If we time out, drop
3550  * all requests from the tree to avoid dangling dentry refs.
3551  */
3552 static void wait_requests(struct ceph_mds_client *mdsc)
3553 {
3554         struct ceph_options *opts = mdsc->fsc->client->options;
3555         struct ceph_mds_request *req;
3556
3557         mutex_lock(&mdsc->mutex);
3558         if (__get_oldest_req(mdsc)) {
3559                 mutex_unlock(&mdsc->mutex);
3560
3561                 dout("wait_requests waiting for requests\n");
3562                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3563                                     ceph_timeout_jiffies(opts->mount_timeout));
3564
3565                 /* tear down remaining requests */
3566                 mutex_lock(&mdsc->mutex);
3567                 while ((req = __get_oldest_req(mdsc))) {
3568                         dout("wait_requests timed out on tid %llu\n",
3569                              req->r_tid);
3570                         __unregister_request(mdsc, req);
3571                 }
3572         }
3573         mutex_unlock(&mdsc->mutex);
3574         dout("wait_requests done\n");
3575 }
3576
3577 /*
3578  * called before mount is ro, and before dentries are torn down.
3579  * (hmm, does this still race with new lookups?)
3580  */
3581 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3582 {
3583         dout("pre_umount\n");
3584         mdsc->stopping = 1;
3585
3586         drop_leases(mdsc);
3587         ceph_flush_dirty_caps(mdsc);
3588         wait_requests(mdsc);
3589
3590         /*
3591          * wait for reply handlers to drop their request refs and
3592          * their inode/dcache refs
3593          */
3594         ceph_msgr_flush();
3595 }
3596
3597 /*
3598  * wait for all write mds requests to flush.
3599  */
3600 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3601 {
3602         struct ceph_mds_request *req = NULL, *nextreq;
3603         struct rb_node *n;
3604
3605         mutex_lock(&mdsc->mutex);
3606         dout("wait_unsafe_requests want %lld\n", want_tid);
3607 restart:
3608         req = __get_oldest_req(mdsc);
3609         while (req && req->r_tid <= want_tid) {
3610                 /* find next request */
3611                 n = rb_next(&req->r_node);
3612                 if (n)
3613                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3614                 else
3615                         nextreq = NULL;
3616                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3617                     (req->r_op & CEPH_MDS_OP_WRITE)) {
3618                         /* write op */
3619                         ceph_mdsc_get_request(req);
3620                         if (nextreq)
3621                                 ceph_mdsc_get_request(nextreq);
3622                         mutex_unlock(&mdsc->mutex);
3623                         dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
3624                              req->r_tid, want_tid);
3625                         wait_for_completion(&req->r_safe_completion);
3626                         mutex_lock(&mdsc->mutex);
3627                         ceph_mdsc_put_request(req);
3628                         if (!nextreq)
3629                                 break;  /* next dne before, so we're done! */
3630                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
3631                                 /* next request was removed from tree */
3632                                 ceph_mdsc_put_request(nextreq);
3633                                 goto restart;
3634                         }
3635                         ceph_mdsc_put_request(nextreq);  /* won't go away */
3636                 }
3637                 req = nextreq;
3638         }
3639         mutex_unlock(&mdsc->mutex);
3640         dout("wait_unsafe_requests done\n");
3641 }
3642
3643 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3644 {
3645         u64 want_tid, want_flush;
3646
3647         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3648                 return;
3649
3650         dout("sync\n");
3651         mutex_lock(&mdsc->mutex);
3652         want_tid = mdsc->last_tid;
3653         mutex_unlock(&mdsc->mutex);
3654
3655         ceph_flush_dirty_caps(mdsc);
3656         spin_lock(&mdsc->cap_dirty_lock);
3657         want_flush = mdsc->last_cap_flush_tid;
3658         if (!list_empty(&mdsc->cap_flush_list)) {
3659                 struct ceph_cap_flush *cf =
3660                         list_last_entry(&mdsc->cap_flush_list,
3661                                         struct ceph_cap_flush, g_list);
3662                 cf->wake = true;
3663         }
3664         spin_unlock(&mdsc->cap_dirty_lock);
3665
3666         dout("sync want tid %lld flush_seq %lld\n",
3667              want_tid, want_flush);
3668
3669         wait_unsafe_requests(mdsc, want_tid);
3670         wait_caps_flush(mdsc, want_flush);
3671 }
3672
3673 /*
3674  * true if all sessions are closed, or we force unmount
3675  */
3676 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
3677 {
3678         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3679                 return true;
3680         return atomic_read(&mdsc->num_sessions) <= skipped;
3681 }
3682
3683 /*
3684  * called after sb is ro.
3685  */
3686 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3687 {
3688         struct ceph_options *opts = mdsc->fsc->client->options;
3689         struct ceph_mds_session *session;
3690         int i;
3691         int skipped = 0;
3692
3693         dout("close_sessions\n");
3694
3695         /* close sessions */
3696         mutex_lock(&mdsc->mutex);
3697         for (i = 0; i < mdsc->max_sessions; i++) {
3698                 session = __ceph_lookup_mds_session(mdsc, i);
3699                 if (!session)
3700                         continue;
3701                 mutex_unlock(&mdsc->mutex);
3702                 mutex_lock(&session->s_mutex);
3703                 if (__close_session(mdsc, session) <= 0)
3704                         skipped++;
3705                 mutex_unlock(&session->s_mutex);
3706                 ceph_put_mds_session(session);
3707                 mutex_lock(&mdsc->mutex);
3708         }
3709         mutex_unlock(&mdsc->mutex);
3710
3711         dout("waiting for sessions to close\n");
3712         wait_event_timeout(mdsc->session_close_wq,
3713                            done_closing_sessions(mdsc, skipped),
3714                            ceph_timeout_jiffies(opts->mount_timeout));
3715
3716         /* tear down remaining sessions */
3717         mutex_lock(&mdsc->mutex);
3718         for (i = 0; i < mdsc->max_sessions; i++) {
3719                 if (mdsc->sessions[i]) {
3720                         session = get_session(mdsc->sessions[i]);
3721                         __unregister_session(mdsc, session);
3722                         mutex_unlock(&mdsc->mutex);
3723                         mutex_lock(&session->s_mutex);
3724                         remove_session_caps(session);
3725                         mutex_unlock(&session->s_mutex);
3726                         ceph_put_mds_session(session);
3727                         mutex_lock(&mdsc->mutex);
3728                 }
3729         }
3730         WARN_ON(!list_empty(&mdsc->cap_delay_list));
3731         mutex_unlock(&mdsc->mutex);
3732
3733         ceph_cleanup_empty_realms(mdsc);
3734
3735         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3736
3737         dout("stopped\n");
3738 }
3739
3740 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3741 {
3742         struct ceph_mds_session *session;
3743         int mds;
3744
3745         dout("force umount\n");
3746
3747         mutex_lock(&mdsc->mutex);
3748         for (mds = 0; mds < mdsc->max_sessions; mds++) {
3749                 session = __ceph_lookup_mds_session(mdsc, mds);
3750                 if (!session)
3751                         continue;
3752                 mutex_unlock(&mdsc->mutex);
3753                 mutex_lock(&session->s_mutex);
3754                 __close_session(mdsc, session);
3755                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3756                         cleanup_session_requests(mdsc, session);
3757                         remove_session_caps(session);
3758                 }
3759                 mutex_unlock(&session->s_mutex);
3760                 ceph_put_mds_session(session);
3761                 mutex_lock(&mdsc->mutex);
3762                 kick_requests(mdsc, mds);
3763         }
3764         __wake_requests(mdsc, &mdsc->waiting_for_map);
3765         mutex_unlock(&mdsc->mutex);
3766 }
3767
3768 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3769 {
3770         dout("stop\n");
3771         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3772         if (mdsc->mdsmap)
3773                 ceph_mdsmap_destroy(mdsc->mdsmap);
3774         kfree(mdsc->sessions);
3775         ceph_caps_finalize(mdsc);
3776         ceph_pool_perm_destroy(mdsc);
3777 }
3778
3779 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3780 {
3781         struct ceph_mds_client *mdsc = fsc->mdsc;
3782         dout("mdsc_destroy %p\n", mdsc);
3783
3784         /* flush out any connection work with references to us */
3785         ceph_msgr_flush();
3786
3787         ceph_mdsc_stop(mdsc);
3788
3789         fsc->mdsc = NULL;
3790         kfree(mdsc);
3791         dout("mdsc_destroy %p done\n", mdsc);
3792 }
3793
3794 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3795 {
3796         struct ceph_fs_client *fsc = mdsc->fsc;
3797         const char *mds_namespace = fsc->mount_options->mds_namespace;
3798         void *p = msg->front.iov_base;
3799         void *end = p + msg->front.iov_len;
3800         u32 epoch;
3801         u32 map_len;
3802         u32 num_fs;
3803         u32 mount_fscid = (u32)-1;
3804         u8 struct_v, struct_cv;
3805         int err = -EINVAL;
3806
3807         ceph_decode_need(&p, end, sizeof(u32), bad);
3808         epoch = ceph_decode_32(&p);
3809
3810         dout("handle_fsmap epoch %u\n", epoch);
3811
3812         ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3813         struct_v = ceph_decode_8(&p);
3814         struct_cv = ceph_decode_8(&p);
3815         map_len = ceph_decode_32(&p);
3816
3817         ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
3818         p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
3819
3820         num_fs = ceph_decode_32(&p);
3821         while (num_fs-- > 0) {
3822                 void *info_p, *info_end;
3823                 u32 info_len;
3824                 u8 info_v, info_cv;
3825                 u32 fscid, namelen;
3826
3827                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3828                 info_v = ceph_decode_8(&p);
3829                 info_cv = ceph_decode_8(&p);
3830                 info_len = ceph_decode_32(&p);
3831                 ceph_decode_need(&p, end, info_len, bad);
3832                 info_p = p;
3833                 info_end = p + info_len;
3834                 p = info_end;
3835
3836                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
3837                 fscid = ceph_decode_32(&info_p);
3838                 namelen = ceph_decode_32(&info_p);
3839                 ceph_decode_need(&info_p, info_end, namelen, bad);
3840
3841                 if (mds_namespace &&
3842                     strlen(mds_namespace) == namelen &&
3843                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
3844                         mount_fscid = fscid;
3845                         break;
3846                 }
3847         }
3848
3849         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
3850         if (mount_fscid != (u32)-1) {
3851                 fsc->client->monc.fs_cluster_id = mount_fscid;
3852                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
3853                                    0, true);
3854                 ceph_monc_renew_subs(&fsc->client->monc);
3855         } else {
3856                 err = -ENOENT;
3857                 goto err_out;
3858         }
3859         return;
3860 bad:
3861         pr_err("error decoding fsmap\n");
3862 err_out:
3863         mutex_lock(&mdsc->mutex);
3864         mdsc->mdsmap_err = -ENOENT;
3865         __wake_requests(mdsc, &mdsc->waiting_for_map);
3866         mutex_unlock(&mdsc->mutex);
3867         return;
3868 }
3869
3870 /*
3871  * handle mds map update.
3872  */
3873 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3874 {
3875         u32 epoch;
3876         u32 maplen;
3877         void *p = msg->front.iov_base;
3878         void *end = p + msg->front.iov_len;
3879         struct ceph_mdsmap *newmap, *oldmap;
3880         struct ceph_fsid fsid;
3881         int err = -EINVAL;
3882
3883         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3884         ceph_decode_copy(&p, &fsid, sizeof(fsid));
3885         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3886                 return;
3887         epoch = ceph_decode_32(&p);
3888         maplen = ceph_decode_32(&p);
3889         dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3890
3891         /* do we need it? */
3892         mutex_lock(&mdsc->mutex);
3893         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3894                 dout("handle_map epoch %u <= our %u\n",
3895                      epoch, mdsc->mdsmap->m_epoch);
3896                 mutex_unlock(&mdsc->mutex);
3897                 return;
3898         }
3899
3900         newmap = ceph_mdsmap_decode(&p, end);
3901         if (IS_ERR(newmap)) {
3902                 err = PTR_ERR(newmap);
3903                 goto bad_unlock;
3904         }
3905
3906         /* swap into place */
3907         if (mdsc->mdsmap) {
3908                 oldmap = mdsc->mdsmap;
3909                 mdsc->mdsmap = newmap;
3910                 check_new_map(mdsc, newmap, oldmap);
3911                 ceph_mdsmap_destroy(oldmap);
3912         } else {
3913                 mdsc->mdsmap = newmap;  /* first mds map */
3914         }
3915         mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3916
3917         __wake_requests(mdsc, &mdsc->waiting_for_map);
3918         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
3919                           mdsc->mdsmap->m_epoch);
3920
3921         mutex_unlock(&mdsc->mutex);
3922         schedule_delayed(mdsc);
3923         return;
3924
3925 bad_unlock:
3926         mutex_unlock(&mdsc->mutex);
3927 bad:
3928         pr_err("error decoding mdsmap %d\n", err);
3929         return;
3930 }
3931
3932 static struct ceph_connection *con_get(struct ceph_connection *con)
3933 {
3934         struct ceph_mds_session *s = con->private;
3935
3936         if (get_session(s)) {
3937                 dout("mdsc con_get %p ok (%d)\n", s, refcount_read(&s->s_ref));
3938                 return con;
3939         }
3940         dout("mdsc con_get %p FAIL\n", s);
3941         return NULL;
3942 }
3943
3944 static void con_put(struct ceph_connection *con)
3945 {
3946         struct ceph_mds_session *s = con->private;
3947
3948         dout("mdsc con_put %p (%d)\n", s, refcount_read(&s->s_ref) - 1);
3949         ceph_put_mds_session(s);
3950 }
3951
3952 /*
3953  * if the client is unresponsive for long enough, the mds will kill
3954  * the session entirely.
3955  */
3956 static void peer_reset(struct ceph_connection *con)
3957 {
3958         struct ceph_mds_session *s = con->private;
3959         struct ceph_mds_client *mdsc = s->s_mdsc;
3960
3961         pr_warn("mds%d closed our session\n", s->s_mds);
3962         send_mds_reconnect(mdsc, s);
3963 }
3964
3965 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3966 {
3967         struct ceph_mds_session *s = con->private;
3968         struct ceph_mds_client *mdsc = s->s_mdsc;
3969         int type = le16_to_cpu(msg->hdr.type);
3970
3971         mutex_lock(&mdsc->mutex);
3972         if (__verify_registered_session(mdsc, s) < 0) {
3973                 mutex_unlock(&mdsc->mutex);
3974                 goto out;
3975         }
3976         mutex_unlock(&mdsc->mutex);
3977
3978         switch (type) {
3979         case CEPH_MSG_MDS_MAP:
3980                 ceph_mdsc_handle_mdsmap(mdsc, msg);
3981                 break;
3982         case CEPH_MSG_FS_MAP_USER:
3983                 ceph_mdsc_handle_fsmap(mdsc, msg);
3984                 break;
3985         case CEPH_MSG_CLIENT_SESSION:
3986                 handle_session(s, msg);
3987                 break;
3988         case CEPH_MSG_CLIENT_REPLY:
3989                 handle_reply(s, msg);
3990                 break;
3991         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3992                 handle_forward(mdsc, s, msg);
3993                 break;
3994         case CEPH_MSG_CLIENT_CAPS:
3995                 ceph_handle_caps(s, msg);
3996                 break;
3997         case CEPH_MSG_CLIENT_SNAP:
3998                 ceph_handle_snap(mdsc, s, msg);
3999                 break;
4000         case CEPH_MSG_CLIENT_LEASE:
4001                 handle_lease(mdsc, s, msg);
4002                 break;
4003
4004         default:
4005                 pr_err("received unknown message type %d %s\n", type,
4006                        ceph_msg_type_name(type));
4007         }
4008 out:
4009         ceph_msg_put(msg);
4010 }
4011
4012 /*
4013  * authentication
4014  */
4015
4016 /*
4017  * Note: returned pointer is the address of a structure that's
4018  * managed separately.  Caller must *not* attempt to free it.
4019  */
4020 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4021                                         int *proto, int force_new)
4022 {
4023         struct ceph_mds_session *s = con->private;
4024         struct ceph_mds_client *mdsc = s->s_mdsc;
4025         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4026         struct ceph_auth_handshake *auth = &s->s_auth;
4027
4028         if (force_new && auth->authorizer) {
4029                 ceph_auth_destroy_authorizer(auth->authorizer);
4030                 auth->authorizer = NULL;
4031         }
4032         if (!auth->authorizer) {
4033                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4034                                                       auth);
4035                 if (ret)
4036                         return ERR_PTR(ret);
4037         } else {
4038                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4039                                                       auth);
4040                 if (ret)
4041                         return ERR_PTR(ret);
4042         }
4043         *proto = ac->protocol;
4044
4045         return auth;
4046 }
4047
4048
4049 static int verify_authorizer_reply(struct ceph_connection *con)
4050 {
4051         struct ceph_mds_session *s = con->private;
4052         struct ceph_mds_client *mdsc = s->s_mdsc;
4053         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4054
4055         return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4056 }
4057
4058 static int invalidate_authorizer(struct ceph_connection *con)
4059 {
4060         struct ceph_mds_session *s = con->private;
4061         struct ceph_mds_client *mdsc = s->s_mdsc;
4062         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4063
4064         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4065
4066         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4067 }
4068
4069 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4070                                 struct ceph_msg_header *hdr, int *skip)
4071 {
4072         struct ceph_msg *msg;
4073         int type = (int) le16_to_cpu(hdr->type);
4074         int front_len = (int) le32_to_cpu(hdr->front_len);
4075
4076         if (con->in_msg)
4077                 return con->in_msg;
4078
4079         *skip = 0;
4080         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4081         if (!msg) {
4082                 pr_err("unable to allocate msg type %d len %d\n",
4083                        type, front_len);
4084                 return NULL;
4085         }
4086
4087         return msg;
4088 }
4089
4090 static int mds_sign_message(struct ceph_msg *msg)
4091 {
4092        struct ceph_mds_session *s = msg->con->private;
4093        struct ceph_auth_handshake *auth = &s->s_auth;
4094
4095        return ceph_auth_sign_message(auth, msg);
4096 }
4097
4098 static int mds_check_message_signature(struct ceph_msg *msg)
4099 {
4100        struct ceph_mds_session *s = msg->con->private;
4101        struct ceph_auth_handshake *auth = &s->s_auth;
4102
4103        return ceph_auth_check_message_signature(auth, msg);
4104 }
4105
4106 static const struct ceph_connection_operations mds_con_ops = {
4107         .get = con_get,
4108         .put = con_put,
4109         .dispatch = dispatch,
4110         .get_authorizer = get_authorizer,
4111         .verify_authorizer_reply = verify_authorizer_reply,
4112         .invalidate_authorizer = invalidate_authorizer,
4113         .peer_reset = peer_reset,
4114         .alloc_msg = mds_alloc_msg,
4115         .sign_message = mds_sign_message,
4116         .check_message_signature = mds_check_message_signature,
4117 };
4118
4119 /* eof */