fs/ceph/file.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/ceph/ceph_debug.h>
   3 #include <linux/ceph/striper.h>
   4
   5 #include <linux/module.h>
   6 #include <linux/sched.h>
   7 #include <linux/slab.h>
   8 #include <linux/file.h>
   9 #include <linux/mount.h>
  10 #include <linux/namei.h>
  11 #include <linux/writeback.h>
  12 #include <linux/falloc.h>
  13 #include <linux/iversion.h>
  14 #include <linux/ktime.h>
  15 #include <linux/splice.h>
  16
  17 #include "super.h"
  18 #include "mds_client.h"
  19 #include "cache.h"
  20 #include "io.h"
  21 #include "metric.h"
  22
  23 static __le32 ceph_flags_sys2wire(struct ceph_mds_client *mdsc, u32 flags)
  24 {
  25         struct ceph_client *cl = mdsc->fsc->client;
  26         u32 wire_flags = 0;
  27
  28         switch (flags & O_ACCMODE) {
  29         case O_RDONLY:
  30                 wire_flags |= CEPH_O_RDONLY;
  31                 break;
  32         case O_WRONLY:
  33                 wire_flags |= CEPH_O_WRONLY;
  34                 break;
  35         case O_RDWR:
  36                 wire_flags |= CEPH_O_RDWR;
  37                 break;
  38         }
  39
  40         flags &= ~O_ACCMODE;
  41
  42 #define ceph_sys2wire(a) if (flags & a) { wire_flags |= CEPH_##a; flags &= ~a; }
  43
  44         ceph_sys2wire(O_CREAT);
  45         ceph_sys2wire(O_EXCL);
  46         ceph_sys2wire(O_TRUNC);
  47         ceph_sys2wire(O_DIRECTORY);
  48         ceph_sys2wire(O_NOFOLLOW);
  49
  50 #undef ceph_sys2wire
  51
  52         if (flags)
  53                 doutc(cl, "unused open flags: %x\n", flags);
  54
  55         return cpu_to_le32(wire_flags);
  56 }
  57
  58 /*
  59  * Ceph file operations
  60  *
  61  * Implement basic open/close functionality, and implement
  62  * read/write.
  63  *
  64  * We implement three modes of file I/O:
  65  *  - buffered uses the generic_file_aio_{read,write} helpers
  66  *
  67  *  - synchronous is used when there is multi-client read/write
  68  *    sharing, avoids the page cache, and synchronously waits for an
  69  *    ack from the OSD.
  70  *
  71  *  - direct io takes the variant of the sync path that references
  72  *    user pages directly.
  73  *
  74  * fsync() flushes and waits on dirty pages, but just queues metadata
  75  * for writeback: since the MDS can recover size and mtime there is no
  76  * need to wait for MDS acknowledgement.
  77  */
  78
  79 /*
  80  * How many pages to get in one call to iov_iter_get_pages().  This
  81  * determines the size of the on-stack array used as a buffer.
  82  */
  83 #define ITER_GET_BVECS_PAGES    64
  84
  85 static ssize_t __iter_get_bvecs(struct iov_iter *iter, size_t maxsize,
  86                                 struct bio_vec *bvecs)
  87 {
  88         size_t size = 0;
  89         int bvec_idx = 0;
  90
  91         if (maxsize > iov_iter_count(iter))
  92                 maxsize = iov_iter_count(iter);
  93
  94         while (size < maxsize) {
  95                 struct page *pages[ITER_GET_BVECS_PAGES];
  96                 ssize_t bytes;
  97                 size_t start;
  98                 int idx = 0;
  99
 100                 bytes = iov_iter_get_pages2(iter, pages, maxsize - size,
 101                                            ITER_GET_BVECS_PAGES, &start);
 102                 if (bytes < 0)
 103                         return size ?: bytes;
 104
 105                 size += bytes;
 106
 107                 for ( ; bytes; idx++, bvec_idx++) {
 108                         int len = min_t(int, bytes, PAGE_SIZE - start);
 109
 110                         bvec_set_page(&bvecs[bvec_idx], pages[idx], len, start);
 111                         bytes -= len;
 112                         start = 0;
 113                 }
 114         }
 115
 116         return size;
 117 }
 118
 119 /*
 120  * iov_iter_get_pages() only considers one iov_iter segment, no matter
 121  * what maxsize or maxpages are given.  For ITER_BVEC that is a single
 122  * page.
 123  *
 124  * Attempt to get up to @maxsize bytes worth of pages from @iter.
 125  * Return the number of bytes in the created bio_vec array, or an error.
 126  */
 127 static ssize_t iter_get_bvecs_alloc(struct iov_iter *iter, size_t maxsize,
 128                                     struct bio_vec **bvecs, int *num_bvecs)
 129 {
 130         struct bio_vec *bv;
 131         size_t orig_count = iov_iter_count(iter);
 132         ssize_t bytes;
 133         int npages;
 134
 135         iov_iter_truncate(iter, maxsize);
 136         npages = iov_iter_npages(iter, INT_MAX);
 137         iov_iter_reexpand(iter, orig_count);
 138
 139         /*
 140          * __iter_get_bvecs() may populate only part of the array -- zero it
 141          * out.
 142          */
 143         bv = kvmalloc_array(npages, sizeof(*bv), GFP_KERNEL | __GFP_ZERO);
 144         if (!bv)
 145                 return -ENOMEM;
 146
 147         bytes = __iter_get_bvecs(iter, maxsize, bv);
 148         if (bytes < 0) {
 149                 /*
 150                  * No pages were pinned -- just free the array.
 151                  */
 152                 kvfree(bv);
 153                 return bytes;
 154         }
 155
 156         *bvecs = bv;
 157         *num_bvecs = npages;
 158         return bytes;
 159 }
 160
 161 static void put_bvecs(struct bio_vec *bvecs, int num_bvecs, bool should_dirty)
 162 {
 163         int i;
 164
 165         for (i = 0; i < num_bvecs; i++) {
 166                 if (bvecs[i].bv_page) {
 167                         if (should_dirty)
 168                                 set_page_dirty_lock(bvecs[i].bv_page);
 169                         put_page(bvecs[i].bv_page);
 170                 }
 171         }
 172         kvfree(bvecs);
 173 }
 174
 175 /*
 176  * Prepare an open request.  Preallocate ceph_cap to avoid an
 177  * inopportune ENOMEM later.
 178  */
 179 static struct ceph_mds_request *
 180 prepare_open_request(struct super_block *sb, int flags, int create_mode)
 181 {
 182         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(sb);
 183         struct ceph_mds_request *req;
 184         int want_auth = USE_ANY_MDS;
 185         int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
 186
 187         if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
 188                 want_auth = USE_AUTH_MDS;
 189
 190         req = ceph_mdsc_create_request(mdsc, op, want_auth);
 191         if (IS_ERR(req))
 192                 goto out;
 193         req->r_fmode = ceph_flags_to_mode(flags);
 194         req->r_args.open.flags = ceph_flags_sys2wire(mdsc, flags);
 195         req->r_args.open.mode = cpu_to_le32(create_mode);
 196 out:
 197         return req;
 198 }
 199
 200 static int ceph_init_file_info(struct inode *inode, struct file *file,
 201                                         int fmode, bool isdir)
 202 {
 203         struct ceph_inode_info *ci = ceph_inode(inode);
 204         struct ceph_mount_options *opt =
 205                 ceph_inode_to_fs_client(&ci->netfs.inode)->mount_options;
 206         struct ceph_client *cl = ceph_inode_to_client(inode);
 207         struct ceph_file_info *fi;
 208         int ret;
 209
 210         doutc(cl, "%p %llx.%llx %p 0%o (%s)\n", inode, ceph_vinop(inode),
 211               file, inode->i_mode, isdir ? "dir" : "regular");
 212         BUG_ON(inode->i_fop->release != ceph_release);
 213
 214         if (isdir) {
 215                 struct ceph_dir_file_info *dfi =
 216                         kmem_cache_zalloc(ceph_dir_file_cachep, GFP_KERNEL);
 217                 if (!dfi)
 218                         return -ENOMEM;
 219
 220                 file->private_data = dfi;
 221                 fi = &dfi->file_info;
 222                 dfi->next_offset = 2;
 223                 dfi->readdir_cache_idx = -1;
 224         } else {
 225                 fi = kmem_cache_zalloc(ceph_file_cachep, GFP_KERNEL);
 226                 if (!fi)
 227                         return -ENOMEM;
 228
 229                 if (opt->flags & CEPH_MOUNT_OPT_NOPAGECACHE)
 230                         fi->flags |= CEPH_F_SYNC;
 231
 232                 file->private_data = fi;
 233         }
 234
 235         ceph_get_fmode(ci, fmode, 1);
 236         fi->fmode = fmode;
 237
 238         spin_lock_init(&fi->rw_contexts_lock);
 239         INIT_LIST_HEAD(&fi->rw_contexts);
 240         fi->filp_gen = READ_ONCE(ceph_inode_to_fs_client(inode)->filp_gen);
 241
 242         if ((file->f_mode & FMODE_WRITE) && ceph_has_inline_data(ci)) {
 243                 ret = ceph_uninline_data(file);
 244                 if (ret < 0)
 245                         goto error;
 246         }
 247
 248         return 0;
 249
 250 error:
 251         ceph_fscache_unuse_cookie(inode, file->f_mode & FMODE_WRITE);
 252         ceph_put_fmode(ci, fi->fmode, 1);
 253         kmem_cache_free(ceph_file_cachep, fi);
 254         /* wake up anyone waiting for caps on this inode */
 255         wake_up_all(&ci->i_cap_wq);
 256         return ret;
 257 }
 258
 259 /*
 260  * initialize private struct file data.
 261  * if we fail, clean up by dropping fmode reference on the ceph_inode
 262  */
 263 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
 264 {
 265         struct ceph_client *cl = ceph_inode_to_client(inode);
 266         int ret = 0;
 267
 268         switch (inode->i_mode & S_IFMT) {
 269         case S_IFREG:
 270                 ceph_fscache_use_cookie(inode, file->f_mode & FMODE_WRITE);
 271                 fallthrough;
 272         case S_IFDIR:
 273                 ret = ceph_init_file_info(inode, file, fmode,
 274                                                 S_ISDIR(inode->i_mode));
 275                 break;
 276
 277         case S_IFLNK:
 278                 doutc(cl, "%p %llx.%llx %p 0%o (symlink)\n", inode,
 279                       ceph_vinop(inode), file, inode->i_mode);
 280                 break;
 281
 282         default:
 283                 doutc(cl, "%p %llx.%llx %p 0%o (special)\n", inode,
 284                       ceph_vinop(inode), file, inode->i_mode);
 285                 /*
 286                  * we need to drop the open ref now, since we don't
 287                  * have .release set to ceph_release.
 288                  */
 289                 BUG_ON(inode->i_fop->release == ceph_release);
 290
 291                 /* call the proper open fop */
 292                 ret = inode->i_fop->open(inode, file);
 293         }
 294         return ret;
 295 }
 296
 297 /*
 298  * try renew caps after session gets killed.
 299  */
 300 int ceph_renew_caps(struct inode *inode, int fmode)
 301 {
 302         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
 303         struct ceph_client *cl = mdsc->fsc->client;
 304         struct ceph_inode_info *ci = ceph_inode(inode);
 305         struct ceph_mds_request *req;
 306         int err, flags, wanted;
 307
 308         spin_lock(&ci->i_ceph_lock);
 309         __ceph_touch_fmode(ci, mdsc, fmode);
 310         wanted = __ceph_caps_file_wanted(ci);
 311         if (__ceph_is_any_real_caps(ci) &&
 312             (!(wanted & CEPH_CAP_ANY_WR) || ci->i_auth_cap)) {
 313                 int issued = __ceph_caps_issued(ci, NULL);
 314                 spin_unlock(&ci->i_ceph_lock);
 315                 doutc(cl, "%p %llx.%llx want %s issued %s updating mds_wanted\n",
 316                       inode, ceph_vinop(inode), ceph_cap_string(wanted),
 317                       ceph_cap_string(issued));
 318                 ceph_check_caps(ci, 0);
 319                 return 0;
 320         }
 321         spin_unlock(&ci->i_ceph_lock);
 322
 323         flags = 0;
 324         if ((wanted & CEPH_CAP_FILE_RD) && (wanted & CEPH_CAP_FILE_WR))
 325                 flags = O_RDWR;
 326         else if (wanted & CEPH_CAP_FILE_RD)
 327                 flags = O_RDONLY;
 328         else if (wanted & CEPH_CAP_FILE_WR)
 329                 flags = O_WRONLY;
 330 #ifdef O_LAZY
 331         if (wanted & CEPH_CAP_FILE_LAZYIO)
 332                 flags |= O_LAZY;
 333 #endif
 334
 335         req = prepare_open_request(inode->i_sb, flags, 0);
 336         if (IS_ERR(req)) {
 337                 err = PTR_ERR(req);
 338                 goto out;
 339         }
 340
 341         req->r_inode = inode;
 342         ihold(inode);
 343         req->r_num_caps = 1;
 344
 345         err = ceph_mdsc_do_request(mdsc, NULL, req);
 346         ceph_mdsc_put_request(req);
 347 out:
 348         doutc(cl, "%p %llx.%llx open result=%d\n", inode, ceph_vinop(inode),
 349               err);
 350         return err < 0 ? err : 0;
 351 }
 352
 353 /*
 354  * If we already have the requisite capabilities, we can satisfy
 355  * the open request locally (no need to request new caps from the
 356  * MDS).  We do, however, need to inform the MDS (asynchronously)
 357  * if our wanted caps set expands.
 358  */
 359 int ceph_open(struct inode *inode, struct file *file)
 360 {
 361         struct ceph_inode_info *ci = ceph_inode(inode);
 362         struct ceph_fs_client *fsc = ceph_sb_to_fs_client(inode->i_sb);
 363         struct ceph_client *cl = fsc->client;
 364         struct ceph_mds_client *mdsc = fsc->mdsc;
 365         struct ceph_mds_request *req;
 366         struct ceph_file_info *fi = file->private_data;
 367         int err;
 368         int flags, fmode, wanted;
 369         struct dentry *dentry;
 370         char *path;
 371         int pathlen;
 372         u64 pathbase;
 373         bool do_sync = false;
 374         int mask = MAY_READ;
 375
 376         if (fi) {
 377                 doutc(cl, "file %p is already opened\n", file);
 378                 return 0;
 379         }
 380
 381         /* filter out O_CREAT|O_EXCL; vfs did that already.  yuck. */
 382         flags = file->f_flags & ~(O_CREAT|O_EXCL);
 383         if (S_ISDIR(inode->i_mode)) {
 384                 flags = O_DIRECTORY;  /* mds likes to know */
 385         } else if (S_ISREG(inode->i_mode)) {
 386                 err = fscrypt_file_open(inode, file);
 387                 if (err)
 388                         return err;
 389         }
 390
 391         doutc(cl, "%p %llx.%llx file %p flags %d (%d)\n", inode,
 392               ceph_vinop(inode), file, flags, file->f_flags);
 393         fmode = ceph_flags_to_mode(flags);
 394         wanted = ceph_caps_for_mode(fmode);
 395
 396         if (fmode & CEPH_FILE_MODE_WR)
 397                 mask |= MAY_WRITE;
 398         dentry = d_find_alias(inode);
 399         if (!dentry) {
 400                 do_sync = true;
 401         } else {
 402                 path = ceph_mdsc_build_path(mdsc, dentry, &pathlen, &pathbase, 0);
 403                 if (IS_ERR(path)) {
 404                         do_sync = true;
 405                         err = 0;
 406                 } else {
 407                         err = ceph_mds_check_access(mdsc, path, mask);
 408                 }
 409                 ceph_mdsc_free_path(path, pathlen);
 410                 dput(dentry);
 411
 412                 /* For none EACCES cases will let the MDS do the mds auth check */
 413                 if (err == -EACCES) {
 414                         return err;
 415                 } else if (err < 0) {
 416                         do_sync = true;
 417                         err = 0;
 418                 }
 419         }
 420
 421         /* snapped files are read-only */
 422         if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
 423                 return -EROFS;
 424
 425         /* trivially open snapdir */
 426         if (ceph_snap(inode) == CEPH_SNAPDIR) {
 427                 return ceph_init_file(inode, file, fmode);
 428         }
 429
 430         /*
 431          * No need to block if we have caps on the auth MDS (for
 432          * write) or any MDS (for read).  Update wanted set
 433          * asynchronously.
 434          */
 435         spin_lock(&ci->i_ceph_lock);
 436         if (!do_sync && __ceph_is_any_real_caps(ci) &&
 437             (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
 438                 int mds_wanted = __ceph_caps_mds_wanted(ci, true);
 439                 int issued = __ceph_caps_issued(ci, NULL);
 440
 441                 doutc(cl, "open %p fmode %d want %s issued %s using existing\n",
 442                       inode, fmode, ceph_cap_string(wanted),
 443                       ceph_cap_string(issued));
 444                 __ceph_touch_fmode(ci, mdsc, fmode);
 445                 spin_unlock(&ci->i_ceph_lock);
 446
 447                 /* adjust wanted? */
 448                 if ((issued & wanted) != wanted &&
 449                     (mds_wanted & wanted) != wanted &&
 450                     ceph_snap(inode) != CEPH_SNAPDIR)
 451                         ceph_check_caps(ci, 0);
 452
 453                 return ceph_init_file(inode, file, fmode);
 454         } else if (!do_sync && ceph_snap(inode) != CEPH_NOSNAP &&
 455                    (ci->i_snap_caps & wanted) == wanted) {
 456                 __ceph_touch_fmode(ci, mdsc, fmode);
 457                 spin_unlock(&ci->i_ceph_lock);
 458                 return ceph_init_file(inode, file, fmode);
 459         }
 460
 461         spin_unlock(&ci->i_ceph_lock);
 462
 463         doutc(cl, "open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
 464         req = prepare_open_request(inode->i_sb, flags, 0);
 465         if (IS_ERR(req)) {
 466                 err = PTR_ERR(req);
 467                 goto out;
 468         }
 469         req->r_inode = inode;
 470         ihold(inode);
 471
 472         req->r_num_caps = 1;
 473         err = ceph_mdsc_do_request(mdsc, NULL, req);
 474         if (!err)
 475                 err = ceph_init_file(inode, file, req->r_fmode);
 476         ceph_mdsc_put_request(req);
 477         doutc(cl, "open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
 478 out:
 479         return err;
 480 }
 481
 482 /* Clone the layout from a synchronous create, if the dir now has Dc caps */
 483 static void
 484 cache_file_layout(struct inode *dst, struct inode *src)
 485 {
 486         struct ceph_inode_info *cdst = ceph_inode(dst);
 487         struct ceph_inode_info *csrc = ceph_inode(src);
 488
 489         spin_lock(&cdst->i_ceph_lock);
 490         if ((__ceph_caps_issued(cdst, NULL) & CEPH_CAP_DIR_CREATE) &&
 491             !ceph_file_layout_is_valid(&cdst->i_cached_layout)) {
 492                 memcpy(&cdst->i_cached_layout, &csrc->i_layout,
 493                         sizeof(cdst->i_cached_layout));
 494                 rcu_assign_pointer(cdst->i_cached_layout.pool_ns,
 495                                    ceph_try_get_string(csrc->i_layout.pool_ns));
 496         }
 497         spin_unlock(&cdst->i_ceph_lock);
 498 }
 499
 500 /*
 501  * Try to set up an async create. We need caps, a file layout, and inode number,
 502  * and either a lease on the dentry or complete dir info. If any of those
 503  * criteria are not satisfied, then return false and the caller can go
 504  * synchronous.
 505  */
 506 static int try_prep_async_create(struct inode *dir, struct dentry *dentry,
 507                                  struct ceph_file_layout *lo, u64 *pino)
 508 {
 509         struct ceph_inode_info *ci = ceph_inode(dir);
 510         struct ceph_dentry_info *di = ceph_dentry(dentry);
 511         int got = 0, want = CEPH_CAP_FILE_EXCL | CEPH_CAP_DIR_CREATE;
 512         u64 ino;
 513
 514         spin_lock(&ci->i_ceph_lock);
 515         /* No auth cap means no chance for Dc caps */
 516         if (!ci->i_auth_cap)
 517                 goto no_async;
 518
 519         /* Any delegated inos? */
 520         if (xa_empty(&ci->i_auth_cap->session->s_delegated_inos))
 521                 goto no_async;
 522
 523         if (!ceph_file_layout_is_valid(&ci->i_cached_layout))
 524                 goto no_async;
 525
 526         if ((__ceph_caps_issued(ci, NULL) & want) != want)
 527                 goto no_async;
 528
 529         if (d_in_lookup(dentry)) {
 530                 if (!__ceph_dir_is_complete(ci))
 531                         goto no_async;
 532                 spin_lock(&dentry->d_lock);
 533                 di->lease_shared_gen = atomic_read(&ci->i_shared_gen);
 534                 spin_unlock(&dentry->d_lock);
 535         } else if (atomic_read(&ci->i_shared_gen) !=
 536                    READ_ONCE(di->lease_shared_gen)) {
 537                 goto no_async;
 538         }
 539
 540         ino = ceph_get_deleg_ino(ci->i_auth_cap->session);
 541         if (!ino)
 542                 goto no_async;
 543
 544         *pino = ino;
 545         ceph_take_cap_refs(ci, want, false);
 546         memcpy(lo, &ci->i_cached_layout, sizeof(*lo));
 547         rcu_assign_pointer(lo->pool_ns,
 548                            ceph_try_get_string(ci->i_cached_layout.pool_ns));
 549         got = want;
 550 no_async:
 551         spin_unlock(&ci->i_ceph_lock);
 552         return got;
 553 }
 554
 555 static void restore_deleg_ino(struct inode *dir, u64 ino)
 556 {
 557         struct ceph_client *cl = ceph_inode_to_client(dir);
 558         struct ceph_inode_info *ci = ceph_inode(dir);
 559         struct ceph_mds_session *s = NULL;
 560
 561         spin_lock(&ci->i_ceph_lock);
 562         if (ci->i_auth_cap)
 563                 s = ceph_get_mds_session(ci->i_auth_cap->session);
 564         spin_unlock(&ci->i_ceph_lock);
 565         if (s) {
 566                 int err = ceph_restore_deleg_ino(s, ino);
 567                 if (err)
 568                         pr_warn_client(cl,
 569                                 "unable to restore delegated ino 0x%llx to session: %d\n",
 570                                 ino, err);
 571                 ceph_put_mds_session(s);
 572         }
 573 }
 574
 575 static void wake_async_create_waiters(struct inode *inode,
 576                                       struct ceph_mds_session *session)
 577 {
 578         struct ceph_inode_info *ci = ceph_inode(inode);
 579         bool check_cap = false;
 580
 581         spin_lock(&ci->i_ceph_lock);
 582         if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE) {
 583                 ci->i_ceph_flags &= ~CEPH_I_ASYNC_CREATE;
 584                 wake_up_bit(&ci->i_ceph_flags, CEPH_ASYNC_CREATE_BIT);
 585
 586                 if (ci->i_ceph_flags & CEPH_I_ASYNC_CHECK_CAPS) {
 587                         ci->i_ceph_flags &= ~CEPH_I_ASYNC_CHECK_CAPS;
 588                         check_cap = true;
 589                 }
 590         }
 591         ceph_kick_flushing_inode_caps(session, ci);
 592         spin_unlock(&ci->i_ceph_lock);
 593
 594         if (check_cap)
 595                 ceph_check_caps(ci, CHECK_CAPS_FLUSH);
 596 }
 597
 598 static void ceph_async_create_cb(struct ceph_mds_client *mdsc,
 599                                  struct ceph_mds_request *req)
 600 {
 601         struct ceph_client *cl = mdsc->fsc->client;
 602         struct dentry *dentry = req->r_dentry;
 603         struct inode *dinode = d_inode(dentry);
 604         struct inode *tinode = req->r_target_inode;
 605         int result = req->r_err ? req->r_err :
 606                         le32_to_cpu(req->r_reply_info.head->result);
 607
 608         WARN_ON_ONCE(dinode && tinode && dinode != tinode);
 609
 610         /* MDS changed -- caller must resubmit */
 611         if (result == -EJUKEBOX)
 612                 goto out;
 613
 614         mapping_set_error(req->r_parent->i_mapping, result);
 615
 616         if (result) {
 617                 int pathlen = 0;
 618                 u64 base = 0;
 619                 char *path = ceph_mdsc_build_path(mdsc, req->r_dentry, &pathlen,
 620                                                   &base, 0);
 621
 622                 pr_warn_client(cl,
 623                         "async create failure path=(%llx)%s result=%d!\n",
 624                         base, IS_ERR(path) ? "<<bad>>" : path, result);
 625                 ceph_mdsc_free_path(path, pathlen);
 626
 627                 ceph_dir_clear_complete(req->r_parent);
 628                 if (!d_unhashed(dentry))
 629                         d_drop(dentry);
 630
 631                 if (dinode) {
 632                         mapping_set_error(dinode->i_mapping, result);
 633                         ceph_inode_shutdown(dinode);
 634                         wake_async_create_waiters(dinode, req->r_session);
 635                 }
 636         }
 637
 638         if (tinode) {
 639                 u64 ino = ceph_vino(tinode).ino;
 640
 641                 if (req->r_deleg_ino != ino)
 642                         pr_warn_client(cl,
 643                                 "inode number mismatch! err=%d deleg_ino=0x%llx target=0x%llx\n",
 644                                 req->r_err, req->r_deleg_ino, ino);
 645
 646                 mapping_set_error(tinode->i_mapping, result);
 647                 wake_async_create_waiters(tinode, req->r_session);
 648         } else if (!result) {
 649                 pr_warn_client(cl, "no req->r_target_inode for 0x%llx\n",
 650                                req->r_deleg_ino);
 651         }
 652 out:
 653         ceph_mdsc_release_dir_caps(req);
 654 }
 655
 656 static int ceph_finish_async_create(struct inode *dir, struct inode *inode,
 657                                     struct dentry *dentry,
 658                                     struct file *file, umode_t mode,
 659                                     struct ceph_mds_request *req,
 660                                     struct ceph_acl_sec_ctx *as_ctx,
 661                                     struct ceph_file_layout *lo)
 662 {
 663         int ret;
 664         char xattr_buf[4];
 665         struct ceph_mds_reply_inode in = { };
 666         struct ceph_mds_reply_info_in iinfo = { .in = &in };
 667         struct ceph_inode_info *ci = ceph_inode(dir);
 668         struct ceph_dentry_info *di = ceph_dentry(dentry);
 669         struct timespec64 now;
 670         struct ceph_string *pool_ns;
 671         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 672         struct ceph_client *cl = mdsc->fsc->client;
 673         struct ceph_vino vino = { .ino = req->r_deleg_ino,
 674                                   .snap = CEPH_NOSNAP };
 675
 676         ktime_get_real_ts64(&now);
 677
 678         iinfo.inline_version = CEPH_INLINE_NONE;
 679         iinfo.change_attr = 1;
 680         ceph_encode_timespec64(&iinfo.btime, &now);
 681
 682         if (req->r_pagelist) {
 683                 iinfo.xattr_len = req->r_pagelist->length;
 684                 iinfo.xattr_data = req->r_pagelist->mapped_tail;
 685         } else {
 686                 /* fake it */
 687                 iinfo.xattr_len = ARRAY_SIZE(xattr_buf);
 688                 iinfo.xattr_data = xattr_buf;
 689                 memset(iinfo.xattr_data, 0, iinfo.xattr_len);
 690         }
 691
 692         in.ino = cpu_to_le64(vino.ino);
 693         in.snapid = cpu_to_le64(CEPH_NOSNAP);
 694         in.version = cpu_to_le64(1);    // ???
 695         in.cap.caps = in.cap.wanted = cpu_to_le32(CEPH_CAP_ALL_FILE);
 696         in.cap.cap_id = cpu_to_le64(1);
 697         in.cap.realm = cpu_to_le64(ci->i_snap_realm->ino);
 698         in.cap.flags = CEPH_CAP_FLAG_AUTH;
 699         in.ctime = in.mtime = in.atime = iinfo.btime;
 700         in.truncate_seq = cpu_to_le32(1);
 701         in.truncate_size = cpu_to_le64(-1ULL);
 702         in.xattr_version = cpu_to_le64(1);
 703         in.uid = cpu_to_le32(from_kuid(&init_user_ns,
 704                                        mapped_fsuid(req->r_mnt_idmap,
 705                                                     &init_user_ns)));
 706         if (dir->i_mode & S_ISGID) {
 707                 in.gid = cpu_to_le32(from_kgid(&init_user_ns, dir->i_gid));
 708
 709                 /* Directories always inherit the setgid bit. */
 710                 if (S_ISDIR(mode))
 711                         mode |= S_ISGID;
 712         } else {
 713                 in.gid = cpu_to_le32(from_kgid(&init_user_ns,
 714                                      mapped_fsgid(req->r_mnt_idmap,
 715                                                   &init_user_ns)));
 716         }
 717         in.mode = cpu_to_le32((u32)mode);
 718
 719         in.nlink = cpu_to_le32(1);
 720         in.max_size = cpu_to_le64(lo->stripe_unit);
 721
 722         ceph_file_layout_to_legacy(lo, &in.layout);
 723         /* lo is private, so pool_ns can't change */
 724         pool_ns = rcu_dereference_raw(lo->pool_ns);
 725         if (pool_ns) {
 726                 iinfo.pool_ns_len = pool_ns->len;
 727                 iinfo.pool_ns_data = pool_ns->str;
 728         }
 729
 730         down_read(&mdsc->snap_rwsem);
 731         ret = ceph_fill_inode(inode, NULL, &iinfo, NULL, req->r_session,
 732                               req->r_fmode, NULL);
 733         up_read(&mdsc->snap_rwsem);
 734         if (ret) {
 735                 doutc(cl, "failed to fill inode: %d\n", ret);
 736                 ceph_dir_clear_complete(dir);
 737                 if (!d_unhashed(dentry))
 738                         d_drop(dentry);
 739                 discard_new_inode(inode);
 740         } else {
 741                 struct dentry *dn;
 742
 743                 doutc(cl, "d_adding new inode 0x%llx to 0x%llx/%s\n",
 744                       vino.ino, ceph_ino(dir), dentry->d_name.name);
 745                 ceph_dir_clear_ordered(dir);
 746                 ceph_init_inode_acls(inode, as_ctx);
 747                 if (inode->i_state & I_NEW) {
 748                         /*
 749                          * If it's not I_NEW, then someone created this before
 750                          * we got here. Assume the server is aware of it at
 751                          * that point and don't worry about setting
 752                          * CEPH_I_ASYNC_CREATE.
 753                          */
 754                         ceph_inode(inode)->i_ceph_flags = CEPH_I_ASYNC_CREATE;
 755                         unlock_new_inode(inode);
 756                 }
 757                 if (d_in_lookup(dentry) || d_really_is_negative(dentry)) {
 758                         if (!d_unhashed(dentry))
 759                                 d_drop(dentry);
 760                         dn = d_splice_alias(inode, dentry);
 761                         WARN_ON_ONCE(dn && dn != dentry);
 762                 }
 763                 file->f_mode |= FMODE_CREATED;
 764                 ret = finish_open(file, dentry, ceph_open);
 765         }
 766
 767         spin_lock(&dentry->d_lock);
 768         di->flags &= ~CEPH_DENTRY_ASYNC_CREATE;
 769         wake_up_bit(&di->flags, CEPH_DENTRY_ASYNC_CREATE_BIT);
 770         spin_unlock(&dentry->d_lock);
 771
 772         return ret;
 773 }
 774
 775 /*
 776  * Do a lookup + open with a single request.  If we get a non-existent
 777  * file or symlink, return 1 so the VFS can retry.
 778  */
 779 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
 780                      struct file *file, unsigned flags, umode_t mode)
 781 {
 782         struct mnt_idmap *idmap = file_mnt_idmap(file);
 783         struct ceph_fs_client *fsc = ceph_sb_to_fs_client(dir->i_sb);
 784         struct ceph_client *cl = fsc->client;
 785         struct ceph_mds_client *mdsc = fsc->mdsc;
 786         struct ceph_mds_request *req;
 787         struct inode *new_inode = NULL;
 788         struct dentry *dn;
 789         struct ceph_acl_sec_ctx as_ctx = {};
 790         bool try_async = ceph_test_mount_opt(fsc, ASYNC_DIROPS);
 791         int mask;
 792         int err;
 793         char *path;
 794         int pathlen;
 795         u64 pathbase;
 796
 797         doutc(cl, "%p %llx.%llx dentry %p '%pd' %s flags %d mode 0%o\n",
 798               dir, ceph_vinop(dir), dentry, dentry,
 799               d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
 800
 801         if (dentry->d_name.len > NAME_MAX)
 802                 return -ENAMETOOLONG;
 803
 804         err = ceph_wait_on_conflict_unlink(dentry);
 805         if (err)
 806                 return err;
 807         /*
 808          * Do not truncate the file, since atomic_open is called before the
 809          * permission check. The caller will do the truncation afterward.
 810          */
 811         flags &= ~O_TRUNC;
 812
 813         dn = d_find_alias(dir);
 814         if (!dn) {
 815                 try_async = false;
 816         } else {
 817                 path = ceph_mdsc_build_path(mdsc, dn, &pathlen, &pathbase, 0);
 818                 if (IS_ERR(path)) {
 819                         try_async = false;
 820                         err = 0;
 821                 } else {
 822                         int fmode = ceph_flags_to_mode(flags);
 823
 824                         mask = MAY_READ;
 825                         if (fmode & CEPH_FILE_MODE_WR)
 826                                 mask |= MAY_WRITE;
 827                         err = ceph_mds_check_access(mdsc, path, mask);
 828                 }
 829                 ceph_mdsc_free_path(path, pathlen);
 830                 dput(dn);
 831
 832                 /* For none EACCES cases will let the MDS do the mds auth check */
 833                 if (err == -EACCES) {
 834                         return err;
 835                 } else if (err < 0) {
 836                         try_async = false;
 837                         err = 0;
 838                 }
 839         }
 840
 841 retry:
 842         if (flags & O_CREAT) {
 843                 if (ceph_quota_is_max_files_exceeded(dir))
 844                         return -EDQUOT;
 845
 846                 new_inode = ceph_new_inode(dir, dentry, &mode, &as_ctx);
 847                 if (IS_ERR(new_inode)) {
 848                         err = PTR_ERR(new_inode);
 849                         goto out_ctx;
 850                 }
 851                 /* Async create can't handle more than a page of xattrs */
 852                 if (as_ctx.pagelist &&
 853                     !list_is_singular(&as_ctx.pagelist->head))
 854                         try_async = false;
 855         } else if (!d_in_lookup(dentry)) {
 856                 /* If it's not being looked up, it's negative */
 857                 return -ENOENT;
 858         }
 859
 860         /* do the open */
 861         req = prepare_open_request(dir->i_sb, flags, mode);
 862         if (IS_ERR(req)) {
 863                 err = PTR_ERR(req);
 864                 goto out_ctx;
 865         }
 866         req->r_dentry = dget(dentry);
 867         req->r_num_caps = 2;
 868         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 869         if (ceph_security_xattr_wanted(dir))
 870                 mask |= CEPH_CAP_XATTR_SHARED;
 871         req->r_args.open.mask = cpu_to_le32(mask);
 872         req->r_parent = dir;
 873         if (req->r_op == CEPH_MDS_OP_CREATE)
 874                 req->r_mnt_idmap = mnt_idmap_get(idmap);
 875         ihold(dir);
 876         if (IS_ENCRYPTED(dir)) {
 877                 set_bit(CEPH_MDS_R_FSCRYPT_FILE, &req->r_req_flags);
 878                 err = fscrypt_prepare_lookup_partial(dir, dentry);
 879                 if (err < 0)
 880                         goto out_req;
 881         }
 882
 883         if (flags & O_CREAT) {
 884                 struct ceph_file_layout lo;
 885
 886                 req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL |
 887                                      CEPH_CAP_XATTR_EXCL;
 888                 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 889
 890                 ceph_as_ctx_to_req(req, &as_ctx);
 891
 892                 if (try_async && (req->r_dir_caps =
 893                                   try_prep_async_create(dir, dentry, &lo,
 894                                                         &req->r_deleg_ino))) {
 895                         struct ceph_vino vino = { .ino = req->r_deleg_ino,
 896                                                   .snap = CEPH_NOSNAP };
 897                         struct ceph_dentry_info *di = ceph_dentry(dentry);
 898
 899                         set_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags);
 900                         req->r_args.open.flags |= cpu_to_le32(CEPH_O_EXCL);
 901                         req->r_callback = ceph_async_create_cb;
 902
 903                         /* Hash inode before RPC */
 904                         new_inode = ceph_get_inode(dir->i_sb, vino, new_inode);
 905                         if (IS_ERR(new_inode)) {
 906                                 err = PTR_ERR(new_inode);
 907                                 new_inode = NULL;
 908                                 goto out_req;
 909                         }
 910                         WARN_ON_ONCE(!(new_inode->i_state & I_NEW));
 911
 912                         spin_lock(&dentry->d_lock);
 913                         di->flags |= CEPH_DENTRY_ASYNC_CREATE;
 914                         spin_unlock(&dentry->d_lock);
 915
 916                         err = ceph_mdsc_submit_request(mdsc, dir, req);
 917                         if (!err) {
 918                                 err = ceph_finish_async_create(dir, new_inode,
 919                                                                dentry, file,
 920                                                                mode, req,
 921                                                                &as_ctx, &lo);
 922                                 new_inode = NULL;
 923                         } else if (err == -EJUKEBOX) {
 924                                 restore_deleg_ino(dir, req->r_deleg_ino);
 925                                 ceph_mdsc_put_request(req);
 926                                 discard_new_inode(new_inode);
 927                                 ceph_release_acl_sec_ctx(&as_ctx);
 928                                 memset(&as_ctx, 0, sizeof(as_ctx));
 929                                 new_inode = NULL;
 930                                 try_async = false;
 931                                 ceph_put_string(rcu_dereference_raw(lo.pool_ns));
 932                                 goto retry;
 933                         }
 934                         ceph_put_string(rcu_dereference_raw(lo.pool_ns));
 935                         goto out_req;
 936                 }
 937         }
 938
 939         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 940         req->r_new_inode = new_inode;
 941         new_inode = NULL;
 942         err = ceph_mdsc_do_request(mdsc, (flags & O_CREAT) ? dir : NULL, req);
 943         if (err == -ENOENT) {
 944                 dentry = ceph_handle_snapdir(req, dentry);
 945                 if (IS_ERR(dentry)) {
 946                         err = PTR_ERR(dentry);
 947                         goto out_req;
 948                 }
 949                 err = 0;
 950         }
 951
 952         if (!err && (flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
 953                 err = ceph_handle_notrace_create(dir, dentry);
 954
 955         if (d_in_lookup(dentry)) {
 956                 dn = ceph_finish_lookup(req, dentry, err);
 957                 if (IS_ERR(dn))
 958                         err = PTR_ERR(dn);
 959         } else {
 960                 /* we were given a hashed negative dentry */
 961                 dn = NULL;
 962         }
 963         if (err)
 964                 goto out_req;
 965         if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
 966                 /* make vfs retry on splice, ENOENT, or symlink */
 967                 doutc(cl, "finish_no_open on dn %p\n", dn);
 968                 err = finish_no_open(file, dn);
 969         } else {
 970                 if (IS_ENCRYPTED(dir) &&
 971                     !fscrypt_has_permitted_context(dir, d_inode(dentry))) {
 972                         pr_warn_client(cl,
 973                                 "Inconsistent encryption context (parent %llx:%llx child %llx:%llx)\n",
 974                                 ceph_vinop(dir), ceph_vinop(d_inode(dentry)));
 975                         goto out_req;
 976                 }
 977
 978                 doutc(cl, "finish_open on dn %p\n", dn);
 979                 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
 980                         struct inode *newino = d_inode(dentry);
 981
 982                         cache_file_layout(dir, newino);
 983                         ceph_init_inode_acls(newino, &as_ctx);
 984                         file->f_mode |= FMODE_CREATED;
 985                 }
 986                 err = finish_open(file, dentry, ceph_open);
 987         }
 988 out_req:
 989         ceph_mdsc_put_request(req);
 990         iput(new_inode);
 991 out_ctx:
 992         ceph_release_acl_sec_ctx(&as_ctx);
 993         doutc(cl, "result=%d\n", err);
 994         return err;
 995 }
 996
 997 int ceph_release(struct inode *inode, struct file *file)
 998 {
 999         struct ceph_client *cl = ceph_inode_to_client(inode);
1000         struct ceph_inode_info *ci = ceph_inode(inode);
1001
1002         if (S_ISDIR(inode->i_mode)) {
1003                 struct ceph_dir_file_info *dfi = file->private_data;
1004                 doutc(cl, "%p %llx.%llx dir file %p\n", inode,
1005                       ceph_vinop(inode), file);
1006                 WARN_ON(!list_empty(&dfi->file_info.rw_contexts));
1007
1008                 ceph_put_fmode(ci, dfi->file_info.fmode, 1);
1009
1010                 if (dfi->last_readdir)
1011                         ceph_mdsc_put_request(dfi->last_readdir);
1012                 kfree(dfi->last_name);
1013                 kfree(dfi->dir_info);
1014                 kmem_cache_free(ceph_dir_file_cachep, dfi);
1015         } else {
1016                 struct ceph_file_info *fi = file->private_data;
1017                 doutc(cl, "%p %llx.%llx regular file %p\n", inode,
1018                       ceph_vinop(inode), file);
1019                 WARN_ON(!list_empty(&fi->rw_contexts));
1020
1021                 ceph_fscache_unuse_cookie(inode, file->f_mode & FMODE_WRITE);
1022                 ceph_put_fmode(ci, fi->fmode, 1);
1023
1024                 kmem_cache_free(ceph_file_cachep, fi);
1025         }
1026
1027         /* wake up anyone waiting for caps on this inode */
1028         wake_up_all(&ci->i_cap_wq);
1029         return 0;
1030 }
1031
1032 enum {
1033         HAVE_RETRIED = 1,
1034         CHECK_EOF =    2,
1035         READ_INLINE =  3,
1036 };
1037
1038 /*
1039  * Completely synchronous read and write methods.  Direct from __user
1040  * buffer to osd, or directly to user pages (if O_DIRECT).
1041  *
1042  * If the read spans object boundary, just do multiple reads.  (That's not
1043  * atomic, but good enough for now.)
1044  *
1045  * If we get a short result from the OSD, check against i_size; we need to
1046  * only return a short read to the caller if we hit EOF.
1047  */
1048 ssize_t __ceph_sync_read(struct inode *inode, loff_t *ki_pos,
1049                          struct iov_iter *to, int *retry_op,
1050                          u64 *last_objver)
1051 {
1052         struct ceph_inode_info *ci = ceph_inode(inode);
1053         struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
1054         struct ceph_client *cl = fsc->client;
1055         struct ceph_osd_client *osdc = &fsc->client->osdc;
1056         ssize_t ret;
1057         u64 off = *ki_pos;
1058         u64 len = iov_iter_count(to);
1059         u64 i_size = i_size_read(inode);
1060         bool sparse = IS_ENCRYPTED(inode) || ceph_test_mount_opt(fsc, SPARSEREAD);
1061         u64 objver = 0;
1062
1063         doutc(cl, "on inode %p %llx.%llx %llx~%llx\n", inode,
1064               ceph_vinop(inode), *ki_pos, len);
1065
1066         if (ceph_inode_is_shutdown(inode))
1067                 return -EIO;
1068
1069         if (!len || !i_size)
1070                 return 0;
1071         /*
1072          * flush any page cache pages in this range.  this
1073          * will make concurrent normal and sync io slow,
1074          * but it will at least behave sensibly when they are
1075          * in sequence.
1076          */
1077         ret = filemap_write_and_wait_range(inode->i_mapping,
1078                                            off, off + len - 1);
1079         if (ret < 0)
1080                 return ret;
1081
1082         ret = 0;
1083         while ((len = iov_iter_count(to)) > 0) {
1084                 struct ceph_osd_request *req;
1085                 struct page **pages;
1086                 int num_pages;
1087                 size_t page_off;
1088                 bool more;
1089                 int idx = 0;
1090                 size_t left;
1091                 struct ceph_osd_req_op *op;
1092                 u64 read_off = off;
1093                 u64 read_len = len;
1094                 int extent_cnt;
1095
1096                 /* determine new offset/length if encrypted */
1097                 ceph_fscrypt_adjust_off_and_len(inode, &read_off, &read_len);
1098
1099                 doutc(cl, "orig %llu~%llu reading %llu~%llu", off, len,
1100                       read_off, read_len);
1101
1102                 req = ceph_osdc_new_request(osdc, &ci->i_layout,
1103                                         ci->i_vino, read_off, &read_len, 0, 1,
1104                                         sparse ? CEPH_OSD_OP_SPARSE_READ :
1105                                                  CEPH_OSD_OP_READ,
1106                                         CEPH_OSD_FLAG_READ,
1107                                         NULL, ci->i_truncate_seq,
1108                                         ci->i_truncate_size, false);
1109                 if (IS_ERR(req)) {
1110                         ret = PTR_ERR(req);
1111                         break;
1112                 }
1113
1114                 /* adjust len downward if the request truncated the len */
1115                 if (off + len > read_off + read_len)
1116                         len = read_off + read_len - off;
1117                 more = len < iov_iter_count(to);
1118
1119                 op = &req->r_ops[0];
1120                 if (sparse) {
1121                         extent_cnt = __ceph_sparse_read_ext_count(inode, read_len);
1122                         ret = ceph_alloc_sparse_ext_map(op, extent_cnt);
1123                         if (ret) {
1124                                 ceph_osdc_put_request(req);
1125                                 break;
1126                         }
1127                 }
1128
1129                 num_pages = calc_pages_for(read_off, read_len);
1130                 page_off = offset_in_page(off);
1131                 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1132                 if (IS_ERR(pages)) {
1133                         ceph_osdc_put_request(req);
1134                         ret = PTR_ERR(pages);
1135                         break;
1136                 }
1137
1138                 osd_req_op_extent_osd_data_pages(req, 0, pages, read_len,
1139                                                  offset_in_page(read_off),
1140                                                  false, true);
1141
1142                 ceph_osdc_start_request(osdc, req);
1143                 ret = ceph_osdc_wait_request(osdc, req);
1144
1145                 ceph_update_read_metrics(&fsc->mdsc->metric,
1146                                          req->r_start_latency,
1147                                          req->r_end_latency,
1148                                          read_len, ret);
1149
1150                 if (ret > 0)
1151                         objver = req->r_version;
1152
1153                 i_size = i_size_read(inode);
1154                 doutc(cl, "%llu~%llu got %zd i_size %llu%s\n", off, len,
1155                       ret, i_size, (more ? " MORE" : ""));
1156
1157                 /* Fix it to go to end of extent map */
1158                 if (sparse && ret >= 0)
1159                         ret = ceph_sparse_ext_map_end(op);
1160                 else if (ret == -ENOENT)
1161                         ret = 0;
1162
1163                 if (ret < 0) {
1164                         ceph_osdc_put_request(req);
1165                         if (ret == -EBLOCKLISTED)
1166                                 fsc->blocklisted = true;
1167                         break;
1168                 }
1169
1170                 if (IS_ENCRYPTED(inode)) {
1171                         int fret;
1172
1173                         fret = ceph_fscrypt_decrypt_extents(inode, pages,
1174                                         read_off, op->extent.sparse_ext,
1175                                         op->extent.sparse_ext_cnt);
1176                         if (fret < 0) {
1177                                 ret = fret;
1178                                 ceph_osdc_put_request(req);
1179                                 break;
1180                         }
1181
1182                         /* account for any partial block at the beginning */
1183                         fret -= (off - read_off);
1184
1185                         /*
1186                          * Short read after big offset adjustment?
1187                          * Nothing is usable, just call it a zero
1188                          * len read.
1189                          */
1190                         fret = max(fret, 0);
1191
1192                         /* account for partial block at the end */
1193                         ret = min_t(ssize_t, fret, len);
1194                 }
1195
1196                 /* Short read but not EOF? Zero out the remainder. */
1197                 if (ret < len && (off + ret < i_size)) {
1198                         int zlen = min(len - ret, i_size - off - ret);
1199                         int zoff = page_off + ret;
1200
1201                         doutc(cl, "zero gap %llu~%llu\n", off + ret,
1202                               off + ret + zlen);
1203                         ceph_zero_page_vector_range(zoff, zlen, pages);
1204                         ret += zlen;
1205                 }
1206
1207                 if (off + ret > i_size)
1208                         left = (i_size > off) ? i_size - off : 0;
1209                 else
1210                         left = ret;
1211
1212                 while (left > 0) {
1213                         size_t plen, copied;
1214
1215                         plen = min_t(size_t, left, PAGE_SIZE - page_off);
1216                         SetPageUptodate(pages[idx]);
1217                         copied = copy_page_to_iter(pages[idx++],
1218                                                    page_off, plen, to);
1219                         off += copied;
1220                         left -= copied;
1221                         page_off = 0;
1222                         if (copied < plen) {
1223                                 ret = -EFAULT;
1224                                 break;
1225                         }
1226                 }
1227
1228                 ceph_osdc_put_request(req);
1229
1230                 if (off >= i_size || !more)
1231                         break;
1232         }
1233
1234         if (ret > 0) {
1235                 if (off >= i_size) {
1236                         *retry_op = CHECK_EOF;
1237                         ret = i_size - *ki_pos;
1238                         *ki_pos = i_size;
1239                 } else {
1240                         ret = off - *ki_pos;
1241                         *ki_pos = off;
1242                 }
1243
1244                 if (last_objver)
1245                         *last_objver = objver;
1246         }
1247         doutc(cl, "result %zd retry_op %d\n", ret, *retry_op);
1248         return ret;
1249 }
1250
1251 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *to,
1252                               int *retry_op)
1253 {
1254         struct file *file = iocb->ki_filp;
1255         struct inode *inode = file_inode(file);
1256         struct ceph_client *cl = ceph_inode_to_client(inode);
1257
1258         doutc(cl, "on file %p %llx~%zx %s\n", file, iocb->ki_pos,
1259               iov_iter_count(to),
1260               (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
1261
1262         return __ceph_sync_read(inode, &iocb->ki_pos, to, retry_op, NULL);
1263 }
1264
1265 struct ceph_aio_request {
1266         struct kiocb *iocb;
1267         size_t total_len;
1268         bool write;
1269         bool should_dirty;
1270         int error;
1271         struct list_head osd_reqs;
1272         unsigned num_reqs;
1273         atomic_t pending_reqs;
1274         struct timespec64 mtime;
1275         struct ceph_cap_flush *prealloc_cf;
1276 };
1277
1278 struct ceph_aio_work {
1279         struct work_struct work;
1280         struct ceph_osd_request *req;
1281 };
1282
1283 static void ceph_aio_retry_work(struct work_struct *work);
1284
1285 static void ceph_aio_complete(struct inode *inode,
1286                               struct ceph_aio_request *aio_req)
1287 {
1288         struct ceph_client *cl = ceph_inode_to_client(inode);
1289         struct ceph_inode_info *ci = ceph_inode(inode);
1290         int ret;
1291
1292         if (!atomic_dec_and_test(&aio_req->pending_reqs))
1293                 return;
1294
1295         if (aio_req->iocb->ki_flags & IOCB_DIRECT)
1296                 inode_dio_end(inode);
1297
1298         ret = aio_req->error;
1299         if (!ret)
1300                 ret = aio_req->total_len;
1301
1302         doutc(cl, "%p %llx.%llx rc %d\n", inode, ceph_vinop(inode), ret);
1303
1304         if (ret >= 0 && aio_req->write) {
1305                 int dirty;
1306
1307                 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
1308                 if (endoff > i_size_read(inode)) {
1309                         if (ceph_inode_set_size(inode, endoff))
1310                                 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY);
1311                 }
1312
1313                 spin_lock(&ci->i_ceph_lock);
1314                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1315                                                &aio_req->prealloc_cf);
1316                 spin_unlock(&ci->i_ceph_lock);
1317                 if (dirty)
1318                         __mark_inode_dirty(inode, dirty);
1319
1320         }
1321
1322         ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
1323                                                 CEPH_CAP_FILE_RD));
1324
1325         aio_req->iocb->ki_complete(aio_req->iocb, ret);
1326
1327         ceph_free_cap_flush(aio_req->prealloc_cf);
1328         kfree(aio_req);
1329 }
1330
1331 static void ceph_aio_complete_req(struct ceph_osd_request *req)
1332 {
1333         int rc = req->r_result;
1334         struct inode *inode = req->r_inode;
1335         struct ceph_aio_request *aio_req = req->r_priv;
1336         struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
1337         struct ceph_osd_req_op *op = &req->r_ops[0];
1338         struct ceph_client_metric *metric = &ceph_sb_to_mdsc(inode->i_sb)->metric;
1339         unsigned int len = osd_data->bvec_pos.iter.bi_size;
1340         bool sparse = (op->op == CEPH_OSD_OP_SPARSE_READ);
1341         struct ceph_client *cl = ceph_inode_to_client(inode);
1342
1343         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_BVECS);
1344         BUG_ON(!osd_data->num_bvecs);
1345
1346         doutc(cl, "req %p inode %p %llx.%llx, rc %d bytes %u\n", req,
1347               inode, ceph_vinop(inode), rc, len);
1348
1349         if (rc == -EOLDSNAPC) {
1350                 struct ceph_aio_work *aio_work;
1351                 BUG_ON(!aio_req->write);
1352
1353                 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
1354                 if (aio_work) {
1355                         INIT_WORK(&aio_work->work, ceph_aio_retry_work);
1356                         aio_work->req = req;
1357                         queue_work(ceph_inode_to_fs_client(inode)->inode_wq,
1358                                    &aio_work->work);
1359                         return;
1360                 }
1361                 rc = -ENOMEM;
1362         } else if (!aio_req->write) {
1363                 if (sparse && rc >= 0)
1364                         rc = ceph_sparse_ext_map_end(op);
1365                 if (rc == -ENOENT)
1366                         rc = 0;
1367                 if (rc >= 0 && len > rc) {
1368                         struct iov_iter i;
1369                         int zlen = len - rc;
1370
1371                         /*
1372                          * If read is satisfied by single OSD request,
1373                          * it can pass EOF. Otherwise read is within
1374                          * i_size.
1375                          */
1376                         if (aio_req->num_reqs == 1) {
1377                                 loff_t i_size = i_size_read(inode);
1378                                 loff_t endoff = aio_req->iocb->ki_pos + rc;
1379                                 if (endoff < i_size)
1380                                         zlen = min_t(size_t, zlen,
1381                                                      i_size - endoff);
1382                                 aio_req->total_len = rc + zlen;
1383                         }
1384
1385                         iov_iter_bvec(&i, ITER_DEST, osd_data->bvec_pos.bvecs,
1386                                       osd_data->num_bvecs, len);
1387                         iov_iter_advance(&i, rc);
1388                         iov_iter_zero(zlen, &i);
1389                 }
1390         }
1391
1392         /* r_start_latency == 0 means the request was not submitted */
1393         if (req->r_start_latency) {
1394                 if (aio_req->write)
1395                         ceph_update_write_metrics(metric, req->r_start_latency,
1396                                                   req->r_end_latency, len, rc);
1397                 else
1398                         ceph_update_read_metrics(metric, req->r_start_latency,
1399                                                  req->r_end_latency, len, rc);
1400         }
1401
1402         put_bvecs(osd_data->bvec_pos.bvecs, osd_data->num_bvecs,
1403                   aio_req->should_dirty);
1404         ceph_osdc_put_request(req);
1405
1406         if (rc < 0)
1407                 cmpxchg(&aio_req->error, 0, rc);
1408
1409         ceph_aio_complete(inode, aio_req);
1410         return;
1411 }
1412
1413 static void ceph_aio_retry_work(struct work_struct *work)
1414 {
1415         struct ceph_aio_work *aio_work =
1416                 container_of(work, struct ceph_aio_work, work);
1417         struct ceph_osd_request *orig_req = aio_work->req;
1418         struct ceph_aio_request *aio_req = orig_req->r_priv;
1419         struct inode *inode = orig_req->r_inode;
1420         struct ceph_inode_info *ci = ceph_inode(inode);
1421         struct ceph_snap_context *snapc;
1422         struct ceph_osd_request *req;
1423         int ret;
1424
1425         spin_lock(&ci->i_ceph_lock);
1426         if (__ceph_have_pending_cap_snap(ci)) {
1427                 struct ceph_cap_snap *capsnap =
1428                         list_last_entry(&ci->i_cap_snaps,
1429                                         struct ceph_cap_snap,
1430                                         ci_item);
1431                 snapc = ceph_get_snap_context(capsnap->context);
1432         } else {
1433                 BUG_ON(!ci->i_head_snapc);
1434                 snapc = ceph_get_snap_context(ci->i_head_snapc);
1435         }
1436         spin_unlock(&ci->i_ceph_lock);
1437
1438         req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 1,
1439                         false, GFP_NOFS);
1440         if (!req) {
1441                 ret = -ENOMEM;
1442                 req = orig_req;
1443                 goto out;
1444         }
1445
1446         req->r_flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1447         ceph_oloc_copy(&req->r_base_oloc, &orig_req->r_base_oloc);
1448         ceph_oid_copy(&req->r_base_oid, &orig_req->r_base_oid);
1449
1450         req->r_ops[0] = orig_req->r_ops[0];
1451
1452         req->r_mtime = aio_req->mtime;
1453         req->r_data_offset = req->r_ops[0].extent.offset;
1454
1455         ret = ceph_osdc_alloc_messages(req, GFP_NOFS);
1456         if (ret) {
1457                 ceph_osdc_put_request(req);
1458                 req = orig_req;
1459                 goto out;
1460         }
1461
1462         ceph_osdc_put_request(orig_req);
1463
1464         req->r_callback = ceph_aio_complete_req;
1465         req->r_inode = inode;
1466         req->r_priv = aio_req;
1467
1468         ceph_osdc_start_request(req->r_osdc, req);
1469 out:
1470         if (ret < 0) {
1471                 req->r_result = ret;
1472                 ceph_aio_complete_req(req);
1473         }
1474
1475         ceph_put_snap_context(snapc);
1476         kfree(aio_work);
1477 }
1478
1479 static ssize_t
1480 ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
1481                        struct ceph_snap_context *snapc,
1482                        struct ceph_cap_flush **pcf)
1483 {
1484         struct file *file = iocb->ki_filp;
1485         struct inode *inode = file_inode(file);
1486         struct ceph_inode_info *ci = ceph_inode(inode);
1487         struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
1488         struct ceph_client *cl = fsc->client;
1489         struct ceph_client_metric *metric = &fsc->mdsc->metric;
1490         struct ceph_vino vino;
1491         struct ceph_osd_request *req;
1492         struct bio_vec *bvecs;
1493         struct ceph_aio_request *aio_req = NULL;
1494         int num_pages = 0;
1495         int flags;
1496         int ret = 0;
1497         struct timespec64 mtime = current_time(inode);
1498         size_t count = iov_iter_count(iter);
1499         loff_t pos = iocb->ki_pos;
1500         bool write = iov_iter_rw(iter) == WRITE;
1501         bool should_dirty = !write && user_backed_iter(iter);
1502         bool sparse = ceph_test_mount_opt(fsc, SPARSEREAD);
1503
1504         if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1505                 return -EROFS;
1506
1507         doutc(cl, "sync_direct_%s on file %p %lld~%u snapc %p seq %lld\n",
1508               (write ? "write" : "read"), file, pos, (unsigned)count,
1509               snapc, snapc ? snapc->seq : 0);
1510
1511         if (write) {
1512                 int ret2;
1513
1514                 ceph_fscache_invalidate(inode, true);
1515
1516                 ret2 = invalidate_inode_pages2_range(inode->i_mapping,
1517                                         pos >> PAGE_SHIFT,
1518                                         (pos + count - 1) >> PAGE_SHIFT);
1519                 if (ret2 < 0)
1520                         doutc(cl, "invalidate_inode_pages2_range returned %d\n",
1521                               ret2);
1522
1523                 flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1524         } else {
1525                 flags = CEPH_OSD_FLAG_READ;
1526         }
1527
1528         while (iov_iter_count(iter) > 0) {
1529                 u64 size = iov_iter_count(iter);
1530                 ssize_t len;
1531                 struct ceph_osd_req_op *op;
1532                 int readop = sparse ? CEPH_OSD_OP_SPARSE_READ : CEPH_OSD_OP_READ;
1533                 int extent_cnt;
1534
1535                 if (write)
1536                         size = min_t(u64, size, fsc->mount_options->wsize);
1537                 else
1538                         size = min_t(u64, size, fsc->mount_options->rsize);
1539
1540                 vino = ceph_vino(inode);
1541                 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1542                                             vino, pos, &size, 0,
1543                                             1,
1544                                             write ? CEPH_OSD_OP_WRITE : readop,
1545                                             flags, snapc,
1546                                             ci->i_truncate_seq,
1547                                             ci->i_truncate_size,
1548                                             false);
1549                 if (IS_ERR(req)) {
1550                         ret = PTR_ERR(req);
1551                         break;
1552                 }
1553
1554                 op = &req->r_ops[0];
1555                 if (!write && sparse) {
1556                         extent_cnt = __ceph_sparse_read_ext_count(inode, size);
1557                         ret = ceph_alloc_sparse_ext_map(op, extent_cnt);
1558                         if (ret) {
1559                                 ceph_osdc_put_request(req);
1560                                 break;
1561                         }
1562                 }
1563
1564                 len = iter_get_bvecs_alloc(iter, size, &bvecs, &num_pages);
1565                 if (len < 0) {
1566                         ceph_osdc_put_request(req);
1567                         ret = len;
1568                         break;
1569                 }
1570                 if (len != size)
1571                         osd_req_op_extent_update(req, 0, len);
1572
1573                 osd_req_op_extent_osd_data_bvecs(req, 0, bvecs, num_pages, len);
1574
1575                 /*
1576                  * To simplify error handling, allow AIO when IO within i_size
1577                  * or IO can be satisfied by single OSD request.
1578                  */
1579                 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
1580                     (len == count || pos + count <= i_size_read(inode))) {
1581                         aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
1582                         if (aio_req) {
1583                                 aio_req->iocb = iocb;
1584                                 aio_req->write = write;
1585                                 aio_req->should_dirty = should_dirty;
1586                                 INIT_LIST_HEAD(&aio_req->osd_reqs);
1587                                 if (write) {
1588                                         aio_req->mtime = mtime;
1589                                         swap(aio_req->prealloc_cf, *pcf);
1590                                 }
1591                         }
1592                         /* ignore error */
1593                 }
1594
1595                 if (write) {
1596                         /*
1597                          * throw out any page cache pages in this range. this
1598                          * may block.
1599                          */
1600                         truncate_inode_pages_range(inode->i_mapping, pos,
1601                                                    PAGE_ALIGN(pos + len) - 1);
1602
1603                         req->r_mtime = mtime;
1604                 }
1605
1606                 if (aio_req) {
1607                         aio_req->total_len += len;
1608                         aio_req->num_reqs++;
1609                         atomic_inc(&aio_req->pending_reqs);
1610
1611                         req->r_callback = ceph_aio_complete_req;
1612                         req->r_inode = inode;
1613                         req->r_priv = aio_req;
1614                         list_add_tail(&req->r_private_item, &aio_req->osd_reqs);
1615
1616                         pos += len;
1617                         continue;
1618                 }
1619
1620                 ceph_osdc_start_request(req->r_osdc, req);
1621                 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1622
1623                 if (write)
1624                         ceph_update_write_metrics(metric, req->r_start_latency,
1625                                                   req->r_end_latency, len, ret);
1626                 else
1627                         ceph_update_read_metrics(metric, req->r_start_latency,
1628                                                  req->r_end_latency, len, ret);
1629
1630                 size = i_size_read(inode);
1631                 if (!write) {
1632                         if (sparse && ret >= 0)
1633                                 ret = ceph_sparse_ext_map_end(op);
1634                         else if (ret == -ENOENT)
1635                                 ret = 0;
1636
1637                         if (ret >= 0 && ret < len && pos + ret < size) {
1638                                 struct iov_iter i;
1639                                 int zlen = min_t(size_t, len - ret,
1640                                                  size - pos - ret);
1641
1642                                 iov_iter_bvec(&i, ITER_DEST, bvecs, num_pages, len);
1643                                 iov_iter_advance(&i, ret);
1644                                 iov_iter_zero(zlen, &i);
1645                                 ret += zlen;
1646                         }
1647                         if (ret >= 0)
1648                                 len = ret;
1649                 }
1650
1651                 put_bvecs(bvecs, num_pages, should_dirty);
1652                 ceph_osdc_put_request(req);
1653                 if (ret < 0)
1654                         break;
1655
1656                 pos += len;
1657                 if (!write && pos >= size)
1658                         break;
1659
1660                 if (write && pos > size) {
1661                         if (ceph_inode_set_size(inode, pos))
1662                                 ceph_check_caps(ceph_inode(inode),
1663                                                 CHECK_CAPS_AUTHONLY);
1664                 }
1665         }
1666
1667         if (aio_req) {
1668                 LIST_HEAD(osd_reqs);
1669
1670                 if (aio_req->num_reqs == 0) {
1671                         kfree(aio_req);
1672                         return ret;
1673                 }
1674
1675                 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
1676                                               CEPH_CAP_FILE_RD);
1677
1678                 list_splice(&aio_req->osd_reqs, &osd_reqs);
1679                 inode_dio_begin(inode);
1680                 while (!list_empty(&osd_reqs)) {
1681                         req = list_first_entry(&osd_reqs,
1682                                                struct ceph_osd_request,
1683                                                r_private_item);
1684                         list_del_init(&req->r_private_item);
1685                         if (ret >= 0)
1686                                 ceph_osdc_start_request(req->r_osdc, req);
1687                         if (ret < 0) {
1688                                 req->r_result = ret;
1689                                 ceph_aio_complete_req(req);
1690                         }
1691                 }
1692                 return -EIOCBQUEUED;
1693         }
1694
1695         if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
1696                 ret = pos - iocb->ki_pos;
1697                 iocb->ki_pos = pos;
1698         }
1699         return ret;
1700 }
1701
1702 /*
1703  * Synchronous write, straight from __user pointer or user pages.
1704  *
1705  * If write spans object boundary, just do multiple writes.  (For a
1706  * correct atomic write, we should e.g. take write locks on all
1707  * objects, rollback on failure, etc.)
1708  */
1709 static ssize_t
1710 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
1711                 struct ceph_snap_context *snapc)
1712 {
1713         struct file *file = iocb->ki_filp;
1714         struct inode *inode = file_inode(file);
1715         struct ceph_inode_info *ci = ceph_inode(inode);
1716         struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
1717         struct ceph_client *cl = fsc->client;
1718         struct ceph_osd_client *osdc = &fsc->client->osdc;
1719         struct ceph_osd_request *req;
1720         struct page **pages;
1721         u64 len;
1722         int num_pages;
1723         int written = 0;
1724         int ret;
1725         bool check_caps = false;
1726         struct timespec64 mtime = current_time(inode);
1727         size_t count = iov_iter_count(from);
1728
1729         if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1730                 return -EROFS;
1731
1732         doutc(cl, "on file %p %lld~%u snapc %p seq %lld\n", file, pos,
1733               (unsigned)count, snapc, snapc->seq);
1734
1735         ret = filemap_write_and_wait_range(inode->i_mapping,
1736                                            pos, pos + count - 1);
1737         if (ret < 0)
1738                 return ret;
1739
1740         ceph_fscache_invalidate(inode, false);
1741
1742         while ((len = iov_iter_count(from)) > 0) {
1743                 size_t left;
1744                 int n;
1745                 u64 write_pos = pos;
1746                 u64 write_len = len;
1747                 u64 objnum, objoff;
1748                 u32 xlen;
1749                 u64 assert_ver = 0;
1750                 bool rmw;
1751                 bool first, last;
1752                 struct iov_iter saved_iter = *from;
1753                 size_t off;
1754
1755                 ceph_fscrypt_adjust_off_and_len(inode, &write_pos, &write_len);
1756
1757                 /* clamp the length to the end of first object */
1758                 ceph_calc_file_object_mapping(&ci->i_layout, write_pos,
1759                                               write_len, &objnum, &objoff,
1760                                               &xlen);
1761                 write_len = xlen;
1762
1763                 /* adjust len downward if it goes beyond current object */
1764                 if (pos + len > write_pos + write_len)
1765                         len = write_pos + write_len - pos;
1766
1767                 /*
1768                  * If we had to adjust the length or position to align with a
1769                  * crypto block, then we must do a read/modify/write cycle. We
1770                  * use a version assertion to redrive the thing if something
1771                  * changes in between.
1772                  */
1773                 first = pos != write_pos;
1774                 last = (pos + len) != (write_pos + write_len);
1775                 rmw = first || last;
1776
1777                 doutc(cl, "ino %llx %lld~%llu adjusted %lld~%llu -- %srmw\n",
1778                       ci->i_vino.ino, pos, len, write_pos, write_len,
1779                       rmw ? "" : "no ");
1780
1781                 /*
1782                  * The data is emplaced into the page as it would be if it were
1783                  * in an array of pagecache pages.
1784                  */
1785                 num_pages = calc_pages_for(write_pos, write_len);
1786                 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1787                 if (IS_ERR(pages)) {
1788                         ret = PTR_ERR(pages);
1789                         break;
1790                 }
1791
1792                 /* Do we need to preload the pages? */
1793                 if (rmw) {
1794                         u64 first_pos = write_pos;
1795                         u64 last_pos = (write_pos + write_len) - CEPH_FSCRYPT_BLOCK_SIZE;
1796                         u64 read_len = CEPH_FSCRYPT_BLOCK_SIZE;
1797                         struct ceph_osd_req_op *op;
1798
1799                         /* We should only need to do this for encrypted inodes */
1800                         WARN_ON_ONCE(!IS_ENCRYPTED(inode));
1801
1802                         /* No need to do two reads if first and last blocks are same */
1803                         if (first && last_pos == first_pos)
1804                                 last = false;
1805
1806                         /*
1807                          * Allocate a read request for one or two extents,
1808                          * depending on how the request was aligned.
1809                          */
1810                         req = ceph_osdc_new_request(osdc, &ci->i_layout,
1811                                         ci->i_vino, first ? first_pos : last_pos,
1812                                         &read_len, 0, (first && last) ? 2 : 1,
1813                                         CEPH_OSD_OP_SPARSE_READ, CEPH_OSD_FLAG_READ,
1814                                         NULL, ci->i_truncate_seq,
1815                                         ci->i_truncate_size, false);
1816                         if (IS_ERR(req)) {
1817                                 ceph_release_page_vector(pages, num_pages);
1818                                 ret = PTR_ERR(req);
1819                                 break;
1820                         }
1821
1822                         /* Something is misaligned! */
1823                         if (read_len != CEPH_FSCRYPT_BLOCK_SIZE) {
1824                                 ceph_osdc_put_request(req);
1825                                 ceph_release_page_vector(pages, num_pages);
1826                                 ret = -EIO;
1827                                 break;
1828                         }
1829
1830                         /* Add extent for first block? */
1831                         op = &req->r_ops[0];
1832
1833                         if (first) {
1834                                 osd_req_op_extent_osd_data_pages(req, 0, pages,
1835                                                          CEPH_FSCRYPT_BLOCK_SIZE,
1836                                                          offset_in_page(first_pos),
1837                                                          false, false);
1838                                 /* We only expect a single extent here */
1839                                 ret = __ceph_alloc_sparse_ext_map(op, 1);
1840                                 if (ret) {
1841                                         ceph_osdc_put_request(req);
1842                                         ceph_release_page_vector(pages, num_pages);
1843                                         break;
1844                                 }
1845                         }
1846
1847                         /* Add extent for last block */
1848                         if (last) {
1849                                 /* Init the other extent if first extent has been used */
1850                                 if (first) {
1851                                         op = &req->r_ops[1];
1852                                         osd_req_op_extent_init(req, 1,
1853                                                         CEPH_OSD_OP_SPARSE_READ,
1854                                                         last_pos, CEPH_FSCRYPT_BLOCK_SIZE,
1855                                                         ci->i_truncate_size,
1856                                                         ci->i_truncate_seq);
1857                                 }
1858
1859                                 ret = __ceph_alloc_sparse_ext_map(op, 1);
1860                                 if (ret) {
1861                                         ceph_osdc_put_request(req);
1862                                         ceph_release_page_vector(pages, num_pages);
1863                                         break;
1864                                 }
1865
1866                                 osd_req_op_extent_osd_data_pages(req, first ? 1 : 0,
1867                                                         &pages[num_pages - 1],
1868                                                         CEPH_FSCRYPT_BLOCK_SIZE,
1869                                                         offset_in_page(last_pos),
1870                                                         false, false);
1871                         }
1872
1873                         ceph_osdc_start_request(osdc, req);
1874                         ret = ceph_osdc_wait_request(osdc, req);
1875
1876                         /* FIXME: length field is wrong if there are 2 extents */
1877                         ceph_update_read_metrics(&fsc->mdsc->metric,
1878                                                  req->r_start_latency,
1879                                                  req->r_end_latency,
1880                                                  read_len, ret);
1881
1882                         /* Ok if object is not already present */
1883                         if (ret == -ENOENT) {
1884                                 /*
1885                                  * If there is no object, then we can't assert
1886                                  * on its version. Set it to 0, and we'll use an
1887                                  * exclusive create instead.
1888                                  */
1889                                 ceph_osdc_put_request(req);
1890                                 ret = 0;
1891
1892                                 /*
1893                                  * zero out the soon-to-be uncopied parts of the
1894                                  * first and last pages.
1895                                  */
1896                                 if (first)
1897                                         zero_user_segment(pages[0], 0,
1898                                                           offset_in_page(first_pos));
1899                                 if (last)
1900                                         zero_user_segment(pages[num_pages - 1],
1901                                                           offset_in_page(last_pos),
1902                                                           PAGE_SIZE);
1903                         } else {
1904                                 if (ret < 0) {
1905                                         ceph_osdc_put_request(req);
1906                                         ceph_release_page_vector(pages, num_pages);
1907                                         break;
1908                                 }
1909
1910                                 op = &req->r_ops[0];
1911                                 if (op->extent.sparse_ext_cnt == 0) {
1912                                         if (first)
1913                                                 zero_user_segment(pages[0], 0,
1914                                                                   offset_in_page(first_pos));
1915                                         else
1916                                                 zero_user_segment(pages[num_pages - 1],
1917                                                                   offset_in_page(last_pos),
1918                                                                   PAGE_SIZE);
1919                                 } else if (op->extent.sparse_ext_cnt != 1 ||
1920                                            ceph_sparse_ext_map_end(op) !=
1921                                                 CEPH_FSCRYPT_BLOCK_SIZE) {
1922                                         ret = -EIO;
1923                                         ceph_osdc_put_request(req);
1924                                         ceph_release_page_vector(pages, num_pages);
1925                                         break;
1926                                 }
1927
1928                                 if (first && last) {
1929                                         op = &req->r_ops[1];
1930                                         if (op->extent.sparse_ext_cnt == 0) {
1931                                                 zero_user_segment(pages[num_pages - 1],
1932                                                                   offset_in_page(last_pos),
1933                                                                   PAGE_SIZE);
1934                                         } else if (op->extent.sparse_ext_cnt != 1 ||
1935                                                    ceph_sparse_ext_map_end(op) !=
1936                                                         CEPH_FSCRYPT_BLOCK_SIZE) {
1937                                                 ret = -EIO;
1938                                                 ceph_osdc_put_request(req);
1939                                                 ceph_release_page_vector(pages, num_pages);
1940                                                 break;
1941                                         }
1942                                 }
1943
1944                                 /* Grab assert version. It must be non-zero. */
1945                                 assert_ver = req->r_version;
1946                                 WARN_ON_ONCE(ret > 0 && assert_ver == 0);
1947
1948                                 ceph_osdc_put_request(req);
1949                                 if (first) {
1950                                         ret = ceph_fscrypt_decrypt_block_inplace(inode,
1951                                                         pages[0], CEPH_FSCRYPT_BLOCK_SIZE,
1952                                                         offset_in_page(first_pos),
1953                                                         first_pos >> CEPH_FSCRYPT_BLOCK_SHIFT);
1954                                         if (ret < 0) {
1955                                                 ceph_release_page_vector(pages, num_pages);
1956                                                 break;
1957                                         }
1958                                 }
1959                                 if (last) {
1960                                         ret = ceph_fscrypt_decrypt_block_inplace(inode,
1961                                                         pages[num_pages - 1],
1962                                                         CEPH_FSCRYPT_BLOCK_SIZE,
1963                                                         offset_in_page(last_pos),
1964                                                         last_pos >> CEPH_FSCRYPT_BLOCK_SHIFT);
1965                                         if (ret < 0) {
1966                                                 ceph_release_page_vector(pages, num_pages);
1967                                                 break;
1968                                         }
1969                                 }
1970                         }
1971                 }
1972
1973                 left = len;
1974                 off = offset_in_page(pos);
1975                 for (n = 0; n < num_pages; n++) {
1976                         size_t plen = min_t(size_t, left, PAGE_SIZE - off);
1977
1978                         /* copy the data */
1979                         ret = copy_page_from_iter(pages[n], off, plen, from);
1980                         if (ret != plen) {
1981                                 ret = -EFAULT;
1982                                 break;
1983                         }
1984                         off = 0;
1985                         left -= ret;
1986                 }
1987                 if (ret < 0) {
1988                         doutc(cl, "write failed with %d\n", ret);
1989                         ceph_release_page_vector(pages, num_pages);
1990                         break;
1991                 }
1992
1993                 if (IS_ENCRYPTED(inode)) {
1994                         ret = ceph_fscrypt_encrypt_pages(inode, pages,
1995                                                          write_pos, write_len,
1996                                                          GFP_KERNEL);
1997                         if (ret < 0) {
1998                                 doutc(cl, "encryption failed with %d\n", ret);
1999                                 ceph_release_page_vector(pages, num_pages);
2000                                 break;
2001                         }
2002                 }
2003
2004                 req = ceph_osdc_new_request(osdc, &ci->i_layout,
2005                                             ci->i_vino, write_pos, &write_len,
2006                                             rmw ? 1 : 0, rmw ? 2 : 1,
2007                                             CEPH_OSD_OP_WRITE,
2008                                             CEPH_OSD_FLAG_WRITE,
2009                                             snapc, ci->i_truncate_seq,
2010                                             ci->i_truncate_size, false);
2011                 if (IS_ERR(req)) {
2012                         ret = PTR_ERR(req);
2013                         ceph_release_page_vector(pages, num_pages);
2014                         break;
2015                 }
2016
2017                 doutc(cl, "write op %lld~%llu\n", write_pos, write_len);
2018                 osd_req_op_extent_osd_data_pages(req, rmw ? 1 : 0, pages, write_len,
2019                                                  offset_in_page(write_pos), false,
2020                                                  true);
2021                 req->r_inode = inode;
2022                 req->r_mtime = mtime;
2023
2024                 /* Set up the assertion */
2025                 if (rmw) {
2026                         /*
2027                          * Set up the assertion. If we don't have a version
2028                          * number, then the object doesn't exist yet. Use an
2029                          * exclusive create instead of a version assertion in
2030                          * that case.
2031                          */
2032                         if (assert_ver) {
2033                                 osd_req_op_init(req, 0, CEPH_OSD_OP_ASSERT_VER, 0);
2034                                 req->r_ops[0].assert_ver.ver = assert_ver;
2035                         } else {
2036                                 osd_req_op_init(req, 0, CEPH_OSD_OP_CREATE,
2037                                                 CEPH_OSD_OP_FLAG_EXCL);
2038                         }
2039                 }
2040
2041                 ceph_osdc_start_request(osdc, req);
2042                 ret = ceph_osdc_wait_request(osdc, req);
2043
2044                 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
2045                                           req->r_end_latency, len, ret);
2046                 ceph_osdc_put_request(req);
2047                 if (ret != 0) {
2048                         doutc(cl, "osd write returned %d\n", ret);
2049                         /* Version changed! Must re-do the rmw cycle */
2050                         if ((assert_ver && (ret == -ERANGE || ret == -EOVERFLOW)) ||
2051                             (!assert_ver && ret == -EEXIST)) {
2052                                 /* We should only ever see this on a rmw */
2053                                 WARN_ON_ONCE(!rmw);
2054
2055                                 /* The version should never go backward */
2056                                 WARN_ON_ONCE(ret == -EOVERFLOW);
2057
2058                                 *from = saved_iter;
2059
2060                                 /* FIXME: limit number of times we loop? */
2061                                 continue;
2062                         }
2063                         ceph_set_error_write(ci);
2064                         break;
2065                 }
2066
2067                 ceph_clear_error_write(ci);
2068
2069                 /*
2070                  * We successfully wrote to a range of the file. Declare
2071                  * that region of the pagecache invalid.
2072                  */
2073                 ret = invalidate_inode_pages2_range(
2074                                 inode->i_mapping,
2075                                 pos >> PAGE_SHIFT,
2076                                 (pos + len - 1) >> PAGE_SHIFT);
2077                 if (ret < 0) {
2078                         doutc(cl, "invalidate_inode_pages2_range returned %d\n",
2079                               ret);
2080                         ret = 0;
2081                 }
2082                 pos += len;
2083                 written += len;
2084                 doutc(cl, "written %d\n", written);
2085                 if (pos > i_size_read(inode)) {
2086                         check_caps = ceph_inode_set_size(inode, pos);
2087                         if (check_caps)
2088                                 ceph_check_caps(ceph_inode(inode),
2089                                                 CHECK_CAPS_AUTHONLY);
2090                 }
2091
2092         }
2093
2094         if (ret != -EOLDSNAPC && written > 0) {
2095                 ret = written;
2096                 iocb->ki_pos = pos;
2097         }
2098         doutc(cl, "returning %d\n", ret);
2099         return ret;
2100 }
2101
2102 /*
2103  * Wrap generic_file_aio_read with checks for cap bits on the inode.
2104  * Atomically grab references, so that those bits are not released
2105  * back to the MDS mid-read.
2106  *
2107  * Hmm, the sync read case isn't actually async... should it be?
2108  */
2109 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
2110 {
2111         struct file *filp = iocb->ki_filp;
2112         struct ceph_file_info *fi = filp->private_data;
2113         size_t len = iov_iter_count(to);
2114         struct inode *inode = file_inode(filp);
2115         struct ceph_inode_info *ci = ceph_inode(inode);
2116         bool direct_lock = iocb->ki_flags & IOCB_DIRECT;
2117         struct ceph_client *cl = ceph_inode_to_client(inode);
2118         ssize_t ret;
2119         int want = 0, got = 0;
2120         int retry_op = 0, read = 0;
2121
2122 again:
2123         doutc(cl, "%llu~%u trying to get caps on %p %llx.%llx\n",
2124               iocb->ki_pos, (unsigned)len, inode, ceph_vinop(inode));
2125
2126         if (ceph_inode_is_shutdown(inode))
2127                 return -ESTALE;
2128
2129         if (direct_lock)
2130                 ceph_start_io_direct(inode);
2131         else
2132                 ceph_start_io_read(inode);
2133
2134         if (!(fi->flags & CEPH_F_SYNC) && !direct_lock)
2135                 want |= CEPH_CAP_FILE_CACHE;
2136         if (fi->fmode & CEPH_FILE_MODE_LAZY)
2137                 want |= CEPH_CAP_FILE_LAZYIO;
2138
2139         ret = ceph_get_caps(filp, CEPH_CAP_FILE_RD, want, -1, &got);
2140         if (ret < 0) {
2141                 if (direct_lock)
2142                         ceph_end_io_direct(inode);
2143                 else
2144                         ceph_end_io_read(inode);
2145                 return ret;
2146         }
2147
2148         if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
2149             (iocb->ki_flags & IOCB_DIRECT) ||
2150             (fi->flags & CEPH_F_SYNC)) {
2151
2152                 doutc(cl, "sync %p %llx.%llx %llu~%u got cap refs on %s\n",
2153                       inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
2154                       ceph_cap_string(got));
2155
2156                 if (!ceph_has_inline_data(ci)) {
2157                         if (!retry_op &&
2158                             (iocb->ki_flags & IOCB_DIRECT) &&
2159                             !IS_ENCRYPTED(inode)) {
2160                                 ret = ceph_direct_read_write(iocb, to,
2161                                                              NULL, NULL);
2162                                 if (ret >= 0 && ret < len)
2163                                         retry_op = CHECK_EOF;
2164                         } else {
2165                                 ret = ceph_sync_read(iocb, to, &retry_op);
2166                         }
2167                 } else {
2168                         retry_op = READ_INLINE;
2169                 }
2170         } else {
2171                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
2172                 doutc(cl, "async %p %llx.%llx %llu~%u got cap refs on %s\n",
2173                       inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
2174                       ceph_cap_string(got));
2175                 ceph_add_rw_context(fi, &rw_ctx);
2176                 ret = generic_file_read_iter(iocb, to);
2177                 ceph_del_rw_context(fi, &rw_ctx);
2178         }
2179
2180         doutc(cl, "%p %llx.%llx dropping cap refs on %s = %d\n",
2181               inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
2182         ceph_put_cap_refs(ci, got);
2183
2184         if (direct_lock)
2185                 ceph_end_io_direct(inode);
2186         else
2187                 ceph_end_io_read(inode);
2188
2189         if (retry_op > HAVE_RETRIED && ret >= 0) {
2190                 int statret;
2191                 struct page *page = NULL;
2192                 loff_t i_size;
2193                 int mask = CEPH_STAT_CAP_SIZE;
2194                 if (retry_op == READ_INLINE) {
2195                         page = __page_cache_alloc(GFP_KERNEL);
2196                         if (!page)
2197                                 return -ENOMEM;
2198
2199                         mask = CEPH_STAT_CAP_INLINE_DATA;
2200                 }
2201
2202                 statret = __ceph_do_getattr(inode, page, mask, !!page);
2203                 if (statret < 0) {
2204                         if (page)
2205                                 __free_page(page);
2206                         if (statret == -ENODATA) {
2207                                 BUG_ON(retry_op != READ_INLINE);
2208                                 goto again;
2209                         }
2210                         return statret;
2211                 }
2212
2213                 i_size = i_size_read(inode);
2214                 if (retry_op == READ_INLINE) {
2215                         BUG_ON(ret > 0 || read > 0);
2216                         if (iocb->ki_pos < i_size &&
2217                             iocb->ki_pos < PAGE_SIZE) {
2218                                 loff_t end = min_t(loff_t, i_size,
2219                                                    iocb->ki_pos + len);
2220                                 end = min_t(loff_t, end, PAGE_SIZE);
2221                                 if (statret < end)
2222                                         zero_user_segment(page, statret, end);
2223                                 ret = copy_page_to_iter(page,
2224                                                 iocb->ki_pos & ~PAGE_MASK,
2225                                                 end - iocb->ki_pos, to);
2226                                 iocb->ki_pos += ret;
2227                                 read += ret;
2228                         }
2229                         if (iocb->ki_pos < i_size && read < len) {
2230                                 size_t zlen = min_t(size_t, len - read,
2231                                                     i_size - iocb->ki_pos);
2232                                 ret = iov_iter_zero(zlen, to);
2233                                 iocb->ki_pos += ret;
2234                                 read += ret;
2235                         }
2236                         __free_pages(page, 0);
2237                         return read;
2238                 }
2239
2240                 /* hit EOF or hole? */
2241                 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
2242                     ret < len) {
2243                         doutc(cl, "may hit hole, ppos %lld < size %lld, reading more\n",
2244                               iocb->ki_pos, i_size);
2245
2246                         read += ret;
2247                         len -= ret;
2248                         retry_op = HAVE_RETRIED;
2249                         goto again;
2250                 }
2251         }
2252
2253         if (ret >= 0)
2254                 ret += read;
2255
2256         return ret;
2257 }
2258
2259 /*
2260  * Wrap filemap_splice_read with checks for cap bits on the inode.
2261  * Atomically grab references, so that those bits are not released
2262  * back to the MDS mid-read.
2263  */
2264 static ssize_t ceph_splice_read(struct file *in, loff_t *ppos,
2265                                 struct pipe_inode_info *pipe,
2266                                 size_t len, unsigned int flags)
2267 {
2268         struct ceph_file_info *fi = in->private_data;
2269         struct inode *inode = file_inode(in);
2270         struct ceph_inode_info *ci = ceph_inode(inode);
2271         ssize_t ret;
2272         int want = 0, got = 0;
2273         CEPH_DEFINE_RW_CONTEXT(rw_ctx, 0);
2274
2275         dout("splice_read %p %llx.%llx %llu~%zu trying to get caps on %p\n",
2276              inode, ceph_vinop(inode), *ppos, len, inode);
2277
2278         if (ceph_inode_is_shutdown(inode))
2279                 return -ESTALE;
2280
2281         if (ceph_has_inline_data(ci) ||
2282             (fi->flags & CEPH_F_SYNC))
2283                 return copy_splice_read(in, ppos, pipe, len, flags);
2284
2285         ceph_start_io_read(inode);
2286
2287         want = CEPH_CAP_FILE_CACHE;
2288         if (fi->fmode & CEPH_FILE_MODE_LAZY)
2289                 want |= CEPH_CAP_FILE_LAZYIO;
2290
2291         ret = ceph_get_caps(in, CEPH_CAP_FILE_RD, want, -1, &got);
2292         if (ret < 0)
2293                 goto out_end;
2294
2295         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) == 0) {
2296                 dout("splice_read/sync %p %llx.%llx %llu~%zu got cap refs on %s\n",
2297                      inode, ceph_vinop(inode), *ppos, len,
2298                      ceph_cap_string(got));
2299
2300                 ceph_put_cap_refs(ci, got);
2301                 ceph_end_io_read(inode);
2302                 return copy_splice_read(in, ppos, pipe, len, flags);
2303         }
2304
2305         dout("splice_read %p %llx.%llx %llu~%zu got cap refs on %s\n",
2306              inode, ceph_vinop(inode), *ppos, len, ceph_cap_string(got));
2307
2308         rw_ctx.caps = got;
2309         ceph_add_rw_context(fi, &rw_ctx);
2310         ret = filemap_splice_read(in, ppos, pipe, len, flags);
2311         ceph_del_rw_context(fi, &rw_ctx);
2312
2313         dout("splice_read %p %llx.%llx dropping cap refs on %s = %zd\n",
2314              inode, ceph_vinop(inode), ceph_cap_string(got), ret);
2315
2316         ceph_put_cap_refs(ci, got);
2317 out_end:
2318         ceph_end_io_read(inode);
2319         return ret;
2320 }
2321
2322 /*
2323  * Take cap references to avoid releasing caps to MDS mid-write.
2324  *
2325  * If we are synchronous, and write with an old snap context, the OSD
2326  * may return EOLDSNAPC.  In that case, retry the write.. _after_
2327  * dropping our cap refs and allowing the pending snap to logically
2328  * complete _before_ this write occurs.
2329  *
2330  * If we are near ENOSPC, write synchronously.
2331  */
2332 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
2333 {
2334         struct file *file = iocb->ki_filp;
2335         struct ceph_file_info *fi = file->private_data;
2336         struct inode *inode = file_inode(file);
2337         struct ceph_inode_info *ci = ceph_inode(inode);
2338         struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
2339         struct ceph_client *cl = fsc->client;
2340         struct ceph_osd_client *osdc = &fsc->client->osdc;
2341         struct ceph_cap_flush *prealloc_cf;
2342         ssize_t count, written = 0;
2343         int err, want = 0, got;
2344         bool direct_lock = false;
2345         u32 map_flags;
2346         u64 pool_flags;
2347         loff_t pos;
2348         loff_t limit = max(i_size_read(inode), fsc->max_file_size);
2349
2350         if (ceph_inode_is_shutdown(inode))
2351                 return -ESTALE;
2352
2353         if (ceph_snap(inode) != CEPH_NOSNAP)
2354                 return -EROFS;
2355
2356         prealloc_cf = ceph_alloc_cap_flush();
2357         if (!prealloc_cf)
2358                 return -ENOMEM;
2359
2360         if ((iocb->ki_flags & (IOCB_DIRECT | IOCB_APPEND)) == IOCB_DIRECT)
2361                 direct_lock = true;
2362
2363 retry_snap:
2364         if (direct_lock)
2365                 ceph_start_io_direct(inode);
2366         else
2367                 ceph_start_io_write(inode);
2368
2369         if (iocb->ki_flags & IOCB_APPEND) {
2370                 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
2371                 if (err < 0)
2372                         goto out;
2373         }
2374
2375         err = generic_write_checks(iocb, from);
2376         if (err <= 0)
2377                 goto out;
2378
2379         pos = iocb->ki_pos;
2380         if (unlikely(pos >= limit)) {
2381                 err = -EFBIG;
2382                 goto out;
2383         } else {
2384                 iov_iter_truncate(from, limit - pos);
2385         }
2386
2387         count = iov_iter_count(from);
2388         if (ceph_quota_is_max_bytes_exceeded(inode, pos + count)) {
2389                 err = -EDQUOT;
2390                 goto out;
2391         }
2392
2393         down_read(&osdc->lock);
2394         map_flags = osdc->osdmap->flags;
2395         pool_flags = ceph_pg_pool_flags(osdc->osdmap, ci->i_layout.pool_id);
2396         up_read(&osdc->lock);
2397         if ((map_flags & CEPH_OSDMAP_FULL) ||
2398             (pool_flags & CEPH_POOL_FLAG_FULL)) {
2399                 err = -ENOSPC;
2400                 goto out;
2401         }
2402
2403         err = file_remove_privs(file);
2404         if (err)
2405                 goto out;
2406
2407         doutc(cl, "%p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
2408               inode, ceph_vinop(inode), pos, count,
2409               i_size_read(inode));
2410         if (!(fi->flags & CEPH_F_SYNC) && !direct_lock)
2411                 want |= CEPH_CAP_FILE_BUFFER;
2412         if (fi->fmode & CEPH_FILE_MODE_LAZY)
2413                 want |= CEPH_CAP_FILE_LAZYIO;
2414         got = 0;
2415         err = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, pos + count, &got);
2416         if (err < 0)
2417                 goto out;
2418
2419         err = file_update_time(file);
2420         if (err)
2421                 goto out_caps;
2422
2423         inode_inc_iversion_raw(inode);
2424
2425         doutc(cl, "%p %llx.%llx %llu~%zd got cap refs on %s\n",
2426               inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
2427
2428         if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
2429             (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC) ||
2430             (ci->i_ceph_flags & CEPH_I_ERROR_WRITE)) {
2431                 struct ceph_snap_context *snapc;
2432                 struct iov_iter data;
2433
2434                 spin_lock(&ci->i_ceph_lock);
2435                 if (__ceph_have_pending_cap_snap(ci)) {
2436                         struct ceph_cap_snap *capsnap =
2437                                         list_last_entry(&ci->i_cap_snaps,
2438                                                         struct ceph_cap_snap,
2439                                                         ci_item);
2440                         snapc = ceph_get_snap_context(capsnap->context);
2441                 } else {
2442                         BUG_ON(!ci->i_head_snapc);
2443                         snapc = ceph_get_snap_context(ci->i_head_snapc);
2444                 }
2445                 spin_unlock(&ci->i_ceph_lock);
2446
2447                 /* we might need to revert back to that point */
2448                 data = *from;
2449                 if ((iocb->ki_flags & IOCB_DIRECT) && !IS_ENCRYPTED(inode))
2450                         written = ceph_direct_read_write(iocb, &data, snapc,
2451                                                          &prealloc_cf);
2452                 else
2453                         written = ceph_sync_write(iocb, &data, pos, snapc);
2454                 if (direct_lock)
2455                         ceph_end_io_direct(inode);
2456                 else
2457                         ceph_end_io_write(inode);
2458                 if (written > 0)
2459                         iov_iter_advance(from, written);
2460                 ceph_put_snap_context(snapc);
2461         } else {
2462                 /*
2463                  * No need to acquire the i_truncate_mutex. Because
2464                  * the MDS revokes Fwb caps before sending truncate
2465                  * message to us. We can't get Fwb cap while there
2466                  * are pending vmtruncate. So write and vmtruncate
2467                  * can not run at the same time
2468                  */
2469                 written = generic_perform_write(iocb, from);
2470                 ceph_end_io_write(inode);
2471         }
2472
2473         if (written >= 0) {
2474                 int dirty;
2475
2476                 spin_lock(&ci->i_ceph_lock);
2477                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
2478                                                &prealloc_cf);
2479                 spin_unlock(&ci->i_ceph_lock);
2480                 if (dirty)
2481                         __mark_inode_dirty(inode, dirty);
2482                 if (ceph_quota_is_max_bytes_approaching(inode, iocb->ki_pos))
2483                         ceph_check_caps(ci, CHECK_CAPS_FLUSH);
2484         }
2485
2486         doutc(cl, "%p %llx.%llx %llu~%u  dropping cap refs on %s\n",
2487               inode, ceph_vinop(inode), pos, (unsigned)count,
2488               ceph_cap_string(got));
2489         ceph_put_cap_refs(ci, got);
2490
2491         if (written == -EOLDSNAPC) {
2492                 doutc(cl, "%p %llx.%llx %llu~%u" "got EOLDSNAPC, retrying\n",
2493                       inode, ceph_vinop(inode), pos, (unsigned)count);
2494                 goto retry_snap;
2495         }
2496
2497         if (written >= 0) {
2498                 if ((map_flags & CEPH_OSDMAP_NEARFULL) ||
2499                     (pool_flags & CEPH_POOL_FLAG_NEARFULL))
2500                         iocb->ki_flags |= IOCB_DSYNC;
2501                 written = generic_write_sync(iocb, written);
2502         }
2503
2504         goto out_unlocked;
2505 out_caps:
2506         ceph_put_cap_refs(ci, got);
2507 out:
2508         if (direct_lock)
2509                 ceph_end_io_direct(inode);
2510         else
2511                 ceph_end_io_write(inode);
2512 out_unlocked:
2513         ceph_free_cap_flush(prealloc_cf);
2514         return written ? written : err;
2515 }
2516
2517 /*
2518  * llseek.  be sure to verify file size on SEEK_END.
2519  */
2520 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
2521 {
2522         if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
2523                 struct inode *inode = file_inode(file);
2524                 int ret;
2525
2526                 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
2527                 if (ret < 0)
2528                         return ret;
2529         }
2530         return generic_file_llseek(file, offset, whence);
2531 }
2532
2533 static inline void ceph_zero_partial_page(
2534         struct inode *inode, loff_t offset, unsigned size)
2535 {
2536         struct page *page;
2537         pgoff_t index = offset >> PAGE_SHIFT;
2538
2539         page = find_lock_page(inode->i_mapping, index);
2540         if (page) {
2541                 wait_on_page_writeback(page);
2542                 zero_user(page, offset & (PAGE_SIZE - 1), size);
2543                 unlock_page(page);
2544                 put_page(page);
2545         }
2546 }
2547
2548 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
2549                                       loff_t length)
2550 {
2551         loff_t nearly = round_up(offset, PAGE_SIZE);
2552         if (offset < nearly) {
2553                 loff_t size = nearly - offset;
2554                 if (length < size)
2555                         size = length;
2556                 ceph_zero_partial_page(inode, offset, size);
2557                 offset += size;
2558                 length -= size;
2559         }
2560         if (length >= PAGE_SIZE) {
2561                 loff_t size = round_down(length, PAGE_SIZE);
2562                 truncate_pagecache_range(inode, offset, offset + size - 1);
2563                 offset += size;
2564                 length -= size;
2565         }
2566         if (length)
2567                 ceph_zero_partial_page(inode, offset, length);
2568 }
2569
2570 static int ceph_zero_partial_object(struct inode *inode,
2571                                     loff_t offset, loff_t *length)
2572 {
2573         struct ceph_inode_info *ci = ceph_inode(inode);
2574         struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
2575         struct ceph_osd_request *req;
2576         int ret = 0;
2577         loff_t zero = 0;
2578         int op;
2579
2580         if (ceph_inode_is_shutdown(inode))
2581                 return -EIO;
2582
2583         if (!length) {
2584                 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
2585                 length = &zero;
2586         } else {
2587                 op = CEPH_OSD_OP_ZERO;
2588         }
2589
2590         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
2591                                         ceph_vino(inode),
2592                                         offset, length,
2593                                         0, 1, op,
2594                                         CEPH_OSD_FLAG_WRITE,
2595                                         NULL, 0, 0, false);
2596         if (IS_ERR(req)) {
2597                 ret = PTR_ERR(req);
2598                 goto out;
2599         }
2600
2601         req->r_mtime = inode_get_mtime(inode);
2602         ceph_osdc_start_request(&fsc->client->osdc, req);
2603         ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
2604         if (ret == -ENOENT)
2605                 ret = 0;
2606         ceph_osdc_put_request(req);
2607
2608 out:
2609         return ret;
2610 }
2611
2612 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
2613 {
2614         int ret = 0;
2615         struct ceph_inode_info *ci = ceph_inode(inode);
2616         s32 stripe_unit = ci->i_layout.stripe_unit;
2617         s32 stripe_count = ci->i_layout.stripe_count;
2618         s32 object_size = ci->i_layout.object_size;
2619         u64 object_set_size = (u64) object_size * stripe_count;
2620         u64 nearly, t;
2621
2622         /* round offset up to next period boundary */
2623         nearly = offset + object_set_size - 1;
2624         t = nearly;
2625         nearly -= do_div(t, object_set_size);
2626
2627         while (length && offset < nearly) {
2628                 loff_t size = length;
2629                 ret = ceph_zero_partial_object(inode, offset, &size);
2630                 if (ret < 0)
2631                         return ret;
2632                 offset += size;
2633                 length -= size;
2634         }
2635         while (length >= object_set_size) {
2636                 int i;
2637                 loff_t pos = offset;
2638                 for (i = 0; i < stripe_count; ++i) {
2639                         ret = ceph_zero_partial_object(inode, pos, NULL);
2640                         if (ret < 0)
2641                                 return ret;
2642                         pos += stripe_unit;
2643                 }
2644                 offset += object_set_size;
2645                 length -= object_set_size;
2646         }
2647         while (length) {
2648                 loff_t size = length;
2649                 ret = ceph_zero_partial_object(inode, offset, &size);
2650                 if (ret < 0)
2651                         return ret;
2652                 offset += size;
2653                 length -= size;
2654         }
2655         return ret;
2656 }
2657
2658 static long ceph_fallocate(struct file *file, int mode,
2659                                 loff_t offset, loff_t length)
2660 {
2661         struct ceph_file_info *fi = file->private_data;
2662         struct inode *inode = file_inode(file);
2663         struct ceph_inode_info *ci = ceph_inode(inode);
2664         struct ceph_cap_flush *prealloc_cf;
2665         struct ceph_client *cl = ceph_inode_to_client(inode);
2666         int want, got = 0;
2667         int dirty;
2668         int ret = 0;
2669         loff_t endoff = 0;
2670         loff_t size;
2671
2672         doutc(cl, "%p %llx.%llx mode %x, offset %llu length %llu\n",
2673               inode, ceph_vinop(inode), mode, offset, length);
2674
2675         if (mode != (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2676                 return -EOPNOTSUPP;
2677
2678         if (!S_ISREG(inode->i_mode))
2679                 return -EOPNOTSUPP;
2680
2681         if (IS_ENCRYPTED(inode))
2682                 return -EOPNOTSUPP;
2683
2684         prealloc_cf = ceph_alloc_cap_flush();
2685         if (!prealloc_cf)
2686                 return -ENOMEM;
2687
2688         inode_lock(inode);
2689
2690         if (ceph_snap(inode) != CEPH_NOSNAP) {
2691                 ret = -EROFS;
2692                 goto unlock;
2693         }
2694
2695         size = i_size_read(inode);
2696
2697         /* Are we punching a hole beyond EOF? */
2698         if (offset >= size)
2699                 goto unlock;
2700         if ((offset + length) > size)
2701                 length = size - offset;
2702
2703         if (fi->fmode & CEPH_FILE_MODE_LAZY)
2704                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
2705         else
2706                 want = CEPH_CAP_FILE_BUFFER;
2707
2708         ret = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, endoff, &got);
2709         if (ret < 0)
2710                 goto unlock;
2711
2712         ret = file_modified(file);
2713         if (ret)
2714                 goto put_caps;
2715
2716         filemap_invalidate_lock(inode->i_mapping);
2717         ceph_fscache_invalidate(inode, false);
2718         ceph_zero_pagecache_range(inode, offset, length);
2719         ret = ceph_zero_objects(inode, offset, length);
2720
2721         if (!ret) {
2722                 spin_lock(&ci->i_ceph_lock);
2723                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
2724                                                &prealloc_cf);
2725                 spin_unlock(&ci->i_ceph_lock);
2726                 if (dirty)
2727                         __mark_inode_dirty(inode, dirty);
2728         }
2729         filemap_invalidate_unlock(inode->i_mapping);
2730
2731 put_caps:
2732         ceph_put_cap_refs(ci, got);
2733 unlock:
2734         inode_unlock(inode);
2735         ceph_free_cap_flush(prealloc_cf);
2736         return ret;
2737 }
2738
2739 /*
2740  * This function tries to get FILE_WR capabilities for dst_ci and FILE_RD for
2741  * src_ci.  Two attempts are made to obtain both caps, and an error is return if
2742  * this fails; zero is returned on success.
2743  */
2744 static int get_rd_wr_caps(struct file *src_filp, int *src_got,
2745                           struct file *dst_filp,
2746                           loff_t dst_endoff, int *dst_got)
2747 {
2748         int ret = 0;
2749         bool retrying = false;
2750
2751 retry_caps:
2752         ret = ceph_get_caps(dst_filp, CEPH_CAP_FILE_WR, CEPH_CAP_FILE_BUFFER,
2753                             dst_endoff, dst_got);
2754         if (ret < 0)
2755                 return ret;
2756
2757         /*
2758          * Since we're already holding the FILE_WR capability for the dst file,
2759          * we would risk a deadlock by using ceph_get_caps.  Thus, we'll do some
2760          * retry dance instead to try to get both capabilities.
2761          */
2762         ret = ceph_try_get_caps(file_inode(src_filp),
2763                                 CEPH_CAP_FILE_RD, CEPH_CAP_FILE_SHARED,
2764                                 false, src_got);
2765         if (ret <= 0) {
2766                 /* Start by dropping dst_ci caps and getting src_ci caps */
2767                 ceph_put_cap_refs(ceph_inode(file_inode(dst_filp)), *dst_got);
2768                 if (retrying) {
2769                         if (!ret)
2770                                 /* ceph_try_get_caps masks EAGAIN */
2771                                 ret = -EAGAIN;
2772                         return ret;
2773                 }
2774                 ret = ceph_get_caps(src_filp, CEPH_CAP_FILE_RD,
2775                                     CEPH_CAP_FILE_SHARED, -1, src_got);
2776                 if (ret < 0)
2777                         return ret;
2778                 /*... drop src_ci caps too, and retry */
2779                 ceph_put_cap_refs(ceph_inode(file_inode(src_filp)), *src_got);
2780                 retrying = true;
2781                 goto retry_caps;
2782         }
2783         return ret;
2784 }
2785
2786 static void put_rd_wr_caps(struct ceph_inode_info *src_ci, int src_got,
2787                            struct ceph_inode_info *dst_ci, int dst_got)
2788 {
2789         ceph_put_cap_refs(src_ci, src_got);
2790         ceph_put_cap_refs(dst_ci, dst_got);
2791 }
2792
2793 /*
2794  * This function does several size-related checks, returning an error if:
2795  *  - source file is smaller than off+len
2796  *  - destination file size is not OK (inode_newsize_ok())
2797  *  - max bytes quotas is exceeded
2798  */
2799 static int is_file_size_ok(struct inode *src_inode, struct inode *dst_inode,
2800                            loff_t src_off, loff_t dst_off, size_t len)
2801 {
2802         struct ceph_client *cl = ceph_inode_to_client(src_inode);
2803         loff_t size, endoff;
2804
2805         size = i_size_read(src_inode);
2806         /*
2807          * Don't copy beyond source file EOF.  Instead of simply setting length
2808          * to (size - src_off), just drop to VFS default implementation, as the
2809          * local i_size may be stale due to other clients writing to the source
2810          * inode.
2811          */
2812         if (src_off + len > size) {
2813                 doutc(cl, "Copy beyond EOF (%llu + %zu > %llu)\n", src_off,
2814                       len, size);
2815                 return -EOPNOTSUPP;
2816         }
2817         size = i_size_read(dst_inode);
2818
2819         endoff = dst_off + len;
2820         if (inode_newsize_ok(dst_inode, endoff))
2821                 return -EOPNOTSUPP;
2822
2823         if (ceph_quota_is_max_bytes_exceeded(dst_inode, endoff))
2824                 return -EDQUOT;
2825
2826         return 0;
2827 }
2828
2829 static struct ceph_osd_request *
2830 ceph_alloc_copyfrom_request(struct ceph_osd_client *osdc,
2831                             u64 src_snapid,
2832                             struct ceph_object_id *src_oid,
2833                             struct ceph_object_locator *src_oloc,
2834                             struct ceph_object_id *dst_oid,
2835                             struct ceph_object_locator *dst_oloc,
2836                             u32 truncate_seq, u64 truncate_size)
2837 {
2838         struct ceph_osd_request *req;
2839         int ret;
2840         u32 src_fadvise_flags =
2841                 CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2842                 CEPH_OSD_OP_FLAG_FADVISE_NOCACHE;
2843         u32 dst_fadvise_flags =
2844                 CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2845                 CEPH_OSD_OP_FLAG_FADVISE_DONTNEED;
2846
2847         req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_KERNEL);
2848         if (!req)
2849                 return ERR_PTR(-ENOMEM);
2850
2851         req->r_flags = CEPH_OSD_FLAG_WRITE;
2852
2853         ceph_oloc_copy(&req->r_t.base_oloc, dst_oloc);
2854         ceph_oid_copy(&req->r_t.base_oid, dst_oid);
2855
2856         ret = osd_req_op_copy_from_init(req, src_snapid, 0,
2857                                         src_oid, src_oloc,
2858                                         src_fadvise_flags,
2859                                         dst_fadvise_flags,
2860                                         truncate_seq,
2861                                         truncate_size,
2862                                         CEPH_OSD_COPY_FROM_FLAG_TRUNCATE_SEQ);
2863         if (ret)
2864                 goto out;
2865
2866         ret = ceph_osdc_alloc_messages(req, GFP_KERNEL);
2867         if (ret)
2868                 goto out;
2869
2870         return req;
2871
2872 out:
2873         ceph_osdc_put_request(req);
2874         return ERR_PTR(ret);
2875 }
2876
2877 static ssize_t ceph_do_objects_copy(struct ceph_inode_info *src_ci, u64 *src_off,
2878                                     struct ceph_inode_info *dst_ci, u64 *dst_off,
2879                                     struct ceph_fs_client *fsc,
2880                                     size_t len, unsigned int flags)
2881 {
2882         struct ceph_object_locator src_oloc, dst_oloc;
2883         struct ceph_object_id src_oid, dst_oid;
2884         struct ceph_osd_client *osdc;
2885         struct ceph_osd_request *req;
2886         size_t bytes = 0;
2887         u64 src_objnum, src_objoff, dst_objnum, dst_objoff;
2888         u32 src_objlen, dst_objlen;
2889         u32 object_size = src_ci->i_layout.object_size;
2890         struct ceph_client *cl = fsc->client;
2891         int ret;
2892
2893         src_oloc.pool = src_ci->i_layout.pool_id;
2894         src_oloc.pool_ns = ceph_try_get_string(src_ci->i_layout.pool_ns);
2895         dst_oloc.pool = dst_ci->i_layout.pool_id;
2896         dst_oloc.pool_ns = ceph_try_get_string(dst_ci->i_layout.pool_ns);
2897         osdc = &fsc->client->osdc;
2898
2899         while (len >= object_size) {
2900                 ceph_calc_file_object_mapping(&src_ci->i_layout, *src_off,
2901                                               object_size, &src_objnum,
2902                                               &src_objoff, &src_objlen);
2903                 ceph_calc_file_object_mapping(&dst_ci->i_layout, *dst_off,
2904                                               object_size, &dst_objnum,
2905                                               &dst_objoff, &dst_objlen);
2906                 ceph_oid_init(&src_oid);
2907                 ceph_oid_printf(&src_oid, "%llx.%08llx",
2908                                 src_ci->i_vino.ino, src_objnum);
2909                 ceph_oid_init(&dst_oid);
2910                 ceph_oid_printf(&dst_oid, "%llx.%08llx",
2911                                 dst_ci->i_vino.ino, dst_objnum);
2912                 /* Do an object remote copy */
2913                 req = ceph_alloc_copyfrom_request(osdc, src_ci->i_vino.snap,
2914                                                   &src_oid, &src_oloc,
2915                                                   &dst_oid, &dst_oloc,
2916                                                   dst_ci->i_truncate_seq,
2917                                                   dst_ci->i_truncate_size);
2918                 if (IS_ERR(req))
2919                         ret = PTR_ERR(req);
2920                 else {
2921                         ceph_osdc_start_request(osdc, req);
2922                         ret = ceph_osdc_wait_request(osdc, req);
2923                         ceph_update_copyfrom_metrics(&fsc->mdsc->metric,
2924                                                      req->r_start_latency,
2925                                                      req->r_end_latency,
2926                                                      object_size, ret);
2927                         ceph_osdc_put_request(req);
2928                 }
2929                 if (ret) {
2930                         if (ret == -EOPNOTSUPP) {
2931                                 fsc->have_copy_from2 = false;
2932                                 pr_notice_client(cl,
2933                                         "OSDs don't support copy-from2; disabling copy offload\n");
2934                         }
2935                         doutc(cl, "returned %d\n", ret);
2936                         if (!bytes)
2937                                 bytes = ret;
2938                         goto out;
2939                 }
2940                 len -= object_size;
2941                 bytes += object_size;
2942                 *src_off += object_size;
2943                 *dst_off += object_size;
2944         }
2945
2946 out:
2947         ceph_oloc_destroy(&src_oloc);
2948         ceph_oloc_destroy(&dst_oloc);
2949         return bytes;
2950 }
2951
2952 static ssize_t __ceph_copy_file_range(struct file *src_file, loff_t src_off,
2953                                       struct file *dst_file, loff_t dst_off,
2954                                       size_t len, unsigned int flags)
2955 {
2956         struct inode *src_inode = file_inode(src_file);
2957         struct inode *dst_inode = file_inode(dst_file);
2958         struct ceph_inode_info *src_ci = ceph_inode(src_inode);
2959         struct ceph_inode_info *dst_ci = ceph_inode(dst_inode);
2960         struct ceph_cap_flush *prealloc_cf;
2961         struct ceph_fs_client *src_fsc = ceph_inode_to_fs_client(src_inode);
2962         struct ceph_client *cl = src_fsc->client;
2963         loff_t size;
2964         ssize_t ret = -EIO, bytes;
2965         u64 src_objnum, dst_objnum, src_objoff, dst_objoff;
2966         u32 src_objlen, dst_objlen;
2967         int src_got = 0, dst_got = 0, err, dirty;
2968
2969         if (src_inode->i_sb != dst_inode->i_sb) {
2970                 struct ceph_fs_client *dst_fsc = ceph_inode_to_fs_client(dst_inode);
2971
2972                 if (ceph_fsid_compare(&src_fsc->client->fsid,
2973                                       &dst_fsc->client->fsid)) {
2974                         dout("Copying files across clusters: src: %pU dst: %pU\n",
2975                              &src_fsc->client->fsid, &dst_fsc->client->fsid);
2976                         return -EXDEV;
2977                 }
2978         }
2979         if (ceph_snap(dst_inode) != CEPH_NOSNAP)
2980                 return -EROFS;
2981
2982         /*
2983          * Some of the checks below will return -EOPNOTSUPP, which will force a
2984          * fallback to the default VFS copy_file_range implementation.  This is
2985          * desirable in several cases (for ex, the 'len' is smaller than the
2986          * size of the objects, or in cases where that would be more
2987          * efficient).
2988          */
2989
2990         if (ceph_test_mount_opt(src_fsc, NOCOPYFROM))
2991                 return -EOPNOTSUPP;
2992
2993         if (!src_fsc->have_copy_from2)
2994                 return -EOPNOTSUPP;
2995
2996         /*
2997          * Striped file layouts require that we copy partial objects, but the
2998          * OSD copy-from operation only supports full-object copies.  Limit
2999          * this to non-striped file layouts for now.
3000          */
3001         if ((src_ci->i_layout.stripe_unit != dst_ci->i_layout.stripe_unit) ||
3002             (src_ci->i_layout.stripe_count != 1) ||
3003             (dst_ci->i_layout.stripe_count != 1) ||
3004             (src_ci->i_layout.object_size != dst_ci->i_layout.object_size)) {
3005                 doutc(cl, "Invalid src/dst files layout\n");
3006                 return -EOPNOTSUPP;
3007         }
3008
3009         /* Every encrypted inode gets its own key, so we can't offload them */
3010         if (IS_ENCRYPTED(src_inode) || IS_ENCRYPTED(dst_inode))
3011                 return -EOPNOTSUPP;
3012
3013         if (len < src_ci->i_layout.object_size)
3014                 return -EOPNOTSUPP; /* no remote copy will be done */
3015
3016         prealloc_cf = ceph_alloc_cap_flush();
3017         if (!prealloc_cf)
3018                 return -ENOMEM;
3019
3020         /* Start by sync'ing the source and destination files */
3021         ret = file_write_and_wait_range(src_file, src_off, (src_off + len));
3022         if (ret < 0) {
3023                 doutc(cl, "failed to write src file (%zd)\n", ret);
3024                 goto out;
3025         }
3026         ret = file_write_and_wait_range(dst_file, dst_off, (dst_off + len));
3027         if (ret < 0) {
3028                 doutc(cl, "failed to write dst file (%zd)\n", ret);
3029                 goto out;
3030         }
3031
3032         /*
3033          * We need FILE_WR caps for dst_ci and FILE_RD for src_ci as other
3034          * clients may have dirty data in their caches.  And OSDs know nothing
3035          * about caps, so they can't safely do the remote object copies.
3036          */
3037         err = get_rd_wr_caps(src_file, &src_got,
3038                              dst_file, (dst_off + len), &dst_got);
3039         if (err < 0) {
3040                 doutc(cl, "get_rd_wr_caps returned %d\n", err);
3041                 ret = -EOPNOTSUPP;
3042                 goto out;
3043         }
3044
3045         ret = is_file_size_ok(src_inode, dst_inode, src_off, dst_off, len);
3046         if (ret < 0)
3047                 goto out_caps;
3048
3049         /* Drop dst file cached pages */
3050         ceph_fscache_invalidate(dst_inode, false);
3051         ret = invalidate_inode_pages2_range(dst_inode->i_mapping,
3052                                             dst_off >> PAGE_SHIFT,
3053                                             (dst_off + len) >> PAGE_SHIFT);
3054         if (ret < 0) {
3055                 doutc(cl, "Failed to invalidate inode pages (%zd)\n",
3056                             ret);
3057                 ret = 0; /* XXX */
3058         }
3059         ceph_calc_file_object_mapping(&src_ci->i_layout, src_off,
3060                                       src_ci->i_layout.object_size,
3061                                       &src_objnum, &src_objoff, &src_objlen);
3062         ceph_calc_file_object_mapping(&dst_ci->i_layout, dst_off,
3063                                       dst_ci->i_layout.object_size,
3064                                       &dst_objnum, &dst_objoff, &dst_objlen);
3065         /* object-level offsets need to the same */
3066         if (src_objoff != dst_objoff) {
3067                 ret = -EOPNOTSUPP;
3068                 goto out_caps;
3069         }
3070
3071         /*
3072          * Do a manual copy if the object offset isn't object aligned.
3073          * 'src_objlen' contains the bytes left until the end of the object,
3074          * starting at the src_off
3075          */
3076         if (src_objoff) {
3077                 doutc(cl, "Initial partial copy of %u bytes\n", src_objlen);
3078
3079                 /*
3080                  * we need to temporarily drop all caps as we'll be calling
3081                  * {read,write}_iter, which will get caps again.
3082                  */
3083                 put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
3084                 ret = splice_file_range(src_file, &src_off, dst_file, &dst_off,
3085                                         src_objlen);
3086                 /* Abort on short copies or on error */
3087                 if (ret < (long)src_objlen) {
3088                         doutc(cl, "Failed partial copy (%zd)\n", ret);
3089                         goto out;
3090                 }
3091                 len -= ret;
3092                 err = get_rd_wr_caps(src_file, &src_got,
3093                                      dst_file, (dst_off + len), &dst_got);
3094                 if (err < 0)
3095                         goto out;
3096                 err = is_file_size_ok(src_inode, dst_inode,
3097                                       src_off, dst_off, len);
3098                 if (err < 0)
3099                         goto out_caps;
3100         }
3101
3102         size = i_size_read(dst_inode);
3103         bytes = ceph_do_objects_copy(src_ci, &src_off, dst_ci, &dst_off,
3104                                      src_fsc, len, flags);
3105         if (bytes <= 0) {
3106                 if (!ret)
3107                         ret = bytes;
3108                 goto out_caps;
3109         }
3110         doutc(cl, "Copied %zu bytes out of %zu\n", bytes, len);
3111         len -= bytes;
3112         ret += bytes;
3113
3114         file_update_time(dst_file);
3115         inode_inc_iversion_raw(dst_inode);
3116
3117         if (dst_off > size) {
3118                 /* Let the MDS know about dst file size change */
3119                 if (ceph_inode_set_size(dst_inode, dst_off) ||
3120                     ceph_quota_is_max_bytes_approaching(dst_inode, dst_off))
3121                         ceph_check_caps(dst_ci, CHECK_CAPS_AUTHONLY | CHECK_CAPS_FLUSH);
3122         }
3123         /* Mark Fw dirty */
3124         spin_lock(&dst_ci->i_ceph_lock);
3125         dirty = __ceph_mark_dirty_caps(dst_ci, CEPH_CAP_FILE_WR, &prealloc_cf);
3126         spin_unlock(&dst_ci->i_ceph_lock);
3127         if (dirty)
3128                 __mark_inode_dirty(dst_inode, dirty);
3129
3130 out_caps:
3131         put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
3132
3133         /*
3134          * Do the final manual copy if we still have some bytes left, unless
3135          * there were errors in remote object copies (len >= object_size).
3136          */
3137         if (len && (len < src_ci->i_layout.object_size)) {
3138                 doutc(cl, "Final partial copy of %zu bytes\n", len);
3139                 bytes = splice_file_range(src_file, &src_off, dst_file,
3140                                           &dst_off, len);
3141                 if (bytes > 0)
3142                         ret += bytes;
3143                 else
3144                         doutc(cl, "Failed partial copy (%zd)\n", bytes);
3145         }
3146
3147 out:
3148         ceph_free_cap_flush(prealloc_cf);
3149
3150         return ret;
3151 }
3152
3153 static ssize_t ceph_copy_file_range(struct file *src_file, loff_t src_off,
3154                                     struct file *dst_file, loff_t dst_off,
3155                                     size_t len, unsigned int flags)
3156 {
3157         ssize_t ret;
3158
3159         ret = __ceph_copy_file_range(src_file, src_off, dst_file, dst_off,
3160                                      len, flags);
3161
3162         if (ret == -EOPNOTSUPP || ret == -EXDEV)
3163                 ret = splice_copy_file_range(src_file, src_off, dst_file,
3164                                              dst_off, len);
3165         return ret;
3166 }
3167
3168 const struct file_operations ceph_file_fops = {
3169         .open = ceph_open,
3170         .release = ceph_release,
3171         .llseek = ceph_llseek,
3172         .read_iter = ceph_read_iter,
3173         .write_iter = ceph_write_iter,
3174         .mmap = ceph_mmap,
3175         .fsync = ceph_fsync,
3176         .lock = ceph_lock,
3177         .setlease = simple_nosetlease,
3178         .flock = ceph_flock,
3179         .splice_read = ceph_splice_read,
3180         .splice_write = iter_file_splice_write,
3181         .unlocked_ioctl = ceph_ioctl,
3182         .compat_ioctl = compat_ptr_ioctl,
3183         .fallocate      = ceph_fallocate,
3184         .copy_file_range = ceph_copy_file_range,
3185 };