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