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_pages(iter, pages, maxsize - size,
  99                                            ITER_GET_BVECS_PAGES, &start);
 100                 if (bytes < 0)
 101                         return size ?: bytes;
 102
 103                 iov_iter_advance(iter, bytes);
 104                 size += bytes;
 105
 106                 for ( ; bytes; idx++, bvec_idx++) {
 107                         struct bio_vec bv = {
 108                                 .bv_page = pages[idx],
 109                                 .bv_len = min_t(int, bytes, PAGE_SIZE - start),
 110                                 .bv_offset = start,
 111                         };
 112
 113                         bvecs[bvec_idx] = bv;
 114                         bytes -= bv.bv_len;
 115                         start = 0;
 116                 }
 117         }
 118
 119         return size;
 120 }
 121
 122 /*
 123  * iov_iter_get_pages() only considers one iov_iter segment, no matter
 124  * what maxsize or maxpages are given.  For ITER_BVEC that is a single
 125  * page.
 126  *
 127  * Attempt to get up to @maxsize bytes worth of pages from @iter.
 128  * Return the number of bytes in the created bio_vec array, or an error.
 129  */
 130 static ssize_t iter_get_bvecs_alloc(struct iov_iter *iter, size_t maxsize,
 131                                     struct bio_vec **bvecs, int *num_bvecs)
 132 {
 133         struct bio_vec *bv;
 134         size_t orig_count = iov_iter_count(iter);
 135         ssize_t bytes;
 136         int npages;
 137
 138         iov_iter_truncate(iter, maxsize);
 139         npages = iov_iter_npages(iter, INT_MAX);
 140         iov_iter_reexpand(iter, orig_count);
 141
 142         /*
 143          * __iter_get_bvecs() may populate only part of the array -- zero it
 144          * out.
 145          */
 146         bv = kvmalloc_array(npages, sizeof(*bv), GFP_KERNEL | __GFP_ZERO);
 147         if (!bv)
 148                 return -ENOMEM;
 149
 150         bytes = __iter_get_bvecs(iter, maxsize, bv);
 151         if (bytes < 0) {
 152                 /*
 153                  * No pages were pinned -- just free the array.
 154                  */
 155                 kvfree(bv);
 156                 return bytes;
 157         }
 158
 159         *bvecs = bv;
 160         *num_bvecs = npages;
 161         return bytes;
 162 }
 163
 164 static void put_bvecs(struct bio_vec *bvecs, int num_bvecs, bool should_dirty)
 165 {
 166         int i;
 167
 168         for (i = 0; i < num_bvecs; i++) {
 169                 if (bvecs[i].bv_page) {
 170                         if (should_dirty)
 171                                 set_page_dirty_lock(bvecs[i].bv_page);
 172                         put_page(bvecs[i].bv_page);
 173                 }
 174         }
 175         kvfree(bvecs);
 176 }
 177
 178 /*
 179  * Prepare an open request.  Preallocate ceph_cap to avoid an
 180  * inopportune ENOMEM later.
 181  */
 182 static struct ceph_mds_request *
 183 prepare_open_request(struct super_block *sb, int flags, int create_mode)
 184 {
 185         struct ceph_fs_client *fsc = ceph_sb_to_client(sb);
 186         struct ceph_mds_client *mdsc = fsc->mdsc;
 187         struct ceph_mds_request *req;
 188         int want_auth = USE_ANY_MDS;
 189         int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
 190
 191         if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
 192                 want_auth = USE_AUTH_MDS;
 193
 194         req = ceph_mdsc_create_request(mdsc, op, want_auth);
 195         if (IS_ERR(req))
 196                 goto out;
 197         req->r_fmode = ceph_flags_to_mode(flags);
 198         req->r_args.open.flags = ceph_flags_sys2wire(flags);
 199         req->r_args.open.mode = cpu_to_le32(create_mode);
 200 out:
 201         return req;
 202 }
 203
 204 static int ceph_init_file_info(struct inode *inode, struct file *file,
 205                                         int fmode, bool isdir)
 206 {
 207         struct ceph_inode_info *ci = ceph_inode(inode);
 208         struct ceph_file_info *fi;
 209
 210         dout("%s %p %p 0%o (%s)\n", __func__, inode, file,
 211                         inode->i_mode, isdir ? "dir" : "regular");
 212         BUG_ON(inode->i_fop->release != ceph_release);
 213
 214         if (isdir) {
 215                 struct ceph_dir_file_info *dfi =
 216                         kmem_cache_zalloc(ceph_dir_file_cachep, GFP_KERNEL);
 217                 if (!dfi)
 218                         return -ENOMEM;
 219
 220                 file->private_data = dfi;
 221                 fi = &dfi->file_info;
 222                 dfi->next_offset = 2;
 223                 dfi->readdir_cache_idx = -1;
 224         } else {
 225                 fi = kmem_cache_zalloc(ceph_file_cachep, GFP_KERNEL);
 226                 if (!fi)
 227                         return -ENOMEM;
 228
 229                 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->meta_err = errseq_sample(&ci->i_meta_err);
 238         fi->filp_gen = READ_ONCE(ceph_inode_to_client(inode)->filp_gen);
 239
 240         return 0;
 241 }
 242
 243 /*
 244  * initialize private struct file data.
 245  * if we fail, clean up by dropping fmode reference on the ceph_inode
 246  */
 247 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
 248 {
 249         int ret = 0;
 250
 251         switch (inode->i_mode & S_IFMT) {
 252         case S_IFREG:
 253                 ceph_fscache_register_inode_cookie(inode);
 254                 ceph_fscache_file_set_cookie(inode, file);
 255                 fallthrough;
 256         case S_IFDIR:
 257                 ret = ceph_init_file_info(inode, file, fmode,
 258                                                 S_ISDIR(inode->i_mode));
 259                 break;
 260
 261         case S_IFLNK:
 262                 dout("init_file %p %p 0%o (symlink)\n", inode, file,
 263                      inode->i_mode);
 264                 break;
 265
 266         default:
 267                 dout("init_file %p %p 0%o (special)\n", inode, file,
 268                      inode->i_mode);
 269                 /*
 270                  * we need to drop the open ref now, since we don't
 271                  * have .release set to ceph_release.
 272                  */
 273                 BUG_ON(inode->i_fop->release == ceph_release);
 274
 275                 /* call the proper open fop */
 276                 ret = inode->i_fop->open(inode, file);
 277         }
 278         return ret;
 279 }
 280
 281 /*
 282  * try renew caps after session gets killed.
 283  */
 284 int ceph_renew_caps(struct inode *inode, int fmode)
 285 {
 286         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
 287         struct ceph_inode_info *ci = ceph_inode(inode);
 288         struct ceph_mds_request *req;
 289         int err, flags, wanted;
 290
 291         spin_lock(&ci->i_ceph_lock);
 292         __ceph_touch_fmode(ci, mdsc, fmode);
 293         wanted = __ceph_caps_file_wanted(ci);
 294         if (__ceph_is_any_real_caps(ci) &&
 295             (!(wanted & CEPH_CAP_ANY_WR) || ci->i_auth_cap)) {
 296                 int issued = __ceph_caps_issued(ci, NULL);
 297                 spin_unlock(&ci->i_ceph_lock);
 298                 dout("renew caps %p want %s issued %s updating mds_wanted\n",
 299                      inode, ceph_cap_string(wanted), ceph_cap_string(issued));
 300                 ceph_check_caps(ci, 0, NULL);
 301                 return 0;
 302         }
 303         spin_unlock(&ci->i_ceph_lock);
 304
 305         flags = 0;
 306         if ((wanted & CEPH_CAP_FILE_RD) && (wanted & CEPH_CAP_FILE_WR))
 307                 flags = O_RDWR;
 308         else if (wanted & CEPH_CAP_FILE_RD)
 309                 flags = O_RDONLY;
 310         else if (wanted & CEPH_CAP_FILE_WR)
 311                 flags = O_WRONLY;
 312 #ifdef O_LAZY
 313         if (wanted & CEPH_CAP_FILE_LAZYIO)
 314                 flags |= O_LAZY;
 315 #endif
 316
 317         req = prepare_open_request(inode->i_sb, flags, 0);
 318         if (IS_ERR(req)) {
 319                 err = PTR_ERR(req);
 320                 goto out;
 321         }
 322
 323         req->r_inode = inode;
 324         ihold(inode);
 325         req->r_num_caps = 1;
 326
 327         err = ceph_mdsc_do_request(mdsc, NULL, req);
 328         ceph_mdsc_put_request(req);
 329 out:
 330         dout("renew caps %p open result=%d\n", inode, err);
 331         return err < 0 ? err : 0;
 332 }
 333
 334 /*
 335  * If we already have the requisite capabilities, we can satisfy
 336  * the open request locally (no need to request new caps from the
 337  * MDS).  We do, however, need to inform the MDS (asynchronously)
 338  * if our wanted caps set expands.
 339  */
 340 int ceph_open(struct inode *inode, struct file *file)
 341 {
 342         struct ceph_inode_info *ci = ceph_inode(inode);
 343         struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
 344         struct ceph_mds_client *mdsc = fsc->mdsc;
 345         struct ceph_mds_request *req;
 346         struct ceph_file_info *fi = file->private_data;
 347         int err;
 348         int flags, fmode, wanted;
 349
 350         if (fi) {
 351                 dout("open file %p is already opened\n", file);
 352                 return 0;
 353         }
 354
 355         /* filter out O_CREAT|O_EXCL; vfs did that already.  yuck. */
 356         flags = file->f_flags & ~(O_CREAT|O_EXCL);
 357         if (S_ISDIR(inode->i_mode))
 358                 flags = O_DIRECTORY;  /* mds likes to know */
 359
 360         dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
 361              ceph_vinop(inode), file, flags, file->f_flags);
 362         fmode = ceph_flags_to_mode(flags);
 363         wanted = ceph_caps_for_mode(fmode);
 364
 365         /* snapped files are read-only */
 366         if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
 367                 return -EROFS;
 368
 369         /* trivially open snapdir */
 370         if (ceph_snap(inode) == CEPH_SNAPDIR) {
 371                 return ceph_init_file(inode, file, fmode);
 372         }
 373
 374         /*
 375          * No need to block if we have caps on the auth MDS (for
 376          * write) or any MDS (for read).  Update wanted set
 377          * asynchronously.
 378          */
 379         spin_lock(&ci->i_ceph_lock);
 380         if (__ceph_is_any_real_caps(ci) &&
 381             (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
 382                 int mds_wanted = __ceph_caps_mds_wanted(ci, true);
 383                 int issued = __ceph_caps_issued(ci, NULL);
 384
 385                 dout("open %p fmode %d want %s issued %s using existing\n",
 386                      inode, fmode, ceph_cap_string(wanted),
 387                      ceph_cap_string(issued));
 388                 __ceph_touch_fmode(ci, mdsc, fmode);
 389                 spin_unlock(&ci->i_ceph_lock);
 390
 391                 /* adjust wanted? */
 392                 if ((issued & wanted) != wanted &&
 393                     (mds_wanted & wanted) != wanted &&
 394                     ceph_snap(inode) != CEPH_SNAPDIR)
 395                         ceph_check_caps(ci, 0, NULL);
 396
 397                 return ceph_init_file(inode, file, fmode);
 398         } else if (ceph_snap(inode) != CEPH_NOSNAP &&
 399                    (ci->i_snap_caps & wanted) == wanted) {
 400                 __ceph_touch_fmode(ci, mdsc, fmode);
 401                 spin_unlock(&ci->i_ceph_lock);
 402                 return ceph_init_file(inode, file, fmode);
 403         }
 404
 405         spin_unlock(&ci->i_ceph_lock);
 406
 407         dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
 408         req = prepare_open_request(inode->i_sb, flags, 0);
 409         if (IS_ERR(req)) {
 410                 err = PTR_ERR(req);
 411                 goto out;
 412         }
 413         req->r_inode = inode;
 414         ihold(inode);
 415
 416         req->r_num_caps = 1;
 417         err = ceph_mdsc_do_request(mdsc, NULL, req);
 418         if (!err)
 419                 err = ceph_init_file(inode, file, req->r_fmode);
 420         ceph_mdsc_put_request(req);
 421         dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
 422 out:
 423         return err;
 424 }
 425
 426 /* Clone the layout from a synchronous create, if the dir now has Dc caps */
 427 static void
 428 cache_file_layout(struct inode *dst, struct inode *src)
 429 {
 430         struct ceph_inode_info *cdst = ceph_inode(dst);
 431         struct ceph_inode_info *csrc = ceph_inode(src);
 432
 433         spin_lock(&cdst->i_ceph_lock);
 434         if ((__ceph_caps_issued(cdst, NULL) & CEPH_CAP_DIR_CREATE) &&
 435             !ceph_file_layout_is_valid(&cdst->i_cached_layout)) {
 436                 memcpy(&cdst->i_cached_layout, &csrc->i_layout,
 437                         sizeof(cdst->i_cached_layout));
 438                 rcu_assign_pointer(cdst->i_cached_layout.pool_ns,
 439                                    ceph_try_get_string(csrc->i_layout.pool_ns));
 440         }
 441         spin_unlock(&cdst->i_ceph_lock);
 442 }
 443
 444 /*
 445  * Try to set up an async create. We need caps, a file layout, and inode number,
 446  * and either a lease on the dentry or complete dir info. If any of those
 447  * criteria are not satisfied, then return false and the caller can go
 448  * synchronous.
 449  */
 450 static int try_prep_async_create(struct inode *dir, struct dentry *dentry,
 451                                  struct ceph_file_layout *lo, u64 *pino)
 452 {
 453         struct ceph_inode_info *ci = ceph_inode(dir);
 454         struct ceph_dentry_info *di = ceph_dentry(dentry);
 455         int got = 0, want = CEPH_CAP_FILE_EXCL | CEPH_CAP_DIR_CREATE;
 456         u64 ino;
 457
 458         spin_lock(&ci->i_ceph_lock);
 459         /* No auth cap means no chance for Dc caps */
 460         if (!ci->i_auth_cap)
 461                 goto no_async;
 462
 463         /* Any delegated inos? */
 464         if (xa_empty(&ci->i_auth_cap->session->s_delegated_inos))
 465                 goto no_async;
 466
 467         if (!ceph_file_layout_is_valid(&ci->i_cached_layout))
 468                 goto no_async;
 469
 470         if ((__ceph_caps_issued(ci, NULL) & want) != want)
 471                 goto no_async;
 472
 473         if (d_in_lookup(dentry)) {
 474                 if (!__ceph_dir_is_complete(ci))
 475                         goto no_async;
 476                 spin_lock(&dentry->d_lock);
 477                 di->lease_shared_gen = atomic_read(&ci->i_shared_gen);
 478                 spin_unlock(&dentry->d_lock);
 479         } else if (atomic_read(&ci->i_shared_gen) !=
 480                    READ_ONCE(di->lease_shared_gen)) {
 481                 goto no_async;
 482         }
 483
 484         ino = ceph_get_deleg_ino(ci->i_auth_cap->session);
 485         if (!ino)
 486                 goto no_async;
 487
 488         *pino = ino;
 489         ceph_take_cap_refs(ci, want, false);
 490         memcpy(lo, &ci->i_cached_layout, sizeof(*lo));
 491         rcu_assign_pointer(lo->pool_ns,
 492                            ceph_try_get_string(ci->i_cached_layout.pool_ns));
 493         got = want;
 494 no_async:
 495         spin_unlock(&ci->i_ceph_lock);
 496         return got;
 497 }
 498
 499 static void restore_deleg_ino(struct inode *dir, u64 ino)
 500 {
 501         struct ceph_inode_info *ci = ceph_inode(dir);
 502         struct ceph_mds_session *s = NULL;
 503
 504         spin_lock(&ci->i_ceph_lock);
 505         if (ci->i_auth_cap)
 506                 s = ceph_get_mds_session(ci->i_auth_cap->session);
 507         spin_unlock(&ci->i_ceph_lock);
 508         if (s) {
 509                 int err = ceph_restore_deleg_ino(s, ino);
 510                 if (err)
 511                         pr_warn("ceph: unable to restore delegated ino 0x%llx to session: %d\n",
 512                                 ino, err);
 513                 ceph_put_mds_session(s);
 514         }
 515 }
 516
 517 static void ceph_async_create_cb(struct ceph_mds_client *mdsc,
 518                                  struct ceph_mds_request *req)
 519 {
 520         int result = req->r_err ? req->r_err :
 521                         le32_to_cpu(req->r_reply_info.head->result);
 522
 523         if (result == -EJUKEBOX)
 524                 goto out;
 525
 526         mapping_set_error(req->r_parent->i_mapping, result);
 527
 528         if (result) {
 529                 struct dentry *dentry = req->r_dentry;
 530                 int pathlen = 0;
 531                 u64 base = 0;
 532                 char *path = ceph_mdsc_build_path(req->r_dentry, &pathlen,
 533                                                   &base, 0);
 534
 535                 ceph_dir_clear_complete(req->r_parent);
 536                 if (!d_unhashed(dentry))
 537                         d_drop(dentry);
 538
 539                 /* FIXME: start returning I/O errors on all accesses? */
 540                 pr_warn("ceph: async create failure path=(%llx)%s result=%d!\n",
 541                         base, IS_ERR(path) ? "<<bad>>" : path, result);
 542                 ceph_mdsc_free_path(path, pathlen);
 543         }
 544
 545         if (req->r_target_inode) {
 546                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
 547                 u64 ino = ceph_vino(req->r_target_inode).ino;
 548
 549                 if (req->r_deleg_ino != ino)
 550                         pr_warn("%s: inode number mismatch! err=%d deleg_ino=0x%llx target=0x%llx\n",
 551                                 __func__, req->r_err, req->r_deleg_ino, ino);
 552                 mapping_set_error(req->r_target_inode->i_mapping, result);
 553
 554                 spin_lock(&ci->i_ceph_lock);
 555                 if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE) {
 556                         ci->i_ceph_flags &= ~CEPH_I_ASYNC_CREATE;
 557                         wake_up_bit(&ci->i_ceph_flags, CEPH_ASYNC_CREATE_BIT);
 558                 }
 559                 ceph_kick_flushing_inode_caps(req->r_session, ci);
 560                 spin_unlock(&ci->i_ceph_lock);
 561         } else {
 562                 pr_warn("%s: no req->r_target_inode for 0x%llx\n", __func__,
 563                         req->r_deleg_ino);
 564         }
 565 out:
 566         ceph_mdsc_release_dir_caps(req);
 567 }
 568
 569 static int ceph_finish_async_create(struct inode *dir, struct dentry *dentry,
 570                                     struct file *file, umode_t mode,
 571                                     struct ceph_mds_request *req,
 572                                     struct ceph_acl_sec_ctx *as_ctx,
 573                                     struct ceph_file_layout *lo)
 574 {
 575         int ret;
 576         char xattr_buf[4];
 577         struct ceph_mds_reply_inode in = { };
 578         struct ceph_mds_reply_info_in iinfo = { .in = &in };
 579         struct ceph_inode_info *ci = ceph_inode(dir);
 580         struct inode *inode;
 581         struct timespec64 now;
 582         struct ceph_vino vino = { .ino = req->r_deleg_ino,
 583                                   .snap = CEPH_NOSNAP };
 584
 585         ktime_get_real_ts64(&now);
 586
 587         inode = ceph_get_inode(dentry->d_sb, vino);
 588         if (IS_ERR(inode))
 589                 return PTR_ERR(inode);
 590
 591         iinfo.inline_version = CEPH_INLINE_NONE;
 592         iinfo.change_attr = 1;
 593         ceph_encode_timespec64(&iinfo.btime, &now);
 594
 595         iinfo.xattr_len = ARRAY_SIZE(xattr_buf);
 596         iinfo.xattr_data = xattr_buf;
 597         memset(iinfo.xattr_data, 0, iinfo.xattr_len);
 598
 599         in.ino = cpu_to_le64(vino.ino);
 600         in.snapid = cpu_to_le64(CEPH_NOSNAP);
 601         in.version = cpu_to_le64(1);    // ???
 602         in.cap.caps = in.cap.wanted = cpu_to_le32(CEPH_CAP_ALL_FILE);
 603         in.cap.cap_id = cpu_to_le64(1);
 604         in.cap.realm = cpu_to_le64(ci->i_snap_realm->ino);
 605         in.cap.flags = CEPH_CAP_FLAG_AUTH;
 606         in.ctime = in.mtime = in.atime = iinfo.btime;
 607         in.mode = cpu_to_le32((u32)mode);
 608         in.truncate_seq = cpu_to_le32(1);
 609         in.truncate_size = cpu_to_le64(-1ULL);
 610         in.xattr_version = cpu_to_le64(1);
 611         in.uid = cpu_to_le32(from_kuid(&init_user_ns, current_fsuid()));
 612         in.gid = cpu_to_le32(from_kgid(&init_user_ns, dir->i_mode & S_ISGID ?
 613                                 dir->i_gid : current_fsgid()));
 614         in.nlink = cpu_to_le32(1);
 615         in.max_size = cpu_to_le64(lo->stripe_unit);
 616
 617         ceph_file_layout_to_legacy(lo, &in.layout);
 618
 619         ret = ceph_fill_inode(inode, NULL, &iinfo, NULL, req->r_session,
 620                               req->r_fmode, NULL);
 621         if (ret) {
 622                 dout("%s failed to fill inode: %d\n", __func__, ret);
 623                 ceph_dir_clear_complete(dir);
 624                 if (!d_unhashed(dentry))
 625                         d_drop(dentry);
 626                 if (inode->i_state & I_NEW)
 627                         discard_new_inode(inode);
 628         } else {
 629                 struct dentry *dn;
 630
 631                 dout("%s d_adding new inode 0x%llx to 0x%llx/%s\n", __func__,
 632                         vino.ino, ceph_ino(dir), dentry->d_name.name);
 633                 ceph_dir_clear_ordered(dir);
 634                 ceph_init_inode_acls(inode, as_ctx);
 635                 if (inode->i_state & I_NEW) {
 636                         /*
 637                          * If it's not I_NEW, then someone created this before
 638                          * we got here. Assume the server is aware of it at
 639                          * that point and don't worry about setting
 640                          * CEPH_I_ASYNC_CREATE.
 641                          */
 642                         ceph_inode(inode)->i_ceph_flags = CEPH_I_ASYNC_CREATE;
 643                         unlock_new_inode(inode);
 644                 }
 645                 if (d_in_lookup(dentry) || d_really_is_negative(dentry)) {
 646                         if (!d_unhashed(dentry))
 647                                 d_drop(dentry);
 648                         dn = d_splice_alias(inode, dentry);
 649                         WARN_ON_ONCE(dn && dn != dentry);
 650                 }
 651                 file->f_mode |= FMODE_CREATED;
 652                 ret = finish_open(file, dentry, ceph_open);
 653         }
 654         return ret;
 655 }
 656
 657 /*
 658  * Do a lookup + open with a single request.  If we get a non-existent
 659  * file or symlink, return 1 so the VFS can retry.
 660  */
 661 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
 662                      struct file *file, unsigned flags, umode_t mode)
 663 {
 664         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 665         struct ceph_mds_client *mdsc = fsc->mdsc;
 666         struct ceph_mds_request *req;
 667         struct dentry *dn;
 668         struct ceph_acl_sec_ctx as_ctx = {};
 669         bool try_async = ceph_test_mount_opt(fsc, ASYNC_DIROPS);
 670         int mask;
 671         int err;
 672
 673         dout("atomic_open %p dentry %p '%pd' %s flags %d mode 0%o\n",
 674              dir, dentry, dentry,
 675              d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
 676
 677         if (dentry->d_name.len > NAME_MAX)
 678                 return -ENAMETOOLONG;
 679
 680         if (flags & O_CREAT) {
 681                 if (ceph_quota_is_max_files_exceeded(dir))
 682                         return -EDQUOT;
 683                 err = ceph_pre_init_acls(dir, &mode, &as_ctx);
 684                 if (err < 0)
 685                         return err;
 686                 err = ceph_security_init_secctx(dentry, mode, &as_ctx);
 687                 if (err < 0)
 688                         goto out_ctx;
 689         } else if (!d_in_lookup(dentry)) {
 690                 /* If it's not being looked up, it's negative */
 691                 return -ENOENT;
 692         }
 693 retry:
 694         /* do the open */
 695         req = prepare_open_request(dir->i_sb, flags, mode);
 696         if (IS_ERR(req)) {
 697                 err = PTR_ERR(req);
 698                 goto out_ctx;
 699         }
 700         req->r_dentry = dget(dentry);
 701         req->r_num_caps = 2;
 702         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 703         if (ceph_security_xattr_wanted(dir))
 704                 mask |= CEPH_CAP_XATTR_SHARED;
 705         req->r_args.open.mask = cpu_to_le32(mask);
 706         req->r_parent = dir;
 707
 708         if (flags & O_CREAT) {
 709                 struct ceph_file_layout lo;
 710
 711                 req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 712                 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 713                 if (as_ctx.pagelist) {
 714                         req->r_pagelist = as_ctx.pagelist;
 715                         as_ctx.pagelist = NULL;
 716                 }
 717                 if (try_async &&
 718                     (req->r_dir_caps =
 719                       try_prep_async_create(dir, dentry, &lo,
 720                                             &req->r_deleg_ino))) {
 721                         set_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags);
 722                         req->r_args.open.flags |= cpu_to_le32(CEPH_O_EXCL);
 723                         req->r_callback = ceph_async_create_cb;
 724                         err = ceph_mdsc_submit_request(mdsc, dir, req);
 725                         if (!err) {
 726                                 err = ceph_finish_async_create(dir, dentry,
 727                                                         file, mode, req,
 728                                                         &as_ctx, &lo);
 729                         } else if (err == -EJUKEBOX) {
 730                                 restore_deleg_ino(dir, req->r_deleg_ino);
 731                                 ceph_mdsc_put_request(req);
 732                                 try_async = false;
 733                                 goto retry;
 734                         }
 735                         goto out_req;
 736                 }
 737         }
 738
 739         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 740         err = ceph_mdsc_do_request(mdsc,
 741                                    (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
 742                                    req);
 743         err = ceph_handle_snapdir(req, dentry, err);
 744         if (err)
 745                 goto out_req;
 746
 747         if ((flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
 748                 err = ceph_handle_notrace_create(dir, dentry);
 749
 750         if (d_in_lookup(dentry)) {
 751                 dn = ceph_finish_lookup(req, dentry, err);
 752                 if (IS_ERR(dn))
 753                         err = PTR_ERR(dn);
 754         } else {
 755                 /* we were given a hashed negative dentry */
 756                 dn = NULL;
 757         }
 758         if (err)
 759                 goto out_req;
 760         if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
 761                 /* make vfs retry on splice, ENOENT, or symlink */
 762                 dout("atomic_open finish_no_open on dn %p\n", dn);
 763                 err = finish_no_open(file, dn);
 764         } else {
 765                 dout("atomic_open finish_open on dn %p\n", dn);
 766                 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
 767                         struct inode *newino = d_inode(dentry);
 768
 769                         cache_file_layout(dir, newino);
 770                         ceph_init_inode_acls(newino, &as_ctx);
 771                         file->f_mode |= FMODE_CREATED;
 772                 }
 773                 err = finish_open(file, dentry, ceph_open);
 774         }
 775 out_req:
 776         ceph_mdsc_put_request(req);
 777 out_ctx:
 778         ceph_release_acl_sec_ctx(&as_ctx);
 779         dout("atomic_open result=%d\n", err);
 780         return err;
 781 }
 782
 783 int ceph_release(struct inode *inode, struct file *file)
 784 {
 785         struct ceph_inode_info *ci = ceph_inode(inode);
 786
 787         if (S_ISDIR(inode->i_mode)) {
 788                 struct ceph_dir_file_info *dfi = file->private_data;
 789                 dout("release inode %p dir file %p\n", inode, file);
 790                 WARN_ON(!list_empty(&dfi->file_info.rw_contexts));
 791
 792                 ceph_put_fmode(ci, dfi->file_info.fmode, 1);
 793
 794                 if (dfi->last_readdir)
 795                         ceph_mdsc_put_request(dfi->last_readdir);
 796                 kfree(dfi->last_name);
 797                 kfree(dfi->dir_info);
 798                 kmem_cache_free(ceph_dir_file_cachep, dfi);
 799         } else {
 800                 struct ceph_file_info *fi = file->private_data;
 801                 dout("release inode %p regular file %p\n", inode, file);
 802                 WARN_ON(!list_empty(&fi->rw_contexts));
 803
 804                 ceph_put_fmode(ci, fi->fmode, 1);
 805
 806                 kmem_cache_free(ceph_file_cachep, fi);
 807         }
 808
 809         /* wake up anyone waiting for caps on this inode */
 810         wake_up_all(&ci->i_cap_wq);
 811         return 0;
 812 }
 813
 814 enum {
 815         HAVE_RETRIED = 1,
 816         CHECK_EOF =    2,
 817         READ_INLINE =  3,
 818 };
 819
 820 /*
 821  * Completely synchronous read and write methods.  Direct from __user
 822  * buffer to osd, or directly to user pages (if O_DIRECT).
 823  *
 824  * If the read spans object boundary, just do multiple reads.  (That's not
 825  * atomic, but good enough for now.)
 826  *
 827  * If we get a short result from the OSD, check against i_size; we need to
 828  * only return a short read to the caller if we hit EOF.
 829  */
 830 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *to,
 831                               int *retry_op)
 832 {
 833         struct file *file = iocb->ki_filp;
 834         struct inode *inode = file_inode(file);
 835         struct ceph_inode_info *ci = ceph_inode(inode);
 836         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 837         struct ceph_osd_client *osdc = &fsc->client->osdc;
 838         ssize_t ret;
 839         u64 off = iocb->ki_pos;
 840         u64 len = iov_iter_count(to);
 841
 842         dout("sync_read on file %p %llu~%u %s\n", file, off, (unsigned)len,
 843              (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
 844
 845         if (!len)
 846                 return 0;
 847         /*
 848          * flush any page cache pages in this range.  this
 849          * will make concurrent normal and sync io slow,
 850          * but it will at least behave sensibly when they are
 851          * in sequence.
 852          */
 853         ret = filemap_write_and_wait_range(inode->i_mapping,
 854                                            off, off + len - 1);
 855         if (ret < 0)
 856                 return ret;
 857
 858         ret = 0;
 859         while ((len = iov_iter_count(to)) > 0) {
 860                 struct ceph_osd_request *req;
 861                 struct page **pages;
 862                 int num_pages;
 863                 size_t page_off;
 864                 u64 i_size;
 865                 bool more;
 866
 867                 req = ceph_osdc_new_request(osdc, &ci->i_layout,
 868                                         ci->i_vino, off, &len, 0, 1,
 869                                         CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
 870                                         NULL, ci->i_truncate_seq,
 871                                         ci->i_truncate_size, false);
 872                 if (IS_ERR(req)) {
 873                         ret = PTR_ERR(req);
 874                         break;
 875                 }
 876
 877                 more = len < iov_iter_count(to);
 878
 879                 if (unlikely(iov_iter_is_pipe(to))) {
 880                         ret = iov_iter_get_pages_alloc(to, &pages, len,
 881                                                        &page_off);
 882                         if (ret <= 0) {
 883                                 ceph_osdc_put_request(req);
 884                                 ret = -ENOMEM;
 885                                 break;
 886                         }
 887                         num_pages = DIV_ROUND_UP(ret + page_off, PAGE_SIZE);
 888                         if (ret < len) {
 889                                 len = ret;
 890                                 osd_req_op_extent_update(req, 0, len);
 891                                 more = false;
 892                         }
 893                 } else {
 894                         num_pages = calc_pages_for(off, len);
 895                         page_off = off & ~PAGE_MASK;
 896                         pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
 897                         if (IS_ERR(pages)) {
 898                                 ceph_osdc_put_request(req);
 899                                 ret = PTR_ERR(pages);
 900                                 break;
 901                         }
 902                 }
 903
 904                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_off,
 905                                                  false, false);
 906                 ret = ceph_osdc_start_request(osdc, req, false);
 907                 if (!ret)
 908                         ret = ceph_osdc_wait_request(osdc, req);
 909
 910                 ceph_update_read_latency(&fsc->mdsc->metric,
 911                                          req->r_start_latency,
 912                                          req->r_end_latency,
 913                                          ret);
 914
 915                 ceph_osdc_put_request(req);
 916
 917                 i_size = i_size_read(inode);
 918                 dout("sync_read %llu~%llu got %zd i_size %llu%s\n",
 919                      off, len, ret, i_size, (more ? " MORE" : ""));
 920
 921                 if (ret == -ENOENT)
 922                         ret = 0;
 923                 if (ret >= 0 && ret < len && (off + ret < i_size)) {
 924                         int zlen = min(len - ret, i_size - off - ret);
 925                         int zoff = page_off + ret;
 926                         dout("sync_read zero gap %llu~%llu\n",
 927                              off + ret, off + ret + zlen);
 928                         ceph_zero_page_vector_range(zoff, zlen, pages);
 929                         ret += zlen;
 930                 }
 931
 932                 if (unlikely(iov_iter_is_pipe(to))) {
 933                         if (ret > 0) {
 934                                 iov_iter_advance(to, ret);
 935                                 off += ret;
 936                         } else {
 937                                 iov_iter_advance(to, 0);
 938                         }
 939                         ceph_put_page_vector(pages, num_pages, false);
 940                 } else {
 941                         int idx = 0;
 942                         size_t left = ret > 0 ? ret : 0;
 943                         while (left > 0) {
 944                                 size_t len, copied;
 945                                 page_off = off & ~PAGE_MASK;
 946                                 len = min_t(size_t, left, PAGE_SIZE - page_off);
 947                                 copied = copy_page_to_iter(pages[idx++],
 948                                                            page_off, len, to);
 949                                 off += copied;
 950                                 left -= copied;
 951                                 if (copied < len) {
 952                                         ret = -EFAULT;
 953                                         break;
 954                                 }
 955                         }
 956                         ceph_release_page_vector(pages, num_pages);
 957                 }
 958
 959                 if (ret < 0) {
 960                         if (ret == -EBLACKLISTED)
 961                                 fsc->blacklisted = true;
 962                         break;
 963                 }
 964
 965                 if (off >= i_size || !more)
 966                         break;
 967         }
 968
 969         if (off > iocb->ki_pos) {
 970                 if (ret >= 0 &&
 971                     iov_iter_count(to) > 0 && off >= i_size_read(inode))
 972                         *retry_op = CHECK_EOF;
 973                 ret = off - iocb->ki_pos;
 974                 iocb->ki_pos = off;
 975         }
 976
 977         dout("sync_read result %zd retry_op %d\n", ret, *retry_op);
 978         return ret;
 979 }
 980
 981 struct ceph_aio_request {
 982         struct kiocb *iocb;
 983         size_t total_len;
 984         bool write;
 985         bool should_dirty;
 986         int error;
 987         struct list_head osd_reqs;
 988         unsigned num_reqs;
 989         atomic_t pending_reqs;
 990         struct timespec64 mtime;
 991         struct ceph_cap_flush *prealloc_cf;
 992 };
 993
 994 struct ceph_aio_work {
 995         struct work_struct work;
 996         struct ceph_osd_request *req;
 997 };
 998
 999 static void ceph_aio_retry_work(struct work_struct *work);
1000
1001 static void ceph_aio_complete(struct inode *inode,
1002                               struct ceph_aio_request *aio_req)
1003 {
1004         struct ceph_inode_info *ci = ceph_inode(inode);
1005         int ret;
1006
1007         if (!atomic_dec_and_test(&aio_req->pending_reqs))
1008                 return;
1009
1010         if (aio_req->iocb->ki_flags & IOCB_DIRECT)
1011                 inode_dio_end(inode);
1012
1013         ret = aio_req->error;
1014         if (!ret)
1015                 ret = aio_req->total_len;
1016
1017         dout("ceph_aio_complete %p rc %d\n", inode, ret);
1018
1019         if (ret >= 0 && aio_req->write) {
1020                 int dirty;
1021
1022                 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
1023                 if (endoff > i_size_read(inode)) {
1024                         if (ceph_inode_set_size(inode, endoff))
1025                                 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
1026                 }
1027
1028                 spin_lock(&ci->i_ceph_lock);
1029                 ci->i_inline_version = CEPH_INLINE_NONE;
1030                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1031                                                &aio_req->prealloc_cf);
1032                 spin_unlock(&ci->i_ceph_lock);
1033                 if (dirty)
1034                         __mark_inode_dirty(inode, dirty);
1035
1036         }
1037
1038         ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
1039                                                 CEPH_CAP_FILE_RD));
1040
1041         aio_req->iocb->ki_complete(aio_req->iocb, ret, 0);
1042
1043         ceph_free_cap_flush(aio_req->prealloc_cf);
1044         kfree(aio_req);
1045 }
1046
1047 static void ceph_aio_complete_req(struct ceph_osd_request *req)
1048 {
1049         int rc = req->r_result;
1050         struct inode *inode = req->r_inode;
1051         struct ceph_aio_request *aio_req = req->r_priv;
1052         struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
1053         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1054         struct ceph_client_metric *metric = &fsc->mdsc->metric;
1055
1056         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_BVECS);
1057         BUG_ON(!osd_data->num_bvecs);
1058
1059         dout("ceph_aio_complete_req %p rc %d bytes %u\n",
1060              inode, rc, osd_data->bvec_pos.iter.bi_size);
1061
1062         /* r_start_latency == 0 means the request was not submitted */
1063         if (req->r_start_latency) {
1064                 if (aio_req->write)
1065                         ceph_update_write_latency(metric, req->r_start_latency,
1066                                                   req->r_end_latency, rc);
1067                 else
1068                         ceph_update_read_latency(metric, req->r_start_latency,
1069                                                  req->r_end_latency, rc);
1070         }
1071
1072         if (rc == -EOLDSNAPC) {
1073                 struct ceph_aio_work *aio_work;
1074                 BUG_ON(!aio_req->write);
1075
1076                 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
1077                 if (aio_work) {
1078                         INIT_WORK(&aio_work->work, ceph_aio_retry_work);
1079                         aio_work->req = req;
1080                         queue_work(ceph_inode_to_client(inode)->inode_wq,
1081                                    &aio_work->work);
1082                         return;
1083                 }
1084                 rc = -ENOMEM;
1085         } else if (!aio_req->write) {
1086                 if (rc == -ENOENT)
1087                         rc = 0;
1088                 if (rc >= 0 && osd_data->bvec_pos.iter.bi_size > rc) {
1089                         struct iov_iter i;
1090                         int zlen = osd_data->bvec_pos.iter.bi_size - rc;
1091
1092                         /*
1093                          * If read is satisfied by single OSD request,
1094                          * it can pass EOF. Otherwise read is within
1095                          * i_size.
1096                          */
1097                         if (aio_req->num_reqs == 1) {
1098                                 loff_t i_size = i_size_read(inode);
1099                                 loff_t endoff = aio_req->iocb->ki_pos + rc;
1100                                 if (endoff < i_size)
1101                                         zlen = min_t(size_t, zlen,
1102                                                      i_size - endoff);
1103                                 aio_req->total_len = rc + zlen;
1104                         }
1105
1106                         iov_iter_bvec(&i, READ, osd_data->bvec_pos.bvecs,
1107                                       osd_data->num_bvecs,
1108                                       osd_data->bvec_pos.iter.bi_size);
1109                         iov_iter_advance(&i, rc);
1110                         iov_iter_zero(zlen, &i);
1111                 }
1112         }
1113
1114         put_bvecs(osd_data->bvec_pos.bvecs, osd_data->num_bvecs,
1115                   aio_req->should_dirty);
1116         ceph_osdc_put_request(req);
1117
1118         if (rc < 0)
1119                 cmpxchg(&aio_req->error, 0, rc);
1120
1121         ceph_aio_complete(inode, aio_req);
1122         return;
1123 }
1124
1125 static void ceph_aio_retry_work(struct work_struct *work)
1126 {
1127         struct ceph_aio_work *aio_work =
1128                 container_of(work, struct ceph_aio_work, work);
1129         struct ceph_osd_request *orig_req = aio_work->req;
1130         struct ceph_aio_request *aio_req = orig_req->r_priv;
1131         struct inode *inode = orig_req->r_inode;
1132         struct ceph_inode_info *ci = ceph_inode(inode);
1133         struct ceph_snap_context *snapc;
1134         struct ceph_osd_request *req;
1135         int ret;
1136
1137         spin_lock(&ci->i_ceph_lock);
1138         if (__ceph_have_pending_cap_snap(ci)) {
1139                 struct ceph_cap_snap *capsnap =
1140                         list_last_entry(&ci->i_cap_snaps,
1141                                         struct ceph_cap_snap,
1142                                         ci_item);
1143                 snapc = ceph_get_snap_context(capsnap->context);
1144         } else {
1145                 BUG_ON(!ci->i_head_snapc);
1146                 snapc = ceph_get_snap_context(ci->i_head_snapc);
1147         }
1148         spin_unlock(&ci->i_ceph_lock);
1149
1150         req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 1,
1151                         false, GFP_NOFS);
1152         if (!req) {
1153                 ret = -ENOMEM;
1154                 req = orig_req;
1155                 goto out;
1156         }
1157
1158         req->r_flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1159         ceph_oloc_copy(&req->r_base_oloc, &orig_req->r_base_oloc);
1160         ceph_oid_copy(&req->r_base_oid, &orig_req->r_base_oid);
1161
1162         req->r_ops[0] = orig_req->r_ops[0];
1163
1164         req->r_mtime = aio_req->mtime;
1165         req->r_data_offset = req->r_ops[0].extent.offset;
1166
1167         ret = ceph_osdc_alloc_messages(req, GFP_NOFS);
1168         if (ret) {
1169                 ceph_osdc_put_request(req);
1170                 req = orig_req;
1171                 goto out;
1172         }
1173
1174         ceph_osdc_put_request(orig_req);
1175
1176         req->r_callback = ceph_aio_complete_req;
1177         req->r_inode = inode;
1178         req->r_priv = aio_req;
1179
1180         ret = ceph_osdc_start_request(req->r_osdc, req, false);
1181 out:
1182         if (ret < 0) {
1183                 req->r_result = ret;
1184                 ceph_aio_complete_req(req);
1185         }
1186
1187         ceph_put_snap_context(snapc);
1188         kfree(aio_work);
1189 }
1190
1191 static ssize_t
1192 ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
1193                        struct ceph_snap_context *snapc,
1194                        struct ceph_cap_flush **pcf)
1195 {
1196         struct file *file = iocb->ki_filp;
1197         struct inode *inode = file_inode(file);
1198         struct ceph_inode_info *ci = ceph_inode(inode);
1199         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1200         struct ceph_client_metric *metric = &fsc->mdsc->metric;
1201         struct ceph_vino vino;
1202         struct ceph_osd_request *req;
1203         struct bio_vec *bvecs;
1204         struct ceph_aio_request *aio_req = NULL;
1205         int num_pages = 0;
1206         int flags;
1207         int ret = 0;
1208         struct timespec64 mtime = current_time(inode);
1209         size_t count = iov_iter_count(iter);
1210         loff_t pos = iocb->ki_pos;
1211         bool write = iov_iter_rw(iter) == WRITE;
1212         bool should_dirty = !write && iter_is_iovec(iter);
1213
1214         if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1215                 return -EROFS;
1216
1217         dout("sync_direct_%s on file %p %lld~%u snapc %p seq %lld\n",
1218              (write ? "write" : "read"), file, pos, (unsigned)count,
1219              snapc, snapc ? snapc->seq : 0);
1220
1221         if (write) {
1222                 int ret2 = invalidate_inode_pages2_range(inode->i_mapping,
1223                                         pos >> PAGE_SHIFT,
1224                                         (pos + count - 1) >> PAGE_SHIFT);
1225                 if (ret2 < 0)
1226                         dout("invalidate_inode_pages2_range returned %d\n", ret2);
1227
1228                 flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1229         } else {
1230                 flags = CEPH_OSD_FLAG_READ;
1231         }
1232
1233         while (iov_iter_count(iter) > 0) {
1234                 u64 size = iov_iter_count(iter);
1235                 ssize_t len;
1236
1237                 if (write)
1238                         size = min_t(u64, size, fsc->mount_options->wsize);
1239                 else
1240                         size = min_t(u64, size, fsc->mount_options->rsize);
1241
1242                 vino = ceph_vino(inode);
1243                 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1244                                             vino, pos, &size, 0,
1245                                             1,
1246                                             write ? CEPH_OSD_OP_WRITE :
1247                                                     CEPH_OSD_OP_READ,
1248                                             flags, snapc,
1249                                             ci->i_truncate_seq,
1250                                             ci->i_truncate_size,
1251                                             false);
1252                 if (IS_ERR(req)) {
1253                         ret = PTR_ERR(req);
1254                         break;
1255                 }
1256
1257                 len = iter_get_bvecs_alloc(iter, size, &bvecs, &num_pages);
1258                 if (len < 0) {
1259                         ceph_osdc_put_request(req);
1260                         ret = len;
1261                         break;
1262                 }
1263                 if (len != size)
1264                         osd_req_op_extent_update(req, 0, len);
1265
1266                 /*
1267                  * To simplify error handling, allow AIO when IO within i_size
1268                  * or IO can be satisfied by single OSD request.
1269                  */
1270                 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
1271                     (len == count || pos + count <= i_size_read(inode))) {
1272                         aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
1273                         if (aio_req) {
1274                                 aio_req->iocb = iocb;
1275                                 aio_req->write = write;
1276                                 aio_req->should_dirty = should_dirty;
1277                                 INIT_LIST_HEAD(&aio_req->osd_reqs);
1278                                 if (write) {
1279                                         aio_req->mtime = mtime;
1280                                         swap(aio_req->prealloc_cf, *pcf);
1281                                 }
1282                         }
1283                         /* ignore error */
1284                 }
1285
1286                 if (write) {
1287                         /*
1288                          * throw out any page cache pages in this range. this
1289                          * may block.
1290                          */
1291                         truncate_inode_pages_range(inode->i_mapping, pos,
1292                                                    PAGE_ALIGN(pos + len) - 1);
1293
1294                         req->r_mtime = mtime;
1295                 }
1296
1297                 osd_req_op_extent_osd_data_bvecs(req, 0, bvecs, num_pages, len);
1298
1299                 if (aio_req) {
1300                         aio_req->total_len += len;
1301                         aio_req->num_reqs++;
1302                         atomic_inc(&aio_req->pending_reqs);
1303
1304                         req->r_callback = ceph_aio_complete_req;
1305                         req->r_inode = inode;
1306                         req->r_priv = aio_req;
1307                         list_add_tail(&req->r_private_item, &aio_req->osd_reqs);
1308
1309                         pos += len;
1310                         continue;
1311                 }
1312
1313                 ret = ceph_osdc_start_request(req->r_osdc, req, false);
1314                 if (!ret)
1315                         ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1316
1317                 if (write)
1318                         ceph_update_write_latency(metric, req->r_start_latency,
1319                                                   req->r_end_latency, ret);
1320                 else
1321                         ceph_update_read_latency(metric, req->r_start_latency,
1322                                                  req->r_end_latency, ret);
1323
1324                 size = i_size_read(inode);
1325                 if (!write) {
1326                         if (ret == -ENOENT)
1327                                 ret = 0;
1328                         if (ret >= 0 && ret < len && pos + ret < size) {
1329                                 struct iov_iter i;
1330                                 int zlen = min_t(size_t, len - ret,
1331                                                  size - pos - ret);
1332
1333                                 iov_iter_bvec(&i, READ, bvecs, num_pages, len);
1334                                 iov_iter_advance(&i, ret);
1335                                 iov_iter_zero(zlen, &i);
1336                                 ret += zlen;
1337                         }
1338                         if (ret >= 0)
1339                                 len = ret;
1340                 }
1341
1342                 put_bvecs(bvecs, num_pages, should_dirty);
1343                 ceph_osdc_put_request(req);
1344                 if (ret < 0)
1345                         break;
1346
1347                 pos += len;
1348                 if (!write && pos >= size)
1349                         break;
1350
1351                 if (write && pos > size) {
1352                         if (ceph_inode_set_size(inode, pos))
1353                                 ceph_check_caps(ceph_inode(inode),
1354                                                 CHECK_CAPS_AUTHONLY,
1355                                                 NULL);
1356                 }
1357         }
1358
1359         if (aio_req) {
1360                 LIST_HEAD(osd_reqs);
1361
1362                 if (aio_req->num_reqs == 0) {
1363                         kfree(aio_req);
1364                         return ret;
1365                 }
1366
1367                 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
1368                                               CEPH_CAP_FILE_RD);
1369
1370                 list_splice(&aio_req->osd_reqs, &osd_reqs);
1371                 inode_dio_begin(inode);
1372                 while (!list_empty(&osd_reqs)) {
1373                         req = list_first_entry(&osd_reqs,
1374                                                struct ceph_osd_request,
1375                                                r_private_item);
1376                         list_del_init(&req->r_private_item);
1377                         if (ret >= 0)
1378                                 ret = ceph_osdc_start_request(req->r_osdc,
1379                                                               req, false);
1380                         if (ret < 0) {
1381                                 req->r_result = ret;
1382                                 ceph_aio_complete_req(req);
1383                         }
1384                 }
1385                 return -EIOCBQUEUED;
1386         }
1387
1388         if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
1389                 ret = pos - iocb->ki_pos;
1390                 iocb->ki_pos = pos;
1391         }
1392         return ret;
1393 }
1394
1395 /*
1396  * Synchronous write, straight from __user pointer or user pages.
1397  *
1398  * If write spans object boundary, just do multiple writes.  (For a
1399  * correct atomic write, we should e.g. take write locks on all
1400  * objects, rollback on failure, etc.)
1401  */
1402 static ssize_t
1403 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
1404                 struct ceph_snap_context *snapc)
1405 {
1406         struct file *file = iocb->ki_filp;
1407         struct inode *inode = file_inode(file);
1408         struct ceph_inode_info *ci = ceph_inode(inode);
1409         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1410         struct ceph_vino vino;
1411         struct ceph_osd_request *req;
1412         struct page **pages;
1413         u64 len;
1414         int num_pages;
1415         int written = 0;
1416         int flags;
1417         int ret;
1418         bool check_caps = false;
1419         struct timespec64 mtime = current_time(inode);
1420         size_t count = iov_iter_count(from);
1421
1422         if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1423                 return -EROFS;
1424
1425         dout("sync_write on file %p %lld~%u snapc %p seq %lld\n",
1426              file, pos, (unsigned)count, snapc, snapc->seq);
1427
1428         ret = filemap_write_and_wait_range(inode->i_mapping,
1429                                            pos, pos + count - 1);
1430         if (ret < 0)
1431                 return ret;
1432
1433         ret = invalidate_inode_pages2_range(inode->i_mapping,
1434                                             pos >> PAGE_SHIFT,
1435                                             (pos + count - 1) >> PAGE_SHIFT);
1436         if (ret < 0)
1437                 dout("invalidate_inode_pages2_range returned %d\n", ret);
1438
1439         flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1440
1441         while ((len = iov_iter_count(from)) > 0) {
1442                 size_t left;
1443                 int n;
1444
1445                 vino = ceph_vino(inode);
1446                 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1447                                             vino, pos, &len, 0, 1,
1448                                             CEPH_OSD_OP_WRITE, flags, snapc,
1449                                             ci->i_truncate_seq,
1450                                             ci->i_truncate_size,
1451                                             false);
1452                 if (IS_ERR(req)) {
1453                         ret = PTR_ERR(req);
1454                         break;
1455                 }
1456
1457                 /*
1458                  * write from beginning of first page,
1459                  * regardless of io alignment
1460                  */
1461                 num_pages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1462
1463                 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1464                 if (IS_ERR(pages)) {
1465                         ret = PTR_ERR(pages);
1466                         goto out;
1467                 }
1468
1469                 left = len;
1470                 for (n = 0; n < num_pages; n++) {
1471                         size_t plen = min_t(size_t, left, PAGE_SIZE);
1472                         ret = copy_page_from_iter(pages[n], 0, plen, from);
1473                         if (ret != plen) {
1474                                 ret = -EFAULT;
1475                                 break;
1476                         }
1477                         left -= ret;
1478                 }
1479
1480                 if (ret < 0) {
1481                         ceph_release_page_vector(pages, num_pages);
1482                         goto out;
1483                 }
1484
1485                 req->r_inode = inode;
1486
1487                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
1488                                                 false, true);
1489
1490                 req->r_mtime = mtime;
1491                 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1492                 if (!ret)
1493                         ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1494
1495                 ceph_update_write_latency(&fsc->mdsc->metric, req->r_start_latency,
1496                                           req->r_end_latency, ret);
1497 out:
1498                 ceph_osdc_put_request(req);
1499                 if (ret != 0) {
1500                         ceph_set_error_write(ci);
1501                         break;
1502                 }
1503
1504                 ceph_clear_error_write(ci);
1505                 pos += len;
1506                 written += len;
1507                 if (pos > i_size_read(inode)) {
1508                         check_caps = ceph_inode_set_size(inode, pos);
1509                         if (check_caps)
1510                                 ceph_check_caps(ceph_inode(inode),
1511                                                 CHECK_CAPS_AUTHONLY,
1512                                                 NULL);
1513                 }
1514
1515         }
1516
1517         if (ret != -EOLDSNAPC && written > 0) {
1518                 ret = written;
1519                 iocb->ki_pos = pos;
1520         }
1521         return ret;
1522 }
1523
1524 /*
1525  * Wrap generic_file_aio_read with checks for cap bits on the inode.
1526  * Atomically grab references, so that those bits are not released
1527  * back to the MDS mid-read.
1528  *
1529  * Hmm, the sync read case isn't actually async... should it be?
1530  */
1531 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
1532 {
1533         struct file *filp = iocb->ki_filp;
1534         struct ceph_file_info *fi = filp->private_data;
1535         size_t len = iov_iter_count(to);
1536         struct inode *inode = file_inode(filp);
1537         struct ceph_inode_info *ci = ceph_inode(inode);
1538         struct page *pinned_page = NULL;
1539         bool direct_lock = iocb->ki_flags & IOCB_DIRECT;
1540         ssize_t ret;
1541         int want, got = 0;
1542         int retry_op = 0, read = 0;
1543
1544 again:
1545         dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
1546              inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
1547
1548         if (direct_lock)
1549                 ceph_start_io_direct(inode);
1550         else
1551                 ceph_start_io_read(inode);
1552
1553         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1554                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1555         else
1556                 want = CEPH_CAP_FILE_CACHE;
1557         ret = ceph_get_caps(filp, CEPH_CAP_FILE_RD, want, -1,
1558                             &got, &pinned_page);
1559         if (ret < 0) {
1560                 if (iocb->ki_flags & IOCB_DIRECT)
1561                         ceph_end_io_direct(inode);
1562                 else
1563                         ceph_end_io_read(inode);
1564                 return ret;
1565         }
1566
1567         if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1568             (iocb->ki_flags & IOCB_DIRECT) ||
1569             (fi->flags & CEPH_F_SYNC)) {
1570
1571                 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1572                      inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1573                      ceph_cap_string(got));
1574
1575                 if (ci->i_inline_version == CEPH_INLINE_NONE) {
1576                         if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
1577                                 ret = ceph_direct_read_write(iocb, to,
1578                                                              NULL, NULL);
1579                                 if (ret >= 0 && ret < len)
1580                                         retry_op = CHECK_EOF;
1581                         } else {
1582                                 ret = ceph_sync_read(iocb, to, &retry_op);
1583                         }
1584                 } else {
1585                         retry_op = READ_INLINE;
1586                 }
1587         } else {
1588                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1589                 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1590                      inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1591                      ceph_cap_string(got));
1592                 ceph_add_rw_context(fi, &rw_ctx);
1593                 ret = generic_file_read_iter(iocb, to);
1594                 ceph_del_rw_context(fi, &rw_ctx);
1595         }
1596
1597         dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
1598              inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
1599         if (pinned_page) {
1600                 put_page(pinned_page);
1601                 pinned_page = NULL;
1602         }
1603         ceph_put_cap_refs(ci, got);
1604
1605         if (direct_lock)
1606                 ceph_end_io_direct(inode);
1607         else
1608                 ceph_end_io_read(inode);
1609
1610         if (retry_op > HAVE_RETRIED && ret >= 0) {
1611                 int statret;
1612                 struct page *page = NULL;
1613                 loff_t i_size;
1614                 if (retry_op == READ_INLINE) {
1615                         page = __page_cache_alloc(GFP_KERNEL);
1616                         if (!page)
1617                                 return -ENOMEM;
1618                 }
1619
1620                 statret = __ceph_do_getattr(inode, page,
1621                                             CEPH_STAT_CAP_INLINE_DATA, !!page);
1622                 if (statret < 0) {
1623                         if (page)
1624                                 __free_page(page);
1625                         if (statret == -ENODATA) {
1626                                 BUG_ON(retry_op != READ_INLINE);
1627                                 goto again;
1628                         }
1629                         return statret;
1630                 }
1631
1632                 i_size = i_size_read(inode);
1633                 if (retry_op == READ_INLINE) {
1634                         BUG_ON(ret > 0 || read > 0);
1635                         if (iocb->ki_pos < i_size &&
1636                             iocb->ki_pos < PAGE_SIZE) {
1637                                 loff_t end = min_t(loff_t, i_size,
1638                                                    iocb->ki_pos + len);
1639                                 end = min_t(loff_t, end, PAGE_SIZE);
1640                                 if (statret < end)
1641                                         zero_user_segment(page, statret, end);
1642                                 ret = copy_page_to_iter(page,
1643                                                 iocb->ki_pos & ~PAGE_MASK,
1644                                                 end - iocb->ki_pos, to);
1645                                 iocb->ki_pos += ret;
1646                                 read += ret;
1647                         }
1648                         if (iocb->ki_pos < i_size && read < len) {
1649                                 size_t zlen = min_t(size_t, len - read,
1650                                                     i_size - iocb->ki_pos);
1651                                 ret = iov_iter_zero(zlen, to);
1652                                 iocb->ki_pos += ret;
1653                                 read += ret;
1654                         }
1655                         __free_pages(page, 0);
1656                         return read;
1657                 }
1658
1659                 /* hit EOF or hole? */
1660                 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
1661                     ret < len) {
1662                         dout("sync_read hit hole, ppos %lld < size %lld"
1663                              ", reading more\n", iocb->ki_pos, i_size);
1664
1665                         read += ret;
1666                         len -= ret;
1667                         retry_op = HAVE_RETRIED;
1668                         goto again;
1669                 }
1670         }
1671
1672         if (ret >= 0)
1673                 ret += read;
1674
1675         return ret;
1676 }
1677
1678 /*
1679  * Take cap references to avoid releasing caps to MDS mid-write.
1680  *
1681  * If we are synchronous, and write with an old snap context, the OSD
1682  * may return EOLDSNAPC.  In that case, retry the write.. _after_
1683  * dropping our cap refs and allowing the pending snap to logically
1684  * complete _before_ this write occurs.
1685  *
1686  * If we are near ENOSPC, write synchronously.
1687  */
1688 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
1689 {
1690         struct file *file = iocb->ki_filp;
1691         struct ceph_file_info *fi = file->private_data;
1692         struct inode *inode = file_inode(file);
1693         struct ceph_inode_info *ci = ceph_inode(inode);
1694         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1695         struct ceph_osd_client *osdc = &fsc->client->osdc;
1696         struct ceph_cap_flush *prealloc_cf;
1697         ssize_t count, written = 0;
1698         int err, want, got;
1699         bool direct_lock = false;
1700         u32 map_flags;
1701         u64 pool_flags;
1702         loff_t pos;
1703         loff_t limit = max(i_size_read(inode), fsc->max_file_size);
1704
1705         if (ceph_snap(inode) != CEPH_NOSNAP)
1706                 return -EROFS;
1707
1708         prealloc_cf = ceph_alloc_cap_flush();
1709         if (!prealloc_cf)
1710                 return -ENOMEM;
1711
1712         if ((iocb->ki_flags & (IOCB_DIRECT | IOCB_APPEND)) == IOCB_DIRECT)
1713                 direct_lock = true;
1714
1715 retry_snap:
1716         if (direct_lock)
1717                 ceph_start_io_direct(inode);
1718         else
1719                 ceph_start_io_write(inode);
1720
1721         /* We can write back this queue in page reclaim */
1722         current->backing_dev_info = inode_to_bdi(inode);
1723
1724         if (iocb->ki_flags & IOCB_APPEND) {
1725                 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1726                 if (err < 0)
1727                         goto out;
1728         }
1729
1730         err = generic_write_checks(iocb, from);
1731         if (err <= 0)
1732                 goto out;
1733
1734         pos = iocb->ki_pos;
1735         if (unlikely(pos >= limit)) {
1736                 err = -EFBIG;
1737                 goto out;
1738         } else {
1739                 iov_iter_truncate(from, limit - pos);
1740         }
1741
1742         count = iov_iter_count(from);
1743         if (ceph_quota_is_max_bytes_exceeded(inode, pos + count)) {
1744                 err = -EDQUOT;
1745                 goto out;
1746         }
1747
1748         err = file_remove_privs(file);
1749         if (err)
1750                 goto out;
1751
1752         err = file_update_time(file);
1753         if (err)
1754                 goto out;
1755
1756         inode_inc_iversion_raw(inode);
1757
1758         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1759                 err = ceph_uninline_data(file, NULL);
1760                 if (err < 0)
1761                         goto out;
1762         }
1763
1764         down_read(&osdc->lock);
1765         map_flags = osdc->osdmap->flags;
1766         pool_flags = ceph_pg_pool_flags(osdc->osdmap, ci->i_layout.pool_id);
1767         up_read(&osdc->lock);
1768         if ((map_flags & CEPH_OSDMAP_FULL) ||
1769             (pool_flags & CEPH_POOL_FLAG_FULL)) {
1770                 err = -ENOSPC;
1771                 goto out;
1772         }
1773
1774         dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
1775              inode, ceph_vinop(inode), pos, count, i_size_read(inode));
1776         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1777                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1778         else
1779                 want = CEPH_CAP_FILE_BUFFER;
1780         got = 0;
1781         err = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, pos + count,
1782                             &got, NULL);
1783         if (err < 0)
1784                 goto out;
1785
1786         dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
1787              inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
1788
1789         if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1790             (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC) ||
1791             (ci->i_ceph_flags & CEPH_I_ERROR_WRITE)) {
1792                 struct ceph_snap_context *snapc;
1793                 struct iov_iter data;
1794
1795                 spin_lock(&ci->i_ceph_lock);
1796                 if (__ceph_have_pending_cap_snap(ci)) {
1797                         struct ceph_cap_snap *capsnap =
1798                                         list_last_entry(&ci->i_cap_snaps,
1799                                                         struct ceph_cap_snap,
1800                                                         ci_item);
1801                         snapc = ceph_get_snap_context(capsnap->context);
1802                 } else {
1803                         BUG_ON(!ci->i_head_snapc);
1804                         snapc = ceph_get_snap_context(ci->i_head_snapc);
1805                 }
1806                 spin_unlock(&ci->i_ceph_lock);
1807
1808                 /* we might need to revert back to that point */
1809                 data = *from;
1810                 if (iocb->ki_flags & IOCB_DIRECT)
1811                         written = ceph_direct_read_write(iocb, &data, snapc,
1812                                                          &prealloc_cf);
1813                 else
1814                         written = ceph_sync_write(iocb, &data, pos, snapc);
1815                 if (direct_lock)
1816                         ceph_end_io_direct(inode);
1817                 else
1818                         ceph_end_io_write(inode);
1819                 if (written > 0)
1820                         iov_iter_advance(from, written);
1821                 ceph_put_snap_context(snapc);
1822         } else {
1823                 /*
1824                  * No need to acquire the i_truncate_mutex. Because
1825                  * the MDS revokes Fwb caps before sending truncate
1826                  * message to us. We can't get Fwb cap while there
1827                  * are pending vmtruncate. So write and vmtruncate
1828                  * can not run at the same time
1829                  */
1830                 written = generic_perform_write(file, from, pos);
1831                 if (likely(written >= 0))
1832                         iocb->ki_pos = pos + written;
1833                 ceph_end_io_write(inode);
1834         }
1835
1836         if (written >= 0) {
1837                 int dirty;
1838
1839                 spin_lock(&ci->i_ceph_lock);
1840                 ci->i_inline_version = CEPH_INLINE_NONE;
1841                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1842                                                &prealloc_cf);
1843                 spin_unlock(&ci->i_ceph_lock);
1844                 if (dirty)
1845                         __mark_inode_dirty(inode, dirty);
1846                 if (ceph_quota_is_max_bytes_approaching(inode, iocb->ki_pos))
1847                         ceph_check_caps(ci, 0, NULL);
1848         }
1849
1850         dout("aio_write %p %llx.%llx %llu~%u  dropping cap refs on %s\n",
1851              inode, ceph_vinop(inode), pos, (unsigned)count,
1852              ceph_cap_string(got));
1853         ceph_put_cap_refs(ci, got);
1854
1855         if (written == -EOLDSNAPC) {
1856                 dout("aio_write %p %llx.%llx %llu~%u" "got EOLDSNAPC, retrying\n",
1857                      inode, ceph_vinop(inode), pos, (unsigned)count);
1858                 goto retry_snap;
1859         }
1860
1861         if (written >= 0) {
1862                 if ((map_flags & CEPH_OSDMAP_NEARFULL) ||
1863                     (pool_flags & CEPH_POOL_FLAG_NEARFULL))
1864                         iocb->ki_flags |= IOCB_DSYNC;
1865                 written = generic_write_sync(iocb, written);
1866         }
1867
1868         goto out_unlocked;
1869 out:
1870         if (direct_lock)
1871                 ceph_end_io_direct(inode);
1872         else
1873                 ceph_end_io_write(inode);
1874 out_unlocked:
1875         ceph_free_cap_flush(prealloc_cf);
1876         current->backing_dev_info = NULL;
1877         return written ? written : err;
1878 }
1879
1880 /*
1881  * llseek.  be sure to verify file size on SEEK_END.
1882  */
1883 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
1884 {
1885         struct inode *inode = file->f_mapping->host;
1886         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1887         loff_t i_size;
1888         loff_t ret;
1889
1890         inode_lock(inode);
1891
1892         if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
1893                 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1894                 if (ret < 0)
1895                         goto out;
1896         }
1897
1898         i_size = i_size_read(inode);
1899         switch (whence) {
1900         case SEEK_END:
1901                 offset += i_size;
1902                 break;
1903         case SEEK_CUR:
1904                 /*
1905                  * Here we special-case the lseek(fd, 0, SEEK_CUR)
1906                  * position-querying operation.  Avoid rewriting the "same"
1907                  * f_pos value back to the file because a concurrent read(),
1908                  * write() or lseek() might have altered it
1909                  */
1910                 if (offset == 0) {
1911                         ret = file->f_pos;
1912                         goto out;
1913                 }
1914                 offset += file->f_pos;
1915                 break;
1916         case SEEK_DATA:
1917                 if (offset < 0 || offset >= i_size) {
1918                         ret = -ENXIO;
1919                         goto out;
1920                 }
1921                 break;
1922         case SEEK_HOLE:
1923                 if (offset < 0 || offset >= i_size) {
1924                         ret = -ENXIO;
1925                         goto out;
1926                 }
1927                 offset = i_size;
1928                 break;
1929         }
1930
1931         ret = vfs_setpos(file, offset, max(i_size, fsc->max_file_size));
1932
1933 out:
1934         inode_unlock(inode);
1935         return ret;
1936 }
1937
1938 static inline void ceph_zero_partial_page(
1939         struct inode *inode, loff_t offset, unsigned size)
1940 {
1941         struct page *page;
1942         pgoff_t index = offset >> PAGE_SHIFT;
1943
1944         page = find_lock_page(inode->i_mapping, index);
1945         if (page) {
1946                 wait_on_page_writeback(page);
1947                 zero_user(page, offset & (PAGE_SIZE - 1), size);
1948                 unlock_page(page);
1949                 put_page(page);
1950         }
1951 }
1952
1953 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
1954                                       loff_t length)
1955 {
1956         loff_t nearly = round_up(offset, PAGE_SIZE);
1957         if (offset < nearly) {
1958                 loff_t size = nearly - offset;
1959                 if (length < size)
1960                         size = length;
1961                 ceph_zero_partial_page(inode, offset, size);
1962                 offset += size;
1963                 length -= size;
1964         }
1965         if (length >= PAGE_SIZE) {
1966                 loff_t size = round_down(length, PAGE_SIZE);
1967                 truncate_pagecache_range(inode, offset, offset + size - 1);
1968                 offset += size;
1969                 length -= size;
1970         }
1971         if (length)
1972                 ceph_zero_partial_page(inode, offset, length);
1973 }
1974
1975 static int ceph_zero_partial_object(struct inode *inode,
1976                                     loff_t offset, loff_t *length)
1977 {
1978         struct ceph_inode_info *ci = ceph_inode(inode);
1979         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1980         struct ceph_osd_request *req;
1981         int ret = 0;
1982         loff_t zero = 0;
1983         int op;
1984
1985         if (!length) {
1986                 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
1987                 length = &zero;
1988         } else {
1989                 op = CEPH_OSD_OP_ZERO;
1990         }
1991
1992         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1993                                         ceph_vino(inode),
1994                                         offset, length,
1995                                         0, 1, op,
1996                                         CEPH_OSD_FLAG_WRITE,
1997                                         NULL, 0, 0, false);
1998         if (IS_ERR(req)) {
1999                 ret = PTR_ERR(req);
2000                 goto out;
2001         }
2002
2003         req->r_mtime = inode->i_mtime;
2004         ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
2005         if (!ret) {
2006                 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
2007                 if (ret == -ENOENT)
2008                         ret = 0;
2009         }
2010         ceph_osdc_put_request(req);
2011
2012 out:
2013         return ret;
2014 }
2015
2016 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
2017 {
2018         int ret = 0;
2019         struct ceph_inode_info *ci = ceph_inode(inode);
2020         s32 stripe_unit = ci->i_layout.stripe_unit;
2021         s32 stripe_count = ci->i_layout.stripe_count;
2022         s32 object_size = ci->i_layout.object_size;
2023         u64 object_set_size = object_size * stripe_count;
2024         u64 nearly, t;
2025
2026         /* round offset up to next period boundary */
2027         nearly = offset + object_set_size - 1;
2028         t = nearly;
2029         nearly -= do_div(t, object_set_size);
2030
2031         while (length && offset < nearly) {
2032                 loff_t size = length;
2033                 ret = ceph_zero_partial_object(inode, offset, &size);
2034                 if (ret < 0)
2035                         return ret;
2036                 offset += size;
2037                 length -= size;
2038         }
2039         while (length >= object_set_size) {
2040                 int i;
2041                 loff_t pos = offset;
2042                 for (i = 0; i < stripe_count; ++i) {
2043                         ret = ceph_zero_partial_object(inode, pos, NULL);
2044                         if (ret < 0)
2045                                 return ret;
2046                         pos += stripe_unit;
2047                 }
2048                 offset += object_set_size;
2049                 length -= object_set_size;
2050         }
2051         while (length) {
2052                 loff_t size = length;
2053                 ret = ceph_zero_partial_object(inode, offset, &size);
2054                 if (ret < 0)
2055                         return ret;
2056                 offset += size;
2057                 length -= size;
2058         }
2059         return ret;
2060 }
2061
2062 static long ceph_fallocate(struct file *file, int mode,
2063                                 loff_t offset, loff_t length)
2064 {
2065         struct ceph_file_info *fi = file->private_data;
2066         struct inode *inode = file_inode(file);
2067         struct ceph_inode_info *ci = ceph_inode(inode);
2068         struct ceph_cap_flush *prealloc_cf;
2069         int want, got = 0;
2070         int dirty;
2071         int ret = 0;
2072         loff_t endoff = 0;
2073         loff_t size;
2074
2075         if (mode != (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2076                 return -EOPNOTSUPP;
2077
2078         if (!S_ISREG(inode->i_mode))
2079                 return -EOPNOTSUPP;
2080
2081         prealloc_cf = ceph_alloc_cap_flush();
2082         if (!prealloc_cf)
2083                 return -ENOMEM;
2084
2085         inode_lock(inode);
2086
2087         if (ceph_snap(inode) != CEPH_NOSNAP) {
2088                 ret = -EROFS;
2089                 goto unlock;
2090         }
2091
2092         if (ci->i_inline_version != CEPH_INLINE_NONE) {
2093                 ret = ceph_uninline_data(file, NULL);
2094                 if (ret < 0)
2095                         goto unlock;
2096         }
2097
2098         size = i_size_read(inode);
2099
2100         /* Are we punching a hole beyond EOF? */
2101         if (offset >= size)
2102                 goto unlock;
2103         if ((offset + length) > size)
2104                 length = size - offset;
2105
2106         if (fi->fmode & CEPH_FILE_MODE_LAZY)
2107                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
2108         else
2109                 want = CEPH_CAP_FILE_BUFFER;
2110
2111         ret = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, endoff, &got, NULL);
2112         if (ret < 0)
2113                 goto unlock;
2114
2115         ceph_zero_pagecache_range(inode, offset, length);
2116         ret = ceph_zero_objects(inode, offset, length);
2117
2118         if (!ret) {
2119                 spin_lock(&ci->i_ceph_lock);
2120                 ci->i_inline_version = CEPH_INLINE_NONE;
2121                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
2122                                                &prealloc_cf);
2123                 spin_unlock(&ci->i_ceph_lock);
2124                 if (dirty)
2125                         __mark_inode_dirty(inode, dirty);
2126         }
2127
2128         ceph_put_cap_refs(ci, got);
2129 unlock:
2130         inode_unlock(inode);
2131         ceph_free_cap_flush(prealloc_cf);
2132         return ret;
2133 }
2134
2135 /*
2136  * This function tries to get FILE_WR capabilities for dst_ci and FILE_RD for
2137  * src_ci.  Two attempts are made to obtain both caps, and an error is return if
2138  * this fails; zero is returned on success.
2139  */
2140 static int get_rd_wr_caps(struct file *src_filp, int *src_got,
2141                           struct file *dst_filp,
2142                           loff_t dst_endoff, int *dst_got)
2143 {
2144         int ret = 0;
2145         bool retrying = false;
2146
2147 retry_caps:
2148         ret = ceph_get_caps(dst_filp, CEPH_CAP_FILE_WR, CEPH_CAP_FILE_BUFFER,
2149                             dst_endoff, dst_got, NULL);
2150         if (ret < 0)
2151                 return ret;
2152
2153         /*
2154          * Since we're already holding the FILE_WR capability for the dst file,
2155          * we would risk a deadlock by using ceph_get_caps.  Thus, we'll do some
2156          * retry dance instead to try to get both capabilities.
2157          */
2158         ret = ceph_try_get_caps(file_inode(src_filp),
2159                                 CEPH_CAP_FILE_RD, CEPH_CAP_FILE_SHARED,
2160                                 false, src_got);
2161         if (ret <= 0) {
2162                 /* Start by dropping dst_ci caps and getting src_ci caps */
2163                 ceph_put_cap_refs(ceph_inode(file_inode(dst_filp)), *dst_got);
2164                 if (retrying) {
2165                         if (!ret)
2166                                 /* ceph_try_get_caps masks EAGAIN */
2167                                 ret = -EAGAIN;
2168                         return ret;
2169                 }
2170                 ret = ceph_get_caps(src_filp, CEPH_CAP_FILE_RD,
2171                                     CEPH_CAP_FILE_SHARED, -1, src_got, NULL);
2172                 if (ret < 0)
2173                         return ret;
2174                 /*... drop src_ci caps too, and retry */
2175                 ceph_put_cap_refs(ceph_inode(file_inode(src_filp)), *src_got);
2176                 retrying = true;
2177                 goto retry_caps;
2178         }
2179         return ret;
2180 }
2181
2182 static void put_rd_wr_caps(struct ceph_inode_info *src_ci, int src_got,
2183                            struct ceph_inode_info *dst_ci, int dst_got)
2184 {
2185         ceph_put_cap_refs(src_ci, src_got);
2186         ceph_put_cap_refs(dst_ci, dst_got);
2187 }
2188
2189 /*
2190  * This function does several size-related checks, returning an error if:
2191  *  - source file is smaller than off+len
2192  *  - destination file size is not OK (inode_newsize_ok())
2193  *  - max bytes quotas is exceeded
2194  */
2195 static int is_file_size_ok(struct inode *src_inode, struct inode *dst_inode,
2196                            loff_t src_off, loff_t dst_off, size_t len)
2197 {
2198         loff_t size, endoff;
2199
2200         size = i_size_read(src_inode);
2201         /*
2202          * Don't copy beyond source file EOF.  Instead of simply setting length
2203          * to (size - src_off), just drop to VFS default implementation, as the
2204          * local i_size may be stale due to other clients writing to the source
2205          * inode.
2206          */
2207         if (src_off + len > size) {
2208                 dout("Copy beyond EOF (%llu + %zu > %llu)\n",
2209                      src_off, len, size);
2210                 return -EOPNOTSUPP;
2211         }
2212         size = i_size_read(dst_inode);
2213
2214         endoff = dst_off + len;
2215         if (inode_newsize_ok(dst_inode, endoff))
2216                 return -EOPNOTSUPP;
2217
2218         if (ceph_quota_is_max_bytes_exceeded(dst_inode, endoff))
2219                 return -EDQUOT;
2220
2221         return 0;
2222 }
2223
2224 static ssize_t ceph_do_objects_copy(struct ceph_inode_info *src_ci, u64 *src_off,
2225                                     struct ceph_inode_info *dst_ci, u64 *dst_off,
2226                                     struct ceph_fs_client *fsc,
2227                                     size_t len, unsigned int flags)
2228 {
2229         struct ceph_object_locator src_oloc, dst_oloc;
2230         struct ceph_object_id src_oid, dst_oid;
2231         size_t bytes = 0;
2232         u64 src_objnum, src_objoff, dst_objnum, dst_objoff;
2233         u32 src_objlen, dst_objlen;
2234         u32 object_size = src_ci->i_layout.object_size;
2235         int ret;
2236
2237         src_oloc.pool = src_ci->i_layout.pool_id;
2238         src_oloc.pool_ns = ceph_try_get_string(src_ci->i_layout.pool_ns);
2239         dst_oloc.pool = dst_ci->i_layout.pool_id;
2240         dst_oloc.pool_ns = ceph_try_get_string(dst_ci->i_layout.pool_ns);
2241
2242         while (len >= object_size) {
2243                 ceph_calc_file_object_mapping(&src_ci->i_layout, *src_off,
2244                                               object_size, &src_objnum,
2245                                               &src_objoff, &src_objlen);
2246                 ceph_calc_file_object_mapping(&dst_ci->i_layout, *dst_off,
2247                                               object_size, &dst_objnum,
2248                                               &dst_objoff, &dst_objlen);
2249                 ceph_oid_init(&src_oid);
2250                 ceph_oid_printf(&src_oid, "%llx.%08llx",
2251                                 src_ci->i_vino.ino, src_objnum);
2252                 ceph_oid_init(&dst_oid);
2253                 ceph_oid_printf(&dst_oid, "%llx.%08llx",
2254                                 dst_ci->i_vino.ino, dst_objnum);
2255                 /* Do an object remote copy */
2256                 ret = ceph_osdc_copy_from(&fsc->client->osdc,
2257                                           src_ci->i_vino.snap, 0,
2258                                           &src_oid, &src_oloc,
2259                                           CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2260                                           CEPH_OSD_OP_FLAG_FADVISE_NOCACHE,
2261                                           &dst_oid, &dst_oloc,
2262                                           CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2263                                           CEPH_OSD_OP_FLAG_FADVISE_DONTNEED,
2264                                           dst_ci->i_truncate_seq,
2265                                           dst_ci->i_truncate_size,
2266                                           CEPH_OSD_COPY_FROM_FLAG_TRUNCATE_SEQ);
2267                 if (ret) {
2268                         if (ret == -EOPNOTSUPP) {
2269                                 fsc->have_copy_from2 = false;
2270                                 pr_notice("OSDs don't support copy-from2; disabling copy offload\n");
2271                         }
2272                         dout("ceph_osdc_copy_from returned %d\n", ret);
2273                         if (!bytes)
2274                                 bytes = ret;
2275                         goto out;
2276                 }
2277                 len -= object_size;
2278                 bytes += object_size;
2279                 *src_off += object_size;
2280                 *dst_off += object_size;
2281         }
2282
2283 out:
2284         ceph_oloc_destroy(&src_oloc);
2285         ceph_oloc_destroy(&dst_oloc);
2286         return bytes;
2287 }
2288
2289 static ssize_t __ceph_copy_file_range(struct file *src_file, loff_t src_off,
2290                                       struct file *dst_file, loff_t dst_off,
2291                                       size_t len, unsigned int flags)
2292 {
2293         struct inode *src_inode = file_inode(src_file);
2294         struct inode *dst_inode = file_inode(dst_file);
2295         struct ceph_inode_info *src_ci = ceph_inode(src_inode);
2296         struct ceph_inode_info *dst_ci = ceph_inode(dst_inode);
2297         struct ceph_cap_flush *prealloc_cf;
2298         struct ceph_fs_client *src_fsc = ceph_inode_to_client(src_inode);
2299         loff_t size;
2300         ssize_t ret = -EIO, bytes;
2301         u64 src_objnum, dst_objnum, src_objoff, dst_objoff;
2302         u32 src_objlen, dst_objlen;
2303         int src_got = 0, dst_got = 0, err, dirty;
2304
2305         if (src_inode->i_sb != dst_inode->i_sb) {
2306                 struct ceph_fs_client *dst_fsc = ceph_inode_to_client(dst_inode);
2307
2308                 if (ceph_fsid_compare(&src_fsc->client->fsid,
2309                                       &dst_fsc->client->fsid)) {
2310                         dout("Copying files across clusters: src: %pU dst: %pU\n",
2311                              &src_fsc->client->fsid, &dst_fsc->client->fsid);
2312                         return -EXDEV;
2313                 }
2314         }
2315         if (ceph_snap(dst_inode) != CEPH_NOSNAP)
2316                 return -EROFS;
2317
2318         /*
2319          * Some of the checks below will return -EOPNOTSUPP, which will force a
2320          * fallback to the default VFS copy_file_range implementation.  This is
2321          * desirable in several cases (for ex, the 'len' is smaller than the
2322          * size of the objects, or in cases where that would be more
2323          * efficient).
2324          */
2325
2326         if (ceph_test_mount_opt(src_fsc, NOCOPYFROM))
2327                 return -EOPNOTSUPP;
2328
2329         if (!src_fsc->have_copy_from2)
2330                 return -EOPNOTSUPP;
2331
2332         /*
2333          * Striped file layouts require that we copy partial objects, but the
2334          * OSD copy-from operation only supports full-object copies.  Limit
2335          * this to non-striped file layouts for now.
2336          */
2337         if ((src_ci->i_layout.stripe_unit != dst_ci->i_layout.stripe_unit) ||
2338             (src_ci->i_layout.stripe_count != 1) ||
2339             (dst_ci->i_layout.stripe_count != 1) ||
2340             (src_ci->i_layout.object_size != dst_ci->i_layout.object_size)) {
2341                 dout("Invalid src/dst files layout\n");
2342                 return -EOPNOTSUPP;
2343         }
2344
2345         if (len < src_ci->i_layout.object_size)
2346                 return -EOPNOTSUPP; /* no remote copy will be done */
2347
2348         prealloc_cf = ceph_alloc_cap_flush();
2349         if (!prealloc_cf)
2350                 return -ENOMEM;
2351
2352         /* Start by sync'ing the source and destination files */
2353         ret = file_write_and_wait_range(src_file, src_off, (src_off + len));
2354         if (ret < 0) {
2355                 dout("failed to write src file (%zd)\n", ret);
2356                 goto out;
2357         }
2358         ret = file_write_and_wait_range(dst_file, dst_off, (dst_off + len));
2359         if (ret < 0) {
2360                 dout("failed to write dst file (%zd)\n", ret);
2361                 goto out;
2362         }
2363
2364         /*
2365          * We need FILE_WR caps for dst_ci and FILE_RD for src_ci as other
2366          * clients may have dirty data in their caches.  And OSDs know nothing
2367          * about caps, so they can't safely do the remote object copies.
2368          */
2369         err = get_rd_wr_caps(src_file, &src_got,
2370                              dst_file, (dst_off + len), &dst_got);
2371         if (err < 0) {
2372                 dout("get_rd_wr_caps returned %d\n", err);
2373                 ret = -EOPNOTSUPP;
2374                 goto out;
2375         }
2376
2377         ret = is_file_size_ok(src_inode, dst_inode, src_off, dst_off, len);
2378         if (ret < 0)
2379                 goto out_caps;
2380
2381         /* Drop dst file cached pages */
2382         ret = invalidate_inode_pages2_range(dst_inode->i_mapping,
2383                                             dst_off >> PAGE_SHIFT,
2384                                             (dst_off + len) >> PAGE_SHIFT);
2385         if (ret < 0) {
2386                 dout("Failed to invalidate inode pages (%zd)\n", ret);
2387                 ret = 0; /* XXX */
2388         }
2389         ceph_calc_file_object_mapping(&src_ci->i_layout, src_off,
2390                                       src_ci->i_layout.object_size,
2391                                       &src_objnum, &src_objoff, &src_objlen);
2392         ceph_calc_file_object_mapping(&dst_ci->i_layout, dst_off,
2393                                       dst_ci->i_layout.object_size,
2394                                       &dst_objnum, &dst_objoff, &dst_objlen);
2395         /* object-level offsets need to the same */
2396         if (src_objoff != dst_objoff) {
2397                 ret = -EOPNOTSUPP;
2398                 goto out_caps;
2399         }
2400
2401         /*
2402          * Do a manual copy if the object offset isn't object aligned.
2403          * 'src_objlen' contains the bytes left until the end of the object,
2404          * starting at the src_off
2405          */
2406         if (src_objoff) {
2407                 dout("Initial partial copy of %u bytes\n", src_objlen);
2408
2409                 /*
2410                  * we need to temporarily drop all caps as we'll be calling
2411                  * {read,write}_iter, which will get caps again.
2412                  */
2413                 put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2414                 ret = do_splice_direct(src_file, &src_off, dst_file,
2415                                        &dst_off, src_objlen, flags);
2416                 /* Abort on short copies or on error */
2417                 if (ret < src_objlen) {
2418                         dout("Failed partial copy (%zd)\n", ret);
2419                         goto out;
2420                 }
2421                 len -= ret;
2422                 err = get_rd_wr_caps(src_file, &src_got,
2423                                      dst_file, (dst_off + len), &dst_got);
2424                 if (err < 0)
2425                         goto out;
2426                 err = is_file_size_ok(src_inode, dst_inode,
2427                                       src_off, dst_off, len);
2428                 if (err < 0)
2429                         goto out_caps;
2430         }
2431
2432         size = i_size_read(dst_inode);
2433         bytes = ceph_do_objects_copy(src_ci, &src_off, dst_ci, &dst_off,
2434                                      src_fsc, len, flags);
2435         if (bytes <= 0) {
2436                 if (!ret)
2437                         ret = bytes;
2438                 goto out_caps;
2439         }
2440         dout("Copied %zu bytes out of %zu\n", bytes, len);
2441         len -= bytes;
2442         ret += bytes;
2443
2444         file_update_time(dst_file);
2445         inode_inc_iversion_raw(dst_inode);
2446
2447         if (dst_off > size) {
2448                 /* Let the MDS know about dst file size change */
2449                 if (ceph_inode_set_size(dst_inode, dst_off) ||
2450                     ceph_quota_is_max_bytes_approaching(dst_inode, dst_off))
2451                         ceph_check_caps(dst_ci, CHECK_CAPS_AUTHONLY, NULL);
2452         }
2453         /* Mark Fw dirty */
2454         spin_lock(&dst_ci->i_ceph_lock);
2455         dst_ci->i_inline_version = CEPH_INLINE_NONE;
2456         dirty = __ceph_mark_dirty_caps(dst_ci, CEPH_CAP_FILE_WR, &prealloc_cf);
2457         spin_unlock(&dst_ci->i_ceph_lock);
2458         if (dirty)
2459                 __mark_inode_dirty(dst_inode, dirty);
2460
2461 out_caps:
2462         put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2463
2464         /*
2465          * Do the final manual copy if we still have some bytes left, unless
2466          * there were errors in remote object copies (len >= object_size).
2467          */
2468         if (len && (len < src_ci->i_layout.object_size)) {
2469                 dout("Final partial copy of %zu bytes\n", len);
2470                 bytes = do_splice_direct(src_file, &src_off, dst_file,
2471                                          &dst_off, len, flags);
2472                 if (bytes > 0)
2473                         ret += bytes;
2474                 else
2475                         dout("Failed partial copy (%zd)\n", bytes);
2476         }
2477
2478 out:
2479         ceph_free_cap_flush(prealloc_cf);
2480
2481         return ret;
2482 }
2483
2484 static ssize_t ceph_copy_file_range(struct file *src_file, loff_t src_off,
2485                                     struct file *dst_file, loff_t dst_off,
2486                                     size_t len, unsigned int flags)
2487 {
2488         ssize_t ret;
2489
2490         ret = __ceph_copy_file_range(src_file, src_off, dst_file, dst_off,
2491                                      len, flags);
2492
2493         if (ret == -EOPNOTSUPP || ret == -EXDEV)
2494                 ret = generic_copy_file_range(src_file, src_off, dst_file,
2495                                               dst_off, len, flags);
2496         return ret;
2497 }
2498
2499 const struct file_operations ceph_file_fops = {
2500         .open = ceph_open,
2501         .release = ceph_release,
2502         .llseek = ceph_llseek,
2503         .read_iter = ceph_read_iter,
2504         .write_iter = ceph_write_iter,
2505         .mmap = ceph_mmap,
2506         .fsync = ceph_fsync,
2507         .lock = ceph_lock,
2508         .setlease = simple_nosetlease,
2509         .flock = ceph_flock,
2510         .splice_read = generic_file_splice_read,
2511         .splice_write = iter_file_splice_write,
2512         .unlocked_ioctl = ceph_ioctl,
2513         .compat_ioctl = compat_ptr_ioctl,
2514         .fallocate      = ceph_fallocate,
2515         .copy_file_range = ceph_copy_file_range,
2516 };