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_mds_client *mdsc = ceph_sb_to_mdsc(sb);
 186         struct ceph_mds_request *req;
 187         int want_auth = USE_ANY_MDS;
 188         int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
 189
 190         if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
 191                 want_auth = USE_AUTH_MDS;
 192
 193         req = ceph_mdsc_create_request(mdsc, op, want_auth);
 194         if (IS_ERR(req))
 195                 goto out;
 196         req->r_fmode = ceph_flags_to_mode(flags);
 197         req->r_args.open.flags = ceph_flags_sys2wire(flags);
 198         req->r_args.open.mode = cpu_to_le32(create_mode);
 199 out:
 200         return req;
 201 }
 202
 203 static int ceph_init_file_info(struct inode *inode, struct file *file,
 204                                         int fmode, bool isdir)
 205 {
 206         struct ceph_inode_info *ci = ceph_inode(inode);
 207         struct ceph_file_info *fi;
 208
 209         dout("%s %p %p 0%o (%s)\n", __func__, inode, file,
 210                         inode->i_mode, isdir ? "dir" : "regular");
 211         BUG_ON(inode->i_fop->release != ceph_release);
 212
 213         if (isdir) {
 214                 struct ceph_dir_file_info *dfi =
 215                         kmem_cache_zalloc(ceph_dir_file_cachep, GFP_KERNEL);
 216                 if (!dfi)
 217                         return -ENOMEM;
 218
 219                 file->private_data = dfi;
 220                 fi = &dfi->file_info;
 221                 dfi->next_offset = 2;
 222                 dfi->readdir_cache_idx = -1;
 223         } else {
 224                 fi = kmem_cache_zalloc(ceph_file_cachep, GFP_KERNEL);
 225                 if (!fi)
 226                         return -ENOMEM;
 227
 228                 file->private_data = fi;
 229         }
 230
 231         ceph_get_fmode(ci, fmode, 1);
 232         fi->fmode = fmode;
 233
 234         spin_lock_init(&fi->rw_contexts_lock);
 235         INIT_LIST_HEAD(&fi->rw_contexts);
 236         fi->filp_gen = READ_ONCE(ceph_inode_to_client(inode)->filp_gen);
 237
 238         return 0;
 239 }
 240
 241 /*
 242  * initialize private struct file data.
 243  * if we fail, clean up by dropping fmode reference on the ceph_inode
 244  */
 245 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
 246 {
 247         int ret = 0;
 248
 249         switch (inode->i_mode & S_IFMT) {
 250         case S_IFREG:
 251                 ceph_fscache_register_inode_cookie(inode);
 252                 ceph_fscache_file_set_cookie(inode, file);
 253                 fallthrough;
 254         case S_IFDIR:
 255                 ret = ceph_init_file_info(inode, file, fmode,
 256                                                 S_ISDIR(inode->i_mode));
 257                 break;
 258
 259         case S_IFLNK:
 260                 dout("init_file %p %p 0%o (symlink)\n", inode, file,
 261                      inode->i_mode);
 262                 break;
 263
 264         default:
 265                 dout("init_file %p %p 0%o (special)\n", inode, file,
 266                      inode->i_mode);
 267                 /*
 268                  * we need to drop the open ref now, since we don't
 269                  * have .release set to ceph_release.
 270                  */
 271                 BUG_ON(inode->i_fop->release == ceph_release);
 272
 273                 /* call the proper open fop */
 274                 ret = inode->i_fop->open(inode, file);
 275         }
 276         return ret;
 277 }
 278
 279 /*
 280  * try renew caps after session gets killed.
 281  */
 282 int ceph_renew_caps(struct inode *inode, int fmode)
 283 {
 284         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
 285         struct ceph_inode_info *ci = ceph_inode(inode);
 286         struct ceph_mds_request *req;
 287         int err, flags, wanted;
 288
 289         spin_lock(&ci->i_ceph_lock);
 290         __ceph_touch_fmode(ci, mdsc, fmode);
 291         wanted = __ceph_caps_file_wanted(ci);
 292         if (__ceph_is_any_real_caps(ci) &&
 293             (!(wanted & CEPH_CAP_ANY_WR) || ci->i_auth_cap)) {
 294                 int issued = __ceph_caps_issued(ci, NULL);
 295                 spin_unlock(&ci->i_ceph_lock);
 296                 dout("renew caps %p want %s issued %s updating mds_wanted\n",
 297                      inode, ceph_cap_string(wanted), ceph_cap_string(issued));
 298                 ceph_check_caps(ci, 0, NULL);
 299                 return 0;
 300         }
 301         spin_unlock(&ci->i_ceph_lock);
 302
 303         flags = 0;
 304         if ((wanted & CEPH_CAP_FILE_RD) && (wanted & CEPH_CAP_FILE_WR))
 305                 flags = O_RDWR;
 306         else if (wanted & CEPH_CAP_FILE_RD)
 307                 flags = O_RDONLY;
 308         else if (wanted & CEPH_CAP_FILE_WR)
 309                 flags = O_WRONLY;
 310 #ifdef O_LAZY
 311         if (wanted & CEPH_CAP_FILE_LAZYIO)
 312                 flags |= O_LAZY;
 313 #endif
 314
 315         req = prepare_open_request(inode->i_sb, flags, 0);
 316         if (IS_ERR(req)) {
 317                 err = PTR_ERR(req);
 318                 goto out;
 319         }
 320
 321         req->r_inode = inode;
 322         ihold(inode);
 323         req->r_num_caps = 1;
 324
 325         err = ceph_mdsc_do_request(mdsc, NULL, req);
 326         ceph_mdsc_put_request(req);
 327 out:
 328         dout("renew caps %p open result=%d\n", inode, err);
 329         return err < 0 ? err : 0;
 330 }
 331
 332 /*
 333  * If we already have the requisite capabilities, we can satisfy
 334  * the open request locally (no need to request new caps from the
 335  * MDS).  We do, however, need to inform the MDS (asynchronously)
 336  * if our wanted caps set expands.
 337  */
 338 int ceph_open(struct inode *inode, struct file *file)
 339 {
 340         struct ceph_inode_info *ci = ceph_inode(inode);
 341         struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
 342         struct ceph_mds_client *mdsc = fsc->mdsc;
 343         struct ceph_mds_request *req;
 344         struct ceph_file_info *fi = file->private_data;
 345         int err;
 346         int flags, fmode, wanted;
 347
 348         if (fi) {
 349                 dout("open file %p is already opened\n", file);
 350                 return 0;
 351         }
 352
 353         /* filter out O_CREAT|O_EXCL; vfs did that already.  yuck. */
 354         flags = file->f_flags & ~(O_CREAT|O_EXCL);
 355         if (S_ISDIR(inode->i_mode))
 356                 flags = O_DIRECTORY;  /* mds likes to know */
 357
 358         dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
 359              ceph_vinop(inode), file, flags, file->f_flags);
 360         fmode = ceph_flags_to_mode(flags);
 361         wanted = ceph_caps_for_mode(fmode);
 362
 363         /* snapped files are read-only */
 364         if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
 365                 return -EROFS;
 366
 367         /* trivially open snapdir */
 368         if (ceph_snap(inode) == CEPH_SNAPDIR) {
 369                 return ceph_init_file(inode, file, fmode);
 370         }
 371
 372         /*
 373          * No need to block if we have caps on the auth MDS (for
 374          * write) or any MDS (for read).  Update wanted set
 375          * asynchronously.
 376          */
 377         spin_lock(&ci->i_ceph_lock);
 378         if (__ceph_is_any_real_caps(ci) &&
 379             (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
 380                 int mds_wanted = __ceph_caps_mds_wanted(ci, true);
 381                 int issued = __ceph_caps_issued(ci, NULL);
 382
 383                 dout("open %p fmode %d want %s issued %s using existing\n",
 384                      inode, fmode, ceph_cap_string(wanted),
 385                      ceph_cap_string(issued));
 386                 __ceph_touch_fmode(ci, mdsc, fmode);
 387                 spin_unlock(&ci->i_ceph_lock);
 388
 389                 /* adjust wanted? */
 390                 if ((issued & wanted) != wanted &&
 391                     (mds_wanted & wanted) != wanted &&
 392                     ceph_snap(inode) != CEPH_SNAPDIR)
 393                         ceph_check_caps(ci, 0, NULL);
 394
 395                 return ceph_init_file(inode, file, fmode);
 396         } else if (ceph_snap(inode) != CEPH_NOSNAP &&
 397                    (ci->i_snap_caps & wanted) == wanted) {
 398                 __ceph_touch_fmode(ci, mdsc, fmode);
 399                 spin_unlock(&ci->i_ceph_lock);
 400                 return ceph_init_file(inode, file, fmode);
 401         }
 402
 403         spin_unlock(&ci->i_ceph_lock);
 404
 405         dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
 406         req = prepare_open_request(inode->i_sb, flags, 0);
 407         if (IS_ERR(req)) {
 408                 err = PTR_ERR(req);
 409                 goto out;
 410         }
 411         req->r_inode = inode;
 412         ihold(inode);
 413
 414         req->r_num_caps = 1;
 415         err = ceph_mdsc_do_request(mdsc, NULL, req);
 416         if (!err)
 417                 err = ceph_init_file(inode, file, req->r_fmode);
 418         ceph_mdsc_put_request(req);
 419         dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
 420 out:
 421         return err;
 422 }
 423
 424 /* Clone the layout from a synchronous create, if the dir now has Dc caps */
 425 static void
 426 cache_file_layout(struct inode *dst, struct inode *src)
 427 {
 428         struct ceph_inode_info *cdst = ceph_inode(dst);
 429         struct ceph_inode_info *csrc = ceph_inode(src);
 430
 431         spin_lock(&cdst->i_ceph_lock);
 432         if ((__ceph_caps_issued(cdst, NULL) & CEPH_CAP_DIR_CREATE) &&
 433             !ceph_file_layout_is_valid(&cdst->i_cached_layout)) {
 434                 memcpy(&cdst->i_cached_layout, &csrc->i_layout,
 435                         sizeof(cdst->i_cached_layout));
 436                 rcu_assign_pointer(cdst->i_cached_layout.pool_ns,
 437                                    ceph_try_get_string(csrc->i_layout.pool_ns));
 438         }
 439         spin_unlock(&cdst->i_ceph_lock);
 440 }
 441
 442 /*
 443  * Try to set up an async create. We need caps, a file layout, and inode number,
 444  * and either a lease on the dentry or complete dir info. If any of those
 445  * criteria are not satisfied, then return false and the caller can go
 446  * synchronous.
 447  */
 448 static int try_prep_async_create(struct inode *dir, struct dentry *dentry,
 449                                  struct ceph_file_layout *lo, u64 *pino)
 450 {
 451         struct ceph_inode_info *ci = ceph_inode(dir);
 452         struct ceph_dentry_info *di = ceph_dentry(dentry);
 453         int got = 0, want = CEPH_CAP_FILE_EXCL | CEPH_CAP_DIR_CREATE;
 454         u64 ino;
 455
 456         spin_lock(&ci->i_ceph_lock);
 457         /* No auth cap means no chance for Dc caps */
 458         if (!ci->i_auth_cap)
 459                 goto no_async;
 460
 461         /* Any delegated inos? */
 462         if (xa_empty(&ci->i_auth_cap->session->s_delegated_inos))
 463                 goto no_async;
 464
 465         if (!ceph_file_layout_is_valid(&ci->i_cached_layout))
 466                 goto no_async;
 467
 468         if ((__ceph_caps_issued(ci, NULL) & want) != want)
 469                 goto no_async;
 470
 471         if (d_in_lookup(dentry)) {
 472                 if (!__ceph_dir_is_complete(ci))
 473                         goto no_async;
 474                 spin_lock(&dentry->d_lock);
 475                 di->lease_shared_gen = atomic_read(&ci->i_shared_gen);
 476                 spin_unlock(&dentry->d_lock);
 477         } else if (atomic_read(&ci->i_shared_gen) !=
 478                    READ_ONCE(di->lease_shared_gen)) {
 479                 goto no_async;
 480         }
 481
 482         ino = ceph_get_deleg_ino(ci->i_auth_cap->session);
 483         if (!ino)
 484                 goto no_async;
 485
 486         *pino = ino;
 487         ceph_take_cap_refs(ci, want, false);
 488         memcpy(lo, &ci->i_cached_layout, sizeof(*lo));
 489         rcu_assign_pointer(lo->pool_ns,
 490                            ceph_try_get_string(ci->i_cached_layout.pool_ns));
 491         got = want;
 492 no_async:
 493         spin_unlock(&ci->i_ceph_lock);
 494         return got;
 495 }
 496
 497 static void restore_deleg_ino(struct inode *dir, u64 ino)
 498 {
 499         struct ceph_inode_info *ci = ceph_inode(dir);
 500         struct ceph_mds_session *s = NULL;
 501
 502         spin_lock(&ci->i_ceph_lock);
 503         if (ci->i_auth_cap)
 504                 s = ceph_get_mds_session(ci->i_auth_cap->session);
 505         spin_unlock(&ci->i_ceph_lock);
 506         if (s) {
 507                 int err = ceph_restore_deleg_ino(s, ino);
 508                 if (err)
 509                         pr_warn("ceph: unable to restore delegated ino 0x%llx to session: %d\n",
 510                                 ino, err);
 511                 ceph_put_mds_session(s);
 512         }
 513 }
 514
 515 static void ceph_async_create_cb(struct ceph_mds_client *mdsc,
 516                                  struct ceph_mds_request *req)
 517 {
 518         int result = req->r_err ? req->r_err :
 519                         le32_to_cpu(req->r_reply_info.head->result);
 520
 521         if (result == -EJUKEBOX)
 522                 goto out;
 523
 524         mapping_set_error(req->r_parent->i_mapping, result);
 525
 526         if (result) {
 527                 struct dentry *dentry = req->r_dentry;
 528                 struct inode *inode = d_inode(dentry);
 529                 int pathlen = 0;
 530                 u64 base = 0;
 531                 char *path = ceph_mdsc_build_path(req->r_dentry, &pathlen,
 532                                                   &base, 0);
 533
 534                 ceph_dir_clear_complete(req->r_parent);
 535                 if (!d_unhashed(dentry))
 536                         d_drop(dentry);
 537
 538                 ceph_inode_shutdown(inode);
 539
 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 if (!result) {
 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_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 583         struct ceph_vino vino = { .ino = req->r_deleg_ino,
 584                                   .snap = CEPH_NOSNAP };
 585
 586         ktime_get_real_ts64(&now);
 587
 588         inode = ceph_get_inode(dentry->d_sb, vino);
 589         if (IS_ERR(inode))
 590                 return PTR_ERR(inode);
 591
 592         iinfo.inline_version = CEPH_INLINE_NONE;
 593         iinfo.change_attr = 1;
 594         ceph_encode_timespec64(&iinfo.btime, &now);
 595
 596         iinfo.xattr_len = ARRAY_SIZE(xattr_buf);
 597         iinfo.xattr_data = xattr_buf;
 598         memset(iinfo.xattr_data, 0, iinfo.xattr_len);
 599
 600         in.ino = cpu_to_le64(vino.ino);
 601         in.snapid = cpu_to_le64(CEPH_NOSNAP);
 602         in.version = cpu_to_le64(1);    // ???
 603         in.cap.caps = in.cap.wanted = cpu_to_le32(CEPH_CAP_ALL_FILE);
 604         in.cap.cap_id = cpu_to_le64(1);
 605         in.cap.realm = cpu_to_le64(ci->i_snap_realm->ino);
 606         in.cap.flags = CEPH_CAP_FLAG_AUTH;
 607         in.ctime = in.mtime = in.atime = iinfo.btime;
 608         in.mode = cpu_to_le32((u32)mode);
 609         in.truncate_seq = cpu_to_le32(1);
 610         in.truncate_size = cpu_to_le64(-1ULL);
 611         in.xattr_version = cpu_to_le64(1);
 612         in.uid = cpu_to_le32(from_kuid(&init_user_ns, current_fsuid()));
 613         in.gid = cpu_to_le32(from_kgid(&init_user_ns, dir->i_mode & S_ISGID ?
 614                                 dir->i_gid : current_fsgid()));
 615         in.nlink = cpu_to_le32(1);
 616         in.max_size = cpu_to_le64(lo->stripe_unit);
 617
 618         ceph_file_layout_to_legacy(lo, &in.layout);
 619
 620         down_read(&mdsc->snap_rwsem);
 621         ret = ceph_fill_inode(inode, NULL, &iinfo, NULL, req->r_session,
 622                               req->r_fmode, NULL);
 623         up_read(&mdsc->snap_rwsem);
 624         if (ret) {
 625                 dout("%s failed to fill inode: %d\n", __func__, ret);
 626                 ceph_dir_clear_complete(dir);
 627                 if (!d_unhashed(dentry))
 628                         d_drop(dentry);
 629                 if (inode->i_state & I_NEW)
 630                         discard_new_inode(inode);
 631         } else {
 632                 struct dentry *dn;
 633
 634                 dout("%s d_adding new inode 0x%llx to 0x%llx/%s\n", __func__,
 635                         vino.ino, ceph_ino(dir), dentry->d_name.name);
 636                 ceph_dir_clear_ordered(dir);
 637                 ceph_init_inode_acls(inode, as_ctx);
 638                 if (inode->i_state & I_NEW) {
 639                         /*
 640                          * If it's not I_NEW, then someone created this before
 641                          * we got here. Assume the server is aware of it at
 642                          * that point and don't worry about setting
 643                          * CEPH_I_ASYNC_CREATE.
 644                          */
 645                         ceph_inode(inode)->i_ceph_flags = CEPH_I_ASYNC_CREATE;
 646                         unlock_new_inode(inode);
 647                 }
 648                 if (d_in_lookup(dentry) || d_really_is_negative(dentry)) {
 649                         if (!d_unhashed(dentry))
 650                                 d_drop(dentry);
 651                         dn = d_splice_alias(inode, dentry);
 652                         WARN_ON_ONCE(dn && dn != dentry);
 653                 }
 654                 file->f_mode |= FMODE_CREATED;
 655                 ret = finish_open(file, dentry, ceph_open);
 656         }
 657         return ret;
 658 }
 659
 660 /*
 661  * Do a lookup + open with a single request.  If we get a non-existent
 662  * file or symlink, return 1 so the VFS can retry.
 663  */
 664 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
 665                      struct file *file, unsigned flags, umode_t mode)
 666 {
 667         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 668         struct ceph_mds_client *mdsc = fsc->mdsc;
 669         struct ceph_mds_request *req;
 670         struct dentry *dn;
 671         struct ceph_acl_sec_ctx as_ctx = {};
 672         bool try_async = ceph_test_mount_opt(fsc, ASYNC_DIROPS);
 673         int mask;
 674         int err;
 675
 676         dout("atomic_open %p dentry %p '%pd' %s flags %d mode 0%o\n",
 677              dir, dentry, dentry,
 678              d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
 679
 680         if (dentry->d_name.len > NAME_MAX)
 681                 return -ENAMETOOLONG;
 682
 683         if (flags & O_CREAT) {
 684                 if (ceph_quota_is_max_files_exceeded(dir))
 685                         return -EDQUOT;
 686                 err = ceph_pre_init_acls(dir, &mode, &as_ctx);
 687                 if (err < 0)
 688                         return err;
 689                 err = ceph_security_init_secctx(dentry, mode, &as_ctx);
 690                 if (err < 0)
 691                         goto out_ctx;
 692         } else if (!d_in_lookup(dentry)) {
 693                 /* If it's not being looked up, it's negative */
 694                 return -ENOENT;
 695         }
 696 retry:
 697         /* do the open */
 698         req = prepare_open_request(dir->i_sb, flags, mode);
 699         if (IS_ERR(req)) {
 700                 err = PTR_ERR(req);
 701                 goto out_ctx;
 702         }
 703         req->r_dentry = dget(dentry);
 704         req->r_num_caps = 2;
 705         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 706         if (ceph_security_xattr_wanted(dir))
 707                 mask |= CEPH_CAP_XATTR_SHARED;
 708         req->r_args.open.mask = cpu_to_le32(mask);
 709         req->r_parent = dir;
 710         ihold(dir);
 711
 712         if (flags & O_CREAT) {
 713                 struct ceph_file_layout lo;
 714
 715                 req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 716                 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 717                 if (as_ctx.pagelist) {
 718                         req->r_pagelist = as_ctx.pagelist;
 719                         as_ctx.pagelist = NULL;
 720                 }
 721                 if (try_async &&
 722                     (req->r_dir_caps =
 723                       try_prep_async_create(dir, dentry, &lo,
 724                                             &req->r_deleg_ino))) {
 725                         set_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags);
 726                         req->r_args.open.flags |= cpu_to_le32(CEPH_O_EXCL);
 727                         req->r_callback = ceph_async_create_cb;
 728                         err = ceph_mdsc_submit_request(mdsc, dir, req);
 729                         if (!err) {
 730                                 err = ceph_finish_async_create(dir, dentry,
 731                                                         file, mode, req,
 732                                                         &as_ctx, &lo);
 733                         } else if (err == -EJUKEBOX) {
 734                                 restore_deleg_ino(dir, req->r_deleg_ino);
 735                                 ceph_mdsc_put_request(req);
 736                                 try_async = false;
 737                                 goto retry;
 738                         }
 739                         goto out_req;
 740                 }
 741         }
 742
 743         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 744         err = ceph_mdsc_do_request(mdsc,
 745                                    (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
 746                                    req);
 747         if (err == -ENOENT) {
 748                 dentry = ceph_handle_snapdir(req, dentry);
 749                 if (IS_ERR(dentry)) {
 750                         err = PTR_ERR(dentry);
 751                         goto out_req;
 752                 }
 753                 err = 0;
 754         }
 755
 756         if (!err && (flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
 757                 err = ceph_handle_notrace_create(dir, dentry);
 758
 759         if (d_in_lookup(dentry)) {
 760                 dn = ceph_finish_lookup(req, dentry, err);
 761                 if (IS_ERR(dn))
 762                         err = PTR_ERR(dn);
 763         } else {
 764                 /* we were given a hashed negative dentry */
 765                 dn = NULL;
 766         }
 767         if (err)
 768                 goto out_req;
 769         if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
 770                 /* make vfs retry on splice, ENOENT, or symlink */
 771                 dout("atomic_open finish_no_open on dn %p\n", dn);
 772                 err = finish_no_open(file, dn);
 773         } else {
 774                 dout("atomic_open finish_open on dn %p\n", dn);
 775                 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
 776                         struct inode *newino = d_inode(dentry);
 777
 778                         cache_file_layout(dir, newino);
 779                         ceph_init_inode_acls(newino, &as_ctx);
 780                         file->f_mode |= FMODE_CREATED;
 781                 }
 782                 err = finish_open(file, dentry, ceph_open);
 783         }
 784 out_req:
 785         ceph_mdsc_put_request(req);
 786 out_ctx:
 787         ceph_release_acl_sec_ctx(&as_ctx);
 788         dout("atomic_open result=%d\n", err);
 789         return err;
 790 }
 791
 792 int ceph_release(struct inode *inode, struct file *file)
 793 {
 794         struct ceph_inode_info *ci = ceph_inode(inode);
 795
 796         if (S_ISDIR(inode->i_mode)) {
 797                 struct ceph_dir_file_info *dfi = file->private_data;
 798                 dout("release inode %p dir file %p\n", inode, file);
 799                 WARN_ON(!list_empty(&dfi->file_info.rw_contexts));
 800
 801                 ceph_put_fmode(ci, dfi->file_info.fmode, 1);
 802
 803                 if (dfi->last_readdir)
 804                         ceph_mdsc_put_request(dfi->last_readdir);
 805                 kfree(dfi->last_name);
 806                 kfree(dfi->dir_info);
 807                 kmem_cache_free(ceph_dir_file_cachep, dfi);
 808         } else {
 809                 struct ceph_file_info *fi = file->private_data;
 810                 dout("release inode %p regular file %p\n", inode, file);
 811                 WARN_ON(!list_empty(&fi->rw_contexts));
 812
 813                 ceph_put_fmode(ci, fi->fmode, 1);
 814
 815                 kmem_cache_free(ceph_file_cachep, fi);
 816         }
 817
 818         /* wake up anyone waiting for caps on this inode */
 819         wake_up_all(&ci->i_cap_wq);
 820         return 0;
 821 }
 822
 823 enum {
 824         HAVE_RETRIED = 1,
 825         CHECK_EOF =    2,
 826         READ_INLINE =  3,
 827 };
 828
 829 /*
 830  * Completely synchronous read and write methods.  Direct from __user
 831  * buffer to osd, or directly to user pages (if O_DIRECT).
 832  *
 833  * If the read spans object boundary, just do multiple reads.  (That's not
 834  * atomic, but good enough for now.)
 835  *
 836  * If we get a short result from the OSD, check against i_size; we need to
 837  * only return a short read to the caller if we hit EOF.
 838  */
 839 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *to,
 840                               int *retry_op)
 841 {
 842         struct file *file = iocb->ki_filp;
 843         struct inode *inode = file_inode(file);
 844         struct ceph_inode_info *ci = ceph_inode(inode);
 845         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 846         struct ceph_osd_client *osdc = &fsc->client->osdc;
 847         ssize_t ret;
 848         u64 off = iocb->ki_pos;
 849         u64 len = iov_iter_count(to);
 850
 851         dout("sync_read on file %p %llu~%u %s\n", file, off, (unsigned)len,
 852              (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
 853
 854         if (!len)
 855                 return 0;
 856         /*
 857          * flush any page cache pages in this range.  this
 858          * will make concurrent normal and sync io slow,
 859          * but it will at least behave sensibly when they are
 860          * in sequence.
 861          */
 862         ret = filemap_write_and_wait_range(inode->i_mapping,
 863                                            off, off + len - 1);
 864         if (ret < 0)
 865                 return ret;
 866
 867         ret = 0;
 868         while ((len = iov_iter_count(to)) > 0) {
 869                 struct ceph_osd_request *req;
 870                 struct page **pages;
 871                 int num_pages;
 872                 size_t page_off;
 873                 u64 i_size;
 874                 bool more;
 875                 int idx;
 876                 size_t left;
 877
 878                 req = ceph_osdc_new_request(osdc, &ci->i_layout,
 879                                         ci->i_vino, off, &len, 0, 1,
 880                                         CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
 881                                         NULL, ci->i_truncate_seq,
 882                                         ci->i_truncate_size, false);
 883                 if (IS_ERR(req)) {
 884                         ret = PTR_ERR(req);
 885                         break;
 886                 }
 887
 888                 more = len < iov_iter_count(to);
 889
 890                 num_pages = calc_pages_for(off, len);
 891                 page_off = off & ~PAGE_MASK;
 892                 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
 893                 if (IS_ERR(pages)) {
 894                         ceph_osdc_put_request(req);
 895                         ret = PTR_ERR(pages);
 896                         break;
 897                 }
 898
 899                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_off,
 900                                                  false, false);
 901                 ret = ceph_osdc_start_request(osdc, req, false);
 902                 if (!ret)
 903                         ret = ceph_osdc_wait_request(osdc, req);
 904
 905                 ceph_update_read_metrics(&fsc->mdsc->metric,
 906                                          req->r_start_latency,
 907                                          req->r_end_latency,
 908                                          len, ret);
 909
 910                 ceph_osdc_put_request(req);
 911
 912                 i_size = i_size_read(inode);
 913                 dout("sync_read %llu~%llu got %zd i_size %llu%s\n",
 914                      off, len, ret, i_size, (more ? " MORE" : ""));
 915
 916                 if (ret == -ENOENT)
 917                         ret = 0;
 918                 if (ret >= 0 && ret < len && (off + ret < i_size)) {
 919                         int zlen = min(len - ret, i_size - off - ret);
 920                         int zoff = page_off + ret;
 921                         dout("sync_read zero gap %llu~%llu\n",
 922                              off + ret, off + ret + zlen);
 923                         ceph_zero_page_vector_range(zoff, zlen, pages);
 924                         ret += zlen;
 925                 }
 926
 927                 idx = 0;
 928                 left = ret > 0 ? ret : 0;
 929                 while (left > 0) {
 930                         size_t len, copied;
 931                         page_off = off & ~PAGE_MASK;
 932                         len = min_t(size_t, left, PAGE_SIZE - page_off);
 933                         SetPageUptodate(pages[idx]);
 934                         copied = copy_page_to_iter(pages[idx++],
 935                                                    page_off, len, to);
 936                         off += copied;
 937                         left -= copied;
 938                         if (copied < len) {
 939                                 ret = -EFAULT;
 940                                 break;
 941                         }
 942                 }
 943                 ceph_release_page_vector(pages, num_pages);
 944
 945                 if (ret < 0) {
 946                         if (ret == -EBLOCKLISTED)
 947                                 fsc->blocklisted = true;
 948                         break;
 949                 }
 950
 951                 if (off >= i_size || !more)
 952                         break;
 953         }
 954
 955         if (off > iocb->ki_pos) {
 956                 if (ret >= 0 &&
 957                     iov_iter_count(to) > 0 && off >= i_size_read(inode))
 958                         *retry_op = CHECK_EOF;
 959                 ret = off - iocb->ki_pos;
 960                 iocb->ki_pos = off;
 961         }
 962
 963         dout("sync_read result %zd retry_op %d\n", ret, *retry_op);
 964         return ret;
 965 }
 966
 967 struct ceph_aio_request {
 968         struct kiocb *iocb;
 969         size_t total_len;
 970         bool write;
 971         bool should_dirty;
 972         int error;
 973         struct list_head osd_reqs;
 974         unsigned num_reqs;
 975         atomic_t pending_reqs;
 976         struct timespec64 mtime;
 977         struct ceph_cap_flush *prealloc_cf;
 978 };
 979
 980 struct ceph_aio_work {
 981         struct work_struct work;
 982         struct ceph_osd_request *req;
 983 };
 984
 985 static void ceph_aio_retry_work(struct work_struct *work);
 986
 987 static void ceph_aio_complete(struct inode *inode,
 988                               struct ceph_aio_request *aio_req)
 989 {
 990         struct ceph_inode_info *ci = ceph_inode(inode);
 991         int ret;
 992
 993         if (!atomic_dec_and_test(&aio_req->pending_reqs))
 994                 return;
 995
 996         if (aio_req->iocb->ki_flags & IOCB_DIRECT)
 997                 inode_dio_end(inode);
 998
 999         ret = aio_req->error;
1000         if (!ret)
1001                 ret = aio_req->total_len;
1002
1003         dout("ceph_aio_complete %p rc %d\n", inode, ret);
1004
1005         if (ret >= 0 && aio_req->write) {
1006                 int dirty;
1007
1008                 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
1009                 if (endoff > i_size_read(inode)) {
1010                         if (ceph_inode_set_size(inode, endoff))
1011                                 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
1012                 }
1013
1014                 spin_lock(&ci->i_ceph_lock);
1015                 ci->i_inline_version = CEPH_INLINE_NONE;
1016                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1017                                                &aio_req->prealloc_cf);
1018                 spin_unlock(&ci->i_ceph_lock);
1019                 if (dirty)
1020                         __mark_inode_dirty(inode, dirty);
1021
1022         }
1023
1024         ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
1025                                                 CEPH_CAP_FILE_RD));
1026
1027         aio_req->iocb->ki_complete(aio_req->iocb, ret, 0);
1028
1029         ceph_free_cap_flush(aio_req->prealloc_cf);
1030         kfree(aio_req);
1031 }
1032
1033 static void ceph_aio_complete_req(struct ceph_osd_request *req)
1034 {
1035         int rc = req->r_result;
1036         struct inode *inode = req->r_inode;
1037         struct ceph_aio_request *aio_req = req->r_priv;
1038         struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
1039         struct ceph_client_metric *metric = &ceph_sb_to_mdsc(inode->i_sb)->metric;
1040         unsigned int len = osd_data->bvec_pos.iter.bi_size;
1041
1042         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_BVECS);
1043         BUG_ON(!osd_data->num_bvecs);
1044
1045         dout("ceph_aio_complete_req %p rc %d bytes %u\n", inode, rc, len);
1046
1047         if (rc == -EOLDSNAPC) {
1048                 struct ceph_aio_work *aio_work;
1049                 BUG_ON(!aio_req->write);
1050
1051                 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
1052                 if (aio_work) {
1053                         INIT_WORK(&aio_work->work, ceph_aio_retry_work);
1054                         aio_work->req = req;
1055                         queue_work(ceph_inode_to_client(inode)->inode_wq,
1056                                    &aio_work->work);
1057                         return;
1058                 }
1059                 rc = -ENOMEM;
1060         } else if (!aio_req->write) {
1061                 if (rc == -ENOENT)
1062                         rc = 0;
1063                 if (rc >= 0 && len > rc) {
1064                         struct iov_iter i;
1065                         int zlen = len - rc;
1066
1067                         /*
1068                          * If read is satisfied by single OSD request,
1069                          * it can pass EOF. Otherwise read is within
1070                          * i_size.
1071                          */
1072                         if (aio_req->num_reqs == 1) {
1073                                 loff_t i_size = i_size_read(inode);
1074                                 loff_t endoff = aio_req->iocb->ki_pos + rc;
1075                                 if (endoff < i_size)
1076                                         zlen = min_t(size_t, zlen,
1077                                                      i_size - endoff);
1078                                 aio_req->total_len = rc + zlen;
1079                         }
1080
1081                         iov_iter_bvec(&i, READ, osd_data->bvec_pos.bvecs,
1082                                       osd_data->num_bvecs, len);
1083                         iov_iter_advance(&i, rc);
1084                         iov_iter_zero(zlen, &i);
1085                 }
1086         }
1087
1088         /* r_start_latency == 0 means the request was not submitted */
1089         if (req->r_start_latency) {
1090                 if (aio_req->write)
1091                         ceph_update_write_metrics(metric, req->r_start_latency,
1092                                                   req->r_end_latency, len, rc);
1093                 else
1094                         ceph_update_read_metrics(metric, req->r_start_latency,
1095                                                  req->r_end_latency, len, rc);
1096         }
1097
1098         put_bvecs(osd_data->bvec_pos.bvecs, osd_data->num_bvecs,
1099                   aio_req->should_dirty);
1100         ceph_osdc_put_request(req);
1101
1102         if (rc < 0)
1103                 cmpxchg(&aio_req->error, 0, rc);
1104
1105         ceph_aio_complete(inode, aio_req);
1106         return;
1107 }
1108
1109 static void ceph_aio_retry_work(struct work_struct *work)
1110 {
1111         struct ceph_aio_work *aio_work =
1112                 container_of(work, struct ceph_aio_work, work);
1113         struct ceph_osd_request *orig_req = aio_work->req;
1114         struct ceph_aio_request *aio_req = orig_req->r_priv;
1115         struct inode *inode = orig_req->r_inode;
1116         struct ceph_inode_info *ci = ceph_inode(inode);
1117         struct ceph_snap_context *snapc;
1118         struct ceph_osd_request *req;
1119         int ret;
1120
1121         spin_lock(&ci->i_ceph_lock);
1122         if (__ceph_have_pending_cap_snap(ci)) {
1123                 struct ceph_cap_snap *capsnap =
1124                         list_last_entry(&ci->i_cap_snaps,
1125                                         struct ceph_cap_snap,
1126                                         ci_item);
1127                 snapc = ceph_get_snap_context(capsnap->context);
1128         } else {
1129                 BUG_ON(!ci->i_head_snapc);
1130                 snapc = ceph_get_snap_context(ci->i_head_snapc);
1131         }
1132         spin_unlock(&ci->i_ceph_lock);
1133
1134         req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 1,
1135                         false, GFP_NOFS);
1136         if (!req) {
1137                 ret = -ENOMEM;
1138                 req = orig_req;
1139                 goto out;
1140         }
1141
1142         req->r_flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1143         ceph_oloc_copy(&req->r_base_oloc, &orig_req->r_base_oloc);
1144         ceph_oid_copy(&req->r_base_oid, &orig_req->r_base_oid);
1145
1146         req->r_ops[0] = orig_req->r_ops[0];
1147
1148         req->r_mtime = aio_req->mtime;
1149         req->r_data_offset = req->r_ops[0].extent.offset;
1150
1151         ret = ceph_osdc_alloc_messages(req, GFP_NOFS);
1152         if (ret) {
1153                 ceph_osdc_put_request(req);
1154                 req = orig_req;
1155                 goto out;
1156         }
1157
1158         ceph_osdc_put_request(orig_req);
1159
1160         req->r_callback = ceph_aio_complete_req;
1161         req->r_inode = inode;
1162         req->r_priv = aio_req;
1163
1164         ret = ceph_osdc_start_request(req->r_osdc, req, false);
1165 out:
1166         if (ret < 0) {
1167                 req->r_result = ret;
1168                 ceph_aio_complete_req(req);
1169         }
1170
1171         ceph_put_snap_context(snapc);
1172         kfree(aio_work);
1173 }
1174
1175 static ssize_t
1176 ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
1177                        struct ceph_snap_context *snapc,
1178                        struct ceph_cap_flush **pcf)
1179 {
1180         struct file *file = iocb->ki_filp;
1181         struct inode *inode = file_inode(file);
1182         struct ceph_inode_info *ci = ceph_inode(inode);
1183         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1184         struct ceph_client_metric *metric = &fsc->mdsc->metric;
1185         struct ceph_vino vino;
1186         struct ceph_osd_request *req;
1187         struct bio_vec *bvecs;
1188         struct ceph_aio_request *aio_req = NULL;
1189         int num_pages = 0;
1190         int flags;
1191         int ret = 0;
1192         struct timespec64 mtime = current_time(inode);
1193         size_t count = iov_iter_count(iter);
1194         loff_t pos = iocb->ki_pos;
1195         bool write = iov_iter_rw(iter) == WRITE;
1196         bool should_dirty = !write && iter_is_iovec(iter);
1197
1198         if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1199                 return -EROFS;
1200
1201         dout("sync_direct_%s on file %p %lld~%u snapc %p seq %lld\n",
1202              (write ? "write" : "read"), file, pos, (unsigned)count,
1203              snapc, snapc ? snapc->seq : 0);
1204
1205         if (write) {
1206                 int ret2 = invalidate_inode_pages2_range(inode->i_mapping,
1207                                         pos >> PAGE_SHIFT,
1208                                         (pos + count - 1) >> PAGE_SHIFT);
1209                 if (ret2 < 0)
1210                         dout("invalidate_inode_pages2_range returned %d\n", ret2);
1211
1212                 flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1213         } else {
1214                 flags = CEPH_OSD_FLAG_READ;
1215         }
1216
1217         while (iov_iter_count(iter) > 0) {
1218                 u64 size = iov_iter_count(iter);
1219                 ssize_t len;
1220
1221                 if (write)
1222                         size = min_t(u64, size, fsc->mount_options->wsize);
1223                 else
1224                         size = min_t(u64, size, fsc->mount_options->rsize);
1225
1226                 vino = ceph_vino(inode);
1227                 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1228                                             vino, pos, &size, 0,
1229                                             1,
1230                                             write ? CEPH_OSD_OP_WRITE :
1231                                                     CEPH_OSD_OP_READ,
1232                                             flags, snapc,
1233                                             ci->i_truncate_seq,
1234                                             ci->i_truncate_size,
1235                                             false);
1236                 if (IS_ERR(req)) {
1237                         ret = PTR_ERR(req);
1238                         break;
1239                 }
1240
1241                 len = iter_get_bvecs_alloc(iter, size, &bvecs, &num_pages);
1242                 if (len < 0) {
1243                         ceph_osdc_put_request(req);
1244                         ret = len;
1245                         break;
1246                 }
1247                 if (len != size)
1248                         osd_req_op_extent_update(req, 0, len);
1249
1250                 /*
1251                  * To simplify error handling, allow AIO when IO within i_size
1252                  * or IO can be satisfied by single OSD request.
1253                  */
1254                 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
1255                     (len == count || pos + count <= i_size_read(inode))) {
1256                         aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
1257                         if (aio_req) {
1258                                 aio_req->iocb = iocb;
1259                                 aio_req->write = write;
1260                                 aio_req->should_dirty = should_dirty;
1261                                 INIT_LIST_HEAD(&aio_req->osd_reqs);
1262                                 if (write) {
1263                                         aio_req->mtime = mtime;
1264                                         swap(aio_req->prealloc_cf, *pcf);
1265                                 }
1266                         }
1267                         /* ignore error */
1268                 }
1269
1270                 if (write) {
1271                         /*
1272                          * throw out any page cache pages in this range. this
1273                          * may block.
1274                          */
1275                         truncate_inode_pages_range(inode->i_mapping, pos,
1276                                                    PAGE_ALIGN(pos + len) - 1);
1277
1278                         req->r_mtime = mtime;
1279                 }
1280
1281                 osd_req_op_extent_osd_data_bvecs(req, 0, bvecs, num_pages, len);
1282
1283                 if (aio_req) {
1284                         aio_req->total_len += len;
1285                         aio_req->num_reqs++;
1286                         atomic_inc(&aio_req->pending_reqs);
1287
1288                         req->r_callback = ceph_aio_complete_req;
1289                         req->r_inode = inode;
1290                         req->r_priv = aio_req;
1291                         list_add_tail(&req->r_private_item, &aio_req->osd_reqs);
1292
1293                         pos += len;
1294                         continue;
1295                 }
1296
1297                 ret = ceph_osdc_start_request(req->r_osdc, req, false);
1298                 if (!ret)
1299                         ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1300
1301                 if (write)
1302                         ceph_update_write_metrics(metric, req->r_start_latency,
1303                                                   req->r_end_latency, len, ret);
1304                 else
1305                         ceph_update_read_metrics(metric, req->r_start_latency,
1306                                                  req->r_end_latency, len, ret);
1307
1308                 size = i_size_read(inode);
1309                 if (!write) {
1310                         if (ret == -ENOENT)
1311                                 ret = 0;
1312                         if (ret >= 0 && ret < len && pos + ret < size) {
1313                                 struct iov_iter i;
1314                                 int zlen = min_t(size_t, len - ret,
1315                                                  size - pos - ret);
1316
1317                                 iov_iter_bvec(&i, READ, bvecs, num_pages, len);
1318                                 iov_iter_advance(&i, ret);
1319                                 iov_iter_zero(zlen, &i);
1320                                 ret += zlen;
1321                         }
1322                         if (ret >= 0)
1323                                 len = ret;
1324                 }
1325
1326                 put_bvecs(bvecs, num_pages, should_dirty);
1327                 ceph_osdc_put_request(req);
1328                 if (ret < 0)
1329                         break;
1330
1331                 pos += len;
1332                 if (!write && pos >= size)
1333                         break;
1334
1335                 if (write && pos > size) {
1336                         if (ceph_inode_set_size(inode, pos))
1337                                 ceph_check_caps(ceph_inode(inode),
1338                                                 CHECK_CAPS_AUTHONLY,
1339                                                 NULL);
1340                 }
1341         }
1342
1343         if (aio_req) {
1344                 LIST_HEAD(osd_reqs);
1345
1346                 if (aio_req->num_reqs == 0) {
1347                         kfree(aio_req);
1348                         return ret;
1349                 }
1350
1351                 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
1352                                               CEPH_CAP_FILE_RD);
1353
1354                 list_splice(&aio_req->osd_reqs, &osd_reqs);
1355                 inode_dio_begin(inode);
1356                 while (!list_empty(&osd_reqs)) {
1357                         req = list_first_entry(&osd_reqs,
1358                                                struct ceph_osd_request,
1359                                                r_private_item);
1360                         list_del_init(&req->r_private_item);
1361                         if (ret >= 0)
1362                                 ret = ceph_osdc_start_request(req->r_osdc,
1363                                                               req, false);
1364                         if (ret < 0) {
1365                                 req->r_result = ret;
1366                                 ceph_aio_complete_req(req);
1367                         }
1368                 }
1369                 return -EIOCBQUEUED;
1370         }
1371
1372         if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
1373                 ret = pos - iocb->ki_pos;
1374                 iocb->ki_pos = pos;
1375         }
1376         return ret;
1377 }
1378
1379 /*
1380  * Synchronous write, straight from __user pointer or user pages.
1381  *
1382  * If write spans object boundary, just do multiple writes.  (For a
1383  * correct atomic write, we should e.g. take write locks on all
1384  * objects, rollback on failure, etc.)
1385  */
1386 static ssize_t
1387 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
1388                 struct ceph_snap_context *snapc)
1389 {
1390         struct file *file = iocb->ki_filp;
1391         struct inode *inode = file_inode(file);
1392         struct ceph_inode_info *ci = ceph_inode(inode);
1393         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1394         struct ceph_vino vino;
1395         struct ceph_osd_request *req;
1396         struct page **pages;
1397         u64 len;
1398         int num_pages;
1399         int written = 0;
1400         int flags;
1401         int ret;
1402         bool check_caps = false;
1403         struct timespec64 mtime = current_time(inode);
1404         size_t count = iov_iter_count(from);
1405
1406         if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1407                 return -EROFS;
1408
1409         dout("sync_write on file %p %lld~%u snapc %p seq %lld\n",
1410              file, pos, (unsigned)count, snapc, snapc->seq);
1411
1412         ret = filemap_write_and_wait_range(inode->i_mapping,
1413                                            pos, pos + count - 1);
1414         if (ret < 0)
1415                 return ret;
1416
1417         ret = invalidate_inode_pages2_range(inode->i_mapping,
1418                                             pos >> PAGE_SHIFT,
1419                                             (pos + count - 1) >> PAGE_SHIFT);
1420         if (ret < 0)
1421                 dout("invalidate_inode_pages2_range returned %d\n", ret);
1422
1423         flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1424
1425         while ((len = iov_iter_count(from)) > 0) {
1426                 size_t left;
1427                 int n;
1428
1429                 vino = ceph_vino(inode);
1430                 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1431                                             vino, pos, &len, 0, 1,
1432                                             CEPH_OSD_OP_WRITE, flags, snapc,
1433                                             ci->i_truncate_seq,
1434                                             ci->i_truncate_size,
1435                                             false);
1436                 if (IS_ERR(req)) {
1437                         ret = PTR_ERR(req);
1438                         break;
1439                 }
1440
1441                 /*
1442                  * write from beginning of first page,
1443                  * regardless of io alignment
1444                  */
1445                 num_pages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1446
1447                 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1448                 if (IS_ERR(pages)) {
1449                         ret = PTR_ERR(pages);
1450                         goto out;
1451                 }
1452
1453                 left = len;
1454                 for (n = 0; n < num_pages; n++) {
1455                         size_t plen = min_t(size_t, left, PAGE_SIZE);
1456                         ret = copy_page_from_iter(pages[n], 0, plen, from);
1457                         if (ret != plen) {
1458                                 ret = -EFAULT;
1459                                 break;
1460                         }
1461                         left -= ret;
1462                 }
1463
1464                 if (ret < 0) {
1465                         ceph_release_page_vector(pages, num_pages);
1466                         goto out;
1467                 }
1468
1469                 req->r_inode = inode;
1470
1471                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
1472                                                 false, true);
1473
1474                 req->r_mtime = mtime;
1475                 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1476                 if (!ret)
1477                         ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1478
1479                 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1480                                           req->r_end_latency, len, ret);
1481 out:
1482                 ceph_osdc_put_request(req);
1483                 if (ret != 0) {
1484                         ceph_set_error_write(ci);
1485                         break;
1486                 }
1487
1488                 ceph_clear_error_write(ci);
1489                 pos += len;
1490                 written += len;
1491                 if (pos > i_size_read(inode)) {
1492                         check_caps = ceph_inode_set_size(inode, pos);
1493                         if (check_caps)
1494                                 ceph_check_caps(ceph_inode(inode),
1495                                                 CHECK_CAPS_AUTHONLY,
1496                                                 NULL);
1497                 }
1498
1499         }
1500
1501         if (ret != -EOLDSNAPC && written > 0) {
1502                 ret = written;
1503                 iocb->ki_pos = pos;
1504         }
1505         return ret;
1506 }
1507
1508 /*
1509  * Wrap generic_file_aio_read with checks for cap bits on the inode.
1510  * Atomically grab references, so that those bits are not released
1511  * back to the MDS mid-read.
1512  *
1513  * Hmm, the sync read case isn't actually async... should it be?
1514  */
1515 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
1516 {
1517         struct file *filp = iocb->ki_filp;
1518         struct ceph_file_info *fi = filp->private_data;
1519         size_t len = iov_iter_count(to);
1520         struct inode *inode = file_inode(filp);
1521         struct ceph_inode_info *ci = ceph_inode(inode);
1522         bool direct_lock = iocb->ki_flags & IOCB_DIRECT;
1523         ssize_t ret;
1524         int want, got = 0;
1525         int retry_op = 0, read = 0;
1526
1527 again:
1528         dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
1529              inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
1530
1531         if (ceph_inode_is_shutdown(inode))
1532                 return -ESTALE;
1533
1534         if (direct_lock)
1535                 ceph_start_io_direct(inode);
1536         else
1537                 ceph_start_io_read(inode);
1538
1539         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1540                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1541         else
1542                 want = CEPH_CAP_FILE_CACHE;
1543         ret = ceph_get_caps(filp, CEPH_CAP_FILE_RD, want, -1, &got);
1544         if (ret < 0) {
1545                 if (iocb->ki_flags & IOCB_DIRECT)
1546                         ceph_end_io_direct(inode);
1547                 else
1548                         ceph_end_io_read(inode);
1549                 return ret;
1550         }
1551
1552         if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1553             (iocb->ki_flags & IOCB_DIRECT) ||
1554             (fi->flags & CEPH_F_SYNC)) {
1555
1556                 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1557                      inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1558                      ceph_cap_string(got));
1559
1560                 if (ci->i_inline_version == CEPH_INLINE_NONE) {
1561                         if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
1562                                 ret = ceph_direct_read_write(iocb, to,
1563                                                              NULL, NULL);
1564                                 if (ret >= 0 && ret < len)
1565                                         retry_op = CHECK_EOF;
1566                         } else {
1567                                 ret = ceph_sync_read(iocb, to, &retry_op);
1568                         }
1569                 } else {
1570                         retry_op = READ_INLINE;
1571                 }
1572         } else {
1573                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1574                 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1575                      inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1576                      ceph_cap_string(got));
1577                 ceph_add_rw_context(fi, &rw_ctx);
1578                 ret = generic_file_read_iter(iocb, to);
1579                 ceph_del_rw_context(fi, &rw_ctx);
1580         }
1581
1582         dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
1583              inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
1584         ceph_put_cap_refs(ci, got);
1585
1586         if (direct_lock)
1587                 ceph_end_io_direct(inode);
1588         else
1589                 ceph_end_io_read(inode);
1590
1591         if (retry_op > HAVE_RETRIED && ret >= 0) {
1592                 int statret;
1593                 struct page *page = NULL;
1594                 loff_t i_size;
1595                 if (retry_op == READ_INLINE) {
1596                         page = __page_cache_alloc(GFP_KERNEL);
1597                         if (!page)
1598                                 return -ENOMEM;
1599                 }
1600
1601                 statret = __ceph_do_getattr(inode, page,
1602                                             CEPH_STAT_CAP_INLINE_DATA, !!page);
1603                 if (statret < 0) {
1604                         if (page)
1605                                 __free_page(page);
1606                         if (statret == -ENODATA) {
1607                                 BUG_ON(retry_op != READ_INLINE);
1608                                 goto again;
1609                         }
1610                         return statret;
1611                 }
1612
1613                 i_size = i_size_read(inode);
1614                 if (retry_op == READ_INLINE) {
1615                         BUG_ON(ret > 0 || read > 0);
1616                         if (iocb->ki_pos < i_size &&
1617                             iocb->ki_pos < PAGE_SIZE) {
1618                                 loff_t end = min_t(loff_t, i_size,
1619                                                    iocb->ki_pos + len);
1620                                 end = min_t(loff_t, end, PAGE_SIZE);
1621                                 if (statret < end)
1622                                         zero_user_segment(page, statret, end);
1623                                 ret = copy_page_to_iter(page,
1624                                                 iocb->ki_pos & ~PAGE_MASK,
1625                                                 end - iocb->ki_pos, to);
1626                                 iocb->ki_pos += ret;
1627                                 read += ret;
1628                         }
1629                         if (iocb->ki_pos < i_size && read < len) {
1630                                 size_t zlen = min_t(size_t, len - read,
1631                                                     i_size - iocb->ki_pos);
1632                                 ret = iov_iter_zero(zlen, to);
1633                                 iocb->ki_pos += ret;
1634                                 read += ret;
1635                         }
1636                         __free_pages(page, 0);
1637                         return read;
1638                 }
1639
1640                 /* hit EOF or hole? */
1641                 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
1642                     ret < len) {
1643                         dout("sync_read hit hole, ppos %lld < size %lld"
1644                              ", reading more\n", iocb->ki_pos, i_size);
1645
1646                         read += ret;
1647                         len -= ret;
1648                         retry_op = HAVE_RETRIED;
1649                         goto again;
1650                 }
1651         }
1652
1653         if (ret >= 0)
1654                 ret += read;
1655
1656         return ret;
1657 }
1658
1659 /*
1660  * Take cap references to avoid releasing caps to MDS mid-write.
1661  *
1662  * If we are synchronous, and write with an old snap context, the OSD
1663  * may return EOLDSNAPC.  In that case, retry the write.. _after_
1664  * dropping our cap refs and allowing the pending snap to logically
1665  * complete _before_ this write occurs.
1666  *
1667  * If we are near ENOSPC, write synchronously.
1668  */
1669 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
1670 {
1671         struct file *file = iocb->ki_filp;
1672         struct ceph_file_info *fi = file->private_data;
1673         struct inode *inode = file_inode(file);
1674         struct ceph_inode_info *ci = ceph_inode(inode);
1675         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1676         struct ceph_osd_client *osdc = &fsc->client->osdc;
1677         struct ceph_cap_flush *prealloc_cf;
1678         ssize_t count, written = 0;
1679         int err, want, got;
1680         bool direct_lock = false;
1681         u32 map_flags;
1682         u64 pool_flags;
1683         loff_t pos;
1684         loff_t limit = max(i_size_read(inode), fsc->max_file_size);
1685
1686         if (ceph_inode_is_shutdown(inode))
1687                 return -ESTALE;
1688
1689         if (ceph_snap(inode) != CEPH_NOSNAP)
1690                 return -EROFS;
1691
1692         prealloc_cf = ceph_alloc_cap_flush();
1693         if (!prealloc_cf)
1694                 return -ENOMEM;
1695
1696         if ((iocb->ki_flags & (IOCB_DIRECT | IOCB_APPEND)) == IOCB_DIRECT)
1697                 direct_lock = true;
1698
1699 retry_snap:
1700         if (direct_lock)
1701                 ceph_start_io_direct(inode);
1702         else
1703                 ceph_start_io_write(inode);
1704
1705         /* We can write back this queue in page reclaim */
1706         current->backing_dev_info = inode_to_bdi(inode);
1707
1708         if (iocb->ki_flags & IOCB_APPEND) {
1709                 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1710                 if (err < 0)
1711                         goto out;
1712         }
1713
1714         err = generic_write_checks(iocb, from);
1715         if (err <= 0)
1716                 goto out;
1717
1718         pos = iocb->ki_pos;
1719         if (unlikely(pos >= limit)) {
1720                 err = -EFBIG;
1721                 goto out;
1722         } else {
1723                 iov_iter_truncate(from, limit - pos);
1724         }
1725
1726         count = iov_iter_count(from);
1727         if (ceph_quota_is_max_bytes_exceeded(inode, pos + count)) {
1728                 err = -EDQUOT;
1729                 goto out;
1730         }
1731
1732         down_read(&osdc->lock);
1733         map_flags = osdc->osdmap->flags;
1734         pool_flags = ceph_pg_pool_flags(osdc->osdmap, ci->i_layout.pool_id);
1735         up_read(&osdc->lock);
1736         if ((map_flags & CEPH_OSDMAP_FULL) ||
1737             (pool_flags & CEPH_POOL_FLAG_FULL)) {
1738                 err = -ENOSPC;
1739                 goto out;
1740         }
1741
1742         err = file_remove_privs(file);
1743         if (err)
1744                 goto out;
1745
1746         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1747                 err = ceph_uninline_data(file, NULL);
1748                 if (err < 0)
1749                         goto out;
1750         }
1751
1752         dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
1753              inode, ceph_vinop(inode), pos, count, i_size_read(inode));
1754         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1755                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1756         else
1757                 want = CEPH_CAP_FILE_BUFFER;
1758         got = 0;
1759         err = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, pos + count, &got);
1760         if (err < 0)
1761                 goto out;
1762
1763         err = file_update_time(file);
1764         if (err)
1765                 goto out_caps;
1766
1767         inode_inc_iversion_raw(inode);
1768
1769         dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
1770              inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
1771
1772         if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1773             (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC) ||
1774             (ci->i_ceph_flags & CEPH_I_ERROR_WRITE)) {
1775                 struct ceph_snap_context *snapc;
1776                 struct iov_iter data;
1777
1778                 spin_lock(&ci->i_ceph_lock);
1779                 if (__ceph_have_pending_cap_snap(ci)) {
1780                         struct ceph_cap_snap *capsnap =
1781                                         list_last_entry(&ci->i_cap_snaps,
1782                                                         struct ceph_cap_snap,
1783                                                         ci_item);
1784                         snapc = ceph_get_snap_context(capsnap->context);
1785                 } else {
1786                         BUG_ON(!ci->i_head_snapc);
1787                         snapc = ceph_get_snap_context(ci->i_head_snapc);
1788                 }
1789                 spin_unlock(&ci->i_ceph_lock);
1790
1791                 /* we might need to revert back to that point */
1792                 data = *from;
1793                 if (iocb->ki_flags & IOCB_DIRECT)
1794                         written = ceph_direct_read_write(iocb, &data, snapc,
1795                                                          &prealloc_cf);
1796                 else
1797                         written = ceph_sync_write(iocb, &data, pos, snapc);
1798                 if (direct_lock)
1799                         ceph_end_io_direct(inode);
1800                 else
1801                         ceph_end_io_write(inode);
1802                 if (written > 0)
1803                         iov_iter_advance(from, written);
1804                 ceph_put_snap_context(snapc);
1805         } else {
1806                 /*
1807                  * No need to acquire the i_truncate_mutex. Because
1808                  * the MDS revokes Fwb caps before sending truncate
1809                  * message to us. We can't get Fwb cap while there
1810                  * are pending vmtruncate. So write and vmtruncate
1811                  * can not run at the same time
1812                  */
1813                 written = generic_perform_write(file, from, pos);
1814                 if (likely(written >= 0))
1815                         iocb->ki_pos = pos + written;
1816                 ceph_end_io_write(inode);
1817         }
1818
1819         if (written >= 0) {
1820                 int dirty;
1821
1822                 spin_lock(&ci->i_ceph_lock);
1823                 ci->i_inline_version = CEPH_INLINE_NONE;
1824                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1825                                                &prealloc_cf);
1826                 spin_unlock(&ci->i_ceph_lock);
1827                 if (dirty)
1828                         __mark_inode_dirty(inode, dirty);
1829                 if (ceph_quota_is_max_bytes_approaching(inode, iocb->ki_pos))
1830                         ceph_check_caps(ci, 0, NULL);
1831         }
1832
1833         dout("aio_write %p %llx.%llx %llu~%u  dropping cap refs on %s\n",
1834              inode, ceph_vinop(inode), pos, (unsigned)count,
1835              ceph_cap_string(got));
1836         ceph_put_cap_refs(ci, got);
1837
1838         if (written == -EOLDSNAPC) {
1839                 dout("aio_write %p %llx.%llx %llu~%u" "got EOLDSNAPC, retrying\n",
1840                      inode, ceph_vinop(inode), pos, (unsigned)count);
1841                 goto retry_snap;
1842         }
1843
1844         if (written >= 0) {
1845                 if ((map_flags & CEPH_OSDMAP_NEARFULL) ||
1846                     (pool_flags & CEPH_POOL_FLAG_NEARFULL))
1847                         iocb->ki_flags |= IOCB_DSYNC;
1848                 written = generic_write_sync(iocb, written);
1849         }
1850
1851         goto out_unlocked;
1852 out_caps:
1853         ceph_put_cap_refs(ci, got);
1854 out:
1855         if (direct_lock)
1856                 ceph_end_io_direct(inode);
1857         else
1858                 ceph_end_io_write(inode);
1859 out_unlocked:
1860         ceph_free_cap_flush(prealloc_cf);
1861         current->backing_dev_info = NULL;
1862         return written ? written : err;
1863 }
1864
1865 /*
1866  * llseek.  be sure to verify file size on SEEK_END.
1867  */
1868 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
1869 {
1870         struct inode *inode = file->f_mapping->host;
1871         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1872         loff_t i_size;
1873         loff_t ret;
1874
1875         inode_lock(inode);
1876
1877         if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
1878                 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1879                 if (ret < 0)
1880                         goto out;
1881         }
1882
1883         i_size = i_size_read(inode);
1884         switch (whence) {
1885         case SEEK_END:
1886                 offset += i_size;
1887                 break;
1888         case SEEK_CUR:
1889                 /*
1890                  * Here we special-case the lseek(fd, 0, SEEK_CUR)
1891                  * position-querying operation.  Avoid rewriting the "same"
1892                  * f_pos value back to the file because a concurrent read(),
1893                  * write() or lseek() might have altered it
1894                  */
1895                 if (offset == 0) {
1896                         ret = file->f_pos;
1897                         goto out;
1898                 }
1899                 offset += file->f_pos;
1900                 break;
1901         case SEEK_DATA:
1902                 if (offset < 0 || offset >= i_size) {
1903                         ret = -ENXIO;
1904                         goto out;
1905                 }
1906                 break;
1907         case SEEK_HOLE:
1908                 if (offset < 0 || offset >= i_size) {
1909                         ret = -ENXIO;
1910                         goto out;
1911                 }
1912                 offset = i_size;
1913                 break;
1914         }
1915
1916         ret = vfs_setpos(file, offset, max(i_size, fsc->max_file_size));
1917
1918 out:
1919         inode_unlock(inode);
1920         return ret;
1921 }
1922
1923 static inline void ceph_zero_partial_page(
1924         struct inode *inode, loff_t offset, unsigned size)
1925 {
1926         struct page *page;
1927         pgoff_t index = offset >> PAGE_SHIFT;
1928
1929         page = find_lock_page(inode->i_mapping, index);
1930         if (page) {
1931                 wait_on_page_writeback(page);
1932                 zero_user(page, offset & (PAGE_SIZE - 1), size);
1933                 unlock_page(page);
1934                 put_page(page);
1935         }
1936 }
1937
1938 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
1939                                       loff_t length)
1940 {
1941         loff_t nearly = round_up(offset, PAGE_SIZE);
1942         if (offset < nearly) {
1943                 loff_t size = nearly - offset;
1944                 if (length < size)
1945                         size = length;
1946                 ceph_zero_partial_page(inode, offset, size);
1947                 offset += size;
1948                 length -= size;
1949         }
1950         if (length >= PAGE_SIZE) {
1951                 loff_t size = round_down(length, PAGE_SIZE);
1952                 truncate_pagecache_range(inode, offset, offset + size - 1);
1953                 offset += size;
1954                 length -= size;
1955         }
1956         if (length)
1957                 ceph_zero_partial_page(inode, offset, length);
1958 }
1959
1960 static int ceph_zero_partial_object(struct inode *inode,
1961                                     loff_t offset, loff_t *length)
1962 {
1963         struct ceph_inode_info *ci = ceph_inode(inode);
1964         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1965         struct ceph_osd_request *req;
1966         int ret = 0;
1967         loff_t zero = 0;
1968         int op;
1969
1970         if (!length) {
1971                 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
1972                 length = &zero;
1973         } else {
1974                 op = CEPH_OSD_OP_ZERO;
1975         }
1976
1977         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1978                                         ceph_vino(inode),
1979                                         offset, length,
1980                                         0, 1, op,
1981                                         CEPH_OSD_FLAG_WRITE,
1982                                         NULL, 0, 0, false);
1983         if (IS_ERR(req)) {
1984                 ret = PTR_ERR(req);
1985                 goto out;
1986         }
1987
1988         req->r_mtime = inode->i_mtime;
1989         ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1990         if (!ret) {
1991                 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1992                 if (ret == -ENOENT)
1993                         ret = 0;
1994         }
1995         ceph_osdc_put_request(req);
1996
1997 out:
1998         return ret;
1999 }
2000
2001 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
2002 {
2003         int ret = 0;
2004         struct ceph_inode_info *ci = ceph_inode(inode);
2005         s32 stripe_unit = ci->i_layout.stripe_unit;
2006         s32 stripe_count = ci->i_layout.stripe_count;
2007         s32 object_size = ci->i_layout.object_size;
2008         u64 object_set_size = object_size * stripe_count;
2009         u64 nearly, t;
2010
2011         /* round offset up to next period boundary */
2012         nearly = offset + object_set_size - 1;
2013         t = nearly;
2014         nearly -= do_div(t, object_set_size);
2015
2016         while (length && offset < nearly) {
2017                 loff_t size = length;
2018                 ret = ceph_zero_partial_object(inode, offset, &size);
2019                 if (ret < 0)
2020                         return ret;
2021                 offset += size;
2022                 length -= size;
2023         }
2024         while (length >= object_set_size) {
2025                 int i;
2026                 loff_t pos = offset;
2027                 for (i = 0; i < stripe_count; ++i) {
2028                         ret = ceph_zero_partial_object(inode, pos, NULL);
2029                         if (ret < 0)
2030                                 return ret;
2031                         pos += stripe_unit;
2032                 }
2033                 offset += object_set_size;
2034                 length -= object_set_size;
2035         }
2036         while (length) {
2037                 loff_t size = length;
2038                 ret = ceph_zero_partial_object(inode, offset, &size);
2039                 if (ret < 0)
2040                         return ret;
2041                 offset += size;
2042                 length -= size;
2043         }
2044         return ret;
2045 }
2046
2047 static long ceph_fallocate(struct file *file, int mode,
2048                                 loff_t offset, loff_t length)
2049 {
2050         struct ceph_file_info *fi = file->private_data;
2051         struct inode *inode = file_inode(file);
2052         struct ceph_inode_info *ci = ceph_inode(inode);
2053         struct ceph_cap_flush *prealloc_cf;
2054         int want, got = 0;
2055         int dirty;
2056         int ret = 0;
2057         loff_t endoff = 0;
2058         loff_t size;
2059
2060         if (mode != (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2061                 return -EOPNOTSUPP;
2062
2063         if (!S_ISREG(inode->i_mode))
2064                 return -EOPNOTSUPP;
2065
2066         prealloc_cf = ceph_alloc_cap_flush();
2067         if (!prealloc_cf)
2068                 return -ENOMEM;
2069
2070         inode_lock(inode);
2071
2072         if (ceph_snap(inode) != CEPH_NOSNAP) {
2073                 ret = -EROFS;
2074                 goto unlock;
2075         }
2076
2077         if (ci->i_inline_version != CEPH_INLINE_NONE) {
2078                 ret = ceph_uninline_data(file, NULL);
2079                 if (ret < 0)
2080                         goto unlock;
2081         }
2082
2083         size = i_size_read(inode);
2084
2085         /* Are we punching a hole beyond EOF? */
2086         if (offset >= size)
2087                 goto unlock;
2088         if ((offset + length) > size)
2089                 length = size - offset;
2090
2091         if (fi->fmode & CEPH_FILE_MODE_LAZY)
2092                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
2093         else
2094                 want = CEPH_CAP_FILE_BUFFER;
2095
2096         ret = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, endoff, &got);
2097         if (ret < 0)
2098                 goto unlock;
2099
2100         filemap_invalidate_lock(inode->i_mapping);
2101         ceph_zero_pagecache_range(inode, offset, length);
2102         ret = ceph_zero_objects(inode, offset, length);
2103
2104         if (!ret) {
2105                 spin_lock(&ci->i_ceph_lock);
2106                 ci->i_inline_version = CEPH_INLINE_NONE;
2107                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
2108                                                &prealloc_cf);
2109                 spin_unlock(&ci->i_ceph_lock);
2110                 if (dirty)
2111                         __mark_inode_dirty(inode, dirty);
2112         }
2113         filemap_invalidate_unlock(inode->i_mapping);
2114
2115         ceph_put_cap_refs(ci, got);
2116 unlock:
2117         inode_unlock(inode);
2118         ceph_free_cap_flush(prealloc_cf);
2119         return ret;
2120 }
2121
2122 /*
2123  * This function tries to get FILE_WR capabilities for dst_ci and FILE_RD for
2124  * src_ci.  Two attempts are made to obtain both caps, and an error is return if
2125  * this fails; zero is returned on success.
2126  */
2127 static int get_rd_wr_caps(struct file *src_filp, int *src_got,
2128                           struct file *dst_filp,
2129                           loff_t dst_endoff, int *dst_got)
2130 {
2131         int ret = 0;
2132         bool retrying = false;
2133
2134 retry_caps:
2135         ret = ceph_get_caps(dst_filp, CEPH_CAP_FILE_WR, CEPH_CAP_FILE_BUFFER,
2136                             dst_endoff, dst_got);
2137         if (ret < 0)
2138                 return ret;
2139
2140         /*
2141          * Since we're already holding the FILE_WR capability for the dst file,
2142          * we would risk a deadlock by using ceph_get_caps.  Thus, we'll do some
2143          * retry dance instead to try to get both capabilities.
2144          */
2145         ret = ceph_try_get_caps(file_inode(src_filp),
2146                                 CEPH_CAP_FILE_RD, CEPH_CAP_FILE_SHARED,
2147                                 false, src_got);
2148         if (ret <= 0) {
2149                 /* Start by dropping dst_ci caps and getting src_ci caps */
2150                 ceph_put_cap_refs(ceph_inode(file_inode(dst_filp)), *dst_got);
2151                 if (retrying) {
2152                         if (!ret)
2153                                 /* ceph_try_get_caps masks EAGAIN */
2154                                 ret = -EAGAIN;
2155                         return ret;
2156                 }
2157                 ret = ceph_get_caps(src_filp, CEPH_CAP_FILE_RD,
2158                                     CEPH_CAP_FILE_SHARED, -1, src_got);
2159                 if (ret < 0)
2160                         return ret;
2161                 /*... drop src_ci caps too, and retry */
2162                 ceph_put_cap_refs(ceph_inode(file_inode(src_filp)), *src_got);
2163                 retrying = true;
2164                 goto retry_caps;
2165         }
2166         return ret;
2167 }
2168
2169 static void put_rd_wr_caps(struct ceph_inode_info *src_ci, int src_got,
2170                            struct ceph_inode_info *dst_ci, int dst_got)
2171 {
2172         ceph_put_cap_refs(src_ci, src_got);
2173         ceph_put_cap_refs(dst_ci, dst_got);
2174 }
2175
2176 /*
2177  * This function does several size-related checks, returning an error if:
2178  *  - source file is smaller than off+len
2179  *  - destination file size is not OK (inode_newsize_ok())
2180  *  - max bytes quotas is exceeded
2181  */
2182 static int is_file_size_ok(struct inode *src_inode, struct inode *dst_inode,
2183                            loff_t src_off, loff_t dst_off, size_t len)
2184 {
2185         loff_t size, endoff;
2186
2187         size = i_size_read(src_inode);
2188         /*
2189          * Don't copy beyond source file EOF.  Instead of simply setting length
2190          * to (size - src_off), just drop to VFS default implementation, as the
2191          * local i_size may be stale due to other clients writing to the source
2192          * inode.
2193          */
2194         if (src_off + len > size) {
2195                 dout("Copy beyond EOF (%llu + %zu > %llu)\n",
2196                      src_off, len, size);
2197                 return -EOPNOTSUPP;
2198         }
2199         size = i_size_read(dst_inode);
2200
2201         endoff = dst_off + len;
2202         if (inode_newsize_ok(dst_inode, endoff))
2203                 return -EOPNOTSUPP;
2204
2205         if (ceph_quota_is_max_bytes_exceeded(dst_inode, endoff))
2206                 return -EDQUOT;
2207
2208         return 0;
2209 }
2210
2211 static ssize_t ceph_do_objects_copy(struct ceph_inode_info *src_ci, u64 *src_off,
2212                                     struct ceph_inode_info *dst_ci, u64 *dst_off,
2213                                     struct ceph_fs_client *fsc,
2214                                     size_t len, unsigned int flags)
2215 {
2216         struct ceph_object_locator src_oloc, dst_oloc;
2217         struct ceph_object_id src_oid, dst_oid;
2218         size_t bytes = 0;
2219         u64 src_objnum, src_objoff, dst_objnum, dst_objoff;
2220         u32 src_objlen, dst_objlen;
2221         u32 object_size = src_ci->i_layout.object_size;
2222         int ret;
2223
2224         src_oloc.pool = src_ci->i_layout.pool_id;
2225         src_oloc.pool_ns = ceph_try_get_string(src_ci->i_layout.pool_ns);
2226         dst_oloc.pool = dst_ci->i_layout.pool_id;
2227         dst_oloc.pool_ns = ceph_try_get_string(dst_ci->i_layout.pool_ns);
2228
2229         while (len >= object_size) {
2230                 ceph_calc_file_object_mapping(&src_ci->i_layout, *src_off,
2231                                               object_size, &src_objnum,
2232                                               &src_objoff, &src_objlen);
2233                 ceph_calc_file_object_mapping(&dst_ci->i_layout, *dst_off,
2234                                               object_size, &dst_objnum,
2235                                               &dst_objoff, &dst_objlen);
2236                 ceph_oid_init(&src_oid);
2237                 ceph_oid_printf(&src_oid, "%llx.%08llx",
2238                                 src_ci->i_vino.ino, src_objnum);
2239                 ceph_oid_init(&dst_oid);
2240                 ceph_oid_printf(&dst_oid, "%llx.%08llx",
2241                                 dst_ci->i_vino.ino, dst_objnum);
2242                 /* Do an object remote copy */
2243                 ret = ceph_osdc_copy_from(&fsc->client->osdc,
2244                                           src_ci->i_vino.snap, 0,
2245                                           &src_oid, &src_oloc,
2246                                           CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2247                                           CEPH_OSD_OP_FLAG_FADVISE_NOCACHE,
2248                                           &dst_oid, &dst_oloc,
2249                                           CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2250                                           CEPH_OSD_OP_FLAG_FADVISE_DONTNEED,
2251                                           dst_ci->i_truncate_seq,
2252                                           dst_ci->i_truncate_size,
2253                                           CEPH_OSD_COPY_FROM_FLAG_TRUNCATE_SEQ);
2254                 if (ret) {
2255                         if (ret == -EOPNOTSUPP) {
2256                                 fsc->have_copy_from2 = false;
2257                                 pr_notice("OSDs don't support copy-from2; disabling copy offload\n");
2258                         }
2259                         dout("ceph_osdc_copy_from returned %d\n", ret);
2260                         if (!bytes)
2261                                 bytes = ret;
2262                         goto out;
2263                 }
2264                 len -= object_size;
2265                 bytes += object_size;
2266                 *src_off += object_size;
2267                 *dst_off += object_size;
2268         }
2269
2270 out:
2271         ceph_oloc_destroy(&src_oloc);
2272         ceph_oloc_destroy(&dst_oloc);
2273         return bytes;
2274 }
2275
2276 static ssize_t __ceph_copy_file_range(struct file *src_file, loff_t src_off,
2277                                       struct file *dst_file, loff_t dst_off,
2278                                       size_t len, unsigned int flags)
2279 {
2280         struct inode *src_inode = file_inode(src_file);
2281         struct inode *dst_inode = file_inode(dst_file);
2282         struct ceph_inode_info *src_ci = ceph_inode(src_inode);
2283         struct ceph_inode_info *dst_ci = ceph_inode(dst_inode);
2284         struct ceph_cap_flush *prealloc_cf;
2285         struct ceph_fs_client *src_fsc = ceph_inode_to_client(src_inode);
2286         loff_t size;
2287         ssize_t ret = -EIO, bytes;
2288         u64 src_objnum, dst_objnum, src_objoff, dst_objoff;
2289         u32 src_objlen, dst_objlen;
2290         int src_got = 0, dst_got = 0, err, dirty;
2291
2292         if (src_inode->i_sb != dst_inode->i_sb) {
2293                 struct ceph_fs_client *dst_fsc = ceph_inode_to_client(dst_inode);
2294
2295                 if (ceph_fsid_compare(&src_fsc->client->fsid,
2296                                       &dst_fsc->client->fsid)) {
2297                         dout("Copying files across clusters: src: %pU dst: %pU\n",
2298                              &src_fsc->client->fsid, &dst_fsc->client->fsid);
2299                         return -EXDEV;
2300                 }
2301         }
2302         if (ceph_snap(dst_inode) != CEPH_NOSNAP)
2303                 return -EROFS;
2304
2305         /*
2306          * Some of the checks below will return -EOPNOTSUPP, which will force a
2307          * fallback to the default VFS copy_file_range implementation.  This is
2308          * desirable in several cases (for ex, the 'len' is smaller than the
2309          * size of the objects, or in cases where that would be more
2310          * efficient).
2311          */
2312
2313         if (ceph_test_mount_opt(src_fsc, NOCOPYFROM))
2314                 return -EOPNOTSUPP;
2315
2316         if (!src_fsc->have_copy_from2)
2317                 return -EOPNOTSUPP;
2318
2319         /*
2320          * Striped file layouts require that we copy partial objects, but the
2321          * OSD copy-from operation only supports full-object copies.  Limit
2322          * this to non-striped file layouts for now.
2323          */
2324         if ((src_ci->i_layout.stripe_unit != dst_ci->i_layout.stripe_unit) ||
2325             (src_ci->i_layout.stripe_count != 1) ||
2326             (dst_ci->i_layout.stripe_count != 1) ||
2327             (src_ci->i_layout.object_size != dst_ci->i_layout.object_size)) {
2328                 dout("Invalid src/dst files layout\n");
2329                 return -EOPNOTSUPP;
2330         }
2331
2332         if (len < src_ci->i_layout.object_size)
2333                 return -EOPNOTSUPP; /* no remote copy will be done */
2334
2335         prealloc_cf = ceph_alloc_cap_flush();
2336         if (!prealloc_cf)
2337                 return -ENOMEM;
2338
2339         /* Start by sync'ing the source and destination files */
2340         ret = file_write_and_wait_range(src_file, src_off, (src_off + len));
2341         if (ret < 0) {
2342                 dout("failed to write src file (%zd)\n", ret);
2343                 goto out;
2344         }
2345         ret = file_write_and_wait_range(dst_file, dst_off, (dst_off + len));
2346         if (ret < 0) {
2347                 dout("failed to write dst file (%zd)\n", ret);
2348                 goto out;
2349         }
2350
2351         /*
2352          * We need FILE_WR caps for dst_ci and FILE_RD for src_ci as other
2353          * clients may have dirty data in their caches.  And OSDs know nothing
2354          * about caps, so they can't safely do the remote object copies.
2355          */
2356         err = get_rd_wr_caps(src_file, &src_got,
2357                              dst_file, (dst_off + len), &dst_got);
2358         if (err < 0) {
2359                 dout("get_rd_wr_caps returned %d\n", err);
2360                 ret = -EOPNOTSUPP;
2361                 goto out;
2362         }
2363
2364         ret = is_file_size_ok(src_inode, dst_inode, src_off, dst_off, len);
2365         if (ret < 0)
2366                 goto out_caps;
2367
2368         /* Drop dst file cached pages */
2369         ret = invalidate_inode_pages2_range(dst_inode->i_mapping,
2370                                             dst_off >> PAGE_SHIFT,
2371                                             (dst_off + len) >> PAGE_SHIFT);
2372         if (ret < 0) {
2373                 dout("Failed to invalidate inode pages (%zd)\n", ret);
2374                 ret = 0; /* XXX */
2375         }
2376         ceph_calc_file_object_mapping(&src_ci->i_layout, src_off,
2377                                       src_ci->i_layout.object_size,
2378                                       &src_objnum, &src_objoff, &src_objlen);
2379         ceph_calc_file_object_mapping(&dst_ci->i_layout, dst_off,
2380                                       dst_ci->i_layout.object_size,
2381                                       &dst_objnum, &dst_objoff, &dst_objlen);
2382         /* object-level offsets need to the same */
2383         if (src_objoff != dst_objoff) {
2384                 ret = -EOPNOTSUPP;
2385                 goto out_caps;
2386         }
2387
2388         /*
2389          * Do a manual copy if the object offset isn't object aligned.
2390          * 'src_objlen' contains the bytes left until the end of the object,
2391          * starting at the src_off
2392          */
2393         if (src_objoff) {
2394                 dout("Initial partial copy of %u bytes\n", src_objlen);
2395
2396                 /*
2397                  * we need to temporarily drop all caps as we'll be calling
2398                  * {read,write}_iter, which will get caps again.
2399                  */
2400                 put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2401                 ret = do_splice_direct(src_file, &src_off, dst_file,
2402                                        &dst_off, src_objlen, flags);
2403                 /* Abort on short copies or on error */
2404                 if (ret < src_objlen) {
2405                         dout("Failed partial copy (%zd)\n", ret);
2406                         goto out;
2407                 }
2408                 len -= ret;
2409                 err = get_rd_wr_caps(src_file, &src_got,
2410                                      dst_file, (dst_off + len), &dst_got);
2411                 if (err < 0)
2412                         goto out;
2413                 err = is_file_size_ok(src_inode, dst_inode,
2414                                       src_off, dst_off, len);
2415                 if (err < 0)
2416                         goto out_caps;
2417         }
2418
2419         size = i_size_read(dst_inode);
2420         bytes = ceph_do_objects_copy(src_ci, &src_off, dst_ci, &dst_off,
2421                                      src_fsc, len, flags);
2422         if (bytes <= 0) {
2423                 if (!ret)
2424                         ret = bytes;
2425                 goto out_caps;
2426         }
2427         dout("Copied %zu bytes out of %zu\n", bytes, len);
2428         len -= bytes;
2429         ret += bytes;
2430
2431         file_update_time(dst_file);
2432         inode_inc_iversion_raw(dst_inode);
2433
2434         if (dst_off > size) {
2435                 /* Let the MDS know about dst file size change */
2436                 if (ceph_inode_set_size(dst_inode, dst_off) ||
2437                     ceph_quota_is_max_bytes_approaching(dst_inode, dst_off))
2438                         ceph_check_caps(dst_ci, CHECK_CAPS_AUTHONLY, NULL);
2439         }
2440         /* Mark Fw dirty */
2441         spin_lock(&dst_ci->i_ceph_lock);
2442         dst_ci->i_inline_version = CEPH_INLINE_NONE;
2443         dirty = __ceph_mark_dirty_caps(dst_ci, CEPH_CAP_FILE_WR, &prealloc_cf);
2444         spin_unlock(&dst_ci->i_ceph_lock);
2445         if (dirty)
2446                 __mark_inode_dirty(dst_inode, dirty);
2447
2448 out_caps:
2449         put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2450
2451         /*
2452          * Do the final manual copy if we still have some bytes left, unless
2453          * there were errors in remote object copies (len >= object_size).
2454          */
2455         if (len && (len < src_ci->i_layout.object_size)) {
2456                 dout("Final partial copy of %zu bytes\n", len);
2457                 bytes = do_splice_direct(src_file, &src_off, dst_file,
2458                                          &dst_off, len, flags);
2459                 if (bytes > 0)
2460                         ret += bytes;
2461                 else
2462                         dout("Failed partial copy (%zd)\n", bytes);
2463         }
2464
2465 out:
2466         ceph_free_cap_flush(prealloc_cf);
2467
2468         return ret;
2469 }
2470
2471 static ssize_t ceph_copy_file_range(struct file *src_file, loff_t src_off,
2472                                     struct file *dst_file, loff_t dst_off,
2473                                     size_t len, unsigned int flags)
2474 {
2475         ssize_t ret;
2476
2477         ret = __ceph_copy_file_range(src_file, src_off, dst_file, dst_off,
2478                                      len, flags);
2479
2480         if (ret == -EOPNOTSUPP || ret == -EXDEV)
2481                 ret = generic_copy_file_range(src_file, src_off, dst_file,
2482                                               dst_off, len, flags);
2483         return ret;
2484 }
2485
2486 const struct file_operations ceph_file_fops = {
2487         .open = ceph_open,
2488         .release = ceph_release,
2489         .llseek = ceph_llseek,
2490         .read_iter = ceph_read_iter,
2491         .write_iter = ceph_write_iter,
2492         .mmap = ceph_mmap,
2493         .fsync = ceph_fsync,
2494         .lock = ceph_lock,
2495         .setlease = simple_nosetlease,
2496         .flock = ceph_flock,
2497         .splice_read = generic_file_splice_read,
2498         .splice_write = iter_file_splice_write,
2499         .unlocked_ioctl = ceph_ioctl,
2500         .compat_ioctl = compat_ptr_ioctl,
2501         .fallocate      = ceph_fallocate,
2502         .copy_file_range = ceph_copy_file_range,
2503 };