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