fs/ceph/file.c

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