fs/ceph/file.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/ceph/ceph_debug.h>
   3 #include <linux/ceph/striper.h>
   4
   5 #include <linux/module.h>
   6 #include <linux/sched.h>
   7 #include <linux/slab.h>
   8 #include <linux/file.h>
   9 #include <linux/mount.h>
  10 #include <linux/namei.h>
  11 #include <linux/writeback.h>
  12 #include <linux/falloc.h>
  13 #include <linux/iversion.h>
  14 #include <linux/ktime.h>
  15
  16 #include "super.h"
  17 #include "mds_client.h"
  18 #include "cache.h"
  19 #include "io.h"
  20 #include "metric.h"
  21
  22 static __le32 ceph_flags_sys2wire(u32 flags)
  23 {
  24         u32 wire_flags = 0;
  25
  26         switch (flags & O_ACCMODE) {
  27         case O_RDONLY:
  28                 wire_flags |= CEPH_O_RDONLY;
  29                 break;
  30         case O_WRONLY:
  31                 wire_flags |= CEPH_O_WRONLY;
  32                 break;
  33         case O_RDWR:
  34                 wire_flags |= CEPH_O_RDWR;
  35                 break;
  36         }
  37
  38         flags &= ~O_ACCMODE;
  39
  40 #define ceph_sys2wire(a) if (flags & a) { wire_flags |= CEPH_##a; flags &= ~a; }
  41
  42         ceph_sys2wire(O_CREAT);
  43         ceph_sys2wire(O_EXCL);
  44         ceph_sys2wire(O_TRUNC);
  45         ceph_sys2wire(O_DIRECTORY);
  46         ceph_sys2wire(O_NOFOLLOW);
  47
  48 #undef ceph_sys2wire
  49
  50         if (flags)
  51                 dout("unused open flags: %x\n", flags);
  52
  53         return cpu_to_le32(wire_flags);
  54 }
  55
  56 /*
  57  * Ceph file operations
  58  *
  59  * Implement basic open/close functionality, and implement
  60  * read/write.
  61  *
  62  * We implement three modes of file I/O:
  63  *  - buffered uses the generic_file_aio_{read,write} helpers
  64  *
  65  *  - synchronous is used when there is multi-client read/write
  66  *    sharing, avoids the page cache, and synchronously waits for an
  67  *    ack from the OSD.
  68  *
  69  *  - direct io takes the variant of the sync path that references
  70  *    user pages directly.
  71  *
  72  * fsync() flushes and waits on dirty pages, but just queues metadata
  73  * for writeback: since the MDS can recover size and mtime there is no
  74  * need to wait for MDS acknowledgement.
  75  */
  76
  77 /*
  78  * How many pages to get in one call to iov_iter_get_pages().  This
  79  * determines the size of the on-stack array used as a buffer.
  80  */
  81 #define ITER_GET_BVECS_PAGES    64
  82
  83 static ssize_t __iter_get_bvecs(struct iov_iter *iter, size_t maxsize,
  84                                 struct bio_vec *bvecs)
  85 {
  86         size_t size = 0;
  87         int bvec_idx = 0;
  88
  89         if (maxsize > iov_iter_count(iter))
  90                 maxsize = iov_iter_count(iter);
  91
  92         while (size < maxsize) {
  93                 struct page *pages[ITER_GET_BVECS_PAGES];
  94                 ssize_t bytes;
  95                 size_t start;
  96                 int idx = 0;
  97
  98                 bytes = iov_iter_get_pages2(iter, pages, maxsize - size,
  99                                            ITER_GET_BVECS_PAGES, &start);
 100                 if (bytes < 0)
 101                         return size ?: bytes;
 102
 103                 size += bytes;
 104
 105                 for ( ; bytes; idx++, bvec_idx++) {
 106                         int len = min_t(int, bytes, PAGE_SIZE - start);
 107
 108                         bvec_set_page(&bvecs[bvec_idx], pages[idx], len, start);
 109                         bytes -= len;
 110                         start = 0;
 111                 }
 112         }
 113
 114         return size;
 115 }
 116
 117 /*
 118  * iov_iter_get_pages() only considers one iov_iter segment, no matter
 119  * what maxsize or maxpages are given.  For ITER_BVEC that is a single
 120  * page.
 121  *
 122  * Attempt to get up to @maxsize bytes worth of pages from @iter.
 123  * Return the number of bytes in the created bio_vec array, or an error.
 124  */
 125 static ssize_t iter_get_bvecs_alloc(struct iov_iter *iter, size_t maxsize,
 126                                     struct bio_vec **bvecs, int *num_bvecs)
 127 {
 128         struct bio_vec *bv;
 129         size_t orig_count = iov_iter_count(iter);
 130         ssize_t bytes;
 131         int npages;
 132
 133         iov_iter_truncate(iter, maxsize);
 134         npages = iov_iter_npages(iter, INT_MAX);
 135         iov_iter_reexpand(iter, orig_count);
 136
 137         /*
 138          * __iter_get_bvecs() may populate only part of the array -- zero it
 139          * out.
 140          */
 141         bv = kvmalloc_array(npages, sizeof(*bv), GFP_KERNEL | __GFP_ZERO);
 142         if (!bv)
 143                 return -ENOMEM;
 144
 145         bytes = __iter_get_bvecs(iter, maxsize, bv);
 146         if (bytes < 0) {
 147                 /*
 148                  * No pages were pinned -- just free the array.
 149                  */
 150                 kvfree(bv);
 151                 return bytes;
 152         }
 153
 154         *bvecs = bv;
 155         *num_bvecs = npages;
 156         return bytes;
 157 }
 158
 159 static void put_bvecs(struct bio_vec *bvecs, int num_bvecs, bool should_dirty)
 160 {
 161         int i;
 162
 163         for (i = 0; i < num_bvecs; i++) {
 164                 if (bvecs[i].bv_page) {
 165                         if (should_dirty)
 166                                 set_page_dirty_lock(bvecs[i].bv_page);
 167                         put_page(bvecs[i].bv_page);
 168                 }
 169         }
 170         kvfree(bvecs);
 171 }
 172
 173 /*
 174  * Prepare an open request.  Preallocate ceph_cap to avoid an
 175  * inopportune ENOMEM later.
 176  */
 177 static struct ceph_mds_request *
 178 prepare_open_request(struct super_block *sb, int flags, int create_mode)
 179 {
 180         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(sb);
 181         struct ceph_mds_request *req;
 182         int want_auth = USE_ANY_MDS;
 183         int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
 184
 185         if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
 186                 want_auth = USE_AUTH_MDS;
 187
 188         req = ceph_mdsc_create_request(mdsc, op, want_auth);
 189         if (IS_ERR(req))
 190                 goto out;
 191         req->r_fmode = ceph_flags_to_mode(flags);
 192         req->r_args.open.flags = ceph_flags_sys2wire(flags);
 193         req->r_args.open.mode = cpu_to_le32(create_mode);
 194 out:
 195         return req;
 196 }
 197
 198 static int ceph_init_file_info(struct inode *inode, struct file *file,
 199                                         int fmode, bool isdir)
 200 {
 201         struct ceph_inode_info *ci = ceph_inode(inode);
 202         struct ceph_mount_options *opt =
 203                 ceph_inode_to_client(&ci->netfs.inode)->mount_options;
 204         struct ceph_file_info *fi;
 205         int ret;
 206
 207         dout("%s %p %p 0%o (%s)\n", __func__, inode, file,
 208                         inode->i_mode, isdir ? "dir" : "regular");
 209         BUG_ON(inode->i_fop->release != ceph_release);
 210
 211         if (isdir) {
 212                 struct ceph_dir_file_info *dfi =
 213                         kmem_cache_zalloc(ceph_dir_file_cachep, GFP_KERNEL);
 214                 if (!dfi)
 215                         return -ENOMEM;
 216
 217                 file->private_data = dfi;
 218                 fi = &dfi->file_info;
 219                 dfi->next_offset = 2;
 220                 dfi->readdir_cache_idx = -1;
 221         } else {
 222                 fi = kmem_cache_zalloc(ceph_file_cachep, GFP_KERNEL);
 223                 if (!fi)
 224                         return -ENOMEM;
 225
 226                 if (opt->flags & CEPH_MOUNT_OPT_NOPAGECACHE)
 227                         fi->flags |= CEPH_F_SYNC;
 228
 229                 file->private_data = fi;
 230         }
 231
 232         ceph_get_fmode(ci, fmode, 1);
 233         fi->fmode = fmode;
 234
 235         spin_lock_init(&fi->rw_contexts_lock);
 236         INIT_LIST_HEAD(&fi->rw_contexts);
 237         fi->filp_gen = READ_ONCE(ceph_inode_to_client(inode)->filp_gen);
 238
 239         if ((file->f_mode & FMODE_WRITE) && ceph_has_inline_data(ci)) {
 240                 ret = ceph_uninline_data(file);
 241                 if (ret < 0)
 242                         goto error;
 243         }
 244
 245         return 0;
 246
 247 error:
 248         ceph_fscache_unuse_cookie(inode, file->f_mode & FMODE_WRITE);
 249         ceph_put_fmode(ci, fi->fmode, 1);
 250         kmem_cache_free(ceph_file_cachep, fi);
 251         /* wake up anyone waiting for caps on this inode */
 252         wake_up_all(&ci->i_cap_wq);
 253         return ret;
 254 }
 255
 256 /*
 257  * initialize private struct file data.
 258  * if we fail, clean up by dropping fmode reference on the ceph_inode
 259  */
 260 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
 261 {
 262         int ret = 0;
 263
 264         switch (inode->i_mode & S_IFMT) {
 265         case S_IFREG:
 266                 ceph_fscache_use_cookie(inode, file->f_mode & FMODE_WRITE);
 267                 fallthrough;
 268         case S_IFDIR:
 269                 ret = ceph_init_file_info(inode, file, fmode,
 270                                                 S_ISDIR(inode->i_mode));
 271                 break;
 272
 273         case S_IFLNK:
 274                 dout("init_file %p %p 0%o (symlink)\n", inode, file,
 275                      inode->i_mode);
 276                 break;
 277
 278         default:
 279                 dout("init_file %p %p 0%o (special)\n", inode, file,
 280                      inode->i_mode);
 281                 /*
 282                  * we need to drop the open ref now, since we don't
 283                  * have .release set to ceph_release.
 284                  */
 285                 BUG_ON(inode->i_fop->release == ceph_release);
 286
 287                 /* call the proper open fop */
 288                 ret = inode->i_fop->open(inode, file);
 289         }
 290         return ret;
 291 }
 292
 293 /*
 294  * try renew caps after session gets killed.
 295  */
 296 int ceph_renew_caps(struct inode *inode, int fmode)
 297 {
 298         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
 299         struct ceph_inode_info *ci = ceph_inode(inode);
 300         struct ceph_mds_request *req;
 301         int err, flags, wanted;
 302
 303         spin_lock(&ci->i_ceph_lock);
 304         __ceph_touch_fmode(ci, mdsc, fmode);
 305         wanted = __ceph_caps_file_wanted(ci);
 306         if (__ceph_is_any_real_caps(ci) &&
 307             (!(wanted & CEPH_CAP_ANY_WR) || ci->i_auth_cap)) {
 308                 int issued = __ceph_caps_issued(ci, NULL);
 309                 spin_unlock(&ci->i_ceph_lock);
 310                 dout("renew caps %p want %s issued %s updating mds_wanted\n",
 311                      inode, ceph_cap_string(wanted), ceph_cap_string(issued));
 312                 ceph_check_caps(ci, 0);
 313                 return 0;
 314         }
 315         spin_unlock(&ci->i_ceph_lock);
 316
 317         flags = 0;
 318         if ((wanted & CEPH_CAP_FILE_RD) && (wanted & CEPH_CAP_FILE_WR))
 319                 flags = O_RDWR;
 320         else if (wanted & CEPH_CAP_FILE_RD)
 321                 flags = O_RDONLY;
 322         else if (wanted & CEPH_CAP_FILE_WR)
 323                 flags = O_WRONLY;
 324 #ifdef O_LAZY
 325         if (wanted & CEPH_CAP_FILE_LAZYIO)
 326                 flags |= O_LAZY;
 327 #endif
 328
 329         req = prepare_open_request(inode->i_sb, flags, 0);
 330         if (IS_ERR(req)) {
 331                 err = PTR_ERR(req);
 332                 goto out;
 333         }
 334
 335         req->r_inode = inode;
 336         ihold(inode);
 337         req->r_num_caps = 1;
 338
 339         err = ceph_mdsc_do_request(mdsc, NULL, req);
 340         ceph_mdsc_put_request(req);
 341 out:
 342         dout("renew caps %p open result=%d\n", inode, err);
 343         return err < 0 ? err : 0;
 344 }
 345
 346 /*
 347  * If we already have the requisite capabilities, we can satisfy
 348  * the open request locally (no need to request new caps from the
 349  * MDS).  We do, however, need to inform the MDS (asynchronously)
 350  * if our wanted caps set expands.
 351  */
 352 int ceph_open(struct inode *inode, struct file *file)
 353 {
 354         struct ceph_inode_info *ci = ceph_inode(inode);
 355         struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
 356         struct ceph_mds_client *mdsc = fsc->mdsc;
 357         struct ceph_mds_request *req;
 358         struct ceph_file_info *fi = file->private_data;
 359         int err;
 360         int flags, fmode, wanted;
 361
 362         if (fi) {
 363                 dout("open file %p is already opened\n", file);
 364                 return 0;
 365         }
 366
 367         /* filter out O_CREAT|O_EXCL; vfs did that already.  yuck. */
 368         flags = file->f_flags & ~(O_CREAT|O_EXCL);
 369         if (S_ISDIR(inode->i_mode)) {
 370                 flags = O_DIRECTORY;  /* mds likes to know */
 371         } else if (S_ISREG(inode->i_mode)) {
 372                 err = fscrypt_file_open(inode, file);
 373                 if (err)
 374                         return err;
 375         }
 376
 377         dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
 378              ceph_vinop(inode), file, flags, file->f_flags);
 379         fmode = ceph_flags_to_mode(flags);
 380         wanted = ceph_caps_for_mode(fmode);
 381
 382         /* snapped files are read-only */
 383         if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
 384                 return -EROFS;
 385
 386         /* trivially open snapdir */
 387         if (ceph_snap(inode) == CEPH_SNAPDIR) {
 388                 return ceph_init_file(inode, file, fmode);
 389         }
 390
 391         /*
 392          * No need to block if we have caps on the auth MDS (for
 393          * write) or any MDS (for read).  Update wanted set
 394          * asynchronously.
 395          */
 396         spin_lock(&ci->i_ceph_lock);
 397         if (__ceph_is_any_real_caps(ci) &&
 398             (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
 399                 int mds_wanted = __ceph_caps_mds_wanted(ci, true);
 400                 int issued = __ceph_caps_issued(ci, NULL);
 401
 402                 dout("open %p fmode %d want %s issued %s using existing\n",
 403                      inode, fmode, ceph_cap_string(wanted),
 404                      ceph_cap_string(issued));
 405                 __ceph_touch_fmode(ci, mdsc, fmode);
 406                 spin_unlock(&ci->i_ceph_lock);
 407
 408                 /* adjust wanted? */
 409                 if ((issued & wanted) != wanted &&
 410                     (mds_wanted & wanted) != wanted &&
 411                     ceph_snap(inode) != CEPH_SNAPDIR)
 412                         ceph_check_caps(ci, 0);
 413
 414                 return ceph_init_file(inode, file, fmode);
 415         } else if (ceph_snap(inode) != CEPH_NOSNAP &&
 416                    (ci->i_snap_caps & wanted) == wanted) {
 417                 __ceph_touch_fmode(ci, mdsc, fmode);
 418                 spin_unlock(&ci->i_ceph_lock);
 419                 return ceph_init_file(inode, file, fmode);
 420         }
 421
 422         spin_unlock(&ci->i_ceph_lock);
 423
 424         dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
 425         req = prepare_open_request(inode->i_sb, flags, 0);
 426         if (IS_ERR(req)) {
 427                 err = PTR_ERR(req);
 428                 goto out;
 429         }
 430         req->r_inode = inode;
 431         ihold(inode);
 432
 433         req->r_num_caps = 1;
 434         err = ceph_mdsc_do_request(mdsc, NULL, req);
 435         if (!err)
 436                 err = ceph_init_file(inode, file, req->r_fmode);
 437         ceph_mdsc_put_request(req);
 438         dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
 439 out:
 440         return err;
 441 }
 442
 443 /* Clone the layout from a synchronous create, if the dir now has Dc caps */
 444 static void
 445 cache_file_layout(struct inode *dst, struct inode *src)
 446 {
 447         struct ceph_inode_info *cdst = ceph_inode(dst);
 448         struct ceph_inode_info *csrc = ceph_inode(src);
 449
 450         spin_lock(&cdst->i_ceph_lock);
 451         if ((__ceph_caps_issued(cdst, NULL) & CEPH_CAP_DIR_CREATE) &&
 452             !ceph_file_layout_is_valid(&cdst->i_cached_layout)) {
 453                 memcpy(&cdst->i_cached_layout, &csrc->i_layout,
 454                         sizeof(cdst->i_cached_layout));
 455                 rcu_assign_pointer(cdst->i_cached_layout.pool_ns,
 456                                    ceph_try_get_string(csrc->i_layout.pool_ns));
 457         }
 458         spin_unlock(&cdst->i_ceph_lock);
 459 }
 460
 461 /*
 462  * Try to set up an async create. We need caps, a file layout, and inode number,
 463  * and either a lease on the dentry or complete dir info. If any of those
 464  * criteria are not satisfied, then return false and the caller can go
 465  * synchronous.
 466  */
 467 static int try_prep_async_create(struct inode *dir, struct dentry *dentry,
 468                                  struct ceph_file_layout *lo, u64 *pino)
 469 {
 470         struct ceph_inode_info *ci = ceph_inode(dir);
 471         struct ceph_dentry_info *di = ceph_dentry(dentry);
 472         int got = 0, want = CEPH_CAP_FILE_EXCL | CEPH_CAP_DIR_CREATE;
 473         u64 ino;
 474
 475         spin_lock(&ci->i_ceph_lock);
 476         /* No auth cap means no chance for Dc caps */
 477         if (!ci->i_auth_cap)
 478                 goto no_async;
 479
 480         /* Any delegated inos? */
 481         if (xa_empty(&ci->i_auth_cap->session->s_delegated_inos))
 482                 goto no_async;
 483
 484         if (!ceph_file_layout_is_valid(&ci->i_cached_layout))
 485                 goto no_async;
 486
 487         if ((__ceph_caps_issued(ci, NULL) & want) != want)
 488                 goto no_async;
 489
 490         if (d_in_lookup(dentry)) {
 491                 if (!__ceph_dir_is_complete(ci))
 492                         goto no_async;
 493                 spin_lock(&dentry->d_lock);
 494                 di->lease_shared_gen = atomic_read(&ci->i_shared_gen);
 495                 spin_unlock(&dentry->d_lock);
 496         } else if (atomic_read(&ci->i_shared_gen) !=
 497                    READ_ONCE(di->lease_shared_gen)) {
 498                 goto no_async;
 499         }
 500
 501         ino = ceph_get_deleg_ino(ci->i_auth_cap->session);
 502         if (!ino)
 503                 goto no_async;
 504
 505         *pino = ino;
 506         ceph_take_cap_refs(ci, want, false);
 507         memcpy(lo, &ci->i_cached_layout, sizeof(*lo));
 508         rcu_assign_pointer(lo->pool_ns,
 509                            ceph_try_get_string(ci->i_cached_layout.pool_ns));
 510         got = want;
 511 no_async:
 512         spin_unlock(&ci->i_ceph_lock);
 513         return got;
 514 }
 515
 516 static void restore_deleg_ino(struct inode *dir, u64 ino)
 517 {
 518         struct ceph_inode_info *ci = ceph_inode(dir);
 519         struct ceph_mds_session *s = NULL;
 520
 521         spin_lock(&ci->i_ceph_lock);
 522         if (ci->i_auth_cap)
 523                 s = ceph_get_mds_session(ci->i_auth_cap->session);
 524         spin_unlock(&ci->i_ceph_lock);
 525         if (s) {
 526                 int err = ceph_restore_deleg_ino(s, ino);
 527                 if (err)
 528                         pr_warn("ceph: unable to restore delegated ino 0x%llx to session: %d\n",
 529                                 ino, err);
 530                 ceph_put_mds_session(s);
 531         }
 532 }
 533
 534 static void wake_async_create_waiters(struct inode *inode,
 535                                       struct ceph_mds_session *session)
 536 {
 537         struct ceph_inode_info *ci = ceph_inode(inode);
 538         bool check_cap = false;
 539
 540         spin_lock(&ci->i_ceph_lock);
 541         if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE) {
 542                 ci->i_ceph_flags &= ~CEPH_I_ASYNC_CREATE;
 543                 wake_up_bit(&ci->i_ceph_flags, CEPH_ASYNC_CREATE_BIT);
 544
 545                 if (ci->i_ceph_flags & CEPH_I_ASYNC_CHECK_CAPS) {
 546                         ci->i_ceph_flags &= ~CEPH_I_ASYNC_CHECK_CAPS;
 547                         check_cap = true;
 548                 }
 549         }
 550         ceph_kick_flushing_inode_caps(session, ci);
 551         spin_unlock(&ci->i_ceph_lock);
 552
 553         if (check_cap)
 554                 ceph_check_caps(ci, CHECK_CAPS_FLUSH);
 555 }
 556
 557 static void ceph_async_create_cb(struct ceph_mds_client *mdsc,
 558                                  struct ceph_mds_request *req)
 559 {
 560         struct dentry *dentry = req->r_dentry;
 561         struct inode *dinode = d_inode(dentry);
 562         struct inode *tinode = req->r_target_inode;
 563         int result = req->r_err ? req->r_err :
 564                         le32_to_cpu(req->r_reply_info.head->result);
 565
 566         WARN_ON_ONCE(dinode && tinode && dinode != tinode);
 567
 568         /* MDS changed -- caller must resubmit */
 569         if (result == -EJUKEBOX)
 570                 goto out;
 571
 572         mapping_set_error(req->r_parent->i_mapping, result);
 573
 574         if (result) {
 575                 int pathlen = 0;
 576                 u64 base = 0;
 577                 char *path = ceph_mdsc_build_path(req->r_dentry, &pathlen,
 578                                                   &base, 0);
 579
 580                 pr_warn("async create failure path=(%llx)%s result=%d!\n",
 581                         base, IS_ERR(path) ? "<<bad>>" : path, result);
 582                 ceph_mdsc_free_path(path, pathlen);
 583
 584                 ceph_dir_clear_complete(req->r_parent);
 585                 if (!d_unhashed(dentry))
 586                         d_drop(dentry);
 587
 588                 if (dinode) {
 589                         mapping_set_error(dinode->i_mapping, result);
 590                         ceph_inode_shutdown(dinode);
 591                         wake_async_create_waiters(dinode, req->r_session);
 592                 }
 593         }
 594
 595         if (tinode) {
 596                 u64 ino = ceph_vino(tinode).ino;
 597
 598                 if (req->r_deleg_ino != ino)
 599                         pr_warn("%s: inode number mismatch! err=%d deleg_ino=0x%llx target=0x%llx\n",
 600                                 __func__, req->r_err, req->r_deleg_ino, ino);
 601
 602                 mapping_set_error(tinode->i_mapping, result);
 603                 wake_async_create_waiters(tinode, req->r_session);
 604         } else if (!result) {
 605                 pr_warn("%s: no req->r_target_inode for 0x%llx\n", __func__,
 606                         req->r_deleg_ino);
 607         }
 608 out:
 609         ceph_mdsc_release_dir_caps(req);
 610 }
 611
 612 static int ceph_finish_async_create(struct inode *dir, struct inode *inode,
 613                                     struct dentry *dentry,
 614                                     struct file *file, umode_t mode,
 615                                     struct ceph_mds_request *req,
 616                                     struct ceph_acl_sec_ctx *as_ctx,
 617                                     struct ceph_file_layout *lo)
 618 {
 619         int ret;
 620         char xattr_buf[4];
 621         struct ceph_mds_reply_inode in = { };
 622         struct ceph_mds_reply_info_in iinfo = { .in = &in };
 623         struct ceph_inode_info *ci = ceph_inode(dir);
 624         struct ceph_dentry_info *di = ceph_dentry(dentry);
 625         struct timespec64 now;
 626         struct ceph_string *pool_ns;
 627         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 628         struct ceph_vino vino = { .ino = req->r_deleg_ino,
 629                                   .snap = CEPH_NOSNAP };
 630
 631         ktime_get_real_ts64(&now);
 632
 633         iinfo.inline_version = CEPH_INLINE_NONE;
 634         iinfo.change_attr = 1;
 635         ceph_encode_timespec64(&iinfo.btime, &now);
 636
 637         if (req->r_pagelist) {
 638                 iinfo.xattr_len = req->r_pagelist->length;
 639                 iinfo.xattr_data = req->r_pagelist->mapped_tail;
 640         } else {
 641                 /* fake it */
 642                 iinfo.xattr_len = ARRAY_SIZE(xattr_buf);
 643                 iinfo.xattr_data = xattr_buf;
 644                 memset(iinfo.xattr_data, 0, iinfo.xattr_len);
 645         }
 646
 647         in.ino = cpu_to_le64(vino.ino);
 648         in.snapid = cpu_to_le64(CEPH_NOSNAP);
 649         in.version = cpu_to_le64(1);    // ???
 650         in.cap.caps = in.cap.wanted = cpu_to_le32(CEPH_CAP_ALL_FILE);
 651         in.cap.cap_id = cpu_to_le64(1);
 652         in.cap.realm = cpu_to_le64(ci->i_snap_realm->ino);
 653         in.cap.flags = CEPH_CAP_FLAG_AUTH;
 654         in.ctime = in.mtime = in.atime = iinfo.btime;
 655         in.truncate_seq = cpu_to_le32(1);
 656         in.truncate_size = cpu_to_le64(-1ULL);
 657         in.xattr_version = cpu_to_le64(1);
 658         in.uid = cpu_to_le32(from_kuid(&init_user_ns, current_fsuid()));
 659         if (dir->i_mode & S_ISGID) {
 660                 in.gid = cpu_to_le32(from_kgid(&init_user_ns, dir->i_gid));
 661
 662                 /* Directories always inherit the setgid bit. */
 663                 if (S_ISDIR(mode))
 664                         mode |= S_ISGID;
 665         } else {
 666                 in.gid = cpu_to_le32(from_kgid(&init_user_ns, current_fsgid()));
 667         }
 668         in.mode = cpu_to_le32((u32)mode);
 669
 670         in.nlink = cpu_to_le32(1);
 671         in.max_size = cpu_to_le64(lo->stripe_unit);
 672
 673         ceph_file_layout_to_legacy(lo, &in.layout);
 674         /* lo is private, so pool_ns can't change */
 675         pool_ns = rcu_dereference_raw(lo->pool_ns);
 676         if (pool_ns) {
 677                 iinfo.pool_ns_len = pool_ns->len;
 678                 iinfo.pool_ns_data = pool_ns->str;
 679         }
 680
 681         down_read(&mdsc->snap_rwsem);
 682         ret = ceph_fill_inode(inode, NULL, &iinfo, NULL, req->r_session,
 683                               req->r_fmode, NULL);
 684         up_read(&mdsc->snap_rwsem);
 685         if (ret) {
 686                 dout("%s failed to fill inode: %d\n", __func__, ret);
 687                 ceph_dir_clear_complete(dir);
 688                 if (!d_unhashed(dentry))
 689                         d_drop(dentry);
 690                 discard_new_inode(inode);
 691         } else {
 692                 struct dentry *dn;
 693
 694                 dout("%s d_adding new inode 0x%llx to 0x%llx/%s\n", __func__,
 695                         vino.ino, ceph_ino(dir), dentry->d_name.name);
 696                 ceph_dir_clear_ordered(dir);
 697                 ceph_init_inode_acls(inode, as_ctx);
 698                 if (inode->i_state & I_NEW) {
 699                         /*
 700                          * If it's not I_NEW, then someone created this before
 701                          * we got here. Assume the server is aware of it at
 702                          * that point and don't worry about setting
 703                          * CEPH_I_ASYNC_CREATE.
 704                          */
 705                         ceph_inode(inode)->i_ceph_flags = CEPH_I_ASYNC_CREATE;
 706                         unlock_new_inode(inode);
 707                 }
 708                 if (d_in_lookup(dentry) || d_really_is_negative(dentry)) {
 709                         if (!d_unhashed(dentry))
 710                                 d_drop(dentry);
 711                         dn = d_splice_alias(inode, dentry);
 712                         WARN_ON_ONCE(dn && dn != dentry);
 713                 }
 714                 file->f_mode |= FMODE_CREATED;
 715                 ret = finish_open(file, dentry, ceph_open);
 716         }
 717
 718         spin_lock(&dentry->d_lock);
 719         di->flags &= ~CEPH_DENTRY_ASYNC_CREATE;
 720         wake_up_bit(&di->flags, CEPH_DENTRY_ASYNC_CREATE_BIT);
 721         spin_unlock(&dentry->d_lock);
 722
 723         return ret;
 724 }
 725
 726 /*
 727  * Do a lookup + open with a single request.  If we get a non-existent
 728  * file or symlink, return 1 so the VFS can retry.
 729  */
 730 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
 731                      struct file *file, unsigned flags, umode_t mode)
 732 {
 733         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 734         struct ceph_mds_client *mdsc = fsc->mdsc;
 735         struct ceph_mds_request *req;
 736         struct inode *new_inode = NULL;
 737         struct dentry *dn;
 738         struct ceph_acl_sec_ctx as_ctx = {};
 739         bool try_async = ceph_test_mount_opt(fsc, ASYNC_DIROPS);
 740         int mask;
 741         int err;
 742
 743         dout("atomic_open %p dentry %p '%pd' %s flags %d mode 0%o\n",
 744              dir, dentry, dentry,
 745              d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
 746
 747         if (dentry->d_name.len > NAME_MAX)
 748                 return -ENAMETOOLONG;
 749
 750         err = ceph_wait_on_conflict_unlink(dentry);
 751         if (err)
 752                 return err;
 753         /*
 754          * Do not truncate the file, since atomic_open is called before the
 755          * permission check. The caller will do the truncation afterward.
 756          */
 757         flags &= ~O_TRUNC;
 758
 759 retry:
 760         if (flags & O_CREAT) {
 761                 if (ceph_quota_is_max_files_exceeded(dir))
 762                         return -EDQUOT;
 763
 764                 new_inode = ceph_new_inode(dir, dentry, &mode, &as_ctx);
 765                 if (IS_ERR(new_inode)) {
 766                         err = PTR_ERR(new_inode);
 767                         goto out_ctx;
 768                 }
 769                 /* Async create can't handle more than a page of xattrs */
 770                 if (as_ctx.pagelist &&
 771                     !list_is_singular(&as_ctx.pagelist->head))
 772                         try_async = false;
 773         } else if (!d_in_lookup(dentry)) {
 774                 /* If it's not being looked up, it's negative */
 775                 return -ENOENT;
 776         }
 777
 778         /* do the open */
 779         req = prepare_open_request(dir->i_sb, flags, mode);
 780         if (IS_ERR(req)) {
 781                 err = PTR_ERR(req);
 782                 goto out_ctx;
 783         }
 784         req->r_dentry = dget(dentry);
 785         req->r_num_caps = 2;
 786         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 787         if (ceph_security_xattr_wanted(dir))
 788                 mask |= CEPH_CAP_XATTR_SHARED;
 789         req->r_args.open.mask = cpu_to_le32(mask);
 790         req->r_parent = dir;
 791         ihold(dir);
 792         if (IS_ENCRYPTED(dir)) {
 793                 if (!fscrypt_has_encryption_key(dir)) {
 794                         spin_lock(&dentry->d_lock);
 795                         dentry->d_flags |= DCACHE_NOKEY_NAME;
 796                         spin_unlock(&dentry->d_lock);
 797                 }
 798         }
 799
 800         if (flags & O_CREAT) {
 801                 struct ceph_file_layout lo;
 802
 803                 req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL |
 804                                      CEPH_CAP_XATTR_EXCL;
 805                 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 806
 807                 ceph_as_ctx_to_req(req, &as_ctx);
 808
 809                 if (try_async && (req->r_dir_caps =
 810                                   try_prep_async_create(dir, dentry, &lo,
 811                                                         &req->r_deleg_ino))) {
 812                         struct ceph_vino vino = { .ino = req->r_deleg_ino,
 813                                                   .snap = CEPH_NOSNAP };
 814                         struct ceph_dentry_info *di = ceph_dentry(dentry);
 815
 816                         set_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags);
 817                         req->r_args.open.flags |= cpu_to_le32(CEPH_O_EXCL);
 818                         req->r_callback = ceph_async_create_cb;
 819
 820                         /* Hash inode before RPC */
 821                         new_inode = ceph_get_inode(dir->i_sb, vino, new_inode);
 822                         if (IS_ERR(new_inode)) {
 823                                 err = PTR_ERR(new_inode);
 824                                 new_inode = NULL;
 825                                 goto out_req;
 826                         }
 827                         WARN_ON_ONCE(!(new_inode->i_state & I_NEW));
 828
 829                         spin_lock(&dentry->d_lock);
 830                         di->flags |= CEPH_DENTRY_ASYNC_CREATE;
 831                         spin_unlock(&dentry->d_lock);
 832
 833                         err = ceph_mdsc_submit_request(mdsc, dir, req);
 834                         if (!err) {
 835                                 err = ceph_finish_async_create(dir, new_inode,
 836                                                                dentry, file,
 837                                                                mode, req,
 838                                                                &as_ctx, &lo);
 839                                 new_inode = NULL;
 840                         } else if (err == -EJUKEBOX) {
 841                                 restore_deleg_ino(dir, req->r_deleg_ino);
 842                                 ceph_mdsc_put_request(req);
 843                                 discard_new_inode(new_inode);
 844                                 ceph_release_acl_sec_ctx(&as_ctx);
 845                                 memset(&as_ctx, 0, sizeof(as_ctx));
 846                                 new_inode = NULL;
 847                                 try_async = false;
 848                                 ceph_put_string(rcu_dereference_raw(lo.pool_ns));
 849                                 goto retry;
 850                         }
 851                         ceph_put_string(rcu_dereference_raw(lo.pool_ns));
 852                         goto out_req;
 853                 }
 854         }
 855
 856         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 857         req->r_new_inode = new_inode;
 858         new_inode = NULL;
 859         err = ceph_mdsc_do_request(mdsc, (flags & O_CREAT) ? dir : NULL, req);
 860         if (err == -ENOENT) {
 861                 dentry = ceph_handle_snapdir(req, dentry);
 862                 if (IS_ERR(dentry)) {
 863                         err = PTR_ERR(dentry);
 864                         goto out_req;
 865                 }
 866                 err = 0;
 867         }
 868
 869         if (!err && (flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
 870                 err = ceph_handle_notrace_create(dir, dentry);
 871
 872         if (d_in_lookup(dentry)) {
 873                 dn = ceph_finish_lookup(req, dentry, err);
 874                 if (IS_ERR(dn))
 875                         err = PTR_ERR(dn);
 876         } else {
 877                 /* we were given a hashed negative dentry */
 878                 dn = NULL;
 879         }
 880         if (err)
 881                 goto out_req;
 882         if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
 883                 /* make vfs retry on splice, ENOENT, or symlink */
 884                 dout("atomic_open finish_no_open on dn %p\n", dn);
 885                 err = finish_no_open(file, dn);
 886         } else {
 887                 if (IS_ENCRYPTED(dir) &&
 888                     !fscrypt_has_permitted_context(dir, d_inode(dentry))) {
 889                         pr_warn("Inconsistent encryption context (parent %llx:%llx child %llx:%llx)\n",
 890                                 ceph_vinop(dir), ceph_vinop(d_inode(dentry)));
 891                         goto out_req;
 892                 }
 893
 894                 dout("atomic_open finish_open on dn %p\n", dn);
 895                 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
 896                         struct inode *newino = d_inode(dentry);
 897
 898                         cache_file_layout(dir, newino);
 899                         ceph_init_inode_acls(newino, &as_ctx);
 900                         file->f_mode |= FMODE_CREATED;
 901                 }
 902                 err = finish_open(file, dentry, ceph_open);
 903         }
 904 out_req:
 905         ceph_mdsc_put_request(req);
 906         iput(new_inode);
 907 out_ctx:
 908         ceph_release_acl_sec_ctx(&as_ctx);
 909         dout("atomic_open result=%d\n", err);
 910         return err;
 911 }
 912
 913 int ceph_release(struct inode *inode, struct file *file)
 914 {
 915         struct ceph_inode_info *ci = ceph_inode(inode);
 916
 917         if (S_ISDIR(inode->i_mode)) {
 918                 struct ceph_dir_file_info *dfi = file->private_data;
 919                 dout("release inode %p dir file %p\n", inode, file);
 920                 WARN_ON(!list_empty(&dfi->file_info.rw_contexts));
 921
 922                 ceph_put_fmode(ci, dfi->file_info.fmode, 1);
 923
 924                 if (dfi->last_readdir)
 925                         ceph_mdsc_put_request(dfi->last_readdir);
 926                 kfree(dfi->last_name);
 927                 kfree(dfi->dir_info);
 928                 kmem_cache_free(ceph_dir_file_cachep, dfi);
 929         } else {
 930                 struct ceph_file_info *fi = file->private_data;
 931                 dout("release inode %p regular file %p\n", inode, file);
 932                 WARN_ON(!list_empty(&fi->rw_contexts));
 933
 934                 ceph_fscache_unuse_cookie(inode, file->f_mode & FMODE_WRITE);
 935                 ceph_put_fmode(ci, fi->fmode, 1);
 936
 937                 kmem_cache_free(ceph_file_cachep, fi);
 938         }
 939
 940         /* wake up anyone waiting for caps on this inode */
 941         wake_up_all(&ci->i_cap_wq);
 942         return 0;
 943 }
 944
 945 enum {
 946         HAVE_RETRIED = 1,
 947         CHECK_EOF =    2,
 948         READ_INLINE =  3,
 949 };
 950
 951 /*
 952  * Completely synchronous read and write methods.  Direct from __user
 953  * buffer to osd, or directly to user pages (if O_DIRECT).
 954  *
 955  * If the read spans object boundary, just do multiple reads.  (That's not
 956  * atomic, but good enough for now.)
 957  *
 958  * If we get a short result from the OSD, check against i_size; we need to
 959  * only return a short read to the caller if we hit EOF.
 960  */
 961 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *to,
 962                               int *retry_op)
 963 {
 964         struct file *file = iocb->ki_filp;
 965         struct inode *inode = file_inode(file);
 966         struct ceph_inode_info *ci = ceph_inode(inode);
 967         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 968         struct ceph_osd_client *osdc = &fsc->client->osdc;
 969         ssize_t ret;
 970         u64 off = iocb->ki_pos;
 971         u64 len = iov_iter_count(to);
 972         u64 i_size = i_size_read(inode);
 973         bool sparse = ceph_test_mount_opt(fsc, SPARSEREAD);
 974
 975         dout("sync_read on file %p %llu~%u %s\n", file, off, (unsigned)len,
 976              (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
 977
 978         if (!len)
 979                 return 0;
 980         /*
 981          * flush any page cache pages in this range.  this
 982          * will make concurrent normal and sync io slow,
 983          * but it will at least behave sensibly when they are
 984          * in sequence.
 985          */
 986         ret = filemap_write_and_wait_range(inode->i_mapping,
 987                                            off, off + len - 1);
 988         if (ret < 0)
 989                 return ret;
 990
 991         ret = 0;
 992         while ((len = iov_iter_count(to)) > 0) {
 993                 struct ceph_osd_request *req;
 994                 struct page **pages;
 995                 int num_pages;
 996                 size_t page_off;
 997                 bool more;
 998                 int idx;
 999                 size_t left;
1000                 struct ceph_osd_req_op *op;
1001
1002                 req = ceph_osdc_new_request(osdc, &ci->i_layout,
1003                                         ci->i_vino, off, &len, 0, 1,
1004                                         sparse ? CEPH_OSD_OP_SPARSE_READ : CEPH_OSD_OP_READ,
1005                                         CEPH_OSD_FLAG_READ,
1006                                         NULL, ci->i_truncate_seq,
1007                                         ci->i_truncate_size, false);
1008                 if (IS_ERR(req)) {
1009                         ret = PTR_ERR(req);
1010                         break;
1011                 }
1012
1013                 more = len < iov_iter_count(to);
1014
1015                 num_pages = calc_pages_for(off, len);
1016                 page_off = off & ~PAGE_MASK;
1017                 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1018                 if (IS_ERR(pages)) {
1019                         ceph_osdc_put_request(req);
1020                         ret = PTR_ERR(pages);
1021                         break;
1022                 }
1023
1024                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_off,
1025                                                  false, false);
1026
1027                 op = &req->r_ops[0];
1028                 if (sparse) {
1029                         ret = ceph_alloc_sparse_ext_map(op);
1030                         if (ret) {
1031                                 ceph_osdc_put_request(req);
1032                                 break;
1033                         }
1034                 }
1035
1036                 ceph_osdc_start_request(osdc, req);
1037                 ret = ceph_osdc_wait_request(osdc, req);
1038
1039                 ceph_update_read_metrics(&fsc->mdsc->metric,
1040                                          req->r_start_latency,
1041                                          req->r_end_latency,
1042                                          len, ret);
1043
1044                 i_size = i_size_read(inode);
1045                 dout("sync_read %llu~%llu got %zd i_size %llu%s\n",
1046                      off, len, ret, i_size, (more ? " MORE" : ""));
1047
1048                 /* Fix it to go to end of extent map */
1049                 if (sparse && ret >= 0)
1050                         ret = ceph_sparse_ext_map_end(op);
1051                 else if (ret == -ENOENT)
1052                         ret = 0;
1053
1054                 ceph_osdc_put_request(req);
1055
1056                 if (ret >= 0 && ret < len && (off + ret < i_size)) {
1057                         int zlen = min(len - ret, i_size - off - ret);
1058                         int zoff = page_off + ret;
1059
1060                         dout("sync_read zero gap %llu~%llu\n",
1061                                 off + ret, off + ret + zlen);
1062                         ceph_zero_page_vector_range(zoff, zlen, pages);
1063                         ret += zlen;
1064                 }
1065
1066                 idx = 0;
1067                 left = ret > 0 ? ret : 0;
1068                 while (left > 0) {
1069                         size_t len, copied;
1070                         page_off = off & ~PAGE_MASK;
1071                         len = min_t(size_t, left, PAGE_SIZE - page_off);
1072                         SetPageUptodate(pages[idx]);
1073                         copied = copy_page_to_iter(pages[idx++],
1074                                                    page_off, len, to);
1075                         off += copied;
1076                         left -= copied;
1077                         if (copied < len) {
1078                                 ret = -EFAULT;
1079                                 break;
1080                         }
1081                 }
1082                 ceph_release_page_vector(pages, num_pages);
1083
1084                 if (ret < 0) {
1085                         if (ret == -EBLOCKLISTED)
1086                                 fsc->blocklisted = true;
1087                         break;
1088                 }
1089
1090                 if (off >= i_size || !more)
1091                         break;
1092         }
1093
1094         if (off > iocb->ki_pos) {
1095                 if (off >= i_size) {
1096                         *retry_op = CHECK_EOF;
1097                         ret = i_size - iocb->ki_pos;
1098                         iocb->ki_pos = i_size;
1099                 } else {
1100                         ret = off - iocb->ki_pos;
1101                         iocb->ki_pos = off;
1102                 }
1103         }
1104
1105         dout("sync_read result %zd retry_op %d\n", ret, *retry_op);
1106         return ret;
1107 }
1108
1109 struct ceph_aio_request {
1110         struct kiocb *iocb;
1111         size_t total_len;
1112         bool write;
1113         bool should_dirty;
1114         int error;
1115         struct list_head osd_reqs;
1116         unsigned num_reqs;
1117         atomic_t pending_reqs;
1118         struct timespec64 mtime;
1119         struct ceph_cap_flush *prealloc_cf;
1120 };
1121
1122 struct ceph_aio_work {
1123         struct work_struct work;
1124         struct ceph_osd_request *req;
1125 };
1126
1127 static void ceph_aio_retry_work(struct work_struct *work);
1128
1129 static void ceph_aio_complete(struct inode *inode,
1130                               struct ceph_aio_request *aio_req)
1131 {
1132         struct ceph_inode_info *ci = ceph_inode(inode);
1133         int ret;
1134
1135         if (!atomic_dec_and_test(&aio_req->pending_reqs))
1136                 return;
1137
1138         if (aio_req->iocb->ki_flags & IOCB_DIRECT)
1139                 inode_dio_end(inode);
1140
1141         ret = aio_req->error;
1142         if (!ret)
1143                 ret = aio_req->total_len;
1144
1145         dout("ceph_aio_complete %p rc %d\n", inode, ret);
1146
1147         if (ret >= 0 && aio_req->write) {
1148                 int dirty;
1149
1150                 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
1151                 if (endoff > i_size_read(inode)) {
1152                         if (ceph_inode_set_size(inode, endoff))
1153                                 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY);
1154                 }
1155
1156                 spin_lock(&ci->i_ceph_lock);
1157                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1158                                                &aio_req->prealloc_cf);
1159                 spin_unlock(&ci->i_ceph_lock);
1160                 if (dirty)
1161                         __mark_inode_dirty(inode, dirty);
1162
1163         }
1164
1165         ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
1166                                                 CEPH_CAP_FILE_RD));
1167
1168         aio_req->iocb->ki_complete(aio_req->iocb, ret);
1169
1170         ceph_free_cap_flush(aio_req->prealloc_cf);
1171         kfree(aio_req);
1172 }
1173
1174 static void ceph_aio_complete_req(struct ceph_osd_request *req)
1175 {
1176         int rc = req->r_result;
1177         struct inode *inode = req->r_inode;
1178         struct ceph_aio_request *aio_req = req->r_priv;
1179         struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
1180         struct ceph_osd_req_op *op = &req->r_ops[0];
1181         struct ceph_client_metric *metric = &ceph_sb_to_mdsc(inode->i_sb)->metric;
1182         unsigned int len = osd_data->bvec_pos.iter.bi_size;
1183         bool sparse = (op->op == CEPH_OSD_OP_SPARSE_READ);
1184
1185         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_BVECS);
1186         BUG_ON(!osd_data->num_bvecs);
1187
1188         dout("ceph_aio_complete_req %p rc %d bytes %u\n", inode, rc, len);
1189
1190         if (rc == -EOLDSNAPC) {
1191                 struct ceph_aio_work *aio_work;
1192                 BUG_ON(!aio_req->write);
1193
1194                 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
1195                 if (aio_work) {
1196                         INIT_WORK(&aio_work->work, ceph_aio_retry_work);
1197                         aio_work->req = req;
1198                         queue_work(ceph_inode_to_client(inode)->inode_wq,
1199                                    &aio_work->work);
1200                         return;
1201                 }
1202                 rc = -ENOMEM;
1203         } else if (!aio_req->write) {
1204                 if (sparse && rc >= 0)
1205                         rc = ceph_sparse_ext_map_end(op);
1206                 if (rc == -ENOENT)
1207                         rc = 0;
1208                 if (rc >= 0 && len > rc) {
1209                         struct iov_iter i;
1210                         int zlen = len - rc;
1211
1212                         /*
1213                          * If read is satisfied by single OSD request,
1214                          * it can pass EOF. Otherwise read is within
1215                          * i_size.
1216                          */
1217                         if (aio_req->num_reqs == 1) {
1218                                 loff_t i_size = i_size_read(inode);
1219                                 loff_t endoff = aio_req->iocb->ki_pos + rc;
1220                                 if (endoff < i_size)
1221                                         zlen = min_t(size_t, zlen,
1222                                                      i_size - endoff);
1223                                 aio_req->total_len = rc + zlen;
1224                         }
1225
1226                         iov_iter_bvec(&i, ITER_DEST, osd_data->bvec_pos.bvecs,
1227                                       osd_data->num_bvecs, len);
1228                         iov_iter_advance(&i, rc);
1229                         iov_iter_zero(zlen, &i);
1230                 }
1231         }
1232
1233         /* r_start_latency == 0 means the request was not submitted */
1234         if (req->r_start_latency) {
1235                 if (aio_req->write)
1236                         ceph_update_write_metrics(metric, req->r_start_latency,
1237                                                   req->r_end_latency, len, rc);
1238                 else
1239                         ceph_update_read_metrics(metric, req->r_start_latency,
1240                                                  req->r_end_latency, len, rc);
1241         }
1242
1243         put_bvecs(osd_data->bvec_pos.bvecs, osd_data->num_bvecs,
1244                   aio_req->should_dirty);
1245         ceph_osdc_put_request(req);
1246
1247         if (rc < 0)
1248                 cmpxchg(&aio_req->error, 0, rc);
1249
1250         ceph_aio_complete(inode, aio_req);
1251         return;
1252 }
1253
1254 static void ceph_aio_retry_work(struct work_struct *work)
1255 {
1256         struct ceph_aio_work *aio_work =
1257                 container_of(work, struct ceph_aio_work, work);
1258         struct ceph_osd_request *orig_req = aio_work->req;
1259         struct ceph_aio_request *aio_req = orig_req->r_priv;
1260         struct inode *inode = orig_req->r_inode;
1261         struct ceph_inode_info *ci = ceph_inode(inode);
1262         struct ceph_snap_context *snapc;
1263         struct ceph_osd_request *req;
1264         int ret;
1265
1266         spin_lock(&ci->i_ceph_lock);
1267         if (__ceph_have_pending_cap_snap(ci)) {
1268                 struct ceph_cap_snap *capsnap =
1269                         list_last_entry(&ci->i_cap_snaps,
1270                                         struct ceph_cap_snap,
1271                                         ci_item);
1272                 snapc = ceph_get_snap_context(capsnap->context);
1273         } else {
1274                 BUG_ON(!ci->i_head_snapc);
1275                 snapc = ceph_get_snap_context(ci->i_head_snapc);
1276         }
1277         spin_unlock(&ci->i_ceph_lock);
1278
1279         req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 1,
1280                         false, GFP_NOFS);
1281         if (!req) {
1282                 ret = -ENOMEM;
1283                 req = orig_req;
1284                 goto out;
1285         }
1286
1287         req->r_flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1288         ceph_oloc_copy(&req->r_base_oloc, &orig_req->r_base_oloc);
1289         ceph_oid_copy(&req->r_base_oid, &orig_req->r_base_oid);
1290
1291         req->r_ops[0] = orig_req->r_ops[0];
1292
1293         req->r_mtime = aio_req->mtime;
1294         req->r_data_offset = req->r_ops[0].extent.offset;
1295
1296         ret = ceph_osdc_alloc_messages(req, GFP_NOFS);
1297         if (ret) {
1298                 ceph_osdc_put_request(req);
1299                 req = orig_req;
1300                 goto out;
1301         }
1302
1303         ceph_osdc_put_request(orig_req);
1304
1305         req->r_callback = ceph_aio_complete_req;
1306         req->r_inode = inode;
1307         req->r_priv = aio_req;
1308
1309         ceph_osdc_start_request(req->r_osdc, req);
1310 out:
1311         if (ret < 0) {
1312                 req->r_result = ret;
1313                 ceph_aio_complete_req(req);
1314         }
1315
1316         ceph_put_snap_context(snapc);
1317         kfree(aio_work);
1318 }
1319
1320 static ssize_t
1321 ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
1322                        struct ceph_snap_context *snapc,
1323                        struct ceph_cap_flush **pcf)
1324 {
1325         struct file *file = iocb->ki_filp;
1326         struct inode *inode = file_inode(file);
1327         struct ceph_inode_info *ci = ceph_inode(inode);
1328         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1329         struct ceph_client_metric *metric = &fsc->mdsc->metric;
1330         struct ceph_vino vino;
1331         struct ceph_osd_request *req;
1332         struct bio_vec *bvecs;
1333         struct ceph_aio_request *aio_req = NULL;
1334         int num_pages = 0;
1335         int flags;
1336         int ret = 0;
1337         struct timespec64 mtime = current_time(inode);
1338         size_t count = iov_iter_count(iter);
1339         loff_t pos = iocb->ki_pos;
1340         bool write = iov_iter_rw(iter) == WRITE;
1341         bool should_dirty = !write && user_backed_iter(iter);
1342         bool sparse = ceph_test_mount_opt(fsc, SPARSEREAD);
1343
1344         if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1345                 return -EROFS;
1346
1347         dout("sync_direct_%s on file %p %lld~%u snapc %p seq %lld\n",
1348              (write ? "write" : "read"), file, pos, (unsigned)count,
1349              snapc, snapc ? snapc->seq : 0);
1350
1351         if (write) {
1352                 int ret2;
1353
1354                 ceph_fscache_invalidate(inode, true);
1355
1356                 ret2 = invalidate_inode_pages2_range(inode->i_mapping,
1357                                         pos >> PAGE_SHIFT,
1358                                         (pos + count - 1) >> PAGE_SHIFT);
1359                 if (ret2 < 0)
1360                         dout("invalidate_inode_pages2_range returned %d\n", ret2);
1361
1362                 flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1363         } else {
1364                 flags = CEPH_OSD_FLAG_READ;
1365         }
1366
1367         while (iov_iter_count(iter) > 0) {
1368                 u64 size = iov_iter_count(iter);
1369                 ssize_t len;
1370                 struct ceph_osd_req_op *op;
1371                 int readop = sparse ? CEPH_OSD_OP_SPARSE_READ : CEPH_OSD_OP_READ;
1372
1373                 if (write)
1374                         size = min_t(u64, size, fsc->mount_options->wsize);
1375                 else
1376                         size = min_t(u64, size, fsc->mount_options->rsize);
1377
1378                 vino = ceph_vino(inode);
1379                 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1380                                             vino, pos, &size, 0,
1381                                             1,
1382                                             write ? CEPH_OSD_OP_WRITE : readop,
1383                                             flags, snapc,
1384                                             ci->i_truncate_seq,
1385                                             ci->i_truncate_size,
1386                                             false);
1387                 if (IS_ERR(req)) {
1388                         ret = PTR_ERR(req);
1389                         break;
1390                 }
1391
1392                 len = iter_get_bvecs_alloc(iter, size, &bvecs, &num_pages);
1393                 if (len < 0) {
1394                         ceph_osdc_put_request(req);
1395                         ret = len;
1396                         break;
1397                 }
1398                 if (len != size)
1399                         osd_req_op_extent_update(req, 0, len);
1400
1401                 /*
1402                  * To simplify error handling, allow AIO when IO within i_size
1403                  * or IO can be satisfied by single OSD request.
1404                  */
1405                 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
1406                     (len == count || pos + count <= i_size_read(inode))) {
1407                         aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
1408                         if (aio_req) {
1409                                 aio_req->iocb = iocb;
1410                                 aio_req->write = write;
1411                                 aio_req->should_dirty = should_dirty;
1412                                 INIT_LIST_HEAD(&aio_req->osd_reqs);
1413                                 if (write) {
1414                                         aio_req->mtime = mtime;
1415                                         swap(aio_req->prealloc_cf, *pcf);
1416                                 }
1417                         }
1418                         /* ignore error */
1419                 }
1420
1421                 if (write) {
1422                         /*
1423                          * throw out any page cache pages in this range. this
1424                          * may block.
1425                          */
1426                         truncate_inode_pages_range(inode->i_mapping, pos,
1427                                                    PAGE_ALIGN(pos + len) - 1);
1428
1429                         req->r_mtime = mtime;
1430                 }
1431
1432                 osd_req_op_extent_osd_data_bvecs(req, 0, bvecs, num_pages, len);
1433                 op = &req->r_ops[0];
1434                 if (sparse) {
1435                         ret = ceph_alloc_sparse_ext_map(op);
1436                         if (ret) {
1437                                 ceph_osdc_put_request(req);
1438                                 break;
1439                         }
1440                 }
1441
1442                 if (aio_req) {
1443                         aio_req->total_len += len;
1444                         aio_req->num_reqs++;
1445                         atomic_inc(&aio_req->pending_reqs);
1446
1447                         req->r_callback = ceph_aio_complete_req;
1448                         req->r_inode = inode;
1449                         req->r_priv = aio_req;
1450                         list_add_tail(&req->r_private_item, &aio_req->osd_reqs);
1451
1452                         pos += len;
1453                         continue;
1454                 }
1455
1456                 ceph_osdc_start_request(req->r_osdc, req);
1457                 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1458
1459                 if (write)
1460                         ceph_update_write_metrics(metric, req->r_start_latency,
1461                                                   req->r_end_latency, len, ret);
1462                 else
1463                         ceph_update_read_metrics(metric, req->r_start_latency,
1464                                                  req->r_end_latency, len, ret);
1465
1466                 size = i_size_read(inode);
1467                 if (!write) {
1468                         if (sparse && ret >= 0)
1469                                 ret = ceph_sparse_ext_map_end(op);
1470                         else if (ret == -ENOENT)
1471                                 ret = 0;
1472
1473                         if (ret >= 0 && ret < len && pos + ret < size) {
1474                                 struct iov_iter i;
1475                                 int zlen = min_t(size_t, len - ret,
1476                                                  size - pos - ret);
1477
1478                                 iov_iter_bvec(&i, ITER_DEST, bvecs, num_pages, len);
1479                                 iov_iter_advance(&i, ret);
1480                                 iov_iter_zero(zlen, &i);
1481                                 ret += zlen;
1482                         }
1483                         if (ret >= 0)
1484                                 len = ret;
1485                 }
1486
1487                 put_bvecs(bvecs, num_pages, should_dirty);
1488                 ceph_osdc_put_request(req);
1489                 if (ret < 0)
1490                         break;
1491
1492                 pos += len;
1493                 if (!write && pos >= size)
1494                         break;
1495
1496                 if (write && pos > size) {
1497                         if (ceph_inode_set_size(inode, pos))
1498                                 ceph_check_caps(ceph_inode(inode),
1499                                                 CHECK_CAPS_AUTHONLY);
1500                 }
1501         }
1502
1503         if (aio_req) {
1504                 LIST_HEAD(osd_reqs);
1505
1506                 if (aio_req->num_reqs == 0) {
1507                         kfree(aio_req);
1508                         return ret;
1509                 }
1510
1511                 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
1512                                               CEPH_CAP_FILE_RD);
1513
1514                 list_splice(&aio_req->osd_reqs, &osd_reqs);
1515                 inode_dio_begin(inode);
1516                 while (!list_empty(&osd_reqs)) {
1517                         req = list_first_entry(&osd_reqs,
1518                                                struct ceph_osd_request,
1519                                                r_private_item);
1520                         list_del_init(&req->r_private_item);
1521                         if (ret >= 0)
1522                                 ceph_osdc_start_request(req->r_osdc, req);
1523                         if (ret < 0) {
1524                                 req->r_result = ret;
1525                                 ceph_aio_complete_req(req);
1526                         }
1527                 }
1528                 return -EIOCBQUEUED;
1529         }
1530
1531         if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
1532                 ret = pos - iocb->ki_pos;
1533                 iocb->ki_pos = pos;
1534         }
1535         return ret;
1536 }
1537
1538 /*
1539  * Synchronous write, straight from __user pointer or user pages.
1540  *
1541  * If write spans object boundary, just do multiple writes.  (For a
1542  * correct atomic write, we should e.g. take write locks on all
1543  * objects, rollback on failure, etc.)
1544  */
1545 static ssize_t
1546 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
1547                 struct ceph_snap_context *snapc)
1548 {
1549         struct file *file = iocb->ki_filp;
1550         struct inode *inode = file_inode(file);
1551         struct ceph_inode_info *ci = ceph_inode(inode);
1552         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1553         struct ceph_vino vino;
1554         struct ceph_osd_request *req;
1555         struct page **pages;
1556         u64 len;
1557         int num_pages;
1558         int written = 0;
1559         int flags;
1560         int ret;
1561         bool check_caps = false;
1562         struct timespec64 mtime = current_time(inode);
1563         size_t count = iov_iter_count(from);
1564
1565         if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1566                 return -EROFS;
1567
1568         dout("sync_write on file %p %lld~%u snapc %p seq %lld\n",
1569              file, pos, (unsigned)count, snapc, snapc->seq);
1570
1571         ret = filemap_write_and_wait_range(inode->i_mapping,
1572                                            pos, pos + count - 1);
1573         if (ret < 0)
1574                 return ret;
1575
1576         ceph_fscache_invalidate(inode, false);
1577         ret = invalidate_inode_pages2_range(inode->i_mapping,
1578                                             pos >> PAGE_SHIFT,
1579                                             (pos + count - 1) >> PAGE_SHIFT);
1580         if (ret < 0)
1581                 dout("invalidate_inode_pages2_range returned %d\n", ret);
1582
1583         flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1584
1585         while ((len = iov_iter_count(from)) > 0) {
1586                 size_t left;
1587                 int n;
1588
1589                 vino = ceph_vino(inode);
1590                 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1591                                             vino, pos, &len, 0, 1,
1592                                             CEPH_OSD_OP_WRITE, flags, snapc,
1593                                             ci->i_truncate_seq,
1594                                             ci->i_truncate_size,
1595                                             false);
1596                 if (IS_ERR(req)) {
1597                         ret = PTR_ERR(req);
1598                         break;
1599                 }
1600
1601                 /*
1602                  * write from beginning of first page,
1603                  * regardless of io alignment
1604                  */
1605                 num_pages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1606
1607                 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1608                 if (IS_ERR(pages)) {
1609                         ret = PTR_ERR(pages);
1610                         goto out;
1611                 }
1612
1613                 left = len;
1614                 for (n = 0; n < num_pages; n++) {
1615                         size_t plen = min_t(size_t, left, PAGE_SIZE);
1616                         ret = copy_page_from_iter(pages[n], 0, plen, from);
1617                         if (ret != plen) {
1618                                 ret = -EFAULT;
1619                                 break;
1620                         }
1621                         left -= ret;
1622                 }
1623
1624                 if (ret < 0) {
1625                         ceph_release_page_vector(pages, num_pages);
1626                         goto out;
1627                 }
1628
1629                 req->r_inode = inode;
1630
1631                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
1632                                                 false, true);
1633
1634                 req->r_mtime = mtime;
1635                 ceph_osdc_start_request(&fsc->client->osdc, req);
1636                 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1637
1638                 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1639                                           req->r_end_latency, len, ret);
1640 out:
1641                 ceph_osdc_put_request(req);
1642                 if (ret != 0) {
1643                         ceph_set_error_write(ci);
1644                         break;
1645                 }
1646
1647                 ceph_clear_error_write(ci);
1648                 pos += len;
1649                 written += len;
1650                 if (pos > i_size_read(inode)) {
1651                         check_caps = ceph_inode_set_size(inode, pos);
1652                         if (check_caps)
1653                                 ceph_check_caps(ceph_inode(inode),
1654                                                 CHECK_CAPS_AUTHONLY);
1655                 }
1656
1657         }
1658
1659         if (ret != -EOLDSNAPC && written > 0) {
1660                 ret = written;
1661                 iocb->ki_pos = pos;
1662         }
1663         return ret;
1664 }
1665
1666 /*
1667  * Wrap generic_file_aio_read with checks for cap bits on the inode.
1668  * Atomically grab references, so that those bits are not released
1669  * back to the MDS mid-read.
1670  *
1671  * Hmm, the sync read case isn't actually async... should it be?
1672  */
1673 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
1674 {
1675         struct file *filp = iocb->ki_filp;
1676         struct ceph_file_info *fi = filp->private_data;
1677         size_t len = iov_iter_count(to);
1678         struct inode *inode = file_inode(filp);
1679         struct ceph_inode_info *ci = ceph_inode(inode);
1680         bool direct_lock = iocb->ki_flags & IOCB_DIRECT;
1681         ssize_t ret;
1682         int want = 0, got = 0;
1683         int retry_op = 0, read = 0;
1684
1685 again:
1686         dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
1687              inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
1688
1689         if (ceph_inode_is_shutdown(inode))
1690                 return -ESTALE;
1691
1692         if (direct_lock)
1693                 ceph_start_io_direct(inode);
1694         else
1695                 ceph_start_io_read(inode);
1696
1697         if (!(fi->flags & CEPH_F_SYNC) && !direct_lock)
1698                 want |= CEPH_CAP_FILE_CACHE;
1699         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1700                 want |= CEPH_CAP_FILE_LAZYIO;
1701
1702         ret = ceph_get_caps(filp, CEPH_CAP_FILE_RD, want, -1, &got);
1703         if (ret < 0) {
1704                 if (direct_lock)
1705                         ceph_end_io_direct(inode);
1706                 else
1707                         ceph_end_io_read(inode);
1708                 return ret;
1709         }
1710
1711         if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1712             (iocb->ki_flags & IOCB_DIRECT) ||
1713             (fi->flags & CEPH_F_SYNC)) {
1714
1715                 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1716                      inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1717                      ceph_cap_string(got));
1718
1719                 if (!ceph_has_inline_data(ci)) {
1720                         if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
1721                                 ret = ceph_direct_read_write(iocb, to,
1722                                                              NULL, NULL);
1723                                 if (ret >= 0 && ret < len)
1724                                         retry_op = CHECK_EOF;
1725                         } else {
1726                                 ret = ceph_sync_read(iocb, to, &retry_op);
1727                         }
1728                 } else {
1729                         retry_op = READ_INLINE;
1730                 }
1731         } else {
1732                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1733                 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1734                      inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1735                      ceph_cap_string(got));
1736                 ceph_add_rw_context(fi, &rw_ctx);
1737                 ret = generic_file_read_iter(iocb, to);
1738                 ceph_del_rw_context(fi, &rw_ctx);
1739         }
1740
1741         dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
1742              inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
1743         ceph_put_cap_refs(ci, got);
1744
1745         if (direct_lock)
1746                 ceph_end_io_direct(inode);
1747         else
1748                 ceph_end_io_read(inode);
1749
1750         if (retry_op > HAVE_RETRIED && ret >= 0) {
1751                 int statret;
1752                 struct page *page = NULL;
1753                 loff_t i_size;
1754                 if (retry_op == READ_INLINE) {
1755                         page = __page_cache_alloc(GFP_KERNEL);
1756                         if (!page)
1757                                 return -ENOMEM;
1758                 }
1759
1760                 statret = __ceph_do_getattr(inode, page,
1761                                             CEPH_STAT_CAP_INLINE_DATA, !!page);
1762                 if (statret < 0) {
1763                         if (page)
1764                                 __free_page(page);
1765                         if (statret == -ENODATA) {
1766                                 BUG_ON(retry_op != READ_INLINE);
1767                                 goto again;
1768                         }
1769                         return statret;
1770                 }
1771
1772                 i_size = i_size_read(inode);
1773                 if (retry_op == READ_INLINE) {
1774                         BUG_ON(ret > 0 || read > 0);
1775                         if (iocb->ki_pos < i_size &&
1776                             iocb->ki_pos < PAGE_SIZE) {
1777                                 loff_t end = min_t(loff_t, i_size,
1778                                                    iocb->ki_pos + len);
1779                                 end = min_t(loff_t, end, PAGE_SIZE);
1780                                 if (statret < end)
1781                                         zero_user_segment(page, statret, end);
1782                                 ret = copy_page_to_iter(page,
1783                                                 iocb->ki_pos & ~PAGE_MASK,
1784                                                 end - iocb->ki_pos, to);
1785                                 iocb->ki_pos += ret;
1786                                 read += ret;
1787                         }
1788                         if (iocb->ki_pos < i_size && read < len) {
1789                                 size_t zlen = min_t(size_t, len - read,
1790                                                     i_size - iocb->ki_pos);
1791                                 ret = iov_iter_zero(zlen, to);
1792                                 iocb->ki_pos += ret;
1793                                 read += ret;
1794                         }
1795                         __free_pages(page, 0);
1796                         return read;
1797                 }
1798
1799                 /* hit EOF or hole? */
1800                 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
1801                     ret < len) {
1802                         dout("sync_read hit hole, ppos %lld < size %lld"
1803                              ", reading more\n", iocb->ki_pos, i_size);
1804
1805                         read += ret;
1806                         len -= ret;
1807                         retry_op = HAVE_RETRIED;
1808                         goto again;
1809                 }
1810         }
1811
1812         if (ret >= 0)
1813                 ret += read;
1814
1815         return ret;
1816 }
1817
1818 /*
1819  * Wrap filemap_splice_read with checks for cap bits on the inode.
1820  * Atomically grab references, so that those bits are not released
1821  * back to the MDS mid-read.
1822  */
1823 static ssize_t ceph_splice_read(struct file *in, loff_t *ppos,
1824                                 struct pipe_inode_info *pipe,
1825                                 size_t len, unsigned int flags)
1826 {
1827         struct ceph_file_info *fi = in->private_data;
1828         struct inode *inode = file_inode(in);
1829         struct ceph_inode_info *ci = ceph_inode(inode);
1830         ssize_t ret;
1831         int want = 0, got = 0;
1832         CEPH_DEFINE_RW_CONTEXT(rw_ctx, 0);
1833
1834         dout("splice_read %p %llx.%llx %llu~%zu trying to get caps on %p\n",
1835              inode, ceph_vinop(inode), *ppos, len, inode);
1836
1837         if (ceph_inode_is_shutdown(inode))
1838                 return -ESTALE;
1839
1840         if (ceph_has_inline_data(ci) ||
1841             (fi->flags & CEPH_F_SYNC))
1842                 return copy_splice_read(in, ppos, pipe, len, flags);
1843
1844         ceph_start_io_read(inode);
1845
1846         want = CEPH_CAP_FILE_CACHE;
1847         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1848                 want |= CEPH_CAP_FILE_LAZYIO;
1849
1850         ret = ceph_get_caps(in, CEPH_CAP_FILE_RD, want, -1, &got);
1851         if (ret < 0)
1852                 goto out_end;
1853
1854         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) == 0) {
1855                 dout("splice_read/sync %p %llx.%llx %llu~%zu got cap refs on %s\n",
1856                      inode, ceph_vinop(inode), *ppos, len,
1857                      ceph_cap_string(got));
1858
1859                 ceph_put_cap_refs(ci, got);
1860                 ceph_end_io_read(inode);
1861                 return copy_splice_read(in, ppos, pipe, len, flags);
1862         }
1863
1864         dout("splice_read %p %llx.%llx %llu~%zu got cap refs on %s\n",
1865              inode, ceph_vinop(inode), *ppos, len, ceph_cap_string(got));
1866
1867         rw_ctx.caps = got;
1868         ceph_add_rw_context(fi, &rw_ctx);
1869         ret = filemap_splice_read(in, ppos, pipe, len, flags);
1870         ceph_del_rw_context(fi, &rw_ctx);
1871
1872         dout("splice_read %p %llx.%llx dropping cap refs on %s = %zd\n",
1873              inode, ceph_vinop(inode), ceph_cap_string(got), ret);
1874
1875         ceph_put_cap_refs(ci, got);
1876 out_end:
1877         ceph_end_io_read(inode);
1878         return ret;
1879 }
1880
1881 /*
1882  * Take cap references to avoid releasing caps to MDS mid-write.
1883  *
1884  * If we are synchronous, and write with an old snap context, the OSD
1885  * may return EOLDSNAPC.  In that case, retry the write.. _after_
1886  * dropping our cap refs and allowing the pending snap to logically
1887  * complete _before_ this write occurs.
1888  *
1889  * If we are near ENOSPC, write synchronously.
1890  */
1891 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
1892 {
1893         struct file *file = iocb->ki_filp;
1894         struct ceph_file_info *fi = file->private_data;
1895         struct inode *inode = file_inode(file);
1896         struct ceph_inode_info *ci = ceph_inode(inode);
1897         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1898         struct ceph_osd_client *osdc = &fsc->client->osdc;
1899         struct ceph_cap_flush *prealloc_cf;
1900         ssize_t count, written = 0;
1901         int err, want = 0, got;
1902         bool direct_lock = false;
1903         u32 map_flags;
1904         u64 pool_flags;
1905         loff_t pos;
1906         loff_t limit = max(i_size_read(inode), fsc->max_file_size);
1907
1908         if (ceph_inode_is_shutdown(inode))
1909                 return -ESTALE;
1910
1911         if (ceph_snap(inode) != CEPH_NOSNAP)
1912                 return -EROFS;
1913
1914         prealloc_cf = ceph_alloc_cap_flush();
1915         if (!prealloc_cf)
1916                 return -ENOMEM;
1917
1918         if ((iocb->ki_flags & (IOCB_DIRECT | IOCB_APPEND)) == IOCB_DIRECT)
1919                 direct_lock = true;
1920
1921 retry_snap:
1922         if (direct_lock)
1923                 ceph_start_io_direct(inode);
1924         else
1925                 ceph_start_io_write(inode);
1926
1927         if (iocb->ki_flags & IOCB_APPEND) {
1928                 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1929                 if (err < 0)
1930                         goto out;
1931         }
1932
1933         err = generic_write_checks(iocb, from);
1934         if (err <= 0)
1935                 goto out;
1936
1937         pos = iocb->ki_pos;
1938         if (unlikely(pos >= limit)) {
1939                 err = -EFBIG;
1940                 goto out;
1941         } else {
1942                 iov_iter_truncate(from, limit - pos);
1943         }
1944
1945         count = iov_iter_count(from);
1946         if (ceph_quota_is_max_bytes_exceeded(inode, pos + count)) {
1947                 err = -EDQUOT;
1948                 goto out;
1949         }
1950
1951         down_read(&osdc->lock);
1952         map_flags = osdc->osdmap->flags;
1953         pool_flags = ceph_pg_pool_flags(osdc->osdmap, ci->i_layout.pool_id);
1954         up_read(&osdc->lock);
1955         if ((map_flags & CEPH_OSDMAP_FULL) ||
1956             (pool_flags & CEPH_POOL_FLAG_FULL)) {
1957                 err = -ENOSPC;
1958                 goto out;
1959         }
1960
1961         err = file_remove_privs(file);
1962         if (err)
1963                 goto out;
1964
1965         dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
1966              inode, ceph_vinop(inode), pos, count, i_size_read(inode));
1967         if (!(fi->flags & CEPH_F_SYNC) && !direct_lock)
1968                 want |= CEPH_CAP_FILE_BUFFER;
1969         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1970                 want |= CEPH_CAP_FILE_LAZYIO;
1971         got = 0;
1972         err = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, pos + count, &got);
1973         if (err < 0)
1974                 goto out;
1975
1976         err = file_update_time(file);
1977         if (err)
1978                 goto out_caps;
1979
1980         inode_inc_iversion_raw(inode);
1981
1982         dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
1983              inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
1984
1985         if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1986             (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC) ||
1987             (ci->i_ceph_flags & CEPH_I_ERROR_WRITE)) {
1988                 struct ceph_snap_context *snapc;
1989                 struct iov_iter data;
1990
1991                 spin_lock(&ci->i_ceph_lock);
1992                 if (__ceph_have_pending_cap_snap(ci)) {
1993                         struct ceph_cap_snap *capsnap =
1994                                         list_last_entry(&ci->i_cap_snaps,
1995                                                         struct ceph_cap_snap,
1996                                                         ci_item);
1997                         snapc = ceph_get_snap_context(capsnap->context);
1998                 } else {
1999                         BUG_ON(!ci->i_head_snapc);
2000                         snapc = ceph_get_snap_context(ci->i_head_snapc);
2001                 }
2002                 spin_unlock(&ci->i_ceph_lock);
2003
2004                 /* we might need to revert back to that point */
2005                 data = *from;
2006                 if (iocb->ki_flags & IOCB_DIRECT)
2007                         written = ceph_direct_read_write(iocb, &data, snapc,
2008                                                          &prealloc_cf);
2009                 else
2010                         written = ceph_sync_write(iocb, &data, pos, snapc);
2011                 if (direct_lock)
2012                         ceph_end_io_direct(inode);
2013                 else
2014                         ceph_end_io_write(inode);
2015                 if (written > 0)
2016                         iov_iter_advance(from, written);
2017                 ceph_put_snap_context(snapc);
2018         } else {
2019                 /*
2020                  * No need to acquire the i_truncate_mutex. Because
2021                  * the MDS revokes Fwb caps before sending truncate
2022                  * message to us. We can't get Fwb cap while there
2023                  * are pending vmtruncate. So write and vmtruncate
2024                  * can not run at the same time
2025                  */
2026                 written = generic_perform_write(iocb, from);
2027                 ceph_end_io_write(inode);
2028         }
2029
2030         if (written >= 0) {
2031                 int dirty;
2032
2033                 spin_lock(&ci->i_ceph_lock);
2034                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
2035                                                &prealloc_cf);
2036                 spin_unlock(&ci->i_ceph_lock);
2037                 if (dirty)
2038                         __mark_inode_dirty(inode, dirty);
2039                 if (ceph_quota_is_max_bytes_approaching(inode, iocb->ki_pos))
2040                         ceph_check_caps(ci, CHECK_CAPS_FLUSH);
2041         }
2042
2043         dout("aio_write %p %llx.%llx %llu~%u  dropping cap refs on %s\n",
2044              inode, ceph_vinop(inode), pos, (unsigned)count,
2045              ceph_cap_string(got));
2046         ceph_put_cap_refs(ci, got);
2047
2048         if (written == -EOLDSNAPC) {
2049                 dout("aio_write %p %llx.%llx %llu~%u" "got EOLDSNAPC, retrying\n",
2050                      inode, ceph_vinop(inode), pos, (unsigned)count);
2051                 goto retry_snap;
2052         }
2053
2054         if (written >= 0) {
2055                 if ((map_flags & CEPH_OSDMAP_NEARFULL) ||
2056                     (pool_flags & CEPH_POOL_FLAG_NEARFULL))
2057                         iocb->ki_flags |= IOCB_DSYNC;
2058                 written = generic_write_sync(iocb, written);
2059         }
2060
2061         goto out_unlocked;
2062 out_caps:
2063         ceph_put_cap_refs(ci, got);
2064 out:
2065         if (direct_lock)
2066                 ceph_end_io_direct(inode);
2067         else
2068                 ceph_end_io_write(inode);
2069 out_unlocked:
2070         ceph_free_cap_flush(prealloc_cf);
2071         return written ? written : err;
2072 }
2073
2074 /*
2075  * llseek.  be sure to verify file size on SEEK_END.
2076  */
2077 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
2078 {
2079         if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
2080                 struct inode *inode = file_inode(file);
2081                 int ret;
2082
2083                 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
2084                 if (ret < 0)
2085                         return ret;
2086         }
2087         return generic_file_llseek(file, offset, whence);
2088 }
2089
2090 static inline void ceph_zero_partial_page(
2091         struct inode *inode, loff_t offset, unsigned size)
2092 {
2093         struct page *page;
2094         pgoff_t index = offset >> PAGE_SHIFT;
2095
2096         page = find_lock_page(inode->i_mapping, index);
2097         if (page) {
2098                 wait_on_page_writeback(page);
2099                 zero_user(page, offset & (PAGE_SIZE - 1), size);
2100                 unlock_page(page);
2101                 put_page(page);
2102         }
2103 }
2104
2105 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
2106                                       loff_t length)
2107 {
2108         loff_t nearly = round_up(offset, PAGE_SIZE);
2109         if (offset < nearly) {
2110                 loff_t size = nearly - offset;
2111                 if (length < size)
2112                         size = length;
2113                 ceph_zero_partial_page(inode, offset, size);
2114                 offset += size;
2115                 length -= size;
2116         }
2117         if (length >= PAGE_SIZE) {
2118                 loff_t size = round_down(length, PAGE_SIZE);
2119                 truncate_pagecache_range(inode, offset, offset + size - 1);
2120                 offset += size;
2121                 length -= size;
2122         }
2123         if (length)
2124                 ceph_zero_partial_page(inode, offset, length);
2125 }
2126
2127 static int ceph_zero_partial_object(struct inode *inode,
2128                                     loff_t offset, loff_t *length)
2129 {
2130         struct ceph_inode_info *ci = ceph_inode(inode);
2131         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
2132         struct ceph_osd_request *req;
2133         int ret = 0;
2134         loff_t zero = 0;
2135         int op;
2136
2137         if (ceph_inode_is_shutdown(inode))
2138                 return -EIO;
2139
2140         if (!length) {
2141                 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
2142                 length = &zero;
2143         } else {
2144                 op = CEPH_OSD_OP_ZERO;
2145         }
2146
2147         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
2148                                         ceph_vino(inode),
2149                                         offset, length,
2150                                         0, 1, op,
2151                                         CEPH_OSD_FLAG_WRITE,
2152                                         NULL, 0, 0, false);
2153         if (IS_ERR(req)) {
2154                 ret = PTR_ERR(req);
2155                 goto out;
2156         }
2157
2158         req->r_mtime = inode->i_mtime;
2159         ceph_osdc_start_request(&fsc->client->osdc, req);
2160         ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
2161         if (ret == -ENOENT)
2162                 ret = 0;
2163         ceph_osdc_put_request(req);
2164
2165 out:
2166         return ret;
2167 }
2168
2169 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
2170 {
2171         int ret = 0;
2172         struct ceph_inode_info *ci = ceph_inode(inode);
2173         s32 stripe_unit = ci->i_layout.stripe_unit;
2174         s32 stripe_count = ci->i_layout.stripe_count;
2175         s32 object_size = ci->i_layout.object_size;
2176         u64 object_set_size = object_size * stripe_count;
2177         u64 nearly, t;
2178
2179         /* round offset up to next period boundary */
2180         nearly = offset + object_set_size - 1;
2181         t = nearly;
2182         nearly -= do_div(t, object_set_size);
2183
2184         while (length && offset < nearly) {
2185                 loff_t size = length;
2186                 ret = ceph_zero_partial_object(inode, offset, &size);
2187                 if (ret < 0)
2188                         return ret;
2189                 offset += size;
2190                 length -= size;
2191         }
2192         while (length >= object_set_size) {
2193                 int i;
2194                 loff_t pos = offset;
2195                 for (i = 0; i < stripe_count; ++i) {
2196                         ret = ceph_zero_partial_object(inode, pos, NULL);
2197                         if (ret < 0)
2198                                 return ret;
2199                         pos += stripe_unit;
2200                 }
2201                 offset += object_set_size;
2202                 length -= object_set_size;
2203         }
2204         while (length) {
2205                 loff_t size = length;
2206                 ret = ceph_zero_partial_object(inode, offset, &size);
2207                 if (ret < 0)
2208                         return ret;
2209                 offset += size;
2210                 length -= size;
2211         }
2212         return ret;
2213 }
2214
2215 static long ceph_fallocate(struct file *file, int mode,
2216                                 loff_t offset, loff_t length)
2217 {
2218         struct ceph_file_info *fi = file->private_data;
2219         struct inode *inode = file_inode(file);
2220         struct ceph_inode_info *ci = ceph_inode(inode);
2221         struct ceph_cap_flush *prealloc_cf;
2222         int want, got = 0;
2223         int dirty;
2224         int ret = 0;
2225         loff_t endoff = 0;
2226         loff_t size;
2227
2228         dout("%s %p %llx.%llx mode %x, offset %llu length %llu\n", __func__,
2229              inode, ceph_vinop(inode), mode, offset, length);
2230
2231         if (mode != (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2232                 return -EOPNOTSUPP;
2233
2234         if (!S_ISREG(inode->i_mode))
2235                 return -EOPNOTSUPP;
2236
2237         if (IS_ENCRYPTED(inode))
2238                 return -EOPNOTSUPP;
2239
2240         prealloc_cf = ceph_alloc_cap_flush();
2241         if (!prealloc_cf)
2242                 return -ENOMEM;
2243
2244         inode_lock(inode);
2245
2246         if (ceph_snap(inode) != CEPH_NOSNAP) {
2247                 ret = -EROFS;
2248                 goto unlock;
2249         }
2250
2251         size = i_size_read(inode);
2252
2253         /* Are we punching a hole beyond EOF? */
2254         if (offset >= size)
2255                 goto unlock;
2256         if ((offset + length) > size)
2257                 length = size - offset;
2258
2259         if (fi->fmode & CEPH_FILE_MODE_LAZY)
2260                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
2261         else
2262                 want = CEPH_CAP_FILE_BUFFER;
2263
2264         ret = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, endoff, &got);
2265         if (ret < 0)
2266                 goto unlock;
2267
2268         ret = file_modified(file);
2269         if (ret)
2270                 goto put_caps;
2271
2272         filemap_invalidate_lock(inode->i_mapping);
2273         ceph_fscache_invalidate(inode, false);
2274         ceph_zero_pagecache_range(inode, offset, length);
2275         ret = ceph_zero_objects(inode, offset, length);
2276
2277         if (!ret) {
2278                 spin_lock(&ci->i_ceph_lock);
2279                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
2280                                                &prealloc_cf);
2281                 spin_unlock(&ci->i_ceph_lock);
2282                 if (dirty)
2283                         __mark_inode_dirty(inode, dirty);
2284         }
2285         filemap_invalidate_unlock(inode->i_mapping);
2286
2287 put_caps:
2288         ceph_put_cap_refs(ci, got);
2289 unlock:
2290         inode_unlock(inode);
2291         ceph_free_cap_flush(prealloc_cf);
2292         return ret;
2293 }
2294
2295 /*
2296  * This function tries to get FILE_WR capabilities for dst_ci and FILE_RD for
2297  * src_ci.  Two attempts are made to obtain both caps, and an error is return if
2298  * this fails; zero is returned on success.
2299  */
2300 static int get_rd_wr_caps(struct file *src_filp, int *src_got,
2301                           struct file *dst_filp,
2302                           loff_t dst_endoff, int *dst_got)
2303 {
2304         int ret = 0;
2305         bool retrying = false;
2306
2307 retry_caps:
2308         ret = ceph_get_caps(dst_filp, CEPH_CAP_FILE_WR, CEPH_CAP_FILE_BUFFER,
2309                             dst_endoff, dst_got);
2310         if (ret < 0)
2311                 return ret;
2312
2313         /*
2314          * Since we're already holding the FILE_WR capability for the dst file,
2315          * we would risk a deadlock by using ceph_get_caps.  Thus, we'll do some
2316          * retry dance instead to try to get both capabilities.
2317          */
2318         ret = ceph_try_get_caps(file_inode(src_filp),
2319                                 CEPH_CAP_FILE_RD, CEPH_CAP_FILE_SHARED,
2320                                 false, src_got);
2321         if (ret <= 0) {
2322                 /* Start by dropping dst_ci caps and getting src_ci caps */
2323                 ceph_put_cap_refs(ceph_inode(file_inode(dst_filp)), *dst_got);
2324                 if (retrying) {
2325                         if (!ret)
2326                                 /* ceph_try_get_caps masks EAGAIN */
2327                                 ret = -EAGAIN;
2328                         return ret;
2329                 }
2330                 ret = ceph_get_caps(src_filp, CEPH_CAP_FILE_RD,
2331                                     CEPH_CAP_FILE_SHARED, -1, src_got);
2332                 if (ret < 0)
2333                         return ret;
2334                 /*... drop src_ci caps too, and retry */
2335                 ceph_put_cap_refs(ceph_inode(file_inode(src_filp)), *src_got);
2336                 retrying = true;
2337                 goto retry_caps;
2338         }
2339         return ret;
2340 }
2341
2342 static void put_rd_wr_caps(struct ceph_inode_info *src_ci, int src_got,
2343                            struct ceph_inode_info *dst_ci, int dst_got)
2344 {
2345         ceph_put_cap_refs(src_ci, src_got);
2346         ceph_put_cap_refs(dst_ci, dst_got);
2347 }
2348
2349 /*
2350  * This function does several size-related checks, returning an error if:
2351  *  - source file is smaller than off+len
2352  *  - destination file size is not OK (inode_newsize_ok())
2353  *  - max bytes quotas is exceeded
2354  */
2355 static int is_file_size_ok(struct inode *src_inode, struct inode *dst_inode,
2356                            loff_t src_off, loff_t dst_off, size_t len)
2357 {
2358         loff_t size, endoff;
2359
2360         size = i_size_read(src_inode);
2361         /*
2362          * Don't copy beyond source file EOF.  Instead of simply setting length
2363          * to (size - src_off), just drop to VFS default implementation, as the
2364          * local i_size may be stale due to other clients writing to the source
2365          * inode.
2366          */
2367         if (src_off + len > size) {
2368                 dout("Copy beyond EOF (%llu + %zu > %llu)\n",
2369                      src_off, len, size);
2370                 return -EOPNOTSUPP;
2371         }
2372         size = i_size_read(dst_inode);
2373
2374         endoff = dst_off + len;
2375         if (inode_newsize_ok(dst_inode, endoff))
2376                 return -EOPNOTSUPP;
2377
2378         if (ceph_quota_is_max_bytes_exceeded(dst_inode, endoff))
2379                 return -EDQUOT;
2380
2381         return 0;
2382 }
2383
2384 static struct ceph_osd_request *
2385 ceph_alloc_copyfrom_request(struct ceph_osd_client *osdc,
2386                             u64 src_snapid,
2387                             struct ceph_object_id *src_oid,
2388                             struct ceph_object_locator *src_oloc,
2389                             struct ceph_object_id *dst_oid,
2390                             struct ceph_object_locator *dst_oloc,
2391                             u32 truncate_seq, u64 truncate_size)
2392 {
2393         struct ceph_osd_request *req;
2394         int ret;
2395         u32 src_fadvise_flags =
2396                 CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2397                 CEPH_OSD_OP_FLAG_FADVISE_NOCACHE;
2398         u32 dst_fadvise_flags =
2399                 CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2400                 CEPH_OSD_OP_FLAG_FADVISE_DONTNEED;
2401
2402         req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_KERNEL);
2403         if (!req)
2404                 return ERR_PTR(-ENOMEM);
2405
2406         req->r_flags = CEPH_OSD_FLAG_WRITE;
2407
2408         ceph_oloc_copy(&req->r_t.base_oloc, dst_oloc);
2409         ceph_oid_copy(&req->r_t.base_oid, dst_oid);
2410
2411         ret = osd_req_op_copy_from_init(req, src_snapid, 0,
2412                                         src_oid, src_oloc,
2413                                         src_fadvise_flags,
2414                                         dst_fadvise_flags,
2415                                         truncate_seq,
2416                                         truncate_size,
2417                                         CEPH_OSD_COPY_FROM_FLAG_TRUNCATE_SEQ);
2418         if (ret)
2419                 goto out;
2420
2421         ret = ceph_osdc_alloc_messages(req, GFP_KERNEL);
2422         if (ret)
2423                 goto out;
2424
2425         return req;
2426
2427 out:
2428         ceph_osdc_put_request(req);
2429         return ERR_PTR(ret);
2430 }
2431
2432 static ssize_t ceph_do_objects_copy(struct ceph_inode_info *src_ci, u64 *src_off,
2433                                     struct ceph_inode_info *dst_ci, u64 *dst_off,
2434                                     struct ceph_fs_client *fsc,
2435                                     size_t len, unsigned int flags)
2436 {
2437         struct ceph_object_locator src_oloc, dst_oloc;
2438         struct ceph_object_id src_oid, dst_oid;
2439         struct ceph_osd_client *osdc;
2440         struct ceph_osd_request *req;
2441         size_t bytes = 0;
2442         u64 src_objnum, src_objoff, dst_objnum, dst_objoff;
2443         u32 src_objlen, dst_objlen;
2444         u32 object_size = src_ci->i_layout.object_size;
2445         int ret;
2446
2447         src_oloc.pool = src_ci->i_layout.pool_id;
2448         src_oloc.pool_ns = ceph_try_get_string(src_ci->i_layout.pool_ns);
2449         dst_oloc.pool = dst_ci->i_layout.pool_id;
2450         dst_oloc.pool_ns = ceph_try_get_string(dst_ci->i_layout.pool_ns);
2451         osdc = &fsc->client->osdc;
2452
2453         while (len >= object_size) {
2454                 ceph_calc_file_object_mapping(&src_ci->i_layout, *src_off,
2455                                               object_size, &src_objnum,
2456                                               &src_objoff, &src_objlen);
2457                 ceph_calc_file_object_mapping(&dst_ci->i_layout, *dst_off,
2458                                               object_size, &dst_objnum,
2459                                               &dst_objoff, &dst_objlen);
2460                 ceph_oid_init(&src_oid);
2461                 ceph_oid_printf(&src_oid, "%llx.%08llx",
2462                                 src_ci->i_vino.ino, src_objnum);
2463                 ceph_oid_init(&dst_oid);
2464                 ceph_oid_printf(&dst_oid, "%llx.%08llx",
2465                                 dst_ci->i_vino.ino, dst_objnum);
2466                 /* Do an object remote copy */
2467                 req = ceph_alloc_copyfrom_request(osdc, src_ci->i_vino.snap,
2468                                                   &src_oid, &src_oloc,
2469                                                   &dst_oid, &dst_oloc,
2470                                                   dst_ci->i_truncate_seq,
2471                                                   dst_ci->i_truncate_size);
2472                 if (IS_ERR(req))
2473                         ret = PTR_ERR(req);
2474                 else {
2475                         ceph_osdc_start_request(osdc, req);
2476                         ret = ceph_osdc_wait_request(osdc, req);
2477                         ceph_update_copyfrom_metrics(&fsc->mdsc->metric,
2478                                                      req->r_start_latency,
2479                                                      req->r_end_latency,
2480                                                      object_size, ret);
2481                         ceph_osdc_put_request(req);
2482                 }
2483                 if (ret) {
2484                         if (ret == -EOPNOTSUPP) {
2485                                 fsc->have_copy_from2 = false;
2486                                 pr_notice("OSDs don't support copy-from2; disabling copy offload\n");
2487                         }
2488                         dout("ceph_osdc_copy_from returned %d\n", ret);
2489                         if (!bytes)
2490                                 bytes = ret;
2491                         goto out;
2492                 }
2493                 len -= object_size;
2494                 bytes += object_size;
2495                 *src_off += object_size;
2496                 *dst_off += object_size;
2497         }
2498
2499 out:
2500         ceph_oloc_destroy(&src_oloc);
2501         ceph_oloc_destroy(&dst_oloc);
2502         return bytes;
2503 }
2504
2505 static ssize_t __ceph_copy_file_range(struct file *src_file, loff_t src_off,
2506                                       struct file *dst_file, loff_t dst_off,
2507                                       size_t len, unsigned int flags)
2508 {
2509         struct inode *src_inode = file_inode(src_file);
2510         struct inode *dst_inode = file_inode(dst_file);
2511         struct ceph_inode_info *src_ci = ceph_inode(src_inode);
2512         struct ceph_inode_info *dst_ci = ceph_inode(dst_inode);
2513         struct ceph_cap_flush *prealloc_cf;
2514         struct ceph_fs_client *src_fsc = ceph_inode_to_client(src_inode);
2515         loff_t size;
2516         ssize_t ret = -EIO, bytes;
2517         u64 src_objnum, dst_objnum, src_objoff, dst_objoff;
2518         u32 src_objlen, dst_objlen;
2519         int src_got = 0, dst_got = 0, err, dirty;
2520
2521         if (src_inode->i_sb != dst_inode->i_sb) {
2522                 struct ceph_fs_client *dst_fsc = ceph_inode_to_client(dst_inode);
2523
2524                 if (ceph_fsid_compare(&src_fsc->client->fsid,
2525                                       &dst_fsc->client->fsid)) {
2526                         dout("Copying files across clusters: src: %pU dst: %pU\n",
2527                              &src_fsc->client->fsid, &dst_fsc->client->fsid);
2528                         return -EXDEV;
2529                 }
2530         }
2531         if (ceph_snap(dst_inode) != CEPH_NOSNAP)
2532                 return -EROFS;
2533
2534         /*
2535          * Some of the checks below will return -EOPNOTSUPP, which will force a
2536          * fallback to the default VFS copy_file_range implementation.  This is
2537          * desirable in several cases (for ex, the 'len' is smaller than the
2538          * size of the objects, or in cases where that would be more
2539          * efficient).
2540          */
2541
2542         if (ceph_test_mount_opt(src_fsc, NOCOPYFROM))
2543                 return -EOPNOTSUPP;
2544
2545         if (!src_fsc->have_copy_from2)
2546                 return -EOPNOTSUPP;
2547
2548         /*
2549          * Striped file layouts require that we copy partial objects, but the
2550          * OSD copy-from operation only supports full-object copies.  Limit
2551          * this to non-striped file layouts for now.
2552          */
2553         if ((src_ci->i_layout.stripe_unit != dst_ci->i_layout.stripe_unit) ||
2554             (src_ci->i_layout.stripe_count != 1) ||
2555             (dst_ci->i_layout.stripe_count != 1) ||
2556             (src_ci->i_layout.object_size != dst_ci->i_layout.object_size)) {
2557                 dout("Invalid src/dst files layout\n");
2558                 return -EOPNOTSUPP;
2559         }
2560
2561         /* Every encrypted inode gets its own key, so we can't offload them */
2562         if (IS_ENCRYPTED(src_inode) || IS_ENCRYPTED(dst_inode))
2563                 return -EOPNOTSUPP;
2564
2565         if (len < src_ci->i_layout.object_size)
2566                 return -EOPNOTSUPP; /* no remote copy will be done */
2567
2568         prealloc_cf = ceph_alloc_cap_flush();
2569         if (!prealloc_cf)
2570                 return -ENOMEM;
2571
2572         /* Start by sync'ing the source and destination files */
2573         ret = file_write_and_wait_range(src_file, src_off, (src_off + len));
2574         if (ret < 0) {
2575                 dout("failed to write src file (%zd)\n", ret);
2576                 goto out;
2577         }
2578         ret = file_write_and_wait_range(dst_file, dst_off, (dst_off + len));
2579         if (ret < 0) {
2580                 dout("failed to write dst file (%zd)\n", ret);
2581                 goto out;
2582         }
2583
2584         /*
2585          * We need FILE_WR caps for dst_ci and FILE_RD for src_ci as other
2586          * clients may have dirty data in their caches.  And OSDs know nothing
2587          * about caps, so they can't safely do the remote object copies.
2588          */
2589         err = get_rd_wr_caps(src_file, &src_got,
2590                              dst_file, (dst_off + len), &dst_got);
2591         if (err < 0) {
2592                 dout("get_rd_wr_caps returned %d\n", err);
2593                 ret = -EOPNOTSUPP;
2594                 goto out;
2595         }
2596
2597         ret = is_file_size_ok(src_inode, dst_inode, src_off, dst_off, len);
2598         if (ret < 0)
2599                 goto out_caps;
2600
2601         /* Drop dst file cached pages */
2602         ceph_fscache_invalidate(dst_inode, false);
2603         ret = invalidate_inode_pages2_range(dst_inode->i_mapping,
2604                                             dst_off >> PAGE_SHIFT,
2605                                             (dst_off + len) >> PAGE_SHIFT);
2606         if (ret < 0) {
2607                 dout("Failed to invalidate inode pages (%zd)\n", ret);
2608                 ret = 0; /* XXX */
2609         }
2610         ceph_calc_file_object_mapping(&src_ci->i_layout, src_off,
2611                                       src_ci->i_layout.object_size,
2612                                       &src_objnum, &src_objoff, &src_objlen);
2613         ceph_calc_file_object_mapping(&dst_ci->i_layout, dst_off,
2614                                       dst_ci->i_layout.object_size,
2615                                       &dst_objnum, &dst_objoff, &dst_objlen);
2616         /* object-level offsets need to the same */
2617         if (src_objoff != dst_objoff) {
2618                 ret = -EOPNOTSUPP;
2619                 goto out_caps;
2620         }
2621
2622         /*
2623          * Do a manual copy if the object offset isn't object aligned.
2624          * 'src_objlen' contains the bytes left until the end of the object,
2625          * starting at the src_off
2626          */
2627         if (src_objoff) {
2628                 dout("Initial partial copy of %u bytes\n", src_objlen);
2629
2630                 /*
2631                  * we need to temporarily drop all caps as we'll be calling
2632                  * {read,write}_iter, which will get caps again.
2633                  */
2634                 put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2635                 ret = do_splice_direct(src_file, &src_off, dst_file,
2636                                        &dst_off, src_objlen, flags);
2637                 /* Abort on short copies or on error */
2638                 if (ret < src_objlen) {
2639                         dout("Failed partial copy (%zd)\n", ret);
2640                         goto out;
2641                 }
2642                 len -= ret;
2643                 err = get_rd_wr_caps(src_file, &src_got,
2644                                      dst_file, (dst_off + len), &dst_got);
2645                 if (err < 0)
2646                         goto out;
2647                 err = is_file_size_ok(src_inode, dst_inode,
2648                                       src_off, dst_off, len);
2649                 if (err < 0)
2650                         goto out_caps;
2651         }
2652
2653         size = i_size_read(dst_inode);
2654         bytes = ceph_do_objects_copy(src_ci, &src_off, dst_ci, &dst_off,
2655                                      src_fsc, len, flags);
2656         if (bytes <= 0) {
2657                 if (!ret)
2658                         ret = bytes;
2659                 goto out_caps;
2660         }
2661         dout("Copied %zu bytes out of %zu\n", bytes, len);
2662         len -= bytes;
2663         ret += bytes;
2664
2665         file_update_time(dst_file);
2666         inode_inc_iversion_raw(dst_inode);
2667
2668         if (dst_off > size) {
2669                 /* Let the MDS know about dst file size change */
2670                 if (ceph_inode_set_size(dst_inode, dst_off) ||
2671                     ceph_quota_is_max_bytes_approaching(dst_inode, dst_off))
2672                         ceph_check_caps(dst_ci, CHECK_CAPS_AUTHONLY | CHECK_CAPS_FLUSH);
2673         }
2674         /* Mark Fw dirty */
2675         spin_lock(&dst_ci->i_ceph_lock);
2676         dirty = __ceph_mark_dirty_caps(dst_ci, CEPH_CAP_FILE_WR, &prealloc_cf);
2677         spin_unlock(&dst_ci->i_ceph_lock);
2678         if (dirty)
2679                 __mark_inode_dirty(dst_inode, dirty);
2680
2681 out_caps:
2682         put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2683
2684         /*
2685          * Do the final manual copy if we still have some bytes left, unless
2686          * there were errors in remote object copies (len >= object_size).
2687          */
2688         if (len && (len < src_ci->i_layout.object_size)) {
2689                 dout("Final partial copy of %zu bytes\n", len);
2690                 bytes = do_splice_direct(src_file, &src_off, dst_file,
2691                                          &dst_off, len, flags);
2692                 if (bytes > 0)
2693                         ret += bytes;
2694                 else
2695                         dout("Failed partial copy (%zd)\n", bytes);
2696         }
2697
2698 out:
2699         ceph_free_cap_flush(prealloc_cf);
2700
2701         return ret;
2702 }
2703
2704 static ssize_t ceph_copy_file_range(struct file *src_file, loff_t src_off,
2705                                     struct file *dst_file, loff_t dst_off,
2706                                     size_t len, unsigned int flags)
2707 {
2708         ssize_t ret;
2709
2710         ret = __ceph_copy_file_range(src_file, src_off, dst_file, dst_off,
2711                                      len, flags);
2712
2713         if (ret == -EOPNOTSUPP || ret == -EXDEV)
2714                 ret = generic_copy_file_range(src_file, src_off, dst_file,
2715                                               dst_off, len, flags);
2716         return ret;
2717 }
2718
2719 const struct file_operations ceph_file_fops = {
2720         .open = ceph_open,
2721         .release = ceph_release,
2722         .llseek = ceph_llseek,
2723         .read_iter = ceph_read_iter,
2724         .write_iter = ceph_write_iter,
2725         .mmap = ceph_mmap,
2726         .fsync = ceph_fsync,
2727         .lock = ceph_lock,
2728         .setlease = simple_nosetlease,
2729         .flock = ceph_flock,
2730         .splice_read = ceph_splice_read,
2731         .splice_write = iter_file_splice_write,
2732         .unlocked_ioctl = ceph_ioctl,
2733         .compat_ioctl = compat_ptr_ioctl,
2734         .fallocate      = ceph_fallocate,
2735         .copy_file_range = ceph_copy_file_range,
2736 };