fs/ceph/file.c

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