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