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