fs/ceph/file.c

   1 #include <linux/ceph/ceph_debug.h>
   2
   3 #include <linux/module.h>
   4 #include <linux/sched.h>
   5 #include <linux/slab.h>
   6 #include <linux/file.h>
   7 #include <linux/namei.h>
   8 #include <linux/writeback.h>
   9
  10 #include "super.h"
  11 #include "mds_client.h"
  12
  13 /*
  14  * Ceph file operations
  15  *
  16  * Implement basic open/close functionality, and implement
  17  * read/write.
  18  *
  19  * We implement three modes of file I/O:
  20  *  - buffered uses the generic_file_aio_{read,write} helpers
  21  *
  22  *  - synchronous is used when there is multi-client read/write
  23  *    sharing, avoids the page cache, and synchronously waits for an
  24  *    ack from the OSD.
  25  *
  26  *  - direct io takes the variant of the sync path that references
  27  *    user pages directly.
  28  *
  29  * fsync() flushes and waits on dirty pages, but just queues metadata
  30  * for writeback: since the MDS can recover size and mtime there is no
  31  * need to wait for MDS acknowledgement.
  32  */
  33
  34
  35 /*
  36  * Prepare an open request.  Preallocate ceph_cap to avoid an
  37  * inopportune ENOMEM later.
  38  */
  39 static struct ceph_mds_request *
  40 prepare_open_request(struct super_block *sb, int flags, int create_mode)
  41 {
  42         struct ceph_fs_client *fsc = ceph_sb_to_client(sb);
  43         struct ceph_mds_client *mdsc = fsc->mdsc;
  44         struct ceph_mds_request *req;
  45         int want_auth = USE_ANY_MDS;
  46         int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
  47
  48         if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
  49                 want_auth = USE_AUTH_MDS;
  50
  51         req = ceph_mdsc_create_request(mdsc, op, want_auth);
  52         if (IS_ERR(req))
  53                 goto out;
  54         req->r_fmode = ceph_flags_to_mode(flags);
  55         req->r_args.open.flags = cpu_to_le32(flags);
  56         req->r_args.open.mode = cpu_to_le32(create_mode);
  57         req->r_args.open.preferred = cpu_to_le32(-1);
  58 out:
  59         return req;
  60 }
  61
  62 /*
  63  * initialize private struct file data.
  64  * if we fail, clean up by dropping fmode reference on the ceph_inode
  65  */
  66 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
  67 {
  68         struct ceph_file_info *cf;
  69         int ret = 0;
  70
  71         switch (inode->i_mode & S_IFMT) {
  72         case S_IFREG:
  73         case S_IFDIR:
  74                 dout("init_file %p %p 0%o (regular)\n", inode, file,
  75                      inode->i_mode);
  76                 cf = kmem_cache_alloc(ceph_file_cachep, GFP_NOFS | __GFP_ZERO);
  77                 if (cf == NULL) {
  78                         ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
  79                         return -ENOMEM;
  80                 }
  81                 cf->fmode = fmode;
  82                 cf->next_offset = 2;
  83                 file->private_data = cf;
  84                 BUG_ON(inode->i_fop->release != ceph_release);
  85                 break;
  86
  87         case S_IFLNK:
  88                 dout("init_file %p %p 0%o (symlink)\n", inode, file,
  89                      inode->i_mode);
  90                 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
  91                 break;
  92
  93         default:
  94                 dout("init_file %p %p 0%o (special)\n", inode, file,
  95                      inode->i_mode);
  96                 /*
  97                  * we need to drop the open ref now, since we don't
  98                  * have .release set to ceph_release.
  99                  */
 100                 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
 101                 BUG_ON(inode->i_fop->release == ceph_release);
 102
 103                 /* call the proper open fop */
 104                 ret = inode->i_fop->open(inode, file);
 105         }
 106         return ret;
 107 }
 108
 109 /*
 110  * If the filp already has private_data, that means the file was
 111  * already opened by intent during lookup, and we do nothing.
 112  *
 113  * If we already have the requisite capabilities, we can satisfy
 114  * the open request locally (no need to request new caps from the
 115  * MDS).  We do, however, need to inform the MDS (asynchronously)
 116  * if our wanted caps set expands.
 117  */
 118 int ceph_open(struct inode *inode, struct file *file)
 119 {
 120         struct ceph_inode_info *ci = ceph_inode(inode);
 121         struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
 122         struct ceph_mds_client *mdsc = fsc->mdsc;
 123         struct ceph_mds_request *req;
 124         struct ceph_file_info *cf = file->private_data;
 125         struct inode *parent_inode = file->f_dentry->d_parent->d_inode;
 126         int err;
 127         int flags, fmode, wanted;
 128
 129         if (cf) {
 130                 dout("open file %p is already opened\n", file);
 131                 return 0;
 132         }
 133
 134         /* filter out O_CREAT|O_EXCL; vfs did that already.  yuck. */
 135         flags = file->f_flags & ~(O_CREAT|O_EXCL);
 136         if (S_ISDIR(inode->i_mode))
 137                 flags = O_DIRECTORY;  /* mds likes to know */
 138
 139         dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
 140              ceph_vinop(inode), file, flags, file->f_flags);
 141         fmode = ceph_flags_to_mode(flags);
 142         wanted = ceph_caps_for_mode(fmode);
 143
 144         /* snapped files are read-only */
 145         if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
 146                 return -EROFS;
 147
 148         /* trivially open snapdir */
 149         if (ceph_snap(inode) == CEPH_SNAPDIR) {
 150                 spin_lock(&inode->i_lock);
 151                 __ceph_get_fmode(ci, fmode);
 152                 spin_unlock(&inode->i_lock);
 153                 return ceph_init_file(inode, file, fmode);
 154         }
 155
 156         /*
 157          * No need to block if we have any caps.  Update wanted set
 158          * asynchronously.
 159          */
 160         spin_lock(&inode->i_lock);
 161         if (__ceph_is_any_real_caps(ci)) {
 162                 int mds_wanted = __ceph_caps_mds_wanted(ci);
 163                 int issued = __ceph_caps_issued(ci, NULL);
 164
 165                 dout("open %p fmode %d want %s issued %s using existing\n",
 166                      inode, fmode, ceph_cap_string(wanted),
 167                      ceph_cap_string(issued));
 168                 __ceph_get_fmode(ci, fmode);
 169                 spin_unlock(&inode->i_lock);
 170
 171                 /* adjust wanted? */
 172                 if ((issued & wanted) != wanted &&
 173                     (mds_wanted & wanted) != wanted &&
 174                     ceph_snap(inode) != CEPH_SNAPDIR)
 175                         ceph_check_caps(ci, 0, NULL);
 176
 177                 return ceph_init_file(inode, file, fmode);
 178         } else if (ceph_snap(inode) != CEPH_NOSNAP &&
 179                    (ci->i_snap_caps & wanted) == wanted) {
 180                 __ceph_get_fmode(ci, fmode);
 181                 spin_unlock(&inode->i_lock);
 182                 return ceph_init_file(inode, file, fmode);
 183         }
 184         spin_unlock(&inode->i_lock);
 185
 186         dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
 187         req = prepare_open_request(inode->i_sb, flags, 0);
 188         if (IS_ERR(req)) {
 189                 err = PTR_ERR(req);
 190                 goto out;
 191         }
 192         req->r_inode = igrab(inode);
 193         req->r_num_caps = 1;
 194         err = ceph_mdsc_do_request(mdsc, parent_inode, req);
 195         if (!err)
 196                 err = ceph_init_file(inode, file, req->r_fmode);
 197         ceph_mdsc_put_request(req);
 198         dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
 199 out:
 200         return err;
 201 }
 202
 203
 204 /*
 205  * Do a lookup + open with a single request.
 206  *
 207  * If this succeeds, but some subsequent check in the vfs
 208  * may_open() fails, the struct *file gets cleaned up (i.e.
 209  * ceph_release gets called).  So fear not!
 210  */
 211 /*
 212  * flags
 213  *  path_lookup_open   -> LOOKUP_OPEN
 214  *  path_lookup_create -> LOOKUP_OPEN|LOOKUP_CREATE
 215  */
 216 struct dentry *ceph_lookup_open(struct inode *dir, struct dentry *dentry,
 217                                 struct nameidata *nd, int mode,
 218                                 int locked_dir)
 219 {
 220         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 221         struct ceph_mds_client *mdsc = fsc->mdsc;
 222         struct file *file = nd->intent.open.file;
 223         struct inode *parent_inode = get_dentry_parent_inode(file->f_dentry);
 224         struct ceph_mds_request *req;
 225         int err;
 226         int flags = nd->intent.open.flags - 1;  /* silly vfs! */
 227
 228         dout("ceph_lookup_open dentry %p '%.*s' flags %d mode 0%o\n",
 229              dentry, dentry->d_name.len, dentry->d_name.name, flags, mode);
 230
 231         /* do the open */
 232         req = prepare_open_request(dir->i_sb, flags, mode);
 233         if (IS_ERR(req))
 234                 return ERR_CAST(req);
 235         req->r_dentry = dget(dentry);
 236         req->r_num_caps = 2;
 237         if (flags & O_CREAT) {
 238                 req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
 239                 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 240         }
 241         req->r_locked_dir = dir;           /* caller holds dir->i_mutex */
 242         err = ceph_mdsc_do_request(mdsc, parent_inode, req);
 243         dentry = ceph_finish_lookup(req, dentry, err);
 244         if (!err && (flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
 245                 err = ceph_handle_notrace_create(dir, dentry);
 246         if (!err)
 247                 err = ceph_init_file(req->r_dentry->d_inode, file,
 248                                      req->r_fmode);
 249         ceph_mdsc_put_request(req);
 250         dout("ceph_lookup_open result=%p\n", dentry);
 251         return dentry;
 252 }
 253
 254 int ceph_release(struct inode *inode, struct file *file)
 255 {
 256         struct ceph_inode_info *ci = ceph_inode(inode);
 257         struct ceph_file_info *cf = file->private_data;
 258
 259         dout("release inode %p file %p\n", inode, file);
 260         ceph_put_fmode(ci, cf->fmode);
 261         if (cf->last_readdir)
 262                 ceph_mdsc_put_request(cf->last_readdir);
 263         kfree(cf->last_name);
 264         kfree(cf->dir_info);
 265         dput(cf->dentry);
 266         kmem_cache_free(ceph_file_cachep, cf);
 267
 268         /* wake up anyone waiting for caps on this inode */
 269         wake_up_all(&ci->i_cap_wq);
 270         return 0;
 271 }
 272
 273 /*
 274  * Read a range of bytes striped over one or more objects.  Iterate over
 275  * objects we stripe over.  (That's not atomic, but good enough for now.)
 276  *
 277  * If we get a short result from the OSD, check against i_size; we need to
 278  * only return a short read to the caller if we hit EOF.
 279  */
 280 static int striped_read(struct inode *inode,
 281                         u64 off, u64 len,
 282                         struct page **pages, int num_pages,
 283                         int *checkeof)
 284 {
 285         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 286         struct ceph_inode_info *ci = ceph_inode(inode);
 287         u64 pos, this_len;
 288         int page_off = off & ~PAGE_CACHE_MASK; /* first byte's offset in page */
 289         int left, pages_left;
 290         int read;
 291         struct page **page_pos;
 292         int ret;
 293         bool hit_stripe, was_short;
 294
 295         /*
 296          * we may need to do multiple reads.  not atomic, unfortunately.
 297          */
 298         pos = off;
 299         left = len;
 300         page_pos = pages;
 301         pages_left = num_pages;
 302         read = 0;
 303
 304 more:
 305         this_len = left;
 306         ret = ceph_osdc_readpages(&fsc->client->osdc, ceph_vino(inode),
 307                                   &ci->i_layout, pos, &this_len,
 308                                   ci->i_truncate_seq,
 309                                   ci->i_truncate_size,
 310                                   page_pos, pages_left);
 311         hit_stripe = this_len < left;
 312         was_short = ret >= 0 && ret < this_len;
 313         if (ret == -ENOENT)
 314                 ret = 0;
 315         dout("striped_read %llu~%u (read %u) got %d%s%s\n", pos, left, read,
 316              ret, hit_stripe ? " HITSTRIPE" : "", was_short ? " SHORT" : "");
 317
 318         if (ret > 0) {
 319                 int didpages =
 320                         ((pos & ~PAGE_CACHE_MASK) + ret) >> PAGE_CACHE_SHIFT;
 321
 322                 if (read < pos - off) {
 323                         dout(" zero gap %llu to %llu\n", off + read, pos);
 324                         ceph_zero_page_vector_range(page_off + read,
 325                                                     pos - off - read, pages);
 326                 }
 327                 pos += ret;
 328                 read = pos - off;
 329                 left -= ret;
 330                 page_pos += didpages;
 331                 pages_left -= didpages;
 332
 333                 /* hit stripe? */
 334                 if (left && hit_stripe)
 335                         goto more;
 336         }
 337
 338         if (was_short) {
 339                 /* was original extent fully inside i_size? */
 340                 if (pos + left <= inode->i_size) {
 341                         dout("zero tail\n");
 342                         ceph_zero_page_vector_range(page_off + read, len - read,
 343                                                     pages);
 344                         read = len;
 345                         goto out;
 346                 }
 347
 348                 /* check i_size */
 349                 *checkeof = 1;
 350         }
 351
 352 out:
 353         if (ret >= 0)
 354                 ret = read;
 355         dout("striped_read returns %d\n", ret);
 356         return ret;
 357 }
 358
 359 /*
 360  * Completely synchronous read and write methods.  Direct from __user
 361  * buffer to osd, or directly to user pages (if O_DIRECT).
 362  *
 363  * If the read spans object boundary, just do multiple reads.
 364  */
 365 static ssize_t ceph_sync_read(struct file *file, char __user *data,
 366                               unsigned len, loff_t *poff, int *checkeof)
 367 {
 368         struct inode *inode = file->f_dentry->d_inode;
 369         struct page **pages;
 370         u64 off = *poff;
 371         int num_pages = calc_pages_for(off, len);
 372         int ret;
 373
 374         dout("sync_read on file %p %llu~%u %s\n", file, off, len,
 375              (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
 376
 377         if (file->f_flags & O_DIRECT) {
 378                 pages = ceph_get_direct_page_vector(data, num_pages, off, len);
 379
 380                 /*
 381                  * flush any page cache pages in this range.  this
 382                  * will make concurrent normal and O_DIRECT io slow,
 383                  * but it will at least behave sensibly when they are
 384                  * in sequence.
 385                  */
 386         } else {
 387                 pages = ceph_alloc_page_vector(num_pages, GFP_NOFS);
 388         }
 389         if (IS_ERR(pages))
 390                 return PTR_ERR(pages);
 391
 392         ret = filemap_write_and_wait(inode->i_mapping);
 393         if (ret < 0)
 394                 goto done;
 395
 396         ret = striped_read(inode, off, len, pages, num_pages, checkeof);
 397
 398         if (ret >= 0 && (file->f_flags & O_DIRECT) == 0)
 399                 ret = ceph_copy_page_vector_to_user(pages, data, off, ret);
 400         if (ret >= 0)
 401                 *poff = off + ret;
 402
 403 done:
 404         if (file->f_flags & O_DIRECT)
 405                 ceph_put_page_vector(pages, num_pages);
 406         else
 407                 ceph_release_page_vector(pages, num_pages);
 408         dout("sync_read result %d\n", ret);
 409         return ret;
 410 }
 411
 412 /*
 413  * Write commit callback, called if we requested both an ACK and
 414  * ONDISK commit reply from the OSD.
 415  */
 416 static void sync_write_commit(struct ceph_osd_request *req,
 417                               struct ceph_msg *msg)
 418 {
 419         struct ceph_inode_info *ci = ceph_inode(req->r_inode);
 420
 421         dout("sync_write_commit %p tid %llu\n", req, req->r_tid);
 422         spin_lock(&ci->i_unsafe_lock);
 423         list_del_init(&req->r_unsafe_item);
 424         spin_unlock(&ci->i_unsafe_lock);
 425         ceph_put_cap_refs(ci, CEPH_CAP_FILE_WR);
 426 }
 427
 428 /*
 429  * Synchronous write, straight from __user pointer or user pages (if
 430  * O_DIRECT).
 431  *
 432  * If write spans object boundary, just do multiple writes.  (For a
 433  * correct atomic write, we should e.g. take write locks on all
 434  * objects, rollback on failure, etc.)
 435  */
 436 static ssize_t ceph_sync_write(struct file *file, const char __user *data,
 437                                size_t left, loff_t *offset)
 438 {
 439         struct inode *inode = file->f_dentry->d_inode;
 440         struct ceph_inode_info *ci = ceph_inode(inode);
 441         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 442         struct ceph_osd_request *req;
 443         struct page **pages;
 444         int num_pages;
 445         long long unsigned pos;
 446         u64 len;
 447         int written = 0;
 448         int flags;
 449         int do_sync = 0;
 450         int check_caps = 0;
 451         int ret;
 452         struct timespec mtime = CURRENT_TIME;
 453
 454         if (ceph_snap(file->f_dentry->d_inode) != CEPH_NOSNAP)
 455                 return -EROFS;
 456
 457         dout("sync_write on file %p %lld~%u %s\n", file, *offset,
 458              (unsigned)left, (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
 459
 460         if (file->f_flags & O_APPEND)
 461                 pos = i_size_read(inode);
 462         else
 463                 pos = *offset;
 464
 465         ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + left);
 466         if (ret < 0)
 467                 return ret;
 468
 469         ret = invalidate_inode_pages2_range(inode->i_mapping,
 470                                             pos >> PAGE_CACHE_SHIFT,
 471                                             (pos + left) >> PAGE_CACHE_SHIFT);
 472         if (ret < 0)
 473                 dout("invalidate_inode_pages2_range returned %d\n", ret);
 474
 475         flags = CEPH_OSD_FLAG_ORDERSNAP |
 476                 CEPH_OSD_FLAG_ONDISK |
 477                 CEPH_OSD_FLAG_WRITE;
 478         if ((file->f_flags & (O_SYNC|O_DIRECT)) == 0)
 479                 flags |= CEPH_OSD_FLAG_ACK;
 480         else
 481                 do_sync = 1;
 482
 483         /*
 484          * we may need to do multiple writes here if we span an object
 485          * boundary.  this isn't atomic, unfortunately.  :(
 486          */
 487 more:
 488         len = left;
 489         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
 490                                     ceph_vino(inode), pos, &len,
 491                                     CEPH_OSD_OP_WRITE, flags,
 492                                     ci->i_snap_realm->cached_context,
 493                                     do_sync,
 494                                     ci->i_truncate_seq, ci->i_truncate_size,
 495                                     &mtime, false, 2);
 496         if (!req)
 497                 return -ENOMEM;
 498
 499         num_pages = calc_pages_for(pos, len);
 500
 501         if (file->f_flags & O_DIRECT) {
 502                 pages = ceph_get_direct_page_vector(data, num_pages, pos, len);
 503                 if (IS_ERR(pages)) {
 504                         ret = PTR_ERR(pages);
 505                         goto out;
 506                 }
 507
 508                 /*
 509                  * throw out any page cache pages in this range. this
 510                  * may block.
 511                  */
 512                 truncate_inode_pages_range(inode->i_mapping, pos,
 513                                            (pos+len) | (PAGE_CACHE_SIZE-1));
 514         } else {
 515                 pages = ceph_alloc_page_vector(num_pages, GFP_NOFS);
 516                 if (IS_ERR(pages)) {
 517                         ret = PTR_ERR(pages);
 518                         goto out;
 519                 }
 520                 ret = ceph_copy_user_to_page_vector(pages, data, pos, len);
 521                 if (ret < 0) {
 522                         ceph_release_page_vector(pages, num_pages);
 523                         goto out;
 524                 }
 525
 526                 if ((file->f_flags & O_SYNC) == 0) {
 527                         /* get a second commit callback */
 528                         req->r_safe_callback = sync_write_commit;
 529                         req->r_own_pages = 1;
 530                 }
 531         }
 532         req->r_pages = pages;
 533         req->r_num_pages = num_pages;
 534         req->r_inode = inode;
 535
 536         ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
 537         if (!ret) {
 538                 if (req->r_safe_callback) {
 539                         /*
 540                          * Add to inode unsafe list only after we
 541                          * start_request so that a tid has been assigned.
 542                          */
 543                         spin_lock(&ci->i_unsafe_lock);
 544                         list_add(&req->r_unsafe_item, &ci->i_unsafe_writes);
 545                         spin_unlock(&ci->i_unsafe_lock);
 546                         ceph_get_cap_refs(ci, CEPH_CAP_FILE_WR);
 547                 }
 548                 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
 549         }
 550
 551         if (file->f_flags & O_DIRECT)
 552                 ceph_put_page_vector(pages, num_pages);
 553         else if (file->f_flags & O_SYNC)
 554                 ceph_release_page_vector(pages, num_pages);
 555
 556 out:
 557         ceph_osdc_put_request(req);
 558         if (ret == 0) {
 559                 pos += len;
 560                 written += len;
 561                 left -= len;
 562                 if (left)
 563                         goto more;
 564
 565                 ret = written;
 566                 *offset = pos;
 567                 if (pos > i_size_read(inode))
 568                         check_caps = ceph_inode_set_size(inode, pos);
 569                 if (check_caps)
 570                         ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY,
 571                                         NULL);
 572         }
 573         return ret;
 574 }
 575
 576 /*
 577  * Wrap generic_file_aio_read with checks for cap bits on the inode.
 578  * Atomically grab references, so that those bits are not released
 579  * back to the MDS mid-read.
 580  *
 581  * Hmm, the sync read case isn't actually async... should it be?
 582  */
 583 static ssize_t ceph_aio_read(struct kiocb *iocb, const struct iovec *iov,
 584                              unsigned long nr_segs, loff_t pos)
 585 {
 586         struct file *filp = iocb->ki_filp;
 587         struct ceph_file_info *fi = filp->private_data;
 588         loff_t *ppos = &iocb->ki_pos;
 589         size_t len = iov->iov_len;
 590         struct inode *inode = filp->f_dentry->d_inode;
 591         struct ceph_inode_info *ci = ceph_inode(inode);
 592         void __user *base = iov->iov_base;
 593         ssize_t ret;
 594         int want, got = 0;
 595         int checkeof = 0, read = 0;
 596
 597         dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
 598              inode, ceph_vinop(inode), pos, (unsigned)len, inode);
 599 again:
 600         __ceph_do_pending_vmtruncate(inode);
 601         if (fi->fmode & CEPH_FILE_MODE_LAZY)
 602                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
 603         else
 604                 want = CEPH_CAP_FILE_CACHE;
 605         ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, &got, -1);
 606         if (ret < 0)
 607                 goto out;
 608         dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
 609              inode, ceph_vinop(inode), pos, (unsigned)len,
 610              ceph_cap_string(got));
 611
 612         if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
 613             (iocb->ki_filp->f_flags & O_DIRECT) ||
 614             (inode->i_sb->s_flags & MS_SYNCHRONOUS))
 615                 /* hmm, this isn't really async... */
 616                 ret = ceph_sync_read(filp, base, len, ppos, &checkeof);
 617         else
 618                 ret = generic_file_aio_read(iocb, iov, nr_segs, pos);
 619
 620 out:
 621         dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
 622              inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
 623         ceph_put_cap_refs(ci, got);
 624
 625         if (checkeof && ret >= 0) {
 626                 int statret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE);
 627
 628                 /* hit EOF or hole? */
 629                 if (statret == 0 && *ppos < inode->i_size) {
 630                         dout("aio_read sync_read hit hole, reading more\n");
 631                         read += ret;
 632                         base += ret;
 633                         len -= ret;
 634                         checkeof = 0;
 635                         goto again;
 636                 }
 637         }
 638         if (ret >= 0)
 639                 ret += read;
 640
 641         return ret;
 642 }
 643
 644 /*
 645  * Take cap references to avoid releasing caps to MDS mid-write.
 646  *
 647  * If we are synchronous, and write with an old snap context, the OSD
 648  * may return EOLDSNAPC.  In that case, retry the write.. _after_
 649  * dropping our cap refs and allowing the pending snap to logically
 650  * complete _before_ this write occurs.
 651  *
 652  * If we are near ENOSPC, write synchronously.
 653  */
 654 static ssize_t ceph_aio_write(struct kiocb *iocb, const struct iovec *iov,
 655                        unsigned long nr_segs, loff_t pos)
 656 {
 657         struct file *file = iocb->ki_filp;
 658         struct ceph_file_info *fi = file->private_data;
 659         struct inode *inode = file->f_dentry->d_inode;
 660         struct ceph_inode_info *ci = ceph_inode(inode);
 661         struct ceph_osd_client *osdc =
 662                 &ceph_sb_to_client(inode->i_sb)->client->osdc;
 663         loff_t endoff = pos + iov->iov_len;
 664         int want, got = 0;
 665         int ret, err;
 666
 667         if (ceph_snap(inode) != CEPH_NOSNAP)
 668                 return -EROFS;
 669
 670 retry_snap:
 671         if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL))
 672                 return -ENOSPC;
 673         __ceph_do_pending_vmtruncate(inode);
 674         dout("aio_write %p %llx.%llx %llu~%u getting caps. i_size %llu\n",
 675              inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len,
 676              inode->i_size);
 677         if (fi->fmode & CEPH_FILE_MODE_LAZY)
 678                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
 679         else
 680                 want = CEPH_CAP_FILE_BUFFER;
 681         ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, &got, endoff);
 682         if (ret < 0)
 683                 goto out;
 684
 685         dout("aio_write %p %llx.%llx %llu~%u  got cap refs on %s\n",
 686              inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len,
 687              ceph_cap_string(got));
 688
 689         if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
 690             (iocb->ki_filp->f_flags & O_DIRECT) ||
 691             (inode->i_sb->s_flags & MS_SYNCHRONOUS)) {
 692                 ret = ceph_sync_write(file, iov->iov_base, iov->iov_len,
 693                         &iocb->ki_pos);
 694         } else {
 695                 ret = generic_file_aio_write(iocb, iov, nr_segs, pos);
 696
 697                 if ((ret >= 0 || ret == -EIOCBQUEUED) &&
 698                     ((file->f_flags & O_SYNC) || IS_SYNC(file->f_mapping->host)
 699                      || ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_NEARFULL))) {
 700                         err = vfs_fsync_range(file, pos, pos + ret - 1, 1);
 701                         if (err < 0)
 702                                 ret = err;
 703                 }
 704         }
 705         if (ret >= 0) {
 706                 spin_lock(&inode->i_lock);
 707                 __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR);
 708                 spin_unlock(&inode->i_lock);
 709         }
 710
 711 out:
 712         dout("aio_write %p %llx.%llx %llu~%u  dropping cap refs on %s\n",
 713              inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len,
 714              ceph_cap_string(got));
 715         ceph_put_cap_refs(ci, got);
 716
 717         if (ret == -EOLDSNAPC) {
 718                 dout("aio_write %p %llx.%llx %llu~%u got EOLDSNAPC, retrying\n",
 719                      inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len);
 720                 goto retry_snap;
 721         }
 722
 723         return ret;
 724 }
 725
 726 /*
 727  * llseek.  be sure to verify file size on SEEK_END.
 728  */
 729 static loff_t ceph_llseek(struct file *file, loff_t offset, int origin)
 730 {
 731         struct inode *inode = file->f_mapping->host;
 732         int ret;
 733
 734         mutex_lock(&inode->i_mutex);
 735         __ceph_do_pending_vmtruncate(inode);
 736         switch (origin) {
 737         case SEEK_END:
 738                 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE);
 739                 if (ret < 0) {
 740                         offset = ret;
 741                         goto out;
 742                 }
 743                 offset += inode->i_size;
 744                 break;
 745         case SEEK_CUR:
 746                 /*
 747                  * Here we special-case the lseek(fd, 0, SEEK_CUR)
 748                  * position-querying operation.  Avoid rewriting the "same"
 749                  * f_pos value back to the file because a concurrent read(),
 750                  * write() or lseek() might have altered it
 751                  */
 752                 if (offset == 0) {
 753                         offset = file->f_pos;
 754                         goto out;
 755                 }
 756                 offset += file->f_pos;
 757                 break;
 758         }
 759
 760         if (offset < 0 || offset > inode->i_sb->s_maxbytes) {
 761                 offset = -EINVAL;
 762                 goto out;
 763         }
 764
 765         /* Special lock needed here? */
 766         if (offset != file->f_pos) {
 767                 file->f_pos = offset;
 768                 file->f_version = 0;
 769         }
 770
 771 out:
 772         mutex_unlock(&inode->i_mutex);
 773         return offset;
 774 }
 775
 776 const struct file_operations ceph_file_fops = {
 777         .open = ceph_open,
 778         .release = ceph_release,
 779         .llseek = ceph_llseek,
 780         .read = do_sync_read,
 781         .write = do_sync_write,
 782         .aio_read = ceph_aio_read,
 783         .aio_write = ceph_aio_write,
 784         .mmap = ceph_mmap,
 785         .fsync = ceph_fsync,
 786         .lock = ceph_lock,
 787         .flock = ceph_flock,
 788         .splice_read = generic_file_splice_read,
 789         .splice_write = generic_file_splice_write,
 790         .unlocked_ioctl = ceph_ioctl,
 791         .compat_ioctl   = ceph_ioctl,
 792 };
 793