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 out:
  58         return req;
  59 }
  60
  61 /*
  62  * initialize private struct file data.
  63  * if we fail, clean up by dropping fmode reference on the ceph_inode
  64  */
  65 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
  66 {
  67         struct ceph_file_info *cf;
  68         int ret = 0;
  69
  70         switch (inode->i_mode & S_IFMT) {
  71         case S_IFREG:
  72         case S_IFDIR:
  73                 dout("init_file %p %p 0%o (regular)\n", inode, file,
  74                      inode->i_mode);
  75                 cf = kmem_cache_alloc(ceph_file_cachep, GFP_NOFS | __GFP_ZERO);
  76                 if (cf == NULL) {
  77                         ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
  78                         return -ENOMEM;
  79                 }
  80                 cf->fmode = fmode;
  81                 cf->next_offset = 2;
  82                 file->private_data = cf;
  83                 BUG_ON(inode->i_fop->release != ceph_release);
  84                 break;
  85
  86         case S_IFLNK:
  87                 dout("init_file %p %p 0%o (symlink)\n", inode, file,
  88                      inode->i_mode);
  89                 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
  90                 break;
  91
  92         default:
  93                 dout("init_file %p %p 0%o (special)\n", inode, file,
  94                      inode->i_mode);
  95                 /*
  96                  * we need to drop the open ref now, since we don't
  97                  * have .release set to ceph_release.
  98                  */
  99                 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
 100                 BUG_ON(inode->i_fop->release == ceph_release);
 101
 102                 /* call the proper open fop */
 103                 ret = inode->i_fop->open(inode, file);
 104         }
 105         return ret;
 106 }
 107
 108 /*
 109  * If the filp already has private_data, that means the file was
 110  * already opened by intent during lookup, and we do nothing.
 111  *
 112  * If we already have the requisite capabilities, we can satisfy
 113  * the open request locally (no need to request new caps from the
 114  * MDS).  We do, however, need to inform the MDS (asynchronously)
 115  * if our wanted caps set expands.
 116  */
 117 int ceph_open(struct inode *inode, struct file *file)
 118 {
 119         struct ceph_inode_info *ci = ceph_inode(inode);
 120         struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
 121         struct ceph_mds_client *mdsc = fsc->mdsc;
 122         struct ceph_mds_request *req;
 123         struct ceph_file_info *cf = file->private_data;
 124         struct inode *parent_inode = NULL;
 125         int err;
 126         int flags, fmode, wanted;
 127
 128         if (cf) {
 129                 dout("open file %p is already opened\n", file);
 130                 return 0;
 131         }
 132
 133         /* filter out O_CREAT|O_EXCL; vfs did that already.  yuck. */
 134         flags = file->f_flags & ~(O_CREAT|O_EXCL);
 135         if (S_ISDIR(inode->i_mode))
 136                 flags = O_DIRECTORY;  /* mds likes to know */
 137
 138         dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
 139              ceph_vinop(inode), file, flags, file->f_flags);
 140         fmode = ceph_flags_to_mode(flags);
 141         wanted = ceph_caps_for_mode(fmode);
 142
 143         /* snapped files are read-only */
 144         if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
 145                 return -EROFS;
 146
 147         /* trivially open snapdir */
 148         if (ceph_snap(inode) == CEPH_SNAPDIR) {
 149                 spin_lock(&ci->i_ceph_lock);
 150                 __ceph_get_fmode(ci, fmode);
 151                 spin_unlock(&ci->i_ceph_lock);
 152                 return ceph_init_file(inode, file, fmode);
 153         }
 154
 155         /*
 156          * No need to block if we have caps on the auth MDS (for
 157          * write) or any MDS (for read).  Update wanted set
 158          * asynchronously.
 159          */
 160         spin_lock(&ci->i_ceph_lock);
 161         if (__ceph_is_any_real_caps(ci) &&
 162             (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
 163                 int mds_wanted = __ceph_caps_mds_wanted(ci);
 164                 int issued = __ceph_caps_issued(ci, NULL);
 165
 166                 dout("open %p fmode %d want %s issued %s using existing\n",
 167                      inode, fmode, ceph_cap_string(wanted),
 168                      ceph_cap_string(issued));
 169                 __ceph_get_fmode(ci, fmode);
 170                 spin_unlock(&ci->i_ceph_lock);
 171
 172                 /* adjust wanted? */
 173                 if ((issued & wanted) != wanted &&
 174                     (mds_wanted & wanted) != wanted &&
 175                     ceph_snap(inode) != CEPH_SNAPDIR)
 176                         ceph_check_caps(ci, 0, NULL);
 177
 178                 return ceph_init_file(inode, file, fmode);
 179         } else if (ceph_snap(inode) != CEPH_NOSNAP &&
 180                    (ci->i_snap_caps & wanted) == wanted) {
 181                 __ceph_get_fmode(ci, fmode);
 182                 spin_unlock(&ci->i_ceph_lock);
 183                 return ceph_init_file(inode, file, fmode);
 184         }
 185         spin_unlock(&ci->i_ceph_lock);
 186
 187         dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
 188         req = prepare_open_request(inode->i_sb, flags, 0);
 189         if (IS_ERR(req)) {
 190                 err = PTR_ERR(req);
 191                 goto out;
 192         }
 193         req->r_inode = inode;
 194         ihold(inode);
 195         req->r_num_caps = 1;
 196         if (flags & (O_CREAT|O_TRUNC))
 197                 parent_inode = ceph_get_dentry_parent_inode(file->f_dentry);
 198         err = ceph_mdsc_do_request(mdsc, parent_inode, req);
 199         iput(parent_inode);
 200         if (!err)
 201                 err = ceph_init_file(inode, file, req->r_fmode);
 202         ceph_mdsc_put_request(req);
 203         dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
 204 out:
 205         return err;
 206 }
 207
 208
 209 /*
 210  * Do a lookup + open with a single request.
 211  *
 212  * If this succeeds, but some subsequent check in the vfs
 213  * may_open() fails, the struct *file gets cleaned up (i.e.
 214  * ceph_release gets called).  So fear not!
 215  */
 216 struct file *ceph_lookup_open(struct inode *dir, struct dentry *dentry,
 217                               struct opendata *od, unsigned flags, umode_t mode)
 218 {
 219         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 220         struct ceph_mds_client *mdsc = fsc->mdsc;
 221         struct file *file = NULL;
 222         struct ceph_mds_request *req;
 223         struct dentry *ret;
 224         int err;
 225
 226         dout("ceph_lookup_open dentry %p '%.*s' flags %d mode 0%o\n",
 227              dentry, dentry->d_name.len, dentry->d_name.name, flags, mode);
 228
 229         /* do the open */
 230         req = prepare_open_request(dir->i_sb, flags, mode);
 231         if (IS_ERR(req))
 232                 return ERR_CAST(req);
 233         req->r_dentry = dget(dentry);
 234         req->r_num_caps = 2;
 235         if (flags & O_CREAT) {
 236                 req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
 237                 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 238         }
 239         req->r_locked_dir = dir;           /* caller holds dir->i_mutex */
 240         err = ceph_mdsc_do_request(mdsc,
 241                                    (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
 242                                    req);
 243         err = ceph_handle_snapdir(req, dentry, err);
 244         if (err)
 245                 goto out;
 246         if ((flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
 247                 err = ceph_handle_notrace_create(dir, dentry);
 248         if (err)
 249                 goto out;
 250         file = finish_open(od, req->r_dentry, ceph_open);
 251         if (IS_ERR(file))
 252                 err = PTR_ERR(file);
 253 out:
 254         ret = ceph_finish_lookup(req, dentry, err);
 255         ceph_mdsc_put_request(req);
 256         dout("ceph_lookup_open result=%p\n", ret);
 257
 258         if (IS_ERR(ret))
 259                 return ERR_CAST(ret);
 260
 261         dput(ret);
 262         return err ? ERR_PTR(err) : file;
 263 }
 264
 265 int ceph_release(struct inode *inode, struct file *file)
 266 {
 267         struct ceph_inode_info *ci = ceph_inode(inode);
 268         struct ceph_file_info *cf = file->private_data;
 269
 270         dout("release inode %p file %p\n", inode, file);
 271         ceph_put_fmode(ci, cf->fmode);
 272         if (cf->last_readdir)
 273                 ceph_mdsc_put_request(cf->last_readdir);
 274         kfree(cf->last_name);
 275         kfree(cf->dir_info);
 276         dput(cf->dentry);
 277         kmem_cache_free(ceph_file_cachep, cf);
 278
 279         /* wake up anyone waiting for caps on this inode */
 280         wake_up_all(&ci->i_cap_wq);
 281         return 0;
 282 }
 283
 284 /*
 285  * Read a range of bytes striped over one or more objects.  Iterate over
 286  * objects we stripe over.  (That's not atomic, but good enough for now.)
 287  *
 288  * If we get a short result from the OSD, check against i_size; we need to
 289  * only return a short read to the caller if we hit EOF.
 290  */
 291 static int striped_read(struct inode *inode,
 292                         u64 off, u64 len,
 293                         struct page **pages, int num_pages,
 294                         int *checkeof, bool o_direct,
 295                         unsigned long buf_align)
 296 {
 297         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 298         struct ceph_inode_info *ci = ceph_inode(inode);
 299         u64 pos, this_len;
 300         int io_align, page_align;
 301         int left, pages_left;
 302         int read;
 303         struct page **page_pos;
 304         int ret;
 305         bool hit_stripe, was_short;
 306
 307         /*
 308          * we may need to do multiple reads.  not atomic, unfortunately.
 309          */
 310         pos = off;
 311         left = len;
 312         page_pos = pages;
 313         pages_left = num_pages;
 314         read = 0;
 315         io_align = off & ~PAGE_MASK;
 316
 317 more:
 318         if (o_direct)
 319                 page_align = (pos - io_align + buf_align) & ~PAGE_MASK;
 320         else
 321                 page_align = pos & ~PAGE_MASK;
 322         this_len = left;
 323         ret = ceph_osdc_readpages(&fsc->client->osdc, ceph_vino(inode),
 324                                   &ci->i_layout, pos, &this_len,
 325                                   ci->i_truncate_seq,
 326                                   ci->i_truncate_size,
 327                                   page_pos, pages_left, page_align);
 328         if (ret == -ENOENT)
 329                 ret = 0;
 330         hit_stripe = this_len < left;
 331         was_short = ret >= 0 && ret < this_len;
 332         dout("striped_read %llu~%u (read %u) got %d%s%s\n", pos, left, read,
 333              ret, hit_stripe ? " HITSTRIPE" : "", was_short ? " SHORT" : "");
 334
 335         if (ret > 0) {
 336                 int didpages = (page_align + ret) >> PAGE_CACHE_SHIFT;
 337
 338                 if (read < pos - off) {
 339                         dout(" zero gap %llu to %llu\n", off + read, pos);
 340                         ceph_zero_page_vector_range(page_align + read,
 341                                                     pos - off - read, pages);
 342                 }
 343                 pos += ret;
 344                 read = pos - off;
 345                 left -= ret;
 346                 page_pos += didpages;
 347                 pages_left -= didpages;
 348
 349                 /* hit stripe? */
 350                 if (left && hit_stripe)
 351                         goto more;
 352         }
 353
 354         if (was_short) {
 355                 /* did we bounce off eof? */
 356                 if (pos + left > inode->i_size)
 357                         *checkeof = 1;
 358
 359                 /* zero trailing bytes (inside i_size) */
 360                 if (left > 0 && pos < inode->i_size) {
 361                         if (pos + left > inode->i_size)
 362                                 left = inode->i_size - pos;
 363
 364                         dout("zero tail %d\n", left);
 365                         ceph_zero_page_vector_range(page_align + read, left,
 366                                                     pages);
 367                         read += left;
 368                 }
 369         }
 370
 371         if (ret >= 0)
 372                 ret = read;
 373         dout("striped_read returns %d\n", ret);
 374         return ret;
 375 }
 376
 377 /*
 378  * Completely synchronous read and write methods.  Direct from __user
 379  * buffer to osd, or directly to user pages (if O_DIRECT).
 380  *
 381  * If the read spans object boundary, just do multiple reads.
 382  */
 383 static ssize_t ceph_sync_read(struct file *file, char __user *data,
 384                               unsigned len, loff_t *poff, int *checkeof)
 385 {
 386         struct inode *inode = file->f_dentry->d_inode;
 387         struct page **pages;
 388         u64 off = *poff;
 389         int num_pages, ret;
 390
 391         dout("sync_read on file %p %llu~%u %s\n", file, off, len,
 392              (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
 393
 394         if (file->f_flags & O_DIRECT) {
 395                 num_pages = calc_pages_for((unsigned long)data, len);
 396                 pages = ceph_get_direct_page_vector(data, num_pages, true);
 397         } else {
 398                 num_pages = calc_pages_for(off, len);
 399                 pages = ceph_alloc_page_vector(num_pages, GFP_NOFS);
 400         }
 401         if (IS_ERR(pages))
 402                 return PTR_ERR(pages);
 403
 404         /*
 405          * flush any page cache pages in this range.  this
 406          * will make concurrent normal and sync io slow,
 407          * but it will at least behave sensibly when they are
 408          * in sequence.
 409          */
 410         ret = filemap_write_and_wait(inode->i_mapping);
 411         if (ret < 0)
 412                 goto done;
 413
 414         ret = striped_read(inode, off, len, pages, num_pages, checkeof,
 415                            file->f_flags & O_DIRECT,
 416                            (unsigned long)data & ~PAGE_MASK);
 417
 418         if (ret >= 0 && (file->f_flags & O_DIRECT) == 0)
 419                 ret = ceph_copy_page_vector_to_user(pages, data, off, ret);
 420         if (ret >= 0)
 421                 *poff = off + ret;
 422
 423 done:
 424         if (file->f_flags & O_DIRECT)
 425                 ceph_put_page_vector(pages, num_pages, true);
 426         else
 427                 ceph_release_page_vector(pages, num_pages);
 428         dout("sync_read result %d\n", ret);
 429         return ret;
 430 }
 431
 432 /*
 433  * Write commit callback, called if we requested both an ACK and
 434  * ONDISK commit reply from the OSD.
 435  */
 436 static void sync_write_commit(struct ceph_osd_request *req,
 437                               struct ceph_msg *msg)
 438 {
 439         struct ceph_inode_info *ci = ceph_inode(req->r_inode);
 440
 441         dout("sync_write_commit %p tid %llu\n", req, req->r_tid);
 442         spin_lock(&ci->i_unsafe_lock);
 443         list_del_init(&req->r_unsafe_item);
 444         spin_unlock(&ci->i_unsafe_lock);
 445         ceph_put_cap_refs(ci, CEPH_CAP_FILE_WR);
 446 }
 447
 448 /*
 449  * Synchronous write, straight from __user pointer or user pages (if
 450  * O_DIRECT).
 451  *
 452  * If write spans object boundary, just do multiple writes.  (For a
 453  * correct atomic write, we should e.g. take write locks on all
 454  * objects, rollback on failure, etc.)
 455  */
 456 static ssize_t ceph_sync_write(struct file *file, const char __user *data,
 457                                size_t left, loff_t *offset)
 458 {
 459         struct inode *inode = file->f_dentry->d_inode;
 460         struct ceph_inode_info *ci = ceph_inode(inode);
 461         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 462         struct ceph_osd_request *req;
 463         struct page **pages;
 464         int num_pages;
 465         long long unsigned pos;
 466         u64 len;
 467         int written = 0;
 468         int flags;
 469         int do_sync = 0;
 470         int check_caps = 0;
 471         int page_align, io_align;
 472         unsigned long buf_align;
 473         int ret;
 474         struct timespec mtime = CURRENT_TIME;
 475
 476         if (ceph_snap(file->f_dentry->d_inode) != CEPH_NOSNAP)
 477                 return -EROFS;
 478
 479         dout("sync_write on file %p %lld~%u %s\n", file, *offset,
 480              (unsigned)left, (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
 481
 482         if (file->f_flags & O_APPEND)
 483                 pos = i_size_read(inode);
 484         else
 485                 pos = *offset;
 486
 487         ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + left);
 488         if (ret < 0)
 489                 return ret;
 490
 491         ret = invalidate_inode_pages2_range(inode->i_mapping,
 492                                             pos >> PAGE_CACHE_SHIFT,
 493                                             (pos + left) >> PAGE_CACHE_SHIFT);
 494         if (ret < 0)
 495                 dout("invalidate_inode_pages2_range returned %d\n", ret);
 496
 497         flags = CEPH_OSD_FLAG_ORDERSNAP |
 498                 CEPH_OSD_FLAG_ONDISK |
 499                 CEPH_OSD_FLAG_WRITE;
 500         if ((file->f_flags & (O_SYNC|O_DIRECT)) == 0)
 501                 flags |= CEPH_OSD_FLAG_ACK;
 502         else
 503                 do_sync = 1;
 504
 505         /*
 506          * we may need to do multiple writes here if we span an object
 507          * boundary.  this isn't atomic, unfortunately.  :(
 508          */
 509 more:
 510         io_align = pos & ~PAGE_MASK;
 511         buf_align = (unsigned long)data & ~PAGE_MASK;
 512         len = left;
 513         if (file->f_flags & O_DIRECT) {
 514                 /* write from beginning of first page, regardless of
 515                    io alignment */
 516                 page_align = (pos - io_align + buf_align) & ~PAGE_MASK;
 517                 num_pages = calc_pages_for((unsigned long)data, len);
 518         } else {
 519                 page_align = pos & ~PAGE_MASK;
 520                 num_pages = calc_pages_for(pos, len);
 521         }
 522         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
 523                                     ceph_vino(inode), pos, &len,
 524                                     CEPH_OSD_OP_WRITE, flags,
 525                                     ci->i_snap_realm->cached_context,
 526                                     do_sync,
 527                                     ci->i_truncate_seq, ci->i_truncate_size,
 528                                     &mtime, false, 2, page_align);
 529         if (!req)
 530                 return -ENOMEM;
 531
 532         if (file->f_flags & O_DIRECT) {
 533                 pages = ceph_get_direct_page_vector(data, num_pages, false);
 534                 if (IS_ERR(pages)) {
 535                         ret = PTR_ERR(pages);
 536                         goto out;
 537                 }
 538
 539                 /*
 540                  * throw out any page cache pages in this range. this
 541                  * may block.
 542                  */
 543                 truncate_inode_pages_range(inode->i_mapping, pos,
 544                                            (pos+len) | (PAGE_CACHE_SIZE-1));
 545         } else {
 546                 pages = ceph_alloc_page_vector(num_pages, GFP_NOFS);
 547                 if (IS_ERR(pages)) {
 548                         ret = PTR_ERR(pages);
 549                         goto out;
 550                 }
 551                 ret = ceph_copy_user_to_page_vector(pages, data, pos, len);
 552                 if (ret < 0) {
 553                         ceph_release_page_vector(pages, num_pages);
 554                         goto out;
 555                 }
 556
 557                 if ((file->f_flags & O_SYNC) == 0) {
 558                         /* get a second commit callback */
 559                         req->r_safe_callback = sync_write_commit;
 560                         req->r_own_pages = 1;
 561                 }
 562         }
 563         req->r_pages = pages;
 564         req->r_num_pages = num_pages;
 565         req->r_inode = inode;
 566
 567         ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
 568         if (!ret) {
 569                 if (req->r_safe_callback) {
 570                         /*
 571                          * Add to inode unsafe list only after we
 572                          * start_request so that a tid has been assigned.
 573                          */
 574                         spin_lock(&ci->i_unsafe_lock);
 575                         list_add_tail(&req->r_unsafe_item,
 576                                       &ci->i_unsafe_writes);
 577                         spin_unlock(&ci->i_unsafe_lock);
 578                         ceph_get_cap_refs(ci, CEPH_CAP_FILE_WR);
 579                 }
 580
 581                 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
 582                 if (ret < 0 && req->r_safe_callback) {
 583                         spin_lock(&ci->i_unsafe_lock);
 584                         list_del_init(&req->r_unsafe_item);
 585                         spin_unlock(&ci->i_unsafe_lock);
 586                         ceph_put_cap_refs(ci, CEPH_CAP_FILE_WR);
 587                 }
 588         }
 589
 590         if (file->f_flags & O_DIRECT)
 591                 ceph_put_page_vector(pages, num_pages, false);
 592         else if (file->f_flags & O_SYNC)
 593                 ceph_release_page_vector(pages, num_pages);
 594
 595 out:
 596         ceph_osdc_put_request(req);
 597         if (ret == 0) {
 598                 pos += len;
 599                 written += len;
 600                 left -= len;
 601                 data += written;
 602                 if (left)
 603                         goto more;
 604
 605                 ret = written;
 606                 *offset = pos;
 607                 if (pos > i_size_read(inode))
 608                         check_caps = ceph_inode_set_size(inode, pos);
 609                 if (check_caps)
 610                         ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY,
 611                                         NULL);
 612         }
 613         return ret;
 614 }
 615
 616 /*
 617  * Wrap generic_file_aio_read with checks for cap bits on the inode.
 618  * Atomically grab references, so that those bits are not released
 619  * back to the MDS mid-read.
 620  *
 621  * Hmm, the sync read case isn't actually async... should it be?
 622  */
 623 static ssize_t ceph_aio_read(struct kiocb *iocb, const struct iovec *iov,
 624                              unsigned long nr_segs, loff_t pos)
 625 {
 626         struct file *filp = iocb->ki_filp;
 627         struct ceph_file_info *fi = filp->private_data;
 628         loff_t *ppos = &iocb->ki_pos;
 629         size_t len = iov->iov_len;
 630         struct inode *inode = filp->f_dentry->d_inode;
 631         struct ceph_inode_info *ci = ceph_inode(inode);
 632         void __user *base = iov->iov_base;
 633         ssize_t ret;
 634         int want, got = 0;
 635         int checkeof = 0, read = 0;
 636
 637         dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
 638              inode, ceph_vinop(inode), pos, (unsigned)len, inode);
 639 again:
 640         __ceph_do_pending_vmtruncate(inode);
 641         if (fi->fmode & CEPH_FILE_MODE_LAZY)
 642                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
 643         else
 644                 want = CEPH_CAP_FILE_CACHE;
 645         ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, &got, -1);
 646         if (ret < 0)
 647                 goto out;
 648         dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
 649              inode, ceph_vinop(inode), pos, (unsigned)len,
 650              ceph_cap_string(got));
 651
 652         if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
 653             (iocb->ki_filp->f_flags & O_DIRECT) ||
 654             (inode->i_sb->s_flags & MS_SYNCHRONOUS) ||
 655             (fi->flags & CEPH_F_SYNC))
 656                 /* hmm, this isn't really async... */
 657                 ret = ceph_sync_read(filp, base, len, ppos, &checkeof);
 658         else
 659                 ret = generic_file_aio_read(iocb, iov, nr_segs, pos);
 660
 661 out:
 662         dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
 663              inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
 664         ceph_put_cap_refs(ci, got);
 665
 666         if (checkeof && ret >= 0) {
 667                 int statret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE);
 668
 669                 /* hit EOF or hole? */
 670                 if (statret == 0 && *ppos < inode->i_size) {
 671                         dout("aio_read sync_read hit hole, ppos %lld < size %lld, reading more\n", *ppos, inode->i_size);
 672                         read += ret;
 673                         base += ret;
 674                         len -= ret;
 675                         checkeof = 0;
 676                         goto again;
 677                 }
 678         }
 679         if (ret >= 0)
 680                 ret += read;
 681
 682         return ret;
 683 }
 684
 685 /*
 686  * Take cap references to avoid releasing caps to MDS mid-write.
 687  *
 688  * If we are synchronous, and write with an old snap context, the OSD
 689  * may return EOLDSNAPC.  In that case, retry the write.. _after_
 690  * dropping our cap refs and allowing the pending snap to logically
 691  * complete _before_ this write occurs.
 692  *
 693  * If we are near ENOSPC, write synchronously.
 694  */
 695 static ssize_t ceph_aio_write(struct kiocb *iocb, const struct iovec *iov,
 696                        unsigned long nr_segs, loff_t pos)
 697 {
 698         struct file *file = iocb->ki_filp;
 699         struct ceph_file_info *fi = file->private_data;
 700         struct inode *inode = file->f_dentry->d_inode;
 701         struct ceph_inode_info *ci = ceph_inode(inode);
 702         struct ceph_osd_client *osdc =
 703                 &ceph_sb_to_client(inode->i_sb)->client->osdc;
 704         loff_t endoff = pos + iov->iov_len;
 705         int want, got = 0;
 706         int ret, err;
 707
 708         if (ceph_snap(inode) != CEPH_NOSNAP)
 709                 return -EROFS;
 710
 711 retry_snap:
 712         if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL))
 713                 return -ENOSPC;
 714         __ceph_do_pending_vmtruncate(inode);
 715         dout("aio_write %p %llx.%llx %llu~%u getting caps. i_size %llu\n",
 716              inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len,
 717              inode->i_size);
 718         if (fi->fmode & CEPH_FILE_MODE_LAZY)
 719                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
 720         else
 721                 want = CEPH_CAP_FILE_BUFFER;
 722         ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, &got, endoff);
 723         if (ret < 0)
 724                 goto out_put;
 725
 726         dout("aio_write %p %llx.%llx %llu~%u  got cap refs on %s\n",
 727              inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len,
 728              ceph_cap_string(got));
 729
 730         if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
 731             (iocb->ki_filp->f_flags & O_DIRECT) ||
 732             (inode->i_sb->s_flags & MS_SYNCHRONOUS) ||
 733             (fi->flags & CEPH_F_SYNC)) {
 734                 ret = ceph_sync_write(file, iov->iov_base, iov->iov_len,
 735                         &iocb->ki_pos);
 736         } else {
 737                 /*
 738                  * buffered write; drop Fw early to avoid slow
 739                  * revocation if we get stuck on balance_dirty_pages
 740                  */
 741                 int dirty;
 742
 743                 spin_lock(&ci->i_ceph_lock);
 744                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR);
 745                 spin_unlock(&ci->i_ceph_lock);
 746                 ceph_put_cap_refs(ci, got);
 747
 748                 ret = generic_file_aio_write(iocb, iov, nr_segs, pos);
 749                 if ((ret >= 0 || ret == -EIOCBQUEUED) &&
 750                     ((file->f_flags & O_SYNC) || IS_SYNC(file->f_mapping->host)
 751                      || ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_NEARFULL))) {
 752                         err = vfs_fsync_range(file, pos, pos + ret - 1, 1);
 753                         if (err < 0)
 754                                 ret = err;
 755                 }
 756
 757                 if (dirty)
 758                         __mark_inode_dirty(inode, dirty);
 759                 goto out;
 760         }
 761
 762         if (ret >= 0) {
 763                 int dirty;
 764                 spin_lock(&ci->i_ceph_lock);
 765                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR);
 766                 spin_unlock(&ci->i_ceph_lock);
 767                 if (dirty)
 768                         __mark_inode_dirty(inode, dirty);
 769         }
 770
 771 out_put:
 772         dout("aio_write %p %llx.%llx %llu~%u  dropping cap refs on %s\n",
 773              inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len,
 774              ceph_cap_string(got));
 775         ceph_put_cap_refs(ci, got);
 776
 777 out:
 778         if (ret == -EOLDSNAPC) {
 779                 dout("aio_write %p %llx.%llx %llu~%u got EOLDSNAPC, retrying\n",
 780                      inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len);
 781                 goto retry_snap;
 782         }
 783
 784         return ret;
 785 }
 786
 787 /*
 788  * llseek.  be sure to verify file size on SEEK_END.
 789  */
 790 static loff_t ceph_llseek(struct file *file, loff_t offset, int origin)
 791 {
 792         struct inode *inode = file->f_mapping->host;
 793         int ret;
 794
 795         mutex_lock(&inode->i_mutex);
 796         __ceph_do_pending_vmtruncate(inode);
 797
 798         if (origin == SEEK_END || origin == SEEK_DATA || origin == SEEK_HOLE) {
 799                 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE);
 800                 if (ret < 0) {
 801                         offset = ret;
 802                         goto out;
 803                 }
 804         }
 805
 806         switch (origin) {
 807         case SEEK_END:
 808                 offset += inode->i_size;
 809                 break;
 810         case SEEK_CUR:
 811                 /*
 812                  * Here we special-case the lseek(fd, 0, SEEK_CUR)
 813                  * position-querying operation.  Avoid rewriting the "same"
 814                  * f_pos value back to the file because a concurrent read(),
 815                  * write() or lseek() might have altered it
 816                  */
 817                 if (offset == 0) {
 818                         offset = file->f_pos;
 819                         goto out;
 820                 }
 821                 offset += file->f_pos;
 822                 break;
 823         case SEEK_DATA:
 824                 if (offset >= inode->i_size) {
 825                         ret = -ENXIO;
 826                         goto out;
 827                 }
 828                 break;
 829         case SEEK_HOLE:
 830                 if (offset >= inode->i_size) {
 831                         ret = -ENXIO;
 832                         goto out;
 833                 }
 834                 offset = inode->i_size;
 835                 break;
 836         }
 837
 838         if (offset < 0 || offset > inode->i_sb->s_maxbytes) {
 839                 offset = -EINVAL;
 840                 goto out;
 841         }
 842
 843         /* Special lock needed here? */
 844         if (offset != file->f_pos) {
 845                 file->f_pos = offset;
 846                 file->f_version = 0;
 847         }
 848
 849 out:
 850         mutex_unlock(&inode->i_mutex);
 851         return offset;
 852 }
 853
 854 const struct file_operations ceph_file_fops = {
 855         .open = ceph_open,
 856         .release = ceph_release,
 857         .llseek = ceph_llseek,
 858         .read = do_sync_read,
 859         .write = do_sync_write,
 860         .aio_read = ceph_aio_read,
 861         .aio_write = ceph_aio_write,
 862         .mmap = ceph_mmap,
 863         .fsync = ceph_fsync,
 864         .lock = ceph_lock,
 865         .flock = ceph_flock,
 866         .splice_read = generic_file_splice_read,
 867         .splice_write = generic_file_splice_write,
 868         .unlocked_ioctl = ceph_ioctl,
 869         .compat_ioctl   = ceph_ioctl,
 870 };
 871