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/mount.h> 8 #include <linux/namei.h> 9 #include <linux/writeback.h> 10 #include <linux/aio.h> 11 #include <linux/falloc.h> 12 13 #include "super.h" 14 #include "mds_client.h" 15 #include "cache.h" 16 17 /* 18 * Ceph file operations 19 * 20 * Implement basic open/close functionality, and implement 21 * read/write. 22 * 23 * We implement three modes of file I/O: 24 * - buffered uses the generic_file_aio_{read,write} helpers 25 * 26 * - synchronous is used when there is multi-client read/write 27 * sharing, avoids the page cache, and synchronously waits for an 28 * ack from the OSD. 29 * 30 * - direct io takes the variant of the sync path that references 31 * user pages directly. 32 * 33 * fsync() flushes and waits on dirty pages, but just queues metadata 34 * for writeback: since the MDS can recover size and mtime there is no 35 * need to wait for MDS acknowledgement. 36 */ 37 38 39 /* 40 * Prepare an open request. Preallocate ceph_cap to avoid an 41 * inopportune ENOMEM later. 42 */ 43 static struct ceph_mds_request * 44 prepare_open_request(struct super_block *sb, int flags, int create_mode) 45 { 46 struct ceph_fs_client *fsc = ceph_sb_to_client(sb); 47 struct ceph_mds_client *mdsc = fsc->mdsc; 48 struct ceph_mds_request *req; 49 int want_auth = USE_ANY_MDS; 50 int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN; 51 52 if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC)) 53 want_auth = USE_AUTH_MDS; 54 55 req = ceph_mdsc_create_request(mdsc, op, want_auth); 56 if (IS_ERR(req)) 57 goto out; 58 req->r_fmode = ceph_flags_to_mode(flags); 59 req->r_args.open.flags = cpu_to_le32(flags); 60 req->r_args.open.mode = cpu_to_le32(create_mode); 61 out: 62 return req; 63 } 64 65 /* 66 * initialize private struct file data. 67 * if we fail, clean up by dropping fmode reference on the ceph_inode 68 */ 69 static int ceph_init_file(struct inode *inode, struct file *file, int fmode) 70 { 71 struct ceph_file_info *cf; 72 int ret = 0; 73 struct ceph_inode_info *ci = ceph_inode(inode); 74 struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb); 75 struct ceph_mds_client *mdsc = fsc->mdsc; 76 77 switch (inode->i_mode & S_IFMT) { 78 case S_IFREG: 79 /* First file open request creates the cookie, we want to keep 80 * this cookie around for the filetime of the inode as not to 81 * have to worry about fscache register / revoke / operation 82 * races. 83 * 84 * Also, if we know the operation is going to invalidate data 85 * (non readonly) just nuke the cache right away. 86 */ 87 ceph_fscache_register_inode_cookie(mdsc->fsc, ci); 88 if ((fmode & CEPH_FILE_MODE_WR)) 89 ceph_fscache_invalidate(inode); 90 case S_IFDIR: 91 dout("init_file %p %p 0%o (regular)\n", inode, file, 92 inode->i_mode); 93 cf = kmem_cache_alloc(ceph_file_cachep, GFP_NOFS | __GFP_ZERO); 94 if (cf == NULL) { 95 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */ 96 return -ENOMEM; 97 } 98 cf->fmode = fmode; 99 cf->next_offset = 2; 100 file->private_data = cf; 101 BUG_ON(inode->i_fop->release != ceph_release); 102 break; 103 104 case S_IFLNK: 105 dout("init_file %p %p 0%o (symlink)\n", inode, file, 106 inode->i_mode); 107 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */ 108 break; 109 110 default: 111 dout("init_file %p %p 0%o (special)\n", inode, file, 112 inode->i_mode); 113 /* 114 * we need to drop the open ref now, since we don't 115 * have .release set to ceph_release. 116 */ 117 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */ 118 BUG_ON(inode->i_fop->release == ceph_release); 119 120 /* call the proper open fop */ 121 ret = inode->i_fop->open(inode, file); 122 } 123 return ret; 124 } 125 126 /* 127 * If we already have the requisite capabilities, we can satisfy 128 * the open request locally (no need to request new caps from the 129 * MDS). We do, however, need to inform the MDS (asynchronously) 130 * if our wanted caps set expands. 131 */ 132 int ceph_open(struct inode *inode, struct file *file) 133 { 134 struct ceph_inode_info *ci = ceph_inode(inode); 135 struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb); 136 struct ceph_mds_client *mdsc = fsc->mdsc; 137 struct ceph_mds_request *req; 138 struct ceph_file_info *cf = file->private_data; 139 struct inode *parent_inode = NULL; 140 int err; 141 int flags, fmode, wanted; 142 143 if (cf) { 144 dout("open file %p is already opened\n", file); 145 return 0; 146 } 147 148 /* filter out O_CREAT|O_EXCL; vfs did that already. yuck. */ 149 flags = file->f_flags & ~(O_CREAT|O_EXCL); 150 if (S_ISDIR(inode->i_mode)) 151 flags = O_DIRECTORY; /* mds likes to know */ 152 153 dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode, 154 ceph_vinop(inode), file, flags, file->f_flags); 155 fmode = ceph_flags_to_mode(flags); 156 wanted = ceph_caps_for_mode(fmode); 157 158 /* snapped files are read-only */ 159 if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE)) 160 return -EROFS; 161 162 /* trivially open snapdir */ 163 if (ceph_snap(inode) == CEPH_SNAPDIR) { 164 spin_lock(&ci->i_ceph_lock); 165 __ceph_get_fmode(ci, fmode); 166 spin_unlock(&ci->i_ceph_lock); 167 return ceph_init_file(inode, file, fmode); 168 } 169 170 /* 171 * No need to block if we have caps on the auth MDS (for 172 * write) or any MDS (for read). Update wanted set 173 * asynchronously. 174 */ 175 spin_lock(&ci->i_ceph_lock); 176 if (__ceph_is_any_real_caps(ci) && 177 (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) { 178 int mds_wanted = __ceph_caps_mds_wanted(ci); 179 int issued = __ceph_caps_issued(ci, NULL); 180 181 dout("open %p fmode %d want %s issued %s using existing\n", 182 inode, fmode, ceph_cap_string(wanted), 183 ceph_cap_string(issued)); 184 __ceph_get_fmode(ci, fmode); 185 spin_unlock(&ci->i_ceph_lock); 186 187 /* adjust wanted? */ 188 if ((issued & wanted) != wanted && 189 (mds_wanted & wanted) != wanted && 190 ceph_snap(inode) != CEPH_SNAPDIR) 191 ceph_check_caps(ci, 0, NULL); 192 193 return ceph_init_file(inode, file, fmode); 194 } else if (ceph_snap(inode) != CEPH_NOSNAP && 195 (ci->i_snap_caps & wanted) == wanted) { 196 __ceph_get_fmode(ci, fmode); 197 spin_unlock(&ci->i_ceph_lock); 198 return ceph_init_file(inode, file, fmode); 199 } 200 201 spin_unlock(&ci->i_ceph_lock); 202 203 dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted)); 204 req = prepare_open_request(inode->i_sb, flags, 0); 205 if (IS_ERR(req)) { 206 err = PTR_ERR(req); 207 goto out; 208 } 209 req->r_inode = inode; 210 ihold(inode); 211 212 req->r_num_caps = 1; 213 if (flags & O_CREAT) 214 parent_inode = ceph_get_dentry_parent_inode(file->f_dentry); 215 err = ceph_mdsc_do_request(mdsc, parent_inode, req); 216 iput(parent_inode); 217 if (!err) 218 err = ceph_init_file(inode, file, req->r_fmode); 219 ceph_mdsc_put_request(req); 220 dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode)); 221 out: 222 return err; 223 } 224 225 226 /* 227 * Do a lookup + open with a single request. If we get a non-existent 228 * file or symlink, return 1 so the VFS can retry. 229 */ 230 int ceph_atomic_open(struct inode *dir, struct dentry *dentry, 231 struct file *file, unsigned flags, umode_t mode, 232 int *opened) 233 { 234 struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb); 235 struct ceph_mds_client *mdsc = fsc->mdsc; 236 struct ceph_mds_request *req; 237 struct dentry *dn; 238 int err; 239 240 dout("atomic_open %p dentry %p '%.*s' %s flags %d mode 0%o\n", 241 dir, dentry, dentry->d_name.len, dentry->d_name.name, 242 d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode); 243 244 if (dentry->d_name.len > NAME_MAX) 245 return -ENAMETOOLONG; 246 247 err = ceph_init_dentry(dentry); 248 if (err < 0) 249 return err; 250 251 /* do the open */ 252 req = prepare_open_request(dir->i_sb, flags, mode); 253 if (IS_ERR(req)) 254 return PTR_ERR(req); 255 req->r_dentry = dget(dentry); 256 req->r_num_caps = 2; 257 if (flags & O_CREAT) { 258 req->r_dentry_drop = CEPH_CAP_FILE_SHARED; 259 req->r_dentry_unless = CEPH_CAP_FILE_EXCL; 260 } 261 req->r_locked_dir = dir; /* caller holds dir->i_mutex */ 262 err = ceph_mdsc_do_request(mdsc, 263 (flags & (O_CREAT|O_TRUNC)) ? dir : NULL, 264 req); 265 if (err) 266 goto out_err; 267 268 err = ceph_handle_snapdir(req, dentry, err); 269 if (err == 0 && (flags & O_CREAT) && !req->r_reply_info.head->is_dentry) 270 err = ceph_handle_notrace_create(dir, dentry); 271 272 if (d_unhashed(dentry)) { 273 dn = ceph_finish_lookup(req, dentry, err); 274 if (IS_ERR(dn)) 275 err = PTR_ERR(dn); 276 } else { 277 /* we were given a hashed negative dentry */ 278 dn = NULL; 279 } 280 if (err) 281 goto out_err; 282 if (dn || dentry->d_inode == NULL || S_ISLNK(dentry->d_inode->i_mode)) { 283 /* make vfs retry on splice, ENOENT, or symlink */ 284 dout("atomic_open finish_no_open on dn %p\n", dn); 285 err = finish_no_open(file, dn); 286 } else { 287 dout("atomic_open finish_open on dn %p\n", dn); 288 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) { 289 ceph_init_acl(dentry, dentry->d_inode, dir); 290 *opened |= FILE_CREATED; 291 } 292 err = finish_open(file, dentry, ceph_open, opened); 293 } 294 out_err: 295 if (!req->r_err && req->r_target_inode) 296 ceph_put_fmode(ceph_inode(req->r_target_inode), req->r_fmode); 297 ceph_mdsc_put_request(req); 298 dout("atomic_open result=%d\n", err); 299 return err; 300 } 301 302 int ceph_release(struct inode *inode, struct file *file) 303 { 304 struct ceph_inode_info *ci = ceph_inode(inode); 305 struct ceph_file_info *cf = file->private_data; 306 307 dout("release inode %p file %p\n", inode, file); 308 ceph_put_fmode(ci, cf->fmode); 309 if (cf->last_readdir) 310 ceph_mdsc_put_request(cf->last_readdir); 311 kfree(cf->last_name); 312 kfree(cf->dir_info); 313 dput(cf->dentry); 314 kmem_cache_free(ceph_file_cachep, cf); 315 316 /* wake up anyone waiting for caps on this inode */ 317 wake_up_all(&ci->i_cap_wq); 318 return 0; 319 } 320 321 /* 322 * Read a range of bytes striped over one or more objects. Iterate over 323 * objects we stripe over. (That's not atomic, but good enough for now.) 324 * 325 * If we get a short result from the OSD, check against i_size; we need to 326 * only return a short read to the caller if we hit EOF. 327 */ 328 static int striped_read(struct inode *inode, 329 u64 off, u64 len, 330 struct page **pages, int num_pages, 331 int *checkeof, bool o_direct, 332 unsigned long buf_align) 333 { 334 struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 335 struct ceph_inode_info *ci = ceph_inode(inode); 336 u64 pos, this_len, left; 337 int io_align, page_align; 338 int pages_left; 339 int read; 340 struct page **page_pos; 341 int ret; 342 bool hit_stripe, was_short; 343 344 /* 345 * we may need to do multiple reads. not atomic, unfortunately. 346 */ 347 pos = off; 348 left = len; 349 page_pos = pages; 350 pages_left = num_pages; 351 read = 0; 352 io_align = off & ~PAGE_MASK; 353 354 more: 355 if (o_direct) 356 page_align = (pos - io_align + buf_align) & ~PAGE_MASK; 357 else 358 page_align = pos & ~PAGE_MASK; 359 this_len = left; 360 ret = ceph_osdc_readpages(&fsc->client->osdc, ceph_vino(inode), 361 &ci->i_layout, pos, &this_len, 362 ci->i_truncate_seq, 363 ci->i_truncate_size, 364 page_pos, pages_left, page_align); 365 if (ret == -ENOENT) 366 ret = 0; 367 hit_stripe = this_len < left; 368 was_short = ret >= 0 && ret < this_len; 369 dout("striped_read %llu~%llu (read %u) got %d%s%s\n", pos, left, read, 370 ret, hit_stripe ? " HITSTRIPE" : "", was_short ? " SHORT" : ""); 371 372 if (ret >= 0) { 373 int didpages; 374 if (was_short && (pos + ret < inode->i_size)) { 375 u64 tmp = min(this_len - ret, 376 inode->i_size - pos - ret); 377 dout(" zero gap %llu to %llu\n", 378 pos + ret, pos + ret + tmp); 379 ceph_zero_page_vector_range(page_align + read + ret, 380 tmp, pages); 381 ret += tmp; 382 } 383 384 didpages = (page_align + ret) >> PAGE_CACHE_SHIFT; 385 pos += ret; 386 read = pos - off; 387 left -= ret; 388 page_pos += didpages; 389 pages_left -= didpages; 390 391 /* hit stripe and need continue*/ 392 if (left && hit_stripe && pos < inode->i_size) 393 goto more; 394 } 395 396 if (read > 0) { 397 ret = read; 398 /* did we bounce off eof? */ 399 if (pos + left > inode->i_size) 400 *checkeof = 1; 401 } 402 403 dout("striped_read returns %d\n", ret); 404 return ret; 405 } 406 407 /* 408 * Completely synchronous read and write methods. Direct from __user 409 * buffer to osd, or directly to user pages (if O_DIRECT). 410 * 411 * If the read spans object boundary, just do multiple reads. 412 */ 413 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *i, 414 int *checkeof) 415 { 416 struct file *file = iocb->ki_filp; 417 struct inode *inode = file_inode(file); 418 struct page **pages; 419 u64 off = iocb->ki_pos; 420 int num_pages, ret; 421 size_t len = i->count; 422 423 dout("sync_read on file %p %llu~%u %s\n", file, off, 424 (unsigned)len, 425 (file->f_flags & O_DIRECT) ? "O_DIRECT" : ""); 426 /* 427 * flush any page cache pages in this range. this 428 * will make concurrent normal and sync io slow, 429 * but it will at least behave sensibly when they are 430 * in sequence. 431 */ 432 ret = filemap_write_and_wait_range(inode->i_mapping, off, 433 off + len); 434 if (ret < 0) 435 return ret; 436 437 if (file->f_flags & O_DIRECT) { 438 while (iov_iter_count(i)) { 439 void __user *data = i->iov[0].iov_base + i->iov_offset; 440 size_t len = i->iov[0].iov_len - i->iov_offset; 441 442 num_pages = calc_pages_for((unsigned long)data, len); 443 pages = ceph_get_direct_page_vector(data, 444 num_pages, true); 445 if (IS_ERR(pages)) 446 return PTR_ERR(pages); 447 448 ret = striped_read(inode, off, len, 449 pages, num_pages, checkeof, 450 1, (unsigned long)data & ~PAGE_MASK); 451 ceph_put_page_vector(pages, num_pages, true); 452 453 if (ret <= 0) 454 break; 455 off += ret; 456 iov_iter_advance(i, ret); 457 if (ret < len) 458 break; 459 } 460 } else { 461 num_pages = calc_pages_for(off, len); 462 pages = ceph_alloc_page_vector(num_pages, GFP_NOFS); 463 if (IS_ERR(pages)) 464 return PTR_ERR(pages); 465 ret = striped_read(inode, off, len, pages, 466 num_pages, checkeof, 0, 0); 467 if (ret > 0) { 468 int l, k = 0; 469 size_t left = len = ret; 470 471 while (left) { 472 void __user *data = i->iov[0].iov_base 473 + i->iov_offset; 474 l = min(i->iov[0].iov_len - i->iov_offset, 475 left); 476 477 ret = ceph_copy_page_vector_to_user(&pages[k], 478 data, off, 479 l); 480 if (ret > 0) { 481 iov_iter_advance(i, ret); 482 left -= ret; 483 off += ret; 484 k = calc_pages_for(iocb->ki_pos, 485 len - left + 1) - 1; 486 BUG_ON(k >= num_pages && left); 487 } else 488 break; 489 } 490 } 491 ceph_release_page_vector(pages, num_pages); 492 } 493 494 if (off > iocb->ki_pos) { 495 ret = off - iocb->ki_pos; 496 iocb->ki_pos = off; 497 } 498 499 dout("sync_read result %d\n", ret); 500 return ret; 501 } 502 503 /* 504 * Write commit request unsafe callback, called to tell us when a 505 * request is unsafe (that is, in flight--has been handed to the 506 * messenger to send to its target osd). It is called again when 507 * we've received a response message indicating the request is 508 * "safe" (its CEPH_OSD_FLAG_ONDISK flag is set), or when a request 509 * is completed early (and unsuccessfully) due to a timeout or 510 * interrupt. 511 * 512 * This is used if we requested both an ACK and ONDISK commit reply 513 * from the OSD. 514 */ 515 static void ceph_sync_write_unsafe(struct ceph_osd_request *req, bool unsafe) 516 { 517 struct ceph_inode_info *ci = ceph_inode(req->r_inode); 518 519 dout("%s %p tid %llu %ssafe\n", __func__, req, req->r_tid, 520 unsafe ? "un" : ""); 521 if (unsafe) { 522 ceph_get_cap_refs(ci, CEPH_CAP_FILE_WR); 523 spin_lock(&ci->i_unsafe_lock); 524 list_add_tail(&req->r_unsafe_item, 525 &ci->i_unsafe_writes); 526 spin_unlock(&ci->i_unsafe_lock); 527 } else { 528 spin_lock(&ci->i_unsafe_lock); 529 list_del_init(&req->r_unsafe_item); 530 spin_unlock(&ci->i_unsafe_lock); 531 ceph_put_cap_refs(ci, CEPH_CAP_FILE_WR); 532 } 533 } 534 535 536 /* 537 * Synchronous write, straight from __user pointer or user pages. 538 * 539 * If write spans object boundary, just do multiple writes. (For a 540 * correct atomic write, we should e.g. take write locks on all 541 * objects, rollback on failure, etc.) 542 */ 543 static ssize_t 544 ceph_sync_direct_write(struct kiocb *iocb, const struct iovec *iov, 545 unsigned long nr_segs, size_t count) 546 { 547 struct file *file = iocb->ki_filp; 548 struct inode *inode = file_inode(file); 549 struct ceph_inode_info *ci = ceph_inode(inode); 550 struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 551 struct ceph_snap_context *snapc; 552 struct ceph_vino vino; 553 struct ceph_osd_request *req; 554 struct page **pages; 555 int num_pages; 556 int written = 0; 557 int flags; 558 int check_caps = 0; 559 int page_align; 560 int ret; 561 struct timespec mtime = CURRENT_TIME; 562 loff_t pos = iocb->ki_pos; 563 struct iov_iter i; 564 565 if (ceph_snap(file_inode(file)) != CEPH_NOSNAP) 566 return -EROFS; 567 568 dout("sync_direct_write on file %p %lld~%u\n", file, pos, 569 (unsigned)count); 570 571 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count); 572 if (ret < 0) 573 return ret; 574 575 ret = invalidate_inode_pages2_range(inode->i_mapping, 576 pos >> PAGE_CACHE_SHIFT, 577 (pos + count) >> PAGE_CACHE_SHIFT); 578 if (ret < 0) 579 dout("invalidate_inode_pages2_range returned %d\n", ret); 580 581 flags = CEPH_OSD_FLAG_ORDERSNAP | 582 CEPH_OSD_FLAG_ONDISK | 583 CEPH_OSD_FLAG_WRITE; 584 585 iov_iter_init(&i, iov, nr_segs, count, 0); 586 587 while (iov_iter_count(&i) > 0) { 588 void __user *data = i.iov->iov_base + i.iov_offset; 589 u64 len = i.iov->iov_len - i.iov_offset; 590 591 page_align = (unsigned long)data & ~PAGE_MASK; 592 593 snapc = ci->i_snap_realm->cached_context; 594 vino = ceph_vino(inode); 595 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, 596 vino, pos, &len, 597 2,/*include a 'startsync' command*/ 598 CEPH_OSD_OP_WRITE, flags, snapc, 599 ci->i_truncate_seq, 600 ci->i_truncate_size, 601 false); 602 if (IS_ERR(req)) { 603 ret = PTR_ERR(req); 604 break; 605 } 606 607 num_pages = calc_pages_for(page_align, len); 608 pages = ceph_get_direct_page_vector(data, num_pages, false); 609 if (IS_ERR(pages)) { 610 ret = PTR_ERR(pages); 611 goto out; 612 } 613 614 /* 615 * throw out any page cache pages in this range. this 616 * may block. 617 */ 618 truncate_inode_pages_range(inode->i_mapping, pos, 619 (pos+len) | (PAGE_CACHE_SIZE-1)); 620 osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align, 621 false, false); 622 623 /* BUG_ON(vino.snap != CEPH_NOSNAP); */ 624 ceph_osdc_build_request(req, pos, snapc, vino.snap, &mtime); 625 626 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false); 627 if (!ret) 628 ret = ceph_osdc_wait_request(&fsc->client->osdc, req); 629 630 ceph_put_page_vector(pages, num_pages, false); 631 632 out: 633 ceph_osdc_put_request(req); 634 if (ret == 0) { 635 pos += len; 636 written += len; 637 iov_iter_advance(&i, (size_t)len); 638 639 if (pos > i_size_read(inode)) { 640 check_caps = ceph_inode_set_size(inode, pos); 641 if (check_caps) 642 ceph_check_caps(ceph_inode(inode), 643 CHECK_CAPS_AUTHONLY, 644 NULL); 645 } 646 } else 647 break; 648 } 649 650 if (ret != -EOLDSNAPC && written > 0) { 651 iocb->ki_pos = pos; 652 ret = written; 653 } 654 return ret; 655 } 656 657 658 /* 659 * Synchronous write, straight from __user pointer or user pages. 660 * 661 * If write spans object boundary, just do multiple writes. (For a 662 * correct atomic write, we should e.g. take write locks on all 663 * objects, rollback on failure, etc.) 664 */ 665 static ssize_t ceph_sync_write(struct kiocb *iocb, const struct iovec *iov, 666 unsigned long nr_segs, size_t count) 667 { 668 struct file *file = iocb->ki_filp; 669 struct inode *inode = file_inode(file); 670 struct ceph_inode_info *ci = ceph_inode(inode); 671 struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 672 struct ceph_snap_context *snapc; 673 struct ceph_vino vino; 674 struct ceph_osd_request *req; 675 struct page **pages; 676 u64 len; 677 int num_pages; 678 int written = 0; 679 int flags; 680 int check_caps = 0; 681 int ret; 682 struct timespec mtime = CURRENT_TIME; 683 loff_t pos = iocb->ki_pos; 684 struct iov_iter i; 685 686 if (ceph_snap(file_inode(file)) != CEPH_NOSNAP) 687 return -EROFS; 688 689 dout("sync_write on file %p %lld~%u\n", file, pos, (unsigned)count); 690 691 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count); 692 if (ret < 0) 693 return ret; 694 695 ret = invalidate_inode_pages2_range(inode->i_mapping, 696 pos >> PAGE_CACHE_SHIFT, 697 (pos + count) >> PAGE_CACHE_SHIFT); 698 if (ret < 0) 699 dout("invalidate_inode_pages2_range returned %d\n", ret); 700 701 flags = CEPH_OSD_FLAG_ORDERSNAP | 702 CEPH_OSD_FLAG_ONDISK | 703 CEPH_OSD_FLAG_WRITE | 704 CEPH_OSD_FLAG_ACK; 705 706 iov_iter_init(&i, iov, nr_segs, count, 0); 707 708 while ((len = iov_iter_count(&i)) > 0) { 709 size_t left; 710 int n; 711 712 snapc = ci->i_snap_realm->cached_context; 713 vino = ceph_vino(inode); 714 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, 715 vino, pos, &len, 1, 716 CEPH_OSD_OP_WRITE, flags, snapc, 717 ci->i_truncate_seq, 718 ci->i_truncate_size, 719 false); 720 if (IS_ERR(req)) { 721 ret = PTR_ERR(req); 722 break; 723 } 724 725 /* 726 * write from beginning of first page, 727 * regardless of io alignment 728 */ 729 num_pages = (len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT; 730 731 pages = ceph_alloc_page_vector(num_pages, GFP_NOFS); 732 if (IS_ERR(pages)) { 733 ret = PTR_ERR(pages); 734 goto out; 735 } 736 737 left = len; 738 for (n = 0; n < num_pages; n++) { 739 size_t plen = min_t(size_t, left, PAGE_SIZE); 740 ret = iov_iter_copy_from_user(pages[n], &i, 0, plen); 741 if (ret != plen) { 742 ret = -EFAULT; 743 break; 744 } 745 left -= ret; 746 iov_iter_advance(&i, ret); 747 } 748 749 if (ret < 0) { 750 ceph_release_page_vector(pages, num_pages); 751 goto out; 752 } 753 754 /* get a second commit callback */ 755 req->r_unsafe_callback = ceph_sync_write_unsafe; 756 req->r_inode = inode; 757 758 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, 759 false, true); 760 761 /* BUG_ON(vino.snap != CEPH_NOSNAP); */ 762 ceph_osdc_build_request(req, pos, snapc, vino.snap, &mtime); 763 764 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false); 765 if (!ret) 766 ret = ceph_osdc_wait_request(&fsc->client->osdc, req); 767 768 out: 769 ceph_osdc_put_request(req); 770 if (ret == 0) { 771 pos += len; 772 written += len; 773 774 if (pos > i_size_read(inode)) { 775 check_caps = ceph_inode_set_size(inode, pos); 776 if (check_caps) 777 ceph_check_caps(ceph_inode(inode), 778 CHECK_CAPS_AUTHONLY, 779 NULL); 780 } 781 } else 782 break; 783 } 784 785 if (ret != -EOLDSNAPC && written > 0) { 786 ret = written; 787 iocb->ki_pos = pos; 788 } 789 return ret; 790 } 791 792 /* 793 * Wrap generic_file_aio_read with checks for cap bits on the inode. 794 * Atomically grab references, so that those bits are not released 795 * back to the MDS mid-read. 796 * 797 * Hmm, the sync read case isn't actually async... should it be? 798 */ 799 static ssize_t ceph_aio_read(struct kiocb *iocb, const struct iovec *iov, 800 unsigned long nr_segs, loff_t pos) 801 { 802 struct file *filp = iocb->ki_filp; 803 struct ceph_file_info *fi = filp->private_data; 804 size_t len = iocb->ki_nbytes; 805 struct inode *inode = file_inode(filp); 806 struct ceph_inode_info *ci = ceph_inode(inode); 807 ssize_t ret; 808 int want, got = 0; 809 int checkeof = 0, read = 0; 810 811 again: 812 dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n", 813 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode); 814 815 if (fi->fmode & CEPH_FILE_MODE_LAZY) 816 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO; 817 else 818 want = CEPH_CAP_FILE_CACHE; 819 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, &got, -1); 820 if (ret < 0) 821 return ret; 822 823 if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 || 824 (iocb->ki_filp->f_flags & O_DIRECT) || 825 (fi->flags & CEPH_F_SYNC)) { 826 struct iov_iter i; 827 828 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n", 829 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, 830 ceph_cap_string(got)); 831 832 if (!read) { 833 ret = generic_segment_checks(iov, &nr_segs, 834 &len, VERIFY_WRITE); 835 if (ret) 836 goto out; 837 } 838 839 iov_iter_init(&i, iov, nr_segs, len, read); 840 841 /* hmm, this isn't really async... */ 842 ret = ceph_sync_read(iocb, &i, &checkeof); 843 } else { 844 /* 845 * We can't modify the content of iov, 846 * so we only read from beginning. 847 */ 848 if (read) { 849 iocb->ki_pos = pos; 850 len = iocb->ki_nbytes; 851 read = 0; 852 } 853 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n", 854 inode, ceph_vinop(inode), pos, (unsigned)len, 855 ceph_cap_string(got)); 856 857 ret = generic_file_aio_read(iocb, iov, nr_segs, pos); 858 } 859 out: 860 dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n", 861 inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret); 862 ceph_put_cap_refs(ci, got); 863 864 if (checkeof && ret >= 0) { 865 int statret = ceph_do_getattr(inode, 866 CEPH_STAT_CAP_SIZE); 867 868 /* hit EOF or hole? */ 869 if (statret == 0 && iocb->ki_pos < inode->i_size && 870 ret < len) { 871 dout("sync_read hit hole, ppos %lld < size %lld" 872 ", reading more\n", iocb->ki_pos, 873 inode->i_size); 874 875 read += ret; 876 len -= ret; 877 checkeof = 0; 878 goto again; 879 } 880 } 881 882 if (ret >= 0) 883 ret += read; 884 885 return ret; 886 } 887 888 /* 889 * Take cap references to avoid releasing caps to MDS mid-write. 890 * 891 * If we are synchronous, and write with an old snap context, the OSD 892 * may return EOLDSNAPC. In that case, retry the write.. _after_ 893 * dropping our cap refs and allowing the pending snap to logically 894 * complete _before_ this write occurs. 895 * 896 * If we are near ENOSPC, write synchronously. 897 */ 898 static ssize_t ceph_aio_write(struct kiocb *iocb, const struct iovec *iov, 899 unsigned long nr_segs, loff_t pos) 900 { 901 struct file *file = iocb->ki_filp; 902 struct ceph_file_info *fi = file->private_data; 903 struct inode *inode = file_inode(file); 904 struct ceph_inode_info *ci = ceph_inode(inode); 905 struct ceph_osd_client *osdc = 906 &ceph_sb_to_client(inode->i_sb)->client->osdc; 907 ssize_t count, written = 0; 908 int err, want, got; 909 910 if (ceph_snap(inode) != CEPH_NOSNAP) 911 return -EROFS; 912 913 mutex_lock(&inode->i_mutex); 914 915 err = generic_segment_checks(iov, &nr_segs, &count, VERIFY_READ); 916 if (err) 917 goto out; 918 919 /* We can write back this queue in page reclaim */ 920 current->backing_dev_info = file->f_mapping->backing_dev_info; 921 922 err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode)); 923 if (err) 924 goto out; 925 926 if (count == 0) 927 goto out; 928 929 err = file_remove_suid(file); 930 if (err) 931 goto out; 932 933 err = file_update_time(file); 934 if (err) 935 goto out; 936 937 retry_snap: 938 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL)) { 939 err = -ENOSPC; 940 goto out; 941 } 942 943 dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n", 944 inode, ceph_vinop(inode), pos, count, inode->i_size); 945 if (fi->fmode & CEPH_FILE_MODE_LAZY) 946 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO; 947 else 948 want = CEPH_CAP_FILE_BUFFER; 949 got = 0; 950 err = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, &got, pos + count); 951 if (err < 0) 952 goto out; 953 954 dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n", 955 inode, ceph_vinop(inode), pos, count, ceph_cap_string(got)); 956 957 if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 || 958 (file->f_flags & O_DIRECT) || (fi->flags & CEPH_F_SYNC)) { 959 mutex_unlock(&inode->i_mutex); 960 if (file->f_flags & O_DIRECT) 961 written = ceph_sync_direct_write(iocb, iov, 962 nr_segs, count); 963 else 964 written = ceph_sync_write(iocb, iov, nr_segs, count); 965 if (written == -EOLDSNAPC) { 966 dout("aio_write %p %llx.%llx %llu~%u" 967 "got EOLDSNAPC, retrying\n", 968 inode, ceph_vinop(inode), 969 pos, (unsigned)iov->iov_len); 970 mutex_lock(&inode->i_mutex); 971 goto retry_snap; 972 } 973 } else { 974 loff_t old_size = inode->i_size; 975 struct iov_iter from; 976 /* 977 * No need to acquire the i_truncate_mutex. Because 978 * the MDS revokes Fwb caps before sending truncate 979 * message to us. We can't get Fwb cap while there 980 * are pending vmtruncate. So write and vmtruncate 981 * can not run at the same time 982 */ 983 iov_iter_init(&from, iov, nr_segs, count, 0); 984 written = generic_perform_write(file, &from, pos); 985 if (likely(written >= 0)) 986 iocb->ki_pos = pos + written; 987 if (inode->i_size > old_size) 988 ceph_fscache_update_objectsize(inode); 989 mutex_unlock(&inode->i_mutex); 990 } 991 992 if (written >= 0) { 993 int dirty; 994 spin_lock(&ci->i_ceph_lock); 995 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR); 996 spin_unlock(&ci->i_ceph_lock); 997 if (dirty) 998 __mark_inode_dirty(inode, dirty); 999 } 1000 1001 dout("aio_write %p %llx.%llx %llu~%u dropping cap refs on %s\n", 1002 inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len, 1003 ceph_cap_string(got)); 1004 ceph_put_cap_refs(ci, got); 1005 1006 if (written >= 0 && 1007 ((file->f_flags & O_SYNC) || IS_SYNC(file->f_mapping->host) || 1008 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_NEARFULL))) { 1009 err = vfs_fsync_range(file, pos, pos + written - 1, 1); 1010 if (err < 0) 1011 written = err; 1012 } 1013 1014 goto out_unlocked; 1015 1016 out: 1017 mutex_unlock(&inode->i_mutex); 1018 out_unlocked: 1019 current->backing_dev_info = NULL; 1020 return written ? written : err; 1021 } 1022 1023 /* 1024 * llseek. be sure to verify file size on SEEK_END. 1025 */ 1026 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence) 1027 { 1028 struct inode *inode = file->f_mapping->host; 1029 int ret; 1030 1031 mutex_lock(&inode->i_mutex); 1032 1033 if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) { 1034 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE); 1035 if (ret < 0) { 1036 offset = ret; 1037 goto out; 1038 } 1039 } 1040 1041 switch (whence) { 1042 case SEEK_END: 1043 offset += inode->i_size; 1044 break; 1045 case SEEK_CUR: 1046 /* 1047 * Here we special-case the lseek(fd, 0, SEEK_CUR) 1048 * position-querying operation. Avoid rewriting the "same" 1049 * f_pos value back to the file because a concurrent read(), 1050 * write() or lseek() might have altered it 1051 */ 1052 if (offset == 0) { 1053 offset = file->f_pos; 1054 goto out; 1055 } 1056 offset += file->f_pos; 1057 break; 1058 case SEEK_DATA: 1059 if (offset >= inode->i_size) { 1060 ret = -ENXIO; 1061 goto out; 1062 } 1063 break; 1064 case SEEK_HOLE: 1065 if (offset >= inode->i_size) { 1066 ret = -ENXIO; 1067 goto out; 1068 } 1069 offset = inode->i_size; 1070 break; 1071 } 1072 1073 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes); 1074 1075 out: 1076 mutex_unlock(&inode->i_mutex); 1077 return offset; 1078 } 1079 1080 static inline void ceph_zero_partial_page( 1081 struct inode *inode, loff_t offset, unsigned size) 1082 { 1083 struct page *page; 1084 pgoff_t index = offset >> PAGE_CACHE_SHIFT; 1085 1086 page = find_lock_page(inode->i_mapping, index); 1087 if (page) { 1088 wait_on_page_writeback(page); 1089 zero_user(page, offset & (PAGE_CACHE_SIZE - 1), size); 1090 unlock_page(page); 1091 page_cache_release(page); 1092 } 1093 } 1094 1095 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset, 1096 loff_t length) 1097 { 1098 loff_t nearly = round_up(offset, PAGE_CACHE_SIZE); 1099 if (offset < nearly) { 1100 loff_t size = nearly - offset; 1101 if (length < size) 1102 size = length; 1103 ceph_zero_partial_page(inode, offset, size); 1104 offset += size; 1105 length -= size; 1106 } 1107 if (length >= PAGE_CACHE_SIZE) { 1108 loff_t size = round_down(length, PAGE_CACHE_SIZE); 1109 truncate_pagecache_range(inode, offset, offset + size - 1); 1110 offset += size; 1111 length -= size; 1112 } 1113 if (length) 1114 ceph_zero_partial_page(inode, offset, length); 1115 } 1116 1117 static int ceph_zero_partial_object(struct inode *inode, 1118 loff_t offset, loff_t *length) 1119 { 1120 struct ceph_inode_info *ci = ceph_inode(inode); 1121 struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 1122 struct ceph_osd_request *req; 1123 int ret = 0; 1124 loff_t zero = 0; 1125 int op; 1126 1127 if (!length) { 1128 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE; 1129 length = &zero; 1130 } else { 1131 op = CEPH_OSD_OP_ZERO; 1132 } 1133 1134 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, 1135 ceph_vino(inode), 1136 offset, length, 1137 1, op, 1138 CEPH_OSD_FLAG_WRITE | 1139 CEPH_OSD_FLAG_ONDISK, 1140 NULL, 0, 0, false); 1141 if (IS_ERR(req)) { 1142 ret = PTR_ERR(req); 1143 goto out; 1144 } 1145 1146 ceph_osdc_build_request(req, offset, NULL, ceph_vino(inode).snap, 1147 &inode->i_mtime); 1148 1149 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false); 1150 if (!ret) { 1151 ret = ceph_osdc_wait_request(&fsc->client->osdc, req); 1152 if (ret == -ENOENT) 1153 ret = 0; 1154 } 1155 ceph_osdc_put_request(req); 1156 1157 out: 1158 return ret; 1159 } 1160 1161 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length) 1162 { 1163 int ret = 0; 1164 struct ceph_inode_info *ci = ceph_inode(inode); 1165 s32 stripe_unit = ceph_file_layout_su(ci->i_layout); 1166 s32 stripe_count = ceph_file_layout_stripe_count(ci->i_layout); 1167 s32 object_size = ceph_file_layout_object_size(ci->i_layout); 1168 u64 object_set_size = object_size * stripe_count; 1169 u64 nearly, t; 1170 1171 /* round offset up to next period boundary */ 1172 nearly = offset + object_set_size - 1; 1173 t = nearly; 1174 nearly -= do_div(t, object_set_size); 1175 1176 while (length && offset < nearly) { 1177 loff_t size = length; 1178 ret = ceph_zero_partial_object(inode, offset, &size); 1179 if (ret < 0) 1180 return ret; 1181 offset += size; 1182 length -= size; 1183 } 1184 while (length >= object_set_size) { 1185 int i; 1186 loff_t pos = offset; 1187 for (i = 0; i < stripe_count; ++i) { 1188 ret = ceph_zero_partial_object(inode, pos, NULL); 1189 if (ret < 0) 1190 return ret; 1191 pos += stripe_unit; 1192 } 1193 offset += object_set_size; 1194 length -= object_set_size; 1195 } 1196 while (length) { 1197 loff_t size = length; 1198 ret = ceph_zero_partial_object(inode, offset, &size); 1199 if (ret < 0) 1200 return ret; 1201 offset += size; 1202 length -= size; 1203 } 1204 return ret; 1205 } 1206 1207 static long ceph_fallocate(struct file *file, int mode, 1208 loff_t offset, loff_t length) 1209 { 1210 struct ceph_file_info *fi = file->private_data; 1211 struct inode *inode = file_inode(file); 1212 struct ceph_inode_info *ci = ceph_inode(inode); 1213 struct ceph_osd_client *osdc = 1214 &ceph_inode_to_client(inode)->client->osdc; 1215 int want, got = 0; 1216 int dirty; 1217 int ret = 0; 1218 loff_t endoff = 0; 1219 loff_t size; 1220 1221 if (!S_ISREG(inode->i_mode)) 1222 return -EOPNOTSUPP; 1223 1224 if (IS_SWAPFILE(inode)) 1225 return -ETXTBSY; 1226 1227 mutex_lock(&inode->i_mutex); 1228 1229 if (ceph_snap(inode) != CEPH_NOSNAP) { 1230 ret = -EROFS; 1231 goto unlock; 1232 } 1233 1234 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) && 1235 !(mode & FALLOC_FL_PUNCH_HOLE)) { 1236 ret = -ENOSPC; 1237 goto unlock; 1238 } 1239 1240 size = i_size_read(inode); 1241 if (!(mode & FALLOC_FL_KEEP_SIZE)) 1242 endoff = offset + length; 1243 1244 if (fi->fmode & CEPH_FILE_MODE_LAZY) 1245 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO; 1246 else 1247 want = CEPH_CAP_FILE_BUFFER; 1248 1249 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, &got, endoff); 1250 if (ret < 0) 1251 goto unlock; 1252 1253 if (mode & FALLOC_FL_PUNCH_HOLE) { 1254 if (offset < size) 1255 ceph_zero_pagecache_range(inode, offset, length); 1256 ret = ceph_zero_objects(inode, offset, length); 1257 } else if (endoff > size) { 1258 truncate_pagecache_range(inode, size, -1); 1259 if (ceph_inode_set_size(inode, endoff)) 1260 ceph_check_caps(ceph_inode(inode), 1261 CHECK_CAPS_AUTHONLY, NULL); 1262 } 1263 1264 if (!ret) { 1265 spin_lock(&ci->i_ceph_lock); 1266 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR); 1267 spin_unlock(&ci->i_ceph_lock); 1268 if (dirty) 1269 __mark_inode_dirty(inode, dirty); 1270 } 1271 1272 ceph_put_cap_refs(ci, got); 1273 unlock: 1274 mutex_unlock(&inode->i_mutex); 1275 return ret; 1276 } 1277 1278 const struct file_operations ceph_file_fops = { 1279 .open = ceph_open, 1280 .release = ceph_release, 1281 .llseek = ceph_llseek, 1282 .read = do_sync_read, 1283 .write = do_sync_write, 1284 .aio_read = ceph_aio_read, 1285 .aio_write = ceph_aio_write, 1286 .mmap = ceph_mmap, 1287 .fsync = ceph_fsync, 1288 .lock = ceph_lock, 1289 .flock = ceph_flock, 1290 .splice_read = generic_file_splice_read, 1291 .splice_write = generic_file_splice_write, 1292 .unlocked_ioctl = ceph_ioctl, 1293 .compat_ioctl = ceph_ioctl, 1294 .fallocate = ceph_fallocate, 1295 }; 1296 1297