1 #include "ceph_debug.h" 2 3 #include <linux/backing-dev.h> 4 #include <linux/fs.h> 5 #include <linux/mm.h> 6 #include <linux/pagemap.h> 7 #include <linux/writeback.h> /* generic_writepages */ 8 #include <linux/pagevec.h> 9 #include <linux/task_io_accounting_ops.h> 10 11 #include "super.h" 12 #include "osd_client.h" 13 14 /* 15 * Ceph address space ops. 16 * 17 * There are a few funny things going on here. 18 * 19 * The page->private field is used to reference a struct 20 * ceph_snap_context for _every_ dirty page. This indicates which 21 * snapshot the page was logically dirtied in, and thus which snap 22 * context needs to be associated with the osd write during writeback. 23 * 24 * Similarly, struct ceph_inode_info maintains a set of counters to 25 * count dirty pages on the inode. In the absense of snapshots, 26 * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count. 27 * 28 * When a snapshot is taken (that is, when the client receives 29 * notification that a snapshot was taken), each inode with caps and 30 * with dirty pages (dirty pages implies there is a cap) gets a new 31 * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending 32 * order, new snaps go to the tail). The i_wrbuffer_ref_head count is 33 * moved to capsnap->dirty. (Unless a sync write is currently in 34 * progress. In that case, the capsnap is said to be "pending", new 35 * writes cannot start, and the capsnap isn't "finalized" until the 36 * write completes (or fails) and a final size/mtime for the inode for 37 * that snap can be settled upon.) i_wrbuffer_ref_head is reset to 0. 38 * 39 * On writeback, we must submit writes to the osd IN SNAP ORDER. So, 40 * we look for the first capsnap in i_cap_snaps and write out pages in 41 * that snap context _only_. Then we move on to the next capsnap, 42 * eventually reaching the "live" or "head" context (i.e., pages that 43 * are not yet snapped) and are writing the most recently dirtied 44 * pages. 45 * 46 * Invalidate and so forth must take care to ensure the dirty page 47 * accounting is preserved. 48 */ 49 50 51 /* 52 * Dirty a page. Optimistically adjust accounting, on the assumption 53 * that we won't race with invalidate. If we do, readjust. 54 */ 55 static int ceph_set_page_dirty(struct page *page) 56 { 57 struct address_space *mapping = page->mapping; 58 struct inode *inode; 59 struct ceph_inode_info *ci; 60 int undo = 0; 61 struct ceph_snap_context *snapc; 62 63 if (unlikely(!mapping)) 64 return !TestSetPageDirty(page); 65 66 if (TestSetPageDirty(page)) { 67 dout("%p set_page_dirty %p idx %lu -- already dirty\n", 68 mapping->host, page, page->index); 69 return 0; 70 } 71 72 inode = mapping->host; 73 ci = ceph_inode(inode); 74 75 /* 76 * Note that we're grabbing a snapc ref here without holding 77 * any locks! 78 */ 79 snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context); 80 81 /* dirty the head */ 82 spin_lock(&inode->i_lock); 83 if (ci->i_wrbuffer_ref_head == 0) 84 ci->i_head_snapc = ceph_get_snap_context(snapc); 85 ++ci->i_wrbuffer_ref_head; 86 if (ci->i_wrbuffer_ref == 0) 87 igrab(inode); 88 ++ci->i_wrbuffer_ref; 89 dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d " 90 "snapc %p seq %lld (%d snaps)\n", 91 mapping->host, page, page->index, 92 ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1, 93 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head, 94 snapc, snapc->seq, snapc->num_snaps); 95 spin_unlock(&inode->i_lock); 96 97 /* now adjust page */ 98 spin_lock_irq(&mapping->tree_lock); 99 if (page->mapping) { /* Race with truncate? */ 100 WARN_ON_ONCE(!PageUptodate(page)); 101 102 if (mapping_cap_account_dirty(mapping)) { 103 __inc_zone_page_state(page, NR_FILE_DIRTY); 104 __inc_bdi_stat(mapping->backing_dev_info, 105 BDI_RECLAIMABLE); 106 task_io_account_write(PAGE_CACHE_SIZE); 107 } 108 radix_tree_tag_set(&mapping->page_tree, 109 page_index(page), PAGECACHE_TAG_DIRTY); 110 111 /* 112 * Reference snap context in page->private. Also set 113 * PagePrivate so that we get invalidatepage callback. 114 */ 115 page->private = (unsigned long)snapc; 116 SetPagePrivate(page); 117 } else { 118 dout("ANON set_page_dirty %p (raced truncate?)\n", page); 119 undo = 1; 120 } 121 122 spin_unlock_irq(&mapping->tree_lock); 123 124 if (undo) 125 /* whoops, we failed to dirty the page */ 126 ceph_put_wrbuffer_cap_refs(ci, 1, snapc); 127 128 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); 129 130 BUG_ON(!PageDirty(page)); 131 return 1; 132 } 133 134 /* 135 * If we are truncating the full page (i.e. offset == 0), adjust the 136 * dirty page counters appropriately. Only called if there is private 137 * data on the page. 138 */ 139 static void ceph_invalidatepage(struct page *page, unsigned long offset) 140 { 141 struct inode *inode = page->mapping->host; 142 struct ceph_inode_info *ci; 143 struct ceph_snap_context *snapc = (void *)page->private; 144 145 BUG_ON(!PageLocked(page)); 146 BUG_ON(!page->private); 147 BUG_ON(!PagePrivate(page)); 148 BUG_ON(!page->mapping); 149 150 /* 151 * We can get non-dirty pages here due to races between 152 * set_page_dirty and truncate_complete_page; just spit out a 153 * warning, in case we end up with accounting problems later. 154 */ 155 if (!PageDirty(page)) 156 pr_err("%p invalidatepage %p page not dirty\n", inode, page); 157 158 if (offset == 0) 159 ClearPageChecked(page); 160 161 ci = ceph_inode(inode); 162 if (offset == 0) { 163 dout("%p invalidatepage %p idx %lu full dirty page %lu\n", 164 inode, page, page->index, offset); 165 ceph_put_wrbuffer_cap_refs(ci, 1, snapc); 166 ceph_put_snap_context(snapc); 167 page->private = 0; 168 ClearPagePrivate(page); 169 } else { 170 dout("%p invalidatepage %p idx %lu partial dirty page\n", 171 inode, page, page->index); 172 } 173 } 174 175 /* just a sanity check */ 176 static int ceph_releasepage(struct page *page, gfp_t g) 177 { 178 struct inode *inode = page->mapping ? page->mapping->host : NULL; 179 dout("%p releasepage %p idx %lu\n", inode, page, page->index); 180 WARN_ON(PageDirty(page)); 181 WARN_ON(page->private); 182 WARN_ON(PagePrivate(page)); 183 return 0; 184 } 185 186 /* 187 * read a single page, without unlocking it. 188 */ 189 static int readpage_nounlock(struct file *filp, struct page *page) 190 { 191 struct inode *inode = filp->f_dentry->d_inode; 192 struct ceph_inode_info *ci = ceph_inode(inode); 193 struct ceph_osd_client *osdc = &ceph_inode_to_client(inode)->osdc; 194 int err = 0; 195 u64 len = PAGE_CACHE_SIZE; 196 197 dout("readpage inode %p file %p page %p index %lu\n", 198 inode, filp, page, page->index); 199 err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout, 200 page->index << PAGE_CACHE_SHIFT, &len, 201 ci->i_truncate_seq, ci->i_truncate_size, 202 &page, 1); 203 if (err == -ENOENT) 204 err = 0; 205 if (err < 0) { 206 SetPageError(page); 207 goto out; 208 } else if (err < PAGE_CACHE_SIZE) { 209 /* zero fill remainder of page */ 210 zero_user_segment(page, err, PAGE_CACHE_SIZE); 211 } 212 SetPageUptodate(page); 213 214 out: 215 return err < 0 ? err : 0; 216 } 217 218 static int ceph_readpage(struct file *filp, struct page *page) 219 { 220 int r = readpage_nounlock(filp, page); 221 unlock_page(page); 222 return r; 223 } 224 225 /* 226 * Build a vector of contiguous pages from the provided page list. 227 */ 228 static struct page **page_vector_from_list(struct list_head *page_list, 229 unsigned *nr_pages) 230 { 231 struct page **pages; 232 struct page *page; 233 int next_index, contig_pages = 0; 234 235 /* build page vector */ 236 pages = kmalloc(sizeof(*pages) * *nr_pages, GFP_NOFS); 237 if (!pages) 238 return ERR_PTR(-ENOMEM); 239 240 BUG_ON(list_empty(page_list)); 241 next_index = list_entry(page_list->prev, struct page, lru)->index; 242 list_for_each_entry_reverse(page, page_list, lru) { 243 if (page->index == next_index) { 244 dout("readpages page %d %p\n", contig_pages, page); 245 pages[contig_pages] = page; 246 contig_pages++; 247 next_index++; 248 } else { 249 break; 250 } 251 } 252 *nr_pages = contig_pages; 253 return pages; 254 } 255 256 /* 257 * Read multiple pages. Leave pages we don't read + unlock in page_list; 258 * the caller (VM) cleans them up. 259 */ 260 static int ceph_readpages(struct file *file, struct address_space *mapping, 261 struct list_head *page_list, unsigned nr_pages) 262 { 263 struct inode *inode = file->f_dentry->d_inode; 264 struct ceph_inode_info *ci = ceph_inode(inode); 265 struct ceph_osd_client *osdc = &ceph_inode_to_client(inode)->osdc; 266 int rc = 0; 267 struct page **pages; 268 struct pagevec pvec; 269 loff_t offset; 270 u64 len; 271 272 dout("readpages %p file %p nr_pages %d\n", 273 inode, file, nr_pages); 274 275 pages = page_vector_from_list(page_list, &nr_pages); 276 if (IS_ERR(pages)) 277 return PTR_ERR(pages); 278 279 /* guess read extent */ 280 offset = pages[0]->index << PAGE_CACHE_SHIFT; 281 len = nr_pages << PAGE_CACHE_SHIFT; 282 rc = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout, 283 offset, &len, 284 ci->i_truncate_seq, ci->i_truncate_size, 285 pages, nr_pages); 286 if (rc == -ENOENT) 287 rc = 0; 288 if (rc < 0) 289 goto out; 290 291 /* set uptodate and add to lru in pagevec-sized chunks */ 292 pagevec_init(&pvec, 0); 293 for (; !list_empty(page_list) && len > 0; 294 rc -= PAGE_CACHE_SIZE, len -= PAGE_CACHE_SIZE) { 295 struct page *page = 296 list_entry(page_list->prev, struct page, lru); 297 298 list_del(&page->lru); 299 300 if (rc < (int)PAGE_CACHE_SIZE) { 301 /* zero (remainder of) page */ 302 int s = rc < 0 ? 0 : rc; 303 zero_user_segment(page, s, PAGE_CACHE_SIZE); 304 } 305 306 if (add_to_page_cache(page, mapping, page->index, GFP_NOFS)) { 307 page_cache_release(page); 308 dout("readpages %p add_to_page_cache failed %p\n", 309 inode, page); 310 continue; 311 } 312 dout("readpages %p adding %p idx %lu\n", inode, page, 313 page->index); 314 flush_dcache_page(page); 315 SetPageUptodate(page); 316 unlock_page(page); 317 if (pagevec_add(&pvec, page) == 0) 318 pagevec_lru_add_file(&pvec); /* add to lru */ 319 } 320 pagevec_lru_add_file(&pvec); 321 rc = 0; 322 323 out: 324 kfree(pages); 325 return rc; 326 } 327 328 /* 329 * Get ref for the oldest snapc for an inode with dirty data... that is, the 330 * only snap context we are allowed to write back. 331 * 332 * Caller holds i_lock. 333 */ 334 static struct ceph_snap_context *__get_oldest_context(struct inode *inode, 335 u64 *snap_size) 336 { 337 struct ceph_inode_info *ci = ceph_inode(inode); 338 struct ceph_snap_context *snapc = NULL; 339 struct ceph_cap_snap *capsnap = NULL; 340 341 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { 342 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap, 343 capsnap->context, capsnap->dirty_pages); 344 if (capsnap->dirty_pages) { 345 snapc = ceph_get_snap_context(capsnap->context); 346 if (snap_size) 347 *snap_size = capsnap->size; 348 break; 349 } 350 } 351 if (!snapc && ci->i_snap_realm) { 352 snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context); 353 dout(" head snapc %p has %d dirty pages\n", 354 snapc, ci->i_wrbuffer_ref_head); 355 } 356 return snapc; 357 } 358 359 static struct ceph_snap_context *get_oldest_context(struct inode *inode, 360 u64 *snap_size) 361 { 362 struct ceph_snap_context *snapc = NULL; 363 364 spin_lock(&inode->i_lock); 365 snapc = __get_oldest_context(inode, snap_size); 366 spin_unlock(&inode->i_lock); 367 return snapc; 368 } 369 370 /* 371 * Write a single page, but leave the page locked. 372 * 373 * If we get a write error, set the page error bit, but still adjust the 374 * dirty page accounting (i.e., page is no longer dirty). 375 */ 376 static int writepage_nounlock(struct page *page, struct writeback_control *wbc) 377 { 378 struct inode *inode; 379 struct ceph_inode_info *ci; 380 struct ceph_osd_client *osdc; 381 loff_t page_off = page->index << PAGE_CACHE_SHIFT; 382 int len = PAGE_CACHE_SIZE; 383 loff_t i_size; 384 int err = 0; 385 struct ceph_snap_context *snapc; 386 u64 snap_size = 0; 387 388 dout("writepage %p idx %lu\n", page, page->index); 389 390 if (!page->mapping || !page->mapping->host) { 391 dout("writepage %p - no mapping\n", page); 392 return -EFAULT; 393 } 394 inode = page->mapping->host; 395 ci = ceph_inode(inode); 396 osdc = &ceph_inode_to_client(inode)->osdc; 397 398 /* verify this is a writeable snap context */ 399 snapc = (void *)page->private; 400 if (snapc == NULL) { 401 dout("writepage %p page %p not dirty?\n", inode, page); 402 goto out; 403 } 404 if (snapc != get_oldest_context(inode, &snap_size)) { 405 dout("writepage %p page %p snapc %p not writeable - noop\n", 406 inode, page, (void *)page->private); 407 /* we should only noop if called by kswapd */ 408 WARN_ON((current->flags & PF_MEMALLOC) == 0); 409 goto out; 410 } 411 412 /* is this a partial page at end of file? */ 413 if (snap_size) 414 i_size = snap_size; 415 else 416 i_size = i_size_read(inode); 417 if (i_size < page_off + len) 418 len = i_size - page_off; 419 420 dout("writepage %p page %p index %lu on %llu~%u\n", 421 inode, page, page->index, page_off, len); 422 423 set_page_writeback(page); 424 err = ceph_osdc_writepages(osdc, ceph_vino(inode), 425 &ci->i_layout, snapc, 426 page_off, len, 427 ci->i_truncate_seq, ci->i_truncate_size, 428 &inode->i_mtime, 429 &page, 1, 0, 0, true); 430 if (err < 0) { 431 dout("writepage setting page/mapping error %d %p\n", err, page); 432 SetPageError(page); 433 mapping_set_error(&inode->i_data, err); 434 if (wbc) 435 wbc->pages_skipped++; 436 } else { 437 dout("writepage cleaned page %p\n", page); 438 err = 0; /* vfs expects us to return 0 */ 439 } 440 page->private = 0; 441 ClearPagePrivate(page); 442 end_page_writeback(page); 443 ceph_put_wrbuffer_cap_refs(ci, 1, snapc); 444 ceph_put_snap_context(snapc); 445 out: 446 return err; 447 } 448 449 static int ceph_writepage(struct page *page, struct writeback_control *wbc) 450 { 451 int err; 452 struct inode *inode = page->mapping->host; 453 BUG_ON(!inode); 454 igrab(inode); 455 err = writepage_nounlock(page, wbc); 456 unlock_page(page); 457 iput(inode); 458 return err; 459 } 460 461 462 /* 463 * lame release_pages helper. release_pages() isn't exported to 464 * modules. 465 */ 466 static void ceph_release_pages(struct page **pages, int num) 467 { 468 struct pagevec pvec; 469 int i; 470 471 pagevec_init(&pvec, 0); 472 for (i = 0; i < num; i++) { 473 if (pagevec_add(&pvec, pages[i]) == 0) 474 pagevec_release(&pvec); 475 } 476 pagevec_release(&pvec); 477 } 478 479 480 /* 481 * async writeback completion handler. 482 * 483 * If we get an error, set the mapping error bit, but not the individual 484 * page error bits. 485 */ 486 static void writepages_finish(struct ceph_osd_request *req, 487 struct ceph_msg *msg) 488 { 489 struct inode *inode = req->r_inode; 490 struct ceph_osd_reply_head *replyhead; 491 struct ceph_osd_op *op; 492 struct ceph_inode_info *ci = ceph_inode(inode); 493 unsigned wrote; 494 loff_t offset = req->r_pages[0]->index << PAGE_CACHE_SHIFT; 495 struct page *page; 496 int i; 497 struct ceph_snap_context *snapc = req->r_snapc; 498 struct address_space *mapping = inode->i_mapping; 499 struct writeback_control *wbc = req->r_wbc; 500 __s32 rc = -EIO; 501 u64 bytes = 0; 502 503 /* parse reply */ 504 replyhead = msg->front.iov_base; 505 WARN_ON(le32_to_cpu(replyhead->num_ops) == 0); 506 op = (void *)(replyhead + 1); 507 rc = le32_to_cpu(replyhead->result); 508 bytes = le64_to_cpu(op->extent.length); 509 510 if (rc >= 0) { 511 wrote = (bytes + (offset & ~PAGE_CACHE_MASK) + ~PAGE_CACHE_MASK) 512 >> PAGE_CACHE_SHIFT; 513 WARN_ON(wrote != req->r_num_pages); 514 } else { 515 wrote = 0; 516 mapping_set_error(mapping, rc); 517 } 518 dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n", 519 inode, rc, bytes, wrote); 520 521 /* clean all pages */ 522 for (i = 0; i < req->r_num_pages; i++) { 523 page = req->r_pages[i]; 524 BUG_ON(!page); 525 WARN_ON(!PageUptodate(page)); 526 527 if (i >= wrote) { 528 dout("inode %p skipping page %p\n", inode, page); 529 wbc->pages_skipped++; 530 } 531 page->private = 0; 532 ClearPagePrivate(page); 533 ceph_put_snap_context(snapc); 534 dout("unlocking %d %p\n", i, page); 535 end_page_writeback(page); 536 unlock_page(page); 537 } 538 dout("%p wrote+cleaned %d pages\n", inode, wrote); 539 ceph_put_wrbuffer_cap_refs(ci, req->r_num_pages, snapc); 540 541 ceph_release_pages(req->r_pages, req->r_num_pages); 542 if (req->r_pages_from_pool) 543 mempool_free(req->r_pages, 544 ceph_client(inode->i_sb)->wb_pagevec_pool); 545 else 546 kfree(req->r_pages); 547 ceph_osdc_put_request(req); 548 } 549 550 /* 551 * allocate a page vec, either directly, or if necessary, via a the 552 * mempool. we avoid the mempool if we can because req->r_num_pages 553 * may be less than the maximum write size. 554 */ 555 static void alloc_page_vec(struct ceph_client *client, 556 struct ceph_osd_request *req) 557 { 558 req->r_pages = kmalloc(sizeof(struct page *) * req->r_num_pages, 559 GFP_NOFS); 560 if (!req->r_pages) { 561 req->r_pages = mempool_alloc(client->wb_pagevec_pool, GFP_NOFS); 562 req->r_pages_from_pool = 1; 563 WARN_ON(!req->r_pages); 564 } 565 } 566 567 /* 568 * initiate async writeback 569 */ 570 static int ceph_writepages_start(struct address_space *mapping, 571 struct writeback_control *wbc) 572 { 573 struct inode *inode = mapping->host; 574 struct backing_dev_info *bdi = mapping->backing_dev_info; 575 struct ceph_inode_info *ci = ceph_inode(inode); 576 struct ceph_client *client = ceph_inode_to_client(inode); 577 pgoff_t index, start, end; 578 int range_whole = 0; 579 int should_loop = 1; 580 pgoff_t max_pages = 0, max_pages_ever = 0; 581 struct ceph_snap_context *snapc = NULL, *last_snapc = NULL; 582 struct pagevec pvec; 583 int done = 0; 584 int rc = 0; 585 unsigned wsize = 1 << inode->i_blkbits; 586 struct ceph_osd_request *req = NULL; 587 int do_sync; 588 u64 snap_size = 0; 589 590 /* 591 * Include a 'sync' in the OSD request if this is a data 592 * integrity write (e.g., O_SYNC write or fsync()), or if our 593 * cap is being revoked. 594 */ 595 do_sync = wbc->sync_mode == WB_SYNC_ALL; 596 if (ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER)) 597 do_sync = 1; 598 dout("writepages_start %p dosync=%d (mode=%s)\n", 599 inode, do_sync, 600 wbc->sync_mode == WB_SYNC_NONE ? "NONE" : 601 (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD")); 602 603 client = ceph_inode_to_client(inode); 604 if (client->mount_state == CEPH_MOUNT_SHUTDOWN) { 605 pr_warning("writepage_start %p on forced umount\n", inode); 606 return -EIO; /* we're in a forced umount, don't write! */ 607 } 608 if (client->mount_args->wsize && client->mount_args->wsize < wsize) 609 wsize = client->mount_args->wsize; 610 if (wsize < PAGE_CACHE_SIZE) 611 wsize = PAGE_CACHE_SIZE; 612 max_pages_ever = wsize >> PAGE_CACHE_SHIFT; 613 614 pagevec_init(&pvec, 0); 615 616 /* ?? */ 617 if (wbc->nonblocking && bdi_write_congested(bdi)) { 618 dout(" writepages congested\n"); 619 wbc->encountered_congestion = 1; 620 goto out_final; 621 } 622 623 /* where to start/end? */ 624 if (wbc->range_cyclic) { 625 start = mapping->writeback_index; /* Start from prev offset */ 626 end = -1; 627 dout(" cyclic, start at %lu\n", start); 628 } else { 629 start = wbc->range_start >> PAGE_CACHE_SHIFT; 630 end = wbc->range_end >> PAGE_CACHE_SHIFT; 631 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX) 632 range_whole = 1; 633 should_loop = 0; 634 dout(" not cyclic, %lu to %lu\n", start, end); 635 } 636 index = start; 637 638 retry: 639 /* find oldest snap context with dirty data */ 640 ceph_put_snap_context(snapc); 641 snapc = get_oldest_context(inode, &snap_size); 642 if (!snapc) { 643 /* hmm, why does writepages get called when there 644 is no dirty data? */ 645 dout(" no snap context with dirty data?\n"); 646 goto out; 647 } 648 dout(" oldest snapc is %p seq %lld (%d snaps)\n", 649 snapc, snapc->seq, snapc->num_snaps); 650 if (last_snapc && snapc != last_snapc) { 651 /* if we switched to a newer snapc, restart our scan at the 652 * start of the original file range. */ 653 dout(" snapc differs from last pass, restarting at %lu\n", 654 index); 655 index = start; 656 } 657 last_snapc = snapc; 658 659 while (!done && index <= end) { 660 unsigned i; 661 int first; 662 pgoff_t next; 663 int pvec_pages, locked_pages; 664 struct page *page; 665 int want; 666 u64 offset, len; 667 struct ceph_osd_request_head *reqhead; 668 struct ceph_osd_op *op; 669 670 next = 0; 671 locked_pages = 0; 672 max_pages = max_pages_ever; 673 674 get_more_pages: 675 first = -1; 676 want = min(end - index, 677 min((pgoff_t)PAGEVEC_SIZE, 678 max_pages - (pgoff_t)locked_pages) - 1) 679 + 1; 680 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index, 681 PAGECACHE_TAG_DIRTY, 682 want); 683 dout("pagevec_lookup_tag got %d\n", pvec_pages); 684 if (!pvec_pages && !locked_pages) 685 break; 686 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) { 687 page = pvec.pages[i]; 688 dout("? %p idx %lu\n", page, page->index); 689 if (locked_pages == 0) 690 lock_page(page); /* first page */ 691 else if (!trylock_page(page)) 692 break; 693 694 /* only dirty pages, or our accounting breaks */ 695 if (unlikely(!PageDirty(page)) || 696 unlikely(page->mapping != mapping)) { 697 dout("!dirty or !mapping %p\n", page); 698 unlock_page(page); 699 break; 700 } 701 if (!wbc->range_cyclic && page->index > end) { 702 dout("end of range %p\n", page); 703 done = 1; 704 unlock_page(page); 705 break; 706 } 707 if (next && (page->index != next)) { 708 dout("not consecutive %p\n", page); 709 unlock_page(page); 710 break; 711 } 712 if (wbc->sync_mode != WB_SYNC_NONE) { 713 dout("waiting on writeback %p\n", page); 714 wait_on_page_writeback(page); 715 } 716 if ((snap_size && page_offset(page) > snap_size) || 717 (!snap_size && 718 page_offset(page) > i_size_read(inode))) { 719 dout("%p page eof %llu\n", page, snap_size ? 720 snap_size : i_size_read(inode)); 721 done = 1; 722 unlock_page(page); 723 break; 724 } 725 if (PageWriteback(page)) { 726 dout("%p under writeback\n", page); 727 unlock_page(page); 728 break; 729 } 730 731 /* only if matching snap context */ 732 if (snapc != (void *)page->private) { 733 dout("page snapc %p != oldest %p\n", 734 (void *)page->private, snapc); 735 unlock_page(page); 736 if (!locked_pages) 737 continue; /* keep looking for snap */ 738 break; 739 } 740 741 if (!clear_page_dirty_for_io(page)) { 742 dout("%p !clear_page_dirty_for_io\n", page); 743 unlock_page(page); 744 break; 745 } 746 747 /* ok */ 748 if (locked_pages == 0) { 749 /* prepare async write request */ 750 offset = page->index << PAGE_CACHE_SHIFT; 751 len = wsize; 752 req = ceph_osdc_new_request(&client->osdc, 753 &ci->i_layout, 754 ceph_vino(inode), 755 offset, &len, 756 CEPH_OSD_OP_WRITE, 757 CEPH_OSD_FLAG_WRITE | 758 CEPH_OSD_FLAG_ONDISK, 759 snapc, do_sync, 760 ci->i_truncate_seq, 761 ci->i_truncate_size, 762 &inode->i_mtime, true, 1); 763 max_pages = req->r_num_pages; 764 765 alloc_page_vec(client, req); 766 req->r_callback = writepages_finish; 767 req->r_inode = inode; 768 req->r_wbc = wbc; 769 } 770 771 /* note position of first page in pvec */ 772 if (first < 0) 773 first = i; 774 dout("%p will write page %p idx %lu\n", 775 inode, page, page->index); 776 set_page_writeback(page); 777 req->r_pages[locked_pages] = page; 778 locked_pages++; 779 next = page->index + 1; 780 } 781 782 /* did we get anything? */ 783 if (!locked_pages) 784 goto release_pvec_pages; 785 if (i) { 786 int j; 787 BUG_ON(!locked_pages || first < 0); 788 789 if (pvec_pages && i == pvec_pages && 790 locked_pages < max_pages) { 791 dout("reached end pvec, trying for more\n"); 792 pagevec_reinit(&pvec); 793 goto get_more_pages; 794 } 795 796 /* shift unused pages over in the pvec... we 797 * will need to release them below. */ 798 for (j = i; j < pvec_pages; j++) { 799 dout(" pvec leftover page %p\n", 800 pvec.pages[j]); 801 pvec.pages[j-i+first] = pvec.pages[j]; 802 } 803 pvec.nr -= i-first; 804 } 805 806 /* submit the write */ 807 offset = req->r_pages[0]->index << PAGE_CACHE_SHIFT; 808 len = min((snap_size ? snap_size : i_size_read(inode)) - offset, 809 (u64)locked_pages << PAGE_CACHE_SHIFT); 810 dout("writepages got %d pages at %llu~%llu\n", 811 locked_pages, offset, len); 812 813 /* revise final length, page count */ 814 req->r_num_pages = locked_pages; 815 reqhead = req->r_request->front.iov_base; 816 op = (void *)(reqhead + 1); 817 op->extent.length = cpu_to_le64(len); 818 op->payload_len = cpu_to_le32(len); 819 req->r_request->hdr.data_len = cpu_to_le32(len); 820 821 ceph_osdc_start_request(&client->osdc, req, true); 822 req = NULL; 823 824 /* continue? */ 825 index = next; 826 wbc->nr_to_write -= locked_pages; 827 if (wbc->nr_to_write <= 0) 828 done = 1; 829 830 release_pvec_pages: 831 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr, 832 pvec.nr ? pvec.pages[0] : NULL); 833 pagevec_release(&pvec); 834 835 if (locked_pages && !done) 836 goto retry; 837 } 838 839 if (should_loop && !done) { 840 /* more to do; loop back to beginning of file */ 841 dout("writepages looping back to beginning of file\n"); 842 should_loop = 0; 843 index = 0; 844 goto retry; 845 } 846 847 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) 848 mapping->writeback_index = index; 849 850 out: 851 if (req) 852 ceph_osdc_put_request(req); 853 if (rc > 0) 854 rc = 0; /* vfs expects us to return 0 */ 855 ceph_put_snap_context(snapc); 856 dout("writepages done, rc = %d\n", rc); 857 out_final: 858 return rc; 859 } 860 861 862 863 /* 864 * See if a given @snapc is either writeable, or already written. 865 */ 866 static int context_is_writeable_or_written(struct inode *inode, 867 struct ceph_snap_context *snapc) 868 { 869 struct ceph_snap_context *oldest = get_oldest_context(inode, NULL); 870 return !oldest || snapc->seq <= oldest->seq; 871 } 872 873 /* 874 * We are only allowed to write into/dirty the page if the page is 875 * clean, or already dirty within the same snap context. 876 */ 877 static int ceph_write_begin(struct file *file, struct address_space *mapping, 878 loff_t pos, unsigned len, unsigned flags, 879 struct page **pagep, void **fsdata) 880 { 881 struct inode *inode = file->f_dentry->d_inode; 882 struct ceph_inode_info *ci = ceph_inode(inode); 883 struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc; 884 struct page *page; 885 pgoff_t index = pos >> PAGE_CACHE_SHIFT; 886 loff_t page_off = pos & PAGE_CACHE_MASK; 887 int pos_in_page = pos & ~PAGE_CACHE_MASK; 888 int end_in_page = pos_in_page + len; 889 loff_t i_size; 890 struct ceph_snap_context *snapc; 891 int r; 892 893 /* get a page*/ 894 retry: 895 page = grab_cache_page_write_begin(mapping, index, 0); 896 if (!page) 897 return -ENOMEM; 898 *pagep = page; 899 900 dout("write_begin file %p inode %p page %p %d~%d\n", file, 901 inode, page, (int)pos, (int)len); 902 903 retry_locked: 904 /* writepages currently holds page lock, but if we change that later, */ 905 wait_on_page_writeback(page); 906 907 /* check snap context */ 908 BUG_ON(!ci->i_snap_realm); 909 down_read(&mdsc->snap_rwsem); 910 BUG_ON(!ci->i_snap_realm->cached_context); 911 if (page->private && 912 (void *)page->private != ci->i_snap_realm->cached_context) { 913 /* 914 * this page is already dirty in another (older) snap 915 * context! is it writeable now? 916 */ 917 snapc = get_oldest_context(inode, NULL); 918 up_read(&mdsc->snap_rwsem); 919 920 if (snapc != (void *)page->private) { 921 dout(" page %p snapc %p not current or oldest\n", 922 page, (void *)page->private); 923 /* 924 * queue for writeback, and wait for snapc to 925 * be writeable or written 926 */ 927 snapc = ceph_get_snap_context((void *)page->private); 928 unlock_page(page); 929 if (ceph_queue_writeback(inode)) 930 igrab(inode); 931 wait_event_interruptible(ci->i_cap_wq, 932 context_is_writeable_or_written(inode, snapc)); 933 ceph_put_snap_context(snapc); 934 goto retry; 935 } 936 937 /* yay, writeable, do it now (without dropping page lock) */ 938 dout(" page %p snapc %p not current, but oldest\n", 939 page, snapc); 940 if (!clear_page_dirty_for_io(page)) 941 goto retry_locked; 942 r = writepage_nounlock(page, NULL); 943 if (r < 0) 944 goto fail_nosnap; 945 goto retry_locked; 946 } 947 948 if (PageUptodate(page)) { 949 dout(" page %p already uptodate\n", page); 950 return 0; 951 } 952 953 /* full page? */ 954 if (pos_in_page == 0 && len == PAGE_CACHE_SIZE) 955 return 0; 956 957 /* past end of file? */ 958 i_size = inode->i_size; /* caller holds i_mutex */ 959 960 if (i_size + len > inode->i_sb->s_maxbytes) { 961 /* file is too big */ 962 r = -EINVAL; 963 goto fail; 964 } 965 966 if (page_off >= i_size || 967 (pos_in_page == 0 && (pos+len) >= i_size && 968 end_in_page - pos_in_page != PAGE_CACHE_SIZE)) { 969 dout(" zeroing %p 0 - %d and %d - %d\n", 970 page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE); 971 zero_user_segments(page, 972 0, pos_in_page, 973 end_in_page, PAGE_CACHE_SIZE); 974 return 0; 975 } 976 977 /* we need to read it. */ 978 up_read(&mdsc->snap_rwsem); 979 r = readpage_nounlock(file, page); 980 if (r < 0) 981 goto fail_nosnap; 982 goto retry_locked; 983 984 fail: 985 up_read(&mdsc->snap_rwsem); 986 fail_nosnap: 987 unlock_page(page); 988 return r; 989 } 990 991 /* 992 * we don't do anything in here that simple_write_end doesn't do 993 * except adjust dirty page accounting and drop read lock on 994 * mdsc->snap_rwsem. 995 */ 996 static int ceph_write_end(struct file *file, struct address_space *mapping, 997 loff_t pos, unsigned len, unsigned copied, 998 struct page *page, void *fsdata) 999 { 1000 struct inode *inode = file->f_dentry->d_inode; 1001 struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc; 1002 unsigned from = pos & (PAGE_CACHE_SIZE - 1); 1003 int check_cap = 0; 1004 1005 dout("write_end file %p inode %p page %p %d~%d (%d)\n", file, 1006 inode, page, (int)pos, (int)copied, (int)len); 1007 1008 /* zero the stale part of the page if we did a short copy */ 1009 if (copied < len) 1010 zero_user_segment(page, from+copied, len); 1011 1012 /* did file size increase? */ 1013 /* (no need for i_size_read(); we caller holds i_mutex */ 1014 if (pos+copied > inode->i_size) 1015 check_cap = ceph_inode_set_size(inode, pos+copied); 1016 1017 if (!PageUptodate(page)) 1018 SetPageUptodate(page); 1019 1020 set_page_dirty(page); 1021 1022 unlock_page(page); 1023 up_read(&mdsc->snap_rwsem); 1024 page_cache_release(page); 1025 1026 if (check_cap) 1027 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL); 1028 1029 return copied; 1030 } 1031 1032 /* 1033 * we set .direct_IO to indicate direct io is supported, but since we 1034 * intercept O_DIRECT reads and writes early, this function should 1035 * never get called. 1036 */ 1037 static ssize_t ceph_direct_io(int rw, struct kiocb *iocb, 1038 const struct iovec *iov, 1039 loff_t pos, unsigned long nr_segs) 1040 { 1041 WARN_ON(1); 1042 return -EINVAL; 1043 } 1044 1045 const struct address_space_operations ceph_aops = { 1046 .readpage = ceph_readpage, 1047 .readpages = ceph_readpages, 1048 .writepage = ceph_writepage, 1049 .writepages = ceph_writepages_start, 1050 .write_begin = ceph_write_begin, 1051 .write_end = ceph_write_end, 1052 .set_page_dirty = ceph_set_page_dirty, 1053 .invalidatepage = ceph_invalidatepage, 1054 .releasepage = ceph_releasepage, 1055 .direct_IO = ceph_direct_io, 1056 }; 1057 1058 1059 /* 1060 * vm ops 1061 */ 1062 1063 /* 1064 * Reuse write_begin here for simplicity. 1065 */ 1066 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf) 1067 { 1068 struct inode *inode = vma->vm_file->f_dentry->d_inode; 1069 struct page *page = vmf->page; 1070 struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc; 1071 loff_t off = page->index << PAGE_CACHE_SHIFT; 1072 loff_t size, len; 1073 struct page *locked_page = NULL; 1074 void *fsdata = NULL; 1075 int ret; 1076 1077 size = i_size_read(inode); 1078 if (off + PAGE_CACHE_SIZE <= size) 1079 len = PAGE_CACHE_SIZE; 1080 else 1081 len = size & ~PAGE_CACHE_MASK; 1082 1083 dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode, 1084 off, len, page, page->index); 1085 ret = ceph_write_begin(vma->vm_file, inode->i_mapping, off, len, 0, 1086 &locked_page, &fsdata); 1087 WARN_ON(page != locked_page); 1088 if (!ret) { 1089 /* 1090 * doing the following, instead of calling 1091 * ceph_write_end. Note that we keep the 1092 * page locked 1093 */ 1094 set_page_dirty(page); 1095 up_read(&mdsc->snap_rwsem); 1096 page_cache_release(page); 1097 ret = VM_FAULT_LOCKED; 1098 } else { 1099 ret = VM_FAULT_SIGBUS; 1100 } 1101 dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret); 1102 return ret; 1103 } 1104 1105 static struct vm_operations_struct ceph_vmops = { 1106 .fault = filemap_fault, 1107 .page_mkwrite = ceph_page_mkwrite, 1108 }; 1109 1110 int ceph_mmap(struct file *file, struct vm_area_struct *vma) 1111 { 1112 struct address_space *mapping = file->f_mapping; 1113 1114 if (!mapping->a_ops->readpage) 1115 return -ENOEXEC; 1116 file_accessed(file); 1117 vma->vm_ops = &ceph_vmops; 1118 vma->vm_flags |= VM_CAN_NONLINEAR; 1119 return 0; 1120 } 1121