1 #include <linux/ceph/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/slab.h> 9 #include <linux/pagevec.h> 10 #include <linux/task_io_accounting_ops.h> 11 12 #include "super.h" 13 #include "mds_client.h" 14 #include <linux/ceph/osd_client.h> 15 16 /* 17 * Ceph address space ops. 18 * 19 * There are a few funny things going on here. 20 * 21 * The page->private field is used to reference a struct 22 * ceph_snap_context for _every_ dirty page. This indicates which 23 * snapshot the page was logically dirtied in, and thus which snap 24 * context needs to be associated with the osd write during writeback. 25 * 26 * Similarly, struct ceph_inode_info maintains a set of counters to 27 * count dirty pages on the inode. In the absence of snapshots, 28 * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count. 29 * 30 * When a snapshot is taken (that is, when the client receives 31 * notification that a snapshot was taken), each inode with caps and 32 * with dirty pages (dirty pages implies there is a cap) gets a new 33 * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending 34 * order, new snaps go to the tail). The i_wrbuffer_ref_head count is 35 * moved to capsnap->dirty. (Unless a sync write is currently in 36 * progress. In that case, the capsnap is said to be "pending", new 37 * writes cannot start, and the capsnap isn't "finalized" until the 38 * write completes (or fails) and a final size/mtime for the inode for 39 * that snap can be settled upon.) i_wrbuffer_ref_head is reset to 0. 40 * 41 * On writeback, we must submit writes to the osd IN SNAP ORDER. So, 42 * we look for the first capsnap in i_cap_snaps and write out pages in 43 * that snap context _only_. Then we move on to the next capsnap, 44 * eventually reaching the "live" or "head" context (i.e., pages that 45 * are not yet snapped) and are writing the most recently dirtied 46 * pages. 47 * 48 * Invalidate and so forth must take care to ensure the dirty page 49 * accounting is preserved. 50 */ 51 52 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10)) 53 #define CONGESTION_OFF_THRESH(congestion_kb) \ 54 (CONGESTION_ON_THRESH(congestion_kb) - \ 55 (CONGESTION_ON_THRESH(congestion_kb) >> 2)) 56 57 58 59 /* 60 * Dirty a page. Optimistically adjust accounting, on the assumption 61 * that we won't race with invalidate. If we do, readjust. 62 */ 63 static int ceph_set_page_dirty(struct page *page) 64 { 65 struct address_space *mapping = page->mapping; 66 struct inode *inode; 67 struct ceph_inode_info *ci; 68 int undo = 0; 69 struct ceph_snap_context *snapc; 70 71 if (unlikely(!mapping)) 72 return !TestSetPageDirty(page); 73 74 if (TestSetPageDirty(page)) { 75 dout("%p set_page_dirty %p idx %lu -- already dirty\n", 76 mapping->host, page, page->index); 77 return 0; 78 } 79 80 inode = mapping->host; 81 ci = ceph_inode(inode); 82 83 /* 84 * Note that we're grabbing a snapc ref here without holding 85 * any locks! 86 */ 87 snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context); 88 89 /* dirty the head */ 90 spin_lock(&inode->i_lock); 91 if (ci->i_head_snapc == NULL) 92 ci->i_head_snapc = ceph_get_snap_context(snapc); 93 ++ci->i_wrbuffer_ref_head; 94 if (ci->i_wrbuffer_ref == 0) 95 ihold(inode); 96 ++ci->i_wrbuffer_ref; 97 dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d " 98 "snapc %p seq %lld (%d snaps)\n", 99 mapping->host, page, page->index, 100 ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1, 101 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head, 102 snapc, snapc->seq, snapc->num_snaps); 103 spin_unlock(&inode->i_lock); 104 105 /* now adjust page */ 106 spin_lock_irq(&mapping->tree_lock); 107 if (page->mapping) { /* Race with truncate? */ 108 WARN_ON_ONCE(!PageUptodate(page)); 109 account_page_dirtied(page, page->mapping); 110 radix_tree_tag_set(&mapping->page_tree, 111 page_index(page), PAGECACHE_TAG_DIRTY); 112 113 /* 114 * Reference snap context in page->private. Also set 115 * PagePrivate so that we get invalidatepage callback. 116 */ 117 page->private = (unsigned long)snapc; 118 SetPagePrivate(page); 119 } else { 120 dout("ANON set_page_dirty %p (raced truncate?)\n", page); 121 undo = 1; 122 } 123 124 spin_unlock_irq(&mapping->tree_lock); 125 126 if (undo) 127 /* whoops, we failed to dirty the page */ 128 ceph_put_wrbuffer_cap_refs(ci, 1, snapc); 129 130 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); 131 132 BUG_ON(!PageDirty(page)); 133 return 1; 134 } 135 136 /* 137 * If we are truncating the full page (i.e. offset == 0), adjust the 138 * dirty page counters appropriately. Only called if there is private 139 * data on the page. 140 */ 141 static void ceph_invalidatepage(struct page *page, unsigned long offset) 142 { 143 struct inode *inode; 144 struct ceph_inode_info *ci; 145 struct ceph_snap_context *snapc = (void *)page->private; 146 147 BUG_ON(!PageLocked(page)); 148 BUG_ON(!page->private); 149 BUG_ON(!PagePrivate(page)); 150 BUG_ON(!page->mapping); 151 152 inode = page->mapping->host; 153 154 /* 155 * We can get non-dirty pages here due to races between 156 * set_page_dirty and truncate_complete_page; just spit out a 157 * warning, in case we end up with accounting problems later. 158 */ 159 if (!PageDirty(page)) 160 pr_err("%p invalidatepage %p page not dirty\n", inode, page); 161 162 if (offset == 0) 163 ClearPageChecked(page); 164 165 ci = ceph_inode(inode); 166 if (offset == 0) { 167 dout("%p invalidatepage %p idx %lu full dirty page %lu\n", 168 inode, page, page->index, offset); 169 ceph_put_wrbuffer_cap_refs(ci, 1, snapc); 170 ceph_put_snap_context(snapc); 171 page->private = 0; 172 ClearPagePrivate(page); 173 } else { 174 dout("%p invalidatepage %p idx %lu partial dirty page\n", 175 inode, page, page->index); 176 } 177 } 178 179 /* just a sanity check */ 180 static int ceph_releasepage(struct page *page, gfp_t g) 181 { 182 struct inode *inode = page->mapping ? page->mapping->host : NULL; 183 dout("%p releasepage %p idx %lu\n", inode, page, page->index); 184 WARN_ON(PageDirty(page)); 185 WARN_ON(page->private); 186 WARN_ON(PagePrivate(page)); 187 return 0; 188 } 189 190 /* 191 * read a single page, without unlocking it. 192 */ 193 static int readpage_nounlock(struct file *filp, struct page *page) 194 { 195 struct inode *inode = filp->f_dentry->d_inode; 196 struct ceph_inode_info *ci = ceph_inode(inode); 197 struct ceph_osd_client *osdc = 198 &ceph_inode_to_client(inode)->client->osdc; 199 int err = 0; 200 u64 len = PAGE_CACHE_SIZE; 201 202 dout("readpage inode %p file %p page %p index %lu\n", 203 inode, filp, page, page->index); 204 err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout, 205 page->index << PAGE_CACHE_SHIFT, &len, 206 ci->i_truncate_seq, ci->i_truncate_size, 207 &page, 1, 0); 208 if (err == -ENOENT) 209 err = 0; 210 if (err < 0) { 211 SetPageError(page); 212 goto out; 213 } else if (err < PAGE_CACHE_SIZE) { 214 /* zero fill remainder of page */ 215 zero_user_segment(page, err, PAGE_CACHE_SIZE); 216 } 217 SetPageUptodate(page); 218 219 out: 220 return err < 0 ? err : 0; 221 } 222 223 static int ceph_readpage(struct file *filp, struct page *page) 224 { 225 int r = readpage_nounlock(filp, page); 226 unlock_page(page); 227 return r; 228 } 229 230 /* 231 * Finish an async read(ahead) op. 232 */ 233 static void finish_read(struct ceph_osd_request *req, struct ceph_msg *msg) 234 { 235 struct inode *inode = req->r_inode; 236 struct ceph_osd_reply_head *replyhead; 237 int rc, bytes; 238 int i; 239 240 /* parse reply */ 241 replyhead = msg->front.iov_base; 242 WARN_ON(le32_to_cpu(replyhead->num_ops) == 0); 243 rc = le32_to_cpu(replyhead->result); 244 bytes = le32_to_cpu(msg->hdr.data_len); 245 246 dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes); 247 248 /* unlock all pages, zeroing any data we didn't read */ 249 for (i = 0; i < req->r_num_pages; i++, bytes -= PAGE_CACHE_SIZE) { 250 struct page *page = req->r_pages[i]; 251 252 if (bytes < (int)PAGE_CACHE_SIZE) { 253 /* zero (remainder of) page */ 254 int s = bytes < 0 ? 0 : bytes; 255 zero_user_segment(page, s, PAGE_CACHE_SIZE); 256 } 257 dout("finish_read %p uptodate %p idx %lu\n", inode, page, 258 page->index); 259 flush_dcache_page(page); 260 SetPageUptodate(page); 261 unlock_page(page); 262 page_cache_release(page); 263 } 264 kfree(req->r_pages); 265 } 266 267 /* 268 * start an async read(ahead) operation. return nr_pages we submitted 269 * a read for on success, or negative error code. 270 */ 271 static int start_read(struct inode *inode, struct list_head *page_list) 272 { 273 struct ceph_osd_client *osdc = 274 &ceph_inode_to_client(inode)->client->osdc; 275 struct ceph_inode_info *ci = ceph_inode(inode); 276 struct page *page = list_entry(page_list->prev, struct page, lru); 277 struct ceph_osd_request *req; 278 u64 off; 279 u64 len; 280 int i; 281 struct page **pages; 282 pgoff_t next_index; 283 int nr_pages = 0; 284 int ret; 285 286 off = page->index << PAGE_CACHE_SHIFT; 287 288 /* count pages */ 289 next_index = page->index; 290 list_for_each_entry_reverse(page, page_list, lru) { 291 if (page->index != next_index) 292 break; 293 nr_pages++; 294 next_index++; 295 } 296 len = nr_pages << PAGE_CACHE_SHIFT; 297 dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages, 298 off, len); 299 300 req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode), 301 off, &len, 302 CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, 303 NULL, 0, 304 ci->i_truncate_seq, ci->i_truncate_size, 305 NULL, false, 1, 0); 306 if (!req) 307 return -ENOMEM; 308 309 /* build page vector */ 310 nr_pages = len >> PAGE_CACHE_SHIFT; 311 pages = kmalloc(sizeof(*pages) * nr_pages, GFP_NOFS); 312 ret = -ENOMEM; 313 if (!pages) 314 goto out; 315 for (i = 0; i < nr_pages; ++i) { 316 page = list_entry(page_list->prev, struct page, lru); 317 BUG_ON(PageLocked(page)); 318 list_del(&page->lru); 319 320 dout("start_read %p adding %p idx %lu\n", inode, page, 321 page->index); 322 if (add_to_page_cache_lru(page, &inode->i_data, page->index, 323 GFP_NOFS)) { 324 page_cache_release(page); 325 dout("start_read %p add_to_page_cache failed %p\n", 326 inode, page); 327 nr_pages = i; 328 goto out_pages; 329 } 330 pages[i] = page; 331 } 332 req->r_pages = pages; 333 req->r_num_pages = nr_pages; 334 req->r_callback = finish_read; 335 req->r_inode = inode; 336 337 dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len); 338 ret = ceph_osdc_start_request(osdc, req, false); 339 if (ret < 0) 340 goto out_pages; 341 ceph_osdc_put_request(req); 342 return nr_pages; 343 344 out_pages: 345 ceph_release_page_vector(pages, nr_pages); 346 kfree(pages); 347 out: 348 ceph_osdc_put_request(req); 349 return ret; 350 } 351 352 353 /* 354 * Read multiple pages. Leave pages we don't read + unlock in page_list; 355 * the caller (VM) cleans them up. 356 */ 357 static int ceph_readpages(struct file *file, struct address_space *mapping, 358 struct list_head *page_list, unsigned nr_pages) 359 { 360 struct inode *inode = file->f_dentry->d_inode; 361 int rc = 0; 362 363 dout("readpages %p file %p nr_pages %d\n", inode, file, nr_pages); 364 while (!list_empty(page_list)) { 365 rc = start_read(inode, page_list); 366 if (rc < 0) 367 goto out; 368 BUG_ON(rc == 0); 369 } 370 out: 371 dout("readpages %p file %p ret %d\n", inode, file, rc); 372 return rc; 373 } 374 375 /* 376 * Get ref for the oldest snapc for an inode with dirty data... that is, the 377 * only snap context we are allowed to write back. 378 */ 379 static struct ceph_snap_context *get_oldest_context(struct inode *inode, 380 u64 *snap_size) 381 { 382 struct ceph_inode_info *ci = ceph_inode(inode); 383 struct ceph_snap_context *snapc = NULL; 384 struct ceph_cap_snap *capsnap = NULL; 385 386 spin_lock(&inode->i_lock); 387 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { 388 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap, 389 capsnap->context, capsnap->dirty_pages); 390 if (capsnap->dirty_pages) { 391 snapc = ceph_get_snap_context(capsnap->context); 392 if (snap_size) 393 *snap_size = capsnap->size; 394 break; 395 } 396 } 397 if (!snapc && ci->i_wrbuffer_ref_head) { 398 snapc = ceph_get_snap_context(ci->i_head_snapc); 399 dout(" head snapc %p has %d dirty pages\n", 400 snapc, ci->i_wrbuffer_ref_head); 401 } 402 spin_unlock(&inode->i_lock); 403 return snapc; 404 } 405 406 /* 407 * Write a single page, but leave the page locked. 408 * 409 * If we get a write error, set the page error bit, but still adjust the 410 * dirty page accounting (i.e., page is no longer dirty). 411 */ 412 static int writepage_nounlock(struct page *page, struct writeback_control *wbc) 413 { 414 struct inode *inode; 415 struct ceph_inode_info *ci; 416 struct ceph_fs_client *fsc; 417 struct ceph_osd_client *osdc; 418 loff_t page_off = page->index << PAGE_CACHE_SHIFT; 419 int len = PAGE_CACHE_SIZE; 420 loff_t i_size; 421 int err = 0; 422 struct ceph_snap_context *snapc, *oldest; 423 u64 snap_size = 0; 424 long writeback_stat; 425 426 dout("writepage %p idx %lu\n", page, page->index); 427 428 if (!page->mapping || !page->mapping->host) { 429 dout("writepage %p - no mapping\n", page); 430 return -EFAULT; 431 } 432 inode = page->mapping->host; 433 ci = ceph_inode(inode); 434 fsc = ceph_inode_to_client(inode); 435 osdc = &fsc->client->osdc; 436 437 /* verify this is a writeable snap context */ 438 snapc = (void *)page->private; 439 if (snapc == NULL) { 440 dout("writepage %p page %p not dirty?\n", inode, page); 441 goto out; 442 } 443 oldest = get_oldest_context(inode, &snap_size); 444 if (snapc->seq > oldest->seq) { 445 dout("writepage %p page %p snapc %p not writeable - noop\n", 446 inode, page, (void *)page->private); 447 /* we should only noop if called by kswapd */ 448 WARN_ON((current->flags & PF_MEMALLOC) == 0); 449 ceph_put_snap_context(oldest); 450 goto out; 451 } 452 ceph_put_snap_context(oldest); 453 454 /* is this a partial page at end of file? */ 455 if (snap_size) 456 i_size = snap_size; 457 else 458 i_size = i_size_read(inode); 459 if (i_size < page_off + len) 460 len = i_size - page_off; 461 462 dout("writepage %p page %p index %lu on %llu~%u snapc %p\n", 463 inode, page, page->index, page_off, len, snapc); 464 465 writeback_stat = atomic_long_inc_return(&fsc->writeback_count); 466 if (writeback_stat > 467 CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb)) 468 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC); 469 470 set_page_writeback(page); 471 err = ceph_osdc_writepages(osdc, ceph_vino(inode), 472 &ci->i_layout, snapc, 473 page_off, len, 474 ci->i_truncate_seq, ci->i_truncate_size, 475 &inode->i_mtime, 476 &page, 1, 0, 0, true); 477 if (err < 0) { 478 dout("writepage setting page/mapping error %d %p\n", err, page); 479 SetPageError(page); 480 mapping_set_error(&inode->i_data, err); 481 if (wbc) 482 wbc->pages_skipped++; 483 } else { 484 dout("writepage cleaned page %p\n", page); 485 err = 0; /* vfs expects us to return 0 */ 486 } 487 page->private = 0; 488 ClearPagePrivate(page); 489 end_page_writeback(page); 490 ceph_put_wrbuffer_cap_refs(ci, 1, snapc); 491 ceph_put_snap_context(snapc); /* page's reference */ 492 out: 493 return err; 494 } 495 496 static int ceph_writepage(struct page *page, struct writeback_control *wbc) 497 { 498 int err; 499 struct inode *inode = page->mapping->host; 500 BUG_ON(!inode); 501 ihold(inode); 502 err = writepage_nounlock(page, wbc); 503 unlock_page(page); 504 iput(inode); 505 return err; 506 } 507 508 509 /* 510 * lame release_pages helper. release_pages() isn't exported to 511 * modules. 512 */ 513 static void ceph_release_pages(struct page **pages, int num) 514 { 515 struct pagevec pvec; 516 int i; 517 518 pagevec_init(&pvec, 0); 519 for (i = 0; i < num; i++) { 520 if (pagevec_add(&pvec, pages[i]) == 0) 521 pagevec_release(&pvec); 522 } 523 pagevec_release(&pvec); 524 } 525 526 527 /* 528 * async writeback completion handler. 529 * 530 * If we get an error, set the mapping error bit, but not the individual 531 * page error bits. 532 */ 533 static void writepages_finish(struct ceph_osd_request *req, 534 struct ceph_msg *msg) 535 { 536 struct inode *inode = req->r_inode; 537 struct ceph_osd_reply_head *replyhead; 538 struct ceph_osd_op *op; 539 struct ceph_inode_info *ci = ceph_inode(inode); 540 unsigned wrote; 541 struct page *page; 542 int i; 543 struct ceph_snap_context *snapc = req->r_snapc; 544 struct address_space *mapping = inode->i_mapping; 545 __s32 rc = -EIO; 546 u64 bytes = 0; 547 struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 548 long writeback_stat; 549 unsigned issued = ceph_caps_issued(ci); 550 551 /* parse reply */ 552 replyhead = msg->front.iov_base; 553 WARN_ON(le32_to_cpu(replyhead->num_ops) == 0); 554 op = (void *)(replyhead + 1); 555 rc = le32_to_cpu(replyhead->result); 556 bytes = le64_to_cpu(op->extent.length); 557 558 if (rc >= 0) { 559 /* 560 * Assume we wrote the pages we originally sent. The 561 * osd might reply with fewer pages if our writeback 562 * raced with a truncation and was adjusted at the osd, 563 * so don't believe the reply. 564 */ 565 wrote = req->r_num_pages; 566 } else { 567 wrote = 0; 568 mapping_set_error(mapping, rc); 569 } 570 dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n", 571 inode, rc, bytes, wrote); 572 573 /* clean all pages */ 574 for (i = 0; i < req->r_num_pages; i++) { 575 page = req->r_pages[i]; 576 BUG_ON(!page); 577 WARN_ON(!PageUptodate(page)); 578 579 writeback_stat = 580 atomic_long_dec_return(&fsc->writeback_count); 581 if (writeback_stat < 582 CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb)) 583 clear_bdi_congested(&fsc->backing_dev_info, 584 BLK_RW_ASYNC); 585 586 ceph_put_snap_context((void *)page->private); 587 page->private = 0; 588 ClearPagePrivate(page); 589 dout("unlocking %d %p\n", i, page); 590 end_page_writeback(page); 591 592 /* 593 * We lost the cache cap, need to truncate the page before 594 * it is unlocked, otherwise we'd truncate it later in the 595 * page truncation thread, possibly losing some data that 596 * raced its way in 597 */ 598 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) 599 generic_error_remove_page(inode->i_mapping, page); 600 601 unlock_page(page); 602 } 603 dout("%p wrote+cleaned %d pages\n", inode, wrote); 604 ceph_put_wrbuffer_cap_refs(ci, req->r_num_pages, snapc); 605 606 ceph_release_pages(req->r_pages, req->r_num_pages); 607 if (req->r_pages_from_pool) 608 mempool_free(req->r_pages, 609 ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool); 610 else 611 kfree(req->r_pages); 612 ceph_osdc_put_request(req); 613 } 614 615 /* 616 * allocate a page vec, either directly, or if necessary, via a the 617 * mempool. we avoid the mempool if we can because req->r_num_pages 618 * may be less than the maximum write size. 619 */ 620 static void alloc_page_vec(struct ceph_fs_client *fsc, 621 struct ceph_osd_request *req) 622 { 623 req->r_pages = kmalloc(sizeof(struct page *) * req->r_num_pages, 624 GFP_NOFS); 625 if (!req->r_pages) { 626 req->r_pages = mempool_alloc(fsc->wb_pagevec_pool, GFP_NOFS); 627 req->r_pages_from_pool = 1; 628 WARN_ON(!req->r_pages); 629 } 630 } 631 632 /* 633 * initiate async writeback 634 */ 635 static int ceph_writepages_start(struct address_space *mapping, 636 struct writeback_control *wbc) 637 { 638 struct inode *inode = mapping->host; 639 struct ceph_inode_info *ci = ceph_inode(inode); 640 struct ceph_fs_client *fsc; 641 pgoff_t index, start, end; 642 int range_whole = 0; 643 int should_loop = 1; 644 pgoff_t max_pages = 0, max_pages_ever = 0; 645 struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc; 646 struct pagevec pvec; 647 int done = 0; 648 int rc = 0; 649 unsigned wsize = 1 << inode->i_blkbits; 650 struct ceph_osd_request *req = NULL; 651 int do_sync; 652 u64 snap_size = 0; 653 654 /* 655 * Include a 'sync' in the OSD request if this is a data 656 * integrity write (e.g., O_SYNC write or fsync()), or if our 657 * cap is being revoked. 658 */ 659 do_sync = wbc->sync_mode == WB_SYNC_ALL; 660 if (ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER)) 661 do_sync = 1; 662 dout("writepages_start %p dosync=%d (mode=%s)\n", 663 inode, do_sync, 664 wbc->sync_mode == WB_SYNC_NONE ? "NONE" : 665 (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD")); 666 667 fsc = ceph_inode_to_client(inode); 668 if (fsc->mount_state == CEPH_MOUNT_SHUTDOWN) { 669 pr_warning("writepage_start %p on forced umount\n", inode); 670 return -EIO; /* we're in a forced umount, don't write! */ 671 } 672 if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize) 673 wsize = fsc->mount_options->wsize; 674 if (wsize < PAGE_CACHE_SIZE) 675 wsize = PAGE_CACHE_SIZE; 676 max_pages_ever = wsize >> PAGE_CACHE_SHIFT; 677 678 pagevec_init(&pvec, 0); 679 680 /* where to start/end? */ 681 if (wbc->range_cyclic) { 682 start = mapping->writeback_index; /* Start from prev offset */ 683 end = -1; 684 dout(" cyclic, start at %lu\n", start); 685 } else { 686 start = wbc->range_start >> PAGE_CACHE_SHIFT; 687 end = wbc->range_end >> PAGE_CACHE_SHIFT; 688 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX) 689 range_whole = 1; 690 should_loop = 0; 691 dout(" not cyclic, %lu to %lu\n", start, end); 692 } 693 index = start; 694 695 retry: 696 /* find oldest snap context with dirty data */ 697 ceph_put_snap_context(snapc); 698 snapc = get_oldest_context(inode, &snap_size); 699 if (!snapc) { 700 /* hmm, why does writepages get called when there 701 is no dirty data? */ 702 dout(" no snap context with dirty data?\n"); 703 goto out; 704 } 705 dout(" oldest snapc is %p seq %lld (%d snaps)\n", 706 snapc, snapc->seq, snapc->num_snaps); 707 if (last_snapc && snapc != last_snapc) { 708 /* if we switched to a newer snapc, restart our scan at the 709 * start of the original file range. */ 710 dout(" snapc differs from last pass, restarting at %lu\n", 711 index); 712 index = start; 713 } 714 last_snapc = snapc; 715 716 while (!done && index <= end) { 717 unsigned i; 718 int first; 719 pgoff_t next; 720 int pvec_pages, locked_pages; 721 struct page *page; 722 int want; 723 u64 offset, len; 724 struct ceph_osd_request_head *reqhead; 725 struct ceph_osd_op *op; 726 long writeback_stat; 727 728 next = 0; 729 locked_pages = 0; 730 max_pages = max_pages_ever; 731 732 get_more_pages: 733 first = -1; 734 want = min(end - index, 735 min((pgoff_t)PAGEVEC_SIZE, 736 max_pages - (pgoff_t)locked_pages) - 1) 737 + 1; 738 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index, 739 PAGECACHE_TAG_DIRTY, 740 want); 741 dout("pagevec_lookup_tag got %d\n", pvec_pages); 742 if (!pvec_pages && !locked_pages) 743 break; 744 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) { 745 page = pvec.pages[i]; 746 dout("? %p idx %lu\n", page, page->index); 747 if (locked_pages == 0) 748 lock_page(page); /* first page */ 749 else if (!trylock_page(page)) 750 break; 751 752 /* only dirty pages, or our accounting breaks */ 753 if (unlikely(!PageDirty(page)) || 754 unlikely(page->mapping != mapping)) { 755 dout("!dirty or !mapping %p\n", page); 756 unlock_page(page); 757 break; 758 } 759 if (!wbc->range_cyclic && page->index > end) { 760 dout("end of range %p\n", page); 761 done = 1; 762 unlock_page(page); 763 break; 764 } 765 if (next && (page->index != next)) { 766 dout("not consecutive %p\n", page); 767 unlock_page(page); 768 break; 769 } 770 if (wbc->sync_mode != WB_SYNC_NONE) { 771 dout("waiting on writeback %p\n", page); 772 wait_on_page_writeback(page); 773 } 774 if ((snap_size && page_offset(page) > snap_size) || 775 (!snap_size && 776 page_offset(page) > i_size_read(inode))) { 777 dout("%p page eof %llu\n", page, snap_size ? 778 snap_size : i_size_read(inode)); 779 done = 1; 780 unlock_page(page); 781 break; 782 } 783 if (PageWriteback(page)) { 784 dout("%p under writeback\n", page); 785 unlock_page(page); 786 break; 787 } 788 789 /* only if matching snap context */ 790 pgsnapc = (void *)page->private; 791 if (pgsnapc->seq > snapc->seq) { 792 dout("page snapc %p %lld > oldest %p %lld\n", 793 pgsnapc, pgsnapc->seq, snapc, snapc->seq); 794 unlock_page(page); 795 if (!locked_pages) 796 continue; /* keep looking for snap */ 797 break; 798 } 799 800 if (!clear_page_dirty_for_io(page)) { 801 dout("%p !clear_page_dirty_for_io\n", page); 802 unlock_page(page); 803 break; 804 } 805 806 /* ok */ 807 if (locked_pages == 0) { 808 /* prepare async write request */ 809 offset = (unsigned long long)page->index 810 << PAGE_CACHE_SHIFT; 811 len = wsize; 812 req = ceph_osdc_new_request(&fsc->client->osdc, 813 &ci->i_layout, 814 ceph_vino(inode), 815 offset, &len, 816 CEPH_OSD_OP_WRITE, 817 CEPH_OSD_FLAG_WRITE | 818 CEPH_OSD_FLAG_ONDISK, 819 snapc, do_sync, 820 ci->i_truncate_seq, 821 ci->i_truncate_size, 822 &inode->i_mtime, true, 1, 0); 823 824 if (!req) { 825 rc = -ENOMEM; 826 unlock_page(page); 827 break; 828 } 829 830 max_pages = req->r_num_pages; 831 832 alloc_page_vec(fsc, req); 833 req->r_callback = writepages_finish; 834 req->r_inode = inode; 835 } 836 837 /* note position of first page in pvec */ 838 if (first < 0) 839 first = i; 840 dout("%p will write page %p idx %lu\n", 841 inode, page, page->index); 842 843 writeback_stat = 844 atomic_long_inc_return(&fsc->writeback_count); 845 if (writeback_stat > CONGESTION_ON_THRESH( 846 fsc->mount_options->congestion_kb)) { 847 set_bdi_congested(&fsc->backing_dev_info, 848 BLK_RW_ASYNC); 849 } 850 851 set_page_writeback(page); 852 req->r_pages[locked_pages] = page; 853 locked_pages++; 854 next = page->index + 1; 855 } 856 857 /* did we get anything? */ 858 if (!locked_pages) 859 goto release_pvec_pages; 860 if (i) { 861 int j; 862 BUG_ON(!locked_pages || first < 0); 863 864 if (pvec_pages && i == pvec_pages && 865 locked_pages < max_pages) { 866 dout("reached end pvec, trying for more\n"); 867 pagevec_reinit(&pvec); 868 goto get_more_pages; 869 } 870 871 /* shift unused pages over in the pvec... we 872 * will need to release them below. */ 873 for (j = i; j < pvec_pages; j++) { 874 dout(" pvec leftover page %p\n", 875 pvec.pages[j]); 876 pvec.pages[j-i+first] = pvec.pages[j]; 877 } 878 pvec.nr -= i-first; 879 } 880 881 /* submit the write */ 882 offset = req->r_pages[0]->index << PAGE_CACHE_SHIFT; 883 len = min((snap_size ? snap_size : i_size_read(inode)) - offset, 884 (u64)locked_pages << PAGE_CACHE_SHIFT); 885 dout("writepages got %d pages at %llu~%llu\n", 886 locked_pages, offset, len); 887 888 /* revise final length, page count */ 889 req->r_num_pages = locked_pages; 890 reqhead = req->r_request->front.iov_base; 891 op = (void *)(reqhead + 1); 892 op->extent.length = cpu_to_le64(len); 893 op->payload_len = cpu_to_le32(len); 894 req->r_request->hdr.data_len = cpu_to_le32(len); 895 896 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true); 897 BUG_ON(rc); 898 req = NULL; 899 900 /* continue? */ 901 index = next; 902 wbc->nr_to_write -= locked_pages; 903 if (wbc->nr_to_write <= 0) 904 done = 1; 905 906 release_pvec_pages: 907 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr, 908 pvec.nr ? pvec.pages[0] : NULL); 909 pagevec_release(&pvec); 910 911 if (locked_pages && !done) 912 goto retry; 913 } 914 915 if (should_loop && !done) { 916 /* more to do; loop back to beginning of file */ 917 dout("writepages looping back to beginning of file\n"); 918 should_loop = 0; 919 index = 0; 920 goto retry; 921 } 922 923 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) 924 mapping->writeback_index = index; 925 926 out: 927 if (req) 928 ceph_osdc_put_request(req); 929 ceph_put_snap_context(snapc); 930 dout("writepages done, rc = %d\n", rc); 931 return rc; 932 } 933 934 935 936 /* 937 * See if a given @snapc is either writeable, or already written. 938 */ 939 static int context_is_writeable_or_written(struct inode *inode, 940 struct ceph_snap_context *snapc) 941 { 942 struct ceph_snap_context *oldest = get_oldest_context(inode, NULL); 943 int ret = !oldest || snapc->seq <= oldest->seq; 944 945 ceph_put_snap_context(oldest); 946 return ret; 947 } 948 949 /* 950 * We are only allowed to write into/dirty the page if the page is 951 * clean, or already dirty within the same snap context. 952 * 953 * called with page locked. 954 * return success with page locked, 955 * or any failure (incl -EAGAIN) with page unlocked. 956 */ 957 static int ceph_update_writeable_page(struct file *file, 958 loff_t pos, unsigned len, 959 struct page *page) 960 { 961 struct inode *inode = file->f_dentry->d_inode; 962 struct ceph_inode_info *ci = ceph_inode(inode); 963 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc; 964 loff_t page_off = pos & PAGE_CACHE_MASK; 965 int pos_in_page = pos & ~PAGE_CACHE_MASK; 966 int end_in_page = pos_in_page + len; 967 loff_t i_size; 968 int r; 969 struct ceph_snap_context *snapc, *oldest; 970 971 retry_locked: 972 /* writepages currently holds page lock, but if we change that later, */ 973 wait_on_page_writeback(page); 974 975 /* check snap context */ 976 BUG_ON(!ci->i_snap_realm); 977 down_read(&mdsc->snap_rwsem); 978 BUG_ON(!ci->i_snap_realm->cached_context); 979 snapc = (void *)page->private; 980 if (snapc && snapc != ci->i_head_snapc) { 981 /* 982 * this page is already dirty in another (older) snap 983 * context! is it writeable now? 984 */ 985 oldest = get_oldest_context(inode, NULL); 986 up_read(&mdsc->snap_rwsem); 987 988 if (snapc->seq > oldest->seq) { 989 ceph_put_snap_context(oldest); 990 dout(" page %p snapc %p not current or oldest\n", 991 page, snapc); 992 /* 993 * queue for writeback, and wait for snapc to 994 * be writeable or written 995 */ 996 snapc = ceph_get_snap_context(snapc); 997 unlock_page(page); 998 ceph_queue_writeback(inode); 999 r = wait_event_interruptible(ci->i_cap_wq, 1000 context_is_writeable_or_written(inode, snapc)); 1001 ceph_put_snap_context(snapc); 1002 if (r == -ERESTARTSYS) 1003 return r; 1004 return -EAGAIN; 1005 } 1006 ceph_put_snap_context(oldest); 1007 1008 /* yay, writeable, do it now (without dropping page lock) */ 1009 dout(" page %p snapc %p not current, but oldest\n", 1010 page, snapc); 1011 if (!clear_page_dirty_for_io(page)) 1012 goto retry_locked; 1013 r = writepage_nounlock(page, NULL); 1014 if (r < 0) 1015 goto fail_nosnap; 1016 goto retry_locked; 1017 } 1018 1019 if (PageUptodate(page)) { 1020 dout(" page %p already uptodate\n", page); 1021 return 0; 1022 } 1023 1024 /* full page? */ 1025 if (pos_in_page == 0 && len == PAGE_CACHE_SIZE) 1026 return 0; 1027 1028 /* past end of file? */ 1029 i_size = inode->i_size; /* caller holds i_mutex */ 1030 1031 if (i_size + len > inode->i_sb->s_maxbytes) { 1032 /* file is too big */ 1033 r = -EINVAL; 1034 goto fail; 1035 } 1036 1037 if (page_off >= i_size || 1038 (pos_in_page == 0 && (pos+len) >= i_size && 1039 end_in_page - pos_in_page != PAGE_CACHE_SIZE)) { 1040 dout(" zeroing %p 0 - %d and %d - %d\n", 1041 page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE); 1042 zero_user_segments(page, 1043 0, pos_in_page, 1044 end_in_page, PAGE_CACHE_SIZE); 1045 return 0; 1046 } 1047 1048 /* we need to read it. */ 1049 up_read(&mdsc->snap_rwsem); 1050 r = readpage_nounlock(file, page); 1051 if (r < 0) 1052 goto fail_nosnap; 1053 goto retry_locked; 1054 1055 fail: 1056 up_read(&mdsc->snap_rwsem); 1057 fail_nosnap: 1058 unlock_page(page); 1059 return r; 1060 } 1061 1062 /* 1063 * We are only allowed to write into/dirty the page if the page is 1064 * clean, or already dirty within the same snap context. 1065 */ 1066 static int ceph_write_begin(struct file *file, struct address_space *mapping, 1067 loff_t pos, unsigned len, unsigned flags, 1068 struct page **pagep, void **fsdata) 1069 { 1070 struct inode *inode = file->f_dentry->d_inode; 1071 struct page *page; 1072 pgoff_t index = pos >> PAGE_CACHE_SHIFT; 1073 int r; 1074 1075 do { 1076 /* get a page */ 1077 page = grab_cache_page_write_begin(mapping, index, 0); 1078 if (!page) 1079 return -ENOMEM; 1080 *pagep = page; 1081 1082 dout("write_begin file %p inode %p page %p %d~%d\n", file, 1083 inode, page, (int)pos, (int)len); 1084 1085 r = ceph_update_writeable_page(file, pos, len, page); 1086 } while (r == -EAGAIN); 1087 1088 return r; 1089 } 1090 1091 /* 1092 * we don't do anything in here that simple_write_end doesn't do 1093 * except adjust dirty page accounting and drop read lock on 1094 * mdsc->snap_rwsem. 1095 */ 1096 static int ceph_write_end(struct file *file, struct address_space *mapping, 1097 loff_t pos, unsigned len, unsigned copied, 1098 struct page *page, void *fsdata) 1099 { 1100 struct inode *inode = file->f_dentry->d_inode; 1101 struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 1102 struct ceph_mds_client *mdsc = fsc->mdsc; 1103 unsigned from = pos & (PAGE_CACHE_SIZE - 1); 1104 int check_cap = 0; 1105 1106 dout("write_end file %p inode %p page %p %d~%d (%d)\n", file, 1107 inode, page, (int)pos, (int)copied, (int)len); 1108 1109 /* zero the stale part of the page if we did a short copy */ 1110 if (copied < len) 1111 zero_user_segment(page, from+copied, len); 1112 1113 /* did file size increase? */ 1114 /* (no need for i_size_read(); we caller holds i_mutex */ 1115 if (pos+copied > inode->i_size) 1116 check_cap = ceph_inode_set_size(inode, pos+copied); 1117 1118 if (!PageUptodate(page)) 1119 SetPageUptodate(page); 1120 1121 set_page_dirty(page); 1122 1123 unlock_page(page); 1124 up_read(&mdsc->snap_rwsem); 1125 page_cache_release(page); 1126 1127 if (check_cap) 1128 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL); 1129 1130 return copied; 1131 } 1132 1133 /* 1134 * we set .direct_IO to indicate direct io is supported, but since we 1135 * intercept O_DIRECT reads and writes early, this function should 1136 * never get called. 1137 */ 1138 static ssize_t ceph_direct_io(int rw, struct kiocb *iocb, 1139 const struct iovec *iov, 1140 loff_t pos, unsigned long nr_segs) 1141 { 1142 WARN_ON(1); 1143 return -EINVAL; 1144 } 1145 1146 const struct address_space_operations ceph_aops = { 1147 .readpage = ceph_readpage, 1148 .readpages = ceph_readpages, 1149 .writepage = ceph_writepage, 1150 .writepages = ceph_writepages_start, 1151 .write_begin = ceph_write_begin, 1152 .write_end = ceph_write_end, 1153 .set_page_dirty = ceph_set_page_dirty, 1154 .invalidatepage = ceph_invalidatepage, 1155 .releasepage = ceph_releasepage, 1156 .direct_IO = ceph_direct_io, 1157 }; 1158 1159 1160 /* 1161 * vm ops 1162 */ 1163 1164 /* 1165 * Reuse write_begin here for simplicity. 1166 */ 1167 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf) 1168 { 1169 struct inode *inode = vma->vm_file->f_dentry->d_inode; 1170 struct page *page = vmf->page; 1171 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc; 1172 loff_t off = page->index << PAGE_CACHE_SHIFT; 1173 loff_t size, len; 1174 int ret; 1175 1176 size = i_size_read(inode); 1177 if (off + PAGE_CACHE_SIZE <= size) 1178 len = PAGE_CACHE_SIZE; 1179 else 1180 len = size & ~PAGE_CACHE_MASK; 1181 1182 dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode, 1183 off, len, page, page->index); 1184 1185 lock_page(page); 1186 1187 ret = VM_FAULT_NOPAGE; 1188 if ((off > size) || 1189 (page->mapping != inode->i_mapping)) 1190 goto out; 1191 1192 ret = ceph_update_writeable_page(vma->vm_file, off, len, page); 1193 if (ret == 0) { 1194 /* success. we'll keep the page locked. */ 1195 set_page_dirty(page); 1196 up_read(&mdsc->snap_rwsem); 1197 ret = VM_FAULT_LOCKED; 1198 } else { 1199 if (ret == -ENOMEM) 1200 ret = VM_FAULT_OOM; 1201 else 1202 ret = VM_FAULT_SIGBUS; 1203 } 1204 out: 1205 dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret); 1206 if (ret != VM_FAULT_LOCKED) 1207 unlock_page(page); 1208 return ret; 1209 } 1210 1211 static struct vm_operations_struct ceph_vmops = { 1212 .fault = filemap_fault, 1213 .page_mkwrite = ceph_page_mkwrite, 1214 }; 1215 1216 int ceph_mmap(struct file *file, struct vm_area_struct *vma) 1217 { 1218 struct address_space *mapping = file->f_mapping; 1219 1220 if (!mapping->a_ops->readpage) 1221 return -ENOEXEC; 1222 file_accessed(file); 1223 vma->vm_ops = &ceph_vmops; 1224 vma->vm_flags |= VM_CAN_NONLINEAR; 1225 return 0; 1226 } 1227