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 static inline struct ceph_snap_context *page_snap_context(struct page *page) 58 { 59 if (PagePrivate(page)) 60 return (void *)page->private; 61 return NULL; 62 } 63 64 /* 65 * Dirty a page. Optimistically adjust accounting, on the assumption 66 * that we won't race with invalidate. If we do, readjust. 67 */ 68 static int ceph_set_page_dirty(struct page *page) 69 { 70 struct address_space *mapping = page->mapping; 71 struct inode *inode; 72 struct ceph_inode_info *ci; 73 int undo = 0; 74 struct ceph_snap_context *snapc; 75 76 if (unlikely(!mapping)) 77 return !TestSetPageDirty(page); 78 79 if (TestSetPageDirty(page)) { 80 dout("%p set_page_dirty %p idx %lu -- already dirty\n", 81 mapping->host, page, page->index); 82 return 0; 83 } 84 85 inode = mapping->host; 86 ci = ceph_inode(inode); 87 88 /* 89 * Note that we're grabbing a snapc ref here without holding 90 * any locks! 91 */ 92 snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context); 93 94 /* dirty the head */ 95 spin_lock(&ci->i_ceph_lock); 96 if (ci->i_head_snapc == NULL) 97 ci->i_head_snapc = ceph_get_snap_context(snapc); 98 ++ci->i_wrbuffer_ref_head; 99 if (ci->i_wrbuffer_ref == 0) 100 ihold(inode); 101 ++ci->i_wrbuffer_ref; 102 dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d " 103 "snapc %p seq %lld (%d snaps)\n", 104 mapping->host, page, page->index, 105 ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1, 106 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head, 107 snapc, snapc->seq, snapc->num_snaps); 108 spin_unlock(&ci->i_ceph_lock); 109 110 /* now adjust page */ 111 spin_lock_irq(&mapping->tree_lock); 112 if (page->mapping) { /* Race with truncate? */ 113 WARN_ON_ONCE(!PageUptodate(page)); 114 account_page_dirtied(page, page->mapping); 115 radix_tree_tag_set(&mapping->page_tree, 116 page_index(page), PAGECACHE_TAG_DIRTY); 117 118 /* 119 * Reference snap context in page->private. Also set 120 * PagePrivate so that we get invalidatepage callback. 121 */ 122 page->private = (unsigned long)snapc; 123 SetPagePrivate(page); 124 } else { 125 dout("ANON set_page_dirty %p (raced truncate?)\n", page); 126 undo = 1; 127 } 128 129 spin_unlock_irq(&mapping->tree_lock); 130 131 if (undo) 132 /* whoops, we failed to dirty the page */ 133 ceph_put_wrbuffer_cap_refs(ci, 1, snapc); 134 135 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); 136 137 BUG_ON(!PageDirty(page)); 138 return 1; 139 } 140 141 /* 142 * If we are truncating the full page (i.e. offset == 0), adjust the 143 * dirty page counters appropriately. Only called if there is private 144 * data on the page. 145 */ 146 static void ceph_invalidatepage(struct page *page, unsigned long offset) 147 { 148 struct inode *inode; 149 struct ceph_inode_info *ci; 150 struct ceph_snap_context *snapc = page_snap_context(page); 151 152 BUG_ON(!PageLocked(page)); 153 BUG_ON(!PagePrivate(page)); 154 BUG_ON(!page->mapping); 155 156 inode = page->mapping->host; 157 158 /* 159 * We can get non-dirty pages here due to races between 160 * set_page_dirty and truncate_complete_page; just spit out a 161 * warning, in case we end up with accounting problems later. 162 */ 163 if (!PageDirty(page)) 164 pr_err("%p invalidatepage %p page not dirty\n", inode, page); 165 166 if (offset == 0) 167 ClearPageChecked(page); 168 169 ci = ceph_inode(inode); 170 if (offset == 0) { 171 dout("%p invalidatepage %p idx %lu full dirty page %lu\n", 172 inode, page, page->index, offset); 173 ceph_put_wrbuffer_cap_refs(ci, 1, snapc); 174 ceph_put_snap_context(snapc); 175 page->private = 0; 176 ClearPagePrivate(page); 177 } else { 178 dout("%p invalidatepage %p idx %lu partial dirty page\n", 179 inode, page, page->index); 180 } 181 } 182 183 /* just a sanity check */ 184 static int ceph_releasepage(struct page *page, gfp_t g) 185 { 186 struct inode *inode = page->mapping ? page->mapping->host : NULL; 187 dout("%p releasepage %p idx %lu\n", inode, page, page->index); 188 WARN_ON(PageDirty(page)); 189 WARN_ON(PagePrivate(page)); 190 return 0; 191 } 192 193 /* 194 * read a single page, without unlocking it. 195 */ 196 static int readpage_nounlock(struct file *filp, struct page *page) 197 { 198 struct inode *inode = filp->f_dentry->d_inode; 199 struct ceph_inode_info *ci = ceph_inode(inode); 200 struct ceph_osd_client *osdc = 201 &ceph_inode_to_client(inode)->client->osdc; 202 int err = 0; 203 u64 len = PAGE_CACHE_SIZE; 204 205 dout("readpage inode %p file %p page %p index %lu\n", 206 inode, filp, page, page->index); 207 err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout, 208 (u64) page_offset(page), &len, 209 ci->i_truncate_seq, ci->i_truncate_size, 210 &page, 1, 0); 211 if (err == -ENOENT) 212 err = 0; 213 if (err < 0) { 214 SetPageError(page); 215 goto out; 216 } else if (err < PAGE_CACHE_SIZE) { 217 /* zero fill remainder of page */ 218 zero_user_segment(page, err, PAGE_CACHE_SIZE); 219 } 220 SetPageUptodate(page); 221 222 out: 223 return err < 0 ? err : 0; 224 } 225 226 static int ceph_readpage(struct file *filp, struct page *page) 227 { 228 int r = readpage_nounlock(filp, page); 229 unlock_page(page); 230 return r; 231 } 232 233 /* 234 * Finish an async read(ahead) op. 235 */ 236 static void finish_read(struct ceph_osd_request *req, struct ceph_msg *msg) 237 { 238 struct inode *inode = req->r_inode; 239 struct ceph_osd_reply_head *replyhead; 240 int rc, bytes; 241 int i; 242 243 /* parse reply */ 244 replyhead = msg->front.iov_base; 245 WARN_ON(le32_to_cpu(replyhead->num_ops) == 0); 246 rc = le32_to_cpu(replyhead->result); 247 bytes = le32_to_cpu(msg->hdr.data_len); 248 249 dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes); 250 251 /* unlock all pages, zeroing any data we didn't read */ 252 for (i = 0; i < req->r_num_pages; i++, bytes -= PAGE_CACHE_SIZE) { 253 struct page *page = req->r_pages[i]; 254 255 if (bytes < (int)PAGE_CACHE_SIZE) { 256 /* zero (remainder of) page */ 257 int s = bytes < 0 ? 0 : bytes; 258 zero_user_segment(page, s, PAGE_CACHE_SIZE); 259 } 260 dout("finish_read %p uptodate %p idx %lu\n", inode, page, 261 page->index); 262 flush_dcache_page(page); 263 SetPageUptodate(page); 264 unlock_page(page); 265 page_cache_release(page); 266 } 267 kfree(req->r_pages); 268 } 269 270 static void ceph_unlock_page_vector(struct page **pages, int num_pages) 271 { 272 int i; 273 274 for (i = 0; i < num_pages; i++) 275 unlock_page(pages[i]); 276 } 277 278 /* 279 * start an async read(ahead) operation. return nr_pages we submitted 280 * a read for on success, or negative error code. 281 */ 282 static int start_read(struct inode *inode, struct list_head *page_list, int max) 283 { 284 struct ceph_osd_client *osdc = 285 &ceph_inode_to_client(inode)->client->osdc; 286 struct ceph_inode_info *ci = ceph_inode(inode); 287 struct page *page = list_entry(page_list->prev, struct page, lru); 288 struct ceph_osd_request *req; 289 u64 off; 290 u64 len; 291 int i; 292 struct page **pages; 293 pgoff_t next_index; 294 int nr_pages = 0; 295 int ret; 296 297 off = (u64) page_offset(page); 298 299 /* count pages */ 300 next_index = page->index; 301 list_for_each_entry_reverse(page, page_list, lru) { 302 if (page->index != next_index) 303 break; 304 nr_pages++; 305 next_index++; 306 if (max && nr_pages == max) 307 break; 308 } 309 len = nr_pages << PAGE_CACHE_SHIFT; 310 dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages, 311 off, len); 312 313 req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode), 314 off, &len, 315 CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, 316 NULL, 0, 317 ci->i_truncate_seq, ci->i_truncate_size, 318 NULL, false, 1, 0); 319 if (IS_ERR(req)) 320 return PTR_ERR(req); 321 322 /* build page vector */ 323 nr_pages = len >> PAGE_CACHE_SHIFT; 324 pages = kmalloc(sizeof(*pages) * nr_pages, GFP_NOFS); 325 ret = -ENOMEM; 326 if (!pages) 327 goto out; 328 for (i = 0; i < nr_pages; ++i) { 329 page = list_entry(page_list->prev, struct page, lru); 330 BUG_ON(PageLocked(page)); 331 list_del(&page->lru); 332 333 dout("start_read %p adding %p idx %lu\n", inode, page, 334 page->index); 335 if (add_to_page_cache_lru(page, &inode->i_data, page->index, 336 GFP_NOFS)) { 337 page_cache_release(page); 338 dout("start_read %p add_to_page_cache failed %p\n", 339 inode, page); 340 nr_pages = i; 341 goto out_pages; 342 } 343 pages[i] = page; 344 } 345 req->r_pages = pages; 346 req->r_num_pages = nr_pages; 347 req->r_callback = finish_read; 348 req->r_inode = inode; 349 350 dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len); 351 ret = ceph_osdc_start_request(osdc, req, false); 352 if (ret < 0) 353 goto out_pages; 354 ceph_osdc_put_request(req); 355 return nr_pages; 356 357 out_pages: 358 ceph_unlock_page_vector(pages, nr_pages); 359 ceph_release_page_vector(pages, nr_pages); 360 out: 361 ceph_osdc_put_request(req); 362 return ret; 363 } 364 365 366 /* 367 * Read multiple pages. Leave pages we don't read + unlock in page_list; 368 * the caller (VM) cleans them up. 369 */ 370 static int ceph_readpages(struct file *file, struct address_space *mapping, 371 struct list_head *page_list, unsigned nr_pages) 372 { 373 struct inode *inode = file->f_dentry->d_inode; 374 struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 375 int rc = 0; 376 int max = 0; 377 378 if (fsc->mount_options->rsize >= PAGE_CACHE_SIZE) 379 max = (fsc->mount_options->rsize + PAGE_CACHE_SIZE - 1) 380 >> PAGE_SHIFT; 381 382 dout("readpages %p file %p nr_pages %d max %d\n", inode, file, nr_pages, 383 max); 384 while (!list_empty(page_list)) { 385 rc = start_read(inode, page_list, max); 386 if (rc < 0) 387 goto out; 388 BUG_ON(rc == 0); 389 } 390 out: 391 dout("readpages %p file %p ret %d\n", inode, file, rc); 392 return rc; 393 } 394 395 /* 396 * Get ref for the oldest snapc for an inode with dirty data... that is, the 397 * only snap context we are allowed to write back. 398 */ 399 static struct ceph_snap_context *get_oldest_context(struct inode *inode, 400 u64 *snap_size) 401 { 402 struct ceph_inode_info *ci = ceph_inode(inode); 403 struct ceph_snap_context *snapc = NULL; 404 struct ceph_cap_snap *capsnap = NULL; 405 406 spin_lock(&ci->i_ceph_lock); 407 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { 408 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap, 409 capsnap->context, capsnap->dirty_pages); 410 if (capsnap->dirty_pages) { 411 snapc = ceph_get_snap_context(capsnap->context); 412 if (snap_size) 413 *snap_size = capsnap->size; 414 break; 415 } 416 } 417 if (!snapc && ci->i_wrbuffer_ref_head) { 418 snapc = ceph_get_snap_context(ci->i_head_snapc); 419 dout(" head snapc %p has %d dirty pages\n", 420 snapc, ci->i_wrbuffer_ref_head); 421 } 422 spin_unlock(&ci->i_ceph_lock); 423 return snapc; 424 } 425 426 /* 427 * Write a single page, but leave the page locked. 428 * 429 * If we get a write error, set the page error bit, but still adjust the 430 * dirty page accounting (i.e., page is no longer dirty). 431 */ 432 static int writepage_nounlock(struct page *page, struct writeback_control *wbc) 433 { 434 struct inode *inode; 435 struct ceph_inode_info *ci; 436 struct ceph_fs_client *fsc; 437 struct ceph_osd_client *osdc; 438 loff_t page_off = page_offset(page); 439 int len = PAGE_CACHE_SIZE; 440 loff_t i_size; 441 int err = 0; 442 struct ceph_snap_context *snapc, *oldest; 443 u64 snap_size = 0; 444 long writeback_stat; 445 446 dout("writepage %p idx %lu\n", page, page->index); 447 448 if (!page->mapping || !page->mapping->host) { 449 dout("writepage %p - no mapping\n", page); 450 return -EFAULT; 451 } 452 inode = page->mapping->host; 453 ci = ceph_inode(inode); 454 fsc = ceph_inode_to_client(inode); 455 osdc = &fsc->client->osdc; 456 457 /* verify this is a writeable snap context */ 458 snapc = page_snap_context(page); 459 if (snapc == NULL) { 460 dout("writepage %p page %p not dirty?\n", inode, page); 461 goto out; 462 } 463 oldest = get_oldest_context(inode, &snap_size); 464 if (snapc->seq > oldest->seq) { 465 dout("writepage %p page %p snapc %p not writeable - noop\n", 466 inode, page, snapc); 467 /* we should only noop if called by kswapd */ 468 WARN_ON((current->flags & PF_MEMALLOC) == 0); 469 ceph_put_snap_context(oldest); 470 goto out; 471 } 472 ceph_put_snap_context(oldest); 473 474 /* is this a partial page at end of file? */ 475 if (snap_size) 476 i_size = snap_size; 477 else 478 i_size = i_size_read(inode); 479 if (i_size < page_off + len) 480 len = i_size - page_off; 481 482 dout("writepage %p page %p index %lu on %llu~%u snapc %p\n", 483 inode, page, page->index, page_off, len, snapc); 484 485 writeback_stat = atomic_long_inc_return(&fsc->writeback_count); 486 if (writeback_stat > 487 CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb)) 488 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC); 489 490 set_page_writeback(page); 491 err = ceph_osdc_writepages(osdc, ceph_vino(inode), 492 &ci->i_layout, snapc, 493 page_off, len, 494 ci->i_truncate_seq, ci->i_truncate_size, 495 &inode->i_mtime, &page, 1); 496 if (err < 0) { 497 dout("writepage setting page/mapping error %d %p\n", err, page); 498 SetPageError(page); 499 mapping_set_error(&inode->i_data, err); 500 if (wbc) 501 wbc->pages_skipped++; 502 } else { 503 dout("writepage cleaned page %p\n", page); 504 err = 0; /* vfs expects us to return 0 */ 505 } 506 page->private = 0; 507 ClearPagePrivate(page); 508 end_page_writeback(page); 509 ceph_put_wrbuffer_cap_refs(ci, 1, snapc); 510 ceph_put_snap_context(snapc); /* page's reference */ 511 out: 512 return err; 513 } 514 515 static int ceph_writepage(struct page *page, struct writeback_control *wbc) 516 { 517 int err; 518 struct inode *inode = page->mapping->host; 519 BUG_ON(!inode); 520 ihold(inode); 521 err = writepage_nounlock(page, wbc); 522 unlock_page(page); 523 iput(inode); 524 return err; 525 } 526 527 528 /* 529 * lame release_pages helper. release_pages() isn't exported to 530 * modules. 531 */ 532 static void ceph_release_pages(struct page **pages, int num) 533 { 534 struct pagevec pvec; 535 int i; 536 537 pagevec_init(&pvec, 0); 538 for (i = 0; i < num; i++) { 539 if (pagevec_add(&pvec, pages[i]) == 0) 540 pagevec_release(&pvec); 541 } 542 pagevec_release(&pvec); 543 } 544 545 546 /* 547 * async writeback completion handler. 548 * 549 * If we get an error, set the mapping error bit, but not the individual 550 * page error bits. 551 */ 552 static void writepages_finish(struct ceph_osd_request *req, 553 struct ceph_msg *msg) 554 { 555 struct inode *inode = req->r_inode; 556 struct ceph_osd_reply_head *replyhead; 557 struct ceph_osd_op *op; 558 struct ceph_inode_info *ci = ceph_inode(inode); 559 unsigned wrote; 560 struct page *page; 561 int i; 562 struct ceph_snap_context *snapc = req->r_snapc; 563 struct address_space *mapping = inode->i_mapping; 564 __s32 rc = -EIO; 565 u64 bytes = 0; 566 struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 567 long writeback_stat; 568 unsigned issued = ceph_caps_issued(ci); 569 570 /* parse reply */ 571 replyhead = msg->front.iov_base; 572 WARN_ON(le32_to_cpu(replyhead->num_ops) == 0); 573 op = (void *)(replyhead + 1); 574 rc = le32_to_cpu(replyhead->result); 575 bytes = le64_to_cpu(op->extent.length); 576 577 if (rc >= 0) { 578 /* 579 * Assume we wrote the pages we originally sent. The 580 * osd might reply with fewer pages if our writeback 581 * raced with a truncation and was adjusted at the osd, 582 * so don't believe the reply. 583 */ 584 wrote = req->r_num_pages; 585 } else { 586 wrote = 0; 587 mapping_set_error(mapping, rc); 588 } 589 dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n", 590 inode, rc, bytes, wrote); 591 592 /* clean all pages */ 593 for (i = 0; i < req->r_num_pages; i++) { 594 page = req->r_pages[i]; 595 BUG_ON(!page); 596 WARN_ON(!PageUptodate(page)); 597 598 writeback_stat = 599 atomic_long_dec_return(&fsc->writeback_count); 600 if (writeback_stat < 601 CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb)) 602 clear_bdi_congested(&fsc->backing_dev_info, 603 BLK_RW_ASYNC); 604 605 ceph_put_snap_context(page_snap_context(page)); 606 page->private = 0; 607 ClearPagePrivate(page); 608 dout("unlocking %d %p\n", i, page); 609 end_page_writeback(page); 610 611 /* 612 * We lost the cache cap, need to truncate the page before 613 * it is unlocked, otherwise we'd truncate it later in the 614 * page truncation thread, possibly losing some data that 615 * raced its way in 616 */ 617 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) 618 generic_error_remove_page(inode->i_mapping, page); 619 620 unlock_page(page); 621 } 622 dout("%p wrote+cleaned %d pages\n", inode, wrote); 623 ceph_put_wrbuffer_cap_refs(ci, req->r_num_pages, snapc); 624 625 ceph_release_pages(req->r_pages, req->r_num_pages); 626 if (req->r_pages_from_pool) 627 mempool_free(req->r_pages, 628 ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool); 629 else 630 kfree(req->r_pages); 631 ceph_osdc_put_request(req); 632 } 633 634 /* 635 * allocate a page vec, either directly, or if necessary, via a the 636 * mempool. we avoid the mempool if we can because req->r_num_pages 637 * may be less than the maximum write size. 638 */ 639 static void alloc_page_vec(struct ceph_fs_client *fsc, 640 struct ceph_osd_request *req) 641 { 642 req->r_pages = kmalloc(sizeof(struct page *) * req->r_num_pages, 643 GFP_NOFS); 644 if (!req->r_pages) { 645 req->r_pages = mempool_alloc(fsc->wb_pagevec_pool, GFP_NOFS); 646 req->r_pages_from_pool = 1; 647 WARN_ON(!req->r_pages); 648 } 649 } 650 651 /* 652 * initiate async writeback 653 */ 654 static int ceph_writepages_start(struct address_space *mapping, 655 struct writeback_control *wbc) 656 { 657 struct inode *inode = mapping->host; 658 struct ceph_inode_info *ci = ceph_inode(inode); 659 struct ceph_fs_client *fsc; 660 pgoff_t index, start, end; 661 int range_whole = 0; 662 int should_loop = 1; 663 pgoff_t max_pages = 0, max_pages_ever = 0; 664 struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc; 665 struct pagevec pvec; 666 int done = 0; 667 int rc = 0; 668 unsigned wsize = 1 << inode->i_blkbits; 669 struct ceph_osd_request *req = NULL; 670 int do_sync; 671 u64 snap_size = 0; 672 673 /* 674 * Include a 'sync' in the OSD request if this is a data 675 * integrity write (e.g., O_SYNC write or fsync()), or if our 676 * cap is being revoked. 677 */ 678 do_sync = wbc->sync_mode == WB_SYNC_ALL; 679 if (ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER)) 680 do_sync = 1; 681 dout("writepages_start %p dosync=%d (mode=%s)\n", 682 inode, do_sync, 683 wbc->sync_mode == WB_SYNC_NONE ? "NONE" : 684 (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD")); 685 686 fsc = ceph_inode_to_client(inode); 687 if (fsc->mount_state == CEPH_MOUNT_SHUTDOWN) { 688 pr_warning("writepage_start %p on forced umount\n", inode); 689 return -EIO; /* we're in a forced umount, don't write! */ 690 } 691 if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize) 692 wsize = fsc->mount_options->wsize; 693 if (wsize < PAGE_CACHE_SIZE) 694 wsize = PAGE_CACHE_SIZE; 695 max_pages_ever = wsize >> PAGE_CACHE_SHIFT; 696 697 pagevec_init(&pvec, 0); 698 699 /* where to start/end? */ 700 if (wbc->range_cyclic) { 701 start = mapping->writeback_index; /* Start from prev offset */ 702 end = -1; 703 dout(" cyclic, start at %lu\n", start); 704 } else { 705 start = wbc->range_start >> PAGE_CACHE_SHIFT; 706 end = wbc->range_end >> PAGE_CACHE_SHIFT; 707 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX) 708 range_whole = 1; 709 should_loop = 0; 710 dout(" not cyclic, %lu to %lu\n", start, end); 711 } 712 index = start; 713 714 retry: 715 /* find oldest snap context with dirty data */ 716 ceph_put_snap_context(snapc); 717 snapc = get_oldest_context(inode, &snap_size); 718 if (!snapc) { 719 /* hmm, why does writepages get called when there 720 is no dirty data? */ 721 dout(" no snap context with dirty data?\n"); 722 goto out; 723 } 724 dout(" oldest snapc is %p seq %lld (%d snaps)\n", 725 snapc, snapc->seq, snapc->num_snaps); 726 if (last_snapc && snapc != last_snapc) { 727 /* if we switched to a newer snapc, restart our scan at the 728 * start of the original file range. */ 729 dout(" snapc differs from last pass, restarting at %lu\n", 730 index); 731 index = start; 732 } 733 last_snapc = snapc; 734 735 while (!done && index <= end) { 736 unsigned i; 737 int first; 738 pgoff_t next; 739 int pvec_pages, locked_pages; 740 struct page *page; 741 int want; 742 u64 offset, len; 743 struct ceph_osd_request_head *reqhead; 744 struct ceph_osd_op *op; 745 long writeback_stat; 746 747 next = 0; 748 locked_pages = 0; 749 max_pages = max_pages_ever; 750 751 get_more_pages: 752 first = -1; 753 want = min(end - index, 754 min((pgoff_t)PAGEVEC_SIZE, 755 max_pages - (pgoff_t)locked_pages) - 1) 756 + 1; 757 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index, 758 PAGECACHE_TAG_DIRTY, 759 want); 760 dout("pagevec_lookup_tag got %d\n", pvec_pages); 761 if (!pvec_pages && !locked_pages) 762 break; 763 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) { 764 page = pvec.pages[i]; 765 dout("? %p idx %lu\n", page, page->index); 766 if (locked_pages == 0) 767 lock_page(page); /* first page */ 768 else if (!trylock_page(page)) 769 break; 770 771 /* only dirty pages, or our accounting breaks */ 772 if (unlikely(!PageDirty(page)) || 773 unlikely(page->mapping != mapping)) { 774 dout("!dirty or !mapping %p\n", page); 775 unlock_page(page); 776 break; 777 } 778 if (!wbc->range_cyclic && page->index > end) { 779 dout("end of range %p\n", page); 780 done = 1; 781 unlock_page(page); 782 break; 783 } 784 if (next && (page->index != next)) { 785 dout("not consecutive %p\n", page); 786 unlock_page(page); 787 break; 788 } 789 if (wbc->sync_mode != WB_SYNC_NONE) { 790 dout("waiting on writeback %p\n", page); 791 wait_on_page_writeback(page); 792 } 793 if ((snap_size && page_offset(page) > snap_size) || 794 (!snap_size && 795 page_offset(page) > i_size_read(inode))) { 796 dout("%p page eof %llu\n", page, snap_size ? 797 snap_size : i_size_read(inode)); 798 done = 1; 799 unlock_page(page); 800 break; 801 } 802 if (PageWriteback(page)) { 803 dout("%p under writeback\n", page); 804 unlock_page(page); 805 break; 806 } 807 808 /* only if matching snap context */ 809 pgsnapc = page_snap_context(page); 810 if (pgsnapc->seq > snapc->seq) { 811 dout("page snapc %p %lld > oldest %p %lld\n", 812 pgsnapc, pgsnapc->seq, snapc, snapc->seq); 813 unlock_page(page); 814 if (!locked_pages) 815 continue; /* keep looking for snap */ 816 break; 817 } 818 819 if (!clear_page_dirty_for_io(page)) { 820 dout("%p !clear_page_dirty_for_io\n", page); 821 unlock_page(page); 822 break; 823 } 824 825 /* ok */ 826 if (locked_pages == 0) { 827 /* prepare async write request */ 828 offset = (u64) page_offset(page); 829 len = wsize; 830 req = ceph_osdc_new_request(&fsc->client->osdc, 831 &ci->i_layout, 832 ceph_vino(inode), 833 offset, &len, 834 CEPH_OSD_OP_WRITE, 835 CEPH_OSD_FLAG_WRITE | 836 CEPH_OSD_FLAG_ONDISK, 837 snapc, do_sync, 838 ci->i_truncate_seq, 839 ci->i_truncate_size, 840 &inode->i_mtime, true, 1, 0); 841 842 if (IS_ERR(req)) { 843 rc = PTR_ERR(req); 844 unlock_page(page); 845 break; 846 } 847 848 max_pages = req->r_num_pages; 849 850 alloc_page_vec(fsc, req); 851 req->r_callback = writepages_finish; 852 req->r_inode = inode; 853 } 854 855 /* note position of first page in pvec */ 856 if (first < 0) 857 first = i; 858 dout("%p will write page %p idx %lu\n", 859 inode, page, page->index); 860 861 writeback_stat = 862 atomic_long_inc_return(&fsc->writeback_count); 863 if (writeback_stat > CONGESTION_ON_THRESH( 864 fsc->mount_options->congestion_kb)) { 865 set_bdi_congested(&fsc->backing_dev_info, 866 BLK_RW_ASYNC); 867 } 868 869 set_page_writeback(page); 870 req->r_pages[locked_pages] = page; 871 locked_pages++; 872 next = page->index + 1; 873 } 874 875 /* did we get anything? */ 876 if (!locked_pages) 877 goto release_pvec_pages; 878 if (i) { 879 int j; 880 BUG_ON(!locked_pages || first < 0); 881 882 if (pvec_pages && i == pvec_pages && 883 locked_pages < max_pages) { 884 dout("reached end pvec, trying for more\n"); 885 pagevec_reinit(&pvec); 886 goto get_more_pages; 887 } 888 889 /* shift unused pages over in the pvec... we 890 * will need to release them below. */ 891 for (j = i; j < pvec_pages; j++) { 892 dout(" pvec leftover page %p\n", 893 pvec.pages[j]); 894 pvec.pages[j-i+first] = pvec.pages[j]; 895 } 896 pvec.nr -= i-first; 897 } 898 899 /* submit the write */ 900 offset = req->r_pages[0]->index << PAGE_CACHE_SHIFT; 901 len = min((snap_size ? snap_size : i_size_read(inode)) - offset, 902 (u64)locked_pages << PAGE_CACHE_SHIFT); 903 dout("writepages got %d pages at %llu~%llu\n", 904 locked_pages, offset, len); 905 906 /* revise final length, page count */ 907 req->r_num_pages = locked_pages; 908 reqhead = req->r_request->front.iov_base; 909 op = (void *)(reqhead + 1); 910 op->extent.length = cpu_to_le64(len); 911 op->payload_len = cpu_to_le32(len); 912 req->r_request->hdr.data_len = cpu_to_le32(len); 913 914 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true); 915 BUG_ON(rc); 916 req = NULL; 917 918 /* continue? */ 919 index = next; 920 wbc->nr_to_write -= locked_pages; 921 if (wbc->nr_to_write <= 0) 922 done = 1; 923 924 release_pvec_pages: 925 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr, 926 pvec.nr ? pvec.pages[0] : NULL); 927 pagevec_release(&pvec); 928 929 if (locked_pages && !done) 930 goto retry; 931 } 932 933 if (should_loop && !done) { 934 /* more to do; loop back to beginning of file */ 935 dout("writepages looping back to beginning of file\n"); 936 should_loop = 0; 937 index = 0; 938 goto retry; 939 } 940 941 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) 942 mapping->writeback_index = index; 943 944 out: 945 if (req) 946 ceph_osdc_put_request(req); 947 ceph_put_snap_context(snapc); 948 dout("writepages done, rc = %d\n", rc); 949 return rc; 950 } 951 952 953 954 /* 955 * See if a given @snapc is either writeable, or already written. 956 */ 957 static int context_is_writeable_or_written(struct inode *inode, 958 struct ceph_snap_context *snapc) 959 { 960 struct ceph_snap_context *oldest = get_oldest_context(inode, NULL); 961 int ret = !oldest || snapc->seq <= oldest->seq; 962 963 ceph_put_snap_context(oldest); 964 return ret; 965 } 966 967 /* 968 * We are only allowed to write into/dirty the page if the page is 969 * clean, or already dirty within the same snap context. 970 * 971 * called with page locked. 972 * return success with page locked, 973 * or any failure (incl -EAGAIN) with page unlocked. 974 */ 975 static int ceph_update_writeable_page(struct file *file, 976 loff_t pos, unsigned len, 977 struct page *page) 978 { 979 struct inode *inode = file->f_dentry->d_inode; 980 struct ceph_inode_info *ci = ceph_inode(inode); 981 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc; 982 loff_t page_off = pos & PAGE_CACHE_MASK; 983 int pos_in_page = pos & ~PAGE_CACHE_MASK; 984 int end_in_page = pos_in_page + len; 985 loff_t i_size; 986 int r; 987 struct ceph_snap_context *snapc, *oldest; 988 989 retry_locked: 990 /* writepages currently holds page lock, but if we change that later, */ 991 wait_on_page_writeback(page); 992 993 /* check snap context */ 994 BUG_ON(!ci->i_snap_realm); 995 down_read(&mdsc->snap_rwsem); 996 BUG_ON(!ci->i_snap_realm->cached_context); 997 snapc = page_snap_context(page); 998 if (snapc && snapc != ci->i_head_snapc) { 999 /* 1000 * this page is already dirty in another (older) snap 1001 * context! is it writeable now? 1002 */ 1003 oldest = get_oldest_context(inode, NULL); 1004 up_read(&mdsc->snap_rwsem); 1005 1006 if (snapc->seq > oldest->seq) { 1007 ceph_put_snap_context(oldest); 1008 dout(" page %p snapc %p not current or oldest\n", 1009 page, snapc); 1010 /* 1011 * queue for writeback, and wait for snapc to 1012 * be writeable or written 1013 */ 1014 snapc = ceph_get_snap_context(snapc); 1015 unlock_page(page); 1016 ceph_queue_writeback(inode); 1017 r = wait_event_interruptible(ci->i_cap_wq, 1018 context_is_writeable_or_written(inode, snapc)); 1019 ceph_put_snap_context(snapc); 1020 if (r == -ERESTARTSYS) 1021 return r; 1022 return -EAGAIN; 1023 } 1024 ceph_put_snap_context(oldest); 1025 1026 /* yay, writeable, do it now (without dropping page lock) */ 1027 dout(" page %p snapc %p not current, but oldest\n", 1028 page, snapc); 1029 if (!clear_page_dirty_for_io(page)) 1030 goto retry_locked; 1031 r = writepage_nounlock(page, NULL); 1032 if (r < 0) 1033 goto fail_nosnap; 1034 goto retry_locked; 1035 } 1036 1037 if (PageUptodate(page)) { 1038 dout(" page %p already uptodate\n", page); 1039 return 0; 1040 } 1041 1042 /* full page? */ 1043 if (pos_in_page == 0 && len == PAGE_CACHE_SIZE) 1044 return 0; 1045 1046 /* past end of file? */ 1047 i_size = inode->i_size; /* caller holds i_mutex */ 1048 1049 if (i_size + len > inode->i_sb->s_maxbytes) { 1050 /* file is too big */ 1051 r = -EINVAL; 1052 goto fail; 1053 } 1054 1055 if (page_off >= i_size || 1056 (pos_in_page == 0 && (pos+len) >= i_size && 1057 end_in_page - pos_in_page != PAGE_CACHE_SIZE)) { 1058 dout(" zeroing %p 0 - %d and %d - %d\n", 1059 page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE); 1060 zero_user_segments(page, 1061 0, pos_in_page, 1062 end_in_page, PAGE_CACHE_SIZE); 1063 return 0; 1064 } 1065 1066 /* we need to read it. */ 1067 up_read(&mdsc->snap_rwsem); 1068 r = readpage_nounlock(file, page); 1069 if (r < 0) 1070 goto fail_nosnap; 1071 goto retry_locked; 1072 1073 fail: 1074 up_read(&mdsc->snap_rwsem); 1075 fail_nosnap: 1076 unlock_page(page); 1077 return r; 1078 } 1079 1080 /* 1081 * We are only allowed to write into/dirty the page if the page is 1082 * clean, or already dirty within the same snap context. 1083 */ 1084 static int ceph_write_begin(struct file *file, struct address_space *mapping, 1085 loff_t pos, unsigned len, unsigned flags, 1086 struct page **pagep, void **fsdata) 1087 { 1088 struct inode *inode = file->f_dentry->d_inode; 1089 struct ceph_inode_info *ci = ceph_inode(inode); 1090 struct ceph_file_info *fi = file->private_data; 1091 struct page *page; 1092 pgoff_t index = pos >> PAGE_CACHE_SHIFT; 1093 int r, want, got = 0; 1094 1095 if (fi->fmode & CEPH_FILE_MODE_LAZY) 1096 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO; 1097 else 1098 want = CEPH_CAP_FILE_BUFFER; 1099 1100 dout("write_begin %p %llx.%llx %llu~%u getting caps. i_size %llu\n", 1101 inode, ceph_vinop(inode), pos, len, inode->i_size); 1102 r = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, &got, pos+len); 1103 if (r < 0) 1104 return r; 1105 dout("write_begin %p %llx.%llx %llu~%u got cap refs on %s\n", 1106 inode, ceph_vinop(inode), pos, len, ceph_cap_string(got)); 1107 if (!(got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO))) { 1108 ceph_put_cap_refs(ci, got); 1109 return -EAGAIN; 1110 } 1111 1112 do { 1113 /* get a page */ 1114 page = grab_cache_page_write_begin(mapping, index, 0); 1115 if (!page) { 1116 r = -ENOMEM; 1117 break; 1118 } 1119 1120 dout("write_begin file %p inode %p page %p %d~%d\n", file, 1121 inode, page, (int)pos, (int)len); 1122 1123 r = ceph_update_writeable_page(file, pos, len, page); 1124 if (r) 1125 page_cache_release(page); 1126 } while (r == -EAGAIN); 1127 1128 if (r) { 1129 ceph_put_cap_refs(ci, got); 1130 } else { 1131 *pagep = page; 1132 *(int *)fsdata = got; 1133 } 1134 return r; 1135 } 1136 1137 /* 1138 * we don't do anything in here that simple_write_end doesn't do 1139 * except adjust dirty page accounting and drop read lock on 1140 * mdsc->snap_rwsem. 1141 */ 1142 static int ceph_write_end(struct file *file, struct address_space *mapping, 1143 loff_t pos, unsigned len, unsigned copied, 1144 struct page *page, void *fsdata) 1145 { 1146 struct inode *inode = file->f_dentry->d_inode; 1147 struct ceph_inode_info *ci = ceph_inode(inode); 1148 struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 1149 struct ceph_mds_client *mdsc = fsc->mdsc; 1150 unsigned from = pos & (PAGE_CACHE_SIZE - 1); 1151 int check_cap = 0; 1152 int got = (unsigned long)fsdata; 1153 1154 dout("write_end file %p inode %p page %p %d~%d (%d)\n", file, 1155 inode, page, (int)pos, (int)copied, (int)len); 1156 1157 /* zero the stale part of the page if we did a short copy */ 1158 if (copied < len) 1159 zero_user_segment(page, from+copied, len); 1160 1161 /* did file size increase? */ 1162 /* (no need for i_size_read(); we caller holds i_mutex */ 1163 if (pos+copied > inode->i_size) 1164 check_cap = ceph_inode_set_size(inode, pos+copied); 1165 1166 if (!PageUptodate(page)) 1167 SetPageUptodate(page); 1168 1169 set_page_dirty(page); 1170 1171 unlock_page(page); 1172 up_read(&mdsc->snap_rwsem); 1173 page_cache_release(page); 1174 1175 if (copied > 0) { 1176 int dirty; 1177 spin_lock(&ci->i_ceph_lock); 1178 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR); 1179 spin_unlock(&ci->i_ceph_lock); 1180 if (dirty) 1181 __mark_inode_dirty(inode, dirty); 1182 } 1183 1184 dout("write_end %p %llx.%llx %llu~%u dropping cap refs on %s\n", 1185 inode, ceph_vinop(inode), pos, len, ceph_cap_string(got)); 1186 ceph_put_cap_refs(ci, got); 1187 1188 if (check_cap) 1189 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL); 1190 1191 return copied; 1192 } 1193 1194 /* 1195 * we set .direct_IO to indicate direct io is supported, but since we 1196 * intercept O_DIRECT reads and writes early, this function should 1197 * never get called. 1198 */ 1199 static ssize_t ceph_direct_io(int rw, struct kiocb *iocb, 1200 const struct iovec *iov, 1201 loff_t pos, unsigned long nr_segs) 1202 { 1203 WARN_ON(1); 1204 return -EINVAL; 1205 } 1206 1207 const struct address_space_operations ceph_aops = { 1208 .readpage = ceph_readpage, 1209 .readpages = ceph_readpages, 1210 .writepage = ceph_writepage, 1211 .writepages = ceph_writepages_start, 1212 .write_begin = ceph_write_begin, 1213 .write_end = ceph_write_end, 1214 .set_page_dirty = ceph_set_page_dirty, 1215 .invalidatepage = ceph_invalidatepage, 1216 .releasepage = ceph_releasepage, 1217 .direct_IO = ceph_direct_io, 1218 }; 1219 1220 1221 /* 1222 * vm ops 1223 */ 1224 1225 /* 1226 * Reuse write_begin here for simplicity. 1227 */ 1228 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf) 1229 { 1230 struct inode *inode = vma->vm_file->f_dentry->d_inode; 1231 struct page *page = vmf->page; 1232 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc; 1233 loff_t off = page_offset(page); 1234 loff_t size, len; 1235 int ret; 1236 1237 /* Update time before taking page lock */ 1238 file_update_time(vma->vm_file); 1239 1240 size = i_size_read(inode); 1241 if (off + PAGE_CACHE_SIZE <= size) 1242 len = PAGE_CACHE_SIZE; 1243 else 1244 len = size & ~PAGE_CACHE_MASK; 1245 1246 dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode, 1247 off, len, page, page->index); 1248 1249 lock_page(page); 1250 1251 ret = VM_FAULT_NOPAGE; 1252 if ((off > size) || 1253 (page->mapping != inode->i_mapping)) 1254 goto out; 1255 1256 ret = ceph_update_writeable_page(vma->vm_file, off, len, page); 1257 if (ret == 0) { 1258 /* success. we'll keep the page locked. */ 1259 set_page_dirty(page); 1260 up_read(&mdsc->snap_rwsem); 1261 ret = VM_FAULT_LOCKED; 1262 } else { 1263 if (ret == -ENOMEM) 1264 ret = VM_FAULT_OOM; 1265 else 1266 ret = VM_FAULT_SIGBUS; 1267 } 1268 out: 1269 dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret); 1270 if (ret != VM_FAULT_LOCKED) 1271 unlock_page(page); 1272 return ret; 1273 } 1274 1275 static struct vm_operations_struct ceph_vmops = { 1276 .fault = filemap_fault, 1277 .page_mkwrite = ceph_page_mkwrite, 1278 .remap_pages = generic_file_remap_pages, 1279 }; 1280 1281 int ceph_mmap(struct file *file, struct vm_area_struct *vma) 1282 { 1283 struct address_space *mapping = file->f_mapping; 1284 1285 if (!mapping->a_ops->readpage) 1286 return -ENOEXEC; 1287 file_accessed(file); 1288 vma->vm_ops = &ceph_vmops; 1289 return 0; 1290 } 1291