1 // SPDX-License-Identifier: GPL-2.0 2 #include <linux/ceph/ceph_debug.h> 3 4 #include <linux/backing-dev.h> 5 #include <linux/fs.h> 6 #include <linux/mm.h> 7 #include <linux/pagemap.h> 8 #include <linux/writeback.h> /* generic_writepages */ 9 #include <linux/slab.h> 10 #include <linux/pagevec.h> 11 #include <linux/task_io_accounting_ops.h> 12 #include <linux/signal.h> 13 #include <linux/iversion.h> 14 #include <linux/ktime.h> 15 16 #include "super.h" 17 #include "mds_client.h" 18 #include "cache.h" 19 #include "metric.h" 20 #include <linux/ceph/osd_client.h> 21 #include <linux/ceph/striper.h> 22 23 /* 24 * Ceph address space ops. 25 * 26 * There are a few funny things going on here. 27 * 28 * The page->private field is used to reference a struct 29 * ceph_snap_context for _every_ dirty page. This indicates which 30 * snapshot the page was logically dirtied in, and thus which snap 31 * context needs to be associated with the osd write during writeback. 32 * 33 * Similarly, struct ceph_inode_info maintains a set of counters to 34 * count dirty pages on the inode. In the absence of snapshots, 35 * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count. 36 * 37 * When a snapshot is taken (that is, when the client receives 38 * notification that a snapshot was taken), each inode with caps and 39 * with dirty pages (dirty pages implies there is a cap) gets a new 40 * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending 41 * order, new snaps go to the tail). The i_wrbuffer_ref_head count is 42 * moved to capsnap->dirty. (Unless a sync write is currently in 43 * progress. In that case, the capsnap is said to be "pending", new 44 * writes cannot start, and the capsnap isn't "finalized" until the 45 * write completes (or fails) and a final size/mtime for the inode for 46 * that snap can be settled upon.) i_wrbuffer_ref_head is reset to 0. 47 * 48 * On writeback, we must submit writes to the osd IN SNAP ORDER. So, 49 * we look for the first capsnap in i_cap_snaps and write out pages in 50 * that snap context _only_. Then we move on to the next capsnap, 51 * eventually reaching the "live" or "head" context (i.e., pages that 52 * are not yet snapped) and are writing the most recently dirtied 53 * pages. 54 * 55 * Invalidate and so forth must take care to ensure the dirty page 56 * accounting is preserved. 57 */ 58 59 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10)) 60 #define CONGESTION_OFF_THRESH(congestion_kb) \ 61 (CONGESTION_ON_THRESH(congestion_kb) - \ 62 (CONGESTION_ON_THRESH(congestion_kb) >> 2)) 63 64 static inline struct ceph_snap_context *page_snap_context(struct page *page) 65 { 66 if (PagePrivate(page)) 67 return (void *)page->private; 68 return NULL; 69 } 70 71 /* 72 * Dirty a page. Optimistically adjust accounting, on the assumption 73 * that we won't race with invalidate. If we do, readjust. 74 */ 75 static int ceph_set_page_dirty(struct page *page) 76 { 77 struct address_space *mapping = page->mapping; 78 struct inode *inode; 79 struct ceph_inode_info *ci; 80 struct ceph_snap_context *snapc; 81 int ret; 82 83 if (unlikely(!mapping)) 84 return !TestSetPageDirty(page); 85 86 if (PageDirty(page)) { 87 dout("%p set_page_dirty %p idx %lu -- already dirty\n", 88 mapping->host, page, page->index); 89 BUG_ON(!PagePrivate(page)); 90 return 0; 91 } 92 93 inode = mapping->host; 94 ci = ceph_inode(inode); 95 96 /* dirty the head */ 97 spin_lock(&ci->i_ceph_lock); 98 BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference 99 if (__ceph_have_pending_cap_snap(ci)) { 100 struct ceph_cap_snap *capsnap = 101 list_last_entry(&ci->i_cap_snaps, 102 struct ceph_cap_snap, 103 ci_item); 104 snapc = ceph_get_snap_context(capsnap->context); 105 capsnap->dirty_pages++; 106 } else { 107 BUG_ON(!ci->i_head_snapc); 108 snapc = ceph_get_snap_context(ci->i_head_snapc); 109 ++ci->i_wrbuffer_ref_head; 110 } 111 if (ci->i_wrbuffer_ref == 0) 112 ihold(inode); 113 ++ci->i_wrbuffer_ref; 114 dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d " 115 "snapc %p seq %lld (%d snaps)\n", 116 mapping->host, page, page->index, 117 ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1, 118 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head, 119 snapc, snapc->seq, snapc->num_snaps); 120 spin_unlock(&ci->i_ceph_lock); 121 122 /* 123 * Reference snap context in page->private. Also set 124 * PagePrivate so that we get invalidatepage callback. 125 */ 126 BUG_ON(PagePrivate(page)); 127 page->private = (unsigned long)snapc; 128 SetPagePrivate(page); 129 130 ret = __set_page_dirty_nobuffers(page); 131 WARN_ON(!PageLocked(page)); 132 WARN_ON(!page->mapping); 133 134 return ret; 135 } 136 137 /* 138 * If we are truncating the full page (i.e. offset == 0), adjust the 139 * dirty page counters appropriately. Only called if there is private 140 * data on the page. 141 */ 142 static void ceph_invalidatepage(struct page *page, unsigned int offset, 143 unsigned int length) 144 { 145 struct inode *inode; 146 struct ceph_inode_info *ci; 147 struct ceph_snap_context *snapc = page_snap_context(page); 148 149 inode = page->mapping->host; 150 ci = ceph_inode(inode); 151 152 if (offset != 0 || length != PAGE_SIZE) { 153 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n", 154 inode, page, page->index, offset, length); 155 return; 156 } 157 158 ceph_invalidate_fscache_page(inode, page); 159 160 WARN_ON(!PageLocked(page)); 161 if (!PagePrivate(page)) 162 return; 163 164 dout("%p invalidatepage %p idx %lu full dirty page\n", 165 inode, page, page->index); 166 167 ceph_put_wrbuffer_cap_refs(ci, 1, snapc); 168 ceph_put_snap_context(snapc); 169 page->private = 0; 170 ClearPagePrivate(page); 171 } 172 173 static int ceph_releasepage(struct page *page, gfp_t g) 174 { 175 dout("%p releasepage %p idx %lu (%sdirty)\n", page->mapping->host, 176 page, page->index, PageDirty(page) ? "" : "not "); 177 178 /* Can we release the page from the cache? */ 179 if (!ceph_release_fscache_page(page, g)) 180 return 0; 181 182 return !PagePrivate(page); 183 } 184 185 /* read a single page, without unlocking it. */ 186 static int ceph_do_readpage(struct file *filp, struct page *page) 187 { 188 struct inode *inode = file_inode(filp); 189 struct ceph_inode_info *ci = ceph_inode(inode); 190 struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 191 struct ceph_osd_client *osdc = &fsc->client->osdc; 192 struct ceph_osd_request *req; 193 struct ceph_vino vino = ceph_vino(inode); 194 int err = 0; 195 u64 off = page_offset(page); 196 u64 len = PAGE_SIZE; 197 198 if (off >= i_size_read(inode)) { 199 zero_user_segment(page, 0, PAGE_SIZE); 200 SetPageUptodate(page); 201 return 0; 202 } 203 204 if (ci->i_inline_version != CEPH_INLINE_NONE) { 205 /* 206 * Uptodate inline data should have been added 207 * into page cache while getting Fcr caps. 208 */ 209 if (off == 0) 210 return -EINVAL; 211 zero_user_segment(page, 0, PAGE_SIZE); 212 SetPageUptodate(page); 213 return 0; 214 } 215 216 err = ceph_readpage_from_fscache(inode, page); 217 if (err == 0) 218 return -EINPROGRESS; 219 220 dout("readpage ino %llx.%llx file %p off %llu len %llu page %p index %lu\n", 221 vino.ino, vino.snap, filp, off, len, page, page->index); 222 req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len, 0, 1, 223 CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, NULL, 224 ci->i_truncate_seq, ci->i_truncate_size, 225 false); 226 if (IS_ERR(req)) 227 return PTR_ERR(req); 228 229 osd_req_op_extent_osd_data_pages(req, 0, &page, len, 0, false, false); 230 231 err = ceph_osdc_start_request(osdc, req, false); 232 if (!err) 233 err = ceph_osdc_wait_request(osdc, req); 234 235 ceph_update_read_latency(&fsc->mdsc->metric, req->r_start_latency, 236 req->r_end_latency, err); 237 238 ceph_osdc_put_request(req); 239 dout("readpage result %d\n", err); 240 241 if (err == -ENOENT) 242 err = 0; 243 if (err < 0) { 244 SetPageError(page); 245 ceph_fscache_readpage_cancel(inode, page); 246 if (err == -EBLOCKLISTED) 247 fsc->blocklisted = true; 248 goto out; 249 } 250 if (err < PAGE_SIZE) 251 /* zero fill remainder of page */ 252 zero_user_segment(page, err, PAGE_SIZE); 253 else 254 flush_dcache_page(page); 255 256 SetPageUptodate(page); 257 ceph_readpage_to_fscache(inode, page); 258 259 out: 260 return err < 0 ? err : 0; 261 } 262 263 static int ceph_readpage(struct file *filp, struct page *page) 264 { 265 int r = ceph_do_readpage(filp, page); 266 if (r != -EINPROGRESS) 267 unlock_page(page); 268 else 269 r = 0; 270 return r; 271 } 272 273 /* 274 * Finish an async read(ahead) op. 275 */ 276 static void finish_read(struct ceph_osd_request *req) 277 { 278 struct inode *inode = req->r_inode; 279 struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 280 struct ceph_osd_data *osd_data; 281 int rc = req->r_result <= 0 ? req->r_result : 0; 282 int bytes = req->r_result >= 0 ? req->r_result : 0; 283 int num_pages; 284 int i; 285 286 dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes); 287 if (rc == -EBLOCKLISTED) 288 ceph_inode_to_client(inode)->blocklisted = true; 289 290 /* unlock all pages, zeroing any data we didn't read */ 291 osd_data = osd_req_op_extent_osd_data(req, 0); 292 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES); 293 num_pages = calc_pages_for((u64)osd_data->alignment, 294 (u64)osd_data->length); 295 for (i = 0; i < num_pages; i++) { 296 struct page *page = osd_data->pages[i]; 297 298 if (rc < 0 && rc != -ENOENT) { 299 ceph_fscache_readpage_cancel(inode, page); 300 goto unlock; 301 } 302 if (bytes < (int)PAGE_SIZE) { 303 /* zero (remainder of) page */ 304 int s = bytes < 0 ? 0 : bytes; 305 zero_user_segment(page, s, PAGE_SIZE); 306 } 307 dout("finish_read %p uptodate %p idx %lu\n", inode, page, 308 page->index); 309 flush_dcache_page(page); 310 SetPageUptodate(page); 311 ceph_readpage_to_fscache(inode, page); 312 unlock: 313 unlock_page(page); 314 put_page(page); 315 bytes -= PAGE_SIZE; 316 } 317 318 ceph_update_read_latency(&fsc->mdsc->metric, req->r_start_latency, 319 req->r_end_latency, rc); 320 321 kfree(osd_data->pages); 322 } 323 324 /* 325 * start an async read(ahead) operation. return nr_pages we submitted 326 * a read for on success, or negative error code. 327 */ 328 static int start_read(struct inode *inode, struct ceph_rw_context *rw_ctx, 329 struct list_head *page_list, int max) 330 { 331 struct ceph_osd_client *osdc = 332 &ceph_inode_to_client(inode)->client->osdc; 333 struct ceph_inode_info *ci = ceph_inode(inode); 334 struct page *page = lru_to_page(page_list); 335 struct ceph_vino vino; 336 struct ceph_osd_request *req; 337 u64 off; 338 u64 len; 339 int i; 340 struct page **pages; 341 pgoff_t next_index; 342 int nr_pages = 0; 343 int got = 0; 344 int ret = 0; 345 346 if (!rw_ctx) { 347 /* caller of readpages does not hold buffer and read caps 348 * (fadvise, madvise and readahead cases) */ 349 int want = CEPH_CAP_FILE_CACHE; 350 ret = ceph_try_get_caps(inode, CEPH_CAP_FILE_RD, want, 351 true, &got); 352 if (ret < 0) { 353 dout("start_read %p, error getting cap\n", inode); 354 } else if (!(got & want)) { 355 dout("start_read %p, no cache cap\n", inode); 356 ret = 0; 357 } 358 if (ret <= 0) { 359 if (got) 360 ceph_put_cap_refs(ci, got); 361 while (!list_empty(page_list)) { 362 page = lru_to_page(page_list); 363 list_del(&page->lru); 364 put_page(page); 365 } 366 return ret; 367 } 368 } 369 370 off = (u64) page_offset(page); 371 372 /* count pages */ 373 next_index = page->index; 374 list_for_each_entry_reverse(page, page_list, lru) { 375 if (page->index != next_index) 376 break; 377 nr_pages++; 378 next_index++; 379 if (max && nr_pages == max) 380 break; 381 } 382 len = nr_pages << PAGE_SHIFT; 383 dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages, 384 off, len); 385 vino = ceph_vino(inode); 386 req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len, 387 0, 1, CEPH_OSD_OP_READ, 388 CEPH_OSD_FLAG_READ, NULL, 389 ci->i_truncate_seq, ci->i_truncate_size, 390 false); 391 if (IS_ERR(req)) { 392 ret = PTR_ERR(req); 393 goto out; 394 } 395 396 /* build page vector */ 397 nr_pages = calc_pages_for(0, len); 398 pages = kmalloc_array(nr_pages, sizeof(*pages), GFP_KERNEL); 399 if (!pages) { 400 ret = -ENOMEM; 401 goto out_put; 402 } 403 for (i = 0; i < nr_pages; ++i) { 404 page = list_entry(page_list->prev, struct page, lru); 405 BUG_ON(PageLocked(page)); 406 list_del(&page->lru); 407 408 dout("start_read %p adding %p idx %lu\n", inode, page, 409 page->index); 410 if (add_to_page_cache_lru(page, &inode->i_data, page->index, 411 GFP_KERNEL)) { 412 ceph_fscache_uncache_page(inode, page); 413 put_page(page); 414 dout("start_read %p add_to_page_cache failed %p\n", 415 inode, page); 416 nr_pages = i; 417 if (nr_pages > 0) { 418 len = nr_pages << PAGE_SHIFT; 419 osd_req_op_extent_update(req, 0, len); 420 break; 421 } 422 goto out_pages; 423 } 424 pages[i] = page; 425 } 426 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false); 427 req->r_callback = finish_read; 428 req->r_inode = inode; 429 430 dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len); 431 ret = ceph_osdc_start_request(osdc, req, false); 432 if (ret < 0) 433 goto out_pages; 434 ceph_osdc_put_request(req); 435 436 /* After adding locked pages to page cache, the inode holds cache cap. 437 * So we can drop our cap refs. */ 438 if (got) 439 ceph_put_cap_refs(ci, got); 440 441 return nr_pages; 442 443 out_pages: 444 for (i = 0; i < nr_pages; ++i) { 445 ceph_fscache_readpage_cancel(inode, pages[i]); 446 unlock_page(pages[i]); 447 } 448 ceph_put_page_vector(pages, nr_pages, false); 449 out_put: 450 ceph_osdc_put_request(req); 451 out: 452 if (got) 453 ceph_put_cap_refs(ci, got); 454 return ret; 455 } 456 457 458 /* 459 * Read multiple pages. Leave pages we don't read + unlock in page_list; 460 * the caller (VM) cleans them up. 461 */ 462 static int ceph_readpages(struct file *file, struct address_space *mapping, 463 struct list_head *page_list, unsigned nr_pages) 464 { 465 struct inode *inode = file_inode(file); 466 struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 467 struct ceph_file_info *fi = file->private_data; 468 struct ceph_rw_context *rw_ctx; 469 int rc = 0; 470 int max = 0; 471 472 if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE) 473 return -EINVAL; 474 475 rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list, 476 &nr_pages); 477 478 if (rc == 0) 479 goto out; 480 481 rw_ctx = ceph_find_rw_context(fi); 482 max = fsc->mount_options->rsize >> PAGE_SHIFT; 483 dout("readpages %p file %p ctx %p nr_pages %d max %d\n", 484 inode, file, rw_ctx, nr_pages, max); 485 while (!list_empty(page_list)) { 486 rc = start_read(inode, rw_ctx, page_list, max); 487 if (rc < 0) 488 goto out; 489 } 490 out: 491 ceph_fscache_readpages_cancel(inode, page_list); 492 493 dout("readpages %p file %p ret %d\n", inode, file, rc); 494 return rc; 495 } 496 497 struct ceph_writeback_ctl 498 { 499 loff_t i_size; 500 u64 truncate_size; 501 u32 truncate_seq; 502 bool size_stable; 503 bool head_snapc; 504 }; 505 506 /* 507 * Get ref for the oldest snapc for an inode with dirty data... that is, the 508 * only snap context we are allowed to write back. 509 */ 510 static struct ceph_snap_context * 511 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl, 512 struct ceph_snap_context *page_snapc) 513 { 514 struct ceph_inode_info *ci = ceph_inode(inode); 515 struct ceph_snap_context *snapc = NULL; 516 struct ceph_cap_snap *capsnap = NULL; 517 518 spin_lock(&ci->i_ceph_lock); 519 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { 520 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap, 521 capsnap->context, capsnap->dirty_pages); 522 if (!capsnap->dirty_pages) 523 continue; 524 525 /* get i_size, truncate_{seq,size} for page_snapc? */ 526 if (snapc && capsnap->context != page_snapc) 527 continue; 528 529 if (ctl) { 530 if (capsnap->writing) { 531 ctl->i_size = i_size_read(inode); 532 ctl->size_stable = false; 533 } else { 534 ctl->i_size = capsnap->size; 535 ctl->size_stable = true; 536 } 537 ctl->truncate_size = capsnap->truncate_size; 538 ctl->truncate_seq = capsnap->truncate_seq; 539 ctl->head_snapc = false; 540 } 541 542 if (snapc) 543 break; 544 545 snapc = ceph_get_snap_context(capsnap->context); 546 if (!page_snapc || 547 page_snapc == snapc || 548 page_snapc->seq > snapc->seq) 549 break; 550 } 551 if (!snapc && ci->i_wrbuffer_ref_head) { 552 snapc = ceph_get_snap_context(ci->i_head_snapc); 553 dout(" head snapc %p has %d dirty pages\n", 554 snapc, ci->i_wrbuffer_ref_head); 555 if (ctl) { 556 ctl->i_size = i_size_read(inode); 557 ctl->truncate_size = ci->i_truncate_size; 558 ctl->truncate_seq = ci->i_truncate_seq; 559 ctl->size_stable = false; 560 ctl->head_snapc = true; 561 } 562 } 563 spin_unlock(&ci->i_ceph_lock); 564 return snapc; 565 } 566 567 static u64 get_writepages_data_length(struct inode *inode, 568 struct page *page, u64 start) 569 { 570 struct ceph_inode_info *ci = ceph_inode(inode); 571 struct ceph_snap_context *snapc = page_snap_context(page); 572 struct ceph_cap_snap *capsnap = NULL; 573 u64 end = i_size_read(inode); 574 575 if (snapc != ci->i_head_snapc) { 576 bool found = false; 577 spin_lock(&ci->i_ceph_lock); 578 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { 579 if (capsnap->context == snapc) { 580 if (!capsnap->writing) 581 end = capsnap->size; 582 found = true; 583 break; 584 } 585 } 586 spin_unlock(&ci->i_ceph_lock); 587 WARN_ON(!found); 588 } 589 if (end > page_offset(page) + PAGE_SIZE) 590 end = page_offset(page) + PAGE_SIZE; 591 return end > start ? end - start : 0; 592 } 593 594 /* 595 * do a synchronous write on N pages 596 */ 597 static int ceph_sync_writepages(struct ceph_fs_client *fsc, 598 struct ceph_vino vino, 599 struct ceph_file_layout *layout, 600 struct ceph_snap_context *snapc, 601 u64 off, u64 len, 602 u32 truncate_seq, u64 truncate_size, 603 struct timespec64 *mtime, 604 struct page **pages, int num_pages) 605 { 606 struct ceph_osd_client *osdc = &fsc->client->osdc; 607 struct ceph_osd_request *req; 608 int rc = 0; 609 int page_align = off & ~PAGE_MASK; 610 611 req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 0, 1, 612 CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE, 613 snapc, truncate_seq, truncate_size, 614 true); 615 if (IS_ERR(req)) 616 return PTR_ERR(req); 617 618 /* it may be a short write due to an object boundary */ 619 osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align, 620 false, false); 621 dout("writepages %llu~%llu (%llu bytes)\n", off, len, len); 622 623 req->r_mtime = *mtime; 624 rc = ceph_osdc_start_request(osdc, req, true); 625 if (!rc) 626 rc = ceph_osdc_wait_request(osdc, req); 627 628 ceph_update_write_latency(&fsc->mdsc->metric, req->r_start_latency, 629 req->r_end_latency, rc); 630 631 ceph_osdc_put_request(req); 632 if (rc == 0) 633 rc = len; 634 dout("writepages result %d\n", rc); 635 return rc; 636 } 637 638 /* 639 * Write a single page, but leave the page locked. 640 * 641 * If we get a write error, mark the mapping for error, but still adjust the 642 * dirty page accounting (i.e., page is no longer dirty). 643 */ 644 static int writepage_nounlock(struct page *page, struct writeback_control *wbc) 645 { 646 struct inode *inode; 647 struct ceph_inode_info *ci; 648 struct ceph_fs_client *fsc; 649 struct ceph_snap_context *snapc, *oldest; 650 loff_t page_off = page_offset(page); 651 int err, len = PAGE_SIZE; 652 struct ceph_writeback_ctl ceph_wbc; 653 654 dout("writepage %p idx %lu\n", page, page->index); 655 656 inode = page->mapping->host; 657 ci = ceph_inode(inode); 658 fsc = ceph_inode_to_client(inode); 659 660 /* verify this is a writeable snap context */ 661 snapc = page_snap_context(page); 662 if (!snapc) { 663 dout("writepage %p page %p not dirty?\n", inode, page); 664 return 0; 665 } 666 oldest = get_oldest_context(inode, &ceph_wbc, snapc); 667 if (snapc->seq > oldest->seq) { 668 dout("writepage %p page %p snapc %p not writeable - noop\n", 669 inode, page, snapc); 670 /* we should only noop if called by kswapd */ 671 WARN_ON(!(current->flags & PF_MEMALLOC)); 672 ceph_put_snap_context(oldest); 673 redirty_page_for_writepage(wbc, page); 674 return 0; 675 } 676 ceph_put_snap_context(oldest); 677 678 /* is this a partial page at end of file? */ 679 if (page_off >= ceph_wbc.i_size) { 680 dout("%p page eof %llu\n", page, ceph_wbc.i_size); 681 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE); 682 return 0; 683 } 684 685 if (ceph_wbc.i_size < page_off + len) 686 len = ceph_wbc.i_size - page_off; 687 688 dout("writepage %p page %p index %lu on %llu~%u snapc %p seq %lld\n", 689 inode, page, page->index, page_off, len, snapc, snapc->seq); 690 691 if (atomic_long_inc_return(&fsc->writeback_count) > 692 CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb)) 693 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC); 694 695 set_page_writeback(page); 696 err = ceph_sync_writepages(fsc, ceph_vino(inode), 697 &ci->i_layout, snapc, page_off, len, 698 ceph_wbc.truncate_seq, 699 ceph_wbc.truncate_size, 700 &inode->i_mtime, &page, 1); 701 if (err < 0) { 702 struct writeback_control tmp_wbc; 703 if (!wbc) 704 wbc = &tmp_wbc; 705 if (err == -ERESTARTSYS) { 706 /* killed by SIGKILL */ 707 dout("writepage interrupted page %p\n", page); 708 redirty_page_for_writepage(wbc, page); 709 end_page_writeback(page); 710 return err; 711 } 712 if (err == -EBLOCKLISTED) 713 fsc->blocklisted = true; 714 dout("writepage setting page/mapping error %d %p\n", 715 err, page); 716 mapping_set_error(&inode->i_data, err); 717 wbc->pages_skipped++; 718 } else { 719 dout("writepage cleaned page %p\n", page); 720 err = 0; /* vfs expects us to return 0 */ 721 } 722 page->private = 0; 723 ClearPagePrivate(page); 724 end_page_writeback(page); 725 ceph_put_wrbuffer_cap_refs(ci, 1, snapc); 726 ceph_put_snap_context(snapc); /* page's reference */ 727 728 if (atomic_long_dec_return(&fsc->writeback_count) < 729 CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb)) 730 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC); 731 732 return err; 733 } 734 735 static int ceph_writepage(struct page *page, struct writeback_control *wbc) 736 { 737 int err; 738 struct inode *inode = page->mapping->host; 739 BUG_ON(!inode); 740 ihold(inode); 741 err = writepage_nounlock(page, wbc); 742 if (err == -ERESTARTSYS) { 743 /* direct memory reclaimer was killed by SIGKILL. return 0 744 * to prevent caller from setting mapping/page error */ 745 err = 0; 746 } 747 unlock_page(page); 748 iput(inode); 749 return err; 750 } 751 752 /* 753 * async writeback completion handler. 754 * 755 * If we get an error, set the mapping error bit, but not the individual 756 * page error bits. 757 */ 758 static void writepages_finish(struct ceph_osd_request *req) 759 { 760 struct inode *inode = req->r_inode; 761 struct ceph_inode_info *ci = ceph_inode(inode); 762 struct ceph_osd_data *osd_data; 763 struct page *page; 764 int num_pages, total_pages = 0; 765 int i, j; 766 int rc = req->r_result; 767 struct ceph_snap_context *snapc = req->r_snapc; 768 struct address_space *mapping = inode->i_mapping; 769 struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 770 bool remove_page; 771 772 dout("writepages_finish %p rc %d\n", inode, rc); 773 if (rc < 0) { 774 mapping_set_error(mapping, rc); 775 ceph_set_error_write(ci); 776 if (rc == -EBLOCKLISTED) 777 fsc->blocklisted = true; 778 } else { 779 ceph_clear_error_write(ci); 780 } 781 782 ceph_update_write_latency(&fsc->mdsc->metric, req->r_start_latency, 783 req->r_end_latency, rc); 784 785 /* 786 * We lost the cache cap, need to truncate the page before 787 * it is unlocked, otherwise we'd truncate it later in the 788 * page truncation thread, possibly losing some data that 789 * raced its way in 790 */ 791 remove_page = !(ceph_caps_issued(ci) & 792 (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)); 793 794 /* clean all pages */ 795 for (i = 0; i < req->r_num_ops; i++) { 796 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE) 797 break; 798 799 osd_data = osd_req_op_extent_osd_data(req, i); 800 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES); 801 num_pages = calc_pages_for((u64)osd_data->alignment, 802 (u64)osd_data->length); 803 total_pages += num_pages; 804 for (j = 0; j < num_pages; j++) { 805 page = osd_data->pages[j]; 806 BUG_ON(!page); 807 WARN_ON(!PageUptodate(page)); 808 809 if (atomic_long_dec_return(&fsc->writeback_count) < 810 CONGESTION_OFF_THRESH( 811 fsc->mount_options->congestion_kb)) 812 clear_bdi_congested(inode_to_bdi(inode), 813 BLK_RW_ASYNC); 814 815 ceph_put_snap_context(page_snap_context(page)); 816 page->private = 0; 817 ClearPagePrivate(page); 818 dout("unlocking %p\n", page); 819 end_page_writeback(page); 820 821 if (remove_page) 822 generic_error_remove_page(inode->i_mapping, 823 page); 824 825 unlock_page(page); 826 } 827 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n", 828 inode, osd_data->length, rc >= 0 ? num_pages : 0); 829 830 release_pages(osd_data->pages, num_pages); 831 } 832 833 ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc); 834 835 osd_data = osd_req_op_extent_osd_data(req, 0); 836 if (osd_data->pages_from_pool) 837 mempool_free(osd_data->pages, ceph_wb_pagevec_pool); 838 else 839 kfree(osd_data->pages); 840 ceph_osdc_put_request(req); 841 } 842 843 /* 844 * initiate async writeback 845 */ 846 static int ceph_writepages_start(struct address_space *mapping, 847 struct writeback_control *wbc) 848 { 849 struct inode *inode = mapping->host; 850 struct ceph_inode_info *ci = ceph_inode(inode); 851 struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 852 struct ceph_vino vino = ceph_vino(inode); 853 pgoff_t index, start_index, end = -1; 854 struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc; 855 struct pagevec pvec; 856 int rc = 0; 857 unsigned int wsize = i_blocksize(inode); 858 struct ceph_osd_request *req = NULL; 859 struct ceph_writeback_ctl ceph_wbc; 860 bool should_loop, range_whole = false; 861 bool done = false; 862 863 dout("writepages_start %p (mode=%s)\n", inode, 864 wbc->sync_mode == WB_SYNC_NONE ? "NONE" : 865 (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD")); 866 867 if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) { 868 if (ci->i_wrbuffer_ref > 0) { 869 pr_warn_ratelimited( 870 "writepage_start %p %lld forced umount\n", 871 inode, ceph_ino(inode)); 872 } 873 mapping_set_error(mapping, -EIO); 874 return -EIO; /* we're in a forced umount, don't write! */ 875 } 876 if (fsc->mount_options->wsize < wsize) 877 wsize = fsc->mount_options->wsize; 878 879 pagevec_init(&pvec); 880 881 start_index = wbc->range_cyclic ? mapping->writeback_index : 0; 882 index = start_index; 883 884 retry: 885 /* find oldest snap context with dirty data */ 886 snapc = get_oldest_context(inode, &ceph_wbc, NULL); 887 if (!snapc) { 888 /* hmm, why does writepages get called when there 889 is no dirty data? */ 890 dout(" no snap context with dirty data?\n"); 891 goto out; 892 } 893 dout(" oldest snapc is %p seq %lld (%d snaps)\n", 894 snapc, snapc->seq, snapc->num_snaps); 895 896 should_loop = false; 897 if (ceph_wbc.head_snapc && snapc != last_snapc) { 898 /* where to start/end? */ 899 if (wbc->range_cyclic) { 900 index = start_index; 901 end = -1; 902 if (index > 0) 903 should_loop = true; 904 dout(" cyclic, start at %lu\n", index); 905 } else { 906 index = wbc->range_start >> PAGE_SHIFT; 907 end = wbc->range_end >> PAGE_SHIFT; 908 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX) 909 range_whole = true; 910 dout(" not cyclic, %lu to %lu\n", index, end); 911 } 912 } else if (!ceph_wbc.head_snapc) { 913 /* Do not respect wbc->range_{start,end}. Dirty pages 914 * in that range can be associated with newer snapc. 915 * They are not writeable until we write all dirty pages 916 * associated with 'snapc' get written */ 917 if (index > 0) 918 should_loop = true; 919 dout(" non-head snapc, range whole\n"); 920 } 921 922 ceph_put_snap_context(last_snapc); 923 last_snapc = snapc; 924 925 while (!done && index <= end) { 926 int num_ops = 0, op_idx; 927 unsigned i, pvec_pages, max_pages, locked_pages = 0; 928 struct page **pages = NULL, **data_pages; 929 struct page *page; 930 pgoff_t strip_unit_end = 0; 931 u64 offset = 0, len = 0; 932 bool from_pool = false; 933 934 max_pages = wsize >> PAGE_SHIFT; 935 936 get_more_pages: 937 pvec_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, 938 end, PAGECACHE_TAG_DIRTY); 939 dout("pagevec_lookup_range_tag got %d\n", pvec_pages); 940 if (!pvec_pages && !locked_pages) 941 break; 942 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) { 943 page = pvec.pages[i]; 944 dout("? %p idx %lu\n", page, page->index); 945 if (locked_pages == 0) 946 lock_page(page); /* first page */ 947 else if (!trylock_page(page)) 948 break; 949 950 /* only dirty pages, or our accounting breaks */ 951 if (unlikely(!PageDirty(page)) || 952 unlikely(page->mapping != mapping)) { 953 dout("!dirty or !mapping %p\n", page); 954 unlock_page(page); 955 continue; 956 } 957 /* only if matching snap context */ 958 pgsnapc = page_snap_context(page); 959 if (pgsnapc != snapc) { 960 dout("page snapc %p %lld != oldest %p %lld\n", 961 pgsnapc, pgsnapc->seq, snapc, snapc->seq); 962 if (!should_loop && 963 !ceph_wbc.head_snapc && 964 wbc->sync_mode != WB_SYNC_NONE) 965 should_loop = true; 966 unlock_page(page); 967 continue; 968 } 969 if (page_offset(page) >= ceph_wbc.i_size) { 970 dout("%p page eof %llu\n", 971 page, ceph_wbc.i_size); 972 if ((ceph_wbc.size_stable || 973 page_offset(page) >= i_size_read(inode)) && 974 clear_page_dirty_for_io(page)) 975 mapping->a_ops->invalidatepage(page, 976 0, PAGE_SIZE); 977 unlock_page(page); 978 continue; 979 } 980 if (strip_unit_end && (page->index > strip_unit_end)) { 981 dout("end of strip unit %p\n", page); 982 unlock_page(page); 983 break; 984 } 985 if (PageWriteback(page)) { 986 if (wbc->sync_mode == WB_SYNC_NONE) { 987 dout("%p under writeback\n", page); 988 unlock_page(page); 989 continue; 990 } 991 dout("waiting on writeback %p\n", page); 992 wait_on_page_writeback(page); 993 } 994 995 if (!clear_page_dirty_for_io(page)) { 996 dout("%p !clear_page_dirty_for_io\n", page); 997 unlock_page(page); 998 continue; 999 } 1000 1001 /* 1002 * We have something to write. If this is 1003 * the first locked page this time through, 1004 * calculate max possinle write size and 1005 * allocate a page array 1006 */ 1007 if (locked_pages == 0) { 1008 u64 objnum; 1009 u64 objoff; 1010 u32 xlen; 1011 1012 /* prepare async write request */ 1013 offset = (u64)page_offset(page); 1014 ceph_calc_file_object_mapping(&ci->i_layout, 1015 offset, wsize, 1016 &objnum, &objoff, 1017 &xlen); 1018 len = xlen; 1019 1020 num_ops = 1; 1021 strip_unit_end = page->index + 1022 ((len - 1) >> PAGE_SHIFT); 1023 1024 BUG_ON(pages); 1025 max_pages = calc_pages_for(0, (u64)len); 1026 pages = kmalloc_array(max_pages, 1027 sizeof(*pages), 1028 GFP_NOFS); 1029 if (!pages) { 1030 from_pool = true; 1031 pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS); 1032 BUG_ON(!pages); 1033 } 1034 1035 len = 0; 1036 } else if (page->index != 1037 (offset + len) >> PAGE_SHIFT) { 1038 if (num_ops >= (from_pool ? CEPH_OSD_SLAB_OPS : 1039 CEPH_OSD_MAX_OPS)) { 1040 redirty_page_for_writepage(wbc, page); 1041 unlock_page(page); 1042 break; 1043 } 1044 1045 num_ops++; 1046 offset = (u64)page_offset(page); 1047 len = 0; 1048 } 1049 1050 /* note position of first page in pvec */ 1051 dout("%p will write page %p idx %lu\n", 1052 inode, page, page->index); 1053 1054 if (atomic_long_inc_return(&fsc->writeback_count) > 1055 CONGESTION_ON_THRESH( 1056 fsc->mount_options->congestion_kb)) { 1057 set_bdi_congested(inode_to_bdi(inode), 1058 BLK_RW_ASYNC); 1059 } 1060 1061 1062 pages[locked_pages++] = page; 1063 pvec.pages[i] = NULL; 1064 1065 len += PAGE_SIZE; 1066 } 1067 1068 /* did we get anything? */ 1069 if (!locked_pages) 1070 goto release_pvec_pages; 1071 if (i) { 1072 unsigned j, n = 0; 1073 /* shift unused page to beginning of pvec */ 1074 for (j = 0; j < pvec_pages; j++) { 1075 if (!pvec.pages[j]) 1076 continue; 1077 if (n < j) 1078 pvec.pages[n] = pvec.pages[j]; 1079 n++; 1080 } 1081 pvec.nr = n; 1082 1083 if (pvec_pages && i == pvec_pages && 1084 locked_pages < max_pages) { 1085 dout("reached end pvec, trying for more\n"); 1086 pagevec_release(&pvec); 1087 goto get_more_pages; 1088 } 1089 } 1090 1091 new_request: 1092 offset = page_offset(pages[0]); 1093 len = wsize; 1094 1095 req = ceph_osdc_new_request(&fsc->client->osdc, 1096 &ci->i_layout, vino, 1097 offset, &len, 0, num_ops, 1098 CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE, 1099 snapc, ceph_wbc.truncate_seq, 1100 ceph_wbc.truncate_size, false); 1101 if (IS_ERR(req)) { 1102 req = ceph_osdc_new_request(&fsc->client->osdc, 1103 &ci->i_layout, vino, 1104 offset, &len, 0, 1105 min(num_ops, 1106 CEPH_OSD_SLAB_OPS), 1107 CEPH_OSD_OP_WRITE, 1108 CEPH_OSD_FLAG_WRITE, 1109 snapc, ceph_wbc.truncate_seq, 1110 ceph_wbc.truncate_size, true); 1111 BUG_ON(IS_ERR(req)); 1112 } 1113 BUG_ON(len < page_offset(pages[locked_pages - 1]) + 1114 PAGE_SIZE - offset); 1115 1116 req->r_callback = writepages_finish; 1117 req->r_inode = inode; 1118 1119 /* Format the osd request message and submit the write */ 1120 len = 0; 1121 data_pages = pages; 1122 op_idx = 0; 1123 for (i = 0; i < locked_pages; i++) { 1124 u64 cur_offset = page_offset(pages[i]); 1125 if (offset + len != cur_offset) { 1126 if (op_idx + 1 == req->r_num_ops) 1127 break; 1128 osd_req_op_extent_dup_last(req, op_idx, 1129 cur_offset - offset); 1130 dout("writepages got pages at %llu~%llu\n", 1131 offset, len); 1132 osd_req_op_extent_osd_data_pages(req, op_idx, 1133 data_pages, len, 0, 1134 from_pool, false); 1135 osd_req_op_extent_update(req, op_idx, len); 1136 1137 len = 0; 1138 offset = cur_offset; 1139 data_pages = pages + i; 1140 op_idx++; 1141 } 1142 1143 set_page_writeback(pages[i]); 1144 len += PAGE_SIZE; 1145 } 1146 1147 if (ceph_wbc.size_stable) { 1148 len = min(len, ceph_wbc.i_size - offset); 1149 } else if (i == locked_pages) { 1150 /* writepages_finish() clears writeback pages 1151 * according to the data length, so make sure 1152 * data length covers all locked pages */ 1153 u64 min_len = len + 1 - PAGE_SIZE; 1154 len = get_writepages_data_length(inode, pages[i - 1], 1155 offset); 1156 len = max(len, min_len); 1157 } 1158 dout("writepages got pages at %llu~%llu\n", offset, len); 1159 1160 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len, 1161 0, from_pool, false); 1162 osd_req_op_extent_update(req, op_idx, len); 1163 1164 BUG_ON(op_idx + 1 != req->r_num_ops); 1165 1166 from_pool = false; 1167 if (i < locked_pages) { 1168 BUG_ON(num_ops <= req->r_num_ops); 1169 num_ops -= req->r_num_ops; 1170 locked_pages -= i; 1171 1172 /* allocate new pages array for next request */ 1173 data_pages = pages; 1174 pages = kmalloc_array(locked_pages, sizeof(*pages), 1175 GFP_NOFS); 1176 if (!pages) { 1177 from_pool = true; 1178 pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS); 1179 BUG_ON(!pages); 1180 } 1181 memcpy(pages, data_pages + i, 1182 locked_pages * sizeof(*pages)); 1183 memset(data_pages + i, 0, 1184 locked_pages * sizeof(*pages)); 1185 } else { 1186 BUG_ON(num_ops != req->r_num_ops); 1187 index = pages[i - 1]->index + 1; 1188 /* request message now owns the pages array */ 1189 pages = NULL; 1190 } 1191 1192 req->r_mtime = inode->i_mtime; 1193 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true); 1194 BUG_ON(rc); 1195 req = NULL; 1196 1197 wbc->nr_to_write -= i; 1198 if (pages) 1199 goto new_request; 1200 1201 /* 1202 * We stop writing back only if we are not doing 1203 * integrity sync. In case of integrity sync we have to 1204 * keep going until we have written all the pages 1205 * we tagged for writeback prior to entering this loop. 1206 */ 1207 if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE) 1208 done = true; 1209 1210 release_pvec_pages: 1211 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr, 1212 pvec.nr ? pvec.pages[0] : NULL); 1213 pagevec_release(&pvec); 1214 } 1215 1216 if (should_loop && !done) { 1217 /* more to do; loop back to beginning of file */ 1218 dout("writepages looping back to beginning of file\n"); 1219 end = start_index - 1; /* OK even when start_index == 0 */ 1220 1221 /* to write dirty pages associated with next snapc, 1222 * we need to wait until current writes complete */ 1223 if (wbc->sync_mode != WB_SYNC_NONE && 1224 start_index == 0 && /* all dirty pages were checked */ 1225 !ceph_wbc.head_snapc) { 1226 struct page *page; 1227 unsigned i, nr; 1228 index = 0; 1229 while ((index <= end) && 1230 (nr = pagevec_lookup_tag(&pvec, mapping, &index, 1231 PAGECACHE_TAG_WRITEBACK))) { 1232 for (i = 0; i < nr; i++) { 1233 page = pvec.pages[i]; 1234 if (page_snap_context(page) != snapc) 1235 continue; 1236 wait_on_page_writeback(page); 1237 } 1238 pagevec_release(&pvec); 1239 cond_resched(); 1240 } 1241 } 1242 1243 start_index = 0; 1244 index = 0; 1245 goto retry; 1246 } 1247 1248 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) 1249 mapping->writeback_index = index; 1250 1251 out: 1252 ceph_osdc_put_request(req); 1253 ceph_put_snap_context(last_snapc); 1254 dout("writepages dend - startone, rc = %d\n", rc); 1255 return rc; 1256 } 1257 1258 1259 1260 /* 1261 * See if a given @snapc is either writeable, or already written. 1262 */ 1263 static int context_is_writeable_or_written(struct inode *inode, 1264 struct ceph_snap_context *snapc) 1265 { 1266 struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL); 1267 int ret = !oldest || snapc->seq <= oldest->seq; 1268 1269 ceph_put_snap_context(oldest); 1270 return ret; 1271 } 1272 1273 /** 1274 * ceph_find_incompatible - find an incompatible context and return it 1275 * @page: page being dirtied 1276 * 1277 * We are only allowed to write into/dirty a page if the page is 1278 * clean, or already dirty within the same snap context. Returns a 1279 * conflicting context if there is one, NULL if there isn't, or a 1280 * negative error code on other errors. 1281 * 1282 * Must be called with page lock held. 1283 */ 1284 static struct ceph_snap_context * 1285 ceph_find_incompatible(struct page *page) 1286 { 1287 struct inode *inode = page->mapping->host; 1288 struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 1289 struct ceph_inode_info *ci = ceph_inode(inode); 1290 1291 if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) { 1292 dout(" page %p forced umount\n", page); 1293 return ERR_PTR(-EIO); 1294 } 1295 1296 for (;;) { 1297 struct ceph_snap_context *snapc, *oldest; 1298 1299 wait_on_page_writeback(page); 1300 1301 snapc = page_snap_context(page); 1302 if (!snapc || snapc == ci->i_head_snapc) 1303 break; 1304 1305 /* 1306 * this page is already dirty in another (older) snap 1307 * context! is it writeable now? 1308 */ 1309 oldest = get_oldest_context(inode, NULL, NULL); 1310 if (snapc->seq > oldest->seq) { 1311 /* not writeable -- return it for the caller to deal with */ 1312 ceph_put_snap_context(oldest); 1313 dout(" page %p snapc %p not current or oldest\n", page, snapc); 1314 return ceph_get_snap_context(snapc); 1315 } 1316 ceph_put_snap_context(oldest); 1317 1318 /* yay, writeable, do it now (without dropping page lock) */ 1319 dout(" page %p snapc %p not current, but oldest\n", page, snapc); 1320 if (clear_page_dirty_for_io(page)) { 1321 int r = writepage_nounlock(page, NULL); 1322 if (r < 0) 1323 return ERR_PTR(r); 1324 } 1325 } 1326 return NULL; 1327 } 1328 1329 /* 1330 * We are only allowed to write into/dirty the page if the page is 1331 * clean, or already dirty within the same snap context. 1332 * 1333 * called with page locked. 1334 * return success with page locked, 1335 * or any failure (incl -EAGAIN) with page unlocked. 1336 */ 1337 static int ceph_update_writeable_page(struct file *file, 1338 loff_t pos, unsigned len, 1339 struct page *page) 1340 { 1341 struct inode *inode = file_inode(file); 1342 struct ceph_inode_info *ci = ceph_inode(inode); 1343 struct ceph_snap_context *snapc; 1344 loff_t page_off = pos & PAGE_MASK; 1345 int pos_in_page = pos & ~PAGE_MASK; 1346 int end_in_page = pos_in_page + len; 1347 loff_t i_size; 1348 int r; 1349 1350 retry_locked: 1351 snapc = ceph_find_incompatible(page); 1352 if (snapc) { 1353 if (IS_ERR(snapc)) { 1354 r = PTR_ERR(snapc); 1355 goto fail_unlock; 1356 } 1357 unlock_page(page); 1358 ceph_queue_writeback(inode); 1359 r = wait_event_killable(ci->i_cap_wq, 1360 context_is_writeable_or_written(inode, snapc)); 1361 ceph_put_snap_context(snapc); 1362 return -EAGAIN; 1363 } 1364 1365 if (PageUptodate(page)) { 1366 dout(" page %p already uptodate\n", page); 1367 return 0; 1368 } 1369 1370 /* full page? */ 1371 if (pos_in_page == 0 && len == PAGE_SIZE) 1372 return 0; 1373 1374 /* past end of file? */ 1375 i_size = i_size_read(inode); 1376 1377 if (page_off >= i_size || 1378 (pos_in_page == 0 && (pos+len) >= i_size && 1379 end_in_page - pos_in_page != PAGE_SIZE)) { 1380 dout(" zeroing %p 0 - %d and %d - %d\n", 1381 page, pos_in_page, end_in_page, (int)PAGE_SIZE); 1382 zero_user_segments(page, 1383 0, pos_in_page, 1384 end_in_page, PAGE_SIZE); 1385 return 0; 1386 } 1387 1388 /* we need to read it. */ 1389 r = ceph_do_readpage(file, page); 1390 if (r < 0) { 1391 if (r == -EINPROGRESS) 1392 return -EAGAIN; 1393 goto fail_unlock; 1394 } 1395 goto retry_locked; 1396 fail_unlock: 1397 unlock_page(page); 1398 return r; 1399 } 1400 1401 /* 1402 * We are only allowed to write into/dirty the page if the page is 1403 * clean, or already dirty within the same snap context. 1404 */ 1405 static int ceph_write_begin(struct file *file, struct address_space *mapping, 1406 loff_t pos, unsigned len, unsigned flags, 1407 struct page **pagep, void **fsdata) 1408 { 1409 struct inode *inode = file_inode(file); 1410 struct page *page; 1411 pgoff_t index = pos >> PAGE_SHIFT; 1412 int r; 1413 1414 do { 1415 /* get a page */ 1416 page = grab_cache_page_write_begin(mapping, index, 0); 1417 if (!page) 1418 return -ENOMEM; 1419 1420 dout("write_begin file %p inode %p page %p %d~%d\n", file, 1421 inode, page, (int)pos, (int)len); 1422 1423 r = ceph_update_writeable_page(file, pos, len, page); 1424 if (r < 0) 1425 put_page(page); 1426 else 1427 *pagep = page; 1428 } while (r == -EAGAIN); 1429 1430 return r; 1431 } 1432 1433 /* 1434 * we don't do anything in here that simple_write_end doesn't do 1435 * except adjust dirty page accounting 1436 */ 1437 static int ceph_write_end(struct file *file, struct address_space *mapping, 1438 loff_t pos, unsigned len, unsigned copied, 1439 struct page *page, void *fsdata) 1440 { 1441 struct inode *inode = file_inode(file); 1442 bool check_cap = false; 1443 1444 dout("write_end file %p inode %p page %p %d~%d (%d)\n", file, 1445 inode, page, (int)pos, (int)copied, (int)len); 1446 1447 /* zero the stale part of the page if we did a short copy */ 1448 if (!PageUptodate(page)) { 1449 if (copied < len) { 1450 copied = 0; 1451 goto out; 1452 } 1453 SetPageUptodate(page); 1454 } 1455 1456 /* did file size increase? */ 1457 if (pos+copied > i_size_read(inode)) 1458 check_cap = ceph_inode_set_size(inode, pos+copied); 1459 1460 set_page_dirty(page); 1461 1462 out: 1463 unlock_page(page); 1464 put_page(page); 1465 1466 if (check_cap) 1467 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL); 1468 1469 return copied; 1470 } 1471 1472 /* 1473 * we set .direct_IO to indicate direct io is supported, but since we 1474 * intercept O_DIRECT reads and writes early, this function should 1475 * never get called. 1476 */ 1477 static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter) 1478 { 1479 WARN_ON(1); 1480 return -EINVAL; 1481 } 1482 1483 const struct address_space_operations ceph_aops = { 1484 .readpage = ceph_readpage, 1485 .readpages = ceph_readpages, 1486 .writepage = ceph_writepage, 1487 .writepages = ceph_writepages_start, 1488 .write_begin = ceph_write_begin, 1489 .write_end = ceph_write_end, 1490 .set_page_dirty = ceph_set_page_dirty, 1491 .invalidatepage = ceph_invalidatepage, 1492 .releasepage = ceph_releasepage, 1493 .direct_IO = ceph_direct_io, 1494 }; 1495 1496 static void ceph_block_sigs(sigset_t *oldset) 1497 { 1498 sigset_t mask; 1499 siginitsetinv(&mask, sigmask(SIGKILL)); 1500 sigprocmask(SIG_BLOCK, &mask, oldset); 1501 } 1502 1503 static void ceph_restore_sigs(sigset_t *oldset) 1504 { 1505 sigprocmask(SIG_SETMASK, oldset, NULL); 1506 } 1507 1508 /* 1509 * vm ops 1510 */ 1511 static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf) 1512 { 1513 struct vm_area_struct *vma = vmf->vma; 1514 struct inode *inode = file_inode(vma->vm_file); 1515 struct ceph_inode_info *ci = ceph_inode(inode); 1516 struct ceph_file_info *fi = vma->vm_file->private_data; 1517 struct page *pinned_page = NULL; 1518 loff_t off = vmf->pgoff << PAGE_SHIFT; 1519 int want, got, err; 1520 sigset_t oldset; 1521 vm_fault_t ret = VM_FAULT_SIGBUS; 1522 1523 ceph_block_sigs(&oldset); 1524 1525 dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n", 1526 inode, ceph_vinop(inode), off, (size_t)PAGE_SIZE); 1527 if (fi->fmode & CEPH_FILE_MODE_LAZY) 1528 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO; 1529 else 1530 want = CEPH_CAP_FILE_CACHE; 1531 1532 got = 0; 1533 err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_RD, want, -1, 1534 &got, &pinned_page); 1535 if (err < 0) 1536 goto out_restore; 1537 1538 dout("filemap_fault %p %llu~%zd got cap refs on %s\n", 1539 inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got)); 1540 1541 if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) || 1542 ci->i_inline_version == CEPH_INLINE_NONE) { 1543 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got); 1544 ceph_add_rw_context(fi, &rw_ctx); 1545 ret = filemap_fault(vmf); 1546 ceph_del_rw_context(fi, &rw_ctx); 1547 dout("filemap_fault %p %llu~%zd drop cap refs %s ret %x\n", 1548 inode, off, (size_t)PAGE_SIZE, 1549 ceph_cap_string(got), ret); 1550 } else 1551 err = -EAGAIN; 1552 1553 if (pinned_page) 1554 put_page(pinned_page); 1555 ceph_put_cap_refs(ci, got); 1556 1557 if (err != -EAGAIN) 1558 goto out_restore; 1559 1560 /* read inline data */ 1561 if (off >= PAGE_SIZE) { 1562 /* does not support inline data > PAGE_SIZE */ 1563 ret = VM_FAULT_SIGBUS; 1564 } else { 1565 struct address_space *mapping = inode->i_mapping; 1566 struct page *page = find_or_create_page(mapping, 0, 1567 mapping_gfp_constraint(mapping, 1568 ~__GFP_FS)); 1569 if (!page) { 1570 ret = VM_FAULT_OOM; 1571 goto out_inline; 1572 } 1573 err = __ceph_do_getattr(inode, page, 1574 CEPH_STAT_CAP_INLINE_DATA, true); 1575 if (err < 0 || off >= i_size_read(inode)) { 1576 unlock_page(page); 1577 put_page(page); 1578 ret = vmf_error(err); 1579 goto out_inline; 1580 } 1581 if (err < PAGE_SIZE) 1582 zero_user_segment(page, err, PAGE_SIZE); 1583 else 1584 flush_dcache_page(page); 1585 SetPageUptodate(page); 1586 vmf->page = page; 1587 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED; 1588 out_inline: 1589 dout("filemap_fault %p %llu~%zd read inline data ret %x\n", 1590 inode, off, (size_t)PAGE_SIZE, ret); 1591 } 1592 out_restore: 1593 ceph_restore_sigs(&oldset); 1594 if (err < 0) 1595 ret = vmf_error(err); 1596 1597 return ret; 1598 } 1599 1600 /* 1601 * Reuse write_begin here for simplicity. 1602 */ 1603 static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf) 1604 { 1605 struct vm_area_struct *vma = vmf->vma; 1606 struct inode *inode = file_inode(vma->vm_file); 1607 struct ceph_inode_info *ci = ceph_inode(inode); 1608 struct ceph_file_info *fi = vma->vm_file->private_data; 1609 struct ceph_cap_flush *prealloc_cf; 1610 struct page *page = vmf->page; 1611 loff_t off = page_offset(page); 1612 loff_t size = i_size_read(inode); 1613 size_t len; 1614 int want, got, err; 1615 sigset_t oldset; 1616 vm_fault_t ret = VM_FAULT_SIGBUS; 1617 1618 prealloc_cf = ceph_alloc_cap_flush(); 1619 if (!prealloc_cf) 1620 return VM_FAULT_OOM; 1621 1622 sb_start_pagefault(inode->i_sb); 1623 ceph_block_sigs(&oldset); 1624 1625 if (ci->i_inline_version != CEPH_INLINE_NONE) { 1626 struct page *locked_page = NULL; 1627 if (off == 0) { 1628 lock_page(page); 1629 locked_page = page; 1630 } 1631 err = ceph_uninline_data(vma->vm_file, locked_page); 1632 if (locked_page) 1633 unlock_page(locked_page); 1634 if (err < 0) 1635 goto out_free; 1636 } 1637 1638 if (off + PAGE_SIZE <= size) 1639 len = PAGE_SIZE; 1640 else 1641 len = size & ~PAGE_MASK; 1642 1643 dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n", 1644 inode, ceph_vinop(inode), off, len, size); 1645 if (fi->fmode & CEPH_FILE_MODE_LAZY) 1646 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO; 1647 else 1648 want = CEPH_CAP_FILE_BUFFER; 1649 1650 got = 0; 1651 err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_WR, want, off + len, 1652 &got, NULL); 1653 if (err < 0) 1654 goto out_free; 1655 1656 dout("page_mkwrite %p %llu~%zd got cap refs on %s\n", 1657 inode, off, len, ceph_cap_string(got)); 1658 1659 /* Update time before taking page lock */ 1660 file_update_time(vma->vm_file); 1661 inode_inc_iversion_raw(inode); 1662 1663 do { 1664 struct ceph_snap_context *snapc; 1665 1666 lock_page(page); 1667 1668 if (page_mkwrite_check_truncate(page, inode) < 0) { 1669 unlock_page(page); 1670 ret = VM_FAULT_NOPAGE; 1671 break; 1672 } 1673 1674 snapc = ceph_find_incompatible(page); 1675 if (!snapc) { 1676 /* success. we'll keep the page locked. */ 1677 set_page_dirty(page); 1678 ret = VM_FAULT_LOCKED; 1679 break; 1680 } 1681 1682 unlock_page(page); 1683 1684 if (IS_ERR(snapc)) { 1685 ret = VM_FAULT_SIGBUS; 1686 break; 1687 } 1688 1689 ceph_queue_writeback(inode); 1690 err = wait_event_killable(ci->i_cap_wq, 1691 context_is_writeable_or_written(inode, snapc)); 1692 ceph_put_snap_context(snapc); 1693 } while (err == 0); 1694 1695 if (ret == VM_FAULT_LOCKED || 1696 ci->i_inline_version != CEPH_INLINE_NONE) { 1697 int dirty; 1698 spin_lock(&ci->i_ceph_lock); 1699 ci->i_inline_version = CEPH_INLINE_NONE; 1700 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR, 1701 &prealloc_cf); 1702 spin_unlock(&ci->i_ceph_lock); 1703 if (dirty) 1704 __mark_inode_dirty(inode, dirty); 1705 } 1706 1707 dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %x\n", 1708 inode, off, len, ceph_cap_string(got), ret); 1709 ceph_put_cap_refs(ci, got); 1710 out_free: 1711 ceph_restore_sigs(&oldset); 1712 sb_end_pagefault(inode->i_sb); 1713 ceph_free_cap_flush(prealloc_cf); 1714 if (err < 0) 1715 ret = vmf_error(err); 1716 return ret; 1717 } 1718 1719 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page, 1720 char *data, size_t len) 1721 { 1722 struct address_space *mapping = inode->i_mapping; 1723 struct page *page; 1724 1725 if (locked_page) { 1726 page = locked_page; 1727 } else { 1728 if (i_size_read(inode) == 0) 1729 return; 1730 page = find_or_create_page(mapping, 0, 1731 mapping_gfp_constraint(mapping, 1732 ~__GFP_FS)); 1733 if (!page) 1734 return; 1735 if (PageUptodate(page)) { 1736 unlock_page(page); 1737 put_page(page); 1738 return; 1739 } 1740 } 1741 1742 dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n", 1743 inode, ceph_vinop(inode), len, locked_page); 1744 1745 if (len > 0) { 1746 void *kaddr = kmap_atomic(page); 1747 memcpy(kaddr, data, len); 1748 kunmap_atomic(kaddr); 1749 } 1750 1751 if (page != locked_page) { 1752 if (len < PAGE_SIZE) 1753 zero_user_segment(page, len, PAGE_SIZE); 1754 else 1755 flush_dcache_page(page); 1756 1757 SetPageUptodate(page); 1758 unlock_page(page); 1759 put_page(page); 1760 } 1761 } 1762 1763 int ceph_uninline_data(struct file *filp, struct page *locked_page) 1764 { 1765 struct inode *inode = file_inode(filp); 1766 struct ceph_inode_info *ci = ceph_inode(inode); 1767 struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 1768 struct ceph_osd_request *req; 1769 struct page *page = NULL; 1770 u64 len, inline_version; 1771 int err = 0; 1772 bool from_pagecache = false; 1773 1774 spin_lock(&ci->i_ceph_lock); 1775 inline_version = ci->i_inline_version; 1776 spin_unlock(&ci->i_ceph_lock); 1777 1778 dout("uninline_data %p %llx.%llx inline_version %llu\n", 1779 inode, ceph_vinop(inode), inline_version); 1780 1781 if (inline_version == 1 || /* initial version, no data */ 1782 inline_version == CEPH_INLINE_NONE) 1783 goto out; 1784 1785 if (locked_page) { 1786 page = locked_page; 1787 WARN_ON(!PageUptodate(page)); 1788 } else if (ceph_caps_issued(ci) & 1789 (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) { 1790 page = find_get_page(inode->i_mapping, 0); 1791 if (page) { 1792 if (PageUptodate(page)) { 1793 from_pagecache = true; 1794 lock_page(page); 1795 } else { 1796 put_page(page); 1797 page = NULL; 1798 } 1799 } 1800 } 1801 1802 if (page) { 1803 len = i_size_read(inode); 1804 if (len > PAGE_SIZE) 1805 len = PAGE_SIZE; 1806 } else { 1807 page = __page_cache_alloc(GFP_NOFS); 1808 if (!page) { 1809 err = -ENOMEM; 1810 goto out; 1811 } 1812 err = __ceph_do_getattr(inode, page, 1813 CEPH_STAT_CAP_INLINE_DATA, true); 1814 if (err < 0) { 1815 /* no inline data */ 1816 if (err == -ENODATA) 1817 err = 0; 1818 goto out; 1819 } 1820 len = err; 1821 } 1822 1823 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, 1824 ceph_vino(inode), 0, &len, 0, 1, 1825 CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE, 1826 NULL, 0, 0, false); 1827 if (IS_ERR(req)) { 1828 err = PTR_ERR(req); 1829 goto out; 1830 } 1831 1832 req->r_mtime = inode->i_mtime; 1833 err = ceph_osdc_start_request(&fsc->client->osdc, req, false); 1834 if (!err) 1835 err = ceph_osdc_wait_request(&fsc->client->osdc, req); 1836 ceph_osdc_put_request(req); 1837 if (err < 0) 1838 goto out; 1839 1840 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, 1841 ceph_vino(inode), 0, &len, 1, 3, 1842 CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE, 1843 NULL, ci->i_truncate_seq, 1844 ci->i_truncate_size, false); 1845 if (IS_ERR(req)) { 1846 err = PTR_ERR(req); 1847 goto out; 1848 } 1849 1850 osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false); 1851 1852 { 1853 __le64 xattr_buf = cpu_to_le64(inline_version); 1854 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR, 1855 "inline_version", &xattr_buf, 1856 sizeof(xattr_buf), 1857 CEPH_OSD_CMPXATTR_OP_GT, 1858 CEPH_OSD_CMPXATTR_MODE_U64); 1859 if (err) 1860 goto out_put; 1861 } 1862 1863 { 1864 char xattr_buf[32]; 1865 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf), 1866 "%llu", inline_version); 1867 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR, 1868 "inline_version", 1869 xattr_buf, xattr_len, 0, 0); 1870 if (err) 1871 goto out_put; 1872 } 1873 1874 req->r_mtime = inode->i_mtime; 1875 err = ceph_osdc_start_request(&fsc->client->osdc, req, false); 1876 if (!err) 1877 err = ceph_osdc_wait_request(&fsc->client->osdc, req); 1878 1879 ceph_update_write_latency(&fsc->mdsc->metric, req->r_start_latency, 1880 req->r_end_latency, err); 1881 1882 out_put: 1883 ceph_osdc_put_request(req); 1884 if (err == -ECANCELED) 1885 err = 0; 1886 out: 1887 if (page && page != locked_page) { 1888 if (from_pagecache) { 1889 unlock_page(page); 1890 put_page(page); 1891 } else 1892 __free_pages(page, 0); 1893 } 1894 1895 dout("uninline_data %p %llx.%llx inline_version %llu = %d\n", 1896 inode, ceph_vinop(inode), inline_version, err); 1897 return err; 1898 } 1899 1900 static const struct vm_operations_struct ceph_vmops = { 1901 .fault = ceph_filemap_fault, 1902 .page_mkwrite = ceph_page_mkwrite, 1903 }; 1904 1905 int ceph_mmap(struct file *file, struct vm_area_struct *vma) 1906 { 1907 struct address_space *mapping = file->f_mapping; 1908 1909 if (!mapping->a_ops->readpage) 1910 return -ENOEXEC; 1911 file_accessed(file); 1912 vma->vm_ops = &ceph_vmops; 1913 return 0; 1914 } 1915 1916 enum { 1917 POOL_READ = 1, 1918 POOL_WRITE = 2, 1919 }; 1920 1921 static int __ceph_pool_perm_get(struct ceph_inode_info *ci, 1922 s64 pool, struct ceph_string *pool_ns) 1923 { 1924 struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode); 1925 struct ceph_mds_client *mdsc = fsc->mdsc; 1926 struct ceph_osd_request *rd_req = NULL, *wr_req = NULL; 1927 struct rb_node **p, *parent; 1928 struct ceph_pool_perm *perm; 1929 struct page **pages; 1930 size_t pool_ns_len; 1931 int err = 0, err2 = 0, have = 0; 1932 1933 down_read(&mdsc->pool_perm_rwsem); 1934 p = &mdsc->pool_perm_tree.rb_node; 1935 while (*p) { 1936 perm = rb_entry(*p, struct ceph_pool_perm, node); 1937 if (pool < perm->pool) 1938 p = &(*p)->rb_left; 1939 else if (pool > perm->pool) 1940 p = &(*p)->rb_right; 1941 else { 1942 int ret = ceph_compare_string(pool_ns, 1943 perm->pool_ns, 1944 perm->pool_ns_len); 1945 if (ret < 0) 1946 p = &(*p)->rb_left; 1947 else if (ret > 0) 1948 p = &(*p)->rb_right; 1949 else { 1950 have = perm->perm; 1951 break; 1952 } 1953 } 1954 } 1955 up_read(&mdsc->pool_perm_rwsem); 1956 if (*p) 1957 goto out; 1958 1959 if (pool_ns) 1960 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n", 1961 pool, (int)pool_ns->len, pool_ns->str); 1962 else 1963 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool); 1964 1965 down_write(&mdsc->pool_perm_rwsem); 1966 p = &mdsc->pool_perm_tree.rb_node; 1967 parent = NULL; 1968 while (*p) { 1969 parent = *p; 1970 perm = rb_entry(parent, struct ceph_pool_perm, node); 1971 if (pool < perm->pool) 1972 p = &(*p)->rb_left; 1973 else if (pool > perm->pool) 1974 p = &(*p)->rb_right; 1975 else { 1976 int ret = ceph_compare_string(pool_ns, 1977 perm->pool_ns, 1978 perm->pool_ns_len); 1979 if (ret < 0) 1980 p = &(*p)->rb_left; 1981 else if (ret > 0) 1982 p = &(*p)->rb_right; 1983 else { 1984 have = perm->perm; 1985 break; 1986 } 1987 } 1988 } 1989 if (*p) { 1990 up_write(&mdsc->pool_perm_rwsem); 1991 goto out; 1992 } 1993 1994 rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL, 1995 1, false, GFP_NOFS); 1996 if (!rd_req) { 1997 err = -ENOMEM; 1998 goto out_unlock; 1999 } 2000 2001 rd_req->r_flags = CEPH_OSD_FLAG_READ; 2002 osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0); 2003 rd_req->r_base_oloc.pool = pool; 2004 if (pool_ns) 2005 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns); 2006 ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino); 2007 2008 err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS); 2009 if (err) 2010 goto out_unlock; 2011 2012 wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL, 2013 1, false, GFP_NOFS); 2014 if (!wr_req) { 2015 err = -ENOMEM; 2016 goto out_unlock; 2017 } 2018 2019 wr_req->r_flags = CEPH_OSD_FLAG_WRITE; 2020 osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL); 2021 ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc); 2022 ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid); 2023 2024 err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS); 2025 if (err) 2026 goto out_unlock; 2027 2028 /* one page should be large enough for STAT data */ 2029 pages = ceph_alloc_page_vector(1, GFP_KERNEL); 2030 if (IS_ERR(pages)) { 2031 err = PTR_ERR(pages); 2032 goto out_unlock; 2033 } 2034 2035 osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE, 2036 0, false, true); 2037 err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false); 2038 2039 wr_req->r_mtime = ci->vfs_inode.i_mtime; 2040 err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false); 2041 2042 if (!err) 2043 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req); 2044 if (!err2) 2045 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req); 2046 2047 if (err >= 0 || err == -ENOENT) 2048 have |= POOL_READ; 2049 else if (err != -EPERM) { 2050 if (err == -EBLOCKLISTED) 2051 fsc->blocklisted = true; 2052 goto out_unlock; 2053 } 2054 2055 if (err2 == 0 || err2 == -EEXIST) 2056 have |= POOL_WRITE; 2057 else if (err2 != -EPERM) { 2058 if (err2 == -EBLOCKLISTED) 2059 fsc->blocklisted = true; 2060 err = err2; 2061 goto out_unlock; 2062 } 2063 2064 pool_ns_len = pool_ns ? pool_ns->len : 0; 2065 perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS); 2066 if (!perm) { 2067 err = -ENOMEM; 2068 goto out_unlock; 2069 } 2070 2071 perm->pool = pool; 2072 perm->perm = have; 2073 perm->pool_ns_len = pool_ns_len; 2074 if (pool_ns_len > 0) 2075 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len); 2076 perm->pool_ns[pool_ns_len] = 0; 2077 2078 rb_link_node(&perm->node, parent, p); 2079 rb_insert_color(&perm->node, &mdsc->pool_perm_tree); 2080 err = 0; 2081 out_unlock: 2082 up_write(&mdsc->pool_perm_rwsem); 2083 2084 ceph_osdc_put_request(rd_req); 2085 ceph_osdc_put_request(wr_req); 2086 out: 2087 if (!err) 2088 err = have; 2089 if (pool_ns) 2090 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n", 2091 pool, (int)pool_ns->len, pool_ns->str, err); 2092 else 2093 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err); 2094 return err; 2095 } 2096 2097 int ceph_pool_perm_check(struct inode *inode, int need) 2098 { 2099 struct ceph_inode_info *ci = ceph_inode(inode); 2100 struct ceph_string *pool_ns; 2101 s64 pool; 2102 int ret, flags; 2103 2104 if (ci->i_vino.snap != CEPH_NOSNAP) { 2105 /* 2106 * Pool permission check needs to write to the first object. 2107 * But for snapshot, head of the first object may have alread 2108 * been deleted. Skip check to avoid creating orphan object. 2109 */ 2110 return 0; 2111 } 2112 2113 if (ceph_test_mount_opt(ceph_inode_to_client(inode), 2114 NOPOOLPERM)) 2115 return 0; 2116 2117 spin_lock(&ci->i_ceph_lock); 2118 flags = ci->i_ceph_flags; 2119 pool = ci->i_layout.pool_id; 2120 spin_unlock(&ci->i_ceph_lock); 2121 check: 2122 if (flags & CEPH_I_POOL_PERM) { 2123 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) { 2124 dout("ceph_pool_perm_check pool %lld no read perm\n", 2125 pool); 2126 return -EPERM; 2127 } 2128 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) { 2129 dout("ceph_pool_perm_check pool %lld no write perm\n", 2130 pool); 2131 return -EPERM; 2132 } 2133 return 0; 2134 } 2135 2136 pool_ns = ceph_try_get_string(ci->i_layout.pool_ns); 2137 ret = __ceph_pool_perm_get(ci, pool, pool_ns); 2138 ceph_put_string(pool_ns); 2139 if (ret < 0) 2140 return ret; 2141 2142 flags = CEPH_I_POOL_PERM; 2143 if (ret & POOL_READ) 2144 flags |= CEPH_I_POOL_RD; 2145 if (ret & POOL_WRITE) 2146 flags |= CEPH_I_POOL_WR; 2147 2148 spin_lock(&ci->i_ceph_lock); 2149 if (pool == ci->i_layout.pool_id && 2150 pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) { 2151 ci->i_ceph_flags |= flags; 2152 } else { 2153 pool = ci->i_layout.pool_id; 2154 flags = ci->i_ceph_flags; 2155 } 2156 spin_unlock(&ci->i_ceph_lock); 2157 goto check; 2158 } 2159 2160 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc) 2161 { 2162 struct ceph_pool_perm *perm; 2163 struct rb_node *n; 2164 2165 while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) { 2166 n = rb_first(&mdsc->pool_perm_tree); 2167 perm = rb_entry(n, struct ceph_pool_perm, node); 2168 rb_erase(n, &mdsc->pool_perm_tree); 2169 kfree(perm); 2170 } 2171 } 2172