1 /* -*- mode: c; c-basic-offset: 8; -*- 2 * vim: noexpandtab sw=8 ts=8 sts=0: 3 * 4 * Copyright (C) 2002, 2004 Oracle. All rights reserved. 5 * 6 * This program is free software; you can redistribute it and/or 7 * modify it under the terms of the GNU General Public 8 * License as published by the Free Software Foundation; either 9 * version 2 of the License, or (at your option) any later version. 10 * 11 * This program is distributed in the hope that it will be useful, 12 * but WITHOUT ANY WARRANTY; without even the implied warranty of 13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 14 * General Public License for more details. 15 * 16 * You should have received a copy of the GNU General Public 17 * License along with this program; if not, write to the 18 * Free Software Foundation, Inc., 59 Temple Place - Suite 330, 19 * Boston, MA 021110-1307, USA. 20 */ 21 22 #include <linux/fs.h> 23 #include <linux/slab.h> 24 #include <linux/highmem.h> 25 #include <linux/pagemap.h> 26 #include <asm/byteorder.h> 27 #include <linux/swap.h> 28 #include <linux/pipe_fs_i.h> 29 #include <linux/mpage.h> 30 #include <linux/quotaops.h> 31 #include <linux/blkdev.h> 32 #include <linux/uio.h> 33 34 #include <cluster/masklog.h> 35 36 #include "ocfs2.h" 37 38 #include "alloc.h" 39 #include "aops.h" 40 #include "dlmglue.h" 41 #include "extent_map.h" 42 #include "file.h" 43 #include "inode.h" 44 #include "journal.h" 45 #include "suballoc.h" 46 #include "super.h" 47 #include "symlink.h" 48 #include "refcounttree.h" 49 #include "ocfs2_trace.h" 50 51 #include "buffer_head_io.h" 52 #include "dir.h" 53 #include "namei.h" 54 #include "sysfile.h" 55 56 static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock, 57 struct buffer_head *bh_result, int create) 58 { 59 int err = -EIO; 60 int status; 61 struct ocfs2_dinode *fe = NULL; 62 struct buffer_head *bh = NULL; 63 struct buffer_head *buffer_cache_bh = NULL; 64 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 65 void *kaddr; 66 67 trace_ocfs2_symlink_get_block( 68 (unsigned long long)OCFS2_I(inode)->ip_blkno, 69 (unsigned long long)iblock, bh_result, create); 70 71 BUG_ON(ocfs2_inode_is_fast_symlink(inode)); 72 73 if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) { 74 mlog(ML_ERROR, "block offset > PATH_MAX: %llu", 75 (unsigned long long)iblock); 76 goto bail; 77 } 78 79 status = ocfs2_read_inode_block(inode, &bh); 80 if (status < 0) { 81 mlog_errno(status); 82 goto bail; 83 } 84 fe = (struct ocfs2_dinode *) bh->b_data; 85 86 if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb, 87 le32_to_cpu(fe->i_clusters))) { 88 err = -ENOMEM; 89 mlog(ML_ERROR, "block offset is outside the allocated size: " 90 "%llu\n", (unsigned long long)iblock); 91 goto bail; 92 } 93 94 /* We don't use the page cache to create symlink data, so if 95 * need be, copy it over from the buffer cache. */ 96 if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) { 97 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + 98 iblock; 99 buffer_cache_bh = sb_getblk(osb->sb, blkno); 100 if (!buffer_cache_bh) { 101 err = -ENOMEM; 102 mlog(ML_ERROR, "couldn't getblock for symlink!\n"); 103 goto bail; 104 } 105 106 /* we haven't locked out transactions, so a commit 107 * could've happened. Since we've got a reference on 108 * the bh, even if it commits while we're doing the 109 * copy, the data is still good. */ 110 if (buffer_jbd(buffer_cache_bh) 111 && ocfs2_inode_is_new(inode)) { 112 kaddr = kmap_atomic(bh_result->b_page); 113 if (!kaddr) { 114 mlog(ML_ERROR, "couldn't kmap!\n"); 115 goto bail; 116 } 117 memcpy(kaddr + (bh_result->b_size * iblock), 118 buffer_cache_bh->b_data, 119 bh_result->b_size); 120 kunmap_atomic(kaddr); 121 set_buffer_uptodate(bh_result); 122 } 123 brelse(buffer_cache_bh); 124 } 125 126 map_bh(bh_result, inode->i_sb, 127 le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock); 128 129 err = 0; 130 131 bail: 132 brelse(bh); 133 134 return err; 135 } 136 137 int ocfs2_get_block(struct inode *inode, sector_t iblock, 138 struct buffer_head *bh_result, int create) 139 { 140 int err = 0; 141 unsigned int ext_flags; 142 u64 max_blocks = bh_result->b_size >> inode->i_blkbits; 143 u64 p_blkno, count, past_eof; 144 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 145 146 trace_ocfs2_get_block((unsigned long long)OCFS2_I(inode)->ip_blkno, 147 (unsigned long long)iblock, bh_result, create); 148 149 if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE) 150 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n", 151 inode, inode->i_ino); 152 153 if (S_ISLNK(inode->i_mode)) { 154 /* this always does I/O for some reason. */ 155 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create); 156 goto bail; 157 } 158 159 err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, &count, 160 &ext_flags); 161 if (err) { 162 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, " 163 "%llu, NULL)\n", err, inode, (unsigned long long)iblock, 164 (unsigned long long)p_blkno); 165 goto bail; 166 } 167 168 if (max_blocks < count) 169 count = max_blocks; 170 171 /* 172 * ocfs2 never allocates in this function - the only time we 173 * need to use BH_New is when we're extending i_size on a file 174 * system which doesn't support holes, in which case BH_New 175 * allows __block_write_begin() to zero. 176 * 177 * If we see this on a sparse file system, then a truncate has 178 * raced us and removed the cluster. In this case, we clear 179 * the buffers dirty and uptodate bits and let the buffer code 180 * ignore it as a hole. 181 */ 182 if (create && p_blkno == 0 && ocfs2_sparse_alloc(osb)) { 183 clear_buffer_dirty(bh_result); 184 clear_buffer_uptodate(bh_result); 185 goto bail; 186 } 187 188 /* Treat the unwritten extent as a hole for zeroing purposes. */ 189 if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN)) 190 map_bh(bh_result, inode->i_sb, p_blkno); 191 192 bh_result->b_size = count << inode->i_blkbits; 193 194 if (!ocfs2_sparse_alloc(osb)) { 195 if (p_blkno == 0) { 196 err = -EIO; 197 mlog(ML_ERROR, 198 "iblock = %llu p_blkno = %llu blkno=(%llu)\n", 199 (unsigned long long)iblock, 200 (unsigned long long)p_blkno, 201 (unsigned long long)OCFS2_I(inode)->ip_blkno); 202 mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters); 203 dump_stack(); 204 goto bail; 205 } 206 } 207 208 past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode)); 209 210 trace_ocfs2_get_block_end((unsigned long long)OCFS2_I(inode)->ip_blkno, 211 (unsigned long long)past_eof); 212 if (create && (iblock >= past_eof)) 213 set_buffer_new(bh_result); 214 215 bail: 216 if (err < 0) 217 err = -EIO; 218 219 return err; 220 } 221 222 int ocfs2_read_inline_data(struct inode *inode, struct page *page, 223 struct buffer_head *di_bh) 224 { 225 void *kaddr; 226 loff_t size; 227 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 228 229 if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) { 230 ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag", 231 (unsigned long long)OCFS2_I(inode)->ip_blkno); 232 return -EROFS; 233 } 234 235 size = i_size_read(inode); 236 237 if (size > PAGE_CACHE_SIZE || 238 size > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) { 239 ocfs2_error(inode->i_sb, 240 "Inode %llu has with inline data has bad size: %Lu", 241 (unsigned long long)OCFS2_I(inode)->ip_blkno, 242 (unsigned long long)size); 243 return -EROFS; 244 } 245 246 kaddr = kmap_atomic(page); 247 if (size) 248 memcpy(kaddr, di->id2.i_data.id_data, size); 249 /* Clear the remaining part of the page */ 250 memset(kaddr + size, 0, PAGE_CACHE_SIZE - size); 251 flush_dcache_page(page); 252 kunmap_atomic(kaddr); 253 254 SetPageUptodate(page); 255 256 return 0; 257 } 258 259 static int ocfs2_readpage_inline(struct inode *inode, struct page *page) 260 { 261 int ret; 262 struct buffer_head *di_bh = NULL; 263 264 BUG_ON(!PageLocked(page)); 265 BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)); 266 267 ret = ocfs2_read_inode_block(inode, &di_bh); 268 if (ret) { 269 mlog_errno(ret); 270 goto out; 271 } 272 273 ret = ocfs2_read_inline_data(inode, page, di_bh); 274 out: 275 unlock_page(page); 276 277 brelse(di_bh); 278 return ret; 279 } 280 281 static int ocfs2_readpage(struct file *file, struct page *page) 282 { 283 struct inode *inode = page->mapping->host; 284 struct ocfs2_inode_info *oi = OCFS2_I(inode); 285 loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT; 286 int ret, unlock = 1; 287 288 trace_ocfs2_readpage((unsigned long long)oi->ip_blkno, 289 (page ? page->index : 0)); 290 291 ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page); 292 if (ret != 0) { 293 if (ret == AOP_TRUNCATED_PAGE) 294 unlock = 0; 295 mlog_errno(ret); 296 goto out; 297 } 298 299 if (down_read_trylock(&oi->ip_alloc_sem) == 0) { 300 /* 301 * Unlock the page and cycle ip_alloc_sem so that we don't 302 * busyloop waiting for ip_alloc_sem to unlock 303 */ 304 ret = AOP_TRUNCATED_PAGE; 305 unlock_page(page); 306 unlock = 0; 307 down_read(&oi->ip_alloc_sem); 308 up_read(&oi->ip_alloc_sem); 309 goto out_inode_unlock; 310 } 311 312 /* 313 * i_size might have just been updated as we grabed the meta lock. We 314 * might now be discovering a truncate that hit on another node. 315 * block_read_full_page->get_block freaks out if it is asked to read 316 * beyond the end of a file, so we check here. Callers 317 * (generic_file_read, vm_ops->fault) are clever enough to check i_size 318 * and notice that the page they just read isn't needed. 319 * 320 * XXX sys_readahead() seems to get that wrong? 321 */ 322 if (start >= i_size_read(inode)) { 323 zero_user(page, 0, PAGE_SIZE); 324 SetPageUptodate(page); 325 ret = 0; 326 goto out_alloc; 327 } 328 329 if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) 330 ret = ocfs2_readpage_inline(inode, page); 331 else 332 ret = block_read_full_page(page, ocfs2_get_block); 333 unlock = 0; 334 335 out_alloc: 336 up_read(&OCFS2_I(inode)->ip_alloc_sem); 337 out_inode_unlock: 338 ocfs2_inode_unlock(inode, 0); 339 out: 340 if (unlock) 341 unlock_page(page); 342 return ret; 343 } 344 345 /* 346 * This is used only for read-ahead. Failures or difficult to handle 347 * situations are safe to ignore. 348 * 349 * Right now, we don't bother with BH_Boundary - in-inode extent lists 350 * are quite large (243 extents on 4k blocks), so most inodes don't 351 * grow out to a tree. If need be, detecting boundary extents could 352 * trivially be added in a future version of ocfs2_get_block(). 353 */ 354 static int ocfs2_readpages(struct file *filp, struct address_space *mapping, 355 struct list_head *pages, unsigned nr_pages) 356 { 357 int ret, err = -EIO; 358 struct inode *inode = mapping->host; 359 struct ocfs2_inode_info *oi = OCFS2_I(inode); 360 loff_t start; 361 struct page *last; 362 363 /* 364 * Use the nonblocking flag for the dlm code to avoid page 365 * lock inversion, but don't bother with retrying. 366 */ 367 ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK); 368 if (ret) 369 return err; 370 371 if (down_read_trylock(&oi->ip_alloc_sem) == 0) { 372 ocfs2_inode_unlock(inode, 0); 373 return err; 374 } 375 376 /* 377 * Don't bother with inline-data. There isn't anything 378 * to read-ahead in that case anyway... 379 */ 380 if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) 381 goto out_unlock; 382 383 /* 384 * Check whether a remote node truncated this file - we just 385 * drop out in that case as it's not worth handling here. 386 */ 387 last = list_entry(pages->prev, struct page, lru); 388 start = (loff_t)last->index << PAGE_CACHE_SHIFT; 389 if (start >= i_size_read(inode)) 390 goto out_unlock; 391 392 err = mpage_readpages(mapping, pages, nr_pages, ocfs2_get_block); 393 394 out_unlock: 395 up_read(&oi->ip_alloc_sem); 396 ocfs2_inode_unlock(inode, 0); 397 398 return err; 399 } 400 401 /* Note: Because we don't support holes, our allocation has 402 * already happened (allocation writes zeros to the file data) 403 * so we don't have to worry about ordered writes in 404 * ocfs2_writepage. 405 * 406 * ->writepage is called during the process of invalidating the page cache 407 * during blocked lock processing. It can't block on any cluster locks 408 * to during block mapping. It's relying on the fact that the block 409 * mapping can't have disappeared under the dirty pages that it is 410 * being asked to write back. 411 */ 412 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc) 413 { 414 trace_ocfs2_writepage( 415 (unsigned long long)OCFS2_I(page->mapping->host)->ip_blkno, 416 page->index); 417 418 return block_write_full_page(page, ocfs2_get_block, wbc); 419 } 420 421 /* Taken from ext3. We don't necessarily need the full blown 422 * functionality yet, but IMHO it's better to cut and paste the whole 423 * thing so we can avoid introducing our own bugs (and easily pick up 424 * their fixes when they happen) --Mark */ 425 int walk_page_buffers( handle_t *handle, 426 struct buffer_head *head, 427 unsigned from, 428 unsigned to, 429 int *partial, 430 int (*fn)( handle_t *handle, 431 struct buffer_head *bh)) 432 { 433 struct buffer_head *bh; 434 unsigned block_start, block_end; 435 unsigned blocksize = head->b_size; 436 int err, ret = 0; 437 struct buffer_head *next; 438 439 for ( bh = head, block_start = 0; 440 ret == 0 && (bh != head || !block_start); 441 block_start = block_end, bh = next) 442 { 443 next = bh->b_this_page; 444 block_end = block_start + blocksize; 445 if (block_end <= from || block_start >= to) { 446 if (partial && !buffer_uptodate(bh)) 447 *partial = 1; 448 continue; 449 } 450 err = (*fn)(handle, bh); 451 if (!ret) 452 ret = err; 453 } 454 return ret; 455 } 456 457 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block) 458 { 459 sector_t status; 460 u64 p_blkno = 0; 461 int err = 0; 462 struct inode *inode = mapping->host; 463 464 trace_ocfs2_bmap((unsigned long long)OCFS2_I(inode)->ip_blkno, 465 (unsigned long long)block); 466 467 /* We don't need to lock journal system files, since they aren't 468 * accessed concurrently from multiple nodes. 469 */ 470 if (!INODE_JOURNAL(inode)) { 471 err = ocfs2_inode_lock(inode, NULL, 0); 472 if (err) { 473 if (err != -ENOENT) 474 mlog_errno(err); 475 goto bail; 476 } 477 down_read(&OCFS2_I(inode)->ip_alloc_sem); 478 } 479 480 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)) 481 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL, 482 NULL); 483 484 if (!INODE_JOURNAL(inode)) { 485 up_read(&OCFS2_I(inode)->ip_alloc_sem); 486 ocfs2_inode_unlock(inode, 0); 487 } 488 489 if (err) { 490 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n", 491 (unsigned long long)block); 492 mlog_errno(err); 493 goto bail; 494 } 495 496 bail: 497 status = err ? 0 : p_blkno; 498 499 return status; 500 } 501 502 /* 503 * TODO: Make this into a generic get_blocks function. 504 * 505 * From do_direct_io in direct-io.c: 506 * "So what we do is to permit the ->get_blocks function to populate 507 * bh.b_size with the size of IO which is permitted at this offset and 508 * this i_blkbits." 509 * 510 * This function is called directly from get_more_blocks in direct-io.c. 511 * 512 * called like this: dio->get_blocks(dio->inode, fs_startblk, 513 * fs_count, map_bh, dio->rw == WRITE); 514 */ 515 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock, 516 struct buffer_head *bh_result, int create) 517 { 518 int ret; 519 u32 cpos = 0; 520 int alloc_locked = 0; 521 u64 p_blkno, inode_blocks, contig_blocks; 522 unsigned int ext_flags; 523 unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits; 524 unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits; 525 unsigned long len = bh_result->b_size; 526 unsigned int clusters_to_alloc = 0, contig_clusters = 0; 527 528 cpos = ocfs2_blocks_to_clusters(inode->i_sb, iblock); 529 530 /* This function won't even be called if the request isn't all 531 * nicely aligned and of the right size, so there's no need 532 * for us to check any of that. */ 533 534 inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode)); 535 536 /* This figures out the size of the next contiguous block, and 537 * our logical offset */ 538 ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, 539 &contig_blocks, &ext_flags); 540 if (ret) { 541 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n", 542 (unsigned long long)iblock); 543 ret = -EIO; 544 goto bail; 545 } 546 547 /* We should already CoW the refcounted extent in case of create. */ 548 BUG_ON(create && (ext_flags & OCFS2_EXT_REFCOUNTED)); 549 550 /* allocate blocks if no p_blkno is found, and create == 1 */ 551 if (!p_blkno && create) { 552 ret = ocfs2_inode_lock(inode, NULL, 1); 553 if (ret < 0) { 554 mlog_errno(ret); 555 goto bail; 556 } 557 558 alloc_locked = 1; 559 560 /* fill hole, allocate blocks can't be larger than the size 561 * of the hole */ 562 clusters_to_alloc = ocfs2_clusters_for_bytes(inode->i_sb, len); 563 contig_clusters = ocfs2_clusters_for_blocks(inode->i_sb, 564 contig_blocks); 565 if (clusters_to_alloc > contig_clusters) 566 clusters_to_alloc = contig_clusters; 567 568 /* allocate extent and insert them into the extent tree */ 569 ret = ocfs2_extend_allocation(inode, cpos, 570 clusters_to_alloc, 0); 571 if (ret < 0) { 572 mlog_errno(ret); 573 goto bail; 574 } 575 576 ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, 577 &contig_blocks, &ext_flags); 578 if (ret < 0) { 579 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n", 580 (unsigned long long)iblock); 581 ret = -EIO; 582 goto bail; 583 } 584 } 585 586 /* 587 * get_more_blocks() expects us to describe a hole by clearing 588 * the mapped bit on bh_result(). 589 * 590 * Consider an unwritten extent as a hole. 591 */ 592 if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN)) 593 map_bh(bh_result, inode->i_sb, p_blkno); 594 else 595 clear_buffer_mapped(bh_result); 596 597 /* make sure we don't map more than max_blocks blocks here as 598 that's all the kernel will handle at this point. */ 599 if (max_blocks < contig_blocks) 600 contig_blocks = max_blocks; 601 bh_result->b_size = contig_blocks << blocksize_bits; 602 bail: 603 if (alloc_locked) 604 ocfs2_inode_unlock(inode, 1); 605 return ret; 606 } 607 608 /* 609 * ocfs2_dio_end_io is called by the dio core when a dio is finished. We're 610 * particularly interested in the aio/dio case. We use the rw_lock DLM lock 611 * to protect io on one node from truncation on another. 612 */ 613 static void ocfs2_dio_end_io(struct kiocb *iocb, 614 loff_t offset, 615 ssize_t bytes, 616 void *private) 617 { 618 struct inode *inode = file_inode(iocb->ki_filp); 619 int level; 620 621 /* this io's submitter should not have unlocked this before we could */ 622 BUG_ON(!ocfs2_iocb_is_rw_locked(iocb)); 623 624 if (ocfs2_iocb_is_unaligned_aio(iocb)) { 625 ocfs2_iocb_clear_unaligned_aio(iocb); 626 627 mutex_unlock(&OCFS2_I(inode)->ip_unaligned_aio); 628 } 629 630 ocfs2_iocb_clear_rw_locked(iocb); 631 632 level = ocfs2_iocb_rw_locked_level(iocb); 633 ocfs2_rw_unlock(inode, level); 634 } 635 636 static int ocfs2_releasepage(struct page *page, gfp_t wait) 637 { 638 if (!page_has_buffers(page)) 639 return 0; 640 return try_to_free_buffers(page); 641 } 642 643 static int ocfs2_is_overwrite(struct ocfs2_super *osb, 644 struct inode *inode, loff_t offset) 645 { 646 int ret = 0; 647 u32 v_cpos = 0; 648 u32 p_cpos = 0; 649 unsigned int num_clusters = 0; 650 unsigned int ext_flags = 0; 651 652 v_cpos = ocfs2_bytes_to_clusters(osb->sb, offset); 653 ret = ocfs2_get_clusters(inode, v_cpos, &p_cpos, 654 &num_clusters, &ext_flags); 655 if (ret < 0) { 656 mlog_errno(ret); 657 return ret; 658 } 659 660 if (p_cpos && !(ext_flags & OCFS2_EXT_UNWRITTEN)) 661 return 1; 662 663 return 0; 664 } 665 666 static int ocfs2_direct_IO_zero_extend(struct ocfs2_super *osb, 667 struct inode *inode, loff_t offset, 668 u64 zero_len, int cluster_align) 669 { 670 u32 p_cpos = 0; 671 u32 v_cpos = ocfs2_bytes_to_clusters(osb->sb, i_size_read(inode)); 672 unsigned int num_clusters = 0; 673 unsigned int ext_flags = 0; 674 int ret = 0; 675 676 if (offset <= i_size_read(inode) || cluster_align) 677 return 0; 678 679 ret = ocfs2_get_clusters(inode, v_cpos, &p_cpos, &num_clusters, 680 &ext_flags); 681 if (ret < 0) { 682 mlog_errno(ret); 683 return ret; 684 } 685 686 if (p_cpos && !(ext_flags & OCFS2_EXT_UNWRITTEN)) { 687 u64 s = i_size_read(inode); 688 sector_t sector = ((u64)p_cpos << (osb->s_clustersize_bits - 9)) + 689 (do_div(s, osb->s_clustersize) >> 9); 690 691 ret = blkdev_issue_zeroout(osb->sb->s_bdev, sector, 692 zero_len >> 9, GFP_NOFS, false); 693 if (ret < 0) 694 mlog_errno(ret); 695 } 696 697 return ret; 698 } 699 700 static int ocfs2_direct_IO_extend_no_holes(struct ocfs2_super *osb, 701 struct inode *inode, loff_t offset) 702 { 703 u64 zero_start, zero_len, total_zero_len; 704 u32 p_cpos = 0, clusters_to_add; 705 u32 v_cpos = ocfs2_bytes_to_clusters(osb->sb, i_size_read(inode)); 706 unsigned int num_clusters = 0; 707 unsigned int ext_flags = 0; 708 u32 size_div, offset_div; 709 int ret = 0; 710 711 { 712 u64 o = offset; 713 u64 s = i_size_read(inode); 714 715 offset_div = do_div(o, osb->s_clustersize); 716 size_div = do_div(s, osb->s_clustersize); 717 } 718 719 if (offset <= i_size_read(inode)) 720 return 0; 721 722 clusters_to_add = ocfs2_bytes_to_clusters(inode->i_sb, offset) - 723 ocfs2_bytes_to_clusters(inode->i_sb, i_size_read(inode)); 724 total_zero_len = offset - i_size_read(inode); 725 if (clusters_to_add) 726 total_zero_len -= offset_div; 727 728 /* Allocate clusters to fill out holes, and this is only needed 729 * when we add more than one clusters. Otherwise the cluster will 730 * be allocated during direct IO */ 731 if (clusters_to_add > 1) { 732 ret = ocfs2_extend_allocation(inode, 733 OCFS2_I(inode)->ip_clusters, 734 clusters_to_add - 1, 0); 735 if (ret) { 736 mlog_errno(ret); 737 goto out; 738 } 739 } 740 741 while (total_zero_len) { 742 ret = ocfs2_get_clusters(inode, v_cpos, &p_cpos, &num_clusters, 743 &ext_flags); 744 if (ret < 0) { 745 mlog_errno(ret); 746 goto out; 747 } 748 749 zero_start = ocfs2_clusters_to_bytes(osb->sb, p_cpos) + 750 size_div; 751 zero_len = ocfs2_clusters_to_bytes(osb->sb, num_clusters) - 752 size_div; 753 zero_len = min(total_zero_len, zero_len); 754 755 if (p_cpos && !(ext_flags & OCFS2_EXT_UNWRITTEN)) { 756 ret = blkdev_issue_zeroout(osb->sb->s_bdev, 757 zero_start >> 9, zero_len >> 9, 758 GFP_NOFS, false); 759 if (ret < 0) { 760 mlog_errno(ret); 761 goto out; 762 } 763 } 764 765 total_zero_len -= zero_len; 766 v_cpos += ocfs2_bytes_to_clusters(osb->sb, zero_len + size_div); 767 768 /* Only at first iteration can be cluster not aligned. 769 * So set size_div to 0 for the rest */ 770 size_div = 0; 771 } 772 773 out: 774 return ret; 775 } 776 777 static ssize_t ocfs2_direct_IO_write(struct kiocb *iocb, 778 struct iov_iter *iter, 779 loff_t offset) 780 { 781 ssize_t ret = 0; 782 ssize_t written = 0; 783 bool orphaned = false; 784 int is_overwrite = 0; 785 struct file *file = iocb->ki_filp; 786 struct inode *inode = file_inode(file)->i_mapping->host; 787 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 788 struct buffer_head *di_bh = NULL; 789 size_t count = iter->count; 790 journal_t *journal = osb->journal->j_journal; 791 u64 zero_len_head, zero_len_tail; 792 int cluster_align_head, cluster_align_tail; 793 loff_t final_size = offset + count; 794 int append_write = offset >= i_size_read(inode) ? 1 : 0; 795 unsigned int num_clusters = 0; 796 unsigned int ext_flags = 0; 797 798 { 799 u64 o = offset; 800 u64 s = i_size_read(inode); 801 802 zero_len_head = do_div(o, 1 << osb->s_clustersize_bits); 803 cluster_align_head = !zero_len_head; 804 805 zero_len_tail = osb->s_clustersize - 806 do_div(s, osb->s_clustersize); 807 if ((offset - i_size_read(inode)) < zero_len_tail) 808 zero_len_tail = offset - i_size_read(inode); 809 cluster_align_tail = !zero_len_tail; 810 } 811 812 /* 813 * when final_size > inode->i_size, inode->i_size will be 814 * updated after direct write, so add the inode to orphan 815 * dir first. 816 */ 817 if (final_size > i_size_read(inode)) { 818 ret = ocfs2_add_inode_to_orphan(osb, inode); 819 if (ret < 0) { 820 mlog_errno(ret); 821 goto out; 822 } 823 orphaned = true; 824 } 825 826 if (append_write) { 827 ret = ocfs2_inode_lock(inode, NULL, 1); 828 if (ret < 0) { 829 mlog_errno(ret); 830 goto clean_orphan; 831 } 832 833 /* zeroing out the previously allocated cluster tail 834 * that but not zeroed */ 835 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) 836 ret = ocfs2_direct_IO_zero_extend(osb, inode, offset, 837 zero_len_tail, cluster_align_tail); 838 else 839 ret = ocfs2_direct_IO_extend_no_holes(osb, inode, 840 offset); 841 if (ret < 0) { 842 mlog_errno(ret); 843 ocfs2_inode_unlock(inode, 1); 844 goto clean_orphan; 845 } 846 847 is_overwrite = ocfs2_is_overwrite(osb, inode, offset); 848 if (is_overwrite < 0) { 849 mlog_errno(is_overwrite); 850 ocfs2_inode_unlock(inode, 1); 851 goto clean_orphan; 852 } 853 854 ocfs2_inode_unlock(inode, 1); 855 } 856 857 written = __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev, iter, 858 offset, ocfs2_direct_IO_get_blocks, 859 ocfs2_dio_end_io, NULL, 0); 860 if (unlikely(written < 0)) { 861 loff_t i_size = i_size_read(inode); 862 863 if (offset + count > i_size) { 864 ret = ocfs2_inode_lock(inode, &di_bh, 1); 865 if (ret < 0) { 866 mlog_errno(ret); 867 goto clean_orphan; 868 } 869 870 if (i_size == i_size_read(inode)) { 871 ret = ocfs2_truncate_file(inode, di_bh, 872 i_size); 873 if (ret < 0) { 874 if (ret != -ENOSPC) 875 mlog_errno(ret); 876 877 ocfs2_inode_unlock(inode, 1); 878 brelse(di_bh); 879 goto clean_orphan; 880 } 881 } 882 883 ocfs2_inode_unlock(inode, 1); 884 brelse(di_bh); 885 886 ret = jbd2_journal_force_commit(journal); 887 if (ret < 0) 888 mlog_errno(ret); 889 } 890 } else if (written > 0 && append_write && !is_overwrite && 891 !cluster_align_head) { 892 /* zeroing out the allocated cluster head */ 893 u32 p_cpos = 0; 894 u32 v_cpos = ocfs2_bytes_to_clusters(osb->sb, offset); 895 896 ret = ocfs2_inode_lock(inode, NULL, 0); 897 if (ret < 0) { 898 mlog_errno(ret); 899 goto clean_orphan; 900 } 901 902 ret = ocfs2_get_clusters(inode, v_cpos, &p_cpos, 903 &num_clusters, &ext_flags); 904 if (ret < 0) { 905 mlog_errno(ret); 906 ocfs2_inode_unlock(inode, 0); 907 goto clean_orphan; 908 } 909 910 BUG_ON(!p_cpos || (ext_flags & OCFS2_EXT_UNWRITTEN)); 911 912 ret = blkdev_issue_zeroout(osb->sb->s_bdev, 913 (u64)p_cpos << (osb->s_clustersize_bits - 9), 914 zero_len_head >> 9, GFP_NOFS, false); 915 if (ret < 0) 916 mlog_errno(ret); 917 918 ocfs2_inode_unlock(inode, 0); 919 } 920 921 clean_orphan: 922 if (orphaned) { 923 int tmp_ret; 924 int update_isize = written > 0 ? 1 : 0; 925 loff_t end = update_isize ? offset + written : 0; 926 927 tmp_ret = ocfs2_inode_lock(inode, &di_bh, 1); 928 if (tmp_ret < 0) { 929 ret = tmp_ret; 930 mlog_errno(ret); 931 goto out; 932 } 933 934 tmp_ret = ocfs2_del_inode_from_orphan(osb, inode, di_bh, 935 update_isize, end); 936 if (tmp_ret < 0) { 937 ret = tmp_ret; 938 mlog_errno(ret); 939 goto out; 940 } 941 942 ocfs2_inode_unlock(inode, 1); 943 944 tmp_ret = jbd2_journal_force_commit(journal); 945 if (tmp_ret < 0) { 946 ret = tmp_ret; 947 mlog_errno(tmp_ret); 948 } 949 } 950 951 out: 952 if (ret >= 0) 953 ret = written; 954 return ret; 955 } 956 957 static ssize_t ocfs2_direct_IO(struct kiocb *iocb, struct iov_iter *iter, 958 loff_t offset) 959 { 960 struct file *file = iocb->ki_filp; 961 struct inode *inode = file_inode(file)->i_mapping->host; 962 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 963 int full_coherency = !(osb->s_mount_opt & 964 OCFS2_MOUNT_COHERENCY_BUFFERED); 965 966 /* 967 * Fallback to buffered I/O if we see an inode without 968 * extents. 969 */ 970 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) 971 return 0; 972 973 /* Fallback to buffered I/O if we are appending and 974 * concurrent O_DIRECT writes are allowed. 975 */ 976 if (i_size_read(inode) <= offset && !full_coherency) 977 return 0; 978 979 if (iov_iter_rw(iter) == READ) 980 return __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev, 981 iter, offset, 982 ocfs2_direct_IO_get_blocks, 983 ocfs2_dio_end_io, NULL, 0); 984 else 985 return ocfs2_direct_IO_write(iocb, iter, offset); 986 } 987 988 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb, 989 u32 cpos, 990 unsigned int *start, 991 unsigned int *end) 992 { 993 unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE; 994 995 if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) { 996 unsigned int cpp; 997 998 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits); 999 1000 cluster_start = cpos % cpp; 1001 cluster_start = cluster_start << osb->s_clustersize_bits; 1002 1003 cluster_end = cluster_start + osb->s_clustersize; 1004 } 1005 1006 BUG_ON(cluster_start > PAGE_SIZE); 1007 BUG_ON(cluster_end > PAGE_SIZE); 1008 1009 if (start) 1010 *start = cluster_start; 1011 if (end) 1012 *end = cluster_end; 1013 } 1014 1015 /* 1016 * 'from' and 'to' are the region in the page to avoid zeroing. 1017 * 1018 * If pagesize > clustersize, this function will avoid zeroing outside 1019 * of the cluster boundary. 1020 * 1021 * from == to == 0 is code for "zero the entire cluster region" 1022 */ 1023 static void ocfs2_clear_page_regions(struct page *page, 1024 struct ocfs2_super *osb, u32 cpos, 1025 unsigned from, unsigned to) 1026 { 1027 void *kaddr; 1028 unsigned int cluster_start, cluster_end; 1029 1030 ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end); 1031 1032 kaddr = kmap_atomic(page); 1033 1034 if (from || to) { 1035 if (from > cluster_start) 1036 memset(kaddr + cluster_start, 0, from - cluster_start); 1037 if (to < cluster_end) 1038 memset(kaddr + to, 0, cluster_end - to); 1039 } else { 1040 memset(kaddr + cluster_start, 0, cluster_end - cluster_start); 1041 } 1042 1043 kunmap_atomic(kaddr); 1044 } 1045 1046 /* 1047 * Nonsparse file systems fully allocate before we get to the write 1048 * code. This prevents ocfs2_write() from tagging the write as an 1049 * allocating one, which means ocfs2_map_page_blocks() might try to 1050 * read-in the blocks at the tail of our file. Avoid reading them by 1051 * testing i_size against each block offset. 1052 */ 1053 static int ocfs2_should_read_blk(struct inode *inode, struct page *page, 1054 unsigned int block_start) 1055 { 1056 u64 offset = page_offset(page) + block_start; 1057 1058 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) 1059 return 1; 1060 1061 if (i_size_read(inode) > offset) 1062 return 1; 1063 1064 return 0; 1065 } 1066 1067 /* 1068 * Some of this taken from __block_write_begin(). We already have our 1069 * mapping by now though, and the entire write will be allocating or 1070 * it won't, so not much need to use BH_New. 1071 * 1072 * This will also skip zeroing, which is handled externally. 1073 */ 1074 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno, 1075 struct inode *inode, unsigned int from, 1076 unsigned int to, int new) 1077 { 1078 int ret = 0; 1079 struct buffer_head *head, *bh, *wait[2], **wait_bh = wait; 1080 unsigned int block_end, block_start; 1081 unsigned int bsize = 1 << inode->i_blkbits; 1082 1083 if (!page_has_buffers(page)) 1084 create_empty_buffers(page, bsize, 0); 1085 1086 head = page_buffers(page); 1087 for (bh = head, block_start = 0; bh != head || !block_start; 1088 bh = bh->b_this_page, block_start += bsize) { 1089 block_end = block_start + bsize; 1090 1091 clear_buffer_new(bh); 1092 1093 /* 1094 * Ignore blocks outside of our i/o range - 1095 * they may belong to unallocated clusters. 1096 */ 1097 if (block_start >= to || block_end <= from) { 1098 if (PageUptodate(page)) 1099 set_buffer_uptodate(bh); 1100 continue; 1101 } 1102 1103 /* 1104 * For an allocating write with cluster size >= page 1105 * size, we always write the entire page. 1106 */ 1107 if (new) 1108 set_buffer_new(bh); 1109 1110 if (!buffer_mapped(bh)) { 1111 map_bh(bh, inode->i_sb, *p_blkno); 1112 unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr); 1113 } 1114 1115 if (PageUptodate(page)) { 1116 if (!buffer_uptodate(bh)) 1117 set_buffer_uptodate(bh); 1118 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) && 1119 !buffer_new(bh) && 1120 ocfs2_should_read_blk(inode, page, block_start) && 1121 (block_start < from || block_end > to)) { 1122 ll_rw_block(READ, 1, &bh); 1123 *wait_bh++=bh; 1124 } 1125 1126 *p_blkno = *p_blkno + 1; 1127 } 1128 1129 /* 1130 * If we issued read requests - let them complete. 1131 */ 1132 while(wait_bh > wait) { 1133 wait_on_buffer(*--wait_bh); 1134 if (!buffer_uptodate(*wait_bh)) 1135 ret = -EIO; 1136 } 1137 1138 if (ret == 0 || !new) 1139 return ret; 1140 1141 /* 1142 * If we get -EIO above, zero out any newly allocated blocks 1143 * to avoid exposing stale data. 1144 */ 1145 bh = head; 1146 block_start = 0; 1147 do { 1148 block_end = block_start + bsize; 1149 if (block_end <= from) 1150 goto next_bh; 1151 if (block_start >= to) 1152 break; 1153 1154 zero_user(page, block_start, bh->b_size); 1155 set_buffer_uptodate(bh); 1156 mark_buffer_dirty(bh); 1157 1158 next_bh: 1159 block_start = block_end; 1160 bh = bh->b_this_page; 1161 } while (bh != head); 1162 1163 return ret; 1164 } 1165 1166 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE) 1167 #define OCFS2_MAX_CTXT_PAGES 1 1168 #else 1169 #define OCFS2_MAX_CTXT_PAGES (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE) 1170 #endif 1171 1172 #define OCFS2_MAX_CLUSTERS_PER_PAGE (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE) 1173 1174 /* 1175 * Describe the state of a single cluster to be written to. 1176 */ 1177 struct ocfs2_write_cluster_desc { 1178 u32 c_cpos; 1179 u32 c_phys; 1180 /* 1181 * Give this a unique field because c_phys eventually gets 1182 * filled. 1183 */ 1184 unsigned c_new; 1185 unsigned c_unwritten; 1186 unsigned c_needs_zero; 1187 }; 1188 1189 struct ocfs2_write_ctxt { 1190 /* Logical cluster position / len of write */ 1191 u32 w_cpos; 1192 u32 w_clen; 1193 1194 /* First cluster allocated in a nonsparse extend */ 1195 u32 w_first_new_cpos; 1196 1197 struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE]; 1198 1199 /* 1200 * This is true if page_size > cluster_size. 1201 * 1202 * It triggers a set of special cases during write which might 1203 * have to deal with allocating writes to partial pages. 1204 */ 1205 unsigned int w_large_pages; 1206 1207 /* 1208 * Pages involved in this write. 1209 * 1210 * w_target_page is the page being written to by the user. 1211 * 1212 * w_pages is an array of pages which always contains 1213 * w_target_page, and in the case of an allocating write with 1214 * page_size < cluster size, it will contain zero'd and mapped 1215 * pages adjacent to w_target_page which need to be written 1216 * out in so that future reads from that region will get 1217 * zero's. 1218 */ 1219 unsigned int w_num_pages; 1220 struct page *w_pages[OCFS2_MAX_CTXT_PAGES]; 1221 struct page *w_target_page; 1222 1223 /* 1224 * w_target_locked is used for page_mkwrite path indicating no unlocking 1225 * against w_target_page in ocfs2_write_end_nolock. 1226 */ 1227 unsigned int w_target_locked:1; 1228 1229 /* 1230 * ocfs2_write_end() uses this to know what the real range to 1231 * write in the target should be. 1232 */ 1233 unsigned int w_target_from; 1234 unsigned int w_target_to; 1235 1236 /* 1237 * We could use journal_current_handle() but this is cleaner, 1238 * IMHO -Mark 1239 */ 1240 handle_t *w_handle; 1241 1242 struct buffer_head *w_di_bh; 1243 1244 struct ocfs2_cached_dealloc_ctxt w_dealloc; 1245 }; 1246 1247 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages) 1248 { 1249 int i; 1250 1251 for(i = 0; i < num_pages; i++) { 1252 if (pages[i]) { 1253 unlock_page(pages[i]); 1254 mark_page_accessed(pages[i]); 1255 page_cache_release(pages[i]); 1256 } 1257 } 1258 } 1259 1260 static void ocfs2_unlock_pages(struct ocfs2_write_ctxt *wc) 1261 { 1262 int i; 1263 1264 /* 1265 * w_target_locked is only set to true in the page_mkwrite() case. 1266 * The intent is to allow us to lock the target page from write_begin() 1267 * to write_end(). The caller must hold a ref on w_target_page. 1268 */ 1269 if (wc->w_target_locked) { 1270 BUG_ON(!wc->w_target_page); 1271 for (i = 0; i < wc->w_num_pages; i++) { 1272 if (wc->w_target_page == wc->w_pages[i]) { 1273 wc->w_pages[i] = NULL; 1274 break; 1275 } 1276 } 1277 mark_page_accessed(wc->w_target_page); 1278 page_cache_release(wc->w_target_page); 1279 } 1280 ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages); 1281 } 1282 1283 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc) 1284 { 1285 ocfs2_unlock_pages(wc); 1286 brelse(wc->w_di_bh); 1287 kfree(wc); 1288 } 1289 1290 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp, 1291 struct ocfs2_super *osb, loff_t pos, 1292 unsigned len, struct buffer_head *di_bh) 1293 { 1294 u32 cend; 1295 struct ocfs2_write_ctxt *wc; 1296 1297 wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS); 1298 if (!wc) 1299 return -ENOMEM; 1300 1301 wc->w_cpos = pos >> osb->s_clustersize_bits; 1302 wc->w_first_new_cpos = UINT_MAX; 1303 cend = (pos + len - 1) >> osb->s_clustersize_bits; 1304 wc->w_clen = cend - wc->w_cpos + 1; 1305 get_bh(di_bh); 1306 wc->w_di_bh = di_bh; 1307 1308 if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) 1309 wc->w_large_pages = 1; 1310 else 1311 wc->w_large_pages = 0; 1312 1313 ocfs2_init_dealloc_ctxt(&wc->w_dealloc); 1314 1315 *wcp = wc; 1316 1317 return 0; 1318 } 1319 1320 /* 1321 * If a page has any new buffers, zero them out here, and mark them uptodate 1322 * and dirty so they'll be written out (in order to prevent uninitialised 1323 * block data from leaking). And clear the new bit. 1324 */ 1325 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to) 1326 { 1327 unsigned int block_start, block_end; 1328 struct buffer_head *head, *bh; 1329 1330 BUG_ON(!PageLocked(page)); 1331 if (!page_has_buffers(page)) 1332 return; 1333 1334 bh = head = page_buffers(page); 1335 block_start = 0; 1336 do { 1337 block_end = block_start + bh->b_size; 1338 1339 if (buffer_new(bh)) { 1340 if (block_end > from && block_start < to) { 1341 if (!PageUptodate(page)) { 1342 unsigned start, end; 1343 1344 start = max(from, block_start); 1345 end = min(to, block_end); 1346 1347 zero_user_segment(page, start, end); 1348 set_buffer_uptodate(bh); 1349 } 1350 1351 clear_buffer_new(bh); 1352 mark_buffer_dirty(bh); 1353 } 1354 } 1355 1356 block_start = block_end; 1357 bh = bh->b_this_page; 1358 } while (bh != head); 1359 } 1360 1361 /* 1362 * Only called when we have a failure during allocating write to write 1363 * zero's to the newly allocated region. 1364 */ 1365 static void ocfs2_write_failure(struct inode *inode, 1366 struct ocfs2_write_ctxt *wc, 1367 loff_t user_pos, unsigned user_len) 1368 { 1369 int i; 1370 unsigned from = user_pos & (PAGE_CACHE_SIZE - 1), 1371 to = user_pos + user_len; 1372 struct page *tmppage; 1373 1374 ocfs2_zero_new_buffers(wc->w_target_page, from, to); 1375 1376 for(i = 0; i < wc->w_num_pages; i++) { 1377 tmppage = wc->w_pages[i]; 1378 1379 if (page_has_buffers(tmppage)) { 1380 if (ocfs2_should_order_data(inode)) 1381 ocfs2_jbd2_file_inode(wc->w_handle, inode); 1382 1383 block_commit_write(tmppage, from, to); 1384 } 1385 } 1386 } 1387 1388 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno, 1389 struct ocfs2_write_ctxt *wc, 1390 struct page *page, u32 cpos, 1391 loff_t user_pos, unsigned user_len, 1392 int new) 1393 { 1394 int ret; 1395 unsigned int map_from = 0, map_to = 0; 1396 unsigned int cluster_start, cluster_end; 1397 unsigned int user_data_from = 0, user_data_to = 0; 1398 1399 ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos, 1400 &cluster_start, &cluster_end); 1401 1402 /* treat the write as new if the a hole/lseek spanned across 1403 * the page boundary. 1404 */ 1405 new = new | ((i_size_read(inode) <= page_offset(page)) && 1406 (page_offset(page) <= user_pos)); 1407 1408 if (page == wc->w_target_page) { 1409 map_from = user_pos & (PAGE_CACHE_SIZE - 1); 1410 map_to = map_from + user_len; 1411 1412 if (new) 1413 ret = ocfs2_map_page_blocks(page, p_blkno, inode, 1414 cluster_start, cluster_end, 1415 new); 1416 else 1417 ret = ocfs2_map_page_blocks(page, p_blkno, inode, 1418 map_from, map_to, new); 1419 if (ret) { 1420 mlog_errno(ret); 1421 goto out; 1422 } 1423 1424 user_data_from = map_from; 1425 user_data_to = map_to; 1426 if (new) { 1427 map_from = cluster_start; 1428 map_to = cluster_end; 1429 } 1430 } else { 1431 /* 1432 * If we haven't allocated the new page yet, we 1433 * shouldn't be writing it out without copying user 1434 * data. This is likely a math error from the caller. 1435 */ 1436 BUG_ON(!new); 1437 1438 map_from = cluster_start; 1439 map_to = cluster_end; 1440 1441 ret = ocfs2_map_page_blocks(page, p_blkno, inode, 1442 cluster_start, cluster_end, new); 1443 if (ret) { 1444 mlog_errno(ret); 1445 goto out; 1446 } 1447 } 1448 1449 /* 1450 * Parts of newly allocated pages need to be zero'd. 1451 * 1452 * Above, we have also rewritten 'to' and 'from' - as far as 1453 * the rest of the function is concerned, the entire cluster 1454 * range inside of a page needs to be written. 1455 * 1456 * We can skip this if the page is up to date - it's already 1457 * been zero'd from being read in as a hole. 1458 */ 1459 if (new && !PageUptodate(page)) 1460 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb), 1461 cpos, user_data_from, user_data_to); 1462 1463 flush_dcache_page(page); 1464 1465 out: 1466 return ret; 1467 } 1468 1469 /* 1470 * This function will only grab one clusters worth of pages. 1471 */ 1472 static int ocfs2_grab_pages_for_write(struct address_space *mapping, 1473 struct ocfs2_write_ctxt *wc, 1474 u32 cpos, loff_t user_pos, 1475 unsigned user_len, int new, 1476 struct page *mmap_page) 1477 { 1478 int ret = 0, i; 1479 unsigned long start, target_index, end_index, index; 1480 struct inode *inode = mapping->host; 1481 loff_t last_byte; 1482 1483 target_index = user_pos >> PAGE_CACHE_SHIFT; 1484 1485 /* 1486 * Figure out how many pages we'll be manipulating here. For 1487 * non allocating write, we just change the one 1488 * page. Otherwise, we'll need a whole clusters worth. If we're 1489 * writing past i_size, we only need enough pages to cover the 1490 * last page of the write. 1491 */ 1492 if (new) { 1493 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb); 1494 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos); 1495 /* 1496 * We need the index *past* the last page we could possibly 1497 * touch. This is the page past the end of the write or 1498 * i_size, whichever is greater. 1499 */ 1500 last_byte = max(user_pos + user_len, i_size_read(inode)); 1501 BUG_ON(last_byte < 1); 1502 end_index = ((last_byte - 1) >> PAGE_CACHE_SHIFT) + 1; 1503 if ((start + wc->w_num_pages) > end_index) 1504 wc->w_num_pages = end_index - start; 1505 } else { 1506 wc->w_num_pages = 1; 1507 start = target_index; 1508 } 1509 1510 for(i = 0; i < wc->w_num_pages; i++) { 1511 index = start + i; 1512 1513 if (index == target_index && mmap_page) { 1514 /* 1515 * ocfs2_pagemkwrite() is a little different 1516 * and wants us to directly use the page 1517 * passed in. 1518 */ 1519 lock_page(mmap_page); 1520 1521 /* Exit and let the caller retry */ 1522 if (mmap_page->mapping != mapping) { 1523 WARN_ON(mmap_page->mapping); 1524 unlock_page(mmap_page); 1525 ret = -EAGAIN; 1526 goto out; 1527 } 1528 1529 page_cache_get(mmap_page); 1530 wc->w_pages[i] = mmap_page; 1531 wc->w_target_locked = true; 1532 } else { 1533 wc->w_pages[i] = find_or_create_page(mapping, index, 1534 GFP_NOFS); 1535 if (!wc->w_pages[i]) { 1536 ret = -ENOMEM; 1537 mlog_errno(ret); 1538 goto out; 1539 } 1540 } 1541 wait_for_stable_page(wc->w_pages[i]); 1542 1543 if (index == target_index) 1544 wc->w_target_page = wc->w_pages[i]; 1545 } 1546 out: 1547 if (ret) 1548 wc->w_target_locked = false; 1549 return ret; 1550 } 1551 1552 /* 1553 * Prepare a single cluster for write one cluster into the file. 1554 */ 1555 static int ocfs2_write_cluster(struct address_space *mapping, 1556 u32 phys, unsigned int unwritten, 1557 unsigned int should_zero, 1558 struct ocfs2_alloc_context *data_ac, 1559 struct ocfs2_alloc_context *meta_ac, 1560 struct ocfs2_write_ctxt *wc, u32 cpos, 1561 loff_t user_pos, unsigned user_len) 1562 { 1563 int ret, i, new; 1564 u64 v_blkno, p_blkno; 1565 struct inode *inode = mapping->host; 1566 struct ocfs2_extent_tree et; 1567 1568 new = phys == 0 ? 1 : 0; 1569 if (new) { 1570 u32 tmp_pos; 1571 1572 /* 1573 * This is safe to call with the page locks - it won't take 1574 * any additional semaphores or cluster locks. 1575 */ 1576 tmp_pos = cpos; 1577 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode, 1578 &tmp_pos, 1, 0, wc->w_di_bh, 1579 wc->w_handle, data_ac, 1580 meta_ac, NULL); 1581 /* 1582 * This shouldn't happen because we must have already 1583 * calculated the correct meta data allocation required. The 1584 * internal tree allocation code should know how to increase 1585 * transaction credits itself. 1586 * 1587 * If need be, we could handle -EAGAIN for a 1588 * RESTART_TRANS here. 1589 */ 1590 mlog_bug_on_msg(ret == -EAGAIN, 1591 "Inode %llu: EAGAIN return during allocation.\n", 1592 (unsigned long long)OCFS2_I(inode)->ip_blkno); 1593 if (ret < 0) { 1594 mlog_errno(ret); 1595 goto out; 1596 } 1597 } else if (unwritten) { 1598 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), 1599 wc->w_di_bh); 1600 ret = ocfs2_mark_extent_written(inode, &et, 1601 wc->w_handle, cpos, 1, phys, 1602 meta_ac, &wc->w_dealloc); 1603 if (ret < 0) { 1604 mlog_errno(ret); 1605 goto out; 1606 } 1607 } 1608 1609 if (should_zero) 1610 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos); 1611 else 1612 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits; 1613 1614 /* 1615 * The only reason this should fail is due to an inability to 1616 * find the extent added. 1617 */ 1618 ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL, 1619 NULL); 1620 if (ret < 0) { 1621 mlog(ML_ERROR, "Get physical blkno failed for inode %llu, " 1622 "at logical block %llu", 1623 (unsigned long long)OCFS2_I(inode)->ip_blkno, 1624 (unsigned long long)v_blkno); 1625 goto out; 1626 } 1627 1628 BUG_ON(p_blkno == 0); 1629 1630 for(i = 0; i < wc->w_num_pages; i++) { 1631 int tmpret; 1632 1633 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc, 1634 wc->w_pages[i], cpos, 1635 user_pos, user_len, 1636 should_zero); 1637 if (tmpret) { 1638 mlog_errno(tmpret); 1639 if (ret == 0) 1640 ret = tmpret; 1641 } 1642 } 1643 1644 /* 1645 * We only have cleanup to do in case of allocating write. 1646 */ 1647 if (ret && new) 1648 ocfs2_write_failure(inode, wc, user_pos, user_len); 1649 1650 out: 1651 1652 return ret; 1653 } 1654 1655 static int ocfs2_write_cluster_by_desc(struct address_space *mapping, 1656 struct ocfs2_alloc_context *data_ac, 1657 struct ocfs2_alloc_context *meta_ac, 1658 struct ocfs2_write_ctxt *wc, 1659 loff_t pos, unsigned len) 1660 { 1661 int ret, i; 1662 loff_t cluster_off; 1663 unsigned int local_len = len; 1664 struct ocfs2_write_cluster_desc *desc; 1665 struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb); 1666 1667 for (i = 0; i < wc->w_clen; i++) { 1668 desc = &wc->w_desc[i]; 1669 1670 /* 1671 * We have to make sure that the total write passed in 1672 * doesn't extend past a single cluster. 1673 */ 1674 local_len = len; 1675 cluster_off = pos & (osb->s_clustersize - 1); 1676 if ((cluster_off + local_len) > osb->s_clustersize) 1677 local_len = osb->s_clustersize - cluster_off; 1678 1679 ret = ocfs2_write_cluster(mapping, desc->c_phys, 1680 desc->c_unwritten, 1681 desc->c_needs_zero, 1682 data_ac, meta_ac, 1683 wc, desc->c_cpos, pos, local_len); 1684 if (ret) { 1685 mlog_errno(ret); 1686 goto out; 1687 } 1688 1689 len -= local_len; 1690 pos += local_len; 1691 } 1692 1693 ret = 0; 1694 out: 1695 return ret; 1696 } 1697 1698 /* 1699 * ocfs2_write_end() wants to know which parts of the target page it 1700 * should complete the write on. It's easiest to compute them ahead of 1701 * time when a more complete view of the write is available. 1702 */ 1703 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb, 1704 struct ocfs2_write_ctxt *wc, 1705 loff_t pos, unsigned len, int alloc) 1706 { 1707 struct ocfs2_write_cluster_desc *desc; 1708 1709 wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1); 1710 wc->w_target_to = wc->w_target_from + len; 1711 1712 if (alloc == 0) 1713 return; 1714 1715 /* 1716 * Allocating write - we may have different boundaries based 1717 * on page size and cluster size. 1718 * 1719 * NOTE: We can no longer compute one value from the other as 1720 * the actual write length and user provided length may be 1721 * different. 1722 */ 1723 1724 if (wc->w_large_pages) { 1725 /* 1726 * We only care about the 1st and last cluster within 1727 * our range and whether they should be zero'd or not. Either 1728 * value may be extended out to the start/end of a 1729 * newly allocated cluster. 1730 */ 1731 desc = &wc->w_desc[0]; 1732 if (desc->c_needs_zero) 1733 ocfs2_figure_cluster_boundaries(osb, 1734 desc->c_cpos, 1735 &wc->w_target_from, 1736 NULL); 1737 1738 desc = &wc->w_desc[wc->w_clen - 1]; 1739 if (desc->c_needs_zero) 1740 ocfs2_figure_cluster_boundaries(osb, 1741 desc->c_cpos, 1742 NULL, 1743 &wc->w_target_to); 1744 } else { 1745 wc->w_target_from = 0; 1746 wc->w_target_to = PAGE_CACHE_SIZE; 1747 } 1748 } 1749 1750 /* 1751 * Populate each single-cluster write descriptor in the write context 1752 * with information about the i/o to be done. 1753 * 1754 * Returns the number of clusters that will have to be allocated, as 1755 * well as a worst case estimate of the number of extent records that 1756 * would have to be created during a write to an unwritten region. 1757 */ 1758 static int ocfs2_populate_write_desc(struct inode *inode, 1759 struct ocfs2_write_ctxt *wc, 1760 unsigned int *clusters_to_alloc, 1761 unsigned int *extents_to_split) 1762 { 1763 int ret; 1764 struct ocfs2_write_cluster_desc *desc; 1765 unsigned int num_clusters = 0; 1766 unsigned int ext_flags = 0; 1767 u32 phys = 0; 1768 int i; 1769 1770 *clusters_to_alloc = 0; 1771 *extents_to_split = 0; 1772 1773 for (i = 0; i < wc->w_clen; i++) { 1774 desc = &wc->w_desc[i]; 1775 desc->c_cpos = wc->w_cpos + i; 1776 1777 if (num_clusters == 0) { 1778 /* 1779 * Need to look up the next extent record. 1780 */ 1781 ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys, 1782 &num_clusters, &ext_flags); 1783 if (ret) { 1784 mlog_errno(ret); 1785 goto out; 1786 } 1787 1788 /* We should already CoW the refcountd extent. */ 1789 BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED); 1790 1791 /* 1792 * Assume worst case - that we're writing in 1793 * the middle of the extent. 1794 * 1795 * We can assume that the write proceeds from 1796 * left to right, in which case the extent 1797 * insert code is smart enough to coalesce the 1798 * next splits into the previous records created. 1799 */ 1800 if (ext_flags & OCFS2_EXT_UNWRITTEN) 1801 *extents_to_split = *extents_to_split + 2; 1802 } else if (phys) { 1803 /* 1804 * Only increment phys if it doesn't describe 1805 * a hole. 1806 */ 1807 phys++; 1808 } 1809 1810 /* 1811 * If w_first_new_cpos is < UINT_MAX, we have a non-sparse 1812 * file that got extended. w_first_new_cpos tells us 1813 * where the newly allocated clusters are so we can 1814 * zero them. 1815 */ 1816 if (desc->c_cpos >= wc->w_first_new_cpos) { 1817 BUG_ON(phys == 0); 1818 desc->c_needs_zero = 1; 1819 } 1820 1821 desc->c_phys = phys; 1822 if (phys == 0) { 1823 desc->c_new = 1; 1824 desc->c_needs_zero = 1; 1825 *clusters_to_alloc = *clusters_to_alloc + 1; 1826 } 1827 1828 if (ext_flags & OCFS2_EXT_UNWRITTEN) { 1829 desc->c_unwritten = 1; 1830 desc->c_needs_zero = 1; 1831 } 1832 1833 num_clusters--; 1834 } 1835 1836 ret = 0; 1837 out: 1838 return ret; 1839 } 1840 1841 static int ocfs2_write_begin_inline(struct address_space *mapping, 1842 struct inode *inode, 1843 struct ocfs2_write_ctxt *wc) 1844 { 1845 int ret; 1846 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 1847 struct page *page; 1848 handle_t *handle; 1849 struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data; 1850 1851 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS); 1852 if (IS_ERR(handle)) { 1853 ret = PTR_ERR(handle); 1854 mlog_errno(ret); 1855 goto out; 1856 } 1857 1858 page = find_or_create_page(mapping, 0, GFP_NOFS); 1859 if (!page) { 1860 ocfs2_commit_trans(osb, handle); 1861 ret = -ENOMEM; 1862 mlog_errno(ret); 1863 goto out; 1864 } 1865 /* 1866 * If we don't set w_num_pages then this page won't get unlocked 1867 * and freed on cleanup of the write context. 1868 */ 1869 wc->w_pages[0] = wc->w_target_page = page; 1870 wc->w_num_pages = 1; 1871 1872 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh, 1873 OCFS2_JOURNAL_ACCESS_WRITE); 1874 if (ret) { 1875 ocfs2_commit_trans(osb, handle); 1876 1877 mlog_errno(ret); 1878 goto out; 1879 } 1880 1881 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)) 1882 ocfs2_set_inode_data_inline(inode, di); 1883 1884 if (!PageUptodate(page)) { 1885 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh); 1886 if (ret) { 1887 ocfs2_commit_trans(osb, handle); 1888 1889 goto out; 1890 } 1891 } 1892 1893 wc->w_handle = handle; 1894 out: 1895 return ret; 1896 } 1897 1898 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size) 1899 { 1900 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 1901 1902 if (new_size <= le16_to_cpu(di->id2.i_data.id_count)) 1903 return 1; 1904 return 0; 1905 } 1906 1907 static int ocfs2_try_to_write_inline_data(struct address_space *mapping, 1908 struct inode *inode, loff_t pos, 1909 unsigned len, struct page *mmap_page, 1910 struct ocfs2_write_ctxt *wc) 1911 { 1912 int ret, written = 0; 1913 loff_t end = pos + len; 1914 struct ocfs2_inode_info *oi = OCFS2_I(inode); 1915 struct ocfs2_dinode *di = NULL; 1916 1917 trace_ocfs2_try_to_write_inline_data((unsigned long long)oi->ip_blkno, 1918 len, (unsigned long long)pos, 1919 oi->ip_dyn_features); 1920 1921 /* 1922 * Handle inodes which already have inline data 1st. 1923 */ 1924 if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) { 1925 if (mmap_page == NULL && 1926 ocfs2_size_fits_inline_data(wc->w_di_bh, end)) 1927 goto do_inline_write; 1928 1929 /* 1930 * The write won't fit - we have to give this inode an 1931 * inline extent list now. 1932 */ 1933 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh); 1934 if (ret) 1935 mlog_errno(ret); 1936 goto out; 1937 } 1938 1939 /* 1940 * Check whether the inode can accept inline data. 1941 */ 1942 if (oi->ip_clusters != 0 || i_size_read(inode) != 0) 1943 return 0; 1944 1945 /* 1946 * Check whether the write can fit. 1947 */ 1948 di = (struct ocfs2_dinode *)wc->w_di_bh->b_data; 1949 if (mmap_page || 1950 end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) 1951 return 0; 1952 1953 do_inline_write: 1954 ret = ocfs2_write_begin_inline(mapping, inode, wc); 1955 if (ret) { 1956 mlog_errno(ret); 1957 goto out; 1958 } 1959 1960 /* 1961 * This signals to the caller that the data can be written 1962 * inline. 1963 */ 1964 written = 1; 1965 out: 1966 return written ? written : ret; 1967 } 1968 1969 /* 1970 * This function only does anything for file systems which can't 1971 * handle sparse files. 1972 * 1973 * What we want to do here is fill in any hole between the current end 1974 * of allocation and the end of our write. That way the rest of the 1975 * write path can treat it as an non-allocating write, which has no 1976 * special case code for sparse/nonsparse files. 1977 */ 1978 static int ocfs2_expand_nonsparse_inode(struct inode *inode, 1979 struct buffer_head *di_bh, 1980 loff_t pos, unsigned len, 1981 struct ocfs2_write_ctxt *wc) 1982 { 1983 int ret; 1984 loff_t newsize = pos + len; 1985 1986 BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))); 1987 1988 if (newsize <= i_size_read(inode)) 1989 return 0; 1990 1991 ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos); 1992 if (ret) 1993 mlog_errno(ret); 1994 1995 wc->w_first_new_cpos = 1996 ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode)); 1997 1998 return ret; 1999 } 2000 2001 static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh, 2002 loff_t pos) 2003 { 2004 int ret = 0; 2005 2006 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))); 2007 if (pos > i_size_read(inode)) 2008 ret = ocfs2_zero_extend(inode, di_bh, pos); 2009 2010 return ret; 2011 } 2012 2013 /* 2014 * Try to flush truncate logs if we can free enough clusters from it. 2015 * As for return value, "< 0" means error, "0" no space and "1" means 2016 * we have freed enough spaces and let the caller try to allocate again. 2017 */ 2018 static int ocfs2_try_to_free_truncate_log(struct ocfs2_super *osb, 2019 unsigned int needed) 2020 { 2021 tid_t target; 2022 int ret = 0; 2023 unsigned int truncated_clusters; 2024 2025 mutex_lock(&osb->osb_tl_inode->i_mutex); 2026 truncated_clusters = osb->truncated_clusters; 2027 mutex_unlock(&osb->osb_tl_inode->i_mutex); 2028 2029 /* 2030 * Check whether we can succeed in allocating if we free 2031 * the truncate log. 2032 */ 2033 if (truncated_clusters < needed) 2034 goto out; 2035 2036 ret = ocfs2_flush_truncate_log(osb); 2037 if (ret) { 2038 mlog_errno(ret); 2039 goto out; 2040 } 2041 2042 if (jbd2_journal_start_commit(osb->journal->j_journal, &target)) { 2043 jbd2_log_wait_commit(osb->journal->j_journal, target); 2044 ret = 1; 2045 } 2046 out: 2047 return ret; 2048 } 2049 2050 int ocfs2_write_begin_nolock(struct file *filp, 2051 struct address_space *mapping, 2052 loff_t pos, unsigned len, unsigned flags, 2053 struct page **pagep, void **fsdata, 2054 struct buffer_head *di_bh, struct page *mmap_page) 2055 { 2056 int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS; 2057 unsigned int clusters_to_alloc, extents_to_split, clusters_need = 0; 2058 struct ocfs2_write_ctxt *wc; 2059 struct inode *inode = mapping->host; 2060 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 2061 struct ocfs2_dinode *di; 2062 struct ocfs2_alloc_context *data_ac = NULL; 2063 struct ocfs2_alloc_context *meta_ac = NULL; 2064 handle_t *handle; 2065 struct ocfs2_extent_tree et; 2066 int try_free = 1, ret1; 2067 2068 try_again: 2069 ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh); 2070 if (ret) { 2071 mlog_errno(ret); 2072 return ret; 2073 } 2074 2075 if (ocfs2_supports_inline_data(osb)) { 2076 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len, 2077 mmap_page, wc); 2078 if (ret == 1) { 2079 ret = 0; 2080 goto success; 2081 } 2082 if (ret < 0) { 2083 mlog_errno(ret); 2084 goto out; 2085 } 2086 } 2087 2088 if (ocfs2_sparse_alloc(osb)) 2089 ret = ocfs2_zero_tail(inode, di_bh, pos); 2090 else 2091 ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos, len, 2092 wc); 2093 if (ret) { 2094 mlog_errno(ret); 2095 goto out; 2096 } 2097 2098 ret = ocfs2_check_range_for_refcount(inode, pos, len); 2099 if (ret < 0) { 2100 mlog_errno(ret); 2101 goto out; 2102 } else if (ret == 1) { 2103 clusters_need = wc->w_clen; 2104 ret = ocfs2_refcount_cow(inode, di_bh, 2105 wc->w_cpos, wc->w_clen, UINT_MAX); 2106 if (ret) { 2107 mlog_errno(ret); 2108 goto out; 2109 } 2110 } 2111 2112 ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc, 2113 &extents_to_split); 2114 if (ret) { 2115 mlog_errno(ret); 2116 goto out; 2117 } 2118 clusters_need += clusters_to_alloc; 2119 2120 di = (struct ocfs2_dinode *)wc->w_di_bh->b_data; 2121 2122 trace_ocfs2_write_begin_nolock( 2123 (unsigned long long)OCFS2_I(inode)->ip_blkno, 2124 (long long)i_size_read(inode), 2125 le32_to_cpu(di->i_clusters), 2126 pos, len, flags, mmap_page, 2127 clusters_to_alloc, extents_to_split); 2128 2129 /* 2130 * We set w_target_from, w_target_to here so that 2131 * ocfs2_write_end() knows which range in the target page to 2132 * write out. An allocation requires that we write the entire 2133 * cluster range. 2134 */ 2135 if (clusters_to_alloc || extents_to_split) { 2136 /* 2137 * XXX: We are stretching the limits of 2138 * ocfs2_lock_allocators(). It greatly over-estimates 2139 * the work to be done. 2140 */ 2141 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), 2142 wc->w_di_bh); 2143 ret = ocfs2_lock_allocators(inode, &et, 2144 clusters_to_alloc, extents_to_split, 2145 &data_ac, &meta_ac); 2146 if (ret) { 2147 mlog_errno(ret); 2148 goto out; 2149 } 2150 2151 if (data_ac) 2152 data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv; 2153 2154 credits = ocfs2_calc_extend_credits(inode->i_sb, 2155 &di->id2.i_list); 2156 2157 } 2158 2159 /* 2160 * We have to zero sparse allocated clusters, unwritten extent clusters, 2161 * and non-sparse clusters we just extended. For non-sparse writes, 2162 * we know zeros will only be needed in the first and/or last cluster. 2163 */ 2164 if (clusters_to_alloc || extents_to_split || 2165 (wc->w_clen && (wc->w_desc[0].c_needs_zero || 2166 wc->w_desc[wc->w_clen - 1].c_needs_zero))) 2167 cluster_of_pages = 1; 2168 else 2169 cluster_of_pages = 0; 2170 2171 ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages); 2172 2173 handle = ocfs2_start_trans(osb, credits); 2174 if (IS_ERR(handle)) { 2175 ret = PTR_ERR(handle); 2176 mlog_errno(ret); 2177 goto out; 2178 } 2179 2180 wc->w_handle = handle; 2181 2182 if (clusters_to_alloc) { 2183 ret = dquot_alloc_space_nodirty(inode, 2184 ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc)); 2185 if (ret) 2186 goto out_commit; 2187 } 2188 /* 2189 * We don't want this to fail in ocfs2_write_end(), so do it 2190 * here. 2191 */ 2192 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh, 2193 OCFS2_JOURNAL_ACCESS_WRITE); 2194 if (ret) { 2195 mlog_errno(ret); 2196 goto out_quota; 2197 } 2198 2199 /* 2200 * Fill our page array first. That way we've grabbed enough so 2201 * that we can zero and flush if we error after adding the 2202 * extent. 2203 */ 2204 ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len, 2205 cluster_of_pages, mmap_page); 2206 if (ret && ret != -EAGAIN) { 2207 mlog_errno(ret); 2208 goto out_quota; 2209 } 2210 2211 /* 2212 * ocfs2_grab_pages_for_write() returns -EAGAIN if it could not lock 2213 * the target page. In this case, we exit with no error and no target 2214 * page. This will trigger the caller, page_mkwrite(), to re-try 2215 * the operation. 2216 */ 2217 if (ret == -EAGAIN) { 2218 BUG_ON(wc->w_target_page); 2219 ret = 0; 2220 goto out_quota; 2221 } 2222 2223 ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos, 2224 len); 2225 if (ret) { 2226 mlog_errno(ret); 2227 goto out_quota; 2228 } 2229 2230 if (data_ac) 2231 ocfs2_free_alloc_context(data_ac); 2232 if (meta_ac) 2233 ocfs2_free_alloc_context(meta_ac); 2234 2235 success: 2236 *pagep = wc->w_target_page; 2237 *fsdata = wc; 2238 return 0; 2239 out_quota: 2240 if (clusters_to_alloc) 2241 dquot_free_space(inode, 2242 ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc)); 2243 out_commit: 2244 ocfs2_commit_trans(osb, handle); 2245 2246 out: 2247 ocfs2_free_write_ctxt(wc); 2248 2249 if (data_ac) { 2250 ocfs2_free_alloc_context(data_ac); 2251 data_ac = NULL; 2252 } 2253 if (meta_ac) { 2254 ocfs2_free_alloc_context(meta_ac); 2255 meta_ac = NULL; 2256 } 2257 2258 if (ret == -ENOSPC && try_free) { 2259 /* 2260 * Try to free some truncate log so that we can have enough 2261 * clusters to allocate. 2262 */ 2263 try_free = 0; 2264 2265 ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need); 2266 if (ret1 == 1) 2267 goto try_again; 2268 2269 if (ret1 < 0) 2270 mlog_errno(ret1); 2271 } 2272 2273 return ret; 2274 } 2275 2276 static int ocfs2_write_begin(struct file *file, struct address_space *mapping, 2277 loff_t pos, unsigned len, unsigned flags, 2278 struct page **pagep, void **fsdata) 2279 { 2280 int ret; 2281 struct buffer_head *di_bh = NULL; 2282 struct inode *inode = mapping->host; 2283 2284 ret = ocfs2_inode_lock(inode, &di_bh, 1); 2285 if (ret) { 2286 mlog_errno(ret); 2287 return ret; 2288 } 2289 2290 /* 2291 * Take alloc sem here to prevent concurrent lookups. That way 2292 * the mapping, zeroing and tree manipulation within 2293 * ocfs2_write() will be safe against ->readpage(). This 2294 * should also serve to lock out allocation from a shared 2295 * writeable region. 2296 */ 2297 down_write(&OCFS2_I(inode)->ip_alloc_sem); 2298 2299 ret = ocfs2_write_begin_nolock(file, mapping, pos, len, flags, pagep, 2300 fsdata, di_bh, NULL); 2301 if (ret) { 2302 mlog_errno(ret); 2303 goto out_fail; 2304 } 2305 2306 brelse(di_bh); 2307 2308 return 0; 2309 2310 out_fail: 2311 up_write(&OCFS2_I(inode)->ip_alloc_sem); 2312 2313 brelse(di_bh); 2314 ocfs2_inode_unlock(inode, 1); 2315 2316 return ret; 2317 } 2318 2319 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos, 2320 unsigned len, unsigned *copied, 2321 struct ocfs2_dinode *di, 2322 struct ocfs2_write_ctxt *wc) 2323 { 2324 void *kaddr; 2325 2326 if (unlikely(*copied < len)) { 2327 if (!PageUptodate(wc->w_target_page)) { 2328 *copied = 0; 2329 return; 2330 } 2331 } 2332 2333 kaddr = kmap_atomic(wc->w_target_page); 2334 memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied); 2335 kunmap_atomic(kaddr); 2336 2337 trace_ocfs2_write_end_inline( 2338 (unsigned long long)OCFS2_I(inode)->ip_blkno, 2339 (unsigned long long)pos, *copied, 2340 le16_to_cpu(di->id2.i_data.id_count), 2341 le16_to_cpu(di->i_dyn_features)); 2342 } 2343 2344 int ocfs2_write_end_nolock(struct address_space *mapping, 2345 loff_t pos, unsigned len, unsigned copied, 2346 struct page *page, void *fsdata) 2347 { 2348 int i; 2349 unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1); 2350 struct inode *inode = mapping->host; 2351 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 2352 struct ocfs2_write_ctxt *wc = fsdata; 2353 struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data; 2354 handle_t *handle = wc->w_handle; 2355 struct page *tmppage; 2356 2357 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) { 2358 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc); 2359 goto out_write_size; 2360 } 2361 2362 if (unlikely(copied < len)) { 2363 if (!PageUptodate(wc->w_target_page)) 2364 copied = 0; 2365 2366 ocfs2_zero_new_buffers(wc->w_target_page, start+copied, 2367 start+len); 2368 } 2369 flush_dcache_page(wc->w_target_page); 2370 2371 for(i = 0; i < wc->w_num_pages; i++) { 2372 tmppage = wc->w_pages[i]; 2373 2374 if (tmppage == wc->w_target_page) { 2375 from = wc->w_target_from; 2376 to = wc->w_target_to; 2377 2378 BUG_ON(from > PAGE_CACHE_SIZE || 2379 to > PAGE_CACHE_SIZE || 2380 to < from); 2381 } else { 2382 /* 2383 * Pages adjacent to the target (if any) imply 2384 * a hole-filling write in which case we want 2385 * to flush their entire range. 2386 */ 2387 from = 0; 2388 to = PAGE_CACHE_SIZE; 2389 } 2390 2391 if (page_has_buffers(tmppage)) { 2392 if (ocfs2_should_order_data(inode)) 2393 ocfs2_jbd2_file_inode(wc->w_handle, inode); 2394 block_commit_write(tmppage, from, to); 2395 } 2396 } 2397 2398 out_write_size: 2399 pos += copied; 2400 if (pos > i_size_read(inode)) { 2401 i_size_write(inode, pos); 2402 mark_inode_dirty(inode); 2403 } 2404 inode->i_blocks = ocfs2_inode_sector_count(inode); 2405 di->i_size = cpu_to_le64((u64)i_size_read(inode)); 2406 inode->i_mtime = inode->i_ctime = CURRENT_TIME; 2407 di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec); 2408 di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec); 2409 ocfs2_update_inode_fsync_trans(handle, inode, 1); 2410 ocfs2_journal_dirty(handle, wc->w_di_bh); 2411 2412 /* unlock pages before dealloc since it needs acquiring j_trans_barrier 2413 * lock, or it will cause a deadlock since journal commit threads holds 2414 * this lock and will ask for the page lock when flushing the data. 2415 * put it here to preserve the unlock order. 2416 */ 2417 ocfs2_unlock_pages(wc); 2418 2419 ocfs2_commit_trans(osb, handle); 2420 2421 ocfs2_run_deallocs(osb, &wc->w_dealloc); 2422 2423 brelse(wc->w_di_bh); 2424 kfree(wc); 2425 2426 return copied; 2427 } 2428 2429 static int ocfs2_write_end(struct file *file, struct address_space *mapping, 2430 loff_t pos, unsigned len, unsigned copied, 2431 struct page *page, void *fsdata) 2432 { 2433 int ret; 2434 struct inode *inode = mapping->host; 2435 2436 ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata); 2437 2438 up_write(&OCFS2_I(inode)->ip_alloc_sem); 2439 ocfs2_inode_unlock(inode, 1); 2440 2441 return ret; 2442 } 2443 2444 const struct address_space_operations ocfs2_aops = { 2445 .readpage = ocfs2_readpage, 2446 .readpages = ocfs2_readpages, 2447 .writepage = ocfs2_writepage, 2448 .write_begin = ocfs2_write_begin, 2449 .write_end = ocfs2_write_end, 2450 .bmap = ocfs2_bmap, 2451 .direct_IO = ocfs2_direct_IO, 2452 .invalidatepage = block_invalidatepage, 2453 .releasepage = ocfs2_releasepage, 2454 .migratepage = buffer_migrate_page, 2455 .is_partially_uptodate = block_is_partially_uptodate, 2456 .error_remove_page = generic_error_remove_page, 2457 }; 2458