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 32 #define MLOG_MASK_PREFIX ML_FILE_IO 33 #include <cluster/masklog.h> 34 35 #include "ocfs2.h" 36 37 #include "alloc.h" 38 #include "aops.h" 39 #include "dlmglue.h" 40 #include "extent_map.h" 41 #include "file.h" 42 #include "inode.h" 43 #include "journal.h" 44 #include "suballoc.h" 45 #include "super.h" 46 #include "symlink.h" 47 #include "refcounttree.h" 48 49 #include "buffer_head_io.h" 50 51 static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock, 52 struct buffer_head *bh_result, int create) 53 { 54 int err = -EIO; 55 int status; 56 struct ocfs2_dinode *fe = NULL; 57 struct buffer_head *bh = NULL; 58 struct buffer_head *buffer_cache_bh = NULL; 59 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 60 void *kaddr; 61 62 mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode, 63 (unsigned long long)iblock, bh_result, create); 64 65 BUG_ON(ocfs2_inode_is_fast_symlink(inode)); 66 67 if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) { 68 mlog(ML_ERROR, "block offset > PATH_MAX: %llu", 69 (unsigned long long)iblock); 70 goto bail; 71 } 72 73 status = ocfs2_read_inode_block(inode, &bh); 74 if (status < 0) { 75 mlog_errno(status); 76 goto bail; 77 } 78 fe = (struct ocfs2_dinode *) bh->b_data; 79 80 if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb, 81 le32_to_cpu(fe->i_clusters))) { 82 mlog(ML_ERROR, "block offset is outside the allocated size: " 83 "%llu\n", (unsigned long long)iblock); 84 goto bail; 85 } 86 87 /* We don't use the page cache to create symlink data, so if 88 * need be, copy it over from the buffer cache. */ 89 if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) { 90 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + 91 iblock; 92 buffer_cache_bh = sb_getblk(osb->sb, blkno); 93 if (!buffer_cache_bh) { 94 mlog(ML_ERROR, "couldn't getblock for symlink!\n"); 95 goto bail; 96 } 97 98 /* we haven't locked out transactions, so a commit 99 * could've happened. Since we've got a reference on 100 * the bh, even if it commits while we're doing the 101 * copy, the data is still good. */ 102 if (buffer_jbd(buffer_cache_bh) 103 && ocfs2_inode_is_new(inode)) { 104 kaddr = kmap_atomic(bh_result->b_page, KM_USER0); 105 if (!kaddr) { 106 mlog(ML_ERROR, "couldn't kmap!\n"); 107 goto bail; 108 } 109 memcpy(kaddr + (bh_result->b_size * iblock), 110 buffer_cache_bh->b_data, 111 bh_result->b_size); 112 kunmap_atomic(kaddr, KM_USER0); 113 set_buffer_uptodate(bh_result); 114 } 115 brelse(buffer_cache_bh); 116 } 117 118 map_bh(bh_result, inode->i_sb, 119 le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock); 120 121 err = 0; 122 123 bail: 124 brelse(bh); 125 126 mlog_exit(err); 127 return err; 128 } 129 130 int ocfs2_get_block(struct inode *inode, sector_t iblock, 131 struct buffer_head *bh_result, int create) 132 { 133 int err = 0; 134 unsigned int ext_flags; 135 u64 max_blocks = bh_result->b_size >> inode->i_blkbits; 136 u64 p_blkno, count, past_eof; 137 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 138 139 mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode, 140 (unsigned long long)iblock, bh_result, create); 141 142 if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE) 143 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n", 144 inode, inode->i_ino); 145 146 if (S_ISLNK(inode->i_mode)) { 147 /* this always does I/O for some reason. */ 148 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create); 149 goto bail; 150 } 151 152 err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, &count, 153 &ext_flags); 154 if (err) { 155 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, " 156 "%llu, NULL)\n", err, inode, (unsigned long long)iblock, 157 (unsigned long long)p_blkno); 158 goto bail; 159 } 160 161 if (max_blocks < count) 162 count = max_blocks; 163 164 /* 165 * ocfs2 never allocates in this function - the only time we 166 * need to use BH_New is when we're extending i_size on a file 167 * system which doesn't support holes, in which case BH_New 168 * allows __block_write_begin() to zero. 169 * 170 * If we see this on a sparse file system, then a truncate has 171 * raced us and removed the cluster. In this case, we clear 172 * the buffers dirty and uptodate bits and let the buffer code 173 * ignore it as a hole. 174 */ 175 if (create && p_blkno == 0 && ocfs2_sparse_alloc(osb)) { 176 clear_buffer_dirty(bh_result); 177 clear_buffer_uptodate(bh_result); 178 goto bail; 179 } 180 181 /* Treat the unwritten extent as a hole for zeroing purposes. */ 182 if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN)) 183 map_bh(bh_result, inode->i_sb, p_blkno); 184 185 bh_result->b_size = count << inode->i_blkbits; 186 187 if (!ocfs2_sparse_alloc(osb)) { 188 if (p_blkno == 0) { 189 err = -EIO; 190 mlog(ML_ERROR, 191 "iblock = %llu p_blkno = %llu blkno=(%llu)\n", 192 (unsigned long long)iblock, 193 (unsigned long long)p_blkno, 194 (unsigned long long)OCFS2_I(inode)->ip_blkno); 195 mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters); 196 dump_stack(); 197 goto bail; 198 } 199 } 200 201 past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode)); 202 mlog(0, "Inode %lu, past_eof = %llu\n", inode->i_ino, 203 (unsigned long long)past_eof); 204 if (create && (iblock >= past_eof)) 205 set_buffer_new(bh_result); 206 207 bail: 208 if (err < 0) 209 err = -EIO; 210 211 mlog_exit(err); 212 return err; 213 } 214 215 int ocfs2_read_inline_data(struct inode *inode, struct page *page, 216 struct buffer_head *di_bh) 217 { 218 void *kaddr; 219 loff_t size; 220 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 221 222 if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) { 223 ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag", 224 (unsigned long long)OCFS2_I(inode)->ip_blkno); 225 return -EROFS; 226 } 227 228 size = i_size_read(inode); 229 230 if (size > PAGE_CACHE_SIZE || 231 size > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) { 232 ocfs2_error(inode->i_sb, 233 "Inode %llu has with inline data has bad size: %Lu", 234 (unsigned long long)OCFS2_I(inode)->ip_blkno, 235 (unsigned long long)size); 236 return -EROFS; 237 } 238 239 kaddr = kmap_atomic(page, KM_USER0); 240 if (size) 241 memcpy(kaddr, di->id2.i_data.id_data, size); 242 /* Clear the remaining part of the page */ 243 memset(kaddr + size, 0, PAGE_CACHE_SIZE - size); 244 flush_dcache_page(page); 245 kunmap_atomic(kaddr, KM_USER0); 246 247 SetPageUptodate(page); 248 249 return 0; 250 } 251 252 static int ocfs2_readpage_inline(struct inode *inode, struct page *page) 253 { 254 int ret; 255 struct buffer_head *di_bh = NULL; 256 257 BUG_ON(!PageLocked(page)); 258 BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)); 259 260 ret = ocfs2_read_inode_block(inode, &di_bh); 261 if (ret) { 262 mlog_errno(ret); 263 goto out; 264 } 265 266 ret = ocfs2_read_inline_data(inode, page, di_bh); 267 out: 268 unlock_page(page); 269 270 brelse(di_bh); 271 return ret; 272 } 273 274 static int ocfs2_readpage(struct file *file, struct page *page) 275 { 276 struct inode *inode = page->mapping->host; 277 struct ocfs2_inode_info *oi = OCFS2_I(inode); 278 loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT; 279 int ret, unlock = 1; 280 281 mlog_entry("(0x%p, %lu)\n", file, (page ? page->index : 0)); 282 283 ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page); 284 if (ret != 0) { 285 if (ret == AOP_TRUNCATED_PAGE) 286 unlock = 0; 287 mlog_errno(ret); 288 goto out; 289 } 290 291 if (down_read_trylock(&oi->ip_alloc_sem) == 0) { 292 ret = AOP_TRUNCATED_PAGE; 293 goto out_inode_unlock; 294 } 295 296 /* 297 * i_size might have just been updated as we grabed the meta lock. We 298 * might now be discovering a truncate that hit on another node. 299 * block_read_full_page->get_block freaks out if it is asked to read 300 * beyond the end of a file, so we check here. Callers 301 * (generic_file_read, vm_ops->fault) are clever enough to check i_size 302 * and notice that the page they just read isn't needed. 303 * 304 * XXX sys_readahead() seems to get that wrong? 305 */ 306 if (start >= i_size_read(inode)) { 307 zero_user(page, 0, PAGE_SIZE); 308 SetPageUptodate(page); 309 ret = 0; 310 goto out_alloc; 311 } 312 313 if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) 314 ret = ocfs2_readpage_inline(inode, page); 315 else 316 ret = block_read_full_page(page, ocfs2_get_block); 317 unlock = 0; 318 319 out_alloc: 320 up_read(&OCFS2_I(inode)->ip_alloc_sem); 321 out_inode_unlock: 322 ocfs2_inode_unlock(inode, 0); 323 out: 324 if (unlock) 325 unlock_page(page); 326 mlog_exit(ret); 327 return ret; 328 } 329 330 /* 331 * This is used only for read-ahead. Failures or difficult to handle 332 * situations are safe to ignore. 333 * 334 * Right now, we don't bother with BH_Boundary - in-inode extent lists 335 * are quite large (243 extents on 4k blocks), so most inodes don't 336 * grow out to a tree. If need be, detecting boundary extents could 337 * trivially be added in a future version of ocfs2_get_block(). 338 */ 339 static int ocfs2_readpages(struct file *filp, struct address_space *mapping, 340 struct list_head *pages, unsigned nr_pages) 341 { 342 int ret, err = -EIO; 343 struct inode *inode = mapping->host; 344 struct ocfs2_inode_info *oi = OCFS2_I(inode); 345 loff_t start; 346 struct page *last; 347 348 /* 349 * Use the nonblocking flag for the dlm code to avoid page 350 * lock inversion, but don't bother with retrying. 351 */ 352 ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK); 353 if (ret) 354 return err; 355 356 if (down_read_trylock(&oi->ip_alloc_sem) == 0) { 357 ocfs2_inode_unlock(inode, 0); 358 return err; 359 } 360 361 /* 362 * Don't bother with inline-data. There isn't anything 363 * to read-ahead in that case anyway... 364 */ 365 if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) 366 goto out_unlock; 367 368 /* 369 * Check whether a remote node truncated this file - we just 370 * drop out in that case as it's not worth handling here. 371 */ 372 last = list_entry(pages->prev, struct page, lru); 373 start = (loff_t)last->index << PAGE_CACHE_SHIFT; 374 if (start >= i_size_read(inode)) 375 goto out_unlock; 376 377 err = mpage_readpages(mapping, pages, nr_pages, ocfs2_get_block); 378 379 out_unlock: 380 up_read(&oi->ip_alloc_sem); 381 ocfs2_inode_unlock(inode, 0); 382 383 return err; 384 } 385 386 /* Note: Because we don't support holes, our allocation has 387 * already happened (allocation writes zeros to the file data) 388 * so we don't have to worry about ordered writes in 389 * ocfs2_writepage. 390 * 391 * ->writepage is called during the process of invalidating the page cache 392 * during blocked lock processing. It can't block on any cluster locks 393 * to during block mapping. It's relying on the fact that the block 394 * mapping can't have disappeared under the dirty pages that it is 395 * being asked to write back. 396 */ 397 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc) 398 { 399 int ret; 400 401 mlog_entry("(0x%p)\n", page); 402 403 ret = block_write_full_page(page, ocfs2_get_block, wbc); 404 405 mlog_exit(ret); 406 407 return ret; 408 } 409 410 /* Taken from ext3. We don't necessarily need the full blown 411 * functionality yet, but IMHO it's better to cut and paste the whole 412 * thing so we can avoid introducing our own bugs (and easily pick up 413 * their fixes when they happen) --Mark */ 414 int walk_page_buffers( handle_t *handle, 415 struct buffer_head *head, 416 unsigned from, 417 unsigned to, 418 int *partial, 419 int (*fn)( handle_t *handle, 420 struct buffer_head *bh)) 421 { 422 struct buffer_head *bh; 423 unsigned block_start, block_end; 424 unsigned blocksize = head->b_size; 425 int err, ret = 0; 426 struct buffer_head *next; 427 428 for ( bh = head, block_start = 0; 429 ret == 0 && (bh != head || !block_start); 430 block_start = block_end, bh = next) 431 { 432 next = bh->b_this_page; 433 block_end = block_start + blocksize; 434 if (block_end <= from || block_start >= to) { 435 if (partial && !buffer_uptodate(bh)) 436 *partial = 1; 437 continue; 438 } 439 err = (*fn)(handle, bh); 440 if (!ret) 441 ret = err; 442 } 443 return ret; 444 } 445 446 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block) 447 { 448 sector_t status; 449 u64 p_blkno = 0; 450 int err = 0; 451 struct inode *inode = mapping->host; 452 453 mlog_entry("(block = %llu)\n", (unsigned long long)block); 454 455 /* We don't need to lock journal system files, since they aren't 456 * accessed concurrently from multiple nodes. 457 */ 458 if (!INODE_JOURNAL(inode)) { 459 err = ocfs2_inode_lock(inode, NULL, 0); 460 if (err) { 461 if (err != -ENOENT) 462 mlog_errno(err); 463 goto bail; 464 } 465 down_read(&OCFS2_I(inode)->ip_alloc_sem); 466 } 467 468 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)) 469 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL, 470 NULL); 471 472 if (!INODE_JOURNAL(inode)) { 473 up_read(&OCFS2_I(inode)->ip_alloc_sem); 474 ocfs2_inode_unlock(inode, 0); 475 } 476 477 if (err) { 478 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n", 479 (unsigned long long)block); 480 mlog_errno(err); 481 goto bail; 482 } 483 484 bail: 485 status = err ? 0 : p_blkno; 486 487 mlog_exit((int)status); 488 489 return status; 490 } 491 492 /* 493 * TODO: Make this into a generic get_blocks function. 494 * 495 * From do_direct_io in direct-io.c: 496 * "So what we do is to permit the ->get_blocks function to populate 497 * bh.b_size with the size of IO which is permitted at this offset and 498 * this i_blkbits." 499 * 500 * This function is called directly from get_more_blocks in direct-io.c. 501 * 502 * called like this: dio->get_blocks(dio->inode, fs_startblk, 503 * fs_count, map_bh, dio->rw == WRITE); 504 * 505 * Note that we never bother to allocate blocks here, and thus ignore the 506 * create argument. 507 */ 508 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock, 509 struct buffer_head *bh_result, int create) 510 { 511 int ret; 512 u64 p_blkno, inode_blocks, contig_blocks; 513 unsigned int ext_flags; 514 unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits; 515 unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits; 516 517 /* This function won't even be called if the request isn't all 518 * nicely aligned and of the right size, so there's no need 519 * for us to check any of that. */ 520 521 inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode)); 522 523 /* This figures out the size of the next contiguous block, and 524 * our logical offset */ 525 ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, 526 &contig_blocks, &ext_flags); 527 if (ret) { 528 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n", 529 (unsigned long long)iblock); 530 ret = -EIO; 531 goto bail; 532 } 533 534 /* We should already CoW the refcounted extent in case of create. */ 535 BUG_ON(create && (ext_flags & OCFS2_EXT_REFCOUNTED)); 536 537 /* 538 * get_more_blocks() expects us to describe a hole by clearing 539 * the mapped bit on bh_result(). 540 * 541 * Consider an unwritten extent as a hole. 542 */ 543 if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN)) 544 map_bh(bh_result, inode->i_sb, p_blkno); 545 else 546 clear_buffer_mapped(bh_result); 547 548 /* make sure we don't map more than max_blocks blocks here as 549 that's all the kernel will handle at this point. */ 550 if (max_blocks < contig_blocks) 551 contig_blocks = max_blocks; 552 bh_result->b_size = contig_blocks << blocksize_bits; 553 bail: 554 return ret; 555 } 556 557 /* 558 * ocfs2_dio_end_io is called by the dio core when a dio is finished. We're 559 * particularly interested in the aio/dio case. Like the core uses 560 * i_alloc_sem, we use the rw_lock DLM lock to protect io on one node from 561 * truncation on another. 562 */ 563 static void ocfs2_dio_end_io(struct kiocb *iocb, 564 loff_t offset, 565 ssize_t bytes, 566 void *private, 567 int ret, 568 bool is_async) 569 { 570 struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode; 571 int level; 572 573 /* this io's submitter should not have unlocked this before we could */ 574 BUG_ON(!ocfs2_iocb_is_rw_locked(iocb)); 575 576 if (ocfs2_iocb_is_sem_locked(iocb)) { 577 up_read(&inode->i_alloc_sem); 578 ocfs2_iocb_clear_sem_locked(iocb); 579 } 580 581 ocfs2_iocb_clear_rw_locked(iocb); 582 583 level = ocfs2_iocb_rw_locked_level(iocb); 584 ocfs2_rw_unlock(inode, level); 585 586 if (is_async) 587 aio_complete(iocb, ret, 0); 588 } 589 590 /* 591 * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen 592 * from ext3. PageChecked() bits have been removed as OCFS2 does not 593 * do journalled data. 594 */ 595 static void ocfs2_invalidatepage(struct page *page, unsigned long offset) 596 { 597 journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal; 598 599 jbd2_journal_invalidatepage(journal, page, offset); 600 } 601 602 static int ocfs2_releasepage(struct page *page, gfp_t wait) 603 { 604 journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal; 605 606 if (!page_has_buffers(page)) 607 return 0; 608 return jbd2_journal_try_to_free_buffers(journal, page, wait); 609 } 610 611 static ssize_t ocfs2_direct_IO(int rw, 612 struct kiocb *iocb, 613 const struct iovec *iov, 614 loff_t offset, 615 unsigned long nr_segs) 616 { 617 struct file *file = iocb->ki_filp; 618 struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host; 619 int ret; 620 621 mlog_entry_void(); 622 623 /* 624 * Fallback to buffered I/O if we see an inode without 625 * extents. 626 */ 627 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) 628 return 0; 629 630 /* Fallback to buffered I/O if we are appending. */ 631 if (i_size_read(inode) <= offset) 632 return 0; 633 634 ret = __blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, 635 iov, offset, nr_segs, 636 ocfs2_direct_IO_get_blocks, 637 ocfs2_dio_end_io, NULL, 0); 638 639 mlog_exit(ret); 640 return ret; 641 } 642 643 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb, 644 u32 cpos, 645 unsigned int *start, 646 unsigned int *end) 647 { 648 unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE; 649 650 if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) { 651 unsigned int cpp; 652 653 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits); 654 655 cluster_start = cpos % cpp; 656 cluster_start = cluster_start << osb->s_clustersize_bits; 657 658 cluster_end = cluster_start + osb->s_clustersize; 659 } 660 661 BUG_ON(cluster_start > PAGE_SIZE); 662 BUG_ON(cluster_end > PAGE_SIZE); 663 664 if (start) 665 *start = cluster_start; 666 if (end) 667 *end = cluster_end; 668 } 669 670 /* 671 * 'from' and 'to' are the region in the page to avoid zeroing. 672 * 673 * If pagesize > clustersize, this function will avoid zeroing outside 674 * of the cluster boundary. 675 * 676 * from == to == 0 is code for "zero the entire cluster region" 677 */ 678 static void ocfs2_clear_page_regions(struct page *page, 679 struct ocfs2_super *osb, u32 cpos, 680 unsigned from, unsigned to) 681 { 682 void *kaddr; 683 unsigned int cluster_start, cluster_end; 684 685 ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end); 686 687 kaddr = kmap_atomic(page, KM_USER0); 688 689 if (from || to) { 690 if (from > cluster_start) 691 memset(kaddr + cluster_start, 0, from - cluster_start); 692 if (to < cluster_end) 693 memset(kaddr + to, 0, cluster_end - to); 694 } else { 695 memset(kaddr + cluster_start, 0, cluster_end - cluster_start); 696 } 697 698 kunmap_atomic(kaddr, KM_USER0); 699 } 700 701 /* 702 * Nonsparse file systems fully allocate before we get to the write 703 * code. This prevents ocfs2_write() from tagging the write as an 704 * allocating one, which means ocfs2_map_page_blocks() might try to 705 * read-in the blocks at the tail of our file. Avoid reading them by 706 * testing i_size against each block offset. 707 */ 708 static int ocfs2_should_read_blk(struct inode *inode, struct page *page, 709 unsigned int block_start) 710 { 711 u64 offset = page_offset(page) + block_start; 712 713 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) 714 return 1; 715 716 if (i_size_read(inode) > offset) 717 return 1; 718 719 return 0; 720 } 721 722 /* 723 * Some of this taken from __block_write_begin(). We already have our 724 * mapping by now though, and the entire write will be allocating or 725 * it won't, so not much need to use BH_New. 726 * 727 * This will also skip zeroing, which is handled externally. 728 */ 729 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno, 730 struct inode *inode, unsigned int from, 731 unsigned int to, int new) 732 { 733 int ret = 0; 734 struct buffer_head *head, *bh, *wait[2], **wait_bh = wait; 735 unsigned int block_end, block_start; 736 unsigned int bsize = 1 << inode->i_blkbits; 737 738 if (!page_has_buffers(page)) 739 create_empty_buffers(page, bsize, 0); 740 741 head = page_buffers(page); 742 for (bh = head, block_start = 0; bh != head || !block_start; 743 bh = bh->b_this_page, block_start += bsize) { 744 block_end = block_start + bsize; 745 746 clear_buffer_new(bh); 747 748 /* 749 * Ignore blocks outside of our i/o range - 750 * they may belong to unallocated clusters. 751 */ 752 if (block_start >= to || block_end <= from) { 753 if (PageUptodate(page)) 754 set_buffer_uptodate(bh); 755 continue; 756 } 757 758 /* 759 * For an allocating write with cluster size >= page 760 * size, we always write the entire page. 761 */ 762 if (new) 763 set_buffer_new(bh); 764 765 if (!buffer_mapped(bh)) { 766 map_bh(bh, inode->i_sb, *p_blkno); 767 unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr); 768 } 769 770 if (PageUptodate(page)) { 771 if (!buffer_uptodate(bh)) 772 set_buffer_uptodate(bh); 773 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) && 774 !buffer_new(bh) && 775 ocfs2_should_read_blk(inode, page, block_start) && 776 (block_start < from || block_end > to)) { 777 ll_rw_block(READ, 1, &bh); 778 *wait_bh++=bh; 779 } 780 781 *p_blkno = *p_blkno + 1; 782 } 783 784 /* 785 * If we issued read requests - let them complete. 786 */ 787 while(wait_bh > wait) { 788 wait_on_buffer(*--wait_bh); 789 if (!buffer_uptodate(*wait_bh)) 790 ret = -EIO; 791 } 792 793 if (ret == 0 || !new) 794 return ret; 795 796 /* 797 * If we get -EIO above, zero out any newly allocated blocks 798 * to avoid exposing stale data. 799 */ 800 bh = head; 801 block_start = 0; 802 do { 803 block_end = block_start + bsize; 804 if (block_end <= from) 805 goto next_bh; 806 if (block_start >= to) 807 break; 808 809 zero_user(page, block_start, bh->b_size); 810 set_buffer_uptodate(bh); 811 mark_buffer_dirty(bh); 812 813 next_bh: 814 block_start = block_end; 815 bh = bh->b_this_page; 816 } while (bh != head); 817 818 return ret; 819 } 820 821 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE) 822 #define OCFS2_MAX_CTXT_PAGES 1 823 #else 824 #define OCFS2_MAX_CTXT_PAGES (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE) 825 #endif 826 827 #define OCFS2_MAX_CLUSTERS_PER_PAGE (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE) 828 829 /* 830 * Describe the state of a single cluster to be written to. 831 */ 832 struct ocfs2_write_cluster_desc { 833 u32 c_cpos; 834 u32 c_phys; 835 /* 836 * Give this a unique field because c_phys eventually gets 837 * filled. 838 */ 839 unsigned c_new; 840 unsigned c_unwritten; 841 unsigned c_needs_zero; 842 }; 843 844 struct ocfs2_write_ctxt { 845 /* Logical cluster position / len of write */ 846 u32 w_cpos; 847 u32 w_clen; 848 849 /* First cluster allocated in a nonsparse extend */ 850 u32 w_first_new_cpos; 851 852 struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE]; 853 854 /* 855 * This is true if page_size > cluster_size. 856 * 857 * It triggers a set of special cases during write which might 858 * have to deal with allocating writes to partial pages. 859 */ 860 unsigned int w_large_pages; 861 862 /* 863 * Pages involved in this write. 864 * 865 * w_target_page is the page being written to by the user. 866 * 867 * w_pages is an array of pages which always contains 868 * w_target_page, and in the case of an allocating write with 869 * page_size < cluster size, it will contain zero'd and mapped 870 * pages adjacent to w_target_page which need to be written 871 * out in so that future reads from that region will get 872 * zero's. 873 */ 874 unsigned int w_num_pages; 875 struct page *w_pages[OCFS2_MAX_CTXT_PAGES]; 876 struct page *w_target_page; 877 878 /* 879 * ocfs2_write_end() uses this to know what the real range to 880 * write in the target should be. 881 */ 882 unsigned int w_target_from; 883 unsigned int w_target_to; 884 885 /* 886 * We could use journal_current_handle() but this is cleaner, 887 * IMHO -Mark 888 */ 889 handle_t *w_handle; 890 891 struct buffer_head *w_di_bh; 892 893 struct ocfs2_cached_dealloc_ctxt w_dealloc; 894 }; 895 896 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages) 897 { 898 int i; 899 900 for(i = 0; i < num_pages; i++) { 901 if (pages[i]) { 902 unlock_page(pages[i]); 903 mark_page_accessed(pages[i]); 904 page_cache_release(pages[i]); 905 } 906 } 907 } 908 909 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc) 910 { 911 ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages); 912 913 brelse(wc->w_di_bh); 914 kfree(wc); 915 } 916 917 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp, 918 struct ocfs2_super *osb, loff_t pos, 919 unsigned len, struct buffer_head *di_bh) 920 { 921 u32 cend; 922 struct ocfs2_write_ctxt *wc; 923 924 wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS); 925 if (!wc) 926 return -ENOMEM; 927 928 wc->w_cpos = pos >> osb->s_clustersize_bits; 929 wc->w_first_new_cpos = UINT_MAX; 930 cend = (pos + len - 1) >> osb->s_clustersize_bits; 931 wc->w_clen = cend - wc->w_cpos + 1; 932 get_bh(di_bh); 933 wc->w_di_bh = di_bh; 934 935 if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) 936 wc->w_large_pages = 1; 937 else 938 wc->w_large_pages = 0; 939 940 ocfs2_init_dealloc_ctxt(&wc->w_dealloc); 941 942 *wcp = wc; 943 944 return 0; 945 } 946 947 /* 948 * If a page has any new buffers, zero them out here, and mark them uptodate 949 * and dirty so they'll be written out (in order to prevent uninitialised 950 * block data from leaking). And clear the new bit. 951 */ 952 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to) 953 { 954 unsigned int block_start, block_end; 955 struct buffer_head *head, *bh; 956 957 BUG_ON(!PageLocked(page)); 958 if (!page_has_buffers(page)) 959 return; 960 961 bh = head = page_buffers(page); 962 block_start = 0; 963 do { 964 block_end = block_start + bh->b_size; 965 966 if (buffer_new(bh)) { 967 if (block_end > from && block_start < to) { 968 if (!PageUptodate(page)) { 969 unsigned start, end; 970 971 start = max(from, block_start); 972 end = min(to, block_end); 973 974 zero_user_segment(page, start, end); 975 set_buffer_uptodate(bh); 976 } 977 978 clear_buffer_new(bh); 979 mark_buffer_dirty(bh); 980 } 981 } 982 983 block_start = block_end; 984 bh = bh->b_this_page; 985 } while (bh != head); 986 } 987 988 /* 989 * Only called when we have a failure during allocating write to write 990 * zero's to the newly allocated region. 991 */ 992 static void ocfs2_write_failure(struct inode *inode, 993 struct ocfs2_write_ctxt *wc, 994 loff_t user_pos, unsigned user_len) 995 { 996 int i; 997 unsigned from = user_pos & (PAGE_CACHE_SIZE - 1), 998 to = user_pos + user_len; 999 struct page *tmppage; 1000 1001 ocfs2_zero_new_buffers(wc->w_target_page, from, to); 1002 1003 for(i = 0; i < wc->w_num_pages; i++) { 1004 tmppage = wc->w_pages[i]; 1005 1006 if (page_has_buffers(tmppage)) { 1007 if (ocfs2_should_order_data(inode)) 1008 ocfs2_jbd2_file_inode(wc->w_handle, inode); 1009 1010 block_commit_write(tmppage, from, to); 1011 } 1012 } 1013 } 1014 1015 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno, 1016 struct ocfs2_write_ctxt *wc, 1017 struct page *page, u32 cpos, 1018 loff_t user_pos, unsigned user_len, 1019 int new) 1020 { 1021 int ret; 1022 unsigned int map_from = 0, map_to = 0; 1023 unsigned int cluster_start, cluster_end; 1024 unsigned int user_data_from = 0, user_data_to = 0; 1025 1026 ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos, 1027 &cluster_start, &cluster_end); 1028 1029 if (page == wc->w_target_page) { 1030 map_from = user_pos & (PAGE_CACHE_SIZE - 1); 1031 map_to = map_from + user_len; 1032 1033 if (new) 1034 ret = ocfs2_map_page_blocks(page, p_blkno, inode, 1035 cluster_start, cluster_end, 1036 new); 1037 else 1038 ret = ocfs2_map_page_blocks(page, p_blkno, inode, 1039 map_from, map_to, new); 1040 if (ret) { 1041 mlog_errno(ret); 1042 goto out; 1043 } 1044 1045 user_data_from = map_from; 1046 user_data_to = map_to; 1047 if (new) { 1048 map_from = cluster_start; 1049 map_to = cluster_end; 1050 } 1051 } else { 1052 /* 1053 * If we haven't allocated the new page yet, we 1054 * shouldn't be writing it out without copying user 1055 * data. This is likely a math error from the caller. 1056 */ 1057 BUG_ON(!new); 1058 1059 map_from = cluster_start; 1060 map_to = cluster_end; 1061 1062 ret = ocfs2_map_page_blocks(page, p_blkno, inode, 1063 cluster_start, cluster_end, new); 1064 if (ret) { 1065 mlog_errno(ret); 1066 goto out; 1067 } 1068 } 1069 1070 /* 1071 * Parts of newly allocated pages need to be zero'd. 1072 * 1073 * Above, we have also rewritten 'to' and 'from' - as far as 1074 * the rest of the function is concerned, the entire cluster 1075 * range inside of a page needs to be written. 1076 * 1077 * We can skip this if the page is up to date - it's already 1078 * been zero'd from being read in as a hole. 1079 */ 1080 if (new && !PageUptodate(page)) 1081 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb), 1082 cpos, user_data_from, user_data_to); 1083 1084 flush_dcache_page(page); 1085 1086 out: 1087 return ret; 1088 } 1089 1090 /* 1091 * This function will only grab one clusters worth of pages. 1092 */ 1093 static int ocfs2_grab_pages_for_write(struct address_space *mapping, 1094 struct ocfs2_write_ctxt *wc, 1095 u32 cpos, loff_t user_pos, 1096 unsigned user_len, int new, 1097 struct page *mmap_page) 1098 { 1099 int ret = 0, i; 1100 unsigned long start, target_index, end_index, index; 1101 struct inode *inode = mapping->host; 1102 loff_t last_byte; 1103 1104 target_index = user_pos >> PAGE_CACHE_SHIFT; 1105 1106 /* 1107 * Figure out how many pages we'll be manipulating here. For 1108 * non allocating write, we just change the one 1109 * page. Otherwise, we'll need a whole clusters worth. If we're 1110 * writing past i_size, we only need enough pages to cover the 1111 * last page of the write. 1112 */ 1113 if (new) { 1114 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb); 1115 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos); 1116 /* 1117 * We need the index *past* the last page we could possibly 1118 * touch. This is the page past the end of the write or 1119 * i_size, whichever is greater. 1120 */ 1121 last_byte = max(user_pos + user_len, i_size_read(inode)); 1122 BUG_ON(last_byte < 1); 1123 end_index = ((last_byte - 1) >> PAGE_CACHE_SHIFT) + 1; 1124 if ((start + wc->w_num_pages) > end_index) 1125 wc->w_num_pages = end_index - start; 1126 } else { 1127 wc->w_num_pages = 1; 1128 start = target_index; 1129 } 1130 1131 for(i = 0; i < wc->w_num_pages; i++) { 1132 index = start + i; 1133 1134 if (index == target_index && mmap_page) { 1135 /* 1136 * ocfs2_pagemkwrite() is a little different 1137 * and wants us to directly use the page 1138 * passed in. 1139 */ 1140 lock_page(mmap_page); 1141 1142 if (mmap_page->mapping != mapping) { 1143 unlock_page(mmap_page); 1144 /* 1145 * Sanity check - the locking in 1146 * ocfs2_pagemkwrite() should ensure 1147 * that this code doesn't trigger. 1148 */ 1149 ret = -EINVAL; 1150 mlog_errno(ret); 1151 goto out; 1152 } 1153 1154 page_cache_get(mmap_page); 1155 wc->w_pages[i] = mmap_page; 1156 } else { 1157 wc->w_pages[i] = find_or_create_page(mapping, index, 1158 GFP_NOFS); 1159 if (!wc->w_pages[i]) { 1160 ret = -ENOMEM; 1161 mlog_errno(ret); 1162 goto out; 1163 } 1164 } 1165 1166 if (index == target_index) 1167 wc->w_target_page = wc->w_pages[i]; 1168 } 1169 out: 1170 return ret; 1171 } 1172 1173 /* 1174 * Prepare a single cluster for write one cluster into the file. 1175 */ 1176 static int ocfs2_write_cluster(struct address_space *mapping, 1177 u32 phys, unsigned int unwritten, 1178 unsigned int should_zero, 1179 struct ocfs2_alloc_context *data_ac, 1180 struct ocfs2_alloc_context *meta_ac, 1181 struct ocfs2_write_ctxt *wc, u32 cpos, 1182 loff_t user_pos, unsigned user_len) 1183 { 1184 int ret, i, new; 1185 u64 v_blkno, p_blkno; 1186 struct inode *inode = mapping->host; 1187 struct ocfs2_extent_tree et; 1188 1189 new = phys == 0 ? 1 : 0; 1190 if (new) { 1191 u32 tmp_pos; 1192 1193 /* 1194 * This is safe to call with the page locks - it won't take 1195 * any additional semaphores or cluster locks. 1196 */ 1197 tmp_pos = cpos; 1198 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode, 1199 &tmp_pos, 1, 0, wc->w_di_bh, 1200 wc->w_handle, data_ac, 1201 meta_ac, NULL); 1202 /* 1203 * This shouldn't happen because we must have already 1204 * calculated the correct meta data allocation required. The 1205 * internal tree allocation code should know how to increase 1206 * transaction credits itself. 1207 * 1208 * If need be, we could handle -EAGAIN for a 1209 * RESTART_TRANS here. 1210 */ 1211 mlog_bug_on_msg(ret == -EAGAIN, 1212 "Inode %llu: EAGAIN return during allocation.\n", 1213 (unsigned long long)OCFS2_I(inode)->ip_blkno); 1214 if (ret < 0) { 1215 mlog_errno(ret); 1216 goto out; 1217 } 1218 } else if (unwritten) { 1219 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), 1220 wc->w_di_bh); 1221 ret = ocfs2_mark_extent_written(inode, &et, 1222 wc->w_handle, cpos, 1, phys, 1223 meta_ac, &wc->w_dealloc); 1224 if (ret < 0) { 1225 mlog_errno(ret); 1226 goto out; 1227 } 1228 } 1229 1230 if (should_zero) 1231 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos); 1232 else 1233 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits; 1234 1235 /* 1236 * The only reason this should fail is due to an inability to 1237 * find the extent added. 1238 */ 1239 ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL, 1240 NULL); 1241 if (ret < 0) { 1242 ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, " 1243 "at logical block %llu", 1244 (unsigned long long)OCFS2_I(inode)->ip_blkno, 1245 (unsigned long long)v_blkno); 1246 goto out; 1247 } 1248 1249 BUG_ON(p_blkno == 0); 1250 1251 for(i = 0; i < wc->w_num_pages; i++) { 1252 int tmpret; 1253 1254 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc, 1255 wc->w_pages[i], cpos, 1256 user_pos, user_len, 1257 should_zero); 1258 if (tmpret) { 1259 mlog_errno(tmpret); 1260 if (ret == 0) 1261 ret = tmpret; 1262 } 1263 } 1264 1265 /* 1266 * We only have cleanup to do in case of allocating write. 1267 */ 1268 if (ret && new) 1269 ocfs2_write_failure(inode, wc, user_pos, user_len); 1270 1271 out: 1272 1273 return ret; 1274 } 1275 1276 static int ocfs2_write_cluster_by_desc(struct address_space *mapping, 1277 struct ocfs2_alloc_context *data_ac, 1278 struct ocfs2_alloc_context *meta_ac, 1279 struct ocfs2_write_ctxt *wc, 1280 loff_t pos, unsigned len) 1281 { 1282 int ret, i; 1283 loff_t cluster_off; 1284 unsigned int local_len = len; 1285 struct ocfs2_write_cluster_desc *desc; 1286 struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb); 1287 1288 for (i = 0; i < wc->w_clen; i++) { 1289 desc = &wc->w_desc[i]; 1290 1291 /* 1292 * We have to make sure that the total write passed in 1293 * doesn't extend past a single cluster. 1294 */ 1295 local_len = len; 1296 cluster_off = pos & (osb->s_clustersize - 1); 1297 if ((cluster_off + local_len) > osb->s_clustersize) 1298 local_len = osb->s_clustersize - cluster_off; 1299 1300 ret = ocfs2_write_cluster(mapping, desc->c_phys, 1301 desc->c_unwritten, 1302 desc->c_needs_zero, 1303 data_ac, meta_ac, 1304 wc, desc->c_cpos, pos, local_len); 1305 if (ret) { 1306 mlog_errno(ret); 1307 goto out; 1308 } 1309 1310 len -= local_len; 1311 pos += local_len; 1312 } 1313 1314 ret = 0; 1315 out: 1316 return ret; 1317 } 1318 1319 /* 1320 * ocfs2_write_end() wants to know which parts of the target page it 1321 * should complete the write on. It's easiest to compute them ahead of 1322 * time when a more complete view of the write is available. 1323 */ 1324 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb, 1325 struct ocfs2_write_ctxt *wc, 1326 loff_t pos, unsigned len, int alloc) 1327 { 1328 struct ocfs2_write_cluster_desc *desc; 1329 1330 wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1); 1331 wc->w_target_to = wc->w_target_from + len; 1332 1333 if (alloc == 0) 1334 return; 1335 1336 /* 1337 * Allocating write - we may have different boundaries based 1338 * on page size and cluster size. 1339 * 1340 * NOTE: We can no longer compute one value from the other as 1341 * the actual write length and user provided length may be 1342 * different. 1343 */ 1344 1345 if (wc->w_large_pages) { 1346 /* 1347 * We only care about the 1st and last cluster within 1348 * our range and whether they should be zero'd or not. Either 1349 * value may be extended out to the start/end of a 1350 * newly allocated cluster. 1351 */ 1352 desc = &wc->w_desc[0]; 1353 if (desc->c_needs_zero) 1354 ocfs2_figure_cluster_boundaries(osb, 1355 desc->c_cpos, 1356 &wc->w_target_from, 1357 NULL); 1358 1359 desc = &wc->w_desc[wc->w_clen - 1]; 1360 if (desc->c_needs_zero) 1361 ocfs2_figure_cluster_boundaries(osb, 1362 desc->c_cpos, 1363 NULL, 1364 &wc->w_target_to); 1365 } else { 1366 wc->w_target_from = 0; 1367 wc->w_target_to = PAGE_CACHE_SIZE; 1368 } 1369 } 1370 1371 /* 1372 * Populate each single-cluster write descriptor in the write context 1373 * with information about the i/o to be done. 1374 * 1375 * Returns the number of clusters that will have to be allocated, as 1376 * well as a worst case estimate of the number of extent records that 1377 * would have to be created during a write to an unwritten region. 1378 */ 1379 static int ocfs2_populate_write_desc(struct inode *inode, 1380 struct ocfs2_write_ctxt *wc, 1381 unsigned int *clusters_to_alloc, 1382 unsigned int *extents_to_split) 1383 { 1384 int ret; 1385 struct ocfs2_write_cluster_desc *desc; 1386 unsigned int num_clusters = 0; 1387 unsigned int ext_flags = 0; 1388 u32 phys = 0; 1389 int i; 1390 1391 *clusters_to_alloc = 0; 1392 *extents_to_split = 0; 1393 1394 for (i = 0; i < wc->w_clen; i++) { 1395 desc = &wc->w_desc[i]; 1396 desc->c_cpos = wc->w_cpos + i; 1397 1398 if (num_clusters == 0) { 1399 /* 1400 * Need to look up the next extent record. 1401 */ 1402 ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys, 1403 &num_clusters, &ext_flags); 1404 if (ret) { 1405 mlog_errno(ret); 1406 goto out; 1407 } 1408 1409 /* We should already CoW the refcountd extent. */ 1410 BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED); 1411 1412 /* 1413 * Assume worst case - that we're writing in 1414 * the middle of the extent. 1415 * 1416 * We can assume that the write proceeds from 1417 * left to right, in which case the extent 1418 * insert code is smart enough to coalesce the 1419 * next splits into the previous records created. 1420 */ 1421 if (ext_flags & OCFS2_EXT_UNWRITTEN) 1422 *extents_to_split = *extents_to_split + 2; 1423 } else if (phys) { 1424 /* 1425 * Only increment phys if it doesn't describe 1426 * a hole. 1427 */ 1428 phys++; 1429 } 1430 1431 /* 1432 * If w_first_new_cpos is < UINT_MAX, we have a non-sparse 1433 * file that got extended. w_first_new_cpos tells us 1434 * where the newly allocated clusters are so we can 1435 * zero them. 1436 */ 1437 if (desc->c_cpos >= wc->w_first_new_cpos) { 1438 BUG_ON(phys == 0); 1439 desc->c_needs_zero = 1; 1440 } 1441 1442 desc->c_phys = phys; 1443 if (phys == 0) { 1444 desc->c_new = 1; 1445 desc->c_needs_zero = 1; 1446 *clusters_to_alloc = *clusters_to_alloc + 1; 1447 } 1448 1449 if (ext_flags & OCFS2_EXT_UNWRITTEN) { 1450 desc->c_unwritten = 1; 1451 desc->c_needs_zero = 1; 1452 } 1453 1454 num_clusters--; 1455 } 1456 1457 ret = 0; 1458 out: 1459 return ret; 1460 } 1461 1462 static int ocfs2_write_begin_inline(struct address_space *mapping, 1463 struct inode *inode, 1464 struct ocfs2_write_ctxt *wc) 1465 { 1466 int ret; 1467 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 1468 struct page *page; 1469 handle_t *handle; 1470 struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data; 1471 1472 page = find_or_create_page(mapping, 0, GFP_NOFS); 1473 if (!page) { 1474 ret = -ENOMEM; 1475 mlog_errno(ret); 1476 goto out; 1477 } 1478 /* 1479 * If we don't set w_num_pages then this page won't get unlocked 1480 * and freed on cleanup of the write context. 1481 */ 1482 wc->w_pages[0] = wc->w_target_page = page; 1483 wc->w_num_pages = 1; 1484 1485 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS); 1486 if (IS_ERR(handle)) { 1487 ret = PTR_ERR(handle); 1488 mlog_errno(ret); 1489 goto out; 1490 } 1491 1492 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh, 1493 OCFS2_JOURNAL_ACCESS_WRITE); 1494 if (ret) { 1495 ocfs2_commit_trans(osb, handle); 1496 1497 mlog_errno(ret); 1498 goto out; 1499 } 1500 1501 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)) 1502 ocfs2_set_inode_data_inline(inode, di); 1503 1504 if (!PageUptodate(page)) { 1505 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh); 1506 if (ret) { 1507 ocfs2_commit_trans(osb, handle); 1508 1509 goto out; 1510 } 1511 } 1512 1513 wc->w_handle = handle; 1514 out: 1515 return ret; 1516 } 1517 1518 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size) 1519 { 1520 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 1521 1522 if (new_size <= le16_to_cpu(di->id2.i_data.id_count)) 1523 return 1; 1524 return 0; 1525 } 1526 1527 static int ocfs2_try_to_write_inline_data(struct address_space *mapping, 1528 struct inode *inode, loff_t pos, 1529 unsigned len, struct page *mmap_page, 1530 struct ocfs2_write_ctxt *wc) 1531 { 1532 int ret, written = 0; 1533 loff_t end = pos + len; 1534 struct ocfs2_inode_info *oi = OCFS2_I(inode); 1535 struct ocfs2_dinode *di = NULL; 1536 1537 mlog(0, "Inode %llu, write of %u bytes at off %llu. features: 0x%x\n", 1538 (unsigned long long)oi->ip_blkno, len, (unsigned long long)pos, 1539 oi->ip_dyn_features); 1540 1541 /* 1542 * Handle inodes which already have inline data 1st. 1543 */ 1544 if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) { 1545 if (mmap_page == NULL && 1546 ocfs2_size_fits_inline_data(wc->w_di_bh, end)) 1547 goto do_inline_write; 1548 1549 /* 1550 * The write won't fit - we have to give this inode an 1551 * inline extent list now. 1552 */ 1553 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh); 1554 if (ret) 1555 mlog_errno(ret); 1556 goto out; 1557 } 1558 1559 /* 1560 * Check whether the inode can accept inline data. 1561 */ 1562 if (oi->ip_clusters != 0 || i_size_read(inode) != 0) 1563 return 0; 1564 1565 /* 1566 * Check whether the write can fit. 1567 */ 1568 di = (struct ocfs2_dinode *)wc->w_di_bh->b_data; 1569 if (mmap_page || 1570 end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) 1571 return 0; 1572 1573 do_inline_write: 1574 ret = ocfs2_write_begin_inline(mapping, inode, wc); 1575 if (ret) { 1576 mlog_errno(ret); 1577 goto out; 1578 } 1579 1580 /* 1581 * This signals to the caller that the data can be written 1582 * inline. 1583 */ 1584 written = 1; 1585 out: 1586 return written ? written : ret; 1587 } 1588 1589 /* 1590 * This function only does anything for file systems which can't 1591 * handle sparse files. 1592 * 1593 * What we want to do here is fill in any hole between the current end 1594 * of allocation and the end of our write. That way the rest of the 1595 * write path can treat it as an non-allocating write, which has no 1596 * special case code for sparse/nonsparse files. 1597 */ 1598 static int ocfs2_expand_nonsparse_inode(struct inode *inode, 1599 struct buffer_head *di_bh, 1600 loff_t pos, unsigned len, 1601 struct ocfs2_write_ctxt *wc) 1602 { 1603 int ret; 1604 loff_t newsize = pos + len; 1605 1606 BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))); 1607 1608 if (newsize <= i_size_read(inode)) 1609 return 0; 1610 1611 ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos); 1612 if (ret) 1613 mlog_errno(ret); 1614 1615 wc->w_first_new_cpos = 1616 ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode)); 1617 1618 return ret; 1619 } 1620 1621 static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh, 1622 loff_t pos) 1623 { 1624 int ret = 0; 1625 1626 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))); 1627 if (pos > i_size_read(inode)) 1628 ret = ocfs2_zero_extend(inode, di_bh, pos); 1629 1630 return ret; 1631 } 1632 1633 /* 1634 * Try to flush truncate logs if we can free enough clusters from it. 1635 * As for return value, "< 0" means error, "0" no space and "1" means 1636 * we have freed enough spaces and let the caller try to allocate again. 1637 */ 1638 static int ocfs2_try_to_free_truncate_log(struct ocfs2_super *osb, 1639 unsigned int needed) 1640 { 1641 tid_t target; 1642 int ret = 0; 1643 unsigned int truncated_clusters; 1644 1645 mutex_lock(&osb->osb_tl_inode->i_mutex); 1646 truncated_clusters = osb->truncated_clusters; 1647 mutex_unlock(&osb->osb_tl_inode->i_mutex); 1648 1649 /* 1650 * Check whether we can succeed in allocating if we free 1651 * the truncate log. 1652 */ 1653 if (truncated_clusters < needed) 1654 goto out; 1655 1656 ret = ocfs2_flush_truncate_log(osb); 1657 if (ret) { 1658 mlog_errno(ret); 1659 goto out; 1660 } 1661 1662 if (jbd2_journal_start_commit(osb->journal->j_journal, &target)) { 1663 jbd2_log_wait_commit(osb->journal->j_journal, target); 1664 ret = 1; 1665 } 1666 out: 1667 return ret; 1668 } 1669 1670 int ocfs2_write_begin_nolock(struct file *filp, 1671 struct address_space *mapping, 1672 loff_t pos, unsigned len, unsigned flags, 1673 struct page **pagep, void **fsdata, 1674 struct buffer_head *di_bh, struct page *mmap_page) 1675 { 1676 int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS; 1677 unsigned int clusters_to_alloc, extents_to_split, clusters_need = 0; 1678 struct ocfs2_write_ctxt *wc; 1679 struct inode *inode = mapping->host; 1680 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 1681 struct ocfs2_dinode *di; 1682 struct ocfs2_alloc_context *data_ac = NULL; 1683 struct ocfs2_alloc_context *meta_ac = NULL; 1684 handle_t *handle; 1685 struct ocfs2_extent_tree et; 1686 int try_free = 1, ret1; 1687 1688 try_again: 1689 ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh); 1690 if (ret) { 1691 mlog_errno(ret); 1692 return ret; 1693 } 1694 1695 if (ocfs2_supports_inline_data(osb)) { 1696 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len, 1697 mmap_page, wc); 1698 if (ret == 1) { 1699 ret = 0; 1700 goto success; 1701 } 1702 if (ret < 0) { 1703 mlog_errno(ret); 1704 goto out; 1705 } 1706 } 1707 1708 if (ocfs2_sparse_alloc(osb)) 1709 ret = ocfs2_zero_tail(inode, di_bh, pos); 1710 else 1711 ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos, len, 1712 wc); 1713 if (ret) { 1714 mlog_errno(ret); 1715 goto out; 1716 } 1717 1718 ret = ocfs2_check_range_for_refcount(inode, pos, len); 1719 if (ret < 0) { 1720 mlog_errno(ret); 1721 goto out; 1722 } else if (ret == 1) { 1723 clusters_need = wc->w_clen; 1724 ret = ocfs2_refcount_cow(inode, filp, di_bh, 1725 wc->w_cpos, wc->w_clen, UINT_MAX); 1726 if (ret) { 1727 mlog_errno(ret); 1728 goto out; 1729 } 1730 } 1731 1732 ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc, 1733 &extents_to_split); 1734 if (ret) { 1735 mlog_errno(ret); 1736 goto out; 1737 } 1738 clusters_need += clusters_to_alloc; 1739 1740 di = (struct ocfs2_dinode *)wc->w_di_bh->b_data; 1741 1742 /* 1743 * We set w_target_from, w_target_to here so that 1744 * ocfs2_write_end() knows which range in the target page to 1745 * write out. An allocation requires that we write the entire 1746 * cluster range. 1747 */ 1748 if (clusters_to_alloc || extents_to_split) { 1749 /* 1750 * XXX: We are stretching the limits of 1751 * ocfs2_lock_allocators(). It greatly over-estimates 1752 * the work to be done. 1753 */ 1754 mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u," 1755 " clusters_to_add = %u, extents_to_split = %u\n", 1756 (unsigned long long)OCFS2_I(inode)->ip_blkno, 1757 (long long)i_size_read(inode), le32_to_cpu(di->i_clusters), 1758 clusters_to_alloc, extents_to_split); 1759 1760 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), 1761 wc->w_di_bh); 1762 ret = ocfs2_lock_allocators(inode, &et, 1763 clusters_to_alloc, extents_to_split, 1764 &data_ac, &meta_ac); 1765 if (ret) { 1766 mlog_errno(ret); 1767 goto out; 1768 } 1769 1770 if (data_ac) 1771 data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv; 1772 1773 credits = ocfs2_calc_extend_credits(inode->i_sb, 1774 &di->id2.i_list, 1775 clusters_to_alloc); 1776 1777 } 1778 1779 /* 1780 * We have to zero sparse allocated clusters, unwritten extent clusters, 1781 * and non-sparse clusters we just extended. For non-sparse writes, 1782 * we know zeros will only be needed in the first and/or last cluster. 1783 */ 1784 if (clusters_to_alloc || extents_to_split || 1785 (wc->w_clen && (wc->w_desc[0].c_needs_zero || 1786 wc->w_desc[wc->w_clen - 1].c_needs_zero))) 1787 cluster_of_pages = 1; 1788 else 1789 cluster_of_pages = 0; 1790 1791 ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages); 1792 1793 handle = ocfs2_start_trans(osb, credits); 1794 if (IS_ERR(handle)) { 1795 ret = PTR_ERR(handle); 1796 mlog_errno(ret); 1797 goto out; 1798 } 1799 1800 wc->w_handle = handle; 1801 1802 if (clusters_to_alloc) { 1803 ret = dquot_alloc_space_nodirty(inode, 1804 ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc)); 1805 if (ret) 1806 goto out_commit; 1807 } 1808 /* 1809 * We don't want this to fail in ocfs2_write_end(), so do it 1810 * here. 1811 */ 1812 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh, 1813 OCFS2_JOURNAL_ACCESS_WRITE); 1814 if (ret) { 1815 mlog_errno(ret); 1816 goto out_quota; 1817 } 1818 1819 /* 1820 * Fill our page array first. That way we've grabbed enough so 1821 * that we can zero and flush if we error after adding the 1822 * extent. 1823 */ 1824 ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len, 1825 cluster_of_pages, mmap_page); 1826 if (ret) { 1827 mlog_errno(ret); 1828 goto out_quota; 1829 } 1830 1831 ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos, 1832 len); 1833 if (ret) { 1834 mlog_errno(ret); 1835 goto out_quota; 1836 } 1837 1838 if (data_ac) 1839 ocfs2_free_alloc_context(data_ac); 1840 if (meta_ac) 1841 ocfs2_free_alloc_context(meta_ac); 1842 1843 success: 1844 *pagep = wc->w_target_page; 1845 *fsdata = wc; 1846 return 0; 1847 out_quota: 1848 if (clusters_to_alloc) 1849 dquot_free_space(inode, 1850 ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc)); 1851 out_commit: 1852 ocfs2_commit_trans(osb, handle); 1853 1854 out: 1855 ocfs2_free_write_ctxt(wc); 1856 1857 if (data_ac) 1858 ocfs2_free_alloc_context(data_ac); 1859 if (meta_ac) 1860 ocfs2_free_alloc_context(meta_ac); 1861 1862 if (ret == -ENOSPC && try_free) { 1863 /* 1864 * Try to free some truncate log so that we can have enough 1865 * clusters to allocate. 1866 */ 1867 try_free = 0; 1868 1869 ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need); 1870 if (ret1 == 1) 1871 goto try_again; 1872 1873 if (ret1 < 0) 1874 mlog_errno(ret1); 1875 } 1876 1877 return ret; 1878 } 1879 1880 static int ocfs2_write_begin(struct file *file, struct address_space *mapping, 1881 loff_t pos, unsigned len, unsigned flags, 1882 struct page **pagep, void **fsdata) 1883 { 1884 int ret; 1885 struct buffer_head *di_bh = NULL; 1886 struct inode *inode = mapping->host; 1887 1888 ret = ocfs2_inode_lock(inode, &di_bh, 1); 1889 if (ret) { 1890 mlog_errno(ret); 1891 return ret; 1892 } 1893 1894 /* 1895 * Take alloc sem here to prevent concurrent lookups. That way 1896 * the mapping, zeroing and tree manipulation within 1897 * ocfs2_write() will be safe against ->readpage(). This 1898 * should also serve to lock out allocation from a shared 1899 * writeable region. 1900 */ 1901 down_write(&OCFS2_I(inode)->ip_alloc_sem); 1902 1903 ret = ocfs2_write_begin_nolock(file, mapping, pos, len, flags, pagep, 1904 fsdata, di_bh, NULL); 1905 if (ret) { 1906 mlog_errno(ret); 1907 goto out_fail; 1908 } 1909 1910 brelse(di_bh); 1911 1912 return 0; 1913 1914 out_fail: 1915 up_write(&OCFS2_I(inode)->ip_alloc_sem); 1916 1917 brelse(di_bh); 1918 ocfs2_inode_unlock(inode, 1); 1919 1920 return ret; 1921 } 1922 1923 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos, 1924 unsigned len, unsigned *copied, 1925 struct ocfs2_dinode *di, 1926 struct ocfs2_write_ctxt *wc) 1927 { 1928 void *kaddr; 1929 1930 if (unlikely(*copied < len)) { 1931 if (!PageUptodate(wc->w_target_page)) { 1932 *copied = 0; 1933 return; 1934 } 1935 } 1936 1937 kaddr = kmap_atomic(wc->w_target_page, KM_USER0); 1938 memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied); 1939 kunmap_atomic(kaddr, KM_USER0); 1940 1941 mlog(0, "Data written to inode at offset %llu. " 1942 "id_count = %u, copied = %u, i_dyn_features = 0x%x\n", 1943 (unsigned long long)pos, *copied, 1944 le16_to_cpu(di->id2.i_data.id_count), 1945 le16_to_cpu(di->i_dyn_features)); 1946 } 1947 1948 int ocfs2_write_end_nolock(struct address_space *mapping, 1949 loff_t pos, unsigned len, unsigned copied, 1950 struct page *page, void *fsdata) 1951 { 1952 int i; 1953 unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1); 1954 struct inode *inode = mapping->host; 1955 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 1956 struct ocfs2_write_ctxt *wc = fsdata; 1957 struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data; 1958 handle_t *handle = wc->w_handle; 1959 struct page *tmppage; 1960 1961 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) { 1962 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc); 1963 goto out_write_size; 1964 } 1965 1966 if (unlikely(copied < len)) { 1967 if (!PageUptodate(wc->w_target_page)) 1968 copied = 0; 1969 1970 ocfs2_zero_new_buffers(wc->w_target_page, start+copied, 1971 start+len); 1972 } 1973 flush_dcache_page(wc->w_target_page); 1974 1975 for(i = 0; i < wc->w_num_pages; i++) { 1976 tmppage = wc->w_pages[i]; 1977 1978 if (tmppage == wc->w_target_page) { 1979 from = wc->w_target_from; 1980 to = wc->w_target_to; 1981 1982 BUG_ON(from > PAGE_CACHE_SIZE || 1983 to > PAGE_CACHE_SIZE || 1984 to < from); 1985 } else { 1986 /* 1987 * Pages adjacent to the target (if any) imply 1988 * a hole-filling write in which case we want 1989 * to flush their entire range. 1990 */ 1991 from = 0; 1992 to = PAGE_CACHE_SIZE; 1993 } 1994 1995 if (page_has_buffers(tmppage)) { 1996 if (ocfs2_should_order_data(inode)) 1997 ocfs2_jbd2_file_inode(wc->w_handle, inode); 1998 block_commit_write(tmppage, from, to); 1999 } 2000 } 2001 2002 out_write_size: 2003 pos += copied; 2004 if (pos > inode->i_size) { 2005 i_size_write(inode, pos); 2006 mark_inode_dirty(inode); 2007 } 2008 inode->i_blocks = ocfs2_inode_sector_count(inode); 2009 di->i_size = cpu_to_le64((u64)i_size_read(inode)); 2010 inode->i_mtime = inode->i_ctime = CURRENT_TIME; 2011 di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec); 2012 di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec); 2013 ocfs2_journal_dirty(handle, wc->w_di_bh); 2014 2015 ocfs2_commit_trans(osb, handle); 2016 2017 ocfs2_run_deallocs(osb, &wc->w_dealloc); 2018 2019 ocfs2_free_write_ctxt(wc); 2020 2021 return copied; 2022 } 2023 2024 static int ocfs2_write_end(struct file *file, struct address_space *mapping, 2025 loff_t pos, unsigned len, unsigned copied, 2026 struct page *page, void *fsdata) 2027 { 2028 int ret; 2029 struct inode *inode = mapping->host; 2030 2031 ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata); 2032 2033 up_write(&OCFS2_I(inode)->ip_alloc_sem); 2034 ocfs2_inode_unlock(inode, 1); 2035 2036 return ret; 2037 } 2038 2039 const struct address_space_operations ocfs2_aops = { 2040 .readpage = ocfs2_readpage, 2041 .readpages = ocfs2_readpages, 2042 .writepage = ocfs2_writepage, 2043 .write_begin = ocfs2_write_begin, 2044 .write_end = ocfs2_write_end, 2045 .bmap = ocfs2_bmap, 2046 .sync_page = block_sync_page, 2047 .direct_IO = ocfs2_direct_IO, 2048 .invalidatepage = ocfs2_invalidatepage, 2049 .releasepage = ocfs2_releasepage, 2050 .migratepage = buffer_migrate_page, 2051 .is_partially_uptodate = block_is_partially_uptodate, 2052 .error_remove_page = generic_error_remove_page, 2053 }; 2054