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 #include <cluster/masklog.h> 33 34 #include "ocfs2.h" 35 36 #include "alloc.h" 37 #include "aops.h" 38 #include "dlmglue.h" 39 #include "extent_map.h" 40 #include "file.h" 41 #include "inode.h" 42 #include "journal.h" 43 #include "suballoc.h" 44 #include "super.h" 45 #include "symlink.h" 46 #include "refcounttree.h" 47 #include "ocfs2_trace.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 trace_ocfs2_symlink_get_block( 63 (unsigned long long)OCFS2_I(inode)->ip_blkno, 64 (unsigned long long)iblock, bh_result, create); 65 66 BUG_ON(ocfs2_inode_is_fast_symlink(inode)); 67 68 if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) { 69 mlog(ML_ERROR, "block offset > PATH_MAX: %llu", 70 (unsigned long long)iblock); 71 goto bail; 72 } 73 74 status = ocfs2_read_inode_block(inode, &bh); 75 if (status < 0) { 76 mlog_errno(status); 77 goto bail; 78 } 79 fe = (struct ocfs2_dinode *) bh->b_data; 80 81 if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb, 82 le32_to_cpu(fe->i_clusters))) { 83 mlog(ML_ERROR, "block offset is outside the allocated size: " 84 "%llu\n", (unsigned long long)iblock); 85 goto bail; 86 } 87 88 /* We don't use the page cache to create symlink data, so if 89 * need be, copy it over from the buffer cache. */ 90 if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) { 91 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + 92 iblock; 93 buffer_cache_bh = sb_getblk(osb->sb, blkno); 94 if (!buffer_cache_bh) { 95 mlog(ML_ERROR, "couldn't getblock for symlink!\n"); 96 goto bail; 97 } 98 99 /* we haven't locked out transactions, so a commit 100 * could've happened. Since we've got a reference on 101 * the bh, even if it commits while we're doing the 102 * copy, the data is still good. */ 103 if (buffer_jbd(buffer_cache_bh) 104 && ocfs2_inode_is_new(inode)) { 105 kaddr = kmap_atomic(bh_result->b_page, KM_USER0); 106 if (!kaddr) { 107 mlog(ML_ERROR, "couldn't kmap!\n"); 108 goto bail; 109 } 110 memcpy(kaddr + (bh_result->b_size * iblock), 111 buffer_cache_bh->b_data, 112 bh_result->b_size); 113 kunmap_atomic(kaddr, KM_USER0); 114 set_buffer_uptodate(bh_result); 115 } 116 brelse(buffer_cache_bh); 117 } 118 119 map_bh(bh_result, inode->i_sb, 120 le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock); 121 122 err = 0; 123 124 bail: 125 brelse(bh); 126 127 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 trace_ocfs2_get_block((unsigned long long)OCFS2_I(inode)->ip_blkno, 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)); 202 203 trace_ocfs2_get_block_end((unsigned long long)OCFS2_I(inode)->ip_blkno, 204 (unsigned long long)past_eof); 205 if (create && (iblock >= past_eof)) 206 set_buffer_new(bh_result); 207 208 bail: 209 if (err < 0) 210 err = -EIO; 211 212 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 trace_ocfs2_readpage((unsigned long long)oi->ip_blkno, 282 (page ? page->index : 0)); 283 284 ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page); 285 if (ret != 0) { 286 if (ret == AOP_TRUNCATED_PAGE) 287 unlock = 0; 288 mlog_errno(ret); 289 goto out; 290 } 291 292 if (down_read_trylock(&oi->ip_alloc_sem) == 0) { 293 ret = AOP_TRUNCATED_PAGE; 294 goto out_inode_unlock; 295 } 296 297 /* 298 * i_size might have just been updated as we grabed the meta lock. We 299 * might now be discovering a truncate that hit on another node. 300 * block_read_full_page->get_block freaks out if it is asked to read 301 * beyond the end of a file, so we check here. Callers 302 * (generic_file_read, vm_ops->fault) are clever enough to check i_size 303 * and notice that the page they just read isn't needed. 304 * 305 * XXX sys_readahead() seems to get that wrong? 306 */ 307 if (start >= i_size_read(inode)) { 308 zero_user(page, 0, PAGE_SIZE); 309 SetPageUptodate(page); 310 ret = 0; 311 goto out_alloc; 312 } 313 314 if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) 315 ret = ocfs2_readpage_inline(inode, page); 316 else 317 ret = block_read_full_page(page, ocfs2_get_block); 318 unlock = 0; 319 320 out_alloc: 321 up_read(&OCFS2_I(inode)->ip_alloc_sem); 322 out_inode_unlock: 323 ocfs2_inode_unlock(inode, 0); 324 out: 325 if (unlock) 326 unlock_page(page); 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 trace_ocfs2_writepage( 400 (unsigned long long)OCFS2_I(page->mapping->host)->ip_blkno, 401 page->index); 402 403 return block_write_full_page(page, ocfs2_get_block, wbc); 404 } 405 406 /* Taken from ext3. We don't necessarily need the full blown 407 * functionality yet, but IMHO it's better to cut and paste the whole 408 * thing so we can avoid introducing our own bugs (and easily pick up 409 * their fixes when they happen) --Mark */ 410 int walk_page_buffers( handle_t *handle, 411 struct buffer_head *head, 412 unsigned from, 413 unsigned to, 414 int *partial, 415 int (*fn)( handle_t *handle, 416 struct buffer_head *bh)) 417 { 418 struct buffer_head *bh; 419 unsigned block_start, block_end; 420 unsigned blocksize = head->b_size; 421 int err, ret = 0; 422 struct buffer_head *next; 423 424 for ( bh = head, block_start = 0; 425 ret == 0 && (bh != head || !block_start); 426 block_start = block_end, bh = next) 427 { 428 next = bh->b_this_page; 429 block_end = block_start + blocksize; 430 if (block_end <= from || block_start >= to) { 431 if (partial && !buffer_uptodate(bh)) 432 *partial = 1; 433 continue; 434 } 435 err = (*fn)(handle, bh); 436 if (!ret) 437 ret = err; 438 } 439 return ret; 440 } 441 442 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block) 443 { 444 sector_t status; 445 u64 p_blkno = 0; 446 int err = 0; 447 struct inode *inode = mapping->host; 448 449 trace_ocfs2_bmap((unsigned long long)OCFS2_I(inode)->ip_blkno, 450 (unsigned long long)block); 451 452 /* We don't need to lock journal system files, since they aren't 453 * accessed concurrently from multiple nodes. 454 */ 455 if (!INODE_JOURNAL(inode)) { 456 err = ocfs2_inode_lock(inode, NULL, 0); 457 if (err) { 458 if (err != -ENOENT) 459 mlog_errno(err); 460 goto bail; 461 } 462 down_read(&OCFS2_I(inode)->ip_alloc_sem); 463 } 464 465 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)) 466 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL, 467 NULL); 468 469 if (!INODE_JOURNAL(inode)) { 470 up_read(&OCFS2_I(inode)->ip_alloc_sem); 471 ocfs2_inode_unlock(inode, 0); 472 } 473 474 if (err) { 475 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n", 476 (unsigned long long)block); 477 mlog_errno(err); 478 goto bail; 479 } 480 481 bail: 482 status = err ? 0 : p_blkno; 483 484 return status; 485 } 486 487 /* 488 * TODO: Make this into a generic get_blocks function. 489 * 490 * From do_direct_io in direct-io.c: 491 * "So what we do is to permit the ->get_blocks function to populate 492 * bh.b_size with the size of IO which is permitted at this offset and 493 * this i_blkbits." 494 * 495 * This function is called directly from get_more_blocks in direct-io.c. 496 * 497 * called like this: dio->get_blocks(dio->inode, fs_startblk, 498 * fs_count, map_bh, dio->rw == WRITE); 499 * 500 * Note that we never bother to allocate blocks here, and thus ignore the 501 * create argument. 502 */ 503 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock, 504 struct buffer_head *bh_result, int create) 505 { 506 int ret; 507 u64 p_blkno, inode_blocks, contig_blocks; 508 unsigned int ext_flags; 509 unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits; 510 unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits; 511 512 /* This function won't even be called if the request isn't all 513 * nicely aligned and of the right size, so there's no need 514 * for us to check any of that. */ 515 516 inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode)); 517 518 /* This figures out the size of the next contiguous block, and 519 * our logical offset */ 520 ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, 521 &contig_blocks, &ext_flags); 522 if (ret) { 523 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n", 524 (unsigned long long)iblock); 525 ret = -EIO; 526 goto bail; 527 } 528 529 /* We should already CoW the refcounted extent in case of create. */ 530 BUG_ON(create && (ext_flags & OCFS2_EXT_REFCOUNTED)); 531 532 /* 533 * get_more_blocks() expects us to describe a hole by clearing 534 * the mapped bit on bh_result(). 535 * 536 * Consider an unwritten extent as a hole. 537 */ 538 if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN)) 539 map_bh(bh_result, inode->i_sb, p_blkno); 540 else 541 clear_buffer_mapped(bh_result); 542 543 /* make sure we don't map more than max_blocks blocks here as 544 that's all the kernel will handle at this point. */ 545 if (max_blocks < contig_blocks) 546 contig_blocks = max_blocks; 547 bh_result->b_size = contig_blocks << blocksize_bits; 548 bail: 549 return ret; 550 } 551 552 /* 553 * ocfs2_dio_end_io is called by the dio core when a dio is finished. We're 554 * particularly interested in the aio/dio case. We use the rw_lock DLM lock 555 * to protect io on one node from truncation on another. 556 */ 557 static void ocfs2_dio_end_io(struct kiocb *iocb, 558 loff_t offset, 559 ssize_t bytes, 560 void *private, 561 int ret, 562 bool is_async) 563 { 564 struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode; 565 int level; 566 567 /* this io's submitter should not have unlocked this before we could */ 568 BUG_ON(!ocfs2_iocb_is_rw_locked(iocb)); 569 570 if (ocfs2_iocb_is_sem_locked(iocb)) 571 ocfs2_iocb_clear_sem_locked(iocb); 572 573 ocfs2_iocb_clear_rw_locked(iocb); 574 575 level = ocfs2_iocb_rw_locked_level(iocb); 576 ocfs2_rw_unlock(inode, level); 577 578 if (is_async) 579 aio_complete(iocb, ret, 0); 580 inode_dio_done(inode); 581 } 582 583 /* 584 * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen 585 * from ext3. PageChecked() bits have been removed as OCFS2 does not 586 * do journalled data. 587 */ 588 static void ocfs2_invalidatepage(struct page *page, unsigned long offset) 589 { 590 journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal; 591 592 jbd2_journal_invalidatepage(journal, page, offset); 593 } 594 595 static int ocfs2_releasepage(struct page *page, gfp_t wait) 596 { 597 journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal; 598 599 if (!page_has_buffers(page)) 600 return 0; 601 return jbd2_journal_try_to_free_buffers(journal, page, wait); 602 } 603 604 static ssize_t ocfs2_direct_IO(int rw, 605 struct kiocb *iocb, 606 const struct iovec *iov, 607 loff_t offset, 608 unsigned long nr_segs) 609 { 610 struct file *file = iocb->ki_filp; 611 struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host; 612 613 /* 614 * Fallback to buffered I/O if we see an inode without 615 * extents. 616 */ 617 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) 618 return 0; 619 620 /* Fallback to buffered I/O if we are appending. */ 621 if (i_size_read(inode) <= offset) 622 return 0; 623 624 return __blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, 625 iov, offset, nr_segs, 626 ocfs2_direct_IO_get_blocks, 627 ocfs2_dio_end_io, NULL, 0); 628 } 629 630 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb, 631 u32 cpos, 632 unsigned int *start, 633 unsigned int *end) 634 { 635 unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE; 636 637 if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) { 638 unsigned int cpp; 639 640 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits); 641 642 cluster_start = cpos % cpp; 643 cluster_start = cluster_start << osb->s_clustersize_bits; 644 645 cluster_end = cluster_start + osb->s_clustersize; 646 } 647 648 BUG_ON(cluster_start > PAGE_SIZE); 649 BUG_ON(cluster_end > PAGE_SIZE); 650 651 if (start) 652 *start = cluster_start; 653 if (end) 654 *end = cluster_end; 655 } 656 657 /* 658 * 'from' and 'to' are the region in the page to avoid zeroing. 659 * 660 * If pagesize > clustersize, this function will avoid zeroing outside 661 * of the cluster boundary. 662 * 663 * from == to == 0 is code for "zero the entire cluster region" 664 */ 665 static void ocfs2_clear_page_regions(struct page *page, 666 struct ocfs2_super *osb, u32 cpos, 667 unsigned from, unsigned to) 668 { 669 void *kaddr; 670 unsigned int cluster_start, cluster_end; 671 672 ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end); 673 674 kaddr = kmap_atomic(page, KM_USER0); 675 676 if (from || to) { 677 if (from > cluster_start) 678 memset(kaddr + cluster_start, 0, from - cluster_start); 679 if (to < cluster_end) 680 memset(kaddr + to, 0, cluster_end - to); 681 } else { 682 memset(kaddr + cluster_start, 0, cluster_end - cluster_start); 683 } 684 685 kunmap_atomic(kaddr, KM_USER0); 686 } 687 688 /* 689 * Nonsparse file systems fully allocate before we get to the write 690 * code. This prevents ocfs2_write() from tagging the write as an 691 * allocating one, which means ocfs2_map_page_blocks() might try to 692 * read-in the blocks at the tail of our file. Avoid reading them by 693 * testing i_size against each block offset. 694 */ 695 static int ocfs2_should_read_blk(struct inode *inode, struct page *page, 696 unsigned int block_start) 697 { 698 u64 offset = page_offset(page) + block_start; 699 700 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) 701 return 1; 702 703 if (i_size_read(inode) > offset) 704 return 1; 705 706 return 0; 707 } 708 709 /* 710 * Some of this taken from __block_write_begin(). We already have our 711 * mapping by now though, and the entire write will be allocating or 712 * it won't, so not much need to use BH_New. 713 * 714 * This will also skip zeroing, which is handled externally. 715 */ 716 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno, 717 struct inode *inode, unsigned int from, 718 unsigned int to, int new) 719 { 720 int ret = 0; 721 struct buffer_head *head, *bh, *wait[2], **wait_bh = wait; 722 unsigned int block_end, block_start; 723 unsigned int bsize = 1 << inode->i_blkbits; 724 725 if (!page_has_buffers(page)) 726 create_empty_buffers(page, bsize, 0); 727 728 head = page_buffers(page); 729 for (bh = head, block_start = 0; bh != head || !block_start; 730 bh = bh->b_this_page, block_start += bsize) { 731 block_end = block_start + bsize; 732 733 clear_buffer_new(bh); 734 735 /* 736 * Ignore blocks outside of our i/o range - 737 * they may belong to unallocated clusters. 738 */ 739 if (block_start >= to || block_end <= from) { 740 if (PageUptodate(page)) 741 set_buffer_uptodate(bh); 742 continue; 743 } 744 745 /* 746 * For an allocating write with cluster size >= page 747 * size, we always write the entire page. 748 */ 749 if (new) 750 set_buffer_new(bh); 751 752 if (!buffer_mapped(bh)) { 753 map_bh(bh, inode->i_sb, *p_blkno); 754 unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr); 755 } 756 757 if (PageUptodate(page)) { 758 if (!buffer_uptodate(bh)) 759 set_buffer_uptodate(bh); 760 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) && 761 !buffer_new(bh) && 762 ocfs2_should_read_blk(inode, page, block_start) && 763 (block_start < from || block_end > to)) { 764 ll_rw_block(READ, 1, &bh); 765 *wait_bh++=bh; 766 } 767 768 *p_blkno = *p_blkno + 1; 769 } 770 771 /* 772 * If we issued read requests - let them complete. 773 */ 774 while(wait_bh > wait) { 775 wait_on_buffer(*--wait_bh); 776 if (!buffer_uptodate(*wait_bh)) 777 ret = -EIO; 778 } 779 780 if (ret == 0 || !new) 781 return ret; 782 783 /* 784 * If we get -EIO above, zero out any newly allocated blocks 785 * to avoid exposing stale data. 786 */ 787 bh = head; 788 block_start = 0; 789 do { 790 block_end = block_start + bsize; 791 if (block_end <= from) 792 goto next_bh; 793 if (block_start >= to) 794 break; 795 796 zero_user(page, block_start, bh->b_size); 797 set_buffer_uptodate(bh); 798 mark_buffer_dirty(bh); 799 800 next_bh: 801 block_start = block_end; 802 bh = bh->b_this_page; 803 } while (bh != head); 804 805 return ret; 806 } 807 808 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE) 809 #define OCFS2_MAX_CTXT_PAGES 1 810 #else 811 #define OCFS2_MAX_CTXT_PAGES (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE) 812 #endif 813 814 #define OCFS2_MAX_CLUSTERS_PER_PAGE (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE) 815 816 /* 817 * Describe the state of a single cluster to be written to. 818 */ 819 struct ocfs2_write_cluster_desc { 820 u32 c_cpos; 821 u32 c_phys; 822 /* 823 * Give this a unique field because c_phys eventually gets 824 * filled. 825 */ 826 unsigned c_new; 827 unsigned c_unwritten; 828 unsigned c_needs_zero; 829 }; 830 831 struct ocfs2_write_ctxt { 832 /* Logical cluster position / len of write */ 833 u32 w_cpos; 834 u32 w_clen; 835 836 /* First cluster allocated in a nonsparse extend */ 837 u32 w_first_new_cpos; 838 839 struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE]; 840 841 /* 842 * This is true if page_size > cluster_size. 843 * 844 * It triggers a set of special cases during write which might 845 * have to deal with allocating writes to partial pages. 846 */ 847 unsigned int w_large_pages; 848 849 /* 850 * Pages involved in this write. 851 * 852 * w_target_page is the page being written to by the user. 853 * 854 * w_pages is an array of pages which always contains 855 * w_target_page, and in the case of an allocating write with 856 * page_size < cluster size, it will contain zero'd and mapped 857 * pages adjacent to w_target_page which need to be written 858 * out in so that future reads from that region will get 859 * zero's. 860 */ 861 unsigned int w_num_pages; 862 struct page *w_pages[OCFS2_MAX_CTXT_PAGES]; 863 struct page *w_target_page; 864 865 /* 866 * ocfs2_write_end() uses this to know what the real range to 867 * write in the target should be. 868 */ 869 unsigned int w_target_from; 870 unsigned int w_target_to; 871 872 /* 873 * We could use journal_current_handle() but this is cleaner, 874 * IMHO -Mark 875 */ 876 handle_t *w_handle; 877 878 struct buffer_head *w_di_bh; 879 880 struct ocfs2_cached_dealloc_ctxt w_dealloc; 881 }; 882 883 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages) 884 { 885 int i; 886 887 for(i = 0; i < num_pages; i++) { 888 if (pages[i]) { 889 unlock_page(pages[i]); 890 mark_page_accessed(pages[i]); 891 page_cache_release(pages[i]); 892 } 893 } 894 } 895 896 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc) 897 { 898 ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages); 899 900 brelse(wc->w_di_bh); 901 kfree(wc); 902 } 903 904 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp, 905 struct ocfs2_super *osb, loff_t pos, 906 unsigned len, struct buffer_head *di_bh) 907 { 908 u32 cend; 909 struct ocfs2_write_ctxt *wc; 910 911 wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS); 912 if (!wc) 913 return -ENOMEM; 914 915 wc->w_cpos = pos >> osb->s_clustersize_bits; 916 wc->w_first_new_cpos = UINT_MAX; 917 cend = (pos + len - 1) >> osb->s_clustersize_bits; 918 wc->w_clen = cend - wc->w_cpos + 1; 919 get_bh(di_bh); 920 wc->w_di_bh = di_bh; 921 922 if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) 923 wc->w_large_pages = 1; 924 else 925 wc->w_large_pages = 0; 926 927 ocfs2_init_dealloc_ctxt(&wc->w_dealloc); 928 929 *wcp = wc; 930 931 return 0; 932 } 933 934 /* 935 * If a page has any new buffers, zero them out here, and mark them uptodate 936 * and dirty so they'll be written out (in order to prevent uninitialised 937 * block data from leaking). And clear the new bit. 938 */ 939 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to) 940 { 941 unsigned int block_start, block_end; 942 struct buffer_head *head, *bh; 943 944 BUG_ON(!PageLocked(page)); 945 if (!page_has_buffers(page)) 946 return; 947 948 bh = head = page_buffers(page); 949 block_start = 0; 950 do { 951 block_end = block_start + bh->b_size; 952 953 if (buffer_new(bh)) { 954 if (block_end > from && block_start < to) { 955 if (!PageUptodate(page)) { 956 unsigned start, end; 957 958 start = max(from, block_start); 959 end = min(to, block_end); 960 961 zero_user_segment(page, start, end); 962 set_buffer_uptodate(bh); 963 } 964 965 clear_buffer_new(bh); 966 mark_buffer_dirty(bh); 967 } 968 } 969 970 block_start = block_end; 971 bh = bh->b_this_page; 972 } while (bh != head); 973 } 974 975 /* 976 * Only called when we have a failure during allocating write to write 977 * zero's to the newly allocated region. 978 */ 979 static void ocfs2_write_failure(struct inode *inode, 980 struct ocfs2_write_ctxt *wc, 981 loff_t user_pos, unsigned user_len) 982 { 983 int i; 984 unsigned from = user_pos & (PAGE_CACHE_SIZE - 1), 985 to = user_pos + user_len; 986 struct page *tmppage; 987 988 ocfs2_zero_new_buffers(wc->w_target_page, from, to); 989 990 for(i = 0; i < wc->w_num_pages; i++) { 991 tmppage = wc->w_pages[i]; 992 993 if (page_has_buffers(tmppage)) { 994 if (ocfs2_should_order_data(inode)) 995 ocfs2_jbd2_file_inode(wc->w_handle, inode); 996 997 block_commit_write(tmppage, from, to); 998 } 999 } 1000 } 1001 1002 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno, 1003 struct ocfs2_write_ctxt *wc, 1004 struct page *page, u32 cpos, 1005 loff_t user_pos, unsigned user_len, 1006 int new) 1007 { 1008 int ret; 1009 unsigned int map_from = 0, map_to = 0; 1010 unsigned int cluster_start, cluster_end; 1011 unsigned int user_data_from = 0, user_data_to = 0; 1012 1013 ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos, 1014 &cluster_start, &cluster_end); 1015 1016 /* treat the write as new if the a hole/lseek spanned across 1017 * the page boundary. 1018 */ 1019 new = new | ((i_size_read(inode) <= page_offset(page)) && 1020 (page_offset(page) <= user_pos)); 1021 1022 if (page == wc->w_target_page) { 1023 map_from = user_pos & (PAGE_CACHE_SIZE - 1); 1024 map_to = map_from + user_len; 1025 1026 if (new) 1027 ret = ocfs2_map_page_blocks(page, p_blkno, inode, 1028 cluster_start, cluster_end, 1029 new); 1030 else 1031 ret = ocfs2_map_page_blocks(page, p_blkno, inode, 1032 map_from, map_to, new); 1033 if (ret) { 1034 mlog_errno(ret); 1035 goto out; 1036 } 1037 1038 user_data_from = map_from; 1039 user_data_to = map_to; 1040 if (new) { 1041 map_from = cluster_start; 1042 map_to = cluster_end; 1043 } 1044 } else { 1045 /* 1046 * If we haven't allocated the new page yet, we 1047 * shouldn't be writing it out without copying user 1048 * data. This is likely a math error from the caller. 1049 */ 1050 BUG_ON(!new); 1051 1052 map_from = cluster_start; 1053 map_to = cluster_end; 1054 1055 ret = ocfs2_map_page_blocks(page, p_blkno, inode, 1056 cluster_start, cluster_end, new); 1057 if (ret) { 1058 mlog_errno(ret); 1059 goto out; 1060 } 1061 } 1062 1063 /* 1064 * Parts of newly allocated pages need to be zero'd. 1065 * 1066 * Above, we have also rewritten 'to' and 'from' - as far as 1067 * the rest of the function is concerned, the entire cluster 1068 * range inside of a page needs to be written. 1069 * 1070 * We can skip this if the page is up to date - it's already 1071 * been zero'd from being read in as a hole. 1072 */ 1073 if (new && !PageUptodate(page)) 1074 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb), 1075 cpos, user_data_from, user_data_to); 1076 1077 flush_dcache_page(page); 1078 1079 out: 1080 return ret; 1081 } 1082 1083 /* 1084 * This function will only grab one clusters worth of pages. 1085 */ 1086 static int ocfs2_grab_pages_for_write(struct address_space *mapping, 1087 struct ocfs2_write_ctxt *wc, 1088 u32 cpos, loff_t user_pos, 1089 unsigned user_len, int new, 1090 struct page *mmap_page) 1091 { 1092 int ret = 0, i; 1093 unsigned long start, target_index, end_index, index; 1094 struct inode *inode = mapping->host; 1095 loff_t last_byte; 1096 1097 target_index = user_pos >> PAGE_CACHE_SHIFT; 1098 1099 /* 1100 * Figure out how many pages we'll be manipulating here. For 1101 * non allocating write, we just change the one 1102 * page. Otherwise, we'll need a whole clusters worth. If we're 1103 * writing past i_size, we only need enough pages to cover the 1104 * last page of the write. 1105 */ 1106 if (new) { 1107 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb); 1108 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos); 1109 /* 1110 * We need the index *past* the last page we could possibly 1111 * touch. This is the page past the end of the write or 1112 * i_size, whichever is greater. 1113 */ 1114 last_byte = max(user_pos + user_len, i_size_read(inode)); 1115 BUG_ON(last_byte < 1); 1116 end_index = ((last_byte - 1) >> PAGE_CACHE_SHIFT) + 1; 1117 if ((start + wc->w_num_pages) > end_index) 1118 wc->w_num_pages = end_index - start; 1119 } else { 1120 wc->w_num_pages = 1; 1121 start = target_index; 1122 } 1123 1124 for(i = 0; i < wc->w_num_pages; i++) { 1125 index = start + i; 1126 1127 if (index == target_index && mmap_page) { 1128 /* 1129 * ocfs2_pagemkwrite() is a little different 1130 * and wants us to directly use the page 1131 * passed in. 1132 */ 1133 lock_page(mmap_page); 1134 1135 if (mmap_page->mapping != mapping) { 1136 unlock_page(mmap_page); 1137 /* 1138 * Sanity check - the locking in 1139 * ocfs2_pagemkwrite() should ensure 1140 * that this code doesn't trigger. 1141 */ 1142 ret = -EINVAL; 1143 mlog_errno(ret); 1144 goto out; 1145 } 1146 1147 page_cache_get(mmap_page); 1148 wc->w_pages[i] = mmap_page; 1149 } else { 1150 wc->w_pages[i] = find_or_create_page(mapping, index, 1151 GFP_NOFS); 1152 if (!wc->w_pages[i]) { 1153 ret = -ENOMEM; 1154 mlog_errno(ret); 1155 goto out; 1156 } 1157 } 1158 1159 if (index == target_index) 1160 wc->w_target_page = wc->w_pages[i]; 1161 } 1162 out: 1163 return ret; 1164 } 1165 1166 /* 1167 * Prepare a single cluster for write one cluster into the file. 1168 */ 1169 static int ocfs2_write_cluster(struct address_space *mapping, 1170 u32 phys, unsigned int unwritten, 1171 unsigned int should_zero, 1172 struct ocfs2_alloc_context *data_ac, 1173 struct ocfs2_alloc_context *meta_ac, 1174 struct ocfs2_write_ctxt *wc, u32 cpos, 1175 loff_t user_pos, unsigned user_len) 1176 { 1177 int ret, i, new; 1178 u64 v_blkno, p_blkno; 1179 struct inode *inode = mapping->host; 1180 struct ocfs2_extent_tree et; 1181 1182 new = phys == 0 ? 1 : 0; 1183 if (new) { 1184 u32 tmp_pos; 1185 1186 /* 1187 * This is safe to call with the page locks - it won't take 1188 * any additional semaphores or cluster locks. 1189 */ 1190 tmp_pos = cpos; 1191 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode, 1192 &tmp_pos, 1, 0, wc->w_di_bh, 1193 wc->w_handle, data_ac, 1194 meta_ac, NULL); 1195 /* 1196 * This shouldn't happen because we must have already 1197 * calculated the correct meta data allocation required. The 1198 * internal tree allocation code should know how to increase 1199 * transaction credits itself. 1200 * 1201 * If need be, we could handle -EAGAIN for a 1202 * RESTART_TRANS here. 1203 */ 1204 mlog_bug_on_msg(ret == -EAGAIN, 1205 "Inode %llu: EAGAIN return during allocation.\n", 1206 (unsigned long long)OCFS2_I(inode)->ip_blkno); 1207 if (ret < 0) { 1208 mlog_errno(ret); 1209 goto out; 1210 } 1211 } else if (unwritten) { 1212 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), 1213 wc->w_di_bh); 1214 ret = ocfs2_mark_extent_written(inode, &et, 1215 wc->w_handle, cpos, 1, phys, 1216 meta_ac, &wc->w_dealloc); 1217 if (ret < 0) { 1218 mlog_errno(ret); 1219 goto out; 1220 } 1221 } 1222 1223 if (should_zero) 1224 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos); 1225 else 1226 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits; 1227 1228 /* 1229 * The only reason this should fail is due to an inability to 1230 * find the extent added. 1231 */ 1232 ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL, 1233 NULL); 1234 if (ret < 0) { 1235 ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, " 1236 "at logical block %llu", 1237 (unsigned long long)OCFS2_I(inode)->ip_blkno, 1238 (unsigned long long)v_blkno); 1239 goto out; 1240 } 1241 1242 BUG_ON(p_blkno == 0); 1243 1244 for(i = 0; i < wc->w_num_pages; i++) { 1245 int tmpret; 1246 1247 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc, 1248 wc->w_pages[i], cpos, 1249 user_pos, user_len, 1250 should_zero); 1251 if (tmpret) { 1252 mlog_errno(tmpret); 1253 if (ret == 0) 1254 ret = tmpret; 1255 } 1256 } 1257 1258 /* 1259 * We only have cleanup to do in case of allocating write. 1260 */ 1261 if (ret && new) 1262 ocfs2_write_failure(inode, wc, user_pos, user_len); 1263 1264 out: 1265 1266 return ret; 1267 } 1268 1269 static int ocfs2_write_cluster_by_desc(struct address_space *mapping, 1270 struct ocfs2_alloc_context *data_ac, 1271 struct ocfs2_alloc_context *meta_ac, 1272 struct ocfs2_write_ctxt *wc, 1273 loff_t pos, unsigned len) 1274 { 1275 int ret, i; 1276 loff_t cluster_off; 1277 unsigned int local_len = len; 1278 struct ocfs2_write_cluster_desc *desc; 1279 struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb); 1280 1281 for (i = 0; i < wc->w_clen; i++) { 1282 desc = &wc->w_desc[i]; 1283 1284 /* 1285 * We have to make sure that the total write passed in 1286 * doesn't extend past a single cluster. 1287 */ 1288 local_len = len; 1289 cluster_off = pos & (osb->s_clustersize - 1); 1290 if ((cluster_off + local_len) > osb->s_clustersize) 1291 local_len = osb->s_clustersize - cluster_off; 1292 1293 ret = ocfs2_write_cluster(mapping, desc->c_phys, 1294 desc->c_unwritten, 1295 desc->c_needs_zero, 1296 data_ac, meta_ac, 1297 wc, desc->c_cpos, pos, local_len); 1298 if (ret) { 1299 mlog_errno(ret); 1300 goto out; 1301 } 1302 1303 len -= local_len; 1304 pos += local_len; 1305 } 1306 1307 ret = 0; 1308 out: 1309 return ret; 1310 } 1311 1312 /* 1313 * ocfs2_write_end() wants to know which parts of the target page it 1314 * should complete the write on. It's easiest to compute them ahead of 1315 * time when a more complete view of the write is available. 1316 */ 1317 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb, 1318 struct ocfs2_write_ctxt *wc, 1319 loff_t pos, unsigned len, int alloc) 1320 { 1321 struct ocfs2_write_cluster_desc *desc; 1322 1323 wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1); 1324 wc->w_target_to = wc->w_target_from + len; 1325 1326 if (alloc == 0) 1327 return; 1328 1329 /* 1330 * Allocating write - we may have different boundaries based 1331 * on page size and cluster size. 1332 * 1333 * NOTE: We can no longer compute one value from the other as 1334 * the actual write length and user provided length may be 1335 * different. 1336 */ 1337 1338 if (wc->w_large_pages) { 1339 /* 1340 * We only care about the 1st and last cluster within 1341 * our range and whether they should be zero'd or not. Either 1342 * value may be extended out to the start/end of a 1343 * newly allocated cluster. 1344 */ 1345 desc = &wc->w_desc[0]; 1346 if (desc->c_needs_zero) 1347 ocfs2_figure_cluster_boundaries(osb, 1348 desc->c_cpos, 1349 &wc->w_target_from, 1350 NULL); 1351 1352 desc = &wc->w_desc[wc->w_clen - 1]; 1353 if (desc->c_needs_zero) 1354 ocfs2_figure_cluster_boundaries(osb, 1355 desc->c_cpos, 1356 NULL, 1357 &wc->w_target_to); 1358 } else { 1359 wc->w_target_from = 0; 1360 wc->w_target_to = PAGE_CACHE_SIZE; 1361 } 1362 } 1363 1364 /* 1365 * Populate each single-cluster write descriptor in the write context 1366 * with information about the i/o to be done. 1367 * 1368 * Returns the number of clusters that will have to be allocated, as 1369 * well as a worst case estimate of the number of extent records that 1370 * would have to be created during a write to an unwritten region. 1371 */ 1372 static int ocfs2_populate_write_desc(struct inode *inode, 1373 struct ocfs2_write_ctxt *wc, 1374 unsigned int *clusters_to_alloc, 1375 unsigned int *extents_to_split) 1376 { 1377 int ret; 1378 struct ocfs2_write_cluster_desc *desc; 1379 unsigned int num_clusters = 0; 1380 unsigned int ext_flags = 0; 1381 u32 phys = 0; 1382 int i; 1383 1384 *clusters_to_alloc = 0; 1385 *extents_to_split = 0; 1386 1387 for (i = 0; i < wc->w_clen; i++) { 1388 desc = &wc->w_desc[i]; 1389 desc->c_cpos = wc->w_cpos + i; 1390 1391 if (num_clusters == 0) { 1392 /* 1393 * Need to look up the next extent record. 1394 */ 1395 ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys, 1396 &num_clusters, &ext_flags); 1397 if (ret) { 1398 mlog_errno(ret); 1399 goto out; 1400 } 1401 1402 /* We should already CoW the refcountd extent. */ 1403 BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED); 1404 1405 /* 1406 * Assume worst case - that we're writing in 1407 * the middle of the extent. 1408 * 1409 * We can assume that the write proceeds from 1410 * left to right, in which case the extent 1411 * insert code is smart enough to coalesce the 1412 * next splits into the previous records created. 1413 */ 1414 if (ext_flags & OCFS2_EXT_UNWRITTEN) 1415 *extents_to_split = *extents_to_split + 2; 1416 } else if (phys) { 1417 /* 1418 * Only increment phys if it doesn't describe 1419 * a hole. 1420 */ 1421 phys++; 1422 } 1423 1424 /* 1425 * If w_first_new_cpos is < UINT_MAX, we have a non-sparse 1426 * file that got extended. w_first_new_cpos tells us 1427 * where the newly allocated clusters are so we can 1428 * zero them. 1429 */ 1430 if (desc->c_cpos >= wc->w_first_new_cpos) { 1431 BUG_ON(phys == 0); 1432 desc->c_needs_zero = 1; 1433 } 1434 1435 desc->c_phys = phys; 1436 if (phys == 0) { 1437 desc->c_new = 1; 1438 desc->c_needs_zero = 1; 1439 *clusters_to_alloc = *clusters_to_alloc + 1; 1440 } 1441 1442 if (ext_flags & OCFS2_EXT_UNWRITTEN) { 1443 desc->c_unwritten = 1; 1444 desc->c_needs_zero = 1; 1445 } 1446 1447 num_clusters--; 1448 } 1449 1450 ret = 0; 1451 out: 1452 return ret; 1453 } 1454 1455 static int ocfs2_write_begin_inline(struct address_space *mapping, 1456 struct inode *inode, 1457 struct ocfs2_write_ctxt *wc) 1458 { 1459 int ret; 1460 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 1461 struct page *page; 1462 handle_t *handle; 1463 struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data; 1464 1465 page = find_or_create_page(mapping, 0, GFP_NOFS); 1466 if (!page) { 1467 ret = -ENOMEM; 1468 mlog_errno(ret); 1469 goto out; 1470 } 1471 /* 1472 * If we don't set w_num_pages then this page won't get unlocked 1473 * and freed on cleanup of the write context. 1474 */ 1475 wc->w_pages[0] = wc->w_target_page = page; 1476 wc->w_num_pages = 1; 1477 1478 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS); 1479 if (IS_ERR(handle)) { 1480 ret = PTR_ERR(handle); 1481 mlog_errno(ret); 1482 goto out; 1483 } 1484 1485 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh, 1486 OCFS2_JOURNAL_ACCESS_WRITE); 1487 if (ret) { 1488 ocfs2_commit_trans(osb, handle); 1489 1490 mlog_errno(ret); 1491 goto out; 1492 } 1493 1494 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)) 1495 ocfs2_set_inode_data_inline(inode, di); 1496 1497 if (!PageUptodate(page)) { 1498 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh); 1499 if (ret) { 1500 ocfs2_commit_trans(osb, handle); 1501 1502 goto out; 1503 } 1504 } 1505 1506 wc->w_handle = handle; 1507 out: 1508 return ret; 1509 } 1510 1511 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size) 1512 { 1513 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 1514 1515 if (new_size <= le16_to_cpu(di->id2.i_data.id_count)) 1516 return 1; 1517 return 0; 1518 } 1519 1520 static int ocfs2_try_to_write_inline_data(struct address_space *mapping, 1521 struct inode *inode, loff_t pos, 1522 unsigned len, struct page *mmap_page, 1523 struct ocfs2_write_ctxt *wc) 1524 { 1525 int ret, written = 0; 1526 loff_t end = pos + len; 1527 struct ocfs2_inode_info *oi = OCFS2_I(inode); 1528 struct ocfs2_dinode *di = NULL; 1529 1530 trace_ocfs2_try_to_write_inline_data((unsigned long long)oi->ip_blkno, 1531 len, (unsigned long long)pos, 1532 oi->ip_dyn_features); 1533 1534 /* 1535 * Handle inodes which already have inline data 1st. 1536 */ 1537 if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) { 1538 if (mmap_page == NULL && 1539 ocfs2_size_fits_inline_data(wc->w_di_bh, end)) 1540 goto do_inline_write; 1541 1542 /* 1543 * The write won't fit - we have to give this inode an 1544 * inline extent list now. 1545 */ 1546 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh); 1547 if (ret) 1548 mlog_errno(ret); 1549 goto out; 1550 } 1551 1552 /* 1553 * Check whether the inode can accept inline data. 1554 */ 1555 if (oi->ip_clusters != 0 || i_size_read(inode) != 0) 1556 return 0; 1557 1558 /* 1559 * Check whether the write can fit. 1560 */ 1561 di = (struct ocfs2_dinode *)wc->w_di_bh->b_data; 1562 if (mmap_page || 1563 end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) 1564 return 0; 1565 1566 do_inline_write: 1567 ret = ocfs2_write_begin_inline(mapping, inode, wc); 1568 if (ret) { 1569 mlog_errno(ret); 1570 goto out; 1571 } 1572 1573 /* 1574 * This signals to the caller that the data can be written 1575 * inline. 1576 */ 1577 written = 1; 1578 out: 1579 return written ? written : ret; 1580 } 1581 1582 /* 1583 * This function only does anything for file systems which can't 1584 * handle sparse files. 1585 * 1586 * What we want to do here is fill in any hole between the current end 1587 * of allocation and the end of our write. That way the rest of the 1588 * write path can treat it as an non-allocating write, which has no 1589 * special case code for sparse/nonsparse files. 1590 */ 1591 static int ocfs2_expand_nonsparse_inode(struct inode *inode, 1592 struct buffer_head *di_bh, 1593 loff_t pos, unsigned len, 1594 struct ocfs2_write_ctxt *wc) 1595 { 1596 int ret; 1597 loff_t newsize = pos + len; 1598 1599 BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))); 1600 1601 if (newsize <= i_size_read(inode)) 1602 return 0; 1603 1604 ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos); 1605 if (ret) 1606 mlog_errno(ret); 1607 1608 wc->w_first_new_cpos = 1609 ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode)); 1610 1611 return ret; 1612 } 1613 1614 static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh, 1615 loff_t pos) 1616 { 1617 int ret = 0; 1618 1619 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))); 1620 if (pos > i_size_read(inode)) 1621 ret = ocfs2_zero_extend(inode, di_bh, pos); 1622 1623 return ret; 1624 } 1625 1626 /* 1627 * Try to flush truncate logs if we can free enough clusters from it. 1628 * As for return value, "< 0" means error, "0" no space and "1" means 1629 * we have freed enough spaces and let the caller try to allocate again. 1630 */ 1631 static int ocfs2_try_to_free_truncate_log(struct ocfs2_super *osb, 1632 unsigned int needed) 1633 { 1634 tid_t target; 1635 int ret = 0; 1636 unsigned int truncated_clusters; 1637 1638 mutex_lock(&osb->osb_tl_inode->i_mutex); 1639 truncated_clusters = osb->truncated_clusters; 1640 mutex_unlock(&osb->osb_tl_inode->i_mutex); 1641 1642 /* 1643 * Check whether we can succeed in allocating if we free 1644 * the truncate log. 1645 */ 1646 if (truncated_clusters < needed) 1647 goto out; 1648 1649 ret = ocfs2_flush_truncate_log(osb); 1650 if (ret) { 1651 mlog_errno(ret); 1652 goto out; 1653 } 1654 1655 if (jbd2_journal_start_commit(osb->journal->j_journal, &target)) { 1656 jbd2_log_wait_commit(osb->journal->j_journal, target); 1657 ret = 1; 1658 } 1659 out: 1660 return ret; 1661 } 1662 1663 int ocfs2_write_begin_nolock(struct file *filp, 1664 struct address_space *mapping, 1665 loff_t pos, unsigned len, unsigned flags, 1666 struct page **pagep, void **fsdata, 1667 struct buffer_head *di_bh, struct page *mmap_page) 1668 { 1669 int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS; 1670 unsigned int clusters_to_alloc, extents_to_split, clusters_need = 0; 1671 struct ocfs2_write_ctxt *wc; 1672 struct inode *inode = mapping->host; 1673 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 1674 struct ocfs2_dinode *di; 1675 struct ocfs2_alloc_context *data_ac = NULL; 1676 struct ocfs2_alloc_context *meta_ac = NULL; 1677 handle_t *handle; 1678 struct ocfs2_extent_tree et; 1679 int try_free = 1, ret1; 1680 1681 try_again: 1682 ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh); 1683 if (ret) { 1684 mlog_errno(ret); 1685 return ret; 1686 } 1687 1688 if (ocfs2_supports_inline_data(osb)) { 1689 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len, 1690 mmap_page, wc); 1691 if (ret == 1) { 1692 ret = 0; 1693 goto success; 1694 } 1695 if (ret < 0) { 1696 mlog_errno(ret); 1697 goto out; 1698 } 1699 } 1700 1701 if (ocfs2_sparse_alloc(osb)) 1702 ret = ocfs2_zero_tail(inode, di_bh, pos); 1703 else 1704 ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos, len, 1705 wc); 1706 if (ret) { 1707 mlog_errno(ret); 1708 goto out; 1709 } 1710 1711 ret = ocfs2_check_range_for_refcount(inode, pos, len); 1712 if (ret < 0) { 1713 mlog_errno(ret); 1714 goto out; 1715 } else if (ret == 1) { 1716 clusters_need = wc->w_clen; 1717 ret = ocfs2_refcount_cow(inode, filp, di_bh, 1718 wc->w_cpos, wc->w_clen, UINT_MAX); 1719 if (ret) { 1720 mlog_errno(ret); 1721 goto out; 1722 } 1723 } 1724 1725 ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc, 1726 &extents_to_split); 1727 if (ret) { 1728 mlog_errno(ret); 1729 goto out; 1730 } 1731 clusters_need += clusters_to_alloc; 1732 1733 di = (struct ocfs2_dinode *)wc->w_di_bh->b_data; 1734 1735 trace_ocfs2_write_begin_nolock( 1736 (unsigned long long)OCFS2_I(inode)->ip_blkno, 1737 (long long)i_size_read(inode), 1738 le32_to_cpu(di->i_clusters), 1739 pos, len, flags, mmap_page, 1740 clusters_to_alloc, extents_to_split); 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 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), 1755 wc->w_di_bh); 1756 ret = ocfs2_lock_allocators(inode, &et, 1757 clusters_to_alloc, extents_to_split, 1758 &data_ac, &meta_ac); 1759 if (ret) { 1760 mlog_errno(ret); 1761 goto out; 1762 } 1763 1764 if (data_ac) 1765 data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv; 1766 1767 credits = ocfs2_calc_extend_credits(inode->i_sb, 1768 &di->id2.i_list, 1769 clusters_to_alloc); 1770 1771 } 1772 1773 /* 1774 * We have to zero sparse allocated clusters, unwritten extent clusters, 1775 * and non-sparse clusters we just extended. For non-sparse writes, 1776 * we know zeros will only be needed in the first and/or last cluster. 1777 */ 1778 if (clusters_to_alloc || extents_to_split || 1779 (wc->w_clen && (wc->w_desc[0].c_needs_zero || 1780 wc->w_desc[wc->w_clen - 1].c_needs_zero))) 1781 cluster_of_pages = 1; 1782 else 1783 cluster_of_pages = 0; 1784 1785 ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages); 1786 1787 handle = ocfs2_start_trans(osb, credits); 1788 if (IS_ERR(handle)) { 1789 ret = PTR_ERR(handle); 1790 mlog_errno(ret); 1791 goto out; 1792 } 1793 1794 wc->w_handle = handle; 1795 1796 if (clusters_to_alloc) { 1797 ret = dquot_alloc_space_nodirty(inode, 1798 ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc)); 1799 if (ret) 1800 goto out_commit; 1801 } 1802 /* 1803 * We don't want this to fail in ocfs2_write_end(), so do it 1804 * here. 1805 */ 1806 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh, 1807 OCFS2_JOURNAL_ACCESS_WRITE); 1808 if (ret) { 1809 mlog_errno(ret); 1810 goto out_quota; 1811 } 1812 1813 /* 1814 * Fill our page array first. That way we've grabbed enough so 1815 * that we can zero and flush if we error after adding the 1816 * extent. 1817 */ 1818 ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len, 1819 cluster_of_pages, mmap_page); 1820 if (ret) { 1821 mlog_errno(ret); 1822 goto out_quota; 1823 } 1824 1825 ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos, 1826 len); 1827 if (ret) { 1828 mlog_errno(ret); 1829 goto out_quota; 1830 } 1831 1832 if (data_ac) 1833 ocfs2_free_alloc_context(data_ac); 1834 if (meta_ac) 1835 ocfs2_free_alloc_context(meta_ac); 1836 1837 success: 1838 *pagep = wc->w_target_page; 1839 *fsdata = wc; 1840 return 0; 1841 out_quota: 1842 if (clusters_to_alloc) 1843 dquot_free_space(inode, 1844 ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc)); 1845 out_commit: 1846 ocfs2_commit_trans(osb, handle); 1847 1848 out: 1849 ocfs2_free_write_ctxt(wc); 1850 1851 if (data_ac) 1852 ocfs2_free_alloc_context(data_ac); 1853 if (meta_ac) 1854 ocfs2_free_alloc_context(meta_ac); 1855 1856 if (ret == -ENOSPC && try_free) { 1857 /* 1858 * Try to free some truncate log so that we can have enough 1859 * clusters to allocate. 1860 */ 1861 try_free = 0; 1862 1863 ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need); 1864 if (ret1 == 1) 1865 goto try_again; 1866 1867 if (ret1 < 0) 1868 mlog_errno(ret1); 1869 } 1870 1871 return ret; 1872 } 1873 1874 static int ocfs2_write_begin(struct file *file, struct address_space *mapping, 1875 loff_t pos, unsigned len, unsigned flags, 1876 struct page **pagep, void **fsdata) 1877 { 1878 int ret; 1879 struct buffer_head *di_bh = NULL; 1880 struct inode *inode = mapping->host; 1881 1882 ret = ocfs2_inode_lock(inode, &di_bh, 1); 1883 if (ret) { 1884 mlog_errno(ret); 1885 return ret; 1886 } 1887 1888 /* 1889 * Take alloc sem here to prevent concurrent lookups. That way 1890 * the mapping, zeroing and tree manipulation within 1891 * ocfs2_write() will be safe against ->readpage(). This 1892 * should also serve to lock out allocation from a shared 1893 * writeable region. 1894 */ 1895 down_write(&OCFS2_I(inode)->ip_alloc_sem); 1896 1897 ret = ocfs2_write_begin_nolock(file, mapping, pos, len, flags, pagep, 1898 fsdata, di_bh, NULL); 1899 if (ret) { 1900 mlog_errno(ret); 1901 goto out_fail; 1902 } 1903 1904 brelse(di_bh); 1905 1906 return 0; 1907 1908 out_fail: 1909 up_write(&OCFS2_I(inode)->ip_alloc_sem); 1910 1911 brelse(di_bh); 1912 ocfs2_inode_unlock(inode, 1); 1913 1914 return ret; 1915 } 1916 1917 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos, 1918 unsigned len, unsigned *copied, 1919 struct ocfs2_dinode *di, 1920 struct ocfs2_write_ctxt *wc) 1921 { 1922 void *kaddr; 1923 1924 if (unlikely(*copied < len)) { 1925 if (!PageUptodate(wc->w_target_page)) { 1926 *copied = 0; 1927 return; 1928 } 1929 } 1930 1931 kaddr = kmap_atomic(wc->w_target_page, KM_USER0); 1932 memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied); 1933 kunmap_atomic(kaddr, KM_USER0); 1934 1935 trace_ocfs2_write_end_inline( 1936 (unsigned long long)OCFS2_I(inode)->ip_blkno, 1937 (unsigned long long)pos, *copied, 1938 le16_to_cpu(di->id2.i_data.id_count), 1939 le16_to_cpu(di->i_dyn_features)); 1940 } 1941 1942 int ocfs2_write_end_nolock(struct address_space *mapping, 1943 loff_t pos, unsigned len, unsigned copied, 1944 struct page *page, void *fsdata) 1945 { 1946 int i; 1947 unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1); 1948 struct inode *inode = mapping->host; 1949 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 1950 struct ocfs2_write_ctxt *wc = fsdata; 1951 struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data; 1952 handle_t *handle = wc->w_handle; 1953 struct page *tmppage; 1954 1955 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) { 1956 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc); 1957 goto out_write_size; 1958 } 1959 1960 if (unlikely(copied < len)) { 1961 if (!PageUptodate(wc->w_target_page)) 1962 copied = 0; 1963 1964 ocfs2_zero_new_buffers(wc->w_target_page, start+copied, 1965 start+len); 1966 } 1967 flush_dcache_page(wc->w_target_page); 1968 1969 for(i = 0; i < wc->w_num_pages; i++) { 1970 tmppage = wc->w_pages[i]; 1971 1972 if (tmppage == wc->w_target_page) { 1973 from = wc->w_target_from; 1974 to = wc->w_target_to; 1975 1976 BUG_ON(from > PAGE_CACHE_SIZE || 1977 to > PAGE_CACHE_SIZE || 1978 to < from); 1979 } else { 1980 /* 1981 * Pages adjacent to the target (if any) imply 1982 * a hole-filling write in which case we want 1983 * to flush their entire range. 1984 */ 1985 from = 0; 1986 to = PAGE_CACHE_SIZE; 1987 } 1988 1989 if (page_has_buffers(tmppage)) { 1990 if (ocfs2_should_order_data(inode)) 1991 ocfs2_jbd2_file_inode(wc->w_handle, inode); 1992 block_commit_write(tmppage, from, to); 1993 } 1994 } 1995 1996 out_write_size: 1997 pos += copied; 1998 if (pos > inode->i_size) { 1999 i_size_write(inode, pos); 2000 mark_inode_dirty(inode); 2001 } 2002 inode->i_blocks = ocfs2_inode_sector_count(inode); 2003 di->i_size = cpu_to_le64((u64)i_size_read(inode)); 2004 inode->i_mtime = inode->i_ctime = CURRENT_TIME; 2005 di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec); 2006 di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec); 2007 ocfs2_journal_dirty(handle, wc->w_di_bh); 2008 2009 ocfs2_commit_trans(osb, handle); 2010 2011 ocfs2_run_deallocs(osb, &wc->w_dealloc); 2012 2013 ocfs2_free_write_ctxt(wc); 2014 2015 return copied; 2016 } 2017 2018 static int ocfs2_write_end(struct file *file, struct address_space *mapping, 2019 loff_t pos, unsigned len, unsigned copied, 2020 struct page *page, void *fsdata) 2021 { 2022 int ret; 2023 struct inode *inode = mapping->host; 2024 2025 ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata); 2026 2027 up_write(&OCFS2_I(inode)->ip_alloc_sem); 2028 ocfs2_inode_unlock(inode, 1); 2029 2030 return ret; 2031 } 2032 2033 const struct address_space_operations ocfs2_aops = { 2034 .readpage = ocfs2_readpage, 2035 .readpages = ocfs2_readpages, 2036 .writepage = ocfs2_writepage, 2037 .write_begin = ocfs2_write_begin, 2038 .write_end = ocfs2_write_end, 2039 .bmap = ocfs2_bmap, 2040 .direct_IO = ocfs2_direct_IO, 2041 .invalidatepage = ocfs2_invalidatepage, 2042 .releasepage = ocfs2_releasepage, 2043 .migratepage = buffer_migrate_page, 2044 .is_partially_uptodate = block_is_partially_uptodate, 2045 .error_remove_page = generic_error_remove_page, 2046 }; 2047