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 51static 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 123bail: 124 brelse(bh); 125 126 mlog_exit(err); 127 return err; 128} 129 130int 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_prepare_write() 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 207bail: 208 if (err < 0) 209 err = -EIO; 210 211 mlog_exit(err); 212 return err; 213} 214 215int 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 252static 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); 267out: 268 unlock_page(page); 269 270 brelse(di_bh); 271 return ret; 272} 273 274static 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 if (start >= i_size_read(inode)) { 297 zero_user(page, 0, PAGE_SIZE); 298 SetPageUptodate(page); 299 ret = 0; 300 goto out_alloc; 301 } 302 303 if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) 304 ret = ocfs2_readpage_inline(inode, page); 305 else 306 ret = block_read_full_page(page, ocfs2_get_block); 307 unlock = 0; 308 309out_alloc: 310 up_read(&OCFS2_I(inode)->ip_alloc_sem); 311out_inode_unlock: 312 ocfs2_inode_unlock(inode, 0); 313out: 314 if (unlock) 315 unlock_page(page); 316 mlog_exit(ret); 317 return ret; 318} 319 320/* 321 * This is used only for read-ahead. Failures or difficult to handle 322 * situations are safe to ignore. 323 * 324 * Right now, we don't bother with BH_Boundary - in-inode extent lists 325 * are quite large (243 extents on 4k blocks), so most inodes don't 326 * grow out to a tree. If need be, detecting boundary extents could 327 * trivially be added in a future version of ocfs2_get_block(). 328 */ 329static int ocfs2_readpages(struct file *filp, struct address_space *mapping, 330 struct list_head *pages, unsigned nr_pages) 331{ 332 int ret, err = -EIO; 333 struct inode *inode = mapping->host; 334 struct ocfs2_inode_info *oi = OCFS2_I(inode); 335 loff_t start; 336 struct page *last; 337 338 /* 339 * Use the nonblocking flag for the dlm code to avoid page 340 * lock inversion, but don't bother with retrying. 341 */ 342 ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK); 343 if (ret) 344 return err; 345 346 if (down_read_trylock(&oi->ip_alloc_sem) == 0) { 347 ocfs2_inode_unlock(inode, 0); 348 return err; 349 } 350 351 /* 352 * Don't bother with inline-data. There isn't anything 353 * to read-ahead in that case anyway... 354 */ 355 if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) 356 goto out_unlock; 357 358 /* 359 * Check whether a remote node truncated this file - we just 360 * drop out in that case as it's not worth handling here. 361 */ 362 last = list_entry(pages->prev, struct page, lru); 363 start = (loff_t)last->index << PAGE_CACHE_SHIFT; 364 if (start >= i_size_read(inode)) 365 goto out_unlock; 366 367 err = mpage_readpages(mapping, pages, nr_pages, ocfs2_get_block); 368 369out_unlock: 370 up_read(&oi->ip_alloc_sem); 371 ocfs2_inode_unlock(inode, 0); 372 373 return err; 374} 375 376/* Note: Because we don't support holes, our allocation has 377 * already happened (allocation writes zeros to the file data) 378 * so we don't have to worry about ordered writes in 379 * ocfs2_writepage. 380 * 381 * ->writepage is called during the process of invalidating the page cache 382 * during blocked lock processing. It can't block on any cluster locks 383 * to during block mapping. It's relying on the fact that the block 384 * mapping can't have disappeared under the dirty pages that it is 385 * being asked to write back. 386 */ 387static int ocfs2_writepage(struct page *page, struct writeback_control *wbc) 388{ 389 int ret; 390 391 mlog_entry("(0x%p)\n", page); 392 393 ret = block_write_full_page(page, ocfs2_get_block, wbc); 394 395 mlog_exit(ret); 396 397 return ret; 398} 399 400/* 401 * This is called from ocfs2_write_zero_page() which has handled it's 402 * own cluster locking and has ensured allocation exists for those 403 * blocks to be written. 404 */ 405int ocfs2_prepare_write_nolock(struct inode *inode, struct page *page, 406 unsigned from, unsigned to) 407{ 408 int ret; 409 410 ret = block_prepare_write(page, from, to, ocfs2_get_block); 411 412 return ret; 413} 414 415/* Taken from ext3. We don't necessarily need the full blown 416 * functionality yet, but IMHO it's better to cut and paste the whole 417 * thing so we can avoid introducing our own bugs (and easily pick up 418 * their fixes when they happen) --Mark */ 419int walk_page_buffers( handle_t *handle, 420 struct buffer_head *head, 421 unsigned from, 422 unsigned to, 423 int *partial, 424 int (*fn)( handle_t *handle, 425 struct buffer_head *bh)) 426{ 427 struct buffer_head *bh; 428 unsigned block_start, block_end; 429 unsigned blocksize = head->b_size; 430 int err, ret = 0; 431 struct buffer_head *next; 432 433 for ( bh = head, block_start = 0; 434 ret == 0 && (bh != head || !block_start); 435 block_start = block_end, bh = next) 436 { 437 next = bh->b_this_page; 438 block_end = block_start + blocksize; 439 if (block_end <= from || block_start >= to) { 440 if (partial && !buffer_uptodate(bh)) 441 *partial = 1; 442 continue; 443 } 444 err = (*fn)(handle, bh); 445 if (!ret) 446 ret = err; 447 } 448 return ret; 449} 450 451static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block) 452{ 453 sector_t status; 454 u64 p_blkno = 0; 455 int err = 0; 456 struct inode *inode = mapping->host; 457 458 mlog_entry("(block = %llu)\n", (unsigned long long)block); 459 460 /* We don't need to lock journal system files, since they aren't 461 * accessed concurrently from multiple nodes. 462 */ 463 if (!INODE_JOURNAL(inode)) { 464 err = ocfs2_inode_lock(inode, NULL, 0); 465 if (err) { 466 if (err != -ENOENT) 467 mlog_errno(err); 468 goto bail; 469 } 470 down_read(&OCFS2_I(inode)->ip_alloc_sem); 471 } 472 473 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)) 474 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL, 475 NULL); 476 477 if (!INODE_JOURNAL(inode)) { 478 up_read(&OCFS2_I(inode)->ip_alloc_sem); 479 ocfs2_inode_unlock(inode, 0); 480 } 481 482 if (err) { 483 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n", 484 (unsigned long long)block); 485 mlog_errno(err); 486 goto bail; 487 } 488 489bail: 490 status = err ? 0 : p_blkno; 491 492 mlog_exit((int)status); 493 494 return status; 495} 496 497/* 498 * TODO: Make this into a generic get_blocks function. 499 * 500 * From do_direct_io in direct-io.c: 501 * "So what we do is to permit the ->get_blocks function to populate 502 * bh.b_size with the size of IO which is permitted at this offset and 503 * this i_blkbits." 504 * 505 * This function is called directly from get_more_blocks in direct-io.c. 506 * 507 * called like this: dio->get_blocks(dio->inode, fs_startblk, 508 * fs_count, map_bh, dio->rw == WRITE); 509 * 510 * Note that we never bother to allocate blocks here, and thus ignore the 511 * create argument. 512 */ 513static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock, 514 struct buffer_head *bh_result, int create) 515{ 516 int ret; 517 u64 p_blkno, inode_blocks, contig_blocks; 518 unsigned int ext_flags; 519 unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits; 520 unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits; 521 522 /* This function won't even be called if the request isn't all 523 * nicely aligned and of the right size, so there's no need 524 * for us to check any of that. */ 525 526 inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode)); 527 528 /* This figures out the size of the next contiguous block, and 529 * our logical offset */ 530 ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, 531 &contig_blocks, &ext_flags); 532 if (ret) { 533 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n", 534 (unsigned long long)iblock); 535 ret = -EIO; 536 goto bail; 537 } 538 539 /* We should already CoW the refcounted extent in case of create. */ 540 BUG_ON(create && (ext_flags & OCFS2_EXT_REFCOUNTED)); 541 542 /* 543 * get_more_blocks() expects us to describe a hole by clearing 544 * the mapped bit on bh_result(). 545 * 546 * Consider an unwritten extent as a hole. 547 */ 548 if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN)) 549 map_bh(bh_result, inode->i_sb, p_blkno); 550 else 551 clear_buffer_mapped(bh_result); 552 553 /* make sure we don't map more than max_blocks blocks here as 554 that's all the kernel will handle at this point. */ 555 if (max_blocks < contig_blocks) 556 contig_blocks = max_blocks; 557 bh_result->b_size = contig_blocks << blocksize_bits; 558bail: 559 return ret; 560} 561 562/* 563 * ocfs2_dio_end_io is called by the dio core when a dio is finished. We're 564 * particularly interested in the aio/dio case. Like the core uses 565 * i_alloc_sem, we use the rw_lock DLM lock to protect io on one node from 566 * truncation on another. 567 */ 568static void ocfs2_dio_end_io(struct kiocb *iocb, 569 loff_t offset, 570 ssize_t bytes, 571 void *private, 572 int ret, 573 bool is_async) 574{ 575 struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode; 576 int level; 577 578 /* this io's submitter should not have unlocked this before we could */ 579 BUG_ON(!ocfs2_iocb_is_rw_locked(iocb)); 580 581 ocfs2_iocb_clear_rw_locked(iocb); 582 583 level = ocfs2_iocb_rw_locked_level(iocb); 584 if (!level) 585 up_read(&inode->i_alloc_sem); 586 ocfs2_rw_unlock(inode, level); 587 588 if (is_async) 589 aio_complete(iocb, ret, 0); 590} 591 592/* 593 * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen 594 * from ext3. PageChecked() bits have been removed as OCFS2 does not 595 * do journalled data. 596 */ 597static void ocfs2_invalidatepage(struct page *page, unsigned long offset) 598{ 599 journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal; 600 601 jbd2_journal_invalidatepage(journal, page, offset); 602} 603 604static int ocfs2_releasepage(struct page *page, gfp_t wait) 605{ 606 journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal; 607 608 if (!page_has_buffers(page)) 609 return 0; 610 return jbd2_journal_try_to_free_buffers(journal, page, wait); 611} 612 613static ssize_t ocfs2_direct_IO(int rw, 614 struct kiocb *iocb, 615 const struct iovec *iov, 616 loff_t offset, 617 unsigned long nr_segs) 618{ 619 struct file *file = iocb->ki_filp; 620 struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host; 621 int ret; 622 623 mlog_entry_void(); 624 625 /* 626 * Fallback to buffered I/O if we see an inode without 627 * extents. 628 */ 629 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) 630 return 0; 631 632 /* Fallback to buffered I/O if we are appending. */ 633 if (i_size_read(inode) <= offset) 634 return 0; 635 636 ret = __blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, 637 iov, offset, nr_segs, 638 ocfs2_direct_IO_get_blocks, 639 ocfs2_dio_end_io, NULL, 0); 640 641 mlog_exit(ret); 642 return ret; 643} 644 645static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb, 646 u32 cpos, 647 unsigned int *start, 648 unsigned int *end) 649{ 650 unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE; 651 652 if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) { 653 unsigned int cpp; 654 655 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits); 656 657 cluster_start = cpos % cpp; 658 cluster_start = cluster_start << osb->s_clustersize_bits; 659 660 cluster_end = cluster_start + osb->s_clustersize; 661 } 662 663 BUG_ON(cluster_start > PAGE_SIZE); 664 BUG_ON(cluster_end > PAGE_SIZE); 665 666 if (start) 667 *start = cluster_start; 668 if (end) 669 *end = cluster_end; 670} 671 672/* 673 * 'from' and 'to' are the region in the page to avoid zeroing. 674 * 675 * If pagesize > clustersize, this function will avoid zeroing outside 676 * of the cluster boundary. 677 * 678 * from == to == 0 is code for "zero the entire cluster region" 679 */ 680static void ocfs2_clear_page_regions(struct page *page, 681 struct ocfs2_super *osb, u32 cpos, 682 unsigned from, unsigned to) 683{ 684 void *kaddr; 685 unsigned int cluster_start, cluster_end; 686 687 ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end); 688 689 kaddr = kmap_atomic(page, KM_USER0); 690 691 if (from || to) { 692 if (from > cluster_start) 693 memset(kaddr + cluster_start, 0, from - cluster_start); 694 if (to < cluster_end) 695 memset(kaddr + to, 0, cluster_end - to); 696 } else { 697 memset(kaddr + cluster_start, 0, cluster_end - cluster_start); 698 } 699 700 kunmap_atomic(kaddr, KM_USER0); 701} 702 703/* 704 * Nonsparse file systems fully allocate before we get to the write 705 * code. This prevents ocfs2_write() from tagging the write as an 706 * allocating one, which means ocfs2_map_page_blocks() might try to 707 * read-in the blocks at the tail of our file. Avoid reading them by 708 * testing i_size against each block offset. 709 */ 710static int ocfs2_should_read_blk(struct inode *inode, struct page *page, 711 unsigned int block_start) 712{ 713 u64 offset = page_offset(page) + block_start; 714 715 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) 716 return 1; 717 718 if (i_size_read(inode) > offset) 719 return 1; 720 721 return 0; 722} 723 724/* 725 * Some of this taken from block_prepare_write(). We already have our 726 * mapping by now though, and the entire write will be allocating or 727 * it won't, so not much need to use BH_New. 728 * 729 * This will also skip zeroing, which is handled externally. 730 */ 731int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno, 732 struct inode *inode, unsigned int from, 733 unsigned int to, int new) 734{ 735 int ret = 0; 736 struct buffer_head *head, *bh, *wait[2], **wait_bh = wait; 737 unsigned int block_end, block_start; 738 unsigned int bsize = 1 << inode->i_blkbits; 739 740 if (!page_has_buffers(page)) 741 create_empty_buffers(page, bsize, 0); 742 743 head = page_buffers(page); 744 for (bh = head, block_start = 0; bh != head || !block_start; 745 bh = bh->b_this_page, block_start += bsize) { 746 block_end = block_start + bsize; 747 748 clear_buffer_new(bh); 749 750 /* 751 * Ignore blocks outside of our i/o range - 752 * they may belong to unallocated clusters. 753 */ 754 if (block_start >= to || block_end <= from) { 755 if (PageUptodate(page)) 756 set_buffer_uptodate(bh); 757 continue; 758 } 759 760 /* 761 * For an allocating write with cluster size >= page 762 * size, we always write the entire page. 763 */ 764 if (new) 765 set_buffer_new(bh); 766 767 if (!buffer_mapped(bh)) { 768 map_bh(bh, inode->i_sb, *p_blkno); 769 unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr); 770 } 771 772 if (PageUptodate(page)) { 773 if (!buffer_uptodate(bh)) 774 set_buffer_uptodate(bh); 775 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) && 776 !buffer_new(bh) && 777 ocfs2_should_read_blk(inode, page, block_start) && 778 (block_start < from || block_end > to)) { 779 ll_rw_block(READ, 1, &bh); 780 *wait_bh++=bh; 781 } 782 783 *p_blkno = *p_blkno + 1; 784 } 785 786 /* 787 * If we issued read requests - let them complete. 788 */ 789 while(wait_bh > wait) { 790 wait_on_buffer(*--wait_bh); 791 if (!buffer_uptodate(*wait_bh)) 792 ret = -EIO; 793 } 794 795 if (ret == 0 || !new) 796 return ret; 797 798 /* 799 * If we get -EIO above, zero out any newly allocated blocks 800 * to avoid exposing stale data. 801 */ 802 bh = head; 803 block_start = 0; 804 do { 805 block_end = block_start + bsize; 806 if (block_end <= from) 807 goto next_bh; 808 if (block_start >= to) 809 break; 810 811 zero_user(page, block_start, bh->b_size); 812 set_buffer_uptodate(bh); 813 mark_buffer_dirty(bh); 814 815next_bh: 816 block_start = block_end; 817 bh = bh->b_this_page; 818 } while (bh != head); 819 820 return ret; 821} 822 823#if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE) 824#define OCFS2_MAX_CTXT_PAGES 1 825#else 826#define OCFS2_MAX_CTXT_PAGES (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE) 827#endif 828 829#define OCFS2_MAX_CLUSTERS_PER_PAGE (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE) 830 831/* 832 * Describe the state of a single cluster to be written to. 833 */ 834struct ocfs2_write_cluster_desc { 835 u32 c_cpos; 836 u32 c_phys; 837 /* 838 * Give this a unique field because c_phys eventually gets 839 * filled. 840 */ 841 unsigned c_new; 842 unsigned c_unwritten; 843 unsigned c_needs_zero; 844}; 845 846struct ocfs2_write_ctxt { 847 /* Logical cluster position / len of write */ 848 u32 w_cpos; 849 u32 w_clen; 850 851 /* First cluster allocated in a nonsparse extend */ 852 u32 w_first_new_cpos; 853 854 struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE]; 855 856 /* 857 * This is true if page_size > cluster_size. 858 * 859 * It triggers a set of special cases during write which might 860 * have to deal with allocating writes to partial pages. 861 */ 862 unsigned int w_large_pages; 863 864 /* 865 * Pages involved in this write. 866 * 867 * w_target_page is the page being written to by the user. 868 * 869 * w_pages is an array of pages which always contains 870 * w_target_page, and in the case of an allocating write with 871 * page_size < cluster size, it will contain zero'd and mapped 872 * pages adjacent to w_target_page which need to be written 873 * out in so that future reads from that region will get 874 * zero's. 875 */ 876 struct page *w_pages[OCFS2_MAX_CTXT_PAGES]; 877 unsigned int w_num_pages; 878 struct page *w_target_page; 879 880 /* 881 * ocfs2_write_end() uses this to know what the real range to 882 * write in the target should be. 883 */ 884 unsigned int w_target_from; 885 unsigned int w_target_to; 886 887 /* 888 * We could use journal_current_handle() but this is cleaner, 889 * IMHO -Mark 890 */ 891 handle_t *w_handle; 892 893 struct buffer_head *w_di_bh; 894 895 struct ocfs2_cached_dealloc_ctxt w_dealloc; 896}; 897 898void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages) 899{ 900 int i; 901 902 for(i = 0; i < num_pages; i++) { 903 if (pages[i]) { 904 unlock_page(pages[i]); 905 mark_page_accessed(pages[i]); 906 page_cache_release(pages[i]); 907 } 908 } 909} 910 911static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc) 912{ 913 ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages); 914 915 brelse(wc->w_di_bh); 916 kfree(wc); 917} 918 919static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp, 920 struct ocfs2_super *osb, loff_t pos, 921 unsigned len, struct buffer_head *di_bh) 922{ 923 u32 cend; 924 struct ocfs2_write_ctxt *wc; 925 926 wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS); 927 if (!wc) 928 return -ENOMEM; 929 930 wc->w_cpos = pos >> osb->s_clustersize_bits; 931 wc->w_first_new_cpos = UINT_MAX; 932 cend = (pos + len - 1) >> osb->s_clustersize_bits; 933 wc->w_clen = cend - wc->w_cpos + 1; 934 get_bh(di_bh); 935 wc->w_di_bh = di_bh; 936 937 if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) 938 wc->w_large_pages = 1; 939 else 940 wc->w_large_pages = 0; 941 942 ocfs2_init_dealloc_ctxt(&wc->w_dealloc); 943 944 *wcp = wc; 945 946 return 0; 947} 948 949/* 950 * If a page has any new buffers, zero them out here, and mark them uptodate 951 * and dirty so they'll be written out (in order to prevent uninitialised 952 * block data from leaking). And clear the new bit. 953 */ 954static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to) 955{ 956 unsigned int block_start, block_end; 957 struct buffer_head *head, *bh; 958 959 BUG_ON(!PageLocked(page)); 960 if (!page_has_buffers(page)) 961 return; 962 963 bh = head = page_buffers(page); 964 block_start = 0; 965 do { 966 block_end = block_start + bh->b_size; 967 968 if (buffer_new(bh)) { 969 if (block_end > from && block_start < to) { 970 if (!PageUptodate(page)) { 971 unsigned start, end; 972 973 start = max(from, block_start); 974 end = min(to, block_end); 975 976 zero_user_segment(page, start, end); 977 set_buffer_uptodate(bh); 978 } 979 980 clear_buffer_new(bh); 981 mark_buffer_dirty(bh); 982 } 983 } 984 985 block_start = block_end; 986 bh = bh->b_this_page; 987 } while (bh != head); 988} 989 990/* 991 * Only called when we have a failure during allocating write to write 992 * zero's to the newly allocated region. 993 */ 994static void ocfs2_write_failure(struct inode *inode, 995 struct ocfs2_write_ctxt *wc, 996 loff_t user_pos, unsigned user_len) 997{ 998 int i; 999 unsigned from = user_pos & (PAGE_CACHE_SIZE - 1), 1000 to = user_pos + user_len; 1001 struct page *tmppage; 1002 1003 ocfs2_zero_new_buffers(wc->w_target_page, from, to); 1004 1005 for(i = 0; i < wc->w_num_pages; i++) { 1006 tmppage = wc->w_pages[i]; 1007 1008 if (page_has_buffers(tmppage)) { 1009 if (ocfs2_should_order_data(inode)) 1010 ocfs2_jbd2_file_inode(wc->w_handle, inode); 1011 1012 block_commit_write(tmppage, from, to); 1013 } 1014 } 1015} 1016 1017static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno, 1018 struct ocfs2_write_ctxt *wc, 1019 struct page *page, u32 cpos, 1020 loff_t user_pos, unsigned user_len, 1021 int new) 1022{ 1023 int ret; 1024 unsigned int map_from = 0, map_to = 0; 1025 unsigned int cluster_start, cluster_end; 1026 unsigned int user_data_from = 0, user_data_to = 0; 1027 1028 ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos, 1029 &cluster_start, &cluster_end); 1030 1031 if (page == wc->w_target_page) { 1032 map_from = user_pos & (PAGE_CACHE_SIZE - 1); 1033 map_to = map_from + user_len; 1034 1035 if (new) 1036 ret = ocfs2_map_page_blocks(page, p_blkno, inode, 1037 cluster_start, cluster_end, 1038 new); 1039 else 1040 ret = ocfs2_map_page_blocks(page, p_blkno, inode, 1041 map_from, map_to, new); 1042 if (ret) { 1043 mlog_errno(ret); 1044 goto out; 1045 } 1046 1047 user_data_from = map_from; 1048 user_data_to = map_to; 1049 if (new) { 1050 map_from = cluster_start; 1051 map_to = cluster_end; 1052 } 1053 } else { 1054 /* 1055 * If we haven't allocated the new page yet, we 1056 * shouldn't be writing it out without copying user 1057 * data. This is likely a math error from the caller. 1058 */ 1059 BUG_ON(!new); 1060 1061 map_from = cluster_start; 1062 map_to = cluster_end; 1063 1064 ret = ocfs2_map_page_blocks(page, p_blkno, inode, 1065 cluster_start, cluster_end, new); 1066 if (ret) { 1067 mlog_errno(ret); 1068 goto out; 1069 } 1070 } 1071 1072 /* 1073 * Parts of newly allocated pages need to be zero'd. 1074 * 1075 * Above, we have also rewritten 'to' and 'from' - as far as 1076 * the rest of the function is concerned, the entire cluster 1077 * range inside of a page needs to be written. 1078 * 1079 * We can skip this if the page is up to date - it's already 1080 * been zero'd from being read in as a hole. 1081 */ 1082 if (new && !PageUptodate(page)) 1083 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb), 1084 cpos, user_data_from, user_data_to); 1085 1086 flush_dcache_page(page); 1087 1088out: 1089 return ret; 1090} 1091 1092/* 1093 * This function will only grab one clusters worth of pages. 1094 */ 1095static int ocfs2_grab_pages_for_write(struct address_space *mapping, 1096 struct ocfs2_write_ctxt *wc, 1097 u32 cpos, loff_t user_pos, 1098 unsigned user_len, int new, 1099 struct page *mmap_page) 1100{ 1101 int ret = 0, i; 1102 unsigned long start, target_index, end_index, index; 1103 struct inode *inode = mapping->host; 1104 loff_t last_byte; 1105 1106 target_index = user_pos >> PAGE_CACHE_SHIFT; 1107 1108 /* 1109 * Figure out how many pages we'll be manipulating here. For 1110 * non allocating write, we just change the one 1111 * page. Otherwise, we'll need a whole clusters worth. If we're 1112 * writing past i_size, we only need enough pages to cover the 1113 * last page of the write. 1114 */ 1115 if (new) { 1116 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb); 1117 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos); 1118 /* 1119 * We need the index *past* the last page we could possibly 1120 * touch. This is the page past the end of the write or 1121 * i_size, whichever is greater. 1122 */ 1123 last_byte = max(user_pos + user_len, i_size_read(inode)); 1124 BUG_ON(last_byte < 1); 1125 end_index = ((last_byte - 1) >> PAGE_CACHE_SHIFT) + 1; 1126 if ((start + wc->w_num_pages) > end_index) 1127 wc->w_num_pages = end_index - start; 1128 } else { 1129 wc->w_num_pages = 1; 1130 start = target_index; 1131 } 1132 1133 for(i = 0; i < wc->w_num_pages; i++) { 1134 index = start + i; 1135 1136 if (index == target_index && mmap_page) { 1137 /* 1138 * ocfs2_pagemkwrite() is a little different 1139 * and wants us to directly use the page 1140 * passed in. 1141 */ 1142 lock_page(mmap_page); 1143 1144 if (mmap_page->mapping != mapping) { 1145 unlock_page(mmap_page); 1146 /* 1147 * Sanity check - the locking in 1148 * ocfs2_pagemkwrite() should ensure 1149 * that this code doesn't trigger. 1150 */ 1151 ret = -EINVAL; 1152 mlog_errno(ret); 1153 goto out; 1154 } 1155 1156 page_cache_get(mmap_page); 1157 wc->w_pages[i] = mmap_page; 1158 } else { 1159 wc->w_pages[i] = find_or_create_page(mapping, index, 1160 GFP_NOFS); 1161 if (!wc->w_pages[i]) { 1162 ret = -ENOMEM; 1163 mlog_errno(ret); 1164 goto out; 1165 } 1166 } 1167 1168 if (index == target_index) 1169 wc->w_target_page = wc->w_pages[i]; 1170 } 1171out: 1172 return ret; 1173} 1174 1175/* 1176 * Prepare a single cluster for write one cluster into the file. 1177 */ 1178static int ocfs2_write_cluster(struct address_space *mapping, 1179 u32 phys, unsigned int unwritten, 1180 unsigned int should_zero, 1181 struct ocfs2_alloc_context *data_ac, 1182 struct ocfs2_alloc_context *meta_ac, 1183 struct ocfs2_write_ctxt *wc, u32 cpos, 1184 loff_t user_pos, unsigned user_len) 1185{ 1186 int ret, i, new; 1187 u64 v_blkno, p_blkno; 1188 struct inode *inode = mapping->host; 1189 struct ocfs2_extent_tree et; 1190 1191 new = phys == 0 ? 1 : 0; 1192 if (new) { 1193 u32 tmp_pos; 1194 1195 /* 1196 * This is safe to call with the page locks - it won't take 1197 * any additional semaphores or cluster locks. 1198 */ 1199 tmp_pos = cpos; 1200 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode, 1201 &tmp_pos, 1, 0, wc->w_di_bh, 1202 wc->w_handle, data_ac, 1203 meta_ac, NULL); 1204 /* 1205 * This shouldn't happen because we must have already 1206 * calculated the correct meta data allocation required. The 1207 * internal tree allocation code should know how to increase 1208 * transaction credits itself. 1209 * 1210 * If need be, we could handle -EAGAIN for a 1211 * RESTART_TRANS here. 1212 */ 1213 mlog_bug_on_msg(ret == -EAGAIN, 1214 "Inode %llu: EAGAIN return during allocation.\n", 1215 (unsigned long long)OCFS2_I(inode)->ip_blkno); 1216 if (ret < 0) { 1217 mlog_errno(ret); 1218 goto out; 1219 } 1220 } else if (unwritten) { 1221 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), 1222 wc->w_di_bh); 1223 ret = ocfs2_mark_extent_written(inode, &et, 1224 wc->w_handle, cpos, 1, phys, 1225 meta_ac, &wc->w_dealloc); 1226 if (ret < 0) { 1227 mlog_errno(ret); 1228 goto out; 1229 } 1230 } 1231 1232 if (should_zero) 1233 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos); 1234 else 1235 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits; 1236 1237 /* 1238 * The only reason this should fail is due to an inability to 1239 * find the extent added. 1240 */ 1241 ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL, 1242 NULL); 1243 if (ret < 0) { 1244 ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, " 1245 "at logical block %llu", 1246 (unsigned long long)OCFS2_I(inode)->ip_blkno, 1247 (unsigned long long)v_blkno); 1248 goto out; 1249 } 1250 1251 BUG_ON(p_blkno == 0); 1252 1253 for(i = 0; i < wc->w_num_pages; i++) { 1254 int tmpret; 1255 1256 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc, 1257 wc->w_pages[i], cpos, 1258 user_pos, user_len, 1259 should_zero); 1260 if (tmpret) { 1261 mlog_errno(tmpret); 1262 if (ret == 0) 1263 ret = tmpret; 1264 } 1265 } 1266 1267 /* 1268 * We only have cleanup to do in case of allocating write. 1269 */ 1270 if (ret && new) 1271 ocfs2_write_failure(inode, wc, user_pos, user_len); 1272 1273out: 1274 1275 return ret; 1276} 1277 1278static int ocfs2_write_cluster_by_desc(struct address_space *mapping, 1279 struct ocfs2_alloc_context *data_ac, 1280 struct ocfs2_alloc_context *meta_ac, 1281 struct ocfs2_write_ctxt *wc, 1282 loff_t pos, unsigned len) 1283{ 1284 int ret, i; 1285 loff_t cluster_off; 1286 unsigned int local_len = len; 1287 struct ocfs2_write_cluster_desc *desc; 1288 struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb); 1289 1290 for (i = 0; i < wc->w_clen; i++) { 1291 desc = &wc->w_desc[i]; 1292 1293 /* 1294 * We have to make sure that the total write passed in 1295 * doesn't extend past a single cluster. 1296 */ 1297 local_len = len; 1298 cluster_off = pos & (osb->s_clustersize - 1); 1299 if ((cluster_off + local_len) > osb->s_clustersize) 1300 local_len = osb->s_clustersize - cluster_off; 1301 1302 ret = ocfs2_write_cluster(mapping, desc->c_phys, 1303 desc->c_unwritten, 1304 desc->c_needs_zero, 1305 data_ac, meta_ac, 1306 wc, desc->c_cpos, pos, local_len); 1307 if (ret) { 1308 mlog_errno(ret); 1309 goto out; 1310 } 1311 1312 len -= local_len; 1313 pos += local_len; 1314 } 1315 1316 ret = 0; 1317out: 1318 return ret; 1319} 1320 1321/* 1322 * ocfs2_write_end() wants to know which parts of the target page it 1323 * should complete the write on. It's easiest to compute them ahead of 1324 * time when a more complete view of the write is available. 1325 */ 1326static void ocfs2_set_target_boundaries(struct ocfs2_super *osb, 1327 struct ocfs2_write_ctxt *wc, 1328 loff_t pos, unsigned len, int alloc) 1329{ 1330 struct ocfs2_write_cluster_desc *desc; 1331 1332 wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1); 1333 wc->w_target_to = wc->w_target_from + len; 1334 1335 if (alloc == 0) 1336 return; 1337 1338 /* 1339 * Allocating write - we may have different boundaries based 1340 * on page size and cluster size. 1341 * 1342 * NOTE: We can no longer compute one value from the other as 1343 * the actual write length and user provided length may be 1344 * different. 1345 */ 1346 1347 if (wc->w_large_pages) { 1348 /* 1349 * We only care about the 1st and last cluster within 1350 * our range and whether they should be zero'd or not. Either 1351 * value may be extended out to the start/end of a 1352 * newly allocated cluster. 1353 */ 1354 desc = &wc->w_desc[0]; 1355 if (desc->c_needs_zero) 1356 ocfs2_figure_cluster_boundaries(osb, 1357 desc->c_cpos, 1358 &wc->w_target_from, 1359 NULL); 1360 1361 desc = &wc->w_desc[wc->w_clen - 1]; 1362 if (desc->c_needs_zero) 1363 ocfs2_figure_cluster_boundaries(osb, 1364 desc->c_cpos, 1365 NULL, 1366 &wc->w_target_to); 1367 } else { 1368 wc->w_target_from = 0; 1369 wc->w_target_to = PAGE_CACHE_SIZE; 1370 } 1371} 1372 1373/* 1374 * Populate each single-cluster write descriptor in the write context 1375 * with information about the i/o to be done. 1376 * 1377 * Returns the number of clusters that will have to be allocated, as 1378 * well as a worst case estimate of the number of extent records that 1379 * would have to be created during a write to an unwritten region. 1380 */ 1381static int ocfs2_populate_write_desc(struct inode *inode, 1382 struct ocfs2_write_ctxt *wc, 1383 unsigned int *clusters_to_alloc, 1384 unsigned int *extents_to_split) 1385{ 1386 int ret; 1387 struct ocfs2_write_cluster_desc *desc; 1388 unsigned int num_clusters = 0; 1389 unsigned int ext_flags = 0; 1390 u32 phys = 0; 1391 int i; 1392 1393 *clusters_to_alloc = 0; 1394 *extents_to_split = 0; 1395 1396 for (i = 0; i < wc->w_clen; i++) { 1397 desc = &wc->w_desc[i]; 1398 desc->c_cpos = wc->w_cpos + i; 1399 1400 if (num_clusters == 0) { 1401 /* 1402 * Need to look up the next extent record. 1403 */ 1404 ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys, 1405 &num_clusters, &ext_flags); 1406 if (ret) { 1407 mlog_errno(ret); 1408 goto out; 1409 } 1410 1411 /* We should already CoW the refcountd extent. */ 1412 BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED); 1413 1414 /* 1415 * Assume worst case - that we're writing in 1416 * the middle of the extent. 1417 * 1418 * We can assume that the write proceeds from 1419 * left to right, in which case the extent 1420 * insert code is smart enough to coalesce the 1421 * next splits into the previous records created. 1422 */ 1423 if (ext_flags & OCFS2_EXT_UNWRITTEN) 1424 *extents_to_split = *extents_to_split + 2; 1425 } else if (phys) { 1426 /* 1427 * Only increment phys if it doesn't describe 1428 * a hole. 1429 */ 1430 phys++; 1431 } 1432 1433 /* 1434 * If w_first_new_cpos is < UINT_MAX, we have a non-sparse 1435 * file that got extended. w_first_new_cpos tells us 1436 * where the newly allocated clusters are so we can 1437 * zero them. 1438 */ 1439 if (desc->c_cpos >= wc->w_first_new_cpos) { 1440 BUG_ON(phys == 0); 1441 desc->c_needs_zero = 1; 1442 } 1443 1444 desc->c_phys = phys; 1445 if (phys == 0) { 1446 desc->c_new = 1; 1447 desc->c_needs_zero = 1; 1448 *clusters_to_alloc = *clusters_to_alloc + 1; 1449 } 1450 1451 if (ext_flags & OCFS2_EXT_UNWRITTEN) { 1452 desc->c_unwritten = 1; 1453 desc->c_needs_zero = 1; 1454 } 1455 1456 num_clusters--; 1457 } 1458 1459 ret = 0; 1460out: 1461 return ret; 1462} 1463 1464static int ocfs2_write_begin_inline(struct address_space *mapping, 1465 struct inode *inode, 1466 struct ocfs2_write_ctxt *wc) 1467{ 1468 int ret; 1469 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 1470 struct page *page; 1471 handle_t *handle; 1472 struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data; 1473 1474 page = find_or_create_page(mapping, 0, GFP_NOFS); 1475 if (!page) { 1476 ret = -ENOMEM; 1477 mlog_errno(ret); 1478 goto out; 1479 } 1480 /* 1481 * If we don't set w_num_pages then this page won't get unlocked 1482 * and freed on cleanup of the write context. 1483 */ 1484 wc->w_pages[0] = wc->w_target_page = page; 1485 wc->w_num_pages = 1; 1486 1487 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS); 1488 if (IS_ERR(handle)) { 1489 ret = PTR_ERR(handle); 1490 mlog_errno(ret); 1491 goto out; 1492 } 1493 1494 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh, 1495 OCFS2_JOURNAL_ACCESS_WRITE); 1496 if (ret) { 1497 ocfs2_commit_trans(osb, handle); 1498 1499 mlog_errno(ret); 1500 goto out; 1501 } 1502 1503 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)) 1504 ocfs2_set_inode_data_inline(inode, di); 1505 1506 if (!PageUptodate(page)) { 1507 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh); 1508 if (ret) { 1509 ocfs2_commit_trans(osb, handle); 1510 1511 goto out; 1512 } 1513 } 1514 1515 wc->w_handle = handle; 1516out: 1517 return ret; 1518} 1519 1520int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size) 1521{ 1522 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 1523 1524 if (new_size <= le16_to_cpu(di->id2.i_data.id_count)) 1525 return 1; 1526 return 0; 1527} 1528 1529static int ocfs2_try_to_write_inline_data(struct address_space *mapping, 1530 struct inode *inode, loff_t pos, 1531 unsigned len, struct page *mmap_page, 1532 struct ocfs2_write_ctxt *wc) 1533{ 1534 int ret, written = 0; 1535 loff_t end = pos + len; 1536 struct ocfs2_inode_info *oi = OCFS2_I(inode); 1537 struct ocfs2_dinode *di = NULL; 1538 1539 mlog(0, "Inode %llu, write of %u bytes at off %llu. features: 0x%x\n", 1540 (unsigned long long)oi->ip_blkno, len, (unsigned long long)pos, 1541 oi->ip_dyn_features); 1542 1543 /* 1544 * Handle inodes which already have inline data 1st. 1545 */ 1546 if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) { 1547 if (mmap_page == NULL && 1548 ocfs2_size_fits_inline_data(wc->w_di_bh, end)) 1549 goto do_inline_write; 1550 1551 /* 1552 * The write won't fit - we have to give this inode an 1553 * inline extent list now. 1554 */ 1555 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh); 1556 if (ret) 1557 mlog_errno(ret); 1558 goto out; 1559 } 1560 1561 /* 1562 * Check whether the inode can accept inline data. 1563 */ 1564 if (oi->ip_clusters != 0 || i_size_read(inode) != 0) 1565 return 0; 1566 1567 /* 1568 * Check whether the write can fit. 1569 */ 1570 di = (struct ocfs2_dinode *)wc->w_di_bh->b_data; 1571 if (mmap_page || 1572 end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) 1573 return 0; 1574 1575do_inline_write: 1576 ret = ocfs2_write_begin_inline(mapping, inode, wc); 1577 if (ret) { 1578 mlog_errno(ret); 1579 goto out; 1580 } 1581 1582 /* 1583 * This signals to the caller that the data can be written 1584 * inline. 1585 */ 1586 written = 1; 1587out: 1588 return written ? written : ret; 1589} 1590 1591/* 1592 * This function only does anything for file systems which can't 1593 * handle sparse files. 1594 * 1595 * What we want to do here is fill in any hole between the current end 1596 * of allocation and the end of our write. That way the rest of the 1597 * write path can treat it as an non-allocating write, which has no 1598 * special case code for sparse/nonsparse files. 1599 */ 1600static int ocfs2_expand_nonsparse_inode(struct inode *inode, 1601 struct buffer_head *di_bh, 1602 loff_t pos, unsigned len, 1603 struct ocfs2_write_ctxt *wc) 1604{ 1605 int ret; 1606 loff_t newsize = pos + len; 1607 1608 BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))); 1609 1610 if (newsize <= i_size_read(inode)) 1611 return 0; 1612 1613 ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos); 1614 if (ret) 1615 mlog_errno(ret); 1616 1617 wc->w_first_new_cpos = 1618 ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode)); 1619 1620 return ret; 1621} 1622 1623static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh, 1624 loff_t pos) 1625{ 1626 int ret = 0; 1627 1628 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))); 1629 if (pos > i_size_read(inode)) 1630 ret = ocfs2_zero_extend(inode, di_bh, pos); 1631 1632 return ret; 1633} 1634 1635int ocfs2_write_begin_nolock(struct address_space *mapping, 1636 loff_t pos, unsigned len, unsigned flags, 1637 struct page **pagep, void **fsdata, 1638 struct buffer_head *di_bh, struct page *mmap_page) 1639{ 1640 int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS; 1641 unsigned int clusters_to_alloc, extents_to_split; 1642 struct ocfs2_write_ctxt *wc; 1643 struct inode *inode = mapping->host; 1644 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 1645 struct ocfs2_dinode *di; 1646 struct ocfs2_alloc_context *data_ac = NULL; 1647 struct ocfs2_alloc_context *meta_ac = NULL; 1648 handle_t *handle; 1649 struct ocfs2_extent_tree et; 1650 1651 ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh); 1652 if (ret) { 1653 mlog_errno(ret); 1654 return ret; 1655 } 1656 1657 if (ocfs2_supports_inline_data(osb)) { 1658 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len, 1659 mmap_page, wc); 1660 if (ret == 1) { 1661 ret = 0; 1662 goto success; 1663 } 1664 if (ret < 0) { 1665 mlog_errno(ret); 1666 goto out; 1667 } 1668 } 1669 1670 if (ocfs2_sparse_alloc(osb)) 1671 ret = ocfs2_zero_tail(inode, di_bh, pos); 1672 else 1673 ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos, len, 1674 wc); 1675 if (ret) { 1676 mlog_errno(ret); 1677 goto out; 1678 } 1679 1680 ret = ocfs2_check_range_for_refcount(inode, pos, len); 1681 if (ret < 0) { 1682 mlog_errno(ret); 1683 goto out; 1684 } else if (ret == 1) { 1685 ret = ocfs2_refcount_cow(inode, di_bh, 1686 wc->w_cpos, wc->w_clen, UINT_MAX); 1687 if (ret) { 1688 mlog_errno(ret); 1689 goto out; 1690 } 1691 } 1692 1693 ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc, 1694 &extents_to_split); 1695 if (ret) { 1696 mlog_errno(ret); 1697 goto out; 1698 } 1699 1700 di = (struct ocfs2_dinode *)wc->w_di_bh->b_data; 1701 1702 /* 1703 * We set w_target_from, w_target_to here so that 1704 * ocfs2_write_end() knows which range in the target page to 1705 * write out. An allocation requires that we write the entire 1706 * cluster range. 1707 */ 1708 if (clusters_to_alloc || extents_to_split) { 1709 mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u," 1710 " clusters_to_add = %u, extents_to_split = %u\n", 1711 (unsigned long long)OCFS2_I(inode)->ip_blkno, 1712 (long long)i_size_read(inode), le32_to_cpu(di->i_clusters), 1713 clusters_to_alloc, extents_to_split); 1714 1715 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), 1716 wc->w_di_bh); 1717 ret = ocfs2_lock_allocators(inode, &et, 1718 clusters_to_alloc, extents_to_split, 1719 &data_ac, &meta_ac); 1720 if (ret) { 1721 mlog_errno(ret); 1722 goto out; 1723 } 1724 1725 if (data_ac) 1726 data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv; 1727 1728 credits = ocfs2_calc_extend_credits(inode->i_sb, 1729 &di->id2.i_list, 1730 clusters_to_alloc); 1731 1732 } 1733 1734 /* 1735 * We have to zero sparse allocated clusters, unwritten extent clusters, 1736 * and non-sparse clusters we just extended. For non-sparse writes, 1737 * we know zeros will only be needed in the first and/or last cluster. 1738 */ 1739 if (clusters_to_alloc || extents_to_split || 1740 (wc->w_clen && (wc->w_desc[0].c_needs_zero || 1741 wc->w_desc[wc->w_clen - 1].c_needs_zero))) 1742 cluster_of_pages = 1; 1743 else 1744 cluster_of_pages = 0; 1745 1746 ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages); 1747 1748 handle = ocfs2_start_trans(osb, credits); 1749 if (IS_ERR(handle)) { 1750 ret = PTR_ERR(handle); 1751 mlog_errno(ret); 1752 goto out; 1753 } 1754 1755 wc->w_handle = handle; 1756 1757 if (clusters_to_alloc) { 1758 ret = dquot_alloc_space_nodirty(inode, 1759 ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc)); 1760 if (ret) 1761 goto out_commit; 1762 } 1763 /* 1764 * We don't want this to fail in ocfs2_write_end(), so do it 1765 * here. 1766 */ 1767 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh, 1768 OCFS2_JOURNAL_ACCESS_WRITE); 1769 if (ret) { 1770 mlog_errno(ret); 1771 goto out_quota; 1772 } 1773 1774 /* 1775 * Fill our page array first. That way we've grabbed enough so 1776 * that we can zero and flush if we error after adding the 1777 * extent. 1778 */ 1779 ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len, 1780 cluster_of_pages, mmap_page); 1781 if (ret) { 1782 mlog_errno(ret); 1783 goto out_quota; 1784 } 1785 1786 ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos, 1787 len); 1788 if (ret) { 1789 mlog_errno(ret); 1790 goto out_quota; 1791 } 1792 1793 if (data_ac) 1794 ocfs2_free_alloc_context(data_ac); 1795 if (meta_ac) 1796 ocfs2_free_alloc_context(meta_ac); 1797 1798success: 1799 *pagep = wc->w_target_page; 1800 *fsdata = wc; 1801 return 0; 1802out_quota: 1803 if (clusters_to_alloc) 1804 dquot_free_space(inode, 1805 ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc)); 1806out_commit: 1807 ocfs2_commit_trans(osb, handle); 1808 1809out: 1810 ocfs2_free_write_ctxt(wc); 1811 1812 if (data_ac) 1813 ocfs2_free_alloc_context(data_ac); 1814 if (meta_ac) 1815 ocfs2_free_alloc_context(meta_ac); 1816 return ret; 1817} 1818 1819static int ocfs2_write_begin(struct file *file, struct address_space *mapping, 1820 loff_t pos, unsigned len, unsigned flags, 1821 struct page **pagep, void **fsdata) 1822{ 1823 int ret; 1824 struct buffer_head *di_bh = NULL; 1825 struct inode *inode = mapping->host; 1826 1827 ret = ocfs2_inode_lock(inode, &di_bh, 1); 1828 if (ret) { 1829 mlog_errno(ret); 1830 return ret; 1831 } 1832 1833 /* 1834 * Take alloc sem here to prevent concurrent lookups. That way 1835 * the mapping, zeroing and tree manipulation within 1836 * ocfs2_write() will be safe against ->readpage(). This 1837 * should also serve to lock out allocation from a shared 1838 * writeable region. 1839 */ 1840 down_write(&OCFS2_I(inode)->ip_alloc_sem); 1841 1842 ret = ocfs2_write_begin_nolock(mapping, pos, len, flags, pagep, 1843 fsdata, di_bh, NULL); 1844 if (ret) { 1845 mlog_errno(ret); 1846 goto out_fail; 1847 } 1848 1849 brelse(di_bh); 1850 1851 return 0; 1852 1853out_fail: 1854 up_write(&OCFS2_I(inode)->ip_alloc_sem); 1855 1856 brelse(di_bh); 1857 ocfs2_inode_unlock(inode, 1); 1858 1859 return ret; 1860} 1861 1862static void ocfs2_write_end_inline(struct inode *inode, loff_t pos, 1863 unsigned len, unsigned *copied, 1864 struct ocfs2_dinode *di, 1865 struct ocfs2_write_ctxt *wc) 1866{ 1867 void *kaddr; 1868 1869 if (unlikely(*copied < len)) { 1870 if (!PageUptodate(wc->w_target_page)) { 1871 *copied = 0; 1872 return; 1873 } 1874 } 1875 1876 kaddr = kmap_atomic(wc->w_target_page, KM_USER0); 1877 memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied); 1878 kunmap_atomic(kaddr, KM_USER0); 1879 1880 mlog(0, "Data written to inode at offset %llu. " 1881 "id_count = %u, copied = %u, i_dyn_features = 0x%x\n", 1882 (unsigned long long)pos, *copied, 1883 le16_to_cpu(di->id2.i_data.id_count), 1884 le16_to_cpu(di->i_dyn_features)); 1885} 1886 1887int ocfs2_write_end_nolock(struct address_space *mapping, 1888 loff_t pos, unsigned len, unsigned copied, 1889 struct page *page, void *fsdata) 1890{ 1891 int i; 1892 unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1); 1893 struct inode *inode = mapping->host; 1894 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 1895 struct ocfs2_write_ctxt *wc = fsdata; 1896 struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data; 1897 handle_t *handle = wc->w_handle; 1898 struct page *tmppage; 1899 1900 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) { 1901 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc); 1902 goto out_write_size; 1903 } 1904 1905 if (unlikely(copied < len)) { 1906 if (!PageUptodate(wc->w_target_page)) 1907 copied = 0; 1908 1909 ocfs2_zero_new_buffers(wc->w_target_page, start+copied, 1910 start+len); 1911 } 1912 flush_dcache_page(wc->w_target_page); 1913 1914 for(i = 0; i < wc->w_num_pages; i++) { 1915 tmppage = wc->w_pages[i]; 1916 1917 if (tmppage == wc->w_target_page) { 1918 from = wc->w_target_from; 1919 to = wc->w_target_to; 1920 1921 BUG_ON(from > PAGE_CACHE_SIZE || 1922 to > PAGE_CACHE_SIZE || 1923 to < from); 1924 } else { 1925 /* 1926 * Pages adjacent to the target (if any) imply 1927 * a hole-filling write in which case we want 1928 * to flush their entire range. 1929 */ 1930 from = 0; 1931 to = PAGE_CACHE_SIZE; 1932 } 1933 1934 if (page_has_buffers(tmppage)) { 1935 if (ocfs2_should_order_data(inode)) 1936 ocfs2_jbd2_file_inode(wc->w_handle, inode); 1937 block_commit_write(tmppage, from, to); 1938 } 1939 } 1940 1941out_write_size: 1942 pos += copied; 1943 if (pos > inode->i_size) { 1944 i_size_write(inode, pos); 1945 mark_inode_dirty(inode); 1946 } 1947 inode->i_blocks = ocfs2_inode_sector_count(inode); 1948 di->i_size = cpu_to_le64((u64)i_size_read(inode)); 1949 inode->i_mtime = inode->i_ctime = CURRENT_TIME; 1950 di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec); 1951 di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec); 1952 ocfs2_journal_dirty(handle, wc->w_di_bh); 1953 1954 ocfs2_commit_trans(osb, handle); 1955 1956 ocfs2_run_deallocs(osb, &wc->w_dealloc); 1957 1958 ocfs2_free_write_ctxt(wc); 1959 1960 return copied; 1961} 1962 1963static int ocfs2_write_end(struct file *file, struct address_space *mapping, 1964 loff_t pos, unsigned len, unsigned copied, 1965 struct page *page, void *fsdata) 1966{ 1967 int ret; 1968 struct inode *inode = mapping->host; 1969 1970 ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata); 1971 1972 up_write(&OCFS2_I(inode)->ip_alloc_sem); 1973 ocfs2_inode_unlock(inode, 1); 1974 1975 return ret; 1976} 1977 1978const struct address_space_operations ocfs2_aops = { 1979 .readpage = ocfs2_readpage, 1980 .readpages = ocfs2_readpages, 1981 .writepage = ocfs2_writepage, 1982 .write_begin = ocfs2_write_begin, 1983 .write_end = ocfs2_write_end, 1984 .bmap = ocfs2_bmap, 1985 .sync_page = block_sync_page, 1986 .direct_IO = ocfs2_direct_IO, 1987 .invalidatepage = ocfs2_invalidatepage, 1988 .releasepage = ocfs2_releasepage, 1989 .migratepage = buffer_migrate_page, 1990 .is_partially_uptodate = block_is_partially_uptodate, 1991 .error_remove_page = generic_error_remove_page, 1992}; 1993