aops.c revision 49cb8d2d496ce06869ccca2ab368ed6b0b5b979d
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 30#define MLOG_MASK_PREFIX ML_FILE_IO 31#include <cluster/masklog.h> 32 33#include "ocfs2.h" 34 35#include "alloc.h" 36#include "aops.h" 37#include "dlmglue.h" 38#include "extent_map.h" 39#include "file.h" 40#include "inode.h" 41#include "journal.h" 42#include "suballoc.h" 43#include "super.h" 44#include "symlink.h" 45 46#include "buffer_head_io.h" 47 48static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock, 49 struct buffer_head *bh_result, int create) 50{ 51 int err = -EIO; 52 int status; 53 struct ocfs2_dinode *fe = NULL; 54 struct buffer_head *bh = NULL; 55 struct buffer_head *buffer_cache_bh = NULL; 56 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 57 void *kaddr; 58 59 mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode, 60 (unsigned long long)iblock, bh_result, create); 61 62 BUG_ON(ocfs2_inode_is_fast_symlink(inode)); 63 64 if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) { 65 mlog(ML_ERROR, "block offset > PATH_MAX: %llu", 66 (unsigned long long)iblock); 67 goto bail; 68 } 69 70 status = ocfs2_read_block(OCFS2_SB(inode->i_sb), 71 OCFS2_I(inode)->ip_blkno, 72 &bh, OCFS2_BH_CACHED, inode); 73 if (status < 0) { 74 mlog_errno(status); 75 goto bail; 76 } 77 fe = (struct ocfs2_dinode *) bh->b_data; 78 79 if (!OCFS2_IS_VALID_DINODE(fe)) { 80 mlog(ML_ERROR, "Invalid dinode #%llu: signature = %.*s\n", 81 (unsigned long long)fe->i_blkno, 7, fe->i_signature); 82 goto bail; 83 } 84 85 if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb, 86 le32_to_cpu(fe->i_clusters))) { 87 mlog(ML_ERROR, "block offset is outside the allocated size: " 88 "%llu\n", (unsigned long long)iblock); 89 goto bail; 90 } 91 92 /* We don't use the page cache to create symlink data, so if 93 * need be, copy it over from the buffer cache. */ 94 if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) { 95 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + 96 iblock; 97 buffer_cache_bh = sb_getblk(osb->sb, blkno); 98 if (!buffer_cache_bh) { 99 mlog(ML_ERROR, "couldn't getblock for symlink!\n"); 100 goto bail; 101 } 102 103 /* we haven't locked out transactions, so a commit 104 * could've happened. Since we've got a reference on 105 * the bh, even if it commits while we're doing the 106 * copy, the data is still good. */ 107 if (buffer_jbd(buffer_cache_bh) 108 && ocfs2_inode_is_new(inode)) { 109 kaddr = kmap_atomic(bh_result->b_page, KM_USER0); 110 if (!kaddr) { 111 mlog(ML_ERROR, "couldn't kmap!\n"); 112 goto bail; 113 } 114 memcpy(kaddr + (bh_result->b_size * iblock), 115 buffer_cache_bh->b_data, 116 bh_result->b_size); 117 kunmap_atomic(kaddr, KM_USER0); 118 set_buffer_uptodate(bh_result); 119 } 120 brelse(buffer_cache_bh); 121 } 122 123 map_bh(bh_result, inode->i_sb, 124 le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock); 125 126 err = 0; 127 128bail: 129 if (bh) 130 brelse(bh); 131 132 mlog_exit(err); 133 return err; 134} 135 136static int ocfs2_get_block(struct inode *inode, sector_t iblock, 137 struct buffer_head *bh_result, int create) 138{ 139 int err = 0; 140 unsigned int ext_flags; 141 u64 p_blkno, past_eof; 142 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 143 144 mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode, 145 (unsigned long long)iblock, bh_result, create); 146 147 if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE) 148 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n", 149 inode, inode->i_ino); 150 151 if (S_ISLNK(inode->i_mode)) { 152 /* this always does I/O for some reason. */ 153 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create); 154 goto bail; 155 } 156 157 err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, NULL, 158 &ext_flags); 159 if (err) { 160 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, " 161 "%llu, NULL)\n", err, inode, (unsigned long long)iblock, 162 (unsigned long long)p_blkno); 163 goto bail; 164 } 165 166 /* 167 * ocfs2 never allocates in this function - the only time we 168 * need to use BH_New is when we're extending i_size on a file 169 * system which doesn't support holes, in which case BH_New 170 * allows block_prepare_write() to zero. 171 */ 172 mlog_bug_on_msg(create && p_blkno == 0 && ocfs2_sparse_alloc(osb), 173 "ino %lu, iblock %llu\n", inode->i_ino, 174 (unsigned long long)iblock); 175 176 /* Treat the unwritten extent as a hole for zeroing purposes. */ 177 if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN)) 178 map_bh(bh_result, inode->i_sb, p_blkno); 179 180 if (!ocfs2_sparse_alloc(osb)) { 181 if (p_blkno == 0) { 182 err = -EIO; 183 mlog(ML_ERROR, 184 "iblock = %llu p_blkno = %llu blkno=(%llu)\n", 185 (unsigned long long)iblock, 186 (unsigned long long)p_blkno, 187 (unsigned long long)OCFS2_I(inode)->ip_blkno); 188 mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters); 189 dump_stack(); 190 } 191 192 past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode)); 193 mlog(0, "Inode %lu, past_eof = %llu\n", inode->i_ino, 194 (unsigned long long)past_eof); 195 196 if (create && (iblock >= past_eof)) 197 set_buffer_new(bh_result); 198 } 199 200bail: 201 if (err < 0) 202 err = -EIO; 203 204 mlog_exit(err); 205 return err; 206} 207 208static int ocfs2_readpage(struct file *file, struct page *page) 209{ 210 struct inode *inode = page->mapping->host; 211 loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT; 212 int ret, unlock = 1; 213 214 mlog_entry("(0x%p, %lu)\n", file, (page ? page->index : 0)); 215 216 ret = ocfs2_meta_lock_with_page(inode, NULL, 0, page); 217 if (ret != 0) { 218 if (ret == AOP_TRUNCATED_PAGE) 219 unlock = 0; 220 mlog_errno(ret); 221 goto out; 222 } 223 224 down_read(&OCFS2_I(inode)->ip_alloc_sem); 225 226 /* 227 * i_size might have just been updated as we grabed the meta lock. We 228 * might now be discovering a truncate that hit on another node. 229 * block_read_full_page->get_block freaks out if it is asked to read 230 * beyond the end of a file, so we check here. Callers 231 * (generic_file_read, fault->nopage) are clever enough to check i_size 232 * and notice that the page they just read isn't needed. 233 * 234 * XXX sys_readahead() seems to get that wrong? 235 */ 236 if (start >= i_size_read(inode)) { 237 char *addr = kmap(page); 238 memset(addr, 0, PAGE_SIZE); 239 flush_dcache_page(page); 240 kunmap(page); 241 SetPageUptodate(page); 242 ret = 0; 243 goto out_alloc; 244 } 245 246 ret = ocfs2_data_lock_with_page(inode, 0, page); 247 if (ret != 0) { 248 if (ret == AOP_TRUNCATED_PAGE) 249 unlock = 0; 250 mlog_errno(ret); 251 goto out_alloc; 252 } 253 254 ret = block_read_full_page(page, ocfs2_get_block); 255 unlock = 0; 256 257 ocfs2_data_unlock(inode, 0); 258out_alloc: 259 up_read(&OCFS2_I(inode)->ip_alloc_sem); 260 ocfs2_meta_unlock(inode, 0); 261out: 262 if (unlock) 263 unlock_page(page); 264 mlog_exit(ret); 265 return ret; 266} 267 268/* Note: Because we don't support holes, our allocation has 269 * already happened (allocation writes zeros to the file data) 270 * so we don't have to worry about ordered writes in 271 * ocfs2_writepage. 272 * 273 * ->writepage is called during the process of invalidating the page cache 274 * during blocked lock processing. It can't block on any cluster locks 275 * to during block mapping. It's relying on the fact that the block 276 * mapping can't have disappeared under the dirty pages that it is 277 * being asked to write back. 278 */ 279static int ocfs2_writepage(struct page *page, struct writeback_control *wbc) 280{ 281 int ret; 282 283 mlog_entry("(0x%p)\n", page); 284 285 ret = block_write_full_page(page, ocfs2_get_block, wbc); 286 287 mlog_exit(ret); 288 289 return ret; 290} 291 292/* 293 * This is called from ocfs2_write_zero_page() which has handled it's 294 * own cluster locking and has ensured allocation exists for those 295 * blocks to be written. 296 */ 297int ocfs2_prepare_write_nolock(struct inode *inode, struct page *page, 298 unsigned from, unsigned to) 299{ 300 int ret; 301 302 down_read(&OCFS2_I(inode)->ip_alloc_sem); 303 304 ret = block_prepare_write(page, from, to, ocfs2_get_block); 305 306 up_read(&OCFS2_I(inode)->ip_alloc_sem); 307 308 return ret; 309} 310 311/* Taken from ext3. We don't necessarily need the full blown 312 * functionality yet, but IMHO it's better to cut and paste the whole 313 * thing so we can avoid introducing our own bugs (and easily pick up 314 * their fixes when they happen) --Mark */ 315int walk_page_buffers( handle_t *handle, 316 struct buffer_head *head, 317 unsigned from, 318 unsigned to, 319 int *partial, 320 int (*fn)( handle_t *handle, 321 struct buffer_head *bh)) 322{ 323 struct buffer_head *bh; 324 unsigned block_start, block_end; 325 unsigned blocksize = head->b_size; 326 int err, ret = 0; 327 struct buffer_head *next; 328 329 for ( bh = head, block_start = 0; 330 ret == 0 && (bh != head || !block_start); 331 block_start = block_end, bh = next) 332 { 333 next = bh->b_this_page; 334 block_end = block_start + blocksize; 335 if (block_end <= from || block_start >= to) { 336 if (partial && !buffer_uptodate(bh)) 337 *partial = 1; 338 continue; 339 } 340 err = (*fn)(handle, bh); 341 if (!ret) 342 ret = err; 343 } 344 return ret; 345} 346 347handle_t *ocfs2_start_walk_page_trans(struct inode *inode, 348 struct page *page, 349 unsigned from, 350 unsigned to) 351{ 352 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 353 handle_t *handle = NULL; 354 int ret = 0; 355 356 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS); 357 if (!handle) { 358 ret = -ENOMEM; 359 mlog_errno(ret); 360 goto out; 361 } 362 363 if (ocfs2_should_order_data(inode)) { 364 ret = walk_page_buffers(handle, 365 page_buffers(page), 366 from, to, NULL, 367 ocfs2_journal_dirty_data); 368 if (ret < 0) 369 mlog_errno(ret); 370 } 371out: 372 if (ret) { 373 if (handle) 374 ocfs2_commit_trans(osb, handle); 375 handle = ERR_PTR(ret); 376 } 377 return handle; 378} 379 380static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block) 381{ 382 sector_t status; 383 u64 p_blkno = 0; 384 int err = 0; 385 struct inode *inode = mapping->host; 386 387 mlog_entry("(block = %llu)\n", (unsigned long long)block); 388 389 /* We don't need to lock journal system files, since they aren't 390 * accessed concurrently from multiple nodes. 391 */ 392 if (!INODE_JOURNAL(inode)) { 393 err = ocfs2_meta_lock(inode, NULL, 0); 394 if (err) { 395 if (err != -ENOENT) 396 mlog_errno(err); 397 goto bail; 398 } 399 down_read(&OCFS2_I(inode)->ip_alloc_sem); 400 } 401 402 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL, NULL); 403 404 if (!INODE_JOURNAL(inode)) { 405 up_read(&OCFS2_I(inode)->ip_alloc_sem); 406 ocfs2_meta_unlock(inode, 0); 407 } 408 409 if (err) { 410 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n", 411 (unsigned long long)block); 412 mlog_errno(err); 413 goto bail; 414 } 415 416 417bail: 418 status = err ? 0 : p_blkno; 419 420 mlog_exit((int)status); 421 422 return status; 423} 424 425/* 426 * TODO: Make this into a generic get_blocks function. 427 * 428 * From do_direct_io in direct-io.c: 429 * "So what we do is to permit the ->get_blocks function to populate 430 * bh.b_size with the size of IO which is permitted at this offset and 431 * this i_blkbits." 432 * 433 * This function is called directly from get_more_blocks in direct-io.c. 434 * 435 * called like this: dio->get_blocks(dio->inode, fs_startblk, 436 * fs_count, map_bh, dio->rw == WRITE); 437 */ 438static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock, 439 struct buffer_head *bh_result, int create) 440{ 441 int ret; 442 u64 p_blkno, inode_blocks; 443 int contig_blocks; 444 unsigned int ext_flags; 445 unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits; 446 unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits; 447 448 /* This function won't even be called if the request isn't all 449 * nicely aligned and of the right size, so there's no need 450 * for us to check any of that. */ 451 452 inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode)); 453 454 /* 455 * Any write past EOF is not allowed because we'd be extending. 456 */ 457 if (create && (iblock + max_blocks) > inode_blocks) { 458 ret = -EIO; 459 goto bail; 460 } 461 462 /* This figures out the size of the next contiguous block, and 463 * our logical offset */ 464 ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, 465 &contig_blocks, &ext_flags); 466 if (ret) { 467 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n", 468 (unsigned long long)iblock); 469 ret = -EIO; 470 goto bail; 471 } 472 473 if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)) && !p_blkno) { 474 ocfs2_error(inode->i_sb, 475 "Inode %llu has a hole at block %llu\n", 476 (unsigned long long)OCFS2_I(inode)->ip_blkno, 477 (unsigned long long)iblock); 478 ret = -EROFS; 479 goto bail; 480 } 481 482 /* 483 * get_more_blocks() expects us to describe a hole by clearing 484 * the mapped bit on bh_result(). 485 * 486 * Consider an unwritten extent as a hole. 487 */ 488 if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN)) 489 map_bh(bh_result, inode->i_sb, p_blkno); 490 else { 491 /* 492 * ocfs2_prepare_inode_for_write() should have caught 493 * the case where we'd be filling a hole and triggered 494 * a buffered write instead. 495 */ 496 if (create) { 497 ret = -EIO; 498 mlog_errno(ret); 499 goto bail; 500 } 501 502 clear_buffer_mapped(bh_result); 503 } 504 505 /* make sure we don't map more than max_blocks blocks here as 506 that's all the kernel will handle at this point. */ 507 if (max_blocks < contig_blocks) 508 contig_blocks = max_blocks; 509 bh_result->b_size = contig_blocks << blocksize_bits; 510bail: 511 return ret; 512} 513 514/* 515 * ocfs2_dio_end_io is called by the dio core when a dio is finished. We're 516 * particularly interested in the aio/dio case. Like the core uses 517 * i_alloc_sem, we use the rw_lock DLM lock to protect io on one node from 518 * truncation on another. 519 */ 520static void ocfs2_dio_end_io(struct kiocb *iocb, 521 loff_t offset, 522 ssize_t bytes, 523 void *private) 524{ 525 struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode; 526 527 /* this io's submitter should not have unlocked this before we could */ 528 BUG_ON(!ocfs2_iocb_is_rw_locked(iocb)); 529 ocfs2_iocb_clear_rw_locked(iocb); 530 up_read(&inode->i_alloc_sem); 531 ocfs2_rw_unlock(inode, 0); 532} 533 534/* 535 * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen 536 * from ext3. PageChecked() bits have been removed as OCFS2 does not 537 * do journalled data. 538 */ 539static void ocfs2_invalidatepage(struct page *page, unsigned long offset) 540{ 541 journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal; 542 543 journal_invalidatepage(journal, page, offset); 544} 545 546static int ocfs2_releasepage(struct page *page, gfp_t wait) 547{ 548 journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal; 549 550 if (!page_has_buffers(page)) 551 return 0; 552 return journal_try_to_free_buffers(journal, page, wait); 553} 554 555static ssize_t ocfs2_direct_IO(int rw, 556 struct kiocb *iocb, 557 const struct iovec *iov, 558 loff_t offset, 559 unsigned long nr_segs) 560{ 561 struct file *file = iocb->ki_filp; 562 struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host; 563 int ret; 564 565 mlog_entry_void(); 566 567 if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) { 568 /* 569 * We get PR data locks even for O_DIRECT. This 570 * allows concurrent O_DIRECT I/O but doesn't let 571 * O_DIRECT with extending and buffered zeroing writes 572 * race. If they did race then the buffered zeroing 573 * could be written back after the O_DIRECT I/O. It's 574 * one thing to tell people not to mix buffered and 575 * O_DIRECT writes, but expecting them to understand 576 * that file extension is also an implicit buffered 577 * write is too much. By getting the PR we force 578 * writeback of the buffered zeroing before 579 * proceeding. 580 */ 581 ret = ocfs2_data_lock(inode, 0); 582 if (ret < 0) { 583 mlog_errno(ret); 584 goto out; 585 } 586 ocfs2_data_unlock(inode, 0); 587 } 588 589 ret = blockdev_direct_IO_no_locking(rw, iocb, inode, 590 inode->i_sb->s_bdev, iov, offset, 591 nr_segs, 592 ocfs2_direct_IO_get_blocks, 593 ocfs2_dio_end_io); 594out: 595 mlog_exit(ret); 596 return ret; 597} 598 599static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb, 600 u32 cpos, 601 unsigned int *start, 602 unsigned int *end) 603{ 604 unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE; 605 606 if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) { 607 unsigned int cpp; 608 609 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits); 610 611 cluster_start = cpos % cpp; 612 cluster_start = cluster_start << osb->s_clustersize_bits; 613 614 cluster_end = cluster_start + osb->s_clustersize; 615 } 616 617 BUG_ON(cluster_start > PAGE_SIZE); 618 BUG_ON(cluster_end > PAGE_SIZE); 619 620 if (start) 621 *start = cluster_start; 622 if (end) 623 *end = cluster_end; 624} 625 626/* 627 * 'from' and 'to' are the region in the page to avoid zeroing. 628 * 629 * If pagesize > clustersize, this function will avoid zeroing outside 630 * of the cluster boundary. 631 * 632 * from == to == 0 is code for "zero the entire cluster region" 633 */ 634static void ocfs2_clear_page_regions(struct page *page, 635 struct ocfs2_super *osb, u32 cpos, 636 unsigned from, unsigned to) 637{ 638 void *kaddr; 639 unsigned int cluster_start, cluster_end; 640 641 ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end); 642 643 kaddr = kmap_atomic(page, KM_USER0); 644 645 if (from || to) { 646 if (from > cluster_start) 647 memset(kaddr + cluster_start, 0, from - cluster_start); 648 if (to < cluster_end) 649 memset(kaddr + to, 0, cluster_end - to); 650 } else { 651 memset(kaddr + cluster_start, 0, cluster_end - cluster_start); 652 } 653 654 kunmap_atomic(kaddr, KM_USER0); 655} 656 657/* 658 * Some of this taken from block_prepare_write(). We already have our 659 * mapping by now though, and the entire write will be allocating or 660 * it won't, so not much need to use BH_New. 661 * 662 * This will also skip zeroing, which is handled externally. 663 */ 664int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno, 665 struct inode *inode, unsigned int from, 666 unsigned int to, int new) 667{ 668 int ret = 0; 669 struct buffer_head *head, *bh, *wait[2], **wait_bh = wait; 670 unsigned int block_end, block_start; 671 unsigned int bsize = 1 << inode->i_blkbits; 672 673 if (!page_has_buffers(page)) 674 create_empty_buffers(page, bsize, 0); 675 676 head = page_buffers(page); 677 for (bh = head, block_start = 0; bh != head || !block_start; 678 bh = bh->b_this_page, block_start += bsize) { 679 block_end = block_start + bsize; 680 681 /* 682 * Ignore blocks outside of our i/o range - 683 * they may belong to unallocated clusters. 684 */ 685 if (block_start >= to || block_end <= from) { 686 if (PageUptodate(page)) 687 set_buffer_uptodate(bh); 688 continue; 689 } 690 691 /* 692 * For an allocating write with cluster size >= page 693 * size, we always write the entire page. 694 */ 695 696 if (buffer_new(bh)) 697 clear_buffer_new(bh); 698 699 if (!buffer_mapped(bh)) { 700 map_bh(bh, inode->i_sb, *p_blkno); 701 unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr); 702 } 703 704 if (PageUptodate(page)) { 705 if (!buffer_uptodate(bh)) 706 set_buffer_uptodate(bh); 707 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) && 708 (block_start < from || block_end > to)) { 709 ll_rw_block(READ, 1, &bh); 710 *wait_bh++=bh; 711 } 712 713 *p_blkno = *p_blkno + 1; 714 } 715 716 /* 717 * If we issued read requests - let them complete. 718 */ 719 while(wait_bh > wait) { 720 wait_on_buffer(*--wait_bh); 721 if (!buffer_uptodate(*wait_bh)) 722 ret = -EIO; 723 } 724 725 if (ret == 0 || !new) 726 return ret; 727 728 /* 729 * If we get -EIO above, zero out any newly allocated blocks 730 * to avoid exposing stale data. 731 */ 732 bh = head; 733 block_start = 0; 734 do { 735 void *kaddr; 736 737 block_end = block_start + bsize; 738 if (block_end <= from) 739 goto next_bh; 740 if (block_start >= to) 741 break; 742 743 kaddr = kmap_atomic(page, KM_USER0); 744 memset(kaddr+block_start, 0, bh->b_size); 745 flush_dcache_page(page); 746 kunmap_atomic(kaddr, KM_USER0); 747 set_buffer_uptodate(bh); 748 mark_buffer_dirty(bh); 749 750next_bh: 751 block_start = block_end; 752 bh = bh->b_this_page; 753 } while (bh != head); 754 755 return ret; 756} 757 758/* 759 * This will copy user data from the buffer page in the splice 760 * context. 761 * 762 * For now, we ignore SPLICE_F_MOVE as that would require some extra 763 * communication out all the way to ocfs2_write(). 764 */ 765int ocfs2_map_and_write_splice_data(struct inode *inode, 766 struct ocfs2_write_ctxt *wc, u64 *p_blkno, 767 unsigned int *ret_from, unsigned int *ret_to) 768{ 769 int ret; 770 unsigned int to, from, cluster_start, cluster_end; 771 char *src, *dst; 772 struct ocfs2_splice_write_priv *sp = wc->w_private; 773 struct pipe_buffer *buf = sp->s_buf; 774 unsigned long bytes, src_from; 775 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 776 777 ocfs2_figure_cluster_boundaries(osb, wc->w_cpos, &cluster_start, 778 &cluster_end); 779 780 from = sp->s_offset; 781 src_from = sp->s_buf_offset; 782 bytes = wc->w_count; 783 784 if (wc->w_large_pages) { 785 /* 786 * For cluster size < page size, we have to 787 * calculate pos within the cluster and obey 788 * the rightmost boundary. 789 */ 790 bytes = min(bytes, (unsigned long)(osb->s_clustersize 791 - (wc->w_pos & (osb->s_clustersize - 1)))); 792 } 793 to = from + bytes; 794 795 if (wc->w_this_page_new) 796 ret = ocfs2_map_page_blocks(wc->w_this_page, p_blkno, inode, 797 cluster_start, cluster_end, 1); 798 else 799 ret = ocfs2_map_page_blocks(wc->w_this_page, p_blkno, inode, 800 from, to, 0); 801 if (ret) { 802 mlog_errno(ret); 803 goto out; 804 } 805 806 BUG_ON(from > PAGE_CACHE_SIZE); 807 BUG_ON(to > PAGE_CACHE_SIZE); 808 BUG_ON(from > osb->s_clustersize); 809 BUG_ON(to > osb->s_clustersize); 810 811 src = buf->ops->map(sp->s_pipe, buf, 1); 812 dst = kmap_atomic(wc->w_this_page, KM_USER1); 813 memcpy(dst + from, src + src_from, bytes); 814 kunmap_atomic(wc->w_this_page, KM_USER1); 815 buf->ops->unmap(sp->s_pipe, buf, src); 816 817 wc->w_finished_copy = 1; 818 819 *ret_from = from; 820 *ret_to = to; 821out: 822 823 return bytes ? (unsigned int)bytes : ret; 824} 825 826/* 827 * This will copy user data from the iovec in the buffered write 828 * context. 829 */ 830int ocfs2_map_and_write_user_data(struct inode *inode, 831 struct ocfs2_write_ctxt *wc, u64 *p_blkno, 832 unsigned int *ret_from, unsigned int *ret_to) 833{ 834 int ret; 835 unsigned int to, from, cluster_start, cluster_end; 836 unsigned long bytes, src_from; 837 char *dst; 838 struct ocfs2_buffered_write_priv *bp = wc->w_private; 839 const struct iovec *cur_iov = bp->b_cur_iov; 840 char __user *buf; 841 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 842 843 ocfs2_figure_cluster_boundaries(osb, wc->w_cpos, &cluster_start, 844 &cluster_end); 845 846 buf = cur_iov->iov_base + bp->b_cur_off; 847 src_from = (unsigned long)buf & ~PAGE_CACHE_MASK; 848 849 from = wc->w_pos & (PAGE_CACHE_SIZE - 1); 850 851 /* 852 * This is a lot of comparisons, but it reads quite 853 * easily, which is important here. 854 */ 855 /* Stay within the src page */ 856 bytes = PAGE_SIZE - src_from; 857 /* Stay within the vector */ 858 bytes = min(bytes, 859 (unsigned long)(cur_iov->iov_len - bp->b_cur_off)); 860 /* Stay within count */ 861 bytes = min(bytes, (unsigned long)wc->w_count); 862 /* 863 * For clustersize > page size, just stay within 864 * target page, otherwise we have to calculate pos 865 * within the cluster and obey the rightmost 866 * boundary. 867 */ 868 if (wc->w_large_pages) { 869 /* 870 * For cluster size < page size, we have to 871 * calculate pos within the cluster and obey 872 * the rightmost boundary. 873 */ 874 bytes = min(bytes, (unsigned long)(osb->s_clustersize 875 - (wc->w_pos & (osb->s_clustersize - 1)))); 876 } else { 877 /* 878 * cluster size > page size is the most common 879 * case - we just stay within the target page 880 * boundary. 881 */ 882 bytes = min(bytes, PAGE_CACHE_SIZE - from); 883 } 884 885 to = from + bytes; 886 887 if (wc->w_this_page_new) 888 ret = ocfs2_map_page_blocks(wc->w_this_page, p_blkno, inode, 889 cluster_start, cluster_end, 1); 890 else 891 ret = ocfs2_map_page_blocks(wc->w_this_page, p_blkno, inode, 892 from, to, 0); 893 if (ret) { 894 mlog_errno(ret); 895 goto out; 896 } 897 898 BUG_ON(from > PAGE_CACHE_SIZE); 899 BUG_ON(to > PAGE_CACHE_SIZE); 900 BUG_ON(from > osb->s_clustersize); 901 BUG_ON(to > osb->s_clustersize); 902 903 dst = kmap(wc->w_this_page); 904 memcpy(dst + from, bp->b_src_buf + src_from, bytes); 905 kunmap(wc->w_this_page); 906 907 /* 908 * XXX: This is slow, but simple. The caller of 909 * ocfs2_buffered_write_cluster() is responsible for 910 * passing through the iovecs, so it's difficult to 911 * predict what our next step is in here after our 912 * initial write. A future version should be pushing 913 * that iovec manipulation further down. 914 * 915 * By setting this, we indicate that a copy from user 916 * data was done, and subsequent calls for this 917 * cluster will skip copying more data. 918 */ 919 wc->w_finished_copy = 1; 920 921 *ret_from = from; 922 *ret_to = to; 923out: 924 925 return bytes ? (unsigned int)bytes : ret; 926} 927 928/* 929 * Map, fill and write a page to disk. 930 * 931 * The work of copying data is done via callback. Newly allocated 932 * pages which don't take user data will be zero'd (set 'new' to 933 * indicate an allocating write) 934 * 935 * Returns a negative error code or the number of bytes copied into 936 * the page. 937 */ 938int ocfs2_write_data_page(struct inode *inode, handle_t *handle, 939 u64 *p_blkno, struct page *page, 940 struct ocfs2_write_ctxt *wc, int new) 941{ 942 int ret, copied = 0; 943 unsigned int from = 0, to = 0; 944 unsigned int cluster_start, cluster_end; 945 unsigned int zero_from = 0, zero_to = 0; 946 947 ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), wc->w_cpos, 948 &cluster_start, &cluster_end); 949 950 if ((wc->w_pos >> PAGE_CACHE_SHIFT) == page->index 951 && !wc->w_finished_copy) { 952 953 wc->w_this_page = page; 954 wc->w_this_page_new = new; 955 ret = wc->w_write_data_page(inode, wc, p_blkno, &from, &to); 956 if (ret < 0) { 957 mlog_errno(ret); 958 goto out; 959 } 960 961 copied = ret; 962 963 zero_from = from; 964 zero_to = to; 965 if (new) { 966 from = cluster_start; 967 to = cluster_end; 968 } 969 } else { 970 /* 971 * If we haven't allocated the new page yet, we 972 * shouldn't be writing it out without copying user 973 * data. This is likely a math error from the caller. 974 */ 975 BUG_ON(!new); 976 977 from = cluster_start; 978 to = cluster_end; 979 980 ret = ocfs2_map_page_blocks(page, p_blkno, inode, 981 cluster_start, cluster_end, 1); 982 if (ret) { 983 mlog_errno(ret); 984 goto out; 985 } 986 } 987 988 /* 989 * Parts of newly allocated pages need to be zero'd. 990 * 991 * Above, we have also rewritten 'to' and 'from' - as far as 992 * the rest of the function is concerned, the entire cluster 993 * range inside of a page needs to be written. 994 * 995 * We can skip this if the page is up to date - it's already 996 * been zero'd from being read in as a hole. 997 */ 998 if (new && !PageUptodate(page)) 999 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb), 1000 wc->w_cpos, zero_from, zero_to); 1001 1002 flush_dcache_page(page); 1003 1004 if (ocfs2_should_order_data(inode)) { 1005 ret = walk_page_buffers(handle, 1006 page_buffers(page), 1007 from, to, NULL, 1008 ocfs2_journal_dirty_data); 1009 if (ret < 0) 1010 mlog_errno(ret); 1011 } 1012 1013 /* 1014 * We don't use generic_commit_write() because we need to 1015 * handle our own i_size update. 1016 */ 1017 ret = block_commit_write(page, from, to); 1018 if (ret) 1019 mlog_errno(ret); 1020out: 1021 1022 return copied ? copied : ret; 1023} 1024 1025/* 1026 * Do the actual write of some data into an inode. Optionally allocate 1027 * in order to fulfill the write. 1028 * 1029 * cpos is the logical cluster offset within the file to write at 1030 * 1031 * 'phys' is the physical mapping of that offset. a 'phys' value of 1032 * zero indicates that allocation is required. In this case, data_ac 1033 * and meta_ac should be valid (meta_ac can be null if metadata 1034 * allocation isn't required). 1035 */ 1036static ssize_t ocfs2_write(struct file *file, u32 phys, handle_t *handle, 1037 struct buffer_head *di_bh, 1038 struct ocfs2_alloc_context *data_ac, 1039 struct ocfs2_alloc_context *meta_ac, 1040 struct ocfs2_write_ctxt *wc) 1041{ 1042 int ret, i, numpages = 1, new; 1043 unsigned int copied = 0; 1044 u32 tmp_pos; 1045 u64 v_blkno, p_blkno; 1046 struct address_space *mapping = file->f_mapping; 1047 struct inode *inode = mapping->host; 1048 unsigned long index, start; 1049 struct page **cpages; 1050 1051 new = phys == 0 ? 1 : 0; 1052 1053 /* 1054 * Figure out how many pages we'll be manipulating here. For 1055 * non allocating write, we just change the one 1056 * page. Otherwise, we'll need a whole clusters worth. 1057 */ 1058 if (new) 1059 numpages = ocfs2_pages_per_cluster(inode->i_sb); 1060 1061 cpages = kzalloc(sizeof(*cpages) * numpages, GFP_NOFS); 1062 if (!cpages) { 1063 ret = -ENOMEM; 1064 mlog_errno(ret); 1065 return ret; 1066 } 1067 1068 /* 1069 * Fill our page array first. That way we've grabbed enough so 1070 * that we can zero and flush if we error after adding the 1071 * extent. 1072 */ 1073 if (new) { 1074 start = ocfs2_align_clusters_to_page_index(inode->i_sb, 1075 wc->w_cpos); 1076 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, wc->w_cpos); 1077 } else { 1078 start = wc->w_pos >> PAGE_CACHE_SHIFT; 1079 v_blkno = wc->w_pos >> inode->i_sb->s_blocksize_bits; 1080 } 1081 1082 for(i = 0; i < numpages; i++) { 1083 index = start + i; 1084 1085 cpages[i] = grab_cache_page(mapping, index); 1086 if (!cpages[i]) { 1087 ret = -ENOMEM; 1088 mlog_errno(ret); 1089 goto out; 1090 } 1091 } 1092 1093 if (new) { 1094 /* 1095 * This is safe to call with the page locks - it won't take 1096 * any additional semaphores or cluster locks. 1097 */ 1098 tmp_pos = wc->w_cpos; 1099 ret = ocfs2_do_extend_allocation(OCFS2_SB(inode->i_sb), inode, 1100 &tmp_pos, 1, di_bh, handle, 1101 data_ac, meta_ac, NULL); 1102 /* 1103 * This shouldn't happen because we must have already 1104 * calculated the correct meta data allocation required. The 1105 * internal tree allocation code should know how to increase 1106 * transaction credits itself. 1107 * 1108 * If need be, we could handle -EAGAIN for a 1109 * RESTART_TRANS here. 1110 */ 1111 mlog_bug_on_msg(ret == -EAGAIN, 1112 "Inode %llu: EAGAIN return during allocation.\n", 1113 (unsigned long long)OCFS2_I(inode)->ip_blkno); 1114 if (ret < 0) { 1115 mlog_errno(ret); 1116 goto out; 1117 } 1118 } 1119 1120 ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL, 1121 NULL); 1122 if (ret < 0) { 1123 1124 /* 1125 * XXX: Should we go readonly here? 1126 */ 1127 1128 mlog_errno(ret); 1129 goto out; 1130 } 1131 1132 BUG_ON(p_blkno == 0); 1133 1134 for(i = 0; i < numpages; i++) { 1135 ret = ocfs2_write_data_page(inode, handle, &p_blkno, cpages[i], 1136 wc, new); 1137 if (ret < 0) { 1138 mlog_errno(ret); 1139 goto out; 1140 } 1141 1142 copied += ret; 1143 } 1144 1145out: 1146 for(i = 0; i < numpages; i++) { 1147 unlock_page(cpages[i]); 1148 mark_page_accessed(cpages[i]); 1149 page_cache_release(cpages[i]); 1150 } 1151 kfree(cpages); 1152 1153 return copied ? copied : ret; 1154} 1155 1156static void ocfs2_write_ctxt_init(struct ocfs2_write_ctxt *wc, 1157 struct ocfs2_super *osb, loff_t pos, 1158 size_t count, ocfs2_page_writer *cb, 1159 void *cb_priv) 1160{ 1161 wc->w_count = count; 1162 wc->w_pos = pos; 1163 wc->w_cpos = wc->w_pos >> osb->s_clustersize_bits; 1164 wc->w_finished_copy = 0; 1165 1166 if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) 1167 wc->w_large_pages = 1; 1168 else 1169 wc->w_large_pages = 0; 1170 1171 wc->w_write_data_page = cb; 1172 wc->w_private = cb_priv; 1173} 1174 1175/* 1176 * Write a cluster to an inode. The cluster may not be allocated yet, 1177 * in which case it will be. This only exists for buffered writes - 1178 * O_DIRECT takes a more "traditional" path through the kernel. 1179 * 1180 * The caller is responsible for incrementing pos, written counts, etc 1181 * 1182 * For file systems that don't support sparse files, pre-allocation 1183 * and page zeroing up until cpos should be done prior to this 1184 * function call. 1185 * 1186 * Callers should be holding i_sem, and the rw cluster lock. 1187 * 1188 * Returns the number of user bytes written, or less than zero for 1189 * error. 1190 */ 1191ssize_t ocfs2_buffered_write_cluster(struct file *file, loff_t pos, 1192 size_t count, ocfs2_page_writer *actor, 1193 void *priv) 1194{ 1195 int ret, credits = OCFS2_INODE_UPDATE_CREDITS; 1196 ssize_t written = 0; 1197 u32 phys; 1198 struct inode *inode = file->f_mapping->host; 1199 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 1200 struct buffer_head *di_bh = NULL; 1201 struct ocfs2_dinode *di; 1202 struct ocfs2_alloc_context *data_ac = NULL; 1203 struct ocfs2_alloc_context *meta_ac = NULL; 1204 handle_t *handle; 1205 struct ocfs2_write_ctxt wc; 1206 1207 ocfs2_write_ctxt_init(&wc, osb, pos, count, actor, priv); 1208 1209 ret = ocfs2_meta_lock(inode, &di_bh, 1); 1210 if (ret) { 1211 mlog_errno(ret); 1212 goto out; 1213 } 1214 di = (struct ocfs2_dinode *)di_bh->b_data; 1215 1216 /* 1217 * Take alloc sem here to prevent concurrent lookups. That way 1218 * the mapping, zeroing and tree manipulation within 1219 * ocfs2_write() will be safe against ->readpage(). This 1220 * should also serve to lock out allocation from a shared 1221 * writeable region. 1222 */ 1223 down_write(&OCFS2_I(inode)->ip_alloc_sem); 1224 1225 ret = ocfs2_get_clusters(inode, wc.w_cpos, &phys, NULL, NULL); 1226 if (ret) { 1227 mlog_errno(ret); 1228 goto out_meta; 1229 } 1230 1231 /* phys == 0 means that allocation is required. */ 1232 if (phys == 0) { 1233 ret = ocfs2_lock_allocators(inode, di, 1, &data_ac, &meta_ac); 1234 if (ret) { 1235 mlog_errno(ret); 1236 goto out_meta; 1237 } 1238 1239 credits = ocfs2_calc_extend_credits(inode->i_sb, di, 1); 1240 } 1241 1242 ret = ocfs2_data_lock(inode, 1); 1243 if (ret) { 1244 mlog_errno(ret); 1245 goto out_meta; 1246 } 1247 1248 handle = ocfs2_start_trans(osb, credits); 1249 if (IS_ERR(handle)) { 1250 ret = PTR_ERR(handle); 1251 mlog_errno(ret); 1252 goto out_data; 1253 } 1254 1255 written = ocfs2_write(file, phys, handle, di_bh, data_ac, 1256 meta_ac, &wc); 1257 if (written < 0) { 1258 ret = written; 1259 mlog_errno(ret); 1260 goto out_commit; 1261 } 1262 1263 ret = ocfs2_journal_access(handle, inode, di_bh, 1264 OCFS2_JOURNAL_ACCESS_WRITE); 1265 if (ret) { 1266 mlog_errno(ret); 1267 goto out_commit; 1268 } 1269 1270 pos += written; 1271 if (pos > inode->i_size) { 1272 i_size_write(inode, pos); 1273 mark_inode_dirty(inode); 1274 } 1275 inode->i_blocks = ocfs2_align_bytes_to_sectors((u64)(i_size_read(inode))); 1276 di->i_size = cpu_to_le64((u64)i_size_read(inode)); 1277 inode->i_mtime = inode->i_ctime = CURRENT_TIME; 1278 di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec); 1279 di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec); 1280 1281 ret = ocfs2_journal_dirty(handle, di_bh); 1282 if (ret) 1283 mlog_errno(ret); 1284 1285out_commit: 1286 ocfs2_commit_trans(osb, handle); 1287 1288out_data: 1289 ocfs2_data_unlock(inode, 1); 1290 1291out_meta: 1292 up_write(&OCFS2_I(inode)->ip_alloc_sem); 1293 ocfs2_meta_unlock(inode, 1); 1294 1295out: 1296 brelse(di_bh); 1297 if (data_ac) 1298 ocfs2_free_alloc_context(data_ac); 1299 if (meta_ac) 1300 ocfs2_free_alloc_context(meta_ac); 1301 1302 return written ? written : ret; 1303} 1304 1305const struct address_space_operations ocfs2_aops = { 1306 .readpage = ocfs2_readpage, 1307 .writepage = ocfs2_writepage, 1308 .bmap = ocfs2_bmap, 1309 .sync_page = block_sync_page, 1310 .direct_IO = ocfs2_direct_IO, 1311 .invalidatepage = ocfs2_invalidatepage, 1312 .releasepage = ocfs2_releasepage, 1313 .migratepage = buffer_migrate_page, 1314}; 1315