aops.c revision d28c91740ae2cd1d963f9e4e3889789894cb6d52
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 28#define MLOG_MASK_PREFIX ML_FILE_IO 29#include <cluster/masklog.h> 30 31#include "ocfs2.h" 32 33#include "alloc.h" 34#include "aops.h" 35#include "dlmglue.h" 36#include "extent_map.h" 37#include "file.h" 38#include "inode.h" 39#include "journal.h" 40#include "super.h" 41#include "symlink.h" 42 43#include "buffer_head_io.h" 44 45static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock, 46 struct buffer_head *bh_result, int create) 47{ 48 int err = -EIO; 49 int status; 50 struct ocfs2_dinode *fe = NULL; 51 struct buffer_head *bh = NULL; 52 struct buffer_head *buffer_cache_bh = NULL; 53 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 54 void *kaddr; 55 56 mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode, 57 (unsigned long long)iblock, bh_result, create); 58 59 BUG_ON(ocfs2_inode_is_fast_symlink(inode)); 60 61 if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) { 62 mlog(ML_ERROR, "block offset > PATH_MAX: %llu", 63 (unsigned long long)iblock); 64 goto bail; 65 } 66 67 status = ocfs2_read_block(OCFS2_SB(inode->i_sb), 68 OCFS2_I(inode)->ip_blkno, 69 &bh, OCFS2_BH_CACHED, inode); 70 if (status < 0) { 71 mlog_errno(status); 72 goto bail; 73 } 74 fe = (struct ocfs2_dinode *) bh->b_data; 75 76 if (!OCFS2_IS_VALID_DINODE(fe)) { 77 mlog(ML_ERROR, "Invalid dinode #%llu: signature = %.*s\n", 78 (unsigned long long)fe->i_blkno, 7, fe->i_signature); 79 goto bail; 80 } 81 82 if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb, 83 le32_to_cpu(fe->i_clusters))) { 84 mlog(ML_ERROR, "block offset is outside the allocated size: " 85 "%llu\n", (unsigned long long)iblock); 86 goto bail; 87 } 88 89 /* We don't use the page cache to create symlink data, so if 90 * need be, copy it over from the buffer cache. */ 91 if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) { 92 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + 93 iblock; 94 buffer_cache_bh = sb_getblk(osb->sb, blkno); 95 if (!buffer_cache_bh) { 96 mlog(ML_ERROR, "couldn't getblock for symlink!\n"); 97 goto bail; 98 } 99 100 /* we haven't locked out transactions, so a commit 101 * could've happened. Since we've got a reference on 102 * the bh, even if it commits while we're doing the 103 * copy, the data is still good. */ 104 if (buffer_jbd(buffer_cache_bh) 105 && ocfs2_inode_is_new(inode)) { 106 kaddr = kmap_atomic(bh_result->b_page, KM_USER0); 107 if (!kaddr) { 108 mlog(ML_ERROR, "couldn't kmap!\n"); 109 goto bail; 110 } 111 memcpy(kaddr + (bh_result->b_size * iblock), 112 buffer_cache_bh->b_data, 113 bh_result->b_size); 114 kunmap_atomic(kaddr, KM_USER0); 115 set_buffer_uptodate(bh_result); 116 } 117 brelse(buffer_cache_bh); 118 } 119 120 map_bh(bh_result, inode->i_sb, 121 le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock); 122 123 err = 0; 124 125bail: 126 if (bh) 127 brelse(bh); 128 129 mlog_exit(err); 130 return err; 131} 132 133static int ocfs2_get_block(struct inode *inode, sector_t iblock, 134 struct buffer_head *bh_result, int create) 135{ 136 int err = 0; 137 u64 p_blkno, past_eof; 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 /* this can happen if another node truncs after our extend! */ 153 spin_lock(&OCFS2_I(inode)->ip_lock); 154 if (iblock >= ocfs2_clusters_to_blocks(inode->i_sb, 155 OCFS2_I(inode)->ip_clusters)) 156 err = -EIO; 157 spin_unlock(&OCFS2_I(inode)->ip_lock); 158 if (err) 159 goto bail; 160 161 err = ocfs2_extent_map_get_blocks(inode, iblock, 1, &p_blkno, 162 NULL); 163 if (err) { 164 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, " 165 "%llu, NULL)\n", err, inode, (unsigned long long)iblock, 166 (unsigned long long)p_blkno); 167 goto bail; 168 } 169 170 map_bh(bh_result, inode->i_sb, p_blkno); 171 172 if (bh_result->b_blocknr == 0) { 173 err = -EIO; 174 mlog(ML_ERROR, "iblock = %llu p_blkno = %llu blkno=(%llu)\n", 175 (unsigned long long)iblock, 176 (unsigned long long)p_blkno, 177 (unsigned long long)OCFS2_I(inode)->ip_blkno); 178 } 179 180 past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode)); 181 mlog(0, "Inode %lu, past_eof = %llu\n", inode->i_ino, 182 (unsigned long long)past_eof); 183 184 if (create && (iblock >= past_eof)) 185 set_buffer_new(bh_result); 186 187bail: 188 if (err < 0) 189 err = -EIO; 190 191 mlog_exit(err); 192 return err; 193} 194 195static int ocfs2_readpage(struct file *file, struct page *page) 196{ 197 struct inode *inode = page->mapping->host; 198 loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT; 199 int ret, unlock = 1; 200 201 mlog_entry("(0x%p, %lu)\n", file, (page ? page->index : 0)); 202 203 ret = ocfs2_meta_lock_with_page(inode, NULL, 0, page); 204 if (ret != 0) { 205 if (ret == AOP_TRUNCATED_PAGE) 206 unlock = 0; 207 mlog_errno(ret); 208 goto out; 209 } 210 211 down_read(&OCFS2_I(inode)->ip_alloc_sem); 212 213 /* 214 * i_size might have just been updated as we grabed the meta lock. We 215 * might now be discovering a truncate that hit on another node. 216 * block_read_full_page->get_block freaks out if it is asked to read 217 * beyond the end of a file, so we check here. Callers 218 * (generic_file_read, fault->nopage) are clever enough to check i_size 219 * and notice that the page they just read isn't needed. 220 * 221 * XXX sys_readahead() seems to get that wrong? 222 */ 223 if (start >= i_size_read(inode)) { 224 char *addr = kmap(page); 225 memset(addr, 0, PAGE_SIZE); 226 flush_dcache_page(page); 227 kunmap(page); 228 SetPageUptodate(page); 229 ret = 0; 230 goto out_alloc; 231 } 232 233 ret = ocfs2_data_lock_with_page(inode, 0, page); 234 if (ret != 0) { 235 if (ret == AOP_TRUNCATED_PAGE) 236 unlock = 0; 237 mlog_errno(ret); 238 goto out_alloc; 239 } 240 241 ret = block_read_full_page(page, ocfs2_get_block); 242 unlock = 0; 243 244 ocfs2_data_unlock(inode, 0); 245out_alloc: 246 up_read(&OCFS2_I(inode)->ip_alloc_sem); 247 ocfs2_meta_unlock(inode, 0); 248out: 249 if (unlock) 250 unlock_page(page); 251 mlog_exit(ret); 252 return ret; 253} 254 255/* Note: Because we don't support holes, our allocation has 256 * already happened (allocation writes zeros to the file data) 257 * so we don't have to worry about ordered writes in 258 * ocfs2_writepage. 259 * 260 * ->writepage is called during the process of invalidating the page cache 261 * during blocked lock processing. It can't block on any cluster locks 262 * to during block mapping. It's relying on the fact that the block 263 * mapping can't have disappeared under the dirty pages that it is 264 * being asked to write back. 265 */ 266static int ocfs2_writepage(struct page *page, struct writeback_control *wbc) 267{ 268 int ret; 269 270 mlog_entry("(0x%p)\n", page); 271 272 ret = block_write_full_page(page, ocfs2_get_block, wbc); 273 274 mlog_exit(ret); 275 276 return ret; 277} 278 279/* This can also be called from ocfs2_write_zero_page() which has done 280 * it's own cluster locking. */ 281int ocfs2_prepare_write_nolock(struct inode *inode, struct page *page, 282 unsigned from, unsigned to) 283{ 284 int ret; 285 286 down_read(&OCFS2_I(inode)->ip_alloc_sem); 287 288 ret = block_prepare_write(page, from, to, ocfs2_get_block); 289 290 up_read(&OCFS2_I(inode)->ip_alloc_sem); 291 292 return ret; 293} 294 295/* 296 * ocfs2_prepare_write() can be an outer-most ocfs2 call when it is called 297 * from loopback. It must be able to perform its own locking around 298 * ocfs2_get_block(). 299 */ 300static int ocfs2_prepare_write(struct file *file, struct page *page, 301 unsigned from, unsigned to) 302{ 303 struct inode *inode = page->mapping->host; 304 int ret; 305 306 mlog_entry("(0x%p, 0x%p, %u, %u)\n", file, page, from, to); 307 308 ret = ocfs2_meta_lock_with_page(inode, NULL, 0, page); 309 if (ret != 0) { 310 mlog_errno(ret); 311 goto out; 312 } 313 314 ret = ocfs2_prepare_write_nolock(inode, page, from, to); 315 316 ocfs2_meta_unlock(inode, 0); 317out: 318 mlog_exit(ret); 319 return ret; 320} 321 322/* Taken from ext3. We don't necessarily need the full blown 323 * functionality yet, but IMHO it's better to cut and paste the whole 324 * thing so we can avoid introducing our own bugs (and easily pick up 325 * their fixes when they happen) --Mark */ 326static int walk_page_buffers( handle_t *handle, 327 struct buffer_head *head, 328 unsigned from, 329 unsigned to, 330 int *partial, 331 int (*fn)( handle_t *handle, 332 struct buffer_head *bh)) 333{ 334 struct buffer_head *bh; 335 unsigned block_start, block_end; 336 unsigned blocksize = head->b_size; 337 int err, ret = 0; 338 struct buffer_head *next; 339 340 for ( bh = head, block_start = 0; 341 ret == 0 && (bh != head || !block_start); 342 block_start = block_end, bh = next) 343 { 344 next = bh->b_this_page; 345 block_end = block_start + blocksize; 346 if (block_end <= from || block_start >= to) { 347 if (partial && !buffer_uptodate(bh)) 348 *partial = 1; 349 continue; 350 } 351 err = (*fn)(handle, bh); 352 if (!ret) 353 ret = err; 354 } 355 return ret; 356} 357 358handle_t *ocfs2_start_walk_page_trans(struct inode *inode, 359 struct page *page, 360 unsigned from, 361 unsigned to) 362{ 363 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 364 handle_t *handle = NULL; 365 int ret = 0; 366 367 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS); 368 if (!handle) { 369 ret = -ENOMEM; 370 mlog_errno(ret); 371 goto out; 372 } 373 374 if (ocfs2_should_order_data(inode)) { 375 ret = walk_page_buffers(handle, 376 page_buffers(page), 377 from, to, NULL, 378 ocfs2_journal_dirty_data); 379 if (ret < 0) 380 mlog_errno(ret); 381 } 382out: 383 if (ret) { 384 if (handle) 385 ocfs2_commit_trans(osb, handle); 386 handle = ERR_PTR(ret); 387 } 388 return handle; 389} 390 391static int ocfs2_commit_write(struct file *file, struct page *page, 392 unsigned from, unsigned to) 393{ 394 int ret; 395 struct buffer_head *di_bh = NULL; 396 struct inode *inode = page->mapping->host; 397 handle_t *handle = NULL; 398 struct ocfs2_dinode *di; 399 400 mlog_entry("(0x%p, 0x%p, %u, %u)\n", file, page, from, to); 401 402 /* NOTE: ocfs2_file_aio_write has ensured that it's safe for 403 * us to continue here without rechecking the I/O against 404 * changed inode values. 405 * 406 * 1) We're currently holding the inode alloc lock, so no 407 * nodes can change it underneath us. 408 * 409 * 2) We've had to take the metadata lock at least once 410 * already to check for extending writes, suid removal, etc. 411 * The meta data update code then ensures that we don't get a 412 * stale inode allocation image (i_size, i_clusters, etc). 413 */ 414 415 ret = ocfs2_meta_lock_with_page(inode, &di_bh, 1, page); 416 if (ret != 0) { 417 mlog_errno(ret); 418 goto out; 419 } 420 421 ret = ocfs2_data_lock_with_page(inode, 1, page); 422 if (ret != 0) { 423 mlog_errno(ret); 424 goto out_unlock_meta; 425 } 426 427 handle = ocfs2_start_walk_page_trans(inode, page, from, to); 428 if (IS_ERR(handle)) { 429 ret = PTR_ERR(handle); 430 goto out_unlock_data; 431 } 432 433 /* Mark our buffer early. We'd rather catch this error up here 434 * as opposed to after a successful commit_write which would 435 * require us to set back inode->i_size. */ 436 ret = ocfs2_journal_access(handle, inode, di_bh, 437 OCFS2_JOURNAL_ACCESS_WRITE); 438 if (ret < 0) { 439 mlog_errno(ret); 440 goto out_commit; 441 } 442 443 /* might update i_size */ 444 ret = generic_commit_write(file, page, from, to); 445 if (ret < 0) { 446 mlog_errno(ret); 447 goto out_commit; 448 } 449 450 di = (struct ocfs2_dinode *)di_bh->b_data; 451 452 /* ocfs2_mark_inode_dirty() is too heavy to use here. */ 453 inode->i_mtime = inode->i_ctime = CURRENT_TIME; 454 di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec); 455 di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec); 456 457 inode->i_blocks = ocfs2_align_bytes_to_sectors((u64)(i_size_read(inode))); 458 di->i_size = cpu_to_le64((u64)i_size_read(inode)); 459 460 ret = ocfs2_journal_dirty(handle, di_bh); 461 if (ret < 0) { 462 mlog_errno(ret); 463 goto out_commit; 464 } 465 466out_commit: 467 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle); 468out_unlock_data: 469 ocfs2_data_unlock(inode, 1); 470out_unlock_meta: 471 ocfs2_meta_unlock(inode, 1); 472out: 473 if (di_bh) 474 brelse(di_bh); 475 476 mlog_exit(ret); 477 return ret; 478} 479 480static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block) 481{ 482 sector_t status; 483 u64 p_blkno = 0; 484 int err = 0; 485 struct inode *inode = mapping->host; 486 487 mlog_entry("(block = %llu)\n", (unsigned long long)block); 488 489 /* We don't need to lock journal system files, since they aren't 490 * accessed concurrently from multiple nodes. 491 */ 492 if (!INODE_JOURNAL(inode)) { 493 err = ocfs2_meta_lock(inode, NULL, 0); 494 if (err) { 495 if (err != -ENOENT) 496 mlog_errno(err); 497 goto bail; 498 } 499 down_read(&OCFS2_I(inode)->ip_alloc_sem); 500 } 501 502 err = ocfs2_extent_map_get_blocks(inode, block, 1, &p_blkno, 503 NULL); 504 505 if (!INODE_JOURNAL(inode)) { 506 up_read(&OCFS2_I(inode)->ip_alloc_sem); 507 ocfs2_meta_unlock(inode, 0); 508 } 509 510 if (err) { 511 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n", 512 (unsigned long long)block); 513 mlog_errno(err); 514 goto bail; 515 } 516 517 518bail: 519 status = err ? 0 : p_blkno; 520 521 mlog_exit((int)status); 522 523 return status; 524} 525 526/* 527 * TODO: Make this into a generic get_blocks function. 528 * 529 * From do_direct_io in direct-io.c: 530 * "So what we do is to permit the ->get_blocks function to populate 531 * bh.b_size with the size of IO which is permitted at this offset and 532 * this i_blkbits." 533 * 534 * This function is called directly from get_more_blocks in direct-io.c. 535 * 536 * called like this: dio->get_blocks(dio->inode, fs_startblk, 537 * fs_count, map_bh, dio->rw == WRITE); 538 */ 539static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock, 540 struct buffer_head *bh_result, int create) 541{ 542 int ret; 543 u64 vbo_max; /* file offset, max_blocks from iblock */ 544 u64 p_blkno; 545 int contig_blocks; 546 unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits; 547 unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits; 548 549 /* This function won't even be called if the request isn't all 550 * nicely aligned and of the right size, so there's no need 551 * for us to check any of that. */ 552 553 vbo_max = ((u64)iblock + max_blocks) << blocksize_bits; 554 555 spin_lock(&OCFS2_I(inode)->ip_lock); 556 if ((iblock + max_blocks) > 557 ocfs2_clusters_to_blocks(inode->i_sb, 558 OCFS2_I(inode)->ip_clusters)) { 559 spin_unlock(&OCFS2_I(inode)->ip_lock); 560 ret = -EIO; 561 goto bail; 562 } 563 spin_unlock(&OCFS2_I(inode)->ip_lock); 564 565 /* This figures out the size of the next contiguous block, and 566 * our logical offset */ 567 ret = ocfs2_extent_map_get_blocks(inode, iblock, 1, &p_blkno, 568 &contig_blocks); 569 if (ret) { 570 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n", 571 (unsigned long long)iblock); 572 ret = -EIO; 573 goto bail; 574 } 575 576 map_bh(bh_result, inode->i_sb, p_blkno); 577 578 /* make sure we don't map more than max_blocks blocks here as 579 that's all the kernel will handle at this point. */ 580 if (max_blocks < contig_blocks) 581 contig_blocks = max_blocks; 582 bh_result->b_size = contig_blocks << blocksize_bits; 583bail: 584 return ret; 585} 586 587/* 588 * ocfs2_dio_end_io is called by the dio core when a dio is finished. We're 589 * particularly interested in the aio/dio case. Like the core uses 590 * i_alloc_sem, we use the rw_lock DLM lock to protect io on one node from 591 * truncation on another. 592 */ 593static void ocfs2_dio_end_io(struct kiocb *iocb, 594 loff_t offset, 595 ssize_t bytes, 596 void *private) 597{ 598 struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode; 599 600 /* this io's submitter should not have unlocked this before we could */ 601 BUG_ON(!ocfs2_iocb_is_rw_locked(iocb)); 602 ocfs2_iocb_clear_rw_locked(iocb); 603 up_read(&inode->i_alloc_sem); 604 ocfs2_rw_unlock(inode, 0); 605} 606 607static ssize_t ocfs2_direct_IO(int rw, 608 struct kiocb *iocb, 609 const struct iovec *iov, 610 loff_t offset, 611 unsigned long nr_segs) 612{ 613 struct file *file = iocb->ki_filp; 614 struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host; 615 int ret; 616 617 mlog_entry_void(); 618 619 /* 620 * We get PR data locks even for O_DIRECT. This allows 621 * concurrent O_DIRECT I/O but doesn't let O_DIRECT with 622 * extending and buffered zeroing writes race. If they did 623 * race then the buffered zeroing could be written back after 624 * the O_DIRECT I/O. It's one thing to tell people not to mix 625 * buffered and O_DIRECT writes, but expecting them to 626 * understand that file extension is also an implicit buffered 627 * write is too much. By getting the PR we force writeback of 628 * the buffered zeroing before proceeding. 629 */ 630 ret = ocfs2_data_lock(inode, 0); 631 if (ret < 0) { 632 mlog_errno(ret); 633 goto out; 634 } 635 ocfs2_data_unlock(inode, 0); 636 637 ret = blockdev_direct_IO_no_locking(rw, iocb, inode, 638 inode->i_sb->s_bdev, iov, offset, 639 nr_segs, 640 ocfs2_direct_IO_get_blocks, 641 ocfs2_dio_end_io); 642out: 643 mlog_exit(ret); 644 return ret; 645} 646 647const struct address_space_operations ocfs2_aops = { 648 .readpage = ocfs2_readpage, 649 .writepage = ocfs2_writepage, 650 .prepare_write = ocfs2_prepare_write, 651 .commit_write = ocfs2_commit_write, 652 .bmap = ocfs2_bmap, 653 .sync_page = block_sync_page, 654 .direct_IO = ocfs2_direct_IO 655}; 656