| 1 | /* |
| 2 | * Copyright (c) 2000-2005 Silicon Graphics, Inc. |
| 3 | * All Rights Reserved. |
| 4 | * |
| 5 | * This program is free software; you can redistribute it and/or |
| 6 | * modify it under the terms of the GNU General Public License as |
| 7 | * published by the Free Software Foundation. |
| 8 | * |
| 9 | * This program is distributed in the hope that it would be useful, |
| 10 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 11 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 12 | * GNU General Public License for more details. |
| 13 | * |
| 14 | * You should have received a copy of the GNU General Public License |
| 15 | * along with this program; if not, write the Free Software Foundation, |
| 16 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA |
| 17 | */ |
| 18 | #include "xfs.h" |
| 19 | #include "xfs_shared.h" |
| 20 | #include "xfs_format.h" |
| 21 | #include "xfs_log_format.h" |
| 22 | #include "xfs_trans_resv.h" |
| 23 | #include "xfs_mount.h" |
| 24 | #include "xfs_inode.h" |
| 25 | #include "xfs_trans.h" |
| 26 | #include "xfs_inode_item.h" |
| 27 | #include "xfs_alloc.h" |
| 28 | #include "xfs_error.h" |
| 29 | #include "xfs_iomap.h" |
| 30 | #include "xfs_trace.h" |
| 31 | #include "xfs_bmap.h" |
| 32 | #include "xfs_bmap_util.h" |
| 33 | #include "xfs_bmap_btree.h" |
| 34 | #include <linux/gfp.h> |
| 35 | #include <linux/mpage.h> |
| 36 | #include <linux/pagevec.h> |
| 37 | #include <linux/writeback.h> |
| 38 | |
| 39 | void |
| 40 | xfs_count_page_state( |
| 41 | struct page *page, |
| 42 | int *delalloc, |
| 43 | int *unwritten) |
| 44 | { |
| 45 | struct buffer_head *bh, *head; |
| 46 | |
| 47 | *delalloc = *unwritten = 0; |
| 48 | |
| 49 | bh = head = page_buffers(page); |
| 50 | do { |
| 51 | if (buffer_unwritten(bh)) |
| 52 | (*unwritten) = 1; |
| 53 | else if (buffer_delay(bh)) |
| 54 | (*delalloc) = 1; |
| 55 | } while ((bh = bh->b_this_page) != head); |
| 56 | } |
| 57 | |
| 58 | STATIC struct block_device * |
| 59 | xfs_find_bdev_for_inode( |
| 60 | struct inode *inode) |
| 61 | { |
| 62 | struct xfs_inode *ip = XFS_I(inode); |
| 63 | struct xfs_mount *mp = ip->i_mount; |
| 64 | |
| 65 | if (XFS_IS_REALTIME_INODE(ip)) |
| 66 | return mp->m_rtdev_targp->bt_bdev; |
| 67 | else |
| 68 | return mp->m_ddev_targp->bt_bdev; |
| 69 | } |
| 70 | |
| 71 | /* |
| 72 | * We're now finished for good with this ioend structure. |
| 73 | * Update the page state via the associated buffer_heads, |
| 74 | * release holds on the inode and bio, and finally free |
| 75 | * up memory. Do not use the ioend after this. |
| 76 | */ |
| 77 | STATIC void |
| 78 | xfs_destroy_ioend( |
| 79 | xfs_ioend_t *ioend) |
| 80 | { |
| 81 | struct buffer_head *bh, *next; |
| 82 | |
| 83 | for (bh = ioend->io_buffer_head; bh; bh = next) { |
| 84 | next = bh->b_private; |
| 85 | bh->b_end_io(bh, !ioend->io_error); |
| 86 | } |
| 87 | |
| 88 | mempool_free(ioend, xfs_ioend_pool); |
| 89 | } |
| 90 | |
| 91 | /* |
| 92 | * Fast and loose check if this write could update the on-disk inode size. |
| 93 | */ |
| 94 | static inline bool xfs_ioend_is_append(struct xfs_ioend *ioend) |
| 95 | { |
| 96 | return ioend->io_offset + ioend->io_size > |
| 97 | XFS_I(ioend->io_inode)->i_d.di_size; |
| 98 | } |
| 99 | |
| 100 | STATIC int |
| 101 | xfs_setfilesize_trans_alloc( |
| 102 | struct xfs_ioend *ioend) |
| 103 | { |
| 104 | struct xfs_mount *mp = XFS_I(ioend->io_inode)->i_mount; |
| 105 | struct xfs_trans *tp; |
| 106 | int error; |
| 107 | |
| 108 | tp = xfs_trans_alloc(mp, XFS_TRANS_FSYNC_TS); |
| 109 | |
| 110 | error = xfs_trans_reserve(tp, &M_RES(mp)->tr_fsyncts, 0, 0); |
| 111 | if (error) { |
| 112 | xfs_trans_cancel(tp); |
| 113 | return error; |
| 114 | } |
| 115 | |
| 116 | ioend->io_append_trans = tp; |
| 117 | |
| 118 | /* |
| 119 | * We may pass freeze protection with a transaction. So tell lockdep |
| 120 | * we released it. |
| 121 | */ |
| 122 | __sb_writers_release(ioend->io_inode->i_sb, SB_FREEZE_FS); |
| 123 | /* |
| 124 | * We hand off the transaction to the completion thread now, so |
| 125 | * clear the flag here. |
| 126 | */ |
| 127 | current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS); |
| 128 | return 0; |
| 129 | } |
| 130 | |
| 131 | /* |
| 132 | * Update on-disk file size now that data has been written to disk. |
| 133 | */ |
| 134 | STATIC int |
| 135 | xfs_setfilesize( |
| 136 | struct xfs_inode *ip, |
| 137 | struct xfs_trans *tp, |
| 138 | xfs_off_t offset, |
| 139 | size_t size) |
| 140 | { |
| 141 | xfs_fsize_t isize; |
| 142 | |
| 143 | xfs_ilock(ip, XFS_ILOCK_EXCL); |
| 144 | isize = xfs_new_eof(ip, offset + size); |
| 145 | if (!isize) { |
| 146 | xfs_iunlock(ip, XFS_ILOCK_EXCL); |
| 147 | xfs_trans_cancel(tp); |
| 148 | return 0; |
| 149 | } |
| 150 | |
| 151 | trace_xfs_setfilesize(ip, offset, size); |
| 152 | |
| 153 | ip->i_d.di_size = isize; |
| 154 | xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL); |
| 155 | xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE); |
| 156 | |
| 157 | return xfs_trans_commit(tp); |
| 158 | } |
| 159 | |
| 160 | STATIC int |
| 161 | xfs_setfilesize_ioend( |
| 162 | struct xfs_ioend *ioend) |
| 163 | { |
| 164 | struct xfs_inode *ip = XFS_I(ioend->io_inode); |
| 165 | struct xfs_trans *tp = ioend->io_append_trans; |
| 166 | |
| 167 | /* |
| 168 | * The transaction may have been allocated in the I/O submission thread, |
| 169 | * thus we need to mark ourselves as being in a transaction manually. |
| 170 | * Similarly for freeze protection. |
| 171 | */ |
| 172 | current_set_flags_nested(&tp->t_pflags, PF_FSTRANS); |
| 173 | __sb_writers_acquired(VFS_I(ip)->i_sb, SB_FREEZE_FS); |
| 174 | |
| 175 | /* we abort the update if there was an IO error */ |
| 176 | if (ioend->io_error) { |
| 177 | xfs_trans_cancel(tp); |
| 178 | return ioend->io_error; |
| 179 | } |
| 180 | |
| 181 | return xfs_setfilesize(ip, tp, ioend->io_offset, ioend->io_size); |
| 182 | } |
| 183 | |
| 184 | /* |
| 185 | * Schedule IO completion handling on the final put of an ioend. |
| 186 | * |
| 187 | * If there is no work to do we might as well call it a day and free the |
| 188 | * ioend right now. |
| 189 | */ |
| 190 | STATIC void |
| 191 | xfs_finish_ioend( |
| 192 | struct xfs_ioend *ioend) |
| 193 | { |
| 194 | if (atomic_dec_and_test(&ioend->io_remaining)) { |
| 195 | struct xfs_mount *mp = XFS_I(ioend->io_inode)->i_mount; |
| 196 | |
| 197 | if (ioend->io_type == XFS_IO_UNWRITTEN) |
| 198 | queue_work(mp->m_unwritten_workqueue, &ioend->io_work); |
| 199 | else if (ioend->io_append_trans) |
| 200 | queue_work(mp->m_data_workqueue, &ioend->io_work); |
| 201 | else |
| 202 | xfs_destroy_ioend(ioend); |
| 203 | } |
| 204 | } |
| 205 | |
| 206 | /* |
| 207 | * IO write completion. |
| 208 | */ |
| 209 | STATIC void |
| 210 | xfs_end_io( |
| 211 | struct work_struct *work) |
| 212 | { |
| 213 | xfs_ioend_t *ioend = container_of(work, xfs_ioend_t, io_work); |
| 214 | struct xfs_inode *ip = XFS_I(ioend->io_inode); |
| 215 | int error = 0; |
| 216 | |
| 217 | if (XFS_FORCED_SHUTDOWN(ip->i_mount)) { |
| 218 | ioend->io_error = -EIO; |
| 219 | goto done; |
| 220 | } |
| 221 | |
| 222 | /* |
| 223 | * For unwritten extents we need to issue transactions to convert a |
| 224 | * range to normal written extens after the data I/O has finished. |
| 225 | * Detecting and handling completion IO errors is done individually |
| 226 | * for each case as different cleanup operations need to be performed |
| 227 | * on error. |
| 228 | */ |
| 229 | if (ioend->io_type == XFS_IO_UNWRITTEN) { |
| 230 | if (ioend->io_error) |
| 231 | goto done; |
| 232 | error = xfs_iomap_write_unwritten(ip, ioend->io_offset, |
| 233 | ioend->io_size); |
| 234 | } else if (ioend->io_append_trans) { |
| 235 | error = xfs_setfilesize_ioend(ioend); |
| 236 | } else { |
| 237 | ASSERT(!xfs_ioend_is_append(ioend)); |
| 238 | } |
| 239 | |
| 240 | done: |
| 241 | if (error) |
| 242 | ioend->io_error = error; |
| 243 | xfs_destroy_ioend(ioend); |
| 244 | } |
| 245 | |
| 246 | /* |
| 247 | * Allocate and initialise an IO completion structure. |
| 248 | * We need to track unwritten extent write completion here initially. |
| 249 | * We'll need to extend this for updating the ondisk inode size later |
| 250 | * (vs. incore size). |
| 251 | */ |
| 252 | STATIC xfs_ioend_t * |
| 253 | xfs_alloc_ioend( |
| 254 | struct inode *inode, |
| 255 | unsigned int type) |
| 256 | { |
| 257 | xfs_ioend_t *ioend; |
| 258 | |
| 259 | ioend = mempool_alloc(xfs_ioend_pool, GFP_NOFS); |
| 260 | |
| 261 | /* |
| 262 | * Set the count to 1 initially, which will prevent an I/O |
| 263 | * completion callback from happening before we have started |
| 264 | * all the I/O from calling the completion routine too early. |
| 265 | */ |
| 266 | atomic_set(&ioend->io_remaining, 1); |
| 267 | ioend->io_error = 0; |
| 268 | ioend->io_list = NULL; |
| 269 | ioend->io_type = type; |
| 270 | ioend->io_inode = inode; |
| 271 | ioend->io_buffer_head = NULL; |
| 272 | ioend->io_buffer_tail = NULL; |
| 273 | ioend->io_offset = 0; |
| 274 | ioend->io_size = 0; |
| 275 | ioend->io_append_trans = NULL; |
| 276 | |
| 277 | INIT_WORK(&ioend->io_work, xfs_end_io); |
| 278 | return ioend; |
| 279 | } |
| 280 | |
| 281 | STATIC int |
| 282 | xfs_map_blocks( |
| 283 | struct inode *inode, |
| 284 | loff_t offset, |
| 285 | struct xfs_bmbt_irec *imap, |
| 286 | int type, |
| 287 | int nonblocking) |
| 288 | { |
| 289 | struct xfs_inode *ip = XFS_I(inode); |
| 290 | struct xfs_mount *mp = ip->i_mount; |
| 291 | ssize_t count = i_blocksize(inode); |
| 292 | xfs_fileoff_t offset_fsb, end_fsb; |
| 293 | int error = 0; |
| 294 | int bmapi_flags = XFS_BMAPI_ENTIRE; |
| 295 | int nimaps = 1; |
| 296 | |
| 297 | if (XFS_FORCED_SHUTDOWN(mp)) |
| 298 | return -EIO; |
| 299 | |
| 300 | if (type == XFS_IO_UNWRITTEN) |
| 301 | bmapi_flags |= XFS_BMAPI_IGSTATE; |
| 302 | |
| 303 | if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) { |
| 304 | if (nonblocking) |
| 305 | return -EAGAIN; |
| 306 | xfs_ilock(ip, XFS_ILOCK_SHARED); |
| 307 | } |
| 308 | |
| 309 | ASSERT(ip->i_d.di_format != XFS_DINODE_FMT_BTREE || |
| 310 | (ip->i_df.if_flags & XFS_IFEXTENTS)); |
| 311 | ASSERT(offset <= mp->m_super->s_maxbytes); |
| 312 | |
| 313 | if (offset + count > mp->m_super->s_maxbytes) |
| 314 | count = mp->m_super->s_maxbytes - offset; |
| 315 | end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + count); |
| 316 | offset_fsb = XFS_B_TO_FSBT(mp, offset); |
| 317 | error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb, |
| 318 | imap, &nimaps, bmapi_flags); |
| 319 | xfs_iunlock(ip, XFS_ILOCK_SHARED); |
| 320 | |
| 321 | if (error) |
| 322 | return error; |
| 323 | |
| 324 | if (type == XFS_IO_DELALLOC && |
| 325 | (!nimaps || isnullstartblock(imap->br_startblock))) { |
| 326 | error = xfs_iomap_write_allocate(ip, offset, imap); |
| 327 | if (!error) |
| 328 | trace_xfs_map_blocks_alloc(ip, offset, count, type, imap); |
| 329 | return error; |
| 330 | } |
| 331 | |
| 332 | #ifdef DEBUG |
| 333 | if (type == XFS_IO_UNWRITTEN) { |
| 334 | ASSERT(nimaps); |
| 335 | ASSERT(imap->br_startblock != HOLESTARTBLOCK); |
| 336 | ASSERT(imap->br_startblock != DELAYSTARTBLOCK); |
| 337 | } |
| 338 | #endif |
| 339 | if (nimaps) |
| 340 | trace_xfs_map_blocks_found(ip, offset, count, type, imap); |
| 341 | return 0; |
| 342 | } |
| 343 | |
| 344 | STATIC int |
| 345 | xfs_imap_valid( |
| 346 | struct inode *inode, |
| 347 | struct xfs_bmbt_irec *imap, |
| 348 | xfs_off_t offset) |
| 349 | { |
| 350 | offset >>= inode->i_blkbits; |
| 351 | |
| 352 | return offset >= imap->br_startoff && |
| 353 | offset < imap->br_startoff + imap->br_blockcount; |
| 354 | } |
| 355 | |
| 356 | /* |
| 357 | * BIO completion handler for buffered IO. |
| 358 | */ |
| 359 | STATIC void |
| 360 | xfs_end_bio( |
| 361 | struct bio *bio) |
| 362 | { |
| 363 | xfs_ioend_t *ioend = bio->bi_private; |
| 364 | |
| 365 | if (!ioend->io_error) |
| 366 | ioend->io_error = bio->bi_error; |
| 367 | |
| 368 | /* Toss bio and pass work off to an xfsdatad thread */ |
| 369 | bio->bi_private = NULL; |
| 370 | bio->bi_end_io = NULL; |
| 371 | bio_put(bio); |
| 372 | |
| 373 | xfs_finish_ioend(ioend); |
| 374 | } |
| 375 | |
| 376 | STATIC void |
| 377 | xfs_submit_ioend_bio( |
| 378 | struct writeback_control *wbc, |
| 379 | xfs_ioend_t *ioend, |
| 380 | struct bio *bio) |
| 381 | { |
| 382 | atomic_inc(&ioend->io_remaining); |
| 383 | bio->bi_private = ioend; |
| 384 | bio->bi_end_io = xfs_end_bio; |
| 385 | submit_bio(wbc->sync_mode == WB_SYNC_ALL ? WRITE_SYNC : WRITE, bio); |
| 386 | } |
| 387 | |
| 388 | STATIC struct bio * |
| 389 | xfs_alloc_ioend_bio( |
| 390 | struct buffer_head *bh) |
| 391 | { |
| 392 | struct bio *bio = bio_alloc(GFP_NOIO, BIO_MAX_PAGES); |
| 393 | |
| 394 | ASSERT(bio->bi_private == NULL); |
| 395 | bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9); |
| 396 | bio->bi_bdev = bh->b_bdev; |
| 397 | return bio; |
| 398 | } |
| 399 | |
| 400 | STATIC void |
| 401 | xfs_start_buffer_writeback( |
| 402 | struct buffer_head *bh) |
| 403 | { |
| 404 | ASSERT(buffer_mapped(bh)); |
| 405 | ASSERT(buffer_locked(bh)); |
| 406 | ASSERT(!buffer_delay(bh)); |
| 407 | ASSERT(!buffer_unwritten(bh)); |
| 408 | |
| 409 | mark_buffer_async_write(bh); |
| 410 | set_buffer_uptodate(bh); |
| 411 | clear_buffer_dirty(bh); |
| 412 | } |
| 413 | |
| 414 | STATIC void |
| 415 | xfs_start_page_writeback( |
| 416 | struct page *page, |
| 417 | int clear_dirty, |
| 418 | int buffers) |
| 419 | { |
| 420 | ASSERT(PageLocked(page)); |
| 421 | ASSERT(!PageWriteback(page)); |
| 422 | |
| 423 | /* |
| 424 | * if the page was not fully cleaned, we need to ensure that the higher |
| 425 | * layers come back to it correctly. That means we need to keep the page |
| 426 | * dirty, and for WB_SYNC_ALL writeback we need to ensure the |
| 427 | * PAGECACHE_TAG_TOWRITE index mark is not removed so another attempt to |
| 428 | * write this page in this writeback sweep will be made. |
| 429 | */ |
| 430 | if (clear_dirty) { |
| 431 | clear_page_dirty_for_io(page); |
| 432 | set_page_writeback(page); |
| 433 | } else |
| 434 | set_page_writeback_keepwrite(page); |
| 435 | |
| 436 | unlock_page(page); |
| 437 | |
| 438 | /* If no buffers on the page are to be written, finish it here */ |
| 439 | if (!buffers) |
| 440 | end_page_writeback(page); |
| 441 | } |
| 442 | |
| 443 | static inline int xfs_bio_add_buffer(struct bio *bio, struct buffer_head *bh) |
| 444 | { |
| 445 | return bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh)); |
| 446 | } |
| 447 | |
| 448 | /* |
| 449 | * Submit all of the bios for all of the ioends we have saved up, covering the |
| 450 | * initial writepage page and also any probed pages. |
| 451 | * |
| 452 | * Because we may have multiple ioends spanning a page, we need to start |
| 453 | * writeback on all the buffers before we submit them for I/O. If we mark the |
| 454 | * buffers as we got, then we can end up with a page that only has buffers |
| 455 | * marked async write and I/O complete on can occur before we mark the other |
| 456 | * buffers async write. |
| 457 | * |
| 458 | * The end result of this is that we trip a bug in end_page_writeback() because |
| 459 | * we call it twice for the one page as the code in end_buffer_async_write() |
| 460 | * assumes that all buffers on the page are started at the same time. |
| 461 | * |
| 462 | * The fix is two passes across the ioend list - one to start writeback on the |
| 463 | * buffer_heads, and then submit them for I/O on the second pass. |
| 464 | * |
| 465 | * If @fail is non-zero, it means that we have a situation where some part of |
| 466 | * the submission process has failed after we have marked paged for writeback |
| 467 | * and unlocked them. In this situation, we need to fail the ioend chain rather |
| 468 | * than submit it to IO. This typically only happens on a filesystem shutdown. |
| 469 | */ |
| 470 | STATIC void |
| 471 | xfs_submit_ioend( |
| 472 | struct writeback_control *wbc, |
| 473 | xfs_ioend_t *ioend, |
| 474 | int fail) |
| 475 | { |
| 476 | xfs_ioend_t *head = ioend; |
| 477 | xfs_ioend_t *next; |
| 478 | struct buffer_head *bh; |
| 479 | struct bio *bio; |
| 480 | sector_t lastblock = 0; |
| 481 | |
| 482 | /* Pass 1 - start writeback */ |
| 483 | do { |
| 484 | next = ioend->io_list; |
| 485 | for (bh = ioend->io_buffer_head; bh; bh = bh->b_private) |
| 486 | xfs_start_buffer_writeback(bh); |
| 487 | } while ((ioend = next) != NULL); |
| 488 | |
| 489 | /* Pass 2 - submit I/O */ |
| 490 | ioend = head; |
| 491 | do { |
| 492 | next = ioend->io_list; |
| 493 | bio = NULL; |
| 494 | |
| 495 | /* |
| 496 | * If we are failing the IO now, just mark the ioend with an |
| 497 | * error and finish it. This will run IO completion immediately |
| 498 | * as there is only one reference to the ioend at this point in |
| 499 | * time. |
| 500 | */ |
| 501 | if (fail) { |
| 502 | ioend->io_error = fail; |
| 503 | xfs_finish_ioend(ioend); |
| 504 | continue; |
| 505 | } |
| 506 | |
| 507 | for (bh = ioend->io_buffer_head; bh; bh = bh->b_private) { |
| 508 | |
| 509 | if (!bio) { |
| 510 | retry: |
| 511 | bio = xfs_alloc_ioend_bio(bh); |
| 512 | } else if (bh->b_blocknr != lastblock + 1) { |
| 513 | xfs_submit_ioend_bio(wbc, ioend, bio); |
| 514 | goto retry; |
| 515 | } |
| 516 | |
| 517 | if (xfs_bio_add_buffer(bio, bh) != bh->b_size) { |
| 518 | xfs_submit_ioend_bio(wbc, ioend, bio); |
| 519 | goto retry; |
| 520 | } |
| 521 | |
| 522 | lastblock = bh->b_blocknr; |
| 523 | } |
| 524 | if (bio) |
| 525 | xfs_submit_ioend_bio(wbc, ioend, bio); |
| 526 | xfs_finish_ioend(ioend); |
| 527 | } while ((ioend = next) != NULL); |
| 528 | } |
| 529 | |
| 530 | /* |
| 531 | * Cancel submission of all buffer_heads so far in this endio. |
| 532 | * Toss the endio too. Only ever called for the initial page |
| 533 | * in a writepage request, so only ever one page. |
| 534 | */ |
| 535 | STATIC void |
| 536 | xfs_cancel_ioend( |
| 537 | xfs_ioend_t *ioend) |
| 538 | { |
| 539 | xfs_ioend_t *next; |
| 540 | struct buffer_head *bh, *next_bh; |
| 541 | |
| 542 | do { |
| 543 | next = ioend->io_list; |
| 544 | bh = ioend->io_buffer_head; |
| 545 | do { |
| 546 | next_bh = bh->b_private; |
| 547 | clear_buffer_async_write(bh); |
| 548 | /* |
| 549 | * The unwritten flag is cleared when added to the |
| 550 | * ioend. We're not submitting for I/O so mark the |
| 551 | * buffer unwritten again for next time around. |
| 552 | */ |
| 553 | if (ioend->io_type == XFS_IO_UNWRITTEN) |
| 554 | set_buffer_unwritten(bh); |
| 555 | unlock_buffer(bh); |
| 556 | } while ((bh = next_bh) != NULL); |
| 557 | |
| 558 | mempool_free(ioend, xfs_ioend_pool); |
| 559 | } while ((ioend = next) != NULL); |
| 560 | } |
| 561 | |
| 562 | /* |
| 563 | * Test to see if we've been building up a completion structure for |
| 564 | * earlier buffers -- if so, we try to append to this ioend if we |
| 565 | * can, otherwise we finish off any current ioend and start another. |
| 566 | * Return true if we've finished the given ioend. |
| 567 | */ |
| 568 | STATIC void |
| 569 | xfs_add_to_ioend( |
| 570 | struct inode *inode, |
| 571 | struct buffer_head *bh, |
| 572 | xfs_off_t offset, |
| 573 | unsigned int type, |
| 574 | xfs_ioend_t **result, |
| 575 | int need_ioend) |
| 576 | { |
| 577 | xfs_ioend_t *ioend = *result; |
| 578 | |
| 579 | if (!ioend || need_ioend || type != ioend->io_type) { |
| 580 | xfs_ioend_t *previous = *result; |
| 581 | |
| 582 | ioend = xfs_alloc_ioend(inode, type); |
| 583 | ioend->io_offset = offset; |
| 584 | ioend->io_buffer_head = bh; |
| 585 | ioend->io_buffer_tail = bh; |
| 586 | if (previous) |
| 587 | previous->io_list = ioend; |
| 588 | *result = ioend; |
| 589 | } else { |
| 590 | ioend->io_buffer_tail->b_private = bh; |
| 591 | ioend->io_buffer_tail = bh; |
| 592 | } |
| 593 | |
| 594 | bh->b_private = NULL; |
| 595 | ioend->io_size += bh->b_size; |
| 596 | } |
| 597 | |
| 598 | STATIC void |
| 599 | xfs_map_buffer( |
| 600 | struct inode *inode, |
| 601 | struct buffer_head *bh, |
| 602 | struct xfs_bmbt_irec *imap, |
| 603 | xfs_off_t offset) |
| 604 | { |
| 605 | sector_t bn; |
| 606 | struct xfs_mount *m = XFS_I(inode)->i_mount; |
| 607 | xfs_off_t iomap_offset = XFS_FSB_TO_B(m, imap->br_startoff); |
| 608 | xfs_daddr_t iomap_bn = xfs_fsb_to_db(XFS_I(inode), imap->br_startblock); |
| 609 | |
| 610 | ASSERT(imap->br_startblock != HOLESTARTBLOCK); |
| 611 | ASSERT(imap->br_startblock != DELAYSTARTBLOCK); |
| 612 | |
| 613 | bn = (iomap_bn >> (inode->i_blkbits - BBSHIFT)) + |
| 614 | ((offset - iomap_offset) >> inode->i_blkbits); |
| 615 | |
| 616 | ASSERT(bn || XFS_IS_REALTIME_INODE(XFS_I(inode))); |
| 617 | |
| 618 | bh->b_blocknr = bn; |
| 619 | set_buffer_mapped(bh); |
| 620 | } |
| 621 | |
| 622 | STATIC void |
| 623 | xfs_map_at_offset( |
| 624 | struct inode *inode, |
| 625 | struct buffer_head *bh, |
| 626 | struct xfs_bmbt_irec *imap, |
| 627 | xfs_off_t offset) |
| 628 | { |
| 629 | ASSERT(imap->br_startblock != HOLESTARTBLOCK); |
| 630 | ASSERT(imap->br_startblock != DELAYSTARTBLOCK); |
| 631 | |
| 632 | xfs_map_buffer(inode, bh, imap, offset); |
| 633 | set_buffer_mapped(bh); |
| 634 | clear_buffer_delay(bh); |
| 635 | clear_buffer_unwritten(bh); |
| 636 | } |
| 637 | |
| 638 | /* |
| 639 | * Test if a given page contains at least one buffer of a given @type. |
| 640 | * If @check_all_buffers is true, then we walk all the buffers in the page to |
| 641 | * try to find one of the type passed in. If it is not set, then the caller only |
| 642 | * needs to check the first buffer on the page for a match. |
| 643 | */ |
| 644 | STATIC bool |
| 645 | xfs_check_page_type( |
| 646 | struct page *page, |
| 647 | unsigned int type, |
| 648 | bool check_all_buffers) |
| 649 | { |
| 650 | struct buffer_head *bh; |
| 651 | struct buffer_head *head; |
| 652 | |
| 653 | if (PageWriteback(page)) |
| 654 | return false; |
| 655 | if (!page->mapping) |
| 656 | return false; |
| 657 | if (!page_has_buffers(page)) |
| 658 | return false; |
| 659 | |
| 660 | bh = head = page_buffers(page); |
| 661 | do { |
| 662 | if (buffer_unwritten(bh)) { |
| 663 | if (type == XFS_IO_UNWRITTEN) |
| 664 | return true; |
| 665 | } else if (buffer_delay(bh)) { |
| 666 | if (type == XFS_IO_DELALLOC) |
| 667 | return true; |
| 668 | } else if (buffer_dirty(bh) && buffer_mapped(bh)) { |
| 669 | if (type == XFS_IO_OVERWRITE) |
| 670 | return true; |
| 671 | } |
| 672 | |
| 673 | /* If we are only checking the first buffer, we are done now. */ |
| 674 | if (!check_all_buffers) |
| 675 | break; |
| 676 | } while ((bh = bh->b_this_page) != head); |
| 677 | |
| 678 | return false; |
| 679 | } |
| 680 | |
| 681 | /* |
| 682 | * Allocate & map buffers for page given the extent map. Write it out. |
| 683 | * except for the original page of a writepage, this is called on |
| 684 | * delalloc/unwritten pages only, for the original page it is possible |
| 685 | * that the page has no mapping at all. |
| 686 | */ |
| 687 | STATIC int |
| 688 | xfs_convert_page( |
| 689 | struct inode *inode, |
| 690 | struct page *page, |
| 691 | loff_t tindex, |
| 692 | struct xfs_bmbt_irec *imap, |
| 693 | xfs_ioend_t **ioendp, |
| 694 | struct writeback_control *wbc) |
| 695 | { |
| 696 | struct buffer_head *bh, *head; |
| 697 | xfs_off_t end_offset; |
| 698 | unsigned long p_offset; |
| 699 | unsigned int type; |
| 700 | int len, page_dirty; |
| 701 | int count = 0, done = 0, uptodate = 1; |
| 702 | xfs_off_t offset = page_offset(page); |
| 703 | |
| 704 | if (page->index != tindex) |
| 705 | goto fail; |
| 706 | if (!trylock_page(page)) |
| 707 | goto fail; |
| 708 | if (PageWriteback(page)) |
| 709 | goto fail_unlock_page; |
| 710 | if (page->mapping != inode->i_mapping) |
| 711 | goto fail_unlock_page; |
| 712 | if (!xfs_check_page_type(page, (*ioendp)->io_type, false)) |
| 713 | goto fail_unlock_page; |
| 714 | |
| 715 | /* |
| 716 | * page_dirty is initially a count of buffers on the page before |
| 717 | * EOF and is decremented as we move each into a cleanable state. |
| 718 | * |
| 719 | * Derivation: |
| 720 | * |
| 721 | * End offset is the highest offset that this page should represent. |
| 722 | * If we are on the last page, (end_offset & (PAGE_CACHE_SIZE - 1)) |
| 723 | * will evaluate non-zero and be less than PAGE_CACHE_SIZE and |
| 724 | * hence give us the correct page_dirty count. On any other page, |
| 725 | * it will be zero and in that case we need page_dirty to be the |
| 726 | * count of buffers on the page. |
| 727 | */ |
| 728 | end_offset = min_t(unsigned long long, |
| 729 | (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT, |
| 730 | i_size_read(inode)); |
| 731 | |
| 732 | /* |
| 733 | * If the current map does not span the entire page we are about to try |
| 734 | * to write, then give up. The only way we can write a page that spans |
| 735 | * multiple mappings in a single writeback iteration is via the |
| 736 | * xfs_vm_writepage() function. Data integrity writeback requires the |
| 737 | * entire page to be written in a single attempt, otherwise the part of |
| 738 | * the page we don't write here doesn't get written as part of the data |
| 739 | * integrity sync. |
| 740 | * |
| 741 | * For normal writeback, we also don't attempt to write partial pages |
| 742 | * here as it simply means that write_cache_pages() will see it under |
| 743 | * writeback and ignore the page until some point in the future, at |
| 744 | * which time this will be the only page in the file that needs |
| 745 | * writeback. Hence for more optimal IO patterns, we should always |
| 746 | * avoid partial page writeback due to multiple mappings on a page here. |
| 747 | */ |
| 748 | if (!xfs_imap_valid(inode, imap, end_offset)) |
| 749 | goto fail_unlock_page; |
| 750 | |
| 751 | len = 1 << inode->i_blkbits; |
| 752 | p_offset = min_t(unsigned long, end_offset & (PAGE_CACHE_SIZE - 1), |
| 753 | PAGE_CACHE_SIZE); |
| 754 | p_offset = p_offset ? roundup(p_offset, len) : PAGE_CACHE_SIZE; |
| 755 | page_dirty = p_offset / len; |
| 756 | |
| 757 | /* |
| 758 | * The moment we find a buffer that doesn't match our current type |
| 759 | * specification or can't be written, abort the loop and start |
| 760 | * writeback. As per the above xfs_imap_valid() check, only |
| 761 | * xfs_vm_writepage() can handle partial page writeback fully - we are |
| 762 | * limited here to the buffers that are contiguous with the current |
| 763 | * ioend, and hence a buffer we can't write breaks that contiguity and |
| 764 | * we have to defer the rest of the IO to xfs_vm_writepage(). |
| 765 | */ |
| 766 | bh = head = page_buffers(page); |
| 767 | do { |
| 768 | if (offset >= end_offset) |
| 769 | break; |
| 770 | if (!buffer_uptodate(bh)) |
| 771 | uptodate = 0; |
| 772 | if (!(PageUptodate(page) || buffer_uptodate(bh))) { |
| 773 | done = 1; |
| 774 | break; |
| 775 | } |
| 776 | |
| 777 | if (buffer_unwritten(bh) || buffer_delay(bh) || |
| 778 | buffer_mapped(bh)) { |
| 779 | if (buffer_unwritten(bh)) |
| 780 | type = XFS_IO_UNWRITTEN; |
| 781 | else if (buffer_delay(bh)) |
| 782 | type = XFS_IO_DELALLOC; |
| 783 | else |
| 784 | type = XFS_IO_OVERWRITE; |
| 785 | |
| 786 | /* |
| 787 | * imap should always be valid because of the above |
| 788 | * partial page end_offset check on the imap. |
| 789 | */ |
| 790 | ASSERT(xfs_imap_valid(inode, imap, offset)); |
| 791 | |
| 792 | lock_buffer(bh); |
| 793 | if (type != XFS_IO_OVERWRITE) |
| 794 | xfs_map_at_offset(inode, bh, imap, offset); |
| 795 | xfs_add_to_ioend(inode, bh, offset, type, |
| 796 | ioendp, done); |
| 797 | |
| 798 | page_dirty--; |
| 799 | count++; |
| 800 | } else { |
| 801 | done = 1; |
| 802 | break; |
| 803 | } |
| 804 | } while (offset += len, (bh = bh->b_this_page) != head); |
| 805 | |
| 806 | if (uptodate && bh == head) |
| 807 | SetPageUptodate(page); |
| 808 | |
| 809 | if (count) { |
| 810 | if (--wbc->nr_to_write <= 0 && |
| 811 | wbc->sync_mode == WB_SYNC_NONE) |
| 812 | done = 1; |
| 813 | } |
| 814 | xfs_start_page_writeback(page, !page_dirty, count); |
| 815 | |
| 816 | return done; |
| 817 | fail_unlock_page: |
| 818 | unlock_page(page); |
| 819 | fail: |
| 820 | return 1; |
| 821 | } |
| 822 | |
| 823 | /* |
| 824 | * Convert & write out a cluster of pages in the same extent as defined |
| 825 | * by mp and following the start page. |
| 826 | */ |
| 827 | STATIC void |
| 828 | xfs_cluster_write( |
| 829 | struct inode *inode, |
| 830 | pgoff_t tindex, |
| 831 | struct xfs_bmbt_irec *imap, |
| 832 | xfs_ioend_t **ioendp, |
| 833 | struct writeback_control *wbc, |
| 834 | pgoff_t tlast) |
| 835 | { |
| 836 | struct pagevec pvec; |
| 837 | int done = 0, i; |
| 838 | |
| 839 | pagevec_init(&pvec, 0); |
| 840 | while (!done && tindex <= tlast) { |
| 841 | unsigned len = min_t(pgoff_t, PAGEVEC_SIZE, tlast - tindex + 1); |
| 842 | |
| 843 | if (!pagevec_lookup(&pvec, inode->i_mapping, tindex, len)) |
| 844 | break; |
| 845 | |
| 846 | for (i = 0; i < pagevec_count(&pvec); i++) { |
| 847 | done = xfs_convert_page(inode, pvec.pages[i], tindex++, |
| 848 | imap, ioendp, wbc); |
| 849 | if (done) |
| 850 | break; |
| 851 | } |
| 852 | |
| 853 | pagevec_release(&pvec); |
| 854 | cond_resched(); |
| 855 | } |
| 856 | } |
| 857 | |
| 858 | STATIC void |
| 859 | xfs_vm_invalidatepage( |
| 860 | struct page *page, |
| 861 | unsigned int offset, |
| 862 | unsigned int length) |
| 863 | { |
| 864 | trace_xfs_invalidatepage(page->mapping->host, page, offset, |
| 865 | length); |
| 866 | block_invalidatepage(page, offset, length); |
| 867 | } |
| 868 | |
| 869 | /* |
| 870 | * If the page has delalloc buffers on it, we need to punch them out before we |
| 871 | * invalidate the page. If we don't, we leave a stale delalloc mapping on the |
| 872 | * inode that can trip a BUG() in xfs_get_blocks() later on if a direct IO read |
| 873 | * is done on that same region - the delalloc extent is returned when none is |
| 874 | * supposed to be there. |
| 875 | * |
| 876 | * We prevent this by truncating away the delalloc regions on the page before |
| 877 | * invalidating it. Because they are delalloc, we can do this without needing a |
| 878 | * transaction. Indeed - if we get ENOSPC errors, we have to be able to do this |
| 879 | * truncation without a transaction as there is no space left for block |
| 880 | * reservation (typically why we see a ENOSPC in writeback). |
| 881 | * |
| 882 | * This is not a performance critical path, so for now just do the punching a |
| 883 | * buffer head at a time. |
| 884 | */ |
| 885 | STATIC void |
| 886 | xfs_aops_discard_page( |
| 887 | struct page *page) |
| 888 | { |
| 889 | struct inode *inode = page->mapping->host; |
| 890 | struct xfs_inode *ip = XFS_I(inode); |
| 891 | struct buffer_head *bh, *head; |
| 892 | loff_t offset = page_offset(page); |
| 893 | |
| 894 | if (!xfs_check_page_type(page, XFS_IO_DELALLOC, true)) |
| 895 | goto out_invalidate; |
| 896 | |
| 897 | if (XFS_FORCED_SHUTDOWN(ip->i_mount)) |
| 898 | goto out_invalidate; |
| 899 | |
| 900 | xfs_alert(ip->i_mount, |
| 901 | "page discard on page %p, inode 0x%llx, offset %llu.", |
| 902 | page, ip->i_ino, offset); |
| 903 | |
| 904 | xfs_ilock(ip, XFS_ILOCK_EXCL); |
| 905 | bh = head = page_buffers(page); |
| 906 | do { |
| 907 | int error; |
| 908 | xfs_fileoff_t start_fsb; |
| 909 | |
| 910 | if (!buffer_delay(bh)) |
| 911 | goto next_buffer; |
| 912 | |
| 913 | start_fsb = XFS_B_TO_FSBT(ip->i_mount, offset); |
| 914 | error = xfs_bmap_punch_delalloc_range(ip, start_fsb, 1); |
| 915 | if (error) { |
| 916 | /* something screwed, just bail */ |
| 917 | if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) { |
| 918 | xfs_alert(ip->i_mount, |
| 919 | "page discard unable to remove delalloc mapping."); |
| 920 | } |
| 921 | break; |
| 922 | } |
| 923 | next_buffer: |
| 924 | offset += i_blocksize(inode); |
| 925 | |
| 926 | } while ((bh = bh->b_this_page) != head); |
| 927 | |
| 928 | xfs_iunlock(ip, XFS_ILOCK_EXCL); |
| 929 | out_invalidate: |
| 930 | xfs_vm_invalidatepage(page, 0, PAGE_CACHE_SIZE); |
| 931 | return; |
| 932 | } |
| 933 | |
| 934 | /* |
| 935 | * Write out a dirty page. |
| 936 | * |
| 937 | * For delalloc space on the page we need to allocate space and flush it. |
| 938 | * For unwritten space on the page we need to start the conversion to |
| 939 | * regular allocated space. |
| 940 | * For any other dirty buffer heads on the page we should flush them. |
| 941 | */ |
| 942 | STATIC int |
| 943 | xfs_vm_writepage( |
| 944 | struct page *page, |
| 945 | struct writeback_control *wbc) |
| 946 | { |
| 947 | struct inode *inode = page->mapping->host; |
| 948 | struct buffer_head *bh, *head; |
| 949 | struct xfs_bmbt_irec imap; |
| 950 | xfs_ioend_t *ioend = NULL, *iohead = NULL; |
| 951 | loff_t offset; |
| 952 | unsigned int type; |
| 953 | __uint64_t end_offset; |
| 954 | pgoff_t end_index, last_index; |
| 955 | ssize_t len; |
| 956 | int err, imap_valid = 0, uptodate = 1; |
| 957 | int count = 0; |
| 958 | int nonblocking = 0; |
| 959 | |
| 960 | trace_xfs_writepage(inode, page, 0, 0); |
| 961 | |
| 962 | ASSERT(page_has_buffers(page)); |
| 963 | |
| 964 | /* |
| 965 | * Refuse to write the page out if we are called from reclaim context. |
| 966 | * |
| 967 | * This avoids stack overflows when called from deeply used stacks in |
| 968 | * random callers for direct reclaim or memcg reclaim. We explicitly |
| 969 | * allow reclaim from kswapd as the stack usage there is relatively low. |
| 970 | * |
| 971 | * This should never happen except in the case of a VM regression so |
| 972 | * warn about it. |
| 973 | */ |
| 974 | if (WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD)) == |
| 975 | PF_MEMALLOC)) |
| 976 | goto redirty; |
| 977 | |
| 978 | /* |
| 979 | * Given that we do not allow direct reclaim to call us, we should |
| 980 | * never be called while in a filesystem transaction. |
| 981 | */ |
| 982 | if (WARN_ON_ONCE(current->flags & PF_FSTRANS)) |
| 983 | goto redirty; |
| 984 | |
| 985 | /* Is this page beyond the end of the file? */ |
| 986 | offset = i_size_read(inode); |
| 987 | end_index = offset >> PAGE_CACHE_SHIFT; |
| 988 | last_index = (offset - 1) >> PAGE_CACHE_SHIFT; |
| 989 | |
| 990 | /* |
| 991 | * The page index is less than the end_index, adjust the end_offset |
| 992 | * to the highest offset that this page should represent. |
| 993 | * ----------------------------------------------------- |
| 994 | * | file mapping | <EOF> | |
| 995 | * ----------------------------------------------------- |
| 996 | * | Page ... | Page N-2 | Page N-1 | Page N | | |
| 997 | * ^--------------------------------^----------|-------- |
| 998 | * | desired writeback range | see else | |
| 999 | * ---------------------------------^------------------| |
| 1000 | */ |
| 1001 | if (page->index < end_index) |
| 1002 | end_offset = (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT; |
| 1003 | else { |
| 1004 | /* |
| 1005 | * Check whether the page to write out is beyond or straddles |
| 1006 | * i_size or not. |
| 1007 | * ------------------------------------------------------- |
| 1008 | * | file mapping | <EOF> | |
| 1009 | * ------------------------------------------------------- |
| 1010 | * | Page ... | Page N-2 | Page N-1 | Page N | Beyond | |
| 1011 | * ^--------------------------------^-----------|--------- |
| 1012 | * | | Straddles | |
| 1013 | * ---------------------------------^-----------|--------| |
| 1014 | */ |
| 1015 | unsigned offset_into_page = offset & (PAGE_CACHE_SIZE - 1); |
| 1016 | |
| 1017 | /* |
| 1018 | * Skip the page if it is fully outside i_size, e.g. due to a |
| 1019 | * truncate operation that is in progress. We must redirty the |
| 1020 | * page so that reclaim stops reclaiming it. Otherwise |
| 1021 | * xfs_vm_releasepage() is called on it and gets confused. |
| 1022 | * |
| 1023 | * Note that the end_index is unsigned long, it would overflow |
| 1024 | * if the given offset is greater than 16TB on 32-bit system |
| 1025 | * and if we do check the page is fully outside i_size or not |
| 1026 | * via "if (page->index >= end_index + 1)" as "end_index + 1" |
| 1027 | * will be evaluated to 0. Hence this page will be redirtied |
| 1028 | * and be written out repeatedly which would result in an |
| 1029 | * infinite loop, the user program that perform this operation |
| 1030 | * will hang. Instead, we can verify this situation by checking |
| 1031 | * if the page to write is totally beyond the i_size or if it's |
| 1032 | * offset is just equal to the EOF. |
| 1033 | */ |
| 1034 | if (page->index > end_index || |
| 1035 | (page->index == end_index && offset_into_page == 0)) |
| 1036 | goto redirty; |
| 1037 | |
| 1038 | /* |
| 1039 | * The page straddles i_size. It must be zeroed out on each |
| 1040 | * and every writepage invocation because it may be mmapped. |
| 1041 | * "A file is mapped in multiples of the page size. For a file |
| 1042 | * that is not a multiple of the page size, the remaining |
| 1043 | * memory is zeroed when mapped, and writes to that region are |
| 1044 | * not written out to the file." |
| 1045 | */ |
| 1046 | zero_user_segment(page, offset_into_page, PAGE_CACHE_SIZE); |
| 1047 | |
| 1048 | /* Adjust the end_offset to the end of file */ |
| 1049 | end_offset = offset; |
| 1050 | } |
| 1051 | |
| 1052 | len = 1 << inode->i_blkbits; |
| 1053 | |
| 1054 | bh = head = page_buffers(page); |
| 1055 | offset = page_offset(page); |
| 1056 | type = XFS_IO_OVERWRITE; |
| 1057 | |
| 1058 | if (wbc->sync_mode == WB_SYNC_NONE) |
| 1059 | nonblocking = 1; |
| 1060 | |
| 1061 | do { |
| 1062 | int new_ioend = 0; |
| 1063 | |
| 1064 | if (offset >= end_offset) |
| 1065 | break; |
| 1066 | if (!buffer_uptodate(bh)) |
| 1067 | uptodate = 0; |
| 1068 | |
| 1069 | /* |
| 1070 | * set_page_dirty dirties all buffers in a page, independent |
| 1071 | * of their state. The dirty state however is entirely |
| 1072 | * meaningless for holes (!mapped && uptodate), so skip |
| 1073 | * buffers covering holes here. |
| 1074 | */ |
| 1075 | if (!buffer_mapped(bh) && buffer_uptodate(bh)) { |
| 1076 | imap_valid = 0; |
| 1077 | continue; |
| 1078 | } |
| 1079 | |
| 1080 | if (buffer_unwritten(bh)) { |
| 1081 | if (type != XFS_IO_UNWRITTEN) { |
| 1082 | type = XFS_IO_UNWRITTEN; |
| 1083 | imap_valid = 0; |
| 1084 | } |
| 1085 | } else if (buffer_delay(bh)) { |
| 1086 | if (type != XFS_IO_DELALLOC) { |
| 1087 | type = XFS_IO_DELALLOC; |
| 1088 | imap_valid = 0; |
| 1089 | } |
| 1090 | } else if (buffer_uptodate(bh)) { |
| 1091 | if (type != XFS_IO_OVERWRITE) { |
| 1092 | type = XFS_IO_OVERWRITE; |
| 1093 | imap_valid = 0; |
| 1094 | } |
| 1095 | } else { |
| 1096 | if (PageUptodate(page)) |
| 1097 | ASSERT(buffer_mapped(bh)); |
| 1098 | /* |
| 1099 | * This buffer is not uptodate and will not be |
| 1100 | * written to disk. Ensure that we will put any |
| 1101 | * subsequent writeable buffers into a new |
| 1102 | * ioend. |
| 1103 | */ |
| 1104 | imap_valid = 0; |
| 1105 | continue; |
| 1106 | } |
| 1107 | |
| 1108 | if (imap_valid) |
| 1109 | imap_valid = xfs_imap_valid(inode, &imap, offset); |
| 1110 | if (!imap_valid) { |
| 1111 | /* |
| 1112 | * If we didn't have a valid mapping then we need to |
| 1113 | * put the new mapping into a separate ioend structure. |
| 1114 | * This ensures non-contiguous extents always have |
| 1115 | * separate ioends, which is particularly important |
| 1116 | * for unwritten extent conversion at I/O completion |
| 1117 | * time. |
| 1118 | */ |
| 1119 | new_ioend = 1; |
| 1120 | err = xfs_map_blocks(inode, offset, &imap, type, |
| 1121 | nonblocking); |
| 1122 | if (err) |
| 1123 | goto error; |
| 1124 | imap_valid = xfs_imap_valid(inode, &imap, offset); |
| 1125 | } |
| 1126 | if (imap_valid) { |
| 1127 | lock_buffer(bh); |
| 1128 | if (type != XFS_IO_OVERWRITE) |
| 1129 | xfs_map_at_offset(inode, bh, &imap, offset); |
| 1130 | xfs_add_to_ioend(inode, bh, offset, type, &ioend, |
| 1131 | new_ioend); |
| 1132 | count++; |
| 1133 | } |
| 1134 | |
| 1135 | if (!iohead) |
| 1136 | iohead = ioend; |
| 1137 | |
| 1138 | } while (offset += len, ((bh = bh->b_this_page) != head)); |
| 1139 | |
| 1140 | if (uptodate && bh == head) |
| 1141 | SetPageUptodate(page); |
| 1142 | |
| 1143 | xfs_start_page_writeback(page, 1, count); |
| 1144 | |
| 1145 | /* if there is no IO to be submitted for this page, we are done */ |
| 1146 | if (!ioend) |
| 1147 | return 0; |
| 1148 | |
| 1149 | ASSERT(iohead); |
| 1150 | |
| 1151 | /* |
| 1152 | * Any errors from this point onwards need tobe reported through the IO |
| 1153 | * completion path as we have marked the initial page as under writeback |
| 1154 | * and unlocked it. |
| 1155 | */ |
| 1156 | if (imap_valid) { |
| 1157 | xfs_off_t end_index; |
| 1158 | |
| 1159 | end_index = imap.br_startoff + imap.br_blockcount; |
| 1160 | |
| 1161 | /* to bytes */ |
| 1162 | end_index <<= inode->i_blkbits; |
| 1163 | |
| 1164 | /* to pages */ |
| 1165 | end_index = (end_index - 1) >> PAGE_CACHE_SHIFT; |
| 1166 | |
| 1167 | /* check against file size */ |
| 1168 | if (end_index > last_index) |
| 1169 | end_index = last_index; |
| 1170 | |
| 1171 | xfs_cluster_write(inode, page->index + 1, &imap, &ioend, |
| 1172 | wbc, end_index); |
| 1173 | } |
| 1174 | |
| 1175 | |
| 1176 | /* |
| 1177 | * Reserve log space if we might write beyond the on-disk inode size. |
| 1178 | */ |
| 1179 | err = 0; |
| 1180 | if (ioend->io_type != XFS_IO_UNWRITTEN && xfs_ioend_is_append(ioend)) |
| 1181 | err = xfs_setfilesize_trans_alloc(ioend); |
| 1182 | |
| 1183 | xfs_submit_ioend(wbc, iohead, err); |
| 1184 | |
| 1185 | return 0; |
| 1186 | |
| 1187 | error: |
| 1188 | if (iohead) |
| 1189 | xfs_cancel_ioend(iohead); |
| 1190 | |
| 1191 | if (err == -EAGAIN) |
| 1192 | goto redirty; |
| 1193 | |
| 1194 | xfs_aops_discard_page(page); |
| 1195 | ClearPageUptodate(page); |
| 1196 | unlock_page(page); |
| 1197 | return err; |
| 1198 | |
| 1199 | redirty: |
| 1200 | redirty_page_for_writepage(wbc, page); |
| 1201 | unlock_page(page); |
| 1202 | return 0; |
| 1203 | } |
| 1204 | |
| 1205 | STATIC int |
| 1206 | xfs_vm_writepages( |
| 1207 | struct address_space *mapping, |
| 1208 | struct writeback_control *wbc) |
| 1209 | { |
| 1210 | xfs_iflags_clear(XFS_I(mapping->host), XFS_ITRUNCATED); |
| 1211 | return generic_writepages(mapping, wbc); |
| 1212 | } |
| 1213 | |
| 1214 | /* |
| 1215 | * Called to move a page into cleanable state - and from there |
| 1216 | * to be released. The page should already be clean. We always |
| 1217 | * have buffer heads in this call. |
| 1218 | * |
| 1219 | * Returns 1 if the page is ok to release, 0 otherwise. |
| 1220 | */ |
| 1221 | STATIC int |
| 1222 | xfs_vm_releasepage( |
| 1223 | struct page *page, |
| 1224 | gfp_t gfp_mask) |
| 1225 | { |
| 1226 | int delalloc, unwritten; |
| 1227 | |
| 1228 | trace_xfs_releasepage(page->mapping->host, page, 0, 0); |
| 1229 | |
| 1230 | xfs_count_page_state(page, &delalloc, &unwritten); |
| 1231 | |
| 1232 | if (WARN_ON_ONCE(delalloc)) |
| 1233 | return 0; |
| 1234 | if (WARN_ON_ONCE(unwritten)) |
| 1235 | return 0; |
| 1236 | |
| 1237 | return try_to_free_buffers(page); |
| 1238 | } |
| 1239 | |
| 1240 | /* |
| 1241 | * When we map a DIO buffer, we may need to attach an ioend that describes the |
| 1242 | * type of write IO we are doing. This passes to the completion function the |
| 1243 | * operations it needs to perform. If the mapping is for an overwrite wholly |
| 1244 | * within the EOF then we don't need an ioend and so we don't allocate one. |
| 1245 | * This avoids the unnecessary overhead of allocating and freeing ioends for |
| 1246 | * workloads that don't require transactions on IO completion. |
| 1247 | * |
| 1248 | * If we get multiple mappings in a single IO, we might be mapping different |
| 1249 | * types. But because the direct IO can only have a single private pointer, we |
| 1250 | * need to ensure that: |
| 1251 | * |
| 1252 | * a) i) the ioend spans the entire region of unwritten mappings; or |
| 1253 | * ii) the ioend spans all the mappings that cross or are beyond EOF; and |
| 1254 | * b) if it contains unwritten extents, it is *permanently* marked as such |
| 1255 | * |
| 1256 | * We could do this by chaining ioends like buffered IO does, but we only |
| 1257 | * actually get one IO completion callback from the direct IO, and that spans |
| 1258 | * the entire IO regardless of how many mappings and IOs are needed to complete |
| 1259 | * the DIO. There is only going to be one reference to the ioend and its life |
| 1260 | * cycle is constrained by the DIO completion code. hence we don't need |
| 1261 | * reference counting here. |
| 1262 | * |
| 1263 | * Note that for DIO, an IO to the highest supported file block offset (i.e. |
| 1264 | * 2^63 - 1FSB bytes) will result in the offset + count overflowing a signed 64 |
| 1265 | * bit variable. Hence if we see this overflow, we have to assume that the IO is |
| 1266 | * extending the file size. We won't know for sure until IO completion is run |
| 1267 | * and the actual max write offset is communicated to the IO completion |
| 1268 | * routine. |
| 1269 | * |
| 1270 | * For DAX page faults, we are preparing to never see unwritten extents here, |
| 1271 | * nor should we ever extend the inode size. Hence we will soon have nothing to |
| 1272 | * do here for this case, ensuring we don't have to provide an IO completion |
| 1273 | * callback to free an ioend that we don't actually need for a fault into the |
| 1274 | * page at offset (2^63 - 1FSB) bytes. |
| 1275 | */ |
| 1276 | |
| 1277 | static void |
| 1278 | xfs_map_direct( |
| 1279 | struct inode *inode, |
| 1280 | struct buffer_head *bh_result, |
| 1281 | struct xfs_bmbt_irec *imap, |
| 1282 | xfs_off_t offset, |
| 1283 | bool dax_fault) |
| 1284 | { |
| 1285 | struct xfs_ioend *ioend; |
| 1286 | xfs_off_t size = bh_result->b_size; |
| 1287 | int type; |
| 1288 | |
| 1289 | if (ISUNWRITTEN(imap)) |
| 1290 | type = XFS_IO_UNWRITTEN; |
| 1291 | else |
| 1292 | type = XFS_IO_OVERWRITE; |
| 1293 | |
| 1294 | trace_xfs_gbmap_direct(XFS_I(inode), offset, size, type, imap); |
| 1295 | |
| 1296 | if (dax_fault) { |
| 1297 | ASSERT(type == XFS_IO_OVERWRITE); |
| 1298 | trace_xfs_gbmap_direct_none(XFS_I(inode), offset, size, type, |
| 1299 | imap); |
| 1300 | return; |
| 1301 | } |
| 1302 | |
| 1303 | if (bh_result->b_private) { |
| 1304 | ioend = bh_result->b_private; |
| 1305 | ASSERT(ioend->io_size > 0); |
| 1306 | ASSERT(offset >= ioend->io_offset); |
| 1307 | if (offset + size > ioend->io_offset + ioend->io_size) |
| 1308 | ioend->io_size = offset - ioend->io_offset + size; |
| 1309 | |
| 1310 | if (type == XFS_IO_UNWRITTEN && type != ioend->io_type) |
| 1311 | ioend->io_type = XFS_IO_UNWRITTEN; |
| 1312 | |
| 1313 | trace_xfs_gbmap_direct_update(XFS_I(inode), ioend->io_offset, |
| 1314 | ioend->io_size, ioend->io_type, |
| 1315 | imap); |
| 1316 | } else if (type == XFS_IO_UNWRITTEN || |
| 1317 | offset + size > i_size_read(inode) || |
| 1318 | offset + size < 0) { |
| 1319 | ioend = xfs_alloc_ioend(inode, type); |
| 1320 | ioend->io_offset = offset; |
| 1321 | ioend->io_size = size; |
| 1322 | |
| 1323 | bh_result->b_private = ioend; |
| 1324 | set_buffer_defer_completion(bh_result); |
| 1325 | |
| 1326 | trace_xfs_gbmap_direct_new(XFS_I(inode), offset, size, type, |
| 1327 | imap); |
| 1328 | } else { |
| 1329 | trace_xfs_gbmap_direct_none(XFS_I(inode), offset, size, type, |
| 1330 | imap); |
| 1331 | } |
| 1332 | } |
| 1333 | |
| 1334 | /* |
| 1335 | * If this is O_DIRECT or the mpage code calling tell them how large the mapping |
| 1336 | * is, so that we can avoid repeated get_blocks calls. |
| 1337 | * |
| 1338 | * If the mapping spans EOF, then we have to break the mapping up as the mapping |
| 1339 | * for blocks beyond EOF must be marked new so that sub block regions can be |
| 1340 | * correctly zeroed. We can't do this for mappings within EOF unless the mapping |
| 1341 | * was just allocated or is unwritten, otherwise the callers would overwrite |
| 1342 | * existing data with zeros. Hence we have to split the mapping into a range up |
| 1343 | * to and including EOF, and a second mapping for beyond EOF. |
| 1344 | */ |
| 1345 | static void |
| 1346 | xfs_map_trim_size( |
| 1347 | struct inode *inode, |
| 1348 | sector_t iblock, |
| 1349 | struct buffer_head *bh_result, |
| 1350 | struct xfs_bmbt_irec *imap, |
| 1351 | xfs_off_t offset, |
| 1352 | ssize_t size) |
| 1353 | { |
| 1354 | xfs_off_t mapping_size; |
| 1355 | |
| 1356 | mapping_size = imap->br_startoff + imap->br_blockcount - iblock; |
| 1357 | mapping_size <<= inode->i_blkbits; |
| 1358 | |
| 1359 | ASSERT(mapping_size > 0); |
| 1360 | if (mapping_size > size) |
| 1361 | mapping_size = size; |
| 1362 | if (offset < i_size_read(inode) && |
| 1363 | offset + mapping_size >= i_size_read(inode)) { |
| 1364 | /* limit mapping to block that spans EOF */ |
| 1365 | mapping_size = roundup_64(i_size_read(inode) - offset, |
| 1366 | i_blocksize(inode)); |
| 1367 | } |
| 1368 | if (mapping_size > LONG_MAX) |
| 1369 | mapping_size = LONG_MAX; |
| 1370 | |
| 1371 | bh_result->b_size = mapping_size; |
| 1372 | } |
| 1373 | |
| 1374 | STATIC int |
| 1375 | __xfs_get_blocks( |
| 1376 | struct inode *inode, |
| 1377 | sector_t iblock, |
| 1378 | struct buffer_head *bh_result, |
| 1379 | int create, |
| 1380 | bool direct, |
| 1381 | bool dax_fault) |
| 1382 | { |
| 1383 | struct xfs_inode *ip = XFS_I(inode); |
| 1384 | struct xfs_mount *mp = ip->i_mount; |
| 1385 | xfs_fileoff_t offset_fsb, end_fsb; |
| 1386 | int error = 0; |
| 1387 | int lockmode = 0; |
| 1388 | struct xfs_bmbt_irec imap; |
| 1389 | int nimaps = 1; |
| 1390 | xfs_off_t offset; |
| 1391 | ssize_t size; |
| 1392 | int new = 0; |
| 1393 | |
| 1394 | if (XFS_FORCED_SHUTDOWN(mp)) |
| 1395 | return -EIO; |
| 1396 | |
| 1397 | offset = (xfs_off_t)iblock << inode->i_blkbits; |
| 1398 | ASSERT(bh_result->b_size >= i_blocksize(inode)); |
| 1399 | size = bh_result->b_size; |
| 1400 | |
| 1401 | if (!create && direct && offset >= i_size_read(inode)) |
| 1402 | return 0; |
| 1403 | |
| 1404 | /* |
| 1405 | * Direct I/O is usually done on preallocated files, so try getting |
| 1406 | * a block mapping without an exclusive lock first. For buffered |
| 1407 | * writes we already have the exclusive iolock anyway, so avoiding |
| 1408 | * a lock roundtrip here by taking the ilock exclusive from the |
| 1409 | * beginning is a useful micro optimization. |
| 1410 | */ |
| 1411 | if (create && !direct) { |
| 1412 | lockmode = XFS_ILOCK_EXCL; |
| 1413 | xfs_ilock(ip, lockmode); |
| 1414 | } else { |
| 1415 | lockmode = xfs_ilock_data_map_shared(ip); |
| 1416 | } |
| 1417 | |
| 1418 | ASSERT(offset <= mp->m_super->s_maxbytes); |
| 1419 | if (offset + size > mp->m_super->s_maxbytes) |
| 1420 | size = mp->m_super->s_maxbytes - offset; |
| 1421 | end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + size); |
| 1422 | offset_fsb = XFS_B_TO_FSBT(mp, offset); |
| 1423 | |
| 1424 | error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb, |
| 1425 | &imap, &nimaps, XFS_BMAPI_ENTIRE); |
| 1426 | if (error) |
| 1427 | goto out_unlock; |
| 1428 | |
| 1429 | /* |
| 1430 | * The only time we can ever safely find delalloc blocks on direct I/O |
| 1431 | * is a dio write to post-eof speculative preallocation. All other |
| 1432 | * scenarios are indicative of a problem or misuse (such as mixing |
| 1433 | * direct and mapped I/O). |
| 1434 | * |
| 1435 | * The file may be unmapped by the time we get here so we cannot |
| 1436 | * reliably fail the I/O based on mapping. Instead, fail the I/O if this |
| 1437 | * is a read or a write within eof. Otherwise, carry on but warn as a |
| 1438 | * precuation if the file happens to be mapped. |
| 1439 | */ |
| 1440 | if (direct && imap.br_startblock == DELAYSTARTBLOCK) { |
| 1441 | if (!create || offset < i_size_read(VFS_I(ip))) { |
| 1442 | WARN_ON_ONCE(1); |
| 1443 | error = -EIO; |
| 1444 | goto out_unlock; |
| 1445 | } |
| 1446 | WARN_ON_ONCE(mapping_mapped(VFS_I(ip)->i_mapping)); |
| 1447 | } |
| 1448 | |
| 1449 | /* for DAX, we convert unwritten extents directly */ |
| 1450 | if (create && |
| 1451 | (!nimaps || |
| 1452 | (imap.br_startblock == HOLESTARTBLOCK || |
| 1453 | imap.br_startblock == DELAYSTARTBLOCK) || |
| 1454 | (IS_DAX(inode) && ISUNWRITTEN(&imap)))) { |
| 1455 | if (direct || xfs_get_extsz_hint(ip)) { |
| 1456 | /* |
| 1457 | * xfs_iomap_write_direct() expects the shared lock. It |
| 1458 | * is unlocked on return. |
| 1459 | */ |
| 1460 | if (lockmode == XFS_ILOCK_EXCL) |
| 1461 | xfs_ilock_demote(ip, lockmode); |
| 1462 | |
| 1463 | error = xfs_iomap_write_direct(ip, offset, size, |
| 1464 | &imap, nimaps); |
| 1465 | if (error) |
| 1466 | return error; |
| 1467 | new = 1; |
| 1468 | |
| 1469 | } else { |
| 1470 | /* |
| 1471 | * Delalloc reservations do not require a transaction, |
| 1472 | * we can go on without dropping the lock here. If we |
| 1473 | * are allocating a new delalloc block, make sure that |
| 1474 | * we set the new flag so that we mark the buffer new so |
| 1475 | * that we know that it is newly allocated if the write |
| 1476 | * fails. |
| 1477 | */ |
| 1478 | if (nimaps && imap.br_startblock == HOLESTARTBLOCK) |
| 1479 | new = 1; |
| 1480 | error = xfs_iomap_write_delay(ip, offset, size, &imap); |
| 1481 | if (error) |
| 1482 | goto out_unlock; |
| 1483 | |
| 1484 | xfs_iunlock(ip, lockmode); |
| 1485 | } |
| 1486 | trace_xfs_get_blocks_alloc(ip, offset, size, |
| 1487 | ISUNWRITTEN(&imap) ? XFS_IO_UNWRITTEN |
| 1488 | : XFS_IO_DELALLOC, &imap); |
| 1489 | } else if (nimaps) { |
| 1490 | trace_xfs_get_blocks_found(ip, offset, size, |
| 1491 | ISUNWRITTEN(&imap) ? XFS_IO_UNWRITTEN |
| 1492 | : XFS_IO_OVERWRITE, &imap); |
| 1493 | xfs_iunlock(ip, lockmode); |
| 1494 | } else { |
| 1495 | trace_xfs_get_blocks_notfound(ip, offset, size); |
| 1496 | goto out_unlock; |
| 1497 | } |
| 1498 | |
| 1499 | if (IS_DAX(inode) && create) { |
| 1500 | ASSERT(!ISUNWRITTEN(&imap)); |
| 1501 | /* zeroing is not needed at a higher layer */ |
| 1502 | new = 0; |
| 1503 | } |
| 1504 | |
| 1505 | /* trim mapping down to size requested */ |
| 1506 | if (direct || size > (1 << inode->i_blkbits)) |
| 1507 | xfs_map_trim_size(inode, iblock, bh_result, |
| 1508 | &imap, offset, size); |
| 1509 | |
| 1510 | /* |
| 1511 | * For unwritten extents do not report a disk address in the buffered |
| 1512 | * read case (treat as if we're reading into a hole). |
| 1513 | */ |
| 1514 | if (imap.br_startblock != HOLESTARTBLOCK && |
| 1515 | imap.br_startblock != DELAYSTARTBLOCK && |
| 1516 | (create || !ISUNWRITTEN(&imap))) { |
| 1517 | xfs_map_buffer(inode, bh_result, &imap, offset); |
| 1518 | if (ISUNWRITTEN(&imap)) |
| 1519 | set_buffer_unwritten(bh_result); |
| 1520 | /* direct IO needs special help */ |
| 1521 | if (create && direct) |
| 1522 | xfs_map_direct(inode, bh_result, &imap, offset, |
| 1523 | dax_fault); |
| 1524 | } |
| 1525 | |
| 1526 | /* |
| 1527 | * If this is a realtime file, data may be on a different device. |
| 1528 | * to that pointed to from the buffer_head b_bdev currently. |
| 1529 | */ |
| 1530 | bh_result->b_bdev = xfs_find_bdev_for_inode(inode); |
| 1531 | |
| 1532 | /* |
| 1533 | * If we previously allocated a block out beyond eof and we are now |
| 1534 | * coming back to use it then we will need to flag it as new even if it |
| 1535 | * has a disk address. |
| 1536 | * |
| 1537 | * With sub-block writes into unwritten extents we also need to mark |
| 1538 | * the buffer as new so that the unwritten parts of the buffer gets |
| 1539 | * correctly zeroed. |
| 1540 | */ |
| 1541 | if (create && |
| 1542 | ((!buffer_mapped(bh_result) && !buffer_uptodate(bh_result)) || |
| 1543 | (offset >= i_size_read(inode)) || |
| 1544 | (new || ISUNWRITTEN(&imap)))) |
| 1545 | set_buffer_new(bh_result); |
| 1546 | |
| 1547 | if (imap.br_startblock == DELAYSTARTBLOCK) { |
| 1548 | if (create) { |
| 1549 | set_buffer_uptodate(bh_result); |
| 1550 | set_buffer_mapped(bh_result); |
| 1551 | set_buffer_delay(bh_result); |
| 1552 | } |
| 1553 | } |
| 1554 | |
| 1555 | return 0; |
| 1556 | |
| 1557 | out_unlock: |
| 1558 | xfs_iunlock(ip, lockmode); |
| 1559 | return error; |
| 1560 | } |
| 1561 | |
| 1562 | int |
| 1563 | xfs_get_blocks( |
| 1564 | struct inode *inode, |
| 1565 | sector_t iblock, |
| 1566 | struct buffer_head *bh_result, |
| 1567 | int create) |
| 1568 | { |
| 1569 | return __xfs_get_blocks(inode, iblock, bh_result, create, false, false); |
| 1570 | } |
| 1571 | |
| 1572 | int |
| 1573 | xfs_get_blocks_direct( |
| 1574 | struct inode *inode, |
| 1575 | sector_t iblock, |
| 1576 | struct buffer_head *bh_result, |
| 1577 | int create) |
| 1578 | { |
| 1579 | return __xfs_get_blocks(inode, iblock, bh_result, create, true, false); |
| 1580 | } |
| 1581 | |
| 1582 | int |
| 1583 | xfs_get_blocks_dax_fault( |
| 1584 | struct inode *inode, |
| 1585 | sector_t iblock, |
| 1586 | struct buffer_head *bh_result, |
| 1587 | int create) |
| 1588 | { |
| 1589 | return __xfs_get_blocks(inode, iblock, bh_result, create, true, true); |
| 1590 | } |
| 1591 | |
| 1592 | static void |
| 1593 | __xfs_end_io_direct_write( |
| 1594 | struct inode *inode, |
| 1595 | struct xfs_ioend *ioend, |
| 1596 | loff_t offset, |
| 1597 | ssize_t size) |
| 1598 | { |
| 1599 | struct xfs_mount *mp = XFS_I(inode)->i_mount; |
| 1600 | |
| 1601 | if (XFS_FORCED_SHUTDOWN(mp) || ioend->io_error) |
| 1602 | goto out_end_io; |
| 1603 | |
| 1604 | /* |
| 1605 | * dio completion end_io functions are only called on writes if more |
| 1606 | * than 0 bytes was written. |
| 1607 | */ |
| 1608 | ASSERT(size > 0); |
| 1609 | |
| 1610 | /* |
| 1611 | * The ioend only maps whole blocks, while the IO may be sector aligned. |
| 1612 | * Hence the ioend offset/size may not match the IO offset/size exactly. |
| 1613 | * Because we don't map overwrites within EOF into the ioend, the offset |
| 1614 | * may not match, but only if the endio spans EOF. Either way, write |
| 1615 | * the IO sizes into the ioend so that completion processing does the |
| 1616 | * right thing. |
| 1617 | */ |
| 1618 | ASSERT(offset + size <= ioend->io_offset + ioend->io_size); |
| 1619 | ioend->io_size = size; |
| 1620 | ioend->io_offset = offset; |
| 1621 | |
| 1622 | /* |
| 1623 | * The ioend tells us whether we are doing unwritten extent conversion |
| 1624 | * or an append transaction that updates the on-disk file size. These |
| 1625 | * cases are the only cases where we should *potentially* be needing |
| 1626 | * to update the VFS inode size. |
| 1627 | * |
| 1628 | * We need to update the in-core inode size here so that we don't end up |
| 1629 | * with the on-disk inode size being outside the in-core inode size. We |
| 1630 | * have no other method of updating EOF for AIO, so always do it here |
| 1631 | * if necessary. |
| 1632 | * |
| 1633 | * We need to lock the test/set EOF update as we can be racing with |
| 1634 | * other IO completions here to update the EOF. Failing to serialise |
| 1635 | * here can result in EOF moving backwards and Bad Things Happen when |
| 1636 | * that occurs. |
| 1637 | */ |
| 1638 | spin_lock(&XFS_I(inode)->i_flags_lock); |
| 1639 | if (offset + size > i_size_read(inode)) |
| 1640 | i_size_write(inode, offset + size); |
| 1641 | spin_unlock(&XFS_I(inode)->i_flags_lock); |
| 1642 | |
| 1643 | /* |
| 1644 | * If we are doing an append IO that needs to update the EOF on disk, |
| 1645 | * do the transaction reserve now so we can use common end io |
| 1646 | * processing. Stashing the error (if there is one) in the ioend will |
| 1647 | * result in the ioend processing passing on the error if it is |
| 1648 | * possible as we can't return it from here. |
| 1649 | */ |
| 1650 | if (ioend->io_type == XFS_IO_OVERWRITE) |
| 1651 | ioend->io_error = xfs_setfilesize_trans_alloc(ioend); |
| 1652 | |
| 1653 | out_end_io: |
| 1654 | xfs_end_io(&ioend->io_work); |
| 1655 | return; |
| 1656 | } |
| 1657 | |
| 1658 | /* |
| 1659 | * Complete a direct I/O write request. |
| 1660 | * |
| 1661 | * The ioend structure is passed from __xfs_get_blocks() to tell us what to do. |
| 1662 | * If no ioend exists (i.e. @private == NULL) then the write IO is an overwrite |
| 1663 | * wholly within the EOF and so there is nothing for us to do. Note that in this |
| 1664 | * case the completion can be called in interrupt context, whereas if we have an |
| 1665 | * ioend we will always be called in task context (i.e. from a workqueue). |
| 1666 | */ |
| 1667 | STATIC void |
| 1668 | xfs_end_io_direct_write( |
| 1669 | struct kiocb *iocb, |
| 1670 | loff_t offset, |
| 1671 | ssize_t size, |
| 1672 | void *private) |
| 1673 | { |
| 1674 | struct inode *inode = file_inode(iocb->ki_filp); |
| 1675 | struct xfs_ioend *ioend = private; |
| 1676 | |
| 1677 | trace_xfs_gbmap_direct_endio(XFS_I(inode), offset, size, |
| 1678 | ioend ? ioend->io_type : 0, NULL); |
| 1679 | |
| 1680 | if (!ioend) { |
| 1681 | ASSERT(offset + size <= i_size_read(inode)); |
| 1682 | return; |
| 1683 | } |
| 1684 | |
| 1685 | __xfs_end_io_direct_write(inode, ioend, offset, size); |
| 1686 | } |
| 1687 | |
| 1688 | static inline ssize_t |
| 1689 | xfs_vm_do_dio( |
| 1690 | struct inode *inode, |
| 1691 | struct kiocb *iocb, |
| 1692 | struct iov_iter *iter, |
| 1693 | loff_t offset, |
| 1694 | void (*endio)(struct kiocb *iocb, |
| 1695 | loff_t offset, |
| 1696 | ssize_t size, |
| 1697 | void *private), |
| 1698 | int flags) |
| 1699 | { |
| 1700 | struct block_device *bdev; |
| 1701 | |
| 1702 | if (IS_DAX(inode)) |
| 1703 | return dax_do_io(iocb, inode, iter, offset, |
| 1704 | xfs_get_blocks_direct, endio, 0); |
| 1705 | |
| 1706 | bdev = xfs_find_bdev_for_inode(inode); |
| 1707 | return __blockdev_direct_IO(iocb, inode, bdev, iter, offset, |
| 1708 | xfs_get_blocks_direct, endio, NULL, flags); |
| 1709 | } |
| 1710 | |
| 1711 | STATIC ssize_t |
| 1712 | xfs_vm_direct_IO( |
| 1713 | struct kiocb *iocb, |
| 1714 | struct iov_iter *iter, |
| 1715 | loff_t offset) |
| 1716 | { |
| 1717 | struct inode *inode = iocb->ki_filp->f_mapping->host; |
| 1718 | |
| 1719 | if (iov_iter_rw(iter) == WRITE) |
| 1720 | return xfs_vm_do_dio(inode, iocb, iter, offset, |
| 1721 | xfs_end_io_direct_write, DIO_ASYNC_EXTEND); |
| 1722 | return xfs_vm_do_dio(inode, iocb, iter, offset, NULL, 0); |
| 1723 | } |
| 1724 | |
| 1725 | /* |
| 1726 | * Punch out the delalloc blocks we have already allocated. |
| 1727 | * |
| 1728 | * Don't bother with xfs_setattr given that nothing can have made it to disk yet |
| 1729 | * as the page is still locked at this point. |
| 1730 | */ |
| 1731 | STATIC void |
| 1732 | xfs_vm_kill_delalloc_range( |
| 1733 | struct inode *inode, |
| 1734 | loff_t start, |
| 1735 | loff_t end) |
| 1736 | { |
| 1737 | struct xfs_inode *ip = XFS_I(inode); |
| 1738 | xfs_fileoff_t start_fsb; |
| 1739 | xfs_fileoff_t end_fsb; |
| 1740 | int error; |
| 1741 | |
| 1742 | start_fsb = XFS_B_TO_FSB(ip->i_mount, start); |
| 1743 | end_fsb = XFS_B_TO_FSB(ip->i_mount, end); |
| 1744 | if (end_fsb <= start_fsb) |
| 1745 | return; |
| 1746 | |
| 1747 | xfs_ilock(ip, XFS_ILOCK_EXCL); |
| 1748 | error = xfs_bmap_punch_delalloc_range(ip, start_fsb, |
| 1749 | end_fsb - start_fsb); |
| 1750 | if (error) { |
| 1751 | /* something screwed, just bail */ |
| 1752 | if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) { |
| 1753 | xfs_alert(ip->i_mount, |
| 1754 | "xfs_vm_write_failed: unable to clean up ino %lld", |
| 1755 | ip->i_ino); |
| 1756 | } |
| 1757 | } |
| 1758 | xfs_iunlock(ip, XFS_ILOCK_EXCL); |
| 1759 | } |
| 1760 | |
| 1761 | STATIC void |
| 1762 | xfs_vm_write_failed( |
| 1763 | struct inode *inode, |
| 1764 | struct page *page, |
| 1765 | loff_t pos, |
| 1766 | unsigned len) |
| 1767 | { |
| 1768 | loff_t block_offset; |
| 1769 | loff_t block_start; |
| 1770 | loff_t block_end; |
| 1771 | loff_t from = pos & (PAGE_CACHE_SIZE - 1); |
| 1772 | loff_t to = from + len; |
| 1773 | struct buffer_head *bh, *head; |
| 1774 | |
| 1775 | /* |
| 1776 | * The request pos offset might be 32 or 64 bit, this is all fine |
| 1777 | * on 64-bit platform. However, for 64-bit pos request on 32-bit |
| 1778 | * platform, the high 32-bit will be masked off if we evaluate the |
| 1779 | * block_offset via (pos & PAGE_MASK) because the PAGE_MASK is |
| 1780 | * 0xfffff000 as an unsigned long, hence the result is incorrect |
| 1781 | * which could cause the following ASSERT failed in most cases. |
| 1782 | * In order to avoid this, we can evaluate the block_offset of the |
| 1783 | * start of the page by using shifts rather than masks the mismatch |
| 1784 | * problem. |
| 1785 | */ |
| 1786 | block_offset = (pos >> PAGE_CACHE_SHIFT) << PAGE_CACHE_SHIFT; |
| 1787 | |
| 1788 | ASSERT(block_offset + from == pos); |
| 1789 | |
| 1790 | head = page_buffers(page); |
| 1791 | block_start = 0; |
| 1792 | for (bh = head; bh != head || !block_start; |
| 1793 | bh = bh->b_this_page, block_start = block_end, |
| 1794 | block_offset += bh->b_size) { |
| 1795 | block_end = block_start + bh->b_size; |
| 1796 | |
| 1797 | /* skip buffers before the write */ |
| 1798 | if (block_end <= from) |
| 1799 | continue; |
| 1800 | |
| 1801 | /* if the buffer is after the write, we're done */ |
| 1802 | if (block_start >= to) |
| 1803 | break; |
| 1804 | |
| 1805 | if (!buffer_delay(bh)) |
| 1806 | continue; |
| 1807 | |
| 1808 | if (!buffer_new(bh) && block_offset < i_size_read(inode)) |
| 1809 | continue; |
| 1810 | |
| 1811 | xfs_vm_kill_delalloc_range(inode, block_offset, |
| 1812 | block_offset + bh->b_size); |
| 1813 | |
| 1814 | /* |
| 1815 | * This buffer does not contain data anymore. make sure anyone |
| 1816 | * who finds it knows that for certain. |
| 1817 | */ |
| 1818 | clear_buffer_delay(bh); |
| 1819 | clear_buffer_uptodate(bh); |
| 1820 | clear_buffer_mapped(bh); |
| 1821 | clear_buffer_new(bh); |
| 1822 | clear_buffer_dirty(bh); |
| 1823 | } |
| 1824 | |
| 1825 | } |
| 1826 | |
| 1827 | /* |
| 1828 | * This used to call block_write_begin(), but it unlocks and releases the page |
| 1829 | * on error, and we need that page to be able to punch stale delalloc blocks out |
| 1830 | * on failure. hence we copy-n-waste it here and call xfs_vm_write_failed() at |
| 1831 | * the appropriate point. |
| 1832 | */ |
| 1833 | STATIC int |
| 1834 | xfs_vm_write_begin( |
| 1835 | struct file *file, |
| 1836 | struct address_space *mapping, |
| 1837 | loff_t pos, |
| 1838 | unsigned len, |
| 1839 | unsigned flags, |
| 1840 | struct page **pagep, |
| 1841 | void **fsdata) |
| 1842 | { |
| 1843 | pgoff_t index = pos >> PAGE_CACHE_SHIFT; |
| 1844 | struct page *page; |
| 1845 | int status; |
| 1846 | |
| 1847 | ASSERT(len <= PAGE_CACHE_SIZE); |
| 1848 | |
| 1849 | page = grab_cache_page_write_begin(mapping, index, flags); |
| 1850 | if (!page) |
| 1851 | return -ENOMEM; |
| 1852 | |
| 1853 | status = __block_write_begin(page, pos, len, xfs_get_blocks); |
| 1854 | if (unlikely(status)) { |
| 1855 | struct inode *inode = mapping->host; |
| 1856 | size_t isize = i_size_read(inode); |
| 1857 | |
| 1858 | xfs_vm_write_failed(inode, page, pos, len); |
| 1859 | unlock_page(page); |
| 1860 | |
| 1861 | /* |
| 1862 | * If the write is beyond EOF, we only want to kill blocks |
| 1863 | * allocated in this write, not blocks that were previously |
| 1864 | * written successfully. |
| 1865 | */ |
| 1866 | if (pos + len > isize) { |
| 1867 | ssize_t start = max_t(ssize_t, pos, isize); |
| 1868 | |
| 1869 | truncate_pagecache_range(inode, start, pos + len); |
| 1870 | } |
| 1871 | |
| 1872 | page_cache_release(page); |
| 1873 | page = NULL; |
| 1874 | } |
| 1875 | |
| 1876 | *pagep = page; |
| 1877 | return status; |
| 1878 | } |
| 1879 | |
| 1880 | /* |
| 1881 | * On failure, we only need to kill delalloc blocks beyond EOF in the range of |
| 1882 | * this specific write because they will never be written. Previous writes |
| 1883 | * beyond EOF where block allocation succeeded do not need to be trashed, so |
| 1884 | * only new blocks from this write should be trashed. For blocks within |
| 1885 | * EOF, generic_write_end() zeros them so they are safe to leave alone and be |
| 1886 | * written with all the other valid data. |
| 1887 | */ |
| 1888 | STATIC int |
| 1889 | xfs_vm_write_end( |
| 1890 | struct file *file, |
| 1891 | struct address_space *mapping, |
| 1892 | loff_t pos, |
| 1893 | unsigned len, |
| 1894 | unsigned copied, |
| 1895 | struct page *page, |
| 1896 | void *fsdata) |
| 1897 | { |
| 1898 | int ret; |
| 1899 | |
| 1900 | ASSERT(len <= PAGE_CACHE_SIZE); |
| 1901 | |
| 1902 | ret = generic_write_end(file, mapping, pos, len, copied, page, fsdata); |
| 1903 | if (unlikely(ret < len)) { |
| 1904 | struct inode *inode = mapping->host; |
| 1905 | size_t isize = i_size_read(inode); |
| 1906 | loff_t to = pos + len; |
| 1907 | |
| 1908 | if (to > isize) { |
| 1909 | /* only kill blocks in this write beyond EOF */ |
| 1910 | if (pos > isize) |
| 1911 | isize = pos; |
| 1912 | xfs_vm_kill_delalloc_range(inode, isize, to); |
| 1913 | truncate_pagecache_range(inode, isize, to); |
| 1914 | } |
| 1915 | } |
| 1916 | return ret; |
| 1917 | } |
| 1918 | |
| 1919 | STATIC sector_t |
| 1920 | xfs_vm_bmap( |
| 1921 | struct address_space *mapping, |
| 1922 | sector_t block) |
| 1923 | { |
| 1924 | struct inode *inode = (struct inode *)mapping->host; |
| 1925 | struct xfs_inode *ip = XFS_I(inode); |
| 1926 | |
| 1927 | trace_xfs_vm_bmap(XFS_I(inode)); |
| 1928 | xfs_ilock(ip, XFS_IOLOCK_SHARED); |
| 1929 | filemap_write_and_wait(mapping); |
| 1930 | xfs_iunlock(ip, XFS_IOLOCK_SHARED); |
| 1931 | return generic_block_bmap(mapping, block, xfs_get_blocks); |
| 1932 | } |
| 1933 | |
| 1934 | STATIC int |
| 1935 | xfs_vm_readpage( |
| 1936 | struct file *unused, |
| 1937 | struct page *page) |
| 1938 | { |
| 1939 | return mpage_readpage(page, xfs_get_blocks); |
| 1940 | } |
| 1941 | |
| 1942 | STATIC int |
| 1943 | xfs_vm_readpages( |
| 1944 | struct file *unused, |
| 1945 | struct address_space *mapping, |
| 1946 | struct list_head *pages, |
| 1947 | unsigned nr_pages) |
| 1948 | { |
| 1949 | return mpage_readpages(mapping, pages, nr_pages, xfs_get_blocks); |
| 1950 | } |
| 1951 | |
| 1952 | /* |
| 1953 | * This is basically a copy of __set_page_dirty_buffers() with one |
| 1954 | * small tweak: buffers beyond EOF do not get marked dirty. If we mark them |
| 1955 | * dirty, we'll never be able to clean them because we don't write buffers |
| 1956 | * beyond EOF, and that means we can't invalidate pages that span EOF |
| 1957 | * that have been marked dirty. Further, the dirty state can leak into |
| 1958 | * the file interior if the file is extended, resulting in all sorts of |
| 1959 | * bad things happening as the state does not match the underlying data. |
| 1960 | * |
| 1961 | * XXX: this really indicates that bufferheads in XFS need to die. Warts like |
| 1962 | * this only exist because of bufferheads and how the generic code manages them. |
| 1963 | */ |
| 1964 | STATIC int |
| 1965 | xfs_vm_set_page_dirty( |
| 1966 | struct page *page) |
| 1967 | { |
| 1968 | struct address_space *mapping = page->mapping; |
| 1969 | struct inode *inode = mapping->host; |
| 1970 | loff_t end_offset; |
| 1971 | loff_t offset; |
| 1972 | int newly_dirty; |
| 1973 | struct mem_cgroup *memcg; |
| 1974 | |
| 1975 | if (unlikely(!mapping)) |
| 1976 | return !TestSetPageDirty(page); |
| 1977 | |
| 1978 | end_offset = i_size_read(inode); |
| 1979 | offset = page_offset(page); |
| 1980 | |
| 1981 | spin_lock(&mapping->private_lock); |
| 1982 | if (page_has_buffers(page)) { |
| 1983 | struct buffer_head *head = page_buffers(page); |
| 1984 | struct buffer_head *bh = head; |
| 1985 | |
| 1986 | do { |
| 1987 | if (offset < end_offset) |
| 1988 | set_buffer_dirty(bh); |
| 1989 | bh = bh->b_this_page; |
| 1990 | offset += i_blocksize(inode); |
| 1991 | } while (bh != head); |
| 1992 | } |
| 1993 | /* |
| 1994 | * Use mem_group_begin_page_stat() to keep PageDirty synchronized with |
| 1995 | * per-memcg dirty page counters. |
| 1996 | */ |
| 1997 | memcg = mem_cgroup_begin_page_stat(page); |
| 1998 | newly_dirty = !TestSetPageDirty(page); |
| 1999 | spin_unlock(&mapping->private_lock); |
| 2000 | |
| 2001 | if (newly_dirty) { |
| 2002 | /* sigh - __set_page_dirty() is static, so copy it here, too */ |
| 2003 | unsigned long flags; |
| 2004 | |
| 2005 | spin_lock_irqsave(&mapping->tree_lock, flags); |
| 2006 | if (page->mapping) { /* Race with truncate? */ |
| 2007 | WARN_ON_ONCE(!PageUptodate(page)); |
| 2008 | account_page_dirtied(page, mapping, memcg); |
| 2009 | radix_tree_tag_set(&mapping->page_tree, |
| 2010 | page_index(page), PAGECACHE_TAG_DIRTY); |
| 2011 | } |
| 2012 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
| 2013 | } |
| 2014 | mem_cgroup_end_page_stat(memcg); |
| 2015 | if (newly_dirty) |
| 2016 | __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); |
| 2017 | return newly_dirty; |
| 2018 | } |
| 2019 | |
| 2020 | const struct address_space_operations xfs_address_space_operations = { |
| 2021 | .readpage = xfs_vm_readpage, |
| 2022 | .readpages = xfs_vm_readpages, |
| 2023 | .writepage = xfs_vm_writepage, |
| 2024 | .writepages = xfs_vm_writepages, |
| 2025 | .set_page_dirty = xfs_vm_set_page_dirty, |
| 2026 | .releasepage = xfs_vm_releasepage, |
| 2027 | .invalidatepage = xfs_vm_invalidatepage, |
| 2028 | .write_begin = xfs_vm_write_begin, |
| 2029 | .write_end = xfs_vm_write_end, |
| 2030 | .bmap = xfs_vm_bmap, |
| 2031 | .direct_IO = xfs_vm_direct_IO, |
| 2032 | .migratepage = buffer_migrate_page, |
| 2033 | .is_partially_uptodate = block_is_partially_uptodate, |
| 2034 | .error_remove_page = generic_error_remove_page, |
| 2035 | }; |