| 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_log.h" |
| 20 | #include "xfs_sb.h" |
| 21 | #include "xfs_ag.h" |
| 22 | #include "xfs_trans.h" |
| 23 | #include "xfs_mount.h" |
| 24 | #include "xfs_bmap_btree.h" |
| 25 | #include "xfs_dinode.h" |
| 26 | #include "xfs_inode.h" |
| 27 | #include "xfs_inode_item.h" |
| 28 | #include "xfs_alloc.h" |
| 29 | #include "xfs_error.h" |
| 30 | #include "xfs_iomap.h" |
| 31 | #include "xfs_vnodeops.h" |
| 32 | #include "xfs_trace.h" |
| 33 | #include "xfs_bmap.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 | if (ioend->io_iocb) { |
| 89 | if (ioend->io_isasync) { |
| 90 | aio_complete(ioend->io_iocb, ioend->io_error ? |
| 91 | ioend->io_error : ioend->io_result, 0); |
| 92 | } |
| 93 | inode_dio_done(ioend->io_inode); |
| 94 | } |
| 95 | |
| 96 | mempool_free(ioend, xfs_ioend_pool); |
| 97 | } |
| 98 | |
| 99 | /* |
| 100 | * Fast and loose check if this write could update the on-disk inode size. |
| 101 | */ |
| 102 | static inline bool xfs_ioend_is_append(struct xfs_ioend *ioend) |
| 103 | { |
| 104 | return ioend->io_offset + ioend->io_size > |
| 105 | XFS_I(ioend->io_inode)->i_d.di_size; |
| 106 | } |
| 107 | |
| 108 | STATIC int |
| 109 | xfs_setfilesize_trans_alloc( |
| 110 | struct xfs_ioend *ioend) |
| 111 | { |
| 112 | struct xfs_mount *mp = XFS_I(ioend->io_inode)->i_mount; |
| 113 | struct xfs_trans *tp; |
| 114 | int error; |
| 115 | |
| 116 | tp = xfs_trans_alloc(mp, XFS_TRANS_FSYNC_TS); |
| 117 | |
| 118 | error = xfs_trans_reserve(tp, 0, XFS_FSYNC_TS_LOG_RES(mp), 0, 0, 0); |
| 119 | if (error) { |
| 120 | xfs_trans_cancel(tp, 0); |
| 121 | return error; |
| 122 | } |
| 123 | |
| 124 | ioend->io_append_trans = tp; |
| 125 | |
| 126 | /* |
| 127 | * We hand off the transaction to the completion thread now, so |
| 128 | * clear the flag here. |
| 129 | */ |
| 130 | current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS); |
| 131 | return 0; |
| 132 | } |
| 133 | |
| 134 | /* |
| 135 | * Update on-disk file size now that data has been written to disk. |
| 136 | */ |
| 137 | STATIC int |
| 138 | xfs_setfilesize( |
| 139 | struct xfs_ioend *ioend) |
| 140 | { |
| 141 | struct xfs_inode *ip = XFS_I(ioend->io_inode); |
| 142 | struct xfs_trans *tp = ioend->io_append_trans; |
| 143 | xfs_fsize_t isize; |
| 144 | |
| 145 | /* |
| 146 | * The transaction was allocated in the I/O submission thread, |
| 147 | * thus we need to mark ourselves as beeing in a transaction |
| 148 | * manually. |
| 149 | */ |
| 150 | current_set_flags_nested(&tp->t_pflags, PF_FSTRANS); |
| 151 | |
| 152 | xfs_ilock(ip, XFS_ILOCK_EXCL); |
| 153 | isize = xfs_new_eof(ip, ioend->io_offset + ioend->io_size); |
| 154 | if (!isize) { |
| 155 | xfs_iunlock(ip, XFS_ILOCK_EXCL); |
| 156 | xfs_trans_cancel(tp, 0); |
| 157 | return 0; |
| 158 | } |
| 159 | |
| 160 | trace_xfs_setfilesize(ip, ioend->io_offset, ioend->io_size); |
| 161 | |
| 162 | ip->i_d.di_size = isize; |
| 163 | xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL); |
| 164 | xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE); |
| 165 | |
| 166 | return xfs_trans_commit(tp, 0); |
| 167 | } |
| 168 | |
| 169 | /* |
| 170 | * Schedule IO completion handling on the final put of an ioend. |
| 171 | * |
| 172 | * If there is no work to do we might as well call it a day and free the |
| 173 | * ioend right now. |
| 174 | */ |
| 175 | STATIC void |
| 176 | xfs_finish_ioend( |
| 177 | struct xfs_ioend *ioend) |
| 178 | { |
| 179 | if (atomic_dec_and_test(&ioend->io_remaining)) { |
| 180 | struct xfs_mount *mp = XFS_I(ioend->io_inode)->i_mount; |
| 181 | |
| 182 | if (ioend->io_type == IO_UNWRITTEN) |
| 183 | queue_work(mp->m_unwritten_workqueue, &ioend->io_work); |
| 184 | else if (ioend->io_append_trans) |
| 185 | queue_work(mp->m_data_workqueue, &ioend->io_work); |
| 186 | else |
| 187 | xfs_destroy_ioend(ioend); |
| 188 | } |
| 189 | } |
| 190 | |
| 191 | /* |
| 192 | * IO write completion. |
| 193 | */ |
| 194 | STATIC void |
| 195 | xfs_end_io( |
| 196 | struct work_struct *work) |
| 197 | { |
| 198 | xfs_ioend_t *ioend = container_of(work, xfs_ioend_t, io_work); |
| 199 | struct xfs_inode *ip = XFS_I(ioend->io_inode); |
| 200 | int error = 0; |
| 201 | |
| 202 | if (XFS_FORCED_SHUTDOWN(ip->i_mount)) { |
| 203 | ioend->io_error = -EIO; |
| 204 | goto done; |
| 205 | } |
| 206 | if (ioend->io_error) |
| 207 | goto done; |
| 208 | |
| 209 | /* |
| 210 | * For unwritten extents we need to issue transactions to convert a |
| 211 | * range to normal written extens after the data I/O has finished. |
| 212 | */ |
| 213 | if (ioend->io_type == IO_UNWRITTEN) { |
| 214 | /* |
| 215 | * For buffered I/O we never preallocate a transaction when |
| 216 | * doing the unwritten extent conversion, but for direct I/O |
| 217 | * we do not know if we are converting an unwritten extent |
| 218 | * or not at the point where we preallocate the transaction. |
| 219 | */ |
| 220 | if (ioend->io_append_trans) { |
| 221 | ASSERT(ioend->io_isdirect); |
| 222 | |
| 223 | current_set_flags_nested( |
| 224 | &ioend->io_append_trans->t_pflags, PF_FSTRANS); |
| 225 | xfs_trans_cancel(ioend->io_append_trans, 0); |
| 226 | } |
| 227 | |
| 228 | error = xfs_iomap_write_unwritten(ip, ioend->io_offset, |
| 229 | ioend->io_size); |
| 230 | if (error) { |
| 231 | ioend->io_error = -error; |
| 232 | goto done; |
| 233 | } |
| 234 | } else if (ioend->io_append_trans) { |
| 235 | error = xfs_setfilesize(ioend); |
| 236 | if (error) |
| 237 | ioend->io_error = -error; |
| 238 | } else { |
| 239 | ASSERT(!xfs_ioend_is_append(ioend)); |
| 240 | } |
| 241 | |
| 242 | done: |
| 243 | xfs_destroy_ioend(ioend); |
| 244 | } |
| 245 | |
| 246 | /* |
| 247 | * Call IO completion handling in caller context on the final put of an ioend. |
| 248 | */ |
| 249 | STATIC void |
| 250 | xfs_finish_ioend_sync( |
| 251 | struct xfs_ioend *ioend) |
| 252 | { |
| 253 | if (atomic_dec_and_test(&ioend->io_remaining)) |
| 254 | xfs_end_io(&ioend->io_work); |
| 255 | } |
| 256 | |
| 257 | /* |
| 258 | * Allocate and initialise an IO completion structure. |
| 259 | * We need to track unwritten extent write completion here initially. |
| 260 | * We'll need to extend this for updating the ondisk inode size later |
| 261 | * (vs. incore size). |
| 262 | */ |
| 263 | STATIC xfs_ioend_t * |
| 264 | xfs_alloc_ioend( |
| 265 | struct inode *inode, |
| 266 | unsigned int type) |
| 267 | { |
| 268 | xfs_ioend_t *ioend; |
| 269 | |
| 270 | ioend = mempool_alloc(xfs_ioend_pool, GFP_NOFS); |
| 271 | |
| 272 | /* |
| 273 | * Set the count to 1 initially, which will prevent an I/O |
| 274 | * completion callback from happening before we have started |
| 275 | * all the I/O from calling the completion routine too early. |
| 276 | */ |
| 277 | atomic_set(&ioend->io_remaining, 1); |
| 278 | ioend->io_isasync = 0; |
| 279 | ioend->io_isdirect = 0; |
| 280 | ioend->io_error = 0; |
| 281 | ioend->io_list = NULL; |
| 282 | ioend->io_type = type; |
| 283 | ioend->io_inode = inode; |
| 284 | ioend->io_buffer_head = NULL; |
| 285 | ioend->io_buffer_tail = NULL; |
| 286 | ioend->io_offset = 0; |
| 287 | ioend->io_size = 0; |
| 288 | ioend->io_iocb = NULL; |
| 289 | ioend->io_result = 0; |
| 290 | ioend->io_append_trans = NULL; |
| 291 | |
| 292 | INIT_WORK(&ioend->io_work, xfs_end_io); |
| 293 | return ioend; |
| 294 | } |
| 295 | |
| 296 | STATIC int |
| 297 | xfs_map_blocks( |
| 298 | struct inode *inode, |
| 299 | loff_t offset, |
| 300 | struct xfs_bmbt_irec *imap, |
| 301 | int type, |
| 302 | int nonblocking) |
| 303 | { |
| 304 | struct xfs_inode *ip = XFS_I(inode); |
| 305 | struct xfs_mount *mp = ip->i_mount; |
| 306 | ssize_t count = 1 << inode->i_blkbits; |
| 307 | xfs_fileoff_t offset_fsb, end_fsb; |
| 308 | int error = 0; |
| 309 | int bmapi_flags = XFS_BMAPI_ENTIRE; |
| 310 | int nimaps = 1; |
| 311 | |
| 312 | if (XFS_FORCED_SHUTDOWN(mp)) |
| 313 | return -XFS_ERROR(EIO); |
| 314 | |
| 315 | if (type == IO_UNWRITTEN) |
| 316 | bmapi_flags |= XFS_BMAPI_IGSTATE; |
| 317 | |
| 318 | if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) { |
| 319 | if (nonblocking) |
| 320 | return -XFS_ERROR(EAGAIN); |
| 321 | xfs_ilock(ip, XFS_ILOCK_SHARED); |
| 322 | } |
| 323 | |
| 324 | ASSERT(ip->i_d.di_format != XFS_DINODE_FMT_BTREE || |
| 325 | (ip->i_df.if_flags & XFS_IFEXTENTS)); |
| 326 | ASSERT(offset <= mp->m_maxioffset); |
| 327 | |
| 328 | if (offset + count > mp->m_maxioffset) |
| 329 | count = mp->m_maxioffset - offset; |
| 330 | end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + count); |
| 331 | offset_fsb = XFS_B_TO_FSBT(mp, offset); |
| 332 | error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb, |
| 333 | imap, &nimaps, bmapi_flags); |
| 334 | xfs_iunlock(ip, XFS_ILOCK_SHARED); |
| 335 | |
| 336 | if (error) |
| 337 | return -XFS_ERROR(error); |
| 338 | |
| 339 | if (type == IO_DELALLOC && |
| 340 | (!nimaps || isnullstartblock(imap->br_startblock))) { |
| 341 | error = xfs_iomap_write_allocate(ip, offset, count, imap); |
| 342 | if (!error) |
| 343 | trace_xfs_map_blocks_alloc(ip, offset, count, type, imap); |
| 344 | return -XFS_ERROR(error); |
| 345 | } |
| 346 | |
| 347 | #ifdef DEBUG |
| 348 | if (type == IO_UNWRITTEN) { |
| 349 | ASSERT(nimaps); |
| 350 | ASSERT(imap->br_startblock != HOLESTARTBLOCK); |
| 351 | ASSERT(imap->br_startblock != DELAYSTARTBLOCK); |
| 352 | } |
| 353 | #endif |
| 354 | if (nimaps) |
| 355 | trace_xfs_map_blocks_found(ip, offset, count, type, imap); |
| 356 | return 0; |
| 357 | } |
| 358 | |
| 359 | STATIC int |
| 360 | xfs_imap_valid( |
| 361 | struct inode *inode, |
| 362 | struct xfs_bmbt_irec *imap, |
| 363 | xfs_off_t offset) |
| 364 | { |
| 365 | offset >>= inode->i_blkbits; |
| 366 | |
| 367 | return offset >= imap->br_startoff && |
| 368 | offset < imap->br_startoff + imap->br_blockcount; |
| 369 | } |
| 370 | |
| 371 | /* |
| 372 | * BIO completion handler for buffered IO. |
| 373 | */ |
| 374 | STATIC void |
| 375 | xfs_end_bio( |
| 376 | struct bio *bio, |
| 377 | int error) |
| 378 | { |
| 379 | xfs_ioend_t *ioend = bio->bi_private; |
| 380 | |
| 381 | ASSERT(atomic_read(&bio->bi_cnt) >= 1); |
| 382 | ioend->io_error = test_bit(BIO_UPTODATE, &bio->bi_flags) ? 0 : error; |
| 383 | |
| 384 | /* Toss bio and pass work off to an xfsdatad thread */ |
| 385 | bio->bi_private = NULL; |
| 386 | bio->bi_end_io = NULL; |
| 387 | bio_put(bio); |
| 388 | |
| 389 | xfs_finish_ioend(ioend); |
| 390 | } |
| 391 | |
| 392 | STATIC void |
| 393 | xfs_submit_ioend_bio( |
| 394 | struct writeback_control *wbc, |
| 395 | xfs_ioend_t *ioend, |
| 396 | struct bio *bio) |
| 397 | { |
| 398 | atomic_inc(&ioend->io_remaining); |
| 399 | bio->bi_private = ioend; |
| 400 | bio->bi_end_io = xfs_end_bio; |
| 401 | submit_bio(wbc->sync_mode == WB_SYNC_ALL ? WRITE_SYNC : WRITE, bio); |
| 402 | } |
| 403 | |
| 404 | STATIC struct bio * |
| 405 | xfs_alloc_ioend_bio( |
| 406 | struct buffer_head *bh) |
| 407 | { |
| 408 | int nvecs = bio_get_nr_vecs(bh->b_bdev); |
| 409 | struct bio *bio = bio_alloc(GFP_NOIO, nvecs); |
| 410 | |
| 411 | ASSERT(bio->bi_private == NULL); |
| 412 | bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9); |
| 413 | bio->bi_bdev = bh->b_bdev; |
| 414 | return bio; |
| 415 | } |
| 416 | |
| 417 | STATIC void |
| 418 | xfs_start_buffer_writeback( |
| 419 | struct buffer_head *bh) |
| 420 | { |
| 421 | ASSERT(buffer_mapped(bh)); |
| 422 | ASSERT(buffer_locked(bh)); |
| 423 | ASSERT(!buffer_delay(bh)); |
| 424 | ASSERT(!buffer_unwritten(bh)); |
| 425 | |
| 426 | mark_buffer_async_write(bh); |
| 427 | set_buffer_uptodate(bh); |
| 428 | clear_buffer_dirty(bh); |
| 429 | } |
| 430 | |
| 431 | STATIC void |
| 432 | xfs_start_page_writeback( |
| 433 | struct page *page, |
| 434 | int clear_dirty, |
| 435 | int buffers) |
| 436 | { |
| 437 | ASSERT(PageLocked(page)); |
| 438 | ASSERT(!PageWriteback(page)); |
| 439 | if (clear_dirty) |
| 440 | clear_page_dirty_for_io(page); |
| 441 | set_page_writeback(page); |
| 442 | unlock_page(page); |
| 443 | /* If no buffers on the page are to be written, finish it here */ |
| 444 | if (!buffers) |
| 445 | end_page_writeback(page); |
| 446 | } |
| 447 | |
| 448 | static inline int bio_add_buffer(struct bio *bio, struct buffer_head *bh) |
| 449 | { |
| 450 | return bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh)); |
| 451 | } |
| 452 | |
| 453 | /* |
| 454 | * Submit all of the bios for all of the ioends we have saved up, covering the |
| 455 | * initial writepage page and also any probed pages. |
| 456 | * |
| 457 | * Because we may have multiple ioends spanning a page, we need to start |
| 458 | * writeback on all the buffers before we submit them for I/O. If we mark the |
| 459 | * buffers as we got, then we can end up with a page that only has buffers |
| 460 | * marked async write and I/O complete on can occur before we mark the other |
| 461 | * buffers async write. |
| 462 | * |
| 463 | * The end result of this is that we trip a bug in end_page_writeback() because |
| 464 | * we call it twice for the one page as the code in end_buffer_async_write() |
| 465 | * assumes that all buffers on the page are started at the same time. |
| 466 | * |
| 467 | * The fix is two passes across the ioend list - one to start writeback on the |
| 468 | * buffer_heads, and then submit them for I/O on the second pass. |
| 469 | */ |
| 470 | STATIC void |
| 471 | xfs_submit_ioend( |
| 472 | struct writeback_control *wbc, |
| 473 | xfs_ioend_t *ioend) |
| 474 | { |
| 475 | xfs_ioend_t *head = ioend; |
| 476 | xfs_ioend_t *next; |
| 477 | struct buffer_head *bh; |
| 478 | struct bio *bio; |
| 479 | sector_t lastblock = 0; |
| 480 | |
| 481 | /* Pass 1 - start writeback */ |
| 482 | do { |
| 483 | next = ioend->io_list; |
| 484 | for (bh = ioend->io_buffer_head; bh; bh = bh->b_private) |
| 485 | xfs_start_buffer_writeback(bh); |
| 486 | } while ((ioend = next) != NULL); |
| 487 | |
| 488 | /* Pass 2 - submit I/O */ |
| 489 | ioend = head; |
| 490 | do { |
| 491 | next = ioend->io_list; |
| 492 | bio = NULL; |
| 493 | |
| 494 | for (bh = ioend->io_buffer_head; bh; bh = bh->b_private) { |
| 495 | |
| 496 | if (!bio) { |
| 497 | retry: |
| 498 | bio = xfs_alloc_ioend_bio(bh); |
| 499 | } else if (bh->b_blocknr != lastblock + 1) { |
| 500 | xfs_submit_ioend_bio(wbc, ioend, bio); |
| 501 | goto retry; |
| 502 | } |
| 503 | |
| 504 | if (bio_add_buffer(bio, bh) != bh->b_size) { |
| 505 | xfs_submit_ioend_bio(wbc, ioend, bio); |
| 506 | goto retry; |
| 507 | } |
| 508 | |
| 509 | lastblock = bh->b_blocknr; |
| 510 | } |
| 511 | if (bio) |
| 512 | xfs_submit_ioend_bio(wbc, ioend, bio); |
| 513 | xfs_finish_ioend(ioend); |
| 514 | } while ((ioend = next) != NULL); |
| 515 | } |
| 516 | |
| 517 | /* |
| 518 | * Cancel submission of all buffer_heads so far in this endio. |
| 519 | * Toss the endio too. Only ever called for the initial page |
| 520 | * in a writepage request, so only ever one page. |
| 521 | */ |
| 522 | STATIC void |
| 523 | xfs_cancel_ioend( |
| 524 | xfs_ioend_t *ioend) |
| 525 | { |
| 526 | xfs_ioend_t *next; |
| 527 | struct buffer_head *bh, *next_bh; |
| 528 | |
| 529 | do { |
| 530 | next = ioend->io_list; |
| 531 | bh = ioend->io_buffer_head; |
| 532 | do { |
| 533 | next_bh = bh->b_private; |
| 534 | clear_buffer_async_write(bh); |
| 535 | unlock_buffer(bh); |
| 536 | } while ((bh = next_bh) != NULL); |
| 537 | |
| 538 | mempool_free(ioend, xfs_ioend_pool); |
| 539 | } while ((ioend = next) != NULL); |
| 540 | } |
| 541 | |
| 542 | /* |
| 543 | * Test to see if we've been building up a completion structure for |
| 544 | * earlier buffers -- if so, we try to append to this ioend if we |
| 545 | * can, otherwise we finish off any current ioend and start another. |
| 546 | * Return true if we've finished the given ioend. |
| 547 | */ |
| 548 | STATIC void |
| 549 | xfs_add_to_ioend( |
| 550 | struct inode *inode, |
| 551 | struct buffer_head *bh, |
| 552 | xfs_off_t offset, |
| 553 | unsigned int type, |
| 554 | xfs_ioend_t **result, |
| 555 | int need_ioend) |
| 556 | { |
| 557 | xfs_ioend_t *ioend = *result; |
| 558 | |
| 559 | if (!ioend || need_ioend || type != ioend->io_type) { |
| 560 | xfs_ioend_t *previous = *result; |
| 561 | |
| 562 | ioend = xfs_alloc_ioend(inode, type); |
| 563 | ioend->io_offset = offset; |
| 564 | ioend->io_buffer_head = bh; |
| 565 | ioend->io_buffer_tail = bh; |
| 566 | if (previous) |
| 567 | previous->io_list = ioend; |
| 568 | *result = ioend; |
| 569 | } else { |
| 570 | ioend->io_buffer_tail->b_private = bh; |
| 571 | ioend->io_buffer_tail = bh; |
| 572 | } |
| 573 | |
| 574 | bh->b_private = NULL; |
| 575 | ioend->io_size += bh->b_size; |
| 576 | } |
| 577 | |
| 578 | STATIC void |
| 579 | xfs_map_buffer( |
| 580 | struct inode *inode, |
| 581 | struct buffer_head *bh, |
| 582 | struct xfs_bmbt_irec *imap, |
| 583 | xfs_off_t offset) |
| 584 | { |
| 585 | sector_t bn; |
| 586 | struct xfs_mount *m = XFS_I(inode)->i_mount; |
| 587 | xfs_off_t iomap_offset = XFS_FSB_TO_B(m, imap->br_startoff); |
| 588 | xfs_daddr_t iomap_bn = xfs_fsb_to_db(XFS_I(inode), imap->br_startblock); |
| 589 | |
| 590 | ASSERT(imap->br_startblock != HOLESTARTBLOCK); |
| 591 | ASSERT(imap->br_startblock != DELAYSTARTBLOCK); |
| 592 | |
| 593 | bn = (iomap_bn >> (inode->i_blkbits - BBSHIFT)) + |
| 594 | ((offset - iomap_offset) >> inode->i_blkbits); |
| 595 | |
| 596 | ASSERT(bn || XFS_IS_REALTIME_INODE(XFS_I(inode))); |
| 597 | |
| 598 | bh->b_blocknr = bn; |
| 599 | set_buffer_mapped(bh); |
| 600 | } |
| 601 | |
| 602 | STATIC void |
| 603 | xfs_map_at_offset( |
| 604 | struct inode *inode, |
| 605 | struct buffer_head *bh, |
| 606 | struct xfs_bmbt_irec *imap, |
| 607 | xfs_off_t offset) |
| 608 | { |
| 609 | ASSERT(imap->br_startblock != HOLESTARTBLOCK); |
| 610 | ASSERT(imap->br_startblock != DELAYSTARTBLOCK); |
| 611 | |
| 612 | xfs_map_buffer(inode, bh, imap, offset); |
| 613 | set_buffer_mapped(bh); |
| 614 | clear_buffer_delay(bh); |
| 615 | clear_buffer_unwritten(bh); |
| 616 | } |
| 617 | |
| 618 | /* |
| 619 | * Test if a given page is suitable for writing as part of an unwritten |
| 620 | * or delayed allocate extent. |
| 621 | */ |
| 622 | STATIC int |
| 623 | xfs_check_page_type( |
| 624 | struct page *page, |
| 625 | unsigned int type) |
| 626 | { |
| 627 | if (PageWriteback(page)) |
| 628 | return 0; |
| 629 | |
| 630 | if (page->mapping && page_has_buffers(page)) { |
| 631 | struct buffer_head *bh, *head; |
| 632 | int acceptable = 0; |
| 633 | |
| 634 | bh = head = page_buffers(page); |
| 635 | do { |
| 636 | if (buffer_unwritten(bh)) |
| 637 | acceptable += (type == IO_UNWRITTEN); |
| 638 | else if (buffer_delay(bh)) |
| 639 | acceptable += (type == IO_DELALLOC); |
| 640 | else if (buffer_dirty(bh) && buffer_mapped(bh)) |
| 641 | acceptable += (type == IO_OVERWRITE); |
| 642 | else |
| 643 | break; |
| 644 | } while ((bh = bh->b_this_page) != head); |
| 645 | |
| 646 | if (acceptable) |
| 647 | return 1; |
| 648 | } |
| 649 | |
| 650 | return 0; |
| 651 | } |
| 652 | |
| 653 | /* |
| 654 | * Allocate & map buffers for page given the extent map. Write it out. |
| 655 | * except for the original page of a writepage, this is called on |
| 656 | * delalloc/unwritten pages only, for the original page it is possible |
| 657 | * that the page has no mapping at all. |
| 658 | */ |
| 659 | STATIC int |
| 660 | xfs_convert_page( |
| 661 | struct inode *inode, |
| 662 | struct page *page, |
| 663 | loff_t tindex, |
| 664 | struct xfs_bmbt_irec *imap, |
| 665 | xfs_ioend_t **ioendp, |
| 666 | struct writeback_control *wbc) |
| 667 | { |
| 668 | struct buffer_head *bh, *head; |
| 669 | xfs_off_t end_offset; |
| 670 | unsigned long p_offset; |
| 671 | unsigned int type; |
| 672 | int len, page_dirty; |
| 673 | int count = 0, done = 0, uptodate = 1; |
| 674 | xfs_off_t offset = page_offset(page); |
| 675 | |
| 676 | if (page->index != tindex) |
| 677 | goto fail; |
| 678 | if (!trylock_page(page)) |
| 679 | goto fail; |
| 680 | if (PageWriteback(page)) |
| 681 | goto fail_unlock_page; |
| 682 | if (page->mapping != inode->i_mapping) |
| 683 | goto fail_unlock_page; |
| 684 | if (!xfs_check_page_type(page, (*ioendp)->io_type)) |
| 685 | goto fail_unlock_page; |
| 686 | |
| 687 | /* |
| 688 | * page_dirty is initially a count of buffers on the page before |
| 689 | * EOF and is decremented as we move each into a cleanable state. |
| 690 | * |
| 691 | * Derivation: |
| 692 | * |
| 693 | * End offset is the highest offset that this page should represent. |
| 694 | * If we are on the last page, (end_offset & (PAGE_CACHE_SIZE - 1)) |
| 695 | * will evaluate non-zero and be less than PAGE_CACHE_SIZE and |
| 696 | * hence give us the correct page_dirty count. On any other page, |
| 697 | * it will be zero and in that case we need page_dirty to be the |
| 698 | * count of buffers on the page. |
| 699 | */ |
| 700 | end_offset = min_t(unsigned long long, |
| 701 | (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT, |
| 702 | i_size_read(inode)); |
| 703 | |
| 704 | len = 1 << inode->i_blkbits; |
| 705 | p_offset = min_t(unsigned long, end_offset & (PAGE_CACHE_SIZE - 1), |
| 706 | PAGE_CACHE_SIZE); |
| 707 | p_offset = p_offset ? roundup(p_offset, len) : PAGE_CACHE_SIZE; |
| 708 | page_dirty = p_offset / len; |
| 709 | |
| 710 | bh = head = page_buffers(page); |
| 711 | do { |
| 712 | if (offset >= end_offset) |
| 713 | break; |
| 714 | if (!buffer_uptodate(bh)) |
| 715 | uptodate = 0; |
| 716 | if (!(PageUptodate(page) || buffer_uptodate(bh))) { |
| 717 | done = 1; |
| 718 | continue; |
| 719 | } |
| 720 | |
| 721 | if (buffer_unwritten(bh) || buffer_delay(bh) || |
| 722 | buffer_mapped(bh)) { |
| 723 | if (buffer_unwritten(bh)) |
| 724 | type = IO_UNWRITTEN; |
| 725 | else if (buffer_delay(bh)) |
| 726 | type = IO_DELALLOC; |
| 727 | else |
| 728 | type = IO_OVERWRITE; |
| 729 | |
| 730 | if (!xfs_imap_valid(inode, imap, offset)) { |
| 731 | done = 1; |
| 732 | continue; |
| 733 | } |
| 734 | |
| 735 | lock_buffer(bh); |
| 736 | if (type != IO_OVERWRITE) |
| 737 | xfs_map_at_offset(inode, bh, imap, offset); |
| 738 | xfs_add_to_ioend(inode, bh, offset, type, |
| 739 | ioendp, done); |
| 740 | |
| 741 | page_dirty--; |
| 742 | count++; |
| 743 | } else { |
| 744 | done = 1; |
| 745 | } |
| 746 | } while (offset += len, (bh = bh->b_this_page) != head); |
| 747 | |
| 748 | if (uptodate && bh == head) |
| 749 | SetPageUptodate(page); |
| 750 | |
| 751 | if (count) { |
| 752 | if (--wbc->nr_to_write <= 0 && |
| 753 | wbc->sync_mode == WB_SYNC_NONE) |
| 754 | done = 1; |
| 755 | } |
| 756 | xfs_start_page_writeback(page, !page_dirty, count); |
| 757 | |
| 758 | return done; |
| 759 | fail_unlock_page: |
| 760 | unlock_page(page); |
| 761 | fail: |
| 762 | return 1; |
| 763 | } |
| 764 | |
| 765 | /* |
| 766 | * Convert & write out a cluster of pages in the same extent as defined |
| 767 | * by mp and following the start page. |
| 768 | */ |
| 769 | STATIC void |
| 770 | xfs_cluster_write( |
| 771 | struct inode *inode, |
| 772 | pgoff_t tindex, |
| 773 | struct xfs_bmbt_irec *imap, |
| 774 | xfs_ioend_t **ioendp, |
| 775 | struct writeback_control *wbc, |
| 776 | pgoff_t tlast) |
| 777 | { |
| 778 | struct pagevec pvec; |
| 779 | int done = 0, i; |
| 780 | |
| 781 | pagevec_init(&pvec, 0); |
| 782 | while (!done && tindex <= tlast) { |
| 783 | unsigned len = min_t(pgoff_t, PAGEVEC_SIZE, tlast - tindex + 1); |
| 784 | |
| 785 | if (!pagevec_lookup(&pvec, inode->i_mapping, tindex, len)) |
| 786 | break; |
| 787 | |
| 788 | for (i = 0; i < pagevec_count(&pvec); i++) { |
| 789 | done = xfs_convert_page(inode, pvec.pages[i], tindex++, |
| 790 | imap, ioendp, wbc); |
| 791 | if (done) |
| 792 | break; |
| 793 | } |
| 794 | |
| 795 | pagevec_release(&pvec); |
| 796 | cond_resched(); |
| 797 | } |
| 798 | } |
| 799 | |
| 800 | STATIC void |
| 801 | xfs_vm_invalidatepage( |
| 802 | struct page *page, |
| 803 | unsigned long offset) |
| 804 | { |
| 805 | trace_xfs_invalidatepage(page->mapping->host, page, offset); |
| 806 | block_invalidatepage(page, offset); |
| 807 | } |
| 808 | |
| 809 | /* |
| 810 | * If the page has delalloc buffers on it, we need to punch them out before we |
| 811 | * invalidate the page. If we don't, we leave a stale delalloc mapping on the |
| 812 | * inode that can trip a BUG() in xfs_get_blocks() later on if a direct IO read |
| 813 | * is done on that same region - the delalloc extent is returned when none is |
| 814 | * supposed to be there. |
| 815 | * |
| 816 | * We prevent this by truncating away the delalloc regions on the page before |
| 817 | * invalidating it. Because they are delalloc, we can do this without needing a |
| 818 | * transaction. Indeed - if we get ENOSPC errors, we have to be able to do this |
| 819 | * truncation without a transaction as there is no space left for block |
| 820 | * reservation (typically why we see a ENOSPC in writeback). |
| 821 | * |
| 822 | * This is not a performance critical path, so for now just do the punching a |
| 823 | * buffer head at a time. |
| 824 | */ |
| 825 | STATIC void |
| 826 | xfs_aops_discard_page( |
| 827 | struct page *page) |
| 828 | { |
| 829 | struct inode *inode = page->mapping->host; |
| 830 | struct xfs_inode *ip = XFS_I(inode); |
| 831 | struct buffer_head *bh, *head; |
| 832 | loff_t offset = page_offset(page); |
| 833 | |
| 834 | if (!xfs_check_page_type(page, IO_DELALLOC)) |
| 835 | goto out_invalidate; |
| 836 | |
| 837 | if (XFS_FORCED_SHUTDOWN(ip->i_mount)) |
| 838 | goto out_invalidate; |
| 839 | |
| 840 | xfs_alert(ip->i_mount, |
| 841 | "page discard on page %p, inode 0x%llx, offset %llu.", |
| 842 | page, ip->i_ino, offset); |
| 843 | |
| 844 | xfs_ilock(ip, XFS_ILOCK_EXCL); |
| 845 | bh = head = page_buffers(page); |
| 846 | do { |
| 847 | int error; |
| 848 | xfs_fileoff_t start_fsb; |
| 849 | |
| 850 | if (!buffer_delay(bh)) |
| 851 | goto next_buffer; |
| 852 | |
| 853 | start_fsb = XFS_B_TO_FSBT(ip->i_mount, offset); |
| 854 | error = xfs_bmap_punch_delalloc_range(ip, start_fsb, 1); |
| 855 | if (error) { |
| 856 | /* something screwed, just bail */ |
| 857 | if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) { |
| 858 | xfs_alert(ip->i_mount, |
| 859 | "page discard unable to remove delalloc mapping."); |
| 860 | } |
| 861 | break; |
| 862 | } |
| 863 | next_buffer: |
| 864 | offset += 1 << inode->i_blkbits; |
| 865 | |
| 866 | } while ((bh = bh->b_this_page) != head); |
| 867 | |
| 868 | xfs_iunlock(ip, XFS_ILOCK_EXCL); |
| 869 | out_invalidate: |
| 870 | xfs_vm_invalidatepage(page, 0); |
| 871 | return; |
| 872 | } |
| 873 | |
| 874 | /* |
| 875 | * Write out a dirty page. |
| 876 | * |
| 877 | * For delalloc space on the page we need to allocate space and flush it. |
| 878 | * For unwritten space on the page we need to start the conversion to |
| 879 | * regular allocated space. |
| 880 | * For any other dirty buffer heads on the page we should flush them. |
| 881 | */ |
| 882 | STATIC int |
| 883 | xfs_vm_writepage( |
| 884 | struct page *page, |
| 885 | struct writeback_control *wbc) |
| 886 | { |
| 887 | struct inode *inode = page->mapping->host; |
| 888 | struct buffer_head *bh, *head; |
| 889 | struct xfs_bmbt_irec imap; |
| 890 | xfs_ioend_t *ioend = NULL, *iohead = NULL; |
| 891 | loff_t offset; |
| 892 | unsigned int type; |
| 893 | __uint64_t end_offset; |
| 894 | pgoff_t end_index, last_index; |
| 895 | ssize_t len; |
| 896 | int err, imap_valid = 0, uptodate = 1; |
| 897 | int count = 0; |
| 898 | int nonblocking = 0; |
| 899 | |
| 900 | trace_xfs_writepage(inode, page, 0); |
| 901 | |
| 902 | ASSERT(page_has_buffers(page)); |
| 903 | |
| 904 | /* |
| 905 | * Refuse to write the page out if we are called from reclaim context. |
| 906 | * |
| 907 | * This avoids stack overflows when called from deeply used stacks in |
| 908 | * random callers for direct reclaim or memcg reclaim. We explicitly |
| 909 | * allow reclaim from kswapd as the stack usage there is relatively low. |
| 910 | * |
| 911 | * This should never happen except in the case of a VM regression so |
| 912 | * warn about it. |
| 913 | */ |
| 914 | if (WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD)) == |
| 915 | PF_MEMALLOC)) |
| 916 | goto redirty; |
| 917 | |
| 918 | /* |
| 919 | * Given that we do not allow direct reclaim to call us, we should |
| 920 | * never be called while in a filesystem transaction. |
| 921 | */ |
| 922 | if (WARN_ON(current->flags & PF_FSTRANS)) |
| 923 | goto redirty; |
| 924 | |
| 925 | /* Is this page beyond the end of the file? */ |
| 926 | offset = i_size_read(inode); |
| 927 | end_index = offset >> PAGE_CACHE_SHIFT; |
| 928 | last_index = (offset - 1) >> PAGE_CACHE_SHIFT; |
| 929 | if (page->index >= end_index) { |
| 930 | if ((page->index >= end_index + 1) || |
| 931 | !(i_size_read(inode) & (PAGE_CACHE_SIZE - 1))) { |
| 932 | unlock_page(page); |
| 933 | return 0; |
| 934 | } |
| 935 | } |
| 936 | |
| 937 | end_offset = min_t(unsigned long long, |
| 938 | (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT, |
| 939 | offset); |
| 940 | len = 1 << inode->i_blkbits; |
| 941 | |
| 942 | bh = head = page_buffers(page); |
| 943 | offset = page_offset(page); |
| 944 | type = IO_OVERWRITE; |
| 945 | |
| 946 | if (wbc->sync_mode == WB_SYNC_NONE) |
| 947 | nonblocking = 1; |
| 948 | |
| 949 | do { |
| 950 | int new_ioend = 0; |
| 951 | |
| 952 | if (offset >= end_offset) |
| 953 | break; |
| 954 | if (!buffer_uptodate(bh)) |
| 955 | uptodate = 0; |
| 956 | |
| 957 | /* |
| 958 | * set_page_dirty dirties all buffers in a page, independent |
| 959 | * of their state. The dirty state however is entirely |
| 960 | * meaningless for holes (!mapped && uptodate), so skip |
| 961 | * buffers covering holes here. |
| 962 | */ |
| 963 | if (!buffer_mapped(bh) && buffer_uptodate(bh)) { |
| 964 | imap_valid = 0; |
| 965 | continue; |
| 966 | } |
| 967 | |
| 968 | if (buffer_unwritten(bh)) { |
| 969 | if (type != IO_UNWRITTEN) { |
| 970 | type = IO_UNWRITTEN; |
| 971 | imap_valid = 0; |
| 972 | } |
| 973 | } else if (buffer_delay(bh)) { |
| 974 | if (type != IO_DELALLOC) { |
| 975 | type = IO_DELALLOC; |
| 976 | imap_valid = 0; |
| 977 | } |
| 978 | } else if (buffer_uptodate(bh)) { |
| 979 | if (type != IO_OVERWRITE) { |
| 980 | type = IO_OVERWRITE; |
| 981 | imap_valid = 0; |
| 982 | } |
| 983 | } else { |
| 984 | if (PageUptodate(page)) |
| 985 | ASSERT(buffer_mapped(bh)); |
| 986 | /* |
| 987 | * This buffer is not uptodate and will not be |
| 988 | * written to disk. Ensure that we will put any |
| 989 | * subsequent writeable buffers into a new |
| 990 | * ioend. |
| 991 | */ |
| 992 | imap_valid = 0; |
| 993 | continue; |
| 994 | } |
| 995 | |
| 996 | if (imap_valid) |
| 997 | imap_valid = xfs_imap_valid(inode, &imap, offset); |
| 998 | if (!imap_valid) { |
| 999 | /* |
| 1000 | * If we didn't have a valid mapping then we need to |
| 1001 | * put the new mapping into a separate ioend structure. |
| 1002 | * This ensures non-contiguous extents always have |
| 1003 | * separate ioends, which is particularly important |
| 1004 | * for unwritten extent conversion at I/O completion |
| 1005 | * time. |
| 1006 | */ |
| 1007 | new_ioend = 1; |
| 1008 | err = xfs_map_blocks(inode, offset, &imap, type, |
| 1009 | nonblocking); |
| 1010 | if (err) |
| 1011 | goto error; |
| 1012 | imap_valid = xfs_imap_valid(inode, &imap, offset); |
| 1013 | } |
| 1014 | if (imap_valid) { |
| 1015 | lock_buffer(bh); |
| 1016 | if (type != IO_OVERWRITE) |
| 1017 | xfs_map_at_offset(inode, bh, &imap, offset); |
| 1018 | xfs_add_to_ioend(inode, bh, offset, type, &ioend, |
| 1019 | new_ioend); |
| 1020 | count++; |
| 1021 | } |
| 1022 | |
| 1023 | if (!iohead) |
| 1024 | iohead = ioend; |
| 1025 | |
| 1026 | } while (offset += len, ((bh = bh->b_this_page) != head)); |
| 1027 | |
| 1028 | if (uptodate && bh == head) |
| 1029 | SetPageUptodate(page); |
| 1030 | |
| 1031 | xfs_start_page_writeback(page, 1, count); |
| 1032 | |
| 1033 | if (ioend && imap_valid) { |
| 1034 | xfs_off_t end_index; |
| 1035 | |
| 1036 | end_index = imap.br_startoff + imap.br_blockcount; |
| 1037 | |
| 1038 | /* to bytes */ |
| 1039 | end_index <<= inode->i_blkbits; |
| 1040 | |
| 1041 | /* to pages */ |
| 1042 | end_index = (end_index - 1) >> PAGE_CACHE_SHIFT; |
| 1043 | |
| 1044 | /* check against file size */ |
| 1045 | if (end_index > last_index) |
| 1046 | end_index = last_index; |
| 1047 | |
| 1048 | xfs_cluster_write(inode, page->index + 1, &imap, &ioend, |
| 1049 | wbc, end_index); |
| 1050 | } |
| 1051 | |
| 1052 | if (iohead) { |
| 1053 | /* |
| 1054 | * Reserve log space if we might write beyond the on-disk |
| 1055 | * inode size. |
| 1056 | */ |
| 1057 | if (ioend->io_type != IO_UNWRITTEN && |
| 1058 | xfs_ioend_is_append(ioend)) { |
| 1059 | err = xfs_setfilesize_trans_alloc(ioend); |
| 1060 | if (err) |
| 1061 | goto error; |
| 1062 | } |
| 1063 | |
| 1064 | xfs_submit_ioend(wbc, iohead); |
| 1065 | } |
| 1066 | |
| 1067 | return 0; |
| 1068 | |
| 1069 | error: |
| 1070 | if (iohead) |
| 1071 | xfs_cancel_ioend(iohead); |
| 1072 | |
| 1073 | if (err == -EAGAIN) |
| 1074 | goto redirty; |
| 1075 | |
| 1076 | xfs_aops_discard_page(page); |
| 1077 | ClearPageUptodate(page); |
| 1078 | unlock_page(page); |
| 1079 | return err; |
| 1080 | |
| 1081 | redirty: |
| 1082 | redirty_page_for_writepage(wbc, page); |
| 1083 | unlock_page(page); |
| 1084 | return 0; |
| 1085 | } |
| 1086 | |
| 1087 | STATIC int |
| 1088 | xfs_vm_writepages( |
| 1089 | struct address_space *mapping, |
| 1090 | struct writeback_control *wbc) |
| 1091 | { |
| 1092 | xfs_iflags_clear(XFS_I(mapping->host), XFS_ITRUNCATED); |
| 1093 | return generic_writepages(mapping, wbc); |
| 1094 | } |
| 1095 | |
| 1096 | /* |
| 1097 | * Called to move a page into cleanable state - and from there |
| 1098 | * to be released. The page should already be clean. We always |
| 1099 | * have buffer heads in this call. |
| 1100 | * |
| 1101 | * Returns 1 if the page is ok to release, 0 otherwise. |
| 1102 | */ |
| 1103 | STATIC int |
| 1104 | xfs_vm_releasepage( |
| 1105 | struct page *page, |
| 1106 | gfp_t gfp_mask) |
| 1107 | { |
| 1108 | int delalloc, unwritten; |
| 1109 | |
| 1110 | trace_xfs_releasepage(page->mapping->host, page, 0); |
| 1111 | |
| 1112 | xfs_count_page_state(page, &delalloc, &unwritten); |
| 1113 | |
| 1114 | if (WARN_ON(delalloc)) |
| 1115 | return 0; |
| 1116 | if (WARN_ON(unwritten)) |
| 1117 | return 0; |
| 1118 | |
| 1119 | return try_to_free_buffers(page); |
| 1120 | } |
| 1121 | |
| 1122 | STATIC int |
| 1123 | __xfs_get_blocks( |
| 1124 | struct inode *inode, |
| 1125 | sector_t iblock, |
| 1126 | struct buffer_head *bh_result, |
| 1127 | int create, |
| 1128 | int direct) |
| 1129 | { |
| 1130 | struct xfs_inode *ip = XFS_I(inode); |
| 1131 | struct xfs_mount *mp = ip->i_mount; |
| 1132 | xfs_fileoff_t offset_fsb, end_fsb; |
| 1133 | int error = 0; |
| 1134 | int lockmode = 0; |
| 1135 | struct xfs_bmbt_irec imap; |
| 1136 | int nimaps = 1; |
| 1137 | xfs_off_t offset; |
| 1138 | ssize_t size; |
| 1139 | int new = 0; |
| 1140 | |
| 1141 | if (XFS_FORCED_SHUTDOWN(mp)) |
| 1142 | return -XFS_ERROR(EIO); |
| 1143 | |
| 1144 | offset = (xfs_off_t)iblock << inode->i_blkbits; |
| 1145 | ASSERT(bh_result->b_size >= (1 << inode->i_blkbits)); |
| 1146 | size = bh_result->b_size; |
| 1147 | |
| 1148 | if (!create && direct && offset >= i_size_read(inode)) |
| 1149 | return 0; |
| 1150 | |
| 1151 | /* |
| 1152 | * Direct I/O is usually done on preallocated files, so try getting |
| 1153 | * a block mapping without an exclusive lock first. For buffered |
| 1154 | * writes we already have the exclusive iolock anyway, so avoiding |
| 1155 | * a lock roundtrip here by taking the ilock exclusive from the |
| 1156 | * beginning is a useful micro optimization. |
| 1157 | */ |
| 1158 | if (create && !direct) { |
| 1159 | lockmode = XFS_ILOCK_EXCL; |
| 1160 | xfs_ilock(ip, lockmode); |
| 1161 | } else { |
| 1162 | lockmode = xfs_ilock_map_shared(ip); |
| 1163 | } |
| 1164 | |
| 1165 | ASSERT(offset <= mp->m_maxioffset); |
| 1166 | if (offset + size > mp->m_maxioffset) |
| 1167 | size = mp->m_maxioffset - offset; |
| 1168 | end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + size); |
| 1169 | offset_fsb = XFS_B_TO_FSBT(mp, offset); |
| 1170 | |
| 1171 | error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb, |
| 1172 | &imap, &nimaps, XFS_BMAPI_ENTIRE); |
| 1173 | if (error) |
| 1174 | goto out_unlock; |
| 1175 | |
| 1176 | if (create && |
| 1177 | (!nimaps || |
| 1178 | (imap.br_startblock == HOLESTARTBLOCK || |
| 1179 | imap.br_startblock == DELAYSTARTBLOCK))) { |
| 1180 | if (direct || xfs_get_extsz_hint(ip)) { |
| 1181 | /* |
| 1182 | * Drop the ilock in preparation for starting the block |
| 1183 | * allocation transaction. It will be retaken |
| 1184 | * exclusively inside xfs_iomap_write_direct for the |
| 1185 | * actual allocation. |
| 1186 | */ |
| 1187 | xfs_iunlock(ip, lockmode); |
| 1188 | error = xfs_iomap_write_direct(ip, offset, size, |
| 1189 | &imap, nimaps); |
| 1190 | if (error) |
| 1191 | return -error; |
| 1192 | new = 1; |
| 1193 | } else { |
| 1194 | /* |
| 1195 | * Delalloc reservations do not require a transaction, |
| 1196 | * we can go on without dropping the lock here. If we |
| 1197 | * are allocating a new delalloc block, make sure that |
| 1198 | * we set the new flag so that we mark the buffer new so |
| 1199 | * that we know that it is newly allocated if the write |
| 1200 | * fails. |
| 1201 | */ |
| 1202 | if (nimaps && imap.br_startblock == HOLESTARTBLOCK) |
| 1203 | new = 1; |
| 1204 | error = xfs_iomap_write_delay(ip, offset, size, &imap); |
| 1205 | if (error) |
| 1206 | goto out_unlock; |
| 1207 | |
| 1208 | xfs_iunlock(ip, lockmode); |
| 1209 | } |
| 1210 | |
| 1211 | trace_xfs_get_blocks_alloc(ip, offset, size, 0, &imap); |
| 1212 | } else if (nimaps) { |
| 1213 | trace_xfs_get_blocks_found(ip, offset, size, 0, &imap); |
| 1214 | xfs_iunlock(ip, lockmode); |
| 1215 | } else { |
| 1216 | trace_xfs_get_blocks_notfound(ip, offset, size); |
| 1217 | goto out_unlock; |
| 1218 | } |
| 1219 | |
| 1220 | if (imap.br_startblock != HOLESTARTBLOCK && |
| 1221 | imap.br_startblock != DELAYSTARTBLOCK) { |
| 1222 | /* |
| 1223 | * For unwritten extents do not report a disk address on |
| 1224 | * the read case (treat as if we're reading into a hole). |
| 1225 | */ |
| 1226 | if (create || !ISUNWRITTEN(&imap)) |
| 1227 | xfs_map_buffer(inode, bh_result, &imap, offset); |
| 1228 | if (create && ISUNWRITTEN(&imap)) { |
| 1229 | if (direct) |
| 1230 | bh_result->b_private = inode; |
| 1231 | set_buffer_unwritten(bh_result); |
| 1232 | } |
| 1233 | } |
| 1234 | |
| 1235 | /* |
| 1236 | * If this is a realtime file, data may be on a different device. |
| 1237 | * to that pointed to from the buffer_head b_bdev currently. |
| 1238 | */ |
| 1239 | bh_result->b_bdev = xfs_find_bdev_for_inode(inode); |
| 1240 | |
| 1241 | /* |
| 1242 | * If we previously allocated a block out beyond eof and we are now |
| 1243 | * coming back to use it then we will need to flag it as new even if it |
| 1244 | * has a disk address. |
| 1245 | * |
| 1246 | * With sub-block writes into unwritten extents we also need to mark |
| 1247 | * the buffer as new so that the unwritten parts of the buffer gets |
| 1248 | * correctly zeroed. |
| 1249 | */ |
| 1250 | if (create && |
| 1251 | ((!buffer_mapped(bh_result) && !buffer_uptodate(bh_result)) || |
| 1252 | (offset >= i_size_read(inode)) || |
| 1253 | (new || ISUNWRITTEN(&imap)))) |
| 1254 | set_buffer_new(bh_result); |
| 1255 | |
| 1256 | if (imap.br_startblock == DELAYSTARTBLOCK) { |
| 1257 | BUG_ON(direct); |
| 1258 | if (create) { |
| 1259 | set_buffer_uptodate(bh_result); |
| 1260 | set_buffer_mapped(bh_result); |
| 1261 | set_buffer_delay(bh_result); |
| 1262 | } |
| 1263 | } |
| 1264 | |
| 1265 | /* |
| 1266 | * If this is O_DIRECT or the mpage code calling tell them how large |
| 1267 | * the mapping is, so that we can avoid repeated get_blocks calls. |
| 1268 | */ |
| 1269 | if (direct || size > (1 << inode->i_blkbits)) { |
| 1270 | xfs_off_t mapping_size; |
| 1271 | |
| 1272 | mapping_size = imap.br_startoff + imap.br_blockcount - iblock; |
| 1273 | mapping_size <<= inode->i_blkbits; |
| 1274 | |
| 1275 | ASSERT(mapping_size > 0); |
| 1276 | if (mapping_size > size) |
| 1277 | mapping_size = size; |
| 1278 | if (mapping_size > LONG_MAX) |
| 1279 | mapping_size = LONG_MAX; |
| 1280 | |
| 1281 | bh_result->b_size = mapping_size; |
| 1282 | } |
| 1283 | |
| 1284 | return 0; |
| 1285 | |
| 1286 | out_unlock: |
| 1287 | xfs_iunlock(ip, lockmode); |
| 1288 | return -error; |
| 1289 | } |
| 1290 | |
| 1291 | int |
| 1292 | xfs_get_blocks( |
| 1293 | struct inode *inode, |
| 1294 | sector_t iblock, |
| 1295 | struct buffer_head *bh_result, |
| 1296 | int create) |
| 1297 | { |
| 1298 | return __xfs_get_blocks(inode, iblock, bh_result, create, 0); |
| 1299 | } |
| 1300 | |
| 1301 | STATIC int |
| 1302 | xfs_get_blocks_direct( |
| 1303 | struct inode *inode, |
| 1304 | sector_t iblock, |
| 1305 | struct buffer_head *bh_result, |
| 1306 | int create) |
| 1307 | { |
| 1308 | return __xfs_get_blocks(inode, iblock, bh_result, create, 1); |
| 1309 | } |
| 1310 | |
| 1311 | /* |
| 1312 | * Complete a direct I/O write request. |
| 1313 | * |
| 1314 | * If the private argument is non-NULL __xfs_get_blocks signals us that we |
| 1315 | * need to issue a transaction to convert the range from unwritten to written |
| 1316 | * extents. In case this is regular synchronous I/O we just call xfs_end_io |
| 1317 | * to do this and we are done. But in case this was a successful AIO |
| 1318 | * request this handler is called from interrupt context, from which we |
| 1319 | * can't start transactions. In that case offload the I/O completion to |
| 1320 | * the workqueues we also use for buffered I/O completion. |
| 1321 | */ |
| 1322 | STATIC void |
| 1323 | xfs_end_io_direct_write( |
| 1324 | struct kiocb *iocb, |
| 1325 | loff_t offset, |
| 1326 | ssize_t size, |
| 1327 | void *private, |
| 1328 | int ret, |
| 1329 | bool is_async) |
| 1330 | { |
| 1331 | struct xfs_ioend *ioend = iocb->private; |
| 1332 | |
| 1333 | /* |
| 1334 | * While the generic direct I/O code updates the inode size, it does |
| 1335 | * so only after the end_io handler is called, which means our |
| 1336 | * end_io handler thinks the on-disk size is outside the in-core |
| 1337 | * size. To prevent this just update it a little bit earlier here. |
| 1338 | */ |
| 1339 | if (offset + size > i_size_read(ioend->io_inode)) |
| 1340 | i_size_write(ioend->io_inode, offset + size); |
| 1341 | |
| 1342 | /* |
| 1343 | * blockdev_direct_IO can return an error even after the I/O |
| 1344 | * completion handler was called. Thus we need to protect |
| 1345 | * against double-freeing. |
| 1346 | */ |
| 1347 | iocb->private = NULL; |
| 1348 | |
| 1349 | ioend->io_offset = offset; |
| 1350 | ioend->io_size = size; |
| 1351 | ioend->io_iocb = iocb; |
| 1352 | ioend->io_result = ret; |
| 1353 | if (private && size > 0) |
| 1354 | ioend->io_type = IO_UNWRITTEN; |
| 1355 | |
| 1356 | if (is_async) { |
| 1357 | ioend->io_isasync = 1; |
| 1358 | xfs_finish_ioend(ioend); |
| 1359 | } else { |
| 1360 | xfs_finish_ioend_sync(ioend); |
| 1361 | } |
| 1362 | } |
| 1363 | |
| 1364 | STATIC ssize_t |
| 1365 | xfs_vm_direct_IO( |
| 1366 | int rw, |
| 1367 | struct kiocb *iocb, |
| 1368 | const struct iovec *iov, |
| 1369 | loff_t offset, |
| 1370 | unsigned long nr_segs) |
| 1371 | { |
| 1372 | struct inode *inode = iocb->ki_filp->f_mapping->host; |
| 1373 | struct block_device *bdev = xfs_find_bdev_for_inode(inode); |
| 1374 | struct xfs_ioend *ioend = NULL; |
| 1375 | ssize_t ret; |
| 1376 | |
| 1377 | if (rw & WRITE) { |
| 1378 | size_t size = iov_length(iov, nr_segs); |
| 1379 | |
| 1380 | /* |
| 1381 | * We need to preallocate a transaction for a size update |
| 1382 | * here. In the case that this write both updates the size |
| 1383 | * and converts at least on unwritten extent we will cancel |
| 1384 | * the still clean transaction after the I/O has finished. |
| 1385 | */ |
| 1386 | iocb->private = ioend = xfs_alloc_ioend(inode, IO_DIRECT); |
| 1387 | if (offset + size > XFS_I(inode)->i_d.di_size) { |
| 1388 | ret = xfs_setfilesize_trans_alloc(ioend); |
| 1389 | if (ret) |
| 1390 | goto out_destroy_ioend; |
| 1391 | ioend->io_isdirect = 1; |
| 1392 | } |
| 1393 | |
| 1394 | ret = __blockdev_direct_IO(rw, iocb, inode, bdev, iov, |
| 1395 | offset, nr_segs, |
| 1396 | xfs_get_blocks_direct, |
| 1397 | xfs_end_io_direct_write, NULL, 0); |
| 1398 | if (ret != -EIOCBQUEUED && iocb->private) |
| 1399 | goto out_trans_cancel; |
| 1400 | } else { |
| 1401 | ret = __blockdev_direct_IO(rw, iocb, inode, bdev, iov, |
| 1402 | offset, nr_segs, |
| 1403 | xfs_get_blocks_direct, |
| 1404 | NULL, NULL, 0); |
| 1405 | } |
| 1406 | |
| 1407 | return ret; |
| 1408 | |
| 1409 | out_trans_cancel: |
| 1410 | if (ioend->io_append_trans) { |
| 1411 | current_set_flags_nested(&ioend->io_append_trans->t_pflags, |
| 1412 | PF_FSTRANS); |
| 1413 | xfs_trans_cancel(ioend->io_append_trans, 0); |
| 1414 | } |
| 1415 | out_destroy_ioend: |
| 1416 | xfs_destroy_ioend(ioend); |
| 1417 | return ret; |
| 1418 | } |
| 1419 | |
| 1420 | /* |
| 1421 | * Punch out the delalloc blocks we have already allocated. |
| 1422 | * |
| 1423 | * Don't bother with xfs_setattr given that nothing can have made it to disk yet |
| 1424 | * as the page is still locked at this point. |
| 1425 | */ |
| 1426 | STATIC void |
| 1427 | xfs_vm_kill_delalloc_range( |
| 1428 | struct inode *inode, |
| 1429 | loff_t start, |
| 1430 | loff_t end) |
| 1431 | { |
| 1432 | struct xfs_inode *ip = XFS_I(inode); |
| 1433 | xfs_fileoff_t start_fsb; |
| 1434 | xfs_fileoff_t end_fsb; |
| 1435 | int error; |
| 1436 | |
| 1437 | start_fsb = XFS_B_TO_FSB(ip->i_mount, start); |
| 1438 | end_fsb = XFS_B_TO_FSB(ip->i_mount, end); |
| 1439 | if (end_fsb <= start_fsb) |
| 1440 | return; |
| 1441 | |
| 1442 | xfs_ilock(ip, XFS_ILOCK_EXCL); |
| 1443 | error = xfs_bmap_punch_delalloc_range(ip, start_fsb, |
| 1444 | end_fsb - start_fsb); |
| 1445 | if (error) { |
| 1446 | /* something screwed, just bail */ |
| 1447 | if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) { |
| 1448 | xfs_alert(ip->i_mount, |
| 1449 | "xfs_vm_write_failed: unable to clean up ino %lld", |
| 1450 | ip->i_ino); |
| 1451 | } |
| 1452 | } |
| 1453 | xfs_iunlock(ip, XFS_ILOCK_EXCL); |
| 1454 | } |
| 1455 | |
| 1456 | STATIC void |
| 1457 | xfs_vm_write_failed( |
| 1458 | struct inode *inode, |
| 1459 | struct page *page, |
| 1460 | loff_t pos, |
| 1461 | unsigned len) |
| 1462 | { |
| 1463 | loff_t block_offset = pos & PAGE_MASK; |
| 1464 | loff_t block_start; |
| 1465 | loff_t block_end; |
| 1466 | loff_t from = pos & (PAGE_CACHE_SIZE - 1); |
| 1467 | loff_t to = from + len; |
| 1468 | struct buffer_head *bh, *head; |
| 1469 | |
| 1470 | ASSERT(block_offset + from == pos); |
| 1471 | |
| 1472 | head = page_buffers(page); |
| 1473 | block_start = 0; |
| 1474 | for (bh = head; bh != head || !block_start; |
| 1475 | bh = bh->b_this_page, block_start = block_end, |
| 1476 | block_offset += bh->b_size) { |
| 1477 | block_end = block_start + bh->b_size; |
| 1478 | |
| 1479 | /* skip buffers before the write */ |
| 1480 | if (block_end <= from) |
| 1481 | continue; |
| 1482 | |
| 1483 | /* if the buffer is after the write, we're done */ |
| 1484 | if (block_start >= to) |
| 1485 | break; |
| 1486 | |
| 1487 | if (!buffer_delay(bh)) |
| 1488 | continue; |
| 1489 | |
| 1490 | if (!buffer_new(bh) && block_offset < i_size_read(inode)) |
| 1491 | continue; |
| 1492 | |
| 1493 | xfs_vm_kill_delalloc_range(inode, block_offset, |
| 1494 | block_offset + bh->b_size); |
| 1495 | } |
| 1496 | |
| 1497 | } |
| 1498 | |
| 1499 | /* |
| 1500 | * This used to call block_write_begin(), but it unlocks and releases the page |
| 1501 | * on error, and we need that page to be able to punch stale delalloc blocks out |
| 1502 | * on failure. hence we copy-n-waste it here and call xfs_vm_write_failed() at |
| 1503 | * the appropriate point. |
| 1504 | */ |
| 1505 | STATIC int |
| 1506 | xfs_vm_write_begin( |
| 1507 | struct file *file, |
| 1508 | struct address_space *mapping, |
| 1509 | loff_t pos, |
| 1510 | unsigned len, |
| 1511 | unsigned flags, |
| 1512 | struct page **pagep, |
| 1513 | void **fsdata) |
| 1514 | { |
| 1515 | pgoff_t index = pos >> PAGE_CACHE_SHIFT; |
| 1516 | struct page *page; |
| 1517 | int status; |
| 1518 | |
| 1519 | ASSERT(len <= PAGE_CACHE_SIZE); |
| 1520 | |
| 1521 | page = grab_cache_page_write_begin(mapping, index, |
| 1522 | flags | AOP_FLAG_NOFS); |
| 1523 | if (!page) |
| 1524 | return -ENOMEM; |
| 1525 | |
| 1526 | status = __block_write_begin(page, pos, len, xfs_get_blocks); |
| 1527 | if (unlikely(status)) { |
| 1528 | struct inode *inode = mapping->host; |
| 1529 | |
| 1530 | xfs_vm_write_failed(inode, page, pos, len); |
| 1531 | unlock_page(page); |
| 1532 | |
| 1533 | if (pos + len > i_size_read(inode)) |
| 1534 | truncate_pagecache(inode, pos + len, i_size_read(inode)); |
| 1535 | |
| 1536 | page_cache_release(page); |
| 1537 | page = NULL; |
| 1538 | } |
| 1539 | |
| 1540 | *pagep = page; |
| 1541 | return status; |
| 1542 | } |
| 1543 | |
| 1544 | /* |
| 1545 | * On failure, we only need to kill delalloc blocks beyond EOF because they |
| 1546 | * will never be written. For blocks within EOF, generic_write_end() zeros them |
| 1547 | * so they are safe to leave alone and be written with all the other valid data. |
| 1548 | */ |
| 1549 | STATIC int |
| 1550 | xfs_vm_write_end( |
| 1551 | struct file *file, |
| 1552 | struct address_space *mapping, |
| 1553 | loff_t pos, |
| 1554 | unsigned len, |
| 1555 | unsigned copied, |
| 1556 | struct page *page, |
| 1557 | void *fsdata) |
| 1558 | { |
| 1559 | int ret; |
| 1560 | |
| 1561 | ASSERT(len <= PAGE_CACHE_SIZE); |
| 1562 | |
| 1563 | ret = generic_write_end(file, mapping, pos, len, copied, page, fsdata); |
| 1564 | if (unlikely(ret < len)) { |
| 1565 | struct inode *inode = mapping->host; |
| 1566 | size_t isize = i_size_read(inode); |
| 1567 | loff_t to = pos + len; |
| 1568 | |
| 1569 | if (to > isize) { |
| 1570 | truncate_pagecache(inode, to, isize); |
| 1571 | xfs_vm_kill_delalloc_range(inode, isize, to); |
| 1572 | } |
| 1573 | } |
| 1574 | return ret; |
| 1575 | } |
| 1576 | |
| 1577 | STATIC sector_t |
| 1578 | xfs_vm_bmap( |
| 1579 | struct address_space *mapping, |
| 1580 | sector_t block) |
| 1581 | { |
| 1582 | struct inode *inode = (struct inode *)mapping->host; |
| 1583 | struct xfs_inode *ip = XFS_I(inode); |
| 1584 | |
| 1585 | trace_xfs_vm_bmap(XFS_I(inode)); |
| 1586 | xfs_ilock(ip, XFS_IOLOCK_SHARED); |
| 1587 | xfs_flush_pages(ip, (xfs_off_t)0, -1, 0, FI_REMAPF); |
| 1588 | xfs_iunlock(ip, XFS_IOLOCK_SHARED); |
| 1589 | return generic_block_bmap(mapping, block, xfs_get_blocks); |
| 1590 | } |
| 1591 | |
| 1592 | STATIC int |
| 1593 | xfs_vm_readpage( |
| 1594 | struct file *unused, |
| 1595 | struct page *page) |
| 1596 | { |
| 1597 | return mpage_readpage(page, xfs_get_blocks); |
| 1598 | } |
| 1599 | |
| 1600 | STATIC int |
| 1601 | xfs_vm_readpages( |
| 1602 | struct file *unused, |
| 1603 | struct address_space *mapping, |
| 1604 | struct list_head *pages, |
| 1605 | unsigned nr_pages) |
| 1606 | { |
| 1607 | return mpage_readpages(mapping, pages, nr_pages, xfs_get_blocks); |
| 1608 | } |
| 1609 | |
| 1610 | const struct address_space_operations xfs_address_space_operations = { |
| 1611 | .readpage = xfs_vm_readpage, |
| 1612 | .readpages = xfs_vm_readpages, |
| 1613 | .writepage = xfs_vm_writepage, |
| 1614 | .writepages = xfs_vm_writepages, |
| 1615 | .releasepage = xfs_vm_releasepage, |
| 1616 | .invalidatepage = xfs_vm_invalidatepage, |
| 1617 | .write_begin = xfs_vm_write_begin, |
| 1618 | .write_end = xfs_vm_write_end, |
| 1619 | .bmap = xfs_vm_bmap, |
| 1620 | .direct_IO = xfs_vm_direct_IO, |
| 1621 | .migratepage = buffer_migrate_page, |
| 1622 | .is_partially_uptodate = block_is_partially_uptodate, |
| 1623 | .error_remove_page = generic_error_remove_page, |
| 1624 | }; |