linux/fs/xfs/xfs_aops.c
<<
>>
Prefs
   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
  39void
  40xfs_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
  58STATIC struct block_device *
  59xfs_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 */
  77STATIC void
  78xfs_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 */
 102static 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
 108STATIC int
 109xfs_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 will pass freeze protection with a transaction.  So tell lockdep
 128         * we released it.
 129         */
 130        rwsem_release(&ioend->io_inode->i_sb->s_writers.lock_map[SB_FREEZE_FS-1],
 131                      1, _THIS_IP_);
 132        /*
 133         * We hand off the transaction to the completion thread now, so
 134         * clear the flag here.
 135         */
 136        current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
 137        return 0;
 138}
 139
 140/*
 141 * Update on-disk file size now that data has been written to disk.
 142 */
 143STATIC int
 144xfs_setfilesize(
 145        struct xfs_ioend        *ioend)
 146{
 147        struct xfs_inode        *ip = XFS_I(ioend->io_inode);
 148        struct xfs_trans        *tp = ioend->io_append_trans;
 149        xfs_fsize_t             isize;
 150
 151        /*
 152         * The transaction was allocated in the I/O submission thread,
 153         * thus we need to mark ourselves as beeing in a transaction
 154         * manually.
 155         */
 156        current_set_flags_nested(&tp->t_pflags, PF_FSTRANS);
 157
 158        xfs_ilock(ip, XFS_ILOCK_EXCL);
 159        isize = xfs_new_eof(ip, ioend->io_offset + ioend->io_size);
 160        if (!isize) {
 161                xfs_iunlock(ip, XFS_ILOCK_EXCL);
 162                xfs_trans_cancel(tp, 0);
 163                return 0;
 164        }
 165
 166        trace_xfs_setfilesize(ip, ioend->io_offset, ioend->io_size);
 167
 168        ip->i_d.di_size = isize;
 169        xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
 170        xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 171
 172        return xfs_trans_commit(tp, 0);
 173}
 174
 175/*
 176 * Schedule IO completion handling on the final put of an ioend.
 177 *
 178 * If there is no work to do we might as well call it a day and free the
 179 * ioend right now.
 180 */
 181STATIC void
 182xfs_finish_ioend(
 183        struct xfs_ioend        *ioend)
 184{
 185        if (atomic_dec_and_test(&ioend->io_remaining)) {
 186                struct xfs_mount        *mp = XFS_I(ioend->io_inode)->i_mount;
 187
 188                if (ioend->io_type == XFS_IO_UNWRITTEN)
 189                        queue_work(mp->m_unwritten_workqueue, &ioend->io_work);
 190                else if (ioend->io_append_trans)
 191                        queue_work(mp->m_data_workqueue, &ioend->io_work);
 192                else
 193                        xfs_destroy_ioend(ioend);
 194        }
 195}
 196
 197/*
 198 * IO write completion.
 199 */
 200STATIC void
 201xfs_end_io(
 202        struct work_struct *work)
 203{
 204        xfs_ioend_t     *ioend = container_of(work, xfs_ioend_t, io_work);
 205        struct xfs_inode *ip = XFS_I(ioend->io_inode);
 206        int             error = 0;
 207
 208        if (ioend->io_append_trans) {
 209                /*
 210                 * We've got freeze protection passed with the transaction.
 211                 * Tell lockdep about it.
 212                 */
 213                rwsem_acquire_read(
 214                        &ioend->io_inode->i_sb->s_writers.lock_map[SB_FREEZE_FS-1],
 215                        0, 1, _THIS_IP_);
 216        }
 217        if (XFS_FORCED_SHUTDOWN(ip->i_mount)) {
 218                ioend->io_error = -EIO;
 219                goto done;
 220        }
 221        if (ioend->io_error)
 222                goto done;
 223
 224        /*
 225         * For unwritten extents we need to issue transactions to convert a
 226         * range to normal written extens after the data I/O has finished.
 227         */
 228        if (ioend->io_type == XFS_IO_UNWRITTEN) {
 229                /*
 230                 * For buffered I/O we never preallocate a transaction when
 231                 * doing the unwritten extent conversion, but for direct I/O
 232                 * we do not know if we are converting an unwritten extent
 233                 * or not at the point where we preallocate the transaction.
 234                 */
 235                if (ioend->io_append_trans) {
 236                        ASSERT(ioend->io_isdirect);
 237
 238                        current_set_flags_nested(
 239                                &ioend->io_append_trans->t_pflags, PF_FSTRANS);
 240                        xfs_trans_cancel(ioend->io_append_trans, 0);
 241                }
 242
 243                error = xfs_iomap_write_unwritten(ip, ioend->io_offset,
 244                                                 ioend->io_size);
 245                if (error) {
 246                        ioend->io_error = -error;
 247                        goto done;
 248                }
 249        } else if (ioend->io_append_trans) {
 250                error = xfs_setfilesize(ioend);
 251                if (error)
 252                        ioend->io_error = -error;
 253        } else {
 254                ASSERT(!xfs_ioend_is_append(ioend));
 255        }
 256
 257done:
 258        xfs_destroy_ioend(ioend);
 259}
 260
 261/*
 262 * Call IO completion handling in caller context on the final put of an ioend.
 263 */
 264STATIC void
 265xfs_finish_ioend_sync(
 266        struct xfs_ioend        *ioend)
 267{
 268        if (atomic_dec_and_test(&ioend->io_remaining))
 269                xfs_end_io(&ioend->io_work);
 270}
 271
 272/*
 273 * Allocate and initialise an IO completion structure.
 274 * We need to track unwritten extent write completion here initially.
 275 * We'll need to extend this for updating the ondisk inode size later
 276 * (vs. incore size).
 277 */
 278STATIC xfs_ioend_t *
 279xfs_alloc_ioend(
 280        struct inode            *inode,
 281        unsigned int            type)
 282{
 283        xfs_ioend_t             *ioend;
 284
 285        ioend = mempool_alloc(xfs_ioend_pool, GFP_NOFS);
 286
 287        /*
 288         * Set the count to 1 initially, which will prevent an I/O
 289         * completion callback from happening before we have started
 290         * all the I/O from calling the completion routine too early.
 291         */
 292        atomic_set(&ioend->io_remaining, 1);
 293        ioend->io_isasync = 0;
 294        ioend->io_isdirect = 0;
 295        ioend->io_error = 0;
 296        ioend->io_list = NULL;
 297        ioend->io_type = type;
 298        ioend->io_inode = inode;
 299        ioend->io_buffer_head = NULL;
 300        ioend->io_buffer_tail = NULL;
 301        ioend->io_offset = 0;
 302        ioend->io_size = 0;
 303        ioend->io_iocb = NULL;
 304        ioend->io_result = 0;
 305        ioend->io_append_trans = NULL;
 306
 307        INIT_WORK(&ioend->io_work, xfs_end_io);
 308        return ioend;
 309}
 310
 311STATIC int
 312xfs_map_blocks(
 313        struct inode            *inode,
 314        loff_t                  offset,
 315        struct xfs_bmbt_irec    *imap,
 316        int                     type,
 317        int                     nonblocking)
 318{
 319        struct xfs_inode        *ip = XFS_I(inode);
 320        struct xfs_mount        *mp = ip->i_mount;
 321        ssize_t                 count = 1 << inode->i_blkbits;
 322        xfs_fileoff_t           offset_fsb, end_fsb;
 323        int                     error = 0;
 324        int                     bmapi_flags = XFS_BMAPI_ENTIRE;
 325        int                     nimaps = 1;
 326
 327        if (XFS_FORCED_SHUTDOWN(mp))
 328                return -XFS_ERROR(EIO);
 329
 330        if (type == XFS_IO_UNWRITTEN)
 331                bmapi_flags |= XFS_BMAPI_IGSTATE;
 332
 333        if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
 334                if (nonblocking)
 335                        return -XFS_ERROR(EAGAIN);
 336                xfs_ilock(ip, XFS_ILOCK_SHARED);
 337        }
 338
 339        ASSERT(ip->i_d.di_format != XFS_DINODE_FMT_BTREE ||
 340               (ip->i_df.if_flags & XFS_IFEXTENTS));
 341        ASSERT(offset <= mp->m_super->s_maxbytes);
 342
 343        if (offset + count > mp->m_super->s_maxbytes)
 344                count = mp->m_super->s_maxbytes - offset;
 345        end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + count);
 346        offset_fsb = XFS_B_TO_FSBT(mp, offset);
 347        error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb,
 348                                imap, &nimaps, bmapi_flags);
 349        xfs_iunlock(ip, XFS_ILOCK_SHARED);
 350
 351        if (error)
 352                return -XFS_ERROR(error);
 353
 354        if (type == XFS_IO_DELALLOC &&
 355            (!nimaps || isnullstartblock(imap->br_startblock))) {
 356                error = xfs_iomap_write_allocate(ip, offset, count, imap);
 357                if (!error)
 358                        trace_xfs_map_blocks_alloc(ip, offset, count, type, imap);
 359                return -XFS_ERROR(error);
 360        }
 361
 362#ifdef DEBUG
 363        if (type == XFS_IO_UNWRITTEN) {
 364                ASSERT(nimaps);
 365                ASSERT(imap->br_startblock != HOLESTARTBLOCK);
 366                ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
 367        }
 368#endif
 369        if (nimaps)
 370                trace_xfs_map_blocks_found(ip, offset, count, type, imap);
 371        return 0;
 372}
 373
 374STATIC int
 375xfs_imap_valid(
 376        struct inode            *inode,
 377        struct xfs_bmbt_irec    *imap,
 378        xfs_off_t               offset)
 379{
 380        offset >>= inode->i_blkbits;
 381
 382        return offset >= imap->br_startoff &&
 383                offset < imap->br_startoff + imap->br_blockcount;
 384}
 385
 386/*
 387 * BIO completion handler for buffered IO.
 388 */
 389STATIC void
 390xfs_end_bio(
 391        struct bio              *bio,
 392        int                     error)
 393{
 394        xfs_ioend_t             *ioend = bio->bi_private;
 395
 396        ASSERT(atomic_read(&bio->bi_cnt) >= 1);
 397        ioend->io_error = test_bit(BIO_UPTODATE, &bio->bi_flags) ? 0 : error;
 398
 399        /* Toss bio and pass work off to an xfsdatad thread */
 400        bio->bi_private = NULL;
 401        bio->bi_end_io = NULL;
 402        bio_put(bio);
 403
 404        xfs_finish_ioend(ioend);
 405}
 406
 407STATIC void
 408xfs_submit_ioend_bio(
 409        struct writeback_control *wbc,
 410        xfs_ioend_t             *ioend,
 411        struct bio              *bio)
 412{
 413        atomic_inc(&ioend->io_remaining);
 414        bio->bi_private = ioend;
 415        bio->bi_end_io = xfs_end_bio;
 416        submit_bio(wbc->sync_mode == WB_SYNC_ALL ? WRITE_SYNC : WRITE, bio);
 417}
 418
 419STATIC struct bio *
 420xfs_alloc_ioend_bio(
 421        struct buffer_head      *bh)
 422{
 423        int                     nvecs = bio_get_nr_vecs(bh->b_bdev);
 424        struct bio              *bio = bio_alloc(GFP_NOIO, nvecs);
 425
 426        ASSERT(bio->bi_private == NULL);
 427        bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
 428        bio->bi_bdev = bh->b_bdev;
 429        return bio;
 430}
 431
 432STATIC void
 433xfs_start_buffer_writeback(
 434        struct buffer_head      *bh)
 435{
 436        ASSERT(buffer_mapped(bh));
 437        ASSERT(buffer_locked(bh));
 438        ASSERT(!buffer_delay(bh));
 439        ASSERT(!buffer_unwritten(bh));
 440
 441        mark_buffer_async_write(bh);
 442        set_buffer_uptodate(bh);
 443        clear_buffer_dirty(bh);
 444}
 445
 446STATIC void
 447xfs_start_page_writeback(
 448        struct page             *page,
 449        int                     clear_dirty,
 450        int                     buffers)
 451{
 452        ASSERT(PageLocked(page));
 453        ASSERT(!PageWriteback(page));
 454        if (clear_dirty)
 455                clear_page_dirty_for_io(page);
 456        set_page_writeback(page);
 457        unlock_page(page);
 458        /* If no buffers on the page are to be written, finish it here */
 459        if (!buffers)
 460                end_page_writeback(page);
 461}
 462
 463static inline int bio_add_buffer(struct bio *bio, struct buffer_head *bh)
 464{
 465        return bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
 466}
 467
 468/*
 469 * Submit all of the bios for all of the ioends we have saved up, covering the
 470 * initial writepage page and also any probed pages.
 471 *
 472 * Because we may have multiple ioends spanning a page, we need to start
 473 * writeback on all the buffers before we submit them for I/O. If we mark the
 474 * buffers as we got, then we can end up with a page that only has buffers
 475 * marked async write and I/O complete on can occur before we mark the other
 476 * buffers async write.
 477 *
 478 * The end result of this is that we trip a bug in end_page_writeback() because
 479 * we call it twice for the one page as the code in end_buffer_async_write()
 480 * assumes that all buffers on the page are started at the same time.
 481 *
 482 * The fix is two passes across the ioend list - one to start writeback on the
 483 * buffer_heads, and then submit them for I/O on the second pass.
 484 *
 485 * If @fail is non-zero, it means that we have a situation where some part of
 486 * the submission process has failed after we have marked paged for writeback
 487 * and unlocked them. In this situation, we need to fail the ioend chain rather
 488 * than submit it to IO. This typically only happens on a filesystem shutdown.
 489 */
 490STATIC void
 491xfs_submit_ioend(
 492        struct writeback_control *wbc,
 493        xfs_ioend_t             *ioend,
 494        int                     fail)
 495{
 496        xfs_ioend_t             *head = ioend;
 497        xfs_ioend_t             *next;
 498        struct buffer_head      *bh;
 499        struct bio              *bio;
 500        sector_t                lastblock = 0;
 501
 502        /* Pass 1 - start writeback */
 503        do {
 504                next = ioend->io_list;
 505                for (bh = ioend->io_buffer_head; bh; bh = bh->b_private)
 506                        xfs_start_buffer_writeback(bh);
 507        } while ((ioend = next) != NULL);
 508
 509        /* Pass 2 - submit I/O */
 510        ioend = head;
 511        do {
 512                next = ioend->io_list;
 513                bio = NULL;
 514
 515                /*
 516                 * If we are failing the IO now, just mark the ioend with an
 517                 * error and finish it. This will run IO completion immediately
 518                 * as there is only one reference to the ioend at this point in
 519                 * time.
 520                 */
 521                if (fail) {
 522                        ioend->io_error = -fail;
 523                        xfs_finish_ioend(ioend);
 524                        continue;
 525                }
 526
 527                for (bh = ioend->io_buffer_head; bh; bh = bh->b_private) {
 528
 529                        if (!bio) {
 530 retry:
 531                                bio = xfs_alloc_ioend_bio(bh);
 532                        } else if (bh->b_blocknr != lastblock + 1) {
 533                                xfs_submit_ioend_bio(wbc, ioend, bio);
 534                                goto retry;
 535                        }
 536
 537                        if (bio_add_buffer(bio, bh) != bh->b_size) {
 538                                xfs_submit_ioend_bio(wbc, ioend, bio);
 539                                goto retry;
 540                        }
 541
 542                        lastblock = bh->b_blocknr;
 543                }
 544                if (bio)
 545                        xfs_submit_ioend_bio(wbc, ioend, bio);
 546                xfs_finish_ioend(ioend);
 547        } while ((ioend = next) != NULL);
 548}
 549
 550/*
 551 * Cancel submission of all buffer_heads so far in this endio.
 552 * Toss the endio too.  Only ever called for the initial page
 553 * in a writepage request, so only ever one page.
 554 */
 555STATIC void
 556xfs_cancel_ioend(
 557        xfs_ioend_t             *ioend)
 558{
 559        xfs_ioend_t             *next;
 560        struct buffer_head      *bh, *next_bh;
 561
 562        do {
 563                next = ioend->io_list;
 564                bh = ioend->io_buffer_head;
 565                do {
 566                        next_bh = bh->b_private;
 567                        clear_buffer_async_write(bh);
 568                        unlock_buffer(bh);
 569                } while ((bh = next_bh) != NULL);
 570
 571                mempool_free(ioend, xfs_ioend_pool);
 572        } while ((ioend = next) != NULL);
 573}
 574
 575/*
 576 * Test to see if we've been building up a completion structure for
 577 * earlier buffers -- if so, we try to append to this ioend if we
 578 * can, otherwise we finish off any current ioend and start another.
 579 * Return true if we've finished the given ioend.
 580 */
 581STATIC void
 582xfs_add_to_ioend(
 583        struct inode            *inode,
 584        struct buffer_head      *bh,
 585        xfs_off_t               offset,
 586        unsigned int            type,
 587        xfs_ioend_t             **result,
 588        int                     need_ioend)
 589{
 590        xfs_ioend_t             *ioend = *result;
 591
 592        if (!ioend || need_ioend || type != ioend->io_type) {
 593                xfs_ioend_t     *previous = *result;
 594
 595                ioend = xfs_alloc_ioend(inode, type);
 596                ioend->io_offset = offset;
 597                ioend->io_buffer_head = bh;
 598                ioend->io_buffer_tail = bh;
 599                if (previous)
 600                        previous->io_list = ioend;
 601                *result = ioend;
 602        } else {
 603                ioend->io_buffer_tail->b_private = bh;
 604                ioend->io_buffer_tail = bh;
 605        }
 606
 607        bh->b_private = NULL;
 608        ioend->io_size += bh->b_size;
 609}
 610
 611STATIC void
 612xfs_map_buffer(
 613        struct inode            *inode,
 614        struct buffer_head      *bh,
 615        struct xfs_bmbt_irec    *imap,
 616        xfs_off_t               offset)
 617{
 618        sector_t                bn;
 619        struct xfs_mount        *m = XFS_I(inode)->i_mount;
 620        xfs_off_t               iomap_offset = XFS_FSB_TO_B(m, imap->br_startoff);
 621        xfs_daddr_t             iomap_bn = xfs_fsb_to_db(XFS_I(inode), imap->br_startblock);
 622
 623        ASSERT(imap->br_startblock != HOLESTARTBLOCK);
 624        ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
 625
 626        bn = (iomap_bn >> (inode->i_blkbits - BBSHIFT)) +
 627              ((offset - iomap_offset) >> inode->i_blkbits);
 628
 629        ASSERT(bn || XFS_IS_REALTIME_INODE(XFS_I(inode)));
 630
 631        bh->b_blocknr = bn;
 632        set_buffer_mapped(bh);
 633}
 634
 635STATIC void
 636xfs_map_at_offset(
 637        struct inode            *inode,
 638        struct buffer_head      *bh,
 639        struct xfs_bmbt_irec    *imap,
 640        xfs_off_t               offset)
 641{
 642        ASSERT(imap->br_startblock != HOLESTARTBLOCK);
 643        ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
 644
 645        xfs_map_buffer(inode, bh, imap, offset);
 646        set_buffer_mapped(bh);
 647        clear_buffer_delay(bh);
 648        clear_buffer_unwritten(bh);
 649}
 650
 651/*
 652 * Test if a given page is suitable for writing as part of an unwritten
 653 * or delayed allocate extent.
 654 */
 655STATIC int
 656xfs_check_page_type(
 657        struct page             *page,
 658        unsigned int            type)
 659{
 660        if (PageWriteback(page))
 661                return 0;
 662
 663        if (page->mapping && page_has_buffers(page)) {
 664                struct buffer_head      *bh, *head;
 665                int                     acceptable = 0;
 666
 667                bh = head = page_buffers(page);
 668                do {
 669                        if (buffer_unwritten(bh))
 670                                acceptable += (type == XFS_IO_UNWRITTEN);
 671                        else if (buffer_delay(bh))
 672                                acceptable += (type == XFS_IO_DELALLOC);
 673                        else if (buffer_dirty(bh) && buffer_mapped(bh))
 674                                acceptable += (type == XFS_IO_OVERWRITE);
 675                        else
 676                                break;
 677                } while ((bh = bh->b_this_page) != head);
 678
 679                if (acceptable)
 680                        return 1;
 681        }
 682
 683        return 0;
 684}
 685
 686/*
 687 * Allocate & map buffers for page given the extent map. Write it out.
 688 * except for the original page of a writepage, this is called on
 689 * delalloc/unwritten pages only, for the original page it is possible
 690 * that the page has no mapping at all.
 691 */
 692STATIC int
 693xfs_convert_page(
 694        struct inode            *inode,
 695        struct page             *page,
 696        loff_t                  tindex,
 697        struct xfs_bmbt_irec    *imap,
 698        xfs_ioend_t             **ioendp,
 699        struct writeback_control *wbc)
 700{
 701        struct buffer_head      *bh, *head;
 702        xfs_off_t               end_offset;
 703        unsigned long           p_offset;
 704        unsigned int            type;
 705        int                     len, page_dirty;
 706        int                     count = 0, done = 0, uptodate = 1;
 707        xfs_off_t               offset = page_offset(page);
 708
 709        if (page->index != tindex)
 710                goto fail;
 711        if (!trylock_page(page))
 712                goto fail;
 713        if (PageWriteback(page))
 714                goto fail_unlock_page;
 715        if (page->mapping != inode->i_mapping)
 716                goto fail_unlock_page;
 717        if (!xfs_check_page_type(page, (*ioendp)->io_type))
 718                goto fail_unlock_page;
 719
 720        /*
 721         * page_dirty is initially a count of buffers on the page before
 722         * EOF and is decremented as we move each into a cleanable state.
 723         *
 724         * Derivation:
 725         *
 726         * End offset is the highest offset that this page should represent.
 727         * If we are on the last page, (end_offset & (PAGE_CACHE_SIZE - 1))
 728         * will evaluate non-zero and be less than PAGE_CACHE_SIZE and
 729         * hence give us the correct page_dirty count. On any other page,
 730         * it will be zero and in that case we need page_dirty to be the
 731         * count of buffers on the page.
 732         */
 733        end_offset = min_t(unsigned long long,
 734                        (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT,
 735                        i_size_read(inode));
 736
 737        len = 1 << inode->i_blkbits;
 738        p_offset = min_t(unsigned long, end_offset & (PAGE_CACHE_SIZE - 1),
 739                                        PAGE_CACHE_SIZE);
 740        p_offset = p_offset ? roundup(p_offset, len) : PAGE_CACHE_SIZE;
 741        page_dirty = p_offset / len;
 742
 743        bh = head = page_buffers(page);
 744        do {
 745                if (offset >= end_offset)
 746                        break;
 747                if (!buffer_uptodate(bh))
 748                        uptodate = 0;
 749                if (!(PageUptodate(page) || buffer_uptodate(bh))) {
 750                        done = 1;
 751                        continue;
 752                }
 753
 754                if (buffer_unwritten(bh) || buffer_delay(bh) ||
 755                    buffer_mapped(bh)) {
 756                        if (buffer_unwritten(bh))
 757                                type = XFS_IO_UNWRITTEN;
 758                        else if (buffer_delay(bh))
 759                                type = XFS_IO_DELALLOC;
 760                        else
 761                                type = XFS_IO_OVERWRITE;
 762
 763                        if (!xfs_imap_valid(inode, imap, offset)) {
 764                                done = 1;
 765                                continue;
 766                        }
 767
 768                        lock_buffer(bh);
 769                        if (type != XFS_IO_OVERWRITE)
 770                                xfs_map_at_offset(inode, bh, imap, offset);
 771                        xfs_add_to_ioend(inode, bh, offset, type,
 772                                         ioendp, done);
 773
 774                        page_dirty--;
 775                        count++;
 776                } else {
 777                        done = 1;
 778                }
 779        } while (offset += len, (bh = bh->b_this_page) != head);
 780
 781        if (uptodate && bh == head)
 782                SetPageUptodate(page);
 783
 784        if (count) {
 785                if (--wbc->nr_to_write <= 0 &&
 786                    wbc->sync_mode == WB_SYNC_NONE)
 787                        done = 1;
 788        }
 789        xfs_start_page_writeback(page, !page_dirty, count);
 790
 791        return done;
 792 fail_unlock_page:
 793        unlock_page(page);
 794 fail:
 795        return 1;
 796}
 797
 798/*
 799 * Convert & write out a cluster of pages in the same extent as defined
 800 * by mp and following the start page.
 801 */
 802STATIC void
 803xfs_cluster_write(
 804        struct inode            *inode,
 805        pgoff_t                 tindex,
 806        struct xfs_bmbt_irec    *imap,
 807        xfs_ioend_t             **ioendp,
 808        struct writeback_control *wbc,
 809        pgoff_t                 tlast)
 810{
 811        struct pagevec          pvec;
 812        int                     done = 0, i;
 813
 814        pagevec_init(&pvec, 0);
 815        while (!done && tindex <= tlast) {
 816                unsigned len = min_t(pgoff_t, PAGEVEC_SIZE, tlast - tindex + 1);
 817
 818                if (!pagevec_lookup(&pvec, inode->i_mapping, tindex, len))
 819                        break;
 820
 821                for (i = 0; i < pagevec_count(&pvec); i++) {
 822                        done = xfs_convert_page(inode, pvec.pages[i], tindex++,
 823                                        imap, ioendp, wbc);
 824                        if (done)
 825                                break;
 826                }
 827
 828                pagevec_release(&pvec);
 829                cond_resched();
 830        }
 831}
 832
 833STATIC void
 834xfs_vm_invalidatepage(
 835        struct page             *page,
 836        unsigned long           offset)
 837{
 838        trace_xfs_invalidatepage(page->mapping->host, page, offset);
 839        block_invalidatepage(page, offset);
 840}
 841
 842/*
 843 * If the page has delalloc buffers on it, we need to punch them out before we
 844 * invalidate the page. If we don't, we leave a stale delalloc mapping on the
 845 * inode that can trip a BUG() in xfs_get_blocks() later on if a direct IO read
 846 * is done on that same region - the delalloc extent is returned when none is
 847 * supposed to be there.
 848 *
 849 * We prevent this by truncating away the delalloc regions on the page before
 850 * invalidating it. Because they are delalloc, we can do this without needing a
 851 * transaction. Indeed - if we get ENOSPC errors, we have to be able to do this
 852 * truncation without a transaction as there is no space left for block
 853 * reservation (typically why we see a ENOSPC in writeback).
 854 *
 855 * This is not a performance critical path, so for now just do the punching a
 856 * buffer head at a time.
 857 */
 858STATIC void
 859xfs_aops_discard_page(
 860        struct page             *page)
 861{
 862        struct inode            *inode = page->mapping->host;
 863        struct xfs_inode        *ip = XFS_I(inode);
 864        struct buffer_head      *bh, *head;
 865        loff_t                  offset = page_offset(page);
 866
 867        if (!xfs_check_page_type(page, XFS_IO_DELALLOC))
 868                goto out_invalidate;
 869
 870        if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 871                goto out_invalidate;
 872
 873        xfs_alert(ip->i_mount,
 874                "page discard on page %p, inode 0x%llx, offset %llu.",
 875                        page, ip->i_ino, offset);
 876
 877        xfs_ilock(ip, XFS_ILOCK_EXCL);
 878        bh = head = page_buffers(page);
 879        do {
 880                int             error;
 881                xfs_fileoff_t   start_fsb;
 882
 883                if (!buffer_delay(bh))
 884                        goto next_buffer;
 885
 886                start_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
 887                error = xfs_bmap_punch_delalloc_range(ip, start_fsb, 1);
 888                if (error) {
 889                        /* something screwed, just bail */
 890                        if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
 891                                xfs_alert(ip->i_mount,
 892                        "page discard unable to remove delalloc mapping.");
 893                        }
 894                        break;
 895                }
 896next_buffer:
 897                offset += 1 << inode->i_blkbits;
 898
 899        } while ((bh = bh->b_this_page) != head);
 900
 901        xfs_iunlock(ip, XFS_ILOCK_EXCL);
 902out_invalidate:
 903        xfs_vm_invalidatepage(page, 0);
 904        return;
 905}
 906
 907/*
 908 * Write out a dirty page.
 909 *
 910 * For delalloc space on the page we need to allocate space and flush it.
 911 * For unwritten space on the page we need to start the conversion to
 912 * regular allocated space.
 913 * For any other dirty buffer heads on the page we should flush them.
 914 */
 915STATIC int
 916xfs_vm_writepage(
 917        struct page             *page,
 918        struct writeback_control *wbc)
 919{
 920        struct inode            *inode = page->mapping->host;
 921        struct buffer_head      *bh, *head;
 922        struct xfs_bmbt_irec    imap;
 923        xfs_ioend_t             *ioend = NULL, *iohead = NULL;
 924        loff_t                  offset;
 925        unsigned int            type;
 926        __uint64_t              end_offset;
 927        pgoff_t                 end_index, last_index;
 928        ssize_t                 len;
 929        int                     err, imap_valid = 0, uptodate = 1;
 930        int                     count = 0;
 931        int                     nonblocking = 0;
 932
 933        trace_xfs_writepage(inode, page, 0);
 934
 935        ASSERT(page_has_buffers(page));
 936
 937        /*
 938         * Refuse to write the page out if we are called from reclaim context.
 939         *
 940         * This avoids stack overflows when called from deeply used stacks in
 941         * random callers for direct reclaim or memcg reclaim.  We explicitly
 942         * allow reclaim from kswapd as the stack usage there is relatively low.
 943         *
 944         * This should never happen except in the case of a VM regression so
 945         * warn about it.
 946         */
 947        if (WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD)) ==
 948                        PF_MEMALLOC))
 949                goto redirty;
 950
 951        /*
 952         * Given that we do not allow direct reclaim to call us, we should
 953         * never be called while in a filesystem transaction.
 954         */
 955        if (WARN_ON(current->flags & PF_FSTRANS))
 956                goto redirty;
 957
 958        /* Is this page beyond the end of the file? */
 959        offset = i_size_read(inode);
 960        end_index = offset >> PAGE_CACHE_SHIFT;
 961        last_index = (offset - 1) >> PAGE_CACHE_SHIFT;
 962        if (page->index >= end_index) {
 963                unsigned offset_into_page = offset & (PAGE_CACHE_SIZE - 1);
 964
 965                /*
 966                 * Just skip the page if it is fully outside i_size, e.g. due
 967                 * to a truncate operation that is in progress.
 968                 */
 969                if (page->index >= end_index + 1 || offset_into_page == 0) {
 970                        unlock_page(page);
 971                        return 0;
 972                }
 973
 974                /*
 975                 * The page straddles i_size.  It must be zeroed out on each
 976                 * and every writepage invocation because it may be mmapped.
 977                 * "A file is mapped in multiples of the page size.  For a file
 978                 * that is not a multiple of the  page size, the remaining
 979                 * memory is zeroed when mapped, and writes to that region are
 980                 * not written out to the file."
 981                 */
 982                zero_user_segment(page, offset_into_page, PAGE_CACHE_SIZE);
 983        }
 984
 985        end_offset = min_t(unsigned long long,
 986                        (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT,
 987                        offset);
 988        len = 1 << inode->i_blkbits;
 989
 990        bh = head = page_buffers(page);
 991        offset = page_offset(page);
 992        type = XFS_IO_OVERWRITE;
 993
 994        if (wbc->sync_mode == WB_SYNC_NONE)
 995                nonblocking = 1;
 996
 997        do {
 998                int new_ioend = 0;
 999
1000                if (offset >= end_offset)
1001                        break;
1002                if (!buffer_uptodate(bh))
1003                        uptodate = 0;
1004
1005                /*
1006                 * set_page_dirty dirties all buffers in a page, independent
1007                 * of their state.  The dirty state however is entirely
1008                 * meaningless for holes (!mapped && uptodate), so skip
1009                 * buffers covering holes here.
1010                 */
1011                if (!buffer_mapped(bh) && buffer_uptodate(bh)) {
1012                        imap_valid = 0;
1013                        continue;
1014                }
1015
1016                if (buffer_unwritten(bh)) {
1017                        if (type != XFS_IO_UNWRITTEN) {
1018                                type = XFS_IO_UNWRITTEN;
1019                                imap_valid = 0;
1020                        }
1021                } else if (buffer_delay(bh)) {
1022                        if (type != XFS_IO_DELALLOC) {
1023                                type = XFS_IO_DELALLOC;
1024                                imap_valid = 0;
1025                        }
1026                } else if (buffer_uptodate(bh)) {
1027                        if (type != XFS_IO_OVERWRITE) {
1028                                type = XFS_IO_OVERWRITE;
1029                                imap_valid = 0;
1030                        }
1031                } else {
1032                        if (PageUptodate(page))
1033                                ASSERT(buffer_mapped(bh));
1034                        /*
1035                         * This buffer is not uptodate and will not be
1036                         * written to disk.  Ensure that we will put any
1037                         * subsequent writeable buffers into a new
1038                         * ioend.
1039                         */
1040                        imap_valid = 0;
1041                        continue;
1042                }
1043
1044                if (imap_valid)
1045                        imap_valid = xfs_imap_valid(inode, &imap, offset);
1046                if (!imap_valid) {
1047                        /*
1048                         * If we didn't have a valid mapping then we need to
1049                         * put the new mapping into a separate ioend structure.
1050                         * This ensures non-contiguous extents always have
1051                         * separate ioends, which is particularly important
1052                         * for unwritten extent conversion at I/O completion
1053                         * time.
1054                         */
1055                        new_ioend = 1;
1056                        err = xfs_map_blocks(inode, offset, &imap, type,
1057                                             nonblocking);
1058                        if (err)
1059                                goto error;
1060                        imap_valid = xfs_imap_valid(inode, &imap, offset);
1061                }
1062                if (imap_valid) {
1063                        lock_buffer(bh);
1064                        if (type != XFS_IO_OVERWRITE)
1065                                xfs_map_at_offset(inode, bh, &imap, offset);
1066                        xfs_add_to_ioend(inode, bh, offset, type, &ioend,
1067                                         new_ioend);
1068                        count++;
1069                }
1070
1071                if (!iohead)
1072                        iohead = ioend;
1073
1074        } while (offset += len, ((bh = bh->b_this_page) != head));
1075
1076        if (uptodate && bh == head)
1077                SetPageUptodate(page);
1078
1079        xfs_start_page_writeback(page, 1, count);
1080
1081        /* if there is no IO to be submitted for this page, we are done */
1082        if (!ioend)
1083                return 0;
1084
1085        ASSERT(iohead);
1086
1087        /*
1088         * Any errors from this point onwards need tobe reported through the IO
1089         * completion path as we have marked the initial page as under writeback
1090         * and unlocked it.
1091         */
1092        if (imap_valid) {
1093                xfs_off_t               end_index;
1094
1095                end_index = imap.br_startoff + imap.br_blockcount;
1096
1097                /* to bytes */
1098                end_index <<= inode->i_blkbits;
1099
1100                /* to pages */
1101                end_index = (end_index - 1) >> PAGE_CACHE_SHIFT;
1102
1103                /* check against file size */
1104                if (end_index > last_index)
1105                        end_index = last_index;
1106
1107                xfs_cluster_write(inode, page->index + 1, &imap, &ioend,
1108                                  wbc, end_index);
1109        }
1110
1111
1112        /*
1113         * Reserve log space if we might write beyond the on-disk inode size.
1114         */
1115        err = 0;
1116        if (ioend->io_type != XFS_IO_UNWRITTEN && xfs_ioend_is_append(ioend))
1117                err = xfs_setfilesize_trans_alloc(ioend);
1118
1119        xfs_submit_ioend(wbc, iohead, err);
1120
1121        return 0;
1122
1123error:
1124        if (iohead)
1125                xfs_cancel_ioend(iohead);
1126
1127        if (err == -EAGAIN)
1128                goto redirty;
1129
1130        xfs_aops_discard_page(page);
1131        ClearPageUptodate(page);
1132        unlock_page(page);
1133        return err;
1134
1135redirty:
1136        redirty_page_for_writepage(wbc, page);
1137        unlock_page(page);
1138        return 0;
1139}
1140
1141STATIC int
1142xfs_vm_writepages(
1143        struct address_space    *mapping,
1144        struct writeback_control *wbc)
1145{
1146        xfs_iflags_clear(XFS_I(mapping->host), XFS_ITRUNCATED);
1147        return generic_writepages(mapping, wbc);
1148}
1149
1150/*
1151 * Called to move a page into cleanable state - and from there
1152 * to be released. The page should already be clean. We always
1153 * have buffer heads in this call.
1154 *
1155 * Returns 1 if the page is ok to release, 0 otherwise.
1156 */
1157STATIC int
1158xfs_vm_releasepage(
1159        struct page             *page,
1160        gfp_t                   gfp_mask)
1161{
1162        int                     delalloc, unwritten;
1163
1164        trace_xfs_releasepage(page->mapping->host, page, 0);
1165
1166        xfs_count_page_state(page, &delalloc, &unwritten);
1167
1168        if (WARN_ON(delalloc))
1169                return 0;
1170        if (WARN_ON(unwritten))
1171                return 0;
1172
1173        return try_to_free_buffers(page);
1174}
1175
1176STATIC int
1177__xfs_get_blocks(
1178        struct inode            *inode,
1179        sector_t                iblock,
1180        struct buffer_head      *bh_result,
1181        int                     create,
1182        int                     direct)
1183{
1184        struct xfs_inode        *ip = XFS_I(inode);
1185        struct xfs_mount        *mp = ip->i_mount;
1186        xfs_fileoff_t           offset_fsb, end_fsb;
1187        int                     error = 0;
1188        int                     lockmode = 0;
1189        struct xfs_bmbt_irec    imap;
1190        int                     nimaps = 1;
1191        xfs_off_t               offset;
1192        ssize_t                 size;
1193        int                     new = 0;
1194
1195        if (XFS_FORCED_SHUTDOWN(mp))
1196                return -XFS_ERROR(EIO);
1197
1198        offset = (xfs_off_t)iblock << inode->i_blkbits;
1199        ASSERT(bh_result->b_size >= (1 << inode->i_blkbits));
1200        size = bh_result->b_size;
1201
1202        if (!create && direct && offset >= i_size_read(inode))
1203                return 0;
1204
1205        /*
1206         * Direct I/O is usually done on preallocated files, so try getting
1207         * a block mapping without an exclusive lock first.  For buffered
1208         * writes we already have the exclusive iolock anyway, so avoiding
1209         * a lock roundtrip here by taking the ilock exclusive from the
1210         * beginning is a useful micro optimization.
1211         */
1212        if (create && !direct) {
1213                lockmode = XFS_ILOCK_EXCL;
1214                xfs_ilock(ip, lockmode);
1215        } else {
1216                lockmode = xfs_ilock_map_shared(ip);
1217        }
1218
1219        ASSERT(offset <= mp->m_super->s_maxbytes);
1220        if (offset + size > mp->m_super->s_maxbytes)
1221                size = mp->m_super->s_maxbytes - offset;
1222        end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + size);
1223        offset_fsb = XFS_B_TO_FSBT(mp, offset);
1224
1225        error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb,
1226                                &imap, &nimaps, XFS_BMAPI_ENTIRE);
1227        if (error)
1228                goto out_unlock;
1229
1230        if (create &&
1231            (!nimaps ||
1232             (imap.br_startblock == HOLESTARTBLOCK ||
1233              imap.br_startblock == DELAYSTARTBLOCK))) {
1234                if (direct || xfs_get_extsz_hint(ip)) {
1235                        /*
1236                         * Drop the ilock in preparation for starting the block
1237                         * allocation transaction.  It will be retaken
1238                         * exclusively inside xfs_iomap_write_direct for the
1239                         * actual allocation.
1240                         */
1241                        xfs_iunlock(ip, lockmode);
1242                        error = xfs_iomap_write_direct(ip, offset, size,
1243                                                       &imap, nimaps);
1244                        if (error)
1245                                return -error;
1246                        new = 1;
1247                } else {
1248                        /*
1249                         * Delalloc reservations do not require a transaction,
1250                         * we can go on without dropping the lock here. If we
1251                         * are allocating a new delalloc block, make sure that
1252                         * we set the new flag so that we mark the buffer new so
1253                         * that we know that it is newly allocated if the write
1254                         * fails.
1255                         */
1256                        if (nimaps && imap.br_startblock == HOLESTARTBLOCK)
1257                                new = 1;
1258                        error = xfs_iomap_write_delay(ip, offset, size, &imap);
1259                        if (error)
1260                                goto out_unlock;
1261
1262                        xfs_iunlock(ip, lockmode);
1263                }
1264
1265                trace_xfs_get_blocks_alloc(ip, offset, size, 0, &imap);
1266        } else if (nimaps) {
1267                trace_xfs_get_blocks_found(ip, offset, size, 0, &imap);
1268                xfs_iunlock(ip, lockmode);
1269        } else {
1270                trace_xfs_get_blocks_notfound(ip, offset, size);
1271                goto out_unlock;
1272        }
1273
1274        if (imap.br_startblock != HOLESTARTBLOCK &&
1275            imap.br_startblock != DELAYSTARTBLOCK) {
1276                /*
1277                 * For unwritten extents do not report a disk address on
1278                 * the read case (treat as if we're reading into a hole).
1279                 */
1280                if (create || !ISUNWRITTEN(&imap))
1281                        xfs_map_buffer(inode, bh_result, &imap, offset);
1282                if (create && ISUNWRITTEN(&imap)) {
1283                        if (direct)
1284                                bh_result->b_private = inode;
1285                        set_buffer_unwritten(bh_result);
1286                }
1287        }
1288
1289        /*
1290         * If this is a realtime file, data may be on a different device.
1291         * to that pointed to from the buffer_head b_bdev currently.
1292         */
1293        bh_result->b_bdev = xfs_find_bdev_for_inode(inode);
1294
1295        /*
1296         * If we previously allocated a block out beyond eof and we are now
1297         * coming back to use it then we will need to flag it as new even if it
1298         * has a disk address.
1299         *
1300         * With sub-block writes into unwritten extents we also need to mark
1301         * the buffer as new so that the unwritten parts of the buffer gets
1302         * correctly zeroed.
1303         */
1304        if (create &&
1305            ((!buffer_mapped(bh_result) && !buffer_uptodate(bh_result)) ||
1306             (offset >= i_size_read(inode)) ||
1307             (new || ISUNWRITTEN(&imap))))
1308                set_buffer_new(bh_result);
1309
1310        if (imap.br_startblock == DELAYSTARTBLOCK) {
1311                BUG_ON(direct);
1312                if (create) {
1313                        set_buffer_uptodate(bh_result);
1314                        set_buffer_mapped(bh_result);
1315                        set_buffer_delay(bh_result);
1316                }
1317        }
1318
1319        /*
1320         * If this is O_DIRECT or the mpage code calling tell them how large
1321         * the mapping is, so that we can avoid repeated get_blocks calls.
1322         */
1323        if (direct || size > (1 << inode->i_blkbits)) {
1324                xfs_off_t               mapping_size;
1325
1326                mapping_size = imap.br_startoff + imap.br_blockcount - iblock;
1327                mapping_size <<= inode->i_blkbits;
1328
1329                ASSERT(mapping_size > 0);
1330                if (mapping_size > size)
1331                        mapping_size = size;
1332                if (mapping_size > LONG_MAX)
1333                        mapping_size = LONG_MAX;
1334
1335                bh_result->b_size = mapping_size;
1336        }
1337
1338        return 0;
1339
1340out_unlock:
1341        xfs_iunlock(ip, lockmode);
1342        return -error;
1343}
1344
1345int
1346xfs_get_blocks(
1347        struct inode            *inode,
1348        sector_t                iblock,
1349        struct buffer_head      *bh_result,
1350        int                     create)
1351{
1352        return __xfs_get_blocks(inode, iblock, bh_result, create, 0);
1353}
1354
1355STATIC int
1356xfs_get_blocks_direct(
1357        struct inode            *inode,
1358        sector_t                iblock,
1359        struct buffer_head      *bh_result,
1360        int                     create)
1361{
1362        return __xfs_get_blocks(inode, iblock, bh_result, create, 1);
1363}
1364
1365/*
1366 * Complete a direct I/O write request.
1367 *
1368 * If the private argument is non-NULL __xfs_get_blocks signals us that we
1369 * need to issue a transaction to convert the range from unwritten to written
1370 * extents.  In case this is regular synchronous I/O we just call xfs_end_io
1371 * to do this and we are done.  But in case this was a successful AIO
1372 * request this handler is called from interrupt context, from which we
1373 * can't start transactions.  In that case offload the I/O completion to
1374 * the workqueues we also use for buffered I/O completion.
1375 */
1376STATIC void
1377xfs_end_io_direct_write(
1378        struct kiocb            *iocb,
1379        loff_t                  offset,
1380        ssize_t                 size,
1381        void                    *private,
1382        int                     ret,
1383        bool                    is_async)
1384{
1385        struct xfs_ioend        *ioend = iocb->private;
1386
1387        /*
1388         * While the generic direct I/O code updates the inode size, it does
1389         * so only after the end_io handler is called, which means our
1390         * end_io handler thinks the on-disk size is outside the in-core
1391         * size.  To prevent this just update it a little bit earlier here.
1392         */
1393        if (offset + size > i_size_read(ioend->io_inode))
1394                i_size_write(ioend->io_inode, offset + size);
1395
1396        /*
1397         * blockdev_direct_IO can return an error even after the I/O
1398         * completion handler was called.  Thus we need to protect
1399         * against double-freeing.
1400         */
1401        iocb->private = NULL;
1402
1403        ioend->io_offset = offset;
1404        ioend->io_size = size;
1405        ioend->io_iocb = iocb;
1406        ioend->io_result = ret;
1407        if (private && size > 0)
1408                ioend->io_type = XFS_IO_UNWRITTEN;
1409
1410        if (is_async) {
1411                ioend->io_isasync = 1;
1412                xfs_finish_ioend(ioend);
1413        } else {
1414                xfs_finish_ioend_sync(ioend);
1415        }
1416}
1417
1418STATIC ssize_t
1419xfs_vm_direct_IO(
1420        int                     rw,
1421        struct kiocb            *iocb,
1422        const struct iovec      *iov,
1423        loff_t                  offset,
1424        unsigned long           nr_segs)
1425{
1426        struct inode            *inode = iocb->ki_filp->f_mapping->host;
1427        struct block_device     *bdev = xfs_find_bdev_for_inode(inode);
1428        struct xfs_ioend        *ioend = NULL;
1429        ssize_t                 ret;
1430
1431        if (rw & WRITE) {
1432                size_t size = iov_length(iov, nr_segs);
1433
1434                /*
1435                 * We need to preallocate a transaction for a size update
1436                 * here.  In the case that this write both updates the size
1437                 * and converts at least on unwritten extent we will cancel
1438                 * the still clean transaction after the I/O has finished.
1439                 */
1440                iocb->private = ioend = xfs_alloc_ioend(inode, XFS_IO_DIRECT);
1441                if (offset + size > XFS_I(inode)->i_d.di_size) {
1442                        ret = xfs_setfilesize_trans_alloc(ioend);
1443                        if (ret)
1444                                goto out_destroy_ioend;
1445                        ioend->io_isdirect = 1;
1446                }
1447
1448                ret = __blockdev_direct_IO(rw, iocb, inode, bdev, iov,
1449                                            offset, nr_segs,
1450                                            xfs_get_blocks_direct,
1451                                            xfs_end_io_direct_write, NULL, 0);
1452                if (ret != -EIOCBQUEUED && iocb->private)
1453                        goto out_trans_cancel;
1454        } else {
1455                ret = __blockdev_direct_IO(rw, iocb, inode, bdev, iov,
1456                                            offset, nr_segs,
1457                                            xfs_get_blocks_direct,
1458                                            NULL, NULL, 0);
1459        }
1460
1461        return ret;
1462
1463out_trans_cancel:
1464        if (ioend->io_append_trans) {
1465                current_set_flags_nested(&ioend->io_append_trans->t_pflags,
1466                                         PF_FSTRANS);
1467                rwsem_acquire_read(
1468                        &inode->i_sb->s_writers.lock_map[SB_FREEZE_FS-1],
1469                        0, 1, _THIS_IP_);
1470                xfs_trans_cancel(ioend->io_append_trans, 0);
1471        }
1472out_destroy_ioend:
1473        xfs_destroy_ioend(ioend);
1474        return ret;
1475}
1476
1477/*
1478 * Punch out the delalloc blocks we have already allocated.
1479 *
1480 * Don't bother with xfs_setattr given that nothing can have made it to disk yet
1481 * as the page is still locked at this point.
1482 */
1483STATIC void
1484xfs_vm_kill_delalloc_range(
1485        struct inode            *inode,
1486        loff_t                  start,
1487        loff_t                  end)
1488{
1489        struct xfs_inode        *ip = XFS_I(inode);
1490        xfs_fileoff_t           start_fsb;
1491        xfs_fileoff_t           end_fsb;
1492        int                     error;
1493
1494        start_fsb = XFS_B_TO_FSB(ip->i_mount, start);
1495        end_fsb = XFS_B_TO_FSB(ip->i_mount, end);
1496        if (end_fsb <= start_fsb)
1497                return;
1498
1499        xfs_ilock(ip, XFS_ILOCK_EXCL);
1500        error = xfs_bmap_punch_delalloc_range(ip, start_fsb,
1501                                                end_fsb - start_fsb);
1502        if (error) {
1503                /* something screwed, just bail */
1504                if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
1505                        xfs_alert(ip->i_mount,
1506                "xfs_vm_write_failed: unable to clean up ino %lld",
1507                                        ip->i_ino);
1508                }
1509        }
1510        xfs_iunlock(ip, XFS_ILOCK_EXCL);
1511}
1512
1513STATIC void
1514xfs_vm_write_failed(
1515        struct inode            *inode,
1516        struct page             *page,
1517        loff_t                  pos,
1518        unsigned                len)
1519{
1520        loff_t                  block_offset = pos & PAGE_MASK;
1521        loff_t                  block_start;
1522        loff_t                  block_end;
1523        loff_t                  from = pos & (PAGE_CACHE_SIZE - 1);
1524        loff_t                  to = from + len;
1525        struct buffer_head      *bh, *head;
1526
1527        ASSERT(block_offset + from == pos);
1528
1529        head = page_buffers(page);
1530        block_start = 0;
1531        for (bh = head; bh != head || !block_start;
1532             bh = bh->b_this_page, block_start = block_end,
1533                                   block_offset += bh->b_size) {
1534                block_end = block_start + bh->b_size;
1535
1536                /* skip buffers before the write */
1537                if (block_end <= from)
1538                        continue;
1539
1540                /* if the buffer is after the write, we're done */
1541                if (block_start >= to)
1542                        break;
1543
1544                if (!buffer_delay(bh))
1545                        continue;
1546
1547                if (!buffer_new(bh) && block_offset < i_size_read(inode))
1548                        continue;
1549
1550                xfs_vm_kill_delalloc_range(inode, block_offset,
1551                                           block_offset + bh->b_size);
1552        }
1553
1554}
1555
1556/*
1557 * This used to call block_write_begin(), but it unlocks and releases the page
1558 * on error, and we need that page to be able to punch stale delalloc blocks out
1559 * on failure. hence we copy-n-waste it here and call xfs_vm_write_failed() at
1560 * the appropriate point.
1561 */
1562STATIC int
1563xfs_vm_write_begin(
1564        struct file             *file,
1565        struct address_space    *mapping,
1566        loff_t                  pos,
1567        unsigned                len,
1568        unsigned                flags,
1569        struct page             **pagep,
1570        void                    **fsdata)
1571{
1572        pgoff_t                 index = pos >> PAGE_CACHE_SHIFT;
1573        struct page             *page;
1574        int                     status;
1575
1576        ASSERT(len <= PAGE_CACHE_SIZE);
1577
1578        page = grab_cache_page_write_begin(mapping, index,
1579                                           flags | AOP_FLAG_NOFS);
1580        if (!page)
1581                return -ENOMEM;
1582
1583        status = __block_write_begin(page, pos, len, xfs_get_blocks);
1584        if (unlikely(status)) {
1585                struct inode    *inode = mapping->host;
1586
1587                xfs_vm_write_failed(inode, page, pos, len);
1588                unlock_page(page);
1589
1590                if (pos + len > i_size_read(inode))
1591                        truncate_pagecache(inode, pos + len, i_size_read(inode));
1592
1593                page_cache_release(page);
1594                page = NULL;
1595        }
1596
1597        *pagep = page;
1598        return status;
1599}
1600
1601/*
1602 * On failure, we only need to kill delalloc blocks beyond EOF because they
1603 * will never be written. For blocks within EOF, generic_write_end() zeros them
1604 * so they are safe to leave alone and be written with all the other valid data.
1605 */
1606STATIC int
1607xfs_vm_write_end(
1608        struct file             *file,
1609        struct address_space    *mapping,
1610        loff_t                  pos,
1611        unsigned                len,
1612        unsigned                copied,
1613        struct page             *page,
1614        void                    *fsdata)
1615{
1616        int                     ret;
1617
1618        ASSERT(len <= PAGE_CACHE_SIZE);
1619
1620        ret = generic_write_end(file, mapping, pos, len, copied, page, fsdata);
1621        if (unlikely(ret < len)) {
1622                struct inode    *inode = mapping->host;
1623                size_t          isize = i_size_read(inode);
1624                loff_t          to = pos + len;
1625
1626                if (to > isize) {
1627                        truncate_pagecache(inode, to, isize);
1628                        xfs_vm_kill_delalloc_range(inode, isize, to);
1629                }
1630        }
1631        return ret;
1632}
1633
1634STATIC sector_t
1635xfs_vm_bmap(
1636        struct address_space    *mapping,
1637        sector_t                block)
1638{
1639        struct inode            *inode = (struct inode *)mapping->host;
1640        struct xfs_inode        *ip = XFS_I(inode);
1641
1642        trace_xfs_vm_bmap(XFS_I(inode));
1643        xfs_ilock(ip, XFS_IOLOCK_SHARED);
1644        xfs_flush_pages(ip, (xfs_off_t)0, -1, 0, FI_REMAPF);
1645        xfs_iunlock(ip, XFS_IOLOCK_SHARED);
1646        return generic_block_bmap(mapping, block, xfs_get_blocks);
1647}
1648
1649STATIC int
1650xfs_vm_readpage(
1651        struct file             *unused,
1652        struct page             *page)
1653{
1654        return mpage_readpage(page, xfs_get_blocks);
1655}
1656
1657STATIC int
1658xfs_vm_readpages(
1659        struct file             *unused,
1660        struct address_space    *mapping,
1661        struct list_head        *pages,
1662        unsigned                nr_pages)
1663{
1664        return mpage_readpages(mapping, pages, nr_pages, xfs_get_blocks);
1665}
1666
1667const struct address_space_operations xfs_address_space_operations = {
1668        .readpage               = xfs_vm_readpage,
1669        .readpages              = xfs_vm_readpages,
1670        .writepage              = xfs_vm_writepage,
1671        .writepages             = xfs_vm_writepages,
1672        .releasepage            = xfs_vm_releasepage,
1673        .invalidatepage         = xfs_vm_invalidatepage,
1674        .write_begin            = xfs_vm_write_begin,
1675        .write_end              = xfs_vm_write_end,
1676        .bmap                   = xfs_vm_bmap,
1677        .direct_IO              = xfs_vm_direct_IO,
1678        .migratepage            = buffer_migrate_page,
1679        .is_partially_uptodate  = block_is_partially_uptodate,
1680        .error_remove_page      = generic_error_remove_page,
1681};
1682
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.