diff options
Diffstat (limited to 'fs/xfs/xfs_reflink.c')
| -rw-r--r-- | fs/xfs/xfs_reflink.c | 116 |
1 files changed, 108 insertions, 8 deletions
diff --git a/fs/xfs/xfs_reflink.c b/fs/xfs/xfs_reflink.c index 07593a362cd0..8c8c4f4676da 100644 --- a/fs/xfs/xfs_reflink.c +++ b/fs/xfs/xfs_reflink.c @@ -82,11 +82,22 @@ * mappings are a reservation against the free space in the filesystem; * adjacent mappings can also be combined into fewer larger mappings. * + * As an optimization, the CoW extent size hint (cowextsz) creates + * outsized aligned delalloc reservations in the hope of landing out of + * order nearby CoW writes in a single extent on disk, thereby reducing + * fragmentation and improving future performance. + * + * D: --RRRRRRSSSRRRRRRRR--- (data fork) + * C: ------DDDDDDD--------- (CoW fork) + * * When dirty pages are being written out (typically in writepage), the - * delalloc reservations are converted into real mappings by allocating - * blocks and replacing the delalloc mapping with real ones. A delalloc - * mapping can be replaced by several real ones if the free space is - * fragmented. + * delalloc reservations are converted into unwritten mappings by + * allocating blocks and replacing the delalloc mapping with real ones. + * A delalloc mapping can be replaced by several unwritten ones if the + * free space is fragmented. + * + * D: --RRRRRRSSSRRRRRRRR--- + * C: ------UUUUUUU--------- * * We want to adapt the delalloc mechanism for copy-on-write, since the * write paths are similar. The first two steps (creating the reservation @@ -101,13 +112,29 @@ * Block-aligned directio writes will use the same mechanism as buffered * writes. * + * Just prior to submitting the actual disk write requests, we convert + * the extents representing the range of the file actually being written + * (as opposed to extra pieces created for the cowextsize hint) to real + * extents. This will become important in the next step: + * + * D: --RRRRRRSSSRRRRRRRR--- + * C: ------UUrrUUU--------- + * * CoW remapping must be done after the data block write completes, * because we don't want to destroy the old data fork map until we're sure * the new block has been written. Since the new mappings are kept in a * separate fork, we can simply iterate these mappings to find the ones * that cover the file blocks that we just CoW'd. For each extent, simply * unmap the corresponding range in the data fork, map the new range into - * the data fork, and remove the extent from the CoW fork. + * the data fork, and remove the extent from the CoW fork. Because of + * the presence of the cowextsize hint, however, we must be careful + * only to remap the blocks that we've actually written out -- we must + * never remap delalloc reservations nor CoW staging blocks that have + * yet to be written. This corresponds exactly to the real extents in + * the CoW fork: + * + * D: --RRRRRRrrSRRRRRRRR--- + * C: ------UU--UUU--------- * * Since the remapping operation can be applied to an arbitrary file * range, we record the need for the remap step as a flag in the ioend @@ -296,6 +323,65 @@ xfs_reflink_reserve_cow( return 0; } +/* Convert part of an unwritten CoW extent to a real one. */ +STATIC int +xfs_reflink_convert_cow_extent( + struct xfs_inode *ip, + struct xfs_bmbt_irec *imap, + xfs_fileoff_t offset_fsb, + xfs_filblks_t count_fsb, + struct xfs_defer_ops *dfops) +{ + struct xfs_bmbt_irec irec = *imap; + xfs_fsblock_t first_block; + int nimaps = 1; + + if (imap->br_state == XFS_EXT_NORM) + return 0; + + xfs_trim_extent(&irec, offset_fsb, count_fsb); + trace_xfs_reflink_convert_cow(ip, &irec); + if (irec.br_blockcount == 0) + return 0; + return xfs_bmapi_write(NULL, ip, irec.br_startoff, irec.br_blockcount, + XFS_BMAPI_COWFORK | XFS_BMAPI_CONVERT, &first_block, + 0, &irec, &nimaps, dfops); +} + +/* Convert all of the unwritten CoW extents in a file's range to real ones. */ +int +xfs_reflink_convert_cow( + struct xfs_inode *ip, + xfs_off_t offset, + xfs_off_t count) +{ + struct xfs_bmbt_irec got; + struct xfs_defer_ops dfops; + struct xfs_mount *mp = ip->i_mount; + struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK); + xfs_fileoff_t offset_fsb = XFS_B_TO_FSBT(mp, offset); + xfs_fileoff_t end_fsb = XFS_B_TO_FSB(mp, offset + count); + xfs_extnum_t idx; + bool found; + int error; + + xfs_ilock(ip, XFS_ILOCK_EXCL); + + /* Convert all the extents to real from unwritten. */ + for (found = xfs_iext_lookup_extent(ip, ifp, offset_fsb, &idx, &got); + found && got.br_startoff < end_fsb; + found = xfs_iext_get_extent(ifp, ++idx, &got)) { + error = xfs_reflink_convert_cow_extent(ip, &got, offset_fsb, + end_fsb - offset_fsb, &dfops); + if (error) + break; + } + + /* Finish up. */ + xfs_iunlock(ip, XFS_ILOCK_EXCL); + return error; +} + /* Allocate all CoW reservations covering a range of blocks in a file. */ static int __xfs_reflink_allocate_cow( @@ -328,6 +414,7 @@ __xfs_reflink_allocate_cow( goto out_unlock; ASSERT(nimaps == 1); + /* Make sure there's a CoW reservation for it. */ error = xfs_reflink_reserve_cow(ip, &imap, &shared); if (error) goto out_trans_cancel; @@ -337,14 +424,16 @@ __xfs_reflink_allocate_cow( goto out_trans_cancel; } + /* Allocate the entire reservation as unwritten blocks. */ xfs_trans_ijoin(tp, ip, 0); error = xfs_bmapi_write(tp, ip, imap.br_startoff, imap.br_blockcount, - XFS_BMAPI_COWFORK, &first_block, + XFS_BMAPI_COWFORK | XFS_BMAPI_PREALLOC, &first_block, XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK), &imap, &nimaps, &dfops); if (error) goto out_trans_cancel; + /* Finish up. */ error = xfs_defer_finish(&tp, &dfops, NULL); if (error) goto out_trans_cancel; @@ -389,11 +478,12 @@ xfs_reflink_allocate_cow_range( if (error) { trace_xfs_reflink_allocate_cow_range_error(ip, error, _RET_IP_); - break; + return error; } } - return error; + /* Convert the CoW extents to regular. */ + return xfs_reflink_convert_cow(ip, offset, count); } /* @@ -641,6 +731,16 @@ xfs_reflink_end_cow( ASSERT(!isnullstartblock(got.br_startblock)); + /* + * Don't remap unwritten extents; these are + * speculatively preallocated CoW extents that have been + * allocated but have not yet been involved in a write. + */ + if (got.br_state == XFS_EXT_UNWRITTEN) { + idx--; + goto next_extent; + } + /* Unmap the old blocks in the data fork. */ xfs_defer_init(&dfops, &firstfsb); rlen = del.br_blockcount; |