xfs: optimize writes to reflink files
authorChristoph Hellwig <hch@lst.de>
Thu, 20 Oct 2016 04:53:50 +0000 (15:53 +1100)
committerDave Chinner <david@fromorbit.com>
Thu, 20 Oct 2016 04:53:50 +0000 (15:53 +1100)
Instead of reserving space as the first thing in write_begin move it past
reading the extent in the data fork.  That way we only have to read from
the data fork once and can reuse that information for trimming the extent
to the shared/unshared boundary.  Additionally this allows to easily
limit the actual write size to said boundary, and avoid a roundtrip on the
ilock.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
fs/xfs/xfs_iomap.c
fs/xfs/xfs_reflink.c
fs/xfs/xfs_reflink.h
fs/xfs/xfs_trace.h

index 1dabf2e..436e109 100644 (file)
@@ -566,6 +566,17 @@ xfs_file_iomap_begin_delay(
        xfs_bmap_search_extents(ip, offset_fsb, XFS_DATA_FORK, &eof, &idx,
                        &got, &prev);
        if (!eof && got.br_startoff <= offset_fsb) {
+               if (xfs_is_reflink_inode(ip)) {
+                       bool            shared;
+
+                       end_fsb = min(XFS_B_TO_FSB(mp, offset + count),
+                                       maxbytes_fsb);
+                       xfs_trim_extent(&got, offset_fsb, end_fsb - offset_fsb);
+                       error = xfs_reflink_reserve_cow(ip, &got, &shared);
+                       if (error)
+                               goto out_unlock;
+               }
+
                trace_xfs_iomap_found(ip, offset, count, 0, &got);
                goto done;
        }
@@ -961,19 +972,13 @@ xfs_file_iomap_begin(
        struct xfs_mount        *mp = ip->i_mount;
        struct xfs_bmbt_irec    imap;
        xfs_fileoff_t           offset_fsb, end_fsb;
-       bool                    shared, trimmed;
        int                     nimaps = 1, error = 0;
+       bool                    shared = false, trimmed = false;
        unsigned                lockmode;
 
        if (XFS_FORCED_SHUTDOWN(mp))
                return -EIO;
 
-       if ((flags & (IOMAP_WRITE | IOMAP_ZERO)) && xfs_is_reflink_inode(ip)) {
-               error = xfs_reflink_reserve_cow_range(ip, offset, length);
-               if (error < 0)
-                       return error;
-       }
-
        if ((flags & IOMAP_WRITE) && !IS_DAX(inode) &&
                   !xfs_get_extsz_hint(ip)) {
                /* Reserve delalloc blocks for regular writeback. */
@@ -981,7 +986,16 @@ xfs_file_iomap_begin(
                                iomap);
        }
 
-       lockmode = xfs_ilock_data_map_shared(ip);
+       /*
+        * COW writes will allocate delalloc space, so we need to make sure
+        * to take the lock exclusively here.
+        */
+       if ((flags & (IOMAP_WRITE | IOMAP_ZERO)) && xfs_is_reflink_inode(ip)) {
+               lockmode = XFS_ILOCK_EXCL;
+               xfs_ilock(ip, XFS_ILOCK_EXCL);
+       } else {
+               lockmode = xfs_ilock_data_map_shared(ip);
+       }
 
        ASSERT(offset <= mp->m_super->s_maxbytes);
        if ((xfs_fsize_t)offset + length > mp->m_super->s_maxbytes)
@@ -991,19 +1005,24 @@ xfs_file_iomap_begin(
 
        error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb, &imap,
                               &nimaps, 0);
-       if (error) {
-               xfs_iunlock(ip, lockmode);
-               return error;
-       }
+       if (error)
+               goto out_unlock;
 
-       if (flags & (IOMAP_WRITE | IOMAP_ZERO | IOMAP_REPORT)) {
+       if (flags & IOMAP_REPORT) {
                /* Trim the mapping to the nearest shared extent boundary. */
                error = xfs_reflink_trim_around_shared(ip, &imap, &shared,
                                &trimmed);
-               if (error) {
-                       xfs_iunlock(ip, lockmode);
-                       return error;
-               }
+               if (error)
+                       goto out_unlock;
+       }
+
+       if ((flags & (IOMAP_WRITE | IOMAP_ZERO)) && xfs_is_reflink_inode(ip)) {
+               error = xfs_reflink_reserve_cow(ip, &imap, &shared);
+               if (error)
+                       goto out_unlock;
+
+               end_fsb = imap.br_startoff + imap.br_blockcount;
+               length = XFS_FSB_TO_B(mp, end_fsb) - offset;
        }
 
        if ((flags & IOMAP_WRITE) && imap_needs_alloc(inode, &imap, nimaps)) {
@@ -1042,6 +1061,9 @@ xfs_file_iomap_begin(
        if (shared)
                iomap->flags |= IOMAP_F_SHARED;
        return 0;
+out_unlock:
+       xfs_iunlock(ip, lockmode);
+       return error;
 }
 
 static int
index 6c4c215..9c477de 100644 (file)
@@ -228,50 +228,54 @@ xfs_reflink_trim_around_shared(
        }
 }
 
-/* Create a CoW reservation for a range of blocks within a file. */
-static int
-__xfs_reflink_reserve_cow(
+/*
+ * Trim the passed in imap to the next shared/unshared extent boundary, and
+ * if imap->br_startoff points to a shared extent reserve space for it in the
+ * COW fork.  In this case *shared is set to true, else to false.
+ *
+ * Note that imap will always contain the block numbers for the existing blocks
+ * in the data fork, as the upper layers need them for read-modify-write
+ * operations.
+ */
+int
+xfs_reflink_reserve_cow(
        struct xfs_inode        *ip,
-       xfs_fileoff_t           *offset_fsb,
-       xfs_fileoff_t           end_fsb,
-       bool                    *skipped)
+       struct xfs_bmbt_irec    *imap,
+       bool                    *shared)
 {
-       struct xfs_bmbt_irec    got, prev, imap;
-       xfs_fileoff_t           orig_end_fsb;
-       int                     nimaps, eof = 0, error = 0;
-       bool                    shared = false, trimmed = false;
+       struct xfs_bmbt_irec    got, prev;
+       xfs_fileoff_t           end_fsb, orig_end_fsb;
+       int                     eof = 0, error = 0;
+       bool                    trimmed;
        xfs_extnum_t            idx;
        xfs_extlen_t            align;
 
-       /* Already reserved?  Skip the refcount btree access. */
-       xfs_bmap_search_extents(ip, *offset_fsb, XFS_COW_FORK, &eof, &idx,
+       /*
+        * Search the COW fork extent list first.  This serves two purposes:
+        * first this implement the speculative preallocation using cowextisze,
+        * so that we also unshared block adjacent to shared blocks instead
+        * of just the shared blocks themselves.  Second the lookup in the
+        * extent list is generally faster than going out to the shared extent
+        * tree.
+        */
+       xfs_bmap_search_extents(ip, imap->br_startoff, XFS_COW_FORK, &eof, &idx,
                        &got, &prev);
-       if (!eof && got.br_startoff <= *offset_fsb) {
-               end_fsb = orig_end_fsb = got.br_startoff + got.br_blockcount;
-               trace_xfs_reflink_cow_found(ip, &got);
-               goto done;
-       }
+       if (!eof && got.br_startoff <= imap->br_startoff) {
+               trace_xfs_reflink_cow_found(ip, imap);
+               xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
 
-       /* Read extent from the source file. */
-       nimaps = 1;
-       error = xfs_bmapi_read(ip, *offset_fsb, end_fsb - *offset_fsb,
-                       &imap, &nimaps, 0);
-       if (error)
-               goto out_unlock;
-       ASSERT(nimaps == 1);
+               *shared = true;
+               return 0;
+       }
 
        /* Trim the mapping to the nearest shared extent boundary. */
-       error = xfs_reflink_trim_around_shared(ip, &imap, &shared, &trimmed);
+       error = xfs_reflink_trim_around_shared(ip, imap, shared, &trimmed);
        if (error)
-               goto out_unlock;
-
-       end_fsb = orig_end_fsb = imap.br_startoff + imap.br_blockcount;
+               return error;
 
        /* Not shared?  Just report the (potentially capped) extent. */
-       if (!shared) {
-               *skipped = true;
-               goto done;
-       }
+       if (!*shared)
+               return 0;
 
        /*
         * Fork all the shared blocks from our write offset until the end of
@@ -279,72 +283,38 @@ __xfs_reflink_reserve_cow(
         */
        error = xfs_qm_dqattach_locked(ip, 0);
        if (error)
-               goto out_unlock;
+               return error;
+
+       end_fsb = orig_end_fsb = imap->br_startoff + imap->br_blockcount;
 
        align = xfs_eof_alignment(ip, xfs_get_cowextsz_hint(ip));
        if (align)
                end_fsb = roundup_64(end_fsb, align);
 
 retry:
-       error = xfs_bmapi_reserve_delalloc(ip, XFS_COW_FORK, *offset_fsb,
-                       end_fsb - *offset_fsb, &got,
-                       &prev, &idx, eof);
+       error = xfs_bmapi_reserve_delalloc(ip, XFS_COW_FORK, imap->br_startoff,
+                       end_fsb - imap->br_startoff, &got, &prev, &idx, eof);
        switch (error) {
        case 0:
                break;
        case -ENOSPC:
        case -EDQUOT:
                /* retry without any preallocation */
-               trace_xfs_reflink_cow_enospc(ip, &imap);
+               trace_xfs_reflink_cow_enospc(ip, imap);
                if (end_fsb != orig_end_fsb) {
                        end_fsb = orig_end_fsb;
                        goto retry;
                }
                /*FALLTHRU*/
        default:
-               goto out_unlock;
+               return error;
        }
 
        if (end_fsb != orig_end_fsb)
                xfs_inode_set_cowblocks_tag(ip);
 
        trace_xfs_reflink_cow_alloc(ip, &got);
-done:
-       *offset_fsb = end_fsb;
-out_unlock:
-       return error;
-}
-
-/* Create a CoW reservation for part of a file. */
-int
-xfs_reflink_reserve_cow_range(
-       struct xfs_inode        *ip,
-       xfs_off_t               offset,
-       xfs_off_t               count)
-{
-       struct xfs_mount        *mp = ip->i_mount;
-       xfs_fileoff_t           offset_fsb, end_fsb;
-       bool                    skipped = false;
-       int                     error = 0;
-
-       trace_xfs_reflink_reserve_cow_range(ip, offset, count);
-
-       offset_fsb = XFS_B_TO_FSBT(mp, offset);
-       end_fsb = XFS_B_TO_FSB(mp, offset + count);
-
-       xfs_ilock(ip, XFS_ILOCK_EXCL);
-       while (offset_fsb < end_fsb) {
-               error = __xfs_reflink_reserve_cow(ip, &offset_fsb, end_fsb,
-                               &skipped);
-               if (error) {
-                       trace_xfs_reflink_reserve_cow_range_error(ip, error,
-                               _RET_IP_);
-                       break;
-               }
-       }
-       xfs_iunlock(ip, XFS_ILOCK_EXCL);
-
-       return error;
+       return 0;
 }
 
 /* Allocate all CoW reservations covering a range of blocks in a file. */
@@ -359,9 +329,8 @@ __xfs_reflink_allocate_cow(
        struct xfs_defer_ops    dfops;
        struct xfs_trans        *tp;
        xfs_fsblock_t           first_block;
-       xfs_fileoff_t           next_fsb;
        int                     nimaps = 1, error;
-       bool                    skipped = false;
+       bool                    shared;
 
        xfs_defer_init(&dfops, &first_block);
 
@@ -372,33 +341,38 @@ __xfs_reflink_allocate_cow(
 
        xfs_ilock(ip, XFS_ILOCK_EXCL);
 
-       next_fsb = *offset_fsb;
-       error = __xfs_reflink_reserve_cow(ip, &next_fsb, end_fsb, &skipped);
+       /* Read extent from the source file. */
+       nimaps = 1;
+       error = xfs_bmapi_read(ip, *offset_fsb, end_fsb - *offset_fsb,
+                       &imap, &nimaps, 0);
+       if (error)
+               goto out_unlock;
+       ASSERT(nimaps == 1);
+
+       error = xfs_reflink_reserve_cow(ip, &imap, &shared);
        if (error)
                goto out_trans_cancel;
 
-       if (skipped) {
-               *offset_fsb = next_fsb;
+       if (!shared) {
+               *offset_fsb = imap.br_startoff + imap.br_blockcount;
                goto out_trans_cancel;
        }
 
        xfs_trans_ijoin(tp, ip, 0);
-       error = xfs_bmapi_write(tp, ip, *offset_fsb, next_fsb - *offset_fsb,
+       error = xfs_bmapi_write(tp, ip, imap.br_startoff, imap.br_blockcount,
                        XFS_BMAPI_COWFORK, &first_block,
                        XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK),
                        &imap, &nimaps, &dfops);
        if (error)
                goto out_trans_cancel;
 
-       /* We might not have been able to map the whole delalloc extent */
-       *offset_fsb = min(*offset_fsb + imap.br_blockcount, next_fsb);
-
        error = xfs_defer_finish(&tp, &dfops, NULL);
        if (error)
                goto out_trans_cancel;
 
        error = xfs_trans_commit(tp);
 
+       *offset_fsb = imap.br_startoff + imap.br_blockcount;
 out_unlock:
        xfs_iunlock(ip, XFS_ILOCK_EXCL);
        return error;
index 7ddd9f6..fad1160 100644 (file)
@@ -26,8 +26,8 @@ extern int xfs_reflink_find_shared(struct xfs_mount *mp, xfs_agnumber_t agno,
 extern int xfs_reflink_trim_around_shared(struct xfs_inode *ip,
                struct xfs_bmbt_irec *irec, bool *shared, bool *trimmed);
 
-extern int xfs_reflink_reserve_cow_range(struct xfs_inode *ip,
-               xfs_off_t offset, xfs_off_t count);
+extern int xfs_reflink_reserve_cow(struct xfs_inode *ip,
+               struct xfs_bmbt_irec *imap, bool *shared);
 extern int xfs_reflink_allocate_cow_range(struct xfs_inode *ip,
                xfs_off_t offset, xfs_off_t count);
 extern bool xfs_reflink_find_cow_mapping(struct xfs_inode *ip, xfs_off_t offset,
index ad188d3..72f9f6b 100644 (file)
@@ -3346,7 +3346,7 @@ DEFINE_INODE_IREC_EVENT(xfs_reflink_cow_alloc);
 DEFINE_INODE_IREC_EVENT(xfs_reflink_cow_found);
 DEFINE_INODE_IREC_EVENT(xfs_reflink_cow_enospc);
 
-DEFINE_RW_EVENT(xfs_reflink_reserve_cow_range);
+DEFINE_RW_EVENT(xfs_reflink_reserve_cow);
 DEFINE_RW_EVENT(xfs_reflink_allocate_cow_range);
 
 DEFINE_INODE_IREC_EVENT(xfs_reflink_bounce_dio_write);
@@ -3358,7 +3358,6 @@ DEFINE_SIMPLE_IO_EVENT(xfs_reflink_end_cow);
 DEFINE_INODE_IREC_EVENT(xfs_reflink_cow_remap);
 DEFINE_INODE_IREC_EVENT(xfs_reflink_cow_remap_piece);
 
-DEFINE_INODE_ERROR_EVENT(xfs_reflink_reserve_cow_range_error);
 DEFINE_INODE_ERROR_EVENT(xfs_reflink_allocate_cow_range_error);
 DEFINE_INODE_ERROR_EVENT(xfs_reflink_cancel_cow_range_error);
 DEFINE_INODE_ERROR_EVENT(xfs_reflink_end_cow_error);