xfs: remove xfs_map_cow
authorChristoph Hellwig <hch@lst.de>
Thu, 12 Jul 2018 05:25:59 +0000 (22:25 -0700)
committerDarrick J. Wong <darrick.wong@oracle.com>
Thu, 12 Jul 2018 05:25:59 +0000 (22:25 -0700)
We can handle the existing cow mapping case as a special case directly
in xfs_writepage_map, and share code for allocating delalloc blocks
with regular I/O in xfs_map_blocks.  This means we need to always
call xfs_map_blocks for reflink inodes, but we can still skip most of
the work if it turns out that there is no COW mapping overlapping the
current block.

As a subtle detail we need to start caching holes in the wpc to deal
with the case of COW reservations between EOF.  But we'll need that
infrastructure later anyway, so this is no big deal.

Based on a patch from Dave Chinner.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
fs/xfs/xfs_aops.c
fs/xfs/xfs_aops.h

index 08605432c497a8c6f03169f150b4a33afcf5b81e..65454a4f4d93302739f1c5058c206f5ef645ce8f 100644 (file)
@@ -363,70 +363,107 @@ xfs_end_bio(
 
 STATIC int
 xfs_map_blocks(
+       struct xfs_writepage_ctx *wpc,
        struct inode            *inode,
-       loff_t                  offset,
-       struct xfs_bmbt_irec    *imap,
-       int                     type)
+       loff_t                  offset)
 {
        struct xfs_inode        *ip = XFS_I(inode);
        struct xfs_mount        *mp = ip->i_mount;
        ssize_t                 count = i_blocksize(inode);
-       xfs_fileoff_t           offset_fsb, end_fsb;
+       xfs_fileoff_t           offset_fsb = XFS_B_TO_FSBT(mp, offset), end_fsb;
+       struct xfs_bmbt_irec    imap;
+       int                     whichfork = XFS_DATA_FORK;
        int                     error = 0;
        int                     nimaps = 1;
 
        if (XFS_FORCED_SHUTDOWN(mp))
                return -EIO;
 
-       /*
-        * Truncate can race with writeback since writeback doesn't take the
-        * iolock and truncate decreases the file size before it starts
-        * truncating the pages between new_size and old_size.  Therefore, we
-        * can end up in the situation where writeback gets a CoW fork mapping
-        * but the truncate makes the mapping invalid and we end up in here
-        * trying to get a new mapping.  Bail out here so that we simply never
-        * get a valid mapping and so we drop the write altogether.  The page
-        * truncation will kill the contents anyway.
-        */
-       if (type == XFS_IO_COW && offset > i_size_read(inode))
-               return 0;
-
-       ASSERT(type != XFS_IO_COW);
-
        xfs_ilock(ip, XFS_ILOCK_SHARED);
        ASSERT(ip->i_d.di_format != XFS_DINODE_FMT_BTREE ||
               (ip->i_df.if_flags & XFS_IFEXTENTS));
        ASSERT(offset <= mp->m_super->s_maxbytes);
 
+       if (xfs_is_reflink_inode(ip) &&
+           xfs_reflink_find_cow_mapping(ip, offset, &imap)) {
+               xfs_iunlock(ip, XFS_ILOCK_SHARED);
+               /*
+                * Truncate can race with writeback since writeback doesn't
+                * take the iolock and truncate decreases the file size before
+                * it starts truncating the pages between new_size and old_size.
+                * Therefore, we can end up in the situation where writeback
+                * gets a CoW fork mapping but the truncate makes the mapping
+                * invalid and we end up in here trying to get a new mapping.
+                * bail out here so that we simply never get a valid mapping
+                * and so we drop the write altogether.  The page truncation
+                * will kill the contents anyway.
+                */
+               if (offset > i_size_read(inode)) {
+                       wpc->io_type = XFS_IO_HOLE;
+                       return 0;
+               }
+               whichfork = XFS_COW_FORK;
+               wpc->io_type = XFS_IO_COW;
+               goto allocate_blocks;
+       }
+
+       /*
+        * Map valid and no COW extent in the way?  We're done.
+        */
+       if (wpc->imap_valid) {
+               xfs_iunlock(ip, XFS_ILOCK_SHARED);
+               return 0;
+       }
+
+       /*
+        * If we don't have a valid map, now it's time to get a new one for this
+        * offset.  This will convert delayed allocations (including COW ones)
+        * into real extents.
+        */
        if (offset > mp->m_super->s_maxbytes - count)
                count = mp->m_super->s_maxbytes - offset;
        end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + count);
        offset_fsb = XFS_B_TO_FSBT(mp, offset);
        error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb,
-                               imap, &nimaps, XFS_BMAPI_ENTIRE);
+                               &imap, &nimaps, XFS_BMAPI_ENTIRE);
        xfs_iunlock(ip, XFS_ILOCK_SHARED);
-
        if (error)
                return error;
 
-       if (type == XFS_IO_DELALLOC &&
-           (!nimaps || isnullstartblock(imap->br_startblock))) {
-               error = xfs_iomap_write_allocate(ip, XFS_DATA_FORK, offset,
-                               imap);
-               if (!error)
-                       trace_xfs_map_blocks_alloc(ip, offset, count, type, imap);
-               return error;
+       if (!nimaps) {
+               /*
+                * Lookup returns no match? Beyond eof? regardless,
+                * return it as a hole so we don't write it
+                */
+               imap.br_startoff = offset_fsb;
+               imap.br_blockcount = end_fsb - offset_fsb;
+               imap.br_startblock = HOLESTARTBLOCK;
+               wpc->io_type = XFS_IO_HOLE;
+       } else if (imap.br_startblock == HOLESTARTBLOCK) {
+               /* landed in a hole */
+               wpc->io_type = XFS_IO_HOLE;
        }
 
+       if (wpc->io_type == XFS_IO_DELALLOC &&
+           (!nimaps || isnullstartblock(imap.br_startblock)))
+               goto allocate_blocks;
+
 #ifdef DEBUG
-       if (type == XFS_IO_UNWRITTEN) {
+       if (wpc->io_type == XFS_IO_UNWRITTEN) {
                ASSERT(nimaps);
-               ASSERT(imap->br_startblock != HOLESTARTBLOCK);
-               ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
+               ASSERT(imap.br_startblock != HOLESTARTBLOCK);
+               ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
        }
 #endif
-       if (nimaps)
-               trace_xfs_map_blocks_found(ip, offset, count, type, imap);
+       wpc->imap = imap;
+       trace_xfs_map_blocks_found(ip, offset, count, wpc->io_type, &imap);
+       return 0;
+allocate_blocks:
+       error = xfs_iomap_write_allocate(ip, whichfork, offset, &imap);
+       if (error)
+               return error;
+       wpc->imap = imap;
+       trace_xfs_map_blocks_alloc(ip, offset, count, wpc->io_type, &imap);
        return 0;
 }
 
@@ -758,56 +795,6 @@ out_invalidate:
        xfs_vm_invalidatepage(page, 0, PAGE_SIZE);
 }
 
-static int
-xfs_map_cow(
-       struct xfs_writepage_ctx *wpc,
-       struct inode            *inode,
-       loff_t                  offset,
-       unsigned int            *new_type)
-{
-       struct xfs_inode        *ip = XFS_I(inode);
-       struct xfs_bmbt_irec    imap;
-       bool                    is_cow = false;
-       int                     error;
-
-       /*
-        * If we already have a valid COW mapping keep using it.
-        */
-       if (wpc->io_type == XFS_IO_COW) {
-               wpc->imap_valid = xfs_imap_valid(inode, &wpc->imap, offset);
-               if (wpc->imap_valid) {
-                       *new_type = XFS_IO_COW;
-                       return 0;
-               }
-       }
-
-       /*
-        * Else we need to check if there is a COW mapping at this offset.
-        */
-       xfs_ilock(ip, XFS_ILOCK_SHARED);
-       is_cow = xfs_reflink_find_cow_mapping(ip, offset, &imap);
-       xfs_iunlock(ip, XFS_ILOCK_SHARED);
-
-       if (!is_cow)
-               return 0;
-
-       /*
-        * And if the COW mapping has a delayed extent here we need to
-        * allocate real space for it now.
-        */
-       if (isnullstartblock(imap.br_startblock)) {
-               error = xfs_iomap_write_allocate(ip, XFS_COW_FORK, offset,
-                               &imap);
-               if (error)
-                       return error;
-       }
-
-       wpc->io_type = *new_type = XFS_IO_COW;
-       wpc->imap_valid = true;
-       wpc->imap = imap;
-       return 0;
-}
-
 /*
  * We implement an immediate ioend submission policy here to avoid needing to
  * chain multiple ioends and hence nest mempool allocations which can violate
@@ -836,7 +823,7 @@ xfs_writepage_map(
        struct xfs_ioend        *ioend, *next;
        struct buffer_head      *bh, *head;
        ssize_t                 len = i_blocksize(inode);
-       uint64_t                offset;
+       uint64_t                offset; /* file offset of page */
        int                     error = 0;
        int                     count = 0;
        unsigned int            new_type;
@@ -872,10 +859,13 @@ xfs_writepage_map(
                        continue;
                }
 
-               if (xfs_is_reflink_inode(XFS_I(inode))) {
-                       error = xfs_map_cow(wpc, inode, offset, &new_type);
-                       if (error)
-                               goto out;
+               /*
+                * If we already have a valid COW mapping keep using it.
+                */
+               if (wpc->io_type == XFS_IO_COW &&
+                   xfs_imap_valid(inode, &wpc->imap, offset)) {
+                       wpc->imap_valid = true;
+                       new_type = XFS_IO_COW;
                }
 
                if (wpc->io_type != new_type) {
@@ -886,22 +876,31 @@ xfs_writepage_map(
                if (wpc->imap_valid)
                        wpc->imap_valid = xfs_imap_valid(inode, &wpc->imap,
                                                         offset);
-               if (!wpc->imap_valid) {
-                       error = xfs_map_blocks(inode, offset, &wpc->imap,
-                                            wpc->io_type);
+
+               /*
+                * COW fork blocks can overlap data fork blocks even if the
+                * blocks aren't shared. COW I/O always takes precedent, so we
+                * must always check for overlap on reflink inodes unless the
+                * mapping is already a COW one.
+                */
+               if (!wpc->imap_valid ||
+                   (xfs_is_reflink_inode(XFS_I(inode)) &&
+                    wpc->io_type != XFS_IO_COW)) {
+                       error = xfs_map_blocks(wpc, inode, offset);
                        if (error)
                                goto out;
                        wpc->imap_valid = xfs_imap_valid(inode, &wpc->imap,
                                                         offset);
                }
-               if (wpc->imap_valid) {
-                       lock_buffer(bh);
-                       if (wpc->io_type != XFS_IO_OVERWRITE)
-                               xfs_map_at_offset(inode, bh, &wpc->imap, offset);
-                       xfs_add_to_ioend(inode, bh, offset, wpc, wbc, &submit_list);
-                       count++;
-               }
 
+               if (!wpc->imap_valid || wpc->io_type == XFS_IO_HOLE)
+                       continue;
+
+               lock_buffer(bh);
+               if (wpc->io_type != XFS_IO_OVERWRITE)
+                       xfs_map_at_offset(inode, bh, &wpc->imap, offset);
+               xfs_add_to_ioend(inode, bh, offset, wpc, wbc, &submit_list);
+               count++;
        } while (offset += len, ((bh = bh->b_this_page) != head));
 
        ASSERT(wpc->ioend || list_empty(&submit_list));
index 25bc6d4a12313e7b056867e894eac8df063079de..9af867951a1077a2fdab44badb8bd743ef75b2b7 100644 (file)
@@ -17,6 +17,7 @@ enum {
        XFS_IO_UNWRITTEN,       /* covers allocated but uninitialized data */
        XFS_IO_OVERWRITE,       /* covers already allocated extent */
        XFS_IO_COW,             /* covers copy-on-write extent */
+       XFS_IO_HOLE,            /* covers region without any block allocation */
 };
 
 #define XFS_IO_TYPES \
@@ -24,7 +25,8 @@ enum {
        { XFS_IO_DELALLOC,              "delalloc" }, \
        { XFS_IO_UNWRITTEN,             "unwritten" }, \
        { XFS_IO_OVERWRITE,             "overwrite" }, \
-       { XFS_IO_COW,                   "CoW" }
+       { XFS_IO_COW,                   "CoW" }, \
+       { XFS_IO_HOLE,                  "hole" }
 
 /*
  * Structure for buffered I/O completions.