xfs: move mappings from cow fork to data fork after copy-write
authorDarrick J. Wong <darrick.wong@oracle.com>
Mon, 3 Oct 2016 16:11:35 +0000 (09:11 -0700)
committerDarrick J. Wong <darrick.wong@oracle.com>
Wed, 5 Oct 2016 20:55:40 +0000 (13:55 -0700)
After the write component of a copy-write operation finishes, clean up
the bookkeeping left behind.  On error, we simply free the new blocks
and pass the error up.  If we succeed, however, then we must remove
the old data fork mapping and move the cow fork mapping to the data
fork.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
[hch: Call the CoW failure function during xfs_cancel_ioend]
Signed-off-by: Christoph Hellwig <hch@lst.de>
fs/xfs/xfs_aops.c
fs/xfs/xfs_reflink.c
fs/xfs/xfs_reflink.h

index c76d7a6..f6f5bbc 100644 (file)
@@ -287,6 +287,25 @@ xfs_end_io(
        if (XFS_FORCED_SHUTDOWN(ip->i_mount))
                error = -EIO;
 
+       /*
+        * For a CoW extent, we need to move the mapping from the CoW fork
+        * to the data fork.  If instead an error happened, just dump the
+        * new blocks.
+        */
+       if (ioend->io_type == XFS_IO_COW) {
+               if (error)
+                       goto done;
+               if (ioend->io_bio->bi_error) {
+                       error = xfs_reflink_cancel_cow_range(ip,
+                                       ioend->io_offset, ioend->io_size);
+                       goto done;
+               }
+               error = xfs_reflink_end_cow(ip, ioend->io_offset,
+                               ioend->io_size);
+               if (error)
+                       goto done;
+       }
+
        /*
         * For unwritten extents we need to issue transactions to convert a
         * range to normal written extens after the data I/O has finished.
@@ -302,7 +321,8 @@ xfs_end_io(
        } else if (ioend->io_append_trans) {
                error = xfs_setfilesize_ioend(ioend, error);
        } else {
-               ASSERT(!xfs_ioend_is_append(ioend));
+               ASSERT(!xfs_ioend_is_append(ioend) ||
+                      ioend->io_type == XFS_IO_COW);
        }
 
 done:
@@ -316,7 +336,7 @@ xfs_end_bio(
        struct xfs_ioend        *ioend = bio->bi_private;
        struct xfs_mount        *mp = XFS_I(ioend->io_inode)->i_mount;
 
-       if (ioend->io_type == XFS_IO_UNWRITTEN)
+       if (ioend->io_type == XFS_IO_UNWRITTEN || ioend->io_type == XFS_IO_COW)
                queue_work(mp->m_unwritten_workqueue, &ioend->io_work);
        else if (ioend->io_append_trans)
                queue_work(mp->m_data_workqueue, &ioend->io_work);
index d953df3..8c3211c 100644 (file)
@@ -52,6 +52,7 @@
 #include "xfs_bmap_btree.h"
 #include "xfs_reflink.h"
 #include "xfs_iomap.h"
+#include "xfs_rmap_btree.h"
 
 /*
  * Copy on Write of Shared Blocks
@@ -412,3 +413,244 @@ xfs_reflink_trim_irec_to_next_cow(
 
        return 0;
 }
+
+/*
+ * Cancel all pending CoW reservations for some block range of an inode.
+ */
+int
+xfs_reflink_cancel_cow_blocks(
+       struct xfs_inode                *ip,
+       struct xfs_trans                **tpp,
+       xfs_fileoff_t                   offset_fsb,
+       xfs_fileoff_t                   end_fsb)
+{
+       struct xfs_bmbt_irec            irec;
+       xfs_filblks_t                   count_fsb;
+       xfs_fsblock_t                   firstfsb;
+       struct xfs_defer_ops            dfops;
+       int                             error = 0;
+       int                             nimaps;
+
+       if (!xfs_is_reflink_inode(ip))
+               return 0;
+
+       /* Go find the old extent in the CoW fork. */
+       while (offset_fsb < end_fsb) {
+               nimaps = 1;
+               count_fsb = (xfs_filblks_t)(end_fsb - offset_fsb);
+               error = xfs_bmapi_read(ip, offset_fsb, count_fsb, &irec,
+                               &nimaps, XFS_BMAPI_COWFORK);
+               if (error)
+                       break;
+               ASSERT(nimaps == 1);
+
+               trace_xfs_reflink_cancel_cow(ip, &irec);
+
+               if (irec.br_startblock == DELAYSTARTBLOCK) {
+                       /* Free a delayed allocation. */
+                       xfs_mod_fdblocks(ip->i_mount, irec.br_blockcount,
+                                       false);
+                       ip->i_delayed_blks -= irec.br_blockcount;
+
+                       /* Remove the mapping from the CoW fork. */
+                       error = xfs_bunmapi_cow(ip, &irec);
+                       if (error)
+                               break;
+               } else if (irec.br_startblock == HOLESTARTBLOCK) {
+                       /* empty */
+               } else {
+                       xfs_trans_ijoin(*tpp, ip, 0);
+                       xfs_defer_init(&dfops, &firstfsb);
+
+                       xfs_bmap_add_free(ip->i_mount, &dfops,
+                                       irec.br_startblock, irec.br_blockcount,
+                                       NULL);
+
+                       /* Update quota accounting */
+                       xfs_trans_mod_dquot_byino(*tpp, ip, XFS_TRANS_DQ_BCOUNT,
+                                       -(long)irec.br_blockcount);
+
+                       /* Roll the transaction */
+                       error = xfs_defer_finish(tpp, &dfops, ip);
+                       if (error) {
+                               xfs_defer_cancel(&dfops);
+                               break;
+                       }
+
+                       /* Remove the mapping from the CoW fork. */
+                       error = xfs_bunmapi_cow(ip, &irec);
+                       if (error)
+                               break;
+               }
+
+               /* Roll on... */
+               offset_fsb = irec.br_startoff + irec.br_blockcount;
+       }
+
+       return error;
+}
+
+/*
+ * Cancel all pending CoW reservations for some byte range of an inode.
+ */
+int
+xfs_reflink_cancel_cow_range(
+       struct xfs_inode        *ip,
+       xfs_off_t               offset,
+       xfs_off_t               count)
+{
+       struct xfs_trans        *tp;
+       xfs_fileoff_t           offset_fsb;
+       xfs_fileoff_t           end_fsb;
+       int                     error;
+
+       trace_xfs_reflink_cancel_cow_range(ip, offset, count);
+
+       offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
+       if (count == NULLFILEOFF)
+               end_fsb = NULLFILEOFF;
+       else
+               end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
+
+       /* Start a rolling transaction to remove the mappings */
+       error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
+                       0, 0, 0, &tp);
+       if (error)
+               goto out;
+
+       xfs_ilock(ip, XFS_ILOCK_EXCL);
+       xfs_trans_ijoin(tp, ip, 0);
+
+       /* Scrape out the old CoW reservations */
+       error = xfs_reflink_cancel_cow_blocks(ip, &tp, offset_fsb, end_fsb);
+       if (error)
+               goto out_cancel;
+
+       error = xfs_trans_commit(tp);
+
+       xfs_iunlock(ip, XFS_ILOCK_EXCL);
+       return error;
+
+out_cancel:
+       xfs_trans_cancel(tp);
+       xfs_iunlock(ip, XFS_ILOCK_EXCL);
+out:
+       trace_xfs_reflink_cancel_cow_range_error(ip, error, _RET_IP_);
+       return error;
+}
+
+/*
+ * Remap parts of a file's data fork after a successful CoW.
+ */
+int
+xfs_reflink_end_cow(
+       struct xfs_inode                *ip,
+       xfs_off_t                       offset,
+       xfs_off_t                       count)
+{
+       struct xfs_bmbt_irec            irec;
+       struct xfs_bmbt_irec            uirec;
+       struct xfs_trans                *tp;
+       xfs_fileoff_t                   offset_fsb;
+       xfs_fileoff_t                   end_fsb;
+       xfs_filblks_t                   count_fsb;
+       xfs_fsblock_t                   firstfsb;
+       struct xfs_defer_ops            dfops;
+       int                             error;
+       unsigned int                    resblks;
+       xfs_filblks_t                   ilen;
+       xfs_filblks_t                   rlen;
+       int                             nimaps;
+
+       trace_xfs_reflink_end_cow(ip, offset, count);
+
+       offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
+       end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
+       count_fsb = (xfs_filblks_t)(end_fsb - offset_fsb);
+
+       /* Start a rolling transaction to switch the mappings */
+       resblks = XFS_EXTENTADD_SPACE_RES(ip->i_mount, XFS_DATA_FORK);
+       error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
+                       resblks, 0, 0, &tp);
+       if (error)
+               goto out;
+
+       xfs_ilock(ip, XFS_ILOCK_EXCL);
+       xfs_trans_ijoin(tp, ip, 0);
+
+       /* Go find the old extent in the CoW fork. */
+       while (offset_fsb < end_fsb) {
+               /* Read extent from the source file */
+               nimaps = 1;
+               count_fsb = (xfs_filblks_t)(end_fsb - offset_fsb);
+               error = xfs_bmapi_read(ip, offset_fsb, count_fsb, &irec,
+                               &nimaps, XFS_BMAPI_COWFORK);
+               if (error)
+                       goto out_cancel;
+               ASSERT(nimaps == 1);
+
+               ASSERT(irec.br_startblock != DELAYSTARTBLOCK);
+               trace_xfs_reflink_cow_remap(ip, &irec);
+
+               /*
+                * We can have a hole in the CoW fork if part of a directio
+                * write is CoW but part of it isn't.
+                */
+               rlen = ilen = irec.br_blockcount;
+               if (irec.br_startblock == HOLESTARTBLOCK)
+                       goto next_extent;
+
+               /* Unmap the old blocks in the data fork. */
+               while (rlen) {
+                       xfs_defer_init(&dfops, &firstfsb);
+                       error = __xfs_bunmapi(tp, ip, irec.br_startoff,
+                                       &rlen, 0, 1, &firstfsb, &dfops);
+                       if (error)
+                               goto out_defer;
+
+                       /*
+                        * Trim the extent to whatever got unmapped.
+                        * Remember, bunmapi works backwards.
+                        */
+                       uirec.br_startblock = irec.br_startblock + rlen;
+                       uirec.br_startoff = irec.br_startoff + rlen;
+                       uirec.br_blockcount = irec.br_blockcount - rlen;
+                       irec.br_blockcount = rlen;
+                       trace_xfs_reflink_cow_remap_piece(ip, &uirec);
+
+                       /* Map the new blocks into the data fork. */
+                       error = xfs_bmap_map_extent(tp->t_mountp, &dfops,
+                                       ip, &uirec);
+                       if (error)
+                               goto out_defer;
+
+                       /* Remove the mapping from the CoW fork. */
+                       error = xfs_bunmapi_cow(ip, &uirec);
+                       if (error)
+                               goto out_defer;
+
+                       error = xfs_defer_finish(&tp, &dfops, ip);
+                       if (error)
+                               goto out_defer;
+               }
+
+next_extent:
+               /* Roll on... */
+               offset_fsb = irec.br_startoff + ilen;
+       }
+
+       error = xfs_trans_commit(tp);
+       xfs_iunlock(ip, XFS_ILOCK_EXCL);
+       if (error)
+               goto out;
+       return 0;
+
+out_defer:
+       xfs_defer_cancel(&dfops);
+out_cancel:
+       xfs_trans_cancel(tp);
+       xfs_iunlock(ip, XFS_ILOCK_EXCL);
+out:
+       trace_xfs_reflink_end_cow_error(ip, error, _RET_IP_);
+       return error;
+}
index 11408c0..bffa4be 100644 (file)
@@ -33,4 +33,12 @@ extern bool xfs_reflink_find_cow_mapping(struct xfs_inode *ip, xfs_off_t offset,
 extern int xfs_reflink_trim_irec_to_next_cow(struct xfs_inode *ip,
                xfs_fileoff_t offset_fsb, struct xfs_bmbt_irec *imap);
 
+extern int xfs_reflink_cancel_cow_blocks(struct xfs_inode *ip,
+               struct xfs_trans **tpp, xfs_fileoff_t offset_fsb,
+               xfs_fileoff_t end_fsb);
+extern int xfs_reflink_cancel_cow_range(struct xfs_inode *ip, xfs_off_t offset,
+               xfs_off_t count);
+extern int xfs_reflink_end_cow(struct xfs_inode *ip, xfs_off_t offset,
+               xfs_off_t count);
+
 #endif /* __XFS_REFLINK_H */