xfs: rework xfs_buf_bio_endio error handling

Currently the report of a bio error from completion
immediately marks the buffer with an error. The issue is that this
is racy w.r.t. synchronous IO - the submitter can see b_error being
set before the IO is complete, and hence we cannot differentiate
between submission failures and completion failures.

Add an internal b_io_error field protected by the b_lock to catch IO
completion errors, and only propagate that to the buffer during
final IO completion handling. Hence we can tell in xfs_buf_iorequest
if we've had a submission failure bey checking bp->b_error before
dropping our b_io_remaining reference - that reference will prevent
b_io_error values from being propagated to b_error in the event that
completion races with submission.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>

2 files changed, 17 insertions, 2 deletions

diff --git a/fs/xfs/xfs_buf.c b/fs/xfs/xfs_buf.c
index a046149e6099..170d6c0afe71 100644
--- a/fs/xfs/xfs_buf.c
+++ b/fs/xfs/xfs_buf.c
@@ -1008,6 +1008,13 @@ xfs_buf_ioend(
 
         bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
 
+        /*
+         * Pull in IO completion errors now. We are guaranteed to be running
+         * single threaded, so we don't need the lock to read b_io_error.
+         */
+        if (!bp->b_error && bp->b_io_error)
+                xfs_buf_ioerror(bp, bp->b_io_error);
+
         /* Only validate buffers that were read without errors */
         if (read && !bp->b_error && bp->b_ops) {
                 ASSERT(!bp->b_iodone);
@@ -1192,8 +1199,12 @@ xfs_buf_bio_end_io(
          * don't overwrite existing errors - otherwise we can lose errors on
          * buffers that require multiple bios to complete.
          */
-        if (!bp->b_error)
-                xfs_buf_ioerror(bp, error);
+        if (error) {
+                spin_lock(&bp->b_lock);
+                if (!bp->b_io_error)
+                        bp->b_io_error = error;
+                spin_unlock(&bp->b_lock);
+        }
 
         if (!bp->b_error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
                 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
@@ -1379,6 +1390,9 @@ xfs_buf_iorequest(
         if (bp->b_flags & XBF_WRITE)
                 xfs_buf_wait_unpin(bp);
 
+        /* clear the internal error state to avoid spurious errors */
+        bp->b_io_error = 0;
+
         /*
          * Take references to the buffer. For XBF_ASYNC buffers, holding a
          * reference for as long as submission takes is all that is necessary
diff --git a/fs/xfs/xfs_buf.h b/fs/xfs/xfs_buf.h
index 4585c1595a98..44db8cd67bda 100644
--- a/fs/xfs/xfs_buf.h
+++ b/fs/xfs/xfs_buf.h
@@ -158,6 +158,7 @@ typedef struct xfs_buf {
         struct list_head        b_lru;          /* lru list */
         spinlock_t              b_lock;         /* internal state lock */
         unsigned int            b_state;        /* internal state flags */
+        int                     b_io_error;     /* internal IO error state */
         wait_queue_head_t       b_waiters;      /* unpin waiters */
         struct list_head        b_list;
         struct xfs_perag        *b_pag;         /* contains rbtree root */