I have a patch in my THP development tree that fails 10% of the readahead
I/Os in order to make sure the fallback paths are tested. This has
exposed three distinct problems so far, resulted in three patches that
are scheduled for 5.10:
iomap: Set all uptodate bits for an Uptodate page
iomap: Mark read blocks uptodate in write_begin
iomap: Clear page error before beginning a write
I've hit a fourth problem when running generic/127:
XFS (sdb): Internal error isnullstartblock(got.br_startblock) at line 5807 of file fs/xfs/libxfs/xfs_bmap.c. Caller xfs_bmap_collapse_extents+0x2bd/0x370
CPU: 4 PID: 4493 Comm: fsx Kdump: loaded Not tainted 5.9.0-rc3-00178-g35daf53935c9-dirty #765
Call Trace:
xfs_corruption_error+0x7c/0x80
xfs_bmap_collapse_extents+0x2e7/0x370
xfs_collapse_file_space+0x133/0x1e0
xfs_file_fallocate+0x110/0x480
vfs_fallocate+0x128/0x270
That finally persuaded me to port the patch to the current iomap for-next
tree (see below). Unfortunately, it doesn't reproduce, but I wonder
if it's simply that a 4kB page size is too small. Would anyone like to
give this a shot on a 64kB page size system? It usually takes less than
15 minutes to reproduce with my THP patchset, but doesn't reproduce in
2 hours without the THP patchset.
diff --git a/fs/iomap/buffered-io.c b/fs/iomap/buffered-io.c
index 8180061b9e16..2e67631a12ce 100644
--- a/fs/iomap/buffered-io.c
+++ b/fs/iomap/buffered-io.c
@@ -193,6 +193,8 @@ iomap_read_end_io(struct bio *bio)
struct bio_vec *bvec;
struct bvec_iter_all iter_all;
+ if (bio->bi_private == (void *)7)
+ error = -EIO;
bio_for_each_segment_all(bvec, bio, iter_all)
iomap_read_page_end_io(bvec, error);
bio_put(bio);
@@ -286,6 +288,12 @@ iomap_readpage_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
if (ctx->rac) /* same as readahead_gfp_mask */
gfp |= __GFP_NORETRY | __GFP_NOWARN;
ctx->bio = bio_alloc(gfp, min(BIO_MAX_PAGES, nr_vecs));
+ if (ctx->rac) {
+ static int error = 0;
+ ctx->bio->bi_private = (void *)(error++);
+ if (error == 10)
+ error = 0;
+ }
/*
* If the bio_alloc fails, try it again for a single page to
* avoid having to deal with partial page reads. This emulates
