summaryrefslogtreecommitdiff
path: root/fs
diff options
context:
space:
mode:
authorLinus Torvalds <torvalds@linux-foundation.org>2022-02-05 20:04:43 +0300
committerLinus Torvalds <torvalds@linux-foundation.org>2022-02-05 20:04:43 +0300
commit524446e2179855534b425647dfc250757905aad8 (patch)
tree6e2c3237389b856c49e28c8394c97d6206158bbd /fs
parent0457e5153e0e8420134f60921349099e907264ca (diff)
parentebb7fb1557b1d03b906b668aa2164b51e6b7d19a (diff)
downloadlinux-524446e2179855534b425647dfc250757905aad8.tar.xz
Merge tag 'iomap-5.17-fixes-1' of git://git.kernel.org/pub/scm/fs/xfs/xfs-linux
Pull iomap fix from Darrick Wong: "A single bugfix for iomap. The fix should eliminate occasional complaints about stall warnings when a lot of writeback IO completes all at once and we have to then go clearing status on a large number of folios. Summary: - Limit the length of ioend chains in writeback so that we don't trip the softlockup watchdog and to limit long tail latency on clearing PageWriteback" * tag 'iomap-5.17-fixes-1' of git://git.kernel.org/pub/scm/fs/xfs/xfs-linux: xfs, iomap: limit individual ioend chain lengths in writeback
Diffstat (limited to 'fs')
-rw-r--r--fs/iomap/buffered-io.c52
-rw-r--r--fs/xfs/xfs_aops.c16
2 files changed, 63 insertions, 5 deletions
diff --git a/fs/iomap/buffered-io.c b/fs/iomap/buffered-io.c
index c938bbad075e..6c51a75d0be6 100644
--- a/fs/iomap/buffered-io.c
+++ b/fs/iomap/buffered-io.c
@@ -21,6 +21,8 @@
#include "../internal.h"
+#define IOEND_BATCH_SIZE 4096
+
/*
* Structure allocated for each folio when block size < folio size
* to track sub-folio uptodate status and I/O completions.
@@ -1039,7 +1041,7 @@ static void iomap_finish_folio_write(struct inode *inode, struct folio *folio,
* state, release holds on bios, and finally free up memory. Do not use the
* ioend after this.
*/
-static void
+static u32
iomap_finish_ioend(struct iomap_ioend *ioend, int error)
{
struct inode *inode = ioend->io_inode;
@@ -1048,6 +1050,7 @@ iomap_finish_ioend(struct iomap_ioend *ioend, int error)
u64 start = bio->bi_iter.bi_sector;
loff_t offset = ioend->io_offset;
bool quiet = bio_flagged(bio, BIO_QUIET);
+ u32 folio_count = 0;
for (bio = &ioend->io_inline_bio; bio; bio = next) {
struct folio_iter fi;
@@ -1062,9 +1065,11 @@ iomap_finish_ioend(struct iomap_ioend *ioend, int error)
next = bio->bi_private;
/* walk all folios in bio, ending page IO on them */
- bio_for_each_folio_all(fi, bio)
+ bio_for_each_folio_all(fi, bio) {
iomap_finish_folio_write(inode, fi.folio, fi.length,
error);
+ folio_count++;
+ }
bio_put(bio);
}
/* The ioend has been freed by bio_put() */
@@ -1074,20 +1079,36 @@ iomap_finish_ioend(struct iomap_ioend *ioend, int error)
"%s: writeback error on inode %lu, offset %lld, sector %llu",
inode->i_sb->s_id, inode->i_ino, offset, start);
}
+ return folio_count;
}
+/*
+ * Ioend completion routine for merged bios. This can only be called from task
+ * contexts as merged ioends can be of unbound length. Hence we have to break up
+ * the writeback completions into manageable chunks to avoid long scheduler
+ * holdoffs. We aim to keep scheduler holdoffs down below 10ms so that we get
+ * good batch processing throughput without creating adverse scheduler latency
+ * conditions.
+ */
void
iomap_finish_ioends(struct iomap_ioend *ioend, int error)
{
struct list_head tmp;
+ u32 completions;
+
+ might_sleep();
list_replace_init(&ioend->io_list, &tmp);
- iomap_finish_ioend(ioend, error);
+ completions = iomap_finish_ioend(ioend, error);
while (!list_empty(&tmp)) {
+ if (completions > IOEND_BATCH_SIZE * 8) {
+ cond_resched();
+ completions = 0;
+ }
ioend = list_first_entry(&tmp, struct iomap_ioend, io_list);
list_del_init(&ioend->io_list);
- iomap_finish_ioend(ioend, error);
+ completions += iomap_finish_ioend(ioend, error);
}
}
EXPORT_SYMBOL_GPL(iomap_finish_ioends);
@@ -1108,6 +1129,18 @@ iomap_ioend_can_merge(struct iomap_ioend *ioend, struct iomap_ioend *next)
return false;
if (ioend->io_offset + ioend->io_size != next->io_offset)
return false;
+ /*
+ * Do not merge physically discontiguous ioends. The filesystem
+ * completion functions will have to iterate the physical
+ * discontiguities even if we merge the ioends at a logical level, so
+ * we don't gain anything by merging physical discontiguities here.
+ *
+ * We cannot use bio->bi_iter.bi_sector here as it is modified during
+ * submission so does not point to the start sector of the bio at
+ * completion.
+ */
+ if (ioend->io_sector + (ioend->io_size >> 9) != next->io_sector)
+ return false;
return true;
}
@@ -1209,8 +1242,10 @@ iomap_alloc_ioend(struct inode *inode, struct iomap_writepage_ctx *wpc,
ioend->io_flags = wpc->iomap.flags;
ioend->io_inode = inode;
ioend->io_size = 0;
+ ioend->io_folios = 0;
ioend->io_offset = offset;
ioend->io_bio = bio;
+ ioend->io_sector = sector;
return ioend;
}
@@ -1251,6 +1286,13 @@ iomap_can_add_to_ioend(struct iomap_writepage_ctx *wpc, loff_t offset,
return false;
if (sector != bio_end_sector(wpc->ioend->io_bio))
return false;
+ /*
+ * Limit ioend bio chain lengths to minimise IO completion latency. This
+ * also prevents long tight loops ending page writeback on all the
+ * folios in the ioend.
+ */
+ if (wpc->ioend->io_folios >= IOEND_BATCH_SIZE)
+ return false;
return true;
}
@@ -1335,6 +1377,8 @@ iomap_writepage_map(struct iomap_writepage_ctx *wpc,
&submit_list);
count++;
}
+ if (count)
+ wpc->ioend->io_folios++;
WARN_ON_ONCE(!wpc->ioend && !list_empty(&submit_list));
WARN_ON_ONCE(!folio_test_locked(folio));
diff --git a/fs/xfs/xfs_aops.c b/fs/xfs/xfs_aops.c
index 2705f91bdd0d..9d6a67c7d227 100644
--- a/fs/xfs/xfs_aops.c
+++ b/fs/xfs/xfs_aops.c
@@ -136,7 +136,20 @@ done:
memalloc_nofs_restore(nofs_flag);
}
-/* Finish all pending io completions. */
+/*
+ * Finish all pending IO completions that require transactional modifications.
+ *
+ * We try to merge physical and logically contiguous ioends before completion to
+ * minimise the number of transactions we need to perform during IO completion.
+ * Both unwritten extent conversion and COW remapping need to iterate and modify
+ * one physical extent at a time, so we gain nothing by merging physically
+ * discontiguous extents here.
+ *
+ * The ioend chain length that we can be processing here is largely unbound in
+ * length and we may have to perform significant amounts of work on each ioend
+ * to complete it. Hence we have to be careful about holding the CPU for too
+ * long in this loop.
+ */
void
xfs_end_io(
struct work_struct *work)
@@ -157,6 +170,7 @@ xfs_end_io(
list_del_init(&ioend->io_list);
iomap_ioend_try_merge(ioend, &tmp);
xfs_end_ioend(ioend);
+ cond_resched();
}
}