]> git.baikalelectronics.ru Git - kernel.git/commitdiff
blk-iolatency: fix IO hang due to negative inflight counter
authorLiu Bo <bo.liu@linux.alibaba.com>
Fri, 25 Jan 2019 00:12:47 +0000 (08:12 +0800)
committerJens Axboe <axboe@kernel.dk>
Fri, 8 Feb 2019 19:42:26 +0000 (12:42 -0700)
Our test reported the following stack, and vmcore showed that
->inflight counter is -1.

[ffffc9003fcc38d0] __schedule at ffffffff8173d95d
[ffffc9003fcc3958] schedule at ffffffff8173de26
[ffffc9003fcc3970] io_schedule at ffffffff810bb6b6
[ffffc9003fcc3988] blkcg_iolatency_throttle at ffffffff813911cb
[ffffc9003fcc3a20] rq_qos_throttle at ffffffff813847f3
[ffffc9003fcc3a48] blk_mq_make_request at ffffffff8137468a
[ffffc9003fcc3b08] generic_make_request at ffffffff81368b49
[ffffc9003fcc3b68] submit_bio at ffffffff81368d7d
[ffffc9003fcc3bb8] ext4_io_submit at ffffffffa031be00 [ext4]
[ffffc9003fcc3c00] ext4_writepages at ffffffffa03163de [ext4]
[ffffc9003fcc3d68] do_writepages at ffffffff811c49ae
[ffffc9003fcc3d78] __filemap_fdatawrite_range at ffffffff811b6188
[ffffc9003fcc3e30] filemap_write_and_wait_range at ffffffff811b6301
[ffffc9003fcc3e60] ext4_sync_file at ffffffffa030cee8 [ext4]
[ffffc9003fcc3ea8] vfs_fsync_range at ffffffff8128594b
[ffffc9003fcc3ee8] do_fsync at ffffffff81285abd
[ffffc9003fcc3f18] sys_fsync at ffffffff81285d50
[ffffc9003fcc3f28] do_syscall_64 at ffffffff81003c04
[ffffc9003fcc3f50] entry_SYSCALL_64_after_swapgs at ffffffff81742b8e

The ->inflight counter may be negative (-1) if

1) blk-iolatency was disabled when the IO was issued,

2) blk-iolatency was enabled before this IO reached its endio,

3) the ->inflight counter is decreased from 0 to -1 in endio()

In fact the hang can be easily reproduced by the below script,

H=/sys/fs/cgroup/unified/
P=/sys/fs/cgroup/unified/test

echo "+io" > $H/cgroup.subtree_control
mkdir -p $P

echo $$ > $P/cgroup.procs

xfs_io -f -d -c "pwrite 0 4k" /dev/sdg

echo "`cat /sys/block/sdg/dev` target=1000000" > $P/io.latency

xfs_io -f -d -c "pwrite 0 4k" /dev/sdg

This fixes the problem by freezing the queue so that while
enabling/disabling iolatency, there is no inflight rq running.

Note that quiesce_queue is not needed as this only updating iolatency
configuration about which dispatching request_queue doesn't care.

Signed-off-by: Liu Bo <bo.liu@linux.alibaba.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
block/blk-iolatency.c

index fc714ef402a6a5d61840659e8f3e32c2301ae172..1893686a9c1f3b3fcbfdf1275193e5a34da6673c 100644 (file)
@@ -72,6 +72,7 @@
 #include <linux/sched/loadavg.h>
 #include <linux/sched/signal.h>
 #include <trace/events/block.h>
+#include <linux/blk-mq.h>
 #include "blk-rq-qos.h"
 #include "blk-stat.h"
 
@@ -601,6 +602,9 @@ static void blkcg_iolatency_done_bio(struct rq_qos *rqos, struct bio *bio)
                return;
 
        enabled = blk_iolatency_enabled(iolat->blkiolat);
+       if (!enabled)
+               return;
+
        while (blkg && blkg->parent) {
                iolat = blkg_to_lat(blkg);
                if (!iolat) {
@@ -610,7 +614,7 @@ static void blkcg_iolatency_done_bio(struct rq_qos *rqos, struct bio *bio)
                rqw = &iolat->rq_wait;
 
                atomic_dec(&rqw->inflight);
-               if (!enabled || iolat->min_lat_nsec == 0)
+               if (iolat->min_lat_nsec == 0)
                        goto next;
                iolatency_record_time(iolat, &bio->bi_issue, now,
                                      issue_as_root);
@@ -754,10 +758,13 @@ int blk_iolatency_init(struct request_queue *q)
        return 0;
 }
 
-static void iolatency_set_min_lat_nsec(struct blkcg_gq *blkg, u64 val)
+/*
+ * return 1 for enabling iolatency, return -1 for disabling iolatency, otherwise
+ * return 0.
+ */
+static int iolatency_set_min_lat_nsec(struct blkcg_gq *blkg, u64 val)
 {
        struct iolatency_grp *iolat = blkg_to_lat(blkg);
-       struct blk_iolatency *blkiolat = iolat->blkiolat;
        u64 oldval = iolat->min_lat_nsec;
 
        iolat->min_lat_nsec = val;
@@ -766,9 +773,10 @@ static void iolatency_set_min_lat_nsec(struct blkcg_gq *blkg, u64 val)
                                    BLKIOLATENCY_MAX_WIN_SIZE);
 
        if (!oldval && val)
-               atomic_inc(&blkiolat->enabled);
+               return 1;
        if (oldval && !val)
-               atomic_dec(&blkiolat->enabled);
+               return -1;
+       return 0;
 }
 
 static void iolatency_clear_scaling(struct blkcg_gq *blkg)
@@ -800,6 +808,7 @@ static ssize_t iolatency_set_limit(struct kernfs_open_file *of, char *buf,
        u64 lat_val = 0;
        u64 oldval;
        int ret;
+       int enable = 0;
 
        ret = blkg_conf_prep(blkcg, &blkcg_policy_iolatency, buf, &ctx);
        if (ret)
@@ -834,7 +843,12 @@ static ssize_t iolatency_set_limit(struct kernfs_open_file *of, char *buf,
        blkg = ctx.blkg;
        oldval = iolat->min_lat_nsec;
 
-       iolatency_set_min_lat_nsec(blkg, lat_val);
+       enable = iolatency_set_min_lat_nsec(blkg, lat_val);
+       if (enable) {
+               WARN_ON_ONCE(!blk_get_queue(blkg->q));
+               blkg_get(blkg);
+       }
+
        if (oldval != iolat->min_lat_nsec) {
                iolatency_clear_scaling(blkg);
        }
@@ -842,6 +856,24 @@ static ssize_t iolatency_set_limit(struct kernfs_open_file *of, char *buf,
        ret = 0;
 out:
        blkg_conf_finish(&ctx);
+       if (ret == 0 && enable) {
+               struct iolatency_grp *tmp = blkg_to_lat(blkg);
+               struct blk_iolatency *blkiolat = tmp->blkiolat;
+
+               blk_mq_freeze_queue(blkg->q);
+
+               if (enable == 1)
+                       atomic_inc(&blkiolat->enabled);
+               else if (enable == -1)
+                       atomic_dec(&blkiolat->enabled);
+               else
+                       WARN_ON_ONCE(1);
+
+               blk_mq_unfreeze_queue(blkg->q);
+
+               blkg_put(blkg);
+               blk_put_queue(blkg->q);
+       }
        return ret ?: nbytes;
 }
 
@@ -977,8 +1009,14 @@ static void iolatency_pd_offline(struct blkg_policy_data *pd)
 {
        struct iolatency_grp *iolat = pd_to_lat(pd);
        struct blkcg_gq *blkg = lat_to_blkg(iolat);
+       struct blk_iolatency *blkiolat = iolat->blkiolat;
+       int ret;
 
-       iolatency_set_min_lat_nsec(blkg, 0);
+       ret = iolatency_set_min_lat_nsec(blkg, 0);
+       if (ret == 1)
+               atomic_inc(&blkiolat->enabled);
+       if (ret == -1)
+               atomic_dec(&blkiolat->enabled);
        iolatency_clear_scaling(blkg);
 }