@@ -954,6 +954,7 @@ int blk_register_queue(struct gendisk *disk)
blk_queue_flag_set(QUEUE_FLAG_REGISTERED, q);
wbt_enable_default(q);
blk_throtl_register_queue(q);
+ blk_queue_flag_set(QUEUE_FLAG_THROTL_INIT_DONE, q);
/* Now everything is ready and send out KOBJ_ADD uevent */
kobject_uevent(&q->kobj, KOBJ_ADD);
@@ -986,6 +987,7 @@ void blk_unregister_queue(struct gendisk *disk)
if (!blk_queue_registered(q))
return;
+ blk_queue_flag_clear(QUEUE_FLAG_THROTL_INIT_DONE, q);
/*
* Since sysfs_remove_dir() prevents adding new directory entries
* before removal of existing entries starts, protect against
@@ -11,6 +11,8 @@
#include <linux/bio.h>
#include <linux/blktrace_api.h>
#include <linux/blk-cgroup.h>
+#include <linux/sched/signal.h>
+#include <linux/delay.h>
#include "blk.h"
/* Max dispatch from a group in 1 round */
@@ -1428,6 +1430,31 @@ static void tg_conf_updated(struct throtl_grp *tg, bool global)
}
}
+static inline int throtl_check_init_done(struct request_queue *q)
+{
+ if (test_bit(QUEUE_FLAG_THROTL_INIT_DONE, &q->queue_flags))
+ return 0;
+
+ return blk_queue_dying(q) ? -ENODEV : -EBUSY;
+}
+
+/*
+ * If throtl_check_init_done() return -EBUSY, we should retry after a short
+ * msleep(), since that throttle init will be completed in blk_register_queue()
+ * soon.
+ */
+static inline int throtl_restart_syscall_when_busy(int errno)
+{
+ int ret = errno;
+
+ if (ret == -EBUSY) {
+ msleep(10);
+ ret = restart_syscall();
+ }
+
+ return ret;
+}
+
static ssize_t tg_set_conf(struct kernfs_open_file *of,
char *buf, size_t nbytes, loff_t off, bool is_u64)
{
@@ -1441,6 +1468,10 @@ static ssize_t tg_set_conf(struct kernfs_open_file *of,
if (ret)
return ret;
+ ret = throtl_check_init_done(ctx.disk->queue);
+ if (ret)
+ goto out_finish;
+
ret = -EINVAL;
if (sscanf(ctx.body, "%llu", &v) != 1)
goto out_finish;
@@ -1448,7 +1479,6 @@ static ssize_t tg_set_conf(struct kernfs_open_file *of,
v = U64_MAX;
tg = blkg_to_tg(ctx.blkg);
-
if (is_u64)
*(u64 *)((void *)tg + of_cft(of)->private) = v;
else
@@ -1458,6 +1488,8 @@ static ssize_t tg_set_conf(struct kernfs_open_file *of,
ret = 0;
out_finish:
blkg_conf_finish(&ctx);
+ ret = throtl_restart_syscall_when_busy(ret);
+
return ret ?: nbytes;
}
@@ -1607,8 +1639,11 @@ static ssize_t tg_set_limit(struct kernfs_open_file *of,
if (ret)
return ret;
- tg = blkg_to_tg(ctx.blkg);
+ ret = throtl_check_init_done(ctx.disk->queue);
+ if (ret)
+ goto out_finish;
+ tg = blkg_to_tg(ctx.blkg);
v[0] = tg->bps_conf[READ][index];
v[1] = tg->bps_conf[WRITE][index];
v[2] = tg->iops_conf[READ][index];
@@ -1704,6 +1739,8 @@ static ssize_t tg_set_limit(struct kernfs_open_file *of,
ret = 0;
out_finish:
blkg_conf_finish(&ctx);
+ ret = throtl_restart_syscall_when_busy(ret);
+
return ret ?: nbytes;
}
@@ -684,6 +684,7 @@ struct request_queue {
#define QUEUE_FLAG_NOMERGES 5 /* disable merge attempts */
#define QUEUE_FLAG_SAME_COMP 6 /* complete on same CPU-group */
#define QUEUE_FLAG_FAIL_IO 7 /* fake timeout */
+#define QUEUE_FLAG_THROTL_INIT_DONE 8 /* io throttle can be online */
#define QUEUE_FLAG_NONROT 9 /* non-rotational device (SSD) */
#define QUEUE_FLAG_VIRT QUEUE_FLAG_NONROT /* paravirt device */
#define QUEUE_FLAG_IO_STAT 10 /* do IO stats */
If blk-throttle is enabled and io is issued before blk_throtl_register_queue() is done. Divide by zero crash will be triggered in tg_may_dispatch() because 'throtl_slice' is uninitialized. The problem is fixed in commit 75f4dca59694 ("block: call blk_register_queue earlier in device_add_disk") from mainline, however it's too hard to backport this patch due to lots of refactoring. Thus introduce a new flag QUEUE_FLAG_THROTL_INIT_DONE. It will be set after blk_throtl_register_queue() is done, and will be checked before applying any config. Signed-off-by: Yu Kuai <yukuai3@huawei.com> --- block/blk-sysfs.c | 2 ++ block/blk-throttle.c | 41 +++++++++++++++++++++++++++++++++++++++-- include/linux/blkdev.h | 1 + 3 files changed, 42 insertions(+), 2 deletions(-)