blk-wbt: introduce a new disable state to prevent false positive by rwb_enabled()
authorZhang Yi <yi.zhang@huawei.com>
Sat, 19 Jun 2021 09:36:59 +0000 (17:36 +0800)
committerJens Axboe <axboe@kernel.dk>
Mon, 21 Jun 2021 21:03:41 +0000 (15:03 -0600)
Now that we disable wbt by simply zero out rwb->wb_normal in
wbt_disable_default() when switch elevator to bfq, but it's not safe
because it will become false positive if we change queue depth. If it
become false positive between wbt_wait() and wbt_track() when submit
write request, it will lead to drop rqw->inflight to -1 in wbt_done(),
which will end up trigger IO hung. Fix this issue by introduce a new
state which mean the wbt was disabled.

Fixes: a79050434b45 ("blk-rq-qos: refactor out common elements of blk-wbt")
Signed-off-by: Zhang Yi <yi.zhang@huawei.com>
Link: https://lore.kernel.org/r/20210619093700.920393-2-yi.zhang@huawei.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
block/blk-wbt.c
block/blk-wbt.h

index b363b05..0ce0883 100644 (file)
@@ -77,7 +77,8 @@ enum {
 
 static inline bool rwb_enabled(struct rq_wb *rwb)
 {
-       return rwb && rwb->wb_normal != 0;
+       return rwb && rwb->enable_state != WBT_STATE_OFF_DEFAULT &&
+                     rwb->wb_normal != 0;
 }
 
 static void wb_timestamp(struct rq_wb *rwb, unsigned long *var)
@@ -701,7 +702,7 @@ void wbt_disable_default(struct request_queue *q)
        rwb = RQWB(rqos);
        if (rwb->enable_state == WBT_STATE_ON_DEFAULT) {
                blk_stat_deactivate(rwb->cb);
-               rwb->wb_normal = 0;
+               rwb->enable_state = WBT_STATE_OFF_DEFAULT;
        }
 }
 EXPORT_SYMBOL_GPL(wbt_disable_default);
index 16bdc85..2eb01be 100644 (file)
@@ -34,6 +34,7 @@ enum {
 enum {
        WBT_STATE_ON_DEFAULT    = 1,
        WBT_STATE_ON_MANUAL     = 2,
+       WBT_STATE_OFF_DEFAULT
 };
 
 struct rq_wb {