On Mon, 2018-09-24 at 13:13 -0600, Keith Busch wrote: +AD4 diff --git a/block/blk-mq.c b/block/blk-mq.c +AD4 index 85a1c1a59c72..28d128450621 100644 +AD4 --- a/block/blk-mq.c +AD4 +-+-+- b/block/blk-mq.c +AD4 +AEAAQA -848,22 +-848,6 +AEAAQA static void blk+AF8-mq+AF8-timeout+AF8-work(struct work+AF8-struct +ACo-work) +AD4 struct blk+AF8-mq+AF8-hw+AF8-ctx +ACo-hctx+ADs +AD4 int i+ADs +AD4 +AD4 - /+ACo A deadlock might occur if a request is stuck requiring a +AD4 - +ACo timeout at the same time a queue freeze is waiting +AD4 - +ACo completion, since the timeout code would not be able to +AD4 - +ACo acquire the queue reference here. +AD4 - +ACo +AD4 - +ACo That's why we don't use blk+AF8-queue+AF8-enter here+ADs instead, we use +AD4 - +ACo percpu+AF8-ref+AF8-tryget directly, because we need to be able to +AD4 - +ACo obtain a reference even in the short window between the queue +AD4 - +ACo starting to freeze, by dropping the first reference in +AD4 - +ACo blk+AF8-freeze+AF8-queue+AF8-start, and the moment the last request is +AD4 - +ACo consumed, marked by the instant q+AF8-usage+AF8-counter reaches +AD4 - +ACo zero. +AD4 - +ACo-/ +AD4 - if (+ACE-percpu+AF8-ref+AF8-tryget(+ACY-q-+AD4-q+AF8-usage+AF8-counter)) +AD4 - return+ADs +AD4 - +AD4 blk+AF8-mq+AF8-queue+AF8-tag+AF8-busy+AF8-iter(q, blk+AF8-mq+AF8-check+AF8-expired, +ACY-next)+ADs +AD4 +AD4 if (next +ACEAPQ 0) +AHs +AD4 +AEAAQA -881,7 +-865,6 +AEAAQA static void blk+AF8-mq+AF8-timeout+AF8-work(struct work+AF8-struct +ACo-work) +AD4 blk+AF8-mq+AF8-tag+AF8-idle(hctx)+ADs +AD4 +AH0 +AD4 +AH0 +AD4 - blk+AF8-queue+AF8-exit(q)+ADs +AD4 +AH0 Hi Keith, The above introduces a behavior change: if the percpu+AF8-ref+AF8-tryget() call inside blk+AF8-mq+AF8-queue+AF8-tag+AF8-busy+AF8-iter() fails then blk+AF8-mq+AF8-timeout+AF8-work() will now call blk+AF8-mq+AF8-tag+AF8-idle(). I think that's wrong if the percpu+AF8-ref+AF8-tryget() call fails due to the queue having been frozen. Please make blk+AF8-mq+AF8-queue+AF8-tag+AF8-busy+AF8-iter() return a bool that indicates whether or not it has iterated over the request queue. Thanks, Bart.
