On Sat, 2017-09-02 at 21:12 +0800, Ming Lei wrote: > Please let us know if V3 addresses your previous concern about calling > blk_queue_enter_live() during preempt freezing. Do you understand how request queue cleanup works? The algorithm used for request queue cleanup is as follows: * Set the DYING flag. This flag makes all later blk_get_request() calls fail. * Wait until all pending requests have finished. * Set the DEAD flag. For the traditional block layer, this flag causes blk_run_queue() not to call .request_fn() anymore. For blk-mq it is guaranteed in another way that .queue_rq() won't be called anymore after this flag has been set. Allowing blk_get_request() to succeed after the DYING flag has been set is completely wrong because that could result in a request being queued after the DEAD flag has been set, resulting in either a hanging request or a kernel crash. This is why it's completely wrong to add a blk_queue_enter_live() call in blk_old_get_request() or blk_mq_alloc_request(). Hence my NAK for any patch that adds a blk_queue_enter_live() call to any function called from blk_get_request(). That includes the patch at the start of this e-mail thread. Bart.
