From e8c7d14ac6c37c173ec606907d38802b00302988 Mon Sep 17 00:00:00 2001 From: Luis Chamberlain Date: Fri, 19 Jun 2020 20:47:25 +0000 Subject: block: revert back to synchronous request_queue removal Commit dc9edc44de6c ("block: Fix a blk_exit_rl() regression") merged on v4.12 moved the work behind blk_release_queue() into a workqueue after a splat floated around which indicated some work on blk_release_queue() could sleep in blk_exit_rl(). This splat would be possible when a driver called blk_put_queue() or blk_cleanup_queue() (which calls blk_put_queue() as its final call) from an atomic context. blk_put_queue() decrements the refcount for the request_queue kobject, and upon reaching 0 blk_release_queue() is called. Although blk_exit_rl() is now removed through commit db6d99523560 ("block: remove request_list code") on v5.0, we reserve the right to be able to sleep within blk_release_queue() context. The last reference for the request_queue must not be called from atomic context. *When* the last reference to the request_queue reaches 0 varies, and so let's take the opportunity to document when that is expected to happen and also document the context of the related calls as best as possible so we can avoid future issues, and with the hopes that the synchronous request_queue removal sticks. We revert back to synchronous request_queue removal because asynchronous removal creates a regression with expected userspace interaction with several drivers. An example is when removing the loopback driver, one uses ioctls from userspace to do so, but upon return and if successful, one expects the device to be removed. Likewise if one races to add another device the new one may not be added as it is still being removed. This was expected behavior before and it now fails as the device is still present and busy still. Moving to asynchronous request_queue removal could have broken many scripts which relied on the removal to have been completed if there was no error. Document this expectation as well so that this doesn't regress userspace again. Using asynchronous request_queue removal however has helped us find other bugs. In the future we can test what could break with this arrangement by enabling CONFIG_DEBUG_KOBJECT_RELEASE. While at it, update the docs with the context expectations for the request_queue / gendisk refcount decrement, and make these expectations explicit by using might_sleep(). Fixes: dc9edc44de6c ("block: Fix a blk_exit_rl() regression") Suggested-by: Nicolai Stange Signed-off-by: Luis Chamberlain Reviewed-by: Christoph Hellwig Reviewed-by: Bart Van Assche Cc: Bart Van Assche Cc: Omar Sandoval Cc: Hannes Reinecke Cc: Nicolai Stange Cc: Greg Kroah-Hartman Cc: Michal Hocko Cc: yu kuai Signed-off-by: Jens Axboe --- block/blk-core.c | 8 ++++++++ 1 file changed, 8 insertions(+) (limited to 'block/blk-core.c') diff --git a/block/blk-core.c b/block/blk-core.c index f68398cb2ef6..a99b22fac38a 100644 --- a/block/blk-core.c +++ b/block/blk-core.c @@ -327,6 +327,9 @@ EXPORT_SYMBOL_GPL(blk_clear_pm_only); * * Decrements the refcount of the request_queue kobject. When this reaches 0 * we'll have blk_release_queue() called. + * + * Context: Any context, but the last reference must not be dropped from + * atomic context. */ void blk_put_queue(struct request_queue *q) { @@ -359,9 +362,14 @@ EXPORT_SYMBOL_GPL(blk_set_queue_dying); * * Mark @q DYING, drain all pending requests, mark @q DEAD, destroy and * put it. All future requests will be failed immediately with -ENODEV. + * + * Context: can sleep */ void blk_cleanup_queue(struct request_queue *q) { + /* cannot be called from atomic context */ + might_sleep(); + WARN_ON_ONCE(blk_queue_registered(q)); /* mark @q DYING, no new request or merges will be allowed afterwards */ -- cgit v1.2.3-58-ga151