Switching block device elevators
After commit 1abec4fdbb142e3ccb6ce99832fae42129134a96, "block: make blk_init_free_list and elevator_init idempotent", we're seeing kernel panics in our s390 tape block device driver. The panic is triggered because our driver tries to replace the default elevator with a noop elevator by calling elevator_exit() directly followed by elevator_init().
Since the commit, elevator_init() returns 0 if request_queue->elevator is non-null, even though it does not install a new elevator. As a result, the next access to the elevator finds a pointer to the old one which was already freed and a panic is triggered. Our current fix consists of setting the elevator pointer to NULL after elevator_exit().
There is at least one other driver where the problem currently exists (drivers/block/mg_disk.c, author on cc) and another s390 driver where the problem was only accidentally fixed before 2.6.35. I'm wondering if there's a better solution (apart from not forcing an elevator) and would like to hear everyone's opinion on this matter. How about declaring elevator_switch() non-static, for example?
Regards,
Peter Oberparleiter
Quoted 6 lines · Peter Oberparleiter
After commit
1abec4fdbb142e3ccb6ce99832fae42129134a96, "block: makeblk_init_free_listandelevator_initidempotent", we're seeing kernel panics in our s390 tape block device driver. The panic is triggered because our driver tries to replace the default elevator with a noop elevator by callingelevator_exit()directly followed byelevator_init().
Maybe we should look to export elevator_switch() -- rather than confining its use to the sysfs interface.
Quoted 6 lines
Since the commit,
elevator_init()returns 0 ifrequest_queue->elevatoris non-null, even though it does not install a new elevator. As a result, the next access to the elevator finds a pointer to the old one which was already freed and a panic is triggered. Our current fix consists of setting the elevator pointer to NULL afterelevator_exit().
elevator_exit() triggers a call, via kobj, to elevator_release() which doesn't have access to the request_queue to reset it.
Unfortunately, commit 1abec4fdbb imposes that the elevator_exit() caller must take care to reset q->elevator to NULL -- like dasd_alloc_queue() does.
Though I suppose we _could_ pass request_queue to elevator_exit.
Quoted 7 lines
There is at least one other driver where the problem currently exists (
drivers/block/mg_disk.c, author on cc) and another s390 driver where the problem was only accidentally fixed before 2.6.35. I'm wondering if there's a better solution (apart from not forcing an elevator) and would like to hear everyone's opinion on this matter. How about declaringelevator_switch()non-static, for example?
Right, updating drivers/block/mg_disk.c and drivers/s390/block/dasd.c to use elevator_switch would work.
Mike