Skip to contentSkip to conversation

PATCH: dangling pointer when switching to noop elevator in obscure block drivers.

PATCH Kevin Vigor 2 participants 2 messages Aug 20, 2010 – Aug 23, 2010

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

Linux 2.6.35 introduced a test in the beginning of elevator_init(), like so:

     if (unlikely(q->elevator))
         return 0;

So the following code sequence, which appears in two (obscure) block drivers is now a serious error:

        elevator_exit(q);
        elevator_init(q, "noop");

The intent is to cleanup the default system elevator and replace it with the noop elevator. Instead, elevator_exit() frees the existing elevator object, but leaves q->elevator pointing to it; elevator_init() then silently fails since q->elevator is non-NULL, and the queue is left with the elevator pointer invalid, This leads to untold woe and segfaults later.

The fix is trivial: zero the q->elevator pointer before calling elevator_init(). Note that drivers/s390/block/dasd.c already follows this pattern.

I do not have the hardware to actually test either of the two afflicted drivers, but I believe the fix to be sufficiently obvious. The following patches are against the current version of Linus' tree.

Thanks,
Kevin Vigor

drivers/s390/char/tape_block.c +1 -0
index b7de025..a1028c9 100644
@@ -218,6 +218,7 @@ tapeblock_setup_device(struct tape_device * device)
 		return -ENOMEM;
 
 	elevator_exit(blkdat->request_queue->elevator);
+	blkdat->request_queue->elevator = NULL;
 	rc = elevator_init(blkdat->request_queue, "noop");
 	if (rc)
 		goto cleanup_queue;
drivers/block/mg_disk.c +1 -0
index b82c5ce..c553404 100644
@@ -975,6 +975,7 @@ static int mg_probe(struct platform_device *plat_dev)
 
 	/* mflash is random device, thanx for the noop */
 	elevator_exit(host->breq->elevator);
+	host->breq->elevator = NULL;
 	err = elevator_init(host->breq, "noop");
 	if (err) {
 		printk(KERN_ERR "%s:%d (elevator_init) fail\n",
maintainer
#1 lore raw
Quoted 19 lines · Kevin Vigor

Linux 2.6.35 introduced a test in the beginning of elevator_init(), like so:

     if (unlikely(q->elevator))
         return 0;

So the following code sequence, which appears in two (obscure) block drivers is now a serious error:

        elevator_exit(q);
        elevator_init(q, "noop");

The intent is to cleanup the default system elevator and replace it with the noop elevator. Instead, elevator_exit() frees the existing elevator object, but leaves q->elevator pointing to it; elevator_init() then silently fails since q->elevator is non-NULL, and the queue is left with the elevator pointer invalid, This leads to untold woe and segfaults later.

Quoted 7 lines

The fix is trivial: zero the q->elevator pointer before calling elevator_init(). Note that drivers/s390/block/dasd.c already follows this pattern.

I do not have the hardware to actually test either of the two afflicted drivers, but I believe the fix to be sufficiently obvious. The following patches are against the current version of Linus' tree.

Thanks, this was also posted last week. I think the best course of action is to provide an API for switching the scheduler, instead of having drivers rely on doing it manually.

Signature · 1 line
Jens Axboe