On Fri 28-04-17 17:46:47, Tejun Heo wrote: > On Fri, Apr 21, 2017 at 09:49:17PM +0000, Bart Van Assche wrote: > > On Thu, 2017-04-20 at 15:18 -0600, Scott Bauer wrote: > > > [ 642.638860] BUG: KASAN: use-after-free in scsi_exit_rq+0xf3/0x120 at addr ffff8802b7fedf00 > > > [ 642.639362] Read of size 1 by task rcuos/5/53 > > > [ 642.639713] CPU: 7 PID: 53 Comm: rcuos/6 Not tainted 4.11.0-rc5+ #13 > > > [ 642.640170] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.7.5-0-ge51488c-20140602_164612-nilsson.home.kraxel.org 04/01/2014 > > > [ 642.640923] Call Trace: > > > [ 642.641080] dump_stack+0x63/0x8f > > > [ 642.641289] kasan_object_err+0x21/0x70 > > > [ 642.641531] kasan_report.part.1+0x231/0x500 > > > [ 642.641823] ? scsi_exit_rq+0xf3/0x120 > > > [ 642.642054] ? _raw_spin_unlock_irqrestore+0xe/0x10 > > > [ 642.642353] ? free_percpu+0x1b7/0x340 > > > [ 642.642586] ? put_task_stack+0x117/0x2b0 > > > [ 642.642837] __asan_report_load1_noabort+0x2e/0x30 > > > [ 642.643138] scsi_exit_rq+0xf3/0x120 > > > [ 642.643366] free_request_size+0x44/0x60 > > > [ 642.643614] mempool_destroy.part.6+0x9b/0x150 > > > [ 642.643899] ? kasan_slab_free+0x87/0xb0 > > > [ 642.644152] mempool_destroy+0x13/0x20 > > > [ 642.644394] blk_exit_rl+0x36/0x40 > > > [ 642.644614] blkg_free+0x146/0x200 > > > [ 642.644836] __blkg_release_rcu+0x121/0x220 > > > [ 642.645112] rcu_nocb_kthread+0x61f/0xca0 > > > [ 642.645376] ? get_state_synchronize_rcu+0x20/0x20 > > > [ 642.645690] ? pci_mmcfg_check_reserved+0x110/0x110 > > > [ 642.646011] kthread+0x298/0x390 > > > [ 642.646224] ? get_state_synchronize_rcu+0x20/0x20 > > > [ 642.646535] ? kthread_park+0x160/0x160 > > > [ 642.646787] ret_from_fork+0x2c/0x40 > > > > I'm not familiar with cgroups but seeing this makes me wonder whether it would > > be possible to move the blk_exit_rl() calls from blk_release_queue() into > > blk_cleanup_queue()? The SCSI core frees a SCSI host after blk_cleanup_queue() > > has finished for all associated SCSI devices. This is why I think that calling > > blk_exit_rl() earlier would be sufficient to avoid that scsi_exit_rq() > > dereferences a SCSI host pointer after it has been freed. > > Hmm... I see. Didn't know request put path involved derefing those > structs. The blk_exit_rl() call just replaced open coded > mempool_destroy call, so the destruction order was always like this. > We shouldn't have any request in flight by cleanup, so moving it there > should be fine. So I'm also not aware of any particular breakage this would cause. However logically the freeing of request mempools really belongs to blk_release_queue() so it seems a bit dumb to move blk_exit_rl() just because SCSI stores the fact from which slab cache it has allocated the sense buffer in a structure (shost) that it frees under its hands by the time blk_release_queue() is called. :-| Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
