On 12/30/19 9:01 AM, Mimi Zohar wrote:
Hi Mimi,
Hi Lakshmi,
On Fri, 2019-12-27 at 22:23 +0800, kernel test robot wrote:
[ 333.455345] BUG: sleeping function called from invalid context at kernel/locking/mutex.c:281
[ 333.457243] in_atomic(): 1, irqs_disabled(): 0, non_block: 0, pid: 12395, name: userfaultfd
[ 333.458888] CPU: 1 PID: 12395 Comm: userfaultfd Not tainted 5.5.0-rc1-00011-g11b771ffff8fc #1
[ 333.461096] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1 04/01/2014
[ 333.463893] Call Trace:
[ 333.465287] <IRQ>
[ 333.466351] dump_stack+0x66/0x8b
[ 333.467346] ___might_sleep+0x102/0x120
[ 333.468385] mutex_lock+0x1c/0x40
[ 333.469421] ima_process_queued_keys+0x24/0x110
[ 333.470529] ? ima_process_queued_keys+0x110/0x110
[ 333.471656] call_timer_fn+0x2d/0x140
[ 333.472707] run_timer_softirq+0x46f/0x4b0
[ 333.473752] ? enqueue_hrtimer+0x39/0xa0
[ 333.474780] __do_softirq+0xe3/0x2f8
[ 333.475768] irq_exit+0xd5/0xe0
[ 333.476738] smp_apic_timer_interrupt+0x74/0x140
[ 333.477834] apic_timer_interrupt+0xf/0x20
[ 333.478858] </IRQ>
I think this is an instance where defining timer_expired as atomic and
then testing it using atomic_dec_and_test() would help. Either the
queued keys would be deleted in ima_timer_handler() or measured in
ima_process_queued_keys().
Mimi
Would it be better to use a spinlock (instead of a mutex) to synchronize
access to the queued keys list? That would work for timer callback
function also.
Before submitting the patch for freeing the keys I had tested the case
where queued keys are freed in timer callback. But I did not hit the
above break.
I will follow the steps given by Rong and validate the fix.
thanks,
-lakshmi
