Re: [Intel-gfx] [PATCH] drm/i915: Fix syncmap memory leak

[John Harrison <john.c.harrison@intel.com>]

On 8/6/2021 11:29, Matthew Brost wrote:
> On Fri, Aug 06, 2021 at 11:23:06AM -0700, John Harrison wrote:
>> On 7/30/2021 12:53, Matthew Brost wrote:
>>> A small race exists between intel_gt_retire_requests_timeout and
>>> intel_timeline_exit which could result in the syncmap not getting
>>> free'd. Rather than work to hard to seal this race, simply cleanup the
>> free'd -> freed
>>
> Sure.
>
>>> syncmap on fini.
>>>
>>> unreferenced object 0xffff88813bc53b18 (size 96):
>>>     comm "gem_close_race", pid 5410, jiffies 4294917818 (age 1105.600s)
>>>     hex dump (first 32 bytes):
>>>       01 00 00 00 00 00 00 00 00 00 00 00 0a 00 00 00  ................
>>>       00 00 00 00 00 00 00 00 6b 6b 6b 6b 06 00 00 00  ........kkkk....
>>>     backtrace:
>>>       [<00000000120b863a>] __sync_alloc_leaf+0x1e/0x40 [i915]
>>>       [<00000000042f6959>] __sync_set+0x1bb/0x240 [i915]
>>>       [<0000000090f0e90f>] i915_request_await_dma_fence+0x1c7/0x400 [i915]
>>>       [<0000000056a48219>] i915_request_await_object+0x222/0x360 [i915]
>>>       [<00000000aaac4ee3>] i915_gem_do_execbuffer+0x1bd0/0x2250 [i915]
>>>       [<000000003c9d830f>] i915_gem_execbuffer2_ioctl+0x405/0xce0 [i915]
>>>       [<00000000fd7a8e68>] drm_ioctl_kernel+0xb0/0xf0 [drm]
>>>       [<00000000e721ee87>] drm_ioctl+0x305/0x3c0 [drm]
>>>       [<000000008b0d8986>] __x64_sys_ioctl+0x71/0xb0
>>>       [<0000000076c362a4>] do_syscall_64+0x33/0x80
>>>       [<00000000eb7a4831>] entry_SYSCALL_64_after_hwframe+0x44/0xa9
>>>
>>> Signed-off-by: Matthew Brost <matthew.brost@intel.com>
>>> Fixes: 531958f6f357 ("drm/i915/gt: Track timeline activeness in enter/exit")
>>> Cc: <stable@vger.kernel.org>
>>> ---
>>>    drivers/gpu/drm/i915/gt/intel_timeline.c | 9 +++++++++
>>>    1 file changed, 9 insertions(+)
>>>
>>> diff --git a/drivers/gpu/drm/i915/gt/intel_timeline.c b/drivers/gpu/drm/i915/gt/intel_timeline.c
>>> index c4a126c8caef..1257f4f11e66 100644
>>> --- a/drivers/gpu/drm/i915/gt/intel_timeline.c
>>> +++ b/drivers/gpu/drm/i915/gt/intel_timeline.c
>>> @@ -127,6 +127,15 @@ static void intel_timeline_fini(struct rcu_head *rcu)
>>>    	i915_vma_put(timeline->hwsp_ggtt);
>>>    	i915_active_fini(&timeline->active);
>>> +
>>> +	/*
>>> +	 * A small race exists between intel_gt_retire_requests_timeout and
>>> +	 * intel_timeline_exit which could result in the syncmap not getting
>>> +	 * free'd. Rather than work to hard to seal this race, simply cleanup
>>> +	 * the syncmap on fini.
>> What is the race? I'm going round in circles just trying to work out how
>> intel_gt_retire_requests_timeout is supposed to get to intel_timeline_exit
>> in the first place.
>>
> intel_gt_retire_requests_timeout increments tl->active_count, active_count == 2
> intel_timeline_exit is called, returns on atomic_add_unless, active_count == 1
> intel_gt_retire_requests_timeout decrements tl->active_count, active_count == 0
> i915_syncmap_free is never called, memory leak
>
> Matt
Okay. Think I follow it now.

Seems like the syncmap free should have been in timeline_fini instead of 
timeline_exit in the first place?

Reviewed-by: John Harrison <John.C.Harrison@Intel.com>


>
>> Also, free'd -> freed.
>>
>> John.
>>
>>
>>> +	 */
>>> +	i915_syncmap_free(&timeline->sync);
>>> +
>>>    	kfree(timeline);
>>>    }