Re: [PATCH] io_uring: execute task_work_run() before dropping mm

[Xiaoguang Wang <xiaoguang.wang@linux.alibaba.com>]

hi,

> On 06/06/2020 18:12, Xiaoguang Wang wrote:
>> While testing io_uring in our internal kernel, note it's not upstream
>> kernel, we see below panic:
>> [  872.498723] x29: ffff00002d553cf0 x28: 0000000000000000
>> [  872.508973] x27: ffff807ef691a0e0 x26: 0000000000000000
>> [  872.519116] x25: 0000000000000000 x24: ffff0000090a7980
>> [  872.529184] x23: ffff000009272060 x22: 0000000100022b11
>> [  872.539144] x21: 0000000046aa5668 x20: ffff80bee8562b18
>> [  872.549000] x19: ffff80bee8562080 x18: 0000000000000000
>> [  872.558876] x17: 0000000000000000 x16: 0000000000000000
>> [  872.568976] x15: 0000000000000000 x14: 0000000000000000
>> [  872.578762] x13: 0000000000000000 x12: 0000000000000000
>> [  872.588474] x11: 0000000000000000 x10: 0000000000000c40
>> [  872.598324] x9 : ffff000008100c00 x8 : 000000007ffff000
>> [  872.608014] x7 : ffff80bee8562080 x6 : ffff80beea862d30
>> [  872.617709] x5 : 0000000000000000 x4 : ffff80beea862d48
>> [  872.627399] x3 : ffff80bee8562b18 x2 : 0000000000000000
>> [  872.637044] x1 : ffff0000090a7000 x0 : 0000000000208040
>> [  872.646575] Call trace:
>> [  872.653139]  task_numa_work+0x4c/0x310
>> [  872.660916]  task_work_run+0xb0/0xe0
>> [  872.668400]  io_sq_thread+0x164/0x388
>> [  872.675829]  kthread+0x108/0x138
>>
>> The reason is that once io_sq_thread has a valid mm, schedule subsystem
>> may call task_tick_numa() adding a task_numa_work() callback, which will
>> visit mm, then above panic will happen.
>>
>> To fix this bug, only call task_work_run() before dropping mm.
> 
> So, the problem is that poll/async paths re-issue requests with
> __io_queue_sqe(), which doesn't care about current->mm, and which
> can be NULL for io_sq_thread(). Right?
No, above panic is not triggered by poll/async paths.
See below code path:
==> task_tick_fair()
====> task_tick_numa()
======> task_work_add, work is task_numa_work, which will visit mm.

In sqpoll mode, there maybe are sqes that need mm, then above codes
maybe executed by schedule subsystem. In io_sq_thread, we drop mm before
task_work_run, if there is a task_numa_work, panic occurs.

Regards,
Xiaoguang Wang
> 
>>
>> Signed-off-by: Xiaoguang Wang <xiaoguang.wang@linux.alibaba.com>
>> ---
>>   fs/io_uring.c | 15 ++++++++-------
>>   1 file changed, 8 insertions(+), 7 deletions(-)
>>
>> diff --git a/fs/io_uring.c b/fs/io_uring.c
>> index 6391a00ff8b7..32381984b2a6 100644
>> --- a/fs/io_uring.c
>> +++ b/fs/io_uring.c
>> @@ -6134,6 +6134,13 @@ static int io_sq_thread(void *data)
>>   		 * to enter the kernel to reap and flush events.
>>   		 */
>>   		if (!to_submit || ret == -EBUSY) {
>> +			/*
>> +			 * Current task context may already have valid mm, that
>> +			 * means some works that visit mm may have been queued,
>> +			 * so we must execute the works before dropping mm.
>> +			 */
>> +			if (current->task_works)
>> +				task_work_run();
> 
> Even though you're not dropping mm, the thread might not have it in the first
> place. see how it's done in io_init_req(). How about setting mm either lazily
> in io_poll_task_func()/io_async_task_func(), or before task_work_run() in
> io_sq_thread().
> 
>>   			/*
>>   			 * Drop cur_mm before scheduling, we can't hold it for
>>   			 * long periods (or over schedule()). Do this before
>> @@ -6152,8 +6159,6 @@ static int io_sq_thread(void *data)
>>   			if (!list_empty(&ctx->poll_list) ||
>>   			    (!time_after(jiffies, timeout) && ret != -EBUSY &&
>>   			    !percpu_ref_is_dying(&ctx->refs))) {
>> -				if (current->task_works)
>> -					task_work_run();
>>   				cond_resched();
>>   				continue;
>>   			}
>> @@ -6185,11 +6190,7 @@ static int io_sq_thread(void *data)
>>   					finish_wait(&ctx->sqo_wait, &wait);
>>   					break;
>>   				}
>> -				if (current->task_works) {
>> -					task_work_run();
>> -					finish_wait(&ctx->sqo_wait, &wait);
>> -					continue;
>> -				}
>> +
>>   				if (signal_pending(current))
>>   					flush_signals(current);
>>   				schedule();
>>
>