Re: [PATCH v2] ACPI: APEI: do not add task_work to kernel thread to avoid memory leak

[Shuai Xue <xueshuai@linux.alibaba.com>]

在 2022/9/25 AM1:17, Rafael J. Wysocki 写道:
> On Sat, Sep 24, 2022 at 9:50 AM Shuai Xue <xueshuai@linux.alibaba.com> wrote:
>>
>> If an error is detected as a result of user-space process accessing a
>> corrupt memory location, the CPU may take an abort. Then the platform
>> firmware reports kernel via NMI like notifications, e.g. NOTIFY_SEA,
>> NOTIFY_SOFTWARE_DELEGATED, etc.
>>
>> For NMI like notifications, commit 7f17b4a121d0 ("ACPI: APEI: Kick the
>> memory_failure() queue for synchronous errors") keep track of whether
>> memory_failure() work was queued, and make task_work pending to flush out
>> the queue so that the work is processed before return to user-space.
>>
>> The code use init_mm to check whether the error occurs in user space:
>>
>>     if (current->mm != &init_mm)
>>
>> The condition is always true, becase _nobody_ ever has "init_mm" as a real
>> VM any more.
>>
>> In addition to abort, errors can also be signaled as asynchronous
>> exceptions, such as interrupt and SError. In such case, the interrupted
>> current process could be any kind of thread. When a kernel thread is
>> interrupted, the work ghes_kick_task_work deferred to task_work will never
>> be processed because entry_handler returns to call ret_to_kernel() instead
>> of ret_to_user(). Consequently, the estatus_node alloced from
>> ghes_estatus_pool in ghes_in_nmi_queue_one_entry() will not be freed.
>> After around 200 allocations in our platform, the ghes_estatus_pool will
>> run of memory and ghes_in_nmi_queue_one_entry() returns ENOMEM. As a
>> result, the event failed to be processed.
>>
>>     sdei: event 805 on CPU 113 failed with error: -2
>>
>> Finally, a lot of unhandled events may cause platform firmware to exceed
>> some threshold and reboot.
>>
>> The condition should generally just do
>>
>>     if (current->mm)
>>
>> as described in active_mm.rst documentation.
>>
>> Then if an asynchronous error is detected when a kernel thread is running,
>> (e.g. when detected by a background scrubber), do not add task_work to it
>> as the original patch intends to do.
>>
>> Fixes: 7f17b4a121d0 ("ACPI: APEI: Kick the memory_failure() queue for synchronous errors")
>> Signed-off-by: Shuai Xue <xueshuai@linux.alibaba.com>
> 
> I need the APEI code reviewers to tell me that this is all OK.

Thank you for your reply. OK, let's wait the reviewers comments.

Best Regards,
Shuai


> 
>> ---
>> changes since v1:
>> - add description the side effect and give more details
>>
>>  drivers/acpi/apei/ghes.c | 2 +-
>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/drivers/acpi/apei/ghes.c b/drivers/acpi/apei/ghes.c
>> index d91ad378c00d..80ad530583c9 100644
>> --- a/drivers/acpi/apei/ghes.c
>> +++ b/drivers/acpi/apei/ghes.c
>> @@ -985,7 +985,7 @@ static void ghes_proc_in_irq(struct irq_work *irq_work)
>>                                 ghes_estatus_cache_add(generic, estatus);
>>                 }
>>
>> -               if (task_work_pending && current->mm != &init_mm) {
>> +               if (task_work_pending && current->mm) {
>>                         estatus_node->task_work.func = ghes_kick_task_work;
>>                         estatus_node->task_work_cpu = smp_processor_id();
>>                         ret = task_work_add(current, &estatus_node->task_work,
>> --
>> 2.20.1.12.g72788fdb
>>