On 03/22/2019 07:16 AM, Oleg Nesterov wrote:
> On 03/21, Matthew Wilcox wrote:
>> On Thu, Mar 21, 2019 at 05:45:10PM -0400, Waiman Long wrote:
>>
>>> To avoid this dire condition and reduce lock hold time of tasklist_lock,
>>> flush_sigqueue() is modified to pass in a freeing queue pointer so that
>>> the actual freeing of memory objects can be deferred until after the
>>> tasklist_lock is released and irq re-enabled.
>> I think this is a really bad solution. It looks kind of generic,
>> but isn't. It's terribly inefficient, and all it's really doing is
>> deferring the debugging code until we've re-enabled interrupts.
> Agreed.
Thanks for looking into that. As I am not knowledgeable enough about the
signal handling code path, I choose the lowest risk approach of not
trying to change the code flow while deferring memory deallocation after
releasing the tasklist_lock.
>> We'd be much better off just having a list_head in the caller
>> and list_splice() the queue->list onto that caller. Then call
>> __sigqueue_free() for each signal on the queue.
> This won't work, note the comment which explains the race with sigqueue_free().
>
> Let me think about it... at least we can do something like
>
> close_the_race_with_sigqueue_free(struct sigpending *queue)
> {
> struct sigqueue *q, *t;
>
> list_for_each_entry_safe(q, t, ...) {
> if (q->flags & SIGQUEUE_PREALLOC)
> list_del_init(&q->list);
> }
>
> called with ->siglock held, tasklist_lock is not needed.
>
> After that flush_sigqueue() can be called lockless in release_task() release_task.
>
> I'll try to make the patch tomorrow.
>
> Oleg.
>
I am looking forward to it.
Thanks,
Longman
