On 06/02/2020 23:16, Jens Axboe wrote:
> On 2/6/20 1:00 PM, Pavel Begunkov wrote:
>> On 06/02/2020 22:56, Jens Axboe wrote:
>>> On 2/6/20 10:16 AM, Pavel Begunkov wrote:
>>>> On 06/02/2020 20:04, Pavel Begunkov wrote:
>>>>> On 06/02/2020 19:51, Pavel Begunkov wrote:
>>>>>> After defer, a request will be prepared, that includes allocating iovec
>>>>>> if needed, and then submitted through io_wq_submit_work() but not custom
>>>>>> handler (e.g. io_rw_async()/io_sendrecv_async()). However, it'll leak
>>>>>> iovec, as it's in io-wq and the code goes as follows:
>>>>>>
>>>>>> io_read() {
>>>>>> if (!io_wq_current_is_worker())
>>>>>> kfree(iovec);
>>>>>> }
>>>>>>
>>>>>> Put all deallocation logic in io_{read,write,send,recv}(), which will
>>>>>> leave the memory, if going async with -EAGAIN.
>>>>>>
>>>>> Interestingly, this will fail badly if it returns -EAGAIN from io-wq context.
>>>>> Apparently, I need to do v2.
>>>>>
>>>> Or not...
>>>> Jens, can you please explain what's with the -EAGAIN handling in
>>>> io_wq_submit_work()? Checking the code, it seems neither of
>>>> read/write/recv/send can return -EAGAIN from async context (i.e.
>>>> force_nonblock=false). Are there other ops that can do it?
>>>
>>> Nobody should return -EAGAIN with force_nonblock=false, they should
>>> end the io_kiocb inline for that.
>>>
>>
>> If so for those 4, then the patch should work well.
>
> Maybe I'm dense, but I'm not seeing the leak? We have two cases here:
>
> - The number of vecs is less than UIO_FASTIOV, in which case we use the
> on-stack inline_vecs. If we need to go async, we copy that inline vec
> to our async_ctx area.
>
> - The number of vecs is more than UIO_FASTIOV, this is where iovec is
> allocated by the vec import. If we make it to completion here, we
> free it at the end of eg io_read(). If we need to go async, we stash
> that pointer in our async_ctx area and free it when the work item
> has run and completed.
>
BTW, there are plenty of ways to leak even with this applied. E.g. double
io_read_prep() call with ->io, and that may happen. Or by not punting in
__io_queue_sqe() after io_issue_sqe()==-EAGAIN.
That's the next patch I'm preparing, and then I'm good for splice(2).
--
Pavel Begunkov
Attachment:
signature.asc
Description: OpenPGP digital signature
