On 10/10/19 8:24 PM, yangerkun wrote:
> 
> 
> On 2019/10/9 9:19, Jackie Liu wrote:
>> __io_get_deferred_req is used to get all defer lists, including defer_list
>> and timeout_list, but io_sequence_defer should be only cares about the 
>> sequence.
>>
>> Signed-off-by: Jackie Liu <[email protected]>
>> ---
>>    fs/io_uring.c | 13 +++++--------
>>    1 file changed, 5 insertions(+), 8 deletions(-)
>>
>> diff --git a/fs/io_uring.c b/fs/io_uring.c
>> index 8a0381f1a43b..8ec2443eb019 100644
>> --- a/fs/io_uring.c
>> +++ b/fs/io_uring.c
>> @@ -418,9 +418,7 @@ static struct io_ring_ctx *io_ring_ctx_alloc(struct 
>> io_uring_params *p)
>>    static inline bool io_sequence_defer(struct io_ring_ctx *ctx,
>>                                   struct io_kiocb *req)
>>    {
>> -    /* timeout requests always honor sequence */
>> -    if (!(req->flags & REQ_F_TIMEOUT) &&
>> -        (req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) != REQ_F_IO_DRAIN)
>> +    if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) != REQ_F_IO_DRAIN)
>>              return false;
>>    
>>      return req->sequence != ctx->cached_cq_tail + ctx->rings->sq_dropped;
>> @@ -435,12 +433,11 @@ static struct io_kiocb *__io_get_deferred_req(struct 
>> io_ring_ctx *ctx,
>>              return NULL;
>>    
>>      req = list_first_entry(list, struct io_kiocb, list);
>> -    if (!io_sequence_defer(ctx, req)) {
>> -            list_del_init(&req->list);
>> -            return req;
>> -    }
>> +    if (!(req->flags & REQ_F_TIMEOUT) && io_sequence_defer(ctx, req))
>> +            return NULL;
> Hi,
> 
> For timeout req, we should also compare the sequence to determine to
> return NULL or the req. But now we will return req directly. Actually,
> no need to compare req->flags with REQ_F_TIMEOUT.

Yes, not sure how I missed this, the patch is broken as-is. I will kill
it, cleanup can be done differently.

> Another problem, before this change, a timeout req can also drain the
> sqe(io_queue_sqe->io_req_defer and add to refer list), i am not sure is
> it a right or wrong logic, but your patch fix that.

We should factor the cq_timeouts into that.

-- 
Jens Axboe

Reply via email to