On 4/30/20 1:51 PM, Jens Axboe wrote:
> On 4/30/20 12:47 PM, Jens Axboe wrote:
>> On 4/30/20 11:58 AM, Matthew Wilcox wrote:
>>> On Thu, Apr 30, 2020 at 10:24:46AM -0600, Jens Axboe wrote:
>>>> Pipe read/write only checks for the file O_NONBLOCK flag, but we should
>>>> also check for IOCB_NOWAIT for whether or not we should handle this read
>>>> or write in a non-blocking fashion. If we don't, then we will block on
>>>> data or space for iocbs that explicitly asked for non-blocking
>>>> operation. This messes up callers that explicitly ask for non-blocking
>>>> operations.
>>>>
>>>> Signed-off-by: Jens Axboe <axboe@xxxxxxxxx>
>>>
>>> Wouldn't this be better?
>>
>> Yeah, that's probably a better idea. Care to send a "proper" patch?
>
> I take that back, running into issues going with a whole-sale conversion
> like that:
>
> mkdir("/run/dhcpcd", 0755) = -1 EEXIST (File exists)
> openat(AT_FDCWD, "/run/dhcpcd/ens7.pid", O_WRONLY|O_CREAT|O_NONBLOCK|O_CLOEXEC, 0644) = 4
> flock(4, LOCK_EX|LOCK_NB) = 0
> getpid() = 214
> ftruncate(4, 0) = 0
> lseek(4, 0, SEEK_SET) = 0
> fstat(4, {st_mode=S_IFREG|0644, st_size=0, ...}) = 0
> lseek(4, 0, SEEK_CUR) = 0
> write(4, "214\n", 4) = -1 EINVAL (Invalid argument)
>
> which I don't know where is coming from yet, but it's definitely
> breakage by auto setting IOCB_NOWAIT if O_NONBLOCK is set.
>
> I'd prefer to go your route, but I also would like this fixed for pipes
> for 5.7. So I'd suggest we go with mine, and then investigate why this
> is breaking stuff and go with the all-in approach for 5.8 if feasible.
OK, it's the old classic in generic_write_checks(), is my guess:
if ((iocb->ki_flags & IOCB_NOWAIT) && !(iocb->ki_flags & IOCB_DIRECT))
return -EINVAL;
so we definitely can't just flip the switch on O_NONBLOCK -> IOCB_NOWAIT
in general, at least not for writes.
--
Jens Axboe
