2020-04-26 17:47:15

by Linus Torvalds

[permalink] [raw]
Subject: Re: [LTP] [PATCH v4 3/3] syscalls/pipe2_03: Add new test for pipe2 O_DIRECT flag

On Sun, Apr 26, 2020 at 4:59 AM Li Wang <[email protected]> wrote:
>
> From kernel code seems you are right. The pipe indeed takes use of PAGE_SIZE(ppc64le: 64kB) to split the writes data in the packetized mode (marked by O_DIRECT). But in the manual page, O_DIRECT indicates us the PIPE_BUF is the correct atomic unit.

The manual is correct.

PIPE_BUF is the size we _guarantee_ can be used atomically.

The fact that in practice we do have bigger buffers on some platforms
is an implementation detail.

Yes, that implementation detail can be visible, but basically any test
code that tries to test for "what if we use a bigger bug that
PIPE_BUF" is buggy. It's simply not guaranteed to work any more.

O_DIRECT is kind of immaterial, except it's just one of those things
where the atomic size is slightly more visible. But basically,
packetized pipes with bigger packets than PIPE_BUF is random behavior.
It may work. It may not.

Linus


2020-04-30 03:27:45

by Yang Xu

[permalink] [raw]
Subject: Re: [LTP] [PATCH v4 3/3] syscalls/pipe2_03: Add new test for pipe2 O_DIRECT flag

Hi Linus


> On Sun, Apr 26, 2020 at 4:59 AM Li Wang <[email protected]> wrote:
>>
>> From kernel code seems you are right. The pipe indeed takes use of PAGE_SIZE(ppc64le: 64kB) to split the writes data in the packetized mode (marked by O_DIRECT). But in the manual page, O_DIRECT indicates us the PIPE_BUF is the correct atomic unit.
>
> The manual is correct.
>
> PIPE_BUF is the size we _guarantee_ can be used atomically.
>
> The fact that in practice we do have bigger buffers on some platforms
> is an implementation detail.
>
> Yes, that implementation detail can be visible, but basically any test
> code that tries to test for "what if we use a bigger bug that
> PIPE_BUF" is buggy. It's simply not guaranteed to work any more.
>
> O_DIRECT is kind of immaterial, except it's just one of those things
> where the atomic size is slightly more visible. But basically,
> packetized pipes with bigger packets than PIPE_BUF is random behavior.
> It may work. It may not.
Thanks for your explanation. I am more curious about the user scene of
this flag.

@Li, so how to design this test? In this test, we don't have complex
scene to test this automic unit.

Best Regards
Yang Xu
>
> Linus
>
>