[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <5b970999-c714-6bfb-0b02-ed206bafced4@yandex-team.ru>
Date: Mon, 28 Oct 2019 12:09:01 +0300
From: Konstantin Khlebnikov <khlebnikov@...dex-team.ru>
To: Linus Torvalds <torvalds@...ux-foundation.org>
Cc: David Howells <dhowells@...hat.com>,
Peter Zijlstra <peterz@...radead.org>,
Linux Kernel Mailing List <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH] pipe: wakeup writer only if pipe buffer is at least half
empty
On 27/10/2019 19.12, Linus Torvalds wrote:
> On Sun, Oct 27, 2019 at 11:46 AM Konstantin Khlebnikov
> <khlebnikov@...dex-team.ru> wrote:
>>
>> There is no reason to wakeup writer if pipe has only one empty page.
>> This means reader consumes data slower then writer produces it.
>>
>> This patch waits until buffer is at least half empty before waking writer.
>
> This is a bit dangerous, at least with David's other changes.
>
> In particular, there's now a "max_usage" in his series means that the
> writer might be blocked even if there's lots of free slots, because
> the writer is only allowed to use part of those slots.
>
> So I'd rather not see this logic particularly now that David is
> working on modifying the overall pipe logic.
>
> I do agree with the overall idea, but I'm not entirely happy about the
> "half full" logic, because it gets subtle with David's changes.
>
> Also, I'm a bit worried about cases where the readers and writers
> block on each other, and depend on "there's enough space in the pipe
> that we won't deadlock". Maybe the writer is blocked (because it
> filled the pipe), the reader reads just part of the pipe, and then the
> reader blocks on the writer doing something else, knowing that it just
> free'd up resources for the writer. But the writer is still blocked,
> and not woken up, because the pipe is still more than half full. See
> what I'm saying?
>
> I'm not sure anything like that exists, but it's an example of a "hmm"
> condition.
>
Ok. This breakage scenario is doubtful but such weird software really might exist.
What about making this thing tunable via fcntl like size of pipe buffer?
fcntl(fd, F_SETPIPE_WRITE_SZ, size)
Powered by blists - more mailing lists