GNU bug report logs - #53041
29.0.50; TRAMP spins the CPU by polling the child processes without a delay

Previous Next

Full log

Message #8 received at 53041 <at> debbugs.gnu.org (full text, mbox):

From: Michael Albinus <michael.albinus <at> gmx.de>
To: Dima Kogan <dima <at> secretsauce.net>
Cc: 53041 <at> debbugs.gnu.org
Subject: Re: bug#53041: 29.0.50; TRAMP spins the CPU by polling the child
 processes without a delay
Date: Sun, 09 Jan 2022 14:46:04 +0100

Dima Kogan <dima <at> secretsauce.net> writes:

> Hi.

Hi Dima,

> I use TRAMP regularly, and I often see it redline my CPU, which
> shouldn't be happening.
>
> The cause in all cases I've seen is TRAMP expecting some output from the
> child process, and looking for this output in a delay-less loop. For
> instance (tramp-process-one-action) looks like this:
>
>   (defun tramp-process-one-action (proc vec actions)
>       ....
>     (while (not found)
>       (while (tramp-accept-process-output proc 0))
>       .... )
>
> The (while (tramp-accept-process-output proc 0)) form does
>
>   Read all available data; returns immediately if none is available
>
> So here we spin the CPU until there's some data to look at AND until the
> incoming data meets some condition we're looking for. In order to not
> spin, at least one of the (tramp-accept-process-output) calls needs to
> block. The simplest thing to do to fix this is to replace
>
>   (while (tramp-accept-process-output proc 0))
>
> with
>
>   (tramp-accept-process-output proc nil)
>
> Here we block until we get SOME data back. I think this is probably
> good-enough, since the outer loop will get more data, if it's needed. If
> we really want to replace the original logic with blocking, we can do
> this instead:
>
>   (let (timeout)
>     (while
>         (prog1
>             (tramp-accept-process-output proc timeout)
>           (setq timeout 0))))
>
> Either one of these makes most of these issues disappear. There are more
> places in the code where we call (tramp-accept-process-output ... 0),
> and I think they're all wrong: we should always block. I can send a
> patch, but let's agree on the approach first. My preference is to
> replace all the (while (tramp-accept-process-output proc 0)) with
> (tramp-accept-process-output proc nil) unless there's a specific reason
> not to.
>
> One easy way to reproduce one such behavior:
>
> 1. Start up emacs
> 2. open /ssh:SERVER:FILE
> 3. Break the network connection (I'm on a laptop. Leaving the wifi area
>    is enough)
> 4. Try to type into the buffer visiting FILE
> 5. See emacs block the user while spinning the CPU.

This was discussed several times already. The most recent discussion wrt
Tramp starts at <https://lists.gnu.org/archive/html/emacs-devel/2019-01/msg00301.html>.

The pattern (while (accept-process-output p) was proposed by Stefan
Monnier in <https://lists.gnu.org/archive/html/emacs-devel/2019-01/msg00338.html>,
so this is used in Tramp. I do not want to reopen this can of worms, really.

To fix your problem of a broken connection, the Tramp manual recommends
to add "ServerAliveInterval 5" in your ~/.ssh/config, see (info "(tramp)
Frequently Asked Questions") . Additionally, you might set "ServerAliveCountMax 2".

> Thanks

Best regards, Michael.

Previous Next

GNU bug tracking system
Copyright (C) 1999 Darren O. Benham, 1997,2003 nCipher Corporation Ltd, 1994-97 Ian Jackson.