[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-Id: <1231538061.29452.8.camel@twins>
Date: Fri, 09 Jan 2009 22:54:21 +0100
From: Peter Zijlstra <peterz@...radead.org>
To: Brad Parker <brad@...ltoe.com>
Cc: linux-kernel@...r.kernel.org,
Linus Torvalds <torvalds@...ux-foundation.org>,
Ingo Molnar <mingo@...e.hu>
Subject: Re: copy_{to,from}_user
On Fri, 2009-01-09 at 12:52 -0500, Brad Parker wrote:
> I have a question about copy_{to,from}_user.
>
> Most implementations I've seen do in-order copies and notice when an
> exception occurs and report back the progress. This is straight
> forward.
>
> (but to be honest, I have suspicions about how just how accurate those
> reports are i.e. +/- 1-3 bytes on some architectures)
>
> On some cpu's it is advantageous to do an out-of-order copy to take
> advantage of various cache fill mechanisms.
>
> The problem is that the out-of-order copy makes it impossible to know
> where the exception occurred (in terms of progress).
>
> Would it be permissible to have a version of copy_{to,from}_user which
> does an out-of-order copy and when an exception occurs, restarts the
> copy from the beginning using a simple in-order copy, to make it
> possible to identify where the exception occurs?
>
> The idea is that exceptions are rare and so the performance hit of doing
> the "recopy" would be minimal and would provide the required accuracy.
x86_64 already does some unrolling and is inaccurate as to where exactly
it happens. The only thing that is very important is that you _never_
say you copied more than you actually did.
That was the source of a data corruption bug a while ago, the code did
something like sequences: read 8 words, write 8 words. And reported the
number of bytes read, instead of bytes written, which is an
over-estimation.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
Powered by blists - more mailing lists