This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH] Fix __lll_timedlock_wait busy-wait issue
- From: Torvald Riegel <triegel at redhat dot com>
- To: "Joseph S. Myers" <joseph at codesourcery dot com>
- Cc: Will Newton <will dot newton at linaro dot org>, Maxim Kuvyrkov <maxim dot kuvyrkov at linaro dot org>, bniebuhr at efjohnson dot com, uclibc at uclibc dot org, "libc-ports at sourceware dot org" <libc-ports at sourceware dot org>, libc-alpha <libc-alpha at sourceware dot org>
- Date: Fri, 28 Mar 2014 19:24:58 +0100
- Subject: Re: [PATCH] Fix __lll_timedlock_wait busy-wait issue
- Authentication-results: sourceware.org; auth=none
- References: <1395409800-4457-1-git-send-email-bniebuhr at efjohnson dot com> <09F962CB-595F-4FAB-9435-52C237DB402C at linaro dot org> <CANu=Dmhn-a-PUNA88OfpoAQyvUZGd3UH+LB+cf=fLWyUNyyU+A at mail dot gmail dot com> <Pine dot LNX dot 4 dot 64 dot 1403272157560 dot 25264 at digraph dot polyomino dot org dot uk>
On Thu, 2014-03-27 at 22:01 +0000, Joseph S. Myers wrote:
> I don't know how this might relate to
> <https://sourceware.org/bugzilla/show_bug.cgi?id=15119> (see
> <https://sourceware.org/ml/libc-ports/2013-01/msg00084.html> and
> <https://sourceware.org/ml/libc-ports/2013-02/msg00021.html> and the rest
> of that thread). But my preference for how to address this is definitely
> to move to unifying lowlevellock.[ch] files across as many architectures
> as possible - which requires someone to understand the differences and
> produce a careful analysis that shows what the best form for generic files
> is and what cases actually require architecture-specific files to override
> those generic files (preferably overriding only the bits that need
> overriding).
>
I agree. My gut feeling is that the locks should eventually become
unified C code, using atomics to do the synchronization;
architecture-specific code should be either in the atomics or in more
generally useful spin-waiting code (which could be used by other sync
constructs as well). The futex syscall is really on the slowpath; if
you hit it, you will have had at least a cache miss on the futex var,
and doing the syscall will likely give you more cache misses.
Therefore, I don't see a reason why the futex syscall needs to have
custom asm implementations such as on x86 currently.