This is the mail archive of the
mailing list for the glibc project.
Re: tilegx32 NPTL nptl/tst-cancel17 failure
- From: "Paul E. Murphy" <murphyp at linux dot vnet dot ibm dot com>
- To: Chris Metcalf <cmetcalf at ezchip dot com>, GNU C Library <libc-alpha at sourceware dot org>
- Date: Wed, 20 Jan 2016 11:03:28 -0600
- Subject: Re: tilegx32 NPTL nptl/tst-cancel17 failure
- Authentication-results: sourceware.org; auth=none
- References: <569FB90D dot 3010006 at ezchip dot com>
Yes, I've seen a similar failure on ppc64le too. It's on
my list of failures to investigate. It seems to be race
condition of some sort. It doesn't happen often though.
Thanks for looking into the failure.
On 01/20/2016 10:42 AM, Chris Metcalf wrote:
> So, this test works fine in 64-bit mode on tilegx, but fails in
> the compat 32-bit mode. I looked a little bit at it and concluded
> I didn't have any great ideas. The test emits this:
> going to cancel tf in-time
> going to cancel tf2 in-time
> in-time cancellation succeeded
> aio_cancel failed
> going to cancel tf early
> going to cancel tf2 early
> Segmentation fault
> The segmentation fault is due to the last AIO_MISC_NOTIFY() in
> __aio_notify() having a "waitlist->counterp" pointer pointing into
> the text of the binary where it isn't writable (it's actually pointing
> at the tf() function start point).
> Anyone else seen anything like this or have any suggestions?
> I realized I'm not sure how long this has been a problem, since
> we have an outstanding bug for tilegx for exceptions aren't
> properly caught for tst-cancelx17 and others, and I may have been
> missing tst-cancel17 when I do my 32-bit compat test since it
> got swept up with the other similarly-named ones.