From: Justin Chen <justinpopo6@gmail.com>
To: libc-help@sourceware.org
Cc: Florian Fainelli <f.fainelli@gmail.com>
Subject: Glibc pthread_rwlock_timed*() Optimization Bug
Date: Tue, 7 Sep 2021 14:31:58 -0700 [thread overview]
Message-ID: <CAJx26kU_T=uam6uZOcHJ9_a-G2onePtHCU9q2F2g0OhoJzR51w@mail.gmail.com> (raw)
Hello all,
We are seeing a faulty compiler optimization with
pthread_rwlock_timed*() when cross-compiling glibc for arm. This is
leading to deadlocks with some of our software.
We are using glibc-2.27, but I see the same issue with the current
master branch. This is built with the master branch using the
following configure.
../configure --host=arm-unknown-linux-gnueabihf
--prefix=/local/users/jc957059/source/glibc/test_lib/lib-built/
make && make install
The code in question is in nptl/pthread_rwlock_common.c line 490-501
/* We still need to wait for explicit hand-over, but we must
not use futex_wait anymore because we would just time out
in this case and thus make the spin-waiting we need
unnecessarily expensive. */
while ((atomic_load_relaxed (&rwlock->__data.__wrphase_futex)
| PTHREAD_RWLOCK_FUTEX_USED)
== (1 | PTHREAD_RWLOCK_FUTEX_USED))
{
/* TODO Back-off? */
}
ready = true;
break;
The compiled ASM is the following
5dc98: f043 0302 orr.w r3, r3, #2
atomic_thread_fence_acquire ();
/* We still need to wait for explicit hand-over, but we must
not use futex_wait anymore because we would just time out
in this case and thus make the spin-waiting we need
unnecessarily expensive. */
while ((atomic_load_relaxed (&rwlock->__data.__wrphase_futex)
5dc9c: 2b03 cmp r3, #3
5dc9e: d184 bne.n 5dbaa
<__pthread_rwlock_timedrdlock64+0x6e>
5dca0: e7fe b.n 5dca0
<__pthread_rwlock_timedrdlock64+0x164>
We only read __wrphase_futex once then hit an infinite loop.
Adding volatile seems to do the trick.
diff --git a/sysdeps/nptl/bits/struct_rwlock.h
b/sysdeps/nptl/bits/struct_rwlock.h
index 2f8b7ac..cd47bd2 100644
--- a/sysdeps/nptl/bits/struct_rwlock.h
+++ b/sysdeps/nptl/bits/struct_rwlock.h
@@ -30,7 +30,7 @@ struct __pthread_rwlock_arch_t
{
unsigned int __readers;
unsigned int __writers;
- unsigned int __wrphase_futex;
+ volatile unsigned int __wrphase_futex;
unsigned int __writers_futex;
unsigned int __pad3;
unsigned int __pad4;
The compiled ASM with this change (and with a few declarations
corrected with the volatile type)
5d2ca: f043 0302 orr.w r3, r3, #2
atomic_thread_fence_acquire ();
/* We still need to wait for explicit hand-over, but we must
not use futex_wait anymore because we would just time out
in this case and thus make the spin-waiting we need
unnecessarily expensive. */
while ((atomic_load_relaxed (&rwlock->__data.__wrphase_futex)
5d2ce: 2b03 cmp r3, #3
5d2d0: d0fa beq.n 5d2c8
<__pthread_rwlock_clockrdlock64+0x168>
5d2d2: e783 b.n 5d1dc
<__pthread_rwlock_clockrdlock64+0x7c>
No longer have infinite loop here.
It seems like the compiler is incorrectly optimizing the loop because
it is not informed that the value of __wrphase_futex can be changed in
another context, which I believe should be done with the volatile
attribute.
Does this analysis look correct?
Thank You,
Justin
next reply other threads:[~2021-09-07 21:32 UTC|newest]
Thread overview: 4+ messages / expand[flat|nested] mbox.gz Atom feed top
2021-09-07 21:31 Justin Chen [this message]
2021-09-07 21:33 ` Florian Fainelli
2021-09-08 6:14 ` Florian Weimer
2021-09-08 6:20 ` Florian Weimer
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to='CAJx26kU_T=uam6uZOcHJ9_a-G2onePtHCU9q2F2g0OhoJzR51w@mail.gmail.com' \
--to=justinpopo6@gmail.com \
--cc=f.fainelli@gmail.com \
--cc=libc-help@sourceware.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).