lkml.org 
[lkml]   [2010]   [Aug]   [27]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
Subjectfs hang in 2.6.27.y | Fwd: [Bug 15658] New: [PATCH] x86 constant_test_bit() prone to misoptimization with gcc-4.4
	Hello,
I was told that lkml is still highly preferred over bugzilla,
so reposting the bugreport here.

The problem first manifested as "ext3 hang" in 2.6.27 based
kernel; the cause appeared to be somewhat more involved,
see the analysis below. hpa@ confirmed he's seen similar
problems elsewhere.

This has been worked around in mainline since but our folks
consider the fix potentially broken and propose the real one
(this seems to have not been fixed for -stable, not sure;
the patch is for 2.6.27.y).

PS: please CC me, not on the list [yet] -- and rather a proxy
than a developer. Thanks in advance.

----- Forwarded message from bugzilla-daemon/bugzilla.kernel.org -----

Date: Tue, 6 Jul 2010 11:38:15 GMT
From: bugzilla-daemon/bugzilla.kernel.org
To: shigorin/gmail.com
Subject: [Bug 15658] New: [PATCH] x86 constant_test_bit() prone to misoptimization with gcc-4.4

https://bugzilla.kernel.org/show_bug.cgi?id=15658

AssignedTo: platform_x86_64/kernel-bugs.osdl.org
CC: greg/kroah.com, vsu/altlinux.org, mingo/elte.hu


Created an attachment (id=25776)
--> (https://bugzilla.kernel.org/attachment.cgi?id=25776)
bluntly drop overcasting instead of fiddling with gcc's sense of humour

While debugging bit_spin_lock() hang, it was tracked down to gcc-4.4
misoptimization of constant_test_bit() when 'const volatile unsigned long
*addr' cast to 'unsigned long *' with subsequent unconditional jump to pause
(and not to the test) leading to hang.

Compiling with gcc-4.3 or disabling CONFIG_OPTIMIZE_INLINING yields inlined
constant_test_bit() and correct jump.

Other arches than asm-x86 (it's include/asm-x86/bitops.h) may implement this
slightly differently; 2.6.29 mitigates the misoptimization by changing the
function prototype (commit c4295fbb6048d85f0b41c5ced5cbf63f6811c46c) but
probably fixing the issue itself would be better.

Here's my translation of original analysis which led to the above conclusion
(Russian text in internal bugzilla; I'm no kernel hacker so take it with grain
of salt and blame me, not them):

---
Gory details
============

Slightly simplified bit_spin_lock() main part:
-------------------include/linux/bit_spinlock.h-----------------------

static inline void bit_spin_lock(int bitnum, unsigned long *addr)
{
<...>
while (test_and_set_bit_lock(bitnum, addr)) {
while (test_bit(bitnum, addr)) {
<...>
cpu_relax();
<...>
}
}
<...>
}
----------------------------------------------------------------------

Outer loop is an attempt to set the needed bit in atomic manner.
If the bit was already set, inner loop waits to reset it and so forth.

Assembly code generated by gcc-4.4 for the internal loop:
----------------------------------------------------------------------
$ rpm2cpio kernel-image-tmc-srv-2.6.27-tmc24.x86_64.rpm | cpio -idmv *jbd.ko
<...>
$ cd ./lib/modules/2.6.27-tmc-srv-tmc24/kernel/fs/jbd/
$ objdump -d --no-show-raw-insn jbd.ko | grep -A92 "<journal_get_undo_access>:"
| tail -7
240a: mov %rbx,%rsi
240d: mov $0x16,%edi
2412: callq 30 <constant_test_bit>
2417: test %eax,%eax
2419: je 2328 <journal_get_undo_access+0x48>
241f: pause
2421: jmp 241f <journal_get_undo_access+0x13f>
----------------------------------------------------------------------

Offsets:
240a, 240d - preparing arguments (bit no. 32, bitmap address in rsi)
2412 - call to constant_test_bit()
2417 - returned value test
2419 - break if 0
241f - otherwise pause and then
2421 - unconditional jump *but* not to the test but to the pause

Thus endless loop. That is, in C equivalent:
----------------------------------------------------------------------
static inline void bit_spin_lock(int bitnum, unsigned long *addr)
{
<...>
while (test_and_set_bit_lock(bitnum, addr)) {
if (test_bit(bitnum, addr))
while (1)
cpu_relax();
}
<...>
}
----------------------------------------------------------------------

When using gcc-4.3 or disabling CONFIG_OPTIMIZE_INLINING kernel configuration
parameter constant_test_bit() function gets inlined and the unconditional jump
in the inner loop goes before condition test as it should.

It is worth mentioning that bit_spin_lock() is used in several places and
not all of them result in wrong code being generated. For example,
end_buffer_async_read() has no problem with the inner loop:
----------------------------------------------------------------------
$ rpm2cpio kernel-modules-oprofile-tmc-srv-2.6.27-tmc24.x86_64.rpm | cpio -idmv
*vmlinux
<...>
$ cd ./lib/modules/2.6.27-tmc-srv-tmc24/
$ objdump -d --no-show-raw-insn vmlinux | grep -A106 "<end_buffer_async_read>:"
| tail -7
ffffffff802f55e2: mov %r13,%rsi
ffffffff802f55e5: mov $0x4,%edi
ffffffff802f55ea: callq ffffffff802f4360 <constant_test_bit>
ffffffff802f55ef: test %eax,%eax
ffffffff802f55f1: je ffffffff802f54f9 <end_buffer_async_read+0x49>
ffffffff802f55f7: pause
ffffffff802f55f9: jmp ffffffff802f55e2 <end_buffer_async_read+0x132>
----------------------------------------------------------------------


Hang scenario
=============
The below is my take at what leads from the above problem to the "hung"
system state observed.

1. At some moment the process changing the filesystem tries to "lock"
struct buffer_head by means of bit_spin_lock() but fails at first
and then goes to internal loop to wait for the buffer being freed.

2. Due to the problem described above the wait comes a bit long.

3. kjournald kernel thread wakes (because current transaction times out
or gets oversized) to commit the transaction into on-disk journal.
But upon changing it to T_LOCKED state kjournald notices that not all
atomic operations on the transactions are complete. Thus it cannot
continue and goes to sleep (in journal_commit_transaction()). Of course,
it waits for the process looping indefinitely as described in (1).

4. Subsequent filesystem change requests result in a new active (T_RUNNING)
transaction. As free space in that transaction ends up, all the processes
trying to change the filesystem march to sleep until kjournald would commit
that transaction. But it's not done with the previous one yet as in (3).

And so it all "hangs".


Solutions
=========
IMHO it looks very much like a bug in gcc-4.4. Adding 'volatile' qualifier
to the datatype of the data addressed by the second parameter to
bit_spin_lock() doesn't help. Maybe there's an upstream fix. Maybe it
gets fixed by some flags. Disabling CONFIG_OPTIMIZE_INLINING does help.
gcc-4.3 does compile things right, at least for JBD. Maybe there are other
possibilities... not me to judge.
---

I wasn't able to find anything similar upon a /quick/ google search, so posting
a new bug.

These folks from Massive Computing worked on this one:
* Volodymyr G. Lukiianyk (volodymyrgl/gmail.com) -- analysis
* Alexander Chumachenko (ledest/gmail.com) -- pin-up and fix

--- Comment #1 from Michael Shigorin <shigorin/gmail.com> 2010-07-06 11:38:07 ---
ping

--- Comment #2 from H. Peter Anvin <hpa@zytor.com> 2010-07-06 23:01:48 ---
I have seen in other contexts that gcc 4.4 seems to mishandle "const volatile".

----- End forwarded message -----

--
---- WBR, Michael Shigorin <mike@altlinux.ru>
------ Linux.Kiev http://www.linux.kiev.ua/
From ef5e7cca16541bd0d9fb36d13c7e40c6ecc4fd77 Mon Sep 17 00:00:00 2001
From: Led <led@altlinux.ru>
Date: Wed, 31 Mar 2010 21:09:20 +0300
Subject: [PATCH] x86: avoid 'constant_test_bit()' misoptimization due to cast to non-volatile

While debugging bit_spin_lock() hang, it was tracked down to gcc-4.4
misoptimization of constant_test_bit() when 'const volatile unsigned long *addr'
cast to 'unsigned long *' with subsequent unconditional jump to pause
(and not to the test) leading to hang.

Compiling with gcc-4.3 or disabling CONFIG_OPTIMIZE_INLINING yields inlined
constant_test_bit() and correct jump.

Other arches than asm-x86 may implement this slightly differently;
2.6.29 mitigates the misoptimization by changing the function prototype
(commit c4295fbb6048d85f0b41c5ced5cbf63f6811c46c) but probably fixing the issue
itself is better.
---
include/asm-x86/bitops.h | 2 +-
1 files changed, 1 insertions(+), 1 deletions(-)

diff --git a/include/asm-x86/bitops.h b/include/asm-x86/bitops.h
index cfb2b64..89352ed 100644
--- a/include/asm-x86/bitops.h
+++ b/include/asm-x86/bitops.h
@@ -295,7 +295,7 @@ static inline int test_and_change_bit(int nr, volatile unsigned long *addr)
static inline int constant_test_bit(int nr, const volatile unsigned long *addr)
{
return ((1UL << (nr % BITS_PER_LONG)) &
- (((unsigned long *)addr)[nr / BITS_PER_LONG])) != 0;
+ (addr[nr / BITS_PER_LONG])) != 0;
}

static inline int variable_test_bit(int nr, volatile const unsigned long *addr)
--
1.7.0.3
[unhandled content-type:application/pgp-signature]
\
 
 \ /
  Last update: 2010-08-28 00:11    [W:0.088 / U:42.500 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site