Date:	Thu, 10 Jul 2014 17:00:03 +0200
From:	Vlastimil Babka <vbabka@...e.cz>
To:	Jakub Jelinek <jakub@...hat.com>, mingo@...nel.org, hpa@...or.com,
	linux-kernel@...r.kernel.org, torvalds@...ux-foundation.org,
	peterz@...radead.org, stable@...r.kernel.org,
	steven@...inklabs.net, rostedt@...dmis.org,
	akpm@...ux-foundation.org, tglx@...utronix.de, oleg@...hat.com,
	rth@...ddle.net
CC:	linux-tip-commits@...r.kernel.org, Michal Hocko <mhocko@...e.cz>
Subject: Re: [tip:x86/urgent] compiler/gcc4: Make quirk for asm_volatile_goto(
 ) unconditional

On 02/13/2014 12:55 PM, Jakub Jelinek wrote:
> On Thu, Feb 13, 2014 at 03:37:08AM -0800, tip-bot for Steven Noonan wrote:
>> Commit-ID:  a9f180345f5378ac87d80ed0bea55ba421d83859
>> Gitweb:     http://git.kernel.org/tip/a9f180345f5378ac87d80ed0bea55ba421d83859
>> Author:     Steven Noonan <steven@...inklabs.net>
>> AuthorDate: Wed, 12 Feb 2014 23:01:07 -0800
>> Committer:  Ingo Molnar <mingo@...nel.org>
>> CommitDate: Thu, 13 Feb 2014 12:34:05 +0100
>> 
>> compiler/gcc4: Make quirk for asm_volatile_goto() unconditional
>> 
>> I started noticing problems with KVM guest destruction on Linux
>> 3.12+, where guest memory wasn't being cleaned up. I bisected it
>> down to the commit introducing the new 'asm goto'-based atomics,
>> and found this quirk was later applied to those.
>> 
>> Unfortunately, even with GCC 4.8.2 (which ostensibly fixed the
>> known 'asm goto' bug) I am still getting some kind of
>> miscompilation. If I enable the asm_volatile_goto quirk for my
>> compiler, KVM guests are destroyed correctly and the memory is
>> cleaned up.
> 
> BTW, which exact 4.8.2 were you using?
> The last known asm goto bug has been fixed on October, 10th, 2013:
> http://gcc.gnu.org/PR58670

FYI, we have hit a very similar kind of memory leak (orphaned THP pages staying on LRU
with elevated page_count) due to the quirk patch missing in a backport, and tracked the
problem down to put_compound_page() which contains this:

                if (put_page_testzero(page_head))
                        VM_BUG_ON_PAGE(1, page_head);

The problem is that with DEBUG_VM disabled, the 'then' part of this 'if' is a no-op which
makes gcc optimize out the whole put_page_testzero operation. The quirk happens to prevent
this.

There is a new gcc bug filed for this:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61772	

> so before the October, 16th, 2013 4.8.2 release.  But already since
> May 31th, 2013 the tip of the 4.8 GCC branch has been announcing itself
> as 4.8.2 prerelease.  While some distribution versions of GCC announce
> themselves as the new version only starting from the release date,
> i.e. snapshots in between 4.8.1 release and 4.8.2 release announce
> themselves as 4.8.1, in other distributions or upstream it announces itself
> as 4.8.2.  So, if you are using the latter and a snapshot in between May
> 31th, 2013 and October, 10th, 2013, then you could see gcc patchlevel 2,
> yet have a gcc with that bug unfixed.
> So, if the kernel doesn't use a runtime test/configure test to check for
> this issue, but instead just relies on the patchlevel version, the only
> safe way would be to look for GCC >= 4.9 or GCC 4.8 with patchlevel > 2
> rather than > 1.
> 
> 	Jakub
> --
> 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/
> 

--
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