[RFC] xfs: work around unlikely() profiler glitch
diff mbox

Message ID 20170125140821.2677725-1-arnd@arndb.de
State New
Headers show

Commit Message

Arnd Bergmann Jan. 25, 2017, 2:08 p.m. UTC
After a patch from Steven Rostedt to rework the profiling of
unlikely(), we get a couple of link errors in XFS on ARM:

fs/xfs/libxfs/xfs_bmap.o: In function `xfs_bmap_btalloc':
xfs_bmap.c:(.text.xfs_bmap_btalloc+0x8cc): undefined reference to `____ilog2_NaN'
xfs_bmap.c:(.text.xfs_bmap_btalloc+0xb34): undefined reference to `__aeabi_uldivmod'
xfs_bmap.c:(.text.xfs_bmap_btalloc+0xb7c): undefined reference to `__aeabi_uldivmod'

As I understand the problem, this is a result of the combination
of __builtin_constant_p() usage in three places: unlikely(),
xfs_bmap_btalloc(), and __div64_32(), where the latter attempts to use
__builtin_constant_p() to decide whether to either call an out-of-line
function for variable divisors or an optimized version for constant
divisors.

As I understand it, here we end up with a case where gcc has determined
that the argument is constant in some cases but variable in other cases,
and tries to split up the handling by implementing both paths and deciding
at runtime.

This seems to be related to the problem fixed in commit b33c8ff4431a
("tracing: Fix freak link error caused by branch tracer").

I don't know why exactly it goes wrong now, but we end up getting object
code emitted for the case that should have been eliminated after the
__builtin_constant_p() check. I don't think we'd ever run into the branches
to ____ilog2_NaN or __aeabi_uldivmod, but we get a link error anyway.

Fixes: d45ae1f7041a ("tracing: Process constants for (un)likely() profiler")i
Cc: Steven Rostedt (VMware) <rostedt@goodmis.org>
Cc: Nicolas Pitre <nico@linaro.org>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
---
 fs/xfs/libxfs/xfs_bmap.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Comments

Christoph Hellwig Jan. 25, 2017, 3:09 p.m. UTC | #1
On Wed, Jan 25, 2017 at 03:08:10PM +0100, Arnd Bergmann wrote:
> diff --git a/fs/xfs/libxfs/xfs_bmap.c b/fs/xfs/libxfs/xfs_bmap.c
> index d22f7930eb75..dca3ddd737d4 100644
> --- a/fs/xfs/libxfs/xfs_bmap.c
> +++ b/fs/xfs/libxfs/xfs_bmap.c
> @@ -3629,7 +3629,7 @@ xfs_bmap_btalloc(
>  		align = xfs_get_cowextsz_hint(ap->ip);
>  	else if (xfs_alloc_is_userdata(ap->datatype))
>  		align = xfs_get_extsz_hint(ap->ip);
> -	if (unlikely(align)) {
> +	if (unlikely_notrace(align)) {
>  		error = xfs_bmap_extsize_align(mp, &ap->got, &ap->prev,
>  						align, 0, ap->eof, 0, ap->conv,
>  						&ap->offset, &ap->length);

The unlikely calls on align in xfs_bmap_btalloc should simply be
removed.  They aren't actually unlikely for many workloads.  I have
a patch in my queue that I can expedite based on your report.
--
To unsubscribe from this list: send the line "unsubscribe linux-xfs" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Arnd Bergmann Jan. 25, 2017, 3:15 p.m. UTC | #2
On Wed, Jan 25, 2017 at 4:09 PM, Christoph Hellwig <hch@infradead.org> wrote:
> On Wed, Jan 25, 2017 at 03:08:10PM +0100, Arnd Bergmann wrote:
>> diff --git a/fs/xfs/libxfs/xfs_bmap.c b/fs/xfs/libxfs/xfs_bmap.c
>> index d22f7930eb75..dca3ddd737d4 100644
>> --- a/fs/xfs/libxfs/xfs_bmap.c
>> +++ b/fs/xfs/libxfs/xfs_bmap.c
>> @@ -3629,7 +3629,7 @@ xfs_bmap_btalloc(
>>               align = xfs_get_cowextsz_hint(ap->ip);
>>       else if (xfs_alloc_is_userdata(ap->datatype))
>>               align = xfs_get_extsz_hint(ap->ip);
>> -     if (unlikely(align)) {
>> +     if (unlikely_notrace(align)) {
>>               error = xfs_bmap_extsize_align(mp, &ap->got, &ap->prev,
>>                                               align, 0, ap->eof, 0, ap->conv,
>>                                               &ap->offset, &ap->length);
>
> The unlikely calls on align in xfs_bmap_btalloc should simply be
> removed.  They aren't actually unlikely for many workloads.  I have
> a patch in my queue that I can expedite based on your report.

That would defines help, thanks!

I also noticed that my patch wouldn't work, as unlikely_notrace() is not
defined unless we are actually tracing, so while it fixes some rare
configurations, it breaks all the configurations that matter.

    Arnd
--
To unsubscribe from this list: send the line "unsubscribe linux-xfs" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Darrick J. Wong Jan. 25, 2017, 4:24 p.m. UTC | #3
On Wed, Jan 25, 2017 at 07:09:29AM -0800, Christoph Hellwig wrote:
> On Wed, Jan 25, 2017 at 03:08:10PM +0100, Arnd Bergmann wrote:
> > diff --git a/fs/xfs/libxfs/xfs_bmap.c b/fs/xfs/libxfs/xfs_bmap.c
> > index d22f7930eb75..dca3ddd737d4 100644
> > --- a/fs/xfs/libxfs/xfs_bmap.c
> > +++ b/fs/xfs/libxfs/xfs_bmap.c
> > @@ -3629,7 +3629,7 @@ xfs_bmap_btalloc(
> >  		align = xfs_get_cowextsz_hint(ap->ip);
> >  	else if (xfs_alloc_is_userdata(ap->datatype))
> >  		align = xfs_get_extsz_hint(ap->ip);
> > -	if (unlikely(align)) {
> > +	if (unlikely_notrace(align)) {
> >  		error = xfs_bmap_extsize_align(mp, &ap->got, &ap->prev,
> >  						align, 0, ap->eof, 0, ap->conv,
> >  						&ap->offset, &ap->length);
> 
> The unlikely calls on align in xfs_bmap_btalloc should simply be
> removed.  They aren't actually unlikely for many workloads.  I have
> a patch in my queue that I can expedite based on your report.

I was thinking exactly the same thing.  Since it breaks the build
somewhere, can you send a oneliner patch so I can roll it into the rc6
fixes?

--D

> --
> To unsubscribe from this list: send the line "unsubscribe linux-xfs" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-xfs" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Patch
diff mbox

diff --git a/fs/xfs/libxfs/xfs_bmap.c b/fs/xfs/libxfs/xfs_bmap.c
index d22f7930eb75..dca3ddd737d4 100644
--- a/fs/xfs/libxfs/xfs_bmap.c
+++ b/fs/xfs/libxfs/xfs_bmap.c
@@ -3629,7 +3629,7 @@  xfs_bmap_btalloc(
 		align = xfs_get_cowextsz_hint(ap->ip);
 	else if (xfs_alloc_is_userdata(ap->datatype))
 		align = xfs_get_extsz_hint(ap->ip);
-	if (unlikely(align)) {
+	if (unlikely_notrace(align)) {
 		error = xfs_bmap_extsize_align(mp, &ap->got, &ap->prev,
 						align, 0, ap->eof, 0, ap->conv,
 						&ap->offset, &ap->length);