2014-04-30 01:11:03

by Behan Webster

[permalink] [raw]
Subject: [PATCH] mbcache: LLVMLinux: Remove double calculation from mbcache

From: Mark Charlebois <[email protected]>

The call to __builtin_log2 presumes there is a
double log2(double x) function defined in the kernel.

The call to hash_log is a call to hash_64 which is
defined in include/linux/hash.h

static __always_inline u64 hash_64(u64 val, unsigned int bits)

That means that __builtin_log2(NR_BG_LOCKS) is converting
NR_BG_LOCKS to a double and returning a double and then that
is converted to an unsigned int.

Using ilog2 is much more appropriate and efficient.

Another side effect of using __builtin_log2 is that is uses
__aeabi_* functions for ARM that require linking with libgcc.a.

Author: Mark Charlebois <[email protected]>
Signed-off-by: Mark Charlebois <[email protected]>
Signed-off-by: Behan Webster <[email protected]>
---
fs/mbcache.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/fs/mbcache.c b/fs/mbcache.c
index bf166e3..2c0752b 100644
--- a/fs/mbcache.c
+++ b/fs/mbcache.c
@@ -93,7 +93,7 @@

#define MB_CACHE_WRITER ((unsigned short)~0U >> 1)

-#define MB_CACHE_ENTRY_LOCK_BITS __builtin_log2(NR_BG_LOCKS)
+#define MB_CACHE_ENTRY_LOCK_BITS ilog2(NR_BG_LOCKS)
#define MB_CACHE_ENTRY_LOCK_INDEX(ce) \
(hash_long((unsigned long)ce, MB_CACHE_ENTRY_LOCK_BITS))

--
1.8.3.2


2014-04-30 01:14:32

by H. Peter Anvin

[permalink] [raw]
Subject: Re: [PATCH] mbcache: LLVMLinux: Remove double calculation from mbcache

I suspect this will generate FP instructions on x86 which will corrupt user space state. This is thus a critical bug!!

On April 29, 2014 6:10:39 PM PDT, [email protected] wrote:
>From: Mark Charlebois <[email protected]>
>
>The call to __builtin_log2 presumes there is a
>double log2(double x) function defined in the kernel.
>
>The call to hash_log is a call to hash_64 which is
>defined in include/linux/hash.h
>
>static __always_inline u64 hash_64(u64 val, unsigned int bits)
>
>That means that __builtin_log2(NR_BG_LOCKS) is converting
>NR_BG_LOCKS to a double and returning a double and then that
>is converted to an unsigned int.
>
>Using ilog2 is much more appropriate and efficient.
>
>Another side effect of using __builtin_log2 is that is uses
>__aeabi_* functions for ARM that require linking with libgcc.a.
>
>Author: Mark Charlebois <[email protected]>
>Signed-off-by: Mark Charlebois <[email protected]>
>Signed-off-by: Behan Webster <[email protected]>
>---
> fs/mbcache.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
>diff --git a/fs/mbcache.c b/fs/mbcache.c
>index bf166e3..2c0752b 100644
>--- a/fs/mbcache.c
>+++ b/fs/mbcache.c
>@@ -93,7 +93,7 @@
>
> #define MB_CACHE_WRITER ((unsigned short)~0U >> 1)
>
>-#define MB_CACHE_ENTRY_LOCK_BITS __builtin_log2(NR_BG_LOCKS)
>+#define MB_CACHE_ENTRY_LOCK_BITS ilog2(NR_BG_LOCKS)
> #define MB_CACHE_ENTRY_LOCK_INDEX(ce) \
> (hash_long((unsigned long)ce, MB_CACHE_ENTRY_LOCK_BITS))
>

--
Sent from my mobile phone. Please pardon brevity and lack of formatting.

2014-04-30 05:19:56

by H. Peter Anvin

[permalink] [raw]
Subject: Re: [PATCH] mbcache: LLVMLinux: Remove double calculation from mbcache

On 04/29/2014 06:13 PM, H. Peter Anvin wrote:
> I suspect this will generate FP instructions on x86 which will corrupt user space state. This is thus a critical bug!!

It looks like at least x86-64 gcc does the evaluation at compile time
and there aren't any fp instructions left in the code. However, it
should still be fixed.

-hpa