Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1572641 > unrolled thread

[PATCH] x86, locking: Inline *_unlock_bh & *_unlock_irqrestore

Back to article view | Back to linux.kernel

Contents

  [PATCH] x86, locking: Inline *_unlock_bh & *_unlock_irqrestore Waiman Long <longman@redhat.com> - 2017-02-02 19:50 +0100
    Re: [PATCH] x86, locking: Inline *_unlock_bh & *_unlock_irqrestore Thomas Gleixner <tglx@linutronix.de> - 2017-02-10 17:10 +0100

#1572641 — [PATCH] x86, locking: Inline *_unlock_bh & *_unlock_irqrestore

For spinlock and read-write locks, both *_unlock() and *_unlock_irq()
functions are inlined if !PREEMPT. Whereas the *_unlock_bh() and
*unlock_irqrestore() are not inlined as that will increase the size
of the kernel binary.

Given the fact that the PV qspinlock unlock call is a callee-saved
function pointer, the unlock function is essentially a leaf function
if !PREEMPT is true under all circumstances. Inlining it will enable
the compiler to do much better optimization around the the unlock
call sites. Similarly for the read-write locks.

To unleash this additional performance, all these unlock functions
are now inlined for x86-64 kernel where kernel size is usually less
of a concern. For 32-bit kernel, it is assumed that the focus will
be a bit more on kernel size optimization and so those functions are
not inlined.

With 4.9.6 kernel source and gcc 4.8.5-11 compiler, the text size
of the base vmlinux binary increased from 8527918 bytes to 8462235
bytes. That was an increase of 0.78%.

On a 2-socket 24-core 48-thread system, the performance of the AIM7
high_systime workload (1000 users) increased from 209818.43 jobs/min
to 213333.33 jobs/min. That was an increase of 1.68%.

Signed-off-by: Waiman Long <longman@redhat.com>
---
 arch/x86/Kconfig | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
index 7b6fd68..a0fbed7 100644
--- a/arch/x86/Kconfig
+++ b/arch/x86/Kconfig
@@ -23,6 +23,12 @@ config X86_64
 	depends on 64BIT
 	# Options that are inherently 64-bit kernel only:
 	select ARCH_HAS_GIGANTIC_PAGE
+	select ARCH_INLINE_READ_UNLOCK_BH		if !PREEMPT
+	select ARCH_INLINE_READ_UNLOCK_IRQRESTORE	if !PREEMPT
+	select ARCH_INLINE_SPIN_UNLOCK_BH		if !PREEMPT
+	select ARCH_INLINE_SPIN_UNLOCK_IRQRESTORE	if !PREEMPT
+	select ARCH_INLINE_WRITE_UNLOCK_BH		if !PREEMPT
+	select ARCH_INLINE_WRITE_UNLOCK_IRQRESTORE	if !PREEMPT
 	select ARCH_SUPPORTS_INT128
 	select ARCH_USE_CMPXCHG_LOCKREF
 	select HAVE_ARCH_SOFT_DIRTY
-- 
1.8.3.1

[toc] | [next] | [standalone]

#1578579

On Thu, 2 Feb 2017, Waiman Long wrote:

> For spinlock and read-write locks, both *_unlock() and *_unlock_irq()
> functions are inlined if !PREEMPT. Whereas the *_unlock_bh() and
> *unlock_irqrestore() are not inlined as that will increase the size
> of the kernel binary.
> 
> Given the fact that the PV qspinlock unlock call is a callee-saved
> function pointer, the unlock function is essentially a leaf function
> if !PREEMPT is true under all circumstances. Inlining it will enable
> the compiler to do much better optimization around the the unlock
> call sites. Similarly for the read-write locks.
> 
> To unleash this additional performance, all these unlock functions
> are now inlined for x86-64 kernel where kernel size is usually less
> of a concern. For 32-bit kernel, it is assumed that the focus will
> be a bit more on kernel size optimization and so those functions are
> not inlined.
> 
> With 4.9.6 kernel source and gcc 4.8.5-11 compiler, the text size
> of the base vmlinux binary increased from 8527918 bytes to 8462235
> bytes. That was an increase of 0.78%.

Increased from 8527918 bytes to 8462235? 

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web