Optimize virt_spin_lock() to use simpler and faster:
atomic_try_cmpxchg(*ptr, &val, new)
instead of:
atomic_cmpxchg(*ptr, val, new) == val.
The x86 CMPXCHG instruction returns success in the ZF flag, so
this change saves a compare after the CMPXCHG.
Also optimize retry loop a bit. atomic_try_cmpxchg() fails iff
&lock->val != 0, so there is no need to load and compare the
lock value again - cpu_relax() can be unconditinally called in
this case. This allows us to generate optimized:
1f: ba 01 00 00 00 mov $0x1,%edx
24: 8b 03 mov (%rbx),%eax
26: 85 c0 test %eax,%eax
28: 75 63 jne 8d <...>
2a: f0 0f b1 13 lock cmpxchg %edx,(%rbx)
2e: 75 5d jne 8d <...>
..
8d: f3 90 pause
8f: eb 93 jmp 24 <...>
instead of:
1f: ba 01 00 00 00 mov $0x1,%edx
24: 8b 03 mov (%rbx),%eax
26: 85 c0 test %eax,%eax
28: 75 13 jne 3d <...>
2a: f0 0f b1 13 lock cmpxchg %edx,(%rbx)
2e: 85 c0 test %eax,%eax
30: 75 f2 jne 24 <...>
..
3d: f3 90 pause
3f: eb e3 jmp 24 <...>
Signed-off-by: Uros Bizjak <[email protected]>
Cc: Thomas Gleixner <[email protected]>
Cc: Ingo Molnar <[email protected]>
Cc: Borislav Petkov <[email protected]>
Cc: Dave Hansen <[email protected]>
Cc: "H. Peter Anvin" <[email protected]>
---
arch/x86/include/asm/qspinlock.h | 13 +++++++++----
1 file changed, 9 insertions(+), 4 deletions(-)
diff --git a/arch/x86/include/asm/qspinlock.h b/arch/x86/include/asm/qspinlock.h
index cde8357bb226..a053c1293975 100644
--- a/arch/x86/include/asm/qspinlock.h
+++ b/arch/x86/include/asm/qspinlock.h
@@ -85,6 +85,8 @@ DECLARE_STATIC_KEY_TRUE(virt_spin_lock_key);
#define virt_spin_lock virt_spin_lock
static inline bool virt_spin_lock(struct qspinlock *lock)
{
+ int val;
+
if (!static_branch_likely(&virt_spin_lock_key))
return false;
@@ -94,10 +96,13 @@ static inline bool virt_spin_lock(struct qspinlock *lock)
* horrible lock 'holder' preemption issues.
*/
- do {
- while (atomic_read(&lock->val) != 0)
- cpu_relax();
- } while (atomic_cmpxchg(&lock->val, 0, _Q_LOCKED_VAL) != 0);
+ __retry:
+ val = atomic_read(&lock->val);
+
+ if (val || !atomic_try_cmpxchg(&lock->val, &val, _Q_LOCKED_VAL)) {
+ cpu_relax();
+ goto __retry;
+ }
return true;
}
--
2.44.0
The following commit has been merged into the locking/core branch of tip:
Commit-ID: 94af3a04e3f386d4f060d903826e85aa006ce252
Gitweb: https://git.kernel.org/tip/94af3a04e3f386d4f060d903826e85aa006ce252
Author: Uros Bizjak <[email protected]>
AuthorDate: Mon, 22 Apr 2024 14:00:38 +02:00
Committer: Ingo Molnar <[email protected]>
CommitterDate: Wed, 24 Apr 2024 11:46:28 +02:00
locking/qspinlock/x86: Micro-optimize virt_spin_lock()
Optimize virt_spin_lock() to use simpler and faster:
atomic_try_cmpxchg(*ptr, &val, new)
instead of:
atomic_cmpxchg(*ptr, val, new) == val
The x86 CMPXCHG instruction returns success in the ZF flag, so
this change saves a compare after the CMPXCHG.
Also optimize retry loop a bit. atomic_try_cmpxchg() fails iff
&lock->val != 0, so there is no need to load and compare the
lock value again - cpu_relax() can be unconditinally called in
this case. This allows us to generate optimized:
1f: ba 01 00 00 00 mov $0x1,%edx
24: 8b 03 mov (%rbx),%eax
26: 85 c0 test %eax,%eax
28: 75 63 jne 8d <...>
2a: f0 0f b1 13 lock cmpxchg %edx,(%rbx)
2e: 75 5d jne 8d <...>
..
8d: f3 90 pause
8f: eb 93 jmp 24 <...>
instead of:
1f: ba 01 00 00 00 mov $0x1,%edx
24: 8b 03 mov (%rbx),%eax
26: 85 c0 test %eax,%eax
28: 75 13 jne 3d <...>
2a: f0 0f b1 13 lock cmpxchg %edx,(%rbx)
2e: 85 c0 test %eax,%eax
30: 75 f2 jne 24 <...>
..
3d: f3 90 pause
3f: eb e3 jmp 24 <...>
Signed-off-by: Uros Bizjak <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>
Cc: Waiman Long <[email protected]>
Cc: Linus Torvalds <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
---
arch/x86/include/asm/qspinlock.h | 13 +++++++++----
1 file changed, 9 insertions(+), 4 deletions(-)
diff --git a/arch/x86/include/asm/qspinlock.h b/arch/x86/include/asm/qspinlock.h
index cde8357..a053c12 100644
--- a/arch/x86/include/asm/qspinlock.h
+++ b/arch/x86/include/asm/qspinlock.h
@@ -85,6 +85,8 @@ DECLARE_STATIC_KEY_TRUE(virt_spin_lock_key);
#define virt_spin_lock virt_spin_lock
static inline bool virt_spin_lock(struct qspinlock *lock)
{
+ int val;
+
if (!static_branch_likely(&virt_spin_lock_key))
return false;
@@ -94,10 +96,13 @@ static inline bool virt_spin_lock(struct qspinlock *lock)
* horrible lock 'holder' preemption issues.
*/
- do {
- while (atomic_read(&lock->val) != 0)
- cpu_relax();
- } while (atomic_cmpxchg(&lock->val, 0, _Q_LOCKED_VAL) != 0);
+ __retry:
+ val = atomic_read(&lock->val);
+
+ if (val || !atomic_try_cmpxchg(&lock->val, &val, _Q_LOCKED_VAL)) {
+ cpu_relax();
+ goto __retry;
+ }
return true;
}