Paravirt spinlock clears slowpath flag after doing unlock.
As explained by Linus currently it does:
prev = *lock;
add_smp(&lock->tickets.head, TICKET_LOCK_INC);
/* add_smp() is a full mb() */
if (unlikely(lock->tickets.tail & TICKET_SLOWPATH_FLAG))
__ticket_unlock_slowpath(lock, prev);
which is *exactly* the kind of things you cannot do with spinlocks,
because after you've done the "add_smp()" and released the spinlock
for the fast-path, you can't access the spinlock any more. Exactly
because a fast-path lock might come in, and release the whole data
structure.
Linus suggested that we should not do any writes to lock after unlock(),
and we can move slowpath clearing to fastpath lock.
So this patch implements the fix with:
1. Moving slowpath flag to head (Oleg).
2. use xadd to avoid read/write after unlock that checks the need for
unlock_kick (Linus).
Result:
setup: 16core (32 cpu +ht sandy bridge 8GB 16vcpu guest)
benchmark overcommit %improve
kernbench 1x -0.13
kernbench 2x 0.02
dbench 1x -1.77
dbench 2x -0.63
[Jeremy: hinted missing TICKET_LOCK_INC for kick]
[Oleg: Moving slowpath flag to head, ticket_equals idea]
Reported-by: Sasha Levin <[email protected]>
Suggested-by: Linus Torvalds <[email protected]>
Signed-off-by: Raghavendra K T <[email protected]>
---
arch/x86/include/asm/spinlock.h | 87 ++++++++++++++++++++---------------------
arch/x86/kernel/kvm.c | 4 +-
2 files changed, 45 insertions(+), 46 deletions(-)
potential TODO:
* The whole patch be splitted into, 1. move slowpath flag
2. fix memory corruption in completion problem ??
* May be we could directly pass inc for __ticket_check_and_clear_slowpath
but I hope current code is more readable.
Changes since V2:
- Move the slowpath flag to head, this enables xadd usage in unlock code
and inturn we can get rid of read/write after unlock (Oleg)
- usage of ticket_equals (Oleg)
Changes since V1:
- Add missing TICKET_LOCK_INC before unlock kick (fixes hang in overcommit: Jeremy).
- Remove SLOWPATH_FLAG clearing in fast lock. (Jeremy)
- clear SLOWPATH_FLAG in arch_spin_value_unlocked during comparison.
Note: The current implementation is still based on avoid writing after unlock.
we could still have potential invalid memory read. (Sasha)
Result:
setup: 16core (32 cpu +ht sandy bridge 8GB 16vcpu guest)
base = 3_19_rc7
3_19_rc7_spinfix_v3
+-----------+-----------+-----------+------------+-----------+
kernbench (Time taken in sec lower is better)
+-----------+-----------+-----------+------------+-----------+
base %stdev patched %stdev %improve
+-----------+-----------+-----------+------------+-----------+
1x 54.2300 3.0652 54.3008 4.0366 -0.13056
2x 90.1883 5.5509 90.1650 6.4336 0.02583
+-----------+-----------+-----------+------------+-----------+
+-----------+-----------+-----------+------------+-----------+
dbench (Throughput higher is better)
+-----------+-----------+-----------+------------+-----------+
base %stdev patched %stdev %improve
+-----------+-----------+-----------+------------+-----------+
1x 7029.9188 2.5952 6905.0712 4.4737 -1.77595
2x 3254.2075 14.8291 3233.7137 26.8784 -0.62976
+-----------+-----------+-----------+------------+-----------+
(here is the result I got from the patches, I believe there may
be some small overhead from xadd etc, but overall looks fine but
a thorough test may be needed)
diff --git a/arch/x86/include/asm/spinlock.h b/arch/x86/include/asm/spinlock.h
index 625660f..9697b45 100644
--- a/arch/x86/include/asm/spinlock.h
+++ b/arch/x86/include/asm/spinlock.h
@@ -46,7 +46,7 @@ static __always_inline bool static_key_false(struct static_key *key);
static inline void __ticket_enter_slowpath(arch_spinlock_t *lock)
{
- set_bit(0, (volatile unsigned long *)&lock->tickets.tail);
+ set_bit(0, (volatile unsigned long *)&lock->tickets.head);
}
#else /* !CONFIG_PARAVIRT_SPINLOCKS */
@@ -60,10 +60,30 @@ static inline void __ticket_unlock_kick(arch_spinlock_t *lock,
}
#endif /* CONFIG_PARAVIRT_SPINLOCKS */
+static inline int __tickets_equal(__ticket_t one, __ticket_t two)
+{
+ return !((one ^ two) & ~TICKET_SLOWPATH_FLAG);
+}
+
+static inline void __ticket_check_and_clear_slowpath(arch_spinlock_t *lock,
+ __ticket_t head)
+{
+ if (head & TICKET_SLOWPATH_FLAG) {
+ arch_spinlock_t old, new;
+
+ old.tickets.head = head;
+ new.tickets.head = head & ~TICKET_SLOWPATH_FLAG;
+ old.tickets.tail = new.tickets.head + TICKET_LOCK_INC;
+ new.tickets.tail = old.tickets.tail;
+
+ /* try to clear slowpath flag when there are no contenders */
+ cmpxchg(&lock->head_tail, old.head_tail, new.head_tail);
+ }
+}
static __always_inline int arch_spin_value_unlocked(arch_spinlock_t lock)
{
- return lock.tickets.head == lock.tickets.tail;
+ return __tickets_equal(lock.tickets.head, lock.tickets.tail);
}
/*
@@ -87,18 +107,22 @@ static __always_inline void arch_spin_lock(arch_spinlock_t *lock)
if (likely(inc.head == inc.tail))
goto out;
- inc.tail &= ~TICKET_SLOWPATH_FLAG;
for (;;) {
unsigned count = SPIN_THRESHOLD;
do {
- if (READ_ONCE(lock->tickets.head) == inc.tail)
- goto out;
+ inc.head = READ_ONCE(lock->tickets.head);
+ if (__tickets_equal(inc.head, inc.tail))
+ goto clear_slowpath;
cpu_relax();
} while (--count);
__ticket_lock_spinning(lock, inc.tail);
}
-out: barrier(); /* make sure nothing creeps before the lock is taken */
+clear_slowpath:
+ if (TICKET_SLOWPATH_FLAG)
+ __ticket_check_and_clear_slowpath(lock, inc.head);
+out:
+ barrier(); /* make sure nothing creeps before the lock is taken */
}
static __always_inline int arch_spin_trylock(arch_spinlock_t *lock)
@@ -106,56 +130,30 @@ static __always_inline int arch_spin_trylock(arch_spinlock_t *lock)
arch_spinlock_t old, new;
old.tickets = READ_ONCE(lock->tickets);
- if (old.tickets.head != (old.tickets.tail & ~TICKET_SLOWPATH_FLAG))
+ if (!__tickets_equal(old.tickets.head, old.tickets.tail))
return 0;
new.head_tail = old.head_tail + (TICKET_LOCK_INC << TICKET_SHIFT);
+ new.head_tail &= ~TICKET_SLOWPATH_FLAG;
/* cmpxchg is a full barrier, so nothing can move before it */
return cmpxchg(&lock->head_tail, old.head_tail, new.head_tail) == old.head_tail;
}
-static inline void __ticket_unlock_slowpath(arch_spinlock_t *lock,
- arch_spinlock_t old)
-{
- arch_spinlock_t new;
-
- BUILD_BUG_ON(((__ticket_t)NR_CPUS) != NR_CPUS);
-
- /* Perform the unlock on the "before" copy */
- old.tickets.head += TICKET_LOCK_INC;
-
- /* Clear the slowpath flag */
- new.head_tail = old.head_tail & ~(TICKET_SLOWPATH_FLAG << TICKET_SHIFT);
-
- /*
- * If the lock is uncontended, clear the flag - use cmpxchg in
- * case it changes behind our back though.
- */
- if (new.tickets.head != new.tickets.tail ||
- cmpxchg(&lock->head_tail, old.head_tail,
- new.head_tail) != old.head_tail) {
- /*
- * Lock still has someone queued for it, so wake up an
- * appropriate waiter.
- */
- __ticket_unlock_kick(lock, old.tickets.head);
- }
-}
-
static __always_inline void arch_spin_unlock(arch_spinlock_t *lock)
{
if (TICKET_SLOWPATH_FLAG &&
- static_key_false(¶virt_ticketlocks_enabled)) {
- arch_spinlock_t prev;
+ static_key_false(¶virt_ticketlocks_enabled)) {
+ __ticket_t head;
- prev = *lock;
- add_smp(&lock->tickets.head, TICKET_LOCK_INC);
+ BUILD_BUG_ON(((__ticket_t)NR_CPUS) != NR_CPUS);
- /* add_smp() is a full mb() */
+ head = xadd(&lock->tickets.head, TICKET_LOCK_INC);
- if (unlikely(lock->tickets.tail & TICKET_SLOWPATH_FLAG))
- __ticket_unlock_slowpath(lock, prev);
+ if (unlikely(head & TICKET_SLOWPATH_FLAG)) {
+ head &= ~TICKET_SLOWPATH_FLAG;
+ __ticket_unlock_kick(lock, (head + TICKET_LOCK_INC));
+ }
} else
__add(&lock->tickets.head, TICKET_LOCK_INC, UNLOCK_LOCK_PREFIX);
}
@@ -164,7 +162,7 @@ static inline int arch_spin_is_locked(arch_spinlock_t *lock)
{
struct __raw_tickets tmp = READ_ONCE(lock->tickets);
- return tmp.tail != tmp.head;
+ return tmp.tail != (tmp.head & ~TICKET_SLOWPATH_FLAG);
}
static inline int arch_spin_is_contended(arch_spinlock_t *lock)
@@ -191,8 +189,7 @@ static inline void arch_spin_unlock_wait(arch_spinlock_t *lock)
* We need to check "unlocked" in a loop, tmp.head == head
* can be false positive because of overflow.
*/
- if (tmp.head == (tmp.tail & ~TICKET_SLOWPATH_FLAG) ||
- tmp.head != head)
+ if (__tickets_equal(tmp.head, tmp.tail) || tmp.head != head)
break;
cpu_relax();
diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c
index 94f6434..e758b46 100644
--- a/arch/x86/kernel/kvm.c
+++ b/arch/x86/kernel/kvm.c
@@ -727,6 +727,7 @@ __visible void kvm_lock_spinning(struct arch_spinlock *lock, __ticket_t want)
int cpu;
u64 start;
unsigned long flags;
+ __ticket_t head;
if (in_nmi())
return;
@@ -772,7 +773,8 @@ __visible void kvm_lock_spinning(struct arch_spinlock *lock, __ticket_t want)
* check again make sure it didn't become free while
* we weren't looking.
*/
- if (ACCESS_ONCE(lock->tickets.head) == want) {
+ head = ACCESS_ONCE(lock->tickets.head);
+ if (__tickets_equal(head, want)) {
add_stats(TAKEN_SLOW_PICKUP, 1);
goto out;
}
--
1.7.11.7
On 02/12, Raghavendra K T wrote:
>
> @@ -772,7 +773,8 @@ __visible void kvm_lock_spinning(struct arch_spinlock *lock, __ticket_t want)
> * check again make sure it didn't become free while
> * we weren't looking.
> */
> - if (ACCESS_ONCE(lock->tickets.head) == want) {
> + head = ACCESS_ONCE(lock->tickets.head);
> + if (__tickets_equal(head, want)) {
> add_stats(TAKEN_SLOW_PICKUP, 1);
While at it, perhaps it makes sense to s/ACCESS_ONCE/READ_ONCE/ but this
is cosmetic.
We also need to change another user of enter_slow_path, xen_lock_spinning()
in arch/x86/xen/spinlock.c.
Other than that looks correct at first glance... but this is up to
maintainers.
Oleg.
On 02/12, Raghavendra K T wrote:
>
> @@ -191,8 +189,7 @@ static inline void arch_spin_unlock_wait(arch_spinlock_t *lock)
> * We need to check "unlocked" in a loop, tmp.head == head
> * can be false positive because of overflow.
> */
> - if (tmp.head == (tmp.tail & ~TICKET_SLOWPATH_FLAG) ||
> - tmp.head != head)
> + if (__tickets_equal(tmp.head, tmp.tail) || tmp.head != head)
> break;
Ah, it seems that "tmp.head != head" should be turned into
!__tickets_equal(), no?
Suppose that TICKET_SLOWPATH_FLAG is set after the first ACCESS_ONCE(head),
then tmp.head != head will be true before the first unlock we are waiting
for.
And perhaps you can turn these ACCESS_ONCE into READ_ONCE as well.
Oleg.
Damn, sorry for noise, forgot to mention...
On 02/12, Raghavendra K T wrote:
>
> +static inline void __ticket_check_and_clear_slowpath(arch_spinlock_t *lock,
> + __ticket_t head)
> +{
> + if (head & TICKET_SLOWPATH_FLAG) {
> + arch_spinlock_t old, new;
> +
> + old.tickets.head = head;
> + new.tickets.head = head & ~TICKET_SLOWPATH_FLAG;
> + old.tickets.tail = new.tickets.head + TICKET_LOCK_INC;
> + new.tickets.tail = old.tickets.tail;
> +
> + /* try to clear slowpath flag when there are no contenders */
> + cmpxchg(&lock->head_tail, old.head_tail, new.head_tail);
> + }
> +}
...
> +clear_slowpath:
> + if (TICKET_SLOWPATH_FLAG)
> + __ticket_check_and_clear_slowpath(lock, inc.head);
I think you can remove this "if (TICKET_SLOWPATH_FLAG)" check. If it is
defined as zero, gcc should optimize out the code under "if (head & 0)".
But this is purely cosmetic and subjective, feel free to ignore.
Oleg.
On 02/12/2015 07:07 PM, Oleg Nesterov wrote:
> On 02/12, Raghavendra K T wrote:
>>
>> @@ -772,7 +773,8 @@ __visible void kvm_lock_spinning(struct arch_spinlock *lock, __ticket_t want)
>> * check again make sure it didn't become free while
>> * we weren't looking.
>> */
>> - if (ACCESS_ONCE(lock->tickets.head) == want) {
>> + head = ACCESS_ONCE(lock->tickets.head);
>> + if (__tickets_equal(head, want)) {
>> add_stats(TAKEN_SLOW_PICKUP, 1);
>
> While at it, perhaps it makes sense to s/ACCESS_ONCE/READ_ONCE/ but this
> is cosmetic.
yes.. will do.
>
> We also need to change another user of enter_slow_path, xen_lock_spinning()
> in arch/x86/xen/spinlock.c.
>
Had in mind but forgot before sending. will update resend.
> Other than that looks correct at first glance... but this is up to
> maintainers.
>
Thanks
On 02/12/2015 07:20 PM, Oleg Nesterov wrote:
> On 02/12, Raghavendra K T wrote:
>>
>> @@ -191,8 +189,7 @@ static inline void arch_spin_unlock_wait(arch_spinlock_t *lock)
>> * We need to check "unlocked" in a loop, tmp.head == head
>> * can be false positive because of overflow.
>> */
>> - if (tmp.head == (tmp.tail & ~TICKET_SLOWPATH_FLAG) ||
>> - tmp.head != head)
>> + if (__tickets_equal(tmp.head, tmp.tail) || tmp.head != head)
>> break;
>
> Ah, it seems that "tmp.head != head" should be turned into
> !__tickets_equal(), no?
>
> Suppose that TICKET_SLOWPATH_FLAG is set after the first ACCESS_ONCE(head),
> then tmp.head != head will be true before the first unlock we are waiting
> for.
Good catch. othewise we would wrongly break out even when somebody
does halt wait.
>
> And perhaps you can turn these ACCESS_ONCE into READ_ONCE as well.
Yes again :)
On 02/12/2015 07:32 PM, Oleg Nesterov wrote:
> Damn, sorry for noise, forgot to mention...
>
> On 02/12, Raghavendra K T wrote:
>>
>> +static inline void __ticket_check_and_clear_slowpath(arch_spinlock_t *lock,
>> + __ticket_t head)
>> +{
>> + if (head & TICKET_SLOWPATH_FLAG) {
>> + arch_spinlock_t old, new;
>> +
>> + old.tickets.head = head;
>> + new.tickets.head = head & ~TICKET_SLOWPATH_FLAG;
>> + old.tickets.tail = new.tickets.head + TICKET_LOCK_INC;
>> + new.tickets.tail = old.tickets.tail;
>> +
>> + /* try to clear slowpath flag when there are no contenders */
>> + cmpxchg(&lock->head_tail, old.head_tail, new.head_tail);
>> + }
>> +}
>
> ...
>
>> +clear_slowpath:
>> + if (TICKET_SLOWPATH_FLAG)
>> + __ticket_check_and_clear_slowpath(lock, inc.head);
>
> I think you can remove this "if (TICKET_SLOWPATH_FLAG)" check. If it is
> defined as zero, gcc should optimize out the code under "if (head & 0)".
>
right, the above if ( ) is unnecesary, though we would have same code
at the end, getting rid of that makes code more clean.
On Thu, Feb 12, 2015 at 05:17:27PM +0530, Raghavendra K T wrote:
> Paravirt spinlock clears slowpath flag after doing unlock.
> As explained by Linus currently it does:
> prev = *lock;
> add_smp(&lock->tickets.head, TICKET_LOCK_INC);
>
> /* add_smp() is a full mb() */
>
> if (unlikely(lock->tickets.tail & TICKET_SLOWPATH_FLAG))
> __ticket_unlock_slowpath(lock, prev);
>
>
> which is *exactly* the kind of things you cannot do with spinlocks,
> because after you've done the "add_smp()" and released the spinlock
> for the fast-path, you can't access the spinlock any more. Exactly
> because a fast-path lock might come in, and release the whole data
> structure.
>
> Linus suggested that we should not do any writes to lock after unlock(),
> and we can move slowpath clearing to fastpath lock.
>
> So this patch implements the fix with:
> 1. Moving slowpath flag to head (Oleg).
> 2. use xadd to avoid read/write after unlock that checks the need for
> unlock_kick (Linus).
Maybe spend a few more words explaining these things; something like:
Unlocked locks don't care about the slowpath flag; therefore we can keep
it set after the last unlock, as long as we clear it again on the first
(try)lock -- this removes the write after unlock.
By moving the slowpath flag from the tail to the head ticket we avoid
the need to access both the head and tail tickets on unlock.
We can further avoid the need for a read-after-release by using xadd;
the prev head value will include the slowpath flag and indicate if we
need to do PV kicking of suspended spinners -- on modern chips xadd
isn't (much) more expensive than an add + load.
Its 'obvious' now, but maybe not so much after we've all not looked at
this for a few months.
On 02/12/2015 08:30 PM, Peter Zijlstra wrote:
> On Thu, Feb 12, 2015 at 05:17:27PM +0530, Raghavendra K T wrote:
[...]
>>
>> Linus suggested that we should not do any writes to lock after unlock(),
>> and we can move slowpath clearing to fastpath lock.
>>
>> So this patch implements the fix with:
>> 1. Moving slowpath flag to head (Oleg).
>> 2. use xadd to avoid read/write after unlock that checks the need for
>> unlock_kick (Linus).
>
> Maybe spend a few more words explaining these things; something like:
>
> Unlocked locks don't care about the slowpath flag; therefore we can keep
> it set after the last unlock, as long as we clear it again on the first
> (try)lock -- this removes the write after unlock.
Nit: I 'll reword a bit here since slowpath flag would result in
unnecessary kick but otherwise harmless IMO.
something like:
Unlocked locks don't care about the slowpath flag; therefore we can keep
it set after the last unlock, and clear it again on the first (try)lock.
-- this removes the write after unlock. note that keeping slowpath flag
would result in unnecessary kicks.
> By moving the slowpath flag from the tail to the head ticket we avoid
> the need to access both the head and tail tickets on unlock.
>
> We can further avoid the need for a read-after-release by using xadd;
> the prev head value will include the slowpath flag and indicate if we
> need to do PV kicking of suspended spinners -- on modern chips xadd
> isn't (much) more expensive than an add + load.
>
> Its 'obvious' now, but maybe not so much after we've all not looked at
> this for a few months.
>
Absolutely correct. Thanks Peter for the detailed and very helpful
writeup.