In virtualized environment there are mainly three problems related to spinlocks that affect performance. 1. LHP (lock holder preemption) 2. Lock Waiter Preemption (LWP) 3. Starvation/fairness Though ticketlocks solve the fairness problem, it worsens LWP, LHP problems. pv-ticketlocks tried to address this. But we can further improve at the cost of relaxed fairness. In this patch, we form a batch of eligible lock holders and we serve the eligible (to hold the lock) batches in FIFO, but the lock-waiters within eligible batch are served in an unfair manner. This increases probability of any eligible lock-holder being in running state (to an average of (batch_size/2)-1). It also provides needed bounded starvation since any lock requester can not acquire more than batch_size times repeatedly during contention. On the negetive side we would need an extra cmpxchg. The patch has the batch size of 4. (As we know increasing batch size means we are closer to unfair locks and batch size of 1 = ticketlock). Result: Test system: 32cpu 2node machine w/ 64GB each (16 pcpu machine +ht). Guests: 8GB 16vcpu guests (average of 8 iterations) % Improvements with kvm guests (batch size = 4): ebizzy_0.5x 4.3 ebizzy_1.0x 7.8 ebizzy_1.5x 23.4 ebizzy_2.0x 48.6 Baremetal: ebizzy showed very high stdev, kernbench result was good but both of them did not show much difference. ebizzy: rec/sec higher is better base 50452 patched 50703 kernbench time in sec (lesser is better) base 48.9 sec patched 48.8 sec Signed-off-by: Raghavendra K T <raghavendra.kt at linux.vnet.ibm.com> --- arch/x86/include/asm/spinlock.h | 35 +++++++++++++++++++++++++---------- arch/x86/include/asm/spinlock_types.h | 14 ++++++++++---- arch/x86/kernel/kvm.c | 6 ++++-- 3 files changed, 39 insertions(+), 16 deletions(-) TODO: - we need an intelligent way to nullify the effect of batching for baremetal (because extra cmpxchg is not required). - we may have to make batch size as kernel arg to solve above problem (to run same kernel for host/guest). Increasing batch size also seem to help virtualized guest more, so we will have flexibility of tuning depending on vm size. - My kernbench/ebizzy test on baremetal (32 cpu +ht sandybridge) did not seem to show the impact of extra cmpxchg. but there should be effect of extra cmpxchg. - virtualized guest had slight impact on 1x cases of some benchmarks but we have got impressive performance for >1x cases. So overall, patch needs exhaustive tesing. - we can further add dynamically changing batch_size implementation (inspiration and hint by Paul McKenney) as necessary. I have found that increasing batch size gives excellent improvements for overcommitted guests. I understand that we need more exhaustive testing. Please provide your suggestion and comments. diff --git a/arch/x86/include/asm/spinlock.h b/arch/x86/include/asm/spinlock.h index 0f62f54..87685f1 100644 --- a/arch/x86/include/asm/spinlock.h +++ b/arch/x86/include/asm/spinlock.h @@ -81,23 +81,36 @@ static __always_inline int arch_spin_value_unlocked(arch_spinlock_t lock) */ static __always_inline void arch_spin_lock(arch_spinlock_t *lock) { - register struct __raw_tickets inc = { .tail = TICKET_LOCK_INC }; + register struct __raw_tickets inc = { .tail = TICKET_LOCK_TAIL_INC }; + struct __raw_tickets new; inc = xadd(&lock->tickets, inc); - if (likely(inc.head == inc.tail)) - goto out; inc.tail &= ~TICKET_SLOWPATH_FLAG; for (;;) { unsigned count = SPIN_THRESHOLD; do { - if (ACCESS_ONCE(lock->tickets.head) == inc.tail) - goto out; + if ((inc.head & TICKET_LOCK_BATCH_MASK) == (inc.tail & + TICKET_LOCK_BATCH_MASK)) + goto spin; cpu_relax(); + inc.head = ACCESS_ONCE(lock->tickets.head); } while (--count); __ticket_lock_spinning(lock, inc.tail); } +spin: + for (;;) { + inc.head = ACCESS_ONCE(lock->tickets.head); + if (!(inc.head & TICKET_LOCK_HEAD_INC)) { + new.head = inc.head | TICKET_LOCK_HEAD_INC; + if (cmpxchg(&lock->tickets.head, inc.head, new.head) + == inc.head) + goto out; + } + cpu_relax(); + } + out: barrier(); /* make sure nothing creeps before the lock is taken */ } @@ -109,7 +122,8 @@ static __always_inline int arch_spin_trylock(arch_spinlock_t *lock) if (old.tickets.head != (old.tickets.tail & ~TICKET_SLOWPATH_FLAG)) return 0; - new.head_tail = old.head_tail + (TICKET_LOCK_INC << TICKET_SHIFT); + new.head_tail = old.head_tail + (TICKET_LOCK_TAIL_INC << TICKET_SHIFT); + new.head_tail |= TICKET_LOCK_HEAD_INC; /* 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; @@ -123,7 +137,7 @@ static inline void __ticket_unlock_slowpath(arch_spinlock_t *lock, BUILD_BUG_ON(((__ticket_t)NR_CPUS) != NR_CPUS); /* Perform the unlock on the "before" copy */ - old.tickets.head += TICKET_LOCK_INC; + old.tickets.head += TICKET_LOCK_HEAD_INC; /* Clear the slowpath flag */ new.head_tail = old.head_tail & ~(TICKET_SLOWPATH_FLAG << TICKET_SHIFT); @@ -150,14 +164,15 @@ static __always_inline void arch_spin_unlock(arch_spinlock_t *lock) arch_spinlock_t prev; prev = *lock; - add_smp(&lock->tickets.head, TICKET_LOCK_INC); + add_smp(&lock->tickets.head, TICKET_LOCK_HEAD_INC); /* add_smp() is a full mb() */ if (unlikely(lock->tickets.tail & TICKET_SLOWPATH_FLAG)) __ticket_unlock_slowpath(lock, prev); } else - __add(&lock->tickets.head, TICKET_LOCK_INC, UNLOCK_LOCK_PREFIX); + __add(&lock->tickets.head, + TICKET_LOCK_HEAD_INC, UNLOCK_LOCK_PREFIX); } static inline int arch_spin_is_locked(arch_spinlock_t *lock) @@ -171,7 +186,7 @@ static inline int arch_spin_is_contended(arch_spinlock_t *lock) { struct __raw_tickets tmp = ACCESS_ONCE(lock->tickets); - return (__ticket_t)(tmp.tail - tmp.head) > TICKET_LOCK_INC; + return (__ticket_t)(tmp.tail - tmp.head) > TICKET_LOCK_TAIL_INC; } #define arch_spin_is_contended arch_spin_is_contended diff --git a/arch/x86/include/asm/spinlock_types.h b/arch/x86/include/asm/spinlock_types.h index 4f1bea1..b04c03d 100644 --- a/arch/x86/include/asm/spinlock_types.h +++ b/arch/x86/include/asm/spinlock_types.h @@ -3,15 +3,16 @@ #include <linux/types.h> +#define TICKET_LOCK_INC_SHIFT 1 +#define __TICKET_LOCK_TAIL_INC (1<<TICKET_LOCK_INC_SHIFT) + #ifdef CONFIG_PARAVIRT_SPINLOCKS -#define __TICKET_LOCK_INC 2 #define TICKET_SLOWPATH_FLAG ((__ticket_t)1) #else -#define __TICKET_LOCK_INC 1 #define TICKET_SLOWPATH_FLAG ((__ticket_t)0) #endif -#if (CONFIG_NR_CPUS < (256 / __TICKET_LOCK_INC)) +#if (CONFIG_NR_CPUS < (256 / __TICKET_LOCK_TAIL_INC)) typedef u8 __ticket_t; typedef u16 __ticketpair_t; #else @@ -19,7 +20,12 @@ typedef u16 __ticket_t; typedef u32 __ticketpair_t; #endif -#define TICKET_LOCK_INC ((__ticket_t)__TICKET_LOCK_INC) +#define TICKET_LOCK_TAIL_INC ((__ticket_t)__TICKET_LOCK_TAIL_INC) + +#define TICKET_LOCK_HEAD_INC ((__ticket_t)1) +#define TICKET_BATCH 0x4 /* 4 waiters can contend simultaneously */ +#define TICKET_LOCK_BATCH_MASK (~(TICKET_BATCH<<TICKET_LOCK_INC_SHIFT) + \ + TICKET_LOCK_TAIL_INC - 1) #define TICKET_SHIFT (sizeof(__ticket_t) * 8) diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c index 0331cb3..3dd41f7 100644 --- a/arch/x86/kernel/kvm.c +++ b/arch/x86/kernel/kvm.c @@ -755,7 +755,8 @@ __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) { + if ((ACCESS_ONCE(lock->tickets.head) & TICKET_LOCK_BATCH_MASK) =+ (want & TICKET_LOCK_BATCH_MASK)) { add_stats(TAKEN_SLOW_PICKUP, 1); goto out; } @@ -787,7 +788,8 @@ static void kvm_unlock_kick(struct arch_spinlock *lock, __ticket_t ticket) for_each_cpu(cpu, &waiting_cpus) { const struct kvm_lock_waiting *w = &per_cpu(klock_waiting, cpu); if (ACCESS_ONCE(w->lock) == lock && - ACCESS_ONCE(w->want) == ticket) { + (ACCESS_ONCE(w->want) & TICKET_LOCK_BATCH_MASK) =+ (ticket & TICKET_LOCK_BATCH_MASK)) { add_stats(RELEASED_SLOW_KICKED, 1); kvm_kick_cpu(cpu); break; -- 1.7.11.7
On 05/28/2014 08:16 AM, Raghavendra K T wrote: This patch looks very promising.> TODO: > - we need an intelligent way to nullify the effect of batching for baremetal > (because extra cmpxchg is not required).On (larger?) NUMA systems, the unfairness may be a nice performance benefit, reducing cache line bouncing through the system, and it could well outweigh the extra cmpxchg at times.> - we may have to make batch size as kernel arg to solve above problem > (to run same kernel for host/guest). Increasing batch size also seem to help > virtualized guest more, so we will have flexibility of tuning depending on vm size. > > - My kernbench/ebizzy test on baremetal (32 cpu +ht sandybridge) did not seem to > show the impact of extra cmpxchg. but there should be effect of extra cmpxchg.Canceled out by better NUMA locality? Or maybe cmpxchg is cheap once you already own the cache line exclusively?> - virtualized guest had slight impact on 1x cases of some benchmarks but we have got > impressive performance for >1x cases. So overall, patch needs exhaustive tesing. > > - we can further add dynamically changing batch_size implementation (inspiration and > hint by Paul McKenney) as necessary.I could see a larger batch size being beneficial. Currently the maximum wait time for a spinlock on a system with N CPUs is N times the length of the largest critical section. Having the batch size set equal to the number of CPUs would only double that, and better locality (CPUs local to the current lock holder winning the spinlock operation) might speed things up enough to cancel that part of that out again...> I have found that increasing batch size gives excellent improvements for > overcommitted guests. I understand that we need more exhaustive testing. > > Please provide your suggestion and comments.I have only nitpicks so far...> diff --git a/arch/x86/include/asm/spinlock_types.h b/arch/x86/include/asm/spinlock_types.h > index 4f1bea1..b04c03d 100644 > --- a/arch/x86/include/asm/spinlock_types.h > +++ b/arch/x86/include/asm/spinlock_types.h > @@ -3,15 +3,16 @@ > > #include <linux/types.h> > > +#define TICKET_LOCK_INC_SHIFT 1 > +#define __TICKET_LOCK_TAIL_INC (1<<TICKET_LOCK_INC_SHIFT) > + > #ifdef CONFIG_PARAVIRT_SPINLOCKS > -#define __TICKET_LOCK_INC 2 > #define TICKET_SLOWPATH_FLAG ((__ticket_t)1) > #else > -#define __TICKET_LOCK_INC 1 > #define TICKET_SLOWPATH_FLAG ((__ticket_t)0) > #endifFor the !CONFIG_PARAVIRT case, TICKET_LOCK_INC_SHIFT used to be 0, now you are making it one. Probably not an issue, since even people who compile with 128 < CONFIG_NR_CPUS <= 256 will likely have their spinlocks padded out to 32 or 64 bits anyway in most data structures.> -#if (CONFIG_NR_CPUS < (256 / __TICKET_LOCK_INC)) > +#if (CONFIG_NR_CPUS < (256 / __TICKET_LOCK_TAIL_INC)) > typedef u8 __ticket_t; > typedef u16 __ticketpair_t; > #else > @@ -19,7 +20,12 @@ typedef u16 __ticket_t; > typedef u32 __ticketpair_t; > #endif > > -#define TICKET_LOCK_INC ((__ticket_t)__TICKET_LOCK_INC) > +#define TICKET_LOCK_TAIL_INC ((__ticket_t)__TICKET_LOCK_TAIL_INC) > + > +#define TICKET_LOCK_HEAD_INC ((__ticket_t)1) > +#define TICKET_BATCH 0x4 /* 4 waiters can contend simultaneously */ > +#define TICKET_LOCK_BATCH_MASK (~(TICKET_BATCH<<TICKET_LOCK_INC_SHIFT) + \ > + TICKET_LOCK_TAIL_INC - 1)I do not see the value in having TICKET_BATCH declared with a hexadecimal number, and it may be worth making sure the code does not compile if someone tried a TICKET_BATCH value that is not a power of 2. -- All rights reversed
On Wed, May 28, 2014 at 2:55 PM, Rik van Riel <riel at redhat.com> wrote:> > Or maybe cmpxchg is cheap once you already own the cache line > exclusively?A locked cmpxchg ends up being anything between ~15-50 cycles depending on microarchitecture if things are already exclusively in the cache (with the P4 being an outlier, and all locked instructions tend to take ~100+ cycles, but I can't say I can really find it in myself to even care about netburst any more). The most noticeable downside we've seen has been when we've used "read-op-cmpxchg" as a _replacement_ for something like "lock [x]add", when that "read+cmpxchg" has caused two cacheline ops (cacheline first loaded shared by the read, then exclusive by the cmpxchg). That's bad. But if preceded by a write (or, in this case, an xadd), that doesn't happen. Still, those roughly 15-50 cycles can certainly be noticeable (especially at the high end), but you need to have some load that doesn't bounce the lock, and largely fit in the caches to see it. And you probably want to test one of the older CPU's, I think Haswell is the lower end (ie in the ~20 cycle range). If somebody has a P4 still, that's likely the worst case by far. Linus
On Wed, May 28, 2014 at 05:46:39PM +0530, Raghavendra K T wrote:> In virtualized environment there are mainly three problems > related to spinlocks that affect performance. > 1. LHP (lock holder preemption) > 2. Lock Waiter Preemption (LWP) > 3. Starvation/fairness > > Though ticketlocks solve the fairness problem, it worsens LWP, LHP problems. > pv-ticketlocks tried to address this. But we can further improve at the > cost of relaxed fairness.So I really hate the idea of having different locks for paravirt and normal kernels. And we're looking to move to that queued lock for normal kernels.
On 05/29/2014 03:25 AM, Rik van Riel wrote:> On 05/28/2014 08:16 AM, Raghavendra K T wrote: > > This patch looks very promising.Thank you Rik. [...]>> >> - My kernbench/ebizzy test on baremetal (32 cpu +ht sandybridge) did not seem to >> show the impact of extra cmpxchg. but there should be effect of extra cmpxchg. > > Canceled out by better NUMA locality?Yes perhaps. it was even slightly better. [...]>> - we can further add dynamically changing batch_size implementation (inspiration and >> hint by Paul McKenney) as necessary. > > I could see a larger batch size being beneficial. > > Currently the maximum wait time for a spinlock on a system > with N CPUs is N times the length of the largest critical > section. > > Having the batch size set equal to the number of CPUs would only > double that, and better locality (CPUs local to the current > lock holder winning the spinlock operation) might speed things > up enough to cancel that part of that out again...having batch size = number of cpus would definitely help contended cases especially on larger machines (by my experience with testing on a 4 node 32 core machine). +ht case should make it even more beneficial. My only botheration was overhead in undercommit cases because of extra cmpxchg. So may be batch_size = total cpus / numa node be optimal?... [...]>> +#define TICKET_LOCK_INC_SHIFT 1 >> +#define __TICKET_LOCK_TAIL_INC (1<<TICKET_LOCK_INC_SHIFT) >> + >> #ifdef CONFIG_PARAVIRT_SPINLOCKS >> -#define __TICKET_LOCK_INC 2 >> #define TICKET_SLOWPATH_FLAG ((__ticket_t)1) >> #else >> -#define __TICKET_LOCK_INC 1 >> #define TICKET_SLOWPATH_FLAG ((__ticket_t)0) >> #endif > > For the !CONFIG_PARAVIRT case, TICKET_LOCK_INC_SHIFT used to be 0, > now you are making it one. Probably not an issue, since even people > who compile with 128 < CONFIG_NR_CPUS <= 256 will likely have their > spinlocks padded out to 32 or 64 bits anyway in most data structures.Yes.. [...]>> +#define TICKET_BATCH 0x4 /* 4 waiters can contend simultaneously */ >> +#define TICKET_LOCK_BATCH_MASK (~(TICKET_BATCH<<TICKET_LOCK_INC_SHIFT) + \ >> + TICKET_LOCK_TAIL_INC - 1) > > I do not see the value in having TICKET_BATCH declared with a > hexadecimal number,yes.. It had only helped me to make the idea readable to myself, I could get rid of this if needed. and it may be worth making sure the code> does not compile if someone tried a TICKET_BATCH value that > is not a power of 2.I agree. will have BUILD_BUG for not power of 2 in next version. But yes it reminds me that I wanted to have TICKET_BATCH = 1 for !CONFIG_PARAVIRT so that we continue to have original fair lock version. Does that make sense? I left it after thinking about same kernel running on host/guest which would anyway will have CONFIG_PARAVIRT on.
On 05/29/2014 12:16 PM, Peter Zijlstra wrote:> On Wed, May 28, 2014 at 05:46:39PM +0530, Raghavendra K T wrote: >> In virtualized environment there are mainly three problems >> related to spinlocks that affect performance. >> 1. LHP (lock holder preemption) >> 2. Lock Waiter Preemption (LWP) >> 3. Starvation/fairness >> >> Though ticketlocks solve the fairness problem, it worsens LWP, LHP problems. >> pv-ticketlocks tried to address this. But we can further improve at the >> cost of relaxed fairness. > > So I really hate the idea of having different locks for paravirt and > normal kernels.Yes. I understand that queued lock for normal kernel and unfair version of queued spinlock for virtual guest would do better. Since strict serialization of lockwaiters (in both ticketlock/queued spinlock) does not work well for virtualized guest, my idea was to give an alternate idea which has bounded starvation and performs as good as unfair version to virtualized guest.> > And we're looking to move to that queued lock for normal kernels.Agree. and I have tested that too.
On 05/28/2014 08:16 AM, Raghavendra K T wrote:> > TODO: > - we need an intelligent way to nullify the effect of batching for baremetal > (because extra cmpxchg is not required).To do this, you will need to have 2 slightly different algorithms depending on the paravirt_ticketlocks_enabled jump label.> > - My kernbench/ebizzy test on baremetal (32 cpu +ht sandybridge) did not seem to > show the impact of extra cmpxchg. but there should be effect of extra cmpxchg.It will depend on the micro-benchmark and the test system used. I had seen the a test case that extra cmpxchg did not really impact performance on a Westmere system but had noticeable adverse impact on an IvyBridge system with the same micro-benchmark.> Please provide your suggestion and comments. > > diff --git a/arch/x86/include/asm/spinlock.h b/arch/x86/include/asm/spinlock.h > index 0f62f54..87685f1 100644 > --- a/arch/x86/include/asm/spinlock.h > +++ b/arch/x86/include/asm/spinlock.h > @@ -81,23 +81,36 @@ static __always_inline int arch_spin_value_unlocked(arch_spinlock_t lock) > */ > static __always_inline void arch_spin_lock(arch_spinlock_t *lock) > { > - register struct __raw_tickets inc = { .tail = TICKET_LOCK_INC }; > + register struct __raw_tickets inc = { .tail = TICKET_LOCK_TAIL_INC }; > + struct __raw_tickets new; > > inc = xadd(&lock->tickets, inc); > - if (likely(inc.head == inc.tail)) > - goto out; > > inc.tail&= ~TICKET_SLOWPATH_FLAG; > for (;;) { > unsigned count = SPIN_THRESHOLD; > > do { > - if (ACCESS_ONCE(lock->tickets.head) == inc.tail) > - goto out; > + if ((inc.head& TICKET_LOCK_BATCH_MASK) == (inc.tail& > + TICKET_LOCK_BATCH_MASK)) > + goto spin; > cpu_relax(); > + inc.head = ACCESS_ONCE(lock->tickets.head); > } while (--count); > __ticket_lock_spinning(lock, inc.tail); > } > +spin: > + for (;;) { > + inc.head = ACCESS_ONCE(lock->tickets.head); > + if (!(inc.head& TICKET_LOCK_HEAD_INC)) { > + new.head = inc.head | TICKET_LOCK_HEAD_INC; > + if (cmpxchg(&lock->tickets.head, inc.head, new.head) > + == inc.head) > + goto out; > + } > + cpu_relax(); > + } > +It had taken me some time to figure out the the LSB of inc.head is used as a bit lock for the contending tasks in the spin loop. I would suggest adding some comment here to make it easier to look at.> diff --git a/arch/x86/include/asm/spinlock_types.h b/arch/x86/include/asm/spinlock_types.h > index 4f1bea1..b04c03d 100644 > --- a/arch/x86/include/asm/spinlock_types.h > +++ b/arch/x86/include/asm/spinlock_types.h > @@ -3,15 +3,16 @@ > > #include<linux/types.h> > > +#define TICKET_LOCK_INC_SHIFT 1 > +#define __TICKET_LOCK_TAIL_INC (1<<TICKET_LOCK_INC_SHIFT) > + > #ifdef CONFIG_PARAVIRT_SPINLOCKS > -#define __TICKET_LOCK_INC 2 > #define TICKET_SLOWPATH_FLAG ((__ticket_t)1) > #else > -#define __TICKET_LOCK_INC 1 > #define TICKET_SLOWPATH_FLAG ((__ticket_t)0) > #endif > > -#if (CONFIG_NR_CPUS< (256 / __TICKET_LOCK_INC)) > +#if (CONFIG_NR_CPUS< (256 / __TICKET_LOCK_TAIL_INC)) > typedef u8 __ticket_t; > typedef u16 __ticketpair_t; > #else > @@ -19,7 +20,12 @@ typedef u16 __ticket_t; > typedef u32 __ticketpair_t; > #endif > > -#define TICKET_LOCK_INC ((__ticket_t)__TICKET_LOCK_INC) > +#define TICKET_LOCK_TAIL_INC ((__ticket_t)__TICKET_LOCK_TAIL_INC) > + > +#define TICKET_LOCK_HEAD_INC ((__ticket_t)1) > +#define TICKET_BATCH 0x4 /* 4 waiters can contend simultaneously */ > +#define TICKET_LOCK_BATCH_MASK (~(TICKET_BATCH<<TICKET_LOCK_INC_SHIFT) + \ > + TICKET_LOCK_TAIL_INC - 1)I don't think TAIL_INC has anything to do with setting the BATCH_MASK. It works here because TAIL_INC is 2. I think it is clearer to define it as either "(~(TICKET_BATCH<<TICKET_LOCK_INC_SHIFT) + 1)" or (~((TICKET_BATCH<<TICKET_LOCK_INC_SHIFT) - 1)). -Longman
On 05/30/2014 04:15 AM, Waiman Long wrote:> On 05/28/2014 08:16 AM, Raghavendra K T wrote: >> - we need an intelligent way to nullify the effect of batching for >> baremetal >> (because extra cmpxchg is not required). > > To do this, you will need to have 2 slightly different algorithms > depending on the paravirt_ticketlocks_enabled jump label.Thanks for the hint Waiman. [...]>> +spin: >> + for (;;) { >> + inc.head = ACCESS_ONCE(lock->tickets.head); >> + if (!(inc.head& TICKET_LOCK_HEAD_INC)) { >> + new.head = inc.head | TICKET_LOCK_HEAD_INC; >> + if (cmpxchg(&lock->tickets.head, inc.head, new.head) >> + == inc.head) >> + goto out; >> + } >> + cpu_relax(); >> + } >> + > > It had taken me some time to figure out the the LSB of inc.head is used > as a bit lock for the contending tasks in the spin loop. I would suggest > adding some comment here to make it easier to look at.Agree. 'll add a comment. [...]>> +#define TICKET_BATCH 0x4 /* 4 waiters can contend simultaneously */ >> +#define TICKET_LOCK_BATCH_MASK >> (~(TICKET_BATCH<<TICKET_LOCK_INC_SHIFT) + \ >> + TICKET_LOCK_TAIL_INC - 1) > > I don't think TAIL_INC has anything to do with setting the BATCH_MASK. > It works here because TAIL_INC is 2. I think it is clearer to define it > as either "(~(TICKET_BATCH<<TICKET_LOCK_INC_SHIFT) + 1)" or > (~((TICKET_BATCH<<TICKET_LOCK_INC_SHIFT) - 1)).You are right. Thanks for pointing out. Your expression is simple and clearer. 'll use one of them.