linux/arch/x86/include/asm/spinlock.h

287 lines
7.5 KiB
C
Raw Normal View History

#ifndef _ASM_X86_SPINLOCK_H
#define _ASM_X86_SPINLOCK_H
x86, ticketlock: Add slowpath logic Maintain a flag in the LSB of the ticket lock tail which indicates whether anyone is in the lock slowpath and may need kicking when the current holder unlocks. The flags are set when the first locker enters the slowpath, and cleared when unlocking to an empty queue (ie, no contention). In the specific implementation of lock_spinning(), make sure to set the slowpath flags on the lock just before blocking. We must do this before the last-chance pickup test to prevent a deadlock with the unlocker: Unlocker Locker test for lock pickup -> fail unlock test slowpath -> false set slowpath flags block Whereas this works in any ordering: Unlocker Locker set slowpath flags test for lock pickup -> fail block unlock test slowpath -> true, kick If the unlocker finds that the lock has the slowpath flag set but it is actually uncontended (ie, head == tail, so nobody is waiting), then it clears the slowpath flag. The unlock code uses a locked add to update the head counter. This also acts as a full memory barrier so that its safe to subsequently read back the slowflag state, knowing that the updated lock is visible to the other CPUs. If it were an unlocked add, then the flag read may just be forwarded from the store buffer before it was visible to the other CPUs, which could result in a deadlock. Unfortunately this means we need to do a locked instruction when unlocking with PV ticketlocks. However, if PV ticketlocks are not enabled, then the old non-locked "add" is the only unlocking code. Note: this code relies on gcc making sure that unlikely() code is out of line of the fastpath, which only happens when OPTIMIZE_SIZE=n. If it doesn't the generated code isn't too bad, but its definitely suboptimal. Thanks to Srivatsa Vaddagiri for providing a bugfix to the original version of this change, which has been folded in. Thanks to Stephan Diestelhorst for commenting on some code which relied on an inaccurate reading of the x86 memory ordering rules. Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-11-git-send-email-raghavendra.kt@linux.vnet.ibm.com Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Stephan Diestelhorst <stephan.diestelhorst@amd.com> Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:58 +08:00
#include <linux/jump_label.h>
#include <linux/atomic.h>
#include <asm/page.h>
#include <asm/processor.h>
x86: FIFO ticket spinlocks Introduce ticket lock spinlocks for x86 which are FIFO. The implementation is described in the comments. The straight-line lock/unlock instruction sequence is slightly slower than the dec based locks on modern x86 CPUs, however the difference is quite small on Core2 and Opteron when working out of cache, and becomes almost insignificant even on P4 when the lock misses cache. trylock is more significantly slower, but they are relatively rare. On an 8 core (2 socket) Opteron, spinlock unfairness is extremely noticable, with a userspace test having a difference of up to 2x runtime per thread, and some threads are starved or "unfairly" granted the lock up to 1 000 000 (!) times. After this patch, all threads appear to finish at exactly the same time. The memory ordering of the lock does conform to x86 standards, and the implementation has been reviewed by Intel and AMD engineers. The algorithm also tells us how many CPUs are contending the lock, so lockbreak becomes trivial and we no longer have to waste 4 bytes per spinlock for it. After this, we can no longer spin on any locks with preempt enabled and cannot reenable interrupts when spinning on an irq safe lock, because at that point we have already taken a ticket and the would deadlock if the same CPU tries to take the lock again. These are questionable anyway: if the lock happens to be called under a preempt or interrupt disabled section, then it will just have the same latency problems. The real fix is to keep critical sections short, and ensure locks are reasonably fair (which this patch does). Signed-off-by: Nick Piggin <npiggin@suse.de> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-01-30 20:31:21 +08:00
#include <linux/compiler.h>
x86/paravirt: add hooks for spinlock operations Ticket spinlocks have absolutely ghastly worst-case performance characteristics in a virtual environment. If there is any contention for physical CPUs (ie, there are more runnable vcpus than cpus), then ticket locks can cause the system to end up spending 90+% of its time spinning. The problem is that (v)cpus waiting on a ticket spinlock will be granted access to the lock in strict order they got their tickets. If the hypervisor scheduler doesn't give the vcpus time in that order, they will burn timeslices waiting for the scheduler to give the right vcpu some time. In the worst case it could take O(n^2) vcpu scheduler timeslices for everyone waiting on the lock to get it, not counting new cpus trying to take the lock while the log-jam is sorted out. These hooks allow a paravirt backend to replace the spinlock implementation. At the very least, this could revert the implementation back to the old lock algorithm, which allows the next scheduled vcpu to take the lock, and has basically fairly good performance. It also allows the spinlocks to take advantages of the hypervisor features to make locks more efficient (spin and block, for example). The cost to native execution is an extra direct call when using a spinlock function. There's no overhead if CONFIG_PARAVIRT is turned off. The lock structure is fixed at a single "unsigned int", initialized to zero, but the spinlock implementation can use it as it wishes. Thanks to Thomas Friebel's Xen Summit talk "Preventing Guests from Spinning Around" for pointing out this problem. Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Christoph Lameter <clameter@linux-foundation.org> Cc: Petr Tesarik <ptesarik@suse.cz> Cc: Virtualization <virtualization@lists.linux-foundation.org> Cc: Xen devel <xen-devel@lists.xensource.com> Cc: Thomas Friebel <thomas.friebel@amd.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-07-08 03:07:50 +08:00
#include <asm/paravirt.h>
x86, ticketlock: Add slowpath logic Maintain a flag in the LSB of the ticket lock tail which indicates whether anyone is in the lock slowpath and may need kicking when the current holder unlocks. The flags are set when the first locker enters the slowpath, and cleared when unlocking to an empty queue (ie, no contention). In the specific implementation of lock_spinning(), make sure to set the slowpath flags on the lock just before blocking. We must do this before the last-chance pickup test to prevent a deadlock with the unlocker: Unlocker Locker test for lock pickup -> fail unlock test slowpath -> false set slowpath flags block Whereas this works in any ordering: Unlocker Locker set slowpath flags test for lock pickup -> fail block unlock test slowpath -> true, kick If the unlocker finds that the lock has the slowpath flag set but it is actually uncontended (ie, head == tail, so nobody is waiting), then it clears the slowpath flag. The unlock code uses a locked add to update the head counter. This also acts as a full memory barrier so that its safe to subsequently read back the slowflag state, knowing that the updated lock is visible to the other CPUs. If it were an unlocked add, then the flag read may just be forwarded from the store buffer before it was visible to the other CPUs, which could result in a deadlock. Unfortunately this means we need to do a locked instruction when unlocking with PV ticketlocks. However, if PV ticketlocks are not enabled, then the old non-locked "add" is the only unlocking code. Note: this code relies on gcc making sure that unlikely() code is out of line of the fastpath, which only happens when OPTIMIZE_SIZE=n. If it doesn't the generated code isn't too bad, but its definitely suboptimal. Thanks to Srivatsa Vaddagiri for providing a bugfix to the original version of this change, which has been folded in. Thanks to Stephan Diestelhorst for commenting on some code which relied on an inaccurate reading of the x86 memory ordering rules. Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-11-git-send-email-raghavendra.kt@linux.vnet.ibm.com Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Stephan Diestelhorst <stephan.diestelhorst@amd.com> Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:58 +08:00
#include <asm/bitops.h>
/*
* Your basic SMP spinlocks, allowing only a single CPU anywhere
*
* Simple spin lock operations. There are two variants, one clears IRQ's
* on the local processor, one does not.
*
* These are fair FIFO ticket locks, which support up to 2^16 CPUs.
*
* (the type definitions are in asm/spinlock_types.h)
*/
#ifdef CONFIG_X86_32
# define LOCK_PTR_REG "a"
#else
# define LOCK_PTR_REG "D"
#endif
#if defined(CONFIG_X86_32) && (defined(CONFIG_X86_PPRO_FENCE))
/*
* On PPro SMP, we use a locked operation to unlock
* (PPro errata 66, 92)
*/
# define UNLOCK_LOCK_PREFIX LOCK_PREFIX
#else
# define UNLOCK_LOCK_PREFIX
x86: FIFO ticket spinlocks Introduce ticket lock spinlocks for x86 which are FIFO. The implementation is described in the comments. The straight-line lock/unlock instruction sequence is slightly slower than the dec based locks on modern x86 CPUs, however the difference is quite small on Core2 and Opteron when working out of cache, and becomes almost insignificant even on P4 when the lock misses cache. trylock is more significantly slower, but they are relatively rare. On an 8 core (2 socket) Opteron, spinlock unfairness is extremely noticable, with a userspace test having a difference of up to 2x runtime per thread, and some threads are starved or "unfairly" granted the lock up to 1 000 000 (!) times. After this patch, all threads appear to finish at exactly the same time. The memory ordering of the lock does conform to x86 standards, and the implementation has been reviewed by Intel and AMD engineers. The algorithm also tells us how many CPUs are contending the lock, so lockbreak becomes trivial and we no longer have to waste 4 bytes per spinlock for it. After this, we can no longer spin on any locks with preempt enabled and cannot reenable interrupts when spinning on an irq safe lock, because at that point we have already taken a ticket and the would deadlock if the same CPU tries to take the lock again. These are questionable anyway: if the lock happens to be called under a preempt or interrupt disabled section, then it will just have the same latency problems. The real fix is to keep critical sections short, and ensure locks are reasonably fair (which this patch does). Signed-off-by: Nick Piggin <npiggin@suse.de> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-01-30 20:31:21 +08:00
#endif
x86, spinlock: Replace pv spinlocks with pv ticketlocks Rather than outright replacing the entire spinlock implementation in order to paravirtualize it, keep the ticket lock implementation but add a couple of pvops hooks on the slow patch (long spin on lock, unlocking a contended lock). Ticket locks have a number of nice properties, but they also have some surprising behaviours in virtual environments. They enforce a strict FIFO ordering on cpus trying to take a lock; however, if the hypervisor scheduler does not schedule the cpus in the correct order, the system can waste a huge amount of time spinning until the next cpu can take the lock. (See Thomas Friebel's talk "Prevent Guests from Spinning Around" http://www.xen.org/files/xensummitboston08/LHP.pdf for more details.) To address this, we add two hooks: - __ticket_spin_lock which is called after the cpu has been spinning on the lock for a significant number of iterations but has failed to take the lock (presumably because the cpu holding the lock has been descheduled). The lock_spinning pvop is expected to block the cpu until it has been kicked by the current lock holder. - __ticket_spin_unlock, which on releasing a contended lock (there are more cpus with tail tickets), it looks to see if the next cpu is blocked and wakes it if so. When compiled with CONFIG_PARAVIRT_SPINLOCKS disabled, a set of stub functions causes all the extra code to go away. Results: ======= setup: 32 core machine with 32 vcpu KVM guest (HT off) with 8GB RAM base = 3.11-rc patched = base + pvspinlock V12 +-----------------+----------------+--------+ dbench (Throughput in MB/sec. Higher is better) +-----------------+----------------+--------+ | base (stdev %)|patched(stdev%) | %gain | +-----------------+----------------+--------+ | 15035.3 (0.3) |15150.0 (0.6) | 0.8 | | 1470.0 (2.2) | 1713.7 (1.9) | 16.6 | | 848.6 (4.3) | 967.8 (4.3) | 14.0 | | 652.9 (3.5) | 685.3 (3.7) | 5.0 | +-----------------+----------------+--------+ pvspinlock shows benefits for overcommit ratio > 1 for PLE enabled cases, and undercommits results are flat Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-2-git-send-email-raghavendra.kt@linux.vnet.ibm.com Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Tested-by: Attilio Rao <attilio.rao@citrix.com> [ Raghavendra: Changed SPIN_THRESHOLD, fixed redefinition of arch_spinlock_t] Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:49 +08:00
/* How long a lock should spin before we consider blocking */
#define SPIN_THRESHOLD (1 << 15)
x86, ticketlock: Add slowpath logic Maintain a flag in the LSB of the ticket lock tail which indicates whether anyone is in the lock slowpath and may need kicking when the current holder unlocks. The flags are set when the first locker enters the slowpath, and cleared when unlocking to an empty queue (ie, no contention). In the specific implementation of lock_spinning(), make sure to set the slowpath flags on the lock just before blocking. We must do this before the last-chance pickup test to prevent a deadlock with the unlocker: Unlocker Locker test for lock pickup -> fail unlock test slowpath -> false set slowpath flags block Whereas this works in any ordering: Unlocker Locker set slowpath flags test for lock pickup -> fail block unlock test slowpath -> true, kick If the unlocker finds that the lock has the slowpath flag set but it is actually uncontended (ie, head == tail, so nobody is waiting), then it clears the slowpath flag. The unlock code uses a locked add to update the head counter. This also acts as a full memory barrier so that its safe to subsequently read back the slowflag state, knowing that the updated lock is visible to the other CPUs. If it were an unlocked add, then the flag read may just be forwarded from the store buffer before it was visible to the other CPUs, which could result in a deadlock. Unfortunately this means we need to do a locked instruction when unlocking with PV ticketlocks. However, if PV ticketlocks are not enabled, then the old non-locked "add" is the only unlocking code. Note: this code relies on gcc making sure that unlikely() code is out of line of the fastpath, which only happens when OPTIMIZE_SIZE=n. If it doesn't the generated code isn't too bad, but its definitely suboptimal. Thanks to Srivatsa Vaddagiri for providing a bugfix to the original version of this change, which has been folded in. Thanks to Stephan Diestelhorst for commenting on some code which relied on an inaccurate reading of the x86 memory ordering rules. Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-11-git-send-email-raghavendra.kt@linux.vnet.ibm.com Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Stephan Diestelhorst <stephan.diestelhorst@amd.com> Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:58 +08:00
extern struct static_key paravirt_ticketlocks_enabled;
static __always_inline bool static_key_false(struct static_key *key);
x86, spinlock: Replace pv spinlocks with pv ticketlocks Rather than outright replacing the entire spinlock implementation in order to paravirtualize it, keep the ticket lock implementation but add a couple of pvops hooks on the slow patch (long spin on lock, unlocking a contended lock). Ticket locks have a number of nice properties, but they also have some surprising behaviours in virtual environments. They enforce a strict FIFO ordering on cpus trying to take a lock; however, if the hypervisor scheduler does not schedule the cpus in the correct order, the system can waste a huge amount of time spinning until the next cpu can take the lock. (See Thomas Friebel's talk "Prevent Guests from Spinning Around" http://www.xen.org/files/xensummitboston08/LHP.pdf for more details.) To address this, we add two hooks: - __ticket_spin_lock which is called after the cpu has been spinning on the lock for a significant number of iterations but has failed to take the lock (presumably because the cpu holding the lock has been descheduled). The lock_spinning pvop is expected to block the cpu until it has been kicked by the current lock holder. - __ticket_spin_unlock, which on releasing a contended lock (there are more cpus with tail tickets), it looks to see if the next cpu is blocked and wakes it if so. When compiled with CONFIG_PARAVIRT_SPINLOCKS disabled, a set of stub functions causes all the extra code to go away. Results: ======= setup: 32 core machine with 32 vcpu KVM guest (HT off) with 8GB RAM base = 3.11-rc patched = base + pvspinlock V12 +-----------------+----------------+--------+ dbench (Throughput in MB/sec. Higher is better) +-----------------+----------------+--------+ | base (stdev %)|patched(stdev%) | %gain | +-----------------+----------------+--------+ | 15035.3 (0.3) |15150.0 (0.6) | 0.8 | | 1470.0 (2.2) | 1713.7 (1.9) | 16.6 | | 848.6 (4.3) | 967.8 (4.3) | 14.0 | | 652.9 (3.5) | 685.3 (3.7) | 5.0 | +-----------------+----------------+--------+ pvspinlock shows benefits for overcommit ratio > 1 for PLE enabled cases, and undercommits results are flat Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-2-git-send-email-raghavendra.kt@linux.vnet.ibm.com Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Tested-by: Attilio Rao <attilio.rao@citrix.com> [ Raghavendra: Changed SPIN_THRESHOLD, fixed redefinition of arch_spinlock_t] Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:49 +08:00
x86, ticketlock: Add slowpath logic Maintain a flag in the LSB of the ticket lock tail which indicates whether anyone is in the lock slowpath and may need kicking when the current holder unlocks. The flags are set when the first locker enters the slowpath, and cleared when unlocking to an empty queue (ie, no contention). In the specific implementation of lock_spinning(), make sure to set the slowpath flags on the lock just before blocking. We must do this before the last-chance pickup test to prevent a deadlock with the unlocker: Unlocker Locker test for lock pickup -> fail unlock test slowpath -> false set slowpath flags block Whereas this works in any ordering: Unlocker Locker set slowpath flags test for lock pickup -> fail block unlock test slowpath -> true, kick If the unlocker finds that the lock has the slowpath flag set but it is actually uncontended (ie, head == tail, so nobody is waiting), then it clears the slowpath flag. The unlock code uses a locked add to update the head counter. This also acts as a full memory barrier so that its safe to subsequently read back the slowflag state, knowing that the updated lock is visible to the other CPUs. If it were an unlocked add, then the flag read may just be forwarded from the store buffer before it was visible to the other CPUs, which could result in a deadlock. Unfortunately this means we need to do a locked instruction when unlocking with PV ticketlocks. However, if PV ticketlocks are not enabled, then the old non-locked "add" is the only unlocking code. Note: this code relies on gcc making sure that unlikely() code is out of line of the fastpath, which only happens when OPTIMIZE_SIZE=n. If it doesn't the generated code isn't too bad, but its definitely suboptimal. Thanks to Srivatsa Vaddagiri for providing a bugfix to the original version of this change, which has been folded in. Thanks to Stephan Diestelhorst for commenting on some code which relied on an inaccurate reading of the x86 memory ordering rules. Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-11-git-send-email-raghavendra.kt@linux.vnet.ibm.com Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Stephan Diestelhorst <stephan.diestelhorst@amd.com> Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:58 +08:00
#ifdef CONFIG_PARAVIRT_SPINLOCKS
static inline void __ticket_enter_slowpath(arch_spinlock_t *lock)
x86, spinlock: Replace pv spinlocks with pv ticketlocks Rather than outright replacing the entire spinlock implementation in order to paravirtualize it, keep the ticket lock implementation but add a couple of pvops hooks on the slow patch (long spin on lock, unlocking a contended lock). Ticket locks have a number of nice properties, but they also have some surprising behaviours in virtual environments. They enforce a strict FIFO ordering on cpus trying to take a lock; however, if the hypervisor scheduler does not schedule the cpus in the correct order, the system can waste a huge amount of time spinning until the next cpu can take the lock. (See Thomas Friebel's talk "Prevent Guests from Spinning Around" http://www.xen.org/files/xensummitboston08/LHP.pdf for more details.) To address this, we add two hooks: - __ticket_spin_lock which is called after the cpu has been spinning on the lock for a significant number of iterations but has failed to take the lock (presumably because the cpu holding the lock has been descheduled). The lock_spinning pvop is expected to block the cpu until it has been kicked by the current lock holder. - __ticket_spin_unlock, which on releasing a contended lock (there are more cpus with tail tickets), it looks to see if the next cpu is blocked and wakes it if so. When compiled with CONFIG_PARAVIRT_SPINLOCKS disabled, a set of stub functions causes all the extra code to go away. Results: ======= setup: 32 core machine with 32 vcpu KVM guest (HT off) with 8GB RAM base = 3.11-rc patched = base + pvspinlock V12 +-----------------+----------------+--------+ dbench (Throughput in MB/sec. Higher is better) +-----------------+----------------+--------+ | base (stdev %)|patched(stdev%) | %gain | +-----------------+----------------+--------+ | 15035.3 (0.3) |15150.0 (0.6) | 0.8 | | 1470.0 (2.2) | 1713.7 (1.9) | 16.6 | | 848.6 (4.3) | 967.8 (4.3) | 14.0 | | 652.9 (3.5) | 685.3 (3.7) | 5.0 | +-----------------+----------------+--------+ pvspinlock shows benefits for overcommit ratio > 1 for PLE enabled cases, and undercommits results are flat Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-2-git-send-email-raghavendra.kt@linux.vnet.ibm.com Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Tested-by: Attilio Rao <attilio.rao@citrix.com> [ Raghavendra: Changed SPIN_THRESHOLD, fixed redefinition of arch_spinlock_t] Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:49 +08:00
{
x86, ticketlock: Add slowpath logic Maintain a flag in the LSB of the ticket lock tail which indicates whether anyone is in the lock slowpath and may need kicking when the current holder unlocks. The flags are set when the first locker enters the slowpath, and cleared when unlocking to an empty queue (ie, no contention). In the specific implementation of lock_spinning(), make sure to set the slowpath flags on the lock just before blocking. We must do this before the last-chance pickup test to prevent a deadlock with the unlocker: Unlocker Locker test for lock pickup -> fail unlock test slowpath -> false set slowpath flags block Whereas this works in any ordering: Unlocker Locker set slowpath flags test for lock pickup -> fail block unlock test slowpath -> true, kick If the unlocker finds that the lock has the slowpath flag set but it is actually uncontended (ie, head == tail, so nobody is waiting), then it clears the slowpath flag. The unlock code uses a locked add to update the head counter. This also acts as a full memory barrier so that its safe to subsequently read back the slowflag state, knowing that the updated lock is visible to the other CPUs. If it were an unlocked add, then the flag read may just be forwarded from the store buffer before it was visible to the other CPUs, which could result in a deadlock. Unfortunately this means we need to do a locked instruction when unlocking with PV ticketlocks. However, if PV ticketlocks are not enabled, then the old non-locked "add" is the only unlocking code. Note: this code relies on gcc making sure that unlikely() code is out of line of the fastpath, which only happens when OPTIMIZE_SIZE=n. If it doesn't the generated code isn't too bad, but its definitely suboptimal. Thanks to Srivatsa Vaddagiri for providing a bugfix to the original version of this change, which has been folded in. Thanks to Stephan Diestelhorst for commenting on some code which relied on an inaccurate reading of the x86 memory ordering rules. Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-11-git-send-email-raghavendra.kt@linux.vnet.ibm.com Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Stephan Diestelhorst <stephan.diestelhorst@amd.com> Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:58 +08:00
set_bit(0, (volatile unsigned long *)&lock->tickets.tail);
x86, spinlock: Replace pv spinlocks with pv ticketlocks Rather than outright replacing the entire spinlock implementation in order to paravirtualize it, keep the ticket lock implementation but add a couple of pvops hooks on the slow patch (long spin on lock, unlocking a contended lock). Ticket locks have a number of nice properties, but they also have some surprising behaviours in virtual environments. They enforce a strict FIFO ordering on cpus trying to take a lock; however, if the hypervisor scheduler does not schedule the cpus in the correct order, the system can waste a huge amount of time spinning until the next cpu can take the lock. (See Thomas Friebel's talk "Prevent Guests from Spinning Around" http://www.xen.org/files/xensummitboston08/LHP.pdf for more details.) To address this, we add two hooks: - __ticket_spin_lock which is called after the cpu has been spinning on the lock for a significant number of iterations but has failed to take the lock (presumably because the cpu holding the lock has been descheduled). The lock_spinning pvop is expected to block the cpu until it has been kicked by the current lock holder. - __ticket_spin_unlock, which on releasing a contended lock (there are more cpus with tail tickets), it looks to see if the next cpu is blocked and wakes it if so. When compiled with CONFIG_PARAVIRT_SPINLOCKS disabled, a set of stub functions causes all the extra code to go away. Results: ======= setup: 32 core machine with 32 vcpu KVM guest (HT off) with 8GB RAM base = 3.11-rc patched = base + pvspinlock V12 +-----------------+----------------+--------+ dbench (Throughput in MB/sec. Higher is better) +-----------------+----------------+--------+ | base (stdev %)|patched(stdev%) | %gain | +-----------------+----------------+--------+ | 15035.3 (0.3) |15150.0 (0.6) | 0.8 | | 1470.0 (2.2) | 1713.7 (1.9) | 16.6 | | 848.6 (4.3) | 967.8 (4.3) | 14.0 | | 652.9 (3.5) | 685.3 (3.7) | 5.0 | +-----------------+----------------+--------+ pvspinlock shows benefits for overcommit ratio > 1 for PLE enabled cases, and undercommits results are flat Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-2-git-send-email-raghavendra.kt@linux.vnet.ibm.com Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Tested-by: Attilio Rao <attilio.rao@citrix.com> [ Raghavendra: Changed SPIN_THRESHOLD, fixed redefinition of arch_spinlock_t] Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:49 +08:00
}
x86, ticketlock: Add slowpath logic Maintain a flag in the LSB of the ticket lock tail which indicates whether anyone is in the lock slowpath and may need kicking when the current holder unlocks. The flags are set when the first locker enters the slowpath, and cleared when unlocking to an empty queue (ie, no contention). In the specific implementation of lock_spinning(), make sure to set the slowpath flags on the lock just before blocking. We must do this before the last-chance pickup test to prevent a deadlock with the unlocker: Unlocker Locker test for lock pickup -> fail unlock test slowpath -> false set slowpath flags block Whereas this works in any ordering: Unlocker Locker set slowpath flags test for lock pickup -> fail block unlock test slowpath -> true, kick If the unlocker finds that the lock has the slowpath flag set but it is actually uncontended (ie, head == tail, so nobody is waiting), then it clears the slowpath flag. The unlock code uses a locked add to update the head counter. This also acts as a full memory barrier so that its safe to subsequently read back the slowflag state, knowing that the updated lock is visible to the other CPUs. If it were an unlocked add, then the flag read may just be forwarded from the store buffer before it was visible to the other CPUs, which could result in a deadlock. Unfortunately this means we need to do a locked instruction when unlocking with PV ticketlocks. However, if PV ticketlocks are not enabled, then the old non-locked "add" is the only unlocking code. Note: this code relies on gcc making sure that unlikely() code is out of line of the fastpath, which only happens when OPTIMIZE_SIZE=n. If it doesn't the generated code isn't too bad, but its definitely suboptimal. Thanks to Srivatsa Vaddagiri for providing a bugfix to the original version of this change, which has been folded in. Thanks to Stephan Diestelhorst for commenting on some code which relied on an inaccurate reading of the x86 memory ordering rules. Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-11-git-send-email-raghavendra.kt@linux.vnet.ibm.com Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Stephan Diestelhorst <stephan.diestelhorst@amd.com> Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:58 +08:00
#else /* !CONFIG_PARAVIRT_SPINLOCKS */
static __always_inline void __ticket_lock_spinning(arch_spinlock_t *lock,
__ticket_t ticket)
x86, spinlock: Replace pv spinlocks with pv ticketlocks Rather than outright replacing the entire spinlock implementation in order to paravirtualize it, keep the ticket lock implementation but add a couple of pvops hooks on the slow patch (long spin on lock, unlocking a contended lock). Ticket locks have a number of nice properties, but they also have some surprising behaviours in virtual environments. They enforce a strict FIFO ordering on cpus trying to take a lock; however, if the hypervisor scheduler does not schedule the cpus in the correct order, the system can waste a huge amount of time spinning until the next cpu can take the lock. (See Thomas Friebel's talk "Prevent Guests from Spinning Around" http://www.xen.org/files/xensummitboston08/LHP.pdf for more details.) To address this, we add two hooks: - __ticket_spin_lock which is called after the cpu has been spinning on the lock for a significant number of iterations but has failed to take the lock (presumably because the cpu holding the lock has been descheduled). The lock_spinning pvop is expected to block the cpu until it has been kicked by the current lock holder. - __ticket_spin_unlock, which on releasing a contended lock (there are more cpus with tail tickets), it looks to see if the next cpu is blocked and wakes it if so. When compiled with CONFIG_PARAVIRT_SPINLOCKS disabled, a set of stub functions causes all the extra code to go away. Results: ======= setup: 32 core machine with 32 vcpu KVM guest (HT off) with 8GB RAM base = 3.11-rc patched = base + pvspinlock V12 +-----------------+----------------+--------+ dbench (Throughput in MB/sec. Higher is better) +-----------------+----------------+--------+ | base (stdev %)|patched(stdev%) | %gain | +-----------------+----------------+--------+ | 15035.3 (0.3) |15150.0 (0.6) | 0.8 | | 1470.0 (2.2) | 1713.7 (1.9) | 16.6 | | 848.6 (4.3) | 967.8 (4.3) | 14.0 | | 652.9 (3.5) | 685.3 (3.7) | 5.0 | +-----------------+----------------+--------+ pvspinlock shows benefits for overcommit ratio > 1 for PLE enabled cases, and undercommits results are flat Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-2-git-send-email-raghavendra.kt@linux.vnet.ibm.com Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Tested-by: Attilio Rao <attilio.rao@citrix.com> [ Raghavendra: Changed SPIN_THRESHOLD, fixed redefinition of arch_spinlock_t] Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:49 +08:00
{
}
x86, ticketlock: Add slowpath logic Maintain a flag in the LSB of the ticket lock tail which indicates whether anyone is in the lock slowpath and may need kicking when the current holder unlocks. The flags are set when the first locker enters the slowpath, and cleared when unlocking to an empty queue (ie, no contention). In the specific implementation of lock_spinning(), make sure to set the slowpath flags on the lock just before blocking. We must do this before the last-chance pickup test to prevent a deadlock with the unlocker: Unlocker Locker test for lock pickup -> fail unlock test slowpath -> false set slowpath flags block Whereas this works in any ordering: Unlocker Locker set slowpath flags test for lock pickup -> fail block unlock test slowpath -> true, kick If the unlocker finds that the lock has the slowpath flag set but it is actually uncontended (ie, head == tail, so nobody is waiting), then it clears the slowpath flag. The unlock code uses a locked add to update the head counter. This also acts as a full memory barrier so that its safe to subsequently read back the slowflag state, knowing that the updated lock is visible to the other CPUs. If it were an unlocked add, then the flag read may just be forwarded from the store buffer before it was visible to the other CPUs, which could result in a deadlock. Unfortunately this means we need to do a locked instruction when unlocking with PV ticketlocks. However, if PV ticketlocks are not enabled, then the old non-locked "add" is the only unlocking code. Note: this code relies on gcc making sure that unlikely() code is out of line of the fastpath, which only happens when OPTIMIZE_SIZE=n. If it doesn't the generated code isn't too bad, but its definitely suboptimal. Thanks to Srivatsa Vaddagiri for providing a bugfix to the original version of this change, which has been folded in. Thanks to Stephan Diestelhorst for commenting on some code which relied on an inaccurate reading of the x86 memory ordering rules. Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-11-git-send-email-raghavendra.kt@linux.vnet.ibm.com Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Stephan Diestelhorst <stephan.diestelhorst@amd.com> Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:58 +08:00
static inline void __ticket_unlock_kick(arch_spinlock_t *lock,
__ticket_t ticket)
x86, spinlock: Replace pv spinlocks with pv ticketlocks Rather than outright replacing the entire spinlock implementation in order to paravirtualize it, keep the ticket lock implementation but add a couple of pvops hooks on the slow patch (long spin on lock, unlocking a contended lock). Ticket locks have a number of nice properties, but they also have some surprising behaviours in virtual environments. They enforce a strict FIFO ordering on cpus trying to take a lock; however, if the hypervisor scheduler does not schedule the cpus in the correct order, the system can waste a huge amount of time spinning until the next cpu can take the lock. (See Thomas Friebel's talk "Prevent Guests from Spinning Around" http://www.xen.org/files/xensummitboston08/LHP.pdf for more details.) To address this, we add two hooks: - __ticket_spin_lock which is called after the cpu has been spinning on the lock for a significant number of iterations but has failed to take the lock (presumably because the cpu holding the lock has been descheduled). The lock_spinning pvop is expected to block the cpu until it has been kicked by the current lock holder. - __ticket_spin_unlock, which on releasing a contended lock (there are more cpus with tail tickets), it looks to see if the next cpu is blocked and wakes it if so. When compiled with CONFIG_PARAVIRT_SPINLOCKS disabled, a set of stub functions causes all the extra code to go away. Results: ======= setup: 32 core machine with 32 vcpu KVM guest (HT off) with 8GB RAM base = 3.11-rc patched = base + pvspinlock V12 +-----------------+----------------+--------+ dbench (Throughput in MB/sec. Higher is better) +-----------------+----------------+--------+ | base (stdev %)|patched(stdev%) | %gain | +-----------------+----------------+--------+ | 15035.3 (0.3) |15150.0 (0.6) | 0.8 | | 1470.0 (2.2) | 1713.7 (1.9) | 16.6 | | 848.6 (4.3) | 967.8 (4.3) | 14.0 | | 652.9 (3.5) | 685.3 (3.7) | 5.0 | +-----------------+----------------+--------+ pvspinlock shows benefits for overcommit ratio > 1 for PLE enabled cases, and undercommits results are flat Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-2-git-send-email-raghavendra.kt@linux.vnet.ibm.com Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Tested-by: Attilio Rao <attilio.rao@citrix.com> [ Raghavendra: Changed SPIN_THRESHOLD, fixed redefinition of arch_spinlock_t] Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:49 +08:00
{
}
x86, ticketlock: Add slowpath logic Maintain a flag in the LSB of the ticket lock tail which indicates whether anyone is in the lock slowpath and may need kicking when the current holder unlocks. The flags are set when the first locker enters the slowpath, and cleared when unlocking to an empty queue (ie, no contention). In the specific implementation of lock_spinning(), make sure to set the slowpath flags on the lock just before blocking. We must do this before the last-chance pickup test to prevent a deadlock with the unlocker: Unlocker Locker test for lock pickup -> fail unlock test slowpath -> false set slowpath flags block Whereas this works in any ordering: Unlocker Locker set slowpath flags test for lock pickup -> fail block unlock test slowpath -> true, kick If the unlocker finds that the lock has the slowpath flag set but it is actually uncontended (ie, head == tail, so nobody is waiting), then it clears the slowpath flag. The unlock code uses a locked add to update the head counter. This also acts as a full memory barrier so that its safe to subsequently read back the slowflag state, knowing that the updated lock is visible to the other CPUs. If it were an unlocked add, then the flag read may just be forwarded from the store buffer before it was visible to the other CPUs, which could result in a deadlock. Unfortunately this means we need to do a locked instruction when unlocking with PV ticketlocks. However, if PV ticketlocks are not enabled, then the old non-locked "add" is the only unlocking code. Note: this code relies on gcc making sure that unlikely() code is out of line of the fastpath, which only happens when OPTIMIZE_SIZE=n. If it doesn't the generated code isn't too bad, but its definitely suboptimal. Thanks to Srivatsa Vaddagiri for providing a bugfix to the original version of this change, which has been folded in. Thanks to Stephan Diestelhorst for commenting on some code which relied on an inaccurate reading of the x86 memory ordering rules. Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-11-git-send-email-raghavendra.kt@linux.vnet.ibm.com Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Stephan Diestelhorst <stephan.diestelhorst@amd.com> Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:58 +08:00
#endif /* CONFIG_PARAVIRT_SPINLOCKS */
lockref: implement lockless reference count updates using cmpxchg() Instead of taking the spinlock, the lockless versions atomically check that the lock is not taken, and do the reference count update using a cmpxchg() loop. This is semantically identical to doing the reference count update protected by the lock, but avoids the "wait for lock" contention that you get when accesses to the reference count are contended. Note that a "lockref" is absolutely _not_ equivalent to an atomic_t. Even when the lockref reference counts are updated atomically with cmpxchg, the fact that they also verify the state of the spinlock means that the lockless updates can never happen while somebody else holds the spinlock. So while "lockref_put_or_lock()" looks a lot like just another name for "atomic_dec_and_lock()", and both optimize to lockless updates, they are fundamentally different: the decrement done by atomic_dec_and_lock() is truly independent of any lock (as long as it doesn't decrement to zero), so a locked region can still see the count change. The lockref structure, in contrast, really is a *locked* reference count. If you hold the spinlock, the reference count will be stable and you can modify the reference count without using atomics, because even the lockless updates will see and respect the state of the lock. In order to enable the cmpxchg lockless code, the architecture needs to do three things: (1) Make sure that the "arch_spinlock_t" and an "unsigned int" can fit in an aligned u64, and have a "cmpxchg()" implementation that works on such a u64 data type. (2) define a helper function to test for a spinlock being unlocked ("arch_spin_value_unlocked()") (3) select the "ARCH_USE_CMPXCHG_LOCKREF" config variable in its Kconfig file. This enables it for x86-64 (but not 32-bit, we'd need to make sure cmpxchg() turns into the proper cmpxchg8b in order to enable it for 32-bit mode). Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-03 03:12:15 +08:00
static __always_inline int arch_spin_value_unlocked(arch_spinlock_t lock)
{
return lock.tickets.head == lock.tickets.tail;
}
/*
* Ticket locks are conceptually two parts, one indicating the current head of
* the queue, and the other indicating the current tail. The lock is acquired
* by atomically noting the tail and incrementing it by one (thus adding
* ourself to the queue and noting our position), then waiting until the head
* becomes equal to the the initial value of the tail.
*
* We use an xadd covering *both* parts of the lock, to increment the tail and
* also load the position of the head, which takes care of memory ordering
* issues and should be optimal for the uncontended case. Note the tail must be
* in the high part, because a wide xadd increment of the low part would carry
* up and contaminate the high part.
*/
x86, ticketlock: Add slowpath logic Maintain a flag in the LSB of the ticket lock tail which indicates whether anyone is in the lock slowpath and may need kicking when the current holder unlocks. The flags are set when the first locker enters the slowpath, and cleared when unlocking to an empty queue (ie, no contention). In the specific implementation of lock_spinning(), make sure to set the slowpath flags on the lock just before blocking. We must do this before the last-chance pickup test to prevent a deadlock with the unlocker: Unlocker Locker test for lock pickup -> fail unlock test slowpath -> false set slowpath flags block Whereas this works in any ordering: Unlocker Locker set slowpath flags test for lock pickup -> fail block unlock test slowpath -> true, kick If the unlocker finds that the lock has the slowpath flag set but it is actually uncontended (ie, head == tail, so nobody is waiting), then it clears the slowpath flag. The unlock code uses a locked add to update the head counter. This also acts as a full memory barrier so that its safe to subsequently read back the slowflag state, knowing that the updated lock is visible to the other CPUs. If it were an unlocked add, then the flag read may just be forwarded from the store buffer before it was visible to the other CPUs, which could result in a deadlock. Unfortunately this means we need to do a locked instruction when unlocking with PV ticketlocks. However, if PV ticketlocks are not enabled, then the old non-locked "add" is the only unlocking code. Note: this code relies on gcc making sure that unlikely() code is out of line of the fastpath, which only happens when OPTIMIZE_SIZE=n. If it doesn't the generated code isn't too bad, but its definitely suboptimal. Thanks to Srivatsa Vaddagiri for providing a bugfix to the original version of this change, which has been folded in. Thanks to Stephan Diestelhorst for commenting on some code which relied on an inaccurate reading of the x86 memory ordering rules. Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-11-git-send-email-raghavendra.kt@linux.vnet.ibm.com Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Stephan Diestelhorst <stephan.diestelhorst@amd.com> Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:58 +08:00
static __always_inline void arch_spin_lock(arch_spinlock_t *lock)
{
register struct __raw_tickets inc = { .tail = TICKET_LOCK_INC };
x86: FIFO ticket spinlocks Introduce ticket lock spinlocks for x86 which are FIFO. The implementation is described in the comments. The straight-line lock/unlock instruction sequence is slightly slower than the dec based locks on modern x86 CPUs, however the difference is quite small on Core2 and Opteron when working out of cache, and becomes almost insignificant even on P4 when the lock misses cache. trylock is more significantly slower, but they are relatively rare. On an 8 core (2 socket) Opteron, spinlock unfairness is extremely noticable, with a userspace test having a difference of up to 2x runtime per thread, and some threads are starved or "unfairly" granted the lock up to 1 000 000 (!) times. After this patch, all threads appear to finish at exactly the same time. The memory ordering of the lock does conform to x86 standards, and the implementation has been reviewed by Intel and AMD engineers. The algorithm also tells us how many CPUs are contending the lock, so lockbreak becomes trivial and we no longer have to waste 4 bytes per spinlock for it. After this, we can no longer spin on any locks with preempt enabled and cannot reenable interrupts when spinning on an irq safe lock, because at that point we have already taken a ticket and the would deadlock if the same CPU tries to take the lock again. These are questionable anyway: if the lock happens to be called under a preempt or interrupt disabled section, then it will just have the same latency problems. The real fix is to keep critical sections short, and ensure locks are reasonably fair (which this patch does). Signed-off-by: Nick Piggin <npiggin@suse.de> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-01-30 20:31:21 +08:00
inc = xadd(&lock->tickets, inc);
x86, ticketlock: Add slowpath logic Maintain a flag in the LSB of the ticket lock tail which indicates whether anyone is in the lock slowpath and may need kicking when the current holder unlocks. The flags are set when the first locker enters the slowpath, and cleared when unlocking to an empty queue (ie, no contention). In the specific implementation of lock_spinning(), make sure to set the slowpath flags on the lock just before blocking. We must do this before the last-chance pickup test to prevent a deadlock with the unlocker: Unlocker Locker test for lock pickup -> fail unlock test slowpath -> false set slowpath flags block Whereas this works in any ordering: Unlocker Locker set slowpath flags test for lock pickup -> fail block unlock test slowpath -> true, kick If the unlocker finds that the lock has the slowpath flag set but it is actually uncontended (ie, head == tail, so nobody is waiting), then it clears the slowpath flag. The unlock code uses a locked add to update the head counter. This also acts as a full memory barrier so that its safe to subsequently read back the slowflag state, knowing that the updated lock is visible to the other CPUs. If it were an unlocked add, then the flag read may just be forwarded from the store buffer before it was visible to the other CPUs, which could result in a deadlock. Unfortunately this means we need to do a locked instruction when unlocking with PV ticketlocks. However, if PV ticketlocks are not enabled, then the old non-locked "add" is the only unlocking code. Note: this code relies on gcc making sure that unlikely() code is out of line of the fastpath, which only happens when OPTIMIZE_SIZE=n. If it doesn't the generated code isn't too bad, but its definitely suboptimal. Thanks to Srivatsa Vaddagiri for providing a bugfix to the original version of this change, which has been folded in. Thanks to Stephan Diestelhorst for commenting on some code which relied on an inaccurate reading of the x86 memory ordering rules. Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-11-git-send-email-raghavendra.kt@linux.vnet.ibm.com Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Stephan Diestelhorst <stephan.diestelhorst@amd.com> Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:58 +08:00
if (likely(inc.head == inc.tail))
goto out;
x86, ticketlock: Add slowpath logic Maintain a flag in the LSB of the ticket lock tail which indicates whether anyone is in the lock slowpath and may need kicking when the current holder unlocks. The flags are set when the first locker enters the slowpath, and cleared when unlocking to an empty queue (ie, no contention). In the specific implementation of lock_spinning(), make sure to set the slowpath flags on the lock just before blocking. We must do this before the last-chance pickup test to prevent a deadlock with the unlocker: Unlocker Locker test for lock pickup -> fail unlock test slowpath -> false set slowpath flags block Whereas this works in any ordering: Unlocker Locker set slowpath flags test for lock pickup -> fail block unlock test slowpath -> true, kick If the unlocker finds that the lock has the slowpath flag set but it is actually uncontended (ie, head == tail, so nobody is waiting), then it clears the slowpath flag. The unlock code uses a locked add to update the head counter. This also acts as a full memory barrier so that its safe to subsequently read back the slowflag state, knowing that the updated lock is visible to the other CPUs. If it were an unlocked add, then the flag read may just be forwarded from the store buffer before it was visible to the other CPUs, which could result in a deadlock. Unfortunately this means we need to do a locked instruction when unlocking with PV ticketlocks. However, if PV ticketlocks are not enabled, then the old non-locked "add" is the only unlocking code. Note: this code relies on gcc making sure that unlikely() code is out of line of the fastpath, which only happens when OPTIMIZE_SIZE=n. If it doesn't the generated code isn't too bad, but its definitely suboptimal. Thanks to Srivatsa Vaddagiri for providing a bugfix to the original version of this change, which has been folded in. Thanks to Stephan Diestelhorst for commenting on some code which relied on an inaccurate reading of the x86 memory ordering rules. Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-11-git-send-email-raghavendra.kt@linux.vnet.ibm.com Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Stephan Diestelhorst <stephan.diestelhorst@amd.com> Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:58 +08:00
inc.tail &= ~TICKET_SLOWPATH_FLAG;
for (;;) {
x86, spinlock: Replace pv spinlocks with pv ticketlocks Rather than outright replacing the entire spinlock implementation in order to paravirtualize it, keep the ticket lock implementation but add a couple of pvops hooks on the slow patch (long spin on lock, unlocking a contended lock). Ticket locks have a number of nice properties, but they also have some surprising behaviours in virtual environments. They enforce a strict FIFO ordering on cpus trying to take a lock; however, if the hypervisor scheduler does not schedule the cpus in the correct order, the system can waste a huge amount of time spinning until the next cpu can take the lock. (See Thomas Friebel's talk "Prevent Guests from Spinning Around" http://www.xen.org/files/xensummitboston08/LHP.pdf for more details.) To address this, we add two hooks: - __ticket_spin_lock which is called after the cpu has been spinning on the lock for a significant number of iterations but has failed to take the lock (presumably because the cpu holding the lock has been descheduled). The lock_spinning pvop is expected to block the cpu until it has been kicked by the current lock holder. - __ticket_spin_unlock, which on releasing a contended lock (there are more cpus with tail tickets), it looks to see if the next cpu is blocked and wakes it if so. When compiled with CONFIG_PARAVIRT_SPINLOCKS disabled, a set of stub functions causes all the extra code to go away. Results: ======= setup: 32 core machine with 32 vcpu KVM guest (HT off) with 8GB RAM base = 3.11-rc patched = base + pvspinlock V12 +-----------------+----------------+--------+ dbench (Throughput in MB/sec. Higher is better) +-----------------+----------------+--------+ | base (stdev %)|patched(stdev%) | %gain | +-----------------+----------------+--------+ | 15035.3 (0.3) |15150.0 (0.6) | 0.8 | | 1470.0 (2.2) | 1713.7 (1.9) | 16.6 | | 848.6 (4.3) | 967.8 (4.3) | 14.0 | | 652.9 (3.5) | 685.3 (3.7) | 5.0 | +-----------------+----------------+--------+ pvspinlock shows benefits for overcommit ratio > 1 for PLE enabled cases, and undercommits results are flat Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-2-git-send-email-raghavendra.kt@linux.vnet.ibm.com Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Tested-by: Attilio Rao <attilio.rao@citrix.com> [ Raghavendra: Changed SPIN_THRESHOLD, fixed redefinition of arch_spinlock_t] Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:49 +08:00
unsigned count = SPIN_THRESHOLD;
do {
x86, ticketlock: Add slowpath logic Maintain a flag in the LSB of the ticket lock tail which indicates whether anyone is in the lock slowpath and may need kicking when the current holder unlocks. The flags are set when the first locker enters the slowpath, and cleared when unlocking to an empty queue (ie, no contention). In the specific implementation of lock_spinning(), make sure to set the slowpath flags on the lock just before blocking. We must do this before the last-chance pickup test to prevent a deadlock with the unlocker: Unlocker Locker test for lock pickup -> fail unlock test slowpath -> false set slowpath flags block Whereas this works in any ordering: Unlocker Locker set slowpath flags test for lock pickup -> fail block unlock test slowpath -> true, kick If the unlocker finds that the lock has the slowpath flag set but it is actually uncontended (ie, head == tail, so nobody is waiting), then it clears the slowpath flag. The unlock code uses a locked add to update the head counter. This also acts as a full memory barrier so that its safe to subsequently read back the slowflag state, knowing that the updated lock is visible to the other CPUs. If it were an unlocked add, then the flag read may just be forwarded from the store buffer before it was visible to the other CPUs, which could result in a deadlock. Unfortunately this means we need to do a locked instruction when unlocking with PV ticketlocks. However, if PV ticketlocks are not enabled, then the old non-locked "add" is the only unlocking code. Note: this code relies on gcc making sure that unlikely() code is out of line of the fastpath, which only happens when OPTIMIZE_SIZE=n. If it doesn't the generated code isn't too bad, but its definitely suboptimal. Thanks to Srivatsa Vaddagiri for providing a bugfix to the original version of this change, which has been folded in. Thanks to Stephan Diestelhorst for commenting on some code which relied on an inaccurate reading of the x86 memory ordering rules. Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-11-git-send-email-raghavendra.kt@linux.vnet.ibm.com Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Stephan Diestelhorst <stephan.diestelhorst@amd.com> Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:58 +08:00
if (ACCESS_ONCE(lock->tickets.head) == inc.tail)
x86, spinlock: Replace pv spinlocks with pv ticketlocks Rather than outright replacing the entire spinlock implementation in order to paravirtualize it, keep the ticket lock implementation but add a couple of pvops hooks on the slow patch (long spin on lock, unlocking a contended lock). Ticket locks have a number of nice properties, but they also have some surprising behaviours in virtual environments. They enforce a strict FIFO ordering on cpus trying to take a lock; however, if the hypervisor scheduler does not schedule the cpus in the correct order, the system can waste a huge amount of time spinning until the next cpu can take the lock. (See Thomas Friebel's talk "Prevent Guests from Spinning Around" http://www.xen.org/files/xensummitboston08/LHP.pdf for more details.) To address this, we add two hooks: - __ticket_spin_lock which is called after the cpu has been spinning on the lock for a significant number of iterations but has failed to take the lock (presumably because the cpu holding the lock has been descheduled). The lock_spinning pvop is expected to block the cpu until it has been kicked by the current lock holder. - __ticket_spin_unlock, which on releasing a contended lock (there are more cpus with tail tickets), it looks to see if the next cpu is blocked and wakes it if so. When compiled with CONFIG_PARAVIRT_SPINLOCKS disabled, a set of stub functions causes all the extra code to go away. Results: ======= setup: 32 core machine with 32 vcpu KVM guest (HT off) with 8GB RAM base = 3.11-rc patched = base + pvspinlock V12 +-----------------+----------------+--------+ dbench (Throughput in MB/sec. Higher is better) +-----------------+----------------+--------+ | base (stdev %)|patched(stdev%) | %gain | +-----------------+----------------+--------+ | 15035.3 (0.3) |15150.0 (0.6) | 0.8 | | 1470.0 (2.2) | 1713.7 (1.9) | 16.6 | | 848.6 (4.3) | 967.8 (4.3) | 14.0 | | 652.9 (3.5) | 685.3 (3.7) | 5.0 | +-----------------+----------------+--------+ pvspinlock shows benefits for overcommit ratio > 1 for PLE enabled cases, and undercommits results are flat Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-2-git-send-email-raghavendra.kt@linux.vnet.ibm.com Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Tested-by: Attilio Rao <attilio.rao@citrix.com> [ Raghavendra: Changed SPIN_THRESHOLD, fixed redefinition of arch_spinlock_t] Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:49 +08:00
goto out;
cpu_relax();
} while (--count);
__ticket_lock_spinning(lock, inc.tail);
}
x86, spinlock: Replace pv spinlocks with pv ticketlocks Rather than outright replacing the entire spinlock implementation in order to paravirtualize it, keep the ticket lock implementation but add a couple of pvops hooks on the slow patch (long spin on lock, unlocking a contended lock). Ticket locks have a number of nice properties, but they also have some surprising behaviours in virtual environments. They enforce a strict FIFO ordering on cpus trying to take a lock; however, if the hypervisor scheduler does not schedule the cpus in the correct order, the system can waste a huge amount of time spinning until the next cpu can take the lock. (See Thomas Friebel's talk "Prevent Guests from Spinning Around" http://www.xen.org/files/xensummitboston08/LHP.pdf for more details.) To address this, we add two hooks: - __ticket_spin_lock which is called after the cpu has been spinning on the lock for a significant number of iterations but has failed to take the lock (presumably because the cpu holding the lock has been descheduled). The lock_spinning pvop is expected to block the cpu until it has been kicked by the current lock holder. - __ticket_spin_unlock, which on releasing a contended lock (there are more cpus with tail tickets), it looks to see if the next cpu is blocked and wakes it if so. When compiled with CONFIG_PARAVIRT_SPINLOCKS disabled, a set of stub functions causes all the extra code to go away. Results: ======= setup: 32 core machine with 32 vcpu KVM guest (HT off) with 8GB RAM base = 3.11-rc patched = base + pvspinlock V12 +-----------------+----------------+--------+ dbench (Throughput in MB/sec. Higher is better) +-----------------+----------------+--------+ | base (stdev %)|patched(stdev%) | %gain | +-----------------+----------------+--------+ | 15035.3 (0.3) |15150.0 (0.6) | 0.8 | | 1470.0 (2.2) | 1713.7 (1.9) | 16.6 | | 848.6 (4.3) | 967.8 (4.3) | 14.0 | | 652.9 (3.5) | 685.3 (3.7) | 5.0 | +-----------------+----------------+--------+ pvspinlock shows benefits for overcommit ratio > 1 for PLE enabled cases, and undercommits results are flat Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-2-git-send-email-raghavendra.kt@linux.vnet.ibm.com Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Tested-by: Attilio Rao <attilio.rao@citrix.com> [ Raghavendra: Changed SPIN_THRESHOLD, fixed redefinition of arch_spinlock_t] Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:49 +08:00
out: barrier(); /* make sure nothing creeps before the lock is taken */
}
x86: FIFO ticket spinlocks Introduce ticket lock spinlocks for x86 which are FIFO. The implementation is described in the comments. The straight-line lock/unlock instruction sequence is slightly slower than the dec based locks on modern x86 CPUs, however the difference is quite small on Core2 and Opteron when working out of cache, and becomes almost insignificant even on P4 when the lock misses cache. trylock is more significantly slower, but they are relatively rare. On an 8 core (2 socket) Opteron, spinlock unfairness is extremely noticable, with a userspace test having a difference of up to 2x runtime per thread, and some threads are starved or "unfairly" granted the lock up to 1 000 000 (!) times. After this patch, all threads appear to finish at exactly the same time. The memory ordering of the lock does conform to x86 standards, and the implementation has been reviewed by Intel and AMD engineers. The algorithm also tells us how many CPUs are contending the lock, so lockbreak becomes trivial and we no longer have to waste 4 bytes per spinlock for it. After this, we can no longer spin on any locks with preempt enabled and cannot reenable interrupts when spinning on an irq safe lock, because at that point we have already taken a ticket and the would deadlock if the same CPU tries to take the lock again. These are questionable anyway: if the lock happens to be called under a preempt or interrupt disabled section, then it will just have the same latency problems. The real fix is to keep critical sections short, and ensure locks are reasonably fair (which this patch does). Signed-off-by: Nick Piggin <npiggin@suse.de> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-01-30 20:31:21 +08:00
static __always_inline int arch_spin_trylock(arch_spinlock_t *lock)
{
arch_spinlock_t old, new;
old.tickets = ACCESS_ONCE(lock->tickets);
x86, ticketlock: Add slowpath logic Maintain a flag in the LSB of the ticket lock tail which indicates whether anyone is in the lock slowpath and may need kicking when the current holder unlocks. The flags are set when the first locker enters the slowpath, and cleared when unlocking to an empty queue (ie, no contention). In the specific implementation of lock_spinning(), make sure to set the slowpath flags on the lock just before blocking. We must do this before the last-chance pickup test to prevent a deadlock with the unlocker: Unlocker Locker test for lock pickup -> fail unlock test slowpath -> false set slowpath flags block Whereas this works in any ordering: Unlocker Locker set slowpath flags test for lock pickup -> fail block unlock test slowpath -> true, kick If the unlocker finds that the lock has the slowpath flag set but it is actually uncontended (ie, head == tail, so nobody is waiting), then it clears the slowpath flag. The unlock code uses a locked add to update the head counter. This also acts as a full memory barrier so that its safe to subsequently read back the slowflag state, knowing that the updated lock is visible to the other CPUs. If it were an unlocked add, then the flag read may just be forwarded from the store buffer before it was visible to the other CPUs, which could result in a deadlock. Unfortunately this means we need to do a locked instruction when unlocking with PV ticketlocks. However, if PV ticketlocks are not enabled, then the old non-locked "add" is the only unlocking code. Note: this code relies on gcc making sure that unlikely() code is out of line of the fastpath, which only happens when OPTIMIZE_SIZE=n. If it doesn't the generated code isn't too bad, but its definitely suboptimal. Thanks to Srivatsa Vaddagiri for providing a bugfix to the original version of this change, which has been folded in. Thanks to Stephan Diestelhorst for commenting on some code which relied on an inaccurate reading of the x86 memory ordering rules. Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-11-git-send-email-raghavendra.kt@linux.vnet.ibm.com Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Stephan Diestelhorst <stephan.diestelhorst@amd.com> Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:58 +08:00
if (old.tickets.head != (old.tickets.tail & ~TICKET_SLOWPATH_FLAG))
return 0;
x86: FIFO ticket spinlocks Introduce ticket lock spinlocks for x86 which are FIFO. The implementation is described in the comments. The straight-line lock/unlock instruction sequence is slightly slower than the dec based locks on modern x86 CPUs, however the difference is quite small on Core2 and Opteron when working out of cache, and becomes almost insignificant even on P4 when the lock misses cache. trylock is more significantly slower, but they are relatively rare. On an 8 core (2 socket) Opteron, spinlock unfairness is extremely noticable, with a userspace test having a difference of up to 2x runtime per thread, and some threads are starved or "unfairly" granted the lock up to 1 000 000 (!) times. After this patch, all threads appear to finish at exactly the same time. The memory ordering of the lock does conform to x86 standards, and the implementation has been reviewed by Intel and AMD engineers. The algorithm also tells us how many CPUs are contending the lock, so lockbreak becomes trivial and we no longer have to waste 4 bytes per spinlock for it. After this, we can no longer spin on any locks with preempt enabled and cannot reenable interrupts when spinning on an irq safe lock, because at that point we have already taken a ticket and the would deadlock if the same CPU tries to take the lock again. These are questionable anyway: if the lock happens to be called under a preempt or interrupt disabled section, then it will just have the same latency problems. The real fix is to keep critical sections short, and ensure locks are reasonably fair (which this patch does). Signed-off-by: Nick Piggin <npiggin@suse.de> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-01-30 20:31:21 +08:00
new.head_tail = old.head_tail + (TICKET_LOCK_INC << TICKET_SHIFT);
/* 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;
}
x86, ticketlock: Add slowpath logic Maintain a flag in the LSB of the ticket lock tail which indicates whether anyone is in the lock slowpath and may need kicking when the current holder unlocks. The flags are set when the first locker enters the slowpath, and cleared when unlocking to an empty queue (ie, no contention). In the specific implementation of lock_spinning(), make sure to set the slowpath flags on the lock just before blocking. We must do this before the last-chance pickup test to prevent a deadlock with the unlocker: Unlocker Locker test for lock pickup -> fail unlock test slowpath -> false set slowpath flags block Whereas this works in any ordering: Unlocker Locker set slowpath flags test for lock pickup -> fail block unlock test slowpath -> true, kick If the unlocker finds that the lock has the slowpath flag set but it is actually uncontended (ie, head == tail, so nobody is waiting), then it clears the slowpath flag. The unlock code uses a locked add to update the head counter. This also acts as a full memory barrier so that its safe to subsequently read back the slowflag state, knowing that the updated lock is visible to the other CPUs. If it were an unlocked add, then the flag read may just be forwarded from the store buffer before it was visible to the other CPUs, which could result in a deadlock. Unfortunately this means we need to do a locked instruction when unlocking with PV ticketlocks. However, if PV ticketlocks are not enabled, then the old non-locked "add" is the only unlocking code. Note: this code relies on gcc making sure that unlikely() code is out of line of the fastpath, which only happens when OPTIMIZE_SIZE=n. If it doesn't the generated code isn't too bad, but its definitely suboptimal. Thanks to Srivatsa Vaddagiri for providing a bugfix to the original version of this change, which has been folded in. Thanks to Stephan Diestelhorst for commenting on some code which relied on an inaccurate reading of the x86 memory ordering rules. Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-11-git-send-email-raghavendra.kt@linux.vnet.ibm.com Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Stephan Diestelhorst <stephan.diestelhorst@amd.com> Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:58 +08:00
static inline void __ticket_unlock_slowpath(arch_spinlock_t *lock,
arch_spinlock_t old)
{
x86, ticketlock: Add slowpath logic Maintain a flag in the LSB of the ticket lock tail which indicates whether anyone is in the lock slowpath and may need kicking when the current holder unlocks. The flags are set when the first locker enters the slowpath, and cleared when unlocking to an empty queue (ie, no contention). In the specific implementation of lock_spinning(), make sure to set the slowpath flags on the lock just before blocking. We must do this before the last-chance pickup test to prevent a deadlock with the unlocker: Unlocker Locker test for lock pickup -> fail unlock test slowpath -> false set slowpath flags block Whereas this works in any ordering: Unlocker Locker set slowpath flags test for lock pickup -> fail block unlock test slowpath -> true, kick If the unlocker finds that the lock has the slowpath flag set but it is actually uncontended (ie, head == tail, so nobody is waiting), then it clears the slowpath flag. The unlock code uses a locked add to update the head counter. This also acts as a full memory barrier so that its safe to subsequently read back the slowflag state, knowing that the updated lock is visible to the other CPUs. If it were an unlocked add, then the flag read may just be forwarded from the store buffer before it was visible to the other CPUs, which could result in a deadlock. Unfortunately this means we need to do a locked instruction when unlocking with PV ticketlocks. However, if PV ticketlocks are not enabled, then the old non-locked "add" is the only unlocking code. Note: this code relies on gcc making sure that unlikely() code is out of line of the fastpath, which only happens when OPTIMIZE_SIZE=n. If it doesn't the generated code isn't too bad, but its definitely suboptimal. Thanks to Srivatsa Vaddagiri for providing a bugfix to the original version of this change, which has been folded in. Thanks to Stephan Diestelhorst for commenting on some code which relied on an inaccurate reading of the x86 memory ordering rules. Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-11-git-send-email-raghavendra.kt@linux.vnet.ibm.com Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Stephan Diestelhorst <stephan.diestelhorst@amd.com> Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:58 +08:00
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)
{
x86, ticketlock: Add slowpath logic Maintain a flag in the LSB of the ticket lock tail which indicates whether anyone is in the lock slowpath and may need kicking when the current holder unlocks. The flags are set when the first locker enters the slowpath, and cleared when unlocking to an empty queue (ie, no contention). In the specific implementation of lock_spinning(), make sure to set the slowpath flags on the lock just before blocking. We must do this before the last-chance pickup test to prevent a deadlock with the unlocker: Unlocker Locker test for lock pickup -> fail unlock test slowpath -> false set slowpath flags block Whereas this works in any ordering: Unlocker Locker set slowpath flags test for lock pickup -> fail block unlock test slowpath -> true, kick If the unlocker finds that the lock has the slowpath flag set but it is actually uncontended (ie, head == tail, so nobody is waiting), then it clears the slowpath flag. The unlock code uses a locked add to update the head counter. This also acts as a full memory barrier so that its safe to subsequently read back the slowflag state, knowing that the updated lock is visible to the other CPUs. If it were an unlocked add, then the flag read may just be forwarded from the store buffer before it was visible to the other CPUs, which could result in a deadlock. Unfortunately this means we need to do a locked instruction when unlocking with PV ticketlocks. However, if PV ticketlocks are not enabled, then the old non-locked "add" is the only unlocking code. Note: this code relies on gcc making sure that unlikely() code is out of line of the fastpath, which only happens when OPTIMIZE_SIZE=n. If it doesn't the generated code isn't too bad, but its definitely suboptimal. Thanks to Srivatsa Vaddagiri for providing a bugfix to the original version of this change, which has been folded in. Thanks to Stephan Diestelhorst for commenting on some code which relied on an inaccurate reading of the x86 memory ordering rules. Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-11-git-send-email-raghavendra.kt@linux.vnet.ibm.com Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Stephan Diestelhorst <stephan.diestelhorst@amd.com> Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:58 +08:00
if (TICKET_SLOWPATH_FLAG &&
static_key_false(&paravirt_ticketlocks_enabled)) {
arch_spinlock_t prev;
x86, ticketlock: Add slowpath logic Maintain a flag in the LSB of the ticket lock tail which indicates whether anyone is in the lock slowpath and may need kicking when the current holder unlocks. The flags are set when the first locker enters the slowpath, and cleared when unlocking to an empty queue (ie, no contention). In the specific implementation of lock_spinning(), make sure to set the slowpath flags on the lock just before blocking. We must do this before the last-chance pickup test to prevent a deadlock with the unlocker: Unlocker Locker test for lock pickup -> fail unlock test slowpath -> false set slowpath flags block Whereas this works in any ordering: Unlocker Locker set slowpath flags test for lock pickup -> fail block unlock test slowpath -> true, kick If the unlocker finds that the lock has the slowpath flag set but it is actually uncontended (ie, head == tail, so nobody is waiting), then it clears the slowpath flag. The unlock code uses a locked add to update the head counter. This also acts as a full memory barrier so that its safe to subsequently read back the slowflag state, knowing that the updated lock is visible to the other CPUs. If it were an unlocked add, then the flag read may just be forwarded from the store buffer before it was visible to the other CPUs, which could result in a deadlock. Unfortunately this means we need to do a locked instruction when unlocking with PV ticketlocks. However, if PV ticketlocks are not enabled, then the old non-locked "add" is the only unlocking code. Note: this code relies on gcc making sure that unlikely() code is out of line of the fastpath, which only happens when OPTIMIZE_SIZE=n. If it doesn't the generated code isn't too bad, but its definitely suboptimal. Thanks to Srivatsa Vaddagiri for providing a bugfix to the original version of this change, which has been folded in. Thanks to Stephan Diestelhorst for commenting on some code which relied on an inaccurate reading of the x86 memory ordering rules. Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-11-git-send-email-raghavendra.kt@linux.vnet.ibm.com Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Stephan Diestelhorst <stephan.diestelhorst@amd.com> Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:58 +08:00
prev = *lock;
add_smp(&lock->tickets.head, TICKET_LOCK_INC);
x86, ticketlock: Add slowpath logic Maintain a flag in the LSB of the ticket lock tail which indicates whether anyone is in the lock slowpath and may need kicking when the current holder unlocks. The flags are set when the first locker enters the slowpath, and cleared when unlocking to an empty queue (ie, no contention). In the specific implementation of lock_spinning(), make sure to set the slowpath flags on the lock just before blocking. We must do this before the last-chance pickup test to prevent a deadlock with the unlocker: Unlocker Locker test for lock pickup -> fail unlock test slowpath -> false set slowpath flags block Whereas this works in any ordering: Unlocker Locker set slowpath flags test for lock pickup -> fail block unlock test slowpath -> true, kick If the unlocker finds that the lock has the slowpath flag set but it is actually uncontended (ie, head == tail, so nobody is waiting), then it clears the slowpath flag. The unlock code uses a locked add to update the head counter. This also acts as a full memory barrier so that its safe to subsequently read back the slowflag state, knowing that the updated lock is visible to the other CPUs. If it were an unlocked add, then the flag read may just be forwarded from the store buffer before it was visible to the other CPUs, which could result in a deadlock. Unfortunately this means we need to do a locked instruction when unlocking with PV ticketlocks. However, if PV ticketlocks are not enabled, then the old non-locked "add" is the only unlocking code. Note: this code relies on gcc making sure that unlikely() code is out of line of the fastpath, which only happens when OPTIMIZE_SIZE=n. If it doesn't the generated code isn't too bad, but its definitely suboptimal. Thanks to Srivatsa Vaddagiri for providing a bugfix to the original version of this change, which has been folded in. Thanks to Stephan Diestelhorst for commenting on some code which relied on an inaccurate reading of the x86 memory ordering rules. Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-11-git-send-email-raghavendra.kt@linux.vnet.ibm.com Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Stephan Diestelhorst <stephan.diestelhorst@amd.com> Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:58 +08:00
/* add_smp() is a full mb() */
x86, ticketlock: Add slowpath logic Maintain a flag in the LSB of the ticket lock tail which indicates whether anyone is in the lock slowpath and may need kicking when the current holder unlocks. The flags are set when the first locker enters the slowpath, and cleared when unlocking to an empty queue (ie, no contention). In the specific implementation of lock_spinning(), make sure to set the slowpath flags on the lock just before blocking. We must do this before the last-chance pickup test to prevent a deadlock with the unlocker: Unlocker Locker test for lock pickup -> fail unlock test slowpath -> false set slowpath flags block Whereas this works in any ordering: Unlocker Locker set slowpath flags test for lock pickup -> fail block unlock test slowpath -> true, kick If the unlocker finds that the lock has the slowpath flag set but it is actually uncontended (ie, head == tail, so nobody is waiting), then it clears the slowpath flag. The unlock code uses a locked add to update the head counter. This also acts as a full memory barrier so that its safe to subsequently read back the slowflag state, knowing that the updated lock is visible to the other CPUs. If it were an unlocked add, then the flag read may just be forwarded from the store buffer before it was visible to the other CPUs, which could result in a deadlock. Unfortunately this means we need to do a locked instruction when unlocking with PV ticketlocks. However, if PV ticketlocks are not enabled, then the old non-locked "add" is the only unlocking code. Note: this code relies on gcc making sure that unlikely() code is out of line of the fastpath, which only happens when OPTIMIZE_SIZE=n. If it doesn't the generated code isn't too bad, but its definitely suboptimal. Thanks to Srivatsa Vaddagiri for providing a bugfix to the original version of this change, which has been folded in. Thanks to Stephan Diestelhorst for commenting on some code which relied on an inaccurate reading of the x86 memory ordering rules. Signed-off-by: Jeremy Fitzhardinge <jeremy@goop.org> Link: http://lkml.kernel.org/r/1376058122-8248-11-git-send-email-raghavendra.kt@linux.vnet.ibm.com Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Stephan Diestelhorst <stephan.diestelhorst@amd.com> Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2013-08-09 22:21:58 +08:00
if (unlikely(lock->tickets.tail & TICKET_SLOWPATH_FLAG))
__ticket_unlock_slowpath(lock, prev);
} else
__add(&lock->tickets.head, TICKET_LOCK_INC, UNLOCK_LOCK_PREFIX);
}
x86/paravirt: add hooks for spinlock operations Ticket spinlocks have absolutely ghastly worst-case performance characteristics in a virtual environment. If there is any contention for physical CPUs (ie, there are more runnable vcpus than cpus), then ticket locks can cause the system to end up spending 90+% of its time spinning. The problem is that (v)cpus waiting on a ticket spinlock will be granted access to the lock in strict order they got their tickets. If the hypervisor scheduler doesn't give the vcpus time in that order, they will burn timeslices waiting for the scheduler to give the right vcpu some time. In the worst case it could take O(n^2) vcpu scheduler timeslices for everyone waiting on the lock to get it, not counting new cpus trying to take the lock while the log-jam is sorted out. These hooks allow a paravirt backend to replace the spinlock implementation. At the very least, this could revert the implementation back to the old lock algorithm, which allows the next scheduled vcpu to take the lock, and has basically fairly good performance. It also allows the spinlocks to take advantages of the hypervisor features to make locks more efficient (spin and block, for example). The cost to native execution is an extra direct call when using a spinlock function. There's no overhead if CONFIG_PARAVIRT is turned off. The lock structure is fixed at a single "unsigned int", initialized to zero, but the spinlock implementation can use it as it wishes. Thanks to Thomas Friebel's Xen Summit talk "Preventing Guests from Spinning Around" for pointing out this problem. Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Christoph Lameter <clameter@linux-foundation.org> Cc: Petr Tesarik <ptesarik@suse.cz> Cc: Virtualization <virtualization@lists.linux-foundation.org> Cc: Xen devel <xen-devel@lists.xensource.com> Cc: Thomas Friebel <thomas.friebel@amd.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-07-08 03:07:50 +08:00
static inline int arch_spin_is_locked(arch_spinlock_t *lock)
x86/paravirt: add hooks for spinlock operations Ticket spinlocks have absolutely ghastly worst-case performance characteristics in a virtual environment. If there is any contention for physical CPUs (ie, there are more runnable vcpus than cpus), then ticket locks can cause the system to end up spending 90+% of its time spinning. The problem is that (v)cpus waiting on a ticket spinlock will be granted access to the lock in strict order they got their tickets. If the hypervisor scheduler doesn't give the vcpus time in that order, they will burn timeslices waiting for the scheduler to give the right vcpu some time. In the worst case it could take O(n^2) vcpu scheduler timeslices for everyone waiting on the lock to get it, not counting new cpus trying to take the lock while the log-jam is sorted out. These hooks allow a paravirt backend to replace the spinlock implementation. At the very least, this could revert the implementation back to the old lock algorithm, which allows the next scheduled vcpu to take the lock, and has basically fairly good performance. It also allows the spinlocks to take advantages of the hypervisor features to make locks more efficient (spin and block, for example). The cost to native execution is an extra direct call when using a spinlock function. There's no overhead if CONFIG_PARAVIRT is turned off. The lock structure is fixed at a single "unsigned int", initialized to zero, but the spinlock implementation can use it as it wishes. Thanks to Thomas Friebel's Xen Summit talk "Preventing Guests from Spinning Around" for pointing out this problem. Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Christoph Lameter <clameter@linux-foundation.org> Cc: Petr Tesarik <ptesarik@suse.cz> Cc: Virtualization <virtualization@lists.linux-foundation.org> Cc: Xen devel <xen-devel@lists.xensource.com> Cc: Thomas Friebel <thomas.friebel@amd.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-07-08 03:07:50 +08:00
{
struct __raw_tickets tmp = ACCESS_ONCE(lock->tickets);
x86/paravirt: add hooks for spinlock operations Ticket spinlocks have absolutely ghastly worst-case performance characteristics in a virtual environment. If there is any contention for physical CPUs (ie, there are more runnable vcpus than cpus), then ticket locks can cause the system to end up spending 90+% of its time spinning. The problem is that (v)cpus waiting on a ticket spinlock will be granted access to the lock in strict order they got their tickets. If the hypervisor scheduler doesn't give the vcpus time in that order, they will burn timeslices waiting for the scheduler to give the right vcpu some time. In the worst case it could take O(n^2) vcpu scheduler timeslices for everyone waiting on the lock to get it, not counting new cpus trying to take the lock while the log-jam is sorted out. These hooks allow a paravirt backend to replace the spinlock implementation. At the very least, this could revert the implementation back to the old lock algorithm, which allows the next scheduled vcpu to take the lock, and has basically fairly good performance. It also allows the spinlocks to take advantages of the hypervisor features to make locks more efficient (spin and block, for example). The cost to native execution is an extra direct call when using a spinlock function. There's no overhead if CONFIG_PARAVIRT is turned off. The lock structure is fixed at a single "unsigned int", initialized to zero, but the spinlock implementation can use it as it wishes. Thanks to Thomas Friebel's Xen Summit talk "Preventing Guests from Spinning Around" for pointing out this problem. Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Christoph Lameter <clameter@linux-foundation.org> Cc: Petr Tesarik <ptesarik@suse.cz> Cc: Virtualization <virtualization@lists.linux-foundation.org> Cc: Xen devel <xen-devel@lists.xensource.com> Cc: Thomas Friebel <thomas.friebel@amd.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-07-08 03:07:50 +08:00
return tmp.tail != tmp.head;
x86/paravirt: add hooks for spinlock operations Ticket spinlocks have absolutely ghastly worst-case performance characteristics in a virtual environment. If there is any contention for physical CPUs (ie, there are more runnable vcpus than cpus), then ticket locks can cause the system to end up spending 90+% of its time spinning. The problem is that (v)cpus waiting on a ticket spinlock will be granted access to the lock in strict order they got their tickets. If the hypervisor scheduler doesn't give the vcpus time in that order, they will burn timeslices waiting for the scheduler to give the right vcpu some time. In the worst case it could take O(n^2) vcpu scheduler timeslices for everyone waiting on the lock to get it, not counting new cpus trying to take the lock while the log-jam is sorted out. These hooks allow a paravirt backend to replace the spinlock implementation. At the very least, this could revert the implementation back to the old lock algorithm, which allows the next scheduled vcpu to take the lock, and has basically fairly good performance. It also allows the spinlocks to take advantages of the hypervisor features to make locks more efficient (spin and block, for example). The cost to native execution is an extra direct call when using a spinlock function. There's no overhead if CONFIG_PARAVIRT is turned off. The lock structure is fixed at a single "unsigned int", initialized to zero, but the spinlock implementation can use it as it wishes. Thanks to Thomas Friebel's Xen Summit talk "Preventing Guests from Spinning Around" for pointing out this problem. Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Christoph Lameter <clameter@linux-foundation.org> Cc: Petr Tesarik <ptesarik@suse.cz> Cc: Virtualization <virtualization@lists.linux-foundation.org> Cc: Xen devel <xen-devel@lists.xensource.com> Cc: Thomas Friebel <thomas.friebel@amd.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-07-08 03:07:50 +08:00
}
static inline int arch_spin_is_contended(arch_spinlock_t *lock)
x86/paravirt: add hooks for spinlock operations Ticket spinlocks have absolutely ghastly worst-case performance characteristics in a virtual environment. If there is any contention for physical CPUs (ie, there are more runnable vcpus than cpus), then ticket locks can cause the system to end up spending 90+% of its time spinning. The problem is that (v)cpus waiting on a ticket spinlock will be granted access to the lock in strict order they got their tickets. If the hypervisor scheduler doesn't give the vcpus time in that order, they will burn timeslices waiting for the scheduler to give the right vcpu some time. In the worst case it could take O(n^2) vcpu scheduler timeslices for everyone waiting on the lock to get it, not counting new cpus trying to take the lock while the log-jam is sorted out. These hooks allow a paravirt backend to replace the spinlock implementation. At the very least, this could revert the implementation back to the old lock algorithm, which allows the next scheduled vcpu to take the lock, and has basically fairly good performance. It also allows the spinlocks to take advantages of the hypervisor features to make locks more efficient (spin and block, for example). The cost to native execution is an extra direct call when using a spinlock function. There's no overhead if CONFIG_PARAVIRT is turned off. The lock structure is fixed at a single "unsigned int", initialized to zero, but the spinlock implementation can use it as it wishes. Thanks to Thomas Friebel's Xen Summit talk "Preventing Guests from Spinning Around" for pointing out this problem. Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Christoph Lameter <clameter@linux-foundation.org> Cc: Petr Tesarik <ptesarik@suse.cz> Cc: Virtualization <virtualization@lists.linux-foundation.org> Cc: Xen devel <xen-devel@lists.xensource.com> Cc: Thomas Friebel <thomas.friebel@amd.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-07-08 03:07:50 +08:00
{
struct __raw_tickets tmp = ACCESS_ONCE(lock->tickets);
x86/paravirt: add hooks for spinlock operations Ticket spinlocks have absolutely ghastly worst-case performance characteristics in a virtual environment. If there is any contention for physical CPUs (ie, there are more runnable vcpus than cpus), then ticket locks can cause the system to end up spending 90+% of its time spinning. The problem is that (v)cpus waiting on a ticket spinlock will be granted access to the lock in strict order they got their tickets. If the hypervisor scheduler doesn't give the vcpus time in that order, they will burn timeslices waiting for the scheduler to give the right vcpu some time. In the worst case it could take O(n^2) vcpu scheduler timeslices for everyone waiting on the lock to get it, not counting new cpus trying to take the lock while the log-jam is sorted out. These hooks allow a paravirt backend to replace the spinlock implementation. At the very least, this could revert the implementation back to the old lock algorithm, which allows the next scheduled vcpu to take the lock, and has basically fairly good performance. It also allows the spinlocks to take advantages of the hypervisor features to make locks more efficient (spin and block, for example). The cost to native execution is an extra direct call when using a spinlock function. There's no overhead if CONFIG_PARAVIRT is turned off. The lock structure is fixed at a single "unsigned int", initialized to zero, but the spinlock implementation can use it as it wishes. Thanks to Thomas Friebel's Xen Summit talk "Preventing Guests from Spinning Around" for pointing out this problem. Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Christoph Lameter <clameter@linux-foundation.org> Cc: Petr Tesarik <ptesarik@suse.cz> Cc: Virtualization <virtualization@lists.linux-foundation.org> Cc: Xen devel <xen-devel@lists.xensource.com> Cc: Thomas Friebel <thomas.friebel@amd.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-07-08 03:07:50 +08:00
return (__ticket_t)(tmp.tail - tmp.head) > TICKET_LOCK_INC;
x86/paravirt: add hooks for spinlock operations Ticket spinlocks have absolutely ghastly worst-case performance characteristics in a virtual environment. If there is any contention for physical CPUs (ie, there are more runnable vcpus than cpus), then ticket locks can cause the system to end up spending 90+% of its time spinning. The problem is that (v)cpus waiting on a ticket spinlock will be granted access to the lock in strict order they got their tickets. If the hypervisor scheduler doesn't give the vcpus time in that order, they will burn timeslices waiting for the scheduler to give the right vcpu some time. In the worst case it could take O(n^2) vcpu scheduler timeslices for everyone waiting on the lock to get it, not counting new cpus trying to take the lock while the log-jam is sorted out. These hooks allow a paravirt backend to replace the spinlock implementation. At the very least, this could revert the implementation back to the old lock algorithm, which allows the next scheduled vcpu to take the lock, and has basically fairly good performance. It also allows the spinlocks to take advantages of the hypervisor features to make locks more efficient (spin and block, for example). The cost to native execution is an extra direct call when using a spinlock function. There's no overhead if CONFIG_PARAVIRT is turned off. The lock structure is fixed at a single "unsigned int", initialized to zero, but the spinlock implementation can use it as it wishes. Thanks to Thomas Friebel's Xen Summit talk "Preventing Guests from Spinning Around" for pointing out this problem. Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Christoph Lameter <clameter@linux-foundation.org> Cc: Petr Tesarik <ptesarik@suse.cz> Cc: Virtualization <virtualization@lists.linux-foundation.org> Cc: Xen devel <xen-devel@lists.xensource.com> Cc: Thomas Friebel <thomas.friebel@amd.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-07-08 03:07:50 +08:00
}
#define arch_spin_is_contended arch_spin_is_contended
static __always_inline void arch_spin_lock_flags(arch_spinlock_t *lock,
unsigned long flags)
{
arch_spin_lock(lock);
}
static inline void arch_spin_unlock_wait(arch_spinlock_t *lock)
{
while (arch_spin_is_locked(lock))
cpu_relax();
}
/*
* Read-write spinlocks, allowing multiple readers
* but only one writer.
*
* NOTE! it is quite common to have readers in interrupts
* but no interrupt writers. For those circumstances we
* can "mix" irq-safe locks - any writer needs to get a
* irq-safe write-lock, but readers can get non-irqsafe
* read-locks.
*
* On x86, we implement read-write locks as a 32-bit counter
* with the high bit (sign) being the "contended" bit.
*/
x86: FIFO ticket spinlocks Introduce ticket lock spinlocks for x86 which are FIFO. The implementation is described in the comments. The straight-line lock/unlock instruction sequence is slightly slower than the dec based locks on modern x86 CPUs, however the difference is quite small on Core2 and Opteron when working out of cache, and becomes almost insignificant even on P4 when the lock misses cache. trylock is more significantly slower, but they are relatively rare. On an 8 core (2 socket) Opteron, spinlock unfairness is extremely noticable, with a userspace test having a difference of up to 2x runtime per thread, and some threads are starved or "unfairly" granted the lock up to 1 000 000 (!) times. After this patch, all threads appear to finish at exactly the same time. The memory ordering of the lock does conform to x86 standards, and the implementation has been reviewed by Intel and AMD engineers. The algorithm also tells us how many CPUs are contending the lock, so lockbreak becomes trivial and we no longer have to waste 4 bytes per spinlock for it. After this, we can no longer spin on any locks with preempt enabled and cannot reenable interrupts when spinning on an irq safe lock, because at that point we have already taken a ticket and the would deadlock if the same CPU tries to take the lock again. These are questionable anyway: if the lock happens to be called under a preempt or interrupt disabled section, then it will just have the same latency problems. The real fix is to keep critical sections short, and ensure locks are reasonably fair (which this patch does). Signed-off-by: Nick Piggin <npiggin@suse.de> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-01-30 20:31:21 +08:00
/**
* read_can_lock - would read_trylock() succeed?
* @lock: the rwlock in question.
*/
static inline int arch_read_can_lock(arch_rwlock_t *lock)
{
return lock->lock > 0;
}
x86: FIFO ticket spinlocks Introduce ticket lock spinlocks for x86 which are FIFO. The implementation is described in the comments. The straight-line lock/unlock instruction sequence is slightly slower than the dec based locks on modern x86 CPUs, however the difference is quite small on Core2 and Opteron when working out of cache, and becomes almost insignificant even on P4 when the lock misses cache. trylock is more significantly slower, but they are relatively rare. On an 8 core (2 socket) Opteron, spinlock unfairness is extremely noticable, with a userspace test having a difference of up to 2x runtime per thread, and some threads are starved or "unfairly" granted the lock up to 1 000 000 (!) times. After this patch, all threads appear to finish at exactly the same time. The memory ordering of the lock does conform to x86 standards, and the implementation has been reviewed by Intel and AMD engineers. The algorithm also tells us how many CPUs are contending the lock, so lockbreak becomes trivial and we no longer have to waste 4 bytes per spinlock for it. After this, we can no longer spin on any locks with preempt enabled and cannot reenable interrupts when spinning on an irq safe lock, because at that point we have already taken a ticket and the would deadlock if the same CPU tries to take the lock again. These are questionable anyway: if the lock happens to be called under a preempt or interrupt disabled section, then it will just have the same latency problems. The real fix is to keep critical sections short, and ensure locks are reasonably fair (which this patch does). Signed-off-by: Nick Piggin <npiggin@suse.de> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-01-30 20:31:21 +08:00
/**
* write_can_lock - would write_trylock() succeed?
* @lock: the rwlock in question.
*/
static inline int arch_write_can_lock(arch_rwlock_t *lock)
{
return lock->write == WRITE_LOCK_CMP;
}
static inline void arch_read_lock(arch_rwlock_t *rw)
{
asm volatile(LOCK_PREFIX READ_LOCK_SIZE(dec) " (%0)\n\t"
"jns 1f\n"
"call __read_lock_failed\n\t"
"1:\n"
::LOCK_PTR_REG (rw) : "memory");
}
static inline void arch_write_lock(arch_rwlock_t *rw)
{
asm volatile(LOCK_PREFIX WRITE_LOCK_SUB(%1) "(%0)\n\t"
"jz 1f\n"
"call __write_lock_failed\n\t"
"1:\n"
::LOCK_PTR_REG (&rw->write), "i" (RW_LOCK_BIAS)
: "memory");
}
static inline int arch_read_trylock(arch_rwlock_t *lock)
{
READ_LOCK_ATOMIC(t) *count = (READ_LOCK_ATOMIC(t) *)lock;
if (READ_LOCK_ATOMIC(dec_return)(count) >= 0)
return 1;
READ_LOCK_ATOMIC(inc)(count);
return 0;
}
static inline int arch_write_trylock(arch_rwlock_t *lock)
{
atomic_t *count = (atomic_t *)&lock->write;
if (atomic_sub_and_test(WRITE_LOCK_CMP, count))
return 1;
atomic_add(WRITE_LOCK_CMP, count);
return 0;
}
static inline void arch_read_unlock(arch_rwlock_t *rw)
{
asm volatile(LOCK_PREFIX READ_LOCK_SIZE(inc) " %0"
:"+m" (rw->lock) : : "memory");
}
static inline void arch_write_unlock(arch_rwlock_t *rw)
{
asm volatile(LOCK_PREFIX WRITE_LOCK_ADD(%1) "%0"
: "+m" (rw->write) : "i" (RW_LOCK_BIAS) : "memory");
}
#define arch_read_lock_flags(lock, flags) arch_read_lock(lock)
#define arch_write_lock_flags(lock, flags) arch_write_lock(lock)
#undef READ_LOCK_SIZE
#undef READ_LOCK_ATOMIC
#undef WRITE_LOCK_ADD
#undef WRITE_LOCK_SUB
#undef WRITE_LOCK_CMP
#define arch_spin_relax(lock) cpu_relax()
#define arch_read_relax(lock) cpu_relax()
#define arch_write_relax(lock) cpu_relax()
#endif /* _ASM_X86_SPINLOCK_H */