linux/kernel/locking/qspinlock_paravirt.h
Waiman Long 11752adb68 locking/pvqspinlock: Implement hybrid PV queued/unfair locks
Currently, all the lock waiters entering the slowpath will do one
lock stealing attempt to acquire the lock. That helps performance,
especially in VMs with over-committed vCPUs. However, the current
pvqspinlocks still don't perform as good as unfair locks in many cases.
On the other hands, unfair locks do have the problem of lock starvation
that pvqspinlocks don't have.

This patch combines the best attributes of an unfair lock and a
pvqspinlock into a hybrid lock with 2 modes - queued mode & unfair
mode. A lock waiter goes into the unfair mode when there are waiters
in the wait queue but the pending bit isn't set. Otherwise, it will
go into the queued mode waiting in the queue for its turn.

On a 2-socket 36-core E5-2699 v3 system (HT off), a kernel build
(make -j<n>) was done in a VM with unpinned vCPUs 3 times with the
best time selected and <n> is the number of vCPUs available. The build
times of the original pvqspinlock, hybrid pvqspinlock and unfair lock
with various number of vCPUs are as follows:

  vCPUs    pvqlock     hybrid pvqlock    unfair lock
  -----    -------     --------------    -----------
    30      342.1s         329.1s          329.1s
    36      314.1s         305.3s          307.3s
    45      345.0s         302.1s          306.6s
    54      365.4s         308.6s          307.8s
    72      358.9s         293.6s          303.9s
   108      343.0s         285.9s          304.2s

The hybrid pvqspinlock performs better or comparable to the unfair
lock.

By turning on QUEUED_LOCK_STAT, the table below showed the number
of lock acquisitions in unfair mode and queue mode after a kernel
build with various number of vCPUs.

  vCPUs    queued mode  unfair mode
  -----    -----------  -----------
    30      9,130,518      294,954
    36     10,856,614      386,809
    45      8,467,264   11,475,373
    54      6,409,987   19,670,855
    72      4,782,063   25,712,180

It can be seen that as the VM became more and more over-committed,
the ratio of locks acquired in unfair mode increases. This is all
done automatically to get the best overall performance as possible.

Using a kernel locking microbenchmark with number of locking
threads equals to the number of vCPUs available on the same machine,
the minimum, average and maximum (min/avg/max) numbers of locking
operations done per thread in a 5-second testing interval are shown
below:

  vCPUs         hybrid pvqlock             unfair lock
  -----         --------------             -----------
    36     822,135/881,063/950,363    75,570/313,496/  690,465
    54     542,435/581,664/625,937    35,460/204,280/  457,172
    72     397,500/428,177/499,299    17,933/150,679/  708,001
   108     257,898/288,150/340,871     3,085/181,176/1,257,109

It can be seen that the hybrid pvqspinlocks are more fair and
performant than the unfair locks in this test.

The table below shows the kernel build times on a smaller 2-socket
16-core 32-thread E5-2620 v4 system.

  vCPUs    pvqlock     hybrid pvqlock    unfair lock
  -----    -------     --------------    -----------
    16      436.8s         433.4s          435.6s
    36      366.2s         364.8s          364.5s
    48      423.6s         376.3s          370.2s
    64      433.1s         376.6s          376.8s

Again, the performance of the hybrid pvqspinlock was comparable to
that of the unfair lock.

Signed-off-by: Waiman Long <longman@redhat.com>
Reviewed-by: Juergen Gross <jgross@suse.com>
Reviewed-by: Eduardo Valentin <eduval@amazon.com>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Radim Krčmář <rkrcmar@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/1510089486-3466-1-git-send-email-longman@redhat.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-11-08 10:10:04 +01:00

590 lines
17 KiB
C

/* SPDX-License-Identifier: GPL-2.0 */
#ifndef _GEN_PV_LOCK_SLOWPATH
#error "do not include this file"
#endif
#include <linux/hash.h>
#include <linux/bootmem.h>
#include <linux/debug_locks.h>
/*
* Implement paravirt qspinlocks; the general idea is to halt the vcpus instead
* of spinning them.
*
* This relies on the architecture to provide two paravirt hypercalls:
*
* pv_wait(u8 *ptr, u8 val) -- suspends the vcpu if *ptr == val
* pv_kick(cpu) -- wakes a suspended vcpu
*
* Using these we implement __pv_queued_spin_lock_slowpath() and
* __pv_queued_spin_unlock() to replace native_queued_spin_lock_slowpath() and
* native_queued_spin_unlock().
*/
#define _Q_SLOW_VAL (3U << _Q_LOCKED_OFFSET)
/*
* Queue Node Adaptive Spinning
*
* A queue node vCPU will stop spinning if the vCPU in the previous node is
* not running. The one lock stealing attempt allowed at slowpath entry
* mitigates the slight slowdown for non-overcommitted guest with this
* aggressive wait-early mechanism.
*
* The status of the previous node will be checked at fixed interval
* controlled by PV_PREV_CHECK_MASK. This is to ensure that we won't
* pound on the cacheline of the previous node too heavily.
*/
#define PV_PREV_CHECK_MASK 0xff
/*
* Queue node uses: vcpu_running & vcpu_halted.
* Queue head uses: vcpu_running & vcpu_hashed.
*/
enum vcpu_state {
vcpu_running = 0,
vcpu_halted, /* Used only in pv_wait_node */
vcpu_hashed, /* = pv_hash'ed + vcpu_halted */
};
struct pv_node {
struct mcs_spinlock mcs;
struct mcs_spinlock __res[3];
int cpu;
u8 state;
};
/*
* Include queued spinlock statistics code
*/
#include "qspinlock_stat.h"
/*
* Hybrid PV queued/unfair lock
*
* By replacing the regular queued_spin_trylock() with the function below,
* it will be called once when a lock waiter enter the PV slowpath before
* being queued.
*
* The pending bit is set by the queue head vCPU of the MCS wait queue in
* pv_wait_head_or_lock() to signal that it is ready to spin on the lock.
* When that bit becomes visible to the incoming waiters, no lock stealing
* is allowed. The function will return immediately to make the waiters
* enter the MCS wait queue. So lock starvation shouldn't happen as long
* as the queued mode vCPUs are actively running to set the pending bit
* and hence disabling lock stealing.
*
* When the pending bit isn't set, the lock waiters will stay in the unfair
* mode spinning on the lock unless the MCS wait queue is empty. In this
* case, the lock waiters will enter the queued mode slowpath trying to
* become the queue head and set the pending bit.
*
* This hybrid PV queued/unfair lock combines the best attributes of a
* queued lock (no lock starvation) and an unfair lock (good performance
* on not heavily contended locks).
*/
#define queued_spin_trylock(l) pv_hybrid_queued_unfair_trylock(l)
static inline bool pv_hybrid_queued_unfair_trylock(struct qspinlock *lock)
{
struct __qspinlock *l = (void *)lock;
/*
* Stay in unfair lock mode as long as queued mode waiters are
* present in the MCS wait queue but the pending bit isn't set.
*/
for (;;) {
int val = atomic_read(&lock->val);
if (!(val & _Q_LOCKED_PENDING_MASK) &&
(cmpxchg_acquire(&l->locked, 0, _Q_LOCKED_VAL) == 0)) {
qstat_inc(qstat_pv_lock_stealing, true);
return true;
}
if (!(val & _Q_TAIL_MASK) || (val & _Q_PENDING_MASK))
break;
cpu_relax();
}
return false;
}
/*
* The pending bit is used by the queue head vCPU to indicate that it
* is actively spinning on the lock and no lock stealing is allowed.
*/
#if _Q_PENDING_BITS == 8
static __always_inline void set_pending(struct qspinlock *lock)
{
struct __qspinlock *l = (void *)lock;
WRITE_ONCE(l->pending, 1);
}
static __always_inline void clear_pending(struct qspinlock *lock)
{
struct __qspinlock *l = (void *)lock;
WRITE_ONCE(l->pending, 0);
}
/*
* The pending bit check in pv_queued_spin_steal_lock() isn't a memory
* barrier. Therefore, an atomic cmpxchg_acquire() is used to acquire the
* lock just to be sure that it will get it.
*/
static __always_inline int trylock_clear_pending(struct qspinlock *lock)
{
struct __qspinlock *l = (void *)lock;
return !READ_ONCE(l->locked) &&
(cmpxchg_acquire(&l->locked_pending, _Q_PENDING_VAL,
_Q_LOCKED_VAL) == _Q_PENDING_VAL);
}
#else /* _Q_PENDING_BITS == 8 */
static __always_inline void set_pending(struct qspinlock *lock)
{
atomic_or(_Q_PENDING_VAL, &lock->val);
}
static __always_inline void clear_pending(struct qspinlock *lock)
{
atomic_andnot(_Q_PENDING_VAL, &lock->val);
}
static __always_inline int trylock_clear_pending(struct qspinlock *lock)
{
int val = atomic_read(&lock->val);
for (;;) {
int old, new;
if (val & _Q_LOCKED_MASK)
break;
/*
* Try to clear pending bit & set locked bit
*/
old = val;
new = (val & ~_Q_PENDING_MASK) | _Q_LOCKED_VAL;
val = atomic_cmpxchg_acquire(&lock->val, old, new);
if (val == old)
return 1;
}
return 0;
}
#endif /* _Q_PENDING_BITS == 8 */
/*
* Lock and MCS node addresses hash table for fast lookup
*
* Hashing is done on a per-cacheline basis to minimize the need to access
* more than one cacheline.
*
* Dynamically allocate a hash table big enough to hold at least 4X the
* number of possible cpus in the system. Allocation is done on page
* granularity. So the minimum number of hash buckets should be at least
* 256 (64-bit) or 512 (32-bit) to fully utilize a 4k page.
*
* Since we should not be holding locks from NMI context (very rare indeed) the
* max load factor is 0.75, which is around the point where open addressing
* breaks down.
*
*/
struct pv_hash_entry {
struct qspinlock *lock;
struct pv_node *node;
};
#define PV_HE_PER_LINE (SMP_CACHE_BYTES / sizeof(struct pv_hash_entry))
#define PV_HE_MIN (PAGE_SIZE / sizeof(struct pv_hash_entry))
static struct pv_hash_entry *pv_lock_hash;
static unsigned int pv_lock_hash_bits __read_mostly;
/*
* Allocate memory for the PV qspinlock hash buckets
*
* This function should be called from the paravirt spinlock initialization
* routine.
*/
void __init __pv_init_lock_hash(void)
{
int pv_hash_size = ALIGN(4 * num_possible_cpus(), PV_HE_PER_LINE);
if (pv_hash_size < PV_HE_MIN)
pv_hash_size = PV_HE_MIN;
/*
* Allocate space from bootmem which should be page-size aligned
* and hence cacheline aligned.
*/
pv_lock_hash = alloc_large_system_hash("PV qspinlock",
sizeof(struct pv_hash_entry),
pv_hash_size, 0,
HASH_EARLY | HASH_ZERO,
&pv_lock_hash_bits, NULL,
pv_hash_size, pv_hash_size);
}
#define for_each_hash_entry(he, offset, hash) \
for (hash &= ~(PV_HE_PER_LINE - 1), he = &pv_lock_hash[hash], offset = 0; \
offset < (1 << pv_lock_hash_bits); \
offset++, he = &pv_lock_hash[(hash + offset) & ((1 << pv_lock_hash_bits) - 1)])
static struct qspinlock **pv_hash(struct qspinlock *lock, struct pv_node *node)
{
unsigned long offset, hash = hash_ptr(lock, pv_lock_hash_bits);
struct pv_hash_entry *he;
int hopcnt = 0;
for_each_hash_entry(he, offset, hash) {
hopcnt++;
if (!cmpxchg(&he->lock, NULL, lock)) {
WRITE_ONCE(he->node, node);
qstat_hop(hopcnt);
return &he->lock;
}
}
/*
* Hard assume there is a free entry for us.
*
* This is guaranteed by ensuring every blocked lock only ever consumes
* a single entry, and since we only have 4 nesting levels per CPU
* and allocated 4*nr_possible_cpus(), this must be so.
*
* The single entry is guaranteed by having the lock owner unhash
* before it releases.
*/
BUG();
}
static struct pv_node *pv_unhash(struct qspinlock *lock)
{
unsigned long offset, hash = hash_ptr(lock, pv_lock_hash_bits);
struct pv_hash_entry *he;
struct pv_node *node;
for_each_hash_entry(he, offset, hash) {
if (READ_ONCE(he->lock) == lock) {
node = READ_ONCE(he->node);
WRITE_ONCE(he->lock, NULL);
return node;
}
}
/*
* Hard assume we'll find an entry.
*
* This guarantees a limited lookup time and is itself guaranteed by
* having the lock owner do the unhash -- IFF the unlock sees the
* SLOW flag, there MUST be a hash entry.
*/
BUG();
}
/*
* Return true if when it is time to check the previous node which is not
* in a running state.
*/
static inline bool
pv_wait_early(struct pv_node *prev, int loop)
{
if ((loop & PV_PREV_CHECK_MASK) != 0)
return false;
return READ_ONCE(prev->state) != vcpu_running || vcpu_is_preempted(prev->cpu);
}
/*
* Initialize the PV part of the mcs_spinlock node.
*/
static void pv_init_node(struct mcs_spinlock *node)
{
struct pv_node *pn = (struct pv_node *)node;
BUILD_BUG_ON(sizeof(struct pv_node) > 5*sizeof(struct mcs_spinlock));
pn->cpu = smp_processor_id();
pn->state = vcpu_running;
}
/*
* Wait for node->locked to become true, halt the vcpu after a short spin.
* pv_kick_node() is used to set _Q_SLOW_VAL and fill in hash table on its
* behalf.
*/
static void pv_wait_node(struct mcs_spinlock *node, struct mcs_spinlock *prev)
{
struct pv_node *pn = (struct pv_node *)node;
struct pv_node *pp = (struct pv_node *)prev;
int loop;
bool wait_early;
for (;;) {
for (wait_early = false, loop = SPIN_THRESHOLD; loop; loop--) {
if (READ_ONCE(node->locked))
return;
if (pv_wait_early(pp, loop)) {
wait_early = true;
break;
}
cpu_relax();
}
/*
* Order pn->state vs pn->locked thusly:
*
* [S] pn->state = vcpu_halted [S] next->locked = 1
* MB MB
* [L] pn->locked [RmW] pn->state = vcpu_hashed
*
* Matches the cmpxchg() from pv_kick_node().
*/
smp_store_mb(pn->state, vcpu_halted);
if (!READ_ONCE(node->locked)) {
qstat_inc(qstat_pv_wait_node, true);
qstat_inc(qstat_pv_wait_early, wait_early);
pv_wait(&pn->state, vcpu_halted);
}
/*
* If pv_kick_node() changed us to vcpu_hashed, retain that
* value so that pv_wait_head_or_lock() knows to not also try
* to hash this lock.
*/
cmpxchg(&pn->state, vcpu_halted, vcpu_running);
/*
* If the locked flag is still not set after wakeup, it is a
* spurious wakeup and the vCPU should wait again. However,
* there is a pretty high overhead for CPU halting and kicking.
* So it is better to spin for a while in the hope that the
* MCS lock will be released soon.
*/
qstat_inc(qstat_pv_spurious_wakeup, !READ_ONCE(node->locked));
}
/*
* By now our node->locked should be 1 and our caller will not actually
* spin-wait for it. We do however rely on our caller to do a
* load-acquire for us.
*/
}
/*
* Called after setting next->locked = 1 when we're the lock owner.
*
* Instead of waking the waiters stuck in pv_wait_node() advance their state
* such that they're waiting in pv_wait_head_or_lock(), this avoids a
* wake/sleep cycle.
*/
static void pv_kick_node(struct qspinlock *lock, struct mcs_spinlock *node)
{
struct pv_node *pn = (struct pv_node *)node;
struct __qspinlock *l = (void *)lock;
/*
* If the vCPU is indeed halted, advance its state to match that of
* pv_wait_node(). If OTOH this fails, the vCPU was running and will
* observe its next->locked value and advance itself.
*
* Matches with smp_store_mb() and cmpxchg() in pv_wait_node()
*
* The write to next->locked in arch_mcs_spin_unlock_contended()
* must be ordered before the read of pn->state in the cmpxchg()
* below for the code to work correctly. To guarantee full ordering
* irrespective of the success or failure of the cmpxchg(),
* a relaxed version with explicit barrier is used. The control
* dependency will order the reading of pn->state before any
* subsequent writes.
*/
smp_mb__before_atomic();
if (cmpxchg_relaxed(&pn->state, vcpu_halted, vcpu_hashed)
!= vcpu_halted)
return;
/*
* Put the lock into the hash table and set the _Q_SLOW_VAL.
*
* As this is the same vCPU that will check the _Q_SLOW_VAL value and
* the hash table later on at unlock time, no atomic instruction is
* needed.
*/
WRITE_ONCE(l->locked, _Q_SLOW_VAL);
(void)pv_hash(lock, pn);
}
/*
* Wait for l->locked to become clear and acquire the lock;
* halt the vcpu after a short spin.
* __pv_queued_spin_unlock() will wake us.
*
* The current value of the lock will be returned for additional processing.
*/
static u32
pv_wait_head_or_lock(struct qspinlock *lock, struct mcs_spinlock *node)
{
struct pv_node *pn = (struct pv_node *)node;
struct __qspinlock *l = (void *)lock;
struct qspinlock **lp = NULL;
int waitcnt = 0;
int loop;
/*
* If pv_kick_node() already advanced our state, we don't need to
* insert ourselves into the hash table anymore.
*/
if (READ_ONCE(pn->state) == vcpu_hashed)
lp = (struct qspinlock **)1;
/*
* Tracking # of slowpath locking operations
*/
qstat_inc(qstat_pv_lock_slowpath, true);
for (;; waitcnt++) {
/*
* Set correct vCPU state to be used by queue node wait-early
* mechanism.
*/
WRITE_ONCE(pn->state, vcpu_running);
/*
* Set the pending bit in the active lock spinning loop to
* disable lock stealing before attempting to acquire the lock.
*/
set_pending(lock);
for (loop = SPIN_THRESHOLD; loop; loop--) {
if (trylock_clear_pending(lock))
goto gotlock;
cpu_relax();
}
clear_pending(lock);
if (!lp) { /* ONCE */
lp = pv_hash(lock, pn);
/*
* We must hash before setting _Q_SLOW_VAL, such that
* when we observe _Q_SLOW_VAL in __pv_queued_spin_unlock()
* we'll be sure to be able to observe our hash entry.
*
* [S] <hash> [Rmw] l->locked == _Q_SLOW_VAL
* MB RMB
* [RmW] l->locked = _Q_SLOW_VAL [L] <unhash>
*
* Matches the smp_rmb() in __pv_queued_spin_unlock().
*/
if (xchg(&l->locked, _Q_SLOW_VAL) == 0) {
/*
* The lock was free and now we own the lock.
* Change the lock value back to _Q_LOCKED_VAL
* and unhash the table.
*/
WRITE_ONCE(l->locked, _Q_LOCKED_VAL);
WRITE_ONCE(*lp, NULL);
goto gotlock;
}
}
WRITE_ONCE(pn->state, vcpu_hashed);
qstat_inc(qstat_pv_wait_head, true);
qstat_inc(qstat_pv_wait_again, waitcnt);
pv_wait(&l->locked, _Q_SLOW_VAL);
/*
* Because of lock stealing, the queue head vCPU may not be
* able to acquire the lock before it has to wait again.
*/
}
/*
* The cmpxchg() or xchg() call before coming here provides the
* acquire semantics for locking. The dummy ORing of _Q_LOCKED_VAL
* here is to indicate to the compiler that the value will always
* be nozero to enable better code optimization.
*/
gotlock:
return (u32)(atomic_read(&lock->val) | _Q_LOCKED_VAL);
}
/*
* PV versions of the unlock fastpath and slowpath functions to be used
* instead of queued_spin_unlock().
*/
__visible void
__pv_queued_spin_unlock_slowpath(struct qspinlock *lock, u8 locked)
{
struct __qspinlock *l = (void *)lock;
struct pv_node *node;
if (unlikely(locked != _Q_SLOW_VAL)) {
WARN(!debug_locks_silent,
"pvqspinlock: lock 0x%lx has corrupted value 0x%x!\n",
(unsigned long)lock, atomic_read(&lock->val));
return;
}
/*
* A failed cmpxchg doesn't provide any memory-ordering guarantees,
* so we need a barrier to order the read of the node data in
* pv_unhash *after* we've read the lock being _Q_SLOW_VAL.
*
* Matches the cmpxchg() in pv_wait_head_or_lock() setting _Q_SLOW_VAL.
*/
smp_rmb();
/*
* Since the above failed to release, this must be the SLOW path.
* Therefore start by looking up the blocked node and unhashing it.
*/
node = pv_unhash(lock);
/*
* Now that we have a reference to the (likely) blocked pv_node,
* release the lock.
*/
smp_store_release(&l->locked, 0);
/*
* At this point the memory pointed at by lock can be freed/reused,
* however we can still use the pv_node to kick the CPU.
* The other vCPU may not really be halted, but kicking an active
* vCPU is harmless other than the additional latency in completing
* the unlock.
*/
qstat_inc(qstat_pv_kick_unlock, true);
pv_kick(node->cpu);
}
/*
* Include the architecture specific callee-save thunk of the
* __pv_queued_spin_unlock(). This thunk is put together with
* __pv_queued_spin_unlock() to make the callee-save thunk and the real unlock
* function close to each other sharing consecutive instruction cachelines.
* Alternatively, architecture specific version of __pv_queued_spin_unlock()
* can be defined.
*/
#include <asm/qspinlock_paravirt.h>
#ifndef __pv_queued_spin_unlock
__visible void __pv_queued_spin_unlock(struct qspinlock *lock)
{
struct __qspinlock *l = (void *)lock;
u8 locked;
/*
* We must not unlock if SLOW, because in that case we must first
* unhash. Otherwise it would be possible to have multiple @lock
* entries, which would be BAD.
*/
locked = cmpxchg_release(&l->locked, _Q_LOCKED_VAL, 0);
if (likely(locked == _Q_LOCKED_VAL))
return;
__pv_queued_spin_unlock_slowpath(lock, locked);
}
#endif /* __pv_queued_spin_unlock */