Re: [PATCH v16 08/14] pvqspinlock: Implement simple paravirt support for the qspinlock

[Peter Zijlstra <peterz@infradead.org>]

I changed it to the below; I've not gotten around to compiling or even
running it yet :-(

The biggest change is the pv_hash/pv_unhash functions, which I've
rewritten to hopefully be clearer (and also hopefully not wrecked them).
I took out the cacheline sized structure which takes out that double
loop and simplifies things. I've also added some comments which
hopefully explain how/why we ended up with this exact scheme.

I've also moved the __pv_queue_spin_unlock() function to the tail, such
that we keep the 'wait'/'kick' order for both node and head.

In any case, like I just wrote on the other email, I've stuck some
things in my queue (up to and including patch 11) and if it all works
out we can continue from there.

---
Subject: pvqspinlock: Implement simple paravirt support for the qspinlock
From: Waiman Long <Waiman.Long@hp.com>
Date: Fri, 24 Apr 2015 14:56:37 -0400

Provide a separate (second) version of the spin_lock_slowpath for
paravirt along with a special unlock path.

The second slowpath is generated by adding a few pv hooks to the
normal slowpath, but where those will compile away for the native
case, they expand into special wait/wake code for the pv version.

The actual MCS queue can use extra storage in the mcs_nodes[] array to
keep track of state and therefore uses directed wakeups.

The head contender has no such storage directly visible to the
unlocker.  So the unlocker searches a hash table with open addressing
using a simple binary Galois linear feedback shift register.

Cc: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com>
Cc: David Vrabel <david.vrabel@citrix.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Daniel J Blueman <daniel@numascale.com>
Cc: Paolo Bonzini <paolo.bonzini@gmail.com>
Cc: Scott J Norton <scott.norton@hp.com>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Cc: Douglas Hatch <doug.hatch@hp.com>
Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com>
Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Rik van Riel <riel@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Waiman Long <Waiman.Long@hp.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1429901803-29771-9-git-send-email-Waiman.Long@hp.com
---
 kernel/locking/qspinlock.c          |   68 +++++++
 kernel/locking/qspinlock_paravirt.h |  325 ++++++++++++++++++++++++++++++++++++
 2 files changed, 392 insertions(+), 1 deletion(-)
 create mode 100644 kernel/locking/qspinlock_paravirt.h

--- a/kernel/locking/qspinlock.c
+++ b/kernel/locking/qspinlock.c
@@ -18,6 +18,9 @@
  * Authors: Waiman Long <waiman.long@hp.com>
  *          Peter Zijlstra <peterz@infradead.org>
  */
+
+#ifndef _GEN_PV_LOCK_SLOWPATH
+
 #include <linux/smp.h>
 #include <linux/bug.h>
 #include <linux/cpumask.h>
@@ -65,13 +68,21 @@
 
 #include "mcs_spinlock.h"
 
+#ifdef CONFIG_PARAVIRT_SPINLOCKS
+#define MAX_NODES	8
+#else
+#define MAX_NODES	4
+#endif
+
 /*
  * Per-CPU queue node structures; we can never have more than 4 nested
  * contexts: task, softirq, hardirq, nmi.
  *
  * Exactly fits one 64-byte cacheline on a 64-bit architecture.
+ *
+ * PV doubles the storage and uses the second cacheline for PV state.
  */
-static DEFINE_PER_CPU_ALIGNED(struct mcs_spinlock, mcs_nodes[4]);
+static DEFINE_PER_CPU_ALIGNED(struct mcs_spinlock, mcs_nodes[MAX_NODES]);
 
 /*
  * We must be able to distinguish between no-tail and the tail at 0:0,
@@ -220,6 +231,32 @@ static __always_inline void set_locked(s
 	WRITE_ONCE(l->locked, _Q_LOCKED_VAL);
 }
 
+
+/*
+ * Generate the native code for queue_spin_unlock_slowpath(); provide NOPs for
+ * all the PV callbacks.
+ */
+
+static __always_inline void __pv_init_node(struct mcs_spinlock *node) { }
+static __always_inline void __pv_wait_node(struct mcs_spinlock *node) { }
+static __always_inline void __pv_kick_node(struct mcs_spinlock *node) { }
+
+static __always_inline void __pv_wait_head(struct qspinlock *lock,
+					   struct mcs_spinlock *node) { }
+
+#define pv_enabled()		false
+
+#define pv_init_node		__pv_init_node
+#define pv_wait_node		__pv_wait_node
+#define pv_kick_node		__pv_kick_node
+#define pv_wait_head		__pv_wait_head
+
+#ifdef CONFIG_PARAVIRT_SPINLOCKS
+#define queue_spin_lock_slowpath	native_queue_spin_lock_slowpath
+#endif
+
+#endif /* _GEN_PV_LOCK_SLOWPATH */
+
 /**
  * queue_spin_lock_slowpath - acquire the queue spinlock
  * @lock: Pointer to queue spinlock structure
@@ -249,6 +286,9 @@ void queue_spin_lock_slowpath(struct qsp
 
 	BUILD_BUG_ON(CONFIG_NR_CPUS >= (1U << _Q_TAIL_CPU_BITS));
 
+	if (pv_enabled())
+		goto queue;
+
 	if (virt_queue_spin_lock(lock))
 		return;
 
@@ -325,6 +365,7 @@ void queue_spin_lock_slowpath(struct qsp
 	node += idx;
 	node->locked = 0;
 	node->next = NULL;
+	pv_init_node(node);
 
 	/*
 	 * We touched a (possibly) cold cacheline in the per-cpu queue node;
@@ -350,6 +391,7 @@ void queue_spin_lock_slowpath(struct qsp
 		prev = decode_tail(old);
 		WRITE_ONCE(prev->next, node);
 
+		pv_wait_node(node);
 		arch_mcs_spin_lock_contended(&node->locked);
 	}
 
@@ -365,6 +407,7 @@ void queue_spin_lock_slowpath(struct qsp
 	 * does not imply a full barrier.
 	 *
 	 */
+	pv_wait_head(lock, node);
 	while ((val = smp_load_acquire(&lock->val.counter)) & _Q_LOCKED_PENDING_MASK)
 		cpu_relax();
 
@@ -397,6 +440,7 @@ void queue_spin_lock_slowpath(struct qsp
 		cpu_relax();
 
 	arch_mcs_spin_unlock_contended(&next->locked);
+	pv_kick_node(next);
 
 release:
 	/*
@@ -405,3 +449,25 @@ void queue_spin_lock_slowpath(struct qsp
 	this_cpu_dec(mcs_nodes[0].count);
 }
 EXPORT_SYMBOL(queue_spin_lock_slowpath);
+
+/*
+ * Generate the paravirt code for queue_spin_unlock_slowpath().
+ */
+#if !defined(_GEN_PV_LOCK_SLOWPATH) && defined(CONFIG_PARAVIRT_SPINLOCKS)
+#define _GEN_PV_LOCK_SLOWPATH
+
+#undef  pv_enabled
+#define pv_enabled()	true
+
+#undef pv_init_node
+#undef pv_wait_node
+#undef pv_kick_node
+#undef pv_wait_head
+
+#undef  queue_spin_lock_slowpath
+#define queue_spin_lock_slowpath	__pv_queue_spin_lock_slowpath
+
+#include "qspinlock_paravirt.h"
+#include "qspinlock.c"
+
+#endif
--- /dev/null
+++ b/kernel/locking/qspinlock_paravirt.h
@@ -0,0 +1,325 @@
+#ifndef _GEN_PV_LOCK_SLOWPATH
+#error "do not include this file"
+#endif
+
+#include <linux/hash.h>
+#include <linux/bootmem.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_queue_spin_lock_slowpath() and
+ * __pv_queue_spin_unlock() to replace native_queue_spin_lock_slowpath() and
+ * native_queue_spin_unlock().
+ */
+
+#define _Q_SLOW_VAL	(3U << _Q_LOCKED_OFFSET)
+
+enum vcpu_state {
+	vcpu_running = 0,
+	vcpu_halted,
+};
+
+struct pv_node {
+	struct mcs_spinlock	mcs;
+	struct mcs_spinlock	__res[3];
+
+	int			cpu;
+	u8			state;
+};
+
+/*
+ * 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.
+ *
+ */
+#define PV_HE_PER_LINE	(SMP_CACHE_BYTES / sizeof(struct pv_hash_entry))
+#define PV_HE_MIN	(PAGE_SIZE / sizeof(struct pv_hash_entry))
+
+struct pv_hash_entry {
+	struct qspinlock *lock;
+	struct pv_node   *node;
+};
+
+static struct pv_hash_bucket *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,
+					       &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;
+
+	for_each_hash_entry(he, offset, hash) {
+		if (!cmpxchg(&he->lock, NULL, lock)) {
+			WRITE_ONCE(he->node, node);
+			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();
+}
+
+/*
+ * 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 wake the vcpu again.
+ */
+static void pv_wait_node(struct mcs_spinlock *node)
+{
+	struct pv_node *pn = (struct pv_node *)node;
+	int loop;
+
+	for (;;) {
+		for (loop = SPIN_THRESHOLD; loop; loop--) {
+			if (READ_ONCE(node->locked))
+				return;
+			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_running
+		 *
+		 * Matches the xchg() from pv_kick_node().
+		 */
+		(void)xchg(&pn->state, vcpu_halted);
+
+		if (!READ_ONCE(node->locked))
+			pv_wait(&pn->state, vcpu_halted);
+
+		/*
+		 * Reset the vCPU state to avoid unncessary CPU kicking
+		 */
+		WRITE_ONCE(pn->state, 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.
+		 */
+	}
+	/*
+	 * 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, used to wake those stuck in
+ * pv_wait_node().
+ */
+static void pv_kick_node(struct mcs_spinlock *node)
+{
+	struct pv_node *pn = (struct pv_node *)node;
+
+	/*
+	 * Note that because node->locked is already set, this actual
+	 * mcs_spinlock entry could be re-used already.
+	 *
+	 * This should be fine however, kicking people for no reason is
+	 * harmless.
+	 *
+	 * See the comment in pv_wait_node().
+	 */
+	if (xchg(&pn->state, vcpu_running) == vcpu_halted)
+		pv_kick(pn->cpu);
+}
+
+/*
+ * Wait for l->locked to become clear; halt the vcpu after a short spin.
+ * __pv_queue_spin_unlock() will wake us.
+ */
+static void pv_wait_head(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 loop;
+
+	for (;;) {
+		for (loop = SPIN_THRESHOLD; loop; loop--) {
+			if (!READ_ONCE(l->locked))
+				return;
+			cpu_relax();
+		}
+
+		WRITE_ONCE(pn->state, vcpu_halted);
+		if (!lp) { /* ONCE */
+			lp = pv_hash(lock, pn);
+			/*
+			 * lp must be set before setting _Q_SLOW_VAL
+			 *
+			 * [S] lp = lock                [RmW] l = l->locked = 0
+			 *     MB                             MB
+			 * [S] l->locked = _Q_SLOW_VAL  [L]   lp
+			 *
+			 * Matches the cmpxchg() in __pv_queue_spin_unlock().
+			 */
+			if (!cmpxchg(&l->locked, _Q_LOCKED_VAL, _Q_SLOW_VAL)) {
+				/*
+				 * The lock is free and _Q_SLOW_VAL has never
+				 * been set. Therefore we need to unhash before
+				 * getting the lock.
+				 */
+				WRITE_ONCE(*lp, NULL);
+				return;
+			}
+		}
+		pv_wait(&l->locked, _Q_SLOW_VAL);
+
+		/*
+		 * The unlocker should have freed the lock before kicking the
+		 * CPU. So if the lock is still not free, it is a spurious
+		 * wakeup and so the vCPU should wait again after spinning for
+		 * a while.
+		 */
+	}
+
+	/*
+	 * Lock is unlocked now; the caller will acquire it without waiting.
+	 * As with pv_wait_node() we rely on the caller to do a load-acquire
+	 * for us.
+	 */
+}
+
+/*
+ * PV version of the unlock function to be used in stead of
+ * queue_spin_unlock().
+ */
+__visible void __pv_queue_spin_unlock(struct qspinlock *lock)
+{
+	struct __qspinlock *l = (void *)lock;
+	struct pv_node *node;
+
+	/*
+	 * 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.
+	 */
+	if (likely(cmpxchg(&l->locked, _Q_LOCKED_VAL, 0) == _Q_LOCKED_VAL))
+		return;
+
+	/*
+	 * 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.
+	 */
+	if (READ_ONCE(node->state) == vcpu_halted)
+		pv_kick(node->cpu);
+}
+/*
+ * Include the architecture specific callee-save thunk of the
+ * __pv_queue_spin_unlock(). This thunk is put together with
+ * __pv_queue_spin_unlock() near the top of the file to make sure
+ * that the callee-save thunk and the real unlock function are close
+ * to each other sharing consecutive instruction cachelines.
+ */
+#include <asm/qspinlock_paravirt.h>
+