Index: head/sys/dev/hwpmc/hwpmc_amd.c
===================================================================
--- head/sys/dev/hwpmc/hwpmc_amd.c
+++ head/sys/dev/hwpmc/hwpmc_amd.c
@@ -694,8 +694,10 @@
 			wrmsr(evsel, config);
 	}
 
-	atomic_add_int(retval ? &pmc_stats.pm_intr_processed :
-	    &pmc_stats.pm_intr_ignored, 1);
+	if (retval)
+		counter_u64_add(pmc_stats.pm_intr_processed, 1);
+	else
+		counter_u64_add(pmc_stats.pm_intr_ignored, 1);
 
 	PMCDBG1(MDP,INT,2, "retval=%d", retval);
 	return (retval);
Index: head/sys/dev/hwpmc/hwpmc_core.c
===================================================================
--- head/sys/dev/hwpmc/hwpmc_core.c
+++ head/sys/dev/hwpmc/hwpmc_core.c
@@ -2831,8 +2831,10 @@
 	if (found_interrupt)
 		lapic_reenable_pmc();
 
-	atomic_add_int(found_interrupt ? &pmc_stats.pm_intr_processed :
-	    &pmc_stats.pm_intr_ignored, 1);
+	if (found_interrupt)
+		counter_u64_add(pmc_stats.pm_intr_processed, 1);
+	else
+		counter_u64_add(pmc_stats.pm_intr_ignored, 1);
 
 	return (found_interrupt);
 }
@@ -2896,6 +2898,7 @@
 
 		error = pmc_process_interrupt(cpu, PMC_HR, pm, tf,
 		    TRAPF_USERMODE(tf));
+
 		if (error)
 			intrenable &= ~flag;
 
@@ -2955,8 +2958,10 @@
 	if (found_interrupt)
 		lapic_reenable_pmc();
 
-	atomic_add_int(found_interrupt ? &pmc_stats.pm_intr_processed :
-	    &pmc_stats.pm_intr_ignored, 1);
+	if (found_interrupt)
+		counter_u64_add(pmc_stats.pm_intr_processed, 1);
+	else
+		counter_u64_add(pmc_stats.pm_intr_ignored, 1);
 
 	return (found_interrupt);
 }
Index: head/sys/dev/hwpmc/hwpmc_e500.c
===================================================================
--- head/sys/dev/hwpmc/hwpmc_e500.c
+++ head/sys/dev/hwpmc/hwpmc_e500.c
@@ -616,8 +616,10 @@
 		e500_write_pmc(cpu, i, pm->pm_sc.pm_reloadcount);
 	}
 
-	atomic_add_int(retval ? &pmc_stats.pm_intr_processed :
-	    &pmc_stats.pm_intr_ignored, 1);
+	if (retval)
+		counter_u64_add(pmc_stats.pm_intr_processed, 1);
+	else
+		counter_u64_add(pmc_stats.pm_intr_ignored, 1);
 
 	/* Re-enable PERF exceptions. */
 	if (retval)
Index: head/sys/dev/hwpmc/hwpmc_intel.c
===================================================================
--- head/sys/dev/hwpmc/hwpmc_intel.c
+++ head/sys/dev/hwpmc/hwpmc_intel.c
@@ -94,6 +94,8 @@
 	model = ((cpu_id & 0xF0000) >> 12) | ((cpu_id & 0xF0) >> 4);
 	stepping = cpu_id & 0xF;
 
+	snprintf(pmc_cpuid, sizeof(pmc_cpuid), "GenuineIntel-%d-%02X",
+			 (cpu_id & 0xF00) >> 8, model);
 	switch (cpu_id & 0xF00) {
 #if	defined(__i386__)
 	case 0x500:		/* Pentium family processors */
Index: head/sys/dev/hwpmc/hwpmc_logging.c
===================================================================
--- head/sys/dev/hwpmc/hwpmc_logging.c
+++ head/sys/dev/hwpmc/hwpmc_logging.c
@@ -3,6 +3,7 @@
  *
  * Copyright (c) 2005-2007 Joseph Koshy
  * Copyright (c) 2007 The FreeBSD Foundation
+ * Copyright (c) 2018 Matthew Macy
  * All rights reserved.
  *
  * Portions of this software were developed by A. Joseph Koshy under
@@ -50,7 +51,9 @@
 #include <sys/pmckern.h>
 #include <sys/pmclog.h>
 #include <sys/proc.h>
+#include <sys/sched.h>
 #include <sys/signalvar.h>
+#include <sys/smp.h>
 #include <sys/syscallsubr.h>
 #include <sys/sysctl.h>
 #include <sys/systm.h>
@@ -79,31 +82,28 @@
  * kern.hwpmc.nbuffer -- number of global log buffers
  */
 
-static int pmc_nlogbuffers = PMC_NLOGBUFFERS;
+static int pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU;
 #if (__FreeBSD_version < 1100000)
-TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "nbuffers", &pmc_nlogbuffers);
+TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "nbuffers", &pmc_nlogbuffers_pcpu);
 #endif
-SYSCTL_INT(_kern_hwpmc, OID_AUTO, nbuffers, CTLFLAG_RDTUN,
-    &pmc_nlogbuffers, 0, "number of global log buffers");
+SYSCTL_INT(_kern_hwpmc, OID_AUTO, nbuffers_pcpu, CTLFLAG_RDTUN,
+    &pmc_nlogbuffers_pcpu, 0, "number of log buffers per cpu");
 
 /*
  * Global log buffer list and associated spin lock.
  */
 
-TAILQ_HEAD(, pmclog_buffer) pmc_bufferlist =
-	TAILQ_HEAD_INITIALIZER(pmc_bufferlist);
-static struct mtx pmc_bufferlist_mtx;	/* spin lock */
 static struct mtx pmc_kthread_mtx;	/* sleep lock */
 
-#define	PMCLOG_INIT_BUFFER_DESCRIPTOR(D) do {				\
-		const int __roundup = roundup(sizeof(*D),		\
-			sizeof(uint32_t));				\
-		(D)->plb_fence = ((char *) (D)) +			\
-			 1024*pmclog_buffer_size;			\
-		(D)->plb_base  = (D)->plb_ptr = ((char *) (D)) +	\
-			__roundup;					\
+#define	PMCLOG_INIT_BUFFER_DESCRIPTOR(D, buf, domain) do {						\
+		(D)->plb_fence = ((char *) (buf)) +	1024*pmclog_buffer_size;			\
+		(D)->plb_base  = (D)->plb_ptr = ((char *) (buf));				\
+		(D)->plb_domain = domain; \
 	} while (0)
 
+#define	PMCLOG_RESET_BUFFER_DESCRIPTOR(D) do {			\
+		(D)->plb_ptr  = (D)->plb_base; \
+	} while (0)
 
 /*
  * Log file record constructors.
@@ -114,15 +114,29 @@
 	 ((L) & 0xFFFF))
 
 /* reserve LEN bytes of space and initialize the entry header */
-#define	_PMCLOG_RESERVE(PO,TYPE,LEN,ACTION) do {			\
+#define	_PMCLOG_RESERVE_SAFE(PO,TYPE,LEN,ACTION) do {			\
 		uint32_t *_le;						\
-		int _len = roundup((LEN), sizeof(uint32_t));		\
+		int _len = roundup((LEN), sizeof(uint32_t));	\
 		if ((_le = pmclog_reserve((PO), _len)) == NULL) {	\
 			ACTION;						\
 		}							\
 		*_le = _PMCLOG_TO_HEADER(TYPE,_len);			\
 		_le += 3	/* skip over timestamp */
 
+/* reserve LEN bytes of space and initialize the entry header */
+#define	_PMCLOG_RESERVE(PO,TYPE,LEN,ACTION) do {			\
+		uint32_t *_le;						\
+		int _len = roundup((LEN), sizeof(uint32_t));		\
+		spinlock_enter();									\
+		if ((_le = pmclog_reserve((PO), _len)) == NULL) {	\
+			spinlock_exit();								\
+			ACTION;											\
+		}												\
+		*_le = _PMCLOG_TO_HEADER(TYPE,_len);			\
+		_le += 3	/* skip over timestamp */
+
+
+#define	PMCLOG_RESERVE_SAFE(P,T,L)		_PMCLOG_RESERVE_SAFE(P,T,L,return)
 #define	PMCLOG_RESERVE(P,T,L)		_PMCLOG_RESERVE(P,T,L,return)
 #define	PMCLOG_RESERVE_WITH_ERROR(P,T,L) _PMCLOG_RESERVE(P,T,L,		\
 	error=ENOMEM;goto error)
@@ -138,11 +152,21 @@
 #define	PMCLOG_EMITSTRING(S,L)	do { bcopy((S), _le, (L)); } while (0)
 #define	PMCLOG_EMITNULLSTRING(L) do { bzero(_le, (L)); } while (0)
 
-#define	PMCLOG_DESPATCH(PO)						\
-		pmclog_release((PO));					\
+#define	PMCLOG_DESPATCH_SAFE(PO)						\
+	    pmclog_release((PO));						\
 	} while (0)
 
+#define	PMCLOG_DESPATCH(PO)							\
+	    pmclog_release((PO));						\
+		spinlock_exit();							\
+	} while (0)
 
+#define	PMCLOG_DESPATCH_SYNC(PO)						\
+	    pmclog_schedule_io((PO));						\
+		spinlock_exit();								\
+		} while (0)
+
+
 /*
  * Assertions about the log file format.
  */
@@ -180,8 +204,19 @@
 	char 		*plb_base;
 	char		*plb_ptr;
 	char 		*plb_fence;
-};
+	uint16_t	 plb_domain;
+} __aligned(CACHE_LINE_SIZE);
 
+struct pmc_domain_buffer_header {
+	struct mtx pdbh_mtx;
+	TAILQ_HEAD(, pmclog_buffer) pdbh_head;
+	struct pmclog_buffer *pdbh_plbs;
+	int pdbh_ncpus;
+} __aligned(CACHE_LINE_SIZE);
+
+struct pmc_domain_buffer_header *pmc_dom_hdrs[MAXMEMDOM];
+
+
 /*
  * Prototypes
  */
@@ -191,12 +226,28 @@
 static void pmclog_release(struct pmc_owner *po);
 static uint32_t *pmclog_reserve(struct pmc_owner *po, int length);
 static void pmclog_schedule_io(struct pmc_owner *po);
+static void pmclog_schedule_all(struct pmc_owner *po);
 static void pmclog_stop_kthread(struct pmc_owner *po);
 
 /*
  * Helper functions
  */
 
+static inline void
+pmc_plb_rele_unlocked(struct pmclog_buffer *plb)
+{
+	TAILQ_INSERT_HEAD(&pmc_dom_hdrs[plb->plb_domain]->pdbh_head, plb, plb_next);
+}
+
+static inline void
+pmc_plb_rele(struct pmclog_buffer *plb)
+{
+	mtx_lock_spin(&pmc_dom_hdrs[plb->plb_domain]->pdbh_mtx);
+	pmc_plb_rele_unlocked(plb);
+	mtx_unlock_spin(&pmc_dom_hdrs[plb->plb_domain]->pdbh_mtx);
+}
+
+
 /*
  * Get a log buffer
  */
@@ -205,16 +256,16 @@
 pmclog_get_buffer(struct pmc_owner *po)
 {
 	struct pmclog_buffer *plb;
+	int domain;
 
-	mtx_assert(&po->po_mtx, MA_OWNED);
-
-	KASSERT(po->po_curbuf == NULL,
+	KASSERT(po->po_curbuf[curcpu] == NULL,
 	    ("[pmclog,%d] po=%p current buffer still valid", __LINE__, po));
 
-	mtx_lock_spin(&pmc_bufferlist_mtx);
-	if ((plb = TAILQ_FIRST(&pmc_bufferlist)) != NULL)
-		TAILQ_REMOVE(&pmc_bufferlist, plb, plb_next);
-	mtx_unlock_spin(&pmc_bufferlist_mtx);
+	domain = PCPU_GET(domain);
+	mtx_lock_spin(&pmc_dom_hdrs[domain]->pdbh_mtx);
+	if ((plb = TAILQ_FIRST(&pmc_dom_hdrs[domain]->pdbh_head)) != NULL)
+		TAILQ_REMOVE(&pmc_dom_hdrs[domain]->pdbh_head, plb, plb_next);
+	mtx_unlock_spin(&pmc_dom_hdrs[domain]->pdbh_mtx);
 
 	PMCDBG2(LOG,GTB,1, "po=%p plb=%p", po, plb);
 
@@ -227,12 +278,12 @@
 		    plb->plb_base, plb->plb_fence));
 #endif
 
-	po->po_curbuf = plb;
+	po->po_curbuf[curcpu] = plb;
 
 	/* update stats */
-	atomic_add_int(&pmc_stats.pm_buffer_requests, 1);
+	counter_u64_add(pmc_stats.pm_buffer_requests, 1);
 	if (plb == NULL)
-		atomic_add_int(&pmc_stats.pm_buffer_requests_failed, 1);
+		counter_u64_add(pmc_stats.pm_buffer_requests_failed, 1);
 
 	return (plb ? 0 : ENOMEM);
 }
@@ -421,12 +472,9 @@
 		mtx_lock(&pmc_kthread_mtx);
 
 		/* put the used buffer back into the global pool */
-		PMCLOG_INIT_BUFFER_DESCRIPTOR(lb);
+		PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
 
-		mtx_lock_spin(&pmc_bufferlist_mtx);
-		TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next);
-		mtx_unlock_spin(&pmc_bufferlist_mtx);
-
+		pmc_plb_rele(lb);
 		lb = NULL;
 	}
 
@@ -437,11 +485,9 @@
 
 	/* return the current I/O buffer to the global pool */
 	if (lb) {
-		PMCLOG_INIT_BUFFER_DESCRIPTOR(lb);
+		PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
 
-		mtx_lock_spin(&pmc_bufferlist_mtx);
-		TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next);
-		mtx_unlock_spin(&pmc_bufferlist_mtx);
+		pmc_plb_rele(lb);
 	}
 
 	/*
@@ -460,19 +506,20 @@
 static void
 pmclog_release(struct pmc_owner *po)
 {
-	KASSERT(po->po_curbuf->plb_ptr >= po->po_curbuf->plb_base,
+	struct pmclog_buffer *plb;
+
+	plb = po->po_curbuf[curcpu];
+	KASSERT(plb->plb_ptr >= plb->plb_base,
 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
-		po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base));
-	KASSERT(po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence,
+		po, plb->plb_ptr, plb->plb_base));
+	KASSERT(plb->plb_ptr <= plb->plb_fence,
 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
-		po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_fence));
+		po, plb->plb_ptr, plb->plb_fence));
 
 	/* schedule an I/O if we've filled a buffer */
-	if (po->po_curbuf->plb_ptr >= po->po_curbuf->plb_fence)
+	if (plb->plb_ptr >= plb->plb_fence)
 		pmclog_schedule_io(po);
 
-	mtx_unlock_spin(&po->po_mtx);
-
 	PMCDBG1(LOG,REL,1, "po=%p", po);
 }
 
@@ -492,36 +539,32 @@
 	uintptr_t newptr, oldptr;
 	uint32_t *lh;
 	struct timespec ts;
+	struct pmclog_buffer *plb, **pplb;
 
 	PMCDBG2(LOG,ALL,1, "po=%p len=%d", po, length);
 
 	KASSERT(length % sizeof(uint32_t) == 0,
 	    ("[pmclog,%d] length not a multiple of word size", __LINE__));
 
-	mtx_lock_spin(&po->po_mtx);
-
 	/* No more data when shutdown in progress. */
-	if (po->po_flags & PMC_PO_SHUTDOWN) {
-		mtx_unlock_spin(&po->po_mtx);
+	if (po->po_flags & PMC_PO_SHUTDOWN)
 		return (NULL);
-	}
 
-	if (po->po_curbuf == NULL)
-		if (pmclog_get_buffer(po) != 0) {
-			mtx_unlock_spin(&po->po_mtx);
-			return (NULL);
-		}
+	pplb = &po->po_curbuf[curcpu];
+	if (*pplb == NULL && pmclog_get_buffer(po) != 0)
+		goto fail;
 
-	KASSERT(po->po_curbuf != NULL,
+	KASSERT(*pplb != NULL,
 	    ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
 
-	KASSERT(po->po_curbuf->plb_ptr >= po->po_curbuf->plb_base &&
-	    po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence,
+	plb = *pplb;
+	KASSERT(plb->plb_ptr >= plb->plb_base &&
+	    plb->plb_ptr <= plb->plb_fence,
 	    ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
-		__LINE__, po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base,
-		po->po_curbuf->plb_fence));
+		__LINE__, po, plb->plb_ptr, plb->plb_base,
+		plb->plb_fence));
 
-	oldptr = (uintptr_t) po->po_curbuf->plb_ptr;
+	oldptr = (uintptr_t) plb->plb_ptr;
 	newptr = oldptr + length;
 
 	KASSERT(oldptr != (uintptr_t) NULL,
@@ -531,8 +574,8 @@
 	 * If we have space in the current buffer, return a pointer to
 	 * available space with the PO structure locked.
 	 */
-	if (newptr <= (uintptr_t) po->po_curbuf->plb_fence) {
-		po->po_curbuf->plb_ptr = (char *) newptr;
+	if (newptr <= (uintptr_t) plb->plb_fence) {
+		plb->plb_ptr = (char *) newptr;
 		goto done;
 	}
 
@@ -542,24 +585,23 @@
 	 */
 	pmclog_schedule_io(po);
 
-	if (pmclog_get_buffer(po) != 0) {
-		mtx_unlock_spin(&po->po_mtx);
-		return (NULL);
-	}
+	if (pmclog_get_buffer(po) != 0)
+		goto fail;
 
-	KASSERT(po->po_curbuf != NULL,
+	plb = *pplb;
+	KASSERT(plb != NULL,
 	    ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
 
-	KASSERT(po->po_curbuf->plb_ptr != NULL,
+	KASSERT(plb->plb_ptr != NULL,
 	    ("[pmclog,%d] null return from pmc_get_log_buffer", __LINE__));
 
-	KASSERT(po->po_curbuf->plb_ptr == po->po_curbuf->plb_base &&
-	    po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence,
+	KASSERT(plb->plb_ptr == plb->plb_base &&
+	    plb->plb_ptr <= plb->plb_fence,
 	    ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
-		__LINE__, po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base,
-		po->po_curbuf->plb_fence));
+		__LINE__, po, plb->plb_ptr, plb->plb_base,
+		plb->plb_fence));
 
-	oldptr = (uintptr_t) po->po_curbuf->plb_ptr;
+	oldptr = (uintptr_t) plb->plb_ptr;
 
  done:
 	lh = (uint32_t *) oldptr;
@@ -568,6 +610,8 @@
 	*lh++ = ts.tv_sec & 0xFFFFFFFF;
 	*lh++ = ts.tv_nsec & 0xFFFFFFF;
 	return ((uint32_t *) oldptr);
+ fail:
+	return (NULL);
 }
 
 /*
@@ -579,26 +623,28 @@
 static void
 pmclog_schedule_io(struct pmc_owner *po)
 {
-	KASSERT(po->po_curbuf != NULL,
-	    ("[pmclog,%d] schedule_io with null buffer po=%p", __LINE__, po));
+	struct pmclog_buffer *plb;
 
-	KASSERT(po->po_curbuf->plb_ptr >= po->po_curbuf->plb_base,
+	plb = po->po_curbuf[curcpu];
+	po->po_curbuf[curcpu] = NULL;
+	KASSERT(plb != NULL,
+	    ("[pmclog,%d] schedule_io with null buffer po=%p", __LINE__, po));
+	KASSERT(plb->plb_ptr >= plb->plb_base,
 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
-		po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base));
-	KASSERT(po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence,
+		po, plb->plb_ptr, plb->plb_base));
+	KASSERT(plb->plb_ptr <= plb->plb_fence,
 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
-		po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_fence));
+		po, plb->plb_ptr, plb->plb_fence));
 
 	PMCDBG1(LOG,SIO, 1, "po=%p", po);
 
-	mtx_assert(&po->po_mtx, MA_OWNED);
-
 	/*
 	 * Add the current buffer to the tail of the buffer list and
 	 * wakeup the helper.
 	 */
-	TAILQ_INSERT_TAIL(&po->po_logbuffers, po->po_curbuf, plb_next);
-	po->po_curbuf = NULL;
+	mtx_lock_spin(&po->po_mtx);
+	TAILQ_INSERT_TAIL(&po->po_logbuffers, plb, plb_next);
+	mtx_unlock_spin(&po->po_mtx);
 	wakeup_one(po);
 }
 
@@ -671,7 +717,7 @@
 	    sizeof(struct pmclog_initialize));
 	PMCLOG_EMIT32(PMC_VERSION);
 	PMCLOG_EMIT32(md->pmd_cputype);
-	PMCLOG_DESPATCH(po);
+	PMCLOG_DESPATCH_SYNC(po);
 
 	return (0);
 
@@ -719,19 +765,22 @@
 	/* return all queued log buffers to the global pool */
 	while ((lb = TAILQ_FIRST(&po->po_logbuffers)) != NULL) {
 		TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
-		PMCLOG_INIT_BUFFER_DESCRIPTOR(lb);
-		mtx_lock_spin(&pmc_bufferlist_mtx);
-		TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next);
-		mtx_unlock_spin(&pmc_bufferlist_mtx);
+		PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
+		pmc_plb_rele(lb);
 	}
-
-	/* return the 'current' buffer to the global pool */
-	if ((lb = po->po_curbuf) != NULL) {
-		PMCLOG_INIT_BUFFER_DESCRIPTOR(lb);
-		mtx_lock_spin(&pmc_bufferlist_mtx);
-		TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next);
-		mtx_unlock_spin(&pmc_bufferlist_mtx);
+	for (int i = 0; i < mp_ncpus; i++) {
+		thread_lock(curthread);
+		sched_bind(curthread, i);
+		thread_unlock(curthread);
+		/* return the 'current' buffer to the global pool */
+		if ((lb = po->po_curbuf[curcpu]) != NULL) {
+			PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
+			pmc_plb_rele(lb);
+		}
 	}
+	thread_lock(curthread);
+	sched_unbind(curthread);
+	thread_unlock(curthread);
 
 	/* drop a reference to the fd */
 	if (po->po_file != NULL) {
@@ -752,7 +801,6 @@
 pmclog_flush(struct pmc_owner *po)
 {
 	int error;
-	struct pmclog_buffer *lb;
 
 	PMCDBG1(LOG,FLS,1, "po=%p", po);
 
@@ -774,23 +822,45 @@
 		goto error;
 	}
 
-	/*
-	 * Schedule the current buffer if any and not empty.
-	 */
-	mtx_lock_spin(&po->po_mtx);
-	lb = po->po_curbuf;
-	if (lb && lb->plb_ptr != lb->plb_base) {
-		pmclog_schedule_io(po);
-	} else
-		error = ENOBUFS;
-	mtx_unlock_spin(&po->po_mtx);
-
+	pmclog_schedule_all(po);
  error:
 	mtx_unlock(&pmc_kthread_mtx);
 
 	return (error);
 }
 
+static void
+pmclog_schedule_one_cond(void *arg)
+{
+	struct pmc_owner *po = arg;
+	struct pmclog_buffer *plb;
+
+	spinlock_enter();
+	/* tell hardclock not to run again */
+	DPCPU_SET(pmc_sampled, 0);
+	plb = po->po_curbuf[curcpu];
+	if (plb && plb->plb_ptr != plb->plb_base)
+		pmclog_schedule_io(po);
+	spinlock_exit();
+}
+
+static void
+pmclog_schedule_all(struct pmc_owner *po)
+{
+	/*
+	 * Schedule the current buffer if any and not empty.
+	 */
+	for (int i = 0; i < mp_ncpus; i++) {
+		thread_lock(curthread);
+		sched_bind(curthread, i);
+		thread_unlock(curthread);
+		pmclog_schedule_one_cond(po);
+	}
+	thread_lock(curthread);
+	sched_unbind(curthread);
+	thread_unlock(curthread);
+}
+
 int
 pmclog_close(struct pmc_owner *po)
 {
@@ -804,19 +874,14 @@
 	/*
 	 * Schedule the current buffer.
 	 */
-	mtx_lock_spin(&po->po_mtx);
-	if (po->po_curbuf)
-		pmclog_schedule_io(po);
-	else
-		wakeup_one(po);
-	mtx_unlock_spin(&po->po_mtx);
+	pmclog_schedule_all(po);
+	wakeup_one(po);
 
 	/*
 	 * Initiate shutdown: no new data queued,
 	 * thread will close file on last block.
 	 */
 	po->po_flags |= PMC_PO_SHUTDOWN;
-
 	mtx_unlock(&pmc_kthread_mtx);
 
 	return (0);
@@ -836,20 +901,20 @@
 	    ps->ps_nsamples * sizeof(uintfptr_t);
 	po = pm->pm_owner;
 	flags = PMC_CALLCHAIN_TO_CPUFLAGS(ps->ps_cpu,ps->ps_flags);
-	PMCLOG_RESERVE(po, CALLCHAIN, recordlen);
+	PMCLOG_RESERVE_SAFE(po, CALLCHAIN, recordlen);
 	PMCLOG_EMIT32(ps->ps_pid);
 	PMCLOG_EMIT32(pm->pm_id);
 	PMCLOG_EMIT32(flags);
 	for (n = 0; n < ps->ps_nsamples; n++)
 		PMCLOG_EMITADDR(ps->ps_pc[n]);
-	PMCLOG_DESPATCH(po);
+	PMCLOG_DESPATCH_SAFE(po);
 }
 
 void
 pmclog_process_closelog(struct pmc_owner *po)
 {
 	PMCLOG_RESERVE(po,CLOSELOG,sizeof(struct pmclog_closelog));
-	PMCLOG_DESPATCH(po);
+	PMCLOG_DESPATCH_SYNC(po);
 }
 
 void
@@ -913,14 +978,14 @@
 		else
 			PMCLOG_EMITNULLSTRING(PMC_NAME_MAX);
 		pmc_soft_ev_release(ps);
-		PMCLOG_DESPATCH(po);
+		PMCLOG_DESPATCH_SYNC(po);
 	} else {
 		PMCLOG_RESERVE(po, PMCALLOCATE,
 		    sizeof(struct pmclog_pmcallocate));
 		PMCLOG_EMIT32(pm->pm_id);
 		PMCLOG_EMIT32(pm->pm_event);
 		PMCLOG_EMIT32(pm->pm_flags);
-		PMCLOG_DESPATCH(po);
+		PMCLOG_DESPATCH_SYNC(po);
 	}
 }
 
@@ -941,7 +1006,7 @@
 	PMCLOG_EMIT32(pm->pm_id);
 	PMCLOG_EMIT32(pid);
 	PMCLOG_EMITSTRING(path, pathlen);
-	PMCLOG_DESPATCH(po);
+	PMCLOG_DESPATCH_SYNC(po);
 }
 
 void
@@ -956,7 +1021,7 @@
 	PMCLOG_RESERVE(po, PMCDETACH, sizeof(struct pmclog_pmcdetach));
 	PMCLOG_EMIT32(pm->pm_id);
 	PMCLOG_EMIT32(pid);
-	PMCLOG_DESPATCH(po);
+	PMCLOG_DESPATCH_SYNC(po);
 }
 
 /*
@@ -1081,30 +1146,57 @@
 void
 pmclog_initialize()
 {
-	int n;
+	int domain, cpu;
+	struct pcpu *pc;
 	struct pmclog_buffer *plb;
 
-	if (pmclog_buffer_size <= 0) {
+	if (pmclog_buffer_size <= 0 || pmclog_buffer_size > 16*1024) {
 		(void) printf("hwpmc: tunable logbuffersize=%d must be "
-		    "greater than zero.\n", pmclog_buffer_size);
+					  "greater than zero and less than or equal to 16MB.\n",
+					  pmclog_buffer_size);
 		pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
 	}
 
-	if (pmc_nlogbuffers <= 0) {
+	if (pmc_nlogbuffers_pcpu <= 0) {
 		(void) printf("hwpmc: tunable nlogbuffers=%d must be greater "
-		    "than zero.\n", pmc_nlogbuffers);
-		pmc_nlogbuffers = PMC_NLOGBUFFERS;
+					  "than zero.\n", pmc_nlogbuffers_pcpu);
+		pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU;
 	}
 
-	/* create global pool of log buffers */
-	for (n = 0; n < pmc_nlogbuffers; n++) {
-		plb = malloc(1024 * pmclog_buffer_size, M_PMC,
-		    M_WAITOK|M_ZERO);
-		PMCLOG_INIT_BUFFER_DESCRIPTOR(plb);
-		TAILQ_INSERT_HEAD(&pmc_bufferlist, plb, plb_next);
+	if (pmc_nlogbuffers_pcpu*pmclog_buffer_size > 32*1024) {
+		(void) printf("hwpmc: memory allocated pcpu must be less than 32MB (is %dK).\n",
+					  pmc_nlogbuffers_pcpu*pmclog_buffer_size);
+		pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU;
+		pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
 	}
-	mtx_init(&pmc_bufferlist_mtx, "pmc-buffer-list", "pmc-leaf",
-	    MTX_SPIN);
+	for (domain = 0; domain < vm_ndomains; domain++) {
+		pmc_dom_hdrs[domain] = malloc_domain(sizeof(struct pmc_domain_buffer_header), M_PMC, domain,
+										M_WAITOK|M_ZERO);
+		mtx_init(&pmc_dom_hdrs[domain]->pdbh_mtx, "pmc_bufferlist_mtx", "pmc-leaf", MTX_SPIN);
+		TAILQ_INIT(&pmc_dom_hdrs[domain]->pdbh_head);
+	}
+	CPU_FOREACH(cpu) {
+		if (CPU_ABSENT(cpu))
+			continue;
+		pc = pcpu_find(cpu);
+		domain = pc->pc_domain;
+		pmc_dom_hdrs[domain]->pdbh_ncpus++;
+	}
+	for (domain = 0; domain < vm_ndomains; domain++) {
+		int ncpus = pmc_dom_hdrs[domain]->pdbh_ncpus;
+		int total = ncpus*pmc_nlogbuffers_pcpu;
+
+		plb = malloc_domain(sizeof(struct pmclog_buffer)*total, M_PMC, domain, M_WAITOK|M_ZERO);
+		pmc_dom_hdrs[domain]->pdbh_plbs = plb;
+		for (int i = 0; i < total; i++, plb++) {
+			void *buf;
+
+			buf = malloc_domain(1024 * pmclog_buffer_size, M_PMC, domain,
+								M_WAITOK|M_ZERO);
+			PMCLOG_INIT_BUFFER_DESCRIPTOR(plb, buf, domain);
+			pmc_plb_rele_unlocked(plb);
+		}
+	}
 	mtx_init(&pmc_kthread_mtx, "pmc-kthread", "pmc-sleep", MTX_DEF);
 }
 
@@ -1118,12 +1210,17 @@
 pmclog_shutdown()
 {
 	struct pmclog_buffer *plb;
+	int domain;
 
 	mtx_destroy(&pmc_kthread_mtx);
-	mtx_destroy(&pmc_bufferlist_mtx);
 
-	while ((plb = TAILQ_FIRST(&pmc_bufferlist)) != NULL) {
-		TAILQ_REMOVE(&pmc_bufferlist, plb, plb_next);
-		free(plb, M_PMC);
+	for (domain = 0; domain < vm_ndomains; domain++) {
+		mtx_destroy(&pmc_dom_hdrs[domain]->pdbh_mtx);
+		while ((plb = TAILQ_FIRST(&pmc_dom_hdrs[domain]->pdbh_head)) != NULL) {
+			TAILQ_REMOVE(&pmc_dom_hdrs[domain]->pdbh_head, plb, plb_next);
+			free(plb->plb_base, M_PMC);
+		}
+		free(pmc_dom_hdrs[domain]->pdbh_plbs, M_PMC);
+		free(pmc_dom_hdrs[domain], M_PMC);
 	}
 }
Index: head/sys/dev/hwpmc/hwpmc_mod.c
===================================================================
--- head/sys/dev/hwpmc/hwpmc_mod.c
+++ head/sys/dev/hwpmc/hwpmc_mod.c
@@ -3,6 +3,7 @@
  *
  * Copyright (c) 2003-2008 Joseph Koshy
  * Copyright (c) 2007 The FreeBSD Foundation
+ * Copyright (c) 2018 Matthew Macy
  * All rights reserved.
  *
  * Portions of this software were developed by A. Joseph Koshy under
@@ -138,8 +139,9 @@
     pmc_kld_unload_tag;
 
 /* Module statistics */
-struct pmc_op_getdriverstats pmc_stats;
+struct pmc_driverstats pmc_stats;
 
+
 /* Machine/processor dependent operations */
 static struct pmc_mdep  *md;
 
@@ -235,11 +237,34 @@
  */
 
 SYSCTL_DECL(_kern_hwpmc);
+SYSCTL_NODE(_kern_hwpmc, OID_AUTO, stats, CTLFLAG_RW, 0, "HWPMC stats");
 
+
+/* Stats. */
+SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_ignored, CTLFLAG_RW,
+				   &pmc_stats.pm_intr_ignored, "# of interrupts ignored");
+SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_processed, CTLFLAG_RW,
+				   &pmc_stats.pm_intr_processed, "# of interrupts processed");
+SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_bufferfull, CTLFLAG_RW,
+				   &pmc_stats.pm_intr_bufferfull, "# of interrupts where buffer was full");
+SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscalls, CTLFLAG_RW,
+				   &pmc_stats.pm_syscalls, "# of syscalls");
+SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscall_errors, CTLFLAG_RW,
+				   &pmc_stats.pm_syscall_errors, "# of syscall_errors");
+SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests, CTLFLAG_RW,
+				   &pmc_stats.pm_buffer_requests, "# of buffer requests");
+SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests_failed, CTLFLAG_RW,
+				   &pmc_stats.pm_buffer_requests_failed, "# of buffer requests which failed");
+SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, log_sweeps, CTLFLAG_RW,
+				   &pmc_stats.pm_log_sweeps, "# of ?");
+
 static int pmc_callchaindepth = PMC_CALLCHAIN_DEPTH;
 SYSCTL_INT(_kern_hwpmc, OID_AUTO, callchaindepth, CTLFLAG_RDTUN,
     &pmc_callchaindepth, 0, "depth of call chain records");
 
+char pmc_cpuid[64];
+SYSCTL_STRING(_kern_hwpmc, OID_AUTO, cpuid, CTLFLAG_RD,
+	pmc_cpuid, 0, "cpu version string");
 #ifdef	HWPMC_DEBUG
 struct pmc_debugflags pmc_debugflags = PMC_DEBUG_DEFAULT_FLAGS;
 char	pmc_debugstr[PMC_DEBUG_STRSIZE];
@@ -250,6 +275,7 @@
     0, 0, pmc_debugflags_sysctl_handler, "A", "debug flags");
 #endif
 
+
 /*
  * kern.hwpmc.hashrows -- determines the number of rows in the
  * of the hash table used to look up threads
@@ -1260,7 +1286,7 @@
 			continue;
 
 		/* increment PMC runcount */
-		atomic_add_rel_int(&pm->pm_runcount, 1);
+		counter_u64_add(pm->pm_runcount, 1);
 
 		/* configure the HWPMC we are going to use. */
 		pcd = pmc_ri_to_classdep(md, ri, &adjri);
@@ -1311,10 +1337,10 @@
 
 		/* If a sampling mode PMC, reset stalled state. */
 		if (PMC_TO_MODE(pm) == PMC_MODE_TS)
-			CPU_CLR_ATOMIC(cpu, &pm->pm_stalled);
+			pm->pm_pcpu_state[cpu].pps_stalled = 0;
 
 		/* Indicate that we desire this to run. */
-		CPU_SET_ATOMIC(cpu, &pm->pm_cpustate);
+		pm->pm_pcpu_state[cpu].pps_cpustate = 1;
 
 		/* Start the PMC. */
 		pcd->pcd_start_pmc(cpu, adjri);
@@ -1417,12 +1443,12 @@
 		 * an interrupt re-enables the PMC after this code has
 		 * already checked the pm_stalled flag.
 		 */
-		CPU_CLR_ATOMIC(cpu, &pm->pm_cpustate);
-		if (!CPU_ISSET(cpu, &pm->pm_stalled))
+		pm->pm_pcpu_state[cpu].pps_cpustate = 0;
+		if (pm->pm_pcpu_state[cpu].pps_stalled == 0)
 			pcd->pcd_stop_pmc(cpu, adjri);
 
 		/* reduce this PMC's runcount */
-		atomic_subtract_rel_int(&pm->pm_runcount, 1);
+		counter_u64_add(pm->pm_runcount, -1);
 
 		/*
 		 * If this PMC is associated with this process,
@@ -1537,7 +1563,7 @@
 	/* Inform owners of all system-wide sampling PMCs. */
 	LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
-		pmclog_process_map_in(po, pid, pkm->pm_address, fullpath);
+			pmclog_process_map_in(po, pid, pkm->pm_address, fullpath);
 
 	if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
 		goto done;
@@ -1993,7 +2019,7 @@
 		 * had already processed the interrupt).  We don't
 		 * lose the interrupt sample.
 		 */
-		CPU_CLR_ATOMIC(PCPU_GET(cpuid), &pmc_cpumask);
+		DPCPU_SET(pmc_sampled, 0);
 		pmc_process_samples(PCPU_GET(cpuid), PMC_HR);
 		pmc_process_samples(PCPU_GET(cpuid), PMC_SR);
 		break;
@@ -2191,7 +2217,8 @@
 	struct pmc *pmc;
 
 	pmc = malloc(sizeof(struct pmc), M_PMC, M_WAITOK|M_ZERO);
-
+	pmc->pm_runcount = counter_u64_alloc(M_WAITOK);
+	pmc->pm_pcpu_state = malloc(sizeof(struct pmc_pcpu_state)*mp_ncpus, M_PMC, M_WAITOK|M_ZERO);
 	PMCDBG1(PMC,ALL,1, "allocate-pmc -> pmc=%p", pmc);
 
 	return pmc;
@@ -2212,10 +2239,12 @@
 	    ("[pmc,%d] destroying pmc with targets", __LINE__));
 	KASSERT(pm->pm_owner == NULL,
 	    ("[pmc,%d] destroying pmc attached to an owner", __LINE__));
-	KASSERT(pm->pm_runcount == 0,
-	    ("[pmc,%d] pmc has non-zero run count %d", __LINE__,
-		pm->pm_runcount));
+	KASSERT(counter_u64_fetch(pm->pm_runcount) == 0,
+	    ("[pmc,%d] pmc has non-zero run count %ld", __LINE__,
+		 (unsigned long)counter_u64_fetch(pm->pm_runcount)));
 
+	counter_u64_free(pm->pm_runcount);
+	free(pm->pm_pcpu_state, M_PMC);
 	free(pm, M_PMC);
 }
 
@@ -2231,13 +2260,13 @@
 	 * Loop (with a forced context switch) till the PMC's runcount
 	 * comes down to zero.
 	 */
-	while (atomic_load_acq_32(&pm->pm_runcount) > 0) {
+	while (counter_u64_fetch(pm->pm_runcount) > 0) {
 #ifdef HWPMC_DEBUG
 		maxloop--;
 		KASSERT(maxloop > 0,
-		    ("[pmc,%d] (ri%d, rc%d) waiting too long for "
+		    ("[pmc,%d] (ri%d, rc%ld) waiting too long for "
 			"pmc to be free", __LINE__,
-			PMC_TO_ROWINDEX(pm), pm->pm_runcount));
+			 PMC_TO_ROWINDEX(pm), (unsigned long)counter_u64_fetch(pm->pm_runcount)));
 #endif
 		pmc_force_context_switch();
 	}
@@ -2295,9 +2324,9 @@
 		pmc_select_cpu(cpu);
 
 		/* switch off non-stalled CPUs */
-		CPU_CLR_ATOMIC(cpu, &pm->pm_cpustate);
+		pm->pm_pcpu_state[cpu].pps_cpustate = 0;
 		if (pm->pm_state == PMC_STATE_RUNNING &&
-		    !CPU_ISSET(cpu, &pm->pm_stalled)) {
+			pm->pm_pcpu_state[cpu].pps_stalled == 0) {
 
 			phw = pmc_pcpu[cpu]->pc_hwpmcs[ri];
 
@@ -2735,10 +2764,10 @@
 		 pm->pm_sc.pm_initial)) == 0) {
 		/* If a sampling mode PMC, reset stalled state. */
 		if (PMC_IS_SAMPLING_MODE(mode))
-			CPU_CLR_ATOMIC(cpu, &pm->pm_stalled);
+			pm->pm_pcpu_state[cpu].pps_stalled = 0;
 
 		/* Indicate that we desire this to run. Start it. */
-		CPU_SET_ATOMIC(cpu, &pm->pm_cpustate);
+		pm->pm_pcpu_state[cpu].pps_cpustate = 1;
 		error = pcd->pcd_start_pmc(cpu, adjri);
 	}
 	critical_exit();
@@ -2802,7 +2831,7 @@
 	ri = PMC_TO_ROWINDEX(pm);
 	pcd = pmc_ri_to_classdep(md, ri, &adjri);
 
-	CPU_CLR_ATOMIC(cpu, &pm->pm_cpustate);
+	pm->pm_pcpu_state[cpu].pps_cpustate = 0;
 	critical_enter();
 	if ((error = pcd->pcd_stop_pmc(cpu, adjri)) == 0)
 		error = pcd->pcd_read_pmc(cpu, adjri, &pm->pm_sc.pm_initial);
@@ -2884,7 +2913,7 @@
 	    pmc_op_to_name[op], arg);
 
 	error = 0;
-	atomic_add_int(&pmc_stats.pm_syscalls, 1);
+	counter_u64_add(pmc_stats.pm_syscalls, 1);
 
 	switch (op) {
 
@@ -3063,8 +3092,16 @@
 	case PMC_OP_GETDRIVERSTATS:
 	{
 		struct pmc_op_getdriverstats gms;
-
-		bcopy(&pmc_stats, &gms, sizeof(gms));
+#define CFETCH(a, b, field) a.field = counter_u64_fetch(b.field)
+		CFETCH(gms, pmc_stats, pm_intr_ignored);
+		CFETCH(gms, pmc_stats, pm_intr_processed);
+		CFETCH(gms, pmc_stats, pm_intr_bufferfull);
+		CFETCH(gms, pmc_stats, pm_syscalls);
+		CFETCH(gms, pmc_stats, pm_syscall_errors);
+		CFETCH(gms, pmc_stats, pm_buffer_requests);
+		CFETCH(gms, pmc_stats, pm_buffer_requests_failed);
+		CFETCH(gms, pmc_stats, pm_log_sweeps);
+#undef CFETCH
 		error = copyout(&gms, arg, sizeof(gms));
 	}
 	break;
@@ -4040,7 +4077,7 @@
 		sx_xunlock(&pmc_sx);
 done_syscall:
 	if (error)
-		atomic_add_int(&pmc_stats.pm_syscall_errors, 1);
+		counter_u64_add(pmc_stats.pm_syscall_errors, 1);
 
 	return (error);
 }
@@ -4115,8 +4152,8 @@
 
 	ps = psb->ps_write;
 	if (ps->ps_nsamples) {	/* in use, reader hasn't caught up */
-		CPU_SET_ATOMIC(cpu, &pm->pm_stalled);
-		atomic_add_int(&pmc_stats.pm_intr_bufferfull, 1);
+		pm->pm_pcpu_state[cpu].pps_stalled = 1;
+		counter_u64_add(pmc_stats.pm_intr_bufferfull, 1);
 		PMCDBG6(SAM,INT,1,"(spc) cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d",
 		    cpu, pm, (void *) tf, inuserspace,
 		    (int) (psb->ps_write - psb->ps_samples),
@@ -4133,11 +4170,11 @@
 	    (int) (psb->ps_write - psb->ps_samples),
 	    (int) (psb->ps_read - psb->ps_samples));
 
-	KASSERT(pm->pm_runcount >= 0,
-	    ("[pmc,%d] pm=%p runcount %d", __LINE__, (void *) pm,
-		pm->pm_runcount));
+	KASSERT(counter_u64_fetch(pm->pm_runcount) >= 0,
+	    ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm,
+		 (unsigned long)counter_u64_fetch(pm->pm_runcount)));
 
-	atomic_add_rel_int(&pm->pm_runcount, 1);	/* hold onto PMC */
+	counter_u64_add(pm->pm_runcount, 1);	/* hold onto PMC */
 
 	ps->ps_pmc = pm;
 	if ((td = curthread) && td->td_proc)
@@ -4180,7 +4217,7 @@
  done:
 	/* mark CPU as needing processing */
 	if (callchaindepth != PMC_SAMPLE_INUSE)
-		CPU_SET_ATOMIC(cpu, &pmc_cpumask);
+		DPCPU_SET(pmc_sampled, 1);
 
 	return (error);
 }
@@ -4244,8 +4281,8 @@
 		    ("[pmc,%d] Retrieving callchain for PMC that doesn't "
 			"want it", __LINE__));
 
-		KASSERT(pm->pm_runcount > 0,
-		    ("[pmc,%d] runcount %d", __LINE__, pm->pm_runcount));
+		KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
+		    ("[pmc,%d] runcount %ld", __LINE__, (unsigned long)counter_u64_fetch(pm->pm_runcount)));
 
 		/*
 		 * Retrieve the callchain and mark the sample buffer
@@ -4275,9 +4312,7 @@
 	sched_unpin();	/* Can migrate safely now. */
 
 	/* mark CPU as needing processing */
-	CPU_SET_ATOMIC(cpu, &pmc_cpumask);
-
-	return;
+	DPCPU_SET(pmc_sampled, 1);
 }
 
 /*
@@ -4309,9 +4344,9 @@
 
 		pm = ps->ps_pmc;
 
-		KASSERT(pm->pm_runcount > 0,
-		    ("[pmc,%d] pm=%p runcount %d", __LINE__, (void *) pm,
-			pm->pm_runcount));
+		KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
+		    ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm,
+			 (unsigned long)counter_u64_fetch(pm->pm_runcount)));
 
 		po = pm->pm_owner;
 
@@ -4326,7 +4361,7 @@
 		/* If there is a pending AST wait for completion */
 		if (ps->ps_nsamples == PMC_SAMPLE_INUSE) {
 			/* Need a rescan at a later time. */
-			CPU_SET_ATOMIC(cpu, &pmc_cpumask);
+			DPCPU_SET(pmc_sampled, 1);
 			break;
 		}
 
@@ -4359,7 +4394,7 @@
 
 	entrydone:
 		ps->ps_nsamples = 0; /* mark entry as free */
-		atomic_subtract_rel_int(&pm->pm_runcount, 1);
+		counter_u64_add(pm->pm_runcount, -1);
 
 		/* increment read pointer, modulo sample size */
 		if (++ps == psb->ps_fence)
@@ -4368,7 +4403,7 @@
 			psb->ps_read = ps;
 	}
 
-	atomic_add_int(&pmc_stats.pm_log_sweeps, 1);
+	counter_u64_add(pmc_stats.pm_log_sweeps, 1);
 
 	/* Do not re-enable stalled PMCs if we failed to process any samples */
 	if (n == 0)
@@ -4390,11 +4425,11 @@
 		if (pm == NULL ||			 /* !cfg'ed */
 		    pm->pm_state != PMC_STATE_RUNNING || /* !active */
 		    !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) || /* !sampling */
-		    !CPU_ISSET(cpu, &pm->pm_cpustate) || /* !desired */
-		    !CPU_ISSET(cpu, &pm->pm_stalled)) /* !stalled */
+			!pm->pm_pcpu_state[cpu].pps_cpustate  || /* !desired */
+		    !pm->pm_pcpu_state[cpu].pps_stalled) /* !stalled */
 			continue;
 
-		CPU_CLR_ATOMIC(cpu, &pm->pm_stalled);
+		pm->pm_pcpu_state[cpu].pps_stalled = 0;
 		(*pcd->pcd_start_pmc)(cpu, adjri);
 	}
 }
@@ -4513,9 +4548,9 @@
 			    ("[pmc,%d] pm %p != pp_pmcs[%d] %p",
 				__LINE__, pm, ri, pp->pp_pmcs[ri].pp_pmc));
 
-			KASSERT(pm->pm_runcount > 0,
-			    ("[pmc,%d] bad runcount ri %d rc %d",
-				__LINE__, ri, pm->pm_runcount));
+			KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
+			    ("[pmc,%d] bad runcount ri %d rc %ld",
+				 __LINE__, ri, (unsigned long)counter_u64_fetch(pm->pm_runcount)));
 
 			/*
 			 * Change desired state, and then stop if not
@@ -4524,9 +4559,9 @@
 			 * the PMC after this code has already checked
 			 * the pm_stalled flag.
 			 */
-			if (CPU_ISSET(cpu, &pm->pm_cpustate)) {
-				CPU_CLR_ATOMIC(cpu, &pm->pm_cpustate);
-				if (!CPU_ISSET(cpu, &pm->pm_stalled)) {
+			if (pm->pm_pcpu_state[cpu].pps_cpustate) {
+				pm->pm_pcpu_state[cpu].pps_cpustate = 0;
+				if (!pm->pm_pcpu_state[cpu].pps_stalled) {
 					(void) pcd->pcd_stop_pmc(cpu, adjri);
 					pcd->pcd_read_pmc(cpu, adjri,
 					    &newvalue);
@@ -4540,9 +4575,9 @@
 				}
 			}
 
-			atomic_subtract_rel_int(&pm->pm_runcount,1);
+			counter_u64_add(pm->pm_runcount, -1);
 
-			KASSERT((int) pm->pm_runcount >= 0,
+			KASSERT((int) counter_u64_fetch(pm->pm_runcount) >= 0,
 			    ("[pmc,%d] runcount is %d", __LINE__, ri));
 
 			(void) pcd->pcd_config_pmc(cpu, adjri, NULL);
@@ -4811,7 +4846,8 @@
 pmc_initialize(void)
 {
 	int c, cpu, error, n, ri;
-	unsigned int maxcpu;
+	unsigned int maxcpu, domain;
+	struct pcpu *pc;
 	struct pmc_binding pb;
 	struct pmc_sample *ps;
 	struct pmc_classdep *pcd;
@@ -4820,6 +4856,15 @@
 	md = NULL;
 	error = 0;
 
+	pmc_stats.pm_intr_ignored = counter_u64_alloc(M_WAITOK);
+	pmc_stats.pm_intr_processed = counter_u64_alloc(M_WAITOK);
+	pmc_stats.pm_intr_bufferfull = counter_u64_alloc(M_WAITOK);
+	pmc_stats.pm_syscalls = counter_u64_alloc(M_WAITOK);
+	pmc_stats.pm_syscall_errors = counter_u64_alloc(M_WAITOK);
+	pmc_stats.pm_buffer_requests = counter_u64_alloc(M_WAITOK);
+	pmc_stats.pm_buffer_requests_failed = counter_u64_alloc(M_WAITOK);
+	pmc_stats.pm_log_sweeps = counter_u64_alloc(M_WAITOK);
+
 #ifdef	HWPMC_DEBUG
 	/* parse debug flags first */
 	if (TUNABLE_STR_FETCH(PMC_SYSCTL_NAME_PREFIX "debugflags",
@@ -4927,9 +4972,10 @@
 	for (cpu = 0; cpu < maxcpu; cpu++) {
 		if (!pmc_cpu_is_active(cpu))
 			continue;
-
-		sb = malloc(sizeof(struct pmc_samplebuffer) +
-		    pmc_nsamples * sizeof(struct pmc_sample), M_PMC,
+		pc = pcpu_find(cpu);
+		domain = pc->pc_domain;
+		sb = malloc_domain(sizeof(struct pmc_samplebuffer) +
+			pmc_nsamples * sizeof(struct pmc_sample), M_PMC, domain,
 		    M_WAITOK|M_ZERO);
 		sb->ps_read = sb->ps_write = sb->ps_samples;
 		sb->ps_fence = sb->ps_samples + pmc_nsamples;
@@ -4937,8 +4983,8 @@
 		KASSERT(pmc_pcpu[cpu] != NULL,
 		    ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
 
-		sb->ps_callchains = malloc(pmc_callchaindepth * pmc_nsamples *
-		    sizeof(uintptr_t), M_PMC, M_WAITOK|M_ZERO);
+		sb->ps_callchains = malloc_domain(pmc_callchaindepth * pmc_nsamples *
+			sizeof(uintptr_t), M_PMC, domain, M_WAITOK|M_ZERO);
 
 		for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
 			ps->ps_pc = sb->ps_callchains +
@@ -4946,8 +4992,8 @@
 
 		pmc_pcpu[cpu]->pc_sb[PMC_HR] = sb;
 
-		sb = malloc(sizeof(struct pmc_samplebuffer) +
-		    pmc_nsamples * sizeof(struct pmc_sample), M_PMC,
+		sb = malloc_domain(sizeof(struct pmc_samplebuffer) +
+			pmc_nsamples * sizeof(struct pmc_sample), M_PMC, domain,
 		    M_WAITOK|M_ZERO);
 		sb->ps_read = sb->ps_write = sb->ps_samples;
 		sb->ps_fence = sb->ps_samples + pmc_nsamples;
@@ -4955,8 +5001,8 @@
 		KASSERT(pmc_pcpu[cpu] != NULL,
 		    ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
 
-		sb->ps_callchains = malloc(pmc_callchaindepth * pmc_nsamples *
-		    sizeof(uintptr_t), M_PMC, M_WAITOK|M_ZERO);
+		sb->ps_callchains = malloc_domain(pmc_callchaindepth * pmc_nsamples *
+			sizeof(uintptr_t), M_PMC, domain, M_WAITOK|M_ZERO);
 
 		for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
 			ps->ps_pc = sb->ps_callchains +
@@ -5048,7 +5094,8 @@
 	PMCDBG0(MOD,INI,0, "cleanup");
 
 	/* switch off sampling */
-	CPU_ZERO(&pmc_cpumask);
+	CPU_FOREACH(cpu)
+		DPCPU_ID_SET(cpu, pmc_sampled, 0);
 	pmc_intr = NULL;
 
 	sx_xlock(&pmc_sx);
@@ -5157,11 +5204,11 @@
 		KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_SR] != NULL,
 		    ("[pmc,%d] Null sw cpu sample buffer cpu=%d", __LINE__,
 			cpu));
-		free(pmc_pcpu[cpu]->pc_sb[PMC_HR]->ps_callchains, M_PMC);
-		free(pmc_pcpu[cpu]->pc_sb[PMC_HR], M_PMC);
-		free(pmc_pcpu[cpu]->pc_sb[PMC_SR]->ps_callchains, M_PMC);
-		free(pmc_pcpu[cpu]->pc_sb[PMC_SR], M_PMC);
-		free(pmc_pcpu[cpu], M_PMC);
+		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_HR]->ps_callchains, M_PMC);
+		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_HR], M_PMC);
+		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_SR]->ps_callchains, M_PMC);
+		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_SR], M_PMC);
+		free_domain(pmc_pcpu[cpu], M_PMC);
 	}
 
 	free(pmc_pcpu, M_PMC);
@@ -5181,7 +5228,14 @@
 	}
 
 	pmclog_shutdown();
-
+	counter_u64_free(pmc_stats.pm_intr_ignored);
+	counter_u64_free(pmc_stats.pm_intr_processed);
+	counter_u64_free(pmc_stats.pm_intr_bufferfull);
+	counter_u64_free(pmc_stats.pm_syscalls);
+	counter_u64_free(pmc_stats.pm_syscall_errors);
+	counter_u64_free(pmc_stats.pm_buffer_requests);
+	counter_u64_free(pmc_stats.pm_buffer_requests_failed);
+	counter_u64_free(pmc_stats.pm_log_sweeps);
 	sx_xunlock(&pmc_sx); 	/* we are done */
 }
 
Index: head/sys/dev/hwpmc/hwpmc_mpc7xxx.c
===================================================================
--- head/sys/dev/hwpmc/hwpmc_mpc7xxx.c
+++ head/sys/dev/hwpmc/hwpmc_mpc7xxx.c
@@ -710,9 +710,10 @@
 		/* reload count. */
 		mpc7xxx_write_pmc(cpu, i, pm->pm_sc.pm_reloadcount);
 	}
-
-	atomic_add_int(retval ? &pmc_stats.pm_intr_processed :
-	    &pmc_stats.pm_intr_ignored, 1);
+	if (retval)
+		counter_u64_add(pmc_stats.pm_intr_processed, 1);
+	else
+		counter_u64_add(pmc_stats.pm_intr_ignored, 1);
 
 	/* Re-enable PERF exceptions. */
 	if (retval)
Index: head/sys/dev/hwpmc/hwpmc_piv.c
===================================================================
--- head/sys/dev/hwpmc/hwpmc_piv.c
+++ head/sys/dev/hwpmc/hwpmc_piv.c
@@ -1545,8 +1545,10 @@
 	if (did_interrupt)
 		lapic_reenable_pmc();
 
-	atomic_add_int(did_interrupt ? &pmc_stats.pm_intr_processed :
-	    &pmc_stats.pm_intr_ignored, 1);
+	if (did_interrupt)
+		counter_u64_add(pmc_stats.pm_intr_processed, 1);
+	else
+		counter_u64_add(pmc_stats.pm_intr_ignored, 1);
 
 	return (did_interrupt);
 }
Index: head/sys/dev/hwpmc/hwpmc_ppc970.c
===================================================================
--- head/sys/dev/hwpmc/hwpmc_ppc970.c
+++ head/sys/dev/hwpmc/hwpmc_ppc970.c
@@ -528,8 +528,10 @@
 		ppc970_write_pmc(cpu, i, pm->pm_sc.pm_reloadcount);
 	}
 
-	atomic_add_int(retval ? &pmc_stats.pm_intr_processed :
-	    &pmc_stats.pm_intr_ignored, 1);
+	if (retval)
+		counter_u64_add(pmc_stats.pm_intr_processed, 1);
+	else
+		counter_u64_add(pmc_stats.pm_intr_ignored, 1);
 
 	/* Re-enable PERF exceptions. */
 	if (retval)
Index: head/sys/dev/hwpmc/hwpmc_ppro.c
===================================================================
--- head/sys/dev/hwpmc/hwpmc_ppro.c
+++ head/sys/dev/hwpmc/hwpmc_ppro.c
@@ -725,8 +725,10 @@
 	if (retval)
 		lapic_reenable_pmc();
 
-	atomic_add_int(retval ? &pmc_stats.pm_intr_processed :
-	    &pmc_stats.pm_intr_ignored, 1);
+	if (retval)
+		counter_u64_add(pmc_stats.pm_intr_processed, 1);
+	else
+		counter_u64_add(pmc_stats.pm_intr_ignored, 1);
 
 	/* restart counters that can be restarted */
 	P6_SYNC_CTR_STATE(pc);
Index: head/sys/dev/hwpmc/hwpmc_soft.c
===================================================================
--- head/sys/dev/hwpmc/hwpmc_soft.c
+++ head/sys/dev/hwpmc/hwpmc_soft.c
@@ -441,9 +441,10 @@
 		} else
 			pc->soft_values[ri]++;
 	}
-
-	atomic_add_int(processed ? &pmc_stats.pm_intr_processed :
-	    &pmc_stats.pm_intr_ignored, 1);
+	if (processed)
+		counter_u64_add(pmc_stats.pm_intr_processed, 1);
+	else
+		counter_u64_add(pmc_stats.pm_intr_ignored, 1);
 
 	return (processed);
 }
Index: head/sys/kern/kern_pmc.c
===================================================================
--- head/sys/kern/kern_pmc.c
+++ head/sys/kern/kern_pmc.c
@@ -66,8 +66,7 @@
 /* Interrupt handler */
 int __read_mostly (*pmc_intr)(int cpu, struct trapframe *tf) = NULL;
 
-/* Bitmask of CPUs requiring servicing at hardclock time */
-volatile cpuset_t pmc_cpumask;
+DPCPU_DEFINE(uint8_t, pmc_sampled);
 
 /*
  * A global count of SS mode PMCs.  When non-zero, this means that
Index: head/sys/sys/pmc.h
===================================================================
--- head/sys/sys/pmc.h
+++ head/sys/sys/pmc.h
@@ -36,7 +36,8 @@
 #define	_SYS_PMC_H_
 
 #include <dev/hwpmc/pmc_events.h>
-
+#include <sys/proc.h>
+#include <sys/counter.h>
 #include <machine/pmc_mdep.h>
 #include <machine/profile.h>
 
@@ -552,6 +553,19 @@
  *
  * Retrieve pmc(4) driver-wide statistics.
  */
+#ifdef _KERNEL
+struct pmc_driverstats {
+	counter_u64_t	pm_intr_ignored;	/* #interrupts ignored */
+	counter_u64_t	pm_intr_processed;	/* #interrupts processed */
+	counter_u64_t	pm_intr_bufferfull;	/* #interrupts with ENOSPC */
+	counter_u64_t	pm_syscalls;		/* #syscalls */
+	counter_u64_t	pm_syscall_errors;	/* #syscalls with errors */
+	counter_u64_t	pm_buffer_requests;	/* #buffer requests */
+	counter_u64_t	pm_buffer_requests_failed; /* #failed buffer requests */
+	counter_u64_t	pm_log_sweeps;		/* #sample buffer processing
+						   passes */
+};
+#endif
 
 struct pmc_op_getdriverstats {
 	unsigned int	pm_intr_ignored;	/* #interrupts ignored */
@@ -625,9 +639,9 @@
 
 #define	PMC_HASH_SIZE				1024
 #define	PMC_MTXPOOL_SIZE			2048
-#define	PMC_LOG_BUFFER_SIZE			4
-#define	PMC_NLOGBUFFERS				1024
-#define	PMC_NSAMPLES				1024
+#define	PMC_LOG_BUFFER_SIZE			128
+#define	PMC_NLOGBUFFERS_PCPU		8
+#define	PMC_NSAMPLES				64
 #define	PMC_CALLCHAIN_DEPTH			32
 
 #define PMC_SYSCTL_NAME_PREFIX "kern." PMC_MODULE_NAME "."
@@ -701,7 +715,10 @@
  * field is '0'.
  *
  */
-
+struct pmc_pcpu_state {
+	uint8_t pps_stalled;
+	uint8_t pps_cpustate;
+} __aligned(CACHE_LINE_SIZE);
 struct pmc {
 	LIST_HEAD(,pmc_target)	pm_targets;	/* list of target processes */
 	LIST_ENTRY(pmc)		pm_next;	/* owner's list */
@@ -735,13 +752,13 @@
 		pmc_value_t	pm_initial;	/* counting PMC modes */
 	} pm_sc;
 
-	volatile cpuset_t pm_stalled;	/* marks stalled sampling PMCs */
+	struct pmc_pcpu_state *pm_pcpu_state;
 	volatile cpuset_t pm_cpustate;	/* CPUs where PMC should be active */
 	uint32_t	pm_caps;	/* PMC capabilities */
 	enum pmc_event	pm_event;	/* event being measured */
 	uint32_t	pm_flags;	/* additional flags PMC_F_... */
 	struct pmc_owner *pm_owner;	/* owner thread state */
-	int		pm_runcount;	/* #cpus currently on */
+	counter_u64_t		pm_runcount;	/* #cpus currently on */
 	enum pmc_state	pm_state;	/* current PMC state */
 	uint32_t	pm_overflowcnt;	/* count overflow interrupts */
 
@@ -816,11 +833,11 @@
 	struct proc		*po_owner;	/* owner proc */
 	uint32_t		po_flags;	/* (k) flags PMC_PO_* */
 	struct proc		*po_kthread;	/* (k) helper kthread */
-	struct pmclog_buffer	*po_curbuf;	/* current log buffer */
 	struct file		*po_file;	/* file reference */
 	int			po_error;	/* recorded error */
 	short			po_sscount;	/* # SS PMCs owned */
 	short			po_logprocmaps;	/* global mappings done */
+	struct pmclog_buffer	*po_curbuf[MAXCPU];	/* current log buffer */
 };
 
 #define	PMC_PO_OWNS_LOGFILE		0x00000001 /* has a log file */
@@ -1012,7 +1029,10 @@
 extern struct pmc_cpu **pmc_pcpu;
 
 /* driver statistics */
-extern struct pmc_op_getdriverstats pmc_stats;
+extern struct pmc_driverstats pmc_stats;
+
+/* cpu model name for pmu lookup */
+extern char pmc_cpuid[64];
 
 #if	defined(HWPMC_DEBUG)
 #include <sys/ktr.h>
Index: head/sys/sys/pmckern.h
===================================================================
--- head/sys/sys/pmckern.h
+++ head/sys/sys/pmckern.h
@@ -165,7 +165,7 @@
 extern struct sx pmc_sx;
 
 /* Per-cpu flags indicating availability of sampling data */
-extern volatile cpuset_t pmc_cpumask;
+DPCPU_DECLARE(uint8_t, pmc_sampled);
 
 /* Count of system-wide sampling PMCs in existence */
 extern volatile int pmc_ss_count;
@@ -220,7 +220,7 @@
 #define	PMC_SYSTEM_SAMPLING_ACTIVE()		(pmc_ss_count > 0)
 
 /* Check if a CPU has recorded samples. */
-#define	PMC_CPU_HAS_SAMPLES(C)	(__predict_false(CPU_ISSET(C, &pmc_cpumask)))
+#define	PMC_CPU_HAS_SAMPLES(C)	(__predict_false(DPCPU_ID_GET((C), pmc_sampled)))
 
 /*
  * Helper functions.