MFC r290811:

  Fix hwpmc "stalled" behavior

  Currently, there is a single pm_stalled flag that tracks whether a
  performance monitor was "stalled" due to insufficent ring buffer
  space for samples. However, because the same performance monitor
  can run on multiple processes or threads at the same time, a single
  pm_stalled flag that impacts them all seems insufficient.

  In particular, you can hit corner cases where the code fails to stop
  performance monitors during a context switch out, because it thinks
  the performance monitor is already stopped. However, in reality,
  it may be that only the monitor running on a different CPU was stalled.

  This patch attempts to fix that behavior by tracking on a per-CPU basis
  whether a PM desires to run and whether it is "stalled". This lets the
  code make better decisions about when to stop PMs and when to try to
  restart them. Ideally, we should avoid the case where the code fails
  to stop a PM during a context switch out.

MFC r290813:
  Optimizations to the way hwpmc gathers user callchains

  Changes to the code to gather user stacks:
  * Delay setting pmc_cpumask until we actually have the stack.
  * When recording user stack traces, only walk the portion of the ring
    that should have samples for us.

MFC r290929:
  Change the driver stats to what they really are: unsigned values.

  When pmcstat exits after some samples were dropped, give the user an
  idea of how many were lost. (Granted, these are global numbers, but
  they may still help quantify the scope of the loss.)

MFC r290930:
  Improve accuracy of PMC sampling frequency

  The code tracks a counter which is the number of events until the next
  sample. On context switch in, it loads the saved counter. On context
  switch out, it tries to calculate a new saved counter.

  Problems:

  1. The saved counter was shared by all threads in a process. However, this
  means that all threads would be initially loaded with the same saved
  counter. However, that could result in sampling more often than once every
  X number of events.

  2. The calculation to determine a new saved counter was backwards. It
  added when it should have subtracted, and subtracted when it should have
  added. Assume a single-threaded process with a reload count of 1000
  events. Assuming the counter on context switch in was 100 and the counter
  on context switch out was 50 (meaning the thread has "consumed" 50 more
  events), the code would calculate a new saved counter of 150 (instead of
  the proper 50).

  Fix:

  1. As soon as the saved counter is used to initialize a monitor for a
  thread on context switch in, set the saved counter to the reload count.
  That way, subsequent threads to use the saved counter will get the full
  reload count, assuring we sample at least once every X number of events
  (across all threads).

  2. Change the calculation of the saved counter. Due to the change to the
  saved counter in #1, we simply need to add (modulo the reload count) the
  remaining counter time we retrieve from the CPU when a thread is context
  switched out.

MFC r291016:
  Support a wider history counter in pmcstat(8) gmon output

  pmcstat(8) contains an option to output sampling data in a gmon format
  compatible with gprof(1). Currently, it uses the default histcounter,
  which is an (unsigned short). With large sets of sampling data, it
  is possible to overflow the maximum value provided by an (unsigned
  short).

  This change adds the -e argument to pmcstat. If -e and -g are both
  specified, pmcstat will use a histcounter type of uint64_t.

MFC r291017:
  Fix the date on the pmcstat(8) man page from r291016.

lib/libpmc/pmc.h

@@ -36,14 +36,15 @@
  * Driver statistics.
  */
 struct pmc_driverstats {
-	int	pm_intr_ignored;	/* #interrupts ignored */
-	int	pm_intr_processed;	/* #interrupts processed */
-	int	pm_intr_bufferfull;	/* #interrupts with ENOSPC */
-	int	pm_syscalls;		/* #syscalls */
-	int	pm_syscall_errors;	/* #syscalls with errors */
-	int	pm_buffer_requests;	/* #buffer requests */
-	int	pm_buffer_requests_failed; /* #failed buffer requests */
-	int	pm_log_sweeps;		/* #sample buffer processing passes */
+	unsigned int	pm_intr_ignored;	/* #interrupts ignored */
+	unsigned int	pm_intr_processed;	/* #interrupts processed */
+	unsigned int	pm_intr_bufferfull;	/* #interrupts with ENOSPC */
+	unsigned int	pm_syscalls;		/* #syscalls */
+	unsigned int	pm_syscall_errors;	/* #syscalls with errors */
+	unsigned int	pm_buffer_requests;	/* #buffer requests */
+	unsigned int	pm_buffer_requests_failed; /* #failed buffer requests */
+	unsigned int	pm_log_sweeps;		/* #sample buffer processing
+						   passes */
 };
 
 /*

sys/dev/hwpmc/hwpmc_mod.c

@@ -1287,8 +1287,16 @@ pmc_process_csw_in(struct thread *td)
 		 */
 		if (PMC_TO_MODE(pm) == PMC_MODE_TS) {
 			mtx_pool_lock_spin(pmc_mtxpool, pm);
+
+			/*
+			 * Use the saved value calculated after the most recent
+			 * thread switch out to start this counter.  Reset
+			 * the saved count in case another thread from this
+			 * process switches in before any threads switch out.
+			 */
 			newvalue = PMC_PCPU_SAVED(cpu,ri) =
 			    pp->pp_pmcs[ri].pp_pmcval;
+			pp->pp_pmcs[ri].pp_pmcval = pm->pm_sc.pm_reloadcount;
 			mtx_pool_unlock_spin(pmc_mtxpool, pm);
 		} else {
 			KASSERT(PMC_TO_MODE(pm) == PMC_MODE_TC,
@@ -1303,6 +1311,15 @@ pmc_process_csw_in(struct thread *td)
 		PMCDBG3(CSW,SWI,1,"cpu=%d ri=%d new=%jd", cpu, ri, newvalue);
 
 		pcd->pcd_write_pmc(cpu, adjri, newvalue);
+
+		/* If a sampling mode PMC, reset stalled state. */
+		if (PMC_TO_MODE(pm) == PMC_MODE_TS)
+			CPU_CLR_ATOMIC(cpu, &pm->pm_stalled);
+
+		/* Indicate that we desire this to run. */
+		CPU_SET_ATOMIC(cpu, &pm->pm_cpustate);
+
+		/* Start the PMC. */
 		pcd->pcd_start_pmc(cpu, adjri);
 	}
 
@@ -1397,8 +1414,14 @@ pmc_process_csw_out(struct thread *td)
 		    ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
 			__LINE__, PMC_TO_ROWINDEX(pm), ri));
 
-		/* Stop hardware if not already stopped */
-		if (pm->pm_stalled == 0)
+		/*
+		 * Change desired state, and then stop if not stalled.
+		 * This two-step dance should avoid race conditions where
+		 * 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))
 			pcd->pcd_stop_pmc(cpu, adjri);
 
 		/* reduce this PMC's runcount */
@@ -1421,31 +1444,43 @@ pmc_process_csw_out(struct thread *td)
 
 			pcd->pcd_read_pmc(cpu, adjri, &newvalue);
 
-			tmp = newvalue - PMC_PCPU_SAVED(cpu,ri);
-
-			PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d tmp=%jd", cpu, ri,
-			    tmp);
-
 			if (mode == PMC_MODE_TS) {
+				PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d tmp=%jd (samp)",
+				    cpu, ri, PMC_PCPU_SAVED(cpu,ri) - newvalue);
 
 				/*
 				 * For sampling process-virtual PMCs,
-				 * we expect the count to be
-				 * decreasing as the 'value'
-				 * programmed into the PMC is the
-				 * number of events to be seen till
-				 * the next sampling interrupt.
+				 * newvalue is the number of events to be seen
+				 * until the next sampling interrupt.
+				 * We can just add the events left from this
+				 * invocation to the counter, then adjust
+				 * in case we overflow our range.
+				 *
+				 * (Recall that we reload the counter every
+				 * time we use it.)
 				 */
-				if (tmp < 0)
-					tmp += pm->pm_sc.pm_reloadcount;
 				mtx_pool_lock_spin(pmc_mtxpool, pm);
-				pp->pp_pmcs[ri].pp_pmcval -= tmp;
-				if ((int64_t) pp->pp_pmcs[ri].pp_pmcval <= 0)
-					pp->pp_pmcs[ri].pp_pmcval +=
+
+				pp->pp_pmcs[ri].pp_pmcval += newvalue;
+				if (pp->pp_pmcs[ri].pp_pmcval >
+				    pm->pm_sc.pm_reloadcount)
+					pp->pp_pmcs[ri].pp_pmcval -=
 					    pm->pm_sc.pm_reloadcount;
+				KASSERT(pp->pp_pmcs[ri].pp_pmcval > 0 &&
+				    pp->pp_pmcs[ri].pp_pmcval <=
+				    pm->pm_sc.pm_reloadcount,
+				    ("[pmc,%d] pp_pmcval outside of expected "
+				    "range cpu=%d ri=%d pp_pmcval=%jx "
+				    "pm_reloadcount=%jx", __LINE__, cpu, ri,
+				    pp->pp_pmcs[ri].pp_pmcval,
+				    pm->pm_sc.pm_reloadcount));
 				mtx_pool_unlock_spin(pmc_mtxpool, pm);
 
 			} else {
+				tmp = newvalue - PMC_PCPU_SAVED(cpu,ri);
+
+				PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d tmp=%jd (count)",
+				    cpu, ri, tmp);
 
 				/*
 				 * For counting process-virtual PMCs,
@@ -2263,8 +2298,9 @@ pmc_release_pmc_descriptor(struct pmc *pm)
 		pmc_select_cpu(cpu);
 
 		/* switch off non-stalled CPUs */
+		CPU_CLR_ATOMIC(cpu, &pm->pm_cpustate);
 		if (pm->pm_state == PMC_STATE_RUNNING &&
-		    pm->pm_stalled == 0) {
+		    !CPU_ISSET(cpu, &pm->pm_stalled)) {
 
 			phw = pmc_pcpu[cpu]->pc_hwpmcs[ri];
 
@@ -2678,8 +2714,15 @@ pmc_start(struct pmc *pm)
 	if ((error = pcd->pcd_write_pmc(cpu, adjri,
 		 PMC_IS_SAMPLING_MODE(mode) ?
 		 pm->pm_sc.pm_reloadcount :
-		 pm->pm_sc.pm_initial)) == 0)
+		 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);
+
+		/* Indicate that we desire this to run. Start it. */
+		CPU_SET_ATOMIC(cpu, &pm->pm_cpustate);
 		error = pcd->pcd_start_pmc(cpu, adjri);
+	}
 	critical_exit();
 
 	pmc_restore_cpu_binding(&pb);
@@ -2741,6 +2784,7 @@ pmc_stop(struct pmc *pm)
 	ri = PMC_TO_ROWINDEX(pm);
 	pcd = pmc_ri_to_classdep(md, ri, &adjri);
 
+	CPU_CLR_ATOMIC(cpu, &pm->pm_cpustate);
 	critical_enter();
 	if ((error = pcd->pcd_stop_pmc(cpu, adjri)) == 0)
 		error = pcd->pcd_read_pmc(cpu, adjri, &pm->pm_sc.pm_initial);
@@ -4049,12 +4093,13 @@ pmc_process_interrupt(int cpu, int ring, struct pmc *pm, struct trapframe *tf,
 
 	ps = psb->ps_write;
 	if (ps->ps_nsamples) {	/* in use, reader hasn't caught up */
-		pm->pm_stalled = 1;
+		CPU_SET_ATOMIC(cpu, &pm->pm_stalled);
 		atomic_add_int(&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),
 		    (int) (psb->ps_read - psb->ps_samples));
+		callchaindepth = 1;
 		error = ENOMEM;
 		goto done;
 	}
@@ -4112,7 +4157,8 @@ pmc_process_interrupt(int cpu, int ring, struct pmc *pm, struct trapframe *tf,
 
  done:
 	/* mark CPU as needing processing */
-	CPU_SET_ATOMIC(cpu, &pmc_cpumask);
+	if (callchaindepth != PMC_SAMPLE_INUSE)
+		CPU_SET_ATOMIC(cpu, &pmc_cpumask);
 
 	return (error);
 }
@@ -4126,10 +4172,9 @@ pmc_process_interrupt(int cpu, int ring, struct pmc *pm, struct trapframe *tf,
 static void
 pmc_capture_user_callchain(int cpu, int ring, struct trapframe *tf)
 {
-	int i;
 	struct pmc *pm;
 	struct thread *td;
-	struct pmc_sample *ps;
+	struct pmc_sample *ps, *ps_end;
 	struct pmc_samplebuffer *psb;
 #ifdef	INVARIANTS
 	int ncallchains;
@@ -4148,15 +4193,17 @@ pmc_capture_user_callchain(int cpu, int ring, struct trapframe *tf)
 
 	/*
 	 * Iterate through all deferred callchain requests.
+	 * Walk from the current read pointer to the current
+	 * write pointer.
 	 */
 
-	ps = psb->ps_samples;
-	for (i = 0; i < pmc_nsamples; i++, ps++) {
-
+	ps = psb->ps_read;
+	ps_end = psb->ps_write;
+	do {
 		if (ps->ps_nsamples != PMC_SAMPLE_INUSE)
-			continue;
+			goto next;
 		if (ps->ps_td != td)
-			continue;
+			goto next;
 
 		KASSERT(ps->ps_cpu == cpu,
 		    ("[pmc,%d] cpu mismatch ps_cpu=%d pcpu=%d", __LINE__,
@@ -4181,7 +4228,12 @@ pmc_capture_user_callchain(int cpu, int ring, struct trapframe *tf)
 #ifdef	INVARIANTS
 		ncallchains++;
 #endif
-	}
+
+next:
+		/* increment the pointer, modulo sample ring size */
+		if (++ps == psb->ps_fence)
+			ps = psb->ps_samples;
+	} while (ps != ps_end);
 
 	KASSERT(ncallchains > 0,
 	    ("[pmc,%d] cpu %d didn't find a sample to collect", __LINE__,
@@ -4191,6 +4243,9 @@ pmc_capture_user_callchain(int cpu, int ring, struct trapframe *tf)
 	    ("[pmc,%d] invalid td_pinned value", __LINE__));
 	sched_unpin();	/* Can migrate safely now. */
 
+	/* mark CPU as needing processing */
+	CPU_SET_ATOMIC(cpu, &pmc_cpumask);
+
 	return;
 }
 
@@ -4304,10 +4359,11 @@ pmc_process_samples(int cpu, int ring)
 		if (pm == NULL ||			 /* !cfg'ed */
 		    pm->pm_state != PMC_STATE_RUNNING || /* !active */
 		    !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) || /* !sampling */
-		    pm->pm_stalled == 0) /* !stalled */
+		    !CPU_ISSET(cpu, &pm->pm_cpustate) || /* !desired */
+		    !CPU_ISSET(cpu, &pm->pm_stalled)) /* !stalled */
 			continue;
 
-		pm->pm_stalled = 0;
+		CPU_CLR_ATOMIC(cpu, &pm->pm_stalled);
 		(*pcd->pcd_start_pmc)(cpu, adjri);
 	}
 }
@@ -4426,23 +4482,31 @@ pmc_process_exit(void *arg __unused, struct proc *p)
 			    ("[pmc,%d] pm %p != pp_pmcs[%d] %p",
 				__LINE__, pm, ri, pp->pp_pmcs[ri].pp_pmc));
 
-			(void) pcd->pcd_stop_pmc(cpu, adjri);
-
 			KASSERT(pm->pm_runcount > 0,
 			    ("[pmc,%d] bad runcount ri %d rc %d",
 				__LINE__, ri, pm->pm_runcount));
 
-			/* Stop hardware only if it is actually running */
-			if (pm->pm_state == PMC_STATE_RUNNING &&
-			    pm->pm_stalled == 0) {
-				pcd->pcd_read_pmc(cpu, adjri, &newvalue);
-				tmp = newvalue -
-				    PMC_PCPU_SAVED(cpu,ri);
-
-				mtx_pool_lock_spin(pmc_mtxpool, pm);
-				pm->pm_gv.pm_savedvalue += tmp;
-				pp->pp_pmcs[ri].pp_pmcval += tmp;
-				mtx_pool_unlock_spin(pmc_mtxpool, pm);
+			/*
+			 * Change desired state, and then stop if not
+			 * stalled. This two-step dance should avoid
+			 * race conditions where an interrupt re-enables
+			 * 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)) {
+					(void) pcd->pcd_stop_pmc(cpu, adjri);
+					pcd->pcd_read_pmc(cpu, adjri,
+					    &newvalue);
+					tmp = newvalue -
+					    PMC_PCPU_SAVED(cpu,ri);
+
+					mtx_pool_lock_spin(pmc_mtxpool, pm);
+					pm->pm_gv.pm_savedvalue += tmp;
+					pp->pp_pmcs[ri].pp_pmcval += tmp;
+					mtx_pool_unlock_spin(pmc_mtxpool, pm);
+				}
 			}
 
 			atomic_subtract_rel_int(&pm->pm_runcount,1);

sys/sys/pmc.h

@@ -534,14 +534,15 @@ struct pmc_op_configurelog {
  */
 
 struct pmc_op_getdriverstats {
-	int	pm_intr_ignored;	/* #interrupts ignored */
-	int	pm_intr_processed;	/* #interrupts processed */
-	int	pm_intr_bufferfull;	/* #interrupts with ENOSPC */
-	int	pm_syscalls;		/* #syscalls */
-	int	pm_syscall_errors;	/* #syscalls with errors */
-	int	pm_buffer_requests;	/* #buffer requests */
-	int	pm_buffer_requests_failed; /* #failed buffer requests */
-	int	pm_log_sweeps;		/* #sample buffer processing passes */
+	unsigned int	pm_intr_ignored;	/* #interrupts ignored */
+	unsigned int	pm_intr_processed;	/* #interrupts processed */
+	unsigned int	pm_intr_bufferfull;	/* #interrupts with ENOSPC */
+	unsigned int	pm_syscalls;		/* #syscalls */
+	unsigned int	pm_syscall_errors;	/* #syscalls with errors */
+	unsigned int	pm_buffer_requests;	/* #buffer requests */
+	unsigned int	pm_buffer_requests_failed; /* #failed buffer requests */
+	unsigned int	pm_log_sweeps;		/* #sample buffer processing
+						   passes */
 };
 
 /*
@@ -598,6 +599,7 @@ struct pmc_op_getdyneventinfo {
 
 #include <sys/malloc.h>
 #include <sys/sysctl.h>
+#include <sys/_cpuset.h>
 
 #include <machine/frame.h>
 
@@ -713,7 +715,8 @@ struct pmc {
 		pmc_value_t	pm_initial;	/* counting PMC modes */
 	} pm_sc;
 
-	uint32_t	pm_stalled;	/* marks stalled sampling PMCs */
+	volatile cpuset_t pm_stalled;	/* marks stalled sampling PMCs */
+	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_... */