powerpc/perf: Add lost exception workaround

Some power8 revisions have a hardware bug where we can lose a PMU
exception, this commit adds a workaround to detect the bad condition and
rectify the situation.

See the comment in the commit for a full description.

Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Signed-off-by: Paul Mackerras <paulus@samba.org>

3 files changed, 89 insertions, 1 deletions

diff --git a/arch/powerpc/include/asm/reg.h b/arch/powerpc/include/asm/reg.h
index 9a414786955f70..92e8df6cf5ccbc 100644
--- a/arch/powerpc/include/asm/reg.h
+++ b/arch/powerpc/include/asm/reg.h
@@ -676,6 +676,7 @@
 #define   MMCR0_PMC1CE	0x00008000UL /* PMC1 count enable*/
 #define   MMCR0_PMCjCE	0x00004000UL /* PMCj count enable*/
 #define   MMCR0_TRIGGER	0x00002000UL /* TRIGGER enable */
+#define   MMCR0_PMAO_SYNC 0x00000800UL /* PMU interrupt is synchronous */
 #define   MMCR0_PMAO	0x00000080UL /* performance monitor alert has occurred, set to 0 after handling exception */
 #define   MMCR0_SHRFC	0x00000040UL /* SHRre freeze conditions between threads */
 #define   MMCR0_FC56	0x00000010UL /* freeze counters 5 and 6 */
diff --git a/arch/powerpc/perf/core-book3s.c b/arch/powerpc/perf/core-book3s.c
index d3ee2e50a3a6fb..ba200939bdf936 100644
--- a/arch/powerpc/perf/core-book3s.c
+++ b/arch/powerpc/perf/core-book3s.c
@@ -108,6 +108,7 @@ static inline void power_pmu_bhrb_enable(struct perf_event *event) {}
 static inline void power_pmu_bhrb_disable(struct perf_event *event) {}
 void power_pmu_flush_branch_stack(void) {}
 static inline void power_pmu_bhrb_read(struct cpu_hw_events *cpuhw) {}
+static void pmao_restore_workaround(void) { }
 #endif /* CONFIG_PPC32 */
 
 static bool regs_use_siar(struct pt_regs *regs)
@@ -464,6 +465,81 @@ void power_pmu_bhrb_read(struct cpu_hw_events *cpuhw)
 	return;
 }
 
+static void pmao_restore_workaround(void)
+{
+	unsigned pmcs[6];
+
+	if (!cpu_has_feature(CPU_FTR_PMAO_BUG))
+		return;
+
+	/*
+	 * On POWER8E there is a hardware defect which affects the PMU context
+	 * switch logic, ie. power_pmu_disable/enable().
+	 *
+	 * When a counter overflows PMXE is cleared and FC/PMAO is set in MMCR0
+	 * by the hardware. Sometime later the actual PMU exception is
+	 * delivered.
+	 *
+	 * If we context switch, or simply disable/enable, the PMU prior to the
+	 * exception arriving, the exception will be lost when we clear PMAO.
+	 *
+	 * When we reenable the PMU, we will write the saved MMCR0 with PMAO
+	 * set, and this _should_ generate an exception. However because of the
+	 * defect no exception is generated when we write PMAO, and we get
+	 * stuck with no counters counting but no exception delivered.
+	 *
+	 * The workaround is to detect this case and tweak the hardware to
+	 * create another pending PMU exception.
+	 *
+	 * We do that by setting up PMC6 (cycles) for an imminent overflow and
+	 * enabling the PMU. That causes a new exception to be generated in the
+	 * chip, but we don't take it yet because we have interrupts hard
+	 * disabled. We then write back the PMU state as we want it to be seen
+	 * by the exception handler. When we reenable interrupts the exception
+	 * handler will be called and see the correct state.
+	 */
+
+	/* Only if PMAO is set and PMAO_SYNC is clear */
+	if ((current->thread.mmcr0 & (MMCR0_PMAO | MMCR0_PMAO_SYNC)) != MMCR0_PMAO)
+		return;
+
+	/*
+	 * We are already soft-disabled in power_pmu_enable(). We need to hard
+	 * enable to actually prevent the PMU exception from firing.
+	 */
+	hard_irq_disable();
+
+	/*
+	 * This is a bit gross, but we know we're on POWER8E and have 6 PMCs.
+	 * Using read/write_pmc() in a for loop adds 12 function calls and
+	 * almost doubles our code size.
+	 */
+	pmcs[0] = mfspr(SPRN_PMC1);
+	pmcs[1] = mfspr(SPRN_PMC2);
+	pmcs[2] = mfspr(SPRN_PMC3);
+	pmcs[3] = mfspr(SPRN_PMC4);
+	pmcs[4] = mfspr(SPRN_PMC5);
+	pmcs[5] = mfspr(SPRN_PMC6);
+
+	/* Ensure all freeze bits are unset */
+	mtspr(SPRN_MMCR2, 0);
+
+	/* Set up PMC6 to overflow in one cycle */
+	mtspr(SPRN_PMC6, 0x7FFFFFFE);
+
+	/* Enable exceptions and unfreeze PMC6 */
+	mtspr(SPRN_MMCR0, MMCR0_PMXE | MMCR0_PMCjCE | MMCR0_PMAO);
+
+	/* Now we need to refreeze and restore the PMCs */
+	mtspr(SPRN_MMCR0, MMCR0_FC | MMCR0_PMAO);
+
+	mtspr(SPRN_PMC1, pmcs[0]);
+	mtspr(SPRN_PMC2, pmcs[1]);
+	mtspr(SPRN_PMC3, pmcs[2]);
+	mtspr(SPRN_PMC4, pmcs[3]);
+	mtspr(SPRN_PMC5, pmcs[4]);
+	mtspr(SPRN_PMC6, pmcs[5]);
+}
 #endif /* CONFIG_PPC64 */
 
 static void perf_event_interrupt(struct pt_regs *regs);
@@ -1025,6 +1101,8 @@ static void power_pmu_enable(struct pmu *pmu)
 	cpuhw->mmcr[0] |= MMCR0_PMXE | MMCR0_FCECE;
 
  out_enable:
+	pmao_restore_workaround();
+
 	mb();
 	write_mmcr0(cpuhw, cpuhw->mmcr[0]);
 
diff --git a/arch/powerpc/perf/power8-pmu.c b/arch/powerpc/perf/power8-pmu.c
index 9aefaebedef14b..530d74e0a2225a 100644
--- a/arch/powerpc/perf/power8-pmu.c
+++ b/arch/powerpc/perf/power8-pmu.c
@@ -601,10 +601,19 @@ static struct power_pmu power8_pmu = {
 
 static int __init init_power8_pmu(void)
 {
+	int err;
+
 	if (!cur_cpu_spec->oprofile_cpu_type ||
 	    strcmp(cur_cpu_spec->oprofile_cpu_type, "ppc64/power8"))
 		return -ENODEV;
 
-	return register_power_pmu(&power8_pmu);
+	err = register_power_pmu(&power8_pmu);
+	if (err)
+		return err;
+
+	if (cpu_has_feature(CPU_FTR_PMAO_BUG))
+		pr_info("PMAO restore workaround active.\n");
+
+	return 0;
 }
 early_initcall(init_power8_pmu);