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: Benjamin Herrenschmidt <benh@kernel.crashing.org>

2 files changed, 103 insertions, 2 deletions

diff --git a/arch/powerpc/perf/core-book3s.c b/arch/powerpc/perf/core-book3s.c
index 704d1c5..53ac1b2 100644
--- a/arch/powerpc/perf/core-book3s.c
+++ b/arch/powerpc/perf/core-book3s.c
@@ -120,6 +120,7 @@ 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(bool ebb) { }
 #endif /* CONFIG_PPC32 */
 
 static bool regs_use_siar(struct pt_regs *regs)
@@ -545,10 +546,18 @@ static unsigned long ebb_switch_in(bool ebb, unsigned long mmcr0)
 	/* Enable EBB and read/write to all 6 PMCs for userspace */
 	mmcr0 |= MMCR0_EBE | MMCR0_PMCC_U6;
 
-	/* Add any bits from the user reg, FC or PMAO */
+	/*
+	 * Add any bits from the user MMCR0, FC or PMAO. This is compatible
+	 * with pmao_restore_workaround() because we may add PMAO but we never
+	 * clear it here.
+	 */
 	mmcr0 |= current->thread.mmcr0;
 
-	/* Be careful not to set PMXE if userspace had it cleared */
+	/*
+	 * Be careful not to set PMXE if userspace had it cleared. This is also
+	 * compatible with pmao_restore_workaround() because it has already
+	 * cleared PMXE and we leave PMAO alone.
+	 */
 	if (!(current->thread.mmcr0 & MMCR0_PMXE))
 		mmcr0 &= ~MMCR0_PMXE;
 
@@ -559,6 +568,91 @@ static unsigned long ebb_switch_in(bool ebb, unsigned long mmcr0)
 out:
 	return mmcr0;
 }
+
+static void pmao_restore_workaround(bool ebb)
+{
+	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.
+	 *
+	 * The logic is the same for EBB, except that the exception is gated by
+	 * us having interrupts hard disabled as well as the fact that we are
+	 * not in userspace. The exception is finally delivered when we return
+	 * to userspace.
+	 */
+
+	/* Only if PMAO is set and PMAO_SYNC is clear */
+	if ((current->thread.mmcr0 & (MMCR0_PMAO | MMCR0_PMAO_SYNC)) != MMCR0_PMAO)
+		return;
+
+	/* If we're doing EBB, only if BESCR[GE] is set */
+	if (ebb && !(current->thread.bescr & BESCR_GE))
+		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);
@@ -1191,6 +1285,8 @@ static void power_pmu_enable(struct pmu *pmu)
 	cpuhw->mmcr[0] |= MMCR0_PMXE | MMCR0_FCECE;
 
  out_enable:
+	pmao_restore_workaround(ebb);
+
 	mmcr0 = ebb_switch_in(ebb, cpuhw->mmcr[0]);
 
 	mb();
diff --git a/arch/powerpc/perf/power8-pmu.c b/arch/powerpc/perf/power8-pmu.c
index 96cee20..64f04cf 100644
--- a/arch/powerpc/perf/power8-pmu.c
+++ b/arch/powerpc/perf/power8-pmu.c
@@ -10,6 +10,8 @@
  * 2 of the License, or (at your option) any later version.
  */
 
+#define pr_fmt(fmt)	"power8-pmu: " fmt
+
 #include <linux/kernel.h>
 #include <linux/perf_event.h>
 #include <asm/firmware.h>
@@ -774,6 +776,9 @@ static int __init init_power8_pmu(void)
 	/* Tell userspace that EBB is supported */
 	cur_cpu_spec->cpu_user_features2 |= PPC_FEATURE2_EBB;
 
+	if (cpu_has_feature(CPU_FTR_PMAO_BUG))
+		pr_info("PMAO restore workaround active.\n");
+
 	return 0;
 }
 early_initcall(init_power8_pmu);