From patchwork Fri Feb 28 06:29:06 2014 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Michael Ellerman X-Patchwork-Id: 325041 Return-Path: X-Original-To: patchwork-incoming@ozlabs.org Delivered-To: patchwork-incoming@ozlabs.org Received: from ozlabs.org (localhost [IPv6:::1]) by ozlabs.org (Postfix) with ESMTP id 706DD2C0267 for ; Fri, 28 Feb 2014 17:30:15 +1100 (EST) Received: by ozlabs.org (Postfix) id D5D8E2C00A0; Fri, 28 Feb 2014 17:29:12 +1100 (EST) Delivered-To: linuxppc-dev@ozlabs.org Received: by ozlabs.org (Postfix, from userid 1034) id CB7D12C00A2; Fri, 28 Feb 2014 17:29:12 +1100 (EST) From: Michael Ellerman To: Subject: [PATCH 2/2] powerpc/perf: Add lost exception workaround Date: Fri, 28 Feb 2014 17:29:06 +1100 Message-Id: <1393568946-4136-2-git-send-email-mpe@ellerman.id.au> X-Mailer: git-send-email 1.8.3.2 In-Reply-To: <1393568946-4136-1-git-send-email-mpe@ellerman.id.au> References: <1393568946-4136-1-git-send-email-mpe@ellerman.id.au> X-BeenThere: linuxppc-dev@lists.ozlabs.org X-Mailman-Version: 2.1.16 Precedence: list List-Id: Linux on PowerPC Developers Mail List List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , MIME-Version: 1.0 Errors-To: linuxppc-dev-bounces+patchwork-incoming=ozlabs.org@lists.ozlabs.org Sender: "Linuxppc-dev" 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 --- arch/powerpc/include/asm/reg.h | 2 + arch/powerpc/perf/core-book3s.c | 100 +++++++++++++++++++++++++++++++++++++++- arch/powerpc/perf/power8-pmu.c | 5 ++ 3 files changed, 105 insertions(+), 2 deletions(-) diff --git a/arch/powerpc/include/asm/reg.h b/arch/powerpc/include/asm/reg.h index 90c06ec..3003472 100644 --- a/arch/powerpc/include/asm/reg.h +++ b/arch/powerpc/include/asm/reg.h @@ -670,6 +670,7 @@ #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 */ @@ -703,6 +704,7 @@ #define SPRN_EBBHR 804 /* Event based branch handler register */ #define SPRN_EBBRR 805 /* Event based branch return register */ #define SPRN_BESCR 806 /* Branch event status and control register */ +#define BESCR_GE 0x8000000000000000ULL /* Global Enable */ #define SPRN_WORT 895 /* Workload optimization register - thread */ #define SPRN_PMC1 787 diff --git a/arch/powerpc/perf/core-book3s.c b/arch/powerpc/perf/core-book3s.c index 67cf220..72cdd05 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 POWER8 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 POWER8 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); @@ -1144,6 +1238,8 @@ 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 #include #include @@ -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);