From: Cyrill Gorcunov Date: Fri, 7 Jan 2011 18:42:06 +0000 (+0300) Subject: perf, x86: P4 PMU - Fix unflagged overflows handling X-Git-Url: https://git.stricted.de/?a=commitdiff_plain;h=047a3772feaae8e43d81d790f3d3f80dae8ae676;p=GitHub%2Fexynos8895%2Fandroid_kernel_samsung_universal8895.git perf, x86: P4 PMU - Fix unflagged overflows handling Don found that P4 PMU reads CCCR register instead of counter itself (in attempt to catch unflagged event) this makes P4 NMI handler to consume all NMIs it observes. So the other NMI users such as kgdb simply have no chance to get NMI on their hands. Side note: at moment there is no way to run nmi-watchdog together with perf tool. This is because both 'perf top' and nmi-watchdog use same event. So while nmi-watchdog reserves one event/counter for own needs there is no room for perf tool left (there is a way to disable nmi-watchdog on boot of course). Ming has tested this patch with the following results | 1. watchdog disabled | | kgdb tests on boot OK | perf works OK | | 2. watchdog enabled, without patch perf-x86-p4-nmi-4 | | kgdb tests on boot hang | | 3. watchdog enabled, without patch perf-x86-p4-nmi-4 and do not run kgdb | tests on boot | | "perf top" partialy works | cpu-cycles no | instructions yes | cache-references no | cache-misses no | branch-instructions no | branch-misses yes | bus-cycles no | | 4. watchdog enabled, with patch perf-x86-p4-nmi-4 applied | | kgdb tests on boot OK | perf does not work, NMI "Dazed and confused" messages show up | Which means we still have problems with p4 box due to 'unknown' nmi happens but at least it should fix kgdb test cases. Reported-by: Jason Wessel Reported-by: Don Zickus Signed-off-by: Cyrill Gorcunov Acked-by: Don Zickus Acked-by: Lin Ming Cc: Stephane Eranian Cc: Peter Zijlstra LKML-Reference: <4D275E7E.3040903@gmail.com> Signed-off-by: Ingo Molnar --- diff --git a/arch/x86/include/asm/perf_event_p4.h b/arch/x86/include/asm/perf_event_p4.h index 295e2ff18a6a..e2f6a99f14ab 100644 --- a/arch/x86/include/asm/perf_event_p4.h +++ b/arch/x86/include/asm/perf_event_p4.h @@ -20,6 +20,9 @@ #define ARCH_P4_MAX_ESCR (ARCH_P4_TOTAL_ESCR - ARCH_P4_RESERVED_ESCR) #define ARCH_P4_MAX_CCCR (18) +#define ARCH_P4_CNTRVAL_BITS (40) +#define ARCH_P4_CNTRVAL_MASK ((1ULL << ARCH_P4_CNTRVAL_BITS) - 1) + #define P4_ESCR_EVENT_MASK 0x7e000000U #define P4_ESCR_EVENT_SHIFT 25 #define P4_ESCR_EVENTMASK_MASK 0x01fffe00U diff --git a/arch/x86/kernel/cpu/perf_event_p4.c b/arch/x86/kernel/cpu/perf_event_p4.c index 81400b93e694..e56b9bfbabd1 100644 --- a/arch/x86/kernel/cpu/perf_event_p4.c +++ b/arch/x86/kernel/cpu/perf_event_p4.c @@ -753,19 +753,21 @@ out: static inline int p4_pmu_clear_cccr_ovf(struct hw_perf_event *hwc) { - int overflow = 0; - u32 low, high; + u64 v; - rdmsr(hwc->config_base + hwc->idx, low, high); - - /* we need to check high bit for unflagged overflows */ - if ((low & P4_CCCR_OVF) || !(high & (1 << 31))) { - overflow = 1; - (void)checking_wrmsrl(hwc->config_base + hwc->idx, - ((u64)low) & ~P4_CCCR_OVF); + /* an official way for overflow indication */ + rdmsrl(hwc->config_base + hwc->idx, v); + if (v & P4_CCCR_OVF) { + wrmsrl(hwc->config_base + hwc->idx, v & ~P4_CCCR_OVF); + return 1; } - return overflow; + /* it might be unflagged overflow */ + rdmsrl(hwc->event_base + hwc->idx, v); + if (!(v & ARCH_P4_CNTRVAL_MASK)) + return 1; + + return 0; } static void p4_pmu_disable_pebs(void) @@ -1152,9 +1154,9 @@ static __initconst const struct x86_pmu p4_pmu = { */ .num_counters = ARCH_P4_MAX_CCCR, .apic = 1, - .cntval_bits = 40, - .cntval_mask = (1ULL << 40) - 1, - .max_period = (1ULL << 39) - 1, + .cntval_bits = ARCH_P4_CNTRVAL_BITS, + .cntval_mask = ARCH_P4_CNTRVAL_MASK, + .max_period = (1ULL << (ARCH_P4_CNTRVAL_BITS - 1)) - 1, .hw_config = p4_hw_config, .schedule_events = p4_pmu_schedule_events, /*