]> Pileus Git - ~andy/linux/commitdiff
MIPS: perf: Add perf functionality for BMIPS5000
authorAl Cooper <alcooperx@gmail.com>
Fri, 13 Jul 2012 20:44:54 +0000 (16:44 -0400)
committerRalf Baechle <ralf@linux-mips.org>
Thu, 11 Oct 2012 09:04:34 +0000 (11:04 +0200)
Add hardware performance counter support to kernel "perf" code for
BMIPS5000. The BMIPS5000 performance counters are similar to MIPS
MTI cores, so the changes were mostly made in perf_event_mipsxx.c
which is typically for MTI cores.

Signed-off-by: Al Cooper <alcooperx@gmail.com>
Cc: linux-mips@linux-mips.org
Cc: linux-kernel@vger.kernel.org
Patchwork: https://patchwork.linux-mips.org/patch/4109/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
arch/mips/kernel/perf_event_mipsxx.c

index cb213089ed2b68d9a09a4b527e993e278db5b13f..a9b995dcf69165b39b7dec0ddeefc9ed06b9980f 100644 (file)
@@ -28,6 +28,8 @@
 #include <asm/time.h> /* For perf_irq */
 
 #define MIPS_MAX_HWEVENTS 4
+#define MIPS_TCS_PER_COUNTER 2
+#define MIPS_CPUID_TO_COUNTER_MASK (MIPS_TCS_PER_COUNTER - 1)
 
 struct cpu_hw_events {
        /* Array of events on this cpu. */
@@ -108,13 +110,20 @@ static struct mips_pmu mipspmu;
 #define M_PERFCTL_INTERRUPT_ENABLE     (1      <<  4)
 #define M_PERFCTL_EVENT(event)         (((event) & 0x3ff)  << 5)
 #define M_PERFCTL_VPEID(vpe)           ((vpe)    << 16)
+
+#ifdef CONFIG_CPU_BMIPS5000
+#define M_PERFCTL_MT_EN(filter)                0
+#else /* !CONFIG_CPU_BMIPS5000 */
 #define M_PERFCTL_MT_EN(filter)                ((filter) << 20)
+#endif /* CONFIG_CPU_BMIPS5000 */
+
 #define    M_TC_EN_ALL                 M_PERFCTL_MT_EN(0)
 #define    M_TC_EN_VPE                 M_PERFCTL_MT_EN(1)
 #define    M_TC_EN_TC                  M_PERFCTL_MT_EN(2)
 #define M_PERFCTL_TCID(tcid)           ((tcid)   << 22)
 #define M_PERFCTL_WIDE                 (1      << 30)
 #define M_PERFCTL_MORE                 (1      << 31)
+#define M_PERFCTL_TC                   (1      << 30)
 
 #define M_PERFCTL_COUNT_EVENT_WHENEVER (M_PERFCTL_EXL |                \
                                        M_PERFCTL_KERNEL |              \
@@ -135,12 +144,17 @@ static int cpu_has_mipsmt_pertccounters;
 
 static DEFINE_RWLOCK(pmuint_rwlock);
 
+#if defined(CONFIG_CPU_BMIPS5000)
+#define vpe_id()       (cpu_has_mipsmt_pertccounters ? \
+                        0 : (smp_processor_id() & MIPS_CPUID_TO_COUNTER_MASK))
+#else
 /*
  * FIXME: For VSMP, vpe_id() is redefined for Perf-events, because
  * cpu_data[cpuid].vpe_id reports 0 for _both_ CPUs.
  */
 #define vpe_id()       (cpu_has_mipsmt_pertccounters ? \
-                       0 : smp_processor_id())
+                        0 : smp_processor_id())
+#endif
 
 /* Copied from op_model_mipsxx.c */
 static unsigned int vpe_shift(void)
@@ -334,6 +348,11 @@ static void mipsxx_pmu_enable_event(struct hw_perf_event *evt, int idx)
                (evt->config_base & M_PERFCTL_CONFIG_MASK) |
                /* Make sure interrupt enabled. */
                M_PERFCTL_INTERRUPT_ENABLE;
+       if (IS_ENABLED(CONFIG_CPU_BMIPS5000))
+               /* enable the counter for the calling thread */
+               cpuc->saved_ctrl[idx] |=
+                       (1 << (12 + vpe_id())) | M_PERFCTL_TC;
+
        /*
         * We do not actually let the counter run. Leave it until start().
         */
@@ -814,6 +833,13 @@ static const struct mips_perf_event octeon_event_map[PERF_COUNT_HW_MAX] = {
        [PERF_COUNT_HW_BUS_CYCLES] = { 0x25, CNTR_ALL },
 };
 
+static const struct mips_perf_event bmips5000_event_map
+                               [PERF_COUNT_HW_MAX] = {
+       [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, T },
+       [PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T },
+       [PERF_COUNT_HW_BRANCH_MISSES] = { 0x02, CNTR_ODD, T },
+};
+
 /* 24K/34K/1004K cores can share the same cache event map. */
 static const struct mips_perf_event mipsxxcore_cache_map
                                [PERF_COUNT_HW_CACHE_MAX]
@@ -966,6 +992,65 @@ static const struct mips_perf_event mipsxx74Kcore_cache_map
 },
 };
 
+/* BMIPS5000 */
+static const struct mips_perf_event bmips5000_cache_map
+                               [PERF_COUNT_HW_CACHE_MAX]
+                               [PERF_COUNT_HW_CACHE_OP_MAX]
+                               [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
+[C(L1D)] = {
+       /*
+        * Like some other architectures (e.g. ARM), the performance
+        * counters don't differentiate between read and write
+        * accesses/misses, so this isn't strictly correct, but it's the
+        * best we can do. Writes and reads get combined.
+        */
+       [C(OP_READ)] = {
+               [C(RESULT_ACCESS)]      = { 12, CNTR_EVEN, T },
+               [C(RESULT_MISS)]        = { 12, CNTR_ODD, T },
+       },
+       [C(OP_WRITE)] = {
+               [C(RESULT_ACCESS)]      = { 12, CNTR_EVEN, T },
+               [C(RESULT_MISS)]        = { 12, CNTR_ODD, T },
+       },
+},
+[C(L1I)] = {
+       [C(OP_READ)] = {
+               [C(RESULT_ACCESS)]      = { 10, CNTR_EVEN, T },
+               [C(RESULT_MISS)]        = { 10, CNTR_ODD, T },
+       },
+       [C(OP_WRITE)] = {
+               [C(RESULT_ACCESS)]      = { 10, CNTR_EVEN, T },
+               [C(RESULT_MISS)]        = { 10, CNTR_ODD, T },
+       },
+       [C(OP_PREFETCH)] = {
+               [C(RESULT_ACCESS)]      = { 23, CNTR_EVEN, T },
+               /*
+                * Note that MIPS has only "hit" events countable for
+                * the prefetch operation.
+                */
+       },
+},
+[C(LL)] = {
+       [C(OP_READ)] = {
+               [C(RESULT_ACCESS)]      = { 28, CNTR_EVEN, P },
+               [C(RESULT_MISS)]        = { 28, CNTR_ODD, P },
+       },
+       [C(OP_WRITE)] = {
+               [C(RESULT_ACCESS)]      = { 28, CNTR_EVEN, P },
+               [C(RESULT_MISS)]        = { 28, CNTR_ODD, P },
+       },
+},
+[C(BPU)] = {
+       /* Using the same code for *HW_BRANCH* */
+       [C(OP_READ)] = {
+               [C(RESULT_MISS)]        = { 0x02, CNTR_ODD, T },
+       },
+       [C(OP_WRITE)] = {
+               [C(RESULT_MISS)]        = { 0x02, CNTR_ODD, T },
+       },
+},
+};
+
 
 static const struct mips_perf_event octeon_cache_map
                                [PERF_COUNT_HW_CACHE_MAX]
@@ -1240,6 +1325,11 @@ static irqreturn_t mipsxx_pmu_handle_irq(int irq, void *dev)
 #define IS_RANGE_V_1004K_EVENT(r)      ((r) == 47)
 #endif
 
+/* BMIPS5000 */
+#define IS_BOTH_COUNTERS_BMIPS5000_EVENT(b)                            \
+       ((b) == 0 || (b) == 1)
+
+
 /*
  * User can use 0-255 raw events, where 0-127 for the events of even
  * counters, and 128-255 for odd counters. Note that bit 7 is used to
@@ -1310,6 +1400,12 @@ static const struct mips_perf_event *mipsxx_pmu_map_raw_event(u64 config)
                        raw_event.range = T;
 #endif
                break;
+       case CPU_BMIPS5000:
+               if (IS_BOTH_COUNTERS_BMIPS5000_EVENT(base_id))
+                       raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
+               else
+                       raw_event.cntr_mask =
+                               raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
        }
 
        return &raw_event;
@@ -1421,6 +1517,11 @@ init_hw_perf_events(void)
                mipspmu.cache_event_map = &octeon_cache_map;
                mipspmu.map_raw_event = octeon_pmu_map_raw_event;
                break;
+       case CPU_BMIPS5000:
+               mipspmu.name = "BMIPS5000";
+               mipspmu.general_event_map = &bmips5000_event_map;
+               mipspmu.cache_event_map = &bmips5000_cache_map;
+               break;
        default:
                pr_cont("Either hardware does not support performance "
                        "counters, or not yet implemented.\n");