arch/powerpc/perf/e500-pmu.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Performance counter support for e500 family processors.
   4  *
   5  * Copyright 2008-2009 Paul Mackerras, IBM Corporation.
   6  * Copyright 2010 Freescale Semiconductor, Inc.
   7  */
   8 #include <linux/string.h>
   9 #include <linux/perf_event.h>
  10 #include <asm/reg.h>
  11 #include <asm/cputable.h>
  12
  13 /*
  14  * Map of generic hardware event types to hardware events
  15  * Zero if unsupported
  16  */
  17 static int e500_generic_events[] = {
  18         [PERF_COUNT_HW_CPU_CYCLES] = 1,
  19         [PERF_COUNT_HW_INSTRUCTIONS] = 2,
  20         [PERF_COUNT_HW_CACHE_MISSES] = 41, /* Data L1 cache reloads */
  21         [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 12,
  22         [PERF_COUNT_HW_BRANCH_MISSES] = 15,
  23         [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 18,
  24         [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 19,
  25 };
  26
  27 #define C(x)    PERF_COUNT_HW_CACHE_##x
  28
  29 /*
  30  * Table of generalized cache-related events.
  31  * 0 means not supported, -1 means nonsensical, other values
  32  * are event codes.
  33  */
  34 static int e500_cache_events[C(MAX)][C(OP_MAX)][C(RESULT_MAX)] = {
  35         /*
  36          * D-cache misses are not split into read/write/prefetch;
  37          * use raw event 41.
  38          */
  39         [C(L1D)] = {            /*      RESULT_ACCESS   RESULT_MISS */
  40                 [C(OP_READ)] = {        27,             0       },
  41                 [C(OP_WRITE)] = {       28,             0       },
  42                 [C(OP_PREFETCH)] = {    29,             0       },
  43         },
  44         [C(L1I)] = {            /*      RESULT_ACCESS   RESULT_MISS */
  45                 [C(OP_READ)] = {        2,              60      },
  46                 [C(OP_WRITE)] = {       -1,             -1      },
  47                 [C(OP_PREFETCH)] = {    0,              0       },
  48         },
  49         /*
  50          * Assuming LL means L2, it's not a good match for this model.
  51          * It allocates only on L1 castout or explicit prefetch, and
  52          * does not have separate read/write events (but it does have
  53          * separate instruction/data events).
  54          */
  55         [C(LL)] = {             /*      RESULT_ACCESS   RESULT_MISS */
  56                 [C(OP_READ)] = {        0,              0       },
  57                 [C(OP_WRITE)] = {       0,              0       },
  58                 [C(OP_PREFETCH)] = {    0,              0       },
  59         },
  60         /*
  61          * There are data/instruction MMU misses, but that's a miss on
  62          * the chip's internal level-one TLB which is probably not
  63          * what the user wants.  Instead, unified level-two TLB misses
  64          * are reported here.
  65          */
  66         [C(DTLB)] = {           /*      RESULT_ACCESS   RESULT_MISS */
  67                 [C(OP_READ)] = {        26,             66      },
  68                 [C(OP_WRITE)] = {       -1,             -1      },
  69                 [C(OP_PREFETCH)] = {    -1,             -1      },
  70         },
  71         [C(BPU)] = {            /*      RESULT_ACCESS   RESULT_MISS */
  72                 [C(OP_READ)] = {        12,             15      },
  73                 [C(OP_WRITE)] = {       -1,             -1      },
  74                 [C(OP_PREFETCH)] = {    -1,             -1      },
  75         },
  76         [C(NODE)] = {           /*      RESULT_ACCESS   RESULT_MISS */
  77                 [C(OP_READ)] = {        -1,             -1      },
  78                 [C(OP_WRITE)] = {       -1,             -1      },
  79                 [C(OP_PREFETCH)] = {    -1,             -1      },
  80         },
  81 };
  82
  83 static int num_events = 128;
  84
  85 /* Upper half of event id is PMLCb, for threshold events */
  86 static u64 e500_xlate_event(u64 event_id)
  87 {
  88         u32 event_low = (u32)event_id;
  89         u64 ret;
  90
  91         if (event_low >= num_events)
  92                 return 0;
  93
  94         ret = FSL_EMB_EVENT_VALID;
  95
  96         if (event_low >= 76 && event_low <= 81) {
  97                 ret |= FSL_EMB_EVENT_RESTRICTED;
  98                 ret |= event_id &
  99                        (FSL_EMB_EVENT_THRESHMUL | FSL_EMB_EVENT_THRESH);
 100         } else if (event_id &
 101                    (FSL_EMB_EVENT_THRESHMUL | FSL_EMB_EVENT_THRESH)) {
 102                 /* Threshold requested on non-threshold event */
 103                 return 0;
 104         }
 105
 106         return ret;
 107 }
 108
 109 static struct fsl_emb_pmu e500_pmu = {
 110         .name                   = "e500 family",
 111         .n_counter              = 4,
 112         .n_restricted           = 2,
 113         .xlate_event            = e500_xlate_event,
 114         .n_generic              = ARRAY_SIZE(e500_generic_events),
 115         .generic_events         = e500_generic_events,
 116         .cache_events           = &e500_cache_events,
 117 };
 118
 119 static int init_e500_pmu(void)
 120 {
 121         if (!cur_cpu_spec->oprofile_cpu_type)
 122                 return -ENODEV;
 123
 124         if (!strcmp(cur_cpu_spec->oprofile_cpu_type, "ppc/e500mc"))
 125                 num_events = 256;
 126         else if (strcmp(cur_cpu_spec->oprofile_cpu_type, "ppc/e500"))
 127                 return -ENODEV;
 128
 129         return register_fsl_emb_pmu(&e500_pmu);
 130 }
 131
 132 early_initcall(init_e500_pmu);