root/drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c

DEFINITIONS

1. amdgpu_perf_event_init
2. amdgpu_perf_start
3. amdgpu_perf_read
4. amdgpu_perf_stop
5. amdgpu_perf_add
6. amdgpu_perf_del
7. init_pmu_by_type
8. amdgpu_pmu_init
9. amdgpu_pmu_fini

   1 /*
   2  * Copyright 2019 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  *
  22  * Author: Jonathan Kim <jonathan.kim@amd.com>
  23  *
  24  */
  25 
  26 #include <linux/perf_event.h>
  27 #include <linux/init.h>
  28 #include "amdgpu.h"
  29 #include "amdgpu_pmu.h"
  30 #include "df_v3_6.h"
  31 
  32 #define PMU_NAME_SIZE 32
  33 
  34 /* record to keep track of pmu entry per pmu type per device */
  35 struct amdgpu_pmu_entry {
  36         struct list_head entry;
  37         struct amdgpu_device *adev;
  38         struct pmu pmu;
  39         unsigned int pmu_perf_type;
  40 };
  41 
  42 static LIST_HEAD(amdgpu_pmu_list);
  43 
  44 
  45 /* initialize perf counter */
  46 static int amdgpu_perf_event_init(struct perf_event *event)
  47 {
  48         struct hw_perf_event *hwc = &event->hw;
  49 
  50         /* test the event attr type check for PMU enumeration */
  51         if (event->attr.type != event->pmu->type)
  52                 return -ENOENT;
  53 
  54         /* update the hw_perf_event struct with config data */
  55         hwc->conf = event->attr.config;
  56 
  57         return 0;
  58 }
  59 
  60 /* start perf counter */
  61 static void amdgpu_perf_start(struct perf_event *event, int flags)
  62 {
  63         struct hw_perf_event *hwc = &event->hw;
  64         struct amdgpu_pmu_entry *pe = container_of(event->pmu,
  65                                                   struct amdgpu_pmu_entry,
  66                                                   pmu);
  67 
  68         if (WARN_ON_ONCE(!(hwc->state & PERF_HES_STOPPED)))
  69                 return;
  70 
  71         WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
  72         hwc->state = 0;
  73 
  74         switch (pe->pmu_perf_type) {
  75         case PERF_TYPE_AMDGPU_DF:
  76                 if (!(flags & PERF_EF_RELOAD))
  77                         pe->adev->df_funcs->pmc_start(pe->adev, hwc->conf, 1);
  78 
  79                 pe->adev->df_funcs->pmc_start(pe->adev, hwc->conf, 0);
  80                 break;
  81         default:
  82                 break;
  83         }
  84 
  85         perf_event_update_userpage(event);
  86 
  87 }
  88 
  89 /* read perf counter */
  90 static void amdgpu_perf_read(struct perf_event *event)
  91 {
  92         struct hw_perf_event *hwc = &event->hw;
  93         struct amdgpu_pmu_entry *pe = container_of(event->pmu,
  94                                                   struct amdgpu_pmu_entry,
  95                                                   pmu);
  96 
  97         u64 count, prev;
  98 
  99         do {
 100                 prev = local64_read(&hwc->prev_count);
 101 
 102                 switch (pe->pmu_perf_type) {
 103                 case PERF_TYPE_AMDGPU_DF:
 104                         pe->adev->df_funcs->pmc_get_count(pe->adev, hwc->conf,
 105                                                           &count);
 106                         break;
 107                 default:
 108                         count = 0;
 109                         break;
 110                 };
 111         } while (local64_cmpxchg(&hwc->prev_count, prev, count) != prev);
 112 
 113         local64_add(count - prev, &event->count);
 114 }
 115 
 116 /* stop perf counter */
 117 static void amdgpu_perf_stop(struct perf_event *event, int flags)
 118 {
 119         struct hw_perf_event *hwc = &event->hw;
 120         struct amdgpu_pmu_entry *pe = container_of(event->pmu,
 121                                                   struct amdgpu_pmu_entry,
 122                                                   pmu);
 123 
 124         if (hwc->state & PERF_HES_UPTODATE)
 125                 return;
 126 
 127         switch (pe->pmu_perf_type) {
 128         case PERF_TYPE_AMDGPU_DF:
 129                 pe->adev->df_funcs->pmc_stop(pe->adev, hwc->conf, 0);
 130                 break;
 131         default:
 132                 break;
 133         };
 134 
 135         WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
 136         hwc->state |= PERF_HES_STOPPED;
 137 
 138         if (hwc->state & PERF_HES_UPTODATE)
 139                 return;
 140 
 141         amdgpu_perf_read(event);
 142         hwc->state |= PERF_HES_UPTODATE;
 143 }
 144 
 145 /* add perf counter  */
 146 static int amdgpu_perf_add(struct perf_event *event, int flags)
 147 {
 148         struct hw_perf_event *hwc = &event->hw;
 149         int retval;
 150 
 151         struct amdgpu_pmu_entry *pe = container_of(event->pmu,
 152                                                   struct amdgpu_pmu_entry,
 153                                                   pmu);
 154 
 155         event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
 156 
 157         switch (pe->pmu_perf_type) {
 158         case PERF_TYPE_AMDGPU_DF:
 159                 retval = pe->adev->df_funcs->pmc_start(pe->adev, hwc->conf, 1);
 160                 break;
 161         default:
 162                 return 0;
 163         };
 164 
 165         if (retval)
 166                 return retval;
 167 
 168         if (flags & PERF_EF_START)
 169                 amdgpu_perf_start(event, PERF_EF_RELOAD);
 170 
 171         return retval;
 172 
 173 }
 174 
 175 /* delete perf counter  */
 176 static void amdgpu_perf_del(struct perf_event *event, int flags)
 177 {
 178         struct hw_perf_event *hwc = &event->hw;
 179         struct amdgpu_pmu_entry *pe = container_of(event->pmu,
 180                                                   struct amdgpu_pmu_entry,
 181                                                   pmu);
 182 
 183         amdgpu_perf_stop(event, PERF_EF_UPDATE);
 184 
 185         switch (pe->pmu_perf_type) {
 186         case PERF_TYPE_AMDGPU_DF:
 187                 pe->adev->df_funcs->pmc_stop(pe->adev, hwc->conf, 1);
 188                 break;
 189         default:
 190                 break;
 191         };
 192 
 193         perf_event_update_userpage(event);
 194 }
 195 
 196 /* vega20 pmus */
 197 
 198 /* init pmu tracking per pmu type */
 199 static int init_pmu_by_type(struct amdgpu_device *adev,
 200                   const struct attribute_group *attr_groups[],
 201                   char *pmu_type_name, char *pmu_file_prefix,
 202                   unsigned int pmu_perf_type,
 203                   unsigned int num_counters)
 204 {
 205         char pmu_name[PMU_NAME_SIZE];
 206         struct amdgpu_pmu_entry *pmu_entry;
 207         int ret = 0;
 208 
 209         pmu_entry = kzalloc(sizeof(struct amdgpu_pmu_entry), GFP_KERNEL);
 210 
 211         if (!pmu_entry)
 212                 return -ENOMEM;
 213 
 214         pmu_entry->adev = adev;
 215         pmu_entry->pmu = (struct pmu){
 216                 .event_init = amdgpu_perf_event_init,
 217                 .add = amdgpu_perf_add,
 218                 .del = amdgpu_perf_del,
 219                 .start = amdgpu_perf_start,
 220                 .stop = amdgpu_perf_stop,
 221                 .read = amdgpu_perf_read,
 222                 .task_ctx_nr = perf_invalid_context,
 223         };
 224 
 225         pmu_entry->pmu.attr_groups = attr_groups;
 226         pmu_entry->pmu_perf_type = pmu_perf_type;
 227         snprintf(pmu_name, PMU_NAME_SIZE, "%s_%d",
 228                                 pmu_file_prefix, adev->ddev->primary->index);
 229 
 230         ret = perf_pmu_register(&pmu_entry->pmu, pmu_name, -1);
 231 
 232         if (ret) {
 233                 kfree(pmu_entry);
 234                 pr_warn("Error initializing AMDGPU %s PMUs.\n", pmu_type_name);
 235                 return ret;
 236         }
 237 
 238         pr_info("Detected AMDGPU %s Counters. # of Counters = %d.\n",
 239                         pmu_type_name, num_counters);
 240 
 241         list_add_tail(&pmu_entry->entry, &amdgpu_pmu_list);
 242 
 243         return 0;
 244 }
 245 
 246 /* init amdgpu_pmu */
 247 int amdgpu_pmu_init(struct amdgpu_device *adev)
 248 {
 249         int ret = 0;
 250 
 251         switch (adev->asic_type) {
 252         case CHIP_VEGA20:
 253                 /* init df */
 254                 ret = init_pmu_by_type(adev, df_v3_6_attr_groups,
 255                                        "DF", "amdgpu_df", PERF_TYPE_AMDGPU_DF,
 256                                        DF_V3_6_MAX_COUNTERS);
 257 
 258                 /* other pmu types go here*/
 259                 break;
 260         default:
 261                 return 0;
 262         }
 263 
 264         return 0;
 265 }
 266 
 267 
 268 /* destroy all pmu data associated with target device */
 269 void amdgpu_pmu_fini(struct amdgpu_device *adev)
 270 {
 271         struct amdgpu_pmu_entry *pe, *temp;
 272 
 273         list_for_each_entry_safe(pe, temp, &amdgpu_pmu_list, entry) {
 274                 if (pe->adev == adev) {
 275                         list_del(&pe->entry);
 276                         perf_pmu_unregister(&pe->pmu);
 277                         kfree(pe);
 278                 }
 279         }
 280 }