drivers/gpu/drm/i915/gt/selftest_engine_heartbeat.c

   1 // SPDX-License-Identifier: MIT
   2 /*
   3  * Copyright © 2018 Intel Corporation
   4  */
   5
   6 #include <linux/sort.h>
   7
   8 #include "i915_drv.h"
   9
  10 #include "intel_gt_requests.h"
  11 #include "i915_selftest.h"
  12 #include "selftest_engine_heartbeat.h"
  13
  14 static void reset_heartbeat(struct intel_engine_cs *engine)
  15 {
  16         intel_engine_set_heartbeat(engine,
  17                                    engine->defaults.heartbeat_interval_ms);
  18 }
  19
  20 static int timeline_sync(struct intel_timeline *tl)
  21 {
  22         struct dma_fence *fence;
  23         long timeout;
  24
  25         fence = i915_active_fence_get(&tl->last_request);
  26         if (!fence)
  27                 return 0;
  28
  29         timeout = dma_fence_wait_timeout(fence, true, HZ / 2);
  30         dma_fence_put(fence);
  31         if (timeout < 0)
  32                 return timeout;
  33
  34         return 0;
  35 }
  36
  37 static int engine_sync_barrier(struct intel_engine_cs *engine)
  38 {
  39         return timeline_sync(engine->kernel_context->timeline);
  40 }
  41
  42 struct pulse {
  43         struct i915_active active;
  44         struct kref kref;
  45 };
  46
  47 static int pulse_active(struct i915_active *active)
  48 {
  49         kref_get(&container_of(active, struct pulse, active)->kref);
  50         return 0;
  51 }
  52
  53 static void pulse_free(struct kref *kref)
  54 {
  55         struct pulse *p = container_of(kref, typeof(*p), kref);
  56
  57         i915_active_fini(&p->active);
  58         kfree(p);
  59 }
  60
  61 static void pulse_put(struct pulse *p)
  62 {
  63         kref_put(&p->kref, pulse_free);
  64 }
  65
  66 static void pulse_retire(struct i915_active *active)
  67 {
  68         pulse_put(container_of(active, struct pulse, active));
  69 }
  70
  71 static struct pulse *pulse_create(void)
  72 {
  73         struct pulse *p;
  74
  75         p = kmalloc(sizeof(*p), GFP_KERNEL);
  76         if (!p)
  77                 return p;
  78
  79         kref_init(&p->kref);
  80         i915_active_init(&p->active, pulse_active, pulse_retire, 0);
  81
  82         return p;
  83 }
  84
  85 static void pulse_unlock_wait(struct pulse *p)
  86 {
  87         wait_var_event_timeout(&p->active, i915_active_is_idle(&p->active), HZ);
  88 }
  89
  90 static int __live_idle_pulse(struct intel_engine_cs *engine,
  91                              int (*fn)(struct intel_engine_cs *cs))
  92 {
  93         struct pulse *p;
  94         int err;
  95
  96         GEM_BUG_ON(!intel_engine_pm_is_awake(engine));
  97
  98         p = pulse_create();
  99         if (!p)
 100                 return -ENOMEM;
 101
 102         err = i915_active_acquire(&p->active);
 103         if (err)
 104                 goto out;
 105
 106         err = i915_active_acquire_preallocate_barrier(&p->active, engine);
 107         if (err) {
 108                 i915_active_release(&p->active);
 109                 goto out;
 110         }
 111
 112         i915_active_acquire_barrier(&p->active);
 113         i915_active_release(&p->active);
 114
 115         GEM_BUG_ON(i915_active_is_idle(&p->active));
 116         GEM_BUG_ON(llist_empty(&engine->barrier_tasks));
 117
 118         err = fn(engine);
 119         if (err)
 120                 goto out;
 121
 122         GEM_BUG_ON(!llist_empty(&engine->barrier_tasks));
 123
 124         if (engine_sync_barrier(engine)) {
 125                 struct drm_printer m = drm_err_printer(&engine->i915->drm, "pulse");
 126
 127                 drm_printf(&m, "%s: no heartbeat pulse?\n", engine->name);
 128                 intel_engine_dump(engine, &m, "%s", engine->name);
 129
 130                 err = -ETIME;
 131                 goto out;
 132         }
 133
 134         GEM_BUG_ON(READ_ONCE(engine->serial) != engine->wakeref_serial);
 135
 136         pulse_unlock_wait(p); /* synchronize with the retirement callback */
 137
 138         if (!i915_active_is_idle(&p->active)) {
 139                 struct drm_printer m = drm_err_printer(&engine->i915->drm, "pulse");
 140
 141                 drm_printf(&m, "%s: heartbeat pulse did not flush idle tasks\n",
 142                            engine->name);
 143                 i915_active_print(&p->active, &m);
 144
 145                 err = -EINVAL;
 146                 goto out;
 147         }
 148
 149 out:
 150         pulse_put(p);
 151         return err;
 152 }
 153
 154 static int live_idle_flush(void *arg)
 155 {
 156         struct intel_gt *gt = arg;
 157         struct intel_engine_cs *engine;
 158         enum intel_engine_id id;
 159         int err = 0;
 160
 161         /* Check that we can flush the idle barriers */
 162
 163         for_each_engine(engine, gt, id) {
 164                 st_engine_heartbeat_disable(engine);
 165                 err = __live_idle_pulse(engine, intel_engine_flush_barriers);
 166                 st_engine_heartbeat_enable(engine);
 167                 if (err)
 168                         break;
 169         }
 170
 171         return err;
 172 }
 173
 174 static int live_idle_pulse(void *arg)
 175 {
 176         struct intel_gt *gt = arg;
 177         struct intel_engine_cs *engine;
 178         enum intel_engine_id id;
 179         int err = 0;
 180
 181         /* Check that heartbeat pulses flush the idle barriers */
 182
 183         for_each_engine(engine, gt, id) {
 184                 st_engine_heartbeat_disable(engine);
 185                 err = __live_idle_pulse(engine, intel_engine_pulse);
 186                 st_engine_heartbeat_enable(engine);
 187                 if (err && err != -ENODEV)
 188                         break;
 189
 190                 err = 0;
 191         }
 192
 193         return err;
 194 }
 195
 196 static int __live_heartbeat_off(struct intel_engine_cs *engine)
 197 {
 198         int err;
 199
 200         intel_engine_pm_get(engine);
 201
 202         engine->serial++;
 203         flush_delayed_work(&engine->heartbeat.work);
 204         if (!delayed_work_pending(&engine->heartbeat.work)) {
 205                 pr_err("%s: heartbeat not running\n",
 206                        engine->name);
 207                 err = -EINVAL;
 208                 goto err_pm;
 209         }
 210
 211         err = intel_engine_set_heartbeat(engine, 0);
 212         if (err)
 213                 goto err_pm;
 214
 215         engine->serial++;
 216         flush_delayed_work(&engine->heartbeat.work);
 217         if (delayed_work_pending(&engine->heartbeat.work)) {
 218                 pr_err("%s: heartbeat still running\n",
 219                        engine->name);
 220                 err = -EINVAL;
 221                 goto err_beat;
 222         }
 223
 224         if (READ_ONCE(engine->heartbeat.systole)) {
 225                 pr_err("%s: heartbeat still allocated\n",
 226                        engine->name);
 227                 err = -EINVAL;
 228                 goto err_beat;
 229         }
 230
 231 err_beat:
 232         reset_heartbeat(engine);
 233 err_pm:
 234         intel_engine_pm_put(engine);
 235         return err;
 236 }
 237
 238 static int live_heartbeat_off(void *arg)
 239 {
 240         struct intel_gt *gt = arg;
 241         struct intel_engine_cs *engine;
 242         enum intel_engine_id id;
 243         int err = 0;
 244
 245         /* Check that we can turn off heartbeat and not interrupt VIP */
 246         if (!CONFIG_DRM_I915_HEARTBEAT_INTERVAL)
 247                 return 0;
 248
 249         for_each_engine(engine, gt, id) {
 250                 if (!intel_engine_has_preemption(engine))
 251                         continue;
 252
 253                 err = __live_heartbeat_off(engine);
 254                 if (err)
 255                         break;
 256         }
 257
 258         return err;
 259 }
 260
 261 int intel_heartbeat_live_selftests(struct drm_i915_private *i915)
 262 {
 263         static const struct i915_subtest tests[] = {
 264                 SUBTEST(live_idle_flush),
 265                 SUBTEST(live_idle_pulse),
 266                 SUBTEST(live_heartbeat_off),
 267         };
 268         int saved_hangcheck;
 269         int err;
 270
 271         if (intel_gt_is_wedged(to_gt(i915)))
 272                 return 0;
 273
 274         saved_hangcheck = i915->params.enable_hangcheck;
 275         i915->params.enable_hangcheck = INT_MAX;
 276
 277         err = intel_gt_live_subtests(tests, to_gt(i915));
 278
 279         i915->params.enable_hangcheck = saved_hangcheck;
 280         return err;
 281 }
 282
 283 void st_engine_heartbeat_disable(struct intel_engine_cs *engine)
 284 {
 285         engine->props.heartbeat_interval_ms = 0;
 286
 287         intel_engine_pm_get(engine);
 288         intel_engine_park_heartbeat(engine);
 289 }
 290
 291 void st_engine_heartbeat_enable(struct intel_engine_cs *engine)
 292 {
 293         intel_engine_pm_put(engine);
 294
 295         engine->props.heartbeat_interval_ms =
 296                 engine->defaults.heartbeat_interval_ms;
 297 }
 298
 299 void st_engine_heartbeat_disable_no_pm(struct intel_engine_cs *engine)
 300 {
 301         engine->props.heartbeat_interval_ms = 0;
 302
 303         /*
 304          * Park the heartbeat but without holding the PM lock as that
 305          * makes the engines appear not-idle. Note that if/when unpark
 306          * is called due to the PM lock being acquired later the
 307          * heartbeat still won't be enabled because of the above = 0.
 308          */
 309         if (intel_engine_pm_get_if_awake(engine)) {
 310                 intel_engine_park_heartbeat(engine);
 311                 intel_engine_pm_put(engine);
 312         }
 313 }
 314
 315 void st_engine_heartbeat_enable_no_pm(struct intel_engine_cs *engine)
 316 {
 317         engine->props.heartbeat_interval_ms =
 318                 engine->defaults.heartbeat_interval_ms;
 319 }