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

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright © 2019 Intel Corporation
   4  */
   5
   6 #include "i915_selftest.h"
   7 #include "intel_engine_heartbeat.h"
   8 #include "intel_engine_pm.h"
   9 #include "intel_gt.h"
  10
  11 #include "gem/selftests/mock_context.h"
  12 #include "selftests/igt_flush_test.h"
  13 #include "selftests/mock_drm.h"
  14
  15 static int request_sync(struct i915_request *rq)
  16 {
  17         struct intel_timeline *tl = i915_request_timeline(rq);
  18         long timeout;
  19         int err = 0;
  20
  21         intel_timeline_get(tl);
  22         i915_request_get(rq);
  23
  24         /* Opencode i915_request_add() so we can keep the timeline locked. */
  25         __i915_request_commit(rq);
  26         rq->sched.attr.priority = I915_PRIORITY_BARRIER;
  27         __i915_request_queue_bh(rq);
  28
  29         timeout = i915_request_wait(rq, 0, HZ / 10);
  30         if (timeout < 0)
  31                 err = timeout;
  32         else
  33                 i915_request_retire_upto(rq);
  34
  35         lockdep_unpin_lock(&tl->mutex, rq->cookie);
  36         mutex_unlock(&tl->mutex);
  37
  38         i915_request_put(rq);
  39         intel_timeline_put(tl);
  40
  41         return err;
  42 }
  43
  44 static int context_sync(struct intel_context *ce)
  45 {
  46         struct intel_timeline *tl = ce->timeline;
  47         int err = 0;
  48
  49         mutex_lock(&tl->mutex);
  50         do {
  51                 struct i915_request *rq;
  52                 long timeout;
  53
  54                 if (list_empty(&tl->requests))
  55                         break;
  56
  57                 rq = list_last_entry(&tl->requests, typeof(*rq), link);
  58                 i915_request_get(rq);
  59
  60                 timeout = i915_request_wait(rq, 0, HZ / 10);
  61                 if (timeout < 0)
  62                         err = timeout;
  63                 else
  64                         i915_request_retire_upto(rq);
  65
  66                 i915_request_put(rq);
  67         } while (!err);
  68         mutex_unlock(&tl->mutex);
  69
  70         /* Wait for all barriers to complete (remote CPU) before we check */
  71         i915_active_unlock_wait(&ce->active);
  72         return err;
  73 }
  74
  75 static int __live_context_size(struct intel_engine_cs *engine)
  76 {
  77         struct intel_context *ce;
  78         struct i915_request *rq;
  79         void *vaddr;
  80         int err;
  81
  82         ce = intel_context_create(engine);
  83         if (IS_ERR(ce))
  84                 return PTR_ERR(ce);
  85
  86         err = intel_context_pin(ce);
  87         if (err)
  88                 goto err;
  89
  90         vaddr = i915_gem_object_pin_map_unlocked(ce->state->obj,
  91                                                  i915_coherent_map_type(engine->i915,
  92                                                                         ce->state->obj, false));
  93         if (IS_ERR(vaddr)) {
  94                 err = PTR_ERR(vaddr);
  95                 intel_context_unpin(ce);
  96                 goto err;
  97         }
  98
  99         /*
 100          * Note that execlists also applies a redzone which it checks on
 101          * context unpin when debugging. We are using the same location
 102          * and same poison value so that our checks overlap. Despite the
 103          * redundancy, we want to keep this little selftest so that we
 104          * get coverage of any and all submission backends, and we can
 105          * always extend this test to ensure we trick the HW into a
 106          * compromising position wrt to the various sections that need
 107          * to be written into the context state.
 108          *
 109          * TLDR; this overlaps with the execlists redzone.
 110          */
 111         vaddr += engine->context_size - I915_GTT_PAGE_SIZE;
 112         memset(vaddr, POISON_INUSE, I915_GTT_PAGE_SIZE);
 113
 114         rq = intel_context_create_request(ce);
 115         intel_context_unpin(ce);
 116         if (IS_ERR(rq)) {
 117                 err = PTR_ERR(rq);
 118                 goto err_unpin;
 119         }
 120
 121         err = request_sync(rq);
 122         if (err)
 123                 goto err_unpin;
 124
 125         /* Force the context switch */
 126         rq = intel_engine_create_kernel_request(engine);
 127         if (IS_ERR(rq)) {
 128                 err = PTR_ERR(rq);
 129                 goto err_unpin;
 130         }
 131         err = request_sync(rq);
 132         if (err)
 133                 goto err_unpin;
 134
 135         if (memchr_inv(vaddr, POISON_INUSE, I915_GTT_PAGE_SIZE)) {
 136                 pr_err("%s context overwrote trailing red-zone!", engine->name);
 137                 err = -EINVAL;
 138         }
 139
 140 err_unpin:
 141         i915_gem_object_unpin_map(ce->state->obj);
 142 err:
 143         intel_context_put(ce);
 144         return err;
 145 }
 146
 147 static int live_context_size(void *arg)
 148 {
 149         struct intel_gt *gt = arg;
 150         struct intel_engine_cs *engine;
 151         enum intel_engine_id id;
 152         int err = 0;
 153
 154         /*
 155          * Check that our context sizes are correct by seeing if the
 156          * HW tries to write past the end of one.
 157          */
 158
 159         for_each_engine(engine, gt, id) {
 160                 struct file *saved;
 161
 162                 if (!engine->context_size)
 163                         continue;
 164
 165                 intel_engine_pm_get(engine);
 166
 167                 /*
 168                  * Hide the old default state -- we lie about the context size
 169                  * and get confused when the default state is smaller than
 170                  * expected. For our do nothing request, inheriting the
 171                  * active state is sufficient, we are only checking that we
 172                  * don't use more than we planned.
 173                  */
 174                 saved = fetch_and_zero(&engine->default_state);
 175
 176                 /* Overlaps with the execlists redzone */
 177                 engine->context_size += I915_GTT_PAGE_SIZE;
 178
 179                 err = __live_context_size(engine);
 180
 181                 engine->context_size -= I915_GTT_PAGE_SIZE;
 182
 183                 engine->default_state = saved;
 184
 185                 intel_engine_pm_put(engine);
 186
 187                 if (err)
 188                         break;
 189         }
 190
 191         return err;
 192 }
 193
 194 static int __live_active_context(struct intel_engine_cs *engine)
 195 {
 196         unsigned long saved_heartbeat;
 197         struct intel_context *ce;
 198         int pass;
 199         int err;
 200
 201         /*
 202          * We keep active contexts alive until after a subsequent context
 203          * switch as the final write from the context-save will be after
 204          * we retire the final request. We track when we unpin the context,
 205          * under the presumption that the final pin is from the last request,
 206          * and instead of immediately unpinning the context, we add a task
 207          * to unpin the context from the next idle-barrier.
 208          *
 209          * This test makes sure that the context is kept alive until a
 210          * subsequent idle-barrier (emitted when the engine wakeref hits 0
 211          * with no more outstanding requests).
 212          *
 213          * In GuC submission mode we don't use idle barriers and we instead
 214          * get a message from the GuC to signal that it is safe to unpin the
 215          * context from memory.
 216          */
 217         if (intel_engine_uses_guc(engine))
 218                 return 0;
 219
 220         if (intel_engine_pm_is_awake(engine)) {
 221                 pr_err("%s is awake before starting %s!\n",
 222                        engine->name, __func__);
 223                 return -EINVAL;
 224         }
 225
 226         ce = intel_context_create(engine);
 227         if (IS_ERR(ce))
 228                 return PTR_ERR(ce);
 229
 230         saved_heartbeat = engine->props.heartbeat_interval_ms;
 231         engine->props.heartbeat_interval_ms = 0;
 232
 233         for (pass = 0; pass <= 2; pass++) {
 234                 struct i915_request *rq;
 235
 236                 intel_engine_pm_get(engine);
 237
 238                 rq = intel_context_create_request(ce);
 239                 if (IS_ERR(rq)) {
 240                         err = PTR_ERR(rq);
 241                         goto out_engine;
 242                 }
 243
 244                 err = request_sync(rq);
 245                 if (err)
 246                         goto out_engine;
 247
 248                 /* Context will be kept active until after an idle-barrier. */
 249                 if (i915_active_is_idle(&ce->active)) {
 250                         pr_err("context is not active; expected idle-barrier (%s pass %d)\n",
 251                                engine->name, pass);
 252                         err = -EINVAL;
 253                         goto out_engine;
 254                 }
 255
 256                 if (!intel_engine_pm_is_awake(engine)) {
 257                         pr_err("%s is asleep before idle-barrier\n",
 258                                engine->name);
 259                         err = -EINVAL;
 260                         goto out_engine;
 261                 }
 262
 263 out_engine:
 264                 intel_engine_pm_put(engine);
 265                 if (err)
 266                         goto err;
 267         }
 268
 269         /* Now make sure our idle-barriers are flushed */
 270         err = intel_engine_flush_barriers(engine);
 271         if (err)
 272                 goto err;
 273
 274         /* Wait for the barrier and in the process wait for engine to park */
 275         err = context_sync(engine->kernel_context);
 276         if (err)
 277                 goto err;
 278
 279         if (!i915_active_is_idle(&ce->active)) {
 280                 pr_err("context is still active!");
 281                 err = -EINVAL;
 282         }
 283
 284         intel_engine_pm_flush(engine);
 285
 286         if (intel_engine_pm_is_awake(engine)) {
 287                 struct drm_printer p = drm_debug_printer(__func__);
 288
 289                 intel_engine_dump(engine, &p,
 290                                   "%s is still awake:%d after idle-barriers\n",
 291                                   engine->name,
 292                                   atomic_read(&engine->wakeref.count));
 293                 GEM_TRACE_DUMP();
 294
 295                 err = -EINVAL;
 296                 goto err;
 297         }
 298
 299 err:
 300         engine->props.heartbeat_interval_ms = saved_heartbeat;
 301         intel_context_put(ce);
 302         return err;
 303 }
 304
 305 static int live_active_context(void *arg)
 306 {
 307         struct intel_gt *gt = arg;
 308         struct intel_engine_cs *engine;
 309         enum intel_engine_id id;
 310         int err = 0;
 311
 312         for_each_engine(engine, gt, id) {
 313                 err = __live_active_context(engine);
 314                 if (err)
 315                         break;
 316
 317                 err = igt_flush_test(gt->i915);
 318                 if (err)
 319                         break;
 320         }
 321
 322         return err;
 323 }
 324
 325 static int __remote_sync(struct intel_context *ce, struct intel_context *remote)
 326 {
 327         struct i915_request *rq;
 328         int err;
 329
 330         err = intel_context_pin(remote);
 331         if (err)
 332                 return err;
 333
 334         rq = intel_context_create_request(ce);
 335         if (IS_ERR(rq)) {
 336                 err = PTR_ERR(rq);
 337                 goto unpin;
 338         }
 339
 340         err = intel_context_prepare_remote_request(remote, rq);
 341         if (err) {
 342                 i915_request_add(rq);
 343                 goto unpin;
 344         }
 345
 346         err = request_sync(rq);
 347
 348 unpin:
 349         intel_context_unpin(remote);
 350         return err;
 351 }
 352
 353 static int __live_remote_context(struct intel_engine_cs *engine)
 354 {
 355         struct intel_context *local, *remote;
 356         unsigned long saved_heartbeat;
 357         int pass;
 358         int err;
 359
 360         /*
 361          * Check that our idle barriers do not interfere with normal
 362          * activity tracking. In particular, check that operating
 363          * on the context image remotely (intel_context_prepare_remote_request),
 364          * which inserts foreign fences into intel_context.active, does not
 365          * clobber the idle-barrier.
 366          *
 367          * In GuC submission mode we don't use idle barriers.
 368          */
 369         if (intel_engine_uses_guc(engine))
 370                 return 0;
 371
 372         if (intel_engine_pm_is_awake(engine)) {
 373                 pr_err("%s is awake before starting %s!\n",
 374                        engine->name, __func__);
 375                 return -EINVAL;
 376         }
 377
 378         remote = intel_context_create(engine);
 379         if (IS_ERR(remote))
 380                 return PTR_ERR(remote);
 381
 382         local = intel_context_create(engine);
 383         if (IS_ERR(local)) {
 384                 err = PTR_ERR(local);
 385                 goto err_remote;
 386         }
 387
 388         saved_heartbeat = engine->props.heartbeat_interval_ms;
 389         engine->props.heartbeat_interval_ms = 0;
 390         intel_engine_pm_get(engine);
 391
 392         for (pass = 0; pass <= 2; pass++) {
 393                 err = __remote_sync(local, remote);
 394                 if (err)
 395                         break;
 396
 397                 err = __remote_sync(engine->kernel_context, remote);
 398                 if (err)
 399                         break;
 400
 401                 if (i915_active_is_idle(&remote->active)) {
 402                         pr_err("remote context is not active; expected idle-barrier (%s pass %d)\n",
 403                                engine->name, pass);
 404                         err = -EINVAL;
 405                         break;
 406                 }
 407         }
 408
 409         intel_engine_pm_put(engine);
 410         engine->props.heartbeat_interval_ms = saved_heartbeat;
 411
 412         intel_context_put(local);
 413 err_remote:
 414         intel_context_put(remote);
 415         return err;
 416 }
 417
 418 static int live_remote_context(void *arg)
 419 {
 420         struct intel_gt *gt = arg;
 421         struct intel_engine_cs *engine;
 422         enum intel_engine_id id;
 423         int err = 0;
 424
 425         for_each_engine(engine, gt, id) {
 426                 err = __live_remote_context(engine);
 427                 if (err)
 428                         break;
 429
 430                 err = igt_flush_test(gt->i915);
 431                 if (err)
 432                         break;
 433         }
 434
 435         return err;
 436 }
 437
 438 int intel_context_live_selftests(struct drm_i915_private *i915)
 439 {
 440         static const struct i915_subtest tests[] = {
 441                 SUBTEST(live_context_size),
 442                 SUBTEST(live_active_context),
 443                 SUBTEST(live_remote_context),
 444         };
 445         struct intel_gt *gt = to_gt(i915);
 446
 447         if (intel_gt_is_wedged(gt))
 448                 return 0;
 449
 450         return intel_gt_live_subtests(tests, gt);
 451 }