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                                                  intel_gt_coherent_map_type(engine->gt,
  92                                                                             ce->state->obj,
  93                                                                             false));
  94         if (IS_ERR(vaddr)) {
  95                 err = PTR_ERR(vaddr);
  96                 intel_context_unpin(ce);
  97                 goto err;
  98         }
  99
 100         /*
 101          * Note that execlists also applies a redzone which it checks on
 102          * context unpin when debugging. We are using the same location
 103          * and same poison value so that our checks overlap. Despite the
 104          * redundancy, we want to keep this little selftest so that we
 105          * get coverage of any and all submission backends, and we can
 106          * always extend this test to ensure we trick the HW into a
 107          * compromising position wrt to the various sections that need
 108          * to be written into the context state.
 109          *
 110          * TLDR; this overlaps with the execlists redzone.
 111          */
 112         vaddr += engine->context_size - I915_GTT_PAGE_SIZE;
 113         memset(vaddr, POISON_INUSE, I915_GTT_PAGE_SIZE);
 114
 115         rq = intel_context_create_request(ce);
 116         intel_context_unpin(ce);
 117         if (IS_ERR(rq)) {
 118                 err = PTR_ERR(rq);
 119                 goto err_unpin;
 120         }
 121
 122         err = request_sync(rq);
 123         if (err)
 124                 goto err_unpin;
 125
 126         /* Force the context switch */
 127         rq = intel_engine_create_kernel_request(engine);
 128         if (IS_ERR(rq)) {
 129                 err = PTR_ERR(rq);
 130                 goto err_unpin;
 131         }
 132         err = request_sync(rq);
 133         if (err)
 134                 goto err_unpin;
 135
 136         if (memchr_inv(vaddr, POISON_INUSE, I915_GTT_PAGE_SIZE)) {
 137                 pr_err("%s context overwrote trailing red-zone!", engine->name);
 138                 err = -EINVAL;
 139         }
 140
 141 err_unpin:
 142         i915_gem_object_unpin_map(ce->state->obj);
 143 err:
 144         intel_context_put(ce);
 145         return err;
 146 }
 147
 148 static int live_context_size(void *arg)
 149 {
 150         struct intel_gt *gt = arg;
 151         struct intel_engine_cs *engine;
 152         enum intel_engine_id id;
 153         int err = 0;
 154
 155         /*
 156          * Check that our context sizes are correct by seeing if the
 157          * HW tries to write past the end of one.
 158          */
 159
 160         for_each_engine(engine, gt, id) {
 161                 struct file *saved;
 162
 163                 if (!engine->context_size)
 164                         continue;
 165
 166                 intel_engine_pm_get(engine);
 167
 168                 /*
 169                  * Hide the old default state -- we lie about the context size
 170                  * and get confused when the default state is smaller than
 171                  * expected. For our do nothing request, inheriting the
 172                  * active state is sufficient, we are only checking that we
 173                  * don't use more than we planned.
 174                  */
 175                 saved = fetch_and_zero(&engine->default_state);
 176
 177                 /* Overlaps with the execlists redzone */
 178                 engine->context_size += I915_GTT_PAGE_SIZE;
 179
 180                 err = __live_context_size(engine);
 181
 182                 engine->context_size -= I915_GTT_PAGE_SIZE;
 183
 184                 engine->default_state = saved;
 185
 186                 intel_engine_pm_put(engine);
 187
 188                 if (err)
 189                         break;
 190         }
 191
 192         return err;
 193 }
 194
 195 static int __live_active_context(struct intel_engine_cs *engine)
 196 {
 197         unsigned long saved_heartbeat;
 198         struct intel_context *ce;
 199         int pass;
 200         int err;
 201
 202         /*
 203          * We keep active contexts alive until after a subsequent context
 204          * switch as the final write from the context-save will be after
 205          * we retire the final request. We track when we unpin the context,
 206          * under the presumption that the final pin is from the last request,
 207          * and instead of immediately unpinning the context, we add a task
 208          * to unpin the context from the next idle-barrier.
 209          *
 210          * This test makes sure that the context is kept alive until a
 211          * subsequent idle-barrier (emitted when the engine wakeref hits 0
 212          * with no more outstanding requests).
 213          *
 214          * In GuC submission mode we don't use idle barriers and we instead
 215          * get a message from the GuC to signal that it is safe to unpin the
 216          * context from memory.
 217          */
 218         if (intel_engine_uses_guc(engine))
 219                 return 0;
 220
 221         if (intel_engine_pm_is_awake(engine)) {
 222                 pr_err("%s is awake before starting %s!\n",
 223                        engine->name, __func__);
 224                 return -EINVAL;
 225         }
 226
 227         ce = intel_context_create(engine);
 228         if (IS_ERR(ce))
 229                 return PTR_ERR(ce);
 230
 231         saved_heartbeat = engine->props.heartbeat_interval_ms;
 232         engine->props.heartbeat_interval_ms = 0;
 233
 234         for (pass = 0; pass <= 2; pass++) {
 235                 struct i915_request *rq;
 236
 237                 intel_engine_pm_get(engine);
 238
 239                 rq = intel_context_create_request(ce);
 240                 if (IS_ERR(rq)) {
 241                         err = PTR_ERR(rq);
 242                         goto out_engine;
 243                 }
 244
 245                 err = request_sync(rq);
 246                 if (err)
 247                         goto out_engine;
 248
 249                 /* Context will be kept active until after an idle-barrier. */
 250                 if (i915_active_is_idle(&ce->active)) {
 251                         pr_err("context is not active; expected idle-barrier (%s pass %d)\n",
 252                                engine->name, pass);
 253                         err = -EINVAL;
 254                         goto out_engine;
 255                 }
 256
 257                 if (!intel_engine_pm_is_awake(engine)) {
 258                         pr_err("%s is asleep before idle-barrier\n",
 259                                engine->name);
 260                         err = -EINVAL;
 261                         goto out_engine;
 262                 }
 263
 264 out_engine:
 265                 intel_engine_pm_put(engine);
 266                 if (err)
 267                         goto err;
 268         }
 269
 270         /* Now make sure our idle-barriers are flushed */
 271         err = intel_engine_flush_barriers(engine);
 272         if (err)
 273                 goto err;
 274
 275         /* Wait for the barrier and in the process wait for engine to park */
 276         err = context_sync(engine->kernel_context);
 277         if (err)
 278                 goto err;
 279
 280         if (!i915_active_is_idle(&ce->active)) {
 281                 pr_err("context is still active!");
 282                 err = -EINVAL;
 283         }
 284
 285         intel_engine_pm_flush(engine);
 286
 287         if (intel_engine_pm_is_awake(engine)) {
 288                 struct drm_printer p = drm_debug_printer(__func__);
 289
 290                 intel_engine_dump(engine, &p,
 291                                   "%s is still awake:%d after idle-barriers\n",
 292                                   engine->name,
 293                                   atomic_read(&engine->wakeref.count));
 294                 GEM_TRACE_DUMP();
 295
 296                 err = -EINVAL;
 297                 goto err;
 298         }
 299
 300 err:
 301         engine->props.heartbeat_interval_ms = saved_heartbeat;
 302         intel_context_put(ce);
 303         return err;
 304 }
 305
 306 static int live_active_context(void *arg)
 307 {
 308         struct intel_gt *gt = arg;
 309         struct intel_engine_cs *engine;
 310         enum intel_engine_id id;
 311         int err = 0;
 312
 313         for_each_engine(engine, gt, id) {
 314                 err = __live_active_context(engine);
 315                 if (err)
 316                         break;
 317
 318                 err = igt_flush_test(gt->i915);
 319                 if (err)
 320                         break;
 321         }
 322
 323         return err;
 324 }
 325
 326 static int __remote_sync(struct intel_context *ce, struct intel_context *remote)
 327 {
 328         struct i915_request *rq;
 329         int err;
 330
 331         err = intel_context_pin(remote);
 332         if (err)
 333                 return err;
 334
 335         rq = intel_context_create_request(ce);
 336         if (IS_ERR(rq)) {
 337                 err = PTR_ERR(rq);
 338                 goto unpin;
 339         }
 340
 341         err = intel_context_prepare_remote_request(remote, rq);
 342         if (err) {
 343                 i915_request_add(rq);
 344                 goto unpin;
 345         }
 346
 347         err = request_sync(rq);
 348
 349 unpin:
 350         intel_context_unpin(remote);
 351         return err;
 352 }
 353
 354 static int __live_remote_context(struct intel_engine_cs *engine)
 355 {
 356         struct intel_context *local, *remote;
 357         unsigned long saved_heartbeat;
 358         int pass;
 359         int err;
 360
 361         /*
 362          * Check that our idle barriers do not interfere with normal
 363          * activity tracking. In particular, check that operating
 364          * on the context image remotely (intel_context_prepare_remote_request),
 365          * which inserts foreign fences into intel_context.active, does not
 366          * clobber the idle-barrier.
 367          *
 368          * In GuC submission mode we don't use idle barriers.
 369          */
 370         if (intel_engine_uses_guc(engine))
 371                 return 0;
 372
 373         if (intel_engine_pm_is_awake(engine)) {
 374                 pr_err("%s is awake before starting %s!\n",
 375                        engine->name, __func__);
 376                 return -EINVAL;
 377         }
 378
 379         remote = intel_context_create(engine);
 380         if (IS_ERR(remote))
 381                 return PTR_ERR(remote);
 382
 383         local = intel_context_create(engine);
 384         if (IS_ERR(local)) {
 385                 err = PTR_ERR(local);
 386                 goto err_remote;
 387         }
 388
 389         saved_heartbeat = engine->props.heartbeat_interval_ms;
 390         engine->props.heartbeat_interval_ms = 0;
 391         intel_engine_pm_get(engine);
 392
 393         for (pass = 0; pass <= 2; pass++) {
 394                 err = __remote_sync(local, remote);
 395                 if (err)
 396                         break;
 397
 398                 err = __remote_sync(engine->kernel_context, remote);
 399                 if (err)
 400                         break;
 401
 402                 if (i915_active_is_idle(&remote->active)) {
 403                         pr_err("remote context is not active; expected idle-barrier (%s pass %d)\n",
 404                                engine->name, pass);
 405                         err = -EINVAL;
 406                         break;
 407                 }
 408         }
 409
 410         intel_engine_pm_put(engine);
 411         engine->props.heartbeat_interval_ms = saved_heartbeat;
 412
 413         intel_context_put(local);
 414 err_remote:
 415         intel_context_put(remote);
 416         return err;
 417 }
 418
 419 static int live_remote_context(void *arg)
 420 {
 421         struct intel_gt *gt = arg;
 422         struct intel_engine_cs *engine;
 423         enum intel_engine_id id;
 424         int err = 0;
 425
 426         for_each_engine(engine, gt, id) {
 427                 err = __live_remote_context(engine);
 428                 if (err)
 429                         break;
 430
 431                 err = igt_flush_test(gt->i915);
 432                 if (err)
 433                         break;
 434         }
 435
 436         return err;
 437 }
 438
 439 int intel_context_live_selftests(struct drm_i915_private *i915)
 440 {
 441         static const struct i915_subtest tests[] = {
 442                 SUBTEST(live_context_size),
 443                 SUBTEST(live_active_context),
 444                 SUBTEST(live_remote_context),
 445         };
 446         struct intel_gt *gt = to_gt(i915);
 447
 448         if (intel_gt_is_wedged(gt))
 449                 return 0;
 450
 451         return intel_gt_live_subtests(tests, gt);
 452 }