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