drivers/gpu/drm/i915/selftests/i915_gem_context.c

   1 /*
   2  * Copyright © 2017 Intel Corporation
   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 (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  *
  23  */
  24
  25 #include "../i915_selftest.h"
  26
  27 #include "mock_drm.h"
  28 #include "huge_gem_object.h"
  29
  30 #define DW_PER_PAGE (PAGE_SIZE / sizeof(u32))
  31
  32 static struct i915_vma *
  33 gpu_fill_dw(struct i915_vma *vma, u64 offset, unsigned long count, u32 value)
  34 {
  35         struct drm_i915_gem_object *obj;
  36         const int gen = INTEL_GEN(vma->vm->i915);
  37         unsigned long n, size;
  38         u32 *cmd;
  39         int err;
  40
  41         GEM_BUG_ON(!igt_can_mi_store_dword_imm(vma->vm->i915));
  42
  43         size = (4 * count + 1) * sizeof(u32);
  44         size = round_up(size, PAGE_SIZE);
  45         obj = i915_gem_object_create_internal(vma->vm->i915, size);
  46         if (IS_ERR(obj))
  47                 return ERR_CAST(obj);
  48
  49         cmd = i915_gem_object_pin_map(obj, I915_MAP_WB);
  50         if (IS_ERR(cmd)) {
  51                 err = PTR_ERR(cmd);
  52                 goto err;
  53         }
  54
  55         GEM_BUG_ON(offset + (count - 1) * PAGE_SIZE > vma->node.size);
  56         offset += vma->node.start;
  57
  58         for (n = 0; n < count; n++) {
  59                 if (gen >= 8) {
  60                         *cmd++ = MI_STORE_DWORD_IMM_GEN4;
  61                         *cmd++ = lower_32_bits(offset);
  62                         *cmd++ = upper_32_bits(offset);
  63                         *cmd++ = value;
  64                 } else if (gen >= 4) {
  65                         *cmd++ = MI_STORE_DWORD_IMM_GEN4 |
  66                                 (gen < 6 ? 1 << 22 : 0);
  67                         *cmd++ = 0;
  68                         *cmd++ = offset;
  69                         *cmd++ = value;
  70                 } else {
  71                         *cmd++ = MI_STORE_DWORD_IMM | 1 << 22;
  72                         *cmd++ = offset;
  73                         *cmd++ = value;
  74                 }
  75                 offset += PAGE_SIZE;
  76         }
  77         *cmd = MI_BATCH_BUFFER_END;
  78         i915_gem_object_unpin_map(obj);
  79
  80         err = i915_gem_object_set_to_gtt_domain(obj, false);
  81         if (err)
  82                 goto err;
  83
  84         vma = i915_vma_instance(obj, vma->vm, NULL);
  85         if (IS_ERR(vma)) {
  86                 err = PTR_ERR(vma);
  87                 goto err;
  88         }
  89
  90         err = i915_vma_pin(vma, 0, 0, PIN_USER);
  91         if (err)
  92                 goto err;
  93
  94         return vma;
  95
  96 err:
  97         i915_gem_object_put(obj);
  98         return ERR_PTR(err);
  99 }
 100
 101 static unsigned long real_page_count(struct drm_i915_gem_object *obj)
 102 {
 103         return huge_gem_object_phys_size(obj) >> PAGE_SHIFT;
 104 }
 105
 106 static unsigned long fake_page_count(struct drm_i915_gem_object *obj)
 107 {
 108         return huge_gem_object_dma_size(obj) >> PAGE_SHIFT;
 109 }
 110
 111 static int gpu_fill(struct drm_i915_gem_object *obj,
 112                     struct i915_gem_context *ctx,
 113                     struct intel_engine_cs *engine,
 114                     unsigned int dw)
 115 {
 116         struct drm_i915_private *i915 = to_i915(obj->base.dev);
 117         struct i915_address_space *vm =
 118                 ctx->ppgtt ? &ctx->ppgtt->base : &i915->ggtt.base;
 119         struct drm_i915_gem_request *rq;
 120         struct i915_vma *vma;
 121         struct i915_vma *batch;
 122         unsigned int flags;
 123         int err;
 124
 125         GEM_BUG_ON(obj->base.size > vm->total);
 126
 127         vma = i915_vma_instance(obj, vm, NULL);
 128         if (IS_ERR(vma))
 129                 return PTR_ERR(vma);
 130
 131         err = i915_gem_object_set_to_gtt_domain(obj, false);
 132         if (err)
 133                 return err;
 134
 135         err = i915_vma_pin(vma, 0, 0, PIN_HIGH | PIN_USER);
 136         if (err)
 137                 return err;
 138
 139         /* Within the GTT the huge objects maps every page onto
 140          * its 1024 real pages (using phys_pfn = dma_pfn % 1024).
 141          * We set the nth dword within the page using the nth
 142          * mapping via the GTT - this should exercise the GTT mapping
 143          * whilst checking that each context provides a unique view
 144          * into the object.
 145          */
 146         batch = gpu_fill_dw(vma,
 147                             (dw * real_page_count(obj)) << PAGE_SHIFT |
 148                             (dw * sizeof(u32)),
 149                             real_page_count(obj),
 150                             dw);
 151         if (IS_ERR(batch)) {
 152                 err = PTR_ERR(batch);
 153                 goto err_vma;
 154         }
 155
 156         rq = i915_gem_request_alloc(engine, ctx);
 157         if (IS_ERR(rq)) {
 158                 err = PTR_ERR(rq);
 159                 goto err_batch;
 160         }
 161
 162         err = engine->emit_flush(rq, EMIT_INVALIDATE);
 163         if (err)
 164                 goto err_request;
 165
 166         err = i915_switch_context(rq);
 167         if (err)
 168                 goto err_request;
 169
 170         flags = 0;
 171         if (INTEL_GEN(vm->i915) <= 5)
 172                 flags |= I915_DISPATCH_SECURE;
 173
 174         err = engine->emit_bb_start(rq,
 175                                     batch->node.start, batch->node.size,
 176                                     flags);
 177         if (err)
 178                 goto err_request;
 179
 180         i915_vma_move_to_active(batch, rq, 0);
 181         i915_gem_object_set_active_reference(batch->obj);
 182         i915_vma_unpin(batch);
 183         i915_vma_close(batch);
 184
 185         i915_vma_move_to_active(vma, rq, 0);
 186         i915_vma_unpin(vma);
 187
 188         reservation_object_lock(obj->resv, NULL);
 189         reservation_object_add_excl_fence(obj->resv, &rq->fence);
 190         reservation_object_unlock(obj->resv);
 191
 192         __i915_add_request(rq, true);
 193
 194         return 0;
 195
 196 err_request:
 197         __i915_add_request(rq, false);
 198 err_batch:
 199         i915_vma_unpin(batch);
 200 err_vma:
 201         i915_vma_unpin(vma);
 202         return err;
 203 }
 204
 205 static int cpu_fill(struct drm_i915_gem_object *obj, u32 value)
 206 {
 207         const bool has_llc = HAS_LLC(to_i915(obj->base.dev));
 208         unsigned int n, m, need_flush;
 209         int err;
 210
 211         err = i915_gem_obj_prepare_shmem_write(obj, &need_flush);
 212         if (err)
 213                 return err;
 214
 215         for (n = 0; n < real_page_count(obj); n++) {
 216                 u32 *map;
 217
 218                 map = kmap_atomic(i915_gem_object_get_page(obj, n));
 219                 for (m = 0; m < DW_PER_PAGE; m++)
 220                         map[m] = value;
 221                 if (!has_llc)
 222                         drm_clflush_virt_range(map, PAGE_SIZE);
 223                 kunmap_atomic(map);
 224         }
 225
 226         i915_gem_obj_finish_shmem_access(obj);
 227         obj->base.read_domains = I915_GEM_DOMAIN_GTT | I915_GEM_DOMAIN_CPU;
 228         obj->base.write_domain = 0;
 229         return 0;
 230 }
 231
 232 static int cpu_check(struct drm_i915_gem_object *obj, unsigned int max)
 233 {
 234         unsigned int n, m, needs_flush;
 235         int err;
 236
 237         err = i915_gem_obj_prepare_shmem_read(obj, &needs_flush);
 238         if (err)
 239                 return err;
 240
 241         for (n = 0; n < real_page_count(obj); n++) {
 242                 u32 *map;
 243
 244                 map = kmap_atomic(i915_gem_object_get_page(obj, n));
 245                 if (needs_flush & CLFLUSH_BEFORE)
 246                         drm_clflush_virt_range(map, PAGE_SIZE);
 247
 248                 for (m = 0; m < max; m++) {
 249                         if (map[m] != m) {
 250                                 pr_err("Invalid value at page %d, offset %d: found %x expected %x\n",
 251                                        n, m, map[m], m);
 252                                 err = -EINVAL;
 253                                 goto out_unmap;
 254                         }
 255                 }
 256
 257                 for (; m < DW_PER_PAGE; m++) {
 258                         if (map[m] != 0xdeadbeef) {
 259                                 pr_err("Invalid value at page %d, offset %d: found %x expected %x\n",
 260                                        n, m, map[m], 0xdeadbeef);
 261                                 err = -EINVAL;
 262                                 goto out_unmap;
 263                         }
 264                 }
 265
 266 out_unmap:
 267                 kunmap_atomic(map);
 268                 if (err)
 269                         break;
 270         }
 271
 272         i915_gem_obj_finish_shmem_access(obj);
 273         return err;
 274 }
 275
 276 static struct drm_i915_gem_object *
 277 create_test_object(struct i915_gem_context *ctx,
 278                    struct drm_file *file,
 279                    struct list_head *objects)
 280 {
 281         struct drm_i915_gem_object *obj;
 282         struct i915_address_space *vm =
 283                 ctx->ppgtt ? &ctx->ppgtt->base : &ctx->i915->ggtt.base;
 284         u64 size;
 285         u32 handle;
 286         int err;
 287
 288         size = min(vm->total / 2, 1024ull * DW_PER_PAGE * PAGE_SIZE);
 289         size = round_down(size, DW_PER_PAGE * PAGE_SIZE);
 290
 291         obj = huge_gem_object(ctx->i915, DW_PER_PAGE * PAGE_SIZE, size);
 292         if (IS_ERR(obj))
 293                 return obj;
 294
 295         /* tie the handle to the drm_file for easy reaping */
 296         err = drm_gem_handle_create(file, &obj->base, &handle);
 297         i915_gem_object_put(obj);
 298         if (err)
 299                 return ERR_PTR(err);
 300
 301         err = cpu_fill(obj, 0xdeadbeef);
 302         if (err) {
 303                 pr_err("Failed to fill object with cpu, err=%d\n",
 304                        err);
 305                 return ERR_PTR(err);
 306         }
 307
 308         list_add_tail(&obj->st_link, objects);
 309         return obj;
 310 }
 311
 312 static unsigned long max_dwords(struct drm_i915_gem_object *obj)
 313 {
 314         unsigned long npages = fake_page_count(obj);
 315
 316         GEM_BUG_ON(!IS_ALIGNED(npages, DW_PER_PAGE));
 317         return npages / DW_PER_PAGE;
 318 }
 319
 320 static int igt_ctx_exec(void *arg)
 321 {
 322         struct drm_i915_private *i915 = arg;
 323         struct drm_i915_gem_object *obj = NULL;
 324         struct drm_file *file;
 325         IGT_TIMEOUT(end_time);
 326         LIST_HEAD(objects);
 327         unsigned long ncontexts, ndwords, dw;
 328         bool first_shared_gtt = true;
 329         int err;
 330
 331         /* Create a few different contexts (with different mm) and write
 332          * through each ctx/mm using the GPU making sure those writes end
 333          * up in the expected pages of our obj.
 334          */
 335
 336         file = mock_file(i915);
 337         if (IS_ERR(file))
 338                 return PTR_ERR(file);
 339
 340         mutex_lock(&i915->drm.struct_mutex);
 341
 342         ncontexts = 0;
 343         ndwords = 0;
 344         dw = 0;
 345         while (!time_after(jiffies, end_time)) {
 346                 struct intel_engine_cs *engine;
 347                 struct i915_gem_context *ctx;
 348                 unsigned int id;
 349
 350                 if (first_shared_gtt) {
 351                         ctx = __create_hw_context(i915, file->driver_priv);
 352                         first_shared_gtt = false;
 353                 } else {
 354                         ctx = i915_gem_create_context(i915, file->driver_priv);
 355                 }
 356                 if (IS_ERR(ctx)) {
 357                         err = PTR_ERR(ctx);
 358                         goto out_unlock;
 359                 }
 360
 361                 for_each_engine(engine, i915, id) {
 362                         if (!obj) {
 363                                 obj = create_test_object(ctx, file, &objects);
 364                                 if (IS_ERR(obj)) {
 365                                         err = PTR_ERR(obj);
 366                                         goto out_unlock;
 367                                 }
 368                         }
 369
 370                         err = gpu_fill(obj, ctx, engine, dw);
 371                         if (err) {
 372                                 pr_err("Failed to fill dword %lu [%lu/%lu] with gpu (%s) in ctx %u [full-ppgtt? %s], err=%d\n",
 373                                        ndwords, dw, max_dwords(obj),
 374                                        engine->name, ctx->hw_id,
 375                                        yesno(!!ctx->ppgtt), err);
 376                                 goto out_unlock;
 377                         }
 378
 379                         if (++dw == max_dwords(obj)) {
 380                                 obj = NULL;
 381                                 dw = 0;
 382                         }
 383                         ndwords++;
 384                 }
 385                 ncontexts++;
 386         }
 387         pr_info("Submitted %lu contexts (across %u engines), filling %lu dwords\n",
 388                 ncontexts, INTEL_INFO(i915)->num_rings, ndwords);
 389
 390         dw = 0;
 391         list_for_each_entry(obj, &objects, st_link) {
 392                 unsigned int rem =
 393                         min_t(unsigned int, ndwords - dw, max_dwords(obj));
 394
 395                 err = cpu_check(obj, rem);
 396                 if (err)
 397                         break;
 398
 399                 dw += rem;
 400         }
 401
 402 out_unlock:
 403         mutex_unlock(&i915->drm.struct_mutex);
 404
 405         mock_file_free(i915, file);
 406         return err;
 407 }
 408
 409 static int fake_aliasing_ppgtt_enable(struct drm_i915_private *i915)
 410 {
 411         struct drm_i915_gem_object *obj;
 412         int err;
 413
 414         err = i915_gem_init_aliasing_ppgtt(i915);
 415         if (err)
 416                 return err;
 417
 418         list_for_each_entry(obj, &i915->mm.bound_list, global_link) {
 419                 struct i915_vma *vma;
 420
 421                 vma = i915_vma_instance(obj, &i915->ggtt.base, NULL);
 422                 if (IS_ERR(vma))
 423                         continue;
 424
 425                 vma->flags &= ~I915_VMA_LOCAL_BIND;
 426         }
 427
 428         return 0;
 429 }
 430
 431 static void fake_aliasing_ppgtt_disable(struct drm_i915_private *i915)
 432 {
 433         i915_gem_fini_aliasing_ppgtt(i915);
 434 }
 435
 436 int i915_gem_context_live_selftests(struct drm_i915_private *dev_priv)
 437 {
 438         static const struct i915_subtest tests[] = {
 439                 SUBTEST(igt_ctx_exec),
 440         };
 441         bool fake_alias = false;
 442         int err;
 443
 444         /* Install a fake aliasing gtt for exercise */
 445         if (USES_PPGTT(dev_priv) && !dev_priv->mm.aliasing_ppgtt) {
 446                 mutex_lock(&dev_priv->drm.struct_mutex);
 447                 err = fake_aliasing_ppgtt_enable(dev_priv);
 448                 mutex_unlock(&dev_priv->drm.struct_mutex);
 449                 if (err)
 450                         return err;
 451
 452                 GEM_BUG_ON(!dev_priv->mm.aliasing_ppgtt);
 453                 fake_alias = true;
 454         }
 455
 456         err = i915_subtests(tests, dev_priv);
 457
 458         if (fake_alias) {
 459                 mutex_lock(&dev_priv->drm.struct_mutex);
 460                 fake_aliasing_ppgtt_disable(dev_priv);
 461                 mutex_unlock(&dev_priv->drm.struct_mutex);
 462         }
 463
 464         return err;
 465 }