linux/drivers/gpu/drm/i915/gem/selftests/i915_gem_execbuffer.c

187 lines
3.9 KiB
C
Raw Normal View History

// SPDX-License-Identifier: MIT
/*
* Copyright © 2020 Intel Corporation
*/
#include "i915_selftest.h"
#include "gt/intel_engine_pm.h"
#include "selftests/igt_flush_test.h"
static u64 read_reloc(const u32 *map, int x, const u64 mask)
{
u64 reloc;
memcpy(&reloc, &map[x], sizeof(reloc));
return reloc & mask;
}
static int __igt_gpu_reloc(struct i915_execbuffer *eb,
struct drm_i915_gem_object *obj)
{
const unsigned int offsets[] = { 8, 3, 0 };
const u64 mask =
GENMASK_ULL(eb->reloc_cache.use_64bit_reloc ? 63 : 31, 0);
const u32 *map = page_mask_bits(obj->mm.mapping);
struct i915_request *rq;
struct i915_vma *vma;
int err;
int i;
vma = i915_vma_instance(obj, eb->context->vm, NULL);
if (IS_ERR(vma))
return PTR_ERR(vma);
drm/i915: Use per object locking in execbuf, v12. Now that we changed execbuf submission slightly to allow us to do all pinning in one place, we can now simply add ww versions on top of struct_mutex. All we have to do is a separate path for -EDEADLK handling, which needs to unpin all gem bo's before dropping the lock, then starting over. This finally allows us to do parallel submission, but because not all of the pinning code uses the ww ctx yet, we cannot completely drop struct_mutex yet. Changes since v1: - Keep struct_mutex for now. :( Changes since v2: - Make sure we always lock the ww context in slowpath. Changes since v3: - Don't call __eb_unreserve_vma in eb_move_to_gpu now; this can be done on normal unlock path. - Unconditionally release vmas and context. Changes since v4: - Rebased on top of struct_mutex reduction. Changes since v5: - Remove training wheels. Changes since v6: - Fix accidentally broken -ENOSPC handling. Changes since v7: - Handle gt buffer pool better. Changes since v8: - Properly clear variables, to make -EDEADLK handling not BUG. Change since v9: - Fix unpinning fence on pnv and below. Changes since v10: - Make relocation gpu chaining working again. Changes since v11: - Remove relocation chaining, pain to make it work. Signed-off-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Reviewed-by: Thomas Hellström <thomas.hellstrom@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20200819140904.1708856-9-maarten.lankhorst@linux.intel.com Signed-off-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
2020-08-19 14:08:48 +00:00
err = i915_gem_object_lock(obj, &eb->ww);
if (err)
return err;
err = i915_vma_pin_ww(vma, &eb->ww, 0, 0, PIN_USER | PIN_HIGH);
if (err)
return err;
/* 8-Byte aligned */
drm/i915: Use per object locking in execbuf, v12. Now that we changed execbuf submission slightly to allow us to do all pinning in one place, we can now simply add ww versions on top of struct_mutex. All we have to do is a separate path for -EDEADLK handling, which needs to unpin all gem bo's before dropping the lock, then starting over. This finally allows us to do parallel submission, but because not all of the pinning code uses the ww ctx yet, we cannot completely drop struct_mutex yet. Changes since v1: - Keep struct_mutex for now. :( Changes since v2: - Make sure we always lock the ww context in slowpath. Changes since v3: - Don't call __eb_unreserve_vma in eb_move_to_gpu now; this can be done on normal unlock path. - Unconditionally release vmas and context. Changes since v4: - Rebased on top of struct_mutex reduction. Changes since v5: - Remove training wheels. Changes since v6: - Fix accidentally broken -ENOSPC handling. Changes since v7: - Handle gt buffer pool better. Changes since v8: - Properly clear variables, to make -EDEADLK handling not BUG. Change since v9: - Fix unpinning fence on pnv and below. Changes since v10: - Make relocation gpu chaining working again. Changes since v11: - Remove relocation chaining, pain to make it work. Signed-off-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Reviewed-by: Thomas Hellström <thomas.hellstrom@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20200819140904.1708856-9-maarten.lankhorst@linux.intel.com Signed-off-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
2020-08-19 14:08:48 +00:00
err = __reloc_entry_gpu(eb, vma, offsets[0] * sizeof(u32), 0);
if (err <= 0)
goto reloc_err;
/* !8-Byte aligned */
drm/i915: Use per object locking in execbuf, v12. Now that we changed execbuf submission slightly to allow us to do all pinning in one place, we can now simply add ww versions on top of struct_mutex. All we have to do is a separate path for -EDEADLK handling, which needs to unpin all gem bo's before dropping the lock, then starting over. This finally allows us to do parallel submission, but because not all of the pinning code uses the ww ctx yet, we cannot completely drop struct_mutex yet. Changes since v1: - Keep struct_mutex for now. :( Changes since v2: - Make sure we always lock the ww context in slowpath. Changes since v3: - Don't call __eb_unreserve_vma in eb_move_to_gpu now; this can be done on normal unlock path. - Unconditionally release vmas and context. Changes since v4: - Rebased on top of struct_mutex reduction. Changes since v5: - Remove training wheels. Changes since v6: - Fix accidentally broken -ENOSPC handling. Changes since v7: - Handle gt buffer pool better. Changes since v8: - Properly clear variables, to make -EDEADLK handling not BUG. Change since v9: - Fix unpinning fence on pnv and below. Changes since v10: - Make relocation gpu chaining working again. Changes since v11: - Remove relocation chaining, pain to make it work. Signed-off-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Reviewed-by: Thomas Hellström <thomas.hellstrom@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20200819140904.1708856-9-maarten.lankhorst@linux.intel.com Signed-off-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
2020-08-19 14:08:48 +00:00
err = __reloc_entry_gpu(eb, vma, offsets[1] * sizeof(u32), 1);
if (err <= 0)
goto reloc_err;
/* Skip to the end of the cmd page */
i = PAGE_SIZE / sizeof(u32) - 1;
i -= eb->reloc_cache.rq_size;
memset32(eb->reloc_cache.rq_cmd + eb->reloc_cache.rq_size,
MI_NOOP, i);
eb->reloc_cache.rq_size += i;
/* Force next batch */
drm/i915: Use per object locking in execbuf, v12. Now that we changed execbuf submission slightly to allow us to do all pinning in one place, we can now simply add ww versions on top of struct_mutex. All we have to do is a separate path for -EDEADLK handling, which needs to unpin all gem bo's before dropping the lock, then starting over. This finally allows us to do parallel submission, but because not all of the pinning code uses the ww ctx yet, we cannot completely drop struct_mutex yet. Changes since v1: - Keep struct_mutex for now. :( Changes since v2: - Make sure we always lock the ww context in slowpath. Changes since v3: - Don't call __eb_unreserve_vma in eb_move_to_gpu now; this can be done on normal unlock path. - Unconditionally release vmas and context. Changes since v4: - Rebased on top of struct_mutex reduction. Changes since v5: - Remove training wheels. Changes since v6: - Fix accidentally broken -ENOSPC handling. Changes since v7: - Handle gt buffer pool better. Changes since v8: - Properly clear variables, to make -EDEADLK handling not BUG. Change since v9: - Fix unpinning fence on pnv and below. Changes since v10: - Make relocation gpu chaining working again. Changes since v11: - Remove relocation chaining, pain to make it work. Signed-off-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Reviewed-by: Thomas Hellström <thomas.hellstrom@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20200819140904.1708856-9-maarten.lankhorst@linux.intel.com Signed-off-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
2020-08-19 14:08:48 +00:00
err = __reloc_entry_gpu(eb, vma, offsets[2] * sizeof(u32), 2);
if (err <= 0)
goto reloc_err;
GEM_BUG_ON(!eb->reloc_cache.rq);
rq = i915_request_get(eb->reloc_cache.rq);
drm/i915: Use per object locking in execbuf, v12. Now that we changed execbuf submission slightly to allow us to do all pinning in one place, we can now simply add ww versions on top of struct_mutex. All we have to do is a separate path for -EDEADLK handling, which needs to unpin all gem bo's before dropping the lock, then starting over. This finally allows us to do parallel submission, but because not all of the pinning code uses the ww ctx yet, we cannot completely drop struct_mutex yet. Changes since v1: - Keep struct_mutex for now. :( Changes since v2: - Make sure we always lock the ww context in slowpath. Changes since v3: - Don't call __eb_unreserve_vma in eb_move_to_gpu now; this can be done on normal unlock path. - Unconditionally release vmas and context. Changes since v4: - Rebased on top of struct_mutex reduction. Changes since v5: - Remove training wheels. Changes since v6: - Fix accidentally broken -ENOSPC handling. Changes since v7: - Handle gt buffer pool better. Changes since v8: - Properly clear variables, to make -EDEADLK handling not BUG. Change since v9: - Fix unpinning fence on pnv and below. Changes since v10: - Make relocation gpu chaining working again. Changes since v11: - Remove relocation chaining, pain to make it work. Signed-off-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Reviewed-by: Thomas Hellström <thomas.hellstrom@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20200819140904.1708856-9-maarten.lankhorst@linux.intel.com Signed-off-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
2020-08-19 14:08:48 +00:00
reloc_gpu_flush(eb, &eb->reloc_cache);
GEM_BUG_ON(eb->reloc_cache.rq);
err = i915_gem_object_wait(obj, I915_WAIT_INTERRUPTIBLE, HZ / 2);
if (err) {
intel_gt_set_wedged(eb->engine->gt);
goto put_rq;
}
if (!i915_request_completed(rq)) {
pr_err("%s: did not wait for relocations!\n", eb->engine->name);
err = -EINVAL;
goto put_rq;
}
for (i = 0; i < ARRAY_SIZE(offsets); i++) {
u64 reloc = read_reloc(map, offsets[i], mask);
if (reloc != i) {
pr_err("%s[%d]: map[%d] %llx != %x\n",
eb->engine->name, i, offsets[i], reloc, i);
err = -EINVAL;
}
}
if (err)
igt_hexdump(map, 4096);
put_rq:
i915_request_put(rq);
unpin_vma:
i915_vma_unpin(vma);
return err;
drm/i915: Use per object locking in execbuf, v12. Now that we changed execbuf submission slightly to allow us to do all pinning in one place, we can now simply add ww versions on top of struct_mutex. All we have to do is a separate path for -EDEADLK handling, which needs to unpin all gem bo's before dropping the lock, then starting over. This finally allows us to do parallel submission, but because not all of the pinning code uses the ww ctx yet, we cannot completely drop struct_mutex yet. Changes since v1: - Keep struct_mutex for now. :( Changes since v2: - Make sure we always lock the ww context in slowpath. Changes since v3: - Don't call __eb_unreserve_vma in eb_move_to_gpu now; this can be done on normal unlock path. - Unconditionally release vmas and context. Changes since v4: - Rebased on top of struct_mutex reduction. Changes since v5: - Remove training wheels. Changes since v6: - Fix accidentally broken -ENOSPC handling. Changes since v7: - Handle gt buffer pool better. Changes since v8: - Properly clear variables, to make -EDEADLK handling not BUG. Change since v9: - Fix unpinning fence on pnv and below. Changes since v10: - Make relocation gpu chaining working again. Changes since v11: - Remove relocation chaining, pain to make it work. Signed-off-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Reviewed-by: Thomas Hellström <thomas.hellstrom@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20200819140904.1708856-9-maarten.lankhorst@linux.intel.com Signed-off-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
2020-08-19 14:08:48 +00:00
reloc_err:
if (!err)
err = -EIO;
goto unpin_vma;
}
static int igt_gpu_reloc(void *arg)
{
struct i915_execbuffer eb;
struct drm_i915_gem_object *scratch;
int err = 0;
u32 *map;
eb.i915 = arg;
scratch = i915_gem_object_create_internal(eb.i915, 4096);
if (IS_ERR(scratch))
return PTR_ERR(scratch);
map = i915_gem_object_pin_map(scratch, I915_MAP_WC);
if (IS_ERR(map)) {
err = PTR_ERR(map);
goto err_scratch;
}
drm/i915: Use per object locking in execbuf, v12. Now that we changed execbuf submission slightly to allow us to do all pinning in one place, we can now simply add ww versions on top of struct_mutex. All we have to do is a separate path for -EDEADLK handling, which needs to unpin all gem bo's before dropping the lock, then starting over. This finally allows us to do parallel submission, but because not all of the pinning code uses the ww ctx yet, we cannot completely drop struct_mutex yet. Changes since v1: - Keep struct_mutex for now. :( Changes since v2: - Make sure we always lock the ww context in slowpath. Changes since v3: - Don't call __eb_unreserve_vma in eb_move_to_gpu now; this can be done on normal unlock path. - Unconditionally release vmas and context. Changes since v4: - Rebased on top of struct_mutex reduction. Changes since v5: - Remove training wheels. Changes since v6: - Fix accidentally broken -ENOSPC handling. Changes since v7: - Handle gt buffer pool better. Changes since v8: - Properly clear variables, to make -EDEADLK handling not BUG. Change since v9: - Fix unpinning fence on pnv and below. Changes since v10: - Make relocation gpu chaining working again. Changes since v11: - Remove relocation chaining, pain to make it work. Signed-off-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Reviewed-by: Thomas Hellström <thomas.hellstrom@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20200819140904.1708856-9-maarten.lankhorst@linux.intel.com Signed-off-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
2020-08-19 14:08:48 +00:00
intel_gt_pm_get(&eb.i915->gt);
for_each_uabi_engine(eb.engine, eb.i915) {
reloc_cache_init(&eb.reloc_cache, eb.i915);
memset(map, POISON_INUSE, 4096);
intel_engine_pm_get(eb.engine);
eb.context = intel_context_create(eb.engine);
if (IS_ERR(eb.context)) {
err = PTR_ERR(eb.context);
goto err_pm;
}
drm/i915: Use per object locking in execbuf, v12. Now that we changed execbuf submission slightly to allow us to do all pinning in one place, we can now simply add ww versions on top of struct_mutex. All we have to do is a separate path for -EDEADLK handling, which needs to unpin all gem bo's before dropping the lock, then starting over. This finally allows us to do parallel submission, but because not all of the pinning code uses the ww ctx yet, we cannot completely drop struct_mutex yet. Changes since v1: - Keep struct_mutex for now. :( Changes since v2: - Make sure we always lock the ww context in slowpath. Changes since v3: - Don't call __eb_unreserve_vma in eb_move_to_gpu now; this can be done on normal unlock path. - Unconditionally release vmas and context. Changes since v4: - Rebased on top of struct_mutex reduction. Changes since v5: - Remove training wheels. Changes since v6: - Fix accidentally broken -ENOSPC handling. Changes since v7: - Handle gt buffer pool better. Changes since v8: - Properly clear variables, to make -EDEADLK handling not BUG. Change since v9: - Fix unpinning fence on pnv and below. Changes since v10: - Make relocation gpu chaining working again. Changes since v11: - Remove relocation chaining, pain to make it work. Signed-off-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Reviewed-by: Thomas Hellström <thomas.hellstrom@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20200819140904.1708856-9-maarten.lankhorst@linux.intel.com Signed-off-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
2020-08-19 14:08:48 +00:00
eb.reloc_pool = NULL;
eb.reloc_context = NULL;
drm/i915: Use per object locking in execbuf, v12. Now that we changed execbuf submission slightly to allow us to do all pinning in one place, we can now simply add ww versions on top of struct_mutex. All we have to do is a separate path for -EDEADLK handling, which needs to unpin all gem bo's before dropping the lock, then starting over. This finally allows us to do parallel submission, but because not all of the pinning code uses the ww ctx yet, we cannot completely drop struct_mutex yet. Changes since v1: - Keep struct_mutex for now. :( Changes since v2: - Make sure we always lock the ww context in slowpath. Changes since v3: - Don't call __eb_unreserve_vma in eb_move_to_gpu now; this can be done on normal unlock path. - Unconditionally release vmas and context. Changes since v4: - Rebased on top of struct_mutex reduction. Changes since v5: - Remove training wheels. Changes since v6: - Fix accidentally broken -ENOSPC handling. Changes since v7: - Handle gt buffer pool better. Changes since v8: - Properly clear variables, to make -EDEADLK handling not BUG. Change since v9: - Fix unpinning fence on pnv and below. Changes since v10: - Make relocation gpu chaining working again. Changes since v11: - Remove relocation chaining, pain to make it work. Signed-off-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Reviewed-by: Thomas Hellström <thomas.hellstrom@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20200819140904.1708856-9-maarten.lankhorst@linux.intel.com Signed-off-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
2020-08-19 14:08:48 +00:00
i915_gem_ww_ctx_init(&eb.ww, false);
retry:
err = intel_context_pin_ww(eb.context, &eb.ww);
drm/i915: Use per object locking in execbuf, v12. Now that we changed execbuf submission slightly to allow us to do all pinning in one place, we can now simply add ww versions on top of struct_mutex. All we have to do is a separate path for -EDEADLK handling, which needs to unpin all gem bo's before dropping the lock, then starting over. This finally allows us to do parallel submission, but because not all of the pinning code uses the ww ctx yet, we cannot completely drop struct_mutex yet. Changes since v1: - Keep struct_mutex for now. :( Changes since v2: - Make sure we always lock the ww context in slowpath. Changes since v3: - Don't call __eb_unreserve_vma in eb_move_to_gpu now; this can be done on normal unlock path. - Unconditionally release vmas and context. Changes since v4: - Rebased on top of struct_mutex reduction. Changes since v5: - Remove training wheels. Changes since v6: - Fix accidentally broken -ENOSPC handling. Changes since v7: - Handle gt buffer pool better. Changes since v8: - Properly clear variables, to make -EDEADLK handling not BUG. Change since v9: - Fix unpinning fence on pnv and below. Changes since v10: - Make relocation gpu chaining working again. Changes since v11: - Remove relocation chaining, pain to make it work. Signed-off-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Reviewed-by: Thomas Hellström <thomas.hellstrom@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20200819140904.1708856-9-maarten.lankhorst@linux.intel.com Signed-off-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
2020-08-19 14:08:48 +00:00
if (!err) {
err = __igt_gpu_reloc(&eb, scratch);
intel_context_unpin(eb.context);
}
if (err == -EDEADLK) {
err = i915_gem_ww_ctx_backoff(&eb.ww);
if (!err)
goto retry;
}
i915_gem_ww_ctx_fini(&eb.ww);
drm/i915: Use per object locking in execbuf, v12. Now that we changed execbuf submission slightly to allow us to do all pinning in one place, we can now simply add ww versions on top of struct_mutex. All we have to do is a separate path for -EDEADLK handling, which needs to unpin all gem bo's before dropping the lock, then starting over. This finally allows us to do parallel submission, but because not all of the pinning code uses the ww ctx yet, we cannot completely drop struct_mutex yet. Changes since v1: - Keep struct_mutex for now. :( Changes since v2: - Make sure we always lock the ww context in slowpath. Changes since v3: - Don't call __eb_unreserve_vma in eb_move_to_gpu now; this can be done on normal unlock path. - Unconditionally release vmas and context. Changes since v4: - Rebased on top of struct_mutex reduction. Changes since v5: - Remove training wheels. Changes since v6: - Fix accidentally broken -ENOSPC handling. Changes since v7: - Handle gt buffer pool better. Changes since v8: - Properly clear variables, to make -EDEADLK handling not BUG. Change since v9: - Fix unpinning fence on pnv and below. Changes since v10: - Make relocation gpu chaining working again. Changes since v11: - Remove relocation chaining, pain to make it work. Signed-off-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Reviewed-by: Thomas Hellström <thomas.hellstrom@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20200819140904.1708856-9-maarten.lankhorst@linux.intel.com Signed-off-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
2020-08-19 14:08:48 +00:00
if (eb.reloc_pool)
intel_gt_buffer_pool_put(eb.reloc_pool);
if (eb.reloc_context)
intel_context_put(eb.reloc_context);
intel_context_put(eb.context);
err_pm:
intel_engine_pm_put(eb.engine);
if (err)
break;
}
if (igt_flush_test(eb.i915))
err = -EIO;
drm/i915: Use per object locking in execbuf, v12. Now that we changed execbuf submission slightly to allow us to do all pinning in one place, we can now simply add ww versions on top of struct_mutex. All we have to do is a separate path for -EDEADLK handling, which needs to unpin all gem bo's before dropping the lock, then starting over. This finally allows us to do parallel submission, but because not all of the pinning code uses the ww ctx yet, we cannot completely drop struct_mutex yet. Changes since v1: - Keep struct_mutex for now. :( Changes since v2: - Make sure we always lock the ww context in slowpath. Changes since v3: - Don't call __eb_unreserve_vma in eb_move_to_gpu now; this can be done on normal unlock path. - Unconditionally release vmas and context. Changes since v4: - Rebased on top of struct_mutex reduction. Changes since v5: - Remove training wheels. Changes since v6: - Fix accidentally broken -ENOSPC handling. Changes since v7: - Handle gt buffer pool better. Changes since v8: - Properly clear variables, to make -EDEADLK handling not BUG. Change since v9: - Fix unpinning fence on pnv and below. Changes since v10: - Make relocation gpu chaining working again. Changes since v11: - Remove relocation chaining, pain to make it work. Signed-off-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Reviewed-by: Thomas Hellström <thomas.hellstrom@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20200819140904.1708856-9-maarten.lankhorst@linux.intel.com Signed-off-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
2020-08-19 14:08:48 +00:00
intel_gt_pm_put(&eb.i915->gt);
err_scratch:
i915_gem_object_put(scratch);
return err;
}
int i915_gem_execbuffer_live_selftests(struct drm_i915_private *i915)
{
static const struct i915_subtest tests[] = {
SUBTEST(igt_gpu_reloc),
};
if (intel_gt_is_wedged(&i915->gt))
return 0;
return i915_live_subtests(tests, i915);
}