| Message ID | 20190809222643.23142-4-matthew.auld@intel.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | Introduce memory region concept (including device local memory) | expand |
Quoting Matthew Auld (2019-08-09 23:26:09) > diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c > index 6ff01a404346..8735dea74809 100644 > --- a/drivers/gpu/drm/i915/i915_gem.c > +++ b/drivers/gpu/drm/i915/i915_gem.c > @@ -1105,6 +1105,23 @@ i915_gem_madvise_ioctl(struct drm_device *dev, void *data, > !i915_gem_object_has_pages(obj)) > i915_gem_object_truncate(obj); > > + if (obj->mm.region) { > + mutex_lock(&obj->mm.region->obj_lock); > + > + switch (obj->mm.madv) { > + case I915_MADV_WILLNEED: > + list_move(&obj->mm.region_link, > + &obj->mm.region->objects); > + break; > + default: > + list_move(&obj->mm.region_link, > + &obj->mm.region->purgeable); > + break; > + } > + > + mutex_unlock(&obj->mm.region->obj_lock); > + } > + > args->retained = obj->mm.madv != __I915_MADV_PURGED; Little bit of an impedance mismatch, I hope this turns out fine when everything is a memory region. > mutex_unlock(&obj->mm.lock); > > diff --git a/drivers/gpu/drm/i915/intel_memory_region.c b/drivers/gpu/drm/i915/intel_memory_region.c > index ef12e462acb8..3a3caaadea1f 100644 > --- a/drivers/gpu/drm/i915/intel_memory_region.c > +++ b/drivers/gpu/drm/i915/intel_memory_region.c > @@ -12,6 +12,51 @@ const u32 intel_region_map[] = { > [INTEL_MEMORY_STOLEN] = BIT(INTEL_STOLEN + INTEL_MEMORY_TYPE_SHIFT) | BIT(0), > }; > > +static int > +intel_memory_region_evict(struct intel_memory_region *mem, > + resource_size_t target, > + unsigned int flags) > +{ > + struct drm_i915_gem_object *obj; > + resource_size_t found; > + int err; > + > + err = 0; > + found = 0; > + > + mutex_lock(&mem->obj_lock); > + list_for_each_entry(obj, &mem->purgeable, mm.region_link) { > + if (!i915_gem_object_has_pages(obj)) > + continue; > + > + if (READ_ONCE(obj->pin_global)) > + continue; > + > + if (atomic_read(&obj->bind_count)) > + continue; > + > + mutex_unlock(&mem->obj_lock); > + > + __i915_gem_object_put_pages(obj, I915_MM_SHRINKER); So we don't really care about the object being bound then? As all we care about is the page's pin_count. So instead of obj->pin_global, obj->bind_bound, you just want if (atomic_read(&obj->pages.pin_count)) continue; as the quick check to see if it is worth preceding. > + mutex_lock_nested(&obj->mm.lock, I915_MM_SHRINKER); > + if (!i915_gem_object_has_pages(obj)) { > + obj->mm.madv = __I915_MADV_PURGED; > + found += obj->base.size; > + } > + mutex_unlock(&obj->mm.lock); The locking here accomplishes what? You just want a boolean from put_pages(). > + > + if (found >= target) > + return 0; > + > + mutex_lock(&mem->obj_lock); > + } > + > + err = -ENOSPC; > + mutex_unlock(&mem->obj_lock); > + return err; > +}
> -----Original Message----- > From: Intel-gfx [mailto:intel-gfx-bounces@lists.freedesktop.org] On Behalf > Of Chris Wilson > Sent: Saturday, August 10, 2019 3:19 AM > To: Auld, Matthew <matthew.auld@intel.com>; intel- > gfx@lists.freedesktop.org > Cc: dri-devel@lists.freedesktop.org > Subject: Re: [Intel-gfx] [PATCH v3 03/37] drm/i915/region: support basic > eviction > > Quoting Matthew Auld (2019-08-09 23:26:09) > > diff --git a/drivers/gpu/drm/i915/i915_gem.c > > b/drivers/gpu/drm/i915/i915_gem.c index 6ff01a404346..8735dea74809 > > 100644 > > --- a/drivers/gpu/drm/i915/i915_gem.c > > +++ b/drivers/gpu/drm/i915/i915_gem.c > > @@ -1105,6 +1105,23 @@ i915_gem_madvise_ioctl(struct drm_device > *dev, void *data, > > !i915_gem_object_has_pages(obj)) > > i915_gem_object_truncate(obj); > > > > + if (obj->mm.region) { > > + mutex_lock(&obj->mm.region->obj_lock); > > + > > + switch (obj->mm.madv) { > > + case I915_MADV_WILLNEED: > > + list_move(&obj->mm.region_link, > > + &obj->mm.region->objects); > > + break; > > + default: > > + list_move(&obj->mm.region_link, > > + &obj->mm.region->purgeable); > > + break; > > + } > > + > > + mutex_unlock(&obj->mm.region->obj_lock); > > + } > > + > > args->retained = obj->mm.madv != __I915_MADV_PURGED; > > Little bit of an impedance mismatch, I hope this turns out fine when > everything is a memory region. > > > mutex_unlock(&obj->mm.lock); > > > > diff --git a/drivers/gpu/drm/i915/intel_memory_region.c > > b/drivers/gpu/drm/i915/intel_memory_region.c > > index ef12e462acb8..3a3caaadea1f 100644 > > --- a/drivers/gpu/drm/i915/intel_memory_region.c > > +++ b/drivers/gpu/drm/i915/intel_memory_region.c > > @@ -12,6 +12,51 @@ const u32 intel_region_map[] = { > > [INTEL_MEMORY_STOLEN] = BIT(INTEL_STOLEN + > > INTEL_MEMORY_TYPE_SHIFT) | BIT(0), }; > > > > +static int > > +intel_memory_region_evict(struct intel_memory_region *mem, > > + resource_size_t target, > > + unsigned int flags) { > > + struct drm_i915_gem_object *obj; > > + resource_size_t found; > > + int err; > > + > > + err = 0; > > + found = 0; > > + > > + mutex_lock(&mem->obj_lock); > > + list_for_each_entry(obj, &mem->purgeable, mm.region_link) { > > + if (!i915_gem_object_has_pages(obj)) > > + continue; > > + > > + if (READ_ONCE(obj->pin_global)) > > + continue; > > + > > + if (atomic_read(&obj->bind_count)) > > + continue; > > + > > + mutex_unlock(&mem->obj_lock); > > + > > + __i915_gem_object_put_pages(obj, I915_MM_SHRINKER); > > So we don't really care about the object being bound then? As all we care > about is the page's pin_count. > > So instead of obj->pin_global, obj->bind_bound, you just want > > if (atomic_read(&obj->pages.pin_count)) > continue; > > as the quick check to see if it is worth preceding. > > > + mutex_lock_nested(&obj->mm.lock, I915_MM_SHRINKER); > > + if (!i915_gem_object_has_pages(obj)) { > > + obj->mm.madv = __I915_MADV_PURGED; > > + found += obj->base.size; > > + } > > + mutex_unlock(&obj->mm.lock); > > The locking here accomplishes what? You just want a boolean from > put_pages(). I have the same question. But looked the i915_gem_shrink() function, it has similar code. Do we prevent any race condition here? I want to use this function for swapping so hope to understand more. --CQ > > > + > > + if (found >= target) > > + return 0; > > + > > + mutex_lock(&mem->obj_lock); > > + } > > + > > + err = -ENOSPC; > > + mutex_unlock(&mem->obj_lock); > > + return err; > > +} > _______________________________________________ > Intel-gfx mailing list > Intel-gfx@lists.freedesktop.org > https://lists.freedesktop.org/mailman/listinfo/intel-gfx
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object_types.h b/drivers/gpu/drm/i915/gem/i915_gem_object_types.h index a32066e66271..5e2fa37e9bc0 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_object_types.h +++ b/drivers/gpu/drm/i915/gem/i915_gem_object_types.h @@ -168,6 +168,13 @@ struct drm_i915_gem_object { * List of memory region blocks allocated for this object. */ struct list_head blocks; + /** + * Element within memory_region->objects or + * memory_region->purgeable if the object is marked as + * DONTNEED. Access is protected by memory_region->obj_lock. + */ + struct list_head region_link; + struct list_head tmp_link; struct sg_table *pages; void *mapping; diff --git a/drivers/gpu/drm/i915/gem/i915_gem_region.c b/drivers/gpu/drm/i915/gem/i915_gem_region.c index 3cd1bf15e25b..be126e70c90f 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_region.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_region.c @@ -102,6 +102,17 @@ void i915_gem_object_init_memory_region(struct drm_i915_gem_object *obj, { INIT_LIST_HEAD(&obj->mm.blocks); obj->mm.region= mem; + + mutex_lock(&mem->obj_lock); + list_add(&obj->mm.region_link, &mem->objects); + mutex_unlock(&mem->obj_lock); +} + +void i915_gem_object_release_memory_region(struct drm_i915_gem_object *obj) +{ + mutex_lock(&obj->mm.region->obj_lock); + list_del(&obj->mm.region_link); + mutex_unlock(&obj->mm.region->obj_lock); } struct drm_i915_gem_object * diff --git a/drivers/gpu/drm/i915/gem/i915_gem_region.h b/drivers/gpu/drm/i915/gem/i915_gem_region.h index da5a2ca1a0fb..ebddc86d78f7 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_region.h +++ b/drivers/gpu/drm/i915/gem/i915_gem_region.h @@ -18,6 +18,7 @@ void i915_gem_object_put_pages_buddy(struct drm_i915_gem_object *obj, void i915_gem_object_init_memory_region(struct drm_i915_gem_object *obj, struct intel_memory_region *mem); +void i915_gem_object_release_memory_region(struct drm_i915_gem_object *obj); struct drm_i915_gem_object * i915_gem_object_create_region(struct intel_memory_region *mem, diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c index 6ff01a404346..8735dea74809 100644 --- a/drivers/gpu/drm/i915/i915_gem.c +++ b/drivers/gpu/drm/i915/i915_gem.c @@ -1105,6 +1105,23 @@ i915_gem_madvise_ioctl(struct drm_device *dev, void *data, !i915_gem_object_has_pages(obj)) i915_gem_object_truncate(obj); + if (obj->mm.region) { + mutex_lock(&obj->mm.region->obj_lock); + + switch (obj->mm.madv) { + case I915_MADV_WILLNEED: + list_move(&obj->mm.region_link, + &obj->mm.region->objects); + break; + default: + list_move(&obj->mm.region_link, + &obj->mm.region->purgeable); + break; + } + + mutex_unlock(&obj->mm.region->obj_lock); + } + args->retained = obj->mm.madv != __I915_MADV_PURGED; mutex_unlock(&obj->mm.lock); diff --git a/drivers/gpu/drm/i915/intel_memory_region.c b/drivers/gpu/drm/i915/intel_memory_region.c index ef12e462acb8..3a3caaadea1f 100644 --- a/drivers/gpu/drm/i915/intel_memory_region.c +++ b/drivers/gpu/drm/i915/intel_memory_region.c @@ -12,6 +12,51 @@ const u32 intel_region_map[] = { [INTEL_MEMORY_STOLEN] = BIT(INTEL_STOLEN + INTEL_MEMORY_TYPE_SHIFT) | BIT(0), }; +static int +intel_memory_region_evict(struct intel_memory_region *mem, + resource_size_t target, + unsigned int flags) +{ + struct drm_i915_gem_object *obj; + resource_size_t found; + int err; + + err = 0; + found = 0; + + mutex_lock(&mem->obj_lock); + list_for_each_entry(obj, &mem->purgeable, mm.region_link) { + if (!i915_gem_object_has_pages(obj)) + continue; + + if (READ_ONCE(obj->pin_global)) + continue; + + if (atomic_read(&obj->bind_count)) + continue; + + mutex_unlock(&mem->obj_lock); + + __i915_gem_object_put_pages(obj, I915_MM_SHRINKER); + + mutex_lock_nested(&obj->mm.lock, I915_MM_SHRINKER); + if (!i915_gem_object_has_pages(obj)) { + obj->mm.madv = __I915_MADV_PURGED; + found += obj->base.size; + } + mutex_unlock(&obj->mm.lock); + + if (found >= target) + return 0; + + mutex_lock(&mem->obj_lock); + } + + err = -ENOSPC; + mutex_unlock(&mem->obj_lock); + return err; +} + static u64 intel_memory_region_free_pages(struct intel_memory_region *mem, struct list_head *blocks) @@ -63,7 +108,8 @@ __intel_memory_region_get_pages_buddy(struct intel_memory_region *mem, do { struct i915_buddy_block *block; unsigned int order; - + bool retry = true; +retry: order = fls(n_pages) - 1; GEM_BUG_ON(order > mem->mm.max_order); @@ -72,9 +118,24 @@ __intel_memory_region_get_pages_buddy(struct intel_memory_region *mem, if (!IS_ERR(block)) break; - /* XXX: some kind of eviction pass, local to the device */ - if (flags & I915_ALLOC_CONTIGUOUS || !order--) - goto err_free_blocks; + if (flags & I915_ALLOC_CONTIGUOUS || !order--) { + resource_size_t target; + int err; + + if (!retry) + goto err_free_blocks; + + target = n_pages * mem->mm.chunk_size; + + mutex_unlock(&mem->mm_lock); + err = intel_memory_region_evict(mem, target, 0); + mutex_lock(&mem->mm_lock); + if (err) + goto err_free_blocks; + + retry = false; + goto retry; + } } while (1); n_pages -= BIT(order); @@ -147,6 +208,10 @@ intel_memory_region_create(struct drm_i915_private *i915, mem->min_page_size = min_page_size; mem->ops = ops; + mutex_init(&mem->obj_lock); + INIT_LIST_HEAD(&mem->objects); + INIT_LIST_HEAD(&mem->purgeable); + mutex_init(&mem->mm_lock); if (ops->init) { diff --git a/drivers/gpu/drm/i915/intel_memory_region.h b/drivers/gpu/drm/i915/intel_memory_region.h index d299fed169e9..340411dcf86b 100644 --- a/drivers/gpu/drm/i915/intel_memory_region.h +++ b/drivers/gpu/drm/i915/intel_memory_region.h @@ -78,6 +78,11 @@ struct intel_memory_region { unsigned int type; unsigned int instance; unsigned int id; + + /* Protects access to objects and purgeable */ + struct mutex obj_lock; + struct list_head objects; + struct list_head purgeable; }; int intel_memory_region_init_buddy(struct intel_memory_region *mem); diff --git a/drivers/gpu/drm/i915/selftests/intel_memory_region.c b/drivers/gpu/drm/i915/selftests/intel_memory_region.c index 84daa92bf92c..2f13e4c1d999 100644 --- a/drivers/gpu/drm/i915/selftests/intel_memory_region.c +++ b/drivers/gpu/drm/i915/selftests/intel_memory_region.c @@ -79,10 +79,86 @@ static int igt_mock_fill(void *arg) return err; } +static void igt_mark_evictable(struct drm_i915_gem_object *obj) +{ + i915_gem_object_unpin_pages(obj); + obj->mm.madv = I915_MADV_DONTNEED; + list_move(&obj->mm.region_link, &obj->mm.region->purgeable); +} + +static int igt_mock_shrink(void *arg) +{ + struct intel_memory_region *mem = arg; + struct drm_i915_gem_object *obj; + unsigned long n_objects; + LIST_HEAD(objects); + resource_size_t target; + resource_size_t total; + int err = 0; + + target = mem->mm.chunk_size; + total = resource_size(&mem->region); + n_objects = total / target; + + while (n_objects--) { + obj = i915_gem_object_create_region(mem, + target, + 0); + if (IS_ERR(obj)) { + err = PTR_ERR(obj); + goto err_close_objects; + } + + list_add(&obj->st_link, &objects); + + err = i915_gem_object_pin_pages(obj); + if (err) + goto err_close_objects; + + /* + * Make half of the region evictable, though do so in a + * horribly fragmented fashion. + */ + if (n_objects % 2) + igt_mark_evictable(obj); + } + + while (target <= total / 2) { + obj = i915_gem_object_create_region(mem, target, 0); + if (IS_ERR(obj)) { + err = PTR_ERR(obj); + goto err_close_objects; + } + + list_add(&obj->st_link, &objects); + + /* Provoke the shrinker to start violently swinging its axe! */ + err = i915_gem_object_pin_pages(obj); + if (err) { + pr_err("failed to shrink for target=%pa", &target); + goto err_close_objects; + } + + /* Again, half of the region should remain evictable */ + igt_mark_evictable(obj); + + target <<= 1; + } + +err_close_objects: + close_objects(&objects); + + if (err == -ENOMEM) + err = 0; + + return err; +} + int intel_memory_region_mock_selftests(void) { static const struct i915_subtest tests[] = { SUBTEST(igt_mock_fill), + SUBTEST(igt_mock_shrink), }; struct intel_memory_region *mem; struct drm_i915_private *i915; diff --git a/drivers/gpu/drm/i915/selftests/mock_region.c b/drivers/gpu/drm/i915/selftests/mock_region.c index fbde890385c7..cc97250dca62 100644 --- a/drivers/gpu/drm/i915/selftests/mock_region.c +++ b/drivers/gpu/drm/i915/selftests/mock_region.c @@ -11,6 +11,7 @@ static const struct drm_i915_gem_object_ops mock_region_obj_ops = { .get_pages = i915_gem_object_get_pages_buddy, .put_pages = i915_gem_object_put_pages_buddy, + .release = i915_gem_object_release_memory_region, }; static struct drm_i915_gem_object *
Support basic eviction for regions. Signed-off-by: Matthew Auld <matthew.auld@intel.com> Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Cc: Abdiel Janulgue <abdiel.janulgue@linux.intel.com> --- .../gpu/drm/i915/gem/i915_gem_object_types.h | 7 ++ drivers/gpu/drm/i915/gem/i915_gem_region.c | 11 +++ drivers/gpu/drm/i915/gem/i915_gem_region.h | 1 + drivers/gpu/drm/i915/i915_gem.c | 17 +++++ drivers/gpu/drm/i915/intel_memory_region.c | 73 +++++++++++++++++- drivers/gpu/drm/i915/intel_memory_region.h | 5 ++ .../drm/i915/selftests/intel_memory_region.c | 76 +++++++++++++++++++ drivers/gpu/drm/i915/selftests/mock_region.c | 1 + 8 files changed, 187 insertions(+), 4 deletions(-)