Skip to contentSkip to conversation

PATCH In review Jerome Glisse 6 participants 10 messages Nov 20, 2009 – Dec 4, 2009

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

This patch series add ttm range validation function. Aim is to include this in 2.6.33 so i have time to iron out issue, comments.

ttm:
I duplicated a bunch of ttm functions but now i think, best would be to add range to all function and use free list if range cover all the manager space. Doing so we might also be able to simplify mem_space alocation into a simpler function like ttm_bo_mem_space_range

radeon:
The second patch is a rework/cleanup of radeon object, it solves few issues along the way (i can't remember them now after fews days testing the patches). Biggest change is that we now rely on BO being validated before doing any change to radeon bo structure. As with any big patch i might introduce regressions, so far after testing on AGP:R1XX,R2XX,R3XX,R6XX PCIE:R3XX,R4XX,R5XX,R6XX,R7XX and RS480,RS690 i didn't found anythings obvious (test being X + glxgears + compiz(on hw which support it) + suspend/resume).

Last patch is smaller, it just use the interface introduced by the first patch.

#1 lore raw

PATCH 1/3 drm/ttm: Add range validation function

Add a function to validate buffer in a given range of memory. This is needed by some hw for some of their buffer (scanout buffer, cursor ...).

---
 drivers/gpu/drm/ttm/ttm_bo.c    |  305 ++++++++++++++++++++++++++++++++++++++-
 include/drm/ttm/ttm_bo_api.h    |    5 +
 include/drm/ttm/ttm_bo_driver.h |    1 +
 3 files changed, 310 insertions(+), 1 deletions(-)
drivers/gpu/drm/ttm/ttm_bo.c +304 -1
index 87c0625..6b0e7e8 100644
@@ -247,7 +247,6 @@ EXPORT_SYMBOL(ttm_bo_unreserve);
 /*
  * Call bo->mutex locked.
  */
-
 static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc)
 {
 	struct ttm_bo_device *bdev = bo->bdev;
@@ -418,6 +417,7 @@ static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo, bool remove_all)
 			kref_put(&bo->list_kref, ttm_bo_ref_bug);
 		}
 		if (bo->mem.mm_node) {
+			bo->mem.mm_node->private = NULL;
 			drm_mm_put_block(bo->mem.mm_node);
 			bo->mem.mm_node = NULL;
 		}
@@ -610,6 +610,7 @@ static int ttm_bo_evict(struct ttm_buffer_object *bo, unsigned mem_type,
 
 	spin_lock(&glob->lru_lock);
 	if (evict_mem.mm_node) {
+		evict_mem.mm_node->private = NULL;
 		drm_mm_put_block(evict_mem.mm_node);
 		evict_mem.mm_node = NULL;
 	}
@@ -826,6 +827,7 @@ int ttm_bo_mem_space(struct ttm_buffer_object *bo,
 		mem->mm_node = node;
 		mem->mem_type = mem_type;
 		mem->placement = cur_flags;
+		node->private = bo;
 		return 0;
 	}
 
@@ -856,6 +858,7 @@ int ttm_bo_mem_space(struct ttm_buffer_object *bo,
 
 		if (ret == 0 && mem->mm_node) {
 			mem->placement = cur_flags;
+			mem->mm_node->private = bo;
 			return 0;
 		}
 
@@ -868,6 +871,173 @@ int ttm_bo_mem_space(struct ttm_buffer_object *bo,
 }
 EXPORT_SYMBOL(ttm_bo_mem_space);
 
+static unsigned long ttm_bo_free_size_if_evicted(struct ttm_buffer_object *bo)
+{
+	struct ttm_mem_type_manager *man = &bo->bdev->man[bo->mem.mem_type];
+	struct drm_mm_node *node;
+	unsigned long size;
+
+	size = bo->mem.mm_node->size;
+	if (bo->mem.mm_node->ml_entry.prev != &man->manager.ml_entry) {
+		node = list_entry(bo->mem.mm_node->ml_entry.prev,
+					struct drm_mm_node, ml_entry);
+		if (node->free)
+			size += node->size;
+	}
+	if (bo->mem.mm_node->ml_entry.next != &man->manager.ml_entry) {
+		node = list_entry(bo->mem.mm_node->ml_entry.next,
+					struct drm_mm_node, ml_entry);
+		if (node->free)
+			size += node->size;
+	}
+	return size;
+}
+
+static void ttm_manager_evict_first(struct ttm_buffer_object *bo)
+{
+	struct ttm_mem_type_manager *man = &bo->bdev->man[bo->mem.mem_type];
+	unsigned long free_size_bo, free_size_bo_first;
+	struct ttm_buffer_object *bo_first;
+
+	/* BO is not on lru list, don't add it */
+	if (!list_empty(&bo->lru))
+		return;
+	bo_first = list_first_entry(&man->lru, struct ttm_buffer_object, lru);
+	free_size_bo = ttm_bo_free_size_if_evicted(bo);
+	free_size_bo_first = ttm_bo_free_size_if_evicted(bo_first);
+	if (free_size_bo > free_size_bo_first) {
+		list_del_init(&bo->lru);
+		list_add(&bo->lru, &man->lru);
+	}
+}
+
+static int ttm_manager_space_range(struct ttm_buffer_object *bo,
+					uint32_t mem_type,
+					struct ttm_mem_reg *mem,
+					unsigned long start, unsigned long end,
+					bool interruptible, bool no_wait)
+{
+	struct ttm_bo_global *glob = bo->glob;
+	struct drm_mm_node *entry;
+	struct ttm_mem_type_manager *man = &bo->bdev->man[mem_type];
+	struct drm_mm *mm = &man->manager;
+	unsigned size = end - start;
+	struct ttm_buffer_object *tbo;
+	unsigned wasted;
+	int ret;
+
+	mem->mm_node = NULL;
+	ret = drm_mm_pre_get(&man->manager);
+	if (unlikely(ret))
+		return ret;
+	spin_lock(&glob->lru_lock);
+	list_for_each_entry(entry, &mm->ml_entry, ml_entry) {
+		wasted = 0;
+		if (mem->page_alignment) {
+			unsigned tmp = entry->start % mem->page_alignment;
+			if (tmp)
+				wasted += mem->page_alignment - tmp;
+		}
+		if (entry->start < end && (entry->start+entry->size) > start) {
+			if (!entry->free) {
+				tbo = (struct ttm_buffer_object*)entry->private;
+				ttm_manager_evict_first(tbo);
+			} else {
+				if ((entry->size - wasted) <= size) {
+					/* Found a place */
+					entry = drm_mm_get_block_atomic(entry,
+							mem->num_pages,
+							mem->page_alignment);
+					mem->mm_node = entry;
+					mem->mem_type = mem_type;
+					entry->private = bo;
+					break;
+				}
+			}
+		}
+	}
+	spin_unlock(&glob->lru_lock);
+	return 0;
+}
+
+static int ttm_bo_mem_space_range(struct ttm_buffer_object *bo,
+				uint32_t proposed_placement,
+				struct ttm_mem_reg *mem,
+				unsigned long start, unsigned long end,
+				bool interruptible, bool no_wait)
+{
+	struct ttm_bo_device *bdev = bo->bdev;
+	struct ttm_bo_global *glob = bo->glob;
+	struct ttm_mem_type_manager *man;
+	uint32_t num_prios = bdev->driver->num_mem_type_prio;
+	const uint32_t *prios = bdev->driver->mem_type_prio;
+	uint32_t i;
+	uint32_t mem_type = TTM_PL_SYSTEM;
+	uint32_t cur_flags = 0;
+	bool type_found = false;
+	bool type_ok = false;
+	bool evicted;
+	int ret;
+
+	mem->mm_node = NULL;
+retry:
+	for (i = 0; i < num_prios; ++i) {
+		mem_type = prios[i];
+		man = &bdev->man[mem_type];
+		type_ok = ttm_bo_mt_compatible(man,
+						bo->type == ttm_bo_type_user,
+						mem_type, proposed_placement,
+						&cur_flags);
+		if (!type_ok)
+			continue;
+		if (start > (man->offset + man->size) || end < man->offset)
+			continue;
+		cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
+							cur_flags);
+		if (mem_type == TTM_PL_SYSTEM)
+			break;
+		if (man->has_type && man->use_type) {
+			type_found = true;
+			mem->placement = cur_flags;
+			ret = ttm_manager_space_range(bo, mem_type, mem,
+							start, end,
+							interruptible, no_wait);
+			if (ret)
+				return ret;
+		}
+		if (mem->mm_node)
+			break;
+	}
+	if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
+		mem->mem_type = mem_type;
+		mem->placement = cur_flags;
+		return 0;
+	}
+	for (i = 0, evicted = false; i < num_prios; ++i) {
+		struct ttm_buffer_object *tmp;
+
+		mem_type = prios[i];
+		man = &bdev->man[mem_type];
+		type_ok = ttm_bo_mt_compatible(man,
+						bo->type == ttm_bo_type_user,
+						mem_type, proposed_placement,
+						&cur_flags);
+		if (!type_ok)
+			continue;
+		if (start > (man->offset + man->size) || end < man->offset)
+			continue;
+		spin_lock(&glob->lru_lock);
+		tmp = list_first_entry(&man->lru, struct ttm_buffer_object, lru);
+		spin_unlock(&glob->lru_lock);
+		ret = ttm_bo_evict(tmp, mem_type, interruptible, no_wait);
+		if (unlikely(ret != 0))
+			return ret;
+	}
+	if (!evicted)
+		return -ENOMEM;
+	goto retry;
+}
+
 int ttm_bo_wait_cpu(struct ttm_buffer_object *bo, bool no_wait)
 {
 	int ret = 0;
@@ -925,6 +1095,7 @@ int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
 out_unlock:
 	if (ret && mem.mm_node) {
 		spin_lock(&glob->lru_lock);
+		mem.mm_node->private = NULL;
 		drm_mm_put_block(mem.mm_node);
 		spin_unlock(&glob->lru_lock);
 	}
@@ -998,6 +1169,137 @@ int ttm_buffer_object_validate(struct ttm_buffer_object *bo,
 }
 EXPORT_SYMBOL(ttm_buffer_object_validate);
 
+int ttm_bo_move_buffer_range(struct ttm_buffer_object *bo,
+				uint32_t proposed_placement,
+				unsigned long start, unsigned long end,
+				bool interruptible, bool no_wait)
+{
+	struct ttm_bo_global *glob = bo->glob;
+	int ret = 0;
+	struct ttm_mem_reg mem;
+
+	BUG_ON(!atomic_read(&bo->reserved));
+
+	/*
+	 * FIXME: It's possible to pipeline buffer moves.
+	 * Have the driver move function wait for idle when necessary,
+	 * instead of doing it here.
+	 */
+
+	spin_lock(&bo->lock);
+	ret = ttm_bo_wait(bo, false, interruptible, no_wait);
+	spin_unlock(&bo->lock);
+
+	if (ret)
+		return ret;
+
+	mem.num_pages = bo->num_pages;
+	mem.size = mem.num_pages << PAGE_SHIFT;
+	mem.page_alignment = bo->mem.page_alignment;
+
+	/*
+	 * Determine where to move the buffer.
+	 */
+
+	ret = ttm_bo_mem_space_range(bo, proposed_placement, &mem,
+					start, end, interruptible, no_wait);
+	if (ret)
+		goto out_unlock;
+
+	ret = ttm_bo_handle_move_mem(bo, &mem, false, interruptible, no_wait);
+
+out_unlock:
+	if (ret && mem.mm_node) {
+		spin_lock(&glob->lru_lock);
+		mem.mm_node->private = NULL;
+		drm_mm_put_block(mem.mm_node);
+		spin_unlock(&glob->lru_lock);
+	}
+	return ret;
+}
+
+static int ttm_bo_mem_compat_range(uint32_t proposed_placement,
+					struct ttm_mem_reg *mem,
+					unsigned long start, unsigned long end)
+{
+	if ((proposed_placement & mem->placement & TTM_PL_MASK_MEM) == 0)
+		return 0;
+	if ((proposed_placement & mem->placement & TTM_PL_MASK_CACHING) == 0)
+		return 0;
+	if (mem->mm_node) {
+		unsigned long eend = mem->mm_node->start + mem->mm_node->size;
+		if (mem->mm_node->start < start || eend > end)
+			return 0;
+	}
+
+	return 1;
+}
+
+int ttm_buffer_object_validate_range(struct ttm_buffer_object *bo,
+					uint32_t proposed_placement,
+					unsigned long start, unsigned long end,
+					bool interruptible, bool no_wait)
+{
+	int ret;
+
+	/* Convert to page */
+	start = start >> PAGE_SHIFT;
+	end = end >> PAGE_SHIFT;
+	/* Check that range is valid */
+	if (start > end || (end - start) < bo->num_pages)
+		return -EINVAL;
+
+	BUG_ON(!atomic_read(&bo->reserved));
+	bo->proposed_placement = proposed_placement;
+
+	TTM_DEBUG("Proposed placement 0x%08lx, Old flags 0x%08lx\n",
+		  (unsigned long)proposed_placement,
+		  (unsigned long)bo->mem.placement);
+
+	/*
+	 * Check whether we need to move buffer.
+	 */
+
+	if (!ttm_bo_mem_compat_range(bo->proposed_placement, &bo->mem,
+					start, end)) {
+		ret = ttm_bo_move_buffer_range(bo, bo->proposed_placement,
+						start, end,
+						interruptible, no_wait);
+		if (ret) {
+			if (ret != -ERESTART)
+				printk(KERN_ERR TTM_PFX
+					"Failed moving buffer. "
+					"Proposed placement 0x%08x\n",
+					bo->proposed_placement);
+			if (ret == -ENOMEM)
+				printk(KERN_ERR TTM_PFX
+					"Out of aperture space or "
+					"DRM memory quota.\n");
+			return ret;
+		}
+	}
+
+	/*
+	 * We might need to add a TTM.
+	 */
+	if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
+		ret = ttm_bo_add_ttm(bo, true);
+		if (ret)
+			return ret;
+	}
+	/*
+	 * Validation has succeeded, move the access and other
+	 * non-mapping-related flag bits from the proposed flags to
+	 * the active flags
+	 */
+
+	ttm_flag_masked(&bo->mem.placement, bo->proposed_placement,
+			~TTM_PL_MASK_MEMTYPE);
+
+	return 0;
+}
+EXPORT_SYMBOL(ttm_buffer_object_validate_range);
+
 int
 ttm_bo_check_placement(struct ttm_buffer_object *bo,
 		       uint32_t set_flags, uint32_t clr_flags)
@@ -1321,6 +1623,7 @@ int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
 	man->has_type = true;
 	man->use_type = true;
 	man->size = p_size;
+	man->offset = p_offset;
 
 	INIT_LIST_HEAD(&man->lru);
 
include/drm/ttm/ttm_bo_api.h +5 -0
index 4911461..3756970 100644
@@ -308,6 +308,11 @@ extern int ttm_bo_wait(struct ttm_buffer_object *bo, bool lazy,
 extern int ttm_buffer_object_validate(struct ttm_buffer_object *bo,
 				      uint32_t proposed_placement,
 				      bool interruptible, bool no_wait);
+extern int ttm_buffer_object_validate_range(struct ttm_buffer_object *bo,
+					uint32_t proposed_placement,
+					unsigned long start, unsigned long end,
+					bool interruptible, bool no_wait);
+
 /**
  * ttm_bo_unref
  *
include/drm/ttm/ttm_bo_driver.h +1 -0
index e8cd6d2..f00ba12 100644
@@ -224,6 +224,7 @@ struct ttm_mem_type_manager {
 	unsigned long gpu_offset;
 	unsigned long io_offset;
 	unsigned long io_size;
+	unsigned long offset;
 	void *io_addr;
 	uint64_t size;
 	uint32_t available_caching;
Signature · 1 line
1.6.5.2
#2 lore raw

PATCH 2/3 drm/radeon/kms: Rework radeon object handling

The locking & protection of radeon object was somewhat messy. This patch completely rework it to now use ttm reserve as a protection for the radeon object structure member. It also shrink down the various radeon object structure by removing field which were redondant with the ttm information. Last it converts few simple functions to inline which should with performances.

---
 drivers/gpu/drm/radeon/atombios_crtc.c      |   30 +-
 drivers/gpu/drm/radeon/r100.c               |   90 +++--
 drivers/gpu/drm/radeon/r100_track.h         |   10 +-
 drivers/gpu/drm/radeon/r300.c               |   15 +-
 drivers/gpu/drm/radeon/r420.c               |    4 +-
 drivers/gpu/drm/radeon/r520.c               |    2 +-
 drivers/gpu/drm/radeon/r600.c               |   65 +++-
 drivers/gpu/drm/radeon/r600_blit_kms.c      |   30 +-
 drivers/gpu/drm/radeon/radeon.h             |  111 +++---
 drivers/gpu/drm/radeon/radeon_benchmark.c   |   36 ++-
 drivers/gpu/drm/radeon/radeon_cs.c          |   13 +-
 drivers/gpu/drm/radeon/radeon_device.c      |   16 +-
 drivers/gpu/drm/radeon/radeon_fb.c          |   63 ++--
 drivers/gpu/drm/radeon/radeon_gart.c        |   42 ++-
 drivers/gpu/drm/radeon/radeon_gem.c         |   90 +++---
 drivers/gpu/drm/radeon/radeon_legacy_crtc.c |   27 +-
 drivers/gpu/drm/radeon/radeon_object.c      |  538 +++++++++++----------------
 drivers/gpu/drm/radeon/radeon_object.h      |  157 +++++++-
 drivers/gpu/drm/radeon/radeon_ring.c        |   67 ++--
 drivers/gpu/drm/radeon/radeon_test.c        |   55 ++--
 drivers/gpu/drm/radeon/radeon_ttm.c         |   36 ++-
 drivers/gpu/drm/radeon/rs400.c              |    4 +-
 drivers/gpu/drm/radeon/rs600.c              |   15 +-
 drivers/gpu/drm/radeon/rs690.c              |    4 +-
 drivers/gpu/drm/radeon/rv515.c              |    6 +-
 drivers/gpu/drm/radeon/rv770.c              |   30 ++-
 26 files changed, 873 insertions(+), 683 deletions(-)
drivers/gpu/drm/radeon/atombios_crtc.c +20 -10
index c15287a..80d7f46 100644
@@ -574,21 +574,31 @@ int atombios_crtc_set_base(struct drm_crtc *crtc, int x, int y,
 	struct radeon_device *rdev = dev->dev_private;
 	struct radeon_framebuffer *radeon_fb;
 	struct drm_gem_object *obj;
-	struct drm_radeon_gem_object *obj_priv;
+	struct radeon_bo *rbo;
 	uint64_t fb_location;
 	uint32_t fb_format, fb_pitch_pixels, tiling_flags;
+	int r;
 
 	if (!crtc->fb)
 		return -EINVAL;
 
 	radeon_fb = to_radeon_framebuffer(crtc->fb);
 
+	/* Pin framebuffer & get tilling informations */
 	obj = radeon_fb->obj;
-	obj_priv = obj->driver_private;
-
-	if (radeon_gem_object_pin(obj, RADEON_GEM_DOMAIN_VRAM, &fb_location)) {
+	rbo = obj->driver_private;
+	r = radeon_bo_reserve(rbo, false);
+	if (unlikely(r != 0))
+		return r;
+	r = radeon_bo_pin(rbo, RADEON_GEM_DOMAIN_VRAM, &fb_location);
+	if (unlikely(r != 0)) {
+		radeon_bo_unreserve(rbo);
 		return -EINVAL;
 	}
+	radeon_bo_get_tiling_flags(rbo, &tiling_flags, NULL);
+	radeon_bo_unreserve(rbo);
+	if (tiling_flags & RADEON_TILING_MACRO)
+		fb_format |= AVIVO_D1GRPH_MACRO_ADDRESS_MODE;
 
 	switch (crtc->fb->bits_per_pixel) {
 	case 8:
@@ -618,11 +628,6 @@ int atombios_crtc_set_base(struct drm_crtc *crtc, int x, int y,
 		return -EINVAL;
 	}
 
-	radeon_object_get_tiling_flags(obj->driver_private,
-				       &tiling_flags, NULL);
-	if (tiling_flags & RADEON_TILING_MACRO)
-		fb_format |= AVIVO_D1GRPH_MACRO_ADDRESS_MODE;
-
 	if (tiling_flags & RADEON_TILING_MICRO)
 		fb_format |= AVIVO_D1GRPH_TILED;
 
@@ -674,7 +679,12 @@ int atombios_crtc_set_base(struct drm_crtc *crtc, int x, int y,
 
 	if (old_fb && old_fb != crtc->fb) {
 		radeon_fb = to_radeon_framebuffer(old_fb);
-		radeon_gem_object_unpin(radeon_fb->obj);
+		rbo = radeon_fb->obj->driver_private;
+		r = radeon_bo_reserve(rbo, false);
+		if (unlikely(r != 0))
+			return r;
+		radeon_bo_unpin(rbo);
+		radeon_bo_unreserve(rbo);
 	}
 
 	/* Bytes per pixel may have changed */
drivers/gpu/drm/radeon/r100.c +52 -38

191 lines not on this page · raw message

drivers/gpu/drm/radeon/r100_track.h +5 -5
index 0daf0d7..ca50903 100644
@@ -10,26 +10,26 @@
  * CS functions
  */
 struct r100_cs_track_cb {
-	struct radeon_object	*robj;
+	struct radeon_bo	*robj;
 	unsigned		pitch;
 	unsigned		cpp;
 	unsigned		offset;
 };
 
 struct r100_cs_track_array {
-	struct radeon_object	*robj;
+	struct radeon_bo	*robj;
 	unsigned		esize;
 };
 
 struct r100_cs_cube_info {
-	struct radeon_object	*robj;
-	unsigned                offset;
+	struct radeon_bo	*robj;
+	unsigned		offset;
 	unsigned		width;
 	unsigned		height;
 };
 
 struct r100_cs_track_texture {
-	struct radeon_object	*robj;
+	struct radeon_bo	*robj;
 	struct r100_cs_cube_info cube_info[5]; /* info for 5 non-primary faces */
 	unsigned		pitch;
 	unsigned		width;
drivers/gpu/drm/radeon/r300.c +10 -5
index 2f43ee8..b76fe28 100644
@@ -137,14 +137,19 @@ int rv370_pcie_gart_enable(struct radeon_device *rdev)
 
 void rv370_pcie_gart_disable(struct radeon_device *rdev)
 {
-	uint32_t tmp;
+	u32 tmp;
+	int r;
 
 	tmp = RREG32_PCIE(RADEON_PCIE_TX_GART_CNTL);
 	tmp |= RADEON_PCIE_TX_GART_UNMAPPED_ACCESS_DISCARD;
 	WREG32_PCIE(RADEON_PCIE_TX_GART_CNTL, tmp & ~RADEON_PCIE_TX_GART_EN);
 	if (rdev->gart.table.vram.robj) {
-		radeon_object_kunmap(rdev->gart.table.vram.robj);
-		radeon_object_unpin(rdev->gart.table.vram.robj);
+		r = radeon_bo_reserve(rdev->gart.table.vram.robj, false);
+		if (likely(r == 0)) {
+			radeon_bo_kunmap(rdev->gart.table.vram.robj);
+			radeon_bo_unpin(rdev->gart.table.vram.robj);
+			radeon_bo_unreserve(rdev->gart.table.vram.robj);
+		}
 	}
 }
 
@@ -1265,7 +1270,7 @@ void r300_fini(struct radeon_device *rdev)
 		r100_pci_gart_fini(rdev);
 	radeon_irq_kms_fini(rdev);
 	radeon_fence_driver_fini(rdev);
-	radeon_object_fini(rdev);
+	radeon_bo_fini(rdev);
 	radeon_atombios_fini(rdev);
 	kfree(rdev->bios);
 	rdev->bios = NULL;
@@ -1325,7 +1330,7 @@ int r300_init(struct radeon_device *rdev)
 	if (r)
 		return r;
 	/* Memory manager */
-	r = radeon_object_init(rdev);
+	r = radeon_bo_init(rdev);
 	if (r)
 		return r;
 	if (rdev->flags & RADEON_IS_PCIE) {
drivers/gpu/drm/radeon/r420.c +2 -2
index 1cefdbc..dc5e93a 100644
@@ -258,7 +258,7 @@ void r420_fini(struct radeon_device *rdev)
 	radeon_agp_fini(rdev);
 	radeon_irq_kms_fini(rdev);
 	radeon_fence_driver_fini(rdev);
-	radeon_object_fini(rdev);
+	radeon_bo_fini(rdev);
 	if (rdev->is_atom_bios) {
 		radeon_atombios_fini(rdev);
 	} else {
@@ -331,7 +331,7 @@ int r420_init(struct radeon_device *rdev)
 		return r;
 	}
 	/* Memory manager */
-	r = radeon_object_init(rdev);
+	r = radeon_bo_init(rdev);
 	if (r) {
 		return r;
 	}
drivers/gpu/drm/radeon/r520.c +1 -1
index f743518..a41e9c7 100644
@@ -277,7 +277,7 @@ int r520_init(struct radeon_device *rdev)
 	if (r)
 		return r;
 	/* Memory manager */
-	r = radeon_object_init(rdev);
+	r = radeon_bo_init(rdev);
 	if (r)
 		return r;
 	r = rv370_pcie_gart_init(rdev);
drivers/gpu/drm/radeon/r600.c +46 -19

137 lines not on this page · raw message

drivers/gpu/drm/radeon/r600_blit_kms.c +19 -11
index dbf716e..f42943a 100644
@@ -473,9 +473,8 @@ int r600_blit_init(struct radeon_device *rdev)
 	obj_size += r6xx_ps_size * 4;
 	obj_size = ALIGN(obj_size, 256);
 
-	r = radeon_object_create(rdev, NULL, obj_size,
-				 true, RADEON_GEM_DOMAIN_VRAM,
-				 false, &rdev->r600_blit.shader_obj);
+	r = radeon_bo_create(rdev, NULL, obj_size, true, RADEON_GEM_DOMAIN_VRAM,
+				&rdev->r600_blit.shader_obj);
 	if (r) {
 		DRM_ERROR("r600 failed to allocate shader\n");
 		return r;
@@ -485,12 +484,14 @@ int r600_blit_init(struct radeon_device *rdev)
 		  obj_size,
 		  rdev->r600_blit.vs_offset, rdev->r600_blit.ps_offset);
 
-	r = radeon_object_kmap(rdev->r600_blit.shader_obj, &ptr);
+	r = radeon_bo_reserve(rdev->r600_blit.shader_obj, false);
+	if (unlikely(r != 0))
+		return r;
+	r = radeon_bo_kmap(rdev->r600_blit.shader_obj, &ptr);
 	if (r) {
 		DRM_ERROR("failed to map blit object %d\n", r);
 		return r;
 	}
-
 	if (rdev->family >= CHIP_RV770)
 		memcpy_toio(ptr + rdev->r600_blit.state_offset,
 			    r7xx_default_state, rdev->r600_blit.state_len * 4);
@@ -500,19 +501,26 @@ int r600_blit_init(struct radeon_device *rdev)
 	if (num_packet2s)
 		memcpy_toio(ptr + rdev->r600_blit.state_offset + (rdev->r600_blit.state_len * 4),
 			    packet2s, num_packet2s * 4);
-
-
 	memcpy(ptr + rdev->r600_blit.vs_offset, r6xx_vs, r6xx_vs_size * 4);
 	memcpy(ptr + rdev->r600_blit.ps_offset, r6xx_ps, r6xx_ps_size * 4);
-
-	radeon_object_kunmap(rdev->r600_blit.shader_obj);
+	radeon_bo_kunmap(rdev->r600_blit.shader_obj);
+	radeon_bo_unreserve(rdev->r600_blit.shader_obj);
 	return 0;
 }
 
 void r600_blit_fini(struct radeon_device *rdev)
 {
-	radeon_object_unpin(rdev->r600_blit.shader_obj);
-	radeon_object_unref(&rdev->r600_blit.shader_obj);
+	int r;
+
+	r = radeon_bo_reserve(rdev->r600_blit.shader_obj, false);
+	if (unlikely(r != 0)) {
+		dev_err(rdev->dev, "(%d) can't finish r600 blit\n", r);
+		goto out_unref;
+	}
+	radeon_bo_unpin(rdev->r600_blit.shader_obj);
+	radeon_bo_unreserve(rdev->r600_blit.shader_obj);
+out_unref:
+	radeon_bo_unref(&rdev->r600_blit.shader_obj);
 }
 
 int r600_vb_ib_get(struct radeon_device *rdev)
drivers/gpu/drm/radeon/radeon.h +52 -59

211 lines not on this page · raw message

drivers/gpu/drm/radeon/radeon_benchmark.c +26 -10
index 10bd50a..4ddfd4b 100644
@@ -29,8 +29,8 @@
 void radeon_benchmark_move(struct radeon_device *rdev, unsigned bsize,
 			   unsigned sdomain, unsigned ddomain)
 {
-	struct radeon_object *dobj = NULL;
-	struct radeon_object *sobj = NULL;
+	struct radeon_bo *dobj = NULL;
+	struct radeon_bo *sobj = NULL;
 	struct radeon_fence *fence = NULL;
 	uint64_t saddr, daddr;
 	unsigned long start_jiffies;
@@ -41,19 +41,27 @@ void radeon_benchmark_move(struct radeon_device *rdev, unsigned bsize,
 
 	size = bsize;
 	n = 1024;
-	r = radeon_object_create(rdev, NULL, size, true, sdomain, false, &sobj);
+	r = radeon_bo_create(rdev, NULL, size, true, sdomain, &sobj);
 	if (r) {
 		goto out_cleanup;
 	}
-	r = radeon_object_pin(sobj, sdomain, &saddr);
+	r = radeon_bo_reserve(sobj, false);
+	if (unlikely(r != 0))
+		goto out_cleanup;
+	r = radeon_bo_pin(sobj, sdomain, &saddr);
+	radeon_bo_unreserve(sobj);
 	if (r) {
 		goto out_cleanup;
 	}
-	r = radeon_object_create(rdev, NULL, size, true, ddomain, false, &dobj);
+	r = radeon_bo_create(rdev, NULL, size, true, ddomain, &dobj);
 	if (r) {
 		goto out_cleanup;
 	}
-	r = radeon_object_pin(dobj, ddomain, &daddr);
+	r = radeon_bo_reserve(dobj, false);
+	if (unlikely(r != 0))
+		goto out_cleanup;
+	r = radeon_bo_pin(dobj, ddomain, &daddr);
+	radeon_bo_unreserve(dobj);
 	if (r) {
 		goto out_cleanup;
 	}
@@ -109,12 +117,20 @@ void radeon_benchmark_move(struct radeon_device *rdev, unsigned bsize,
 	}
 out_cleanup:
 	if (sobj) {
-		radeon_object_unpin(sobj);
-		radeon_object_unref(&sobj);
+		r = radeon_bo_reserve(sobj, false);
+		if (likely(r == 0)) {
+			radeon_bo_unpin(sobj);
+			radeon_bo_unreserve(sobj);
+		}
+		radeon_bo_unref(&sobj);
 	}
 	if (dobj) {
-		radeon_object_unpin(dobj);
-		radeon_object_unref(&dobj);
+		r = radeon_bo_reserve(dobj, false);
+		if (likely(r == 0)) {
+			radeon_bo_unpin(dobj);
+			radeon_bo_unreserve(dobj);
+		}
+		radeon_bo_unref(&dobj);
 	}
 	if (fence) {
 		radeon_fence_unref(&fence);
drivers/gpu/drm/radeon/radeon_cs.c +7 -6
index 5ab2cf9..65590a0 100644
@@ -76,17 +76,17 @@ int radeon_cs_parser_relocs(struct radeon_cs_parser *p)
 			}
 			p->relocs_ptr[i] = &p->relocs[i];
 			p->relocs[i].robj = p->relocs[i].gobj->driver_private;
-			p->relocs[i].lobj.robj = p->relocs[i].robj;
+			p->relocs[i].lobj.bo = p->relocs[i].robj;
 			p->relocs[i].lobj.rdomain = r->read_domains;
 			p->relocs[i].lobj.wdomain = r->write_domain;
 			p->relocs[i].handle = r->handle;
 			p->relocs[i].flags = r->flags;
 			INIT_LIST_HEAD(&p->relocs[i].lobj.list);
-			radeon_object_list_add_object(&p->relocs[i].lobj,
-						      &p->validated);
+			radeon_bo_list_add_object(&p->relocs[i].lobj,
+						&p->validated);
 		}
 	}
-	return radeon_object_list_validate(&p->validated, p->ib->fence);
+	return radeon_bo_list_validate(&p->validated, p->ib->fence);
 }
 
 int radeon_cs_parser_init(struct radeon_cs_parser *p, void *data)
@@ -190,9 +190,10 @@ static void radeon_cs_parser_fini(struct radeon_cs_parser *parser, int error)
 	unsigned i;
 
 	if (error) {
-		radeon_object_list_unvalidate(&parser->validated);
+		radeon_bo_list_unvalidate(&parser->validated,
+						parser->ib->fence);
 	} else {
-		radeon_object_list_clean(&parser->validated);
+		radeon_bo_list_unreserve(&parser->validated);
 	}
 	for (i = 0; i < parser->nrelocs; i++) {
 		if (parser->relocs[i].gobj) {
drivers/gpu/drm/radeon/radeon_device.c +11 -5
index 88c1907..e13f723 100644
@@ -531,6 +531,7 @@ int radeon_device_init(struct radeon_device *rdev,
 	mutex_init(&rdev->cs_mutex);
 	mutex_init(&rdev->ib_pool.mutex);
 	mutex_init(&rdev->cp.mutex);
+	mutex_init(&rdev->gem.mutex);
 	rwlock_init(&rdev->fence_drv.lock);
 	INIT_LIST_HEAD(&rdev->gem.objects);
 
@@ -614,6 +615,7 @@ int radeon_suspend_kms(struct drm_device *dev, pm_message_t state)
 {
 	struct radeon_device *rdev = dev->dev_private;
 	struct drm_crtc *crtc;
+	int r;
 
 	if (dev == NULL || rdev == NULL) {
 		return -ENODEV;
@@ -624,18 +626,22 @@ int radeon_suspend_kms(struct drm_device *dev, pm_message_t state)
 	/* unpin the front buffers */
 	list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
 		struct radeon_framebuffer *rfb = to_radeon_framebuffer(crtc->fb);
-		struct radeon_object *robj;
+		struct radeon_bo *robj;
 
 		if (rfb == NULL || rfb->obj == NULL) {
 			continue;
 		}
 		robj = rfb->obj->driver_private;
-		if (robj != rdev->fbdev_robj) {
-			radeon_object_unpin(robj);
+		if (robj != rdev->fbdev_rbo) {
+			r = radeon_bo_reserve(robj, false);
+			if (unlikely(r == 0)) {
+				radeon_bo_unpin(robj);
+				radeon_bo_unreserve(robj);
+			}
 		}
 	}
 	/* evict vram memory */
-	radeon_object_evict_vram(rdev);
+	radeon_bo_evict_vram(rdev);
 	/* wait for gpu to finish processing current batch */
 	radeon_fence_wait_last(rdev);
 
@@ -643,7 +649,7 @@ int radeon_suspend_kms(struct drm_device *dev, pm_message_t state)
 
 	radeon_suspend(rdev);
 	/* evict remaining vram memory */
-	radeon_object_evict_vram(rdev);
+	radeon_bo_evict_vram(rdev);
 
 	pci_save_state(dev->pdev);
 	if (state.event == PM_EVENT_SUSPEND) {
drivers/gpu/drm/radeon/radeon_fb.c +39 -24

125 lines not on this page · raw message

drivers/gpu/drm/radeon/radeon_gart.c +23 -19

66 lines not on this page · raw message

drivers/gpu/drm/radeon/radeon_gem.c +45 -45

214 lines not on this page · raw message

drivers/gpu/drm/radeon/radeon_legacy_crtc.c +21 -6
index 8d0b7aa..e8a984d 100644
@@ -400,12 +400,14 @@ int radeon_crtc_set_base(struct drm_crtc *crtc, int x, int y,
 	struct radeon_crtc *radeon_crtc = to_radeon_crtc(crtc);
 	struct radeon_framebuffer *radeon_fb;
 	struct drm_gem_object *obj;
+	struct radeon_bo *rbo;
 	uint64_t base;
 	uint32_t crtc_offset, crtc_offset_cntl, crtc_tile_x0_y0 = 0;
 	uint32_t crtc_pitch, pitch_pixels;
 	uint32_t tiling_flags;
 	int format;
 	uint32_t gen_cntl_reg, gen_cntl_val;
+	int r;
 
 	DRM_DEBUG("\n");
 
@@ -431,10 +433,22 @@ int radeon_crtc_set_base(struct drm_crtc *crtc, int x, int y,
 		return false;
 	}
 
+	/* Pin framebuffer & get tilling informations */
 	obj = radeon_fb->obj;
-	if (radeon_gem_object_pin(obj, RADEON_GEM_DOMAIN_VRAM, &base)) {
+	rbo = obj->driver_private;
+	r = radeon_bo_reserve(rbo, false);
+	if (unlikely(r != 0))
+		return r;
+	r = radeon_bo_pin(rbo, RADEON_GEM_DOMAIN_VRAM, &base);
+	if (unlikely(r != 0)) {
+		radeon_bo_unreserve(rbo);
 		return -EINVAL;
 	}
+	radeon_bo_get_tiling_flags(rbo, &tiling_flags, NULL);
+	radeon_bo_unreserve(rbo);
+	if (tiling_flags & RADEON_TILING_MICRO)
+		DRM_ERROR("trying to scanout microtiled buffer\n");
+
 	/* if scanout was in GTT this really wouldn't work */
 	/* crtc offset is from display base addr not FB location */
 	radeon_crtc->legacy_display_base_addr = rdev->mc.vram_location;
@@ -449,10 +463,6 @@ int radeon_crtc_set_base(struct drm_crtc *crtc, int x, int y,
 		       (crtc->fb->bits_per_pixel * 8));
 	crtc_pitch |= crtc_pitch << 16;
 
-	radeon_object_get_tiling_flags(obj->driver_private,
-				       &tiling_flags, NULL);
-	if (tiling_flags & RADEON_TILING_MICRO)
-		DRM_ERROR("trying to scanout microtiled buffer\n");
 
 	if (tiling_flags & RADEON_TILING_MACRO) {
 		if (ASIC_IS_R300(rdev))
@@ -530,7 +540,12 @@ int radeon_crtc_set_base(struct drm_crtc *crtc, int x, int y,
 
 	if (old_fb && old_fb != crtc->fb) {
 		radeon_fb = to_radeon_framebuffer(old_fb);
-		radeon_gem_object_unpin(radeon_fb->obj);
+		rbo = radeon_fb->obj->driver_private;
+		r = radeon_bo_reserve(rbo, false);
+		if (unlikely(r != 0))
+			return r;
+		radeon_bo_unpin(rbo);
+		radeon_bo_unreserve(rbo);
 	}
 
 	/* Bytes per pixel may have changed */
drivers/gpu/drm/radeon/radeon_object.c +213 -325

777 lines not on this page · raw message

drivers/gpu/drm/radeon/radeon_object.h +145 -12

164 lines not on this page · raw message

drivers/gpu/drm/radeon/radeon_ring.c +42 -25

115 lines not on this page · raw message

drivers/gpu/drm/radeon/radeon_test.c +32 -23

142 lines not on this page · raw message

drivers/gpu/drm/radeon/radeon_ttm.c +23 -13
index f489c0d..6b6b459 100644
@@ -150,7 +150,7 @@ static int radeon_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
 		man->default_caching = TTM_PL_FLAG_CACHED;
 		break;
 	case TTM_PL_TT:
-		man->gpu_offset = 0;
+		man->gpu_offset = rdev->mc.gtt_location;
 		man->available_caching = TTM_PL_MASK_CACHING;
 		man->default_caching = TTM_PL_FLAG_CACHED;
 		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE | TTM_MEMTYPE_FLAG_CMA;
@@ -180,7 +180,7 @@ static int radeon_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
 		break;
 	case TTM_PL_VRAM:
 		/* "On-card" video ram */
-		man->gpu_offset = 0;
+		man->gpu_offset = rdev->mc.vram_location;
 		man->flags = TTM_MEMTYPE_FLAG_FIXED |
 			     TTM_MEMTYPE_FLAG_NEEDS_IOREMAP |
 			     TTM_MEMTYPE_FLAG_MAPPABLE;
@@ -482,27 +482,31 @@ int radeon_ttm_init(struct radeon_device *rdev)
 		DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
 		return r;
 	}
-	r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_VRAM, 0,
-			   ((rdev->mc.real_vram_size) >> PAGE_SHIFT));
+	r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_VRAM,
+				0, rdev->mc.real_vram_size >> PAGE_SHIFT);
 	if (r) {
 		DRM_ERROR("Failed initializing VRAM heap.\n");
 		return r;
 	}
-	r = radeon_object_create(rdev, NULL, 256 * 1024, true,
-				 RADEON_GEM_DOMAIN_VRAM, false,
-				 &rdev->stollen_vga_memory);
+	r = radeon_bo_create(rdev, NULL, 256 * 1024, true,
+				RADEON_GEM_DOMAIN_VRAM,
+				&rdev->stollen_vga_memory);
 	if (r) {
 		return r;
 	}
-	r = radeon_object_pin(rdev->stollen_vga_memory, RADEON_GEM_DOMAIN_VRAM, NULL);
+	r = radeon_bo_reserve(rdev->stollen_vga_memory, false);
+	if (r)
+		return r;
+	r = radeon_bo_pin(rdev->stollen_vga_memory, RADEON_GEM_DOMAIN_VRAM, NULL);
+	radeon_bo_unreserve(rdev->stollen_vga_memory);
 	if (r) {
-		radeon_object_unref(&rdev->stollen_vga_memory);
+		radeon_bo_unref(&rdev->stollen_vga_memory);
 		return r;
 	}
 	DRM_INFO("radeon: %uM of VRAM memory ready\n",
 		 (unsigned)rdev->mc.real_vram_size / (1024 * 1024));
-	r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_TT, 0,
-			   ((rdev->mc.gtt_size) >> PAGE_SHIFT));
+	r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_TT,
+				0, rdev->mc.gtt_size >> PAGE_SHIFT);
 	if (r) {
 		DRM_ERROR("Failed initializing GTT heap.\n");
 		return r;
@@ -523,9 +527,15 @@ int radeon_ttm_init(struct radeon_device *rdev)
 
 void radeon_ttm_fini(struct radeon_device *rdev)
 {
+	int r;
+
 	if (rdev->stollen_vga_memory) {
-		radeon_object_unpin(rdev->stollen_vga_memory);
-		radeon_object_unref(&rdev->stollen_vga_memory);
+		r = radeon_bo_reserve(rdev->stollen_vga_memory, false);
+		if (r == 0) {
+			radeon_bo_unpin(rdev->stollen_vga_memory);
+			radeon_bo_unreserve(rdev->stollen_vga_memory);
+		}
+		radeon_bo_unref(&rdev->stollen_vga_memory);
 	}
 	ttm_bo_clean_mm(&rdev->mman.bdev, TTM_PL_VRAM);
 	ttm_bo_clean_mm(&rdev->mman.bdev, TTM_PL_TT);
drivers/gpu/drm/radeon/rs400.c +2 -2
index ca03716..968016b 100644
@@ -452,7 +452,7 @@ void rs400_fini(struct radeon_device *rdev)
 	rs400_gart_fini(rdev);
 	radeon_irq_kms_fini(rdev);
 	radeon_fence_driver_fini(rdev);
-	radeon_object_fini(rdev);
+	radeon_bo_fini(rdev);
 	radeon_atombios_fini(rdev);
 	kfree(rdev->bios);
 	rdev->bios = NULL;
@@ -510,7 +510,7 @@ int rs400_init(struct radeon_device *rdev)
 	if (r)
 		return r;
 	/* Memory manager */
-	r = radeon_object_init(rdev);
+	r = radeon_bo_init(rdev);
 	if (r)
 		return r;
 	r = rs400_gart_init(rdev);
drivers/gpu/drm/radeon/rs600.c +10 -5
index 5f117cd..30dc65d 100644
@@ -146,15 +146,20 @@ int rs600_gart_enable(struct radeon_device *rdev)
 
 void rs600_gart_disable(struct radeon_device *rdev)
 {
-	uint32_t tmp;
+	u32 tmp;
+	int r;
 
 	/* FIXME: disable out of gart access */
 	WREG32_MC(R_000100_MC_PT0_CNTL, 0);
 	tmp = RREG32_MC(R_000009_MC_CNTL1);
 	WREG32_MC(R_000009_MC_CNTL1, tmp & C_000009_ENABLE_PAGE_TABLES);
 	if (rdev->gart.table.vram.robj) {
-		radeon_object_kunmap(rdev->gart.table.vram.robj);
-		radeon_object_unpin(rdev->gart.table.vram.robj);
+		r = radeon_bo_reserve(rdev->gart.table.vram.robj, false);
+		if (r == 0) {
+			radeon_bo_kunmap(rdev->gart.table.vram.robj);
+			radeon_bo_unpin(rdev->gart.table.vram.robj);
+			radeon_bo_unreserve(rdev->gart.table.vram.robj);
+		}
 	}
 }
 
@@ -445,7 +450,7 @@ void rs600_fini(struct radeon_device *rdev)
 	rs600_gart_fini(rdev);
 	radeon_irq_kms_fini(rdev);
 	radeon_fence_driver_fini(rdev);
-	radeon_object_fini(rdev);
+	radeon_bo_fini(rdev);
 	radeon_atombios_fini(rdev);
 	kfree(rdev->bios);
 	rdev->bios = NULL;
@@ -505,7 +510,7 @@ int rs600_init(struct radeon_device *rdev)
 	if (r)
 		return r;
 	/* Memory manager */
-	r = radeon_object_init(rdev);
+	r = radeon_bo_init(rdev);
 	if (r)
 		return r;
 	r = rs600_gart_init(rdev);
drivers/gpu/drm/radeon/rs690.c +2 -2
index 2754717..2d576af 100644
@@ -662,7 +662,7 @@ void rs690_fini(struct radeon_device *rdev)
 	rs400_gart_fini(rdev);
 	radeon_irq_kms_fini(rdev);
 	radeon_fence_driver_fini(rdev);
-	radeon_object_fini(rdev);
+	radeon_bo_fini(rdev);
 	radeon_atombios_fini(rdev);
 	kfree(rdev->bios);
 	rdev->bios = NULL;
@@ -723,7 +723,7 @@ int rs690_init(struct radeon_device *rdev)
 	if (r)
 		return r;
 	/* Memory manager */
-	r = radeon_object_init(rdev);
+	r = radeon_bo_init(rdev);
 	if (r)
 		return r;
 	r = rs400_gart_init(rdev);
drivers/gpu/drm/radeon/rv515.c +3 -3
index 7935f79..10b10c1 100644
@@ -541,11 +541,11 @@ void rv515_fini(struct radeon_device *rdev)
 	r100_wb_fini(rdev);
 	r100_ib_fini(rdev);
 	radeon_gem_fini(rdev);
-    rv370_pcie_gart_fini(rdev);
+	rv370_pcie_gart_fini(rdev);
 	radeon_agp_fini(rdev);
 	radeon_irq_kms_fini(rdev);
 	radeon_fence_driver_fini(rdev);
-	radeon_object_fini(rdev);
+	radeon_bo_fini(rdev);
 	radeon_atombios_fini(rdev);
 	kfree(rdev->bios);
 	rdev->bios = NULL;
@@ -604,7 +604,7 @@ int rv515_init(struct radeon_device *rdev)
 	if (r)
 		return r;
 	/* Memory manager */
-	r = radeon_object_init(rdev);
+	r = radeon_bo_init(rdev);
 	if (r)
 		return r;
 	r = rv370_pcie_gart_init(rdev);
drivers/gpu/drm/radeon/rv770.c +22 -8
index b0efd0d..c90c7c3 100644
@@ -92,7 +92,7 @@ int rv770_pcie_gart_enable(struct radeon_device *rdev)
 void rv770_pcie_gart_disable(struct radeon_device *rdev)
 {
 	u32 tmp;
-	int i;
+	int i, r;
 
 	/* Disable all tables */
 	for (i = 0; i < 7; i++)
@@ -113,8 +113,12 @@ void rv770_pcie_gart_disable(struct radeon_device *rdev)
 	WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp);
 	WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp);
 	if (rdev->gart.table.vram.robj) {
-		radeon_object_kunmap(rdev->gart.table.vram.robj);
-		radeon_object_unpin(rdev->gart.table.vram.robj);
+		r = radeon_bo_reserve(rdev->gart.table.vram.robj, false);
+		if (likely(r == 0)) {
+			radeon_bo_kunmap(rdev->gart.table.vram.robj);
+			radeon_bo_unpin(rdev->gart.table.vram.robj);
+			radeon_bo_unreserve(rdev->gart.table.vram.robj);
+		}
 	}
 }
 
@@ -880,8 +884,12 @@ static int rv770_startup(struct radeon_device *rdev)
 	}
 	rv770_gpu_init(rdev);
 
-	r = radeon_object_pin(rdev->r600_blit.shader_obj, RADEON_GEM_DOMAIN_VRAM,
-			      &rdev->r600_blit.shader_gpu_addr);
+	r = radeon_bo_reserve(rdev->r600_blit.shader_obj, false);
+	if (unlikely(r != 0))
+		return r;
+	r = radeon_bo_pin(rdev->r600_blit.shader_obj, RADEON_GEM_DOMAIN_VRAM,
+			&rdev->r600_blit.shader_gpu_addr);
+	radeon_bo_unreserve(rdev->r600_blit.shader_obj);
 	if (r) {
 		DRM_ERROR("failed to pin blit object %d\n", r);
 		return r;
@@ -934,13 +942,19 @@ int rv770_resume(struct radeon_device *rdev)
 
 int rv770_suspend(struct radeon_device *rdev)
 {
+	int r;
+
 	/* FIXME: we should wait for ring to be empty */
 	r700_cp_stop(rdev);
 	rdev->cp.ready = false;
 	r600_wb_disable(rdev);
 	rv770_pcie_gart_disable(rdev);
 	/* unpin shaders bo */
-        radeon_object_unpin(rdev->r600_blit.shader_obj);
+	r = radeon_bo_reserve(rdev->r600_blit.shader_obj, false);
+	if (likely(r == 0)) {
+		radeon_bo_unpin(rdev->r600_blit.shader_obj);
+		radeon_bo_unreserve(rdev->r600_blit.shader_obj);
+	}
 	return 0;
 }
 
@@ -998,7 +1012,7 @@ int rv770_init(struct radeon_device *rdev)
 	if (r)
 		return r;
 	/* Memory manager */
-	r = radeon_object_init(rdev);
+	r = radeon_bo_init(rdev);
 	if (r)
 		return r;
 	rdev->cp.ring_obj = NULL;
@@ -1059,7 +1073,7 @@ void rv770_fini(struct radeon_device *rdev)
 	radeon_clocks_fini(rdev);
 	if (rdev->flags & RADEON_IS_AGP)
 		radeon_agp_fini(rdev);
-	radeon_object_fini(rdev);
+	radeon_bo_fini(rdev);
 	radeon_atombios_fini(rdev);
 	kfree(rdev->bios);
 	rdev->bios = NULL;
Signature · 1 line
1.6.5.2
#3 lore raw

PATCH 3/3 drm/radeon/kms: Add range pinning to crtc/cursor bo

We force the crtc & cursor bo to be in the first 64M, this will help on legacy modesetting hw where the offset of scanout buffer and cursor are relative to a base address.

---
 drivers/gpu/drm/radeon/radeon.h             |    3 ++
 drivers/gpu/drm/radeon/radeon_cursor.c      |    3 +-
 drivers/gpu/drm/radeon/radeon_gem.c         |   15 +++++++++++
 drivers/gpu/drm/radeon/radeon_legacy_crtc.c |    6 ++++-
 drivers/gpu/drm/radeon/radeon_object.c      |   35 +++++++++++++++++++++++++++
 drivers/gpu/drm/radeon/radeon_object.h      |    2 +
 6 files changed, 62 insertions(+), 2 deletions(-)
drivers/gpu/drm/radeon/radeon.h +3 -0
index 4b5d86b..5cb1b79 100644
@@ -245,6 +245,9 @@ int radeon_gem_object_create(struct radeon_device *rdev, int size,
 				struct drm_gem_object **obj);
 int radeon_gem_object_pin(struct drm_gem_object *obj, uint32_t pin_domain,
 			  uint64_t *gpu_addr);
+extern int radeon_gem_object_pin_range(struct drm_gem_object *obj,
+				uint32_t pin_domain, uint64_t *gpu_addr,
+				unsigned long start, unsigned long end);
 void radeon_gem_object_unpin(struct drm_gem_object *obj);
 
 
drivers/gpu/drm/radeon/radeon_cursor.c +2 -1
index 28772a3..205fca1 100644
@@ -156,7 +156,8 @@ int radeon_crtc_cursor_set(struct drm_crtc *crtc,
 		return -EINVAL;
 	}
 
-	ret = radeon_gem_object_pin(obj, RADEON_GEM_DOMAIN_VRAM, &gpu_addr);
+	ret = radeon_gem_object_pin_range(obj, RADEON_GEM_DOMAIN_VRAM,
+					&gpu_addr, 0, 64 * 1024 * 1024);
 	if (ret)
 		goto fail;
 
drivers/gpu/drm/radeon/radeon_gem.c +15 -0
index f3fb2bf..7db8fcb 100644
@@ -92,6 +92,21 @@ int radeon_gem_object_pin(struct drm_gem_object *obj, uint32_t pin_domain,
 	return r;
 }
 
+int radeon_gem_object_pin_range(struct drm_gem_object *obj,
+				uint32_t pin_domain, uint64_t *gpu_addr,
+				unsigned long start, unsigned long end)
+{
+	struct radeon_bo *robj = obj->driver_private;
+	int r;
+
+	r = radeon_bo_reserve(robj, false);
+	if (unlikely(r != 0))
+		return r;
+	r = radeon_bo_pin_range(robj, pin_domain, gpu_addr, start, end);
+	radeon_bo_unreserve(robj);
+	return r;
+}
+
 void radeon_gem_object_unpin(struct drm_gem_object *obj)
 {
 	struct radeon_bo *robj = obj->driver_private;
drivers/gpu/drm/radeon/radeon_legacy_crtc.c +5 -1
index e8a984d..3a2fb68 100644
@@ -439,7 +439,11 @@ int radeon_crtc_set_base(struct drm_crtc *crtc, int x, int y,
 	r = radeon_bo_reserve(rbo, false);
 	if (unlikely(r != 0))
 		return r;
-	r = radeon_bo_pin(rbo, RADEON_GEM_DOMAIN_VRAM, &base);
+	/* Pin buffer in the first 64M and always set crtc_base to start of
+	 * vram
+	 */
+	r = radeon_bo_pin_range(rbo, RADEON_GEM_DOMAIN_VRAM, &base,
+				0, 64 * 1024 * 1024);
 	if (unlikely(r != 0)) {
 		radeon_bo_unreserve(rbo);
 		return -EINVAL;
drivers/gpu/drm/radeon/radeon_object.c +35 -0
index bec4943..15399e6 100644
@@ -200,6 +200,41 @@ retry:
 	return r;
 }
 
+int radeon_bo_pin_range(struct radeon_bo *bo, u32 domain, u64 *gpu_addr,
+			unsigned long start, unsigned long end)
+{
+	u32 flags;
+	u32 tmp;
+	int r;
+
+	flags = radeon_ttm_flags_from_domain(domain);
+	if (bo->pin_count) {
+		bo->pin_count++;
+		if (gpu_addr)
+			*gpu_addr = radeon_bo_gpu_offset(bo);
+		return 0;
+	}
+	tmp = bo->tbo.mem.placement;
+	ttm_flag_masked(&tmp, flags, TTM_PL_MASK_MEM);
+	bo->tbo.proposed_placement = tmp | TTM_PL_FLAG_NO_EVICT |
+					TTM_PL_MASK_CACHING;
+retry:
+	r = ttm_buffer_object_validate_range(&bo->tbo,
+						bo->tbo.proposed_placement,
+						start, end, true, false);
+	if (likely(r == 0)) {
+		bo->pin_count = 1;
+		if (gpu_addr != NULL)
+			*gpu_addr = radeon_bo_gpu_offset(bo);
+	}
+	if (unlikely(r != 0)) {
+		if (r == -ERESTART)
+			goto retry;
+		dev_err(bo->rdev->dev, "%p pin failed\n", bo);
+	}
+	return r;
+}
+
 int radeon_bo_unpin(struct radeon_bo *bo)
 {
 	int r;
drivers/gpu/drm/radeon/radeon_object.h +2 -0
index e9da130..a40bf67 100644
@@ -153,6 +153,8 @@ extern int radeon_bo_kmap(struct radeon_bo *bo, void **ptr);
 extern void radeon_bo_kunmap(struct radeon_bo *bo);
 extern void radeon_bo_unref(struct radeon_bo **bo);
 extern int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr);
+extern int radeon_bo_pin_range(struct radeon_bo *bo, u32 domain, u64 *gpu_addr,
+				unsigned long start, unsigned long end);
 extern int radeon_bo_unpin(struct radeon_bo *bo);
 extern int radeon_bo_evict_vram(struct radeon_device *rdev);
 extern void radeon_bo_force_delete(struct radeon_device *rdev);
Signature · 1 line
1.6.5.2
#4 lore raw

Re: [PATCH 3/3] drm/radeon/kms: Add range pinning to crtc/cursor bo

Quoted 3 lines · Jerome Glisse

We force the crtc & cursor bo to be in the first 64M, this will help on legacy modesetting hw where the offset of scanout buffer and cursor are relative to a base address.

This limitation only applies to pre-avivo (r1xx-r4xx) chips, there's no need on newer hardware. Also the cursor and crtc have to be within 128 MB of each other, so you could bump the limits to 128 MB. Another thing to consider down the road might be to treat memory over 128 MB as "invisible" on these pre-avivo chips (once we support that in general) then we no longer need this limit.

Alex

Quoted 156 lines

---
 drivers/gpu/drm/radeon/radeon.h             |    3 ++ drivers/gpu/drm/radeon/radeon_cursor.c      |    3 +- drivers/gpu/drm/radeon/radeon_gem.c         |   15 +++++++++++ drivers/gpu/drm/radeon/radeon_legacy_crtc.c |    6 ++++- drivers/gpu/drm/radeon/radeon_object.c      |   35 +++++++++++++++++++++++++++ drivers/gpu/drm/radeon/radeon_object.h      |    2 +

 6 files changed, 62 insertions(+), 2 deletions(-)
drivers/gpu/drm/radeon/radeon.h +3 -0
index 4b5d86b..5cb1b79 100644
@@ -245,6 +245,9 @@ int radeon_gem_object_create(struct radeon_device *rdev, int size,
                                struct drm_gem_object **obj);
  int radeon_gem_object_pin(struct drm_gem_object *obj, uint32_t pin_domain,
                          uint64_t *gpu_addr);
+extern int radeon_gem_object_pin_range(struct drm_gem_object *obj,
+                               uint32_t pin_domain, uint64_t *gpu_addr,
+                               unsigned long start, unsigned long end);
  void radeon_gem_object_unpin(struct drm_gem_object *obj);
 
 
drivers/gpu/drm/radeon/radeon_cursor.c +2 -1
index 28772a3..205fca1 100644
@@ -156,7 +156,8 @@ int radeon_crtc_cursor_set(struct drm_crtc *crtc,
                return -EINVAL;
        }
 
-       ret = radeon_gem_object_pin(obj, RADEON_GEM_DOMAIN_VRAM, &gpu_addr);
+       ret = radeon_gem_object_pin_range(obj, RADEON_GEM_DOMAIN_VRAM,
+                                       &gpu_addr, 0, 64 * 1024 * 1024);
        if (ret)
                goto fail;
 
drivers/gpu/drm/radeon/radeon_gem.c +15 -0
index f3fb2bf..7db8fcb 100644
@@ -92,6 +92,21 @@ int radeon_gem_object_pin(struct drm_gem_object *obj, uint32_t pin_domain,
        return r;
  }
 
+int radeon_gem_object_pin_range(struct drm_gem_object *obj,
+                               uint32_t pin_domain, uint64_t *gpu_addr,
+                               unsigned long start, unsigned long end)
+{
+       struct radeon_bo *robj = obj->driver_private;
+       int r;
+
+       r = radeon_bo_reserve(robj, false);
+       if (unlikely(r != 0))
+               return r;
+       r = radeon_bo_pin_range(robj, pin_domain, gpu_addr, start, end);
+       radeon_bo_unreserve(robj);
+       return r;
+}
+
  void radeon_gem_object_unpin(struct drm_gem_object *obj)
  {
        struct radeon_bo *robj = obj->driver_private;
drivers/gpu/drm/radeon/radeon_legacy_crtc.c +5 -1
index e8a984d..3a2fb68 100644
@@ -439,7 +439,11 @@ int radeon_crtc_set_base(struct drm_crtc *crtc, int x, int y,
        r = radeon_bo_reserve(rbo, false);
        if (unlikely(r != 0))
                return r;
-       r = radeon_bo_pin(rbo, RADEON_GEM_DOMAIN_VRAM, &base);
+       /* Pin buffer in the first 64M and always set crtc_base to start of
+        * vram
+        */
+       r = radeon_bo_pin_range(rbo, RADEON_GEM_DOMAIN_VRAM, &base,
+                               0, 64 * 1024 * 1024);
        if (unlikely(r != 0)) {
                radeon_bo_unreserve(rbo);
                return -EINVAL;
drivers/gpu/drm/radeon/radeon_object.c +35 -0
index bec4943..15399e6 100644
@@ -200,6 +200,41 @@ retry:
        return r;
  }
 
+int radeon_bo_pin_range(struct radeon_bo *bo, u32 domain, u64 *gpu_addr,
+                       unsigned long start, unsigned long end)
+{
+       u32 flags;
+       u32 tmp;
+       int r;
+
+       flags = radeon_ttm_flags_from_domain(domain);
+       if (bo->pin_count) {
+               bo->pin_count++;
+               if (gpu_addr)
+                       *gpu_addr = radeon_bo_gpu_offset(bo);
+               return 0;
+       }
+       tmp = bo->tbo.mem.placement;
+       ttm_flag_masked(&tmp, flags, TTM_PL_MASK_MEM);
+       bo->tbo.proposed_placement = tmp | TTM_PL_FLAG_NO_EVICT |
+                                       TTM_PL_MASK_CACHING;
+retry:
+       r = ttm_buffer_object_validate_range(&bo->tbo,
+                                               bo->tbo.proposed_placement,
+                                               start, end, true, false);
+       if (likely(r == 0)) {
+               bo->pin_count = 1;
+               if (gpu_addr != NULL)
+                       *gpu_addr = radeon_bo_gpu_offset(bo);
+       }
+       if (unlikely(r != 0)) {
+               if (r == -ERESTART)
+                       goto retry;
+               dev_err(bo->rdev->dev, "%p pin failed\n", bo);
+       }
+       return r;
+}
+
  int radeon_bo_unpin(struct radeon_bo *bo)
  {
        int r;
drivers/gpu/drm/radeon/radeon_object.h +2 -1
index e9da130..a40bf67 100644
@@ -153,6 +153,8 @@ extern int radeon_bo_kmap(struct radeon_bo *bo, void **ptr);
  extern void radeon_bo_kunmap(struct radeon_bo *bo);
  extern void radeon_bo_unref(struct radeon_bo **bo);
  extern int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr);
+extern int radeon_bo_pin_range(struct radeon_bo *bo, u32 domain, u64 *gpu_addr,
+                               unsigned long start, unsigned long end);
  extern int radeon_bo_unpin(struct radeon_bo *bo);
  extern int radeon_bo_evict_vram(struct radeon_device *rdev);
  extern void radeon_bo_force_delete(struct radeon_device *rdev);
--

1.6.5.2

------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day trial. Simplify your report design, integration and deployment - and focus on what you do best, core application coding. Discover what's new with Crystal Reports now.  http://p.sf.net/sfu/bobj-july
--
_______________________________________________
Dri-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/dri-devel

#5 lore raw

Re: [PATCH 3/3] drm/radeon/kms: Add range pinning to crtc/cursor bo

Quoted 5 lines · Alex Deucher
Quoted 3 lines · Jerome Glisse

We force the crtc & cursor bo to be in the first 64M, this will help on legacy modesetting hw where the offset of scanout buffer and cursor are relative to a base address.

Quoted 9 lines

This limitation only applies to pre-avivo (r1xx-r4xx) chips, there's no need on newer hardware. Also the cursor and crtc have to be within 128 MB of each other, so you could bump the limits to 128 MB. Another thing to consider down the road might be to treat memory over 128 MB as "invisible" on these pre-avivo chips (once we support that in general) then we no longer need this limit.

Alex

I choosed 64M in hope to disminish VRAM fragmentation a little, this is also why i did not bother to make this conditional for legacy modesetting hw. For supporting non visible vram see my other ttm thread. I am working on adding this knowledge to ttm.

Cheers,
Jerome

maintainer
#6 lore raw

Re: [PATCH 2/3] drm/radeon/kms: Rework radeon object handling

In reply to Jerome Glisse
Quoted 24 lines · Jerome Glisse

The locking & protection of radeon object was somewhat messy. This patch completely rework it to now use ttm reserve as a protection for the radeon object structure member. It also shrink down the various radeon object structure by removing field which were redondant with the ttm information. Last it converts few simple functions to inline which should with performances.

 int radeon_gem_busy_ioctl(struct drm_device *dev, void *data,

drivers/gpu/drm/radeon/radeon_gem.c +1 -1
@@ -264,7 +261,7 @@ int radeon_gem_busy_ioctl(struct drm_device *dev, void *data,
  {
        struct drm_radeon_gem_busy *args = data;
        struct drm_gem_object *gobj;
-       struct radeon_object *robj;
+       struct radeon_bo *robj;
        int r;
        uint32_t cur_placement;
 

drivers/gpu/drm/radeon/radeon_gem.c int radeon_gem_busy_ioctl(struct drm_device *dev, void *data, · line 273

Quoted 5 lines
drivers/gpu/drm/radeon/radeon_gem.c +1 -1
@@ -273,7 +270,7 @@ int radeon_gem_busy_ioctl(struct drm_device *dev, void *data,
                return -EINVAL;
        }
        robj = gobj->driver_private;
-       r = radeon_object_busy_domain(robj, &cur_placement);
+       r = radeon_bo_wait(robj, &cur_placement, true);

Here we pass an uninitialised cur_placement to busy ioctl.

Quoted 14 lines
drivers/gpu/drm/radeon/radeon_object.h +14 -0
+static inline int radeon_bo_wait(struct radeon_bo *bo, u32 *mem_type,
+                                       bool no_wait)
+{
+       int r;
+
+retry:
+       r = ttm_bo_reserve(&bo->tbo, true, no_wait, false, 0);
+       if (unlikely(r != 0)) {
+               if (r == -ERESTART)
+                       goto retry;
+               dev_err(bo->rdev->dev, "%p reserve failed for wait\n", bo);
+               return r;
+       }
+       spin_lock(&bo->tbo.lock);

drivers/gpu/drm/radeon/radeon_object.h line 146

Quoted 10 lines
drivers/gpu/drm/radeon/radeon_object.h +10 -0
+       if (mem_type)
+               *mem_type = bo->tbo.mem.mem_type;
+       if (bo->tbo.sync_obj)
+               r = ttm_bo_wait(&bo->tbo, true, true, no_wait);
+       spin_unlock(&bo->tbo.lock);
+       ttm_bo_unreserve(&bo->tbo);
+       if (unlikely(r == -ERESTART))
+               goto retry;
+       return r;
+}

If this return early mem_type never gets set to anything useful.

The compiler warns on this here.

Can you send a fix up patch as I've merged this to drm-radeon-next already.

Dave.

#7 lore raw

Re: [PATCH 1/3] drm/ttm: Add range validation function

In reply to Jerome Glisse

Jerome,
Please see comments inline.

Quoted 16 lines · Jerome Glisse

Add a function to validate buffer in a given range of memory. This is needed by some hw for some of their buffer (scanout buffer, cursor ...).

---
 drivers/gpu/drm/ttm/ttm_bo.c    |  305 ++++++++++++++++++++++++++++++++++++++-
 include/drm/ttm/ttm_bo_api.h    |    5 +
 include/drm/ttm/ttm_bo_driver.h |    1 +
 3 files changed, 310 insertions(+), 1 deletions(-)
drivers/gpu/drm/ttm/ttm_bo.c +0 -0
index 87c0625..6b0e7e8 100644

drivers/gpu/drm/ttm/ttm_bo.c EXPORT_SYMBOL(ttm_bo_unreserve); · line 250

Quoted 5 lines
drivers/gpu/drm/ttm/ttm_bo.c +0 -1
@@ -247,7 +247,6 @@ EXPORT_SYMBOL(ttm_bo_unreserve);
 /*
  * Call bo->mutex locked.
  */
-
 

Whitespace.

Quoted 22 lines
drivers/gpu/drm/ttm/ttm_bo.c +0 -0
 static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc)
 {
drivers/gpu/drm/ttm/ttm_bo.c +2 -0
 	struct ttm_bo_device *bdev = bo->bdev;
@@ -418,6 +417,7 @@ static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo, bool remove_all)
 			kref_put(&bo->list_kref, ttm_bo_ref_bug);
 		}
 		if (bo->mem.mm_node) {
+			bo->mem.mm_node->private = NULL;
 			drm_mm_put_block(bo->mem.mm_node);
 			bo->mem.mm_node = NULL;
 		}
@@ -610,6 +610,7 @@ static int ttm_bo_evict(struct ttm_buffer_object *bo, unsigned mem_type,
 
 	spin_lock(&glob->lru_lock);
 	if (evict_mem.mm_node) {
+		evict_mem.mm_node->private = NULL;
 		drm_mm_put_block(evict_mem.mm_node);
 		evict_mem.mm_node = NULL;
 	}

drivers/gpu/drm/ttm/ttm_bo.c int ttm_bo_mem_space(struct ttm_buffer_object *bo, · line 833

Quoted 7 lines
drivers/gpu/drm/ttm/ttm_bo.c +1 -0
@@ -826,6 +827,7 @@ int ttm_bo_mem_space(struct ttm_buffer_object *bo,
 		mem->mm_node = node;
 		mem->mem_type = mem_type;
 		mem->placement = cur_flags;
+		node->private = bo;
 		return 0;
 	}
 

I'd like a big warning here somewhere, because since the pointer to the bo is not refcounted, it may never be dereferenced outside of the lru spinlock, because as soon as you release the lru spinlock, someone else may destroy the buffer. The current usage is valid AFAICT. If you choose to refcount it, use bo::list_kref.

Quoted 43 lines
drivers/gpu/drm/ttm/ttm_bo.c +32 -0
@@ -856,6 +858,7 @@ int ttm_bo_mem_space(struct ttm_buffer_object *bo,
 
 		if (ret == 0 && mem->mm_node) {
 			mem->placement = cur_flags;
+			mem->mm_node->private = bo;
 			return 0;
 		}
 
@@ -868,6 +871,173 @@ int ttm_bo_mem_space(struct ttm_buffer_object *bo,
 }
 EXPORT_SYMBOL(ttm_bo_mem_space);
 
+static unsigned long ttm_bo_free_size_if_evicted(struct ttm_buffer_object *bo)
+{
+	struct ttm_mem_type_manager *man = &bo->bdev->man[bo->mem.mem_type];
+	struct drm_mm_node *node;
+	unsigned long size;
+
+	size = bo->mem.mm_node->size;
+	if (bo->mem.mm_node->ml_entry.prev != &man->manager.ml_entry) {
+		node = list_entry(bo->mem.mm_node->ml_entry.prev,
+					struct drm_mm_node, ml_entry);
+		if (node->free)
+			size += node->size;
+	}
+	if (bo->mem.mm_node->ml_entry.next != &man->manager.ml_entry) {
+		node = list_entry(bo->mem.mm_node->ml_entry.next,
+					struct drm_mm_node, ml_entry);
+		if (node->free)
+			size += node->size;
+	}
+	return size;
+}
+
+static void ttm_manager_evict_first(struct ttm_buffer_object *bo)
+{
+	struct ttm_mem_type_manager *man = &bo->bdev->man[bo->mem.mem_type];
+	unsigned long free_size_bo, free_size_bo_first;
+	struct ttm_buffer_object *bo_first;
+
+	/* BO is not on lru list, don't add it */
+	if (!list_empty(&bo->lru))
+		return;

drivers/gpu/drm/ttm/ttm_bo.c int ttm_bo_mem_space(struct ttm_buffer_object *bo, · line 914

Quoted 10 lines
drivers/gpu/drm/ttm/ttm_bo.c +9 -0
+	bo_first = list_first_entry(&man->lru, struct ttm_buffer_object, lru);
+	free_size_bo = ttm_bo_free_size_if_evicted(bo);
+	free_size_bo_first = ttm_bo_free_size_if_evicted(bo_first);
+	if (free_size_bo > free_size_bo_first) {
+		list_del_init(&bo->lru);
+		list_add(&bo->lru, &man->lru);
+	}
+}
+
 

Hmm, Can you explain why we'd ever *not* want to put bo first on the lru list?
I mean , bo_first might not even be in the correct range? That would also save the size computations.

Quoted 39 lines
drivers/gpu/drm/ttm/ttm_bo.c +39 -0
+static int ttm_manager_space_range(struct ttm_buffer_object *bo,
+					uint32_t mem_type,
+					struct ttm_mem_reg *mem,
+					unsigned long start, unsigned long end,
+					bool interruptible, bool no_wait)
+{
+	struct ttm_bo_global *glob = bo->glob;
+	struct drm_mm_node *entry;
+	struct ttm_mem_type_manager *man = &bo->bdev->man[mem_type];
+	struct drm_mm *mm = &man->manager;
+	unsigned size = end - start;
+	struct ttm_buffer_object *tbo;
+	unsigned wasted;
+	int ret;
+
+	mem->mm_node = NULL;
+	ret = drm_mm_pre_get(&man->manager);
+	if (unlikely(ret))
+		return ret;
+	spin_lock(&glob->lru_lock);
+	list_for_each_entry(entry, &mm->ml_entry, ml_entry) {
+		wasted = 0;
+		if (mem->page_alignment) {
+			unsigned tmp = entry->start % mem->page_alignment;
+			if (tmp)
+				wasted += mem->page_alignment - tmp;
+		}
+		if (entry->start < end && (entry->start+entry->size) > start) {
+			if (!entry->free) {
+				tbo = (struct ttm_buffer_object*)entry->private;
+				ttm_manager_evict_first(tbo);
+			} else {
+				if ((entry->size - wasted) <= size) {
+					/* Found a place */
+					entry = drm_mm_get_block_atomic(entry,
+							mem->num_pages,
+							mem->page_alignment);
+					mem->mm_node = entry;
+					mem->mem_type = mem_type;

drivers/gpu/drm/ttm/ttm_bo.c int ttm_bo_mem_space(struct ttm_buffer_object *bo,

Quoted 10 lines
drivers/gpu/drm/ttm/ttm_bo.c +9 -0
+					entry->private = bo;
+					break;
+				}
+			}
+		}
+	}
+	spin_unlock(&glob->lru_lock);
+	return 0;
+}
 

There's stuff here that belongs in drm_mm.c rather in the ttm code! What about
drm_mm_search_free_range() // Returns a free block in the given range if any. and perhaps functions to return the first non-free entry?

Quoted 64 lines
drivers/gpu/drm/ttm/ttm_bo.c +63 -0
+
+static int ttm_bo_mem_space_range(struct ttm_buffer_object *bo,
+				uint32_t proposed_placement,
+				struct ttm_mem_reg *mem,
+				unsigned long start, unsigned long end,
+				bool interruptible, bool no_wait)
+{
+	struct ttm_bo_device *bdev = bo->bdev;
+	struct ttm_bo_global *glob = bo->glob;
+	struct ttm_mem_type_manager *man;
+	uint32_t num_prios = bdev->driver->num_mem_type_prio;
+	const uint32_t *prios = bdev->driver->mem_type_prio;
+	uint32_t i;
+	uint32_t mem_type = TTM_PL_SYSTEM;
+	uint32_t cur_flags = 0;
+	bool type_found = false;
+	bool type_ok = false;
+	bool evicted;
+	int ret;
+
+	mem->mm_node = NULL;
+retry:
+	for (i = 0; i < num_prios; ++i) {
+		mem_type = prios[i];
+		man = &bdev->man[mem_type];
+		type_ok = ttm_bo_mt_compatible(man,
+						bo->type == ttm_bo_type_user,
+						mem_type, proposed_placement,
+						&cur_flags);
+		if (!type_ok)
+			continue;
+		if (start > (man->offset + man->size) || end < man->offset)
+			continue;
+		cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
+							cur_flags);
+		if (mem_type == TTM_PL_SYSTEM)
+			break;
+		if (man->has_type && man->use_type) {
+			type_found = true;
+			mem->placement = cur_flags;
+			ret = ttm_manager_space_range(bo, mem_type, mem,
+							start, end,
+							interruptible, no_wait);
+			if (ret)
+				return ret;
+		}
+		if (mem->mm_node)
+			break;
+	}
+	if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
+		mem->mem_type = mem_type;
+		mem->placement = cur_flags;
+		return 0;
+	}
+	for (i = 0, evicted = false; i < num_prios; ++i) {
+		struct ttm_buffer_object *tmp;
+
+		mem_type = prios[i];
+		man = &bdev->man[mem_type];
+		type_ok = ttm_bo_mt_compatible(man,
+						bo->type == ttm_bo_type_user,
+						mem_type, proposed_placement,
+						&cur_flags);

drivers/gpu/drm/ttm/ttm_bo.c int ttm_bo_mem_space(struct ttm_buffer_object *bo, · line 1032

Quoted 10 lines
drivers/gpu/drm/ttm/ttm_bo.c +8 -0
+		if (!type_ok)
+			continue;
+		if (start > (man->offset + man->size) || end < man->offset)
+			continue;
 
 
+		spin_lock(&glob->lru_lock);
+		tmp = list_first_entry(&man->lru, struct ttm_buffer_object, lru);
+		spin_unlock(&glob->lru_lock);
+		ret = ttm_bo_evict(tmp, mem_type, interruptible, no_wait);

^^^^This is incorrect. As soon as you release the lru lock, tmp may be destroyed by another process. Furthermore, tmp must be reserved before calling ttm_bo_evict. Please see how this is done in ttm_bo_mem_force_space.

Quoted 170 lines
drivers/gpu/drm/ttm/ttm_bo.c +141 -0
+		if (unlikely(ret != 0))
+			return ret;
+	}
+	if (!evicted)
+		return -ENOMEM;
+	goto retry;
+}
+
 int ttm_bo_wait_cpu(struct ttm_buffer_object *bo, bool no_wait)
 {
 	int ret = 0;
@@ -925,6 +1095,7 @@ int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
 out_unlock:
 	if (ret && mem.mm_node) {
 		spin_lock(&glob->lru_lock);
+		mem.mm_node->private = NULL;
 		drm_mm_put_block(mem.mm_node);
 		spin_unlock(&glob->lru_lock);
 	}
@@ -998,6 +1169,137 @@ int ttm_buffer_object_validate(struct ttm_buffer_object *bo,
 }
 EXPORT_SYMBOL(ttm_buffer_object_validate);
 
+int ttm_bo_move_buffer_range(struct ttm_buffer_object *bo,
+				uint32_t proposed_placement,
+				unsigned long start, unsigned long end,
+				bool interruptible, bool no_wait)
+{
+	struct ttm_bo_global *glob = bo->glob;
+	int ret = 0;
+	struct ttm_mem_reg mem;
+
+	BUG_ON(!atomic_read(&bo->reserved));
+
+	/*
+	 * FIXME: It's possible to pipeline buffer moves.
+	 * Have the driver move function wait for idle when necessary,
+	 * instead of doing it here.
+	 */
+
+	spin_lock(&bo->lock);
+	ret = ttm_bo_wait(bo, false, interruptible, no_wait);
+	spin_unlock(&bo->lock);
+
+	if (ret)
+		return ret;
+
+	mem.num_pages = bo->num_pages;
+	mem.size = mem.num_pages << PAGE_SHIFT;
+	mem.page_alignment = bo->mem.page_alignment;
+
+	/*
+	 * Determine where to move the buffer.
+	 */
+
+	ret = ttm_bo_mem_space_range(bo, proposed_placement, &mem,
+					start, end, interruptible, no_wait);
+	if (ret)
+		goto out_unlock;
+
+	ret = ttm_bo_handle_move_mem(bo, &mem, false, interruptible, no_wait);
+
+out_unlock:
+	if (ret && mem.mm_node) {
+		spin_lock(&glob->lru_lock);
+		mem.mm_node->private = NULL;
+		drm_mm_put_block(mem.mm_node);
+		spin_unlock(&glob->lru_lock);
+	}
+	return ret;
+}
+
+static int ttm_bo_mem_compat_range(uint32_t proposed_placement,
+					struct ttm_mem_reg *mem,
+					unsigned long start, unsigned long end)
+{
+	if ((proposed_placement & mem->placement & TTM_PL_MASK_MEM) == 0)
+		return 0;
+	if ((proposed_placement & mem->placement & TTM_PL_MASK_CACHING) == 0)
+		return 0;
+	if (mem->mm_node) {
+		unsigned long eend = mem->mm_node->start + mem->mm_node->size;
+		if (mem->mm_node->start < start || eend > end)
+			return 0;
+	}
+
+	return 1;
+}
+
+int ttm_buffer_object_validate_range(struct ttm_buffer_object *bo,
+					uint32_t proposed_placement,
+					unsigned long start, unsigned long end,
+					bool interruptible, bool no_wait)
+{
+	int ret;
+
+	/* Convert to page */
+	start = start >> PAGE_SHIFT;
+	end = end >> PAGE_SHIFT;
+	/* Check that range is valid */
+	if (start > end || (end - start) < bo->num_pages)
+		return -EINVAL;
+
+	BUG_ON(!atomic_read(&bo->reserved));
+	bo->proposed_placement = proposed_placement;
+
+	TTM_DEBUG("Proposed placement 0x%08lx, Old flags 0x%08lx\n",
+		  (unsigned long)proposed_placement,
+		  (unsigned long)bo->mem.placement);
+
+	/*
+	 * Check whether we need to move buffer.
+	 */
+
+	if (!ttm_bo_mem_compat_range(bo->proposed_placement, &bo->mem,
+					start, end)) {
+		ret = ttm_bo_move_buffer_range(bo, bo->proposed_placement,
+						start, end,
+						interruptible, no_wait);
+		if (ret) {
+			if (ret != -ERESTART)
+				printk(KERN_ERR TTM_PFX
+					"Failed moving buffer. "
+					"Proposed placement 0x%08x\n",
+					bo->proposed_placement);
+			if (ret == -ENOMEM)
+				printk(KERN_ERR TTM_PFX
+					"Out of aperture space or "
+					"DRM memory quota.\n");
+			return ret;
+		}
+	}
+
+	/*
+	 * We might need to add a TTM.
+	 */
+	if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
+		ret = ttm_bo_add_ttm(bo, true);
+		if (ret)
+			return ret;
+	}
+	/*
+	 * Validation has succeeded, move the access and other
+	 * non-mapping-related flag bits from the proposed flags to
+	 * the active flags
+	 */
+
+	ttm_flag_masked(&bo->mem.placement, bo->proposed_placement,
+			~TTM_PL_MASK_MEMTYPE);
+
+	return 0;
+}
+EXPORT_SYMBOL(ttm_buffer_object_validate_range);
+
 int
 ttm_bo_check_placement(struct ttm_buffer_object *bo,
 		       uint32_t set_flags, uint32_t clr_flags)
@@ -1321,6 +1623,7 @@ int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
 	man->has_type = true;
 	man->use_type = true;
 	man->size = p_size;
+	man->offset = p_offset;
 
 	INIT_LIST_HEAD(&man->lru);
 
include/drm/ttm/ttm_bo_api.h +0 -0
index 4911461..3756970 100644

include/drm/ttm/ttm_bo_api.h extern int ttm_bo_wait(struct ttm_buffer_object *bo, bool lazy, · line 315

Quoted 8 lines
include/drm/ttm/ttm_bo_api.h +4 -0
@@ -308,6 +308,11 @@ extern int ttm_bo_wait(struct ttm_buffer_object *bo, bool lazy,
 extern int ttm_buffer_object_validate(struct ttm_buffer_object *bo,
 				      uint32_t proposed_placement,
 				      bool interruptible, bool no_wait);
+extern int ttm_buffer_object_validate_range(struct ttm_buffer_object *bo,
+					uint32_t proposed_placement,
+					unsigned long start, unsigned long end,
+					bool interruptible, bool no_wait);
 

Please add a documentation entry in the header.

Quoted 10 lines
include/drm/ttm/ttm_bo_api.h +1 -0
+
 /**
  * ttm_bo_unref

*

include/drm/ttm/ttm_bo_driver.h +0 -0
index e8cd6d2..f00ba12 100644

include/drm/ttm/ttm_bo_driver.h struct ttm_mem_type_manager { · line 230

Quoted 7 lines
include/drm/ttm/ttm_bo_driver.h +1 -0
@@ -224,6 +224,7 @@ struct ttm_mem_type_manager {
 	unsigned long gpu_offset;
 	unsigned long io_offset;
 	unsigned long io_size;
+	unsigned long offset;
 	void *io_addr;
 	uint64_t size;
 	uint32_t available_caching;

Please add a doc entry in the header also here.
@offset is a bit confusing given all other offset members. What about managed_offset or something similar?

maintainer
#8 lore raw
In reply to Jerome Glisse
Quoted 2 lines · Jerome Glisse

This patch series add ttm range validation function. Aim is to include this in 2.6.33 so i have time to iron out issue, comments.

I missed these first time around,

Thomas if you have any opinions on the TTM stuff please see if you can take a look.

Quoted 6 lines

ttm:
I duplicated a bunch of ttm functions but now i think, best would be to add range to all function and use free list if range cover all the manager space. Doing so we might also be able to simplify mem_space alocation into a simpler function like ttm_bo_mem_space_range

Quoted 9 lines

radeon:
The second patch is a rework/cleanup of radeon object, it solves few issues along the way (i can't remember them now after fews days testing the patches). Biggest change is that we now rely on BO being validated before doing any change to radeon bo structure. As with any big patch i might introduce regressions, so far after testing on AGP:R1XX,R2XX,R3XX,R6XX PCIE:R3XX,R4XX,R5XX,R6XX,R7XX and RS480,RS690 i didn't found anythings obvious (test being X + glxgears + compiz(on hw which support it) + suspend/resume).

I'll try and take a look at this here it doesn't seem to depend on the first patch so I can push it separate if needed.

Dave.

Quoted 5 lines

Last patch is smaller, it just use the interface introduced by the first patch.

#9 lore raw
Quoted 12 lines · Dave Airlie
Quoted 3 lines · Jerome Glisse

This patch series add ttm range validation function. Aim is to include this in 2.6.33 so i have time to iron out issue, comments.

I missed these first time around,

Thomas if you have any opinions on the TTM stuff please see if you can take a look.

Sure, I'll take a look.

/Thomas