Skip to contentSkip to conversation

Add BO reservation to GEM VRAM pin/unpin/push_to_system

PATCH In review Thomas Zimmermann 4 participants 19 messages May 16, 2019 – May 21, 2019

Subsystems (6)
lore
Share & tools
lore

Thomas Zimmermann

author maintainer
#0 lore raw

cover Add BO reservation to GEM VRAM pin/unpin/push_to_system

A kernel test bot reported a problem with the locktorture testcase that was triggered by the GEM VRAM helpers.

...

  [   10.004734] RIP: 0010:ttm_bo_validate+0x41/0x141 [ttm]
  ...
  [   10.015669]  ? kvm_sched_clock_read+0x5/0xd
  [   10.016157]  ? get_lock_stats+0x11/0x3f
  [   10.016607]  drm_gem_vram_pin+0x77/0xa2 [drm_vram_helper]
  [   10.017229]  drm_gem_vram_driver_gem_prime_vmap+0xe/0x39 [drm_vram_helper]
  [   10.018015]  drm_gem_vmap+0x36/0x43 [drm]
  [   10.018491]  drm_client_framebuffer_create+0xc6/0x1ca [drm]
  [   10.019143]  drm_fb_helper_generic_probe+0x4c/0x157 [drm_kms_helper]
  [   10.019864]  __drm_fb_helper_initial_config_and_unlock+0x307/0x442 [drm_kms_helper]
  [   10.020739]  drm_fbdev_client_hotplug+0xc8/0x115 [drm_kms_helper]
  [   10.021442]  drm_fbdev_generic_setup+0xc4/0xf1 [drm_kms_helper]
  [   10.022120]  bochs_pci_probe+0x123/0x143 [bochs_drm]
  [   10.022688]  local_pci_probe+0x34/0x75
  [   10.023127]  pci_device_probe+0xf8/0x150A
  ...

It turns out that the bochs and vbox drivers automatically reserved and unreserved the BO from within their pin and unpin functions. The other drivers; ast, hibmc and mgag200; performed reservation explicitly. With the GEM VRAM conversion, automatic BO reservation within pin and unpin functions accidentally got lost. So for bochs and vbox, ttm_bo_validate() worked on unlocked BOs.

This patch set fixes the problem by adding automatic reservation to the implementation of drm_gem_vram_{pin,unpin,push_to_system}() to fix bochs and vbox. It removes explicit BO reservation around the pin, unpin and push-to-system calls in the ast, hibmc and mgag200 drivers.

The only exception is the cursor handling of mgag200. In this case, the mgag200 driver now calls drm_gem_vram_{pin,unpin}_reserved(), which works with reserved BOs. The respective code should be refactored in a future patch to work with the regular pin and unpin functions.

Thomas Zimmermann (2):
drm: Add drm_gem_vram_{pin/unpin}_reserved() and convert mgag200
drm: Reserve/unreserve GEM VRAM BOs from within pin/unpin functions

 drivers/gpu/drm/ast/ast_mode.c                |  24 +---
 drivers/gpu/drm/drm_gem_vram_helper.c         | 115 +++++++++++++++++-
 .../gpu/drm/hisilicon/hibmc/hibmc_drm_de.c    |   6 -
 .../gpu/drm/hisilicon/hibmc/hibmc_drm_fbdev.c |  17 +--
 drivers/gpu/drm/mgag200/mgag200_cursor.c      |  18 +--
 drivers/gpu/drm/mgag200/mgag200_mode.c        |  19 +--
 include/drm/drm_gem_vram_helper.h             |   3 +
 7 files changed, 126 insertions(+), 76 deletions(-)

--
2.21.0

author maintainer
#1 lore raw

PATCH 1/2 drm: Add drm_gem_vram_{pin/unpin}_reserved() and convert mgag200

The new interfaces drm_gem_vram_{pin/unpin}_reserved() are variants of the GEM VRAM pin/unpin functions that do not reserve the BO during validation. The mgag200 driver requires this behavior for its cursor handling. The patch also converts the driver to use the new interfaces.

---
 drivers/gpu/drm/drm_gem_vram_helper.c    | 75 ++++++++++++++++++++++++
 drivers/gpu/drm/mgag200/mgag200_cursor.c | 18 +++---
 include/drm/drm_gem_vram_helper.h        |  3 +
 3 files changed, 88 insertions(+), 8 deletions(-)
drivers/gpu/drm/drm_gem_vram_helper.c +75 -0
index 8f142b810eb4..a002c03eaf4c 100644
@@ -254,6 +254,47 @@ int drm_gem_vram_pin(struct drm_gem_vram_object *gbo, unsigned long pl_flag)
 }
 EXPORT_SYMBOL(drm_gem_vram_pin);
 
+/**
+ * drm_gem_vram_pin_reserved() - Pins a GEM VRAM object in a region.
+ * @gbo:	the GEM VRAM object
+ * @pl_flag:	a bitmask of possible memory regions
+ *
+ * Pinning a buffer object ensures that it is not evicted from
+ * a memory region. A pinned buffer object has to be unpinned before
+ * it can be pinned to another region.
+ *
+ * This function pins a GEM VRAM object that has already been
+ * reserved. Use drm_gem_vram_pin() if possible.
+ *
+ * Returns:
+ * 0 on success, or
+ * a negative error code otherwise.
+ */
+int drm_gem_vram_pin_reserved(struct drm_gem_vram_object *gbo,
+			      unsigned long pl_flag)
+{
+	int i, ret;
+	struct ttm_operation_ctx ctx = { false, false };
+
+	if (gbo->pin_count) {
+		++gbo->pin_count;
+		return 0;
+	}
+
+	drm_gem_vram_placement(gbo, pl_flag);
+	for (i = 0; i < gbo->placement.num_placement; ++i)
+		gbo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
+
+	ret = ttm_bo_validate(&gbo->bo, &gbo->placement, &ctx);
+	if (ret < 0)
+		return ret;
+
+	gbo->pin_count = 1;
+
+	return 0;
+}
+EXPORT_SYMBOL(drm_gem_vram_pin_reserved);
+
 /**
  * drm_gem_vram_unpin() - Unpins a GEM VRAM object
  * @gbo:	the GEM VRAM object
@@ -285,6 +326,40 @@ int drm_gem_vram_unpin(struct drm_gem_vram_object *gbo)
 }
 EXPORT_SYMBOL(drm_gem_vram_unpin);
 
+/**
+ * drm_gem_vram_unpin_reserved() - Unpins a GEM VRAM object
+ * @gbo:	the GEM VRAM object
+ *
+ * This function unpins a GEM VRAM object that has already been
+ * reserved. Use drm_gem_vram_unpin() if possible.
+ *
+ * Returns:
+ * 0 on success, or
+ * a negative error code otherwise.
+ */
+int drm_gem_vram_unpin_reserved(struct drm_gem_vram_object *gbo)
+{
+	int i, ret;
+	struct ttm_operation_ctx ctx = { false, false };
+
+	if (WARN_ON_ONCE(!gbo->pin_count))
+		return 0;
+
+	--gbo->pin_count;
+	if (gbo->pin_count)
+		return 0;
+
+	for (i = 0; i < gbo->placement.num_placement ; ++i)
+		gbo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
+
+	ret = ttm_bo_validate(&gbo->bo, &gbo->placement, &ctx);
+	if (ret < 0)
+		return ret;
+
+	return 0;
+}
+EXPORT_SYMBOL(drm_gem_vram_unpin_reserved);
+
 /**
  * drm_gem_vram_push_to_system() - \
 	Unpins a GEM VRAM object and moves it to system memory
drivers/gpu/drm/mgag200/mgag200_cursor.c +10 -8
index 6c1a9d724d85..1c4fc85315a0 100644
@@ -23,9 +23,9 @@ static void mga_hide_cursor(struct mga_device *mdev)
 	WREG8(MGA_CURPOSXL, 0);
 	WREG8(MGA_CURPOSXH, 0);
 	if (mdev->cursor.pixels_1->pin_count)
-		drm_gem_vram_unpin(mdev->cursor.pixels_1);
+		drm_gem_vram_unpin_reserved(mdev->cursor.pixels_1);
 	if (mdev->cursor.pixels_2->pin_count)
-		drm_gem_vram_unpin(mdev->cursor.pixels_2);
+		drm_gem_vram_unpin_reserved(mdev->cursor.pixels_2);
 }
 
 int mga_crtc_cursor_set(struct drm_crtc *crtc,
@@ -96,26 +96,28 @@ int mga_crtc_cursor_set(struct drm_crtc *crtc,
 
 	/* Move cursor buffers into VRAM if they aren't already */
 	if (!pixels_1->pin_count) {
-		ret = drm_gem_vram_pin(pixels_1, DRM_GEM_VRAM_PL_FLAG_VRAM);
+		ret = drm_gem_vram_pin_reserved(pixels_1,
+						DRM_GEM_VRAM_PL_FLAG_VRAM);
 		if (ret)
 			goto out1;
 		gpu_addr = drm_gem_vram_offset(pixels_1);
 		if (gpu_addr < 0) {
-			drm_gem_vram_unpin(pixels_1);
+			drm_gem_vram_unpin_reserved(pixels_1);
 			goto out1;
 		}
 		mdev->cursor.pixels_1_gpu_addr = gpu_addr;
 	}
 	if (!pixels_2->pin_count) {
-		ret = drm_gem_vram_pin(pixels_2, DRM_GEM_VRAM_PL_FLAG_VRAM);
+		ret = drm_gem_vram_pin_reserved(pixels_2,
+						DRM_GEM_VRAM_PL_FLAG_VRAM);
 		if (ret) {
-			drm_gem_vram_unpin(pixels_1);
+			drm_gem_vram_unpin_reserved(pixels_1);
 			goto out1;
 		}
 		gpu_addr = drm_gem_vram_offset(pixels_2);
 		if (gpu_addr < 0) {
-			drm_gem_vram_unpin(pixels_1);
-			drm_gem_vram_unpin(pixels_2);
+			drm_gem_vram_unpin_reserved(pixels_1);
+			drm_gem_vram_unpin_reserved(pixels_2);
 			goto out1;
 		}
 		mdev->cursor.pixels_2_gpu_addr = gpu_addr;
include/drm/drm_gem_vram_helper.h +3 -0
index b056f189ba62..ff1a81761543 100644
@@ -82,7 +82,10 @@ void drm_gem_vram_unreserve(struct drm_gem_vram_object *gbo);
 u64 drm_gem_vram_mmap_offset(struct drm_gem_vram_object *gbo);
 s64 drm_gem_vram_offset(struct drm_gem_vram_object *gbo);
 int drm_gem_vram_pin(struct drm_gem_vram_object *gbo, unsigned long pl_flag);
+int drm_gem_vram_pin_reserved(struct drm_gem_vram_object *gbo,
+			      unsigned long pl_flag);
 int drm_gem_vram_unpin(struct drm_gem_vram_object *gbo);
+int drm_gem_vram_unpin_reserved(struct drm_gem_vram_object *gbo);
 int drm_gem_vram_push_to_system(struct drm_gem_vram_object *gbo);
 void *drm_gem_vram_kmap_at(struct drm_gem_vram_object *gbo, bool map,
 			   bool *is_iomem, struct ttm_bo_kmap_obj *kmap);
Signature · 1 line
2.21.0
author maintainer
#2 lore raw

PATCH 1/2 drm: Add drm_gem_vram_{pin/unpin}_reserved() and convert mgag200

In reply to Thomas Zimmermann

The new interfaces drm_gem_vram_{pin/unpin}_reserved() are variants of the GEM VRAM pin/unpin functions that do not reserve the BO during validation. The mgag200 driver requires this behavior for its cursor handling. The patch also converts the driver to use the new interfaces.

---
 drivers/gpu/drm/drm_gem_vram_helper.c    | 75 ++++++++++++++++++++++++
 drivers/gpu/drm/mgag200/mgag200_cursor.c | 18 +++---
 include/drm/drm_gem_vram_helper.h        |  3 +
 3 files changed, 88 insertions(+), 8 deletions(-)
drivers/gpu/drm/drm_gem_vram_helper.c +75 -0
index 8f142b810eb4..a002c03eaf4c 100644
@@ -254,6 +254,47 @@ int drm_gem_vram_pin(struct drm_gem_vram_object *gbo, unsigned long pl_flag)
 }
 EXPORT_SYMBOL(drm_gem_vram_pin);
 
+/**
+ * drm_gem_vram_pin_reserved() - Pins a GEM VRAM object in a region.
+ * @gbo:	the GEM VRAM object
+ * @pl_flag:	a bitmask of possible memory regions
+ *
+ * Pinning a buffer object ensures that it is not evicted from
+ * a memory region. A pinned buffer object has to be unpinned before
+ * it can be pinned to another region.
+ *
+ * This function pins a GEM VRAM object that has already been
+ * reserved. Use drm_gem_vram_pin() if possible.
+ *
+ * Returns:
+ * 0 on success, or
+ * a negative error code otherwise.
+ */
+int drm_gem_vram_pin_reserved(struct drm_gem_vram_object *gbo,
+			      unsigned long pl_flag)
+{
+	int i, ret;
+	struct ttm_operation_ctx ctx = { false, false };
+
+	if (gbo->pin_count) {
+		++gbo->pin_count;
+		return 0;
+	}
+
+	drm_gem_vram_placement(gbo, pl_flag);
+	for (i = 0; i < gbo->placement.num_placement; ++i)
+		gbo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
+
+	ret = ttm_bo_validate(&gbo->bo, &gbo->placement, &ctx);
+	if (ret < 0)
+		return ret;
+
+	gbo->pin_count = 1;
+
+	return 0;
+}
+EXPORT_SYMBOL(drm_gem_vram_pin_reserved);
+
 /**
  * drm_gem_vram_unpin() - Unpins a GEM VRAM object
  * @gbo:	the GEM VRAM object
@@ -285,6 +326,40 @@ int drm_gem_vram_unpin(struct drm_gem_vram_object *gbo)
 }
 EXPORT_SYMBOL(drm_gem_vram_unpin);
 
+/**
+ * drm_gem_vram_unpin_reserved() - Unpins a GEM VRAM object
+ * @gbo:	the GEM VRAM object
+ *
+ * This function unpins a GEM VRAM object that has already been
+ * reserved. Use drm_gem_vram_unpin() if possible.
+ *
+ * Returns:
+ * 0 on success, or
+ * a negative error code otherwise.
+ */
+int drm_gem_vram_unpin_reserved(struct drm_gem_vram_object *gbo)
+{
+	int i, ret;
+	struct ttm_operation_ctx ctx = { false, false };
+
+	if (WARN_ON_ONCE(!gbo->pin_count))
+		return 0;
+
+	--gbo->pin_count;
+	if (gbo->pin_count)
+		return 0;
+
+	for (i = 0; i < gbo->placement.num_placement ; ++i)
+		gbo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
+
+	ret = ttm_bo_validate(&gbo->bo, &gbo->placement, &ctx);
+	if (ret < 0)
+		return ret;
+
+	return 0;
+}
+EXPORT_SYMBOL(drm_gem_vram_unpin_reserved);
+
 /**
  * drm_gem_vram_push_to_system() - \
 	Unpins a GEM VRAM object and moves it to system memory
drivers/gpu/drm/mgag200/mgag200_cursor.c +10 -8
index 6c1a9d724d85..1c4fc85315a0 100644
@@ -23,9 +23,9 @@ static void mga_hide_cursor(struct mga_device *mdev)
 	WREG8(MGA_CURPOSXL, 0);
 	WREG8(MGA_CURPOSXH, 0);
 	if (mdev->cursor.pixels_1->pin_count)
-		drm_gem_vram_unpin(mdev->cursor.pixels_1);
+		drm_gem_vram_unpin_reserved(mdev->cursor.pixels_1);
 	if (mdev->cursor.pixels_2->pin_count)
-		drm_gem_vram_unpin(mdev->cursor.pixels_2);
+		drm_gem_vram_unpin_reserved(mdev->cursor.pixels_2);
 }
 
 int mga_crtc_cursor_set(struct drm_crtc *crtc,
@@ -96,26 +96,28 @@ int mga_crtc_cursor_set(struct drm_crtc *crtc,
 
 	/* Move cursor buffers into VRAM if they aren't already */
 	if (!pixels_1->pin_count) {
-		ret = drm_gem_vram_pin(pixels_1, DRM_GEM_VRAM_PL_FLAG_VRAM);
+		ret = drm_gem_vram_pin_reserved(pixels_1,
+						DRM_GEM_VRAM_PL_FLAG_VRAM);
 		if (ret)
 			goto out1;
 		gpu_addr = drm_gem_vram_offset(pixels_1);
 		if (gpu_addr < 0) {
-			drm_gem_vram_unpin(pixels_1);
+			drm_gem_vram_unpin_reserved(pixels_1);
 			goto out1;
 		}
 		mdev->cursor.pixels_1_gpu_addr = gpu_addr;
 	}
 	if (!pixels_2->pin_count) {
-		ret = drm_gem_vram_pin(pixels_2, DRM_GEM_VRAM_PL_FLAG_VRAM);
+		ret = drm_gem_vram_pin_reserved(pixels_2,
+						DRM_GEM_VRAM_PL_FLAG_VRAM);
 		if (ret) {
-			drm_gem_vram_unpin(pixels_1);
+			drm_gem_vram_unpin_reserved(pixels_1);
 			goto out1;
 		}
 		gpu_addr = drm_gem_vram_offset(pixels_2);
 		if (gpu_addr < 0) {
-			drm_gem_vram_unpin(pixels_1);
-			drm_gem_vram_unpin(pixels_2);
+			drm_gem_vram_unpin_reserved(pixels_1);
+			drm_gem_vram_unpin_reserved(pixels_2);
 			goto out1;
 		}
 		mdev->cursor.pixels_2_gpu_addr = gpu_addr;
include/drm/drm_gem_vram_helper.h +3 -0
index b056f189ba62..ff1a81761543 100644
@@ -82,7 +82,10 @@ void drm_gem_vram_unreserve(struct drm_gem_vram_object *gbo);
 u64 drm_gem_vram_mmap_offset(struct drm_gem_vram_object *gbo);
 s64 drm_gem_vram_offset(struct drm_gem_vram_object *gbo);
 int drm_gem_vram_pin(struct drm_gem_vram_object *gbo, unsigned long pl_flag);
+int drm_gem_vram_pin_reserved(struct drm_gem_vram_object *gbo,
+			      unsigned long pl_flag);
 int drm_gem_vram_unpin(struct drm_gem_vram_object *gbo);
+int drm_gem_vram_unpin_reserved(struct drm_gem_vram_object *gbo);
 int drm_gem_vram_push_to_system(struct drm_gem_vram_object *gbo);
 void *drm_gem_vram_kmap_at(struct drm_gem_vram_object *gbo, bool map,
 			   bool *is_iomem, struct ttm_bo_kmap_obj *kmap);
Signature · 6 lines
2.21.0

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
#3 lore raw

Re: [PATCH 1/2] drm: Add drm_gem_vram_{pin/unpin}_reserved() and convert mgag200

Quoted 31 lines · Thomas Zimmermann

The new interfaces drm_gem_vram_{pin/unpin}_reserved() are variants of the GEM VRAM pin/unpin functions that do not reserve the BO during validation. The mgag200 driver requires this behavior for its cursor handling. The patch also converts the driver to use the new interfaces.

---
 drivers/gpu/drm/drm_gem_vram_helper.c    | 75 ++++++++++++++++++++++++
 drivers/gpu/drm/mgag200/mgag200_cursor.c | 18 +++---
 include/drm/drm_gem_vram_helper.h        |  3 +
 3 files changed, 88 insertions(+), 8 deletions(-)
drivers/gpu/drm/drm_gem_vram_helper.c +11 -0
index 8f142b810eb4..a002c03eaf4c 100644
@@ -254,6 +254,47 @@ int drm_gem_vram_pin(struct drm_gem_vram_object *gbo, unsigned long pl_flag)
 }
 EXPORT_SYMBOL(drm_gem_vram_pin);
 
+/**
+ * drm_gem_vram_pin_reserved() - Pins a GEM VRAM object in a region.
+ * @gbo:	the GEM VRAM object
+ * @pl_flag:	a bitmask of possible memory regions
+ *
+ * Pinning a buffer object ensures that it is not evicted from
+ * a memory region. A pinned buffer object has to be unpinned before
+ * it can be pinned to another region.
+ *
+ * This function pins a GEM VRAM object that has already been
+ * reserved. Use drm_gem_vram_pin() if possible.

drivers/gpu/drm/drm_gem_vram_helper.c int drm_gem_vram_pin(struct drm_gem_vram_object *gbo, unsigned long pl_flag) · line 277

Quoted 10 lines
drivers/gpu/drm/drm_gem_vram_helper.c +10 -0
+ *
+ * Returns:
+ * 0 on success, or
+ * a negative error code otherwise.
+ */
+int drm_gem_vram_pin_reserved(struct drm_gem_vram_object *gbo,
+			      unsigned long pl_flag)
+{
+	int i, ret;
+	struct ttm_operation_ctx ctx = { false, false };

I think would be good to have a lockdep_assert_held here for the ww_mutex.

Also general thing: _reserved is kinda ttm lingo, for dma-buf reservations we call the structure tracking the fences+lock the "reservation", but the naming scheme used is _lock/_unlock.

I think would be good to be consistent with that, and use _locked here. Especially for a very simplified vram helper like this one I expect that's going to lead to less wtf moments by driver writers :-)

Maybe we should also do a large-scale s/reserve/lock/ within ttm, to align more with what we now have in dma-buf.

Cheers, Daniel

Quoted 133 lines
drivers/gpu/drm/drm_gem_vram_helper.c +54 -0
+
+	if (gbo->pin_count) {
+		++gbo->pin_count;
+		return 0;
+	}
+
+	drm_gem_vram_placement(gbo, pl_flag);
+	for (i = 0; i < gbo->placement.num_placement; ++i)
+		gbo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
+
+	ret = ttm_bo_validate(&gbo->bo, &gbo->placement, &ctx);
+	if (ret < 0)
+		return ret;
+
+	gbo->pin_count = 1;
+
+	return 0;
+}
+EXPORT_SYMBOL(drm_gem_vram_pin_reserved);
+
 /**
  * drm_gem_vram_unpin() - Unpins a GEM VRAM object
  * @gbo:	the GEM VRAM object
@@ -285,6 +326,40 @@ int drm_gem_vram_unpin(struct drm_gem_vram_object *gbo)
 }
 EXPORT_SYMBOL(drm_gem_vram_unpin);
 
+/**
+ * drm_gem_vram_unpin_reserved() - Unpins a GEM VRAM object
+ * @gbo:	the GEM VRAM object
+ *
+ * This function unpins a GEM VRAM object that has already been
+ * reserved. Use drm_gem_vram_unpin() if possible.
+ *
+ * Returns:
+ * 0 on success, or
+ * a negative error code otherwise.
+ */
+int drm_gem_vram_unpin_reserved(struct drm_gem_vram_object *gbo)
+{
+	int i, ret;
+	struct ttm_operation_ctx ctx = { false, false };
+
+	if (WARN_ON_ONCE(!gbo->pin_count))
+		return 0;
+
+	--gbo->pin_count;
+	if (gbo->pin_count)
+		return 0;
+
+	for (i = 0; i < gbo->placement.num_placement ; ++i)
+		gbo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
+
+	ret = ttm_bo_validate(&gbo->bo, &gbo->placement, &ctx);
+	if (ret < 0)
+		return ret;
+
+	return 0;
+}
+EXPORT_SYMBOL(drm_gem_vram_unpin_reserved);
+
 /**
  * drm_gem_vram_push_to_system() - \
 	Unpins a GEM VRAM object and moves it to system memory
drivers/gpu/drm/mgag200/mgag200_cursor.c +10 -8
index 6c1a9d724d85..1c4fc85315a0 100644
@@ -23,9 +23,9 @@ static void mga_hide_cursor(struct mga_device *mdev)
 	WREG8(MGA_CURPOSXL, 0);
 	WREG8(MGA_CURPOSXH, 0);
 	if (mdev->cursor.pixels_1->pin_count)
-		drm_gem_vram_unpin(mdev->cursor.pixels_1);
+		drm_gem_vram_unpin_reserved(mdev->cursor.pixels_1);
 	if (mdev->cursor.pixels_2->pin_count)
-		drm_gem_vram_unpin(mdev->cursor.pixels_2);
+		drm_gem_vram_unpin_reserved(mdev->cursor.pixels_2);
 }
 
 int mga_crtc_cursor_set(struct drm_crtc *crtc,
@@ -96,26 +96,28 @@ int mga_crtc_cursor_set(struct drm_crtc *crtc,
 
 	/* Move cursor buffers into VRAM if they aren't already */
 	if (!pixels_1->pin_count) {
-		ret = drm_gem_vram_pin(pixels_1, DRM_GEM_VRAM_PL_FLAG_VRAM);
+		ret = drm_gem_vram_pin_reserved(pixels_1,
+						DRM_GEM_VRAM_PL_FLAG_VRAM);
 		if (ret)
 			goto out1;
 		gpu_addr = drm_gem_vram_offset(pixels_1);
 		if (gpu_addr < 0) {
-			drm_gem_vram_unpin(pixels_1);
+			drm_gem_vram_unpin_reserved(pixels_1);
 			goto out1;
 		}
 		mdev->cursor.pixels_1_gpu_addr = gpu_addr;
 	}
 	if (!pixels_2->pin_count) {
-		ret = drm_gem_vram_pin(pixels_2, DRM_GEM_VRAM_PL_FLAG_VRAM);
+		ret = drm_gem_vram_pin_reserved(pixels_2,
+						DRM_GEM_VRAM_PL_FLAG_VRAM);
 		if (ret) {
-			drm_gem_vram_unpin(pixels_1);
+			drm_gem_vram_unpin_reserved(pixels_1);
 			goto out1;
 		}
 		gpu_addr = drm_gem_vram_offset(pixels_2);
 		if (gpu_addr < 0) {
-			drm_gem_vram_unpin(pixels_1);
-			drm_gem_vram_unpin(pixels_2);
+			drm_gem_vram_unpin_reserved(pixels_1);
+			drm_gem_vram_unpin_reserved(pixels_2);
 			goto out1;
 		}
 		mdev->cursor.pixels_2_gpu_addr = gpu_addr;
include/drm/drm_gem_vram_helper.h +3 -0
index b056f189ba62..ff1a81761543 100644
@@ -82,7 +82,10 @@ void drm_gem_vram_unreserve(struct drm_gem_vram_object *gbo);
 u64 drm_gem_vram_mmap_offset(struct drm_gem_vram_object *gbo);
 s64 drm_gem_vram_offset(struct drm_gem_vram_object *gbo);
 int drm_gem_vram_pin(struct drm_gem_vram_object *gbo, unsigned long pl_flag);
+int drm_gem_vram_pin_reserved(struct drm_gem_vram_object *gbo,
+			      unsigned long pl_flag);
 int drm_gem_vram_unpin(struct drm_gem_vram_object *gbo);
+int drm_gem_vram_unpin_reserved(struct drm_gem_vram_object *gbo);
 int drm_gem_vram_push_to_system(struct drm_gem_vram_object *gbo);
 void *drm_gem_vram_kmap_at(struct drm_gem_vram_object *gbo, bool map,
 			   bool *is_iomem, struct ttm_bo_kmap_obj *kmap);
Signature · 1 line
2.21.0
Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
#4 lore raw

Re: [PATCH 1/2] drm: Add drm_gem_vram_{pin/unpin}_reserved() and convert mgag200

Quoted 49 lines · Daniel Vetter
Quoted 31 lines · Thomas Zimmermann

The new interfaces drm_gem_vram_{pin/unpin}_reserved() are variants of the GEM VRAM pin/unpin functions that do not reserve the BO during validation. The mgag200 driver requires this behavior for its cursor handling. The patch also converts the driver to use the new interfaces.

---
 drivers/gpu/drm/drm_gem_vram_helper.c    | 75 ++++++++++++++++++++++++
 drivers/gpu/drm/mgag200/mgag200_cursor.c | 18 +++---
 include/drm/drm_gem_vram_helper.h        |  3 +
 3 files changed, 88 insertions(+), 8 deletions(-)
drivers/gpu/drm/drm_gem_vram_helper.c +11 -0
index 8f142b810eb4..a002c03eaf4c 100644
@@ -254,6 +254,47 @@ int drm_gem_vram_pin(struct drm_gem_vram_object *gbo, unsigned long pl_flag)
 }
 EXPORT_SYMBOL(drm_gem_vram_pin);
 
+/**
+ * drm_gem_vram_pin_reserved() - Pins a GEM VRAM object in a region.
+ * @gbo:	the GEM VRAM object
+ * @pl_flag:	a bitmask of possible memory regions
+ *
+ * Pinning a buffer object ensures that it is not evicted from
+ * a memory region. A pinned buffer object has to be unpinned before
+ * it can be pinned to another region.
+ *
+ * This function pins a GEM VRAM object that has already been
+ * reserved. Use drm_gem_vram_pin() if possible.

drivers/gpu/drm/drm_gem_vram_helper.c int drm_gem_vram_pin(struct drm_gem_vram_object *gbo, unsigned long pl_flag) · line 277

Quoted 10 lines
drivers/gpu/drm/drm_gem_vram_helper.c +10 -0
+ *
+ * Returns:
+ * 0 on success, or
+ * a negative error code otherwise.
+ */
+int drm_gem_vram_pin_reserved(struct drm_gem_vram_object *gbo,
+			      unsigned long pl_flag)
+{
+	int i, ret;
+	struct ttm_operation_ctx ctx = { false, false };

I think would be good to have a lockdep_assert_held here for the ww_mutex.

Also general thing: _reserved is kinda ttm lingo, for dma-buf reservations we call the structure tracking the fences+lock the "reservation", but the naming scheme used is _lock/_unlock.

Quoted 6 lines

I think would be good to be consistent with that, and use _locked here. Especially for a very simplified vram helper like this one I expect that's going to lead to less wtf moments by driver writers :-)

Maybe we should also do a large-scale s/reserve/lock/ within ttm, to align more with what we now have in dma-buf.

More aside:

Could be a good move to demidlayer this an essentially remove ttm_bo_reserve as a wrapper around the linux/reservation.h functions. Not sure whether that aligns with Christian's plans. TODO.rst patch might be a good step to get that discussion started.
-Daniel

Quoted 141 lines

Cheers, Daniel

Quoted 133 lines
drivers/gpu/drm/drm_gem_vram_helper.c +54 -0
+
+	if (gbo->pin_count) {
+		++gbo->pin_count;
+		return 0;
+	}
+
+	drm_gem_vram_placement(gbo, pl_flag);
+	for (i = 0; i < gbo->placement.num_placement; ++i)
+		gbo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
+
+	ret = ttm_bo_validate(&gbo->bo, &gbo->placement, &ctx);
+	if (ret < 0)
+		return ret;
+
+	gbo->pin_count = 1;
+
+	return 0;
+}
+EXPORT_SYMBOL(drm_gem_vram_pin_reserved);
+
 /**
  * drm_gem_vram_unpin() - Unpins a GEM VRAM object
  * @gbo:	the GEM VRAM object
@@ -285,6 +326,40 @@ int drm_gem_vram_unpin(struct drm_gem_vram_object *gbo)
 }
 EXPORT_SYMBOL(drm_gem_vram_unpin);
 
+/**
+ * drm_gem_vram_unpin_reserved() - Unpins a GEM VRAM object
+ * @gbo:	the GEM VRAM object
+ *
+ * This function unpins a GEM VRAM object that has already been
+ * reserved. Use drm_gem_vram_unpin() if possible.
+ *
+ * Returns:
+ * 0 on success, or
+ * a negative error code otherwise.
+ */
+int drm_gem_vram_unpin_reserved(struct drm_gem_vram_object *gbo)
+{
+	int i, ret;
+	struct ttm_operation_ctx ctx = { false, false };
+
+	if (WARN_ON_ONCE(!gbo->pin_count))
+		return 0;
+
+	--gbo->pin_count;
+	if (gbo->pin_count)
+		return 0;
+
+	for (i = 0; i < gbo->placement.num_placement ; ++i)
+		gbo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
+
+	ret = ttm_bo_validate(&gbo->bo, &gbo->placement, &ctx);
+	if (ret < 0)
+		return ret;
+
+	return 0;
+}
+EXPORT_SYMBOL(drm_gem_vram_unpin_reserved);
+
 /**
  * drm_gem_vram_push_to_system() - \
 	Unpins a GEM VRAM object and moves it to system memory
drivers/gpu/drm/mgag200/mgag200_cursor.c +10 -8
index 6c1a9d724d85..1c4fc85315a0 100644
@@ -23,9 +23,9 @@ static void mga_hide_cursor(struct mga_device *mdev)
 	WREG8(MGA_CURPOSXL, 0);
 	WREG8(MGA_CURPOSXH, 0);
 	if (mdev->cursor.pixels_1->pin_count)
-		drm_gem_vram_unpin(mdev->cursor.pixels_1);
+		drm_gem_vram_unpin_reserved(mdev->cursor.pixels_1);
 	if (mdev->cursor.pixels_2->pin_count)
-		drm_gem_vram_unpin(mdev->cursor.pixels_2);
+		drm_gem_vram_unpin_reserved(mdev->cursor.pixels_2);
 }
 
 int mga_crtc_cursor_set(struct drm_crtc *crtc,
@@ -96,26 +96,28 @@ int mga_crtc_cursor_set(struct drm_crtc *crtc,
 
 	/* Move cursor buffers into VRAM if they aren't already */
 	if (!pixels_1->pin_count) {
-		ret = drm_gem_vram_pin(pixels_1, DRM_GEM_VRAM_PL_FLAG_VRAM);
+		ret = drm_gem_vram_pin_reserved(pixels_1,
+						DRM_GEM_VRAM_PL_FLAG_VRAM);
 		if (ret)
 			goto out1;
 		gpu_addr = drm_gem_vram_offset(pixels_1);
 		if (gpu_addr < 0) {
-			drm_gem_vram_unpin(pixels_1);
+			drm_gem_vram_unpin_reserved(pixels_1);
 			goto out1;
 		}
 		mdev->cursor.pixels_1_gpu_addr = gpu_addr;
 	}
 	if (!pixels_2->pin_count) {
-		ret = drm_gem_vram_pin(pixels_2, DRM_GEM_VRAM_PL_FLAG_VRAM);
+		ret = drm_gem_vram_pin_reserved(pixels_2,
+						DRM_GEM_VRAM_PL_FLAG_VRAM);
 		if (ret) {
-			drm_gem_vram_unpin(pixels_1);
+			drm_gem_vram_unpin_reserved(pixels_1);
 			goto out1;
 		}
 		gpu_addr = drm_gem_vram_offset(pixels_2);
 		if (gpu_addr < 0) {
-			drm_gem_vram_unpin(pixels_1);
-			drm_gem_vram_unpin(pixels_2);
+			drm_gem_vram_unpin_reserved(pixels_1);
+			drm_gem_vram_unpin_reserved(pixels_2);
 			goto out1;
 		}
 		mdev->cursor.pixels_2_gpu_addr = gpu_addr;
include/drm/drm_gem_vram_helper.h +3 -0
index b056f189ba62..ff1a81761543 100644
@@ -82,7 +82,10 @@ void drm_gem_vram_unreserve(struct drm_gem_vram_object *gbo);
 u64 drm_gem_vram_mmap_offset(struct drm_gem_vram_object *gbo);
 s64 drm_gem_vram_offset(struct drm_gem_vram_object *gbo);
 int drm_gem_vram_pin(struct drm_gem_vram_object *gbo, unsigned long pl_flag);
+int drm_gem_vram_pin_reserved(struct drm_gem_vram_object *gbo,
+			      unsigned long pl_flag);
 int drm_gem_vram_unpin(struct drm_gem_vram_object *gbo);
+int drm_gem_vram_unpin_reserved(struct drm_gem_vram_object *gbo);
 int drm_gem_vram_push_to_system(struct drm_gem_vram_object *gbo);
 void *drm_gem_vram_kmap_at(struct drm_gem_vram_object *gbo, bool map,
 			   bool *is_iomem, struct ttm_bo_kmap_obj *kmap);
Signature · 1 line
2.21.0
Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
#5 lore raw

Re: [PATCH 1/2] drm: Add drm_gem_vram_{pin/unpin}_reserved() and convert mgag200

In reply to Daniel Vetter
Quoted 49 lines · Daniel Vetter
Quoted 31 lines · Thomas Zimmermann

The new interfaces drm_gem_vram_{pin/unpin}_reserved() are variants of the GEM VRAM pin/unpin functions that do not reserve the BO during validation. The mgag200 driver requires this behavior for its cursor handling. The patch also converts the driver to use the new interfaces.

---
 drivers/gpu/drm/drm_gem_vram_helper.c    | 75 ++++++++++++++++++++++++
 drivers/gpu/drm/mgag200/mgag200_cursor.c | 18 +++---
 include/drm/drm_gem_vram_helper.h        |  3 +
 3 files changed, 88 insertions(+), 8 deletions(-)
drivers/gpu/drm/drm_gem_vram_helper.c +11 -0
index 8f142b810eb4..a002c03eaf4c 100644
@@ -254,6 +254,47 @@ int drm_gem_vram_pin(struct drm_gem_vram_object *gbo, unsigned long pl_flag)
 }
 EXPORT_SYMBOL(drm_gem_vram_pin);
 
+/**
+ * drm_gem_vram_pin_reserved() - Pins a GEM VRAM object in a region.
+ * @gbo:	the GEM VRAM object
+ * @pl_flag:	a bitmask of possible memory regions
+ *
+ * Pinning a buffer object ensures that it is not evicted from
+ * a memory region. A pinned buffer object has to be unpinned before
+ * it can be pinned to another region.
+ *
+ * This function pins a GEM VRAM object that has already been
+ * reserved. Use drm_gem_vram_pin() if possible.

drivers/gpu/drm/drm_gem_vram_helper.c int drm_gem_vram_pin(struct drm_gem_vram_object *gbo, unsigned long pl_flag) · line 277

Quoted 10 lines
drivers/gpu/drm/drm_gem_vram_helper.c +10 -0
+ *
+ * Returns:
+ * 0 on success, or
+ * a negative error code otherwise.
+ */
+int drm_gem_vram_pin_reserved(struct drm_gem_vram_object *gbo,
+			      unsigned long pl_flag)
+{
+	int i, ret;
+	struct ttm_operation_ctx ctx = { false, false };

I think would be good to have a lockdep_assert_held here for the ww_mutex.

Also general thing: _reserved is kinda ttm lingo, for dma-buf reservations we call the structure tracking the fences+lock the "reservation", but the naming scheme used is _lock/_unlock.

Quoted 6 lines

I think would be good to be consistent with that, and use _locked here. Especially for a very simplified vram helper like this one I expect that's going to lead to less wtf moments by driver writers :-)

Maybe we should also do a large-scale s/reserve/lock/ within ttm, to align more with what we now have in dma-buf.

More aside:

Could be a good move to demidlayer this an essentially remove ttm_bo_reserve as a wrapper around the linux/reservation.h functions. Not sure whether that aligns with Christian's plans. TODO.rst patch might be a good step to get that discussion started.
-Daniel

Quoted 141 lines

Cheers, Daniel

Quoted 133 lines
drivers/gpu/drm/drm_gem_vram_helper.c +54 -0
+
+	if (gbo->pin_count) {
+		++gbo->pin_count;
+		return 0;
+	}
+
+	drm_gem_vram_placement(gbo, pl_flag);
+	for (i = 0; i < gbo->placement.num_placement; ++i)
+		gbo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
+
+	ret = ttm_bo_validate(&gbo->bo, &gbo->placement, &ctx);
+	if (ret < 0)
+		return ret;
+
+	gbo->pin_count = 1;
+
+	return 0;
+}
+EXPORT_SYMBOL(drm_gem_vram_pin_reserved);
+
 /**
  * drm_gem_vram_unpin() - Unpins a GEM VRAM object
  * @gbo:	the GEM VRAM object
@@ -285,6 +326,40 @@ int drm_gem_vram_unpin(struct drm_gem_vram_object *gbo)
 }
 EXPORT_SYMBOL(drm_gem_vram_unpin);
 
+/**
+ * drm_gem_vram_unpin_reserved() - Unpins a GEM VRAM object
+ * @gbo:	the GEM VRAM object
+ *
+ * This function unpins a GEM VRAM object that has already been
+ * reserved. Use drm_gem_vram_unpin() if possible.
+ *
+ * Returns:
+ * 0 on success, or
+ * a negative error code otherwise.
+ */
+int drm_gem_vram_unpin_reserved(struct drm_gem_vram_object *gbo)
+{
+	int i, ret;
+	struct ttm_operation_ctx ctx = { false, false };
+
+	if (WARN_ON_ONCE(!gbo->pin_count))
+		return 0;
+
+	--gbo->pin_count;
+	if (gbo->pin_count)
+		return 0;
+
+	for (i = 0; i < gbo->placement.num_placement ; ++i)
+		gbo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
+
+	ret = ttm_bo_validate(&gbo->bo, &gbo->placement, &ctx);
+	if (ret < 0)
+		return ret;
+
+	return 0;
+}
+EXPORT_SYMBOL(drm_gem_vram_unpin_reserved);
+
 /**
  * drm_gem_vram_push_to_system() - \
 	Unpins a GEM VRAM object and moves it to system memory
drivers/gpu/drm/mgag200/mgag200_cursor.c +10 -8
index 6c1a9d724d85..1c4fc85315a0 100644
@@ -23,9 +23,9 @@ static void mga_hide_cursor(struct mga_device *mdev)
 	WREG8(MGA_CURPOSXL, 0);
 	WREG8(MGA_CURPOSXH, 0);
 	if (mdev->cursor.pixels_1->pin_count)
-		drm_gem_vram_unpin(mdev->cursor.pixels_1);
+		drm_gem_vram_unpin_reserved(mdev->cursor.pixels_1);
 	if (mdev->cursor.pixels_2->pin_count)
-		drm_gem_vram_unpin(mdev->cursor.pixels_2);
+		drm_gem_vram_unpin_reserved(mdev->cursor.pixels_2);
 }
 
 int mga_crtc_cursor_set(struct drm_crtc *crtc,
@@ -96,26 +96,28 @@ int mga_crtc_cursor_set(struct drm_crtc *crtc,
 
 	/* Move cursor buffers into VRAM if they aren't already */
 	if (!pixels_1->pin_count) {
-		ret = drm_gem_vram_pin(pixels_1, DRM_GEM_VRAM_PL_FLAG_VRAM);
+		ret = drm_gem_vram_pin_reserved(pixels_1,
+						DRM_GEM_VRAM_PL_FLAG_VRAM);
 		if (ret)
 			goto out1;
 		gpu_addr = drm_gem_vram_offset(pixels_1);
 		if (gpu_addr < 0) {
-			drm_gem_vram_unpin(pixels_1);
+			drm_gem_vram_unpin_reserved(pixels_1);
 			goto out1;
 		}
 		mdev->cursor.pixels_1_gpu_addr = gpu_addr;
 	}
 	if (!pixels_2->pin_count) {
-		ret = drm_gem_vram_pin(pixels_2, DRM_GEM_VRAM_PL_FLAG_VRAM);
+		ret = drm_gem_vram_pin_reserved(pixels_2,
+						DRM_GEM_VRAM_PL_FLAG_VRAM);
 		if (ret) {
-			drm_gem_vram_unpin(pixels_1);
+			drm_gem_vram_unpin_reserved(pixels_1);
 			goto out1;
 		}
 		gpu_addr = drm_gem_vram_offset(pixels_2);
 		if (gpu_addr < 0) {
-			drm_gem_vram_unpin(pixels_1);
-			drm_gem_vram_unpin(pixels_2);
+			drm_gem_vram_unpin_reserved(pixels_1);
+			drm_gem_vram_unpin_reserved(pixels_2);
 			goto out1;
 		}
 		mdev->cursor.pixels_2_gpu_addr = gpu_addr;
include/drm/drm_gem_vram_helper.h +3 -0
index b056f189ba62..ff1a81761543 100644
@@ -82,7 +82,10 @@ void drm_gem_vram_unreserve(struct drm_gem_vram_object *gbo);
 u64 drm_gem_vram_mmap_offset(struct drm_gem_vram_object *gbo);
 s64 drm_gem_vram_offset(struct drm_gem_vram_object *gbo);
 int drm_gem_vram_pin(struct drm_gem_vram_object *gbo, unsigned long pl_flag);
+int drm_gem_vram_pin_reserved(struct drm_gem_vram_object *gbo,
+			      unsigned long pl_flag);
 int drm_gem_vram_unpin(struct drm_gem_vram_object *gbo);
+int drm_gem_vram_unpin_reserved(struct drm_gem_vram_object *gbo);
 int drm_gem_vram_push_to_system(struct drm_gem_vram_object *gbo);
 void *drm_gem_vram_kmap_at(struct drm_gem_vram_object *gbo, bool map,
 			   bool *is_iomem, struct ttm_bo_kmap_obj *kmap);
Signature · 1 line
2.21.0
Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
Signature · 7 lines
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
#6 lore raw

Re: [PATCH 1/2] drm: Add drm_gem_vram_{pin/unpin}_reserved() and convert mgag200

Am 20.05.19 um 18:26 schrieb Daniel Vetter:

Quoted 55 lines

[CAUTION: External Email]

Quoted 52 lines · Daniel Vetter
Quoted 31 lines · Thomas Zimmermann

The new interfaces drm_gem_vram_{pin/unpin}_reserved() are variants of the GEM VRAM pin/unpin functions that do not reserve the BO during validation. The mgag200 driver requires this behavior for its cursor handling. The patch also converts the driver to use the new interfaces.

---
  drivers/gpu/drm/drm_gem_vram_helper.c    | 75 ++++++++++++++++++++++++
  drivers/gpu/drm/mgag200/mgag200_cursor.c | 18 +++---
  include/drm/drm_gem_vram_helper.h        |  3 +
  3 files changed, 88 insertions(+), 8 deletions(-)
drivers/gpu/drm/drm_gem_vram_helper.c +11 -0
index 8f142b810eb4..a002c03eaf4c 100644
@@ -254,6 +254,47 @@ int drm_gem_vram_pin(struct drm_gem_vram_object *gbo, unsigned long pl_flag)
  }
  EXPORT_SYMBOL(drm_gem_vram_pin);
 
+/**
+ * drm_gem_vram_pin_reserved() - Pins a GEM VRAM object in a region.
+ * @gbo:   the GEM VRAM object
+ * @pl_flag:       a bitmask of possible memory regions
+ *
+ * Pinning a buffer object ensures that it is not evicted from
+ * a memory region. A pinned buffer object has to be unpinned before
+ * it can be pinned to another region.
+ *
+ * This function pins a GEM VRAM object that has already been
+ * reserved. Use drm_gem_vram_pin() if possible.

drivers/gpu/drm/drm_gem_vram_helper.c int drm_gem_vram_pin(struct drm_gem_vram_object *gbo, unsigned long pl_flag) · line 277

Quoted 10 lines
drivers/gpu/drm/drm_gem_vram_helper.c +10 -0
+ *
+ * Returns:
+ * 0 on success, or
+ * a negative error code otherwise.
+ */
+int drm_gem_vram_pin_reserved(struct drm_gem_vram_object *gbo,
+                         unsigned long pl_flag)
+{
+   int i, ret;
+   struct ttm_operation_ctx ctx = { false, false };

I think would be good to have a lockdep_assert_held here for the ww_mutex.

Also general thing: _reserved is kinda ttm lingo, for dma-buf reservations we call the structure tracking the fences+lock the "reservation", but the naming scheme used is _lock/_unlock.

I think would be good to be consistent with that, and use _locked here. Especially for a very simplified vram helper like this one I expect that's going to lead to less wtf moments by driver writers :-)

Quoted 8 lines
Quoted 2 lines

Maybe we should also do a large-scale s/reserve/lock/ within ttm, to align more with what we now have in dma-buf.

More aside:

Could be a good move to demidlayer this an essentially remove ttm_bo_reserve as a wrapper around the linux/reservation.h functions. Not sure whether that aligns with Christian's plans. TODO.rst patch might be a good step to get that discussion started.

Well what ttm_bo_reserve()/_unreserve() does is a) lock/unlock the reservation object and b) remove the BO from the LRU.

Since I'm desperately trying to get rid of the LRU removal right now we sooner or later should be able to remove ttmo_bo_reserve()/_unreserve() as well (or at least replace them with tiny ttm_bo_lock() wrappers.

Christian.

Quoted 145 lines

-Daniel

Quoted 131 lines

Cheers, Daniel

Quoted 129 lines
drivers/gpu/drm/drm_gem_vram_helper.c +54 -0
+
+   if (gbo->pin_count) {
+           ++gbo->pin_count;
+           return 0;
+   }
+
+   drm_gem_vram_placement(gbo, pl_flag);
+   for (i = 0; i < gbo->placement.num_placement; ++i)
+           gbo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
+
+   ret = ttm_bo_validate(&gbo->bo, &gbo->placement, &ctx);
+   if (ret < 0)
+           return ret;
+
+   gbo->pin_count = 1;
+
+   return 0;
+}
+EXPORT_SYMBOL(drm_gem_vram_pin_reserved);
+
  /**
   * drm_gem_vram_unpin() - Unpins a GEM VRAM object
   * @gbo:   the GEM VRAM object
@@ -285,6 +326,40 @@ int drm_gem_vram_unpin(struct drm_gem_vram_object *gbo)
  }
  EXPORT_SYMBOL(drm_gem_vram_unpin);
 
+/**
+ * drm_gem_vram_unpin_reserved() - Unpins a GEM VRAM object
+ * @gbo:   the GEM VRAM object
+ *
+ * This function unpins a GEM VRAM object that has already been
+ * reserved. Use drm_gem_vram_unpin() if possible.
+ *
+ * Returns:
+ * 0 on success, or
+ * a negative error code otherwise.
+ */
+int drm_gem_vram_unpin_reserved(struct drm_gem_vram_object *gbo)
+{
+   int i, ret;
+   struct ttm_operation_ctx ctx = { false, false };
+
+   if (WARN_ON_ONCE(!gbo->pin_count))
+           return 0;
+
+   --gbo->pin_count;
+   if (gbo->pin_count)
+           return 0;
+
+   for (i = 0; i < gbo->placement.num_placement ; ++i)
+           gbo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
+
+   ret = ttm_bo_validate(&gbo->bo, &gbo->placement, &ctx);
+   if (ret < 0)
+           return ret;
+
+   return 0;
+}
+EXPORT_SYMBOL(drm_gem_vram_unpin_reserved);
+
  /**
   * drm_gem_vram_push_to_system() - \
     Unpins a GEM VRAM object and moves it to system memory
drivers/gpu/drm/mgag200/mgag200_cursor.c +10 -8
index 6c1a9d724d85..1c4fc85315a0 100644
@@ -23,9 +23,9 @@ static void mga_hide_cursor(struct mga_device *mdev)
     WREG8(MGA_CURPOSXL, 0);
     WREG8(MGA_CURPOSXH, 0);
     if (mdev->cursor.pixels_1->pin_count)
-           drm_gem_vram_unpin(mdev->cursor.pixels_1);
+           drm_gem_vram_unpin_reserved(mdev->cursor.pixels_1);
     if (mdev->cursor.pixels_2->pin_count)
-           drm_gem_vram_unpin(mdev->cursor.pixels_2);
+           drm_gem_vram_unpin_reserved(mdev->cursor.pixels_2);
  }
 
  int mga_crtc_cursor_set(struct drm_crtc *crtc,
@@ -96,26 +96,28 @@ int mga_crtc_cursor_set(struct drm_crtc *crtc,
 
     /* Move cursor buffers into VRAM if they aren't already */
     if (!pixels_1->pin_count) {
-           ret = drm_gem_vram_pin(pixels_1, DRM_GEM_VRAM_PL_FLAG_VRAM);
+           ret = drm_gem_vram_pin_reserved(pixels_1,
+                                           DRM_GEM_VRAM_PL_FLAG_VRAM);
             if (ret)
                     goto out1;
             gpu_addr = drm_gem_vram_offset(pixels_1);
             if (gpu_addr < 0) {
-                   drm_gem_vram_unpin(pixels_1);
+                   drm_gem_vram_unpin_reserved(pixels_1);
                     goto out1;
             }
             mdev->cursor.pixels_1_gpu_addr = gpu_addr;
     }
     if (!pixels_2->pin_count) {
-           ret = drm_gem_vram_pin(pixels_2, DRM_GEM_VRAM_PL_FLAG_VRAM);
+           ret = drm_gem_vram_pin_reserved(pixels_2,
+                                           DRM_GEM_VRAM_PL_FLAG_VRAM);
             if (ret) {
-                   drm_gem_vram_unpin(pixels_1);
+                   drm_gem_vram_unpin_reserved(pixels_1);
                     goto out1;
             }
             gpu_addr = drm_gem_vram_offset(pixels_2);
             if (gpu_addr < 0) {
-                   drm_gem_vram_unpin(pixels_1);
-                   drm_gem_vram_unpin(pixels_2);
+                   drm_gem_vram_unpin_reserved(pixels_1);
+                   drm_gem_vram_unpin_reserved(pixels_2);
                     goto out1;
             }
             mdev->cursor.pixels_2_gpu_addr = gpu_addr;
include/drm/drm_gem_vram_helper.h +3 -0
index b056f189ba62..ff1a81761543 100644
@@ -82,7 +82,10 @@ void drm_gem_vram_unreserve(struct drm_gem_vram_object *gbo);
  u64 drm_gem_vram_mmap_offset(struct drm_gem_vram_object *gbo);
  s64 drm_gem_vram_offset(struct drm_gem_vram_object *gbo);
  int drm_gem_vram_pin(struct drm_gem_vram_object *gbo, unsigned long pl_flag);
+int drm_gem_vram_pin_reserved(struct drm_gem_vram_object *gbo,
+                         unsigned long pl_flag);
  int drm_gem_vram_unpin(struct drm_gem_vram_object *gbo);
+int drm_gem_vram_unpin_reserved(struct drm_gem_vram_object *gbo);
  int drm_gem_vram_push_to_system(struct drm_gem_vram_object *gbo);
  void *drm_gem_vram_kmap_at(struct drm_gem_vram_object *gbo, bool map,
Quoted 8 lines
Quoted 4 lines
include/drm/drm_gem_vram_helper.h +0 -1
                        bool *is_iomem, struct ttm_bo_kmap_obj *kmap);
--

2.21.0

--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

#7 lore raw

Re: [PATCH 1/2] drm: Add drm_gem_vram_{pin/unpin}_reserved() and convert mgag200

In reply to Daniel Vetter

Am 20.05.19 um 18:26 schrieb Daniel Vetter:

Quoted 55 lines

[CAUTION: External Email]

Quoted 52 lines · Daniel Vetter
Quoted 31 lines · Thomas Zimmermann

The new interfaces drm_gem_vram_{pin/unpin}_reserved() are variants of the GEM VRAM pin/unpin functions that do not reserve the BO during validation. The mgag200 driver requires this behavior for its cursor handling. The patch also converts the driver to use the new interfaces.

---
  drivers/gpu/drm/drm_gem_vram_helper.c    | 75 ++++++++++++++++++++++++
  drivers/gpu/drm/mgag200/mgag200_cursor.c | 18 +++---
  include/drm/drm_gem_vram_helper.h        |  3 +
  3 files changed, 88 insertions(+), 8 deletions(-)
drivers/gpu/drm/drm_gem_vram_helper.c +11 -0
index 8f142b810eb4..a002c03eaf4c 100644
@@ -254,6 +254,47 @@ int drm_gem_vram_pin(struct drm_gem_vram_object *gbo, unsigned long pl_flag)
  }
  EXPORT_SYMBOL(drm_gem_vram_pin);
 
+/**
+ * drm_gem_vram_pin_reserved() - Pins a GEM VRAM object in a region.
+ * @gbo:   the GEM VRAM object
+ * @pl_flag:       a bitmask of possible memory regions
+ *
+ * Pinning a buffer object ensures that it is not evicted from
+ * a memory region. A pinned buffer object has to be unpinned before
+ * it can be pinned to another region.
+ *
+ * This function pins a GEM VRAM object that has already been
+ * reserved. Use drm_gem_vram_pin() if possible.

drivers/gpu/drm/drm_gem_vram_helper.c int drm_gem_vram_pin(struct drm_gem_vram_object *gbo, unsigned long pl_flag) · line 277

Quoted 10 lines
drivers/gpu/drm/drm_gem_vram_helper.c +10 -0
+ *
+ * Returns:
+ * 0 on success, or
+ * a negative error code otherwise.
+ */
+int drm_gem_vram_pin_reserved(struct drm_gem_vram_object *gbo,
+                         unsigned long pl_flag)
+{
+   int i, ret;
+   struct ttm_operation_ctx ctx = { false, false };

I think would be good to have a lockdep_assert_held here for the ww_mutex.

Also general thing: _reserved is kinda ttm lingo, for dma-buf reservations we call the structure tracking the fences+lock the "reservation", but the naming scheme used is _lock/_unlock.

I think would be good to be consistent with that, and use _locked here. Especially for a very simplified vram helper like this one I expect that's going to lead to less wtf moments by driver writers :-)

Quoted 8 lines
Quoted 2 lines

Maybe we should also do a large-scale s/reserve/lock/ within ttm, to align more with what we now have in dma-buf.

More aside:

Could be a good move to demidlayer this an essentially remove ttm_bo_reserve as a wrapper around the linux/reservation.h functions. Not sure whether that aligns with Christian's plans. TODO.rst patch might be a good step to get that discussion started.

Well what ttm_bo_reserve()/_unreserve() does is a) lock/unlock the reservation object and b) remove the BO from the LRU.

Since I'm desperately trying to get rid of the LRU removal right now we sooner or later should be able to remove ttmo_bo_reserve()/_unreserve() as well (or at least replace them with tiny ttm_bo_lock() wrappers.

Christian.

Quoted 141 lines

-Daniel

Quoted 131 lines

Cheers, Daniel

Quoted 129 lines
drivers/gpu/drm/drm_gem_vram_helper.c +54 -0
+
+   if (gbo->pin_count) {
+           ++gbo->pin_count;
+           return 0;
+   }
+
+   drm_gem_vram_placement(gbo, pl_flag);
+   for (i = 0; i < gbo->placement.num_placement; ++i)
+           gbo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
+
+   ret = ttm_bo_validate(&gbo->bo, &gbo->placement, &ctx);
+   if (ret < 0)
+           return ret;
+
+   gbo->pin_count = 1;
+
+   return 0;
+}
+EXPORT_SYMBOL(drm_gem_vram_pin_reserved);
+
  /**
   * drm_gem_vram_unpin() - Unpins a GEM VRAM object
   * @gbo:   the GEM VRAM object
@@ -285,6 +326,40 @@ int drm_gem_vram_unpin(struct drm_gem_vram_object *gbo)
  }
  EXPORT_SYMBOL(drm_gem_vram_unpin);
 
+/**
+ * drm_gem_vram_unpin_reserved() - Unpins a GEM VRAM object
+ * @gbo:   the GEM VRAM object
+ *
+ * This function unpins a GEM VRAM object that has already been
+ * reserved. Use drm_gem_vram_unpin() if possible.
+ *
+ * Returns:
+ * 0 on success, or
+ * a negative error code otherwise.
+ */
+int drm_gem_vram_unpin_reserved(struct drm_gem_vram_object *gbo)
+{
+   int i, ret;
+   struct ttm_operation_ctx ctx = { false, false };
+
+   if (WARN_ON_ONCE(!gbo->pin_count))
+           return 0;
+
+   --gbo->pin_count;
+   if (gbo->pin_count)
+           return 0;
+
+   for (i = 0; i < gbo->placement.num_placement ; ++i)
+           gbo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
+
+   ret = ttm_bo_validate(&gbo->bo, &gbo->placement, &ctx);
+   if (ret < 0)
+           return ret;
+
+   return 0;
+}
+EXPORT_SYMBOL(drm_gem_vram_unpin_reserved);
+
  /**
   * drm_gem_vram_push_to_system() - \
     Unpins a GEM VRAM object and moves it to system memory
drivers/gpu/drm/mgag200/mgag200_cursor.c +10 -8
index 6c1a9d724d85..1c4fc85315a0 100644
@@ -23,9 +23,9 @@ static void mga_hide_cursor(struct mga_device *mdev)
     WREG8(MGA_CURPOSXL, 0);
     WREG8(MGA_CURPOSXH, 0);
     if (mdev->cursor.pixels_1->pin_count)
-           drm_gem_vram_unpin(mdev->cursor.pixels_1);
+           drm_gem_vram_unpin_reserved(mdev->cursor.pixels_1);
     if (mdev->cursor.pixels_2->pin_count)
-           drm_gem_vram_unpin(mdev->cursor.pixels_2);
+           drm_gem_vram_unpin_reserved(mdev->cursor.pixels_2);
  }
 
  int mga_crtc_cursor_set(struct drm_crtc *crtc,
@@ -96,26 +96,28 @@ int mga_crtc_cursor_set(struct drm_crtc *crtc,
 
     /* Move cursor buffers into VRAM if they aren't already */
     if (!pixels_1->pin_count) {
-           ret = drm_gem_vram_pin(pixels_1, DRM_GEM_VRAM_PL_FLAG_VRAM);
+           ret = drm_gem_vram_pin_reserved(pixels_1,
+                                           DRM_GEM_VRAM_PL_FLAG_VRAM);
             if (ret)
                     goto out1;
             gpu_addr = drm_gem_vram_offset(pixels_1);
             if (gpu_addr < 0) {
-                   drm_gem_vram_unpin(pixels_1);
+                   drm_gem_vram_unpin_reserved(pixels_1);
                     goto out1;
             }
             mdev->cursor.pixels_1_gpu_addr = gpu_addr;
     }
     if (!pixels_2->pin_count) {
-           ret = drm_gem_vram_pin(pixels_2, DRM_GEM_VRAM_PL_FLAG_VRAM);
+           ret = drm_gem_vram_pin_reserved(pixels_2,
+                                           DRM_GEM_VRAM_PL_FLAG_VRAM);
             if (ret) {
-                   drm_gem_vram_unpin(pixels_1);
+                   drm_gem_vram_unpin_reserved(pixels_1);
                     goto out1;
             }
             gpu_addr = drm_gem_vram_offset(pixels_2);
             if (gpu_addr < 0) {
-                   drm_gem_vram_unpin(pixels_1);
-                   drm_gem_vram_unpin(pixels_2);
+                   drm_gem_vram_unpin_reserved(pixels_1);
+                   drm_gem_vram_unpin_reserved(pixels_2);
                     goto out1;
             }
             mdev->cursor.pixels_2_gpu_addr = gpu_addr;
include/drm/drm_gem_vram_helper.h +3 -0
index b056f189ba62..ff1a81761543 100644
@@ -82,7 +82,10 @@ void drm_gem_vram_unreserve(struct drm_gem_vram_object *gbo);
  u64 drm_gem_vram_mmap_offset(struct drm_gem_vram_object *gbo);
  s64 drm_gem_vram_offset(struct drm_gem_vram_object *gbo);
  int drm_gem_vram_pin(struct drm_gem_vram_object *gbo, unsigned long pl_flag);
+int drm_gem_vram_pin_reserved(struct drm_gem_vram_object *gbo,
+                         unsigned long pl_flag);
  int drm_gem_vram_unpin(struct drm_gem_vram_object *gbo);
+int drm_gem_vram_unpin_reserved(struct drm_gem_vram_object *gbo);
  int drm_gem_vram_push_to_system(struct drm_gem_vram_object *gbo);
  void *drm_gem_vram_kmap_at(struct drm_gem_vram_object *gbo, bool map,
Quoted 8 lines
Quoted 4 lines
include/drm/drm_gem_vram_helper.h +0 -1
                        bool *is_iomem, struct ttm_bo_kmap_obj *kmap);
--

2.21.0

--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

Quoted 4 lines

--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

#8 lore raw

Re: [PATCH 1/2] drm: Add drm_gem_vram_{pin/unpin}_reserved() and convert mgag200

Quoted 66 lines · Christian

Am 20.05.19 um 18:26 schrieb Daniel Vetter:

Quoted 55 lines

[CAUTION: External Email]

Quoted 52 lines · Daniel Vetter
Quoted 31 lines · Thomas Zimmermann

The new interfaces drm_gem_vram_{pin/unpin}_reserved() are variants of the GEM VRAM pin/unpin functions that do not reserve the BO during validation. The mgag200 driver requires this behavior for its cursor handling. The patch also converts the driver to use the new interfaces.

---
  drivers/gpu/drm/drm_gem_vram_helper.c    | 75 ++++++++++++++++++++++++
  drivers/gpu/drm/mgag200/mgag200_cursor.c | 18 +++---
  include/drm/drm_gem_vram_helper.h        |  3 +
  3 files changed, 88 insertions(+), 8 deletions(-)
drivers/gpu/drm/drm_gem_vram_helper.c +11 -0
index 8f142b810eb4..a002c03eaf4c 100644
@@ -254,6 +254,47 @@ int drm_gem_vram_pin(struct drm_gem_vram_object *gbo, unsigned long pl_flag)
  }
  EXPORT_SYMBOL(drm_gem_vram_pin);
 
+/**
+ * drm_gem_vram_pin_reserved() - Pins a GEM VRAM object in a region.
+ * @gbo:   the GEM VRAM object
+ * @pl_flag:       a bitmask of possible memory regions
+ *
+ * Pinning a buffer object ensures that it is not evicted from
+ * a memory region. A pinned buffer object has to be unpinned before
+ * it can be pinned to another region.
+ *
+ * This function pins a GEM VRAM object that has already been
+ * reserved. Use drm_gem_vram_pin() if possible.

drivers/gpu/drm/drm_gem_vram_helper.c int drm_gem_vram_pin(struct drm_gem_vram_object *gbo, unsigned long pl_flag) · line 277

Quoted 10 lines
drivers/gpu/drm/drm_gem_vram_helper.c +10 -0
+ *
+ * Returns:
+ * 0 on success, or
+ * a negative error code otherwise.
+ */
+int drm_gem_vram_pin_reserved(struct drm_gem_vram_object *gbo,
+                         unsigned long pl_flag)
+{
+   int i, ret;
+   struct ttm_operation_ctx ctx = { false, false };

I think would be good to have a lockdep_assert_held here for the ww_mutex.

Also general thing: _reserved is kinda ttm lingo, for dma-buf reservations we call the structure tracking the fences+lock the "reservation", but the naming scheme used is _lock/_unlock.

I think would be good to be consistent with that, and use _locked here. Especially for a very simplified vram helper like this one I expect that's going to lead to less wtf moments by driver writers :-)

Quoted 8 lines
Quoted 2 lines

Maybe we should also do a large-scale s/reserve/lock/ within ttm, to align more with what we now have in dma-buf.

More aside:

Could be a good move to demidlayer this an essentially remove ttm_bo_reserve as a wrapper around the linux/reservation.h functions. Not sure whether that aligns with Christian's plans. TODO.rst patch might be a good step to get that discussion started.

Quoted 6 lines

Well what ttm_bo_reserve()/_unreserve() does is a) lock/unlock the reservation object and b) remove the BO from the LRU.

Since I'm desperately trying to get rid of the LRU removal right now we sooner or later should be able to remove ttmo_bo_reserve()/_unreserve() as well (or at least replace them with tiny ttm_bo_lock() wrappers.

Yeah that's what I meant, would be nice to ditch that bit of midlayering, at least from drivers.

I guess within ttm we'd still need ttm_bo_lock/unlock, since interrupteble/timedout/lock modes in general is passed around throughout the entire stack as parameters.
-Daniel

Quoted 149 lines

Christian.

Quoted 145 lines

-Daniel

Quoted 131 lines

Cheers, Daniel

Quoted 129 lines
drivers/gpu/drm/drm_gem_vram_helper.c +54 -0
+
+   if (gbo->pin_count) {
+           ++gbo->pin_count;
+           return 0;
+   }
+
+   drm_gem_vram_placement(gbo, pl_flag);
+   for (i = 0; i < gbo->placement.num_placement; ++i)
+           gbo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
+
+   ret = ttm_bo_validate(&gbo->bo, &gbo->placement, &ctx);
+   if (ret < 0)
+           return ret;
+
+   gbo->pin_count = 1;
+
+   return 0;
+}
+EXPORT_SYMBOL(drm_gem_vram_pin_reserved);
+
  /**
   * drm_gem_vram_unpin() - Unpins a GEM VRAM object
   * @gbo:   the GEM VRAM object
@@ -285,6 +326,40 @@ int drm_gem_vram_unpin(struct drm_gem_vram_object *gbo)
  }
  EXPORT_SYMBOL(drm_gem_vram_unpin);
 
+/**
+ * drm_gem_vram_unpin_reserved() - Unpins a GEM VRAM object
+ * @gbo:   the GEM VRAM object
+ *
+ * This function unpins a GEM VRAM object that has already been
+ * reserved. Use drm_gem_vram_unpin() if possible.
+ *
+ * Returns:
+ * 0 on success, or
+ * a negative error code otherwise.
+ */
+int drm_gem_vram_unpin_reserved(struct drm_gem_vram_object *gbo)
+{
+   int i, ret;
+   struct ttm_operation_ctx ctx = { false, false };
+
+   if (WARN_ON_ONCE(!gbo->pin_count))
+           return 0;
+
+   --gbo->pin_count;
+   if (gbo->pin_count)
+           return 0;
+
+   for (i = 0; i < gbo->placement.num_placement ; ++i)
+           gbo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
+
+   ret = ttm_bo_validate(&gbo->bo, &gbo->placement, &ctx);
+   if (ret < 0)
+           return ret;
+
+   return 0;
+}
+EXPORT_SYMBOL(drm_gem_vram_unpin_reserved);
+
  /**
   * drm_gem_vram_push_to_system() - \
     Unpins a GEM VRAM object and moves it to system memory
drivers/gpu/drm/mgag200/mgag200_cursor.c +10 -8
index 6c1a9d724d85..1c4fc85315a0 100644
@@ -23,9 +23,9 @@ static void mga_hide_cursor(struct mga_device *mdev)
     WREG8(MGA_CURPOSXL, 0);
     WREG8(MGA_CURPOSXH, 0);
     if (mdev->cursor.pixels_1->pin_count)
-           drm_gem_vram_unpin(mdev->cursor.pixels_1);
+           drm_gem_vram_unpin_reserved(mdev->cursor.pixels_1);
     if (mdev->cursor.pixels_2->pin_count)
-           drm_gem_vram_unpin(mdev->cursor.pixels_2);
+           drm_gem_vram_unpin_reserved(mdev->cursor.pixels_2);
  }
 
  int mga_crtc_cursor_set(struct drm_crtc *crtc,
@@ -96,26 +96,28 @@ int mga_crtc_cursor_set(struct drm_crtc *crtc,
 
     /* Move cursor buffers into VRAM if they aren't already */
     if (!pixels_1->pin_count) {
-           ret = drm_gem_vram_pin(pixels_1, DRM_GEM_VRAM_PL_FLAG_VRAM);
+           ret = drm_gem_vram_pin_reserved(pixels_1,
+                                           DRM_GEM_VRAM_PL_FLAG_VRAM);
             if (ret)
                     goto out1;
             gpu_addr = drm_gem_vram_offset(pixels_1);
             if (gpu_addr < 0) {
-                   drm_gem_vram_unpin(pixels_1);
+                   drm_gem_vram_unpin_reserved(pixels_1);
                     goto out1;
             }
             mdev->cursor.pixels_1_gpu_addr = gpu_addr;
     }
     if (!pixels_2->pin_count) {
-           ret = drm_gem_vram_pin(pixels_2, DRM_GEM_VRAM_PL_FLAG_VRAM);
+           ret = drm_gem_vram_pin_reserved(pixels_2,
+                                           DRM_GEM_VRAM_PL_FLAG_VRAM);
             if (ret) {
-                   drm_gem_vram_unpin(pixels_1);
+                   drm_gem_vram_unpin_reserved(pixels_1);
                     goto out1;
             }
             gpu_addr = drm_gem_vram_offset(pixels_2);
             if (gpu_addr < 0) {
-                   drm_gem_vram_unpin(pixels_1);
-                   drm_gem_vram_unpin(pixels_2);
+                   drm_gem_vram_unpin_reserved(pixels_1);
+                   drm_gem_vram_unpin_reserved(pixels_2);
                     goto out1;
             }
             mdev->cursor.pixels_2_gpu_addr = gpu_addr;
include/drm/drm_gem_vram_helper.h +3 -0
index b056f189ba62..ff1a81761543 100644
@@ -82,7 +82,10 @@ void drm_gem_vram_unreserve(struct drm_gem_vram_object *gbo);
  u64 drm_gem_vram_mmap_offset(struct drm_gem_vram_object *gbo);
  s64 drm_gem_vram_offset(struct drm_gem_vram_object *gbo);
  int drm_gem_vram_pin(struct drm_gem_vram_object *gbo, unsigned long pl_flag);
+int drm_gem_vram_pin_reserved(struct drm_gem_vram_object *gbo,
+                         unsigned long pl_flag);
  int drm_gem_vram_unpin(struct drm_gem_vram_object *gbo);
+int drm_gem_vram_unpin_reserved(struct drm_gem_vram_object *gbo);
  int drm_gem_vram_push_to_system(struct drm_gem_vram_object *gbo);
  void *drm_gem_vram_kmap_at(struct drm_gem_vram_object *gbo, bool map,
Quoted 8 lines
Quoted 4 lines
include/drm/drm_gem_vram_helper.h +0 -1
                        bool *is_iomem, struct ttm_bo_kmap_obj *kmap);
--

2.21.0

--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
#9 lore raw

Re: [PATCH 1/2] drm: Add drm_gem_vram_{pin/unpin}_reserved() and convert mgag200

In reply to Daniel Vetter

Hi,

Quoted 6 lines

I think would be good to have a lockdep_assert_held here for the ww_mutex.

Also general thing: _reserved is kinda ttm lingo, for dma-buf reservations we call the structure tracking the fences+lock the "reservation", but the naming scheme used is _lock/_unlock.

Quoted 6 lines

I think would be good to be consistent with that, and use _locked here. Especially for a very simplified vram helper like this one I expect that's going to lead to less wtf moments by driver writers :-)

Maybe we should also do a large-scale s/reserve/lock/ within ttm, to align more with what we now have in dma-buf.

Given that mgag200 is the only user I think the best way forward is to improve the mgag200 cursor handling so we can just drop the _reserved variants ...

When looking at mga_crtc_cursor_set() I suspect the easierst way to do that would be to simply pin the cursor bo's at driver_load time, then we don't have to bother with pinning in mga_crtc_cursor_set() at all.

Thomas, as you have test hardware, can you look into this?

thanks,
Gerd

#10 lore raw

Re: [PATCH 1/2] drm: Add drm_gem_vram_{pin/unpin}_reserved() and convert mgag200

In reply to Daniel Vetter

Hi,

Quoted 6 lines

I think would be good to have a lockdep_assert_held here for the ww_mutex.

Also general thing: _reserved is kinda ttm lingo, for dma-buf reservations we call the structure tracking the fences+lock the "reservation", but the naming scheme used is _lock/_unlock.

Quoted 6 lines

I think would be good to be consistent with that, and use _locked here. Especially for a very simplified vram helper like this one I expect that's going to lead to less wtf moments by driver writers :-)

Maybe we should also do a large-scale s/reserve/lock/ within ttm, to align more with what we now have in dma-buf.

Given that mgag200 is the only user I think the best way forward is to improve the mgag200 cursor handling so we can just drop the _reserved variants ...

When looking at mga_crtc_cursor_set() I suspect the easierst way to do that would be to simply pin the cursor bo's at driver_load time, then we don't have to bother with pinning in mga_crtc_cursor_set() at all.

Thomas, as you have test hardware, can you look into this?

thanks,
Gerd

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

author maintainer
#11 lore raw

Re: [PATCH 1/2] drm: Add drm_gem_vram_{pin/unpin}_reserved() and convert mgag200

Hi

Am 21.05.19 um 12:35 schrieb Gerd Hoffmann:

Quoted 19 lines

Hi,

Quoted 6 lines

I think would be good to have a lockdep_assert_held here for the ww_mutex.

Also general thing: _reserved is kinda ttm lingo, for dma-buf reservations we call the structure tracking the fences+lock the "reservation", but the naming scheme used is _lock/_unlock.

Quoted 6 lines

I think would be good to be consistent with that, and use _locked here. Especially for a very simplified vram helper like this one I expect that's going to lead to less wtf moments by driver writers :-)

Maybe we should also do a large-scale s/reserve/lock/ within ttm, to align more with what we now have in dma-buf.

Given that mgag200 is the only user I think the best way forward is to improve the mgag200 cursor handling so we can just drop the _reserved variants ...

Quoted 5 lines

When looking at mga_crtc_cursor_set() I suspect the easierst way to do that would be to simply pin the cursor bo's at driver_load time, then we don't have to bother with pinning in mga_crtc_cursor_set() at all.

Thomas, as you have test hardware, can you look into this?

Sure, that's the plan. May take a few days, though.

I'd like to see all of the locking/reservation gone from the helper API. By using bochs' PRIME helpers, the generic console could probably be used in mgag200 and ast. That would remove the remaining calls from mgag200 and ast.

Best regards
Thomas

Quoted 8 lines

thanks,
Gerd

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Signature · 5 lines
Thomas Zimmermann
Graphics Driver Developer
SUSE Linux GmbH, Maxfeldstrasse 5, 90409 Nuernberg, Germany
GF: Felix Imendörffer, Mary Higgins, Sri Rasiah
HRB 21284 (AG Nürnberg)

_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

author maintainer
#12 lore raw

Re: [PATCH 1/2] drm: Add drm_gem_vram_{pin/unpin}_reserved() and convert mgag200

In reply to Gerd Hoffmann

Hi

Am 21.05.19 um 12:35 schrieb Gerd Hoffmann:

Quoted 19 lines

Hi,

Quoted 6 lines

I think would be good to have a lockdep_assert_held here for the ww_mutex.

Also general thing: _reserved is kinda ttm lingo, for dma-buf reservations we call the structure tracking the fences+lock the "reservation", but the naming scheme used is _lock/_unlock.

Quoted 6 lines

I think would be good to be consistent with that, and use _locked here. Especially for a very simplified vram helper like this one I expect that's going to lead to less wtf moments by driver writers :-)

Maybe we should also do a large-scale s/reserve/lock/ within ttm, to align more with what we now have in dma-buf.

Given that mgag200 is the only user I think the best way forward is to improve the mgag200 cursor handling so we can just drop the _reserved variants ...

Quoted 5 lines

When looking at mga_crtc_cursor_set() I suspect the easierst way to do that would be to simply pin the cursor bo's at driver_load time, then we don't have to bother with pinning in mga_crtc_cursor_set() at all.

Thomas, as you have test hardware, can you look into this?

Sure, that's the plan. May take a few days, though.

I'd like to see all of the locking/reservation gone from the helper API. By using bochs' PRIME helpers, the generic console could probably be used in mgag200 and ast. That would remove the remaining calls from mgag200 and ast.

Best regards
Thomas

Quoted 8 lines

thanks,
Gerd

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Signature · 5 lines
Thomas Zimmermann
Graphics Driver Developer
SUSE Linux GmbH, Maxfeldstrasse 5, 90409 Nuernberg, Germany
GF: Felix Imendörffer, Mary Higgins, Sri Rasiah
HRB 21284 (AG Nürnberg)

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

author maintainer
#13 lore raw

Re: [PATCH 1/2] drm: Add drm_gem_vram_{pin/unpin}_reserved() and convert mgag200

In reply to Gerd Hoffmann

Hi

Am 21.05.19 um 12:35 schrieb Gerd Hoffmann:

Quoted 15 lines

Hi,

Quoted 6 lines

I think would be good to have a lockdep_assert_held here for the ww_mutex.

Also general thing: _reserved is kinda ttm lingo, for dma-buf reservations we call the structure tracking the fences+lock the "reservation", but the naming scheme used is _lock/_unlock.

Quoted 6 lines

I think would be good to be consistent with that, and use _locked here. Especially for a very simplified vram helper like this one I expect that's going to lead to less wtf moments by driver writers :-)

Maybe we should also do a large-scale s/reserve/lock/ within ttm, to align more with what we now have in dma-buf.

Quoted 7 lines

Given that mgag200 is the only user I think the best way forward is to improve the mgag200 cursor handling so we can just drop the _reserved variants ...

When looking at mga_crtc_cursor_set() I suspect the easierst way to do that would be to simply pin the cursor bo's at driver_load time, then we don't have to bother with pinning in mga_crtc_cursor_set() at all.

I've been thinking about converting ast and mgag200 to atomic mode setting. For that, universal planes would be required. But this seems incompatible with HW cursors.

With universal planes, cursor planes would probably have to be exported as RGBA8 framebuffer, but the HW uses MMIO region and 16-color palette. So there's got to be a translation step and a way of notifying userspace if the palette overflows. In the current driver code, drm_crtc_funcs.set_cursor() does this.

How is this done with universal planes? I know that there's drm_framebuffer_funcs.dirty(), but it doesn't seem to be used much.

Best regards
Thomas

Quoted 10 lines

Thomas, as you have test hardware, can you look into this?

thanks,
Gerd

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Signature · 5 lines
Thomas Zimmermann
Graphics Driver Developer
SUSE Linux GmbH, Maxfeldstrasse 5, 90409 Nuernberg, Germany
GF: Felix Imendörffer, Mary Higgins, Sri Rasiah
HRB 21284 (AG Nürnberg)

_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

author maintainer
#14 lore raw

Re: [PATCH 1/2] drm: Add drm_gem_vram_{pin/unpin}_reserved() and convert mgag200

In reply to Gerd Hoffmann

Hi

Am 21.05.19 um 12:35 schrieb Gerd Hoffmann:

Quoted 15 lines

Hi,

Quoted 6 lines

I think would be good to have a lockdep_assert_held here for the ww_mutex.

Also general thing: _reserved is kinda ttm lingo, for dma-buf reservations we call the structure tracking the fences+lock the "reservation", but the naming scheme used is _lock/_unlock.

Quoted 6 lines

I think would be good to be consistent with that, and use _locked here. Especially for a very simplified vram helper like this one I expect that's going to lead to less wtf moments by driver writers :-)

Maybe we should also do a large-scale s/reserve/lock/ within ttm, to align more with what we now have in dma-buf.

Quoted 7 lines

Given that mgag200 is the only user I think the best way forward is to improve the mgag200 cursor handling so we can just drop the _reserved variants ...

When looking at mga_crtc_cursor_set() I suspect the easierst way to do that would be to simply pin the cursor bo's at driver_load time, then we don't have to bother with pinning in mga_crtc_cursor_set() at all.

I've been thinking about converting ast and mgag200 to atomic mode setting. For that, universal planes would be required. But this seems incompatible with HW cursors.

With universal planes, cursor planes would probably have to be exported as RGBA8 framebuffer, but the HW uses MMIO region and 16-color palette. So there's got to be a translation step and a way of notifying userspace if the palette overflows. In the current driver code, drm_crtc_funcs.set_cursor() does this.

How is this done with universal planes? I know that there's drm_framebuffer_funcs.dirty(), but it doesn't seem to be used much.

Best regards
Thomas

Quoted 10 lines

Thomas, as you have test hardware, can you look into this?

thanks,
Gerd

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Signature · 5 lines
Thomas Zimmermann
Graphics Driver Developer
SUSE Linux GmbH, Maxfeldstrasse 5, 90409 Nuernberg, Germany
GF: Felix Imendörffer, Mary Higgins, Sri Rasiah
HRB 21284 (AG Nürnberg)

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

author maintainer
#15 lore raw

PATCH 2/2 drm: Reserve/unreserve GEM VRAM BOs from within pin/unpin functions

In reply to Thomas Zimmermann

The original bochs and vbox implementations of pin and unpin functions automatically reserved BOs during validation. This functionality got lost while converting the code to a generic implementation. This may result in validating unlocked TTM BOs.

Adding the reserve and unreserve operations to GEM VRAM's pin and unpin functions fixes the bochs and vbox drivers. Additionally the patch changes the mgag200, ast and hibmc drivers to not reserve BOs by themselves.

Fixes a3232987fdbf0bede92a9d7c7e2db99a5084d31b ("drm/bochs: Convert [...]")
---
 drivers/gpu/drm/ast/ast_mode.c                | 24 +--------
 drivers/gpu/drm/drm_gem_vram_helper.c         | 54 ++++++++++++++-----
 .../gpu/drm/hisilicon/hibmc/hibmc_drm_de.c    |  6 ---
 .../gpu/drm/hisilicon/hibmc/hibmc_drm_fbdev.c | 17 +-----
 drivers/gpu/drm/mgag200/mgag200_mode.c        | 19 +------
 5 files changed, 45 insertions(+), 75 deletions(-)
drivers/gpu/drm/ast/ast_mode.c +1 -23
index 3475591a22c3..9aca9135a5cc 100644
@@ -539,24 +539,16 @@ static int ast_crtc_do_set_base(struct drm_crtc *crtc,
 		ast_fb = to_ast_framebuffer(fb);
 		obj = ast_fb->obj;
 		gbo = drm_gem_vram_of_gem(obj);
-		ret = drm_gem_vram_reserve(gbo, false);
-		if (ret)
-			return ret;
 		drm_gem_vram_push_to_system(gbo);
-		drm_gem_vram_unreserve(gbo);
 	}
 
 	ast_fb = to_ast_framebuffer(crtc->primary->fb);
 	obj = ast_fb->obj;
 	gbo = drm_gem_vram_of_gem(obj);
 
-	ret = drm_gem_vram_reserve(gbo, false);
-	if (ret)
-		return ret;
-
 	ret = drm_gem_vram_pin(gbo, DRM_GEM_VRAM_PL_FLAG_VRAM);
 	if (ret)
-		goto err_drm_gem_vram_unreserve;
+		return ret;
 	gpu_addr = drm_gem_vram_offset(gbo);
 	if (gpu_addr < 0) {
 		ret = (int)gpu_addr;
@@ -573,7 +565,6 @@ static int ast_crtc_do_set_base(struct drm_crtc *crtc,
 			ast_fbdev_set_base(ast, gpu_addr);
 		}
 	}
-	drm_gem_vram_unreserve(gbo);
 
 	ast_set_offset_reg(crtc);
 	ast_set_start_address_crt1(crtc, (u32)gpu_addr);
@@ -582,8 +573,6 @@ static int ast_crtc_do_set_base(struct drm_crtc *crtc,
 
 err_drm_gem_vram_unpin:
 	drm_gem_vram_unpin(gbo);
-err_drm_gem_vram_unreserve:
-	drm_gem_vram_unreserve(gbo);
 	return ret;
 }
 
@@ -630,8 +619,6 @@ static int ast_crtc_mode_set(struct drm_crtc *crtc,
 
 static void ast_crtc_disable(struct drm_crtc *crtc)
 {
-	int ret;
-
 	DRM_DEBUG_KMS("\n");
 	ast_crtc_dpms(crtc, DRM_MODE_DPMS_OFF);
 	if (crtc->primary->fb) {
@@ -639,11 +626,7 @@ static void ast_crtc_disable(struct drm_crtc *crtc)
 		struct drm_gem_object *obj = ast_fb->obj;
 		struct drm_gem_vram_object *gbo = drm_gem_vram_of_gem(obj);
 
-		ret = drm_gem_vram_reserve(gbo, false);
-		if (ret)
-			return;
 		drm_gem_vram_push_to_system(gbo);
-		drm_gem_vram_unreserve(gbo);
 	}
 	crtc->primary->fb = NULL;
 }
@@ -939,12 +922,7 @@ static int ast_cursor_init(struct drm_device *dev)
 	if (ret)
 		return ret;
 	gbo = drm_gem_vram_of_gem(obj);
-	ret = drm_gem_vram_reserve(gbo, false);
-	if (unlikely(ret != 0))
-		goto fail;
-
 	ret = drm_gem_vram_pin(gbo, DRM_GEM_VRAM_PL_FLAG_VRAM);
-	drm_gem_vram_unreserve(gbo);
 	if (ret)
 		goto fail;
 	gpu_addr = drm_gem_vram_offset(gbo);
drivers/gpu/drm/drm_gem_vram_helper.c +42 -12

105 lines not on this page · raw message

drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c +0 -6
index db0dfa57844e..fbdf495779e0 100644
@@ -107,14 +107,8 @@ static void hibmc_plane_atomic_update(struct drm_plane *plane,
 
 	hibmc_fb = to_hibmc_framebuffer(state->fb);
 	gbo = drm_gem_vram_of_gem(hibmc_fb->obj);
-	ret = drm_gem_vram_reserve(gbo, false);
-	if (ret) {
-		DRM_ERROR("failed to reserve BO: %d", ret);
-		return;
-	}
 
 	ret = drm_gem_vram_pin(gbo, DRM_GEM_VRAM_PL_FLAG_VRAM);
-	drm_gem_vram_unreserve(gbo);
 	if (ret) {
 		DRM_ERROR("failed to pin bo: %d", ret);
 		return;
drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_fbdev.c +1 -16
index 9d2025fa16f8..bd5fbb23973a 100644
@@ -63,7 +63,6 @@ static int hibmc_drm_fb_create(struct drm_fb_helper *helper,
 	struct drm_mode_fb_cmd2 mode_cmd;
 	struct drm_gem_object *gobj = NULL;
 	int ret = 0;
-	int ret1;
 	size_t size;
 	unsigned int bytes_per_pixel;
 	struct drm_gem_vram_object *gbo = NULL;
@@ -91,16 +90,10 @@ static int hibmc_drm_fb_create(struct drm_fb_helper *helper,
 
 	gbo = drm_gem_vram_of_gem(gobj);
 
-	ret = drm_gem_vram_reserve(gbo, false);
-	if (ret) {
-		DRM_ERROR("failed to reserve bo: %d\n", ret);
-		goto out_unref_gem;
-	}
-
 	ret = drm_gem_vram_pin(gbo, DRM_GEM_VRAM_PL_FLAG_VRAM);
 	if (ret) {
 		DRM_ERROR("failed to pin fbcon: %d\n", ret);
-		goto out_unreserve_ttm_bo;
+		goto out_unref_gem;
 	}
 
 	base = drm_gem_vram_kmap(gbo, true, NULL);
@@ -109,7 +102,6 @@ static int hibmc_drm_fb_create(struct drm_fb_helper *helper,
 		DRM_ERROR("failed to kmap fbcon: %d\n", ret);
 		goto out_unpin_bo;
 	}
-	drm_gem_vram_unreserve(gbo);
 
 	info = drm_fb_helper_alloc_fbi(helper);
 	if (IS_ERR(info)) {
@@ -141,16 +133,9 @@ static int hibmc_drm_fb_create(struct drm_fb_helper *helper,
 	return 0;
 
 out_release_fbi:
-	ret1 = drm_gem_vram_reserve(gbo, false);
-	if (ret1) {
-		DRM_ERROR("failed to rsv ttm_bo when release fbi: %d\n", ret1);
-		goto out_unref_gem;
-	}
 	drm_gem_vram_kunmap(gbo);
 out_unpin_bo:
 	drm_gem_vram_unpin(gbo);
-out_unreserve_ttm_bo:
-	drm_gem_vram_unreserve(gbo);
 out_unref_gem:
 	drm_gem_object_put_unlocked(gobj);
 
drivers/gpu/drm/mgag200/mgag200_mode.c +1 -18
index 3098bf5c1744..e79872c968bf 100644
@@ -877,24 +877,16 @@ static int mga_crtc_do_set_base(struct drm_crtc *crtc,
 		mga_fb = to_mga_framebuffer(fb);
 		obj = mga_fb->obj;
 		gbo = drm_gem_vram_of_gem(obj);
-		ret = drm_gem_vram_reserve(gbo, false);
-		if (ret)
-			return ret;
 		drm_gem_vram_push_to_system(gbo);
-		drm_gem_vram_unreserve(gbo);
 	}
 
 	mga_fb = to_mga_framebuffer(crtc->primary->fb);
 	obj = mga_fb->obj;
 	gbo = drm_gem_vram_of_gem(obj);
 
-	ret = drm_gem_vram_reserve(gbo, false);
-	if (ret)
-		return ret;
-
 	ret = drm_gem_vram_pin(gbo, DRM_GEM_VRAM_PL_FLAG_VRAM);
 	if (ret)
-		goto err_drm_gem_vram_unreserve;
+		return ret;
 	gpu_addr = drm_gem_vram_offset(gbo);
 	if (gpu_addr < 0) {
 		ret = (int)gpu_addr;
@@ -910,16 +902,12 @@ static int mga_crtc_do_set_base(struct drm_crtc *crtc,
 		}
 	}
 
-	drm_gem_vram_unreserve(gbo);
-
 	mga_set_start_address(crtc, (u32)gpu_addr);
 
 	return 0;
 
 err_drm_gem_vram_unpin:
 	drm_gem_vram_unpin(gbo);
-err_drm_gem_vram_unreserve:
-	drm_gem_vram_unreserve(gbo);
 	return ret;
 }
 
@@ -1434,7 +1422,6 @@ static void mga_crtc_destroy(struct drm_crtc *crtc)
 
 static void mga_crtc_disable(struct drm_crtc *crtc)
 {
-	int ret;
 	DRM_DEBUG_KMS("\n");
 	mga_crtc_dpms(crtc, DRM_MODE_DPMS_OFF);
 	if (crtc->primary->fb) {
@@ -1442,11 +1429,7 @@ static void mga_crtc_disable(struct drm_crtc *crtc)
 		struct drm_gem_object *obj = mga_fb->obj;
 		struct drm_gem_vram_object *gbo = drm_gem_vram_of_gem(obj);
 
-		ret = drm_gem_vram_reserve(gbo, false);
-		if (ret)
-			return;
 		drm_gem_vram_push_to_system(gbo);
-		drm_gem_vram_unreserve(gbo);
 	}
 	crtc->primary->fb = NULL;
 }
Signature · 1 line
2.21.0
author maintainer
#16 lore raw

PATCH 2/2 drm: Reserve/unreserve GEM VRAM BOs from within pin/unpin functions

In reply to Thomas Zimmermann

The original bochs and vbox implementations of pin and unpin functions automatically reserved BOs during validation. This functionality got lost while converting the code to a generic implementation. This may result in validating unlocked TTM BOs.

Adding the reserve and unreserve operations to GEM VRAM's pin and unpin functions fixes the bochs and vbox drivers. Additionally the patch changes the mgag200, ast and hibmc drivers to not reserve BOs by themselves.

Fixes a3232987fdbf0bede92a9d7c7e2db99a5084d31b ("drm/bochs: Convert [...]")
---
 drivers/gpu/drm/ast/ast_mode.c                | 24 +--------
 drivers/gpu/drm/drm_gem_vram_helper.c         | 54 ++++++++++++++-----
 .../gpu/drm/hisilicon/hibmc/hibmc_drm_de.c    |  6 ---
 .../gpu/drm/hisilicon/hibmc/hibmc_drm_fbdev.c | 17 +-----
 drivers/gpu/drm/mgag200/mgag200_mode.c        | 19 +------
 5 files changed, 45 insertions(+), 75 deletions(-)
drivers/gpu/drm/ast/ast_mode.c +1 -23
index 3475591a22c3..9aca9135a5cc 100644
@@ -539,24 +539,16 @@ static int ast_crtc_do_set_base(struct drm_crtc *crtc,
 		ast_fb = to_ast_framebuffer(fb);
 		obj = ast_fb->obj;
 		gbo = drm_gem_vram_of_gem(obj);
-		ret = drm_gem_vram_reserve(gbo, false);
-		if (ret)
-			return ret;
 		drm_gem_vram_push_to_system(gbo);
-		drm_gem_vram_unreserve(gbo);
 	}
 
 	ast_fb = to_ast_framebuffer(crtc->primary->fb);
 	obj = ast_fb->obj;
 	gbo = drm_gem_vram_of_gem(obj);
 
-	ret = drm_gem_vram_reserve(gbo, false);
-	if (ret)
-		return ret;
-
 	ret = drm_gem_vram_pin(gbo, DRM_GEM_VRAM_PL_FLAG_VRAM);
 	if (ret)
-		goto err_drm_gem_vram_unreserve;
+		return ret;
 	gpu_addr = drm_gem_vram_offset(gbo);
 	if (gpu_addr < 0) {
 		ret = (int)gpu_addr;
@@ -573,7 +565,6 @@ static int ast_crtc_do_set_base(struct drm_crtc *crtc,
 			ast_fbdev_set_base(ast, gpu_addr);
 		}
 	}
-	drm_gem_vram_unreserve(gbo);
 
 	ast_set_offset_reg(crtc);
 	ast_set_start_address_crt1(crtc, (u32)gpu_addr);
@@ -582,8 +573,6 @@ static int ast_crtc_do_set_base(struct drm_crtc *crtc,
 
 err_drm_gem_vram_unpin:
 	drm_gem_vram_unpin(gbo);
-err_drm_gem_vram_unreserve:
-	drm_gem_vram_unreserve(gbo);
 	return ret;
 }
 
@@ -630,8 +619,6 @@ static int ast_crtc_mode_set(struct drm_crtc *crtc,
 
 static void ast_crtc_disable(struct drm_crtc *crtc)
 {
-	int ret;
-
 	DRM_DEBUG_KMS("\n");
 	ast_crtc_dpms(crtc, DRM_MODE_DPMS_OFF);
 	if (crtc->primary->fb) {
@@ -639,11 +626,7 @@ static void ast_crtc_disable(struct drm_crtc *crtc)
 		struct drm_gem_object *obj = ast_fb->obj;
 		struct drm_gem_vram_object *gbo = drm_gem_vram_of_gem(obj);
 
-		ret = drm_gem_vram_reserve(gbo, false);
-		if (ret)
-			return;
 		drm_gem_vram_push_to_system(gbo);
-		drm_gem_vram_unreserve(gbo);
 	}
 	crtc->primary->fb = NULL;
 }
@@ -939,12 +922,7 @@ static int ast_cursor_init(struct drm_device *dev)
 	if (ret)
 		return ret;
 	gbo = drm_gem_vram_of_gem(obj);
-	ret = drm_gem_vram_reserve(gbo, false);
-	if (unlikely(ret != 0))
-		goto fail;
-
 	ret = drm_gem_vram_pin(gbo, DRM_GEM_VRAM_PL_FLAG_VRAM);
-	drm_gem_vram_unreserve(gbo);
 	if (ret)
 		goto fail;
 	gpu_addr = drm_gem_vram_offset(gbo);
drivers/gpu/drm/drm_gem_vram_helper.c +42 -12

105 lines not on this page · raw message

drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_de.c +0 -6
index db0dfa57844e..fbdf495779e0 100644
@@ -107,14 +107,8 @@ static void hibmc_plane_atomic_update(struct drm_plane *plane,
 
 	hibmc_fb = to_hibmc_framebuffer(state->fb);
 	gbo = drm_gem_vram_of_gem(hibmc_fb->obj);
-	ret = drm_gem_vram_reserve(gbo, false);
-	if (ret) {
-		DRM_ERROR("failed to reserve BO: %d", ret);
-		return;
-	}
 
 	ret = drm_gem_vram_pin(gbo, DRM_GEM_VRAM_PL_FLAG_VRAM);
-	drm_gem_vram_unreserve(gbo);
 	if (ret) {
 		DRM_ERROR("failed to pin bo: %d", ret);
 		return;
drivers/gpu/drm/hisilicon/hibmc/hibmc_drm_fbdev.c +1 -16
index 9d2025fa16f8..bd5fbb23973a 100644
@@ -63,7 +63,6 @@ static int hibmc_drm_fb_create(struct drm_fb_helper *helper,
 	struct drm_mode_fb_cmd2 mode_cmd;
 	struct drm_gem_object *gobj = NULL;
 	int ret = 0;
-	int ret1;
 	size_t size;
 	unsigned int bytes_per_pixel;
 	struct drm_gem_vram_object *gbo = NULL;
@@ -91,16 +90,10 @@ static int hibmc_drm_fb_create(struct drm_fb_helper *helper,
 
 	gbo = drm_gem_vram_of_gem(gobj);
 
-	ret = drm_gem_vram_reserve(gbo, false);
-	if (ret) {
-		DRM_ERROR("failed to reserve bo: %d\n", ret);
-		goto out_unref_gem;
-	}
-
 	ret = drm_gem_vram_pin(gbo, DRM_GEM_VRAM_PL_FLAG_VRAM);
 	if (ret) {
 		DRM_ERROR("failed to pin fbcon: %d\n", ret);
-		goto out_unreserve_ttm_bo;
+		goto out_unref_gem;
 	}
 
 	base = drm_gem_vram_kmap(gbo, true, NULL);
@@ -109,7 +102,6 @@ static int hibmc_drm_fb_create(struct drm_fb_helper *helper,
 		DRM_ERROR("failed to kmap fbcon: %d\n", ret);
 		goto out_unpin_bo;
 	}
-	drm_gem_vram_unreserve(gbo);
 
 	info = drm_fb_helper_alloc_fbi(helper);
 	if (IS_ERR(info)) {
@@ -141,16 +133,9 @@ static int hibmc_drm_fb_create(struct drm_fb_helper *helper,
 	return 0;
 
 out_release_fbi:
-	ret1 = drm_gem_vram_reserve(gbo, false);
-	if (ret1) {
-		DRM_ERROR("failed to rsv ttm_bo when release fbi: %d\n", ret1);
-		goto out_unref_gem;
-	}
 	drm_gem_vram_kunmap(gbo);
 out_unpin_bo:
 	drm_gem_vram_unpin(gbo);
-out_unreserve_ttm_bo:
-	drm_gem_vram_unreserve(gbo);
 out_unref_gem:
 	drm_gem_object_put_unlocked(gobj);
 
drivers/gpu/drm/mgag200/mgag200_mode.c +1 -18
index 3098bf5c1744..e79872c968bf 100644
@@ -877,24 +877,16 @@ static int mga_crtc_do_set_base(struct drm_crtc *crtc,
 		mga_fb = to_mga_framebuffer(fb);
 		obj = mga_fb->obj;
 		gbo = drm_gem_vram_of_gem(obj);
-		ret = drm_gem_vram_reserve(gbo, false);
-		if (ret)
-			return ret;
 		drm_gem_vram_push_to_system(gbo);
-		drm_gem_vram_unreserve(gbo);
 	}
 
 	mga_fb = to_mga_framebuffer(crtc->primary->fb);
 	obj = mga_fb->obj;
 	gbo = drm_gem_vram_of_gem(obj);
 
-	ret = drm_gem_vram_reserve(gbo, false);
-	if (ret)
-		return ret;
-
 	ret = drm_gem_vram_pin(gbo, DRM_GEM_VRAM_PL_FLAG_VRAM);
 	if (ret)
-		goto err_drm_gem_vram_unreserve;
+		return ret;
 	gpu_addr = drm_gem_vram_offset(gbo);
 	if (gpu_addr < 0) {
 		ret = (int)gpu_addr;
@@ -910,16 +902,12 @@ static int mga_crtc_do_set_base(struct drm_crtc *crtc,
 		}
 	}
 
-	drm_gem_vram_unreserve(gbo);
-
 	mga_set_start_address(crtc, (u32)gpu_addr);
 
 	return 0;
 
 err_drm_gem_vram_unpin:
 	drm_gem_vram_unpin(gbo);
-err_drm_gem_vram_unreserve:
-	drm_gem_vram_unreserve(gbo);
 	return ret;
 }
 
@@ -1434,7 +1422,6 @@ static void mga_crtc_destroy(struct drm_crtc *crtc)
 
 static void mga_crtc_disable(struct drm_crtc *crtc)
 {
-	int ret;
 	DRM_DEBUG_KMS("\n");
 	mga_crtc_dpms(crtc, DRM_MODE_DPMS_OFF);
 	if (crtc->primary->fb) {
@@ -1442,11 +1429,7 @@ static void mga_crtc_disable(struct drm_crtc *crtc)
 		struct drm_gem_object *obj = mga_fb->obj;
 		struct drm_gem_vram_object *gbo = drm_gem_vram_of_gem(obj);
 
-		ret = drm_gem_vram_reserve(gbo, false);
-		if (ret)
-			return;
 		drm_gem_vram_push_to_system(gbo);
-		drm_gem_vram_unreserve(gbo);
 	}
 	crtc->primary->fb = NULL;
 }
Signature · 6 lines
2.21.0

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
#17 lore raw
In reply to Thomas Zimmermann

Hi,

Quoted 12 lines

It turns out that the bochs and vbox drivers automatically reserved and unreserved the BO from within their pin and unpin functions. The other drivers; ast, hibmc and mgag200; performed reservation explicitly. With the GEM VRAM conversion, automatic BO reservation within pin and unpin functions accidentally got lost. So for bochs and vbox, ttm_bo_validate() worked on unlocked BOs.

This patch set fixes the problem by adding automatic reservation to the implementation of drm_gem_vram_{pin,unpin,push_to_system}() to fix bochs and vbox. It removes explicit BO reservation around the pin, unpin and push-to-system calls in the ast, hibmc and mgag200 drivers.

Quoted 4 lines

The only exception is the cursor handling of mgag200. In this case, the mgag200 driver now calls drm_gem_vram_{pin,unpin}_reserved(), which works with reserved BOs. The respective code should be refactored in a future patch to work with the regular pin and unpin functions.

Looks good, pushed to drm-misc-next.

thanks,
Gerd

#18 lore raw
Quoted 14 lines · Gerd Hoffmann

Hi,

Quoted 12 lines

It turns out that the bochs and vbox drivers automatically reserved and unreserved the BO from within their pin and unpin functions. The other drivers; ast, hibmc and mgag200; performed reservation explicitly. With the GEM VRAM conversion, automatic BO reservation within pin and unpin functions accidentally got lost. So for bochs and vbox, ttm_bo_validate() worked on unlocked BOs.

This patch set fixes the problem by adding automatic reservation to the implementation of drm_gem_vram_{pin,unpin,push_to_system}() to fix bochs and vbox. It removes explicit BO reservation around the pin, unpin and push-to-system calls in the ast, hibmc and mgag200 drivers.

Quoted 6 lines
Quoted 4 lines

The only exception is the cursor handling of mgag200. In this case, the mgag200 driver now calls drm_gem_vram_{pin,unpin}_reserved(), which works with reserved BOs. The respective code should be refactored in a future patch to work with the regular pin and unpin functions.

Looks good, pushed to drm-misc-next.

I have a bit of design review (replied to patch 1), would be great if either of you could address that as a follow up.

Thanks, Daniel

Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch