[PATCH 0/6] Etnaviv cmdbuf suballocator

[PATCH 0/6] Etnaviv cmdbuf suballocator

[Lucas Stach]

Hi all,

the following patches introduce a cmduf suballocator in the Etnaviv
kernel driver, which has the following benefits:

1. Allocating and freeing a CMA buffer for each user command submission
   is taking a big toll on the CPU, as CMA is not exactly low overhead.
   By suballocating a single buffer we avoid all this overhead.

2. Less TLB flushes on MMUv2. Each time a new buffer gets mapped into
   the GPU address space on MMUv2 the TLBs need to be flushed. Mapping
   the suballocated area once allows to skip the TLB flushes (at least
   as long as userspace re-uses existing buffers).

3. No workarounds for GC3000 required anymore. The FE TLB flush on
   GC3000 doesn't work reliably, which required us to map the cmdbufs
   into the GPU address space at specific positions, which also isn't
   guaranteed to work if the address space is already crowded. Having
   a single static area for the cmdbufs side-steps the erratum completely.

If I can get reviews and/or enough testing, I would like to include this
code in kernel 4.11.

Regards,
Lucas

Lucas Stach (6):
  drm/etnaviv: always flush MMU TLBs on map/unmap
  drm/etnaviv: move cmdbuf de-/allocation into own file
  drm/etnaviv: wire up iova handling in new cmdbuf abstraction
  drm/etnaviv: get cmdbuf physical address through the cmdbuf
    abstraction
  drm/etnaviv: add cmdbuf suballocator
  Revert "drm/etnaviv: trick drm_mm into giving out a low IOVA"

 drivers/gpu/drm/etnaviv/Makefile             |   1 +
 drivers/gpu/drm/etnaviv/etnaviv_buffer.c     |  14 +--
 drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.c     | 153 +++++++++++++++++++++++++++
 drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.h     |  59 +++++++++++
 drivers/gpu/drm/etnaviv/etnaviv_drv.c        |   4 +-
 drivers/gpu/drm/etnaviv/etnaviv_dump.c       |   6 +-
 drivers/gpu/drm/etnaviv/etnaviv_gem_submit.c |   8 +-
 drivers/gpu/drm/etnaviv/etnaviv_gpu.c        |  61 ++++-------
 drivers/gpu/drm/etnaviv/etnaviv_gpu.h        |  28 +----
 drivers/gpu/drm/etnaviv/etnaviv_iommu_v2.c   |   4 +-
 drivers/gpu/drm/etnaviv/etnaviv_mmu.c        |  60 +++++------
 drivers/gpu/drm/etnaviv/etnaviv_mmu.h        |  10 +-
 12 files changed, 288 insertions(+), 120 deletions(-)
 create mode 100644 drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.c
 create mode 100644 drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.h

-- 
2.11.0

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

[PATCH 1/6] drm/etnaviv: always flush MMU TLBs on map/unmap

[Lucas Stach]

This ensures that the GPU isn't able to write into already freed
objects, as doing this in the IOVA reaper isn't enough, as the
gem_free_object path will also cause unmaps to happen.

On MMUv2 this also ensures that stale entries, which may have
been prefetched into the TLB will be purged.

The flush is low overhead, as it gets batched up with the next
user command buffer, so this isn't incuring an overhead for
each buffer map/unmap.

Signed-off-by: Lucas Stach <l.stach@pengutronix.de>
---
 drivers/gpu/drm/etnaviv/etnaviv_mmu.c | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)

diff --git a/drivers/gpu/drm/etnaviv/etnaviv_mmu.c b/drivers/gpu/drm/etnaviv/etnaviv_mmu.c
index fe0e85b41310..d1216516b947 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_mmu.c
+++ b/drivers/gpu/drm/etnaviv/etnaviv_mmu.c
@@ -193,11 +193,8 @@ static int etnaviv_iommu_find_iova(struct etnaviv_iommu *mmu,
 
 		/*
 		 * We removed enough mappings so that the new allocation will
-		 * succeed.  Ensure that the MMU will be flushed before the
-		 * associated commit requesting this mapping, and retry the
-		 * allocation one more time.
+		 * succeed, retry the allocation one more time.
 		 */
-		mmu->need_flush = true;
 	}
 
 	return ret;
@@ -249,6 +246,7 @@ int etnaviv_iommu_map_gem(struct etnaviv_iommu *mmu,
 	}
 
 	list_add_tail(&mapping->mmu_node, &mmu->mappings);
+	mmu->need_flush = true;
 	mutex_unlock(&mmu->lock);
 
 	return ret;
@@ -266,6 +264,7 @@ void etnaviv_iommu_unmap_gem(struct etnaviv_iommu *mmu,
 		etnaviv_iommu_remove_mapping(mmu, mapping);
 
 	list_del(&mapping->mmu_node);
+	mmu->need_flush = true;
 	mutex_unlock(&mmu->lock);
 }
 
-- 
2.11.0

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

[PATCH 2/6] drm/etnaviv: move cmdbuf de-/allocation into own file

[Lucas Stach]

This will get more complex with the following changes, so move it
into its own place.

Signed-off-by: Lucas Stach <l.stach@pengutronix.de>
---
 drivers/gpu/drm/etnaviv/Makefile             |  1 +
 drivers/gpu/drm/etnaviv/etnaviv_buffer.c     |  1 +
 drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.c     | 54 ++++++++++++++++++++++++++++
 drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.h     | 46 ++++++++++++++++++++++++
 drivers/gpu/drm/etnaviv/etnaviv_drv.c        |  1 +
 drivers/gpu/drm/etnaviv/etnaviv_dump.c       |  1 +
 drivers/gpu/drm/etnaviv/etnaviv_gem_submit.c |  5 +--
 drivers/gpu/drm/etnaviv/etnaviv_gpu.c        | 45 ++++-------------------
 drivers/gpu/drm/etnaviv/etnaviv_gpu.h        | 27 ++------------
 drivers/gpu/drm/etnaviv/etnaviv_iommu_v2.c   |  1 +
 drivers/gpu/drm/etnaviv/etnaviv_mmu.c        |  1 +
 11 files changed, 117 insertions(+), 66 deletions(-)
 create mode 100644 drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.c
 create mode 100644 drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.h

diff --git a/drivers/gpu/drm/etnaviv/Makefile b/drivers/gpu/drm/etnaviv/Makefile
index 1086e9876f91..4f76c992043f 100644
--- a/drivers/gpu/drm/etnaviv/Makefile
+++ b/drivers/gpu/drm/etnaviv/Makefile
@@ -1,6 +1,7 @@
 etnaviv-y := \
 	etnaviv_buffer.o \
 	etnaviv_cmd_parser.o \
+	etnaviv_cmdbuf.o \
 	etnaviv_drv.o \
 	etnaviv_dump.o \
 	etnaviv_gem_prime.o \
diff --git a/drivers/gpu/drm/etnaviv/etnaviv_buffer.c b/drivers/gpu/drm/etnaviv/etnaviv_buffer.c
index d9230132dfbc..9813b9b33553 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_buffer.c
+++ b/drivers/gpu/drm/etnaviv/etnaviv_buffer.c
@@ -15,6 +15,7 @@
  * this program.  If not, see <http://www.gnu.org/licenses/>.
  */
 
+#include "etnaviv_cmdbuf.h"
 #include "etnaviv_gpu.h"
 #include "etnaviv_gem.h"
 #include "etnaviv_mmu.h"
diff --git a/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.c b/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.c
new file mode 100644
index 000000000000..08f01ce63ff6
--- /dev/null
+++ b/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.c
@@ -0,0 +1,54 @@
+/*
+ * Copyright (C) 2017 Etnaviv Project
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License version 2 as published by
+ * the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+ * more details.
+ *
+ * You should have received a copy of the GNU General Public License along with
+ * this program.  If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#include "etnaviv_cmdbuf.h"
+#include "etnaviv_gpu.h"
+#include "etnaviv_mmu.h"
+
+struct etnaviv_cmdbuf *etnaviv_cmdbuf_new(struct etnaviv_gpu *gpu, u32 size,
+	size_t nr_bos)
+{
+	struct etnaviv_cmdbuf *cmdbuf;
+	size_t sz = size_vstruct(nr_bos, sizeof(cmdbuf->bo_map[0]),
+				 sizeof(*cmdbuf));
+
+	cmdbuf = kzalloc(sz, GFP_KERNEL);
+	if (!cmdbuf)
+		return NULL;
+
+	if (gpu->mmu->version == ETNAVIV_IOMMU_V2)
+		size = ALIGN(size, SZ_4K);
+
+	cmdbuf->vaddr = dma_alloc_wc(gpu->dev, size, &cmdbuf->paddr,
+				     GFP_KERNEL);
+	if (!cmdbuf->vaddr) {
+		kfree(cmdbuf);
+		return NULL;
+	}
+
+	cmdbuf->gpu = gpu;
+	cmdbuf->size = size;
+
+	return cmdbuf;
+}
+
+void etnaviv_cmdbuf_free(struct etnaviv_cmdbuf *cmdbuf)
+{
+	etnaviv_iommu_put_cmdbuf_va(cmdbuf->gpu, cmdbuf);
+	dma_free_wc(cmdbuf->gpu->dev, cmdbuf->size, cmdbuf->vaddr,
+		    cmdbuf->paddr);
+	kfree(cmdbuf);
+}
diff --git a/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.h b/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.h
new file mode 100644
index 000000000000..883f03960f95
--- /dev/null
+++ b/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.h
@@ -0,0 +1,46 @@
+/*
+ * Copyright (C) 2017 Etnaviv Project
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License version 2 as published by
+ * the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+ * more details.
+ *
+ * You should have received a copy of the GNU General Public License along with
+ * this program.  If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#ifndef __ETNAVIV_CMDBUF_H__
+#define __ETNAVIV_CMDBUF_H__
+
+#include <drm/drm_mm.h>
+#include <linux/types.h>
+
+struct etnaviv_cmdbuf {
+	/* device this cmdbuf is allocated for */
+	struct etnaviv_gpu *gpu;
+	/* user context key, must be unique between all active users */
+	struct etnaviv_file_private *ctx;
+	/* cmdbuf properties */
+	void *vaddr;
+	dma_addr_t paddr;
+	u32 size;
+	u32 user_size;
+	/* vram node used if the cmdbuf is mapped through the MMUv2 */
+	struct drm_mm_node vram_node;
+	/* fence after which this buffer is to be disposed */
+	struct dma_fence *fence;
+	/* target exec state */
+	u32 exec_state;
+	/* per GPU in-flight list */
+	struct list_head node;
+	/* BOs attached to this command buffer */
+	unsigned int nr_bos;
+	struct etnaviv_vram_mapping *bo_map[0];
+};
+
+#endif /* __ETNAVIV_CMDBUF_H__ */
diff --git a/drivers/gpu/drm/etnaviv/etnaviv_drv.c b/drivers/gpu/drm/etnaviv/etnaviv_drv.c
index 00368b14d08d..6b3ca97833e3 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_drv.c
+++ b/drivers/gpu/drm/etnaviv/etnaviv_drv.c
@@ -18,6 +18,7 @@
 #include <linux/of_platform.h>
 #include <drm/drm_of.h>
 
+#include "etnaviv_cmdbuf.h"
 #include "etnaviv_drv.h"
 #include "etnaviv_gpu.h"
 #include "etnaviv_gem.h"
diff --git a/drivers/gpu/drm/etnaviv/etnaviv_dump.c b/drivers/gpu/drm/etnaviv/etnaviv_dump.c
index af65491a78e2..6a2ff6b77f2a 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_dump.c
+++ b/drivers/gpu/drm/etnaviv/etnaviv_dump.c
@@ -15,6 +15,7 @@
  */
 
 #include <linux/devcoredump.h>
+#include "etnaviv_cmdbuf.h"
 #include "etnaviv_dump.h"
 #include "etnaviv_gem.h"
 #include "etnaviv_gpu.h"
diff --git a/drivers/gpu/drm/etnaviv/etnaviv_gem_submit.c b/drivers/gpu/drm/etnaviv/etnaviv_gem_submit.c
index afdd55ddf821..7e1fefef2f2c 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_gem_submit.c
+++ b/drivers/gpu/drm/etnaviv/etnaviv_gem_submit.c
@@ -15,6 +15,7 @@
  */
 
 #include <linux/reservation.h>
+#include "etnaviv_cmdbuf.h"
 #include "etnaviv_drv.h"
 #include "etnaviv_gpu.h"
 #include "etnaviv_gem.h"
@@ -332,7 +333,7 @@ int etnaviv_ioctl_gem_submit(struct drm_device *dev, void *data,
 	bos = drm_malloc_ab(args->nr_bos, sizeof(*bos));
 	relocs = drm_malloc_ab(args->nr_relocs, sizeof(*relocs));
 	stream = drm_malloc_ab(1, args->stream_size);
-	cmdbuf = etnaviv_gpu_cmdbuf_new(gpu, ALIGN(args->stream_size, 8) + 8,
+	cmdbuf = etnaviv_cmdbuf_new(gpu, ALIGN(args->stream_size, 8) + 8,
 					args->nr_bos);
 	if (!bos || !relocs || !stream || !cmdbuf) {
 		ret = -ENOMEM;
@@ -422,7 +423,7 @@ int etnaviv_ioctl_gem_submit(struct drm_device *dev, void *data,
 err_submit_cmds:
 	/* if we still own the cmdbuf */
 	if (cmdbuf)
-		etnaviv_gpu_cmdbuf_free(cmdbuf);
+		etnaviv_cmdbuf_free(cmdbuf);
 	if (stream)
 		drm_free_large(stream);
 	if (bos)
diff --git a/drivers/gpu/drm/etnaviv/etnaviv_gpu.c b/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
index 8bc68aef960c..ee8aacc28944 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
+++ b/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
@@ -18,6 +18,8 @@
 #include <linux/dma-fence.h>
 #include <linux/moduleparam.h>
 #include <linux/of_device.h>
+
+#include "etnaviv_cmdbuf.h"
 #include "etnaviv_dump.h"
 #include "etnaviv_gpu.h"
 #include "etnaviv_gem.h"
@@ -693,7 +695,7 @@ int etnaviv_gpu_init(struct etnaviv_gpu *gpu)
 	}
 
 	/* Create buffer: */
-	gpu->buffer = etnaviv_gpu_cmdbuf_new(gpu, PAGE_SIZE, 0);
+	gpu->buffer = etnaviv_cmdbuf_new(gpu, PAGE_SIZE, 0);
 	if (!gpu->buffer) {
 		ret = -ENOMEM;
 		dev_err(gpu->dev, "could not create command buffer\n");
@@ -728,7 +730,7 @@ int etnaviv_gpu_init(struct etnaviv_gpu *gpu)
 	return 0;
 
 free_buffer:
-	etnaviv_gpu_cmdbuf_free(gpu->buffer);
+	etnaviv_cmdbuf_free(gpu->buffer);
 	gpu->buffer = NULL;
 destroy_iommu:
 	etnaviv_iommu_destroy(gpu->mmu);
@@ -1151,41 +1153,6 @@ static void event_free(struct etnaviv_gpu *gpu, unsigned int event)
  * Cmdstream submission/retirement:
  */
 
-struct etnaviv_cmdbuf *etnaviv_gpu_cmdbuf_new(struct etnaviv_gpu *gpu, u32 size,
-	size_t nr_bos)
-{
-	struct etnaviv_cmdbuf *cmdbuf;
-	size_t sz = size_vstruct(nr_bos, sizeof(cmdbuf->bo_map[0]),
-				 sizeof(*cmdbuf));
-
-	cmdbuf = kzalloc(sz, GFP_KERNEL);
-	if (!cmdbuf)
-		return NULL;
-
-	if (gpu->mmu->version == ETNAVIV_IOMMU_V2)
-		size = ALIGN(size, SZ_4K);
-
-	cmdbuf->vaddr = dma_alloc_wc(gpu->dev, size, &cmdbuf->paddr,
-				     GFP_KERNEL);
-	if (!cmdbuf->vaddr) {
-		kfree(cmdbuf);
-		return NULL;
-	}
-
-	cmdbuf->gpu = gpu;
-	cmdbuf->size = size;
-
-	return cmdbuf;
-}
-
-void etnaviv_gpu_cmdbuf_free(struct etnaviv_cmdbuf *cmdbuf)
-{
-	etnaviv_iommu_put_cmdbuf_va(cmdbuf->gpu, cmdbuf);
-	dma_free_wc(cmdbuf->gpu->dev, cmdbuf->size, cmdbuf->vaddr,
-		    cmdbuf->paddr);
-	kfree(cmdbuf);
-}
-
 static void retire_worker(struct work_struct *work)
 {
 	struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu,
@@ -1211,7 +1178,7 @@ static void retire_worker(struct work_struct *work)
 			etnaviv_gem_mapping_unreference(mapping);
 		}
 
-		etnaviv_gpu_cmdbuf_free(cmdbuf);
+		etnaviv_cmdbuf_free(cmdbuf);
 		/*
 		 * We need to balance the runtime PM count caused by
 		 * each submission.  Upon submission, we increment
@@ -1627,7 +1594,7 @@ static void etnaviv_gpu_unbind(struct device *dev, struct device *master,
 #endif
 
 	if (gpu->buffer) {
-		etnaviv_gpu_cmdbuf_free(gpu->buffer);
+		etnaviv_cmdbuf_free(gpu->buffer);
 		gpu->buffer = NULL;
 	}
 
diff --git a/drivers/gpu/drm/etnaviv/etnaviv_gpu.h b/drivers/gpu/drm/etnaviv/etnaviv_gpu.h
index 8c6b824e9d0a..9c10ffeff77e 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_gpu.h
+++ b/drivers/gpu/drm/etnaviv/etnaviv_gpu.h
@@ -150,29 +150,6 @@ struct etnaviv_gpu {
 	struct work_struct recover_work;
 };
 
-struct etnaviv_cmdbuf {
-	/* device this cmdbuf is allocated for */
-	struct etnaviv_gpu *gpu;
-	/* user context key, must be unique between all active users */
-	struct etnaviv_file_private *ctx;
-	/* cmdbuf properties */
-	void *vaddr;
-	dma_addr_t paddr;
-	u32 size;
-	u32 user_size;
-	/* vram node used if the cmdbuf is mapped through the MMUv2 */
-	struct drm_mm_node vram_node;
-	/* fence after which this buffer is to be disposed */
-	struct dma_fence *fence;
-	/* target exec state */
-	u32 exec_state;
-	/* per GPU in-flight list */
-	struct list_head node;
-	/* BOs attached to this command buffer */
-	unsigned int nr_bos;
-	struct etnaviv_vram_mapping *bo_map[0];
-};
-
 static inline void gpu_write(struct etnaviv_gpu *gpu, u32 reg, u32 data)
 {
 	etnaviv_writel(data, gpu->mmio + reg);
@@ -211,9 +188,9 @@ int etnaviv_gpu_wait_obj_inactive(struct etnaviv_gpu *gpu,
 	struct etnaviv_gem_object *etnaviv_obj, struct timespec *timeout);
 int etnaviv_gpu_submit(struct etnaviv_gpu *gpu,
 	struct etnaviv_gem_submit *submit, struct etnaviv_cmdbuf *cmdbuf);
-struct etnaviv_cmdbuf *etnaviv_gpu_cmdbuf_new(struct etnaviv_gpu *gpu,
+struct etnaviv_cmdbuf *etnaviv_cmdbuf_new(struct etnaviv_gpu *gpu,
 					      u32 size, size_t nr_bos);
-void etnaviv_gpu_cmdbuf_free(struct etnaviv_cmdbuf *cmdbuf);
+void etnaviv_cmdbuf_free(struct etnaviv_cmdbuf *cmdbuf);
 int etnaviv_gpu_pm_get_sync(struct etnaviv_gpu *gpu);
 void etnaviv_gpu_pm_put(struct etnaviv_gpu *gpu);
 int etnaviv_gpu_wait_idle(struct etnaviv_gpu *gpu, unsigned int timeout_ms);
diff --git a/drivers/gpu/drm/etnaviv/etnaviv_iommu_v2.c b/drivers/gpu/drm/etnaviv/etnaviv_iommu_v2.c
index 9b88240d8d7b..4e68e6d75477 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_iommu_v2.c
+++ b/drivers/gpu/drm/etnaviv/etnaviv_iommu_v2.c
@@ -21,6 +21,7 @@
 #include <linux/dma-mapping.h>
 #include <linux/bitops.h>
 
+#include "etnaviv_cmdbuf.h"
 #include "etnaviv_gpu.h"
 #include "etnaviv_mmu.h"
 #include "etnaviv_iommu.h"
diff --git a/drivers/gpu/drm/etnaviv/etnaviv_mmu.c b/drivers/gpu/drm/etnaviv/etnaviv_mmu.c
index d1216516b947..056685bd33b2 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_mmu.c
+++ b/drivers/gpu/drm/etnaviv/etnaviv_mmu.c
@@ -15,6 +15,7 @@
  */
 
 #include "common.xml.h"
+#include "etnaviv_cmdbuf.h"
 #include "etnaviv_drv.h"
 #include "etnaviv_gem.h"
 #include "etnaviv_gpu.h"
-- 
2.11.0

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

[PATCH 3/6] drm/etnaviv: wire up iova handling in new cmdbuf abstraction

[Lucas Stach]

Don't call the IOMMU directly, but go through the new cmdbuf abstraction.

Signed-off-by: Lucas Stach <l.stach@pengutronix.de>
---
 drivers/gpu/drm/etnaviv/etnaviv_buffer.c | 13 +++++++------
 drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.c |  5 +++++
 drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.h |  2 ++
 drivers/gpu/drm/etnaviv/etnaviv_dump.c   |  5 ++---
 drivers/gpu/drm/etnaviv/etnaviv_gpu.c    |  4 ++--
 5 files changed, 18 insertions(+), 11 deletions(-)

diff --git a/drivers/gpu/drm/etnaviv/etnaviv_buffer.c b/drivers/gpu/drm/etnaviv/etnaviv_buffer.c
index 9813b9b33553..ed9588f36bc9 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_buffer.c
+++ b/drivers/gpu/drm/etnaviv/etnaviv_buffer.c
@@ -126,7 +126,7 @@ static void etnaviv_buffer_dump(struct etnaviv_gpu *gpu,
 	u32 *ptr = buf->vaddr + off;
 
 	dev_info(gpu->dev, "virt %p phys 0x%08x free 0x%08x\n",
-			ptr, etnaviv_iommu_get_cmdbuf_va(gpu, buf) + off, size - len * 4 - off);
+			ptr, etnaviv_cmdbuf_get_va(buf) + off, size - len * 4 - off);
 
 	print_hex_dump(KERN_INFO, "cmd ", DUMP_PREFIX_OFFSET, 16, 4,
 			ptr, len * 4, 0);
@@ -159,7 +159,7 @@ static u32 etnaviv_buffer_reserve(struct etnaviv_gpu *gpu,
 	if (buffer->user_size + cmd_dwords * sizeof(u64) > buffer->size)
 		buffer->user_size = 0;
 
-	return etnaviv_iommu_get_cmdbuf_va(gpu, buffer) + buffer->user_size;
+	return etnaviv_cmdbuf_get_va(buffer) + buffer->user_size;
 }
 
 u16 etnaviv_buffer_init(struct etnaviv_gpu *gpu)
@@ -170,7 +170,7 @@ u16 etnaviv_buffer_init(struct etnaviv_gpu *gpu)
 	buffer->user_size = 0;
 
 	CMD_WAIT(buffer);
-	CMD_LINK(buffer, 2, etnaviv_iommu_get_cmdbuf_va(gpu, buffer) +
+	CMD_LINK(buffer, 2, etnaviv_cmdbuf_get_va(buffer) +
 		 buffer->user_size - 4);
 
 	return buffer->user_size / 8;
@@ -262,7 +262,7 @@ void etnaviv_buffer_queue(struct etnaviv_gpu *gpu, unsigned int event,
 	if (drm_debug & DRM_UT_DRIVER)
 		etnaviv_buffer_dump(gpu, buffer, 0, 0x50);
 
-	link_target = etnaviv_iommu_get_cmdbuf_va(gpu, cmdbuf);
+	link_target = etnaviv_cmdbuf_get_va(cmdbuf);
 	link_dwords = cmdbuf->size / 8;
 
 	/*
@@ -356,12 +356,13 @@ void etnaviv_buffer_queue(struct etnaviv_gpu *gpu, unsigned int event,
 	CMD_LOAD_STATE(buffer, VIVS_GL_EVENT, VIVS_GL_EVENT_EVENT_ID(event) |
 		       VIVS_GL_EVENT_FROM_PE);
 	CMD_WAIT(buffer);
-	CMD_LINK(buffer, 2, etnaviv_iommu_get_cmdbuf_va(gpu, buffer) +
+	CMD_LINK(buffer, 2, etnaviv_cmdbuf_get_va(buffer) +
 			    buffer->user_size - 4);
 
 	if (drm_debug & DRM_UT_DRIVER)
 		pr_info("stream link to 0x%08x @ 0x%08x %p\n",
-			return_target, etnaviv_iommu_get_cmdbuf_va(gpu, cmdbuf), cmdbuf->vaddr);
+			return_target, etnaviv_cmdbuf_get_va(cmdbuf),
+			cmdbuf->vaddr);
 
 	if (drm_debug & DRM_UT_DRIVER) {
 		print_hex_dump(KERN_INFO, "cmd ", DUMP_PREFIX_OFFSET, 16, 4,
diff --git a/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.c b/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.c
index 08f01ce63ff6..155b1fddca8e 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.c
+++ b/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.c
@@ -52,3 +52,8 @@ void etnaviv_cmdbuf_free(struct etnaviv_cmdbuf *cmdbuf)
 		    cmdbuf->paddr);
 	kfree(cmdbuf);
 }
+
+u32 etnaviv_cmdbuf_get_va(struct etnaviv_cmdbuf *buf)
+{
+	return etnaviv_iommu_get_cmdbuf_va(buf->gpu, buf);
+}
diff --git a/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.h b/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.h
index 883f03960f95..6d64685cd3e7 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.h
+++ b/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.h
@@ -43,4 +43,6 @@ struct etnaviv_cmdbuf {
 	struct etnaviv_vram_mapping *bo_map[0];
 };
 
+u32 etnaviv_cmdbuf_get_va(struct etnaviv_cmdbuf *buf);
+
 #endif /* __ETNAVIV_CMDBUF_H__ */
diff --git a/drivers/gpu/drm/etnaviv/etnaviv_dump.c b/drivers/gpu/drm/etnaviv/etnaviv_dump.c
index 6a2ff6b77f2a..d019b5e311cc 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_dump.c
+++ b/drivers/gpu/drm/etnaviv/etnaviv_dump.c
@@ -178,12 +178,11 @@ void etnaviv_core_dump(struct etnaviv_gpu *gpu)
 	etnaviv_core_dump_mmu(&iter, gpu, mmu_size);
 	etnaviv_core_dump_mem(&iter, ETDUMP_BUF_RING, gpu->buffer->vaddr,
 			      gpu->buffer->size,
-			      etnaviv_iommu_get_cmdbuf_va(gpu, gpu->buffer));
+			      etnaviv_cmdbuf_get_va(gpu->buffer));
 
 	list_for_each_entry(cmd, &gpu->active_cmd_list, node)
 		etnaviv_core_dump_mem(&iter, ETDUMP_BUF_CMD, cmd->vaddr,
-				      cmd->size,
-				      etnaviv_iommu_get_cmdbuf_va(gpu, cmd));
+				      cmd->size, etnaviv_cmdbuf_get_va(cmd));
 
 	/* Reserve space for the bomap */
 	if (n_bomap_pages) {
diff --git a/drivers/gpu/drm/etnaviv/etnaviv_gpu.c b/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
index ee8aacc28944..3a689c3fbe5b 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
+++ b/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
@@ -629,7 +629,7 @@ static void etnaviv_gpu_hw_init(struct etnaviv_gpu *gpu)
 	prefetch = etnaviv_buffer_init(gpu);
 
 	gpu_write(gpu, VIVS_HI_INTR_ENBL, ~0U);
-	etnaviv_gpu_start_fe(gpu, etnaviv_iommu_get_cmdbuf_va(gpu, gpu->buffer),
+	etnaviv_gpu_start_fe(gpu, etnaviv_cmdbuf_get_va(gpu->buffer),
 			     prefetch);
 }
 
@@ -703,7 +703,7 @@ int etnaviv_gpu_init(struct etnaviv_gpu *gpu)
 	}
 
 	if (gpu->mmu->version == ETNAVIV_IOMMU_V1 &&
-	    gpu->buffer->paddr - gpu->memory_base > 0x80000000) {
+	    etnaviv_cmdbuf_get_va(gpu->buffer) > 0x80000000) {
 		ret = -EINVAL;
 		dev_err(gpu->dev,
 			"command buffer outside valid memory window\n");
-- 
2.11.0

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

[PATCH 4/6] drm/etnaviv: get cmdbuf physical address through the cmdbuf abstraction

[Lucas Stach]

Don't allow IOMMUv2 to peek directly into the cmdbuf, but get the
needed PA through a dedicated function.

Signed-off-by: Lucas Stach <l.stach@pengutronix.de>
---
 drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.c   | 5 +++++
 drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.h   | 1 +
 drivers/gpu/drm/etnaviv/etnaviv_drv.c      | 3 ++-
 drivers/gpu/drm/etnaviv/etnaviv_iommu_v2.c | 3 ++-
 4 files changed, 10 insertions(+), 2 deletions(-)

diff --git a/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.c b/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.c
index 155b1fddca8e..1ad118c6c64e 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.c
+++ b/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.c
@@ -57,3 +57,8 @@ u32 etnaviv_cmdbuf_get_va(struct etnaviv_cmdbuf *buf)
 {
 	return etnaviv_iommu_get_cmdbuf_va(buf->gpu, buf);
 }
+
+dma_addr_t etnaviv_cmdbuf_get_pa(struct etnaviv_cmdbuf *buf)
+{
+	return buf->paddr;
+}
diff --git a/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.h b/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.h
index 6d64685cd3e7..244358778407 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.h
+++ b/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.h
@@ -44,5 +44,6 @@ struct etnaviv_cmdbuf {
 };
 
 u32 etnaviv_cmdbuf_get_va(struct etnaviv_cmdbuf *buf);
+dma_addr_t etnaviv_cmdbuf_get_pa(struct etnaviv_cmdbuf *buf);
 
 #endif /* __ETNAVIV_CMDBUF_H__ */
diff --git a/drivers/gpu/drm/etnaviv/etnaviv_drv.c b/drivers/gpu/drm/etnaviv/etnaviv_drv.c
index 6b3ca97833e3..5b6decf85664 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_drv.c
+++ b/drivers/gpu/drm/etnaviv/etnaviv_drv.c
@@ -176,7 +176,8 @@ static void etnaviv_buffer_dump(struct etnaviv_gpu *gpu, struct seq_file *m)
 	u32 i;
 
 	seq_printf(m, "virt %p - phys 0x%llx - free 0x%08x\n",
-			buf->vaddr, (u64)buf->paddr, size - buf->user_size);
+			buf->vaddr, (u64)etnaviv_cmdbuf_get_pa(buf),
+			size - buf->user_size);
 
 	for (i = 0; i < size / 4; i++) {
 		if (i && !(i % 4))
diff --git a/drivers/gpu/drm/etnaviv/etnaviv_iommu_v2.c b/drivers/gpu/drm/etnaviv/etnaviv_iommu_v2.c
index 4e68e6d75477..cbe447ac5974 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_iommu_v2.c
+++ b/drivers/gpu/drm/etnaviv/etnaviv_iommu_v2.c
@@ -255,7 +255,8 @@ void etnaviv_iommuv2_restore(struct etnaviv_gpu *gpu)
 	prefetch = etnaviv_buffer_config_mmuv2(gpu,
 				(u32)etnaviv_domain->mtlb_dma,
 				(u32)etnaviv_domain->bad_page_dma);
-	etnaviv_gpu_start_fe(gpu, gpu->buffer->paddr, prefetch);
+	etnaviv_gpu_start_fe(gpu, (u32)etnaviv_cmdbuf_get_pa(gpu->buffer),
+			     prefetch);
 	etnaviv_gpu_wait_idle(gpu, 100);
 
 	gpu_write(gpu, VIVS_MMUv2_CONTROL, VIVS_MMUv2_CONTROL_ENABLE);
-- 
2.11.0

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

[PATCH 5/6] drm/etnaviv: add cmdbuf suballocator

[Lucas Stach]

There are 3 big benefits to suballocating a single big DMA buffer
for command submission:

1. Avoid hammering CMA. The old way of allocating and freeing a DMA
   buffer for each submission was hitting some of the real slow
   pathes in CMA, as this allocator was not designed for a concurrent
   small buffers load.

2. Less TLB flushes on IOMMUv2. If a new command buffer is mapped into
   the GPU address space the MMU TLBs need to be flushed. By having
   one big buffer statically mapped to the GPU, a lot of those flushes
   can be avoided.

3. No funky workarounds for GC3000. The FE TLB flush on GC3000 isn't
   reliable. To work around that we tried to lay out the cmdbufs in
   the GPU address space in a way to avoid this issue. This hasn't
   always worked if the address space is crowded. A single statically
   mapped buffer avoids the erratum completely.

Signed-off-by: Lucas Stach <l.stach@pengutronix.de>
---
 drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.c     | 123 +++++++++++++++++++++++----
 drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.h     |  20 +++--
 drivers/gpu/drm/etnaviv/etnaviv_gem_submit.c |   5 +-
 drivers/gpu/drm/etnaviv/etnaviv_gpu.c        |  14 ++-
 drivers/gpu/drm/etnaviv/etnaviv_gpu.h        |   5 +-
 drivers/gpu/drm/etnaviv/etnaviv_mmu.c        |  45 +++++-----
 drivers/gpu/drm/etnaviv/etnaviv_mmu.h        |  10 ++-
 7 files changed, 165 insertions(+), 57 deletions(-)

diff --git a/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.c b/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.c
index 1ad118c6c64e..633e0f07cbac 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.c
+++ b/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.c
@@ -14,51 +14,140 @@
  * this program.  If not, see <http://www.gnu.org/licenses/>.
  */
 
+#include <drm/drm_mm.h>
+
 #include "etnaviv_cmdbuf.h"
 #include "etnaviv_gpu.h"
 #include "etnaviv_mmu.h"
 
-struct etnaviv_cmdbuf *etnaviv_cmdbuf_new(struct etnaviv_gpu *gpu, u32 size,
-	size_t nr_bos)
+#define SUBALLOC_SIZE		SZ_256K
+#define SUBALLOC_GRANULE	SZ_4K
+#define SUBALLOC_GRANULES	(SUBALLOC_SIZE / SUBALLOC_GRANULE)
+
+struct etnaviv_cmdbuf_suballoc {
+	/* suballocated dma buffer properties */
+	struct etnaviv_gpu *gpu;
+	void *vaddr;
+	dma_addr_t paddr;
+
+	/* GPU mapping */
+	u32 iova;
+	struct drm_mm_node vram_node; /* only used on MMUv2 */
+
+	/* allocation management */
+	struct mutex lock;
+	DECLARE_BITMAP(granule_map, SUBALLOC_GRANULES);
+	int free_space;
+	wait_queue_head_t free_event;
+};
+
+struct etnaviv_cmdbuf_suballoc *
+etnaviv_cmdbuf_suballoc_new(struct etnaviv_gpu * gpu)
+{
+	struct etnaviv_cmdbuf_suballoc *suballoc;
+	int ret;
+
+	suballoc = kzalloc(sizeof(*suballoc), GFP_KERNEL);
+	if (!suballoc)
+		return ERR_PTR(-ENOMEM);
+
+	suballoc->gpu = gpu;
+	mutex_init(&suballoc->lock);
+	init_waitqueue_head(&suballoc->free_event);
+
+	suballoc->vaddr = dma_alloc_wc(gpu->dev, SUBALLOC_SIZE,
+				       &suballoc->paddr, GFP_KERNEL);
+	if (!suballoc->vaddr)
+		goto free_suballoc;
+
+	ret = etnaviv_iommu_get_suballoc_va(gpu, suballoc->paddr,
+					    &suballoc->vram_node, SUBALLOC_SIZE,
+					    &suballoc->iova);
+	if (ret)
+		goto free_dma;
+
+	return suballoc;
+
+free_dma:
+	dma_free_wc(gpu->dev, SUBALLOC_SIZE, suballoc->vaddr, suballoc->paddr);
+free_suballoc:
+	kfree(suballoc);
+
+	return NULL;
+}
+
+void etnaviv_cmdbuf_suballoc_destroy(struct etnaviv_cmdbuf_suballoc *suballoc)
+{
+	etnaviv_iommu_put_suballoc_va(suballoc->gpu, &suballoc->vram_node,
+				      SUBALLOC_SIZE, suballoc->iova);
+	dma_free_wc(suballoc->gpu->dev, SUBALLOC_SIZE, suballoc->vaddr,
+		    suballoc->paddr);
+	kfree(suballoc);
+}
+
+struct etnaviv_cmdbuf *
+etnaviv_cmdbuf_new(struct etnaviv_cmdbuf_suballoc *suballoc, u32 size,
+		   size_t nr_bos)
 {
 	struct etnaviv_cmdbuf *cmdbuf;
 	size_t sz = size_vstruct(nr_bos, sizeof(cmdbuf->bo_map[0]),
 				 sizeof(*cmdbuf));
+	int granule_offs, order, ret;
 
 	cmdbuf = kzalloc(sz, GFP_KERNEL);
 	if (!cmdbuf)
 		return NULL;
 
-	if (gpu->mmu->version == ETNAVIV_IOMMU_V2)
-		size = ALIGN(size, SZ_4K);
+	cmdbuf->suballoc = suballoc;
+	cmdbuf->size = size;
 
-	cmdbuf->vaddr = dma_alloc_wc(gpu->dev, size, &cmdbuf->paddr,
-				     GFP_KERNEL);
-	if (!cmdbuf->vaddr) {
-		kfree(cmdbuf);
-		return NULL;
+	order = order_base_2(ALIGN(size, SUBALLOC_GRANULE) / SUBALLOC_GRANULE);
+retry:
+	mutex_lock(&suballoc->lock);
+	granule_offs = bitmap_find_free_region(suballoc->granule_map,
+					SUBALLOC_GRANULES, order);
+	if (granule_offs < 0) {
+		suballoc->free_space = 0;
+		mutex_unlock(&suballoc->lock);
+		ret = wait_event_interruptible_timeout(suballoc->free_event,
+						       suballoc->free_space,
+						       msecs_to_jiffies(10 * 1000));
+		if (!ret) {
+			dev_err(suballoc->gpu->dev,
+				"Timeout waiting for cmdbuf space\n");
+			return NULL;
+		}
+		goto retry;
 	}
-
-	cmdbuf->gpu = gpu;
-	cmdbuf->size = size;
+	mutex_unlock(&suballoc->lock);
+	cmdbuf->suballoc_offset = granule_offs * SUBALLOC_GRANULE;
+	cmdbuf->vaddr = suballoc->vaddr + cmdbuf->suballoc_offset;
 
 	return cmdbuf;
 }
 
 void etnaviv_cmdbuf_free(struct etnaviv_cmdbuf *cmdbuf)
 {
-	etnaviv_iommu_put_cmdbuf_va(cmdbuf->gpu, cmdbuf);
-	dma_free_wc(cmdbuf->gpu->dev, cmdbuf->size, cmdbuf->vaddr,
-		    cmdbuf->paddr);
+	struct etnaviv_cmdbuf_suballoc *suballoc = cmdbuf->suballoc;
+	int order = order_base_2(ALIGN(cmdbuf->size, SUBALLOC_GRANULE) /
+				 SUBALLOC_GRANULE);
+
+	mutex_lock(&suballoc->lock);
+	bitmap_release_region(suballoc->granule_map,
+			      cmdbuf->suballoc_offset / SUBALLOC_GRANULE,
+			      order);
+	suballoc->free_space = 1;
+	mutex_unlock(&suballoc->lock);
+	wake_up_all(&suballoc->free_event);
 	kfree(cmdbuf);
 }
 
 u32 etnaviv_cmdbuf_get_va(struct etnaviv_cmdbuf *buf)
 {
-	return etnaviv_iommu_get_cmdbuf_va(buf->gpu, buf);
+	return buf->suballoc->iova + buf->suballoc_offset;
 }
 
 dma_addr_t etnaviv_cmdbuf_get_pa(struct etnaviv_cmdbuf *buf)
 {
-	return buf->paddr;
+	return buf->suballoc->paddr + buf->suballoc_offset;
 }
diff --git a/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.h b/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.h
index 244358778407..a9a6d4b521fb 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.h
+++ b/drivers/gpu/drm/etnaviv/etnaviv_cmdbuf.h
@@ -20,18 +20,19 @@
 #include <drm/drm_mm.h>
 #include <linux/types.h>
 
+struct etnaviv_gpu;
+struct etnaviv_cmdbuf_suballoc;
+
 struct etnaviv_cmdbuf {
-	/* device this cmdbuf is allocated for */
-	struct etnaviv_gpu *gpu;
+	/* suballocator this cmdbuf is allocated from */
+	struct etnaviv_cmdbuf_suballoc *suballoc;
 	/* user context key, must be unique between all active users */
 	struct etnaviv_file_private *ctx;
 	/* cmdbuf properties */
+	int suballoc_offset;
 	void *vaddr;
-	dma_addr_t paddr;
 	u32 size;
 	u32 user_size;
-	/* vram node used if the cmdbuf is mapped through the MMUv2 */
-	struct drm_mm_node vram_node;
 	/* fence after which this buffer is to be disposed */
 	struct dma_fence *fence;
 	/* target exec state */
@@ -43,6 +44,15 @@ struct etnaviv_cmdbuf {
 	struct etnaviv_vram_mapping *bo_map[0];
 };
 
+struct etnaviv_cmdbuf_suballoc *
+etnaviv_cmdbuf_suballoc_new(struct etnaviv_gpu * gpu);
+void etnaviv_cmdbuf_suballoc_destroy(struct etnaviv_cmdbuf_suballoc *suballoc);
+
+struct etnaviv_cmdbuf *
+etnaviv_cmdbuf_new(struct etnaviv_cmdbuf_suballoc *suballoc, u32 size,
+		   size_t nr_bos);
+void etnaviv_cmdbuf_free(struct etnaviv_cmdbuf *cmdbuf);
+
 u32 etnaviv_cmdbuf_get_va(struct etnaviv_cmdbuf *buf);
 dma_addr_t etnaviv_cmdbuf_get_pa(struct etnaviv_cmdbuf *buf);
 
diff --git a/drivers/gpu/drm/etnaviv/etnaviv_gem_submit.c b/drivers/gpu/drm/etnaviv/etnaviv_gem_submit.c
index 7e1fefef2f2c..726090d7a6ac 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_gem_submit.c
+++ b/drivers/gpu/drm/etnaviv/etnaviv_gem_submit.c
@@ -333,8 +333,9 @@ int etnaviv_ioctl_gem_submit(struct drm_device *dev, void *data,
 	bos = drm_malloc_ab(args->nr_bos, sizeof(*bos));
 	relocs = drm_malloc_ab(args->nr_relocs, sizeof(*relocs));
 	stream = drm_malloc_ab(1, args->stream_size);
-	cmdbuf = etnaviv_cmdbuf_new(gpu, ALIGN(args->stream_size, 8) + 8,
-					args->nr_bos);
+	cmdbuf = etnaviv_cmdbuf_new(gpu->cmdbuf_suballoc,
+				    ALIGN(args->stream_size, 8) + 8,
+				    args->nr_bos);
 	if (!bos || !relocs || !stream || !cmdbuf) {
 		ret = -ENOMEM;
 		goto err_submit_cmds;
diff --git a/drivers/gpu/drm/etnaviv/etnaviv_gpu.c b/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
index 3a689c3fbe5b..130d7d517a19 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
+++ b/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
@@ -694,8 +694,15 @@ int etnaviv_gpu_init(struct etnaviv_gpu *gpu)
 		goto fail;
 	}
 
+	gpu->cmdbuf_suballoc = etnaviv_cmdbuf_suballoc_new(gpu);
+	if (IS_ERR(gpu->cmdbuf_suballoc)) {
+		dev_err(gpu->dev, "Failed to create cmdbuf suballocator\n");
+		ret = PTR_ERR(gpu->cmdbuf_suballoc);
+		goto fail;
+	}
+
 	/* Create buffer: */
-	gpu->buffer = etnaviv_cmdbuf_new(gpu, PAGE_SIZE, 0);
+	gpu->buffer = etnaviv_cmdbuf_new(gpu->cmdbuf_suballoc, PAGE_SIZE, 0);
 	if (!gpu->buffer) {
 		ret = -ENOMEM;
 		dev_err(gpu->dev, "could not create command buffer\n");
@@ -1598,6 +1605,11 @@ static void etnaviv_gpu_unbind(struct device *dev, struct device *master,
 		gpu->buffer = NULL;
 	}
 
+	if (gpu->cmdbuf_suballoc) {
+		etnaviv_cmdbuf_suballoc_destroy(gpu->cmdbuf_suballoc);
+		gpu->cmdbuf_suballoc = NULL;
+	}
+
 	if (gpu->mmu) {
 		etnaviv_iommu_destroy(gpu->mmu);
 		gpu->mmu = NULL;
diff --git a/drivers/gpu/drm/etnaviv/etnaviv_gpu.h b/drivers/gpu/drm/etnaviv/etnaviv_gpu.h
index 9c10ffeff77e..1c0606ea7d5e 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_gpu.h
+++ b/drivers/gpu/drm/etnaviv/etnaviv_gpu.h
@@ -92,6 +92,7 @@ struct etnaviv_event {
 	struct dma_fence *fence;
 };
 
+struct etnaviv_cmdbuf_suballoc;
 struct etnaviv_cmdbuf;
 
 struct etnaviv_gpu {
@@ -135,6 +136,7 @@ struct etnaviv_gpu {
 	int irq;
 
 	struct etnaviv_iommu *mmu;
+	struct etnaviv_cmdbuf_suballoc *cmdbuf_suballoc;
 
 	/* Power Control: */
 	struct clk *clk_bus;
@@ -188,9 +190,6 @@ int etnaviv_gpu_wait_obj_inactive(struct etnaviv_gpu *gpu,
 	struct etnaviv_gem_object *etnaviv_obj, struct timespec *timeout);
 int etnaviv_gpu_submit(struct etnaviv_gpu *gpu,
 	struct etnaviv_gem_submit *submit, struct etnaviv_cmdbuf *cmdbuf);
-struct etnaviv_cmdbuf *etnaviv_cmdbuf_new(struct etnaviv_gpu *gpu,
-					      u32 size, size_t nr_bos);
-void etnaviv_cmdbuf_free(struct etnaviv_cmdbuf *cmdbuf);
 int etnaviv_gpu_pm_get_sync(struct etnaviv_gpu *gpu);
 void etnaviv_gpu_pm_put(struct etnaviv_gpu *gpu);
 int etnaviv_gpu_wait_idle(struct etnaviv_gpu *gpu, unsigned int timeout_ms);
diff --git a/drivers/gpu/drm/etnaviv/etnaviv_mmu.c b/drivers/gpu/drm/etnaviv/etnaviv_mmu.c
index 056685bd33b2..dcc86d8eeb98 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_mmu.c
+++ b/drivers/gpu/drm/etnaviv/etnaviv_mmu.c
@@ -321,55 +321,50 @@ void etnaviv_iommu_restore(struct etnaviv_gpu *gpu)
 		etnaviv_iommuv2_restore(gpu);
 }
 
-u32 etnaviv_iommu_get_cmdbuf_va(struct etnaviv_gpu *gpu,
-				struct etnaviv_cmdbuf *buf)
+int etnaviv_iommu_get_suballoc_va(struct etnaviv_gpu *gpu, dma_addr_t paddr,
+				  struct drm_mm_node *vram_node, size_t size,
+				  u32 *iova)
 {
 	struct etnaviv_iommu *mmu = gpu->mmu;
 
 	if (mmu->version == ETNAVIV_IOMMU_V1) {
-		return buf->paddr - gpu->memory_base;
+		*iova = paddr - gpu->memory_base;
+		return 0;
 	} else {
 		int ret;
 
-		if (buf->vram_node.allocated)
-			return (u32)buf->vram_node.start;
-
 		mutex_lock(&mmu->lock);
-		ret = etnaviv_iommu_find_iova(mmu, &buf->vram_node,
-					      buf->size + SZ_64K);
+		ret = etnaviv_iommu_find_iova(mmu, vram_node, size);
 		if (ret < 0) {
 			mutex_unlock(&mmu->lock);
-			return 0;
+			return ret;
 		}
-		ret = iommu_map(mmu->domain, buf->vram_node.start, buf->paddr,
-				buf->size, IOMMU_READ);
+		ret = iommu_map(mmu->domain, vram_node->start, paddr, size,
+				IOMMU_READ);
 		if (ret < 0) {
-			drm_mm_remove_node(&buf->vram_node);
+			drm_mm_remove_node(vram_node);
 			mutex_unlock(&mmu->lock);
-			return 0;
+			return ret;
 		}
-		/*
-		 * At least on GC3000 the FE MMU doesn't properly flush old TLB
-		 * entries. Make sure to space the command buffers out in a way
-		 * that the FE MMU prefetch won't load invalid entries.
-		 */
-		mmu->last_iova = buf->vram_node.start + buf->size + SZ_64K;
+		mmu->last_iova = vram_node->start + size;
 		gpu->mmu->need_flush = true;
 		mutex_unlock(&mmu->lock);
 
-		return (u32)buf->vram_node.start;
+		*iova = (u32)vram_node->start;
+		return 0;
 	}
 }
 
-void etnaviv_iommu_put_cmdbuf_va(struct etnaviv_gpu *gpu,
-				 struct etnaviv_cmdbuf *buf)
+void etnaviv_iommu_put_suballoc_va(struct etnaviv_gpu *gpu,
+				   struct drm_mm_node *vram_node, size_t size,
+				   u32 iova)
 {
 	struct etnaviv_iommu *mmu = gpu->mmu;
 
-	if (mmu->version == ETNAVIV_IOMMU_V2 && buf->vram_node.allocated) {
+	if (mmu->version == ETNAVIV_IOMMU_V2) {
 		mutex_lock(&mmu->lock);
-		iommu_unmap(mmu->domain, buf->vram_node.start, buf->size);
-		drm_mm_remove_node(&buf->vram_node);
+		iommu_unmap(mmu->domain,iova, size);
+		drm_mm_remove_node(vram_node);
 		mutex_unlock(&mmu->lock);
 	}
 }
diff --git a/drivers/gpu/drm/etnaviv/etnaviv_mmu.h b/drivers/gpu/drm/etnaviv/etnaviv_mmu.h
index e787e49c9693..54be289e5981 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_mmu.h
+++ b/drivers/gpu/drm/etnaviv/etnaviv_mmu.h
@@ -62,10 +62,12 @@ void etnaviv_iommu_unmap_gem(struct etnaviv_iommu *mmu,
 	struct etnaviv_vram_mapping *mapping);
 void etnaviv_iommu_destroy(struct etnaviv_iommu *iommu);
 
-u32 etnaviv_iommu_get_cmdbuf_va(struct etnaviv_gpu *gpu,
-				struct etnaviv_cmdbuf *buf);
-void etnaviv_iommu_put_cmdbuf_va(struct etnaviv_gpu *gpu,
-				 struct etnaviv_cmdbuf *buf);
+int etnaviv_iommu_get_suballoc_va(struct etnaviv_gpu *gpu, dma_addr_t paddr,
+				  struct drm_mm_node *vram_node, size_t size,
+				  u32 *iova);
+void etnaviv_iommu_put_suballoc_va(struct etnaviv_gpu *gpu,
+				   struct drm_mm_node *vram_node, size_t size,
+				   u32 iova);
 
 size_t etnaviv_iommu_dump_size(struct etnaviv_iommu *iommu);
 void etnaviv_iommu_dump(struct etnaviv_iommu *iommu, void *buf);
-- 
2.11.0

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

[PATCH 6/6] Revert "drm/etnaviv: trick drm_mm into giving out a low IOVA"

[Lucas Stach]

Now that commandstreams are handled through the cmdbuf suballocator
the workaround to make the IOVA games work is not needed anymore.

Signed-off-by: Lucas Stach <l.stach@pengutronix.de>
---
 drivers/gpu/drm/etnaviv/etnaviv_mmu.c | 7 +------
 1 file changed, 1 insertion(+), 6 deletions(-)

diff --git a/drivers/gpu/drm/etnaviv/etnaviv_mmu.c b/drivers/gpu/drm/etnaviv/etnaviv_mmu.c
index dcc86d8eeb98..6a8c28e64866 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_mmu.c
+++ b/drivers/gpu/drm/etnaviv/etnaviv_mmu.c
@@ -117,14 +117,9 @@ static int etnaviv_iommu_find_iova(struct etnaviv_iommu *mmu,
 		struct list_head list;
 		bool found;
 
-		/*
-		 * XXX: The DRM_MM_SEARCH_BELOW is really a hack to trick
-		 * drm_mm into giving out a low IOVA after address space
-		 * rollover. This needs a proper fix.
-		 */
 		ret = drm_mm_insert_node_in_range(&mmu->mm, node,
 			size, 0, mmu->last_iova, ~0UL,
-			mmu->last_iova ? DRM_MM_SEARCH_DEFAULT : DRM_MM_SEARCH_BELOW);
+			DRM_MM_SEARCH_DEFAULT);
 
 		if (ret != -ENOSPC)
 			break;
-- 
2.11.0

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: [PATCH 0/6] Etnaviv cmdbuf suballocator

[Lucas Stach]

Am Mittwoch, den 18.01.2017, 12:25 +0100 schrieb Lucas Stach:
> Hi all,
> 
> the following patches introduce a cmduf suballocator in the Etnaviv
> kernel driver, which has the following benefits:
> 
> 1. Allocating and freeing a CMA buffer for each user command submission
>    is taking a big toll on the CPU, as CMA is not exactly low overhead.
>    By suballocating a single buffer we avoid all this overhead.
> 
> 2. Less TLB flushes on MMUv2. Each time a new buffer gets mapped into
>    the GPU address space on MMUv2 the TLBs need to be flushed. Mapping
>    the suballocated area once allows to skip the TLB flushes (at least
>    as long as userspace re-uses existing buffers).
> 
> 3. No workarounds for GC3000 required anymore. The FE TLB flush on
>    GC3000 doesn't work reliably, which required us to map the cmdbufs
>    into the GPU address space at specific positions, which also isn't
>    guaranteed to work if the address space is already crowded. Having
>    a single static area for the cmdbufs side-steps the erratum completely.
> 
> If I can get reviews and/or enough testing, I would like to include this
> code in kernel 4.11.

If this helps in testing this change, I have pushed my queue out into my
public git repo at:

https://git.pengutronix.de/git/lst/linux drm-etnaviv-next

The branch is based on v4.10-rc1, but is trivial to rebase onto
something more recent.

Regards,
Lucas

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: [PATCH 0/6] Etnaviv cmdbuf suballocator

[Christian Gmeiner]

Hi Lucas,

2017-01-30 12:48 GMT+01:00 Lucas Stach <l.stach@pengutronix.de>:
> Am Mittwoch, den 18.01.2017, 12:25 +0100 schrieb Lucas Stach:
>> Hi all,
>>
>> the following patches introduce a cmduf suballocator in the Etnaviv
>> kernel driver, which has the following benefits:
>>
>> 1. Allocating and freeing a CMA buffer for each user command submission
>>    is taking a big toll on the CPU, as CMA is not exactly low overhead.
>>    By suballocating a single buffer we avoid all this overhead.
>>
>> 2. Less TLB flushes on MMUv2. Each time a new buffer gets mapped into
>>    the GPU address space on MMUv2 the TLBs need to be flushed. Mapping
>>    the suballocated area once allows to skip the TLB flushes (at least
>>    as long as userspace re-uses existing buffers).
>>
>> 3. No workarounds for GC3000 required anymore. The FE TLB flush on
>>    GC3000 doesn't work reliably, which required us to map the cmdbufs
>>    into the GPU address space at specific positions, which also isn't
>>    guaranteed to work if the address space is already crowded. Having
>>    a single static area for the cmdbufs side-steps the erratum completely.
>>
>> If I can get reviews and/or enough testing, I would like to include this
>> code in kernel 4.11.
>
> If this helps in testing this change, I have pushed my queue out into my
> public git repo at:
>
> https://git.pengutronix.de/git/lst/linux drm-etnaviv-next
>
> The branch is based on v4.10-rc1, but is trivial to rebase onto
> something more recent.
>

Will have a look at it later the day.

greets
--
Christian Gmeiner, MSc

https://www.youtube.com/user/AloryOFFICIAL
https://soundcloud.com/christian-gmeiner
_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: [PATCH 0/6] Etnaviv cmdbuf suballocator

[Christian Gmeiner]

2017-01-30 12:52 GMT+01:00 Christian Gmeiner <christian.gmeiner@gmail.com>:
> Hi Lucas,
>
> 2017-01-30 12:48 GMT+01:00 Lucas Stach <l.stach@pengutronix.de>:
>> Am Mittwoch, den 18.01.2017, 12:25 +0100 schrieb Lucas Stach:
>>> Hi all,
>>>
>>> the following patches introduce a cmduf suballocator in the Etnaviv
>>> kernel driver, which has the following benefits:
>>>
>>> 1. Allocating and freeing a CMA buffer for each user command submission
>>>    is taking a big toll on the CPU, as CMA is not exactly low overhead.
>>>    By suballocating a single buffer we avoid all this overhead.
>>>
>>> 2. Less TLB flushes on MMUv2. Each time a new buffer gets mapped into
>>>    the GPU address space on MMUv2 the TLBs need to be flushed. Mapping
>>>    the suballocated area once allows to skip the TLB flushes (at least
>>>    as long as userspace re-uses existing buffers).
>>>
>>> 3. No workarounds for GC3000 required anymore. The FE TLB flush on
>>>    GC3000 doesn't work reliably, which required us to map the cmdbufs
>>>    into the GPU address space at specific positions, which also isn't
>>>    guaranteed to work if the address space is already crowded. Having
>>>    a single static area for the cmdbufs side-steps the erratum completely.
>>>
>>> If I can get reviews and/or enough testing, I would like to include this
>>> code in kernel 4.11.
>>
>> If this helps in testing this change, I have pushed my queue out into my
>> public git repo at:
>>
>> https://git.pengutronix.de/git/lst/linux drm-etnaviv-next
>>
>> The branch is based on v4.10-rc1, but is trivial to rebase onto
>> something more recent.
>>
>
> Will have a look at it later the day.
>

I really like the idea of a suballocator - nice work!

Series is Reviewed-by: Christian Gmeiner <christian.gmeiner@gmail.com>

greets
--
Christian Gmeiner, MSc

https://www.youtube.com/user/AloryOFFICIAL
https://soundcloud.com/christian-gmeiner
_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel