winsys/amdgpu: add a new winsys for the new kernel driver

v2: - lots of changes according to Emil Velikov's comments
    - implemented radeon_winsys::read_registers

v3: - a lot of new work, many of them adapt to libdrm interface changes
Squashed patches:
winsys/amdgpu: implement radeon_winsys context support
winsys/amdgpu: add reference counting for contexts
winsys/amdgpu: add userptr support
winsys/amdgpu: allocate IBs like normal buffers
winsys/amdgpu: add IBs to the buffer list, adapt to interface changes
winsys/amdgpu: don't use KMS handles as reloc hash keys
winsys/amdgpu: sync buffer accesses to different rings
winsys/amdgpu: use dependencies instead of waiting for last fence v2
gallium/radeon: unify buffer_wait and buffer_is_busy in the winsys interface (amdgpu part)
winsys/amdgpu: track fences per ring and be thread-safe
winsys/amdgpu: simplify waiting on a variable in amdgpu_fence_wait
gallium/radeon: allow the winsys to choose the IB size (amdgpu part)
winsys/amdgpu: switch to new amdgpu_cs_query_fence_status interface
winsys/amdgpu: handle fence and dependencies merge
winsys/amdgpu follow libdrm change to move user fence into UMD
winsys/amdgpu: use amdgpu_bo_va_op for va map/unmap v2
winsys/amdgpu: use the new tiling flags
winsys/amdgpu: switch to new GTT_USWC definition
winsys/amdgpu: expose amdgpu_cs_query_reset_state to drivers
winsys/amdgpu: fix valgrind warnings
winsys/amdgpu: don't use VRAM with APUs that don't have much of it
winsys/amdgpu: require LLVM 3.6.1 for VI because of bug fixes there
winsys/amdgpu: remove amdgpu_winsys::num_cpus
winsys/amdgpu: align BO size to page size
winsys/amdgpu: reduce BO cache timeout
winsys/amdgpu: remove useless flushing and waiting in amdgpu_bo_set_tiling
winsys/amdgpu: use amdgpu_device_handle as a unique device ID instead of fd
winsys/amdgpu: use safer access to amdgpu_fence_wait::signalled
winsys/amdgpu: allow maximum IB size of 4 MB
winsys/amdgpu: add ip_instance into amdgpu_fence
gallium/radeon: add RING_COMPUTE instead of RADEON_FLUSH_COMPUTE
winsys/amdgpu: set the ring type at CS initilization
winsys/amdgpu: query the GART page size from the kernel
winsys/amdgpu: correctly wait for shared buffers to become idle
winsys/amdgpu: set the amdgpu_cs_fence structure only once at fence creation
winsys/amdgpu: add a specific error message for cs_submit -> -ENOMEM
winsys/amdgpu: check num_active_ioctls before calling amdgpu_bo_wait_for_idle
winsys/amdgpu: clear user fence BO after allocating it
winsys/amdgpu: fix user fences
winsys/amdgpu: make amdgpu_winsys_create public
winsys/amdgpu: remove thread offloading
winsys/amdgpu: flatten the amdgpu_cs_context structure and simplify more

v4: require libdrm 2.4.63

configure.ac

@@ -70,6 +70,7 @@ AC_SUBST([OPENCL_VERSION])
 dnl Versions for external dependencies
 LIBDRM_REQUIRED=2.4.60
 LIBDRM_RADEON_REQUIRED=2.4.56
+LIBDRM_AMDGPU_REQUIRED=2.4.63
 LIBDRM_INTEL_REQUIRED=2.4.61
 LIBDRM_NVVIEUX_REQUIRED=2.4.33
 LIBDRM_NOUVEAU_REQUIRED=2.4.62
@@ -2105,6 +2106,7 @@ if test -n "$with_gallium_drivers"; then
         xradeonsi)
             HAVE_GALLIUM_RADEONSI=yes
             PKG_CHECK_MODULES([RADEON], [libdrm_radeon >= $LIBDRM_RADEON_REQUIRED])
+            PKG_CHECK_MODULES([AMDGPU], [libdrm_amdgpu >= $LIBDRM_AMDGPU_REQUIRED])
             gallium_require_drm "radeonsi"
             gallium_require_drm_loader
             radeon_llvm_check "radeonsi"
@@ -2357,6 +2359,7 @@ AC_CONFIG_FILES([Makefile
 		src/gallium/winsys/intel/drm/Makefile
 		src/gallium/winsys/nouveau/drm/Makefile
 		src/gallium/winsys/radeon/drm/Makefile
+		src/gallium/winsys/amdgpu/drm/Makefile
 		src/gallium/winsys/svga/drm/Makefile
 		src/gallium/winsys/sw/dri/Makefile
 		src/gallium/winsys/sw/kms-dri/Makefile

src/gallium/Android.mk

@@ -72,6 +72,7 @@ SUBDIRS += drivers/r600
 endif
 ifneq ($(filter radeonsi, $(MESA_GPU_DRIVERS)),)
 SUBDIRS += drivers/radeonsi
+SUBDIRS += winsys/amdgpu/drm
 endif
 endif
 endif

src/gallium/Makefile.am

@@ -58,6 +58,7 @@ endif
 ## radeonsi
 if HAVE_GALLIUM_RADEONSI
 SUBDIRS += drivers/radeonsi
+SUBDIRS += winsys/amdgpu/drm
 endif
 
 ## the radeon winsys - linked in by r300, r600 and radeonsi

src/gallium/auxiliary/target-helpers/inline_drm_helper.h

@@ -42,6 +42,7 @@
 #if GALLIUM_RADEONSI
 #include "radeon/radeon_winsys.h"
 #include "radeon/drm/radeon_drm_public.h"
+#include "amdgpu/drm/amdgpu_public.h"
 #include "radeonsi/si_public.h"
 #endif
 
@@ -228,7 +229,12 @@ pipe_radeonsi_create_screen(int fd)
 {
    struct radeon_winsys *rw;
 
-   rw = radeon_drm_winsys_create(fd, radeonsi_screen_create);
+   /* First, try amdgpu. */
+   rw = amdgpu_winsys_create(fd, radeonsi_screen_create);
+
+   if (!rw)
+      rw = radeon_drm_winsys_create(fd, radeonsi_screen_create);
+
    return rw ? debug_screen_wrap(rw->screen) : NULL;
 }
 #endif

src/gallium/drivers/radeon/radeon_winsys.h

@@ -44,8 +44,7 @@
 
 #define RADEON_FLUSH_ASYNC		(1 << 0)
 #define RADEON_FLUSH_KEEP_TILING_FLAGS	(1 << 1) /* needs DRM 2.12.0 */
-#define RADEON_FLUSH_COMPUTE		(1 << 2)
-#define RADEON_FLUSH_END_OF_FRAME       (1 << 3)
+#define RADEON_FLUSH_END_OF_FRAME       (1 << 2)
 
 /* Tiling flags. */
 enum radeon_bo_layout {
@@ -134,6 +133,9 @@ enum radeon_family {
     CHIP_KABINI,
     CHIP_HAWAII,
     CHIP_MULLINS,
+    CHIP_TONGA,
+    CHIP_ICELAND,
+    CHIP_CARRIZO,
     CHIP_LAST,
 };
 
@@ -148,10 +150,12 @@ enum chip_class {
     CAYMAN,
     SI,
     CIK,
+    VI,
 };
 
 enum ring_type {
     RING_GFX = 0,
+    RING_COMPUTE,
     RING_DMA,
     RING_UVD,
     RING_VCE,
@@ -517,6 +521,11 @@ struct radeon_winsys {
      */
     void (*ctx_destroy)(struct radeon_winsys_ctx *ctx);
 
+    /**
+     * Query a GPU reset status.
+     */
+    enum pipe_reset_status (*ctx_query_reset_status)(struct radeon_winsys_ctx *ctx);
+
     /**
      * Create a command stream.
      *

src/gallium/drivers/radeonsi/Automake.inc

@@ -5,10 +5,12 @@ TARGET_CPPFLAGS += -DGALLIUM_RADEONSI
 TARGET_LIB_DEPS += \
 	$(top_builddir)/src/gallium/drivers/radeonsi/libradeonsi.la \
 	$(RADEON_LIBS) \
-	$(LIBDRM_LIBS)
+	$(LIBDRM_LIBS) \
+	$(AMDGPU_LIBS)
 
 TARGET_RADEON_WINSYS = \
-	$(top_builddir)/src/gallium/winsys/radeon/drm/libradeonwinsys.la
+	$(top_builddir)/src/gallium/winsys/radeon/drm/libradeonwinsys.la \
+	$(top_builddir)/src/gallium/winsys/amdgpu/drm/libamdgpuwinsys.la
 
 TARGET_RADEON_COMMON = \
 	$(top_builddir)/src/gallium/drivers/radeon/libradeon.la

src/gallium/targets/dri-vdpau.dyn

@@ -1,4 +1,5 @@
 {
 	nouveau_drm_screen_create;
 	radeon_drm_winsys_create;
+	amdgpu_winsys_create;
 };

src/gallium/targets/dri/dri.sym

@@ -4,6 +4,7 @@
 		__driDriverGetExtensions*;
 		nouveau_drm_screen_create;
 		radeon_drm_winsys_create;
+		amdgpu_winsys_create;
 	local:
 		*;
 };

src/gallium/targets/pipe-loader/Makefile.am

@@ -155,10 +155,12 @@ nodist_EXTRA_pipe_radeonsi_la_SOURCES = dummy.cpp
 pipe_radeonsi_la_LIBADD = \
 	$(PIPE_LIBS) \
 	$(top_builddir)/src/gallium/winsys/radeon/drm/libradeonwinsys.la \
+	$(top_builddir)/src/gallium/winsys/amdgpu/drm/libamdgpuwinsys.la \
 	$(top_builddir)/src/gallium/drivers/radeon/libradeon.la \
 	$(top_builddir)/src/gallium/drivers/radeonsi/libradeonsi.la \
 	$(LIBDRM_LIBS) \
-	$(RADEON_LIBS)
+	$(RADEON_LIBS) \
+	$(AMDGPU_LIBS)
 
 endif
 

src/gallium/targets/pipe-loader/pipe_radeonsi.c

@@ -2,6 +2,7 @@
 #include "target-helpers/inline_debug_helper.h"
 #include "radeon/drm/radeon_drm_public.h"
 #include "radeon/radeon_winsys.h"
+#include "amdgpu/drm/amdgpu_public.h"
 #include "radeonsi/si_public.h"
 
 static struct pipe_screen *
@@ -9,7 +10,12 @@ create_screen(int fd)
 {
    struct radeon_winsys *rw;
 
-   rw = radeon_drm_winsys_create(fd, radeonsi_screen_create);
+   /* First, try amdgpu. */
+   rw = amdgpu_winsys_create(fd, radeonsi_screen_create);
+
+   if (!rw)
+      rw = radeon_drm_winsys_create(fd, radeonsi_screen_create);
+
    return rw ? debug_screen_wrap(rw->screen) : NULL;
 }
 

src/gallium/targets/vdpau/vdpau.sym

@@ -3,6 +3,7 @@
                vdp_imp_device_create_x11;
                nouveau_drm_screen_create;
                radeon_drm_winsys_create;
+               amdgpu_winsys_create;
        local:
                *;
 };

src/gallium/winsys/amdgpu/drm/Android.mk

@@ -0,0 +1,37 @@
+# Mesa 3-D graphics library
+#
+# Copyright (C) 2011 Chia-I Wu <olvaffe@gmail.com>
+# Copyright (C) 2011 LunarG Inc.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a
+# copy of this software and associated documentation files (the "Software"),
+# to deal in the Software without restriction, including without limitation
+# the rights to use, copy, modify, merge, publish, distribute, sublicense,
+# and/or sell copies of the Software, and to permit persons to whom the
+# Software is furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included
+# in all copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+# DEALINGS IN THE SOFTWARE.
+
+LOCAL_PATH := $(call my-dir)
+
+# get C_SOURCES
+include $(LOCAL_PATH)/Makefile.sources
+
+include $(CLEAR_VARS)
+
+LOCAL_SRC_FILES := $(C_SOURCES)
+
+LOCAL_SHARED_LIBRARIES := libdrm libdrm_amdgpu
+LOCAL_MODULE := libmesa_winsys_amdgpu
+
+include $(GALLIUM_COMMON_MK)
+include $(BUILD_STATIC_LIBRARY)

src/gallium/winsys/amdgpu/drm/Makefile.am

@@ -0,0 +1,12 @@
+include Makefile.sources
+include $(top_srcdir)/src/gallium/Automake.inc
+
+AM_CFLAGS = \
+	$(GALLIUM_WINSYS_CFLAGS) \
+	$(AMDGPU_CFLAGS)
+
+AM_CXXFLAGS = $(AM_CFLAGS)
+
+noinst_LTLIBRARIES = libamdgpuwinsys.la
+
+libamdgpuwinsys_la_SOURCES = $(C_SOURCES)

src/gallium/winsys/amdgpu/drm/Makefile.sources

@@ -0,0 +1,8 @@
+C_SOURCES := \
+	amdgpu_bo.c \
+	amdgpu_bo.h \
+	amdgpu_cs.c \
+	amdgpu_cs.h \
+	amdgpu_public.h \
+	amdgpu_winsys.c \
+	amdgpu_winsys.h

src/gallium/winsys/amdgpu/drm/amdgpu_bo.c

@@ -0,0 +1,778 @@
+/*
+ * Copyright © 2011 Marek Olšák <maraeo@gmail.com>
+ * Copyright © 2015 Advanced Micro Devices, Inc.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining
+ * a copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
+ * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS, AUTHORS
+ * AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ */
+/*
+ * Authors:
+ *      Marek Olšák <maraeo@gmail.com>
+ */
+
+#include "amdgpu_cs.h"
+
+#include "os/os_time.h"
+#include "state_tracker/drm_driver.h"
+#include <amdgpu_drm.h>
+#include <xf86drm.h>
+#include <stdio.h>
+
+static const struct pb_vtbl amdgpu_winsys_bo_vtbl;
+
+static inline struct amdgpu_winsys_bo *amdgpu_winsys_bo(struct pb_buffer *bo)
+{
+   assert(bo->vtbl == &amdgpu_winsys_bo_vtbl);
+   return (struct amdgpu_winsys_bo *)bo;
+}
+
+struct amdgpu_bomgr {
+   struct pb_manager base;
+   struct amdgpu_winsys *rws;
+};
+
+static struct amdgpu_winsys *get_winsys(struct pb_manager *mgr)
+{
+   return ((struct amdgpu_bomgr*)mgr)->rws;
+}
+
+static struct amdgpu_winsys_bo *get_amdgpu_winsys_bo(struct pb_buffer *_buf)
+{
+   struct amdgpu_winsys_bo *bo = NULL;
+
+   if (_buf->vtbl == &amdgpu_winsys_bo_vtbl) {
+      bo = amdgpu_winsys_bo(_buf);
+   } else {
+      struct pb_buffer *base_buf;
+      pb_size offset;
+      pb_get_base_buffer(_buf, &base_buf, &offset);
+
+      if (base_buf->vtbl == &amdgpu_winsys_bo_vtbl)
+         bo = amdgpu_winsys_bo(base_buf);
+   }
+
+   return bo;
+}
+
+static bool amdgpu_bo_wait(struct pb_buffer *_buf, uint64_t timeout,
+                           enum radeon_bo_usage usage)
+{
+   struct amdgpu_winsys_bo *bo = get_amdgpu_winsys_bo(_buf);
+   struct amdgpu_winsys *ws = bo->rws;
+   int i;
+
+   if (bo->is_shared) {
+      /* We can't use user fences for shared buffers, because user fences
+       * are local to this process only. If we want to wait for all buffer
+       * uses in all processes, we have to use amdgpu_bo_wait_for_idle.
+       */
+      bool buffer_busy = true;
+      int r;
+
+      r = amdgpu_bo_wait_for_idle(bo->bo, timeout, &buffer_busy);
+      if (r)
+         fprintf(stderr, "%s: amdgpu_bo_wait_for_idle failed %i\n", __func__,
+                 r);
+      return !buffer_busy;
+   }
+
+   if (timeout == 0) {
+      /* Timeout == 0 is quite simple. */
+      pipe_mutex_lock(ws->bo_fence_lock);
+      for (i = 0; i < RING_LAST; i++)
+         if (bo->fence[i]) {
+            if (amdgpu_fence_wait(bo->fence[i], 0, false)) {
+               /* Release the idle fence to avoid checking it again later. */
+               amdgpu_fence_reference(&bo->fence[i], NULL);
+            } else {
+               pipe_mutex_unlock(ws->bo_fence_lock);
+               return false;
+            }
+         }
+      pipe_mutex_unlock(ws->bo_fence_lock);
+      return true;
+
+   } else {
+      struct pipe_fence_handle *fence[RING_LAST] = {};
+      bool fence_idle[RING_LAST] = {};
+      bool buffer_idle = true;
+      int64_t abs_timeout = os_time_get_absolute_timeout(timeout);
+
+      /* Take references to all fences, so that we can wait for them
+       * without the lock. */
+      pipe_mutex_lock(ws->bo_fence_lock);
+      for (i = 0; i < RING_LAST; i++)
+         amdgpu_fence_reference(&fence[i], bo->fence[i]);
+      pipe_mutex_unlock(ws->bo_fence_lock);
+
+      /* Now wait for the fences. */
+      for (i = 0; i < RING_LAST; i++) {
+         if (fence[i]) {
+            if (amdgpu_fence_wait(fence[i], abs_timeout, true))
+               fence_idle[i] = true;
+            else
+               buffer_idle = false;
+         }
+      }
+
+      /* Release idle fences to avoid checking them again later. */
+      pipe_mutex_lock(ws->bo_fence_lock);
+      for (i = 0; i < RING_LAST; i++) {
+         if (fence[i] == bo->fence[i] && fence_idle[i])
+            amdgpu_fence_reference(&bo->fence[i], NULL);
+
+         amdgpu_fence_reference(&fence[i], NULL);
+      }
+      pipe_mutex_unlock(ws->bo_fence_lock);
+
+      return buffer_idle;
+   }
+}
+
+static enum radeon_bo_domain amdgpu_bo_get_initial_domain(
+      struct radeon_winsys_cs_handle *buf)
+{
+   return ((struct amdgpu_winsys_bo*)buf)->initial_domain;
+}
+
+static void amdgpu_bo_destroy(struct pb_buffer *_buf)
+{
+   struct amdgpu_winsys_bo *bo = amdgpu_winsys_bo(_buf);
+   int i;
+
+   amdgpu_bo_va_op(bo->bo, 0, bo->base.size, bo->va, 0, AMDGPU_VA_OP_UNMAP);
+   amdgpu_va_range_free(bo->va_handle);
+   amdgpu_bo_free(bo->bo);
+
+   for (i = 0; i < RING_LAST; i++)
+      amdgpu_fence_reference(&bo->fence[i], NULL);
+
+   if (bo->initial_domain & RADEON_DOMAIN_VRAM)
+      bo->rws->allocated_vram -= align(bo->base.size, bo->rws->gart_page_size);
+   else if (bo->initial_domain & RADEON_DOMAIN_GTT)
+      bo->rws->allocated_gtt -= align(bo->base.size, bo->rws->gart_page_size);
+   FREE(bo);
+}
+
+static void *amdgpu_bo_map(struct radeon_winsys_cs_handle *buf,
+                           struct radeon_winsys_cs *rcs,
+                           enum pipe_transfer_usage usage)
+{
+   struct amdgpu_winsys_bo *bo = (struct amdgpu_winsys_bo*)buf;
+   struct amdgpu_cs *cs = (struct amdgpu_cs*)rcs;
+   int r;
+   void *cpu = NULL;
+
+   /* If it's not unsynchronized bo_map, flush CS if needed and then wait. */
+   if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
+      /* DONTBLOCK doesn't make sense with UNSYNCHRONIZED. */
+      if (usage & PIPE_TRANSFER_DONTBLOCK) {
+         if (!(usage & PIPE_TRANSFER_WRITE)) {
+            /* Mapping for read.
+             *
+             * Since we are mapping for read, we don't need to wait
+             * if the GPU is using the buffer for read too
+             * (neither one is changing it).
+             *
+             * Only check whether the buffer is being used for write. */
+            if (cs && amdgpu_bo_is_referenced_by_cs_with_usage(cs, bo,
+                                                               RADEON_USAGE_WRITE)) {
+               cs->flush_cs(cs->flush_data, RADEON_FLUSH_ASYNC, NULL);
+               return NULL;
+            }
+
+            if (!amdgpu_bo_wait((struct pb_buffer*)bo, 0,
+                                RADEON_USAGE_WRITE)) {
+               return NULL;
+            }
+         } else {
+            if (cs && amdgpu_bo_is_referenced_by_cs(cs, bo)) {
+               cs->flush_cs(cs->flush_data, RADEON_FLUSH_ASYNC, NULL);
+               return NULL;
+            }
+
+            if (!amdgpu_bo_wait((struct pb_buffer*)bo, 0,
+                                RADEON_USAGE_READWRITE)) {
+               return NULL;
+            }
+         }
+      } else {
+         uint64_t time = os_time_get_nano();
+
+         if (!(usage & PIPE_TRANSFER_WRITE)) {
+            /* Mapping for read.
+             *
+             * Since we are mapping for read, we don't need to wait
+             * if the GPU is using the buffer for read too
+             * (neither one is changing it).
+             *
+             * Only check whether the buffer is being used for write. */
+            if (cs && amdgpu_bo_is_referenced_by_cs_with_usage(cs, bo,
+                                                               RADEON_USAGE_WRITE)) {
+               cs->flush_cs(cs->flush_data, 0, NULL);
+            }
+            amdgpu_bo_wait((struct pb_buffer*)bo, PIPE_TIMEOUT_INFINITE,
+                           RADEON_USAGE_WRITE);
+         } else {
+            /* Mapping for write. */
+            if (cs && amdgpu_bo_is_referenced_by_cs(cs, bo))
+               cs->flush_cs(cs->flush_data, 0, NULL);
+
+            amdgpu_bo_wait((struct pb_buffer*)bo, PIPE_TIMEOUT_INFINITE,
+                           RADEON_USAGE_READWRITE);
+         }
+
+         bo->rws->buffer_wait_time += os_time_get_nano() - time;
+      }
+   }
+
+   /* If the buffer is created from user memory, return the user pointer. */
+   if (bo->user_ptr)
+       return bo->user_ptr;
+
+   r = amdgpu_bo_cpu_map(bo->bo, &cpu);
+   return r ? NULL : cpu;
+}
+
+static void amdgpu_bo_unmap(struct radeon_winsys_cs_handle *buf)
+{
+   struct amdgpu_winsys_bo *bo = (struct amdgpu_winsys_bo*)buf;
+
+   amdgpu_bo_cpu_unmap(bo->bo);
+}
+
+static void amdgpu_bo_get_base_buffer(struct pb_buffer *buf,
+                                      struct pb_buffer **base_buf,
+                                      unsigned *offset)
+{
+   *base_buf = buf;
+   *offset = 0;
+}
+
+static enum pipe_error amdgpu_bo_validate(struct pb_buffer *_buf,
+                                          struct pb_validate *vl,
+                                          unsigned flags)
+{
+   /* Always pinned */
+   return PIPE_OK;
+}
+
+static void amdgpu_bo_fence(struct pb_buffer *buf,
+                            struct pipe_fence_handle *fence)
+{
+}
+
+static const struct pb_vtbl amdgpu_winsys_bo_vtbl = {
+   amdgpu_bo_destroy,
+   NULL, /* never called */
+   NULL, /* never called */
+   amdgpu_bo_validate,
+   amdgpu_bo_fence,
+   amdgpu_bo_get_base_buffer,
+};
+
+static struct pb_buffer *amdgpu_bomgr_create_bo(struct pb_manager *_mgr,
+                                                pb_size size,
+                                                const struct pb_desc *desc)
+{
+   struct amdgpu_winsys *rws = get_winsys(_mgr);
+   struct amdgpu_bo_desc *rdesc = (struct amdgpu_bo_desc*)desc;
+   struct amdgpu_bo_alloc_request request = {0};
+   amdgpu_bo_handle buf_handle;
+   uint64_t va = 0;
+   struct amdgpu_winsys_bo *bo;
+   amdgpu_va_handle va_handle;
+   int r;
+
+   assert(rdesc->initial_domain & RADEON_DOMAIN_VRAM_GTT);
+   bo = CALLOC_STRUCT(amdgpu_winsys_bo);
+   if (!bo) {
+      return NULL;
+   }
+
+   request.alloc_size = size;
+   request.phys_alignment = desc->alignment;
+
+   if (rdesc->initial_domain & RADEON_DOMAIN_VRAM) {
+      request.preferred_heap |= AMDGPU_GEM_DOMAIN_VRAM;
+      if (rdesc->flags & RADEON_FLAG_CPU_ACCESS)
+         request.flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
+   }
+   if (rdesc->initial_domain & RADEON_DOMAIN_GTT) {
+      request.preferred_heap |= AMDGPU_GEM_DOMAIN_GTT;
+      if (rdesc->flags & RADEON_FLAG_GTT_WC)
+         request.flags |= AMDGPU_GEM_CREATE_CPU_GTT_USWC;
+   }
+
+   r = amdgpu_bo_alloc(rws->dev, &request, &buf_handle);
+   if (r) {
+      fprintf(stderr, "amdgpu: Failed to allocate a buffer:\n");
+      fprintf(stderr, "amdgpu:    size      : %d bytes\n", size);
+      fprintf(stderr, "amdgpu:    alignment : %d bytes\n", desc->alignment);
+      fprintf(stderr, "amdgpu:    domains   : %d\n", rdesc->initial_domain);
+      goto error_bo_alloc;
+   }
+
+   r = amdgpu_va_range_alloc(rws->dev, amdgpu_gpu_va_range_general,
+                             size, desc->alignment, 0, &va, &va_handle, 0);
+   if (r)
+      goto error_va_alloc;
+
+   r = amdgpu_bo_va_op(buf_handle, 0, size, va, 0, AMDGPU_VA_OP_MAP);
+   if (r)
+      goto error_va_map;
+
+   pipe_reference_init(&bo->base.reference, 1);
+   bo->base.alignment = desc->alignment;
+   bo->base.usage = desc->usage;
+   bo->base.size = size;
+   bo->base.vtbl = &amdgpu_winsys_bo_vtbl;
+   bo->rws = rws;
+   bo->bo = buf_handle;
+   bo->va = va;
+   bo->va_handle = va_handle;
+   bo->initial_domain = rdesc->initial_domain;
+   bo->unique_id = __sync_fetch_and_add(&rws->next_bo_unique_id, 1);
+
+   if (rdesc->initial_domain & RADEON_DOMAIN_VRAM)
+      rws->allocated_vram += align(size, rws->gart_page_size);
+   else if (rdesc->initial_domain & RADEON_DOMAIN_GTT)
+      rws->allocated_gtt += align(size, rws->gart_page_size);
+
+   return &bo->base;
+
+error_va_map:
+   amdgpu_va_range_free(va_handle);
+
+error_va_alloc:
+   amdgpu_bo_free(buf_handle);
+
+error_bo_alloc:
+   FREE(bo);
+   return NULL;
+}
+
+static void amdgpu_bomgr_flush(struct pb_manager *mgr)
+{
+   /* NOP */
+}
+
+/* This is for the cache bufmgr. */
+static boolean amdgpu_bomgr_is_buffer_busy(struct pb_manager *_mgr,
+                                           struct pb_buffer *_buf)
+{
+   struct amdgpu_winsys_bo *bo = amdgpu_winsys_bo(_buf);
+
+   if (amdgpu_bo_is_referenced_by_any_cs(bo)) {
+      return TRUE;
+   }
+
+   if (!amdgpu_bo_wait((struct pb_buffer*)bo, 0, RADEON_USAGE_READWRITE)) {
+      return TRUE;
+   }
+
+   return FALSE;
+}
+
+static void amdgpu_bomgr_destroy(struct pb_manager *mgr)
+{
+   FREE(mgr);
+}
+
+struct pb_manager *amdgpu_bomgr_create(struct amdgpu_winsys *rws)
+{
+   struct amdgpu_bomgr *mgr;
+
+   mgr = CALLOC_STRUCT(amdgpu_bomgr);
+   if (!mgr)
+      return NULL;
+
+   mgr->base.destroy = amdgpu_bomgr_destroy;
+   mgr->base.create_buffer = amdgpu_bomgr_create_bo;
+   mgr->base.flush = amdgpu_bomgr_flush;
+   mgr->base.is_buffer_busy = amdgpu_bomgr_is_buffer_busy;
+
+   mgr->rws = rws;
+   return &mgr->base;
+}
+
+static unsigned eg_tile_split(unsigned tile_split)
+{
+   switch (tile_split) {
+   case 0:     tile_split = 64;    break;
+   case 1:     tile_split = 128;   break;
+   case 2:     tile_split = 256;   break;
+   case 3:     tile_split = 512;   break;
+   default:
+   case 4:     tile_split = 1024;  break;
+   case 5:     tile_split = 2048;  break;
+   case 6:     tile_split = 4096;  break;
+   }
+   return tile_split;
+}
+
+static unsigned eg_tile_split_rev(unsigned eg_tile_split)
+{
+   switch (eg_tile_split) {
+   case 64:    return 0;
+   case 128:   return 1;
+   case 256:   return 2;
+   case 512:   return 3;
+   default:
+   case 1024:  return 4;
+   case 2048:  return 5;
+   case 4096:  return 6;
+   }
+}
+
+static void amdgpu_bo_get_tiling(struct pb_buffer *_buf,
+                                 enum radeon_bo_layout *microtiled,
+                                 enum radeon_bo_layout *macrotiled,
+                                 unsigned *bankw, unsigned *bankh,
+                                 unsigned *tile_split,
+                                 unsigned *stencil_tile_split,
+                                 unsigned *mtilea,
+                                 bool *scanout)
+{
+   struct amdgpu_winsys_bo *bo = get_amdgpu_winsys_bo(_buf);
+   struct amdgpu_bo_info info = {0};
+   uint32_t tiling_flags;
+   int r;
+
+   r = amdgpu_bo_query_info(bo->bo, &info);
+   if (r)
+      return;
+
+   tiling_flags = info.metadata.tiling_info;
+
+   *microtiled = RADEON_LAYOUT_LINEAR;
+   *macrotiled = RADEON_LAYOUT_LINEAR;
+
+   if (AMDGPU_TILING_GET(tiling_flags, ARRAY_MODE) == 4)  /* 2D_TILED_THIN1 */
+      *macrotiled = RADEON_LAYOUT_TILED;
+   else if (AMDGPU_TILING_GET(tiling_flags, ARRAY_MODE) == 2) /* 1D_TILED_THIN1 */
+      *microtiled = RADEON_LAYOUT_TILED;
+
+   if (bankw && tile_split && mtilea && tile_split) {
+      *bankw = 1 << AMDGPU_TILING_GET(tiling_flags, BANK_WIDTH);
+      *bankh = 1 << AMDGPU_TILING_GET(tiling_flags, BANK_HEIGHT);
+      *tile_split = eg_tile_split(AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT));
+      *mtilea = 1 << AMDGPU_TILING_GET(tiling_flags, MACRO_TILE_ASPECT);
+   }
+   if (scanout)
+      *scanout = AMDGPU_TILING_GET(tiling_flags, MICRO_TILE_MODE) == 0; /* DISPLAY */
+}
+
+static void amdgpu_bo_set_tiling(struct pb_buffer *_buf,
+                                 struct radeon_winsys_cs *rcs,
+                                 enum radeon_bo_layout microtiled,
+                                 enum radeon_bo_layout macrotiled,
+                                 unsigned bankw, unsigned bankh,
+                                 unsigned tile_split,
+                                 unsigned stencil_tile_split,
+                                 unsigned mtilea,
+                                 uint32_t pitch,
+                                 bool scanout)
+{
+   struct amdgpu_winsys_bo *bo = get_amdgpu_winsys_bo(_buf);
+   struct amdgpu_bo_metadata metadata = {0};
+   uint32_t tiling_flags = 0;
+
+   if (macrotiled == RADEON_LAYOUT_TILED)
+      tiling_flags |= AMDGPU_TILING_SET(ARRAY_MODE, 4); /* 2D_TILED_THIN1 */
+   else if (microtiled == RADEON_LAYOUT_TILED)
+      tiling_flags |= AMDGPU_TILING_SET(ARRAY_MODE, 2); /* 1D_TILED_THIN1 */
+   else
+      tiling_flags |= AMDGPU_TILING_SET(ARRAY_MODE, 1); /* LINEAR_ALIGNED */
+
+   tiling_flags |= AMDGPU_TILING_SET(BANK_WIDTH, util_logbase2(bankw));
+   tiling_flags |= AMDGPU_TILING_SET(BANK_HEIGHT, util_logbase2(bankh));
+   if (tile_split)
+      tiling_flags |= AMDGPU_TILING_SET(TILE_SPLIT, eg_tile_split_rev(tile_split));
+   tiling_flags |= AMDGPU_TILING_SET(MACRO_TILE_ASPECT, util_logbase2(mtilea));
+
+   if (scanout)
+      tiling_flags |= AMDGPU_TILING_SET(MICRO_TILE_MODE, 0); /* DISPLAY_MICRO_TILING */
+   else
+      tiling_flags |= AMDGPU_TILING_SET(MICRO_TILE_MODE, 1); /* THIN_MICRO_TILING */
+
+   metadata.tiling_info = tiling_flags;
+
+   amdgpu_bo_set_metadata(bo->bo, &metadata);
+}
+
+static struct radeon_winsys_cs_handle *amdgpu_get_cs_handle(struct pb_buffer *_buf)
+{
+   /* return a direct pointer to amdgpu_winsys_bo. */
+   return (struct radeon_winsys_cs_handle*)get_amdgpu_winsys_bo(_buf);
+}
+
+static struct pb_buffer *
+amdgpu_bo_create(struct radeon_winsys *rws,
+                 unsigned size,
+                 unsigned alignment,
+                 boolean use_reusable_pool,
+                 enum radeon_bo_domain domain,
+                 enum radeon_bo_flag flags)
+{
+   struct amdgpu_winsys *ws = amdgpu_winsys(rws);
+   struct amdgpu_bo_desc desc;
+   struct pb_manager *provider;
+   struct pb_buffer *buffer;
+
+   /* Don't use VRAM if the GPU doesn't have much. This is only the initial
+    * domain. The kernel is free to move the buffer if it wants to.
+    *
+    * 64MB means no VRAM by todays standards.
+    */
+   if (domain & RADEON_DOMAIN_VRAM && ws->info.vram_size <= 64*1024*1024) {
+      domain = RADEON_DOMAIN_GTT;
+      flags = RADEON_FLAG_GTT_WC;
+   }
+
+   memset(&desc, 0, sizeof(desc));
+   desc.base.alignment = alignment;
+
+   /* Align size to page size. This is the minimum alignment for normal
+    * BOs. Aligning this here helps the cached bufmgr. Especially small BOs,
+    * like constant/uniform buffers, can benefit from better and more reuse.
+    */
+   size = align(size, ws->gart_page_size);
+
+   /* Only set one usage bit each for domains and flags, or the cache manager
+    * might consider different sets of domains / flags compatible
+    */
+   if (domain == RADEON_DOMAIN_VRAM_GTT)
+      desc.base.usage = 1 << 2;
+   else
+      desc.base.usage = domain >> 1;
+   assert(flags < sizeof(desc.base.usage) * 8 - 3);
+   desc.base.usage |= 1 << (flags + 3);
+
+   desc.initial_domain = domain;
+   desc.flags = flags;
+
+   /* Assign a buffer manager. */
+   if (use_reusable_pool)
+      provider = ws->cman;
+   else
+      provider = ws->kman;
+
+   buffer = provider->create_buffer(provider, size, &desc.base);
+   if (!buffer)
+      return NULL;
+
+   return (struct pb_buffer*)buffer;
+}
+
+static struct pb_buffer *amdgpu_bo_from_handle(struct radeon_winsys *rws,
+                                               struct winsys_handle *whandle,
+                                               unsigned *stride)
+{
+   struct amdgpu_winsys *ws = amdgpu_winsys(rws);
+   struct amdgpu_winsys_bo *bo;
+   enum amdgpu_bo_handle_type type;
+   struct amdgpu_bo_import_result result = {0};
+   uint64_t va;
+   amdgpu_va_handle va_handle;
+   struct amdgpu_bo_info info = {0};
+   enum radeon_bo_domain initial = 0;
+   int r;
+
+   /* Initialize the structure. */
+   bo = CALLOC_STRUCT(amdgpu_winsys_bo);
+   if (!bo) {
+      return NULL;
+   }
+
+   switch (whandle->type) {
+   case DRM_API_HANDLE_TYPE_SHARED:
+      type = amdgpu_bo_handle_type_gem_flink_name;
+      break;
+   case DRM_API_HANDLE_TYPE_FD:
+      type = amdgpu_bo_handle_type_dma_buf_fd;
+      break;
+   default:
+      return NULL;
+   }
+
+   r = amdgpu_bo_import(ws->dev, type, whandle->handle, &result);
+   if (r)
+      goto error;
+
+   /* Get initial domains. */
+   r = amdgpu_bo_query_info(result.buf_handle, &info);
+   if (r)
+      goto error_query;
+
+   r = amdgpu_va_range_alloc(ws->dev, amdgpu_gpu_va_range_general,
+                             result.alloc_size, 1 << 20, 0, &va, &va_handle, 0);
+   if (r)
+      goto error_query;
+
+   r = amdgpu_bo_va_op(result.buf_handle, 0, result.alloc_size, va, 0, AMDGPU_VA_OP_MAP);
+   if (r)
+      goto error_va_map;
+
+   if (info.preferred_heap & AMDGPU_GEM_DOMAIN_VRAM)
+      initial |= RADEON_DOMAIN_VRAM;
+   if (info.preferred_heap & AMDGPU_GEM_DOMAIN_GTT)
+      initial |= RADEON_DOMAIN_GTT;
+
+
+   pipe_reference_init(&bo->base.reference, 1);
+   bo->base.alignment = info.phys_alignment;
+   bo->base.usage = PB_USAGE_GPU_WRITE | PB_USAGE_GPU_READ;
+   bo->bo = result.buf_handle;
+   bo->base.size = result.alloc_size;
+   bo->base.vtbl = &amdgpu_winsys_bo_vtbl;
+   bo->rws = ws;
+   bo->va = va;
+   bo->va_handle = va_handle;
+   bo->initial_domain = initial;
+   bo->unique_id = __sync_fetch_and_add(&ws->next_bo_unique_id, 1);
+   bo->is_shared = true;
+
+   if (stride)
+      *stride = whandle->stride;
+
+   if (bo->initial_domain & RADEON_DOMAIN_VRAM)
+      ws->allocated_vram += align(bo->base.size, ws->gart_page_size);
+   else if (bo->initial_domain & RADEON_DOMAIN_GTT)
+      ws->allocated_gtt += align(bo->base.size, ws->gart_page_size);
+
+   return &bo->base;
+
+error_va_map:
+   amdgpu_va_range_free(va_handle);
+
+error_query:
+   amdgpu_bo_free(result.buf_handle);
+
+error:
+   FREE(bo);
+   return NULL;
+}
+
+static boolean amdgpu_bo_get_handle(struct pb_buffer *buffer,
+                                    unsigned stride,
+                                    struct winsys_handle *whandle)
+{
+   struct amdgpu_winsys_bo *bo = get_amdgpu_winsys_bo(buffer);
+   enum amdgpu_bo_handle_type type;
+   int r;
+
+   switch (whandle->type) {
+   case DRM_API_HANDLE_TYPE_SHARED:
+      type = amdgpu_bo_handle_type_gem_flink_name;
+      break;
+   case DRM_API_HANDLE_TYPE_FD:
+      type = amdgpu_bo_handle_type_dma_buf_fd;
+      break;
+   case DRM_API_HANDLE_TYPE_KMS:
+      type = amdgpu_bo_handle_type_kms;
+      break;
+   default:
+      return FALSE;
+   }
+
+   r = amdgpu_bo_export(bo->bo, type, &whandle->handle);
+   if (r)
+      return FALSE;
+
+   whandle->stride = stride;
+   bo->is_shared = true;
+   return TRUE;
+}
+
+static struct pb_buffer *amdgpu_bo_from_ptr(struct radeon_winsys *rws,
+					    void *pointer, unsigned size)
+{
+    struct amdgpu_winsys *ws = amdgpu_winsys(rws);
+    amdgpu_bo_handle buf_handle;
+    struct amdgpu_winsys_bo *bo;
+    uint64_t va;
+    amdgpu_va_handle va_handle;
+
+    bo = CALLOC_STRUCT(amdgpu_winsys_bo);
+    if (!bo)
+        return NULL;
+
+    if (amdgpu_create_bo_from_user_mem(ws->dev, pointer, size, &buf_handle))
+        goto error;
+
+    if (amdgpu_va_range_alloc(ws->dev, amdgpu_gpu_va_range_general,
+                              size, 1 << 12, 0, &va, &va_handle, 0))
+        goto error_va_alloc;
+
+    if (amdgpu_bo_va_op(buf_handle, 0, size, va, 0, AMDGPU_VA_OP_MAP))
+        goto error_va_map;
+
+    /* Initialize it. */
+    pipe_reference_init(&bo->base.reference, 1);
+    bo->bo = buf_handle;
+    bo->base.alignment = 0;
+    bo->base.usage = PB_USAGE_GPU_WRITE | PB_USAGE_GPU_READ;
+    bo->base.size = size;
+    bo->base.vtbl = &amdgpu_winsys_bo_vtbl;
+    bo->rws = ws;
+    bo->user_ptr = pointer;
+    bo->va = va;
+    bo->va_handle = va_handle;
+    bo->initial_domain = RADEON_DOMAIN_GTT;
+    bo->unique_id = __sync_fetch_and_add(&ws->next_bo_unique_id, 1);
+
+    ws->allocated_gtt += align(bo->base.size, ws->gart_page_size);
+
+    return (struct pb_buffer*)bo;
+
+error_va_map:
+    amdgpu_va_range_free(va_handle);
+
+error_va_alloc:
+    amdgpu_bo_free(buf_handle);
+
+error:
+    FREE(bo);
+    return NULL;
+}
+
+static uint64_t amdgpu_bo_get_va(struct radeon_winsys_cs_handle *buf)
+{
+   return ((struct amdgpu_winsys_bo*)buf)->va;
+}
+
+void amdgpu_bomgr_init_functions(struct amdgpu_winsys *ws)
+{
+   ws->base.buffer_get_cs_handle = amdgpu_get_cs_handle;
+   ws->base.buffer_set_tiling = amdgpu_bo_set_tiling;
+   ws->base.buffer_get_tiling = amdgpu_bo_get_tiling;
+   ws->base.buffer_map = amdgpu_bo_map;
+   ws->base.buffer_unmap = amdgpu_bo_unmap;
+   ws->base.buffer_wait = amdgpu_bo_wait;
+   ws->base.buffer_create = amdgpu_bo_create;
+   ws->base.buffer_from_handle = amdgpu_bo_from_handle;
+   ws->base.buffer_from_ptr = amdgpu_bo_from_ptr;
+   ws->base.buffer_get_handle = amdgpu_bo_get_handle;
+   ws->base.buffer_get_virtual_address = amdgpu_bo_get_va;
+   ws->base.buffer_get_initial_domain = amdgpu_bo_get_initial_domain;
+}

src/gallium/winsys/amdgpu/drm/amdgpu_bo.h

@@ -0,0 +1,80 @@
+/*
+ * Copyright © 2008 Jérôme Glisse
+ * Copyright © 2011 Marek Olšák <maraeo@gmail.com>
+ * Copyright © 2015 Advanced Micro Devices, Inc.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining
+ * a copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
+ * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS, AUTHORS
+ * AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ */
+/*
+ * Authors:
+ *      Marek Olšák <maraeo@gmail.com>
+ */
+
+#ifndef AMDGPU_BO_H
+#define AMDGPU_BO_H
+
+#include "amdgpu_winsys.h"
+#include "pipebuffer/pb_bufmgr.h"
+
+struct amdgpu_bo_desc {
+   struct pb_desc base;
+
+   enum radeon_bo_domain initial_domain;
+   unsigned flags;
+};
+
+struct amdgpu_winsys_bo {
+   struct pb_buffer base;
+
+   struct amdgpu_winsys *rws;
+   void *user_ptr; /* from buffer_from_ptr */
+
+   amdgpu_bo_handle bo;
+   uint32_t unique_id;
+   amdgpu_va_handle va_handle;
+   uint64_t va;
+   enum radeon_bo_domain initial_domain;
+
+   /* how many command streams is this bo referenced in? */
+   int num_cs_references;
+
+   /* whether buffer_get_handle or buffer_from_handle was called,
+    * it can only transition from false to true
+    */
+   volatile int is_shared; /* bool (int for atomicity) */
+
+   /* Fences for buffer synchronization. */
+   struct pipe_fence_handle *fence[RING_LAST];
+};
+
+struct pb_manager *amdgpu_bomgr_create(struct amdgpu_winsys *rws);
+void amdgpu_bomgr_init_functions(struct amdgpu_winsys *ws);
+
+static inline
+void amdgpu_winsys_bo_reference(struct amdgpu_winsys_bo **dst,
+                                struct amdgpu_winsys_bo *src)
+{
+   pb_reference((struct pb_buffer**)dst, (struct pb_buffer*)src);
+}
+
+#endif

src/gallium/winsys/amdgpu/drm/amdgpu_cs.c

@@ -0,0 +1,704 @@
+/*
+ * Copyright © 2008 Jérôme Glisse
+ * Copyright © 2010 Marek Olšák <maraeo@gmail.com>
+ * Copyright © 2015 Advanced Micro Devices, Inc.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining
+ * a copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
+ * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS, AUTHORS
+ * AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ */
+/*
+ * Authors:
+ *      Marek Olšák <maraeo@gmail.com>
+ */
+
+#include "amdgpu_cs.h"
+#include "os/os_time.h"
+#include <stdio.h>
+#include <amdgpu_drm.h>
+
+
+/* FENCES */
+
+static struct pipe_fence_handle *
+amdgpu_fence_create(struct amdgpu_ctx *ctx, unsigned ip_type,
+                    unsigned ip_instance, unsigned ring)
+{
+   struct amdgpu_fence *fence = CALLOC_STRUCT(amdgpu_fence);
+
+   fence->reference.count = 1;
+   fence->ctx = ctx;
+   fence->fence.context = ctx->ctx;
+   fence->fence.ip_type = ip_type;
+   fence->fence.ip_instance = ip_instance;
+   fence->fence.ring = ring;
+   p_atomic_inc(&ctx->refcount);
+   return (struct pipe_fence_handle *)fence;
+}
+
+static void amdgpu_fence_submitted(struct pipe_fence_handle *fence,
+				struct amdgpu_cs_request* request,
+				uint64_t *user_fence_cpu_address)
+{
+   struct amdgpu_fence *rfence = (struct amdgpu_fence*)fence;
+
+   rfence->fence.fence = request->seq_no;
+   rfence->user_fence_cpu_address = user_fence_cpu_address;
+}
+
+static void amdgpu_fence_signalled(struct pipe_fence_handle *fence)
+{
+   struct amdgpu_fence *rfence = (struct amdgpu_fence*)fence;
+
+   rfence->signalled = true;
+}
+
+bool amdgpu_fence_wait(struct pipe_fence_handle *fence, uint64_t timeout,
+                       bool absolute)
+{
+   struct amdgpu_fence *rfence = (struct amdgpu_fence*)fence;
+   uint32_t expired;
+   int64_t abs_timeout;
+   uint64_t *user_fence_cpu;
+   int r;
+
+   if (rfence->signalled)
+      return true;
+
+   if (absolute)
+      abs_timeout = timeout;
+   else
+      abs_timeout = os_time_get_absolute_timeout(timeout);
+
+   user_fence_cpu = rfence->user_fence_cpu_address;
+   if (user_fence_cpu && *user_fence_cpu >= rfence->fence.fence) {
+	rfence->signalled = true;
+	return true;
+   }
+   /* Now use the libdrm query. */
+   r = amdgpu_cs_query_fence_status(&rfence->fence,
+				    abs_timeout,
+				    AMDGPU_QUERY_FENCE_TIMEOUT_IS_ABSOLUTE,
+				    &expired);
+   if (r) {
+      fprintf(stderr, "amdgpu: amdgpu_cs_query_fence_status failed.\n");
+      return FALSE;
+   }
+
+   if (expired) {
+      /* This variable can only transition from false to true, so it doesn't
+       * matter if threads race for it. */
+      rfence->signalled = true;
+      return true;
+   }
+   return false;
+}
+
+static bool amdgpu_fence_wait_rel_timeout(struct radeon_winsys *rws,
+                                          struct pipe_fence_handle *fence,
+                                          uint64_t timeout)
+{
+   return amdgpu_fence_wait(fence, timeout, false);
+}
+
+/* CONTEXTS */
+
+static struct radeon_winsys_ctx *amdgpu_ctx_create(struct radeon_winsys *ws)
+{
+   struct amdgpu_ctx *ctx = CALLOC_STRUCT(amdgpu_ctx);
+   int r;
+   struct amdgpu_bo_alloc_request alloc_buffer = {};
+   amdgpu_bo_handle buf_handle;
+
+   ctx->ws = amdgpu_winsys(ws);
+   ctx->refcount = 1;
+
+   r = amdgpu_cs_ctx_create(ctx->ws->dev, &ctx->ctx);
+   if (r) {
+      fprintf(stderr, "amdgpu: amdgpu_cs_ctx_create failed. (%i)\n", r);
+      FREE(ctx);
+      return NULL;
+   }
+
+   alloc_buffer.alloc_size = 4 * 1024;
+   alloc_buffer.phys_alignment = 4 *1024;
+   alloc_buffer.preferred_heap = AMDGPU_GEM_DOMAIN_GTT;
+
+   r = amdgpu_bo_alloc(ctx->ws->dev, &alloc_buffer, &buf_handle);
+   if (r) {
+      fprintf(stderr, "amdgpu: amdgpu_bo_alloc failed. (%i)\n", r);
+      amdgpu_cs_ctx_free(ctx->ctx);
+      FREE(ctx);
+      return NULL;
+   }
+
+   r = amdgpu_bo_cpu_map(buf_handle, (void**)&ctx->user_fence_cpu_address_base);
+   if (r) {
+      fprintf(stderr, "amdgpu: amdgpu_bo_cpu_map failed. (%i)\n", r);
+      amdgpu_bo_free(buf_handle);
+      amdgpu_cs_ctx_free(ctx->ctx);
+      FREE(ctx);
+      return NULL;
+   }
+
+   memset(ctx->user_fence_cpu_address_base, 0, alloc_buffer.alloc_size);
+   ctx->user_fence_bo = buf_handle;
+
+   return (struct radeon_winsys_ctx*)ctx;
+}
+
+static void amdgpu_ctx_destroy(struct radeon_winsys_ctx *rwctx)
+{
+   amdgpu_ctx_unref((struct amdgpu_ctx*)rwctx);
+}
+
+static enum pipe_reset_status
+amdgpu_ctx_query_reset_status(struct radeon_winsys_ctx *rwctx)
+{
+   struct amdgpu_ctx *ctx = (struct amdgpu_ctx*)rwctx;
+   uint32_t result, hangs;
+   int r;
+
+   r = amdgpu_cs_query_reset_state(ctx->ctx, &result, &hangs);
+   if (r) {
+      fprintf(stderr, "amdgpu: amdgpu_cs_query_reset_state failed. (%i)\n", r);
+      return PIPE_NO_RESET;
+   }
+
+   switch (result) {
+   case AMDGPU_CTX_GUILTY_RESET:
+      return PIPE_GUILTY_CONTEXT_RESET;
+   case AMDGPU_CTX_INNOCENT_RESET:
+      return PIPE_INNOCENT_CONTEXT_RESET;
+   case AMDGPU_CTX_UNKNOWN_RESET:
+      return PIPE_UNKNOWN_CONTEXT_RESET;
+   case AMDGPU_CTX_NO_RESET:
+   default:
+      return PIPE_NO_RESET;
+   }
+}
+
+/* COMMAND SUBMISSION */
+
+static bool amdgpu_get_new_ib(struct amdgpu_cs *cs)
+{
+   /* The maximum size is 4MB - 1B, which is unaligned.
+    * Use aligned size 4MB - 16B. */
+   const unsigned max_ib_size = (1024 * 1024 - 16) * 4;
+   const unsigned min_ib_size = 24 * 1024 * 4;
+
+   cs->base.cdw = 0;
+   cs->base.buf = NULL;
+
+   /* Allocate a new buffer for IBs if the current buffer is all used. */
+   if (!cs->big_ib_buffer ||
+       cs->used_ib_space + min_ib_size > cs->big_ib_buffer->size) {
+      struct radeon_winsys *ws = &cs->ctx->ws->base;
+      struct radeon_winsys_cs_handle *winsys_bo;
+
+      pb_reference(&cs->big_ib_buffer, NULL);
+      cs->big_ib_winsys_buffer = NULL;
+      cs->ib_mapped = NULL;
+      cs->used_ib_space = 0;
+
+      cs->big_ib_buffer = ws->buffer_create(ws, max_ib_size,
+                                            4096, true,
+                                            RADEON_DOMAIN_GTT,
+                                            RADEON_FLAG_CPU_ACCESS);
+      if (!cs->big_ib_buffer)
+         return false;
+
+      winsys_bo = ws->buffer_get_cs_handle(cs->big_ib_buffer);
+
+      cs->ib_mapped = ws->buffer_map(winsys_bo, NULL, PIPE_TRANSFER_WRITE);
+      if (!cs->ib_mapped) {
+         pb_reference(&cs->big_ib_buffer, NULL);
+         return false;
+      }
+
+      cs->big_ib_winsys_buffer = (struct amdgpu_winsys_bo*)winsys_bo;
+   }
+
+   cs->ib.ib_mc_address = cs->big_ib_winsys_buffer->va + cs->used_ib_space;
+   cs->base.buf = (uint32_t*)(cs->ib_mapped + cs->used_ib_space);
+   cs->base.max_dw = (cs->big_ib_buffer->size - cs->used_ib_space) / 4;
+   return true;
+}
+
+static boolean amdgpu_init_cs_context(struct amdgpu_cs *cs,
+                                      enum ring_type ring_type)
+{
+   int i;
+
+   switch (ring_type) {
+   case RING_DMA:
+      cs->request.ip_type = AMDGPU_HW_IP_DMA;
+      break;
+
+   case RING_UVD:
+      cs->request.ip_type = AMDGPU_HW_IP_UVD;
+      break;
+
+   case RING_VCE:
+      cs->request.ip_type = AMDGPU_HW_IP_VCE;
+      break;
+
+   case RING_COMPUTE:
+      cs->request.ip_type = AMDGPU_HW_IP_COMPUTE;
+      break;
+
+   default:
+   case RING_GFX:
+      cs->request.ip_type = AMDGPU_HW_IP_GFX;
+      break;
+   }
+
+   cs->request.number_of_ibs = 1;
+   cs->request.ibs = &cs->ib;
+
+   cs->max_num_buffers = 512;
+   cs->buffers = (struct amdgpu_cs_buffer*)
+                  CALLOC(1, cs->max_num_buffers * sizeof(struct amdgpu_cs_buffer));
+   if (!cs->buffers) {
+      return FALSE;
+   }
+
+   cs->handles = CALLOC(1, cs->max_num_buffers * sizeof(amdgpu_bo_handle));
+   if (!cs->handles) {
+      FREE(cs->buffers);
+      return FALSE;
+   }
+
+   cs->flags = CALLOC(1, cs->max_num_buffers);
+   if (!cs->flags) {
+      FREE(cs->handles);
+      FREE(cs->buffers);
+      return FALSE;
+   }
+
+   for (i = 0; i < Elements(cs->buffer_indices_hashlist); i++) {
+      cs->buffer_indices_hashlist[i] = -1;
+   }
+   return TRUE;
+}
+
+static void amdgpu_cs_context_cleanup(struct amdgpu_cs *cs)
+{
+   unsigned i;
+
+   for (i = 0; i < cs->num_buffers; i++) {
+      p_atomic_dec(&cs->buffers[i].bo->num_cs_references);
+      amdgpu_winsys_bo_reference(&cs->buffers[i].bo, NULL);
+      cs->handles[i] = NULL;
+      cs->flags[i] = 0;
+   }
+
+   cs->num_buffers = 0;
+   cs->used_gart = 0;
+   cs->used_vram = 0;
+
+   for (i = 0; i < Elements(cs->buffer_indices_hashlist); i++) {
+      cs->buffer_indices_hashlist[i] = -1;
+   }
+}
+
+static void amdgpu_destroy_cs_context(struct amdgpu_cs *cs)
+{
+   amdgpu_cs_context_cleanup(cs);
+   FREE(cs->flags);
+   FREE(cs->buffers);
+   FREE(cs->handles);
+   FREE(cs->request.dependencies);
+}
+
+
+static struct radeon_winsys_cs *
+amdgpu_cs_create(struct radeon_winsys_ctx *rwctx,
+                 enum ring_type ring_type,
+                 void (*flush)(void *ctx, unsigned flags,
+                               struct pipe_fence_handle **fence),
+                 void *flush_ctx,
+                 struct radeon_winsys_cs_handle *trace_buf)
+{
+   struct amdgpu_ctx *ctx = (struct amdgpu_ctx*)rwctx;
+   struct amdgpu_cs *cs;
+
+   cs = CALLOC_STRUCT(amdgpu_cs);
+   if (!cs) {
+      return NULL;
+   }
+
+   cs->ctx = ctx;
+   cs->flush_cs = flush;
+   cs->flush_data = flush_ctx;
+   cs->base.ring_type = ring_type;
+
+   if (!amdgpu_init_cs_context(cs, ring_type)) {
+      FREE(cs);
+      return NULL;
+   }
+
+   if (!amdgpu_get_new_ib(cs)) {
+      amdgpu_destroy_cs_context(cs);
+      FREE(cs);
+      return NULL;
+   }
+
+   p_atomic_inc(&ctx->ws->num_cs);
+   return &cs->base;
+}
+
+#define OUT_CS(cs, value) (cs)->buf[(cs)->cdw++] = (value)
+
+int amdgpu_get_reloc(struct amdgpu_cs *cs, struct amdgpu_winsys_bo *bo)
+{
+   unsigned hash = bo->unique_id & (Elements(cs->buffer_indices_hashlist)-1);
+   int i = cs->buffer_indices_hashlist[hash];
+
+   /* not found or found */
+   if (i == -1 || cs->buffers[i].bo == bo)
+      return i;
+
+   /* Hash collision, look for the BO in the list of relocs linearly. */
+   for (i = cs->num_buffers - 1; i >= 0; i--) {
+      if (cs->buffers[i].bo == bo) {
+         /* Put this reloc in the hash list.
+          * This will prevent additional hash collisions if there are
+          * several consecutive get_reloc calls for the same buffer.
+          *
+          * Example: Assuming buffers A,B,C collide in the hash list,
+          * the following sequence of relocs:
+          *         AAAAAAAAAAABBBBBBBBBBBBBBCCCCCCCC
+          * will collide here: ^ and here:   ^,
+          * meaning that we should get very few collisions in the end. */
+         cs->buffer_indices_hashlist[hash] = i;
+         return i;
+      }
+   }
+   return -1;
+}
+
+static unsigned amdgpu_add_reloc(struct amdgpu_cs *cs,
+                                 struct amdgpu_winsys_bo *bo,
+                                 enum radeon_bo_usage usage,
+                                 enum radeon_bo_domain domains,
+                                 unsigned priority,
+                                 enum radeon_bo_domain *added_domains)
+{
+   struct amdgpu_cs_buffer *reloc;
+   unsigned hash = bo->unique_id & (Elements(cs->buffer_indices_hashlist)-1);
+   int i = -1;
+
+   priority = MIN2(priority, 15);
+   *added_domains = 0;
+
+   i = amdgpu_get_reloc(cs, bo);
+
+   if (i >= 0) {
+      reloc = &cs->buffers[i];
+      reloc->usage |= usage;
+      *added_domains = domains & ~reloc->domains;
+      reloc->domains |= domains;
+      cs->flags[i] = MAX2(cs->flags[i], priority);
+      return i;
+   }
+
+   /* New relocation, check if the backing array is large enough. */
+   if (cs->num_buffers >= cs->max_num_buffers) {
+      uint32_t size;
+      cs->max_num_buffers += 10;
+
+      size = cs->max_num_buffers * sizeof(struct amdgpu_cs_buffer);
+      cs->buffers = realloc(cs->buffers, size);
+
+      size = cs->max_num_buffers * sizeof(amdgpu_bo_handle);
+      cs->handles = realloc(cs->handles, size);
+
+      cs->flags = realloc(cs->flags, cs->max_num_buffers);
+   }
+
+   /* Initialize the new relocation. */
+   cs->buffers[cs->num_buffers].bo = NULL;
+   amdgpu_winsys_bo_reference(&cs->buffers[cs->num_buffers].bo, bo);
+   cs->handles[cs->num_buffers] = bo->bo;
+   cs->flags[cs->num_buffers] = priority;
+   p_atomic_inc(&bo->num_cs_references);
+   reloc = &cs->buffers[cs->num_buffers];
+   reloc->bo = bo;
+   reloc->usage = usage;
+   reloc->domains = domains;
+
+   cs->buffer_indices_hashlist[hash] = cs->num_buffers;
+
+   *added_domains = domains;
+   return cs->num_buffers++;
+}
+
+static unsigned amdgpu_cs_add_reloc(struct radeon_winsys_cs *rcs,
+                                    struct radeon_winsys_cs_handle *buf,
+                                    enum radeon_bo_usage usage,
+                                    enum radeon_bo_domain domains,
+                                    enum radeon_bo_priority priority)
+{
+   /* Don't use the "domains" parameter. Amdgpu doesn't support changing
+    * the buffer placement during command submission.
+    */
+   struct amdgpu_cs *cs = amdgpu_cs(rcs);
+   struct amdgpu_winsys_bo *bo = (struct amdgpu_winsys_bo*)buf;
+   enum radeon_bo_domain added_domains;
+   unsigned index = amdgpu_add_reloc(cs, bo, usage, bo->initial_domain,
+                                     priority, &added_domains);
+
+   if (added_domains & RADEON_DOMAIN_GTT)
+      cs->used_gart += bo->base.size;
+   if (added_domains & RADEON_DOMAIN_VRAM)
+      cs->used_vram += bo->base.size;
+
+   return index;
+}
+
+static int amdgpu_cs_get_reloc(struct radeon_winsys_cs *rcs,
+                               struct radeon_winsys_cs_handle *buf)
+{
+   struct amdgpu_cs *cs = amdgpu_cs(rcs);
+
+   return amdgpu_get_reloc(cs, (struct amdgpu_winsys_bo*)buf);
+}
+
+static boolean amdgpu_cs_validate(struct radeon_winsys_cs *rcs)
+{
+   return TRUE;
+}
+
+static boolean amdgpu_cs_memory_below_limit(struct radeon_winsys_cs *rcs, uint64_t vram, uint64_t gtt)
+{
+   struct amdgpu_cs *cs = amdgpu_cs(rcs);
+   boolean status =
+         (cs->used_gart + gtt) < cs->ctx->ws->info.gart_size * 0.7 &&
+         (cs->used_vram + vram) < cs->ctx->ws->info.vram_size * 0.7;
+
+   return status;
+}
+
+static void amdgpu_cs_do_submission(struct amdgpu_cs *cs,
+                                    struct pipe_fence_handle **out_fence)
+{
+   struct amdgpu_winsys *ws = cs->ctx->ws;
+   struct pipe_fence_handle *fence;
+   int i, j, r;
+
+   /* Create a fence. */
+   fence = amdgpu_fence_create(cs->ctx,
+                               cs->request.ip_type,
+                               cs->request.ip_instance,
+                               cs->request.ring);
+   if (out_fence)
+      amdgpu_fence_reference(out_fence, fence);
+
+   cs->request.number_of_dependencies = 0;
+
+   /* Since the kernel driver doesn't synchronize execution between different
+    * rings automatically, we have to add fence dependencies manually. */
+   pipe_mutex_lock(ws->bo_fence_lock);
+   for (i = 0; i < cs->num_buffers; i++) {
+      for (j = 0; j < RING_LAST; j++) {
+         struct amdgpu_cs_fence *dep;
+         unsigned idx;
+
+         struct amdgpu_fence *bo_fence = (void *)cs->buffers[i].bo->fence[j];
+         if (!bo_fence)
+            continue;
+
+         if (bo_fence->ctx == cs->ctx &&
+             bo_fence->fence.ip_type == cs->request.ip_type &&
+             bo_fence->fence.ip_instance == cs->request.ip_instance &&
+             bo_fence->fence.ring == cs->request.ring)
+            continue;
+
+         if (amdgpu_fence_wait((void *)bo_fence, 0, false))
+            continue;
+
+         idx = cs->request.number_of_dependencies++;
+         if (idx >= cs->max_dependencies) {
+            unsigned size;
+
+            cs->max_dependencies = idx + 8;
+            size = cs->max_dependencies * sizeof(struct amdgpu_cs_fence);
+            cs->request.dependencies = realloc(cs->request.dependencies, size);
+         }
+
+         dep = &cs->request.dependencies[idx];
+         memcpy(dep, &bo_fence->fence, sizeof(*dep));
+      }
+   }
+
+   cs->request.fence_info.handle = NULL;
+   if (cs->request.ip_type != AMDGPU_HW_IP_UVD && cs->request.ip_type != AMDGPU_HW_IP_VCE) {
+	cs->request.fence_info.handle = cs->ctx->user_fence_bo;
+	cs->request.fence_info.offset = cs->base.ring_type;
+   }
+
+   r = amdgpu_cs_submit(cs->ctx->ctx, 0, &cs->request, 1);
+   if (r) {
+      if (r == -ENOMEM)
+         fprintf(stderr, "amdgpu: Not enough memory for command submission.\n");
+      else
+         fprintf(stderr, "amdgpu: The CS has been rejected, "
+                 "see dmesg for more information.\n");
+
+      amdgpu_fence_signalled(fence);
+   } else {
+      /* Success. */
+      uint64_t *user_fence = NULL;
+      if (cs->request.ip_type != AMDGPU_HW_IP_UVD && cs->request.ip_type != AMDGPU_HW_IP_VCE)
+         user_fence = cs->ctx->user_fence_cpu_address_base +
+                      cs->request.fence_info.offset;
+      amdgpu_fence_submitted(fence, &cs->request, user_fence);
+
+      for (i = 0; i < cs->num_buffers; i++)
+         amdgpu_fence_reference(&cs->buffers[i].bo->fence[cs->base.ring_type],
+                                fence);
+   }
+   pipe_mutex_unlock(ws->bo_fence_lock);
+   amdgpu_fence_reference(&fence, NULL);
+}
+
+static void amdgpu_cs_sync_flush(struct radeon_winsys_cs *rcs)
+{
+   /* no-op */
+}
+
+DEBUG_GET_ONCE_BOOL_OPTION(noop, "RADEON_NOOP", FALSE)
+
+static void amdgpu_cs_flush(struct radeon_winsys_cs *rcs,
+                            unsigned flags,
+                            struct pipe_fence_handle **fence,
+                            uint32_t cs_trace_id)
+{
+   struct amdgpu_cs *cs = amdgpu_cs(rcs);
+   struct amdgpu_winsys *ws = cs->ctx->ws;
+
+   switch (cs->base.ring_type) {
+   case RING_DMA:
+      /* pad DMA ring to 8 DWs */
+      if (ws->info.chip_class <= SI) {
+         while (rcs->cdw & 7)
+            OUT_CS(&cs->base, 0xf0000000); /* NOP packet */
+      } else {
+         while (rcs->cdw & 7)
+            OUT_CS(&cs->base, 0x00000000); /* NOP packet */
+      }
+      break;
+   case RING_GFX:
+      /* pad DMA ring to 8 DWs to meet CP fetch alignment requirements
+             * r6xx, requires at least 4 dw alignment to avoid a hw bug.
+             */
+      if (ws->info.chip_class <= SI) {
+         while (rcs->cdw & 7)
+            OUT_CS(&cs->base, 0x80000000); /* type2 nop packet */
+      } else {
+         while (rcs->cdw & 7)
+            OUT_CS(&cs->base, 0xffff1000); /* type3 nop packet */
+      }
+      break;
+   case RING_UVD:
+      while (rcs->cdw & 15)
+         OUT_CS(&cs->base, 0x80000000); /* type2 nop packet */
+      break;
+   default:
+      break;
+   }
+
+   if (rcs->cdw > rcs->max_dw) {
+      fprintf(stderr, "amdgpu: command stream overflowed\n");
+   }
+
+   amdgpu_cs_add_reloc(rcs, (void*)cs->big_ib_winsys_buffer,
+		       RADEON_USAGE_READ, 0, RADEON_PRIO_MIN);
+
+   /* If the CS is not empty or overflowed.... */
+   if (cs->base.cdw && cs->base.cdw <= cs->base.max_dw && !debug_get_option_noop()) {
+      int r;
+
+      r = amdgpu_bo_list_create(ws->dev, cs->num_buffers,
+                                cs->handles, cs->flags,
+                                &cs->request.resources);
+
+      if (r) {
+         fprintf(stderr, "amdgpu: resource list creation failed (%d)\n", r);
+         cs->request.resources = NULL;
+	 goto cleanup;
+      }
+
+      cs->ib.size = cs->base.cdw;
+      cs->used_ib_space += cs->base.cdw * 4;
+
+      amdgpu_cs_do_submission(cs, fence);
+
+      /* Cleanup. */
+      if (cs->request.resources)
+         amdgpu_bo_list_destroy(cs->request.resources);
+   }
+
+cleanup:
+   amdgpu_cs_context_cleanup(cs);
+   amdgpu_get_new_ib(cs);
+
+   ws->num_cs_flushes++;
+}
+
+static void amdgpu_cs_destroy(struct radeon_winsys_cs *rcs)
+{
+   struct amdgpu_cs *cs = amdgpu_cs(rcs);
+
+   amdgpu_destroy_cs_context(cs);
+   p_atomic_dec(&cs->ctx->ws->num_cs);
+   pb_reference(&cs->big_ib_buffer, NULL);
+   FREE(cs);
+}
+
+static boolean amdgpu_bo_is_referenced(struct radeon_winsys_cs *rcs,
+                                       struct radeon_winsys_cs_handle *_buf,
+                                       enum radeon_bo_usage usage)
+{
+   struct amdgpu_cs *cs = amdgpu_cs(rcs);
+   struct amdgpu_winsys_bo *bo = (struct amdgpu_winsys_bo*)_buf;
+
+   return amdgpu_bo_is_referenced_by_cs_with_usage(cs, bo, usage);
+}
+
+void amdgpu_cs_init_functions(struct amdgpu_winsys *ws)
+{
+   ws->base.ctx_create = amdgpu_ctx_create;
+   ws->base.ctx_destroy = amdgpu_ctx_destroy;
+   ws->base.ctx_query_reset_status = amdgpu_ctx_query_reset_status;
+   ws->base.cs_create = amdgpu_cs_create;
+   ws->base.cs_destroy = amdgpu_cs_destroy;
+   ws->base.cs_add_reloc = amdgpu_cs_add_reloc;
+   ws->base.cs_get_reloc = amdgpu_cs_get_reloc;
+   ws->base.cs_validate = amdgpu_cs_validate;
+   ws->base.cs_memory_below_limit = amdgpu_cs_memory_below_limit;
+   ws->base.cs_flush = amdgpu_cs_flush;
+   ws->base.cs_is_buffer_referenced = amdgpu_bo_is_referenced;
+   ws->base.cs_sync_flush = amdgpu_cs_sync_flush;
+   ws->base.fence_wait = amdgpu_fence_wait_rel_timeout;
+   ws->base.fence_reference = amdgpu_fence_reference;
+}

src/gallium/winsys/amdgpu/drm/amdgpu_cs.h

@@ -0,0 +1,162 @@
+/*
+ * Copyright © 2011 Marek Olšák <maraeo@gmail.com>
+ * Copyright © 2015 Advanced Micro Devices, Inc.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining
+ * a copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
+ * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS, AUTHORS
+ * AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ */
+/*
+ * Authors:
+ *      Marek Olšák <maraeo@gmail.com>
+ */
+
+#ifndef AMDGPU_CS_H
+#define AMDGPU_CS_H
+
+#include "amdgpu_bo.h"
+#include "util/u_memory.h"
+
+struct amdgpu_ctx {
+   struct amdgpu_winsys *ws;
+   amdgpu_context_handle ctx;
+   amdgpu_bo_handle user_fence_bo;
+   uint64_t *user_fence_cpu_address_base;
+   int refcount;
+};
+
+struct amdgpu_cs_buffer {
+   struct amdgpu_winsys_bo *bo;
+   enum radeon_bo_usage usage;
+   enum radeon_bo_domain domains;
+};
+
+
+struct amdgpu_cs {
+   struct radeon_winsys_cs base;
+   struct amdgpu_ctx *ctx;
+
+   /* Flush CS. */
+   void (*flush_cs)(void *ctx, unsigned flags, struct pipe_fence_handle **fence);
+   void *flush_data;
+
+   /* A buffer out of which new IBs are allocated. */
+   struct pb_buffer *big_ib_buffer; /* for holding the reference */
+   struct amdgpu_winsys_bo *big_ib_winsys_buffer;
+   uint8_t *ib_mapped;
+   unsigned used_ib_space;
+
+   /* amdgpu_cs_submit parameters */
+   struct amdgpu_cs_request    request;
+   struct amdgpu_cs_ib_info    ib;
+
+   /* Relocs. */
+   unsigned                    max_num_buffers;
+   unsigned                    num_buffers;
+   amdgpu_bo_handle            *handles;
+   uint8_t                     *flags;
+   struct amdgpu_cs_buffer     *buffers;
+
+   int                         buffer_indices_hashlist[512];
+
+   unsigned                    used_vram;
+   unsigned                    used_gart;
+
+   unsigned                    max_dependencies;
+};
+
+struct amdgpu_fence {
+   struct pipe_reference reference;
+
+   struct amdgpu_ctx *ctx;  /* submission context */
+   struct amdgpu_cs_fence fence;
+   uint64_t *user_fence_cpu_address;
+
+   volatile int signalled;              /* bool (int for atomicity) */
+};
+
+static inline void amdgpu_ctx_unref(struct amdgpu_ctx *ctx)
+{
+   if (p_atomic_dec_zero(&ctx->refcount)) {
+      amdgpu_cs_ctx_free(ctx->ctx);
+      amdgpu_bo_free(ctx->user_fence_bo);
+      FREE(ctx);
+   }
+}
+
+static inline void amdgpu_fence_reference(struct pipe_fence_handle **dst,
+                                          struct pipe_fence_handle *src)
+{
+   struct amdgpu_fence **rdst = (struct amdgpu_fence **)dst;
+   struct amdgpu_fence *rsrc = (struct amdgpu_fence *)src;
+
+   if (pipe_reference(&(*rdst)->reference, &rsrc->reference)) {
+      amdgpu_ctx_unref((*rdst)->ctx);
+      FREE(*rdst);
+   }
+   *rdst = rsrc;
+}
+
+int amdgpu_get_reloc(struct amdgpu_cs *csc, struct amdgpu_winsys_bo *bo);
+
+static inline struct amdgpu_cs *
+amdgpu_cs(struct radeon_winsys_cs *base)
+{
+   return (struct amdgpu_cs*)base;
+}
+
+static inline boolean
+amdgpu_bo_is_referenced_by_cs(struct amdgpu_cs *cs,
+                              struct amdgpu_winsys_bo *bo)
+{
+   int num_refs = bo->num_cs_references;
+   return num_refs == bo->rws->num_cs ||
+         (num_refs && amdgpu_get_reloc(cs, bo) != -1);
+}
+
+static inline boolean
+amdgpu_bo_is_referenced_by_cs_with_usage(struct amdgpu_cs *cs,
+                                         struct amdgpu_winsys_bo *bo,
+                                         enum radeon_bo_usage usage)
+{
+   int index;
+
+   if (!bo->num_cs_references)
+      return FALSE;
+
+   index = amdgpu_get_reloc(cs, bo);
+   if (index == -1)
+      return FALSE;
+
+   return (cs->buffers[index].usage & usage) != 0;
+}
+
+static inline boolean
+amdgpu_bo_is_referenced_by_any_cs(struct amdgpu_winsys_bo *bo)
+{
+   return bo->num_cs_references != 0;
+}
+
+bool amdgpu_fence_wait(struct pipe_fence_handle *fence, uint64_t timeout,
+                       bool absolute);
+void amdgpu_cs_init_functions(struct amdgpu_winsys *ws);
+
+#endif

src/gallium/winsys/amdgpu/drm/amdgpu_public.h

@@ -0,0 +1,40 @@
+/*
+ * Copyright © 2015 Advanced Micro Devices, Inc.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining
+ * a copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
+ * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS, AUTHORS
+ * AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ */
+
+#ifndef AMDGPU_PUBLIC_H
+#define AMDGPU_PUBLIC_H
+
+#include "pipe/p_defines.h"
+
+struct radeon_winsys;
+struct pipe_screen;
+
+typedef struct pipe_screen *(*radeon_screen_create_t)(struct radeon_winsys *);
+
+struct radeon_winsys *
+amdgpu_winsys_create(int fd, radeon_screen_create_t screen_create);
+
+#endif

src/gallium/winsys/amdgpu/drm/amdgpu_winsys.c

@@ -0,0 +1,448 @@
+/*
+ * Copyright © 2009 Corbin Simpson <MostAwesomeDude@gmail.com>
+ * Copyright © 2009 Joakim Sindholt <opensource@zhasha.com>
+ * Copyright © 2011 Marek Olšák <maraeo@gmail.com>
+ * Copyright © 2015 Advanced Micro Devices, Inc.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining
+ * a copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
+ * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS, AUTHORS
+ * AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ */
+/*
+ * Authors:
+ *      Marek Olšák <maraeo@gmail.com>
+ */
+
+#include "amdgpu_cs.h"
+#include "amdgpu_public.h"
+
+#include "util/u_hash_table.h"
+#include <amdgpu_drm.h>
+#include <xf86drm.h>
+#include <stdio.h>
+#include <sys/stat.h>
+
+#define CIK_TILE_MODE_COLOR_2D			14
+
+#define CIK__GB_TILE_MODE__PIPE_CONFIG(x)        (((x) >> 6) & 0x1f)
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P2               0
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P4_8x16          4
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P4_16x16         5
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P4_16x32         6
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P4_32x32         7
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P8_16x16_8x16    8
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_8x16    9
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_8x16    10
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_16x16   11
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x16   12
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x32   13
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P8_32x64_32x32   14
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_8X16   16
+#define     CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_16X16  17
+
+static struct util_hash_table *dev_tab = NULL;
+pipe_static_mutex(dev_tab_mutex);
+
+static unsigned cik_get_num_tile_pipes(struct amdgpu_gpu_info *info)
+{
+   unsigned mode2d = info->gb_tile_mode[CIK_TILE_MODE_COLOR_2D];
+
+   switch (CIK__GB_TILE_MODE__PIPE_CONFIG(mode2d)) {
+   case CIK__PIPE_CONFIG__ADDR_SURF_P2:
+   default:
+       return 2;
+   case CIK__PIPE_CONFIG__ADDR_SURF_P4_8x16:
+   case CIK__PIPE_CONFIG__ADDR_SURF_P4_16x16:
+   case CIK__PIPE_CONFIG__ADDR_SURF_P4_16x32:
+   case CIK__PIPE_CONFIG__ADDR_SURF_P4_32x32:
+       return 4;
+   case CIK__PIPE_CONFIG__ADDR_SURF_P8_16x16_8x16:
+   case CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_8x16:
+   case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_8x16:
+   case CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_16x16:
+   case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x16:
+   case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x32:
+   case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x64_32x32:
+       return 8;
+   case CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_8X16:
+   case CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_16X16:
+       return 16;
+   }
+}
+
+/* Convert Sea Islands register values GB_ADDR_CFG and MC_ADDR_CFG
+ * into GB_TILING_CONFIG register which is only present on R600-R700. */
+static unsigned r600_get_gb_tiling_config(struct amdgpu_gpu_info *info)
+{
+   unsigned num_pipes = info->gb_addr_cfg & 0x7;
+   unsigned num_banks = info->mc_arb_ramcfg & 0x3;
+   unsigned pipe_interleave_bytes = (info->gb_addr_cfg >> 4) & 0x7;
+   unsigned row_size = (info->gb_addr_cfg >> 28) & 0x3;
+
+   return num_pipes | (num_banks << 4) |
+         (pipe_interleave_bytes << 8) |
+         (row_size << 12);
+}
+
+/* Helper function to do the ioctls needed for setup and init. */
+static boolean do_winsys_init(struct amdgpu_winsys *ws)
+{
+   struct amdgpu_buffer_size_alignments alignment_info = {};
+   struct amdgpu_heap_info vram, gtt;
+   struct drm_amdgpu_info_hw_ip dma = {}, uvd = {}, vce = {};
+   uint32_t vce_version = 0, vce_feature = 0;
+   int r;
+
+   /* Query hardware and driver information. */
+   r = amdgpu_query_gpu_info(ws->dev, &ws->amdinfo);
+   if (r) {
+      fprintf(stderr, "amdgpu: amdgpu_query_gpu_info failed.\n");
+      goto fail;
+   }
+
+   r = amdgpu_query_buffer_size_alignment(ws->dev, &alignment_info);
+   if (r) {
+      fprintf(stderr, "amdgpu: amdgpu_query_buffer_size_alignment failed.\n");
+      goto fail;
+   }
+
+   r = amdgpu_query_heap_info(ws->dev, AMDGPU_GEM_DOMAIN_VRAM, 0, &vram);
+   if (r) {
+      fprintf(stderr, "amdgpu: amdgpu_query_heap_info(vram) failed.\n");
+      goto fail;
+   }
+
+   r = amdgpu_query_heap_info(ws->dev, AMDGPU_GEM_DOMAIN_GTT, 0, &gtt);
+   if (r) {
+      fprintf(stderr, "amdgpu: amdgpu_query_heap_info(gtt) failed.\n");
+      goto fail;
+   }
+
+   r = amdgpu_query_hw_ip_info(ws->dev, AMDGPU_HW_IP_DMA, 0, &dma);
+   if (r) {
+      fprintf(stderr, "amdgpu: amdgpu_query_hw_ip_info(dma) failed.\n");
+      goto fail;
+   }
+
+   r = amdgpu_query_hw_ip_info(ws->dev, AMDGPU_HW_IP_UVD, 0, &uvd);
+   if (r) {
+      fprintf(stderr, "amdgpu: amdgpu_query_hw_ip_info(uvd) failed.\n");
+      goto fail;
+   }
+
+   r = amdgpu_query_hw_ip_info(ws->dev, AMDGPU_HW_IP_VCE, 0, &vce);
+   if (r) {
+      fprintf(stderr, "amdgpu: amdgpu_query_hw_ip_info(vce) failed.\n");
+      goto fail;
+   }
+
+   r = amdgpu_query_firmware_version(ws->dev, AMDGPU_INFO_FW_VCE, 0, 0,
+				     &vce_version, &vce_feature);
+   if (r) {
+      fprintf(stderr, "amdgpu: amdgpu_query_firmware_version(vce) failed.\n");
+      goto fail;
+   }
+
+   /* Set chip identification. */
+   ws->info.pci_id = ws->amdinfo.asic_id; /* TODO: is this correct? */
+
+   switch (ws->info.pci_id) {
+#define CHIPSET(pci_id, name, cfamily) case pci_id: ws->info.family = CHIP_##cfamily; break;
+#include "pci_ids/radeonsi_pci_ids.h"
+#undef CHIPSET
+
+   default:
+      fprintf(stderr, "amdgpu: Invalid PCI ID.\n");
+      goto fail;
+   }
+
+   if (ws->info.family >= CHIP_TONGA)
+      ws->info.chip_class = VI;
+   else if (ws->info.family >= CHIP_BONAIRE)
+      ws->info.chip_class = CIK;
+   else {
+      fprintf(stderr, "amdgpu: Unknown family.\n");
+      goto fail;
+   }
+
+   /* LLVM 3.6 is required for VI. */
+   if (ws->info.chip_class >= VI &&
+       (HAVE_LLVM < 0x0306 ||
+        (HAVE_LLVM == 0x0306 && MESA_LLVM_VERSION_PATCH < 1))) {
+      fprintf(stderr, "amdgpu: LLVM 3.6.1 is required, got LLVM %i.%i.%i\n",
+              HAVE_LLVM >> 8, HAVE_LLVM & 255, MESA_LLVM_VERSION_PATCH);
+      goto fail;
+   }
+
+   /* Set hardware information. */
+   ws->info.gart_size = gtt.heap_size;
+   ws->info.vram_size = vram.heap_size;
+   /* convert the shader clock from KHz to MHz */
+   ws->info.max_sclk = ws->amdinfo.max_engine_clk / 1000;
+   ws->info.max_compute_units = 1; /* TODO */
+   ws->info.max_se = ws->amdinfo.num_shader_engines;
+   ws->info.max_sh_per_se = ws->amdinfo.num_shader_arrays_per_engine;
+   ws->info.has_uvd = uvd.available_rings != 0;
+   ws->info.vce_fw_version =
+         vce.available_rings ? vce_version : 0;
+   ws->info.has_userptr = TRUE;
+   ws->info.r600_num_backends = ws->amdinfo.rb_pipes;
+   ws->info.r600_clock_crystal_freq = ws->amdinfo.gpu_counter_freq;
+   ws->info.r600_tiling_config = r600_get_gb_tiling_config(&ws->amdinfo);
+   ws->info.r600_num_tile_pipes = cik_get_num_tile_pipes(&ws->amdinfo);
+   ws->info.r600_max_pipes = ws->amdinfo.max_quad_shader_pipes; /* TODO: is this correct? */
+   ws->info.r600_virtual_address = TRUE;
+   ws->info.r600_has_dma = dma.available_rings != 0;
+
+   memcpy(ws->info.si_tile_mode_array, ws->amdinfo.gb_tile_mode,
+          sizeof(ws->amdinfo.gb_tile_mode));
+   ws->info.si_tile_mode_array_valid = TRUE;
+   ws->info.si_backend_enabled_mask = ws->amdinfo.enabled_rb_pipes_mask;
+
+   memcpy(ws->info.cik_macrotile_mode_array, ws->amdinfo.gb_macro_tile_mode,
+          sizeof(ws->amdinfo.gb_macro_tile_mode));
+   ws->info.cik_macrotile_mode_array_valid = TRUE;
+
+   ws->gart_page_size = alignment_info.size_remote;
+
+   return TRUE;
+
+fail:
+   amdgpu_device_deinitialize(ws->dev);
+   ws->dev = NULL;
+   return FALSE;
+}
+
+static void amdgpu_winsys_destroy(struct radeon_winsys *rws)
+{
+   struct amdgpu_winsys *ws = (struct amdgpu_winsys*)rws;
+
+   pipe_mutex_destroy(ws->bo_fence_lock);
+
+   ws->cman->destroy(ws->cman);
+   ws->kman->destroy(ws->kman);
+
+   amdgpu_device_deinitialize(ws->dev);
+   FREE(rws);
+}
+
+static void amdgpu_winsys_query_info(struct radeon_winsys *rws,
+                                     struct radeon_info *info)
+{
+   *info = ((struct amdgpu_winsys *)rws)->info;
+}
+
+static boolean amdgpu_cs_request_feature(struct radeon_winsys_cs *rcs,
+                                         enum radeon_feature_id fid,
+                                         boolean enable)
+{
+   return FALSE;
+}
+
+static uint64_t amdgpu_query_value(struct radeon_winsys *rws,
+                                   enum radeon_value_id value)
+{
+   struct amdgpu_winsys *ws = (struct amdgpu_winsys*)rws;
+   struct amdgpu_heap_info heap;
+   uint64_t retval = 0;
+
+   switch (value) {
+   case RADEON_REQUESTED_VRAM_MEMORY:
+      return ws->allocated_vram;
+   case RADEON_REQUESTED_GTT_MEMORY:
+      return ws->allocated_gtt;
+   case RADEON_BUFFER_WAIT_TIME_NS:
+      return ws->buffer_wait_time;
+   case RADEON_TIMESTAMP:
+      amdgpu_query_info(ws->dev, AMDGPU_INFO_TIMESTAMP, 8, &retval);
+      return retval;
+   case RADEON_NUM_CS_FLUSHES:
+      return ws->num_cs_flushes;
+   case RADEON_NUM_BYTES_MOVED:
+      amdgpu_query_info(ws->dev, AMDGPU_INFO_NUM_BYTES_MOVED, 8, &retval);
+      return retval;
+   case RADEON_VRAM_USAGE:
+      amdgpu_query_heap_info(ws->dev, AMDGPU_GEM_DOMAIN_VRAM, 0, &heap);
+      return heap.heap_usage;
+   case RADEON_GTT_USAGE:
+      amdgpu_query_heap_info(ws->dev, AMDGPU_GEM_DOMAIN_GTT, 0, &heap);
+      return heap.heap_usage;
+   case RADEON_GPU_TEMPERATURE:
+   case RADEON_CURRENT_SCLK:
+   case RADEON_CURRENT_MCLK:
+      return 0;
+   case RADEON_GPU_RESET_COUNTER:
+      assert(0);
+      return 0;
+   }
+   return 0;
+}
+
+static void amdgpu_read_registers(struct radeon_winsys *rws,
+                                  unsigned reg_offset,
+                                  unsigned num_registers, uint32_t *out)
+{
+   struct amdgpu_winsys *ws = (struct amdgpu_winsys*)rws;
+
+   amdgpu_read_mm_registers(ws->dev, reg_offset / 4, num_registers,
+                            0xffffffff, 0, out);
+}
+
+static unsigned hash_dev(void *key)
+{
+#if defined(PIPE_ARCH_X86_64)
+   return pointer_to_intptr(key) ^ (pointer_to_intptr(key) >> 32);
+#else
+   return pointer_to_intptr(key);
+#endif
+}
+
+static int compare_dev(void *key1, void *key2)
+{
+   return key1 != key2;
+}
+
+static bool amdgpu_winsys_unref(struct radeon_winsys *ws)
+{
+   struct amdgpu_winsys *rws = (struct amdgpu_winsys*)ws;
+   bool destroy;
+
+   /* When the reference counter drops to zero, remove the device pointer
+    * from the table.
+    * This must happen while the mutex is locked, so that
+    * amdgpu_winsys_create in another thread doesn't get the winsys
+    * from the table when the counter drops to 0. */
+   pipe_mutex_lock(dev_tab_mutex);
+
+   destroy = pipe_reference(&rws->reference, NULL);
+   if (destroy && dev_tab)
+      util_hash_table_remove(dev_tab, rws->dev);
+
+   pipe_mutex_unlock(dev_tab_mutex);
+   return destroy;
+}
+
+PUBLIC struct radeon_winsys *
+amdgpu_winsys_create(int fd, radeon_screen_create_t screen_create)
+{
+   struct amdgpu_winsys *ws;
+   drmVersionPtr version = drmGetVersion(fd);
+   amdgpu_device_handle dev;
+   uint32_t drm_major, drm_minor, r;
+
+   /* The DRM driver version of amdgpu is 3.x.x. */
+   if (version->version_major != 3) {
+      drmFreeVersion(version);
+      return NULL;
+   }
+   drmFreeVersion(version);
+
+   /* Look up the winsys from the dev table. */
+   pipe_mutex_lock(dev_tab_mutex);
+   if (!dev_tab)
+      dev_tab = util_hash_table_create(hash_dev, compare_dev);
+
+   /* Initialize the amdgpu device. This should always return the same pointer
+    * for the same fd. */
+   r = amdgpu_device_initialize(fd, &drm_major, &drm_minor, &dev);
+   if (r) {
+      pipe_mutex_unlock(dev_tab_mutex);
+      fprintf(stderr, "amdgpu: amdgpu_device_initialize failed.\n");
+      return NULL;
+   }
+
+   /* Lookup a winsys if we have already created one for this device. */
+   ws = util_hash_table_get(dev_tab, dev);
+   if (ws) {
+      pipe_reference(NULL, &ws->reference);
+      pipe_mutex_unlock(dev_tab_mutex);
+      return &ws->base;
+   }
+
+   /* Create a new winsys. */
+   ws = CALLOC_STRUCT(amdgpu_winsys);
+   if (!ws) {
+      pipe_mutex_unlock(dev_tab_mutex);
+      return NULL;
+   }
+
+   ws->dev = dev;
+   ws->info.drm_major = drm_major;
+   ws->info.drm_minor = drm_minor;
+
+   if (!do_winsys_init(ws))
+      goto fail;
+
+   /* Create managers. */
+   ws->kman = amdgpu_bomgr_create(ws);
+   if (!ws->kman)
+      goto fail;
+   ws->cman = pb_cache_manager_create(ws->kman, 500000, 2.0f, 0,
+			(ws->info.vram_size + ws->info.gart_size) / 8);
+   if (!ws->cman)
+      goto fail;
+
+   /* init reference */
+   pipe_reference_init(&ws->reference, 1);
+
+   /* Set functions. */
+   ws->base.unref = amdgpu_winsys_unref;
+   ws->base.destroy = amdgpu_winsys_destroy;
+   ws->base.query_info = amdgpu_winsys_query_info;
+   ws->base.cs_request_feature = amdgpu_cs_request_feature;
+   ws->base.query_value = amdgpu_query_value;
+   ws->base.read_registers = amdgpu_read_registers;
+
+   amdgpu_bomgr_init_functions(ws);
+   amdgpu_cs_init_functions(ws);
+
+   pipe_mutex_init(ws->bo_fence_lock);
+
+   /* Create the screen at the end. The winsys must be initialized
+    * completely.
+    *
+    * Alternatively, we could create the screen based on "ws->gen"
+    * and link all drivers into one binary blob. */
+   ws->base.screen = screen_create(&ws->base);
+   if (!ws->base.screen) {
+      amdgpu_winsys_destroy(&ws->base);
+      pipe_mutex_unlock(dev_tab_mutex);
+      return NULL;
+   }
+
+   util_hash_table_set(dev_tab, dev, ws);
+
+   /* We must unlock the mutex once the winsys is fully initialized, so that
+    * other threads attempting to create the winsys from the same fd will
+    * get a fully initialized winsys and not just half-way initialized. */
+   pipe_mutex_unlock(dev_tab_mutex);
+
+   return &ws->base;
+
+fail:
+   pipe_mutex_unlock(dev_tab_mutex);
+   if (ws->cman)
+      ws->cman->destroy(ws->cman);
+   if (ws->kman)
+      ws->kman->destroy(ws->kman);
+   FREE(ws);
+   return NULL;
+}

src/gallium/winsys/amdgpu/drm/amdgpu_winsys.h

@@ -0,0 +1,71 @@
+/*
+ * Copyright © 2009 Corbin Simpson
+ * Copyright © 2015 Advanced Micro Devices, Inc.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining
+ * a copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
+ * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS, AUTHORS
+ * AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ */
+/*
+ * Authors:
+ *      Marek Olšák <maraeo@gmail.com>
+ */
+
+#ifndef AMDGPU_WINSYS_H
+#define AMDGPU_WINSYS_H
+
+#include "gallium/drivers/radeon/radeon_winsys.h"
+#include "os/os_thread.h"
+#include <amdgpu.h>
+
+struct amdgpu_cs;
+
+struct amdgpu_winsys {
+   struct radeon_winsys base;
+   struct pipe_reference reference;
+
+   amdgpu_device_handle dev;
+
+   pipe_mutex bo_fence_lock;
+
+   int num_cs; /* The number of command streams created. */
+   uint32_t next_bo_unique_id;
+   uint64_t allocated_vram;
+   uint64_t allocated_gtt;
+   uint64_t buffer_wait_time; /* time spent in buffer_wait in ns */
+   uint64_t num_cs_flushes;
+   unsigned gart_page_size;
+
+   struct radeon_info info;
+
+   struct pb_manager *kman;
+   struct pb_manager *cman;
+
+   struct amdgpu_gpu_info amdinfo;
+};
+
+static inline struct amdgpu_winsys *
+amdgpu_winsys(struct radeon_winsys *base)
+{
+   return (struct amdgpu_winsys*)base;
+}
+
+#endif

src/gallium/winsys/radeon/drm/radeon_drm_cs.c

@@ -550,6 +550,7 @@ static void radeon_drm_cs_flush(struct radeon_winsys_cs *rcs,
 
         default:
         case RING_GFX:
+        case RING_COMPUTE:
             cs->cst->flags[0] = 0;
             cs->cst->flags[1] = RADEON_CS_RING_GFX;
             cs->cst->cs.num_chunks = 2;
@@ -565,7 +566,7 @@ static void radeon_drm_cs_flush(struct radeon_winsys_cs *rcs,
                 cs->cst->flags[0] |= RADEON_CS_END_OF_FRAME;
                 cs->cst->cs.num_chunks = 3;
             }
-            if (flags & RADEON_FLUSH_COMPUTE) {
+            if (cs->base.ring_type == RING_COMPUTE) {
                 cs->cst->flags[1] = RADEON_CS_RING_COMPUTE;
                 cs->cst->cs.num_chunks = 3;
             }