r300g: rework command submission and resource space checking

The motivation behind this rework is to get some speed by reducing
CPU overhead. The performance increase depends on many factors,
but it's measurable (I think it's about 10% increase in Torcs).

This commit replaces libdrm's radeon_cs_gem with our own implemention.
It's optimized specifically for r300g, but r600g could use it as well.
Reloc writes and space checking are faster and simpler than their
counterparts in libdrm (the time complexity of all the functions
is O(1) in nearly all scenarios, thanks to hashing).
(libdrm's radeon_bo_gem is still being used in the driver.)

It works like this:

cs_add_reloc(cs, buf, read_domain, write_domain) adds a new relocation and
also adds the size of 'buf' to the used_gart and used_vram winsys variables
based on the domains, which are simply or'd for the accounting purposes.
The adding is skipped if the reloc is already present in the list, but it
accounts any newly-referenced domains.

cs_validate is then called, which just checks:
    used_vram/gart < vram/gart_size * 0.8
The 0.8 number allows for some memory fragmentation. If the validation
fails, the pipe driver flushes CS and tries do the validation again,
i.e. it validates only that one operation. If it fails again, it drops
the operation on the floor and prints some nasty message to stderr.

cs_write_reloc(cs, buf) just writes a reloc that has been added using
cs_add_reloc. The read_domain and write_domain parameters have been removed,
because we already specify them in cs_add_reloc.

The space checking has been tested by putting small values in vram/gart_size
variables.

8 files changed, 409 insertions, 163 deletions

diff --git a/src/gallium/winsys/radeon/drm/Makefile b/src/gallium/winsys/radeon/drm/Makefile
index aa73edde34e..7e339a2ecfe 100644
--- a/src/gallium/winsys/radeon/drm/Makefile
+++ b/src/gallium/winsys/radeon/drm/Makefile
@@ -6,6 +6,7 @@ LIBNAME = radeonwinsys
 
 C_SOURCES = \
 	radeon_drm_buffer.c \
+	radeon_drm_cs.c \
 	radeon_drm_common.c \
 	radeon_r300.c
 
diff --git a/src/gallium/winsys/radeon/drm/SConscript b/src/gallium/winsys/radeon/drm/SConscript
index 2dbf61a7ba3..80816621848 100644
--- a/src/gallium/winsys/radeon/drm/SConscript
+++ b/src/gallium/winsys/radeon/drm/SConscript
@@ -4,6 +4,7 @@ env = env.Clone()
 
 radeon_sources = [
     'radeon_drm_buffer.c',
+    'radeon_drm_cs.c',
     'radeon_drm_common.c',
     'radeon_r300.c',
 ]
diff --git a/src/gallium/winsys/radeon/drm/radeon_drm_buffer.c b/src/gallium/winsys/radeon/drm/radeon_drm_buffer.c
index dc4b51747f4..5e14287ec2d 100644
--- a/src/gallium/winsys/radeon/drm/radeon_drm_buffer.c
+++ b/src/gallium/winsys/radeon/drm/radeon_drm_buffer.c
@@ -1,5 +1,5 @@
-#include "radeon_cs_gem.h"
 #include "radeon_drm_buffer.h"
+#include "radeon_drm_cs.h"
 
 #include "util/u_hash_table.h"
 #include "util/u_memory.h"
@@ -155,7 +155,7 @@ radeon_drm_buffer_map_internal(struct pb_buffer *_buf,
      * we cannot flush. */
     assert(cs || !radeon_bo_is_referenced_by_cs(buf->bo, NULL));
 
-    if (cs && radeon_bo_is_referenced_by_cs(buf->bo, cs->cs)) {
+    if (cs && radeon_bo_is_referenced_by_cs(buf->bo, NULL)) {
         cs->flush_cs(cs->flush_data);
     }
 
@@ -463,35 +463,6 @@ static void radeon_drm_buffer_set_tiling(struct r300_winsys_screen *ws,
     radeon_bo_set_tiling(buf->bo, flags, pitch);
 }
 
-static void radeon_drm_bufmgr_add_buffer(struct r300_winsys_cs *rcs,
-                                         struct r300_winsys_cs_buffer *_buf,
-                                         enum r300_buffer_domain rd,
-                                         enum r300_buffer_domain wd)
-{
-    struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
-    struct radeon_bo *bo = (struct radeon_bo*)_buf;
-
-    radeon_cs_space_add_persistent_bo(cs->cs, bo, rd, wd);
-}
-
-static void radeon_drm_bufmgr_write_reloc(struct r300_winsys_cs *rcs,
-                                          struct r300_winsys_cs_buffer *_buf,
-                                          enum r300_buffer_domain rd,
-                                          enum r300_buffer_domain wd)
-{
-    struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
-    struct radeon_bo *bo = (struct radeon_bo*)_buf;
-    int retval;
-
-    cs->cs->cdw = cs->base.cdw;
-    retval = radeon_cs_write_reloc(cs->cs, bo, rd, wd, 0);
-    cs->base.cdw = cs->cs->cdw;
-    if (retval) {
-        fprintf(stderr, "radeon: Relocation of %p (%d, %d, %d) failed!\n",
-                bo, rd, wd, 0);
-    }
-}
-
 static struct r300_winsys_cs_buffer *radeon_drm_get_cs_handle(
         struct r300_winsys_screen *rws,
         struct r300_winsys_buffer *_buf)
@@ -505,12 +476,11 @@ static boolean radeon_drm_is_buffer_referenced(struct r300_winsys_cs *rcs,
                                                struct r300_winsys_cs_buffer *_buf,
                                                enum r300_reference_domain domain)
 {
-    struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
     struct radeon_bo *bo = (struct radeon_bo*)_buf;
     uint32_t tmp;
 
     if (domain & R300_REF_CS) {
-        if (radeon_bo_is_referenced_by_cs(bo, cs->cs)) {
+        if (radeon_bo_is_referenced_by_cs(bo, NULL)) {
             return TRUE;
         }
     }
@@ -559,6 +529,4 @@ void radeon_drm_bufmgr_init_functions(struct radeon_drm_winsys *ws)
     ws->base.buffer_unmap = radeon_drm_buffer_unmap;
     ws->base.buffer_wait = radeon_drm_buffer_wait;
     ws->base.cs_is_buffer_referenced = radeon_drm_is_buffer_referenced;
-    ws->base.cs_add_buffer = radeon_drm_bufmgr_add_buffer;
-    ws->base.cs_write_reloc = radeon_drm_bufmgr_write_reloc;
 }
diff --git a/src/gallium/winsys/radeon/drm/radeon_drm_common.c b/src/gallium/winsys/radeon/drm/radeon_drm_common.c
index edd7cd3b7b8..fe71f080592 100644
--- a/src/gallium/winsys/radeon/drm/radeon_drm_common.c
+++ b/src/gallium/winsys/radeon/drm/radeon_drm_common.c
@@ -31,6 +31,7 @@
 
 #include "radeon_winsys.h"
 #include "radeon_drm_buffer.h"
+#include "radeon_drm_cs.h"
 #include "radeon_drm_public.h"
 
 #include "pipebuffer/pb_bufmgr.h"
@@ -185,7 +186,6 @@ static void radeon_winsys_destroy(struct r300_winsys_screen *rws)
     ws->kman->destroy(ws->kman);
 
     radeon_bo_manager_gem_dtor(ws->bom);
-    radeon_cs_manager_gem_dtor(ws->csm);
     FREE(rws);
 }
 
@@ -207,9 +207,6 @@ struct r300_winsys_screen *r300_drm_winsys_screen_create(int fd)
     ws->bom = radeon_bo_manager_gem_ctor(fd);
     if (!ws->bom)
 	goto fail;
-    ws->csm = radeon_cs_manager_gem_ctor(fd);
-    if (!ws->csm)
-	goto fail;
     ws->kman = radeon_drm_bufmgr_create(ws);
     if (!ws->kman)
 	goto fail;
@@ -221,6 +218,7 @@ struct r300_winsys_screen *r300_drm_winsys_screen_create(int fd)
     ws->base.destroy = radeon_winsys_destroy;
 
     radeon_drm_bufmgr_init_functions(ws);
+    radeon_drm_cs_init_functions(ws);
     radeon_winsys_init_functions(ws);
 
     return &ws->base;
@@ -228,8 +226,6 @@ struct r300_winsys_screen *r300_drm_winsys_screen_create(int fd)
 fail:
     if (ws->bom)
 	radeon_bo_manager_gem_dtor(ws->bom);
-    if (ws->csm)
-	radeon_cs_manager_gem_dtor(ws->csm);
 
     if (ws->cman)
 	ws->cman->destroy(ws->cman);
diff --git a/src/gallium/winsys/radeon/drm/radeon_drm_cs.c b/src/gallium/winsys/radeon/drm/radeon_drm_cs.c
new file mode 100644
index 00000000000..60bc36b0929
--- /dev/null
+++ b/src/gallium/winsys/radeon/drm/radeon_drm_cs.c
@@ -0,0 +1,360 @@
+/*
+ * Copyright © 2008 Jérôme Glisse
+ * Copyright © 2010 Marek Olšák <[email protected]>
+ * 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 <[email protected]>
+ *
+ * Based on work from libdrm_radeon by:
+ *      Aapo Tahkola <[email protected]>
+ *      Nicolai Haehnle <[email protected]>
+ *      Jérôme Glisse <[email protected]>
+ */
+
+/*
+    This file replaces libdrm's radeon_cs_gem with our own implemention.
+    It's optimized specifically for r300g, but r600g could use it as well.
+    Reloc writes and space checking are faster and simpler than their
+    counterparts in libdrm (the time complexity of all the functions
+    is O(1) in nearly all scenarios, thanks to hashing).
+
+    It works like this:
+
+    cs_add_reloc(cs, buf, read_domain, write_domain) adds a new relocation and
+    also adds the size of 'buf' to the used_gart and used_vram winsys variables
+    based on the domains, which are simply or'd for the accounting purposes.
+    The adding is skipped if the reloc is already present in the list, but it
+    accounts any newly-referenced domains.
+
+    cs_validate is then called, which just checks:
+        used_vram/gart < vram/gart_size * 0.8
+    The 0.8 number allows for some memory fragmentation. If the validation
+    fails, the pipe driver flushes CS and tries do the validation again,
+    i.e. it validates only that one operation. If it fails again, it drops
+    the operation on the floor and prints some nasty message to stderr.
+    (done in the pipe driver)
+
+    cs_write_reloc(cs, buf) just writes a reloc that has been added using
+    cs_add_reloc. The read_domain and write_domain parameters have been removed,
+    because we already specify them in cs_add_reloc.
+*/
+
+#include "radeon_drm_cs.h"
+#include "radeon_drm_buffer.h"
+
+#include "util/u_memory.h"
+
+#include <stdlib.h>
+#include <stdint.h>
+#include <radeon_bo.h>
+#include <xf86drm.h>
+
+#define RELOC_DWORDS (sizeof(struct drm_radeon_cs_reloc) / sizeof(uint32_t))
+
+static struct r300_winsys_cs *radeon_drm_cs_create(struct r300_winsys_screen *rws)
+{
+    struct radeon_drm_winsys *ws = radeon_drm_winsys(rws);
+    struct radeon_drm_cs *cs;
+
+    cs = CALLOC_STRUCT(radeon_drm_cs);
+    if (!cs) {
+        return NULL;
+    }
+
+    cs->ws = ws;
+    cs->nrelocs = 256;
+    cs->relocs_bo = (struct radeon_bo**)
+                     CALLOC(1, cs->nrelocs * sizeof(struct radeon_bo*));
+    if (!cs->relocs_bo) {
+        FREE(cs);
+        return NULL;
+    }
+
+    cs->relocs = (struct drm_radeon_cs_reloc*)
+                 CALLOC(1, cs->nrelocs * sizeof(struct drm_radeon_cs_reloc));
+    if (!cs->relocs) {
+        FREE(cs->relocs_bo);
+        FREE(cs);
+        return NULL;
+    }
+
+    cs->chunks[0].chunk_id = RADEON_CHUNK_ID_IB;
+    cs->chunks[0].length_dw = 0;
+    cs->chunks[0].chunk_data = (uint64_t)(uintptr_t)cs->base.buf;
+    cs->chunks[1].chunk_id = RADEON_CHUNK_ID_RELOCS;
+    cs->chunks[1].length_dw = 0;
+    cs->chunks[1].chunk_data = (uint64_t)(uintptr_t)cs->relocs;
+    return &cs->base;
+}
+
+#define OUT_CS(cs, value) (cs)->buf[(cs)->cdw++] = (value)
+
+static inline void update_domains(struct drm_radeon_cs_reloc *reloc,
+                                  enum r300_buffer_domain rd,
+                                  enum r300_buffer_domain wd,
+                                  enum r300_buffer_domain *added_domains)
+{
+    *added_domains = (rd | wd) & ~(reloc->read_domains | reloc->write_domain);
+
+    if (reloc->read_domains & wd) {
+        reloc->read_domains = rd;
+        reloc->write_domain = wd;
+    } else if (rd & reloc->write_domain) {
+        reloc->read_domains = rd;
+        reloc->write_domain |= wd;
+    } else {
+        reloc->read_domains |= rd;
+        reloc->write_domain |= wd;
+    }
+}
+
+static int radeon_get_reloc(struct radeon_drm_cs *cs,
+                            struct radeon_bo *bo)
+{
+    struct drm_radeon_cs_reloc *reloc;
+    unsigned i;
+    unsigned hash = bo->handle & (sizeof(cs->is_handle_added)-1);
+
+    if (cs->is_handle_added[hash]) {
+        reloc = cs->relocs_hashlist[hash];
+        if (reloc->handle == bo->handle) {
+            return cs->reloc_indices_hashlist[hash];
+        }
+
+        /* Hash collision, look for the BO in the list of relocs linearly. */
+        for (i = cs->crelocs; i != 0;) {
+            --i;
+            reloc = &cs->relocs[i];
+            if (reloc->handle == bo->handle) {
+                /* Put this reloc in the hash list.
+                 * This will prevent additional hash collisions if there are
+                 * several subsequent get_reloc calls of 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->relocs_hashlist[hash] = reloc;
+                cs->reloc_indices_hashlist[hash] = i;
+                /*printf("write_reloc collision, hash: %i, handle: %i\n", hash, bo->handle);*/
+                return i;
+            }
+        }
+    }
+
+    return -1;
+}
+
+static void radeon_add_reloc(struct radeon_drm_cs *cs,
+                             struct radeon_bo *bo,
+                             enum r300_buffer_domain rd,
+                             enum r300_buffer_domain wd,
+                             enum r300_buffer_domain *added_domains)
+{
+    struct drm_radeon_cs_reloc *reloc;
+    unsigned i;
+    unsigned hash = bo->handle & (sizeof(cs->is_handle_added)-1);
+
+    if (cs->is_handle_added[hash]) {
+        reloc = cs->relocs_hashlist[hash];
+        if (reloc->handle == bo->handle) {
+            update_domains(reloc, rd, wd, added_domains);
+            return;
+        }
+
+        /* Hash collision, look for the BO in the list of relocs linearly. */
+        for (i = cs->crelocs; i != 0;) {
+            --i;
+            reloc = &cs->relocs[i];
+            if (reloc->handle == bo->handle) {
+                update_domains(reloc, rd, wd, added_domains);
+
+                cs->relocs_hashlist[hash] = reloc;
+                cs->reloc_indices_hashlist[hash] = i;
+                /*printf("write_reloc collision, hash: %i, handle: %i\n", hash, bo->handle);*/
+                return;
+            }
+        }
+    }
+
+    /* New relocation, check if the backing array is large enough. */
+    if (cs->crelocs >= cs->nrelocs) {
+        uint32_t size;
+        cs->nrelocs += 10;
+
+        size = cs->nrelocs * sizeof(struct radeon_bo*);
+        cs->relocs_bo = (struct radeon_bo**)realloc(cs->relocs_bo, size);
+
+        size = cs->nrelocs * sizeof(struct drm_radeon_cs_reloc);
+        cs->relocs = (struct drm_radeon_cs_reloc*)realloc(cs->relocs, size);
+
+        cs->chunks[1].chunk_data = (uint64_t)(uintptr_t)cs->relocs;
+    }
+
+    /* Initialize the new relocation. */
+    radeon_bo_ref(bo);
+    cs->relocs_bo[cs->crelocs] = bo;
+    reloc = &cs->relocs[cs->crelocs];
+    reloc->handle = bo->handle;
+    reloc->read_domains = rd;
+    reloc->write_domain = wd;
+    reloc->flags = 0;
+
+    cs->is_handle_added[hash] = TRUE;
+    cs->relocs_hashlist[hash] = reloc;
+    cs->reloc_indices_hashlist[hash] = cs->crelocs;
+
+    cs->chunks[1].length_dw += RELOC_DWORDS;
+    cs->crelocs++;
+
+    *added_domains = rd | wd;
+}
+
+static void radeon_drm_cs_add_reloc(struct r300_winsys_cs *rcs,
+                                    struct r300_winsys_cs_buffer *buf,
+                                    enum r300_buffer_domain rd,
+                                    enum r300_buffer_domain wd)
+{
+    struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
+    struct radeon_bo *bo = (struct radeon_bo*)buf;
+    enum r300_buffer_domain added_domains;
+
+    radeon_add_reloc(cs, bo, rd, wd, &added_domains);
+
+    if (!added_domains)
+        return;
+
+    if (added_domains & R300_DOMAIN_GTT)
+        cs->used_gart += bo->size;
+    if (added_domains & R300_DOMAIN_VRAM)
+        cs->used_vram += bo->size;
+}
+
+static boolean radeon_drm_cs_validate(struct r300_winsys_cs *rcs)
+{
+    struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
+
+    return cs->used_gart < cs->ws->gart_size * 0.8 &&
+           cs->used_vram < cs->ws->vram_size * 0.8;
+}
+
+static void radeon_drm_cs_write_reloc(struct r300_winsys_cs *rcs,
+                                      struct r300_winsys_cs_buffer *buf)
+{
+    struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
+    struct radeon_bo *bo = (struct radeon_bo*)buf;
+
+    unsigned index = radeon_get_reloc(cs, bo);
+
+    if (index == -1) {
+        fprintf(stderr, "r300: Cannot get a relocation in %s.\n", __func__);
+        return;
+    }
+
+    OUT_CS(&cs->base, 0xc0001000);
+    OUT_CS(&cs->base, index * RELOC_DWORDS);
+}
+
+static void radeon_drm_cs_emit(struct r300_winsys_cs *rcs)
+{
+    struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
+    uint64_t chunk_array[2];
+    unsigned i;
+    int r;
+
+    if (cs->base.cdw) {
+        /* Unmap buffers. */
+        radeon_drm_bufmgr_flush_maps(cs->ws->kman);
+
+        /* Prepare the arguments. */
+        cs->chunks[0].length_dw = cs->base.cdw;
+
+        chunk_array[0] = (uint64_t)(uintptr_t)&cs->chunks[0];
+        chunk_array[1] = (uint64_t)(uintptr_t)&cs->chunks[1];
+
+        cs->cs.num_chunks = 2;
+        cs->cs.chunks = (uint64_t)(uintptr_t)chunk_array;
+
+        /* Emit. */
+        r = drmCommandWriteRead(cs->ws->fd, DRM_RADEON_CS,
+                                &cs->cs, sizeof(struct drm_radeon_cs));
+        if (r) {
+            if (debug_get_bool_option("RADEON_DUMP_CS", FALSE)) {
+                fprintf(stderr, "radeon: The kernel rejected CS, dumping...\n");
+                fprintf(stderr, "VENDORID:DEVICEID 0x%04X:0x%04X\n", 0x1002,
+                        cs->ws->pci_id);
+                for (i = 0; i < cs->base.cdw; i++) {
+                    fprintf(stderr, "0x%08X\n", cs->base.buf[i]);
+                }
+            } else {
+                fprintf(stderr, "radeon: The kernel rejected CS, "
+                                "see dmesg for more information.\n");
+            }
+        }
+    }
+
+    /* Unreference buffers, cleanup. */
+    for (i = 0; i < cs->crelocs; i++) {
+        radeon_bo_unref((struct radeon_bo*)cs->relocs_bo[i]);
+        cs->relocs_bo[i] = NULL;
+    }
+
+    cs->base.cdw = 0;
+    cs->crelocs = 0;
+    cs->chunks[0].length_dw = 0;
+    cs->chunks[1].length_dw = 0;
+    cs->used_gart = 0;
+    cs->used_vram = 0;
+    memset(cs->is_handle_added, 0, sizeof(cs->is_handle_added));
+}
+
+static void radeon_drm_cs_destroy(struct r300_winsys_cs *rcs)
+{
+    struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
+    FREE(cs->relocs_bo);
+    FREE(cs->relocs);
+    FREE(cs);
+}
+
+static void radeon_drm_cs_set_flush(struct r300_winsys_cs *rcs,
+                             void (*flush)(void *), void *user)
+{
+    struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
+    cs->flush_cs = flush;
+    cs->flush_data = user;
+}
+
+void radeon_drm_cs_init_functions(struct radeon_drm_winsys *ws)
+{
+    ws->base.cs_create = radeon_drm_cs_create;
+    ws->base.cs_destroy = radeon_drm_cs_destroy;
+    ws->base.cs_add_reloc = radeon_drm_cs_add_reloc;
+    ws->base.cs_validate = radeon_drm_cs_validate;
+    ws->base.cs_write_reloc = radeon_drm_cs_write_reloc;
+    ws->base.cs_flush = radeon_drm_cs_emit;
+    ws->base.cs_set_flush = radeon_drm_cs_set_flush;
+}
diff --git a/src/gallium/winsys/radeon/drm/radeon_drm_cs.h b/src/gallium/winsys/radeon/drm/radeon_drm_cs.h
new file mode 100644
index 00000000000..76046534b65
--- /dev/null
+++ b/src/gallium/winsys/radeon/drm/radeon_drm_cs.h
@@ -0,0 +1,42 @@
+#ifndef RADEON_DRM_CS_H
+#define RADEON_DRM_CS_H
+
+#include "radeon_winsys.h"
+#include <radeon_drm.h>
+
+struct radeon_drm_cs {
+    struct r300_winsys_cs base;
+
+    /* The winsys. */
+    struct radeon_drm_winsys *ws;
+
+    /* Flush CS. */
+    void (*flush_cs)(void *);
+    void *flush_data;
+
+    /* Relocs. */
+    unsigned                    crelocs;
+    unsigned                    nrelocs;
+    struct drm_radeon_cs_reloc  *relocs;
+    struct radeon_bo            **relocs_bo;
+    struct drm_radeon_cs        cs;
+    struct drm_radeon_cs_chunk  chunks[2];
+
+    unsigned used_vram;
+    unsigned used_gart;
+
+    /* 0 = BO not added, 1 = BO added */
+    char                        is_handle_added[256];
+    struct drm_radeon_cs_reloc  *relocs_hashlist[256];
+    unsigned                    reloc_indices_hashlist[256];
+};
+
+static INLINE struct radeon_drm_cs *
+radeon_drm_cs(struct r300_winsys_cs *base)
+{
+    return (struct radeon_drm_cs*)base;
+}
+
+void radeon_drm_cs_init_functions(struct radeon_drm_winsys *ws);
+
+#endif
diff --git a/src/gallium/winsys/radeon/drm/radeon_r300.c b/src/gallium/winsys/radeon/drm/radeon_r300.c
index ae7020a0639..bacf181b47c 100644
--- a/src/gallium/winsys/radeon/drm/radeon_r300.c
+++ b/src/gallium/winsys/radeon/drm/radeon_r300.c
@@ -25,7 +25,6 @@
 #include "util/u_memory.h"
 #include "pipebuffer/pb_bufmgr.h"
 
-#include "radeon_cs_gem.h"
 #include "state_tracker/drm_driver.h"
 
 static unsigned get_pb_usage_from_create_flags(unsigned bind, unsigned usage,
@@ -134,65 +133,6 @@ static boolean radeon_r300_winsys_buffer_get_handle(struct r300_winsys_screen *r
     return radeon_drm_bufmgr_get_handle(_buf, whandle);
 }
 
-static void radeon_r300_winsys_cs_set_flush(struct r300_winsys_cs *rcs,
-                                            void (*flush)(void *),
-                                            void *user)
-{
-    struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
-    cs->flush_cs = flush;
-    cs->flush_data = user;
-    radeon_cs_space_set_flush(cs->cs, flush, user);
-}
-
-static boolean radeon_r300_winsys_cs_validate(struct r300_winsys_cs *rcs)
-{
-    struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
-
-    return radeon_cs_space_check(cs->cs) >= 0;
-}
-
-static void radeon_r300_winsys_cs_reset_buffers(struct r300_winsys_cs *rcs)
-{
-    struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
-    radeon_cs_space_reset_bos(cs->cs);
-}
-
-static void radeon_r300_winsys_cs_flush(struct r300_winsys_cs *rcs)
-{
-    struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
-    int retval;
-
-    /* Don't flush a zero-sized CS. */
-    if (!cs->base.cdw) {
-        return;
-    }
-
-    cs->cs->cdw = cs->base.cdw;
-
-    radeon_drm_bufmgr_flush_maps(cs->ws->kman);
-
-    /* Emit the CS. */
-    retval = radeon_cs_emit(cs->cs);
-    if (retval) {
-        if (debug_get_bool_option("RADEON_DUMP_CS", FALSE)) {
-            fprintf(stderr, "radeon: The kernel rejected CS, dumping...\n");
-            radeon_cs_print(cs->cs, stderr);
-        } else {
-            fprintf(stderr, "radeon: The kernel rejected CS, "
-                            "see dmesg for more information.\n");
-        }
-    }
-
-    /* Reset CS.
-     * Someday, when we care about performance, we should really find a way
-     * to rotate between two or three CS objects so that the GPU can be
-     * spinning through one CS while another one is being filled. */
-    radeon_cs_erase(cs->cs);
-
-    cs->base.buf = cs->cs->packets;
-    cs->base.cdw = cs->cs->cdw;
-}
-
 static uint32_t radeon_get_value(struct r300_winsys_screen *rws,
                                  enum r300_value_id id)
 {
@@ -219,39 +159,6 @@ static uint32_t radeon_get_value(struct r300_winsys_screen *rws,
     return 0;
 }
 
-static struct r300_winsys_cs *radeon_r300_winsys_cs_create(struct r300_winsys_screen *rws)
-{
-    struct radeon_drm_winsys *ws = radeon_drm_winsys(rws);
-    struct radeon_drm_cs *cs = CALLOC_STRUCT(radeon_drm_cs);
-
-    if (!cs)
-        return NULL;
-
-    /* Size limit on IBs is 64 kibibytes. */
-    cs->cs = radeon_cs_create(ws->csm, 1024 * 64 / 4);
-    if (!cs->cs) {
-        FREE(cs);
-        return NULL;
-    }
-
-    radeon_cs_set_limit(cs->cs,
-            RADEON_GEM_DOMAIN_GTT, ws->gart_size);
-    radeon_cs_set_limit(cs->cs,
-            RADEON_GEM_DOMAIN_VRAM, ws->vram_size);
-
-    cs->ws = ws;
-    cs->base.buf = cs->cs->packets;
-    cs->base.cdw = cs->cs->cdw;
-    return &cs->base;
-}
-
-static void radeon_r300_winsys_cs_destroy(struct r300_winsys_cs *rcs)
-{
-    struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
-    radeon_cs_destroy(cs->cs);
-    FREE(cs);
-}
-
 void radeon_winsys_init_functions(struct radeon_drm_winsys *ws)
 {
     ws->base.get_value = radeon_get_value;
@@ -259,10 +166,4 @@ void radeon_winsys_init_functions(struct radeon_drm_winsys *ws)
     ws->base.buffer_reference = radeon_r300_winsys_buffer_reference;
     ws->base.buffer_from_handle = radeon_r300_winsys_buffer_from_handle;
     ws->base.buffer_get_handle = radeon_r300_winsys_buffer_get_handle;
-    ws->base.cs_create = radeon_r300_winsys_cs_create;
-    ws->base.cs_destroy = radeon_r300_winsys_cs_destroy;
-    ws->base.cs_validate = radeon_r300_winsys_cs_validate;
-    ws->base.cs_flush = radeon_r300_winsys_cs_flush;
-    ws->base.cs_reset_buffers = radeon_r300_winsys_cs_reset_buffers;
-    ws->base.cs_set_flush = radeon_r300_winsys_cs_set_flush;
 }
diff --git a/src/gallium/winsys/radeon/drm/radeon_winsys.h b/src/gallium/winsys/radeon/drm/radeon_winsys.h
index 6f232143f6a..492edfef8c3 100644
--- a/src/gallium/winsys/radeon/drm/radeon_winsys.h
+++ b/src/gallium/winsys/radeon/drm/radeon_winsys.h
@@ -55,31 +55,8 @@ struct radeon_drm_winsys {
     boolean drm_2_8_0;
     /* Hyper-Z user */
     boolean hyperz;
-
-    /* Radeon CS manager. */
-    struct radeon_cs_manager *csm;
 };
 
-struct radeon_drm_cs {
-    struct r300_winsys_cs base;
-
-    /* The winsys. */
-    struct radeon_drm_winsys *ws;
-
-    /* The libdrm command stream. */
-    struct radeon_cs *cs;
-
-    /* Flush CS. */
-    void (*flush_cs)(void *);
-    void *flush_data;
-};
-
-static INLINE struct radeon_drm_cs *
-radeon_drm_cs(struct r300_winsys_cs *base)
-{
-    return (struct radeon_drm_cs*)base;
-}
-
 static INLINE struct radeon_drm_winsys *
 radeon_drm_winsys(struct r300_winsys_screen *base)
 {