libdrm_radeon: Optimize cs_gem_reloc to do less looping.

bo->referenced_in_cs is checked if bo is already in cs. Adding and removing
reference in bo is done with atomic operations to allow parallel access to a
bo from multiple contexts.

cs->id generation code quarentees there is not duplicated ids which limits
number of cs->ids to 32. If there is more cs objects rest will get id 0.

V2:
 - Fix configure to check for atomics operations if libdrm_radeon is only selected.
 - Make atomic operations private to libdrm.

This optimization decreases cs_write_reloc share of torcs profiling from 4.3%
to 2.6%.

Tested-by: Michel Dänzer <michel@daenzer.net>
Signed-off-by: Pauli Nieminen <suokkos@gmail.com>

8 files changed, 138 insertions, 41 deletions

diff --git a/configure.ac b/configure.ac
index 953a7586..b044c966 100644
--- a/configure.ac
+++ b/configure.ac
@@ -56,8 +56,8 @@ AC_ARG_ENABLE(intel,
 
 AC_ARG_ENABLE(radeon,
 	      AS_HELP_STRING([--disable-radeon],
-	      [Enable support for radeon's KMS API (default: enabled)]),
-	      [RADEON=$enableval], [RADEON=yes])
+	      [Enable support for radeon's KMS API (default: auto)]),
+	      [RADEON=$enableval], [RADEON=auto])
 
 AC_ARG_ENABLE(vmwgfx-experimental-api,
 	      AS_HELP_STRING([--enable-vmwgfx-experimental-api],
@@ -173,7 +173,7 @@ if test "x$HAVE_LIBUDEV" = xyes; then
 fi
 AM_CONDITIONAL(HAVE_LIBUDEV, [test "x$HAVE_LIBUDEV" = xyes])
 
-if test "x$INTEL" != "xno"; then
+if test "x$INTEL" != "xno" -o "x$RADEON" != "xno"; then
     # Check for atomic intrinsics
     AC_CACHE_CHECK([for native atomic primitives], drm_cv_atomic_primitives,
     [
@@ -206,13 +206,26 @@ if test "x$INTEL" != "xno"; then
     fi
 
     if test "x$drm_cv_atomic_primitives" = "xnone"; then
-	    if test "x$INTEL" != "xauto"; then
-		    AC_MSG_ERROR([libdrm_intel depends upon atomic operations, which were not found for your compiler/cpu. Try compiling with -march=native, or install the libatomics-op-dev package, or, failing both of those, disable support for Intel GPUs by passing --disable-intel to ./configure])
-	   else
-		    INTEL=no
-	   fi
+           if test "x$INTEL" != "xauto"; then
+                   AC_MSG_ERROR([libdrm_intel depends upon atomic operations, which were not found for your compiler/cpu. Try compiling with -march=native, or install the libatomics-op-dev package, or, failing both of those, disable support for Intel GPUs by passing --disable-intel to ./configure])
+           else
+                   AC_MSG_WARN([Disabling libdrm_intel. It depends on atomic operations, which were not found for your compiler/cpu. Try compiling with -march=native, or install the libatomics-op-dev package.])
+                   INTEL=no
+           fi
+           if test "x$RADEON" != "xauto"; then
+                   AC_MSG_ERROR([libdrm_radeon depends upon atomic operations, which were not found for your compiler/cpu. Try compiling with -march=native, or install the libatomics-op-dev package, or, failing both of those, disable support for Radeon support by passing --disable-radeon to ./configure])
+           else
+                   AC_MSG_WARN([Disabling libdrm_radeon. It depends on atomic operations, which were not found for your compiler/cpu. Try compiling with -march=native, or install the libatomics-op-dev package.])
+                   RADEON=no
+           fi
+
     else
-	   INTEL=yes
+           if test "x$INTEL" != "xno"; then
+                   INTEL=yes
+           fi
+           if test "x$RADEON" != "xno"; then
+                   RADEON=yes
+           fi
     fi
 fi
 
diff --git a/radeon/radeon_bo_gem.c b/radeon/radeon_bo_gem.c
index bc8058d8..081ccb9f 100644
--- a/radeon/radeon_bo_gem.c
+++ b/radeon/radeon_bo_gem.c
@@ -39,6 +39,7 @@
 #include <sys/mman.h>
 #include <errno.h>
 #include "xf86drm.h"
+#include "xf86atomic.h"
 #include "drm.h"
 #include "radeon_drm.h"
 #include "radeon_bo.h"
@@ -49,6 +50,7 @@ struct radeon_bo_gem {
     struct radeon_bo_int base;
     uint32_t            name;
     int                 map_count;
+    atomic_t            reloc_in_cs;
     void *priv_ptr;
 };
 
@@ -80,6 +82,7 @@ static struct radeon_bo *bo_open(struct radeon_bo_manager *bom,
     bo->base.domains = domains;
     bo->base.flags = flags;
     bo->base.ptr = NULL;
+    atomic_set(&bo->reloc_in_cs, 0);
     bo->map_count = 0;
     if (handle) {
         struct drm_gem_open open_arg;
@@ -309,6 +312,12 @@ uint32_t radeon_gem_name_bo(struct radeon_bo *bo)
     return bo_gem->name;
 }
 
+void *radeon_gem_get_reloc_in_cs(struct radeon_bo *bo)
+{
+    struct radeon_bo_gem *bo_gem = (struct radeon_bo_gem*)bo;
+    return &bo_gem->reloc_in_cs;
+}
+
 int radeon_gem_get_kernel_name(struct radeon_bo *bo, uint32_t *name)
 {
     struct radeon_bo_int *boi = (struct radeon_bo_int *)bo;
diff --git a/radeon/radeon_bo_gem.h b/radeon/radeon_bo_gem.h
index c56c58e9..0af8610b 100644
--- a/radeon/radeon_bo_gem.h
+++ b/radeon/radeon_bo_gem.h
@@ -38,6 +38,7 @@ struct radeon_bo_manager *radeon_bo_manager_gem_ctor(int fd);
 void radeon_bo_manager_gem_dtor(struct radeon_bo_manager *bom);
 
 uint32_t radeon_gem_name_bo(struct radeon_bo *bo);
+void *radeon_gem_get_reloc_in_cs(struct radeon_bo *bo);
 int radeon_gem_set_domain(struct radeon_bo *bo, uint32_t read_domains, uint32_t write_domain);
 int radeon_gem_get_kernel_name(struct radeon_bo *bo, uint32_t *name);
 #endif
diff --git a/radeon/radeon_cs.c b/radeon/radeon_cs.c
index cc9be398..d0e922be 100644
--- a/radeon/radeon_cs.c
+++ b/radeon/radeon_cs.c
@@ -88,3 +88,9 @@ void radeon_cs_space_set_flush(struct radeon_cs *cs, void (*fn)(void *), void *d
     csi->space_flush_fn = fn;
     csi->space_flush_data = data;
 }
+
+uint32_t radeon_cs_get_id(struct radeon_cs *cs)
+{
+    struct radeon_cs_int *csi = (struct radeon_cs_int *)cs;
+    return csi->id;
+}
diff --git a/radeon/radeon_cs.h b/radeon/radeon_cs.h
index 49d5d9a6..7f6ee68b 100644
--- a/radeon/radeon_cs.h
+++ b/radeon/radeon_cs.h
@@ -85,7 +85,7 @@ extern int radeon_cs_write_reloc(struct radeon_cs *cs,
                                  uint32_t read_domain,
                                  uint32_t write_domain,
                                  uint32_t flags);
-
+extern uint32_t radeon_cs_get_id(struct radeon_cs *cs);
 /*
  * add a persistent BO to the list
  * a persistent BO is one that will be referenced across flushes,
diff --git a/radeon/radeon_cs_gem.c b/radeon/radeon_cs_gem.c
index 45a219c3..28ef5f64 100644
--- a/radeon/radeon_cs_gem.c
+++ b/radeon/radeon_cs_gem.c
@@ -32,6 +32,7 @@
 #include <assert.h>
 #include <errno.h>
 #include <stdlib.h>
+#include <pthread.h>
 #include <sys/mman.h>
 #include <sys/ioctl.h>
 #include "radeon_cs.h"
@@ -41,6 +42,7 @@
 #include "radeon_bo_gem.h"
 #include "drm.h"
 #include "xf86drm.h"
+#include "xf86atomic.h"
 #include "radeon_drm.h"
 
 struct radeon_cs_manager_gem {
@@ -68,6 +70,50 @@ struct cs_gem {
     struct radeon_bo_int        **relocs_bo;
 };
 
+static pthread_mutex_t id_mutex = PTHREAD_MUTEX_INITIALIZER;
+static uint32_t cs_id_source = 0;
+
+/**
+ * result is undefined if called with ~0
+ */
+static uint32_t get_first_zero(const uint32_t n)
+{
+    /* __builtin_ctz returns number of trailing zeros. */
+    return 1 << __builtin_ctz(~n);
+}
+
+/**
+ * Returns a free id for cs.
+ * If there is no free id we return zero
+ **/
+static uint32_t generate_id(void)
+{
+    uint32_t r = 0;
+    pthread_mutex_lock( &id_mutex );
+    /* check for free ids */
+    if (cs_id_source != ~r) {
+        /* find first zero bit */
+        r = get_first_zero(cs_id_source);
+
+        /* set id as reserved */
+        cs_id_source |= r;
+    }
+    pthread_mutex_unlock( &id_mutex );
+    return r;
+}
+
+/**
+ * Free the id for later reuse
+ **/
+static void free_id(uint32_t id)
+{
+    pthread_mutex_lock( &id_mutex );
+
+    cs_id_source &= ~id;
+
+    pthread_mutex_unlock( &id_mutex );
+}
+
 static struct radeon_cs_int *cs_gem_create(struct radeon_cs_manager *csm,
                                        uint32_t ndw)
 {
@@ -90,6 +136,7 @@ static struct radeon_cs_int *cs_gem_create(struct radeon_cs_manager *csm,
     }
     csg->base.relocs_total_size = 0;
     csg->base.crelocs = 0;
+    csg->base.id = generate_id();
     csg->nrelocs = 4096 / (4 * 4) ;
     csg->relocs_bo = (struct radeon_bo_int**)calloc(1,
                                                 csg->nrelocs*sizeof(void*));
@@ -141,38 +188,45 @@ static int cs_gem_write_reloc(struct radeon_cs_int *cs,
     if (write_domain == RADEON_GEM_DOMAIN_CPU) {
         return -EINVAL;
     }
-    /* check if bo is already referenced */
-    for(i = 0; i < cs->crelocs; i++) {
-        idx = i * RELOC_SIZE;
-        reloc = (struct cs_reloc_gem*)&csg->relocs[idx];
-        if (reloc->handle == bo->handle) {
-            /* Check domains must be in read or write. As we check already
-             * checked that in argument one of the read or write domain was
-             * set we only need to check that if previous reloc as the read
-             * domain set then the read_domain should also be set for this
-             * new relocation.
-             */
-            /* the DDX expects to read and write from same pixmap */
-            if (write_domain && (reloc->read_domain & write_domain)) {
-                reloc->read_domain = 0;
-                reloc->write_domain = write_domain;
-            } else if (read_domain & reloc->write_domain) {
-                reloc->read_domain = 0;
-            } else {
-                if (write_domain != reloc->write_domain)
-                    return -EINVAL;
-                if (read_domain != reloc->read_domain)
-                    return -EINVAL;
+    /* use bit field hash function to determine
+       if this bo is for sure not in this cs.*/
+    if ((atomic_read((atomic_t *)radeon_gem_get_reloc_in_cs(bo)) & cs->id)) {
+        /* check if bo is already referenced.
+	 * Scanning from end to begin reduces cycles with mesa because
+	 * it often relocates same shared dma bo again. */
+        for(i = cs->crelocs; i != 0;) {
+            --i;
+            idx = i * RELOC_SIZE;
+            reloc = (struct cs_reloc_gem*)&csg->relocs[idx];
+            if (reloc->handle == bo->handle) {
+                /* Check domains must be in read or write. As we check already
+                 * checked that in argument one of the read or write domain was
+                 * set we only need to check that if previous reloc as the read
+                 * domain set then the read_domain should also be set for this
+                 * new relocation.
+                 */
+                /* the DDX expects to read and write from same pixmap */
+                if (write_domain && (reloc->read_domain & write_domain)) {
+                    reloc->read_domain = 0;
+                    reloc->write_domain = write_domain;
+                } else if (read_domain & reloc->write_domain) {
+                    reloc->read_domain = 0;
+                } else {
+                    if (write_domain != reloc->write_domain)
+                        return -EINVAL;
+                    if (read_domain != reloc->read_domain)
+                        return -EINVAL;
+                }
+
+                reloc->read_domain |= read_domain;
+                reloc->write_domain |= write_domain;
+                /* update flags */
+                reloc->flags |= (flags & reloc->flags);
+                /* write relocation packet */
+                radeon_cs_write_dword((struct radeon_cs *)cs, 0xc0001000);
+                radeon_cs_write_dword((struct radeon_cs *)cs, idx);
+                return 0;
             }
-
-            reloc->read_domain |= read_domain;
-            reloc->write_domain |= write_domain;
-            /* update flags */
-            reloc->flags |= (flags & reloc->flags);
-            /* write relocation packet */
-            radeon_cs_write_dword((struct radeon_cs *)cs, 0xc0001000);
-            radeon_cs_write_dword((struct radeon_cs *)cs, idx);
-            return 0;
         }
     }
     /* new relocation */
@@ -203,6 +257,8 @@ static int cs_gem_write_reloc(struct radeon_cs_int *cs,
     reloc->flags = flags;
     csg->chunks[1].length_dw += RELOC_SIZE;
     radeon_bo_ref(bo);
+    /* bo might be referenced from another context so have to use atomic opertions */
+    atomic_add((atomic_t *)radeon_gem_get_reloc_in_cs(bo), cs->id);
     cs->relocs_total_size += boi->size;
     radeon_cs_write_dword((struct radeon_cs *)cs, 0xc0001000);
     radeon_cs_write_dword((struct radeon_cs *)cs, idx);
@@ -288,6 +344,8 @@ static int cs_gem_emit(struct radeon_cs_int *cs)
                             &csg->cs, sizeof(struct drm_radeon_cs));
     for (i = 0; i < csg->base.crelocs; i++) {
         csg->relocs_bo[i]->space_accounted = 0;
+        /* bo might be referenced from another context so have to use atomic opertions */
+        atomic_dec((atomic_t *)radeon_gem_get_reloc_in_cs((struct radeon_bo*)csg->relocs_bo[i]), cs->id);
         radeon_bo_unref((struct radeon_bo *)csg->relocs_bo[i]);
         csg->relocs_bo[i] = NULL;
     }
@@ -302,6 +360,7 @@ static int cs_gem_destroy(struct radeon_cs_int *cs)
 {
     struct cs_gem *csg = (struct cs_gem*)cs;
 
+    free_id(cs->id);
     free(csg->relocs_bo);
     free(cs->relocs);
     free(cs->packets);
@@ -317,6 +376,8 @@ static int cs_gem_erase(struct radeon_cs_int *cs)
     if (csg->relocs_bo) {
         for (i = 0; i < csg->base.crelocs; i++) {
             if (csg->relocs_bo[i]) {
+                /* bo might be referenced from another context so have to use atomic opertions */
+                atomic_dec((atomic_t *)radeon_gem_get_reloc_in_cs((struct radeon_bo*)csg->relocs_bo[i]), cs->id);
                 radeon_bo_unref((struct radeon_bo *)csg->relocs_bo[i]);
                 csg->relocs_bo[i] = NULL;
             }
diff --git a/radeon/radeon_cs_int.h b/radeon/radeon_cs_int.h
index 8ba76bf9..6cee5742 100644
--- a/radeon/radeon_cs_int.h
+++ b/radeon/radeon_cs_int.h
@@ -28,6 +28,7 @@ struct radeon_cs_int {
     int                         bo_count;
     void                        (*space_flush_fn)(void *);
     void                        *space_flush_data;
+    uint32_t                    id;
 };
 
 /* cs functions */
diff --git a/xf86atomic.h b/xf86atomic.h
index de8e220e..854187a0 100644
--- a/xf86atomic.h
+++ b/xf86atomic.h
@@ -50,6 +50,8 @@ typedef struct {
 # define atomic_set(x, val) ((x)->atomic = (val))
 # define atomic_inc(x) ((void) __sync_fetch_and_add (&(x)->atomic, 1))
 # define atomic_dec_and_test(x) (__sync_fetch_and_add (&(x)->atomic, -1) == 1)
+# define atomic_add(x, v) ((void) __sync_add_and_fetch(&(x)->atomic, (v)))
+# define atomic_dec(x, v) ((void) __sync_sub_and_fetch(&(x)->atomic, (v)))
 # define atomic_cmpxchg(x, oldv, newv) __sync_val_compare_and_swap (&(x)->atomic, oldv, newv)
 
 #endif
@@ -66,6 +68,8 @@ typedef struct {
 # define atomic_read(x) AO_load_full(&(x)->atomic)
 # define atomic_set(x, val) AO_store_full(&(x)->atomic, (val))
 # define atomic_inc(x) ((void) AO_fetch_and_add1_full(&(x)->atomic))
+# define atomic_add(x, v) ((void) AO_fetch_and_add_full(&(x)->atomic, (v)))
+# define atomic_dec(x, v) ((void) AO_fetch_and_add_full(&(x)->atomic, -(v)))
 # define atomic_dec_and_test(x) (AO_fetch_and_sub1_full(&(x)->atomic) == 1)
 # define atomic_cmpxchg(x, oldv, newv) AO_compare_and_swap_full(&(x)->atomic, oldv, newv)
 
@@ -82,6 +86,8 @@ typedef struct { uint_t atomic; } atomic_t;
 # define atomic_set(x, val) ((x)->atomic = (uint_t)(val))
 # define atomic_inc(x) (atomic_inc_uint (&(x)->atomic))
 # define atomic_dec_and_test(x) (atomic_dec_uint_nv(&(x)->atomic) == 1)
+# define atomic_add(x, v) (atomic_add_uint(&(x)->atomic, (v)))
+# define atomic_dec(x, v) (atomic_dec_uint(&(x)->atomic, (v)))
 # define atomic_cmpxchg(x, oldv, newv) atomic_cas_uint (&(x)->atomic, oldv, newv)
 
 #endif