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],
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,
[
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
#include <assert.h>
#include <errno.h>
#include <stdlib.h>
+#include <pthread.h>
#include <sys/mman.h>
#include <sys/ioctl.h>
#include "radeon_cs.h"
#include "radeon_bo_gem.h"
#include "drm.h"
#include "xf86drm.h"
+#include "xf86atomic.h"
#include "radeon_drm.h"
struct radeon_cs_manager_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)
{
}
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*));
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 */
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);
&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;
}
{
struct cs_gem *csg = (struct cs_gem*)cs;
+ free_id(cs->id);
free(csg->relocs_bo);
free(cs->relocs);
free(cs->packets);
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;
}
# 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
# 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)
# 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