* Nicolai Haehnle <prefect_@gmx.net>
* Jérôme Glisse <glisse@freedesktop.org>
*/
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
#include <assert.h>
#include <errno.h>
#include <stdlib.h>
+#include <string.h>
+#include <pthread.h>
#include <sys/mman.h>
#include <sys/ioctl.h>
#include "radeon_cs.h"
#include "radeon_cs_gem.h"
#include "radeon_bo_gem.h"
#include "drm.h"
+#include "libdrm.h"
#include "xf86drm.h"
+#include "xf86atomic.h"
#include "radeon_drm.h"
+#include "bof.h"
+
+#define CS_BOF_DUMP 0
struct radeon_cs_manager_gem {
- struct radeon_cs_manager base;
- uint32_t device_id;
+ struct radeon_cs_manager base;
+ uint32_t device_id;
+ unsigned nbof;
};
#pragma pack(1)
#define RELOC_SIZE (sizeof(struct cs_reloc_gem) / sizeof(uint32_t))
struct cs_gem {
- struct radeon_cs_int base;
+ struct radeon_cs_int base;
struct drm_radeon_cs cs;
struct drm_radeon_cs_chunk chunks[2];
unsigned nrelocs;
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);
fprintf(stderr, "CS section end at (%s,%s,%d)\n",
file, func, line);
- /* We must reset the section even when there is error. */
- cs->section_ndw = 0;
- return -EPIPE;
+ /* We must reset the section even when there is error. */
+ cs->section_ndw = 0;
+ return -EPIPE;
}
cs->section_ndw = 0;
return 0;
}
+#if CS_BOF_DUMP
+static void cs_gem_dump_bof(struct radeon_cs_int *cs)
+{
+ struct cs_gem *csg = (struct cs_gem*)cs;
+ struct radeon_cs_manager_gem *csm;
+ bof_t *bcs, *blob, *array, *bo, *size, *handle, *device_id, *root;
+ char tmp[256];
+ unsigned i;
+
+ csm = (struct radeon_cs_manager_gem *)cs->csm;
+ root = device_id = bcs = blob = array = bo = size = handle = NULL;
+ root = bof_object();
+ if (root == NULL)
+ goto out_err;
+ device_id = bof_int32(csm->device_id);
+ if (device_id == NULL)
+ return;
+ if (bof_object_set(root, "device_id", device_id))
+ goto out_err;
+ bof_decref(device_id);
+ device_id = NULL;
+ /* dump relocs */
+ blob = bof_blob(csg->nrelocs * 16, csg->relocs);
+ if (blob == NULL)
+ goto out_err;
+ if (bof_object_set(root, "reloc", blob))
+ goto out_err;
+ bof_decref(blob);
+ blob = NULL;
+ /* dump cs */
+ blob = bof_blob(cs->cdw * 4, cs->packets);
+ if (blob == NULL)
+ goto out_err;
+ if (bof_object_set(root, "pm4", blob))
+ goto out_err;
+ bof_decref(blob);
+ blob = NULL;
+ /* dump bo */
+ array = bof_array();
+ if (array == NULL)
+ goto out_err;
+ for (i = 0; i < csg->base.crelocs; i++) {
+ bo = bof_object();
+ if (bo == NULL)
+ goto out_err;
+ size = bof_int32(csg->relocs_bo[i]->size);
+ if (size == NULL)
+ goto out_err;
+ if (bof_object_set(bo, "size", size))
+ goto out_err;
+ bof_decref(size);
+ size = NULL;
+ handle = bof_int32(csg->relocs_bo[i]->handle);
+ if (handle == NULL)
+ goto out_err;
+ if (bof_object_set(bo, "handle", handle))
+ goto out_err;
+ bof_decref(handle);
+ handle = NULL;
+ radeon_bo_map((struct radeon_bo*)csg->relocs_bo[i], 0);
+ blob = bof_blob(csg->relocs_bo[i]->size, csg->relocs_bo[i]->ptr);
+ radeon_bo_unmap((struct radeon_bo*)csg->relocs_bo[i]);
+ if (blob == NULL)
+ goto out_err;
+ if (bof_object_set(bo, "data", blob))
+ goto out_err;
+ bof_decref(blob);
+ blob = NULL;
+ if (bof_array_append(array, bo))
+ goto out_err;
+ bof_decref(bo);
+ bo = NULL;
+ }
+ if (bof_object_set(root, "bo", array))
+ goto out_err;
+ sprintf(tmp, "d-0x%04X-%08d.bof", csm->device_id, csm->nbof++);
+ bof_dump_file(root, tmp);
+out_err:
+ bof_decref(blob);
+ bof_decref(array);
+ bof_decref(bo);
+ bof_decref(size);
+ bof_decref(handle);
+ bof_decref(device_id);
+ bof_decref(root);
+}
+#endif
+
static int cs_gem_emit(struct radeon_cs_int *cs)
{
struct cs_gem *csg = (struct cs_gem*)cs;
unsigned i;
int r;
+ while (cs->cdw & 7)
+ radeon_cs_write_dword((struct radeon_cs *)cs, 0x80000000);
+
+#if CS_BOF_DUMP
+ cs_gem_dump_bof(cs);
+#endif
csg->chunks[0].length_dw = cs->cdw;
chunk_array[0] = (uint64_t)(uintptr_t)&csg->chunks[0];
&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;
}
static int radeon_get_device_id(int fd, uint32_t *device_id)
{
- struct drm_radeon_info info;
+ struct drm_radeon_info info = {};
int r;
*device_id = 0;
info.request = RADEON_INFO_DEVICE_ID;
- info.value = device_id;
+ info.value = (uintptr_t)device_id;
r = drmCommandWriteRead(fd, DRM_RADEON_INFO, &info,
sizeof(struct drm_radeon_info));
return r;
}
-struct radeon_cs_manager *radeon_cs_manager_gem_ctor(int fd)
+drm_public struct radeon_cs_manager *radeon_cs_manager_gem_ctor(int fd)
{
struct radeon_cs_manager_gem *csm;
return &csm->base;
}
-void radeon_cs_manager_gem_dtor(struct radeon_cs_manager *csm)
+drm_public void radeon_cs_manager_gem_dtor(struct radeon_cs_manager *csm)
{
free(csm);
}