#include "pan_util.h"
#include "pandecode/decode.h"
-static void
+void
panfrost_drm_mmap_bo(struct panfrost_screen *screen, struct panfrost_bo *bo)
{
struct drm_panfrost_mmap_bo mmap_bo = { .handle = bo->gem_handle };
ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_MMAP_BO, &mmap_bo);
if (ret) {
- fprintf(stderr, "DRM_IOCTL_PANFROST_MMAP_BO failed: %d\n", ret);
+ fprintf(stderr, "DRM_IOCTL_PANFROST_MMAP_BO failed: %m\n");
assert(0);
}
bo->cpu = os_mmap(NULL, bo->size, PROT_READ | PROT_WRITE, MAP_SHARED,
screen->fd, mmap_bo.offset);
if (bo->cpu == MAP_FAILED) {
- fprintf(stderr, "mmap failed: %p\n", bo->cpu);
+ fprintf(stderr, "mmap failed: %p %m\n", bo->cpu);
assert(0);
}
pandecode_inject_mmap(bo->gpu, bo->cpu, bo->size, NULL);
}
-void
-panfrost_drm_allocate_slab(struct panfrost_screen *screen,
- struct panfrost_memory *mem,
- size_t pages,
- bool same_va,
- int extra_flags,
- int commit_count,
- int extent)
+static void
+panfrost_drm_munmap_bo(struct panfrost_screen *screen, struct panfrost_bo *bo)
{
- struct drm_panfrost_create_bo create_bo = {
- .size = pages * 4096,
- .flags = 0, // TODO figure out proper flags..
- };
- struct drm_panfrost_mmap_bo mmap_bo = {0,};
- int ret;
-
- // TODO cache allocations
- // TODO properly handle errors
- // TODO take into account extra_flags
-
- ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_CREATE_BO, &create_bo);
- if (ret) {
- fprintf(stderr, "DRM_IOCTL_PANFROST_CREATE_BO failed: %d\n", ret);
- assert(0);
- }
-
- mem->gpu = create_bo.offset;
- mem->gem_handle = create_bo.handle;
- mem->stack_bottom = 0;
- mem->size = create_bo.size;
-
- // TODO map and unmap on demand?
- mmap_bo.handle = create_bo.handle;
- ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_MMAP_BO, &mmap_bo);
- if (ret) {
- fprintf(stderr, "DRM_IOCTL_PANFROST_MMAP_BO failed: %d\n", ret);
- assert(0);
- }
-
- mem->cpu = os_mmap(NULL, mem->size, PROT_READ | PROT_WRITE, MAP_SHARED,
- screen->fd, mmap_bo.offset);
- if (mem->cpu == MAP_FAILED) {
- fprintf(stderr, "mmap failed: %p\n", mem->cpu);
- assert(0);
- }
+ if (!bo->cpu)
+ return;
- /* Record the mmap if we're tracing */
- if (pan_debug & PAN_DBG_TRACE)
- pandecode_inject_mmap(mem->gpu, mem->cpu, mem->size, NULL);
+ if (os_munmap((void *) (uintptr_t)bo->cpu, bo->size)) {
+ perror("munmap");
+ abort();
+ }
+
+ bo->cpu = NULL;
+}
+
+struct panfrost_bo *
+panfrost_drm_create_bo(struct panfrost_screen *screen, size_t size,
+ uint32_t flags)
+{
+ struct panfrost_bo *bo;
+
+ /* Kernel will fail (confusingly) with EPERM otherwise */
+ assert(size > 0);
+
+ /* To maximize BO cache usage, don't allocate tiny BOs */
+ size = MAX2(size, 4096);
+
+ /* GROWABLE BOs cannot be mmapped */
+ if (flags & PAN_ALLOCATE_GROWABLE)
+ assert(flags & PAN_ALLOCATE_INVISIBLE);
+
+ unsigned translated_flags = 0;
+
+ if (screen->kernel_version->version_major > 1 ||
+ screen->kernel_version->version_minor >= 1) {
+ if (flags & PAN_ALLOCATE_GROWABLE)
+ translated_flags |= PANFROST_BO_HEAP;
+ if (!(flags & PAN_ALLOCATE_EXECUTE))
+ translated_flags |= PANFROST_BO_NOEXEC;
+ }
+
+ struct drm_panfrost_create_bo create_bo = {
+ .size = size,
+ .flags = translated_flags,
+ };
+
+ /* Before creating a BO, we first want to check the cache */
+
+ bo = panfrost_bo_cache_fetch(screen, size, flags);
+
+ if (bo == NULL) {
+ /* Otherwise, the cache misses and we need to allocate a BO fresh from
+ * the kernel */
+
+ int ret;
+
+ ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_CREATE_BO, &create_bo);
+ if (ret) {
+ fprintf(stderr, "DRM_IOCTL_PANFROST_CREATE_BO failed: %m\n");
+ assert(0);
+ }
+
+ /* We have a BO allocated from the kernel; fill in the userspace
+ * version */
+
+ bo = rzalloc(screen, struct panfrost_bo);
+ bo->size = create_bo.size;
+ bo->gpu = create_bo.offset;
+ bo->gem_handle = create_bo.handle;
+ bo->flags = flags;
+ }
+
+ /* Only mmap now if we know we need to. For CPU-invisible buffers, we
+ * never map since we don't care about their contents; they're purely
+ * for GPU-internal use. But we do trace them anyway. */
+
+ if (!(flags & (PAN_ALLOCATE_INVISIBLE | PAN_ALLOCATE_DELAY_MMAP)))
+ panfrost_drm_mmap_bo(screen, bo);
+ else if (flags & PAN_ALLOCATE_INVISIBLE) {
+ if (pan_debug & PAN_DBG_TRACE)
+ pandecode_inject_mmap(bo->gpu, NULL, bo->size, NULL);
+ }
+
+ pipe_reference_init(&bo->reference, 1);
+ return bo;
}
void
-panfrost_drm_free_slab(struct panfrost_screen *screen, struct panfrost_memory *mem)
+panfrost_drm_release_bo(struct panfrost_screen *screen, struct panfrost_bo *bo, bool cacheable)
{
- struct drm_gem_close gem_close = {
- .handle = mem->gem_handle,
- };
- int ret;
+ if (!bo)
+ return;
- if (os_munmap((void *) (uintptr_t) mem->cpu, mem->size)) {
- perror("munmap");
- abort();
+ struct drm_gem_close gem_close = { .handle = bo->gem_handle };
+ int ret;
+
+ /* Rather than freeing the BO now, we'll cache the BO for later
+ * allocations if we're allowed to */
+
+ panfrost_drm_munmap_bo(screen, bo);
+
+ if (cacheable) {
+ bool cached = panfrost_bo_cache_put(screen, bo);
+
+ if (cached)
+ return;
}
- mem->cpu = NULL;
+ /* Otherwise, if the BO wasn't cached, we'll legitimately free the BO */
- ret = drmIoctl(screen->fd, DRM_IOCTL_GEM_CLOSE, &gem_close);
- if (ret) {
- fprintf(stderr, "DRM_IOCTL_GEM_CLOSE failed: %d\n", ret);
- assert(0);
- }
+ ret = drmIoctl(screen->fd, DRM_IOCTL_GEM_CLOSE, &gem_close);
+ if (ret) {
+ fprintf(stderr, "DRM_IOCTL_GEM_CLOSE failed: %m\n");
+ assert(0);
+ }
+
+ ralloc_free(bo);
+}
+
+void
+panfrost_drm_allocate_slab(struct panfrost_screen *screen,
+ struct panfrost_memory *mem,
+ size_t pages,
+ bool same_va,
+ int extra_flags,
+ int commit_count,
+ int extent)
+{
+ // TODO cache allocations
+ // TODO properly handle errors
+ // TODO take into account extra_flags
+ mem->bo = panfrost_drm_create_bo(screen, pages * 4096, extra_flags);
+ mem->stack_bottom = 0;
+}
- mem->gem_handle = -1;
+void
+panfrost_drm_free_slab(struct panfrost_screen *screen, struct panfrost_memory *mem)
+{
+ panfrost_bo_unreference(&screen->base, mem->bo);
+ mem->bo = NULL;
}
struct panfrost_bo *
panfrost_drm_import_bo(struct panfrost_screen *screen, int fd)
{
- struct panfrost_bo *bo = rzalloc(screen, struct panfrost_bo);
+ struct panfrost_bo *bo = rzalloc(screen, struct panfrost_bo);
struct drm_panfrost_get_bo_offset get_bo_offset = {0,};
- int ret;
+ ASSERTED int ret;
unsigned gem_handle;
- ret = drmPrimeFDToHandle(screen->fd, fd, &gem_handle);
- assert(!ret);
+ ret = drmPrimeFDToHandle(screen->fd, fd, &gem_handle);
+ assert(!ret);
- get_bo_offset.handle = gem_handle;
+ get_bo_offset.handle = gem_handle;
ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_GET_BO_OFFSET, &get_bo_offset);
assert(!ret);
- bo->gem_handle = gem_handle;
+ bo->gem_handle = gem_handle;
bo->gpu = (mali_ptr) get_bo_offset.offset;
bo->size = lseek(fd, 0, SEEK_END);
assert(bo->size > 0);
}
static int
-panfrost_drm_submit_job(struct panfrost_context *ctx, u64 job_desc, int reqs, struct pipe_surface *surf)
+panfrost_drm_submit_job(struct panfrost_context *ctx, u64 job_desc, int reqs)
{
struct pipe_context *gallium = (struct pipe_context *) ctx;
struct panfrost_screen *screen = pan_screen(gallium->screen);
+ struct panfrost_job *job = panfrost_get_job_for_fbo(ctx);
struct drm_panfrost_submit submit = {0,};
- int bo_handles[7];
+ int *bo_handles, ret;
submit.in_syncs = (u64) (uintptr_t) &ctx->out_sync;
submit.in_sync_count = 1;
submit.out_sync = ctx->out_sync;
- submit.jc = job_desc;
- submit.requirements = reqs;
-
- if (surf) {
- struct panfrost_resource *res = pan_resource(surf->texture);
- assert(res->bo->gem_handle > 0);
- bo_handles[submit.bo_handle_count++] = res->bo->gem_handle;
- }
-
- /* TODO: Add here the transient pools */
- /* TODO: Add here the BOs listed in the panfrost_job */
- bo_handles[submit.bo_handle_count++] = ctx->shaders.gem_handle;
- bo_handles[submit.bo_handle_count++] = ctx->scratchpad.gem_handle;
- bo_handles[submit.bo_handle_count++] = ctx->tiler_heap.gem_handle;
- bo_handles[submit.bo_handle_count++] = ctx->varying_mem.gem_handle;
- bo_handles[submit.bo_handle_count++] = ctx->tiler_polygon_list.gem_handle;
- submit.bo_handles = (u64) (uintptr_t) bo_handles;
-
- if (drmIoctl(screen->fd, DRM_IOCTL_PANFROST_SUBMIT, &submit)) {
- fprintf(stderr, "Error submitting: %m\n");
- return errno;
- }
+ submit.jc = job_desc;
+ submit.requirements = reqs;
+
+ bo_handles = calloc(job->bos->entries, sizeof(*bo_handles));
+ assert(bo_handles);
+
+ set_foreach(job->bos, entry) {
+ struct panfrost_bo *bo = (struct panfrost_bo *)entry->key;
+ assert(bo->gem_handle > 0);
+ bo_handles[submit.bo_handle_count++] = bo->gem_handle;
+ }
+
+ submit.bo_handles = (u64) (uintptr_t) bo_handles;
+ ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_SUBMIT, &submit);
+ free(bo_handles);
+ if (ret) {
+ fprintf(stderr, "Error submitting: %m\n");
+ return errno;
+ }
/* Trace the job if we're doing that */
if (pan_debug & PAN_DBG_TRACE) {
/* Wait so we can get errors reported back */
drmSyncobjWait(screen->fd, &ctx->out_sync, 1, INT64_MAX, 0, NULL);
- pandecode_replay_jc(submit.jc, FALSE);
+ pandecode_jc(submit.jc, FALSE);
}
- return 0;
+ return 0;
}
int
-panfrost_drm_submit_vs_fs_job(struct panfrost_context *ctx, bool has_draws, bool is_scanout)
+panfrost_drm_submit_vs_fs_job(struct panfrost_context *ctx, bool has_draws)
{
- struct pipe_surface *surf = ctx->pipe_framebuffer.cbufs[0];
- int ret;
+ int ret = 0;
struct panfrost_job *job = panfrost_get_job_for_fbo(ctx);
+ panfrost_job_add_bo(job, ctx->scratchpad.bo);
+ panfrost_job_add_bo(job, ctx->tiler_heap.bo);
+ panfrost_job_add_bo(job, job->polygon_list);
+
if (job->first_job.gpu) {
- ret = panfrost_drm_submit_job(ctx, job->first_job.gpu, 0, NULL);
- assert(!ret);
- }
+ ret = panfrost_drm_submit_job(ctx, job->first_job.gpu, 0);
+ assert(!ret);
+ }
if (job->first_tiler.gpu || job->clear) {
- ret = panfrost_drm_submit_job(ctx, panfrost_fragment_job(ctx, has_draws), PANFROST_JD_REQ_FS, surf);
+ ret = panfrost_drm_submit_job(ctx, panfrost_fragment_job(ctx, has_draws), PANFROST_JD_REQ_FS);
assert(!ret);
}
*/
drmSyncobjExportSyncFile(screen->fd, ctx->out_sync, &f->fd);
if (f->fd == -1) {
- fprintf(stderr, "export failed\n");
+ fprintf(stderr, "export failed: %m\n");
free(f);
return NULL;
}
void
panfrost_drm_force_flush_fragment(struct panfrost_context *ctx,
- struct pipe_fence_handle **fence)
+ struct pipe_fence_handle **fence)
{
struct pipe_context *gallium = (struct pipe_context *) ctx;
struct panfrost_screen *screen = pan_screen(gallium->screen);
- if (!screen->last_fragment_flushed) {
- drmSyncobjWait(screen->fd, &ctx->out_sync, 1, INT64_MAX, 0, NULL);
- screen->last_fragment_flushed = true;
+ if (!ctx->last_fragment_flushed) {
+ drmSyncobjWait(screen->fd, &ctx->out_sync, 1, INT64_MAX, 0, NULL);
+ ctx->last_fragment_flushed = true;
/* The job finished up, so we're safe to clean it up now */
- panfrost_free_job(ctx, screen->last_job);
- }
+ panfrost_free_job(ctx, ctx->last_job);
+ }
if (fence) {
struct panfrost_fence *f = panfrost_fence_create(ctx);
panfrost_drm_query_gpu_version(struct panfrost_screen *screen)
{
struct drm_panfrost_get_param get_param = {0,};
- int ret;
+ ASSERTED int ret;
- get_param.param = DRM_PANFROST_PARAM_GPU_PROD_ID;
+ get_param.param = DRM_PANFROST_PARAM_GPU_PROD_ID;
ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_GET_PARAM, &get_param);
assert(!ret);
- return get_param.value;
+ return get_param.value;
}
int
void
panfrost_drm_fence_reference(struct pipe_screen *screen,
- struct pipe_fence_handle **ptr,
- struct pipe_fence_handle *fence)
+ struct pipe_fence_handle **ptr,
+ struct pipe_fence_handle *fence)
{
struct panfrost_fence **p = (struct panfrost_fence **)ptr;
struct panfrost_fence *f = (struct panfrost_fence *)fence;
boolean
panfrost_drm_fence_finish(struct pipe_screen *pscreen,
- struct pipe_context *ctx,
- struct pipe_fence_handle *fence,
- uint64_t timeout)
+ struct pipe_context *ctx,
+ struct pipe_fence_handle *fence,
+ uint64_t timeout)
{
struct panfrost_screen *screen = pan_screen(pscreen);
struct panfrost_fence *f = (struct panfrost_fence *)fence;
return false;
}
- drmSyncobjImportSyncFile(screen->fd, syncobj, f->fd);
+ ret = drmSyncobjImportSyncFile(screen->fd, syncobj, f->fd);
if (ret) {
fprintf(stderr, "Failed to import fence to syncobj: %m\n");
return false;
projects / mesa.git / blobdiff