static void
queue(struct vc4_compile *c, uint64_t inst)
{
struct queued_qpu_inst *q = rzalloc(c, struct queued_qpu_inst);
q->inst = inst;
list_addtail(&q->link, &c->qpu_inst_list);
}
void fd_bo_del(struct fd_bo *bo)
{
struct fd_device *dev = bo->dev;
if (!atomic_dec_and_test(&bo->refcnt))
return;
pthread_mutex_lock(&table_lock);
if (bo->bo_reuse) {
struct fd_bo_bucket *bucket = get_bucket(dev, bo->size);
/* see if we can be green and recycle: */
if (bucket) {
struct timespec time;
clock_gettime(CLOCK_MONOTONIC, &time);
bo->free_time = time.tv_sec;
list_addtail(&bo->list, &bucket->list);
fd_cleanup_bo_cache(dev, time.tv_sec);
/* bo's in the bucket cache don't have a ref and
* don't hold a ref to the dev:
*/
goto out;
}
}
bo_del(bo);
out:
fd_device_del_locked(dev);
pthread_mutex_unlock(&table_lock);
}
//添加学生的函数
void addstudent()
{
struct student stu;
int flag;
while(1)
{
flag = 0;
CLEAR_SCREEN();
struct list *find;
printf("\033[33m\033[2;50H%s\033[0m","请输入以下信息");
printf("\033[33m\033[4;50H%s\033[0m","请输入学生的ID：");
scanf("%s",stu.id);
find = list_find(list_student, &stu, comp_id_stu);
if(find != NULL)
{
printf("\033[33m\033[6;50H%s\033[0m","帐号已注册，请按任意键重试。\n");
getchar();
getchar();
}
else
{
flag = 1;
}
if(flag == 1)
{
break;
}
}
printf("\033[33m\033[5;50H%s\033[0m","请输入学生的密码：");
scanf("%s",stu.pass);
printf("\033[33m\033[6;50H%s\033[0m","请输入学生的姓名：");
scanf("%s",stu.name);
printf("\033[33m\033[7;50H%s\033[0m","请输入学生的年龄：");
scanf("%d",&stu.age);
printf("\033[33m\033[8;50H%s\033[0m","请输入学生的班级：");
scanf("%s",stu.class);
printf("\033[33m\033[9;50H%s\033[0m","请输入学生的语文成绩：");
scanf("%d",&stu.grade[0]);
printf("\033[33m\033[10;50H%s\033[0m","请输入学生的数学成绩：");
scanf("%d",&stu.grade[1]);
printf("\033[33m\033[11;50H%s\033[0m","请输入学生的C语言成绩：");
scanf("%d",&stu.grade[2]);
stu.average = (stu.grade[0]+stu.grade[1]+stu.grade[2])/3;
//将stu插到链表中去
if(list_addtail(list_student,&stu) == 0)
{
printf("\033[33m\033[12;50H%s\033[0m","添加学生数据成功！\n");
getchar();
getchar();
}
else
{
printf("\033[33m\033[12;50H%s\033[0m","添加学生数据失败！\n");
getchar();
getchar();
}
}
void ppir_node_add_dep(ppir_node *succ, ppir_node *pred)
{
/* don't add dep for two nodes from different block */
if (succ->block != pred->block)
return;
/* don't add duplicated dep */
ppir_node_foreach_pred(succ, dep) {
if (dep->pred == pred)
return;
}
ppir_dep *dep = ralloc(succ, ppir_dep);
dep->pred = pred;
dep->succ = succ;
list_addtail(&dep->pred_link, &succ->pred_list);
list_addtail(&dep->succ_link, &pred->succ_list);
}
static void
pause_query(struct fd_context *ctx, struct fd_hw_query *hq,
struct fd_ringbuffer *ring)
{
int idx = pidx(hq->provider->query_type);
assert(idx >= 0); /* query never would have been created otherwise */
assert(hq->period && !hq->period->end);
assert(ctx->active_providers & (1 << idx));
hq->period->end = get_sample(ctx, ring, hq->base.type);
list_addtail(&hq->period->list, &hq->current_periods);
hq->period = NULL;
}
void
ir3_insert_by_depth(struct ir3_instruction *instr, struct list_head *list)
{
/* remove from existing spot in list: */
list_delinit(&instr->node);
/* find where to re-insert instruction: */
list_for_each_entry (struct ir3_instruction, pos, list, node) {
if (pos->depth > instr->depth) {
list_add(&instr->node, &pos->node);
return;
}
}
/* if we get here, we didn't find an insertion spot: */
list_addtail(&instr->node, list);
}
struct reloc_worklist_entry *
clif_dump_add_address_to_worklist(struct clif_dump *clif,
enum reloc_worklist_type type,
uint32_t addr)
{
struct reloc_worklist_entry *entry =
rzalloc(clif, struct reloc_worklist_entry);
if (!entry)
return NULL;
entry->type = type;
entry->addr = addr;
list_addtail(&entry->link, &clif->worklist);
return entry;
}
static void
create_object(const char *chipname, const char *featurename,
const sensors_chip_name *chip, const sensors_feature *feature,
int mode)
{
struct sensors_temp_info *sti = CALLOC_STRUCT(sensors_temp_info);
sti->mode = mode;
sti->chip = (sensors_chip_name *) chip;
sti->feature = feature;
strcpy(sti->chipname, chipname);
strcpy(sti->featurename, featurename);
snprintf(sti->name, sizeof(sti->name), "%s.%s", sti->chipname,
sti->featurename);
list_addtail(&sti->list, &gsensors_temp_list);
gsensors_temp_count++;
}
void
etna_resource_used(struct etna_context *ctx, struct pipe_resource *prsc,
enum etna_resource_status status)
{
struct etna_resource *rsc;
if (!prsc)
return;
rsc = etna_resource(prsc);
rsc->status |= status;
/* TODO resources can actually be shared across contexts,
* so I'm not sure a single list-head will do the trick? */
debug_assert((rsc->pending_ctx == ctx) || !rsc->pending_ctx);
list_delinit(&rsc->list);
list_addtail(&rsc->list, &ctx->used_resources);
rsc->pending_ctx = ctx;
}
static void virgl_buffer_transfer_flush_region(struct pipe_context *ctx,
struct pipe_transfer *transfer,
const struct pipe_box *box)
{
struct virgl_context *vctx = virgl_context(ctx);
struct virgl_buffer *vbuf = virgl_buffer(transfer->resource);
if (!vbuf->on_list) {
struct pipe_resource *res = NULL;
list_addtail(&vbuf->flush_list, &vctx->to_flush_bufs);
vbuf->on_list = TRUE;
pipe_resource_reference(&res, &vbuf->base.u.b);
}
util_range_add(&vbuf->valid_buffer_range, transfer->box.x + box->x,
transfer->box.x + box->x + box->width);
vbuf->base.clean = FALSE;
}
//学生数据初始化
int init_student()
{
FILE *fp;
struct student stu;
//创建链表用于存放学生的数据
list_student = list_create(sizeof(struct student));
if(list_student == NULL)
{ //分配链表头结点失败
return -1;
}
//读取student.dat文件，然后保存所有的数据到list_student链表里面
fp = fopen("dat/student.dat","r");
if(fp == NULL)
{
// printf("没有student.dat文件.\n");
fp = fopen("dat/student.dat","w+");
if(fp == NULL)
{
list_destroy(&list_student);
return -1;
}
}
while(1)
{
if(fread(&stu, sizeof(stu), 1, fp) != 1)
{ //读取文件结束
if(feof(fp) || ferror(fp))
{
break;
}
}
if(list_addtail(list_student, &stu) == -1)
{
list_destroy(&list_student);
fclose(fp);
return -1;
}
}
fclose(fp);
return 0;
}
//管理员初始化
int init_manager()
{
FILE *fp;
struct manager man = {"admin" , "admin"};
//创建一个链表用于存放管理员的数据
list_manager = list_create(sizeof(struct manager));
if(list_manager == NULL)
{
return -1;
}
//读取manager.dat文件，然后保存所有的数据到链表里面
fp = fopen("dat/manager.dat","r");
if(fp == NULL) //文件不存在
{
fp = fopen("dat/manager.dat", "w+"); //创建该文件
if(fp == NULL)
{
list_destroy(&list_manager);
return -1;
}
fwrite(&man, sizeof(man), 1, fp);
rewind(fp);
}
while(1)
{
if(fread(&man, sizeof(man), 1, fp) != 1)
{ //读取文件结束
if(feof(fp) || ferror(fp))
break;
}
//把man的信息写道链表里面
if(list_addtail(list_manager, &man) == -1)
{
list_destroy(&list_manager);
fclose(fp);
return -1;
}
}
fclose(fp);
return 0;
}
static struct vc4_mem_area_rec *
vc4_add_mem_area_to_list(struct vc4_mem_area_rec *rec)
{
/* Don't add exact duplicates of memory areas to the list. We have to
* be careful to not compare the list pointers, since the new rec
* won't be in the list.
*/
struct vc4_mem_area_rec compare_a = *rec;
memset(&compare_a.link, 0, sizeof(compare_a.link));
list_for_each_entry(struct vc4_mem_area_rec, list_rec, &dump.mem_areas,
link) {
struct vc4_mem_area_rec compare_b = *list_rec;
memset(&compare_b.link, 0, sizeof(compare_b.link));
if (memcmp(&compare_a, &compare_b, sizeof(compare_a)) == 0)
return list_rec;
}
struct vc4_mem_area_rec *list_rec = malloc(sizeof(*list_rec));
*list_rec = *rec;
list_addtail(&list_rec->link, &dump.mem_areas);
return list_rec;
}
static void
schedule(struct ir3_sched_ctx *ctx, struct ir3_instruction *instr)
{
debug_assert(ctx->block == instr->block);
/* maybe there is a better way to handle this than just stuffing
* a nop.. ideally we'd know about this constraint in the
* scheduling and depth calculation..
*/
if (ctx->scheduled && is_sfu_or_mem(ctx->scheduled) && is_sfu_or_mem(instr))
ir3_NOP(ctx->block);
/* remove from depth list:
*/
list_delinit(&instr->node);
if (writes_addr(instr)) {
debug_assert(ctx->addr == NULL);
ctx->addr = instr;
}
if (writes_pred(instr)) {
debug_assert(ctx->pred == NULL);
ctx->pred = instr;
}
instr->flags |= IR3_INSTR_MARK;
list_addtail(&instr->node, &instr->block->instr_list);
ctx->scheduled = instr;
if (writes_addr(instr) || writes_pred(instr) || is_input(instr)) {
clear_cache(ctx, NULL);
} else {
/* invalidate only the necessary entries.. */
clear_cache(ctx, instr);
}
}
static int vc4_create_rcl_bo(struct drm_device *dev, struct vc4_exec_info *exec,
struct vc4_rcl_setup *setup)
{
struct drm_vc4_submit_cl *args = exec->args;
bool has_bin = args->bin_cl_size != 0;
uint8_t min_x_tile = args->min_x_tile;
uint8_t min_y_tile = args->min_y_tile;
uint8_t max_x_tile = args->max_x_tile;
uint8_t max_y_tile = args->max_y_tile;
uint8_t xtiles = max_x_tile - min_x_tile + 1;
uint8_t ytiles = max_y_tile - min_y_tile + 1;
uint8_t x, y;
uint32_t size, loop_body_size;
size = VC4_PACKET_TILE_RENDERING_MODE_CONFIG_SIZE;
loop_body_size = VC4_PACKET_TILE_COORDINATES_SIZE;
if (args->flags & VC4_SUBMIT_CL_USE_CLEAR_COLOR) {
size += VC4_PACKET_CLEAR_COLORS_SIZE +
VC4_PACKET_TILE_COORDINATES_SIZE +
VC4_PACKET_STORE_TILE_BUFFER_GENERAL_SIZE;
}
if (setup->color_read) {
if (args->color_read.flags &
VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES) {
loop_body_size += VC4_PACKET_LOAD_FULL_RES_TILE_BUFFER_SIZE;
} else {
loop_body_size += VC4_PACKET_LOAD_TILE_BUFFER_GENERAL_SIZE;
}
}
if (setup->zs_read) {
if (args->zs_read.flags &
VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES) {
loop_body_size += VC4_PACKET_LOAD_FULL_RES_TILE_BUFFER_SIZE;
} else {
if (setup->color_read &&
!(args->color_read.flags &
VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES)) {
loop_body_size += VC4_PACKET_TILE_COORDINATES_SIZE;
loop_body_size += VC4_PACKET_STORE_TILE_BUFFER_GENERAL_SIZE;
}
loop_body_size += VC4_PACKET_LOAD_TILE_BUFFER_GENERAL_SIZE;
}
}
if (has_bin) {
size += VC4_PACKET_WAIT_ON_SEMAPHORE_SIZE;
loop_body_size += VC4_PACKET_BRANCH_TO_SUB_LIST_SIZE;
}
if (setup->msaa_color_write)
loop_body_size += VC4_PACKET_STORE_FULL_RES_TILE_BUFFER_SIZE;
if (setup->msaa_zs_write)
loop_body_size += VC4_PACKET_STORE_FULL_RES_TILE_BUFFER_SIZE;
if (setup->zs_write)
loop_body_size += VC4_PACKET_STORE_TILE_BUFFER_GENERAL_SIZE;
if (setup->color_write)
loop_body_size += VC4_PACKET_STORE_MS_TILE_BUFFER_SIZE;
/* We need a VC4_PACKET_TILE_COORDINATES in between each store. */
loop_body_size += VC4_PACKET_TILE_COORDINATES_SIZE *
((setup->msaa_color_write != NULL) +
(setup->msaa_zs_write != NULL) +
(setup->color_write != NULL) +
(setup->zs_write != NULL) - 1);
size += xtiles * ytiles * loop_body_size;
setup->rcl = drm_gem_cma_create(dev, size);
if (!setup->rcl)
return -ENOMEM;
list_addtail(&to_vc4_bo(&setup->rcl->base)->unref_head,
&exec->unref_list);
rcl_u8(setup, VC4_PACKET_TILE_RENDERING_MODE_CONFIG);
rcl_u32(setup,
(setup->color_write ? (setup->color_write->paddr +
args->color_write.offset) :
0));
rcl_u16(setup, args->width);
rcl_u16(setup, args->height);
rcl_u16(setup, args->color_write.bits);
/* The tile buffer gets cleared when the previous tile is stored. If
* the clear values changed between frames, then the tile buffer has
* stale clear values in it, so we have to do a store in None mode (no
* writes) so that we trigger the tile buffer clear.
*/
if (args->flags & VC4_SUBMIT_CL_USE_CLEAR_COLOR) {
rcl_u8(setup, VC4_PACKET_CLEAR_COLORS);
rcl_u32(setup, args->clear_color[0]);
rcl_u32(setup, args->clear_color[1]);
rcl_u32(setup, args->clear_z);
rcl_u8(setup, args->clear_s);
vc4_tile_coordinates(setup, 0, 0);
rcl_u8(setup, VC4_PACKET_STORE_TILE_BUFFER_GENERAL);
rcl_u16(setup, VC4_LOADSTORE_TILE_BUFFER_NONE);
rcl_u32(setup, 0); /* no address, since we're in None mode */
}
for (y = min_y_tile; y <= max_y_tile; y++) {
for (x = min_x_tile; x <= max_x_tile; x++) {
bool first = (x == min_x_tile && y == min_y_tile);
bool last = (x == max_x_tile && y == max_y_tile);
emit_tile(exec, setup, x, y, first, last);
}
}
BUG_ON(setup->next_offset != size);
exec->ct1ca = setup->rcl->paddr;
exec->ct1ea = setup->rcl->paddr + setup->next_offset;
return 0;
}
int
iris_bo_busy(struct iris_bo *bo)
{
struct iris_bufmgr *bufmgr = bo->bufmgr;
struct drm_i915_gem_busy busy = { .handle = bo->gem_handle };
int ret = drm_ioctl(bufmgr->fd, DRM_IOCTL_I915_GEM_BUSY, &busy);
if (ret == 0) {
bo->idle = !busy.busy;
return busy.busy;
}
return false;
}
int
iris_bo_madvise(struct iris_bo *bo, int state)
{
struct drm_i915_gem_madvise madv = {
.handle = bo->gem_handle,
.madv = state,
.retained = 1,
};
drm_ioctl(bo->bufmgr->fd, DRM_IOCTL_I915_GEM_MADVISE, &madv);
return madv.retained;
}
/* drop the oldest entries that have been purged by the kernel */
static void
iris_bo_cache_purge_bucket(struct iris_bufmgr *bufmgr,
struct bo_cache_bucket *bucket)
{
list_for_each_entry_safe(struct iris_bo, bo, &bucket->head, head) {
if (iris_bo_madvise(bo, I915_MADV_DONTNEED))
break;
list_del(&bo->head);
bo_free(bo);
}
}
static struct iris_bo *
bo_calloc(void)
{
struct iris_bo *bo = calloc(1, sizeof(*bo));
if (bo) {
bo->hash = _mesa_hash_pointer(bo);
}
return bo;
}
static struct iris_bo *
bo_alloc_internal(struct iris_bufmgr *bufmgr,
const char *name,
uint64_t size,
enum iris_memory_zone memzone,
unsigned flags,
uint32_t tiling_mode,
uint32_t stride)
{
struct iris_bo *bo;
unsigned int page_size = getpagesize();
int ret;
struct bo_cache_bucket *bucket;
bool alloc_from_cache;
uint64_t bo_size;
bool zeroed = false;
if (flags & BO_ALLOC_ZEROED)
zeroed = true;
if ((flags & BO_ALLOC_COHERENT) && !bufmgr->has_llc) {
bo_size = MAX2(ALIGN(size, page_size), page_size);
bucket = NULL;
goto skip_cache;
}
/* Round the allocated size up to a power of two number of pages. */
bucket = bucket_for_size(bufmgr, size);
/* If we don't have caching at this size, don't actually round the
* allocation up.
*/
if (bucket == NULL) {
bo_size = MAX2(ALIGN(size, page_size), page_size);
} else {
bo_size = bucket->size;
}
mtx_lock(&bufmgr->lock);
/* Get a buffer out of the cache if available */
retry:
alloc_from_cache = false;
if (bucket != NULL && !list_empty(&bucket->head)) {
/* If the last BO in the cache is idle, then reuse it. Otherwise,
* allocate a fresh buffer to avoid stalling.
*/
bo = LIST_ENTRY(struct iris_bo, bucket->head.next, head);
if (!iris_bo_busy(bo)) {
alloc_from_cache = true;
list_del(&bo->head);
}
if (alloc_from_cache) {
if (!iris_bo_madvise(bo, I915_MADV_WILLNEED)) {
bo_free(bo);
iris_bo_cache_purge_bucket(bufmgr, bucket);
goto retry;
}
if (bo_set_tiling_internal(bo, tiling_mode, stride)) {
bo_free(bo);
goto retry;
}
if (zeroed) {
void *map = iris_bo_map(NULL, bo, MAP_WRITE | MAP_RAW);
if (!map) {
bo_free(bo);
goto retry;
}
memset(map, 0, bo_size);
}
}
}
if (alloc_from_cache) {
/* If the cached BO isn't in the right memory zone, free the old
* memory and assign it a new address.
*/
if (memzone != iris_memzone_for_address(bo->gtt_offset)) {
vma_free(bufmgr, bo->gtt_offset, bo->size);
bo->gtt_offset = 0ull;
}
} else {
skip_cache:
bo = bo_calloc();
if (!bo)
goto err;
bo->size = bo_size;
bo->idle = true;
struct drm_i915_gem_create create = { .size = bo_size };
/* All new BOs we get from the kernel are zeroed, so we don't need to
* worry about that here.
*/
ret = drm_ioctl(bufmgr->fd, DRM_IOCTL_I915_GEM_CREATE, &create);
if (ret != 0) {
free(bo);
goto err;
}
bo->gem_handle = create.handle;
bo->bufmgr = bufmgr;
bo->tiling_mode = I915_TILING_NONE;
bo->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
bo->stride = 0;
if (bo_set_tiling_internal(bo, tiling_mode, stride))
goto err_free;
/* Calling set_domain() will allocate pages for the BO outside of the
* struct mutex lock in the kernel, which is more efficient than waiting
* to create them during the first execbuf that uses the BO.
*/
struct drm_i915_gem_set_domain sd = {
.handle = bo->gem_handle,
.read_domains = I915_GEM_DOMAIN_CPU,
.write_domain = 0,
};
if (drm_ioctl(bo->bufmgr->fd, DRM_IOCTL_I915_GEM_SET_DOMAIN, &sd) != 0)
goto err_free;
}
bo->name = name;
p_atomic_set(&bo->refcount, 1);
bo->reusable = bucket && bufmgr->bo_reuse;
bo->cache_coherent = bufmgr->has_llc;
bo->index = -1;
bo->kflags = EXEC_OBJECT_SUPPORTS_48B_ADDRESS | EXEC_OBJECT_PINNED;
/* By default, capture all driver-internal buffers like shader kernels,
* surface states, dynamic states, border colors, and so on.
*/
if (memzone < IRIS_MEMZONE_OTHER)
bo->kflags |= EXEC_OBJECT_CAPTURE;
if (bo->gtt_offset == 0ull) {
bo->gtt_offset = vma_alloc(bufmgr, memzone, bo->size, 1);
if (bo->gtt_offset == 0ull)
goto err_free;
}
mtx_unlock(&bufmgr->lock);
if ((flags & BO_ALLOC_COHERENT) && !bo->cache_coherent) {
struct drm_i915_gem_caching arg = {
.handle = bo->gem_handle,
.caching = 1,
};
if (drm_ioctl(bufmgr->fd, DRM_IOCTL_I915_GEM_SET_CACHING, &arg) == 0) {
bo->cache_coherent = true;
bo->reusable = false;
}
}
DBG("bo_create: buf %d (%s) (%s memzone) %llub\n", bo->gem_handle,
bo->name, memzone_name(memzone), (unsigned long long) size);
return bo;
err_free:
bo_free(bo);
err:
mtx_unlock(&bufmgr->lock);
return NULL;
}
struct iris_bo *
iris_bo_alloc(struct iris_bufmgr *bufmgr,
const char *name,
uint64_t size,
enum iris_memory_zone memzone)
{
return bo_alloc_internal(bufmgr, name, size, memzone,
0, I915_TILING_NONE, 0);
}
struct iris_bo *
iris_bo_alloc_tiled(struct iris_bufmgr *bufmgr, const char *name,
uint64_t size, enum iris_memory_zone memzone,
uint32_t tiling_mode, uint32_t pitch, unsigned flags)
{
return bo_alloc_internal(bufmgr, name, size, memzone,
flags, tiling_mode, pitch);
}
struct iris_bo *
iris_bo_create_userptr(struct iris_bufmgr *bufmgr, const char *name,
void *ptr, size_t size,
enum iris_memory_zone memzone)
{
struct iris_bo *bo;
bo = bo_calloc();
if (!bo)
return NULL;
struct drm_i915_gem_userptr arg = {
.user_ptr = (uintptr_t)ptr,
.user_size = size,
};
if (drm_ioctl(bufmgr->fd, DRM_IOCTL_I915_GEM_USERPTR, &arg))
goto err_free;
bo->gem_handle = arg.handle;
/* Check the buffer for validity before we try and use it in a batch */
struct drm_i915_gem_set_domain sd = {
.handle = bo->gem_handle,
.read_domains = I915_GEM_DOMAIN_CPU,
};
if (drm_ioctl(bufmgr->fd, DRM_IOCTL_I915_GEM_SET_DOMAIN, &sd))
goto err_close;
bo->name = name;
bo->size = size;
bo->map_cpu = ptr;
bo->bufmgr = bufmgr;
bo->kflags = EXEC_OBJECT_SUPPORTS_48B_ADDRESS | EXEC_OBJECT_PINNED;
bo->gtt_offset = vma_alloc(bufmgr, memzone, size, 1);
if (bo->gtt_offset == 0ull)
goto err_close;
p_atomic_set(&bo->refcount, 1);
bo->userptr = true;
bo->cache_coherent = true;
bo->index = -1;
bo->idle = true;
return bo;
err_close:
drm_ioctl(bufmgr->fd, DRM_IOCTL_GEM_CLOSE, &bo->gem_handle);
err_free:
free(bo);
return NULL;
}
/**
* Returns a iris_bo wrapping the given buffer object handle.
*
* This can be used when one application needs to pass a buffer object
* to another.
*/
struct iris_bo *
iris_bo_gem_create_from_name(struct iris_bufmgr *bufmgr,
const char *name, unsigned int handle)
{
struct iris_bo *bo;
/* At the moment most applications only have a few named bo.
* For instance, in a DRI client only the render buffers passed
* between X and the client are named. And since X returns the
* alternating names for the front/back buffer a linear search
* provides a sufficiently fast match.
*/
mtx_lock(&bufmgr->lock);
bo = hash_find_bo(bufmgr->name_table, handle);
if (bo) {
iris_bo_reference(bo);
goto out;
}
struct drm_gem_open open_arg = { .name = handle };
int ret = drm_ioctl(bufmgr->fd, DRM_IOCTL_GEM_OPEN, &open_arg);
if (ret != 0) {
DBG("Couldn't reference %s handle 0x%08x: %s\n",
name, handle, strerror(errno));
bo = NULL;
goto out;
}
/* Now see if someone has used a prime handle to get this
* object from the kernel before by looking through the list
* again for a matching gem_handle
*/
bo = hash_find_bo(bufmgr->handle_table, open_arg.handle);
if (bo) {
iris_bo_reference(bo);
goto out;
}
bo = bo_calloc();
if (!bo)
goto out;
p_atomic_set(&bo->refcount, 1);
bo->size = open_arg.size;
bo->gtt_offset = 0;
bo->bufmgr = bufmgr;
bo->gem_handle = open_arg.handle;
bo->name = name;
bo->global_name = handle;
bo->reusable = false;
bo->external = true;
bo->kflags = EXEC_OBJECT_SUPPORTS_48B_ADDRESS | EXEC_OBJECT_PINNED;
bo->gtt_offset = vma_alloc(bufmgr, IRIS_MEMZONE_OTHER, bo->size, 1);
_mesa_hash_table_insert(bufmgr->handle_table, &bo->gem_handle, bo);
_mesa_hash_table_insert(bufmgr->name_table, &bo->global_name, bo);
struct drm_i915_gem_get_tiling get_tiling = { .handle = bo->gem_handle };
ret = drm_ioctl(bufmgr->fd, DRM_IOCTL_I915_GEM_GET_TILING, &get_tiling);
if (ret != 0)
goto err_unref;
bo->tiling_mode = get_tiling.tiling_mode;
bo->swizzle_mode = get_tiling.swizzle_mode;
/* XXX stride is unknown */
DBG("bo_create_from_handle: %d (%s)\n", handle, bo->name);
out:
mtx_unlock(&bufmgr->lock);
return bo;
err_unref:
bo_free(bo);
mtx_unlock(&bufmgr->lock);
return NULL;
}
static void
bo_free(struct iris_bo *bo)
{
struct iris_bufmgr *bufmgr = bo->bufmgr;
if (bo->map_cpu && !bo->userptr) {
VG_NOACCESS(bo->map_cpu, bo->size);
munmap(bo->map_cpu, bo->size);
}
if (bo->map_wc) {
VG_NOACCESS(bo->map_wc, bo->size);
munmap(bo->map_wc, bo->size);
}
if (bo->map_gtt) {
VG_NOACCESS(bo->map_gtt, bo->size);
munmap(bo->map_gtt, bo->size);
}
if (bo->external) {
struct hash_entry *entry;
if (bo->global_name) {
entry = _mesa_hash_table_search(bufmgr->name_table, &bo->global_name);
_mesa_hash_table_remove(bufmgr->name_table, entry);
}
entry = _mesa_hash_table_search(bufmgr->handle_table, &bo->gem_handle);
_mesa_hash_table_remove(bufmgr->handle_table, entry);
}
/* Close this object */
struct drm_gem_close close = { .handle = bo->gem_handle };
int ret = drm_ioctl(bufmgr->fd, DRM_IOCTL_GEM_CLOSE, &close);
if (ret != 0) {
DBG("DRM_IOCTL_GEM_CLOSE %d failed (%s): %s\n",
bo->gem_handle, bo->name, strerror(errno));
}
vma_free(bo->bufmgr, bo->gtt_offset, bo->size);
free(bo);
}
/** Frees all cached buffers significantly older than @time. */
static void
cleanup_bo_cache(struct iris_bufmgr *bufmgr, time_t time)
{
int i;
if (bufmgr->time == time)
return;
for (i = 0; i < bufmgr->num_buckets; i++) {
struct bo_cache_bucket *bucket = &bufmgr->cache_bucket[i];
list_for_each_entry_safe(struct iris_bo, bo, &bucket->head, head) {
if (time - bo->free_time <= 1)
break;
list_del(&bo->head);
bo_free(bo);
}
}
bufmgr->time = time;
}
static void
bo_unreference_final(struct iris_bo *bo, time_t time)
{
struct iris_bufmgr *bufmgr = bo->bufmgr;
struct bo_cache_bucket *bucket;
DBG("bo_unreference final: %d (%s)\n", bo->gem_handle, bo->name);
bucket = NULL;
if (bo->reusable)
bucket = bucket_for_size(bufmgr, bo->size);
/* Put the buffer into our internal cache for reuse if we can. */
if (bucket && iris_bo_madvise(bo, I915_MADV_DONTNEED)) {
bo->free_time = time;
bo->name = NULL;
list_addtail(&bo->head, &bucket->head);
} else {
bo_free(bo);
}
}
void
iris_bo_unreference(struct iris_bo *bo)
{
if (bo == NULL)
return;
assert(p_atomic_read(&bo->refcount) > 0);
if (atomic_add_unless(&bo->refcount, -1, 1)) {
struct iris_bufmgr *bufmgr = bo->bufmgr;
struct timespec time;
clock_gettime(CLOCK_MONOTONIC, &time);
mtx_lock(&bufmgr->lock);
if (p_atomic_dec_zero(&bo->refcount)) {
bo_unreference_final(bo, time.tv_sec);
cleanup_bo_cache(bufmgr, time.tv_sec);
}
mtx_unlock(&bufmgr->lock);
}
}
static void
bo_wait_with_stall_warning(struct pipe_debug_callback *dbg,
struct iris_bo *bo,
const char *action)
{
bool busy = dbg && !bo->idle;
double elapsed = unlikely(busy) ? -get_time() : 0.0;
iris_bo_wait_rendering(bo);
if (unlikely(busy)) {
elapsed += get_time();
if (elapsed > 1e-5) /* 0.01ms */ {
perf_debug(dbg, "%s a busy \"%s\" BO stalled and took %.03f ms.\n",
action, bo->name, elapsed * 1000);
}
}
}
static void
print_flags(unsigned flags)
{
if (flags & MAP_READ)
DBG("READ ");
if (flags & MAP_WRITE)
DBG("WRITE ");
if (flags & MAP_ASYNC)
DBG("ASYNC ");
if (flags & MAP_PERSISTENT)
DBG("PERSISTENT ");
if (flags & MAP_COHERENT)
DBG("COHERENT ");
if (flags & MAP_RAW)
DBG("RAW ");
DBG("\n");
}
static void *
iris_bo_map_cpu(struct pipe_debug_callback *dbg,
struct iris_bo *bo, unsigned flags)
{
struct iris_bufmgr *bufmgr = bo->bufmgr;
/* We disallow CPU maps for writing to non-coherent buffers, as the
* CPU map can become invalidated when a batch is flushed out, which
* can happen at unpredictable times. You should use WC maps instead.
*/
assert(bo->cache_coherent || !(flags & MAP_WRITE));
if (!bo->map_cpu) {
DBG("iris_bo_map_cpu: %d (%s)\n", bo->gem_handle, bo->name);
struct drm_i915_gem_mmap mmap_arg = {
.handle = bo->gem_handle,
.size = bo->size,
};
int ret = drm_ioctl(bufmgr->fd, DRM_IOCTL_I915_GEM_MMAP, &mmap_arg);
if (ret != 0) {
DBG("%s:%d: Error mapping buffer %d (%s): %s .\n",
__FILE__, __LINE__, bo->gem_handle, bo->name, strerror(errno));
return NULL;
}
void *map = (void *) (uintptr_t) mmap_arg.addr_ptr;
VG_DEFINED(map, bo->size);
if (p_atomic_cmpxchg(&bo->map_cpu, NULL, map)) {
VG_NOACCESS(map, bo->size);
munmap(map, bo->size);
}
}
assert(bo->map_cpu);
DBG("iris_bo_map_cpu: %d (%s) -> %p, ", bo->gem_handle, bo->name,
bo->map_cpu);
print_flags(flags);
if (!(flags & MAP_ASYNC)) {
bo_wait_with_stall_warning(dbg, bo, "CPU mapping");
}
if (!bo->cache_coherent && !bo->bufmgr->has_llc) {
/* If we're reusing an existing CPU mapping, the CPU caches may
* contain stale data from the last time we read from that mapping.
* (With the BO cache, it might even be data from a previous buffer!)
* Even if it's a brand new mapping, the kernel may have zeroed the
* buffer via CPU writes.
*
* We need to invalidate those cachelines so that we see the latest
* contents, and so long as we only read from the CPU mmap we do not
* need to write those cachelines back afterwards.
*
* On LLC, the emprical evidence suggests that writes from the GPU
* that bypass the LLC (i.e. for scanout) do *invalidate* the CPU
* cachelines. (Other reads, such as the display engine, bypass the
* LLC entirely requiring us to keep dirty pixels for the scanout
* out of any cache.)
*/
gen_invalidate_range(bo->map_cpu, bo->size);
}
return bo->map_cpu;
}
static void *
iris_bo_map_wc(struct pipe_debug_callback *dbg,
struct iris_bo *bo, unsigned flags)
{
struct iris_bufmgr *bufmgr = bo->bufmgr;
if (!bo->map_wc) {
DBG("iris_bo_map_wc: %d (%s)\n", bo->gem_handle, bo->name);
struct drm_i915_gem_mmap mmap_arg = {
.handle = bo->gem_handle,
.size = bo->size,
.flags = I915_MMAP_WC,
};
int ret = drm_ioctl(bufmgr->fd, DRM_IOCTL_I915_GEM_MMAP, &mmap_arg);
if (ret != 0) {
DBG("%s:%d: Error mapping buffer %d (%s): %s .\n",
__FILE__, __LINE__, bo->gem_handle, bo->name, strerror(errno));
return NULL;
}
void *map = (void *) (uintptr_t) mmap_arg.addr_ptr;
VG_DEFINED(map, bo->size);
if (p_atomic_cmpxchg(&bo->map_wc, NULL, map)) {
VG_NOACCESS(map, bo->size);
munmap(map, bo->size);
}
}
assert(bo->map_wc);
DBG("iris_bo_map_wc: %d (%s) -> %p\n", bo->gem_handle, bo->name, bo->map_wc);
print_flags(flags);
if (!(flags & MAP_ASYNC)) {
bo_wait_with_stall_warning(dbg, bo, "WC mapping");
}
return bo->map_wc;
}
/**
* Perform an uncached mapping via the GTT.
*
* Write access through the GTT is not quite fully coherent. On low power
* systems especially, like modern Atoms, we can observe reads from RAM before
* the write via GTT has landed. A write memory barrier that flushes the Write
* Combining Buffer (i.e. sfence/mfence) is not sufficient to order the later
* read after the write as the GTT write suffers a small delay through the GTT
* indirection. The kernel uses an uncached mmio read to ensure the GTT write
* is ordered with reads (either by the GPU, WB or WC) and unconditionally
* flushes prior to execbuf submission. However, if we are not informing the
* kernel about our GTT writes, it will not flush before earlier access, such
* as when using the cmdparser. Similarly, we need to be careful if we should
* ever issue a CPU read immediately following a GTT write.
*
* Telling the kernel about write access also has one more important
* side-effect. Upon receiving notification about the write, it cancels any
* scanout buffering for FBC/PSR and friends. Later FBC/PSR is then flushed by
* either SW_FINISH or DIRTYFB. The presumption is that we never write to the
* actual scanout via a mmaping, only to a backbuffer and so all the FBC/PSR
* tracking is handled on the buffer exchange instead.
*/
static void *
iris_bo_map_gtt(struct pipe_debug_callback *dbg,
struct iris_bo *bo, unsigned flags)
{
struct iris_bufmgr *bufmgr = bo->bufmgr;
/* Get a mapping of the buffer if we haven't before. */
if (bo->map_gtt == NULL) {
DBG("bo_map_gtt: mmap %d (%s)\n", bo->gem_handle, bo->name);
struct drm_i915_gem_mmap_gtt mmap_arg = { .handle = bo->gem_handle };
/* Get the fake offset back... */
int ret = drm_ioctl(bufmgr->fd, DRM_IOCTL_I915_GEM_MMAP_GTT, &mmap_arg);
if (ret != 0) {
DBG("%s:%d: Error preparing buffer map %d (%s): %s .\n",
__FILE__, __LINE__, bo->gem_handle, bo->name, strerror(errno));
return NULL;
}
/* and mmap it. */
void *map = mmap(0, bo->size, PROT_READ | PROT_WRITE,
MAP_SHARED, bufmgr->fd, mmap_arg.offset);
if (map == MAP_FAILED) {
DBG("%s:%d: Error mapping buffer %d (%s): %s .\n",
__FILE__, __LINE__, bo->gem_handle, bo->name, strerror(errno));
return NULL;
}
/* We don't need to use VALGRIND_MALLOCLIKE_BLOCK because Valgrind will
* already intercept this mmap call. However, for consistency between
* all the mmap paths, we mark the pointer as defined now and mark it
* as inaccessible afterwards.
*/
VG_DEFINED(map, bo->size);
if (p_atomic_cmpxchg(&bo->map_gtt, NULL, map)) {
VG_NOACCESS(map, bo->size);
munmap(map, bo->size);
}
}
assert(bo->map_gtt);
DBG("bo_map_gtt: %d (%s) -> %p, ", bo->gem_handle, bo->name, bo->map_gtt);
print_flags(flags);
if (!(flags & MAP_ASYNC)) {
bo_wait_with_stall_warning(dbg, bo, "GTT mapping");
}
return bo->map_gtt;
}
static bool
can_map_cpu(struct iris_bo *bo, unsigned flags)
{
if (bo->cache_coherent)
return true;
/* Even if the buffer itself is not cache-coherent (such as a scanout), on
* an LLC platform reads always are coherent (as they are performed via the
* central system agent). It is just the writes that we need to take special
* care to ensure that land in main memory and not stick in the CPU cache.
*/
if (!(flags & MAP_WRITE) && bo->bufmgr->has_llc)
return true;
/* If PERSISTENT or COHERENT are set, the mmapping needs to remain valid
* across batch flushes where the kernel will change cache domains of the
* bo, invalidating continued access to the CPU mmap on non-LLC device.
*
* Similarly, ASYNC typically means that the buffer will be accessed via
* both the CPU and the GPU simultaneously. Batches may be executed that
* use the BO even while it is mapped. While OpenGL technically disallows
* most drawing while non-persistent mappings are active, we may still use
* the GPU for blits or other operations, causing batches to happen at
* inconvenient times.
*
* If RAW is set, we expect the caller to be able to handle a WC buffer
* more efficiently than the involuntary clflushes.
*/
if (flags & (MAP_PERSISTENT | MAP_COHERENT | MAP_ASYNC | MAP_RAW))
return false;
return !(flags & MAP_WRITE);
}
void *
iris_bo_map(struct pipe_debug_callback *dbg,
struct iris_bo *bo, unsigned flags)
{
if (bo->tiling_mode != I915_TILING_NONE && !(flags & MAP_RAW))
return iris_bo_map_gtt(dbg, bo, flags);
void *map;
if (can_map_cpu(bo, flags))
map = iris_bo_map_cpu(dbg, bo, flags);
else
map = iris_bo_map_wc(dbg, bo, flags);
/* Allow the attempt to fail by falling back to the GTT where necessary.
*
* Not every buffer can be mmaped directly using the CPU (or WC), for
* example buffers that wrap stolen memory or are imported from other
* devices. For those, we have little choice but to use a GTT mmapping.
* However, if we use a slow GTT mmapping for reads where we expected fast
* access, that order of magnitude difference in throughput will be clearly
* expressed by angry users.
*
* We skip MAP_RAW because we want to avoid map_gtt's fence detiling.
*/
if (!map && !(flags & MAP_RAW)) {
perf_debug(dbg, "Fallback GTT mapping for %s with access flags %x\n",
bo->name, flags);
map = iris_bo_map_gtt(dbg, bo, flags);
}
return map;
}
/** Waits for all GPU rendering with the object to have completed. */
void
iris_bo_wait_rendering(struct iris_bo *bo)
{
/* We require a kernel recent enough for WAIT_IOCTL support.
* See intel_init_bufmgr()
*/
iris_bo_wait(bo, -1);
}
/**
* Waits on a BO for the given amount of time.
*
* @bo: buffer object to wait for
* @timeout_ns: amount of time to wait in nanoseconds.
* If value is less than 0, an infinite wait will occur.
*
* Returns 0 if the wait was successful ie. the last batch referencing the
* object has completed within the allotted time. Otherwise some negative return
* value describes the error. Of particular interest is -ETIME when the wait has
* failed to yield the desired result.
*
* Similar to iris_bo_wait_rendering except a timeout parameter allows
* the operation to give up after a certain amount of time. Another subtle
* difference is the internal locking semantics are different (this variant does
* not hold the lock for the duration of the wait). This makes the wait subject
* to a larger userspace race window.
*
* The implementation shall wait until the object is no longer actively
* referenced within a batch buffer at the time of the call. The wait will
* not guarantee that the buffer is re-issued via another thread, or an flinked
* handle. Userspace must make sure this race does not occur if such precision
* is important.
*
* Note that some kernels have broken the inifite wait for negative values
* promise, upgrade to latest stable kernels if this is the case.
*/
int
iris_bo_wait(struct iris_bo *bo, int64_t timeout_ns)
{
struct iris_bufmgr *bufmgr = bo->bufmgr;
/* If we know it's idle, don't bother with the kernel round trip */
if (bo->idle && !bo->external)
return 0;
struct drm_i915_gem_wait wait = {
.bo_handle = bo->gem_handle,
.timeout_ns = timeout_ns,
};
int ret = drm_ioctl(bufmgr->fd, DRM_IOCTL_I915_GEM_WAIT, &wait);
if (ret != 0)
return -errno;
bo->idle = true;
return ret;
}
void
iris_bufmgr_destroy(struct iris_bufmgr *bufmgr)
{
mtx_destroy(&bufmgr->lock);
/* Free any cached buffer objects we were going to reuse */
for (int i = 0; i < bufmgr->num_buckets; i++) {
struct bo_cache_bucket *bucket = &bufmgr->cache_bucket[i];
list_for_each_entry_safe(struct iris_bo, bo, &bucket->head, head) {
list_del(&bo->head);
bo_free(bo);
}
}
_mesa_hash_table_destroy(bufmgr->name_table, NULL);
_mesa_hash_table_destroy(bufmgr->handle_table, NULL);
for (int z = 0; z < IRIS_MEMZONE_COUNT; z++) {
if (z != IRIS_MEMZONE_BINDER)
util_vma_heap_finish(&bufmgr->vma_allocator[z]);
}
free(bufmgr);
}
static int
bo_set_tiling_internal(struct iris_bo *bo, uint32_t tiling_mode,
uint32_t stride)
{
struct iris_bufmgr *bufmgr = bo->bufmgr;
struct drm_i915_gem_set_tiling set_tiling;
int ret;
if (bo->global_name == 0 &&
tiling_mode == bo->tiling_mode && stride == bo->stride)
return 0;
memset(&set_tiling, 0, sizeof(set_tiling));
do {
/* set_tiling is slightly broken and overwrites the
* input on the error path, so we have to open code
* drm_ioctl.
*/
set_tiling.handle = bo->gem_handle;
set_tiling.tiling_mode = tiling_mode;
set_tiling.stride = stride;
ret = ioctl(bufmgr->fd, DRM_IOCTL_I915_GEM_SET_TILING, &set_tiling);
} while (ret == -1 && (errno == EINTR || errno == EAGAIN));
if (ret == -1)
return -errno;
bo->tiling_mode = set_tiling.tiling_mode;
bo->swizzle_mode = set_tiling.swizzle_mode;
bo->stride = set_tiling.stride;
return 0;
}
int
iris_bo_get_tiling(struct iris_bo *bo, uint32_t *tiling_mode,
uint32_t *swizzle_mode)
{
*tiling_mode = bo->tiling_mode;
*swizzle_mode = bo->swizzle_mode;
return 0;
}
struct iris_bo *
iris_bo_import_dmabuf(struct iris_bufmgr *bufmgr, int prime_fd)
{
uint32_t handle;
struct iris_bo *bo;
mtx_lock(&bufmgr->lock);
int ret = drmPrimeFDToHandle(bufmgr->fd, prime_fd, &handle);
if (ret) {
DBG("import_dmabuf: failed to obtain handle from fd: %s\n",
strerror(errno));
mtx_unlock(&bufmgr->lock);
return NULL;
}
/*
* See if the kernel has already returned this buffer to us. Just as
* for named buffers, we must not create two bo's pointing at the same
* kernel object
*/
bo = hash_find_bo(bufmgr->handle_table, handle);
if (bo) {
iris_bo_reference(bo);
goto out;
}
bo = bo_calloc();
if (!bo)
goto out;
p_atomic_set(&bo->refcount, 1);
/* Determine size of bo. The fd-to-handle ioctl really should
* return the size, but it doesn't. If we have kernel 3.12 or
* later, we can lseek on the prime fd to get the size. Older
* kernels will just fail, in which case we fall back to the
* provided (estimated or guess size). */
ret = lseek(prime_fd, 0, SEEK_END);
if (ret != -1)
bo->size = ret;
bo->bufmgr = bufmgr;
bo->gem_handle = handle;
_mesa_hash_table_insert(bufmgr->handle_table, &bo->gem_handle, bo);
bo->name = "prime";
bo->reusable = false;
bo->external = true;
bo->kflags = EXEC_OBJECT_SUPPORTS_48B_ADDRESS | EXEC_OBJECT_PINNED;
bo->gtt_offset = vma_alloc(bufmgr, IRIS_MEMZONE_OTHER, bo->size, 1);
struct drm_i915_gem_get_tiling get_tiling = { .handle = bo->gem_handle };
if (drm_ioctl(bufmgr->fd, DRM_IOCTL_I915_GEM_GET_TILING, &get_tiling))
goto err;
bo->tiling_mode = get_tiling.tiling_mode;
bo->swizzle_mode = get_tiling.swizzle_mode;
/* XXX stride is unknown */
out:
mtx_unlock(&bufmgr->lock);
return bo;
err:
bo_free(bo);
mtx_unlock(&bufmgr->lock);
return NULL;
}
static void
iris_bo_make_external_locked(struct iris_bo *bo)
{
if (!bo->external) {
_mesa_hash_table_insert(bo->bufmgr->handle_table, &bo->gem_handle, bo);
bo->external = true;
}
}
static void
iris_bo_make_external(struct iris_bo *bo)
{
struct iris_bufmgr *bufmgr = bo->bufmgr;
if (bo->external)
return;
mtx_lock(&bufmgr->lock);
iris_bo_make_external_locked(bo);
mtx_unlock(&bufmgr->lock);
}
int
iris_bo_export_dmabuf(struct iris_bo *bo, int *prime_fd)
{
struct iris_bufmgr *bufmgr = bo->bufmgr;
iris_bo_make_external(bo);
if (drmPrimeHandleToFD(bufmgr->fd, bo->gem_handle,
DRM_CLOEXEC, prime_fd) != 0)
return -errno;
bo->reusable = false;
return 0;
}
uint32_t
iris_bo_export_gem_handle(struct iris_bo *bo)
{
iris_bo_make_external(bo);
return bo->gem_handle;
}
int
iris_bo_flink(struct iris_bo *bo, uint32_t *name)
{
struct iris_bufmgr *bufmgr = bo->bufmgr;
if (!bo->global_name) {
struct drm_gem_flink flink = { .handle = bo->gem_handle };
if (drm_ioctl(bufmgr->fd, DRM_IOCTL_GEM_FLINK, &flink))
return -errno;
mtx_lock(&bufmgr->lock);
if (!bo->global_name) {
iris_bo_make_external_locked(bo);
bo->global_name = flink.name;
_mesa_hash_table_insert(bufmgr->name_table, &bo->global_name, bo);
}
mtx_unlock(&bufmgr->lock);
bo->reusable = false;
}
*name = bo->global_name;
return 0;
}
static void
add_bucket(struct iris_bufmgr *bufmgr, int size)
{
unsigned int i = bufmgr->num_buckets;
assert(i < ARRAY_SIZE(bufmgr->cache_bucket));
list_inithead(&bufmgr->cache_bucket[i].head);
bufmgr->cache_bucket[i].size = size;
bufmgr->num_buckets++;
assert(bucket_for_size(bufmgr, size) == &bufmgr->cache_bucket[i]);
assert(bucket_for_size(bufmgr, size - 2048) == &bufmgr->cache_bucket[i]);
assert(bucket_for_size(bufmgr, size + 1) != &bufmgr->cache_bucket[i]);
}
static void
init_cache_buckets(struct iris_bufmgr *bufmgr)
{
uint64_t size, cache_max_size = 64 * 1024 * 1024;
/* OK, so power of two buckets was too wasteful of memory.
* Give 3 other sizes between each power of two, to hopefully
* cover things accurately enough. (The alternative is
* probably to just go for exact matching of sizes, and assume
* that for things like composited window resize the tiled
* width/height alignment and rounding of sizes to pages will
* get us useful cache hit rates anyway)
*/
add_bucket(bufmgr, PAGE_SIZE);
add_bucket(bufmgr, PAGE_SIZE * 2);
add_bucket(bufmgr, PAGE_SIZE * 3);
/* Initialize the linked lists for BO reuse cache. */
for (size = 4 * PAGE_SIZE; size <= cache_max_size; size *= 2) {
add_bucket(bufmgr, size);
add_bucket(bufmgr, size + size * 1 / 4);
add_bucket(bufmgr, size + size * 2 / 4);
add_bucket(bufmgr, size + size * 3 / 4);
}
}
/*
* Copies in the user's binning command list and generates the validated bin
* CL, along with associated data (shader records, uniforms).
*/
static int
vc4_get_bcl(struct drm_device *dev, struct vc4_exec_info *exec)
{
struct drm_vc4_submit_cl *args = exec->args;
void *temp = NULL;
void *bin;
int ret = 0;
uint32_t bin_offset = 0;
uint32_t shader_rec_offset = roundup(bin_offset + args->bin_cl_size,
16);
uint32_t uniforms_offset = shader_rec_offset + args->shader_rec_size;
uint32_t exec_size = uniforms_offset + args->uniforms_size;
uint32_t temp_size = exec_size + (sizeof(struct vc4_shader_state) *
args->shader_rec_count);
if (uniforms_offset < shader_rec_offset ||
exec_size < uniforms_offset ||
args->shader_rec_count >= (UINT_MAX /
sizeof(struct vc4_shader_state)) ||
temp_size < exec_size) {
DRM_ERROR("overflow in exec arguments\n");
goto fail;
}
/* Allocate space where we'll store the copied in user command lists
* and shader records.
*
* We don't just copy directly into the BOs because we need to
* read the contents back for validation, and I think the
* bo->vaddr is uncached access.
*/
temp = kmalloc(temp_size, GFP_KERNEL);
if (!temp) {
DRM_ERROR("Failed to allocate storage for copying "
"in bin/render CLs.\n");
ret = -ENOMEM;
goto fail;
}
bin = temp + bin_offset;
exec->shader_rec_u = temp + shader_rec_offset;
exec->uniforms_u = temp + uniforms_offset;
exec->shader_state = temp + exec_size;
exec->shader_state_size = args->shader_rec_count;
ret = copy_from_user(bin,
(void __user *)(uintptr_t)args->bin_cl,
args->bin_cl_size);
if (ret) {
DRM_ERROR("Failed to copy in bin cl\n");
goto fail;
}
ret = copy_from_user(exec->shader_rec_u,
(void __user *)(uintptr_t)args->shader_rec,
args->shader_rec_size);
if (ret) {
DRM_ERROR("Failed to copy in shader recs\n");
goto fail;
}
ret = copy_from_user(exec->uniforms_u,
(void __user *)(uintptr_t)args->uniforms,
args->uniforms_size);
if (ret) {
DRM_ERROR("Failed to copy in uniforms cl\n");
goto fail;
}
exec->exec_bo = drm_gem_cma_create(dev, exec_size);
#if 0
if (IS_ERR(exec->exec_bo)) {
DRM_ERROR("Couldn't allocate BO for exec\n");
ret = PTR_ERR(exec->exec_bo);
exec->exec_bo = NULL;
goto fail;
}
#endif
list_addtail(&to_vc4_bo(&exec->exec_bo->base)->unref_head,
&exec->unref_list);
exec->ct0ca = exec->exec_bo->paddr + bin_offset;
exec->bin_u = bin;
exec->shader_rec_v = exec->exec_bo->vaddr + shader_rec_offset;
exec->shader_rec_p = exec->exec_bo->paddr + shader_rec_offset;
exec->shader_rec_size = args->shader_rec_size;
exec->uniforms_v = exec->exec_bo->vaddr + uniforms_offset;
exec->uniforms_p = exec->exec_bo->paddr + uniforms_offset;
exec->uniforms_size = args->uniforms_size;
ret = vc4_validate_bin_cl(dev,
exec->exec_bo->vaddr + bin_offset,
bin,
exec);
if (ret)
goto fail;
ret = vc4_validate_shader_recs(dev, exec);
fail:
kfree(temp);
return ret;
}
void ppir_node_replace_pred(ppir_dep *dep, ppir_node *new_pred)
{
list_del(&dep->succ_link);
dep->pred = new_pred;
list_addtail(&dep->succ_link, &new_pred->succ_list);
}
