freedreno: use drm_mmap/drm_munmap wrappers
[platform/upstream/libdrm.git] / freedreno / freedreno_bo.c
index 4f566e1..dacb34c 100644 (file)
 #include "freedreno_drmif.h"
 #include "freedreno_priv.h"
 
-#include <linux/fb.h>
+static pthread_mutex_t table_lock = PTHREAD_MUTEX_INITIALIZER;
 
+static void bo_del(struct fd_bo *bo);
+
+/* set buffer name, and add to table, call w/ table_lock held: */
+static void set_name(struct fd_bo *bo, uint32_t name)
+{
+       bo->name = name;
+       /* add ourself into the handle table: */
+       drmHashInsert(bo->dev->name_table, name, bo);
+}
+
+/* lookup a buffer, call w/ table_lock held: */
+static struct fd_bo * lookup_bo(void *tbl, uint32_t key)
+{
+       struct fd_bo *bo = NULL;
+       if (!drmHashLookup(tbl, key, (void **)&bo)) {
+               /* found, incr refcnt and return: */
+               bo = fd_bo_ref(bo);
+       }
+       return bo;
+}
+
+/* allocate a new buffer object, call w/ table_lock held */
 static struct fd_bo * bo_from_handle(struct fd_device *dev,
                uint32_t size, uint32_t handle)
 {
-       unsigned i;
-       struct fd_bo *bo = calloc(1, sizeof(*bo));
-       if (!bo)
+       struct fd_bo *bo;
+
+       bo = dev->funcs->bo_from_handle(dev, size, handle);
+       if (!bo) {
+               struct drm_gem_close req = {
+                               .handle = handle,
+               };
+               drmIoctl(dev->fd, DRM_IOCTL_GEM_CLOSE, &req);
                return NULL;
-       bo->dev = dev;
+       }
+       bo->dev = fd_device_ref(dev);
        bo->size = size;
        bo->handle = handle;
        atomic_set(&bo->refcnt, 1);
-       for (i = 0; i < ARRAY_SIZE(bo->list); i++)
-               list_inithead(&bo->list[i]);
+       list_inithead(&bo->list);
+       /* add ourself into the handle table: */
+       drmHashInsert(dev->handle_table, handle, bo);
        return bo;
 }
 
-static int set_memtype(struct fd_bo *bo, uint32_t flags)
+/* Frees older cached buffers.  Called under table_lock */
+void fd_cleanup_bo_cache(struct fd_device *dev, time_t time)
 {
-       struct drm_kgsl_gem_memtype req = {
-                       .handle = bo->handle,
-                       .type = flags & DRM_FREEDRENO_GEM_TYPE_MEM_MASK,
-       };
+       int i;
+
+       if (dev->time == time)
+               return;
+
+       for (i = 0; i < dev->num_buckets; i++) {
+               struct fd_bo_bucket *bucket = &dev->cache_bucket[i];
+               struct fd_bo *bo;
+
+               while (!LIST_IS_EMPTY(&bucket->list)) {
+                       bo = LIST_ENTRY(struct fd_bo, bucket->list.next, list);
+
+                       /* keep things in cache for at least 1 second: */
+                       if (time && ((time - bo->free_time) <= 1))
+                               break;
+
+                       list_del(&bo->list);
+                       bo_del(bo);
+               }
+       }
 
-       return drmCommandWrite(bo->dev->fd, DRM_KGSL_GEM_SETMEMTYPE,
-                       &req, sizeof(req));
+       dev->time = time;
 }
 
-static int bo_alloc(struct fd_bo *bo)
+static struct fd_bo_bucket * get_bucket(struct fd_device *dev, uint32_t size)
 {
-       if (!bo->offset) {
-               struct drm_kgsl_gem_alloc req = {
-                               .handle = bo->handle,
-               };
-               int ret;
+       int i;
 
-               /* if the buffer is already backed by pages then this
-                * doesn't actually do anything (other than giving us
-                * the offset)
-                */
-               ret = drmCommandWriteRead(bo->dev->fd, DRM_KGSL_GEM_ALLOC,
-                               &req, sizeof(req));
-               if (ret) {
-                       ERROR_MSG("alloc failed: %s", strerror(errno));
-                       return ret;
+       /* hmm, this is what intel does, but I suppose we could calculate our
+        * way to the correct bucket size rather than looping..
+        */
+       for (i = 0; i < dev->num_buckets; i++) {
+               struct fd_bo_bucket *bucket = &dev->cache_bucket[i];
+               if (bucket->size >= size) {
+                       return bucket;
                }
-
-               bo->offset = req.offset;
        }
-       return 0;
+
+       return NULL;
 }
 
-struct fd_bo * fd_bo_new(struct fd_device *dev,
-               uint32_t size, uint32_t flags)
+static int is_idle(struct fd_bo *bo)
+{
+       return fd_bo_cpu_prep(bo, NULL,
+                       DRM_FREEDRENO_PREP_READ |
+                       DRM_FREEDRENO_PREP_WRITE |
+                       DRM_FREEDRENO_PREP_NOSYNC) == 0;
+}
+
+static struct fd_bo *find_in_bucket(struct fd_device *dev,
+               struct fd_bo_bucket *bucket, uint32_t flags)
 {
-       struct drm_kgsl_gem_create req = {
-                       .size = ALIGN(size, 4096),
-       };
        struct fd_bo *bo = NULL;
 
-       if (drmCommandWriteRead(dev->fd, DRM_KGSL_GEM_CREATE,
-                       &req, sizeof(req))) {
-               return NULL;
+       /* TODO .. if we had an ALLOC_FOR_RENDER flag like intel, we could
+        * skip the busy check.. if it is only going to be a render target
+        * then we probably don't need to stall..
+        *
+        * NOTE that intel takes ALLOC_FOR_RENDER bo's from the list tail
+        * (MRU, since likely to be in GPU cache), rather than head (LRU)..
+        */
+       pthread_mutex_lock(&table_lock);
+       while (!LIST_IS_EMPTY(&bucket->list)) {
+               bo = LIST_ENTRY(struct fd_bo, bucket->list.next, list);
+               if (0 /* TODO: if madvise tells us bo is gone... */) {
+                       list_del(&bo->list);
+                       bo_del(bo);
+                       bo = NULL;
+                       continue;
+               }
+               /* TODO check for compatible flags? */
+               if (is_idle(bo)) {
+                       list_del(&bo->list);
+                       break;
+               }
+               bo = NULL;
+               break;
        }
+       pthread_mutex_unlock(&table_lock);
+
+       return bo;
+}
 
-       bo = bo_from_handle(dev, size, req.handle);
-       if (!bo) {
-               goto fail;
-       }
 
-       if (set_memtype(bo, flags)) {
-               goto fail;
+drm_public struct fd_bo *
+fd_bo_new(struct fd_device *dev, uint32_t size, uint32_t flags)
+{
+       struct fd_bo *bo = NULL;
+       struct fd_bo_bucket *bucket;
+       uint32_t handle;
+       int ret;
+
+       size = ALIGN(size, 4096);
+       bucket = get_bucket(dev, size);
+
+       /* see if we can be green and recycle: */
+       if (bucket) {
+               size = bucket->size;
+               bo = find_in_bucket(dev, bucket, flags);
+               if (bo) {
+                       atomic_set(&bo->refcnt, 1);
+                       fd_device_ref(bo->dev);
+                       return bo;
+               }
        }
 
+       ret = dev->funcs->bo_new_handle(dev, size, flags, &handle);
+       if (ret)
+               return NULL;
+
+       pthread_mutex_lock(&table_lock);
+       bo = bo_from_handle(dev, size, handle);
+       bo->bo_reuse = 1;
+       pthread_mutex_unlock(&table_lock);
+
        return bo;
-fail:
-       if (bo)
-               fd_bo_del(bo);
-       return NULL;
 }
 
-/* don't use this... it is just needed to get a bo from the
- * framebuffer (pre-dmabuf)
- */
-struct fd_bo * fd_bo_from_fbdev(struct fd_pipe *pipe,
-               int fbfd, uint32_t size)
+drm_public struct fd_bo *
+fd_bo_from_handle(struct fd_device *dev, uint32_t handle, uint32_t size)
+{
+       struct fd_bo *bo = NULL;
+
+       pthread_mutex_lock(&table_lock);
+       bo = bo_from_handle(dev, size, handle);
+       pthread_mutex_unlock(&table_lock);
+
+       return bo;
+}
+
+drm_public struct fd_bo *
+fd_bo_from_dmabuf(struct fd_device *dev, int fd)
 {
-       struct drm_kgsl_gem_create_fd req = {
-                       .fd = fbfd,
+       struct drm_prime_handle req = {
+                       .fd = fd,
        };
-       struct fd_bo *bo;
+       int ret, size;
 
-       if (drmCommandWriteRead(pipe->dev->fd, DRM_KGSL_GEM_CREATE_FD,
-                       &req, sizeof(req))) {
+       ret = drmIoctl(dev->fd, DRM_IOCTL_PRIME_FD_TO_HANDLE, &req);
+       if (ret) {
                return NULL;
        }
 
-       bo = bo_from_handle(pipe->dev, size, req.handle);
-
-       /* this is fugly, but works around a bug in the kernel..
-        * priv->memdesc.size never gets set, so getbufinfo ioctl
-        * thinks the buffer hasn't be allocate and fails
-        */
-       if (bo && !fd_bo_gpuaddr(bo, 0)) {
-               void *fbmem = mmap(NULL, size, PROT_READ | PROT_WRITE,
-                               MAP_SHARED, fbfd, 0);
-               struct kgsl_map_user_mem req = {
-                               .memtype = KGSL_USER_MEM_TYPE_ADDR,
-                               .len     = size,
-                               .offset  = 0,
-                               .hostptr = (unsigned long)fbmem,
-               };
-               int ret;
-               ret = ioctl(pipe->fd, IOCTL_KGSL_MAP_USER_MEM, &req);
-               if (ret) {
-                       ERROR_MSG("mapping user mem failed: %s",
-                                       strerror(errno));
-                       goto fail;
-               }
-               bo->gpuaddr = req.gpuaddr;
-               bo->map = fbmem;
-       }
+       /* hmm, would be nice if we had a way to figure out the size.. */
+       size = 0;
 
-       return bo;
-fail:
-       if (bo)
-               fd_bo_del(bo);
-       return NULL;
+       return fd_bo_from_handle(dev, req.handle, size);
 }
 
-struct fd_bo * fd_bo_from_name(struct fd_device *dev, uint32_t name)
+drm_public struct fd_bo * fd_bo_from_name(struct fd_device *dev, uint32_t name)
 {
        struct drm_gem_open req = {
                        .name = name,
        };
+       struct fd_bo *bo;
+
+       pthread_mutex_lock(&table_lock);
+
+       /* check name table first, to see if bo is already open: */
+       bo = lookup_bo(dev->name_table, name);
+       if (bo)
+               goto out_unlock;
 
        if (drmIoctl(dev->fd, DRM_IOCTL_GEM_OPEN, &req)) {
-               return NULL;
+               ERROR_MSG("gem-open failed: %s", strerror(errno));
+               goto out_unlock;
        }
 
-       return bo_from_handle(dev, req.size, req.handle);
+       bo = lookup_bo(dev->handle_table, req.handle);
+       if (bo)
+               goto out_unlock;
+
+       bo = bo_from_handle(dev, req.size, req.handle);
+       if (bo)
+               set_name(bo, name);
+
+out_unlock:
+       pthread_mutex_unlock(&table_lock);
+
+       return bo;
 }
 
-struct fd_bo * fd_bo_ref(struct fd_bo *bo)
+drm_public struct fd_bo * fd_bo_ref(struct fd_bo *bo)
 {
        atomic_inc(&bo->refcnt);
        return bo;
 }
 
-void fd_bo_del(struct fd_bo *bo)
+drm_public void fd_bo_del(struct fd_bo *bo)
 {
+       struct fd_device *dev = bo->dev;
+
        if (!atomic_dec_and_test(&bo->refcnt))
                return;
 
+       if (bo->fd) {
+               close(bo->fd);
+               bo->fd = 0;
+       }
+
+       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);
+}
+
+/* Called under table_lock */
+static void bo_del(struct fd_bo *bo)
+{
        if (bo->map)
-               munmap(bo->map, bo->size);
+               drm_munmap(bo->map, bo->size);
+
+       /* TODO probably bo's in bucket list get removed from
+        * handle table??
+        */
 
        if (bo->handle) {
                struct drm_gem_close req = {
                                .handle = bo->handle,
                };
+               drmHashDelete(bo->dev->handle_table, bo->handle);
+               if (bo->name)
+                       drmHashDelete(bo->dev->name_table, bo->name);
                drmIoctl(bo->dev->fd, DRM_IOCTL_GEM_CLOSE, &req);
        }
 
-       free(bo);
+       bo->funcs->destroy(bo);
 }
 
-int fd_bo_get_name(struct fd_bo *bo, uint32_t *name)
+drm_public int fd_bo_get_name(struct fd_bo *bo, uint32_t *name)
 {
        if (!bo->name) {
                struct drm_gem_flink req = {
@@ -210,7 +344,9 @@ int fd_bo_get_name(struct fd_bo *bo, uint32_t *name)
                        return ret;
                }
 
-               bo->name = req.name;
+               pthread_mutex_lock(&table_lock);
+               set_name(bo, req.name);
+               pthread_mutex_unlock(&table_lock);
        }
 
        *name = bo->name;
@@ -218,28 +354,48 @@ int fd_bo_get_name(struct fd_bo *bo, uint32_t *name)
        return 0;
 }
 
-uint32_t fd_bo_handle(struct fd_bo *bo)
+drm_public uint32_t fd_bo_handle(struct fd_bo *bo)
 {
        return bo->handle;
 }
 
-uint32_t fd_bo_size(struct fd_bo *bo)
+drm_public int fd_bo_dmabuf(struct fd_bo *bo)
+{
+       if (!bo->fd) {
+               struct drm_prime_handle req = {
+                               .handle = bo->handle,
+                               .flags = DRM_CLOEXEC,
+               };
+               int ret;
+
+               ret = drmIoctl(bo->dev->fd, DRM_IOCTL_PRIME_HANDLE_TO_FD, &req);
+               if (ret) {
+                       return ret;
+               }
+
+               bo->fd = req.fd;
+       }
+       return dup(bo->fd);
+}
+
+drm_public uint32_t fd_bo_size(struct fd_bo *bo)
 {
        return bo->size;
 }
 
-void * fd_bo_map(struct fd_bo *bo)
+drm_public void * fd_bo_map(struct fd_bo *bo)
 {
        if (!bo->map) {
+               uint64_t offset;
                int ret;
 
-               ret = bo_alloc(bo);
+               ret = bo->funcs->offset(bo, &offset);
                if (ret) {
                        return NULL;
                }
 
-               bo->map = mmap(0, bo->size, PROT_READ | PROT_WRITE, MAP_SHARED,
-                               bo->dev->fd, bo->offset);
+               bo->map = drm_mmap(0, bo->size, PROT_READ | PROT_WRITE, MAP_SHARED,
+                               bo->dev->fd, offset);
                if (bo->map == MAP_FAILED) {
                        ERROR_MSG("mmap failed: %s", strerror(errno));
                        bo->map = NULL;
@@ -248,27 +404,13 @@ void * fd_bo_map(struct fd_bo *bo)
        return bo->map;
 }
 
-uint32_t fd_bo_gpuaddr(struct fd_bo *bo, uint32_t offset)
+/* a bit odd to take the pipe as an arg, but it's a, umm, quirk of kgsl.. */
+drm_public int fd_bo_cpu_prep(struct fd_bo *bo, struct fd_pipe *pipe, uint32_t op)
 {
-       if (!bo->gpuaddr) {
-               struct drm_kgsl_gem_bufinfo req = {
-                               .handle = bo->handle,
-               };
-               int ret;
-
-               ret = bo_alloc(bo);
-               if (ret) {
-                       return ret;
-               }
-
-               ret = drmCommandWriteRead(bo->dev->fd, DRM_KGSL_GEM_GET_BUFINFO,
-                               &req, sizeof(req));
-               if (ret) {
-                       ERROR_MSG("get bufinfo failed: %s", strerror(errno));
-                       return 0;
-               }
+       return bo->funcs->cpu_prep(bo, pipe, op);
+}
 
-               bo->gpuaddr = req.gpuaddr[0];
-       }
-       return bo->gpuaddr + offset;
+drm_public void fd_bo_cpu_fini(struct fd_bo *bo)
+{
+       bo->funcs->cpu_fini(bo);
 }