Module: Mesa Branch: main Commit: fd2b014a12a225573cb864d9e5d20f7893beed32 URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=fd2b014a12a225573cb864d9e5d20f7893beed32
Author: Rob Clark <[email protected]> Date: Thu Feb 16 13:56:14 2023 -0800 turnip: Split out vfuncs for kernel interface To allow support for multiple KMD's to coexist, split out vfuncs. This doesn't quite *yet* let us have kgsl and msm coexist, because of the different vk entrypoints needed by the kgsl backend. Signed-off-by: Rob Clark <[email protected]> Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21394> --- src/freedreno/vulkan/meson.build | 1 + src/freedreno/vulkan/tu_device.h | 4 ++ src/freedreno/vulkan/tu_knl.c | 102 ++++++++++++++++++++++++++++++++++ src/freedreno/vulkan/tu_knl.h | 32 ++++++++--- src/freedreno/vulkan/tu_knl_drm_msm.c | 94 ++++++++++++++++++------------- src/freedreno/vulkan/tu_knl_kgsl.c | 90 ++++++++++++++++++------------ 6 files changed, 240 insertions(+), 83 deletions(-) diff --git a/src/freedreno/vulkan/meson.build b/src/freedreno/vulkan/meson.build index b323b74d954..904a14ce815 100644 --- a/src/freedreno/vulkan/meson.build +++ b/src/freedreno/vulkan/meson.build @@ -24,6 +24,7 @@ libtu_files = files( 'tu_dynamic_rendering.c', 'tu_formats.c', 'tu_image.c', + 'tu_knl.c', 'tu_lrz.c', 'tu_nir_lower_multiview.c', 'tu_pass.c', diff --git a/src/freedreno/vulkan/tu_device.h b/src/freedreno/vulkan/tu_device.h index 4de58d66b05..a07874c5dc2 100644 --- a/src/freedreno/vulkan/tu_device.h +++ b/src/freedreno/vulkan/tu_device.h @@ -114,10 +114,14 @@ struct tu_physical_device VK_DEFINE_HANDLE_CASTS(tu_physical_device, vk.base, VkPhysicalDevice, VK_OBJECT_TYPE_PHYSICAL_DEVICE) +struct tu_knl; + struct tu_instance { struct vk_instance vk; + const struct tu_knl *knl; + uint32_t api_version; struct driOptionCache dri_options; diff --git a/src/freedreno/vulkan/tu_knl.c b/src/freedreno/vulkan/tu_knl.c new file mode 100644 index 00000000000..5f57b7d33f2 --- /dev/null +++ b/src/freedreno/vulkan/tu_knl.c @@ -0,0 +1,102 @@ +/* + * Copyright © 2016 Red Hat. + * Copyright © 2016 Bas Nieuwenhuizen + * SPDX-License-Identifier: MIT + * + * based in part on anv driver which is: + * Copyright © 2015 Intel Corporation + */ + +#include "tu_device.h" +#include "tu_knl.h" + + +VkResult +tu_bo_init_new_explicit_iova(struct tu_device *dev, + struct tu_bo **out_bo, + uint64_t size, + uint64_t client_iova, + enum tu_bo_alloc_flags flags, const char *name) +{ + return dev->instance->knl->bo_init(dev, out_bo, size, client_iova, flags, name); +} + +VkResult +tu_bo_init_dmabuf(struct tu_device *dev, + struct tu_bo **bo, + uint64_t size, + int fd) +{ + return dev->instance->knl->bo_init_dmabuf(dev, bo, size, fd); +} + +int +tu_bo_export_dmabuf(struct tu_device *dev, struct tu_bo *bo) +{ + return dev->instance->knl->bo_export_dmabuf(dev, bo); +} + +void +tu_bo_finish(struct tu_device *dev, struct tu_bo *bo) +{ + dev->instance->knl->bo_finish(dev, bo); +} + +VkResult +tu_bo_map(struct tu_device *dev, struct tu_bo *bo) +{ + return dev->instance->knl->bo_map(dev, bo); +} + +void tu_bo_allow_dump(struct tu_device *dev, struct tu_bo *bo) +{ + dev->instance->knl->bo_allow_dump(dev, bo); +} + +int +tu_device_get_gpu_timestamp(struct tu_device *dev, + uint64_t *ts) +{ + return dev->instance->knl->device_get_gpu_timestamp(dev, ts); +} + +int +tu_device_get_suspend_count(struct tu_device *dev, + uint64_t *suspend_count) +{ + return dev->instance->knl->device_get_suspend_count(dev, suspend_count); +} + +VkResult +tu_device_wait_u_trace(struct tu_device *dev, struct tu_u_trace_syncobj *syncobj) +{ + return dev->instance->knl->device_wait_u_trace(dev, syncobj); +} + +VkResult +tu_device_check_status(struct vk_device *vk_device) +{ + struct tu_device *dev = container_of(vk_device, struct tu_device, vk); + return dev->instance->knl->device_check_status(dev); +} + +int +tu_drm_submitqueue_new(const struct tu_device *dev, + int priority, + uint32_t *queue_id) +{ + return dev->instance->knl->submitqueue_new(dev, priority, queue_id); +} + +void +tu_drm_submitqueue_close(const struct tu_device *dev, uint32_t queue_id) +{ + dev->instance->knl->submitqueue_close(dev, queue_id); +} + +VkResult +tu_queue_submit(struct vk_queue *vk_queue, struct vk_queue_submit *submit) +{ + struct tu_queue *queue = container_of(vk_queue, struct tu_queue, vk); + return queue->device->instance->knl->queue_submit(queue, submit); +} diff --git a/src/freedreno/vulkan/tu_knl.h b/src/freedreno/vulkan/tu_knl.h index e65e1425b5b..5b2a6ca2de5 100644 --- a/src/freedreno/vulkan/tu_knl.h +++ b/src/freedreno/vulkan/tu_knl.h @@ -12,13 +12,6 @@ #include "tu_common.h" -/* Keep tu_syncobj until porting to common code for kgsl too */ -#ifdef TU_USE_KGSL -struct tu_syncobj; -/* for TU_FROM_HANDLE with both VkFence and VkSemaphore: */ -#define tu_syncobj_from_handle(x) ((struct tu_syncobj*) (uintptr_t) (x)) -#endif - struct tu_u_trace_syncobj; enum tu_bo_alloc_flags @@ -45,8 +38,7 @@ enum tu_timeline_sync_state { TU_TIMELINE_SYNC_STATE_SIGNALED, }; -struct tu_bo -{ +struct tu_bo { uint32_t gem_handle; uint64_t size; uint64_t iova; @@ -61,6 +53,28 @@ struct tu_bo bool implicit_sync : 1; }; +struct tu_knl { + const char *name; + + int (*device_get_gpu_timestamp)(struct tu_device *dev, uint64_t *ts); + int (*device_get_suspend_count)(struct tu_device *dev, uint64_t *suspend_count); + VkResult (*device_check_status)(struct tu_device *dev); + int (*submitqueue_new)(const struct tu_device *dev, int priority, uint32_t *queue_id); + void (*submitqueue_close)(const struct tu_device *dev, uint32_t queue_id); + VkResult (*bo_init)(struct tu_device *dev, struct tu_bo **out_bo, uint64_t size, + uint64_t client_iova, enum tu_bo_alloc_flags flags, const char *name); + VkResult (*bo_init_dmabuf)(struct tu_device *dev, struct tu_bo **out_bo, + uint64_t size, int prime_fd); + int (*bo_export_dmabuf)(struct tu_device *dev, struct tu_bo *bo); + VkResult (*bo_map)(struct tu_device *dev, struct tu_bo *bo); + void (*bo_allow_dump)(struct tu_device *dev, struct tu_bo *bo); + void (*bo_finish)(struct tu_device *dev, struct tu_bo *bo); + VkResult (*device_wait_u_trace)(struct tu_device *dev, + struct tu_u_trace_syncobj *syncobj); + VkResult (*queue_submit)(struct tu_queue *queue, + struct vk_queue_submit *submit); +}; + struct tu_timeline_sync { struct vk_sync base; diff --git a/src/freedreno/vulkan/tu_knl_drm_msm.c b/src/freedreno/vulkan/tu_knl_drm_msm.c index b254e841cab..75116289be0 100644 --- a/src/freedreno/vulkan/tu_knl_drm_msm.c +++ b/src/freedreno/vulkan/tu_knl_drm_msm.c @@ -140,23 +140,22 @@ tu_drm_get_priorities(const struct tu_physical_device *dev) return val; } -int -tu_device_get_gpu_timestamp(struct tu_device *dev, uint64_t *ts) +static int +msm_device_get_gpu_timestamp(struct tu_device *dev, uint64_t *ts) { return tu_drm_get_param(dev->physical_device, MSM_PARAM_TIMESTAMP, ts); } -int -tu_device_get_suspend_count(struct tu_device *dev, uint64_t *suspend_count) +static int +msm_device_get_suspend_count(struct tu_device *dev, uint64_t *suspend_count) { int ret = tu_drm_get_param(dev->physical_device, MSM_PARAM_SUSPENDS, suspend_count); return ret; } -VkResult -tu_device_check_status(struct vk_device *vk_device) +static VkResult +msm_device_check_status(struct tu_device *device) { - struct tu_device *device = container_of(vk_device, struct tu_device, vk); struct tu_physical_device *physical_device = device->physical_device; uint64_t last_fault_count = physical_device->fault_count; @@ -170,10 +169,10 @@ tu_device_check_status(struct vk_device *vk_device) return VK_SUCCESS; } -int -tu_drm_submitqueue_new(const struct tu_device *dev, - int priority, - uint32_t *queue_id) +static int +msm_submitqueue_new(const struct tu_device *dev, + int priority, + uint32_t *queue_id) { assert(priority >= 0 && priority < dev->physical_device->submitqueue_priority_count); @@ -191,8 +190,8 @@ tu_drm_submitqueue_new(const struct tu_device *dev, return 0; } -void -tu_drm_submitqueue_close(const struct tu_device *dev, uint32_t queue_id) +static void +msm_submitqueue_close(const struct tu_device *dev, uint32_t queue_id) { drmCommandWrite(dev->fd, DRM_MSM_SUBMITQUEUE_CLOSE, &queue_id, sizeof(uint32_t)); @@ -392,13 +391,13 @@ tu_bo_set_kernel_name(struct tu_device *dev, struct tu_bo *bo, const char *name) } } -VkResult -tu_bo_init_new_explicit_iova(struct tu_device *dev, - struct tu_bo **out_bo, - uint64_t size, - uint64_t client_iova, - enum tu_bo_alloc_flags flags, - const char *name) +static VkResult +msm_bo_init(struct tu_device *dev, + struct tu_bo **out_bo, + uint64_t size, + uint64_t client_iova, + enum tu_bo_alloc_flags flags, + const char *name) { /* TODO: Choose better flags. As of 2018-11-12, freedreno/drm/msm_bo.c * always sets `flags = MSM_BO_WC`, and we copy that behavior here. @@ -433,11 +432,11 @@ tu_bo_init_new_explicit_iova(struct tu_device *dev, return result; } -VkResult -tu_bo_init_dmabuf(struct tu_device *dev, - struct tu_bo **out_bo, - uint64_t size, - int prime_fd) +static VkResult +msm_bo_init_dmabuf(struct tu_device *dev, + struct tu_bo **out_bo, + uint64_t size, + int prime_fd) { /* lseek() to get the real size */ off_t real_size = lseek(prime_fd, 0, SEEK_END); @@ -484,8 +483,8 @@ tu_bo_init_dmabuf(struct tu_device *dev, return result; } -int -tu_bo_export_dmabuf(struct tu_device *dev, struct tu_bo *bo) +static int +msm_bo_export_dmabuf(struct tu_device *dev, struct tu_bo *bo) { int prime_fd; int ret = drmPrimeHandleToFD(dev->fd, bo->gem_handle, @@ -494,8 +493,8 @@ tu_bo_export_dmabuf(struct tu_device *dev, struct tu_bo *bo) return ret == 0 ? prime_fd : -1; } -VkResult -tu_bo_map(struct tu_device *dev, struct tu_bo *bo) +static VkResult +msm_bo_map(struct tu_device *dev, struct tu_bo *bo) { if (bo->map) return VK_SUCCESS; @@ -514,16 +513,16 @@ tu_bo_map(struct tu_device *dev, struct tu_bo *bo) return VK_SUCCESS; } -void -tu_bo_allow_dump(struct tu_device *dev, struct tu_bo *bo) +static void +msm_bo_allow_dump(struct tu_device *dev, struct tu_bo *bo) { mtx_lock(&dev->bo_mutex); dev->bo_list[bo->bo_list_idx].flags |= MSM_SUBMIT_BO_DUMP; mtx_unlock(&dev->bo_mutex); } -void -tu_bo_finish(struct tu_device *dev, struct tu_bo *bo) +static void +msm_bo_finish(struct tu_device *dev, struct tu_bo *bo) { assert(bo->gem_handle); @@ -1074,8 +1073,8 @@ get_abs_timeout(struct drm_msm_timespec *tv, uint64_t ns) tv->tv_nsec = t.tv_nsec + ns % 1000000000; } -VkResult -tu_device_wait_u_trace(struct tu_device *dev, struct tu_u_trace_syncobj *syncobj) +static VkResult +msm_device_wait_u_trace(struct tu_device *dev, struct tu_u_trace_syncobj *syncobj) { struct drm_msm_wait_fence req = { .fence = syncobj->fence, @@ -1094,11 +1093,10 @@ tu_device_wait_u_trace(struct tu_device *dev, struct tu_u_trace_syncobj *syncobj return VK_SUCCESS; } -VkResult -tu_queue_submit(struct vk_queue *vk_queue, struct vk_queue_submit *submit) +static VkResult +msm_queue_submit(struct tu_queue *queue, struct vk_queue_submit *submit) { MESA_TRACE_FUNC(); - struct tu_queue *queue = container_of(vk_queue, struct tu_queue, vk); uint32_t perf_pass_index = queue->device->perfcntrs_pass_cs ? submit->perf_pass_index : ~0; struct tu_queue_submit submit_req; @@ -1155,6 +1153,24 @@ tu_queue_submit(struct vk_queue *vk_queue, struct vk_queue_submit *submit) return VK_SUCCESS; } +static const struct tu_knl msm_knl_funcs = { + .name = "msm", + + .device_get_gpu_timestamp = msm_device_get_gpu_timestamp, + .device_get_suspend_count = msm_device_get_suspend_count, + .device_check_status = msm_device_check_status, + .submitqueue_new = msm_submitqueue_new, + .submitqueue_close = msm_submitqueue_close, + .bo_init = msm_bo_init, + .bo_init_dmabuf = msm_bo_init_dmabuf, + .bo_export_dmabuf = msm_bo_export_dmabuf, + .bo_map = msm_bo_map, + .bo_allow_dump = msm_bo_allow_dump, + .bo_finish = msm_bo_finish, + .device_wait_u_trace = msm_device_wait_u_trace, + .queue_submit = msm_queue_submit, +}; + const struct vk_sync_type tu_timeline_sync_type = { .size = sizeof(struct tu_timeline_sync), .features = VK_SYNC_FEATURE_BINARY | @@ -1339,6 +1355,8 @@ tu_physical_device_try_create(struct vk_instance *vk_instance, device->heap.used = 0u; device->heap.flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT; + instance->knl = &msm_knl_funcs; + result = tu_physical_device_init(device, instance); if (result == VK_SUCCESS) { diff --git a/src/freedreno/vulkan/tu_knl_kgsl.c b/src/freedreno/vulkan/tu_knl_kgsl.c index b056c02eaf6..38b9f670298 100644 --- a/src/freedreno/vulkan/tu_knl_kgsl.c +++ b/src/freedreno/vulkan/tu_knl_kgsl.c @@ -7,6 +7,7 @@ #include <errno.h> #include <fcntl.h> +#include <poll.h> #include <sys/ioctl.h> #include <sys/mman.h> @@ -25,6 +26,7 @@ struct tu_syncobj { uint32_t timestamp; bool timestamp_valid; }; +#define tu_syncobj_from_handle(x) ((struct tu_syncobj*) (uintptr_t) (x)) static int safe_ioctl(int fd, unsigned long request, void *arg) @@ -38,10 +40,10 @@ safe_ioctl(int fd, unsigned long request, void *arg) return ret; } -int -tu_drm_submitqueue_new(const struct tu_device *dev, - int priority, - uint32_t *queue_id) +static int +kgsl_submitqueue_new(const struct tu_device *dev, + int priority, + uint32_t *queue_id) { struct kgsl_drawctxt_create req = { .flags = KGSL_CONTEXT_SAVE_GMEM | @@ -58,8 +60,8 @@ tu_drm_submitqueue_new(const struct tu_device *dev, return 0; } -void -tu_drm_submitqueue_close(const struct tu_device *dev, uint32_t queue_id) +static void +kgsl_submitqueue_close(const struct tu_device *dev, uint32_t queue_id) { struct kgsl_drawctxt_destroy req = { .drawctxt_id = queue_id, @@ -68,13 +70,13 @@ tu_drm_submitqueue_close(const struct tu_device *dev, uint32_t queue_id) safe_ioctl(dev->physical_device->local_fd, IOCTL_KGSL_DRAWCTXT_DESTROY, &req); } -VkResult -tu_bo_init_new_explicit_iova(struct tu_device *dev, - struct tu_bo **out_bo, - uint64_t size, - uint64_t client_iova, - enum tu_bo_alloc_flags flags, - const char *name) +static VkResult +kgsl_bo_init(struct tu_device *dev, + struct tu_bo **out_bo, + uint64_t size, + uint64_t client_iova, + enum tu_bo_alloc_flags flags, + const char *name) { assert(client_iova == 0); @@ -110,11 +112,11 @@ tu_bo_init_new_explicit_iova(struct tu_device *dev, return VK_SUCCESS; } -VkResult -tu_bo_init_dmabuf(struct tu_device *dev, - struct tu_bo **out_bo, - uint64_t size, - int fd) +static VkResult +kgsl_bo_init_dmabuf(struct tu_device *dev, + struct tu_bo **out_bo, + uint64_t size, + int fd) { struct kgsl_gpuobj_import_dma_buf import_dmabuf = { .fd = fd, @@ -159,16 +161,16 @@ tu_bo_init_dmabuf(struct tu_device *dev, return VK_SUCCESS; } -int -tu_bo_export_dmabuf(struct tu_device *dev, struct tu_bo *bo) +static int +kgsl_bo_export_dmabuf(struct tu_device *dev, struct tu_bo *bo) { tu_stub(); return -1; } -VkResult -tu_bo_map(struct tu_device *dev, struct tu_bo *bo) +static VkResult +kgsl_bo_map(struct tu_device *dev, struct tu_bo *bo) { if (bo->map) return VK_SUCCESS; @@ -184,13 +186,13 @@ tu_bo_map(struct tu_device *dev, struct tu_bo *bo) return VK_SUCCESS; } -void -tu_bo_allow_dump(struct tu_device *dev, struct tu_bo *bo) +static void +kgsl_bo_allow_dump(struct tu_device *dev, struct tu_bo *bo) { } -void -tu_bo_finish(struct tu_device *dev, struct tu_bo *bo) +static void +kgsl_bo_finish(struct tu_device *dev, struct tu_bo *bo) { assert(bo->gem_handle); @@ -683,33 +685,31 @@ tu_GetFenceStatus(VkDevice _device, VkFence _fence) return VK_SUCCESS; } -VkResult -tu_device_wait_u_trace(struct tu_device *dev, struct tu_u_trace_syncobj *syncobj) +static VkResult +kgsl_device_wait_u_trace(struct tu_device *dev, struct tu_u_trace_syncobj *syncobj) { tu_finishme("tu_device_wait_u_trace"); return VK_SUCCESS; } -int -tu_device_get_gpu_timestamp(struct tu_device *dev, uint64_t *ts) +static int +kgsl_device_get_gpu_timestamp(struct tu_device *dev, uint64_t *ts) { tu_finishme("tu_device_get_gpu_timestamp"); return 0; } -int -tu_device_get_suspend_count(struct tu_device *dev, uint64_t *suspend_count) +static int +kgsl_device_get_suspend_count(struct tu_device *dev, uint64_t *suspend_count) { /* kgsl doesn't have a way to get it */ *suspend_count = 0; return 0; } -VkResult -tu_device_check_status(struct vk_device *vk_device) +static VkResult +kgsl_device_check_status(struct tu_device *device) { - struct tu_device *device = container_of(vk_device, struct tu_device, vk); - for (unsigned i = 0; i < TU_MAX_QUEUE_FAMILIES; i++) { for (unsigned q = 0; q < device->queue_count[i]; q++) { /* KGSL's KGSL_PROP_GPU_RESET_STAT takes the u32 msm_queue_id and returns a @@ -757,6 +757,22 @@ tu_QueueSignalReleaseImageANDROID(VkQueue _queue, } #endif +static const struct tu_knl kgsl_knl_funcs = { + .name = "kgsl", + + .device_get_gpu_timestamp = kgsl_device_get_gpu_timestamp, + .device_get_suspend_count = kgsl_device_get_suspend_count, + .device_check_status = kgsl_device_check_status, + .submitqueue_new = kgsl_submitqueue_new, + .submitqueue_close = kgsl_submitqueue_close, + .bo_init = kgsl_bo_init, + .bo_init_dmabuf = kgsl_bo_init_dmabuf, + .bo_export_dmabuf = kgsl_bo_export_dmabuf, + .bo_map = kgsl_bo_map, + .bo_allow_dump = kgsl_bo_allow_dump, + .bo_finish = kgsl_bo_finish, + .device_wait_u_trace = kgsl_device_wait_u_trace, +}; VkResult tu_enumerate_devices(struct vk_instance *vk_instance) @@ -820,6 +836,8 @@ tu_enumerate_devices(struct vk_instance *vk_instance) device->heap.used = 0u; device->heap.flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT; + instance->knl = &kgsl_knl_funcs; + if (tu_physical_device_init(device, instance) != VK_SUCCESS) goto fail;
