1 // SPDX-License-Identifier: GPL-2.0+
2 /* Copyright (C) 2015-2018 Broadcom */
4 #include <linux/reservation.h>
5 #include <linux/mm_types.h>
7 #include <drm/drm_encoder.h>
8 #include <drm/drm_gem.h>
9 #include <drm/gpu_scheduler.h>
10 #include "uapi/drm/v3d_drm.h"
12 #define GMP_GRANULARITY (128 * 1024)
14 /* Enum for each of the V3D queues. */
21 #define V3D_MAX_QUEUES (V3D_TFU + 1)
23 struct v3d_queue_state {
24 struct drm_gpu_scheduler sched;
31 struct drm_device drm;
33 /* Short representation (e.g. 33, 41) of the V3D tech version
39 struct platform_device *pdev;
40 void __iomem *hub_regs;
41 void __iomem *core_regs[3];
42 void __iomem *bridge_regs;
43 void __iomem *gca_regs;
46 /* Virtual and DMA addresses of the single shared page table. */
50 /* Virtual and DMA addresses of the MMU's scratch page. When
51 * a read or write is invalid in the MMU, it will be
55 dma_addr_t mmu_scratch_paddr;
57 /* Number of V3D cores. */
60 /* Allocator managing the address space. All units are in
66 struct work_struct overflow_mem_work;
68 struct v3d_exec_info *bin_job;
69 struct v3d_exec_info *render_job;
70 struct v3d_tfu_job *tfu_job;
72 struct v3d_queue_state queue[V3D_MAX_QUEUES];
74 /* Spinlock used to synchronize the overflow memory
75 * management against bin job submission.
79 /* Protects bo_stats */
82 /* Lock taken when resetting the GPU, to keep multiple
83 * processes from trying to park the scheduler threads and
86 struct mutex reset_lock;
88 /* Lock taken when creating and pushing the GPU scheduler
89 * jobs, to keep the sched-fence seqnos in order.
91 struct mutex sched_lock;
99 static inline struct v3d_dev *
100 to_v3d_dev(struct drm_device *dev)
102 return (struct v3d_dev *)dev->dev_private;
105 /* The per-fd struct, which tracks the MMU mappings. */
106 struct v3d_file_priv {
109 struct drm_sched_entity sched_entity[V3D_MAX_QUEUES];
112 /* Tracks a mapping of a BO into a per-fd address space */
114 struct v3d_page_table *pt;
115 struct list_head list; /* entry in v3d_bo.vmas */
119 struct drm_gem_object base;
123 struct drm_mm_node node;
127 struct sg_table *sgt;
130 struct list_head vmas; /* list of v3d_vma */
132 /* List entry for the BO's position in
133 * v3d_exec_info->unref_list
135 struct list_head unref_head;
137 /* normally (resv == &_resv) except for imported bo's */
138 struct reservation_object *resv;
139 struct reservation_object _resv;
142 static inline struct v3d_bo *
143 to_v3d_bo(struct drm_gem_object *bo)
145 return (struct v3d_bo *)bo;
149 struct dma_fence base;
150 struct drm_device *dev;
151 /* v3d seqno for signaled() test */
153 enum v3d_queue queue;
156 static inline struct v3d_fence *
157 to_v3d_fence(struct dma_fence *fence)
159 return (struct v3d_fence *)fence;
162 #define V3D_READ(offset) readl(v3d->hub_regs + offset)
163 #define V3D_WRITE(offset, val) writel(val, v3d->hub_regs + offset)
165 #define V3D_BRIDGE_READ(offset) readl(v3d->bridge_regs + offset)
166 #define V3D_BRIDGE_WRITE(offset, val) writel(val, v3d->bridge_regs + offset)
168 #define V3D_GCA_READ(offset) readl(v3d->gca_regs + offset)
169 #define V3D_GCA_WRITE(offset, val) writel(val, v3d->gca_regs + offset)
171 #define V3D_CORE_READ(core, offset) readl(v3d->core_regs[core] + offset)
172 #define V3D_CORE_WRITE(core, offset, val) writel(val, v3d->core_regs[core] + offset)
175 struct drm_sched_job base;
177 struct v3d_exec_info *exec;
179 /* An optional fence userspace can pass in for the job to depend on. */
180 struct dma_fence *in_fence;
182 /* v3d fence to be signaled by IRQ handler when the job is complete. */
183 struct dma_fence *done_fence;
185 /* GPU virtual addresses of the start/end of the CL job. */
188 u32 timedout_ctca, timedout_ctra;
191 struct v3d_exec_info {
194 struct v3d_job bin, render;
196 /* Fence for when the scheduler considers the binner to be
197 * done, for render to depend on.
199 struct dma_fence *bin_done_fence;
201 /* Fence for when the scheduler considers the render to be
202 * done, for when the BOs reservations should be complete.
204 struct dma_fence *render_done_fence;
206 struct kref refcount;
208 /* This is the array of BOs that were looked up at the start of exec. */
212 /* List of overflow BOs used in the job that need to be
213 * released once the job is complete.
215 struct list_head unref_list;
217 /* Submitted tile memory allocation start/size, tile state. */
222 struct drm_sched_job base;
224 struct drm_v3d_submit_tfu args;
226 /* An optional fence userspace can pass in for the job to depend on. */
227 struct dma_fence *in_fence;
229 /* v3d fence to be signaled by IRQ handler when the job is complete. */
230 struct dma_fence *done_fence;
234 struct kref refcount;
236 /* This is the array of BOs that were looked up at the start of exec. */
237 struct v3d_bo *bo[4];
241 * _wait_for - magic (register) wait macro
243 * Does the right thing for modeset paths when run under kdgb or similar atomic
244 * contexts. Note that it's important that we check the condition again after
245 * having timed out, since the timeout could be due to preemption or similar and
246 * we've never had a chance to check the condition before the timeout.
248 #define wait_for(COND, MS) ({ \
249 unsigned long timeout__ = jiffies + msecs_to_jiffies(MS) + 1; \
252 if (time_after(jiffies, timeout__)) { \
254 ret__ = -ETIMEDOUT; \
262 static inline unsigned long nsecs_to_jiffies_timeout(const u64 n)
264 /* nsecs_to_jiffies64() does not guard against overflow */
265 if (NSEC_PER_SEC % HZ &&
266 div_u64(n, NSEC_PER_SEC) >= MAX_JIFFY_OFFSET / HZ)
267 return MAX_JIFFY_OFFSET;
269 return min_t(u64, MAX_JIFFY_OFFSET, nsecs_to_jiffies64(n) + 1);
273 void v3d_free_object(struct drm_gem_object *gem_obj);
274 struct v3d_bo *v3d_bo_create(struct drm_device *dev, struct drm_file *file_priv,
276 int v3d_create_bo_ioctl(struct drm_device *dev, void *data,
277 struct drm_file *file_priv);
278 int v3d_mmap_bo_ioctl(struct drm_device *dev, void *data,
279 struct drm_file *file_priv);
280 int v3d_get_bo_offset_ioctl(struct drm_device *dev, void *data,
281 struct drm_file *file_priv);
282 vm_fault_t v3d_gem_fault(struct vm_fault *vmf);
283 int v3d_mmap(struct file *filp, struct vm_area_struct *vma);
284 struct reservation_object *v3d_prime_res_obj(struct drm_gem_object *obj);
285 int v3d_prime_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma);
286 struct sg_table *v3d_prime_get_sg_table(struct drm_gem_object *obj);
287 struct drm_gem_object *v3d_prime_import_sg_table(struct drm_device *dev,
288 struct dma_buf_attachment *attach,
289 struct sg_table *sgt);
292 int v3d_debugfs_init(struct drm_minor *minor);
295 extern const struct dma_fence_ops v3d_fence_ops;
296 struct dma_fence *v3d_fence_create(struct v3d_dev *v3d, enum v3d_queue queue);
299 int v3d_gem_init(struct drm_device *dev);
300 void v3d_gem_destroy(struct drm_device *dev);
301 int v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
302 struct drm_file *file_priv);
303 int v3d_submit_tfu_ioctl(struct drm_device *dev, void *data,
304 struct drm_file *file_priv);
305 int v3d_wait_bo_ioctl(struct drm_device *dev, void *data,
306 struct drm_file *file_priv);
307 void v3d_exec_put(struct v3d_exec_info *exec);
308 void v3d_tfu_job_put(struct v3d_tfu_job *exec);
309 void v3d_reset(struct v3d_dev *v3d);
310 void v3d_invalidate_caches(struct v3d_dev *v3d);
311 void v3d_flush_caches(struct v3d_dev *v3d);
314 void v3d_irq_init(struct v3d_dev *v3d);
315 void v3d_irq_enable(struct v3d_dev *v3d);
316 void v3d_irq_disable(struct v3d_dev *v3d);
317 void v3d_irq_reset(struct v3d_dev *v3d);
320 int v3d_mmu_get_offset(struct drm_file *file_priv, struct v3d_bo *bo,
322 int v3d_mmu_set_page_table(struct v3d_dev *v3d);
323 void v3d_mmu_insert_ptes(struct v3d_bo *bo);
324 void v3d_mmu_remove_ptes(struct v3d_bo *bo);
327 int v3d_sched_init(struct v3d_dev *v3d);
328 void v3d_sched_fini(struct v3d_dev *v3d);