xref: /linux/drivers/gpu/drm/v3d/v3d_drv.h (revision ab779466166348eecf17d20f620aa9a47965c934)
1 // SPDX-License-Identifier: GPL-2.0+
2 /* Copyright (C) 2015-2018 Broadcom */
3 
4 #include <linux/delay.h>
5 #include <linux/mutex.h>
6 #include <linux/spinlock_types.h>
7 #include <linux/workqueue.h>
8 
9 #include <drm/drm_encoder.h>
10 #include <drm/drm_gem.h>
11 #include <drm/drm_gem_shmem_helper.h>
12 #include <drm/gpu_scheduler.h>
13 
14 #include "uapi/drm/v3d_drm.h"
15 
16 struct clk;
17 struct platform_device;
18 struct reset_control;
19 
20 #define GMP_GRANULARITY (128 * 1024)
21 
22 #define V3D_MAX_QUEUES (V3D_CACHE_CLEAN + 1)
23 
24 static inline char *v3d_queue_to_string(enum v3d_queue queue)
25 {
26 	switch (queue) {
27 	case V3D_BIN: return "bin";
28 	case V3D_RENDER: return "render";
29 	case V3D_TFU: return "tfu";
30 	case V3D_CSD: return "csd";
31 	case V3D_CACHE_CLEAN: return "cache_clean";
32 	}
33 	return "UNKNOWN";
34 }
35 
36 struct v3d_queue_state {
37 	struct drm_gpu_scheduler sched;
38 
39 	u64 fence_context;
40 	u64 emit_seqno;
41 
42 	u64 start_ns;
43 	u64 enabled_ns;
44 	u64 jobs_sent;
45 };
46 
47 /* Performance monitor object. The perform lifetime is controlled by userspace
48  * using perfmon related ioctls. A perfmon can be attached to a submit_cl
49  * request, and when this is the case, HW perf counters will be activated just
50  * before the submit_cl is submitted to the GPU and disabled when the job is
51  * done. This way, only events related to a specific job will be counted.
52  */
53 struct v3d_perfmon {
54 	/* Tracks the number of users of the perfmon, when this counter reaches
55 	 * zero the perfmon is destroyed.
56 	 */
57 	refcount_t refcnt;
58 
59 	/* Protects perfmon stop, as it can be invoked from multiple places. */
60 	struct mutex lock;
61 
62 	/* Number of counters activated in this perfmon instance
63 	 * (should be less than DRM_V3D_MAX_PERF_COUNTERS).
64 	 */
65 	u8 ncounters;
66 
67 	/* Events counted by the HW perf counters. */
68 	u8 counters[DRM_V3D_MAX_PERF_COUNTERS];
69 
70 	/* Storage for counter values. Counters are incremented by the
71 	 * HW perf counter values every time the perfmon is attached
72 	 * to a GPU job.  This way, perfmon users don't have to
73 	 * retrieve the results after each job if they want to track
74 	 * events covering several submissions.  Note that counter
75 	 * values can't be reset, but you can fake a reset by
76 	 * destroying the perfmon and creating a new one.
77 	 */
78 	u64 values[] __counted_by(ncounters);
79 };
80 
81 struct v3d_dev {
82 	struct drm_device drm;
83 
84 	/* Short representation (e.g. 33, 41) of the V3D tech version
85 	 * and revision.
86 	 */
87 	int ver;
88 	bool single_irq_line;
89 
90 	void __iomem *hub_regs;
91 	void __iomem *core_regs[3];
92 	void __iomem *bridge_regs;
93 	void __iomem *gca_regs;
94 	struct clk *clk;
95 	struct reset_control *reset;
96 
97 	/* Virtual and DMA addresses of the single shared page table. */
98 	volatile u32 *pt;
99 	dma_addr_t pt_paddr;
100 
101 	/* Virtual and DMA addresses of the MMU's scratch page.  When
102 	 * a read or write is invalid in the MMU, it will be
103 	 * redirected here.
104 	 */
105 	void *mmu_scratch;
106 	dma_addr_t mmu_scratch_paddr;
107 	/* virtual address bits from V3D to the MMU. */
108 	int va_width;
109 
110 	/* Number of V3D cores. */
111 	u32 cores;
112 
113 	/* Allocator managing the address space.  All units are in
114 	 * number of pages.
115 	 */
116 	struct drm_mm mm;
117 	spinlock_t mm_lock;
118 
119 	struct work_struct overflow_mem_work;
120 
121 	struct v3d_bin_job *bin_job;
122 	struct v3d_render_job *render_job;
123 	struct v3d_tfu_job *tfu_job;
124 	struct v3d_csd_job *csd_job;
125 
126 	struct v3d_queue_state queue[V3D_MAX_QUEUES];
127 
128 	/* Spinlock used to synchronize the overflow memory
129 	 * management against bin job submission.
130 	 */
131 	spinlock_t job_lock;
132 
133 	/* Used to track the active perfmon if any. */
134 	struct v3d_perfmon *active_perfmon;
135 
136 	/* Protects bo_stats */
137 	struct mutex bo_lock;
138 
139 	/* Lock taken when resetting the GPU, to keep multiple
140 	 * processes from trying to park the scheduler threads and
141 	 * reset at once.
142 	 */
143 	struct mutex reset_lock;
144 
145 	/* Lock taken when creating and pushing the GPU scheduler
146 	 * jobs, to keep the sched-fence seqnos in order.
147 	 */
148 	struct mutex sched_lock;
149 
150 	/* Lock taken during a cache clean and when initiating an L2
151 	 * flush, to keep L2 flushes from interfering with the
152 	 * synchronous L2 cleans.
153 	 */
154 	struct mutex cache_clean_lock;
155 
156 	struct {
157 		u32 num_allocated;
158 		u32 pages_allocated;
159 	} bo_stats;
160 };
161 
162 static inline struct v3d_dev *
163 to_v3d_dev(struct drm_device *dev)
164 {
165 	return container_of(dev, struct v3d_dev, drm);
166 }
167 
168 static inline bool
169 v3d_has_csd(struct v3d_dev *v3d)
170 {
171 	return v3d->ver >= 41;
172 }
173 
174 #define v3d_to_pdev(v3d) to_platform_device((v3d)->drm.dev)
175 
176 /* The per-fd struct, which tracks the MMU mappings. */
177 struct v3d_file_priv {
178 	struct v3d_dev *v3d;
179 
180 	struct {
181 		struct idr idr;
182 		struct mutex lock;
183 	} perfmon;
184 
185 	struct drm_sched_entity sched_entity[V3D_MAX_QUEUES];
186 
187 	u64 start_ns[V3D_MAX_QUEUES];
188 
189 	u64 enabled_ns[V3D_MAX_QUEUES];
190 
191 	u64 jobs_sent[V3D_MAX_QUEUES];
192 };
193 
194 struct v3d_bo {
195 	struct drm_gem_shmem_object base;
196 
197 	struct drm_mm_node node;
198 
199 	/* List entry for the BO's position in
200 	 * v3d_render_job->unref_list
201 	 */
202 	struct list_head unref_head;
203 };
204 
205 static inline struct v3d_bo *
206 to_v3d_bo(struct drm_gem_object *bo)
207 {
208 	return (struct v3d_bo *)bo;
209 }
210 
211 struct v3d_fence {
212 	struct dma_fence base;
213 	struct drm_device *dev;
214 	/* v3d seqno for signaled() test */
215 	u64 seqno;
216 	enum v3d_queue queue;
217 };
218 
219 static inline struct v3d_fence *
220 to_v3d_fence(struct dma_fence *fence)
221 {
222 	return (struct v3d_fence *)fence;
223 }
224 
225 #define V3D_READ(offset) readl(v3d->hub_regs + offset)
226 #define V3D_WRITE(offset, val) writel(val, v3d->hub_regs + offset)
227 
228 #define V3D_BRIDGE_READ(offset) readl(v3d->bridge_regs + offset)
229 #define V3D_BRIDGE_WRITE(offset, val) writel(val, v3d->bridge_regs + offset)
230 
231 #define V3D_GCA_READ(offset) readl(v3d->gca_regs + offset)
232 #define V3D_GCA_WRITE(offset, val) writel(val, v3d->gca_regs + offset)
233 
234 #define V3D_CORE_READ(core, offset) readl(v3d->core_regs[core] + offset)
235 #define V3D_CORE_WRITE(core, offset, val) writel(val, v3d->core_regs[core] + offset)
236 
237 struct v3d_job {
238 	struct drm_sched_job base;
239 
240 	struct kref refcount;
241 
242 	struct v3d_dev *v3d;
243 
244 	/* This is the array of BOs that were looked up at the start
245 	 * of submission.
246 	 */
247 	struct drm_gem_object **bo;
248 	u32 bo_count;
249 
250 	/* v3d fence to be signaled by IRQ handler when the job is complete. */
251 	struct dma_fence *irq_fence;
252 
253 	/* scheduler fence for when the job is considered complete and
254 	 * the BO reservations can be released.
255 	 */
256 	struct dma_fence *done_fence;
257 
258 	/* Pointer to a performance monitor object if the user requested it,
259 	 * NULL otherwise.
260 	 */
261 	struct v3d_perfmon *perfmon;
262 
263 	/* File descriptor of the process that submitted the job that could be used
264 	 * for collecting stats by process of GPU usage.
265 	 */
266 	struct drm_file *file;
267 
268 	/* Callback for the freeing of the job on refcount going to 0. */
269 	void (*free)(struct kref *ref);
270 };
271 
272 struct v3d_bin_job {
273 	struct v3d_job base;
274 
275 	/* GPU virtual addresses of the start/end of the CL job. */
276 	u32 start, end;
277 
278 	u32 timedout_ctca, timedout_ctra;
279 
280 	/* Corresponding render job, for attaching our overflow memory. */
281 	struct v3d_render_job *render;
282 
283 	/* Submitted tile memory allocation start/size, tile state. */
284 	u32 qma, qms, qts;
285 };
286 
287 struct v3d_render_job {
288 	struct v3d_job base;
289 
290 	/* GPU virtual addresses of the start/end of the CL job. */
291 	u32 start, end;
292 
293 	u32 timedout_ctca, timedout_ctra;
294 
295 	/* List of overflow BOs used in the job that need to be
296 	 * released once the job is complete.
297 	 */
298 	struct list_head unref_list;
299 };
300 
301 struct v3d_tfu_job {
302 	struct v3d_job base;
303 
304 	struct drm_v3d_submit_tfu args;
305 };
306 
307 struct v3d_csd_job {
308 	struct v3d_job base;
309 
310 	u32 timedout_batches;
311 
312 	struct drm_v3d_submit_csd args;
313 };
314 
315 struct v3d_submit_outsync {
316 	struct drm_syncobj *syncobj;
317 };
318 
319 struct v3d_submit_ext {
320 	u32 flags;
321 	u32 wait_stage;
322 
323 	u32 in_sync_count;
324 	u64 in_syncs;
325 
326 	u32 out_sync_count;
327 	struct v3d_submit_outsync *out_syncs;
328 };
329 
330 /**
331  * __wait_for - magic wait macro
332  *
333  * Macro to help avoid open coding check/wait/timeout patterns. Note that it's
334  * important that we check the condition again after having timed out, since the
335  * timeout could be due to preemption or similar and we've never had a chance to
336  * check the condition before the timeout.
337  */
338 #define __wait_for(OP, COND, US, Wmin, Wmax) ({ \
339 	const ktime_t end__ = ktime_add_ns(ktime_get_raw(), 1000ll * (US)); \
340 	long wait__ = (Wmin); /* recommended min for usleep is 10 us */	\
341 	int ret__;							\
342 	might_sleep();							\
343 	for (;;) {							\
344 		const bool expired__ = ktime_after(ktime_get_raw(), end__); \
345 		OP;							\
346 		/* Guarantee COND check prior to timeout */		\
347 		barrier();						\
348 		if (COND) {						\
349 			ret__ = 0;					\
350 			break;						\
351 		}							\
352 		if (expired__) {					\
353 			ret__ = -ETIMEDOUT;				\
354 			break;						\
355 		}							\
356 		usleep_range(wait__, wait__ * 2);			\
357 		if (wait__ < (Wmax))					\
358 			wait__ <<= 1;					\
359 	}								\
360 	ret__;								\
361 })
362 
363 #define _wait_for(COND, US, Wmin, Wmax)	__wait_for(, (COND), (US), (Wmin), \
364 						   (Wmax))
365 #define wait_for(COND, MS)		_wait_for((COND), (MS) * 1000, 10, 1000)
366 
367 static inline unsigned long nsecs_to_jiffies_timeout(const u64 n)
368 {
369 	/* nsecs_to_jiffies64() does not guard against overflow */
370 	if ((NSEC_PER_SEC % HZ) != 0 &&
371 	    div_u64(n, NSEC_PER_SEC) >= MAX_JIFFY_OFFSET / HZ)
372 		return MAX_JIFFY_OFFSET;
373 
374 	return min_t(u64, MAX_JIFFY_OFFSET, nsecs_to_jiffies64(n) + 1);
375 }
376 
377 /* v3d_bo.c */
378 struct drm_gem_object *v3d_create_object(struct drm_device *dev, size_t size);
379 void v3d_free_object(struct drm_gem_object *gem_obj);
380 struct v3d_bo *v3d_bo_create(struct drm_device *dev, struct drm_file *file_priv,
381 			     size_t size);
382 int v3d_create_bo_ioctl(struct drm_device *dev, void *data,
383 			struct drm_file *file_priv);
384 int v3d_mmap_bo_ioctl(struct drm_device *dev, void *data,
385 		      struct drm_file *file_priv);
386 int v3d_get_bo_offset_ioctl(struct drm_device *dev, void *data,
387 			    struct drm_file *file_priv);
388 struct drm_gem_object *v3d_prime_import_sg_table(struct drm_device *dev,
389 						 struct dma_buf_attachment *attach,
390 						 struct sg_table *sgt);
391 
392 /* v3d_debugfs.c */
393 void v3d_debugfs_init(struct drm_minor *minor);
394 
395 /* v3d_fence.c */
396 extern const struct dma_fence_ops v3d_fence_ops;
397 struct dma_fence *v3d_fence_create(struct v3d_dev *v3d, enum v3d_queue queue);
398 
399 /* v3d_gem.c */
400 int v3d_gem_init(struct drm_device *dev);
401 void v3d_gem_destroy(struct drm_device *dev);
402 int v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
403 			struct drm_file *file_priv);
404 int v3d_submit_tfu_ioctl(struct drm_device *dev, void *data,
405 			 struct drm_file *file_priv);
406 int v3d_submit_csd_ioctl(struct drm_device *dev, void *data,
407 			 struct drm_file *file_priv);
408 int v3d_wait_bo_ioctl(struct drm_device *dev, void *data,
409 		      struct drm_file *file_priv);
410 void v3d_job_cleanup(struct v3d_job *job);
411 void v3d_job_put(struct v3d_job *job);
412 void v3d_reset(struct v3d_dev *v3d);
413 void v3d_invalidate_caches(struct v3d_dev *v3d);
414 void v3d_clean_caches(struct v3d_dev *v3d);
415 
416 /* v3d_irq.c */
417 int v3d_irq_init(struct v3d_dev *v3d);
418 void v3d_irq_enable(struct v3d_dev *v3d);
419 void v3d_irq_disable(struct v3d_dev *v3d);
420 void v3d_irq_reset(struct v3d_dev *v3d);
421 
422 /* v3d_mmu.c */
423 int v3d_mmu_get_offset(struct drm_file *file_priv, struct v3d_bo *bo,
424 		       u32 *offset);
425 int v3d_mmu_set_page_table(struct v3d_dev *v3d);
426 void v3d_mmu_insert_ptes(struct v3d_bo *bo);
427 void v3d_mmu_remove_ptes(struct v3d_bo *bo);
428 
429 /* v3d_sched.c */
430 int v3d_sched_init(struct v3d_dev *v3d);
431 void v3d_sched_fini(struct v3d_dev *v3d);
432 
433 /* v3d_perfmon.c */
434 void v3d_perfmon_get(struct v3d_perfmon *perfmon);
435 void v3d_perfmon_put(struct v3d_perfmon *perfmon);
436 void v3d_perfmon_start(struct v3d_dev *v3d, struct v3d_perfmon *perfmon);
437 void v3d_perfmon_stop(struct v3d_dev *v3d, struct v3d_perfmon *perfmon,
438 		      bool capture);
439 struct v3d_perfmon *v3d_perfmon_find(struct v3d_file_priv *v3d_priv, int id);
440 void v3d_perfmon_open_file(struct v3d_file_priv *v3d_priv);
441 void v3d_perfmon_close_file(struct v3d_file_priv *v3d_priv);
442 int v3d_perfmon_create_ioctl(struct drm_device *dev, void *data,
443 			     struct drm_file *file_priv);
444 int v3d_perfmon_destroy_ioctl(struct drm_device *dev, void *data,
445 			      struct drm_file *file_priv);
446 int v3d_perfmon_get_values_ioctl(struct drm_device *dev, void *data,
447 				 struct drm_file *file_priv);
448 
449 /* v3d_sysfs.c */
450 int v3d_sysfs_init(struct device *dev);
451 void v3d_sysfs_destroy(struct device *dev);
452