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