Loading...
1/*
2 * Copyright (C) 2015 Broadcom
3 *
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License version 2 as
6 * published by the Free Software Foundation.
7 */
8
9#include <linux/reservation.h>
10#include <drm/drmP.h>
11#include <drm/drm_encoder.h>
12#include <drm/drm_gem_cma_helper.h>
13
14#include "uapi/drm/vc4_drm.h"
15
16/* Don't forget to update vc4_bo.c: bo_type_names[] when adding to
17 * this.
18 */
19enum vc4_kernel_bo_type {
20 /* Any kernel allocation (gem_create_object hook) before it
21 * gets another type set.
22 */
23 VC4_BO_TYPE_KERNEL,
24 VC4_BO_TYPE_V3D,
25 VC4_BO_TYPE_V3D_SHADER,
26 VC4_BO_TYPE_DUMB,
27 VC4_BO_TYPE_BIN,
28 VC4_BO_TYPE_RCL,
29 VC4_BO_TYPE_BCL,
30 VC4_BO_TYPE_KERNEL_CACHE,
31 VC4_BO_TYPE_COUNT
32};
33
34/* Performance monitor object. The perform lifetime is controlled by userspace
35 * using perfmon related ioctls. A perfmon can be attached to a submit_cl
36 * request, and when this is the case, HW perf counters will be activated just
37 * before the submit_cl is submitted to the GPU and disabled when the job is
38 * done. This way, only events related to a specific job will be counted.
39 */
40struct vc4_perfmon {
41 /* Tracks the number of users of the perfmon, when this counter reaches
42 * zero the perfmon is destroyed.
43 */
44 refcount_t refcnt;
45
46 /* Number of counters activated in this perfmon instance
47 * (should be less than DRM_VC4_MAX_PERF_COUNTERS).
48 */
49 u8 ncounters;
50
51 /* Events counted by the HW perf counters. */
52 u8 events[DRM_VC4_MAX_PERF_COUNTERS];
53
54 /* Storage for counter values. Counters are incremented by the HW
55 * perf counter values every time the perfmon is attached to a GPU job.
56 * This way, perfmon users don't have to retrieve the results after
57 * each job if they want to track events covering several submissions.
58 * Note that counter values can't be reset, but you can fake a reset by
59 * destroying the perfmon and creating a new one.
60 */
61 u64 counters[0];
62};
63
64struct vc4_dev {
65 struct drm_device *dev;
66
67 struct vc4_hdmi *hdmi;
68 struct vc4_hvs *hvs;
69 struct vc4_v3d *v3d;
70 struct vc4_dpi *dpi;
71 struct vc4_dsi *dsi1;
72 struct vc4_vec *vec;
73
74 struct vc4_hang_state *hang_state;
75
76 /* The kernel-space BO cache. Tracks buffers that have been
77 * unreferenced by all other users (refcounts of 0!) but not
78 * yet freed, so we can do cheap allocations.
79 */
80 struct vc4_bo_cache {
81 /* Array of list heads for entries in the BO cache,
82 * based on number of pages, so we can do O(1) lookups
83 * in the cache when allocating.
84 */
85 struct list_head *size_list;
86 uint32_t size_list_size;
87
88 /* List of all BOs in the cache, ordered by age, so we
89 * can do O(1) lookups when trying to free old
90 * buffers.
91 */
92 struct list_head time_list;
93 struct work_struct time_work;
94 struct timer_list time_timer;
95 } bo_cache;
96
97 u32 num_labels;
98 struct vc4_label {
99 const char *name;
100 u32 num_allocated;
101 u32 size_allocated;
102 } *bo_labels;
103
104 /* Protects bo_cache and bo_labels. */
105 struct mutex bo_lock;
106
107 /* Purgeable BO pool. All BOs in this pool can have their memory
108 * reclaimed if the driver is unable to allocate new BOs. We also
109 * keep stats related to the purge mechanism here.
110 */
111 struct {
112 struct list_head list;
113 unsigned int num;
114 size_t size;
115 unsigned int purged_num;
116 size_t purged_size;
117 struct mutex lock;
118 } purgeable;
119
120 uint64_t dma_fence_context;
121
122 /* Sequence number for the last job queued in bin_job_list.
123 * Starts at 0 (no jobs emitted).
124 */
125 uint64_t emit_seqno;
126
127 /* Sequence number for the last completed job on the GPU.
128 * Starts at 0 (no jobs completed).
129 */
130 uint64_t finished_seqno;
131
132 /* List of all struct vc4_exec_info for jobs to be executed in
133 * the binner. The first job in the list is the one currently
134 * programmed into ct0ca for execution.
135 */
136 struct list_head bin_job_list;
137
138 /* List of all struct vc4_exec_info for jobs that have
139 * completed binning and are ready for rendering. The first
140 * job in the list is the one currently programmed into ct1ca
141 * for execution.
142 */
143 struct list_head render_job_list;
144
145 /* List of the finished vc4_exec_infos waiting to be freed by
146 * job_done_work.
147 */
148 struct list_head job_done_list;
149 /* Spinlock used to synchronize the job_list and seqno
150 * accesses between the IRQ handler and GEM ioctls.
151 */
152 spinlock_t job_lock;
153 wait_queue_head_t job_wait_queue;
154 struct work_struct job_done_work;
155
156 /* Used to track the active perfmon if any. Access to this field is
157 * protected by job_lock.
158 */
159 struct vc4_perfmon *active_perfmon;
160
161 /* List of struct vc4_seqno_cb for callbacks to be made from a
162 * workqueue when the given seqno is passed.
163 */
164 struct list_head seqno_cb_list;
165
166 /* The memory used for storing binner tile alloc, tile state,
167 * and overflow memory allocations. This is freed when V3D
168 * powers down.
169 */
170 struct vc4_bo *bin_bo;
171
172 /* Size of blocks allocated within bin_bo. */
173 uint32_t bin_alloc_size;
174
175 /* Bitmask of the bin_alloc_size chunks in bin_bo that are
176 * used.
177 */
178 uint32_t bin_alloc_used;
179
180 /* Bitmask of the current bin_alloc used for overflow memory. */
181 uint32_t bin_alloc_overflow;
182
183 struct work_struct overflow_mem_work;
184
185 int power_refcount;
186
187 /* Mutex controlling the power refcount. */
188 struct mutex power_lock;
189
190 struct {
191 struct timer_list timer;
192 struct work_struct reset_work;
193 } hangcheck;
194
195 struct semaphore async_modeset;
196};
197
198static inline struct vc4_dev *
199to_vc4_dev(struct drm_device *dev)
200{
201 return (struct vc4_dev *)dev->dev_private;
202}
203
204struct vc4_bo {
205 struct drm_gem_cma_object base;
206
207 /* seqno of the last job to render using this BO. */
208 uint64_t seqno;
209
210 /* seqno of the last job to use the RCL to write to this BO.
211 *
212 * Note that this doesn't include binner overflow memory
213 * writes.
214 */
215 uint64_t write_seqno;
216
217 bool t_format;
218
219 /* List entry for the BO's position in either
220 * vc4_exec_info->unref_list or vc4_dev->bo_cache.time_list
221 */
222 struct list_head unref_head;
223
224 /* Time in jiffies when the BO was put in vc4->bo_cache. */
225 unsigned long free_time;
226
227 /* List entry for the BO's position in vc4_dev->bo_cache.size_list */
228 struct list_head size_head;
229
230 /* Struct for shader validation state, if created by
231 * DRM_IOCTL_VC4_CREATE_SHADER_BO.
232 */
233 struct vc4_validated_shader_info *validated_shader;
234
235 /* normally (resv == &_resv) except for imported bo's */
236 struct reservation_object *resv;
237 struct reservation_object _resv;
238
239 /* One of enum vc4_kernel_bo_type, or VC4_BO_TYPE_COUNT + i
240 * for user-allocated labels.
241 */
242 int label;
243
244 /* Count the number of active users. This is needed to determine
245 * whether we can move the BO to the purgeable list or not (when the BO
246 * is used by the GPU or the display engine we can't purge it).
247 */
248 refcount_t usecnt;
249
250 /* Store purgeable/purged state here */
251 u32 madv;
252 struct mutex madv_lock;
253};
254
255static inline struct vc4_bo *
256to_vc4_bo(struct drm_gem_object *bo)
257{
258 return (struct vc4_bo *)bo;
259}
260
261struct vc4_fence {
262 struct dma_fence base;
263 struct drm_device *dev;
264 /* vc4 seqno for signaled() test */
265 uint64_t seqno;
266};
267
268static inline struct vc4_fence *
269to_vc4_fence(struct dma_fence *fence)
270{
271 return (struct vc4_fence *)fence;
272}
273
274struct vc4_seqno_cb {
275 struct work_struct work;
276 uint64_t seqno;
277 void (*func)(struct vc4_seqno_cb *cb);
278};
279
280struct vc4_v3d {
281 struct vc4_dev *vc4;
282 struct platform_device *pdev;
283 void __iomem *regs;
284 struct clk *clk;
285};
286
287struct vc4_hvs {
288 struct platform_device *pdev;
289 void __iomem *regs;
290 u32 __iomem *dlist;
291
292 /* Memory manager for CRTCs to allocate space in the display
293 * list. Units are dwords.
294 */
295 struct drm_mm dlist_mm;
296 /* Memory manager for the LBM memory used by HVS scaling. */
297 struct drm_mm lbm_mm;
298 spinlock_t mm_lock;
299
300 struct drm_mm_node mitchell_netravali_filter;
301};
302
303struct vc4_plane {
304 struct drm_plane base;
305};
306
307static inline struct vc4_plane *
308to_vc4_plane(struct drm_plane *plane)
309{
310 return (struct vc4_plane *)plane;
311}
312
313enum vc4_scaling_mode {
314 VC4_SCALING_NONE,
315 VC4_SCALING_TPZ,
316 VC4_SCALING_PPF,
317};
318
319struct vc4_plane_state {
320 struct drm_plane_state base;
321 /* System memory copy of the display list for this element, computed
322 * at atomic_check time.
323 */
324 u32 *dlist;
325 u32 dlist_size; /* Number of dwords allocated for the display list */
326 u32 dlist_count; /* Number of used dwords in the display list. */
327
328 /* Offset in the dlist to various words, for pageflip or
329 * cursor updates.
330 */
331 u32 pos0_offset;
332 u32 pos2_offset;
333 u32 ptr0_offset;
334
335 /* Offset where the plane's dlist was last stored in the
336 * hardware at vc4_crtc_atomic_flush() time.
337 */
338 u32 __iomem *hw_dlist;
339
340 /* Clipped coordinates of the plane on the display. */
341 int crtc_x, crtc_y, crtc_w, crtc_h;
342 /* Clipped area being scanned from in the FB. */
343 u32 src_x, src_y;
344
345 u32 src_w[2], src_h[2];
346
347 /* Scaling selection for the RGB/Y plane and the Cb/Cr planes. */
348 enum vc4_scaling_mode x_scaling[2], y_scaling[2];
349 bool is_unity;
350 bool is_yuv;
351
352 /* Offset to start scanning out from the start of the plane's
353 * BO.
354 */
355 u32 offsets[3];
356
357 /* Our allocation in LBM for temporary storage during scaling. */
358 struct drm_mm_node lbm;
359
360 /* Set when the plane has per-pixel alpha content or does not cover
361 * the entire screen. This is a hint to the CRTC that it might need
362 * to enable background color fill.
363 */
364 bool needs_bg_fill;
365};
366
367static inline struct vc4_plane_state *
368to_vc4_plane_state(struct drm_plane_state *state)
369{
370 return (struct vc4_plane_state *)state;
371}
372
373enum vc4_encoder_type {
374 VC4_ENCODER_TYPE_NONE,
375 VC4_ENCODER_TYPE_HDMI,
376 VC4_ENCODER_TYPE_VEC,
377 VC4_ENCODER_TYPE_DSI0,
378 VC4_ENCODER_TYPE_DSI1,
379 VC4_ENCODER_TYPE_SMI,
380 VC4_ENCODER_TYPE_DPI,
381};
382
383struct vc4_encoder {
384 struct drm_encoder base;
385 enum vc4_encoder_type type;
386 u32 clock_select;
387};
388
389static inline struct vc4_encoder *
390to_vc4_encoder(struct drm_encoder *encoder)
391{
392 return container_of(encoder, struct vc4_encoder, base);
393}
394
395#define V3D_READ(offset) readl(vc4->v3d->regs + offset)
396#define V3D_WRITE(offset, val) writel(val, vc4->v3d->regs + offset)
397#define HVS_READ(offset) readl(vc4->hvs->regs + offset)
398#define HVS_WRITE(offset, val) writel(val, vc4->hvs->regs + offset)
399
400struct vc4_exec_info {
401 /* Sequence number for this bin/render job. */
402 uint64_t seqno;
403
404 /* Latest write_seqno of any BO that binning depends on. */
405 uint64_t bin_dep_seqno;
406
407 struct dma_fence *fence;
408
409 /* Last current addresses the hardware was processing when the
410 * hangcheck timer checked on us.
411 */
412 uint32_t last_ct0ca, last_ct1ca;
413
414 /* Kernel-space copy of the ioctl arguments */
415 struct drm_vc4_submit_cl *args;
416
417 /* This is the array of BOs that were looked up at the start of exec.
418 * Command validation will use indices into this array.
419 */
420 struct drm_gem_cma_object **bo;
421 uint32_t bo_count;
422
423 /* List of BOs that are being written by the RCL. Other than
424 * the binner temporary storage, this is all the BOs written
425 * by the job.
426 */
427 struct drm_gem_cma_object *rcl_write_bo[4];
428 uint32_t rcl_write_bo_count;
429
430 /* Pointers for our position in vc4->job_list */
431 struct list_head head;
432
433 /* List of other BOs used in the job that need to be released
434 * once the job is complete.
435 */
436 struct list_head unref_list;
437
438 /* Current unvalidated indices into @bo loaded by the non-hardware
439 * VC4_PACKET_GEM_HANDLES.
440 */
441 uint32_t bo_index[2];
442
443 /* This is the BO where we store the validated command lists, shader
444 * records, and uniforms.
445 */
446 struct drm_gem_cma_object *exec_bo;
447
448 /**
449 * This tracks the per-shader-record state (packet 64) that
450 * determines the length of the shader record and the offset
451 * it's expected to be found at. It gets read in from the
452 * command lists.
453 */
454 struct vc4_shader_state {
455 uint32_t addr;
456 /* Maximum vertex index referenced by any primitive using this
457 * shader state.
458 */
459 uint32_t max_index;
460 } *shader_state;
461
462 /** How many shader states the user declared they were using. */
463 uint32_t shader_state_size;
464 /** How many shader state records the validator has seen. */
465 uint32_t shader_state_count;
466
467 bool found_tile_binning_mode_config_packet;
468 bool found_start_tile_binning_packet;
469 bool found_increment_semaphore_packet;
470 bool found_flush;
471 uint8_t bin_tiles_x, bin_tiles_y;
472 /* Physical address of the start of the tile alloc array
473 * (where each tile's binned CL will start)
474 */
475 uint32_t tile_alloc_offset;
476 /* Bitmask of which binner slots are freed when this job completes. */
477 uint32_t bin_slots;
478
479 /**
480 * Computed addresses pointing into exec_bo where we start the
481 * bin thread (ct0) and render thread (ct1).
482 */
483 uint32_t ct0ca, ct0ea;
484 uint32_t ct1ca, ct1ea;
485
486 /* Pointer to the unvalidated bin CL (if present). */
487 void *bin_u;
488
489 /* Pointers to the shader recs. These paddr gets incremented as CL
490 * packets are relocated in validate_gl_shader_state, and the vaddrs
491 * (u and v) get incremented and size decremented as the shader recs
492 * themselves are validated.
493 */
494 void *shader_rec_u;
495 void *shader_rec_v;
496 uint32_t shader_rec_p;
497 uint32_t shader_rec_size;
498
499 /* Pointers to the uniform data. These pointers are incremented, and
500 * size decremented, as each batch of uniforms is uploaded.
501 */
502 void *uniforms_u;
503 void *uniforms_v;
504 uint32_t uniforms_p;
505 uint32_t uniforms_size;
506
507 /* Pointer to a performance monitor object if the user requested it,
508 * NULL otherwise.
509 */
510 struct vc4_perfmon *perfmon;
511};
512
513/* Per-open file private data. Any driver-specific resource that has to be
514 * released when the DRM file is closed should be placed here.
515 */
516struct vc4_file {
517 struct {
518 struct idr idr;
519 struct mutex lock;
520 } perfmon;
521};
522
523static inline struct vc4_exec_info *
524vc4_first_bin_job(struct vc4_dev *vc4)
525{
526 return list_first_entry_or_null(&vc4->bin_job_list,
527 struct vc4_exec_info, head);
528}
529
530static inline struct vc4_exec_info *
531vc4_first_render_job(struct vc4_dev *vc4)
532{
533 return list_first_entry_or_null(&vc4->render_job_list,
534 struct vc4_exec_info, head);
535}
536
537static inline struct vc4_exec_info *
538vc4_last_render_job(struct vc4_dev *vc4)
539{
540 if (list_empty(&vc4->render_job_list))
541 return NULL;
542 return list_last_entry(&vc4->render_job_list,
543 struct vc4_exec_info, head);
544}
545
546/**
547 * struct vc4_texture_sample_info - saves the offsets into the UBO for texture
548 * setup parameters.
549 *
550 * This will be used at draw time to relocate the reference to the texture
551 * contents in p0, and validate that the offset combined with
552 * width/height/stride/etc. from p1 and p2/p3 doesn't sample outside the BO.
553 * Note that the hardware treats unprovided config parameters as 0, so not all
554 * of them need to be set up for every texure sample, and we'll store ~0 as
555 * the offset to mark the unused ones.
556 *
557 * See the VC4 3D architecture guide page 41 ("Texture and Memory Lookup Unit
558 * Setup") for definitions of the texture parameters.
559 */
560struct vc4_texture_sample_info {
561 bool is_direct;
562 uint32_t p_offset[4];
563};
564
565/**
566 * struct vc4_validated_shader_info - information about validated shaders that
567 * needs to be used from command list validation.
568 *
569 * For a given shader, each time a shader state record references it, we need
570 * to verify that the shader doesn't read more uniforms than the shader state
571 * record's uniform BO pointer can provide, and we need to apply relocations
572 * and validate the shader state record's uniforms that define the texture
573 * samples.
574 */
575struct vc4_validated_shader_info {
576 uint32_t uniforms_size;
577 uint32_t uniforms_src_size;
578 uint32_t num_texture_samples;
579 struct vc4_texture_sample_info *texture_samples;
580
581 uint32_t num_uniform_addr_offsets;
582 uint32_t *uniform_addr_offsets;
583
584 bool is_threaded;
585};
586
587/**
588 * _wait_for - magic (register) wait macro
589 *
590 * Does the right thing for modeset paths when run under kdgb or similar atomic
591 * contexts. Note that it's important that we check the condition again after
592 * having timed out, since the timeout could be due to preemption or similar and
593 * we've never had a chance to check the condition before the timeout.
594 */
595#define _wait_for(COND, MS, W) ({ \
596 unsigned long timeout__ = jiffies + msecs_to_jiffies(MS) + 1; \
597 int ret__ = 0; \
598 while (!(COND)) { \
599 if (time_after(jiffies, timeout__)) { \
600 if (!(COND)) \
601 ret__ = -ETIMEDOUT; \
602 break; \
603 } \
604 if (W && drm_can_sleep()) { \
605 msleep(W); \
606 } else { \
607 cpu_relax(); \
608 } \
609 } \
610 ret__; \
611})
612
613#define wait_for(COND, MS) _wait_for(COND, MS, 1)
614
615/* vc4_bo.c */
616struct drm_gem_object *vc4_create_object(struct drm_device *dev, size_t size);
617void vc4_free_object(struct drm_gem_object *gem_obj);
618struct vc4_bo *vc4_bo_create(struct drm_device *dev, size_t size,
619 bool from_cache, enum vc4_kernel_bo_type type);
620int vc4_dumb_create(struct drm_file *file_priv,
621 struct drm_device *dev,
622 struct drm_mode_create_dumb *args);
623struct dma_buf *vc4_prime_export(struct drm_device *dev,
624 struct drm_gem_object *obj, int flags);
625int vc4_create_bo_ioctl(struct drm_device *dev, void *data,
626 struct drm_file *file_priv);
627int vc4_create_shader_bo_ioctl(struct drm_device *dev, void *data,
628 struct drm_file *file_priv);
629int vc4_mmap_bo_ioctl(struct drm_device *dev, void *data,
630 struct drm_file *file_priv);
631int vc4_set_tiling_ioctl(struct drm_device *dev, void *data,
632 struct drm_file *file_priv);
633int vc4_get_tiling_ioctl(struct drm_device *dev, void *data,
634 struct drm_file *file_priv);
635int vc4_get_hang_state_ioctl(struct drm_device *dev, void *data,
636 struct drm_file *file_priv);
637int vc4_label_bo_ioctl(struct drm_device *dev, void *data,
638 struct drm_file *file_priv);
639int vc4_fault(struct vm_fault *vmf);
640int vc4_mmap(struct file *filp, struct vm_area_struct *vma);
641struct reservation_object *vc4_prime_res_obj(struct drm_gem_object *obj);
642int vc4_prime_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma);
643struct drm_gem_object *vc4_prime_import_sg_table(struct drm_device *dev,
644 struct dma_buf_attachment *attach,
645 struct sg_table *sgt);
646void *vc4_prime_vmap(struct drm_gem_object *obj);
647int vc4_bo_cache_init(struct drm_device *dev);
648void vc4_bo_cache_destroy(struct drm_device *dev);
649int vc4_bo_stats_debugfs(struct seq_file *m, void *arg);
650int vc4_bo_inc_usecnt(struct vc4_bo *bo);
651void vc4_bo_dec_usecnt(struct vc4_bo *bo);
652void vc4_bo_add_to_purgeable_pool(struct vc4_bo *bo);
653void vc4_bo_remove_from_purgeable_pool(struct vc4_bo *bo);
654
655/* vc4_crtc.c */
656extern struct platform_driver vc4_crtc_driver;
657int vc4_crtc_debugfs_regs(struct seq_file *m, void *arg);
658bool vc4_crtc_get_scanoutpos(struct drm_device *dev, unsigned int crtc_id,
659 bool in_vblank_irq, int *vpos, int *hpos,
660 ktime_t *stime, ktime_t *etime,
661 const struct drm_display_mode *mode);
662
663/* vc4_debugfs.c */
664int vc4_debugfs_init(struct drm_minor *minor);
665
666/* vc4_drv.c */
667void __iomem *vc4_ioremap_regs(struct platform_device *dev, int index);
668
669/* vc4_dpi.c */
670extern struct platform_driver vc4_dpi_driver;
671int vc4_dpi_debugfs_regs(struct seq_file *m, void *unused);
672
673/* vc4_dsi.c */
674extern struct platform_driver vc4_dsi_driver;
675int vc4_dsi_debugfs_regs(struct seq_file *m, void *unused);
676
677/* vc4_fence.c */
678extern const struct dma_fence_ops vc4_fence_ops;
679
680/* vc4_gem.c */
681void vc4_gem_init(struct drm_device *dev);
682void vc4_gem_destroy(struct drm_device *dev);
683int vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
684 struct drm_file *file_priv);
685int vc4_wait_seqno_ioctl(struct drm_device *dev, void *data,
686 struct drm_file *file_priv);
687int vc4_wait_bo_ioctl(struct drm_device *dev, void *data,
688 struct drm_file *file_priv);
689void vc4_submit_next_bin_job(struct drm_device *dev);
690void vc4_submit_next_render_job(struct drm_device *dev);
691void vc4_move_job_to_render(struct drm_device *dev, struct vc4_exec_info *exec);
692int vc4_wait_for_seqno(struct drm_device *dev, uint64_t seqno,
693 uint64_t timeout_ns, bool interruptible);
694void vc4_job_handle_completed(struct vc4_dev *vc4);
695int vc4_queue_seqno_cb(struct drm_device *dev,
696 struct vc4_seqno_cb *cb, uint64_t seqno,
697 void (*func)(struct vc4_seqno_cb *cb));
698int vc4_gem_madvise_ioctl(struct drm_device *dev, void *data,
699 struct drm_file *file_priv);
700
701/* vc4_hdmi.c */
702extern struct platform_driver vc4_hdmi_driver;
703int vc4_hdmi_debugfs_regs(struct seq_file *m, void *unused);
704
705/* vc4_vec.c */
706extern struct platform_driver vc4_vec_driver;
707int vc4_vec_debugfs_regs(struct seq_file *m, void *unused);
708
709/* vc4_irq.c */
710irqreturn_t vc4_irq(int irq, void *arg);
711void vc4_irq_preinstall(struct drm_device *dev);
712int vc4_irq_postinstall(struct drm_device *dev);
713void vc4_irq_uninstall(struct drm_device *dev);
714void vc4_irq_reset(struct drm_device *dev);
715
716/* vc4_hvs.c */
717extern struct platform_driver vc4_hvs_driver;
718void vc4_hvs_dump_state(struct drm_device *dev);
719int vc4_hvs_debugfs_regs(struct seq_file *m, void *unused);
720
721/* vc4_kms.c */
722int vc4_kms_load(struct drm_device *dev);
723
724/* vc4_plane.c */
725struct drm_plane *vc4_plane_init(struct drm_device *dev,
726 enum drm_plane_type type);
727u32 vc4_plane_write_dlist(struct drm_plane *plane, u32 __iomem *dlist);
728u32 vc4_plane_dlist_size(const struct drm_plane_state *state);
729void vc4_plane_async_set_fb(struct drm_plane *plane,
730 struct drm_framebuffer *fb);
731
732/* vc4_v3d.c */
733extern struct platform_driver vc4_v3d_driver;
734int vc4_v3d_debugfs_ident(struct seq_file *m, void *unused);
735int vc4_v3d_debugfs_regs(struct seq_file *m, void *unused);
736int vc4_v3d_get_bin_slot(struct vc4_dev *vc4);
737
738/* vc4_validate.c */
739int
740vc4_validate_bin_cl(struct drm_device *dev,
741 void *validated,
742 void *unvalidated,
743 struct vc4_exec_info *exec);
744
745int
746vc4_validate_shader_recs(struct drm_device *dev, struct vc4_exec_info *exec);
747
748struct drm_gem_cma_object *vc4_use_bo(struct vc4_exec_info *exec,
749 uint32_t hindex);
750
751int vc4_get_rcl(struct drm_device *dev, struct vc4_exec_info *exec);
752
753bool vc4_check_tex_size(struct vc4_exec_info *exec,
754 struct drm_gem_cma_object *fbo,
755 uint32_t offset, uint8_t tiling_format,
756 uint32_t width, uint32_t height, uint8_t cpp);
757
758/* vc4_validate_shader.c */
759struct vc4_validated_shader_info *
760vc4_validate_shader(struct drm_gem_cma_object *shader_obj);
761
762/* vc4_perfmon.c */
763void vc4_perfmon_get(struct vc4_perfmon *perfmon);
764void vc4_perfmon_put(struct vc4_perfmon *perfmon);
765void vc4_perfmon_start(struct vc4_dev *vc4, struct vc4_perfmon *perfmon);
766void vc4_perfmon_stop(struct vc4_dev *vc4, struct vc4_perfmon *perfmon,
767 bool capture);
768struct vc4_perfmon *vc4_perfmon_find(struct vc4_file *vc4file, int id);
769void vc4_perfmon_open_file(struct vc4_file *vc4file);
770void vc4_perfmon_close_file(struct vc4_file *vc4file);
771int vc4_perfmon_create_ioctl(struct drm_device *dev, void *data,
772 struct drm_file *file_priv);
773int vc4_perfmon_destroy_ioctl(struct drm_device *dev, void *data,
774 struct drm_file *file_priv);
775int vc4_perfmon_get_values_ioctl(struct drm_device *dev, void *data,
776 struct drm_file *file_priv);
1/* SPDX-License-Identifier: GPL-2.0-only */
2/*
3 * Copyright (C) 2015 Broadcom
4 */
5#ifndef _VC4_DRV_H_
6#define _VC4_DRV_H_
7
8#include <linux/delay.h>
9#include <linux/of.h>
10#include <linux/refcount.h>
11#include <linux/uaccess.h>
12
13#include <drm/drm_atomic.h>
14#include <drm/drm_debugfs.h>
15#include <drm/drm_device.h>
16#include <drm/drm_encoder.h>
17#include <drm/drm_gem_dma_helper.h>
18#include <drm/drm_managed.h>
19#include <drm/drm_mm.h>
20#include <drm/drm_modeset_lock.h>
21
22#include <kunit/test-bug.h>
23
24#include "uapi/drm/vc4_drm.h"
25
26struct drm_device;
27struct drm_gem_object;
28
29extern const struct drm_driver vc4_drm_driver;
30extern const struct drm_driver vc5_drm_driver;
31
32/* Don't forget to update vc4_bo.c: bo_type_names[] when adding to
33 * this.
34 */
35enum vc4_kernel_bo_type {
36 /* Any kernel allocation (gem_create_object hook) before it
37 * gets another type set.
38 */
39 VC4_BO_TYPE_KERNEL,
40 VC4_BO_TYPE_V3D,
41 VC4_BO_TYPE_V3D_SHADER,
42 VC4_BO_TYPE_DUMB,
43 VC4_BO_TYPE_BIN,
44 VC4_BO_TYPE_RCL,
45 VC4_BO_TYPE_BCL,
46 VC4_BO_TYPE_KERNEL_CACHE,
47 VC4_BO_TYPE_COUNT
48};
49
50/* Performance monitor object. The perform lifetime is controlled by userspace
51 * using perfmon related ioctls. A perfmon can be attached to a submit_cl
52 * request, and when this is the case, HW perf counters will be activated just
53 * before the submit_cl is submitted to the GPU and disabled when the job is
54 * done. This way, only events related to a specific job will be counted.
55 */
56struct vc4_perfmon {
57 struct vc4_dev *dev;
58
59 /* Tracks the number of users of the perfmon, when this counter reaches
60 * zero the perfmon is destroyed.
61 */
62 refcount_t refcnt;
63
64 /* Number of counters activated in this perfmon instance
65 * (should be less than DRM_VC4_MAX_PERF_COUNTERS).
66 */
67 u8 ncounters;
68
69 /* Events counted by the HW perf counters. */
70 u8 events[DRM_VC4_MAX_PERF_COUNTERS];
71
72 /* Storage for counter values. Counters are incremented by the HW
73 * perf counter values every time the perfmon is attached to a GPU job.
74 * This way, perfmon users don't have to retrieve the results after
75 * each job if they want to track events covering several submissions.
76 * Note that counter values can't be reset, but you can fake a reset by
77 * destroying the perfmon and creating a new one.
78 */
79 u64 counters[] __counted_by(ncounters);
80};
81
82struct vc4_dev {
83 struct drm_device base;
84 struct device *dev;
85
86 bool is_vc5;
87
88 unsigned int irq;
89
90 struct vc4_hvs *hvs;
91 struct vc4_v3d *v3d;
92
93 struct vc4_hang_state *hang_state;
94
95 /* The kernel-space BO cache. Tracks buffers that have been
96 * unreferenced by all other users (refcounts of 0!) but not
97 * yet freed, so we can do cheap allocations.
98 */
99 struct vc4_bo_cache {
100 /* Array of list heads for entries in the BO cache,
101 * based on number of pages, so we can do O(1) lookups
102 * in the cache when allocating.
103 */
104 struct list_head *size_list;
105 uint32_t size_list_size;
106
107 /* List of all BOs in the cache, ordered by age, so we
108 * can do O(1) lookups when trying to free old
109 * buffers.
110 */
111 struct list_head time_list;
112 struct work_struct time_work;
113 struct timer_list time_timer;
114 } bo_cache;
115
116 u32 num_labels;
117 struct vc4_label {
118 const char *name;
119 u32 num_allocated;
120 u32 size_allocated;
121 } *bo_labels;
122
123 /* Protects bo_cache and bo_labels. */
124 struct mutex bo_lock;
125
126 /* Purgeable BO pool. All BOs in this pool can have their memory
127 * reclaimed if the driver is unable to allocate new BOs. We also
128 * keep stats related to the purge mechanism here.
129 */
130 struct {
131 struct list_head list;
132 unsigned int num;
133 size_t size;
134 unsigned int purged_num;
135 size_t purged_size;
136 struct mutex lock;
137 } purgeable;
138
139 uint64_t dma_fence_context;
140
141 /* Sequence number for the last job queued in bin_job_list.
142 * Starts at 0 (no jobs emitted).
143 */
144 uint64_t emit_seqno;
145
146 /* Sequence number for the last completed job on the GPU.
147 * Starts at 0 (no jobs completed).
148 */
149 uint64_t finished_seqno;
150
151 /* List of all struct vc4_exec_info for jobs to be executed in
152 * the binner. The first job in the list is the one currently
153 * programmed into ct0ca for execution.
154 */
155 struct list_head bin_job_list;
156
157 /* List of all struct vc4_exec_info for jobs that have
158 * completed binning and are ready for rendering. The first
159 * job in the list is the one currently programmed into ct1ca
160 * for execution.
161 */
162 struct list_head render_job_list;
163
164 /* List of the finished vc4_exec_infos waiting to be freed by
165 * job_done_work.
166 */
167 struct list_head job_done_list;
168 /* Spinlock used to synchronize the job_list and seqno
169 * accesses between the IRQ handler and GEM ioctls.
170 */
171 spinlock_t job_lock;
172 wait_queue_head_t job_wait_queue;
173 struct work_struct job_done_work;
174
175 /* Used to track the active perfmon if any. Access to this field is
176 * protected by job_lock.
177 */
178 struct vc4_perfmon *active_perfmon;
179
180 /* List of struct vc4_seqno_cb for callbacks to be made from a
181 * workqueue when the given seqno is passed.
182 */
183 struct list_head seqno_cb_list;
184
185 /* The memory used for storing binner tile alloc, tile state,
186 * and overflow memory allocations. This is freed when V3D
187 * powers down.
188 */
189 struct vc4_bo *bin_bo;
190
191 /* Size of blocks allocated within bin_bo. */
192 uint32_t bin_alloc_size;
193
194 /* Bitmask of the bin_alloc_size chunks in bin_bo that are
195 * used.
196 */
197 uint32_t bin_alloc_used;
198
199 /* Bitmask of the current bin_alloc used for overflow memory. */
200 uint32_t bin_alloc_overflow;
201
202 /* Incremented when an underrun error happened after an atomic commit.
203 * This is particularly useful to detect when a specific modeset is too
204 * demanding in term of memory or HVS bandwidth which is hard to guess
205 * at atomic check time.
206 */
207 atomic_t underrun;
208
209 struct work_struct overflow_mem_work;
210
211 int power_refcount;
212
213 /* Set to true when the load tracker is active. */
214 bool load_tracker_enabled;
215
216 /* Mutex controlling the power refcount. */
217 struct mutex power_lock;
218
219 struct {
220 struct timer_list timer;
221 struct work_struct reset_work;
222 } hangcheck;
223
224 struct drm_modeset_lock ctm_state_lock;
225 struct drm_private_obj ctm_manager;
226 struct drm_private_obj hvs_channels;
227 struct drm_private_obj load_tracker;
228
229 /* Mutex for binner bo allocation. */
230 struct mutex bin_bo_lock;
231 /* Reference count for our binner bo. */
232 struct kref bin_bo_kref;
233};
234
235#define to_vc4_dev(_dev) \
236 container_of_const(_dev, struct vc4_dev, base)
237
238struct vc4_bo {
239 struct drm_gem_dma_object base;
240
241 /* seqno of the last job to render using this BO. */
242 uint64_t seqno;
243
244 /* seqno of the last job to use the RCL to write to this BO.
245 *
246 * Note that this doesn't include binner overflow memory
247 * writes.
248 */
249 uint64_t write_seqno;
250
251 bool t_format;
252
253 /* List entry for the BO's position in either
254 * vc4_exec_info->unref_list or vc4_dev->bo_cache.time_list
255 */
256 struct list_head unref_head;
257
258 /* Time in jiffies when the BO was put in vc4->bo_cache. */
259 unsigned long free_time;
260
261 /* List entry for the BO's position in vc4_dev->bo_cache.size_list */
262 struct list_head size_head;
263
264 /* Struct for shader validation state, if created by
265 * DRM_IOCTL_VC4_CREATE_SHADER_BO.
266 */
267 struct vc4_validated_shader_info *validated_shader;
268
269 /* One of enum vc4_kernel_bo_type, or VC4_BO_TYPE_COUNT + i
270 * for user-allocated labels.
271 */
272 int label;
273
274 /* Count the number of active users. This is needed to determine
275 * whether we can move the BO to the purgeable list or not (when the BO
276 * is used by the GPU or the display engine we can't purge it).
277 */
278 refcount_t usecnt;
279
280 /* Store purgeable/purged state here */
281 u32 madv;
282 struct mutex madv_lock;
283};
284
285#define to_vc4_bo(_bo) \
286 container_of_const(to_drm_gem_dma_obj(_bo), struct vc4_bo, base)
287
288struct vc4_fence {
289 struct dma_fence base;
290 struct drm_device *dev;
291 /* vc4 seqno for signaled() test */
292 uint64_t seqno;
293};
294
295#define to_vc4_fence(_fence) \
296 container_of_const(_fence, struct vc4_fence, base)
297
298struct vc4_seqno_cb {
299 struct work_struct work;
300 uint64_t seqno;
301 void (*func)(struct vc4_seqno_cb *cb);
302};
303
304struct vc4_v3d {
305 struct vc4_dev *vc4;
306 struct platform_device *pdev;
307 void __iomem *regs;
308 struct clk *clk;
309 struct debugfs_regset32 regset;
310};
311
312struct vc4_hvs {
313 struct vc4_dev *vc4;
314 struct platform_device *pdev;
315 void __iomem *regs;
316 u32 __iomem *dlist;
317
318 struct clk *core_clk;
319
320 unsigned long max_core_rate;
321
322 /* Memory manager for CRTCs to allocate space in the display
323 * list. Units are dwords.
324 */
325 struct drm_mm dlist_mm;
326 /* Memory manager for the LBM memory used by HVS scaling. */
327 struct drm_mm lbm_mm;
328 spinlock_t mm_lock;
329
330 struct drm_mm_node mitchell_netravali_filter;
331
332 struct debugfs_regset32 regset;
333
334 /*
335 * Even if HDMI0 on the RPi4 can output modes requiring a pixel
336 * rate higher than 297MHz, it needs some adjustments in the
337 * config.txt file to be able to do so and thus won't always be
338 * available.
339 */
340 bool vc5_hdmi_enable_hdmi_20;
341
342 /*
343 * 4096x2160@60 requires a core overclock to work, so register
344 * whether that is sufficient.
345 */
346 bool vc5_hdmi_enable_4096by2160;
347};
348
349#define HVS_NUM_CHANNELS 3
350
351struct vc4_hvs_state {
352 struct drm_private_state base;
353 unsigned long core_clock_rate;
354
355 struct {
356 unsigned in_use: 1;
357 unsigned long fifo_load;
358 struct drm_crtc_commit *pending_commit;
359 } fifo_state[HVS_NUM_CHANNELS];
360};
361
362#define to_vc4_hvs_state(_state) \
363 container_of_const(_state, struct vc4_hvs_state, base)
364
365struct vc4_hvs_state *vc4_hvs_get_global_state(struct drm_atomic_state *state);
366struct vc4_hvs_state *vc4_hvs_get_old_global_state(const struct drm_atomic_state *state);
367struct vc4_hvs_state *vc4_hvs_get_new_global_state(const struct drm_atomic_state *state);
368
369struct vc4_plane {
370 struct drm_plane base;
371};
372
373#define to_vc4_plane(_plane) \
374 container_of_const(_plane, struct vc4_plane, base)
375
376enum vc4_scaling_mode {
377 VC4_SCALING_NONE,
378 VC4_SCALING_TPZ,
379 VC4_SCALING_PPF,
380};
381
382struct vc4_plane_state {
383 struct drm_plane_state base;
384 /* System memory copy of the display list for this element, computed
385 * at atomic_check time.
386 */
387 u32 *dlist;
388 u32 dlist_size; /* Number of dwords allocated for the display list */
389 u32 dlist_count; /* Number of used dwords in the display list. */
390
391 /* Offset in the dlist to various words, for pageflip or
392 * cursor updates.
393 */
394 u32 pos0_offset;
395 u32 pos2_offset;
396 u32 ptr0_offset;
397 u32 lbm_offset;
398
399 /* Offset where the plane's dlist was last stored in the
400 * hardware at vc4_crtc_atomic_flush() time.
401 */
402 u32 __iomem *hw_dlist;
403
404 /* Clipped coordinates of the plane on the display. */
405 int crtc_x, crtc_y, crtc_w, crtc_h;
406 /* Clipped area being scanned from in the FB. */
407 u32 src_x, src_y;
408
409 u32 src_w[2], src_h[2];
410
411 /* Scaling selection for the RGB/Y plane and the Cb/Cr planes. */
412 enum vc4_scaling_mode x_scaling[2], y_scaling[2];
413 bool is_unity;
414 bool is_yuv;
415
416 /* Offset to start scanning out from the start of the plane's
417 * BO.
418 */
419 u32 offsets[3];
420
421 /* Our allocation in LBM for temporary storage during scaling. */
422 struct drm_mm_node lbm;
423
424 /* Set when the plane has per-pixel alpha content or does not cover
425 * the entire screen. This is a hint to the CRTC that it might need
426 * to enable background color fill.
427 */
428 bool needs_bg_fill;
429
430 /* Mark the dlist as initialized. Useful to avoid initializing it twice
431 * when async update is not possible.
432 */
433 bool dlist_initialized;
434
435 /* Load of this plane on the HVS block. The load is expressed in HVS
436 * cycles/sec.
437 */
438 u64 hvs_load;
439
440 /* Memory bandwidth needed for this plane. This is expressed in
441 * bytes/sec.
442 */
443 u64 membus_load;
444};
445
446#define to_vc4_plane_state(_state) \
447 container_of_const(_state, struct vc4_plane_state, base)
448
449enum vc4_encoder_type {
450 VC4_ENCODER_TYPE_NONE,
451 VC4_ENCODER_TYPE_HDMI0,
452 VC4_ENCODER_TYPE_HDMI1,
453 VC4_ENCODER_TYPE_VEC,
454 VC4_ENCODER_TYPE_DSI0,
455 VC4_ENCODER_TYPE_DSI1,
456 VC4_ENCODER_TYPE_SMI,
457 VC4_ENCODER_TYPE_DPI,
458 VC4_ENCODER_TYPE_TXP,
459};
460
461struct vc4_encoder {
462 struct drm_encoder base;
463 enum vc4_encoder_type type;
464 u32 clock_select;
465
466 void (*pre_crtc_configure)(struct drm_encoder *encoder, struct drm_atomic_state *state);
467 void (*pre_crtc_enable)(struct drm_encoder *encoder, struct drm_atomic_state *state);
468 void (*post_crtc_enable)(struct drm_encoder *encoder, struct drm_atomic_state *state);
469
470 void (*post_crtc_disable)(struct drm_encoder *encoder, struct drm_atomic_state *state);
471 void (*post_crtc_powerdown)(struct drm_encoder *encoder, struct drm_atomic_state *state);
472};
473
474#define to_vc4_encoder(_encoder) \
475 container_of_const(_encoder, struct vc4_encoder, base)
476
477static inline
478struct drm_encoder *vc4_find_encoder_by_type(struct drm_device *drm,
479 enum vc4_encoder_type type)
480{
481 struct drm_encoder *encoder;
482
483 drm_for_each_encoder(encoder, drm) {
484 struct vc4_encoder *vc4_encoder = to_vc4_encoder(encoder);
485
486 if (vc4_encoder->type == type)
487 return encoder;
488 }
489
490 return NULL;
491}
492
493struct vc4_crtc_data {
494 const char *name;
495
496 const char *debugfs_name;
497
498 /* Bitmask of channels (FIFOs) of the HVS that the output can source from */
499 unsigned int hvs_available_channels;
500
501 /* Which output of the HVS this pixelvalve sources from. */
502 int hvs_output;
503};
504
505extern const struct vc4_crtc_data vc4_txp_crtc_data;
506
507struct vc4_pv_data {
508 struct vc4_crtc_data base;
509
510 /* Depth of the PixelValve FIFO in bytes */
511 unsigned int fifo_depth;
512
513 /* Number of pixels output per clock period */
514 u8 pixels_per_clock;
515
516 enum vc4_encoder_type encoder_types[4];
517};
518
519extern const struct vc4_pv_data bcm2835_pv0_data;
520extern const struct vc4_pv_data bcm2835_pv1_data;
521extern const struct vc4_pv_data bcm2835_pv2_data;
522extern const struct vc4_pv_data bcm2711_pv0_data;
523extern const struct vc4_pv_data bcm2711_pv1_data;
524extern const struct vc4_pv_data bcm2711_pv2_data;
525extern const struct vc4_pv_data bcm2711_pv3_data;
526extern const struct vc4_pv_data bcm2711_pv4_data;
527
528struct vc4_crtc {
529 struct drm_crtc base;
530 struct platform_device *pdev;
531 const struct vc4_crtc_data *data;
532 void __iomem *regs;
533
534 /* Timestamp at start of vblank irq - unaffected by lock delays. */
535 ktime_t t_vblank;
536
537 u8 lut_r[256];
538 u8 lut_g[256];
539 u8 lut_b[256];
540
541 struct drm_pending_vblank_event *event;
542
543 struct debugfs_regset32 regset;
544
545 /**
546 * @feeds_txp: True if the CRTC feeds our writeback controller.
547 */
548 bool feeds_txp;
549
550 /**
551 * @irq_lock: Spinlock protecting the resources shared between
552 * the atomic code and our vblank handler.
553 */
554 spinlock_t irq_lock;
555
556 /**
557 * @current_dlist: Start offset of the display list currently
558 * set in the HVS for that CRTC. Protected by @irq_lock, and
559 * copied in vc4_hvs_update_dlist() for the CRTC interrupt
560 * handler to have access to that value.
561 */
562 unsigned int current_dlist;
563
564 /**
565 * @current_hvs_channel: HVS channel currently assigned to the
566 * CRTC. Protected by @irq_lock, and copied in
567 * vc4_hvs_atomic_begin() for the CRTC interrupt handler to have
568 * access to that value.
569 */
570 unsigned int current_hvs_channel;
571};
572
573#define to_vc4_crtc(_crtc) \
574 container_of_const(_crtc, struct vc4_crtc, base)
575
576static inline const struct vc4_crtc_data *
577vc4_crtc_to_vc4_crtc_data(const struct vc4_crtc *crtc)
578{
579 return crtc->data;
580}
581
582static inline const struct vc4_pv_data *
583vc4_crtc_to_vc4_pv_data(const struct vc4_crtc *crtc)
584{
585 const struct vc4_crtc_data *data = vc4_crtc_to_vc4_crtc_data(crtc);
586
587 return container_of_const(data, struct vc4_pv_data, base);
588}
589
590struct drm_encoder *vc4_get_crtc_encoder(struct drm_crtc *crtc,
591 struct drm_crtc_state *state);
592
593struct vc4_crtc_state {
594 struct drm_crtc_state base;
595 /* Dlist area for this CRTC configuration. */
596 struct drm_mm_node mm;
597 bool txp_armed;
598 unsigned int assigned_channel;
599
600 struct {
601 unsigned int left;
602 unsigned int right;
603 unsigned int top;
604 unsigned int bottom;
605 } margins;
606
607 unsigned long hvs_load;
608
609 /* Transitional state below, only valid during atomic commits */
610 bool update_muxing;
611};
612
613#define VC4_HVS_CHANNEL_DISABLED ((unsigned int)-1)
614
615#define to_vc4_crtc_state(_state) \
616 container_of_const(_state, struct vc4_crtc_state, base)
617
618#define V3D_READ(offset) \
619 ({ \
620 kunit_fail_current_test("Accessing a register in a unit test!\n"); \
621 readl(vc4->v3d->regs + (offset)); \
622 })
623
624#define V3D_WRITE(offset, val) \
625 do { \
626 kunit_fail_current_test("Accessing a register in a unit test!\n"); \
627 writel(val, vc4->v3d->regs + (offset)); \
628 } while (0)
629
630#define HVS_READ(offset) \
631 ({ \
632 kunit_fail_current_test("Accessing a register in a unit test!\n"); \
633 readl(hvs->regs + (offset)); \
634 })
635
636#define HVS_WRITE(offset, val) \
637 do { \
638 kunit_fail_current_test("Accessing a register in a unit test!\n"); \
639 writel(val, hvs->regs + (offset)); \
640 } while (0)
641
642#define VC4_REG32(reg) { .name = #reg, .offset = reg }
643
644struct vc4_exec_info {
645 struct vc4_dev *dev;
646
647 /* Sequence number for this bin/render job. */
648 uint64_t seqno;
649
650 /* Latest write_seqno of any BO that binning depends on. */
651 uint64_t bin_dep_seqno;
652
653 struct dma_fence *fence;
654
655 /* Last current addresses the hardware was processing when the
656 * hangcheck timer checked on us.
657 */
658 uint32_t last_ct0ca, last_ct1ca;
659
660 /* Kernel-space copy of the ioctl arguments */
661 struct drm_vc4_submit_cl *args;
662
663 /* This is the array of BOs that were looked up at the start of exec.
664 * Command validation will use indices into this array.
665 */
666 struct drm_gem_object **bo;
667 uint32_t bo_count;
668
669 /* List of BOs that are being written by the RCL. Other than
670 * the binner temporary storage, this is all the BOs written
671 * by the job.
672 */
673 struct drm_gem_dma_object *rcl_write_bo[4];
674 uint32_t rcl_write_bo_count;
675
676 /* Pointers for our position in vc4->job_list */
677 struct list_head head;
678
679 /* List of other BOs used in the job that need to be released
680 * once the job is complete.
681 */
682 struct list_head unref_list;
683
684 /* Current unvalidated indices into @bo loaded by the non-hardware
685 * VC4_PACKET_GEM_HANDLES.
686 */
687 uint32_t bo_index[2];
688
689 /* This is the BO where we store the validated command lists, shader
690 * records, and uniforms.
691 */
692 struct drm_gem_dma_object *exec_bo;
693
694 /**
695 * This tracks the per-shader-record state (packet 64) that
696 * determines the length of the shader record and the offset
697 * it's expected to be found at. It gets read in from the
698 * command lists.
699 */
700 struct vc4_shader_state {
701 uint32_t addr;
702 /* Maximum vertex index referenced by any primitive using this
703 * shader state.
704 */
705 uint32_t max_index;
706 } *shader_state;
707
708 /** How many shader states the user declared they were using. */
709 uint32_t shader_state_size;
710 /** How many shader state records the validator has seen. */
711 uint32_t shader_state_count;
712
713 bool found_tile_binning_mode_config_packet;
714 bool found_start_tile_binning_packet;
715 bool found_increment_semaphore_packet;
716 bool found_flush;
717 uint8_t bin_tiles_x, bin_tiles_y;
718 /* Physical address of the start of the tile alloc array
719 * (where each tile's binned CL will start)
720 */
721 uint32_t tile_alloc_offset;
722 /* Bitmask of which binner slots are freed when this job completes. */
723 uint32_t bin_slots;
724
725 /**
726 * Computed addresses pointing into exec_bo where we start the
727 * bin thread (ct0) and render thread (ct1).
728 */
729 uint32_t ct0ca, ct0ea;
730 uint32_t ct1ca, ct1ea;
731
732 /* Pointer to the unvalidated bin CL (if present). */
733 void *bin_u;
734
735 /* Pointers to the shader recs. These paddr gets incremented as CL
736 * packets are relocated in validate_gl_shader_state, and the vaddrs
737 * (u and v) get incremented and size decremented as the shader recs
738 * themselves are validated.
739 */
740 void *shader_rec_u;
741 void *shader_rec_v;
742 uint32_t shader_rec_p;
743 uint32_t shader_rec_size;
744
745 /* Pointers to the uniform data. These pointers are incremented, and
746 * size decremented, as each batch of uniforms is uploaded.
747 */
748 void *uniforms_u;
749 void *uniforms_v;
750 uint32_t uniforms_p;
751 uint32_t uniforms_size;
752
753 /* Pointer to a performance monitor object if the user requested it,
754 * NULL otherwise.
755 */
756 struct vc4_perfmon *perfmon;
757
758 /* Whether the exec has taken a reference to the binner BO, which should
759 * happen with a VC4_PACKET_TILE_BINNING_MODE_CONFIG packet.
760 */
761 bool bin_bo_used;
762};
763
764/* Per-open file private data. Any driver-specific resource that has to be
765 * released when the DRM file is closed should be placed here.
766 */
767struct vc4_file {
768 struct vc4_dev *dev;
769
770 struct {
771 struct idr idr;
772 struct mutex lock;
773 } perfmon;
774
775 bool bin_bo_used;
776};
777
778static inline struct vc4_exec_info *
779vc4_first_bin_job(struct vc4_dev *vc4)
780{
781 return list_first_entry_or_null(&vc4->bin_job_list,
782 struct vc4_exec_info, head);
783}
784
785static inline struct vc4_exec_info *
786vc4_first_render_job(struct vc4_dev *vc4)
787{
788 return list_first_entry_or_null(&vc4->render_job_list,
789 struct vc4_exec_info, head);
790}
791
792static inline struct vc4_exec_info *
793vc4_last_render_job(struct vc4_dev *vc4)
794{
795 if (list_empty(&vc4->render_job_list))
796 return NULL;
797 return list_last_entry(&vc4->render_job_list,
798 struct vc4_exec_info, head);
799}
800
801/**
802 * struct vc4_texture_sample_info - saves the offsets into the UBO for texture
803 * setup parameters.
804 *
805 * This will be used at draw time to relocate the reference to the texture
806 * contents in p0, and validate that the offset combined with
807 * width/height/stride/etc. from p1 and p2/p3 doesn't sample outside the BO.
808 * Note that the hardware treats unprovided config parameters as 0, so not all
809 * of them need to be set up for every texure sample, and we'll store ~0 as
810 * the offset to mark the unused ones.
811 *
812 * See the VC4 3D architecture guide page 41 ("Texture and Memory Lookup Unit
813 * Setup") for definitions of the texture parameters.
814 */
815struct vc4_texture_sample_info {
816 bool is_direct;
817 uint32_t p_offset[4];
818};
819
820/**
821 * struct vc4_validated_shader_info - information about validated shaders that
822 * needs to be used from command list validation.
823 *
824 * For a given shader, each time a shader state record references it, we need
825 * to verify that the shader doesn't read more uniforms than the shader state
826 * record's uniform BO pointer can provide, and we need to apply relocations
827 * and validate the shader state record's uniforms that define the texture
828 * samples.
829 */
830struct vc4_validated_shader_info {
831 uint32_t uniforms_size;
832 uint32_t uniforms_src_size;
833 uint32_t num_texture_samples;
834 struct vc4_texture_sample_info *texture_samples;
835
836 uint32_t num_uniform_addr_offsets;
837 uint32_t *uniform_addr_offsets;
838
839 bool is_threaded;
840};
841
842/**
843 * __wait_for - magic wait macro
844 *
845 * Macro to help avoid open coding check/wait/timeout patterns. Note that it's
846 * important that we check the condition again after having timed out, since the
847 * timeout could be due to preemption or similar and we've never had a chance to
848 * check the condition before the timeout.
849 */
850#define __wait_for(OP, COND, US, Wmin, Wmax) ({ \
851 const ktime_t end__ = ktime_add_ns(ktime_get_raw(), 1000ll * (US)); \
852 long wait__ = (Wmin); /* recommended min for usleep is 10 us */ \
853 int ret__; \
854 might_sleep(); \
855 for (;;) { \
856 const bool expired__ = ktime_after(ktime_get_raw(), end__); \
857 OP; \
858 /* Guarantee COND check prior to timeout */ \
859 barrier(); \
860 if (COND) { \
861 ret__ = 0; \
862 break; \
863 } \
864 if (expired__) { \
865 ret__ = -ETIMEDOUT; \
866 break; \
867 } \
868 usleep_range(wait__, wait__ * 2); \
869 if (wait__ < (Wmax)) \
870 wait__ <<= 1; \
871 } \
872 ret__; \
873})
874
875#define _wait_for(COND, US, Wmin, Wmax) __wait_for(, (COND), (US), (Wmin), \
876 (Wmax))
877#define wait_for(COND, MS) _wait_for((COND), (MS) * 1000, 10, 1000)
878
879/* vc4_bo.c */
880struct drm_gem_object *vc4_create_object(struct drm_device *dev, size_t size);
881struct vc4_bo *vc4_bo_create(struct drm_device *dev, size_t size,
882 bool from_cache, enum vc4_kernel_bo_type type);
883int vc4_bo_dumb_create(struct drm_file *file_priv,
884 struct drm_device *dev,
885 struct drm_mode_create_dumb *args);
886int vc4_create_bo_ioctl(struct drm_device *dev, void *data,
887 struct drm_file *file_priv);
888int vc4_create_shader_bo_ioctl(struct drm_device *dev, void *data,
889 struct drm_file *file_priv);
890int vc4_mmap_bo_ioctl(struct drm_device *dev, void *data,
891 struct drm_file *file_priv);
892int vc4_set_tiling_ioctl(struct drm_device *dev, void *data,
893 struct drm_file *file_priv);
894int vc4_get_tiling_ioctl(struct drm_device *dev, void *data,
895 struct drm_file *file_priv);
896int vc4_get_hang_state_ioctl(struct drm_device *dev, void *data,
897 struct drm_file *file_priv);
898int vc4_label_bo_ioctl(struct drm_device *dev, void *data,
899 struct drm_file *file_priv);
900int vc4_bo_cache_init(struct drm_device *dev);
901int vc4_bo_inc_usecnt(struct vc4_bo *bo);
902void vc4_bo_dec_usecnt(struct vc4_bo *bo);
903void vc4_bo_add_to_purgeable_pool(struct vc4_bo *bo);
904void vc4_bo_remove_from_purgeable_pool(struct vc4_bo *bo);
905int vc4_bo_debugfs_init(struct drm_minor *minor);
906
907/* vc4_crtc.c */
908extern struct platform_driver vc4_crtc_driver;
909int vc4_crtc_disable_at_boot(struct drm_crtc *crtc);
910int __vc4_crtc_init(struct drm_device *drm, struct platform_device *pdev,
911 struct vc4_crtc *vc4_crtc, const struct vc4_crtc_data *data,
912 struct drm_plane *primary_plane,
913 const struct drm_crtc_funcs *crtc_funcs,
914 const struct drm_crtc_helper_funcs *crtc_helper_funcs,
915 bool feeds_txp);
916int vc4_crtc_init(struct drm_device *drm, struct platform_device *pdev,
917 struct vc4_crtc *vc4_crtc, const struct vc4_crtc_data *data,
918 const struct drm_crtc_funcs *crtc_funcs,
919 const struct drm_crtc_helper_funcs *crtc_helper_funcs,
920 bool feeds_txp);
921int vc4_page_flip(struct drm_crtc *crtc,
922 struct drm_framebuffer *fb,
923 struct drm_pending_vblank_event *event,
924 uint32_t flags,
925 struct drm_modeset_acquire_ctx *ctx);
926int vc4_crtc_atomic_check(struct drm_crtc *crtc,
927 struct drm_atomic_state *state);
928struct drm_crtc_state *vc4_crtc_duplicate_state(struct drm_crtc *crtc);
929void vc4_crtc_destroy_state(struct drm_crtc *crtc,
930 struct drm_crtc_state *state);
931void vc4_crtc_reset(struct drm_crtc *crtc);
932void vc4_crtc_handle_vblank(struct vc4_crtc *crtc);
933void vc4_crtc_send_vblank(struct drm_crtc *crtc);
934int vc4_crtc_late_register(struct drm_crtc *crtc);
935void vc4_crtc_get_margins(struct drm_crtc_state *state,
936 unsigned int *left, unsigned int *right,
937 unsigned int *top, unsigned int *bottom);
938
939/* vc4_debugfs.c */
940void vc4_debugfs_init(struct drm_minor *minor);
941#ifdef CONFIG_DEBUG_FS
942void vc4_debugfs_add_regset32(struct drm_device *drm,
943 const char *filename,
944 struct debugfs_regset32 *regset);
945#else
946
947static inline void vc4_debugfs_add_regset32(struct drm_device *drm,
948 const char *filename,
949 struct debugfs_regset32 *regset)
950{}
951#endif
952
953/* vc4_drv.c */
954void __iomem *vc4_ioremap_regs(struct platform_device *dev, int index);
955int vc4_dumb_fixup_args(struct drm_mode_create_dumb *args);
956
957/* vc4_dpi.c */
958extern struct platform_driver vc4_dpi_driver;
959
960/* vc4_dsi.c */
961extern struct platform_driver vc4_dsi_driver;
962
963/* vc4_fence.c */
964extern const struct dma_fence_ops vc4_fence_ops;
965
966/* vc4_gem.c */
967int vc4_gem_init(struct drm_device *dev);
968int vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
969 struct drm_file *file_priv);
970int vc4_wait_seqno_ioctl(struct drm_device *dev, void *data,
971 struct drm_file *file_priv);
972int vc4_wait_bo_ioctl(struct drm_device *dev, void *data,
973 struct drm_file *file_priv);
974void vc4_submit_next_bin_job(struct drm_device *dev);
975void vc4_submit_next_render_job(struct drm_device *dev);
976void vc4_move_job_to_render(struct drm_device *dev, struct vc4_exec_info *exec);
977int vc4_wait_for_seqno(struct drm_device *dev, uint64_t seqno,
978 uint64_t timeout_ns, bool interruptible);
979void vc4_job_handle_completed(struct vc4_dev *vc4);
980int vc4_queue_seqno_cb(struct drm_device *dev,
981 struct vc4_seqno_cb *cb, uint64_t seqno,
982 void (*func)(struct vc4_seqno_cb *cb));
983int vc4_gem_madvise_ioctl(struct drm_device *dev, void *data,
984 struct drm_file *file_priv);
985
986/* vc4_hdmi.c */
987extern struct platform_driver vc4_hdmi_driver;
988
989/* vc4_vec.c */
990extern struct platform_driver vc4_vec_driver;
991
992/* vc4_txp.c */
993extern struct platform_driver vc4_txp_driver;
994
995/* vc4_irq.c */
996void vc4_irq_enable(struct drm_device *dev);
997void vc4_irq_disable(struct drm_device *dev);
998int vc4_irq_install(struct drm_device *dev, int irq);
999void vc4_irq_uninstall(struct drm_device *dev);
1000void vc4_irq_reset(struct drm_device *dev);
1001
1002/* vc4_hvs.c */
1003extern struct platform_driver vc4_hvs_driver;
1004struct vc4_hvs *__vc4_hvs_alloc(struct vc4_dev *vc4, struct platform_device *pdev);
1005void vc4_hvs_stop_channel(struct vc4_hvs *hvs, unsigned int output);
1006int vc4_hvs_get_fifo_from_output(struct vc4_hvs *hvs, unsigned int output);
1007u8 vc4_hvs_get_fifo_frame_count(struct vc4_hvs *hvs, unsigned int fifo);
1008int vc4_hvs_atomic_check(struct drm_crtc *crtc, struct drm_atomic_state *state);
1009void vc4_hvs_atomic_begin(struct drm_crtc *crtc, struct drm_atomic_state *state);
1010void vc4_hvs_atomic_enable(struct drm_crtc *crtc, struct drm_atomic_state *state);
1011void vc4_hvs_atomic_disable(struct drm_crtc *crtc, struct drm_atomic_state *state);
1012void vc4_hvs_atomic_flush(struct drm_crtc *crtc, struct drm_atomic_state *state);
1013void vc4_hvs_dump_state(struct vc4_hvs *hvs);
1014void vc4_hvs_unmask_underrun(struct vc4_hvs *hvs, int channel);
1015void vc4_hvs_mask_underrun(struct vc4_hvs *hvs, int channel);
1016int vc4_hvs_debugfs_init(struct drm_minor *minor);
1017
1018/* vc4_kms.c */
1019int vc4_kms_load(struct drm_device *dev);
1020
1021/* vc4_plane.c */
1022struct drm_plane *vc4_plane_init(struct drm_device *dev,
1023 enum drm_plane_type type,
1024 uint32_t possible_crtcs);
1025int vc4_plane_create_additional_planes(struct drm_device *dev);
1026u32 vc4_plane_write_dlist(struct drm_plane *plane, u32 __iomem *dlist);
1027u32 vc4_plane_dlist_size(const struct drm_plane_state *state);
1028void vc4_plane_async_set_fb(struct drm_plane *plane,
1029 struct drm_framebuffer *fb);
1030
1031/* vc4_v3d.c */
1032extern struct platform_driver vc4_v3d_driver;
1033extern const struct of_device_id vc4_v3d_dt_match[];
1034int vc4_v3d_get_bin_slot(struct vc4_dev *vc4);
1035int vc4_v3d_bin_bo_get(struct vc4_dev *vc4, bool *used);
1036void vc4_v3d_bin_bo_put(struct vc4_dev *vc4);
1037int vc4_v3d_pm_get(struct vc4_dev *vc4);
1038void vc4_v3d_pm_put(struct vc4_dev *vc4);
1039int vc4_v3d_debugfs_init(struct drm_minor *minor);
1040
1041/* vc4_validate.c */
1042int
1043vc4_validate_bin_cl(struct drm_device *dev,
1044 void *validated,
1045 void *unvalidated,
1046 struct vc4_exec_info *exec);
1047
1048int
1049vc4_validate_shader_recs(struct drm_device *dev, struct vc4_exec_info *exec);
1050
1051struct drm_gem_dma_object *vc4_use_bo(struct vc4_exec_info *exec,
1052 uint32_t hindex);
1053
1054int vc4_get_rcl(struct drm_device *dev, struct vc4_exec_info *exec);
1055
1056bool vc4_check_tex_size(struct vc4_exec_info *exec,
1057 struct drm_gem_dma_object *fbo,
1058 uint32_t offset, uint8_t tiling_format,
1059 uint32_t width, uint32_t height, uint8_t cpp);
1060
1061/* vc4_validate_shader.c */
1062struct vc4_validated_shader_info *
1063vc4_validate_shader(struct drm_gem_dma_object *shader_obj);
1064
1065/* vc4_perfmon.c */
1066void vc4_perfmon_get(struct vc4_perfmon *perfmon);
1067void vc4_perfmon_put(struct vc4_perfmon *perfmon);
1068void vc4_perfmon_start(struct vc4_dev *vc4, struct vc4_perfmon *perfmon);
1069void vc4_perfmon_stop(struct vc4_dev *vc4, struct vc4_perfmon *perfmon,
1070 bool capture);
1071struct vc4_perfmon *vc4_perfmon_find(struct vc4_file *vc4file, int id);
1072void vc4_perfmon_open_file(struct vc4_file *vc4file);
1073void vc4_perfmon_close_file(struct vc4_file *vc4file);
1074int vc4_perfmon_create_ioctl(struct drm_device *dev, void *data,
1075 struct drm_file *file_priv);
1076int vc4_perfmon_destroy_ioctl(struct drm_device *dev, void *data,
1077 struct drm_file *file_priv);
1078int vc4_perfmon_get_values_ioctl(struct drm_device *dev, void *data,
1079 struct drm_file *file_priv);
1080
1081#endif /* _VC4_DRV_H_ */