Linux Audio

Check our new training course

Loading...
v5.4
  1/* SPDX-License-Identifier: GPL-2.0-only */
  2/*
  3 * Copyright (C) 2015 Broadcom
  4 */
 
 
  5
  6#include <linux/delay.h>
  7#include <linux/refcount.h>
  8#include <linux/uaccess.h>
  9
 10#include <drm/drm_atomic.h>
 11#include <drm/drm_debugfs.h>
 12#include <drm/drm_device.h>
 13#include <drm/drm_encoder.h>
 14#include <drm/drm_gem_cma_helper.h>
 15#include <drm/drm_mm.h>
 16#include <drm/drm_modeset_lock.h>
 17
 18#include "uapi/drm/vc4_drm.h"
 19
 20struct drm_device;
 21struct drm_gem_object;
 22
 23/* Don't forget to update vc4_bo.c: bo_type_names[] when adding to
 24 * this.
 25 */
 26enum vc4_kernel_bo_type {
 27	/* Any kernel allocation (gem_create_object hook) before it
 28	 * gets another type set.
 29	 */
 30	VC4_BO_TYPE_KERNEL,
 31	VC4_BO_TYPE_V3D,
 32	VC4_BO_TYPE_V3D_SHADER,
 33	VC4_BO_TYPE_DUMB,
 34	VC4_BO_TYPE_BIN,
 35	VC4_BO_TYPE_RCL,
 36	VC4_BO_TYPE_BCL,
 37	VC4_BO_TYPE_KERNEL_CACHE,
 38	VC4_BO_TYPE_COUNT
 39};
 40
 41/* Performance monitor object. The perform lifetime is controlled by userspace
 42 * using perfmon related ioctls. A perfmon can be attached to a submit_cl
 43 * request, and when this is the case, HW perf counters will be activated just
 44 * before the submit_cl is submitted to the GPU and disabled when the job is
 45 * done. This way, only events related to a specific job will be counted.
 46 */
 47struct vc4_perfmon {
 48	/* Tracks the number of users of the perfmon, when this counter reaches
 49	 * zero the perfmon is destroyed.
 50	 */
 51	refcount_t refcnt;
 52
 53	/* Number of counters activated in this perfmon instance
 54	 * (should be less than DRM_VC4_MAX_PERF_COUNTERS).
 55	 */
 56	u8 ncounters;
 57
 58	/* Events counted by the HW perf counters. */
 59	u8 events[DRM_VC4_MAX_PERF_COUNTERS];
 60
 61	/* Storage for counter values. Counters are incremented by the HW
 62	 * perf counter values every time the perfmon is attached to a GPU job.
 63	 * This way, perfmon users don't have to retrieve the results after
 64	 * each job if they want to track events covering several submissions.
 65	 * Note that counter values can't be reset, but you can fake a reset by
 66	 * destroying the perfmon and creating a new one.
 67	 */
 68	u64 counters[0];
 69};
 70
 71struct vc4_dev {
 72	struct drm_device *dev;
 73
 74	struct vc4_hdmi *hdmi;
 75	struct vc4_hvs *hvs;
 76	struct vc4_v3d *v3d;
 77	struct vc4_dpi *dpi;
 78	struct vc4_dsi *dsi1;
 79	struct vc4_vec *vec;
 80	struct vc4_txp *txp;
 81
 82	struct vc4_hang_state *hang_state;
 83
 84	/* The kernel-space BO cache.  Tracks buffers that have been
 85	 * unreferenced by all other users (refcounts of 0!) but not
 86	 * yet freed, so we can do cheap allocations.
 87	 */
 88	struct vc4_bo_cache {
 89		/* Array of list heads for entries in the BO cache,
 90		 * based on number of pages, so we can do O(1) lookups
 91		 * in the cache when allocating.
 92		 */
 93		struct list_head *size_list;
 94		uint32_t size_list_size;
 95
 96		/* List of all BOs in the cache, ordered by age, so we
 97		 * can do O(1) lookups when trying to free old
 98		 * buffers.
 99		 */
100		struct list_head time_list;
101		struct work_struct time_work;
102		struct timer_list time_timer;
103	} bo_cache;
104
105	u32 num_labels;
106	struct vc4_label {
107		const char *name;
108		u32 num_allocated;
109		u32 size_allocated;
110	} *bo_labels;
111
112	/* Protects bo_cache and bo_labels. */
113	struct mutex bo_lock;
114
115	/* Purgeable BO pool. All BOs in this pool can have their memory
116	 * reclaimed if the driver is unable to allocate new BOs. We also
117	 * keep stats related to the purge mechanism here.
118	 */
119	struct {
120		struct list_head list;
121		unsigned int num;
122		size_t size;
123		unsigned int purged_num;
124		size_t purged_size;
125		struct mutex lock;
126	} purgeable;
127
128	uint64_t dma_fence_context;
129
130	/* Sequence number for the last job queued in bin_job_list.
131	 * Starts at 0 (no jobs emitted).
132	 */
133	uint64_t emit_seqno;
134
135	/* Sequence number for the last completed job on the GPU.
136	 * Starts at 0 (no jobs completed).
137	 */
138	uint64_t finished_seqno;
139
140	/* List of all struct vc4_exec_info for jobs to be executed in
141	 * the binner.  The first job in the list is the one currently
142	 * programmed into ct0ca for execution.
143	 */
144	struct list_head bin_job_list;
145
146	/* List of all struct vc4_exec_info for jobs that have
147	 * completed binning and are ready for rendering.  The first
148	 * job in the list is the one currently programmed into ct1ca
149	 * for execution.
150	 */
151	struct list_head render_job_list;
152
153	/* List of the finished vc4_exec_infos waiting to be freed by
154	 * job_done_work.
155	 */
156	struct list_head job_done_list;
157	/* Spinlock used to synchronize the job_list and seqno
158	 * accesses between the IRQ handler and GEM ioctls.
159	 */
160	spinlock_t job_lock;
161	wait_queue_head_t job_wait_queue;
162	struct work_struct job_done_work;
163
164	/* Used to track the active perfmon if any. Access to this field is
165	 * protected by job_lock.
166	 */
167	struct vc4_perfmon *active_perfmon;
168
169	/* List of struct vc4_seqno_cb for callbacks to be made from a
170	 * workqueue when the given seqno is passed.
171	 */
172	struct list_head seqno_cb_list;
173
174	/* The memory used for storing binner tile alloc, tile state,
175	 * and overflow memory allocations.  This is freed when V3D
176	 * powers down.
177	 */
178	struct vc4_bo *bin_bo;
179
180	/* Size of blocks allocated within bin_bo. */
181	uint32_t bin_alloc_size;
182
183	/* Bitmask of the bin_alloc_size chunks in bin_bo that are
184	 * used.
185	 */
186	uint32_t bin_alloc_used;
187
188	/* Bitmask of the current bin_alloc used for overflow memory. */
189	uint32_t bin_alloc_overflow;
190
191	/* Incremented when an underrun error happened after an atomic commit.
192	 * This is particularly useful to detect when a specific modeset is too
193	 * demanding in term of memory or HVS bandwidth which is hard to guess
194	 * at atomic check time.
195	 */
196	atomic_t underrun;
197
198	struct work_struct overflow_mem_work;
199
200	int power_refcount;
201
202	/* Set to true when the load tracker is active. */
203	bool load_tracker_enabled;
204
205	/* Mutex controlling the power refcount. */
206	struct mutex power_lock;
207
208	struct {
209		struct timer_list timer;
210		struct work_struct reset_work;
211	} hangcheck;
212
213	struct semaphore async_modeset;
214
215	struct drm_modeset_lock ctm_state_lock;
216	struct drm_private_obj ctm_manager;
217	struct drm_private_obj load_tracker;
218
219	/* List of vc4_debugfs_info_entry for adding to debugfs once
220	 * the minor is available (after drm_dev_register()).
221	 */
222	struct list_head debugfs_list;
223
224	/* Mutex for binner bo allocation. */
225	struct mutex bin_bo_lock;
226	/* Reference count for our binner bo. */
227	struct kref bin_bo_kref;
228};
229
230static inline struct vc4_dev *
231to_vc4_dev(struct drm_device *dev)
232{
233	return (struct vc4_dev *)dev->dev_private;
234}
235
236struct vc4_bo {
237	struct drm_gem_cma_object base;
238
239	/* seqno of the last job to render using this BO. */
240	uint64_t seqno;
241
242	/* seqno of the last job to use the RCL to write to this BO.
243	 *
244	 * Note that this doesn't include binner overflow memory
245	 * writes.
246	 */
247	uint64_t write_seqno;
248
249	bool t_format;
250
251	/* List entry for the BO's position in either
252	 * vc4_exec_info->unref_list or vc4_dev->bo_cache.time_list
253	 */
254	struct list_head unref_head;
255
256	/* Time in jiffies when the BO was put in vc4->bo_cache. */
257	unsigned long free_time;
258
259	/* List entry for the BO's position in vc4_dev->bo_cache.size_list */
260	struct list_head size_head;
261
262	/* Struct for shader validation state, if created by
263	 * DRM_IOCTL_VC4_CREATE_SHADER_BO.
264	 */
265	struct vc4_validated_shader_info *validated_shader;
266
267	/* One of enum vc4_kernel_bo_type, or VC4_BO_TYPE_COUNT + i
268	 * for user-allocated labels.
269	 */
270	int label;
271
272	/* Count the number of active users. This is needed to determine
273	 * whether we can move the BO to the purgeable list or not (when the BO
274	 * is used by the GPU or the display engine we can't purge it).
275	 */
276	refcount_t usecnt;
277
278	/* Store purgeable/purged state here */
279	u32 madv;
280	struct mutex madv_lock;
281};
282
283static inline struct vc4_bo *
284to_vc4_bo(struct drm_gem_object *bo)
285{
286	return (struct vc4_bo *)bo;
287}
288
289struct vc4_fence {
290	struct dma_fence base;
291	struct drm_device *dev;
292	/* vc4 seqno for signaled() test */
293	uint64_t seqno;
294};
295
296static inline struct vc4_fence *
297to_vc4_fence(struct dma_fence *fence)
298{
299	return (struct vc4_fence *)fence;
300}
301
302struct vc4_seqno_cb {
303	struct work_struct work;
304	uint64_t seqno;
305	void (*func)(struct vc4_seqno_cb *cb);
306};
307
308struct vc4_v3d {
309	struct vc4_dev *vc4;
310	struct platform_device *pdev;
311	void __iomem *regs;
312	struct clk *clk;
313	struct debugfs_regset32 regset;
314};
315
316struct vc4_hvs {
317	struct platform_device *pdev;
318	void __iomem *regs;
319	u32 __iomem *dlist;
320
321	/* Memory manager for CRTCs to allocate space in the display
322	 * list.  Units are dwords.
323	 */
324	struct drm_mm dlist_mm;
325	/* Memory manager for the LBM memory used by HVS scaling. */
326	struct drm_mm lbm_mm;
327	spinlock_t mm_lock;
328
329	struct drm_mm_node mitchell_netravali_filter;
330	struct debugfs_regset32 regset;
331};
332
333struct vc4_plane {
334	struct drm_plane base;
335};
336
337static inline struct vc4_plane *
338to_vc4_plane(struct drm_plane *plane)
339{
340	return (struct vc4_plane *)plane;
341}
342
343enum vc4_scaling_mode {
344	VC4_SCALING_NONE,
345	VC4_SCALING_TPZ,
346	VC4_SCALING_PPF,
347};
348
349struct vc4_plane_state {
350	struct drm_plane_state base;
351	/* System memory copy of the display list for this element, computed
352	 * at atomic_check time.
353	 */
354	u32 *dlist;
355	u32 dlist_size; /* Number of dwords allocated for the display list */
356	u32 dlist_count; /* Number of used dwords in the display list. */
357
358	/* Offset in the dlist to various words, for pageflip or
359	 * cursor updates.
360	 */
361	u32 pos0_offset;
362	u32 pos2_offset;
363	u32 ptr0_offset;
364	u32 lbm_offset;
365
366	/* Offset where the plane's dlist was last stored in the
367	 * hardware at vc4_crtc_atomic_flush() time.
368	 */
369	u32 __iomem *hw_dlist;
370
371	/* Clipped coordinates of the plane on the display. */
372	int crtc_x, crtc_y, crtc_w, crtc_h;
373	/* Clipped area being scanned from in the FB. */
374	u32 src_x, src_y;
375
376	u32 src_w[2], src_h[2];
377
378	/* Scaling selection for the RGB/Y plane and the Cb/Cr planes. */
379	enum vc4_scaling_mode x_scaling[2], y_scaling[2];
380	bool is_unity;
381	bool is_yuv;
382
383	/* Offset to start scanning out from the start of the plane's
384	 * BO.
385	 */
386	u32 offsets[3];
387
388	/* Our allocation in LBM for temporary storage during scaling. */
389	struct drm_mm_node lbm;
390
391	/* Set when the plane has per-pixel alpha content or does not cover
392	 * the entire screen. This is a hint to the CRTC that it might need
393	 * to enable background color fill.
394	 */
395	bool needs_bg_fill;
396
397	/* Mark the dlist as initialized. Useful to avoid initializing it twice
398	 * when async update is not possible.
399	 */
400	bool dlist_initialized;
401
402	/* Load of this plane on the HVS block. The load is expressed in HVS
403	 * cycles/sec.
404	 */
405	u64 hvs_load;
406
407	/* Memory bandwidth needed for this plane. This is expressed in
408	 * bytes/sec.
409	 */
410	u64 membus_load;
411};
412
413static inline struct vc4_plane_state *
414to_vc4_plane_state(struct drm_plane_state *state)
415{
416	return (struct vc4_plane_state *)state;
417}
418
419enum vc4_encoder_type {
420	VC4_ENCODER_TYPE_NONE,
421	VC4_ENCODER_TYPE_HDMI,
422	VC4_ENCODER_TYPE_VEC,
423	VC4_ENCODER_TYPE_DSI0,
424	VC4_ENCODER_TYPE_DSI1,
425	VC4_ENCODER_TYPE_SMI,
426	VC4_ENCODER_TYPE_DPI,
427};
428
429struct vc4_encoder {
430	struct drm_encoder base;
431	enum vc4_encoder_type type;
432	u32 clock_select;
433};
434
435static inline struct vc4_encoder *
436to_vc4_encoder(struct drm_encoder *encoder)
437{
438	return container_of(encoder, struct vc4_encoder, base);
439}
440
441struct vc4_crtc_data {
442	/* Which channel of the HVS this pixelvalve sources from. */
443	int hvs_channel;
 
 
 
 
444
445	enum vc4_encoder_type encoder_types[4];
446	const char *debugfs_name;
 
447};
448
449struct vc4_crtc {
450	struct drm_crtc base;
451	struct platform_device *pdev;
452	const struct vc4_crtc_data *data;
453	void __iomem *regs;
454
455	/* Timestamp at start of vblank irq - unaffected by lock delays. */
456	ktime_t t_vblank;
457
458	/* Which HVS channel we're using for our CRTC. */
459	int channel;
460
461	u8 lut_r[256];
462	u8 lut_g[256];
463	u8 lut_b[256];
464	/* Size in pixels of the COB memory allocated to this CRTC. */
465	u32 cob_size;
466
467	struct drm_pending_vblank_event *event;
468
469	struct debugfs_regset32 regset;
470};
471
472static inline struct vc4_crtc *
473to_vc4_crtc(struct drm_crtc *crtc)
474{
475	return (struct vc4_crtc *)crtc;
476}
477
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
478#define V3D_READ(offset) readl(vc4->v3d->regs + offset)
479#define V3D_WRITE(offset, val) writel(val, vc4->v3d->regs + offset)
480#define HVS_READ(offset) readl(vc4->hvs->regs + offset)
481#define HVS_WRITE(offset, val) writel(val, vc4->hvs->regs + offset)
482
483#define VC4_REG32(reg) { .name = #reg, .offset = reg }
484
485struct vc4_exec_info {
486	/* Sequence number for this bin/render job. */
487	uint64_t seqno;
488
489	/* Latest write_seqno of any BO that binning depends on. */
490	uint64_t bin_dep_seqno;
491
492	struct dma_fence *fence;
493
494	/* Last current addresses the hardware was processing when the
495	 * hangcheck timer checked on us.
496	 */
497	uint32_t last_ct0ca, last_ct1ca;
498
499	/* Kernel-space copy of the ioctl arguments */
500	struct drm_vc4_submit_cl *args;
501
502	/* This is the array of BOs that were looked up at the start of exec.
503	 * Command validation will use indices into this array.
504	 */
505	struct drm_gem_cma_object **bo;
506	uint32_t bo_count;
507
508	/* List of BOs that are being written by the RCL.  Other than
509	 * the binner temporary storage, this is all the BOs written
510	 * by the job.
511	 */
512	struct drm_gem_cma_object *rcl_write_bo[4];
513	uint32_t rcl_write_bo_count;
514
515	/* Pointers for our position in vc4->job_list */
516	struct list_head head;
517
518	/* List of other BOs used in the job that need to be released
519	 * once the job is complete.
520	 */
521	struct list_head unref_list;
522
523	/* Current unvalidated indices into @bo loaded by the non-hardware
524	 * VC4_PACKET_GEM_HANDLES.
525	 */
526	uint32_t bo_index[2];
527
528	/* This is the BO where we store the validated command lists, shader
529	 * records, and uniforms.
530	 */
531	struct drm_gem_cma_object *exec_bo;
532
533	/**
534	 * This tracks the per-shader-record state (packet 64) that
535	 * determines the length of the shader record and the offset
536	 * it's expected to be found at.  It gets read in from the
537	 * command lists.
538	 */
539	struct vc4_shader_state {
540		uint32_t addr;
541		/* Maximum vertex index referenced by any primitive using this
542		 * shader state.
543		 */
544		uint32_t max_index;
545	} *shader_state;
546
547	/** How many shader states the user declared they were using. */
548	uint32_t shader_state_size;
549	/** How many shader state records the validator has seen. */
550	uint32_t shader_state_count;
551
552	bool found_tile_binning_mode_config_packet;
553	bool found_start_tile_binning_packet;
554	bool found_increment_semaphore_packet;
555	bool found_flush;
556	uint8_t bin_tiles_x, bin_tiles_y;
557	/* Physical address of the start of the tile alloc array
558	 * (where each tile's binned CL will start)
559	 */
560	uint32_t tile_alloc_offset;
561	/* Bitmask of which binner slots are freed when this job completes. */
562	uint32_t bin_slots;
563
564	/**
565	 * Computed addresses pointing into exec_bo where we start the
566	 * bin thread (ct0) and render thread (ct1).
567	 */
568	uint32_t ct0ca, ct0ea;
569	uint32_t ct1ca, ct1ea;
570
571	/* Pointer to the unvalidated bin CL (if present). */
572	void *bin_u;
573
574	/* Pointers to the shader recs.  These paddr gets incremented as CL
575	 * packets are relocated in validate_gl_shader_state, and the vaddrs
576	 * (u and v) get incremented and size decremented as the shader recs
577	 * themselves are validated.
578	 */
579	void *shader_rec_u;
580	void *shader_rec_v;
581	uint32_t shader_rec_p;
582	uint32_t shader_rec_size;
583
584	/* Pointers to the uniform data.  These pointers are incremented, and
585	 * size decremented, as each batch of uniforms is uploaded.
586	 */
587	void *uniforms_u;
588	void *uniforms_v;
589	uint32_t uniforms_p;
590	uint32_t uniforms_size;
591
592	/* Pointer to a performance monitor object if the user requested it,
593	 * NULL otherwise.
594	 */
595	struct vc4_perfmon *perfmon;
596
597	/* Whether the exec has taken a reference to the binner BO, which should
598	 * happen with a VC4_PACKET_TILE_BINNING_MODE_CONFIG packet.
599	 */
600	bool bin_bo_used;
601};
602
603/* Per-open file private data. Any driver-specific resource that has to be
604 * released when the DRM file is closed should be placed here.
605 */
606struct vc4_file {
607	struct {
608		struct idr idr;
609		struct mutex lock;
610	} perfmon;
611
612	bool bin_bo_used;
613};
614
615static inline struct vc4_exec_info *
616vc4_first_bin_job(struct vc4_dev *vc4)
617{
618	return list_first_entry_or_null(&vc4->bin_job_list,
619					struct vc4_exec_info, head);
620}
621
622static inline struct vc4_exec_info *
623vc4_first_render_job(struct vc4_dev *vc4)
624{
625	return list_first_entry_or_null(&vc4->render_job_list,
626					struct vc4_exec_info, head);
627}
628
629static inline struct vc4_exec_info *
630vc4_last_render_job(struct vc4_dev *vc4)
631{
632	if (list_empty(&vc4->render_job_list))
633		return NULL;
634	return list_last_entry(&vc4->render_job_list,
635			       struct vc4_exec_info, head);
636}
637
638/**
639 * struct vc4_texture_sample_info - saves the offsets into the UBO for texture
640 * setup parameters.
641 *
642 * This will be used at draw time to relocate the reference to the texture
643 * contents in p0, and validate that the offset combined with
644 * width/height/stride/etc. from p1 and p2/p3 doesn't sample outside the BO.
645 * Note that the hardware treats unprovided config parameters as 0, so not all
646 * of them need to be set up for every texure sample, and we'll store ~0 as
647 * the offset to mark the unused ones.
648 *
649 * See the VC4 3D architecture guide page 41 ("Texture and Memory Lookup Unit
650 * Setup") for definitions of the texture parameters.
651 */
652struct vc4_texture_sample_info {
653	bool is_direct;
654	uint32_t p_offset[4];
655};
656
657/**
658 * struct vc4_validated_shader_info - information about validated shaders that
659 * needs to be used from command list validation.
660 *
661 * For a given shader, each time a shader state record references it, we need
662 * to verify that the shader doesn't read more uniforms than the shader state
663 * record's uniform BO pointer can provide, and we need to apply relocations
664 * and validate the shader state record's uniforms that define the texture
665 * samples.
666 */
667struct vc4_validated_shader_info {
668	uint32_t uniforms_size;
669	uint32_t uniforms_src_size;
670	uint32_t num_texture_samples;
671	struct vc4_texture_sample_info *texture_samples;
672
673	uint32_t num_uniform_addr_offsets;
674	uint32_t *uniform_addr_offsets;
675
676	bool is_threaded;
677};
678
679/**
680 * _wait_for - magic (register) wait macro
681 *
682 * Does the right thing for modeset paths when run under kdgb or similar atomic
683 * contexts. Note that it's important that we check the condition again after
684 * having timed out, since the timeout could be due to preemption or similar and
685 * we've never had a chance to check the condition before the timeout.
686 */
687#define _wait_for(COND, MS, W) ({ \
688	unsigned long timeout__ = jiffies + msecs_to_jiffies(MS) + 1;	\
689	int ret__ = 0;							\
690	while (!(COND)) {						\
691		if (time_after(jiffies, timeout__)) {			\
692			if (!(COND))					\
693				ret__ = -ETIMEDOUT;			\
 
 
 
 
 
694			break;						\
695		}							\
696		if (W && drm_can_sleep())  {				\
697			msleep(W);					\
698		} else {						\
699			cpu_relax();					\
700		}							\
 
 
 
701	}								\
702	ret__;								\
703})
704
705#define wait_for(COND, MS) _wait_for(COND, MS, 1)
 
 
706
707/* vc4_bo.c */
708struct drm_gem_object *vc4_create_object(struct drm_device *dev, size_t size);
709void vc4_free_object(struct drm_gem_object *gem_obj);
710struct vc4_bo *vc4_bo_create(struct drm_device *dev, size_t size,
711			     bool from_cache, enum vc4_kernel_bo_type type);
712int vc4_dumb_create(struct drm_file *file_priv,
713		    struct drm_device *dev,
714		    struct drm_mode_create_dumb *args);
715struct dma_buf *vc4_prime_export(struct drm_gem_object *obj, int flags);
716int vc4_create_bo_ioctl(struct drm_device *dev, void *data,
717			struct drm_file *file_priv);
718int vc4_create_shader_bo_ioctl(struct drm_device *dev, void *data,
719			       struct drm_file *file_priv);
720int vc4_mmap_bo_ioctl(struct drm_device *dev, void *data,
721		      struct drm_file *file_priv);
722int vc4_set_tiling_ioctl(struct drm_device *dev, void *data,
723			 struct drm_file *file_priv);
724int vc4_get_tiling_ioctl(struct drm_device *dev, void *data,
725			 struct drm_file *file_priv);
726int vc4_get_hang_state_ioctl(struct drm_device *dev, void *data,
727			     struct drm_file *file_priv);
728int vc4_label_bo_ioctl(struct drm_device *dev, void *data,
729		       struct drm_file *file_priv);
730vm_fault_t vc4_fault(struct vm_fault *vmf);
731int vc4_mmap(struct file *filp, struct vm_area_struct *vma);
732int vc4_prime_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma);
733struct drm_gem_object *vc4_prime_import_sg_table(struct drm_device *dev,
734						 struct dma_buf_attachment *attach,
735						 struct sg_table *sgt);
736void *vc4_prime_vmap(struct drm_gem_object *obj);
737int vc4_bo_cache_init(struct drm_device *dev);
738void vc4_bo_cache_destroy(struct drm_device *dev);
739int vc4_bo_inc_usecnt(struct vc4_bo *bo);
740void vc4_bo_dec_usecnt(struct vc4_bo *bo);
741void vc4_bo_add_to_purgeable_pool(struct vc4_bo *bo);
742void vc4_bo_remove_from_purgeable_pool(struct vc4_bo *bo);
743
744/* vc4_crtc.c */
745extern struct platform_driver vc4_crtc_driver;
746bool vc4_crtc_get_scanoutpos(struct drm_device *dev, unsigned int crtc_id,
747			     bool in_vblank_irq, int *vpos, int *hpos,
748			     ktime_t *stime, ktime_t *etime,
749			     const struct drm_display_mode *mode);
 
 
 
 
 
 
 
 
 
750void vc4_crtc_handle_vblank(struct vc4_crtc *crtc);
751void vc4_crtc_txp_armed(struct drm_crtc_state *state);
752void vc4_crtc_get_margins(struct drm_crtc_state *state,
753			  unsigned int *right, unsigned int *left,
754			  unsigned int *top, unsigned int *bottom);
755
756/* vc4_debugfs.c */
757int vc4_debugfs_init(struct drm_minor *minor);
758#ifdef CONFIG_DEBUG_FS
759void vc4_debugfs_add_file(struct drm_device *drm,
760			  const char *filename,
761			  int (*show)(struct seq_file*, void*),
762			  void *data);
763void vc4_debugfs_add_regset32(struct drm_device *drm,
764			      const char *filename,
765			      struct debugfs_regset32 *regset);
766#else
767static inline void vc4_debugfs_add_file(struct drm_device *drm,
768					const char *filename,
769					int (*show)(struct seq_file*, void*),
770					void *data)
771{
772}
773
774static inline void vc4_debugfs_add_regset32(struct drm_device *drm,
775					    const char *filename,
776					    struct debugfs_regset32 *regset)
777{
778}
779#endif
780
781/* vc4_drv.c */
782void __iomem *vc4_ioremap_regs(struct platform_device *dev, int index);
783
784/* vc4_dpi.c */
785extern struct platform_driver vc4_dpi_driver;
786
787/* vc4_dsi.c */
788extern struct platform_driver vc4_dsi_driver;
789
790/* vc4_fence.c */
791extern const struct dma_fence_ops vc4_fence_ops;
792
793/* vc4_gem.c */
794void vc4_gem_init(struct drm_device *dev);
795void vc4_gem_destroy(struct drm_device *dev);
796int vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
797			struct drm_file *file_priv);
798int vc4_wait_seqno_ioctl(struct drm_device *dev, void *data,
799			 struct drm_file *file_priv);
800int vc4_wait_bo_ioctl(struct drm_device *dev, void *data,
801		      struct drm_file *file_priv);
802void vc4_submit_next_bin_job(struct drm_device *dev);
803void vc4_submit_next_render_job(struct drm_device *dev);
804void vc4_move_job_to_render(struct drm_device *dev, struct vc4_exec_info *exec);
805int vc4_wait_for_seqno(struct drm_device *dev, uint64_t seqno,
806		       uint64_t timeout_ns, bool interruptible);
807void vc4_job_handle_completed(struct vc4_dev *vc4);
808int vc4_queue_seqno_cb(struct drm_device *dev,
809		       struct vc4_seqno_cb *cb, uint64_t seqno,
810		       void (*func)(struct vc4_seqno_cb *cb));
811int vc4_gem_madvise_ioctl(struct drm_device *dev, void *data,
812			  struct drm_file *file_priv);
813
814/* vc4_hdmi.c */
815extern struct platform_driver vc4_hdmi_driver;
816
817/* vc4_vec.c */
818extern struct platform_driver vc4_vec_driver;
819
820/* vc4_txp.c */
821extern struct platform_driver vc4_txp_driver;
822
823/* vc4_irq.c */
824irqreturn_t vc4_irq(int irq, void *arg);
825void vc4_irq_preinstall(struct drm_device *dev);
826int vc4_irq_postinstall(struct drm_device *dev);
827void vc4_irq_uninstall(struct drm_device *dev);
828void vc4_irq_reset(struct drm_device *dev);
829
830/* vc4_hvs.c */
831extern struct platform_driver vc4_hvs_driver;
 
 
 
 
 
832void vc4_hvs_dump_state(struct drm_device *dev);
833void vc4_hvs_unmask_underrun(struct drm_device *dev, int channel);
834void vc4_hvs_mask_underrun(struct drm_device *dev, int channel);
835
836/* vc4_kms.c */
837int vc4_kms_load(struct drm_device *dev);
838
839/* vc4_plane.c */
840struct drm_plane *vc4_plane_init(struct drm_device *dev,
841				 enum drm_plane_type type);
 
842u32 vc4_plane_write_dlist(struct drm_plane *plane, u32 __iomem *dlist);
843u32 vc4_plane_dlist_size(const struct drm_plane_state *state);
844void vc4_plane_async_set_fb(struct drm_plane *plane,
845			    struct drm_framebuffer *fb);
846
847/* vc4_v3d.c */
848extern struct platform_driver vc4_v3d_driver;
849extern const struct of_device_id vc4_v3d_dt_match[];
850int vc4_v3d_get_bin_slot(struct vc4_dev *vc4);
851int vc4_v3d_bin_bo_get(struct vc4_dev *vc4, bool *used);
852void vc4_v3d_bin_bo_put(struct vc4_dev *vc4);
853int vc4_v3d_pm_get(struct vc4_dev *vc4);
854void vc4_v3d_pm_put(struct vc4_dev *vc4);
855
856/* vc4_validate.c */
857int
858vc4_validate_bin_cl(struct drm_device *dev,
859		    void *validated,
860		    void *unvalidated,
861		    struct vc4_exec_info *exec);
862
863int
864vc4_validate_shader_recs(struct drm_device *dev, struct vc4_exec_info *exec);
865
866struct drm_gem_cma_object *vc4_use_bo(struct vc4_exec_info *exec,
867				      uint32_t hindex);
868
869int vc4_get_rcl(struct drm_device *dev, struct vc4_exec_info *exec);
870
871bool vc4_check_tex_size(struct vc4_exec_info *exec,
872			struct drm_gem_cma_object *fbo,
873			uint32_t offset, uint8_t tiling_format,
874			uint32_t width, uint32_t height, uint8_t cpp);
875
876/* vc4_validate_shader.c */
877struct vc4_validated_shader_info *
878vc4_validate_shader(struct drm_gem_cma_object *shader_obj);
879
880/* vc4_perfmon.c */
881void vc4_perfmon_get(struct vc4_perfmon *perfmon);
882void vc4_perfmon_put(struct vc4_perfmon *perfmon);
883void vc4_perfmon_start(struct vc4_dev *vc4, struct vc4_perfmon *perfmon);
884void vc4_perfmon_stop(struct vc4_dev *vc4, struct vc4_perfmon *perfmon,
885		      bool capture);
886struct vc4_perfmon *vc4_perfmon_find(struct vc4_file *vc4file, int id);
887void vc4_perfmon_open_file(struct vc4_file *vc4file);
888void vc4_perfmon_close_file(struct vc4_file *vc4file);
889int vc4_perfmon_create_ioctl(struct drm_device *dev, void *data,
890			     struct drm_file *file_priv);
891int vc4_perfmon_destroy_ioctl(struct drm_device *dev, void *data,
892			      struct drm_file *file_priv);
893int vc4_perfmon_get_values_ioctl(struct drm_device *dev, void *data,
894				 struct drm_file *file_priv);
v5.9
  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/refcount.h>
 10#include <linux/uaccess.h>
 11
 12#include <drm/drm_atomic.h>
 13#include <drm/drm_debugfs.h>
 14#include <drm/drm_device.h>
 15#include <drm/drm_encoder.h>
 16#include <drm/drm_gem_cma_helper.h>
 17#include <drm/drm_mm.h>
 18#include <drm/drm_modeset_lock.h>
 19
 20#include "uapi/drm/vc4_drm.h"
 21
 22struct drm_device;
 23struct drm_gem_object;
 24
 25/* Don't forget to update vc4_bo.c: bo_type_names[] when adding to
 26 * this.
 27 */
 28enum vc4_kernel_bo_type {
 29	/* Any kernel allocation (gem_create_object hook) before it
 30	 * gets another type set.
 31	 */
 32	VC4_BO_TYPE_KERNEL,
 33	VC4_BO_TYPE_V3D,
 34	VC4_BO_TYPE_V3D_SHADER,
 35	VC4_BO_TYPE_DUMB,
 36	VC4_BO_TYPE_BIN,
 37	VC4_BO_TYPE_RCL,
 38	VC4_BO_TYPE_BCL,
 39	VC4_BO_TYPE_KERNEL_CACHE,
 40	VC4_BO_TYPE_COUNT
 41};
 42
 43/* Performance monitor object. The perform lifetime is controlled by userspace
 44 * using perfmon related ioctls. A perfmon can be attached to a submit_cl
 45 * request, and when this is the case, HW perf counters will be activated just
 46 * before the submit_cl is submitted to the GPU and disabled when the job is
 47 * done. This way, only events related to a specific job will be counted.
 48 */
 49struct vc4_perfmon {
 50	/* Tracks the number of users of the perfmon, when this counter reaches
 51	 * zero the perfmon is destroyed.
 52	 */
 53	refcount_t refcnt;
 54
 55	/* Number of counters activated in this perfmon instance
 56	 * (should be less than DRM_VC4_MAX_PERF_COUNTERS).
 57	 */
 58	u8 ncounters;
 59
 60	/* Events counted by the HW perf counters. */
 61	u8 events[DRM_VC4_MAX_PERF_COUNTERS];
 62
 63	/* Storage for counter values. Counters are incremented by the HW
 64	 * perf counter values every time the perfmon is attached to a GPU job.
 65	 * This way, perfmon users don't have to retrieve the results after
 66	 * each job if they want to track events covering several submissions.
 67	 * Note that counter values can't be reset, but you can fake a reset by
 68	 * destroying the perfmon and creating a new one.
 69	 */
 70	u64 counters[];
 71};
 72
 73struct vc4_dev {
 74	struct drm_device *dev;
 75
 76	struct vc4_hdmi *hdmi;
 77	struct vc4_hvs *hvs;
 78	struct vc4_v3d *v3d;
 79	struct vc4_dpi *dpi;
 80	struct vc4_dsi *dsi1;
 81	struct vc4_vec *vec;
 82	struct vc4_txp *txp;
 83
 84	struct vc4_hang_state *hang_state;
 85
 86	/* The kernel-space BO cache.  Tracks buffers that have been
 87	 * unreferenced by all other users (refcounts of 0!) but not
 88	 * yet freed, so we can do cheap allocations.
 89	 */
 90	struct vc4_bo_cache {
 91		/* Array of list heads for entries in the BO cache,
 92		 * based on number of pages, so we can do O(1) lookups
 93		 * in the cache when allocating.
 94		 */
 95		struct list_head *size_list;
 96		uint32_t size_list_size;
 97
 98		/* List of all BOs in the cache, ordered by age, so we
 99		 * can do O(1) lookups when trying to free old
100		 * buffers.
101		 */
102		struct list_head time_list;
103		struct work_struct time_work;
104		struct timer_list time_timer;
105	} bo_cache;
106
107	u32 num_labels;
108	struct vc4_label {
109		const char *name;
110		u32 num_allocated;
111		u32 size_allocated;
112	} *bo_labels;
113
114	/* Protects bo_cache and bo_labels. */
115	struct mutex bo_lock;
116
117	/* Purgeable BO pool. All BOs in this pool can have their memory
118	 * reclaimed if the driver is unable to allocate new BOs. We also
119	 * keep stats related to the purge mechanism here.
120	 */
121	struct {
122		struct list_head list;
123		unsigned int num;
124		size_t size;
125		unsigned int purged_num;
126		size_t purged_size;
127		struct mutex lock;
128	} purgeable;
129
130	uint64_t dma_fence_context;
131
132	/* Sequence number for the last job queued in bin_job_list.
133	 * Starts at 0 (no jobs emitted).
134	 */
135	uint64_t emit_seqno;
136
137	/* Sequence number for the last completed job on the GPU.
138	 * Starts at 0 (no jobs completed).
139	 */
140	uint64_t finished_seqno;
141
142	/* List of all struct vc4_exec_info for jobs to be executed in
143	 * the binner.  The first job in the list is the one currently
144	 * programmed into ct0ca for execution.
145	 */
146	struct list_head bin_job_list;
147
148	/* List of all struct vc4_exec_info for jobs that have
149	 * completed binning and are ready for rendering.  The first
150	 * job in the list is the one currently programmed into ct1ca
151	 * for execution.
152	 */
153	struct list_head render_job_list;
154
155	/* List of the finished vc4_exec_infos waiting to be freed by
156	 * job_done_work.
157	 */
158	struct list_head job_done_list;
159	/* Spinlock used to synchronize the job_list and seqno
160	 * accesses between the IRQ handler and GEM ioctls.
161	 */
162	spinlock_t job_lock;
163	wait_queue_head_t job_wait_queue;
164	struct work_struct job_done_work;
165
166	/* Used to track the active perfmon if any. Access to this field is
167	 * protected by job_lock.
168	 */
169	struct vc4_perfmon *active_perfmon;
170
171	/* List of struct vc4_seqno_cb for callbacks to be made from a
172	 * workqueue when the given seqno is passed.
173	 */
174	struct list_head seqno_cb_list;
175
176	/* The memory used for storing binner tile alloc, tile state,
177	 * and overflow memory allocations.  This is freed when V3D
178	 * powers down.
179	 */
180	struct vc4_bo *bin_bo;
181
182	/* Size of blocks allocated within bin_bo. */
183	uint32_t bin_alloc_size;
184
185	/* Bitmask of the bin_alloc_size chunks in bin_bo that are
186	 * used.
187	 */
188	uint32_t bin_alloc_used;
189
190	/* Bitmask of the current bin_alloc used for overflow memory. */
191	uint32_t bin_alloc_overflow;
192
193	/* Incremented when an underrun error happened after an atomic commit.
194	 * This is particularly useful to detect when a specific modeset is too
195	 * demanding in term of memory or HVS bandwidth which is hard to guess
196	 * at atomic check time.
197	 */
198	atomic_t underrun;
199
200	struct work_struct overflow_mem_work;
201
202	int power_refcount;
203
204	/* Set to true when the load tracker is active. */
205	bool load_tracker_enabled;
206
207	/* Mutex controlling the power refcount. */
208	struct mutex power_lock;
209
210	struct {
211		struct timer_list timer;
212		struct work_struct reset_work;
213	} hangcheck;
214
215	struct semaphore async_modeset;
216
217	struct drm_modeset_lock ctm_state_lock;
218	struct drm_private_obj ctm_manager;
219	struct drm_private_obj load_tracker;
220
221	/* List of vc4_debugfs_info_entry for adding to debugfs once
222	 * the minor is available (after drm_dev_register()).
223	 */
224	struct list_head debugfs_list;
225
226	/* Mutex for binner bo allocation. */
227	struct mutex bin_bo_lock;
228	/* Reference count for our binner bo. */
229	struct kref bin_bo_kref;
230};
231
232static inline struct vc4_dev *
233to_vc4_dev(struct drm_device *dev)
234{
235	return (struct vc4_dev *)dev->dev_private;
236}
237
238struct vc4_bo {
239	struct drm_gem_cma_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
285static inline struct vc4_bo *
286to_vc4_bo(struct drm_gem_object *bo)
287{
288	return (struct vc4_bo *)bo;
289}
290
291struct vc4_fence {
292	struct dma_fence base;
293	struct drm_device *dev;
294	/* vc4 seqno for signaled() test */
295	uint64_t seqno;
296};
297
298static inline struct vc4_fence *
299to_vc4_fence(struct dma_fence *fence)
300{
301	return (struct vc4_fence *)fence;
302}
303
304struct vc4_seqno_cb {
305	struct work_struct work;
306	uint64_t seqno;
307	void (*func)(struct vc4_seqno_cb *cb);
308};
309
310struct vc4_v3d {
311	struct vc4_dev *vc4;
312	struct platform_device *pdev;
313	void __iomem *regs;
314	struct clk *clk;
315	struct debugfs_regset32 regset;
316};
317
318struct vc4_hvs {
319	struct platform_device *pdev;
320	void __iomem *regs;
321	u32 __iomem *dlist;
322
323	/* Memory manager for CRTCs to allocate space in the display
324	 * list.  Units are dwords.
325	 */
326	struct drm_mm dlist_mm;
327	/* Memory manager for the LBM memory used by HVS scaling. */
328	struct drm_mm lbm_mm;
329	spinlock_t mm_lock;
330
331	struct drm_mm_node mitchell_netravali_filter;
332	struct debugfs_regset32 regset;
333};
334
335struct vc4_plane {
336	struct drm_plane base;
337};
338
339static inline struct vc4_plane *
340to_vc4_plane(struct drm_plane *plane)
341{
342	return (struct vc4_plane *)plane;
343}
344
345enum vc4_scaling_mode {
346	VC4_SCALING_NONE,
347	VC4_SCALING_TPZ,
348	VC4_SCALING_PPF,
349};
350
351struct vc4_plane_state {
352	struct drm_plane_state base;
353	/* System memory copy of the display list for this element, computed
354	 * at atomic_check time.
355	 */
356	u32 *dlist;
357	u32 dlist_size; /* Number of dwords allocated for the display list */
358	u32 dlist_count; /* Number of used dwords in the display list. */
359
360	/* Offset in the dlist to various words, for pageflip or
361	 * cursor updates.
362	 */
363	u32 pos0_offset;
364	u32 pos2_offset;
365	u32 ptr0_offset;
366	u32 lbm_offset;
367
368	/* Offset where the plane's dlist was last stored in the
369	 * hardware at vc4_crtc_atomic_flush() time.
370	 */
371	u32 __iomem *hw_dlist;
372
373	/* Clipped coordinates of the plane on the display. */
374	int crtc_x, crtc_y, crtc_w, crtc_h;
375	/* Clipped area being scanned from in the FB. */
376	u32 src_x, src_y;
377
378	u32 src_w[2], src_h[2];
379
380	/* Scaling selection for the RGB/Y plane and the Cb/Cr planes. */
381	enum vc4_scaling_mode x_scaling[2], y_scaling[2];
382	bool is_unity;
383	bool is_yuv;
384
385	/* Offset to start scanning out from the start of the plane's
386	 * BO.
387	 */
388	u32 offsets[3];
389
390	/* Our allocation in LBM for temporary storage during scaling. */
391	struct drm_mm_node lbm;
392
393	/* Set when the plane has per-pixel alpha content or does not cover
394	 * the entire screen. This is a hint to the CRTC that it might need
395	 * to enable background color fill.
396	 */
397	bool needs_bg_fill;
398
399	/* Mark the dlist as initialized. Useful to avoid initializing it twice
400	 * when async update is not possible.
401	 */
402	bool dlist_initialized;
403
404	/* Load of this plane on the HVS block. The load is expressed in HVS
405	 * cycles/sec.
406	 */
407	u64 hvs_load;
408
409	/* Memory bandwidth needed for this plane. This is expressed in
410	 * bytes/sec.
411	 */
412	u64 membus_load;
413};
414
415static inline struct vc4_plane_state *
416to_vc4_plane_state(struct drm_plane_state *state)
417{
418	return (struct vc4_plane_state *)state;
419}
420
421enum vc4_encoder_type {
422	VC4_ENCODER_TYPE_NONE,
423	VC4_ENCODER_TYPE_HDMI,
424	VC4_ENCODER_TYPE_VEC,
425	VC4_ENCODER_TYPE_DSI0,
426	VC4_ENCODER_TYPE_DSI1,
427	VC4_ENCODER_TYPE_SMI,
428	VC4_ENCODER_TYPE_DPI,
429};
430
431struct vc4_encoder {
432	struct drm_encoder base;
433	enum vc4_encoder_type type;
434	u32 clock_select;
435};
436
437static inline struct vc4_encoder *
438to_vc4_encoder(struct drm_encoder *encoder)
439{
440	return container_of(encoder, struct vc4_encoder, base);
441}
442
443struct vc4_crtc_data {
444	/* Which channel of the HVS this pixelvalve sources from. */
445	int hvs_channel;
446};
447
448struct vc4_pv_data {
449	struct vc4_crtc_data	base;
450
451	enum vc4_encoder_type encoder_types[4];
452	const char *debugfs_name;
453
454};
455
456struct vc4_crtc {
457	struct drm_crtc base;
458	struct platform_device *pdev;
459	const struct vc4_crtc_data *data;
460	void __iomem *regs;
461
462	/* Timestamp at start of vblank irq - unaffected by lock delays. */
463	ktime_t t_vblank;
464
465	/* Which HVS channel we're using for our CRTC. */
466	int channel;
467
468	u8 lut_r[256];
469	u8 lut_g[256];
470	u8 lut_b[256];
471	/* Size in pixels of the COB memory allocated to this CRTC. */
472	u32 cob_size;
473
474	struct drm_pending_vblank_event *event;
475
476	struct debugfs_regset32 regset;
477};
478
479static inline struct vc4_crtc *
480to_vc4_crtc(struct drm_crtc *crtc)
481{
482	return (struct vc4_crtc *)crtc;
483}
484
485static inline const struct vc4_crtc_data *
486vc4_crtc_to_vc4_crtc_data(const struct vc4_crtc *crtc)
487{
488	return crtc->data;
489}
490
491static inline const struct vc4_pv_data *
492vc4_crtc_to_vc4_pv_data(const struct vc4_crtc *crtc)
493{
494	const struct vc4_crtc_data *data = vc4_crtc_to_vc4_crtc_data(crtc);
495
496	return container_of(data, struct vc4_pv_data, base);
497}
498
499struct vc4_crtc_state {
500	struct drm_crtc_state base;
501	/* Dlist area for this CRTC configuration. */
502	struct drm_mm_node mm;
503	bool feed_txp;
504	bool txp_armed;
505
506	struct {
507		unsigned int left;
508		unsigned int right;
509		unsigned int top;
510		unsigned int bottom;
511	} margins;
512};
513
514static inline struct vc4_crtc_state *
515to_vc4_crtc_state(struct drm_crtc_state *crtc_state)
516{
517	return (struct vc4_crtc_state *)crtc_state;
518}
519
520#define V3D_READ(offset) readl(vc4->v3d->regs + offset)
521#define V3D_WRITE(offset, val) writel(val, vc4->v3d->regs + offset)
522#define HVS_READ(offset) readl(vc4->hvs->regs + offset)
523#define HVS_WRITE(offset, val) writel(val, vc4->hvs->regs + offset)
524
525#define VC4_REG32(reg) { .name = #reg, .offset = reg }
526
527struct vc4_exec_info {
528	/* Sequence number for this bin/render job. */
529	uint64_t seqno;
530
531	/* Latest write_seqno of any BO that binning depends on. */
532	uint64_t bin_dep_seqno;
533
534	struct dma_fence *fence;
535
536	/* Last current addresses the hardware was processing when the
537	 * hangcheck timer checked on us.
538	 */
539	uint32_t last_ct0ca, last_ct1ca;
540
541	/* Kernel-space copy of the ioctl arguments */
542	struct drm_vc4_submit_cl *args;
543
544	/* This is the array of BOs that were looked up at the start of exec.
545	 * Command validation will use indices into this array.
546	 */
547	struct drm_gem_cma_object **bo;
548	uint32_t bo_count;
549
550	/* List of BOs that are being written by the RCL.  Other than
551	 * the binner temporary storage, this is all the BOs written
552	 * by the job.
553	 */
554	struct drm_gem_cma_object *rcl_write_bo[4];
555	uint32_t rcl_write_bo_count;
556
557	/* Pointers for our position in vc4->job_list */
558	struct list_head head;
559
560	/* List of other BOs used in the job that need to be released
561	 * once the job is complete.
562	 */
563	struct list_head unref_list;
564
565	/* Current unvalidated indices into @bo loaded by the non-hardware
566	 * VC4_PACKET_GEM_HANDLES.
567	 */
568	uint32_t bo_index[2];
569
570	/* This is the BO where we store the validated command lists, shader
571	 * records, and uniforms.
572	 */
573	struct drm_gem_cma_object *exec_bo;
574
575	/**
576	 * This tracks the per-shader-record state (packet 64) that
577	 * determines the length of the shader record and the offset
578	 * it's expected to be found at.  It gets read in from the
579	 * command lists.
580	 */
581	struct vc4_shader_state {
582		uint32_t addr;
583		/* Maximum vertex index referenced by any primitive using this
584		 * shader state.
585		 */
586		uint32_t max_index;
587	} *shader_state;
588
589	/** How many shader states the user declared they were using. */
590	uint32_t shader_state_size;
591	/** How many shader state records the validator has seen. */
592	uint32_t shader_state_count;
593
594	bool found_tile_binning_mode_config_packet;
595	bool found_start_tile_binning_packet;
596	bool found_increment_semaphore_packet;
597	bool found_flush;
598	uint8_t bin_tiles_x, bin_tiles_y;
599	/* Physical address of the start of the tile alloc array
600	 * (where each tile's binned CL will start)
601	 */
602	uint32_t tile_alloc_offset;
603	/* Bitmask of which binner slots are freed when this job completes. */
604	uint32_t bin_slots;
605
606	/**
607	 * Computed addresses pointing into exec_bo where we start the
608	 * bin thread (ct0) and render thread (ct1).
609	 */
610	uint32_t ct0ca, ct0ea;
611	uint32_t ct1ca, ct1ea;
612
613	/* Pointer to the unvalidated bin CL (if present). */
614	void *bin_u;
615
616	/* Pointers to the shader recs.  These paddr gets incremented as CL
617	 * packets are relocated in validate_gl_shader_state, and the vaddrs
618	 * (u and v) get incremented and size decremented as the shader recs
619	 * themselves are validated.
620	 */
621	void *shader_rec_u;
622	void *shader_rec_v;
623	uint32_t shader_rec_p;
624	uint32_t shader_rec_size;
625
626	/* Pointers to the uniform data.  These pointers are incremented, and
627	 * size decremented, as each batch of uniforms is uploaded.
628	 */
629	void *uniforms_u;
630	void *uniforms_v;
631	uint32_t uniforms_p;
632	uint32_t uniforms_size;
633
634	/* Pointer to a performance monitor object if the user requested it,
635	 * NULL otherwise.
636	 */
637	struct vc4_perfmon *perfmon;
638
639	/* Whether the exec has taken a reference to the binner BO, which should
640	 * happen with a VC4_PACKET_TILE_BINNING_MODE_CONFIG packet.
641	 */
642	bool bin_bo_used;
643};
644
645/* Per-open file private data. Any driver-specific resource that has to be
646 * released when the DRM file is closed should be placed here.
647 */
648struct vc4_file {
649	struct {
650		struct idr idr;
651		struct mutex lock;
652	} perfmon;
653
654	bool bin_bo_used;
655};
656
657static inline struct vc4_exec_info *
658vc4_first_bin_job(struct vc4_dev *vc4)
659{
660	return list_first_entry_or_null(&vc4->bin_job_list,
661					struct vc4_exec_info, head);
662}
663
664static inline struct vc4_exec_info *
665vc4_first_render_job(struct vc4_dev *vc4)
666{
667	return list_first_entry_or_null(&vc4->render_job_list,
668					struct vc4_exec_info, head);
669}
670
671static inline struct vc4_exec_info *
672vc4_last_render_job(struct vc4_dev *vc4)
673{
674	if (list_empty(&vc4->render_job_list))
675		return NULL;
676	return list_last_entry(&vc4->render_job_list,
677			       struct vc4_exec_info, head);
678}
679
680/**
681 * struct vc4_texture_sample_info - saves the offsets into the UBO for texture
682 * setup parameters.
683 *
684 * This will be used at draw time to relocate the reference to the texture
685 * contents in p0, and validate that the offset combined with
686 * width/height/stride/etc. from p1 and p2/p3 doesn't sample outside the BO.
687 * Note that the hardware treats unprovided config parameters as 0, so not all
688 * of them need to be set up for every texure sample, and we'll store ~0 as
689 * the offset to mark the unused ones.
690 *
691 * See the VC4 3D architecture guide page 41 ("Texture and Memory Lookup Unit
692 * Setup") for definitions of the texture parameters.
693 */
694struct vc4_texture_sample_info {
695	bool is_direct;
696	uint32_t p_offset[4];
697};
698
699/**
700 * struct vc4_validated_shader_info - information about validated shaders that
701 * needs to be used from command list validation.
702 *
703 * For a given shader, each time a shader state record references it, we need
704 * to verify that the shader doesn't read more uniforms than the shader state
705 * record's uniform BO pointer can provide, and we need to apply relocations
706 * and validate the shader state record's uniforms that define the texture
707 * samples.
708 */
709struct vc4_validated_shader_info {
710	uint32_t uniforms_size;
711	uint32_t uniforms_src_size;
712	uint32_t num_texture_samples;
713	struct vc4_texture_sample_info *texture_samples;
714
715	uint32_t num_uniform_addr_offsets;
716	uint32_t *uniform_addr_offsets;
717
718	bool is_threaded;
719};
720
721/**
722 * __wait_for - magic wait macro
723 *
724 * Macro to help avoid open coding check/wait/timeout patterns. Note that it's
725 * important that we check the condition again after having timed out, since the
726 * timeout could be due to preemption or similar and we've never had a chance to
727 * check the condition before the timeout.
728 */
729#define __wait_for(OP, COND, US, Wmin, Wmax) ({ \
730	const ktime_t end__ = ktime_add_ns(ktime_get_raw(), 1000ll * (US)); \
731	long wait__ = (Wmin); /* recommended min for usleep is 10 us */	\
732	int ret__;							\
733	might_sleep();							\
734	for (;;) {							\
735		const bool expired__ = ktime_after(ktime_get_raw(), end__); \
736		OP;							\
737		/* Guarantee COND check prior to timeout */		\
738		barrier();						\
739		if (COND) {						\
740			ret__ = 0;					\
741			break;						\
742		}							\
743		if (expired__) {					\
744			ret__ = -ETIMEDOUT;				\
745			break;						\
 
746		}							\
747		usleep_range(wait__, wait__ * 2);			\
748		if (wait__ < (Wmax))					\
749			wait__ <<= 1;					\
750	}								\
751	ret__;								\
752})
753
754#define _wait_for(COND, US, Wmin, Wmax)	__wait_for(, (COND), (US), (Wmin), \
755						   (Wmax))
756#define wait_for(COND, MS)		_wait_for((COND), (MS) * 1000, 10, 1000)
757
758/* vc4_bo.c */
759struct drm_gem_object *vc4_create_object(struct drm_device *dev, size_t size);
760void vc4_free_object(struct drm_gem_object *gem_obj);
761struct vc4_bo *vc4_bo_create(struct drm_device *dev, size_t size,
762			     bool from_cache, enum vc4_kernel_bo_type type);
763int vc4_dumb_create(struct drm_file *file_priv,
764		    struct drm_device *dev,
765		    struct drm_mode_create_dumb *args);
766struct dma_buf *vc4_prime_export(struct drm_gem_object *obj, int flags);
767int vc4_create_bo_ioctl(struct drm_device *dev, void *data,
768			struct drm_file *file_priv);
769int vc4_create_shader_bo_ioctl(struct drm_device *dev, void *data,
770			       struct drm_file *file_priv);
771int vc4_mmap_bo_ioctl(struct drm_device *dev, void *data,
772		      struct drm_file *file_priv);
773int vc4_set_tiling_ioctl(struct drm_device *dev, void *data,
774			 struct drm_file *file_priv);
775int vc4_get_tiling_ioctl(struct drm_device *dev, void *data,
776			 struct drm_file *file_priv);
777int vc4_get_hang_state_ioctl(struct drm_device *dev, void *data,
778			     struct drm_file *file_priv);
779int vc4_label_bo_ioctl(struct drm_device *dev, void *data,
780		       struct drm_file *file_priv);
781vm_fault_t vc4_fault(struct vm_fault *vmf);
782int vc4_mmap(struct file *filp, struct vm_area_struct *vma);
783int vc4_prime_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma);
784struct drm_gem_object *vc4_prime_import_sg_table(struct drm_device *dev,
785						 struct dma_buf_attachment *attach,
786						 struct sg_table *sgt);
787void *vc4_prime_vmap(struct drm_gem_object *obj);
788int vc4_bo_cache_init(struct drm_device *dev);
789void vc4_bo_cache_destroy(struct drm_device *dev);
790int vc4_bo_inc_usecnt(struct vc4_bo *bo);
791void vc4_bo_dec_usecnt(struct vc4_bo *bo);
792void vc4_bo_add_to_purgeable_pool(struct vc4_bo *bo);
793void vc4_bo_remove_from_purgeable_pool(struct vc4_bo *bo);
794
795/* vc4_crtc.c */
796extern struct platform_driver vc4_crtc_driver;
797int vc4_crtc_init(struct drm_device *drm, struct vc4_crtc *vc4_crtc,
798		  const struct drm_crtc_funcs *crtc_funcs,
799		  const struct drm_crtc_helper_funcs *crtc_helper_funcs);
800void vc4_crtc_destroy(struct drm_crtc *crtc);
801int vc4_page_flip(struct drm_crtc *crtc,
802		  struct drm_framebuffer *fb,
803		  struct drm_pending_vblank_event *event,
804		  uint32_t flags,
805		  struct drm_modeset_acquire_ctx *ctx);
806struct drm_crtc_state *vc4_crtc_duplicate_state(struct drm_crtc *crtc);
807void vc4_crtc_destroy_state(struct drm_crtc *crtc,
808			    struct drm_crtc_state *state);
809void vc4_crtc_reset(struct drm_crtc *crtc);
810void vc4_crtc_handle_vblank(struct vc4_crtc *crtc);
 
811void vc4_crtc_get_margins(struct drm_crtc_state *state,
812			  unsigned int *right, unsigned int *left,
813			  unsigned int *top, unsigned int *bottom);
814
815/* vc4_debugfs.c */
816void vc4_debugfs_init(struct drm_minor *minor);
817#ifdef CONFIG_DEBUG_FS
818void vc4_debugfs_add_file(struct drm_device *drm,
819			  const char *filename,
820			  int (*show)(struct seq_file*, void*),
821			  void *data);
822void vc4_debugfs_add_regset32(struct drm_device *drm,
823			      const char *filename,
824			      struct debugfs_regset32 *regset);
825#else
826static inline void vc4_debugfs_add_file(struct drm_device *drm,
827					const char *filename,
828					int (*show)(struct seq_file*, void*),
829					void *data)
830{
831}
832
833static inline void vc4_debugfs_add_regset32(struct drm_device *drm,
834					    const char *filename,
835					    struct debugfs_regset32 *regset)
836{
837}
838#endif
839
840/* vc4_drv.c */
841void __iomem *vc4_ioremap_regs(struct platform_device *dev, int index);
842
843/* vc4_dpi.c */
844extern struct platform_driver vc4_dpi_driver;
845
846/* vc4_dsi.c */
847extern struct platform_driver vc4_dsi_driver;
848
849/* vc4_fence.c */
850extern const struct dma_fence_ops vc4_fence_ops;
851
852/* vc4_gem.c */
853void vc4_gem_init(struct drm_device *dev);
854void vc4_gem_destroy(struct drm_device *dev);
855int vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
856			struct drm_file *file_priv);
857int vc4_wait_seqno_ioctl(struct drm_device *dev, void *data,
858			 struct drm_file *file_priv);
859int vc4_wait_bo_ioctl(struct drm_device *dev, void *data,
860		      struct drm_file *file_priv);
861void vc4_submit_next_bin_job(struct drm_device *dev);
862void vc4_submit_next_render_job(struct drm_device *dev);
863void vc4_move_job_to_render(struct drm_device *dev, struct vc4_exec_info *exec);
864int vc4_wait_for_seqno(struct drm_device *dev, uint64_t seqno,
865		       uint64_t timeout_ns, bool interruptible);
866void vc4_job_handle_completed(struct vc4_dev *vc4);
867int vc4_queue_seqno_cb(struct drm_device *dev,
868		       struct vc4_seqno_cb *cb, uint64_t seqno,
869		       void (*func)(struct vc4_seqno_cb *cb));
870int vc4_gem_madvise_ioctl(struct drm_device *dev, void *data,
871			  struct drm_file *file_priv);
872
873/* vc4_hdmi.c */
874extern struct platform_driver vc4_hdmi_driver;
875
876/* vc4_vec.c */
877extern struct platform_driver vc4_vec_driver;
878
879/* vc4_txp.c */
880extern struct platform_driver vc4_txp_driver;
881
882/* vc4_irq.c */
883irqreturn_t vc4_irq(int irq, void *arg);
884void vc4_irq_preinstall(struct drm_device *dev);
885int vc4_irq_postinstall(struct drm_device *dev);
886void vc4_irq_uninstall(struct drm_device *dev);
887void vc4_irq_reset(struct drm_device *dev);
888
889/* vc4_hvs.c */
890extern struct platform_driver vc4_hvs_driver;
891int vc4_hvs_atomic_check(struct drm_crtc *crtc, struct drm_crtc_state *state);
892void vc4_hvs_atomic_enable(struct drm_crtc *crtc, struct drm_crtc_state *old_state);
893void vc4_hvs_atomic_disable(struct drm_crtc *crtc, struct drm_crtc_state *old_state);
894void vc4_hvs_atomic_flush(struct drm_crtc *crtc, struct drm_crtc_state *state);
895void vc4_hvs_mode_set_nofb(struct drm_crtc *crtc);
896void vc4_hvs_dump_state(struct drm_device *dev);
897void vc4_hvs_unmask_underrun(struct drm_device *dev, int channel);
898void vc4_hvs_mask_underrun(struct drm_device *dev, int channel);
899
900/* vc4_kms.c */
901int vc4_kms_load(struct drm_device *dev);
902
903/* vc4_plane.c */
904struct drm_plane *vc4_plane_init(struct drm_device *dev,
905				 enum drm_plane_type type);
906int vc4_plane_create_additional_planes(struct drm_device *dev);
907u32 vc4_plane_write_dlist(struct drm_plane *plane, u32 __iomem *dlist);
908u32 vc4_plane_dlist_size(const struct drm_plane_state *state);
909void vc4_plane_async_set_fb(struct drm_plane *plane,
910			    struct drm_framebuffer *fb);
911
912/* vc4_v3d.c */
913extern struct platform_driver vc4_v3d_driver;
914extern const struct of_device_id vc4_v3d_dt_match[];
915int vc4_v3d_get_bin_slot(struct vc4_dev *vc4);
916int vc4_v3d_bin_bo_get(struct vc4_dev *vc4, bool *used);
917void vc4_v3d_bin_bo_put(struct vc4_dev *vc4);
918int vc4_v3d_pm_get(struct vc4_dev *vc4);
919void vc4_v3d_pm_put(struct vc4_dev *vc4);
920
921/* vc4_validate.c */
922int
923vc4_validate_bin_cl(struct drm_device *dev,
924		    void *validated,
925		    void *unvalidated,
926		    struct vc4_exec_info *exec);
927
928int
929vc4_validate_shader_recs(struct drm_device *dev, struct vc4_exec_info *exec);
930
931struct drm_gem_cma_object *vc4_use_bo(struct vc4_exec_info *exec,
932				      uint32_t hindex);
933
934int vc4_get_rcl(struct drm_device *dev, struct vc4_exec_info *exec);
935
936bool vc4_check_tex_size(struct vc4_exec_info *exec,
937			struct drm_gem_cma_object *fbo,
938			uint32_t offset, uint8_t tiling_format,
939			uint32_t width, uint32_t height, uint8_t cpp);
940
941/* vc4_validate_shader.c */
942struct vc4_validated_shader_info *
943vc4_validate_shader(struct drm_gem_cma_object *shader_obj);
944
945/* vc4_perfmon.c */
946void vc4_perfmon_get(struct vc4_perfmon *perfmon);
947void vc4_perfmon_put(struct vc4_perfmon *perfmon);
948void vc4_perfmon_start(struct vc4_dev *vc4, struct vc4_perfmon *perfmon);
949void vc4_perfmon_stop(struct vc4_dev *vc4, struct vc4_perfmon *perfmon,
950		      bool capture);
951struct vc4_perfmon *vc4_perfmon_find(struct vc4_file *vc4file, int id);
952void vc4_perfmon_open_file(struct vc4_file *vc4file);
953void vc4_perfmon_close_file(struct vc4_file *vc4file);
954int vc4_perfmon_create_ioctl(struct drm_device *dev, void *data,
955			     struct drm_file *file_priv);
956int vc4_perfmon_destroy_ioctl(struct drm_device *dev, void *data,
957			      struct drm_file *file_priv);
958int vc4_perfmon_get_values_ioctl(struct drm_device *dev, void *data,
959				 struct drm_file *file_priv);
960
961#endif /* _VC4_DRV_H_ */