2 * Copyright (C) 2015 Broadcom
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.
9 #include <linux/mm_types.h>
11 #include <drm/drm_util.h>
12 #include <drm/drm_encoder.h>
13 #include <drm/drm_gem_cma_helper.h>
14 #include <drm/drm_atomic.h>
15 #include <drm/drm_syncobj.h>
17 #include "uapi/drm/vc4_drm.h"
19 /* Don't forget to update vc4_bo.c: bo_type_names[] when adding to
22 enum vc4_kernel_bo_type {
23 /* Any kernel allocation (gem_create_object hook) before it
24 * gets another type set.
28 VC4_BO_TYPE_V3D_SHADER,
33 VC4_BO_TYPE_KERNEL_CACHE,
37 /* Performance monitor object. The perform lifetime is controlled by userspace
38 * using perfmon related ioctls. A perfmon can be attached to a submit_cl
39 * request, and when this is the case, HW perf counters will be activated just
40 * before the submit_cl is submitted to the GPU and disabled when the job is
41 * done. This way, only events related to a specific job will be counted.
44 /* Tracks the number of users of the perfmon, when this counter reaches
45 * zero the perfmon is destroyed.
49 /* Number of counters activated in this perfmon instance
50 * (should be less than DRM_VC4_MAX_PERF_COUNTERS).
54 /* Events counted by the HW perf counters. */
55 u8 events[DRM_VC4_MAX_PERF_COUNTERS];
57 /* Storage for counter values. Counters are incremented by the HW
58 * perf counter values every time the perfmon is attached to a GPU job.
59 * This way, perfmon users don't have to retrieve the results after
60 * each job if they want to track events covering several submissions.
61 * Note that counter values can't be reset, but you can fake a reset by
62 * destroying the perfmon and creating a new one.
68 struct drm_device *dev;
70 struct vc4_hdmi *hdmi;
78 struct vc4_hang_state *hang_state;
80 /* The kernel-space BO cache. Tracks buffers that have been
81 * unreferenced by all other users (refcounts of 0!) but not
82 * yet freed, so we can do cheap allocations.
85 /* Array of list heads for entries in the BO cache,
86 * based on number of pages, so we can do O(1) lookups
87 * in the cache when allocating.
89 struct list_head *size_list;
90 uint32_t size_list_size;
92 /* List of all BOs in the cache, ordered by age, so we
93 * can do O(1) lookups when trying to free old
96 struct list_head time_list;
97 struct work_struct time_work;
98 struct timer_list time_timer;
108 /* Protects bo_cache and bo_labels. */
109 struct mutex bo_lock;
111 /* Purgeable BO pool. All BOs in this pool can have their memory
112 * reclaimed if the driver is unable to allocate new BOs. We also
113 * keep stats related to the purge mechanism here.
116 struct list_head list;
119 unsigned int purged_num;
124 uint64_t dma_fence_context;
126 /* Sequence number for the last job queued in bin_job_list.
127 * Starts at 0 (no jobs emitted).
131 /* Sequence number for the last completed job on the GPU.
132 * Starts at 0 (no jobs completed).
134 uint64_t finished_seqno;
136 /* List of all struct vc4_exec_info for jobs to be executed in
137 * the binner. The first job in the list is the one currently
138 * programmed into ct0ca for execution.
140 struct list_head bin_job_list;
142 /* List of all struct vc4_exec_info for jobs that have
143 * completed binning and are ready for rendering. The first
144 * job in the list is the one currently programmed into ct1ca
147 struct list_head render_job_list;
149 /* List of the finished vc4_exec_infos waiting to be freed by
152 struct list_head job_done_list;
153 /* Spinlock used to synchronize the job_list and seqno
154 * accesses between the IRQ handler and GEM ioctls.
157 wait_queue_head_t job_wait_queue;
158 struct work_struct job_done_work;
160 /* Used to track the active perfmon if any. Access to this field is
161 * protected by job_lock.
163 struct vc4_perfmon *active_perfmon;
165 /* List of struct vc4_seqno_cb for callbacks to be made from a
166 * workqueue when the given seqno is passed.
168 struct list_head seqno_cb_list;
170 /* The memory used for storing binner tile alloc, tile state,
171 * and overflow memory allocations. This is freed when V3D
174 struct vc4_bo *bin_bo;
176 /* Size of blocks allocated within bin_bo. */
177 uint32_t bin_alloc_size;
179 /* Bitmask of the bin_alloc_size chunks in bin_bo that are
182 uint32_t bin_alloc_used;
184 /* Bitmask of the current bin_alloc used for overflow memory. */
185 uint32_t bin_alloc_overflow;
187 /* Incremented when an underrun error happened after an atomic commit.
188 * This is particularly useful to detect when a specific modeset is too
189 * demanding in term of memory or HVS bandwidth which is hard to guess
190 * at atomic check time.
194 struct work_struct overflow_mem_work;
198 /* Mutex controlling the power refcount. */
199 struct mutex power_lock;
202 struct timer_list timer;
203 struct work_struct reset_work;
206 struct semaphore async_modeset;
208 struct drm_modeset_lock ctm_state_lock;
209 struct drm_private_obj ctm_manager;
212 static inline struct vc4_dev *
213 to_vc4_dev(struct drm_device *dev)
215 return (struct vc4_dev *)dev->dev_private;
219 struct drm_gem_cma_object base;
221 /* seqno of the last job to render using this BO. */
224 /* seqno of the last job to use the RCL to write to this BO.
226 * Note that this doesn't include binner overflow memory
229 uint64_t write_seqno;
233 /* List entry for the BO's position in either
234 * vc4_exec_info->unref_list or vc4_dev->bo_cache.time_list
236 struct list_head unref_head;
238 /* Time in jiffies when the BO was put in vc4->bo_cache. */
239 unsigned long free_time;
241 /* List entry for the BO's position in vc4_dev->bo_cache.size_list */
242 struct list_head size_head;
244 /* Struct for shader validation state, if created by
245 * DRM_IOCTL_VC4_CREATE_SHADER_BO.
247 struct vc4_validated_shader_info *validated_shader;
249 /* One of enum vc4_kernel_bo_type, or VC4_BO_TYPE_COUNT + i
250 * for user-allocated labels.
254 /* Count the number of active users. This is needed to determine
255 * whether we can move the BO to the purgeable list or not (when the BO
256 * is used by the GPU or the display engine we can't purge it).
260 /* Store purgeable/purged state here */
262 struct mutex madv_lock;
265 static inline struct vc4_bo *
266 to_vc4_bo(struct drm_gem_object *bo)
268 return (struct vc4_bo *)bo;
272 struct dma_fence base;
273 struct drm_device *dev;
274 /* vc4 seqno for signaled() test */
278 static inline struct vc4_fence *
279 to_vc4_fence(struct dma_fence *fence)
281 return (struct vc4_fence *)fence;
284 struct vc4_seqno_cb {
285 struct work_struct work;
287 void (*func)(struct vc4_seqno_cb *cb);
292 struct platform_device *pdev;
298 struct platform_device *pdev;
302 /* Memory manager for CRTCs to allocate space in the display
303 * list. Units are dwords.
305 struct drm_mm dlist_mm;
306 /* Memory manager for the LBM memory used by HVS scaling. */
307 struct drm_mm lbm_mm;
310 struct drm_mm_node mitchell_netravali_filter;
314 struct drm_plane base;
317 static inline struct vc4_plane *
318 to_vc4_plane(struct drm_plane *plane)
320 return (struct vc4_plane *)plane;
323 enum vc4_scaling_mode {
329 struct vc4_plane_state {
330 struct drm_plane_state base;
331 /* System memory copy of the display list for this element, computed
332 * at atomic_check time.
335 u32 dlist_size; /* Number of dwords allocated for the display list */
336 u32 dlist_count; /* Number of used dwords in the display list. */
338 /* Offset in the dlist to various words, for pageflip or
346 /* Offset where the plane's dlist was last stored in the
347 * hardware at vc4_crtc_atomic_flush() time.
349 u32 __iomem *hw_dlist;
351 /* Clipped coordinates of the plane on the display. */
352 int crtc_x, crtc_y, crtc_w, crtc_h;
353 /* Clipped area being scanned from in the FB. */
356 u32 src_w[2], src_h[2];
358 /* Scaling selection for the RGB/Y plane and the Cb/Cr planes. */
359 enum vc4_scaling_mode x_scaling[2], y_scaling[2];
363 /* Offset to start scanning out from the start of the plane's
368 /* Our allocation in LBM for temporary storage during scaling. */
369 struct drm_mm_node lbm;
371 /* Set when the plane has per-pixel alpha content or does not cover
372 * the entire screen. This is a hint to the CRTC that it might need
373 * to enable background color fill.
377 /* Mark the dlist as initialized. Useful to avoid initializing it twice
378 * when async update is not possible.
380 bool dlist_initialized;
383 static inline struct vc4_plane_state *
384 to_vc4_plane_state(struct drm_plane_state *state)
386 return (struct vc4_plane_state *)state;
389 enum vc4_encoder_type {
390 VC4_ENCODER_TYPE_NONE,
391 VC4_ENCODER_TYPE_HDMI,
392 VC4_ENCODER_TYPE_VEC,
393 VC4_ENCODER_TYPE_DSI0,
394 VC4_ENCODER_TYPE_DSI1,
395 VC4_ENCODER_TYPE_SMI,
396 VC4_ENCODER_TYPE_DPI,
400 struct drm_encoder base;
401 enum vc4_encoder_type type;
405 static inline struct vc4_encoder *
406 to_vc4_encoder(struct drm_encoder *encoder)
408 return container_of(encoder, struct vc4_encoder, base);
411 struct vc4_crtc_data {
412 /* Which channel of the HVS this pixelvalve sources from. */
415 enum vc4_encoder_type encoder_types[4];
419 struct drm_crtc base;
420 const struct vc4_crtc_data *data;
423 /* Timestamp at start of vblank irq - unaffected by lock delays. */
426 /* Which HVS channel we're using for our CRTC. */
432 /* Size in pixels of the COB memory allocated to this CRTC. */
435 struct drm_pending_vblank_event *event;
438 static inline struct vc4_crtc *
439 to_vc4_crtc(struct drm_crtc *crtc)
441 return (struct vc4_crtc *)crtc;
444 #define V3D_READ(offset) readl(vc4->v3d->regs + offset)
445 #define V3D_WRITE(offset, val) writel(val, vc4->v3d->regs + offset)
446 #define HVS_READ(offset) readl(vc4->hvs->regs + offset)
447 #define HVS_WRITE(offset, val) writel(val, vc4->hvs->regs + offset)
449 struct vc4_exec_info {
450 /* Sequence number for this bin/render job. */
453 /* Latest write_seqno of any BO that binning depends on. */
454 uint64_t bin_dep_seqno;
456 struct dma_fence *fence;
458 /* Last current addresses the hardware was processing when the
459 * hangcheck timer checked on us.
461 uint32_t last_ct0ca, last_ct1ca;
463 /* Kernel-space copy of the ioctl arguments */
464 struct drm_vc4_submit_cl *args;
466 /* This is the array of BOs that were looked up at the start of exec.
467 * Command validation will use indices into this array.
469 struct drm_gem_cma_object **bo;
472 /* List of BOs that are being written by the RCL. Other than
473 * the binner temporary storage, this is all the BOs written
476 struct drm_gem_cma_object *rcl_write_bo[4];
477 uint32_t rcl_write_bo_count;
479 /* Pointers for our position in vc4->job_list */
480 struct list_head head;
482 /* List of other BOs used in the job that need to be released
483 * once the job is complete.
485 struct list_head unref_list;
487 /* Current unvalidated indices into @bo loaded by the non-hardware
488 * VC4_PACKET_GEM_HANDLES.
490 uint32_t bo_index[2];
492 /* This is the BO where we store the validated command lists, shader
493 * records, and uniforms.
495 struct drm_gem_cma_object *exec_bo;
498 * This tracks the per-shader-record state (packet 64) that
499 * determines the length of the shader record and the offset
500 * it's expected to be found at. It gets read in from the
503 struct vc4_shader_state {
505 /* Maximum vertex index referenced by any primitive using this
511 /** How many shader states the user declared they were using. */
512 uint32_t shader_state_size;
513 /** How many shader state records the validator has seen. */
514 uint32_t shader_state_count;
516 bool found_tile_binning_mode_config_packet;
517 bool found_start_tile_binning_packet;
518 bool found_increment_semaphore_packet;
520 uint8_t bin_tiles_x, bin_tiles_y;
521 /* Physical address of the start of the tile alloc array
522 * (where each tile's binned CL will start)
524 uint32_t tile_alloc_offset;
525 /* Bitmask of which binner slots are freed when this job completes. */
529 * Computed addresses pointing into exec_bo where we start the
530 * bin thread (ct0) and render thread (ct1).
532 uint32_t ct0ca, ct0ea;
533 uint32_t ct1ca, ct1ea;
535 /* Pointer to the unvalidated bin CL (if present). */
538 /* Pointers to the shader recs. These paddr gets incremented as CL
539 * packets are relocated in validate_gl_shader_state, and the vaddrs
540 * (u and v) get incremented and size decremented as the shader recs
541 * themselves are validated.
545 uint32_t shader_rec_p;
546 uint32_t shader_rec_size;
548 /* Pointers to the uniform data. These pointers are incremented, and
549 * size decremented, as each batch of uniforms is uploaded.
554 uint32_t uniforms_size;
556 /* Pointer to a performance monitor object if the user requested it,
559 struct vc4_perfmon *perfmon;
562 /* Per-open file private data. Any driver-specific resource that has to be
563 * released when the DRM file is closed should be placed here.
572 static inline struct vc4_exec_info *
573 vc4_first_bin_job(struct vc4_dev *vc4)
575 return list_first_entry_or_null(&vc4->bin_job_list,
576 struct vc4_exec_info, head);
579 static inline struct vc4_exec_info *
580 vc4_first_render_job(struct vc4_dev *vc4)
582 return list_first_entry_or_null(&vc4->render_job_list,
583 struct vc4_exec_info, head);
586 static inline struct vc4_exec_info *
587 vc4_last_render_job(struct vc4_dev *vc4)
589 if (list_empty(&vc4->render_job_list))
591 return list_last_entry(&vc4->render_job_list,
592 struct vc4_exec_info, head);
596 * struct vc4_texture_sample_info - saves the offsets into the UBO for texture
599 * This will be used at draw time to relocate the reference to the texture
600 * contents in p0, and validate that the offset combined with
601 * width/height/stride/etc. from p1 and p2/p3 doesn't sample outside the BO.
602 * Note that the hardware treats unprovided config parameters as 0, so not all
603 * of them need to be set up for every texure sample, and we'll store ~0 as
604 * the offset to mark the unused ones.
606 * See the VC4 3D architecture guide page 41 ("Texture and Memory Lookup Unit
607 * Setup") for definitions of the texture parameters.
609 struct vc4_texture_sample_info {
611 uint32_t p_offset[4];
615 * struct vc4_validated_shader_info - information about validated shaders that
616 * needs to be used from command list validation.
618 * For a given shader, each time a shader state record references it, we need
619 * to verify that the shader doesn't read more uniforms than the shader state
620 * record's uniform BO pointer can provide, and we need to apply relocations
621 * and validate the shader state record's uniforms that define the texture
624 struct vc4_validated_shader_info {
625 uint32_t uniforms_size;
626 uint32_t uniforms_src_size;
627 uint32_t num_texture_samples;
628 struct vc4_texture_sample_info *texture_samples;
630 uint32_t num_uniform_addr_offsets;
631 uint32_t *uniform_addr_offsets;
637 * _wait_for - magic (register) wait macro
639 * Does the right thing for modeset paths when run under kdgb or similar atomic
640 * contexts. Note that it's important that we check the condition again after
641 * having timed out, since the timeout could be due to preemption or similar and
642 * we've never had a chance to check the condition before the timeout.
644 #define _wait_for(COND, MS, W) ({ \
645 unsigned long timeout__ = jiffies + msecs_to_jiffies(MS) + 1; \
648 if (time_after(jiffies, timeout__)) { \
650 ret__ = -ETIMEDOUT; \
653 if (W && drm_can_sleep()) { \
662 #define wait_for(COND, MS) _wait_for(COND, MS, 1)
665 struct drm_gem_object *vc4_create_object(struct drm_device *dev, size_t size);
666 void vc4_free_object(struct drm_gem_object *gem_obj);
667 struct vc4_bo *vc4_bo_create(struct drm_device *dev, size_t size,
668 bool from_cache, enum vc4_kernel_bo_type type);
669 int vc4_dumb_create(struct drm_file *file_priv,
670 struct drm_device *dev,
671 struct drm_mode_create_dumb *args);
672 struct dma_buf *vc4_prime_export(struct drm_device *dev,
673 struct drm_gem_object *obj, int flags);
674 int vc4_create_bo_ioctl(struct drm_device *dev, void *data,
675 struct drm_file *file_priv);
676 int vc4_create_shader_bo_ioctl(struct drm_device *dev, void *data,
677 struct drm_file *file_priv);
678 int vc4_mmap_bo_ioctl(struct drm_device *dev, void *data,
679 struct drm_file *file_priv);
680 int vc4_set_tiling_ioctl(struct drm_device *dev, void *data,
681 struct drm_file *file_priv);
682 int vc4_get_tiling_ioctl(struct drm_device *dev, void *data,
683 struct drm_file *file_priv);
684 int vc4_get_hang_state_ioctl(struct drm_device *dev, void *data,
685 struct drm_file *file_priv);
686 int vc4_label_bo_ioctl(struct drm_device *dev, void *data,
687 struct drm_file *file_priv);
688 vm_fault_t vc4_fault(struct vm_fault *vmf);
689 int vc4_mmap(struct file *filp, struct vm_area_struct *vma);
690 int vc4_prime_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma);
691 struct drm_gem_object *vc4_prime_import_sg_table(struct drm_device *dev,
692 struct dma_buf_attachment *attach,
693 struct sg_table *sgt);
694 void *vc4_prime_vmap(struct drm_gem_object *obj);
695 int vc4_bo_cache_init(struct drm_device *dev);
696 void vc4_bo_cache_destroy(struct drm_device *dev);
697 int vc4_bo_stats_debugfs(struct seq_file *m, void *arg);
698 int vc4_bo_inc_usecnt(struct vc4_bo *bo);
699 void vc4_bo_dec_usecnt(struct vc4_bo *bo);
700 void vc4_bo_add_to_purgeable_pool(struct vc4_bo *bo);
701 void vc4_bo_remove_from_purgeable_pool(struct vc4_bo *bo);
704 extern struct platform_driver vc4_crtc_driver;
705 int vc4_crtc_debugfs_regs(struct seq_file *m, void *arg);
706 bool vc4_crtc_get_scanoutpos(struct drm_device *dev, unsigned int crtc_id,
707 bool in_vblank_irq, int *vpos, int *hpos,
708 ktime_t *stime, ktime_t *etime,
709 const struct drm_display_mode *mode);
710 void vc4_crtc_handle_vblank(struct vc4_crtc *crtc);
711 void vc4_crtc_txp_armed(struct drm_crtc_state *state);
712 void vc4_crtc_get_margins(struct drm_crtc_state *state,
713 unsigned int *right, unsigned int *left,
714 unsigned int *top, unsigned int *bottom);
717 int vc4_debugfs_init(struct drm_minor *minor);
720 void __iomem *vc4_ioremap_regs(struct platform_device *dev, int index);
723 extern struct platform_driver vc4_dpi_driver;
724 int vc4_dpi_debugfs_regs(struct seq_file *m, void *unused);
727 extern struct platform_driver vc4_dsi_driver;
728 int vc4_dsi_debugfs_regs(struct seq_file *m, void *unused);
731 extern const struct dma_fence_ops vc4_fence_ops;
734 void vc4_gem_init(struct drm_device *dev);
735 void vc4_gem_destroy(struct drm_device *dev);
736 int vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
737 struct drm_file *file_priv);
738 int vc4_wait_seqno_ioctl(struct drm_device *dev, void *data,
739 struct drm_file *file_priv);
740 int vc4_wait_bo_ioctl(struct drm_device *dev, void *data,
741 struct drm_file *file_priv);
742 void vc4_submit_next_bin_job(struct drm_device *dev);
743 void vc4_submit_next_render_job(struct drm_device *dev);
744 void vc4_move_job_to_render(struct drm_device *dev, struct vc4_exec_info *exec);
745 int vc4_wait_for_seqno(struct drm_device *dev, uint64_t seqno,
746 uint64_t timeout_ns, bool interruptible);
747 void vc4_job_handle_completed(struct vc4_dev *vc4);
748 int vc4_queue_seqno_cb(struct drm_device *dev,
749 struct vc4_seqno_cb *cb, uint64_t seqno,
750 void (*func)(struct vc4_seqno_cb *cb));
751 int vc4_gem_madvise_ioctl(struct drm_device *dev, void *data,
752 struct drm_file *file_priv);
755 extern struct platform_driver vc4_hdmi_driver;
756 int vc4_hdmi_debugfs_regs(struct seq_file *m, void *unused);
759 extern struct platform_driver vc4_vec_driver;
760 int vc4_vec_debugfs_regs(struct seq_file *m, void *unused);
763 extern struct platform_driver vc4_txp_driver;
764 int vc4_txp_debugfs_regs(struct seq_file *m, void *unused);
767 irqreturn_t vc4_irq(int irq, void *arg);
768 void vc4_irq_preinstall(struct drm_device *dev);
769 int vc4_irq_postinstall(struct drm_device *dev);
770 void vc4_irq_uninstall(struct drm_device *dev);
771 void vc4_irq_reset(struct drm_device *dev);
774 extern struct platform_driver vc4_hvs_driver;
775 void vc4_hvs_dump_state(struct drm_device *dev);
776 int vc4_hvs_debugfs_regs(struct seq_file *m, void *unused);
777 int vc4_hvs_debugfs_underrun(struct seq_file *m, void *unused);
778 void vc4_hvs_unmask_underrun(struct drm_device *dev, int channel);
779 void vc4_hvs_mask_underrun(struct drm_device *dev, int channel);
782 int vc4_kms_load(struct drm_device *dev);
785 struct drm_plane *vc4_plane_init(struct drm_device *dev,
786 enum drm_plane_type type);
787 u32 vc4_plane_write_dlist(struct drm_plane *plane, u32 __iomem *dlist);
788 u32 vc4_plane_dlist_size(const struct drm_plane_state *state);
789 void vc4_plane_async_set_fb(struct drm_plane *plane,
790 struct drm_framebuffer *fb);
793 extern struct platform_driver vc4_v3d_driver;
794 int vc4_v3d_debugfs_ident(struct seq_file *m, void *unused);
795 int vc4_v3d_debugfs_regs(struct seq_file *m, void *unused);
796 int vc4_v3d_get_bin_slot(struct vc4_dev *vc4);
800 vc4_validate_bin_cl(struct drm_device *dev,
803 struct vc4_exec_info *exec);
806 vc4_validate_shader_recs(struct drm_device *dev, struct vc4_exec_info *exec);
808 struct drm_gem_cma_object *vc4_use_bo(struct vc4_exec_info *exec,
811 int vc4_get_rcl(struct drm_device *dev, struct vc4_exec_info *exec);
813 bool vc4_check_tex_size(struct vc4_exec_info *exec,
814 struct drm_gem_cma_object *fbo,
815 uint32_t offset, uint8_t tiling_format,
816 uint32_t width, uint32_t height, uint8_t cpp);
818 /* vc4_validate_shader.c */
819 struct vc4_validated_shader_info *
820 vc4_validate_shader(struct drm_gem_cma_object *shader_obj);
823 void vc4_perfmon_get(struct vc4_perfmon *perfmon);
824 void vc4_perfmon_put(struct vc4_perfmon *perfmon);
825 void vc4_perfmon_start(struct vc4_dev *vc4, struct vc4_perfmon *perfmon);
826 void vc4_perfmon_stop(struct vc4_dev *vc4, struct vc4_perfmon *perfmon,
828 struct vc4_perfmon *vc4_perfmon_find(struct vc4_file *vc4file, int id);
829 void vc4_perfmon_open_file(struct vc4_file *vc4file);
830 void vc4_perfmon_close_file(struct vc4_file *vc4file);
831 int vc4_perfmon_create_ioctl(struct drm_device *dev, void *data,
832 struct drm_file *file_priv);
833 int vc4_perfmon_destroy_ioctl(struct drm_device *dev, void *data,
834 struct drm_file *file_priv);
835 int vc4_perfmon_get_values_ioctl(struct drm_device *dev, void *data,
836 struct drm_file *file_priv);