xref: /openbmc/linux/drivers/gpu/drm/msm/msm_gpu.h (revision 11c416e3)
1 /* SPDX-License-Identifier: GPL-2.0-only */
2 /*
3  * Copyright (C) 2013 Red Hat
4  * Author: Rob Clark <robdclark@gmail.com>
5  */
6 
7 #ifndef __MSM_GPU_H__
8 #define __MSM_GPU_H__
9 
10 #include <linux/clk.h>
11 #include <linux/interconnect.h>
12 #include <linux/regulator/consumer.h>
13 
14 #include "msm_drv.h"
15 #include "msm_fence.h"
16 #include "msm_ringbuffer.h"
17 
18 struct msm_gem_submit;
19 struct msm_gpu_perfcntr;
20 struct msm_gpu_state;
21 
22 struct msm_gpu_config {
23 	const char *ioname;
24 	unsigned int nr_rings;
25 };
26 
27 /* So far, with hardware that I've seen to date, we can have:
28  *  + zero, one, or two z180 2d cores
29  *  + a3xx or a2xx 3d core, which share a common CP (the firmware
30  *    for the CP seems to implement some different PM4 packet types
31  *    but the basics of cmdstream submission are the same)
32  *
33  * Which means that the eventual complete "class" hierarchy, once
34  * support for all past and present hw is in place, becomes:
35  *  + msm_gpu
36  *    + adreno_gpu
37  *      + a3xx_gpu
38  *      + a2xx_gpu
39  *    + z180_gpu
40  */
41 struct msm_gpu_funcs {
42 	int (*get_param)(struct msm_gpu *gpu, uint32_t param, uint64_t *value);
43 	int (*hw_init)(struct msm_gpu *gpu);
44 	int (*pm_suspend)(struct msm_gpu *gpu);
45 	int (*pm_resume)(struct msm_gpu *gpu);
46 	void (*submit)(struct msm_gpu *gpu, struct msm_gem_submit *submit,
47 			struct msm_file_private *ctx);
48 	void (*flush)(struct msm_gpu *gpu, struct msm_ringbuffer *ring);
49 	irqreturn_t (*irq)(struct msm_gpu *irq);
50 	struct msm_ringbuffer *(*active_ring)(struct msm_gpu *gpu);
51 	void (*recover)(struct msm_gpu *gpu);
52 	void (*destroy)(struct msm_gpu *gpu);
53 #if defined(CONFIG_DEBUG_FS) || defined(CONFIG_DEV_COREDUMP)
54 	/* show GPU status in debugfs: */
55 	void (*show)(struct msm_gpu *gpu, struct msm_gpu_state *state,
56 			struct drm_printer *p);
57 	/* for generation specific debugfs: */
58 	void (*debugfs_init)(struct msm_gpu *gpu, struct drm_minor *minor);
59 #endif
60 	unsigned long (*gpu_busy)(struct msm_gpu *gpu);
61 	struct msm_gpu_state *(*gpu_state_get)(struct msm_gpu *gpu);
62 	int (*gpu_state_put)(struct msm_gpu_state *state);
63 	unsigned long (*gpu_get_freq)(struct msm_gpu *gpu);
64 	void (*gpu_set_freq)(struct msm_gpu *gpu, unsigned long freq);
65 	struct msm_gem_address_space *(*create_address_space)
66 		(struct msm_gpu *gpu, struct platform_device *pdev);
67 };
68 
69 struct msm_gpu {
70 	const char *name;
71 	struct drm_device *dev;
72 	struct platform_device *pdev;
73 	const struct msm_gpu_funcs *funcs;
74 
75 	/* performance counters (hw & sw): */
76 	spinlock_t perf_lock;
77 	bool perfcntr_active;
78 	struct {
79 		bool active;
80 		ktime_t time;
81 	} last_sample;
82 	uint32_t totaltime, activetime;    /* sw counters */
83 	uint32_t last_cntrs[5];            /* hw counters */
84 	const struct msm_gpu_perfcntr *perfcntrs;
85 	uint32_t num_perfcntrs;
86 
87 	struct msm_ringbuffer *rb[MSM_GPU_MAX_RINGS];
88 	int nr_rings;
89 
90 	/* list of GEM active objects: */
91 	struct list_head active_list;
92 
93 	/* does gpu need hw_init? */
94 	bool needs_hw_init;
95 
96 	/* number of GPU hangs (for all contexts) */
97 	int global_faults;
98 
99 	/* worker for handling active-list retiring: */
100 	struct work_struct retire_work;
101 
102 	void __iomem *mmio;
103 	int irq;
104 
105 	struct msm_gem_address_space *aspace;
106 
107 	/* Power Control: */
108 	struct regulator *gpu_reg, *gpu_cx;
109 	struct clk_bulk_data *grp_clks;
110 	int nr_clocks;
111 	struct clk *ebi1_clk, *core_clk, *rbbmtimer_clk;
112 	uint32_t fast_rate;
113 
114 	/* The gfx-mem interconnect path that's used by all GPU types. */
115 	struct icc_path *icc_path;
116 
117 	/*
118 	 * Second interconnect path for some A3xx and all A4xx GPUs to the
119 	 * On Chip MEMory (OCMEM).
120 	 */
121 	struct icc_path *ocmem_icc_path;
122 
123 	/* Hang and Inactivity Detection:
124 	 */
125 #define DRM_MSM_INACTIVE_PERIOD   66 /* in ms (roughly four frames) */
126 
127 #define DRM_MSM_HANGCHECK_PERIOD 500 /* in ms */
128 #define DRM_MSM_HANGCHECK_JIFFIES msecs_to_jiffies(DRM_MSM_HANGCHECK_PERIOD)
129 	struct timer_list hangcheck_timer;
130 	struct work_struct recover_work;
131 
132 	struct drm_gem_object *memptrs_bo;
133 
134 	struct {
135 		struct devfreq *devfreq;
136 		u64 busy_cycles;
137 		ktime_t time;
138 	} devfreq;
139 
140 	struct msm_gpu_state *crashstate;
141 };
142 
143 /* It turns out that all targets use the same ringbuffer size */
144 #define MSM_GPU_RINGBUFFER_SZ SZ_32K
145 #define MSM_GPU_RINGBUFFER_BLKSIZE 32
146 
147 #define MSM_GPU_RB_CNTL_DEFAULT \
148 		(AXXX_CP_RB_CNTL_BUFSZ(ilog2(MSM_GPU_RINGBUFFER_SZ / 8)) | \
149 		AXXX_CP_RB_CNTL_BLKSZ(ilog2(MSM_GPU_RINGBUFFER_BLKSIZE / 8)))
150 
151 static inline bool msm_gpu_active(struct msm_gpu *gpu)
152 {
153 	int i;
154 
155 	for (i = 0; i < gpu->nr_rings; i++) {
156 		struct msm_ringbuffer *ring = gpu->rb[i];
157 
158 		if (ring->seqno > ring->memptrs->fence)
159 			return true;
160 	}
161 
162 	return false;
163 }
164 
165 /* Perf-Counters:
166  * The select_reg and select_val are just there for the benefit of the child
167  * class that actually enables the perf counter..  but msm_gpu base class
168  * will handle sampling/displaying the counters.
169  */
170 
171 struct msm_gpu_perfcntr {
172 	uint32_t select_reg;
173 	uint32_t sample_reg;
174 	uint32_t select_val;
175 	const char *name;
176 };
177 
178 struct msm_gpu_submitqueue {
179 	int id;
180 	u32 flags;
181 	u32 prio;
182 	int faults;
183 	struct list_head node;
184 	struct kref ref;
185 };
186 
187 struct msm_gpu_state_bo {
188 	u64 iova;
189 	size_t size;
190 	void *data;
191 	bool encoded;
192 };
193 
194 struct msm_gpu_state {
195 	struct kref ref;
196 	struct timespec64 time;
197 
198 	struct {
199 		u64 iova;
200 		u32 fence;
201 		u32 seqno;
202 		u32 rptr;
203 		u32 wptr;
204 		void *data;
205 		int data_size;
206 		bool encoded;
207 	} ring[MSM_GPU_MAX_RINGS];
208 
209 	int nr_registers;
210 	u32 *registers;
211 
212 	u32 rbbm_status;
213 
214 	char *comm;
215 	char *cmd;
216 
217 	int nr_bos;
218 	struct msm_gpu_state_bo *bos;
219 };
220 
221 static inline void gpu_write(struct msm_gpu *gpu, u32 reg, u32 data)
222 {
223 	msm_writel(data, gpu->mmio + (reg << 2));
224 }
225 
226 static inline u32 gpu_read(struct msm_gpu *gpu, u32 reg)
227 {
228 	return msm_readl(gpu->mmio + (reg << 2));
229 }
230 
231 static inline void gpu_rmw(struct msm_gpu *gpu, u32 reg, u32 mask, u32 or)
232 {
233 	uint32_t val = gpu_read(gpu, reg);
234 
235 	val &= ~mask;
236 	gpu_write(gpu, reg, val | or);
237 }
238 
239 static inline u64 gpu_read64(struct msm_gpu *gpu, u32 lo, u32 hi)
240 {
241 	u64 val;
242 
243 	/*
244 	 * Why not a readq here? Two reasons: 1) many of the LO registers are
245 	 * not quad word aligned and 2) the GPU hardware designers have a bit
246 	 * of a history of putting registers where they fit, especially in
247 	 * spins. The longer a GPU family goes the higher the chance that
248 	 * we'll get burned.  We could do a series of validity checks if we
249 	 * wanted to, but really is a readq() that much better? Nah.
250 	 */
251 
252 	/*
253 	 * For some lo/hi registers (like perfcounters), the hi value is latched
254 	 * when the lo is read, so make sure to read the lo first to trigger
255 	 * that
256 	 */
257 	val = (u64) msm_readl(gpu->mmio + (lo << 2));
258 	val |= ((u64) msm_readl(gpu->mmio + (hi << 2)) << 32);
259 
260 	return val;
261 }
262 
263 static inline void gpu_write64(struct msm_gpu *gpu, u32 lo, u32 hi, u64 val)
264 {
265 	/* Why not a writeq here? Read the screed above */
266 	msm_writel(lower_32_bits(val), gpu->mmio + (lo << 2));
267 	msm_writel(upper_32_bits(val), gpu->mmio + (hi << 2));
268 }
269 
270 int msm_gpu_pm_suspend(struct msm_gpu *gpu);
271 int msm_gpu_pm_resume(struct msm_gpu *gpu);
272 void msm_gpu_resume_devfreq(struct msm_gpu *gpu);
273 
274 int msm_gpu_hw_init(struct msm_gpu *gpu);
275 
276 void msm_gpu_perfcntr_start(struct msm_gpu *gpu);
277 void msm_gpu_perfcntr_stop(struct msm_gpu *gpu);
278 int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime,
279 		uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs);
280 
281 void msm_gpu_retire(struct msm_gpu *gpu);
282 void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit,
283 		struct msm_file_private *ctx);
284 
285 int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
286 		struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
287 		const char *name, struct msm_gpu_config *config);
288 
289 void msm_gpu_cleanup(struct msm_gpu *gpu);
290 
291 struct msm_gpu *adreno_load_gpu(struct drm_device *dev);
292 void __init adreno_register(void);
293 void __exit adreno_unregister(void);
294 
295 static inline void msm_submitqueue_put(struct msm_gpu_submitqueue *queue)
296 {
297 	if (queue)
298 		kref_put(&queue->ref, msm_submitqueue_destroy);
299 }
300 
301 static inline struct msm_gpu_state *msm_gpu_crashstate_get(struct msm_gpu *gpu)
302 {
303 	struct msm_gpu_state *state = NULL;
304 
305 	mutex_lock(&gpu->dev->struct_mutex);
306 
307 	if (gpu->crashstate) {
308 		kref_get(&gpu->crashstate->ref);
309 		state = gpu->crashstate;
310 	}
311 
312 	mutex_unlock(&gpu->dev->struct_mutex);
313 
314 	return state;
315 }
316 
317 static inline void msm_gpu_crashstate_put(struct msm_gpu *gpu)
318 {
319 	mutex_lock(&gpu->dev->struct_mutex);
320 
321 	if (gpu->crashstate) {
322 		if (gpu->funcs->gpu_state_put(gpu->crashstate))
323 			gpu->crashstate = NULL;
324 	}
325 
326 	mutex_unlock(&gpu->dev->struct_mutex);
327 }
328 
329 #endif /* __MSM_GPU_H__ */
330