xref: /openbmc/linux/drivers/gpu/drm/msm/msm_gpu.h (revision e2028c8e)
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/pm_opp.h>
13 #include <linux/regulator/consumer.h>
14 
15 #include "msm_drv.h"
16 #include "msm_fence.h"
17 #include "msm_ringbuffer.h"
18 
19 struct msm_gem_submit;
20 struct msm_gpu_perfcntr;
21 struct msm_gpu_state;
22 
23 struct msm_gpu_config {
24 	const char *ioname;
25 	unsigned int nr_rings;
26 };
27 
28 /* So far, with hardware that I've seen to date, we can have:
29  *  + zero, one, or two z180 2d cores
30  *  + a3xx or a2xx 3d core, which share a common CP (the firmware
31  *    for the CP seems to implement some different PM4 packet types
32  *    but the basics of cmdstream submission are the same)
33  *
34  * Which means that the eventual complete "class" hierarchy, once
35  * support for all past and present hw is in place, becomes:
36  *  + msm_gpu
37  *    + adreno_gpu
38  *      + a3xx_gpu
39  *      + a2xx_gpu
40  *    + z180_gpu
41  */
42 struct msm_gpu_funcs {
43 	int (*get_param)(struct msm_gpu *gpu, uint32_t param, uint64_t *value);
44 	int (*hw_init)(struct msm_gpu *gpu);
45 	int (*pm_suspend)(struct msm_gpu *gpu);
46 	int (*pm_resume)(struct msm_gpu *gpu);
47 	void (*submit)(struct msm_gpu *gpu, struct msm_gem_submit *submit,
48 			struct msm_file_private *ctx);
49 	void (*flush)(struct msm_gpu *gpu, struct msm_ringbuffer *ring);
50 	irqreturn_t (*irq)(struct msm_gpu *irq);
51 	struct msm_ringbuffer *(*active_ring)(struct msm_gpu *gpu);
52 	void (*recover)(struct msm_gpu *gpu);
53 	void (*destroy)(struct msm_gpu *gpu);
54 #if defined(CONFIG_DEBUG_FS) || defined(CONFIG_DEV_COREDUMP)
55 	/* show GPU status in debugfs: */
56 	void (*show)(struct msm_gpu *gpu, struct msm_gpu_state *state,
57 			struct drm_printer *p);
58 	/* for generation specific debugfs: */
59 	void (*debugfs_init)(struct msm_gpu *gpu, struct drm_minor *minor);
60 #endif
61 	unsigned long (*gpu_busy)(struct msm_gpu *gpu);
62 	struct msm_gpu_state *(*gpu_state_get)(struct msm_gpu *gpu);
63 	int (*gpu_state_put)(struct msm_gpu_state *state);
64 	unsigned long (*gpu_get_freq)(struct msm_gpu *gpu);
65 	void (*gpu_set_freq)(struct msm_gpu *gpu, struct dev_pm_opp *opp);
66 	struct msm_gem_address_space *(*create_address_space)
67 		(struct msm_gpu *gpu, struct platform_device *pdev);
68 };
69 
70 struct msm_gpu {
71 	const char *name;
72 	struct drm_device *dev;
73 	struct platform_device *pdev;
74 	const struct msm_gpu_funcs *funcs;
75 
76 	/* performance counters (hw & sw): */
77 	spinlock_t perf_lock;
78 	bool perfcntr_active;
79 	struct {
80 		bool active;
81 		ktime_t time;
82 	} last_sample;
83 	uint32_t totaltime, activetime;    /* sw counters */
84 	uint32_t last_cntrs[5];            /* hw counters */
85 	const struct msm_gpu_perfcntr *perfcntrs;
86 	uint32_t num_perfcntrs;
87 
88 	struct msm_ringbuffer *rb[MSM_GPU_MAX_RINGS];
89 	int nr_rings;
90 
91 	/* list of GEM active objects: */
92 	struct list_head active_list;
93 
94 	/* does gpu need hw_init? */
95 	bool needs_hw_init;
96 
97 	/* number of GPU hangs (for all contexts) */
98 	int global_faults;
99 
100 	/* worker for handling active-list retiring: */
101 	struct work_struct retire_work;
102 
103 	void __iomem *mmio;
104 	int irq;
105 
106 	struct msm_gem_address_space *aspace;
107 
108 	/* Power Control: */
109 	struct regulator *gpu_reg, *gpu_cx;
110 	struct clk_bulk_data *grp_clks;
111 	int nr_clocks;
112 	struct clk *ebi1_clk, *core_clk, *rbbmtimer_clk;
113 	uint32_t fast_rate;
114 
115 	/* The gfx-mem interconnect path that's used by all GPU types. */
116 	struct icc_path *icc_path;
117 
118 	/*
119 	 * Second interconnect path for some A3xx and all A4xx GPUs to the
120 	 * On Chip MEMory (OCMEM).
121 	 */
122 	struct icc_path *ocmem_icc_path;
123 
124 	/* Hang and Inactivity Detection:
125 	 */
126 #define DRM_MSM_INACTIVE_PERIOD   66 /* in ms (roughly four frames) */
127 
128 #define DRM_MSM_HANGCHECK_PERIOD 500 /* in ms */
129 #define DRM_MSM_HANGCHECK_JIFFIES msecs_to_jiffies(DRM_MSM_HANGCHECK_PERIOD)
130 	struct timer_list hangcheck_timer;
131 	struct work_struct recover_work;
132 
133 	struct drm_gem_object *memptrs_bo;
134 
135 	struct {
136 		struct devfreq *devfreq;
137 		u64 busy_cycles;
138 		ktime_t time;
139 	} devfreq;
140 
141 	struct msm_gpu_state *crashstate;
142 };
143 
144 /* It turns out that all targets use the same ringbuffer size */
145 #define MSM_GPU_RINGBUFFER_SZ SZ_32K
146 #define MSM_GPU_RINGBUFFER_BLKSIZE 32
147 
148 #define MSM_GPU_RB_CNTL_DEFAULT \
149 		(AXXX_CP_RB_CNTL_BUFSZ(ilog2(MSM_GPU_RINGBUFFER_SZ / 8)) | \
150 		AXXX_CP_RB_CNTL_BLKSZ(ilog2(MSM_GPU_RINGBUFFER_BLKSIZE / 8)))
151 
152 static inline bool msm_gpu_active(struct msm_gpu *gpu)
153 {
154 	int i;
155 
156 	for (i = 0; i < gpu->nr_rings; i++) {
157 		struct msm_ringbuffer *ring = gpu->rb[i];
158 
159 		if (ring->seqno > ring->memptrs->fence)
160 			return true;
161 	}
162 
163 	return false;
164 }
165 
166 /* Perf-Counters:
167  * The select_reg and select_val are just there for the benefit of the child
168  * class that actually enables the perf counter..  but msm_gpu base class
169  * will handle sampling/displaying the counters.
170  */
171 
172 struct msm_gpu_perfcntr {
173 	uint32_t select_reg;
174 	uint32_t sample_reg;
175 	uint32_t select_val;
176 	const char *name;
177 };
178 
179 struct msm_gpu_submitqueue {
180 	int id;
181 	u32 flags;
182 	u32 prio;
183 	int faults;
184 	struct list_head node;
185 	struct kref ref;
186 };
187 
188 struct msm_gpu_state_bo {
189 	u64 iova;
190 	size_t size;
191 	void *data;
192 	bool encoded;
193 };
194 
195 struct msm_gpu_state {
196 	struct kref ref;
197 	struct timespec64 time;
198 
199 	struct {
200 		u64 iova;
201 		u32 fence;
202 		u32 seqno;
203 		u32 rptr;
204 		u32 wptr;
205 		void *data;
206 		int data_size;
207 		bool encoded;
208 	} ring[MSM_GPU_MAX_RINGS];
209 
210 	int nr_registers;
211 	u32 *registers;
212 
213 	u32 rbbm_status;
214 
215 	char *comm;
216 	char *cmd;
217 
218 	int nr_bos;
219 	struct msm_gpu_state_bo *bos;
220 };
221 
222 static inline void gpu_write(struct msm_gpu *gpu, u32 reg, u32 data)
223 {
224 	msm_writel(data, gpu->mmio + (reg << 2));
225 }
226 
227 static inline u32 gpu_read(struct msm_gpu *gpu, u32 reg)
228 {
229 	return msm_readl(gpu->mmio + (reg << 2));
230 }
231 
232 static inline void gpu_rmw(struct msm_gpu *gpu, u32 reg, u32 mask, u32 or)
233 {
234 	uint32_t val = gpu_read(gpu, reg);
235 
236 	val &= ~mask;
237 	gpu_write(gpu, reg, val | or);
238 }
239 
240 static inline u64 gpu_read64(struct msm_gpu *gpu, u32 lo, u32 hi)
241 {
242 	u64 val;
243 
244 	/*
245 	 * Why not a readq here? Two reasons: 1) many of the LO registers are
246 	 * not quad word aligned and 2) the GPU hardware designers have a bit
247 	 * of a history of putting registers where they fit, especially in
248 	 * spins. The longer a GPU family goes the higher the chance that
249 	 * we'll get burned.  We could do a series of validity checks if we
250 	 * wanted to, but really is a readq() that much better? Nah.
251 	 */
252 
253 	/*
254 	 * For some lo/hi registers (like perfcounters), the hi value is latched
255 	 * when the lo is read, so make sure to read the lo first to trigger
256 	 * that
257 	 */
258 	val = (u64) msm_readl(gpu->mmio + (lo << 2));
259 	val |= ((u64) msm_readl(gpu->mmio + (hi << 2)) << 32);
260 
261 	return val;
262 }
263 
264 static inline void gpu_write64(struct msm_gpu *gpu, u32 lo, u32 hi, u64 val)
265 {
266 	/* Why not a writeq here? Read the screed above */
267 	msm_writel(lower_32_bits(val), gpu->mmio + (lo << 2));
268 	msm_writel(upper_32_bits(val), gpu->mmio + (hi << 2));
269 }
270 
271 int msm_gpu_pm_suspend(struct msm_gpu *gpu);
272 int msm_gpu_pm_resume(struct msm_gpu *gpu);
273 void msm_gpu_resume_devfreq(struct msm_gpu *gpu);
274 
275 int msm_gpu_hw_init(struct msm_gpu *gpu);
276 
277 void msm_gpu_perfcntr_start(struct msm_gpu *gpu);
278 void msm_gpu_perfcntr_stop(struct msm_gpu *gpu);
279 int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime,
280 		uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs);
281 
282 void msm_gpu_retire(struct msm_gpu *gpu);
283 void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit,
284 		struct msm_file_private *ctx);
285 
286 int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
287 		struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
288 		const char *name, struct msm_gpu_config *config);
289 
290 void msm_gpu_cleanup(struct msm_gpu *gpu);
291 
292 struct msm_gpu *adreno_load_gpu(struct drm_device *dev);
293 void __init adreno_register(void);
294 void __exit adreno_unregister(void);
295 
296 static inline void msm_submitqueue_put(struct msm_gpu_submitqueue *queue)
297 {
298 	if (queue)
299 		kref_put(&queue->ref, msm_submitqueue_destroy);
300 }
301 
302 static inline struct msm_gpu_state *msm_gpu_crashstate_get(struct msm_gpu *gpu)
303 {
304 	struct msm_gpu_state *state = NULL;
305 
306 	mutex_lock(&gpu->dev->struct_mutex);
307 
308 	if (gpu->crashstate) {
309 		kref_get(&gpu->crashstate->ref);
310 		state = gpu->crashstate;
311 	}
312 
313 	mutex_unlock(&gpu->dev->struct_mutex);
314 
315 	return state;
316 }
317 
318 static inline void msm_gpu_crashstate_put(struct msm_gpu *gpu)
319 {
320 	mutex_lock(&gpu->dev->struct_mutex);
321 
322 	if (gpu->crashstate) {
323 		if (gpu->funcs->gpu_state_put(gpu->crashstate))
324 			gpu->crashstate = NULL;
325 	}
326 
327 	mutex_unlock(&gpu->dev->struct_mutex);
328 }
329 
330 #endif /* __MSM_GPU_H__ */
331