xref: /openbmc/linux/drivers/gpu/drm/msm/msm_gpu.h (revision 0bf49ffb)
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/adreno-smmu-priv.h>
11 #include <linux/clk.h>
12 #include <linux/interconnect.h>
13 #include <linux/pm_opp.h>
14 #include <linux/regulator/consumer.h>
15 
16 #include "msm_drv.h"
17 #include "msm_fence.h"
18 #include "msm_ringbuffer.h"
19 #include "msm_gem.h"
20 
21 struct msm_gem_submit;
22 struct msm_gpu_perfcntr;
23 struct msm_gpu_state;
24 
25 struct msm_gpu_config {
26 	const char *ioname;
27 	unsigned int nr_rings;
28 };
29 
30 /* So far, with hardware that I've seen to date, we can have:
31  *  + zero, one, or two z180 2d cores
32  *  + a3xx or a2xx 3d core, which share a common CP (the firmware
33  *    for the CP seems to implement some different PM4 packet types
34  *    but the basics of cmdstream submission are the same)
35  *
36  * Which means that the eventual complete "class" hierarchy, once
37  * support for all past and present hw is in place, becomes:
38  *  + msm_gpu
39  *    + adreno_gpu
40  *      + a3xx_gpu
41  *      + a2xx_gpu
42  *    + z180_gpu
43  */
44 struct msm_gpu_funcs {
45 	int (*get_param)(struct msm_gpu *gpu, uint32_t param, uint64_t *value);
46 	int (*hw_init)(struct msm_gpu *gpu);
47 	int (*pm_suspend)(struct msm_gpu *gpu);
48 	int (*pm_resume)(struct msm_gpu *gpu);
49 	void (*submit)(struct msm_gpu *gpu, struct msm_gem_submit *submit);
50 	void (*flush)(struct msm_gpu *gpu, struct msm_ringbuffer *ring);
51 	irqreturn_t (*irq)(struct msm_gpu *irq);
52 	struct msm_ringbuffer *(*active_ring)(struct msm_gpu *gpu);
53 	void (*recover)(struct msm_gpu *gpu);
54 	void (*destroy)(struct msm_gpu *gpu);
55 #if defined(CONFIG_DEBUG_FS) || defined(CONFIG_DEV_COREDUMP)
56 	/* show GPU status in debugfs: */
57 	void (*show)(struct msm_gpu *gpu, struct msm_gpu_state *state,
58 			struct drm_printer *p);
59 	/* for generation specific debugfs: */
60 	void (*debugfs_init)(struct msm_gpu *gpu, struct drm_minor *minor);
61 #endif
62 	unsigned long (*gpu_busy)(struct msm_gpu *gpu);
63 	struct msm_gpu_state *(*gpu_state_get)(struct msm_gpu *gpu);
64 	int (*gpu_state_put)(struct msm_gpu_state *state);
65 	unsigned long (*gpu_get_freq)(struct msm_gpu *gpu);
66 	void (*gpu_set_freq)(struct msm_gpu *gpu, struct dev_pm_opp *opp);
67 	struct msm_gem_address_space *(*create_address_space)
68 		(struct msm_gpu *gpu, struct platform_device *pdev);
69 	struct msm_gem_address_space *(*create_private_address_space)
70 		(struct msm_gpu *gpu);
71 	uint32_t (*get_rptr)(struct msm_gpu *gpu, struct msm_ringbuffer *ring);
72 };
73 
74 struct msm_gpu {
75 	const char *name;
76 	struct drm_device *dev;
77 	struct platform_device *pdev;
78 	const struct msm_gpu_funcs *funcs;
79 
80 	struct adreno_smmu_priv adreno_smmu;
81 
82 	/* performance counters (hw & sw): */
83 	spinlock_t perf_lock;
84 	bool perfcntr_active;
85 	struct {
86 		bool active;
87 		ktime_t time;
88 	} last_sample;
89 	uint32_t totaltime, activetime;    /* sw counters */
90 	uint32_t last_cntrs[5];            /* hw counters */
91 	const struct msm_gpu_perfcntr *perfcntrs;
92 	uint32_t num_perfcntrs;
93 
94 	struct msm_ringbuffer *rb[MSM_GPU_MAX_RINGS];
95 	int nr_rings;
96 
97 	/*
98 	 * List of GEM active objects on this gpu.  Protected by
99 	 * msm_drm_private::mm_lock
100 	 */
101 	struct list_head active_list;
102 
103 	/* does gpu need hw_init? */
104 	bool needs_hw_init;
105 
106 	/* number of GPU hangs (for all contexts) */
107 	int global_faults;
108 
109 	void __iomem *mmio;
110 	int irq;
111 
112 	struct msm_gem_address_space *aspace;
113 
114 	/* Power Control: */
115 	struct regulator *gpu_reg, *gpu_cx;
116 	struct clk_bulk_data *grp_clks;
117 	int nr_clocks;
118 	struct clk *ebi1_clk, *core_clk, *rbbmtimer_clk;
119 	uint32_t fast_rate;
120 
121 	/* The gfx-mem interconnect path that's used by all GPU types. */
122 	struct icc_path *icc_path;
123 
124 	/*
125 	 * Second interconnect path for some A3xx and all A4xx GPUs to the
126 	 * On Chip MEMory (OCMEM).
127 	 */
128 	struct icc_path *ocmem_icc_path;
129 
130 	/* Hang and Inactivity Detection:
131 	 */
132 #define DRM_MSM_INACTIVE_PERIOD   66 /* in ms (roughly four frames) */
133 
134 #define DRM_MSM_HANGCHECK_PERIOD 500 /* in ms */
135 #define DRM_MSM_HANGCHECK_JIFFIES msecs_to_jiffies(DRM_MSM_HANGCHECK_PERIOD)
136 	struct timer_list hangcheck_timer;
137 
138 	/* work for handling GPU recovery: */
139 	struct kthread_work recover_work;
140 
141 	/* work for handling active-list retiring: */
142 	struct kthread_work retire_work;
143 
144 	/* worker for retire/recover: */
145 	struct kthread_worker *worker;
146 
147 	struct drm_gem_object *memptrs_bo;
148 
149 	struct {
150 		struct devfreq *devfreq;
151 		u64 busy_cycles;
152 		ktime_t time;
153 	} devfreq;
154 
155 	struct msm_gpu_state *crashstate;
156 	/* True if the hardware supports expanded apriv (a650 and newer) */
157 	bool hw_apriv;
158 
159 	struct thermal_cooling_device *cooling;
160 };
161 
162 static inline struct msm_gpu *dev_to_gpu(struct device *dev)
163 {
164 	struct adreno_smmu_priv *adreno_smmu = dev_get_drvdata(dev);
165 	return container_of(adreno_smmu, struct msm_gpu, adreno_smmu);
166 }
167 
168 /* It turns out that all targets use the same ringbuffer size */
169 #define MSM_GPU_RINGBUFFER_SZ SZ_32K
170 #define MSM_GPU_RINGBUFFER_BLKSIZE 32
171 
172 #define MSM_GPU_RB_CNTL_DEFAULT \
173 		(AXXX_CP_RB_CNTL_BUFSZ(ilog2(MSM_GPU_RINGBUFFER_SZ / 8)) | \
174 		AXXX_CP_RB_CNTL_BLKSZ(ilog2(MSM_GPU_RINGBUFFER_BLKSIZE / 8)))
175 
176 static inline bool msm_gpu_active(struct msm_gpu *gpu)
177 {
178 	int i;
179 
180 	for (i = 0; i < gpu->nr_rings; i++) {
181 		struct msm_ringbuffer *ring = gpu->rb[i];
182 
183 		if (ring->seqno > ring->memptrs->fence)
184 			return true;
185 	}
186 
187 	return false;
188 }
189 
190 /* Perf-Counters:
191  * The select_reg and select_val are just there for the benefit of the child
192  * class that actually enables the perf counter..  but msm_gpu base class
193  * will handle sampling/displaying the counters.
194  */
195 
196 struct msm_gpu_perfcntr {
197 	uint32_t select_reg;
198 	uint32_t sample_reg;
199 	uint32_t select_val;
200 	const char *name;
201 };
202 
203 struct msm_gpu_submitqueue {
204 	int id;
205 	u32 flags;
206 	u32 prio;
207 	int faults;
208 	struct msm_file_private *ctx;
209 	struct list_head node;
210 	struct kref ref;
211 };
212 
213 struct msm_gpu_state_bo {
214 	u64 iova;
215 	size_t size;
216 	void *data;
217 	bool encoded;
218 };
219 
220 struct msm_gpu_state {
221 	struct kref ref;
222 	struct timespec64 time;
223 
224 	struct {
225 		u64 iova;
226 		u32 fence;
227 		u32 seqno;
228 		u32 rptr;
229 		u32 wptr;
230 		void *data;
231 		int data_size;
232 		bool encoded;
233 	} ring[MSM_GPU_MAX_RINGS];
234 
235 	int nr_registers;
236 	u32 *registers;
237 
238 	u32 rbbm_status;
239 
240 	char *comm;
241 	char *cmd;
242 
243 	int nr_bos;
244 	struct msm_gpu_state_bo *bos;
245 };
246 
247 static inline void gpu_write(struct msm_gpu *gpu, u32 reg, u32 data)
248 {
249 	msm_writel(data, gpu->mmio + (reg << 2));
250 }
251 
252 static inline u32 gpu_read(struct msm_gpu *gpu, u32 reg)
253 {
254 	return msm_readl(gpu->mmio + (reg << 2));
255 }
256 
257 static inline void gpu_rmw(struct msm_gpu *gpu, u32 reg, u32 mask, u32 or)
258 {
259 	msm_rmw(gpu->mmio + (reg << 2), mask, or);
260 }
261 
262 static inline u64 gpu_read64(struct msm_gpu *gpu, u32 lo, u32 hi)
263 {
264 	u64 val;
265 
266 	/*
267 	 * Why not a readq here? Two reasons: 1) many of the LO registers are
268 	 * not quad word aligned and 2) the GPU hardware designers have a bit
269 	 * of a history of putting registers where they fit, especially in
270 	 * spins. The longer a GPU family goes the higher the chance that
271 	 * we'll get burned.  We could do a series of validity checks if we
272 	 * wanted to, but really is a readq() that much better? Nah.
273 	 */
274 
275 	/*
276 	 * For some lo/hi registers (like perfcounters), the hi value is latched
277 	 * when the lo is read, so make sure to read the lo first to trigger
278 	 * that
279 	 */
280 	val = (u64) msm_readl(gpu->mmio + (lo << 2));
281 	val |= ((u64) msm_readl(gpu->mmio + (hi << 2)) << 32);
282 
283 	return val;
284 }
285 
286 static inline void gpu_write64(struct msm_gpu *gpu, u32 lo, u32 hi, u64 val)
287 {
288 	/* Why not a writeq here? Read the screed above */
289 	msm_writel(lower_32_bits(val), gpu->mmio + (lo << 2));
290 	msm_writel(upper_32_bits(val), gpu->mmio + (hi << 2));
291 }
292 
293 int msm_gpu_pm_suspend(struct msm_gpu *gpu);
294 int msm_gpu_pm_resume(struct msm_gpu *gpu);
295 void msm_gpu_resume_devfreq(struct msm_gpu *gpu);
296 
297 int msm_gpu_hw_init(struct msm_gpu *gpu);
298 
299 void msm_gpu_perfcntr_start(struct msm_gpu *gpu);
300 void msm_gpu_perfcntr_stop(struct msm_gpu *gpu);
301 int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime,
302 		uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs);
303 
304 void msm_gpu_retire(struct msm_gpu *gpu);
305 void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit);
306 
307 int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
308 		struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
309 		const char *name, struct msm_gpu_config *config);
310 
311 struct msm_gem_address_space *
312 msm_gpu_create_private_address_space(struct msm_gpu *gpu, struct task_struct *task);
313 
314 void msm_gpu_cleanup(struct msm_gpu *gpu);
315 
316 struct msm_gpu *adreno_load_gpu(struct drm_device *dev);
317 void __init adreno_register(void);
318 void __exit adreno_unregister(void);
319 
320 static inline void msm_submitqueue_put(struct msm_gpu_submitqueue *queue)
321 {
322 	if (queue)
323 		kref_put(&queue->ref, msm_submitqueue_destroy);
324 }
325 
326 static inline struct msm_gpu_state *msm_gpu_crashstate_get(struct msm_gpu *gpu)
327 {
328 	struct msm_gpu_state *state = NULL;
329 
330 	mutex_lock(&gpu->dev->struct_mutex);
331 
332 	if (gpu->crashstate) {
333 		kref_get(&gpu->crashstate->ref);
334 		state = gpu->crashstate;
335 	}
336 
337 	mutex_unlock(&gpu->dev->struct_mutex);
338 
339 	return state;
340 }
341 
342 static inline void msm_gpu_crashstate_put(struct msm_gpu *gpu)
343 {
344 	mutex_lock(&gpu->dev->struct_mutex);
345 
346 	if (gpu->crashstate) {
347 		if (gpu->funcs->gpu_state_put(gpu->crashstate))
348 			gpu->crashstate = NULL;
349 	}
350 
351 	mutex_unlock(&gpu->dev->struct_mutex);
352 }
353 
354 /*
355  * Simple macro to semi-cleanly add the MAP_PRIV flag for targets that can
356  * support expanded privileges
357  */
358 #define check_apriv(gpu, flags) \
359 	(((gpu)->hw_apriv ? MSM_BO_MAP_PRIV : 0) | (flags))
360 
361 
362 #endif /* __MSM_GPU_H__ */
363