xref: /linux/drivers/gpu/drm/msm/adreno/a6xx_gpu.h (revision e814f3fd16acfb7f9966773953de8f740a1e3202)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 /* Copyright (c) 2017, 2019 The Linux Foundation. All rights reserved. */
3 
4 #ifndef __A6XX_GPU_H__
5 #define __A6XX_GPU_H__
6 
7 
8 #include "adreno_gpu.h"
9 #include "a6xx.xml.h"
10 
11 #include "a6xx_gmu.h"
12 
13 extern bool hang_debug;
14 
15 struct cpu_gpu_lock {
16 	uint32_t gpu_req;
17 	uint32_t cpu_req;
18 	uint32_t turn;
19 	union {
20 		struct {
21 			uint16_t list_length;
22 			uint16_t list_offset;
23 		};
24 		struct {
25 			uint8_t ifpc_list_len;
26 			uint8_t preemption_list_len;
27 			uint16_t dynamic_list_len;
28 		};
29 	};
30 	uint64_t regs[62];
31 };
32 
33 /**
34  * struct a6xx_info - a6xx specific information from device table
35  *
36  * @hwcg: hw clock gating register sequence
37  * @protect: CP_PROTECT settings
38  * @pwrup_reglist pwrup reglist for preemption
39  */
40 struct a6xx_info {
41 	const struct adreno_reglist *hwcg;
42 	const struct adreno_protect *protect;
43 	const struct adreno_reglist_list *pwrup_reglist;
44 	u32 gmu_chipid;
45 	u32 gmu_cgc_mode;
46 	u32 prim_fifo_threshold;
47 	const struct a6xx_bcm *bcms;
48 };
49 
50 struct a6xx_gpu {
51 	struct adreno_gpu base;
52 
53 	struct drm_gem_object *sqe_bo;
54 	uint64_t sqe_iova;
55 
56 	struct msm_ringbuffer *cur_ring;
57 	struct msm_ringbuffer *next_ring;
58 
59 	struct drm_gem_object *preempt_bo[MSM_GPU_MAX_RINGS];
60 	void *preempt[MSM_GPU_MAX_RINGS];
61 	uint64_t preempt_iova[MSM_GPU_MAX_RINGS];
62 	struct drm_gem_object *preempt_smmu_bo[MSM_GPU_MAX_RINGS];
63 	void *preempt_smmu[MSM_GPU_MAX_RINGS];
64 	uint64_t preempt_smmu_iova[MSM_GPU_MAX_RINGS];
65 	uint32_t last_seqno[MSM_GPU_MAX_RINGS];
66 
67 	atomic_t preempt_state;
68 	spinlock_t eval_lock;
69 	struct timer_list preempt_timer;
70 
71 	unsigned int preempt_level;
72 	bool uses_gmem;
73 	bool skip_save_restore;
74 
75 	struct drm_gem_object *preempt_postamble_bo;
76 	void *preempt_postamble_ptr;
77 	uint64_t preempt_postamble_iova;
78 	uint64_t preempt_postamble_len;
79 	bool postamble_enabled;
80 
81 	struct a6xx_gmu gmu;
82 
83 	struct drm_gem_object *shadow_bo;
84 	uint64_t shadow_iova;
85 	uint32_t *shadow;
86 
87 	struct drm_gem_object *pwrup_reglist_bo;
88 	void *pwrup_reglist_ptr;
89 	uint64_t pwrup_reglist_iova;
90 	bool pwrup_reglist_emitted;
91 
92 	bool has_whereami;
93 
94 	void __iomem *llc_mmio;
95 	void *llc_slice;
96 	void *htw_llc_slice;
97 	bool have_mmu500;
98 	bool hung;
99 };
100 
101 #define to_a6xx_gpu(x) container_of(x, struct a6xx_gpu, base)
102 
103 /*
104  * In order to do lockless preemption we use a simple state machine to progress
105  * through the process.
106  *
107  * PREEMPT_NONE - no preemption in progress.  Next state START.
108  * PREEMPT_START - The trigger is evaluating if preemption is possible. Next
109  * states: TRIGGERED, NONE
110  * PREEMPT_FINISH - An intermediate state before moving back to NONE. Next
111  * state: NONE.
112  * PREEMPT_TRIGGERED: A preemption has been executed on the hardware. Next
113  * states: FAULTED, PENDING
114  * PREEMPT_FAULTED: A preemption timed out (never completed). This will trigger
115  * recovery.  Next state: N/A
116  * PREEMPT_PENDING: Preemption complete interrupt fired - the callback is
117  * checking the success of the operation. Next state: FAULTED, NONE.
118  */
119 
120 enum a6xx_preempt_state {
121 	PREEMPT_NONE = 0,
122 	PREEMPT_START,
123 	PREEMPT_FINISH,
124 	PREEMPT_TRIGGERED,
125 	PREEMPT_FAULTED,
126 	PREEMPT_PENDING,
127 };
128 
129 /*
130  * struct a6xx_preempt_record is a shared buffer between the microcode and the
131  * CPU to store the state for preemption. The record itself is much larger
132  * (2112k) but most of that is used by the CP for storage.
133  *
134  * There is a preemption record assigned per ringbuffer. When the CPU triggers a
135  * preemption, it fills out the record with the useful information (wptr, ring
136  * base, etc) and the microcode uses that information to set up the CP following
137  * the preemption.  When a ring is switched out, the CP will save the ringbuffer
138  * state back to the record. In this way, once the records are properly set up
139  * the CPU can quickly switch back and forth between ringbuffers by only
140  * updating a few registers (often only the wptr).
141  *
142  * These are the CPU aware registers in the record:
143  * @magic: Must always be 0xAE399D6EUL
144  * @info: Type of the record - written 0 by the CPU, updated by the CP
145  * @errno: preemption error record
146  * @data: Data field in YIELD and SET_MARKER packets, Written and used by CP
147  * @cntl: Value of RB_CNTL written by CPU, save/restored by CP
148  * @rptr: Value of RB_RPTR written by CPU, save/restored by CP
149  * @wptr: Value of RB_WPTR written by CPU, save/restored by CP
150  * @_pad: Reserved/padding
151  * @rptr_addr: Value of RB_RPTR_ADDR_LO|HI written by CPU, save/restored by CP
152  * @rbase: Value of RB_BASE written by CPU, save/restored by CP
153  * @counter: GPU address of the storage area for the preemption counters
154  * @bv_rptr_addr: Value of BV_RB_RPTR_ADDR_LO|HI written by CPU, save/restored by CP
155  */
156 struct a6xx_preempt_record {
157 	u32 magic;
158 	u32 info;
159 	u32 errno;
160 	u32 data;
161 	u32 cntl;
162 	u32 rptr;
163 	u32 wptr;
164 	u32 _pad;
165 	u64 rptr_addr;
166 	u64 rbase;
167 	u64 counter;
168 	u64 bv_rptr_addr;
169 };
170 
171 #define A6XX_PREEMPT_RECORD_MAGIC 0xAE399D6EUL
172 
173 #define PREEMPT_SMMU_INFO_SIZE 4096
174 
175 #define PREEMPT_RECORD_SIZE(adreno_gpu) \
176 	((adreno_gpu->info->preempt_record_size) == 0 ? \
177 	 4192 * SZ_1K : (adreno_gpu->info->preempt_record_size))
178 
179 /*
180  * The preemption counter block is a storage area for the value of the
181  * preemption counters that are saved immediately before context switch. We
182  * append it on to the end of the allocation for the preemption record.
183  */
184 #define A6XX_PREEMPT_COUNTER_SIZE (16 * 4)
185 
186 struct a7xx_cp_smmu_info {
187 	u32 magic;
188 	u32 _pad4;
189 	u64 ttbr0;
190 	u32 asid;
191 	u32 context_idr;
192 	u32 context_bank;
193 };
194 
195 #define GEN7_CP_SMMU_INFO_MAGIC 0x241350d5UL
196 
197 /*
198  * Given a register and a count, return a value to program into
199  * REG_CP_PROTECT_REG(n) - this will block both reads and writes for
200  * _len + 1 registers starting at _reg.
201  */
202 #define A6XX_PROTECT_NORDWR(_reg, _len) \
203 	((1 << 31) | \
204 	(((_len) & 0x3FFF) << 18) | ((_reg) & 0x3FFFF))
205 
206 /*
207  * Same as above, but allow reads over the range. For areas of mixed use (such
208  * as performance counters) this allows us to protect a much larger range with a
209  * single register
210  */
211 #define A6XX_PROTECT_RDONLY(_reg, _len) \
212 	((((_len) & 0x3FFF) << 18) | ((_reg) & 0x3FFFF))
213 
214 static inline bool a6xx_has_gbif(struct adreno_gpu *gpu)
215 {
216 	if(adreno_is_a630(gpu))
217 		return false;
218 
219 	return true;
220 }
221 
222 static inline void a6xx_llc_rmw(struct a6xx_gpu *a6xx_gpu, u32 reg, u32 mask, u32 or)
223 {
224 	return msm_rmw(a6xx_gpu->llc_mmio + (reg << 2), mask, or);
225 }
226 
227 static inline u32 a6xx_llc_read(struct a6xx_gpu *a6xx_gpu, u32 reg)
228 {
229 	return readl(a6xx_gpu->llc_mmio + (reg << 2));
230 }
231 
232 static inline void a6xx_llc_write(struct a6xx_gpu *a6xx_gpu, u32 reg, u32 value)
233 {
234 	writel(value, a6xx_gpu->llc_mmio + (reg << 2));
235 }
236 
237 #define shadowptr(_a6xx_gpu, _ring) ((_a6xx_gpu)->shadow_iova + \
238 		((_ring)->id * sizeof(uint32_t)))
239 
240 int a6xx_gmu_resume(struct a6xx_gpu *gpu);
241 int a6xx_gmu_stop(struct a6xx_gpu *gpu);
242 
243 int a6xx_gmu_wait_for_idle(struct a6xx_gmu *gmu);
244 
245 bool a6xx_gmu_isidle(struct a6xx_gmu *gmu);
246 
247 int a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state);
248 void a6xx_gmu_clear_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state);
249 
250 int a6xx_gmu_init(struct a6xx_gpu *a6xx_gpu, struct device_node *node);
251 int a6xx_gmu_wrapper_init(struct a6xx_gpu *a6xx_gpu, struct device_node *node);
252 void a6xx_gmu_remove(struct a6xx_gpu *a6xx_gpu);
253 
254 void a6xx_preempt_init(struct msm_gpu *gpu);
255 void a6xx_preempt_hw_init(struct msm_gpu *gpu);
256 void a6xx_preempt_trigger(struct msm_gpu *gpu);
257 void a6xx_preempt_irq(struct msm_gpu *gpu);
258 void a6xx_preempt_fini(struct msm_gpu *gpu);
259 int a6xx_preempt_submitqueue_setup(struct msm_gpu *gpu,
260 		struct msm_gpu_submitqueue *queue);
261 void a6xx_preempt_submitqueue_close(struct msm_gpu *gpu,
262 		struct msm_gpu_submitqueue *queue);
263 
264 /* Return true if we are in a preempt state */
265 static inline bool a6xx_in_preempt(struct a6xx_gpu *a6xx_gpu)
266 {
267 	/*
268 	 * Make sure the read to preempt_state is ordered with respect to reads
269 	 * of other variables before ...
270 	 */
271 	smp_rmb();
272 
273 	int preempt_state = atomic_read(&a6xx_gpu->preempt_state);
274 
275 	/* ... and after. */
276 	smp_rmb();
277 
278 	return !(preempt_state == PREEMPT_NONE ||
279 			preempt_state == PREEMPT_FINISH);
280 }
281 
282 void a6xx_gmu_set_freq(struct msm_gpu *gpu, struct dev_pm_opp *opp,
283 		       bool suspended);
284 unsigned long a6xx_gmu_get_freq(struct msm_gpu *gpu);
285 
286 void a6xx_show(struct msm_gpu *gpu, struct msm_gpu_state *state,
287 		struct drm_printer *p);
288 
289 struct msm_gpu_state *a6xx_gpu_state_get(struct msm_gpu *gpu);
290 int a6xx_gpu_state_put(struct msm_gpu_state *state);
291 
292 void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu, bool gx_off);
293 void a6xx_gpu_sw_reset(struct msm_gpu *gpu, bool assert);
294 
295 #endif /* __A6XX_GPU_H__ */
296