xref: /openbmc/linux/drivers/gpu/drm/msm/adreno/a6xx_gpu.c (revision fbb6b31a)
1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright (c) 2017-2019 The Linux Foundation. All rights reserved. */
3 
4 
5 #include "msm_gem.h"
6 #include "msm_mmu.h"
7 #include "msm_gpu_trace.h"
8 #include "a6xx_gpu.h"
9 #include "a6xx_gmu.xml.h"
10 
11 #include <linux/bitfield.h>
12 #include <linux/devfreq.h>
13 #include <linux/soc/qcom/llcc-qcom.h>
14 
15 #define GPU_PAS_ID 13
16 
17 static inline bool _a6xx_check_idle(struct msm_gpu *gpu)
18 {
19 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
20 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
21 
22 	/* Check that the GMU is idle */
23 	if (!a6xx_gmu_isidle(&a6xx_gpu->gmu))
24 		return false;
25 
26 	/* Check tha the CX master is idle */
27 	if (gpu_read(gpu, REG_A6XX_RBBM_STATUS) &
28 			~A6XX_RBBM_STATUS_CP_AHB_BUSY_CX_MASTER)
29 		return false;
30 
31 	return !(gpu_read(gpu, REG_A6XX_RBBM_INT_0_STATUS) &
32 		A6XX_RBBM_INT_0_MASK_RBBM_HANG_DETECT);
33 }
34 
35 static bool a6xx_idle(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
36 {
37 	/* wait for CP to drain ringbuffer: */
38 	if (!adreno_idle(gpu, ring))
39 		return false;
40 
41 	if (spin_until(_a6xx_check_idle(gpu))) {
42 		DRM_ERROR("%s: %ps: timeout waiting for GPU to idle: status %8.8X irq %8.8X rptr/wptr %d/%d\n",
43 			gpu->name, __builtin_return_address(0),
44 			gpu_read(gpu, REG_A6XX_RBBM_STATUS),
45 			gpu_read(gpu, REG_A6XX_RBBM_INT_0_STATUS),
46 			gpu_read(gpu, REG_A6XX_CP_RB_RPTR),
47 			gpu_read(gpu, REG_A6XX_CP_RB_WPTR));
48 		return false;
49 	}
50 
51 	return true;
52 }
53 
54 static void update_shadow_rptr(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
55 {
56 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
57 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
58 
59 	/* Expanded APRIV doesn't need to issue the WHERE_AM_I opcode */
60 	if (a6xx_gpu->has_whereami && !adreno_gpu->base.hw_apriv) {
61 		OUT_PKT7(ring, CP_WHERE_AM_I, 2);
62 		OUT_RING(ring, lower_32_bits(shadowptr(a6xx_gpu, ring)));
63 		OUT_RING(ring, upper_32_bits(shadowptr(a6xx_gpu, ring)));
64 	}
65 }
66 
67 static void a6xx_flush(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
68 {
69 	uint32_t wptr;
70 	unsigned long flags;
71 
72 	update_shadow_rptr(gpu, ring);
73 
74 	spin_lock_irqsave(&ring->preempt_lock, flags);
75 
76 	/* Copy the shadow to the actual register */
77 	ring->cur = ring->next;
78 
79 	/* Make sure to wrap wptr if we need to */
80 	wptr = get_wptr(ring);
81 
82 	spin_unlock_irqrestore(&ring->preempt_lock, flags);
83 
84 	/* Make sure everything is posted before making a decision */
85 	mb();
86 
87 	gpu_write(gpu, REG_A6XX_CP_RB_WPTR, wptr);
88 }
89 
90 static void get_stats_counter(struct msm_ringbuffer *ring, u32 counter,
91 		u64 iova)
92 {
93 	OUT_PKT7(ring, CP_REG_TO_MEM, 3);
94 	OUT_RING(ring, CP_REG_TO_MEM_0_REG(counter) |
95 		CP_REG_TO_MEM_0_CNT(2) |
96 		CP_REG_TO_MEM_0_64B);
97 	OUT_RING(ring, lower_32_bits(iova));
98 	OUT_RING(ring, upper_32_bits(iova));
99 }
100 
101 static void a6xx_set_pagetable(struct a6xx_gpu *a6xx_gpu,
102 		struct msm_ringbuffer *ring, struct msm_file_private *ctx)
103 {
104 	bool sysprof = refcount_read(&a6xx_gpu->base.base.sysprof_active) > 1;
105 	phys_addr_t ttbr;
106 	u32 asid;
107 	u64 memptr = rbmemptr(ring, ttbr0);
108 
109 	if (ctx->seqno == a6xx_gpu->base.base.cur_ctx_seqno)
110 		return;
111 
112 	if (msm_iommu_pagetable_params(ctx->aspace->mmu, &ttbr, &asid))
113 		return;
114 
115 	if (!sysprof) {
116 		/* Turn off protected mode to write to special registers */
117 		OUT_PKT7(ring, CP_SET_PROTECTED_MODE, 1);
118 		OUT_RING(ring, 0);
119 
120 		OUT_PKT4(ring, REG_A6XX_RBBM_PERFCTR_SRAM_INIT_CMD, 1);
121 		OUT_RING(ring, 1);
122 	}
123 
124 	/* Execute the table update */
125 	OUT_PKT7(ring, CP_SMMU_TABLE_UPDATE, 4);
126 	OUT_RING(ring, CP_SMMU_TABLE_UPDATE_0_TTBR0_LO(lower_32_bits(ttbr)));
127 
128 	OUT_RING(ring,
129 		CP_SMMU_TABLE_UPDATE_1_TTBR0_HI(upper_32_bits(ttbr)) |
130 		CP_SMMU_TABLE_UPDATE_1_ASID(asid));
131 	OUT_RING(ring, CP_SMMU_TABLE_UPDATE_2_CONTEXTIDR(0));
132 	OUT_RING(ring, CP_SMMU_TABLE_UPDATE_3_CONTEXTBANK(0));
133 
134 	/*
135 	 * Write the new TTBR0 to the memstore. This is good for debugging.
136 	 */
137 	OUT_PKT7(ring, CP_MEM_WRITE, 4);
138 	OUT_RING(ring, CP_MEM_WRITE_0_ADDR_LO(lower_32_bits(memptr)));
139 	OUT_RING(ring, CP_MEM_WRITE_1_ADDR_HI(upper_32_bits(memptr)));
140 	OUT_RING(ring, lower_32_bits(ttbr));
141 	OUT_RING(ring, (asid << 16) | upper_32_bits(ttbr));
142 
143 	/*
144 	 * And finally, trigger a uche flush to be sure there isn't anything
145 	 * lingering in that part of the GPU
146 	 */
147 
148 	OUT_PKT7(ring, CP_EVENT_WRITE, 1);
149 	OUT_RING(ring, 0x31);
150 
151 	if (!sysprof) {
152 		/*
153 		 * Wait for SRAM clear after the pgtable update, so the
154 		 * two can happen in parallel:
155 		 */
156 		OUT_PKT7(ring, CP_WAIT_REG_MEM, 6);
157 		OUT_RING(ring, CP_WAIT_REG_MEM_0_FUNCTION(WRITE_EQ));
158 		OUT_RING(ring, CP_WAIT_REG_MEM_1_POLL_ADDR_LO(
159 				REG_A6XX_RBBM_PERFCTR_SRAM_INIT_STATUS));
160 		OUT_RING(ring, CP_WAIT_REG_MEM_2_POLL_ADDR_HI(0));
161 		OUT_RING(ring, CP_WAIT_REG_MEM_3_REF(0x1));
162 		OUT_RING(ring, CP_WAIT_REG_MEM_4_MASK(0x1));
163 		OUT_RING(ring, CP_WAIT_REG_MEM_5_DELAY_LOOP_CYCLES(0));
164 
165 		/* Re-enable protected mode: */
166 		OUT_PKT7(ring, CP_SET_PROTECTED_MODE, 1);
167 		OUT_RING(ring, 1);
168 	}
169 }
170 
171 static void a6xx_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
172 {
173 	unsigned int index = submit->seqno % MSM_GPU_SUBMIT_STATS_COUNT;
174 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
175 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
176 	struct msm_ringbuffer *ring = submit->ring;
177 	unsigned int i, ibs = 0;
178 
179 	a6xx_set_pagetable(a6xx_gpu, ring, submit->queue->ctx);
180 
181 	get_stats_counter(ring, REG_A6XX_RBBM_PERFCTR_CP(0),
182 		rbmemptr_stats(ring, index, cpcycles_start));
183 
184 	/*
185 	 * For PM4 the GMU register offsets are calculated from the base of the
186 	 * GPU registers so we need to add 0x1a800 to the register value on A630
187 	 * to get the right value from PM4.
188 	 */
189 	get_stats_counter(ring, REG_A6XX_CP_ALWAYS_ON_COUNTER_LO,
190 		rbmemptr_stats(ring, index, alwayson_start));
191 
192 	/* Invalidate CCU depth and color */
193 	OUT_PKT7(ring, CP_EVENT_WRITE, 1);
194 	OUT_RING(ring, CP_EVENT_WRITE_0_EVENT(PC_CCU_INVALIDATE_DEPTH));
195 
196 	OUT_PKT7(ring, CP_EVENT_WRITE, 1);
197 	OUT_RING(ring, CP_EVENT_WRITE_0_EVENT(PC_CCU_INVALIDATE_COLOR));
198 
199 	/* Submit the commands */
200 	for (i = 0; i < submit->nr_cmds; i++) {
201 		switch (submit->cmd[i].type) {
202 		case MSM_SUBMIT_CMD_IB_TARGET_BUF:
203 			break;
204 		case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
205 			if (gpu->cur_ctx_seqno == submit->queue->ctx->seqno)
206 				break;
207 			fallthrough;
208 		case MSM_SUBMIT_CMD_BUF:
209 			OUT_PKT7(ring, CP_INDIRECT_BUFFER_PFE, 3);
210 			OUT_RING(ring, lower_32_bits(submit->cmd[i].iova));
211 			OUT_RING(ring, upper_32_bits(submit->cmd[i].iova));
212 			OUT_RING(ring, submit->cmd[i].size);
213 			ibs++;
214 			break;
215 		}
216 
217 		/*
218 		 * Periodically update shadow-wptr if needed, so that we
219 		 * can see partial progress of submits with large # of
220 		 * cmds.. otherwise we could needlessly stall waiting for
221 		 * ringbuffer state, simply due to looking at a shadow
222 		 * rptr value that has not been updated
223 		 */
224 		if ((ibs % 32) == 0)
225 			update_shadow_rptr(gpu, ring);
226 	}
227 
228 	get_stats_counter(ring, REG_A6XX_RBBM_PERFCTR_CP(0),
229 		rbmemptr_stats(ring, index, cpcycles_end));
230 	get_stats_counter(ring, REG_A6XX_CP_ALWAYS_ON_COUNTER_LO,
231 		rbmemptr_stats(ring, index, alwayson_end));
232 
233 	/* Write the fence to the scratch register */
234 	OUT_PKT4(ring, REG_A6XX_CP_SCRATCH_REG(2), 1);
235 	OUT_RING(ring, submit->seqno);
236 
237 	/*
238 	 * Execute a CACHE_FLUSH_TS event. This will ensure that the
239 	 * timestamp is written to the memory and then triggers the interrupt
240 	 */
241 	OUT_PKT7(ring, CP_EVENT_WRITE, 4);
242 	OUT_RING(ring, CP_EVENT_WRITE_0_EVENT(CACHE_FLUSH_TS) |
243 		CP_EVENT_WRITE_0_IRQ);
244 	OUT_RING(ring, lower_32_bits(rbmemptr(ring, fence)));
245 	OUT_RING(ring, upper_32_bits(rbmemptr(ring, fence)));
246 	OUT_RING(ring, submit->seqno);
247 
248 	trace_msm_gpu_submit_flush(submit,
249 		gpu_read64(gpu, REG_A6XX_CP_ALWAYS_ON_COUNTER_LO,
250 			REG_A6XX_CP_ALWAYS_ON_COUNTER_HI));
251 
252 	a6xx_flush(gpu, ring);
253 }
254 
255 const struct adreno_reglist a630_hwcg[] = {
256 	{REG_A6XX_RBBM_CLOCK_CNTL_SP0, 0x22222222},
257 	{REG_A6XX_RBBM_CLOCK_CNTL_SP1, 0x22222222},
258 	{REG_A6XX_RBBM_CLOCK_CNTL_SP2, 0x22222222},
259 	{REG_A6XX_RBBM_CLOCK_CNTL_SP3, 0x22222222},
260 	{REG_A6XX_RBBM_CLOCK_CNTL2_SP0, 0x02022220},
261 	{REG_A6XX_RBBM_CLOCK_CNTL2_SP1, 0x02022220},
262 	{REG_A6XX_RBBM_CLOCK_CNTL2_SP2, 0x02022220},
263 	{REG_A6XX_RBBM_CLOCK_CNTL2_SP3, 0x02022220},
264 	{REG_A6XX_RBBM_CLOCK_DELAY_SP0, 0x00000080},
265 	{REG_A6XX_RBBM_CLOCK_DELAY_SP1, 0x00000080},
266 	{REG_A6XX_RBBM_CLOCK_DELAY_SP2, 0x00000080},
267 	{REG_A6XX_RBBM_CLOCK_DELAY_SP3, 0x00000080},
268 	{REG_A6XX_RBBM_CLOCK_HYST_SP0, 0x0000f3cf},
269 	{REG_A6XX_RBBM_CLOCK_HYST_SP1, 0x0000f3cf},
270 	{REG_A6XX_RBBM_CLOCK_HYST_SP2, 0x0000f3cf},
271 	{REG_A6XX_RBBM_CLOCK_HYST_SP3, 0x0000f3cf},
272 	{REG_A6XX_RBBM_CLOCK_CNTL_TP0, 0x02222222},
273 	{REG_A6XX_RBBM_CLOCK_CNTL_TP1, 0x02222222},
274 	{REG_A6XX_RBBM_CLOCK_CNTL_TP2, 0x02222222},
275 	{REG_A6XX_RBBM_CLOCK_CNTL_TP3, 0x02222222},
276 	{REG_A6XX_RBBM_CLOCK_CNTL2_TP0, 0x22222222},
277 	{REG_A6XX_RBBM_CLOCK_CNTL2_TP1, 0x22222222},
278 	{REG_A6XX_RBBM_CLOCK_CNTL2_TP2, 0x22222222},
279 	{REG_A6XX_RBBM_CLOCK_CNTL2_TP3, 0x22222222},
280 	{REG_A6XX_RBBM_CLOCK_CNTL3_TP0, 0x22222222},
281 	{REG_A6XX_RBBM_CLOCK_CNTL3_TP1, 0x22222222},
282 	{REG_A6XX_RBBM_CLOCK_CNTL3_TP2, 0x22222222},
283 	{REG_A6XX_RBBM_CLOCK_CNTL3_TP3, 0x22222222},
284 	{REG_A6XX_RBBM_CLOCK_CNTL4_TP0, 0x00022222},
285 	{REG_A6XX_RBBM_CLOCK_CNTL4_TP1, 0x00022222},
286 	{REG_A6XX_RBBM_CLOCK_CNTL4_TP2, 0x00022222},
287 	{REG_A6XX_RBBM_CLOCK_CNTL4_TP3, 0x00022222},
288 	{REG_A6XX_RBBM_CLOCK_HYST_TP0, 0x77777777},
289 	{REG_A6XX_RBBM_CLOCK_HYST_TP1, 0x77777777},
290 	{REG_A6XX_RBBM_CLOCK_HYST_TP2, 0x77777777},
291 	{REG_A6XX_RBBM_CLOCK_HYST_TP3, 0x77777777},
292 	{REG_A6XX_RBBM_CLOCK_HYST2_TP0, 0x77777777},
293 	{REG_A6XX_RBBM_CLOCK_HYST2_TP1, 0x77777777},
294 	{REG_A6XX_RBBM_CLOCK_HYST2_TP2, 0x77777777},
295 	{REG_A6XX_RBBM_CLOCK_HYST2_TP3, 0x77777777},
296 	{REG_A6XX_RBBM_CLOCK_HYST3_TP0, 0x77777777},
297 	{REG_A6XX_RBBM_CLOCK_HYST3_TP1, 0x77777777},
298 	{REG_A6XX_RBBM_CLOCK_HYST3_TP2, 0x77777777},
299 	{REG_A6XX_RBBM_CLOCK_HYST3_TP3, 0x77777777},
300 	{REG_A6XX_RBBM_CLOCK_HYST4_TP0, 0x00077777},
301 	{REG_A6XX_RBBM_CLOCK_HYST4_TP1, 0x00077777},
302 	{REG_A6XX_RBBM_CLOCK_HYST4_TP2, 0x00077777},
303 	{REG_A6XX_RBBM_CLOCK_HYST4_TP3, 0x00077777},
304 	{REG_A6XX_RBBM_CLOCK_DELAY_TP0, 0x11111111},
305 	{REG_A6XX_RBBM_CLOCK_DELAY_TP1, 0x11111111},
306 	{REG_A6XX_RBBM_CLOCK_DELAY_TP2, 0x11111111},
307 	{REG_A6XX_RBBM_CLOCK_DELAY_TP3, 0x11111111},
308 	{REG_A6XX_RBBM_CLOCK_DELAY2_TP0, 0x11111111},
309 	{REG_A6XX_RBBM_CLOCK_DELAY2_TP1, 0x11111111},
310 	{REG_A6XX_RBBM_CLOCK_DELAY2_TP2, 0x11111111},
311 	{REG_A6XX_RBBM_CLOCK_DELAY2_TP3, 0x11111111},
312 	{REG_A6XX_RBBM_CLOCK_DELAY3_TP0, 0x11111111},
313 	{REG_A6XX_RBBM_CLOCK_DELAY3_TP1, 0x11111111},
314 	{REG_A6XX_RBBM_CLOCK_DELAY3_TP2, 0x11111111},
315 	{REG_A6XX_RBBM_CLOCK_DELAY3_TP3, 0x11111111},
316 	{REG_A6XX_RBBM_CLOCK_DELAY4_TP0, 0x00011111},
317 	{REG_A6XX_RBBM_CLOCK_DELAY4_TP1, 0x00011111},
318 	{REG_A6XX_RBBM_CLOCK_DELAY4_TP2, 0x00011111},
319 	{REG_A6XX_RBBM_CLOCK_DELAY4_TP3, 0x00011111},
320 	{REG_A6XX_RBBM_CLOCK_CNTL_UCHE, 0x22222222},
321 	{REG_A6XX_RBBM_CLOCK_CNTL2_UCHE, 0x22222222},
322 	{REG_A6XX_RBBM_CLOCK_CNTL3_UCHE, 0x22222222},
323 	{REG_A6XX_RBBM_CLOCK_CNTL4_UCHE, 0x00222222},
324 	{REG_A6XX_RBBM_CLOCK_HYST_UCHE, 0x00000004},
325 	{REG_A6XX_RBBM_CLOCK_DELAY_UCHE, 0x00000002},
326 	{REG_A6XX_RBBM_CLOCK_CNTL_RB0, 0x22222222},
327 	{REG_A6XX_RBBM_CLOCK_CNTL_RB1, 0x22222222},
328 	{REG_A6XX_RBBM_CLOCK_CNTL_RB2, 0x22222222},
329 	{REG_A6XX_RBBM_CLOCK_CNTL_RB3, 0x22222222},
330 	{REG_A6XX_RBBM_CLOCK_CNTL2_RB0, 0x00002222},
331 	{REG_A6XX_RBBM_CLOCK_CNTL2_RB1, 0x00002222},
332 	{REG_A6XX_RBBM_CLOCK_CNTL2_RB2, 0x00002222},
333 	{REG_A6XX_RBBM_CLOCK_CNTL2_RB3, 0x00002222},
334 	{REG_A6XX_RBBM_CLOCK_CNTL_CCU0, 0x00002220},
335 	{REG_A6XX_RBBM_CLOCK_CNTL_CCU1, 0x00002220},
336 	{REG_A6XX_RBBM_CLOCK_CNTL_CCU2, 0x00002220},
337 	{REG_A6XX_RBBM_CLOCK_CNTL_CCU3, 0x00002220},
338 	{REG_A6XX_RBBM_CLOCK_HYST_RB_CCU0, 0x00040f00},
339 	{REG_A6XX_RBBM_CLOCK_HYST_RB_CCU1, 0x00040f00},
340 	{REG_A6XX_RBBM_CLOCK_HYST_RB_CCU2, 0x00040f00},
341 	{REG_A6XX_RBBM_CLOCK_HYST_RB_CCU3, 0x00040f00},
342 	{REG_A6XX_RBBM_CLOCK_CNTL_RAC, 0x05022022},
343 	{REG_A6XX_RBBM_CLOCK_CNTL2_RAC, 0x00005555},
344 	{REG_A6XX_RBBM_CLOCK_DELAY_RAC, 0x00000011},
345 	{REG_A6XX_RBBM_CLOCK_HYST_RAC, 0x00445044},
346 	{REG_A6XX_RBBM_CLOCK_CNTL_TSE_RAS_RBBM, 0x04222222},
347 	{REG_A6XX_RBBM_CLOCK_MODE_GPC, 0x00222222},
348 	{REG_A6XX_RBBM_CLOCK_MODE_VFD, 0x00002222},
349 	{REG_A6XX_RBBM_CLOCK_HYST_TSE_RAS_RBBM, 0x00000000},
350 	{REG_A6XX_RBBM_CLOCK_HYST_GPC, 0x04104004},
351 	{REG_A6XX_RBBM_CLOCK_HYST_VFD, 0x00000000},
352 	{REG_A6XX_RBBM_CLOCK_DELAY_HLSQ, 0x00000000},
353 	{REG_A6XX_RBBM_CLOCK_DELAY_TSE_RAS_RBBM, 0x00004000},
354 	{REG_A6XX_RBBM_CLOCK_DELAY_GPC, 0x00000200},
355 	{REG_A6XX_RBBM_CLOCK_DELAY_VFD, 0x00002222},
356 	{REG_A6XX_RBBM_CLOCK_DELAY_HLSQ_2, 0x00000002},
357 	{REG_A6XX_RBBM_CLOCK_MODE_HLSQ, 0x00002222},
358 	{REG_A6XX_RBBM_CLOCK_CNTL_GMU_GX, 0x00000222},
359 	{REG_A6XX_RBBM_CLOCK_DELAY_GMU_GX, 0x00000111},
360 	{REG_A6XX_RBBM_CLOCK_HYST_GMU_GX, 0x00000555},
361 	{},
362 };
363 
364 const struct adreno_reglist a640_hwcg[] = {
365 	{REG_A6XX_RBBM_CLOCK_CNTL_SP0, 0x02222222},
366 	{REG_A6XX_RBBM_CLOCK_CNTL2_SP0, 0x02222220},
367 	{REG_A6XX_RBBM_CLOCK_DELAY_SP0, 0x00000080},
368 	{REG_A6XX_RBBM_CLOCK_HYST_SP0, 0x0000F3CF},
369 	{REG_A6XX_RBBM_CLOCK_CNTL_TP0, 0x02222222},
370 	{REG_A6XX_RBBM_CLOCK_CNTL2_TP0, 0x22222222},
371 	{REG_A6XX_RBBM_CLOCK_CNTL3_TP0, 0x22222222},
372 	{REG_A6XX_RBBM_CLOCK_CNTL4_TP0, 0x00022222},
373 	{REG_A6XX_RBBM_CLOCK_DELAY_TP0, 0x11111111},
374 	{REG_A6XX_RBBM_CLOCK_DELAY2_TP0, 0x11111111},
375 	{REG_A6XX_RBBM_CLOCK_DELAY3_TP0, 0x11111111},
376 	{REG_A6XX_RBBM_CLOCK_DELAY4_TP0, 0x00011111},
377 	{REG_A6XX_RBBM_CLOCK_HYST_TP0, 0x77777777},
378 	{REG_A6XX_RBBM_CLOCK_HYST2_TP0, 0x77777777},
379 	{REG_A6XX_RBBM_CLOCK_HYST3_TP0, 0x77777777},
380 	{REG_A6XX_RBBM_CLOCK_HYST4_TP0, 0x00077777},
381 	{REG_A6XX_RBBM_CLOCK_CNTL_RB0, 0x22222222},
382 	{REG_A6XX_RBBM_CLOCK_CNTL2_RB0, 0x01002222},
383 	{REG_A6XX_RBBM_CLOCK_CNTL_CCU0, 0x00002220},
384 	{REG_A6XX_RBBM_CLOCK_HYST_RB_CCU0, 0x00040F00},
385 	{REG_A6XX_RBBM_CLOCK_CNTL_RAC, 0x05222022},
386 	{REG_A6XX_RBBM_CLOCK_CNTL2_RAC, 0x00005555},
387 	{REG_A6XX_RBBM_CLOCK_DELAY_RAC, 0x00000011},
388 	{REG_A6XX_RBBM_CLOCK_HYST_RAC, 0x00445044},
389 	{REG_A6XX_RBBM_CLOCK_CNTL_TSE_RAS_RBBM, 0x04222222},
390 	{REG_A6XX_RBBM_CLOCK_MODE_VFD, 0x00002222},
391 	{REG_A6XX_RBBM_CLOCK_MODE_GPC, 0x00222222},
392 	{REG_A6XX_RBBM_CLOCK_DELAY_HLSQ_2, 0x00000002},
393 	{REG_A6XX_RBBM_CLOCK_MODE_HLSQ, 0x00002222},
394 	{REG_A6XX_RBBM_CLOCK_DELAY_TSE_RAS_RBBM, 0x00004000},
395 	{REG_A6XX_RBBM_CLOCK_DELAY_VFD, 0x00002222},
396 	{REG_A6XX_RBBM_CLOCK_DELAY_GPC, 0x00000200},
397 	{REG_A6XX_RBBM_CLOCK_DELAY_HLSQ, 0x00000000},
398 	{REG_A6XX_RBBM_CLOCK_HYST_TSE_RAS_RBBM, 0x00000000},
399 	{REG_A6XX_RBBM_CLOCK_HYST_VFD, 0x00000000},
400 	{REG_A6XX_RBBM_CLOCK_HYST_GPC, 0x04104004},
401 	{REG_A6XX_RBBM_CLOCK_HYST_HLSQ, 0x00000000},
402 	{REG_A6XX_RBBM_CLOCK_CNTL_TEX_FCHE, 0x00000222},
403 	{REG_A6XX_RBBM_CLOCK_DELAY_TEX_FCHE, 0x00000111},
404 	{REG_A6XX_RBBM_CLOCK_HYST_TEX_FCHE, 0x00000000},
405 	{REG_A6XX_RBBM_CLOCK_CNTL_UCHE, 0x22222222},
406 	{REG_A6XX_RBBM_CLOCK_HYST_UCHE, 0x00000004},
407 	{REG_A6XX_RBBM_CLOCK_DELAY_UCHE, 0x00000002},
408 	{REG_A6XX_RBBM_ISDB_CNT, 0x00000182},
409 	{REG_A6XX_RBBM_RAC_THRESHOLD_CNT, 0x00000000},
410 	{REG_A6XX_RBBM_SP_HYST_CNT, 0x00000000},
411 	{REG_A6XX_RBBM_CLOCK_CNTL_GMU_GX, 0x00000222},
412 	{REG_A6XX_RBBM_CLOCK_DELAY_GMU_GX, 0x00000111},
413 	{REG_A6XX_RBBM_CLOCK_HYST_GMU_GX, 0x00000555},
414 	{},
415 };
416 
417 const struct adreno_reglist a650_hwcg[] = {
418 	{REG_A6XX_RBBM_CLOCK_CNTL_SP0, 0x02222222},
419 	{REG_A6XX_RBBM_CLOCK_CNTL2_SP0, 0x02222220},
420 	{REG_A6XX_RBBM_CLOCK_DELAY_SP0, 0x00000080},
421 	{REG_A6XX_RBBM_CLOCK_HYST_SP0, 0x0000F3CF},
422 	{REG_A6XX_RBBM_CLOCK_CNTL_TP0, 0x02222222},
423 	{REG_A6XX_RBBM_CLOCK_CNTL2_TP0, 0x22222222},
424 	{REG_A6XX_RBBM_CLOCK_CNTL3_TP0, 0x22222222},
425 	{REG_A6XX_RBBM_CLOCK_CNTL4_TP0, 0x00022222},
426 	{REG_A6XX_RBBM_CLOCK_DELAY_TP0, 0x11111111},
427 	{REG_A6XX_RBBM_CLOCK_DELAY2_TP0, 0x11111111},
428 	{REG_A6XX_RBBM_CLOCK_DELAY3_TP0, 0x11111111},
429 	{REG_A6XX_RBBM_CLOCK_DELAY4_TP0, 0x00011111},
430 	{REG_A6XX_RBBM_CLOCK_HYST_TP0, 0x77777777},
431 	{REG_A6XX_RBBM_CLOCK_HYST2_TP0, 0x77777777},
432 	{REG_A6XX_RBBM_CLOCK_HYST3_TP0, 0x77777777},
433 	{REG_A6XX_RBBM_CLOCK_HYST4_TP0, 0x00077777},
434 	{REG_A6XX_RBBM_CLOCK_CNTL_RB0, 0x22222222},
435 	{REG_A6XX_RBBM_CLOCK_CNTL2_RB0, 0x01002222},
436 	{REG_A6XX_RBBM_CLOCK_CNTL_CCU0, 0x00002220},
437 	{REG_A6XX_RBBM_CLOCK_HYST_RB_CCU0, 0x00040F00},
438 	{REG_A6XX_RBBM_CLOCK_CNTL_RAC, 0x25222022},
439 	{REG_A6XX_RBBM_CLOCK_CNTL2_RAC, 0x00005555},
440 	{REG_A6XX_RBBM_CLOCK_DELAY_RAC, 0x00000011},
441 	{REG_A6XX_RBBM_CLOCK_HYST_RAC, 0x00445044},
442 	{REG_A6XX_RBBM_CLOCK_CNTL_TSE_RAS_RBBM, 0x04222222},
443 	{REG_A6XX_RBBM_CLOCK_MODE_VFD, 0x00002222},
444 	{REG_A6XX_RBBM_CLOCK_MODE_GPC, 0x00222222},
445 	{REG_A6XX_RBBM_CLOCK_DELAY_HLSQ_2, 0x00000002},
446 	{REG_A6XX_RBBM_CLOCK_MODE_HLSQ, 0x00002222},
447 	{REG_A6XX_RBBM_CLOCK_DELAY_TSE_RAS_RBBM, 0x00004000},
448 	{REG_A6XX_RBBM_CLOCK_DELAY_VFD, 0x00002222},
449 	{REG_A6XX_RBBM_CLOCK_DELAY_GPC, 0x00000200},
450 	{REG_A6XX_RBBM_CLOCK_DELAY_HLSQ, 0x00000000},
451 	{REG_A6XX_RBBM_CLOCK_HYST_TSE_RAS_RBBM, 0x00000000},
452 	{REG_A6XX_RBBM_CLOCK_HYST_VFD, 0x00000000},
453 	{REG_A6XX_RBBM_CLOCK_HYST_GPC, 0x04104004},
454 	{REG_A6XX_RBBM_CLOCK_HYST_HLSQ, 0x00000000},
455 	{REG_A6XX_RBBM_CLOCK_CNTL_TEX_FCHE, 0x00000222},
456 	{REG_A6XX_RBBM_CLOCK_DELAY_TEX_FCHE, 0x00000111},
457 	{REG_A6XX_RBBM_CLOCK_HYST_TEX_FCHE, 0x00000777},
458 	{REG_A6XX_RBBM_CLOCK_CNTL_UCHE, 0x22222222},
459 	{REG_A6XX_RBBM_CLOCK_HYST_UCHE, 0x00000004},
460 	{REG_A6XX_RBBM_CLOCK_DELAY_UCHE, 0x00000002},
461 	{REG_A6XX_RBBM_ISDB_CNT, 0x00000182},
462 	{REG_A6XX_RBBM_RAC_THRESHOLD_CNT, 0x00000000},
463 	{REG_A6XX_RBBM_SP_HYST_CNT, 0x00000000},
464 	{REG_A6XX_RBBM_CLOCK_CNTL_GMU_GX, 0x00000222},
465 	{REG_A6XX_RBBM_CLOCK_DELAY_GMU_GX, 0x00000111},
466 	{REG_A6XX_RBBM_CLOCK_HYST_GMU_GX, 0x00000555},
467 	{},
468 };
469 
470 const struct adreno_reglist a660_hwcg[] = {
471 	{REG_A6XX_RBBM_CLOCK_CNTL_SP0, 0x02222222},
472 	{REG_A6XX_RBBM_CLOCK_CNTL2_SP0, 0x02222220},
473 	{REG_A6XX_RBBM_CLOCK_DELAY_SP0, 0x00000080},
474 	{REG_A6XX_RBBM_CLOCK_HYST_SP0, 0x0000F3CF},
475 	{REG_A6XX_RBBM_CLOCK_CNTL_TP0, 0x22222222},
476 	{REG_A6XX_RBBM_CLOCK_CNTL2_TP0, 0x22222222},
477 	{REG_A6XX_RBBM_CLOCK_CNTL3_TP0, 0x22222222},
478 	{REG_A6XX_RBBM_CLOCK_CNTL4_TP0, 0x00022222},
479 	{REG_A6XX_RBBM_CLOCK_DELAY_TP0, 0x11111111},
480 	{REG_A6XX_RBBM_CLOCK_DELAY2_TP0, 0x11111111},
481 	{REG_A6XX_RBBM_CLOCK_DELAY3_TP0, 0x11111111},
482 	{REG_A6XX_RBBM_CLOCK_DELAY4_TP0, 0x00011111},
483 	{REG_A6XX_RBBM_CLOCK_HYST_TP0, 0x77777777},
484 	{REG_A6XX_RBBM_CLOCK_HYST2_TP0, 0x77777777},
485 	{REG_A6XX_RBBM_CLOCK_HYST3_TP0, 0x77777777},
486 	{REG_A6XX_RBBM_CLOCK_HYST4_TP0, 0x00077777},
487 	{REG_A6XX_RBBM_CLOCK_CNTL_RB0, 0x22222222},
488 	{REG_A6XX_RBBM_CLOCK_CNTL2_RB0, 0x01002222},
489 	{REG_A6XX_RBBM_CLOCK_CNTL_CCU0, 0x00002220},
490 	{REG_A6XX_RBBM_CLOCK_HYST_RB_CCU0, 0x00040F00},
491 	{REG_A6XX_RBBM_CLOCK_CNTL_RAC, 0x25222022},
492 	{REG_A6XX_RBBM_CLOCK_CNTL2_RAC, 0x00005555},
493 	{REG_A6XX_RBBM_CLOCK_DELAY_RAC, 0x00000011},
494 	{REG_A6XX_RBBM_CLOCK_HYST_RAC, 0x00445044},
495 	{REG_A6XX_RBBM_CLOCK_CNTL_TSE_RAS_RBBM, 0x04222222},
496 	{REG_A6XX_RBBM_CLOCK_MODE_VFD, 0x00002222},
497 	{REG_A6XX_RBBM_CLOCK_MODE_GPC, 0x00222222},
498 	{REG_A6XX_RBBM_CLOCK_DELAY_HLSQ_2, 0x00000002},
499 	{REG_A6XX_RBBM_CLOCK_MODE_HLSQ, 0x00002222},
500 	{REG_A6XX_RBBM_CLOCK_DELAY_TSE_RAS_RBBM, 0x00004000},
501 	{REG_A6XX_RBBM_CLOCK_DELAY_VFD, 0x00002222},
502 	{REG_A6XX_RBBM_CLOCK_DELAY_GPC, 0x00000200},
503 	{REG_A6XX_RBBM_CLOCK_DELAY_HLSQ, 0x00000000},
504 	{REG_A6XX_RBBM_CLOCK_HYST_TSE_RAS_RBBM, 0x00000000},
505 	{REG_A6XX_RBBM_CLOCK_HYST_VFD, 0x00000000},
506 	{REG_A6XX_RBBM_CLOCK_HYST_GPC, 0x04104004},
507 	{REG_A6XX_RBBM_CLOCK_HYST_HLSQ, 0x00000000},
508 	{REG_A6XX_RBBM_CLOCK_CNTL_TEX_FCHE, 0x00000222},
509 	{REG_A6XX_RBBM_CLOCK_DELAY_TEX_FCHE, 0x00000111},
510 	{REG_A6XX_RBBM_CLOCK_HYST_TEX_FCHE, 0x00000000},
511 	{REG_A6XX_RBBM_CLOCK_CNTL_UCHE, 0x22222222},
512 	{REG_A6XX_RBBM_CLOCK_HYST_UCHE, 0x00000004},
513 	{REG_A6XX_RBBM_CLOCK_DELAY_UCHE, 0x00000002},
514 	{REG_A6XX_RBBM_ISDB_CNT, 0x00000182},
515 	{REG_A6XX_RBBM_RAC_THRESHOLD_CNT, 0x00000000},
516 	{REG_A6XX_RBBM_SP_HYST_CNT, 0x00000000},
517 	{REG_A6XX_RBBM_CLOCK_CNTL_GMU_GX, 0x00000222},
518 	{REG_A6XX_RBBM_CLOCK_DELAY_GMU_GX, 0x00000111},
519 	{REG_A6XX_RBBM_CLOCK_HYST_GMU_GX, 0x00000555},
520 	{},
521 };
522 
523 static void a6xx_set_hwcg(struct msm_gpu *gpu, bool state)
524 {
525 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
526 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
527 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
528 	const struct adreno_reglist *reg;
529 	unsigned int i;
530 	u32 val, clock_cntl_on;
531 
532 	if (!adreno_gpu->info->hwcg)
533 		return;
534 
535 	if (adreno_is_a630(adreno_gpu))
536 		clock_cntl_on = 0x8aa8aa02;
537 	else
538 		clock_cntl_on = 0x8aa8aa82;
539 
540 	val = gpu_read(gpu, REG_A6XX_RBBM_CLOCK_CNTL);
541 
542 	/* Don't re-program the registers if they are already correct */
543 	if ((!state && !val) || (state && (val == clock_cntl_on)))
544 		return;
545 
546 	/* Disable SP clock before programming HWCG registers */
547 	gmu_rmw(gmu, REG_A6XX_GPU_GMU_GX_SPTPRAC_CLOCK_CONTROL, 1, 0);
548 
549 	for (i = 0; (reg = &adreno_gpu->info->hwcg[i], reg->offset); i++)
550 		gpu_write(gpu, reg->offset, state ? reg->value : 0);
551 
552 	/* Enable SP clock */
553 	gmu_rmw(gmu, REG_A6XX_GPU_GMU_GX_SPTPRAC_CLOCK_CONTROL, 0, 1);
554 
555 	gpu_write(gpu, REG_A6XX_RBBM_CLOCK_CNTL, state ? clock_cntl_on : 0);
556 }
557 
558 /* For a615, a616, a618, A619, a630, a640 and a680 */
559 static const u32 a6xx_protect[] = {
560 	A6XX_PROTECT_RDONLY(0x00000, 0x04ff),
561 	A6XX_PROTECT_RDONLY(0x00501, 0x0005),
562 	A6XX_PROTECT_RDONLY(0x0050b, 0x02f4),
563 	A6XX_PROTECT_NORDWR(0x0050e, 0x0000),
564 	A6XX_PROTECT_NORDWR(0x00510, 0x0000),
565 	A6XX_PROTECT_NORDWR(0x00534, 0x0000),
566 	A6XX_PROTECT_NORDWR(0x00800, 0x0082),
567 	A6XX_PROTECT_NORDWR(0x008a0, 0x0008),
568 	A6XX_PROTECT_NORDWR(0x008ab, 0x0024),
569 	A6XX_PROTECT_RDONLY(0x008de, 0x00ae),
570 	A6XX_PROTECT_NORDWR(0x00900, 0x004d),
571 	A6XX_PROTECT_NORDWR(0x0098d, 0x0272),
572 	A6XX_PROTECT_NORDWR(0x00e00, 0x0001),
573 	A6XX_PROTECT_NORDWR(0x00e03, 0x000c),
574 	A6XX_PROTECT_NORDWR(0x03c00, 0x00c3),
575 	A6XX_PROTECT_RDONLY(0x03cc4, 0x1fff),
576 	A6XX_PROTECT_NORDWR(0x08630, 0x01cf),
577 	A6XX_PROTECT_NORDWR(0x08e00, 0x0000),
578 	A6XX_PROTECT_NORDWR(0x08e08, 0x0000),
579 	A6XX_PROTECT_NORDWR(0x08e50, 0x001f),
580 	A6XX_PROTECT_NORDWR(0x09624, 0x01db),
581 	A6XX_PROTECT_NORDWR(0x09e70, 0x0001),
582 	A6XX_PROTECT_NORDWR(0x09e78, 0x0187),
583 	A6XX_PROTECT_NORDWR(0x0a630, 0x01cf),
584 	A6XX_PROTECT_NORDWR(0x0ae02, 0x0000),
585 	A6XX_PROTECT_NORDWR(0x0ae50, 0x032f),
586 	A6XX_PROTECT_NORDWR(0x0b604, 0x0000),
587 	A6XX_PROTECT_NORDWR(0x0be02, 0x0001),
588 	A6XX_PROTECT_NORDWR(0x0be20, 0x17df),
589 	A6XX_PROTECT_NORDWR(0x0f000, 0x0bff),
590 	A6XX_PROTECT_RDONLY(0x0fc00, 0x1fff),
591 	A6XX_PROTECT_NORDWR(0x11c00, 0x0000), /* note: infinite range */
592 };
593 
594 /* These are for a620 and a650 */
595 static const u32 a650_protect[] = {
596 	A6XX_PROTECT_RDONLY(0x00000, 0x04ff),
597 	A6XX_PROTECT_RDONLY(0x00501, 0x0005),
598 	A6XX_PROTECT_RDONLY(0x0050b, 0x02f4),
599 	A6XX_PROTECT_NORDWR(0x0050e, 0x0000),
600 	A6XX_PROTECT_NORDWR(0x00510, 0x0000),
601 	A6XX_PROTECT_NORDWR(0x00534, 0x0000),
602 	A6XX_PROTECT_NORDWR(0x00800, 0x0082),
603 	A6XX_PROTECT_NORDWR(0x008a0, 0x0008),
604 	A6XX_PROTECT_NORDWR(0x008ab, 0x0024),
605 	A6XX_PROTECT_RDONLY(0x008de, 0x00ae),
606 	A6XX_PROTECT_NORDWR(0x00900, 0x004d),
607 	A6XX_PROTECT_NORDWR(0x0098d, 0x0272),
608 	A6XX_PROTECT_NORDWR(0x00e00, 0x0001),
609 	A6XX_PROTECT_NORDWR(0x00e03, 0x000c),
610 	A6XX_PROTECT_NORDWR(0x03c00, 0x00c3),
611 	A6XX_PROTECT_RDONLY(0x03cc4, 0x1fff),
612 	A6XX_PROTECT_NORDWR(0x08630, 0x01cf),
613 	A6XX_PROTECT_NORDWR(0x08e00, 0x0000),
614 	A6XX_PROTECT_NORDWR(0x08e08, 0x0000),
615 	A6XX_PROTECT_NORDWR(0x08e50, 0x001f),
616 	A6XX_PROTECT_NORDWR(0x08e80, 0x027f),
617 	A6XX_PROTECT_NORDWR(0x09624, 0x01db),
618 	A6XX_PROTECT_NORDWR(0x09e60, 0x0011),
619 	A6XX_PROTECT_NORDWR(0x09e78, 0x0187),
620 	A6XX_PROTECT_NORDWR(0x0a630, 0x01cf),
621 	A6XX_PROTECT_NORDWR(0x0ae02, 0x0000),
622 	A6XX_PROTECT_NORDWR(0x0ae50, 0x032f),
623 	A6XX_PROTECT_NORDWR(0x0b604, 0x0000),
624 	A6XX_PROTECT_NORDWR(0x0b608, 0x0007),
625 	A6XX_PROTECT_NORDWR(0x0be02, 0x0001),
626 	A6XX_PROTECT_NORDWR(0x0be20, 0x17df),
627 	A6XX_PROTECT_NORDWR(0x0f000, 0x0bff),
628 	A6XX_PROTECT_RDONLY(0x0fc00, 0x1fff),
629 	A6XX_PROTECT_NORDWR(0x18400, 0x1fff),
630 	A6XX_PROTECT_NORDWR(0x1a800, 0x1fff),
631 	A6XX_PROTECT_NORDWR(0x1f400, 0x0443),
632 	A6XX_PROTECT_RDONLY(0x1f844, 0x007b),
633 	A6XX_PROTECT_NORDWR(0x1f887, 0x001b),
634 	A6XX_PROTECT_NORDWR(0x1f8c0, 0x0000), /* note: infinite range */
635 };
636 
637 /* These are for a635 and a660 */
638 static const u32 a660_protect[] = {
639 	A6XX_PROTECT_RDONLY(0x00000, 0x04ff),
640 	A6XX_PROTECT_RDONLY(0x00501, 0x0005),
641 	A6XX_PROTECT_RDONLY(0x0050b, 0x02f4),
642 	A6XX_PROTECT_NORDWR(0x0050e, 0x0000),
643 	A6XX_PROTECT_NORDWR(0x00510, 0x0000),
644 	A6XX_PROTECT_NORDWR(0x00534, 0x0000),
645 	A6XX_PROTECT_NORDWR(0x00800, 0x0082),
646 	A6XX_PROTECT_NORDWR(0x008a0, 0x0008),
647 	A6XX_PROTECT_NORDWR(0x008ab, 0x0024),
648 	A6XX_PROTECT_RDONLY(0x008de, 0x00ae),
649 	A6XX_PROTECT_NORDWR(0x00900, 0x004d),
650 	A6XX_PROTECT_NORDWR(0x0098d, 0x0272),
651 	A6XX_PROTECT_NORDWR(0x00e00, 0x0001),
652 	A6XX_PROTECT_NORDWR(0x00e03, 0x000c),
653 	A6XX_PROTECT_NORDWR(0x03c00, 0x00c3),
654 	A6XX_PROTECT_RDONLY(0x03cc4, 0x1fff),
655 	A6XX_PROTECT_NORDWR(0x08630, 0x01cf),
656 	A6XX_PROTECT_NORDWR(0x08e00, 0x0000),
657 	A6XX_PROTECT_NORDWR(0x08e08, 0x0000),
658 	A6XX_PROTECT_NORDWR(0x08e50, 0x001f),
659 	A6XX_PROTECT_NORDWR(0x08e80, 0x027f),
660 	A6XX_PROTECT_NORDWR(0x09624, 0x01db),
661 	A6XX_PROTECT_NORDWR(0x09e60, 0x0011),
662 	A6XX_PROTECT_NORDWR(0x09e78, 0x0187),
663 	A6XX_PROTECT_NORDWR(0x0a630, 0x01cf),
664 	A6XX_PROTECT_NORDWR(0x0ae02, 0x0000),
665 	A6XX_PROTECT_NORDWR(0x0ae50, 0x012f),
666 	A6XX_PROTECT_NORDWR(0x0b604, 0x0000),
667 	A6XX_PROTECT_NORDWR(0x0b608, 0x0006),
668 	A6XX_PROTECT_NORDWR(0x0be02, 0x0001),
669 	A6XX_PROTECT_NORDWR(0x0be20, 0x015f),
670 	A6XX_PROTECT_NORDWR(0x0d000, 0x05ff),
671 	A6XX_PROTECT_NORDWR(0x0f000, 0x0bff),
672 	A6XX_PROTECT_RDONLY(0x0fc00, 0x1fff),
673 	A6XX_PROTECT_NORDWR(0x18400, 0x1fff),
674 	A6XX_PROTECT_NORDWR(0x1a400, 0x1fff),
675 	A6XX_PROTECT_NORDWR(0x1f400, 0x0443),
676 	A6XX_PROTECT_RDONLY(0x1f844, 0x007b),
677 	A6XX_PROTECT_NORDWR(0x1f860, 0x0000),
678 	A6XX_PROTECT_NORDWR(0x1f887, 0x001b),
679 	A6XX_PROTECT_NORDWR(0x1f8c0, 0x0000), /* note: infinite range */
680 };
681 
682 static void a6xx_set_cp_protect(struct msm_gpu *gpu)
683 {
684 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
685 	const u32 *regs = a6xx_protect;
686 	unsigned i, count, count_max;
687 
688 	if (adreno_is_a650(adreno_gpu)) {
689 		regs = a650_protect;
690 		count = ARRAY_SIZE(a650_protect);
691 		count_max = 48;
692 		BUILD_BUG_ON(ARRAY_SIZE(a650_protect) > 48);
693 	} else if (adreno_is_a660_family(adreno_gpu)) {
694 		regs = a660_protect;
695 		count = ARRAY_SIZE(a660_protect);
696 		count_max = 48;
697 		BUILD_BUG_ON(ARRAY_SIZE(a660_protect) > 48);
698 	} else {
699 		regs = a6xx_protect;
700 		count = ARRAY_SIZE(a6xx_protect);
701 		count_max = 32;
702 		BUILD_BUG_ON(ARRAY_SIZE(a6xx_protect) > 32);
703 	}
704 
705 	/*
706 	 * Enable access protection to privileged registers, fault on an access
707 	 * protect violation and select the last span to protect from the start
708 	 * address all the way to the end of the register address space
709 	 */
710 	gpu_write(gpu, REG_A6XX_CP_PROTECT_CNTL, BIT(0) | BIT(1) | BIT(3));
711 
712 	for (i = 0; i < count - 1; i++)
713 		gpu_write(gpu, REG_A6XX_CP_PROTECT(i), regs[i]);
714 	/* last CP_PROTECT to have "infinite" length on the last entry */
715 	gpu_write(gpu, REG_A6XX_CP_PROTECT(count_max - 1), regs[i]);
716 }
717 
718 static void a6xx_set_ubwc_config(struct msm_gpu *gpu)
719 {
720 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
721 	u32 lower_bit = 2;
722 	u32 amsbc = 0;
723 	u32 rgb565_predicator = 0;
724 	u32 uavflagprd_inv = 0;
725 
726 	/* a618 is using the hw default values */
727 	if (adreno_is_a618(adreno_gpu))
728 		return;
729 
730 	if (adreno_is_a640_family(adreno_gpu))
731 		amsbc = 1;
732 
733 	if (adreno_is_a650(adreno_gpu) || adreno_is_a660(adreno_gpu)) {
734 		/* TODO: get ddr type from bootloader and use 2 for LPDDR4 */
735 		lower_bit = 3;
736 		amsbc = 1;
737 		rgb565_predicator = 1;
738 		uavflagprd_inv = 2;
739 	}
740 
741 	if (adreno_is_7c3(adreno_gpu)) {
742 		lower_bit = 1;
743 		amsbc = 1;
744 		rgb565_predicator = 1;
745 		uavflagprd_inv = 2;
746 	}
747 
748 	gpu_write(gpu, REG_A6XX_RB_NC_MODE_CNTL,
749 		rgb565_predicator << 11 | amsbc << 4 | lower_bit << 1);
750 	gpu_write(gpu, REG_A6XX_TPL1_NC_MODE_CNTL, lower_bit << 1);
751 	gpu_write(gpu, REG_A6XX_SP_NC_MODE_CNTL,
752 		uavflagprd_inv << 4 | lower_bit << 1);
753 	gpu_write(gpu, REG_A6XX_UCHE_MODE_CNTL, lower_bit << 21);
754 }
755 
756 static int a6xx_cp_init(struct msm_gpu *gpu)
757 {
758 	struct msm_ringbuffer *ring = gpu->rb[0];
759 
760 	OUT_PKT7(ring, CP_ME_INIT, 8);
761 
762 	OUT_RING(ring, 0x0000002f);
763 
764 	/* Enable multiple hardware contexts */
765 	OUT_RING(ring, 0x00000003);
766 
767 	/* Enable error detection */
768 	OUT_RING(ring, 0x20000000);
769 
770 	/* Don't enable header dump */
771 	OUT_RING(ring, 0x00000000);
772 	OUT_RING(ring, 0x00000000);
773 
774 	/* No workarounds enabled */
775 	OUT_RING(ring, 0x00000000);
776 
777 	/* Pad rest of the cmds with 0's */
778 	OUT_RING(ring, 0x00000000);
779 	OUT_RING(ring, 0x00000000);
780 
781 	a6xx_flush(gpu, ring);
782 	return a6xx_idle(gpu, ring) ? 0 : -EINVAL;
783 }
784 
785 /*
786  * Check that the microcode version is new enough to include several key
787  * security fixes. Return true if the ucode is safe.
788  */
789 static bool a6xx_ucode_check_version(struct a6xx_gpu *a6xx_gpu,
790 		struct drm_gem_object *obj)
791 {
792 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
793 	struct msm_gpu *gpu = &adreno_gpu->base;
794 	const char *sqe_name = adreno_gpu->info->fw[ADRENO_FW_SQE];
795 	u32 *buf = msm_gem_get_vaddr(obj);
796 	bool ret = false;
797 
798 	if (IS_ERR(buf))
799 		return false;
800 
801 	/*
802 	 * Targets up to a640 (a618, a630 and a640) need to check for a
803 	 * microcode version that is patched to support the whereami opcode or
804 	 * one that is new enough to include it by default.
805 	 *
806 	 * a650 tier targets don't need whereami but still need to be
807 	 * equal to or newer than 0.95 for other security fixes
808 	 *
809 	 * a660 targets have all the critical security fixes from the start
810 	 */
811 	if (!strcmp(sqe_name, "a630_sqe.fw")) {
812 		/*
813 		 * If the lowest nibble is 0xa that is an indication that this
814 		 * microcode has been patched. The actual version is in dword
815 		 * [3] but we only care about the patchlevel which is the lowest
816 		 * nibble of dword [3]
817 		 *
818 		 * Otherwise check that the firmware is greater than or equal
819 		 * to 1.90 which was the first version that had this fix built
820 		 * in
821 		 */
822 		if ((((buf[0] & 0xf) == 0xa) && (buf[2] & 0xf) >= 1) ||
823 			(buf[0] & 0xfff) >= 0x190) {
824 			a6xx_gpu->has_whereami = true;
825 			ret = true;
826 			goto out;
827 		}
828 
829 		DRM_DEV_ERROR(&gpu->pdev->dev,
830 			"a630 SQE ucode is too old. Have version %x need at least %x\n",
831 			buf[0] & 0xfff, 0x190);
832 	} else if (!strcmp(sqe_name, "a650_sqe.fw")) {
833 		if ((buf[0] & 0xfff) >= 0x095) {
834 			ret = true;
835 			goto out;
836 		}
837 
838 		DRM_DEV_ERROR(&gpu->pdev->dev,
839 			"a650 SQE ucode is too old. Have version %x need at least %x\n",
840 			buf[0] & 0xfff, 0x095);
841 	} else if (!strcmp(sqe_name, "a660_sqe.fw")) {
842 		ret = true;
843 	} else {
844 		DRM_DEV_ERROR(&gpu->pdev->dev,
845 			"unknown GPU, add it to a6xx_ucode_check_version()!!\n");
846 	}
847 out:
848 	msm_gem_put_vaddr(obj);
849 	return ret;
850 }
851 
852 static int a6xx_ucode_init(struct msm_gpu *gpu)
853 {
854 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
855 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
856 
857 	if (!a6xx_gpu->sqe_bo) {
858 		a6xx_gpu->sqe_bo = adreno_fw_create_bo(gpu,
859 			adreno_gpu->fw[ADRENO_FW_SQE], &a6xx_gpu->sqe_iova);
860 
861 		if (IS_ERR(a6xx_gpu->sqe_bo)) {
862 			int ret = PTR_ERR(a6xx_gpu->sqe_bo);
863 
864 			a6xx_gpu->sqe_bo = NULL;
865 			DRM_DEV_ERROR(&gpu->pdev->dev,
866 				"Could not allocate SQE ucode: %d\n", ret);
867 
868 			return ret;
869 		}
870 
871 		msm_gem_object_set_name(a6xx_gpu->sqe_bo, "sqefw");
872 		if (!a6xx_ucode_check_version(a6xx_gpu, a6xx_gpu->sqe_bo)) {
873 			msm_gem_unpin_iova(a6xx_gpu->sqe_bo, gpu->aspace);
874 			drm_gem_object_put(a6xx_gpu->sqe_bo);
875 
876 			a6xx_gpu->sqe_bo = NULL;
877 			return -EPERM;
878 		}
879 	}
880 
881 	gpu_write64(gpu, REG_A6XX_CP_SQE_INSTR_BASE,
882 		REG_A6XX_CP_SQE_INSTR_BASE+1, a6xx_gpu->sqe_iova);
883 
884 	return 0;
885 }
886 
887 static int a6xx_zap_shader_init(struct msm_gpu *gpu)
888 {
889 	static bool loaded;
890 	int ret;
891 
892 	if (loaded)
893 		return 0;
894 
895 	ret = adreno_zap_shader_load(gpu, GPU_PAS_ID);
896 
897 	loaded = !ret;
898 	return ret;
899 }
900 
901 #define A6XX_INT_MASK (A6XX_RBBM_INT_0_MASK_CP_AHB_ERROR | \
902 	  A6XX_RBBM_INT_0_MASK_RBBM_ATB_ASYNCFIFO_OVERFLOW | \
903 	  A6XX_RBBM_INT_0_MASK_CP_HW_ERROR | \
904 	  A6XX_RBBM_INT_0_MASK_CP_IB2 | \
905 	  A6XX_RBBM_INT_0_MASK_CP_IB1 | \
906 	  A6XX_RBBM_INT_0_MASK_CP_RB | \
907 	  A6XX_RBBM_INT_0_MASK_CP_CACHE_FLUSH_TS | \
908 	  A6XX_RBBM_INT_0_MASK_RBBM_ATB_BUS_OVERFLOW | \
909 	  A6XX_RBBM_INT_0_MASK_RBBM_HANG_DETECT | \
910 	  A6XX_RBBM_INT_0_MASK_UCHE_OOB_ACCESS | \
911 	  A6XX_RBBM_INT_0_MASK_UCHE_TRAP_INTR)
912 
913 static int hw_init(struct msm_gpu *gpu)
914 {
915 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
916 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
917 	int ret;
918 
919 	/* Make sure the GMU keeps the GPU on while we set it up */
920 	a6xx_gmu_set_oob(&a6xx_gpu->gmu, GMU_OOB_GPU_SET);
921 
922 	gpu_write(gpu, REG_A6XX_RBBM_SECVID_TSB_CNTL, 0);
923 
924 	/*
925 	 * Disable the trusted memory range - we don't actually supported secure
926 	 * memory rendering at this point in time and we don't want to block off
927 	 * part of the virtual memory space.
928 	 */
929 	gpu_write64(gpu, REG_A6XX_RBBM_SECVID_TSB_TRUSTED_BASE_LO,
930 		REG_A6XX_RBBM_SECVID_TSB_TRUSTED_BASE_HI, 0x00000000);
931 	gpu_write(gpu, REG_A6XX_RBBM_SECVID_TSB_TRUSTED_SIZE, 0x00000000);
932 
933 	/* Turn on 64 bit addressing for all blocks */
934 	gpu_write(gpu, REG_A6XX_CP_ADDR_MODE_CNTL, 0x1);
935 	gpu_write(gpu, REG_A6XX_VSC_ADDR_MODE_CNTL, 0x1);
936 	gpu_write(gpu, REG_A6XX_GRAS_ADDR_MODE_CNTL, 0x1);
937 	gpu_write(gpu, REG_A6XX_RB_ADDR_MODE_CNTL, 0x1);
938 	gpu_write(gpu, REG_A6XX_PC_ADDR_MODE_CNTL, 0x1);
939 	gpu_write(gpu, REG_A6XX_HLSQ_ADDR_MODE_CNTL, 0x1);
940 	gpu_write(gpu, REG_A6XX_VFD_ADDR_MODE_CNTL, 0x1);
941 	gpu_write(gpu, REG_A6XX_VPC_ADDR_MODE_CNTL, 0x1);
942 	gpu_write(gpu, REG_A6XX_UCHE_ADDR_MODE_CNTL, 0x1);
943 	gpu_write(gpu, REG_A6XX_SP_ADDR_MODE_CNTL, 0x1);
944 	gpu_write(gpu, REG_A6XX_TPL1_ADDR_MODE_CNTL, 0x1);
945 	gpu_write(gpu, REG_A6XX_RBBM_SECVID_TSB_ADDR_MODE_CNTL, 0x1);
946 
947 	/* enable hardware clockgating */
948 	a6xx_set_hwcg(gpu, true);
949 
950 	/* VBIF/GBIF start*/
951 	if (adreno_is_a640_family(adreno_gpu) ||
952 	    adreno_is_a650_family(adreno_gpu)) {
953 		gpu_write(gpu, REG_A6XX_GBIF_QSB_SIDE0, 0x00071620);
954 		gpu_write(gpu, REG_A6XX_GBIF_QSB_SIDE1, 0x00071620);
955 		gpu_write(gpu, REG_A6XX_GBIF_QSB_SIDE2, 0x00071620);
956 		gpu_write(gpu, REG_A6XX_GBIF_QSB_SIDE3, 0x00071620);
957 		gpu_write(gpu, REG_A6XX_GBIF_QSB_SIDE3, 0x00071620);
958 		gpu_write(gpu, REG_A6XX_RBBM_GBIF_CLIENT_QOS_CNTL, 0x3);
959 	} else {
960 		gpu_write(gpu, REG_A6XX_RBBM_VBIF_CLIENT_QOS_CNTL, 0x3);
961 	}
962 
963 	if (adreno_is_a630(adreno_gpu))
964 		gpu_write(gpu, REG_A6XX_VBIF_GATE_OFF_WRREQ_EN, 0x00000009);
965 
966 	/* Make all blocks contribute to the GPU BUSY perf counter */
967 	gpu_write(gpu, REG_A6XX_RBBM_PERFCTR_GPU_BUSY_MASKED, 0xffffffff);
968 
969 	/* Disable L2 bypass in the UCHE */
970 	gpu_write(gpu, REG_A6XX_UCHE_WRITE_RANGE_MAX_LO, 0xffffffc0);
971 	gpu_write(gpu, REG_A6XX_UCHE_WRITE_RANGE_MAX_HI, 0x0001ffff);
972 	gpu_write(gpu, REG_A6XX_UCHE_TRAP_BASE_LO, 0xfffff000);
973 	gpu_write(gpu, REG_A6XX_UCHE_TRAP_BASE_HI, 0x0001ffff);
974 	gpu_write(gpu, REG_A6XX_UCHE_WRITE_THRU_BASE_LO, 0xfffff000);
975 	gpu_write(gpu, REG_A6XX_UCHE_WRITE_THRU_BASE_HI, 0x0001ffff);
976 
977 	if (!adreno_is_a650_family(adreno_gpu)) {
978 		/* Set the GMEM VA range [0x100000:0x100000 + gpu->gmem - 1] */
979 		gpu_write64(gpu, REG_A6XX_UCHE_GMEM_RANGE_MIN_LO,
980 			REG_A6XX_UCHE_GMEM_RANGE_MIN_HI, 0x00100000);
981 
982 		gpu_write64(gpu, REG_A6XX_UCHE_GMEM_RANGE_MAX_LO,
983 			REG_A6XX_UCHE_GMEM_RANGE_MAX_HI,
984 			0x00100000 + adreno_gpu->gmem - 1);
985 	}
986 
987 	gpu_write(gpu, REG_A6XX_UCHE_FILTER_CNTL, 0x804);
988 	gpu_write(gpu, REG_A6XX_UCHE_CACHE_WAYS, 0x4);
989 
990 	if (adreno_is_a640_family(adreno_gpu) ||
991 	    adreno_is_a650_family(adreno_gpu))
992 		gpu_write(gpu, REG_A6XX_CP_ROQ_THRESHOLDS_2, 0x02000140);
993 	else
994 		gpu_write(gpu, REG_A6XX_CP_ROQ_THRESHOLDS_2, 0x010000c0);
995 	gpu_write(gpu, REG_A6XX_CP_ROQ_THRESHOLDS_1, 0x8040362c);
996 
997 	if (adreno_is_a660_family(adreno_gpu))
998 		gpu_write(gpu, REG_A6XX_CP_LPAC_PROG_FIFO_SIZE, 0x00000020);
999 
1000 	/* Setting the mem pool size */
1001 	gpu_write(gpu, REG_A6XX_CP_MEM_POOL_SIZE, 128);
1002 
1003 	/* Setting the primFifo thresholds default values,
1004 	 * and vccCacheSkipDis=1 bit (0x200) for A640 and newer
1005 	*/
1006 	if (adreno_is_a650(adreno_gpu) || adreno_is_a660(adreno_gpu))
1007 		gpu_write(gpu, REG_A6XX_PC_DBG_ECO_CNTL, 0x00300200);
1008 	else if (adreno_is_a640_family(adreno_gpu) || adreno_is_7c3(adreno_gpu))
1009 		gpu_write(gpu, REG_A6XX_PC_DBG_ECO_CNTL, 0x00200200);
1010 	else if (adreno_is_a650(adreno_gpu) || adreno_is_a660(adreno_gpu))
1011 		gpu_write(gpu, REG_A6XX_PC_DBG_ECO_CNTL, 0x00300200);
1012 	else
1013 		gpu_write(gpu, REG_A6XX_PC_DBG_ECO_CNTL, 0x00180000);
1014 
1015 	/* Set the AHB default slave response to "ERROR" */
1016 	gpu_write(gpu, REG_A6XX_CP_AHB_CNTL, 0x1);
1017 
1018 	/* Turn on performance counters */
1019 	gpu_write(gpu, REG_A6XX_RBBM_PERFCTR_CNTL, 0x1);
1020 
1021 	/* Select CP0 to always count cycles */
1022 	gpu_write(gpu, REG_A6XX_CP_PERFCTR_CP_SEL(0), PERF_CP_ALWAYS_COUNT);
1023 
1024 	a6xx_set_ubwc_config(gpu);
1025 
1026 	/* Enable fault detection */
1027 	gpu_write(gpu, REG_A6XX_RBBM_INTERFACE_HANG_INT_CNTL,
1028 		(1 << 30) | 0x1fffff);
1029 
1030 	gpu_write(gpu, REG_A6XX_UCHE_CLIENT_PF, 1);
1031 
1032 	/* Set weights for bicubic filtering */
1033 	if (adreno_is_a650_family(adreno_gpu)) {
1034 		gpu_write(gpu, REG_A6XX_TPL1_BICUBIC_WEIGHTS_TABLE_0, 0);
1035 		gpu_write(gpu, REG_A6XX_TPL1_BICUBIC_WEIGHTS_TABLE_1,
1036 			0x3fe05ff4);
1037 		gpu_write(gpu, REG_A6XX_TPL1_BICUBIC_WEIGHTS_TABLE_2,
1038 			0x3fa0ebee);
1039 		gpu_write(gpu, REG_A6XX_TPL1_BICUBIC_WEIGHTS_TABLE_3,
1040 			0x3f5193ed);
1041 		gpu_write(gpu, REG_A6XX_TPL1_BICUBIC_WEIGHTS_TABLE_4,
1042 			0x3f0243f0);
1043 	}
1044 
1045 	/* Protect registers from the CP */
1046 	a6xx_set_cp_protect(gpu);
1047 
1048 	if (adreno_is_a660_family(adreno_gpu)) {
1049 		gpu_write(gpu, REG_A6XX_CP_CHICKEN_DBG, 0x1);
1050 		gpu_write(gpu, REG_A6XX_RBBM_GBIF_CLIENT_QOS_CNTL, 0x0);
1051 	}
1052 
1053 	/* Set dualQ + disable afull for A660 GPU */
1054 	if (adreno_is_a660(adreno_gpu))
1055 		gpu_write(gpu, REG_A6XX_UCHE_CMDQ_CONFIG, 0x66906);
1056 
1057 	/* Enable expanded apriv for targets that support it */
1058 	if (gpu->hw_apriv) {
1059 		gpu_write(gpu, REG_A6XX_CP_APRIV_CNTL,
1060 			(1 << 6) | (1 << 5) | (1 << 3) | (1 << 2) | (1 << 1));
1061 	}
1062 
1063 	/* Enable interrupts */
1064 	gpu_write(gpu, REG_A6XX_RBBM_INT_0_MASK, A6XX_INT_MASK);
1065 
1066 	ret = adreno_hw_init(gpu);
1067 	if (ret)
1068 		goto out;
1069 
1070 	ret = a6xx_ucode_init(gpu);
1071 	if (ret)
1072 		goto out;
1073 
1074 	/* Set the ringbuffer address */
1075 	gpu_write64(gpu, REG_A6XX_CP_RB_BASE, REG_A6XX_CP_RB_BASE_HI,
1076 		gpu->rb[0]->iova);
1077 
1078 	/* Targets that support extended APRIV can use the RPTR shadow from
1079 	 * hardware but all the other ones need to disable the feature. Targets
1080 	 * that support the WHERE_AM_I opcode can use that instead
1081 	 */
1082 	if (adreno_gpu->base.hw_apriv)
1083 		gpu_write(gpu, REG_A6XX_CP_RB_CNTL, MSM_GPU_RB_CNTL_DEFAULT);
1084 	else
1085 		gpu_write(gpu, REG_A6XX_CP_RB_CNTL,
1086 			MSM_GPU_RB_CNTL_DEFAULT | AXXX_CP_RB_CNTL_NO_UPDATE);
1087 
1088 	/*
1089 	 * Expanded APRIV and targets that support WHERE_AM_I both need a
1090 	 * privileged buffer to store the RPTR shadow
1091 	 */
1092 
1093 	if (adreno_gpu->base.hw_apriv || a6xx_gpu->has_whereami) {
1094 		if (!a6xx_gpu->shadow_bo) {
1095 			a6xx_gpu->shadow = msm_gem_kernel_new(gpu->dev,
1096 				sizeof(u32) * gpu->nr_rings,
1097 				MSM_BO_WC | MSM_BO_MAP_PRIV,
1098 				gpu->aspace, &a6xx_gpu->shadow_bo,
1099 				&a6xx_gpu->shadow_iova);
1100 
1101 			if (IS_ERR(a6xx_gpu->shadow))
1102 				return PTR_ERR(a6xx_gpu->shadow);
1103 
1104 			msm_gem_object_set_name(a6xx_gpu->shadow_bo, "shadow");
1105 		}
1106 
1107 		gpu_write64(gpu, REG_A6XX_CP_RB_RPTR_ADDR_LO,
1108 			REG_A6XX_CP_RB_RPTR_ADDR_HI,
1109 			shadowptr(a6xx_gpu, gpu->rb[0]));
1110 	}
1111 
1112 	/* Always come up on rb 0 */
1113 	a6xx_gpu->cur_ring = gpu->rb[0];
1114 
1115 	gpu->cur_ctx_seqno = 0;
1116 
1117 	/* Enable the SQE_to start the CP engine */
1118 	gpu_write(gpu, REG_A6XX_CP_SQE_CNTL, 1);
1119 
1120 	ret = a6xx_cp_init(gpu);
1121 	if (ret)
1122 		goto out;
1123 
1124 	/*
1125 	 * Try to load a zap shader into the secure world. If successful
1126 	 * we can use the CP to switch out of secure mode. If not then we
1127 	 * have no resource but to try to switch ourselves out manually. If we
1128 	 * guessed wrong then access to the RBBM_SECVID_TRUST_CNTL register will
1129 	 * be blocked and a permissions violation will soon follow.
1130 	 */
1131 	ret = a6xx_zap_shader_init(gpu);
1132 	if (!ret) {
1133 		OUT_PKT7(gpu->rb[0], CP_SET_SECURE_MODE, 1);
1134 		OUT_RING(gpu->rb[0], 0x00000000);
1135 
1136 		a6xx_flush(gpu, gpu->rb[0]);
1137 		if (!a6xx_idle(gpu, gpu->rb[0]))
1138 			return -EINVAL;
1139 	} else if (ret == -ENODEV) {
1140 		/*
1141 		 * This device does not use zap shader (but print a warning
1142 		 * just in case someone got their dt wrong.. hopefully they
1143 		 * have a debug UART to realize the error of their ways...
1144 		 * if you mess this up you are about to crash horribly)
1145 		 */
1146 		dev_warn_once(gpu->dev->dev,
1147 			"Zap shader not enabled - using SECVID_TRUST_CNTL instead\n");
1148 		gpu_write(gpu, REG_A6XX_RBBM_SECVID_TRUST_CNTL, 0x0);
1149 		ret = 0;
1150 	} else {
1151 		return ret;
1152 	}
1153 
1154 out:
1155 	/*
1156 	 * Tell the GMU that we are done touching the GPU and it can start power
1157 	 * management
1158 	 */
1159 	a6xx_gmu_clear_oob(&a6xx_gpu->gmu, GMU_OOB_GPU_SET);
1160 
1161 	if (a6xx_gpu->gmu.legacy) {
1162 		/* Take the GMU out of its special boot mode */
1163 		a6xx_gmu_clear_oob(&a6xx_gpu->gmu, GMU_OOB_BOOT_SLUMBER);
1164 	}
1165 
1166 	return ret;
1167 }
1168 
1169 static int a6xx_hw_init(struct msm_gpu *gpu)
1170 {
1171 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1172 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1173 	int ret;
1174 
1175 	mutex_lock(&a6xx_gpu->gmu.lock);
1176 	ret = hw_init(gpu);
1177 	mutex_unlock(&a6xx_gpu->gmu.lock);
1178 
1179 	return ret;
1180 }
1181 
1182 static void a6xx_dump(struct msm_gpu *gpu)
1183 {
1184 	DRM_DEV_INFO(&gpu->pdev->dev, "status:   %08x\n",
1185 			gpu_read(gpu, REG_A6XX_RBBM_STATUS));
1186 	adreno_dump(gpu);
1187 }
1188 
1189 #define VBIF_RESET_ACK_TIMEOUT	100
1190 #define VBIF_RESET_ACK_MASK	0x00f0
1191 
1192 static void a6xx_recover(struct msm_gpu *gpu)
1193 {
1194 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1195 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1196 	int i;
1197 
1198 	adreno_dump_info(gpu);
1199 
1200 	for (i = 0; i < 8; i++)
1201 		DRM_DEV_INFO(&gpu->pdev->dev, "CP_SCRATCH_REG%d: %u\n", i,
1202 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(i)));
1203 
1204 	if (hang_debug)
1205 		a6xx_dump(gpu);
1206 
1207 	/*
1208 	 * Turn off keep alive that might have been enabled by the hang
1209 	 * interrupt
1210 	 */
1211 	gmu_write(&a6xx_gpu->gmu, REG_A6XX_GMU_GMU_PWR_COL_KEEPALIVE, 0);
1212 
1213 	gpu->funcs->pm_suspend(gpu);
1214 	gpu->funcs->pm_resume(gpu);
1215 
1216 	msm_gpu_hw_init(gpu);
1217 }
1218 
1219 static const char *a6xx_uche_fault_block(struct msm_gpu *gpu, u32 mid)
1220 {
1221 	static const char *uche_clients[7] = {
1222 		"VFD", "SP", "VSC", "VPC", "HLSQ", "PC", "LRZ",
1223 	};
1224 	u32 val;
1225 
1226 	if (mid < 1 || mid > 3)
1227 		return "UNKNOWN";
1228 
1229 	/*
1230 	 * The source of the data depends on the mid ID read from FSYNR1.
1231 	 * and the client ID read from the UCHE block
1232 	 */
1233 	val = gpu_read(gpu, REG_A6XX_UCHE_CLIENT_PF);
1234 
1235 	/* mid = 3 is most precise and refers to only one block per client */
1236 	if (mid == 3)
1237 		return uche_clients[val & 7];
1238 
1239 	/* For mid=2 the source is TP or VFD except when the client id is 0 */
1240 	if (mid == 2)
1241 		return ((val & 7) == 0) ? "TP" : "TP|VFD";
1242 
1243 	/* For mid=1 just return "UCHE" as a catchall for everything else */
1244 	return "UCHE";
1245 }
1246 
1247 static const char *a6xx_fault_block(struct msm_gpu *gpu, u32 id)
1248 {
1249 	if (id == 0)
1250 		return "CP";
1251 	else if (id == 4)
1252 		return "CCU";
1253 	else if (id == 6)
1254 		return "CDP Prefetch";
1255 
1256 	return a6xx_uche_fault_block(gpu, id);
1257 }
1258 
1259 #define ARM_SMMU_FSR_TF                 BIT(1)
1260 #define ARM_SMMU_FSR_PF			BIT(3)
1261 #define ARM_SMMU_FSR_EF			BIT(4)
1262 
1263 static int a6xx_fault_handler(void *arg, unsigned long iova, int flags, void *data)
1264 {
1265 	struct msm_gpu *gpu = arg;
1266 	struct adreno_smmu_fault_info *info = data;
1267 	const char *type = "UNKNOWN";
1268 	const char *block;
1269 	bool do_devcoredump = info && !READ_ONCE(gpu->crashstate);
1270 
1271 	/*
1272 	 * If we aren't going to be resuming later from fault_worker, then do
1273 	 * it now.
1274 	 */
1275 	if (!do_devcoredump) {
1276 		gpu->aspace->mmu->funcs->resume_translation(gpu->aspace->mmu);
1277 	}
1278 
1279 	/*
1280 	 * Print a default message if we couldn't get the data from the
1281 	 * adreno-smmu-priv
1282 	 */
1283 	if (!info) {
1284 		pr_warn_ratelimited("*** gpu fault: iova=%.16lx flags=%d (%u,%u,%u,%u)\n",
1285 			iova, flags,
1286 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(4)),
1287 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(5)),
1288 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(6)),
1289 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(7)));
1290 
1291 		return 0;
1292 	}
1293 
1294 	if (info->fsr & ARM_SMMU_FSR_TF)
1295 		type = "TRANSLATION";
1296 	else if (info->fsr & ARM_SMMU_FSR_PF)
1297 		type = "PERMISSION";
1298 	else if (info->fsr & ARM_SMMU_FSR_EF)
1299 		type = "EXTERNAL";
1300 
1301 	block = a6xx_fault_block(gpu, info->fsynr1 & 0xff);
1302 
1303 	pr_warn_ratelimited("*** gpu fault: ttbr0=%.16llx iova=%.16lx dir=%s type=%s source=%s (%u,%u,%u,%u)\n",
1304 			info->ttbr0, iova,
1305 			flags & IOMMU_FAULT_WRITE ? "WRITE" : "READ",
1306 			type, block,
1307 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(4)),
1308 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(5)),
1309 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(6)),
1310 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(7)));
1311 
1312 	if (do_devcoredump) {
1313 		/* Turn off the hangcheck timer to keep it from bothering us */
1314 		del_timer(&gpu->hangcheck_timer);
1315 
1316 		gpu->fault_info.ttbr0 = info->ttbr0;
1317 		gpu->fault_info.iova  = iova;
1318 		gpu->fault_info.flags = flags;
1319 		gpu->fault_info.type  = type;
1320 		gpu->fault_info.block = block;
1321 
1322 		kthread_queue_work(gpu->worker, &gpu->fault_work);
1323 	}
1324 
1325 	return 0;
1326 }
1327 
1328 static void a6xx_cp_hw_err_irq(struct msm_gpu *gpu)
1329 {
1330 	u32 status = gpu_read(gpu, REG_A6XX_CP_INTERRUPT_STATUS);
1331 
1332 	if (status & A6XX_CP_INT_CP_OPCODE_ERROR) {
1333 		u32 val;
1334 
1335 		gpu_write(gpu, REG_A6XX_CP_SQE_STAT_ADDR, 1);
1336 		val = gpu_read(gpu, REG_A6XX_CP_SQE_STAT_DATA);
1337 		dev_err_ratelimited(&gpu->pdev->dev,
1338 			"CP | opcode error | possible opcode=0x%8.8X\n",
1339 			val);
1340 	}
1341 
1342 	if (status & A6XX_CP_INT_CP_UCODE_ERROR)
1343 		dev_err_ratelimited(&gpu->pdev->dev,
1344 			"CP ucode error interrupt\n");
1345 
1346 	if (status & A6XX_CP_INT_CP_HW_FAULT_ERROR)
1347 		dev_err_ratelimited(&gpu->pdev->dev, "CP | HW fault | status=0x%8.8X\n",
1348 			gpu_read(gpu, REG_A6XX_CP_HW_FAULT));
1349 
1350 	if (status & A6XX_CP_INT_CP_REGISTER_PROTECTION_ERROR) {
1351 		u32 val = gpu_read(gpu, REG_A6XX_CP_PROTECT_STATUS);
1352 
1353 		dev_err_ratelimited(&gpu->pdev->dev,
1354 			"CP | protected mode error | %s | addr=0x%8.8X | status=0x%8.8X\n",
1355 			val & (1 << 20) ? "READ" : "WRITE",
1356 			(val & 0x3ffff), val);
1357 	}
1358 
1359 	if (status & A6XX_CP_INT_CP_AHB_ERROR)
1360 		dev_err_ratelimited(&gpu->pdev->dev, "CP AHB error interrupt\n");
1361 
1362 	if (status & A6XX_CP_INT_CP_VSD_PARITY_ERROR)
1363 		dev_err_ratelimited(&gpu->pdev->dev, "CP VSD decoder parity error\n");
1364 
1365 	if (status & A6XX_CP_INT_CP_ILLEGAL_INSTR_ERROR)
1366 		dev_err_ratelimited(&gpu->pdev->dev, "CP illegal instruction error\n");
1367 
1368 }
1369 
1370 static void a6xx_fault_detect_irq(struct msm_gpu *gpu)
1371 {
1372 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1373 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1374 	struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
1375 
1376 	/*
1377 	 * If stalled on SMMU fault, we could trip the GPU's hang detection,
1378 	 * but the fault handler will trigger the devcore dump, and we want
1379 	 * to otherwise resume normally rather than killing the submit, so
1380 	 * just bail.
1381 	 */
1382 	if (gpu_read(gpu, REG_A6XX_RBBM_STATUS3) & A6XX_RBBM_STATUS3_SMMU_STALLED_ON_FAULT)
1383 		return;
1384 
1385 	/*
1386 	 * Force the GPU to stay on until after we finish
1387 	 * collecting information
1388 	 */
1389 	gmu_write(&a6xx_gpu->gmu, REG_A6XX_GMU_GMU_PWR_COL_KEEPALIVE, 1);
1390 
1391 	DRM_DEV_ERROR(&gpu->pdev->dev,
1392 		"gpu fault ring %d fence %x status %8.8X rb %4.4x/%4.4x ib1 %16.16llX/%4.4x ib2 %16.16llX/%4.4x\n",
1393 		ring ? ring->id : -1, ring ? ring->seqno : 0,
1394 		gpu_read(gpu, REG_A6XX_RBBM_STATUS),
1395 		gpu_read(gpu, REG_A6XX_CP_RB_RPTR),
1396 		gpu_read(gpu, REG_A6XX_CP_RB_WPTR),
1397 		gpu_read64(gpu, REG_A6XX_CP_IB1_BASE, REG_A6XX_CP_IB1_BASE_HI),
1398 		gpu_read(gpu, REG_A6XX_CP_IB1_REM_SIZE),
1399 		gpu_read64(gpu, REG_A6XX_CP_IB2_BASE, REG_A6XX_CP_IB2_BASE_HI),
1400 		gpu_read(gpu, REG_A6XX_CP_IB2_REM_SIZE));
1401 
1402 	/* Turn off the hangcheck timer to keep it from bothering us */
1403 	del_timer(&gpu->hangcheck_timer);
1404 
1405 	kthread_queue_work(gpu->worker, &gpu->recover_work);
1406 }
1407 
1408 static irqreturn_t a6xx_irq(struct msm_gpu *gpu)
1409 {
1410 	struct msm_drm_private *priv = gpu->dev->dev_private;
1411 	u32 status = gpu_read(gpu, REG_A6XX_RBBM_INT_0_STATUS);
1412 
1413 	gpu_write(gpu, REG_A6XX_RBBM_INT_CLEAR_CMD, status);
1414 
1415 	if (priv->disable_err_irq)
1416 		status &= A6XX_RBBM_INT_0_MASK_CP_CACHE_FLUSH_TS;
1417 
1418 	if (status & A6XX_RBBM_INT_0_MASK_RBBM_HANG_DETECT)
1419 		a6xx_fault_detect_irq(gpu);
1420 
1421 	if (status & A6XX_RBBM_INT_0_MASK_CP_AHB_ERROR)
1422 		dev_err_ratelimited(&gpu->pdev->dev, "CP | AHB bus error\n");
1423 
1424 	if (status & A6XX_RBBM_INT_0_MASK_CP_HW_ERROR)
1425 		a6xx_cp_hw_err_irq(gpu);
1426 
1427 	if (status & A6XX_RBBM_INT_0_MASK_RBBM_ATB_ASYNCFIFO_OVERFLOW)
1428 		dev_err_ratelimited(&gpu->pdev->dev, "RBBM | ATB ASYNC overflow\n");
1429 
1430 	if (status & A6XX_RBBM_INT_0_MASK_RBBM_ATB_BUS_OVERFLOW)
1431 		dev_err_ratelimited(&gpu->pdev->dev, "RBBM | ATB bus overflow\n");
1432 
1433 	if (status & A6XX_RBBM_INT_0_MASK_UCHE_OOB_ACCESS)
1434 		dev_err_ratelimited(&gpu->pdev->dev, "UCHE | Out of bounds access\n");
1435 
1436 	if (status & A6XX_RBBM_INT_0_MASK_CP_CACHE_FLUSH_TS)
1437 		msm_gpu_retire(gpu);
1438 
1439 	return IRQ_HANDLED;
1440 }
1441 
1442 static void a6xx_llc_rmw(struct a6xx_gpu *a6xx_gpu, u32 reg, u32 mask, u32 or)
1443 {
1444 	return msm_rmw(a6xx_gpu->llc_mmio + (reg << 2), mask, or);
1445 }
1446 
1447 static void a6xx_llc_write(struct a6xx_gpu *a6xx_gpu, u32 reg, u32 value)
1448 {
1449 	return msm_writel(value, a6xx_gpu->llc_mmio + (reg << 2));
1450 }
1451 
1452 static void a6xx_llc_deactivate(struct a6xx_gpu *a6xx_gpu)
1453 {
1454 	llcc_slice_deactivate(a6xx_gpu->llc_slice);
1455 	llcc_slice_deactivate(a6xx_gpu->htw_llc_slice);
1456 }
1457 
1458 static void a6xx_llc_activate(struct a6xx_gpu *a6xx_gpu)
1459 {
1460 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1461 	struct msm_gpu *gpu = &adreno_gpu->base;
1462 	u32 cntl1_regval = 0;
1463 
1464 	if (IS_ERR(a6xx_gpu->llc_mmio))
1465 		return;
1466 
1467 	if (!llcc_slice_activate(a6xx_gpu->llc_slice)) {
1468 		u32 gpu_scid = llcc_get_slice_id(a6xx_gpu->llc_slice);
1469 
1470 		gpu_scid &= 0x1f;
1471 		cntl1_regval = (gpu_scid << 0) | (gpu_scid << 5) | (gpu_scid << 10) |
1472 			       (gpu_scid << 15) | (gpu_scid << 20);
1473 
1474 		/* On A660, the SCID programming for UCHE traffic is done in
1475 		 * A6XX_GBIF_SCACHE_CNTL0[14:10]
1476 		 */
1477 		if (adreno_is_a660_family(adreno_gpu))
1478 			gpu_rmw(gpu, REG_A6XX_GBIF_SCACHE_CNTL0, (0x1f << 10) |
1479 				(1 << 8), (gpu_scid << 10) | (1 << 8));
1480 	}
1481 
1482 	/*
1483 	 * For targets with a MMU500, activate the slice but don't program the
1484 	 * register.  The XBL will take care of that.
1485 	 */
1486 	if (!llcc_slice_activate(a6xx_gpu->htw_llc_slice)) {
1487 		if (!a6xx_gpu->have_mmu500) {
1488 			u32 gpuhtw_scid = llcc_get_slice_id(a6xx_gpu->htw_llc_slice);
1489 
1490 			gpuhtw_scid &= 0x1f;
1491 			cntl1_regval |= FIELD_PREP(GENMASK(29, 25), gpuhtw_scid);
1492 		}
1493 	}
1494 
1495 	if (!cntl1_regval)
1496 		return;
1497 
1498 	/*
1499 	 * Program the slice IDs for the various GPU blocks and GPU MMU
1500 	 * pagetables
1501 	 */
1502 	if (!a6xx_gpu->have_mmu500) {
1503 		a6xx_llc_write(a6xx_gpu,
1504 			REG_A6XX_CX_MISC_SYSTEM_CACHE_CNTL_1, cntl1_regval);
1505 
1506 		/*
1507 		 * Program cacheability overrides to not allocate cache
1508 		 * lines on a write miss
1509 		 */
1510 		a6xx_llc_rmw(a6xx_gpu,
1511 			REG_A6XX_CX_MISC_SYSTEM_CACHE_CNTL_0, 0xF, 0x03);
1512 		return;
1513 	}
1514 
1515 	gpu_rmw(gpu, REG_A6XX_GBIF_SCACHE_CNTL1, GENMASK(24, 0), cntl1_regval);
1516 }
1517 
1518 static void a6xx_llc_slices_destroy(struct a6xx_gpu *a6xx_gpu)
1519 {
1520 	llcc_slice_putd(a6xx_gpu->llc_slice);
1521 	llcc_slice_putd(a6xx_gpu->htw_llc_slice);
1522 }
1523 
1524 static void a6xx_llc_slices_init(struct platform_device *pdev,
1525 		struct a6xx_gpu *a6xx_gpu)
1526 {
1527 	struct device_node *phandle;
1528 
1529 	/*
1530 	 * There is a different programming path for targets with an mmu500
1531 	 * attached, so detect if that is the case
1532 	 */
1533 	phandle = of_parse_phandle(pdev->dev.of_node, "iommus", 0);
1534 	a6xx_gpu->have_mmu500 = (phandle &&
1535 		of_device_is_compatible(phandle, "arm,mmu-500"));
1536 	of_node_put(phandle);
1537 
1538 	if (a6xx_gpu->have_mmu500)
1539 		a6xx_gpu->llc_mmio = NULL;
1540 	else
1541 		a6xx_gpu->llc_mmio = msm_ioremap(pdev, "cx_mem");
1542 
1543 	a6xx_gpu->llc_slice = llcc_slice_getd(LLCC_GPU);
1544 	a6xx_gpu->htw_llc_slice = llcc_slice_getd(LLCC_GPUHTW);
1545 
1546 	if (IS_ERR_OR_NULL(a6xx_gpu->llc_slice) && IS_ERR_OR_NULL(a6xx_gpu->htw_llc_slice))
1547 		a6xx_gpu->llc_mmio = ERR_PTR(-EINVAL);
1548 }
1549 
1550 static int a6xx_pm_resume(struct msm_gpu *gpu)
1551 {
1552 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1553 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1554 	int ret;
1555 
1556 	gpu->needs_hw_init = true;
1557 
1558 	trace_msm_gpu_resume(0);
1559 
1560 	mutex_lock(&a6xx_gpu->gmu.lock);
1561 	ret = a6xx_gmu_resume(a6xx_gpu);
1562 	mutex_unlock(&a6xx_gpu->gmu.lock);
1563 	if (ret)
1564 		return ret;
1565 
1566 	msm_devfreq_resume(gpu);
1567 
1568 	a6xx_llc_activate(a6xx_gpu);
1569 
1570 	return 0;
1571 }
1572 
1573 static int a6xx_pm_suspend(struct msm_gpu *gpu)
1574 {
1575 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1576 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1577 	int i, ret;
1578 
1579 	trace_msm_gpu_suspend(0);
1580 
1581 	a6xx_llc_deactivate(a6xx_gpu);
1582 
1583 	msm_devfreq_suspend(gpu);
1584 
1585 	mutex_lock(&a6xx_gpu->gmu.lock);
1586 	ret = a6xx_gmu_stop(a6xx_gpu);
1587 	mutex_unlock(&a6xx_gpu->gmu.lock);
1588 	if (ret)
1589 		return ret;
1590 
1591 	if (a6xx_gpu->shadow_bo)
1592 		for (i = 0; i < gpu->nr_rings; i++)
1593 			a6xx_gpu->shadow[i] = 0;
1594 
1595 	gpu->suspend_count++;
1596 
1597 	return 0;
1598 }
1599 
1600 static int a6xx_get_timestamp(struct msm_gpu *gpu, uint64_t *value)
1601 {
1602 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1603 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1604 
1605 	mutex_lock(&a6xx_gpu->gmu.lock);
1606 
1607 	/* Force the GPU power on so we can read this register */
1608 	a6xx_gmu_set_oob(&a6xx_gpu->gmu, GMU_OOB_PERFCOUNTER_SET);
1609 
1610 	*value = gpu_read64(gpu, REG_A6XX_CP_ALWAYS_ON_COUNTER_LO,
1611 			    REG_A6XX_CP_ALWAYS_ON_COUNTER_HI);
1612 
1613 	a6xx_gmu_clear_oob(&a6xx_gpu->gmu, GMU_OOB_PERFCOUNTER_SET);
1614 
1615 	mutex_unlock(&a6xx_gpu->gmu.lock);
1616 
1617 	return 0;
1618 }
1619 
1620 static struct msm_ringbuffer *a6xx_active_ring(struct msm_gpu *gpu)
1621 {
1622 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1623 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1624 
1625 	return a6xx_gpu->cur_ring;
1626 }
1627 
1628 static void a6xx_destroy(struct msm_gpu *gpu)
1629 {
1630 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1631 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1632 
1633 	if (a6xx_gpu->sqe_bo) {
1634 		msm_gem_unpin_iova(a6xx_gpu->sqe_bo, gpu->aspace);
1635 		drm_gem_object_put(a6xx_gpu->sqe_bo);
1636 	}
1637 
1638 	if (a6xx_gpu->shadow_bo) {
1639 		msm_gem_unpin_iova(a6xx_gpu->shadow_bo, gpu->aspace);
1640 		drm_gem_object_put(a6xx_gpu->shadow_bo);
1641 	}
1642 
1643 	a6xx_llc_slices_destroy(a6xx_gpu);
1644 
1645 	a6xx_gmu_remove(a6xx_gpu);
1646 
1647 	adreno_gpu_cleanup(adreno_gpu);
1648 
1649 	kfree(a6xx_gpu);
1650 }
1651 
1652 static unsigned long a6xx_gpu_busy(struct msm_gpu *gpu)
1653 {
1654 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1655 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1656 	u64 busy_cycles, busy_time;
1657 
1658 
1659 	/* Only read the gpu busy if the hardware is already active */
1660 	if (pm_runtime_get_if_in_use(a6xx_gpu->gmu.dev) == 0)
1661 		return 0;
1662 
1663 	busy_cycles = gmu_read64(&a6xx_gpu->gmu,
1664 			REG_A6XX_GMU_CX_GMU_POWER_COUNTER_XOCLK_0_L,
1665 			REG_A6XX_GMU_CX_GMU_POWER_COUNTER_XOCLK_0_H);
1666 
1667 	busy_time = (busy_cycles - gpu->devfreq.busy_cycles) * 10;
1668 	do_div(busy_time, 192);
1669 
1670 	gpu->devfreq.busy_cycles = busy_cycles;
1671 
1672 	pm_runtime_put(a6xx_gpu->gmu.dev);
1673 
1674 	if (WARN_ON(busy_time > ~0LU))
1675 		return ~0LU;
1676 
1677 	return (unsigned long)busy_time;
1678 }
1679 
1680 static void a6xx_gpu_set_freq(struct msm_gpu *gpu, struct dev_pm_opp *opp)
1681 {
1682 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1683 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1684 
1685 	mutex_lock(&a6xx_gpu->gmu.lock);
1686 	a6xx_gmu_set_freq(gpu, opp);
1687 	mutex_unlock(&a6xx_gpu->gmu.lock);
1688 }
1689 
1690 static struct msm_gem_address_space *
1691 a6xx_create_address_space(struct msm_gpu *gpu, struct platform_device *pdev)
1692 {
1693 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1694 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1695 	struct iommu_domain *iommu;
1696 	struct msm_mmu *mmu;
1697 	struct msm_gem_address_space *aspace;
1698 	u64 start, size;
1699 
1700 	iommu = iommu_domain_alloc(&platform_bus_type);
1701 	if (!iommu)
1702 		return NULL;
1703 
1704 	/*
1705 	 * This allows GPU to set the bus attributes required to use system
1706 	 * cache on behalf of the iommu page table walker.
1707 	 */
1708 	if (!IS_ERR_OR_NULL(a6xx_gpu->htw_llc_slice))
1709 		adreno_set_llc_attributes(iommu);
1710 
1711 	mmu = msm_iommu_new(&pdev->dev, iommu);
1712 	if (IS_ERR(mmu)) {
1713 		iommu_domain_free(iommu);
1714 		return ERR_CAST(mmu);
1715 	}
1716 
1717 	/*
1718 	 * Use the aperture start or SZ_16M, whichever is greater. This will
1719 	 * ensure that we align with the allocated pagetable range while still
1720 	 * allowing room in the lower 32 bits for GMEM and whatnot
1721 	 */
1722 	start = max_t(u64, SZ_16M, iommu->geometry.aperture_start);
1723 	size = iommu->geometry.aperture_end - start + 1;
1724 
1725 	aspace = msm_gem_address_space_create(mmu, "gpu",
1726 		start & GENMASK_ULL(48, 0), size);
1727 
1728 	if (IS_ERR(aspace) && !IS_ERR(mmu))
1729 		mmu->funcs->destroy(mmu);
1730 
1731 	return aspace;
1732 }
1733 
1734 static struct msm_gem_address_space *
1735 a6xx_create_private_address_space(struct msm_gpu *gpu)
1736 {
1737 	struct msm_mmu *mmu;
1738 
1739 	mmu = msm_iommu_pagetable_create(gpu->aspace->mmu);
1740 
1741 	if (IS_ERR(mmu))
1742 		return ERR_CAST(mmu);
1743 
1744 	return msm_gem_address_space_create(mmu,
1745 		"gpu", 0x100000000ULL, SZ_4G);
1746 }
1747 
1748 static uint32_t a6xx_get_rptr(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
1749 {
1750 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1751 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1752 
1753 	if (adreno_gpu->base.hw_apriv || a6xx_gpu->has_whereami)
1754 		return a6xx_gpu->shadow[ring->id];
1755 
1756 	return ring->memptrs->rptr = gpu_read(gpu, REG_A6XX_CP_RB_RPTR);
1757 }
1758 
1759 static u32 a618_get_speed_bin(u32 fuse)
1760 {
1761 	if (fuse == 0)
1762 		return 0;
1763 	else if (fuse == 169)
1764 		return 1;
1765 	else if (fuse == 174)
1766 		return 2;
1767 
1768 	return UINT_MAX;
1769 }
1770 
1771 static u32 adreno_7c3_get_speed_bin(u32 fuse)
1772 {
1773 	if (fuse == 0)
1774 		return 0;
1775 	else if (fuse == 117)
1776 		return 0;
1777 	else if (fuse == 190)
1778 		return 1;
1779 
1780 	return UINT_MAX;
1781 }
1782 
1783 static u32 fuse_to_supp_hw(struct device *dev, struct adreno_rev rev, u32 fuse)
1784 {
1785 	u32 val = UINT_MAX;
1786 
1787 	if (adreno_cmp_rev(ADRENO_REV(6, 1, 8, ANY_ID), rev))
1788 		val = a618_get_speed_bin(fuse);
1789 
1790 	if (adreno_cmp_rev(ADRENO_REV(6, 3, 5, ANY_ID), rev))
1791 		val = adreno_7c3_get_speed_bin(fuse);
1792 
1793 	if (val == UINT_MAX) {
1794 		DRM_DEV_ERROR(dev,
1795 			"missing support for speed-bin: %u. Some OPPs may not be supported by hardware",
1796 			fuse);
1797 		return UINT_MAX;
1798 	}
1799 
1800 	return (1 << val);
1801 }
1802 
1803 static int a6xx_set_supported_hw(struct device *dev, struct adreno_rev rev)
1804 {
1805 	u32 supp_hw = UINT_MAX;
1806 	u32 speedbin;
1807 	int ret;
1808 
1809 	ret = adreno_read_speedbin(dev, &speedbin);
1810 	/*
1811 	 * -ENOENT means that the platform doesn't support speedbin which is
1812 	 * fine
1813 	 */
1814 	if (ret == -ENOENT) {
1815 		return 0;
1816 	} else if (ret) {
1817 		DRM_DEV_ERROR(dev,
1818 			      "failed to read speed-bin (%d). Some OPPs may not be supported by hardware",
1819 			      ret);
1820 		goto done;
1821 	}
1822 
1823 	supp_hw = fuse_to_supp_hw(dev, rev, speedbin);
1824 
1825 done:
1826 	ret = devm_pm_opp_set_supported_hw(dev, &supp_hw, 1);
1827 	if (ret)
1828 		return ret;
1829 
1830 	return 0;
1831 }
1832 
1833 static const struct adreno_gpu_funcs funcs = {
1834 	.base = {
1835 		.get_param = adreno_get_param,
1836 		.set_param = adreno_set_param,
1837 		.hw_init = a6xx_hw_init,
1838 		.pm_suspend = a6xx_pm_suspend,
1839 		.pm_resume = a6xx_pm_resume,
1840 		.recover = a6xx_recover,
1841 		.submit = a6xx_submit,
1842 		.active_ring = a6xx_active_ring,
1843 		.irq = a6xx_irq,
1844 		.destroy = a6xx_destroy,
1845 #if defined(CONFIG_DRM_MSM_GPU_STATE)
1846 		.show = a6xx_show,
1847 #endif
1848 		.gpu_busy = a6xx_gpu_busy,
1849 		.gpu_get_freq = a6xx_gmu_get_freq,
1850 		.gpu_set_freq = a6xx_gpu_set_freq,
1851 #if defined(CONFIG_DRM_MSM_GPU_STATE)
1852 		.gpu_state_get = a6xx_gpu_state_get,
1853 		.gpu_state_put = a6xx_gpu_state_put,
1854 #endif
1855 		.create_address_space = a6xx_create_address_space,
1856 		.create_private_address_space = a6xx_create_private_address_space,
1857 		.get_rptr = a6xx_get_rptr,
1858 	},
1859 	.get_timestamp = a6xx_get_timestamp,
1860 };
1861 
1862 struct msm_gpu *a6xx_gpu_init(struct drm_device *dev)
1863 {
1864 	struct msm_drm_private *priv = dev->dev_private;
1865 	struct platform_device *pdev = priv->gpu_pdev;
1866 	struct adreno_platform_config *config = pdev->dev.platform_data;
1867 	const struct adreno_info *info;
1868 	struct device_node *node;
1869 	struct a6xx_gpu *a6xx_gpu;
1870 	struct adreno_gpu *adreno_gpu;
1871 	struct msm_gpu *gpu;
1872 	int ret;
1873 
1874 	a6xx_gpu = kzalloc(sizeof(*a6xx_gpu), GFP_KERNEL);
1875 	if (!a6xx_gpu)
1876 		return ERR_PTR(-ENOMEM);
1877 
1878 	adreno_gpu = &a6xx_gpu->base;
1879 	gpu = &adreno_gpu->base;
1880 
1881 	adreno_gpu->registers = NULL;
1882 
1883 	/*
1884 	 * We need to know the platform type before calling into adreno_gpu_init
1885 	 * so that the hw_apriv flag can be correctly set. Snoop into the info
1886 	 * and grab the revision number
1887 	 */
1888 	info = adreno_info(config->rev);
1889 
1890 	if (info && (info->revn == 650 || info->revn == 660 ||
1891 			adreno_cmp_rev(ADRENO_REV(6, 3, 5, ANY_ID), info->rev)))
1892 		adreno_gpu->base.hw_apriv = true;
1893 
1894 	/*
1895 	 * For now only clamp to idle freq for devices where this is known not
1896 	 * to cause power supply issues:
1897 	 */
1898 	if (info && (info->revn == 618))
1899 		gpu->clamp_to_idle = true;
1900 
1901 	a6xx_llc_slices_init(pdev, a6xx_gpu);
1902 
1903 	ret = a6xx_set_supported_hw(&pdev->dev, config->rev);
1904 	if (ret) {
1905 		a6xx_destroy(&(a6xx_gpu->base.base));
1906 		return ERR_PTR(ret);
1907 	}
1908 
1909 	ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs, 1);
1910 	if (ret) {
1911 		a6xx_destroy(&(a6xx_gpu->base.base));
1912 		return ERR_PTR(ret);
1913 	}
1914 
1915 	/* Check if there is a GMU phandle and set it up */
1916 	node = of_parse_phandle(pdev->dev.of_node, "qcom,gmu", 0);
1917 
1918 	/* FIXME: How do we gracefully handle this? */
1919 	BUG_ON(!node);
1920 
1921 	ret = a6xx_gmu_init(a6xx_gpu, node);
1922 	if (ret) {
1923 		a6xx_destroy(&(a6xx_gpu->base.base));
1924 		return ERR_PTR(ret);
1925 	}
1926 
1927 	if (gpu->aspace)
1928 		msm_mmu_set_fault_handler(gpu->aspace->mmu, gpu,
1929 				a6xx_fault_handler);
1930 
1931 	return gpu;
1932 }
1933