xref: /linux/drivers/gpu/drm/msm/adreno/a5xx_gpu.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2016-2017 The Linux Foundation. All rights reserved.
3  */
4 
5 #include <linux/kernel.h>
6 #include <linux/types.h>
7 #include <linux/cpumask.h>
8 #include <linux/firmware/qcom/qcom_scm.h>
9 #include <linux/pm_opp.h>
10 #include <linux/nvmem-consumer.h>
11 #include <linux/slab.h>
12 #include "msm_gem.h"
13 #include "msm_mmu.h"
14 #include "a5xx_gpu.h"
15 
16 extern bool hang_debug;
17 static void a5xx_dump(struct msm_gpu *gpu);
18 
19 #define GPU_PAS_ID 13
20 
21 static void update_shadow_rptr(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
22 {
23 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
24 	struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu);
25 
26 	if (a5xx_gpu->has_whereami) {
27 		OUT_PKT7(ring, CP_WHERE_AM_I, 2);
28 		OUT_RING(ring, lower_32_bits(shadowptr(a5xx_gpu, ring)));
29 		OUT_RING(ring, upper_32_bits(shadowptr(a5xx_gpu, ring)));
30 	}
31 }
32 
33 void a5xx_flush(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
34 		bool sync)
35 {
36 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
37 	struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu);
38 	uint32_t wptr;
39 	unsigned long flags;
40 
41 	/*
42 	 * Most flush operations need to issue a WHERE_AM_I opcode to sync up
43 	 * the rptr shadow
44 	 */
45 	if (sync)
46 		update_shadow_rptr(gpu, ring);
47 
48 	spin_lock_irqsave(&ring->preempt_lock, flags);
49 
50 	/* Copy the shadow to the actual register */
51 	ring->cur = ring->next;
52 
53 	/* Make sure to wrap wptr if we need to */
54 	wptr = get_wptr(ring);
55 
56 	spin_unlock_irqrestore(&ring->preempt_lock, flags);
57 
58 	/* Make sure everything is posted before making a decision */
59 	mb();
60 
61 	/* Update HW if this is the current ring and we are not in preempt */
62 	if (a5xx_gpu->cur_ring == ring && !a5xx_in_preempt(a5xx_gpu))
63 		gpu_write(gpu, REG_A5XX_CP_RB_WPTR, wptr);
64 }
65 
66 static void a5xx_submit_in_rb(struct msm_gpu *gpu, struct msm_gem_submit *submit)
67 {
68 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
69 	struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu);
70 	struct msm_ringbuffer *ring = submit->ring;
71 	struct drm_gem_object *obj;
72 	uint32_t *ptr, dwords;
73 	unsigned int i;
74 
75 	for (i = 0; i < submit->nr_cmds; i++) {
76 		switch (submit->cmd[i].type) {
77 		case MSM_SUBMIT_CMD_IB_TARGET_BUF:
78 			break;
79 		case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
80 			if (gpu->cur_ctx_seqno == submit->queue->ctx->seqno)
81 				break;
82 			fallthrough;
83 		case MSM_SUBMIT_CMD_BUF:
84 			/* copy commands into RB: */
85 			obj = submit->bos[submit->cmd[i].idx].obj;
86 			dwords = submit->cmd[i].size;
87 
88 			ptr = msm_gem_get_vaddr(obj);
89 
90 			/* _get_vaddr() shouldn't fail at this point,
91 			 * since we've already mapped it once in
92 			 * submit_reloc()
93 			 */
94 			if (WARN_ON(IS_ERR_OR_NULL(ptr)))
95 				return;
96 
97 			for (i = 0; i < dwords; i++) {
98 				/* normally the OUT_PKTn() would wait
99 				 * for space for the packet.  But since
100 				 * we just OUT_RING() the whole thing,
101 				 * need to call adreno_wait_ring()
102 				 * ourself:
103 				 */
104 				adreno_wait_ring(ring, 1);
105 				OUT_RING(ring, ptr[i]);
106 			}
107 
108 			msm_gem_put_vaddr(obj);
109 
110 			break;
111 		}
112 	}
113 
114 	a5xx_gpu->last_seqno[ring->id] = submit->seqno;
115 	a5xx_flush(gpu, ring, true);
116 	a5xx_preempt_trigger(gpu);
117 
118 	/* we might not necessarily have a cmd from userspace to
119 	 * trigger an event to know that submit has completed, so
120 	 * do this manually:
121 	 */
122 	a5xx_idle(gpu, ring);
123 	ring->memptrs->fence = submit->seqno;
124 	msm_gpu_retire(gpu);
125 }
126 
127 static void a5xx_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
128 {
129 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
130 	struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu);
131 	struct msm_ringbuffer *ring = submit->ring;
132 	unsigned int i, ibs = 0;
133 
134 	if (IS_ENABLED(CONFIG_DRM_MSM_GPU_SUDO) && submit->in_rb) {
135 		gpu->cur_ctx_seqno = 0;
136 		a5xx_submit_in_rb(gpu, submit);
137 		return;
138 	}
139 
140 	OUT_PKT7(ring, CP_PREEMPT_ENABLE_GLOBAL, 1);
141 	OUT_RING(ring, 0x02);
142 
143 	/* Turn off protected mode to write to special registers */
144 	OUT_PKT7(ring, CP_SET_PROTECTED_MODE, 1);
145 	OUT_RING(ring, 0);
146 
147 	/* Set the save preemption record for the ring/command */
148 	OUT_PKT4(ring, REG_A5XX_CP_CONTEXT_SWITCH_SAVE_ADDR_LO, 2);
149 	OUT_RING(ring, lower_32_bits(a5xx_gpu->preempt_iova[submit->ring->id]));
150 	OUT_RING(ring, upper_32_bits(a5xx_gpu->preempt_iova[submit->ring->id]));
151 
152 	/* Turn back on protected mode */
153 	OUT_PKT7(ring, CP_SET_PROTECTED_MODE, 1);
154 	OUT_RING(ring, 1);
155 
156 	/*
157 	 * Disable local preemption by default because it requires
158 	 * user-space to be aware of it and provide additional handling
159 	 * to restore rendering state or do various flushes on switch.
160 	 */
161 	OUT_PKT7(ring, CP_PREEMPT_ENABLE_LOCAL, 1);
162 	OUT_RING(ring, 0x0);
163 
164 	/* Allow CP_CONTEXT_SWITCH_YIELD packets in the IB2 */
165 	OUT_PKT7(ring, CP_YIELD_ENABLE, 1);
166 	OUT_RING(ring, 0x02);
167 
168 	/* Submit the commands */
169 	for (i = 0; i < submit->nr_cmds; i++) {
170 		switch (submit->cmd[i].type) {
171 		case MSM_SUBMIT_CMD_IB_TARGET_BUF:
172 			break;
173 		case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
174 			if (gpu->cur_ctx_seqno == submit->queue->ctx->seqno)
175 				break;
176 			fallthrough;
177 		case MSM_SUBMIT_CMD_BUF:
178 			OUT_PKT7(ring, CP_INDIRECT_BUFFER_PFE, 3);
179 			OUT_RING(ring, lower_32_bits(submit->cmd[i].iova));
180 			OUT_RING(ring, upper_32_bits(submit->cmd[i].iova));
181 			OUT_RING(ring, submit->cmd[i].size);
182 			ibs++;
183 			break;
184 		}
185 
186 		/*
187 		 * Periodically update shadow-wptr if needed, so that we
188 		 * can see partial progress of submits with large # of
189 		 * cmds.. otherwise we could needlessly stall waiting for
190 		 * ringbuffer state, simply due to looking at a shadow
191 		 * rptr value that has not been updated
192 		 */
193 		if ((ibs % 32) == 0)
194 			update_shadow_rptr(gpu, ring);
195 	}
196 
197 	/*
198 	 * Write the render mode to NULL (0) to indicate to the CP that the IBs
199 	 * are done rendering - otherwise a lucky preemption would start
200 	 * replaying from the last checkpoint
201 	 */
202 	OUT_PKT7(ring, CP_SET_RENDER_MODE, 5);
203 	OUT_RING(ring, 0);
204 	OUT_RING(ring, 0);
205 	OUT_RING(ring, 0);
206 	OUT_RING(ring, 0);
207 	OUT_RING(ring, 0);
208 
209 	/* Turn off IB level preemptions */
210 	OUT_PKT7(ring, CP_YIELD_ENABLE, 1);
211 	OUT_RING(ring, 0x01);
212 
213 	/* Write the fence to the scratch register */
214 	OUT_PKT4(ring, REG_A5XX_CP_SCRATCH_REG(2), 1);
215 	OUT_RING(ring, submit->seqno);
216 	a5xx_gpu->last_seqno[ring->id] = submit->seqno;
217 
218 	/*
219 	 * Execute a CACHE_FLUSH_TS event. This will ensure that the
220 	 * timestamp is written to the memory and then triggers the interrupt
221 	 */
222 	OUT_PKT7(ring, CP_EVENT_WRITE, 4);
223 	OUT_RING(ring, CP_EVENT_WRITE_0_EVENT(CACHE_FLUSH_TS) |
224 		CP_EVENT_WRITE_0_IRQ);
225 	OUT_RING(ring, lower_32_bits(rbmemptr(ring, fence)));
226 	OUT_RING(ring, upper_32_bits(rbmemptr(ring, fence)));
227 	OUT_RING(ring, submit->seqno);
228 
229 	/* Yield the floor on command completion */
230 	OUT_PKT7(ring, CP_CONTEXT_SWITCH_YIELD, 4);
231 	/*
232 	 * If dword[2:1] are non zero, they specify an address for the CP to
233 	 * write the value of dword[3] to on preemption complete. Write 0 to
234 	 * skip the write
235 	 */
236 	OUT_RING(ring, 0x00);
237 	OUT_RING(ring, 0x00);
238 	/* Data value - not used if the address above is 0 */
239 	OUT_RING(ring, 0x01);
240 	/* Set bit 0 to trigger an interrupt on preempt complete */
241 	OUT_RING(ring, 0x01);
242 
243 	/* A WHERE_AM_I packet is not needed after a YIELD */
244 	a5xx_flush(gpu, ring, false);
245 
246 	/* Check to see if we need to start preemption */
247 	a5xx_preempt_trigger(gpu);
248 }
249 
250 static const struct adreno_five_hwcg_regs {
251 	u32 offset;
252 	u32 value;
253 } a5xx_hwcg[] = {
254 	{REG_A5XX_RBBM_CLOCK_CNTL_SP0, 0x02222222},
255 	{REG_A5XX_RBBM_CLOCK_CNTL_SP1, 0x02222222},
256 	{REG_A5XX_RBBM_CLOCK_CNTL_SP2, 0x02222222},
257 	{REG_A5XX_RBBM_CLOCK_CNTL_SP3, 0x02222222},
258 	{REG_A5XX_RBBM_CLOCK_CNTL2_SP0, 0x02222220},
259 	{REG_A5XX_RBBM_CLOCK_CNTL2_SP1, 0x02222220},
260 	{REG_A5XX_RBBM_CLOCK_CNTL2_SP2, 0x02222220},
261 	{REG_A5XX_RBBM_CLOCK_CNTL2_SP3, 0x02222220},
262 	{REG_A5XX_RBBM_CLOCK_HYST_SP0, 0x0000F3CF},
263 	{REG_A5XX_RBBM_CLOCK_HYST_SP1, 0x0000F3CF},
264 	{REG_A5XX_RBBM_CLOCK_HYST_SP2, 0x0000F3CF},
265 	{REG_A5XX_RBBM_CLOCK_HYST_SP3, 0x0000F3CF},
266 	{REG_A5XX_RBBM_CLOCK_DELAY_SP0, 0x00000080},
267 	{REG_A5XX_RBBM_CLOCK_DELAY_SP1, 0x00000080},
268 	{REG_A5XX_RBBM_CLOCK_DELAY_SP2, 0x00000080},
269 	{REG_A5XX_RBBM_CLOCK_DELAY_SP3, 0x00000080},
270 	{REG_A5XX_RBBM_CLOCK_CNTL_TP0, 0x22222222},
271 	{REG_A5XX_RBBM_CLOCK_CNTL_TP1, 0x22222222},
272 	{REG_A5XX_RBBM_CLOCK_CNTL_TP2, 0x22222222},
273 	{REG_A5XX_RBBM_CLOCK_CNTL_TP3, 0x22222222},
274 	{REG_A5XX_RBBM_CLOCK_CNTL2_TP0, 0x22222222},
275 	{REG_A5XX_RBBM_CLOCK_CNTL2_TP1, 0x22222222},
276 	{REG_A5XX_RBBM_CLOCK_CNTL2_TP2, 0x22222222},
277 	{REG_A5XX_RBBM_CLOCK_CNTL2_TP3, 0x22222222},
278 	{REG_A5XX_RBBM_CLOCK_CNTL3_TP0, 0x00002222},
279 	{REG_A5XX_RBBM_CLOCK_CNTL3_TP1, 0x00002222},
280 	{REG_A5XX_RBBM_CLOCK_CNTL3_TP2, 0x00002222},
281 	{REG_A5XX_RBBM_CLOCK_CNTL3_TP3, 0x00002222},
282 	{REG_A5XX_RBBM_CLOCK_HYST_TP0, 0x77777777},
283 	{REG_A5XX_RBBM_CLOCK_HYST_TP1, 0x77777777},
284 	{REG_A5XX_RBBM_CLOCK_HYST_TP2, 0x77777777},
285 	{REG_A5XX_RBBM_CLOCK_HYST_TP3, 0x77777777},
286 	{REG_A5XX_RBBM_CLOCK_HYST2_TP0, 0x77777777},
287 	{REG_A5XX_RBBM_CLOCK_HYST2_TP1, 0x77777777},
288 	{REG_A5XX_RBBM_CLOCK_HYST2_TP2, 0x77777777},
289 	{REG_A5XX_RBBM_CLOCK_HYST2_TP3, 0x77777777},
290 	{REG_A5XX_RBBM_CLOCK_HYST3_TP0, 0x00007777},
291 	{REG_A5XX_RBBM_CLOCK_HYST3_TP1, 0x00007777},
292 	{REG_A5XX_RBBM_CLOCK_HYST3_TP2, 0x00007777},
293 	{REG_A5XX_RBBM_CLOCK_HYST3_TP3, 0x00007777},
294 	{REG_A5XX_RBBM_CLOCK_DELAY_TP0, 0x11111111},
295 	{REG_A5XX_RBBM_CLOCK_DELAY_TP1, 0x11111111},
296 	{REG_A5XX_RBBM_CLOCK_DELAY_TP2, 0x11111111},
297 	{REG_A5XX_RBBM_CLOCK_DELAY_TP3, 0x11111111},
298 	{REG_A5XX_RBBM_CLOCK_DELAY2_TP0, 0x11111111},
299 	{REG_A5XX_RBBM_CLOCK_DELAY2_TP1, 0x11111111},
300 	{REG_A5XX_RBBM_CLOCK_DELAY2_TP2, 0x11111111},
301 	{REG_A5XX_RBBM_CLOCK_DELAY2_TP3, 0x11111111},
302 	{REG_A5XX_RBBM_CLOCK_DELAY3_TP0, 0x00001111},
303 	{REG_A5XX_RBBM_CLOCK_DELAY3_TP1, 0x00001111},
304 	{REG_A5XX_RBBM_CLOCK_DELAY3_TP2, 0x00001111},
305 	{REG_A5XX_RBBM_CLOCK_DELAY3_TP3, 0x00001111},
306 	{REG_A5XX_RBBM_CLOCK_CNTL_UCHE, 0x22222222},
307 	{REG_A5XX_RBBM_CLOCK_CNTL2_UCHE, 0x22222222},
308 	{REG_A5XX_RBBM_CLOCK_CNTL3_UCHE, 0x22222222},
309 	{REG_A5XX_RBBM_CLOCK_CNTL4_UCHE, 0x00222222},
310 	{REG_A5XX_RBBM_CLOCK_HYST_UCHE, 0x00444444},
311 	{REG_A5XX_RBBM_CLOCK_DELAY_UCHE, 0x00000002},
312 	{REG_A5XX_RBBM_CLOCK_CNTL_RB0, 0x22222222},
313 	{REG_A5XX_RBBM_CLOCK_CNTL_RB1, 0x22222222},
314 	{REG_A5XX_RBBM_CLOCK_CNTL_RB2, 0x22222222},
315 	{REG_A5XX_RBBM_CLOCK_CNTL_RB3, 0x22222222},
316 	{REG_A5XX_RBBM_CLOCK_CNTL2_RB0, 0x00222222},
317 	{REG_A5XX_RBBM_CLOCK_CNTL2_RB1, 0x00222222},
318 	{REG_A5XX_RBBM_CLOCK_CNTL2_RB2, 0x00222222},
319 	{REG_A5XX_RBBM_CLOCK_CNTL2_RB3, 0x00222222},
320 	{REG_A5XX_RBBM_CLOCK_CNTL_CCU0, 0x00022220},
321 	{REG_A5XX_RBBM_CLOCK_CNTL_CCU1, 0x00022220},
322 	{REG_A5XX_RBBM_CLOCK_CNTL_CCU2, 0x00022220},
323 	{REG_A5XX_RBBM_CLOCK_CNTL_CCU3, 0x00022220},
324 	{REG_A5XX_RBBM_CLOCK_CNTL_RAC, 0x05522222},
325 	{REG_A5XX_RBBM_CLOCK_CNTL2_RAC, 0x00505555},
326 	{REG_A5XX_RBBM_CLOCK_HYST_RB_CCU0, 0x04040404},
327 	{REG_A5XX_RBBM_CLOCK_HYST_RB_CCU1, 0x04040404},
328 	{REG_A5XX_RBBM_CLOCK_HYST_RB_CCU2, 0x04040404},
329 	{REG_A5XX_RBBM_CLOCK_HYST_RB_CCU3, 0x04040404},
330 	{REG_A5XX_RBBM_CLOCK_HYST_RAC, 0x07444044},
331 	{REG_A5XX_RBBM_CLOCK_DELAY_RB_CCU_L1_0, 0x00000002},
332 	{REG_A5XX_RBBM_CLOCK_DELAY_RB_CCU_L1_1, 0x00000002},
333 	{REG_A5XX_RBBM_CLOCK_DELAY_RB_CCU_L1_2, 0x00000002},
334 	{REG_A5XX_RBBM_CLOCK_DELAY_RB_CCU_L1_3, 0x00000002},
335 	{REG_A5XX_RBBM_CLOCK_DELAY_RAC, 0x00010011},
336 	{REG_A5XX_RBBM_CLOCK_CNTL_TSE_RAS_RBBM, 0x04222222},
337 	{REG_A5XX_RBBM_CLOCK_MODE_GPC, 0x02222222},
338 	{REG_A5XX_RBBM_CLOCK_MODE_VFD, 0x00002222},
339 	{REG_A5XX_RBBM_CLOCK_HYST_TSE_RAS_RBBM, 0x00000000},
340 	{REG_A5XX_RBBM_CLOCK_HYST_GPC, 0x04104004},
341 	{REG_A5XX_RBBM_CLOCK_HYST_VFD, 0x00000000},
342 	{REG_A5XX_RBBM_CLOCK_DELAY_HLSQ, 0x00000000},
343 	{REG_A5XX_RBBM_CLOCK_DELAY_TSE_RAS_RBBM, 0x00004000},
344 	{REG_A5XX_RBBM_CLOCK_DELAY_GPC, 0x00000200},
345 	{REG_A5XX_RBBM_CLOCK_DELAY_VFD, 0x00002222}
346 }, a50x_hwcg[] = {
347 	{REG_A5XX_RBBM_CLOCK_CNTL_SP0, 0x02222222},
348 	{REG_A5XX_RBBM_CLOCK_CNTL2_SP0, 0x02222220},
349 	{REG_A5XX_RBBM_CLOCK_HYST_SP0, 0x0000F3CF},
350 	{REG_A5XX_RBBM_CLOCK_DELAY_SP0, 0x00000080},
351 	{REG_A5XX_RBBM_CLOCK_CNTL_TP0, 0x22222222},
352 	{REG_A5XX_RBBM_CLOCK_CNTL2_TP0, 0x22222222},
353 	{REG_A5XX_RBBM_CLOCK_CNTL3_TP0, 0x00002222},
354 	{REG_A5XX_RBBM_CLOCK_HYST_TP0, 0x77777777},
355 	{REG_A5XX_RBBM_CLOCK_HYST2_TP0, 0x77777777},
356 	{REG_A5XX_RBBM_CLOCK_HYST3_TP0, 0x00007777},
357 	{REG_A5XX_RBBM_CLOCK_DELAY_TP0, 0x11111111},
358 	{REG_A5XX_RBBM_CLOCK_DELAY2_TP0, 0x11111111},
359 	{REG_A5XX_RBBM_CLOCK_DELAY3_TP0, 0x00001111},
360 	{REG_A5XX_RBBM_CLOCK_CNTL2_UCHE, 0x22222222},
361 	{REG_A5XX_RBBM_CLOCK_CNTL3_UCHE, 0x22222222},
362 	{REG_A5XX_RBBM_CLOCK_CNTL4_UCHE, 0x00222222},
363 	{REG_A5XX_RBBM_CLOCK_CNTL_UCHE, 0x22222222},
364 	{REG_A5XX_RBBM_CLOCK_HYST_UCHE, 0x00FFFFF4},
365 	{REG_A5XX_RBBM_CLOCK_DELAY_UCHE, 0x00000002},
366 	{REG_A5XX_RBBM_CLOCK_CNTL_RB0, 0x22222222},
367 	{REG_A5XX_RBBM_CLOCK_CNTL2_RB0, 0x00222222},
368 	{REG_A5XX_RBBM_CLOCK_CNTL_CCU0, 0x00022220},
369 	{REG_A5XX_RBBM_CLOCK_CNTL_RAC, 0x05522222},
370 	{REG_A5XX_RBBM_CLOCK_CNTL2_RAC, 0x00505555},
371 	{REG_A5XX_RBBM_CLOCK_HYST_RB_CCU0, 0x04040404},
372 	{REG_A5XX_RBBM_CLOCK_HYST_RAC, 0x07444044},
373 	{REG_A5XX_RBBM_CLOCK_DELAY_RB_CCU_L1_0, 0x00000002},
374 	{REG_A5XX_RBBM_CLOCK_DELAY_RAC, 0x00010011},
375 	{REG_A5XX_RBBM_CLOCK_CNTL_TSE_RAS_RBBM, 0x04222222},
376 	{REG_A5XX_RBBM_CLOCK_MODE_GPC, 0x02222222},
377 	{REG_A5XX_RBBM_CLOCK_MODE_VFD, 0x00002222},
378 	{REG_A5XX_RBBM_CLOCK_HYST_TSE_RAS_RBBM, 0x00000000},
379 	{REG_A5XX_RBBM_CLOCK_HYST_GPC, 0x04104004},
380 	{REG_A5XX_RBBM_CLOCK_HYST_VFD, 0x00000000},
381 	{REG_A5XX_RBBM_CLOCK_DELAY_HLSQ, 0x00000000},
382 	{REG_A5XX_RBBM_CLOCK_DELAY_TSE_RAS_RBBM, 0x00004000},
383 	{REG_A5XX_RBBM_CLOCK_DELAY_GPC, 0x00000200},
384 	{REG_A5XX_RBBM_CLOCK_DELAY_VFD, 0x00002222},
385 }, a512_hwcg[] = {
386 	{REG_A5XX_RBBM_CLOCK_CNTL_SP0, 0x02222222},
387 	{REG_A5XX_RBBM_CLOCK_CNTL_SP1, 0x02222222},
388 	{REG_A5XX_RBBM_CLOCK_CNTL2_SP0, 0x02222220},
389 	{REG_A5XX_RBBM_CLOCK_CNTL2_SP1, 0x02222220},
390 	{REG_A5XX_RBBM_CLOCK_HYST_SP0, 0x0000F3CF},
391 	{REG_A5XX_RBBM_CLOCK_HYST_SP1, 0x0000F3CF},
392 	{REG_A5XX_RBBM_CLOCK_DELAY_SP0, 0x00000080},
393 	{REG_A5XX_RBBM_CLOCK_DELAY_SP1, 0x00000080},
394 	{REG_A5XX_RBBM_CLOCK_CNTL_TP0, 0x22222222},
395 	{REG_A5XX_RBBM_CLOCK_CNTL_TP1, 0x22222222},
396 	{REG_A5XX_RBBM_CLOCK_CNTL2_TP0, 0x22222222},
397 	{REG_A5XX_RBBM_CLOCK_CNTL2_TP1, 0x22222222},
398 	{REG_A5XX_RBBM_CLOCK_CNTL3_TP0, 0x00002222},
399 	{REG_A5XX_RBBM_CLOCK_CNTL3_TP1, 0x00002222},
400 	{REG_A5XX_RBBM_CLOCK_HYST_TP0, 0x77777777},
401 	{REG_A5XX_RBBM_CLOCK_HYST_TP1, 0x77777777},
402 	{REG_A5XX_RBBM_CLOCK_HYST2_TP0, 0x77777777},
403 	{REG_A5XX_RBBM_CLOCK_HYST2_TP1, 0x77777777},
404 	{REG_A5XX_RBBM_CLOCK_HYST3_TP0, 0x00007777},
405 	{REG_A5XX_RBBM_CLOCK_HYST3_TP1, 0x00007777},
406 	{REG_A5XX_RBBM_CLOCK_DELAY_TP0, 0x11111111},
407 	{REG_A5XX_RBBM_CLOCK_DELAY_TP1, 0x11111111},
408 	{REG_A5XX_RBBM_CLOCK_DELAY2_TP0, 0x11111111},
409 	{REG_A5XX_RBBM_CLOCK_DELAY2_TP1, 0x11111111},
410 	{REG_A5XX_RBBM_CLOCK_DELAY3_TP0, 0x00001111},
411 	{REG_A5XX_RBBM_CLOCK_DELAY3_TP1, 0x00001111},
412 	{REG_A5XX_RBBM_CLOCK_CNTL_UCHE, 0x22222222},
413 	{REG_A5XX_RBBM_CLOCK_CNTL2_UCHE, 0x22222222},
414 	{REG_A5XX_RBBM_CLOCK_CNTL3_UCHE, 0x22222222},
415 	{REG_A5XX_RBBM_CLOCK_CNTL4_UCHE, 0x00222222},
416 	{REG_A5XX_RBBM_CLOCK_HYST_UCHE, 0x00444444},
417 	{REG_A5XX_RBBM_CLOCK_DELAY_UCHE, 0x00000002},
418 	{REG_A5XX_RBBM_CLOCK_CNTL_RB0, 0x22222222},
419 	{REG_A5XX_RBBM_CLOCK_CNTL_RB1, 0x22222222},
420 	{REG_A5XX_RBBM_CLOCK_CNTL2_RB0, 0x00222222},
421 	{REG_A5XX_RBBM_CLOCK_CNTL2_RB1, 0x00222222},
422 	{REG_A5XX_RBBM_CLOCK_CNTL_CCU0, 0x00022220},
423 	{REG_A5XX_RBBM_CLOCK_CNTL_CCU1, 0x00022220},
424 	{REG_A5XX_RBBM_CLOCK_CNTL_RAC, 0x05522222},
425 	{REG_A5XX_RBBM_CLOCK_CNTL2_RAC, 0x00505555},
426 	{REG_A5XX_RBBM_CLOCK_HYST_RB_CCU0, 0x04040404},
427 	{REG_A5XX_RBBM_CLOCK_HYST_RB_CCU1, 0x04040404},
428 	{REG_A5XX_RBBM_CLOCK_HYST_RAC, 0x07444044},
429 	{REG_A5XX_RBBM_CLOCK_DELAY_RB_CCU_L1_0, 0x00000002},
430 	{REG_A5XX_RBBM_CLOCK_DELAY_RB_CCU_L1_1, 0x00000002},
431 	{REG_A5XX_RBBM_CLOCK_DELAY_RAC, 0x00010011},
432 	{REG_A5XX_RBBM_CLOCK_CNTL_TSE_RAS_RBBM, 0x04222222},
433 	{REG_A5XX_RBBM_CLOCK_MODE_GPC, 0x02222222},
434 	{REG_A5XX_RBBM_CLOCK_MODE_VFD, 0x00002222},
435 	{REG_A5XX_RBBM_CLOCK_HYST_TSE_RAS_RBBM, 0x00000000},
436 	{REG_A5XX_RBBM_CLOCK_HYST_GPC, 0x04104004},
437 	{REG_A5XX_RBBM_CLOCK_HYST_VFD, 0x00000000},
438 	{REG_A5XX_RBBM_CLOCK_DELAY_HLSQ, 0x00000000},
439 	{REG_A5XX_RBBM_CLOCK_DELAY_TSE_RAS_RBBM, 0x00004000},
440 	{REG_A5XX_RBBM_CLOCK_DELAY_GPC, 0x00000200},
441 	{REG_A5XX_RBBM_CLOCK_DELAY_VFD, 0x00002222},
442 };
443 
444 void a5xx_set_hwcg(struct msm_gpu *gpu, bool state)
445 {
446 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
447 	const struct adreno_five_hwcg_regs *regs;
448 	unsigned int i, sz;
449 
450 	if (adreno_is_a505(adreno_gpu) || adreno_is_a506(adreno_gpu) ||
451 	    adreno_is_a508(adreno_gpu)) {
452 		regs = a50x_hwcg;
453 		sz = ARRAY_SIZE(a50x_hwcg);
454 	} else if (adreno_is_a509(adreno_gpu) || adreno_is_a512(adreno_gpu)) {
455 		regs = a512_hwcg;
456 		sz = ARRAY_SIZE(a512_hwcg);
457 	} else {
458 		regs = a5xx_hwcg;
459 		sz = ARRAY_SIZE(a5xx_hwcg);
460 	}
461 
462 	for (i = 0; i < sz; i++)
463 		gpu_write(gpu, regs[i].offset,
464 			  state ? regs[i].value : 0);
465 
466 	if (adreno_is_a540(adreno_gpu)) {
467 		gpu_write(gpu, REG_A5XX_RBBM_CLOCK_DELAY_GPMU, state ? 0x00000770 : 0);
468 		gpu_write(gpu, REG_A5XX_RBBM_CLOCK_HYST_GPMU, state ? 0x00000004 : 0);
469 	}
470 
471 	gpu_write(gpu, REG_A5XX_RBBM_CLOCK_CNTL, state ? 0xAAA8AA00 : 0);
472 	gpu_write(gpu, REG_A5XX_RBBM_ISDB_CNT, state ? 0x182 : 0x180);
473 }
474 
475 static int a5xx_me_init(struct msm_gpu *gpu)
476 {
477 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
478 	struct msm_ringbuffer *ring = gpu->rb[0];
479 
480 	OUT_PKT7(ring, CP_ME_INIT, 8);
481 
482 	OUT_RING(ring, 0x0000002F);
483 
484 	/* Enable multiple hardware contexts */
485 	OUT_RING(ring, 0x00000003);
486 
487 	/* Enable error detection */
488 	OUT_RING(ring, 0x20000000);
489 
490 	/* Don't enable header dump */
491 	OUT_RING(ring, 0x00000000);
492 	OUT_RING(ring, 0x00000000);
493 
494 	/* Specify workarounds for various microcode issues */
495 	if (adreno_is_a505(adreno_gpu) || adreno_is_a506(adreno_gpu) ||
496 	    adreno_is_a530(adreno_gpu)) {
497 		/* Workaround for token end syncs
498 		 * Force a WFI after every direct-render 3D mode draw and every
499 		 * 2D mode 3 draw
500 		 */
501 		OUT_RING(ring, 0x0000000B);
502 	} else if (adreno_is_a510(adreno_gpu)) {
503 		/* Workaround for token and syncs */
504 		OUT_RING(ring, 0x00000001);
505 	} else {
506 		/* No workarounds enabled */
507 		OUT_RING(ring, 0x00000000);
508 	}
509 
510 	OUT_RING(ring, 0x00000000);
511 	OUT_RING(ring, 0x00000000);
512 
513 	a5xx_flush(gpu, ring, true);
514 	return a5xx_idle(gpu, ring) ? 0 : -EINVAL;
515 }
516 
517 static int a5xx_preempt_start(struct msm_gpu *gpu)
518 {
519 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
520 	struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu);
521 	struct msm_ringbuffer *ring = gpu->rb[0];
522 
523 	if (gpu->nr_rings == 1)
524 		return 0;
525 
526 	/* Turn off protected mode to write to special registers */
527 	OUT_PKT7(ring, CP_SET_PROTECTED_MODE, 1);
528 	OUT_RING(ring, 0);
529 
530 	/* Set the save preemption record for the ring/command */
531 	OUT_PKT4(ring, REG_A5XX_CP_CONTEXT_SWITCH_SAVE_ADDR_LO, 2);
532 	OUT_RING(ring, lower_32_bits(a5xx_gpu->preempt_iova[ring->id]));
533 	OUT_RING(ring, upper_32_bits(a5xx_gpu->preempt_iova[ring->id]));
534 
535 	/* Turn back on protected mode */
536 	OUT_PKT7(ring, CP_SET_PROTECTED_MODE, 1);
537 	OUT_RING(ring, 1);
538 
539 	OUT_PKT7(ring, CP_PREEMPT_ENABLE_GLOBAL, 1);
540 	OUT_RING(ring, 0x00);
541 
542 	OUT_PKT7(ring, CP_PREEMPT_ENABLE_LOCAL, 1);
543 	OUT_RING(ring, 0x01);
544 
545 	OUT_PKT7(ring, CP_YIELD_ENABLE, 1);
546 	OUT_RING(ring, 0x01);
547 
548 	/* Yield the floor on command completion */
549 	OUT_PKT7(ring, CP_CONTEXT_SWITCH_YIELD, 4);
550 	OUT_RING(ring, 0x00);
551 	OUT_RING(ring, 0x00);
552 	OUT_RING(ring, 0x01);
553 	OUT_RING(ring, 0x01);
554 
555 	/* The WHERE_AMI_I packet is not needed after a YIELD is issued */
556 	a5xx_flush(gpu, ring, false);
557 
558 	return a5xx_idle(gpu, ring) ? 0 : -EINVAL;
559 }
560 
561 static void a5xx_ucode_check_version(struct a5xx_gpu *a5xx_gpu,
562 		struct drm_gem_object *obj)
563 {
564 	u32 *buf = msm_gem_get_vaddr(obj);
565 
566 	if (IS_ERR(buf))
567 		return;
568 
569 	/*
570 	 * If the lowest nibble is 0xa that is an indication that this microcode
571 	 * has been patched. The actual version is in dword [3] but we only care
572 	 * about the patchlevel which is the lowest nibble of dword [3]
573 	 */
574 	if (((buf[0] & 0xf) == 0xa) && (buf[2] & 0xf) >= 1)
575 		a5xx_gpu->has_whereami = true;
576 
577 	msm_gem_put_vaddr(obj);
578 }
579 
580 static int a5xx_ucode_load(struct msm_gpu *gpu)
581 {
582 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
583 	struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu);
584 	int ret;
585 
586 	if (!a5xx_gpu->pm4_bo) {
587 		a5xx_gpu->pm4_bo = adreno_fw_create_bo(gpu,
588 			adreno_gpu->fw[ADRENO_FW_PM4], &a5xx_gpu->pm4_iova);
589 
590 
591 		if (IS_ERR(a5xx_gpu->pm4_bo)) {
592 			ret = PTR_ERR(a5xx_gpu->pm4_bo);
593 			a5xx_gpu->pm4_bo = NULL;
594 			DRM_DEV_ERROR(gpu->dev->dev, "could not allocate PM4: %d\n",
595 				ret);
596 			return ret;
597 		}
598 
599 		msm_gem_object_set_name(a5xx_gpu->pm4_bo, "pm4fw");
600 	}
601 
602 	if (!a5xx_gpu->pfp_bo) {
603 		a5xx_gpu->pfp_bo = adreno_fw_create_bo(gpu,
604 			adreno_gpu->fw[ADRENO_FW_PFP], &a5xx_gpu->pfp_iova);
605 
606 		if (IS_ERR(a5xx_gpu->pfp_bo)) {
607 			ret = PTR_ERR(a5xx_gpu->pfp_bo);
608 			a5xx_gpu->pfp_bo = NULL;
609 			DRM_DEV_ERROR(gpu->dev->dev, "could not allocate PFP: %d\n",
610 				ret);
611 			return ret;
612 		}
613 
614 		msm_gem_object_set_name(a5xx_gpu->pfp_bo, "pfpfw");
615 		a5xx_ucode_check_version(a5xx_gpu, a5xx_gpu->pfp_bo);
616 	}
617 
618 	if (a5xx_gpu->has_whereami) {
619 		if (!a5xx_gpu->shadow_bo) {
620 			a5xx_gpu->shadow = msm_gem_kernel_new(gpu->dev,
621 				sizeof(u32) * gpu->nr_rings,
622 				MSM_BO_WC | MSM_BO_MAP_PRIV,
623 				gpu->aspace, &a5xx_gpu->shadow_bo,
624 				&a5xx_gpu->shadow_iova);
625 
626 			if (IS_ERR(a5xx_gpu->shadow))
627 				return PTR_ERR(a5xx_gpu->shadow);
628 
629 			msm_gem_object_set_name(a5xx_gpu->shadow_bo, "shadow");
630 		}
631 	} else if (gpu->nr_rings > 1) {
632 		/* Disable preemption if WHERE_AM_I isn't available */
633 		a5xx_preempt_fini(gpu);
634 		gpu->nr_rings = 1;
635 	}
636 
637 	return 0;
638 }
639 
640 #define SCM_GPU_ZAP_SHADER_RESUME 0
641 
642 static int a5xx_zap_shader_resume(struct msm_gpu *gpu)
643 {
644 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
645 	int ret;
646 
647 	/*
648 	 * Adreno 506 have CPZ Retention feature and doesn't require
649 	 * to resume zap shader
650 	 */
651 	if (adreno_is_a506(adreno_gpu))
652 		return 0;
653 
654 	ret = qcom_scm_set_remote_state(SCM_GPU_ZAP_SHADER_RESUME, GPU_PAS_ID);
655 	if (ret)
656 		DRM_ERROR("%s: zap-shader resume failed: %d\n",
657 			gpu->name, ret);
658 
659 	return ret;
660 }
661 
662 static int a5xx_zap_shader_init(struct msm_gpu *gpu)
663 {
664 	static bool loaded;
665 	int ret;
666 
667 	/*
668 	 * If the zap shader is already loaded into memory we just need to kick
669 	 * the remote processor to reinitialize it
670 	 */
671 	if (loaded)
672 		return a5xx_zap_shader_resume(gpu);
673 
674 	ret = adreno_zap_shader_load(gpu, GPU_PAS_ID);
675 
676 	loaded = !ret;
677 	return ret;
678 }
679 
680 #define A5XX_INT_MASK (A5XX_RBBM_INT_0_MASK_RBBM_AHB_ERROR | \
681 	  A5XX_RBBM_INT_0_MASK_RBBM_TRANSFER_TIMEOUT | \
682 	  A5XX_RBBM_INT_0_MASK_RBBM_ME_MS_TIMEOUT | \
683 	  A5XX_RBBM_INT_0_MASK_RBBM_PFP_MS_TIMEOUT | \
684 	  A5XX_RBBM_INT_0_MASK_RBBM_ETS_MS_TIMEOUT | \
685 	  A5XX_RBBM_INT_0_MASK_RBBM_ATB_ASYNC_OVERFLOW | \
686 	  A5XX_RBBM_INT_0_MASK_CP_HW_ERROR | \
687 	  A5XX_RBBM_INT_0_MASK_MISC_HANG_DETECT | \
688 	  A5XX_RBBM_INT_0_MASK_CP_SW | \
689 	  A5XX_RBBM_INT_0_MASK_CP_CACHE_FLUSH_TS | \
690 	  A5XX_RBBM_INT_0_MASK_UCHE_OOB_ACCESS | \
691 	  A5XX_RBBM_INT_0_MASK_GPMU_VOLTAGE_DROOP)
692 
693 static int a5xx_hw_init(struct msm_gpu *gpu)
694 {
695 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
696 	struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu);
697 	u32 hbb;
698 	int ret;
699 
700 	gpu_write(gpu, REG_A5XX_VBIF_ROUND_ROBIN_QOS_ARB, 0x00000003);
701 
702 	if (adreno_is_a509(adreno_gpu) || adreno_is_a512(adreno_gpu) ||
703 	    adreno_is_a540(adreno_gpu))
704 		gpu_write(gpu, REG_A5XX_VBIF_GATE_OFF_WRREQ_EN, 0x00000009);
705 
706 	/* Make all blocks contribute to the GPU BUSY perf counter */
707 	gpu_write(gpu, REG_A5XX_RBBM_PERFCTR_GPU_BUSY_MASKED, 0xFFFFFFFF);
708 
709 	/* Enable RBBM error reporting bits */
710 	gpu_write(gpu, REG_A5XX_RBBM_AHB_CNTL0, 0x00000001);
711 
712 	if (adreno_gpu->info->quirks & ADRENO_QUIRK_FAULT_DETECT_MASK) {
713 		/*
714 		 * Mask out the activity signals from RB1-3 to avoid false
715 		 * positives
716 		 */
717 
718 		gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL11,
719 			0xF0000000);
720 		gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL12,
721 			0xFFFFFFFF);
722 		gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL13,
723 			0xFFFFFFFF);
724 		gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL14,
725 			0xFFFFFFFF);
726 		gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL15,
727 			0xFFFFFFFF);
728 		gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL16,
729 			0xFFFFFFFF);
730 		gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL17,
731 			0xFFFFFFFF);
732 		gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL18,
733 			0xFFFFFFFF);
734 	}
735 
736 	/* Enable fault detection */
737 	gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_INT_CNTL,
738 		(1 << 30) | 0xFFFF);
739 
740 	/* Turn on performance counters */
741 	gpu_write(gpu, REG_A5XX_RBBM_PERFCTR_CNTL, 0x01);
742 
743 	/* Select CP0 to always count cycles */
744 	gpu_write(gpu, REG_A5XX_CP_PERFCTR_CP_SEL_0, PERF_CP_ALWAYS_COUNT);
745 
746 	/* Select RBBM0 to countable 6 to get the busy status for devfreq */
747 	gpu_write(gpu, REG_A5XX_RBBM_PERFCTR_RBBM_SEL_0, 6);
748 
749 	/* Increase VFD cache access so LRZ and other data gets evicted less */
750 	gpu_write(gpu, REG_A5XX_UCHE_CACHE_WAYS, 0x02);
751 
752 	/* Disable L2 bypass in the UCHE */
753 	gpu_write(gpu, REG_A5XX_UCHE_TRAP_BASE_LO, 0xFFFF0000);
754 	gpu_write(gpu, REG_A5XX_UCHE_TRAP_BASE_HI, 0x0001FFFF);
755 	gpu_write(gpu, REG_A5XX_UCHE_WRITE_THRU_BASE_LO, 0xFFFF0000);
756 	gpu_write(gpu, REG_A5XX_UCHE_WRITE_THRU_BASE_HI, 0x0001FFFF);
757 
758 	/* Set the GMEM VA range (0 to gpu->gmem) */
759 	gpu_write(gpu, REG_A5XX_UCHE_GMEM_RANGE_MIN_LO, 0x00100000);
760 	gpu_write(gpu, REG_A5XX_UCHE_GMEM_RANGE_MIN_HI, 0x00000000);
761 	gpu_write(gpu, REG_A5XX_UCHE_GMEM_RANGE_MAX_LO,
762 		0x00100000 + adreno_gpu->info->gmem - 1);
763 	gpu_write(gpu, REG_A5XX_UCHE_GMEM_RANGE_MAX_HI, 0x00000000);
764 
765 	if (adreno_is_a505(adreno_gpu) || adreno_is_a506(adreno_gpu) ||
766 	    adreno_is_a508(adreno_gpu) || adreno_is_a510(adreno_gpu)) {
767 		gpu_write(gpu, REG_A5XX_CP_MEQ_THRESHOLDS, 0x20);
768 		if (adreno_is_a505(adreno_gpu) || adreno_is_a506(adreno_gpu) ||
769 			adreno_is_a508(adreno_gpu))
770 			gpu_write(gpu, REG_A5XX_CP_MERCIU_SIZE, 0x400);
771 		else
772 			gpu_write(gpu, REG_A5XX_CP_MERCIU_SIZE, 0x20);
773 		gpu_write(gpu, REG_A5XX_CP_ROQ_THRESHOLDS_2, 0x40000030);
774 		gpu_write(gpu, REG_A5XX_CP_ROQ_THRESHOLDS_1, 0x20100D0A);
775 	} else {
776 		gpu_write(gpu, REG_A5XX_CP_MEQ_THRESHOLDS, 0x40);
777 		if (adreno_is_a530(adreno_gpu))
778 			gpu_write(gpu, REG_A5XX_CP_MERCIU_SIZE, 0x40);
779 		else
780 			gpu_write(gpu, REG_A5XX_CP_MERCIU_SIZE, 0x400);
781 		gpu_write(gpu, REG_A5XX_CP_ROQ_THRESHOLDS_2, 0x80000060);
782 		gpu_write(gpu, REG_A5XX_CP_ROQ_THRESHOLDS_1, 0x40201B16);
783 	}
784 
785 	if (adreno_is_a505(adreno_gpu) || adreno_is_a506(adreno_gpu) ||
786 	    adreno_is_a508(adreno_gpu))
787 		gpu_write(gpu, REG_A5XX_PC_DBG_ECO_CNTL,
788 			  (0x100 << 11 | 0x100 << 22));
789 	else if (adreno_is_a509(adreno_gpu) || adreno_is_a510(adreno_gpu) ||
790 		 adreno_is_a512(adreno_gpu))
791 		gpu_write(gpu, REG_A5XX_PC_DBG_ECO_CNTL,
792 			  (0x200 << 11 | 0x200 << 22));
793 	else
794 		gpu_write(gpu, REG_A5XX_PC_DBG_ECO_CNTL,
795 			  (0x400 << 11 | 0x300 << 22));
796 
797 	if (adreno_gpu->info->quirks & ADRENO_QUIRK_TWO_PASS_USE_WFI)
798 		gpu_rmw(gpu, REG_A5XX_PC_DBG_ECO_CNTL, 0, (1 << 8));
799 
800 	/*
801 	 * Disable the RB sampler datapath DP2 clock gating optimization
802 	 * for 1-SP GPUs, as it is enabled by default.
803 	 */
804 	if (adreno_is_a505(adreno_gpu) || adreno_is_a506(adreno_gpu) ||
805 	    adreno_is_a508(adreno_gpu) || adreno_is_a509(adreno_gpu) ||
806 	    adreno_is_a512(adreno_gpu))
807 		gpu_rmw(gpu, REG_A5XX_RB_DBG_ECO_CNTL, 0, (1 << 9));
808 
809 	/* Disable UCHE global filter as SP can invalidate/flush independently */
810 	gpu_write(gpu, REG_A5XX_UCHE_MODE_CNTL, BIT(29));
811 
812 	/* Enable USE_RETENTION_FLOPS */
813 	gpu_write(gpu, REG_A5XX_CP_CHICKEN_DBG, 0x02000000);
814 
815 	/* Enable ME/PFP split notification */
816 	gpu_write(gpu, REG_A5XX_RBBM_AHB_CNTL1, 0xA6FFFFFF);
817 
818 	/*
819 	 *  In A5x, CCU can send context_done event of a particular context to
820 	 *  UCHE which ultimately reaches CP even when there is valid
821 	 *  transaction of that context inside CCU. This can let CP to program
822 	 *  config registers, which will make the "valid transaction" inside
823 	 *  CCU to be interpreted differently. This can cause gpu fault. This
824 	 *  bug is fixed in latest A510 revision. To enable this bug fix -
825 	 *  bit[11] of RB_DBG_ECO_CNTL need to be set to 0, default is 1
826 	 *  (disable). For older A510 version this bit is unused.
827 	 */
828 	if (adreno_is_a510(adreno_gpu))
829 		gpu_rmw(gpu, REG_A5XX_RB_DBG_ECO_CNTL, (1 << 11), 0);
830 
831 	/* Enable HWCG */
832 	a5xx_set_hwcg(gpu, true);
833 
834 	gpu_write(gpu, REG_A5XX_RBBM_AHB_CNTL2, 0x0000003F);
835 
836 	BUG_ON(adreno_gpu->ubwc_config.highest_bank_bit < 13);
837 	hbb = adreno_gpu->ubwc_config.highest_bank_bit - 13;
838 
839 	gpu_write(gpu, REG_A5XX_TPL1_MODE_CNTL, hbb << 7);
840 	gpu_write(gpu, REG_A5XX_RB_MODE_CNTL, hbb << 1);
841 
842 	if (adreno_is_a509(adreno_gpu) || adreno_is_a512(adreno_gpu) ||
843 	    adreno_is_a540(adreno_gpu))
844 		gpu_write(gpu, REG_A5XX_UCHE_DBG_ECO_CNTL_2, hbb);
845 
846 	/* Disable All flat shading optimization (ALLFLATOPTDIS) */
847 	gpu_rmw(gpu, REG_A5XX_VPC_DBG_ECO_CNTL, 0, (1 << 10));
848 
849 	/* Protect registers from the CP */
850 	gpu_write(gpu, REG_A5XX_CP_PROTECT_CNTL, 0x00000007);
851 
852 	/* RBBM */
853 	gpu_write(gpu, REG_A5XX_CP_PROTECT(0), ADRENO_PROTECT_RW(0x04, 4));
854 	gpu_write(gpu, REG_A5XX_CP_PROTECT(1), ADRENO_PROTECT_RW(0x08, 8));
855 	gpu_write(gpu, REG_A5XX_CP_PROTECT(2), ADRENO_PROTECT_RW(0x10, 16));
856 	gpu_write(gpu, REG_A5XX_CP_PROTECT(3), ADRENO_PROTECT_RW(0x20, 32));
857 	gpu_write(gpu, REG_A5XX_CP_PROTECT(4), ADRENO_PROTECT_RW(0x40, 64));
858 	gpu_write(gpu, REG_A5XX_CP_PROTECT(5), ADRENO_PROTECT_RW(0x80, 64));
859 
860 	/* Content protect */
861 	gpu_write(gpu, REG_A5XX_CP_PROTECT(6),
862 		ADRENO_PROTECT_RW(REG_A5XX_RBBM_SECVID_TSB_TRUSTED_BASE_LO,
863 			16));
864 	gpu_write(gpu, REG_A5XX_CP_PROTECT(7),
865 		ADRENO_PROTECT_RW(REG_A5XX_RBBM_SECVID_TRUST_CNTL, 2));
866 
867 	/* CP */
868 	gpu_write(gpu, REG_A5XX_CP_PROTECT(8), ADRENO_PROTECT_RW(0x800, 64));
869 	gpu_write(gpu, REG_A5XX_CP_PROTECT(9), ADRENO_PROTECT_RW(0x840, 8));
870 	gpu_write(gpu, REG_A5XX_CP_PROTECT(10), ADRENO_PROTECT_RW(0x880, 32));
871 	gpu_write(gpu, REG_A5XX_CP_PROTECT(11), ADRENO_PROTECT_RW(0xAA0, 1));
872 
873 	/* RB */
874 	gpu_write(gpu, REG_A5XX_CP_PROTECT(12), ADRENO_PROTECT_RW(0xCC0, 1));
875 	gpu_write(gpu, REG_A5XX_CP_PROTECT(13), ADRENO_PROTECT_RW(0xCF0, 2));
876 
877 	/* VPC */
878 	gpu_write(gpu, REG_A5XX_CP_PROTECT(14), ADRENO_PROTECT_RW(0xE68, 8));
879 	gpu_write(gpu, REG_A5XX_CP_PROTECT(15), ADRENO_PROTECT_RW(0xE70, 16));
880 
881 	/* UCHE */
882 	gpu_write(gpu, REG_A5XX_CP_PROTECT(16), ADRENO_PROTECT_RW(0xE80, 16));
883 
884 	/* SMMU */
885 	gpu_write(gpu, REG_A5XX_CP_PROTECT(17),
886 			ADRENO_PROTECT_RW(0x10000, 0x8000));
887 
888 	gpu_write(gpu, REG_A5XX_RBBM_SECVID_TSB_CNTL, 0);
889 	/*
890 	 * Disable the trusted memory range - we don't actually supported secure
891 	 * memory rendering at this point in time and we don't want to block off
892 	 * part of the virtual memory space.
893 	 */
894 	gpu_write64(gpu, REG_A5XX_RBBM_SECVID_TSB_TRUSTED_BASE_LO, 0x00000000);
895 	gpu_write(gpu, REG_A5XX_RBBM_SECVID_TSB_TRUSTED_SIZE, 0x00000000);
896 
897 	/* Put the GPU into 64 bit by default */
898 	gpu_write(gpu, REG_A5XX_CP_ADDR_MODE_CNTL, 0x1);
899 	gpu_write(gpu, REG_A5XX_VSC_ADDR_MODE_CNTL, 0x1);
900 	gpu_write(gpu, REG_A5XX_GRAS_ADDR_MODE_CNTL, 0x1);
901 	gpu_write(gpu, REG_A5XX_RB_ADDR_MODE_CNTL, 0x1);
902 	gpu_write(gpu, REG_A5XX_PC_ADDR_MODE_CNTL, 0x1);
903 	gpu_write(gpu, REG_A5XX_HLSQ_ADDR_MODE_CNTL, 0x1);
904 	gpu_write(gpu, REG_A5XX_VFD_ADDR_MODE_CNTL, 0x1);
905 	gpu_write(gpu, REG_A5XX_VPC_ADDR_MODE_CNTL, 0x1);
906 	gpu_write(gpu, REG_A5XX_UCHE_ADDR_MODE_CNTL, 0x1);
907 	gpu_write(gpu, REG_A5XX_SP_ADDR_MODE_CNTL, 0x1);
908 	gpu_write(gpu, REG_A5XX_TPL1_ADDR_MODE_CNTL, 0x1);
909 	gpu_write(gpu, REG_A5XX_RBBM_SECVID_TSB_ADDR_MODE_CNTL, 0x1);
910 
911 	/*
912 	 * VPC corner case with local memory load kill leads to corrupt
913 	 * internal state. Normal Disable does not work for all a5x chips.
914 	 * So do the following setting to disable it.
915 	 */
916 	if (adreno_gpu->info->quirks & ADRENO_QUIRK_LMLOADKILL_DISABLE) {
917 		gpu_rmw(gpu, REG_A5XX_VPC_DBG_ECO_CNTL, 0, BIT(23));
918 		gpu_rmw(gpu, REG_A5XX_HLSQ_DBG_ECO_CNTL, BIT(18), 0);
919 	}
920 
921 	ret = adreno_hw_init(gpu);
922 	if (ret)
923 		return ret;
924 
925 	if (adreno_is_a530(adreno_gpu) || adreno_is_a540(adreno_gpu))
926 		a5xx_gpmu_ucode_init(gpu);
927 
928 	gpu_write64(gpu, REG_A5XX_CP_ME_INSTR_BASE_LO, a5xx_gpu->pm4_iova);
929 	gpu_write64(gpu, REG_A5XX_CP_PFP_INSTR_BASE_LO, a5xx_gpu->pfp_iova);
930 
931 	/* Set the ringbuffer address */
932 	gpu_write64(gpu, REG_A5XX_CP_RB_BASE, gpu->rb[0]->iova);
933 
934 	/*
935 	 * If the microcode supports the WHERE_AM_I opcode then we can use that
936 	 * in lieu of the RPTR shadow and enable preemption. Otherwise, we
937 	 * can't safely use the RPTR shadow or preemption. In either case, the
938 	 * RPTR shadow should be disabled in hardware.
939 	 */
940 	gpu_write(gpu, REG_A5XX_CP_RB_CNTL,
941 		MSM_GPU_RB_CNTL_DEFAULT | AXXX_CP_RB_CNTL_NO_UPDATE);
942 
943 	/* Configure the RPTR shadow if needed: */
944 	if (a5xx_gpu->shadow_bo) {
945 		gpu_write64(gpu, REG_A5XX_CP_RB_RPTR_ADDR,
946 			    shadowptr(a5xx_gpu, gpu->rb[0]));
947 	}
948 
949 	a5xx_preempt_hw_init(gpu);
950 
951 	/* Disable the interrupts through the initial bringup stage */
952 	gpu_write(gpu, REG_A5XX_RBBM_INT_0_MASK, A5XX_INT_MASK);
953 
954 	/* Clear ME_HALT to start the micro engine */
955 	gpu_write(gpu, REG_A5XX_CP_PFP_ME_CNTL, 0);
956 	ret = a5xx_me_init(gpu);
957 	if (ret)
958 		return ret;
959 
960 	ret = a5xx_power_init(gpu);
961 	if (ret)
962 		return ret;
963 
964 	/*
965 	 * Send a pipeline event stat to get misbehaving counters to start
966 	 * ticking correctly
967 	 */
968 	if (adreno_is_a530(adreno_gpu)) {
969 		OUT_PKT7(gpu->rb[0], CP_EVENT_WRITE, 1);
970 		OUT_RING(gpu->rb[0], CP_EVENT_WRITE_0_EVENT(STAT_EVENT));
971 
972 		a5xx_flush(gpu, gpu->rb[0], true);
973 		if (!a5xx_idle(gpu, gpu->rb[0]))
974 			return -EINVAL;
975 	}
976 
977 	/*
978 	 * If the chip that we are using does support loading one, then
979 	 * try to load a zap shader into the secure world. If successful
980 	 * we can use the CP to switch out of secure mode. If not then we
981 	 * have no resource but to try to switch ourselves out manually. If we
982 	 * guessed wrong then access to the RBBM_SECVID_TRUST_CNTL register will
983 	 * be blocked and a permissions violation will soon follow.
984 	 */
985 	ret = a5xx_zap_shader_init(gpu);
986 	if (!ret) {
987 		OUT_PKT7(gpu->rb[0], CP_SET_SECURE_MODE, 1);
988 		OUT_RING(gpu->rb[0], 0x00000000);
989 
990 		a5xx_flush(gpu, gpu->rb[0], true);
991 		if (!a5xx_idle(gpu, gpu->rb[0]))
992 			return -EINVAL;
993 	} else if (ret == -ENODEV) {
994 		/*
995 		 * This device does not use zap shader (but print a warning
996 		 * just in case someone got their dt wrong.. hopefully they
997 		 * have a debug UART to realize the error of their ways...
998 		 * if you mess this up you are about to crash horribly)
999 		 */
1000 		dev_warn_once(gpu->dev->dev,
1001 			"Zap shader not enabled - using SECVID_TRUST_CNTL instead\n");
1002 		gpu_write(gpu, REG_A5XX_RBBM_SECVID_TRUST_CNTL, 0x0);
1003 	} else {
1004 		return ret;
1005 	}
1006 
1007 	/* Last step - yield the ringbuffer */
1008 	a5xx_preempt_start(gpu);
1009 
1010 	return 0;
1011 }
1012 
1013 static void a5xx_recover(struct msm_gpu *gpu)
1014 {
1015 	int i;
1016 
1017 	adreno_dump_info(gpu);
1018 
1019 	for (i = 0; i < 8; i++) {
1020 		printk("CP_SCRATCH_REG%d: %u\n", i,
1021 			gpu_read(gpu, REG_A5XX_CP_SCRATCH_REG(i)));
1022 	}
1023 
1024 	if (hang_debug)
1025 		a5xx_dump(gpu);
1026 
1027 	gpu_write(gpu, REG_A5XX_RBBM_SW_RESET_CMD, 1);
1028 	gpu_read(gpu, REG_A5XX_RBBM_SW_RESET_CMD);
1029 	gpu_write(gpu, REG_A5XX_RBBM_SW_RESET_CMD, 0);
1030 	adreno_recover(gpu);
1031 }
1032 
1033 static void a5xx_destroy(struct msm_gpu *gpu)
1034 {
1035 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1036 	struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu);
1037 
1038 	DBG("%s", gpu->name);
1039 
1040 	a5xx_preempt_fini(gpu);
1041 
1042 	if (a5xx_gpu->pm4_bo) {
1043 		msm_gem_unpin_iova(a5xx_gpu->pm4_bo, gpu->aspace);
1044 		drm_gem_object_put(a5xx_gpu->pm4_bo);
1045 	}
1046 
1047 	if (a5xx_gpu->pfp_bo) {
1048 		msm_gem_unpin_iova(a5xx_gpu->pfp_bo, gpu->aspace);
1049 		drm_gem_object_put(a5xx_gpu->pfp_bo);
1050 	}
1051 
1052 	if (a5xx_gpu->gpmu_bo) {
1053 		msm_gem_unpin_iova(a5xx_gpu->gpmu_bo, gpu->aspace);
1054 		drm_gem_object_put(a5xx_gpu->gpmu_bo);
1055 	}
1056 
1057 	if (a5xx_gpu->shadow_bo) {
1058 		msm_gem_unpin_iova(a5xx_gpu->shadow_bo, gpu->aspace);
1059 		drm_gem_object_put(a5xx_gpu->shadow_bo);
1060 	}
1061 
1062 	adreno_gpu_cleanup(adreno_gpu);
1063 	kfree(a5xx_gpu);
1064 }
1065 
1066 static inline bool _a5xx_check_idle(struct msm_gpu *gpu)
1067 {
1068 	if (gpu_read(gpu, REG_A5XX_RBBM_STATUS) & ~A5XX_RBBM_STATUS_HI_BUSY)
1069 		return false;
1070 
1071 	/*
1072 	 * Nearly every abnormality ends up pausing the GPU and triggering a
1073 	 * fault so we can safely just watch for this one interrupt to fire
1074 	 */
1075 	return !(gpu_read(gpu, REG_A5XX_RBBM_INT_0_STATUS) &
1076 		A5XX_RBBM_INT_0_MASK_MISC_HANG_DETECT);
1077 }
1078 
1079 bool a5xx_idle(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
1080 {
1081 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1082 	struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu);
1083 
1084 	if (ring != a5xx_gpu->cur_ring) {
1085 		WARN(1, "Tried to idle a non-current ringbuffer\n");
1086 		return false;
1087 	}
1088 
1089 	/* wait for CP to drain ringbuffer: */
1090 	if (!adreno_idle(gpu, ring))
1091 		return false;
1092 
1093 	if (spin_until(_a5xx_check_idle(gpu))) {
1094 		DRM_ERROR("%s: %ps: timeout waiting for GPU to idle: status %8.8X irq %8.8X rptr/wptr %d/%d\n",
1095 			gpu->name, __builtin_return_address(0),
1096 			gpu_read(gpu, REG_A5XX_RBBM_STATUS),
1097 			gpu_read(gpu, REG_A5XX_RBBM_INT_0_STATUS),
1098 			gpu_read(gpu, REG_A5XX_CP_RB_RPTR),
1099 			gpu_read(gpu, REG_A5XX_CP_RB_WPTR));
1100 		return false;
1101 	}
1102 
1103 	return true;
1104 }
1105 
1106 static int a5xx_fault_handler(void *arg, unsigned long iova, int flags, void *data)
1107 {
1108 	struct msm_gpu *gpu = arg;
1109 	struct adreno_smmu_fault_info *info = data;
1110 	char block[12] = "unknown";
1111 	u32 scratch[] = {
1112 			gpu_read(gpu, REG_A5XX_CP_SCRATCH_REG(4)),
1113 			gpu_read(gpu, REG_A5XX_CP_SCRATCH_REG(5)),
1114 			gpu_read(gpu, REG_A5XX_CP_SCRATCH_REG(6)),
1115 			gpu_read(gpu, REG_A5XX_CP_SCRATCH_REG(7)),
1116 	};
1117 
1118 	if (info)
1119 		snprintf(block, sizeof(block), "%x", info->fsynr1);
1120 
1121 	return adreno_fault_handler(gpu, iova, flags, info, block, scratch);
1122 }
1123 
1124 static void a5xx_cp_err_irq(struct msm_gpu *gpu)
1125 {
1126 	u32 status = gpu_read(gpu, REG_A5XX_CP_INTERRUPT_STATUS);
1127 
1128 	if (status & A5XX_CP_INT_CP_OPCODE_ERROR) {
1129 		u32 val;
1130 
1131 		gpu_write(gpu, REG_A5XX_CP_PFP_STAT_ADDR, 0);
1132 
1133 		/*
1134 		 * REG_A5XX_CP_PFP_STAT_DATA is indexed, and we want index 1 so
1135 		 * read it twice
1136 		 */
1137 
1138 		gpu_read(gpu, REG_A5XX_CP_PFP_STAT_DATA);
1139 		val = gpu_read(gpu, REG_A5XX_CP_PFP_STAT_DATA);
1140 
1141 		dev_err_ratelimited(gpu->dev->dev, "CP | opcode error | possible opcode=0x%8.8X\n",
1142 			val);
1143 	}
1144 
1145 	if (status & A5XX_CP_INT_CP_HW_FAULT_ERROR)
1146 		dev_err_ratelimited(gpu->dev->dev, "CP | HW fault | status=0x%8.8X\n",
1147 			gpu_read(gpu, REG_A5XX_CP_HW_FAULT));
1148 
1149 	if (status & A5XX_CP_INT_CP_DMA_ERROR)
1150 		dev_err_ratelimited(gpu->dev->dev, "CP | DMA error\n");
1151 
1152 	if (status & A5XX_CP_INT_CP_REGISTER_PROTECTION_ERROR) {
1153 		u32 val = gpu_read(gpu, REG_A5XX_CP_PROTECT_STATUS);
1154 
1155 		dev_err_ratelimited(gpu->dev->dev,
1156 			"CP | protected mode error | %s | addr=0x%8.8X | status=0x%8.8X\n",
1157 			val & (1 << 24) ? "WRITE" : "READ",
1158 			(val & 0xFFFFF) >> 2, val);
1159 	}
1160 
1161 	if (status & A5XX_CP_INT_CP_AHB_ERROR) {
1162 		u32 status = gpu_read(gpu, REG_A5XX_CP_AHB_FAULT);
1163 		const char *access[16] = { "reserved", "reserved",
1164 			"timestamp lo", "timestamp hi", "pfp read", "pfp write",
1165 			"", "", "me read", "me write", "", "", "crashdump read",
1166 			"crashdump write" };
1167 
1168 		dev_err_ratelimited(gpu->dev->dev,
1169 			"CP | AHB error | addr=%X access=%s error=%d | status=0x%8.8X\n",
1170 			status & 0xFFFFF, access[(status >> 24) & 0xF],
1171 			(status & (1 << 31)), status);
1172 	}
1173 }
1174 
1175 static void a5xx_rbbm_err_irq(struct msm_gpu *gpu, u32 status)
1176 {
1177 	if (status & A5XX_RBBM_INT_0_MASK_RBBM_AHB_ERROR) {
1178 		u32 val = gpu_read(gpu, REG_A5XX_RBBM_AHB_ERROR_STATUS);
1179 
1180 		dev_err_ratelimited(gpu->dev->dev,
1181 			"RBBM | AHB bus error | %s | addr=0x%X | ports=0x%X:0x%X\n",
1182 			val & (1 << 28) ? "WRITE" : "READ",
1183 			(val & 0xFFFFF) >> 2, (val >> 20) & 0x3,
1184 			(val >> 24) & 0xF);
1185 
1186 		/* Clear the error */
1187 		gpu_write(gpu, REG_A5XX_RBBM_AHB_CMD, (1 << 4));
1188 
1189 		/* Clear the interrupt */
1190 		gpu_write(gpu, REG_A5XX_RBBM_INT_CLEAR_CMD,
1191 			A5XX_RBBM_INT_0_MASK_RBBM_AHB_ERROR);
1192 	}
1193 
1194 	if (status & A5XX_RBBM_INT_0_MASK_RBBM_TRANSFER_TIMEOUT)
1195 		dev_err_ratelimited(gpu->dev->dev, "RBBM | AHB transfer timeout\n");
1196 
1197 	if (status & A5XX_RBBM_INT_0_MASK_RBBM_ME_MS_TIMEOUT)
1198 		dev_err_ratelimited(gpu->dev->dev, "RBBM | ME master split | status=0x%X\n",
1199 			gpu_read(gpu, REG_A5XX_RBBM_AHB_ME_SPLIT_STATUS));
1200 
1201 	if (status & A5XX_RBBM_INT_0_MASK_RBBM_PFP_MS_TIMEOUT)
1202 		dev_err_ratelimited(gpu->dev->dev, "RBBM | PFP master split | status=0x%X\n",
1203 			gpu_read(gpu, REG_A5XX_RBBM_AHB_PFP_SPLIT_STATUS));
1204 
1205 	if (status & A5XX_RBBM_INT_0_MASK_RBBM_ETS_MS_TIMEOUT)
1206 		dev_err_ratelimited(gpu->dev->dev, "RBBM | ETS master split | status=0x%X\n",
1207 			gpu_read(gpu, REG_A5XX_RBBM_AHB_ETS_SPLIT_STATUS));
1208 
1209 	if (status & A5XX_RBBM_INT_0_MASK_RBBM_ATB_ASYNC_OVERFLOW)
1210 		dev_err_ratelimited(gpu->dev->dev, "RBBM | ATB ASYNC overflow\n");
1211 
1212 	if (status & A5XX_RBBM_INT_0_MASK_RBBM_ATB_BUS_OVERFLOW)
1213 		dev_err_ratelimited(gpu->dev->dev, "RBBM | ATB bus overflow\n");
1214 }
1215 
1216 static void a5xx_uche_err_irq(struct msm_gpu *gpu)
1217 {
1218 	uint64_t addr = (uint64_t) gpu_read(gpu, REG_A5XX_UCHE_TRAP_LOG_HI);
1219 
1220 	addr |= gpu_read(gpu, REG_A5XX_UCHE_TRAP_LOG_LO);
1221 
1222 	dev_err_ratelimited(gpu->dev->dev, "UCHE | Out of bounds access | addr=0x%llX\n",
1223 		addr);
1224 }
1225 
1226 static void a5xx_gpmu_err_irq(struct msm_gpu *gpu)
1227 {
1228 	dev_err_ratelimited(gpu->dev->dev, "GPMU | voltage droop\n");
1229 }
1230 
1231 static void a5xx_fault_detect_irq(struct msm_gpu *gpu)
1232 {
1233 	struct drm_device *dev = gpu->dev;
1234 	struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
1235 
1236 	/*
1237 	 * If stalled on SMMU fault, we could trip the GPU's hang detection,
1238 	 * but the fault handler will trigger the devcore dump, and we want
1239 	 * to otherwise resume normally rather than killing the submit, so
1240 	 * just bail.
1241 	 */
1242 	if (gpu_read(gpu, REG_A5XX_RBBM_STATUS3) & BIT(24))
1243 		return;
1244 
1245 	DRM_DEV_ERROR(dev->dev, "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",
1246 		ring ? ring->id : -1, ring ? ring->fctx->last_fence : 0,
1247 		gpu_read(gpu, REG_A5XX_RBBM_STATUS),
1248 		gpu_read(gpu, REG_A5XX_CP_RB_RPTR),
1249 		gpu_read(gpu, REG_A5XX_CP_RB_WPTR),
1250 		gpu_read64(gpu, REG_A5XX_CP_IB1_BASE),
1251 		gpu_read(gpu, REG_A5XX_CP_IB1_BUFSZ),
1252 		gpu_read64(gpu, REG_A5XX_CP_IB2_BASE),
1253 		gpu_read(gpu, REG_A5XX_CP_IB2_BUFSZ));
1254 
1255 	/* Turn off the hangcheck timer to keep it from bothering us */
1256 	del_timer(&gpu->hangcheck_timer);
1257 
1258 	kthread_queue_work(gpu->worker, &gpu->recover_work);
1259 }
1260 
1261 #define RBBM_ERROR_MASK \
1262 	(A5XX_RBBM_INT_0_MASK_RBBM_AHB_ERROR | \
1263 	A5XX_RBBM_INT_0_MASK_RBBM_TRANSFER_TIMEOUT | \
1264 	A5XX_RBBM_INT_0_MASK_RBBM_ME_MS_TIMEOUT | \
1265 	A5XX_RBBM_INT_0_MASK_RBBM_PFP_MS_TIMEOUT | \
1266 	A5XX_RBBM_INT_0_MASK_RBBM_ETS_MS_TIMEOUT | \
1267 	A5XX_RBBM_INT_0_MASK_RBBM_ATB_ASYNC_OVERFLOW)
1268 
1269 static irqreturn_t a5xx_irq(struct msm_gpu *gpu)
1270 {
1271 	struct msm_drm_private *priv = gpu->dev->dev_private;
1272 	u32 status = gpu_read(gpu, REG_A5XX_RBBM_INT_0_STATUS);
1273 
1274 	/*
1275 	 * Clear all the interrupts except RBBM_AHB_ERROR - if we clear it
1276 	 * before the source is cleared the interrupt will storm.
1277 	 */
1278 	gpu_write(gpu, REG_A5XX_RBBM_INT_CLEAR_CMD,
1279 		status & ~A5XX_RBBM_INT_0_MASK_RBBM_AHB_ERROR);
1280 
1281 	if (priv->disable_err_irq) {
1282 		status &= A5XX_RBBM_INT_0_MASK_CP_CACHE_FLUSH_TS |
1283 			  A5XX_RBBM_INT_0_MASK_CP_SW;
1284 	}
1285 
1286 	/* Pass status to a5xx_rbbm_err_irq because we've already cleared it */
1287 	if (status & RBBM_ERROR_MASK)
1288 		a5xx_rbbm_err_irq(gpu, status);
1289 
1290 	if (status & A5XX_RBBM_INT_0_MASK_CP_HW_ERROR)
1291 		a5xx_cp_err_irq(gpu);
1292 
1293 	if (status & A5XX_RBBM_INT_0_MASK_MISC_HANG_DETECT)
1294 		a5xx_fault_detect_irq(gpu);
1295 
1296 	if (status & A5XX_RBBM_INT_0_MASK_UCHE_OOB_ACCESS)
1297 		a5xx_uche_err_irq(gpu);
1298 
1299 	if (status & A5XX_RBBM_INT_0_MASK_GPMU_VOLTAGE_DROOP)
1300 		a5xx_gpmu_err_irq(gpu);
1301 
1302 	if (status & A5XX_RBBM_INT_0_MASK_CP_CACHE_FLUSH_TS) {
1303 		a5xx_preempt_trigger(gpu);
1304 		msm_gpu_retire(gpu);
1305 	}
1306 
1307 	if (status & A5XX_RBBM_INT_0_MASK_CP_SW)
1308 		a5xx_preempt_irq(gpu);
1309 
1310 	return IRQ_HANDLED;
1311 }
1312 
1313 static const u32 a5xx_registers[] = {
1314 	0x0000, 0x0002, 0x0004, 0x0020, 0x0022, 0x0026, 0x0029, 0x002B,
1315 	0x002E, 0x0035, 0x0038, 0x0042, 0x0044, 0x0044, 0x0047, 0x0095,
1316 	0x0097, 0x00BB, 0x03A0, 0x0464, 0x0469, 0x046F, 0x04D2, 0x04D3,
1317 	0x04E0, 0x0533, 0x0540, 0x0555, 0x0800, 0x081A, 0x081F, 0x0841,
1318 	0x0860, 0x0860, 0x0880, 0x08A0, 0x0B00, 0x0B12, 0x0B15, 0x0B28,
1319 	0x0B78, 0x0B7F, 0x0BB0, 0x0BBD, 0x0BC0, 0x0BC6, 0x0BD0, 0x0C53,
1320 	0x0C60, 0x0C61, 0x0C80, 0x0C82, 0x0C84, 0x0C85, 0x0C90, 0x0C98,
1321 	0x0CA0, 0x0CA0, 0x0CB0, 0x0CB2, 0x2180, 0x2185, 0x2580, 0x2585,
1322 	0x0CC1, 0x0CC1, 0x0CC4, 0x0CC7, 0x0CCC, 0x0CCC, 0x0CD0, 0x0CD8,
1323 	0x0CE0, 0x0CE5, 0x0CE8, 0x0CE8, 0x0CEC, 0x0CF1, 0x0CFB, 0x0D0E,
1324 	0x2100, 0x211E, 0x2140, 0x2145, 0x2500, 0x251E, 0x2540, 0x2545,
1325 	0x0D10, 0x0D17, 0x0D20, 0x0D23, 0x0D30, 0x0D30, 0x20C0, 0x20C0,
1326 	0x24C0, 0x24C0, 0x0E40, 0x0E43, 0x0E4A, 0x0E4A, 0x0E50, 0x0E57,
1327 	0x0E60, 0x0E7C, 0x0E80, 0x0E8E, 0x0E90, 0x0E96, 0x0EA0, 0x0EA8,
1328 	0x0EB0, 0x0EB2, 0xE140, 0xE147, 0xE150, 0xE187, 0xE1A0, 0xE1A9,
1329 	0xE1B0, 0xE1B6, 0xE1C0, 0xE1C7, 0xE1D0, 0xE1D1, 0xE200, 0xE201,
1330 	0xE210, 0xE21C, 0xE240, 0xE268, 0xE000, 0xE006, 0xE010, 0xE09A,
1331 	0xE0A0, 0xE0A4, 0xE0AA, 0xE0EB, 0xE100, 0xE105, 0xE380, 0xE38F,
1332 	0xE3B0, 0xE3B0, 0xE400, 0xE405, 0xE408, 0xE4E9, 0xE4F0, 0xE4F0,
1333 	0xE280, 0xE280, 0xE282, 0xE2A3, 0xE2A5, 0xE2C2, 0xE940, 0xE947,
1334 	0xE950, 0xE987, 0xE9A0, 0xE9A9, 0xE9B0, 0xE9B6, 0xE9C0, 0xE9C7,
1335 	0xE9D0, 0xE9D1, 0xEA00, 0xEA01, 0xEA10, 0xEA1C, 0xEA40, 0xEA68,
1336 	0xE800, 0xE806, 0xE810, 0xE89A, 0xE8A0, 0xE8A4, 0xE8AA, 0xE8EB,
1337 	0xE900, 0xE905, 0xEB80, 0xEB8F, 0xEBB0, 0xEBB0, 0xEC00, 0xEC05,
1338 	0xEC08, 0xECE9, 0xECF0, 0xECF0, 0xEA80, 0xEA80, 0xEA82, 0xEAA3,
1339 	0xEAA5, 0xEAC2, 0xA800, 0xA800, 0xA820, 0xA828, 0xA840, 0xA87D,
1340 	0XA880, 0xA88D, 0xA890, 0xA8A3, 0xA8D0, 0xA8D8, 0xA8E0, 0xA8F5,
1341 	0xAC60, 0xAC60, ~0,
1342 };
1343 
1344 static void a5xx_dump(struct msm_gpu *gpu)
1345 {
1346 	DRM_DEV_INFO(gpu->dev->dev, "status:   %08x\n",
1347 		gpu_read(gpu, REG_A5XX_RBBM_STATUS));
1348 	adreno_dump(gpu);
1349 }
1350 
1351 static int a5xx_pm_resume(struct msm_gpu *gpu)
1352 {
1353 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1354 	int ret;
1355 
1356 	/* Turn on the core power */
1357 	ret = msm_gpu_pm_resume(gpu);
1358 	if (ret)
1359 		return ret;
1360 
1361 	/* Adreno 505, 506, 508, 509, 510, 512 needs manual RBBM sus/res control */
1362 	if (!(adreno_is_a530(adreno_gpu) || adreno_is_a540(adreno_gpu))) {
1363 		/* Halt the sp_input_clk at HM level */
1364 		gpu_write(gpu, REG_A5XX_RBBM_CLOCK_CNTL, 0x00000055);
1365 		a5xx_set_hwcg(gpu, true);
1366 		/* Turn on sp_input_clk at HM level */
1367 		gpu_rmw(gpu, REG_A5XX_RBBM_CLOCK_CNTL, 0xff, 0);
1368 		return 0;
1369 	}
1370 
1371 	/* Turn the RBCCU domain first to limit the chances of voltage droop */
1372 	gpu_write(gpu, REG_A5XX_GPMU_RBCCU_POWER_CNTL, 0x778000);
1373 
1374 	/* Wait 3 usecs before polling */
1375 	udelay(3);
1376 
1377 	ret = spin_usecs(gpu, 20, REG_A5XX_GPMU_RBCCU_PWR_CLK_STATUS,
1378 		(1 << 20), (1 << 20));
1379 	if (ret) {
1380 		DRM_ERROR("%s: timeout waiting for RBCCU GDSC enable: %X\n",
1381 			gpu->name,
1382 			gpu_read(gpu, REG_A5XX_GPMU_RBCCU_PWR_CLK_STATUS));
1383 		return ret;
1384 	}
1385 
1386 	/* Turn on the SP domain */
1387 	gpu_write(gpu, REG_A5XX_GPMU_SP_POWER_CNTL, 0x778000);
1388 	ret = spin_usecs(gpu, 20, REG_A5XX_GPMU_SP_PWR_CLK_STATUS,
1389 		(1 << 20), (1 << 20));
1390 	if (ret)
1391 		DRM_ERROR("%s: timeout waiting for SP GDSC enable\n",
1392 			gpu->name);
1393 
1394 	return ret;
1395 }
1396 
1397 static int a5xx_pm_suspend(struct msm_gpu *gpu)
1398 {
1399 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1400 	struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu);
1401 	u32 mask = 0xf;
1402 	int i, ret;
1403 
1404 	/* A505, A506, A508, A510 have 3 XIN ports in VBIF */
1405 	if (adreno_is_a505(adreno_gpu) || adreno_is_a506(adreno_gpu) ||
1406 	    adreno_is_a508(adreno_gpu) || adreno_is_a510(adreno_gpu))
1407 		mask = 0x7;
1408 
1409 	/* Clear the VBIF pipe before shutting down */
1410 	gpu_write(gpu, REG_A5XX_VBIF_XIN_HALT_CTRL0, mask);
1411 	spin_until((gpu_read(gpu, REG_A5XX_VBIF_XIN_HALT_CTRL1) &
1412 				mask) == mask);
1413 
1414 	gpu_write(gpu, REG_A5XX_VBIF_XIN_HALT_CTRL0, 0);
1415 
1416 	/*
1417 	 * Reset the VBIF before power collapse to avoid issue with FIFO
1418 	 * entries on Adreno A510 and A530 (the others will tend to lock up)
1419 	 */
1420 	if (adreno_is_a510(adreno_gpu) || adreno_is_a530(adreno_gpu)) {
1421 		gpu_write(gpu, REG_A5XX_RBBM_BLOCK_SW_RESET_CMD, 0x003C0000);
1422 		gpu_write(gpu, REG_A5XX_RBBM_BLOCK_SW_RESET_CMD, 0x00000000);
1423 	}
1424 
1425 	ret = msm_gpu_pm_suspend(gpu);
1426 	if (ret)
1427 		return ret;
1428 
1429 	if (a5xx_gpu->has_whereami)
1430 		for (i = 0; i < gpu->nr_rings; i++)
1431 			a5xx_gpu->shadow[i] = 0;
1432 
1433 	return 0;
1434 }
1435 
1436 static int a5xx_get_timestamp(struct msm_gpu *gpu, uint64_t *value)
1437 {
1438 	*value = gpu_read64(gpu, REG_A5XX_RBBM_ALWAYSON_COUNTER_LO);
1439 
1440 	return 0;
1441 }
1442 
1443 struct a5xx_crashdumper {
1444 	void *ptr;
1445 	struct drm_gem_object *bo;
1446 	u64 iova;
1447 };
1448 
1449 struct a5xx_gpu_state {
1450 	struct msm_gpu_state base;
1451 	u32 *hlsqregs;
1452 };
1453 
1454 static int a5xx_crashdumper_init(struct msm_gpu *gpu,
1455 		struct a5xx_crashdumper *dumper)
1456 {
1457 	dumper->ptr = msm_gem_kernel_new(gpu->dev,
1458 		SZ_1M, MSM_BO_WC, gpu->aspace,
1459 		&dumper->bo, &dumper->iova);
1460 
1461 	if (!IS_ERR(dumper->ptr))
1462 		msm_gem_object_set_name(dumper->bo, "crashdump");
1463 
1464 	return PTR_ERR_OR_ZERO(dumper->ptr);
1465 }
1466 
1467 static int a5xx_crashdumper_run(struct msm_gpu *gpu,
1468 		struct a5xx_crashdumper *dumper)
1469 {
1470 	u32 val;
1471 
1472 	if (IS_ERR_OR_NULL(dumper->ptr))
1473 		return -EINVAL;
1474 
1475 	gpu_write64(gpu, REG_A5XX_CP_CRASH_SCRIPT_BASE_LO, dumper->iova);
1476 
1477 	gpu_write(gpu, REG_A5XX_CP_CRASH_DUMP_CNTL, 1);
1478 
1479 	return gpu_poll_timeout(gpu, REG_A5XX_CP_CRASH_DUMP_CNTL, val,
1480 		val & 0x04, 100, 10000);
1481 }
1482 
1483 /*
1484  * These are a list of the registers that need to be read through the HLSQ
1485  * aperture through the crashdumper.  These are not nominally accessible from
1486  * the CPU on a secure platform.
1487  */
1488 static const struct {
1489 	u32 type;
1490 	u32 regoffset;
1491 	u32 count;
1492 } a5xx_hlsq_aperture_regs[] = {
1493 	{ 0x35, 0xe00, 0x32 },   /* HSLQ non-context */
1494 	{ 0x31, 0x2080, 0x1 },   /* HLSQ 2D context 0 */
1495 	{ 0x33, 0x2480, 0x1 },   /* HLSQ 2D context 1 */
1496 	{ 0x32, 0xe780, 0x62 },  /* HLSQ 3D context 0 */
1497 	{ 0x34, 0xef80, 0x62 },  /* HLSQ 3D context 1 */
1498 	{ 0x3f, 0x0ec0, 0x40 },  /* SP non-context */
1499 	{ 0x3d, 0x2040, 0x1 },   /* SP 2D context 0 */
1500 	{ 0x3b, 0x2440, 0x1 },   /* SP 2D context 1 */
1501 	{ 0x3e, 0xe580, 0x170 }, /* SP 3D context 0 */
1502 	{ 0x3c, 0xed80, 0x170 }, /* SP 3D context 1 */
1503 	{ 0x3a, 0x0f00, 0x1c },  /* TP non-context */
1504 	{ 0x38, 0x2000, 0xa },   /* TP 2D context 0 */
1505 	{ 0x36, 0x2400, 0xa },   /* TP 2D context 1 */
1506 	{ 0x39, 0xe700, 0x80 },  /* TP 3D context 0 */
1507 	{ 0x37, 0xef00, 0x80 },  /* TP 3D context 1 */
1508 };
1509 
1510 static void a5xx_gpu_state_get_hlsq_regs(struct msm_gpu *gpu,
1511 		struct a5xx_gpu_state *a5xx_state)
1512 {
1513 	struct a5xx_crashdumper dumper = { 0 };
1514 	u32 offset, count = 0;
1515 	u64 *ptr;
1516 	int i;
1517 
1518 	if (a5xx_crashdumper_init(gpu, &dumper))
1519 		return;
1520 
1521 	/* The script will be written at offset 0 */
1522 	ptr = dumper.ptr;
1523 
1524 	/* Start writing the data at offset 256k */
1525 	offset = dumper.iova + (256 * SZ_1K);
1526 
1527 	/* Count how many additional registers to get from the HLSQ aperture */
1528 	for (i = 0; i < ARRAY_SIZE(a5xx_hlsq_aperture_regs); i++)
1529 		count += a5xx_hlsq_aperture_regs[i].count;
1530 
1531 	a5xx_state->hlsqregs = kcalloc(count, sizeof(u32), GFP_KERNEL);
1532 	if (!a5xx_state->hlsqregs)
1533 		return;
1534 
1535 	/* Build the crashdump script */
1536 	for (i = 0; i < ARRAY_SIZE(a5xx_hlsq_aperture_regs); i++) {
1537 		u32 type = a5xx_hlsq_aperture_regs[i].type;
1538 		u32 c = a5xx_hlsq_aperture_regs[i].count;
1539 
1540 		/* Write the register to select the desired bank */
1541 		*ptr++ = ((u64) type << 8);
1542 		*ptr++ = (((u64) REG_A5XX_HLSQ_DBG_READ_SEL) << 44) |
1543 			(1 << 21) | 1;
1544 
1545 		*ptr++ = offset;
1546 		*ptr++ = (((u64) REG_A5XX_HLSQ_DBG_AHB_READ_APERTURE) << 44)
1547 			| c;
1548 
1549 		offset += c * sizeof(u32);
1550 	}
1551 
1552 	/* Write two zeros to close off the script */
1553 	*ptr++ = 0;
1554 	*ptr++ = 0;
1555 
1556 	if (a5xx_crashdumper_run(gpu, &dumper)) {
1557 		kfree(a5xx_state->hlsqregs);
1558 		msm_gem_kernel_put(dumper.bo, gpu->aspace);
1559 		return;
1560 	}
1561 
1562 	/* Copy the data from the crashdumper to the state */
1563 	memcpy(a5xx_state->hlsqregs, dumper.ptr + (256 * SZ_1K),
1564 		count * sizeof(u32));
1565 
1566 	msm_gem_kernel_put(dumper.bo, gpu->aspace);
1567 }
1568 
1569 static struct msm_gpu_state *a5xx_gpu_state_get(struct msm_gpu *gpu)
1570 {
1571 	struct a5xx_gpu_state *a5xx_state = kzalloc(sizeof(*a5xx_state),
1572 			GFP_KERNEL);
1573 	bool stalled = !!(gpu_read(gpu, REG_A5XX_RBBM_STATUS3) & BIT(24));
1574 
1575 	if (!a5xx_state)
1576 		return ERR_PTR(-ENOMEM);
1577 
1578 	/* Temporarily disable hardware clock gating before reading the hw */
1579 	a5xx_set_hwcg(gpu, false);
1580 
1581 	/* First get the generic state from the adreno core */
1582 	adreno_gpu_state_get(gpu, &(a5xx_state->base));
1583 
1584 	a5xx_state->base.rbbm_status = gpu_read(gpu, REG_A5XX_RBBM_STATUS);
1585 
1586 	/*
1587 	 * Get the HLSQ regs with the help of the crashdumper, but only if
1588 	 * we are not stalled in an iommu fault (in which case the crashdumper
1589 	 * would not have access to memory)
1590 	 */
1591 	if (!stalled)
1592 		a5xx_gpu_state_get_hlsq_regs(gpu, a5xx_state);
1593 
1594 	a5xx_set_hwcg(gpu, true);
1595 
1596 	return &a5xx_state->base;
1597 }
1598 
1599 static void a5xx_gpu_state_destroy(struct kref *kref)
1600 {
1601 	struct msm_gpu_state *state = container_of(kref,
1602 		struct msm_gpu_state, ref);
1603 	struct a5xx_gpu_state *a5xx_state = container_of(state,
1604 		struct a5xx_gpu_state, base);
1605 
1606 	kfree(a5xx_state->hlsqregs);
1607 
1608 	adreno_gpu_state_destroy(state);
1609 	kfree(a5xx_state);
1610 }
1611 
1612 static int a5xx_gpu_state_put(struct msm_gpu_state *state)
1613 {
1614 	if (IS_ERR_OR_NULL(state))
1615 		return 1;
1616 
1617 	return kref_put(&state->ref, a5xx_gpu_state_destroy);
1618 }
1619 
1620 
1621 #if defined(CONFIG_DEBUG_FS) || defined(CONFIG_DEV_COREDUMP)
1622 static void a5xx_show(struct msm_gpu *gpu, struct msm_gpu_state *state,
1623 		      struct drm_printer *p)
1624 {
1625 	int i, j;
1626 	u32 pos = 0;
1627 	struct a5xx_gpu_state *a5xx_state = container_of(state,
1628 		struct a5xx_gpu_state, base);
1629 
1630 	if (IS_ERR_OR_NULL(state))
1631 		return;
1632 
1633 	adreno_show(gpu, state, p);
1634 
1635 	/* Dump the additional a5xx HLSQ registers */
1636 	if (!a5xx_state->hlsqregs)
1637 		return;
1638 
1639 	drm_printf(p, "registers-hlsq:\n");
1640 
1641 	for (i = 0; i < ARRAY_SIZE(a5xx_hlsq_aperture_regs); i++) {
1642 		u32 o = a5xx_hlsq_aperture_regs[i].regoffset;
1643 		u32 c = a5xx_hlsq_aperture_regs[i].count;
1644 
1645 		for (j = 0; j < c; j++, pos++, o++) {
1646 			/*
1647 			 * To keep the crashdump simple we pull the entire range
1648 			 * for each register type but not all of the registers
1649 			 * in the range are valid. Fortunately invalid registers
1650 			 * stick out like a sore thumb with a value of
1651 			 * 0xdeadbeef
1652 			 */
1653 			if (a5xx_state->hlsqregs[pos] == 0xdeadbeef)
1654 				continue;
1655 
1656 			drm_printf(p, "  - { offset: 0x%04x, value: 0x%08x }\n",
1657 				o << 2, a5xx_state->hlsqregs[pos]);
1658 		}
1659 	}
1660 }
1661 #endif
1662 
1663 static struct msm_ringbuffer *a5xx_active_ring(struct msm_gpu *gpu)
1664 {
1665 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1666 	struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu);
1667 
1668 	return a5xx_gpu->cur_ring;
1669 }
1670 
1671 static u64 a5xx_gpu_busy(struct msm_gpu *gpu, unsigned long *out_sample_rate)
1672 {
1673 	u64 busy_cycles;
1674 
1675 	busy_cycles = gpu_read64(gpu, REG_A5XX_RBBM_PERFCTR_RBBM_0_LO);
1676 	*out_sample_rate = clk_get_rate(gpu->core_clk);
1677 
1678 	return busy_cycles;
1679 }
1680 
1681 static uint32_t a5xx_get_rptr(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
1682 {
1683 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1684 	struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu);
1685 
1686 	if (a5xx_gpu->has_whereami)
1687 		return a5xx_gpu->shadow[ring->id];
1688 
1689 	return ring->memptrs->rptr = gpu_read(gpu, REG_A5XX_CP_RB_RPTR);
1690 }
1691 
1692 static const struct adreno_gpu_funcs funcs = {
1693 	.base = {
1694 		.get_param = adreno_get_param,
1695 		.set_param = adreno_set_param,
1696 		.hw_init = a5xx_hw_init,
1697 		.ucode_load = a5xx_ucode_load,
1698 		.pm_suspend = a5xx_pm_suspend,
1699 		.pm_resume = a5xx_pm_resume,
1700 		.recover = a5xx_recover,
1701 		.submit = a5xx_submit,
1702 		.active_ring = a5xx_active_ring,
1703 		.irq = a5xx_irq,
1704 		.destroy = a5xx_destroy,
1705 #if defined(CONFIG_DEBUG_FS) || defined(CONFIG_DEV_COREDUMP)
1706 		.show = a5xx_show,
1707 #endif
1708 #if defined(CONFIG_DEBUG_FS)
1709 		.debugfs_init = a5xx_debugfs_init,
1710 #endif
1711 		.gpu_busy = a5xx_gpu_busy,
1712 		.gpu_state_get = a5xx_gpu_state_get,
1713 		.gpu_state_put = a5xx_gpu_state_put,
1714 		.create_address_space = adreno_create_address_space,
1715 		.get_rptr = a5xx_get_rptr,
1716 	},
1717 	.get_timestamp = a5xx_get_timestamp,
1718 };
1719 
1720 static void check_speed_bin(struct device *dev)
1721 {
1722 	struct nvmem_cell *cell;
1723 	u32 val;
1724 
1725 	/*
1726 	 * If the OPP table specifies a opp-supported-hw property then we have
1727 	 * to set something with dev_pm_opp_set_supported_hw() or the table
1728 	 * doesn't get populated so pick an arbitrary value that should
1729 	 * ensure the default frequencies are selected but not conflict with any
1730 	 * actual bins
1731 	 */
1732 	val = 0x80;
1733 
1734 	cell = nvmem_cell_get(dev, "speed_bin");
1735 
1736 	if (!IS_ERR(cell)) {
1737 		void *buf = nvmem_cell_read(cell, NULL);
1738 
1739 		if (!IS_ERR(buf)) {
1740 			u8 bin = *((u8 *) buf);
1741 
1742 			val = (1 << bin);
1743 			kfree(buf);
1744 		}
1745 
1746 		nvmem_cell_put(cell);
1747 	}
1748 
1749 	devm_pm_opp_set_supported_hw(dev, &val, 1);
1750 }
1751 
1752 struct msm_gpu *a5xx_gpu_init(struct drm_device *dev)
1753 {
1754 	struct msm_drm_private *priv = dev->dev_private;
1755 	struct platform_device *pdev = priv->gpu_pdev;
1756 	struct adreno_platform_config *config = pdev->dev.platform_data;
1757 	struct a5xx_gpu *a5xx_gpu = NULL;
1758 	struct adreno_gpu *adreno_gpu;
1759 	struct msm_gpu *gpu;
1760 	unsigned int nr_rings;
1761 	int ret;
1762 
1763 	if (!pdev) {
1764 		DRM_DEV_ERROR(dev->dev, "No A5XX device is defined\n");
1765 		return ERR_PTR(-ENXIO);
1766 	}
1767 
1768 	a5xx_gpu = kzalloc(sizeof(*a5xx_gpu), GFP_KERNEL);
1769 	if (!a5xx_gpu)
1770 		return ERR_PTR(-ENOMEM);
1771 
1772 	adreno_gpu = &a5xx_gpu->base;
1773 	gpu = &adreno_gpu->base;
1774 
1775 	adreno_gpu->registers = a5xx_registers;
1776 
1777 	a5xx_gpu->lm_leakage = 0x4E001A;
1778 
1779 	check_speed_bin(&pdev->dev);
1780 
1781 	nr_rings = 4;
1782 
1783 	if (config->info->revn == 510)
1784 		nr_rings = 1;
1785 
1786 	ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs, nr_rings);
1787 	if (ret) {
1788 		a5xx_destroy(&(a5xx_gpu->base.base));
1789 		return ERR_PTR(ret);
1790 	}
1791 
1792 	if (gpu->aspace)
1793 		msm_mmu_set_fault_handler(gpu->aspace->mmu, gpu, a5xx_fault_handler);
1794 
1795 	/* Set up the preemption specific bits and pieces for each ringbuffer */
1796 	a5xx_preempt_init(gpu);
1797 
1798 	/* Set the highest bank bit */
1799 	if (adreno_is_a540(adreno_gpu) || adreno_is_a530(adreno_gpu))
1800 		adreno_gpu->ubwc_config.highest_bank_bit = 15;
1801 	else
1802 		adreno_gpu->ubwc_config.highest_bank_bit = 14;
1803 
1804 	/* a5xx only supports UBWC 1.0, these are not configurable */
1805 	adreno_gpu->ubwc_config.macrotile_mode = 0;
1806 	adreno_gpu->ubwc_config.ubwc_swizzle = 0x7;
1807 
1808 	return gpu;
1809 }
1810