xref: /linux/drivers/gpu/drm/msm/adreno/a6xx_gpu.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
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/firmware/qcom/qcom_scm.h>
14 #include <linux/pm_domain.h>
15 #include <linux/soc/qcom/llcc-qcom.h>
16 
17 #define GPU_PAS_ID 13
18 
19 static inline bool _a6xx_check_idle(struct msm_gpu *gpu)
20 {
21 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
22 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
23 
24 	/* Check that the GMU is idle */
25 	if (!adreno_has_gmu_wrapper(adreno_gpu) && !a6xx_gmu_isidle(&a6xx_gpu->gmu))
26 		return false;
27 
28 	/* Check tha the CX master is idle */
29 	if (gpu_read(gpu, REG_A6XX_RBBM_STATUS) &
30 			~A6XX_RBBM_STATUS_CP_AHB_BUSY_CX_MASTER)
31 		return false;
32 
33 	return !(gpu_read(gpu, REG_A6XX_RBBM_INT_0_STATUS) &
34 		A6XX_RBBM_INT_0_MASK_RBBM_HANG_DETECT);
35 }
36 
37 static bool a6xx_idle(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
38 {
39 	/* wait for CP to drain ringbuffer: */
40 	if (!adreno_idle(gpu, ring))
41 		return false;
42 
43 	if (spin_until(_a6xx_check_idle(gpu))) {
44 		DRM_ERROR("%s: %ps: timeout waiting for GPU to idle: status %8.8X irq %8.8X rptr/wptr %d/%d\n",
45 			gpu->name, __builtin_return_address(0),
46 			gpu_read(gpu, REG_A6XX_RBBM_STATUS),
47 			gpu_read(gpu, REG_A6XX_RBBM_INT_0_STATUS),
48 			gpu_read(gpu, REG_A6XX_CP_RB_RPTR),
49 			gpu_read(gpu, REG_A6XX_CP_RB_WPTR));
50 		return false;
51 	}
52 
53 	return true;
54 }
55 
56 static void update_shadow_rptr(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
57 {
58 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
59 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
60 
61 	/* Expanded APRIV doesn't need to issue the WHERE_AM_I opcode */
62 	if (a6xx_gpu->has_whereami && !adreno_gpu->base.hw_apriv) {
63 		OUT_PKT7(ring, CP_WHERE_AM_I, 2);
64 		OUT_RING(ring, lower_32_bits(shadowptr(a6xx_gpu, ring)));
65 		OUT_RING(ring, upper_32_bits(shadowptr(a6xx_gpu, ring)));
66 	}
67 }
68 
69 static void a6xx_flush(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
70 {
71 	uint32_t wptr;
72 	unsigned long flags;
73 
74 	update_shadow_rptr(gpu, ring);
75 
76 	spin_lock_irqsave(&ring->preempt_lock, flags);
77 
78 	/* Copy the shadow to the actual register */
79 	ring->cur = ring->next;
80 
81 	/* Make sure to wrap wptr if we need to */
82 	wptr = get_wptr(ring);
83 
84 	spin_unlock_irqrestore(&ring->preempt_lock, flags);
85 
86 	/* Make sure everything is posted before making a decision */
87 	mb();
88 
89 	gpu_write(gpu, REG_A6XX_CP_RB_WPTR, wptr);
90 }
91 
92 static void get_stats_counter(struct msm_ringbuffer *ring, u32 counter,
93 		u64 iova)
94 {
95 	OUT_PKT7(ring, CP_REG_TO_MEM, 3);
96 	OUT_RING(ring, CP_REG_TO_MEM_0_REG(counter) |
97 		CP_REG_TO_MEM_0_CNT(2) |
98 		CP_REG_TO_MEM_0_64B);
99 	OUT_RING(ring, lower_32_bits(iova));
100 	OUT_RING(ring, upper_32_bits(iova));
101 }
102 
103 static void a6xx_set_pagetable(struct a6xx_gpu *a6xx_gpu,
104 		struct msm_ringbuffer *ring, struct msm_gem_submit *submit)
105 {
106 	bool sysprof = refcount_read(&a6xx_gpu->base.base.sysprof_active) > 1;
107 	struct msm_file_private *ctx = submit->queue->ctx;
108 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
109 	phys_addr_t ttbr;
110 	u32 asid;
111 	u64 memptr = rbmemptr(ring, ttbr0);
112 
113 	if (ctx->seqno == a6xx_gpu->base.base.cur_ctx_seqno)
114 		return;
115 
116 	if (msm_iommu_pagetable_params(ctx->aspace->mmu, &ttbr, &asid))
117 		return;
118 
119 	if (adreno_gpu->info->family >= ADRENO_7XX_GEN1) {
120 		/* Wait for previous submit to complete before continuing: */
121 		OUT_PKT7(ring, CP_WAIT_TIMESTAMP, 4);
122 		OUT_RING(ring, 0);
123 		OUT_RING(ring, lower_32_bits(rbmemptr(ring, fence)));
124 		OUT_RING(ring, upper_32_bits(rbmemptr(ring, fence)));
125 		OUT_RING(ring, submit->seqno - 1);
126 	}
127 
128 	if (!sysprof) {
129 		if (!adreno_is_a7xx(adreno_gpu)) {
130 			/* Turn off protected mode to write to special registers */
131 			OUT_PKT7(ring, CP_SET_PROTECTED_MODE, 1);
132 			OUT_RING(ring, 0);
133 		}
134 
135 		OUT_PKT4(ring, REG_A6XX_RBBM_PERFCTR_SRAM_INIT_CMD, 1);
136 		OUT_RING(ring, 1);
137 	}
138 
139 	/* Execute the table update */
140 	OUT_PKT7(ring, CP_SMMU_TABLE_UPDATE, 4);
141 	OUT_RING(ring, CP_SMMU_TABLE_UPDATE_0_TTBR0_LO(lower_32_bits(ttbr)));
142 
143 	OUT_RING(ring,
144 		CP_SMMU_TABLE_UPDATE_1_TTBR0_HI(upper_32_bits(ttbr)) |
145 		CP_SMMU_TABLE_UPDATE_1_ASID(asid));
146 	OUT_RING(ring, CP_SMMU_TABLE_UPDATE_2_CONTEXTIDR(0));
147 	OUT_RING(ring, CP_SMMU_TABLE_UPDATE_3_CONTEXTBANK(0));
148 
149 	/*
150 	 * Write the new TTBR0 to the memstore. This is good for debugging.
151 	 */
152 	OUT_PKT7(ring, CP_MEM_WRITE, 4);
153 	OUT_RING(ring, CP_MEM_WRITE_0_ADDR_LO(lower_32_bits(memptr)));
154 	OUT_RING(ring, CP_MEM_WRITE_1_ADDR_HI(upper_32_bits(memptr)));
155 	OUT_RING(ring, lower_32_bits(ttbr));
156 	OUT_RING(ring, (asid << 16) | upper_32_bits(ttbr));
157 
158 	/*
159 	 * Sync both threads after switching pagetables and enable BR only
160 	 * to make sure BV doesn't race ahead while BR is still switching
161 	 * pagetables.
162 	 */
163 	if (adreno_is_a7xx(&a6xx_gpu->base)) {
164 		OUT_PKT7(ring, CP_THREAD_CONTROL, 1);
165 		OUT_RING(ring, CP_THREAD_CONTROL_0_SYNC_THREADS | CP_SET_THREAD_BR);
166 	}
167 
168 	/*
169 	 * And finally, trigger a uche flush to be sure there isn't anything
170 	 * lingering in that part of the GPU
171 	 */
172 
173 	OUT_PKT7(ring, CP_EVENT_WRITE, 1);
174 	OUT_RING(ring, CACHE_INVALIDATE);
175 
176 	if (!sysprof) {
177 		/*
178 		 * Wait for SRAM clear after the pgtable update, so the
179 		 * two can happen in parallel:
180 		 */
181 		OUT_PKT7(ring, CP_WAIT_REG_MEM, 6);
182 		OUT_RING(ring, CP_WAIT_REG_MEM_0_FUNCTION(WRITE_EQ));
183 		OUT_RING(ring, CP_WAIT_REG_MEM_1_POLL_ADDR_LO(
184 				REG_A6XX_RBBM_PERFCTR_SRAM_INIT_STATUS));
185 		OUT_RING(ring, CP_WAIT_REG_MEM_2_POLL_ADDR_HI(0));
186 		OUT_RING(ring, CP_WAIT_REG_MEM_3_REF(0x1));
187 		OUT_RING(ring, CP_WAIT_REG_MEM_4_MASK(0x1));
188 		OUT_RING(ring, CP_WAIT_REG_MEM_5_DELAY_LOOP_CYCLES(0));
189 
190 		if (!adreno_is_a7xx(adreno_gpu)) {
191 			/* Re-enable protected mode: */
192 			OUT_PKT7(ring, CP_SET_PROTECTED_MODE, 1);
193 			OUT_RING(ring, 1);
194 		}
195 	}
196 }
197 
198 static void a6xx_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
199 {
200 	unsigned int index = submit->seqno % MSM_GPU_SUBMIT_STATS_COUNT;
201 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
202 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
203 	struct msm_ringbuffer *ring = submit->ring;
204 	unsigned int i, ibs = 0;
205 
206 	a6xx_set_pagetable(a6xx_gpu, ring, submit);
207 
208 	get_stats_counter(ring, REG_A6XX_RBBM_PERFCTR_CP(0),
209 		rbmemptr_stats(ring, index, cpcycles_start));
210 
211 	/*
212 	 * For PM4 the GMU register offsets are calculated from the base of the
213 	 * GPU registers so we need to add 0x1a800 to the register value on A630
214 	 * to get the right value from PM4.
215 	 */
216 	get_stats_counter(ring, REG_A6XX_CP_ALWAYS_ON_COUNTER,
217 		rbmemptr_stats(ring, index, alwayson_start));
218 
219 	/* Invalidate CCU depth and color */
220 	OUT_PKT7(ring, CP_EVENT_WRITE, 1);
221 	OUT_RING(ring, CP_EVENT_WRITE_0_EVENT(PC_CCU_INVALIDATE_DEPTH));
222 
223 	OUT_PKT7(ring, CP_EVENT_WRITE, 1);
224 	OUT_RING(ring, CP_EVENT_WRITE_0_EVENT(PC_CCU_INVALIDATE_COLOR));
225 
226 	/* Submit the commands */
227 	for (i = 0; i < submit->nr_cmds; i++) {
228 		switch (submit->cmd[i].type) {
229 		case MSM_SUBMIT_CMD_IB_TARGET_BUF:
230 			break;
231 		case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
232 			if (gpu->cur_ctx_seqno == submit->queue->ctx->seqno)
233 				break;
234 			fallthrough;
235 		case MSM_SUBMIT_CMD_BUF:
236 			OUT_PKT7(ring, CP_INDIRECT_BUFFER_PFE, 3);
237 			OUT_RING(ring, lower_32_bits(submit->cmd[i].iova));
238 			OUT_RING(ring, upper_32_bits(submit->cmd[i].iova));
239 			OUT_RING(ring, submit->cmd[i].size);
240 			ibs++;
241 			break;
242 		}
243 
244 		/*
245 		 * Periodically update shadow-wptr if needed, so that we
246 		 * can see partial progress of submits with large # of
247 		 * cmds.. otherwise we could needlessly stall waiting for
248 		 * ringbuffer state, simply due to looking at a shadow
249 		 * rptr value that has not been updated
250 		 */
251 		if ((ibs % 32) == 0)
252 			update_shadow_rptr(gpu, ring);
253 	}
254 
255 	get_stats_counter(ring, REG_A6XX_RBBM_PERFCTR_CP(0),
256 		rbmemptr_stats(ring, index, cpcycles_end));
257 	get_stats_counter(ring, REG_A6XX_CP_ALWAYS_ON_COUNTER,
258 		rbmemptr_stats(ring, index, alwayson_end));
259 
260 	/* Write the fence to the scratch register */
261 	OUT_PKT4(ring, REG_A6XX_CP_SCRATCH_REG(2), 1);
262 	OUT_RING(ring, submit->seqno);
263 
264 	/*
265 	 * Execute a CACHE_FLUSH_TS event. This will ensure that the
266 	 * timestamp is written to the memory and then triggers the interrupt
267 	 */
268 	OUT_PKT7(ring, CP_EVENT_WRITE, 4);
269 	OUT_RING(ring, CP_EVENT_WRITE_0_EVENT(CACHE_FLUSH_TS) |
270 		CP_EVENT_WRITE_0_IRQ);
271 	OUT_RING(ring, lower_32_bits(rbmemptr(ring, fence)));
272 	OUT_RING(ring, upper_32_bits(rbmemptr(ring, fence)));
273 	OUT_RING(ring, submit->seqno);
274 
275 	trace_msm_gpu_submit_flush(submit,
276 		gpu_read64(gpu, REG_A6XX_CP_ALWAYS_ON_COUNTER));
277 
278 	a6xx_flush(gpu, ring);
279 }
280 
281 static void a7xx_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
282 {
283 	unsigned int index = submit->seqno % MSM_GPU_SUBMIT_STATS_COUNT;
284 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
285 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
286 	struct msm_ringbuffer *ring = submit->ring;
287 	unsigned int i, ibs = 0;
288 
289 	/*
290 	 * Toggle concurrent binning for pagetable switch and set the thread to
291 	 * BR since only it can execute the pagetable switch packets.
292 	 */
293 	OUT_PKT7(ring, CP_THREAD_CONTROL, 1);
294 	OUT_RING(ring, CP_THREAD_CONTROL_0_SYNC_THREADS | CP_SET_THREAD_BR);
295 
296 	a6xx_set_pagetable(a6xx_gpu, ring, submit);
297 
298 	get_stats_counter(ring, REG_A7XX_RBBM_PERFCTR_CP(0),
299 		rbmemptr_stats(ring, index, cpcycles_start));
300 	get_stats_counter(ring, REG_A6XX_CP_ALWAYS_ON_COUNTER,
301 		rbmemptr_stats(ring, index, alwayson_start));
302 
303 	OUT_PKT7(ring, CP_THREAD_CONTROL, 1);
304 	OUT_RING(ring, CP_SET_THREAD_BOTH);
305 
306 	OUT_PKT7(ring, CP_SET_MARKER, 1);
307 	OUT_RING(ring, 0x101); /* IFPC disable */
308 
309 	OUT_PKT7(ring, CP_SET_MARKER, 1);
310 	OUT_RING(ring, 0x00d); /* IB1LIST start */
311 
312 	/* Submit the commands */
313 	for (i = 0; i < submit->nr_cmds; i++) {
314 		switch (submit->cmd[i].type) {
315 		case MSM_SUBMIT_CMD_IB_TARGET_BUF:
316 			break;
317 		case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
318 			if (gpu->cur_ctx_seqno == submit->queue->ctx->seqno)
319 				break;
320 			fallthrough;
321 		case MSM_SUBMIT_CMD_BUF:
322 			OUT_PKT7(ring, CP_INDIRECT_BUFFER_PFE, 3);
323 			OUT_RING(ring, lower_32_bits(submit->cmd[i].iova));
324 			OUT_RING(ring, upper_32_bits(submit->cmd[i].iova));
325 			OUT_RING(ring, submit->cmd[i].size);
326 			ibs++;
327 			break;
328 		}
329 
330 		/*
331 		 * Periodically update shadow-wptr if needed, so that we
332 		 * can see partial progress of submits with large # of
333 		 * cmds.. otherwise we could needlessly stall waiting for
334 		 * ringbuffer state, simply due to looking at a shadow
335 		 * rptr value that has not been updated
336 		 */
337 		if ((ibs % 32) == 0)
338 			update_shadow_rptr(gpu, ring);
339 	}
340 
341 	OUT_PKT7(ring, CP_SET_MARKER, 1);
342 	OUT_RING(ring, 0x00e); /* IB1LIST end */
343 
344 	get_stats_counter(ring, REG_A7XX_RBBM_PERFCTR_CP(0),
345 		rbmemptr_stats(ring, index, cpcycles_end));
346 	get_stats_counter(ring, REG_A6XX_CP_ALWAYS_ON_COUNTER,
347 		rbmemptr_stats(ring, index, alwayson_end));
348 
349 	/* Write the fence to the scratch register */
350 	OUT_PKT4(ring, REG_A6XX_CP_SCRATCH_REG(2), 1);
351 	OUT_RING(ring, submit->seqno);
352 
353 	OUT_PKT7(ring, CP_THREAD_CONTROL, 1);
354 	OUT_RING(ring, CP_SET_THREAD_BR);
355 
356 	OUT_PKT7(ring, CP_EVENT_WRITE, 1);
357 	OUT_RING(ring, CCU_INVALIDATE_DEPTH);
358 
359 	OUT_PKT7(ring, CP_EVENT_WRITE, 1);
360 	OUT_RING(ring, CCU_INVALIDATE_COLOR);
361 
362 	OUT_PKT7(ring, CP_THREAD_CONTROL, 1);
363 	OUT_RING(ring, CP_SET_THREAD_BV);
364 
365 	/*
366 	 * Make sure the timestamp is committed once BV pipe is
367 	 * completely done with this submission.
368 	 */
369 	OUT_PKT7(ring, CP_EVENT_WRITE, 4);
370 	OUT_RING(ring, CACHE_CLEAN | BIT(27));
371 	OUT_RING(ring, lower_32_bits(rbmemptr(ring, bv_fence)));
372 	OUT_RING(ring, upper_32_bits(rbmemptr(ring, bv_fence)));
373 	OUT_RING(ring, submit->seqno);
374 
375 	OUT_PKT7(ring, CP_THREAD_CONTROL, 1);
376 	OUT_RING(ring, CP_SET_THREAD_BR);
377 
378 	/*
379 	 * This makes sure that BR doesn't race ahead and commit
380 	 * timestamp to memstore while BV is still processing
381 	 * this submission.
382 	 */
383 	OUT_PKT7(ring, CP_WAIT_TIMESTAMP, 4);
384 	OUT_RING(ring, 0);
385 	OUT_RING(ring, lower_32_bits(rbmemptr(ring, bv_fence)));
386 	OUT_RING(ring, upper_32_bits(rbmemptr(ring, bv_fence)));
387 	OUT_RING(ring, submit->seqno);
388 
389 	/* write the ringbuffer timestamp */
390 	OUT_PKT7(ring, CP_EVENT_WRITE, 4);
391 	OUT_RING(ring, CACHE_CLEAN | CP_EVENT_WRITE_0_IRQ | BIT(27));
392 	OUT_RING(ring, lower_32_bits(rbmemptr(ring, fence)));
393 	OUT_RING(ring, upper_32_bits(rbmemptr(ring, fence)));
394 	OUT_RING(ring, submit->seqno);
395 
396 	OUT_PKT7(ring, CP_THREAD_CONTROL, 1);
397 	OUT_RING(ring, CP_SET_THREAD_BOTH);
398 
399 	OUT_PKT7(ring, CP_SET_MARKER, 1);
400 	OUT_RING(ring, 0x100); /* IFPC enable */
401 
402 	trace_msm_gpu_submit_flush(submit,
403 		gpu_read64(gpu, REG_A6XX_CP_ALWAYS_ON_COUNTER));
404 
405 	a6xx_flush(gpu, ring);
406 }
407 
408 static void a6xx_set_hwcg(struct msm_gpu *gpu, bool state)
409 {
410 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
411 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
412 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
413 	const struct adreno_reglist *reg;
414 	unsigned int i;
415 	u32 cgc_delay, cgc_hyst;
416 	u32 val, clock_cntl_on;
417 
418 	if (!(adreno_gpu->info->a6xx->hwcg || adreno_is_a7xx(adreno_gpu)))
419 		return;
420 
421 	if (adreno_is_a630(adreno_gpu))
422 		clock_cntl_on = 0x8aa8aa02;
423 	else if (adreno_is_a610(adreno_gpu))
424 		clock_cntl_on = 0xaaa8aa82;
425 	else if (adreno_is_a702(adreno_gpu))
426 		clock_cntl_on = 0xaaaaaa82;
427 	else
428 		clock_cntl_on = 0x8aa8aa82;
429 
430 	cgc_delay = adreno_is_a615_family(adreno_gpu) ? 0x111 : 0x10111;
431 	cgc_hyst = adreno_is_a615_family(adreno_gpu) ? 0x555 : 0x5555;
432 
433 	gmu_write(&a6xx_gpu->gmu, REG_A6XX_GPU_GMU_AO_GMU_CGC_MODE_CNTL,
434 			state ? adreno_gpu->info->a6xx->gmu_cgc_mode : 0);
435 	gmu_write(&a6xx_gpu->gmu, REG_A6XX_GPU_GMU_AO_GMU_CGC_DELAY_CNTL,
436 			state ? cgc_delay : 0);
437 	gmu_write(&a6xx_gpu->gmu, REG_A6XX_GPU_GMU_AO_GMU_CGC_HYST_CNTL,
438 			state ? cgc_hyst : 0);
439 
440 	if (!adreno_gpu->info->a6xx->hwcg) {
441 		gpu_write(gpu, REG_A7XX_RBBM_CLOCK_CNTL_GLOBAL, 1);
442 		gpu_write(gpu, REG_A7XX_RBBM_CGC_GLOBAL_LOAD_CMD, state ? 1 : 0);
443 
444 		if (state) {
445 			gpu_write(gpu, REG_A7XX_RBBM_CGC_P2S_TRIG_CMD, 1);
446 
447 			if (gpu_poll_timeout(gpu, REG_A7XX_RBBM_CGC_P2S_STATUS, val,
448 					     val & A7XX_RBBM_CGC_P2S_STATUS_TXDONE, 1, 10)) {
449 				dev_err(&gpu->pdev->dev, "RBBM_CGC_P2S_STATUS TXDONE Poll failed\n");
450 				return;
451 			}
452 
453 			gpu_write(gpu, REG_A7XX_RBBM_CLOCK_CNTL_GLOBAL, 0);
454 		}
455 
456 		return;
457 	}
458 
459 	val = gpu_read(gpu, REG_A6XX_RBBM_CLOCK_CNTL);
460 
461 	/* Don't re-program the registers if they are already correct */
462 	if ((!state && !val) || (state && (val == clock_cntl_on)))
463 		return;
464 
465 	/* Disable SP clock before programming HWCG registers */
466 	if (!adreno_is_a610_family(adreno_gpu) && !adreno_is_a7xx(adreno_gpu))
467 		gmu_rmw(gmu, REG_A6XX_GPU_GMU_GX_SPTPRAC_CLOCK_CONTROL, 1, 0);
468 
469 	for (i = 0; (reg = &adreno_gpu->info->a6xx->hwcg[i], reg->offset); i++)
470 		gpu_write(gpu, reg->offset, state ? reg->value : 0);
471 
472 	/* Enable SP clock */
473 	if (!adreno_is_a610_family(adreno_gpu) && !adreno_is_a7xx(adreno_gpu))
474 		gmu_rmw(gmu, REG_A6XX_GPU_GMU_GX_SPTPRAC_CLOCK_CONTROL, 0, 1);
475 
476 	gpu_write(gpu, REG_A6XX_RBBM_CLOCK_CNTL, state ? clock_cntl_on : 0);
477 }
478 
479 static void a6xx_set_cp_protect(struct msm_gpu *gpu)
480 {
481 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
482 	const struct adreno_protect *protect = adreno_gpu->info->a6xx->protect;
483 	unsigned i;
484 
485 	/*
486 	 * Enable access protection to privileged registers, fault on an access
487 	 * protect violation and select the last span to protect from the start
488 	 * address all the way to the end of the register address space
489 	 */
490 	gpu_write(gpu, REG_A6XX_CP_PROTECT_CNTL,
491 		  A6XX_CP_PROTECT_CNTL_ACCESS_PROT_EN |
492 		  A6XX_CP_PROTECT_CNTL_ACCESS_FAULT_ON_VIOL_EN |
493 		  A6XX_CP_PROTECT_CNTL_LAST_SPAN_INF_RANGE);
494 
495 	for (i = 0; i < protect->count - 1; i++) {
496 		/* Intentionally skip writing to some registers */
497 		if (protect->regs[i])
498 			gpu_write(gpu, REG_A6XX_CP_PROTECT(i), protect->regs[i]);
499 	}
500 	/* last CP_PROTECT to have "infinite" length on the last entry */
501 	gpu_write(gpu, REG_A6XX_CP_PROTECT(protect->count_max - 1), protect->regs[i]);
502 }
503 
504 static void a6xx_calc_ubwc_config(struct adreno_gpu *gpu)
505 {
506 	gpu->ubwc_config.rgb565_predicator = 0;
507 	gpu->ubwc_config.uavflagprd_inv = 0;
508 	gpu->ubwc_config.min_acc_len = 0;
509 	gpu->ubwc_config.ubwc_swizzle = 0x6;
510 	gpu->ubwc_config.macrotile_mode = 0;
511 	gpu->ubwc_config.highest_bank_bit = 15;
512 
513 	if (adreno_is_a610(gpu)) {
514 		gpu->ubwc_config.highest_bank_bit = 13;
515 		gpu->ubwc_config.min_acc_len = 1;
516 		gpu->ubwc_config.ubwc_swizzle = 0x7;
517 	}
518 
519 	if (adreno_is_a618(gpu))
520 		gpu->ubwc_config.highest_bank_bit = 14;
521 
522 	if (adreno_is_a619(gpu))
523 		/* TODO: Should be 14 but causes corruption at e.g. 1920x1200 on DP */
524 		gpu->ubwc_config.highest_bank_bit = 13;
525 
526 	if (adreno_is_a619_holi(gpu))
527 		gpu->ubwc_config.highest_bank_bit = 13;
528 
529 	if (adreno_is_a621(gpu)) {
530 		gpu->ubwc_config.highest_bank_bit = 13;
531 		gpu->ubwc_config.amsbc = 1;
532 		gpu->ubwc_config.uavflagprd_inv = 2;
533 	}
534 
535 	if (adreno_is_a640_family(gpu))
536 		gpu->ubwc_config.amsbc = 1;
537 
538 	if (adreno_is_a680(gpu))
539 		gpu->ubwc_config.macrotile_mode = 1;
540 
541 	if (adreno_is_a650(gpu) ||
542 	    adreno_is_a660(gpu) ||
543 	    adreno_is_a690(gpu) ||
544 	    adreno_is_a730(gpu) ||
545 	    adreno_is_a740_family(gpu)) {
546 		/* TODO: get ddr type from bootloader and use 2 for LPDDR4 */
547 		gpu->ubwc_config.highest_bank_bit = 16;
548 		gpu->ubwc_config.amsbc = 1;
549 		gpu->ubwc_config.rgb565_predicator = 1;
550 		gpu->ubwc_config.uavflagprd_inv = 2;
551 		gpu->ubwc_config.macrotile_mode = 1;
552 	}
553 
554 	if (adreno_is_7c3(gpu)) {
555 		gpu->ubwc_config.highest_bank_bit = 14;
556 		gpu->ubwc_config.amsbc = 1;
557 		gpu->ubwc_config.rgb565_predicator = 1;
558 		gpu->ubwc_config.uavflagprd_inv = 2;
559 		gpu->ubwc_config.macrotile_mode = 1;
560 	}
561 
562 	if (adreno_is_a702(gpu)) {
563 		gpu->ubwc_config.highest_bank_bit = 14;
564 		gpu->ubwc_config.min_acc_len = 1;
565 	}
566 }
567 
568 static void a6xx_set_ubwc_config(struct msm_gpu *gpu)
569 {
570 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
571 	/*
572 	 * We subtract 13 from the highest bank bit (13 is the minimum value
573 	 * allowed by hw) and write the lowest two bits of the remaining value
574 	 * as hbb_lo and the one above it as hbb_hi to the hardware.
575 	 */
576 	BUG_ON(adreno_gpu->ubwc_config.highest_bank_bit < 13);
577 	u32 hbb = adreno_gpu->ubwc_config.highest_bank_bit - 13;
578 	u32 hbb_hi = hbb >> 2;
579 	u32 hbb_lo = hbb & 3;
580 	u32 ubwc_mode = adreno_gpu->ubwc_config.ubwc_swizzle & 1;
581 	u32 level2_swizzling_dis = !(adreno_gpu->ubwc_config.ubwc_swizzle & 2);
582 
583 	gpu_write(gpu, REG_A6XX_RB_NC_MODE_CNTL,
584 		  level2_swizzling_dis << 12 |
585 		  adreno_gpu->ubwc_config.rgb565_predicator << 11 |
586 		  hbb_hi << 10 | adreno_gpu->ubwc_config.amsbc << 4 |
587 		  adreno_gpu->ubwc_config.min_acc_len << 3 |
588 		  hbb_lo << 1 | ubwc_mode);
589 
590 	gpu_write(gpu, REG_A6XX_TPL1_NC_MODE_CNTL,
591 		  level2_swizzling_dis << 6 | hbb_hi << 4 |
592 		  adreno_gpu->ubwc_config.min_acc_len << 3 |
593 		  hbb_lo << 1 | ubwc_mode);
594 
595 	gpu_write(gpu, REG_A6XX_SP_NC_MODE_CNTL,
596 		  level2_swizzling_dis << 12 | hbb_hi << 10 |
597 		  adreno_gpu->ubwc_config.uavflagprd_inv << 4 |
598 		  adreno_gpu->ubwc_config.min_acc_len << 3 |
599 		  hbb_lo << 1 | ubwc_mode);
600 
601 	if (adreno_is_a7xx(adreno_gpu))
602 		gpu_write(gpu, REG_A7XX_GRAS_NC_MODE_CNTL,
603 			  FIELD_PREP(GENMASK(8, 5), hbb_lo));
604 
605 	gpu_write(gpu, REG_A6XX_UCHE_MODE_CNTL,
606 		  adreno_gpu->ubwc_config.min_acc_len << 23 | hbb_lo << 21);
607 
608 	gpu_write(gpu, REG_A6XX_RBBM_NC_MODE_CNTL,
609 		  adreno_gpu->ubwc_config.macrotile_mode);
610 }
611 
612 static int a6xx_cp_init(struct msm_gpu *gpu)
613 {
614 	struct msm_ringbuffer *ring = gpu->rb[0];
615 
616 	OUT_PKT7(ring, CP_ME_INIT, 8);
617 
618 	OUT_RING(ring, 0x0000002f);
619 
620 	/* Enable multiple hardware contexts */
621 	OUT_RING(ring, 0x00000003);
622 
623 	/* Enable error detection */
624 	OUT_RING(ring, 0x20000000);
625 
626 	/* Don't enable header dump */
627 	OUT_RING(ring, 0x00000000);
628 	OUT_RING(ring, 0x00000000);
629 
630 	/* No workarounds enabled */
631 	OUT_RING(ring, 0x00000000);
632 
633 	/* Pad rest of the cmds with 0's */
634 	OUT_RING(ring, 0x00000000);
635 	OUT_RING(ring, 0x00000000);
636 
637 	a6xx_flush(gpu, ring);
638 	return a6xx_idle(gpu, ring) ? 0 : -EINVAL;
639 }
640 
641 static int a7xx_cp_init(struct msm_gpu *gpu)
642 {
643 	struct msm_ringbuffer *ring = gpu->rb[0];
644 	u32 mask;
645 
646 	/* Disable concurrent binning before sending CP init */
647 	OUT_PKT7(ring, CP_THREAD_CONTROL, 1);
648 	OUT_RING(ring, BIT(27));
649 
650 	OUT_PKT7(ring, CP_ME_INIT, 7);
651 
652 	/* Use multiple HW contexts */
653 	mask = BIT(0);
654 
655 	/* Enable error detection */
656 	mask |= BIT(1);
657 
658 	/* Set default reset state */
659 	mask |= BIT(3);
660 
661 	/* Disable save/restore of performance counters across preemption */
662 	mask |= BIT(6);
663 
664 	/* Enable the register init list with the spinlock */
665 	mask |= BIT(8);
666 
667 	OUT_RING(ring, mask);
668 
669 	/* Enable multiple hardware contexts */
670 	OUT_RING(ring, 0x00000003);
671 
672 	/* Enable error detection */
673 	OUT_RING(ring, 0x20000000);
674 
675 	/* Operation mode mask */
676 	OUT_RING(ring, 0x00000002);
677 
678 	/* *Don't* send a power up reg list for concurrent binning (TODO) */
679 	/* Lo address */
680 	OUT_RING(ring, 0x00000000);
681 	/* Hi address */
682 	OUT_RING(ring, 0x00000000);
683 	/* BIT(31) set => read the regs from the list */
684 	OUT_RING(ring, 0x00000000);
685 
686 	a6xx_flush(gpu, ring);
687 	return a6xx_idle(gpu, ring) ? 0 : -EINVAL;
688 }
689 
690 /*
691  * Check that the microcode version is new enough to include several key
692  * security fixes. Return true if the ucode is safe.
693  */
694 static bool a6xx_ucode_check_version(struct a6xx_gpu *a6xx_gpu,
695 		struct drm_gem_object *obj)
696 {
697 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
698 	struct msm_gpu *gpu = &adreno_gpu->base;
699 	const char *sqe_name = adreno_gpu->info->fw[ADRENO_FW_SQE];
700 	u32 *buf = msm_gem_get_vaddr(obj);
701 	bool ret = false;
702 
703 	if (IS_ERR(buf))
704 		return false;
705 
706 	/* A7xx is safe! */
707 	if (adreno_is_a7xx(adreno_gpu) || adreno_is_a702(adreno_gpu))
708 		return true;
709 
710 	/*
711 	 * Targets up to a640 (a618, a630 and a640) need to check for a
712 	 * microcode version that is patched to support the whereami opcode or
713 	 * one that is new enough to include it by default.
714 	 *
715 	 * a650 tier targets don't need whereami but still need to be
716 	 * equal to or newer than 0.95 for other security fixes
717 	 *
718 	 * a660 targets have all the critical security fixes from the start
719 	 */
720 	if (!strcmp(sqe_name, "a630_sqe.fw")) {
721 		/*
722 		 * If the lowest nibble is 0xa that is an indication that this
723 		 * microcode has been patched. The actual version is in dword
724 		 * [3] but we only care about the patchlevel which is the lowest
725 		 * nibble of dword [3]
726 		 *
727 		 * Otherwise check that the firmware is greater than or equal
728 		 * to 1.90 which was the first version that had this fix built
729 		 * in
730 		 */
731 		if ((((buf[0] & 0xf) == 0xa) && (buf[2] & 0xf) >= 1) ||
732 			(buf[0] & 0xfff) >= 0x190) {
733 			a6xx_gpu->has_whereami = true;
734 			ret = true;
735 			goto out;
736 		}
737 
738 		DRM_DEV_ERROR(&gpu->pdev->dev,
739 			"a630 SQE ucode is too old. Have version %x need at least %x\n",
740 			buf[0] & 0xfff, 0x190);
741 	} else if (!strcmp(sqe_name, "a650_sqe.fw")) {
742 		if ((buf[0] & 0xfff) >= 0x095) {
743 			ret = true;
744 			goto out;
745 		}
746 
747 		DRM_DEV_ERROR(&gpu->pdev->dev,
748 			"a650 SQE ucode is too old. Have version %x need at least %x\n",
749 			buf[0] & 0xfff, 0x095);
750 	} else if (!strcmp(sqe_name, "a660_sqe.fw")) {
751 		ret = true;
752 	} else {
753 		DRM_DEV_ERROR(&gpu->pdev->dev,
754 			"unknown GPU, add it to a6xx_ucode_check_version()!!\n");
755 	}
756 out:
757 	msm_gem_put_vaddr(obj);
758 	return ret;
759 }
760 
761 static int a6xx_ucode_load(struct msm_gpu *gpu)
762 {
763 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
764 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
765 
766 	if (!a6xx_gpu->sqe_bo) {
767 		a6xx_gpu->sqe_bo = adreno_fw_create_bo(gpu,
768 			adreno_gpu->fw[ADRENO_FW_SQE], &a6xx_gpu->sqe_iova);
769 
770 		if (IS_ERR(a6xx_gpu->sqe_bo)) {
771 			int ret = PTR_ERR(a6xx_gpu->sqe_bo);
772 
773 			a6xx_gpu->sqe_bo = NULL;
774 			DRM_DEV_ERROR(&gpu->pdev->dev,
775 				"Could not allocate SQE ucode: %d\n", ret);
776 
777 			return ret;
778 		}
779 
780 		msm_gem_object_set_name(a6xx_gpu->sqe_bo, "sqefw");
781 		if (!a6xx_ucode_check_version(a6xx_gpu, a6xx_gpu->sqe_bo)) {
782 			msm_gem_unpin_iova(a6xx_gpu->sqe_bo, gpu->aspace);
783 			drm_gem_object_put(a6xx_gpu->sqe_bo);
784 
785 			a6xx_gpu->sqe_bo = NULL;
786 			return -EPERM;
787 		}
788 	}
789 
790 	/*
791 	 * Expanded APRIV and targets that support WHERE_AM_I both need a
792 	 * privileged buffer to store the RPTR shadow
793 	 */
794 	if ((adreno_gpu->base.hw_apriv || a6xx_gpu->has_whereami) &&
795 	    !a6xx_gpu->shadow_bo) {
796 		a6xx_gpu->shadow = msm_gem_kernel_new(gpu->dev,
797 						      sizeof(u32) * gpu->nr_rings,
798 						      MSM_BO_WC | MSM_BO_MAP_PRIV,
799 						      gpu->aspace, &a6xx_gpu->shadow_bo,
800 						      &a6xx_gpu->shadow_iova);
801 
802 		if (IS_ERR(a6xx_gpu->shadow))
803 			return PTR_ERR(a6xx_gpu->shadow);
804 
805 		msm_gem_object_set_name(a6xx_gpu->shadow_bo, "shadow");
806 	}
807 
808 	return 0;
809 }
810 
811 static int a6xx_zap_shader_init(struct msm_gpu *gpu)
812 {
813 	static bool loaded;
814 	int ret;
815 
816 	if (loaded)
817 		return 0;
818 
819 	ret = adreno_zap_shader_load(gpu, GPU_PAS_ID);
820 
821 	loaded = !ret;
822 	return ret;
823 }
824 
825 #define A6XX_INT_MASK (A6XX_RBBM_INT_0_MASK_CP_AHB_ERROR | \
826 		       A6XX_RBBM_INT_0_MASK_RBBM_ATB_ASYNCFIFO_OVERFLOW | \
827 		       A6XX_RBBM_INT_0_MASK_CP_HW_ERROR | \
828 		       A6XX_RBBM_INT_0_MASK_CP_IB2 | \
829 		       A6XX_RBBM_INT_0_MASK_CP_IB1 | \
830 		       A6XX_RBBM_INT_0_MASK_CP_RB | \
831 		       A6XX_RBBM_INT_0_MASK_CP_CACHE_FLUSH_TS | \
832 		       A6XX_RBBM_INT_0_MASK_RBBM_ATB_BUS_OVERFLOW | \
833 		       A6XX_RBBM_INT_0_MASK_RBBM_HANG_DETECT | \
834 		       A6XX_RBBM_INT_0_MASK_UCHE_OOB_ACCESS | \
835 		       A6XX_RBBM_INT_0_MASK_UCHE_TRAP_INTR)
836 
837 #define A7XX_INT_MASK (A6XX_RBBM_INT_0_MASK_CP_AHB_ERROR | \
838 		       A6XX_RBBM_INT_0_MASK_RBBM_ATB_ASYNCFIFO_OVERFLOW | \
839 		       A6XX_RBBM_INT_0_MASK_RBBM_GPC_ERROR | \
840 		       A6XX_RBBM_INT_0_MASK_CP_SW | \
841 		       A6XX_RBBM_INT_0_MASK_CP_HW_ERROR | \
842 		       A6XX_RBBM_INT_0_MASK_PM4CPINTERRUPT | \
843 		       A6XX_RBBM_INT_0_MASK_CP_RB_DONE_TS | \
844 		       A6XX_RBBM_INT_0_MASK_CP_CACHE_FLUSH_TS | \
845 		       A6XX_RBBM_INT_0_MASK_RBBM_ATB_BUS_OVERFLOW | \
846 		       A6XX_RBBM_INT_0_MASK_RBBM_HANG_DETECT | \
847 		       A6XX_RBBM_INT_0_MASK_UCHE_OOB_ACCESS | \
848 		       A6XX_RBBM_INT_0_MASK_UCHE_TRAP_INTR | \
849 		       A6XX_RBBM_INT_0_MASK_TSBWRITEERROR | \
850 		       A6XX_RBBM_INT_0_MASK_SWFUSEVIOLATION)
851 
852 #define A7XX_APRIV_MASK (A6XX_CP_APRIV_CNTL_ICACHE | \
853 			 A6XX_CP_APRIV_CNTL_RBFETCH | \
854 			 A6XX_CP_APRIV_CNTL_RBPRIVLEVEL | \
855 			 A6XX_CP_APRIV_CNTL_RBRPWB)
856 
857 #define A7XX_BR_APRIVMASK (A7XX_APRIV_MASK | \
858 			   A6XX_CP_APRIV_CNTL_CDREAD | \
859 			   A6XX_CP_APRIV_CNTL_CDWRITE)
860 
861 static int hw_init(struct msm_gpu *gpu)
862 {
863 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
864 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
865 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
866 	u64 gmem_range_min;
867 	int ret;
868 
869 	if (!adreno_has_gmu_wrapper(adreno_gpu)) {
870 		/* Make sure the GMU keeps the GPU on while we set it up */
871 		ret = a6xx_gmu_set_oob(&a6xx_gpu->gmu, GMU_OOB_GPU_SET);
872 		if (ret)
873 			return ret;
874 	}
875 
876 	/* Clear GBIF halt in case GX domain was not collapsed */
877 	if (adreno_is_a619_holi(adreno_gpu)) {
878 		gpu_write(gpu, REG_A6XX_GBIF_HALT, 0);
879 		gpu_read(gpu, REG_A6XX_GBIF_HALT);
880 
881 		gpu_write(gpu, REG_A6XX_RBBM_GPR0_CNTL, 0);
882 		gpu_read(gpu, REG_A6XX_RBBM_GPR0_CNTL);
883 	} else if (a6xx_has_gbif(adreno_gpu)) {
884 		gpu_write(gpu, REG_A6XX_GBIF_HALT, 0);
885 		gpu_read(gpu, REG_A6XX_GBIF_HALT);
886 
887 		gpu_write(gpu, REG_A6XX_RBBM_GBIF_HALT, 0);
888 		gpu_read(gpu, REG_A6XX_RBBM_GBIF_HALT);
889 	}
890 
891 	gpu_write(gpu, REG_A6XX_RBBM_SECVID_TSB_CNTL, 0);
892 
893 	if (adreno_is_a619_holi(adreno_gpu))
894 		a6xx_sptprac_enable(gmu);
895 
896 	/*
897 	 * Disable the trusted memory range - we don't actually supported secure
898 	 * memory rendering at this point in time and we don't want to block off
899 	 * part of the virtual memory space.
900 	 */
901 	gpu_write64(gpu, REG_A6XX_RBBM_SECVID_TSB_TRUSTED_BASE, 0x00000000);
902 	gpu_write(gpu, REG_A6XX_RBBM_SECVID_TSB_TRUSTED_SIZE, 0x00000000);
903 
904 	if (!adreno_is_a7xx(adreno_gpu)) {
905 		/* Turn on 64 bit addressing for all blocks */
906 		gpu_write(gpu, REG_A6XX_CP_ADDR_MODE_CNTL, 0x1);
907 		gpu_write(gpu, REG_A6XX_VSC_ADDR_MODE_CNTL, 0x1);
908 		gpu_write(gpu, REG_A6XX_GRAS_ADDR_MODE_CNTL, 0x1);
909 		gpu_write(gpu, REG_A6XX_RB_ADDR_MODE_CNTL, 0x1);
910 		gpu_write(gpu, REG_A6XX_PC_ADDR_MODE_CNTL, 0x1);
911 		gpu_write(gpu, REG_A6XX_HLSQ_ADDR_MODE_CNTL, 0x1);
912 		gpu_write(gpu, REG_A6XX_VFD_ADDR_MODE_CNTL, 0x1);
913 		gpu_write(gpu, REG_A6XX_VPC_ADDR_MODE_CNTL, 0x1);
914 		gpu_write(gpu, REG_A6XX_UCHE_ADDR_MODE_CNTL, 0x1);
915 		gpu_write(gpu, REG_A6XX_SP_ADDR_MODE_CNTL, 0x1);
916 		gpu_write(gpu, REG_A6XX_TPL1_ADDR_MODE_CNTL, 0x1);
917 		gpu_write(gpu, REG_A6XX_RBBM_SECVID_TSB_ADDR_MODE_CNTL, 0x1);
918 	}
919 
920 	/* enable hardware clockgating */
921 	a6xx_set_hwcg(gpu, true);
922 
923 	/* VBIF/GBIF start*/
924 	if (adreno_is_a610_family(adreno_gpu) ||
925 	    adreno_is_a640_family(adreno_gpu) ||
926 	    adreno_is_a650_family(adreno_gpu) ||
927 	    adreno_is_a7xx(adreno_gpu)) {
928 		gpu_write(gpu, REG_A6XX_GBIF_QSB_SIDE0, 0x00071620);
929 		gpu_write(gpu, REG_A6XX_GBIF_QSB_SIDE1, 0x00071620);
930 		gpu_write(gpu, REG_A6XX_GBIF_QSB_SIDE2, 0x00071620);
931 		gpu_write(gpu, REG_A6XX_GBIF_QSB_SIDE3, 0x00071620);
932 		gpu_write(gpu, REG_A6XX_RBBM_GBIF_CLIENT_QOS_CNTL,
933 			  adreno_is_a7xx(adreno_gpu) ? 0x2120212 : 0x3);
934 	} else {
935 		gpu_write(gpu, REG_A6XX_RBBM_VBIF_CLIENT_QOS_CNTL, 0x3);
936 	}
937 
938 	if (adreno_is_a630(adreno_gpu))
939 		gpu_write(gpu, REG_A6XX_VBIF_GATE_OFF_WRREQ_EN, 0x00000009);
940 
941 	if (adreno_is_a7xx(adreno_gpu))
942 		gpu_write(gpu, REG_A6XX_UCHE_GBIF_GX_CONFIG, 0x10240e0);
943 
944 	/* Make all blocks contribute to the GPU BUSY perf counter */
945 	gpu_write(gpu, REG_A6XX_RBBM_PERFCTR_GPU_BUSY_MASKED, 0xffffffff);
946 
947 	/* Disable L2 bypass in the UCHE */
948 	if (adreno_is_a7xx(adreno_gpu)) {
949 		gpu_write64(gpu, REG_A6XX_UCHE_TRAP_BASE, 0x0001fffffffff000llu);
950 		gpu_write64(gpu, REG_A6XX_UCHE_WRITE_THRU_BASE, 0x0001fffffffff000llu);
951 	} else {
952 		gpu_write64(gpu, REG_A6XX_UCHE_WRITE_RANGE_MAX, 0x0001ffffffffffc0llu);
953 		gpu_write64(gpu, REG_A6XX_UCHE_TRAP_BASE, 0x0001fffffffff000llu);
954 		gpu_write64(gpu, REG_A6XX_UCHE_WRITE_THRU_BASE, 0x0001fffffffff000llu);
955 	}
956 
957 	if (!(adreno_is_a650_family(adreno_gpu) ||
958 	      adreno_is_a702(adreno_gpu) ||
959 	      adreno_is_a730(adreno_gpu))) {
960 		gmem_range_min = adreno_is_a740_family(adreno_gpu) ? SZ_16M : SZ_1M;
961 
962 		/* Set the GMEM VA range [0x100000:0x100000 + gpu->gmem - 1] */
963 		gpu_write64(gpu, REG_A6XX_UCHE_GMEM_RANGE_MIN, gmem_range_min);
964 
965 		gpu_write64(gpu, REG_A6XX_UCHE_GMEM_RANGE_MAX,
966 			gmem_range_min + adreno_gpu->info->gmem - 1);
967 	}
968 
969 	if (adreno_is_a7xx(adreno_gpu))
970 		gpu_write(gpu, REG_A6XX_UCHE_CACHE_WAYS, BIT(23));
971 	else {
972 		gpu_write(gpu, REG_A6XX_UCHE_FILTER_CNTL, 0x804);
973 		gpu_write(gpu, REG_A6XX_UCHE_CACHE_WAYS, 0x4);
974 	}
975 
976 	if (adreno_is_a640_family(adreno_gpu) || adreno_is_a650_family(adreno_gpu)) {
977 		gpu_write(gpu, REG_A6XX_CP_ROQ_THRESHOLDS_2, 0x02000140);
978 		gpu_write(gpu, REG_A6XX_CP_ROQ_THRESHOLDS_1, 0x8040362c);
979 	} else if (adreno_is_a610_family(adreno_gpu)) {
980 		gpu_write(gpu, REG_A6XX_CP_ROQ_THRESHOLDS_2, 0x00800060);
981 		gpu_write(gpu, REG_A6XX_CP_ROQ_THRESHOLDS_1, 0x40201b16);
982 	} else if (!adreno_is_a7xx(adreno_gpu)) {
983 		gpu_write(gpu, REG_A6XX_CP_ROQ_THRESHOLDS_2, 0x010000c0);
984 		gpu_write(gpu, REG_A6XX_CP_ROQ_THRESHOLDS_1, 0x8040362c);
985 	}
986 
987 	if (adreno_is_a660_family(adreno_gpu))
988 		gpu_write(gpu, REG_A6XX_CP_LPAC_PROG_FIFO_SIZE, 0x00000020);
989 
990 	/* Setting the mem pool size */
991 	if (adreno_is_a610(adreno_gpu)) {
992 		gpu_write(gpu, REG_A6XX_CP_MEM_POOL_SIZE, 48);
993 		gpu_write(gpu, REG_A6XX_CP_MEM_POOL_DBG_ADDR, 47);
994 	} else if (adreno_is_a702(adreno_gpu)) {
995 		gpu_write(gpu, REG_A6XX_CP_MEM_POOL_SIZE, 64);
996 		gpu_write(gpu, REG_A6XX_CP_MEM_POOL_DBG_ADDR, 63);
997 	} else if (!adreno_is_a7xx(adreno_gpu))
998 		gpu_write(gpu, REG_A6XX_CP_MEM_POOL_SIZE, 128);
999 
1000 
1001 	/* Set the default primFifo threshold values */
1002 	if (adreno_gpu->info->a6xx->prim_fifo_threshold)
1003 		gpu_write(gpu, REG_A6XX_PC_DBG_ECO_CNTL,
1004 			  adreno_gpu->info->a6xx->prim_fifo_threshold);
1005 
1006 	/* Set the AHB default slave response to "ERROR" */
1007 	gpu_write(gpu, REG_A6XX_CP_AHB_CNTL, 0x1);
1008 
1009 	/* Turn on performance counters */
1010 	gpu_write(gpu, REG_A6XX_RBBM_PERFCTR_CNTL, 0x1);
1011 
1012 	if (adreno_is_a7xx(adreno_gpu)) {
1013 		/* Turn on the IFPC counter (countable 4 on XOCLK4) */
1014 		gmu_write(&a6xx_gpu->gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_SELECT_1,
1015 			  FIELD_PREP(GENMASK(7, 0), 0x4));
1016 	}
1017 
1018 	/* Select CP0 to always count cycles */
1019 	gpu_write(gpu, REG_A6XX_CP_PERFCTR_CP_SEL(0), PERF_CP_ALWAYS_COUNT);
1020 
1021 	a6xx_set_ubwc_config(gpu);
1022 
1023 	/* Enable fault detection */
1024 	if (adreno_is_a730(adreno_gpu) ||
1025 	    adreno_is_a740_family(adreno_gpu))
1026 		gpu_write(gpu, REG_A6XX_RBBM_INTERFACE_HANG_INT_CNTL, (1 << 30) | 0xcfffff);
1027 	else if (adreno_is_a690(adreno_gpu))
1028 		gpu_write(gpu, REG_A6XX_RBBM_INTERFACE_HANG_INT_CNTL, (1 << 30) | 0x4fffff);
1029 	else if (adreno_is_a619(adreno_gpu))
1030 		gpu_write(gpu, REG_A6XX_RBBM_INTERFACE_HANG_INT_CNTL, (1 << 30) | 0x3fffff);
1031 	else if (adreno_is_a610(adreno_gpu) || adreno_is_a702(adreno_gpu))
1032 		gpu_write(gpu, REG_A6XX_RBBM_INTERFACE_HANG_INT_CNTL, (1 << 30) | 0x3ffff);
1033 	else
1034 		gpu_write(gpu, REG_A6XX_RBBM_INTERFACE_HANG_INT_CNTL, (1 << 30) | 0x1fffff);
1035 
1036 	gpu_write(gpu, REG_A6XX_UCHE_CLIENT_PF, BIT(7) | 0x1);
1037 
1038 	/* Set weights for bicubic filtering */
1039 	if (adreno_is_a650_family(adreno_gpu) || adreno_is_x185(adreno_gpu)) {
1040 		gpu_write(gpu, REG_A6XX_TPL1_BICUBIC_WEIGHTS_TABLE_0, 0);
1041 		gpu_write(gpu, REG_A6XX_TPL1_BICUBIC_WEIGHTS_TABLE_1,
1042 			0x3fe05ff4);
1043 		gpu_write(gpu, REG_A6XX_TPL1_BICUBIC_WEIGHTS_TABLE_2,
1044 			0x3fa0ebee);
1045 		gpu_write(gpu, REG_A6XX_TPL1_BICUBIC_WEIGHTS_TABLE_3,
1046 			0x3f5193ed);
1047 		gpu_write(gpu, REG_A6XX_TPL1_BICUBIC_WEIGHTS_TABLE_4,
1048 			0x3f0243f0);
1049 	}
1050 
1051 	/* Set up the CX GMU counter 0 to count busy ticks */
1052 	gmu_write(gmu, REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_MASK, 0xff000000);
1053 
1054 	/* Enable the power counter */
1055 	gmu_rmw(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_SELECT_0, 0xff, BIT(5));
1056 	gmu_write(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_ENABLE, 1);
1057 
1058 	/* Protect registers from the CP */
1059 	a6xx_set_cp_protect(gpu);
1060 
1061 	if (adreno_is_a660_family(adreno_gpu)) {
1062 		if (adreno_is_a690(adreno_gpu))
1063 			gpu_write(gpu, REG_A6XX_CP_CHICKEN_DBG, 0x00028801);
1064 		else
1065 			gpu_write(gpu, REG_A6XX_CP_CHICKEN_DBG, 0x1);
1066 		gpu_write(gpu, REG_A6XX_RBBM_GBIF_CLIENT_QOS_CNTL, 0x0);
1067 	} else if (adreno_is_a702(adreno_gpu)) {
1068 		/* Something to do with the HLSQ cluster */
1069 		gpu_write(gpu, REG_A6XX_CP_CHICKEN_DBG, BIT(24));
1070 	}
1071 
1072 	if (adreno_is_a690(adreno_gpu))
1073 		gpu_write(gpu, REG_A6XX_UCHE_CMDQ_CONFIG, 0x90);
1074 	/* Set dualQ + disable afull for A660 GPU */
1075 	else if (adreno_is_a660(adreno_gpu))
1076 		gpu_write(gpu, REG_A6XX_UCHE_CMDQ_CONFIG, 0x66906);
1077 	else if (adreno_is_a7xx(adreno_gpu))
1078 		gpu_write(gpu, REG_A6XX_UCHE_CMDQ_CONFIG,
1079 			  FIELD_PREP(GENMASK(19, 16), 6) |
1080 			  FIELD_PREP(GENMASK(15, 12), 6) |
1081 			  FIELD_PREP(GENMASK(11, 8), 9) |
1082 			  BIT(3) | BIT(2) |
1083 			  FIELD_PREP(GENMASK(1, 0), 2));
1084 
1085 	/* Enable expanded apriv for targets that support it */
1086 	if (gpu->hw_apriv) {
1087 		if (adreno_is_a7xx(adreno_gpu)) {
1088 			gpu_write(gpu, REG_A6XX_CP_APRIV_CNTL,
1089 				  A7XX_BR_APRIVMASK);
1090 			gpu_write(gpu, REG_A7XX_CP_BV_APRIV_CNTL,
1091 				  A7XX_APRIV_MASK);
1092 			gpu_write(gpu, REG_A7XX_CP_LPAC_APRIV_CNTL,
1093 				  A7XX_APRIV_MASK);
1094 		} else
1095 			gpu_write(gpu, REG_A6XX_CP_APRIV_CNTL,
1096 				  BIT(6) | BIT(5) | BIT(3) | BIT(2) | BIT(1));
1097 	}
1098 
1099 	if (adreno_is_a750(adreno_gpu)) {
1100 		/* Disable ubwc merged UFC request feature */
1101 		gpu_rmw(gpu, REG_A6XX_RB_CMP_DBG_ECO_CNTL, BIT(19), BIT(19));
1102 
1103 		/* Enable TP flaghint and other performance settings */
1104 		gpu_write(gpu, REG_A6XX_TPL1_DBG_ECO_CNTL1, 0xc0700);
1105 	} else if (adreno_is_a7xx(adreno_gpu)) {
1106 		/* Disable non-ubwc read reqs from passing write reqs */
1107 		gpu_rmw(gpu, REG_A6XX_RB_CMP_DBG_ECO_CNTL, BIT(11), BIT(11));
1108 	}
1109 
1110 	/* Enable interrupts */
1111 	gpu_write(gpu, REG_A6XX_RBBM_INT_0_MASK,
1112 		  adreno_is_a7xx(adreno_gpu) ? A7XX_INT_MASK : A6XX_INT_MASK);
1113 
1114 	ret = adreno_hw_init(gpu);
1115 	if (ret)
1116 		goto out;
1117 
1118 	gpu_write64(gpu, REG_A6XX_CP_SQE_INSTR_BASE, a6xx_gpu->sqe_iova);
1119 
1120 	/* Set the ringbuffer address */
1121 	gpu_write64(gpu, REG_A6XX_CP_RB_BASE, gpu->rb[0]->iova);
1122 
1123 	/* Targets that support extended APRIV can use the RPTR shadow from
1124 	 * hardware but all the other ones need to disable the feature. Targets
1125 	 * that support the WHERE_AM_I opcode can use that instead
1126 	 */
1127 	if (adreno_gpu->base.hw_apriv)
1128 		gpu_write(gpu, REG_A6XX_CP_RB_CNTL, MSM_GPU_RB_CNTL_DEFAULT);
1129 	else
1130 		gpu_write(gpu, REG_A6XX_CP_RB_CNTL,
1131 			MSM_GPU_RB_CNTL_DEFAULT | AXXX_CP_RB_CNTL_NO_UPDATE);
1132 
1133 	/* Configure the RPTR shadow if needed: */
1134 	if (a6xx_gpu->shadow_bo) {
1135 		gpu_write64(gpu, REG_A6XX_CP_RB_RPTR_ADDR,
1136 			shadowptr(a6xx_gpu, gpu->rb[0]));
1137 	}
1138 
1139 	/* ..which means "always" on A7xx, also for BV shadow */
1140 	if (adreno_is_a7xx(adreno_gpu)) {
1141 		gpu_write64(gpu, REG_A7XX_CP_BV_RB_RPTR_ADDR,
1142 			    rbmemptr(gpu->rb[0], bv_fence));
1143 	}
1144 
1145 	/* Always come up on rb 0 */
1146 	a6xx_gpu->cur_ring = gpu->rb[0];
1147 
1148 	gpu->cur_ctx_seqno = 0;
1149 
1150 	/* Enable the SQE_to start the CP engine */
1151 	gpu_write(gpu, REG_A6XX_CP_SQE_CNTL, 1);
1152 
1153 	ret = adreno_is_a7xx(adreno_gpu) ? a7xx_cp_init(gpu) : a6xx_cp_init(gpu);
1154 	if (ret)
1155 		goto out;
1156 
1157 	/*
1158 	 * Try to load a zap shader into the secure world. If successful
1159 	 * we can use the CP to switch out of secure mode. If not then we
1160 	 * have no resource but to try to switch ourselves out manually. If we
1161 	 * guessed wrong then access to the RBBM_SECVID_TRUST_CNTL register will
1162 	 * be blocked and a permissions violation will soon follow.
1163 	 */
1164 	ret = a6xx_zap_shader_init(gpu);
1165 	if (!ret) {
1166 		OUT_PKT7(gpu->rb[0], CP_SET_SECURE_MODE, 1);
1167 		OUT_RING(gpu->rb[0], 0x00000000);
1168 
1169 		a6xx_flush(gpu, gpu->rb[0]);
1170 		if (!a6xx_idle(gpu, gpu->rb[0]))
1171 			return -EINVAL;
1172 	} else if (ret == -ENODEV) {
1173 		/*
1174 		 * This device does not use zap shader (but print a warning
1175 		 * just in case someone got their dt wrong.. hopefully they
1176 		 * have a debug UART to realize the error of their ways...
1177 		 * if you mess this up you are about to crash horribly)
1178 		 */
1179 		dev_warn_once(gpu->dev->dev,
1180 			"Zap shader not enabled - using SECVID_TRUST_CNTL instead\n");
1181 		gpu_write(gpu, REG_A6XX_RBBM_SECVID_TRUST_CNTL, 0x0);
1182 		ret = 0;
1183 	} else {
1184 		return ret;
1185 	}
1186 
1187 out:
1188 	if (adreno_has_gmu_wrapper(adreno_gpu))
1189 		return ret;
1190 	/*
1191 	 * Tell the GMU that we are done touching the GPU and it can start power
1192 	 * management
1193 	 */
1194 	a6xx_gmu_clear_oob(&a6xx_gpu->gmu, GMU_OOB_GPU_SET);
1195 
1196 	if (a6xx_gpu->gmu.legacy) {
1197 		/* Take the GMU out of its special boot mode */
1198 		a6xx_gmu_clear_oob(&a6xx_gpu->gmu, GMU_OOB_BOOT_SLUMBER);
1199 	}
1200 
1201 	return ret;
1202 }
1203 
1204 static int a6xx_hw_init(struct msm_gpu *gpu)
1205 {
1206 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1207 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1208 	int ret;
1209 
1210 	mutex_lock(&a6xx_gpu->gmu.lock);
1211 	ret = hw_init(gpu);
1212 	mutex_unlock(&a6xx_gpu->gmu.lock);
1213 
1214 	return ret;
1215 }
1216 
1217 static void a6xx_dump(struct msm_gpu *gpu)
1218 {
1219 	DRM_DEV_INFO(&gpu->pdev->dev, "status:   %08x\n",
1220 			gpu_read(gpu, REG_A6XX_RBBM_STATUS));
1221 	adreno_dump(gpu);
1222 }
1223 
1224 static void a6xx_recover(struct msm_gpu *gpu)
1225 {
1226 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1227 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1228 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1229 	int i, active_submits;
1230 
1231 	adreno_dump_info(gpu);
1232 
1233 	for (i = 0; i < 8; i++)
1234 		DRM_DEV_INFO(&gpu->pdev->dev, "CP_SCRATCH_REG%d: %u\n", i,
1235 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(i)));
1236 
1237 	if (hang_debug)
1238 		a6xx_dump(gpu);
1239 
1240 	/*
1241 	 * To handle recovery specific sequences during the rpm suspend we are
1242 	 * about to trigger
1243 	 */
1244 	a6xx_gpu->hung = true;
1245 
1246 	/* Halt SQE first */
1247 	gpu_write(gpu, REG_A6XX_CP_SQE_CNTL, 3);
1248 
1249 	pm_runtime_dont_use_autosuspend(&gpu->pdev->dev);
1250 
1251 	/* active_submit won't change until we make a submission */
1252 	mutex_lock(&gpu->active_lock);
1253 	active_submits = gpu->active_submits;
1254 
1255 	/*
1256 	 * Temporarily clear active_submits count to silence a WARN() in the
1257 	 * runtime suspend cb
1258 	 */
1259 	gpu->active_submits = 0;
1260 
1261 	if (adreno_has_gmu_wrapper(adreno_gpu)) {
1262 		/* Drain the outstanding traffic on memory buses */
1263 		a6xx_bus_clear_pending_transactions(adreno_gpu, true);
1264 
1265 		/* Reset the GPU to a clean state */
1266 		a6xx_gpu_sw_reset(gpu, true);
1267 		a6xx_gpu_sw_reset(gpu, false);
1268 	}
1269 
1270 	reinit_completion(&gmu->pd_gate);
1271 	dev_pm_genpd_add_notifier(gmu->cxpd, &gmu->pd_nb);
1272 	dev_pm_genpd_synced_poweroff(gmu->cxpd);
1273 
1274 	/* Drop the rpm refcount from active submits */
1275 	if (active_submits)
1276 		pm_runtime_put(&gpu->pdev->dev);
1277 
1278 	/* And the final one from recover worker */
1279 	pm_runtime_put_sync(&gpu->pdev->dev);
1280 
1281 	if (!wait_for_completion_timeout(&gmu->pd_gate, msecs_to_jiffies(1000)))
1282 		DRM_DEV_ERROR(&gpu->pdev->dev, "cx gdsc didn't collapse\n");
1283 
1284 	dev_pm_genpd_remove_notifier(gmu->cxpd);
1285 
1286 	pm_runtime_use_autosuspend(&gpu->pdev->dev);
1287 
1288 	if (active_submits)
1289 		pm_runtime_get(&gpu->pdev->dev);
1290 
1291 	pm_runtime_get_sync(&gpu->pdev->dev);
1292 
1293 	gpu->active_submits = active_submits;
1294 	mutex_unlock(&gpu->active_lock);
1295 
1296 	msm_gpu_hw_init(gpu);
1297 	a6xx_gpu->hung = false;
1298 }
1299 
1300 static const char *a6xx_uche_fault_block(struct msm_gpu *gpu, u32 mid)
1301 {
1302 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1303 	static const char *uche_clients[7] = {
1304 		"VFD", "SP", "VSC", "VPC", "HLSQ", "PC", "LRZ",
1305 	};
1306 	u32 val;
1307 
1308 	if (adreno_is_a7xx(adreno_gpu)) {
1309 		if (mid != 1 && mid != 2 && mid != 3 && mid != 8)
1310 			return "UNKNOWN";
1311 	} else {
1312 		if (mid < 1 || mid > 3)
1313 			return "UNKNOWN";
1314 	}
1315 
1316 	/*
1317 	 * The source of the data depends on the mid ID read from FSYNR1.
1318 	 * and the client ID read from the UCHE block
1319 	 */
1320 	val = gpu_read(gpu, REG_A6XX_UCHE_CLIENT_PF);
1321 
1322 	if (adreno_is_a7xx(adreno_gpu)) {
1323 		/* Bit 3 for mid=3 indicates BR or BV */
1324 		static const char *uche_clients_a7xx[16] = {
1325 			"BR_VFD", "BR_SP", "BR_VSC", "BR_VPC",
1326 			"BR_HLSQ", "BR_PC", "BR_LRZ", "BR_TP",
1327 			"BV_VFD", "BV_SP", "BV_VSC", "BV_VPC",
1328 			"BV_HLSQ", "BV_PC", "BV_LRZ", "BV_TP",
1329 		};
1330 
1331 		/* LPAC has the same clients as BR and BV, but because it is
1332 		 * compute-only some of them do not exist and there are holes
1333 		 * in the array.
1334 		 */
1335 		static const char *uche_clients_lpac_a7xx[8] = {
1336 			"-", "LPAC_SP", "-", "-",
1337 			"LPAC_HLSQ", "-", "-", "LPAC_TP",
1338 		};
1339 
1340 		val &= GENMASK(6, 0);
1341 
1342 		/* mid=3 refers to BR or BV */
1343 		if (mid == 3) {
1344 			if (val < ARRAY_SIZE(uche_clients_a7xx))
1345 				return uche_clients_a7xx[val];
1346 			else
1347 				return "UCHE";
1348 		}
1349 
1350 		/* mid=8 refers to LPAC */
1351 		if (mid == 8) {
1352 			if (val < ARRAY_SIZE(uche_clients_lpac_a7xx))
1353 				return uche_clients_lpac_a7xx[val];
1354 			else
1355 				return "UCHE_LPAC";
1356 		}
1357 
1358 		/* mid=2 is a catchall for everything else in LPAC */
1359 		if (mid == 2)
1360 			return "UCHE_LPAC";
1361 
1362 		/* mid=1 is a catchall for everything else in BR/BV */
1363 		return "UCHE";
1364 	} else if (adreno_is_a660_family(adreno_gpu)) {
1365 		static const char *uche_clients_a660[8] = {
1366 			"VFD", "SP", "VSC", "VPC", "HLSQ", "PC", "LRZ", "TP",
1367 		};
1368 
1369 		static const char *uche_clients_a660_not[8] = {
1370 			"not VFD", "not SP", "not VSC", "not VPC",
1371 			"not HLSQ", "not PC", "not LRZ", "not TP",
1372 		};
1373 
1374 		val &= GENMASK(6, 0);
1375 
1376 		if (mid == 3 && val < ARRAY_SIZE(uche_clients_a660))
1377 			return uche_clients_a660[val];
1378 
1379 		if (mid == 1 && val < ARRAY_SIZE(uche_clients_a660_not))
1380 			return uche_clients_a660_not[val];
1381 
1382 		return "UCHE";
1383 	} else {
1384 		/* mid = 3 is most precise and refers to only one block per client */
1385 		if (mid == 3)
1386 			return uche_clients[val & 7];
1387 
1388 		/* For mid=2 the source is TP or VFD except when the client id is 0 */
1389 		if (mid == 2)
1390 			return ((val & 7) == 0) ? "TP" : "TP|VFD";
1391 
1392 		/* For mid=1 just return "UCHE" as a catchall for everything else */
1393 		return "UCHE";
1394 	}
1395 }
1396 
1397 static const char *a6xx_fault_block(struct msm_gpu *gpu, u32 id)
1398 {
1399 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1400 
1401 	if (id == 0)
1402 		return "CP";
1403 	else if (id == 4)
1404 		return "CCU";
1405 	else if (id == 6)
1406 		return "CDP Prefetch";
1407 	else if (id == 7)
1408 		return "GMU";
1409 	else if (id == 5 && adreno_is_a7xx(adreno_gpu))
1410 		return "Flag cache";
1411 
1412 	return a6xx_uche_fault_block(gpu, id);
1413 }
1414 
1415 static int a6xx_fault_handler(void *arg, unsigned long iova, int flags, void *data)
1416 {
1417 	struct msm_gpu *gpu = arg;
1418 	struct adreno_smmu_fault_info *info = data;
1419 	const char *block = "unknown";
1420 
1421 	u32 scratch[] = {
1422 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(4)),
1423 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(5)),
1424 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(6)),
1425 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(7)),
1426 	};
1427 
1428 	if (info)
1429 		block = a6xx_fault_block(gpu, info->fsynr1 & 0xff);
1430 
1431 	return adreno_fault_handler(gpu, iova, flags, info, block, scratch);
1432 }
1433 
1434 static void a6xx_cp_hw_err_irq(struct msm_gpu *gpu)
1435 {
1436 	u32 status = gpu_read(gpu, REG_A6XX_CP_INTERRUPT_STATUS);
1437 
1438 	if (status & A6XX_CP_INT_CP_OPCODE_ERROR) {
1439 		u32 val;
1440 
1441 		gpu_write(gpu, REG_A6XX_CP_SQE_STAT_ADDR, 1);
1442 		val = gpu_read(gpu, REG_A6XX_CP_SQE_STAT_DATA);
1443 		dev_err_ratelimited(&gpu->pdev->dev,
1444 			"CP | opcode error | possible opcode=0x%8.8X\n",
1445 			val);
1446 	}
1447 
1448 	if (status & A6XX_CP_INT_CP_UCODE_ERROR)
1449 		dev_err_ratelimited(&gpu->pdev->dev,
1450 			"CP ucode error interrupt\n");
1451 
1452 	if (status & A6XX_CP_INT_CP_HW_FAULT_ERROR)
1453 		dev_err_ratelimited(&gpu->pdev->dev, "CP | HW fault | status=0x%8.8X\n",
1454 			gpu_read(gpu, REG_A6XX_CP_HW_FAULT));
1455 
1456 	if (status & A6XX_CP_INT_CP_REGISTER_PROTECTION_ERROR) {
1457 		u32 val = gpu_read(gpu, REG_A6XX_CP_PROTECT_STATUS);
1458 
1459 		dev_err_ratelimited(&gpu->pdev->dev,
1460 			"CP | protected mode error | %s | addr=0x%8.8X | status=0x%8.8X\n",
1461 			val & (1 << 20) ? "READ" : "WRITE",
1462 			(val & 0x3ffff), val);
1463 	}
1464 
1465 	if (status & A6XX_CP_INT_CP_AHB_ERROR && !adreno_is_a7xx(to_adreno_gpu(gpu)))
1466 		dev_err_ratelimited(&gpu->pdev->dev, "CP AHB error interrupt\n");
1467 
1468 	if (status & A6XX_CP_INT_CP_VSD_PARITY_ERROR)
1469 		dev_err_ratelimited(&gpu->pdev->dev, "CP VSD decoder parity error\n");
1470 
1471 	if (status & A6XX_CP_INT_CP_ILLEGAL_INSTR_ERROR)
1472 		dev_err_ratelimited(&gpu->pdev->dev, "CP illegal instruction error\n");
1473 
1474 }
1475 
1476 static void a6xx_fault_detect_irq(struct msm_gpu *gpu)
1477 {
1478 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1479 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1480 	struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
1481 
1482 	/*
1483 	 * If stalled on SMMU fault, we could trip the GPU's hang detection,
1484 	 * but the fault handler will trigger the devcore dump, and we want
1485 	 * to otherwise resume normally rather than killing the submit, so
1486 	 * just bail.
1487 	 */
1488 	if (gpu_read(gpu, REG_A6XX_RBBM_STATUS3) & A6XX_RBBM_STATUS3_SMMU_STALLED_ON_FAULT)
1489 		return;
1490 
1491 	/*
1492 	 * Force the GPU to stay on until after we finish
1493 	 * collecting information
1494 	 */
1495 	if (!adreno_has_gmu_wrapper(adreno_gpu))
1496 		gmu_write(&a6xx_gpu->gmu, REG_A6XX_GMU_GMU_PWR_COL_KEEPALIVE, 1);
1497 
1498 	DRM_DEV_ERROR(&gpu->pdev->dev,
1499 		"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",
1500 		ring ? ring->id : -1, ring ? ring->fctx->last_fence : 0,
1501 		gpu_read(gpu, REG_A6XX_RBBM_STATUS),
1502 		gpu_read(gpu, REG_A6XX_CP_RB_RPTR),
1503 		gpu_read(gpu, REG_A6XX_CP_RB_WPTR),
1504 		gpu_read64(gpu, REG_A6XX_CP_IB1_BASE),
1505 		gpu_read(gpu, REG_A6XX_CP_IB1_REM_SIZE),
1506 		gpu_read64(gpu, REG_A6XX_CP_IB2_BASE),
1507 		gpu_read(gpu, REG_A6XX_CP_IB2_REM_SIZE));
1508 
1509 	/* Turn off the hangcheck timer to keep it from bothering us */
1510 	del_timer(&gpu->hangcheck_timer);
1511 
1512 	kthread_queue_work(gpu->worker, &gpu->recover_work);
1513 }
1514 
1515 static void a7xx_sw_fuse_violation_irq(struct msm_gpu *gpu)
1516 {
1517 	u32 status;
1518 
1519 	status = gpu_read(gpu, REG_A7XX_RBBM_SW_FUSE_INT_STATUS);
1520 	gpu_write(gpu, REG_A7XX_RBBM_SW_FUSE_INT_MASK, 0);
1521 
1522 	dev_err_ratelimited(&gpu->pdev->dev, "SW fuse violation status=%8.8x\n", status);
1523 
1524 	/*
1525 	 * Ignore FASTBLEND violations, because the HW will silently fall back
1526 	 * to legacy blending.
1527 	 */
1528 	if (status & (A7XX_CX_MISC_SW_FUSE_VALUE_RAYTRACING |
1529 		      A7XX_CX_MISC_SW_FUSE_VALUE_LPAC)) {
1530 		del_timer(&gpu->hangcheck_timer);
1531 
1532 		kthread_queue_work(gpu->worker, &gpu->recover_work);
1533 	}
1534 }
1535 
1536 static irqreturn_t a6xx_irq(struct msm_gpu *gpu)
1537 {
1538 	struct msm_drm_private *priv = gpu->dev->dev_private;
1539 	u32 status = gpu_read(gpu, REG_A6XX_RBBM_INT_0_STATUS);
1540 
1541 	gpu_write(gpu, REG_A6XX_RBBM_INT_CLEAR_CMD, status);
1542 
1543 	if (priv->disable_err_irq)
1544 		status &= A6XX_RBBM_INT_0_MASK_CP_CACHE_FLUSH_TS;
1545 
1546 	if (status & A6XX_RBBM_INT_0_MASK_RBBM_HANG_DETECT)
1547 		a6xx_fault_detect_irq(gpu);
1548 
1549 	if (status & A6XX_RBBM_INT_0_MASK_CP_AHB_ERROR)
1550 		dev_err_ratelimited(&gpu->pdev->dev, "CP | AHB bus error\n");
1551 
1552 	if (status & A6XX_RBBM_INT_0_MASK_CP_HW_ERROR)
1553 		a6xx_cp_hw_err_irq(gpu);
1554 
1555 	if (status & A6XX_RBBM_INT_0_MASK_RBBM_ATB_ASYNCFIFO_OVERFLOW)
1556 		dev_err_ratelimited(&gpu->pdev->dev, "RBBM | ATB ASYNC overflow\n");
1557 
1558 	if (status & A6XX_RBBM_INT_0_MASK_RBBM_ATB_BUS_OVERFLOW)
1559 		dev_err_ratelimited(&gpu->pdev->dev, "RBBM | ATB bus overflow\n");
1560 
1561 	if (status & A6XX_RBBM_INT_0_MASK_UCHE_OOB_ACCESS)
1562 		dev_err_ratelimited(&gpu->pdev->dev, "UCHE | Out of bounds access\n");
1563 
1564 	if (status & A6XX_RBBM_INT_0_MASK_SWFUSEVIOLATION)
1565 		a7xx_sw_fuse_violation_irq(gpu);
1566 
1567 	if (status & A6XX_RBBM_INT_0_MASK_CP_CACHE_FLUSH_TS)
1568 		msm_gpu_retire(gpu);
1569 
1570 	return IRQ_HANDLED;
1571 }
1572 
1573 static void a6xx_llc_deactivate(struct a6xx_gpu *a6xx_gpu)
1574 {
1575 	llcc_slice_deactivate(a6xx_gpu->llc_slice);
1576 	llcc_slice_deactivate(a6xx_gpu->htw_llc_slice);
1577 }
1578 
1579 static void a6xx_llc_activate(struct a6xx_gpu *a6xx_gpu)
1580 {
1581 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1582 	struct msm_gpu *gpu = &adreno_gpu->base;
1583 	u32 cntl1_regval = 0;
1584 
1585 	if (IS_ERR(a6xx_gpu->llc_mmio))
1586 		return;
1587 
1588 	if (!llcc_slice_activate(a6xx_gpu->llc_slice)) {
1589 		u32 gpu_scid = llcc_get_slice_id(a6xx_gpu->llc_slice);
1590 
1591 		gpu_scid &= 0x1f;
1592 		cntl1_regval = (gpu_scid << 0) | (gpu_scid << 5) | (gpu_scid << 10) |
1593 			       (gpu_scid << 15) | (gpu_scid << 20);
1594 
1595 		/* On A660, the SCID programming for UCHE traffic is done in
1596 		 * A6XX_GBIF_SCACHE_CNTL0[14:10]
1597 		 */
1598 		if (adreno_is_a660_family(adreno_gpu))
1599 			gpu_rmw(gpu, REG_A6XX_GBIF_SCACHE_CNTL0, (0x1f << 10) |
1600 				(1 << 8), (gpu_scid << 10) | (1 << 8));
1601 	}
1602 
1603 	/*
1604 	 * For targets with a MMU500, activate the slice but don't program the
1605 	 * register.  The XBL will take care of that.
1606 	 */
1607 	if (!llcc_slice_activate(a6xx_gpu->htw_llc_slice)) {
1608 		if (!a6xx_gpu->have_mmu500) {
1609 			u32 gpuhtw_scid = llcc_get_slice_id(a6xx_gpu->htw_llc_slice);
1610 
1611 			gpuhtw_scid &= 0x1f;
1612 			cntl1_regval |= FIELD_PREP(GENMASK(29, 25), gpuhtw_scid);
1613 		}
1614 	}
1615 
1616 	if (!cntl1_regval)
1617 		return;
1618 
1619 	/*
1620 	 * Program the slice IDs for the various GPU blocks and GPU MMU
1621 	 * pagetables
1622 	 */
1623 	if (!a6xx_gpu->have_mmu500) {
1624 		a6xx_llc_write(a6xx_gpu,
1625 			REG_A6XX_CX_MISC_SYSTEM_CACHE_CNTL_1, cntl1_regval);
1626 
1627 		/*
1628 		 * Program cacheability overrides to not allocate cache
1629 		 * lines on a write miss
1630 		 */
1631 		a6xx_llc_rmw(a6xx_gpu,
1632 			REG_A6XX_CX_MISC_SYSTEM_CACHE_CNTL_0, 0xF, 0x03);
1633 		return;
1634 	}
1635 
1636 	gpu_rmw(gpu, REG_A6XX_GBIF_SCACHE_CNTL1, GENMASK(24, 0), cntl1_regval);
1637 }
1638 
1639 static void a7xx_llc_activate(struct a6xx_gpu *a6xx_gpu)
1640 {
1641 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1642 	struct msm_gpu *gpu = &adreno_gpu->base;
1643 
1644 	if (IS_ERR(a6xx_gpu->llc_mmio))
1645 		return;
1646 
1647 	if (!llcc_slice_activate(a6xx_gpu->llc_slice)) {
1648 		u32 gpu_scid = llcc_get_slice_id(a6xx_gpu->llc_slice);
1649 
1650 		gpu_scid &= GENMASK(4, 0);
1651 
1652 		gpu_write(gpu, REG_A6XX_GBIF_SCACHE_CNTL1,
1653 			  FIELD_PREP(GENMASK(29, 25), gpu_scid) |
1654 			  FIELD_PREP(GENMASK(24, 20), gpu_scid) |
1655 			  FIELD_PREP(GENMASK(19, 15), gpu_scid) |
1656 			  FIELD_PREP(GENMASK(14, 10), gpu_scid) |
1657 			  FIELD_PREP(GENMASK(9, 5), gpu_scid) |
1658 			  FIELD_PREP(GENMASK(4, 0), gpu_scid));
1659 
1660 		gpu_write(gpu, REG_A6XX_GBIF_SCACHE_CNTL0,
1661 			  FIELD_PREP(GENMASK(14, 10), gpu_scid) |
1662 			  BIT(8));
1663 	}
1664 
1665 	llcc_slice_activate(a6xx_gpu->htw_llc_slice);
1666 }
1667 
1668 static void a6xx_llc_slices_destroy(struct a6xx_gpu *a6xx_gpu)
1669 {
1670 	/* No LLCC on non-RPMh (and by extension, non-GMU) SoCs */
1671 	if (adreno_has_gmu_wrapper(&a6xx_gpu->base))
1672 		return;
1673 
1674 	llcc_slice_putd(a6xx_gpu->llc_slice);
1675 	llcc_slice_putd(a6xx_gpu->htw_llc_slice);
1676 }
1677 
1678 static void a6xx_llc_slices_init(struct platform_device *pdev,
1679 		struct a6xx_gpu *a6xx_gpu, bool is_a7xx)
1680 {
1681 	struct device_node *phandle;
1682 
1683 	/* No LLCC on non-RPMh (and by extension, non-GMU) SoCs */
1684 	if (adreno_has_gmu_wrapper(&a6xx_gpu->base))
1685 		return;
1686 
1687 	/*
1688 	 * There is a different programming path for A6xx targets with an
1689 	 * mmu500 attached, so detect if that is the case
1690 	 */
1691 	phandle = of_parse_phandle(pdev->dev.of_node, "iommus", 0);
1692 	a6xx_gpu->have_mmu500 = (phandle &&
1693 		of_device_is_compatible(phandle, "arm,mmu-500"));
1694 	of_node_put(phandle);
1695 
1696 	if (is_a7xx || !a6xx_gpu->have_mmu500)
1697 		a6xx_gpu->llc_mmio = msm_ioremap(pdev, "cx_mem");
1698 	else
1699 		a6xx_gpu->llc_mmio = NULL;
1700 
1701 	a6xx_gpu->llc_slice = llcc_slice_getd(LLCC_GPU);
1702 	a6xx_gpu->htw_llc_slice = llcc_slice_getd(LLCC_GPUHTW);
1703 
1704 	if (IS_ERR_OR_NULL(a6xx_gpu->llc_slice) && IS_ERR_OR_NULL(a6xx_gpu->htw_llc_slice))
1705 		a6xx_gpu->llc_mmio = ERR_PTR(-EINVAL);
1706 }
1707 
1708 static int a7xx_cx_mem_init(struct a6xx_gpu *a6xx_gpu)
1709 {
1710 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1711 	struct msm_gpu *gpu = &adreno_gpu->base;
1712 	u32 fuse_val;
1713 	int ret;
1714 
1715 	if (adreno_is_a750(adreno_gpu)) {
1716 		/*
1717 		 * Assume that if qcom scm isn't available, that whatever
1718 		 * replacement allows writing the fuse register ourselves.
1719 		 * Users of alternative firmware need to make sure this
1720 		 * register is writeable or indicate that it's not somehow.
1721 		 * Print a warning because if you mess this up you're about to
1722 		 * crash horribly.
1723 		 */
1724 		if (!qcom_scm_is_available()) {
1725 			dev_warn_once(gpu->dev->dev,
1726 				"SCM is not available, poking fuse register\n");
1727 			a6xx_llc_write(a6xx_gpu, REG_A7XX_CX_MISC_SW_FUSE_VALUE,
1728 				A7XX_CX_MISC_SW_FUSE_VALUE_RAYTRACING |
1729 				A7XX_CX_MISC_SW_FUSE_VALUE_FASTBLEND |
1730 				A7XX_CX_MISC_SW_FUSE_VALUE_LPAC);
1731 			adreno_gpu->has_ray_tracing = true;
1732 			return 0;
1733 		}
1734 
1735 		ret = qcom_scm_gpu_init_regs(QCOM_SCM_GPU_ALWAYS_EN_REQ |
1736 					     QCOM_SCM_GPU_TSENSE_EN_REQ);
1737 		if (ret)
1738 			return ret;
1739 
1740 		/*
1741 		 * On a750 raytracing may be disabled by the firmware, find out
1742 		 * whether that's the case. The scm call above sets the fuse
1743 		 * register.
1744 		 */
1745 		fuse_val = a6xx_llc_read(a6xx_gpu,
1746 					 REG_A7XX_CX_MISC_SW_FUSE_VALUE);
1747 		adreno_gpu->has_ray_tracing =
1748 			!!(fuse_val & A7XX_CX_MISC_SW_FUSE_VALUE_RAYTRACING);
1749 	} else if (adreno_is_a740(adreno_gpu)) {
1750 		/* Raytracing is always enabled on a740 */
1751 		adreno_gpu->has_ray_tracing = true;
1752 	}
1753 
1754 	return 0;
1755 }
1756 
1757 
1758 #define GBIF_CLIENT_HALT_MASK		BIT(0)
1759 #define GBIF_ARB_HALT_MASK		BIT(1)
1760 #define VBIF_XIN_HALT_CTRL0_MASK	GENMASK(3, 0)
1761 #define VBIF_RESET_ACK_MASK		0xF0
1762 #define GPR0_GBIF_HALT_REQUEST		0x1E0
1763 
1764 void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu, bool gx_off)
1765 {
1766 	struct msm_gpu *gpu = &adreno_gpu->base;
1767 
1768 	if (adreno_is_a619_holi(adreno_gpu)) {
1769 		gpu_write(gpu, REG_A6XX_RBBM_GPR0_CNTL, GPR0_GBIF_HALT_REQUEST);
1770 		spin_until((gpu_read(gpu, REG_A6XX_RBBM_VBIF_GX_RESET_STATUS) &
1771 				(VBIF_RESET_ACK_MASK)) == VBIF_RESET_ACK_MASK);
1772 	} else if (!a6xx_has_gbif(adreno_gpu)) {
1773 		gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, VBIF_XIN_HALT_CTRL0_MASK);
1774 		spin_until((gpu_read(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL1) &
1775 				(VBIF_XIN_HALT_CTRL0_MASK)) == VBIF_XIN_HALT_CTRL0_MASK);
1776 		gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, 0);
1777 
1778 		return;
1779 	}
1780 
1781 	if (gx_off) {
1782 		/* Halt the gx side of GBIF */
1783 		gpu_write(gpu, REG_A6XX_RBBM_GBIF_HALT, 1);
1784 		spin_until(gpu_read(gpu, REG_A6XX_RBBM_GBIF_HALT_ACK) & 1);
1785 	}
1786 
1787 	/* Halt new client requests on GBIF */
1788 	gpu_write(gpu, REG_A6XX_GBIF_HALT, GBIF_CLIENT_HALT_MASK);
1789 	spin_until((gpu_read(gpu, REG_A6XX_GBIF_HALT_ACK) &
1790 			(GBIF_CLIENT_HALT_MASK)) == GBIF_CLIENT_HALT_MASK);
1791 
1792 	/* Halt all AXI requests on GBIF */
1793 	gpu_write(gpu, REG_A6XX_GBIF_HALT, GBIF_ARB_HALT_MASK);
1794 	spin_until((gpu_read(gpu,  REG_A6XX_GBIF_HALT_ACK) &
1795 			(GBIF_ARB_HALT_MASK)) == GBIF_ARB_HALT_MASK);
1796 
1797 	/* The GBIF halt needs to be explicitly cleared */
1798 	gpu_write(gpu, REG_A6XX_GBIF_HALT, 0x0);
1799 }
1800 
1801 void a6xx_gpu_sw_reset(struct msm_gpu *gpu, bool assert)
1802 {
1803 	/* 11nm chips (e.g. ones with A610) have hw issues with the reset line! */
1804 	if (adreno_is_a610(to_adreno_gpu(gpu)))
1805 		return;
1806 
1807 	gpu_write(gpu, REG_A6XX_RBBM_SW_RESET_CMD, assert);
1808 	/* Perform a bogus read and add a brief delay to ensure ordering. */
1809 	gpu_read(gpu, REG_A6XX_RBBM_SW_RESET_CMD);
1810 	udelay(1);
1811 
1812 	/* The reset line needs to be asserted for at least 100 us */
1813 	if (assert)
1814 		udelay(100);
1815 }
1816 
1817 static int a6xx_gmu_pm_resume(struct msm_gpu *gpu)
1818 {
1819 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1820 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1821 	int ret;
1822 
1823 	gpu->needs_hw_init = true;
1824 
1825 	trace_msm_gpu_resume(0);
1826 
1827 	mutex_lock(&a6xx_gpu->gmu.lock);
1828 	ret = a6xx_gmu_resume(a6xx_gpu);
1829 	mutex_unlock(&a6xx_gpu->gmu.lock);
1830 	if (ret)
1831 		return ret;
1832 
1833 	msm_devfreq_resume(gpu);
1834 
1835 	adreno_is_a7xx(adreno_gpu) ? a7xx_llc_activate(a6xx_gpu) : a6xx_llc_activate(a6xx_gpu);
1836 
1837 	return ret;
1838 }
1839 
1840 static int a6xx_pm_resume(struct msm_gpu *gpu)
1841 {
1842 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1843 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1844 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1845 	unsigned long freq = gpu->fast_rate;
1846 	struct dev_pm_opp *opp;
1847 	int ret;
1848 
1849 	gpu->needs_hw_init = true;
1850 
1851 	trace_msm_gpu_resume(0);
1852 
1853 	mutex_lock(&a6xx_gpu->gmu.lock);
1854 
1855 	opp = dev_pm_opp_find_freq_ceil(&gpu->pdev->dev, &freq);
1856 	if (IS_ERR(opp)) {
1857 		ret = PTR_ERR(opp);
1858 		goto err_set_opp;
1859 	}
1860 	dev_pm_opp_put(opp);
1861 
1862 	/* Set the core clock and bus bw, having VDD scaling in mind */
1863 	dev_pm_opp_set_opp(&gpu->pdev->dev, opp);
1864 
1865 	pm_runtime_resume_and_get(gmu->dev);
1866 	pm_runtime_resume_and_get(gmu->gxpd);
1867 
1868 	ret = clk_bulk_prepare_enable(gpu->nr_clocks, gpu->grp_clks);
1869 	if (ret)
1870 		goto err_bulk_clk;
1871 
1872 	if (adreno_is_a619_holi(adreno_gpu))
1873 		a6xx_sptprac_enable(gmu);
1874 
1875 	/* If anything goes south, tear the GPU down piece by piece.. */
1876 	if (ret) {
1877 err_bulk_clk:
1878 		pm_runtime_put(gmu->gxpd);
1879 		pm_runtime_put(gmu->dev);
1880 		dev_pm_opp_set_opp(&gpu->pdev->dev, NULL);
1881 	}
1882 err_set_opp:
1883 	mutex_unlock(&a6xx_gpu->gmu.lock);
1884 
1885 	if (!ret)
1886 		msm_devfreq_resume(gpu);
1887 
1888 	return ret;
1889 }
1890 
1891 static int a6xx_gmu_pm_suspend(struct msm_gpu *gpu)
1892 {
1893 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1894 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1895 	int i, ret;
1896 
1897 	trace_msm_gpu_suspend(0);
1898 
1899 	a6xx_llc_deactivate(a6xx_gpu);
1900 
1901 	msm_devfreq_suspend(gpu);
1902 
1903 	mutex_lock(&a6xx_gpu->gmu.lock);
1904 	ret = a6xx_gmu_stop(a6xx_gpu);
1905 	mutex_unlock(&a6xx_gpu->gmu.lock);
1906 	if (ret)
1907 		return ret;
1908 
1909 	if (a6xx_gpu->shadow_bo)
1910 		for (i = 0; i < gpu->nr_rings; i++)
1911 			a6xx_gpu->shadow[i] = 0;
1912 
1913 	gpu->suspend_count++;
1914 
1915 	return 0;
1916 }
1917 
1918 static int a6xx_pm_suspend(struct msm_gpu *gpu)
1919 {
1920 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1921 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1922 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1923 	int i;
1924 
1925 	trace_msm_gpu_suspend(0);
1926 
1927 	msm_devfreq_suspend(gpu);
1928 
1929 	mutex_lock(&a6xx_gpu->gmu.lock);
1930 
1931 	/* Drain the outstanding traffic on memory buses */
1932 	a6xx_bus_clear_pending_transactions(adreno_gpu, true);
1933 
1934 	if (adreno_is_a619_holi(adreno_gpu))
1935 		a6xx_sptprac_disable(gmu);
1936 
1937 	clk_bulk_disable_unprepare(gpu->nr_clocks, gpu->grp_clks);
1938 
1939 	pm_runtime_put_sync(gmu->gxpd);
1940 	dev_pm_opp_set_opp(&gpu->pdev->dev, NULL);
1941 	pm_runtime_put_sync(gmu->dev);
1942 
1943 	mutex_unlock(&a6xx_gpu->gmu.lock);
1944 
1945 	if (a6xx_gpu->shadow_bo)
1946 		for (i = 0; i < gpu->nr_rings; i++)
1947 			a6xx_gpu->shadow[i] = 0;
1948 
1949 	gpu->suspend_count++;
1950 
1951 	return 0;
1952 }
1953 
1954 static int a6xx_gmu_get_timestamp(struct msm_gpu *gpu, uint64_t *value)
1955 {
1956 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1957 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1958 
1959 	mutex_lock(&a6xx_gpu->gmu.lock);
1960 
1961 	/* Force the GPU power on so we can read this register */
1962 	a6xx_gmu_set_oob(&a6xx_gpu->gmu, GMU_OOB_PERFCOUNTER_SET);
1963 
1964 	*value = gpu_read64(gpu, REG_A6XX_CP_ALWAYS_ON_COUNTER);
1965 
1966 	a6xx_gmu_clear_oob(&a6xx_gpu->gmu, GMU_OOB_PERFCOUNTER_SET);
1967 
1968 	mutex_unlock(&a6xx_gpu->gmu.lock);
1969 
1970 	return 0;
1971 }
1972 
1973 static int a6xx_get_timestamp(struct msm_gpu *gpu, uint64_t *value)
1974 {
1975 	*value = gpu_read64(gpu, REG_A6XX_CP_ALWAYS_ON_COUNTER);
1976 	return 0;
1977 }
1978 
1979 static struct msm_ringbuffer *a6xx_active_ring(struct msm_gpu *gpu)
1980 {
1981 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1982 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1983 
1984 	return a6xx_gpu->cur_ring;
1985 }
1986 
1987 static void a6xx_destroy(struct msm_gpu *gpu)
1988 {
1989 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1990 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1991 
1992 	if (a6xx_gpu->sqe_bo) {
1993 		msm_gem_unpin_iova(a6xx_gpu->sqe_bo, gpu->aspace);
1994 		drm_gem_object_put(a6xx_gpu->sqe_bo);
1995 	}
1996 
1997 	if (a6xx_gpu->shadow_bo) {
1998 		msm_gem_unpin_iova(a6xx_gpu->shadow_bo, gpu->aspace);
1999 		drm_gem_object_put(a6xx_gpu->shadow_bo);
2000 	}
2001 
2002 	a6xx_llc_slices_destroy(a6xx_gpu);
2003 
2004 	a6xx_gmu_remove(a6xx_gpu);
2005 
2006 	adreno_gpu_cleanup(adreno_gpu);
2007 
2008 	kfree(a6xx_gpu);
2009 }
2010 
2011 static u64 a6xx_gpu_busy(struct msm_gpu *gpu, unsigned long *out_sample_rate)
2012 {
2013 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
2014 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
2015 	u64 busy_cycles;
2016 
2017 	/* 19.2MHz */
2018 	*out_sample_rate = 19200000;
2019 
2020 	busy_cycles = gmu_read64(&a6xx_gpu->gmu,
2021 			REG_A6XX_GMU_CX_GMU_POWER_COUNTER_XOCLK_0_L,
2022 			REG_A6XX_GMU_CX_GMU_POWER_COUNTER_XOCLK_0_H);
2023 
2024 	return busy_cycles;
2025 }
2026 
2027 static void a6xx_gpu_set_freq(struct msm_gpu *gpu, struct dev_pm_opp *opp,
2028 			      bool suspended)
2029 {
2030 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
2031 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
2032 
2033 	mutex_lock(&a6xx_gpu->gmu.lock);
2034 	a6xx_gmu_set_freq(gpu, opp, suspended);
2035 	mutex_unlock(&a6xx_gpu->gmu.lock);
2036 }
2037 
2038 static struct msm_gem_address_space *
2039 a6xx_create_address_space(struct msm_gpu *gpu, struct platform_device *pdev)
2040 {
2041 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
2042 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
2043 	unsigned long quirks = 0;
2044 
2045 	/*
2046 	 * This allows GPU to set the bus attributes required to use system
2047 	 * cache on behalf of the iommu page table walker.
2048 	 */
2049 	if (!IS_ERR_OR_NULL(a6xx_gpu->htw_llc_slice) &&
2050 	    !device_iommu_capable(&pdev->dev, IOMMU_CAP_CACHE_COHERENCY))
2051 		quirks |= IO_PGTABLE_QUIRK_ARM_OUTER_WBWA;
2052 
2053 	return adreno_iommu_create_address_space(gpu, pdev, quirks);
2054 }
2055 
2056 static struct msm_gem_address_space *
2057 a6xx_create_private_address_space(struct msm_gpu *gpu)
2058 {
2059 	struct msm_mmu *mmu;
2060 
2061 	mmu = msm_iommu_pagetable_create(gpu->aspace->mmu);
2062 
2063 	if (IS_ERR(mmu))
2064 		return ERR_CAST(mmu);
2065 
2066 	return msm_gem_address_space_create(mmu,
2067 		"gpu", 0x100000000ULL,
2068 		adreno_private_address_space_size(gpu));
2069 }
2070 
2071 static uint32_t a6xx_get_rptr(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
2072 {
2073 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
2074 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
2075 
2076 	if (adreno_gpu->base.hw_apriv || a6xx_gpu->has_whereami)
2077 		return a6xx_gpu->shadow[ring->id];
2078 
2079 	return ring->memptrs->rptr = gpu_read(gpu, REG_A6XX_CP_RB_RPTR);
2080 }
2081 
2082 static bool a6xx_progress(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
2083 {
2084 	struct msm_cp_state cp_state = {
2085 		.ib1_base = gpu_read64(gpu, REG_A6XX_CP_IB1_BASE),
2086 		.ib2_base = gpu_read64(gpu, REG_A6XX_CP_IB2_BASE),
2087 		.ib1_rem  = gpu_read(gpu, REG_A6XX_CP_IB1_REM_SIZE),
2088 		.ib2_rem  = gpu_read(gpu, REG_A6XX_CP_IB2_REM_SIZE),
2089 	};
2090 	bool progress;
2091 
2092 	/*
2093 	 * Adjust the remaining data to account for what has already been
2094 	 * fetched from memory, but not yet consumed by the SQE.
2095 	 *
2096 	 * This is not *technically* correct, the amount buffered could
2097 	 * exceed the IB size due to hw prefetching ahead, but:
2098 	 *
2099 	 * (1) We aren't trying to find the exact position, just whether
2100 	 *     progress has been made
2101 	 * (2) The CP_REG_TO_MEM at the end of a submit should be enough
2102 	 *     to prevent prefetching into an unrelated submit.  (And
2103 	 *     either way, at some point the ROQ will be full.)
2104 	 */
2105 	cp_state.ib1_rem += gpu_read(gpu, REG_A6XX_CP_ROQ_AVAIL_IB1) >> 16;
2106 	cp_state.ib2_rem += gpu_read(gpu, REG_A6XX_CP_ROQ_AVAIL_IB2) >> 16;
2107 
2108 	progress = !!memcmp(&cp_state, &ring->last_cp_state, sizeof(cp_state));
2109 
2110 	ring->last_cp_state = cp_state;
2111 
2112 	return progress;
2113 }
2114 
2115 static u32 fuse_to_supp_hw(const struct adreno_info *info, u32 fuse)
2116 {
2117 	if (!info->speedbins)
2118 		return UINT_MAX;
2119 
2120 	for (int i = 0; info->speedbins[i].fuse != SHRT_MAX; i++)
2121 		if (info->speedbins[i].fuse == fuse)
2122 			return BIT(info->speedbins[i].speedbin);
2123 
2124 	return UINT_MAX;
2125 }
2126 
2127 static int a6xx_set_supported_hw(struct device *dev, const struct adreno_info *info)
2128 {
2129 	u32 supp_hw;
2130 	u32 speedbin;
2131 	int ret;
2132 
2133 	ret = adreno_read_speedbin(dev, &speedbin);
2134 	/*
2135 	 * -ENOENT means that the platform doesn't support speedbin which is
2136 	 * fine
2137 	 */
2138 	if (ret == -ENOENT) {
2139 		return 0;
2140 	} else if (ret) {
2141 		dev_err_probe(dev, ret,
2142 			      "failed to read speed-bin. Some OPPs may not be supported by hardware\n");
2143 		return ret;
2144 	}
2145 
2146 	supp_hw = fuse_to_supp_hw(info, speedbin);
2147 
2148 	if (supp_hw == UINT_MAX) {
2149 		DRM_DEV_ERROR(dev,
2150 			"missing support for speed-bin: %u. Some OPPs may not be supported by hardware\n",
2151 			speedbin);
2152 		supp_hw = BIT(0); /* Default */
2153 	}
2154 
2155 	ret = devm_pm_opp_set_supported_hw(dev, &supp_hw, 1);
2156 	if (ret)
2157 		return ret;
2158 
2159 	return 0;
2160 }
2161 
2162 static const struct adreno_gpu_funcs funcs = {
2163 	.base = {
2164 		.get_param = adreno_get_param,
2165 		.set_param = adreno_set_param,
2166 		.hw_init = a6xx_hw_init,
2167 		.ucode_load = a6xx_ucode_load,
2168 		.pm_suspend = a6xx_gmu_pm_suspend,
2169 		.pm_resume = a6xx_gmu_pm_resume,
2170 		.recover = a6xx_recover,
2171 		.submit = a6xx_submit,
2172 		.active_ring = a6xx_active_ring,
2173 		.irq = a6xx_irq,
2174 		.destroy = a6xx_destroy,
2175 #if defined(CONFIG_DRM_MSM_GPU_STATE)
2176 		.show = a6xx_show,
2177 #endif
2178 		.gpu_busy = a6xx_gpu_busy,
2179 		.gpu_get_freq = a6xx_gmu_get_freq,
2180 		.gpu_set_freq = a6xx_gpu_set_freq,
2181 #if defined(CONFIG_DRM_MSM_GPU_STATE)
2182 		.gpu_state_get = a6xx_gpu_state_get,
2183 		.gpu_state_put = a6xx_gpu_state_put,
2184 #endif
2185 		.create_address_space = a6xx_create_address_space,
2186 		.create_private_address_space = a6xx_create_private_address_space,
2187 		.get_rptr = a6xx_get_rptr,
2188 		.progress = a6xx_progress,
2189 	},
2190 	.get_timestamp = a6xx_gmu_get_timestamp,
2191 };
2192 
2193 static const struct adreno_gpu_funcs funcs_gmuwrapper = {
2194 	.base = {
2195 		.get_param = adreno_get_param,
2196 		.set_param = adreno_set_param,
2197 		.hw_init = a6xx_hw_init,
2198 		.ucode_load = a6xx_ucode_load,
2199 		.pm_suspend = a6xx_pm_suspend,
2200 		.pm_resume = a6xx_pm_resume,
2201 		.recover = a6xx_recover,
2202 		.submit = a6xx_submit,
2203 		.active_ring = a6xx_active_ring,
2204 		.irq = a6xx_irq,
2205 		.destroy = a6xx_destroy,
2206 #if defined(CONFIG_DRM_MSM_GPU_STATE)
2207 		.show = a6xx_show,
2208 #endif
2209 		.gpu_busy = a6xx_gpu_busy,
2210 #if defined(CONFIG_DRM_MSM_GPU_STATE)
2211 		.gpu_state_get = a6xx_gpu_state_get,
2212 		.gpu_state_put = a6xx_gpu_state_put,
2213 #endif
2214 		.create_address_space = a6xx_create_address_space,
2215 		.create_private_address_space = a6xx_create_private_address_space,
2216 		.get_rptr = a6xx_get_rptr,
2217 		.progress = a6xx_progress,
2218 	},
2219 	.get_timestamp = a6xx_get_timestamp,
2220 };
2221 
2222 static const struct adreno_gpu_funcs funcs_a7xx = {
2223 	.base = {
2224 		.get_param = adreno_get_param,
2225 		.set_param = adreno_set_param,
2226 		.hw_init = a6xx_hw_init,
2227 		.ucode_load = a6xx_ucode_load,
2228 		.pm_suspend = a6xx_gmu_pm_suspend,
2229 		.pm_resume = a6xx_gmu_pm_resume,
2230 		.recover = a6xx_recover,
2231 		.submit = a7xx_submit,
2232 		.active_ring = a6xx_active_ring,
2233 		.irq = a6xx_irq,
2234 		.destroy = a6xx_destroy,
2235 #if defined(CONFIG_DRM_MSM_GPU_STATE)
2236 		.show = a6xx_show,
2237 #endif
2238 		.gpu_busy = a6xx_gpu_busy,
2239 		.gpu_get_freq = a6xx_gmu_get_freq,
2240 		.gpu_set_freq = a6xx_gpu_set_freq,
2241 #if defined(CONFIG_DRM_MSM_GPU_STATE)
2242 		.gpu_state_get = a6xx_gpu_state_get,
2243 		.gpu_state_put = a6xx_gpu_state_put,
2244 #endif
2245 		.create_address_space = a6xx_create_address_space,
2246 		.create_private_address_space = a6xx_create_private_address_space,
2247 		.get_rptr = a6xx_get_rptr,
2248 		.progress = a6xx_progress,
2249 	},
2250 	.get_timestamp = a6xx_gmu_get_timestamp,
2251 };
2252 
2253 struct msm_gpu *a6xx_gpu_init(struct drm_device *dev)
2254 {
2255 	struct msm_drm_private *priv = dev->dev_private;
2256 	struct platform_device *pdev = priv->gpu_pdev;
2257 	struct adreno_platform_config *config = pdev->dev.platform_data;
2258 	struct device_node *node;
2259 	struct a6xx_gpu *a6xx_gpu;
2260 	struct adreno_gpu *adreno_gpu;
2261 	struct msm_gpu *gpu;
2262 	bool is_a7xx;
2263 	int ret;
2264 
2265 	a6xx_gpu = kzalloc(sizeof(*a6xx_gpu), GFP_KERNEL);
2266 	if (!a6xx_gpu)
2267 		return ERR_PTR(-ENOMEM);
2268 
2269 	adreno_gpu = &a6xx_gpu->base;
2270 	gpu = &adreno_gpu->base;
2271 
2272 	mutex_init(&a6xx_gpu->gmu.lock);
2273 
2274 	adreno_gpu->registers = NULL;
2275 
2276 	/* Check if there is a GMU phandle and set it up */
2277 	node = of_parse_phandle(pdev->dev.of_node, "qcom,gmu", 0);
2278 	/* FIXME: How do we gracefully handle this? */
2279 	BUG_ON(!node);
2280 
2281 	adreno_gpu->gmu_is_wrapper = of_device_is_compatible(node, "qcom,adreno-gmu-wrapper");
2282 
2283 	adreno_gpu->base.hw_apriv =
2284 		!!(config->info->quirks & ADRENO_QUIRK_HAS_HW_APRIV);
2285 
2286 	/* gpu->info only gets assigned in adreno_gpu_init() */
2287 	is_a7xx = config->info->family == ADRENO_7XX_GEN1 ||
2288 		  config->info->family == ADRENO_7XX_GEN2 ||
2289 		  config->info->family == ADRENO_7XX_GEN3;
2290 
2291 	a6xx_llc_slices_init(pdev, a6xx_gpu, is_a7xx);
2292 
2293 	ret = a6xx_set_supported_hw(&pdev->dev, config->info);
2294 	if (ret) {
2295 		a6xx_llc_slices_destroy(a6xx_gpu);
2296 		kfree(a6xx_gpu);
2297 		return ERR_PTR(ret);
2298 	}
2299 
2300 	if (is_a7xx)
2301 		ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs_a7xx, 1);
2302 	else if (adreno_has_gmu_wrapper(adreno_gpu))
2303 		ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs_gmuwrapper, 1);
2304 	else
2305 		ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs, 1);
2306 	if (ret) {
2307 		a6xx_destroy(&(a6xx_gpu->base.base));
2308 		return ERR_PTR(ret);
2309 	}
2310 
2311 	/*
2312 	 * For now only clamp to idle freq for devices where this is known not
2313 	 * to cause power supply issues:
2314 	 */
2315 	if (adreno_is_a618(adreno_gpu) || adreno_is_7c3(adreno_gpu))
2316 		priv->gpu_clamp_to_idle = true;
2317 
2318 	if (adreno_has_gmu_wrapper(adreno_gpu))
2319 		ret = a6xx_gmu_wrapper_init(a6xx_gpu, node);
2320 	else
2321 		ret = a6xx_gmu_init(a6xx_gpu, node);
2322 	of_node_put(node);
2323 	if (ret) {
2324 		a6xx_destroy(&(a6xx_gpu->base.base));
2325 		return ERR_PTR(ret);
2326 	}
2327 
2328 	if (adreno_is_a7xx(adreno_gpu)) {
2329 		ret = a7xx_cx_mem_init(a6xx_gpu);
2330 		if (ret) {
2331 			a6xx_destroy(&(a6xx_gpu->base.base));
2332 			return ERR_PTR(ret);
2333 		}
2334 	}
2335 
2336 	if (gpu->aspace)
2337 		msm_mmu_set_fault_handler(gpu->aspace->mmu, gpu,
2338 				a6xx_fault_handler);
2339 
2340 	a6xx_calc_ubwc_config(adreno_gpu);
2341 
2342 	return gpu;
2343 }
2344