xref: /openbmc/linux/drivers/gpu/drm/msm/adreno/a5xx_gpu.c (revision 217188d9)
1 /* Copyright (c) 2016-2017 The Linux Foundation. All rights reserved.
2  *
3  * This program is free software; you can redistribute it and/or modify
4  * it under the terms of the GNU General Public License version 2 and
5  * only version 2 as published by the Free Software Foundation.
6  *
7  * This program is distributed in the hope that it will be useful,
8  * but WITHOUT ANY WARRANTY; without even the implied warranty of
9  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
10  * GNU General Public License for more details.
11  *
12  */
13 
14 #include <linux/kernel.h>
15 #include <linux/types.h>
16 #include <linux/cpumask.h>
17 #include <linux/qcom_scm.h>
18 #include <linux/dma-mapping.h>
19 #include <linux/of_address.h>
20 #include <linux/soc/qcom/mdt_loader.h>
21 #include <linux/pm_opp.h>
22 #include <linux/nvmem-consumer.h>
23 #include <linux/iopoll.h>
24 #include <linux/slab.h>
25 #include "msm_gem.h"
26 #include "msm_mmu.h"
27 #include "a5xx_gpu.h"
28 
29 extern bool hang_debug;
30 static void a5xx_dump(struct msm_gpu *gpu);
31 
32 #define GPU_PAS_ID 13
33 
34 static int zap_shader_load_mdt(struct msm_gpu *gpu, const char *fwname)
35 {
36 	struct device *dev = &gpu->pdev->dev;
37 	const struct firmware *fw;
38 	struct device_node *np;
39 	struct resource r;
40 	phys_addr_t mem_phys;
41 	ssize_t mem_size;
42 	void *mem_region = NULL;
43 	int ret;
44 
45 	if (!IS_ENABLED(CONFIG_ARCH_QCOM))
46 		return -EINVAL;
47 
48 	np = of_get_child_by_name(dev->of_node, "zap-shader");
49 	if (!np)
50 		return -ENODEV;
51 
52 	np = of_parse_phandle(np, "memory-region", 0);
53 	if (!np)
54 		return -EINVAL;
55 
56 	ret = of_address_to_resource(np, 0, &r);
57 	if (ret)
58 		return ret;
59 
60 	mem_phys = r.start;
61 	mem_size = resource_size(&r);
62 
63 	/* Request the MDT file for the firmware */
64 	fw = adreno_request_fw(to_adreno_gpu(gpu), fwname);
65 	if (IS_ERR(fw)) {
66 		DRM_DEV_ERROR(dev, "Unable to load %s\n", fwname);
67 		return PTR_ERR(fw);
68 	}
69 
70 	/* Figure out how much memory we need */
71 	mem_size = qcom_mdt_get_size(fw);
72 	if (mem_size < 0) {
73 		ret = mem_size;
74 		goto out;
75 	}
76 
77 	/* Allocate memory for the firmware image */
78 	mem_region = memremap(mem_phys, mem_size,  MEMREMAP_WC);
79 	if (!mem_region) {
80 		ret = -ENOMEM;
81 		goto out;
82 	}
83 
84 	/*
85 	 * Load the rest of the MDT
86 	 *
87 	 * Note that we could be dealing with two different paths, since
88 	 * with upstream linux-firmware it would be in a qcom/ subdir..
89 	 * adreno_request_fw() handles this, but qcom_mdt_load() does
90 	 * not.  But since we've already gotten thru adreno_request_fw()
91 	 * we know which of the two cases it is:
92 	 */
93 	if (to_adreno_gpu(gpu)->fwloc == FW_LOCATION_LEGACY) {
94 		ret = qcom_mdt_load(dev, fw, fwname, GPU_PAS_ID,
95 				mem_region, mem_phys, mem_size, NULL);
96 	} else {
97 		char *newname;
98 
99 		newname = kasprintf(GFP_KERNEL, "qcom/%s", fwname);
100 
101 		ret = qcom_mdt_load(dev, fw, newname, GPU_PAS_ID,
102 				mem_region, mem_phys, mem_size, NULL);
103 		kfree(newname);
104 	}
105 	if (ret)
106 		goto out;
107 
108 	/* Send the image to the secure world */
109 	ret = qcom_scm_pas_auth_and_reset(GPU_PAS_ID);
110 	if (ret)
111 		DRM_DEV_ERROR(dev, "Unable to authorize the image\n");
112 
113 out:
114 	if (mem_region)
115 		memunmap(mem_region);
116 
117 	release_firmware(fw);
118 
119 	return ret;
120 }
121 
122 static void a5xx_flush(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
123 {
124 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
125 	struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu);
126 	uint32_t wptr;
127 	unsigned long flags;
128 
129 	spin_lock_irqsave(&ring->lock, flags);
130 
131 	/* Copy the shadow to the actual register */
132 	ring->cur = ring->next;
133 
134 	/* Make sure to wrap wptr if we need to */
135 	wptr = get_wptr(ring);
136 
137 	spin_unlock_irqrestore(&ring->lock, flags);
138 
139 	/* Make sure everything is posted before making a decision */
140 	mb();
141 
142 	/* Update HW if this is the current ring and we are not in preempt */
143 	if (a5xx_gpu->cur_ring == ring && !a5xx_in_preempt(a5xx_gpu))
144 		gpu_write(gpu, REG_A5XX_CP_RB_WPTR, wptr);
145 }
146 
147 static void a5xx_submit_in_rb(struct msm_gpu *gpu, struct msm_gem_submit *submit,
148 	struct msm_file_private *ctx)
149 {
150 	struct msm_drm_private *priv = gpu->dev->dev_private;
151 	struct msm_ringbuffer *ring = submit->ring;
152 	struct msm_gem_object *obj;
153 	uint32_t *ptr, dwords;
154 	unsigned int i;
155 
156 	for (i = 0; i < submit->nr_cmds; i++) {
157 		switch (submit->cmd[i].type) {
158 		case MSM_SUBMIT_CMD_IB_TARGET_BUF:
159 			break;
160 		case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
161 			if (priv->lastctx == ctx)
162 				break;
163 		case MSM_SUBMIT_CMD_BUF:
164 			/* copy commands into RB: */
165 			obj = submit->bos[submit->cmd[i].idx].obj;
166 			dwords = submit->cmd[i].size;
167 
168 			ptr = msm_gem_get_vaddr(&obj->base);
169 
170 			/* _get_vaddr() shouldn't fail at this point,
171 			 * since we've already mapped it once in
172 			 * submit_reloc()
173 			 */
174 			if (WARN_ON(!ptr))
175 				return;
176 
177 			for (i = 0; i < dwords; i++) {
178 				/* normally the OUT_PKTn() would wait
179 				 * for space for the packet.  But since
180 				 * we just OUT_RING() the whole thing,
181 				 * need to call adreno_wait_ring()
182 				 * ourself:
183 				 */
184 				adreno_wait_ring(ring, 1);
185 				OUT_RING(ring, ptr[i]);
186 			}
187 
188 			msm_gem_put_vaddr(&obj->base);
189 
190 			break;
191 		}
192 	}
193 
194 	a5xx_flush(gpu, ring);
195 	a5xx_preempt_trigger(gpu);
196 
197 	/* we might not necessarily have a cmd from userspace to
198 	 * trigger an event to know that submit has completed, so
199 	 * do this manually:
200 	 */
201 	a5xx_idle(gpu, ring);
202 	ring->memptrs->fence = submit->seqno;
203 	msm_gpu_retire(gpu);
204 }
205 
206 static void a5xx_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit,
207 	struct msm_file_private *ctx)
208 {
209 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
210 	struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu);
211 	struct msm_drm_private *priv = gpu->dev->dev_private;
212 	struct msm_ringbuffer *ring = submit->ring;
213 	unsigned int i, ibs = 0;
214 
215 	if (IS_ENABLED(CONFIG_DRM_MSM_GPU_SUDO) && submit->in_rb) {
216 		priv->lastctx = NULL;
217 		a5xx_submit_in_rb(gpu, submit, ctx);
218 		return;
219 	}
220 
221 	OUT_PKT7(ring, CP_PREEMPT_ENABLE_GLOBAL, 1);
222 	OUT_RING(ring, 0x02);
223 
224 	/* Turn off protected mode to write to special registers */
225 	OUT_PKT7(ring, CP_SET_PROTECTED_MODE, 1);
226 	OUT_RING(ring, 0);
227 
228 	/* Set the save preemption record for the ring/command */
229 	OUT_PKT4(ring, REG_A5XX_CP_CONTEXT_SWITCH_SAVE_ADDR_LO, 2);
230 	OUT_RING(ring, lower_32_bits(a5xx_gpu->preempt_iova[submit->ring->id]));
231 	OUT_RING(ring, upper_32_bits(a5xx_gpu->preempt_iova[submit->ring->id]));
232 
233 	/* Turn back on protected mode */
234 	OUT_PKT7(ring, CP_SET_PROTECTED_MODE, 1);
235 	OUT_RING(ring, 1);
236 
237 	/* Enable local preemption for finegrain preemption */
238 	OUT_PKT7(ring, CP_PREEMPT_ENABLE_GLOBAL, 1);
239 	OUT_RING(ring, 0x02);
240 
241 	/* Allow CP_CONTEXT_SWITCH_YIELD packets in the IB2 */
242 	OUT_PKT7(ring, CP_YIELD_ENABLE, 1);
243 	OUT_RING(ring, 0x02);
244 
245 	/* Submit the commands */
246 	for (i = 0; i < submit->nr_cmds; i++) {
247 		switch (submit->cmd[i].type) {
248 		case MSM_SUBMIT_CMD_IB_TARGET_BUF:
249 			break;
250 		case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
251 			if (priv->lastctx == ctx)
252 				break;
253 		case MSM_SUBMIT_CMD_BUF:
254 			OUT_PKT7(ring, CP_INDIRECT_BUFFER_PFE, 3);
255 			OUT_RING(ring, lower_32_bits(submit->cmd[i].iova));
256 			OUT_RING(ring, upper_32_bits(submit->cmd[i].iova));
257 			OUT_RING(ring, submit->cmd[i].size);
258 			ibs++;
259 			break;
260 		}
261 	}
262 
263 	/*
264 	 * Write the render mode to NULL (0) to indicate to the CP that the IBs
265 	 * are done rendering - otherwise a lucky preemption would start
266 	 * replaying from the last checkpoint
267 	 */
268 	OUT_PKT7(ring, CP_SET_RENDER_MODE, 5);
269 	OUT_RING(ring, 0);
270 	OUT_RING(ring, 0);
271 	OUT_RING(ring, 0);
272 	OUT_RING(ring, 0);
273 	OUT_RING(ring, 0);
274 
275 	/* Turn off IB level preemptions */
276 	OUT_PKT7(ring, CP_YIELD_ENABLE, 1);
277 	OUT_RING(ring, 0x01);
278 
279 	/* Write the fence to the scratch register */
280 	OUT_PKT4(ring, REG_A5XX_CP_SCRATCH_REG(2), 1);
281 	OUT_RING(ring, submit->seqno);
282 
283 	/*
284 	 * Execute a CACHE_FLUSH_TS event. This will ensure that the
285 	 * timestamp is written to the memory and then triggers the interrupt
286 	 */
287 	OUT_PKT7(ring, CP_EVENT_WRITE, 4);
288 	OUT_RING(ring, CACHE_FLUSH_TS | (1 << 31));
289 	OUT_RING(ring, lower_32_bits(rbmemptr(ring, fence)));
290 	OUT_RING(ring, upper_32_bits(rbmemptr(ring, fence)));
291 	OUT_RING(ring, submit->seqno);
292 
293 	/* Yield the floor on command completion */
294 	OUT_PKT7(ring, CP_CONTEXT_SWITCH_YIELD, 4);
295 	/*
296 	 * If dword[2:1] are non zero, they specify an address for the CP to
297 	 * write the value of dword[3] to on preemption complete. Write 0 to
298 	 * skip the write
299 	 */
300 	OUT_RING(ring, 0x00);
301 	OUT_RING(ring, 0x00);
302 	/* Data value - not used if the address above is 0 */
303 	OUT_RING(ring, 0x01);
304 	/* Set bit 0 to trigger an interrupt on preempt complete */
305 	OUT_RING(ring, 0x01);
306 
307 	a5xx_flush(gpu, ring);
308 
309 	/* Check to see if we need to start preemption */
310 	a5xx_preempt_trigger(gpu);
311 }
312 
313 static const struct {
314 	u32 offset;
315 	u32 value;
316 } a5xx_hwcg[] = {
317 	{REG_A5XX_RBBM_CLOCK_CNTL_SP0, 0x02222222},
318 	{REG_A5XX_RBBM_CLOCK_CNTL_SP1, 0x02222222},
319 	{REG_A5XX_RBBM_CLOCK_CNTL_SP2, 0x02222222},
320 	{REG_A5XX_RBBM_CLOCK_CNTL_SP3, 0x02222222},
321 	{REG_A5XX_RBBM_CLOCK_CNTL2_SP0, 0x02222220},
322 	{REG_A5XX_RBBM_CLOCK_CNTL2_SP1, 0x02222220},
323 	{REG_A5XX_RBBM_CLOCK_CNTL2_SP2, 0x02222220},
324 	{REG_A5XX_RBBM_CLOCK_CNTL2_SP3, 0x02222220},
325 	{REG_A5XX_RBBM_CLOCK_HYST_SP0, 0x0000F3CF},
326 	{REG_A5XX_RBBM_CLOCK_HYST_SP1, 0x0000F3CF},
327 	{REG_A5XX_RBBM_CLOCK_HYST_SP2, 0x0000F3CF},
328 	{REG_A5XX_RBBM_CLOCK_HYST_SP3, 0x0000F3CF},
329 	{REG_A5XX_RBBM_CLOCK_DELAY_SP0, 0x00000080},
330 	{REG_A5XX_RBBM_CLOCK_DELAY_SP1, 0x00000080},
331 	{REG_A5XX_RBBM_CLOCK_DELAY_SP2, 0x00000080},
332 	{REG_A5XX_RBBM_CLOCK_DELAY_SP3, 0x00000080},
333 	{REG_A5XX_RBBM_CLOCK_CNTL_TP0, 0x22222222},
334 	{REG_A5XX_RBBM_CLOCK_CNTL_TP1, 0x22222222},
335 	{REG_A5XX_RBBM_CLOCK_CNTL_TP2, 0x22222222},
336 	{REG_A5XX_RBBM_CLOCK_CNTL_TP3, 0x22222222},
337 	{REG_A5XX_RBBM_CLOCK_CNTL2_TP0, 0x22222222},
338 	{REG_A5XX_RBBM_CLOCK_CNTL2_TP1, 0x22222222},
339 	{REG_A5XX_RBBM_CLOCK_CNTL2_TP2, 0x22222222},
340 	{REG_A5XX_RBBM_CLOCK_CNTL2_TP3, 0x22222222},
341 	{REG_A5XX_RBBM_CLOCK_CNTL3_TP0, 0x00002222},
342 	{REG_A5XX_RBBM_CLOCK_CNTL3_TP1, 0x00002222},
343 	{REG_A5XX_RBBM_CLOCK_CNTL3_TP2, 0x00002222},
344 	{REG_A5XX_RBBM_CLOCK_CNTL3_TP3, 0x00002222},
345 	{REG_A5XX_RBBM_CLOCK_HYST_TP0, 0x77777777},
346 	{REG_A5XX_RBBM_CLOCK_HYST_TP1, 0x77777777},
347 	{REG_A5XX_RBBM_CLOCK_HYST_TP2, 0x77777777},
348 	{REG_A5XX_RBBM_CLOCK_HYST_TP3, 0x77777777},
349 	{REG_A5XX_RBBM_CLOCK_HYST2_TP0, 0x77777777},
350 	{REG_A5XX_RBBM_CLOCK_HYST2_TP1, 0x77777777},
351 	{REG_A5XX_RBBM_CLOCK_HYST2_TP2, 0x77777777},
352 	{REG_A5XX_RBBM_CLOCK_HYST2_TP3, 0x77777777},
353 	{REG_A5XX_RBBM_CLOCK_HYST3_TP0, 0x00007777},
354 	{REG_A5XX_RBBM_CLOCK_HYST3_TP1, 0x00007777},
355 	{REG_A5XX_RBBM_CLOCK_HYST3_TP2, 0x00007777},
356 	{REG_A5XX_RBBM_CLOCK_HYST3_TP3, 0x00007777},
357 	{REG_A5XX_RBBM_CLOCK_DELAY_TP0, 0x11111111},
358 	{REG_A5XX_RBBM_CLOCK_DELAY_TP1, 0x11111111},
359 	{REG_A5XX_RBBM_CLOCK_DELAY_TP2, 0x11111111},
360 	{REG_A5XX_RBBM_CLOCK_DELAY_TP3, 0x11111111},
361 	{REG_A5XX_RBBM_CLOCK_DELAY2_TP0, 0x11111111},
362 	{REG_A5XX_RBBM_CLOCK_DELAY2_TP1, 0x11111111},
363 	{REG_A5XX_RBBM_CLOCK_DELAY2_TP2, 0x11111111},
364 	{REG_A5XX_RBBM_CLOCK_DELAY2_TP3, 0x11111111},
365 	{REG_A5XX_RBBM_CLOCK_DELAY3_TP0, 0x00001111},
366 	{REG_A5XX_RBBM_CLOCK_DELAY3_TP1, 0x00001111},
367 	{REG_A5XX_RBBM_CLOCK_DELAY3_TP2, 0x00001111},
368 	{REG_A5XX_RBBM_CLOCK_DELAY3_TP3, 0x00001111},
369 	{REG_A5XX_RBBM_CLOCK_CNTL_UCHE, 0x22222222},
370 	{REG_A5XX_RBBM_CLOCK_CNTL2_UCHE, 0x22222222},
371 	{REG_A5XX_RBBM_CLOCK_CNTL3_UCHE, 0x22222222},
372 	{REG_A5XX_RBBM_CLOCK_CNTL4_UCHE, 0x00222222},
373 	{REG_A5XX_RBBM_CLOCK_HYST_UCHE, 0x00444444},
374 	{REG_A5XX_RBBM_CLOCK_DELAY_UCHE, 0x00000002},
375 	{REG_A5XX_RBBM_CLOCK_CNTL_RB0, 0x22222222},
376 	{REG_A5XX_RBBM_CLOCK_CNTL_RB1, 0x22222222},
377 	{REG_A5XX_RBBM_CLOCK_CNTL_RB2, 0x22222222},
378 	{REG_A5XX_RBBM_CLOCK_CNTL_RB3, 0x22222222},
379 	{REG_A5XX_RBBM_CLOCK_CNTL2_RB0, 0x00222222},
380 	{REG_A5XX_RBBM_CLOCK_CNTL2_RB1, 0x00222222},
381 	{REG_A5XX_RBBM_CLOCK_CNTL2_RB2, 0x00222222},
382 	{REG_A5XX_RBBM_CLOCK_CNTL2_RB3, 0x00222222},
383 	{REG_A5XX_RBBM_CLOCK_CNTL_CCU0, 0x00022220},
384 	{REG_A5XX_RBBM_CLOCK_CNTL_CCU1, 0x00022220},
385 	{REG_A5XX_RBBM_CLOCK_CNTL_CCU2, 0x00022220},
386 	{REG_A5XX_RBBM_CLOCK_CNTL_CCU3, 0x00022220},
387 	{REG_A5XX_RBBM_CLOCK_CNTL_RAC, 0x05522222},
388 	{REG_A5XX_RBBM_CLOCK_CNTL2_RAC, 0x00505555},
389 	{REG_A5XX_RBBM_CLOCK_HYST_RB_CCU0, 0x04040404},
390 	{REG_A5XX_RBBM_CLOCK_HYST_RB_CCU1, 0x04040404},
391 	{REG_A5XX_RBBM_CLOCK_HYST_RB_CCU2, 0x04040404},
392 	{REG_A5XX_RBBM_CLOCK_HYST_RB_CCU3, 0x04040404},
393 	{REG_A5XX_RBBM_CLOCK_HYST_RAC, 0x07444044},
394 	{REG_A5XX_RBBM_CLOCK_DELAY_RB_CCU_L1_0, 0x00000002},
395 	{REG_A5XX_RBBM_CLOCK_DELAY_RB_CCU_L1_1, 0x00000002},
396 	{REG_A5XX_RBBM_CLOCK_DELAY_RB_CCU_L1_2, 0x00000002},
397 	{REG_A5XX_RBBM_CLOCK_DELAY_RB_CCU_L1_3, 0x00000002},
398 	{REG_A5XX_RBBM_CLOCK_DELAY_RAC, 0x00010011},
399 	{REG_A5XX_RBBM_CLOCK_CNTL_TSE_RAS_RBBM, 0x04222222},
400 	{REG_A5XX_RBBM_CLOCK_MODE_GPC, 0x02222222},
401 	{REG_A5XX_RBBM_CLOCK_MODE_VFD, 0x00002222},
402 	{REG_A5XX_RBBM_CLOCK_HYST_TSE_RAS_RBBM, 0x00000000},
403 	{REG_A5XX_RBBM_CLOCK_HYST_GPC, 0x04104004},
404 	{REG_A5XX_RBBM_CLOCK_HYST_VFD, 0x00000000},
405 	{REG_A5XX_RBBM_CLOCK_DELAY_HLSQ, 0x00000000},
406 	{REG_A5XX_RBBM_CLOCK_DELAY_TSE_RAS_RBBM, 0x00004000},
407 	{REG_A5XX_RBBM_CLOCK_DELAY_GPC, 0x00000200},
408 	{REG_A5XX_RBBM_CLOCK_DELAY_VFD, 0x00002222}
409 };
410 
411 void a5xx_set_hwcg(struct msm_gpu *gpu, bool state)
412 {
413 	unsigned int i;
414 
415 	for (i = 0; i < ARRAY_SIZE(a5xx_hwcg); i++)
416 		gpu_write(gpu, a5xx_hwcg[i].offset,
417 			state ? a5xx_hwcg[i].value : 0);
418 
419 	gpu_write(gpu, REG_A5XX_RBBM_CLOCK_CNTL, state ? 0xAAA8AA00 : 0);
420 	gpu_write(gpu, REG_A5XX_RBBM_ISDB_CNT, state ? 0x182 : 0x180);
421 }
422 
423 static int a5xx_me_init(struct msm_gpu *gpu)
424 {
425 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
426 	struct msm_ringbuffer *ring = gpu->rb[0];
427 
428 	OUT_PKT7(ring, CP_ME_INIT, 8);
429 
430 	OUT_RING(ring, 0x0000002F);
431 
432 	/* Enable multiple hardware contexts */
433 	OUT_RING(ring, 0x00000003);
434 
435 	/* Enable error detection */
436 	OUT_RING(ring, 0x20000000);
437 
438 	/* Don't enable header dump */
439 	OUT_RING(ring, 0x00000000);
440 	OUT_RING(ring, 0x00000000);
441 
442 	/* Specify workarounds for various microcode issues */
443 	if (adreno_is_a530(adreno_gpu)) {
444 		/* Workaround for token end syncs
445 		 * Force a WFI after every direct-render 3D mode draw and every
446 		 * 2D mode 3 draw
447 		 */
448 		OUT_RING(ring, 0x0000000B);
449 	} else {
450 		/* No workarounds enabled */
451 		OUT_RING(ring, 0x00000000);
452 	}
453 
454 	OUT_RING(ring, 0x00000000);
455 	OUT_RING(ring, 0x00000000);
456 
457 	gpu->funcs->flush(gpu, ring);
458 	return a5xx_idle(gpu, ring) ? 0 : -EINVAL;
459 }
460 
461 static int a5xx_preempt_start(struct msm_gpu *gpu)
462 {
463 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
464 	struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu);
465 	struct msm_ringbuffer *ring = gpu->rb[0];
466 
467 	if (gpu->nr_rings == 1)
468 		return 0;
469 
470 	/* Turn off protected mode to write to special registers */
471 	OUT_PKT7(ring, CP_SET_PROTECTED_MODE, 1);
472 	OUT_RING(ring, 0);
473 
474 	/* Set the save preemption record for the ring/command */
475 	OUT_PKT4(ring, REG_A5XX_CP_CONTEXT_SWITCH_SAVE_ADDR_LO, 2);
476 	OUT_RING(ring, lower_32_bits(a5xx_gpu->preempt_iova[ring->id]));
477 	OUT_RING(ring, upper_32_bits(a5xx_gpu->preempt_iova[ring->id]));
478 
479 	/* Turn back on protected mode */
480 	OUT_PKT7(ring, CP_SET_PROTECTED_MODE, 1);
481 	OUT_RING(ring, 1);
482 
483 	OUT_PKT7(ring, CP_PREEMPT_ENABLE_GLOBAL, 1);
484 	OUT_RING(ring, 0x00);
485 
486 	OUT_PKT7(ring, CP_PREEMPT_ENABLE_LOCAL, 1);
487 	OUT_RING(ring, 0x01);
488 
489 	OUT_PKT7(ring, CP_YIELD_ENABLE, 1);
490 	OUT_RING(ring, 0x01);
491 
492 	/* Yield the floor on command completion */
493 	OUT_PKT7(ring, CP_CONTEXT_SWITCH_YIELD, 4);
494 	OUT_RING(ring, 0x00);
495 	OUT_RING(ring, 0x00);
496 	OUT_RING(ring, 0x01);
497 	OUT_RING(ring, 0x01);
498 
499 	gpu->funcs->flush(gpu, ring);
500 
501 	return a5xx_idle(gpu, ring) ? 0 : -EINVAL;
502 }
503 
504 static int a5xx_ucode_init(struct msm_gpu *gpu)
505 {
506 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
507 	struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu);
508 	int ret;
509 
510 	if (!a5xx_gpu->pm4_bo) {
511 		a5xx_gpu->pm4_bo = adreno_fw_create_bo(gpu,
512 			adreno_gpu->fw[ADRENO_FW_PM4], &a5xx_gpu->pm4_iova);
513 
514 		if (IS_ERR(a5xx_gpu->pm4_bo)) {
515 			ret = PTR_ERR(a5xx_gpu->pm4_bo);
516 			a5xx_gpu->pm4_bo = NULL;
517 			dev_err(gpu->dev->dev, "could not allocate PM4: %d\n",
518 				ret);
519 			return ret;
520 		}
521 	}
522 
523 	if (!a5xx_gpu->pfp_bo) {
524 		a5xx_gpu->pfp_bo = adreno_fw_create_bo(gpu,
525 			adreno_gpu->fw[ADRENO_FW_PFP], &a5xx_gpu->pfp_iova);
526 
527 		if (IS_ERR(a5xx_gpu->pfp_bo)) {
528 			ret = PTR_ERR(a5xx_gpu->pfp_bo);
529 			a5xx_gpu->pfp_bo = NULL;
530 			dev_err(gpu->dev->dev, "could not allocate PFP: %d\n",
531 				ret);
532 			return ret;
533 		}
534 	}
535 
536 	gpu_write64(gpu, REG_A5XX_CP_ME_INSTR_BASE_LO,
537 		REG_A5XX_CP_ME_INSTR_BASE_HI, a5xx_gpu->pm4_iova);
538 
539 	gpu_write64(gpu, REG_A5XX_CP_PFP_INSTR_BASE_LO,
540 		REG_A5XX_CP_PFP_INSTR_BASE_HI, a5xx_gpu->pfp_iova);
541 
542 	return 0;
543 }
544 
545 #define SCM_GPU_ZAP_SHADER_RESUME 0
546 
547 static int a5xx_zap_shader_resume(struct msm_gpu *gpu)
548 {
549 	int ret;
550 
551 	ret = qcom_scm_set_remote_state(SCM_GPU_ZAP_SHADER_RESUME, GPU_PAS_ID);
552 	if (ret)
553 		DRM_ERROR("%s: zap-shader resume failed: %d\n",
554 			gpu->name, ret);
555 
556 	return ret;
557 }
558 
559 static int a5xx_zap_shader_init(struct msm_gpu *gpu)
560 {
561 	static bool loaded;
562 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
563 	struct platform_device *pdev = gpu->pdev;
564 	int ret;
565 
566 	/*
567 	 * If the zap shader is already loaded into memory we just need to kick
568 	 * the remote processor to reinitialize it
569 	 */
570 	if (loaded)
571 		return a5xx_zap_shader_resume(gpu);
572 
573 	/* We need SCM to be able to load the firmware */
574 	if (!qcom_scm_is_available()) {
575 		DRM_DEV_ERROR(&pdev->dev, "SCM is not available\n");
576 		return -EPROBE_DEFER;
577 	}
578 
579 	/* Each GPU has a target specific zap shader firmware name to use */
580 	if (!adreno_gpu->info->zapfw) {
581 		DRM_DEV_ERROR(&pdev->dev,
582 			"Zap shader firmware file not specified for this target\n");
583 		return -ENODEV;
584 	}
585 
586 	ret = zap_shader_load_mdt(gpu, adreno_gpu->info->zapfw);
587 
588 	loaded = !ret;
589 
590 	return ret;
591 }
592 
593 #define A5XX_INT_MASK (A5XX_RBBM_INT_0_MASK_RBBM_AHB_ERROR | \
594 	  A5XX_RBBM_INT_0_MASK_RBBM_TRANSFER_TIMEOUT | \
595 	  A5XX_RBBM_INT_0_MASK_RBBM_ME_MS_TIMEOUT | \
596 	  A5XX_RBBM_INT_0_MASK_RBBM_PFP_MS_TIMEOUT | \
597 	  A5XX_RBBM_INT_0_MASK_RBBM_ETS_MS_TIMEOUT | \
598 	  A5XX_RBBM_INT_0_MASK_RBBM_ATB_ASYNC_OVERFLOW | \
599 	  A5XX_RBBM_INT_0_MASK_CP_HW_ERROR | \
600 	  A5XX_RBBM_INT_0_MASK_MISC_HANG_DETECT | \
601 	  A5XX_RBBM_INT_0_MASK_CP_SW | \
602 	  A5XX_RBBM_INT_0_MASK_CP_CACHE_FLUSH_TS | \
603 	  A5XX_RBBM_INT_0_MASK_UCHE_OOB_ACCESS | \
604 	  A5XX_RBBM_INT_0_MASK_GPMU_VOLTAGE_DROOP)
605 
606 static int a5xx_hw_init(struct msm_gpu *gpu)
607 {
608 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
609 	int ret;
610 
611 	gpu_write(gpu, REG_A5XX_VBIF_ROUND_ROBIN_QOS_ARB, 0x00000003);
612 
613 	/* Make all blocks contribute to the GPU BUSY perf counter */
614 	gpu_write(gpu, REG_A5XX_RBBM_PERFCTR_GPU_BUSY_MASKED, 0xFFFFFFFF);
615 
616 	/* Enable RBBM error reporting bits */
617 	gpu_write(gpu, REG_A5XX_RBBM_AHB_CNTL0, 0x00000001);
618 
619 	if (adreno_gpu->info->quirks & ADRENO_QUIRK_FAULT_DETECT_MASK) {
620 		/*
621 		 * Mask out the activity signals from RB1-3 to avoid false
622 		 * positives
623 		 */
624 
625 		gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL11,
626 			0xF0000000);
627 		gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL12,
628 			0xFFFFFFFF);
629 		gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL13,
630 			0xFFFFFFFF);
631 		gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL14,
632 			0xFFFFFFFF);
633 		gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL15,
634 			0xFFFFFFFF);
635 		gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL16,
636 			0xFFFFFFFF);
637 		gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL17,
638 			0xFFFFFFFF);
639 		gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL18,
640 			0xFFFFFFFF);
641 	}
642 
643 	/* Enable fault detection */
644 	gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_INT_CNTL,
645 		(1 << 30) | 0xFFFF);
646 
647 	/* Turn on performance counters */
648 	gpu_write(gpu, REG_A5XX_RBBM_PERFCTR_CNTL, 0x01);
649 
650 	/* Select CP0 to always count cycles */
651 	gpu_write(gpu, REG_A5XX_CP_PERFCTR_CP_SEL_0, PERF_CP_ALWAYS_COUNT);
652 
653 	/* Select RBBM0 to countable 6 to get the busy status for devfreq */
654 	gpu_write(gpu, REG_A5XX_RBBM_PERFCTR_RBBM_SEL_0, 6);
655 
656 	/* Increase VFD cache access so LRZ and other data gets evicted less */
657 	gpu_write(gpu, REG_A5XX_UCHE_CACHE_WAYS, 0x02);
658 
659 	/* Disable L2 bypass in the UCHE */
660 	gpu_write(gpu, REG_A5XX_UCHE_TRAP_BASE_LO, 0xFFFF0000);
661 	gpu_write(gpu, REG_A5XX_UCHE_TRAP_BASE_HI, 0x0001FFFF);
662 	gpu_write(gpu, REG_A5XX_UCHE_WRITE_THRU_BASE_LO, 0xFFFF0000);
663 	gpu_write(gpu, REG_A5XX_UCHE_WRITE_THRU_BASE_HI, 0x0001FFFF);
664 
665 	/* Set the GMEM VA range (0 to gpu->gmem) */
666 	gpu_write(gpu, REG_A5XX_UCHE_GMEM_RANGE_MIN_LO, 0x00100000);
667 	gpu_write(gpu, REG_A5XX_UCHE_GMEM_RANGE_MIN_HI, 0x00000000);
668 	gpu_write(gpu, REG_A5XX_UCHE_GMEM_RANGE_MAX_LO,
669 		0x00100000 + adreno_gpu->gmem - 1);
670 	gpu_write(gpu, REG_A5XX_UCHE_GMEM_RANGE_MAX_HI, 0x00000000);
671 
672 	gpu_write(gpu, REG_A5XX_CP_MEQ_THRESHOLDS, 0x40);
673 	gpu_write(gpu, REG_A5XX_CP_MERCIU_SIZE, 0x40);
674 	gpu_write(gpu, REG_A5XX_CP_ROQ_THRESHOLDS_2, 0x80000060);
675 	gpu_write(gpu, REG_A5XX_CP_ROQ_THRESHOLDS_1, 0x40201B16);
676 
677 	gpu_write(gpu, REG_A5XX_PC_DBG_ECO_CNTL, (0x400 << 11 | 0x300 << 22));
678 
679 	if (adreno_gpu->info->quirks & ADRENO_QUIRK_TWO_PASS_USE_WFI)
680 		gpu_rmw(gpu, REG_A5XX_PC_DBG_ECO_CNTL, 0, (1 << 8));
681 
682 	gpu_write(gpu, REG_A5XX_PC_DBG_ECO_CNTL, 0xc0200100);
683 
684 	/* Enable USE_RETENTION_FLOPS */
685 	gpu_write(gpu, REG_A5XX_CP_CHICKEN_DBG, 0x02000000);
686 
687 	/* Enable ME/PFP split notification */
688 	gpu_write(gpu, REG_A5XX_RBBM_AHB_CNTL1, 0xA6FFFFFF);
689 
690 	/* Enable HWCG */
691 	a5xx_set_hwcg(gpu, true);
692 
693 	gpu_write(gpu, REG_A5XX_RBBM_AHB_CNTL2, 0x0000003F);
694 
695 	/* Set the highest bank bit */
696 	gpu_write(gpu, REG_A5XX_TPL1_MODE_CNTL, 2 << 7);
697 	gpu_write(gpu, REG_A5XX_RB_MODE_CNTL, 2 << 1);
698 
699 	/* Protect registers from the CP */
700 	gpu_write(gpu, REG_A5XX_CP_PROTECT_CNTL, 0x00000007);
701 
702 	/* RBBM */
703 	gpu_write(gpu, REG_A5XX_CP_PROTECT(0), ADRENO_PROTECT_RW(0x04, 4));
704 	gpu_write(gpu, REG_A5XX_CP_PROTECT(1), ADRENO_PROTECT_RW(0x08, 8));
705 	gpu_write(gpu, REG_A5XX_CP_PROTECT(2), ADRENO_PROTECT_RW(0x10, 16));
706 	gpu_write(gpu, REG_A5XX_CP_PROTECT(3), ADRENO_PROTECT_RW(0x20, 32));
707 	gpu_write(gpu, REG_A5XX_CP_PROTECT(4), ADRENO_PROTECT_RW(0x40, 64));
708 	gpu_write(gpu, REG_A5XX_CP_PROTECT(5), ADRENO_PROTECT_RW(0x80, 64));
709 
710 	/* Content protect */
711 	gpu_write(gpu, REG_A5XX_CP_PROTECT(6),
712 		ADRENO_PROTECT_RW(REG_A5XX_RBBM_SECVID_TSB_TRUSTED_BASE_LO,
713 			16));
714 	gpu_write(gpu, REG_A5XX_CP_PROTECT(7),
715 		ADRENO_PROTECT_RW(REG_A5XX_RBBM_SECVID_TRUST_CNTL, 2));
716 
717 	/* CP */
718 	gpu_write(gpu, REG_A5XX_CP_PROTECT(8), ADRENO_PROTECT_RW(0x800, 64));
719 	gpu_write(gpu, REG_A5XX_CP_PROTECT(9), ADRENO_PROTECT_RW(0x840, 8));
720 	gpu_write(gpu, REG_A5XX_CP_PROTECT(10), ADRENO_PROTECT_RW(0x880, 32));
721 	gpu_write(gpu, REG_A5XX_CP_PROTECT(11), ADRENO_PROTECT_RW(0xAA0, 1));
722 
723 	/* RB */
724 	gpu_write(gpu, REG_A5XX_CP_PROTECT(12), ADRENO_PROTECT_RW(0xCC0, 1));
725 	gpu_write(gpu, REG_A5XX_CP_PROTECT(13), ADRENO_PROTECT_RW(0xCF0, 2));
726 
727 	/* VPC */
728 	gpu_write(gpu, REG_A5XX_CP_PROTECT(14), ADRENO_PROTECT_RW(0xE68, 8));
729 	gpu_write(gpu, REG_A5XX_CP_PROTECT(15), ADRENO_PROTECT_RW(0xE70, 4));
730 
731 	/* UCHE */
732 	gpu_write(gpu, REG_A5XX_CP_PROTECT(16), ADRENO_PROTECT_RW(0xE80, 16));
733 
734 	if (adreno_is_a530(adreno_gpu))
735 		gpu_write(gpu, REG_A5XX_CP_PROTECT(17),
736 			ADRENO_PROTECT_RW(0x10000, 0x8000));
737 
738 	gpu_write(gpu, REG_A5XX_RBBM_SECVID_TSB_CNTL, 0);
739 	/*
740 	 * Disable the trusted memory range - we don't actually supported secure
741 	 * memory rendering at this point in time and we don't want to block off
742 	 * part of the virtual memory space.
743 	 */
744 	gpu_write64(gpu, REG_A5XX_RBBM_SECVID_TSB_TRUSTED_BASE_LO,
745 		REG_A5XX_RBBM_SECVID_TSB_TRUSTED_BASE_HI, 0x00000000);
746 	gpu_write(gpu, REG_A5XX_RBBM_SECVID_TSB_TRUSTED_SIZE, 0x00000000);
747 
748 	ret = adreno_hw_init(gpu);
749 	if (ret)
750 		return ret;
751 
752 	a5xx_preempt_hw_init(gpu);
753 
754 	a5xx_gpmu_ucode_init(gpu);
755 
756 	ret = a5xx_ucode_init(gpu);
757 	if (ret)
758 		return ret;
759 
760 	/* Disable the interrupts through the initial bringup stage */
761 	gpu_write(gpu, REG_A5XX_RBBM_INT_0_MASK, A5XX_INT_MASK);
762 
763 	/* Clear ME_HALT to start the micro engine */
764 	gpu_write(gpu, REG_A5XX_CP_PFP_ME_CNTL, 0);
765 	ret = a5xx_me_init(gpu);
766 	if (ret)
767 		return ret;
768 
769 	ret = a5xx_power_init(gpu);
770 	if (ret)
771 		return ret;
772 
773 	/*
774 	 * Send a pipeline event stat to get misbehaving counters to start
775 	 * ticking correctly
776 	 */
777 	if (adreno_is_a530(adreno_gpu)) {
778 		OUT_PKT7(gpu->rb[0], CP_EVENT_WRITE, 1);
779 		OUT_RING(gpu->rb[0], 0x0F);
780 
781 		gpu->funcs->flush(gpu, gpu->rb[0]);
782 		if (!a5xx_idle(gpu, gpu->rb[0]))
783 			return -EINVAL;
784 	}
785 
786 	/*
787 	 * Try to load a zap shader into the secure world. If successful
788 	 * we can use the CP to switch out of secure mode. If not then we
789 	 * have no resource but to try to switch ourselves out manually. If we
790 	 * guessed wrong then access to the RBBM_SECVID_TRUST_CNTL register will
791 	 * be blocked and a permissions violation will soon follow.
792 	 */
793 	ret = a5xx_zap_shader_init(gpu);
794 	if (!ret) {
795 		OUT_PKT7(gpu->rb[0], CP_SET_SECURE_MODE, 1);
796 		OUT_RING(gpu->rb[0], 0x00000000);
797 
798 		gpu->funcs->flush(gpu, gpu->rb[0]);
799 		if (!a5xx_idle(gpu, gpu->rb[0]))
800 			return -EINVAL;
801 	} else {
802 		/* Print a warning so if we die, we know why */
803 		dev_warn_once(gpu->dev->dev,
804 			"Zap shader not enabled - using SECVID_TRUST_CNTL instead\n");
805 		gpu_write(gpu, REG_A5XX_RBBM_SECVID_TRUST_CNTL, 0x0);
806 	}
807 
808 	/* Last step - yield the ringbuffer */
809 	a5xx_preempt_start(gpu);
810 
811 	return 0;
812 }
813 
814 static void a5xx_recover(struct msm_gpu *gpu)
815 {
816 	int i;
817 
818 	adreno_dump_info(gpu);
819 
820 	for (i = 0; i < 8; i++) {
821 		printk("CP_SCRATCH_REG%d: %u\n", i,
822 			gpu_read(gpu, REG_A5XX_CP_SCRATCH_REG(i)));
823 	}
824 
825 	if (hang_debug)
826 		a5xx_dump(gpu);
827 
828 	gpu_write(gpu, REG_A5XX_RBBM_SW_RESET_CMD, 1);
829 	gpu_read(gpu, REG_A5XX_RBBM_SW_RESET_CMD);
830 	gpu_write(gpu, REG_A5XX_RBBM_SW_RESET_CMD, 0);
831 	adreno_recover(gpu);
832 }
833 
834 static void a5xx_destroy(struct msm_gpu *gpu)
835 {
836 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
837 	struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu);
838 
839 	DBG("%s", gpu->name);
840 
841 	a5xx_preempt_fini(gpu);
842 
843 	if (a5xx_gpu->pm4_bo) {
844 		if (a5xx_gpu->pm4_iova)
845 			msm_gem_put_iova(a5xx_gpu->pm4_bo, gpu->aspace);
846 		drm_gem_object_put_unlocked(a5xx_gpu->pm4_bo);
847 	}
848 
849 	if (a5xx_gpu->pfp_bo) {
850 		if (a5xx_gpu->pfp_iova)
851 			msm_gem_put_iova(a5xx_gpu->pfp_bo, gpu->aspace);
852 		drm_gem_object_put_unlocked(a5xx_gpu->pfp_bo);
853 	}
854 
855 	if (a5xx_gpu->gpmu_bo) {
856 		if (a5xx_gpu->gpmu_iova)
857 			msm_gem_put_iova(a5xx_gpu->gpmu_bo, gpu->aspace);
858 		drm_gem_object_put_unlocked(a5xx_gpu->gpmu_bo);
859 	}
860 
861 	adreno_gpu_cleanup(adreno_gpu);
862 	kfree(a5xx_gpu);
863 }
864 
865 static inline bool _a5xx_check_idle(struct msm_gpu *gpu)
866 {
867 	if (gpu_read(gpu, REG_A5XX_RBBM_STATUS) & ~A5XX_RBBM_STATUS_HI_BUSY)
868 		return false;
869 
870 	/*
871 	 * Nearly every abnormality ends up pausing the GPU and triggering a
872 	 * fault so we can safely just watch for this one interrupt to fire
873 	 */
874 	return !(gpu_read(gpu, REG_A5XX_RBBM_INT_0_STATUS) &
875 		A5XX_RBBM_INT_0_MASK_MISC_HANG_DETECT);
876 }
877 
878 bool a5xx_idle(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
879 {
880 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
881 	struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu);
882 
883 	if (ring != a5xx_gpu->cur_ring) {
884 		WARN(1, "Tried to idle a non-current ringbuffer\n");
885 		return false;
886 	}
887 
888 	/* wait for CP to drain ringbuffer: */
889 	if (!adreno_idle(gpu, ring))
890 		return false;
891 
892 	if (spin_until(_a5xx_check_idle(gpu))) {
893 		DRM_ERROR("%s: %ps: timeout waiting for GPU to idle: status %8.8X irq %8.8X rptr/wptr %d/%d\n",
894 			gpu->name, __builtin_return_address(0),
895 			gpu_read(gpu, REG_A5XX_RBBM_STATUS),
896 			gpu_read(gpu, REG_A5XX_RBBM_INT_0_STATUS),
897 			gpu_read(gpu, REG_A5XX_CP_RB_RPTR),
898 			gpu_read(gpu, REG_A5XX_CP_RB_WPTR));
899 		return false;
900 	}
901 
902 	return true;
903 }
904 
905 static int a5xx_fault_handler(void *arg, unsigned long iova, int flags)
906 {
907 	struct msm_gpu *gpu = arg;
908 	pr_warn_ratelimited("*** gpu fault: iova=%08lx, flags=%d (%u,%u,%u,%u)\n",
909 			iova, flags,
910 			gpu_read(gpu, REG_A5XX_CP_SCRATCH_REG(4)),
911 			gpu_read(gpu, REG_A5XX_CP_SCRATCH_REG(5)),
912 			gpu_read(gpu, REG_A5XX_CP_SCRATCH_REG(6)),
913 			gpu_read(gpu, REG_A5XX_CP_SCRATCH_REG(7)));
914 
915 	return -EFAULT;
916 }
917 
918 static void a5xx_cp_err_irq(struct msm_gpu *gpu)
919 {
920 	u32 status = gpu_read(gpu, REG_A5XX_CP_INTERRUPT_STATUS);
921 
922 	if (status & A5XX_CP_INT_CP_OPCODE_ERROR) {
923 		u32 val;
924 
925 		gpu_write(gpu, REG_A5XX_CP_PFP_STAT_ADDR, 0);
926 
927 		/*
928 		 * REG_A5XX_CP_PFP_STAT_DATA is indexed, and we want index 1 so
929 		 * read it twice
930 		 */
931 
932 		gpu_read(gpu, REG_A5XX_CP_PFP_STAT_DATA);
933 		val = gpu_read(gpu, REG_A5XX_CP_PFP_STAT_DATA);
934 
935 		dev_err_ratelimited(gpu->dev->dev, "CP | opcode error | possible opcode=0x%8.8X\n",
936 			val);
937 	}
938 
939 	if (status & A5XX_CP_INT_CP_HW_FAULT_ERROR)
940 		dev_err_ratelimited(gpu->dev->dev, "CP | HW fault | status=0x%8.8X\n",
941 			gpu_read(gpu, REG_A5XX_CP_HW_FAULT));
942 
943 	if (status & A5XX_CP_INT_CP_DMA_ERROR)
944 		dev_err_ratelimited(gpu->dev->dev, "CP | DMA error\n");
945 
946 	if (status & A5XX_CP_INT_CP_REGISTER_PROTECTION_ERROR) {
947 		u32 val = gpu_read(gpu, REG_A5XX_CP_PROTECT_STATUS);
948 
949 		dev_err_ratelimited(gpu->dev->dev,
950 			"CP | protected mode error | %s | addr=0x%8.8X | status=0x%8.8X\n",
951 			val & (1 << 24) ? "WRITE" : "READ",
952 			(val & 0xFFFFF) >> 2, val);
953 	}
954 
955 	if (status & A5XX_CP_INT_CP_AHB_ERROR) {
956 		u32 status = gpu_read(gpu, REG_A5XX_CP_AHB_FAULT);
957 		const char *access[16] = { "reserved", "reserved",
958 			"timestamp lo", "timestamp hi", "pfp read", "pfp write",
959 			"", "", "me read", "me write", "", "", "crashdump read",
960 			"crashdump write" };
961 
962 		dev_err_ratelimited(gpu->dev->dev,
963 			"CP | AHB error | addr=%X access=%s error=%d | status=0x%8.8X\n",
964 			status & 0xFFFFF, access[(status >> 24) & 0xF],
965 			(status & (1 << 31)), status);
966 	}
967 }
968 
969 static void a5xx_rbbm_err_irq(struct msm_gpu *gpu, u32 status)
970 {
971 	if (status & A5XX_RBBM_INT_0_MASK_RBBM_AHB_ERROR) {
972 		u32 val = gpu_read(gpu, REG_A5XX_RBBM_AHB_ERROR_STATUS);
973 
974 		dev_err_ratelimited(gpu->dev->dev,
975 			"RBBM | AHB bus error | %s | addr=0x%X | ports=0x%X:0x%X\n",
976 			val & (1 << 28) ? "WRITE" : "READ",
977 			(val & 0xFFFFF) >> 2, (val >> 20) & 0x3,
978 			(val >> 24) & 0xF);
979 
980 		/* Clear the error */
981 		gpu_write(gpu, REG_A5XX_RBBM_AHB_CMD, (1 << 4));
982 
983 		/* Clear the interrupt */
984 		gpu_write(gpu, REG_A5XX_RBBM_INT_CLEAR_CMD,
985 			A5XX_RBBM_INT_0_MASK_RBBM_AHB_ERROR);
986 	}
987 
988 	if (status & A5XX_RBBM_INT_0_MASK_RBBM_TRANSFER_TIMEOUT)
989 		dev_err_ratelimited(gpu->dev->dev, "RBBM | AHB transfer timeout\n");
990 
991 	if (status & A5XX_RBBM_INT_0_MASK_RBBM_ME_MS_TIMEOUT)
992 		dev_err_ratelimited(gpu->dev->dev, "RBBM | ME master split | status=0x%X\n",
993 			gpu_read(gpu, REG_A5XX_RBBM_AHB_ME_SPLIT_STATUS));
994 
995 	if (status & A5XX_RBBM_INT_0_MASK_RBBM_PFP_MS_TIMEOUT)
996 		dev_err_ratelimited(gpu->dev->dev, "RBBM | PFP master split | status=0x%X\n",
997 			gpu_read(gpu, REG_A5XX_RBBM_AHB_PFP_SPLIT_STATUS));
998 
999 	if (status & A5XX_RBBM_INT_0_MASK_RBBM_ETS_MS_TIMEOUT)
1000 		dev_err_ratelimited(gpu->dev->dev, "RBBM | ETS master split | status=0x%X\n",
1001 			gpu_read(gpu, REG_A5XX_RBBM_AHB_ETS_SPLIT_STATUS));
1002 
1003 	if (status & A5XX_RBBM_INT_0_MASK_RBBM_ATB_ASYNC_OVERFLOW)
1004 		dev_err_ratelimited(gpu->dev->dev, "RBBM | ATB ASYNC overflow\n");
1005 
1006 	if (status & A5XX_RBBM_INT_0_MASK_RBBM_ATB_BUS_OVERFLOW)
1007 		dev_err_ratelimited(gpu->dev->dev, "RBBM | ATB bus overflow\n");
1008 }
1009 
1010 static void a5xx_uche_err_irq(struct msm_gpu *gpu)
1011 {
1012 	uint64_t addr = (uint64_t) gpu_read(gpu, REG_A5XX_UCHE_TRAP_LOG_HI);
1013 
1014 	addr |= gpu_read(gpu, REG_A5XX_UCHE_TRAP_LOG_LO);
1015 
1016 	dev_err_ratelimited(gpu->dev->dev, "UCHE | Out of bounds access | addr=0x%llX\n",
1017 		addr);
1018 }
1019 
1020 static void a5xx_gpmu_err_irq(struct msm_gpu *gpu)
1021 {
1022 	dev_err_ratelimited(gpu->dev->dev, "GPMU | voltage droop\n");
1023 }
1024 
1025 static void a5xx_fault_detect_irq(struct msm_gpu *gpu)
1026 {
1027 	struct drm_device *dev = gpu->dev;
1028 	struct msm_drm_private *priv = dev->dev_private;
1029 	struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
1030 
1031 	dev_err(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",
1032 		ring ? ring->id : -1, ring ? ring->seqno : 0,
1033 		gpu_read(gpu, REG_A5XX_RBBM_STATUS),
1034 		gpu_read(gpu, REG_A5XX_CP_RB_RPTR),
1035 		gpu_read(gpu, REG_A5XX_CP_RB_WPTR),
1036 		gpu_read64(gpu, REG_A5XX_CP_IB1_BASE, REG_A5XX_CP_IB1_BASE_HI),
1037 		gpu_read(gpu, REG_A5XX_CP_IB1_BUFSZ),
1038 		gpu_read64(gpu, REG_A5XX_CP_IB2_BASE, REG_A5XX_CP_IB2_BASE_HI),
1039 		gpu_read(gpu, REG_A5XX_CP_IB2_BUFSZ));
1040 
1041 	/* Turn off the hangcheck timer to keep it from bothering us */
1042 	del_timer(&gpu->hangcheck_timer);
1043 
1044 	queue_work(priv->wq, &gpu->recover_work);
1045 }
1046 
1047 #define RBBM_ERROR_MASK \
1048 	(A5XX_RBBM_INT_0_MASK_RBBM_AHB_ERROR | \
1049 	A5XX_RBBM_INT_0_MASK_RBBM_TRANSFER_TIMEOUT | \
1050 	A5XX_RBBM_INT_0_MASK_RBBM_ME_MS_TIMEOUT | \
1051 	A5XX_RBBM_INT_0_MASK_RBBM_PFP_MS_TIMEOUT | \
1052 	A5XX_RBBM_INT_0_MASK_RBBM_ETS_MS_TIMEOUT | \
1053 	A5XX_RBBM_INT_0_MASK_RBBM_ATB_ASYNC_OVERFLOW)
1054 
1055 static irqreturn_t a5xx_irq(struct msm_gpu *gpu)
1056 {
1057 	u32 status = gpu_read(gpu, REG_A5XX_RBBM_INT_0_STATUS);
1058 
1059 	/*
1060 	 * Clear all the interrupts except RBBM_AHB_ERROR - if we clear it
1061 	 * before the source is cleared the interrupt will storm.
1062 	 */
1063 	gpu_write(gpu, REG_A5XX_RBBM_INT_CLEAR_CMD,
1064 		status & ~A5XX_RBBM_INT_0_MASK_RBBM_AHB_ERROR);
1065 
1066 	/* Pass status to a5xx_rbbm_err_irq because we've already cleared it */
1067 	if (status & RBBM_ERROR_MASK)
1068 		a5xx_rbbm_err_irq(gpu, status);
1069 
1070 	if (status & A5XX_RBBM_INT_0_MASK_CP_HW_ERROR)
1071 		a5xx_cp_err_irq(gpu);
1072 
1073 	if (status & A5XX_RBBM_INT_0_MASK_MISC_HANG_DETECT)
1074 		a5xx_fault_detect_irq(gpu);
1075 
1076 	if (status & A5XX_RBBM_INT_0_MASK_UCHE_OOB_ACCESS)
1077 		a5xx_uche_err_irq(gpu);
1078 
1079 	if (status & A5XX_RBBM_INT_0_MASK_GPMU_VOLTAGE_DROOP)
1080 		a5xx_gpmu_err_irq(gpu);
1081 
1082 	if (status & A5XX_RBBM_INT_0_MASK_CP_CACHE_FLUSH_TS) {
1083 		a5xx_preempt_trigger(gpu);
1084 		msm_gpu_retire(gpu);
1085 	}
1086 
1087 	if (status & A5XX_RBBM_INT_0_MASK_CP_SW)
1088 		a5xx_preempt_irq(gpu);
1089 
1090 	return IRQ_HANDLED;
1091 }
1092 
1093 static const u32 a5xx_register_offsets[REG_ADRENO_REGISTER_MAX] = {
1094 	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_BASE, REG_A5XX_CP_RB_BASE),
1095 	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_BASE_HI, REG_A5XX_CP_RB_BASE_HI),
1096 	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_RPTR_ADDR, REG_A5XX_CP_RB_RPTR_ADDR),
1097 	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_RPTR_ADDR_HI,
1098 		REG_A5XX_CP_RB_RPTR_ADDR_HI),
1099 	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_RPTR, REG_A5XX_CP_RB_RPTR),
1100 	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_WPTR, REG_A5XX_CP_RB_WPTR),
1101 	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_CNTL, REG_A5XX_CP_RB_CNTL),
1102 };
1103 
1104 static const u32 a5xx_registers[] = {
1105 	0x0000, 0x0002, 0x0004, 0x0020, 0x0022, 0x0026, 0x0029, 0x002B,
1106 	0x002E, 0x0035, 0x0038, 0x0042, 0x0044, 0x0044, 0x0047, 0x0095,
1107 	0x0097, 0x00BB, 0x03A0, 0x0464, 0x0469, 0x046F, 0x04D2, 0x04D3,
1108 	0x04E0, 0x0533, 0x0540, 0x0555, 0x0800, 0x081A, 0x081F, 0x0841,
1109 	0x0860, 0x0860, 0x0880, 0x08A0, 0x0B00, 0x0B12, 0x0B15, 0x0B28,
1110 	0x0B78, 0x0B7F, 0x0BB0, 0x0BBD, 0x0BC0, 0x0BC6, 0x0BD0, 0x0C53,
1111 	0x0C60, 0x0C61, 0x0C80, 0x0C82, 0x0C84, 0x0C85, 0x0C90, 0x0C98,
1112 	0x0CA0, 0x0CA0, 0x0CB0, 0x0CB2, 0x2180, 0x2185, 0x2580, 0x2585,
1113 	0x0CC1, 0x0CC1, 0x0CC4, 0x0CC7, 0x0CCC, 0x0CCC, 0x0CD0, 0x0CD8,
1114 	0x0CE0, 0x0CE5, 0x0CE8, 0x0CE8, 0x0CEC, 0x0CF1, 0x0CFB, 0x0D0E,
1115 	0x2100, 0x211E, 0x2140, 0x2145, 0x2500, 0x251E, 0x2540, 0x2545,
1116 	0x0D10, 0x0D17, 0x0D20, 0x0D23, 0x0D30, 0x0D30, 0x20C0, 0x20C0,
1117 	0x24C0, 0x24C0, 0x0E40, 0x0E43, 0x0E4A, 0x0E4A, 0x0E50, 0x0E57,
1118 	0x0E60, 0x0E7C, 0x0E80, 0x0E8E, 0x0E90, 0x0E96, 0x0EA0, 0x0EA8,
1119 	0x0EB0, 0x0EB2, 0xE140, 0xE147, 0xE150, 0xE187, 0xE1A0, 0xE1A9,
1120 	0xE1B0, 0xE1B6, 0xE1C0, 0xE1C7, 0xE1D0, 0xE1D1, 0xE200, 0xE201,
1121 	0xE210, 0xE21C, 0xE240, 0xE268, 0xE000, 0xE006, 0xE010, 0xE09A,
1122 	0xE0A0, 0xE0A4, 0xE0AA, 0xE0EB, 0xE100, 0xE105, 0xE380, 0xE38F,
1123 	0xE3B0, 0xE3B0, 0xE400, 0xE405, 0xE408, 0xE4E9, 0xE4F0, 0xE4F0,
1124 	0xE280, 0xE280, 0xE282, 0xE2A3, 0xE2A5, 0xE2C2, 0xE940, 0xE947,
1125 	0xE950, 0xE987, 0xE9A0, 0xE9A9, 0xE9B0, 0xE9B6, 0xE9C0, 0xE9C7,
1126 	0xE9D0, 0xE9D1, 0xEA00, 0xEA01, 0xEA10, 0xEA1C, 0xEA40, 0xEA68,
1127 	0xE800, 0xE806, 0xE810, 0xE89A, 0xE8A0, 0xE8A4, 0xE8AA, 0xE8EB,
1128 	0xE900, 0xE905, 0xEB80, 0xEB8F, 0xEBB0, 0xEBB0, 0xEC00, 0xEC05,
1129 	0xEC08, 0xECE9, 0xECF0, 0xECF0, 0xEA80, 0xEA80, 0xEA82, 0xEAA3,
1130 	0xEAA5, 0xEAC2, 0xA800, 0xA800, 0xA820, 0xA828, 0xA840, 0xA87D,
1131 	0XA880, 0xA88D, 0xA890, 0xA8A3, 0xA8D0, 0xA8D8, 0xA8E0, 0xA8F5,
1132 	0xAC60, 0xAC60, ~0,
1133 };
1134 
1135 static void a5xx_dump(struct msm_gpu *gpu)
1136 {
1137 	dev_info(gpu->dev->dev, "status:   %08x\n",
1138 		gpu_read(gpu, REG_A5XX_RBBM_STATUS));
1139 	adreno_dump(gpu);
1140 }
1141 
1142 static int a5xx_pm_resume(struct msm_gpu *gpu)
1143 {
1144 	int ret;
1145 
1146 	/* Turn on the core power */
1147 	ret = msm_gpu_pm_resume(gpu);
1148 	if (ret)
1149 		return ret;
1150 
1151 	/* Turn the RBCCU domain first to limit the chances of voltage droop */
1152 	gpu_write(gpu, REG_A5XX_GPMU_RBCCU_POWER_CNTL, 0x778000);
1153 
1154 	/* Wait 3 usecs before polling */
1155 	udelay(3);
1156 
1157 	ret = spin_usecs(gpu, 20, REG_A5XX_GPMU_RBCCU_PWR_CLK_STATUS,
1158 		(1 << 20), (1 << 20));
1159 	if (ret) {
1160 		DRM_ERROR("%s: timeout waiting for RBCCU GDSC enable: %X\n",
1161 			gpu->name,
1162 			gpu_read(gpu, REG_A5XX_GPMU_RBCCU_PWR_CLK_STATUS));
1163 		return ret;
1164 	}
1165 
1166 	/* Turn on the SP domain */
1167 	gpu_write(gpu, REG_A5XX_GPMU_SP_POWER_CNTL, 0x778000);
1168 	ret = spin_usecs(gpu, 20, REG_A5XX_GPMU_SP_PWR_CLK_STATUS,
1169 		(1 << 20), (1 << 20));
1170 	if (ret)
1171 		DRM_ERROR("%s: timeout waiting for SP GDSC enable\n",
1172 			gpu->name);
1173 
1174 	return ret;
1175 }
1176 
1177 static int a5xx_pm_suspend(struct msm_gpu *gpu)
1178 {
1179 	/* Clear the VBIF pipe before shutting down */
1180 	gpu_write(gpu, REG_A5XX_VBIF_XIN_HALT_CTRL0, 0xF);
1181 	spin_until((gpu_read(gpu, REG_A5XX_VBIF_XIN_HALT_CTRL1) & 0xF) == 0xF);
1182 
1183 	gpu_write(gpu, REG_A5XX_VBIF_XIN_HALT_CTRL0, 0);
1184 
1185 	/*
1186 	 * Reset the VBIF before power collapse to avoid issue with FIFO
1187 	 * entries
1188 	 */
1189 	gpu_write(gpu, REG_A5XX_RBBM_BLOCK_SW_RESET_CMD, 0x003C0000);
1190 	gpu_write(gpu, REG_A5XX_RBBM_BLOCK_SW_RESET_CMD, 0x00000000);
1191 
1192 	return msm_gpu_pm_suspend(gpu);
1193 }
1194 
1195 static int a5xx_get_timestamp(struct msm_gpu *gpu, uint64_t *value)
1196 {
1197 	*value = gpu_read64(gpu, REG_A5XX_RBBM_PERFCTR_CP_0_LO,
1198 		REG_A5XX_RBBM_PERFCTR_CP_0_HI);
1199 
1200 	return 0;
1201 }
1202 
1203 struct a5xx_crashdumper {
1204 	void *ptr;
1205 	struct drm_gem_object *bo;
1206 	u64 iova;
1207 };
1208 
1209 struct a5xx_gpu_state {
1210 	struct msm_gpu_state base;
1211 	u32 *hlsqregs;
1212 };
1213 
1214 #define gpu_poll_timeout(gpu, addr, val, cond, interval, timeout) \
1215 	readl_poll_timeout((gpu)->mmio + ((addr) << 2), val, cond, \
1216 		interval, timeout)
1217 
1218 static int a5xx_crashdumper_init(struct msm_gpu *gpu,
1219 		struct a5xx_crashdumper *dumper)
1220 {
1221 	dumper->ptr = msm_gem_kernel_new_locked(gpu->dev,
1222 		SZ_1M, MSM_BO_UNCACHED, gpu->aspace,
1223 		&dumper->bo, &dumper->iova);
1224 
1225 	if (IS_ERR(dumper->ptr))
1226 		return PTR_ERR(dumper->ptr);
1227 
1228 	return 0;
1229 }
1230 
1231 static void a5xx_crashdumper_free(struct msm_gpu *gpu,
1232 		struct a5xx_crashdumper *dumper)
1233 {
1234 	msm_gem_put_iova(dumper->bo, gpu->aspace);
1235 	msm_gem_put_vaddr(dumper->bo);
1236 
1237 	drm_gem_object_unreference(dumper->bo);
1238 }
1239 
1240 static int a5xx_crashdumper_run(struct msm_gpu *gpu,
1241 		struct a5xx_crashdumper *dumper)
1242 {
1243 	u32 val;
1244 
1245 	if (IS_ERR_OR_NULL(dumper->ptr))
1246 		return -EINVAL;
1247 
1248 	gpu_write64(gpu, REG_A5XX_CP_CRASH_SCRIPT_BASE_LO,
1249 		REG_A5XX_CP_CRASH_SCRIPT_BASE_HI, dumper->iova);
1250 
1251 	gpu_write(gpu, REG_A5XX_CP_CRASH_DUMP_CNTL, 1);
1252 
1253 	return gpu_poll_timeout(gpu, REG_A5XX_CP_CRASH_DUMP_CNTL, val,
1254 		val & 0x04, 100, 10000);
1255 }
1256 
1257 /*
1258  * These are a list of the registers that need to be read through the HLSQ
1259  * aperture through the crashdumper.  These are not nominally accessible from
1260  * the CPU on a secure platform.
1261  */
1262 static const struct {
1263 	u32 type;
1264 	u32 regoffset;
1265 	u32 count;
1266 } a5xx_hlsq_aperture_regs[] = {
1267 	{ 0x35, 0xe00, 0x32 },   /* HSLQ non-context */
1268 	{ 0x31, 0x2080, 0x1 },   /* HLSQ 2D context 0 */
1269 	{ 0x33, 0x2480, 0x1 },   /* HLSQ 2D context 1 */
1270 	{ 0x32, 0xe780, 0x62 },  /* HLSQ 3D context 0 */
1271 	{ 0x34, 0xef80, 0x62 },  /* HLSQ 3D context 1 */
1272 	{ 0x3f, 0x0ec0, 0x40 },  /* SP non-context */
1273 	{ 0x3d, 0x2040, 0x1 },   /* SP 2D context 0 */
1274 	{ 0x3b, 0x2440, 0x1 },   /* SP 2D context 1 */
1275 	{ 0x3e, 0xe580, 0x170 }, /* SP 3D context 0 */
1276 	{ 0x3c, 0xed80, 0x170 }, /* SP 3D context 1 */
1277 	{ 0x3a, 0x0f00, 0x1c },  /* TP non-context */
1278 	{ 0x38, 0x2000, 0xa },   /* TP 2D context 0 */
1279 	{ 0x36, 0x2400, 0xa },   /* TP 2D context 1 */
1280 	{ 0x39, 0xe700, 0x80 },  /* TP 3D context 0 */
1281 	{ 0x37, 0xef00, 0x80 },  /* TP 3D context 1 */
1282 };
1283 
1284 static void a5xx_gpu_state_get_hlsq_regs(struct msm_gpu *gpu,
1285 		struct a5xx_gpu_state *a5xx_state)
1286 {
1287 	struct a5xx_crashdumper dumper = { 0 };
1288 	u32 offset, count = 0;
1289 	u64 *ptr;
1290 	int i;
1291 
1292 	if (a5xx_crashdumper_init(gpu, &dumper))
1293 		return;
1294 
1295 	/* The script will be written at offset 0 */
1296 	ptr = dumper.ptr;
1297 
1298 	/* Start writing the data at offset 256k */
1299 	offset = dumper.iova + (256 * SZ_1K);
1300 
1301 	/* Count how many additional registers to get from the HLSQ aperture */
1302 	for (i = 0; i < ARRAY_SIZE(a5xx_hlsq_aperture_regs); i++)
1303 		count += a5xx_hlsq_aperture_regs[i].count;
1304 
1305 	a5xx_state->hlsqregs = kcalloc(count, sizeof(u32), GFP_KERNEL);
1306 	if (!a5xx_state->hlsqregs)
1307 		return;
1308 
1309 	/* Build the crashdump script */
1310 	for (i = 0; i < ARRAY_SIZE(a5xx_hlsq_aperture_regs); i++) {
1311 		u32 type = a5xx_hlsq_aperture_regs[i].type;
1312 		u32 c = a5xx_hlsq_aperture_regs[i].count;
1313 
1314 		/* Write the register to select the desired bank */
1315 		*ptr++ = ((u64) type << 8);
1316 		*ptr++ = (((u64) REG_A5XX_HLSQ_DBG_READ_SEL) << 44) |
1317 			(1 << 21) | 1;
1318 
1319 		*ptr++ = offset;
1320 		*ptr++ = (((u64) REG_A5XX_HLSQ_DBG_AHB_READ_APERTURE) << 44)
1321 			| c;
1322 
1323 		offset += c * sizeof(u32);
1324 	}
1325 
1326 	/* Write two zeros to close off the script */
1327 	*ptr++ = 0;
1328 	*ptr++ = 0;
1329 
1330 	if (a5xx_crashdumper_run(gpu, &dumper)) {
1331 		kfree(a5xx_state->hlsqregs);
1332 		a5xx_crashdumper_free(gpu, &dumper);
1333 		return;
1334 	}
1335 
1336 	/* Copy the data from the crashdumper to the state */
1337 	memcpy(a5xx_state->hlsqregs, dumper.ptr + (256 * SZ_1K),
1338 		count * sizeof(u32));
1339 
1340 	a5xx_crashdumper_free(gpu, &dumper);
1341 }
1342 
1343 static struct msm_gpu_state *a5xx_gpu_state_get(struct msm_gpu *gpu)
1344 {
1345 	struct a5xx_gpu_state *a5xx_state = kzalloc(sizeof(*a5xx_state),
1346 			GFP_KERNEL);
1347 
1348 	if (!a5xx_state)
1349 		return ERR_PTR(-ENOMEM);
1350 
1351 	/* Temporarily disable hardware clock gating before reading the hw */
1352 	a5xx_set_hwcg(gpu, false);
1353 
1354 	/* First get the generic state from the adreno core */
1355 	adreno_gpu_state_get(gpu, &(a5xx_state->base));
1356 
1357 	a5xx_state->base.rbbm_status = gpu_read(gpu, REG_A5XX_RBBM_STATUS);
1358 
1359 	/* Get the HLSQ regs with the help of the crashdumper */
1360 	a5xx_gpu_state_get_hlsq_regs(gpu, a5xx_state);
1361 
1362 	a5xx_set_hwcg(gpu, true);
1363 
1364 	return &a5xx_state->base;
1365 }
1366 
1367 static void a5xx_gpu_state_destroy(struct kref *kref)
1368 {
1369 	struct msm_gpu_state *state = container_of(kref,
1370 		struct msm_gpu_state, ref);
1371 	struct a5xx_gpu_state *a5xx_state = container_of(state,
1372 		struct a5xx_gpu_state, base);
1373 
1374 	kfree(a5xx_state->hlsqregs);
1375 
1376 	adreno_gpu_state_destroy(state);
1377 	kfree(a5xx_state);
1378 }
1379 
1380 int a5xx_gpu_state_put(struct msm_gpu_state *state)
1381 {
1382 	if (IS_ERR_OR_NULL(state))
1383 		return 1;
1384 
1385 	return kref_put(&state->ref, a5xx_gpu_state_destroy);
1386 }
1387 
1388 
1389 #if defined(CONFIG_DEBUG_FS) || defined(CONFIG_DEV_COREDUMP)
1390 void a5xx_show(struct msm_gpu *gpu, struct msm_gpu_state *state,
1391 		struct drm_printer *p)
1392 {
1393 	int i, j;
1394 	u32 pos = 0;
1395 	struct a5xx_gpu_state *a5xx_state = container_of(state,
1396 		struct a5xx_gpu_state, base);
1397 
1398 	if (IS_ERR_OR_NULL(state))
1399 		return;
1400 
1401 	adreno_show(gpu, state, p);
1402 
1403 	/* Dump the additional a5xx HLSQ registers */
1404 	if (!a5xx_state->hlsqregs)
1405 		return;
1406 
1407 	drm_printf(p, "registers-hlsq:\n");
1408 
1409 	for (i = 0; i < ARRAY_SIZE(a5xx_hlsq_aperture_regs); i++) {
1410 		u32 o = a5xx_hlsq_aperture_regs[i].regoffset;
1411 		u32 c = a5xx_hlsq_aperture_regs[i].count;
1412 
1413 		for (j = 0; j < c; j++, pos++, o++) {
1414 			/*
1415 			 * To keep the crashdump simple we pull the entire range
1416 			 * for each register type but not all of the registers
1417 			 * in the range are valid. Fortunately invalid registers
1418 			 * stick out like a sore thumb with a value of
1419 			 * 0xdeadbeef
1420 			 */
1421 			if (a5xx_state->hlsqregs[pos] == 0xdeadbeef)
1422 				continue;
1423 
1424 			drm_printf(p, "  - { offset: 0x%04x, value: 0x%08x }\n",
1425 				o << 2, a5xx_state->hlsqregs[pos]);
1426 		}
1427 	}
1428 }
1429 #endif
1430 
1431 static struct msm_ringbuffer *a5xx_active_ring(struct msm_gpu *gpu)
1432 {
1433 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1434 	struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu);
1435 
1436 	return a5xx_gpu->cur_ring;
1437 }
1438 
1439 static int a5xx_gpu_busy(struct msm_gpu *gpu, uint64_t *value)
1440 {
1441 	*value = gpu_read64(gpu, REG_A5XX_RBBM_PERFCTR_RBBM_0_LO,
1442 		REG_A5XX_RBBM_PERFCTR_RBBM_0_HI);
1443 
1444 	return 0;
1445 }
1446 
1447 static const struct adreno_gpu_funcs funcs = {
1448 	.base = {
1449 		.get_param = adreno_get_param,
1450 		.hw_init = a5xx_hw_init,
1451 		.pm_suspend = a5xx_pm_suspend,
1452 		.pm_resume = a5xx_pm_resume,
1453 		.recover = a5xx_recover,
1454 		.submit = a5xx_submit,
1455 		.flush = a5xx_flush,
1456 		.active_ring = a5xx_active_ring,
1457 		.irq = a5xx_irq,
1458 		.destroy = a5xx_destroy,
1459 #if defined(CONFIG_DEBUG_FS) || defined(CONFIG_DEV_COREDUMP)
1460 		.show = a5xx_show,
1461 #endif
1462 #if defined(CONFIG_DEBUG_FS)
1463 		.debugfs_init = a5xx_debugfs_init,
1464 #endif
1465 		.gpu_busy = a5xx_gpu_busy,
1466 		.gpu_state_get = a5xx_gpu_state_get,
1467 		.gpu_state_put = a5xx_gpu_state_put,
1468 	},
1469 	.get_timestamp = a5xx_get_timestamp,
1470 };
1471 
1472 static void check_speed_bin(struct device *dev)
1473 {
1474 	struct nvmem_cell *cell;
1475 	u32 bin, val;
1476 
1477 	cell = nvmem_cell_get(dev, "speed_bin");
1478 
1479 	/* If a nvmem cell isn't defined, nothing to do */
1480 	if (IS_ERR(cell))
1481 		return;
1482 
1483 	bin = *((u32 *) nvmem_cell_read(cell, NULL));
1484 	nvmem_cell_put(cell);
1485 
1486 	val = (1 << bin);
1487 
1488 	dev_pm_opp_set_supported_hw(dev, &val, 1);
1489 }
1490 
1491 struct msm_gpu *a5xx_gpu_init(struct drm_device *dev)
1492 {
1493 	struct msm_drm_private *priv = dev->dev_private;
1494 	struct platform_device *pdev = priv->gpu_pdev;
1495 	struct a5xx_gpu *a5xx_gpu = NULL;
1496 	struct adreno_gpu *adreno_gpu;
1497 	struct msm_gpu *gpu;
1498 	int ret;
1499 
1500 	if (!pdev) {
1501 		dev_err(dev->dev, "No A5XX device is defined\n");
1502 		return ERR_PTR(-ENXIO);
1503 	}
1504 
1505 	a5xx_gpu = kzalloc(sizeof(*a5xx_gpu), GFP_KERNEL);
1506 	if (!a5xx_gpu)
1507 		return ERR_PTR(-ENOMEM);
1508 
1509 	adreno_gpu = &a5xx_gpu->base;
1510 	gpu = &adreno_gpu->base;
1511 
1512 	adreno_gpu->registers = a5xx_registers;
1513 	adreno_gpu->reg_offsets = a5xx_register_offsets;
1514 
1515 	a5xx_gpu->lm_leakage = 0x4E001A;
1516 
1517 	check_speed_bin(&pdev->dev);
1518 
1519 	ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs, 4);
1520 	if (ret) {
1521 		a5xx_destroy(&(a5xx_gpu->base.base));
1522 		return ERR_PTR(ret);
1523 	}
1524 
1525 	if (gpu->aspace)
1526 		msm_mmu_set_fault_handler(gpu->aspace->mmu, gpu, a5xx_fault_handler);
1527 
1528 	/* Set up the preemption specific bits and pieces for each ringbuffer */
1529 	a5xx_preempt_init(gpu);
1530 
1531 	return gpu;
1532 }
1533