xref: /openbmc/linux/drivers/gpu/drm/msm/adreno/a6xx_hfi.c (revision d40d48e1)
1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright (c) 2017-2018 The Linux Foundation. All rights reserved. */
3 
4 #include <linux/completion.h>
5 #include <linux/circ_buf.h>
6 #include <linux/list.h>
7 
8 #include "a6xx_gmu.h"
9 #include "a6xx_gmu.xml.h"
10 #include "a6xx_gpu.h"
11 
12 #define HFI_MSG_ID(val) [val] = #val
13 
14 static const char * const a6xx_hfi_msg_id[] = {
15 	HFI_MSG_ID(HFI_H2F_MSG_INIT),
16 	HFI_MSG_ID(HFI_H2F_MSG_FW_VERSION),
17 	HFI_MSG_ID(HFI_H2F_MSG_BW_TABLE),
18 	HFI_MSG_ID(HFI_H2F_MSG_PERF_TABLE),
19 	HFI_MSG_ID(HFI_H2F_MSG_TEST),
20 	HFI_MSG_ID(HFI_H2F_MSG_START),
21 	HFI_MSG_ID(HFI_H2F_MSG_CORE_FW_START),
22 	HFI_MSG_ID(HFI_H2F_MSG_GX_BW_PERF_VOTE),
23 	HFI_MSG_ID(HFI_H2F_MSG_PREPARE_SLUMBER),
24 };
25 
26 static int a6xx_hfi_queue_read(struct a6xx_gmu *gmu,
27 	struct a6xx_hfi_queue *queue, u32 *data, u32 dwords)
28 {
29 	struct a6xx_hfi_queue_header *header = queue->header;
30 	u32 i, hdr, index = header->read_index;
31 
32 	if (header->read_index == header->write_index) {
33 		header->rx_request = 1;
34 		return 0;
35 	}
36 
37 	hdr = queue->data[index];
38 
39 	/*
40 	 * If we are to assume that the GMU firmware is in fact a rational actor
41 	 * and is programmed to not send us a larger response than we expect
42 	 * then we can also assume that if the header size is unexpectedly large
43 	 * that it is due to memory corruption and/or hardware failure. In this
44 	 * case the only reasonable course of action is to BUG() to help harden
45 	 * the failure.
46 	 */
47 
48 	BUG_ON(HFI_HEADER_SIZE(hdr) > dwords);
49 
50 	for (i = 0; i < HFI_HEADER_SIZE(hdr); i++) {
51 		data[i] = queue->data[index];
52 		index = (index + 1) % header->size;
53 	}
54 
55 	if (!gmu->legacy)
56 		index = ALIGN(index, 4) % header->size;
57 
58 	header->read_index = index;
59 	return HFI_HEADER_SIZE(hdr);
60 }
61 
62 static int a6xx_hfi_queue_write(struct a6xx_gmu *gmu,
63 	struct a6xx_hfi_queue *queue, u32 *data, u32 dwords)
64 {
65 	struct a6xx_hfi_queue_header *header = queue->header;
66 	u32 i, space, index = header->write_index;
67 
68 	spin_lock(&queue->lock);
69 
70 	space = CIRC_SPACE(header->write_index, header->read_index,
71 		header->size);
72 	if (space < dwords) {
73 		header->dropped++;
74 		spin_unlock(&queue->lock);
75 		return -ENOSPC;
76 	}
77 
78 	for (i = 0; i < dwords; i++) {
79 		queue->data[index] = data[i];
80 		index = (index + 1) % header->size;
81 	}
82 
83 	/* Cookify any non used data at the end of the write buffer */
84 	if (!gmu->legacy) {
85 		for (; index % 4; index = (index + 1) % header->size)
86 			queue->data[index] = 0xfafafafa;
87 	}
88 
89 	header->write_index = index;
90 	spin_unlock(&queue->lock);
91 
92 	gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET, 0x01);
93 	return 0;
94 }
95 
96 static int a6xx_hfi_wait_for_ack(struct a6xx_gmu *gmu, u32 id, u32 seqnum,
97 		u32 *payload, u32 payload_size)
98 {
99 	struct a6xx_hfi_queue *queue = &gmu->queues[HFI_RESPONSE_QUEUE];
100 	u32 val;
101 	int ret;
102 
103 	/* Wait for a response */
104 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO, val,
105 		val & A6XX_GMU_GMU2HOST_INTR_INFO_MSGQ, 100, 5000);
106 
107 	if (ret) {
108 		DRM_DEV_ERROR(gmu->dev,
109 			"Message %s id %d timed out waiting for response\n",
110 			a6xx_hfi_msg_id[id], seqnum);
111 		return -ETIMEDOUT;
112 	}
113 
114 	/* Clear the interrupt */
115 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR,
116 		A6XX_GMU_GMU2HOST_INTR_INFO_MSGQ);
117 
118 	for (;;) {
119 		struct a6xx_hfi_msg_response resp;
120 
121 		/* Get the next packet */
122 		ret = a6xx_hfi_queue_read(gmu, queue, (u32 *) &resp,
123 			sizeof(resp) >> 2);
124 
125 		/* If the queue is empty our response never made it */
126 		if (!ret) {
127 			DRM_DEV_ERROR(gmu->dev,
128 				"The HFI response queue is unexpectedly empty\n");
129 
130 			return -ENOENT;
131 		}
132 
133 		if (HFI_HEADER_ID(resp.header) == HFI_F2H_MSG_ERROR) {
134 			struct a6xx_hfi_msg_error *error =
135 				(struct a6xx_hfi_msg_error *) &resp;
136 
137 			DRM_DEV_ERROR(gmu->dev, "GMU firmware error %d\n",
138 				error->code);
139 			continue;
140 		}
141 
142 		if (seqnum != HFI_HEADER_SEQNUM(resp.ret_header)) {
143 			DRM_DEV_ERROR(gmu->dev,
144 				"Unexpected message id %d on the response queue\n",
145 				HFI_HEADER_SEQNUM(resp.ret_header));
146 			continue;
147 		}
148 
149 		if (resp.error) {
150 			DRM_DEV_ERROR(gmu->dev,
151 				"Message %s id %d returned error %d\n",
152 				a6xx_hfi_msg_id[id], seqnum, resp.error);
153 			return -EINVAL;
154 		}
155 
156 		/* All is well, copy over the buffer */
157 		if (payload && payload_size)
158 			memcpy(payload, resp.payload,
159 				min_t(u32, payload_size, sizeof(resp.payload)));
160 
161 		return 0;
162 	}
163 }
164 
165 static int a6xx_hfi_send_msg(struct a6xx_gmu *gmu, int id,
166 		void *data, u32 size, u32 *payload, u32 payload_size)
167 {
168 	struct a6xx_hfi_queue *queue = &gmu->queues[HFI_COMMAND_QUEUE];
169 	int ret, dwords = size >> 2;
170 	u32 seqnum;
171 
172 	seqnum = atomic_inc_return(&queue->seqnum) % 0xfff;
173 
174 	/* First dword of the message is the message header - fill it in */
175 	*((u32 *) data) = (seqnum << 20) | (HFI_MSG_CMD << 16) |
176 		(dwords << 8) | id;
177 
178 	ret = a6xx_hfi_queue_write(gmu, queue, data, dwords);
179 	if (ret) {
180 		DRM_DEV_ERROR(gmu->dev, "Unable to send message %s id %d\n",
181 			a6xx_hfi_msg_id[id], seqnum);
182 		return ret;
183 	}
184 
185 	return a6xx_hfi_wait_for_ack(gmu, id, seqnum, payload, payload_size);
186 }
187 
188 static int a6xx_hfi_send_gmu_init(struct a6xx_gmu *gmu, int boot_state)
189 {
190 	struct a6xx_hfi_msg_gmu_init_cmd msg = { 0 };
191 
192 	msg.dbg_buffer_addr = (u32) gmu->debug.iova;
193 	msg.dbg_buffer_size = (u32) gmu->debug.size;
194 	msg.boot_state = boot_state;
195 
196 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_INIT, &msg, sizeof(msg),
197 		NULL, 0);
198 }
199 
200 static int a6xx_hfi_get_fw_version(struct a6xx_gmu *gmu, u32 *version)
201 {
202 	struct a6xx_hfi_msg_fw_version msg = { 0 };
203 
204 	/* Currently supporting version 1.1 */
205 	msg.supported_version = (1 << 28) | (1 << 16);
206 
207 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_FW_VERSION, &msg, sizeof(msg),
208 		version, sizeof(*version));
209 }
210 
211 static int a6xx_hfi_send_perf_table_v1(struct a6xx_gmu *gmu)
212 {
213 	struct a6xx_hfi_msg_perf_table_v1 msg = { 0 };
214 	int i;
215 
216 	msg.num_gpu_levels = gmu->nr_gpu_freqs;
217 	msg.num_gmu_levels = gmu->nr_gmu_freqs;
218 
219 	for (i = 0; i < gmu->nr_gpu_freqs; i++) {
220 		msg.gx_votes[i].vote = gmu->gx_arc_votes[i];
221 		msg.gx_votes[i].freq = gmu->gpu_freqs[i] / 1000;
222 	}
223 
224 	for (i = 0; i < gmu->nr_gmu_freqs; i++) {
225 		msg.cx_votes[i].vote = gmu->cx_arc_votes[i];
226 		msg.cx_votes[i].freq = gmu->gmu_freqs[i] / 1000;
227 	}
228 
229 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_PERF_TABLE, &msg, sizeof(msg),
230 		NULL, 0);
231 }
232 
233 static int a6xx_hfi_send_perf_table(struct a6xx_gmu *gmu)
234 {
235 	struct a6xx_hfi_msg_perf_table msg = { 0 };
236 	int i;
237 
238 	msg.num_gpu_levels = gmu->nr_gpu_freqs;
239 	msg.num_gmu_levels = gmu->nr_gmu_freqs;
240 
241 	for (i = 0; i < gmu->nr_gpu_freqs; i++) {
242 		msg.gx_votes[i].vote = gmu->gx_arc_votes[i];
243 		msg.gx_votes[i].acd = 0xffffffff;
244 		msg.gx_votes[i].freq = gmu->gpu_freqs[i] / 1000;
245 	}
246 
247 	for (i = 0; i < gmu->nr_gmu_freqs; i++) {
248 		msg.cx_votes[i].vote = gmu->cx_arc_votes[i];
249 		msg.cx_votes[i].freq = gmu->gmu_freqs[i] / 1000;
250 	}
251 
252 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_PERF_TABLE, &msg, sizeof(msg),
253 		NULL, 0);
254 }
255 
256 static void a618_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
257 {
258 	/* Send a single "off" entry since the 618 GMU doesn't do bus scaling */
259 	msg->bw_level_num = 1;
260 
261 	msg->ddr_cmds_num = 3;
262 	msg->ddr_wait_bitmask = 0x01;
263 
264 	msg->ddr_cmds_addrs[0] = 0x50000;
265 	msg->ddr_cmds_addrs[1] = 0x5003c;
266 	msg->ddr_cmds_addrs[2] = 0x5000c;
267 
268 	msg->ddr_cmds_data[0][0] =  0x40000000;
269 	msg->ddr_cmds_data[0][1] =  0x40000000;
270 	msg->ddr_cmds_data[0][2] =  0x40000000;
271 
272 	/*
273 	 * These are the CX (CNOC) votes - these are used by the GMU but the
274 	 * votes are known and fixed for the target
275 	 */
276 	msg->cnoc_cmds_num = 1;
277 	msg->cnoc_wait_bitmask = 0x01;
278 
279 	msg->cnoc_cmds_addrs[0] = 0x5007c;
280 	msg->cnoc_cmds_data[0][0] =  0x40000000;
281 	msg->cnoc_cmds_data[1][0] =  0x60000001;
282 }
283 
284 static void a640_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
285 {
286 	/*
287 	 * Send a single "off" entry just to get things running
288 	 * TODO: bus scaling
289 	 */
290 	msg->bw_level_num = 1;
291 
292 	msg->ddr_cmds_num = 3;
293 	msg->ddr_wait_bitmask = 0x01;
294 
295 	msg->ddr_cmds_addrs[0] = 0x50000;
296 	msg->ddr_cmds_addrs[1] = 0x5003c;
297 	msg->ddr_cmds_addrs[2] = 0x5000c;
298 
299 	msg->ddr_cmds_data[0][0] =  0x40000000;
300 	msg->ddr_cmds_data[0][1] =  0x40000000;
301 	msg->ddr_cmds_data[0][2] =  0x40000000;
302 
303 	/*
304 	 * These are the CX (CNOC) votes - these are used by the GMU but the
305 	 * votes are known and fixed for the target
306 	 */
307 	msg->cnoc_cmds_num = 3;
308 	msg->cnoc_wait_bitmask = 0x01;
309 
310 	msg->cnoc_cmds_addrs[0] = 0x50034;
311 	msg->cnoc_cmds_addrs[1] = 0x5007c;
312 	msg->cnoc_cmds_addrs[2] = 0x5004c;
313 
314 	msg->cnoc_cmds_data[0][0] =  0x40000000;
315 	msg->cnoc_cmds_data[0][1] =  0x00000000;
316 	msg->cnoc_cmds_data[0][2] =  0x40000000;
317 
318 	msg->cnoc_cmds_data[1][0] =  0x60000001;
319 	msg->cnoc_cmds_data[1][1] =  0x20000001;
320 	msg->cnoc_cmds_data[1][2] =  0x60000001;
321 }
322 
323 static void a650_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
324 {
325 	/*
326 	 * Send a single "off" entry just to get things running
327 	 * TODO: bus scaling
328 	 */
329 	msg->bw_level_num = 1;
330 
331 	msg->ddr_cmds_num = 3;
332 	msg->ddr_wait_bitmask = 0x01;
333 
334 	msg->ddr_cmds_addrs[0] = 0x50000;
335 	msg->ddr_cmds_addrs[1] = 0x50004;
336 	msg->ddr_cmds_addrs[2] = 0x5007c;
337 
338 	msg->ddr_cmds_data[0][0] =  0x40000000;
339 	msg->ddr_cmds_data[0][1] =  0x40000000;
340 	msg->ddr_cmds_data[0][2] =  0x40000000;
341 
342 	/*
343 	 * These are the CX (CNOC) votes - these are used by the GMU but the
344 	 * votes are known and fixed for the target
345 	 */
346 	msg->cnoc_cmds_num = 1;
347 	msg->cnoc_wait_bitmask = 0x01;
348 
349 	msg->cnoc_cmds_addrs[0] = 0x500a4;
350 	msg->cnoc_cmds_data[0][0] =  0x40000000;
351 	msg->cnoc_cmds_data[1][0] =  0x60000001;
352 }
353 
354 static void a660_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
355 {
356 	/*
357 	 * Send a single "off" entry just to get things running
358 	 * TODO: bus scaling
359 	 */
360 	msg->bw_level_num = 1;
361 
362 	msg->ddr_cmds_num = 3;
363 	msg->ddr_wait_bitmask = 0x01;
364 
365 	msg->ddr_cmds_addrs[0] = 0x50004;
366 	msg->ddr_cmds_addrs[1] = 0x500a0;
367 	msg->ddr_cmds_addrs[2] = 0x50000;
368 
369 	msg->ddr_cmds_data[0][0] =  0x40000000;
370 	msg->ddr_cmds_data[0][1] =  0x40000000;
371 	msg->ddr_cmds_data[0][2] =  0x40000000;
372 
373 	/*
374 	 * These are the CX (CNOC) votes - these are used by the GMU but the
375 	 * votes are known and fixed for the target
376 	 */
377 	msg->cnoc_cmds_num = 1;
378 	msg->cnoc_wait_bitmask = 0x01;
379 
380 	msg->cnoc_cmds_addrs[0] = 0x50070;
381 	msg->cnoc_cmds_data[0][0] =  0x40000000;
382 	msg->cnoc_cmds_data[1][0] =  0x60000001;
383 }
384 
385 static void adreno_7c3_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
386 {
387 	/*
388 	 * Send a single "off" entry just to get things running
389 	 * TODO: bus scaling
390 	 */
391 	msg->bw_level_num = 1;
392 
393 	msg->ddr_cmds_num = 3;
394 	msg->ddr_wait_bitmask = 0x07;
395 
396 	msg->ddr_cmds_addrs[0] = 0x50004;
397 	msg->ddr_cmds_addrs[1] = 0x50000;
398 	msg->ddr_cmds_addrs[2] = 0x50088;
399 
400 	msg->ddr_cmds_data[0][0] =  0x40000000;
401 	msg->ddr_cmds_data[0][1] =  0x40000000;
402 	msg->ddr_cmds_data[0][2] =  0x40000000;
403 
404 	/*
405 	 * These are the CX (CNOC) votes - these are used by the GMU but the
406 	 * votes are known and fixed for the target
407 	 */
408 	msg->cnoc_cmds_num = 1;
409 	msg->cnoc_wait_bitmask = 0x01;
410 
411 	msg->cnoc_cmds_addrs[0] = 0x5006c;
412 	msg->cnoc_cmds_data[0][0] =  0x40000000;
413 	msg->cnoc_cmds_data[1][0] =  0x60000001;
414 }
415 static void a6xx_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
416 {
417 	/* Send a single "off" entry since the 630 GMU doesn't do bus scaling */
418 	msg->bw_level_num = 1;
419 
420 	msg->ddr_cmds_num = 3;
421 	msg->ddr_wait_bitmask = 0x07;
422 
423 	msg->ddr_cmds_addrs[0] = 0x50000;
424 	msg->ddr_cmds_addrs[1] = 0x5005c;
425 	msg->ddr_cmds_addrs[2] = 0x5000c;
426 
427 	msg->ddr_cmds_data[0][0] =  0x40000000;
428 	msg->ddr_cmds_data[0][1] =  0x40000000;
429 	msg->ddr_cmds_data[0][2] =  0x40000000;
430 
431 	/*
432 	 * These are the CX (CNOC) votes.  This is used but the values for the
433 	 * sdm845 GMU are known and fixed so we can hard code them.
434 	 */
435 
436 	msg->cnoc_cmds_num = 3;
437 	msg->cnoc_wait_bitmask = 0x05;
438 
439 	msg->cnoc_cmds_addrs[0] = 0x50034;
440 	msg->cnoc_cmds_addrs[1] = 0x5007c;
441 	msg->cnoc_cmds_addrs[2] = 0x5004c;
442 
443 	msg->cnoc_cmds_data[0][0] =  0x40000000;
444 	msg->cnoc_cmds_data[0][1] =  0x00000000;
445 	msg->cnoc_cmds_data[0][2] =  0x40000000;
446 
447 	msg->cnoc_cmds_data[1][0] =  0x60000001;
448 	msg->cnoc_cmds_data[1][1] =  0x20000001;
449 	msg->cnoc_cmds_data[1][2] =  0x60000001;
450 }
451 
452 
453 static int a6xx_hfi_send_bw_table(struct a6xx_gmu *gmu)
454 {
455 	struct a6xx_hfi_msg_bw_table msg = { 0 };
456 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
457 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
458 
459 	if (adreno_is_a618(adreno_gpu))
460 		a618_build_bw_table(&msg);
461 	else if (adreno_is_a640_family(adreno_gpu))
462 		a640_build_bw_table(&msg);
463 	else if (adreno_is_a650(adreno_gpu))
464 		a650_build_bw_table(&msg);
465 	else if (adreno_is_7c3(adreno_gpu))
466 		adreno_7c3_build_bw_table(&msg);
467 	else if (adreno_is_a660(adreno_gpu))
468 		a660_build_bw_table(&msg);
469 	else
470 		a6xx_build_bw_table(&msg);
471 
472 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_BW_TABLE, &msg, sizeof(msg),
473 		NULL, 0);
474 }
475 
476 static int a6xx_hfi_send_test(struct a6xx_gmu *gmu)
477 {
478 	struct a6xx_hfi_msg_test msg = { 0 };
479 
480 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_TEST, &msg, sizeof(msg),
481 		NULL, 0);
482 }
483 
484 static int a6xx_hfi_send_start(struct a6xx_gmu *gmu)
485 {
486 	struct a6xx_hfi_msg_start msg = { 0 };
487 
488 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_START, &msg, sizeof(msg),
489 		NULL, 0);
490 }
491 
492 static int a6xx_hfi_send_core_fw_start(struct a6xx_gmu *gmu)
493 {
494 	struct a6xx_hfi_msg_core_fw_start msg = { 0 };
495 
496 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_CORE_FW_START, &msg,
497 		sizeof(msg), NULL, 0);
498 }
499 
500 int a6xx_hfi_set_freq(struct a6xx_gmu *gmu, int index)
501 {
502 	struct a6xx_hfi_gx_bw_perf_vote_cmd msg = { 0 };
503 
504 	msg.ack_type = 1; /* blocking */
505 	msg.freq = index;
506 	msg.bw = 0; /* TODO: bus scaling */
507 
508 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_GX_BW_PERF_VOTE, &msg,
509 		sizeof(msg), NULL, 0);
510 }
511 
512 int a6xx_hfi_send_prep_slumber(struct a6xx_gmu *gmu)
513 {
514 	struct a6xx_hfi_prep_slumber_cmd msg = { 0 };
515 
516 	/* TODO: should freq and bw fields be non-zero ? */
517 
518 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_PREPARE_SLUMBER, &msg,
519 		sizeof(msg), NULL, 0);
520 }
521 
522 static int a6xx_hfi_start_v1(struct a6xx_gmu *gmu, int boot_state)
523 {
524 	int ret;
525 
526 	ret = a6xx_hfi_send_gmu_init(gmu, boot_state);
527 	if (ret)
528 		return ret;
529 
530 	ret = a6xx_hfi_get_fw_version(gmu, NULL);
531 	if (ret)
532 		return ret;
533 
534 	/*
535 	 * We have to get exchange version numbers per the sequence but at this
536 	 * point th kernel driver doesn't need to know the exact version of
537 	 * the GMU firmware
538 	 */
539 
540 	ret = a6xx_hfi_send_perf_table_v1(gmu);
541 	if (ret)
542 		return ret;
543 
544 	ret = a6xx_hfi_send_bw_table(gmu);
545 	if (ret)
546 		return ret;
547 
548 	/*
549 	 * Let the GMU know that there won't be any more HFI messages until next
550 	 * boot
551 	 */
552 	a6xx_hfi_send_test(gmu);
553 
554 	return 0;
555 }
556 
557 int a6xx_hfi_start(struct a6xx_gmu *gmu, int boot_state)
558 {
559 	int ret;
560 
561 	if (gmu->legacy)
562 		return a6xx_hfi_start_v1(gmu, boot_state);
563 
564 
565 	ret = a6xx_hfi_send_perf_table(gmu);
566 	if (ret)
567 		return ret;
568 
569 	ret = a6xx_hfi_send_bw_table(gmu);
570 	if (ret)
571 		return ret;
572 
573 	ret = a6xx_hfi_send_core_fw_start(gmu);
574 	if (ret)
575 		return ret;
576 
577 	/*
578 	 * Downstream driver sends this in its "a6xx_hw_init" equivalent,
579 	 * but seems to be no harm in sending it here
580 	 */
581 	ret = a6xx_hfi_send_start(gmu);
582 	if (ret)
583 		return ret;
584 
585 	return 0;
586 }
587 
588 void a6xx_hfi_stop(struct a6xx_gmu *gmu)
589 {
590 	int i;
591 
592 	for (i = 0; i < ARRAY_SIZE(gmu->queues); i++) {
593 		struct a6xx_hfi_queue *queue = &gmu->queues[i];
594 
595 		if (!queue->header)
596 			continue;
597 
598 		if (queue->header->read_index != queue->header->write_index)
599 			DRM_DEV_ERROR(gmu->dev, "HFI queue %d is not empty\n", i);
600 
601 		queue->header->read_index = 0;
602 		queue->header->write_index = 0;
603 	}
604 }
605 
606 static void a6xx_hfi_queue_init(struct a6xx_hfi_queue *queue,
607 		struct a6xx_hfi_queue_header *header, void *virt, u64 iova,
608 		u32 id)
609 {
610 	spin_lock_init(&queue->lock);
611 	queue->header = header;
612 	queue->data = virt;
613 	atomic_set(&queue->seqnum, 0);
614 
615 	/* Set up the shared memory header */
616 	header->iova = iova;
617 	header->type =  10 << 8 | id;
618 	header->status = 1;
619 	header->size = SZ_4K >> 2;
620 	header->msg_size = 0;
621 	header->dropped = 0;
622 	header->rx_watermark = 1;
623 	header->tx_watermark = 1;
624 	header->rx_request = 1;
625 	header->tx_request = 0;
626 	header->read_index = 0;
627 	header->write_index = 0;
628 }
629 
630 void a6xx_hfi_init(struct a6xx_gmu *gmu)
631 {
632 	struct a6xx_gmu_bo *hfi = &gmu->hfi;
633 	struct a6xx_hfi_queue_table_header *table = hfi->virt;
634 	struct a6xx_hfi_queue_header *headers = hfi->virt + sizeof(*table);
635 	u64 offset;
636 	int table_size;
637 
638 	/*
639 	 * The table size is the size of the table header plus all of the queue
640 	 * headers
641 	 */
642 	table_size = sizeof(*table);
643 	table_size += (ARRAY_SIZE(gmu->queues) *
644 		sizeof(struct a6xx_hfi_queue_header));
645 
646 	table->version = 0;
647 	table->size = table_size;
648 	/* First queue header is located immediately after the table header */
649 	table->qhdr0_offset = sizeof(*table) >> 2;
650 	table->qhdr_size = sizeof(struct a6xx_hfi_queue_header) >> 2;
651 	table->num_queues = ARRAY_SIZE(gmu->queues);
652 	table->active_queues = ARRAY_SIZE(gmu->queues);
653 
654 	/* Command queue */
655 	offset = SZ_4K;
656 	a6xx_hfi_queue_init(&gmu->queues[0], &headers[0], hfi->virt + offset,
657 		hfi->iova + offset, 0);
658 
659 	/* GMU response queue */
660 	offset += SZ_4K;
661 	a6xx_hfi_queue_init(&gmu->queues[1], &headers[1], hfi->virt + offset,
662 		hfi->iova + offset, gmu->legacy ? 4 : 1);
663 }
664