xref: /openbmc/linux/drivers/gpu/drm/msm/adreno/a6xx_hfi.c (revision c4849f88)
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 	queue->history[(queue->history_idx++) % HFI_HISTORY_SZ] = index;
40 
41 	/*
42 	 * If we are to assume that the GMU firmware is in fact a rational actor
43 	 * and is programmed to not send us a larger response than we expect
44 	 * then we can also assume that if the header size is unexpectedly large
45 	 * that it is due to memory corruption and/or hardware failure. In this
46 	 * case the only reasonable course of action is to BUG() to help harden
47 	 * the failure.
48 	 */
49 
50 	BUG_ON(HFI_HEADER_SIZE(hdr) > dwords);
51 
52 	for (i = 0; i < HFI_HEADER_SIZE(hdr); i++) {
53 		data[i] = queue->data[index];
54 		index = (index + 1) % header->size;
55 	}
56 
57 	if (!gmu->legacy)
58 		index = ALIGN(index, 4) % header->size;
59 
60 	header->read_index = index;
61 	return HFI_HEADER_SIZE(hdr);
62 }
63 
64 static int a6xx_hfi_queue_write(struct a6xx_gmu *gmu,
65 	struct a6xx_hfi_queue *queue, u32 *data, u32 dwords)
66 {
67 	struct a6xx_hfi_queue_header *header = queue->header;
68 	u32 i, space, index = header->write_index;
69 
70 	spin_lock(&queue->lock);
71 
72 	space = CIRC_SPACE(header->write_index, header->read_index,
73 		header->size);
74 	if (space < dwords) {
75 		header->dropped++;
76 		spin_unlock(&queue->lock);
77 		return -ENOSPC;
78 	}
79 
80 	queue->history[(queue->history_idx++) % HFI_HISTORY_SZ] = index;
81 
82 	for (i = 0; i < dwords; i++) {
83 		queue->data[index] = data[i];
84 		index = (index + 1) % header->size;
85 	}
86 
87 	/* Cookify any non used data at the end of the write buffer */
88 	if (!gmu->legacy) {
89 		for (; index % 4; index = (index + 1) % header->size)
90 			queue->data[index] = 0xfafafafa;
91 	}
92 
93 	header->write_index = index;
94 	spin_unlock(&queue->lock);
95 
96 	gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET, 0x01);
97 	return 0;
98 }
99 
100 static int a6xx_hfi_wait_for_ack(struct a6xx_gmu *gmu, u32 id, u32 seqnum,
101 		u32 *payload, u32 payload_size)
102 {
103 	struct a6xx_hfi_queue *queue = &gmu->queues[HFI_RESPONSE_QUEUE];
104 	u32 val;
105 	int ret;
106 
107 	/* Wait for a response */
108 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO, val,
109 		val & A6XX_GMU_GMU2HOST_INTR_INFO_MSGQ, 100, 5000);
110 
111 	if (ret) {
112 		DRM_DEV_ERROR(gmu->dev,
113 			"Message %s id %d timed out waiting for response\n",
114 			a6xx_hfi_msg_id[id], seqnum);
115 		return -ETIMEDOUT;
116 	}
117 
118 	/* Clear the interrupt */
119 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR,
120 		A6XX_GMU_GMU2HOST_INTR_INFO_MSGQ);
121 
122 	for (;;) {
123 		struct a6xx_hfi_msg_response resp;
124 
125 		/* Get the next packet */
126 		ret = a6xx_hfi_queue_read(gmu, queue, (u32 *) &resp,
127 			sizeof(resp) >> 2);
128 
129 		/* If the queue is empty our response never made it */
130 		if (!ret) {
131 			DRM_DEV_ERROR(gmu->dev,
132 				"The HFI response queue is unexpectedly empty\n");
133 
134 			return -ENOENT;
135 		}
136 
137 		if (HFI_HEADER_ID(resp.header) == HFI_F2H_MSG_ERROR) {
138 			struct a6xx_hfi_msg_error *error =
139 				(struct a6xx_hfi_msg_error *) &resp;
140 
141 			DRM_DEV_ERROR(gmu->dev, "GMU firmware error %d\n",
142 				error->code);
143 			continue;
144 		}
145 
146 		if (seqnum != HFI_HEADER_SEQNUM(resp.ret_header)) {
147 			DRM_DEV_ERROR(gmu->dev,
148 				"Unexpected message id %d on the response queue\n",
149 				HFI_HEADER_SEQNUM(resp.ret_header));
150 			continue;
151 		}
152 
153 		if (resp.error) {
154 			DRM_DEV_ERROR(gmu->dev,
155 				"Message %s id %d returned error %d\n",
156 				a6xx_hfi_msg_id[id], seqnum, resp.error);
157 			return -EINVAL;
158 		}
159 
160 		/* All is well, copy over the buffer */
161 		if (payload && payload_size)
162 			memcpy(payload, resp.payload,
163 				min_t(u32, payload_size, sizeof(resp.payload)));
164 
165 		return 0;
166 	}
167 }
168 
169 static int a6xx_hfi_send_msg(struct a6xx_gmu *gmu, int id,
170 		void *data, u32 size, u32 *payload, u32 payload_size)
171 {
172 	struct a6xx_hfi_queue *queue = &gmu->queues[HFI_COMMAND_QUEUE];
173 	int ret, dwords = size >> 2;
174 	u32 seqnum;
175 
176 	seqnum = atomic_inc_return(&queue->seqnum) % 0xfff;
177 
178 	/* First dword of the message is the message header - fill it in */
179 	*((u32 *) data) = (seqnum << 20) | (HFI_MSG_CMD << 16) |
180 		(dwords << 8) | id;
181 
182 	ret = a6xx_hfi_queue_write(gmu, queue, data, dwords);
183 	if (ret) {
184 		DRM_DEV_ERROR(gmu->dev, "Unable to send message %s id %d\n",
185 			a6xx_hfi_msg_id[id], seqnum);
186 		return ret;
187 	}
188 
189 	return a6xx_hfi_wait_for_ack(gmu, id, seqnum, payload, payload_size);
190 }
191 
192 static int a6xx_hfi_send_gmu_init(struct a6xx_gmu *gmu, int boot_state)
193 {
194 	struct a6xx_hfi_msg_gmu_init_cmd msg = { 0 };
195 
196 	msg.dbg_buffer_addr = (u32) gmu->debug.iova;
197 	msg.dbg_buffer_size = (u32) gmu->debug.size;
198 	msg.boot_state = boot_state;
199 
200 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_INIT, &msg, sizeof(msg),
201 		NULL, 0);
202 }
203 
204 static int a6xx_hfi_get_fw_version(struct a6xx_gmu *gmu, u32 *version)
205 {
206 	struct a6xx_hfi_msg_fw_version msg = { 0 };
207 
208 	/* Currently supporting version 1.1 */
209 	msg.supported_version = (1 << 28) | (1 << 16);
210 
211 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_FW_VERSION, &msg, sizeof(msg),
212 		version, sizeof(*version));
213 }
214 
215 static int a6xx_hfi_send_perf_table_v1(struct a6xx_gmu *gmu)
216 {
217 	struct a6xx_hfi_msg_perf_table_v1 msg = { 0 };
218 	int i;
219 
220 	msg.num_gpu_levels = gmu->nr_gpu_freqs;
221 	msg.num_gmu_levels = gmu->nr_gmu_freqs;
222 
223 	for (i = 0; i < gmu->nr_gpu_freqs; i++) {
224 		msg.gx_votes[i].vote = gmu->gx_arc_votes[i];
225 		msg.gx_votes[i].freq = gmu->gpu_freqs[i] / 1000;
226 	}
227 
228 	for (i = 0; i < gmu->nr_gmu_freqs; i++) {
229 		msg.cx_votes[i].vote = gmu->cx_arc_votes[i];
230 		msg.cx_votes[i].freq = gmu->gmu_freqs[i] / 1000;
231 	}
232 
233 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_PERF_TABLE, &msg, sizeof(msg),
234 		NULL, 0);
235 }
236 
237 static int a6xx_hfi_send_perf_table(struct a6xx_gmu *gmu)
238 {
239 	struct a6xx_hfi_msg_perf_table msg = { 0 };
240 	int i;
241 
242 	msg.num_gpu_levels = gmu->nr_gpu_freqs;
243 	msg.num_gmu_levels = gmu->nr_gmu_freqs;
244 
245 	for (i = 0; i < gmu->nr_gpu_freqs; i++) {
246 		msg.gx_votes[i].vote = gmu->gx_arc_votes[i];
247 		msg.gx_votes[i].acd = 0xffffffff;
248 		msg.gx_votes[i].freq = gmu->gpu_freqs[i] / 1000;
249 	}
250 
251 	for (i = 0; i < gmu->nr_gmu_freqs; i++) {
252 		msg.cx_votes[i].vote = gmu->cx_arc_votes[i];
253 		msg.cx_votes[i].freq = gmu->gmu_freqs[i] / 1000;
254 	}
255 
256 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_PERF_TABLE, &msg, sizeof(msg),
257 		NULL, 0);
258 }
259 
260 static void a618_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
261 {
262 	/* Send a single "off" entry since the 618 GMU doesn't do bus scaling */
263 	msg->bw_level_num = 1;
264 
265 	msg->ddr_cmds_num = 3;
266 	msg->ddr_wait_bitmask = 0x01;
267 
268 	msg->ddr_cmds_addrs[0] = 0x50000;
269 	msg->ddr_cmds_addrs[1] = 0x5003c;
270 	msg->ddr_cmds_addrs[2] = 0x5000c;
271 
272 	msg->ddr_cmds_data[0][0] =  0x40000000;
273 	msg->ddr_cmds_data[0][1] =  0x40000000;
274 	msg->ddr_cmds_data[0][2] =  0x40000000;
275 
276 	/*
277 	 * These are the CX (CNOC) votes - these are used by the GMU but the
278 	 * votes are known and fixed for the target
279 	 */
280 	msg->cnoc_cmds_num = 1;
281 	msg->cnoc_wait_bitmask = 0x01;
282 
283 	msg->cnoc_cmds_addrs[0] = 0x5007c;
284 	msg->cnoc_cmds_data[0][0] =  0x40000000;
285 	msg->cnoc_cmds_data[1][0] =  0x60000001;
286 }
287 
288 static void a640_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
289 {
290 	/*
291 	 * Send a single "off" entry just to get things running
292 	 * TODO: bus scaling
293 	 */
294 	msg->bw_level_num = 1;
295 
296 	msg->ddr_cmds_num = 3;
297 	msg->ddr_wait_bitmask = 0x01;
298 
299 	msg->ddr_cmds_addrs[0] = 0x50000;
300 	msg->ddr_cmds_addrs[1] = 0x5003c;
301 	msg->ddr_cmds_addrs[2] = 0x5000c;
302 
303 	msg->ddr_cmds_data[0][0] =  0x40000000;
304 	msg->ddr_cmds_data[0][1] =  0x40000000;
305 	msg->ddr_cmds_data[0][2] =  0x40000000;
306 
307 	/*
308 	 * These are the CX (CNOC) votes - these are used by the GMU but the
309 	 * votes are known and fixed for the target
310 	 */
311 	msg->cnoc_cmds_num = 3;
312 	msg->cnoc_wait_bitmask = 0x01;
313 
314 	msg->cnoc_cmds_addrs[0] = 0x50034;
315 	msg->cnoc_cmds_addrs[1] = 0x5007c;
316 	msg->cnoc_cmds_addrs[2] = 0x5004c;
317 
318 	msg->cnoc_cmds_data[0][0] =  0x40000000;
319 	msg->cnoc_cmds_data[0][1] =  0x00000000;
320 	msg->cnoc_cmds_data[0][2] =  0x40000000;
321 
322 	msg->cnoc_cmds_data[1][0] =  0x60000001;
323 	msg->cnoc_cmds_data[1][1] =  0x20000001;
324 	msg->cnoc_cmds_data[1][2] =  0x60000001;
325 }
326 
327 static void a650_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
328 {
329 	/*
330 	 * Send a single "off" entry just to get things running
331 	 * TODO: bus scaling
332 	 */
333 	msg->bw_level_num = 1;
334 
335 	msg->ddr_cmds_num = 3;
336 	msg->ddr_wait_bitmask = 0x01;
337 
338 	msg->ddr_cmds_addrs[0] = 0x50000;
339 	msg->ddr_cmds_addrs[1] = 0x50004;
340 	msg->ddr_cmds_addrs[2] = 0x5007c;
341 
342 	msg->ddr_cmds_data[0][0] =  0x40000000;
343 	msg->ddr_cmds_data[0][1] =  0x40000000;
344 	msg->ddr_cmds_data[0][2] =  0x40000000;
345 
346 	/*
347 	 * These are the CX (CNOC) votes - these are used by the GMU but the
348 	 * votes are known and fixed for the target
349 	 */
350 	msg->cnoc_cmds_num = 1;
351 	msg->cnoc_wait_bitmask = 0x01;
352 
353 	msg->cnoc_cmds_addrs[0] = 0x500a4;
354 	msg->cnoc_cmds_data[0][0] =  0x40000000;
355 	msg->cnoc_cmds_data[1][0] =  0x60000001;
356 }
357 
358 static void a660_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
359 {
360 	/*
361 	 * Send a single "off" entry just to get things running
362 	 * TODO: bus scaling
363 	 */
364 	msg->bw_level_num = 1;
365 
366 	msg->ddr_cmds_num = 3;
367 	msg->ddr_wait_bitmask = 0x01;
368 
369 	msg->ddr_cmds_addrs[0] = 0x50004;
370 	msg->ddr_cmds_addrs[1] = 0x500a0;
371 	msg->ddr_cmds_addrs[2] = 0x50000;
372 
373 	msg->ddr_cmds_data[0][0] =  0x40000000;
374 	msg->ddr_cmds_data[0][1] =  0x40000000;
375 	msg->ddr_cmds_data[0][2] =  0x40000000;
376 
377 	/*
378 	 * These are the CX (CNOC) votes - these are used by the GMU but the
379 	 * votes are known and fixed for the target
380 	 */
381 	msg->cnoc_cmds_num = 1;
382 	msg->cnoc_wait_bitmask = 0x01;
383 
384 	msg->cnoc_cmds_addrs[0] = 0x50070;
385 	msg->cnoc_cmds_data[0][0] =  0x40000000;
386 	msg->cnoc_cmds_data[1][0] =  0x60000001;
387 }
388 
389 static void adreno_7c3_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
390 {
391 	/*
392 	 * Send a single "off" entry just to get things running
393 	 * TODO: bus scaling
394 	 */
395 	msg->bw_level_num = 1;
396 
397 	msg->ddr_cmds_num = 3;
398 	msg->ddr_wait_bitmask = 0x07;
399 
400 	msg->ddr_cmds_addrs[0] = 0x50004;
401 	msg->ddr_cmds_addrs[1] = 0x50000;
402 	msg->ddr_cmds_addrs[2] = 0x50088;
403 
404 	msg->ddr_cmds_data[0][0] =  0x40000000;
405 	msg->ddr_cmds_data[0][1] =  0x40000000;
406 	msg->ddr_cmds_data[0][2] =  0x40000000;
407 
408 	/*
409 	 * These are the CX (CNOC) votes - these are used by the GMU but the
410 	 * votes are known and fixed for the target
411 	 */
412 	msg->cnoc_cmds_num = 1;
413 	msg->cnoc_wait_bitmask = 0x01;
414 
415 	msg->cnoc_cmds_addrs[0] = 0x5006c;
416 	msg->cnoc_cmds_data[0][0] =  0x40000000;
417 	msg->cnoc_cmds_data[1][0] =  0x60000001;
418 }
419 static void a6xx_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
420 {
421 	/* Send a single "off" entry since the 630 GMU doesn't do bus scaling */
422 	msg->bw_level_num = 1;
423 
424 	msg->ddr_cmds_num = 3;
425 	msg->ddr_wait_bitmask = 0x07;
426 
427 	msg->ddr_cmds_addrs[0] = 0x50000;
428 	msg->ddr_cmds_addrs[1] = 0x5005c;
429 	msg->ddr_cmds_addrs[2] = 0x5000c;
430 
431 	msg->ddr_cmds_data[0][0] =  0x40000000;
432 	msg->ddr_cmds_data[0][1] =  0x40000000;
433 	msg->ddr_cmds_data[0][2] =  0x40000000;
434 
435 	/*
436 	 * These are the CX (CNOC) votes.  This is used but the values for the
437 	 * sdm845 GMU are known and fixed so we can hard code them.
438 	 */
439 
440 	msg->cnoc_cmds_num = 3;
441 	msg->cnoc_wait_bitmask = 0x05;
442 
443 	msg->cnoc_cmds_addrs[0] = 0x50034;
444 	msg->cnoc_cmds_addrs[1] = 0x5007c;
445 	msg->cnoc_cmds_addrs[2] = 0x5004c;
446 
447 	msg->cnoc_cmds_data[0][0] =  0x40000000;
448 	msg->cnoc_cmds_data[0][1] =  0x00000000;
449 	msg->cnoc_cmds_data[0][2] =  0x40000000;
450 
451 	msg->cnoc_cmds_data[1][0] =  0x60000001;
452 	msg->cnoc_cmds_data[1][1] =  0x20000001;
453 	msg->cnoc_cmds_data[1][2] =  0x60000001;
454 }
455 
456 
457 static int a6xx_hfi_send_bw_table(struct a6xx_gmu *gmu)
458 {
459 	struct a6xx_hfi_msg_bw_table msg = { 0 };
460 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
461 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
462 
463 	if (adreno_is_a618(adreno_gpu))
464 		a618_build_bw_table(&msg);
465 	else if (adreno_is_a640_family(adreno_gpu))
466 		a640_build_bw_table(&msg);
467 	else if (adreno_is_a650(adreno_gpu))
468 		a650_build_bw_table(&msg);
469 	else if (adreno_is_7c3(adreno_gpu))
470 		adreno_7c3_build_bw_table(&msg);
471 	else if (adreno_is_a660(adreno_gpu))
472 		a660_build_bw_table(&msg);
473 	else
474 		a6xx_build_bw_table(&msg);
475 
476 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_BW_TABLE, &msg, sizeof(msg),
477 		NULL, 0);
478 }
479 
480 static int a6xx_hfi_send_test(struct a6xx_gmu *gmu)
481 {
482 	struct a6xx_hfi_msg_test msg = { 0 };
483 
484 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_TEST, &msg, sizeof(msg),
485 		NULL, 0);
486 }
487 
488 static int a6xx_hfi_send_start(struct a6xx_gmu *gmu)
489 {
490 	struct a6xx_hfi_msg_start msg = { 0 };
491 
492 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_START, &msg, sizeof(msg),
493 		NULL, 0);
494 }
495 
496 static int a6xx_hfi_send_core_fw_start(struct a6xx_gmu *gmu)
497 {
498 	struct a6xx_hfi_msg_core_fw_start msg = { 0 };
499 
500 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_CORE_FW_START, &msg,
501 		sizeof(msg), NULL, 0);
502 }
503 
504 int a6xx_hfi_set_freq(struct a6xx_gmu *gmu, int index)
505 {
506 	struct a6xx_hfi_gx_bw_perf_vote_cmd msg = { 0 };
507 
508 	msg.ack_type = 1; /* blocking */
509 	msg.freq = index;
510 	msg.bw = 0; /* TODO: bus scaling */
511 
512 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_GX_BW_PERF_VOTE, &msg,
513 		sizeof(msg), NULL, 0);
514 }
515 
516 int a6xx_hfi_send_prep_slumber(struct a6xx_gmu *gmu)
517 {
518 	struct a6xx_hfi_prep_slumber_cmd msg = { 0 };
519 
520 	/* TODO: should freq and bw fields be non-zero ? */
521 
522 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_PREPARE_SLUMBER, &msg,
523 		sizeof(msg), NULL, 0);
524 }
525 
526 static int a6xx_hfi_start_v1(struct a6xx_gmu *gmu, int boot_state)
527 {
528 	int ret;
529 
530 	ret = a6xx_hfi_send_gmu_init(gmu, boot_state);
531 	if (ret)
532 		return ret;
533 
534 	ret = a6xx_hfi_get_fw_version(gmu, NULL);
535 	if (ret)
536 		return ret;
537 
538 	/*
539 	 * We have to get exchange version numbers per the sequence but at this
540 	 * point th kernel driver doesn't need to know the exact version of
541 	 * the GMU firmware
542 	 */
543 
544 	ret = a6xx_hfi_send_perf_table_v1(gmu);
545 	if (ret)
546 		return ret;
547 
548 	ret = a6xx_hfi_send_bw_table(gmu);
549 	if (ret)
550 		return ret;
551 
552 	/*
553 	 * Let the GMU know that there won't be any more HFI messages until next
554 	 * boot
555 	 */
556 	a6xx_hfi_send_test(gmu);
557 
558 	return 0;
559 }
560 
561 int a6xx_hfi_start(struct a6xx_gmu *gmu, int boot_state)
562 {
563 	int ret;
564 
565 	if (gmu->legacy)
566 		return a6xx_hfi_start_v1(gmu, boot_state);
567 
568 
569 	ret = a6xx_hfi_send_perf_table(gmu);
570 	if (ret)
571 		return ret;
572 
573 	ret = a6xx_hfi_send_bw_table(gmu);
574 	if (ret)
575 		return ret;
576 
577 	ret = a6xx_hfi_send_core_fw_start(gmu);
578 	if (ret)
579 		return ret;
580 
581 	/*
582 	 * Downstream driver sends this in its "a6xx_hw_init" equivalent,
583 	 * but seems to be no harm in sending it here
584 	 */
585 	ret = a6xx_hfi_send_start(gmu);
586 	if (ret)
587 		return ret;
588 
589 	return 0;
590 }
591 
592 void a6xx_hfi_stop(struct a6xx_gmu *gmu)
593 {
594 	int i;
595 
596 	for (i = 0; i < ARRAY_SIZE(gmu->queues); i++) {
597 		struct a6xx_hfi_queue *queue = &gmu->queues[i];
598 
599 		if (!queue->header)
600 			continue;
601 
602 		if (queue->header->read_index != queue->header->write_index)
603 			DRM_DEV_ERROR(gmu->dev, "HFI queue %d is not empty\n", i);
604 
605 		queue->header->read_index = 0;
606 		queue->header->write_index = 0;
607 
608 		memset(&queue->history, 0xff, sizeof(queue->history));
609 		queue->history_idx = 0;
610 	}
611 }
612 
613 static void a6xx_hfi_queue_init(struct a6xx_hfi_queue *queue,
614 		struct a6xx_hfi_queue_header *header, void *virt, u64 iova,
615 		u32 id)
616 {
617 	spin_lock_init(&queue->lock);
618 	queue->header = header;
619 	queue->data = virt;
620 	atomic_set(&queue->seqnum, 0);
621 
622 	memset(&queue->history, 0xff, sizeof(queue->history));
623 	queue->history_idx = 0;
624 
625 	/* Set up the shared memory header */
626 	header->iova = iova;
627 	header->type =  10 << 8 | id;
628 	header->status = 1;
629 	header->size = SZ_4K >> 2;
630 	header->msg_size = 0;
631 	header->dropped = 0;
632 	header->rx_watermark = 1;
633 	header->tx_watermark = 1;
634 	header->rx_request = 1;
635 	header->tx_request = 0;
636 	header->read_index = 0;
637 	header->write_index = 0;
638 }
639 
640 void a6xx_hfi_init(struct a6xx_gmu *gmu)
641 {
642 	struct a6xx_gmu_bo *hfi = &gmu->hfi;
643 	struct a6xx_hfi_queue_table_header *table = hfi->virt;
644 	struct a6xx_hfi_queue_header *headers = hfi->virt + sizeof(*table);
645 	u64 offset;
646 	int table_size;
647 
648 	/*
649 	 * The table size is the size of the table header plus all of the queue
650 	 * headers
651 	 */
652 	table_size = sizeof(*table);
653 	table_size += (ARRAY_SIZE(gmu->queues) *
654 		sizeof(struct a6xx_hfi_queue_header));
655 
656 	table->version = 0;
657 	table->size = table_size;
658 	/* First queue header is located immediately after the table header */
659 	table->qhdr0_offset = sizeof(*table) >> 2;
660 	table->qhdr_size = sizeof(struct a6xx_hfi_queue_header) >> 2;
661 	table->num_queues = ARRAY_SIZE(gmu->queues);
662 	table->active_queues = ARRAY_SIZE(gmu->queues);
663 
664 	/* Command queue */
665 	offset = SZ_4K;
666 	a6xx_hfi_queue_init(&gmu->queues[0], &headers[0], hfi->virt + offset,
667 		hfi->iova + offset, 0);
668 
669 	/* GMU response queue */
670 	offset += SZ_4K;
671 	a6xx_hfi_queue_init(&gmu->queues[1], &headers[1], hfi->virt + offset,
672 		hfi->iova + offset, gmu->legacy ? 4 : 1);
673 }
674