xref: /linux/drivers/gpu/drm/msm/adreno/a6xx_hfi.c (revision 2c142b63c8ee982cdfdba49a616027c266294838)
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 <soc/qcom/cmd-db.h>
9 #include <soc/qcom/tcs.h>
10 
11 #include "a6xx_gmu.h"
12 #include "a6xx_gmu.xml.h"
13 #include "a6xx_gpu.h"
14 
15 #define HFI_MSG_ID(val) [val] = #val
16 
17 static const char * const a6xx_hfi_msg_id[] = {
18 	HFI_MSG_ID(HFI_H2F_MSG_INIT),
19 	HFI_MSG_ID(HFI_H2F_MSG_FW_VERSION),
20 	HFI_MSG_ID(HFI_H2F_MSG_BW_TABLE),
21 	HFI_MSG_ID(HFI_H2F_MSG_PERF_TABLE),
22 	HFI_MSG_ID(HFI_H2F_MSG_TEST),
23 	HFI_MSG_ID(HFI_H2F_MSG_START),
24 	HFI_MSG_ID(HFI_H2F_FEATURE_CTRL),
25 	HFI_MSG_ID(HFI_H2F_MSG_CORE_FW_START),
26 	HFI_MSG_ID(HFI_H2F_MSG_TABLE),
27 	HFI_MSG_ID(HFI_H2F_MSG_GX_BW_PERF_VOTE),
28 	HFI_MSG_ID(HFI_H2F_MSG_PREPARE_SLUMBER),
29 };
30 
a6xx_hfi_queue_read(struct a6xx_gmu * gmu,struct a6xx_hfi_queue * queue,u32 * data,u32 dwords)31 static int a6xx_hfi_queue_read(struct a6xx_gmu *gmu,
32 	struct a6xx_hfi_queue *queue, u32 *data, u32 dwords)
33 {
34 	struct a6xx_hfi_queue_header *header = queue->header;
35 	u32 i, hdr, index = header->read_index;
36 
37 	if (header->read_index == READ_ONCE(header->write_index)) {
38 		header->rx_request = 1;
39 		return 0;
40 	}
41 
42 	hdr = queue->data[index];
43 
44 	queue->history[(queue->history_idx++) % HFI_HISTORY_SZ] = index;
45 
46 	/*
47 	 * If we are to assume that the GMU firmware is in fact a rational actor
48 	 * and is programmed to not send us a larger response than we expect
49 	 * then we can also assume that if the header size is unexpectedly large
50 	 * that it is due to memory corruption and/or hardware failure. In this
51 	 * case the only reasonable course of action is to BUG() to help harden
52 	 * the failure.
53 	 */
54 
55 	BUG_ON(HFI_HEADER_SIZE(hdr) > dwords);
56 
57 	for (i = 0; i < HFI_HEADER_SIZE(hdr); i++) {
58 		data[i] = queue->data[index];
59 		index = (index + 1) % header->size;
60 	}
61 
62 	if (!gmu->legacy)
63 		index = ALIGN(index, 4) % header->size;
64 
65 	/* Ensure all memory operations are complete before updating the read index */
66 	dma_mb();
67 
68 	WRITE_ONCE(header->read_index, index);
69 	return HFI_HEADER_SIZE(hdr);
70 }
71 
a6xx_hfi_queue_write(struct a6xx_gmu * gmu,struct a6xx_hfi_queue * queue,u32 * data,u32 dwords)72 static int a6xx_hfi_queue_write(struct a6xx_gmu *gmu,
73 	struct a6xx_hfi_queue *queue, u32 *data, u32 dwords)
74 {
75 	struct a6xx_hfi_queue_header *header = queue->header;
76 	u32 i, space, index = header->write_index;
77 
78 	spin_lock(&queue->lock);
79 
80 	space = CIRC_SPACE(header->write_index, READ_ONCE(header->read_index),
81 		header->size);
82 	if (space < dwords) {
83 		header->dropped++;
84 		spin_unlock(&queue->lock);
85 		return -ENOSPC;
86 	}
87 
88 	queue->history[(queue->history_idx++) % HFI_HISTORY_SZ] = index;
89 
90 	for (i = 0; i < dwords; i++) {
91 		queue->data[index] = data[i];
92 		index = (index + 1) % header->size;
93 	}
94 
95 	/* Cookify any non used data at the end of the write buffer */
96 	if (!gmu->legacy) {
97 		for (; index % 4; index = (index + 1) % header->size)
98 			queue->data[index] = 0xfafafafa;
99 	}
100 
101 	/* Ensure all memory operations are complete before updating the write index */
102 	dma_mb();
103 
104 	WRITE_ONCE(header->write_index, index);
105 	spin_unlock(&queue->lock);
106 
107 	gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET, 0x01);
108 	return 0;
109 }
110 
a6xx_hfi_wait_for_msg_interrupt(struct a6xx_gmu * gmu,u32 id,u32 seqnum)111 static int a6xx_hfi_wait_for_msg_interrupt(struct a6xx_gmu *gmu, u32 id, u32 seqnum)
112 {
113 	int ret;
114 	u32 val;
115 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
116 
117 	do {
118 		/* Wait for a response */
119 		ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO, val,
120 			val & A6XX_GMU_GMU2HOST_INTR_INFO_MSGQ, 100, 1000000);
121 
122 		if (!ret)
123 			break;
124 
125 		if (completion_done(&a6xx_gpu->base.fault_coredump_done))
126 			break;
127 
128 		/* We may timeout because the GMU is temporarily wedged from
129 		 * pending faults from the GPU and we are taking a devcoredump.
130 		 * Wait until the MMU is resumed and try again.
131 		 */
132 		wait_for_completion(&a6xx_gpu->base.fault_coredump_done);
133 	} while (true);
134 
135 	if (ret) {
136 		DRM_DEV_ERROR(gmu->dev,
137 			"Message %s id %d timed out waiting for response\n",
138 			a6xx_hfi_msg_id[id], seqnum);
139 		return -ETIMEDOUT;
140 	}
141 
142 	/* Clear the interrupt */
143 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR,
144 		A6XX_GMU_GMU2HOST_INTR_INFO_MSGQ);
145 
146 	return 0;
147 }
148 
a6xx_hfi_wait_for_ack(struct a6xx_gmu * gmu,u32 id,u32 seqnum,u32 * payload,u32 payload_size)149 static int a6xx_hfi_wait_for_ack(struct a6xx_gmu *gmu, u32 id, u32 seqnum,
150 		u32 *payload, u32 payload_size)
151 {
152 	struct a6xx_hfi_queue *queue = &gmu->queues[HFI_RESPONSE_QUEUE];
153 	int ret;
154 
155 	ret = a6xx_hfi_wait_for_msg_interrupt(gmu, id, seqnum);
156 	if (ret)
157 		return ret;
158 
159 	for (;;) {
160 		struct a6xx_hfi_msg_response resp;
161 
162 		/* Get the next packet */
163 		ret = a6xx_hfi_queue_read(gmu, queue, (u32 *) &resp,
164 			sizeof(resp) >> 2);
165 
166 		/* If the queue is empty, there may have been previous missed
167 		 * responses that preceded the response to our packet. Wait
168 		 * further before we give up.
169 		 */
170 		if (!ret) {
171 			ret = a6xx_hfi_wait_for_msg_interrupt(gmu, id, seqnum);
172 			if (ret) {
173 				DRM_DEV_ERROR(gmu->dev,
174 					"The HFI response queue is unexpectedly empty\n");
175 				return ret;
176 			}
177 			continue;
178 		}
179 
180 		if (HFI_HEADER_ID(resp.header) == HFI_F2H_MSG_ERROR) {
181 			struct a6xx_hfi_msg_error *error =
182 				(struct a6xx_hfi_msg_error *) &resp;
183 
184 			DRM_DEV_ERROR(gmu->dev, "GMU firmware error %d\n",
185 				error->code);
186 			continue;
187 		}
188 
189 		if (seqnum != HFI_HEADER_SEQNUM(resp.ret_header)) {
190 			DRM_DEV_ERROR(gmu->dev,
191 				"Unexpected message id %d on the response queue\n",
192 				HFI_HEADER_SEQNUM(resp.ret_header));
193 			continue;
194 		}
195 
196 		if (resp.error) {
197 			DRM_DEV_ERROR(gmu->dev,
198 				"Message %s id %d returned error %d\n",
199 				a6xx_hfi_msg_id[id], seqnum, resp.error);
200 			return -EINVAL;
201 		}
202 
203 		/* All is well, copy over the buffer */
204 		if (payload && payload_size)
205 			memcpy(payload, resp.payload,
206 				min_t(u32, payload_size, sizeof(resp.payload)));
207 
208 		return 0;
209 	}
210 }
211 
a6xx_hfi_send_msg(struct a6xx_gmu * gmu,int id,void * data,u32 size,u32 * payload,u32 payload_size)212 static int a6xx_hfi_send_msg(struct a6xx_gmu *gmu, int id,
213 		void *data, u32 size, u32 *payload, u32 payload_size)
214 {
215 	struct a6xx_hfi_queue *queue = &gmu->queues[HFI_COMMAND_QUEUE];
216 	int ret, dwords = size >> 2;
217 	u32 seqnum;
218 
219 	seqnum = atomic_inc_return(&queue->seqnum) % 0xfff;
220 
221 	/* First dword of the message is the message header - fill it in */
222 	*((u32 *) data) = (seqnum << 20) | (HFI_MSG_CMD << 16) |
223 		(dwords << 8) | id;
224 
225 	ret = a6xx_hfi_queue_write(gmu, queue, data, dwords);
226 	if (ret) {
227 		DRM_DEV_ERROR(gmu->dev, "Unable to send message %s id %d\n",
228 			a6xx_hfi_msg_id[id], seqnum);
229 		return ret;
230 	}
231 
232 	return a6xx_hfi_wait_for_ack(gmu, id, seqnum, payload, payload_size);
233 }
234 
a6xx_hfi_send_gmu_init(struct a6xx_gmu * gmu,int boot_state)235 static int a6xx_hfi_send_gmu_init(struct a6xx_gmu *gmu, int boot_state)
236 {
237 	struct a6xx_hfi_msg_gmu_init_cmd msg = { 0 };
238 
239 	msg.dbg_buffer_addr = (u32) gmu->debug.iova;
240 	msg.dbg_buffer_size = (u32) gmu->debug.size;
241 	msg.boot_state = boot_state;
242 
243 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_INIT, &msg, sizeof(msg),
244 		NULL, 0);
245 }
246 
a6xx_hfi_get_fw_version(struct a6xx_gmu * gmu,u32 * version)247 static int a6xx_hfi_get_fw_version(struct a6xx_gmu *gmu, u32 *version)
248 {
249 	struct a6xx_hfi_msg_fw_version msg = { 0 };
250 
251 	/* Currently supporting version 1.10 */
252 	msg.supported_version = (1 << 28) | (1 << 19) | (1 << 17);
253 
254 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_FW_VERSION, &msg, sizeof(msg),
255 		version, sizeof(*version));
256 }
257 
a6xx_hfi_send_perf_table_v1(struct a6xx_gmu * gmu)258 static int a6xx_hfi_send_perf_table_v1(struct a6xx_gmu *gmu)
259 {
260 	struct a6xx_hfi_msg_perf_table_v1 msg = { 0 };
261 	int i;
262 
263 	msg.num_gpu_levels = gmu->nr_gpu_freqs;
264 	msg.num_gmu_levels = gmu->nr_gmu_freqs;
265 
266 	for (i = 0; i < gmu->nr_gpu_freqs; i++) {
267 		msg.gx_votes[i].vote = gmu->gx_arc_votes[i];
268 		msg.gx_votes[i].freq = gmu->gpu_freqs[i] / 1000;
269 	}
270 
271 	for (i = 0; i < gmu->nr_gmu_freqs; i++) {
272 		msg.cx_votes[i].vote = gmu->cx_arc_votes[i];
273 		msg.cx_votes[i].freq = gmu->gmu_freqs[i] / 1000;
274 	}
275 
276 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_PERF_TABLE, &msg, sizeof(msg),
277 		NULL, 0);
278 }
279 
a8xx_hfi_send_perf_table(struct a6xx_gmu * gmu)280 static int a8xx_hfi_send_perf_table(struct a6xx_gmu *gmu)
281 {
282 	unsigned int num_gx_votes = 3, num_cx_votes = 2;
283 	struct a6xx_hfi_table_entry *entry;
284 	struct a6xx_hfi_table *tbl;
285 	int ret, i;
286 	u32 size;
287 
288 	size = sizeof(*tbl) +  (2 * sizeof(tbl->entry[0])) +
289 		(gmu->nr_gpu_freqs * num_gx_votes * sizeof(gmu->gx_arc_votes[0])) +
290 		(gmu->nr_gmu_freqs * num_cx_votes * sizeof(gmu->cx_arc_votes[0]));
291 	tbl = kzalloc(size, GFP_KERNEL);
292 	if (!tbl)
293 		return -ENOMEM;
294 	tbl->type = HFI_TABLE_GPU_PERF;
295 
296 	/* First fill GX votes */
297 	entry = &tbl->entry[0];
298 	entry->count = gmu->nr_gpu_freqs;
299 	entry->stride = num_gx_votes;
300 
301 	for (i = 0; i < gmu->nr_gpu_freqs; i++) {
302 		unsigned int base = i * entry->stride;
303 
304 		entry->data[base+0] = gmu->gx_arc_votes[i];
305 		entry->data[base+1] = gmu->dep_arc_votes[i];
306 		entry->data[base+2] = gmu->gpu_freqs[i] / 1000;
307 	}
308 
309 	/* Then fill CX votes */
310 	entry = (struct a6xx_hfi_table_entry *)
311 		&tbl->entry[0].data[gmu->nr_gpu_freqs * num_gx_votes];
312 
313 	entry->count = gmu->nr_gmu_freqs;
314 	entry->stride = num_cx_votes;
315 
316 	for (i = 0; i < gmu->nr_gmu_freqs; i++) {
317 		unsigned int base = i * entry->stride;
318 
319 		entry->data[base] = gmu->cx_arc_votes[i];
320 		entry->data[base+1] = gmu->gmu_freqs[i] / 1000;
321 	}
322 
323 	ret = a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_TABLE, tbl, size, NULL, 0);
324 
325 	kfree(tbl);
326 	return ret;
327 }
328 
a6xx_hfi_send_perf_table(struct a6xx_gmu * gmu)329 static int a6xx_hfi_send_perf_table(struct a6xx_gmu *gmu)
330 {
331 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
332 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
333 	struct a6xx_hfi_msg_perf_table msg = { 0 };
334 	int i;
335 
336 	if (adreno_is_a8xx(adreno_gpu))
337 		return a8xx_hfi_send_perf_table(gmu);
338 
339 	msg.num_gpu_levels = gmu->nr_gpu_freqs;
340 	msg.num_gmu_levels = gmu->nr_gmu_freqs;
341 
342 	for (i = 0; i < gmu->nr_gpu_freqs; i++) {
343 		msg.gx_votes[i].vote = gmu->gx_arc_votes[i];
344 		msg.gx_votes[i].acd = 0xffffffff;
345 		msg.gx_votes[i].freq = gmu->gpu_freqs[i] / 1000;
346 	}
347 
348 	for (i = 0; i < gmu->nr_gmu_freqs; i++) {
349 		msg.cx_votes[i].vote = gmu->cx_arc_votes[i];
350 		msg.cx_votes[i].freq = gmu->gmu_freqs[i] / 1000;
351 	}
352 
353 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_PERF_TABLE, &msg, sizeof(msg),
354 		NULL, 0);
355 }
356 
a6xx_generate_bw_table(const struct a6xx_info * info,struct a6xx_gmu * gmu,struct a6xx_hfi_msg_bw_table * msg)357 static void a6xx_generate_bw_table(const struct a6xx_info *info, struct a6xx_gmu *gmu,
358 				   struct a6xx_hfi_msg_bw_table *msg)
359 {
360 	unsigned int i, j;
361 
362 	for (i = 0; i < GMU_MAX_BCMS; i++) {
363 		if (!info->bcms[i].name)
364 			break;
365 		msg->ddr_cmds_addrs[i] = cmd_db_read_addr(info->bcms[i].name);
366 	}
367 	msg->ddr_cmds_num = i;
368 
369 	for (i = 0; i < gmu->nr_gpu_bws; ++i)
370 		for (j = 0; j < msg->ddr_cmds_num; j++)
371 			msg->ddr_cmds_data[i][j] = gmu->gpu_ib_votes[i][j];
372 	msg->bw_level_num = gmu->nr_gpu_bws;
373 
374 	/* Compute the wait bitmask with each BCM having the commit bit */
375 	msg->ddr_wait_bitmask = 0;
376 	for (j = 0; j < msg->ddr_cmds_num; j++)
377 		if (msg->ddr_cmds_data[0][j] & BCM_TCS_CMD_COMMIT_MASK)
378 			msg->ddr_wait_bitmask |= BIT(j);
379 
380 	/*
381 	 * These are the CX (CNOC) votes - these are used by the GMU
382 	 * The 'CN0' BCM is used on all targets, and votes are basically
383 	 * 'off' and 'on' states with first bit to enable the path.
384 	 */
385 
386 	msg->cnoc_cmds_addrs[0] = cmd_db_read_addr("CN0");
387 	msg->cnoc_cmds_num = 1;
388 
389 	msg->cnoc_cmds_data[0][0] = BCM_TCS_CMD(true, false, 0, 0);
390 	msg->cnoc_cmds_data[1][0] = BCM_TCS_CMD(true, true, 0, BIT(0));
391 
392 	/* Compute the wait bitmask with each BCM having the commit bit */
393 	msg->cnoc_wait_bitmask = 0;
394 	for (j = 0; j < msg->cnoc_cmds_num; j++)
395 		if (msg->cnoc_cmds_data[0][j] & BCM_TCS_CMD_COMMIT_MASK)
396 			msg->cnoc_wait_bitmask |= BIT(j);
397 }
398 
a618_build_bw_table(struct a6xx_hfi_msg_bw_table * msg)399 static void a618_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
400 {
401 	/* Send a single "off" entry since the 618 GMU doesn't do bus scaling */
402 	msg->bw_level_num = 1;
403 
404 	msg->ddr_cmds_num = 3;
405 	msg->ddr_wait_bitmask = 0x01;
406 
407 	msg->ddr_cmds_addrs[0] = 0x50000;
408 	msg->ddr_cmds_addrs[1] = 0x5003c;
409 	msg->ddr_cmds_addrs[2] = 0x5000c;
410 
411 	msg->ddr_cmds_data[0][0] =  0x40000000;
412 	msg->ddr_cmds_data[0][1] =  0x40000000;
413 	msg->ddr_cmds_data[0][2] =  0x40000000;
414 
415 	/*
416 	 * These are the CX (CNOC) votes - these are used by the GMU but the
417 	 * votes are known and fixed for the target
418 	 */
419 	msg->cnoc_cmds_num = 1;
420 	msg->cnoc_wait_bitmask = 0x01;
421 
422 	msg->cnoc_cmds_addrs[0] = 0x5007c;
423 	msg->cnoc_cmds_data[0][0] =  0x40000000;
424 	msg->cnoc_cmds_data[1][0] =  0x60000001;
425 }
426 
a619_build_bw_table(struct a6xx_hfi_msg_bw_table * msg)427 static void a619_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
428 {
429 	msg->bw_level_num = 13;
430 
431 	msg->ddr_cmds_num = 3;
432 	msg->ddr_wait_bitmask = 0x0;
433 
434 	msg->ddr_cmds_addrs[0] = 0x50000;
435 	msg->ddr_cmds_addrs[1] = 0x50004;
436 	msg->ddr_cmds_addrs[2] = 0x50080;
437 
438 	msg->ddr_cmds_data[0][0]  = 0x40000000;
439 	msg->ddr_cmds_data[0][1]  = 0x40000000;
440 	msg->ddr_cmds_data[0][2]  = 0x40000000;
441 	msg->ddr_cmds_data[1][0]  = 0x6000030c;
442 	msg->ddr_cmds_data[1][1]  = 0x600000db;
443 	msg->ddr_cmds_data[1][2]  = 0x60000008;
444 	msg->ddr_cmds_data[2][0]  = 0x60000618;
445 	msg->ddr_cmds_data[2][1]  = 0x600001b6;
446 	msg->ddr_cmds_data[2][2]  = 0x60000008;
447 	msg->ddr_cmds_data[3][0]  = 0x60000925;
448 	msg->ddr_cmds_data[3][1]  = 0x60000291;
449 	msg->ddr_cmds_data[3][2]  = 0x60000008;
450 	msg->ddr_cmds_data[4][0]  = 0x60000dc1;
451 	msg->ddr_cmds_data[4][1]  = 0x600003dc;
452 	msg->ddr_cmds_data[4][2]  = 0x60000008;
453 	msg->ddr_cmds_data[5][0]  = 0x600010ad;
454 	msg->ddr_cmds_data[5][1]  = 0x600004ae;
455 	msg->ddr_cmds_data[5][2]  = 0x60000008;
456 	msg->ddr_cmds_data[6][0]  = 0x600014c3;
457 	msg->ddr_cmds_data[6][1]  = 0x600005d4;
458 	msg->ddr_cmds_data[6][2]  = 0x60000008;
459 	msg->ddr_cmds_data[7][0]  = 0x6000176a;
460 	msg->ddr_cmds_data[7][1]  = 0x60000693;
461 	msg->ddr_cmds_data[7][2]  = 0x60000008;
462 	msg->ddr_cmds_data[8][0]  = 0x60001f01;
463 	msg->ddr_cmds_data[8][1]  = 0x600008b5;
464 	msg->ddr_cmds_data[8][2]  = 0x60000008;
465 	msg->ddr_cmds_data[9][0]  = 0x60002940;
466 	msg->ddr_cmds_data[9][1]  = 0x60000b95;
467 	msg->ddr_cmds_data[9][2]  = 0x60000008;
468 	msg->ddr_cmds_data[10][0] = 0x60002f68;
469 	msg->ddr_cmds_data[10][1] = 0x60000d50;
470 	msg->ddr_cmds_data[10][2] = 0x60000008;
471 	msg->ddr_cmds_data[11][0] = 0x60003700;
472 	msg->ddr_cmds_data[11][1] = 0x60000f71;
473 	msg->ddr_cmds_data[11][2] = 0x60000008;
474 	msg->ddr_cmds_data[12][0] = 0x60003fce;
475 	msg->ddr_cmds_data[12][1] = 0x600011ea;
476 	msg->ddr_cmds_data[12][2] = 0x60000008;
477 
478 	msg->cnoc_cmds_num = 1;
479 	msg->cnoc_wait_bitmask = 0x0;
480 
481 	msg->cnoc_cmds_addrs[0] = 0x50054;
482 
483 	msg->cnoc_cmds_data[0][0] = 0x40000000;
484 }
485 
a640_build_bw_table(struct a6xx_hfi_msg_bw_table * msg)486 static void a640_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
487 {
488 	/*
489 	 * Send a single "off" entry just to get things running
490 	 * TODO: bus scaling
491 	 */
492 	msg->bw_level_num = 1;
493 
494 	msg->ddr_cmds_num = 3;
495 	msg->ddr_wait_bitmask = 0x01;
496 
497 	msg->ddr_cmds_addrs[0] = 0x50000;
498 	msg->ddr_cmds_addrs[1] = 0x5003c;
499 	msg->ddr_cmds_addrs[2] = 0x5000c;
500 
501 	msg->ddr_cmds_data[0][0] =  0x40000000;
502 	msg->ddr_cmds_data[0][1] =  0x40000000;
503 	msg->ddr_cmds_data[0][2] =  0x40000000;
504 
505 	/*
506 	 * These are the CX (CNOC) votes - these are used by the GMU but the
507 	 * votes are known and fixed for the target
508 	 */
509 	msg->cnoc_cmds_num = 3;
510 	msg->cnoc_wait_bitmask = 0x01;
511 
512 	msg->cnoc_cmds_addrs[0] = 0x50034;
513 	msg->cnoc_cmds_addrs[1] = 0x5007c;
514 	msg->cnoc_cmds_addrs[2] = 0x5004c;
515 
516 	msg->cnoc_cmds_data[0][0] =  0x40000000;
517 	msg->cnoc_cmds_data[0][1] =  0x00000000;
518 	msg->cnoc_cmds_data[0][2] =  0x40000000;
519 
520 	msg->cnoc_cmds_data[1][0] =  0x60000001;
521 	msg->cnoc_cmds_data[1][1] =  0x20000001;
522 	msg->cnoc_cmds_data[1][2] =  0x60000001;
523 }
524 
a650_build_bw_table(struct a6xx_hfi_msg_bw_table * msg)525 static void a650_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
526 {
527 	/*
528 	 * Send a single "off" entry just to get things running
529 	 * TODO: bus scaling
530 	 */
531 	msg->bw_level_num = 1;
532 
533 	msg->ddr_cmds_num = 3;
534 	msg->ddr_wait_bitmask = 0x01;
535 
536 	msg->ddr_cmds_addrs[0] = 0x50000;
537 	msg->ddr_cmds_addrs[1] = 0x50004;
538 	msg->ddr_cmds_addrs[2] = 0x5007c;
539 
540 	msg->ddr_cmds_data[0][0] =  0x40000000;
541 	msg->ddr_cmds_data[0][1] =  0x40000000;
542 	msg->ddr_cmds_data[0][2] =  0x40000000;
543 
544 	/*
545 	 * These are the CX (CNOC) votes - these are used by the GMU but the
546 	 * votes are known and fixed for the target
547 	 */
548 	msg->cnoc_cmds_num = 1;
549 	msg->cnoc_wait_bitmask = 0x01;
550 
551 	msg->cnoc_cmds_addrs[0] = 0x500a4;
552 	msg->cnoc_cmds_data[0][0] =  0x40000000;
553 	msg->cnoc_cmds_data[1][0] =  0x60000001;
554 }
555 
a690_build_bw_table(struct a6xx_hfi_msg_bw_table * msg)556 static void a690_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
557 {
558 	/*
559 	 * Send a single "off" entry just to get things running
560 	 * TODO: bus scaling
561 	 */
562 	msg->bw_level_num = 1;
563 
564 	msg->ddr_cmds_num = 3;
565 	msg->ddr_wait_bitmask = 0x01;
566 
567 	msg->ddr_cmds_addrs[0] = 0x50004;
568 	msg->ddr_cmds_addrs[1] = 0x50000;
569 	msg->ddr_cmds_addrs[2] = 0x500ac;
570 
571 	msg->ddr_cmds_data[0][0] =  0x40000000;
572 	msg->ddr_cmds_data[0][1] =  0x40000000;
573 	msg->ddr_cmds_data[0][2] =  0x40000000;
574 
575 	/*
576 	 * These are the CX (CNOC) votes - these are used by the GMU but the
577 	 * votes are known and fixed for the target
578 	 */
579 	msg->cnoc_cmds_num = 1;
580 	msg->cnoc_wait_bitmask = 0x01;
581 
582 	msg->cnoc_cmds_addrs[0] = 0x5003c;
583 	msg->cnoc_cmds_data[0][0] =  0x40000000;
584 	msg->cnoc_cmds_data[1][0] =  0x60000001;
585 }
586 
a660_build_bw_table(struct a6xx_hfi_msg_bw_table * msg)587 static void a660_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
588 {
589 	/*
590 	 * Send a single "off" entry just to get things running
591 	 * TODO: bus scaling
592 	 */
593 	msg->bw_level_num = 1;
594 
595 	msg->ddr_cmds_num = 3;
596 	msg->ddr_wait_bitmask = 0x01;
597 
598 	msg->ddr_cmds_addrs[0] = 0x50004;
599 	msg->ddr_cmds_addrs[1] = 0x500a0;
600 	msg->ddr_cmds_addrs[2] = 0x50000;
601 
602 	msg->ddr_cmds_data[0][0] =  0x40000000;
603 	msg->ddr_cmds_data[0][1] =  0x40000000;
604 	msg->ddr_cmds_data[0][2] =  0x40000000;
605 
606 	/*
607 	 * These are the CX (CNOC) votes - these are used by the GMU but the
608 	 * votes are known and fixed for the target
609 	 */
610 	msg->cnoc_cmds_num = 1;
611 	msg->cnoc_wait_bitmask = 0x01;
612 
613 	msg->cnoc_cmds_addrs[0] = 0x50070;
614 	msg->cnoc_cmds_data[0][0] =  0x40000000;
615 	msg->cnoc_cmds_data[1][0] =  0x60000001;
616 }
617 
a663_build_bw_table(struct a6xx_hfi_msg_bw_table * msg)618 static void a663_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
619 {
620 	/*
621 	 * Send a single "off" entry just to get things running
622 	 * TODO: bus scaling
623 	 */
624 	msg->bw_level_num = 1;
625 
626 	msg->ddr_cmds_num = 3;
627 	msg->ddr_wait_bitmask = 0x07;
628 
629 	msg->ddr_cmds_addrs[0] = 0x50004;
630 	msg->ddr_cmds_addrs[1] = 0x50000;
631 	msg->ddr_cmds_addrs[2] = 0x500b4;
632 
633 	msg->ddr_cmds_data[0][0] =  0x40000000;
634 	msg->ddr_cmds_data[0][1] =  0x40000000;
635 	msg->ddr_cmds_data[0][2] =  0x40000000;
636 
637 	/*
638 	 * These are the CX (CNOC) votes - these are used by the GMU but the
639 	 * votes are known and fixed for the target
640 	 */
641 	msg->cnoc_cmds_num = 1;
642 	msg->cnoc_wait_bitmask = 0x01;
643 
644 	msg->cnoc_cmds_addrs[0] = 0x50058;
645 	msg->cnoc_cmds_data[0][0] =  0x40000000;
646 	msg->cnoc_cmds_data[1][0] =  0x60000001;
647 }
648 
adreno_7c3_build_bw_table(struct a6xx_hfi_msg_bw_table * msg)649 static void adreno_7c3_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
650 {
651 	/*
652 	 * Send a single "off" entry just to get things running
653 	 * TODO: bus scaling
654 	 */
655 	msg->bw_level_num = 1;
656 
657 	msg->ddr_cmds_num = 3;
658 	msg->ddr_wait_bitmask = 0x07;
659 
660 	msg->ddr_cmds_addrs[0] = 0x50004;
661 	msg->ddr_cmds_addrs[1] = 0x50000;
662 	msg->ddr_cmds_addrs[2] = 0x50088;
663 
664 	msg->ddr_cmds_data[0][0] =  0x40000000;
665 	msg->ddr_cmds_data[0][1] =  0x40000000;
666 	msg->ddr_cmds_data[0][2] =  0x40000000;
667 
668 	/*
669 	 * These are the CX (CNOC) votes - these are used by the GMU but the
670 	 * votes are known and fixed for the target
671 	 */
672 	msg->cnoc_cmds_num = 1;
673 	msg->cnoc_wait_bitmask = 0x01;
674 
675 	msg->cnoc_cmds_addrs[0] = 0x5006c;
676 	msg->cnoc_cmds_data[0][0] =  0x40000000;
677 	msg->cnoc_cmds_data[1][0] =  0x60000001;
678 }
679 
a730_build_bw_table(struct a6xx_hfi_msg_bw_table * msg)680 static void a730_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
681 {
682 	msg->bw_level_num = 12;
683 
684 	msg->ddr_cmds_num = 3;
685 	msg->ddr_wait_bitmask = 0x7;
686 
687 	msg->ddr_cmds_addrs[0] = cmd_db_read_addr("SH0");
688 	msg->ddr_cmds_addrs[1] = cmd_db_read_addr("MC0");
689 	msg->ddr_cmds_addrs[2] = cmd_db_read_addr("ACV");
690 
691 	msg->ddr_cmds_data[0][0] = 0x40000000;
692 	msg->ddr_cmds_data[0][1] = 0x40000000;
693 	msg->ddr_cmds_data[0][2] = 0x40000000;
694 	msg->ddr_cmds_data[1][0] = 0x600002e8;
695 	msg->ddr_cmds_data[1][1] = 0x600003d0;
696 	msg->ddr_cmds_data[1][2] = 0x60000008;
697 	msg->ddr_cmds_data[2][0] = 0x6000068d;
698 	msg->ddr_cmds_data[2][1] = 0x6000089a;
699 	msg->ddr_cmds_data[2][2] = 0x60000008;
700 	msg->ddr_cmds_data[3][0] = 0x600007f2;
701 	msg->ddr_cmds_data[3][1] = 0x60000a6e;
702 	msg->ddr_cmds_data[3][2] = 0x60000008;
703 	msg->ddr_cmds_data[4][0] = 0x600009e5;
704 	msg->ddr_cmds_data[4][1] = 0x60000cfd;
705 	msg->ddr_cmds_data[4][2] = 0x60000008;
706 	msg->ddr_cmds_data[5][0] = 0x60000b29;
707 	msg->ddr_cmds_data[5][1] = 0x60000ea6;
708 	msg->ddr_cmds_data[5][2] = 0x60000008;
709 	msg->ddr_cmds_data[6][0] = 0x60001698;
710 	msg->ddr_cmds_data[6][1] = 0x60001da8;
711 	msg->ddr_cmds_data[6][2] = 0x60000008;
712 	msg->ddr_cmds_data[7][0] = 0x600018d2;
713 	msg->ddr_cmds_data[7][1] = 0x60002093;
714 	msg->ddr_cmds_data[7][2] = 0x60000008;
715 	msg->ddr_cmds_data[8][0] = 0x60001e66;
716 	msg->ddr_cmds_data[8][1] = 0x600027e6;
717 	msg->ddr_cmds_data[8][2] = 0x60000008;
718 	msg->ddr_cmds_data[9][0] = 0x600027c2;
719 	msg->ddr_cmds_data[9][1] = 0x6000342f;
720 	msg->ddr_cmds_data[9][2] = 0x60000008;
721 	msg->ddr_cmds_data[10][0] = 0x60002e71;
722 	msg->ddr_cmds_data[10][1] = 0x60003cf5;
723 	msg->ddr_cmds_data[10][2] = 0x60000008;
724 	msg->ddr_cmds_data[11][0] = 0x600030ae;
725 	msg->ddr_cmds_data[11][1] = 0x60003fe5;
726 	msg->ddr_cmds_data[11][2] = 0x60000008;
727 
728 	msg->cnoc_cmds_num = 1;
729 	msg->cnoc_wait_bitmask = 0x1;
730 
731 	msg->cnoc_cmds_addrs[0] = cmd_db_read_addr("CN0");
732 	msg->cnoc_cmds_data[0][0] = 0x40000000;
733 	msg->cnoc_cmds_data[1][0] = 0x60000001;
734 }
735 
a740_build_bw_table(struct a6xx_hfi_msg_bw_table * msg)736 static void a740_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
737 {
738 	msg->bw_level_num = 1;
739 
740 	msg->ddr_cmds_num = 3;
741 	msg->ddr_wait_bitmask = 0x7;
742 
743 	msg->ddr_cmds_addrs[0] = cmd_db_read_addr("SH0");
744 	msg->ddr_cmds_addrs[1] = cmd_db_read_addr("MC0");
745 	msg->ddr_cmds_addrs[2] = cmd_db_read_addr("ACV");
746 
747 	msg->ddr_cmds_data[0][0] = 0x40000000;
748 	msg->ddr_cmds_data[0][1] = 0x40000000;
749 	msg->ddr_cmds_data[0][2] = 0x40000000;
750 
751 	/* TODO: add a proper dvfs table */
752 
753 	msg->cnoc_cmds_num = 1;
754 	msg->cnoc_wait_bitmask = 0x1;
755 
756 	msg->cnoc_cmds_addrs[0] = cmd_db_read_addr("CN0");
757 	msg->cnoc_cmds_data[0][0] = 0x40000000;
758 	msg->cnoc_cmds_data[1][0] = 0x60000001;
759 }
760 
a6xx_build_bw_table(struct a6xx_hfi_msg_bw_table * msg)761 static void a6xx_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
762 {
763 	/* Send a single "off" entry since the 630 GMU doesn't do bus scaling */
764 	msg->bw_level_num = 1;
765 
766 	msg->ddr_cmds_num = 3;
767 	msg->ddr_wait_bitmask = 0x07;
768 
769 	msg->ddr_cmds_addrs[0] = 0x50000;
770 	msg->ddr_cmds_addrs[1] = 0x5005c;
771 	msg->ddr_cmds_addrs[2] = 0x5000c;
772 
773 	msg->ddr_cmds_data[0][0] =  0x40000000;
774 	msg->ddr_cmds_data[0][1] =  0x40000000;
775 	msg->ddr_cmds_data[0][2] =  0x40000000;
776 
777 	/*
778 	 * These are the CX (CNOC) votes.  This is used but the values for the
779 	 * sdm845 GMU are known and fixed so we can hard code them.
780 	 */
781 
782 	msg->cnoc_cmds_num = 3;
783 	msg->cnoc_wait_bitmask = 0x05;
784 
785 	msg->cnoc_cmds_addrs[0] = 0x50034;
786 	msg->cnoc_cmds_addrs[1] = 0x5007c;
787 	msg->cnoc_cmds_addrs[2] = 0x5004c;
788 
789 	msg->cnoc_cmds_data[0][0] =  0x40000000;
790 	msg->cnoc_cmds_data[0][1] =  0x00000000;
791 	msg->cnoc_cmds_data[0][2] =  0x40000000;
792 
793 	msg->cnoc_cmds_data[1][0] =  0x60000001;
794 	msg->cnoc_cmds_data[1][1] =  0x20000001;
795 	msg->cnoc_cmds_data[1][2] =  0x60000001;
796 }
797 
798 
a6xx_hfi_send_bw_table(struct a6xx_gmu * gmu)799 static int a6xx_hfi_send_bw_table(struct a6xx_gmu *gmu)
800 {
801 	struct a6xx_hfi_msg_bw_table *msg;
802 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
803 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
804 	const struct a6xx_info *info = adreno_gpu->info->a6xx;
805 
806 	if (gmu->bw_table)
807 		goto send;
808 
809 	msg = devm_kzalloc(gmu->dev, sizeof(*msg), GFP_KERNEL);
810 	if (!msg)
811 		return -ENOMEM;
812 
813 	if (info->bcms && gmu->nr_gpu_bws > 1)
814 		a6xx_generate_bw_table(info, gmu, msg);
815 	else if (adreno_is_a618(adreno_gpu))
816 		a618_build_bw_table(msg);
817 	else if (adreno_is_a619(adreno_gpu))
818 		a619_build_bw_table(msg);
819 	else if (adreno_is_a640_family(adreno_gpu))
820 		a640_build_bw_table(msg);
821 	else if (adreno_is_a650(adreno_gpu))
822 		a650_build_bw_table(msg);
823 	else if (adreno_is_7c3(adreno_gpu))
824 		adreno_7c3_build_bw_table(msg);
825 	else if (adreno_is_a660(adreno_gpu))
826 		a660_build_bw_table(msg);
827 	else if (adreno_is_a663(adreno_gpu))
828 		a663_build_bw_table(msg);
829 	else if (adreno_is_a690(adreno_gpu))
830 		a690_build_bw_table(msg);
831 	else if (adreno_is_a730(adreno_gpu))
832 		a730_build_bw_table(msg);
833 	else if (adreno_is_a740_family(adreno_gpu))
834 		a740_build_bw_table(msg);
835 	else
836 		a6xx_build_bw_table(msg);
837 
838 	gmu->bw_table = msg;
839 
840 send:
841 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_BW_TABLE, gmu->bw_table, sizeof(*(gmu->bw_table)),
842 		NULL, 0);
843 }
844 
a6xx_hfi_feature_ctrl_msg(struct a6xx_gmu * gmu,u32 feature,u32 enable,u32 data)845 static int a6xx_hfi_feature_ctrl_msg(struct a6xx_gmu *gmu, u32 feature, u32 enable, u32 data)
846 {
847 	struct a6xx_hfi_msg_feature_ctrl msg = {
848 		.feature = feature,
849 		.enable = enable,
850 		.data = data,
851 	};
852 
853 	return a6xx_hfi_send_msg(gmu, HFI_H2F_FEATURE_CTRL, &msg, sizeof(msg), NULL, 0);
854 }
855 
856 #define IFPC_LONG_HYST 0x1680
857 
a6xx_hfi_enable_ifpc(struct a6xx_gmu * gmu)858 static int a6xx_hfi_enable_ifpc(struct a6xx_gmu *gmu)
859 {
860 	if (gmu->idle_level != GMU_IDLE_STATE_IFPC)
861 		return 0;
862 
863 	return a6xx_hfi_feature_ctrl_msg(gmu, HFI_FEATURE_IFPC, 1, IFPC_LONG_HYST);
864 }
865 
a6xx_hfi_enable_acd(struct a6xx_gmu * gmu)866 static int a6xx_hfi_enable_acd(struct a6xx_gmu *gmu)
867 {
868 	struct a6xx_hfi_acd_table *acd_table = &gmu->acd_table;
869 	int ret;
870 
871 	if (!acd_table->enable_by_level)
872 		return 0;
873 
874 	/* Enable ACD feature at GMU */
875 	ret = a6xx_hfi_feature_ctrl_msg(gmu, HFI_FEATURE_ACD, 1, 0);
876 	if (ret) {
877 		DRM_DEV_ERROR(gmu->dev, "Unable to enable ACD (%d)\n", ret);
878 		return ret;
879 	}
880 
881 	/* Send ACD table to GMU */
882 	ret = a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_ACD, acd_table, sizeof(*acd_table), NULL, 0);
883 	if (ret) {
884 		DRM_DEV_ERROR(gmu->dev, "Unable to ACD table (%d)\n", ret);
885 		return ret;
886 	}
887 
888 	return 0;
889 }
890 
a6xx_hfi_send_test(struct a6xx_gmu * gmu)891 static int a6xx_hfi_send_test(struct a6xx_gmu *gmu)
892 {
893 	struct a6xx_hfi_msg_test msg = { 0 };
894 
895 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_TEST, &msg, sizeof(msg),
896 		NULL, 0);
897 }
898 
a6xx_hfi_send_start(struct a6xx_gmu * gmu)899 static int a6xx_hfi_send_start(struct a6xx_gmu *gmu)
900 {
901 	struct a6xx_hfi_msg_start msg = { 0 };
902 
903 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_START, &msg, sizeof(msg),
904 		NULL, 0);
905 }
906 
a6xx_hfi_send_core_fw_start(struct a6xx_gmu * gmu)907 static int a6xx_hfi_send_core_fw_start(struct a6xx_gmu *gmu)
908 {
909 	struct a6xx_hfi_msg_core_fw_start msg = { 0 };
910 
911 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_CORE_FW_START, &msg,
912 		sizeof(msg), NULL, 0);
913 }
914 
a6xx_hfi_set_freq(struct a6xx_gmu * gmu,u32 freq_index,u32 bw_index)915 int a6xx_hfi_set_freq(struct a6xx_gmu *gmu, u32 freq_index, u32 bw_index)
916 {
917 	struct a6xx_hfi_gx_bw_perf_vote_cmd msg = { 0 };
918 
919 	msg.ack_type = 1; /* blocking */
920 	msg.freq = freq_index;
921 	msg.bw = bw_index;
922 
923 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_GX_BW_PERF_VOTE, &msg,
924 		sizeof(msg), NULL, 0);
925 }
926 
a6xx_hfi_send_prep_slumber(struct a6xx_gmu * gmu)927 int a6xx_hfi_send_prep_slumber(struct a6xx_gmu *gmu)
928 {
929 	struct a6xx_hfi_prep_slumber_cmd msg = { 0 };
930 
931 	/* TODO: should freq and bw fields be non-zero ? */
932 
933 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_PREPARE_SLUMBER, &msg,
934 		sizeof(msg), NULL, 0);
935 }
936 
a6xx_hfi_start_v1(struct a6xx_gmu * gmu,int boot_state)937 static int a6xx_hfi_start_v1(struct a6xx_gmu *gmu, int boot_state)
938 {
939 	int ret;
940 
941 	ret = a6xx_hfi_send_gmu_init(gmu, boot_state);
942 	if (ret)
943 		return ret;
944 
945 	ret = a6xx_hfi_get_fw_version(gmu, NULL);
946 	if (ret)
947 		return ret;
948 
949 	/*
950 	 * We have to get exchange version numbers per the sequence but at this
951 	 * point th kernel driver doesn't need to know the exact version of
952 	 * the GMU firmware
953 	 */
954 
955 	ret = a6xx_hfi_send_perf_table_v1(gmu);
956 	if (ret)
957 		return ret;
958 
959 	ret = a6xx_hfi_send_bw_table(gmu);
960 	if (ret)
961 		return ret;
962 
963 	/*
964 	 * Let the GMU know that there won't be any more HFI messages until next
965 	 * boot
966 	 */
967 	a6xx_hfi_send_test(gmu);
968 
969 	return 0;
970 }
971 
a6xx_hfi_start(struct a6xx_gmu * gmu,int boot_state)972 int a6xx_hfi_start(struct a6xx_gmu *gmu, int boot_state)
973 {
974 	int ret;
975 
976 	if (gmu->legacy)
977 		return a6xx_hfi_start_v1(gmu, boot_state);
978 
979 
980 	ret = a6xx_hfi_send_perf_table(gmu);
981 	if (ret)
982 		return ret;
983 
984 	ret = a6xx_hfi_send_bw_table(gmu);
985 	if (ret)
986 		return ret;
987 
988 	ret = a6xx_hfi_enable_acd(gmu);
989 	if (ret)
990 		return ret;
991 
992 	ret = a6xx_hfi_enable_ifpc(gmu);
993 	if (ret)
994 		return ret;
995 
996 	ret = a6xx_hfi_send_core_fw_start(gmu);
997 	if (ret)
998 		return ret;
999 
1000 	/*
1001 	 * Downstream driver sends this in its "a6xx_hw_init" equivalent,
1002 	 * but seems to be no harm in sending it here
1003 	 */
1004 	ret = a6xx_hfi_send_start(gmu);
1005 	if (ret)
1006 		return ret;
1007 
1008 	return 0;
1009 }
1010 
a6xx_hfi_stop(struct a6xx_gmu * gmu)1011 void a6xx_hfi_stop(struct a6xx_gmu *gmu)
1012 {
1013 	int i;
1014 
1015 	for (i = 0; i < ARRAY_SIZE(gmu->queues); i++) {
1016 		struct a6xx_hfi_queue *queue = &gmu->queues[i];
1017 
1018 		if (!queue->header)
1019 			continue;
1020 
1021 		if (queue->header->read_index != queue->header->write_index)
1022 			DRM_DEV_ERROR(gmu->dev, "HFI queue %d is not empty\n", i);
1023 
1024 		queue->header->read_index = 0;
1025 		queue->header->write_index = 0;
1026 
1027 		memset(&queue->history, 0xff, sizeof(queue->history));
1028 		queue->history_idx = 0;
1029 	}
1030 }
1031 
a6xx_hfi_queue_init(struct a6xx_hfi_queue * queue,struct a6xx_hfi_queue_header * header,void * virt,u64 iova,u32 id)1032 static void a6xx_hfi_queue_init(struct a6xx_hfi_queue *queue,
1033 		struct a6xx_hfi_queue_header *header, void *virt, u64 iova,
1034 		u32 id)
1035 {
1036 	spin_lock_init(&queue->lock);
1037 	queue->header = header;
1038 	queue->data = virt;
1039 	atomic_set(&queue->seqnum, 0);
1040 
1041 	memset(&queue->history, 0xff, sizeof(queue->history));
1042 	queue->history_idx = 0;
1043 
1044 	/* Set up the shared memory header */
1045 	header->iova = iova;
1046 	header->type =  10 << 8 | id;
1047 	header->status = 1;
1048 	header->size = SZ_4K >> 2;
1049 	header->msg_size = 0;
1050 	header->dropped = 0;
1051 	header->rx_watermark = 1;
1052 	header->tx_watermark = 1;
1053 	header->rx_request = 1;
1054 	header->tx_request = 0;
1055 	header->read_index = 0;
1056 	header->write_index = 0;
1057 }
1058 
a6xx_hfi_init(struct a6xx_gmu * gmu)1059 void a6xx_hfi_init(struct a6xx_gmu *gmu)
1060 {
1061 	struct a6xx_gmu_bo *hfi = &gmu->hfi;
1062 	struct a6xx_hfi_queue_table_header *table = hfi->virt;
1063 	struct a6xx_hfi_queue_header *headers = hfi->virt + sizeof(*table);
1064 	int table_size, idx;
1065 	u64 offset;
1066 
1067 	/*
1068 	 * The table size is the size of the table header plus all of the queue
1069 	 * headers
1070 	 */
1071 	table_size = sizeof(*table);
1072 	table_size += (ARRAY_SIZE(gmu->queues) *
1073 		sizeof(struct a6xx_hfi_queue_header));
1074 
1075 	table->version = 0;
1076 	table->size = table_size;
1077 	/* First queue header is located immediately after the table header */
1078 	table->qhdr0_offset = sizeof(*table) >> 2;
1079 	table->qhdr_size = sizeof(struct a6xx_hfi_queue_header) >> 2;
1080 	table->num_queues = ARRAY_SIZE(gmu->queues);
1081 	table->active_queues = ARRAY_SIZE(gmu->queues);
1082 
1083 	/* Command queue */
1084 	idx = 0;
1085 	offset = SZ_4K;
1086 	a6xx_hfi_queue_init(&gmu->queues[idx], &headers[idx], hfi->virt + offset,
1087 		hfi->iova + offset, 0);
1088 
1089 	/* GMU response queue */
1090 	idx++;
1091 	offset += SZ_4K;
1092 	a6xx_hfi_queue_init(&gmu->queues[idx], &headers[idx], hfi->virt + offset,
1093 		hfi->iova + offset, gmu->legacy ? 4 : 1);
1094 
1095 	/* GMU Debug queue */
1096 	idx++;
1097 	offset += SZ_4K;
1098 	a6xx_hfi_queue_init(&gmu->queues[idx], &headers[idx], hfi->virt + offset,
1099 		hfi->iova + offset, gmu->legacy ? 5 : 2);
1100 
1101 	WARN_ON(idx >= HFI_MAX_QUEUES);
1102 }
1103