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
a6xx_hfi_queue_read(struct a6xx_gmu * gmu,struct a6xx_hfi_queue * queue,u32 * data,u32 dwords)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
a6xx_hfi_queue_write(struct a6xx_gmu * gmu,struct a6xx_hfi_queue * queue,u32 * data,u32 dwords)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
a6xx_hfi_wait_for_ack(struct a6xx_gmu * gmu,u32 id,u32 seqnum,u32 * payload,u32 payload_size)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
a6xx_hfi_send_msg(struct a6xx_gmu * gmu,int id,void * data,u32 size,u32 * payload,u32 payload_size)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
a6xx_hfi_send_gmu_init(struct a6xx_gmu * gmu,int boot_state)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
a6xx_hfi_get_fw_version(struct a6xx_gmu * gmu,u32 * version)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 1.10 */
209 msg.supported_version = (1 << 28) | (1 << 19) | (1 << 17);
210
211 return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_FW_VERSION, &msg, sizeof(msg),
212 version, sizeof(*version));
213 }
214
a6xx_hfi_send_perf_table_v1(struct a6xx_gmu * gmu)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
a6xx_hfi_send_perf_table(struct a6xx_gmu * gmu)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
a618_build_bw_table(struct a6xx_hfi_msg_bw_table * msg)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
a619_build_bw_table(struct a6xx_hfi_msg_bw_table * msg)288 static void a619_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
289 {
290 msg->bw_level_num = 13;
291
292 msg->ddr_cmds_num = 3;
293 msg->ddr_wait_bitmask = 0x0;
294
295 msg->ddr_cmds_addrs[0] = 0x50000;
296 msg->ddr_cmds_addrs[1] = 0x50004;
297 msg->ddr_cmds_addrs[2] = 0x50080;
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 msg->ddr_cmds_data[1][0] = 0x6000030c;
303 msg->ddr_cmds_data[1][1] = 0x600000db;
304 msg->ddr_cmds_data[1][2] = 0x60000008;
305 msg->ddr_cmds_data[2][0] = 0x60000618;
306 msg->ddr_cmds_data[2][1] = 0x600001b6;
307 msg->ddr_cmds_data[2][2] = 0x60000008;
308 msg->ddr_cmds_data[3][0] = 0x60000925;
309 msg->ddr_cmds_data[3][1] = 0x60000291;
310 msg->ddr_cmds_data[3][2] = 0x60000008;
311 msg->ddr_cmds_data[4][0] = 0x60000dc1;
312 msg->ddr_cmds_data[4][1] = 0x600003dc;
313 msg->ddr_cmds_data[4][2] = 0x60000008;
314 msg->ddr_cmds_data[5][0] = 0x600010ad;
315 msg->ddr_cmds_data[5][1] = 0x600004ae;
316 msg->ddr_cmds_data[5][2] = 0x60000008;
317 msg->ddr_cmds_data[6][0] = 0x600014c3;
318 msg->ddr_cmds_data[6][1] = 0x600005d4;
319 msg->ddr_cmds_data[6][2] = 0x60000008;
320 msg->ddr_cmds_data[7][0] = 0x6000176a;
321 msg->ddr_cmds_data[7][1] = 0x60000693;
322 msg->ddr_cmds_data[7][2] = 0x60000008;
323 msg->ddr_cmds_data[8][0] = 0x60001f01;
324 msg->ddr_cmds_data[8][1] = 0x600008b5;
325 msg->ddr_cmds_data[8][2] = 0x60000008;
326 msg->ddr_cmds_data[9][0] = 0x60002940;
327 msg->ddr_cmds_data[9][1] = 0x60000b95;
328 msg->ddr_cmds_data[9][2] = 0x60000008;
329 msg->ddr_cmds_data[10][0] = 0x60002f68;
330 msg->ddr_cmds_data[10][1] = 0x60000d50;
331 msg->ddr_cmds_data[10][2] = 0x60000008;
332 msg->ddr_cmds_data[11][0] = 0x60003700;
333 msg->ddr_cmds_data[11][1] = 0x60000f71;
334 msg->ddr_cmds_data[11][2] = 0x60000008;
335 msg->ddr_cmds_data[12][0] = 0x60003fce;
336 msg->ddr_cmds_data[12][1] = 0x600011ea;
337 msg->ddr_cmds_data[12][2] = 0x60000008;
338
339 msg->cnoc_cmds_num = 1;
340 msg->cnoc_wait_bitmask = 0x0;
341
342 msg->cnoc_cmds_addrs[0] = 0x50054;
343
344 msg->cnoc_cmds_data[0][0] = 0x40000000;
345 }
346
a640_build_bw_table(struct a6xx_hfi_msg_bw_table * msg)347 static void a640_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
348 {
349 /*
350 * Send a single "off" entry just to get things running
351 * TODO: bus scaling
352 */
353 msg->bw_level_num = 1;
354
355 msg->ddr_cmds_num = 3;
356 msg->ddr_wait_bitmask = 0x01;
357
358 msg->ddr_cmds_addrs[0] = 0x50000;
359 msg->ddr_cmds_addrs[1] = 0x5003c;
360 msg->ddr_cmds_addrs[2] = 0x5000c;
361
362 msg->ddr_cmds_data[0][0] = 0x40000000;
363 msg->ddr_cmds_data[0][1] = 0x40000000;
364 msg->ddr_cmds_data[0][2] = 0x40000000;
365
366 /*
367 * These are the CX (CNOC) votes - these are used by the GMU but the
368 * votes are known and fixed for the target
369 */
370 msg->cnoc_cmds_num = 3;
371 msg->cnoc_wait_bitmask = 0x01;
372
373 msg->cnoc_cmds_addrs[0] = 0x50034;
374 msg->cnoc_cmds_addrs[1] = 0x5007c;
375 msg->cnoc_cmds_addrs[2] = 0x5004c;
376
377 msg->cnoc_cmds_data[0][0] = 0x40000000;
378 msg->cnoc_cmds_data[0][1] = 0x00000000;
379 msg->cnoc_cmds_data[0][2] = 0x40000000;
380
381 msg->cnoc_cmds_data[1][0] = 0x60000001;
382 msg->cnoc_cmds_data[1][1] = 0x20000001;
383 msg->cnoc_cmds_data[1][2] = 0x60000001;
384 }
385
a650_build_bw_table(struct a6xx_hfi_msg_bw_table * msg)386 static void a650_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
387 {
388 /*
389 * Send a single "off" entry just to get things running
390 * TODO: bus scaling
391 */
392 msg->bw_level_num = 1;
393
394 msg->ddr_cmds_num = 3;
395 msg->ddr_wait_bitmask = 0x01;
396
397 msg->ddr_cmds_addrs[0] = 0x50000;
398 msg->ddr_cmds_addrs[1] = 0x50004;
399 msg->ddr_cmds_addrs[2] = 0x5007c;
400
401 msg->ddr_cmds_data[0][0] = 0x40000000;
402 msg->ddr_cmds_data[0][1] = 0x40000000;
403 msg->ddr_cmds_data[0][2] = 0x40000000;
404
405 /*
406 * These are the CX (CNOC) votes - these are used by the GMU but the
407 * votes are known and fixed for the target
408 */
409 msg->cnoc_cmds_num = 1;
410 msg->cnoc_wait_bitmask = 0x01;
411
412 msg->cnoc_cmds_addrs[0] = 0x500a4;
413 msg->cnoc_cmds_data[0][0] = 0x40000000;
414 msg->cnoc_cmds_data[1][0] = 0x60000001;
415 }
416
a660_build_bw_table(struct a6xx_hfi_msg_bw_table * msg)417 static void a660_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
418 {
419 /*
420 * Send a single "off" entry just to get things running
421 * TODO: bus scaling
422 */
423 msg->bw_level_num = 1;
424
425 msg->ddr_cmds_num = 3;
426 msg->ddr_wait_bitmask = 0x01;
427
428 msg->ddr_cmds_addrs[0] = 0x50004;
429 msg->ddr_cmds_addrs[1] = 0x500a0;
430 msg->ddr_cmds_addrs[2] = 0x50000;
431
432 msg->ddr_cmds_data[0][0] = 0x40000000;
433 msg->ddr_cmds_data[0][1] = 0x40000000;
434 msg->ddr_cmds_data[0][2] = 0x40000000;
435
436 /*
437 * These are the CX (CNOC) votes - these are used by the GMU but the
438 * votes are known and fixed for the target
439 */
440 msg->cnoc_cmds_num = 1;
441 msg->cnoc_wait_bitmask = 0x01;
442
443 msg->cnoc_cmds_addrs[0] = 0x50070;
444 msg->cnoc_cmds_data[0][0] = 0x40000000;
445 msg->cnoc_cmds_data[1][0] = 0x60000001;
446 }
447
adreno_7c3_build_bw_table(struct a6xx_hfi_msg_bw_table * msg)448 static void adreno_7c3_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
449 {
450 /*
451 * Send a single "off" entry just to get things running
452 * TODO: bus scaling
453 */
454 msg->bw_level_num = 1;
455
456 msg->ddr_cmds_num = 3;
457 msg->ddr_wait_bitmask = 0x07;
458
459 msg->ddr_cmds_addrs[0] = 0x50004;
460 msg->ddr_cmds_addrs[1] = 0x50000;
461 msg->ddr_cmds_addrs[2] = 0x50088;
462
463 msg->ddr_cmds_data[0][0] = 0x40000000;
464 msg->ddr_cmds_data[0][1] = 0x40000000;
465 msg->ddr_cmds_data[0][2] = 0x40000000;
466
467 /*
468 * These are the CX (CNOC) votes - these are used by the GMU but the
469 * votes are known and fixed for the target
470 */
471 msg->cnoc_cmds_num = 1;
472 msg->cnoc_wait_bitmask = 0x01;
473
474 msg->cnoc_cmds_addrs[0] = 0x5006c;
475 msg->cnoc_cmds_data[0][0] = 0x40000000;
476 msg->cnoc_cmds_data[1][0] = 0x60000001;
477 }
a6xx_build_bw_table(struct a6xx_hfi_msg_bw_table * msg)478 static void a6xx_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
479 {
480 /* Send a single "off" entry since the 630 GMU doesn't do bus scaling */
481 msg->bw_level_num = 1;
482
483 msg->ddr_cmds_num = 3;
484 msg->ddr_wait_bitmask = 0x07;
485
486 msg->ddr_cmds_addrs[0] = 0x50000;
487 msg->ddr_cmds_addrs[1] = 0x5005c;
488 msg->ddr_cmds_addrs[2] = 0x5000c;
489
490 msg->ddr_cmds_data[0][0] = 0x40000000;
491 msg->ddr_cmds_data[0][1] = 0x40000000;
492 msg->ddr_cmds_data[0][2] = 0x40000000;
493
494 /*
495 * These are the CX (CNOC) votes. This is used but the values for the
496 * sdm845 GMU are known and fixed so we can hard code them.
497 */
498
499 msg->cnoc_cmds_num = 3;
500 msg->cnoc_wait_bitmask = 0x05;
501
502 msg->cnoc_cmds_addrs[0] = 0x50034;
503 msg->cnoc_cmds_addrs[1] = 0x5007c;
504 msg->cnoc_cmds_addrs[2] = 0x5004c;
505
506 msg->cnoc_cmds_data[0][0] = 0x40000000;
507 msg->cnoc_cmds_data[0][1] = 0x00000000;
508 msg->cnoc_cmds_data[0][2] = 0x40000000;
509
510 msg->cnoc_cmds_data[1][0] = 0x60000001;
511 msg->cnoc_cmds_data[1][1] = 0x20000001;
512 msg->cnoc_cmds_data[1][2] = 0x60000001;
513 }
514
515
a6xx_hfi_send_bw_table(struct a6xx_gmu * gmu)516 static int a6xx_hfi_send_bw_table(struct a6xx_gmu *gmu)
517 {
518 struct a6xx_hfi_msg_bw_table msg = { 0 };
519 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
520 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
521
522 if (adreno_is_a618(adreno_gpu))
523 a618_build_bw_table(&msg);
524 else if (adreno_is_a619(adreno_gpu))
525 a619_build_bw_table(&msg);
526 else if (adreno_is_a640_family(adreno_gpu))
527 a640_build_bw_table(&msg);
528 else if (adreno_is_a650(adreno_gpu))
529 a650_build_bw_table(&msg);
530 else if (adreno_is_7c3(adreno_gpu))
531 adreno_7c3_build_bw_table(&msg);
532 else if (adreno_is_a660(adreno_gpu))
533 a660_build_bw_table(&msg);
534 else
535 a6xx_build_bw_table(&msg);
536
537 return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_BW_TABLE, &msg, sizeof(msg),
538 NULL, 0);
539 }
540
a6xx_hfi_send_test(struct a6xx_gmu * gmu)541 static int a6xx_hfi_send_test(struct a6xx_gmu *gmu)
542 {
543 struct a6xx_hfi_msg_test msg = { 0 };
544
545 return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_TEST, &msg, sizeof(msg),
546 NULL, 0);
547 }
548
a6xx_hfi_send_start(struct a6xx_gmu * gmu)549 static int a6xx_hfi_send_start(struct a6xx_gmu *gmu)
550 {
551 struct a6xx_hfi_msg_start msg = { 0 };
552
553 return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_START, &msg, sizeof(msg),
554 NULL, 0);
555 }
556
a6xx_hfi_send_core_fw_start(struct a6xx_gmu * gmu)557 static int a6xx_hfi_send_core_fw_start(struct a6xx_gmu *gmu)
558 {
559 struct a6xx_hfi_msg_core_fw_start msg = { 0 };
560
561 return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_CORE_FW_START, &msg,
562 sizeof(msg), NULL, 0);
563 }
564
a6xx_hfi_set_freq(struct a6xx_gmu * gmu,int index)565 int a6xx_hfi_set_freq(struct a6xx_gmu *gmu, int index)
566 {
567 struct a6xx_hfi_gx_bw_perf_vote_cmd msg = { 0 };
568
569 msg.ack_type = 1; /* blocking */
570 msg.freq = index;
571 msg.bw = 0; /* TODO: bus scaling */
572
573 return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_GX_BW_PERF_VOTE, &msg,
574 sizeof(msg), NULL, 0);
575 }
576
a6xx_hfi_send_prep_slumber(struct a6xx_gmu * gmu)577 int a6xx_hfi_send_prep_slumber(struct a6xx_gmu *gmu)
578 {
579 struct a6xx_hfi_prep_slumber_cmd msg = { 0 };
580
581 /* TODO: should freq and bw fields be non-zero ? */
582
583 return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_PREPARE_SLUMBER, &msg,
584 sizeof(msg), NULL, 0);
585 }
586
a6xx_hfi_start_v1(struct a6xx_gmu * gmu,int boot_state)587 static int a6xx_hfi_start_v1(struct a6xx_gmu *gmu, int boot_state)
588 {
589 int ret;
590
591 ret = a6xx_hfi_send_gmu_init(gmu, boot_state);
592 if (ret)
593 return ret;
594
595 ret = a6xx_hfi_get_fw_version(gmu, NULL);
596 if (ret)
597 return ret;
598
599 /*
600 * We have to get exchange version numbers per the sequence but at this
601 * point th kernel driver doesn't need to know the exact version of
602 * the GMU firmware
603 */
604
605 ret = a6xx_hfi_send_perf_table_v1(gmu);
606 if (ret)
607 return ret;
608
609 ret = a6xx_hfi_send_bw_table(gmu);
610 if (ret)
611 return ret;
612
613 /*
614 * Let the GMU know that there won't be any more HFI messages until next
615 * boot
616 */
617 a6xx_hfi_send_test(gmu);
618
619 return 0;
620 }
621
a6xx_hfi_start(struct a6xx_gmu * gmu,int boot_state)622 int a6xx_hfi_start(struct a6xx_gmu *gmu, int boot_state)
623 {
624 int ret;
625
626 if (gmu->legacy)
627 return a6xx_hfi_start_v1(gmu, boot_state);
628
629
630 ret = a6xx_hfi_send_perf_table(gmu);
631 if (ret)
632 return ret;
633
634 ret = a6xx_hfi_send_bw_table(gmu);
635 if (ret)
636 return ret;
637
638 ret = a6xx_hfi_send_core_fw_start(gmu);
639 if (ret)
640 return ret;
641
642 /*
643 * Downstream driver sends this in its "a6xx_hw_init" equivalent,
644 * but seems to be no harm in sending it here
645 */
646 ret = a6xx_hfi_send_start(gmu);
647 if (ret)
648 return ret;
649
650 return 0;
651 }
652
a6xx_hfi_stop(struct a6xx_gmu * gmu)653 void a6xx_hfi_stop(struct a6xx_gmu *gmu)
654 {
655 int i;
656
657 for (i = 0; i < ARRAY_SIZE(gmu->queues); i++) {
658 struct a6xx_hfi_queue *queue = &gmu->queues[i];
659
660 if (!queue->header)
661 continue;
662
663 if (queue->header->read_index != queue->header->write_index)
664 DRM_DEV_ERROR(gmu->dev, "HFI queue %d is not empty\n", i);
665
666 queue->header->read_index = 0;
667 queue->header->write_index = 0;
668
669 memset(&queue->history, 0xff, sizeof(queue->history));
670 queue->history_idx = 0;
671 }
672 }
673
a6xx_hfi_queue_init(struct a6xx_hfi_queue * queue,struct a6xx_hfi_queue_header * header,void * virt,u64 iova,u32 id)674 static void a6xx_hfi_queue_init(struct a6xx_hfi_queue *queue,
675 struct a6xx_hfi_queue_header *header, void *virt, u64 iova,
676 u32 id)
677 {
678 spin_lock_init(&queue->lock);
679 queue->header = header;
680 queue->data = virt;
681 atomic_set(&queue->seqnum, 0);
682
683 memset(&queue->history, 0xff, sizeof(queue->history));
684 queue->history_idx = 0;
685
686 /* Set up the shared memory header */
687 header->iova = iova;
688 header->type = 10 << 8 | id;
689 header->status = 1;
690 header->size = SZ_4K >> 2;
691 header->msg_size = 0;
692 header->dropped = 0;
693 header->rx_watermark = 1;
694 header->tx_watermark = 1;
695 header->rx_request = 1;
696 header->tx_request = 0;
697 header->read_index = 0;
698 header->write_index = 0;
699 }
700
a6xx_hfi_init(struct a6xx_gmu * gmu)701 void a6xx_hfi_init(struct a6xx_gmu *gmu)
702 {
703 struct a6xx_gmu_bo *hfi = &gmu->hfi;
704 struct a6xx_hfi_queue_table_header *table = hfi->virt;
705 struct a6xx_hfi_queue_header *headers = hfi->virt + sizeof(*table);
706 u64 offset;
707 int table_size;
708
709 /*
710 * The table size is the size of the table header plus all of the queue
711 * headers
712 */
713 table_size = sizeof(*table);
714 table_size += (ARRAY_SIZE(gmu->queues) *
715 sizeof(struct a6xx_hfi_queue_header));
716
717 table->version = 0;
718 table->size = table_size;
719 /* First queue header is located immediately after the table header */
720 table->qhdr0_offset = sizeof(*table) >> 2;
721 table->qhdr_size = sizeof(struct a6xx_hfi_queue_header) >> 2;
722 table->num_queues = ARRAY_SIZE(gmu->queues);
723 table->active_queues = ARRAY_SIZE(gmu->queues);
724
725 /* Command queue */
726 offset = SZ_4K;
727 a6xx_hfi_queue_init(&gmu->queues[0], &headers[0], hfi->virt + offset,
728 hfi->iova + offset, 0);
729
730 /* GMU response queue */
731 offset += SZ_4K;
732 a6xx_hfi_queue_init(&gmu->queues[1], &headers[1], hfi->virt + offset,
733 hfi->iova + offset, gmu->legacy ? 4 : 1);
734 }
735