xref: /openbmc/linux/drivers/gpu/drm/msm/adreno/a6xx_hfi.c (revision 4b565ca5)
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 
11 #define HFI_MSG_ID(val) [val] = #val
12 
13 static const char * const a6xx_hfi_msg_id[] = {
14 	HFI_MSG_ID(HFI_H2F_MSG_INIT),
15 	HFI_MSG_ID(HFI_H2F_MSG_FW_VERSION),
16 	HFI_MSG_ID(HFI_H2F_MSG_BW_TABLE),
17 	HFI_MSG_ID(HFI_H2F_MSG_PERF_TABLE),
18 	HFI_MSG_ID(HFI_H2F_MSG_TEST),
19 };
20 
21 static int a6xx_hfi_queue_read(struct a6xx_hfi_queue *queue, u32 *data,
22 		u32 dwords)
23 {
24 	struct a6xx_hfi_queue_header *header = queue->header;
25 	u32 i, hdr, index = header->read_index;
26 
27 	if (header->read_index == header->write_index) {
28 		header->rx_request = 1;
29 		return 0;
30 	}
31 
32 	hdr = queue->data[index];
33 
34 	/*
35 	 * If we are to assume that the GMU firmware is in fact a rational actor
36 	 * and is programmed to not send us a larger response than we expect
37 	 * then we can also assume that if the header size is unexpectedly large
38 	 * that it is due to memory corruption and/or hardware failure. In this
39 	 * case the only reasonable course of action is to BUG() to help harden
40 	 * the failure.
41 	 */
42 
43 	BUG_ON(HFI_HEADER_SIZE(hdr) > dwords);
44 
45 	for (i = 0; i < HFI_HEADER_SIZE(hdr); i++) {
46 		data[i] = queue->data[index];
47 		index = (index + 1) % header->size;
48 	}
49 
50 	header->read_index = index;
51 	return HFI_HEADER_SIZE(hdr);
52 }
53 
54 static int a6xx_hfi_queue_write(struct a6xx_gmu *gmu,
55 	struct a6xx_hfi_queue *queue, u32 *data, u32 dwords)
56 {
57 	struct a6xx_hfi_queue_header *header = queue->header;
58 	u32 i, space, index = header->write_index;
59 
60 	spin_lock(&queue->lock);
61 
62 	space = CIRC_SPACE(header->write_index, header->read_index,
63 		header->size);
64 	if (space < dwords) {
65 		header->dropped++;
66 		spin_unlock(&queue->lock);
67 		return -ENOSPC;
68 	}
69 
70 	for (i = 0; i < dwords; i++) {
71 		queue->data[index] = data[i];
72 		index = (index + 1) % header->size;
73 	}
74 
75 	header->write_index = index;
76 	spin_unlock(&queue->lock);
77 
78 	gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET, 0x01);
79 	return 0;
80 }
81 
82 struct a6xx_hfi_response {
83 	u32 id;
84 	u32 seqnum;
85 	struct list_head node;
86 	struct completion complete;
87 
88 	u32 error;
89 	u32 payload[16];
90 };
91 
92 /*
93  * Incoming HFI ack messages can come in out of order so we need to store all
94  * the pending messages on a list until they are handled.
95  */
96 static spinlock_t hfi_ack_lock = __SPIN_LOCK_UNLOCKED(message_lock);
97 static LIST_HEAD(hfi_ack_list);
98 
99 static void a6xx_hfi_handle_ack(struct a6xx_gmu *gmu,
100 		struct a6xx_hfi_msg_response *msg)
101 {
102 	struct a6xx_hfi_response *resp;
103 	u32 id, seqnum;
104 
105 	/* msg->ret_header contains the header of the message being acked */
106 	id = HFI_HEADER_ID(msg->ret_header);
107 	seqnum = HFI_HEADER_SEQNUM(msg->ret_header);
108 
109 	spin_lock(&hfi_ack_lock);
110 	list_for_each_entry(resp, &hfi_ack_list, node) {
111 		if (resp->id == id && resp->seqnum == seqnum) {
112 			resp->error = msg->error;
113 			memcpy(resp->payload, msg->payload,
114 				sizeof(resp->payload));
115 
116 			complete(&resp->complete);
117 			spin_unlock(&hfi_ack_lock);
118 			return;
119 		}
120 	}
121 	spin_unlock(&hfi_ack_lock);
122 
123 	dev_err(gmu->dev, "Nobody was waiting for HFI message %d\n", seqnum);
124 }
125 
126 static void a6xx_hfi_handle_error(struct a6xx_gmu *gmu,
127 		struct a6xx_hfi_msg_response *msg)
128 {
129 	struct a6xx_hfi_msg_error *error = (struct a6xx_hfi_msg_error *) msg;
130 
131 	dev_err(gmu->dev, "GMU firmware error %d\n", error->code);
132 }
133 
134 void a6xx_hfi_task(unsigned long data)
135 {
136 	struct a6xx_gmu *gmu = (struct a6xx_gmu *) data;
137 	struct a6xx_hfi_queue *queue = &gmu->queues[HFI_RESPONSE_QUEUE];
138 	struct a6xx_hfi_msg_response resp;
139 
140 	for (;;) {
141 		u32 id;
142 		int ret = a6xx_hfi_queue_read(queue, (u32 *) &resp,
143 			sizeof(resp) >> 2);
144 
145 		/* Returns the number of bytes copied or negative on error */
146 		if (ret <= 0) {
147 			if (ret < 0)
148 				dev_err(gmu->dev,
149 					"Unable to read the HFI message queue\n");
150 			break;
151 		}
152 
153 		id = HFI_HEADER_ID(resp.header);
154 
155 		if (id == HFI_F2H_MSG_ACK)
156 			a6xx_hfi_handle_ack(gmu, &resp);
157 		else if (id == HFI_F2H_MSG_ERROR)
158 			a6xx_hfi_handle_error(gmu, &resp);
159 	}
160 }
161 
162 static int a6xx_hfi_send_msg(struct a6xx_gmu *gmu, int id,
163 		void *data, u32 size, u32 *payload, u32 payload_size)
164 {
165 	struct a6xx_hfi_queue *queue = &gmu->queues[HFI_COMMAND_QUEUE];
166 	struct a6xx_hfi_response resp = { 0 };
167 	int ret, dwords = size >> 2;
168 	u32 seqnum;
169 
170 	seqnum = atomic_inc_return(&queue->seqnum) % 0xfff;
171 
172 	/* First dword of the message is the message header - fill it in */
173 	*((u32 *) data) = (seqnum << 20) | (HFI_MSG_CMD << 16) |
174 		(dwords << 8) | id;
175 
176 	init_completion(&resp.complete);
177 	resp.id = id;
178 	resp.seqnum = seqnum;
179 
180 	spin_lock_bh(&hfi_ack_lock);
181 	list_add_tail(&resp.node, &hfi_ack_list);
182 	spin_unlock_bh(&hfi_ack_lock);
183 
184 	ret = a6xx_hfi_queue_write(gmu, queue, data, dwords);
185 	if (ret) {
186 		dev_err(gmu->dev, "Unable to send message %s id %d\n",
187 			a6xx_hfi_msg_id[id], seqnum);
188 		goto out;
189 	}
190 
191 	/* Wait up to 5 seconds for the response */
192 	ret = wait_for_completion_timeout(&resp.complete,
193 		msecs_to_jiffies(5000));
194 	if (!ret) {
195 		dev_err(gmu->dev,
196 			"Message %s id %d timed out waiting for response\n",
197 			a6xx_hfi_msg_id[id], seqnum);
198 		ret = -ETIMEDOUT;
199 	} else
200 		ret = 0;
201 
202 out:
203 	spin_lock_bh(&hfi_ack_lock);
204 	list_del(&resp.node);
205 	spin_unlock_bh(&hfi_ack_lock);
206 
207 	if (ret)
208 		return ret;
209 
210 	if (resp.error) {
211 		dev_err(gmu->dev, "Message %s id %d returned error %d\n",
212 			a6xx_hfi_msg_id[id], seqnum, resp.error);
213 		return -EINVAL;
214 	}
215 
216 	if (payload && payload_size) {
217 		int copy = min_t(u32, payload_size, sizeof(resp.payload));
218 
219 		memcpy(payload, resp.payload, copy);
220 	}
221 
222 	return 0;
223 }
224 
225 static int a6xx_hfi_send_gmu_init(struct a6xx_gmu *gmu, int boot_state)
226 {
227 	struct a6xx_hfi_msg_gmu_init_cmd msg = { 0 };
228 
229 	msg.dbg_buffer_addr = (u32) gmu->debug->iova;
230 	msg.dbg_buffer_size = (u32) gmu->debug->size;
231 	msg.boot_state = boot_state;
232 
233 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_INIT, &msg, sizeof(msg),
234 		NULL, 0);
235 }
236 
237 static int a6xx_hfi_get_fw_version(struct a6xx_gmu *gmu, u32 *version)
238 {
239 	struct a6xx_hfi_msg_fw_version msg = { 0 };
240 
241 	/* Currently supporting version 1.1 */
242 	msg.supported_version = (1 << 28) | (1 << 16);
243 
244 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_FW_VERSION, &msg, sizeof(msg),
245 		version, sizeof(*version));
246 }
247 
248 static int a6xx_hfi_send_perf_table(struct a6xx_gmu *gmu)
249 {
250 	struct a6xx_hfi_msg_perf_table msg = { 0 };
251 	int i;
252 
253 	msg.num_gpu_levels = gmu->nr_gpu_freqs;
254 	msg.num_gmu_levels = gmu->nr_gmu_freqs;
255 
256 	for (i = 0; i < gmu->nr_gpu_freqs; i++) {
257 		msg.gx_votes[i].vote = gmu->gx_arc_votes[i];
258 		msg.gx_votes[i].freq = gmu->gpu_freqs[i] / 1000;
259 	}
260 
261 	for (i = 0; i < gmu->nr_gmu_freqs; i++) {
262 		msg.cx_votes[i].vote = gmu->cx_arc_votes[i];
263 		msg.cx_votes[i].freq = gmu->gmu_freqs[i] / 1000;
264 	}
265 
266 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_PERF_TABLE, &msg, sizeof(msg),
267 		NULL, 0);
268 }
269 
270 static int a6xx_hfi_send_bw_table(struct a6xx_gmu *gmu)
271 {
272 	struct a6xx_hfi_msg_bw_table msg = { 0 };
273 
274 	/*
275 	 * The sdm845 GMU doesn't do bus frequency scaling on its own but it
276 	 * does need at least one entry in the list because it might be accessed
277 	 * when the GMU is shutting down. Send a single "off" entry.
278 	 */
279 
280 	msg.bw_level_num = 1;
281 
282 	msg.ddr_cmds_num = 3;
283 	msg.ddr_wait_bitmask = 0x07;
284 
285 	msg.ddr_cmds_addrs[0] = 0x50000;
286 	msg.ddr_cmds_addrs[1] = 0x5005c;
287 	msg.ddr_cmds_addrs[2] = 0x5000c;
288 
289 	msg.ddr_cmds_data[0][0] =  0x40000000;
290 	msg.ddr_cmds_data[0][1] =  0x40000000;
291 	msg.ddr_cmds_data[0][2] =  0x40000000;
292 
293 	/*
294 	 * These are the CX (CNOC) votes.  This is used but the values for the
295 	 * sdm845 GMU are known and fixed so we can hard code them.
296 	 */
297 
298 	msg.cnoc_cmds_num = 3;
299 	msg.cnoc_wait_bitmask = 0x05;
300 
301 	msg.cnoc_cmds_addrs[0] = 0x50034;
302 	msg.cnoc_cmds_addrs[1] = 0x5007c;
303 	msg.cnoc_cmds_addrs[2] = 0x5004c;
304 
305 	msg.cnoc_cmds_data[0][0] =  0x40000000;
306 	msg.cnoc_cmds_data[0][1] =  0x00000000;
307 	msg.cnoc_cmds_data[0][2] =  0x40000000;
308 
309 	msg.cnoc_cmds_data[1][0] =  0x60000001;
310 	msg.cnoc_cmds_data[1][1] =  0x20000001;
311 	msg.cnoc_cmds_data[1][2] =  0x60000001;
312 
313 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_BW_TABLE, &msg, sizeof(msg),
314 		NULL, 0);
315 }
316 
317 static int a6xx_hfi_send_test(struct a6xx_gmu *gmu)
318 {
319 	struct a6xx_hfi_msg_test msg = { 0 };
320 
321 	return a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_TEST, &msg, sizeof(msg),
322 		NULL, 0);
323 }
324 
325 int a6xx_hfi_start(struct a6xx_gmu *gmu, int boot_state)
326 {
327 	int ret;
328 
329 	ret = a6xx_hfi_send_gmu_init(gmu, boot_state);
330 	if (ret)
331 		return ret;
332 
333 	ret = a6xx_hfi_get_fw_version(gmu, NULL);
334 	if (ret)
335 		return ret;
336 
337 	/*
338 	 * We have to get exchange version numbers per the sequence but at this
339 	 * point th kernel driver doesn't need to know the exact version of
340 	 * the GMU firmware
341 	 */
342 
343 	ret = a6xx_hfi_send_perf_table(gmu);
344 	if (ret)
345 		return ret;
346 
347 	ret = a6xx_hfi_send_bw_table(gmu);
348 	if (ret)
349 		return ret;
350 
351 	/*
352 	 * Let the GMU know that there won't be any more HFI messages until next
353 	 * boot
354 	 */
355 	a6xx_hfi_send_test(gmu);
356 
357 	return 0;
358 }
359 
360 void a6xx_hfi_stop(struct a6xx_gmu *gmu)
361 {
362 	int i;
363 
364 	for (i = 0; i < ARRAY_SIZE(gmu->queues); i++) {
365 		struct a6xx_hfi_queue *queue = &gmu->queues[i];
366 
367 		if (!queue->header)
368 			continue;
369 
370 		if (queue->header->read_index != queue->header->write_index)
371 			dev_err(gmu->dev, "HFI queue %d is not empty\n", i);
372 
373 		queue->header->read_index = 0;
374 		queue->header->write_index = 0;
375 	}
376 }
377 
378 static void a6xx_hfi_queue_init(struct a6xx_hfi_queue *queue,
379 		struct a6xx_hfi_queue_header *header, void *virt, u64 iova,
380 		u32 id)
381 {
382 	spin_lock_init(&queue->lock);
383 	queue->header = header;
384 	queue->data = virt;
385 	atomic_set(&queue->seqnum, 0);
386 
387 	/* Set up the shared memory header */
388 	header->iova = iova;
389 	header->type =  10 << 8 | id;
390 	header->status = 1;
391 	header->size = SZ_4K >> 2;
392 	header->msg_size = 0;
393 	header->dropped = 0;
394 	header->rx_watermark = 1;
395 	header->tx_watermark = 1;
396 	header->rx_request = 1;
397 	header->tx_request = 0;
398 	header->read_index = 0;
399 	header->write_index = 0;
400 }
401 
402 void a6xx_hfi_init(struct a6xx_gmu *gmu)
403 {
404 	struct a6xx_gmu_bo *hfi = gmu->hfi;
405 	struct a6xx_hfi_queue_table_header *table = hfi->virt;
406 	struct a6xx_hfi_queue_header *headers = hfi->virt + sizeof(*table);
407 	u64 offset;
408 	int table_size;
409 
410 	/*
411 	 * The table size is the size of the table header plus all of the queue
412 	 * headers
413 	 */
414 	table_size = sizeof(*table);
415 	table_size += (ARRAY_SIZE(gmu->queues) *
416 		sizeof(struct a6xx_hfi_queue_header));
417 
418 	table->version = 0;
419 	table->size = table_size;
420 	/* First queue header is located immediately after the table header */
421 	table->qhdr0_offset = sizeof(*table) >> 2;
422 	table->qhdr_size = sizeof(struct a6xx_hfi_queue_header) >> 2;
423 	table->num_queues = ARRAY_SIZE(gmu->queues);
424 	table->active_queues = ARRAY_SIZE(gmu->queues);
425 
426 	/* Command queue */
427 	offset = SZ_4K;
428 	a6xx_hfi_queue_init(&gmu->queues[0], &headers[0], hfi->virt + offset,
429 		hfi->iova + offset, 0);
430 
431 	/* GMU response queue */
432 	offset += SZ_4K;
433 	a6xx_hfi_queue_init(&gmu->queues[1], &headers[1], hfi->virt + offset,
434 		hfi->iova + offset, 4);
435 }
436