xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c (revision e1cd7b80)
1 /*
2  * Copyright 2014 Advanced Micro Devices, Inc.
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20  * OTHER DEALINGS IN THE SOFTWARE.
21  *
22  */
23 
24 #include "amdgpu.h"
25 #include "nbio/nbio_6_1_offset.h"
26 #include "nbio/nbio_6_1_sh_mask.h"
27 #include "gc/gc_9_0_offset.h"
28 #include "gc/gc_9_0_sh_mask.h"
29 #include "mp/mp_9_0_offset.h"
30 #include "soc15.h"
31 #include "vega10_ih.h"
32 #include "soc15_common.h"
33 #include "mxgpu_ai.h"
34 
35 static void xgpu_ai_mailbox_send_ack(struct amdgpu_device *adev)
36 {
37 	WREG8(AI_MAIBOX_CONTROL_RCV_OFFSET_BYTE, 2);
38 }
39 
40 static void xgpu_ai_mailbox_set_valid(struct amdgpu_device *adev, bool val)
41 {
42 	WREG8(AI_MAIBOX_CONTROL_TRN_OFFSET_BYTE, val ? 1 : 0);
43 }
44 
45 /*
46  * this peek_msg could *only* be called in IRQ routine becuase in IRQ routine
47  * RCV_MSG_VALID filed of BIF_BX_PF0_MAILBOX_CONTROL must already be set to 1
48  * by host.
49  *
50  * if called no in IRQ routine, this peek_msg cannot guaranteed to return the
51  * correct value since it doesn't return the RCV_DW0 under the case that
52  * RCV_MSG_VALID is set by host.
53  */
54 static enum idh_event xgpu_ai_mailbox_peek_msg(struct amdgpu_device *adev)
55 {
56 	return RREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0,
57 				mmBIF_BX_PF0_MAILBOX_MSGBUF_RCV_DW0));
58 }
59 
60 
61 static int xgpu_ai_mailbox_rcv_msg(struct amdgpu_device *adev,
62 				   enum idh_event event)
63 {
64 	u32 reg;
65 
66 	reg = RREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0,
67 					     mmBIF_BX_PF0_MAILBOX_MSGBUF_RCV_DW0));
68 	if (reg != event)
69 		return -ENOENT;
70 
71 	xgpu_ai_mailbox_send_ack(adev);
72 
73 	return 0;
74 }
75 
76 static uint8_t xgpu_ai_peek_ack(struct amdgpu_device *adev) {
77 	return RREG8(AI_MAIBOX_CONTROL_TRN_OFFSET_BYTE) & 2;
78 }
79 
80 static int xgpu_ai_poll_ack(struct amdgpu_device *adev)
81 {
82 	int timeout  = AI_MAILBOX_POLL_ACK_TIMEDOUT;
83 	u8 reg;
84 
85 	do {
86 		reg = RREG8(AI_MAIBOX_CONTROL_TRN_OFFSET_BYTE);
87 		if (reg & 2)
88 			return 0;
89 
90 		mdelay(5);
91 		timeout -= 5;
92 	} while (timeout > 1);
93 
94 	pr_err("Doesn't get TRN_MSG_ACK from pf in %d msec\n", AI_MAILBOX_POLL_ACK_TIMEDOUT);
95 
96 	return -ETIME;
97 }
98 
99 static int xgpu_ai_poll_msg(struct amdgpu_device *adev, enum idh_event event)
100 {
101 	int r, timeout = AI_MAILBOX_POLL_MSG_TIMEDOUT;
102 
103 	do {
104 		r = xgpu_ai_mailbox_rcv_msg(adev, event);
105 		if (!r)
106 			return 0;
107 
108 		msleep(10);
109 		timeout -= 10;
110 	} while (timeout > 1);
111 
112 	pr_err("Doesn't get msg:%d from pf, error=%d\n", event, r);
113 
114 	return -ETIME;
115 }
116 
117 static void xgpu_ai_mailbox_trans_msg (struct amdgpu_device *adev,
118 	      enum idh_request req, u32 data1, u32 data2, u32 data3) {
119 	u32 reg;
120 	int r;
121 	uint8_t trn;
122 
123 	/* IMPORTANT:
124 	 * clear TRN_MSG_VALID valid to clear host's RCV_MSG_ACK
125 	 * and with host's RCV_MSG_ACK cleared hw automatically clear host's RCV_MSG_ACK
126 	 * which lead to VF's TRN_MSG_ACK cleared, otherwise below xgpu_ai_poll_ack()
127 	 * will return immediatly
128 	 */
129 	do {
130 		xgpu_ai_mailbox_set_valid(adev, false);
131 		trn = xgpu_ai_peek_ack(adev);
132 		if (trn) {
133 			pr_err("trn=%x ACK should not assert! wait again !\n", trn);
134 			msleep(1);
135 		}
136 	} while(trn);
137 
138 	reg = RREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0,
139 					     mmBIF_BX_PF0_MAILBOX_MSGBUF_TRN_DW0));
140 	reg = REG_SET_FIELD(reg, BIF_BX_PF0_MAILBOX_MSGBUF_TRN_DW0,
141 			    MSGBUF_DATA, req);
142 	WREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF0_MAILBOX_MSGBUF_TRN_DW0),
143 		      reg);
144 	WREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF0_MAILBOX_MSGBUF_TRN_DW1),
145 				data1);
146 	WREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF0_MAILBOX_MSGBUF_TRN_DW2),
147 				data2);
148 	WREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF0_MAILBOX_MSGBUF_TRN_DW3),
149 				data3);
150 
151 	xgpu_ai_mailbox_set_valid(adev, true);
152 
153 	/* start to poll ack */
154 	r = xgpu_ai_poll_ack(adev);
155 	if (r)
156 		pr_err("Doesn't get ack from pf, continue\n");
157 
158 	xgpu_ai_mailbox_set_valid(adev, false);
159 }
160 
161 static int xgpu_ai_get_pp_clk(struct amdgpu_device *adev, u32 type, char *buf)
162 {
163         int r = 0;
164         u32 req, val, size;
165 
166         if (!amdgim_is_hwperf(adev) || buf == NULL)
167                 return -EBADRQC;
168 
169         switch(type) {
170         case PP_SCLK:
171                 req = IDH_IRQ_GET_PP_SCLK;
172                 break;
173         case PP_MCLK:
174                 req = IDH_IRQ_GET_PP_MCLK;
175                 break;
176         default:
177                 return -EBADRQC;
178         }
179 
180         mutex_lock(&adev->virt.dpm_mutex);
181 
182         xgpu_ai_mailbox_trans_msg(adev, req, 0, 0, 0);
183 
184         r = xgpu_ai_poll_msg(adev, IDH_SUCCESS);
185         if (!r && adev->fw_vram_usage.va != NULL) {
186                 val = RREG32_NO_KIQ(
187                         SOC15_REG_OFFSET(NBIO, 0,
188                                          mmBIF_BX_PF0_MAILBOX_MSGBUF_RCV_DW1));
189                 size = strnlen((((char *)adev->virt.fw_reserve.p_pf2vf) +
190                                 val), PAGE_SIZE);
191 
192                 if (size < PAGE_SIZE)
193                         strcpy(buf,((char *)adev->virt.fw_reserve.p_pf2vf + val));
194                 else
195                         size = 0;
196 
197                 r = size;
198                 goto out;
199         }
200 
201         r = xgpu_ai_poll_msg(adev, IDH_FAIL);
202         if(r)
203                 pr_info("%s DPM request failed",
204                         (type == PP_SCLK)? "SCLK" : "MCLK");
205 
206 out:
207         mutex_unlock(&adev->virt.dpm_mutex);
208         return r;
209 }
210 
211 static int xgpu_ai_force_dpm_level(struct amdgpu_device *adev, u32 level)
212 {
213         int r = 0;
214         u32 req = IDH_IRQ_FORCE_DPM_LEVEL;
215 
216         if (!amdgim_is_hwperf(adev))
217                 return -EBADRQC;
218 
219         mutex_lock(&adev->virt.dpm_mutex);
220         xgpu_ai_mailbox_trans_msg(adev, req, level, 0, 0);
221 
222         r = xgpu_ai_poll_msg(adev, IDH_SUCCESS);
223         if (!r)
224                 goto out;
225 
226         r = xgpu_ai_poll_msg(adev, IDH_FAIL);
227         if (!r)
228                 pr_info("DPM request failed");
229         else
230                 pr_info("Mailbox is broken");
231 
232 out:
233         mutex_unlock(&adev->virt.dpm_mutex);
234         return r;
235 }
236 
237 static int xgpu_ai_send_access_requests(struct amdgpu_device *adev,
238 					enum idh_request req)
239 {
240 	int r;
241 
242 	xgpu_ai_mailbox_trans_msg(adev, req, 0, 0, 0);
243 
244 	/* start to check msg if request is idh_req_gpu_init_access */
245 	if (req == IDH_REQ_GPU_INIT_ACCESS ||
246 		req == IDH_REQ_GPU_FINI_ACCESS ||
247 		req == IDH_REQ_GPU_RESET_ACCESS) {
248 		r = xgpu_ai_poll_msg(adev, IDH_READY_TO_ACCESS_GPU);
249 		if (r) {
250 			pr_err("Doesn't get READY_TO_ACCESS_GPU from pf, give up\n");
251 			return r;
252 		}
253 		/* Retrieve checksum from mailbox2 */
254 		if (req == IDH_REQ_GPU_INIT_ACCESS || req == IDH_REQ_GPU_RESET_ACCESS) {
255 			adev->virt.fw_reserve.checksum_key =
256 				RREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0,
257 					mmBIF_BX_PF0_MAILBOX_MSGBUF_RCV_DW2));
258 		}
259 	}
260 
261 	return 0;
262 }
263 
264 static int xgpu_ai_request_reset(struct amdgpu_device *adev)
265 {
266 	return xgpu_ai_send_access_requests(adev, IDH_REQ_GPU_RESET_ACCESS);
267 }
268 
269 static int xgpu_ai_request_full_gpu_access(struct amdgpu_device *adev,
270 					   bool init)
271 {
272 	enum idh_request req;
273 
274 	req = init ? IDH_REQ_GPU_INIT_ACCESS : IDH_REQ_GPU_FINI_ACCESS;
275 	return xgpu_ai_send_access_requests(adev, req);
276 }
277 
278 static int xgpu_ai_release_full_gpu_access(struct amdgpu_device *adev,
279 					   bool init)
280 {
281 	enum idh_request req;
282 	int r = 0;
283 
284 	req = init ? IDH_REL_GPU_INIT_ACCESS : IDH_REL_GPU_FINI_ACCESS;
285 	r = xgpu_ai_send_access_requests(adev, req);
286 
287 	return r;
288 }
289 
290 static int xgpu_ai_mailbox_ack_irq(struct amdgpu_device *adev,
291 					struct amdgpu_irq_src *source,
292 					struct amdgpu_iv_entry *entry)
293 {
294 	DRM_DEBUG("get ack intr and do nothing.\n");
295 	return 0;
296 }
297 
298 static int xgpu_ai_set_mailbox_ack_irq(struct amdgpu_device *adev,
299 					struct amdgpu_irq_src *source,
300 					unsigned type,
301 					enum amdgpu_interrupt_state state)
302 {
303 	u32 tmp = RREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF0_MAILBOX_INT_CNTL));
304 
305 	tmp = REG_SET_FIELD(tmp, BIF_BX_PF0_MAILBOX_INT_CNTL, ACK_INT_EN,
306 				(state == AMDGPU_IRQ_STATE_ENABLE) ? 1 : 0);
307 	WREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF0_MAILBOX_INT_CNTL), tmp);
308 
309 	return 0;
310 }
311 
312 static void xgpu_ai_mailbox_flr_work(struct work_struct *work)
313 {
314 	struct amdgpu_virt *virt = container_of(work, struct amdgpu_virt, flr_work);
315 	struct amdgpu_device *adev = container_of(virt, struct amdgpu_device, virt);
316 	int timeout = AI_MAILBOX_POLL_FLR_TIMEDOUT;
317 	int locked;
318 
319 	/* block amdgpu_gpu_recover till msg FLR COMPLETE received,
320 	 * otherwise the mailbox msg will be ruined/reseted by
321 	 * the VF FLR.
322 	 *
323 	 * we can unlock the lock_reset to allow "amdgpu_job_timedout"
324 	 * to run gpu_recover() after FLR_NOTIFICATION_CMPL received
325 	 * which means host side had finished this VF's FLR.
326 	 */
327 	locked = mutex_trylock(&adev->lock_reset);
328 	if (locked)
329 		adev->in_gpu_reset = 1;
330 
331 	do {
332 		if (xgpu_ai_mailbox_peek_msg(adev) == IDH_FLR_NOTIFICATION_CMPL)
333 			goto flr_done;
334 
335 		msleep(10);
336 		timeout -= 10;
337 	} while (timeout > 1);
338 
339 flr_done:
340 	if (locked) {
341 		adev->in_gpu_reset = 0;
342 		mutex_unlock(&adev->lock_reset);
343 	}
344 
345 	/* Trigger recovery for world switch failure if no TDR */
346 	if (amdgpu_device_should_recover_gpu(adev)
347 		&& adev->sdma_timeout == MAX_SCHEDULE_TIMEOUT)
348 		amdgpu_device_gpu_recover(adev, NULL);
349 }
350 
351 static int xgpu_ai_set_mailbox_rcv_irq(struct amdgpu_device *adev,
352 				       struct amdgpu_irq_src *src,
353 				       unsigned type,
354 				       enum amdgpu_interrupt_state state)
355 {
356 	u32 tmp = RREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF0_MAILBOX_INT_CNTL));
357 
358 	tmp = REG_SET_FIELD(tmp, BIF_BX_PF0_MAILBOX_INT_CNTL, VALID_INT_EN,
359 			    (state == AMDGPU_IRQ_STATE_ENABLE) ? 1 : 0);
360 	WREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF0_MAILBOX_INT_CNTL), tmp);
361 
362 	return 0;
363 }
364 
365 static int xgpu_ai_mailbox_rcv_irq(struct amdgpu_device *adev,
366 				   struct amdgpu_irq_src *source,
367 				   struct amdgpu_iv_entry *entry)
368 {
369 	enum idh_event event = xgpu_ai_mailbox_peek_msg(adev);
370 
371 	switch (event) {
372 		case IDH_FLR_NOTIFICATION:
373 		if (amdgpu_sriov_runtime(adev))
374 			schedule_work(&adev->virt.flr_work);
375 		break;
376 		case IDH_QUERY_ALIVE:
377 			xgpu_ai_mailbox_send_ack(adev);
378 			break;
379 		/* READY_TO_ACCESS_GPU is fetched by kernel polling, IRQ can ignore
380 		 * it byfar since that polling thread will handle it,
381 		 * other msg like flr complete is not handled here.
382 		 */
383 		case IDH_CLR_MSG_BUF:
384 		case IDH_FLR_NOTIFICATION_CMPL:
385 		case IDH_READY_TO_ACCESS_GPU:
386 		default:
387 		break;
388 	}
389 
390 	return 0;
391 }
392 
393 static const struct amdgpu_irq_src_funcs xgpu_ai_mailbox_ack_irq_funcs = {
394 	.set = xgpu_ai_set_mailbox_ack_irq,
395 	.process = xgpu_ai_mailbox_ack_irq,
396 };
397 
398 static const struct amdgpu_irq_src_funcs xgpu_ai_mailbox_rcv_irq_funcs = {
399 	.set = xgpu_ai_set_mailbox_rcv_irq,
400 	.process = xgpu_ai_mailbox_rcv_irq,
401 };
402 
403 void xgpu_ai_mailbox_set_irq_funcs(struct amdgpu_device *adev)
404 {
405 	adev->virt.ack_irq.num_types = 1;
406 	adev->virt.ack_irq.funcs = &xgpu_ai_mailbox_ack_irq_funcs;
407 	adev->virt.rcv_irq.num_types = 1;
408 	adev->virt.rcv_irq.funcs = &xgpu_ai_mailbox_rcv_irq_funcs;
409 }
410 
411 int xgpu_ai_mailbox_add_irq_id(struct amdgpu_device *adev)
412 {
413 	int r;
414 
415 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_BIF, 135, &adev->virt.rcv_irq);
416 	if (r)
417 		return r;
418 
419 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_BIF, 138, &adev->virt.ack_irq);
420 	if (r) {
421 		amdgpu_irq_put(adev, &adev->virt.rcv_irq, 0);
422 		return r;
423 	}
424 
425 	return 0;
426 }
427 
428 int xgpu_ai_mailbox_get_irq(struct amdgpu_device *adev)
429 {
430 	int r;
431 
432 	r = amdgpu_irq_get(adev, &adev->virt.rcv_irq, 0);
433 	if (r)
434 		return r;
435 	r = amdgpu_irq_get(adev, &adev->virt.ack_irq, 0);
436 	if (r) {
437 		amdgpu_irq_put(adev, &adev->virt.rcv_irq, 0);
438 		return r;
439 	}
440 
441 	INIT_WORK(&adev->virt.flr_work, xgpu_ai_mailbox_flr_work);
442 
443 	return 0;
444 }
445 
446 void xgpu_ai_mailbox_put_irq(struct amdgpu_device *adev)
447 {
448 	amdgpu_irq_put(adev, &adev->virt.ack_irq, 0);
449 	amdgpu_irq_put(adev, &adev->virt.rcv_irq, 0);
450 }
451 
452 const struct amdgpu_virt_ops xgpu_ai_virt_ops = {
453 	.req_full_gpu	= xgpu_ai_request_full_gpu_access,
454 	.rel_full_gpu	= xgpu_ai_release_full_gpu_access,
455 	.reset_gpu = xgpu_ai_request_reset,
456 	.wait_reset = NULL,
457 	.trans_msg = xgpu_ai_mailbox_trans_msg,
458 	.get_pp_clk = xgpu_ai_get_pp_clk,
459 	.force_dpm_level = xgpu_ai_force_dpm_level,
460 };
461