xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c (revision 2a598d0b)
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 #include "amdgpu_reset.h"
36 
37 static void xgpu_ai_mailbox_send_ack(struct amdgpu_device *adev)
38 {
39 	WREG8(AI_MAIBOX_CONTROL_RCV_OFFSET_BYTE, 2);
40 }
41 
42 static void xgpu_ai_mailbox_set_valid(struct amdgpu_device *adev, bool val)
43 {
44 	WREG8(AI_MAIBOX_CONTROL_TRN_OFFSET_BYTE, val ? 1 : 0);
45 }
46 
47 /*
48  * this peek_msg could *only* be called in IRQ routine becuase in IRQ routine
49  * RCV_MSG_VALID filed of BIF_BX_PF0_MAILBOX_CONTROL must already be set to 1
50  * by host.
51  *
52  * if called no in IRQ routine, this peek_msg cannot guaranteed to return the
53  * correct value since it doesn't return the RCV_DW0 under the case that
54  * RCV_MSG_VALID is set by host.
55  */
56 static enum idh_event xgpu_ai_mailbox_peek_msg(struct amdgpu_device *adev)
57 {
58 	return RREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0,
59 				mmBIF_BX_PF0_MAILBOX_MSGBUF_RCV_DW0));
60 }
61 
62 
63 static int xgpu_ai_mailbox_rcv_msg(struct amdgpu_device *adev,
64 				   enum idh_event event)
65 {
66 	u32 reg;
67 
68 	reg = RREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0,
69 					     mmBIF_BX_PF0_MAILBOX_MSGBUF_RCV_DW0));
70 	if (reg != event)
71 		return -ENOENT;
72 
73 	xgpu_ai_mailbox_send_ack(adev);
74 
75 	return 0;
76 }
77 
78 static uint8_t xgpu_ai_peek_ack(struct amdgpu_device *adev) {
79 	return RREG8(AI_MAIBOX_CONTROL_TRN_OFFSET_BYTE) & 2;
80 }
81 
82 static int xgpu_ai_poll_ack(struct amdgpu_device *adev)
83 {
84 	int timeout  = AI_MAILBOX_POLL_ACK_TIMEDOUT;
85 	u8 reg;
86 
87 	do {
88 		reg = RREG8(AI_MAIBOX_CONTROL_TRN_OFFSET_BYTE);
89 		if (reg & 2)
90 			return 0;
91 
92 		mdelay(5);
93 		timeout -= 5;
94 	} while (timeout > 1);
95 
96 	pr_err("Doesn't get TRN_MSG_ACK from pf in %d msec\n", AI_MAILBOX_POLL_ACK_TIMEDOUT);
97 
98 	return -ETIME;
99 }
100 
101 static int xgpu_ai_poll_msg(struct amdgpu_device *adev, enum idh_event event)
102 {
103 	int r, timeout = AI_MAILBOX_POLL_MSG_TIMEDOUT;
104 
105 	do {
106 		r = xgpu_ai_mailbox_rcv_msg(adev, event);
107 		if (!r)
108 			return 0;
109 
110 		msleep(10);
111 		timeout -= 10;
112 	} while (timeout > 1);
113 
114 	pr_err("Doesn't get msg:%d from pf, error=%d\n", event, r);
115 
116 	return -ETIME;
117 }
118 
119 static void xgpu_ai_mailbox_trans_msg (struct amdgpu_device *adev,
120 	      enum idh_request req, u32 data1, u32 data2, u32 data3) {
121 	u32 reg;
122 	int r;
123 	uint8_t trn;
124 
125 	/* IMPORTANT:
126 	 * clear TRN_MSG_VALID valid to clear host's RCV_MSG_ACK
127 	 * and with host's RCV_MSG_ACK cleared hw automatically clear host's RCV_MSG_ACK
128 	 * which lead to VF's TRN_MSG_ACK cleared, otherwise below xgpu_ai_poll_ack()
129 	 * will return immediatly
130 	 */
131 	do {
132 		xgpu_ai_mailbox_set_valid(adev, false);
133 		trn = xgpu_ai_peek_ack(adev);
134 		if (trn) {
135 			pr_err("trn=%x ACK should not assert! wait again !\n", trn);
136 			msleep(1);
137 		}
138 	} while(trn);
139 
140 	reg = RREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0,
141 					     mmBIF_BX_PF0_MAILBOX_MSGBUF_TRN_DW0));
142 	reg = REG_SET_FIELD(reg, BIF_BX_PF0_MAILBOX_MSGBUF_TRN_DW0,
143 			    MSGBUF_DATA, req);
144 	WREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF0_MAILBOX_MSGBUF_TRN_DW0),
145 		      reg);
146 	WREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF0_MAILBOX_MSGBUF_TRN_DW1),
147 				data1);
148 	WREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF0_MAILBOX_MSGBUF_TRN_DW2),
149 				data2);
150 	WREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF0_MAILBOX_MSGBUF_TRN_DW3),
151 				data3);
152 
153 	xgpu_ai_mailbox_set_valid(adev, true);
154 
155 	/* start to poll ack */
156 	r = xgpu_ai_poll_ack(adev);
157 	if (r)
158 		pr_err("Doesn't get ack from pf, continue\n");
159 
160 	xgpu_ai_mailbox_set_valid(adev, false);
161 }
162 
163 static int xgpu_ai_send_access_requests(struct amdgpu_device *adev,
164 					enum idh_request req)
165 {
166 	int r;
167 
168 	xgpu_ai_mailbox_trans_msg(adev, req, 0, 0, 0);
169 
170 	/* start to check msg if request is idh_req_gpu_init_access */
171 	if (req == IDH_REQ_GPU_INIT_ACCESS ||
172 		req == IDH_REQ_GPU_FINI_ACCESS ||
173 		req == IDH_REQ_GPU_RESET_ACCESS) {
174 		r = xgpu_ai_poll_msg(adev, IDH_READY_TO_ACCESS_GPU);
175 		if (r) {
176 			pr_err("Doesn't get READY_TO_ACCESS_GPU from pf, give up\n");
177 			return r;
178 		}
179 		/* Retrieve checksum from mailbox2 */
180 		if (req == IDH_REQ_GPU_INIT_ACCESS || req == IDH_REQ_GPU_RESET_ACCESS) {
181 			adev->virt.fw_reserve.checksum_key =
182 				RREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0,
183 					mmBIF_BX_PF0_MAILBOX_MSGBUF_RCV_DW2));
184 		}
185 	} else if (req == IDH_REQ_GPU_INIT_DATA){
186 		/* Dummy REQ_GPU_INIT_DATA handling */
187 		r = xgpu_ai_poll_msg(adev, IDH_REQ_GPU_INIT_DATA_READY);
188 		/* version set to 0 since dummy */
189 		adev->virt.req_init_data_ver = 0;
190 	}
191 
192 	return 0;
193 }
194 
195 static int xgpu_ai_request_reset(struct amdgpu_device *adev)
196 {
197 	int ret, i = 0;
198 
199 	while (i < AI_MAILBOX_POLL_MSG_REP_MAX) {
200 		ret = xgpu_ai_send_access_requests(adev, IDH_REQ_GPU_RESET_ACCESS);
201 		if (!ret)
202 			break;
203 		i++;
204 	}
205 
206 	return ret;
207 }
208 
209 static int xgpu_ai_request_full_gpu_access(struct amdgpu_device *adev,
210 					   bool init)
211 {
212 	enum idh_request req;
213 
214 	req = init ? IDH_REQ_GPU_INIT_ACCESS : IDH_REQ_GPU_FINI_ACCESS;
215 	return xgpu_ai_send_access_requests(adev, req);
216 }
217 
218 static int xgpu_ai_release_full_gpu_access(struct amdgpu_device *adev,
219 					   bool init)
220 {
221 	enum idh_request req;
222 	int r = 0;
223 
224 	req = init ? IDH_REL_GPU_INIT_ACCESS : IDH_REL_GPU_FINI_ACCESS;
225 	r = xgpu_ai_send_access_requests(adev, req);
226 
227 	return r;
228 }
229 
230 static int xgpu_ai_mailbox_ack_irq(struct amdgpu_device *adev,
231 					struct amdgpu_irq_src *source,
232 					struct amdgpu_iv_entry *entry)
233 {
234 	DRM_DEBUG("get ack intr and do nothing.\n");
235 	return 0;
236 }
237 
238 static int xgpu_ai_set_mailbox_ack_irq(struct amdgpu_device *adev,
239 					struct amdgpu_irq_src *source,
240 					unsigned type,
241 					enum amdgpu_interrupt_state state)
242 {
243 	u32 tmp = RREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF0_MAILBOX_INT_CNTL));
244 
245 	tmp = REG_SET_FIELD(tmp, BIF_BX_PF0_MAILBOX_INT_CNTL, ACK_INT_EN,
246 				(state == AMDGPU_IRQ_STATE_ENABLE) ? 1 : 0);
247 	WREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF0_MAILBOX_INT_CNTL), tmp);
248 
249 	return 0;
250 }
251 
252 static void xgpu_ai_mailbox_flr_work(struct work_struct *work)
253 {
254 	struct amdgpu_virt *virt = container_of(work, struct amdgpu_virt, flr_work);
255 	struct amdgpu_device *adev = container_of(virt, struct amdgpu_device, virt);
256 	int timeout = AI_MAILBOX_POLL_FLR_TIMEDOUT;
257 
258 	/* block amdgpu_gpu_recover till msg FLR COMPLETE received,
259 	 * otherwise the mailbox msg will be ruined/reseted by
260 	 * the VF FLR.
261 	 */
262 	if (atomic_cmpxchg(&adev->reset_domain->in_gpu_reset, 0, 1) != 0)
263 		return;
264 
265 	down_write(&adev->reset_domain->sem);
266 
267 	amdgpu_virt_fini_data_exchange(adev);
268 
269 	xgpu_ai_mailbox_trans_msg(adev, IDH_READY_TO_RESET, 0, 0, 0);
270 
271 	do {
272 		if (xgpu_ai_mailbox_peek_msg(adev) == IDH_FLR_NOTIFICATION_CMPL)
273 			goto flr_done;
274 
275 		msleep(10);
276 		timeout -= 10;
277 	} while (timeout > 1);
278 
279 flr_done:
280 	atomic_set(&adev->reset_domain->in_gpu_reset, 0);
281 	up_write(&adev->reset_domain->sem);
282 
283 	/* Trigger recovery for world switch failure if no TDR */
284 	if (amdgpu_device_should_recover_gpu(adev)
285 		&& (!amdgpu_device_has_job_running(adev) ||
286 			adev->sdma_timeout == MAX_SCHEDULE_TIMEOUT)) {
287 		struct amdgpu_reset_context reset_context;
288 		memset(&reset_context, 0, sizeof(reset_context));
289 
290 		reset_context.method = AMD_RESET_METHOD_NONE;
291 		reset_context.reset_req_dev = adev;
292 		clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
293 
294 		amdgpu_device_gpu_recover(adev, NULL, &reset_context);
295 	}
296 }
297 
298 static int xgpu_ai_set_mailbox_rcv_irq(struct amdgpu_device *adev,
299 				       struct amdgpu_irq_src *src,
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, VALID_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 int xgpu_ai_mailbox_rcv_irq(struct amdgpu_device *adev,
313 				   struct amdgpu_irq_src *source,
314 				   struct amdgpu_iv_entry *entry)
315 {
316 	enum idh_event event = xgpu_ai_mailbox_peek_msg(adev);
317 
318 	switch (event) {
319 		case IDH_FLR_NOTIFICATION:
320 		if (amdgpu_sriov_runtime(adev) && !amdgpu_in_reset(adev))
321 			WARN_ONCE(!amdgpu_reset_domain_schedule(adev->reset_domain,
322 								&adev->virt.flr_work),
323 				  "Failed to queue work! at %s",
324 				  __func__);
325 		break;
326 		case IDH_QUERY_ALIVE:
327 			xgpu_ai_mailbox_send_ack(adev);
328 			break;
329 		/* READY_TO_ACCESS_GPU is fetched by kernel polling, IRQ can ignore
330 		 * it byfar since that polling thread will handle it,
331 		 * other msg like flr complete is not handled here.
332 		 */
333 		case IDH_CLR_MSG_BUF:
334 		case IDH_FLR_NOTIFICATION_CMPL:
335 		case IDH_READY_TO_ACCESS_GPU:
336 		default:
337 		break;
338 	}
339 
340 	return 0;
341 }
342 
343 static const struct amdgpu_irq_src_funcs xgpu_ai_mailbox_ack_irq_funcs = {
344 	.set = xgpu_ai_set_mailbox_ack_irq,
345 	.process = xgpu_ai_mailbox_ack_irq,
346 };
347 
348 static const struct amdgpu_irq_src_funcs xgpu_ai_mailbox_rcv_irq_funcs = {
349 	.set = xgpu_ai_set_mailbox_rcv_irq,
350 	.process = xgpu_ai_mailbox_rcv_irq,
351 };
352 
353 void xgpu_ai_mailbox_set_irq_funcs(struct amdgpu_device *adev)
354 {
355 	adev->virt.ack_irq.num_types = 1;
356 	adev->virt.ack_irq.funcs = &xgpu_ai_mailbox_ack_irq_funcs;
357 	adev->virt.rcv_irq.num_types = 1;
358 	adev->virt.rcv_irq.funcs = &xgpu_ai_mailbox_rcv_irq_funcs;
359 }
360 
361 int xgpu_ai_mailbox_add_irq_id(struct amdgpu_device *adev)
362 {
363 	int r;
364 
365 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_BIF, 135, &adev->virt.rcv_irq);
366 	if (r)
367 		return r;
368 
369 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_BIF, 138, &adev->virt.ack_irq);
370 	if (r) {
371 		amdgpu_irq_put(adev, &adev->virt.rcv_irq, 0);
372 		return r;
373 	}
374 
375 	return 0;
376 }
377 
378 int xgpu_ai_mailbox_get_irq(struct amdgpu_device *adev)
379 {
380 	int r;
381 
382 	r = amdgpu_irq_get(adev, &adev->virt.rcv_irq, 0);
383 	if (r)
384 		return r;
385 	r = amdgpu_irq_get(adev, &adev->virt.ack_irq, 0);
386 	if (r) {
387 		amdgpu_irq_put(adev, &adev->virt.rcv_irq, 0);
388 		return r;
389 	}
390 
391 	INIT_WORK(&adev->virt.flr_work, xgpu_ai_mailbox_flr_work);
392 
393 	return 0;
394 }
395 
396 void xgpu_ai_mailbox_put_irq(struct amdgpu_device *adev)
397 {
398 	amdgpu_irq_put(adev, &adev->virt.ack_irq, 0);
399 	amdgpu_irq_put(adev, &adev->virt.rcv_irq, 0);
400 }
401 
402 static int xgpu_ai_request_init_data(struct amdgpu_device *adev)
403 {
404 	return xgpu_ai_send_access_requests(adev, IDH_REQ_GPU_INIT_DATA);
405 }
406 
407 static void xgpu_ai_ras_poison_handler(struct amdgpu_device *adev)
408 {
409 	xgpu_ai_send_access_requests(adev, IDH_RAS_POISON);
410 }
411 
412 const struct amdgpu_virt_ops xgpu_ai_virt_ops = {
413 	.req_full_gpu	= xgpu_ai_request_full_gpu_access,
414 	.rel_full_gpu	= xgpu_ai_release_full_gpu_access,
415 	.reset_gpu = xgpu_ai_request_reset,
416 	.wait_reset = NULL,
417 	.trans_msg = xgpu_ai_mailbox_trans_msg,
418 	.req_init_data  = xgpu_ai_request_init_data,
419 	.ras_poison_handler = xgpu_ai_ras_poison_handler,
420 };
421