xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/mxgpu_nv.c (revision e7bae9bb) 
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_2_3_offset.h"
26 #include "nbio/nbio_2_3_sh_mask.h"
27 #include "gc/gc_10_1_0_offset.h"
28 #include "gc/gc_10_1_0_sh_mask.h"
29 #include "soc15.h"
30 #include "navi10_ih.h"
31 #include "soc15_common.h"
32 #include "mxgpu_nv.h"
33 
34 static void xgpu_nv_mailbox_send_ack(struct amdgpu_device *adev)
35 {
36 	WREG8(NV_MAIBOX_CONTROL_RCV_OFFSET_BYTE, 2);
37 }
38 
39 static void xgpu_nv_mailbox_set_valid(struct amdgpu_device *adev, bool val)
40 {
41 	WREG8(NV_MAIBOX_CONTROL_TRN_OFFSET_BYTE, val ? 1 : 0);
42 }
43 
44 /*
45  * this peek_msg could *only* be called in IRQ routine becuase in IRQ routine
46  * RCV_MSG_VALID filed of BIF_BX_PF_MAILBOX_CONTROL must already be set to 1
47  * by host.
48  *
49  * if called no in IRQ routine, this peek_msg cannot guaranteed to return the
50  * correct value since it doesn't return the RCV_DW0 under the case that
51  * RCV_MSG_VALID is set by host.
52  */
53 static enum idh_event xgpu_nv_mailbox_peek_msg(struct amdgpu_device *adev)
54 {
55 	return RREG32_NO_KIQ(mmMAILBOX_MSGBUF_RCV_DW0);
56 }
57 
58 
59 static int xgpu_nv_mailbox_rcv_msg(struct amdgpu_device *adev,
60 				   enum idh_event event)
61 {
62 	u32 reg;
63 
64 	reg = RREG32_NO_KIQ(mmMAILBOX_MSGBUF_RCV_DW0);
65 	if (reg != event)
66 		return -ENOENT;
67 
68 	xgpu_nv_mailbox_send_ack(adev);
69 
70 	return 0;
71 }
72 
73 static uint8_t xgpu_nv_peek_ack(struct amdgpu_device *adev)
74 {
75 	return RREG8(NV_MAIBOX_CONTROL_TRN_OFFSET_BYTE) & 2;
76 }
77 
78 static int xgpu_nv_poll_ack(struct amdgpu_device *adev)
79 {
80 	int timeout  = NV_MAILBOX_POLL_ACK_TIMEDOUT;
81 	u8 reg;
82 
83 	do {
84 		reg = RREG8(NV_MAIBOX_CONTROL_TRN_OFFSET_BYTE);
85 		if (reg & 2)
86 			return 0;
87 
88 		mdelay(5);
89 		timeout -= 5;
90 	} while (timeout > 1);
91 
92 	pr_err("Doesn't get TRN_MSG_ACK from pf in %d msec\n", NV_MAILBOX_POLL_ACK_TIMEDOUT);
93 
94 	return -ETIME;
95 }
96 
97 static int xgpu_nv_poll_msg(struct amdgpu_device *adev, enum idh_event event)
98 {
99 	int r, timeout = NV_MAILBOX_POLL_MSG_TIMEDOUT;
100 
101 	do {
102 		r = xgpu_nv_mailbox_rcv_msg(adev, event);
103 		if (!r)
104 			return 0;
105 
106 		msleep(10);
107 		timeout -= 10;
108 	} while (timeout > 1);
109 
110 
111 	return -ETIME;
112 }
113 
114 static void xgpu_nv_mailbox_trans_msg (struct amdgpu_device *adev,
115 	      enum idh_request req, u32 data1, u32 data2, u32 data3)
116 {
117 	int r;
118 	uint8_t trn;
119 
120 	/* IMPORTANT:
121 	 * clear TRN_MSG_VALID valid to clear host's RCV_MSG_ACK
122 	 * and with host's RCV_MSG_ACK cleared hw automatically clear host's RCV_MSG_ACK
123 	 * which lead to VF's TRN_MSG_ACK cleared, otherwise below xgpu_nv_poll_ack()
124 	 * will return immediatly
125 	 */
126 	do {
127 		xgpu_nv_mailbox_set_valid(adev, false);
128 		trn = xgpu_nv_peek_ack(adev);
129 		if (trn) {
130 			pr_err("trn=%x ACK should not assert! wait again !\n", trn);
131 			msleep(1);
132 		}
133 	} while (trn);
134 
135 	WREG32_NO_KIQ(mmMAILBOX_MSGBUF_TRN_DW0, req);
136 	WREG32_NO_KIQ(mmMAILBOX_MSGBUF_TRN_DW1, data1);
137 	WREG32_NO_KIQ(mmMAILBOX_MSGBUF_TRN_DW2, data2);
138 	WREG32_NO_KIQ(mmMAILBOX_MSGBUF_TRN_DW3, data3);
139 	xgpu_nv_mailbox_set_valid(adev, true);
140 
141 	/* start to poll ack */
142 	r = xgpu_nv_poll_ack(adev);
143 	if (r)
144 		pr_err("Doesn't get ack from pf, continue\n");
145 
146 	xgpu_nv_mailbox_set_valid(adev, false);
147 }
148 
149 static int xgpu_nv_send_access_requests(struct amdgpu_device *adev,
150 					enum idh_request req)
151 {
152 	int r;
153 	enum idh_event event = -1;
154 
155 	xgpu_nv_mailbox_trans_msg(adev, req, 0, 0, 0);
156 
157 	switch (req) {
158 	case IDH_REQ_GPU_INIT_ACCESS:
159 	case IDH_REQ_GPU_FINI_ACCESS:
160 	case IDH_REQ_GPU_RESET_ACCESS:
161 		event = IDH_READY_TO_ACCESS_GPU;
162 		break;
163 	case IDH_REQ_GPU_INIT_DATA:
164 		event = IDH_REQ_GPU_INIT_DATA_READY;
165 		break;
166 	default:
167 		break;
168 	}
169 
170 	if (event != -1) {
171 		r = xgpu_nv_poll_msg(adev, event);
172 		if (r) {
173 			if (req != IDH_REQ_GPU_INIT_DATA) {
174 				pr_err("Doesn't get msg:%d from pf, error=%d\n", event, r);
175 				return r;
176 			}
177 			else /* host doesn't support REQ_GPU_INIT_DATA handshake */
178 				adev->virt.req_init_data_ver = 0;
179 		} else {
180 			if (req == IDH_REQ_GPU_INIT_DATA)
181 			{
182 				adev->virt.req_init_data_ver =
183 					RREG32_NO_KIQ(mmMAILBOX_MSGBUF_RCV_DW1);
184 
185 				/* assume V1 in case host doesn't set version number */
186 				if (adev->virt.req_init_data_ver < 1)
187 					adev->virt.req_init_data_ver = 1;
188 			}
189 		}
190 
191 		/* Retrieve checksum from mailbox2 */
192 		if (req == IDH_REQ_GPU_INIT_ACCESS || req == IDH_REQ_GPU_RESET_ACCESS) {
193 			adev->virt.fw_reserve.checksum_key =
194 				RREG32_NO_KIQ(mmMAILBOX_MSGBUF_RCV_DW2);
195 		}
196 	}
197 
198 	return 0;
199 }
200 
201 static int xgpu_nv_request_reset(struct amdgpu_device *adev)
202 {
203 	return xgpu_nv_send_access_requests(adev, IDH_REQ_GPU_RESET_ACCESS);
204 }
205 
206 static int xgpu_nv_request_full_gpu_access(struct amdgpu_device *adev,
207 					   bool init)
208 {
209 	enum idh_request req;
210 
211 	req = init ? IDH_REQ_GPU_INIT_ACCESS : IDH_REQ_GPU_FINI_ACCESS;
212 	return xgpu_nv_send_access_requests(adev, req);
213 }
214 
215 static int xgpu_nv_release_full_gpu_access(struct amdgpu_device *adev,
216 					   bool init)
217 {
218 	enum idh_request req;
219 	int r = 0;
220 
221 	req = init ? IDH_REL_GPU_INIT_ACCESS : IDH_REL_GPU_FINI_ACCESS;
222 	r = xgpu_nv_send_access_requests(adev, req);
223 
224 	return r;
225 }
226 
227 static int xgpu_nv_request_init_data(struct amdgpu_device *adev)
228 {
229 	return xgpu_nv_send_access_requests(adev, IDH_REQ_GPU_INIT_DATA);
230 }
231 
232 static int xgpu_nv_mailbox_ack_irq(struct amdgpu_device *adev,
233 					struct amdgpu_irq_src *source,
234 					struct amdgpu_iv_entry *entry)
235 {
236 	DRM_DEBUG("get ack intr and do nothing.\n");
237 	return 0;
238 }
239 
240 static int xgpu_nv_set_mailbox_ack_irq(struct amdgpu_device *adev,
241 					struct amdgpu_irq_src *source,
242 					unsigned type,
243 					enum amdgpu_interrupt_state state)
244 {
245 	u32 tmp = RREG32_NO_KIQ(mmMAILBOX_INT_CNTL);
246 
247 	if (state == AMDGPU_IRQ_STATE_ENABLE)
248 		tmp |= 2;
249 	else
250 		tmp &= ~2;
251 
252 	WREG32_NO_KIQ(mmMAILBOX_INT_CNTL, tmp);
253 
254 	return 0;
255 }
256 
257 static void xgpu_nv_mailbox_flr_work(struct work_struct *work)
258 {
259 	struct amdgpu_virt *virt = container_of(work, struct amdgpu_virt, flr_work);
260 	struct amdgpu_device *adev = container_of(virt, struct amdgpu_device, virt);
261 	int timeout = NV_MAILBOX_POLL_FLR_TIMEDOUT;
262 	int locked;
263 
264 	/* block amdgpu_gpu_recover till msg FLR COMPLETE received,
265 	 * otherwise the mailbox msg will be ruined/reseted by
266 	 * the VF FLR.
267 	 *
268 	 * we can unlock the lock_reset to allow "amdgpu_job_timedout"
269 	 * to run gpu_recover() after FLR_NOTIFICATION_CMPL received
270 	 * which means host side had finished this VF's FLR.
271 	 */
272 	locked = mutex_trylock(&adev->lock_reset);
273 	if (locked)
274 		adev->in_gpu_reset = true;
275 
276 	do {
277 		if (xgpu_nv_mailbox_peek_msg(adev) == IDH_FLR_NOTIFICATION_CMPL)
278 			goto flr_done;
279 
280 		msleep(10);
281 		timeout -= 10;
282 	} while (timeout > 1);
283 
284 flr_done:
285 	if (locked) {
286 		adev->in_gpu_reset = false;
287 		mutex_unlock(&adev->lock_reset);
288 	}
289 
290 	/* Trigger recovery for world switch failure if no TDR */
291 	if (amdgpu_device_should_recover_gpu(adev)
292 		&& (adev->sdma_timeout == MAX_SCHEDULE_TIMEOUT ||
293 		adev->gfx_timeout == MAX_SCHEDULE_TIMEOUT ||
294 		adev->compute_timeout == MAX_SCHEDULE_TIMEOUT ||
295 		adev->video_timeout == MAX_SCHEDULE_TIMEOUT))
296 		amdgpu_device_gpu_recover(adev, NULL);
297 }
298 
299 static int xgpu_nv_set_mailbox_rcv_irq(struct amdgpu_device *adev,
300 				       struct amdgpu_irq_src *src,
301 				       unsigned type,
302 				       enum amdgpu_interrupt_state state)
303 {
304 	u32 tmp = RREG32_NO_KIQ(mmMAILBOX_INT_CNTL);
305 
306 	if (state == AMDGPU_IRQ_STATE_ENABLE)
307 		tmp |= 1;
308 	else
309 		tmp &= ~1;
310 
311 	WREG32_NO_KIQ(mmMAILBOX_INT_CNTL, tmp);
312 
313 	return 0;
314 }
315 
316 static int xgpu_nv_mailbox_rcv_irq(struct amdgpu_device *adev,
317 				   struct amdgpu_irq_src *source,
318 				   struct amdgpu_iv_entry *entry)
319 {
320 	enum idh_event event = xgpu_nv_mailbox_peek_msg(adev);
321 
322 	switch (event) {
323 	case IDH_FLR_NOTIFICATION:
324 		if (amdgpu_sriov_runtime(adev))
325 			schedule_work(&adev->virt.flr_work);
326 		break;
327 		/* READY_TO_ACCESS_GPU is fetched by kernel polling, IRQ can ignore
328 		 * it byfar since that polling thread will handle it,
329 		 * other msg like flr complete is not handled here.
330 		 */
331 	case IDH_CLR_MSG_BUF:
332 	case IDH_FLR_NOTIFICATION_CMPL:
333 	case IDH_READY_TO_ACCESS_GPU:
334 	default:
335 		break;
336 	}
337 
338 	return 0;
339 }
340 
341 static const struct amdgpu_irq_src_funcs xgpu_nv_mailbox_ack_irq_funcs = {
342 	.set = xgpu_nv_set_mailbox_ack_irq,
343 	.process = xgpu_nv_mailbox_ack_irq,
344 };
345 
346 static const struct amdgpu_irq_src_funcs xgpu_nv_mailbox_rcv_irq_funcs = {
347 	.set = xgpu_nv_set_mailbox_rcv_irq,
348 	.process = xgpu_nv_mailbox_rcv_irq,
349 };
350 
351 void xgpu_nv_mailbox_set_irq_funcs(struct amdgpu_device *adev)
352 {
353 	adev->virt.ack_irq.num_types = 1;
354 	adev->virt.ack_irq.funcs = &xgpu_nv_mailbox_ack_irq_funcs;
355 	adev->virt.rcv_irq.num_types = 1;
356 	adev->virt.rcv_irq.funcs = &xgpu_nv_mailbox_rcv_irq_funcs;
357 }
358 
359 int xgpu_nv_mailbox_add_irq_id(struct amdgpu_device *adev)
360 {
361 	int r;
362 
363 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_BIF, 135, &adev->virt.rcv_irq);
364 	if (r)
365 		return r;
366 
367 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_BIF, 138, &adev->virt.ack_irq);
368 	if (r) {
369 		amdgpu_irq_put(adev, &adev->virt.rcv_irq, 0);
370 		return r;
371 	}
372 
373 	return 0;
374 }
375 
376 int xgpu_nv_mailbox_get_irq(struct amdgpu_device *adev)
377 {
378 	int r;
379 
380 	r = amdgpu_irq_get(adev, &adev->virt.rcv_irq, 0);
381 	if (r)
382 		return r;
383 	r = amdgpu_irq_get(adev, &adev->virt.ack_irq, 0);
384 	if (r) {
385 		amdgpu_irq_put(adev, &adev->virt.rcv_irq, 0);
386 		return r;
387 	}
388 
389 	INIT_WORK(&adev->virt.flr_work, xgpu_nv_mailbox_flr_work);
390 
391 	return 0;
392 }
393 
394 void xgpu_nv_mailbox_put_irq(struct amdgpu_device *adev)
395 {
396 	amdgpu_irq_put(adev, &adev->virt.ack_irq, 0);
397 	amdgpu_irq_put(adev, &adev->virt.rcv_irq, 0);
398 }
399 
400 const struct amdgpu_virt_ops xgpu_nv_virt_ops = {
401 	.req_full_gpu	= xgpu_nv_request_full_gpu_access,
402 	.rel_full_gpu	= xgpu_nv_release_full_gpu_access,
403 	.req_init_data  = xgpu_nv_request_init_data,
404 	.reset_gpu = xgpu_nv_request_reset,
405 	.wait_reset = NULL,
406 	.trans_msg = xgpu_nv_mailbox_trans_msg,
407 };
408