xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/mxgpu_nv.c (revision 7f877908) 
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 #include "mxgpu_ai.h"
34 
35 static void xgpu_nv_mailbox_send_ack(struct amdgpu_device *adev)
36 {
37 	WREG8(NV_MAIBOX_CONTROL_RCV_OFFSET_BYTE, 2);
38 }
39 
40 static void xgpu_nv_mailbox_set_valid(struct amdgpu_device *adev, bool val)
41 {
42 	WREG8(NV_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_PF_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_nv_mailbox_peek_msg(struct amdgpu_device *adev)
55 {
56 	return RREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0,
57 				mmBIF_BX_PF_MAILBOX_MSGBUF_RCV_DW0));
58 }
59 
60 
61 static int xgpu_nv_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_PF_MAILBOX_MSGBUF_RCV_DW0));
68 	if (reg != event)
69 		return -ENOENT;
70 
71 	xgpu_nv_mailbox_send_ack(adev);
72 
73 	return 0;
74 }
75 
76 static uint8_t xgpu_nv_peek_ack(struct amdgpu_device *adev)
77 {
78 	return RREG8(NV_MAIBOX_CONTROL_TRN_OFFSET_BYTE) & 2;
79 }
80 
81 static int xgpu_nv_poll_ack(struct amdgpu_device *adev)
82 {
83 	int timeout  = NV_MAILBOX_POLL_ACK_TIMEDOUT;
84 	u8 reg;
85 
86 	do {
87 		reg = RREG8(NV_MAIBOX_CONTROL_TRN_OFFSET_BYTE);
88 		if (reg & 2)
89 			return 0;
90 
91 		mdelay(5);
92 		timeout -= 5;
93 	} while (timeout > 1);
94 
95 	pr_err("Doesn't get TRN_MSG_ACK from pf in %d msec\n", NV_MAILBOX_POLL_ACK_TIMEDOUT);
96 
97 	return -ETIME;
98 }
99 
100 static int xgpu_nv_poll_msg(struct amdgpu_device *adev, enum idh_event event)
101 {
102 	int r, timeout = NV_MAILBOX_POLL_MSG_TIMEDOUT;
103 
104 	do {
105 		r = xgpu_nv_mailbox_rcv_msg(adev, event);
106 		if (!r)
107 			return 0;
108 
109 		msleep(10);
110 		timeout -= 10;
111 	} while (timeout > 1);
112 
113 	pr_err("Doesn't get msg:%d from pf, error=%d\n", event, r);
114 
115 	return -ETIME;
116 }
117 
118 static void xgpu_nv_mailbox_trans_msg (struct amdgpu_device *adev,
119 	      enum idh_request req, u32 data1, u32 data2, u32 data3)
120 {
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_nv_poll_ack()
129 	 * will return immediatly
130 	 */
131 	do {
132 		xgpu_nv_mailbox_set_valid(adev, false);
133 		trn = xgpu_nv_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_PF_MAILBOX_MSGBUF_TRN_DW0));
142 	reg = REG_SET_FIELD(reg, BIF_BX_PF_MAILBOX_MSGBUF_TRN_DW0,
143 			    MSGBUF_DATA, req);
144 	WREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF_MAILBOX_MSGBUF_TRN_DW0),
145 		      reg);
146 	WREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF_MAILBOX_MSGBUF_TRN_DW1),
147 				data1);
148 	WREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF_MAILBOX_MSGBUF_TRN_DW2),
149 				data2);
150 	WREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF_MAILBOX_MSGBUF_TRN_DW3),
151 				data3);
152 
153 	xgpu_nv_mailbox_set_valid(adev, true);
154 
155 	/* start to poll ack */
156 	r = xgpu_nv_poll_ack(adev);
157 	if (r)
158 		pr_err("Doesn't get ack from pf, continue\n");
159 
160 	xgpu_nv_mailbox_set_valid(adev, false);
161 }
162 
163 static int xgpu_nv_send_access_requests(struct amdgpu_device *adev,
164 					enum idh_request req)
165 {
166 	int r;
167 
168 	xgpu_nv_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_nv_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_PF_MAILBOX_MSGBUF_RCV_DW2));
184 		}
185 	}
186 
187 	return 0;
188 }
189 
190 static int xgpu_nv_request_reset(struct amdgpu_device *adev)
191 {
192 	return xgpu_nv_send_access_requests(adev, IDH_REQ_GPU_RESET_ACCESS);
193 }
194 
195 static int xgpu_nv_request_full_gpu_access(struct amdgpu_device *adev,
196 					   bool init)
197 {
198 	enum idh_request req;
199 
200 	req = init ? IDH_REQ_GPU_INIT_ACCESS : IDH_REQ_GPU_FINI_ACCESS;
201 	return xgpu_nv_send_access_requests(adev, req);
202 }
203 
204 static int xgpu_nv_release_full_gpu_access(struct amdgpu_device *adev,
205 					   bool init)
206 {
207 	enum idh_request req;
208 	int r = 0;
209 
210 	req = init ? IDH_REL_GPU_INIT_ACCESS : IDH_REL_GPU_FINI_ACCESS;
211 	r = xgpu_nv_send_access_requests(adev, req);
212 
213 	return r;
214 }
215 
216 static int xgpu_nv_mailbox_ack_irq(struct amdgpu_device *adev,
217 					struct amdgpu_irq_src *source,
218 					struct amdgpu_iv_entry *entry)
219 {
220 	DRM_DEBUG("get ack intr and do nothing.\n");
221 	return 0;
222 }
223 
224 static int xgpu_nv_set_mailbox_ack_irq(struct amdgpu_device *adev,
225 					struct amdgpu_irq_src *source,
226 					unsigned type,
227 					enum amdgpu_interrupt_state state)
228 {
229 	u32 tmp = RREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF_MAILBOX_INT_CNTL));
230 
231 	tmp = REG_SET_FIELD(tmp, BIF_BX_PF_MAILBOX_INT_CNTL, ACK_INT_EN,
232 				(state == AMDGPU_IRQ_STATE_ENABLE) ? 1 : 0);
233 	WREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF_MAILBOX_INT_CNTL), tmp);
234 
235 	return 0;
236 }
237 
238 static void xgpu_nv_mailbox_flr_work(struct work_struct *work)
239 {
240 	struct amdgpu_virt *virt = container_of(work, struct amdgpu_virt, flr_work);
241 	struct amdgpu_device *adev = container_of(virt, struct amdgpu_device, virt);
242 	int timeout = NV_MAILBOX_POLL_FLR_TIMEDOUT;
243 	int locked;
244 
245 	/* block amdgpu_gpu_recover till msg FLR COMPLETE received,
246 	 * otherwise the mailbox msg will be ruined/reseted by
247 	 * the VF FLR.
248 	 *
249 	 * we can unlock the lock_reset to allow "amdgpu_job_timedout"
250 	 * to run gpu_recover() after FLR_NOTIFICATION_CMPL received
251 	 * which means host side had finished this VF's FLR.
252 	 */
253 	locked = mutex_trylock(&adev->lock_reset);
254 	if (locked)
255 		adev->in_gpu_reset = true;
256 
257 	do {
258 		if (xgpu_nv_mailbox_peek_msg(adev) == IDH_FLR_NOTIFICATION_CMPL)
259 			goto flr_done;
260 
261 		msleep(10);
262 		timeout -= 10;
263 	} while (timeout > 1);
264 
265 flr_done:
266 	if (locked) {
267 		adev->in_gpu_reset = false;
268 		mutex_unlock(&adev->lock_reset);
269 	}
270 
271 	/* Trigger recovery for world switch failure if no TDR */
272 	if (amdgpu_device_should_recover_gpu(adev)
273 		&& (adev->sdma_timeout == MAX_SCHEDULE_TIMEOUT ||
274 		adev->gfx_timeout == MAX_SCHEDULE_TIMEOUT ||
275 		adev->compute_timeout == MAX_SCHEDULE_TIMEOUT ||
276 		adev->video_timeout == MAX_SCHEDULE_TIMEOUT))
277 		amdgpu_device_gpu_recover(adev, NULL);
278 }
279 
280 static int xgpu_nv_set_mailbox_rcv_irq(struct amdgpu_device *adev,
281 				       struct amdgpu_irq_src *src,
282 				       unsigned type,
283 				       enum amdgpu_interrupt_state state)
284 {
285 	u32 tmp = RREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF_MAILBOX_INT_CNTL));
286 
287 	tmp = REG_SET_FIELD(tmp, BIF_BX_PF_MAILBOX_INT_CNTL, VALID_INT_EN,
288 			    (state == AMDGPU_IRQ_STATE_ENABLE) ? 1 : 0);
289 	WREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF_MAILBOX_INT_CNTL), tmp);
290 
291 	return 0;
292 }
293 
294 static int xgpu_nv_mailbox_rcv_irq(struct amdgpu_device *adev,
295 				   struct amdgpu_irq_src *source,
296 				   struct amdgpu_iv_entry *entry)
297 {
298 	enum idh_event event = xgpu_nv_mailbox_peek_msg(adev);
299 
300 	switch (event) {
301 	case IDH_FLR_NOTIFICATION:
302 		if (amdgpu_sriov_runtime(adev))
303 			schedule_work(&adev->virt.flr_work);
304 		break;
305 		/* READY_TO_ACCESS_GPU is fetched by kernel polling, IRQ can ignore
306 		 * it byfar since that polling thread will handle it,
307 		 * other msg like flr complete is not handled here.
308 		 */
309 	case IDH_CLR_MSG_BUF:
310 	case IDH_FLR_NOTIFICATION_CMPL:
311 	case IDH_READY_TO_ACCESS_GPU:
312 	default:
313 		break;
314 	}
315 
316 	return 0;
317 }
318 
319 static const struct amdgpu_irq_src_funcs xgpu_nv_mailbox_ack_irq_funcs = {
320 	.set = xgpu_nv_set_mailbox_ack_irq,
321 	.process = xgpu_nv_mailbox_ack_irq,
322 };
323 
324 static const struct amdgpu_irq_src_funcs xgpu_nv_mailbox_rcv_irq_funcs = {
325 	.set = xgpu_nv_set_mailbox_rcv_irq,
326 	.process = xgpu_nv_mailbox_rcv_irq,
327 };
328 
329 void xgpu_nv_mailbox_set_irq_funcs(struct amdgpu_device *adev)
330 {
331 	adev->virt.ack_irq.num_types = 1;
332 	adev->virt.ack_irq.funcs = &xgpu_nv_mailbox_ack_irq_funcs;
333 	adev->virt.rcv_irq.num_types = 1;
334 	adev->virt.rcv_irq.funcs = &xgpu_nv_mailbox_rcv_irq_funcs;
335 }
336 
337 int xgpu_nv_mailbox_add_irq_id(struct amdgpu_device *adev)
338 {
339 	int r;
340 
341 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_BIF, 135, &adev->virt.rcv_irq);
342 	if (r)
343 		return r;
344 
345 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_BIF, 138, &adev->virt.ack_irq);
346 	if (r) {
347 		amdgpu_irq_put(adev, &adev->virt.rcv_irq, 0);
348 		return r;
349 	}
350 
351 	return 0;
352 }
353 
354 int xgpu_nv_mailbox_get_irq(struct amdgpu_device *adev)
355 {
356 	int r;
357 
358 	r = amdgpu_irq_get(adev, &adev->virt.rcv_irq, 0);
359 	if (r)
360 		return r;
361 	r = amdgpu_irq_get(adev, &adev->virt.ack_irq, 0);
362 	if (r) {
363 		amdgpu_irq_put(adev, &adev->virt.rcv_irq, 0);
364 		return r;
365 	}
366 
367 	INIT_WORK(&adev->virt.flr_work, xgpu_nv_mailbox_flr_work);
368 
369 	return 0;
370 }
371 
372 void xgpu_nv_mailbox_put_irq(struct amdgpu_device *adev)
373 {
374 	amdgpu_irq_put(adev, &adev->virt.ack_irq, 0);
375 	amdgpu_irq_put(adev, &adev->virt.rcv_irq, 0);
376 }
377 
378 const struct amdgpu_virt_ops xgpu_nv_virt_ops = {
379 	.req_full_gpu	= xgpu_nv_request_full_gpu_access,
380 	.rel_full_gpu	= xgpu_nv_release_full_gpu_access,
381 	.reset_gpu = xgpu_nv_request_reset,
382 	.wait_reset = NULL,
383 	.trans_msg = xgpu_nv_mailbox_trans_msg,
384 };
385