xref: /dragonfly/sys/dev/drm/amd/amdgpu/mxgpu_ai.c (revision b843c749)
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 "soc15.h"
30 #include "vega10_ih.h"
31 #include "soc15_common.h"
32 #include "mxgpu_ai.h"
33 
xgpu_ai_mailbox_send_ack(struct amdgpu_device * adev)34 static void xgpu_ai_mailbox_send_ack(struct amdgpu_device *adev)
35 {
36 	WREG8(AI_MAIBOX_CONTROL_RCV_OFFSET_BYTE, 2);
37 }
38 
xgpu_ai_mailbox_set_valid(struct amdgpu_device * adev,bool val)39 static void xgpu_ai_mailbox_set_valid(struct amdgpu_device *adev, bool val)
40 {
41 	WREG8(AI_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_PF0_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  */
xgpu_ai_mailbox_peek_msg(struct amdgpu_device * adev)53 static enum idh_event xgpu_ai_mailbox_peek_msg(struct amdgpu_device *adev)
54 {
55 	return RREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0,
56 				mmBIF_BX_PF0_MAILBOX_MSGBUF_RCV_DW0));
57 }
58 
59 
xgpu_ai_mailbox_rcv_msg(struct amdgpu_device * adev,enum idh_event event)60 static int xgpu_ai_mailbox_rcv_msg(struct amdgpu_device *adev,
61 				   enum idh_event event)
62 {
63 	u32 reg;
64 
65 	reg = RREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0,
66 					     mmBIF_BX_PF0_MAILBOX_MSGBUF_RCV_DW0));
67 	if (reg != event)
68 		return -ENOENT;
69 
70 	xgpu_ai_mailbox_send_ack(adev);
71 
72 	return 0;
73 }
74 
xgpu_ai_peek_ack(struct amdgpu_device * adev)75 static uint8_t xgpu_ai_peek_ack(struct amdgpu_device *adev) {
76 	return RREG8(AI_MAIBOX_CONTROL_TRN_OFFSET_BYTE) & 2;
77 }
78 
xgpu_ai_poll_ack(struct amdgpu_device * adev)79 static int xgpu_ai_poll_ack(struct amdgpu_device *adev)
80 {
81 	int timeout  = AI_MAILBOX_POLL_ACK_TIMEDOUT;
82 	u8 reg;
83 
84 	do {
85 		reg = RREG8(AI_MAIBOX_CONTROL_TRN_OFFSET_BYTE);
86 		if (reg & 2)
87 			return 0;
88 
89 		mdelay(5);
90 		timeout -= 5;
91 	} while (timeout > 1);
92 
93 	pr_err("Doesn't get TRN_MSG_ACK from pf in %d msec\n", AI_MAILBOX_POLL_ACK_TIMEDOUT);
94 
95 	return -ETIME;
96 }
97 
xgpu_ai_poll_msg(struct amdgpu_device * adev,enum idh_event event)98 static int xgpu_ai_poll_msg(struct amdgpu_device *adev, enum idh_event event)
99 {
100 	int r, timeout = AI_MAILBOX_POLL_MSG_TIMEDOUT;
101 
102 	do {
103 		r = xgpu_ai_mailbox_rcv_msg(adev, event);
104 		if (!r)
105 			return 0;
106 
107 		msleep(10);
108 		timeout -= 10;
109 	} while (timeout > 1);
110 
111 	pr_err("Doesn't get msg:%d from pf, error=%d\n", event, r);
112 
113 	return -ETIME;
114 }
115 
xgpu_ai_mailbox_trans_msg(struct amdgpu_device * adev,enum idh_request req,u32 data1,u32 data2,u32 data3)116 static void xgpu_ai_mailbox_trans_msg (struct amdgpu_device *adev,
117 	      enum idh_request req, u32 data1, u32 data2, u32 data3) {
118 	u32 reg;
119 	int r;
120 	uint8_t trn;
121 
122 	/* IMPORTANT:
123 	 * clear TRN_MSG_VALID valid to clear host's RCV_MSG_ACK
124 	 * and with host's RCV_MSG_ACK cleared hw automatically clear host's RCV_MSG_ACK
125 	 * which lead to VF's TRN_MSG_ACK cleared, otherwise below xgpu_ai_poll_ack()
126 	 * will return immediatly
127 	 */
128 	do {
129 		xgpu_ai_mailbox_set_valid(adev, false);
130 		trn = xgpu_ai_peek_ack(adev);
131 		if (trn) {
132 			pr_err("trn=%x ACK should not assert! wait again !\n", trn);
133 			msleep(1);
134 		}
135 	} while(trn);
136 
137 	reg = RREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0,
138 					     mmBIF_BX_PF0_MAILBOX_MSGBUF_TRN_DW0));
139 	reg = REG_SET_FIELD(reg, BIF_BX_PF0_MAILBOX_MSGBUF_TRN_DW0,
140 			    MSGBUF_DATA, req);
141 	WREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF0_MAILBOX_MSGBUF_TRN_DW0),
142 		      reg);
143 	WREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF0_MAILBOX_MSGBUF_TRN_DW1),
144 				data1);
145 	WREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF0_MAILBOX_MSGBUF_TRN_DW2),
146 				data2);
147 	WREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF0_MAILBOX_MSGBUF_TRN_DW3),
148 				data3);
149 
150 	xgpu_ai_mailbox_set_valid(adev, true);
151 
152 	/* start to poll ack */
153 	r = xgpu_ai_poll_ack(adev);
154 	if (r)
155 		pr_err("Doesn't get ack from pf, continue\n");
156 
157 	xgpu_ai_mailbox_set_valid(adev, false);
158 }
159 
xgpu_ai_send_access_requests(struct amdgpu_device * adev,enum idh_request req)160 static int xgpu_ai_send_access_requests(struct amdgpu_device *adev,
161 					enum idh_request req)
162 {
163 	int r;
164 
165 	xgpu_ai_mailbox_trans_msg(adev, req, 0, 0, 0);
166 
167 	/* start to check msg if request is idh_req_gpu_init_access */
168 	if (req == IDH_REQ_GPU_INIT_ACCESS ||
169 		req == IDH_REQ_GPU_FINI_ACCESS ||
170 		req == IDH_REQ_GPU_RESET_ACCESS) {
171 		r = xgpu_ai_poll_msg(adev, IDH_READY_TO_ACCESS_GPU);
172 		if (r) {
173 			pr_err("Doesn't get READY_TO_ACCESS_GPU from pf, give up\n");
174 			return r;
175 		}
176 		/* Retrieve checksum from mailbox2 */
177 		if (req == IDH_REQ_GPU_INIT_ACCESS || req == IDH_REQ_GPU_RESET_ACCESS) {
178 			adev->virt.fw_reserve.checksum_key =
179 				RREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0,
180 					mmBIF_BX_PF0_MAILBOX_MSGBUF_RCV_DW2));
181 		}
182 	}
183 
184 	return 0;
185 }
186 
xgpu_ai_request_reset(struct amdgpu_device * adev)187 static int xgpu_ai_request_reset(struct amdgpu_device *adev)
188 {
189 	return xgpu_ai_send_access_requests(adev, IDH_REQ_GPU_RESET_ACCESS);
190 }
191 
xgpu_ai_request_full_gpu_access(struct amdgpu_device * adev,bool init)192 static int xgpu_ai_request_full_gpu_access(struct amdgpu_device *adev,
193 					   bool init)
194 {
195 	enum idh_request req;
196 
197 	req = init ? IDH_REQ_GPU_INIT_ACCESS : IDH_REQ_GPU_FINI_ACCESS;
198 	return xgpu_ai_send_access_requests(adev, req);
199 }
200 
xgpu_ai_release_full_gpu_access(struct amdgpu_device * adev,bool init)201 static int xgpu_ai_release_full_gpu_access(struct amdgpu_device *adev,
202 					   bool init)
203 {
204 	enum idh_request req;
205 	int r = 0;
206 
207 	req = init ? IDH_REL_GPU_INIT_ACCESS : IDH_REL_GPU_FINI_ACCESS;
208 	r = xgpu_ai_send_access_requests(adev, req);
209 
210 	return r;
211 }
212 
xgpu_ai_mailbox_ack_irq(struct amdgpu_device * adev,struct amdgpu_irq_src * source,struct amdgpu_iv_entry * entry)213 static int xgpu_ai_mailbox_ack_irq(struct amdgpu_device *adev,
214 					struct amdgpu_irq_src *source,
215 					struct amdgpu_iv_entry *entry)
216 {
217 	DRM_DEBUG("get ack intr and do nothing.\n");
218 	return 0;
219 }
220 
xgpu_ai_set_mailbox_ack_irq(struct amdgpu_device * adev,struct amdgpu_irq_src * source,unsigned type,enum amdgpu_interrupt_state state)221 static int xgpu_ai_set_mailbox_ack_irq(struct amdgpu_device *adev,
222 					struct amdgpu_irq_src *source,
223 					unsigned type,
224 					enum amdgpu_interrupt_state state)
225 {
226 	u32 tmp = RREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF0_MAILBOX_INT_CNTL));
227 
228 	tmp = REG_SET_FIELD(tmp, BIF_BX_PF0_MAILBOX_INT_CNTL, ACK_INT_EN,
229 				(state == AMDGPU_IRQ_STATE_ENABLE) ? 1 : 0);
230 	WREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF0_MAILBOX_INT_CNTL), tmp);
231 
232 	return 0;
233 }
234 
xgpu_ai_mailbox_flr_work(struct work_struct * work)235 static void xgpu_ai_mailbox_flr_work(struct work_struct *work)
236 {
237 	struct amdgpu_virt *virt = container_of(work, struct amdgpu_virt, flr_work);
238 	struct amdgpu_device *adev = container_of(virt, struct amdgpu_device, virt);
239 	int timeout = AI_MAILBOX_POLL_FLR_TIMEDOUT;
240 	int locked;
241 
242 	/* block amdgpu_gpu_recover till msg FLR COMPLETE received,
243 	 * otherwise the mailbox msg will be ruined/reseted by
244 	 * the VF FLR.
245 	 *
246 	 * we can unlock the lock_reset to allow "amdgpu_job_timedout"
247 	 * to run gpu_recover() after FLR_NOTIFICATION_CMPL received
248 	 * which means host side had finished this VF's FLR.
249 	 */
250 	locked = mutex_trylock(&adev->lock_reset);
251 	if (locked)
252 		adev->in_gpu_reset = 1;
253 
254 	do {
255 		if (xgpu_ai_mailbox_peek_msg(adev) == IDH_FLR_NOTIFICATION_CMPL)
256 			goto flr_done;
257 
258 		msleep(10);
259 		timeout -= 10;
260 	} while (timeout > 1);
261 
262 flr_done:
263 	if (locked) {
264 		adev->in_gpu_reset = 0;
265 		mutex_unlock(&adev->lock_reset);
266 	}
267 
268 	/* Trigger recovery for world switch failure if no TDR */
269 	if (amdgpu_lockup_timeout == 0)
270 		amdgpu_device_gpu_recover(adev, NULL, true);
271 }
272 
xgpu_ai_set_mailbox_rcv_irq(struct amdgpu_device * adev,struct amdgpu_irq_src * src,unsigned type,enum amdgpu_interrupt_state state)273 static int xgpu_ai_set_mailbox_rcv_irq(struct amdgpu_device *adev,
274 				       struct amdgpu_irq_src *src,
275 				       unsigned type,
276 				       enum amdgpu_interrupt_state state)
277 {
278 	u32 tmp = RREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF0_MAILBOX_INT_CNTL));
279 
280 	tmp = REG_SET_FIELD(tmp, BIF_BX_PF0_MAILBOX_INT_CNTL, VALID_INT_EN,
281 			    (state == AMDGPU_IRQ_STATE_ENABLE) ? 1 : 0);
282 	WREG32_NO_KIQ(SOC15_REG_OFFSET(NBIO, 0, mmBIF_BX_PF0_MAILBOX_INT_CNTL), tmp);
283 
284 	return 0;
285 }
286 
xgpu_ai_mailbox_rcv_irq(struct amdgpu_device * adev,struct amdgpu_irq_src * source,struct amdgpu_iv_entry * entry)287 static int xgpu_ai_mailbox_rcv_irq(struct amdgpu_device *adev,
288 				   struct amdgpu_irq_src *source,
289 				   struct amdgpu_iv_entry *entry)
290 {
291 	enum idh_event event = xgpu_ai_mailbox_peek_msg(adev);
292 
293 	switch (event) {
294 		case IDH_FLR_NOTIFICATION:
295 		if (amdgpu_sriov_runtime(adev))
296 			schedule_work(&adev->virt.flr_work);
297 		break;
298 		/* READY_TO_ACCESS_GPU is fetched by kernel polling, IRQ can ignore
299 		 * it byfar since that polling thread will handle it,
300 		 * other msg like flr complete is not handled here.
301 		 */
302 		case IDH_CLR_MSG_BUF:
303 		case IDH_FLR_NOTIFICATION_CMPL:
304 		case IDH_READY_TO_ACCESS_GPU:
305 		default:
306 		break;
307 	}
308 
309 	return 0;
310 }
311 
312 static const struct amdgpu_irq_src_funcs xgpu_ai_mailbox_ack_irq_funcs = {
313 	.set = xgpu_ai_set_mailbox_ack_irq,
314 	.process = xgpu_ai_mailbox_ack_irq,
315 };
316 
317 static const struct amdgpu_irq_src_funcs xgpu_ai_mailbox_rcv_irq_funcs = {
318 	.set = xgpu_ai_set_mailbox_rcv_irq,
319 	.process = xgpu_ai_mailbox_rcv_irq,
320 };
321 
xgpu_ai_mailbox_set_irq_funcs(struct amdgpu_device * adev)322 void xgpu_ai_mailbox_set_irq_funcs(struct amdgpu_device *adev)
323 {
324 	adev->virt.ack_irq.num_types = 1;
325 	adev->virt.ack_irq.funcs = &xgpu_ai_mailbox_ack_irq_funcs;
326 	adev->virt.rcv_irq.num_types = 1;
327 	adev->virt.rcv_irq.funcs = &xgpu_ai_mailbox_rcv_irq_funcs;
328 }
329 
xgpu_ai_mailbox_add_irq_id(struct amdgpu_device * adev)330 int xgpu_ai_mailbox_add_irq_id(struct amdgpu_device *adev)
331 {
332 	int r;
333 
334 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_BIF, 135, &adev->virt.rcv_irq);
335 	if (r)
336 		return r;
337 
338 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_BIF, 138, &adev->virt.ack_irq);
339 	if (r) {
340 		amdgpu_irq_put(adev, &adev->virt.rcv_irq, 0);
341 		return r;
342 	}
343 
344 	return 0;
345 }
346 
xgpu_ai_mailbox_get_irq(struct amdgpu_device * adev)347 int xgpu_ai_mailbox_get_irq(struct amdgpu_device *adev)
348 {
349 	int r;
350 
351 	r = amdgpu_irq_get(adev, &adev->virt.rcv_irq, 0);
352 	if (r)
353 		return r;
354 	r = amdgpu_irq_get(adev, &adev->virt.ack_irq, 0);
355 	if (r) {
356 		amdgpu_irq_put(adev, &adev->virt.rcv_irq, 0);
357 		return r;
358 	}
359 
360 	INIT_WORK(&adev->virt.flr_work, xgpu_ai_mailbox_flr_work);
361 
362 	return 0;
363 }
364 
xgpu_ai_mailbox_put_irq(struct amdgpu_device * adev)365 void xgpu_ai_mailbox_put_irq(struct amdgpu_device *adev)
366 {
367 	amdgpu_irq_put(adev, &adev->virt.ack_irq, 0);
368 	amdgpu_irq_put(adev, &adev->virt.rcv_irq, 0);
369 }
370 
371 const struct amdgpu_virt_ops xgpu_ai_virt_ops = {
372 	.req_full_gpu	= xgpu_ai_request_full_gpu_access,
373 	.rel_full_gpu	= xgpu_ai_release_full_gpu_access,
374 	.reset_gpu = xgpu_ai_request_reset,
375 	.wait_reset = NULL,
376 	.trans_msg = xgpu_ai_mailbox_trans_msg,
377 };
378