Loading...
1/*
2 * Copyright 2016 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 <linux/module.h>
25
26#include <drm/drm_drv.h>
27
28#include "amdgpu.h"
29#include "amdgpu_ras.h"
30#include "vi.h"
31#include "soc15.h"
32#include "nv.h"
33
34bool amdgpu_virt_mmio_blocked(struct amdgpu_device *adev)
35{
36 /* By now all MMIO pages except mailbox are blocked */
37 /* if blocking is enabled in hypervisor. Choose the */
38 /* SCRATCH_REG0 to test. */
39 return RREG32_NO_KIQ(0xc040) == 0xffffffff;
40}
41
42void amdgpu_virt_init_setting(struct amdgpu_device *adev)
43{
44 /* enable virtual display */
45 if (adev->mode_info.num_crtc == 0)
46 adev->mode_info.num_crtc = 1;
47 adev->enable_virtual_display = true;
48 adev->ddev->driver->driver_features &= ~DRIVER_ATOMIC;
49 adev->cg_flags = 0;
50 adev->pg_flags = 0;
51}
52
53void amdgpu_virt_kiq_reg_write_reg_wait(struct amdgpu_device *adev,
54 uint32_t reg0, uint32_t reg1,
55 uint32_t ref, uint32_t mask)
56{
57 struct amdgpu_kiq *kiq = &adev->gfx.kiq;
58 struct amdgpu_ring *ring = &kiq->ring;
59 signed long r, cnt = 0;
60 unsigned long flags;
61 uint32_t seq;
62
63 spin_lock_irqsave(&kiq->ring_lock, flags);
64 amdgpu_ring_alloc(ring, 32);
65 amdgpu_ring_emit_reg_write_reg_wait(ring, reg0, reg1,
66 ref, mask);
67 r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
68 if (r)
69 goto failed_undo;
70
71 amdgpu_ring_commit(ring);
72 spin_unlock_irqrestore(&kiq->ring_lock, flags);
73
74 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
75
76 /* don't wait anymore for IRQ context */
77 if (r < 1 && in_interrupt())
78 goto failed_kiq;
79
80 might_sleep();
81 while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
82
83 msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
84 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
85 }
86
87 if (cnt > MAX_KIQ_REG_TRY)
88 goto failed_kiq;
89
90 return;
91
92failed_undo:
93 amdgpu_ring_undo(ring);
94 spin_unlock_irqrestore(&kiq->ring_lock, flags);
95failed_kiq:
96 pr_err("failed to write reg %x wait reg %x\n", reg0, reg1);
97}
98
99/**
100 * amdgpu_virt_request_full_gpu() - request full gpu access
101 * @amdgpu: amdgpu device.
102 * @init: is driver init time.
103 * When start to init/fini driver, first need to request full gpu access.
104 * Return: Zero if request success, otherwise will return error.
105 */
106int amdgpu_virt_request_full_gpu(struct amdgpu_device *adev, bool init)
107{
108 struct amdgpu_virt *virt = &adev->virt;
109 int r;
110
111 if (virt->ops && virt->ops->req_full_gpu) {
112 r = virt->ops->req_full_gpu(adev, init);
113 if (r)
114 return r;
115
116 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
117 }
118
119 return 0;
120}
121
122/**
123 * amdgpu_virt_release_full_gpu() - release full gpu access
124 * @amdgpu: amdgpu device.
125 * @init: is driver init time.
126 * When finishing driver init/fini, need to release full gpu access.
127 * Return: Zero if release success, otherwise will returen error.
128 */
129int amdgpu_virt_release_full_gpu(struct amdgpu_device *adev, bool init)
130{
131 struct amdgpu_virt *virt = &adev->virt;
132 int r;
133
134 if (virt->ops && virt->ops->rel_full_gpu) {
135 r = virt->ops->rel_full_gpu(adev, init);
136 if (r)
137 return r;
138
139 adev->virt.caps |= AMDGPU_SRIOV_CAPS_RUNTIME;
140 }
141 return 0;
142}
143
144/**
145 * amdgpu_virt_reset_gpu() - reset gpu
146 * @amdgpu: amdgpu device.
147 * Send reset command to GPU hypervisor to reset GPU that VM is using
148 * Return: Zero if reset success, otherwise will return error.
149 */
150int amdgpu_virt_reset_gpu(struct amdgpu_device *adev)
151{
152 struct amdgpu_virt *virt = &adev->virt;
153 int r;
154
155 if (virt->ops && virt->ops->reset_gpu) {
156 r = virt->ops->reset_gpu(adev);
157 if (r)
158 return r;
159
160 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
161 }
162
163 return 0;
164}
165
166void amdgpu_virt_request_init_data(struct amdgpu_device *adev)
167{
168 struct amdgpu_virt *virt = &adev->virt;
169
170 if (virt->ops && virt->ops->req_init_data)
171 virt->ops->req_init_data(adev);
172
173 if (adev->virt.req_init_data_ver > 0)
174 DRM_INFO("host supports REQ_INIT_DATA handshake\n");
175 else
176 DRM_WARN("host doesn't support REQ_INIT_DATA handshake\n");
177}
178
179/**
180 * amdgpu_virt_wait_reset() - wait for reset gpu completed
181 * @amdgpu: amdgpu device.
182 * Wait for GPU reset completed.
183 * Return: Zero if reset success, otherwise will return error.
184 */
185int amdgpu_virt_wait_reset(struct amdgpu_device *adev)
186{
187 struct amdgpu_virt *virt = &adev->virt;
188
189 if (!virt->ops || !virt->ops->wait_reset)
190 return -EINVAL;
191
192 return virt->ops->wait_reset(adev);
193}
194
195/**
196 * amdgpu_virt_alloc_mm_table() - alloc memory for mm table
197 * @amdgpu: amdgpu device.
198 * MM table is used by UVD and VCE for its initialization
199 * Return: Zero if allocate success.
200 */
201int amdgpu_virt_alloc_mm_table(struct amdgpu_device *adev)
202{
203 int r;
204
205 if (!amdgpu_sriov_vf(adev) || adev->virt.mm_table.gpu_addr)
206 return 0;
207
208 r = amdgpu_bo_create_kernel(adev, PAGE_SIZE, PAGE_SIZE,
209 AMDGPU_GEM_DOMAIN_VRAM,
210 &adev->virt.mm_table.bo,
211 &adev->virt.mm_table.gpu_addr,
212 (void *)&adev->virt.mm_table.cpu_addr);
213 if (r) {
214 DRM_ERROR("failed to alloc mm table and error = %d.\n", r);
215 return r;
216 }
217
218 memset((void *)adev->virt.mm_table.cpu_addr, 0, PAGE_SIZE);
219 DRM_INFO("MM table gpu addr = 0x%llx, cpu addr = %p.\n",
220 adev->virt.mm_table.gpu_addr,
221 adev->virt.mm_table.cpu_addr);
222 return 0;
223}
224
225/**
226 * amdgpu_virt_free_mm_table() - free mm table memory
227 * @amdgpu: amdgpu device.
228 * Free MM table memory
229 */
230void amdgpu_virt_free_mm_table(struct amdgpu_device *adev)
231{
232 if (!amdgpu_sriov_vf(adev) || !adev->virt.mm_table.gpu_addr)
233 return;
234
235 amdgpu_bo_free_kernel(&adev->virt.mm_table.bo,
236 &adev->virt.mm_table.gpu_addr,
237 (void *)&adev->virt.mm_table.cpu_addr);
238 adev->virt.mm_table.gpu_addr = 0;
239}
240
241
242int amdgpu_virt_fw_reserve_get_checksum(void *obj,
243 unsigned long obj_size,
244 unsigned int key,
245 unsigned int chksum)
246{
247 unsigned int ret = key;
248 unsigned long i = 0;
249 unsigned char *pos;
250
251 pos = (char *)obj;
252 /* calculate checksum */
253 for (i = 0; i < obj_size; ++i)
254 ret += *(pos + i);
255 /* minus the chksum itself */
256 pos = (char *)&chksum;
257 for (i = 0; i < sizeof(chksum); ++i)
258 ret -= *(pos + i);
259 return ret;
260}
261
262static int amdgpu_virt_init_ras_err_handler_data(struct amdgpu_device *adev)
263{
264 struct amdgpu_virt *virt = &adev->virt;
265 struct amdgpu_virt_ras_err_handler_data **data = &virt->virt_eh_data;
266 /* GPU will be marked bad on host if bp count more then 10,
267 * so alloc 512 is enough.
268 */
269 unsigned int align_space = 512;
270 void *bps = NULL;
271 struct amdgpu_bo **bps_bo = NULL;
272
273 *data = kmalloc(sizeof(struct amdgpu_virt_ras_err_handler_data), GFP_KERNEL);
274 if (!*data)
275 return -ENOMEM;
276
277 bps = kmalloc(align_space * sizeof((*data)->bps), GFP_KERNEL);
278 bps_bo = kmalloc(align_space * sizeof((*data)->bps_bo), GFP_KERNEL);
279
280 if (!bps || !bps_bo) {
281 kfree(bps);
282 kfree(bps_bo);
283 kfree(*data);
284 return -ENOMEM;
285 }
286
287 (*data)->bps = bps;
288 (*data)->bps_bo = bps_bo;
289 (*data)->count = 0;
290 (*data)->last_reserved = 0;
291
292 virt->ras_init_done = true;
293
294 return 0;
295}
296
297static void amdgpu_virt_ras_release_bp(struct amdgpu_device *adev)
298{
299 struct amdgpu_virt *virt = &adev->virt;
300 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data;
301 struct amdgpu_bo *bo;
302 int i;
303
304 if (!data)
305 return;
306
307 for (i = data->last_reserved - 1; i >= 0; i--) {
308 bo = data->bps_bo[i];
309 amdgpu_bo_free_kernel(&bo, NULL, NULL);
310 data->bps_bo[i] = bo;
311 data->last_reserved = i;
312 }
313}
314
315void amdgpu_virt_release_ras_err_handler_data(struct amdgpu_device *adev)
316{
317 struct amdgpu_virt *virt = &adev->virt;
318 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data;
319
320 virt->ras_init_done = false;
321
322 if (!data)
323 return;
324
325 amdgpu_virt_ras_release_bp(adev);
326
327 kfree(data->bps);
328 kfree(data->bps_bo);
329 kfree(data);
330 virt->virt_eh_data = NULL;
331}
332
333static void amdgpu_virt_ras_add_bps(struct amdgpu_device *adev,
334 struct eeprom_table_record *bps, int pages)
335{
336 struct amdgpu_virt *virt = &adev->virt;
337 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data;
338
339 if (!data)
340 return;
341
342 memcpy(&data->bps[data->count], bps, pages * sizeof(*data->bps));
343 data->count += pages;
344}
345
346static void amdgpu_virt_ras_reserve_bps(struct amdgpu_device *adev)
347{
348 struct amdgpu_virt *virt = &adev->virt;
349 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data;
350 struct amdgpu_bo *bo = NULL;
351 uint64_t bp;
352 int i;
353
354 if (!data)
355 return;
356
357 for (i = data->last_reserved; i < data->count; i++) {
358 bp = data->bps[i].retired_page;
359
360 /* There are two cases of reserve error should be ignored:
361 * 1) a ras bad page has been allocated (used by someone);
362 * 2) a ras bad page has been reserved (duplicate error injection
363 * for one page);
364 */
365 if (amdgpu_bo_create_kernel_at(adev, bp << AMDGPU_GPU_PAGE_SHIFT,
366 AMDGPU_GPU_PAGE_SIZE,
367 AMDGPU_GEM_DOMAIN_VRAM,
368 &bo, NULL))
369 DRM_DEBUG("RAS WARN: reserve vram for retired page %llx fail\n", bp);
370
371 data->bps_bo[i] = bo;
372 data->last_reserved = i + 1;
373 bo = NULL;
374 }
375}
376
377static bool amdgpu_virt_ras_check_bad_page(struct amdgpu_device *adev,
378 uint64_t retired_page)
379{
380 struct amdgpu_virt *virt = &adev->virt;
381 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data;
382 int i;
383
384 if (!data)
385 return true;
386
387 for (i = 0; i < data->count; i++)
388 if (retired_page == data->bps[i].retired_page)
389 return true;
390
391 return false;
392}
393
394static void amdgpu_virt_add_bad_page(struct amdgpu_device *adev,
395 uint64_t bp_block_offset, uint32_t bp_block_size)
396{
397 struct eeprom_table_record bp;
398 uint64_t retired_page;
399 uint32_t bp_idx, bp_cnt;
400
401 if (bp_block_size) {
402 bp_cnt = bp_block_size / sizeof(uint64_t);
403 for (bp_idx = 0; bp_idx < bp_cnt; bp_idx++) {
404 retired_page = *(uint64_t *)(adev->fw_vram_usage.va +
405 bp_block_offset + bp_idx * sizeof(uint64_t));
406 bp.retired_page = retired_page;
407
408 if (amdgpu_virt_ras_check_bad_page(adev, retired_page))
409 continue;
410
411 amdgpu_virt_ras_add_bps(adev, &bp, 1);
412
413 amdgpu_virt_ras_reserve_bps(adev);
414 }
415 }
416}
417
418void amdgpu_virt_init_data_exchange(struct amdgpu_device *adev)
419{
420 uint32_t pf2vf_size = 0;
421 uint32_t checksum = 0;
422 uint32_t checkval;
423 char *str;
424 uint64_t bp_block_offset = 0;
425 uint32_t bp_block_size = 0;
426 struct amdgim_pf2vf_info_v2 *pf2vf_v2 = NULL;
427
428 adev->virt.fw_reserve.p_pf2vf = NULL;
429 adev->virt.fw_reserve.p_vf2pf = NULL;
430
431 if (adev->fw_vram_usage.va != NULL) {
432 adev->virt.fw_reserve.p_pf2vf =
433 (struct amd_sriov_msg_pf2vf_info_header *)(
434 adev->fw_vram_usage.va + AMDGIM_DATAEXCHANGE_OFFSET);
435 AMDGPU_FW_VRAM_PF2VF_READ(adev, header.size, &pf2vf_size);
436 AMDGPU_FW_VRAM_PF2VF_READ(adev, checksum, &checksum);
437 AMDGPU_FW_VRAM_PF2VF_READ(adev, feature_flags, &adev->virt.gim_feature);
438
439 /* pf2vf message must be in 4K */
440 if (pf2vf_size > 0 && pf2vf_size < 4096) {
441 if (adev->virt.fw_reserve.p_pf2vf->version == 2) {
442 pf2vf_v2 = (struct amdgim_pf2vf_info_v2 *)adev->virt.fw_reserve.p_pf2vf;
443 bp_block_offset = ((uint64_t)pf2vf_v2->bp_block_offset_L & 0xFFFFFFFF) |
444 ((((uint64_t)pf2vf_v2->bp_block_offset_H) << 32) & 0xFFFFFFFF00000000);
445 bp_block_size = pf2vf_v2->bp_block_size;
446
447 if (bp_block_size && !adev->virt.ras_init_done)
448 amdgpu_virt_init_ras_err_handler_data(adev);
449
450 if (adev->virt.ras_init_done)
451 amdgpu_virt_add_bad_page(adev, bp_block_offset, bp_block_size);
452 }
453
454 checkval = amdgpu_virt_fw_reserve_get_checksum(
455 adev->virt.fw_reserve.p_pf2vf, pf2vf_size,
456 adev->virt.fw_reserve.checksum_key, checksum);
457 if (checkval == checksum) {
458 adev->virt.fw_reserve.p_vf2pf =
459 ((void *)adev->virt.fw_reserve.p_pf2vf +
460 pf2vf_size);
461 memset((void *)adev->virt.fw_reserve.p_vf2pf, 0,
462 sizeof(amdgim_vf2pf_info));
463 AMDGPU_FW_VRAM_VF2PF_WRITE(adev, header.version,
464 AMDGPU_FW_VRAM_VF2PF_VER);
465 AMDGPU_FW_VRAM_VF2PF_WRITE(adev, header.size,
466 sizeof(amdgim_vf2pf_info));
467 AMDGPU_FW_VRAM_VF2PF_READ(adev, driver_version,
468 &str);
469#ifdef MODULE
470 if (THIS_MODULE->version != NULL)
471 strcpy(str, THIS_MODULE->version);
472 else
473#endif
474 strcpy(str, "N/A");
475 AMDGPU_FW_VRAM_VF2PF_WRITE(adev, driver_cert,
476 0);
477 AMDGPU_FW_VRAM_VF2PF_WRITE(adev, checksum,
478 amdgpu_virt_fw_reserve_get_checksum(
479 adev->virt.fw_reserve.p_vf2pf,
480 pf2vf_size,
481 adev->virt.fw_reserve.checksum_key, 0));
482 }
483 }
484 }
485}
486
487void amdgpu_detect_virtualization(struct amdgpu_device *adev)
488{
489 uint32_t reg;
490
491 switch (adev->asic_type) {
492 case CHIP_TONGA:
493 case CHIP_FIJI:
494 reg = RREG32(mmBIF_IOV_FUNC_IDENTIFIER);
495 break;
496 case CHIP_VEGA10:
497 case CHIP_VEGA20:
498 case CHIP_NAVI10:
499 case CHIP_NAVI12:
500 case CHIP_SIENNA_CICHLID:
501 case CHIP_ARCTURUS:
502 reg = RREG32(mmRCC_IOV_FUNC_IDENTIFIER);
503 break;
504 default: /* other chip doesn't support SRIOV */
505 reg = 0;
506 break;
507 }
508
509 if (reg & 1)
510 adev->virt.caps |= AMDGPU_SRIOV_CAPS_IS_VF;
511
512 if (reg & 0x80000000)
513 adev->virt.caps |= AMDGPU_SRIOV_CAPS_ENABLE_IOV;
514
515 if (!reg) {
516 if (is_virtual_machine()) /* passthrough mode exclus sriov mod */
517 adev->virt.caps |= AMDGPU_PASSTHROUGH_MODE;
518 }
519
520 /* we have the ability to check now */
521 if (amdgpu_sriov_vf(adev)) {
522 switch (adev->asic_type) {
523 case CHIP_TONGA:
524 case CHIP_FIJI:
525 vi_set_virt_ops(adev);
526 break;
527 case CHIP_VEGA10:
528 case CHIP_VEGA20:
529 case CHIP_ARCTURUS:
530 soc15_set_virt_ops(adev);
531 break;
532 case CHIP_NAVI10:
533 case CHIP_NAVI12:
534 case CHIP_SIENNA_CICHLID:
535 nv_set_virt_ops(adev);
536 /* try send GPU_INIT_DATA request to host */
537 amdgpu_virt_request_init_data(adev);
538 break;
539 default: /* other chip doesn't support SRIOV */
540 DRM_ERROR("Unknown asic type: %d!\n", adev->asic_type);
541 break;
542 }
543 }
544}
545
546static bool amdgpu_virt_access_debugfs_is_mmio(struct amdgpu_device *adev)
547{
548 return amdgpu_sriov_is_debug(adev) ? true : false;
549}
550
551static bool amdgpu_virt_access_debugfs_is_kiq(struct amdgpu_device *adev)
552{
553 return amdgpu_sriov_is_normal(adev) ? true : false;
554}
555
556int amdgpu_virt_enable_access_debugfs(struct amdgpu_device *adev)
557{
558 if (!amdgpu_sriov_vf(adev) ||
559 amdgpu_virt_access_debugfs_is_kiq(adev))
560 return 0;
561
562 if (amdgpu_virt_access_debugfs_is_mmio(adev))
563 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
564 else
565 return -EPERM;
566
567 return 0;
568}
569
570void amdgpu_virt_disable_access_debugfs(struct amdgpu_device *adev)
571{
572 if (amdgpu_sriov_vf(adev))
573 adev->virt.caps |= AMDGPU_SRIOV_CAPS_RUNTIME;
574}
575
576enum amdgpu_sriov_vf_mode amdgpu_virt_get_sriov_vf_mode(struct amdgpu_device *adev)
577{
578 enum amdgpu_sriov_vf_mode mode;
579
580 if (amdgpu_sriov_vf(adev)) {
581 if (amdgpu_sriov_is_pp_one_vf(adev))
582 mode = SRIOV_VF_MODE_ONE_VF;
583 else
584 mode = SRIOV_VF_MODE_MULTI_VF;
585 } else {
586 mode = SRIOV_VF_MODE_BARE_METAL;
587 }
588
589 return mode;
590}
1/*
2 * Copyright 2016 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 <linux/module.h>
25
26#ifdef CONFIG_X86
27#include <asm/hypervisor.h>
28#endif
29
30#include <drm/drm_drv.h>
31#include <xen/xen.h>
32
33#include "amdgpu.h"
34#include "amdgpu_ras.h"
35#include "vi.h"
36#include "soc15.h"
37#include "nv.h"
38
39#define POPULATE_UCODE_INFO(vf2pf_info, ucode, ver) \
40 do { \
41 vf2pf_info->ucode_info[ucode].id = ucode; \
42 vf2pf_info->ucode_info[ucode].version = ver; \
43 } while (0)
44
45bool amdgpu_virt_mmio_blocked(struct amdgpu_device *adev)
46{
47 /* By now all MMIO pages except mailbox are blocked */
48 /* if blocking is enabled in hypervisor. Choose the */
49 /* SCRATCH_REG0 to test. */
50 return RREG32_NO_KIQ(0xc040) == 0xffffffff;
51}
52
53void amdgpu_virt_init_setting(struct amdgpu_device *adev)
54{
55 struct drm_device *ddev = adev_to_drm(adev);
56
57 /* enable virtual display */
58 if (adev->asic_type != CHIP_ALDEBARAN &&
59 adev->asic_type != CHIP_ARCTURUS) {
60 if (adev->mode_info.num_crtc == 0)
61 adev->mode_info.num_crtc = 1;
62 adev->enable_virtual_display = true;
63 }
64 ddev->driver_features &= ~DRIVER_ATOMIC;
65 adev->cg_flags = 0;
66 adev->pg_flags = 0;
67
68 /* enable mcbp for sriov asic_type before soc21 */
69 amdgpu_mcbp = (adev->asic_type < CHIP_IP_DISCOVERY) ? 1 : 0;
70
71}
72
73void amdgpu_virt_kiq_reg_write_reg_wait(struct amdgpu_device *adev,
74 uint32_t reg0, uint32_t reg1,
75 uint32_t ref, uint32_t mask)
76{
77 struct amdgpu_kiq *kiq = &adev->gfx.kiq;
78 struct amdgpu_ring *ring = &kiq->ring;
79 signed long r, cnt = 0;
80 unsigned long flags;
81 uint32_t seq;
82
83 if (adev->mes.ring.sched.ready) {
84 amdgpu_mes_reg_write_reg_wait(adev, reg0, reg1,
85 ref, mask);
86 return;
87 }
88
89 spin_lock_irqsave(&kiq->ring_lock, flags);
90 amdgpu_ring_alloc(ring, 32);
91 amdgpu_ring_emit_reg_write_reg_wait(ring, reg0, reg1,
92 ref, mask);
93 r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
94 if (r)
95 goto failed_undo;
96
97 amdgpu_ring_commit(ring);
98 spin_unlock_irqrestore(&kiq->ring_lock, flags);
99
100 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
101
102 /* don't wait anymore for IRQ context */
103 if (r < 1 && in_interrupt())
104 goto failed_kiq;
105
106 might_sleep();
107 while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
108
109 msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
110 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
111 }
112
113 if (cnt > MAX_KIQ_REG_TRY)
114 goto failed_kiq;
115
116 return;
117
118failed_undo:
119 amdgpu_ring_undo(ring);
120 spin_unlock_irqrestore(&kiq->ring_lock, flags);
121failed_kiq:
122 dev_err(adev->dev, "failed to write reg %x wait reg %x\n", reg0, reg1);
123}
124
125/**
126 * amdgpu_virt_request_full_gpu() - request full gpu access
127 * @adev: amdgpu device.
128 * @init: is driver init time.
129 * When start to init/fini driver, first need to request full gpu access.
130 * Return: Zero if request success, otherwise will return error.
131 */
132int amdgpu_virt_request_full_gpu(struct amdgpu_device *adev, bool init)
133{
134 struct amdgpu_virt *virt = &adev->virt;
135 int r;
136
137 if (virt->ops && virt->ops->req_full_gpu) {
138 r = virt->ops->req_full_gpu(adev, init);
139 if (r)
140 return r;
141
142 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
143 }
144
145 return 0;
146}
147
148/**
149 * amdgpu_virt_release_full_gpu() - release full gpu access
150 * @adev: amdgpu device.
151 * @init: is driver init time.
152 * When finishing driver init/fini, need to release full gpu access.
153 * Return: Zero if release success, otherwise will returen error.
154 */
155int amdgpu_virt_release_full_gpu(struct amdgpu_device *adev, bool init)
156{
157 struct amdgpu_virt *virt = &adev->virt;
158 int r;
159
160 if (virt->ops && virt->ops->rel_full_gpu) {
161 r = virt->ops->rel_full_gpu(adev, init);
162 if (r)
163 return r;
164
165 adev->virt.caps |= AMDGPU_SRIOV_CAPS_RUNTIME;
166 }
167 return 0;
168}
169
170/**
171 * amdgpu_virt_reset_gpu() - reset gpu
172 * @adev: amdgpu device.
173 * Send reset command to GPU hypervisor to reset GPU that VM is using
174 * Return: Zero if reset success, otherwise will return error.
175 */
176int amdgpu_virt_reset_gpu(struct amdgpu_device *adev)
177{
178 struct amdgpu_virt *virt = &adev->virt;
179 int r;
180
181 if (virt->ops && virt->ops->reset_gpu) {
182 r = virt->ops->reset_gpu(adev);
183 if (r)
184 return r;
185
186 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
187 }
188
189 return 0;
190}
191
192void amdgpu_virt_request_init_data(struct amdgpu_device *adev)
193{
194 struct amdgpu_virt *virt = &adev->virt;
195
196 if (virt->ops && virt->ops->req_init_data)
197 virt->ops->req_init_data(adev);
198
199 if (adev->virt.req_init_data_ver > 0)
200 DRM_INFO("host supports REQ_INIT_DATA handshake\n");
201 else
202 DRM_WARN("host doesn't support REQ_INIT_DATA handshake\n");
203}
204
205/**
206 * amdgpu_virt_wait_reset() - wait for reset gpu completed
207 * @adev: amdgpu device.
208 * Wait for GPU reset completed.
209 * Return: Zero if reset success, otherwise will return error.
210 */
211int amdgpu_virt_wait_reset(struct amdgpu_device *adev)
212{
213 struct amdgpu_virt *virt = &adev->virt;
214
215 if (!virt->ops || !virt->ops->wait_reset)
216 return -EINVAL;
217
218 return virt->ops->wait_reset(adev);
219}
220
221/**
222 * amdgpu_virt_alloc_mm_table() - alloc memory for mm table
223 * @adev: amdgpu device.
224 * MM table is used by UVD and VCE for its initialization
225 * Return: Zero if allocate success.
226 */
227int amdgpu_virt_alloc_mm_table(struct amdgpu_device *adev)
228{
229 int r;
230
231 if (!amdgpu_sriov_vf(adev) || adev->virt.mm_table.gpu_addr)
232 return 0;
233
234 r = amdgpu_bo_create_kernel(adev, PAGE_SIZE, PAGE_SIZE,
235 AMDGPU_GEM_DOMAIN_VRAM,
236 &adev->virt.mm_table.bo,
237 &adev->virt.mm_table.gpu_addr,
238 (void *)&adev->virt.mm_table.cpu_addr);
239 if (r) {
240 DRM_ERROR("failed to alloc mm table and error = %d.\n", r);
241 return r;
242 }
243
244 memset((void *)adev->virt.mm_table.cpu_addr, 0, PAGE_SIZE);
245 DRM_INFO("MM table gpu addr = 0x%llx, cpu addr = %p.\n",
246 adev->virt.mm_table.gpu_addr,
247 adev->virt.mm_table.cpu_addr);
248 return 0;
249}
250
251/**
252 * amdgpu_virt_free_mm_table() - free mm table memory
253 * @adev: amdgpu device.
254 * Free MM table memory
255 */
256void amdgpu_virt_free_mm_table(struct amdgpu_device *adev)
257{
258 if (!amdgpu_sriov_vf(adev) || !adev->virt.mm_table.gpu_addr)
259 return;
260
261 amdgpu_bo_free_kernel(&adev->virt.mm_table.bo,
262 &adev->virt.mm_table.gpu_addr,
263 (void *)&adev->virt.mm_table.cpu_addr);
264 adev->virt.mm_table.gpu_addr = 0;
265}
266
267
268unsigned int amd_sriov_msg_checksum(void *obj,
269 unsigned long obj_size,
270 unsigned int key,
271 unsigned int checksum)
272{
273 unsigned int ret = key;
274 unsigned long i = 0;
275 unsigned char *pos;
276
277 pos = (char *)obj;
278 /* calculate checksum */
279 for (i = 0; i < obj_size; ++i)
280 ret += *(pos + i);
281 /* minus the checksum itself */
282 pos = (char *)&checksum;
283 for (i = 0; i < sizeof(checksum); ++i)
284 ret -= *(pos + i);
285 return ret;
286}
287
288static int amdgpu_virt_init_ras_err_handler_data(struct amdgpu_device *adev)
289{
290 struct amdgpu_virt *virt = &adev->virt;
291 struct amdgpu_virt_ras_err_handler_data **data = &virt->virt_eh_data;
292 /* GPU will be marked bad on host if bp count more then 10,
293 * so alloc 512 is enough.
294 */
295 unsigned int align_space = 512;
296 void *bps = NULL;
297 struct amdgpu_bo **bps_bo = NULL;
298
299 *data = kmalloc(sizeof(struct amdgpu_virt_ras_err_handler_data), GFP_KERNEL);
300 if (!*data)
301 goto data_failure;
302
303 bps = kmalloc_array(align_space, sizeof((*data)->bps), GFP_KERNEL);
304 if (!bps)
305 goto bps_failure;
306
307 bps_bo = kmalloc_array(align_space, sizeof((*data)->bps_bo), GFP_KERNEL);
308 if (!bps_bo)
309 goto bps_bo_failure;
310
311 (*data)->bps = bps;
312 (*data)->bps_bo = bps_bo;
313 (*data)->count = 0;
314 (*data)->last_reserved = 0;
315
316 virt->ras_init_done = true;
317
318 return 0;
319
320bps_bo_failure:
321 kfree(bps);
322bps_failure:
323 kfree(*data);
324data_failure:
325 return -ENOMEM;
326}
327
328static void amdgpu_virt_ras_release_bp(struct amdgpu_device *adev)
329{
330 struct amdgpu_virt *virt = &adev->virt;
331 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data;
332 struct amdgpu_bo *bo;
333 int i;
334
335 if (!data)
336 return;
337
338 for (i = data->last_reserved - 1; i >= 0; i--) {
339 bo = data->bps_bo[i];
340 amdgpu_bo_free_kernel(&bo, NULL, NULL);
341 data->bps_bo[i] = bo;
342 data->last_reserved = i;
343 }
344}
345
346void amdgpu_virt_release_ras_err_handler_data(struct amdgpu_device *adev)
347{
348 struct amdgpu_virt *virt = &adev->virt;
349 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data;
350
351 virt->ras_init_done = false;
352
353 if (!data)
354 return;
355
356 amdgpu_virt_ras_release_bp(adev);
357
358 kfree(data->bps);
359 kfree(data->bps_bo);
360 kfree(data);
361 virt->virt_eh_data = NULL;
362}
363
364static void amdgpu_virt_ras_add_bps(struct amdgpu_device *adev,
365 struct eeprom_table_record *bps, int pages)
366{
367 struct amdgpu_virt *virt = &adev->virt;
368 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data;
369
370 if (!data)
371 return;
372
373 memcpy(&data->bps[data->count], bps, pages * sizeof(*data->bps));
374 data->count += pages;
375}
376
377static void amdgpu_virt_ras_reserve_bps(struct amdgpu_device *adev)
378{
379 struct amdgpu_virt *virt = &adev->virt;
380 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data;
381 struct amdgpu_bo *bo = NULL;
382 uint64_t bp;
383 int i;
384
385 if (!data)
386 return;
387
388 for (i = data->last_reserved; i < data->count; i++) {
389 bp = data->bps[i].retired_page;
390
391 /* There are two cases of reserve error should be ignored:
392 * 1) a ras bad page has been allocated (used by someone);
393 * 2) a ras bad page has been reserved (duplicate error injection
394 * for one page);
395 */
396 if (amdgpu_bo_create_kernel_at(adev, bp << AMDGPU_GPU_PAGE_SHIFT,
397 AMDGPU_GPU_PAGE_SIZE,
398 &bo, NULL))
399 DRM_DEBUG("RAS WARN: reserve vram for retired page %llx fail\n", bp);
400
401 data->bps_bo[i] = bo;
402 data->last_reserved = i + 1;
403 bo = NULL;
404 }
405}
406
407static bool amdgpu_virt_ras_check_bad_page(struct amdgpu_device *adev,
408 uint64_t retired_page)
409{
410 struct amdgpu_virt *virt = &adev->virt;
411 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data;
412 int i;
413
414 if (!data)
415 return true;
416
417 for (i = 0; i < data->count; i++)
418 if (retired_page == data->bps[i].retired_page)
419 return true;
420
421 return false;
422}
423
424static void amdgpu_virt_add_bad_page(struct amdgpu_device *adev,
425 uint64_t bp_block_offset, uint32_t bp_block_size)
426{
427 struct eeprom_table_record bp;
428 uint64_t retired_page;
429 uint32_t bp_idx, bp_cnt;
430 void *vram_usage_va = NULL;
431
432 if (adev->mman.fw_vram_usage_va)
433 vram_usage_va = adev->mman.fw_vram_usage_va;
434 else
435 vram_usage_va = adev->mman.drv_vram_usage_va;
436
437 if (bp_block_size) {
438 bp_cnt = bp_block_size / sizeof(uint64_t);
439 for (bp_idx = 0; bp_idx < bp_cnt; bp_idx++) {
440 retired_page = *(uint64_t *)(vram_usage_va +
441 bp_block_offset + bp_idx * sizeof(uint64_t));
442 bp.retired_page = retired_page;
443
444 if (amdgpu_virt_ras_check_bad_page(adev, retired_page))
445 continue;
446
447 amdgpu_virt_ras_add_bps(adev, &bp, 1);
448
449 amdgpu_virt_ras_reserve_bps(adev);
450 }
451 }
452}
453
454static int amdgpu_virt_read_pf2vf_data(struct amdgpu_device *adev)
455{
456 struct amd_sriov_msg_pf2vf_info_header *pf2vf_info = adev->virt.fw_reserve.p_pf2vf;
457 uint32_t checksum;
458 uint32_t checkval;
459
460 uint32_t i;
461 uint32_t tmp;
462
463 if (adev->virt.fw_reserve.p_pf2vf == NULL)
464 return -EINVAL;
465
466 if (pf2vf_info->size > 1024) {
467 DRM_ERROR("invalid pf2vf message size\n");
468 return -EINVAL;
469 }
470
471 switch (pf2vf_info->version) {
472 case 1:
473 checksum = ((struct amdgim_pf2vf_info_v1 *)pf2vf_info)->checksum;
474 checkval = amd_sriov_msg_checksum(
475 adev->virt.fw_reserve.p_pf2vf, pf2vf_info->size,
476 adev->virt.fw_reserve.checksum_key, checksum);
477 if (checksum != checkval) {
478 DRM_ERROR("invalid pf2vf message\n");
479 return -EINVAL;
480 }
481
482 adev->virt.gim_feature =
483 ((struct amdgim_pf2vf_info_v1 *)pf2vf_info)->feature_flags;
484 break;
485 case 2:
486 /* TODO: missing key, need to add it later */
487 checksum = ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->checksum;
488 checkval = amd_sriov_msg_checksum(
489 adev->virt.fw_reserve.p_pf2vf, pf2vf_info->size,
490 0, checksum);
491 if (checksum != checkval) {
492 DRM_ERROR("invalid pf2vf message\n");
493 return -EINVAL;
494 }
495
496 adev->virt.vf2pf_update_interval_ms =
497 ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->vf2pf_update_interval_ms;
498 adev->virt.gim_feature =
499 ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->feature_flags.all;
500 adev->virt.reg_access =
501 ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->reg_access_flags.all;
502
503 adev->virt.decode_max_dimension_pixels = 0;
504 adev->virt.decode_max_frame_pixels = 0;
505 adev->virt.encode_max_dimension_pixels = 0;
506 adev->virt.encode_max_frame_pixels = 0;
507 adev->virt.is_mm_bw_enabled = false;
508 for (i = 0; i < AMD_SRIOV_MSG_RESERVE_VCN_INST; i++) {
509 tmp = ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->mm_bw_management[i].decode_max_dimension_pixels;
510 adev->virt.decode_max_dimension_pixels = max(tmp, adev->virt.decode_max_dimension_pixels);
511
512 tmp = ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->mm_bw_management[i].decode_max_frame_pixels;
513 adev->virt.decode_max_frame_pixels = max(tmp, adev->virt.decode_max_frame_pixels);
514
515 tmp = ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->mm_bw_management[i].encode_max_dimension_pixels;
516 adev->virt.encode_max_dimension_pixels = max(tmp, adev->virt.encode_max_dimension_pixels);
517
518 tmp = ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->mm_bw_management[i].encode_max_frame_pixels;
519 adev->virt.encode_max_frame_pixels = max(tmp, adev->virt.encode_max_frame_pixels);
520 }
521 if((adev->virt.decode_max_dimension_pixels > 0) || (adev->virt.encode_max_dimension_pixels > 0))
522 adev->virt.is_mm_bw_enabled = true;
523
524 adev->unique_id =
525 ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->uuid;
526 break;
527 default:
528 DRM_ERROR("invalid pf2vf version\n");
529 return -EINVAL;
530 }
531
532 /* correct too large or too little interval value */
533 if (adev->virt.vf2pf_update_interval_ms < 200 || adev->virt.vf2pf_update_interval_ms > 10000)
534 adev->virt.vf2pf_update_interval_ms = 2000;
535
536 return 0;
537}
538
539static void amdgpu_virt_populate_vf2pf_ucode_info(struct amdgpu_device *adev)
540{
541 struct amd_sriov_msg_vf2pf_info *vf2pf_info;
542 vf2pf_info = (struct amd_sriov_msg_vf2pf_info *) adev->virt.fw_reserve.p_vf2pf;
543
544 if (adev->virt.fw_reserve.p_vf2pf == NULL)
545 return;
546
547 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_VCE, adev->vce.fw_version);
548 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_UVD, adev->uvd.fw_version);
549 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_MC, adev->gmc.fw_version);
550 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_ME, adev->gfx.me_fw_version);
551 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_PFP, adev->gfx.pfp_fw_version);
552 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_CE, adev->gfx.ce_fw_version);
553 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_RLC, adev->gfx.rlc_fw_version);
554 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_RLC_SRLC, adev->gfx.rlc_srlc_fw_version);
555 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_RLC_SRLG, adev->gfx.rlc_srlg_fw_version);
556 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_RLC_SRLS, adev->gfx.rlc_srls_fw_version);
557 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_MEC, adev->gfx.mec_fw_version);
558 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_MEC2, adev->gfx.mec2_fw_version);
559 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_IMU, adev->gfx.imu_fw_version);
560 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_SOS, adev->psp.sos.fw_version);
561 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_ASD,
562 adev->psp.asd_context.bin_desc.fw_version);
563 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_TA_RAS,
564 adev->psp.ras_context.context.bin_desc.fw_version);
565 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_TA_XGMI,
566 adev->psp.xgmi_context.context.bin_desc.fw_version);
567 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_SMC, adev->pm.fw_version);
568 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_SDMA, adev->sdma.instance[0].fw_version);
569 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_SDMA2, adev->sdma.instance[1].fw_version);
570 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_VCN, adev->vcn.fw_version);
571 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_DMCU, adev->dm.dmcu_fw_version);
572}
573
574static int amdgpu_virt_write_vf2pf_data(struct amdgpu_device *adev)
575{
576 struct amd_sriov_msg_vf2pf_info *vf2pf_info;
577
578 vf2pf_info = (struct amd_sriov_msg_vf2pf_info *) adev->virt.fw_reserve.p_vf2pf;
579
580 if (adev->virt.fw_reserve.p_vf2pf == NULL)
581 return -EINVAL;
582
583 memset(vf2pf_info, 0, sizeof(struct amd_sriov_msg_vf2pf_info));
584
585 vf2pf_info->header.size = sizeof(struct amd_sriov_msg_vf2pf_info);
586 vf2pf_info->header.version = AMD_SRIOV_MSG_FW_VRAM_VF2PF_VER;
587
588#ifdef MODULE
589 if (THIS_MODULE->version != NULL)
590 strcpy(vf2pf_info->driver_version, THIS_MODULE->version);
591 else
592#endif
593 strcpy(vf2pf_info->driver_version, "N/A");
594
595 vf2pf_info->pf2vf_version_required = 0; // no requirement, guest understands all
596 vf2pf_info->driver_cert = 0;
597 vf2pf_info->os_info.all = 0;
598
599 vf2pf_info->fb_usage =
600 ttm_resource_manager_usage(&adev->mman.vram_mgr.manager) >> 20;
601 vf2pf_info->fb_vis_usage =
602 amdgpu_vram_mgr_vis_usage(&adev->mman.vram_mgr) >> 20;
603 vf2pf_info->fb_size = adev->gmc.real_vram_size >> 20;
604 vf2pf_info->fb_vis_size = adev->gmc.visible_vram_size >> 20;
605
606 amdgpu_virt_populate_vf2pf_ucode_info(adev);
607
608 /* TODO: read dynamic info */
609 vf2pf_info->gfx_usage = 0;
610 vf2pf_info->compute_usage = 0;
611 vf2pf_info->encode_usage = 0;
612 vf2pf_info->decode_usage = 0;
613
614 vf2pf_info->dummy_page_addr = (uint64_t)adev->dummy_page_addr;
615 vf2pf_info->checksum =
616 amd_sriov_msg_checksum(
617 vf2pf_info, vf2pf_info->header.size, 0, 0);
618
619 return 0;
620}
621
622static void amdgpu_virt_update_vf2pf_work_item(struct work_struct *work)
623{
624 struct amdgpu_device *adev = container_of(work, struct amdgpu_device, virt.vf2pf_work.work);
625 int ret;
626
627 ret = amdgpu_virt_read_pf2vf_data(adev);
628 if (ret)
629 goto out;
630 amdgpu_virt_write_vf2pf_data(adev);
631
632out:
633 schedule_delayed_work(&(adev->virt.vf2pf_work), adev->virt.vf2pf_update_interval_ms);
634}
635
636void amdgpu_virt_fini_data_exchange(struct amdgpu_device *adev)
637{
638 if (adev->virt.vf2pf_update_interval_ms != 0) {
639 DRM_INFO("clean up the vf2pf work item\n");
640 cancel_delayed_work_sync(&adev->virt.vf2pf_work);
641 adev->virt.vf2pf_update_interval_ms = 0;
642 }
643}
644
645void amdgpu_virt_init_data_exchange(struct amdgpu_device *adev)
646{
647 adev->virt.fw_reserve.p_pf2vf = NULL;
648 adev->virt.fw_reserve.p_vf2pf = NULL;
649 adev->virt.vf2pf_update_interval_ms = 0;
650
651 if (adev->mman.fw_vram_usage_va && adev->mman.drv_vram_usage_va) {
652 DRM_WARN("Currently fw_vram and drv_vram should not have values at the same time!");
653 } else if (adev->mman.fw_vram_usage_va || adev->mman.drv_vram_usage_va) {
654 /* go through this logic in ip_init and reset to init workqueue*/
655 amdgpu_virt_exchange_data(adev);
656
657 INIT_DELAYED_WORK(&adev->virt.vf2pf_work, amdgpu_virt_update_vf2pf_work_item);
658 schedule_delayed_work(&(adev->virt.vf2pf_work), msecs_to_jiffies(adev->virt.vf2pf_update_interval_ms));
659 } else if (adev->bios != NULL) {
660 /* got through this logic in early init stage to get necessary flags, e.g. rlcg_acc related*/
661 adev->virt.fw_reserve.p_pf2vf =
662 (struct amd_sriov_msg_pf2vf_info_header *)
663 (adev->bios + (AMD_SRIOV_MSG_PF2VF_OFFSET_KB << 10));
664
665 amdgpu_virt_read_pf2vf_data(adev);
666 }
667}
668
669
670void amdgpu_virt_exchange_data(struct amdgpu_device *adev)
671{
672 uint64_t bp_block_offset = 0;
673 uint32_t bp_block_size = 0;
674 struct amd_sriov_msg_pf2vf_info *pf2vf_v2 = NULL;
675
676 if (adev->mman.fw_vram_usage_va || adev->mman.drv_vram_usage_va) {
677 if (adev->mman.fw_vram_usage_va) {
678 adev->virt.fw_reserve.p_pf2vf =
679 (struct amd_sriov_msg_pf2vf_info_header *)
680 (adev->mman.fw_vram_usage_va + (AMD_SRIOV_MSG_PF2VF_OFFSET_KB << 10));
681 adev->virt.fw_reserve.p_vf2pf =
682 (struct amd_sriov_msg_vf2pf_info_header *)
683 (adev->mman.fw_vram_usage_va + (AMD_SRIOV_MSG_VF2PF_OFFSET_KB << 10));
684 } else if (adev->mman.drv_vram_usage_va) {
685 adev->virt.fw_reserve.p_pf2vf =
686 (struct amd_sriov_msg_pf2vf_info_header *)
687 (adev->mman.drv_vram_usage_va + (AMD_SRIOV_MSG_PF2VF_OFFSET_KB << 10));
688 adev->virt.fw_reserve.p_vf2pf =
689 (struct amd_sriov_msg_vf2pf_info_header *)
690 (adev->mman.drv_vram_usage_va + (AMD_SRIOV_MSG_VF2PF_OFFSET_KB << 10));
691 }
692
693 amdgpu_virt_read_pf2vf_data(adev);
694 amdgpu_virt_write_vf2pf_data(adev);
695
696 /* bad page handling for version 2 */
697 if (adev->virt.fw_reserve.p_pf2vf->version == 2) {
698 pf2vf_v2 = (struct amd_sriov_msg_pf2vf_info *)adev->virt.fw_reserve.p_pf2vf;
699
700 bp_block_offset = ((uint64_t)pf2vf_v2->bp_block_offset_low & 0xFFFFFFFF) |
701 ((((uint64_t)pf2vf_v2->bp_block_offset_high) << 32) & 0xFFFFFFFF00000000);
702 bp_block_size = pf2vf_v2->bp_block_size;
703
704 if (bp_block_size && !adev->virt.ras_init_done)
705 amdgpu_virt_init_ras_err_handler_data(adev);
706
707 if (adev->virt.ras_init_done)
708 amdgpu_virt_add_bad_page(adev, bp_block_offset, bp_block_size);
709 }
710 }
711}
712
713void amdgpu_detect_virtualization(struct amdgpu_device *adev)
714{
715 uint32_t reg;
716
717 switch (adev->asic_type) {
718 case CHIP_TONGA:
719 case CHIP_FIJI:
720 reg = RREG32(mmBIF_IOV_FUNC_IDENTIFIER);
721 break;
722 case CHIP_VEGA10:
723 case CHIP_VEGA20:
724 case CHIP_NAVI10:
725 case CHIP_NAVI12:
726 case CHIP_SIENNA_CICHLID:
727 case CHIP_ARCTURUS:
728 case CHIP_ALDEBARAN:
729 case CHIP_IP_DISCOVERY:
730 reg = RREG32(mmRCC_IOV_FUNC_IDENTIFIER);
731 break;
732 default: /* other chip doesn't support SRIOV */
733 reg = 0;
734 break;
735 }
736
737 if (reg & 1)
738 adev->virt.caps |= AMDGPU_SRIOV_CAPS_IS_VF;
739
740 if (reg & 0x80000000)
741 adev->virt.caps |= AMDGPU_SRIOV_CAPS_ENABLE_IOV;
742
743 if (!reg) {
744 /* passthrough mode exclus sriov mod */
745 if (is_virtual_machine() && !xen_initial_domain())
746 adev->virt.caps |= AMDGPU_PASSTHROUGH_MODE;
747 }
748
749 if (amdgpu_sriov_vf(adev) && adev->asic_type == CHIP_SIENNA_CICHLID)
750 /* VF MMIO access (except mailbox range) from CPU
751 * will be blocked during sriov runtime
752 */
753 adev->virt.caps |= AMDGPU_VF_MMIO_ACCESS_PROTECT;
754
755 /* we have the ability to check now */
756 if (amdgpu_sriov_vf(adev)) {
757 switch (adev->asic_type) {
758 case CHIP_TONGA:
759 case CHIP_FIJI:
760 vi_set_virt_ops(adev);
761 break;
762 case CHIP_VEGA10:
763 soc15_set_virt_ops(adev);
764#ifdef CONFIG_X86
765 /* not send GPU_INIT_DATA with MS_HYPERV*/
766 if (!hypervisor_is_type(X86_HYPER_MS_HYPERV))
767#endif
768 /* send a dummy GPU_INIT_DATA request to host on vega10 */
769 amdgpu_virt_request_init_data(adev);
770 break;
771 case CHIP_VEGA20:
772 case CHIP_ARCTURUS:
773 case CHIP_ALDEBARAN:
774 soc15_set_virt_ops(adev);
775 break;
776 case CHIP_NAVI10:
777 case CHIP_NAVI12:
778 case CHIP_SIENNA_CICHLID:
779 case CHIP_IP_DISCOVERY:
780 nv_set_virt_ops(adev);
781 /* try send GPU_INIT_DATA request to host */
782 amdgpu_virt_request_init_data(adev);
783 break;
784 default: /* other chip doesn't support SRIOV */
785 DRM_ERROR("Unknown asic type: %d!\n", adev->asic_type);
786 break;
787 }
788 }
789}
790
791static bool amdgpu_virt_access_debugfs_is_mmio(struct amdgpu_device *adev)
792{
793 return amdgpu_sriov_is_debug(adev) ? true : false;
794}
795
796static bool amdgpu_virt_access_debugfs_is_kiq(struct amdgpu_device *adev)
797{
798 return amdgpu_sriov_is_normal(adev) ? true : false;
799}
800
801int amdgpu_virt_enable_access_debugfs(struct amdgpu_device *adev)
802{
803 if (!amdgpu_sriov_vf(adev) ||
804 amdgpu_virt_access_debugfs_is_kiq(adev))
805 return 0;
806
807 if (amdgpu_virt_access_debugfs_is_mmio(adev))
808 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
809 else
810 return -EPERM;
811
812 return 0;
813}
814
815void amdgpu_virt_disable_access_debugfs(struct amdgpu_device *adev)
816{
817 if (amdgpu_sriov_vf(adev))
818 adev->virt.caps |= AMDGPU_SRIOV_CAPS_RUNTIME;
819}
820
821enum amdgpu_sriov_vf_mode amdgpu_virt_get_sriov_vf_mode(struct amdgpu_device *adev)
822{
823 enum amdgpu_sriov_vf_mode mode;
824
825 if (amdgpu_sriov_vf(adev)) {
826 if (amdgpu_sriov_is_pp_one_vf(adev))
827 mode = SRIOV_VF_MODE_ONE_VF;
828 else
829 mode = SRIOV_VF_MODE_MULTI_VF;
830 } else {
831 mode = SRIOV_VF_MODE_BARE_METAL;
832 }
833
834 return mode;
835}
836
837bool amdgpu_virt_fw_load_skip_check(struct amdgpu_device *adev, uint32_t ucode_id)
838{
839 switch (adev->ip_versions[MP0_HWIP][0]) {
840 case IP_VERSION(13, 0, 0):
841 /* no vf autoload, white list */
842 if (ucode_id == AMDGPU_UCODE_ID_VCN1 ||
843 ucode_id == AMDGPU_UCODE_ID_VCN)
844 return false;
845 else
846 return true;
847 case IP_VERSION(13, 0, 10):
848 /* white list */
849 if (ucode_id == AMDGPU_UCODE_ID_CAP
850 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_PFP
851 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_ME
852 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_MEC
853 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_PFP_P0_STACK
854 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_PFP_P1_STACK
855 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_ME_P0_STACK
856 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_ME_P1_STACK
857 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_MEC_P0_STACK
858 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_MEC_P1_STACK
859 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_MEC_P2_STACK
860 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_MEC_P3_STACK
861 || ucode_id == AMDGPU_UCODE_ID_CP_MES
862 || ucode_id == AMDGPU_UCODE_ID_CP_MES_DATA
863 || ucode_id == AMDGPU_UCODE_ID_CP_MES1
864 || ucode_id == AMDGPU_UCODE_ID_CP_MES1_DATA
865 || ucode_id == AMDGPU_UCODE_ID_VCN1
866 || ucode_id == AMDGPU_UCODE_ID_VCN)
867 return false;
868 else
869 return true;
870 default:
871 /* lagacy black list */
872 if (ucode_id == AMDGPU_UCODE_ID_SDMA0
873 || ucode_id == AMDGPU_UCODE_ID_SDMA1
874 || ucode_id == AMDGPU_UCODE_ID_SDMA2
875 || ucode_id == AMDGPU_UCODE_ID_SDMA3
876 || ucode_id == AMDGPU_UCODE_ID_SDMA4
877 || ucode_id == AMDGPU_UCODE_ID_SDMA5
878 || ucode_id == AMDGPU_UCODE_ID_SDMA6
879 || ucode_id == AMDGPU_UCODE_ID_SDMA7
880 || ucode_id == AMDGPU_UCODE_ID_RLC_G
881 || ucode_id == AMDGPU_UCODE_ID_RLC_RESTORE_LIST_CNTL
882 || ucode_id == AMDGPU_UCODE_ID_RLC_RESTORE_LIST_GPM_MEM
883 || ucode_id == AMDGPU_UCODE_ID_RLC_RESTORE_LIST_SRM_MEM
884 || ucode_id == AMDGPU_UCODE_ID_SMC)
885 return true;
886 else
887 return false;
888 }
889}
890
891void amdgpu_virt_update_sriov_video_codec(struct amdgpu_device *adev,
892 struct amdgpu_video_codec_info *encode, uint32_t encode_array_size,
893 struct amdgpu_video_codec_info *decode, uint32_t decode_array_size)
894{
895 uint32_t i;
896
897 if (!adev->virt.is_mm_bw_enabled)
898 return;
899
900 if (encode) {
901 for (i = 0; i < encode_array_size; i++) {
902 encode[i].max_width = adev->virt.encode_max_dimension_pixels;
903 encode[i].max_pixels_per_frame = adev->virt.encode_max_frame_pixels;
904 if (encode[i].max_width > 0)
905 encode[i].max_height = encode[i].max_pixels_per_frame / encode[i].max_width;
906 else
907 encode[i].max_height = 0;
908 }
909 }
910
911 if (decode) {
912 for (i = 0; i < decode_array_size; i++) {
913 decode[i].max_width = adev->virt.decode_max_dimension_pixels;
914 decode[i].max_pixels_per_frame = adev->virt.decode_max_frame_pixels;
915 if (decode[i].max_width > 0)
916 decode[i].max_height = decode[i].max_pixels_per_frame / decode[i].max_width;
917 else
918 decode[i].max_height = 0;
919 }
920 }
921}
922
923static bool amdgpu_virt_get_rlcg_reg_access_flag(struct amdgpu_device *adev,
924 u32 acc_flags, u32 hwip,
925 bool write, u32 *rlcg_flag)
926{
927 bool ret = false;
928
929 switch (hwip) {
930 case GC_HWIP:
931 if (amdgpu_sriov_reg_indirect_gc(adev)) {
932 *rlcg_flag =
933 write ? AMDGPU_RLCG_GC_WRITE : AMDGPU_RLCG_GC_READ;
934 ret = true;
935 /* only in new version, AMDGPU_REGS_NO_KIQ and
936 * AMDGPU_REGS_RLC are enabled simultaneously */
937 } else if ((acc_flags & AMDGPU_REGS_RLC) &&
938 !(acc_flags & AMDGPU_REGS_NO_KIQ) && write) {
939 *rlcg_flag = AMDGPU_RLCG_GC_WRITE_LEGACY;
940 ret = true;
941 }
942 break;
943 case MMHUB_HWIP:
944 if (amdgpu_sriov_reg_indirect_mmhub(adev) &&
945 (acc_flags & AMDGPU_REGS_RLC) && write) {
946 *rlcg_flag = AMDGPU_RLCG_MMHUB_WRITE;
947 ret = true;
948 }
949 break;
950 default:
951 break;
952 }
953 return ret;
954}
955
956static u32 amdgpu_virt_rlcg_reg_rw(struct amdgpu_device *adev, u32 offset, u32 v, u32 flag)
957{
958 struct amdgpu_rlcg_reg_access_ctrl *reg_access_ctrl;
959 uint32_t timeout = 50000;
960 uint32_t i, tmp;
961 uint32_t ret = 0;
962 void *scratch_reg0;
963 void *scratch_reg1;
964 void *scratch_reg2;
965 void *scratch_reg3;
966 void *spare_int;
967
968 if (!adev->gfx.rlc.rlcg_reg_access_supported) {
969 dev_err(adev->dev,
970 "indirect registers access through rlcg is not available\n");
971 return 0;
972 }
973
974 reg_access_ctrl = &adev->gfx.rlc.reg_access_ctrl;
975 scratch_reg0 = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->scratch_reg0;
976 scratch_reg1 = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->scratch_reg1;
977 scratch_reg2 = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->scratch_reg2;
978 scratch_reg3 = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->scratch_reg3;
979 if (reg_access_ctrl->spare_int)
980 spare_int = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->spare_int;
981
982 if (offset == reg_access_ctrl->grbm_cntl) {
983 /* if the target reg offset is grbm_cntl, write to scratch_reg2 */
984 writel(v, scratch_reg2);
985 writel(v, ((void __iomem *)adev->rmmio) + (offset * 4));
986 } else if (offset == reg_access_ctrl->grbm_idx) {
987 /* if the target reg offset is grbm_idx, write to scratch_reg3 */
988 writel(v, scratch_reg3);
989 writel(v, ((void __iomem *)adev->rmmio) + (offset * 4));
990 } else {
991 /*
992 * SCRATCH_REG0 = read/write value
993 * SCRATCH_REG1[30:28] = command
994 * SCRATCH_REG1[19:0] = address in dword
995 * SCRATCH_REG1[26:24] = Error reporting
996 */
997 writel(v, scratch_reg0);
998 writel((offset | flag), scratch_reg1);
999 if (reg_access_ctrl->spare_int)
1000 writel(1, spare_int);
1001
1002 for (i = 0; i < timeout; i++) {
1003 tmp = readl(scratch_reg1);
1004 if (!(tmp & AMDGPU_RLCG_SCRATCH1_ADDRESS_MASK))
1005 break;
1006 udelay(10);
1007 }
1008
1009 if (i >= timeout) {
1010 if (amdgpu_sriov_rlcg_error_report_enabled(adev)) {
1011 if (tmp & AMDGPU_RLCG_VFGATE_DISABLED) {
1012 dev_err(adev->dev,
1013 "vfgate is disabled, rlcg failed to program reg: 0x%05x\n", offset);
1014 } else if (tmp & AMDGPU_RLCG_WRONG_OPERATION_TYPE) {
1015 dev_err(adev->dev,
1016 "wrong operation type, rlcg failed to program reg: 0x%05x\n", offset);
1017 } else if (tmp & AMDGPU_RLCG_REG_NOT_IN_RANGE) {
1018 dev_err(adev->dev,
1019 "register is not in range, rlcg failed to program reg: 0x%05x\n", offset);
1020 } else {
1021 dev_err(adev->dev,
1022 "unknown error type, rlcg failed to program reg: 0x%05x\n", offset);
1023 }
1024 } else {
1025 dev_err(adev->dev,
1026 "timeout: rlcg faled to program reg: 0x%05x\n", offset);
1027 }
1028 }
1029 }
1030
1031 ret = readl(scratch_reg0);
1032 return ret;
1033}
1034
1035void amdgpu_sriov_wreg(struct amdgpu_device *adev,
1036 u32 offset, u32 value,
1037 u32 acc_flags, u32 hwip)
1038{
1039 u32 rlcg_flag;
1040
1041 if (!amdgpu_sriov_runtime(adev) &&
1042 amdgpu_virt_get_rlcg_reg_access_flag(adev, acc_flags, hwip, true, &rlcg_flag)) {
1043 amdgpu_virt_rlcg_reg_rw(adev, offset, value, rlcg_flag);
1044 return;
1045 }
1046
1047 if (acc_flags & AMDGPU_REGS_NO_KIQ)
1048 WREG32_NO_KIQ(offset, value);
1049 else
1050 WREG32(offset, value);
1051}
1052
1053u32 amdgpu_sriov_rreg(struct amdgpu_device *adev,
1054 u32 offset, u32 acc_flags, u32 hwip)
1055{
1056 u32 rlcg_flag;
1057
1058 if (!amdgpu_sriov_runtime(adev) &&
1059 amdgpu_virt_get_rlcg_reg_access_flag(adev, acc_flags, hwip, false, &rlcg_flag))
1060 return amdgpu_virt_rlcg_reg_rw(adev, offset, 0, rlcg_flag);
1061
1062 if (acc_flags & AMDGPU_REGS_NO_KIQ)
1063 return RREG32_NO_KIQ(offset);
1064 else
1065 return RREG32(offset);
1066}