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#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 ((adev->pdev->class >> 8) != PCI_CLASS_ACCELERATOR_PROCESSING)) {
61 if (adev->mode_info.num_crtc == 0)
62 adev->mode_info.num_crtc = 1;
63 adev->enable_virtual_display = true;
64 }
65 ddev->driver_features &= ~DRIVER_ATOMIC;
66 adev->cg_flags = 0;
67 adev->pg_flags = 0;
68
69 /* Reduce kcq number to 2 to reduce latency */
70 if (amdgpu_num_kcq == -1)
71 amdgpu_num_kcq = 2;
72}
73
74/**
75 * amdgpu_virt_request_full_gpu() - request full gpu access
76 * @adev: amdgpu device.
77 * @init: is driver init time.
78 * When start to init/fini driver, first need to request full gpu access.
79 * Return: Zero if request success, otherwise will return error.
80 */
81int amdgpu_virt_request_full_gpu(struct amdgpu_device *adev, bool init)
82{
83 struct amdgpu_virt *virt = &adev->virt;
84 int r;
85
86 if (virt->ops && virt->ops->req_full_gpu) {
87 r = virt->ops->req_full_gpu(adev, init);
88 if (r)
89 return r;
90
91 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
92 }
93
94 return 0;
95}
96
97/**
98 * amdgpu_virt_release_full_gpu() - release full gpu access
99 * @adev: amdgpu device.
100 * @init: is driver init time.
101 * When finishing driver init/fini, need to release full gpu access.
102 * Return: Zero if release success, otherwise will returen error.
103 */
104int amdgpu_virt_release_full_gpu(struct amdgpu_device *adev, bool init)
105{
106 struct amdgpu_virt *virt = &adev->virt;
107 int r;
108
109 if (virt->ops && virt->ops->rel_full_gpu) {
110 r = virt->ops->rel_full_gpu(adev, init);
111 if (r)
112 return r;
113
114 adev->virt.caps |= AMDGPU_SRIOV_CAPS_RUNTIME;
115 }
116 return 0;
117}
118
119/**
120 * amdgpu_virt_reset_gpu() - reset gpu
121 * @adev: amdgpu device.
122 * Send reset command to GPU hypervisor to reset GPU that VM is using
123 * Return: Zero if reset success, otherwise will return error.
124 */
125int amdgpu_virt_reset_gpu(struct amdgpu_device *adev)
126{
127 struct amdgpu_virt *virt = &adev->virt;
128 int r;
129
130 if (virt->ops && virt->ops->reset_gpu) {
131 r = virt->ops->reset_gpu(adev);
132 if (r)
133 return r;
134
135 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
136 }
137
138 return 0;
139}
140
141void amdgpu_virt_request_init_data(struct amdgpu_device *adev)
142{
143 struct amdgpu_virt *virt = &adev->virt;
144
145 if (virt->ops && virt->ops->req_init_data)
146 virt->ops->req_init_data(adev);
147
148 if (adev->virt.req_init_data_ver > 0)
149 DRM_INFO("host supports REQ_INIT_DATA handshake\n");
150 else
151 DRM_WARN("host doesn't support REQ_INIT_DATA handshake\n");
152}
153
154/**
155 * amdgpu_virt_wait_reset() - wait for reset gpu completed
156 * @adev: amdgpu device.
157 * Wait for GPU reset completed.
158 * Return: Zero if reset success, otherwise will return error.
159 */
160int amdgpu_virt_wait_reset(struct amdgpu_device *adev)
161{
162 struct amdgpu_virt *virt = &adev->virt;
163
164 if (!virt->ops || !virt->ops->wait_reset)
165 return -EINVAL;
166
167 return virt->ops->wait_reset(adev);
168}
169
170/**
171 * amdgpu_virt_alloc_mm_table() - alloc memory for mm table
172 * @adev: amdgpu device.
173 * MM table is used by UVD and VCE for its initialization
174 * Return: Zero if allocate success.
175 */
176int amdgpu_virt_alloc_mm_table(struct amdgpu_device *adev)
177{
178 int r;
179
180 if (!amdgpu_sriov_vf(adev) || adev->virt.mm_table.gpu_addr)
181 return 0;
182
183 r = amdgpu_bo_create_kernel(adev, PAGE_SIZE, PAGE_SIZE,
184 AMDGPU_GEM_DOMAIN_VRAM |
185 AMDGPU_GEM_DOMAIN_GTT,
186 &adev->virt.mm_table.bo,
187 &adev->virt.mm_table.gpu_addr,
188 (void *)&adev->virt.mm_table.cpu_addr);
189 if (r) {
190 DRM_ERROR("failed to alloc mm table and error = %d.\n", r);
191 return r;
192 }
193
194 memset((void *)adev->virt.mm_table.cpu_addr, 0, PAGE_SIZE);
195 DRM_INFO("MM table gpu addr = 0x%llx, cpu addr = %p.\n",
196 adev->virt.mm_table.gpu_addr,
197 adev->virt.mm_table.cpu_addr);
198 return 0;
199}
200
201/**
202 * amdgpu_virt_free_mm_table() - free mm table memory
203 * @adev: amdgpu device.
204 * Free MM table memory
205 */
206void amdgpu_virt_free_mm_table(struct amdgpu_device *adev)
207{
208 if (!amdgpu_sriov_vf(adev) || !adev->virt.mm_table.gpu_addr)
209 return;
210
211 amdgpu_bo_free_kernel(&adev->virt.mm_table.bo,
212 &adev->virt.mm_table.gpu_addr,
213 (void *)&adev->virt.mm_table.cpu_addr);
214 adev->virt.mm_table.gpu_addr = 0;
215}
216
217
218unsigned int amd_sriov_msg_checksum(void *obj,
219 unsigned long obj_size,
220 unsigned int key,
221 unsigned int checksum)
222{
223 unsigned int ret = key;
224 unsigned long i = 0;
225 unsigned char *pos;
226
227 pos = (char *)obj;
228 /* calculate checksum */
229 for (i = 0; i < obj_size; ++i)
230 ret += *(pos + i);
231 /* minus the checksum itself */
232 pos = (char *)&checksum;
233 for (i = 0; i < sizeof(checksum); ++i)
234 ret -= *(pos + i);
235 return ret;
236}
237
238static int amdgpu_virt_init_ras_err_handler_data(struct amdgpu_device *adev)
239{
240 struct amdgpu_virt *virt = &adev->virt;
241 struct amdgpu_virt_ras_err_handler_data **data = &virt->virt_eh_data;
242 /* GPU will be marked bad on host if bp count more then 10,
243 * so alloc 512 is enough.
244 */
245 unsigned int align_space = 512;
246 void *bps = NULL;
247 struct amdgpu_bo **bps_bo = NULL;
248
249 *data = kmalloc(sizeof(struct amdgpu_virt_ras_err_handler_data), GFP_KERNEL);
250 if (!*data)
251 goto data_failure;
252
253 bps = kmalloc_array(align_space, sizeof(*(*data)->bps), GFP_KERNEL);
254 if (!bps)
255 goto bps_failure;
256
257 bps_bo = kmalloc_array(align_space, sizeof(*(*data)->bps_bo), GFP_KERNEL);
258 if (!bps_bo)
259 goto bps_bo_failure;
260
261 (*data)->bps = bps;
262 (*data)->bps_bo = bps_bo;
263 (*data)->count = 0;
264 (*data)->last_reserved = 0;
265
266 virt->ras_init_done = true;
267
268 return 0;
269
270bps_bo_failure:
271 kfree(bps);
272bps_failure:
273 kfree(*data);
274data_failure:
275 return -ENOMEM;
276}
277
278static void amdgpu_virt_ras_release_bp(struct amdgpu_device *adev)
279{
280 struct amdgpu_virt *virt = &adev->virt;
281 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data;
282 struct amdgpu_bo *bo;
283 int i;
284
285 if (!data)
286 return;
287
288 for (i = data->last_reserved - 1; i >= 0; i--) {
289 bo = data->bps_bo[i];
290 if (bo) {
291 amdgpu_bo_free_kernel(&bo, NULL, NULL);
292 data->bps_bo[i] = bo;
293 }
294 data->last_reserved = i;
295 }
296}
297
298void amdgpu_virt_release_ras_err_handler_data(struct amdgpu_device *adev)
299{
300 struct amdgpu_virt *virt = &adev->virt;
301 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data;
302
303 virt->ras_init_done = false;
304
305 if (!data)
306 return;
307
308 amdgpu_virt_ras_release_bp(adev);
309
310 kfree(data->bps);
311 kfree(data->bps_bo);
312 kfree(data);
313 virt->virt_eh_data = NULL;
314}
315
316static void amdgpu_virt_ras_add_bps(struct amdgpu_device *adev,
317 struct eeprom_table_record *bps, int pages)
318{
319 struct amdgpu_virt *virt = &adev->virt;
320 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data;
321
322 if (!data)
323 return;
324
325 memcpy(&data->bps[data->count], bps, pages * sizeof(*data->bps));
326 data->count += pages;
327}
328
329static void amdgpu_virt_ras_reserve_bps(struct amdgpu_device *adev)
330{
331 struct amdgpu_virt *virt = &adev->virt;
332 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data;
333 struct amdgpu_vram_mgr *mgr = &adev->mman.vram_mgr;
334 struct ttm_resource_manager *man = &mgr->manager;
335 struct amdgpu_bo *bo = NULL;
336 uint64_t bp;
337 int i;
338
339 if (!data)
340 return;
341
342 for (i = data->last_reserved; i < data->count; i++) {
343 bp = data->bps[i].retired_page;
344
345 /* There are two cases of reserve error should be ignored:
346 * 1) a ras bad page has been allocated (used by someone);
347 * 2) a ras bad page has been reserved (duplicate error injection
348 * for one page);
349 */
350 if (ttm_resource_manager_used(man)) {
351 amdgpu_vram_mgr_reserve_range(&adev->mman.vram_mgr,
352 bp << AMDGPU_GPU_PAGE_SHIFT,
353 AMDGPU_GPU_PAGE_SIZE);
354 data->bps_bo[i] = NULL;
355 } else {
356 if (amdgpu_bo_create_kernel_at(adev, bp << AMDGPU_GPU_PAGE_SHIFT,
357 AMDGPU_GPU_PAGE_SIZE,
358 &bo, NULL))
359 DRM_DEBUG("RAS WARN: reserve vram for retired page %llx fail\n", bp);
360 data->bps_bo[i] = bo;
361 }
362 data->last_reserved = i + 1;
363 bo = NULL;
364 }
365}
366
367static bool amdgpu_virt_ras_check_bad_page(struct amdgpu_device *adev,
368 uint64_t retired_page)
369{
370 struct amdgpu_virt *virt = &adev->virt;
371 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data;
372 int i;
373
374 if (!data)
375 return true;
376
377 for (i = 0; i < data->count; i++)
378 if (retired_page == data->bps[i].retired_page)
379 return true;
380
381 return false;
382}
383
384static void amdgpu_virt_add_bad_page(struct amdgpu_device *adev,
385 uint64_t bp_block_offset, uint32_t bp_block_size)
386{
387 struct eeprom_table_record bp;
388 uint64_t retired_page;
389 uint32_t bp_idx, bp_cnt;
390 void *vram_usage_va = NULL;
391
392 if (adev->mman.fw_vram_usage_va)
393 vram_usage_va = adev->mman.fw_vram_usage_va;
394 else
395 vram_usage_va = adev->mman.drv_vram_usage_va;
396
397 if (bp_block_size) {
398 bp_cnt = bp_block_size / sizeof(uint64_t);
399 for (bp_idx = 0; bp_idx < bp_cnt; bp_idx++) {
400 retired_page = *(uint64_t *)(vram_usage_va +
401 bp_block_offset + bp_idx * sizeof(uint64_t));
402 bp.retired_page = retired_page;
403
404 if (amdgpu_virt_ras_check_bad_page(adev, retired_page))
405 continue;
406
407 amdgpu_virt_ras_add_bps(adev, &bp, 1);
408
409 amdgpu_virt_ras_reserve_bps(adev);
410 }
411 }
412}
413
414static int amdgpu_virt_read_pf2vf_data(struct amdgpu_device *adev)
415{
416 struct amd_sriov_msg_pf2vf_info_header *pf2vf_info = adev->virt.fw_reserve.p_pf2vf;
417 uint32_t checksum;
418 uint32_t checkval;
419
420 uint32_t i;
421 uint32_t tmp;
422
423 if (adev->virt.fw_reserve.p_pf2vf == NULL)
424 return -EINVAL;
425
426 if (pf2vf_info->size > 1024) {
427 DRM_ERROR("invalid pf2vf message size\n");
428 return -EINVAL;
429 }
430
431 switch (pf2vf_info->version) {
432 case 1:
433 checksum = ((struct amdgim_pf2vf_info_v1 *)pf2vf_info)->checksum;
434 checkval = amd_sriov_msg_checksum(
435 adev->virt.fw_reserve.p_pf2vf, pf2vf_info->size,
436 adev->virt.fw_reserve.checksum_key, checksum);
437 if (checksum != checkval) {
438 DRM_ERROR("invalid pf2vf message\n");
439 return -EINVAL;
440 }
441
442 adev->virt.gim_feature =
443 ((struct amdgim_pf2vf_info_v1 *)pf2vf_info)->feature_flags;
444 break;
445 case 2:
446 /* TODO: missing key, need to add it later */
447 checksum = ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->checksum;
448 checkval = amd_sriov_msg_checksum(
449 adev->virt.fw_reserve.p_pf2vf, pf2vf_info->size,
450 0, checksum);
451 if (checksum != checkval) {
452 DRM_ERROR("invalid pf2vf message\n");
453 return -EINVAL;
454 }
455
456 adev->virt.vf2pf_update_interval_ms =
457 ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->vf2pf_update_interval_ms;
458 adev->virt.gim_feature =
459 ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->feature_flags.all;
460 adev->virt.reg_access =
461 ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->reg_access_flags.all;
462
463 adev->virt.decode_max_dimension_pixels = 0;
464 adev->virt.decode_max_frame_pixels = 0;
465 adev->virt.encode_max_dimension_pixels = 0;
466 adev->virt.encode_max_frame_pixels = 0;
467 adev->virt.is_mm_bw_enabled = false;
468 for (i = 0; i < AMD_SRIOV_MSG_RESERVE_VCN_INST; i++) {
469 tmp = ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->mm_bw_management[i].decode_max_dimension_pixels;
470 adev->virt.decode_max_dimension_pixels = max(tmp, adev->virt.decode_max_dimension_pixels);
471
472 tmp = ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->mm_bw_management[i].decode_max_frame_pixels;
473 adev->virt.decode_max_frame_pixels = max(tmp, adev->virt.decode_max_frame_pixels);
474
475 tmp = ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->mm_bw_management[i].encode_max_dimension_pixels;
476 adev->virt.encode_max_dimension_pixels = max(tmp, adev->virt.encode_max_dimension_pixels);
477
478 tmp = ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->mm_bw_management[i].encode_max_frame_pixels;
479 adev->virt.encode_max_frame_pixels = max(tmp, adev->virt.encode_max_frame_pixels);
480 }
481 if ((adev->virt.decode_max_dimension_pixels > 0) || (adev->virt.encode_max_dimension_pixels > 0))
482 adev->virt.is_mm_bw_enabled = true;
483
484 adev->unique_id =
485 ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->uuid;
486 break;
487 default:
488 DRM_ERROR("invalid pf2vf version\n");
489 return -EINVAL;
490 }
491
492 /* correct too large or too little interval value */
493 if (adev->virt.vf2pf_update_interval_ms < 200 || adev->virt.vf2pf_update_interval_ms > 10000)
494 adev->virt.vf2pf_update_interval_ms = 2000;
495
496 return 0;
497}
498
499static void amdgpu_virt_populate_vf2pf_ucode_info(struct amdgpu_device *adev)
500{
501 struct amd_sriov_msg_vf2pf_info *vf2pf_info;
502 vf2pf_info = (struct amd_sriov_msg_vf2pf_info *) adev->virt.fw_reserve.p_vf2pf;
503
504 if (adev->virt.fw_reserve.p_vf2pf == NULL)
505 return;
506
507 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_VCE, adev->vce.fw_version);
508 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_UVD, adev->uvd.fw_version);
509 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_MC, adev->gmc.fw_version);
510 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_ME, adev->gfx.me_fw_version);
511 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_PFP, adev->gfx.pfp_fw_version);
512 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_CE, adev->gfx.ce_fw_version);
513 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_RLC, adev->gfx.rlc_fw_version);
514 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_RLC_SRLC, adev->gfx.rlc_srlc_fw_version);
515 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_RLC_SRLG, adev->gfx.rlc_srlg_fw_version);
516 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_RLC_SRLS, adev->gfx.rlc_srls_fw_version);
517 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_MEC, adev->gfx.mec_fw_version);
518 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_MEC2, adev->gfx.mec2_fw_version);
519 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_SOS, adev->psp.sos.fw_version);
520 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_ASD,
521 adev->psp.asd_context.bin_desc.fw_version);
522 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_TA_RAS,
523 adev->psp.ras_context.context.bin_desc.fw_version);
524 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_TA_XGMI,
525 adev->psp.xgmi_context.context.bin_desc.fw_version);
526 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_SMC, adev->pm.fw_version);
527 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_SDMA, adev->sdma.instance[0].fw_version);
528 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_SDMA2, adev->sdma.instance[1].fw_version);
529 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_VCN, adev->vcn.fw_version);
530 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_DMCU, adev->dm.dmcu_fw_version);
531}
532
533static int amdgpu_virt_write_vf2pf_data(struct amdgpu_device *adev)
534{
535 struct amd_sriov_msg_vf2pf_info *vf2pf_info;
536
537 vf2pf_info = (struct amd_sriov_msg_vf2pf_info *) adev->virt.fw_reserve.p_vf2pf;
538
539 if (adev->virt.fw_reserve.p_vf2pf == NULL)
540 return -EINVAL;
541
542 memset(vf2pf_info, 0, sizeof(struct amd_sriov_msg_vf2pf_info));
543
544 vf2pf_info->header.size = sizeof(struct amd_sriov_msg_vf2pf_info);
545 vf2pf_info->header.version = AMD_SRIOV_MSG_FW_VRAM_VF2PF_VER;
546
547#ifdef MODULE
548 if (THIS_MODULE->version != NULL)
549 strcpy(vf2pf_info->driver_version, THIS_MODULE->version);
550 else
551#endif
552 strcpy(vf2pf_info->driver_version, "N/A");
553
554 vf2pf_info->pf2vf_version_required = 0; // no requirement, guest understands all
555 vf2pf_info->driver_cert = 0;
556 vf2pf_info->os_info.all = 0;
557
558 vf2pf_info->fb_usage =
559 ttm_resource_manager_usage(&adev->mman.vram_mgr.manager) >> 20;
560 vf2pf_info->fb_vis_usage =
561 amdgpu_vram_mgr_vis_usage(&adev->mman.vram_mgr) >> 20;
562 vf2pf_info->fb_size = adev->gmc.real_vram_size >> 20;
563 vf2pf_info->fb_vis_size = adev->gmc.visible_vram_size >> 20;
564
565 amdgpu_virt_populate_vf2pf_ucode_info(adev);
566
567 /* TODO: read dynamic info */
568 vf2pf_info->gfx_usage = 0;
569 vf2pf_info->compute_usage = 0;
570 vf2pf_info->encode_usage = 0;
571 vf2pf_info->decode_usage = 0;
572
573 vf2pf_info->dummy_page_addr = (uint64_t)adev->dummy_page_addr;
574 vf2pf_info->checksum =
575 amd_sriov_msg_checksum(
576 vf2pf_info, vf2pf_info->header.size, 0, 0);
577
578 return 0;
579}
580
581static void amdgpu_virt_update_vf2pf_work_item(struct work_struct *work)
582{
583 struct amdgpu_device *adev = container_of(work, struct amdgpu_device, virt.vf2pf_work.work);
584 int ret;
585
586 ret = amdgpu_virt_read_pf2vf_data(adev);
587 if (ret)
588 goto out;
589 amdgpu_virt_write_vf2pf_data(adev);
590
591out:
592 schedule_delayed_work(&(adev->virt.vf2pf_work), adev->virt.vf2pf_update_interval_ms);
593}
594
595void amdgpu_virt_fini_data_exchange(struct amdgpu_device *adev)
596{
597 if (adev->virt.vf2pf_update_interval_ms != 0) {
598 DRM_INFO("clean up the vf2pf work item\n");
599 cancel_delayed_work_sync(&adev->virt.vf2pf_work);
600 adev->virt.vf2pf_update_interval_ms = 0;
601 }
602}
603
604void amdgpu_virt_init_data_exchange(struct amdgpu_device *adev)
605{
606 adev->virt.fw_reserve.p_pf2vf = NULL;
607 adev->virt.fw_reserve.p_vf2pf = NULL;
608 adev->virt.vf2pf_update_interval_ms = 0;
609
610 if (adev->mman.fw_vram_usage_va && adev->mman.drv_vram_usage_va) {
611 DRM_WARN("Currently fw_vram and drv_vram should not have values at the same time!");
612 } else if (adev->mman.fw_vram_usage_va || adev->mman.drv_vram_usage_va) {
613 /* go through this logic in ip_init and reset to init workqueue*/
614 amdgpu_virt_exchange_data(adev);
615
616 INIT_DELAYED_WORK(&adev->virt.vf2pf_work, amdgpu_virt_update_vf2pf_work_item);
617 schedule_delayed_work(&(adev->virt.vf2pf_work), msecs_to_jiffies(adev->virt.vf2pf_update_interval_ms));
618 } else if (adev->bios != NULL) {
619 /* got through this logic in early init stage to get necessary flags, e.g. rlcg_acc related*/
620 adev->virt.fw_reserve.p_pf2vf =
621 (struct amd_sriov_msg_pf2vf_info_header *)
622 (adev->bios + (AMD_SRIOV_MSG_PF2VF_OFFSET_KB << 10));
623
624 amdgpu_virt_read_pf2vf_data(adev);
625 }
626}
627
628
629void amdgpu_virt_exchange_data(struct amdgpu_device *adev)
630{
631 uint64_t bp_block_offset = 0;
632 uint32_t bp_block_size = 0;
633 struct amd_sriov_msg_pf2vf_info *pf2vf_v2 = NULL;
634
635 if (adev->mman.fw_vram_usage_va || adev->mman.drv_vram_usage_va) {
636 if (adev->mman.fw_vram_usage_va) {
637 adev->virt.fw_reserve.p_pf2vf =
638 (struct amd_sriov_msg_pf2vf_info_header *)
639 (adev->mman.fw_vram_usage_va + (AMD_SRIOV_MSG_PF2VF_OFFSET_KB << 10));
640 adev->virt.fw_reserve.p_vf2pf =
641 (struct amd_sriov_msg_vf2pf_info_header *)
642 (adev->mman.fw_vram_usage_va + (AMD_SRIOV_MSG_VF2PF_OFFSET_KB << 10));
643 } else if (adev->mman.drv_vram_usage_va) {
644 adev->virt.fw_reserve.p_pf2vf =
645 (struct amd_sriov_msg_pf2vf_info_header *)
646 (adev->mman.drv_vram_usage_va + (AMD_SRIOV_MSG_PF2VF_OFFSET_KB << 10));
647 adev->virt.fw_reserve.p_vf2pf =
648 (struct amd_sriov_msg_vf2pf_info_header *)
649 (adev->mman.drv_vram_usage_va + (AMD_SRIOV_MSG_VF2PF_OFFSET_KB << 10));
650 }
651
652 amdgpu_virt_read_pf2vf_data(adev);
653 amdgpu_virt_write_vf2pf_data(adev);
654
655 /* bad page handling for version 2 */
656 if (adev->virt.fw_reserve.p_pf2vf->version == 2) {
657 pf2vf_v2 = (struct amd_sriov_msg_pf2vf_info *)adev->virt.fw_reserve.p_pf2vf;
658
659 bp_block_offset = ((uint64_t)pf2vf_v2->bp_block_offset_low & 0xFFFFFFFF) |
660 ((((uint64_t)pf2vf_v2->bp_block_offset_high) << 32) & 0xFFFFFFFF00000000);
661 bp_block_size = pf2vf_v2->bp_block_size;
662
663 if (bp_block_size && !adev->virt.ras_init_done)
664 amdgpu_virt_init_ras_err_handler_data(adev);
665
666 if (adev->virt.ras_init_done)
667 amdgpu_virt_add_bad_page(adev, bp_block_offset, bp_block_size);
668 }
669 }
670}
671
672void amdgpu_detect_virtualization(struct amdgpu_device *adev)
673{
674 uint32_t reg;
675
676 switch (adev->asic_type) {
677 case CHIP_TONGA:
678 case CHIP_FIJI:
679 reg = RREG32(mmBIF_IOV_FUNC_IDENTIFIER);
680 break;
681 case CHIP_VEGA10:
682 case CHIP_VEGA20:
683 case CHIP_NAVI10:
684 case CHIP_NAVI12:
685 case CHIP_SIENNA_CICHLID:
686 case CHIP_ARCTURUS:
687 case CHIP_ALDEBARAN:
688 case CHIP_IP_DISCOVERY:
689 reg = RREG32(mmRCC_IOV_FUNC_IDENTIFIER);
690 break;
691 default: /* other chip doesn't support SRIOV */
692 reg = 0;
693 break;
694 }
695
696 if (reg & 1)
697 adev->virt.caps |= AMDGPU_SRIOV_CAPS_IS_VF;
698
699 if (reg & 0x80000000)
700 adev->virt.caps |= AMDGPU_SRIOV_CAPS_ENABLE_IOV;
701
702 if (!reg) {
703 /* passthrough mode exclus sriov mod */
704 if (is_virtual_machine() && !xen_initial_domain())
705 adev->virt.caps |= AMDGPU_PASSTHROUGH_MODE;
706 }
707
708 if (amdgpu_sriov_vf(adev) && adev->asic_type == CHIP_SIENNA_CICHLID)
709 /* VF MMIO access (except mailbox range) from CPU
710 * will be blocked during sriov runtime
711 */
712 adev->virt.caps |= AMDGPU_VF_MMIO_ACCESS_PROTECT;
713
714 /* we have the ability to check now */
715 if (amdgpu_sriov_vf(adev)) {
716 switch (adev->asic_type) {
717 case CHIP_TONGA:
718 case CHIP_FIJI:
719 vi_set_virt_ops(adev);
720 break;
721 case CHIP_VEGA10:
722 soc15_set_virt_ops(adev);
723#ifdef CONFIG_X86
724 /* not send GPU_INIT_DATA with MS_HYPERV*/
725 if (!hypervisor_is_type(X86_HYPER_MS_HYPERV))
726#endif
727 /* send a dummy GPU_INIT_DATA request to host on vega10 */
728 amdgpu_virt_request_init_data(adev);
729 break;
730 case CHIP_VEGA20:
731 case CHIP_ARCTURUS:
732 case CHIP_ALDEBARAN:
733 soc15_set_virt_ops(adev);
734 break;
735 case CHIP_NAVI10:
736 case CHIP_NAVI12:
737 case CHIP_SIENNA_CICHLID:
738 case CHIP_IP_DISCOVERY:
739 nv_set_virt_ops(adev);
740 /* try send GPU_INIT_DATA request to host */
741 amdgpu_virt_request_init_data(adev);
742 break;
743 default: /* other chip doesn't support SRIOV */
744 DRM_ERROR("Unknown asic type: %d!\n", adev->asic_type);
745 break;
746 }
747 }
748}
749
750static bool amdgpu_virt_access_debugfs_is_mmio(struct amdgpu_device *adev)
751{
752 return amdgpu_sriov_is_debug(adev) ? true : false;
753}
754
755static bool amdgpu_virt_access_debugfs_is_kiq(struct amdgpu_device *adev)
756{
757 return amdgpu_sriov_is_normal(adev) ? true : false;
758}
759
760int amdgpu_virt_enable_access_debugfs(struct amdgpu_device *adev)
761{
762 if (!amdgpu_sriov_vf(adev) ||
763 amdgpu_virt_access_debugfs_is_kiq(adev))
764 return 0;
765
766 if (amdgpu_virt_access_debugfs_is_mmio(adev))
767 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
768 else
769 return -EPERM;
770
771 return 0;
772}
773
774void amdgpu_virt_disable_access_debugfs(struct amdgpu_device *adev)
775{
776 if (amdgpu_sriov_vf(adev))
777 adev->virt.caps |= AMDGPU_SRIOV_CAPS_RUNTIME;
778}
779
780enum amdgpu_sriov_vf_mode amdgpu_virt_get_sriov_vf_mode(struct amdgpu_device *adev)
781{
782 enum amdgpu_sriov_vf_mode mode;
783
784 if (amdgpu_sriov_vf(adev)) {
785 if (amdgpu_sriov_is_pp_one_vf(adev))
786 mode = SRIOV_VF_MODE_ONE_VF;
787 else
788 mode = SRIOV_VF_MODE_MULTI_VF;
789 } else {
790 mode = SRIOV_VF_MODE_BARE_METAL;
791 }
792
793 return mode;
794}
795
796void amdgpu_virt_post_reset(struct amdgpu_device *adev)
797{
798 if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(11, 0, 3)) {
799 /* force set to GFXOFF state after reset,
800 * to avoid some invalid operation before GC enable
801 */
802 adev->gfx.is_poweron = false;
803 }
804}
805
806bool amdgpu_virt_fw_load_skip_check(struct amdgpu_device *adev, uint32_t ucode_id)
807{
808 switch (amdgpu_ip_version(adev, MP0_HWIP, 0)) {
809 case IP_VERSION(13, 0, 0):
810 /* no vf autoload, white list */
811 if (ucode_id == AMDGPU_UCODE_ID_VCN1 ||
812 ucode_id == AMDGPU_UCODE_ID_VCN)
813 return false;
814 else
815 return true;
816 case IP_VERSION(11, 0, 9):
817 case IP_VERSION(11, 0, 7):
818 /* black list for CHIP_NAVI12 and CHIP_SIENNA_CICHLID */
819 if (ucode_id == AMDGPU_UCODE_ID_RLC_G
820 || ucode_id == AMDGPU_UCODE_ID_RLC_RESTORE_LIST_CNTL
821 || ucode_id == AMDGPU_UCODE_ID_RLC_RESTORE_LIST_GPM_MEM
822 || ucode_id == AMDGPU_UCODE_ID_RLC_RESTORE_LIST_SRM_MEM
823 || ucode_id == AMDGPU_UCODE_ID_SMC)
824 return true;
825 else
826 return false;
827 case IP_VERSION(13, 0, 10):
828 /* white list */
829 if (ucode_id == AMDGPU_UCODE_ID_CAP
830 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_PFP
831 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_ME
832 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_MEC
833 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_PFP_P0_STACK
834 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_PFP_P1_STACK
835 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_ME_P0_STACK
836 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_ME_P1_STACK
837 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_MEC_P0_STACK
838 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_MEC_P1_STACK
839 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_MEC_P2_STACK
840 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_MEC_P3_STACK
841 || ucode_id == AMDGPU_UCODE_ID_CP_MES
842 || ucode_id == AMDGPU_UCODE_ID_CP_MES_DATA
843 || ucode_id == AMDGPU_UCODE_ID_CP_MES1
844 || ucode_id == AMDGPU_UCODE_ID_CP_MES1_DATA
845 || ucode_id == AMDGPU_UCODE_ID_VCN1
846 || ucode_id == AMDGPU_UCODE_ID_VCN)
847 return false;
848 else
849 return true;
850 default:
851 /* lagacy black list */
852 if (ucode_id == AMDGPU_UCODE_ID_SDMA0
853 || ucode_id == AMDGPU_UCODE_ID_SDMA1
854 || ucode_id == AMDGPU_UCODE_ID_SDMA2
855 || ucode_id == AMDGPU_UCODE_ID_SDMA3
856 || ucode_id == AMDGPU_UCODE_ID_SDMA4
857 || ucode_id == AMDGPU_UCODE_ID_SDMA5
858 || ucode_id == AMDGPU_UCODE_ID_SDMA6
859 || ucode_id == AMDGPU_UCODE_ID_SDMA7
860 || ucode_id == AMDGPU_UCODE_ID_RLC_G
861 || ucode_id == AMDGPU_UCODE_ID_RLC_RESTORE_LIST_CNTL
862 || ucode_id == AMDGPU_UCODE_ID_RLC_RESTORE_LIST_GPM_MEM
863 || ucode_id == AMDGPU_UCODE_ID_RLC_RESTORE_LIST_SRM_MEM
864 || ucode_id == AMDGPU_UCODE_ID_SMC)
865 return true;
866 else
867 return false;
868 }
869}
870
871void amdgpu_virt_update_sriov_video_codec(struct amdgpu_device *adev,
872 struct amdgpu_video_codec_info *encode, uint32_t encode_array_size,
873 struct amdgpu_video_codec_info *decode, uint32_t decode_array_size)
874{
875 uint32_t i;
876
877 if (!adev->virt.is_mm_bw_enabled)
878 return;
879
880 if (encode) {
881 for (i = 0; i < encode_array_size; i++) {
882 encode[i].max_width = adev->virt.encode_max_dimension_pixels;
883 encode[i].max_pixels_per_frame = adev->virt.encode_max_frame_pixels;
884 if (encode[i].max_width > 0)
885 encode[i].max_height = encode[i].max_pixels_per_frame / encode[i].max_width;
886 else
887 encode[i].max_height = 0;
888 }
889 }
890
891 if (decode) {
892 for (i = 0; i < decode_array_size; i++) {
893 decode[i].max_width = adev->virt.decode_max_dimension_pixels;
894 decode[i].max_pixels_per_frame = adev->virt.decode_max_frame_pixels;
895 if (decode[i].max_width > 0)
896 decode[i].max_height = decode[i].max_pixels_per_frame / decode[i].max_width;
897 else
898 decode[i].max_height = 0;
899 }
900 }
901}
902
903bool amdgpu_virt_get_rlcg_reg_access_flag(struct amdgpu_device *adev,
904 u32 acc_flags, u32 hwip,
905 bool write, u32 *rlcg_flag)
906{
907 bool ret = false;
908
909 switch (hwip) {
910 case GC_HWIP:
911 if (amdgpu_sriov_reg_indirect_gc(adev)) {
912 *rlcg_flag =
913 write ? AMDGPU_RLCG_GC_WRITE : AMDGPU_RLCG_GC_READ;
914 ret = true;
915 /* only in new version, AMDGPU_REGS_NO_KIQ and
916 * AMDGPU_REGS_RLC are enabled simultaneously */
917 } else if ((acc_flags & AMDGPU_REGS_RLC) &&
918 !(acc_flags & AMDGPU_REGS_NO_KIQ) && write) {
919 *rlcg_flag = AMDGPU_RLCG_GC_WRITE_LEGACY;
920 ret = true;
921 }
922 break;
923 case MMHUB_HWIP:
924 if (amdgpu_sriov_reg_indirect_mmhub(adev) &&
925 (acc_flags & AMDGPU_REGS_RLC) && write) {
926 *rlcg_flag = AMDGPU_RLCG_MMHUB_WRITE;
927 ret = true;
928 }
929 break;
930 default:
931 break;
932 }
933 return ret;
934}
935
936u32 amdgpu_virt_rlcg_reg_rw(struct amdgpu_device *adev, u32 offset, u32 v, u32 flag, u32 xcc_id)
937{
938 struct amdgpu_rlcg_reg_access_ctrl *reg_access_ctrl;
939 uint32_t timeout = 50000;
940 uint32_t i, tmp;
941 uint32_t ret = 0;
942 void *scratch_reg0;
943 void *scratch_reg1;
944 void *scratch_reg2;
945 void *scratch_reg3;
946 void *spare_int;
947
948 if (!adev->gfx.rlc.rlcg_reg_access_supported) {
949 dev_err(adev->dev,
950 "indirect registers access through rlcg is not available\n");
951 return 0;
952 }
953
954 if (adev->gfx.xcc_mask && (((1 << xcc_id) & adev->gfx.xcc_mask) == 0)) {
955 dev_err(adev->dev, "invalid xcc\n");
956 return 0;
957 }
958
959 reg_access_ctrl = &adev->gfx.rlc.reg_access_ctrl[xcc_id];
960 scratch_reg0 = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->scratch_reg0;
961 scratch_reg1 = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->scratch_reg1;
962 scratch_reg2 = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->scratch_reg2;
963 scratch_reg3 = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->scratch_reg3;
964 if (reg_access_ctrl->spare_int)
965 spare_int = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->spare_int;
966
967 if (offset == reg_access_ctrl->grbm_cntl) {
968 /* if the target reg offset is grbm_cntl, write to scratch_reg2 */
969 writel(v, scratch_reg2);
970 if (flag == AMDGPU_RLCG_GC_WRITE_LEGACY)
971 writel(v, ((void __iomem *)adev->rmmio) + (offset * 4));
972 } else if (offset == reg_access_ctrl->grbm_idx) {
973 /* if the target reg offset is grbm_idx, write to scratch_reg3 */
974 writel(v, scratch_reg3);
975 if (flag == AMDGPU_RLCG_GC_WRITE_LEGACY)
976 writel(v, ((void __iomem *)adev->rmmio) + (offset * 4));
977 } else {
978 /*
979 * SCRATCH_REG0 = read/write value
980 * SCRATCH_REG1[30:28] = command
981 * SCRATCH_REG1[19:0] = address in dword
982 * SCRATCH_REG1[27:24] = Error reporting
983 */
984 writel(v, scratch_reg0);
985 writel((offset | flag), scratch_reg1);
986 if (reg_access_ctrl->spare_int)
987 writel(1, spare_int);
988
989 for (i = 0; i < timeout; i++) {
990 tmp = readl(scratch_reg1);
991 if (!(tmp & AMDGPU_RLCG_SCRATCH1_ADDRESS_MASK))
992 break;
993 udelay(10);
994 }
995
996 tmp = readl(scratch_reg1);
997 if (i >= timeout || (tmp & AMDGPU_RLCG_SCRATCH1_ERROR_MASK) != 0) {
998 if (amdgpu_sriov_rlcg_error_report_enabled(adev)) {
999 if (tmp & AMDGPU_RLCG_VFGATE_DISABLED) {
1000 dev_err(adev->dev,
1001 "vfgate is disabled, rlcg failed to program reg: 0x%05x\n", offset);
1002 } else if (tmp & AMDGPU_RLCG_WRONG_OPERATION_TYPE) {
1003 dev_err(adev->dev,
1004 "wrong operation type, rlcg failed to program reg: 0x%05x\n", offset);
1005 } else if (tmp & AMDGPU_RLCG_REG_NOT_IN_RANGE) {
1006 dev_err(adev->dev,
1007 "register is not in range, rlcg failed to program reg: 0x%05x\n", offset);
1008 } else {
1009 dev_err(adev->dev,
1010 "unknown error type, rlcg failed to program reg: 0x%05x\n", offset);
1011 }
1012 } else {
1013 dev_err(adev->dev,
1014 "timeout: rlcg faled to program reg: 0x%05x\n", offset);
1015 }
1016 }
1017 }
1018
1019 ret = readl(scratch_reg0);
1020 return ret;
1021}
1022
1023void amdgpu_sriov_wreg(struct amdgpu_device *adev,
1024 u32 offset, u32 value,
1025 u32 acc_flags, u32 hwip, u32 xcc_id)
1026{
1027 u32 rlcg_flag;
1028
1029 if (!amdgpu_sriov_runtime(adev) &&
1030 amdgpu_virt_get_rlcg_reg_access_flag(adev, acc_flags, hwip, true, &rlcg_flag)) {
1031 amdgpu_virt_rlcg_reg_rw(adev, offset, value, rlcg_flag, xcc_id);
1032 return;
1033 }
1034
1035 if (acc_flags & AMDGPU_REGS_NO_KIQ)
1036 WREG32_NO_KIQ(offset, value);
1037 else
1038 WREG32(offset, value);
1039}
1040
1041u32 amdgpu_sriov_rreg(struct amdgpu_device *adev,
1042 u32 offset, u32 acc_flags, u32 hwip, u32 xcc_id)
1043{
1044 u32 rlcg_flag;
1045
1046 if (!amdgpu_sriov_runtime(adev) &&
1047 amdgpu_virt_get_rlcg_reg_access_flag(adev, acc_flags, hwip, false, &rlcg_flag))
1048 return amdgpu_virt_rlcg_reg_rw(adev, offset, 0, rlcg_flag, xcc_id);
1049
1050 if (acc_flags & AMDGPU_REGS_NO_KIQ)
1051 return RREG32_NO_KIQ(offset);
1052 else
1053 return RREG32(offset);
1054}
1055
1056bool amdgpu_sriov_xnack_support(struct amdgpu_device *adev)
1057{
1058 bool xnack_mode = true;
1059
1060 if (amdgpu_sriov_vf(adev) &&
1061 amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 2))
1062 xnack_mode = false;
1063
1064 return xnack_mode;
1065}
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 ((adev->pdev->class >> 8) != PCI_CLASS_ACCELERATOR_PROCESSING)) {
61 if (adev->mode_info.num_crtc == 0)
62 adev->mode_info.num_crtc = 1;
63 adev->enable_virtual_display = true;
64 }
65 ddev->driver_features &= ~DRIVER_ATOMIC;
66 adev->cg_flags = 0;
67 adev->pg_flags = 0;
68
69 /* Reduce kcq number to 2 to reduce latency */
70 if (amdgpu_num_kcq == -1)
71 amdgpu_num_kcq = 2;
72}
73
74void amdgpu_virt_kiq_reg_write_reg_wait(struct amdgpu_device *adev,
75 uint32_t reg0, uint32_t reg1,
76 uint32_t ref, uint32_t mask,
77 uint32_t xcc_inst)
78{
79 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_inst];
80 struct amdgpu_ring *ring = &kiq->ring;
81 signed long r, cnt = 0;
82 unsigned long flags;
83 uint32_t seq;
84
85 if (adev->mes.ring.sched.ready) {
86 amdgpu_mes_reg_write_reg_wait(adev, reg0, reg1,
87 ref, mask);
88 return;
89 }
90
91 spin_lock_irqsave(&kiq->ring_lock, flags);
92 amdgpu_ring_alloc(ring, 32);
93 amdgpu_ring_emit_reg_write_reg_wait(ring, reg0, reg1,
94 ref, mask);
95 r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
96 if (r)
97 goto failed_undo;
98
99 amdgpu_ring_commit(ring);
100 spin_unlock_irqrestore(&kiq->ring_lock, flags);
101
102 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
103
104 /* don't wait anymore for IRQ context */
105 if (r < 1 && in_interrupt())
106 goto failed_kiq;
107
108 might_sleep();
109 while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
110
111 msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
112 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
113 }
114
115 if (cnt > MAX_KIQ_REG_TRY)
116 goto failed_kiq;
117
118 return;
119
120failed_undo:
121 amdgpu_ring_undo(ring);
122 spin_unlock_irqrestore(&kiq->ring_lock, flags);
123failed_kiq:
124 dev_err(adev->dev, "failed to write reg %x wait reg %x\n", reg0, reg1);
125}
126
127/**
128 * amdgpu_virt_request_full_gpu() - request full gpu access
129 * @adev: amdgpu device.
130 * @init: is driver init time.
131 * When start to init/fini driver, first need to request full gpu access.
132 * Return: Zero if request success, otherwise will return error.
133 */
134int amdgpu_virt_request_full_gpu(struct amdgpu_device *adev, bool init)
135{
136 struct amdgpu_virt *virt = &adev->virt;
137 int r;
138
139 if (virt->ops && virt->ops->req_full_gpu) {
140 r = virt->ops->req_full_gpu(adev, init);
141 if (r)
142 return r;
143
144 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
145 }
146
147 return 0;
148}
149
150/**
151 * amdgpu_virt_release_full_gpu() - release full gpu access
152 * @adev: amdgpu device.
153 * @init: is driver init time.
154 * When finishing driver init/fini, need to release full gpu access.
155 * Return: Zero if release success, otherwise will returen error.
156 */
157int amdgpu_virt_release_full_gpu(struct amdgpu_device *adev, bool init)
158{
159 struct amdgpu_virt *virt = &adev->virt;
160 int r;
161
162 if (virt->ops && virt->ops->rel_full_gpu) {
163 r = virt->ops->rel_full_gpu(adev, init);
164 if (r)
165 return r;
166
167 adev->virt.caps |= AMDGPU_SRIOV_CAPS_RUNTIME;
168 }
169 return 0;
170}
171
172/**
173 * amdgpu_virt_reset_gpu() - reset gpu
174 * @adev: amdgpu device.
175 * Send reset command to GPU hypervisor to reset GPU that VM is using
176 * Return: Zero if reset success, otherwise will return error.
177 */
178int amdgpu_virt_reset_gpu(struct amdgpu_device *adev)
179{
180 struct amdgpu_virt *virt = &adev->virt;
181 int r;
182
183 if (virt->ops && virt->ops->reset_gpu) {
184 r = virt->ops->reset_gpu(adev);
185 if (r)
186 return r;
187
188 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
189 }
190
191 return 0;
192}
193
194void amdgpu_virt_request_init_data(struct amdgpu_device *adev)
195{
196 struct amdgpu_virt *virt = &adev->virt;
197
198 if (virt->ops && virt->ops->req_init_data)
199 virt->ops->req_init_data(adev);
200
201 if (adev->virt.req_init_data_ver > 0)
202 DRM_INFO("host supports REQ_INIT_DATA handshake\n");
203 else
204 DRM_WARN("host doesn't support REQ_INIT_DATA handshake\n");
205}
206
207/**
208 * amdgpu_virt_wait_reset() - wait for reset gpu completed
209 * @adev: amdgpu device.
210 * Wait for GPU reset completed.
211 * Return: Zero if reset success, otherwise will return error.
212 */
213int amdgpu_virt_wait_reset(struct amdgpu_device *adev)
214{
215 struct amdgpu_virt *virt = &adev->virt;
216
217 if (!virt->ops || !virt->ops->wait_reset)
218 return -EINVAL;
219
220 return virt->ops->wait_reset(adev);
221}
222
223/**
224 * amdgpu_virt_alloc_mm_table() - alloc memory for mm table
225 * @adev: amdgpu device.
226 * MM table is used by UVD and VCE for its initialization
227 * Return: Zero if allocate success.
228 */
229int amdgpu_virt_alloc_mm_table(struct amdgpu_device *adev)
230{
231 int r;
232
233 if (!amdgpu_sriov_vf(adev) || adev->virt.mm_table.gpu_addr)
234 return 0;
235
236 r = amdgpu_bo_create_kernel(adev, PAGE_SIZE, PAGE_SIZE,
237 AMDGPU_GEM_DOMAIN_VRAM |
238 AMDGPU_GEM_DOMAIN_GTT,
239 &adev->virt.mm_table.bo,
240 &adev->virt.mm_table.gpu_addr,
241 (void *)&adev->virt.mm_table.cpu_addr);
242 if (r) {
243 DRM_ERROR("failed to alloc mm table and error = %d.\n", r);
244 return r;
245 }
246
247 memset((void *)adev->virt.mm_table.cpu_addr, 0, PAGE_SIZE);
248 DRM_INFO("MM table gpu addr = 0x%llx, cpu addr = %p.\n",
249 adev->virt.mm_table.gpu_addr,
250 adev->virt.mm_table.cpu_addr);
251 return 0;
252}
253
254/**
255 * amdgpu_virt_free_mm_table() - free mm table memory
256 * @adev: amdgpu device.
257 * Free MM table memory
258 */
259void amdgpu_virt_free_mm_table(struct amdgpu_device *adev)
260{
261 if (!amdgpu_sriov_vf(adev) || !adev->virt.mm_table.gpu_addr)
262 return;
263
264 amdgpu_bo_free_kernel(&adev->virt.mm_table.bo,
265 &adev->virt.mm_table.gpu_addr,
266 (void *)&adev->virt.mm_table.cpu_addr);
267 adev->virt.mm_table.gpu_addr = 0;
268}
269
270
271unsigned int amd_sriov_msg_checksum(void *obj,
272 unsigned long obj_size,
273 unsigned int key,
274 unsigned int checksum)
275{
276 unsigned int ret = key;
277 unsigned long i = 0;
278 unsigned char *pos;
279
280 pos = (char *)obj;
281 /* calculate checksum */
282 for (i = 0; i < obj_size; ++i)
283 ret += *(pos + i);
284 /* minus the checksum itself */
285 pos = (char *)&checksum;
286 for (i = 0; i < sizeof(checksum); ++i)
287 ret -= *(pos + i);
288 return ret;
289}
290
291static int amdgpu_virt_init_ras_err_handler_data(struct amdgpu_device *adev)
292{
293 struct amdgpu_virt *virt = &adev->virt;
294 struct amdgpu_virt_ras_err_handler_data **data = &virt->virt_eh_data;
295 /* GPU will be marked bad on host if bp count more then 10,
296 * so alloc 512 is enough.
297 */
298 unsigned int align_space = 512;
299 void *bps = NULL;
300 struct amdgpu_bo **bps_bo = NULL;
301
302 *data = kmalloc(sizeof(struct amdgpu_virt_ras_err_handler_data), GFP_KERNEL);
303 if (!*data)
304 goto data_failure;
305
306 bps = kmalloc_array(align_space, sizeof((*data)->bps), GFP_KERNEL);
307 if (!bps)
308 goto bps_failure;
309
310 bps_bo = kmalloc_array(align_space, sizeof((*data)->bps_bo), GFP_KERNEL);
311 if (!bps_bo)
312 goto bps_bo_failure;
313
314 (*data)->bps = bps;
315 (*data)->bps_bo = bps_bo;
316 (*data)->count = 0;
317 (*data)->last_reserved = 0;
318
319 virt->ras_init_done = true;
320
321 return 0;
322
323bps_bo_failure:
324 kfree(bps);
325bps_failure:
326 kfree(*data);
327data_failure:
328 return -ENOMEM;
329}
330
331static void amdgpu_virt_ras_release_bp(struct amdgpu_device *adev)
332{
333 struct amdgpu_virt *virt = &adev->virt;
334 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data;
335 struct amdgpu_bo *bo;
336 int i;
337
338 if (!data)
339 return;
340
341 for (i = data->last_reserved - 1; i >= 0; i--) {
342 bo = data->bps_bo[i];
343 amdgpu_bo_free_kernel(&bo, NULL, NULL);
344 data->bps_bo[i] = bo;
345 data->last_reserved = i;
346 }
347}
348
349void amdgpu_virt_release_ras_err_handler_data(struct amdgpu_device *adev)
350{
351 struct amdgpu_virt *virt = &adev->virt;
352 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data;
353
354 virt->ras_init_done = false;
355
356 if (!data)
357 return;
358
359 amdgpu_virt_ras_release_bp(adev);
360
361 kfree(data->bps);
362 kfree(data->bps_bo);
363 kfree(data);
364 virt->virt_eh_data = NULL;
365}
366
367static void amdgpu_virt_ras_add_bps(struct amdgpu_device *adev,
368 struct eeprom_table_record *bps, int pages)
369{
370 struct amdgpu_virt *virt = &adev->virt;
371 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data;
372
373 if (!data)
374 return;
375
376 memcpy(&data->bps[data->count], bps, pages * sizeof(*data->bps));
377 data->count += pages;
378}
379
380static void amdgpu_virt_ras_reserve_bps(struct amdgpu_device *adev)
381{
382 struct amdgpu_virt *virt = &adev->virt;
383 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data;
384 struct amdgpu_bo *bo = NULL;
385 uint64_t bp;
386 int i;
387
388 if (!data)
389 return;
390
391 for (i = data->last_reserved; i < data->count; i++) {
392 bp = data->bps[i].retired_page;
393
394 /* There are two cases of reserve error should be ignored:
395 * 1) a ras bad page has been allocated (used by someone);
396 * 2) a ras bad page has been reserved (duplicate error injection
397 * for one page);
398 */
399 if (amdgpu_bo_create_kernel_at(adev, bp << AMDGPU_GPU_PAGE_SHIFT,
400 AMDGPU_GPU_PAGE_SIZE,
401 &bo, NULL))
402 DRM_DEBUG("RAS WARN: reserve vram for retired page %llx fail\n", bp);
403
404 data->bps_bo[i] = bo;
405 data->last_reserved = i + 1;
406 bo = NULL;
407 }
408}
409
410static bool amdgpu_virt_ras_check_bad_page(struct amdgpu_device *adev,
411 uint64_t retired_page)
412{
413 struct amdgpu_virt *virt = &adev->virt;
414 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data;
415 int i;
416
417 if (!data)
418 return true;
419
420 for (i = 0; i < data->count; i++)
421 if (retired_page == data->bps[i].retired_page)
422 return true;
423
424 return false;
425}
426
427static void amdgpu_virt_add_bad_page(struct amdgpu_device *adev,
428 uint64_t bp_block_offset, uint32_t bp_block_size)
429{
430 struct eeprom_table_record bp;
431 uint64_t retired_page;
432 uint32_t bp_idx, bp_cnt;
433 void *vram_usage_va = NULL;
434
435 if (adev->mman.fw_vram_usage_va)
436 vram_usage_va = adev->mman.fw_vram_usage_va;
437 else
438 vram_usage_va = adev->mman.drv_vram_usage_va;
439
440 if (bp_block_size) {
441 bp_cnt = bp_block_size / sizeof(uint64_t);
442 for (bp_idx = 0; bp_idx < bp_cnt; bp_idx++) {
443 retired_page = *(uint64_t *)(vram_usage_va +
444 bp_block_offset + bp_idx * sizeof(uint64_t));
445 bp.retired_page = retired_page;
446
447 if (amdgpu_virt_ras_check_bad_page(adev, retired_page))
448 continue;
449
450 amdgpu_virt_ras_add_bps(adev, &bp, 1);
451
452 amdgpu_virt_ras_reserve_bps(adev);
453 }
454 }
455}
456
457static int amdgpu_virt_read_pf2vf_data(struct amdgpu_device *adev)
458{
459 struct amd_sriov_msg_pf2vf_info_header *pf2vf_info = adev->virt.fw_reserve.p_pf2vf;
460 uint32_t checksum;
461 uint32_t checkval;
462
463 uint32_t i;
464 uint32_t tmp;
465
466 if (adev->virt.fw_reserve.p_pf2vf == NULL)
467 return -EINVAL;
468
469 if (pf2vf_info->size > 1024) {
470 DRM_ERROR("invalid pf2vf message size\n");
471 return -EINVAL;
472 }
473
474 switch (pf2vf_info->version) {
475 case 1:
476 checksum = ((struct amdgim_pf2vf_info_v1 *)pf2vf_info)->checksum;
477 checkval = amd_sriov_msg_checksum(
478 adev->virt.fw_reserve.p_pf2vf, pf2vf_info->size,
479 adev->virt.fw_reserve.checksum_key, checksum);
480 if (checksum != checkval) {
481 DRM_ERROR("invalid pf2vf message\n");
482 return -EINVAL;
483 }
484
485 adev->virt.gim_feature =
486 ((struct amdgim_pf2vf_info_v1 *)pf2vf_info)->feature_flags;
487 break;
488 case 2:
489 /* TODO: missing key, need to add it later */
490 checksum = ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->checksum;
491 checkval = amd_sriov_msg_checksum(
492 adev->virt.fw_reserve.p_pf2vf, pf2vf_info->size,
493 0, checksum);
494 if (checksum != checkval) {
495 DRM_ERROR("invalid pf2vf message\n");
496 return -EINVAL;
497 }
498
499 adev->virt.vf2pf_update_interval_ms =
500 ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->vf2pf_update_interval_ms;
501 adev->virt.gim_feature =
502 ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->feature_flags.all;
503 adev->virt.reg_access =
504 ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->reg_access_flags.all;
505
506 adev->virt.decode_max_dimension_pixels = 0;
507 adev->virt.decode_max_frame_pixels = 0;
508 adev->virt.encode_max_dimension_pixels = 0;
509 adev->virt.encode_max_frame_pixels = 0;
510 adev->virt.is_mm_bw_enabled = false;
511 for (i = 0; i < AMD_SRIOV_MSG_RESERVE_VCN_INST; i++) {
512 tmp = ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->mm_bw_management[i].decode_max_dimension_pixels;
513 adev->virt.decode_max_dimension_pixels = max(tmp, adev->virt.decode_max_dimension_pixels);
514
515 tmp = ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->mm_bw_management[i].decode_max_frame_pixels;
516 adev->virt.decode_max_frame_pixels = max(tmp, adev->virt.decode_max_frame_pixels);
517
518 tmp = ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->mm_bw_management[i].encode_max_dimension_pixels;
519 adev->virt.encode_max_dimension_pixels = max(tmp, adev->virt.encode_max_dimension_pixels);
520
521 tmp = ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->mm_bw_management[i].encode_max_frame_pixels;
522 adev->virt.encode_max_frame_pixels = max(tmp, adev->virt.encode_max_frame_pixels);
523 }
524 if ((adev->virt.decode_max_dimension_pixels > 0) || (adev->virt.encode_max_dimension_pixels > 0))
525 adev->virt.is_mm_bw_enabled = true;
526
527 adev->unique_id =
528 ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->uuid;
529 break;
530 default:
531 DRM_ERROR("invalid pf2vf version\n");
532 return -EINVAL;
533 }
534
535 /* correct too large or too little interval value */
536 if (adev->virt.vf2pf_update_interval_ms < 200 || adev->virt.vf2pf_update_interval_ms > 10000)
537 adev->virt.vf2pf_update_interval_ms = 2000;
538
539 return 0;
540}
541
542static void amdgpu_virt_populate_vf2pf_ucode_info(struct amdgpu_device *adev)
543{
544 struct amd_sriov_msg_vf2pf_info *vf2pf_info;
545 vf2pf_info = (struct amd_sriov_msg_vf2pf_info *) adev->virt.fw_reserve.p_vf2pf;
546
547 if (adev->virt.fw_reserve.p_vf2pf == NULL)
548 return;
549
550 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_VCE, adev->vce.fw_version);
551 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_UVD, adev->uvd.fw_version);
552 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_MC, adev->gmc.fw_version);
553 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_ME, adev->gfx.me_fw_version);
554 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_PFP, adev->gfx.pfp_fw_version);
555 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_CE, adev->gfx.ce_fw_version);
556 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_RLC, adev->gfx.rlc_fw_version);
557 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_RLC_SRLC, adev->gfx.rlc_srlc_fw_version);
558 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_RLC_SRLG, adev->gfx.rlc_srlg_fw_version);
559 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_RLC_SRLS, adev->gfx.rlc_srls_fw_version);
560 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_MEC, adev->gfx.mec_fw_version);
561 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_MEC2, adev->gfx.mec2_fw_version);
562 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_SOS, adev->psp.sos.fw_version);
563 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_ASD,
564 adev->psp.asd_context.bin_desc.fw_version);
565 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_TA_RAS,
566 adev->psp.ras_context.context.bin_desc.fw_version);
567 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_TA_XGMI,
568 adev->psp.xgmi_context.context.bin_desc.fw_version);
569 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_SMC, adev->pm.fw_version);
570 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_SDMA, adev->sdma.instance[0].fw_version);
571 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_SDMA2, adev->sdma.instance[1].fw_version);
572 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_VCN, adev->vcn.fw_version);
573 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_DMCU, adev->dm.dmcu_fw_version);
574}
575
576static int amdgpu_virt_write_vf2pf_data(struct amdgpu_device *adev)
577{
578 struct amd_sriov_msg_vf2pf_info *vf2pf_info;
579
580 vf2pf_info = (struct amd_sriov_msg_vf2pf_info *) adev->virt.fw_reserve.p_vf2pf;
581
582 if (adev->virt.fw_reserve.p_vf2pf == NULL)
583 return -EINVAL;
584
585 memset(vf2pf_info, 0, sizeof(struct amd_sriov_msg_vf2pf_info));
586
587 vf2pf_info->header.size = sizeof(struct amd_sriov_msg_vf2pf_info);
588 vf2pf_info->header.version = AMD_SRIOV_MSG_FW_VRAM_VF2PF_VER;
589
590#ifdef MODULE
591 if (THIS_MODULE->version != NULL)
592 strcpy(vf2pf_info->driver_version, THIS_MODULE->version);
593 else
594#endif
595 strcpy(vf2pf_info->driver_version, "N/A");
596
597 vf2pf_info->pf2vf_version_required = 0; // no requirement, guest understands all
598 vf2pf_info->driver_cert = 0;
599 vf2pf_info->os_info.all = 0;
600
601 vf2pf_info->fb_usage =
602 ttm_resource_manager_usage(&adev->mman.vram_mgr.manager) >> 20;
603 vf2pf_info->fb_vis_usage =
604 amdgpu_vram_mgr_vis_usage(&adev->mman.vram_mgr) >> 20;
605 vf2pf_info->fb_size = adev->gmc.real_vram_size >> 20;
606 vf2pf_info->fb_vis_size = adev->gmc.visible_vram_size >> 20;
607
608 amdgpu_virt_populate_vf2pf_ucode_info(adev);
609
610 /* TODO: read dynamic info */
611 vf2pf_info->gfx_usage = 0;
612 vf2pf_info->compute_usage = 0;
613 vf2pf_info->encode_usage = 0;
614 vf2pf_info->decode_usage = 0;
615
616 vf2pf_info->dummy_page_addr = (uint64_t)adev->dummy_page_addr;
617 vf2pf_info->checksum =
618 amd_sriov_msg_checksum(
619 vf2pf_info, vf2pf_info->header.size, 0, 0);
620
621 return 0;
622}
623
624static void amdgpu_virt_update_vf2pf_work_item(struct work_struct *work)
625{
626 struct amdgpu_device *adev = container_of(work, struct amdgpu_device, virt.vf2pf_work.work);
627 int ret;
628
629 ret = amdgpu_virt_read_pf2vf_data(adev);
630 if (ret)
631 goto out;
632 amdgpu_virt_write_vf2pf_data(adev);
633
634out:
635 schedule_delayed_work(&(adev->virt.vf2pf_work), adev->virt.vf2pf_update_interval_ms);
636}
637
638void amdgpu_virt_fini_data_exchange(struct amdgpu_device *adev)
639{
640 if (adev->virt.vf2pf_update_interval_ms != 0) {
641 DRM_INFO("clean up the vf2pf work item\n");
642 cancel_delayed_work_sync(&adev->virt.vf2pf_work);
643 adev->virt.vf2pf_update_interval_ms = 0;
644 }
645}
646
647void amdgpu_virt_init_data_exchange(struct amdgpu_device *adev)
648{
649 adev->virt.fw_reserve.p_pf2vf = NULL;
650 adev->virt.fw_reserve.p_vf2pf = NULL;
651 adev->virt.vf2pf_update_interval_ms = 0;
652
653 if (adev->mman.fw_vram_usage_va && adev->mman.drv_vram_usage_va) {
654 DRM_WARN("Currently fw_vram and drv_vram should not have values at the same time!");
655 } else if (adev->mman.fw_vram_usage_va || adev->mman.drv_vram_usage_va) {
656 /* go through this logic in ip_init and reset to init workqueue*/
657 amdgpu_virt_exchange_data(adev);
658
659 INIT_DELAYED_WORK(&adev->virt.vf2pf_work, amdgpu_virt_update_vf2pf_work_item);
660 schedule_delayed_work(&(adev->virt.vf2pf_work), msecs_to_jiffies(adev->virt.vf2pf_update_interval_ms));
661 } else if (adev->bios != NULL) {
662 /* got through this logic in early init stage to get necessary flags, e.g. rlcg_acc related*/
663 adev->virt.fw_reserve.p_pf2vf =
664 (struct amd_sriov_msg_pf2vf_info_header *)
665 (adev->bios + (AMD_SRIOV_MSG_PF2VF_OFFSET_KB << 10));
666
667 amdgpu_virt_read_pf2vf_data(adev);
668 }
669}
670
671
672void amdgpu_virt_exchange_data(struct amdgpu_device *adev)
673{
674 uint64_t bp_block_offset = 0;
675 uint32_t bp_block_size = 0;
676 struct amd_sriov_msg_pf2vf_info *pf2vf_v2 = NULL;
677
678 if (adev->mman.fw_vram_usage_va || adev->mman.drv_vram_usage_va) {
679 if (adev->mman.fw_vram_usage_va) {
680 adev->virt.fw_reserve.p_pf2vf =
681 (struct amd_sriov_msg_pf2vf_info_header *)
682 (adev->mman.fw_vram_usage_va + (AMD_SRIOV_MSG_PF2VF_OFFSET_KB << 10));
683 adev->virt.fw_reserve.p_vf2pf =
684 (struct amd_sriov_msg_vf2pf_info_header *)
685 (adev->mman.fw_vram_usage_va + (AMD_SRIOV_MSG_VF2PF_OFFSET_KB << 10));
686 } else if (adev->mman.drv_vram_usage_va) {
687 adev->virt.fw_reserve.p_pf2vf =
688 (struct amd_sriov_msg_pf2vf_info_header *)
689 (adev->mman.drv_vram_usage_va + (AMD_SRIOV_MSG_PF2VF_OFFSET_KB << 10));
690 adev->virt.fw_reserve.p_vf2pf =
691 (struct amd_sriov_msg_vf2pf_info_header *)
692 (adev->mman.drv_vram_usage_va + (AMD_SRIOV_MSG_VF2PF_OFFSET_KB << 10));
693 }
694
695 amdgpu_virt_read_pf2vf_data(adev);
696 amdgpu_virt_write_vf2pf_data(adev);
697
698 /* bad page handling for version 2 */
699 if (adev->virt.fw_reserve.p_pf2vf->version == 2) {
700 pf2vf_v2 = (struct amd_sriov_msg_pf2vf_info *)adev->virt.fw_reserve.p_pf2vf;
701
702 bp_block_offset = ((uint64_t)pf2vf_v2->bp_block_offset_low & 0xFFFFFFFF) |
703 ((((uint64_t)pf2vf_v2->bp_block_offset_high) << 32) & 0xFFFFFFFF00000000);
704 bp_block_size = pf2vf_v2->bp_block_size;
705
706 if (bp_block_size && !adev->virt.ras_init_done)
707 amdgpu_virt_init_ras_err_handler_data(adev);
708
709 if (adev->virt.ras_init_done)
710 amdgpu_virt_add_bad_page(adev, bp_block_offset, bp_block_size);
711 }
712 }
713}
714
715void amdgpu_detect_virtualization(struct amdgpu_device *adev)
716{
717 uint32_t reg;
718
719 switch (adev->asic_type) {
720 case CHIP_TONGA:
721 case CHIP_FIJI:
722 reg = RREG32(mmBIF_IOV_FUNC_IDENTIFIER);
723 break;
724 case CHIP_VEGA10:
725 case CHIP_VEGA20:
726 case CHIP_NAVI10:
727 case CHIP_NAVI12:
728 case CHIP_SIENNA_CICHLID:
729 case CHIP_ARCTURUS:
730 case CHIP_ALDEBARAN:
731 case CHIP_IP_DISCOVERY:
732 reg = RREG32(mmRCC_IOV_FUNC_IDENTIFIER);
733 break;
734 default: /* other chip doesn't support SRIOV */
735 reg = 0;
736 break;
737 }
738
739 if (reg & 1)
740 adev->virt.caps |= AMDGPU_SRIOV_CAPS_IS_VF;
741
742 if (reg & 0x80000000)
743 adev->virt.caps |= AMDGPU_SRIOV_CAPS_ENABLE_IOV;
744
745 if (!reg) {
746 /* passthrough mode exclus sriov mod */
747 if (is_virtual_machine() && !xen_initial_domain())
748 adev->virt.caps |= AMDGPU_PASSTHROUGH_MODE;
749 }
750
751 if (amdgpu_sriov_vf(adev) && adev->asic_type == CHIP_SIENNA_CICHLID)
752 /* VF MMIO access (except mailbox range) from CPU
753 * will be blocked during sriov runtime
754 */
755 adev->virt.caps |= AMDGPU_VF_MMIO_ACCESS_PROTECT;
756
757 /* we have the ability to check now */
758 if (amdgpu_sriov_vf(adev)) {
759 switch (adev->asic_type) {
760 case CHIP_TONGA:
761 case CHIP_FIJI:
762 vi_set_virt_ops(adev);
763 break;
764 case CHIP_VEGA10:
765 soc15_set_virt_ops(adev);
766#ifdef CONFIG_X86
767 /* not send GPU_INIT_DATA with MS_HYPERV*/
768 if (!hypervisor_is_type(X86_HYPER_MS_HYPERV))
769#endif
770 /* send a dummy GPU_INIT_DATA request to host on vega10 */
771 amdgpu_virt_request_init_data(adev);
772 break;
773 case CHIP_VEGA20:
774 case CHIP_ARCTURUS:
775 case CHIP_ALDEBARAN:
776 soc15_set_virt_ops(adev);
777 break;
778 case CHIP_NAVI10:
779 case CHIP_NAVI12:
780 case CHIP_SIENNA_CICHLID:
781 case CHIP_IP_DISCOVERY:
782 nv_set_virt_ops(adev);
783 /* try send GPU_INIT_DATA request to host */
784 amdgpu_virt_request_init_data(adev);
785 break;
786 default: /* other chip doesn't support SRIOV */
787 DRM_ERROR("Unknown asic type: %d!\n", adev->asic_type);
788 break;
789 }
790 }
791}
792
793static bool amdgpu_virt_access_debugfs_is_mmio(struct amdgpu_device *adev)
794{
795 return amdgpu_sriov_is_debug(adev) ? true : false;
796}
797
798static bool amdgpu_virt_access_debugfs_is_kiq(struct amdgpu_device *adev)
799{
800 return amdgpu_sriov_is_normal(adev) ? true : false;
801}
802
803int amdgpu_virt_enable_access_debugfs(struct amdgpu_device *adev)
804{
805 if (!amdgpu_sriov_vf(adev) ||
806 amdgpu_virt_access_debugfs_is_kiq(adev))
807 return 0;
808
809 if (amdgpu_virt_access_debugfs_is_mmio(adev))
810 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
811 else
812 return -EPERM;
813
814 return 0;
815}
816
817void amdgpu_virt_disable_access_debugfs(struct amdgpu_device *adev)
818{
819 if (amdgpu_sriov_vf(adev))
820 adev->virt.caps |= AMDGPU_SRIOV_CAPS_RUNTIME;
821}
822
823enum amdgpu_sriov_vf_mode amdgpu_virt_get_sriov_vf_mode(struct amdgpu_device *adev)
824{
825 enum amdgpu_sriov_vf_mode mode;
826
827 if (amdgpu_sriov_vf(adev)) {
828 if (amdgpu_sriov_is_pp_one_vf(adev))
829 mode = SRIOV_VF_MODE_ONE_VF;
830 else
831 mode = SRIOV_VF_MODE_MULTI_VF;
832 } else {
833 mode = SRIOV_VF_MODE_BARE_METAL;
834 }
835
836 return mode;
837}
838
839void amdgpu_virt_post_reset(struct amdgpu_device *adev)
840{
841 if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(11, 0, 3)) {
842 /* force set to GFXOFF state after reset,
843 * to avoid some invalid operation before GC enable
844 */
845 adev->gfx.is_poweron = false;
846 }
847}
848
849bool amdgpu_virt_fw_load_skip_check(struct amdgpu_device *adev, uint32_t ucode_id)
850{
851 switch (amdgpu_ip_version(adev, MP0_HWIP, 0)) {
852 case IP_VERSION(13, 0, 0):
853 /* no vf autoload, white list */
854 if (ucode_id == AMDGPU_UCODE_ID_VCN1 ||
855 ucode_id == AMDGPU_UCODE_ID_VCN)
856 return false;
857 else
858 return true;
859 case IP_VERSION(11, 0, 9):
860 case IP_VERSION(11, 0, 7):
861 /* black list for CHIP_NAVI12 and CHIP_SIENNA_CICHLID */
862 if (ucode_id == AMDGPU_UCODE_ID_RLC_G
863 || ucode_id == AMDGPU_UCODE_ID_RLC_RESTORE_LIST_CNTL
864 || ucode_id == AMDGPU_UCODE_ID_RLC_RESTORE_LIST_GPM_MEM
865 || ucode_id == AMDGPU_UCODE_ID_RLC_RESTORE_LIST_SRM_MEM
866 || ucode_id == AMDGPU_UCODE_ID_SMC)
867 return true;
868 else
869 return false;
870 case IP_VERSION(13, 0, 10):
871 /* white list */
872 if (ucode_id == AMDGPU_UCODE_ID_CAP
873 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_PFP
874 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_ME
875 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_MEC
876 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_PFP_P0_STACK
877 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_PFP_P1_STACK
878 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_ME_P0_STACK
879 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_ME_P1_STACK
880 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_MEC_P0_STACK
881 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_MEC_P1_STACK
882 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_MEC_P2_STACK
883 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_MEC_P3_STACK
884 || ucode_id == AMDGPU_UCODE_ID_CP_MES
885 || ucode_id == AMDGPU_UCODE_ID_CP_MES_DATA
886 || ucode_id == AMDGPU_UCODE_ID_CP_MES1
887 || ucode_id == AMDGPU_UCODE_ID_CP_MES1_DATA
888 || ucode_id == AMDGPU_UCODE_ID_VCN1
889 || ucode_id == AMDGPU_UCODE_ID_VCN)
890 return false;
891 else
892 return true;
893 default:
894 /* lagacy black list */
895 if (ucode_id == AMDGPU_UCODE_ID_SDMA0
896 || ucode_id == AMDGPU_UCODE_ID_SDMA1
897 || ucode_id == AMDGPU_UCODE_ID_SDMA2
898 || ucode_id == AMDGPU_UCODE_ID_SDMA3
899 || ucode_id == AMDGPU_UCODE_ID_SDMA4
900 || ucode_id == AMDGPU_UCODE_ID_SDMA5
901 || ucode_id == AMDGPU_UCODE_ID_SDMA6
902 || ucode_id == AMDGPU_UCODE_ID_SDMA7
903 || ucode_id == AMDGPU_UCODE_ID_RLC_G
904 || ucode_id == AMDGPU_UCODE_ID_RLC_RESTORE_LIST_CNTL
905 || ucode_id == AMDGPU_UCODE_ID_RLC_RESTORE_LIST_GPM_MEM
906 || ucode_id == AMDGPU_UCODE_ID_RLC_RESTORE_LIST_SRM_MEM
907 || ucode_id == AMDGPU_UCODE_ID_SMC)
908 return true;
909 else
910 return false;
911 }
912}
913
914void amdgpu_virt_update_sriov_video_codec(struct amdgpu_device *adev,
915 struct amdgpu_video_codec_info *encode, uint32_t encode_array_size,
916 struct amdgpu_video_codec_info *decode, uint32_t decode_array_size)
917{
918 uint32_t i;
919
920 if (!adev->virt.is_mm_bw_enabled)
921 return;
922
923 if (encode) {
924 for (i = 0; i < encode_array_size; i++) {
925 encode[i].max_width = adev->virt.encode_max_dimension_pixels;
926 encode[i].max_pixels_per_frame = adev->virt.encode_max_frame_pixels;
927 if (encode[i].max_width > 0)
928 encode[i].max_height = encode[i].max_pixels_per_frame / encode[i].max_width;
929 else
930 encode[i].max_height = 0;
931 }
932 }
933
934 if (decode) {
935 for (i = 0; i < decode_array_size; i++) {
936 decode[i].max_width = adev->virt.decode_max_dimension_pixels;
937 decode[i].max_pixels_per_frame = adev->virt.decode_max_frame_pixels;
938 if (decode[i].max_width > 0)
939 decode[i].max_height = decode[i].max_pixels_per_frame / decode[i].max_width;
940 else
941 decode[i].max_height = 0;
942 }
943 }
944}
945
946bool amdgpu_virt_get_rlcg_reg_access_flag(struct amdgpu_device *adev,
947 u32 acc_flags, u32 hwip,
948 bool write, u32 *rlcg_flag)
949{
950 bool ret = false;
951
952 switch (hwip) {
953 case GC_HWIP:
954 if (amdgpu_sriov_reg_indirect_gc(adev)) {
955 *rlcg_flag =
956 write ? AMDGPU_RLCG_GC_WRITE : AMDGPU_RLCG_GC_READ;
957 ret = true;
958 /* only in new version, AMDGPU_REGS_NO_KIQ and
959 * AMDGPU_REGS_RLC are enabled simultaneously */
960 } else if ((acc_flags & AMDGPU_REGS_RLC) &&
961 !(acc_flags & AMDGPU_REGS_NO_KIQ) && write) {
962 *rlcg_flag = AMDGPU_RLCG_GC_WRITE_LEGACY;
963 ret = true;
964 }
965 break;
966 case MMHUB_HWIP:
967 if (amdgpu_sriov_reg_indirect_mmhub(adev) &&
968 (acc_flags & AMDGPU_REGS_RLC) && write) {
969 *rlcg_flag = AMDGPU_RLCG_MMHUB_WRITE;
970 ret = true;
971 }
972 break;
973 default:
974 break;
975 }
976 return ret;
977}
978
979u32 amdgpu_virt_rlcg_reg_rw(struct amdgpu_device *adev, u32 offset, u32 v, u32 flag, u32 xcc_id)
980{
981 struct amdgpu_rlcg_reg_access_ctrl *reg_access_ctrl;
982 uint32_t timeout = 50000;
983 uint32_t i, tmp;
984 uint32_t ret = 0;
985 void *scratch_reg0;
986 void *scratch_reg1;
987 void *scratch_reg2;
988 void *scratch_reg3;
989 void *spare_int;
990
991 if (!adev->gfx.rlc.rlcg_reg_access_supported) {
992 dev_err(adev->dev,
993 "indirect registers access through rlcg is not available\n");
994 return 0;
995 }
996
997 if (adev->gfx.xcc_mask && (((1 << xcc_id) & adev->gfx.xcc_mask) == 0)) {
998 dev_err(adev->dev, "invalid xcc\n");
999 return 0;
1000 }
1001
1002 reg_access_ctrl = &adev->gfx.rlc.reg_access_ctrl[xcc_id];
1003 scratch_reg0 = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->scratch_reg0;
1004 scratch_reg1 = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->scratch_reg1;
1005 scratch_reg2 = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->scratch_reg2;
1006 scratch_reg3 = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->scratch_reg3;
1007 if (reg_access_ctrl->spare_int)
1008 spare_int = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->spare_int;
1009
1010 if (offset == reg_access_ctrl->grbm_cntl) {
1011 /* if the target reg offset is grbm_cntl, write to scratch_reg2 */
1012 writel(v, scratch_reg2);
1013 if (flag == AMDGPU_RLCG_GC_WRITE_LEGACY)
1014 writel(v, ((void __iomem *)adev->rmmio) + (offset * 4));
1015 } else if (offset == reg_access_ctrl->grbm_idx) {
1016 /* if the target reg offset is grbm_idx, write to scratch_reg3 */
1017 writel(v, scratch_reg3);
1018 if (flag == AMDGPU_RLCG_GC_WRITE_LEGACY)
1019 writel(v, ((void __iomem *)adev->rmmio) + (offset * 4));
1020 } else {
1021 /*
1022 * SCRATCH_REG0 = read/write value
1023 * SCRATCH_REG1[30:28] = command
1024 * SCRATCH_REG1[19:0] = address in dword
1025 * SCRATCH_REG1[26:24] = Error reporting
1026 */
1027 writel(v, scratch_reg0);
1028 writel((offset | flag), scratch_reg1);
1029 if (reg_access_ctrl->spare_int)
1030 writel(1, spare_int);
1031
1032 for (i = 0; i < timeout; i++) {
1033 tmp = readl(scratch_reg1);
1034 if (!(tmp & AMDGPU_RLCG_SCRATCH1_ADDRESS_MASK))
1035 break;
1036 udelay(10);
1037 }
1038
1039 if (i >= timeout) {
1040 if (amdgpu_sriov_rlcg_error_report_enabled(adev)) {
1041 if (tmp & AMDGPU_RLCG_VFGATE_DISABLED) {
1042 dev_err(adev->dev,
1043 "vfgate is disabled, rlcg failed to program reg: 0x%05x\n", offset);
1044 } else if (tmp & AMDGPU_RLCG_WRONG_OPERATION_TYPE) {
1045 dev_err(adev->dev,
1046 "wrong operation type, rlcg failed to program reg: 0x%05x\n", offset);
1047 } else if (tmp & AMDGPU_RLCG_REG_NOT_IN_RANGE) {
1048 dev_err(adev->dev,
1049 "register is not in range, rlcg failed to program reg: 0x%05x\n", offset);
1050 } else {
1051 dev_err(adev->dev,
1052 "unknown error type, rlcg failed to program reg: 0x%05x\n", offset);
1053 }
1054 } else {
1055 dev_err(adev->dev,
1056 "timeout: rlcg faled to program reg: 0x%05x\n", offset);
1057 }
1058 }
1059 }
1060
1061 ret = readl(scratch_reg0);
1062 return ret;
1063}
1064
1065void amdgpu_sriov_wreg(struct amdgpu_device *adev,
1066 u32 offset, u32 value,
1067 u32 acc_flags, u32 hwip, u32 xcc_id)
1068{
1069 u32 rlcg_flag;
1070
1071 if (!amdgpu_sriov_runtime(adev) &&
1072 amdgpu_virt_get_rlcg_reg_access_flag(adev, acc_flags, hwip, true, &rlcg_flag)) {
1073 amdgpu_virt_rlcg_reg_rw(adev, offset, value, rlcg_flag, xcc_id);
1074 return;
1075 }
1076
1077 if (acc_flags & AMDGPU_REGS_NO_KIQ)
1078 WREG32_NO_KIQ(offset, value);
1079 else
1080 WREG32(offset, value);
1081}
1082
1083u32 amdgpu_sriov_rreg(struct amdgpu_device *adev,
1084 u32 offset, u32 acc_flags, u32 hwip, u32 xcc_id)
1085{
1086 u32 rlcg_flag;
1087
1088 if (!amdgpu_sriov_runtime(adev) &&
1089 amdgpu_virt_get_rlcg_reg_access_flag(adev, acc_flags, hwip, false, &rlcg_flag))
1090 return amdgpu_virt_rlcg_reg_rw(adev, offset, 0, rlcg_flag, xcc_id);
1091
1092 if (acc_flags & AMDGPU_REGS_NO_KIQ)
1093 return RREG32_NO_KIQ(offset);
1094 else
1095 return RREG32(offset);
1096}
1097
1098bool amdgpu_sriov_xnack_support(struct amdgpu_device *adev)
1099{
1100 bool xnack_mode = true;
1101
1102 if (amdgpu_sriov_vf(adev) &&
1103 amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 2))
1104 xnack_mode = false;
1105
1106 return xnack_mode;
1107}