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
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}
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 "amdgpu.h"
25#define MAX_KIQ_REG_WAIT 5000 /* in usecs, 5ms */
26#define MAX_KIQ_REG_BAILOUT_INTERVAL 5 /* in msecs, 5ms */
27#define MAX_KIQ_REG_TRY 20
28
29uint64_t amdgpu_csa_vaddr(struct amdgpu_device *adev)
30{
31 uint64_t addr = adev->vm_manager.max_pfn << AMDGPU_GPU_PAGE_SHIFT;
32
33 addr -= AMDGPU_VA_RESERVED_SIZE;
34
35 if (addr >= AMDGPU_VA_HOLE_START)
36 addr |= AMDGPU_VA_HOLE_END;
37
38 return addr;
39}
40
41bool amdgpu_virt_mmio_blocked(struct amdgpu_device *adev)
42{
43 /* By now all MMIO pages except mailbox are blocked */
44 /* if blocking is enabled in hypervisor. Choose the */
45 /* SCRATCH_REG0 to test. */
46 return RREG32_NO_KIQ(0xc040) == 0xffffffff;
47}
48
49int amdgpu_allocate_static_csa(struct amdgpu_device *adev)
50{
51 int r;
52 void *ptr;
53
54 r = amdgpu_bo_create_kernel(adev, AMDGPU_CSA_SIZE, PAGE_SIZE,
55 AMDGPU_GEM_DOMAIN_VRAM, &adev->virt.csa_obj,
56 &adev->virt.csa_vmid0_addr, &ptr);
57 if (r)
58 return r;
59
60 memset(ptr, 0, AMDGPU_CSA_SIZE);
61 return 0;
62}
63
64void amdgpu_free_static_csa(struct amdgpu_device *adev) {
65 amdgpu_bo_free_kernel(&adev->virt.csa_obj,
66 &adev->virt.csa_vmid0_addr,
67 NULL);
68}
69
70/*
71 * amdgpu_map_static_csa should be called during amdgpu_vm_init
72 * it maps virtual address amdgpu_csa_vaddr() to this VM, and each command
73 * submission of GFX should use this virtual address within META_DATA init
74 * package to support SRIOV gfx preemption.
75 */
76int amdgpu_map_static_csa(struct amdgpu_device *adev, struct amdgpu_vm *vm,
77 struct amdgpu_bo_va **bo_va)
78{
79 uint64_t csa_addr = amdgpu_csa_vaddr(adev) & AMDGPU_VA_HOLE_MASK;
80 struct ww_acquire_ctx ticket;
81 struct list_head list;
82 struct amdgpu_bo_list_entry pd;
83 struct ttm_validate_buffer csa_tv;
84 int r;
85
86 INIT_LIST_HEAD(&list);
87 INIT_LIST_HEAD(&csa_tv.head);
88 csa_tv.bo = &adev->virt.csa_obj->tbo;
89 csa_tv.shared = true;
90
91 list_add(&csa_tv.head, &list);
92 amdgpu_vm_get_pd_bo(vm, &list, &pd);
93
94 r = ttm_eu_reserve_buffers(&ticket, &list, true, NULL);
95 if (r) {
96 DRM_ERROR("failed to reserve CSA,PD BOs: err=%d\n", r);
97 return r;
98 }
99
100 *bo_va = amdgpu_vm_bo_add(adev, vm, adev->virt.csa_obj);
101 if (!*bo_va) {
102 ttm_eu_backoff_reservation(&ticket, &list);
103 DRM_ERROR("failed to create bo_va for static CSA\n");
104 return -ENOMEM;
105 }
106
107 r = amdgpu_vm_alloc_pts(adev, (*bo_va)->base.vm, csa_addr,
108 AMDGPU_CSA_SIZE);
109 if (r) {
110 DRM_ERROR("failed to allocate pts for static CSA, err=%d\n", r);
111 amdgpu_vm_bo_rmv(adev, *bo_va);
112 ttm_eu_backoff_reservation(&ticket, &list);
113 return r;
114 }
115
116 r = amdgpu_vm_bo_map(adev, *bo_va, csa_addr, 0, AMDGPU_CSA_SIZE,
117 AMDGPU_PTE_READABLE | AMDGPU_PTE_WRITEABLE |
118 AMDGPU_PTE_EXECUTABLE);
119
120 if (r) {
121 DRM_ERROR("failed to do bo_map on static CSA, err=%d\n", r);
122 amdgpu_vm_bo_rmv(adev, *bo_va);
123 ttm_eu_backoff_reservation(&ticket, &list);
124 return r;
125 }
126
127 ttm_eu_backoff_reservation(&ticket, &list);
128 return 0;
129}
130
131void amdgpu_virt_init_setting(struct amdgpu_device *adev)
132{
133 /* enable virtual display */
134 adev->mode_info.num_crtc = 1;
135 adev->enable_virtual_display = true;
136 adev->cg_flags = 0;
137 adev->pg_flags = 0;
138}
139
140uint32_t amdgpu_virt_kiq_rreg(struct amdgpu_device *adev, uint32_t reg)
141{
142 signed long r, cnt = 0;
143 unsigned long flags;
144 uint32_t seq;
145 struct amdgpu_kiq *kiq = &adev->gfx.kiq;
146 struct amdgpu_ring *ring = &kiq->ring;
147
148 BUG_ON(!ring->funcs->emit_rreg);
149
150 spin_lock_irqsave(&kiq->ring_lock, flags);
151 amdgpu_ring_alloc(ring, 32);
152 amdgpu_ring_emit_rreg(ring, reg);
153 amdgpu_fence_emit_polling(ring, &seq);
154 amdgpu_ring_commit(ring);
155 spin_unlock_irqrestore(&kiq->ring_lock, flags);
156
157 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
158
159 /* don't wait anymore for gpu reset case because this way may
160 * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
161 * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
162 * never return if we keep waiting in virt_kiq_rreg, which cause
163 * gpu_recover() hang there.
164 *
165 * also don't wait anymore for IRQ context
166 * */
167 if (r < 1 && (adev->in_gpu_reset || in_interrupt()))
168 goto failed_kiq_read;
169
170 if (in_interrupt())
171 might_sleep();
172
173 while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
174 msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
175 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
176 }
177
178 if (cnt > MAX_KIQ_REG_TRY)
179 goto failed_kiq_read;
180
181 return adev->wb.wb[adev->virt.reg_val_offs];
182
183failed_kiq_read:
184 pr_err("failed to read reg:%x\n", reg);
185 return ~0;
186}
187
188void amdgpu_virt_kiq_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
189{
190 signed long r, cnt = 0;
191 unsigned long flags;
192 uint32_t seq;
193 struct amdgpu_kiq *kiq = &adev->gfx.kiq;
194 struct amdgpu_ring *ring = &kiq->ring;
195
196 BUG_ON(!ring->funcs->emit_wreg);
197
198 spin_lock_irqsave(&kiq->ring_lock, flags);
199 amdgpu_ring_alloc(ring, 32);
200 amdgpu_ring_emit_wreg(ring, reg, v);
201 amdgpu_fence_emit_polling(ring, &seq);
202 amdgpu_ring_commit(ring);
203 spin_unlock_irqrestore(&kiq->ring_lock, flags);
204
205 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
206
207 /* don't wait anymore for gpu reset case because this way may
208 * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
209 * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
210 * never return if we keep waiting in virt_kiq_rreg, which cause
211 * gpu_recover() hang there.
212 *
213 * also don't wait anymore for IRQ context
214 * */
215 if (r < 1 && (adev->in_gpu_reset || in_interrupt()))
216 goto failed_kiq_write;
217
218 if (in_interrupt())
219 might_sleep();
220
221 while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
222
223 msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
224 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
225 }
226
227 if (cnt > MAX_KIQ_REG_TRY)
228 goto failed_kiq_write;
229
230 return;
231
232failed_kiq_write:
233 pr_err("failed to write reg:%x\n", reg);
234}
235
236/**
237 * amdgpu_virt_request_full_gpu() - request full gpu access
238 * @amdgpu: amdgpu device.
239 * @init: is driver init time.
240 * When start to init/fini driver, first need to request full gpu access.
241 * Return: Zero if request success, otherwise will return error.
242 */
243int amdgpu_virt_request_full_gpu(struct amdgpu_device *adev, bool init)
244{
245 struct amdgpu_virt *virt = &adev->virt;
246 int r;
247
248 if (virt->ops && virt->ops->req_full_gpu) {
249 r = virt->ops->req_full_gpu(adev, init);
250 if (r)
251 return r;
252
253 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
254 }
255
256 return 0;
257}
258
259/**
260 * amdgpu_virt_release_full_gpu() - release full gpu access
261 * @amdgpu: amdgpu device.
262 * @init: is driver init time.
263 * When finishing driver init/fini, need to release full gpu access.
264 * Return: Zero if release success, otherwise will returen error.
265 */
266int amdgpu_virt_release_full_gpu(struct amdgpu_device *adev, bool init)
267{
268 struct amdgpu_virt *virt = &adev->virt;
269 int r;
270
271 if (virt->ops && virt->ops->rel_full_gpu) {
272 r = virt->ops->rel_full_gpu(adev, init);
273 if (r)
274 return r;
275
276 adev->virt.caps |= AMDGPU_SRIOV_CAPS_RUNTIME;
277 }
278 return 0;
279}
280
281/**
282 * amdgpu_virt_reset_gpu() - reset gpu
283 * @amdgpu: amdgpu device.
284 * Send reset command to GPU hypervisor to reset GPU that VM is using
285 * Return: Zero if reset success, otherwise will return error.
286 */
287int amdgpu_virt_reset_gpu(struct amdgpu_device *adev)
288{
289 struct amdgpu_virt *virt = &adev->virt;
290 int r;
291
292 if (virt->ops && virt->ops->reset_gpu) {
293 r = virt->ops->reset_gpu(adev);
294 if (r)
295 return r;
296
297 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
298 }
299
300 return 0;
301}
302
303/**
304 * amdgpu_virt_wait_reset() - wait for reset gpu completed
305 * @amdgpu: amdgpu device.
306 * Wait for GPU reset completed.
307 * Return: Zero if reset success, otherwise will return error.
308 */
309int amdgpu_virt_wait_reset(struct amdgpu_device *adev)
310{
311 struct amdgpu_virt *virt = &adev->virt;
312
313 if (!virt->ops || !virt->ops->wait_reset)
314 return -EINVAL;
315
316 return virt->ops->wait_reset(adev);
317}
318
319/**
320 * amdgpu_virt_alloc_mm_table() - alloc memory for mm table
321 * @amdgpu: amdgpu device.
322 * MM table is used by UVD and VCE for its initialization
323 * Return: Zero if allocate success.
324 */
325int amdgpu_virt_alloc_mm_table(struct amdgpu_device *adev)
326{
327 int r;
328
329 if (!amdgpu_sriov_vf(adev) || adev->virt.mm_table.gpu_addr)
330 return 0;
331
332 r = amdgpu_bo_create_kernel(adev, PAGE_SIZE, PAGE_SIZE,
333 AMDGPU_GEM_DOMAIN_VRAM,
334 &adev->virt.mm_table.bo,
335 &adev->virt.mm_table.gpu_addr,
336 (void *)&adev->virt.mm_table.cpu_addr);
337 if (r) {
338 DRM_ERROR("failed to alloc mm table and error = %d.\n", r);
339 return r;
340 }
341
342 memset((void *)adev->virt.mm_table.cpu_addr, 0, PAGE_SIZE);
343 DRM_INFO("MM table gpu addr = 0x%llx, cpu addr = %p.\n",
344 adev->virt.mm_table.gpu_addr,
345 adev->virt.mm_table.cpu_addr);
346 return 0;
347}
348
349/**
350 * amdgpu_virt_free_mm_table() - free mm table memory
351 * @amdgpu: amdgpu device.
352 * Free MM table memory
353 */
354void amdgpu_virt_free_mm_table(struct amdgpu_device *adev)
355{
356 if (!amdgpu_sriov_vf(adev) || !adev->virt.mm_table.gpu_addr)
357 return;
358
359 amdgpu_bo_free_kernel(&adev->virt.mm_table.bo,
360 &adev->virt.mm_table.gpu_addr,
361 (void *)&adev->virt.mm_table.cpu_addr);
362 adev->virt.mm_table.gpu_addr = 0;
363}
364
365
366int amdgpu_virt_fw_reserve_get_checksum(void *obj,
367 unsigned long obj_size,
368 unsigned int key,
369 unsigned int chksum)
370{
371 unsigned int ret = key;
372 unsigned long i = 0;
373 unsigned char *pos;
374
375 pos = (char *)obj;
376 /* calculate checksum */
377 for (i = 0; i < obj_size; ++i)
378 ret += *(pos + i);
379 /* minus the chksum itself */
380 pos = (char *)&chksum;
381 for (i = 0; i < sizeof(chksum); ++i)
382 ret -= *(pos + i);
383 return ret;
384}
385
386void amdgpu_virt_init_data_exchange(struct amdgpu_device *adev)
387{
388 uint32_t pf2vf_size = 0;
389 uint32_t checksum = 0;
390 uint32_t checkval;
391 char *str;
392
393 adev->virt.fw_reserve.p_pf2vf = NULL;
394 adev->virt.fw_reserve.p_vf2pf = NULL;
395
396 if (adev->fw_vram_usage.va != NULL) {
397 adev->virt.fw_reserve.p_pf2vf =
398 (struct amdgim_pf2vf_info_header *)(
399 adev->fw_vram_usage.va + AMDGIM_DATAEXCHANGE_OFFSET);
400 AMDGPU_FW_VRAM_PF2VF_READ(adev, header.size, &pf2vf_size);
401 AMDGPU_FW_VRAM_PF2VF_READ(adev, checksum, &checksum);
402 AMDGPU_FW_VRAM_PF2VF_READ(adev, feature_flags, &adev->virt.gim_feature);
403
404 /* pf2vf message must be in 4K */
405 if (pf2vf_size > 0 && pf2vf_size < 4096) {
406 checkval = amdgpu_virt_fw_reserve_get_checksum(
407 adev->virt.fw_reserve.p_pf2vf, pf2vf_size,
408 adev->virt.fw_reserve.checksum_key, checksum);
409 if (checkval == checksum) {
410 adev->virt.fw_reserve.p_vf2pf =
411 ((void *)adev->virt.fw_reserve.p_pf2vf +
412 pf2vf_size);
413 memset((void *)adev->virt.fw_reserve.p_vf2pf, 0,
414 sizeof(amdgim_vf2pf_info));
415 AMDGPU_FW_VRAM_VF2PF_WRITE(adev, header.version,
416 AMDGPU_FW_VRAM_VF2PF_VER);
417 AMDGPU_FW_VRAM_VF2PF_WRITE(adev, header.size,
418 sizeof(amdgim_vf2pf_info));
419 AMDGPU_FW_VRAM_VF2PF_READ(adev, driver_version,
420 &str);
421#ifdef MODULE
422 if (THIS_MODULE->version != NULL)
423 strcpy(str, THIS_MODULE->version);
424 else
425#endif
426 strcpy(str, "N/A");
427 AMDGPU_FW_VRAM_VF2PF_WRITE(adev, driver_cert,
428 0);
429 AMDGPU_FW_VRAM_VF2PF_WRITE(adev, checksum,
430 amdgpu_virt_fw_reserve_get_checksum(
431 adev->virt.fw_reserve.p_vf2pf,
432 pf2vf_size,
433 adev->virt.fw_reserve.checksum_key, 0));
434 }
435 }
436 }
437}
438
439