drm/amdgpu: clear UVD VCPU buffer when err_event_athub generated
[linux-2.6-block.git] / drivers / gpu / drm / amd / amdgpu / amdgpu_device.c
CommitLineData
d38ceaf9
AD
1/*
2 * Copyright 2008 Advanced Micro Devices, Inc.
3 * Copyright 2008 Red Hat Inc.
4 * Copyright 2009 Jerome Glisse.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22 * OTHER DEALINGS IN THE SOFTWARE.
23 *
24 * Authors: Dave Airlie
25 * Alex Deucher
26 * Jerome Glisse
27 */
b1ddf548 28#include <linux/power_supply.h>
0875dc9e 29#include <linux/kthread.h>
fdf2f6c5 30#include <linux/module.h>
d38ceaf9
AD
31#include <linux/console.h>
32#include <linux/slab.h>
fdf2f6c5 33
4562236b 34#include <drm/drm_atomic_helper.h>
fcd70cd3 35#include <drm/drm_probe_helper.h>
d38ceaf9
AD
36#include <drm/amdgpu_drm.h>
37#include <linux/vgaarb.h>
38#include <linux/vga_switcheroo.h>
39#include <linux/efi.h>
40#include "amdgpu.h"
f4b373f4 41#include "amdgpu_trace.h"
d38ceaf9
AD
42#include "amdgpu_i2c.h"
43#include "atom.h"
44#include "amdgpu_atombios.h"
a5bde2f9 45#include "amdgpu_atomfirmware.h"
d0dd7f0c 46#include "amd_pcie.h"
33f34802
KW
47#ifdef CONFIG_DRM_AMDGPU_SI
48#include "si.h"
49#endif
a2e73f56
AD
50#ifdef CONFIG_DRM_AMDGPU_CIK
51#include "cik.h"
52#endif
aaa36a97 53#include "vi.h"
460826e6 54#include "soc15.h"
0a5b8c7b 55#include "nv.h"
d38ceaf9 56#include "bif/bif_4_1_d.h"
9accf2fd 57#include <linux/pci.h>
bec86378 58#include <linux/firmware.h>
89041940 59#include "amdgpu_vf_error.h"
d38ceaf9 60
ba997709 61#include "amdgpu_amdkfd.h"
d2f52ac8 62#include "amdgpu_pm.h"
d38ceaf9 63
5183411b 64#include "amdgpu_xgmi.h"
c030f2e4 65#include "amdgpu_ras.h"
9c7c85f7 66#include "amdgpu_pmu.h"
5183411b 67
d5ea093e
AG
68#include <linux/suspend.h>
69
e2a75f88 70MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin");
3f76dced 71MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin");
2d2e5e7e 72MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin");
ad5a67a7 73MODULE_FIRMWARE("amdgpu/picasso_gpu_info.bin");
54c4d17e 74MODULE_FIRMWARE("amdgpu/raven2_gpu_info.bin");
65e60f6e 75MODULE_FIRMWARE("amdgpu/arcturus_gpu_info.bin");
b51a26a0 76MODULE_FIRMWARE("amdgpu/renoir_gpu_info.bin");
23c6268e 77MODULE_FIRMWARE("amdgpu/navi10_gpu_info.bin");
ed42cfe1 78MODULE_FIRMWARE("amdgpu/navi14_gpu_info.bin");
42b325e5 79MODULE_FIRMWARE("amdgpu/navi12_gpu_info.bin");
e2a75f88 80
2dc80b00
S
81#define AMDGPU_RESUME_MS 2000
82
050091ab 83const char *amdgpu_asic_name[] = {
da69c161
KW
84 "TAHITI",
85 "PITCAIRN",
86 "VERDE",
87 "OLAND",
88 "HAINAN",
d38ceaf9
AD
89 "BONAIRE",
90 "KAVERI",
91 "KABINI",
92 "HAWAII",
93 "MULLINS",
94 "TOPAZ",
95 "TONGA",
48299f95 96 "FIJI",
d38ceaf9 97 "CARRIZO",
139f4917 98 "STONEY",
2cc0c0b5
FC
99 "POLARIS10",
100 "POLARIS11",
c4642a47 101 "POLARIS12",
48ff108d 102 "VEGAM",
d4196f01 103 "VEGA10",
8fab806a 104 "VEGA12",
956fcddc 105 "VEGA20",
2ca8a5d2 106 "RAVEN",
d6c3b24e 107 "ARCTURUS",
1eee4228 108 "RENOIR",
852a6626 109 "NAVI10",
87dbad02 110 "NAVI14",
9802f5d7 111 "NAVI12",
d38ceaf9
AD
112 "LAST",
113};
114
dcea6e65
KR
115/**
116 * DOC: pcie_replay_count
117 *
118 * The amdgpu driver provides a sysfs API for reporting the total number
119 * of PCIe replays (NAKs)
120 * The file pcie_replay_count is used for this and returns the total
121 * number of replays as a sum of the NAKs generated and NAKs received
122 */
123
124static ssize_t amdgpu_device_get_pcie_replay_count(struct device *dev,
125 struct device_attribute *attr, char *buf)
126{
127 struct drm_device *ddev = dev_get_drvdata(dev);
128 struct amdgpu_device *adev = ddev->dev_private;
129 uint64_t cnt = amdgpu_asic_get_pcie_replay_count(adev);
130
131 return snprintf(buf, PAGE_SIZE, "%llu\n", cnt);
132}
133
134static DEVICE_ATTR(pcie_replay_count, S_IRUGO,
135 amdgpu_device_get_pcie_replay_count, NULL);
136
5494d864
AD
137static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev);
138
e3ecdffa
AD
139/**
140 * amdgpu_device_is_px - Is the device is a dGPU with HG/PX power control
141 *
142 * @dev: drm_device pointer
143 *
144 * Returns true if the device is a dGPU with HG/PX power control,
145 * otherwise return false.
146 */
d38ceaf9
AD
147bool amdgpu_device_is_px(struct drm_device *dev)
148{
149 struct amdgpu_device *adev = dev->dev_private;
150
2f7d10b3 151 if (adev->flags & AMD_IS_PX)
d38ceaf9
AD
152 return true;
153 return false;
154}
155
e35e2b11
TY
156/**
157 * VRAM access helper functions.
158 *
159 * amdgpu_device_vram_access - read/write a buffer in vram
160 *
161 * @adev: amdgpu_device pointer
162 * @pos: offset of the buffer in vram
163 * @buf: virtual address of the buffer in system memory
164 * @size: read/write size, sizeof(@buf) must > @size
165 * @write: true - write to vram, otherwise - read from vram
166 */
167void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos,
168 uint32_t *buf, size_t size, bool write)
169{
170 uint64_t last;
171 unsigned long flags;
172
173 last = size - 4;
174 for (last += pos; pos <= last; pos += 4) {
175 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
176 WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)pos) | 0x80000000);
177 WREG32_NO_KIQ(mmMM_INDEX_HI, pos >> 31);
178 if (write)
179 WREG32_NO_KIQ(mmMM_DATA, *buf++);
180 else
181 *buf++ = RREG32_NO_KIQ(mmMM_DATA);
182 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
183 }
184}
185
d38ceaf9
AD
186/*
187 * MMIO register access helper functions.
188 */
e3ecdffa
AD
189/**
190 * amdgpu_mm_rreg - read a memory mapped IO register
191 *
192 * @adev: amdgpu_device pointer
193 * @reg: dword aligned register offset
194 * @acc_flags: access flags which require special behavior
195 *
196 * Returns the 32 bit value from the offset specified.
197 */
d38ceaf9 198uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg,
15d72fd7 199 uint32_t acc_flags)
d38ceaf9 200{
f4b373f4
TSD
201 uint32_t ret;
202
43ca8efa 203 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
bc992ba5 204 return amdgpu_virt_kiq_rreg(adev, reg);
bc992ba5 205
15d72fd7 206 if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
f4b373f4 207 ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
d38ceaf9
AD
208 else {
209 unsigned long flags;
d38ceaf9
AD
210
211 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
212 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
213 ret = readl(((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
214 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
d38ceaf9 215 }
f4b373f4
TSD
216 trace_amdgpu_mm_rreg(adev->pdev->device, reg, ret);
217 return ret;
d38ceaf9
AD
218}
219
421a2a30
ML
220/*
221 * MMIO register read with bytes helper functions
222 * @offset:bytes offset from MMIO start
223 *
224*/
225
e3ecdffa
AD
226/**
227 * amdgpu_mm_rreg8 - read a memory mapped IO register
228 *
229 * @adev: amdgpu_device pointer
230 * @offset: byte aligned register offset
231 *
232 * Returns the 8 bit value from the offset specified.
233 */
421a2a30
ML
234uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset) {
235 if (offset < adev->rmmio_size)
236 return (readb(adev->rmmio + offset));
237 BUG();
238}
239
240/*
241 * MMIO register write with bytes helper functions
242 * @offset:bytes offset from MMIO start
243 * @value: the value want to be written to the register
244 *
245*/
e3ecdffa
AD
246/**
247 * amdgpu_mm_wreg8 - read a memory mapped IO register
248 *
249 * @adev: amdgpu_device pointer
250 * @offset: byte aligned register offset
251 * @value: 8 bit value to write
252 *
253 * Writes the value specified to the offset specified.
254 */
421a2a30
ML
255void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value) {
256 if (offset < adev->rmmio_size)
257 writeb(value, adev->rmmio + offset);
258 else
259 BUG();
260}
261
e3ecdffa
AD
262/**
263 * amdgpu_mm_wreg - write to a memory mapped IO register
264 *
265 * @adev: amdgpu_device pointer
266 * @reg: dword aligned register offset
267 * @v: 32 bit value to write to the register
268 * @acc_flags: access flags which require special behavior
269 *
270 * Writes the value specified to the offset specified.
271 */
d38ceaf9 272void amdgpu_mm_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v,
15d72fd7 273 uint32_t acc_flags)
d38ceaf9 274{
f4b373f4 275 trace_amdgpu_mm_wreg(adev->pdev->device, reg, v);
4e99a44e 276
47ed4e1c
KW
277 if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
278 adev->last_mm_index = v;
279 }
280
43ca8efa 281 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
bc992ba5 282 return amdgpu_virt_kiq_wreg(adev, reg, v);
bc992ba5 283
15d72fd7 284 if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
d38ceaf9
AD
285 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
286 else {
287 unsigned long flags;
288
289 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
290 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
291 writel(v, ((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
292 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
293 }
47ed4e1c
KW
294
295 if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
296 udelay(500);
297 }
d38ceaf9
AD
298}
299
e3ecdffa
AD
300/**
301 * amdgpu_io_rreg - read an IO register
302 *
303 * @adev: amdgpu_device pointer
304 * @reg: dword aligned register offset
305 *
306 * Returns the 32 bit value from the offset specified.
307 */
d38ceaf9
AD
308u32 amdgpu_io_rreg(struct amdgpu_device *adev, u32 reg)
309{
310 if ((reg * 4) < adev->rio_mem_size)
311 return ioread32(adev->rio_mem + (reg * 4));
312 else {
313 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
314 return ioread32(adev->rio_mem + (mmMM_DATA * 4));
315 }
316}
317
e3ecdffa
AD
318/**
319 * amdgpu_io_wreg - write to an IO register
320 *
321 * @adev: amdgpu_device pointer
322 * @reg: dword aligned register offset
323 * @v: 32 bit value to write to the register
324 *
325 * Writes the value specified to the offset specified.
326 */
d38ceaf9
AD
327void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32 v)
328{
47ed4e1c
KW
329 if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
330 adev->last_mm_index = v;
331 }
d38ceaf9
AD
332
333 if ((reg * 4) < adev->rio_mem_size)
334 iowrite32(v, adev->rio_mem + (reg * 4));
335 else {
336 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
337 iowrite32(v, adev->rio_mem + (mmMM_DATA * 4));
338 }
47ed4e1c
KW
339
340 if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
341 udelay(500);
342 }
d38ceaf9
AD
343}
344
345/**
346 * amdgpu_mm_rdoorbell - read a doorbell dword
347 *
348 * @adev: amdgpu_device pointer
349 * @index: doorbell index
350 *
351 * Returns the value in the doorbell aperture at the
352 * requested doorbell index (CIK).
353 */
354u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
355{
356 if (index < adev->doorbell.num_doorbells) {
357 return readl(adev->doorbell.ptr + index);
358 } else {
359 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
360 return 0;
361 }
362}
363
364/**
365 * amdgpu_mm_wdoorbell - write a doorbell dword
366 *
367 * @adev: amdgpu_device pointer
368 * @index: doorbell index
369 * @v: value to write
370 *
371 * Writes @v to the doorbell aperture at the
372 * requested doorbell index (CIK).
373 */
374void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v)
375{
376 if (index < adev->doorbell.num_doorbells) {
377 writel(v, adev->doorbell.ptr + index);
378 } else {
379 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
380 }
381}
382
832be404
KW
383/**
384 * amdgpu_mm_rdoorbell64 - read a doorbell Qword
385 *
386 * @adev: amdgpu_device pointer
387 * @index: doorbell index
388 *
389 * Returns the value in the doorbell aperture at the
390 * requested doorbell index (VEGA10+).
391 */
392u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index)
393{
394 if (index < adev->doorbell.num_doorbells) {
395 return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index));
396 } else {
397 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
398 return 0;
399 }
400}
401
402/**
403 * amdgpu_mm_wdoorbell64 - write a doorbell Qword
404 *
405 * @adev: amdgpu_device pointer
406 * @index: doorbell index
407 * @v: value to write
408 *
409 * Writes @v to the doorbell aperture at the
410 * requested doorbell index (VEGA10+).
411 */
412void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v)
413{
414 if (index < adev->doorbell.num_doorbells) {
415 atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v);
416 } else {
417 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
418 }
419}
420
d38ceaf9
AD
421/**
422 * amdgpu_invalid_rreg - dummy reg read function
423 *
424 * @adev: amdgpu device pointer
425 * @reg: offset of register
426 *
427 * Dummy register read function. Used for register blocks
428 * that certain asics don't have (all asics).
429 * Returns the value in the register.
430 */
431static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
432{
433 DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
434 BUG();
435 return 0;
436}
437
438/**
439 * amdgpu_invalid_wreg - dummy reg write function
440 *
441 * @adev: amdgpu device pointer
442 * @reg: offset of register
443 * @v: value to write to the register
444 *
445 * Dummy register read function. Used for register blocks
446 * that certain asics don't have (all asics).
447 */
448static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
449{
450 DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
451 reg, v);
452 BUG();
453}
454
4fa1c6a6
TZ
455/**
456 * amdgpu_invalid_rreg64 - dummy 64 bit reg read function
457 *
458 * @adev: amdgpu device pointer
459 * @reg: offset of register
460 *
461 * Dummy register read function. Used for register blocks
462 * that certain asics don't have (all asics).
463 * Returns the value in the register.
464 */
465static uint64_t amdgpu_invalid_rreg64(struct amdgpu_device *adev, uint32_t reg)
466{
467 DRM_ERROR("Invalid callback to read 64 bit register 0x%04X\n", reg);
468 BUG();
469 return 0;
470}
471
472/**
473 * amdgpu_invalid_wreg64 - dummy reg write function
474 *
475 * @adev: amdgpu device pointer
476 * @reg: offset of register
477 * @v: value to write to the register
478 *
479 * Dummy register read function. Used for register blocks
480 * that certain asics don't have (all asics).
481 */
482static void amdgpu_invalid_wreg64(struct amdgpu_device *adev, uint32_t reg, uint64_t v)
483{
484 DRM_ERROR("Invalid callback to write 64 bit register 0x%04X with 0x%08llX\n",
485 reg, v);
486 BUG();
487}
488
d38ceaf9
AD
489/**
490 * amdgpu_block_invalid_rreg - dummy reg read function
491 *
492 * @adev: amdgpu device pointer
493 * @block: offset of instance
494 * @reg: offset of register
495 *
496 * Dummy register read function. Used for register blocks
497 * that certain asics don't have (all asics).
498 * Returns the value in the register.
499 */
500static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
501 uint32_t block, uint32_t reg)
502{
503 DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
504 reg, block);
505 BUG();
506 return 0;
507}
508
509/**
510 * amdgpu_block_invalid_wreg - dummy reg write function
511 *
512 * @adev: amdgpu device pointer
513 * @block: offset of instance
514 * @reg: offset of register
515 * @v: value to write to the register
516 *
517 * Dummy register read function. Used for register blocks
518 * that certain asics don't have (all asics).
519 */
520static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
521 uint32_t block,
522 uint32_t reg, uint32_t v)
523{
524 DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
525 reg, block, v);
526 BUG();
527}
528
e3ecdffa
AD
529/**
530 * amdgpu_device_vram_scratch_init - allocate the VRAM scratch page
531 *
532 * @adev: amdgpu device pointer
533 *
534 * Allocates a scratch page of VRAM for use by various things in the
535 * driver.
536 */
06ec9070 537static int amdgpu_device_vram_scratch_init(struct amdgpu_device *adev)
d38ceaf9 538{
a4a02777
CK
539 return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE,
540 PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
541 &adev->vram_scratch.robj,
542 &adev->vram_scratch.gpu_addr,
543 (void **)&adev->vram_scratch.ptr);
d38ceaf9
AD
544}
545
e3ecdffa
AD
546/**
547 * amdgpu_device_vram_scratch_fini - Free the VRAM scratch page
548 *
549 * @adev: amdgpu device pointer
550 *
551 * Frees the VRAM scratch page.
552 */
06ec9070 553static void amdgpu_device_vram_scratch_fini(struct amdgpu_device *adev)
d38ceaf9 554{
078af1a3 555 amdgpu_bo_free_kernel(&adev->vram_scratch.robj, NULL, NULL);
d38ceaf9
AD
556}
557
558/**
9c3f2b54 559 * amdgpu_device_program_register_sequence - program an array of registers.
d38ceaf9
AD
560 *
561 * @adev: amdgpu_device pointer
562 * @registers: pointer to the register array
563 * @array_size: size of the register array
564 *
565 * Programs an array or registers with and and or masks.
566 * This is a helper for setting golden registers.
567 */
9c3f2b54
AD
568void amdgpu_device_program_register_sequence(struct amdgpu_device *adev,
569 const u32 *registers,
570 const u32 array_size)
d38ceaf9
AD
571{
572 u32 tmp, reg, and_mask, or_mask;
573 int i;
574
575 if (array_size % 3)
576 return;
577
578 for (i = 0; i < array_size; i +=3) {
579 reg = registers[i + 0];
580 and_mask = registers[i + 1];
581 or_mask = registers[i + 2];
582
583 if (and_mask == 0xffffffff) {
584 tmp = or_mask;
585 } else {
586 tmp = RREG32(reg);
587 tmp &= ~and_mask;
e0d07657
HZ
588 if (adev->family >= AMDGPU_FAMILY_AI)
589 tmp |= (or_mask & and_mask);
590 else
591 tmp |= or_mask;
d38ceaf9
AD
592 }
593 WREG32(reg, tmp);
594 }
595}
596
e3ecdffa
AD
597/**
598 * amdgpu_device_pci_config_reset - reset the GPU
599 *
600 * @adev: amdgpu_device pointer
601 *
602 * Resets the GPU using the pci config reset sequence.
603 * Only applicable to asics prior to vega10.
604 */
8111c387 605void amdgpu_device_pci_config_reset(struct amdgpu_device *adev)
d38ceaf9
AD
606{
607 pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
608}
609
610/*
611 * GPU doorbell aperture helpers function.
612 */
613/**
06ec9070 614 * amdgpu_device_doorbell_init - Init doorbell driver information.
d38ceaf9
AD
615 *
616 * @adev: amdgpu_device pointer
617 *
618 * Init doorbell driver information (CIK)
619 * Returns 0 on success, error on failure.
620 */
06ec9070 621static int amdgpu_device_doorbell_init(struct amdgpu_device *adev)
d38ceaf9 622{
6585661d 623
705e519e
CK
624 /* No doorbell on SI hardware generation */
625 if (adev->asic_type < CHIP_BONAIRE) {
626 adev->doorbell.base = 0;
627 adev->doorbell.size = 0;
628 adev->doorbell.num_doorbells = 0;
629 adev->doorbell.ptr = NULL;
630 return 0;
631 }
632
d6895ad3
CK
633 if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET)
634 return -EINVAL;
635
22357775
AD
636 amdgpu_asic_init_doorbell_index(adev);
637
d38ceaf9
AD
638 /* doorbell bar mapping */
639 adev->doorbell.base = pci_resource_start(adev->pdev, 2);
640 adev->doorbell.size = pci_resource_len(adev->pdev, 2);
641
edf600da 642 adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32),
9564f192 643 adev->doorbell_index.max_assignment+1);
d38ceaf9
AD
644 if (adev->doorbell.num_doorbells == 0)
645 return -EINVAL;
646
ec3db8a6 647 /* For Vega, reserve and map two pages on doorbell BAR since SDMA
88dc26e4
OZ
648 * paging queue doorbell use the second page. The
649 * AMDGPU_DOORBELL64_MAX_ASSIGNMENT definition assumes all the
650 * doorbells are in the first page. So with paging queue enabled,
651 * the max num_doorbells should + 1 page (0x400 in dword)
ec3db8a6
PY
652 */
653 if (adev->asic_type >= CHIP_VEGA10)
88dc26e4 654 adev->doorbell.num_doorbells += 0x400;
ec3db8a6 655
8972e5d2
CK
656 adev->doorbell.ptr = ioremap(adev->doorbell.base,
657 adev->doorbell.num_doorbells *
658 sizeof(u32));
659 if (adev->doorbell.ptr == NULL)
d38ceaf9 660 return -ENOMEM;
d38ceaf9
AD
661
662 return 0;
663}
664
665/**
06ec9070 666 * amdgpu_device_doorbell_fini - Tear down doorbell driver information.
d38ceaf9
AD
667 *
668 * @adev: amdgpu_device pointer
669 *
670 * Tear down doorbell driver information (CIK)
671 */
06ec9070 672static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev)
d38ceaf9
AD
673{
674 iounmap(adev->doorbell.ptr);
675 adev->doorbell.ptr = NULL;
676}
677
22cb0164 678
d38ceaf9
AD
679
680/*
06ec9070 681 * amdgpu_device_wb_*()
455a7bc2 682 * Writeback is the method by which the GPU updates special pages in memory
ea81a173 683 * with the status of certain GPU events (fences, ring pointers,etc.).
d38ceaf9
AD
684 */
685
686/**
06ec9070 687 * amdgpu_device_wb_fini - Disable Writeback and free memory
d38ceaf9
AD
688 *
689 * @adev: amdgpu_device pointer
690 *
691 * Disables Writeback and frees the Writeback memory (all asics).
692 * Used at driver shutdown.
693 */
06ec9070 694static void amdgpu_device_wb_fini(struct amdgpu_device *adev)
d38ceaf9
AD
695{
696 if (adev->wb.wb_obj) {
a76ed485
AD
697 amdgpu_bo_free_kernel(&adev->wb.wb_obj,
698 &adev->wb.gpu_addr,
699 (void **)&adev->wb.wb);
d38ceaf9
AD
700 adev->wb.wb_obj = NULL;
701 }
702}
703
704/**
06ec9070 705 * amdgpu_device_wb_init- Init Writeback driver info and allocate memory
d38ceaf9
AD
706 *
707 * @adev: amdgpu_device pointer
708 *
455a7bc2 709 * Initializes writeback and allocates writeback memory (all asics).
d38ceaf9
AD
710 * Used at driver startup.
711 * Returns 0 on success or an -error on failure.
712 */
06ec9070 713static int amdgpu_device_wb_init(struct amdgpu_device *adev)
d38ceaf9
AD
714{
715 int r;
716
717 if (adev->wb.wb_obj == NULL) {
97407b63
AD
718 /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */
719 r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8,
a76ed485
AD
720 PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
721 &adev->wb.wb_obj, &adev->wb.gpu_addr,
722 (void **)&adev->wb.wb);
d38ceaf9
AD
723 if (r) {
724 dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
725 return r;
726 }
d38ceaf9
AD
727
728 adev->wb.num_wb = AMDGPU_MAX_WB;
729 memset(&adev->wb.used, 0, sizeof(adev->wb.used));
730
731 /* clear wb memory */
73469585 732 memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8);
d38ceaf9
AD
733 }
734
735 return 0;
736}
737
738/**
131b4b36 739 * amdgpu_device_wb_get - Allocate a wb entry
d38ceaf9
AD
740 *
741 * @adev: amdgpu_device pointer
742 * @wb: wb index
743 *
744 * Allocate a wb slot for use by the driver (all asics).
745 * Returns 0 on success or -EINVAL on failure.
746 */
131b4b36 747int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb)
d38ceaf9
AD
748{
749 unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
d38ceaf9 750
97407b63 751 if (offset < adev->wb.num_wb) {
7014285a 752 __set_bit(offset, adev->wb.used);
63ae07ca 753 *wb = offset << 3; /* convert to dw offset */
0915fdbc
ML
754 return 0;
755 } else {
756 return -EINVAL;
757 }
758}
759
d38ceaf9 760/**
131b4b36 761 * amdgpu_device_wb_free - Free a wb entry
d38ceaf9
AD
762 *
763 * @adev: amdgpu_device pointer
764 * @wb: wb index
765 *
766 * Free a wb slot allocated for use by the driver (all asics)
767 */
131b4b36 768void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb)
d38ceaf9 769{
73469585 770 wb >>= 3;
d38ceaf9 771 if (wb < adev->wb.num_wb)
73469585 772 __clear_bit(wb, adev->wb.used);
d38ceaf9
AD
773}
774
d6895ad3
CK
775/**
776 * amdgpu_device_resize_fb_bar - try to resize FB BAR
777 *
778 * @adev: amdgpu_device pointer
779 *
780 * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not
781 * to fail, but if any of the BARs is not accessible after the size we abort
782 * driver loading by returning -ENODEV.
783 */
784int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev)
785{
770d13b1 786 u64 space_needed = roundup_pow_of_two(adev->gmc.real_vram_size);
d6895ad3 787 u32 rbar_size = order_base_2(((space_needed >> 20) | 1)) - 1;
31b8adab
CK
788 struct pci_bus *root;
789 struct resource *res;
790 unsigned i;
d6895ad3
CK
791 u16 cmd;
792 int r;
793
0c03b912 794 /* Bypass for VF */
795 if (amdgpu_sriov_vf(adev))
796 return 0;
797
31b8adab
CK
798 /* Check if the root BUS has 64bit memory resources */
799 root = adev->pdev->bus;
800 while (root->parent)
801 root = root->parent;
802
803 pci_bus_for_each_resource(root, res, i) {
0ebb7c54 804 if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) &&
31b8adab
CK
805 res->start > 0x100000000ull)
806 break;
807 }
808
809 /* Trying to resize is pointless without a root hub window above 4GB */
810 if (!res)
811 return 0;
812
d6895ad3
CK
813 /* Disable memory decoding while we change the BAR addresses and size */
814 pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd);
815 pci_write_config_word(adev->pdev, PCI_COMMAND,
816 cmd & ~PCI_COMMAND_MEMORY);
817
818 /* Free the VRAM and doorbell BAR, we most likely need to move both. */
06ec9070 819 amdgpu_device_doorbell_fini(adev);
d6895ad3
CK
820 if (adev->asic_type >= CHIP_BONAIRE)
821 pci_release_resource(adev->pdev, 2);
822
823 pci_release_resource(adev->pdev, 0);
824
825 r = pci_resize_resource(adev->pdev, 0, rbar_size);
826 if (r == -ENOSPC)
827 DRM_INFO("Not enough PCI address space for a large BAR.");
828 else if (r && r != -ENOTSUPP)
829 DRM_ERROR("Problem resizing BAR0 (%d).", r);
830
831 pci_assign_unassigned_bus_resources(adev->pdev->bus);
832
833 /* When the doorbell or fb BAR isn't available we have no chance of
834 * using the device.
835 */
06ec9070 836 r = amdgpu_device_doorbell_init(adev);
d6895ad3
CK
837 if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET))
838 return -ENODEV;
839
840 pci_write_config_word(adev->pdev, PCI_COMMAND, cmd);
841
842 return 0;
843}
a05502e5 844
d38ceaf9
AD
845/*
846 * GPU helpers function.
847 */
848/**
39c640c0 849 * amdgpu_device_need_post - check if the hw need post or not
d38ceaf9
AD
850 *
851 * @adev: amdgpu_device pointer
852 *
c836fec5
JQ
853 * Check if the asic has been initialized (all asics) at driver startup
854 * or post is needed if hw reset is performed.
855 * Returns true if need or false if not.
d38ceaf9 856 */
39c640c0 857bool amdgpu_device_need_post(struct amdgpu_device *adev)
d38ceaf9
AD
858{
859 uint32_t reg;
860
bec86378
ML
861 if (amdgpu_sriov_vf(adev))
862 return false;
863
864 if (amdgpu_passthrough(adev)) {
1da2c326
ML
865 /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
866 * some old smc fw still need driver do vPost otherwise gpu hang, while
867 * those smc fw version above 22.15 doesn't have this flaw, so we force
868 * vpost executed for smc version below 22.15
bec86378
ML
869 */
870 if (adev->asic_type == CHIP_FIJI) {
871 int err;
872 uint32_t fw_ver;
873 err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
874 /* force vPost if error occured */
875 if (err)
876 return true;
877
878 fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
1da2c326
ML
879 if (fw_ver < 0x00160e00)
880 return true;
bec86378 881 }
bec86378 882 }
91fe77eb 883
884 if (adev->has_hw_reset) {
885 adev->has_hw_reset = false;
886 return true;
887 }
888
889 /* bios scratch used on CIK+ */
890 if (adev->asic_type >= CHIP_BONAIRE)
891 return amdgpu_atombios_scratch_need_asic_init(adev);
892
893 /* check MEM_SIZE for older asics */
894 reg = amdgpu_asic_get_config_memsize(adev);
895
896 if ((reg != 0) && (reg != 0xffffffff))
897 return false;
898
899 return true;
bec86378
ML
900}
901
d38ceaf9
AD
902/* if we get transitioned to only one device, take VGA back */
903/**
06ec9070 904 * amdgpu_device_vga_set_decode - enable/disable vga decode
d38ceaf9
AD
905 *
906 * @cookie: amdgpu_device pointer
907 * @state: enable/disable vga decode
908 *
909 * Enable/disable vga decode (all asics).
910 * Returns VGA resource flags.
911 */
06ec9070 912static unsigned int amdgpu_device_vga_set_decode(void *cookie, bool state)
d38ceaf9
AD
913{
914 struct amdgpu_device *adev = cookie;
915 amdgpu_asic_set_vga_state(adev, state);
916 if (state)
917 return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
918 VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
919 else
920 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
921}
922
e3ecdffa
AD
923/**
924 * amdgpu_device_check_block_size - validate the vm block size
925 *
926 * @adev: amdgpu_device pointer
927 *
928 * Validates the vm block size specified via module parameter.
929 * The vm block size defines number of bits in page table versus page directory,
930 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
931 * page table and the remaining bits are in the page directory.
932 */
06ec9070 933static void amdgpu_device_check_block_size(struct amdgpu_device *adev)
a1adf8be
CZ
934{
935 /* defines number of bits in page table versus page directory,
936 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
937 * page table and the remaining bits are in the page directory */
bab4fee7
JZ
938 if (amdgpu_vm_block_size == -1)
939 return;
a1adf8be 940
bab4fee7 941 if (amdgpu_vm_block_size < 9) {
a1adf8be
CZ
942 dev_warn(adev->dev, "VM page table size (%d) too small\n",
943 amdgpu_vm_block_size);
97489129 944 amdgpu_vm_block_size = -1;
a1adf8be 945 }
a1adf8be
CZ
946}
947
e3ecdffa
AD
948/**
949 * amdgpu_device_check_vm_size - validate the vm size
950 *
951 * @adev: amdgpu_device pointer
952 *
953 * Validates the vm size in GB specified via module parameter.
954 * The VM size is the size of the GPU virtual memory space in GB.
955 */
06ec9070 956static void amdgpu_device_check_vm_size(struct amdgpu_device *adev)
83ca145d 957{
64dab074
AD
958 /* no need to check the default value */
959 if (amdgpu_vm_size == -1)
960 return;
961
83ca145d
ZJ
962 if (amdgpu_vm_size < 1) {
963 dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
964 amdgpu_vm_size);
f3368128 965 amdgpu_vm_size = -1;
83ca145d 966 }
83ca145d
ZJ
967}
968
7951e376
RZ
969static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev)
970{
971 struct sysinfo si;
972 bool is_os_64 = (sizeof(void *) == 8) ? true : false;
973 uint64_t total_memory;
974 uint64_t dram_size_seven_GB = 0x1B8000000;
975 uint64_t dram_size_three_GB = 0xB8000000;
976
977 if (amdgpu_smu_memory_pool_size == 0)
978 return;
979
980 if (!is_os_64) {
981 DRM_WARN("Not 64-bit OS, feature not supported\n");
982 goto def_value;
983 }
984 si_meminfo(&si);
985 total_memory = (uint64_t)si.totalram * si.mem_unit;
986
987 if ((amdgpu_smu_memory_pool_size == 1) ||
988 (amdgpu_smu_memory_pool_size == 2)) {
989 if (total_memory < dram_size_three_GB)
990 goto def_value1;
991 } else if ((amdgpu_smu_memory_pool_size == 4) ||
992 (amdgpu_smu_memory_pool_size == 8)) {
993 if (total_memory < dram_size_seven_GB)
994 goto def_value1;
995 } else {
996 DRM_WARN("Smu memory pool size not supported\n");
997 goto def_value;
998 }
999 adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28;
1000
1001 return;
1002
1003def_value1:
1004 DRM_WARN("No enough system memory\n");
1005def_value:
1006 adev->pm.smu_prv_buffer_size = 0;
1007}
1008
d38ceaf9 1009/**
06ec9070 1010 * amdgpu_device_check_arguments - validate module params
d38ceaf9
AD
1011 *
1012 * @adev: amdgpu_device pointer
1013 *
1014 * Validates certain module parameters and updates
1015 * the associated values used by the driver (all asics).
1016 */
912dfc84 1017static int amdgpu_device_check_arguments(struct amdgpu_device *adev)
d38ceaf9 1018{
912dfc84
EQ
1019 int ret = 0;
1020
5b011235
CZ
1021 if (amdgpu_sched_jobs < 4) {
1022 dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
1023 amdgpu_sched_jobs);
1024 amdgpu_sched_jobs = 4;
76117507 1025 } else if (!is_power_of_2(amdgpu_sched_jobs)){
5b011235
CZ
1026 dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
1027 amdgpu_sched_jobs);
1028 amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
1029 }
d38ceaf9 1030
83e74db6 1031 if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) {
f9321cc4
CK
1032 /* gart size must be greater or equal to 32M */
1033 dev_warn(adev->dev, "gart size (%d) too small\n",
1034 amdgpu_gart_size);
83e74db6 1035 amdgpu_gart_size = -1;
d38ceaf9
AD
1036 }
1037
36d38372 1038 if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) {
c4e1a13a 1039 /* gtt size must be greater or equal to 32M */
36d38372
CK
1040 dev_warn(adev->dev, "gtt size (%d) too small\n",
1041 amdgpu_gtt_size);
1042 amdgpu_gtt_size = -1;
d38ceaf9
AD
1043 }
1044
d07f14be
RH
1045 /* valid range is between 4 and 9 inclusive */
1046 if (amdgpu_vm_fragment_size != -1 &&
1047 (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) {
1048 dev_warn(adev->dev, "valid range is between 4 and 9\n");
1049 amdgpu_vm_fragment_size = -1;
1050 }
1051
7951e376
RZ
1052 amdgpu_device_check_smu_prv_buffer_size(adev);
1053
06ec9070 1054 amdgpu_device_check_vm_size(adev);
d38ceaf9 1055
06ec9070 1056 amdgpu_device_check_block_size(adev);
6a7f76e7 1057
19aede77 1058 adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type);
912dfc84
EQ
1059
1060 return ret;
d38ceaf9
AD
1061}
1062
1063/**
1064 * amdgpu_switcheroo_set_state - set switcheroo state
1065 *
1066 * @pdev: pci dev pointer
1694467b 1067 * @state: vga_switcheroo state
d38ceaf9
AD
1068 *
1069 * Callback for the switcheroo driver. Suspends or resumes the
1070 * the asics before or after it is powered up using ACPI methods.
1071 */
1072static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, enum vga_switcheroo_state state)
1073{
1074 struct drm_device *dev = pci_get_drvdata(pdev);
1075
1076 if (amdgpu_device_is_px(dev) && state == VGA_SWITCHEROO_OFF)
1077 return;
1078
1079 if (state == VGA_SWITCHEROO_ON) {
7ca85295 1080 pr_info("amdgpu: switched on\n");
d38ceaf9
AD
1081 /* don't suspend or resume card normally */
1082 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1083
810ddc3a 1084 amdgpu_device_resume(dev, true, true);
d38ceaf9 1085
d38ceaf9
AD
1086 dev->switch_power_state = DRM_SWITCH_POWER_ON;
1087 drm_kms_helper_poll_enable(dev);
1088 } else {
7ca85295 1089 pr_info("amdgpu: switched off\n");
d38ceaf9
AD
1090 drm_kms_helper_poll_disable(dev);
1091 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
810ddc3a 1092 amdgpu_device_suspend(dev, true, true);
d38ceaf9
AD
1093 dev->switch_power_state = DRM_SWITCH_POWER_OFF;
1094 }
1095}
1096
1097/**
1098 * amdgpu_switcheroo_can_switch - see if switcheroo state can change
1099 *
1100 * @pdev: pci dev pointer
1101 *
1102 * Callback for the switcheroo driver. Check of the switcheroo
1103 * state can be changed.
1104 * Returns true if the state can be changed, false if not.
1105 */
1106static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
1107{
1108 struct drm_device *dev = pci_get_drvdata(pdev);
1109
1110 /*
1111 * FIXME: open_count is protected by drm_global_mutex but that would lead to
1112 * locking inversion with the driver load path. And the access here is
1113 * completely racy anyway. So don't bother with locking for now.
1114 */
1115 return dev->open_count == 0;
1116}
1117
1118static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
1119 .set_gpu_state = amdgpu_switcheroo_set_state,
1120 .reprobe = NULL,
1121 .can_switch = amdgpu_switcheroo_can_switch,
1122};
1123
e3ecdffa
AD
1124/**
1125 * amdgpu_device_ip_set_clockgating_state - set the CG state
1126 *
87e3f136 1127 * @dev: amdgpu_device pointer
e3ecdffa
AD
1128 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1129 * @state: clockgating state (gate or ungate)
1130 *
1131 * Sets the requested clockgating state for all instances of
1132 * the hardware IP specified.
1133 * Returns the error code from the last instance.
1134 */
43fa561f 1135int amdgpu_device_ip_set_clockgating_state(void *dev,
2990a1fc
AD
1136 enum amd_ip_block_type block_type,
1137 enum amd_clockgating_state state)
d38ceaf9 1138{
43fa561f 1139 struct amdgpu_device *adev = dev;
d38ceaf9
AD
1140 int i, r = 0;
1141
1142 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 1143 if (!adev->ip_blocks[i].status.valid)
9ecbe7f5 1144 continue;
c722865a
RZ
1145 if (adev->ip_blocks[i].version->type != block_type)
1146 continue;
1147 if (!adev->ip_blocks[i].version->funcs->set_clockgating_state)
1148 continue;
1149 r = adev->ip_blocks[i].version->funcs->set_clockgating_state(
1150 (void *)adev, state);
1151 if (r)
1152 DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",
1153 adev->ip_blocks[i].version->funcs->name, r);
d38ceaf9
AD
1154 }
1155 return r;
1156}
1157
e3ecdffa
AD
1158/**
1159 * amdgpu_device_ip_set_powergating_state - set the PG state
1160 *
87e3f136 1161 * @dev: amdgpu_device pointer
e3ecdffa
AD
1162 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1163 * @state: powergating state (gate or ungate)
1164 *
1165 * Sets the requested powergating state for all instances of
1166 * the hardware IP specified.
1167 * Returns the error code from the last instance.
1168 */
43fa561f 1169int amdgpu_device_ip_set_powergating_state(void *dev,
2990a1fc
AD
1170 enum amd_ip_block_type block_type,
1171 enum amd_powergating_state state)
d38ceaf9 1172{
43fa561f 1173 struct amdgpu_device *adev = dev;
d38ceaf9
AD
1174 int i, r = 0;
1175
1176 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 1177 if (!adev->ip_blocks[i].status.valid)
9ecbe7f5 1178 continue;
c722865a
RZ
1179 if (adev->ip_blocks[i].version->type != block_type)
1180 continue;
1181 if (!adev->ip_blocks[i].version->funcs->set_powergating_state)
1182 continue;
1183 r = adev->ip_blocks[i].version->funcs->set_powergating_state(
1184 (void *)adev, state);
1185 if (r)
1186 DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",
1187 adev->ip_blocks[i].version->funcs->name, r);
d38ceaf9
AD
1188 }
1189 return r;
1190}
1191
e3ecdffa
AD
1192/**
1193 * amdgpu_device_ip_get_clockgating_state - get the CG state
1194 *
1195 * @adev: amdgpu_device pointer
1196 * @flags: clockgating feature flags
1197 *
1198 * Walks the list of IPs on the device and updates the clockgating
1199 * flags for each IP.
1200 * Updates @flags with the feature flags for each hardware IP where
1201 * clockgating is enabled.
1202 */
2990a1fc
AD
1203void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev,
1204 u32 *flags)
6cb2d4e4
HR
1205{
1206 int i;
1207
1208 for (i = 0; i < adev->num_ip_blocks; i++) {
1209 if (!adev->ip_blocks[i].status.valid)
1210 continue;
1211 if (adev->ip_blocks[i].version->funcs->get_clockgating_state)
1212 adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags);
1213 }
1214}
1215
e3ecdffa
AD
1216/**
1217 * amdgpu_device_ip_wait_for_idle - wait for idle
1218 *
1219 * @adev: amdgpu_device pointer
1220 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1221 *
1222 * Waits for the request hardware IP to be idle.
1223 * Returns 0 for success or a negative error code on failure.
1224 */
2990a1fc
AD
1225int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev,
1226 enum amd_ip_block_type block_type)
5dbbb60b
AD
1227{
1228 int i, r;
1229
1230 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 1231 if (!adev->ip_blocks[i].status.valid)
9ecbe7f5 1232 continue;
a1255107
AD
1233 if (adev->ip_blocks[i].version->type == block_type) {
1234 r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
5dbbb60b
AD
1235 if (r)
1236 return r;
1237 break;
1238 }
1239 }
1240 return 0;
1241
1242}
1243
e3ecdffa
AD
1244/**
1245 * amdgpu_device_ip_is_idle - is the hardware IP idle
1246 *
1247 * @adev: amdgpu_device pointer
1248 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1249 *
1250 * Check if the hardware IP is idle or not.
1251 * Returns true if it the IP is idle, false if not.
1252 */
2990a1fc
AD
1253bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev,
1254 enum amd_ip_block_type block_type)
5dbbb60b
AD
1255{
1256 int i;
1257
1258 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 1259 if (!adev->ip_blocks[i].status.valid)
9ecbe7f5 1260 continue;
a1255107
AD
1261 if (adev->ip_blocks[i].version->type == block_type)
1262 return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
5dbbb60b
AD
1263 }
1264 return true;
1265
1266}
1267
e3ecdffa
AD
1268/**
1269 * amdgpu_device_ip_get_ip_block - get a hw IP pointer
1270 *
1271 * @adev: amdgpu_device pointer
87e3f136 1272 * @type: Type of hardware IP (SMU, GFX, UVD, etc.)
e3ecdffa
AD
1273 *
1274 * Returns a pointer to the hardware IP block structure
1275 * if it exists for the asic, otherwise NULL.
1276 */
2990a1fc
AD
1277struct amdgpu_ip_block *
1278amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,
1279 enum amd_ip_block_type type)
d38ceaf9
AD
1280{
1281 int i;
1282
1283 for (i = 0; i < adev->num_ip_blocks; i++)
a1255107 1284 if (adev->ip_blocks[i].version->type == type)
d38ceaf9
AD
1285 return &adev->ip_blocks[i];
1286
1287 return NULL;
1288}
1289
1290/**
2990a1fc 1291 * amdgpu_device_ip_block_version_cmp
d38ceaf9
AD
1292 *
1293 * @adev: amdgpu_device pointer
5fc3aeeb 1294 * @type: enum amd_ip_block_type
d38ceaf9
AD
1295 * @major: major version
1296 * @minor: minor version
1297 *
1298 * return 0 if equal or greater
1299 * return 1 if smaller or the ip_block doesn't exist
1300 */
2990a1fc
AD
1301int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev,
1302 enum amd_ip_block_type type,
1303 u32 major, u32 minor)
d38ceaf9 1304{
2990a1fc 1305 struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type);
d38ceaf9 1306
a1255107
AD
1307 if (ip_block && ((ip_block->version->major > major) ||
1308 ((ip_block->version->major == major) &&
1309 (ip_block->version->minor >= minor))))
d38ceaf9
AD
1310 return 0;
1311
1312 return 1;
1313}
1314
a1255107 1315/**
2990a1fc 1316 * amdgpu_device_ip_block_add
a1255107
AD
1317 *
1318 * @adev: amdgpu_device pointer
1319 * @ip_block_version: pointer to the IP to add
1320 *
1321 * Adds the IP block driver information to the collection of IPs
1322 * on the asic.
1323 */
2990a1fc
AD
1324int amdgpu_device_ip_block_add(struct amdgpu_device *adev,
1325 const struct amdgpu_ip_block_version *ip_block_version)
a1255107
AD
1326{
1327 if (!ip_block_version)
1328 return -EINVAL;
1329
e966a725 1330 DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks,
a0bae357
HR
1331 ip_block_version->funcs->name);
1332
a1255107
AD
1333 adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
1334
1335 return 0;
1336}
1337
e3ecdffa
AD
1338/**
1339 * amdgpu_device_enable_virtual_display - enable virtual display feature
1340 *
1341 * @adev: amdgpu_device pointer
1342 *
1343 * Enabled the virtual display feature if the user has enabled it via
1344 * the module parameter virtual_display. This feature provides a virtual
1345 * display hardware on headless boards or in virtualized environments.
1346 * This function parses and validates the configuration string specified by
1347 * the user and configues the virtual display configuration (number of
1348 * virtual connectors, crtcs, etc.) specified.
1349 */
483ef985 1350static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
9accf2fd
ED
1351{
1352 adev->enable_virtual_display = false;
1353
1354 if (amdgpu_virtual_display) {
1355 struct drm_device *ddev = adev->ddev;
1356 const char *pci_address_name = pci_name(ddev->pdev);
0f66356d 1357 char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
9accf2fd
ED
1358
1359 pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
1360 pciaddstr_tmp = pciaddstr;
0f66356d
ED
1361 while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
1362 pciaddname = strsep(&pciaddname_tmp, ",");
967de2a9
YT
1363 if (!strcmp("all", pciaddname)
1364 || !strcmp(pci_address_name, pciaddname)) {
0f66356d
ED
1365 long num_crtc;
1366 int res = -1;
1367
9accf2fd 1368 adev->enable_virtual_display = true;
0f66356d
ED
1369
1370 if (pciaddname_tmp)
1371 res = kstrtol(pciaddname_tmp, 10,
1372 &num_crtc);
1373
1374 if (!res) {
1375 if (num_crtc < 1)
1376 num_crtc = 1;
1377 if (num_crtc > 6)
1378 num_crtc = 6;
1379 adev->mode_info.num_crtc = num_crtc;
1380 } else {
1381 adev->mode_info.num_crtc = 1;
1382 }
9accf2fd
ED
1383 break;
1384 }
1385 }
1386
0f66356d
ED
1387 DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
1388 amdgpu_virtual_display, pci_address_name,
1389 adev->enable_virtual_display, adev->mode_info.num_crtc);
9accf2fd
ED
1390
1391 kfree(pciaddstr);
1392 }
1393}
1394
e3ecdffa
AD
1395/**
1396 * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware
1397 *
1398 * @adev: amdgpu_device pointer
1399 *
1400 * Parses the asic configuration parameters specified in the gpu info
1401 * firmware and makes them availale to the driver for use in configuring
1402 * the asic.
1403 * Returns 0 on success, -EINVAL on failure.
1404 */
e2a75f88
AD
1405static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev)
1406{
e2a75f88
AD
1407 const char *chip_name;
1408 char fw_name[30];
1409 int err;
1410 const struct gpu_info_firmware_header_v1_0 *hdr;
1411
ab4fe3e1
HR
1412 adev->firmware.gpu_info_fw = NULL;
1413
e2a75f88
AD
1414 switch (adev->asic_type) {
1415 case CHIP_TOPAZ:
1416 case CHIP_TONGA:
1417 case CHIP_FIJI:
e2a75f88 1418 case CHIP_POLARIS10:
cc07f18d 1419 case CHIP_POLARIS11:
e2a75f88 1420 case CHIP_POLARIS12:
cc07f18d 1421 case CHIP_VEGAM:
e2a75f88
AD
1422 case CHIP_CARRIZO:
1423 case CHIP_STONEY:
1424#ifdef CONFIG_DRM_AMDGPU_SI
1425 case CHIP_VERDE:
1426 case CHIP_TAHITI:
1427 case CHIP_PITCAIRN:
1428 case CHIP_OLAND:
1429 case CHIP_HAINAN:
1430#endif
1431#ifdef CONFIG_DRM_AMDGPU_CIK
1432 case CHIP_BONAIRE:
1433 case CHIP_HAWAII:
1434 case CHIP_KAVERI:
1435 case CHIP_KABINI:
1436 case CHIP_MULLINS:
1437#endif
27c0bc71 1438 case CHIP_VEGA20:
e2a75f88
AD
1439 default:
1440 return 0;
1441 case CHIP_VEGA10:
1442 chip_name = "vega10";
1443 break;
3f76dced
AD
1444 case CHIP_VEGA12:
1445 chip_name = "vega12";
1446 break;
2d2e5e7e 1447 case CHIP_RAVEN:
54c4d17e
FX
1448 if (adev->rev_id >= 8)
1449 chip_name = "raven2";
741deade
AD
1450 else if (adev->pdev->device == 0x15d8)
1451 chip_name = "picasso";
54c4d17e
FX
1452 else
1453 chip_name = "raven";
2d2e5e7e 1454 break;
65e60f6e
LM
1455 case CHIP_ARCTURUS:
1456 chip_name = "arcturus";
1457 break;
b51a26a0
HR
1458 case CHIP_RENOIR:
1459 chip_name = "renoir";
1460 break;
23c6268e
HR
1461 case CHIP_NAVI10:
1462 chip_name = "navi10";
1463 break;
ed42cfe1
XY
1464 case CHIP_NAVI14:
1465 chip_name = "navi14";
1466 break;
42b325e5
XY
1467 case CHIP_NAVI12:
1468 chip_name = "navi12";
1469 break;
e2a75f88
AD
1470 }
1471
1472 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name);
ab4fe3e1 1473 err = request_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev);
e2a75f88
AD
1474 if (err) {
1475 dev_err(adev->dev,
1476 "Failed to load gpu_info firmware \"%s\"\n",
1477 fw_name);
1478 goto out;
1479 }
ab4fe3e1 1480 err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw);
e2a75f88
AD
1481 if (err) {
1482 dev_err(adev->dev,
1483 "Failed to validate gpu_info firmware \"%s\"\n",
1484 fw_name);
1485 goto out;
1486 }
1487
ab4fe3e1 1488 hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data;
e2a75f88
AD
1489 amdgpu_ucode_print_gpu_info_hdr(&hdr->header);
1490
1491 switch (hdr->version_major) {
1492 case 1:
1493 {
1494 const struct gpu_info_firmware_v1_0 *gpu_info_fw =
ab4fe3e1 1495 (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data +
e2a75f88
AD
1496 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1497
ec51d3fa
XY
1498 if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10)
1499 goto parse_soc_bounding_box;
1500
b5ab16bf
AD
1501 adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se);
1502 adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh);
1503 adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se);
1504 adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se);
e2a75f88 1505 adev->gfx.config.max_texture_channel_caches =
b5ab16bf
AD
1506 le32_to_cpu(gpu_info_fw->gc_num_tccs);
1507 adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs);
1508 adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds);
1509 adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth);
1510 adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth);
e2a75f88 1511 adev->gfx.config.double_offchip_lds_buf =
b5ab16bf
AD
1512 le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer);
1513 adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size);
51fd0370
HZ
1514 adev->gfx.cu_info.max_waves_per_simd =
1515 le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd);
1516 adev->gfx.cu_info.max_scratch_slots_per_cu =
1517 le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu);
1518 adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size);
48321c3d 1519 if (hdr->version_minor >= 1) {
35c2e910
HZ
1520 const struct gpu_info_firmware_v1_1 *gpu_info_fw =
1521 (const struct gpu_info_firmware_v1_1 *)(adev->firmware.gpu_info_fw->data +
1522 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1523 adev->gfx.config.num_sc_per_sh =
1524 le32_to_cpu(gpu_info_fw->num_sc_per_sh);
1525 adev->gfx.config.num_packer_per_sc =
1526 le32_to_cpu(gpu_info_fw->num_packer_per_sc);
1527 }
ec51d3fa
XY
1528
1529parse_soc_bounding_box:
48321c3d 1530#ifdef CONFIG_DRM_AMD_DC_DCN2_0
ec51d3fa
XY
1531 /*
1532 * soc bounding box info is not integrated in disocovery table,
1533 * we always need to parse it from gpu info firmware.
1534 */
48321c3d
HW
1535 if (hdr->version_minor == 2) {
1536 const struct gpu_info_firmware_v1_2 *gpu_info_fw =
1537 (const struct gpu_info_firmware_v1_2 *)(adev->firmware.gpu_info_fw->data +
1538 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1539 adev->dm.soc_bounding_box = &gpu_info_fw->soc_bounding_box;
1540 }
1541#endif
e2a75f88
AD
1542 break;
1543 }
1544 default:
1545 dev_err(adev->dev,
1546 "Unsupported gpu_info table %d\n", hdr->header.ucode_version);
1547 err = -EINVAL;
1548 goto out;
1549 }
1550out:
e2a75f88
AD
1551 return err;
1552}
1553
e3ecdffa
AD
1554/**
1555 * amdgpu_device_ip_early_init - run early init for hardware IPs
1556 *
1557 * @adev: amdgpu_device pointer
1558 *
1559 * Early initialization pass for hardware IPs. The hardware IPs that make
1560 * up each asic are discovered each IP's early_init callback is run. This
1561 * is the first stage in initializing the asic.
1562 * Returns 0 on success, negative error code on failure.
1563 */
06ec9070 1564static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
d38ceaf9 1565{
aaa36a97 1566 int i, r;
d38ceaf9 1567
483ef985 1568 amdgpu_device_enable_virtual_display(adev);
a6be7570 1569
d38ceaf9 1570 switch (adev->asic_type) {
aaa36a97
AD
1571 case CHIP_TOPAZ:
1572 case CHIP_TONGA:
48299f95 1573 case CHIP_FIJI:
2cc0c0b5 1574 case CHIP_POLARIS10:
32cc7e53 1575 case CHIP_POLARIS11:
c4642a47 1576 case CHIP_POLARIS12:
32cc7e53 1577 case CHIP_VEGAM:
aaa36a97 1578 case CHIP_CARRIZO:
39bb0c92
SL
1579 case CHIP_STONEY:
1580 if (adev->asic_type == CHIP_CARRIZO || adev->asic_type == CHIP_STONEY)
aaa36a97
AD
1581 adev->family = AMDGPU_FAMILY_CZ;
1582 else
1583 adev->family = AMDGPU_FAMILY_VI;
1584
1585 r = vi_set_ip_blocks(adev);
1586 if (r)
1587 return r;
1588 break;
33f34802
KW
1589#ifdef CONFIG_DRM_AMDGPU_SI
1590 case CHIP_VERDE:
1591 case CHIP_TAHITI:
1592 case CHIP_PITCAIRN:
1593 case CHIP_OLAND:
1594 case CHIP_HAINAN:
295d0daf 1595 adev->family = AMDGPU_FAMILY_SI;
33f34802
KW
1596 r = si_set_ip_blocks(adev);
1597 if (r)
1598 return r;
1599 break;
1600#endif
a2e73f56
AD
1601#ifdef CONFIG_DRM_AMDGPU_CIK
1602 case CHIP_BONAIRE:
1603 case CHIP_HAWAII:
1604 case CHIP_KAVERI:
1605 case CHIP_KABINI:
1606 case CHIP_MULLINS:
1607 if ((adev->asic_type == CHIP_BONAIRE) || (adev->asic_type == CHIP_HAWAII))
1608 adev->family = AMDGPU_FAMILY_CI;
1609 else
1610 adev->family = AMDGPU_FAMILY_KV;
1611
1612 r = cik_set_ip_blocks(adev);
1613 if (r)
1614 return r;
1615 break;
1616#endif
e48a3cd9
AD
1617 case CHIP_VEGA10:
1618 case CHIP_VEGA12:
e4bd8170 1619 case CHIP_VEGA20:
e48a3cd9 1620 case CHIP_RAVEN:
61cf44c1 1621 case CHIP_ARCTURUS:
b51a26a0
HR
1622 case CHIP_RENOIR:
1623 if (adev->asic_type == CHIP_RAVEN ||
1624 adev->asic_type == CHIP_RENOIR)
2ca8a5d2
CZ
1625 adev->family = AMDGPU_FAMILY_RV;
1626 else
1627 adev->family = AMDGPU_FAMILY_AI;
460826e6
KW
1628
1629 r = soc15_set_ip_blocks(adev);
1630 if (r)
1631 return r;
1632 break;
0a5b8c7b 1633 case CHIP_NAVI10:
7ecb5cd4 1634 case CHIP_NAVI14:
4808cf9c 1635 case CHIP_NAVI12:
0a5b8c7b
HR
1636 adev->family = AMDGPU_FAMILY_NV;
1637
1638 r = nv_set_ip_blocks(adev);
1639 if (r)
1640 return r;
1641 break;
d38ceaf9
AD
1642 default:
1643 /* FIXME: not supported yet */
1644 return -EINVAL;
1645 }
1646
e2a75f88
AD
1647 r = amdgpu_device_parse_gpu_info_fw(adev);
1648 if (r)
1649 return r;
1650
ec51d3fa
XY
1651 if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10)
1652 amdgpu_discovery_get_gfx_info(adev);
1653
1884734a 1654 amdgpu_amdkfd_device_probe(adev);
1655
3149d9da
XY
1656 if (amdgpu_sriov_vf(adev)) {
1657 r = amdgpu_virt_request_full_gpu(adev, true);
1658 if (r)
5ffa61c1 1659 return -EAGAIN;
3149d9da
XY
1660 }
1661
3b94fb10 1662 adev->pm.pp_feature = amdgpu_pp_feature_mask;
a35ad98b 1663 if (amdgpu_sriov_vf(adev) || sched_policy == KFD_SCHED_POLICY_NO_HWS)
00544006 1664 adev->pm.pp_feature &= ~PP_GFXOFF_MASK;
00f54b97 1665
d38ceaf9
AD
1666 for (i = 0; i < adev->num_ip_blocks; i++) {
1667 if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
ed8cf00c
HR
1668 DRM_ERROR("disabled ip block: %d <%s>\n",
1669 i, adev->ip_blocks[i].version->funcs->name);
a1255107 1670 adev->ip_blocks[i].status.valid = false;
d38ceaf9 1671 } else {
a1255107
AD
1672 if (adev->ip_blocks[i].version->funcs->early_init) {
1673 r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
2c1a2784 1674 if (r == -ENOENT) {
a1255107 1675 adev->ip_blocks[i].status.valid = false;
2c1a2784 1676 } else if (r) {
a1255107
AD
1677 DRM_ERROR("early_init of IP block <%s> failed %d\n",
1678 adev->ip_blocks[i].version->funcs->name, r);
d38ceaf9 1679 return r;
2c1a2784 1680 } else {
a1255107 1681 adev->ip_blocks[i].status.valid = true;
2c1a2784 1682 }
974e6b64 1683 } else {
a1255107 1684 adev->ip_blocks[i].status.valid = true;
d38ceaf9 1685 }
d38ceaf9 1686 }
21a249ca
AD
1687 /* get the vbios after the asic_funcs are set up */
1688 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) {
1689 /* Read BIOS */
1690 if (!amdgpu_get_bios(adev))
1691 return -EINVAL;
1692
1693 r = amdgpu_atombios_init(adev);
1694 if (r) {
1695 dev_err(adev->dev, "amdgpu_atombios_init failed\n");
1696 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0);
1697 return r;
1698 }
1699 }
d38ceaf9
AD
1700 }
1701
395d1fb9
NH
1702 adev->cg_flags &= amdgpu_cg_mask;
1703 adev->pg_flags &= amdgpu_pg_mask;
1704
d38ceaf9
AD
1705 return 0;
1706}
1707
0a4f2520
RZ
1708static int amdgpu_device_ip_hw_init_phase1(struct amdgpu_device *adev)
1709{
1710 int i, r;
1711
1712 for (i = 0; i < adev->num_ip_blocks; i++) {
1713 if (!adev->ip_blocks[i].status.sw)
1714 continue;
1715 if (adev->ip_blocks[i].status.hw)
1716 continue;
1717 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2d11fd3f 1718 (amdgpu_sriov_vf(adev) && (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)) ||
0a4f2520
RZ
1719 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
1720 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
1721 if (r) {
1722 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1723 adev->ip_blocks[i].version->funcs->name, r);
1724 return r;
1725 }
1726 adev->ip_blocks[i].status.hw = true;
1727 }
1728 }
1729
1730 return 0;
1731}
1732
1733static int amdgpu_device_ip_hw_init_phase2(struct amdgpu_device *adev)
1734{
1735 int i, r;
1736
1737 for (i = 0; i < adev->num_ip_blocks; i++) {
1738 if (!adev->ip_blocks[i].status.sw)
1739 continue;
1740 if (adev->ip_blocks[i].status.hw)
1741 continue;
1742 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
1743 if (r) {
1744 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1745 adev->ip_blocks[i].version->funcs->name, r);
1746 return r;
1747 }
1748 adev->ip_blocks[i].status.hw = true;
1749 }
1750
1751 return 0;
1752}
1753
7a3e0bb2
RZ
1754static int amdgpu_device_fw_loading(struct amdgpu_device *adev)
1755{
1756 int r = 0;
1757 int i;
80f41f84 1758 uint32_t smu_version;
7a3e0bb2
RZ
1759
1760 if (adev->asic_type >= CHIP_VEGA10) {
1761 for (i = 0; i < adev->num_ip_blocks; i++) {
482f0e53
ML
1762 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_PSP)
1763 continue;
1764
1765 /* no need to do the fw loading again if already done*/
1766 if (adev->ip_blocks[i].status.hw == true)
1767 break;
1768
1769 if (adev->in_gpu_reset || adev->in_suspend) {
1770 r = adev->ip_blocks[i].version->funcs->resume(adev);
1771 if (r) {
1772 DRM_ERROR("resume of IP block <%s> failed %d\n",
7a3e0bb2 1773 adev->ip_blocks[i].version->funcs->name, r);
482f0e53
ML
1774 return r;
1775 }
1776 } else {
1777 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
1778 if (r) {
1779 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1780 adev->ip_blocks[i].version->funcs->name, r);
1781 return r;
7a3e0bb2 1782 }
7a3e0bb2 1783 }
482f0e53
ML
1784
1785 adev->ip_blocks[i].status.hw = true;
1786 break;
7a3e0bb2
RZ
1787 }
1788 }
482f0e53 1789
80f41f84 1790 r = amdgpu_pm_load_smu_firmware(adev, &smu_version);
7a3e0bb2 1791
80f41f84 1792 return r;
7a3e0bb2
RZ
1793}
1794
e3ecdffa
AD
1795/**
1796 * amdgpu_device_ip_init - run init for hardware IPs
1797 *
1798 * @adev: amdgpu_device pointer
1799 *
1800 * Main initialization pass for hardware IPs. The list of all the hardware
1801 * IPs that make up the asic is walked and the sw_init and hw_init callbacks
1802 * are run. sw_init initializes the software state associated with each IP
1803 * and hw_init initializes the hardware associated with each IP.
1804 * Returns 0 on success, negative error code on failure.
1805 */
06ec9070 1806static int amdgpu_device_ip_init(struct amdgpu_device *adev)
d38ceaf9
AD
1807{
1808 int i, r;
1809
c030f2e4 1810 r = amdgpu_ras_init(adev);
1811 if (r)
1812 return r;
1813
d38ceaf9 1814 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 1815 if (!adev->ip_blocks[i].status.valid)
d38ceaf9 1816 continue;
a1255107 1817 r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
2c1a2784 1818 if (r) {
a1255107
AD
1819 DRM_ERROR("sw_init of IP block <%s> failed %d\n",
1820 adev->ip_blocks[i].version->funcs->name, r);
72d3f592 1821 goto init_failed;
2c1a2784 1822 }
a1255107 1823 adev->ip_blocks[i].status.sw = true;
bfca0289 1824
d38ceaf9 1825 /* need to do gmc hw init early so we can allocate gpu mem */
a1255107 1826 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
06ec9070 1827 r = amdgpu_device_vram_scratch_init(adev);
2c1a2784
AD
1828 if (r) {
1829 DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r);
72d3f592 1830 goto init_failed;
2c1a2784 1831 }
a1255107 1832 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
2c1a2784
AD
1833 if (r) {
1834 DRM_ERROR("hw_init %d failed %d\n", i, r);
72d3f592 1835 goto init_failed;
2c1a2784 1836 }
06ec9070 1837 r = amdgpu_device_wb_init(adev);
2c1a2784 1838 if (r) {
06ec9070 1839 DRM_ERROR("amdgpu_device_wb_init failed %d\n", r);
72d3f592 1840 goto init_failed;
2c1a2784 1841 }
a1255107 1842 adev->ip_blocks[i].status.hw = true;
2493664f
ML
1843
1844 /* right after GMC hw init, we create CSA */
f92d5c61 1845 if (amdgpu_mcbp || amdgpu_sriov_vf(adev)) {
1e256e27
RZ
1846 r = amdgpu_allocate_static_csa(adev, &adev->virt.csa_obj,
1847 AMDGPU_GEM_DOMAIN_VRAM,
1848 AMDGPU_CSA_SIZE);
2493664f
ML
1849 if (r) {
1850 DRM_ERROR("allocate CSA failed %d\n", r);
72d3f592 1851 goto init_failed;
2493664f
ML
1852 }
1853 }
d38ceaf9
AD
1854 }
1855 }
1856
533aed27
AG
1857 r = amdgpu_ib_pool_init(adev);
1858 if (r) {
1859 dev_err(adev->dev, "IB initialization failed (%d).\n", r);
1860 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r);
1861 goto init_failed;
1862 }
1863
c8963ea4
RZ
1864 r = amdgpu_ucode_create_bo(adev); /* create ucode bo when sw_init complete*/
1865 if (r)
72d3f592 1866 goto init_failed;
0a4f2520
RZ
1867
1868 r = amdgpu_device_ip_hw_init_phase1(adev);
1869 if (r)
72d3f592 1870 goto init_failed;
0a4f2520 1871
7a3e0bb2
RZ
1872 r = amdgpu_device_fw_loading(adev);
1873 if (r)
72d3f592 1874 goto init_failed;
7a3e0bb2 1875
0a4f2520
RZ
1876 r = amdgpu_device_ip_hw_init_phase2(adev);
1877 if (r)
72d3f592 1878 goto init_failed;
d38ceaf9 1879
121a2bc6
AG
1880 /*
1881 * retired pages will be loaded from eeprom and reserved here,
1882 * it should be called after amdgpu_device_ip_hw_init_phase2 since
1883 * for some ASICs the RAS EEPROM code relies on SMU fully functioning
1884 * for I2C communication which only true at this point.
1885 * recovery_init may fail, but it can free all resources allocated by
1886 * itself and its failure should not stop amdgpu init process.
1887 *
1888 * Note: theoretically, this should be called before all vram allocations
1889 * to protect retired page from abusing
1890 */
1891 amdgpu_ras_recovery_init(adev);
1892
3e2e2ab5
HZ
1893 if (adev->gmc.xgmi.num_physical_nodes > 1)
1894 amdgpu_xgmi_add_device(adev);
1884734a 1895 amdgpu_amdkfd_device_init(adev);
c6332b97 1896
72d3f592 1897init_failed:
d3c117e5 1898 if (amdgpu_sriov_vf(adev)) {
72d3f592
ED
1899 if (!r)
1900 amdgpu_virt_init_data_exchange(adev);
c6332b97 1901 amdgpu_virt_release_full_gpu(adev, true);
d3c117e5 1902 }
c6332b97 1903
72d3f592 1904 return r;
d38ceaf9
AD
1905}
1906
e3ecdffa
AD
1907/**
1908 * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer
1909 *
1910 * @adev: amdgpu_device pointer
1911 *
1912 * Writes a reset magic value to the gart pointer in VRAM. The driver calls
1913 * this function before a GPU reset. If the value is retained after a
1914 * GPU reset, VRAM has not been lost. Some GPU resets may destry VRAM contents.
1915 */
06ec9070 1916static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev)
0c49e0b8
CZ
1917{
1918 memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
1919}
1920
e3ecdffa
AD
1921/**
1922 * amdgpu_device_check_vram_lost - check if vram is valid
1923 *
1924 * @adev: amdgpu_device pointer
1925 *
1926 * Checks the reset magic value written to the gart pointer in VRAM.
1927 * The driver calls this after a GPU reset to see if the contents of
1928 * VRAM is lost or now.
1929 * returns true if vram is lost, false if not.
1930 */
06ec9070 1931static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev)
0c49e0b8
CZ
1932{
1933 return !!memcmp(adev->gart.ptr, adev->reset_magic,
1934 AMDGPU_RESET_MAGIC_NUM);
1935}
1936
e3ecdffa 1937/**
1112a46b 1938 * amdgpu_device_set_cg_state - set clockgating for amdgpu device
e3ecdffa
AD
1939 *
1940 * @adev: amdgpu_device pointer
1941 *
e3ecdffa 1942 * The list of all the hardware IPs that make up the asic is walked and the
1112a46b
RZ
1943 * set_clockgating_state callbacks are run.
1944 * Late initialization pass enabling clockgating for hardware IPs.
1945 * Fini or suspend, pass disabling clockgating for hardware IPs.
e3ecdffa
AD
1946 * Returns 0 on success, negative error code on failure.
1947 */
fdd34271 1948
1112a46b
RZ
1949static int amdgpu_device_set_cg_state(struct amdgpu_device *adev,
1950 enum amd_clockgating_state state)
d38ceaf9 1951{
1112a46b 1952 int i, j, r;
d38ceaf9 1953
4a2ba394
SL
1954 if (amdgpu_emu_mode == 1)
1955 return 0;
1956
1112a46b
RZ
1957 for (j = 0; j < adev->num_ip_blocks; j++) {
1958 i = state == AMD_CG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
a2d31dc3 1959 if (!adev->ip_blocks[i].status.late_initialized)
d38ceaf9 1960 continue;
4a446d55 1961 /* skip CG for VCE/UVD, it's handled specially */
a1255107 1962 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
57716327 1963 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
34319b32 1964 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
57716327 1965 adev->ip_blocks[i].version->funcs->set_clockgating_state) {
4a446d55 1966 /* enable clockgating to save power */
a1255107 1967 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1112a46b 1968 state);
4a446d55
AD
1969 if (r) {
1970 DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
a1255107 1971 adev->ip_blocks[i].version->funcs->name, r);
4a446d55
AD
1972 return r;
1973 }
b0b00ff1 1974 }
d38ceaf9 1975 }
06b18f61 1976
c9f96fd5
RZ
1977 return 0;
1978}
1979
1112a46b 1980static int amdgpu_device_set_pg_state(struct amdgpu_device *adev, enum amd_powergating_state state)
c9f96fd5 1981{
1112a46b 1982 int i, j, r;
06b18f61 1983
c9f96fd5
RZ
1984 if (amdgpu_emu_mode == 1)
1985 return 0;
1986
1112a46b
RZ
1987 for (j = 0; j < adev->num_ip_blocks; j++) {
1988 i = state == AMD_PG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
a2d31dc3 1989 if (!adev->ip_blocks[i].status.late_initialized)
c9f96fd5
RZ
1990 continue;
1991 /* skip CG for VCE/UVD, it's handled specially */
1992 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1993 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1994 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
1995 adev->ip_blocks[i].version->funcs->set_powergating_state) {
1996 /* enable powergating to save power */
1997 r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev,
1112a46b 1998 state);
c9f96fd5
RZ
1999 if (r) {
2000 DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n",
2001 adev->ip_blocks[i].version->funcs->name, r);
2002 return r;
2003 }
2004 }
2005 }
2dc80b00
S
2006 return 0;
2007}
2008
beff74bc
AD
2009static int amdgpu_device_enable_mgpu_fan_boost(void)
2010{
2011 struct amdgpu_gpu_instance *gpu_ins;
2012 struct amdgpu_device *adev;
2013 int i, ret = 0;
2014
2015 mutex_lock(&mgpu_info.mutex);
2016
2017 /*
2018 * MGPU fan boost feature should be enabled
2019 * only when there are two or more dGPUs in
2020 * the system
2021 */
2022 if (mgpu_info.num_dgpu < 2)
2023 goto out;
2024
2025 for (i = 0; i < mgpu_info.num_dgpu; i++) {
2026 gpu_ins = &(mgpu_info.gpu_ins[i]);
2027 adev = gpu_ins->adev;
2028 if (!(adev->flags & AMD_IS_APU) &&
2029 !gpu_ins->mgpu_fan_enabled &&
2030 adev->powerplay.pp_funcs &&
2031 adev->powerplay.pp_funcs->enable_mgpu_fan_boost) {
2032 ret = amdgpu_dpm_enable_mgpu_fan_boost(adev);
2033 if (ret)
2034 break;
2035
2036 gpu_ins->mgpu_fan_enabled = 1;
2037 }
2038 }
2039
2040out:
2041 mutex_unlock(&mgpu_info.mutex);
2042
2043 return ret;
2044}
2045
e3ecdffa
AD
2046/**
2047 * amdgpu_device_ip_late_init - run late init for hardware IPs
2048 *
2049 * @adev: amdgpu_device pointer
2050 *
2051 * Late initialization pass for hardware IPs. The list of all the hardware
2052 * IPs that make up the asic is walked and the late_init callbacks are run.
2053 * late_init covers any special initialization that an IP requires
2054 * after all of the have been initialized or something that needs to happen
2055 * late in the init process.
2056 * Returns 0 on success, negative error code on failure.
2057 */
06ec9070 2058static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
2dc80b00
S
2059{
2060 int i = 0, r;
2061
2062 for (i = 0; i < adev->num_ip_blocks; i++) {
73f847db 2063 if (!adev->ip_blocks[i].status.hw)
2dc80b00
S
2064 continue;
2065 if (adev->ip_blocks[i].version->funcs->late_init) {
2066 r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
2067 if (r) {
2068 DRM_ERROR("late_init of IP block <%s> failed %d\n",
2069 adev->ip_blocks[i].version->funcs->name, r);
2070 return r;
2071 }
2dc80b00 2072 }
73f847db 2073 adev->ip_blocks[i].status.late_initialized = true;
2dc80b00
S
2074 }
2075
1112a46b
RZ
2076 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE);
2077 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE);
916ac57f 2078
06ec9070 2079 amdgpu_device_fill_reset_magic(adev);
d38ceaf9 2080
beff74bc
AD
2081 r = amdgpu_device_enable_mgpu_fan_boost();
2082 if (r)
2083 DRM_ERROR("enable mgpu fan boost failed (%d).\n", r);
2084
2085 /* set to low pstate by default */
2086 amdgpu_xgmi_set_pstate(adev, 0);
2087
d38ceaf9
AD
2088 return 0;
2089}
2090
e3ecdffa
AD
2091/**
2092 * amdgpu_device_ip_fini - run fini for hardware IPs
2093 *
2094 * @adev: amdgpu_device pointer
2095 *
2096 * Main teardown pass for hardware IPs. The list of all the hardware
2097 * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks
2098 * are run. hw_fini tears down the hardware associated with each IP
2099 * and sw_fini tears down any software state associated with each IP.
2100 * Returns 0 on success, negative error code on failure.
2101 */
06ec9070 2102static int amdgpu_device_ip_fini(struct amdgpu_device *adev)
d38ceaf9
AD
2103{
2104 int i, r;
2105
c030f2e4 2106 amdgpu_ras_pre_fini(adev);
2107
a82400b5
AG
2108 if (adev->gmc.xgmi.num_physical_nodes > 1)
2109 amdgpu_xgmi_remove_device(adev);
2110
1884734a 2111 amdgpu_amdkfd_device_fini(adev);
05df1f01
RZ
2112
2113 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
fdd34271
RZ
2114 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
2115
3e96dbfd
AD
2116 /* need to disable SMC first */
2117 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 2118 if (!adev->ip_blocks[i].status.hw)
3e96dbfd 2119 continue;
fdd34271 2120 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
a1255107 2121 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
3e96dbfd
AD
2122 /* XXX handle errors */
2123 if (r) {
2124 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
a1255107 2125 adev->ip_blocks[i].version->funcs->name, r);
3e96dbfd 2126 }
a1255107 2127 adev->ip_blocks[i].status.hw = false;
3e96dbfd
AD
2128 break;
2129 }
2130 }
2131
d38ceaf9 2132 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
a1255107 2133 if (!adev->ip_blocks[i].status.hw)
d38ceaf9 2134 continue;
8201a67a 2135
a1255107 2136 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
d38ceaf9 2137 /* XXX handle errors */
2c1a2784 2138 if (r) {
a1255107
AD
2139 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
2140 adev->ip_blocks[i].version->funcs->name, r);
2c1a2784 2141 }
8201a67a 2142
a1255107 2143 adev->ip_blocks[i].status.hw = false;
d38ceaf9
AD
2144 }
2145
9950cda2 2146
d38ceaf9 2147 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
a1255107 2148 if (!adev->ip_blocks[i].status.sw)
d38ceaf9 2149 continue;
c12aba3a
ML
2150
2151 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
c8963ea4 2152 amdgpu_ucode_free_bo(adev);
1e256e27 2153 amdgpu_free_static_csa(&adev->virt.csa_obj);
c12aba3a
ML
2154 amdgpu_device_wb_fini(adev);
2155 amdgpu_device_vram_scratch_fini(adev);
533aed27 2156 amdgpu_ib_pool_fini(adev);
c12aba3a
ML
2157 }
2158
a1255107 2159 r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
d38ceaf9 2160 /* XXX handle errors */
2c1a2784 2161 if (r) {
a1255107
AD
2162 DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
2163 adev->ip_blocks[i].version->funcs->name, r);
2c1a2784 2164 }
a1255107
AD
2165 adev->ip_blocks[i].status.sw = false;
2166 adev->ip_blocks[i].status.valid = false;
d38ceaf9
AD
2167 }
2168
a6dcfd9c 2169 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
a1255107 2170 if (!adev->ip_blocks[i].status.late_initialized)
8a2eef1d 2171 continue;
a1255107
AD
2172 if (adev->ip_blocks[i].version->funcs->late_fini)
2173 adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
2174 adev->ip_blocks[i].status.late_initialized = false;
a6dcfd9c
ML
2175 }
2176
c030f2e4 2177 amdgpu_ras_fini(adev);
2178
030308fc 2179 if (amdgpu_sriov_vf(adev))
24136135
ML
2180 if (amdgpu_virt_release_full_gpu(adev, false))
2181 DRM_ERROR("failed to release exclusive mode on fini\n");
2493664f 2182
d38ceaf9
AD
2183 return 0;
2184}
2185
e3ecdffa 2186/**
beff74bc 2187 * amdgpu_device_delayed_init_work_handler - work handler for IB tests
e3ecdffa 2188 *
1112a46b 2189 * @work: work_struct.
e3ecdffa 2190 */
beff74bc 2191static void amdgpu_device_delayed_init_work_handler(struct work_struct *work)
2dc80b00
S
2192{
2193 struct amdgpu_device *adev =
beff74bc 2194 container_of(work, struct amdgpu_device, delayed_init_work.work);
916ac57f
RZ
2195 int r;
2196
2197 r = amdgpu_ib_ring_tests(adev);
2198 if (r)
2199 DRM_ERROR("ib ring test failed (%d).\n", r);
2dc80b00
S
2200}
2201
1e317b99
RZ
2202static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work)
2203{
2204 struct amdgpu_device *adev =
2205 container_of(work, struct amdgpu_device, gfx.gfx_off_delay_work.work);
2206
2207 mutex_lock(&adev->gfx.gfx_off_mutex);
2208 if (!adev->gfx.gfx_off_state && !adev->gfx.gfx_off_req_count) {
2209 if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, true))
2210 adev->gfx.gfx_off_state = true;
2211 }
2212 mutex_unlock(&adev->gfx.gfx_off_mutex);
2213}
2214
e3ecdffa 2215/**
e7854a03 2216 * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1)
e3ecdffa
AD
2217 *
2218 * @adev: amdgpu_device pointer
2219 *
2220 * Main suspend function for hardware IPs. The list of all the hardware
2221 * IPs that make up the asic is walked, clockgating is disabled and the
2222 * suspend callbacks are run. suspend puts the hardware and software state
2223 * in each IP into a state suitable for suspend.
2224 * Returns 0 on success, negative error code on failure.
2225 */
e7854a03
AD
2226static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev)
2227{
2228 int i, r;
2229
05df1f01 2230 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
fdd34271 2231 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
05df1f01 2232
e7854a03
AD
2233 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2234 if (!adev->ip_blocks[i].status.valid)
2235 continue;
2236 /* displays are handled separately */
2237 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) {
e7854a03
AD
2238 /* XXX handle errors */
2239 r = adev->ip_blocks[i].version->funcs->suspend(adev);
2240 /* XXX handle errors */
2241 if (r) {
2242 DRM_ERROR("suspend of IP block <%s> failed %d\n",
2243 adev->ip_blocks[i].version->funcs->name, r);
482f0e53 2244 return r;
e7854a03 2245 }
482f0e53 2246 adev->ip_blocks[i].status.hw = false;
e7854a03
AD
2247 }
2248 }
2249
e7854a03
AD
2250 return 0;
2251}
2252
2253/**
2254 * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2)
2255 *
2256 * @adev: amdgpu_device pointer
2257 *
2258 * Main suspend function for hardware IPs. The list of all the hardware
2259 * IPs that make up the asic is walked, clockgating is disabled and the
2260 * suspend callbacks are run. suspend puts the hardware and software state
2261 * in each IP into a state suitable for suspend.
2262 * Returns 0 on success, negative error code on failure.
2263 */
2264static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev)
d38ceaf9
AD
2265{
2266 int i, r;
2267
2268 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
a1255107 2269 if (!adev->ip_blocks[i].status.valid)
d38ceaf9 2270 continue;
e7854a03
AD
2271 /* displays are handled in phase1 */
2272 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE)
2273 continue;
d38ceaf9 2274 /* XXX handle errors */
a1255107 2275 r = adev->ip_blocks[i].version->funcs->suspend(adev);
d38ceaf9 2276 /* XXX handle errors */
2c1a2784 2277 if (r) {
a1255107
AD
2278 DRM_ERROR("suspend of IP block <%s> failed %d\n",
2279 adev->ip_blocks[i].version->funcs->name, r);
2c1a2784 2280 }
876923fb 2281 adev->ip_blocks[i].status.hw = false;
a3a09142
AD
2282 /* handle putting the SMC in the appropriate state */
2283 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
2284 if (is_support_sw_smu(adev)) {
0e0b89c0 2285 r = smu_set_mp1_state(&adev->smu, adev->mp1_state);
a3a09142 2286 } else if (adev->powerplay.pp_funcs &&
482f0e53 2287 adev->powerplay.pp_funcs->set_mp1_state) {
a3a09142
AD
2288 r = adev->powerplay.pp_funcs->set_mp1_state(
2289 adev->powerplay.pp_handle,
2290 adev->mp1_state);
0e0b89c0
EQ
2291 }
2292 if (r) {
2293 DRM_ERROR("SMC failed to set mp1 state %d, %d\n",
2294 adev->mp1_state, r);
2295 return r;
a3a09142
AD
2296 }
2297 }
b5507c7e
AG
2298
2299 adev->ip_blocks[i].status.hw = false;
d38ceaf9
AD
2300 }
2301
2302 return 0;
2303}
2304
e7854a03
AD
2305/**
2306 * amdgpu_device_ip_suspend - run suspend for hardware IPs
2307 *
2308 * @adev: amdgpu_device pointer
2309 *
2310 * Main suspend function for hardware IPs. The list of all the hardware
2311 * IPs that make up the asic is walked, clockgating is disabled and the
2312 * suspend callbacks are run. suspend puts the hardware and software state
2313 * in each IP into a state suitable for suspend.
2314 * Returns 0 on success, negative error code on failure.
2315 */
2316int amdgpu_device_ip_suspend(struct amdgpu_device *adev)
2317{
2318 int r;
2319
e7819644
YT
2320 if (amdgpu_sriov_vf(adev))
2321 amdgpu_virt_request_full_gpu(adev, false);
2322
e7854a03
AD
2323 r = amdgpu_device_ip_suspend_phase1(adev);
2324 if (r)
2325 return r;
2326 r = amdgpu_device_ip_suspend_phase2(adev);
2327
e7819644
YT
2328 if (amdgpu_sriov_vf(adev))
2329 amdgpu_virt_release_full_gpu(adev, false);
2330
e7854a03
AD
2331 return r;
2332}
2333
06ec9070 2334static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
a90ad3c2
ML
2335{
2336 int i, r;
2337
2cb681b6
ML
2338 static enum amd_ip_block_type ip_order[] = {
2339 AMD_IP_BLOCK_TYPE_GMC,
2340 AMD_IP_BLOCK_TYPE_COMMON,
39186aef 2341 AMD_IP_BLOCK_TYPE_PSP,
2cb681b6
ML
2342 AMD_IP_BLOCK_TYPE_IH,
2343 };
a90ad3c2 2344
2cb681b6
ML
2345 for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
2346 int j;
2347 struct amdgpu_ip_block *block;
a90ad3c2 2348
2cb681b6
ML
2349 for (j = 0; j < adev->num_ip_blocks; j++) {
2350 block = &adev->ip_blocks[j];
2351
482f0e53 2352 block->status.hw = false;
2cb681b6
ML
2353 if (block->version->type != ip_order[i] ||
2354 !block->status.valid)
2355 continue;
2356
2357 r = block->version->funcs->hw_init(adev);
0aaeefcc 2358 DRM_INFO("RE-INIT-early: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
c41d1cf6
ML
2359 if (r)
2360 return r;
482f0e53 2361 block->status.hw = true;
a90ad3c2
ML
2362 }
2363 }
2364
2365 return 0;
2366}
2367
06ec9070 2368static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
a90ad3c2
ML
2369{
2370 int i, r;
2371
2cb681b6
ML
2372 static enum amd_ip_block_type ip_order[] = {
2373 AMD_IP_BLOCK_TYPE_SMC,
2374 AMD_IP_BLOCK_TYPE_DCE,
2375 AMD_IP_BLOCK_TYPE_GFX,
2376 AMD_IP_BLOCK_TYPE_SDMA,
257deb8c
FM
2377 AMD_IP_BLOCK_TYPE_UVD,
2378 AMD_IP_BLOCK_TYPE_VCE
2cb681b6 2379 };
a90ad3c2 2380
2cb681b6
ML
2381 for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
2382 int j;
2383 struct amdgpu_ip_block *block;
a90ad3c2 2384
2cb681b6
ML
2385 for (j = 0; j < adev->num_ip_blocks; j++) {
2386 block = &adev->ip_blocks[j];
2387
2388 if (block->version->type != ip_order[i] ||
482f0e53
ML
2389 !block->status.valid ||
2390 block->status.hw)
2cb681b6
ML
2391 continue;
2392
2393 r = block->version->funcs->hw_init(adev);
0aaeefcc 2394 DRM_INFO("RE-INIT-late: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
c41d1cf6
ML
2395 if (r)
2396 return r;
482f0e53 2397 block->status.hw = true;
a90ad3c2
ML
2398 }
2399 }
2400
2401 return 0;
2402}
2403
e3ecdffa
AD
2404/**
2405 * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs
2406 *
2407 * @adev: amdgpu_device pointer
2408 *
2409 * First resume function for hardware IPs. The list of all the hardware
2410 * IPs that make up the asic is walked and the resume callbacks are run for
2411 * COMMON, GMC, and IH. resume puts the hardware into a functional state
2412 * after a suspend and updates the software state as necessary. This
2413 * function is also used for restoring the GPU after a GPU reset.
2414 * Returns 0 on success, negative error code on failure.
2415 */
06ec9070 2416static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)
d38ceaf9
AD
2417{
2418 int i, r;
2419
a90ad3c2 2420 for (i = 0; i < adev->num_ip_blocks; i++) {
482f0e53 2421 if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
a90ad3c2 2422 continue;
a90ad3c2 2423 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
e3ecdffa
AD
2424 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2425 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
482f0e53 2426
fcf0649f
CZ
2427 r = adev->ip_blocks[i].version->funcs->resume(adev);
2428 if (r) {
2429 DRM_ERROR("resume of IP block <%s> failed %d\n",
2430 adev->ip_blocks[i].version->funcs->name, r);
2431 return r;
2432 }
482f0e53 2433 adev->ip_blocks[i].status.hw = true;
a90ad3c2
ML
2434 }
2435 }
2436
2437 return 0;
2438}
2439
e3ecdffa
AD
2440/**
2441 * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs
2442 *
2443 * @adev: amdgpu_device pointer
2444 *
2445 * First resume function for hardware IPs. The list of all the hardware
2446 * IPs that make up the asic is walked and the resume callbacks are run for
2447 * all blocks except COMMON, GMC, and IH. resume puts the hardware into a
2448 * functional state after a suspend and updates the software state as
2449 * necessary. This function is also used for restoring the GPU after a GPU
2450 * reset.
2451 * Returns 0 on success, negative error code on failure.
2452 */
06ec9070 2453static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)
d38ceaf9
AD
2454{
2455 int i, r;
2456
2457 for (i = 0; i < adev->num_ip_blocks; i++) {
482f0e53 2458 if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
d38ceaf9 2459 continue;
fcf0649f 2460 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
e3ecdffa 2461 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
7a3e0bb2
RZ
2462 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH ||
2463 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)
fcf0649f 2464 continue;
a1255107 2465 r = adev->ip_blocks[i].version->funcs->resume(adev);
2c1a2784 2466 if (r) {
a1255107
AD
2467 DRM_ERROR("resume of IP block <%s> failed %d\n",
2468 adev->ip_blocks[i].version->funcs->name, r);
d38ceaf9 2469 return r;
2c1a2784 2470 }
482f0e53 2471 adev->ip_blocks[i].status.hw = true;
d38ceaf9
AD
2472 }
2473
2474 return 0;
2475}
2476
e3ecdffa
AD
2477/**
2478 * amdgpu_device_ip_resume - run resume for hardware IPs
2479 *
2480 * @adev: amdgpu_device pointer
2481 *
2482 * Main resume function for hardware IPs. The hardware IPs
2483 * are split into two resume functions because they are
2484 * are also used in in recovering from a GPU reset and some additional
2485 * steps need to be take between them. In this case (S3/S4) they are
2486 * run sequentially.
2487 * Returns 0 on success, negative error code on failure.
2488 */
06ec9070 2489static int amdgpu_device_ip_resume(struct amdgpu_device *adev)
fcf0649f
CZ
2490{
2491 int r;
2492
06ec9070 2493 r = amdgpu_device_ip_resume_phase1(adev);
fcf0649f
CZ
2494 if (r)
2495 return r;
7a3e0bb2
RZ
2496
2497 r = amdgpu_device_fw_loading(adev);
2498 if (r)
2499 return r;
2500
06ec9070 2501 r = amdgpu_device_ip_resume_phase2(adev);
fcf0649f
CZ
2502
2503 return r;
2504}
2505
e3ecdffa
AD
2506/**
2507 * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV
2508 *
2509 * @adev: amdgpu_device pointer
2510 *
2511 * Query the VBIOS data tables to determine if the board supports SR-IOV.
2512 */
4e99a44e 2513static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
048765ad 2514{
6867e1b5
ML
2515 if (amdgpu_sriov_vf(adev)) {
2516 if (adev->is_atom_fw) {
2517 if (amdgpu_atomfirmware_gpu_supports_virtualization(adev))
2518 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2519 } else {
2520 if (amdgpu_atombios_has_gpu_virtualization_table(adev))
2521 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2522 }
2523
2524 if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
2525 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
a5bde2f9 2526 }
048765ad
AR
2527}
2528
e3ecdffa
AD
2529/**
2530 * amdgpu_device_asic_has_dc_support - determine if DC supports the asic
2531 *
2532 * @asic_type: AMD asic type
2533 *
2534 * Check if there is DC (new modesetting infrastructre) support for an asic.
2535 * returns true if DC has support, false if not.
2536 */
4562236b
HW
2537bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
2538{
2539 switch (asic_type) {
2540#if defined(CONFIG_DRM_AMD_DC)
2541 case CHIP_BONAIRE:
0d6fbccb 2542 case CHIP_KAVERI:
367e6687
AD
2543 case CHIP_KABINI:
2544 case CHIP_MULLINS:
d9fda248
HW
2545 /*
2546 * We have systems in the wild with these ASICs that require
2547 * LVDS and VGA support which is not supported with DC.
2548 *
2549 * Fallback to the non-DC driver here by default so as not to
2550 * cause regressions.
2551 */
2552 return amdgpu_dc > 0;
2553 case CHIP_HAWAII:
4562236b
HW
2554 case CHIP_CARRIZO:
2555 case CHIP_STONEY:
4562236b 2556 case CHIP_POLARIS10:
675fd32b 2557 case CHIP_POLARIS11:
2c8ad2d5 2558 case CHIP_POLARIS12:
675fd32b 2559 case CHIP_VEGAM:
4562236b
HW
2560 case CHIP_TONGA:
2561 case CHIP_FIJI:
42f8ffa1 2562 case CHIP_VEGA10:
dca7b401 2563 case CHIP_VEGA12:
c6034aa2 2564 case CHIP_VEGA20:
dc37a9a0 2565#if defined(CONFIG_DRM_AMD_DC_DCN1_0)
fd187853 2566 case CHIP_RAVEN:
b4f199c7
HW
2567#endif
2568#if defined(CONFIG_DRM_AMD_DC_DCN2_0)
2569 case CHIP_NAVI10:
8fceceb6 2570 case CHIP_NAVI14:
078655d9 2571 case CHIP_NAVI12:
e1c14c43
RL
2572#endif
2573#if defined(CONFIG_DRM_AMD_DC_DCN2_1)
2574 case CHIP_RENOIR:
42f8ffa1 2575#endif
fd187853 2576 return amdgpu_dc != 0;
4562236b
HW
2577#endif
2578 default:
2579 return false;
2580 }
2581}
2582
2583/**
2584 * amdgpu_device_has_dc_support - check if dc is supported
2585 *
2586 * @adev: amdgpu_device_pointer
2587 *
2588 * Returns true for supported, false for not supported
2589 */
2590bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
2591{
2555039d
XY
2592 if (amdgpu_sriov_vf(adev))
2593 return false;
2594
4562236b
HW
2595 return amdgpu_device_asic_has_dc_support(adev->asic_type);
2596}
2597
d4535e2c
AG
2598
2599static void amdgpu_device_xgmi_reset_func(struct work_struct *__work)
2600{
2601 struct amdgpu_device *adev =
2602 container_of(__work, struct amdgpu_device, xgmi_reset_work);
2603
2604 adev->asic_reset_res = amdgpu_asic_reset(adev);
2605 if (adev->asic_reset_res)
fed184e9 2606 DRM_WARN("ASIC reset failed with error, %d for drm dev, %s",
d4535e2c
AG
2607 adev->asic_reset_res, adev->ddev->unique);
2608}
2609
71f98027
AD
2610static int amdgpu_device_get_job_timeout_settings(struct amdgpu_device *adev)
2611{
2612 char *input = amdgpu_lockup_timeout;
2613 char *timeout_setting = NULL;
2614 int index = 0;
2615 long timeout;
2616 int ret = 0;
2617
2618 /*
2619 * By default timeout for non compute jobs is 10000.
2620 * And there is no timeout enforced on compute jobs.
2621 * In SR-IOV or passthrough mode, timeout for compute
2622 * jobs are 10000 by default.
2623 */
2624 adev->gfx_timeout = msecs_to_jiffies(10000);
2625 adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout;
2626 if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev))
2627 adev->compute_timeout = adev->gfx_timeout;
2628 else
2629 adev->compute_timeout = MAX_SCHEDULE_TIMEOUT;
2630
f440ff44 2631 if (strnlen(input, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) {
71f98027 2632 while ((timeout_setting = strsep(&input, ",")) &&
f440ff44 2633 strnlen(timeout_setting, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) {
71f98027
AD
2634 ret = kstrtol(timeout_setting, 0, &timeout);
2635 if (ret)
2636 return ret;
2637
2638 if (timeout == 0) {
2639 index++;
2640 continue;
2641 } else if (timeout < 0) {
2642 timeout = MAX_SCHEDULE_TIMEOUT;
2643 } else {
2644 timeout = msecs_to_jiffies(timeout);
2645 }
2646
2647 switch (index++) {
2648 case 0:
2649 adev->gfx_timeout = timeout;
2650 break;
2651 case 1:
2652 adev->compute_timeout = timeout;
2653 break;
2654 case 2:
2655 adev->sdma_timeout = timeout;
2656 break;
2657 case 3:
2658 adev->video_timeout = timeout;
2659 break;
2660 default:
2661 break;
2662 }
2663 }
2664 /*
2665 * There is only one value specified and
2666 * it should apply to all non-compute jobs.
2667 */
bcccee89 2668 if (index == 1) {
71f98027 2669 adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout;
bcccee89
ED
2670 if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev))
2671 adev->compute_timeout = adev->gfx_timeout;
2672 }
71f98027
AD
2673 }
2674
2675 return ret;
2676}
d4535e2c 2677
d38ceaf9
AD
2678/**
2679 * amdgpu_device_init - initialize the driver
2680 *
2681 * @adev: amdgpu_device pointer
87e3f136 2682 * @ddev: drm dev pointer
d38ceaf9
AD
2683 * @pdev: pci dev pointer
2684 * @flags: driver flags
2685 *
2686 * Initializes the driver info and hw (all asics).
2687 * Returns 0 for success or an error on failure.
2688 * Called at driver startup.
2689 */
2690int amdgpu_device_init(struct amdgpu_device *adev,
2691 struct drm_device *ddev,
2692 struct pci_dev *pdev,
2693 uint32_t flags)
2694{
2695 int r, i;
2696 bool runtime = false;
95844d20 2697 u32 max_MBps;
d38ceaf9
AD
2698
2699 adev->shutdown = false;
2700 adev->dev = &pdev->dev;
2701 adev->ddev = ddev;
2702 adev->pdev = pdev;
2703 adev->flags = flags;
4e66d7d2
YZ
2704
2705 if (amdgpu_force_asic_type >= 0 && amdgpu_force_asic_type < CHIP_LAST)
2706 adev->asic_type = amdgpu_force_asic_type;
2707 else
2708 adev->asic_type = flags & AMD_ASIC_MASK;
2709
d38ceaf9 2710 adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
593aa2d2
SL
2711 if (amdgpu_emu_mode == 1)
2712 adev->usec_timeout *= 2;
770d13b1 2713 adev->gmc.gart_size = 512 * 1024 * 1024;
d38ceaf9
AD
2714 adev->accel_working = false;
2715 adev->num_rings = 0;
2716 adev->mman.buffer_funcs = NULL;
2717 adev->mman.buffer_funcs_ring = NULL;
2718 adev->vm_manager.vm_pte_funcs = NULL;
3798e9a6 2719 adev->vm_manager.vm_pte_num_rqs = 0;
132f34e4 2720 adev->gmc.gmc_funcs = NULL;
f54d1867 2721 adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
b8866c26 2722 bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
d38ceaf9
AD
2723
2724 adev->smc_rreg = &amdgpu_invalid_rreg;
2725 adev->smc_wreg = &amdgpu_invalid_wreg;
2726 adev->pcie_rreg = &amdgpu_invalid_rreg;
2727 adev->pcie_wreg = &amdgpu_invalid_wreg;
36b9a952
HR
2728 adev->pciep_rreg = &amdgpu_invalid_rreg;
2729 adev->pciep_wreg = &amdgpu_invalid_wreg;
4fa1c6a6
TZ
2730 adev->pcie_rreg64 = &amdgpu_invalid_rreg64;
2731 adev->pcie_wreg64 = &amdgpu_invalid_wreg64;
d38ceaf9
AD
2732 adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
2733 adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
2734 adev->didt_rreg = &amdgpu_invalid_rreg;
2735 adev->didt_wreg = &amdgpu_invalid_wreg;
ccdbb20a
RZ
2736 adev->gc_cac_rreg = &amdgpu_invalid_rreg;
2737 adev->gc_cac_wreg = &amdgpu_invalid_wreg;
d38ceaf9
AD
2738 adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
2739 adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
2740
3e39ab90
AD
2741 DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
2742 amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
2743 pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
d38ceaf9
AD
2744
2745 /* mutex initialization are all done here so we
2746 * can recall function without having locking issues */
d38ceaf9 2747 atomic_set(&adev->irq.ih.lock, 0);
0e5ca0d1 2748 mutex_init(&adev->firmware.mutex);
d38ceaf9
AD
2749 mutex_init(&adev->pm.mutex);
2750 mutex_init(&adev->gfx.gpu_clock_mutex);
2751 mutex_init(&adev->srbm_mutex);
b8866c26 2752 mutex_init(&adev->gfx.pipe_reserve_mutex);
d23ee13f 2753 mutex_init(&adev->gfx.gfx_off_mutex);
d38ceaf9 2754 mutex_init(&adev->grbm_idx_mutex);
d38ceaf9 2755 mutex_init(&adev->mn_lock);
e23b74aa 2756 mutex_init(&adev->virt.vf_errors.lock);
d38ceaf9 2757 hash_init(adev->mn_hash);
13a752e3 2758 mutex_init(&adev->lock_reset);
bb5a2bdf 2759 mutex_init(&adev->virt.dpm_mutex);
32eaeae0 2760 mutex_init(&adev->psp.mutex);
d38ceaf9 2761
912dfc84
EQ
2762 r = amdgpu_device_check_arguments(adev);
2763 if (r)
2764 return r;
d38ceaf9 2765
d38ceaf9
AD
2766 spin_lock_init(&adev->mmio_idx_lock);
2767 spin_lock_init(&adev->smc_idx_lock);
2768 spin_lock_init(&adev->pcie_idx_lock);
2769 spin_lock_init(&adev->uvd_ctx_idx_lock);
2770 spin_lock_init(&adev->didt_idx_lock);
ccdbb20a 2771 spin_lock_init(&adev->gc_cac_idx_lock);
16abb5d2 2772 spin_lock_init(&adev->se_cac_idx_lock);
d38ceaf9 2773 spin_lock_init(&adev->audio_endpt_idx_lock);
95844d20 2774 spin_lock_init(&adev->mm_stats.lock);
d38ceaf9 2775
0c4e7fa5
CZ
2776 INIT_LIST_HEAD(&adev->shadow_list);
2777 mutex_init(&adev->shadow_list_lock);
2778
795f2813
AR
2779 INIT_LIST_HEAD(&adev->ring_lru_list);
2780 spin_lock_init(&adev->ring_lru_list_lock);
2781
beff74bc
AD
2782 INIT_DELAYED_WORK(&adev->delayed_init_work,
2783 amdgpu_device_delayed_init_work_handler);
1e317b99
RZ
2784 INIT_DELAYED_WORK(&adev->gfx.gfx_off_delay_work,
2785 amdgpu_device_delay_enable_gfx_off);
2dc80b00 2786
d4535e2c
AG
2787 INIT_WORK(&adev->xgmi_reset_work, amdgpu_device_xgmi_reset_func);
2788
d23ee13f 2789 adev->gfx.gfx_off_req_count = 1;
b1ddf548
RZ
2790 adev->pm.ac_power = power_supply_is_system_supplied() > 0 ? true : false;
2791
0fa49558
AX
2792 /* Registers mapping */
2793 /* TODO: block userspace mapping of io register */
da69c161
KW
2794 if (adev->asic_type >= CHIP_BONAIRE) {
2795 adev->rmmio_base = pci_resource_start(adev->pdev, 5);
2796 adev->rmmio_size = pci_resource_len(adev->pdev, 5);
2797 } else {
2798 adev->rmmio_base = pci_resource_start(adev->pdev, 2);
2799 adev->rmmio_size = pci_resource_len(adev->pdev, 2);
2800 }
d38ceaf9 2801
d38ceaf9
AD
2802 adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
2803 if (adev->rmmio == NULL) {
2804 return -ENOMEM;
2805 }
2806 DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
2807 DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
2808
d38ceaf9
AD
2809 /* io port mapping */
2810 for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
2811 if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) {
2812 adev->rio_mem_size = pci_resource_len(adev->pdev, i);
2813 adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size);
2814 break;
2815 }
2816 }
2817 if (adev->rio_mem == NULL)
b64a18c5 2818 DRM_INFO("PCI I/O BAR is not found.\n");
d38ceaf9 2819
b2109d8e
JX
2820 /* enable PCIE atomic ops */
2821 r = pci_enable_atomic_ops_to_root(adev->pdev,
2822 PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
2823 PCI_EXP_DEVCAP2_ATOMIC_COMP64);
2824 if (r) {
2825 adev->have_atomics_support = false;
2826 DRM_INFO("PCIE atomic ops is not supported\n");
2827 } else {
2828 adev->have_atomics_support = true;
2829 }
2830
5494d864
AD
2831 amdgpu_device_get_pcie_info(adev);
2832
b239c017
JX
2833 if (amdgpu_mcbp)
2834 DRM_INFO("MCBP is enabled\n");
2835
5f84cc63
JX
2836 if (amdgpu_mes && adev->asic_type >= CHIP_NAVI10)
2837 adev->enable_mes = true;
2838
f54eeab4 2839 if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10) {
a190d1c7
XY
2840 r = amdgpu_discovery_init(adev);
2841 if (r) {
2842 dev_err(adev->dev, "amdgpu_discovery_init failed\n");
2843 return r;
2844 }
2845 }
2846
d38ceaf9 2847 /* early init functions */
06ec9070 2848 r = amdgpu_device_ip_early_init(adev);
d38ceaf9
AD
2849 if (r)
2850 return r;
2851
df99ac0f
JZ
2852 r = amdgpu_device_get_job_timeout_settings(adev);
2853 if (r) {
2854 dev_err(adev->dev, "invalid lockup_timeout parameter syntax\n");
2855 return r;
2856 }
2857
6585661d
OZ
2858 /* doorbell bar mapping and doorbell index init*/
2859 amdgpu_device_doorbell_init(adev);
2860
d38ceaf9
AD
2861 /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
2862 /* this will fail for cards that aren't VGA class devices, just
2863 * ignore it */
06ec9070 2864 vga_client_register(adev->pdev, adev, NULL, amdgpu_device_vga_set_decode);
d38ceaf9 2865
e9bef455 2866 if (amdgpu_device_is_px(ddev))
d38ceaf9 2867 runtime = true;
84c8b22e
LW
2868 if (!pci_is_thunderbolt_attached(adev->pdev))
2869 vga_switcheroo_register_client(adev->pdev,
2870 &amdgpu_switcheroo_ops, runtime);
d38ceaf9
AD
2871 if (runtime)
2872 vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
2873
9475a943
SL
2874 if (amdgpu_emu_mode == 1) {
2875 /* post the asic on emulation mode */
2876 emu_soc_asic_init(adev);
bfca0289 2877 goto fence_driver_init;
9475a943 2878 }
bfca0289 2879
4e99a44e
ML
2880 /* detect if we are with an SRIOV vbios */
2881 amdgpu_device_detect_sriov_bios(adev);
048765ad 2882
95e8e59e
AD
2883 /* check if we need to reset the asic
2884 * E.g., driver was not cleanly unloaded previously, etc.
2885 */
f14899fd 2886 if (!amdgpu_sriov_vf(adev) && amdgpu_asic_need_reset_on_init(adev)) {
95e8e59e
AD
2887 r = amdgpu_asic_reset(adev);
2888 if (r) {
2889 dev_err(adev->dev, "asic reset on init failed\n");
2890 goto failed;
2891 }
2892 }
2893
d38ceaf9 2894 /* Post card if necessary */
39c640c0 2895 if (amdgpu_device_need_post(adev)) {
d38ceaf9 2896 if (!adev->bios) {
bec86378 2897 dev_err(adev->dev, "no vBIOS found\n");
83ba126a
AD
2898 r = -EINVAL;
2899 goto failed;
d38ceaf9 2900 }
bec86378 2901 DRM_INFO("GPU posting now...\n");
4e99a44e
ML
2902 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2903 if (r) {
2904 dev_err(adev->dev, "gpu post error!\n");
2905 goto failed;
2906 }
d38ceaf9
AD
2907 }
2908
88b64e95
AD
2909 if (adev->is_atom_fw) {
2910 /* Initialize clocks */
2911 r = amdgpu_atomfirmware_get_clock_info(adev);
2912 if (r) {
2913 dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
e23b74aa 2914 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
88b64e95
AD
2915 goto failed;
2916 }
2917 } else {
a5bde2f9
AD
2918 /* Initialize clocks */
2919 r = amdgpu_atombios_get_clock_info(adev);
2920 if (r) {
2921 dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
e23b74aa 2922 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
89041940 2923 goto failed;
a5bde2f9
AD
2924 }
2925 /* init i2c buses */
4562236b
HW
2926 if (!amdgpu_device_has_dc_support(adev))
2927 amdgpu_atombios_i2c_init(adev);
2c1a2784 2928 }
d38ceaf9 2929
bfca0289 2930fence_driver_init:
d38ceaf9
AD
2931 /* Fence driver */
2932 r = amdgpu_fence_driver_init(adev);
2c1a2784
AD
2933 if (r) {
2934 dev_err(adev->dev, "amdgpu_fence_driver_init failed\n");
e23b74aa 2935 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
83ba126a 2936 goto failed;
2c1a2784 2937 }
d38ceaf9
AD
2938
2939 /* init the mode config */
2940 drm_mode_config_init(adev->ddev);
2941
06ec9070 2942 r = amdgpu_device_ip_init(adev);
d38ceaf9 2943 if (r) {
8840a387 2944 /* failed in exclusive mode due to timeout */
2945 if (amdgpu_sriov_vf(adev) &&
2946 !amdgpu_sriov_runtime(adev) &&
2947 amdgpu_virt_mmio_blocked(adev) &&
2948 !amdgpu_virt_wait_reset(adev)) {
2949 dev_err(adev->dev, "VF exclusive mode timeout\n");
1daee8b4
PD
2950 /* Don't send request since VF is inactive. */
2951 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
2952 adev->virt.ops = NULL;
8840a387 2953 r = -EAGAIN;
2954 goto failed;
2955 }
06ec9070 2956 dev_err(adev->dev, "amdgpu_device_ip_init failed\n");
e23b74aa 2957 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
72d3f592
ED
2958 if (amdgpu_virt_request_full_gpu(adev, false))
2959 amdgpu_virt_release_full_gpu(adev, false);
83ba126a 2960 goto failed;
d38ceaf9
AD
2961 }
2962
2963 adev->accel_working = true;
2964
e59c0205
AX
2965 amdgpu_vm_check_compute_bug(adev);
2966
95844d20
MO
2967 /* Initialize the buffer migration limit. */
2968 if (amdgpu_moverate >= 0)
2969 max_MBps = amdgpu_moverate;
2970 else
2971 max_MBps = 8; /* Allow 8 MB/s. */
2972 /* Get a log2 for easy divisions. */
2973 adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
2974
9bc92b9c
ML
2975 amdgpu_fbdev_init(adev);
2976
e9bc1bf7
YT
2977 if (amdgpu_sriov_vf(adev) && amdgim_is_hwperf(adev))
2978 amdgpu_pm_virt_sysfs_init(adev);
2979
d2f52ac8
RZ
2980 r = amdgpu_pm_sysfs_init(adev);
2981 if (r)
2982 DRM_ERROR("registering pm debugfs failed (%d).\n", r);
2983
5bb23532
OM
2984 r = amdgpu_ucode_sysfs_init(adev);
2985 if (r)
2986 DRM_ERROR("Creating firmware sysfs failed (%d).\n", r);
2987
75758255 2988 r = amdgpu_debugfs_gem_init(adev);
3f14e623 2989 if (r)
d38ceaf9 2990 DRM_ERROR("registering gem debugfs failed (%d).\n", r);
d38ceaf9
AD
2991
2992 r = amdgpu_debugfs_regs_init(adev);
3f14e623 2993 if (r)
d38ceaf9 2994 DRM_ERROR("registering register debugfs failed (%d).\n", r);
d38ceaf9 2995
50ab2533 2996 r = amdgpu_debugfs_firmware_init(adev);
3f14e623 2997 if (r)
50ab2533 2998 DRM_ERROR("registering firmware debugfs failed (%d).\n", r);
50ab2533 2999
763efb6c 3000 r = amdgpu_debugfs_init(adev);
db95e218 3001 if (r)
763efb6c 3002 DRM_ERROR("Creating debugfs files failed (%d).\n", r);
db95e218 3003
d38ceaf9
AD
3004 if ((amdgpu_testing & 1)) {
3005 if (adev->accel_working)
3006 amdgpu_test_moves(adev);
3007 else
3008 DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n");
3009 }
d38ceaf9
AD
3010 if (amdgpu_benchmarking) {
3011 if (adev->accel_working)
3012 amdgpu_benchmark(adev, amdgpu_benchmarking);
3013 else
3014 DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n");
3015 }
3016
3017 /* enable clockgating, etc. after ib tests, etc. since some blocks require
3018 * explicit gating rather than handling it automatically.
3019 */
06ec9070 3020 r = amdgpu_device_ip_late_init(adev);
2c1a2784 3021 if (r) {
06ec9070 3022 dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n");
e23b74aa 3023 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
83ba126a 3024 goto failed;
2c1a2784 3025 }
d38ceaf9 3026
108c6a63 3027 /* must succeed. */
511fdbc3 3028 amdgpu_ras_resume(adev);
108c6a63 3029
beff74bc
AD
3030 queue_delayed_work(system_wq, &adev->delayed_init_work,
3031 msecs_to_jiffies(AMDGPU_RESUME_MS));
3032
dcea6e65
KR
3033 r = device_create_file(adev->dev, &dev_attr_pcie_replay_count);
3034 if (r) {
3035 dev_err(adev->dev, "Could not create pcie_replay_count");
3036 return r;
3037 }
108c6a63 3038
d155bef0
AB
3039 if (IS_ENABLED(CONFIG_PERF_EVENTS))
3040 r = amdgpu_pmu_init(adev);
9c7c85f7
JK
3041 if (r)
3042 dev_err(adev->dev, "amdgpu_pmu_init failed\n");
3043
d38ceaf9 3044 return 0;
83ba126a
AD
3045
3046failed:
89041940 3047 amdgpu_vf_error_trans_all(adev);
83ba126a
AD
3048 if (runtime)
3049 vga_switcheroo_fini_domain_pm_ops(adev->dev);
8840a387 3050
83ba126a 3051 return r;
d38ceaf9
AD
3052}
3053
d38ceaf9
AD
3054/**
3055 * amdgpu_device_fini - tear down the driver
3056 *
3057 * @adev: amdgpu_device pointer
3058 *
3059 * Tear down the driver info (all asics).
3060 * Called at driver shutdown.
3061 */
3062void amdgpu_device_fini(struct amdgpu_device *adev)
3063{
3064 int r;
3065
3066 DRM_INFO("amdgpu: finishing device.\n");
3067 adev->shutdown = true;
e5b03032
ML
3068 /* disable all interrupts */
3069 amdgpu_irq_disable_all(adev);
ff97cba8
ML
3070 if (adev->mode_info.mode_config_initialized){
3071 if (!amdgpu_device_has_dc_support(adev))
c2d88e06 3072 drm_helper_force_disable_all(adev->ddev);
ff97cba8
ML
3073 else
3074 drm_atomic_helper_shutdown(adev->ddev);
3075 }
d38ceaf9 3076 amdgpu_fence_driver_fini(adev);
58e955d9 3077 amdgpu_pm_sysfs_fini(adev);
d38ceaf9 3078 amdgpu_fbdev_fini(adev);
06ec9070 3079 r = amdgpu_device_ip_fini(adev);
ab4fe3e1
HR
3080 if (adev->firmware.gpu_info_fw) {
3081 release_firmware(adev->firmware.gpu_info_fw);
3082 adev->firmware.gpu_info_fw = NULL;
3083 }
d38ceaf9 3084 adev->accel_working = false;
beff74bc 3085 cancel_delayed_work_sync(&adev->delayed_init_work);
d38ceaf9 3086 /* free i2c buses */
4562236b
HW
3087 if (!amdgpu_device_has_dc_support(adev))
3088 amdgpu_i2c_fini(adev);
bfca0289
SL
3089
3090 if (amdgpu_emu_mode != 1)
3091 amdgpu_atombios_fini(adev);
3092
d38ceaf9
AD
3093 kfree(adev->bios);
3094 adev->bios = NULL;
84c8b22e
LW
3095 if (!pci_is_thunderbolt_attached(adev->pdev))
3096 vga_switcheroo_unregister_client(adev->pdev);
83ba126a
AD
3097 if (adev->flags & AMD_IS_PX)
3098 vga_switcheroo_fini_domain_pm_ops(adev->dev);
d38ceaf9
AD
3099 vga_client_register(adev->pdev, NULL, NULL, NULL);
3100 if (adev->rio_mem)
3101 pci_iounmap(adev->pdev, adev->rio_mem);
3102 adev->rio_mem = NULL;
3103 iounmap(adev->rmmio);
3104 adev->rmmio = NULL;
06ec9070 3105 amdgpu_device_doorbell_fini(adev);
e9bc1bf7
YT
3106 if (amdgpu_sriov_vf(adev) && amdgim_is_hwperf(adev))
3107 amdgpu_pm_virt_sysfs_fini(adev);
3108
d38ceaf9 3109 amdgpu_debugfs_regs_cleanup(adev);
dcea6e65 3110 device_remove_file(adev->dev, &dev_attr_pcie_replay_count);
5bb23532 3111 amdgpu_ucode_sysfs_fini(adev);
d155bef0
AB
3112 if (IS_ENABLED(CONFIG_PERF_EVENTS))
3113 amdgpu_pmu_fini(adev);
6698a3d0 3114 amdgpu_debugfs_preempt_cleanup(adev);
f54eeab4 3115 if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10)
a190d1c7 3116 amdgpu_discovery_fini(adev);
d38ceaf9
AD
3117}
3118
3119
3120/*
3121 * Suspend & resume.
3122 */
3123/**
810ddc3a 3124 * amdgpu_device_suspend - initiate device suspend
d38ceaf9 3125 *
87e3f136
DP
3126 * @dev: drm dev pointer
3127 * @suspend: suspend state
3128 * @fbcon : notify the fbdev of suspend
d38ceaf9
AD
3129 *
3130 * Puts the hw in the suspend state (all asics).
3131 * Returns 0 for success or an error on failure.
3132 * Called at driver suspend.
3133 */
810ddc3a 3134int amdgpu_device_suspend(struct drm_device *dev, bool suspend, bool fbcon)
d38ceaf9
AD
3135{
3136 struct amdgpu_device *adev;
3137 struct drm_crtc *crtc;
3138 struct drm_connector *connector;
f8d2d39e 3139 struct drm_connector_list_iter iter;
5ceb54c6 3140 int r;
d38ceaf9
AD
3141
3142 if (dev == NULL || dev->dev_private == NULL) {
3143 return -ENODEV;
3144 }
3145
3146 adev = dev->dev_private;
3147
3148 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
3149 return 0;
3150
44779b43 3151 adev->in_suspend = true;
d38ceaf9
AD
3152 drm_kms_helper_poll_disable(dev);
3153
5f818173
S
3154 if (fbcon)
3155 amdgpu_fbdev_set_suspend(adev, 1);
3156
beff74bc 3157 cancel_delayed_work_sync(&adev->delayed_init_work);
a5459475 3158
4562236b
HW
3159 if (!amdgpu_device_has_dc_support(adev)) {
3160 /* turn off display hw */
3161 drm_modeset_lock_all(dev);
f8d2d39e
LP
3162 drm_connector_list_iter_begin(dev, &iter);
3163 drm_for_each_connector_iter(connector, &iter)
3164 drm_helper_connector_dpms(connector,
3165 DRM_MODE_DPMS_OFF);
3166 drm_connector_list_iter_end(&iter);
4562236b 3167 drm_modeset_unlock_all(dev);
fe1053b7
AD
3168 /* unpin the front buffers and cursors */
3169 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
3170 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
3171 struct drm_framebuffer *fb = crtc->primary->fb;
3172 struct amdgpu_bo *robj;
3173
91334223 3174 if (amdgpu_crtc->cursor_bo && !adev->enable_virtual_display) {
fe1053b7
AD
3175 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
3176 r = amdgpu_bo_reserve(aobj, true);
3177 if (r == 0) {
3178 amdgpu_bo_unpin(aobj);
3179 amdgpu_bo_unreserve(aobj);
3180 }
756e6880 3181 }
756e6880 3182
fe1053b7
AD
3183 if (fb == NULL || fb->obj[0] == NULL) {
3184 continue;
3185 }
3186 robj = gem_to_amdgpu_bo(fb->obj[0]);
3187 /* don't unpin kernel fb objects */
3188 if (!amdgpu_fbdev_robj_is_fb(adev, robj)) {
3189 r = amdgpu_bo_reserve(robj, true);
3190 if (r == 0) {
3191 amdgpu_bo_unpin(robj);
3192 amdgpu_bo_unreserve(robj);
3193 }
d38ceaf9
AD
3194 }
3195 }
3196 }
fe1053b7
AD
3197
3198 amdgpu_amdkfd_suspend(adev);
3199
5e6932fe 3200 amdgpu_ras_suspend(adev);
3201
fe1053b7
AD
3202 r = amdgpu_device_ip_suspend_phase1(adev);
3203
d38ceaf9
AD
3204 /* evict vram memory */
3205 amdgpu_bo_evict_vram(adev);
3206
5ceb54c6 3207 amdgpu_fence_driver_suspend(adev);
d38ceaf9 3208
fe1053b7 3209 r = amdgpu_device_ip_suspend_phase2(adev);
d38ceaf9 3210
a0a71e49
AD
3211 /* evict remaining vram memory
3212 * This second call to evict vram is to evict the gart page table
3213 * using the CPU.
3214 */
d38ceaf9
AD
3215 amdgpu_bo_evict_vram(adev);
3216
d38ceaf9 3217 if (suspend) {
803cc26d 3218 pci_save_state(dev->pdev);
d38ceaf9
AD
3219 /* Shut down the device */
3220 pci_disable_device(dev->pdev);
3221 pci_set_power_state(dev->pdev, PCI_D3hot);
3222 }
3223
d38ceaf9
AD
3224 return 0;
3225}
3226
3227/**
810ddc3a 3228 * amdgpu_device_resume - initiate device resume
d38ceaf9 3229 *
87e3f136
DP
3230 * @dev: drm dev pointer
3231 * @resume: resume state
3232 * @fbcon : notify the fbdev of resume
d38ceaf9
AD
3233 *
3234 * Bring the hw back to operating state (all asics).
3235 * Returns 0 for success or an error on failure.
3236 * Called at driver resume.
3237 */
810ddc3a 3238int amdgpu_device_resume(struct drm_device *dev, bool resume, bool fbcon)
d38ceaf9
AD
3239{
3240 struct drm_connector *connector;
f8d2d39e 3241 struct drm_connector_list_iter iter;
d38ceaf9 3242 struct amdgpu_device *adev = dev->dev_private;
756e6880 3243 struct drm_crtc *crtc;
03161a6e 3244 int r = 0;
d38ceaf9
AD
3245
3246 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
3247 return 0;
3248
d38ceaf9
AD
3249 if (resume) {
3250 pci_set_power_state(dev->pdev, PCI_D0);
3251 pci_restore_state(dev->pdev);
74b0b157 3252 r = pci_enable_device(dev->pdev);
03161a6e 3253 if (r)
4d3b9ae5 3254 return r;
d38ceaf9
AD
3255 }
3256
3257 /* post card */
39c640c0 3258 if (amdgpu_device_need_post(adev)) {
74b0b157 3259 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
3260 if (r)
3261 DRM_ERROR("amdgpu asic init failed\n");
3262 }
d38ceaf9 3263
06ec9070 3264 r = amdgpu_device_ip_resume(adev);
e6707218 3265 if (r) {
06ec9070 3266 DRM_ERROR("amdgpu_device_ip_resume failed (%d).\n", r);
4d3b9ae5 3267 return r;
e6707218 3268 }
5ceb54c6
AD
3269 amdgpu_fence_driver_resume(adev);
3270
d38ceaf9 3271
06ec9070 3272 r = amdgpu_device_ip_late_init(adev);
03161a6e 3273 if (r)
4d3b9ae5 3274 return r;
d38ceaf9 3275
beff74bc
AD
3276 queue_delayed_work(system_wq, &adev->delayed_init_work,
3277 msecs_to_jiffies(AMDGPU_RESUME_MS));
3278
fe1053b7
AD
3279 if (!amdgpu_device_has_dc_support(adev)) {
3280 /* pin cursors */
3281 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
3282 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
3283
91334223 3284 if (amdgpu_crtc->cursor_bo && !adev->enable_virtual_display) {
fe1053b7
AD
3285 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
3286 r = amdgpu_bo_reserve(aobj, true);
3287 if (r == 0) {
3288 r = amdgpu_bo_pin(aobj, AMDGPU_GEM_DOMAIN_VRAM);
3289 if (r != 0)
3290 DRM_ERROR("Failed to pin cursor BO (%d)\n", r);
3291 amdgpu_crtc->cursor_addr = amdgpu_bo_gpu_offset(aobj);
3292 amdgpu_bo_unreserve(aobj);
3293 }
756e6880
AD
3294 }
3295 }
3296 }
ba997709
YZ
3297 r = amdgpu_amdkfd_resume(adev);
3298 if (r)
3299 return r;
756e6880 3300
96a5d8d4 3301 /* Make sure IB tests flushed */
beff74bc 3302 flush_delayed_work(&adev->delayed_init_work);
96a5d8d4 3303
d38ceaf9
AD
3304 /* blat the mode back in */
3305 if (fbcon) {
4562236b
HW
3306 if (!amdgpu_device_has_dc_support(adev)) {
3307 /* pre DCE11 */
3308 drm_helper_resume_force_mode(dev);
3309
3310 /* turn on display hw */
3311 drm_modeset_lock_all(dev);
f8d2d39e
LP
3312
3313 drm_connector_list_iter_begin(dev, &iter);
3314 drm_for_each_connector_iter(connector, &iter)
3315 drm_helper_connector_dpms(connector,
3316 DRM_MODE_DPMS_ON);
3317 drm_connector_list_iter_end(&iter);
3318
4562236b 3319 drm_modeset_unlock_all(dev);
d38ceaf9 3320 }
4d3b9ae5 3321 amdgpu_fbdev_set_suspend(adev, 0);
d38ceaf9
AD
3322 }
3323
3324 drm_kms_helper_poll_enable(dev);
23a1a9e5 3325
5e6932fe 3326 amdgpu_ras_resume(adev);
3327
23a1a9e5
L
3328 /*
3329 * Most of the connector probing functions try to acquire runtime pm
3330 * refs to ensure that the GPU is powered on when connector polling is
3331 * performed. Since we're calling this from a runtime PM callback,
3332 * trying to acquire rpm refs will cause us to deadlock.
3333 *
3334 * Since we're guaranteed to be holding the rpm lock, it's safe to
3335 * temporarily disable the rpm helpers so this doesn't deadlock us.
3336 */
3337#ifdef CONFIG_PM
3338 dev->dev->power.disable_depth++;
3339#endif
4562236b
HW
3340 if (!amdgpu_device_has_dc_support(adev))
3341 drm_helper_hpd_irq_event(dev);
3342 else
3343 drm_kms_helper_hotplug_event(dev);
23a1a9e5
L
3344#ifdef CONFIG_PM
3345 dev->dev->power.disable_depth--;
3346#endif
44779b43
RZ
3347 adev->in_suspend = false;
3348
4d3b9ae5 3349 return 0;
d38ceaf9
AD
3350}
3351
e3ecdffa
AD
3352/**
3353 * amdgpu_device_ip_check_soft_reset - did soft reset succeed
3354 *
3355 * @adev: amdgpu_device pointer
3356 *
3357 * The list of all the hardware IPs that make up the asic is walked and
3358 * the check_soft_reset callbacks are run. check_soft_reset determines
3359 * if the asic is still hung or not.
3360 * Returns true if any of the IPs are still in a hung state, false if not.
3361 */
06ec9070 3362static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)
63fbf42f
CZ
3363{
3364 int i;
3365 bool asic_hang = false;
3366
f993d628
ML
3367 if (amdgpu_sriov_vf(adev))
3368 return true;
3369
8bc04c29
AD
3370 if (amdgpu_asic_need_full_reset(adev))
3371 return true;
3372
63fbf42f 3373 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 3374 if (!adev->ip_blocks[i].status.valid)
63fbf42f 3375 continue;
a1255107
AD
3376 if (adev->ip_blocks[i].version->funcs->check_soft_reset)
3377 adev->ip_blocks[i].status.hang =
3378 adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
3379 if (adev->ip_blocks[i].status.hang) {
3380 DRM_INFO("IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
63fbf42f
CZ
3381 asic_hang = true;
3382 }
3383 }
3384 return asic_hang;
3385}
3386
e3ecdffa
AD
3387/**
3388 * amdgpu_device_ip_pre_soft_reset - prepare for soft reset
3389 *
3390 * @adev: amdgpu_device pointer
3391 *
3392 * The list of all the hardware IPs that make up the asic is walked and the
3393 * pre_soft_reset callbacks are run if the block is hung. pre_soft_reset
3394 * handles any IP specific hardware or software state changes that are
3395 * necessary for a soft reset to succeed.
3396 * Returns 0 on success, negative error code on failure.
3397 */
06ec9070 3398static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)
d31a501e
CZ
3399{
3400 int i, r = 0;
3401
3402 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 3403 if (!adev->ip_blocks[i].status.valid)
d31a501e 3404 continue;
a1255107
AD
3405 if (adev->ip_blocks[i].status.hang &&
3406 adev->ip_blocks[i].version->funcs->pre_soft_reset) {
3407 r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
d31a501e
CZ
3408 if (r)
3409 return r;
3410 }
3411 }
3412
3413 return 0;
3414}
3415
e3ecdffa
AD
3416/**
3417 * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed
3418 *
3419 * @adev: amdgpu_device pointer
3420 *
3421 * Some hardware IPs cannot be soft reset. If they are hung, a full gpu
3422 * reset is necessary to recover.
3423 * Returns true if a full asic reset is required, false if not.
3424 */
06ec9070 3425static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)
35d782fe 3426{
da146d3b
AD
3427 int i;
3428
8bc04c29
AD
3429 if (amdgpu_asic_need_full_reset(adev))
3430 return true;
3431
da146d3b 3432 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 3433 if (!adev->ip_blocks[i].status.valid)
da146d3b 3434 continue;
a1255107
AD
3435 if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
3436 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
3437 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
98512bb8
KW
3438 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
3439 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
a1255107 3440 if (adev->ip_blocks[i].status.hang) {
da146d3b
AD
3441 DRM_INFO("Some block need full reset!\n");
3442 return true;
3443 }
3444 }
35d782fe
CZ
3445 }
3446 return false;
3447}
3448
e3ecdffa
AD
3449/**
3450 * amdgpu_device_ip_soft_reset - do a soft reset
3451 *
3452 * @adev: amdgpu_device pointer
3453 *
3454 * The list of all the hardware IPs that make up the asic is walked and the
3455 * soft_reset callbacks are run if the block is hung. soft_reset handles any
3456 * IP specific hardware or software state changes that are necessary to soft
3457 * reset the IP.
3458 * Returns 0 on success, negative error code on failure.
3459 */
06ec9070 3460static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)
35d782fe
CZ
3461{
3462 int i, r = 0;
3463
3464 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 3465 if (!adev->ip_blocks[i].status.valid)
35d782fe 3466 continue;
a1255107
AD
3467 if (adev->ip_blocks[i].status.hang &&
3468 adev->ip_blocks[i].version->funcs->soft_reset) {
3469 r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
35d782fe
CZ
3470 if (r)
3471 return r;
3472 }
3473 }
3474
3475 return 0;
3476}
3477
e3ecdffa
AD
3478/**
3479 * amdgpu_device_ip_post_soft_reset - clean up from soft reset
3480 *
3481 * @adev: amdgpu_device pointer
3482 *
3483 * The list of all the hardware IPs that make up the asic is walked and the
3484 * post_soft_reset callbacks are run if the asic was hung. post_soft_reset
3485 * handles any IP specific hardware or software state changes that are
3486 * necessary after the IP has been soft reset.
3487 * Returns 0 on success, negative error code on failure.
3488 */
06ec9070 3489static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)
35d782fe
CZ
3490{
3491 int i, r = 0;
3492
3493 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 3494 if (!adev->ip_blocks[i].status.valid)
35d782fe 3495 continue;
a1255107
AD
3496 if (adev->ip_blocks[i].status.hang &&
3497 adev->ip_blocks[i].version->funcs->post_soft_reset)
3498 r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
35d782fe
CZ
3499 if (r)
3500 return r;
3501 }
3502
3503 return 0;
3504}
3505
e3ecdffa 3506/**
c33adbc7 3507 * amdgpu_device_recover_vram - Recover some VRAM contents
e3ecdffa
AD
3508 *
3509 * @adev: amdgpu_device pointer
3510 *
3511 * Restores the contents of VRAM buffers from the shadows in GTT. Used to
3512 * restore things like GPUVM page tables after a GPU reset where
3513 * the contents of VRAM might be lost.
403009bf
CK
3514 *
3515 * Returns:
3516 * 0 on success, negative error code on failure.
e3ecdffa 3517 */
c33adbc7 3518static int amdgpu_device_recover_vram(struct amdgpu_device *adev)
c41d1cf6 3519{
c41d1cf6 3520 struct dma_fence *fence = NULL, *next = NULL;
403009bf
CK
3521 struct amdgpu_bo *shadow;
3522 long r = 1, tmo;
c41d1cf6
ML
3523
3524 if (amdgpu_sriov_runtime(adev))
b045d3af 3525 tmo = msecs_to_jiffies(8000);
c41d1cf6
ML
3526 else
3527 tmo = msecs_to_jiffies(100);
3528
3529 DRM_INFO("recover vram bo from shadow start\n");
3530 mutex_lock(&adev->shadow_list_lock);
403009bf
CK
3531 list_for_each_entry(shadow, &adev->shadow_list, shadow_list) {
3532
3533 /* No need to recover an evicted BO */
3534 if (shadow->tbo.mem.mem_type != TTM_PL_TT ||
b575f10d 3535 shadow->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET ||
403009bf
CK
3536 shadow->parent->tbo.mem.mem_type != TTM_PL_VRAM)
3537 continue;
3538
3539 r = amdgpu_bo_restore_shadow(shadow, &next);
3540 if (r)
3541 break;
3542
c41d1cf6 3543 if (fence) {
1712fb1a 3544 tmo = dma_fence_wait_timeout(fence, false, tmo);
403009bf
CK
3545 dma_fence_put(fence);
3546 fence = next;
1712fb1a 3547 if (tmo == 0) {
3548 r = -ETIMEDOUT;
c41d1cf6 3549 break;
1712fb1a 3550 } else if (tmo < 0) {
3551 r = tmo;
3552 break;
3553 }
403009bf
CK
3554 } else {
3555 fence = next;
c41d1cf6 3556 }
c41d1cf6
ML
3557 }
3558 mutex_unlock(&adev->shadow_list_lock);
3559
403009bf
CK
3560 if (fence)
3561 tmo = dma_fence_wait_timeout(fence, false, tmo);
c41d1cf6
ML
3562 dma_fence_put(fence);
3563
1712fb1a 3564 if (r < 0 || tmo <= 0) {
3565 DRM_ERROR("recover vram bo from shadow failed, r is %ld, tmo is %ld\n", r, tmo);
403009bf
CK
3566 return -EIO;
3567 }
c41d1cf6 3568
403009bf
CK
3569 DRM_INFO("recover vram bo from shadow done\n");
3570 return 0;
c41d1cf6
ML
3571}
3572
a90ad3c2 3573
e3ecdffa 3574/**
06ec9070 3575 * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
5740682e
ML
3576 *
3577 * @adev: amdgpu device pointer
87e3f136 3578 * @from_hypervisor: request from hypervisor
5740682e
ML
3579 *
3580 * do VF FLR and reinitialize Asic
3f48c681 3581 * return 0 means succeeded otherwise failed
e3ecdffa
AD
3582 */
3583static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
3584 bool from_hypervisor)
5740682e
ML
3585{
3586 int r;
3587
3588 if (from_hypervisor)
3589 r = amdgpu_virt_request_full_gpu(adev, true);
3590 else
3591 r = amdgpu_virt_reset_gpu(adev);
3592 if (r)
3593 return r;
a90ad3c2 3594
f81e8d53
WL
3595 amdgpu_amdkfd_pre_reset(adev);
3596
a90ad3c2 3597 /* Resume IP prior to SMC */
06ec9070 3598 r = amdgpu_device_ip_reinit_early_sriov(adev);
5740682e
ML
3599 if (r)
3600 goto error;
a90ad3c2
ML
3601
3602 /* we need recover gart prior to run SMC/CP/SDMA resume */
c1c7ce8f 3603 amdgpu_gtt_mgr_recover(&adev->mman.bdev.man[TTM_PL_TT]);
a90ad3c2 3604
7a3e0bb2
RZ
3605 r = amdgpu_device_fw_loading(adev);
3606 if (r)
3607 return r;
3608
a90ad3c2 3609 /* now we are okay to resume SMC/CP/SDMA */
06ec9070 3610 r = amdgpu_device_ip_reinit_late_sriov(adev);
5740682e
ML
3611 if (r)
3612 goto error;
a90ad3c2
ML
3613
3614 amdgpu_irq_gpu_reset_resume_helper(adev);
5740682e 3615 r = amdgpu_ib_ring_tests(adev);
f81e8d53 3616 amdgpu_amdkfd_post_reset(adev);
a90ad3c2 3617
abc34253 3618error:
d3c117e5 3619 amdgpu_virt_init_data_exchange(adev);
abc34253 3620 amdgpu_virt_release_full_gpu(adev, true);
c41d1cf6 3621 if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
e3526257 3622 amdgpu_inc_vram_lost(adev);
c33adbc7 3623 r = amdgpu_device_recover_vram(adev);
a90ad3c2
ML
3624 }
3625
3626 return r;
3627}
3628
12938fad
CK
3629/**
3630 * amdgpu_device_should_recover_gpu - check if we should try GPU recovery
3631 *
3632 * @adev: amdgpu device pointer
3633 *
3634 * Check amdgpu_gpu_recovery and SRIOV status to see if we should try to recover
3635 * a hung GPU.
3636 */
3637bool amdgpu_device_should_recover_gpu(struct amdgpu_device *adev)
3638{
3639 if (!amdgpu_device_ip_check_soft_reset(adev)) {
3640 DRM_INFO("Timeout, but no hardware hang detected.\n");
3641 return false;
3642 }
3643
3ba7b418
AG
3644 if (amdgpu_gpu_recovery == 0)
3645 goto disabled;
3646
3647 if (amdgpu_sriov_vf(adev))
3648 return true;
3649
3650 if (amdgpu_gpu_recovery == -1) {
3651 switch (adev->asic_type) {
fc42d47c
AG
3652 case CHIP_BONAIRE:
3653 case CHIP_HAWAII:
3ba7b418
AG
3654 case CHIP_TOPAZ:
3655 case CHIP_TONGA:
3656 case CHIP_FIJI:
3657 case CHIP_POLARIS10:
3658 case CHIP_POLARIS11:
3659 case CHIP_POLARIS12:
3660 case CHIP_VEGAM:
3661 case CHIP_VEGA20:
3662 case CHIP_VEGA10:
3663 case CHIP_VEGA12:
c43b849f 3664 case CHIP_RAVEN:
3ba7b418
AG
3665 break;
3666 default:
3667 goto disabled;
3668 }
12938fad
CK
3669 }
3670
3671 return true;
3ba7b418
AG
3672
3673disabled:
3674 DRM_INFO("GPU recovery disabled.\n");
3675 return false;
12938fad
CK
3676}
3677
5c6dd71e 3678
26bc5340
AG
3679static int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,
3680 struct amdgpu_job *job,
3681 bool *need_full_reset_arg)
3682{
3683 int i, r = 0;
3684 bool need_full_reset = *need_full_reset_arg;
71182665 3685
71182665 3686 /* block all schedulers and reset given job's ring */
0875dc9e
CZ
3687 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3688 struct amdgpu_ring *ring = adev->rings[i];
3689
51687759 3690 if (!ring || !ring->sched.thread)
0875dc9e 3691 continue;
5740682e 3692
2f9d4084
ML
3693 /* after all hw jobs are reset, hw fence is meaningless, so force_completion */
3694 amdgpu_fence_driver_force_completion(ring);
0875dc9e 3695 }
d38ceaf9 3696
222b5f04
AG
3697 if(job)
3698 drm_sched_increase_karma(&job->base);
3699
1d721ed6 3700 /* Don't suspend on bare metal if we are not going to HW reset the ASIC */
26bc5340
AG
3701 if (!amdgpu_sriov_vf(adev)) {
3702
3703 if (!need_full_reset)
3704 need_full_reset = amdgpu_device_ip_need_full_reset(adev);
3705
3706 if (!need_full_reset) {
3707 amdgpu_device_ip_pre_soft_reset(adev);
3708 r = amdgpu_device_ip_soft_reset(adev);
3709 amdgpu_device_ip_post_soft_reset(adev);
3710 if (r || amdgpu_device_ip_check_soft_reset(adev)) {
3711 DRM_INFO("soft reset failed, will fallback to full reset!\n");
3712 need_full_reset = true;
3713 }
3714 }
3715
3716 if (need_full_reset)
3717 r = amdgpu_device_ip_suspend(adev);
3718
3719 *need_full_reset_arg = need_full_reset;
3720 }
3721
3722 return r;
3723}
3724
3725static int amdgpu_do_asic_reset(struct amdgpu_hive_info *hive,
3726 struct list_head *device_list_handle,
3727 bool *need_full_reset_arg)
3728{
3729 struct amdgpu_device *tmp_adev = NULL;
3730 bool need_full_reset = *need_full_reset_arg, vram_lost = false;
3731 int r = 0;
3732
3733 /*
3734 * ASIC reset has to be done on all HGMI hive nodes ASAP
3735 * to allow proper links negotiation in FW (within 1 sec)
3736 */
3737 if (need_full_reset) {
3738 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
d4535e2c
AG
3739 /* For XGMI run all resets in parallel to speed up the process */
3740 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
3741 if (!queue_work(system_highpri_wq, &tmp_adev->xgmi_reset_work))
3742 r = -EALREADY;
3743 } else
3744 r = amdgpu_asic_reset(tmp_adev);
3745
3746 if (r) {
fed184e9 3747 DRM_ERROR("ASIC reset failed with error, %d for drm dev, %s",
26bc5340 3748 r, tmp_adev->ddev->unique);
d4535e2c
AG
3749 break;
3750 }
3751 }
3752
3753 /* For XGMI wait for all PSP resets to complete before proceed */
3754 if (!r) {
3755 list_for_each_entry(tmp_adev, device_list_handle,
3756 gmc.xgmi.head) {
3757 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
3758 flush_work(&tmp_adev->xgmi_reset_work);
3759 r = tmp_adev->asic_reset_res;
3760 if (r)
3761 break;
3762 }
3763 }
26bc5340
AG
3764 }
3765 }
3766
3767
3768 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
3769 if (need_full_reset) {
3770 /* post card */
3771 if (amdgpu_atom_asic_init(tmp_adev->mode_info.atom_context))
3772 DRM_WARN("asic atom init failed!");
3773
3774 if (!r) {
3775 dev_info(tmp_adev->dev, "GPU reset succeeded, trying to resume\n");
3776 r = amdgpu_device_ip_resume_phase1(tmp_adev);
3777 if (r)
3778 goto out;
3779
3780 vram_lost = amdgpu_device_check_vram_lost(tmp_adev);
3781 if (vram_lost) {
77e7f829 3782 DRM_INFO("VRAM is lost due to GPU reset!\n");
e3526257 3783 amdgpu_inc_vram_lost(tmp_adev);
26bc5340
AG
3784 }
3785
3786 r = amdgpu_gtt_mgr_recover(
3787 &tmp_adev->mman.bdev.man[TTM_PL_TT]);
3788 if (r)
3789 goto out;
3790
3791 r = amdgpu_device_fw_loading(tmp_adev);
3792 if (r)
3793 return r;
3794
3795 r = amdgpu_device_ip_resume_phase2(tmp_adev);
3796 if (r)
3797 goto out;
3798
3799 if (vram_lost)
3800 amdgpu_device_fill_reset_magic(tmp_adev);
3801
fdafb359
EQ
3802 /*
3803 * Add this ASIC as tracked as reset was already
3804 * complete successfully.
3805 */
3806 amdgpu_register_gpu_instance(tmp_adev);
3807
7c04ca50 3808 r = amdgpu_device_ip_late_init(tmp_adev);
3809 if (r)
3810 goto out;
3811
e79a04d5 3812 /* must succeed. */
511fdbc3 3813 amdgpu_ras_resume(tmp_adev);
e79a04d5 3814
26bc5340
AG
3815 /* Update PSP FW topology after reset */
3816 if (hive && tmp_adev->gmc.xgmi.num_physical_nodes > 1)
3817 r = amdgpu_xgmi_update_topology(hive, tmp_adev);
3818 }
3819 }
3820
3821
3822out:
3823 if (!r) {
3824 amdgpu_irq_gpu_reset_resume_helper(tmp_adev);
3825 r = amdgpu_ib_ring_tests(tmp_adev);
3826 if (r) {
3827 dev_err(tmp_adev->dev, "ib ring test failed (%d).\n", r);
3828 r = amdgpu_device_ip_suspend(tmp_adev);
3829 need_full_reset = true;
3830 r = -EAGAIN;
3831 goto end;
3832 }
3833 }
3834
3835 if (!r)
3836 r = amdgpu_device_recover_vram(tmp_adev);
3837 else
3838 tmp_adev->asic_reset_res = r;
3839 }
3840
3841end:
3842 *need_full_reset_arg = need_full_reset;
3843 return r;
3844}
3845
1d721ed6 3846static bool amdgpu_device_lock_adev(struct amdgpu_device *adev, bool trylock)
26bc5340 3847{
1d721ed6
AG
3848 if (trylock) {
3849 if (!mutex_trylock(&adev->lock_reset))
3850 return false;
3851 } else
3852 mutex_lock(&adev->lock_reset);
5740682e 3853
26bc5340
AG
3854 atomic_inc(&adev->gpu_reset_counter);
3855 adev->in_gpu_reset = 1;
a3a09142
AD
3856 switch (amdgpu_asic_reset_method(adev)) {
3857 case AMD_RESET_METHOD_MODE1:
3858 adev->mp1_state = PP_MP1_STATE_SHUTDOWN;
3859 break;
3860 case AMD_RESET_METHOD_MODE2:
3861 adev->mp1_state = PP_MP1_STATE_RESET;
3862 break;
3863 default:
3864 adev->mp1_state = PP_MP1_STATE_NONE;
3865 break;
3866 }
1d721ed6
AG
3867
3868 return true;
26bc5340 3869}
d38ceaf9 3870
26bc5340
AG
3871static void amdgpu_device_unlock_adev(struct amdgpu_device *adev)
3872{
89041940 3873 amdgpu_vf_error_trans_all(adev);
a3a09142 3874 adev->mp1_state = PP_MP1_STATE_NONE;
13a752e3
ML
3875 adev->in_gpu_reset = 0;
3876 mutex_unlock(&adev->lock_reset);
26bc5340
AG
3877}
3878
26bc5340
AG
3879/**
3880 * amdgpu_device_gpu_recover - reset the asic and recover scheduler
3881 *
3882 * @adev: amdgpu device pointer
3883 * @job: which job trigger hang
3884 *
3885 * Attempt to reset the GPU if it has hung (all asics).
3886 * Attempt to do soft-reset or full-reset and reinitialize Asic
3887 * Returns 0 for success or an error on failure.
3888 */
3889
3890int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
3891 struct amdgpu_job *job)
3892{
1d721ed6
AG
3893 struct list_head device_list, *device_list_handle = NULL;
3894 bool need_full_reset, job_signaled;
26bc5340 3895 struct amdgpu_hive_info *hive = NULL;
26bc5340 3896 struct amdgpu_device *tmp_adev = NULL;
1d721ed6 3897 int i, r = 0;
7c6e68c7 3898 bool in_ras_intr = amdgpu_ras_intr_triggered();
26bc5340 3899
d5ea093e
AG
3900 /*
3901 * Flush RAM to disk so that after reboot
3902 * the user can read log and see why the system rebooted.
3903 */
3904 if (in_ras_intr && amdgpu_ras_get_context(adev)->reboot) {
3905
3906 DRM_WARN("Emergency reboot.");
3907
3908 ksys_sync_helper();
3909 emergency_restart();
3910 }
3911
1d721ed6 3912 need_full_reset = job_signaled = false;
26bc5340
AG
3913 INIT_LIST_HEAD(&device_list);
3914
7c6e68c7 3915 dev_info(adev->dev, "GPU %s begin!\n", in_ras_intr ? "jobs stop":"reset");
26bc5340 3916
beff74bc 3917 cancel_delayed_work_sync(&adev->delayed_init_work);
c53e4db7 3918
1d721ed6
AG
3919 hive = amdgpu_get_xgmi_hive(adev, false);
3920
26bc5340 3921 /*
1d721ed6
AG
3922 * Here we trylock to avoid chain of resets executing from
3923 * either trigger by jobs on different adevs in XGMI hive or jobs on
3924 * different schedulers for same device while this TO handler is running.
3925 * We always reset all schedulers for device and all devices for XGMI
3926 * hive so that should take care of them too.
26bc5340 3927 */
1d721ed6
AG
3928
3929 if (hive && !mutex_trylock(&hive->reset_lock)) {
3930 DRM_INFO("Bailing on TDR for s_job:%llx, hive: %llx as another already in progress",
0b2d2c2e 3931 job ? job->base.id : -1, hive->hive_id);
26bc5340 3932 return 0;
1d721ed6 3933 }
26bc5340
AG
3934
3935 /* Start with adev pre asic reset first for soft reset check.*/
1d721ed6
AG
3936 if (!amdgpu_device_lock_adev(adev, !hive)) {
3937 DRM_INFO("Bailing on TDR for s_job:%llx, as another already in progress",
0b2d2c2e 3938 job ? job->base.id : -1);
1d721ed6 3939 return 0;
26bc5340
AG
3940 }
3941
7c6e68c7
AG
3942 /* Block kfd: SRIOV would do it separately */
3943 if (!amdgpu_sriov_vf(adev))
3944 amdgpu_amdkfd_pre_reset(adev);
3945
26bc5340 3946 /* Build list of devices to reset */
1d721ed6 3947 if (adev->gmc.xgmi.num_physical_nodes > 1) {
26bc5340 3948 if (!hive) {
7c6e68c7
AG
3949 /*unlock kfd: SRIOV would do it separately */
3950 if (!amdgpu_sriov_vf(adev))
3951 amdgpu_amdkfd_post_reset(adev);
26bc5340
AG
3952 amdgpu_device_unlock_adev(adev);
3953 return -ENODEV;
3954 }
3955
3956 /*
3957 * In case we are in XGMI hive mode device reset is done for all the
3958 * nodes in the hive to retrain all XGMI links and hence the reset
3959 * sequence is executed in loop on all nodes.
3960 */
3961 device_list_handle = &hive->device_list;
3962 } else {
3963 list_add_tail(&adev->gmc.xgmi.head, &device_list);
3964 device_list_handle = &device_list;
3965 }
3966
1d721ed6
AG
3967 /* block all schedulers and reset given job's ring */
3968 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
7c6e68c7 3969 if (tmp_adev != adev) {
12ffa55d 3970 amdgpu_device_lock_adev(tmp_adev, false);
7c6e68c7
AG
3971 if (!amdgpu_sriov_vf(tmp_adev))
3972 amdgpu_amdkfd_pre_reset(tmp_adev);
3973 }
3974
12ffa55d
AG
3975 /*
3976 * Mark these ASICs to be reseted as untracked first
3977 * And add them back after reset completed
3978 */
3979 amdgpu_unregister_gpu_instance(tmp_adev);
3980
f1c1314b 3981 /* disable ras on ALL IPs */
7c6e68c7 3982 if (!in_ras_intr && amdgpu_device_ip_need_full_reset(tmp_adev))
f1c1314b 3983 amdgpu_ras_suspend(tmp_adev);
3984
1d721ed6
AG
3985 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3986 struct amdgpu_ring *ring = tmp_adev->rings[i];
3987
3988 if (!ring || !ring->sched.thread)
3989 continue;
3990
0b2d2c2e 3991 drm_sched_stop(&ring->sched, job ? &job->base : NULL);
7c6e68c7
AG
3992
3993 if (in_ras_intr)
3994 amdgpu_job_stop_all_jobs_on_sched(&ring->sched);
1d721ed6
AG
3995 }
3996 }
3997
3998
7c6e68c7
AG
3999 if (in_ras_intr)
4000 goto skip_sched_resume;
4001
1d721ed6
AG
4002 /*
4003 * Must check guilty signal here since after this point all old
4004 * HW fences are force signaled.
4005 *
4006 * job->base holds a reference to parent fence
4007 */
4008 if (job && job->base.s_fence->parent &&
4009 dma_fence_is_signaled(job->base.s_fence->parent))
4010 job_signaled = true;
4011
1d721ed6
AG
4012 if (job_signaled) {
4013 dev_info(adev->dev, "Guilty job already signaled, skipping HW reset");
4014 goto skip_hw_reset;
4015 }
4016
4017
4018 /* Guilty job will be freed after this*/
0b2d2c2e 4019 r = amdgpu_device_pre_asic_reset(adev, job, &need_full_reset);
1d721ed6
AG
4020 if (r) {
4021 /*TODO Should we stop ?*/
4022 DRM_ERROR("GPU pre asic reset failed with err, %d for drm dev, %s ",
4023 r, adev->ddev->unique);
4024 adev->asic_reset_res = r;
4025 }
4026
26bc5340
AG
4027retry: /* Rest of adevs pre asic reset from XGMI hive. */
4028 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
4029
4030 if (tmp_adev == adev)
4031 continue;
4032
26bc5340
AG
4033 r = amdgpu_device_pre_asic_reset(tmp_adev,
4034 NULL,
4035 &need_full_reset);
4036 /*TODO Should we stop ?*/
4037 if (r) {
4038 DRM_ERROR("GPU pre asic reset failed with err, %d for drm dev, %s ",
4039 r, tmp_adev->ddev->unique);
4040 tmp_adev->asic_reset_res = r;
4041 }
4042 }
4043
4044 /* Actual ASIC resets if needed.*/
4045 /* TODO Implement XGMI hive reset logic for SRIOV */
4046 if (amdgpu_sriov_vf(adev)) {
4047 r = amdgpu_device_reset_sriov(adev, job ? false : true);
4048 if (r)
4049 adev->asic_reset_res = r;
4050 } else {
4051 r = amdgpu_do_asic_reset(hive, device_list_handle, &need_full_reset);
4052 if (r && r == -EAGAIN)
4053 goto retry;
4054 }
4055
1d721ed6
AG
4056skip_hw_reset:
4057
26bc5340
AG
4058 /* Post ASIC reset for all devs .*/
4059 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
7c6e68c7 4060
1d721ed6
AG
4061 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
4062 struct amdgpu_ring *ring = tmp_adev->rings[i];
4063
4064 if (!ring || !ring->sched.thread)
4065 continue;
4066
4067 /* No point to resubmit jobs if we didn't HW reset*/
4068 if (!tmp_adev->asic_reset_res && !job_signaled)
4069 drm_sched_resubmit_jobs(&ring->sched);
4070
4071 drm_sched_start(&ring->sched, !tmp_adev->asic_reset_res);
4072 }
4073
4074 if (!amdgpu_device_has_dc_support(tmp_adev) && !job_signaled) {
4075 drm_helper_resume_force_mode(tmp_adev->ddev);
4076 }
4077
4078 tmp_adev->asic_reset_res = 0;
26bc5340
AG
4079
4080 if (r) {
4081 /* bad news, how to tell it to userspace ? */
12ffa55d 4082 dev_info(tmp_adev->dev, "GPU reset(%d) failed\n", atomic_read(&tmp_adev->gpu_reset_counter));
26bc5340
AG
4083 amdgpu_vf_error_put(tmp_adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r);
4084 } else {
12ffa55d 4085 dev_info(tmp_adev->dev, "GPU reset(%d) succeeded!\n", atomic_read(&tmp_adev->gpu_reset_counter));
26bc5340 4086 }
7c6e68c7 4087 }
26bc5340 4088
7c6e68c7
AG
4089skip_sched_resume:
4090 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
4091 /*unlock kfd: SRIOV would do it separately */
4092 if (!in_ras_intr && !amdgpu_sriov_vf(tmp_adev))
4093 amdgpu_amdkfd_post_reset(tmp_adev);
26bc5340
AG
4094 amdgpu_device_unlock_adev(tmp_adev);
4095 }
4096
1d721ed6 4097 if (hive)
22d6575b 4098 mutex_unlock(&hive->reset_lock);
26bc5340
AG
4099
4100 if (r)
4101 dev_info(adev->dev, "GPU reset end with ret = %d\n", r);
d38ceaf9
AD
4102 return r;
4103}
4104
e3ecdffa
AD
4105/**
4106 * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot
4107 *
4108 * @adev: amdgpu_device pointer
4109 *
4110 * Fetchs and stores in the driver the PCIE capabilities (gen speed
4111 * and lanes) of the slot the device is in. Handles APUs and
4112 * virtualized environments where PCIE config space may not be available.
4113 */
5494d864 4114static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev)
d0dd7f0c 4115{
5d9a6330 4116 struct pci_dev *pdev;
c5313457
HK
4117 enum pci_bus_speed speed_cap, platform_speed_cap;
4118 enum pcie_link_width platform_link_width;
d0dd7f0c 4119
cd474ba0
AD
4120 if (amdgpu_pcie_gen_cap)
4121 adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
d0dd7f0c 4122
cd474ba0
AD
4123 if (amdgpu_pcie_lane_cap)
4124 adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
d0dd7f0c 4125
cd474ba0
AD
4126 /* covers APUs as well */
4127 if (pci_is_root_bus(adev->pdev->bus)) {
4128 if (adev->pm.pcie_gen_mask == 0)
4129 adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
4130 if (adev->pm.pcie_mlw_mask == 0)
4131 adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
d0dd7f0c 4132 return;
cd474ba0 4133 }
d0dd7f0c 4134
c5313457
HK
4135 if (adev->pm.pcie_gen_mask && adev->pm.pcie_mlw_mask)
4136 return;
4137
dbaa922b
AD
4138 pcie_bandwidth_available(adev->pdev, NULL,
4139 &platform_speed_cap, &platform_link_width);
c5313457 4140
cd474ba0 4141 if (adev->pm.pcie_gen_mask == 0) {
5d9a6330
AD
4142 /* asic caps */
4143 pdev = adev->pdev;
4144 speed_cap = pcie_get_speed_cap(pdev);
4145 if (speed_cap == PCI_SPEED_UNKNOWN) {
4146 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
cd474ba0
AD
4147 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4148 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
cd474ba0 4149 } else {
5d9a6330
AD
4150 if (speed_cap == PCIE_SPEED_16_0GT)
4151 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4152 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4153 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
4154 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4);
4155 else if (speed_cap == PCIE_SPEED_8_0GT)
4156 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4157 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4158 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
4159 else if (speed_cap == PCIE_SPEED_5_0GT)
4160 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4161 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2);
4162 else
4163 adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1;
4164 }
4165 /* platform caps */
c5313457 4166 if (platform_speed_cap == PCI_SPEED_UNKNOWN) {
5d9a6330
AD
4167 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4168 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
4169 } else {
c5313457 4170 if (platform_speed_cap == PCIE_SPEED_16_0GT)
5d9a6330
AD
4171 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4172 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4173 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
4174 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4);
c5313457 4175 else if (platform_speed_cap == PCIE_SPEED_8_0GT)
5d9a6330
AD
4176 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4177 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4178 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3);
c5313457 4179 else if (platform_speed_cap == PCIE_SPEED_5_0GT)
5d9a6330
AD
4180 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4181 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
4182 else
4183 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
4184
cd474ba0
AD
4185 }
4186 }
4187 if (adev->pm.pcie_mlw_mask == 0) {
c5313457 4188 if (platform_link_width == PCIE_LNK_WIDTH_UNKNOWN) {
5d9a6330
AD
4189 adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK;
4190 } else {
c5313457 4191 switch (platform_link_width) {
5d9a6330 4192 case PCIE_LNK_X32:
cd474ba0
AD
4193 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
4194 CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
4195 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
4196 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4197 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4198 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4199 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4200 break;
5d9a6330 4201 case PCIE_LNK_X16:
cd474ba0
AD
4202 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
4203 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
4204 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4205 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4206 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4207 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4208 break;
5d9a6330 4209 case PCIE_LNK_X12:
cd474ba0
AD
4210 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
4211 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4212 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4213 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4214 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4215 break;
5d9a6330 4216 case PCIE_LNK_X8:
cd474ba0
AD
4217 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4218 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4219 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4220 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4221 break;
5d9a6330 4222 case PCIE_LNK_X4:
cd474ba0
AD
4223 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4224 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4225 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4226 break;
5d9a6330 4227 case PCIE_LNK_X2:
cd474ba0
AD
4228 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4229 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4230 break;
5d9a6330 4231 case PCIE_LNK_X1:
cd474ba0
AD
4232 adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;
4233 break;
4234 default:
4235 break;
4236 }
d0dd7f0c
AD
4237 }
4238 }
4239}
d38ceaf9 4240