drm/amdgpu: Use ARCTURUS in RAS EEPROM.
[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
3e2e2ab5
HZ
1880 if (adev->gmc.xgmi.num_physical_nodes > 1)
1881 amdgpu_xgmi_add_device(adev);
1884734a 1882 amdgpu_amdkfd_device_init(adev);
c6332b97 1883
72d3f592 1884init_failed:
d3c117e5 1885 if (amdgpu_sriov_vf(adev)) {
72d3f592
ED
1886 if (!r)
1887 amdgpu_virt_init_data_exchange(adev);
c6332b97 1888 amdgpu_virt_release_full_gpu(adev, true);
d3c117e5 1889 }
c6332b97 1890
72d3f592 1891 return r;
d38ceaf9
AD
1892}
1893
e3ecdffa
AD
1894/**
1895 * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer
1896 *
1897 * @adev: amdgpu_device pointer
1898 *
1899 * Writes a reset magic value to the gart pointer in VRAM. The driver calls
1900 * this function before a GPU reset. If the value is retained after a
1901 * GPU reset, VRAM has not been lost. Some GPU resets may destry VRAM contents.
1902 */
06ec9070 1903static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev)
0c49e0b8
CZ
1904{
1905 memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
1906}
1907
e3ecdffa
AD
1908/**
1909 * amdgpu_device_check_vram_lost - check if vram is valid
1910 *
1911 * @adev: amdgpu_device pointer
1912 *
1913 * Checks the reset magic value written to the gart pointer in VRAM.
1914 * The driver calls this after a GPU reset to see if the contents of
1915 * VRAM is lost or now.
1916 * returns true if vram is lost, false if not.
1917 */
06ec9070 1918static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev)
0c49e0b8
CZ
1919{
1920 return !!memcmp(adev->gart.ptr, adev->reset_magic,
1921 AMDGPU_RESET_MAGIC_NUM);
1922}
1923
e3ecdffa 1924/**
1112a46b 1925 * amdgpu_device_set_cg_state - set clockgating for amdgpu device
e3ecdffa
AD
1926 *
1927 * @adev: amdgpu_device pointer
1928 *
e3ecdffa 1929 * The list of all the hardware IPs that make up the asic is walked and the
1112a46b
RZ
1930 * set_clockgating_state callbacks are run.
1931 * Late initialization pass enabling clockgating for hardware IPs.
1932 * Fini or suspend, pass disabling clockgating for hardware IPs.
e3ecdffa
AD
1933 * Returns 0 on success, negative error code on failure.
1934 */
fdd34271 1935
1112a46b
RZ
1936static int amdgpu_device_set_cg_state(struct amdgpu_device *adev,
1937 enum amd_clockgating_state state)
d38ceaf9 1938{
1112a46b 1939 int i, j, r;
d38ceaf9 1940
4a2ba394
SL
1941 if (amdgpu_emu_mode == 1)
1942 return 0;
1943
1112a46b
RZ
1944 for (j = 0; j < adev->num_ip_blocks; j++) {
1945 i = state == AMD_CG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
a2d31dc3 1946 if (!adev->ip_blocks[i].status.late_initialized)
d38ceaf9 1947 continue;
4a446d55 1948 /* skip CG for VCE/UVD, it's handled specially */
a1255107 1949 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
57716327 1950 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
34319b32 1951 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
57716327 1952 adev->ip_blocks[i].version->funcs->set_clockgating_state) {
4a446d55 1953 /* enable clockgating to save power */
a1255107 1954 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1112a46b 1955 state);
4a446d55
AD
1956 if (r) {
1957 DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
a1255107 1958 adev->ip_blocks[i].version->funcs->name, r);
4a446d55
AD
1959 return r;
1960 }
b0b00ff1 1961 }
d38ceaf9 1962 }
06b18f61 1963
c9f96fd5
RZ
1964 return 0;
1965}
1966
1112a46b 1967static int amdgpu_device_set_pg_state(struct amdgpu_device *adev, enum amd_powergating_state state)
c9f96fd5 1968{
1112a46b 1969 int i, j, r;
06b18f61 1970
c9f96fd5
RZ
1971 if (amdgpu_emu_mode == 1)
1972 return 0;
1973
1112a46b
RZ
1974 for (j = 0; j < adev->num_ip_blocks; j++) {
1975 i = state == AMD_PG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
a2d31dc3 1976 if (!adev->ip_blocks[i].status.late_initialized)
c9f96fd5
RZ
1977 continue;
1978 /* skip CG for VCE/UVD, it's handled specially */
1979 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1980 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1981 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
1982 adev->ip_blocks[i].version->funcs->set_powergating_state) {
1983 /* enable powergating to save power */
1984 r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev,
1112a46b 1985 state);
c9f96fd5
RZ
1986 if (r) {
1987 DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n",
1988 adev->ip_blocks[i].version->funcs->name, r);
1989 return r;
1990 }
1991 }
1992 }
2dc80b00
S
1993 return 0;
1994}
1995
beff74bc
AD
1996static int amdgpu_device_enable_mgpu_fan_boost(void)
1997{
1998 struct amdgpu_gpu_instance *gpu_ins;
1999 struct amdgpu_device *adev;
2000 int i, ret = 0;
2001
2002 mutex_lock(&mgpu_info.mutex);
2003
2004 /*
2005 * MGPU fan boost feature should be enabled
2006 * only when there are two or more dGPUs in
2007 * the system
2008 */
2009 if (mgpu_info.num_dgpu < 2)
2010 goto out;
2011
2012 for (i = 0; i < mgpu_info.num_dgpu; i++) {
2013 gpu_ins = &(mgpu_info.gpu_ins[i]);
2014 adev = gpu_ins->adev;
2015 if (!(adev->flags & AMD_IS_APU) &&
2016 !gpu_ins->mgpu_fan_enabled &&
2017 adev->powerplay.pp_funcs &&
2018 adev->powerplay.pp_funcs->enable_mgpu_fan_boost) {
2019 ret = amdgpu_dpm_enable_mgpu_fan_boost(adev);
2020 if (ret)
2021 break;
2022
2023 gpu_ins->mgpu_fan_enabled = 1;
2024 }
2025 }
2026
2027out:
2028 mutex_unlock(&mgpu_info.mutex);
2029
2030 return ret;
2031}
2032
e3ecdffa
AD
2033/**
2034 * amdgpu_device_ip_late_init - run late init for hardware IPs
2035 *
2036 * @adev: amdgpu_device pointer
2037 *
2038 * Late initialization pass for hardware IPs. The list of all the hardware
2039 * IPs that make up the asic is walked and the late_init callbacks are run.
2040 * late_init covers any special initialization that an IP requires
2041 * after all of the have been initialized or something that needs to happen
2042 * late in the init process.
2043 * Returns 0 on success, negative error code on failure.
2044 */
06ec9070 2045static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
2dc80b00
S
2046{
2047 int i = 0, r;
2048
2049 for (i = 0; i < adev->num_ip_blocks; i++) {
73f847db 2050 if (!adev->ip_blocks[i].status.hw)
2dc80b00
S
2051 continue;
2052 if (adev->ip_blocks[i].version->funcs->late_init) {
2053 r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
2054 if (r) {
2055 DRM_ERROR("late_init of IP block <%s> failed %d\n",
2056 adev->ip_blocks[i].version->funcs->name, r);
2057 return r;
2058 }
2dc80b00 2059 }
73f847db 2060 adev->ip_blocks[i].status.late_initialized = true;
2dc80b00
S
2061 }
2062
1112a46b
RZ
2063 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE);
2064 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE);
916ac57f 2065
06ec9070 2066 amdgpu_device_fill_reset_magic(adev);
d38ceaf9 2067
beff74bc
AD
2068 r = amdgpu_device_enable_mgpu_fan_boost();
2069 if (r)
2070 DRM_ERROR("enable mgpu fan boost failed (%d).\n", r);
2071
2072 /* set to low pstate by default */
2073 amdgpu_xgmi_set_pstate(adev, 0);
2074
d38ceaf9
AD
2075 return 0;
2076}
2077
e3ecdffa
AD
2078/**
2079 * amdgpu_device_ip_fini - run fini for hardware IPs
2080 *
2081 * @adev: amdgpu_device pointer
2082 *
2083 * Main teardown pass for hardware IPs. The list of all the hardware
2084 * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks
2085 * are run. hw_fini tears down the hardware associated with each IP
2086 * and sw_fini tears down any software state associated with each IP.
2087 * Returns 0 on success, negative error code on failure.
2088 */
06ec9070 2089static int amdgpu_device_ip_fini(struct amdgpu_device *adev)
d38ceaf9
AD
2090{
2091 int i, r;
2092
c030f2e4 2093 amdgpu_ras_pre_fini(adev);
2094
a82400b5
AG
2095 if (adev->gmc.xgmi.num_physical_nodes > 1)
2096 amdgpu_xgmi_remove_device(adev);
2097
1884734a 2098 amdgpu_amdkfd_device_fini(adev);
05df1f01
RZ
2099
2100 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
fdd34271
RZ
2101 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
2102
3e96dbfd
AD
2103 /* need to disable SMC first */
2104 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 2105 if (!adev->ip_blocks[i].status.hw)
3e96dbfd 2106 continue;
fdd34271 2107 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
a1255107 2108 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
3e96dbfd
AD
2109 /* XXX handle errors */
2110 if (r) {
2111 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
a1255107 2112 adev->ip_blocks[i].version->funcs->name, r);
3e96dbfd 2113 }
a1255107 2114 adev->ip_blocks[i].status.hw = false;
3e96dbfd
AD
2115 break;
2116 }
2117 }
2118
d38ceaf9 2119 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
a1255107 2120 if (!adev->ip_blocks[i].status.hw)
d38ceaf9 2121 continue;
8201a67a 2122
a1255107 2123 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
d38ceaf9 2124 /* XXX handle errors */
2c1a2784 2125 if (r) {
a1255107
AD
2126 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
2127 adev->ip_blocks[i].version->funcs->name, r);
2c1a2784 2128 }
8201a67a 2129
a1255107 2130 adev->ip_blocks[i].status.hw = false;
d38ceaf9
AD
2131 }
2132
9950cda2 2133
d38ceaf9 2134 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
a1255107 2135 if (!adev->ip_blocks[i].status.sw)
d38ceaf9 2136 continue;
c12aba3a
ML
2137
2138 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
c8963ea4 2139 amdgpu_ucode_free_bo(adev);
1e256e27 2140 amdgpu_free_static_csa(&adev->virt.csa_obj);
c12aba3a
ML
2141 amdgpu_device_wb_fini(adev);
2142 amdgpu_device_vram_scratch_fini(adev);
533aed27 2143 amdgpu_ib_pool_fini(adev);
c12aba3a
ML
2144 }
2145
a1255107 2146 r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
d38ceaf9 2147 /* XXX handle errors */
2c1a2784 2148 if (r) {
a1255107
AD
2149 DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
2150 adev->ip_blocks[i].version->funcs->name, r);
2c1a2784 2151 }
a1255107
AD
2152 adev->ip_blocks[i].status.sw = false;
2153 adev->ip_blocks[i].status.valid = false;
d38ceaf9
AD
2154 }
2155
a6dcfd9c 2156 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
a1255107 2157 if (!adev->ip_blocks[i].status.late_initialized)
8a2eef1d 2158 continue;
a1255107
AD
2159 if (adev->ip_blocks[i].version->funcs->late_fini)
2160 adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
2161 adev->ip_blocks[i].status.late_initialized = false;
a6dcfd9c
ML
2162 }
2163
c030f2e4 2164 amdgpu_ras_fini(adev);
2165
030308fc 2166 if (amdgpu_sriov_vf(adev))
24136135
ML
2167 if (amdgpu_virt_release_full_gpu(adev, false))
2168 DRM_ERROR("failed to release exclusive mode on fini\n");
2493664f 2169
d38ceaf9
AD
2170 return 0;
2171}
2172
e3ecdffa 2173/**
beff74bc 2174 * amdgpu_device_delayed_init_work_handler - work handler for IB tests
e3ecdffa 2175 *
1112a46b 2176 * @work: work_struct.
e3ecdffa 2177 */
beff74bc 2178static void amdgpu_device_delayed_init_work_handler(struct work_struct *work)
2dc80b00
S
2179{
2180 struct amdgpu_device *adev =
beff74bc 2181 container_of(work, struct amdgpu_device, delayed_init_work.work);
916ac57f
RZ
2182 int r;
2183
2184 r = amdgpu_ib_ring_tests(adev);
2185 if (r)
2186 DRM_ERROR("ib ring test failed (%d).\n", r);
2dc80b00
S
2187}
2188
1e317b99
RZ
2189static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work)
2190{
2191 struct amdgpu_device *adev =
2192 container_of(work, struct amdgpu_device, gfx.gfx_off_delay_work.work);
2193
2194 mutex_lock(&adev->gfx.gfx_off_mutex);
2195 if (!adev->gfx.gfx_off_state && !adev->gfx.gfx_off_req_count) {
2196 if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, true))
2197 adev->gfx.gfx_off_state = true;
2198 }
2199 mutex_unlock(&adev->gfx.gfx_off_mutex);
2200}
2201
e3ecdffa 2202/**
e7854a03 2203 * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1)
e3ecdffa
AD
2204 *
2205 * @adev: amdgpu_device pointer
2206 *
2207 * Main suspend function for hardware IPs. The list of all the hardware
2208 * IPs that make up the asic is walked, clockgating is disabled and the
2209 * suspend callbacks are run. suspend puts the hardware and software state
2210 * in each IP into a state suitable for suspend.
2211 * Returns 0 on success, negative error code on failure.
2212 */
e7854a03
AD
2213static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev)
2214{
2215 int i, r;
2216
05df1f01 2217 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
fdd34271 2218 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
05df1f01 2219
e7854a03
AD
2220 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2221 if (!adev->ip_blocks[i].status.valid)
2222 continue;
2223 /* displays are handled separately */
2224 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) {
e7854a03
AD
2225 /* XXX handle errors */
2226 r = adev->ip_blocks[i].version->funcs->suspend(adev);
2227 /* XXX handle errors */
2228 if (r) {
2229 DRM_ERROR("suspend of IP block <%s> failed %d\n",
2230 adev->ip_blocks[i].version->funcs->name, r);
482f0e53 2231 return r;
e7854a03 2232 }
482f0e53 2233 adev->ip_blocks[i].status.hw = false;
e7854a03
AD
2234 }
2235 }
2236
e7854a03
AD
2237 return 0;
2238}
2239
2240/**
2241 * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2)
2242 *
2243 * @adev: amdgpu_device pointer
2244 *
2245 * Main suspend function for hardware IPs. The list of all the hardware
2246 * IPs that make up the asic is walked, clockgating is disabled and the
2247 * suspend callbacks are run. suspend puts the hardware and software state
2248 * in each IP into a state suitable for suspend.
2249 * Returns 0 on success, negative error code on failure.
2250 */
2251static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev)
d38ceaf9
AD
2252{
2253 int i, r;
2254
2255 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
a1255107 2256 if (!adev->ip_blocks[i].status.valid)
d38ceaf9 2257 continue;
e7854a03
AD
2258 /* displays are handled in phase1 */
2259 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE)
2260 continue;
d38ceaf9 2261 /* XXX handle errors */
a1255107 2262 r = adev->ip_blocks[i].version->funcs->suspend(adev);
d38ceaf9 2263 /* XXX handle errors */
2c1a2784 2264 if (r) {
a1255107
AD
2265 DRM_ERROR("suspend of IP block <%s> failed %d\n",
2266 adev->ip_blocks[i].version->funcs->name, r);
2c1a2784 2267 }
876923fb 2268 adev->ip_blocks[i].status.hw = false;
a3a09142
AD
2269 /* handle putting the SMC in the appropriate state */
2270 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
2271 if (is_support_sw_smu(adev)) {
0e0b89c0 2272 r = smu_set_mp1_state(&adev->smu, adev->mp1_state);
a3a09142 2273 } else if (adev->powerplay.pp_funcs &&
482f0e53 2274 adev->powerplay.pp_funcs->set_mp1_state) {
a3a09142
AD
2275 r = adev->powerplay.pp_funcs->set_mp1_state(
2276 adev->powerplay.pp_handle,
2277 adev->mp1_state);
0e0b89c0
EQ
2278 }
2279 if (r) {
2280 DRM_ERROR("SMC failed to set mp1 state %d, %d\n",
2281 adev->mp1_state, r);
2282 return r;
a3a09142
AD
2283 }
2284 }
b5507c7e
AG
2285
2286 adev->ip_blocks[i].status.hw = false;
d38ceaf9
AD
2287 }
2288
2289 return 0;
2290}
2291
e7854a03
AD
2292/**
2293 * amdgpu_device_ip_suspend - run suspend for hardware IPs
2294 *
2295 * @adev: amdgpu_device pointer
2296 *
2297 * Main suspend function for hardware IPs. The list of all the hardware
2298 * IPs that make up the asic is walked, clockgating is disabled and the
2299 * suspend callbacks are run. suspend puts the hardware and software state
2300 * in each IP into a state suitable for suspend.
2301 * Returns 0 on success, negative error code on failure.
2302 */
2303int amdgpu_device_ip_suspend(struct amdgpu_device *adev)
2304{
2305 int r;
2306
e7819644
YT
2307 if (amdgpu_sriov_vf(adev))
2308 amdgpu_virt_request_full_gpu(adev, false);
2309
e7854a03
AD
2310 r = amdgpu_device_ip_suspend_phase1(adev);
2311 if (r)
2312 return r;
2313 r = amdgpu_device_ip_suspend_phase2(adev);
2314
e7819644
YT
2315 if (amdgpu_sriov_vf(adev))
2316 amdgpu_virt_release_full_gpu(adev, false);
2317
e7854a03
AD
2318 return r;
2319}
2320
06ec9070 2321static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
a90ad3c2
ML
2322{
2323 int i, r;
2324
2cb681b6
ML
2325 static enum amd_ip_block_type ip_order[] = {
2326 AMD_IP_BLOCK_TYPE_GMC,
2327 AMD_IP_BLOCK_TYPE_COMMON,
39186aef 2328 AMD_IP_BLOCK_TYPE_PSP,
2cb681b6
ML
2329 AMD_IP_BLOCK_TYPE_IH,
2330 };
a90ad3c2 2331
2cb681b6
ML
2332 for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
2333 int j;
2334 struct amdgpu_ip_block *block;
a90ad3c2 2335
2cb681b6
ML
2336 for (j = 0; j < adev->num_ip_blocks; j++) {
2337 block = &adev->ip_blocks[j];
2338
482f0e53 2339 block->status.hw = false;
2cb681b6
ML
2340 if (block->version->type != ip_order[i] ||
2341 !block->status.valid)
2342 continue;
2343
2344 r = block->version->funcs->hw_init(adev);
0aaeefcc 2345 DRM_INFO("RE-INIT-early: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
c41d1cf6
ML
2346 if (r)
2347 return r;
482f0e53 2348 block->status.hw = true;
a90ad3c2
ML
2349 }
2350 }
2351
2352 return 0;
2353}
2354
06ec9070 2355static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
a90ad3c2
ML
2356{
2357 int i, r;
2358
2cb681b6
ML
2359 static enum amd_ip_block_type ip_order[] = {
2360 AMD_IP_BLOCK_TYPE_SMC,
2361 AMD_IP_BLOCK_TYPE_DCE,
2362 AMD_IP_BLOCK_TYPE_GFX,
2363 AMD_IP_BLOCK_TYPE_SDMA,
257deb8c
FM
2364 AMD_IP_BLOCK_TYPE_UVD,
2365 AMD_IP_BLOCK_TYPE_VCE
2cb681b6 2366 };
a90ad3c2 2367
2cb681b6
ML
2368 for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
2369 int j;
2370 struct amdgpu_ip_block *block;
a90ad3c2 2371
2cb681b6
ML
2372 for (j = 0; j < adev->num_ip_blocks; j++) {
2373 block = &adev->ip_blocks[j];
2374
2375 if (block->version->type != ip_order[i] ||
482f0e53
ML
2376 !block->status.valid ||
2377 block->status.hw)
2cb681b6
ML
2378 continue;
2379
2380 r = block->version->funcs->hw_init(adev);
0aaeefcc 2381 DRM_INFO("RE-INIT-late: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
c41d1cf6
ML
2382 if (r)
2383 return r;
482f0e53 2384 block->status.hw = true;
a90ad3c2
ML
2385 }
2386 }
2387
2388 return 0;
2389}
2390
e3ecdffa
AD
2391/**
2392 * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs
2393 *
2394 * @adev: amdgpu_device pointer
2395 *
2396 * First resume function for hardware IPs. The list of all the hardware
2397 * IPs that make up the asic is walked and the resume callbacks are run for
2398 * COMMON, GMC, and IH. resume puts the hardware into a functional state
2399 * after a suspend and updates the software state as necessary. This
2400 * function is also used for restoring the GPU after a GPU reset.
2401 * Returns 0 on success, negative error code on failure.
2402 */
06ec9070 2403static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)
d38ceaf9
AD
2404{
2405 int i, r;
2406
a90ad3c2 2407 for (i = 0; i < adev->num_ip_blocks; i++) {
482f0e53 2408 if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
a90ad3c2 2409 continue;
a90ad3c2 2410 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
e3ecdffa
AD
2411 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2412 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
482f0e53 2413
fcf0649f
CZ
2414 r = adev->ip_blocks[i].version->funcs->resume(adev);
2415 if (r) {
2416 DRM_ERROR("resume of IP block <%s> failed %d\n",
2417 adev->ip_blocks[i].version->funcs->name, r);
2418 return r;
2419 }
482f0e53 2420 adev->ip_blocks[i].status.hw = true;
a90ad3c2
ML
2421 }
2422 }
2423
2424 return 0;
2425}
2426
e3ecdffa
AD
2427/**
2428 * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs
2429 *
2430 * @adev: amdgpu_device pointer
2431 *
2432 * First resume function for hardware IPs. The list of all the hardware
2433 * IPs that make up the asic is walked and the resume callbacks are run for
2434 * all blocks except COMMON, GMC, and IH. resume puts the hardware into a
2435 * functional state after a suspend and updates the software state as
2436 * necessary. This function is also used for restoring the GPU after a GPU
2437 * reset.
2438 * Returns 0 on success, negative error code on failure.
2439 */
06ec9070 2440static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)
d38ceaf9
AD
2441{
2442 int i, r;
2443
2444 for (i = 0; i < adev->num_ip_blocks; i++) {
482f0e53 2445 if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
d38ceaf9 2446 continue;
fcf0649f 2447 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
e3ecdffa 2448 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
7a3e0bb2
RZ
2449 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH ||
2450 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)
fcf0649f 2451 continue;
a1255107 2452 r = adev->ip_blocks[i].version->funcs->resume(adev);
2c1a2784 2453 if (r) {
a1255107
AD
2454 DRM_ERROR("resume of IP block <%s> failed %d\n",
2455 adev->ip_blocks[i].version->funcs->name, r);
d38ceaf9 2456 return r;
2c1a2784 2457 }
482f0e53 2458 adev->ip_blocks[i].status.hw = true;
d38ceaf9
AD
2459 }
2460
2461 return 0;
2462}
2463
e3ecdffa
AD
2464/**
2465 * amdgpu_device_ip_resume - run resume for hardware IPs
2466 *
2467 * @adev: amdgpu_device pointer
2468 *
2469 * Main resume function for hardware IPs. The hardware IPs
2470 * are split into two resume functions because they are
2471 * are also used in in recovering from a GPU reset and some additional
2472 * steps need to be take between them. In this case (S3/S4) they are
2473 * run sequentially.
2474 * Returns 0 on success, negative error code on failure.
2475 */
06ec9070 2476static int amdgpu_device_ip_resume(struct amdgpu_device *adev)
fcf0649f
CZ
2477{
2478 int r;
2479
06ec9070 2480 r = amdgpu_device_ip_resume_phase1(adev);
fcf0649f
CZ
2481 if (r)
2482 return r;
7a3e0bb2
RZ
2483
2484 r = amdgpu_device_fw_loading(adev);
2485 if (r)
2486 return r;
2487
06ec9070 2488 r = amdgpu_device_ip_resume_phase2(adev);
fcf0649f
CZ
2489
2490 return r;
2491}
2492
e3ecdffa
AD
2493/**
2494 * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV
2495 *
2496 * @adev: amdgpu_device pointer
2497 *
2498 * Query the VBIOS data tables to determine if the board supports SR-IOV.
2499 */
4e99a44e 2500static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
048765ad 2501{
6867e1b5
ML
2502 if (amdgpu_sriov_vf(adev)) {
2503 if (adev->is_atom_fw) {
2504 if (amdgpu_atomfirmware_gpu_supports_virtualization(adev))
2505 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2506 } else {
2507 if (amdgpu_atombios_has_gpu_virtualization_table(adev))
2508 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2509 }
2510
2511 if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
2512 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
a5bde2f9 2513 }
048765ad
AR
2514}
2515
e3ecdffa
AD
2516/**
2517 * amdgpu_device_asic_has_dc_support - determine if DC supports the asic
2518 *
2519 * @asic_type: AMD asic type
2520 *
2521 * Check if there is DC (new modesetting infrastructre) support for an asic.
2522 * returns true if DC has support, false if not.
2523 */
4562236b
HW
2524bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
2525{
2526 switch (asic_type) {
2527#if defined(CONFIG_DRM_AMD_DC)
2528 case CHIP_BONAIRE:
0d6fbccb 2529 case CHIP_KAVERI:
367e6687
AD
2530 case CHIP_KABINI:
2531 case CHIP_MULLINS:
d9fda248
HW
2532 /*
2533 * We have systems in the wild with these ASICs that require
2534 * LVDS and VGA support which is not supported with DC.
2535 *
2536 * Fallback to the non-DC driver here by default so as not to
2537 * cause regressions.
2538 */
2539 return amdgpu_dc > 0;
2540 case CHIP_HAWAII:
4562236b
HW
2541 case CHIP_CARRIZO:
2542 case CHIP_STONEY:
4562236b 2543 case CHIP_POLARIS10:
675fd32b 2544 case CHIP_POLARIS11:
2c8ad2d5 2545 case CHIP_POLARIS12:
675fd32b 2546 case CHIP_VEGAM:
4562236b
HW
2547 case CHIP_TONGA:
2548 case CHIP_FIJI:
42f8ffa1 2549 case CHIP_VEGA10:
dca7b401 2550 case CHIP_VEGA12:
c6034aa2 2551 case CHIP_VEGA20:
dc37a9a0 2552#if defined(CONFIG_DRM_AMD_DC_DCN1_0)
fd187853 2553 case CHIP_RAVEN:
b4f199c7
HW
2554#endif
2555#if defined(CONFIG_DRM_AMD_DC_DCN2_0)
2556 case CHIP_NAVI10:
8fceceb6 2557 case CHIP_NAVI14:
078655d9 2558 case CHIP_NAVI12:
e1c14c43
RL
2559#endif
2560#if defined(CONFIG_DRM_AMD_DC_DCN2_1)
2561 case CHIP_RENOIR:
42f8ffa1 2562#endif
fd187853 2563 return amdgpu_dc != 0;
4562236b
HW
2564#endif
2565 default:
2566 return false;
2567 }
2568}
2569
2570/**
2571 * amdgpu_device_has_dc_support - check if dc is supported
2572 *
2573 * @adev: amdgpu_device_pointer
2574 *
2575 * Returns true for supported, false for not supported
2576 */
2577bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
2578{
2555039d
XY
2579 if (amdgpu_sriov_vf(adev))
2580 return false;
2581
4562236b
HW
2582 return amdgpu_device_asic_has_dc_support(adev->asic_type);
2583}
2584
d4535e2c
AG
2585
2586static void amdgpu_device_xgmi_reset_func(struct work_struct *__work)
2587{
2588 struct amdgpu_device *adev =
2589 container_of(__work, struct amdgpu_device, xgmi_reset_work);
2590
2591 adev->asic_reset_res = amdgpu_asic_reset(adev);
2592 if (adev->asic_reset_res)
fed184e9 2593 DRM_WARN("ASIC reset failed with error, %d for drm dev, %s",
d4535e2c
AG
2594 adev->asic_reset_res, adev->ddev->unique);
2595}
2596
71f98027
AD
2597static int amdgpu_device_get_job_timeout_settings(struct amdgpu_device *adev)
2598{
2599 char *input = amdgpu_lockup_timeout;
2600 char *timeout_setting = NULL;
2601 int index = 0;
2602 long timeout;
2603 int ret = 0;
2604
2605 /*
2606 * By default timeout for non compute jobs is 10000.
2607 * And there is no timeout enforced on compute jobs.
2608 * In SR-IOV or passthrough mode, timeout for compute
2609 * jobs are 10000 by default.
2610 */
2611 adev->gfx_timeout = msecs_to_jiffies(10000);
2612 adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout;
2613 if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev))
2614 adev->compute_timeout = adev->gfx_timeout;
2615 else
2616 adev->compute_timeout = MAX_SCHEDULE_TIMEOUT;
2617
2618 if (strnlen(input, AMDGPU_MAX_TIMEOUT_PARAM_LENTH)) {
2619 while ((timeout_setting = strsep(&input, ",")) &&
2620 strnlen(timeout_setting, AMDGPU_MAX_TIMEOUT_PARAM_LENTH)) {
2621 ret = kstrtol(timeout_setting, 0, &timeout);
2622 if (ret)
2623 return ret;
2624
2625 if (timeout == 0) {
2626 index++;
2627 continue;
2628 } else if (timeout < 0) {
2629 timeout = MAX_SCHEDULE_TIMEOUT;
2630 } else {
2631 timeout = msecs_to_jiffies(timeout);
2632 }
2633
2634 switch (index++) {
2635 case 0:
2636 adev->gfx_timeout = timeout;
2637 break;
2638 case 1:
2639 adev->compute_timeout = timeout;
2640 break;
2641 case 2:
2642 adev->sdma_timeout = timeout;
2643 break;
2644 case 3:
2645 adev->video_timeout = timeout;
2646 break;
2647 default:
2648 break;
2649 }
2650 }
2651 /*
2652 * There is only one value specified and
2653 * it should apply to all non-compute jobs.
2654 */
bcccee89 2655 if (index == 1) {
71f98027 2656 adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout;
bcccee89
ED
2657 if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev))
2658 adev->compute_timeout = adev->gfx_timeout;
2659 }
71f98027
AD
2660 }
2661
2662 return ret;
2663}
d4535e2c 2664
d38ceaf9
AD
2665/**
2666 * amdgpu_device_init - initialize the driver
2667 *
2668 * @adev: amdgpu_device pointer
87e3f136 2669 * @ddev: drm dev pointer
d38ceaf9
AD
2670 * @pdev: pci dev pointer
2671 * @flags: driver flags
2672 *
2673 * Initializes the driver info and hw (all asics).
2674 * Returns 0 for success or an error on failure.
2675 * Called at driver startup.
2676 */
2677int amdgpu_device_init(struct amdgpu_device *adev,
2678 struct drm_device *ddev,
2679 struct pci_dev *pdev,
2680 uint32_t flags)
2681{
2682 int r, i;
2683 bool runtime = false;
95844d20 2684 u32 max_MBps;
d38ceaf9
AD
2685
2686 adev->shutdown = false;
2687 adev->dev = &pdev->dev;
2688 adev->ddev = ddev;
2689 adev->pdev = pdev;
2690 adev->flags = flags;
4e66d7d2
YZ
2691
2692 if (amdgpu_force_asic_type >= 0 && amdgpu_force_asic_type < CHIP_LAST)
2693 adev->asic_type = amdgpu_force_asic_type;
2694 else
2695 adev->asic_type = flags & AMD_ASIC_MASK;
2696
d38ceaf9 2697 adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
593aa2d2
SL
2698 if (amdgpu_emu_mode == 1)
2699 adev->usec_timeout *= 2;
770d13b1 2700 adev->gmc.gart_size = 512 * 1024 * 1024;
d38ceaf9
AD
2701 adev->accel_working = false;
2702 adev->num_rings = 0;
2703 adev->mman.buffer_funcs = NULL;
2704 adev->mman.buffer_funcs_ring = NULL;
2705 adev->vm_manager.vm_pte_funcs = NULL;
3798e9a6 2706 adev->vm_manager.vm_pte_num_rqs = 0;
132f34e4 2707 adev->gmc.gmc_funcs = NULL;
f54d1867 2708 adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
b8866c26 2709 bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
d38ceaf9
AD
2710
2711 adev->smc_rreg = &amdgpu_invalid_rreg;
2712 adev->smc_wreg = &amdgpu_invalid_wreg;
2713 adev->pcie_rreg = &amdgpu_invalid_rreg;
2714 adev->pcie_wreg = &amdgpu_invalid_wreg;
36b9a952
HR
2715 adev->pciep_rreg = &amdgpu_invalid_rreg;
2716 adev->pciep_wreg = &amdgpu_invalid_wreg;
4fa1c6a6
TZ
2717 adev->pcie_rreg64 = &amdgpu_invalid_rreg64;
2718 adev->pcie_wreg64 = &amdgpu_invalid_wreg64;
d38ceaf9
AD
2719 adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
2720 adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
2721 adev->didt_rreg = &amdgpu_invalid_rreg;
2722 adev->didt_wreg = &amdgpu_invalid_wreg;
ccdbb20a
RZ
2723 adev->gc_cac_rreg = &amdgpu_invalid_rreg;
2724 adev->gc_cac_wreg = &amdgpu_invalid_wreg;
d38ceaf9
AD
2725 adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
2726 adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
2727
3e39ab90
AD
2728 DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
2729 amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
2730 pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
d38ceaf9
AD
2731
2732 /* mutex initialization are all done here so we
2733 * can recall function without having locking issues */
d38ceaf9 2734 atomic_set(&adev->irq.ih.lock, 0);
0e5ca0d1 2735 mutex_init(&adev->firmware.mutex);
d38ceaf9
AD
2736 mutex_init(&adev->pm.mutex);
2737 mutex_init(&adev->gfx.gpu_clock_mutex);
2738 mutex_init(&adev->srbm_mutex);
b8866c26 2739 mutex_init(&adev->gfx.pipe_reserve_mutex);
d23ee13f 2740 mutex_init(&adev->gfx.gfx_off_mutex);
d38ceaf9 2741 mutex_init(&adev->grbm_idx_mutex);
d38ceaf9 2742 mutex_init(&adev->mn_lock);
e23b74aa 2743 mutex_init(&adev->virt.vf_errors.lock);
d38ceaf9 2744 hash_init(adev->mn_hash);
13a752e3 2745 mutex_init(&adev->lock_reset);
bb5a2bdf 2746 mutex_init(&adev->virt.dpm_mutex);
32eaeae0 2747 mutex_init(&adev->psp.mutex);
d38ceaf9 2748
912dfc84
EQ
2749 r = amdgpu_device_check_arguments(adev);
2750 if (r)
2751 return r;
d38ceaf9 2752
d38ceaf9
AD
2753 spin_lock_init(&adev->mmio_idx_lock);
2754 spin_lock_init(&adev->smc_idx_lock);
2755 spin_lock_init(&adev->pcie_idx_lock);
2756 spin_lock_init(&adev->uvd_ctx_idx_lock);
2757 spin_lock_init(&adev->didt_idx_lock);
ccdbb20a 2758 spin_lock_init(&adev->gc_cac_idx_lock);
16abb5d2 2759 spin_lock_init(&adev->se_cac_idx_lock);
d38ceaf9 2760 spin_lock_init(&adev->audio_endpt_idx_lock);
95844d20 2761 spin_lock_init(&adev->mm_stats.lock);
d38ceaf9 2762
0c4e7fa5
CZ
2763 INIT_LIST_HEAD(&adev->shadow_list);
2764 mutex_init(&adev->shadow_list_lock);
2765
795f2813
AR
2766 INIT_LIST_HEAD(&adev->ring_lru_list);
2767 spin_lock_init(&adev->ring_lru_list_lock);
2768
beff74bc
AD
2769 INIT_DELAYED_WORK(&adev->delayed_init_work,
2770 amdgpu_device_delayed_init_work_handler);
1e317b99
RZ
2771 INIT_DELAYED_WORK(&adev->gfx.gfx_off_delay_work,
2772 amdgpu_device_delay_enable_gfx_off);
2dc80b00 2773
d4535e2c
AG
2774 INIT_WORK(&adev->xgmi_reset_work, amdgpu_device_xgmi_reset_func);
2775
d23ee13f 2776 adev->gfx.gfx_off_req_count = 1;
b1ddf548
RZ
2777 adev->pm.ac_power = power_supply_is_system_supplied() > 0 ? true : false;
2778
0fa49558
AX
2779 /* Registers mapping */
2780 /* TODO: block userspace mapping of io register */
da69c161
KW
2781 if (adev->asic_type >= CHIP_BONAIRE) {
2782 adev->rmmio_base = pci_resource_start(adev->pdev, 5);
2783 adev->rmmio_size = pci_resource_len(adev->pdev, 5);
2784 } else {
2785 adev->rmmio_base = pci_resource_start(adev->pdev, 2);
2786 adev->rmmio_size = pci_resource_len(adev->pdev, 2);
2787 }
d38ceaf9 2788
d38ceaf9
AD
2789 adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
2790 if (adev->rmmio == NULL) {
2791 return -ENOMEM;
2792 }
2793 DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
2794 DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
2795
d38ceaf9
AD
2796 /* io port mapping */
2797 for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
2798 if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) {
2799 adev->rio_mem_size = pci_resource_len(adev->pdev, i);
2800 adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size);
2801 break;
2802 }
2803 }
2804 if (adev->rio_mem == NULL)
b64a18c5 2805 DRM_INFO("PCI I/O BAR is not found.\n");
d38ceaf9 2806
b2109d8e
JX
2807 /* enable PCIE atomic ops */
2808 r = pci_enable_atomic_ops_to_root(adev->pdev,
2809 PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
2810 PCI_EXP_DEVCAP2_ATOMIC_COMP64);
2811 if (r) {
2812 adev->have_atomics_support = false;
2813 DRM_INFO("PCIE atomic ops is not supported\n");
2814 } else {
2815 adev->have_atomics_support = true;
2816 }
2817
5494d864
AD
2818 amdgpu_device_get_pcie_info(adev);
2819
b239c017
JX
2820 if (amdgpu_mcbp)
2821 DRM_INFO("MCBP is enabled\n");
2822
5f84cc63
JX
2823 if (amdgpu_mes && adev->asic_type >= CHIP_NAVI10)
2824 adev->enable_mes = true;
2825
f54eeab4 2826 if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10) {
a190d1c7
XY
2827 r = amdgpu_discovery_init(adev);
2828 if (r) {
2829 dev_err(adev->dev, "amdgpu_discovery_init failed\n");
2830 return r;
2831 }
2832 }
2833
d38ceaf9 2834 /* early init functions */
06ec9070 2835 r = amdgpu_device_ip_early_init(adev);
d38ceaf9
AD
2836 if (r)
2837 return r;
2838
df99ac0f
JZ
2839 r = amdgpu_device_get_job_timeout_settings(adev);
2840 if (r) {
2841 dev_err(adev->dev, "invalid lockup_timeout parameter syntax\n");
2842 return r;
2843 }
2844
6585661d
OZ
2845 /* doorbell bar mapping and doorbell index init*/
2846 amdgpu_device_doorbell_init(adev);
2847
d38ceaf9
AD
2848 /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
2849 /* this will fail for cards that aren't VGA class devices, just
2850 * ignore it */
06ec9070 2851 vga_client_register(adev->pdev, adev, NULL, amdgpu_device_vga_set_decode);
d38ceaf9 2852
e9bef455 2853 if (amdgpu_device_is_px(ddev))
d38ceaf9 2854 runtime = true;
84c8b22e
LW
2855 if (!pci_is_thunderbolt_attached(adev->pdev))
2856 vga_switcheroo_register_client(adev->pdev,
2857 &amdgpu_switcheroo_ops, runtime);
d38ceaf9
AD
2858 if (runtime)
2859 vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
2860
9475a943
SL
2861 if (amdgpu_emu_mode == 1) {
2862 /* post the asic on emulation mode */
2863 emu_soc_asic_init(adev);
bfca0289 2864 goto fence_driver_init;
9475a943 2865 }
bfca0289 2866
4e99a44e
ML
2867 /* detect if we are with an SRIOV vbios */
2868 amdgpu_device_detect_sriov_bios(adev);
048765ad 2869
95e8e59e
AD
2870 /* check if we need to reset the asic
2871 * E.g., driver was not cleanly unloaded previously, etc.
2872 */
f14899fd 2873 if (!amdgpu_sriov_vf(adev) && amdgpu_asic_need_reset_on_init(adev)) {
95e8e59e
AD
2874 r = amdgpu_asic_reset(adev);
2875 if (r) {
2876 dev_err(adev->dev, "asic reset on init failed\n");
2877 goto failed;
2878 }
2879 }
2880
d38ceaf9 2881 /* Post card if necessary */
39c640c0 2882 if (amdgpu_device_need_post(adev)) {
d38ceaf9 2883 if (!adev->bios) {
bec86378 2884 dev_err(adev->dev, "no vBIOS found\n");
83ba126a
AD
2885 r = -EINVAL;
2886 goto failed;
d38ceaf9 2887 }
bec86378 2888 DRM_INFO("GPU posting now...\n");
4e99a44e
ML
2889 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2890 if (r) {
2891 dev_err(adev->dev, "gpu post error!\n");
2892 goto failed;
2893 }
d38ceaf9
AD
2894 }
2895
88b64e95
AD
2896 if (adev->is_atom_fw) {
2897 /* Initialize clocks */
2898 r = amdgpu_atomfirmware_get_clock_info(adev);
2899 if (r) {
2900 dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
e23b74aa 2901 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
88b64e95
AD
2902 goto failed;
2903 }
2904 } else {
a5bde2f9
AD
2905 /* Initialize clocks */
2906 r = amdgpu_atombios_get_clock_info(adev);
2907 if (r) {
2908 dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
e23b74aa 2909 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
89041940 2910 goto failed;
a5bde2f9
AD
2911 }
2912 /* init i2c buses */
4562236b
HW
2913 if (!amdgpu_device_has_dc_support(adev))
2914 amdgpu_atombios_i2c_init(adev);
2c1a2784 2915 }
d38ceaf9 2916
bfca0289 2917fence_driver_init:
d38ceaf9
AD
2918 /* Fence driver */
2919 r = amdgpu_fence_driver_init(adev);
2c1a2784
AD
2920 if (r) {
2921 dev_err(adev->dev, "amdgpu_fence_driver_init failed\n");
e23b74aa 2922 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
83ba126a 2923 goto failed;
2c1a2784 2924 }
d38ceaf9
AD
2925
2926 /* init the mode config */
2927 drm_mode_config_init(adev->ddev);
2928
06ec9070 2929 r = amdgpu_device_ip_init(adev);
d38ceaf9 2930 if (r) {
8840a387 2931 /* failed in exclusive mode due to timeout */
2932 if (amdgpu_sriov_vf(adev) &&
2933 !amdgpu_sriov_runtime(adev) &&
2934 amdgpu_virt_mmio_blocked(adev) &&
2935 !amdgpu_virt_wait_reset(adev)) {
2936 dev_err(adev->dev, "VF exclusive mode timeout\n");
1daee8b4
PD
2937 /* Don't send request since VF is inactive. */
2938 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
2939 adev->virt.ops = NULL;
8840a387 2940 r = -EAGAIN;
2941 goto failed;
2942 }
06ec9070 2943 dev_err(adev->dev, "amdgpu_device_ip_init failed\n");
e23b74aa 2944 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
72d3f592
ED
2945 if (amdgpu_virt_request_full_gpu(adev, false))
2946 amdgpu_virt_release_full_gpu(adev, false);
83ba126a 2947 goto failed;
d38ceaf9
AD
2948 }
2949
2950 adev->accel_working = true;
2951
e59c0205
AX
2952 amdgpu_vm_check_compute_bug(adev);
2953
95844d20
MO
2954 /* Initialize the buffer migration limit. */
2955 if (amdgpu_moverate >= 0)
2956 max_MBps = amdgpu_moverate;
2957 else
2958 max_MBps = 8; /* Allow 8 MB/s. */
2959 /* Get a log2 for easy divisions. */
2960 adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
2961
9bc92b9c
ML
2962 amdgpu_fbdev_init(adev);
2963
e9bc1bf7
YT
2964 if (amdgpu_sriov_vf(adev) && amdgim_is_hwperf(adev))
2965 amdgpu_pm_virt_sysfs_init(adev);
2966
d2f52ac8
RZ
2967 r = amdgpu_pm_sysfs_init(adev);
2968 if (r)
2969 DRM_ERROR("registering pm debugfs failed (%d).\n", r);
2970
5bb23532
OM
2971 r = amdgpu_ucode_sysfs_init(adev);
2972 if (r)
2973 DRM_ERROR("Creating firmware sysfs failed (%d).\n", r);
2974
75758255 2975 r = amdgpu_debugfs_gem_init(adev);
3f14e623 2976 if (r)
d38ceaf9 2977 DRM_ERROR("registering gem debugfs failed (%d).\n", r);
d38ceaf9
AD
2978
2979 r = amdgpu_debugfs_regs_init(adev);
3f14e623 2980 if (r)
d38ceaf9 2981 DRM_ERROR("registering register debugfs failed (%d).\n", r);
d38ceaf9 2982
50ab2533 2983 r = amdgpu_debugfs_firmware_init(adev);
3f14e623 2984 if (r)
50ab2533 2985 DRM_ERROR("registering firmware debugfs failed (%d).\n", r);
50ab2533 2986
763efb6c 2987 r = amdgpu_debugfs_init(adev);
db95e218 2988 if (r)
763efb6c 2989 DRM_ERROR("Creating debugfs files failed (%d).\n", r);
db95e218 2990
d38ceaf9
AD
2991 if ((amdgpu_testing & 1)) {
2992 if (adev->accel_working)
2993 amdgpu_test_moves(adev);
2994 else
2995 DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n");
2996 }
d38ceaf9
AD
2997 if (amdgpu_benchmarking) {
2998 if (adev->accel_working)
2999 amdgpu_benchmark(adev, amdgpu_benchmarking);
3000 else
3001 DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n");
3002 }
3003
3004 /* enable clockgating, etc. after ib tests, etc. since some blocks require
3005 * explicit gating rather than handling it automatically.
3006 */
06ec9070 3007 r = amdgpu_device_ip_late_init(adev);
2c1a2784 3008 if (r) {
06ec9070 3009 dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n");
e23b74aa 3010 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
83ba126a 3011 goto failed;
2c1a2784 3012 }
d38ceaf9 3013
108c6a63 3014 /* must succeed. */
511fdbc3 3015 amdgpu_ras_resume(adev);
108c6a63 3016
beff74bc
AD
3017 queue_delayed_work(system_wq, &adev->delayed_init_work,
3018 msecs_to_jiffies(AMDGPU_RESUME_MS));
3019
dcea6e65
KR
3020 r = device_create_file(adev->dev, &dev_attr_pcie_replay_count);
3021 if (r) {
3022 dev_err(adev->dev, "Could not create pcie_replay_count");
3023 return r;
3024 }
108c6a63 3025
d155bef0
AB
3026 if (IS_ENABLED(CONFIG_PERF_EVENTS))
3027 r = amdgpu_pmu_init(adev);
9c7c85f7
JK
3028 if (r)
3029 dev_err(adev->dev, "amdgpu_pmu_init failed\n");
3030
d38ceaf9 3031 return 0;
83ba126a
AD
3032
3033failed:
89041940 3034 amdgpu_vf_error_trans_all(adev);
83ba126a
AD
3035 if (runtime)
3036 vga_switcheroo_fini_domain_pm_ops(adev->dev);
8840a387 3037
83ba126a 3038 return r;
d38ceaf9
AD
3039}
3040
d38ceaf9
AD
3041/**
3042 * amdgpu_device_fini - tear down the driver
3043 *
3044 * @adev: amdgpu_device pointer
3045 *
3046 * Tear down the driver info (all asics).
3047 * Called at driver shutdown.
3048 */
3049void amdgpu_device_fini(struct amdgpu_device *adev)
3050{
3051 int r;
3052
3053 DRM_INFO("amdgpu: finishing device.\n");
3054 adev->shutdown = true;
e5b03032
ML
3055 /* disable all interrupts */
3056 amdgpu_irq_disable_all(adev);
ff97cba8
ML
3057 if (adev->mode_info.mode_config_initialized){
3058 if (!amdgpu_device_has_dc_support(adev))
c2d88e06 3059 drm_helper_force_disable_all(adev->ddev);
ff97cba8
ML
3060 else
3061 drm_atomic_helper_shutdown(adev->ddev);
3062 }
d38ceaf9 3063 amdgpu_fence_driver_fini(adev);
58e955d9 3064 amdgpu_pm_sysfs_fini(adev);
d38ceaf9 3065 amdgpu_fbdev_fini(adev);
06ec9070 3066 r = amdgpu_device_ip_fini(adev);
ab4fe3e1
HR
3067 if (adev->firmware.gpu_info_fw) {
3068 release_firmware(adev->firmware.gpu_info_fw);
3069 adev->firmware.gpu_info_fw = NULL;
3070 }
d38ceaf9 3071 adev->accel_working = false;
beff74bc 3072 cancel_delayed_work_sync(&adev->delayed_init_work);
d38ceaf9 3073 /* free i2c buses */
4562236b
HW
3074 if (!amdgpu_device_has_dc_support(adev))
3075 amdgpu_i2c_fini(adev);
bfca0289
SL
3076
3077 if (amdgpu_emu_mode != 1)
3078 amdgpu_atombios_fini(adev);
3079
d38ceaf9
AD
3080 kfree(adev->bios);
3081 adev->bios = NULL;
84c8b22e
LW
3082 if (!pci_is_thunderbolt_attached(adev->pdev))
3083 vga_switcheroo_unregister_client(adev->pdev);
83ba126a
AD
3084 if (adev->flags & AMD_IS_PX)
3085 vga_switcheroo_fini_domain_pm_ops(adev->dev);
d38ceaf9
AD
3086 vga_client_register(adev->pdev, NULL, NULL, NULL);
3087 if (adev->rio_mem)
3088 pci_iounmap(adev->pdev, adev->rio_mem);
3089 adev->rio_mem = NULL;
3090 iounmap(adev->rmmio);
3091 adev->rmmio = NULL;
06ec9070 3092 amdgpu_device_doorbell_fini(adev);
e9bc1bf7
YT
3093 if (amdgpu_sriov_vf(adev) && amdgim_is_hwperf(adev))
3094 amdgpu_pm_virt_sysfs_fini(adev);
3095
d38ceaf9 3096 amdgpu_debugfs_regs_cleanup(adev);
dcea6e65 3097 device_remove_file(adev->dev, &dev_attr_pcie_replay_count);
5bb23532 3098 amdgpu_ucode_sysfs_fini(adev);
d155bef0
AB
3099 if (IS_ENABLED(CONFIG_PERF_EVENTS))
3100 amdgpu_pmu_fini(adev);
6698a3d0 3101 amdgpu_debugfs_preempt_cleanup(adev);
f54eeab4 3102 if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10)
a190d1c7 3103 amdgpu_discovery_fini(adev);
d38ceaf9
AD
3104}
3105
3106
3107/*
3108 * Suspend & resume.
3109 */
3110/**
810ddc3a 3111 * amdgpu_device_suspend - initiate device suspend
d38ceaf9 3112 *
87e3f136
DP
3113 * @dev: drm dev pointer
3114 * @suspend: suspend state
3115 * @fbcon : notify the fbdev of suspend
d38ceaf9
AD
3116 *
3117 * Puts the hw in the suspend state (all asics).
3118 * Returns 0 for success or an error on failure.
3119 * Called at driver suspend.
3120 */
810ddc3a 3121int amdgpu_device_suspend(struct drm_device *dev, bool suspend, bool fbcon)
d38ceaf9
AD
3122{
3123 struct amdgpu_device *adev;
3124 struct drm_crtc *crtc;
3125 struct drm_connector *connector;
f8d2d39e 3126 struct drm_connector_list_iter iter;
5ceb54c6 3127 int r;
d38ceaf9
AD
3128
3129 if (dev == NULL || dev->dev_private == NULL) {
3130 return -ENODEV;
3131 }
3132
3133 adev = dev->dev_private;
3134
3135 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
3136 return 0;
3137
44779b43 3138 adev->in_suspend = true;
d38ceaf9
AD
3139 drm_kms_helper_poll_disable(dev);
3140
5f818173
S
3141 if (fbcon)
3142 amdgpu_fbdev_set_suspend(adev, 1);
3143
beff74bc 3144 cancel_delayed_work_sync(&adev->delayed_init_work);
a5459475 3145
4562236b
HW
3146 if (!amdgpu_device_has_dc_support(adev)) {
3147 /* turn off display hw */
3148 drm_modeset_lock_all(dev);
f8d2d39e
LP
3149 drm_connector_list_iter_begin(dev, &iter);
3150 drm_for_each_connector_iter(connector, &iter)
3151 drm_helper_connector_dpms(connector,
3152 DRM_MODE_DPMS_OFF);
3153 drm_connector_list_iter_end(&iter);
4562236b 3154 drm_modeset_unlock_all(dev);
fe1053b7
AD
3155 /* unpin the front buffers and cursors */
3156 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
3157 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
3158 struct drm_framebuffer *fb = crtc->primary->fb;
3159 struct amdgpu_bo *robj;
3160
91334223 3161 if (amdgpu_crtc->cursor_bo && !adev->enable_virtual_display) {
fe1053b7
AD
3162 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
3163 r = amdgpu_bo_reserve(aobj, true);
3164 if (r == 0) {
3165 amdgpu_bo_unpin(aobj);
3166 amdgpu_bo_unreserve(aobj);
3167 }
756e6880 3168 }
756e6880 3169
fe1053b7
AD
3170 if (fb == NULL || fb->obj[0] == NULL) {
3171 continue;
3172 }
3173 robj = gem_to_amdgpu_bo(fb->obj[0]);
3174 /* don't unpin kernel fb objects */
3175 if (!amdgpu_fbdev_robj_is_fb(adev, robj)) {
3176 r = amdgpu_bo_reserve(robj, true);
3177 if (r == 0) {
3178 amdgpu_bo_unpin(robj);
3179 amdgpu_bo_unreserve(robj);
3180 }
d38ceaf9
AD
3181 }
3182 }
3183 }
fe1053b7
AD
3184
3185 amdgpu_amdkfd_suspend(adev);
3186
5e6932fe 3187 amdgpu_ras_suspend(adev);
3188
fe1053b7
AD
3189 r = amdgpu_device_ip_suspend_phase1(adev);
3190
d38ceaf9
AD
3191 /* evict vram memory */
3192 amdgpu_bo_evict_vram(adev);
3193
5ceb54c6 3194 amdgpu_fence_driver_suspend(adev);
d38ceaf9 3195
fe1053b7 3196 r = amdgpu_device_ip_suspend_phase2(adev);
d38ceaf9 3197
a0a71e49
AD
3198 /* evict remaining vram memory
3199 * This second call to evict vram is to evict the gart page table
3200 * using the CPU.
3201 */
d38ceaf9
AD
3202 amdgpu_bo_evict_vram(adev);
3203
d38ceaf9 3204 if (suspend) {
803cc26d 3205 pci_save_state(dev->pdev);
d38ceaf9
AD
3206 /* Shut down the device */
3207 pci_disable_device(dev->pdev);
3208 pci_set_power_state(dev->pdev, PCI_D3hot);
3209 }
3210
d38ceaf9
AD
3211 return 0;
3212}
3213
3214/**
810ddc3a 3215 * amdgpu_device_resume - initiate device resume
d38ceaf9 3216 *
87e3f136
DP
3217 * @dev: drm dev pointer
3218 * @resume: resume state
3219 * @fbcon : notify the fbdev of resume
d38ceaf9
AD
3220 *
3221 * Bring the hw back to operating state (all asics).
3222 * Returns 0 for success or an error on failure.
3223 * Called at driver resume.
3224 */
810ddc3a 3225int amdgpu_device_resume(struct drm_device *dev, bool resume, bool fbcon)
d38ceaf9
AD
3226{
3227 struct drm_connector *connector;
f8d2d39e 3228 struct drm_connector_list_iter iter;
d38ceaf9 3229 struct amdgpu_device *adev = dev->dev_private;
756e6880 3230 struct drm_crtc *crtc;
03161a6e 3231 int r = 0;
d38ceaf9
AD
3232
3233 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
3234 return 0;
3235
d38ceaf9
AD
3236 if (resume) {
3237 pci_set_power_state(dev->pdev, PCI_D0);
3238 pci_restore_state(dev->pdev);
74b0b157 3239 r = pci_enable_device(dev->pdev);
03161a6e 3240 if (r)
4d3b9ae5 3241 return r;
d38ceaf9
AD
3242 }
3243
3244 /* post card */
39c640c0 3245 if (amdgpu_device_need_post(adev)) {
74b0b157 3246 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
3247 if (r)
3248 DRM_ERROR("amdgpu asic init failed\n");
3249 }
d38ceaf9 3250
06ec9070 3251 r = amdgpu_device_ip_resume(adev);
e6707218 3252 if (r) {
06ec9070 3253 DRM_ERROR("amdgpu_device_ip_resume failed (%d).\n", r);
4d3b9ae5 3254 return r;
e6707218 3255 }
5ceb54c6
AD
3256 amdgpu_fence_driver_resume(adev);
3257
d38ceaf9 3258
06ec9070 3259 r = amdgpu_device_ip_late_init(adev);
03161a6e 3260 if (r)
4d3b9ae5 3261 return r;
d38ceaf9 3262
beff74bc
AD
3263 queue_delayed_work(system_wq, &adev->delayed_init_work,
3264 msecs_to_jiffies(AMDGPU_RESUME_MS));
3265
fe1053b7
AD
3266 if (!amdgpu_device_has_dc_support(adev)) {
3267 /* pin cursors */
3268 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
3269 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
3270
91334223 3271 if (amdgpu_crtc->cursor_bo && !adev->enable_virtual_display) {
fe1053b7
AD
3272 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
3273 r = amdgpu_bo_reserve(aobj, true);
3274 if (r == 0) {
3275 r = amdgpu_bo_pin(aobj, AMDGPU_GEM_DOMAIN_VRAM);
3276 if (r != 0)
3277 DRM_ERROR("Failed to pin cursor BO (%d)\n", r);
3278 amdgpu_crtc->cursor_addr = amdgpu_bo_gpu_offset(aobj);
3279 amdgpu_bo_unreserve(aobj);
3280 }
756e6880
AD
3281 }
3282 }
3283 }
ba997709
YZ
3284 r = amdgpu_amdkfd_resume(adev);
3285 if (r)
3286 return r;
756e6880 3287
96a5d8d4 3288 /* Make sure IB tests flushed */
beff74bc 3289 flush_delayed_work(&adev->delayed_init_work);
96a5d8d4 3290
d38ceaf9
AD
3291 /* blat the mode back in */
3292 if (fbcon) {
4562236b
HW
3293 if (!amdgpu_device_has_dc_support(adev)) {
3294 /* pre DCE11 */
3295 drm_helper_resume_force_mode(dev);
3296
3297 /* turn on display hw */
3298 drm_modeset_lock_all(dev);
f8d2d39e
LP
3299
3300 drm_connector_list_iter_begin(dev, &iter);
3301 drm_for_each_connector_iter(connector, &iter)
3302 drm_helper_connector_dpms(connector,
3303 DRM_MODE_DPMS_ON);
3304 drm_connector_list_iter_end(&iter);
3305
4562236b 3306 drm_modeset_unlock_all(dev);
d38ceaf9 3307 }
4d3b9ae5 3308 amdgpu_fbdev_set_suspend(adev, 0);
d38ceaf9
AD
3309 }
3310
3311 drm_kms_helper_poll_enable(dev);
23a1a9e5 3312
5e6932fe 3313 amdgpu_ras_resume(adev);
3314
23a1a9e5
L
3315 /*
3316 * Most of the connector probing functions try to acquire runtime pm
3317 * refs to ensure that the GPU is powered on when connector polling is
3318 * performed. Since we're calling this from a runtime PM callback,
3319 * trying to acquire rpm refs will cause us to deadlock.
3320 *
3321 * Since we're guaranteed to be holding the rpm lock, it's safe to
3322 * temporarily disable the rpm helpers so this doesn't deadlock us.
3323 */
3324#ifdef CONFIG_PM
3325 dev->dev->power.disable_depth++;
3326#endif
4562236b
HW
3327 if (!amdgpu_device_has_dc_support(adev))
3328 drm_helper_hpd_irq_event(dev);
3329 else
3330 drm_kms_helper_hotplug_event(dev);
23a1a9e5
L
3331#ifdef CONFIG_PM
3332 dev->dev->power.disable_depth--;
3333#endif
44779b43
RZ
3334 adev->in_suspend = false;
3335
4d3b9ae5 3336 return 0;
d38ceaf9
AD
3337}
3338
e3ecdffa
AD
3339/**
3340 * amdgpu_device_ip_check_soft_reset - did soft reset succeed
3341 *
3342 * @adev: amdgpu_device pointer
3343 *
3344 * The list of all the hardware IPs that make up the asic is walked and
3345 * the check_soft_reset callbacks are run. check_soft_reset determines
3346 * if the asic is still hung or not.
3347 * Returns true if any of the IPs are still in a hung state, false if not.
3348 */
06ec9070 3349static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)
63fbf42f
CZ
3350{
3351 int i;
3352 bool asic_hang = false;
3353
f993d628
ML
3354 if (amdgpu_sriov_vf(adev))
3355 return true;
3356
8bc04c29
AD
3357 if (amdgpu_asic_need_full_reset(adev))
3358 return true;
3359
63fbf42f 3360 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 3361 if (!adev->ip_blocks[i].status.valid)
63fbf42f 3362 continue;
a1255107
AD
3363 if (adev->ip_blocks[i].version->funcs->check_soft_reset)
3364 adev->ip_blocks[i].status.hang =
3365 adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
3366 if (adev->ip_blocks[i].status.hang) {
3367 DRM_INFO("IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
63fbf42f
CZ
3368 asic_hang = true;
3369 }
3370 }
3371 return asic_hang;
3372}
3373
e3ecdffa
AD
3374/**
3375 * amdgpu_device_ip_pre_soft_reset - prepare for soft reset
3376 *
3377 * @adev: amdgpu_device pointer
3378 *
3379 * The list of all the hardware IPs that make up the asic is walked and the
3380 * pre_soft_reset callbacks are run if the block is hung. pre_soft_reset
3381 * handles any IP specific hardware or software state changes that are
3382 * necessary for a soft reset to succeed.
3383 * Returns 0 on success, negative error code on failure.
3384 */
06ec9070 3385static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)
d31a501e
CZ
3386{
3387 int i, r = 0;
3388
3389 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 3390 if (!adev->ip_blocks[i].status.valid)
d31a501e 3391 continue;
a1255107
AD
3392 if (adev->ip_blocks[i].status.hang &&
3393 adev->ip_blocks[i].version->funcs->pre_soft_reset) {
3394 r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
d31a501e
CZ
3395 if (r)
3396 return r;
3397 }
3398 }
3399
3400 return 0;
3401}
3402
e3ecdffa
AD
3403/**
3404 * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed
3405 *
3406 * @adev: amdgpu_device pointer
3407 *
3408 * Some hardware IPs cannot be soft reset. If they are hung, a full gpu
3409 * reset is necessary to recover.
3410 * Returns true if a full asic reset is required, false if not.
3411 */
06ec9070 3412static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)
35d782fe 3413{
da146d3b
AD
3414 int i;
3415
8bc04c29
AD
3416 if (amdgpu_asic_need_full_reset(adev))
3417 return true;
3418
da146d3b 3419 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 3420 if (!adev->ip_blocks[i].status.valid)
da146d3b 3421 continue;
a1255107
AD
3422 if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
3423 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
3424 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
98512bb8
KW
3425 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
3426 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
a1255107 3427 if (adev->ip_blocks[i].status.hang) {
da146d3b
AD
3428 DRM_INFO("Some block need full reset!\n");
3429 return true;
3430 }
3431 }
35d782fe
CZ
3432 }
3433 return false;
3434}
3435
e3ecdffa
AD
3436/**
3437 * amdgpu_device_ip_soft_reset - do a soft reset
3438 *
3439 * @adev: amdgpu_device pointer
3440 *
3441 * The list of all the hardware IPs that make up the asic is walked and the
3442 * soft_reset callbacks are run if the block is hung. soft_reset handles any
3443 * IP specific hardware or software state changes that are necessary to soft
3444 * reset the IP.
3445 * Returns 0 on success, negative error code on failure.
3446 */
06ec9070 3447static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)
35d782fe
CZ
3448{
3449 int i, r = 0;
3450
3451 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 3452 if (!adev->ip_blocks[i].status.valid)
35d782fe 3453 continue;
a1255107
AD
3454 if (adev->ip_blocks[i].status.hang &&
3455 adev->ip_blocks[i].version->funcs->soft_reset) {
3456 r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
35d782fe
CZ
3457 if (r)
3458 return r;
3459 }
3460 }
3461
3462 return 0;
3463}
3464
e3ecdffa
AD
3465/**
3466 * amdgpu_device_ip_post_soft_reset - clean up from soft reset
3467 *
3468 * @adev: amdgpu_device pointer
3469 *
3470 * The list of all the hardware IPs that make up the asic is walked and the
3471 * post_soft_reset callbacks are run if the asic was hung. post_soft_reset
3472 * handles any IP specific hardware or software state changes that are
3473 * necessary after the IP has been soft reset.
3474 * Returns 0 on success, negative error code on failure.
3475 */
06ec9070 3476static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)
35d782fe
CZ
3477{
3478 int i, r = 0;
3479
3480 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 3481 if (!adev->ip_blocks[i].status.valid)
35d782fe 3482 continue;
a1255107
AD
3483 if (adev->ip_blocks[i].status.hang &&
3484 adev->ip_blocks[i].version->funcs->post_soft_reset)
3485 r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
35d782fe
CZ
3486 if (r)
3487 return r;
3488 }
3489
3490 return 0;
3491}
3492
e3ecdffa 3493/**
c33adbc7 3494 * amdgpu_device_recover_vram - Recover some VRAM contents
e3ecdffa
AD
3495 *
3496 * @adev: amdgpu_device pointer
3497 *
3498 * Restores the contents of VRAM buffers from the shadows in GTT. Used to
3499 * restore things like GPUVM page tables after a GPU reset where
3500 * the contents of VRAM might be lost.
403009bf
CK
3501 *
3502 * Returns:
3503 * 0 on success, negative error code on failure.
e3ecdffa 3504 */
c33adbc7 3505static int amdgpu_device_recover_vram(struct amdgpu_device *adev)
c41d1cf6 3506{
c41d1cf6 3507 struct dma_fence *fence = NULL, *next = NULL;
403009bf
CK
3508 struct amdgpu_bo *shadow;
3509 long r = 1, tmo;
c41d1cf6
ML
3510
3511 if (amdgpu_sriov_runtime(adev))
b045d3af 3512 tmo = msecs_to_jiffies(8000);
c41d1cf6
ML
3513 else
3514 tmo = msecs_to_jiffies(100);
3515
3516 DRM_INFO("recover vram bo from shadow start\n");
3517 mutex_lock(&adev->shadow_list_lock);
403009bf
CK
3518 list_for_each_entry(shadow, &adev->shadow_list, shadow_list) {
3519
3520 /* No need to recover an evicted BO */
3521 if (shadow->tbo.mem.mem_type != TTM_PL_TT ||
b575f10d 3522 shadow->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET ||
403009bf
CK
3523 shadow->parent->tbo.mem.mem_type != TTM_PL_VRAM)
3524 continue;
3525
3526 r = amdgpu_bo_restore_shadow(shadow, &next);
3527 if (r)
3528 break;
3529
c41d1cf6 3530 if (fence) {
1712fb1a 3531 tmo = dma_fence_wait_timeout(fence, false, tmo);
403009bf
CK
3532 dma_fence_put(fence);
3533 fence = next;
1712fb1a 3534 if (tmo == 0) {
3535 r = -ETIMEDOUT;
c41d1cf6 3536 break;
1712fb1a 3537 } else if (tmo < 0) {
3538 r = tmo;
3539 break;
3540 }
403009bf
CK
3541 } else {
3542 fence = next;
c41d1cf6 3543 }
c41d1cf6
ML
3544 }
3545 mutex_unlock(&adev->shadow_list_lock);
3546
403009bf
CK
3547 if (fence)
3548 tmo = dma_fence_wait_timeout(fence, false, tmo);
c41d1cf6
ML
3549 dma_fence_put(fence);
3550
1712fb1a 3551 if (r < 0 || tmo <= 0) {
3552 DRM_ERROR("recover vram bo from shadow failed, r is %ld, tmo is %ld\n", r, tmo);
403009bf
CK
3553 return -EIO;
3554 }
c41d1cf6 3555
403009bf
CK
3556 DRM_INFO("recover vram bo from shadow done\n");
3557 return 0;
c41d1cf6
ML
3558}
3559
a90ad3c2 3560
e3ecdffa 3561/**
06ec9070 3562 * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
5740682e
ML
3563 *
3564 * @adev: amdgpu device pointer
87e3f136 3565 * @from_hypervisor: request from hypervisor
5740682e
ML
3566 *
3567 * do VF FLR and reinitialize Asic
3f48c681 3568 * return 0 means succeeded otherwise failed
e3ecdffa
AD
3569 */
3570static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
3571 bool from_hypervisor)
5740682e
ML
3572{
3573 int r;
3574
3575 if (from_hypervisor)
3576 r = amdgpu_virt_request_full_gpu(adev, true);
3577 else
3578 r = amdgpu_virt_reset_gpu(adev);
3579 if (r)
3580 return r;
a90ad3c2 3581
f81e8d53
WL
3582 amdgpu_amdkfd_pre_reset(adev);
3583
a90ad3c2 3584 /* Resume IP prior to SMC */
06ec9070 3585 r = amdgpu_device_ip_reinit_early_sriov(adev);
5740682e
ML
3586 if (r)
3587 goto error;
a90ad3c2
ML
3588
3589 /* we need recover gart prior to run SMC/CP/SDMA resume */
c1c7ce8f 3590 amdgpu_gtt_mgr_recover(&adev->mman.bdev.man[TTM_PL_TT]);
a90ad3c2 3591
7a3e0bb2
RZ
3592 r = amdgpu_device_fw_loading(adev);
3593 if (r)
3594 return r;
3595
a90ad3c2 3596 /* now we are okay to resume SMC/CP/SDMA */
06ec9070 3597 r = amdgpu_device_ip_reinit_late_sriov(adev);
5740682e
ML
3598 if (r)
3599 goto error;
a90ad3c2
ML
3600
3601 amdgpu_irq_gpu_reset_resume_helper(adev);
5740682e 3602 r = amdgpu_ib_ring_tests(adev);
f81e8d53 3603 amdgpu_amdkfd_post_reset(adev);
a90ad3c2 3604
abc34253 3605error:
d3c117e5 3606 amdgpu_virt_init_data_exchange(adev);
abc34253 3607 amdgpu_virt_release_full_gpu(adev, true);
c41d1cf6 3608 if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
e3526257 3609 amdgpu_inc_vram_lost(adev);
c33adbc7 3610 r = amdgpu_device_recover_vram(adev);
a90ad3c2
ML
3611 }
3612
3613 return r;
3614}
3615
12938fad
CK
3616/**
3617 * amdgpu_device_should_recover_gpu - check if we should try GPU recovery
3618 *
3619 * @adev: amdgpu device pointer
3620 *
3621 * Check amdgpu_gpu_recovery and SRIOV status to see if we should try to recover
3622 * a hung GPU.
3623 */
3624bool amdgpu_device_should_recover_gpu(struct amdgpu_device *adev)
3625{
3626 if (!amdgpu_device_ip_check_soft_reset(adev)) {
3627 DRM_INFO("Timeout, but no hardware hang detected.\n");
3628 return false;
3629 }
3630
3ba7b418
AG
3631 if (amdgpu_gpu_recovery == 0)
3632 goto disabled;
3633
3634 if (amdgpu_sriov_vf(adev))
3635 return true;
3636
3637 if (amdgpu_gpu_recovery == -1) {
3638 switch (adev->asic_type) {
fc42d47c
AG
3639 case CHIP_BONAIRE:
3640 case CHIP_HAWAII:
3ba7b418
AG
3641 case CHIP_TOPAZ:
3642 case CHIP_TONGA:
3643 case CHIP_FIJI:
3644 case CHIP_POLARIS10:
3645 case CHIP_POLARIS11:
3646 case CHIP_POLARIS12:
3647 case CHIP_VEGAM:
3648 case CHIP_VEGA20:
3649 case CHIP_VEGA10:
3650 case CHIP_VEGA12:
c43b849f 3651 case CHIP_RAVEN:
3ba7b418
AG
3652 break;
3653 default:
3654 goto disabled;
3655 }
12938fad
CK
3656 }
3657
3658 return true;
3ba7b418
AG
3659
3660disabled:
3661 DRM_INFO("GPU recovery disabled.\n");
3662 return false;
12938fad
CK
3663}
3664
5c6dd71e 3665
26bc5340
AG
3666static int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,
3667 struct amdgpu_job *job,
3668 bool *need_full_reset_arg)
3669{
3670 int i, r = 0;
3671 bool need_full_reset = *need_full_reset_arg;
71182665 3672
71182665 3673 /* block all schedulers and reset given job's ring */
0875dc9e
CZ
3674 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3675 struct amdgpu_ring *ring = adev->rings[i];
3676
51687759 3677 if (!ring || !ring->sched.thread)
0875dc9e 3678 continue;
5740682e 3679
2f9d4084
ML
3680 /* after all hw jobs are reset, hw fence is meaningless, so force_completion */
3681 amdgpu_fence_driver_force_completion(ring);
0875dc9e 3682 }
d38ceaf9 3683
222b5f04
AG
3684 if(job)
3685 drm_sched_increase_karma(&job->base);
3686
1d721ed6 3687 /* Don't suspend on bare metal if we are not going to HW reset the ASIC */
26bc5340
AG
3688 if (!amdgpu_sriov_vf(adev)) {
3689
3690 if (!need_full_reset)
3691 need_full_reset = amdgpu_device_ip_need_full_reset(adev);
3692
3693 if (!need_full_reset) {
3694 amdgpu_device_ip_pre_soft_reset(adev);
3695 r = amdgpu_device_ip_soft_reset(adev);
3696 amdgpu_device_ip_post_soft_reset(adev);
3697 if (r || amdgpu_device_ip_check_soft_reset(adev)) {
3698 DRM_INFO("soft reset failed, will fallback to full reset!\n");
3699 need_full_reset = true;
3700 }
3701 }
3702
3703 if (need_full_reset)
3704 r = amdgpu_device_ip_suspend(adev);
3705
3706 *need_full_reset_arg = need_full_reset;
3707 }
3708
3709 return r;
3710}
3711
3712static int amdgpu_do_asic_reset(struct amdgpu_hive_info *hive,
3713 struct list_head *device_list_handle,
3714 bool *need_full_reset_arg)
3715{
3716 struct amdgpu_device *tmp_adev = NULL;
3717 bool need_full_reset = *need_full_reset_arg, vram_lost = false;
3718 int r = 0;
3719
3720 /*
3721 * ASIC reset has to be done on all HGMI hive nodes ASAP
3722 * to allow proper links negotiation in FW (within 1 sec)
3723 */
3724 if (need_full_reset) {
3725 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
d4535e2c
AG
3726 /* For XGMI run all resets in parallel to speed up the process */
3727 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
3728 if (!queue_work(system_highpri_wq, &tmp_adev->xgmi_reset_work))
3729 r = -EALREADY;
3730 } else
3731 r = amdgpu_asic_reset(tmp_adev);
3732
3733 if (r) {
fed184e9 3734 DRM_ERROR("ASIC reset failed with error, %d for drm dev, %s",
26bc5340 3735 r, tmp_adev->ddev->unique);
d4535e2c
AG
3736 break;
3737 }
3738 }
3739
3740 /* For XGMI wait for all PSP resets to complete before proceed */
3741 if (!r) {
3742 list_for_each_entry(tmp_adev, device_list_handle,
3743 gmc.xgmi.head) {
3744 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
3745 flush_work(&tmp_adev->xgmi_reset_work);
3746 r = tmp_adev->asic_reset_res;
3747 if (r)
3748 break;
3749 }
3750 }
26bc5340
AG
3751 }
3752 }
3753
3754
3755 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
3756 if (need_full_reset) {
3757 /* post card */
3758 if (amdgpu_atom_asic_init(tmp_adev->mode_info.atom_context))
3759 DRM_WARN("asic atom init failed!");
3760
3761 if (!r) {
3762 dev_info(tmp_adev->dev, "GPU reset succeeded, trying to resume\n");
3763 r = amdgpu_device_ip_resume_phase1(tmp_adev);
3764 if (r)
3765 goto out;
3766
3767 vram_lost = amdgpu_device_check_vram_lost(tmp_adev);
3768 if (vram_lost) {
77e7f829 3769 DRM_INFO("VRAM is lost due to GPU reset!\n");
e3526257 3770 amdgpu_inc_vram_lost(tmp_adev);
26bc5340
AG
3771 }
3772
3773 r = amdgpu_gtt_mgr_recover(
3774 &tmp_adev->mman.bdev.man[TTM_PL_TT]);
3775 if (r)
3776 goto out;
3777
3778 r = amdgpu_device_fw_loading(tmp_adev);
3779 if (r)
3780 return r;
3781
3782 r = amdgpu_device_ip_resume_phase2(tmp_adev);
3783 if (r)
3784 goto out;
3785
3786 if (vram_lost)
3787 amdgpu_device_fill_reset_magic(tmp_adev);
3788
fdafb359
EQ
3789 /*
3790 * Add this ASIC as tracked as reset was already
3791 * complete successfully.
3792 */
3793 amdgpu_register_gpu_instance(tmp_adev);
3794
7c04ca50 3795 r = amdgpu_device_ip_late_init(tmp_adev);
3796 if (r)
3797 goto out;
3798
e79a04d5 3799 /* must succeed. */
511fdbc3 3800 amdgpu_ras_resume(tmp_adev);
e79a04d5 3801
26bc5340
AG
3802 /* Update PSP FW topology after reset */
3803 if (hive && tmp_adev->gmc.xgmi.num_physical_nodes > 1)
3804 r = amdgpu_xgmi_update_topology(hive, tmp_adev);
3805 }
3806 }
3807
3808
3809out:
3810 if (!r) {
3811 amdgpu_irq_gpu_reset_resume_helper(tmp_adev);
3812 r = amdgpu_ib_ring_tests(tmp_adev);
3813 if (r) {
3814 dev_err(tmp_adev->dev, "ib ring test failed (%d).\n", r);
3815 r = amdgpu_device_ip_suspend(tmp_adev);
3816 need_full_reset = true;
3817 r = -EAGAIN;
3818 goto end;
3819 }
3820 }
3821
3822 if (!r)
3823 r = amdgpu_device_recover_vram(tmp_adev);
3824 else
3825 tmp_adev->asic_reset_res = r;
3826 }
3827
3828end:
3829 *need_full_reset_arg = need_full_reset;
3830 return r;
3831}
3832
1d721ed6 3833static bool amdgpu_device_lock_adev(struct amdgpu_device *adev, bool trylock)
26bc5340 3834{
1d721ed6
AG
3835 if (trylock) {
3836 if (!mutex_trylock(&adev->lock_reset))
3837 return false;
3838 } else
3839 mutex_lock(&adev->lock_reset);
5740682e 3840
26bc5340
AG
3841 atomic_inc(&adev->gpu_reset_counter);
3842 adev->in_gpu_reset = 1;
a3a09142
AD
3843 switch (amdgpu_asic_reset_method(adev)) {
3844 case AMD_RESET_METHOD_MODE1:
3845 adev->mp1_state = PP_MP1_STATE_SHUTDOWN;
3846 break;
3847 case AMD_RESET_METHOD_MODE2:
3848 adev->mp1_state = PP_MP1_STATE_RESET;
3849 break;
3850 default:
3851 adev->mp1_state = PP_MP1_STATE_NONE;
3852 break;
3853 }
1d721ed6
AG
3854
3855 return true;
26bc5340 3856}
d38ceaf9 3857
26bc5340
AG
3858static void amdgpu_device_unlock_adev(struct amdgpu_device *adev)
3859{
89041940 3860 amdgpu_vf_error_trans_all(adev);
a3a09142 3861 adev->mp1_state = PP_MP1_STATE_NONE;
13a752e3
ML
3862 adev->in_gpu_reset = 0;
3863 mutex_unlock(&adev->lock_reset);
26bc5340
AG
3864}
3865
26bc5340
AG
3866/**
3867 * amdgpu_device_gpu_recover - reset the asic and recover scheduler
3868 *
3869 * @adev: amdgpu device pointer
3870 * @job: which job trigger hang
3871 *
3872 * Attempt to reset the GPU if it has hung (all asics).
3873 * Attempt to do soft-reset or full-reset and reinitialize Asic
3874 * Returns 0 for success or an error on failure.
3875 */
3876
3877int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
3878 struct amdgpu_job *job)
3879{
1d721ed6
AG
3880 struct list_head device_list, *device_list_handle = NULL;
3881 bool need_full_reset, job_signaled;
26bc5340 3882 struct amdgpu_hive_info *hive = NULL;
26bc5340 3883 struct amdgpu_device *tmp_adev = NULL;
1d721ed6 3884 int i, r = 0;
7c6e68c7 3885 bool in_ras_intr = amdgpu_ras_intr_triggered();
26bc5340 3886
d5ea093e
AG
3887 /*
3888 * Flush RAM to disk so that after reboot
3889 * the user can read log and see why the system rebooted.
3890 */
3891 if (in_ras_intr && amdgpu_ras_get_context(adev)->reboot) {
3892
3893 DRM_WARN("Emergency reboot.");
3894
3895 ksys_sync_helper();
3896 emergency_restart();
3897 }
3898
1d721ed6 3899 need_full_reset = job_signaled = false;
26bc5340
AG
3900 INIT_LIST_HEAD(&device_list);
3901
7c6e68c7 3902 dev_info(adev->dev, "GPU %s begin!\n", in_ras_intr ? "jobs stop":"reset");
26bc5340 3903
beff74bc 3904 cancel_delayed_work_sync(&adev->delayed_init_work);
c53e4db7 3905
1d721ed6
AG
3906 hive = amdgpu_get_xgmi_hive(adev, false);
3907
26bc5340 3908 /*
1d721ed6
AG
3909 * Here we trylock to avoid chain of resets executing from
3910 * either trigger by jobs on different adevs in XGMI hive or jobs on
3911 * different schedulers for same device while this TO handler is running.
3912 * We always reset all schedulers for device and all devices for XGMI
3913 * hive so that should take care of them too.
26bc5340 3914 */
1d721ed6
AG
3915
3916 if (hive && !mutex_trylock(&hive->reset_lock)) {
3917 DRM_INFO("Bailing on TDR for s_job:%llx, hive: %llx as another already in progress",
0b2d2c2e 3918 job ? job->base.id : -1, hive->hive_id);
26bc5340 3919 return 0;
1d721ed6 3920 }
26bc5340
AG
3921
3922 /* Start with adev pre asic reset first for soft reset check.*/
1d721ed6
AG
3923 if (!amdgpu_device_lock_adev(adev, !hive)) {
3924 DRM_INFO("Bailing on TDR for s_job:%llx, as another already in progress",
0b2d2c2e 3925 job ? job->base.id : -1);
1d721ed6 3926 return 0;
26bc5340
AG
3927 }
3928
7c6e68c7
AG
3929 /* Block kfd: SRIOV would do it separately */
3930 if (!amdgpu_sriov_vf(adev))
3931 amdgpu_amdkfd_pre_reset(adev);
3932
26bc5340 3933 /* Build list of devices to reset */
1d721ed6 3934 if (adev->gmc.xgmi.num_physical_nodes > 1) {
26bc5340 3935 if (!hive) {
7c6e68c7
AG
3936 /*unlock kfd: SRIOV would do it separately */
3937 if (!amdgpu_sriov_vf(adev))
3938 amdgpu_amdkfd_post_reset(adev);
26bc5340
AG
3939 amdgpu_device_unlock_adev(adev);
3940 return -ENODEV;
3941 }
3942
3943 /*
3944 * In case we are in XGMI hive mode device reset is done for all the
3945 * nodes in the hive to retrain all XGMI links and hence the reset
3946 * sequence is executed in loop on all nodes.
3947 */
3948 device_list_handle = &hive->device_list;
3949 } else {
3950 list_add_tail(&adev->gmc.xgmi.head, &device_list);
3951 device_list_handle = &device_list;
3952 }
3953
1d721ed6
AG
3954 /* block all schedulers and reset given job's ring */
3955 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
7c6e68c7 3956 if (tmp_adev != adev) {
12ffa55d 3957 amdgpu_device_lock_adev(tmp_adev, false);
7c6e68c7
AG
3958 if (!amdgpu_sriov_vf(tmp_adev))
3959 amdgpu_amdkfd_pre_reset(tmp_adev);
3960 }
3961
12ffa55d
AG
3962 /*
3963 * Mark these ASICs to be reseted as untracked first
3964 * And add them back after reset completed
3965 */
3966 amdgpu_unregister_gpu_instance(tmp_adev);
3967
f1c1314b 3968 /* disable ras on ALL IPs */
7c6e68c7 3969 if (!in_ras_intr && amdgpu_device_ip_need_full_reset(tmp_adev))
f1c1314b 3970 amdgpu_ras_suspend(tmp_adev);
3971
1d721ed6
AG
3972 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3973 struct amdgpu_ring *ring = tmp_adev->rings[i];
3974
3975 if (!ring || !ring->sched.thread)
3976 continue;
3977
0b2d2c2e 3978 drm_sched_stop(&ring->sched, job ? &job->base : NULL);
7c6e68c7
AG
3979
3980 if (in_ras_intr)
3981 amdgpu_job_stop_all_jobs_on_sched(&ring->sched);
1d721ed6
AG
3982 }
3983 }
3984
3985
7c6e68c7
AG
3986 if (in_ras_intr)
3987 goto skip_sched_resume;
3988
1d721ed6
AG
3989 /*
3990 * Must check guilty signal here since after this point all old
3991 * HW fences are force signaled.
3992 *
3993 * job->base holds a reference to parent fence
3994 */
3995 if (job && job->base.s_fence->parent &&
3996 dma_fence_is_signaled(job->base.s_fence->parent))
3997 job_signaled = true;
3998
1d721ed6
AG
3999 if (job_signaled) {
4000 dev_info(adev->dev, "Guilty job already signaled, skipping HW reset");
4001 goto skip_hw_reset;
4002 }
4003
4004
4005 /* Guilty job will be freed after this*/
0b2d2c2e 4006 r = amdgpu_device_pre_asic_reset(adev, job, &need_full_reset);
1d721ed6
AG
4007 if (r) {
4008 /*TODO Should we stop ?*/
4009 DRM_ERROR("GPU pre asic reset failed with err, %d for drm dev, %s ",
4010 r, adev->ddev->unique);
4011 adev->asic_reset_res = r;
4012 }
4013
26bc5340
AG
4014retry: /* Rest of adevs pre asic reset from XGMI hive. */
4015 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
4016
4017 if (tmp_adev == adev)
4018 continue;
4019
26bc5340
AG
4020 r = amdgpu_device_pre_asic_reset(tmp_adev,
4021 NULL,
4022 &need_full_reset);
4023 /*TODO Should we stop ?*/
4024 if (r) {
4025 DRM_ERROR("GPU pre asic reset failed with err, %d for drm dev, %s ",
4026 r, tmp_adev->ddev->unique);
4027 tmp_adev->asic_reset_res = r;
4028 }
4029 }
4030
4031 /* Actual ASIC resets if needed.*/
4032 /* TODO Implement XGMI hive reset logic for SRIOV */
4033 if (amdgpu_sriov_vf(adev)) {
4034 r = amdgpu_device_reset_sriov(adev, job ? false : true);
4035 if (r)
4036 adev->asic_reset_res = r;
4037 } else {
4038 r = amdgpu_do_asic_reset(hive, device_list_handle, &need_full_reset);
4039 if (r && r == -EAGAIN)
4040 goto retry;
4041 }
4042
1d721ed6
AG
4043skip_hw_reset:
4044
26bc5340
AG
4045 /* Post ASIC reset for all devs .*/
4046 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
7c6e68c7 4047
1d721ed6
AG
4048 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
4049 struct amdgpu_ring *ring = tmp_adev->rings[i];
4050
4051 if (!ring || !ring->sched.thread)
4052 continue;
4053
4054 /* No point to resubmit jobs if we didn't HW reset*/
4055 if (!tmp_adev->asic_reset_res && !job_signaled)
4056 drm_sched_resubmit_jobs(&ring->sched);
4057
4058 drm_sched_start(&ring->sched, !tmp_adev->asic_reset_res);
4059 }
4060
4061 if (!amdgpu_device_has_dc_support(tmp_adev) && !job_signaled) {
4062 drm_helper_resume_force_mode(tmp_adev->ddev);
4063 }
4064
4065 tmp_adev->asic_reset_res = 0;
26bc5340
AG
4066
4067 if (r) {
4068 /* bad news, how to tell it to userspace ? */
12ffa55d 4069 dev_info(tmp_adev->dev, "GPU reset(%d) failed\n", atomic_read(&tmp_adev->gpu_reset_counter));
26bc5340
AG
4070 amdgpu_vf_error_put(tmp_adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r);
4071 } else {
12ffa55d 4072 dev_info(tmp_adev->dev, "GPU reset(%d) succeeded!\n", atomic_read(&tmp_adev->gpu_reset_counter));
26bc5340 4073 }
7c6e68c7 4074 }
26bc5340 4075
7c6e68c7
AG
4076skip_sched_resume:
4077 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
4078 /*unlock kfd: SRIOV would do it separately */
4079 if (!in_ras_intr && !amdgpu_sriov_vf(tmp_adev))
4080 amdgpu_amdkfd_post_reset(tmp_adev);
26bc5340
AG
4081 amdgpu_device_unlock_adev(tmp_adev);
4082 }
4083
1d721ed6 4084 if (hive)
22d6575b 4085 mutex_unlock(&hive->reset_lock);
26bc5340
AG
4086
4087 if (r)
4088 dev_info(adev->dev, "GPU reset end with ret = %d\n", r);
d38ceaf9
AD
4089 return r;
4090}
4091
e3ecdffa
AD
4092/**
4093 * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot
4094 *
4095 * @adev: amdgpu_device pointer
4096 *
4097 * Fetchs and stores in the driver the PCIE capabilities (gen speed
4098 * and lanes) of the slot the device is in. Handles APUs and
4099 * virtualized environments where PCIE config space may not be available.
4100 */
5494d864 4101static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev)
d0dd7f0c 4102{
5d9a6330 4103 struct pci_dev *pdev;
c5313457
HK
4104 enum pci_bus_speed speed_cap, platform_speed_cap;
4105 enum pcie_link_width platform_link_width;
d0dd7f0c 4106
cd474ba0
AD
4107 if (amdgpu_pcie_gen_cap)
4108 adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
d0dd7f0c 4109
cd474ba0
AD
4110 if (amdgpu_pcie_lane_cap)
4111 adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
d0dd7f0c 4112
cd474ba0
AD
4113 /* covers APUs as well */
4114 if (pci_is_root_bus(adev->pdev->bus)) {
4115 if (adev->pm.pcie_gen_mask == 0)
4116 adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
4117 if (adev->pm.pcie_mlw_mask == 0)
4118 adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
d0dd7f0c 4119 return;
cd474ba0 4120 }
d0dd7f0c 4121
c5313457
HK
4122 if (adev->pm.pcie_gen_mask && adev->pm.pcie_mlw_mask)
4123 return;
4124
dbaa922b
AD
4125 pcie_bandwidth_available(adev->pdev, NULL,
4126 &platform_speed_cap, &platform_link_width);
c5313457 4127
cd474ba0 4128 if (adev->pm.pcie_gen_mask == 0) {
5d9a6330
AD
4129 /* asic caps */
4130 pdev = adev->pdev;
4131 speed_cap = pcie_get_speed_cap(pdev);
4132 if (speed_cap == PCI_SPEED_UNKNOWN) {
4133 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
cd474ba0
AD
4134 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4135 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
cd474ba0 4136 } else {
5d9a6330
AD
4137 if (speed_cap == PCIE_SPEED_16_0GT)
4138 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4139 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4140 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
4141 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4);
4142 else if (speed_cap == PCIE_SPEED_8_0GT)
4143 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4144 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4145 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
4146 else if (speed_cap == PCIE_SPEED_5_0GT)
4147 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4148 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2);
4149 else
4150 adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1;
4151 }
4152 /* platform caps */
c5313457 4153 if (platform_speed_cap == PCI_SPEED_UNKNOWN) {
5d9a6330
AD
4154 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4155 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
4156 } else {
c5313457 4157 if (platform_speed_cap == PCIE_SPEED_16_0GT)
5d9a6330
AD
4158 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4159 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4160 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
4161 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4);
c5313457 4162 else if (platform_speed_cap == PCIE_SPEED_8_0GT)
5d9a6330
AD
4163 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4164 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4165 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3);
c5313457 4166 else if (platform_speed_cap == PCIE_SPEED_5_0GT)
5d9a6330
AD
4167 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4168 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
4169 else
4170 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
4171
cd474ba0
AD
4172 }
4173 }
4174 if (adev->pm.pcie_mlw_mask == 0) {
c5313457 4175 if (platform_link_width == PCIE_LNK_WIDTH_UNKNOWN) {
5d9a6330
AD
4176 adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK;
4177 } else {
c5313457 4178 switch (platform_link_width) {
5d9a6330 4179 case PCIE_LNK_X32:
cd474ba0
AD
4180 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
4181 CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
4182 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
4183 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4184 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4185 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4186 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4187 break;
5d9a6330 4188 case PCIE_LNK_X16:
cd474ba0
AD
4189 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
4190 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
4191 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4192 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4193 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4194 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4195 break;
5d9a6330 4196 case PCIE_LNK_X12:
cd474ba0
AD
4197 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
4198 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4199 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4200 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4201 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4202 break;
5d9a6330 4203 case PCIE_LNK_X8:
cd474ba0
AD
4204 adev->pm.pcie_mlw_mask = (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_X4:
cd474ba0
AD
4210 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4211 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4212 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4213 break;
5d9a6330 4214 case PCIE_LNK_X2:
cd474ba0
AD
4215 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4216 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4217 break;
5d9a6330 4218 case PCIE_LNK_X1:
cd474ba0
AD
4219 adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;
4220 break;
4221 default:
4222 break;
4223 }
d0dd7f0c
AD
4224 }
4225 }
4226}
d38ceaf9 4227