]> git.ipfire.org Git - thirdparty/kernel/stable.git/blame - drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
drm/scheduler: fix param documentation
[thirdparty/kernel/stable.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>
d38ceaf9
AD
30#include <linux/console.h>
31#include <linux/slab.h>
d38ceaf9
AD
32#include <drm/drmP.h>
33#include <drm/drm_crtc_helper.h>
4562236b 34#include <drm/drm_atomic_helper.h>
d38ceaf9
AD
35#include <drm/amdgpu_drm.h>
36#include <linux/vgaarb.h>
37#include <linux/vga_switcheroo.h>
38#include <linux/efi.h>
39#include "amdgpu.h"
f4b373f4 40#include "amdgpu_trace.h"
d38ceaf9
AD
41#include "amdgpu_i2c.h"
42#include "atom.h"
43#include "amdgpu_atombios.h"
a5bde2f9 44#include "amdgpu_atomfirmware.h"
d0dd7f0c 45#include "amd_pcie.h"
33f34802
KW
46#ifdef CONFIG_DRM_AMDGPU_SI
47#include "si.h"
48#endif
a2e73f56
AD
49#ifdef CONFIG_DRM_AMDGPU_CIK
50#include "cik.h"
51#endif
aaa36a97 52#include "vi.h"
460826e6 53#include "soc15.h"
d38ceaf9 54#include "bif/bif_4_1_d.h"
9accf2fd 55#include <linux/pci.h>
bec86378 56#include <linux/firmware.h>
89041940 57#include "amdgpu_vf_error.h"
d38ceaf9 58
ba997709 59#include "amdgpu_amdkfd.h"
d2f52ac8 60#include "amdgpu_pm.h"
d38ceaf9 61
e2a75f88 62MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin");
3f76dced 63MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin");
2d2e5e7e 64MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin");
e2a75f88 65
2dc80b00
S
66#define AMDGPU_RESUME_MS 2000
67
d38ceaf9 68static const char *amdgpu_asic_name[] = {
da69c161
KW
69 "TAHITI",
70 "PITCAIRN",
71 "VERDE",
72 "OLAND",
73 "HAINAN",
d38ceaf9
AD
74 "BONAIRE",
75 "KAVERI",
76 "KABINI",
77 "HAWAII",
78 "MULLINS",
79 "TOPAZ",
80 "TONGA",
48299f95 81 "FIJI",
d38ceaf9 82 "CARRIZO",
139f4917 83 "STONEY",
2cc0c0b5
FC
84 "POLARIS10",
85 "POLARIS11",
c4642a47 86 "POLARIS12",
48ff108d 87 "VEGAM",
d4196f01 88 "VEGA10",
8fab806a 89 "VEGA12",
956fcddc 90 "VEGA20",
2ca8a5d2 91 "RAVEN",
d38ceaf9
AD
92 "LAST",
93};
94
5494d864
AD
95static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev);
96
e3ecdffa
AD
97/**
98 * amdgpu_device_is_px - Is the device is a dGPU with HG/PX power control
99 *
100 * @dev: drm_device pointer
101 *
102 * Returns true if the device is a dGPU with HG/PX power control,
103 * otherwise return false.
104 */
d38ceaf9
AD
105bool amdgpu_device_is_px(struct drm_device *dev)
106{
107 struct amdgpu_device *adev = dev->dev_private;
108
2f7d10b3 109 if (adev->flags & AMD_IS_PX)
d38ceaf9
AD
110 return true;
111 return false;
112}
113
114/*
115 * MMIO register access helper functions.
116 */
e3ecdffa
AD
117/**
118 * amdgpu_mm_rreg - read a memory mapped IO register
119 *
120 * @adev: amdgpu_device pointer
121 * @reg: dword aligned register offset
122 * @acc_flags: access flags which require special behavior
123 *
124 * Returns the 32 bit value from the offset specified.
125 */
d38ceaf9 126uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg,
15d72fd7 127 uint32_t acc_flags)
d38ceaf9 128{
f4b373f4
TSD
129 uint32_t ret;
130
43ca8efa 131 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
bc992ba5 132 return amdgpu_virt_kiq_rreg(adev, reg);
bc992ba5 133
15d72fd7 134 if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
f4b373f4 135 ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
d38ceaf9
AD
136 else {
137 unsigned long flags;
d38ceaf9
AD
138
139 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
140 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
141 ret = readl(((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
142 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
d38ceaf9 143 }
f4b373f4
TSD
144 trace_amdgpu_mm_rreg(adev->pdev->device, reg, ret);
145 return ret;
d38ceaf9
AD
146}
147
421a2a30
ML
148/*
149 * MMIO register read with bytes helper functions
150 * @offset:bytes offset from MMIO start
151 *
152*/
153
e3ecdffa
AD
154/**
155 * amdgpu_mm_rreg8 - read a memory mapped IO register
156 *
157 * @adev: amdgpu_device pointer
158 * @offset: byte aligned register offset
159 *
160 * Returns the 8 bit value from the offset specified.
161 */
421a2a30
ML
162uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset) {
163 if (offset < adev->rmmio_size)
164 return (readb(adev->rmmio + offset));
165 BUG();
166}
167
168/*
169 * MMIO register write with bytes helper functions
170 * @offset:bytes offset from MMIO start
171 * @value: the value want to be written to the register
172 *
173*/
e3ecdffa
AD
174/**
175 * amdgpu_mm_wreg8 - read a memory mapped IO register
176 *
177 * @adev: amdgpu_device pointer
178 * @offset: byte aligned register offset
179 * @value: 8 bit value to write
180 *
181 * Writes the value specified to the offset specified.
182 */
421a2a30
ML
183void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value) {
184 if (offset < adev->rmmio_size)
185 writeb(value, adev->rmmio + offset);
186 else
187 BUG();
188}
189
e3ecdffa
AD
190/**
191 * amdgpu_mm_wreg - write to a memory mapped IO register
192 *
193 * @adev: amdgpu_device pointer
194 * @reg: dword aligned register offset
195 * @v: 32 bit value to write to the register
196 * @acc_flags: access flags which require special behavior
197 *
198 * Writes the value specified to the offset specified.
199 */
d38ceaf9 200void amdgpu_mm_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v,
15d72fd7 201 uint32_t acc_flags)
d38ceaf9 202{
f4b373f4 203 trace_amdgpu_mm_wreg(adev->pdev->device, reg, v);
4e99a44e 204
47ed4e1c
KW
205 if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
206 adev->last_mm_index = v;
207 }
208
43ca8efa 209 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
bc992ba5 210 return amdgpu_virt_kiq_wreg(adev, reg, v);
bc992ba5 211
15d72fd7 212 if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
d38ceaf9
AD
213 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
214 else {
215 unsigned long flags;
216
217 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
218 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
219 writel(v, ((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
220 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
221 }
47ed4e1c
KW
222
223 if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
224 udelay(500);
225 }
d38ceaf9
AD
226}
227
e3ecdffa
AD
228/**
229 * amdgpu_io_rreg - read an IO register
230 *
231 * @adev: amdgpu_device pointer
232 * @reg: dword aligned register offset
233 *
234 * Returns the 32 bit value from the offset specified.
235 */
d38ceaf9
AD
236u32 amdgpu_io_rreg(struct amdgpu_device *adev, u32 reg)
237{
238 if ((reg * 4) < adev->rio_mem_size)
239 return ioread32(adev->rio_mem + (reg * 4));
240 else {
241 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
242 return ioread32(adev->rio_mem + (mmMM_DATA * 4));
243 }
244}
245
e3ecdffa
AD
246/**
247 * amdgpu_io_wreg - write to an IO register
248 *
249 * @adev: amdgpu_device pointer
250 * @reg: dword aligned register offset
251 * @v: 32 bit value to write to the register
252 *
253 * Writes the value specified to the offset specified.
254 */
d38ceaf9
AD
255void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32 v)
256{
47ed4e1c
KW
257 if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
258 adev->last_mm_index = v;
259 }
d38ceaf9
AD
260
261 if ((reg * 4) < adev->rio_mem_size)
262 iowrite32(v, adev->rio_mem + (reg * 4));
263 else {
264 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
265 iowrite32(v, adev->rio_mem + (mmMM_DATA * 4));
266 }
47ed4e1c
KW
267
268 if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
269 udelay(500);
270 }
d38ceaf9
AD
271}
272
273/**
274 * amdgpu_mm_rdoorbell - read a doorbell dword
275 *
276 * @adev: amdgpu_device pointer
277 * @index: doorbell index
278 *
279 * Returns the value in the doorbell aperture at the
280 * requested doorbell index (CIK).
281 */
282u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
283{
284 if (index < adev->doorbell.num_doorbells) {
285 return readl(adev->doorbell.ptr + index);
286 } else {
287 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
288 return 0;
289 }
290}
291
292/**
293 * amdgpu_mm_wdoorbell - write a doorbell dword
294 *
295 * @adev: amdgpu_device pointer
296 * @index: doorbell index
297 * @v: value to write
298 *
299 * Writes @v to the doorbell aperture at the
300 * requested doorbell index (CIK).
301 */
302void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v)
303{
304 if (index < adev->doorbell.num_doorbells) {
305 writel(v, adev->doorbell.ptr + index);
306 } else {
307 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
308 }
309}
310
832be404
KW
311/**
312 * amdgpu_mm_rdoorbell64 - read a doorbell Qword
313 *
314 * @adev: amdgpu_device pointer
315 * @index: doorbell index
316 *
317 * Returns the value in the doorbell aperture at the
318 * requested doorbell index (VEGA10+).
319 */
320u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index)
321{
322 if (index < adev->doorbell.num_doorbells) {
323 return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index));
324 } else {
325 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
326 return 0;
327 }
328}
329
330/**
331 * amdgpu_mm_wdoorbell64 - write a doorbell Qword
332 *
333 * @adev: amdgpu_device pointer
334 * @index: doorbell index
335 * @v: value to write
336 *
337 * Writes @v to the doorbell aperture at the
338 * requested doorbell index (VEGA10+).
339 */
340void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v)
341{
342 if (index < adev->doorbell.num_doorbells) {
343 atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v);
344 } else {
345 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
346 }
347}
348
d38ceaf9
AD
349/**
350 * amdgpu_invalid_rreg - dummy reg read function
351 *
352 * @adev: amdgpu device pointer
353 * @reg: offset of register
354 *
355 * Dummy register read function. Used for register blocks
356 * that certain asics don't have (all asics).
357 * Returns the value in the register.
358 */
359static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
360{
361 DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
362 BUG();
363 return 0;
364}
365
366/**
367 * amdgpu_invalid_wreg - dummy reg write function
368 *
369 * @adev: amdgpu device pointer
370 * @reg: offset of register
371 * @v: value to write to the register
372 *
373 * Dummy register read function. Used for register blocks
374 * that certain asics don't have (all asics).
375 */
376static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
377{
378 DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
379 reg, v);
380 BUG();
381}
382
383/**
384 * amdgpu_block_invalid_rreg - dummy reg read function
385 *
386 * @adev: amdgpu device pointer
387 * @block: offset of instance
388 * @reg: offset of register
389 *
390 * Dummy register read function. Used for register blocks
391 * that certain asics don't have (all asics).
392 * Returns the value in the register.
393 */
394static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
395 uint32_t block, uint32_t reg)
396{
397 DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
398 reg, block);
399 BUG();
400 return 0;
401}
402
403/**
404 * amdgpu_block_invalid_wreg - dummy reg write function
405 *
406 * @adev: amdgpu device pointer
407 * @block: offset of instance
408 * @reg: offset of register
409 * @v: value to write to the register
410 *
411 * Dummy register read function. Used for register blocks
412 * that certain asics don't have (all asics).
413 */
414static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
415 uint32_t block,
416 uint32_t reg, uint32_t v)
417{
418 DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
419 reg, block, v);
420 BUG();
421}
422
e3ecdffa
AD
423/**
424 * amdgpu_device_vram_scratch_init - allocate the VRAM scratch page
425 *
426 * @adev: amdgpu device pointer
427 *
428 * Allocates a scratch page of VRAM for use by various things in the
429 * driver.
430 */
06ec9070 431static int amdgpu_device_vram_scratch_init(struct amdgpu_device *adev)
d38ceaf9 432{
a4a02777
CK
433 return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE,
434 PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
435 &adev->vram_scratch.robj,
436 &adev->vram_scratch.gpu_addr,
437 (void **)&adev->vram_scratch.ptr);
d38ceaf9
AD
438}
439
e3ecdffa
AD
440/**
441 * amdgpu_device_vram_scratch_fini - Free the VRAM scratch page
442 *
443 * @adev: amdgpu device pointer
444 *
445 * Frees the VRAM scratch page.
446 */
06ec9070 447static void amdgpu_device_vram_scratch_fini(struct amdgpu_device *adev)
d38ceaf9 448{
078af1a3 449 amdgpu_bo_free_kernel(&adev->vram_scratch.robj, NULL, NULL);
d38ceaf9
AD
450}
451
452/**
9c3f2b54 453 * amdgpu_device_program_register_sequence - program an array of registers.
d38ceaf9
AD
454 *
455 * @adev: amdgpu_device pointer
456 * @registers: pointer to the register array
457 * @array_size: size of the register array
458 *
459 * Programs an array or registers with and and or masks.
460 * This is a helper for setting golden registers.
461 */
9c3f2b54
AD
462void amdgpu_device_program_register_sequence(struct amdgpu_device *adev,
463 const u32 *registers,
464 const u32 array_size)
d38ceaf9
AD
465{
466 u32 tmp, reg, and_mask, or_mask;
467 int i;
468
469 if (array_size % 3)
470 return;
471
472 for (i = 0; i < array_size; i +=3) {
473 reg = registers[i + 0];
474 and_mask = registers[i + 1];
475 or_mask = registers[i + 2];
476
477 if (and_mask == 0xffffffff) {
478 tmp = or_mask;
479 } else {
480 tmp = RREG32(reg);
481 tmp &= ~and_mask;
482 tmp |= or_mask;
483 }
484 WREG32(reg, tmp);
485 }
486}
487
e3ecdffa
AD
488/**
489 * amdgpu_device_pci_config_reset - reset the GPU
490 *
491 * @adev: amdgpu_device pointer
492 *
493 * Resets the GPU using the pci config reset sequence.
494 * Only applicable to asics prior to vega10.
495 */
8111c387 496void amdgpu_device_pci_config_reset(struct amdgpu_device *adev)
d38ceaf9
AD
497{
498 pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
499}
500
501/*
502 * GPU doorbell aperture helpers function.
503 */
504/**
06ec9070 505 * amdgpu_device_doorbell_init - Init doorbell driver information.
d38ceaf9
AD
506 *
507 * @adev: amdgpu_device pointer
508 *
509 * Init doorbell driver information (CIK)
510 * Returns 0 on success, error on failure.
511 */
06ec9070 512static int amdgpu_device_doorbell_init(struct amdgpu_device *adev)
d38ceaf9 513{
705e519e
CK
514 /* No doorbell on SI hardware generation */
515 if (adev->asic_type < CHIP_BONAIRE) {
516 adev->doorbell.base = 0;
517 adev->doorbell.size = 0;
518 adev->doorbell.num_doorbells = 0;
519 adev->doorbell.ptr = NULL;
520 return 0;
521 }
522
d6895ad3
CK
523 if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET)
524 return -EINVAL;
525
d38ceaf9
AD
526 /* doorbell bar mapping */
527 adev->doorbell.base = pci_resource_start(adev->pdev, 2);
528 adev->doorbell.size = pci_resource_len(adev->pdev, 2);
529
edf600da 530 adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32),
d38ceaf9
AD
531 AMDGPU_DOORBELL_MAX_ASSIGNMENT+1);
532 if (adev->doorbell.num_doorbells == 0)
533 return -EINVAL;
534
8972e5d2
CK
535 adev->doorbell.ptr = ioremap(adev->doorbell.base,
536 adev->doorbell.num_doorbells *
537 sizeof(u32));
538 if (adev->doorbell.ptr == NULL)
d38ceaf9 539 return -ENOMEM;
d38ceaf9
AD
540
541 return 0;
542}
543
544/**
06ec9070 545 * amdgpu_device_doorbell_fini - Tear down doorbell driver information.
d38ceaf9
AD
546 *
547 * @adev: amdgpu_device pointer
548 *
549 * Tear down doorbell driver information (CIK)
550 */
06ec9070 551static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev)
d38ceaf9
AD
552{
553 iounmap(adev->doorbell.ptr);
554 adev->doorbell.ptr = NULL;
555}
556
22cb0164 557
d38ceaf9
AD
558
559/*
06ec9070 560 * amdgpu_device_wb_*()
455a7bc2 561 * Writeback is the method by which the GPU updates special pages in memory
ea81a173 562 * with the status of certain GPU events (fences, ring pointers,etc.).
d38ceaf9
AD
563 */
564
565/**
06ec9070 566 * amdgpu_device_wb_fini - Disable Writeback and free memory
d38ceaf9
AD
567 *
568 * @adev: amdgpu_device pointer
569 *
570 * Disables Writeback and frees the Writeback memory (all asics).
571 * Used at driver shutdown.
572 */
06ec9070 573static void amdgpu_device_wb_fini(struct amdgpu_device *adev)
d38ceaf9
AD
574{
575 if (adev->wb.wb_obj) {
a76ed485
AD
576 amdgpu_bo_free_kernel(&adev->wb.wb_obj,
577 &adev->wb.gpu_addr,
578 (void **)&adev->wb.wb);
d38ceaf9
AD
579 adev->wb.wb_obj = NULL;
580 }
581}
582
583/**
06ec9070 584 * amdgpu_device_wb_init- Init Writeback driver info and allocate memory
d38ceaf9
AD
585 *
586 * @adev: amdgpu_device pointer
587 *
455a7bc2 588 * Initializes writeback and allocates writeback memory (all asics).
d38ceaf9
AD
589 * Used at driver startup.
590 * Returns 0 on success or an -error on failure.
591 */
06ec9070 592static int amdgpu_device_wb_init(struct amdgpu_device *adev)
d38ceaf9
AD
593{
594 int r;
595
596 if (adev->wb.wb_obj == NULL) {
97407b63
AD
597 /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */
598 r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8,
a76ed485
AD
599 PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
600 &adev->wb.wb_obj, &adev->wb.gpu_addr,
601 (void **)&adev->wb.wb);
d38ceaf9
AD
602 if (r) {
603 dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
604 return r;
605 }
d38ceaf9
AD
606
607 adev->wb.num_wb = AMDGPU_MAX_WB;
608 memset(&adev->wb.used, 0, sizeof(adev->wb.used));
609
610 /* clear wb memory */
73469585 611 memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8);
d38ceaf9
AD
612 }
613
614 return 0;
615}
616
617/**
131b4b36 618 * amdgpu_device_wb_get - Allocate a wb entry
d38ceaf9
AD
619 *
620 * @adev: amdgpu_device pointer
621 * @wb: wb index
622 *
623 * Allocate a wb slot for use by the driver (all asics).
624 * Returns 0 on success or -EINVAL on failure.
625 */
131b4b36 626int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb)
d38ceaf9
AD
627{
628 unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
d38ceaf9 629
97407b63 630 if (offset < adev->wb.num_wb) {
7014285a 631 __set_bit(offset, adev->wb.used);
63ae07ca 632 *wb = offset << 3; /* convert to dw offset */
0915fdbc
ML
633 return 0;
634 } else {
635 return -EINVAL;
636 }
637}
638
d38ceaf9 639/**
131b4b36 640 * amdgpu_device_wb_free - Free a wb entry
d38ceaf9
AD
641 *
642 * @adev: amdgpu_device pointer
643 * @wb: wb index
644 *
645 * Free a wb slot allocated for use by the driver (all asics)
646 */
131b4b36 647void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb)
d38ceaf9 648{
73469585 649 wb >>= 3;
d38ceaf9 650 if (wb < adev->wb.num_wb)
73469585 651 __clear_bit(wb, adev->wb.used);
d38ceaf9
AD
652}
653
654/**
2543e28a 655 * amdgpu_device_vram_location - try to find VRAM location
e3ecdffa 656 *
d38ceaf9
AD
657 * @adev: amdgpu device structure holding all necessary informations
658 * @mc: memory controller structure holding memory informations
659 * @base: base address at which to put VRAM
660 *
455a7bc2 661 * Function will try to place VRAM at base address provided
3d647c8f 662 * as parameter.
d38ceaf9 663 */
2543e28a 664void amdgpu_device_vram_location(struct amdgpu_device *adev,
770d13b1 665 struct amdgpu_gmc *mc, u64 base)
d38ceaf9
AD
666{
667 uint64_t limit = (uint64_t)amdgpu_vram_limit << 20;
668
669 mc->vram_start = base;
d38ceaf9
AD
670 mc->vram_end = mc->vram_start + mc->mc_vram_size - 1;
671 if (limit && limit < mc->real_vram_size)
672 mc->real_vram_size = limit;
673 dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
674 mc->mc_vram_size >> 20, mc->vram_start,
675 mc->vram_end, mc->real_vram_size >> 20);
676}
677
678/**
761f58e0 679 * amdgpu_device_gart_location - try to find GART location
e3ecdffa 680 *
d38ceaf9
AD
681 * @adev: amdgpu device structure holding all necessary informations
682 * @mc: memory controller structure holding memory informations
683 *
761f58e0 684 * Function will place try to place GART before or after VRAM.
d38ceaf9 685 *
761f58e0 686 * If GART size is bigger than space left then we ajust GART size.
d38ceaf9 687 * Thus function will never fails.
d38ceaf9 688 */
2543e28a 689void amdgpu_device_gart_location(struct amdgpu_device *adev,
770d13b1 690 struct amdgpu_gmc *mc)
d38ceaf9
AD
691{
692 u64 size_af, size_bf;
693
7951e376
RZ
694 mc->gart_size += adev->pm.smu_prv_buffer_size;
695
770d13b1 696 size_af = adev->gmc.mc_mask - mc->vram_end;
ed21c047 697 size_bf = mc->vram_start;
d38ceaf9 698 if (size_bf > size_af) {
6f02a696 699 if (mc->gart_size > size_bf) {
761f58e0 700 dev_warn(adev->dev, "limiting GART\n");
6f02a696 701 mc->gart_size = size_bf;
d38ceaf9 702 }
6f02a696 703 mc->gart_start = 0;
d38ceaf9 704 } else {
6f02a696 705 if (mc->gart_size > size_af) {
761f58e0 706 dev_warn(adev->dev, "limiting GART\n");
6f02a696 707 mc->gart_size = size_af;
d38ceaf9 708 }
b98f1b9e
CK
709 /* VCE doesn't like it when BOs cross a 4GB segment, so align
710 * the GART base on a 4GB boundary as well.
711 */
712 mc->gart_start = ALIGN(mc->vram_end + 1, 0x100000000ULL);
d38ceaf9 713 }
6f02a696 714 mc->gart_end = mc->gart_start + mc->gart_size - 1;
761f58e0 715 dev_info(adev->dev, "GART: %lluM 0x%016llX - 0x%016llX\n",
6f02a696 716 mc->gart_size >> 20, mc->gart_start, mc->gart_end);
d38ceaf9
AD
717}
718
d6895ad3
CK
719/**
720 * amdgpu_device_resize_fb_bar - try to resize FB BAR
721 *
722 * @adev: amdgpu_device pointer
723 *
724 * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not
725 * to fail, but if any of the BARs is not accessible after the size we abort
726 * driver loading by returning -ENODEV.
727 */
728int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev)
729{
770d13b1 730 u64 space_needed = roundup_pow_of_two(adev->gmc.real_vram_size);
d6895ad3 731 u32 rbar_size = order_base_2(((space_needed >> 20) | 1)) - 1;
31b8adab
CK
732 struct pci_bus *root;
733 struct resource *res;
734 unsigned i;
d6895ad3
CK
735 u16 cmd;
736 int r;
737
0c03b912 738 /* Bypass for VF */
739 if (amdgpu_sriov_vf(adev))
740 return 0;
741
31b8adab
CK
742 /* Check if the root BUS has 64bit memory resources */
743 root = adev->pdev->bus;
744 while (root->parent)
745 root = root->parent;
746
747 pci_bus_for_each_resource(root, res, i) {
0ebb7c54 748 if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) &&
31b8adab
CK
749 res->start > 0x100000000ull)
750 break;
751 }
752
753 /* Trying to resize is pointless without a root hub window above 4GB */
754 if (!res)
755 return 0;
756
d6895ad3
CK
757 /* Disable memory decoding while we change the BAR addresses and size */
758 pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd);
759 pci_write_config_word(adev->pdev, PCI_COMMAND,
760 cmd & ~PCI_COMMAND_MEMORY);
761
762 /* Free the VRAM and doorbell BAR, we most likely need to move both. */
06ec9070 763 amdgpu_device_doorbell_fini(adev);
d6895ad3
CK
764 if (adev->asic_type >= CHIP_BONAIRE)
765 pci_release_resource(adev->pdev, 2);
766
767 pci_release_resource(adev->pdev, 0);
768
769 r = pci_resize_resource(adev->pdev, 0, rbar_size);
770 if (r == -ENOSPC)
771 DRM_INFO("Not enough PCI address space for a large BAR.");
772 else if (r && r != -ENOTSUPP)
773 DRM_ERROR("Problem resizing BAR0 (%d).", r);
774
775 pci_assign_unassigned_bus_resources(adev->pdev->bus);
776
777 /* When the doorbell or fb BAR isn't available we have no chance of
778 * using the device.
779 */
06ec9070 780 r = amdgpu_device_doorbell_init(adev);
d6895ad3
CK
781 if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET))
782 return -ENODEV;
783
784 pci_write_config_word(adev->pdev, PCI_COMMAND, cmd);
785
786 return 0;
787}
a05502e5 788
d38ceaf9
AD
789/*
790 * GPU helpers function.
791 */
792/**
39c640c0 793 * amdgpu_device_need_post - check if the hw need post or not
d38ceaf9
AD
794 *
795 * @adev: amdgpu_device pointer
796 *
c836fec5
JQ
797 * Check if the asic has been initialized (all asics) at driver startup
798 * or post is needed if hw reset is performed.
799 * Returns true if need or false if not.
d38ceaf9 800 */
39c640c0 801bool amdgpu_device_need_post(struct amdgpu_device *adev)
d38ceaf9
AD
802{
803 uint32_t reg;
804
bec86378
ML
805 if (amdgpu_sriov_vf(adev))
806 return false;
807
808 if (amdgpu_passthrough(adev)) {
1da2c326
ML
809 /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
810 * some old smc fw still need driver do vPost otherwise gpu hang, while
811 * those smc fw version above 22.15 doesn't have this flaw, so we force
812 * vpost executed for smc version below 22.15
bec86378
ML
813 */
814 if (adev->asic_type == CHIP_FIJI) {
815 int err;
816 uint32_t fw_ver;
817 err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
818 /* force vPost if error occured */
819 if (err)
820 return true;
821
822 fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
1da2c326
ML
823 if (fw_ver < 0x00160e00)
824 return true;
bec86378 825 }
bec86378 826 }
91fe77eb 827
828 if (adev->has_hw_reset) {
829 adev->has_hw_reset = false;
830 return true;
831 }
832
833 /* bios scratch used on CIK+ */
834 if (adev->asic_type >= CHIP_BONAIRE)
835 return amdgpu_atombios_scratch_need_asic_init(adev);
836
837 /* check MEM_SIZE for older asics */
838 reg = amdgpu_asic_get_config_memsize(adev);
839
840 if ((reg != 0) && (reg != 0xffffffff))
841 return false;
842
843 return true;
bec86378
ML
844}
845
d38ceaf9
AD
846/* if we get transitioned to only one device, take VGA back */
847/**
06ec9070 848 * amdgpu_device_vga_set_decode - enable/disable vga decode
d38ceaf9
AD
849 *
850 * @cookie: amdgpu_device pointer
851 * @state: enable/disable vga decode
852 *
853 * Enable/disable vga decode (all asics).
854 * Returns VGA resource flags.
855 */
06ec9070 856static unsigned int amdgpu_device_vga_set_decode(void *cookie, bool state)
d38ceaf9
AD
857{
858 struct amdgpu_device *adev = cookie;
859 amdgpu_asic_set_vga_state(adev, state);
860 if (state)
861 return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
862 VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
863 else
864 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
865}
866
e3ecdffa
AD
867/**
868 * amdgpu_device_check_block_size - validate the vm block size
869 *
870 * @adev: amdgpu_device pointer
871 *
872 * Validates the vm block size specified via module parameter.
873 * The vm block size defines number of bits in page table versus page directory,
874 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
875 * page table and the remaining bits are in the page directory.
876 */
06ec9070 877static void amdgpu_device_check_block_size(struct amdgpu_device *adev)
a1adf8be
CZ
878{
879 /* defines number of bits in page table versus page directory,
880 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
881 * page table and the remaining bits are in the page directory */
bab4fee7
JZ
882 if (amdgpu_vm_block_size == -1)
883 return;
a1adf8be 884
bab4fee7 885 if (amdgpu_vm_block_size < 9) {
a1adf8be
CZ
886 dev_warn(adev->dev, "VM page table size (%d) too small\n",
887 amdgpu_vm_block_size);
97489129 888 amdgpu_vm_block_size = -1;
a1adf8be 889 }
a1adf8be
CZ
890}
891
e3ecdffa
AD
892/**
893 * amdgpu_device_check_vm_size - validate the vm size
894 *
895 * @adev: amdgpu_device pointer
896 *
897 * Validates the vm size in GB specified via module parameter.
898 * The VM size is the size of the GPU virtual memory space in GB.
899 */
06ec9070 900static void amdgpu_device_check_vm_size(struct amdgpu_device *adev)
83ca145d 901{
64dab074
AD
902 /* no need to check the default value */
903 if (amdgpu_vm_size == -1)
904 return;
905
83ca145d
ZJ
906 if (amdgpu_vm_size < 1) {
907 dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
908 amdgpu_vm_size);
f3368128 909 amdgpu_vm_size = -1;
83ca145d 910 }
83ca145d
ZJ
911}
912
7951e376
RZ
913static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev)
914{
915 struct sysinfo si;
916 bool is_os_64 = (sizeof(void *) == 8) ? true : false;
917 uint64_t total_memory;
918 uint64_t dram_size_seven_GB = 0x1B8000000;
919 uint64_t dram_size_three_GB = 0xB8000000;
920
921 if (amdgpu_smu_memory_pool_size == 0)
922 return;
923
924 if (!is_os_64) {
925 DRM_WARN("Not 64-bit OS, feature not supported\n");
926 goto def_value;
927 }
928 si_meminfo(&si);
929 total_memory = (uint64_t)si.totalram * si.mem_unit;
930
931 if ((amdgpu_smu_memory_pool_size == 1) ||
932 (amdgpu_smu_memory_pool_size == 2)) {
933 if (total_memory < dram_size_three_GB)
934 goto def_value1;
935 } else if ((amdgpu_smu_memory_pool_size == 4) ||
936 (amdgpu_smu_memory_pool_size == 8)) {
937 if (total_memory < dram_size_seven_GB)
938 goto def_value1;
939 } else {
940 DRM_WARN("Smu memory pool size not supported\n");
941 goto def_value;
942 }
943 adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28;
944
945 return;
946
947def_value1:
948 DRM_WARN("No enough system memory\n");
949def_value:
950 adev->pm.smu_prv_buffer_size = 0;
951}
952
d38ceaf9 953/**
06ec9070 954 * amdgpu_device_check_arguments - validate module params
d38ceaf9
AD
955 *
956 * @adev: amdgpu_device pointer
957 *
958 * Validates certain module parameters and updates
959 * the associated values used by the driver (all asics).
960 */
06ec9070 961static void amdgpu_device_check_arguments(struct amdgpu_device *adev)
d38ceaf9 962{
5b011235
CZ
963 if (amdgpu_sched_jobs < 4) {
964 dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
965 amdgpu_sched_jobs);
966 amdgpu_sched_jobs = 4;
76117507 967 } else if (!is_power_of_2(amdgpu_sched_jobs)){
5b011235
CZ
968 dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
969 amdgpu_sched_jobs);
970 amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
971 }
d38ceaf9 972
83e74db6 973 if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) {
f9321cc4
CK
974 /* gart size must be greater or equal to 32M */
975 dev_warn(adev->dev, "gart size (%d) too small\n",
976 amdgpu_gart_size);
83e74db6 977 amdgpu_gart_size = -1;
d38ceaf9
AD
978 }
979
36d38372 980 if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) {
c4e1a13a 981 /* gtt size must be greater or equal to 32M */
36d38372
CK
982 dev_warn(adev->dev, "gtt size (%d) too small\n",
983 amdgpu_gtt_size);
984 amdgpu_gtt_size = -1;
d38ceaf9
AD
985 }
986
d07f14be
RH
987 /* valid range is between 4 and 9 inclusive */
988 if (amdgpu_vm_fragment_size != -1 &&
989 (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) {
990 dev_warn(adev->dev, "valid range is between 4 and 9\n");
991 amdgpu_vm_fragment_size = -1;
992 }
993
7951e376
RZ
994 amdgpu_device_check_smu_prv_buffer_size(adev);
995
06ec9070 996 amdgpu_device_check_vm_size(adev);
d38ceaf9 997
06ec9070 998 amdgpu_device_check_block_size(adev);
6a7f76e7 999
526bae37 1000 if (amdgpu_vram_page_split != -1 && (amdgpu_vram_page_split < 16 ||
76117507 1001 !is_power_of_2(amdgpu_vram_page_split))) {
6a7f76e7
CK
1002 dev_warn(adev->dev, "invalid VRAM page split (%d)\n",
1003 amdgpu_vram_page_split);
1004 amdgpu_vram_page_split = 1024;
1005 }
8854695a
AG
1006
1007 if (amdgpu_lockup_timeout == 0) {
1008 dev_warn(adev->dev, "lockup_timeout msut be > 0, adjusting to 10000\n");
1009 amdgpu_lockup_timeout = 10000;
1010 }
19aede77
AD
1011
1012 adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type);
d38ceaf9
AD
1013}
1014
1015/**
1016 * amdgpu_switcheroo_set_state - set switcheroo state
1017 *
1018 * @pdev: pci dev pointer
1694467b 1019 * @state: vga_switcheroo state
d38ceaf9
AD
1020 *
1021 * Callback for the switcheroo driver. Suspends or resumes the
1022 * the asics before or after it is powered up using ACPI methods.
1023 */
1024static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, enum vga_switcheroo_state state)
1025{
1026 struct drm_device *dev = pci_get_drvdata(pdev);
1027
1028 if (amdgpu_device_is_px(dev) && state == VGA_SWITCHEROO_OFF)
1029 return;
1030
1031 if (state == VGA_SWITCHEROO_ON) {
7ca85295 1032 pr_info("amdgpu: switched on\n");
d38ceaf9
AD
1033 /* don't suspend or resume card normally */
1034 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1035
810ddc3a 1036 amdgpu_device_resume(dev, true, true);
d38ceaf9 1037
d38ceaf9
AD
1038 dev->switch_power_state = DRM_SWITCH_POWER_ON;
1039 drm_kms_helper_poll_enable(dev);
1040 } else {
7ca85295 1041 pr_info("amdgpu: switched off\n");
d38ceaf9
AD
1042 drm_kms_helper_poll_disable(dev);
1043 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
810ddc3a 1044 amdgpu_device_suspend(dev, true, true);
d38ceaf9
AD
1045 dev->switch_power_state = DRM_SWITCH_POWER_OFF;
1046 }
1047}
1048
1049/**
1050 * amdgpu_switcheroo_can_switch - see if switcheroo state can change
1051 *
1052 * @pdev: pci dev pointer
1053 *
1054 * Callback for the switcheroo driver. Check of the switcheroo
1055 * state can be changed.
1056 * Returns true if the state can be changed, false if not.
1057 */
1058static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
1059{
1060 struct drm_device *dev = pci_get_drvdata(pdev);
1061
1062 /*
1063 * FIXME: open_count is protected by drm_global_mutex but that would lead to
1064 * locking inversion with the driver load path. And the access here is
1065 * completely racy anyway. So don't bother with locking for now.
1066 */
1067 return dev->open_count == 0;
1068}
1069
1070static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
1071 .set_gpu_state = amdgpu_switcheroo_set_state,
1072 .reprobe = NULL,
1073 .can_switch = amdgpu_switcheroo_can_switch,
1074};
1075
e3ecdffa
AD
1076/**
1077 * amdgpu_device_ip_set_clockgating_state - set the CG state
1078 *
87e3f136 1079 * @dev: amdgpu_device pointer
e3ecdffa
AD
1080 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1081 * @state: clockgating state (gate or ungate)
1082 *
1083 * Sets the requested clockgating state for all instances of
1084 * the hardware IP specified.
1085 * Returns the error code from the last instance.
1086 */
43fa561f 1087int amdgpu_device_ip_set_clockgating_state(void *dev,
2990a1fc
AD
1088 enum amd_ip_block_type block_type,
1089 enum amd_clockgating_state state)
d38ceaf9 1090{
43fa561f 1091 struct amdgpu_device *adev = dev;
d38ceaf9
AD
1092 int i, r = 0;
1093
1094 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 1095 if (!adev->ip_blocks[i].status.valid)
9ecbe7f5 1096 continue;
c722865a
RZ
1097 if (adev->ip_blocks[i].version->type != block_type)
1098 continue;
1099 if (!adev->ip_blocks[i].version->funcs->set_clockgating_state)
1100 continue;
1101 r = adev->ip_blocks[i].version->funcs->set_clockgating_state(
1102 (void *)adev, state);
1103 if (r)
1104 DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",
1105 adev->ip_blocks[i].version->funcs->name, r);
d38ceaf9
AD
1106 }
1107 return r;
1108}
1109
e3ecdffa
AD
1110/**
1111 * amdgpu_device_ip_set_powergating_state - set the PG state
1112 *
87e3f136 1113 * @dev: amdgpu_device pointer
e3ecdffa
AD
1114 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1115 * @state: powergating state (gate or ungate)
1116 *
1117 * Sets the requested powergating state for all instances of
1118 * the hardware IP specified.
1119 * Returns the error code from the last instance.
1120 */
43fa561f 1121int amdgpu_device_ip_set_powergating_state(void *dev,
2990a1fc
AD
1122 enum amd_ip_block_type block_type,
1123 enum amd_powergating_state state)
d38ceaf9 1124{
43fa561f 1125 struct amdgpu_device *adev = dev;
d38ceaf9
AD
1126 int i, r = 0;
1127
1128 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 1129 if (!adev->ip_blocks[i].status.valid)
9ecbe7f5 1130 continue;
c722865a
RZ
1131 if (adev->ip_blocks[i].version->type != block_type)
1132 continue;
1133 if (!adev->ip_blocks[i].version->funcs->set_powergating_state)
1134 continue;
1135 r = adev->ip_blocks[i].version->funcs->set_powergating_state(
1136 (void *)adev, state);
1137 if (r)
1138 DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",
1139 adev->ip_blocks[i].version->funcs->name, r);
d38ceaf9
AD
1140 }
1141 return r;
1142}
1143
e3ecdffa
AD
1144/**
1145 * amdgpu_device_ip_get_clockgating_state - get the CG state
1146 *
1147 * @adev: amdgpu_device pointer
1148 * @flags: clockgating feature flags
1149 *
1150 * Walks the list of IPs on the device and updates the clockgating
1151 * flags for each IP.
1152 * Updates @flags with the feature flags for each hardware IP where
1153 * clockgating is enabled.
1154 */
2990a1fc
AD
1155void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev,
1156 u32 *flags)
6cb2d4e4
HR
1157{
1158 int i;
1159
1160 for (i = 0; i < adev->num_ip_blocks; i++) {
1161 if (!adev->ip_blocks[i].status.valid)
1162 continue;
1163 if (adev->ip_blocks[i].version->funcs->get_clockgating_state)
1164 adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags);
1165 }
1166}
1167
e3ecdffa
AD
1168/**
1169 * amdgpu_device_ip_wait_for_idle - wait for idle
1170 *
1171 * @adev: amdgpu_device pointer
1172 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1173 *
1174 * Waits for the request hardware IP to be idle.
1175 * Returns 0 for success or a negative error code on failure.
1176 */
2990a1fc
AD
1177int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev,
1178 enum amd_ip_block_type block_type)
5dbbb60b
AD
1179{
1180 int i, r;
1181
1182 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 1183 if (!adev->ip_blocks[i].status.valid)
9ecbe7f5 1184 continue;
a1255107
AD
1185 if (adev->ip_blocks[i].version->type == block_type) {
1186 r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
5dbbb60b
AD
1187 if (r)
1188 return r;
1189 break;
1190 }
1191 }
1192 return 0;
1193
1194}
1195
e3ecdffa
AD
1196/**
1197 * amdgpu_device_ip_is_idle - is the hardware IP idle
1198 *
1199 * @adev: amdgpu_device pointer
1200 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1201 *
1202 * Check if the hardware IP is idle or not.
1203 * Returns true if it the IP is idle, false if not.
1204 */
2990a1fc
AD
1205bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev,
1206 enum amd_ip_block_type block_type)
5dbbb60b
AD
1207{
1208 int i;
1209
1210 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 1211 if (!adev->ip_blocks[i].status.valid)
9ecbe7f5 1212 continue;
a1255107
AD
1213 if (adev->ip_blocks[i].version->type == block_type)
1214 return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
5dbbb60b
AD
1215 }
1216 return true;
1217
1218}
1219
e3ecdffa
AD
1220/**
1221 * amdgpu_device_ip_get_ip_block - get a hw IP pointer
1222 *
1223 * @adev: amdgpu_device pointer
87e3f136 1224 * @type: Type of hardware IP (SMU, GFX, UVD, etc.)
e3ecdffa
AD
1225 *
1226 * Returns a pointer to the hardware IP block structure
1227 * if it exists for the asic, otherwise NULL.
1228 */
2990a1fc
AD
1229struct amdgpu_ip_block *
1230amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,
1231 enum amd_ip_block_type type)
d38ceaf9
AD
1232{
1233 int i;
1234
1235 for (i = 0; i < adev->num_ip_blocks; i++)
a1255107 1236 if (adev->ip_blocks[i].version->type == type)
d38ceaf9
AD
1237 return &adev->ip_blocks[i];
1238
1239 return NULL;
1240}
1241
1242/**
2990a1fc 1243 * amdgpu_device_ip_block_version_cmp
d38ceaf9
AD
1244 *
1245 * @adev: amdgpu_device pointer
5fc3aeeb 1246 * @type: enum amd_ip_block_type
d38ceaf9
AD
1247 * @major: major version
1248 * @minor: minor version
1249 *
1250 * return 0 if equal or greater
1251 * return 1 if smaller or the ip_block doesn't exist
1252 */
2990a1fc
AD
1253int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev,
1254 enum amd_ip_block_type type,
1255 u32 major, u32 minor)
d38ceaf9 1256{
2990a1fc 1257 struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type);
d38ceaf9 1258
a1255107
AD
1259 if (ip_block && ((ip_block->version->major > major) ||
1260 ((ip_block->version->major == major) &&
1261 (ip_block->version->minor >= minor))))
d38ceaf9
AD
1262 return 0;
1263
1264 return 1;
1265}
1266
a1255107 1267/**
2990a1fc 1268 * amdgpu_device_ip_block_add
a1255107
AD
1269 *
1270 * @adev: amdgpu_device pointer
1271 * @ip_block_version: pointer to the IP to add
1272 *
1273 * Adds the IP block driver information to the collection of IPs
1274 * on the asic.
1275 */
2990a1fc
AD
1276int amdgpu_device_ip_block_add(struct amdgpu_device *adev,
1277 const struct amdgpu_ip_block_version *ip_block_version)
a1255107
AD
1278{
1279 if (!ip_block_version)
1280 return -EINVAL;
1281
e966a725 1282 DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks,
a0bae357
HR
1283 ip_block_version->funcs->name);
1284
a1255107
AD
1285 adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
1286
1287 return 0;
1288}
1289
e3ecdffa
AD
1290/**
1291 * amdgpu_device_enable_virtual_display - enable virtual display feature
1292 *
1293 * @adev: amdgpu_device pointer
1294 *
1295 * Enabled the virtual display feature if the user has enabled it via
1296 * the module parameter virtual_display. This feature provides a virtual
1297 * display hardware on headless boards or in virtualized environments.
1298 * This function parses and validates the configuration string specified by
1299 * the user and configues the virtual display configuration (number of
1300 * virtual connectors, crtcs, etc.) specified.
1301 */
483ef985 1302static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
9accf2fd
ED
1303{
1304 adev->enable_virtual_display = false;
1305
1306 if (amdgpu_virtual_display) {
1307 struct drm_device *ddev = adev->ddev;
1308 const char *pci_address_name = pci_name(ddev->pdev);
0f66356d 1309 char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
9accf2fd
ED
1310
1311 pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
1312 pciaddstr_tmp = pciaddstr;
0f66356d
ED
1313 while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
1314 pciaddname = strsep(&pciaddname_tmp, ",");
967de2a9
YT
1315 if (!strcmp("all", pciaddname)
1316 || !strcmp(pci_address_name, pciaddname)) {
0f66356d
ED
1317 long num_crtc;
1318 int res = -1;
1319
9accf2fd 1320 adev->enable_virtual_display = true;
0f66356d
ED
1321
1322 if (pciaddname_tmp)
1323 res = kstrtol(pciaddname_tmp, 10,
1324 &num_crtc);
1325
1326 if (!res) {
1327 if (num_crtc < 1)
1328 num_crtc = 1;
1329 if (num_crtc > 6)
1330 num_crtc = 6;
1331 adev->mode_info.num_crtc = num_crtc;
1332 } else {
1333 adev->mode_info.num_crtc = 1;
1334 }
9accf2fd
ED
1335 break;
1336 }
1337 }
1338
0f66356d
ED
1339 DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
1340 amdgpu_virtual_display, pci_address_name,
1341 adev->enable_virtual_display, adev->mode_info.num_crtc);
9accf2fd
ED
1342
1343 kfree(pciaddstr);
1344 }
1345}
1346
e3ecdffa
AD
1347/**
1348 * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware
1349 *
1350 * @adev: amdgpu_device pointer
1351 *
1352 * Parses the asic configuration parameters specified in the gpu info
1353 * firmware and makes them availale to the driver for use in configuring
1354 * the asic.
1355 * Returns 0 on success, -EINVAL on failure.
1356 */
e2a75f88
AD
1357static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev)
1358{
e2a75f88
AD
1359 const char *chip_name;
1360 char fw_name[30];
1361 int err;
1362 const struct gpu_info_firmware_header_v1_0 *hdr;
1363
ab4fe3e1
HR
1364 adev->firmware.gpu_info_fw = NULL;
1365
e2a75f88
AD
1366 switch (adev->asic_type) {
1367 case CHIP_TOPAZ:
1368 case CHIP_TONGA:
1369 case CHIP_FIJI:
e2a75f88 1370 case CHIP_POLARIS10:
cc07f18d 1371 case CHIP_POLARIS11:
e2a75f88 1372 case CHIP_POLARIS12:
cc07f18d 1373 case CHIP_VEGAM:
e2a75f88
AD
1374 case CHIP_CARRIZO:
1375 case CHIP_STONEY:
1376#ifdef CONFIG_DRM_AMDGPU_SI
1377 case CHIP_VERDE:
1378 case CHIP_TAHITI:
1379 case CHIP_PITCAIRN:
1380 case CHIP_OLAND:
1381 case CHIP_HAINAN:
1382#endif
1383#ifdef CONFIG_DRM_AMDGPU_CIK
1384 case CHIP_BONAIRE:
1385 case CHIP_HAWAII:
1386 case CHIP_KAVERI:
1387 case CHIP_KABINI:
1388 case CHIP_MULLINS:
1389#endif
27c0bc71 1390 case CHIP_VEGA20:
e2a75f88
AD
1391 default:
1392 return 0;
1393 case CHIP_VEGA10:
1394 chip_name = "vega10";
1395 break;
3f76dced
AD
1396 case CHIP_VEGA12:
1397 chip_name = "vega12";
1398 break;
2d2e5e7e
AD
1399 case CHIP_RAVEN:
1400 chip_name = "raven";
1401 break;
e2a75f88
AD
1402 }
1403
1404 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name);
ab4fe3e1 1405 err = request_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev);
e2a75f88
AD
1406 if (err) {
1407 dev_err(adev->dev,
1408 "Failed to load gpu_info firmware \"%s\"\n",
1409 fw_name);
1410 goto out;
1411 }
ab4fe3e1 1412 err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw);
e2a75f88
AD
1413 if (err) {
1414 dev_err(adev->dev,
1415 "Failed to validate gpu_info firmware \"%s\"\n",
1416 fw_name);
1417 goto out;
1418 }
1419
ab4fe3e1 1420 hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data;
e2a75f88
AD
1421 amdgpu_ucode_print_gpu_info_hdr(&hdr->header);
1422
1423 switch (hdr->version_major) {
1424 case 1:
1425 {
1426 const struct gpu_info_firmware_v1_0 *gpu_info_fw =
ab4fe3e1 1427 (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data +
e2a75f88
AD
1428 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1429
b5ab16bf
AD
1430 adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se);
1431 adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh);
1432 adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se);
1433 adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se);
e2a75f88 1434 adev->gfx.config.max_texture_channel_caches =
b5ab16bf
AD
1435 le32_to_cpu(gpu_info_fw->gc_num_tccs);
1436 adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs);
1437 adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds);
1438 adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth);
1439 adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth);
e2a75f88 1440 adev->gfx.config.double_offchip_lds_buf =
b5ab16bf
AD
1441 le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer);
1442 adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size);
51fd0370
HZ
1443 adev->gfx.cu_info.max_waves_per_simd =
1444 le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd);
1445 adev->gfx.cu_info.max_scratch_slots_per_cu =
1446 le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu);
1447 adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size);
e2a75f88
AD
1448 break;
1449 }
1450 default:
1451 dev_err(adev->dev,
1452 "Unsupported gpu_info table %d\n", hdr->header.ucode_version);
1453 err = -EINVAL;
1454 goto out;
1455 }
1456out:
e2a75f88
AD
1457 return err;
1458}
1459
e3ecdffa
AD
1460/**
1461 * amdgpu_device_ip_early_init - run early init for hardware IPs
1462 *
1463 * @adev: amdgpu_device pointer
1464 *
1465 * Early initialization pass for hardware IPs. The hardware IPs that make
1466 * up each asic are discovered each IP's early_init callback is run. This
1467 * is the first stage in initializing the asic.
1468 * Returns 0 on success, negative error code on failure.
1469 */
06ec9070 1470static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
d38ceaf9 1471{
aaa36a97 1472 int i, r;
d38ceaf9 1473
483ef985 1474 amdgpu_device_enable_virtual_display(adev);
a6be7570 1475
d38ceaf9 1476 switch (adev->asic_type) {
aaa36a97
AD
1477 case CHIP_TOPAZ:
1478 case CHIP_TONGA:
48299f95 1479 case CHIP_FIJI:
2cc0c0b5 1480 case CHIP_POLARIS10:
32cc7e53 1481 case CHIP_POLARIS11:
c4642a47 1482 case CHIP_POLARIS12:
32cc7e53 1483 case CHIP_VEGAM:
aaa36a97 1484 case CHIP_CARRIZO:
39bb0c92
SL
1485 case CHIP_STONEY:
1486 if (adev->asic_type == CHIP_CARRIZO || adev->asic_type == CHIP_STONEY)
aaa36a97
AD
1487 adev->family = AMDGPU_FAMILY_CZ;
1488 else
1489 adev->family = AMDGPU_FAMILY_VI;
1490
1491 r = vi_set_ip_blocks(adev);
1492 if (r)
1493 return r;
1494 break;
33f34802
KW
1495#ifdef CONFIG_DRM_AMDGPU_SI
1496 case CHIP_VERDE:
1497 case CHIP_TAHITI:
1498 case CHIP_PITCAIRN:
1499 case CHIP_OLAND:
1500 case CHIP_HAINAN:
295d0daf 1501 adev->family = AMDGPU_FAMILY_SI;
33f34802
KW
1502 r = si_set_ip_blocks(adev);
1503 if (r)
1504 return r;
1505 break;
1506#endif
a2e73f56
AD
1507#ifdef CONFIG_DRM_AMDGPU_CIK
1508 case CHIP_BONAIRE:
1509 case CHIP_HAWAII:
1510 case CHIP_KAVERI:
1511 case CHIP_KABINI:
1512 case CHIP_MULLINS:
1513 if ((adev->asic_type == CHIP_BONAIRE) || (adev->asic_type == CHIP_HAWAII))
1514 adev->family = AMDGPU_FAMILY_CI;
1515 else
1516 adev->family = AMDGPU_FAMILY_KV;
1517
1518 r = cik_set_ip_blocks(adev);
1519 if (r)
1520 return r;
1521 break;
1522#endif
e48a3cd9
AD
1523 case CHIP_VEGA10:
1524 case CHIP_VEGA12:
e4bd8170 1525 case CHIP_VEGA20:
e48a3cd9 1526 case CHIP_RAVEN:
2ca8a5d2
CZ
1527 if (adev->asic_type == CHIP_RAVEN)
1528 adev->family = AMDGPU_FAMILY_RV;
1529 else
1530 adev->family = AMDGPU_FAMILY_AI;
460826e6
KW
1531
1532 r = soc15_set_ip_blocks(adev);
1533 if (r)
1534 return r;
1535 break;
d38ceaf9
AD
1536 default:
1537 /* FIXME: not supported yet */
1538 return -EINVAL;
1539 }
1540
e2a75f88
AD
1541 r = amdgpu_device_parse_gpu_info_fw(adev);
1542 if (r)
1543 return r;
1544
1884734a 1545 amdgpu_amdkfd_device_probe(adev);
1546
3149d9da
XY
1547 if (amdgpu_sriov_vf(adev)) {
1548 r = amdgpu_virt_request_full_gpu(adev, true);
1549 if (r)
5ffa61c1 1550 return -EAGAIN;
3149d9da
XY
1551 }
1552
00f54b97
HR
1553 adev->powerplay.pp_feature = amdgpu_pp_feature_mask;
1554
d38ceaf9
AD
1555 for (i = 0; i < adev->num_ip_blocks; i++) {
1556 if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
ed8cf00c
HR
1557 DRM_ERROR("disabled ip block: %d <%s>\n",
1558 i, adev->ip_blocks[i].version->funcs->name);
a1255107 1559 adev->ip_blocks[i].status.valid = false;
d38ceaf9 1560 } else {
a1255107
AD
1561 if (adev->ip_blocks[i].version->funcs->early_init) {
1562 r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
2c1a2784 1563 if (r == -ENOENT) {
a1255107 1564 adev->ip_blocks[i].status.valid = false;
2c1a2784 1565 } else if (r) {
a1255107
AD
1566 DRM_ERROR("early_init of IP block <%s> failed %d\n",
1567 adev->ip_blocks[i].version->funcs->name, r);
d38ceaf9 1568 return r;
2c1a2784 1569 } else {
a1255107 1570 adev->ip_blocks[i].status.valid = true;
2c1a2784 1571 }
974e6b64 1572 } else {
a1255107 1573 adev->ip_blocks[i].status.valid = true;
d38ceaf9 1574 }
d38ceaf9
AD
1575 }
1576 }
1577
395d1fb9
NH
1578 adev->cg_flags &= amdgpu_cg_mask;
1579 adev->pg_flags &= amdgpu_pg_mask;
1580
d38ceaf9
AD
1581 return 0;
1582}
1583
e3ecdffa
AD
1584/**
1585 * amdgpu_device_ip_init - run init for hardware IPs
1586 *
1587 * @adev: amdgpu_device pointer
1588 *
1589 * Main initialization pass for hardware IPs. The list of all the hardware
1590 * IPs that make up the asic is walked and the sw_init and hw_init callbacks
1591 * are run. sw_init initializes the software state associated with each IP
1592 * and hw_init initializes the hardware associated with each IP.
1593 * Returns 0 on success, negative error code on failure.
1594 */
06ec9070 1595static int amdgpu_device_ip_init(struct amdgpu_device *adev)
d38ceaf9
AD
1596{
1597 int i, r;
1598
1599 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 1600 if (!adev->ip_blocks[i].status.valid)
d38ceaf9 1601 continue;
a1255107 1602 r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
2c1a2784 1603 if (r) {
a1255107
AD
1604 DRM_ERROR("sw_init of IP block <%s> failed %d\n",
1605 adev->ip_blocks[i].version->funcs->name, r);
d38ceaf9 1606 return r;
2c1a2784 1607 }
a1255107 1608 adev->ip_blocks[i].status.sw = true;
bfca0289 1609
d38ceaf9 1610 /* need to do gmc hw init early so we can allocate gpu mem */
a1255107 1611 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
06ec9070 1612 r = amdgpu_device_vram_scratch_init(adev);
2c1a2784
AD
1613 if (r) {
1614 DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r);
d38ceaf9 1615 return r;
2c1a2784 1616 }
a1255107 1617 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
2c1a2784
AD
1618 if (r) {
1619 DRM_ERROR("hw_init %d failed %d\n", i, r);
d38ceaf9 1620 return r;
2c1a2784 1621 }
06ec9070 1622 r = amdgpu_device_wb_init(adev);
2c1a2784 1623 if (r) {
06ec9070 1624 DRM_ERROR("amdgpu_device_wb_init failed %d\n", r);
d38ceaf9 1625 return r;
2c1a2784 1626 }
a1255107 1627 adev->ip_blocks[i].status.hw = true;
2493664f
ML
1628
1629 /* right after GMC hw init, we create CSA */
1630 if (amdgpu_sriov_vf(adev)) {
1631 r = amdgpu_allocate_static_csa(adev);
1632 if (r) {
1633 DRM_ERROR("allocate CSA failed %d\n", r);
1634 return r;
1635 }
1636 }
d38ceaf9
AD
1637 }
1638 }
1639
1640 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 1641 if (!adev->ip_blocks[i].status.sw)
d38ceaf9 1642 continue;
bfca0289 1643 if (adev->ip_blocks[i].status.hw)
d38ceaf9 1644 continue;
a1255107 1645 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
2c1a2784 1646 if (r) {
a1255107
AD
1647 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1648 adev->ip_blocks[i].version->funcs->name, r);
d38ceaf9 1649 return r;
2c1a2784 1650 }
a1255107 1651 adev->ip_blocks[i].status.hw = true;
d38ceaf9
AD
1652 }
1653
1884734a 1654 amdgpu_amdkfd_device_init(adev);
c6332b97 1655
1656 if (amdgpu_sriov_vf(adev))
1657 amdgpu_virt_release_full_gpu(adev, true);
1658
d38ceaf9
AD
1659 return 0;
1660}
1661
e3ecdffa
AD
1662/**
1663 * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer
1664 *
1665 * @adev: amdgpu_device pointer
1666 *
1667 * Writes a reset magic value to the gart pointer in VRAM. The driver calls
1668 * this function before a GPU reset. If the value is retained after a
1669 * GPU reset, VRAM has not been lost. Some GPU resets may destry VRAM contents.
1670 */
06ec9070 1671static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev)
0c49e0b8
CZ
1672{
1673 memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
1674}
1675
e3ecdffa
AD
1676/**
1677 * amdgpu_device_check_vram_lost - check if vram is valid
1678 *
1679 * @adev: amdgpu_device pointer
1680 *
1681 * Checks the reset magic value written to the gart pointer in VRAM.
1682 * The driver calls this after a GPU reset to see if the contents of
1683 * VRAM is lost or now.
1684 * returns true if vram is lost, false if not.
1685 */
06ec9070 1686static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev)
0c49e0b8
CZ
1687{
1688 return !!memcmp(adev->gart.ptr, adev->reset_magic,
1689 AMDGPU_RESET_MAGIC_NUM);
1690}
1691
e3ecdffa
AD
1692/**
1693 * amdgpu_device_ip_late_set_cg_state - late init for clockgating
1694 *
1695 * @adev: amdgpu_device pointer
1696 *
1697 * Late initialization pass enabling clockgating for hardware IPs.
1698 * The list of all the hardware IPs that make up the asic is walked and the
1699 * set_clockgating_state callbacks are run. This stage is run late
1700 * in the init process.
1701 * Returns 0 on success, negative error code on failure.
1702 */
06ec9070 1703static int amdgpu_device_ip_late_set_cg_state(struct amdgpu_device *adev)
d38ceaf9
AD
1704{
1705 int i = 0, r;
1706
4a2ba394
SL
1707 if (amdgpu_emu_mode == 1)
1708 return 0;
1709
d38ceaf9 1710 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 1711 if (!adev->ip_blocks[i].status.valid)
d38ceaf9 1712 continue;
4a446d55 1713 /* skip CG for VCE/UVD, it's handled specially */
a1255107 1714 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
57716327 1715 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
34319b32 1716 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
57716327 1717 adev->ip_blocks[i].version->funcs->set_clockgating_state) {
4a446d55 1718 /* enable clockgating to save power */
a1255107
AD
1719 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1720 AMD_CG_STATE_GATE);
4a446d55
AD
1721 if (r) {
1722 DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
a1255107 1723 adev->ip_blocks[i].version->funcs->name, r);
4a446d55
AD
1724 return r;
1725 }
b0b00ff1 1726 }
d38ceaf9 1727 }
06b18f61 1728
c9f96fd5
RZ
1729 return 0;
1730}
1731
1732static int amdgpu_device_ip_late_set_pg_state(struct amdgpu_device *adev)
1733{
1734 int i = 0, r;
06b18f61 1735
c9f96fd5
RZ
1736 if (amdgpu_emu_mode == 1)
1737 return 0;
1738
1739 for (i = 0; i < adev->num_ip_blocks; i++) {
1740 if (!adev->ip_blocks[i].status.valid)
1741 continue;
1742 /* skip CG for VCE/UVD, it's handled specially */
1743 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1744 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1745 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
1746 adev->ip_blocks[i].version->funcs->set_powergating_state) {
1747 /* enable powergating to save power */
1748 r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev,
1749 AMD_PG_STATE_GATE);
1750 if (r) {
1751 DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n",
1752 adev->ip_blocks[i].version->funcs->name, r);
1753 return r;
1754 }
1755 }
1756 }
2dc80b00
S
1757 return 0;
1758}
1759
e3ecdffa
AD
1760/**
1761 * amdgpu_device_ip_late_init - run late init for hardware IPs
1762 *
1763 * @adev: amdgpu_device pointer
1764 *
1765 * Late initialization pass for hardware IPs. The list of all the hardware
1766 * IPs that make up the asic is walked and the late_init callbacks are run.
1767 * late_init covers any special initialization that an IP requires
1768 * after all of the have been initialized or something that needs to happen
1769 * late in the init process.
1770 * Returns 0 on success, negative error code on failure.
1771 */
06ec9070 1772static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
2dc80b00
S
1773{
1774 int i = 0, r;
1775
1776 for (i = 0; i < adev->num_ip_blocks; i++) {
1777 if (!adev->ip_blocks[i].status.valid)
1778 continue;
1779 if (adev->ip_blocks[i].version->funcs->late_init) {
1780 r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
1781 if (r) {
1782 DRM_ERROR("late_init of IP block <%s> failed %d\n",
1783 adev->ip_blocks[i].version->funcs->name, r);
1784 return r;
1785 }
1786 adev->ip_blocks[i].status.late_initialized = true;
1787 }
1788 }
1789
916ac57f
RZ
1790 amdgpu_device_ip_late_set_cg_state(adev);
1791 amdgpu_device_ip_late_set_pg_state(adev);
1792
2c773de2
S
1793 queue_delayed_work(system_wq, &adev->late_init_work,
1794 msecs_to_jiffies(AMDGPU_RESUME_MS));
d38ceaf9 1795
06ec9070 1796 amdgpu_device_fill_reset_magic(adev);
d38ceaf9
AD
1797
1798 return 0;
1799}
1800
e3ecdffa
AD
1801/**
1802 * amdgpu_device_ip_fini - run fini for hardware IPs
1803 *
1804 * @adev: amdgpu_device pointer
1805 *
1806 * Main teardown pass for hardware IPs. The list of all the hardware
1807 * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks
1808 * are run. hw_fini tears down the hardware associated with each IP
1809 * and sw_fini tears down any software state associated with each IP.
1810 * Returns 0 on success, negative error code on failure.
1811 */
06ec9070 1812static int amdgpu_device_ip_fini(struct amdgpu_device *adev)
d38ceaf9
AD
1813{
1814 int i, r;
1815
1884734a 1816 amdgpu_amdkfd_device_fini(adev);
3e96dbfd
AD
1817 /* need to disable SMC first */
1818 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 1819 if (!adev->ip_blocks[i].status.hw)
3e96dbfd 1820 continue;
57716327
RZ
1821 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC &&
1822 adev->ip_blocks[i].version->funcs->set_clockgating_state) {
3e96dbfd 1823 /* ungate blocks before hw fini so that we can shutdown the blocks safely */
a1255107
AD
1824 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1825 AMD_CG_STATE_UNGATE);
3e96dbfd
AD
1826 if (r) {
1827 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
a1255107 1828 adev->ip_blocks[i].version->funcs->name, r);
3e96dbfd
AD
1829 return r;
1830 }
9134c6d7
RZ
1831 if (adev->powerplay.pp_funcs->set_powergating_by_smu)
1832 amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false);
a1255107 1833 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
3e96dbfd
AD
1834 /* XXX handle errors */
1835 if (r) {
1836 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
a1255107 1837 adev->ip_blocks[i].version->funcs->name, r);
3e96dbfd 1838 }
a1255107 1839 adev->ip_blocks[i].status.hw = false;
3e96dbfd
AD
1840 break;
1841 }
1842 }
1843
d38ceaf9 1844 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
a1255107 1845 if (!adev->ip_blocks[i].status.hw)
d38ceaf9 1846 continue;
8201a67a
RZ
1847
1848 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
81ce8bea 1849 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
34319b32 1850 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
81ce8bea 1851 adev->ip_blocks[i].version->funcs->set_clockgating_state) {
8201a67a
RZ
1852 /* ungate blocks before hw fini so that we can shutdown the blocks safely */
1853 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1854 AMD_CG_STATE_UNGATE);
1855 if (r) {
1856 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1857 adev->ip_blocks[i].version->funcs->name, r);
1858 return r;
1859 }
2c1a2784 1860 }
8201a67a 1861
a1255107 1862 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
d38ceaf9 1863 /* XXX handle errors */
2c1a2784 1864 if (r) {
a1255107
AD
1865 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
1866 adev->ip_blocks[i].version->funcs->name, r);
2c1a2784 1867 }
8201a67a 1868
a1255107 1869 adev->ip_blocks[i].status.hw = false;
d38ceaf9
AD
1870 }
1871
9950cda2 1872
d38ceaf9 1873 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
a1255107 1874 if (!adev->ip_blocks[i].status.sw)
d38ceaf9 1875 continue;
c12aba3a
ML
1876
1877 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1878 amdgpu_free_static_csa(adev);
1879 amdgpu_device_wb_fini(adev);
1880 amdgpu_device_vram_scratch_fini(adev);
1881 }
1882
a1255107 1883 r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
d38ceaf9 1884 /* XXX handle errors */
2c1a2784 1885 if (r) {
a1255107
AD
1886 DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
1887 adev->ip_blocks[i].version->funcs->name, r);
2c1a2784 1888 }
a1255107
AD
1889 adev->ip_blocks[i].status.sw = false;
1890 adev->ip_blocks[i].status.valid = false;
d38ceaf9
AD
1891 }
1892
a6dcfd9c 1893 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
a1255107 1894 if (!adev->ip_blocks[i].status.late_initialized)
8a2eef1d 1895 continue;
a1255107
AD
1896 if (adev->ip_blocks[i].version->funcs->late_fini)
1897 adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
1898 adev->ip_blocks[i].status.late_initialized = false;
a6dcfd9c
ML
1899 }
1900
030308fc 1901 if (amdgpu_sriov_vf(adev))
24136135
ML
1902 if (amdgpu_virt_release_full_gpu(adev, false))
1903 DRM_ERROR("failed to release exclusive mode on fini\n");
2493664f 1904
d38ceaf9
AD
1905 return 0;
1906}
1907
e3ecdffa
AD
1908/**
1909 * amdgpu_device_ip_late_init_func_handler - work handler for clockgating
1910 *
1911 * @work: work_struct
1912 *
1913 * Work handler for amdgpu_device_ip_late_set_cg_state. We put the
1914 * clockgating setup into a worker thread to speed up driver init and
1915 * resume from suspend.
1916 */
06ec9070 1917static void amdgpu_device_ip_late_init_func_handler(struct work_struct *work)
2dc80b00
S
1918{
1919 struct amdgpu_device *adev =
1920 container_of(work, struct amdgpu_device, late_init_work.work);
916ac57f
RZ
1921 int r;
1922
1923 r = amdgpu_ib_ring_tests(adev);
1924 if (r)
1925 DRM_ERROR("ib ring test failed (%d).\n", r);
2dc80b00
S
1926}
1927
e3ecdffa 1928/**
e7854a03 1929 * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1)
e3ecdffa
AD
1930 *
1931 * @adev: amdgpu_device pointer
1932 *
1933 * Main suspend function for hardware IPs. The list of all the hardware
1934 * IPs that make up the asic is walked, clockgating is disabled and the
1935 * suspend callbacks are run. suspend puts the hardware and software state
1936 * in each IP into a state suitable for suspend.
1937 * Returns 0 on success, negative error code on failure.
1938 */
e7854a03
AD
1939static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev)
1940{
1941 int i, r;
1942
1943 if (amdgpu_sriov_vf(adev))
1944 amdgpu_virt_request_full_gpu(adev, false);
1945
1946 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1947 if (!adev->ip_blocks[i].status.valid)
1948 continue;
1949 /* displays are handled separately */
1950 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) {
1951 /* ungate blocks so that suspend can properly shut them down */
1952 if (adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1953 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1954 AMD_CG_STATE_UNGATE);
1955 if (r) {
1956 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1957 adev->ip_blocks[i].version->funcs->name, r);
1958 }
1959 }
1960 /* XXX handle errors */
1961 r = adev->ip_blocks[i].version->funcs->suspend(adev);
1962 /* XXX handle errors */
1963 if (r) {
1964 DRM_ERROR("suspend of IP block <%s> failed %d\n",
1965 adev->ip_blocks[i].version->funcs->name, r);
1966 }
1967 }
1968 }
1969
1970 if (amdgpu_sriov_vf(adev))
1971 amdgpu_virt_release_full_gpu(adev, false);
1972
1973 return 0;
1974}
1975
1976/**
1977 * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2)
1978 *
1979 * @adev: amdgpu_device pointer
1980 *
1981 * Main suspend function for hardware IPs. The list of all the hardware
1982 * IPs that make up the asic is walked, clockgating is disabled and the
1983 * suspend callbacks are run. suspend puts the hardware and software state
1984 * in each IP into a state suitable for suspend.
1985 * Returns 0 on success, negative error code on failure.
1986 */
1987static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev)
d38ceaf9
AD
1988{
1989 int i, r;
1990
e941ea99
XY
1991 if (amdgpu_sriov_vf(adev))
1992 amdgpu_virt_request_full_gpu(adev, false);
1993
c5a93a28 1994 /* ungate SMC block first */
2990a1fc
AD
1995 r = amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_SMC,
1996 AMD_CG_STATE_UNGATE);
c5a93a28 1997 if (r) {
2990a1fc 1998 DRM_ERROR("set_clockgating_state(ungate) SMC failed %d\n", r);
c5a93a28
FC
1999 }
2000
9134c6d7
RZ
2001 /* call smu to disable gfx off feature first when suspend */
2002 if (adev->powerplay.pp_funcs->set_powergating_by_smu)
2003 amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false);
2004
d38ceaf9 2005 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
a1255107 2006 if (!adev->ip_blocks[i].status.valid)
d38ceaf9 2007 continue;
e7854a03
AD
2008 /* displays are handled in phase1 */
2009 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE)
2010 continue;
d38ceaf9 2011 /* ungate blocks so that suspend can properly shut them down */
5b2a3d2c 2012 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_SMC &&
57716327 2013 adev->ip_blocks[i].version->funcs->set_clockgating_state) {
a1255107
AD
2014 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
2015 AMD_CG_STATE_UNGATE);
c5a93a28 2016 if (r) {
a1255107
AD
2017 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
2018 adev->ip_blocks[i].version->funcs->name, r);
c5a93a28 2019 }
2c1a2784 2020 }
d38ceaf9 2021 /* XXX handle errors */
a1255107 2022 r = adev->ip_blocks[i].version->funcs->suspend(adev);
d38ceaf9 2023 /* XXX handle errors */
2c1a2784 2024 if (r) {
a1255107
AD
2025 DRM_ERROR("suspend of IP block <%s> failed %d\n",
2026 adev->ip_blocks[i].version->funcs->name, r);
2c1a2784 2027 }
d38ceaf9
AD
2028 }
2029
e941ea99
XY
2030 if (amdgpu_sriov_vf(adev))
2031 amdgpu_virt_release_full_gpu(adev, false);
2032
d38ceaf9
AD
2033 return 0;
2034}
2035
e7854a03
AD
2036/**
2037 * amdgpu_device_ip_suspend - run suspend for hardware IPs
2038 *
2039 * @adev: amdgpu_device pointer
2040 *
2041 * Main suspend function for hardware IPs. The list of all the hardware
2042 * IPs that make up the asic is walked, clockgating is disabled and the
2043 * suspend callbacks are run. suspend puts the hardware and software state
2044 * in each IP into a state suitable for suspend.
2045 * Returns 0 on success, negative error code on failure.
2046 */
2047int amdgpu_device_ip_suspend(struct amdgpu_device *adev)
2048{
2049 int r;
2050
2051 r = amdgpu_device_ip_suspend_phase1(adev);
2052 if (r)
2053 return r;
2054 r = amdgpu_device_ip_suspend_phase2(adev);
2055
2056 return r;
2057}
2058
06ec9070 2059static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
a90ad3c2
ML
2060{
2061 int i, r;
2062
2cb681b6
ML
2063 static enum amd_ip_block_type ip_order[] = {
2064 AMD_IP_BLOCK_TYPE_GMC,
2065 AMD_IP_BLOCK_TYPE_COMMON,
2cb681b6
ML
2066 AMD_IP_BLOCK_TYPE_IH,
2067 };
a90ad3c2 2068
2cb681b6
ML
2069 for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
2070 int j;
2071 struct amdgpu_ip_block *block;
a90ad3c2 2072
2cb681b6
ML
2073 for (j = 0; j < adev->num_ip_blocks; j++) {
2074 block = &adev->ip_blocks[j];
2075
2076 if (block->version->type != ip_order[i] ||
2077 !block->status.valid)
2078 continue;
2079
2080 r = block->version->funcs->hw_init(adev);
3f48c681 2081 DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
c41d1cf6
ML
2082 if (r)
2083 return r;
a90ad3c2
ML
2084 }
2085 }
2086
2087 return 0;
2088}
2089
06ec9070 2090static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
a90ad3c2
ML
2091{
2092 int i, r;
2093
2cb681b6
ML
2094 static enum amd_ip_block_type ip_order[] = {
2095 AMD_IP_BLOCK_TYPE_SMC,
ef4c166d 2096 AMD_IP_BLOCK_TYPE_PSP,
2cb681b6
ML
2097 AMD_IP_BLOCK_TYPE_DCE,
2098 AMD_IP_BLOCK_TYPE_GFX,
2099 AMD_IP_BLOCK_TYPE_SDMA,
257deb8c
FM
2100 AMD_IP_BLOCK_TYPE_UVD,
2101 AMD_IP_BLOCK_TYPE_VCE
2cb681b6 2102 };
a90ad3c2 2103
2cb681b6
ML
2104 for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
2105 int j;
2106 struct amdgpu_ip_block *block;
a90ad3c2 2107
2cb681b6
ML
2108 for (j = 0; j < adev->num_ip_blocks; j++) {
2109 block = &adev->ip_blocks[j];
2110
2111 if (block->version->type != ip_order[i] ||
2112 !block->status.valid)
2113 continue;
2114
2115 r = block->version->funcs->hw_init(adev);
3f48c681 2116 DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
c41d1cf6
ML
2117 if (r)
2118 return r;
a90ad3c2
ML
2119 }
2120 }
2121
2122 return 0;
2123}
2124
e3ecdffa
AD
2125/**
2126 * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs
2127 *
2128 * @adev: amdgpu_device pointer
2129 *
2130 * First resume function for hardware IPs. The list of all the hardware
2131 * IPs that make up the asic is walked and the resume callbacks are run for
2132 * COMMON, GMC, and IH. resume puts the hardware into a functional state
2133 * after a suspend and updates the software state as necessary. This
2134 * function is also used for restoring the GPU after a GPU reset.
2135 * Returns 0 on success, negative error code on failure.
2136 */
06ec9070 2137static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)
d38ceaf9
AD
2138{
2139 int i, r;
2140
a90ad3c2
ML
2141 for (i = 0; i < adev->num_ip_blocks; i++) {
2142 if (!adev->ip_blocks[i].status.valid)
2143 continue;
a90ad3c2 2144 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
e3ecdffa
AD
2145 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2146 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
fcf0649f
CZ
2147 r = adev->ip_blocks[i].version->funcs->resume(adev);
2148 if (r) {
2149 DRM_ERROR("resume of IP block <%s> failed %d\n",
2150 adev->ip_blocks[i].version->funcs->name, r);
2151 return r;
2152 }
a90ad3c2
ML
2153 }
2154 }
2155
2156 return 0;
2157}
2158
e3ecdffa
AD
2159/**
2160 * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs
2161 *
2162 * @adev: amdgpu_device pointer
2163 *
2164 * First resume function for hardware IPs. The list of all the hardware
2165 * IPs that make up the asic is walked and the resume callbacks are run for
2166 * all blocks except COMMON, GMC, and IH. resume puts the hardware into a
2167 * functional state after a suspend and updates the software state as
2168 * necessary. This function is also used for restoring the GPU after a GPU
2169 * reset.
2170 * Returns 0 on success, negative error code on failure.
2171 */
06ec9070 2172static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)
d38ceaf9
AD
2173{
2174 int i, r;
2175
2176 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 2177 if (!adev->ip_blocks[i].status.valid)
d38ceaf9 2178 continue;
fcf0649f 2179 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
e3ecdffa
AD
2180 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2181 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH)
fcf0649f 2182 continue;
a1255107 2183 r = adev->ip_blocks[i].version->funcs->resume(adev);
2c1a2784 2184 if (r) {
a1255107
AD
2185 DRM_ERROR("resume of IP block <%s> failed %d\n",
2186 adev->ip_blocks[i].version->funcs->name, r);
d38ceaf9 2187 return r;
2c1a2784 2188 }
d38ceaf9
AD
2189 }
2190
2191 return 0;
2192}
2193
e3ecdffa
AD
2194/**
2195 * amdgpu_device_ip_resume - run resume for hardware IPs
2196 *
2197 * @adev: amdgpu_device pointer
2198 *
2199 * Main resume function for hardware IPs. The hardware IPs
2200 * are split into two resume functions because they are
2201 * are also used in in recovering from a GPU reset and some additional
2202 * steps need to be take between them. In this case (S3/S4) they are
2203 * run sequentially.
2204 * Returns 0 on success, negative error code on failure.
2205 */
06ec9070 2206static int amdgpu_device_ip_resume(struct amdgpu_device *adev)
fcf0649f
CZ
2207{
2208 int r;
2209
06ec9070 2210 r = amdgpu_device_ip_resume_phase1(adev);
fcf0649f
CZ
2211 if (r)
2212 return r;
06ec9070 2213 r = amdgpu_device_ip_resume_phase2(adev);
fcf0649f
CZ
2214
2215 return r;
2216}
2217
e3ecdffa
AD
2218/**
2219 * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV
2220 *
2221 * @adev: amdgpu_device pointer
2222 *
2223 * Query the VBIOS data tables to determine if the board supports SR-IOV.
2224 */
4e99a44e 2225static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
048765ad 2226{
6867e1b5
ML
2227 if (amdgpu_sriov_vf(adev)) {
2228 if (adev->is_atom_fw) {
2229 if (amdgpu_atomfirmware_gpu_supports_virtualization(adev))
2230 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2231 } else {
2232 if (amdgpu_atombios_has_gpu_virtualization_table(adev))
2233 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2234 }
2235
2236 if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
2237 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
a5bde2f9 2238 }
048765ad
AR
2239}
2240
e3ecdffa
AD
2241/**
2242 * amdgpu_device_asic_has_dc_support - determine if DC supports the asic
2243 *
2244 * @asic_type: AMD asic type
2245 *
2246 * Check if there is DC (new modesetting infrastructre) support for an asic.
2247 * returns true if DC has support, false if not.
2248 */
4562236b
HW
2249bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
2250{
2251 switch (asic_type) {
2252#if defined(CONFIG_DRM_AMD_DC)
2253 case CHIP_BONAIRE:
0d6fbccb 2254 case CHIP_KAVERI:
367e6687
AD
2255 case CHIP_KABINI:
2256 case CHIP_MULLINS:
d9fda248
HW
2257 /*
2258 * We have systems in the wild with these ASICs that require
2259 * LVDS and VGA support which is not supported with DC.
2260 *
2261 * Fallback to the non-DC driver here by default so as not to
2262 * cause regressions.
2263 */
2264 return amdgpu_dc > 0;
2265 case CHIP_HAWAII:
4562236b
HW
2266 case CHIP_CARRIZO:
2267 case CHIP_STONEY:
4562236b 2268 case CHIP_POLARIS10:
675fd32b 2269 case CHIP_POLARIS11:
2c8ad2d5 2270 case CHIP_POLARIS12:
675fd32b 2271 case CHIP_VEGAM:
4562236b
HW
2272 case CHIP_TONGA:
2273 case CHIP_FIJI:
42f8ffa1 2274 case CHIP_VEGA10:
dca7b401 2275 case CHIP_VEGA12:
c6034aa2 2276 case CHIP_VEGA20:
48412031 2277#ifdef CONFIG_X86
fd187853 2278 case CHIP_RAVEN:
42f8ffa1 2279#endif
fd187853 2280 return amdgpu_dc != 0;
4562236b
HW
2281#endif
2282 default:
2283 return false;
2284 }
2285}
2286
2287/**
2288 * amdgpu_device_has_dc_support - check if dc is supported
2289 *
2290 * @adev: amdgpu_device_pointer
2291 *
2292 * Returns true for supported, false for not supported
2293 */
2294bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
2295{
2555039d
XY
2296 if (amdgpu_sriov_vf(adev))
2297 return false;
2298
4562236b
HW
2299 return amdgpu_device_asic_has_dc_support(adev->asic_type);
2300}
2301
d38ceaf9
AD
2302/**
2303 * amdgpu_device_init - initialize the driver
2304 *
2305 * @adev: amdgpu_device pointer
87e3f136 2306 * @ddev: drm dev pointer
d38ceaf9
AD
2307 * @pdev: pci dev pointer
2308 * @flags: driver flags
2309 *
2310 * Initializes the driver info and hw (all asics).
2311 * Returns 0 for success or an error on failure.
2312 * Called at driver startup.
2313 */
2314int amdgpu_device_init(struct amdgpu_device *adev,
2315 struct drm_device *ddev,
2316 struct pci_dev *pdev,
2317 uint32_t flags)
2318{
2319 int r, i;
2320 bool runtime = false;
95844d20 2321 u32 max_MBps;
d38ceaf9
AD
2322
2323 adev->shutdown = false;
2324 adev->dev = &pdev->dev;
2325 adev->ddev = ddev;
2326 adev->pdev = pdev;
2327 adev->flags = flags;
2f7d10b3 2328 adev->asic_type = flags & AMD_ASIC_MASK;
d38ceaf9 2329 adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
593aa2d2
SL
2330 if (amdgpu_emu_mode == 1)
2331 adev->usec_timeout *= 2;
770d13b1 2332 adev->gmc.gart_size = 512 * 1024 * 1024;
d38ceaf9
AD
2333 adev->accel_working = false;
2334 adev->num_rings = 0;
2335 adev->mman.buffer_funcs = NULL;
2336 adev->mman.buffer_funcs_ring = NULL;
2337 adev->vm_manager.vm_pte_funcs = NULL;
2d55e45a 2338 adev->vm_manager.vm_pte_num_rings = 0;
132f34e4 2339 adev->gmc.gmc_funcs = NULL;
f54d1867 2340 adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
b8866c26 2341 bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
d38ceaf9
AD
2342
2343 adev->smc_rreg = &amdgpu_invalid_rreg;
2344 adev->smc_wreg = &amdgpu_invalid_wreg;
2345 adev->pcie_rreg = &amdgpu_invalid_rreg;
2346 adev->pcie_wreg = &amdgpu_invalid_wreg;
36b9a952
HR
2347 adev->pciep_rreg = &amdgpu_invalid_rreg;
2348 adev->pciep_wreg = &amdgpu_invalid_wreg;
d38ceaf9
AD
2349 adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
2350 adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
2351 adev->didt_rreg = &amdgpu_invalid_rreg;
2352 adev->didt_wreg = &amdgpu_invalid_wreg;
ccdbb20a
RZ
2353 adev->gc_cac_rreg = &amdgpu_invalid_rreg;
2354 adev->gc_cac_wreg = &amdgpu_invalid_wreg;
d38ceaf9
AD
2355 adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
2356 adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
2357
3e39ab90
AD
2358 DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
2359 amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
2360 pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
d38ceaf9
AD
2361
2362 /* mutex initialization are all done here so we
2363 * can recall function without having locking issues */
d38ceaf9 2364 atomic_set(&adev->irq.ih.lock, 0);
0e5ca0d1 2365 mutex_init(&adev->firmware.mutex);
d38ceaf9
AD
2366 mutex_init(&adev->pm.mutex);
2367 mutex_init(&adev->gfx.gpu_clock_mutex);
2368 mutex_init(&adev->srbm_mutex);
b8866c26 2369 mutex_init(&adev->gfx.pipe_reserve_mutex);
d38ceaf9 2370 mutex_init(&adev->grbm_idx_mutex);
d38ceaf9 2371 mutex_init(&adev->mn_lock);
e23b74aa 2372 mutex_init(&adev->virt.vf_errors.lock);
d38ceaf9 2373 hash_init(adev->mn_hash);
13a752e3 2374 mutex_init(&adev->lock_reset);
d38ceaf9 2375
06ec9070 2376 amdgpu_device_check_arguments(adev);
d38ceaf9 2377
d38ceaf9
AD
2378 spin_lock_init(&adev->mmio_idx_lock);
2379 spin_lock_init(&adev->smc_idx_lock);
2380 spin_lock_init(&adev->pcie_idx_lock);
2381 spin_lock_init(&adev->uvd_ctx_idx_lock);
2382 spin_lock_init(&adev->didt_idx_lock);
ccdbb20a 2383 spin_lock_init(&adev->gc_cac_idx_lock);
16abb5d2 2384 spin_lock_init(&adev->se_cac_idx_lock);
d38ceaf9 2385 spin_lock_init(&adev->audio_endpt_idx_lock);
95844d20 2386 spin_lock_init(&adev->mm_stats.lock);
d38ceaf9 2387
0c4e7fa5
CZ
2388 INIT_LIST_HEAD(&adev->shadow_list);
2389 mutex_init(&adev->shadow_list_lock);
2390
795f2813
AR
2391 INIT_LIST_HEAD(&adev->ring_lru_list);
2392 spin_lock_init(&adev->ring_lru_list_lock);
2393
06ec9070
AD
2394 INIT_DELAYED_WORK(&adev->late_init_work,
2395 amdgpu_device_ip_late_init_func_handler);
2dc80b00 2396
b1ddf548
RZ
2397 adev->pm.ac_power = power_supply_is_system_supplied() > 0 ? true : false;
2398
0fa49558
AX
2399 /* Registers mapping */
2400 /* TODO: block userspace mapping of io register */
da69c161
KW
2401 if (adev->asic_type >= CHIP_BONAIRE) {
2402 adev->rmmio_base = pci_resource_start(adev->pdev, 5);
2403 adev->rmmio_size = pci_resource_len(adev->pdev, 5);
2404 } else {
2405 adev->rmmio_base = pci_resource_start(adev->pdev, 2);
2406 adev->rmmio_size = pci_resource_len(adev->pdev, 2);
2407 }
d38ceaf9 2408
d38ceaf9
AD
2409 adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
2410 if (adev->rmmio == NULL) {
2411 return -ENOMEM;
2412 }
2413 DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
2414 DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
2415
705e519e 2416 /* doorbell bar mapping */
06ec9070 2417 amdgpu_device_doorbell_init(adev);
d38ceaf9
AD
2418
2419 /* io port mapping */
2420 for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
2421 if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) {
2422 adev->rio_mem_size = pci_resource_len(adev->pdev, i);
2423 adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size);
2424 break;
2425 }
2426 }
2427 if (adev->rio_mem == NULL)
b64a18c5 2428 DRM_INFO("PCI I/O BAR is not found.\n");
d38ceaf9 2429
5494d864
AD
2430 amdgpu_device_get_pcie_info(adev);
2431
d38ceaf9 2432 /* early init functions */
06ec9070 2433 r = amdgpu_device_ip_early_init(adev);
d38ceaf9
AD
2434 if (r)
2435 return r;
2436
2437 /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
2438 /* this will fail for cards that aren't VGA class devices, just
2439 * ignore it */
06ec9070 2440 vga_client_register(adev->pdev, adev, NULL, amdgpu_device_vga_set_decode);
d38ceaf9 2441
e9bef455 2442 if (amdgpu_device_is_px(ddev))
d38ceaf9 2443 runtime = true;
84c8b22e
LW
2444 if (!pci_is_thunderbolt_attached(adev->pdev))
2445 vga_switcheroo_register_client(adev->pdev,
2446 &amdgpu_switcheroo_ops, runtime);
d38ceaf9
AD
2447 if (runtime)
2448 vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
2449
9475a943
SL
2450 if (amdgpu_emu_mode == 1) {
2451 /* post the asic on emulation mode */
2452 emu_soc_asic_init(adev);
bfca0289 2453 goto fence_driver_init;
9475a943 2454 }
bfca0289 2455
d38ceaf9 2456 /* Read BIOS */
83ba126a
AD
2457 if (!amdgpu_get_bios(adev)) {
2458 r = -EINVAL;
2459 goto failed;
2460 }
f7e9e9fe 2461
d38ceaf9 2462 r = amdgpu_atombios_init(adev);
2c1a2784
AD
2463 if (r) {
2464 dev_err(adev->dev, "amdgpu_atombios_init failed\n");
e23b74aa 2465 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0);
83ba126a 2466 goto failed;
2c1a2784 2467 }
d38ceaf9 2468
4e99a44e
ML
2469 /* detect if we are with an SRIOV vbios */
2470 amdgpu_device_detect_sriov_bios(adev);
048765ad 2471
d38ceaf9 2472 /* Post card if necessary */
39c640c0 2473 if (amdgpu_device_need_post(adev)) {
d38ceaf9 2474 if (!adev->bios) {
bec86378 2475 dev_err(adev->dev, "no vBIOS found\n");
83ba126a
AD
2476 r = -EINVAL;
2477 goto failed;
d38ceaf9 2478 }
bec86378 2479 DRM_INFO("GPU posting now...\n");
4e99a44e
ML
2480 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2481 if (r) {
2482 dev_err(adev->dev, "gpu post error!\n");
2483 goto failed;
2484 }
d38ceaf9
AD
2485 }
2486
88b64e95
AD
2487 if (adev->is_atom_fw) {
2488 /* Initialize clocks */
2489 r = amdgpu_atomfirmware_get_clock_info(adev);
2490 if (r) {
2491 dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
e23b74aa 2492 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
88b64e95
AD
2493 goto failed;
2494 }
2495 } else {
a5bde2f9
AD
2496 /* Initialize clocks */
2497 r = amdgpu_atombios_get_clock_info(adev);
2498 if (r) {
2499 dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
e23b74aa 2500 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
89041940 2501 goto failed;
a5bde2f9
AD
2502 }
2503 /* init i2c buses */
4562236b
HW
2504 if (!amdgpu_device_has_dc_support(adev))
2505 amdgpu_atombios_i2c_init(adev);
2c1a2784 2506 }
d38ceaf9 2507
bfca0289 2508fence_driver_init:
d38ceaf9
AD
2509 /* Fence driver */
2510 r = amdgpu_fence_driver_init(adev);
2c1a2784
AD
2511 if (r) {
2512 dev_err(adev->dev, "amdgpu_fence_driver_init failed\n");
e23b74aa 2513 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
83ba126a 2514 goto failed;
2c1a2784 2515 }
d38ceaf9
AD
2516
2517 /* init the mode config */
2518 drm_mode_config_init(adev->ddev);
2519
06ec9070 2520 r = amdgpu_device_ip_init(adev);
d38ceaf9 2521 if (r) {
8840a387 2522 /* failed in exclusive mode due to timeout */
2523 if (amdgpu_sriov_vf(adev) &&
2524 !amdgpu_sriov_runtime(adev) &&
2525 amdgpu_virt_mmio_blocked(adev) &&
2526 !amdgpu_virt_wait_reset(adev)) {
2527 dev_err(adev->dev, "VF exclusive mode timeout\n");
1daee8b4
PD
2528 /* Don't send request since VF is inactive. */
2529 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
2530 adev->virt.ops = NULL;
8840a387 2531 r = -EAGAIN;
2532 goto failed;
2533 }
06ec9070 2534 dev_err(adev->dev, "amdgpu_device_ip_init failed\n");
e23b74aa 2535 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
83ba126a 2536 goto failed;
d38ceaf9
AD
2537 }
2538
2539 adev->accel_working = true;
2540
e59c0205
AX
2541 amdgpu_vm_check_compute_bug(adev);
2542
95844d20
MO
2543 /* Initialize the buffer migration limit. */
2544 if (amdgpu_moverate >= 0)
2545 max_MBps = amdgpu_moverate;
2546 else
2547 max_MBps = 8; /* Allow 8 MB/s. */
2548 /* Get a log2 for easy divisions. */
2549 adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
2550
d38ceaf9
AD
2551 r = amdgpu_ib_pool_init(adev);
2552 if (r) {
2553 dev_err(adev->dev, "IB initialization failed (%d).\n", r);
e23b74aa 2554 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r);
83ba126a 2555 goto failed;
d38ceaf9
AD
2556 }
2557
2dc8f81e
HC
2558 if (amdgpu_sriov_vf(adev))
2559 amdgpu_virt_init_data_exchange(adev);
2560
9bc92b9c
ML
2561 amdgpu_fbdev_init(adev);
2562
d2f52ac8
RZ
2563 r = amdgpu_pm_sysfs_init(adev);
2564 if (r)
2565 DRM_ERROR("registering pm debugfs failed (%d).\n", r);
2566
75758255 2567 r = amdgpu_debugfs_gem_init(adev);
3f14e623 2568 if (r)
d38ceaf9 2569 DRM_ERROR("registering gem debugfs failed (%d).\n", r);
d38ceaf9
AD
2570
2571 r = amdgpu_debugfs_regs_init(adev);
3f14e623 2572 if (r)
d38ceaf9 2573 DRM_ERROR("registering register debugfs failed (%d).\n", r);
d38ceaf9 2574
50ab2533 2575 r = amdgpu_debugfs_firmware_init(adev);
3f14e623 2576 if (r)
50ab2533 2577 DRM_ERROR("registering firmware debugfs failed (%d).\n", r);
50ab2533 2578
763efb6c 2579 r = amdgpu_debugfs_init(adev);
db95e218 2580 if (r)
763efb6c 2581 DRM_ERROR("Creating debugfs files failed (%d).\n", r);
db95e218 2582
d38ceaf9
AD
2583 if ((amdgpu_testing & 1)) {
2584 if (adev->accel_working)
2585 amdgpu_test_moves(adev);
2586 else
2587 DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n");
2588 }
d38ceaf9
AD
2589 if (amdgpu_benchmarking) {
2590 if (adev->accel_working)
2591 amdgpu_benchmark(adev, amdgpu_benchmarking);
2592 else
2593 DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n");
2594 }
2595
2596 /* enable clockgating, etc. after ib tests, etc. since some blocks require
2597 * explicit gating rather than handling it automatically.
2598 */
06ec9070 2599 r = amdgpu_device_ip_late_init(adev);
2c1a2784 2600 if (r) {
06ec9070 2601 dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n");
e23b74aa 2602 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
83ba126a 2603 goto failed;
2c1a2784 2604 }
d38ceaf9
AD
2605
2606 return 0;
83ba126a
AD
2607
2608failed:
89041940 2609 amdgpu_vf_error_trans_all(adev);
83ba126a
AD
2610 if (runtime)
2611 vga_switcheroo_fini_domain_pm_ops(adev->dev);
8840a387 2612
83ba126a 2613 return r;
d38ceaf9
AD
2614}
2615
d38ceaf9
AD
2616/**
2617 * amdgpu_device_fini - tear down the driver
2618 *
2619 * @adev: amdgpu_device pointer
2620 *
2621 * Tear down the driver info (all asics).
2622 * Called at driver shutdown.
2623 */
2624void amdgpu_device_fini(struct amdgpu_device *adev)
2625{
2626 int r;
2627
2628 DRM_INFO("amdgpu: finishing device.\n");
2629 adev->shutdown = true;
e5b03032
ML
2630 /* disable all interrupts */
2631 amdgpu_irq_disable_all(adev);
ff97cba8
ML
2632 if (adev->mode_info.mode_config_initialized){
2633 if (!amdgpu_device_has_dc_support(adev))
2634 drm_crtc_force_disable_all(adev->ddev);
2635 else
2636 drm_atomic_helper_shutdown(adev->ddev);
2637 }
d38ceaf9
AD
2638 amdgpu_ib_pool_fini(adev);
2639 amdgpu_fence_driver_fini(adev);
58e955d9 2640 amdgpu_pm_sysfs_fini(adev);
d38ceaf9 2641 amdgpu_fbdev_fini(adev);
06ec9070 2642 r = amdgpu_device_ip_fini(adev);
ab4fe3e1
HR
2643 if (adev->firmware.gpu_info_fw) {
2644 release_firmware(adev->firmware.gpu_info_fw);
2645 adev->firmware.gpu_info_fw = NULL;
2646 }
d38ceaf9 2647 adev->accel_working = false;
2dc80b00 2648 cancel_delayed_work_sync(&adev->late_init_work);
d38ceaf9 2649 /* free i2c buses */
4562236b
HW
2650 if (!amdgpu_device_has_dc_support(adev))
2651 amdgpu_i2c_fini(adev);
bfca0289
SL
2652
2653 if (amdgpu_emu_mode != 1)
2654 amdgpu_atombios_fini(adev);
2655
d38ceaf9
AD
2656 kfree(adev->bios);
2657 adev->bios = NULL;
84c8b22e
LW
2658 if (!pci_is_thunderbolt_attached(adev->pdev))
2659 vga_switcheroo_unregister_client(adev->pdev);
83ba126a
AD
2660 if (adev->flags & AMD_IS_PX)
2661 vga_switcheroo_fini_domain_pm_ops(adev->dev);
d38ceaf9
AD
2662 vga_client_register(adev->pdev, NULL, NULL, NULL);
2663 if (adev->rio_mem)
2664 pci_iounmap(adev->pdev, adev->rio_mem);
2665 adev->rio_mem = NULL;
2666 iounmap(adev->rmmio);
2667 adev->rmmio = NULL;
06ec9070 2668 amdgpu_device_doorbell_fini(adev);
d38ceaf9 2669 amdgpu_debugfs_regs_cleanup(adev);
d38ceaf9
AD
2670}
2671
2672
2673/*
2674 * Suspend & resume.
2675 */
2676/**
810ddc3a 2677 * amdgpu_device_suspend - initiate device suspend
d38ceaf9 2678 *
87e3f136
DP
2679 * @dev: drm dev pointer
2680 * @suspend: suspend state
2681 * @fbcon : notify the fbdev of suspend
d38ceaf9
AD
2682 *
2683 * Puts the hw in the suspend state (all asics).
2684 * Returns 0 for success or an error on failure.
2685 * Called at driver suspend.
2686 */
810ddc3a 2687int amdgpu_device_suspend(struct drm_device *dev, bool suspend, bool fbcon)
d38ceaf9
AD
2688{
2689 struct amdgpu_device *adev;
2690 struct drm_crtc *crtc;
2691 struct drm_connector *connector;
5ceb54c6 2692 int r;
d38ceaf9
AD
2693
2694 if (dev == NULL || dev->dev_private == NULL) {
2695 return -ENODEV;
2696 }
2697
2698 adev = dev->dev_private;
2699
2700 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
2701 return 0;
2702
2703 drm_kms_helper_poll_disable(dev);
2704
5f818173
S
2705 if (fbcon)
2706 amdgpu_fbdev_set_suspend(adev, 1);
2707
4562236b
HW
2708 if (!amdgpu_device_has_dc_support(adev)) {
2709 /* turn off display hw */
2710 drm_modeset_lock_all(dev);
2711 list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2712 drm_helper_connector_dpms(connector, DRM_MODE_DPMS_OFF);
2713 }
2714 drm_modeset_unlock_all(dev);
fe1053b7
AD
2715 /* unpin the front buffers and cursors */
2716 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2717 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2718 struct drm_framebuffer *fb = crtc->primary->fb;
2719 struct amdgpu_bo *robj;
2720
2721 if (amdgpu_crtc->cursor_bo) {
2722 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2723 r = amdgpu_bo_reserve(aobj, true);
2724 if (r == 0) {
2725 amdgpu_bo_unpin(aobj);
2726 amdgpu_bo_unreserve(aobj);
2727 }
756e6880 2728 }
756e6880 2729
fe1053b7
AD
2730 if (fb == NULL || fb->obj[0] == NULL) {
2731 continue;
2732 }
2733 robj = gem_to_amdgpu_bo(fb->obj[0]);
2734 /* don't unpin kernel fb objects */
2735 if (!amdgpu_fbdev_robj_is_fb(adev, robj)) {
2736 r = amdgpu_bo_reserve(robj, true);
2737 if (r == 0) {
2738 amdgpu_bo_unpin(robj);
2739 amdgpu_bo_unreserve(robj);
2740 }
d38ceaf9
AD
2741 }
2742 }
2743 }
fe1053b7
AD
2744
2745 amdgpu_amdkfd_suspend(adev);
2746
2747 r = amdgpu_device_ip_suspend_phase1(adev);
2748
d38ceaf9
AD
2749 /* evict vram memory */
2750 amdgpu_bo_evict_vram(adev);
2751
5ceb54c6 2752 amdgpu_fence_driver_suspend(adev);
d38ceaf9 2753
fe1053b7 2754 r = amdgpu_device_ip_suspend_phase2(adev);
d38ceaf9 2755
a0a71e49
AD
2756 /* evict remaining vram memory
2757 * This second call to evict vram is to evict the gart page table
2758 * using the CPU.
2759 */
d38ceaf9
AD
2760 amdgpu_bo_evict_vram(adev);
2761
2762 pci_save_state(dev->pdev);
2763 if (suspend) {
2764 /* Shut down the device */
2765 pci_disable_device(dev->pdev);
2766 pci_set_power_state(dev->pdev, PCI_D3hot);
74b0b157 2767 } else {
2768 r = amdgpu_asic_reset(adev);
2769 if (r)
2770 DRM_ERROR("amdgpu asic reset failed\n");
d38ceaf9
AD
2771 }
2772
d38ceaf9
AD
2773 return 0;
2774}
2775
2776/**
810ddc3a 2777 * amdgpu_device_resume - initiate device resume
d38ceaf9 2778 *
87e3f136
DP
2779 * @dev: drm dev pointer
2780 * @resume: resume state
2781 * @fbcon : notify the fbdev of resume
d38ceaf9
AD
2782 *
2783 * Bring the hw back to operating state (all asics).
2784 * Returns 0 for success or an error on failure.
2785 * Called at driver resume.
2786 */
810ddc3a 2787int amdgpu_device_resume(struct drm_device *dev, bool resume, bool fbcon)
d38ceaf9
AD
2788{
2789 struct drm_connector *connector;
2790 struct amdgpu_device *adev = dev->dev_private;
756e6880 2791 struct drm_crtc *crtc;
03161a6e 2792 int r = 0;
d38ceaf9
AD
2793
2794 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
2795 return 0;
2796
d38ceaf9
AD
2797 if (resume) {
2798 pci_set_power_state(dev->pdev, PCI_D0);
2799 pci_restore_state(dev->pdev);
74b0b157 2800 r = pci_enable_device(dev->pdev);
03161a6e 2801 if (r)
4d3b9ae5 2802 return r;
d38ceaf9
AD
2803 }
2804
2805 /* post card */
39c640c0 2806 if (amdgpu_device_need_post(adev)) {
74b0b157 2807 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2808 if (r)
2809 DRM_ERROR("amdgpu asic init failed\n");
2810 }
d38ceaf9 2811
06ec9070 2812 r = amdgpu_device_ip_resume(adev);
e6707218 2813 if (r) {
06ec9070 2814 DRM_ERROR("amdgpu_device_ip_resume failed (%d).\n", r);
4d3b9ae5 2815 return r;
e6707218 2816 }
5ceb54c6
AD
2817 amdgpu_fence_driver_resume(adev);
2818
d38ceaf9 2819
06ec9070 2820 r = amdgpu_device_ip_late_init(adev);
03161a6e 2821 if (r)
4d3b9ae5 2822 return r;
d38ceaf9 2823
fe1053b7
AD
2824 if (!amdgpu_device_has_dc_support(adev)) {
2825 /* pin cursors */
2826 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2827 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2828
2829 if (amdgpu_crtc->cursor_bo) {
2830 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2831 r = amdgpu_bo_reserve(aobj, true);
2832 if (r == 0) {
2833 r = amdgpu_bo_pin(aobj, AMDGPU_GEM_DOMAIN_VRAM);
2834 if (r != 0)
2835 DRM_ERROR("Failed to pin cursor BO (%d)\n", r);
2836 amdgpu_crtc->cursor_addr = amdgpu_bo_gpu_offset(aobj);
2837 amdgpu_bo_unreserve(aobj);
2838 }
756e6880
AD
2839 }
2840 }
2841 }
ba997709
YZ
2842 r = amdgpu_amdkfd_resume(adev);
2843 if (r)
2844 return r;
756e6880 2845
d38ceaf9
AD
2846 /* blat the mode back in */
2847 if (fbcon) {
4562236b
HW
2848 if (!amdgpu_device_has_dc_support(adev)) {
2849 /* pre DCE11 */
2850 drm_helper_resume_force_mode(dev);
2851
2852 /* turn on display hw */
2853 drm_modeset_lock_all(dev);
2854 list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2855 drm_helper_connector_dpms(connector, DRM_MODE_DPMS_ON);
2856 }
2857 drm_modeset_unlock_all(dev);
d38ceaf9 2858 }
4d3b9ae5 2859 amdgpu_fbdev_set_suspend(adev, 0);
d38ceaf9
AD
2860 }
2861
2862 drm_kms_helper_poll_enable(dev);
23a1a9e5
L
2863
2864 /*
2865 * Most of the connector probing functions try to acquire runtime pm
2866 * refs to ensure that the GPU is powered on when connector polling is
2867 * performed. Since we're calling this from a runtime PM callback,
2868 * trying to acquire rpm refs will cause us to deadlock.
2869 *
2870 * Since we're guaranteed to be holding the rpm lock, it's safe to
2871 * temporarily disable the rpm helpers so this doesn't deadlock us.
2872 */
2873#ifdef CONFIG_PM
2874 dev->dev->power.disable_depth++;
2875#endif
4562236b
HW
2876 if (!amdgpu_device_has_dc_support(adev))
2877 drm_helper_hpd_irq_event(dev);
2878 else
2879 drm_kms_helper_hotplug_event(dev);
23a1a9e5
L
2880#ifdef CONFIG_PM
2881 dev->dev->power.disable_depth--;
2882#endif
4d3b9ae5 2883 return 0;
d38ceaf9
AD
2884}
2885
e3ecdffa
AD
2886/**
2887 * amdgpu_device_ip_check_soft_reset - did soft reset succeed
2888 *
2889 * @adev: amdgpu_device pointer
2890 *
2891 * The list of all the hardware IPs that make up the asic is walked and
2892 * the check_soft_reset callbacks are run. check_soft_reset determines
2893 * if the asic is still hung or not.
2894 * Returns true if any of the IPs are still in a hung state, false if not.
2895 */
06ec9070 2896static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)
63fbf42f
CZ
2897{
2898 int i;
2899 bool asic_hang = false;
2900
f993d628
ML
2901 if (amdgpu_sriov_vf(adev))
2902 return true;
2903
8bc04c29
AD
2904 if (amdgpu_asic_need_full_reset(adev))
2905 return true;
2906
63fbf42f 2907 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 2908 if (!adev->ip_blocks[i].status.valid)
63fbf42f 2909 continue;
a1255107
AD
2910 if (adev->ip_blocks[i].version->funcs->check_soft_reset)
2911 adev->ip_blocks[i].status.hang =
2912 adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
2913 if (adev->ip_blocks[i].status.hang) {
2914 DRM_INFO("IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
63fbf42f
CZ
2915 asic_hang = true;
2916 }
2917 }
2918 return asic_hang;
2919}
2920
e3ecdffa
AD
2921/**
2922 * amdgpu_device_ip_pre_soft_reset - prepare for soft reset
2923 *
2924 * @adev: amdgpu_device pointer
2925 *
2926 * The list of all the hardware IPs that make up the asic is walked and the
2927 * pre_soft_reset callbacks are run if the block is hung. pre_soft_reset
2928 * handles any IP specific hardware or software state changes that are
2929 * necessary for a soft reset to succeed.
2930 * Returns 0 on success, negative error code on failure.
2931 */
06ec9070 2932static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)
d31a501e
CZ
2933{
2934 int i, r = 0;
2935
2936 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 2937 if (!adev->ip_blocks[i].status.valid)
d31a501e 2938 continue;
a1255107
AD
2939 if (adev->ip_blocks[i].status.hang &&
2940 adev->ip_blocks[i].version->funcs->pre_soft_reset) {
2941 r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
d31a501e
CZ
2942 if (r)
2943 return r;
2944 }
2945 }
2946
2947 return 0;
2948}
2949
e3ecdffa
AD
2950/**
2951 * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed
2952 *
2953 * @adev: amdgpu_device pointer
2954 *
2955 * Some hardware IPs cannot be soft reset. If they are hung, a full gpu
2956 * reset is necessary to recover.
2957 * Returns true if a full asic reset is required, false if not.
2958 */
06ec9070 2959static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)
35d782fe 2960{
da146d3b
AD
2961 int i;
2962
8bc04c29
AD
2963 if (amdgpu_asic_need_full_reset(adev))
2964 return true;
2965
da146d3b 2966 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 2967 if (!adev->ip_blocks[i].status.valid)
da146d3b 2968 continue;
a1255107
AD
2969 if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
2970 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
2971 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
98512bb8
KW
2972 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
2973 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
a1255107 2974 if (adev->ip_blocks[i].status.hang) {
da146d3b
AD
2975 DRM_INFO("Some block need full reset!\n");
2976 return true;
2977 }
2978 }
35d782fe
CZ
2979 }
2980 return false;
2981}
2982
e3ecdffa
AD
2983/**
2984 * amdgpu_device_ip_soft_reset - do a soft reset
2985 *
2986 * @adev: amdgpu_device pointer
2987 *
2988 * The list of all the hardware IPs that make up the asic is walked and the
2989 * soft_reset callbacks are run if the block is hung. soft_reset handles any
2990 * IP specific hardware or software state changes that are necessary to soft
2991 * reset the IP.
2992 * Returns 0 on success, negative error code on failure.
2993 */
06ec9070 2994static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)
35d782fe
CZ
2995{
2996 int i, r = 0;
2997
2998 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 2999 if (!adev->ip_blocks[i].status.valid)
35d782fe 3000 continue;
a1255107
AD
3001 if (adev->ip_blocks[i].status.hang &&
3002 adev->ip_blocks[i].version->funcs->soft_reset) {
3003 r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
35d782fe
CZ
3004 if (r)
3005 return r;
3006 }
3007 }
3008
3009 return 0;
3010}
3011
e3ecdffa
AD
3012/**
3013 * amdgpu_device_ip_post_soft_reset - clean up from soft reset
3014 *
3015 * @adev: amdgpu_device pointer
3016 *
3017 * The list of all the hardware IPs that make up the asic is walked and the
3018 * post_soft_reset callbacks are run if the asic was hung. post_soft_reset
3019 * handles any IP specific hardware or software state changes that are
3020 * necessary after the IP has been soft reset.
3021 * Returns 0 on success, negative error code on failure.
3022 */
06ec9070 3023static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)
35d782fe
CZ
3024{
3025 int i, r = 0;
3026
3027 for (i = 0; i < adev->num_ip_blocks; i++) {
a1255107 3028 if (!adev->ip_blocks[i].status.valid)
35d782fe 3029 continue;
a1255107
AD
3030 if (adev->ip_blocks[i].status.hang &&
3031 adev->ip_blocks[i].version->funcs->post_soft_reset)
3032 r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
35d782fe
CZ
3033 if (r)
3034 return r;
3035 }
3036
3037 return 0;
3038}
3039
e3ecdffa
AD
3040/**
3041 * amdgpu_device_recover_vram_from_shadow - restore shadowed VRAM buffers
3042 *
3043 * @adev: amdgpu_device pointer
3044 * @ring: amdgpu_ring for the engine handling the buffer operations
3045 * @bo: amdgpu_bo buffer whose shadow is being restored
3046 * @fence: dma_fence associated with the operation
3047 *
3048 * Restores the VRAM buffer contents from the shadow in GTT. Used to
3049 * restore things like GPUVM page tables after a GPU reset where
3050 * the contents of VRAM might be lost.
3051 * Returns 0 on success, negative error code on failure.
3052 */
06ec9070
AD
3053static int amdgpu_device_recover_vram_from_shadow(struct amdgpu_device *adev,
3054 struct amdgpu_ring *ring,
3055 struct amdgpu_bo *bo,
3056 struct dma_fence **fence)
53cdccd5
CZ
3057{
3058 uint32_t domain;
3059 int r;
3060
23d2e504
RH
3061 if (!bo->shadow)
3062 return 0;
3063
1d284797 3064 r = amdgpu_bo_reserve(bo, true);
23d2e504
RH
3065 if (r)
3066 return r;
3067 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
3068 /* if bo has been evicted, then no need to recover */
3069 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
82521316
RH
3070 r = amdgpu_bo_validate(bo->shadow);
3071 if (r) {
3072 DRM_ERROR("bo validate failed!\n");
3073 goto err;
3074 }
3075
23d2e504 3076 r = amdgpu_bo_restore_from_shadow(adev, ring, bo,
53cdccd5 3077 NULL, fence, true);
23d2e504
RH
3078 if (r) {
3079 DRM_ERROR("recover page table failed!\n");
3080 goto err;
3081 }
3082 }
53cdccd5 3083err:
23d2e504
RH
3084 amdgpu_bo_unreserve(bo);
3085 return r;
53cdccd5
CZ
3086}
3087
e3ecdffa
AD
3088/**
3089 * amdgpu_device_handle_vram_lost - Handle the loss of VRAM contents
3090 *
3091 * @adev: amdgpu_device pointer
3092 *
3093 * Restores the contents of VRAM buffers from the shadows in GTT. Used to
3094 * restore things like GPUVM page tables after a GPU reset where
3095 * the contents of VRAM might be lost.
3096 * Returns 0 on success, 1 on failure.
3097 */
c41d1cf6
ML
3098static int amdgpu_device_handle_vram_lost(struct amdgpu_device *adev)
3099{
3100 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
3101 struct amdgpu_bo *bo, *tmp;
3102 struct dma_fence *fence = NULL, *next = NULL;
3103 long r = 1;
3104 int i = 0;
3105 long tmo;
3106
3107 if (amdgpu_sriov_runtime(adev))
3108 tmo = msecs_to_jiffies(amdgpu_lockup_timeout);
3109 else
3110 tmo = msecs_to_jiffies(100);
3111
3112 DRM_INFO("recover vram bo from shadow start\n");
3113 mutex_lock(&adev->shadow_list_lock);
3114 list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) {
3115 next = NULL;
3116 amdgpu_device_recover_vram_from_shadow(adev, ring, bo, &next);
3117 if (fence) {
3118 r = dma_fence_wait_timeout(fence, false, tmo);
3119 if (r == 0)
3120 pr_err("wait fence %p[%d] timeout\n", fence, i);
3121 else if (r < 0)
3122 pr_err("wait fence %p[%d] interrupted\n", fence, i);
3123 if (r < 1) {
3124 dma_fence_put(fence);
3125 fence = next;
3126 break;
3127 }
3128 i++;
3129 }
3130
3131 dma_fence_put(fence);
3132 fence = next;
3133 }
3134 mutex_unlock(&adev->shadow_list_lock);
3135
3136 if (fence) {
3137 r = dma_fence_wait_timeout(fence, false, tmo);
3138 if (r == 0)
3139 pr_err("wait fence %p[%d] timeout\n", fence, i);
3140 else if (r < 0)
3141 pr_err("wait fence %p[%d] interrupted\n", fence, i);
3142
3143 }
3144 dma_fence_put(fence);
3145
3146 if (r > 0)
3147 DRM_INFO("recover vram bo from shadow done\n");
3148 else
3149 DRM_ERROR("recover vram bo from shadow failed\n");
3150
e3ecdffa 3151 return (r > 0) ? 0 : 1;
c41d1cf6
ML
3152}
3153
e3ecdffa 3154/**
06ec9070 3155 * amdgpu_device_reset - reset ASIC/GPU for bare-metal or passthrough
a90ad3c2
ML
3156 *
3157 * @adev: amdgpu device pointer
a90ad3c2 3158 *
5740682e 3159 * attempt to do soft-reset or full-reset and reinitialize Asic
3f48c681 3160 * return 0 means succeeded otherwise failed
e3ecdffa 3161 */
c41d1cf6 3162static int amdgpu_device_reset(struct amdgpu_device *adev)
a90ad3c2 3163{
5740682e
ML
3164 bool need_full_reset, vram_lost = 0;
3165 int r;
a90ad3c2 3166
06ec9070 3167 need_full_reset = amdgpu_device_ip_need_full_reset(adev);
a90ad3c2 3168
5740682e 3169 if (!need_full_reset) {
06ec9070
AD
3170 amdgpu_device_ip_pre_soft_reset(adev);
3171 r = amdgpu_device_ip_soft_reset(adev);
3172 amdgpu_device_ip_post_soft_reset(adev);
3173 if (r || amdgpu_device_ip_check_soft_reset(adev)) {
5740682e
ML
3174 DRM_INFO("soft reset failed, will fallback to full reset!\n");
3175 need_full_reset = true;
3176 }
5740682e 3177 }
a90ad3c2 3178
5740682e 3179 if (need_full_reset) {
cdd61df6 3180 r = amdgpu_device_ip_suspend(adev);
a90ad3c2 3181
5740682e 3182retry:
5740682e 3183 r = amdgpu_asic_reset(adev);
5740682e
ML
3184 /* post card */
3185 amdgpu_atom_asic_init(adev->mode_info.atom_context);
65781c78 3186
5740682e
ML
3187 if (!r) {
3188 dev_info(adev->dev, "GPU reset succeeded, trying to resume\n");
06ec9070 3189 r = amdgpu_device_ip_resume_phase1(adev);
5740682e
ML
3190 if (r)
3191 goto out;
65781c78 3192
06ec9070 3193 vram_lost = amdgpu_device_check_vram_lost(adev);
5740682e
ML
3194 if (vram_lost) {
3195 DRM_ERROR("VRAM is lost!\n");
3196 atomic_inc(&adev->vram_lost_counter);
3197 }
3198
c1c7ce8f
CK
3199 r = amdgpu_gtt_mgr_recover(
3200 &adev->mman.bdev.man[TTM_PL_TT]);
5740682e
ML
3201 if (r)
3202 goto out;
3203
06ec9070 3204 r = amdgpu_device_ip_resume_phase2(adev);
5740682e
ML
3205 if (r)
3206 goto out;
3207
3208 if (vram_lost)
06ec9070 3209 amdgpu_device_fill_reset_magic(adev);
65781c78 3210 }
5740682e 3211 }
65781c78 3212
5740682e
ML
3213out:
3214 if (!r) {
3215 amdgpu_irq_gpu_reset_resume_helper(adev);
3216 r = amdgpu_ib_ring_tests(adev);
3217 if (r) {
3218 dev_err(adev->dev, "ib ring test failed (%d).\n", r);
cdd61df6 3219 r = amdgpu_device_ip_suspend(adev);
5740682e
ML
3220 need_full_reset = true;
3221 goto retry;
3222 }
3223 }
65781c78 3224
c41d1cf6
ML
3225 if (!r && ((need_full_reset && !(adev->flags & AMD_IS_APU)) || vram_lost))
3226 r = amdgpu_device_handle_vram_lost(adev);
a90ad3c2 3227
5740682e
ML
3228 return r;
3229}
a90ad3c2 3230
e3ecdffa 3231/**
06ec9070 3232 * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
5740682e
ML
3233 *
3234 * @adev: amdgpu device pointer
87e3f136 3235 * @from_hypervisor: request from hypervisor
5740682e
ML
3236 *
3237 * do VF FLR and reinitialize Asic
3f48c681 3238 * return 0 means succeeded otherwise failed
e3ecdffa
AD
3239 */
3240static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
3241 bool from_hypervisor)
5740682e
ML
3242{
3243 int r;
3244
3245 if (from_hypervisor)
3246 r = amdgpu_virt_request_full_gpu(adev, true);
3247 else
3248 r = amdgpu_virt_reset_gpu(adev);
3249 if (r)
3250 return r;
a90ad3c2
ML
3251
3252 /* Resume IP prior to SMC */
06ec9070 3253 r = amdgpu_device_ip_reinit_early_sriov(adev);
5740682e
ML
3254 if (r)
3255 goto error;
a90ad3c2
ML
3256
3257 /* we need recover gart prior to run SMC/CP/SDMA resume */
c1c7ce8f 3258 amdgpu_gtt_mgr_recover(&adev->mman.bdev.man[TTM_PL_TT]);
a90ad3c2
ML
3259
3260 /* now we are okay to resume SMC/CP/SDMA */
06ec9070 3261 r = amdgpu_device_ip_reinit_late_sriov(adev);
5740682e
ML
3262 if (r)
3263 goto error;
a90ad3c2
ML
3264
3265 amdgpu_irq_gpu_reset_resume_helper(adev);
5740682e 3266 r = amdgpu_ib_ring_tests(adev);
a90ad3c2 3267
abc34253
ED
3268error:
3269 amdgpu_virt_release_full_gpu(adev, true);
c41d1cf6
ML
3270 if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
3271 atomic_inc(&adev->vram_lost_counter);
3272 r = amdgpu_device_handle_vram_lost(adev);
a90ad3c2
ML
3273 }
3274
3275 return r;
3276}
3277
d38ceaf9 3278/**
5f152b5e 3279 * amdgpu_device_gpu_recover - reset the asic and recover scheduler
d38ceaf9
AD
3280 *
3281 * @adev: amdgpu device pointer
5740682e 3282 * @job: which job trigger hang
87e3f136 3283 * @force: forces reset regardless of amdgpu_gpu_recovery
d38ceaf9 3284 *
5740682e 3285 * Attempt to reset the GPU if it has hung (all asics).
d38ceaf9
AD
3286 * Returns 0 for success or an error on failure.
3287 */
5f152b5e
AD
3288int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
3289 struct amdgpu_job *job, bool force)
d38ceaf9 3290{
5740682e 3291 int i, r, resched;
fb140b29 3292
54bc1398 3293 if (!force && !amdgpu_device_ip_check_soft_reset(adev)) {
63fbf42f
CZ
3294 DRM_INFO("No hardware hang detected. Did some blocks stall?\n");
3295 return 0;
3296 }
d38ceaf9 3297
dcebf026
AG
3298 if (!force && (amdgpu_gpu_recovery == 0 ||
3299 (amdgpu_gpu_recovery == -1 && !amdgpu_sriov_vf(adev)))) {
3300 DRM_INFO("GPU recovery disabled.\n");
3301 return 0;
3302 }
3303
5740682e
ML
3304 dev_info(adev->dev, "GPU reset begin!\n");
3305
13a752e3 3306 mutex_lock(&adev->lock_reset);
d94aed5a 3307 atomic_inc(&adev->gpu_reset_counter);
13a752e3 3308 adev->in_gpu_reset = 1;
d38ceaf9 3309
a3c47d6b
CZ
3310 /* block TTM */
3311 resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
71182665 3312
71182665 3313 /* block all schedulers and reset given job's ring */
0875dc9e
CZ
3314 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3315 struct amdgpu_ring *ring = adev->rings[i];
3316
51687759 3317 if (!ring || !ring->sched.thread)
0875dc9e 3318 continue;
5740682e 3319
71182665
ML
3320 kthread_park(ring->sched.thread);
3321
3320b8d2 3322 if (job && job->base.sched == &ring->sched)
5740682e
ML
3323 continue;
3324
1b1f42d8 3325 drm_sched_hw_job_reset(&ring->sched, &job->base);
5740682e 3326
2f9d4084
ML
3327 /* after all hw jobs are reset, hw fence is meaningless, so force_completion */
3328 amdgpu_fence_driver_force_completion(ring);
0875dc9e 3329 }
d38ceaf9 3330
5740682e 3331 if (amdgpu_sriov_vf(adev))
c41d1cf6 3332 r = amdgpu_device_reset_sriov(adev, job ? false : true);
5740682e 3333 else
c41d1cf6 3334 r = amdgpu_device_reset(adev);
5740682e 3335
71182665
ML
3336 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3337 struct amdgpu_ring *ring = adev->rings[i];
53cdccd5 3338
71182665
ML
3339 if (!ring || !ring->sched.thread)
3340 continue;
5740682e 3341
71182665
ML
3342 /* only need recovery sched of the given job's ring
3343 * or all rings (in the case @job is NULL)
3344 * after above amdgpu_reset accomplished
3345 */
3320b8d2 3346 if ((!job || job->base.sched == &ring->sched) && !r)
1b1f42d8 3347 drm_sched_job_recovery(&ring->sched);
5740682e 3348
71182665 3349 kthread_unpark(ring->sched.thread);
d38ceaf9
AD
3350 }
3351
bf830604 3352 if (!amdgpu_device_has_dc_support(adev)) {
4562236b 3353 drm_helper_resume_force_mode(adev->ddev);
5740682e 3354 }
d38ceaf9
AD
3355
3356 ttm_bo_unlock_delayed_workqueue(&adev->mman.bdev, resched);
5740682e 3357
89041940 3358 if (r) {
d38ceaf9 3359 /* bad news, how to tell it to userspace ? */
5740682e
ML
3360 dev_info(adev->dev, "GPU reset(%d) failed\n", atomic_read(&adev->gpu_reset_counter));
3361 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r);
3362 } else {
3f48c681 3363 dev_info(adev->dev, "GPU reset(%d) succeeded!\n",atomic_read(&adev->gpu_reset_counter));
89041940 3364 }
d38ceaf9 3365
89041940 3366 amdgpu_vf_error_trans_all(adev);
13a752e3
ML
3367 adev->in_gpu_reset = 0;
3368 mutex_unlock(&adev->lock_reset);
d38ceaf9
AD
3369 return r;
3370}
3371
e3ecdffa
AD
3372/**
3373 * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot
3374 *
3375 * @adev: amdgpu_device pointer
3376 *
3377 * Fetchs and stores in the driver the PCIE capabilities (gen speed
3378 * and lanes) of the slot the device is in. Handles APUs and
3379 * virtualized environments where PCIE config space may not be available.
3380 */
5494d864 3381static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev)
d0dd7f0c 3382{
5d9a6330
AD
3383 struct pci_dev *pdev;
3384 enum pci_bus_speed speed_cap;
3385 enum pcie_link_width link_width;
d0dd7f0c 3386
cd474ba0
AD
3387 if (amdgpu_pcie_gen_cap)
3388 adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
d0dd7f0c 3389
cd474ba0
AD
3390 if (amdgpu_pcie_lane_cap)
3391 adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
d0dd7f0c 3392
cd474ba0
AD
3393 /* covers APUs as well */
3394 if (pci_is_root_bus(adev->pdev->bus)) {
3395 if (adev->pm.pcie_gen_mask == 0)
3396 adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
3397 if (adev->pm.pcie_mlw_mask == 0)
3398 adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
d0dd7f0c 3399 return;
cd474ba0 3400 }
d0dd7f0c 3401
cd474ba0 3402 if (adev->pm.pcie_gen_mask == 0) {
5d9a6330
AD
3403 /* asic caps */
3404 pdev = adev->pdev;
3405 speed_cap = pcie_get_speed_cap(pdev);
3406 if (speed_cap == PCI_SPEED_UNKNOWN) {
3407 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
cd474ba0
AD
3408 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3409 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
cd474ba0 3410 } else {
5d9a6330
AD
3411 if (speed_cap == PCIE_SPEED_16_0GT)
3412 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3413 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3414 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
3415 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4);
3416 else if (speed_cap == PCIE_SPEED_8_0GT)
3417 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3418 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3419 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
3420 else if (speed_cap == PCIE_SPEED_5_0GT)
3421 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3422 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2);
3423 else
3424 adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1;
3425 }
3426 /* platform caps */
3427 pdev = adev->ddev->pdev->bus->self;
3428 speed_cap = pcie_get_speed_cap(pdev);
3429 if (speed_cap == PCI_SPEED_UNKNOWN) {
3430 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3431 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
3432 } else {
3433 if (speed_cap == PCIE_SPEED_16_0GT)
3434 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3435 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3436 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
3437 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4);
3438 else if (speed_cap == PCIE_SPEED_8_0GT)
3439 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3440 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3441 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3);
3442 else if (speed_cap == PCIE_SPEED_5_0GT)
3443 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3444 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
3445 else
3446 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
3447
cd474ba0
AD
3448 }
3449 }
3450 if (adev->pm.pcie_mlw_mask == 0) {
5d9a6330
AD
3451 pdev = adev->ddev->pdev->bus->self;
3452 link_width = pcie_get_width_cap(pdev);
3453 if (link_width == PCIE_LNK_WIDTH_UNKNOWN) {
3454 adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK;
3455 } else {
3456 switch (link_width) {
3457 case PCIE_LNK_X32:
cd474ba0
AD
3458 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
3459 CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
3460 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3461 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3462 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3463 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3464 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3465 break;
5d9a6330 3466 case PCIE_LNK_X16:
cd474ba0
AD
3467 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
3468 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3469 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3470 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3471 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3472 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3473 break;
5d9a6330 3474 case PCIE_LNK_X12:
cd474ba0
AD
3475 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3476 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3477 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3478 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3479 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3480 break;
5d9a6330 3481 case PCIE_LNK_X8:
cd474ba0
AD
3482 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3483 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3484 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3485 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3486 break;
5d9a6330 3487 case PCIE_LNK_X4:
cd474ba0
AD
3488 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3489 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3490 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3491 break;
5d9a6330 3492 case PCIE_LNK_X2:
cd474ba0
AD
3493 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3494 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3495 break;
5d9a6330 3496 case PCIE_LNK_X1:
cd474ba0
AD
3497 adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;
3498 break;
3499 default:
3500 break;
3501 }
d0dd7f0c
AD
3502 }
3503 }
3504}
d38ceaf9 3505