Loading...
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 */
28#include <linux/power_supply.h>
29#include <linux/kthread.h>
30#include <linux/module.h>
31#include <linux/console.h>
32#include <linux/slab.h>
33#include <linux/iommu.h>
34#include <linux/pci.h>
35#include <linux/pci-p2pdma.h>
36#include <linux/apple-gmux.h>
37
38#include <drm/drm_aperture.h>
39#include <drm/drm_atomic_helper.h>
40#include <drm/drm_crtc_helper.h>
41#include <drm/drm_fb_helper.h>
42#include <drm/drm_probe_helper.h>
43#include <drm/amdgpu_drm.h>
44#include <linux/device.h>
45#include <linux/vgaarb.h>
46#include <linux/vga_switcheroo.h>
47#include <linux/efi.h>
48#include "amdgpu.h"
49#include "amdgpu_trace.h"
50#include "amdgpu_i2c.h"
51#include "atom.h"
52#include "amdgpu_atombios.h"
53#include "amdgpu_atomfirmware.h"
54#include "amd_pcie.h"
55#ifdef CONFIG_DRM_AMDGPU_SI
56#include "si.h"
57#endif
58#ifdef CONFIG_DRM_AMDGPU_CIK
59#include "cik.h"
60#endif
61#include "vi.h"
62#include "soc15.h"
63#include "nv.h"
64#include "bif/bif_4_1_d.h"
65#include <linux/firmware.h>
66#include "amdgpu_vf_error.h"
67
68#include "amdgpu_amdkfd.h"
69#include "amdgpu_pm.h"
70
71#include "amdgpu_xgmi.h"
72#include "amdgpu_ras.h"
73#include "amdgpu_pmu.h"
74#include "amdgpu_fru_eeprom.h"
75#include "amdgpu_reset.h"
76#include "amdgpu_virt.h"
77
78#include <linux/suspend.h>
79#include <drm/task_barrier.h>
80#include <linux/pm_runtime.h>
81
82#include <drm/drm_drv.h>
83
84#if IS_ENABLED(CONFIG_X86)
85#include <asm/intel-family.h>
86#endif
87
88MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin");
89MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin");
90MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin");
91MODULE_FIRMWARE("amdgpu/picasso_gpu_info.bin");
92MODULE_FIRMWARE("amdgpu/raven2_gpu_info.bin");
93MODULE_FIRMWARE("amdgpu/arcturus_gpu_info.bin");
94MODULE_FIRMWARE("amdgpu/navi12_gpu_info.bin");
95
96#define AMDGPU_RESUME_MS 2000
97#define AMDGPU_MAX_RETRY_LIMIT 2
98#define AMDGPU_RETRY_SRIOV_RESET(r) ((r) == -EBUSY || (r) == -ETIMEDOUT || (r) == -EINVAL)
99
100static const struct drm_driver amdgpu_kms_driver;
101
102const char *amdgpu_asic_name[] = {
103 "TAHITI",
104 "PITCAIRN",
105 "VERDE",
106 "OLAND",
107 "HAINAN",
108 "BONAIRE",
109 "KAVERI",
110 "KABINI",
111 "HAWAII",
112 "MULLINS",
113 "TOPAZ",
114 "TONGA",
115 "FIJI",
116 "CARRIZO",
117 "STONEY",
118 "POLARIS10",
119 "POLARIS11",
120 "POLARIS12",
121 "VEGAM",
122 "VEGA10",
123 "VEGA12",
124 "VEGA20",
125 "RAVEN",
126 "ARCTURUS",
127 "RENOIR",
128 "ALDEBARAN",
129 "NAVI10",
130 "CYAN_SKILLFISH",
131 "NAVI14",
132 "NAVI12",
133 "SIENNA_CICHLID",
134 "NAVY_FLOUNDER",
135 "VANGOGH",
136 "DIMGREY_CAVEFISH",
137 "BEIGE_GOBY",
138 "YELLOW_CARP",
139 "IP DISCOVERY",
140 "LAST",
141};
142
143/**
144 * DOC: pcie_replay_count
145 *
146 * The amdgpu driver provides a sysfs API for reporting the total number
147 * of PCIe replays (NAKs)
148 * The file pcie_replay_count is used for this and returns the total
149 * number of replays as a sum of the NAKs generated and NAKs received
150 */
151
152static ssize_t amdgpu_device_get_pcie_replay_count(struct device *dev,
153 struct device_attribute *attr, char *buf)
154{
155 struct drm_device *ddev = dev_get_drvdata(dev);
156 struct amdgpu_device *adev = drm_to_adev(ddev);
157 uint64_t cnt = amdgpu_asic_get_pcie_replay_count(adev);
158
159 return sysfs_emit(buf, "%llu\n", cnt);
160}
161
162static DEVICE_ATTR(pcie_replay_count, 0444,
163 amdgpu_device_get_pcie_replay_count, NULL);
164
165static ssize_t amdgpu_sysfs_reg_state_get(struct file *f, struct kobject *kobj,
166 struct bin_attribute *attr, char *buf,
167 loff_t ppos, size_t count)
168{
169 struct device *dev = kobj_to_dev(kobj);
170 struct drm_device *ddev = dev_get_drvdata(dev);
171 struct amdgpu_device *adev = drm_to_adev(ddev);
172 ssize_t bytes_read;
173
174 switch (ppos) {
175 case AMDGPU_SYS_REG_STATE_XGMI:
176 bytes_read = amdgpu_asic_get_reg_state(
177 adev, AMDGPU_REG_STATE_TYPE_XGMI, buf, count);
178 break;
179 case AMDGPU_SYS_REG_STATE_WAFL:
180 bytes_read = amdgpu_asic_get_reg_state(
181 adev, AMDGPU_REG_STATE_TYPE_WAFL, buf, count);
182 break;
183 case AMDGPU_SYS_REG_STATE_PCIE:
184 bytes_read = amdgpu_asic_get_reg_state(
185 adev, AMDGPU_REG_STATE_TYPE_PCIE, buf, count);
186 break;
187 case AMDGPU_SYS_REG_STATE_USR:
188 bytes_read = amdgpu_asic_get_reg_state(
189 adev, AMDGPU_REG_STATE_TYPE_USR, buf, count);
190 break;
191 case AMDGPU_SYS_REG_STATE_USR_1:
192 bytes_read = amdgpu_asic_get_reg_state(
193 adev, AMDGPU_REG_STATE_TYPE_USR_1, buf, count);
194 break;
195 default:
196 return -EINVAL;
197 }
198
199 return bytes_read;
200}
201
202BIN_ATTR(reg_state, 0444, amdgpu_sysfs_reg_state_get, NULL,
203 AMDGPU_SYS_REG_STATE_END);
204
205int amdgpu_reg_state_sysfs_init(struct amdgpu_device *adev)
206{
207 int ret;
208
209 if (!amdgpu_asic_get_reg_state_supported(adev))
210 return 0;
211
212 ret = sysfs_create_bin_file(&adev->dev->kobj, &bin_attr_reg_state);
213
214 return ret;
215}
216
217void amdgpu_reg_state_sysfs_fini(struct amdgpu_device *adev)
218{
219 if (!amdgpu_asic_get_reg_state_supported(adev))
220 return;
221 sysfs_remove_bin_file(&adev->dev->kobj, &bin_attr_reg_state);
222}
223
224/**
225 * DOC: board_info
226 *
227 * The amdgpu driver provides a sysfs API for giving board related information.
228 * It provides the form factor information in the format
229 *
230 * type : form factor
231 *
232 * Possible form factor values
233 *
234 * - "cem" - PCIE CEM card
235 * - "oam" - Open Compute Accelerator Module
236 * - "unknown" - Not known
237 *
238 */
239
240static ssize_t amdgpu_device_get_board_info(struct device *dev,
241 struct device_attribute *attr,
242 char *buf)
243{
244 struct drm_device *ddev = dev_get_drvdata(dev);
245 struct amdgpu_device *adev = drm_to_adev(ddev);
246 enum amdgpu_pkg_type pkg_type = AMDGPU_PKG_TYPE_CEM;
247 const char *pkg;
248
249 if (adev->smuio.funcs && adev->smuio.funcs->get_pkg_type)
250 pkg_type = adev->smuio.funcs->get_pkg_type(adev);
251
252 switch (pkg_type) {
253 case AMDGPU_PKG_TYPE_CEM:
254 pkg = "cem";
255 break;
256 case AMDGPU_PKG_TYPE_OAM:
257 pkg = "oam";
258 break;
259 default:
260 pkg = "unknown";
261 break;
262 }
263
264 return sysfs_emit(buf, "%s : %s\n", "type", pkg);
265}
266
267static DEVICE_ATTR(board_info, 0444, amdgpu_device_get_board_info, NULL);
268
269static struct attribute *amdgpu_board_attrs[] = {
270 &dev_attr_board_info.attr,
271 NULL,
272};
273
274static umode_t amdgpu_board_attrs_is_visible(struct kobject *kobj,
275 struct attribute *attr, int n)
276{
277 struct device *dev = kobj_to_dev(kobj);
278 struct drm_device *ddev = dev_get_drvdata(dev);
279 struct amdgpu_device *adev = drm_to_adev(ddev);
280
281 if (adev->flags & AMD_IS_APU)
282 return 0;
283
284 return attr->mode;
285}
286
287static const struct attribute_group amdgpu_board_attrs_group = {
288 .attrs = amdgpu_board_attrs,
289 .is_visible = amdgpu_board_attrs_is_visible
290};
291
292static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev);
293
294
295/**
296 * amdgpu_device_supports_px - Is the device a dGPU with ATPX power control
297 *
298 * @dev: drm_device pointer
299 *
300 * Returns true if the device is a dGPU with ATPX power control,
301 * otherwise return false.
302 */
303bool amdgpu_device_supports_px(struct drm_device *dev)
304{
305 struct amdgpu_device *adev = drm_to_adev(dev);
306
307 if ((adev->flags & AMD_IS_PX) && !amdgpu_is_atpx_hybrid())
308 return true;
309 return false;
310}
311
312/**
313 * amdgpu_device_supports_boco - Is the device a dGPU with ACPI power resources
314 *
315 * @dev: drm_device pointer
316 *
317 * Returns true if the device is a dGPU with ACPI power control,
318 * otherwise return false.
319 */
320bool amdgpu_device_supports_boco(struct drm_device *dev)
321{
322 struct amdgpu_device *adev = drm_to_adev(dev);
323
324 if (adev->has_pr3 ||
325 ((adev->flags & AMD_IS_PX) && amdgpu_is_atpx_hybrid()))
326 return true;
327 return false;
328}
329
330/**
331 * amdgpu_device_supports_baco - Does the device support BACO
332 *
333 * @dev: drm_device pointer
334 *
335 * Returns true if the device supporte BACO,
336 * otherwise return false.
337 */
338bool amdgpu_device_supports_baco(struct drm_device *dev)
339{
340 struct amdgpu_device *adev = drm_to_adev(dev);
341
342 return amdgpu_asic_supports_baco(adev);
343}
344
345/**
346 * amdgpu_device_supports_smart_shift - Is the device dGPU with
347 * smart shift support
348 *
349 * @dev: drm_device pointer
350 *
351 * Returns true if the device is a dGPU with Smart Shift support,
352 * otherwise returns false.
353 */
354bool amdgpu_device_supports_smart_shift(struct drm_device *dev)
355{
356 return (amdgpu_device_supports_boco(dev) &&
357 amdgpu_acpi_is_power_shift_control_supported());
358}
359
360/*
361 * VRAM access helper functions
362 */
363
364/**
365 * amdgpu_device_mm_access - access vram by MM_INDEX/MM_DATA
366 *
367 * @adev: amdgpu_device pointer
368 * @pos: offset of the buffer in vram
369 * @buf: virtual address of the buffer in system memory
370 * @size: read/write size, sizeof(@buf) must > @size
371 * @write: true - write to vram, otherwise - read from vram
372 */
373void amdgpu_device_mm_access(struct amdgpu_device *adev, loff_t pos,
374 void *buf, size_t size, bool write)
375{
376 unsigned long flags;
377 uint32_t hi = ~0, tmp = 0;
378 uint32_t *data = buf;
379 uint64_t last;
380 int idx;
381
382 if (!drm_dev_enter(adev_to_drm(adev), &idx))
383 return;
384
385 BUG_ON(!IS_ALIGNED(pos, 4) || !IS_ALIGNED(size, 4));
386
387 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
388 for (last = pos + size; pos < last; pos += 4) {
389 tmp = pos >> 31;
390
391 WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)pos) | 0x80000000);
392 if (tmp != hi) {
393 WREG32_NO_KIQ(mmMM_INDEX_HI, tmp);
394 hi = tmp;
395 }
396 if (write)
397 WREG32_NO_KIQ(mmMM_DATA, *data++);
398 else
399 *data++ = RREG32_NO_KIQ(mmMM_DATA);
400 }
401
402 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
403 drm_dev_exit(idx);
404}
405
406/**
407 * amdgpu_device_aper_access - access vram by vram aperature
408 *
409 * @adev: amdgpu_device pointer
410 * @pos: offset of the buffer in vram
411 * @buf: virtual address of the buffer in system memory
412 * @size: read/write size, sizeof(@buf) must > @size
413 * @write: true - write to vram, otherwise - read from vram
414 *
415 * The return value means how many bytes have been transferred.
416 */
417size_t amdgpu_device_aper_access(struct amdgpu_device *adev, loff_t pos,
418 void *buf, size_t size, bool write)
419{
420#ifdef CONFIG_64BIT
421 void __iomem *addr;
422 size_t count = 0;
423 uint64_t last;
424
425 if (!adev->mman.aper_base_kaddr)
426 return 0;
427
428 last = min(pos + size, adev->gmc.visible_vram_size);
429 if (last > pos) {
430 addr = adev->mman.aper_base_kaddr + pos;
431 count = last - pos;
432
433 if (write) {
434 memcpy_toio(addr, buf, count);
435 /* Make sure HDP write cache flush happens without any reordering
436 * after the system memory contents are sent over PCIe device
437 */
438 mb();
439 amdgpu_device_flush_hdp(adev, NULL);
440 } else {
441 amdgpu_device_invalidate_hdp(adev, NULL);
442 /* Make sure HDP read cache is invalidated before issuing a read
443 * to the PCIe device
444 */
445 mb();
446 memcpy_fromio(buf, addr, count);
447 }
448
449 }
450
451 return count;
452#else
453 return 0;
454#endif
455}
456
457/**
458 * amdgpu_device_vram_access - read/write a buffer in vram
459 *
460 * @adev: amdgpu_device pointer
461 * @pos: offset of the buffer in vram
462 * @buf: virtual address of the buffer in system memory
463 * @size: read/write size, sizeof(@buf) must > @size
464 * @write: true - write to vram, otherwise - read from vram
465 */
466void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos,
467 void *buf, size_t size, bool write)
468{
469 size_t count;
470
471 /* try to using vram apreature to access vram first */
472 count = amdgpu_device_aper_access(adev, pos, buf, size, write);
473 size -= count;
474 if (size) {
475 /* using MM to access rest vram */
476 pos += count;
477 buf += count;
478 amdgpu_device_mm_access(adev, pos, buf, size, write);
479 }
480}
481
482/*
483 * register access helper functions.
484 */
485
486/* Check if hw access should be skipped because of hotplug or device error */
487bool amdgpu_device_skip_hw_access(struct amdgpu_device *adev)
488{
489 if (adev->no_hw_access)
490 return true;
491
492#ifdef CONFIG_LOCKDEP
493 /*
494 * This is a bit complicated to understand, so worth a comment. What we assert
495 * here is that the GPU reset is not running on another thread in parallel.
496 *
497 * For this we trylock the read side of the reset semaphore, if that succeeds
498 * we know that the reset is not running in paralell.
499 *
500 * If the trylock fails we assert that we are either already holding the read
501 * side of the lock or are the reset thread itself and hold the write side of
502 * the lock.
503 */
504 if (in_task()) {
505 if (down_read_trylock(&adev->reset_domain->sem))
506 up_read(&adev->reset_domain->sem);
507 else
508 lockdep_assert_held(&adev->reset_domain->sem);
509 }
510#endif
511 return false;
512}
513
514/**
515 * amdgpu_device_rreg - read a memory mapped IO or indirect register
516 *
517 * @adev: amdgpu_device pointer
518 * @reg: dword aligned register offset
519 * @acc_flags: access flags which require special behavior
520 *
521 * Returns the 32 bit value from the offset specified.
522 */
523uint32_t amdgpu_device_rreg(struct amdgpu_device *adev,
524 uint32_t reg, uint32_t acc_flags)
525{
526 uint32_t ret;
527
528 if (amdgpu_device_skip_hw_access(adev))
529 return 0;
530
531 if ((reg * 4) < adev->rmmio_size) {
532 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) &&
533 amdgpu_sriov_runtime(adev) &&
534 down_read_trylock(&adev->reset_domain->sem)) {
535 ret = amdgpu_kiq_rreg(adev, reg, 0);
536 up_read(&adev->reset_domain->sem);
537 } else {
538 ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
539 }
540 } else {
541 ret = adev->pcie_rreg(adev, reg * 4);
542 }
543
544 trace_amdgpu_device_rreg(adev->pdev->device, reg, ret);
545
546 return ret;
547}
548
549/*
550 * MMIO register read with bytes helper functions
551 * @offset:bytes offset from MMIO start
552 */
553
554/**
555 * amdgpu_mm_rreg8 - read a memory mapped IO register
556 *
557 * @adev: amdgpu_device pointer
558 * @offset: byte aligned register offset
559 *
560 * Returns the 8 bit value from the offset specified.
561 */
562uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset)
563{
564 if (amdgpu_device_skip_hw_access(adev))
565 return 0;
566
567 if (offset < adev->rmmio_size)
568 return (readb(adev->rmmio + offset));
569 BUG();
570}
571
572
573/**
574 * amdgpu_device_xcc_rreg - read a memory mapped IO or indirect register with specific XCC
575 *
576 * @adev: amdgpu_device pointer
577 * @reg: dword aligned register offset
578 * @acc_flags: access flags which require special behavior
579 * @xcc_id: xcc accelerated compute core id
580 *
581 * Returns the 32 bit value from the offset specified.
582 */
583uint32_t amdgpu_device_xcc_rreg(struct amdgpu_device *adev,
584 uint32_t reg, uint32_t acc_flags,
585 uint32_t xcc_id)
586{
587 uint32_t ret, rlcg_flag;
588
589 if (amdgpu_device_skip_hw_access(adev))
590 return 0;
591
592 if ((reg * 4) < adev->rmmio_size) {
593 if (amdgpu_sriov_vf(adev) &&
594 !amdgpu_sriov_runtime(adev) &&
595 adev->gfx.rlc.rlcg_reg_access_supported &&
596 amdgpu_virt_get_rlcg_reg_access_flag(adev, acc_flags,
597 GC_HWIP, false,
598 &rlcg_flag)) {
599 ret = amdgpu_virt_rlcg_reg_rw(adev, reg, 0, rlcg_flag, xcc_id);
600 } else if (!(acc_flags & AMDGPU_REGS_NO_KIQ) &&
601 amdgpu_sriov_runtime(adev) &&
602 down_read_trylock(&adev->reset_domain->sem)) {
603 ret = amdgpu_kiq_rreg(adev, reg, xcc_id);
604 up_read(&adev->reset_domain->sem);
605 } else {
606 ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
607 }
608 } else {
609 ret = adev->pcie_rreg(adev, reg * 4);
610 }
611
612 return ret;
613}
614
615/*
616 * MMIO register write with bytes helper functions
617 * @offset:bytes offset from MMIO start
618 * @value: the value want to be written to the register
619 */
620
621/**
622 * amdgpu_mm_wreg8 - read a memory mapped IO register
623 *
624 * @adev: amdgpu_device pointer
625 * @offset: byte aligned register offset
626 * @value: 8 bit value to write
627 *
628 * Writes the value specified to the offset specified.
629 */
630void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value)
631{
632 if (amdgpu_device_skip_hw_access(adev))
633 return;
634
635 if (offset < adev->rmmio_size)
636 writeb(value, adev->rmmio + offset);
637 else
638 BUG();
639}
640
641/**
642 * amdgpu_device_wreg - write to a memory mapped IO or indirect register
643 *
644 * @adev: amdgpu_device pointer
645 * @reg: dword aligned register offset
646 * @v: 32 bit value to write to the register
647 * @acc_flags: access flags which require special behavior
648 *
649 * Writes the value specified to the offset specified.
650 */
651void amdgpu_device_wreg(struct amdgpu_device *adev,
652 uint32_t reg, uint32_t v,
653 uint32_t acc_flags)
654{
655 if (amdgpu_device_skip_hw_access(adev))
656 return;
657
658 if ((reg * 4) < adev->rmmio_size) {
659 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) &&
660 amdgpu_sriov_runtime(adev) &&
661 down_read_trylock(&adev->reset_domain->sem)) {
662 amdgpu_kiq_wreg(adev, reg, v, 0);
663 up_read(&adev->reset_domain->sem);
664 } else {
665 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
666 }
667 } else {
668 adev->pcie_wreg(adev, reg * 4, v);
669 }
670
671 trace_amdgpu_device_wreg(adev->pdev->device, reg, v);
672}
673
674/**
675 * amdgpu_mm_wreg_mmio_rlc - write register either with direct/indirect mmio or with RLC path if in range
676 *
677 * @adev: amdgpu_device pointer
678 * @reg: mmio/rlc register
679 * @v: value to write
680 * @xcc_id: xcc accelerated compute core id
681 *
682 * this function is invoked only for the debugfs register access
683 */
684void amdgpu_mm_wreg_mmio_rlc(struct amdgpu_device *adev,
685 uint32_t reg, uint32_t v,
686 uint32_t xcc_id)
687{
688 if (amdgpu_device_skip_hw_access(adev))
689 return;
690
691 if (amdgpu_sriov_fullaccess(adev) &&
692 adev->gfx.rlc.funcs &&
693 adev->gfx.rlc.funcs->is_rlcg_access_range) {
694 if (adev->gfx.rlc.funcs->is_rlcg_access_range(adev, reg))
695 return amdgpu_sriov_wreg(adev, reg, v, 0, 0, xcc_id);
696 } else if ((reg * 4) >= adev->rmmio_size) {
697 adev->pcie_wreg(adev, reg * 4, v);
698 } else {
699 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
700 }
701}
702
703/**
704 * amdgpu_device_xcc_wreg - write to a memory mapped IO or indirect register with specific XCC
705 *
706 * @adev: amdgpu_device pointer
707 * @reg: dword aligned register offset
708 * @v: 32 bit value to write to the register
709 * @acc_flags: access flags which require special behavior
710 * @xcc_id: xcc accelerated compute core id
711 *
712 * Writes the value specified to the offset specified.
713 */
714void amdgpu_device_xcc_wreg(struct amdgpu_device *adev,
715 uint32_t reg, uint32_t v,
716 uint32_t acc_flags, uint32_t xcc_id)
717{
718 uint32_t rlcg_flag;
719
720 if (amdgpu_device_skip_hw_access(adev))
721 return;
722
723 if ((reg * 4) < adev->rmmio_size) {
724 if (amdgpu_sriov_vf(adev) &&
725 !amdgpu_sriov_runtime(adev) &&
726 adev->gfx.rlc.rlcg_reg_access_supported &&
727 amdgpu_virt_get_rlcg_reg_access_flag(adev, acc_flags,
728 GC_HWIP, true,
729 &rlcg_flag)) {
730 amdgpu_virt_rlcg_reg_rw(adev, reg, v, rlcg_flag, xcc_id);
731 } else if (!(acc_flags & AMDGPU_REGS_NO_KIQ) &&
732 amdgpu_sriov_runtime(adev) &&
733 down_read_trylock(&adev->reset_domain->sem)) {
734 amdgpu_kiq_wreg(adev, reg, v, xcc_id);
735 up_read(&adev->reset_domain->sem);
736 } else {
737 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
738 }
739 } else {
740 adev->pcie_wreg(adev, reg * 4, v);
741 }
742}
743
744/**
745 * amdgpu_device_indirect_rreg - read an indirect register
746 *
747 * @adev: amdgpu_device pointer
748 * @reg_addr: indirect register address to read from
749 *
750 * Returns the value of indirect register @reg_addr
751 */
752u32 amdgpu_device_indirect_rreg(struct amdgpu_device *adev,
753 u32 reg_addr)
754{
755 unsigned long flags, pcie_index, pcie_data;
756 void __iomem *pcie_index_offset;
757 void __iomem *pcie_data_offset;
758 u32 r;
759
760 pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev);
761 pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev);
762
763 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
764 pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
765 pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
766
767 writel(reg_addr, pcie_index_offset);
768 readl(pcie_index_offset);
769 r = readl(pcie_data_offset);
770 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
771
772 return r;
773}
774
775u32 amdgpu_device_indirect_rreg_ext(struct amdgpu_device *adev,
776 u64 reg_addr)
777{
778 unsigned long flags, pcie_index, pcie_index_hi, pcie_data;
779 u32 r;
780 void __iomem *pcie_index_offset;
781 void __iomem *pcie_index_hi_offset;
782 void __iomem *pcie_data_offset;
783
784 pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev);
785 pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev);
786 if ((reg_addr >> 32) && (adev->nbio.funcs->get_pcie_index_hi_offset))
787 pcie_index_hi = adev->nbio.funcs->get_pcie_index_hi_offset(adev);
788 else
789 pcie_index_hi = 0;
790
791 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
792 pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
793 pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
794 if (pcie_index_hi != 0)
795 pcie_index_hi_offset = (void __iomem *)adev->rmmio +
796 pcie_index_hi * 4;
797
798 writel(reg_addr, pcie_index_offset);
799 readl(pcie_index_offset);
800 if (pcie_index_hi != 0) {
801 writel((reg_addr >> 32) & 0xff, pcie_index_hi_offset);
802 readl(pcie_index_hi_offset);
803 }
804 r = readl(pcie_data_offset);
805
806 /* clear the high bits */
807 if (pcie_index_hi != 0) {
808 writel(0, pcie_index_hi_offset);
809 readl(pcie_index_hi_offset);
810 }
811
812 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
813
814 return r;
815}
816
817/**
818 * amdgpu_device_indirect_rreg64 - read a 64bits indirect register
819 *
820 * @adev: amdgpu_device pointer
821 * @reg_addr: indirect register address to read from
822 *
823 * Returns the value of indirect register @reg_addr
824 */
825u64 amdgpu_device_indirect_rreg64(struct amdgpu_device *adev,
826 u32 reg_addr)
827{
828 unsigned long flags, pcie_index, pcie_data;
829 void __iomem *pcie_index_offset;
830 void __iomem *pcie_data_offset;
831 u64 r;
832
833 pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev);
834 pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev);
835
836 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
837 pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
838 pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
839
840 /* read low 32 bits */
841 writel(reg_addr, pcie_index_offset);
842 readl(pcie_index_offset);
843 r = readl(pcie_data_offset);
844 /* read high 32 bits */
845 writel(reg_addr + 4, pcie_index_offset);
846 readl(pcie_index_offset);
847 r |= ((u64)readl(pcie_data_offset) << 32);
848 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
849
850 return r;
851}
852
853u64 amdgpu_device_indirect_rreg64_ext(struct amdgpu_device *adev,
854 u64 reg_addr)
855{
856 unsigned long flags, pcie_index, pcie_data;
857 unsigned long pcie_index_hi = 0;
858 void __iomem *pcie_index_offset;
859 void __iomem *pcie_index_hi_offset;
860 void __iomem *pcie_data_offset;
861 u64 r;
862
863 pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev);
864 pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev);
865 if ((reg_addr >> 32) && (adev->nbio.funcs->get_pcie_index_hi_offset))
866 pcie_index_hi = adev->nbio.funcs->get_pcie_index_hi_offset(adev);
867
868 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
869 pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
870 pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
871 if (pcie_index_hi != 0)
872 pcie_index_hi_offset = (void __iomem *)adev->rmmio +
873 pcie_index_hi * 4;
874
875 /* read low 32 bits */
876 writel(reg_addr, pcie_index_offset);
877 readl(pcie_index_offset);
878 if (pcie_index_hi != 0) {
879 writel((reg_addr >> 32) & 0xff, pcie_index_hi_offset);
880 readl(pcie_index_hi_offset);
881 }
882 r = readl(pcie_data_offset);
883 /* read high 32 bits */
884 writel(reg_addr + 4, pcie_index_offset);
885 readl(pcie_index_offset);
886 if (pcie_index_hi != 0) {
887 writel((reg_addr >> 32) & 0xff, pcie_index_hi_offset);
888 readl(pcie_index_hi_offset);
889 }
890 r |= ((u64)readl(pcie_data_offset) << 32);
891
892 /* clear the high bits */
893 if (pcie_index_hi != 0) {
894 writel(0, pcie_index_hi_offset);
895 readl(pcie_index_hi_offset);
896 }
897
898 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
899
900 return r;
901}
902
903/**
904 * amdgpu_device_indirect_wreg - write an indirect register address
905 *
906 * @adev: amdgpu_device pointer
907 * @reg_addr: indirect register offset
908 * @reg_data: indirect register data
909 *
910 */
911void amdgpu_device_indirect_wreg(struct amdgpu_device *adev,
912 u32 reg_addr, u32 reg_data)
913{
914 unsigned long flags, pcie_index, pcie_data;
915 void __iomem *pcie_index_offset;
916 void __iomem *pcie_data_offset;
917
918 pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev);
919 pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev);
920
921 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
922 pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
923 pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
924
925 writel(reg_addr, pcie_index_offset);
926 readl(pcie_index_offset);
927 writel(reg_data, pcie_data_offset);
928 readl(pcie_data_offset);
929 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
930}
931
932void amdgpu_device_indirect_wreg_ext(struct amdgpu_device *adev,
933 u64 reg_addr, u32 reg_data)
934{
935 unsigned long flags, pcie_index, pcie_index_hi, pcie_data;
936 void __iomem *pcie_index_offset;
937 void __iomem *pcie_index_hi_offset;
938 void __iomem *pcie_data_offset;
939
940 pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev);
941 pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev);
942 if ((reg_addr >> 32) && (adev->nbio.funcs->get_pcie_index_hi_offset))
943 pcie_index_hi = adev->nbio.funcs->get_pcie_index_hi_offset(adev);
944 else
945 pcie_index_hi = 0;
946
947 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
948 pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
949 pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
950 if (pcie_index_hi != 0)
951 pcie_index_hi_offset = (void __iomem *)adev->rmmio +
952 pcie_index_hi * 4;
953
954 writel(reg_addr, pcie_index_offset);
955 readl(pcie_index_offset);
956 if (pcie_index_hi != 0) {
957 writel((reg_addr >> 32) & 0xff, pcie_index_hi_offset);
958 readl(pcie_index_hi_offset);
959 }
960 writel(reg_data, pcie_data_offset);
961 readl(pcie_data_offset);
962
963 /* clear the high bits */
964 if (pcie_index_hi != 0) {
965 writel(0, pcie_index_hi_offset);
966 readl(pcie_index_hi_offset);
967 }
968
969 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
970}
971
972/**
973 * amdgpu_device_indirect_wreg64 - write a 64bits indirect register address
974 *
975 * @adev: amdgpu_device pointer
976 * @reg_addr: indirect register offset
977 * @reg_data: indirect register data
978 *
979 */
980void amdgpu_device_indirect_wreg64(struct amdgpu_device *adev,
981 u32 reg_addr, u64 reg_data)
982{
983 unsigned long flags, pcie_index, pcie_data;
984 void __iomem *pcie_index_offset;
985 void __iomem *pcie_data_offset;
986
987 pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev);
988 pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev);
989
990 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
991 pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
992 pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
993
994 /* write low 32 bits */
995 writel(reg_addr, pcie_index_offset);
996 readl(pcie_index_offset);
997 writel((u32)(reg_data & 0xffffffffULL), pcie_data_offset);
998 readl(pcie_data_offset);
999 /* write high 32 bits */
1000 writel(reg_addr + 4, pcie_index_offset);
1001 readl(pcie_index_offset);
1002 writel((u32)(reg_data >> 32), pcie_data_offset);
1003 readl(pcie_data_offset);
1004 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
1005}
1006
1007void amdgpu_device_indirect_wreg64_ext(struct amdgpu_device *adev,
1008 u64 reg_addr, u64 reg_data)
1009{
1010 unsigned long flags, pcie_index, pcie_data;
1011 unsigned long pcie_index_hi = 0;
1012 void __iomem *pcie_index_offset;
1013 void __iomem *pcie_index_hi_offset;
1014 void __iomem *pcie_data_offset;
1015
1016 pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev);
1017 pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev);
1018 if ((reg_addr >> 32) && (adev->nbio.funcs->get_pcie_index_hi_offset))
1019 pcie_index_hi = adev->nbio.funcs->get_pcie_index_hi_offset(adev);
1020
1021 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
1022 pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
1023 pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
1024 if (pcie_index_hi != 0)
1025 pcie_index_hi_offset = (void __iomem *)adev->rmmio +
1026 pcie_index_hi * 4;
1027
1028 /* write low 32 bits */
1029 writel(reg_addr, pcie_index_offset);
1030 readl(pcie_index_offset);
1031 if (pcie_index_hi != 0) {
1032 writel((reg_addr >> 32) & 0xff, pcie_index_hi_offset);
1033 readl(pcie_index_hi_offset);
1034 }
1035 writel((u32)(reg_data & 0xffffffffULL), pcie_data_offset);
1036 readl(pcie_data_offset);
1037 /* write high 32 bits */
1038 writel(reg_addr + 4, pcie_index_offset);
1039 readl(pcie_index_offset);
1040 if (pcie_index_hi != 0) {
1041 writel((reg_addr >> 32) & 0xff, pcie_index_hi_offset);
1042 readl(pcie_index_hi_offset);
1043 }
1044 writel((u32)(reg_data >> 32), pcie_data_offset);
1045 readl(pcie_data_offset);
1046
1047 /* clear the high bits */
1048 if (pcie_index_hi != 0) {
1049 writel(0, pcie_index_hi_offset);
1050 readl(pcie_index_hi_offset);
1051 }
1052
1053 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
1054}
1055
1056/**
1057 * amdgpu_device_get_rev_id - query device rev_id
1058 *
1059 * @adev: amdgpu_device pointer
1060 *
1061 * Return device rev_id
1062 */
1063u32 amdgpu_device_get_rev_id(struct amdgpu_device *adev)
1064{
1065 return adev->nbio.funcs->get_rev_id(adev);
1066}
1067
1068/**
1069 * amdgpu_invalid_rreg - dummy reg read function
1070 *
1071 * @adev: amdgpu_device pointer
1072 * @reg: offset of register
1073 *
1074 * Dummy register read function. Used for register blocks
1075 * that certain asics don't have (all asics).
1076 * Returns the value in the register.
1077 */
1078static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
1079{
1080 DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
1081 BUG();
1082 return 0;
1083}
1084
1085static uint32_t amdgpu_invalid_rreg_ext(struct amdgpu_device *adev, uint64_t reg)
1086{
1087 DRM_ERROR("Invalid callback to read register 0x%llX\n", reg);
1088 BUG();
1089 return 0;
1090}
1091
1092/**
1093 * amdgpu_invalid_wreg - dummy reg write function
1094 *
1095 * @adev: amdgpu_device pointer
1096 * @reg: offset of register
1097 * @v: value to write to the register
1098 *
1099 * Dummy register read function. Used for register blocks
1100 * that certain asics don't have (all asics).
1101 */
1102static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
1103{
1104 DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
1105 reg, v);
1106 BUG();
1107}
1108
1109static void amdgpu_invalid_wreg_ext(struct amdgpu_device *adev, uint64_t reg, uint32_t v)
1110{
1111 DRM_ERROR("Invalid callback to write register 0x%llX with 0x%08X\n",
1112 reg, v);
1113 BUG();
1114}
1115
1116/**
1117 * amdgpu_invalid_rreg64 - dummy 64 bit reg read function
1118 *
1119 * @adev: amdgpu_device pointer
1120 * @reg: offset of register
1121 *
1122 * Dummy register read function. Used for register blocks
1123 * that certain asics don't have (all asics).
1124 * Returns the value in the register.
1125 */
1126static uint64_t amdgpu_invalid_rreg64(struct amdgpu_device *adev, uint32_t reg)
1127{
1128 DRM_ERROR("Invalid callback to read 64 bit register 0x%04X\n", reg);
1129 BUG();
1130 return 0;
1131}
1132
1133static uint64_t amdgpu_invalid_rreg64_ext(struct amdgpu_device *adev, uint64_t reg)
1134{
1135 DRM_ERROR("Invalid callback to read register 0x%llX\n", reg);
1136 BUG();
1137 return 0;
1138}
1139
1140/**
1141 * amdgpu_invalid_wreg64 - dummy reg write function
1142 *
1143 * @adev: amdgpu_device pointer
1144 * @reg: offset of register
1145 * @v: value to write to the register
1146 *
1147 * Dummy register read function. Used for register blocks
1148 * that certain asics don't have (all asics).
1149 */
1150static void amdgpu_invalid_wreg64(struct amdgpu_device *adev, uint32_t reg, uint64_t v)
1151{
1152 DRM_ERROR("Invalid callback to write 64 bit register 0x%04X with 0x%08llX\n",
1153 reg, v);
1154 BUG();
1155}
1156
1157static void amdgpu_invalid_wreg64_ext(struct amdgpu_device *adev, uint64_t reg, uint64_t v)
1158{
1159 DRM_ERROR("Invalid callback to write 64 bit register 0x%llX with 0x%08llX\n",
1160 reg, v);
1161 BUG();
1162}
1163
1164/**
1165 * amdgpu_block_invalid_rreg - dummy reg read function
1166 *
1167 * @adev: amdgpu_device pointer
1168 * @block: offset of instance
1169 * @reg: offset of register
1170 *
1171 * Dummy register read function. Used for register blocks
1172 * that certain asics don't have (all asics).
1173 * Returns the value in the register.
1174 */
1175static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
1176 uint32_t block, uint32_t reg)
1177{
1178 DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
1179 reg, block);
1180 BUG();
1181 return 0;
1182}
1183
1184/**
1185 * amdgpu_block_invalid_wreg - dummy reg write function
1186 *
1187 * @adev: amdgpu_device pointer
1188 * @block: offset of instance
1189 * @reg: offset of register
1190 * @v: value to write to the register
1191 *
1192 * Dummy register read function. Used for register blocks
1193 * that certain asics don't have (all asics).
1194 */
1195static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
1196 uint32_t block,
1197 uint32_t reg, uint32_t v)
1198{
1199 DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
1200 reg, block, v);
1201 BUG();
1202}
1203
1204/**
1205 * amdgpu_device_asic_init - Wrapper for atom asic_init
1206 *
1207 * @adev: amdgpu_device pointer
1208 *
1209 * Does any asic specific work and then calls atom asic init.
1210 */
1211static int amdgpu_device_asic_init(struct amdgpu_device *adev)
1212{
1213 int ret;
1214
1215 amdgpu_asic_pre_asic_init(adev);
1216
1217 if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3) ||
1218 amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(11, 0, 0)) {
1219 amdgpu_psp_wait_for_bootloader(adev);
1220 ret = amdgpu_atomfirmware_asic_init(adev, true);
1221 /* TODO: check the return val and stop device initialization if boot fails */
1222 amdgpu_psp_query_boot_status(adev);
1223 return ret;
1224 } else {
1225 return amdgpu_atom_asic_init(adev->mode_info.atom_context);
1226 }
1227
1228 return 0;
1229}
1230
1231/**
1232 * amdgpu_device_mem_scratch_init - allocate the VRAM scratch page
1233 *
1234 * @adev: amdgpu_device pointer
1235 *
1236 * Allocates a scratch page of VRAM for use by various things in the
1237 * driver.
1238 */
1239static int amdgpu_device_mem_scratch_init(struct amdgpu_device *adev)
1240{
1241 return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE, PAGE_SIZE,
1242 AMDGPU_GEM_DOMAIN_VRAM |
1243 AMDGPU_GEM_DOMAIN_GTT,
1244 &adev->mem_scratch.robj,
1245 &adev->mem_scratch.gpu_addr,
1246 (void **)&adev->mem_scratch.ptr);
1247}
1248
1249/**
1250 * amdgpu_device_mem_scratch_fini - Free the VRAM scratch page
1251 *
1252 * @adev: amdgpu_device pointer
1253 *
1254 * Frees the VRAM scratch page.
1255 */
1256static void amdgpu_device_mem_scratch_fini(struct amdgpu_device *adev)
1257{
1258 amdgpu_bo_free_kernel(&adev->mem_scratch.robj, NULL, NULL);
1259}
1260
1261/**
1262 * amdgpu_device_program_register_sequence - program an array of registers.
1263 *
1264 * @adev: amdgpu_device pointer
1265 * @registers: pointer to the register array
1266 * @array_size: size of the register array
1267 *
1268 * Programs an array or registers with and or masks.
1269 * This is a helper for setting golden registers.
1270 */
1271void amdgpu_device_program_register_sequence(struct amdgpu_device *adev,
1272 const u32 *registers,
1273 const u32 array_size)
1274{
1275 u32 tmp, reg, and_mask, or_mask;
1276 int i;
1277
1278 if (array_size % 3)
1279 return;
1280
1281 for (i = 0; i < array_size; i += 3) {
1282 reg = registers[i + 0];
1283 and_mask = registers[i + 1];
1284 or_mask = registers[i + 2];
1285
1286 if (and_mask == 0xffffffff) {
1287 tmp = or_mask;
1288 } else {
1289 tmp = RREG32(reg);
1290 tmp &= ~and_mask;
1291 if (adev->family >= AMDGPU_FAMILY_AI)
1292 tmp |= (or_mask & and_mask);
1293 else
1294 tmp |= or_mask;
1295 }
1296 WREG32(reg, tmp);
1297 }
1298}
1299
1300/**
1301 * amdgpu_device_pci_config_reset - reset the GPU
1302 *
1303 * @adev: amdgpu_device pointer
1304 *
1305 * Resets the GPU using the pci config reset sequence.
1306 * Only applicable to asics prior to vega10.
1307 */
1308void amdgpu_device_pci_config_reset(struct amdgpu_device *adev)
1309{
1310 pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
1311}
1312
1313/**
1314 * amdgpu_device_pci_reset - reset the GPU using generic PCI means
1315 *
1316 * @adev: amdgpu_device pointer
1317 *
1318 * Resets the GPU using generic pci reset interfaces (FLR, SBR, etc.).
1319 */
1320int amdgpu_device_pci_reset(struct amdgpu_device *adev)
1321{
1322 return pci_reset_function(adev->pdev);
1323}
1324
1325/*
1326 * amdgpu_device_wb_*()
1327 * Writeback is the method by which the GPU updates special pages in memory
1328 * with the status of certain GPU events (fences, ring pointers,etc.).
1329 */
1330
1331/**
1332 * amdgpu_device_wb_fini - Disable Writeback and free memory
1333 *
1334 * @adev: amdgpu_device pointer
1335 *
1336 * Disables Writeback and frees the Writeback memory (all asics).
1337 * Used at driver shutdown.
1338 */
1339static void amdgpu_device_wb_fini(struct amdgpu_device *adev)
1340{
1341 if (adev->wb.wb_obj) {
1342 amdgpu_bo_free_kernel(&adev->wb.wb_obj,
1343 &adev->wb.gpu_addr,
1344 (void **)&adev->wb.wb);
1345 adev->wb.wb_obj = NULL;
1346 }
1347}
1348
1349/**
1350 * amdgpu_device_wb_init - Init Writeback driver info and allocate memory
1351 *
1352 * @adev: amdgpu_device pointer
1353 *
1354 * Initializes writeback and allocates writeback memory (all asics).
1355 * Used at driver startup.
1356 * Returns 0 on success or an -error on failure.
1357 */
1358static int amdgpu_device_wb_init(struct amdgpu_device *adev)
1359{
1360 int r;
1361
1362 if (adev->wb.wb_obj == NULL) {
1363 /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */
1364 r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8,
1365 PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
1366 &adev->wb.wb_obj, &adev->wb.gpu_addr,
1367 (void **)&adev->wb.wb);
1368 if (r) {
1369 dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
1370 return r;
1371 }
1372
1373 adev->wb.num_wb = AMDGPU_MAX_WB;
1374 memset(&adev->wb.used, 0, sizeof(adev->wb.used));
1375
1376 /* clear wb memory */
1377 memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8);
1378 }
1379
1380 return 0;
1381}
1382
1383/**
1384 * amdgpu_device_wb_get - Allocate a wb entry
1385 *
1386 * @adev: amdgpu_device pointer
1387 * @wb: wb index
1388 *
1389 * Allocate a wb slot for use by the driver (all asics).
1390 * Returns 0 on success or -EINVAL on failure.
1391 */
1392int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb)
1393{
1394 unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
1395
1396 if (offset < adev->wb.num_wb) {
1397 __set_bit(offset, adev->wb.used);
1398 *wb = offset << 3; /* convert to dw offset */
1399 return 0;
1400 } else {
1401 return -EINVAL;
1402 }
1403}
1404
1405/**
1406 * amdgpu_device_wb_free - Free a wb entry
1407 *
1408 * @adev: amdgpu_device pointer
1409 * @wb: wb index
1410 *
1411 * Free a wb slot allocated for use by the driver (all asics)
1412 */
1413void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb)
1414{
1415 wb >>= 3;
1416 if (wb < adev->wb.num_wb)
1417 __clear_bit(wb, adev->wb.used);
1418}
1419
1420/**
1421 * amdgpu_device_resize_fb_bar - try to resize FB BAR
1422 *
1423 * @adev: amdgpu_device pointer
1424 *
1425 * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not
1426 * to fail, but if any of the BARs is not accessible after the size we abort
1427 * driver loading by returning -ENODEV.
1428 */
1429int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev)
1430{
1431 int rbar_size = pci_rebar_bytes_to_size(adev->gmc.real_vram_size);
1432 struct pci_bus *root;
1433 struct resource *res;
1434 unsigned int i;
1435 u16 cmd;
1436 int r;
1437
1438 if (!IS_ENABLED(CONFIG_PHYS_ADDR_T_64BIT))
1439 return 0;
1440
1441 /* Bypass for VF */
1442 if (amdgpu_sriov_vf(adev))
1443 return 0;
1444
1445 /* skip if the bios has already enabled large BAR */
1446 if (adev->gmc.real_vram_size &&
1447 (pci_resource_len(adev->pdev, 0) >= adev->gmc.real_vram_size))
1448 return 0;
1449
1450 /* Check if the root BUS has 64bit memory resources */
1451 root = adev->pdev->bus;
1452 while (root->parent)
1453 root = root->parent;
1454
1455 pci_bus_for_each_resource(root, res, i) {
1456 if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) &&
1457 res->start > 0x100000000ull)
1458 break;
1459 }
1460
1461 /* Trying to resize is pointless without a root hub window above 4GB */
1462 if (!res)
1463 return 0;
1464
1465 /* Limit the BAR size to what is available */
1466 rbar_size = min(fls(pci_rebar_get_possible_sizes(adev->pdev, 0)) - 1,
1467 rbar_size);
1468
1469 /* Disable memory decoding while we change the BAR addresses and size */
1470 pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd);
1471 pci_write_config_word(adev->pdev, PCI_COMMAND,
1472 cmd & ~PCI_COMMAND_MEMORY);
1473
1474 /* Free the VRAM and doorbell BAR, we most likely need to move both. */
1475 amdgpu_doorbell_fini(adev);
1476 if (adev->asic_type >= CHIP_BONAIRE)
1477 pci_release_resource(adev->pdev, 2);
1478
1479 pci_release_resource(adev->pdev, 0);
1480
1481 r = pci_resize_resource(adev->pdev, 0, rbar_size);
1482 if (r == -ENOSPC)
1483 DRM_INFO("Not enough PCI address space for a large BAR.");
1484 else if (r && r != -ENOTSUPP)
1485 DRM_ERROR("Problem resizing BAR0 (%d).", r);
1486
1487 pci_assign_unassigned_bus_resources(adev->pdev->bus);
1488
1489 /* When the doorbell or fb BAR isn't available we have no chance of
1490 * using the device.
1491 */
1492 r = amdgpu_doorbell_init(adev);
1493 if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET))
1494 return -ENODEV;
1495
1496 pci_write_config_word(adev->pdev, PCI_COMMAND, cmd);
1497
1498 return 0;
1499}
1500
1501static bool amdgpu_device_read_bios(struct amdgpu_device *adev)
1502{
1503 if (hweight32(adev->aid_mask) && (adev->flags & AMD_IS_APU))
1504 return false;
1505
1506 return true;
1507}
1508
1509/*
1510 * GPU helpers function.
1511 */
1512/**
1513 * amdgpu_device_need_post - check if the hw need post or not
1514 *
1515 * @adev: amdgpu_device pointer
1516 *
1517 * Check if the asic has been initialized (all asics) at driver startup
1518 * or post is needed if hw reset is performed.
1519 * Returns true if need or false if not.
1520 */
1521bool amdgpu_device_need_post(struct amdgpu_device *adev)
1522{
1523 uint32_t reg;
1524
1525 if (amdgpu_sriov_vf(adev))
1526 return false;
1527
1528 if (!amdgpu_device_read_bios(adev))
1529 return false;
1530
1531 if (amdgpu_passthrough(adev)) {
1532 /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
1533 * some old smc fw still need driver do vPost otherwise gpu hang, while
1534 * those smc fw version above 22.15 doesn't have this flaw, so we force
1535 * vpost executed for smc version below 22.15
1536 */
1537 if (adev->asic_type == CHIP_FIJI) {
1538 int err;
1539 uint32_t fw_ver;
1540
1541 err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
1542 /* force vPost if error occured */
1543 if (err)
1544 return true;
1545
1546 fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
1547 release_firmware(adev->pm.fw);
1548 if (fw_ver < 0x00160e00)
1549 return true;
1550 }
1551 }
1552
1553 /* Don't post if we need to reset whole hive on init */
1554 if (adev->gmc.xgmi.pending_reset)
1555 return false;
1556
1557 if (adev->has_hw_reset) {
1558 adev->has_hw_reset = false;
1559 return true;
1560 }
1561
1562 /* bios scratch used on CIK+ */
1563 if (adev->asic_type >= CHIP_BONAIRE)
1564 return amdgpu_atombios_scratch_need_asic_init(adev);
1565
1566 /* check MEM_SIZE for older asics */
1567 reg = amdgpu_asic_get_config_memsize(adev);
1568
1569 if ((reg != 0) && (reg != 0xffffffff))
1570 return false;
1571
1572 return true;
1573}
1574
1575/*
1576 * Check whether seamless boot is supported.
1577 *
1578 * So far we only support seamless boot on DCE 3.0 or later.
1579 * If users report that it works on older ASICS as well, we may
1580 * loosen this.
1581 */
1582bool amdgpu_device_seamless_boot_supported(struct amdgpu_device *adev)
1583{
1584 switch (amdgpu_seamless) {
1585 case -1:
1586 break;
1587 case 1:
1588 return true;
1589 case 0:
1590 return false;
1591 default:
1592 DRM_ERROR("Invalid value for amdgpu.seamless: %d\n",
1593 amdgpu_seamless);
1594 return false;
1595 }
1596
1597 if (!(adev->flags & AMD_IS_APU))
1598 return false;
1599
1600 if (adev->mman.keep_stolen_vga_memory)
1601 return false;
1602
1603 return amdgpu_ip_version(adev, DCE_HWIP, 0) >= IP_VERSION(3, 0, 0);
1604}
1605
1606/*
1607 * Intel hosts such as Rocket Lake, Alder Lake, Raptor Lake and Sapphire Rapids
1608 * don't support dynamic speed switching. Until we have confirmation from Intel
1609 * that a specific host supports it, it's safer that we keep it disabled for all.
1610 *
1611 * https://edc.intel.com/content/www/us/en/design/products/platforms/details/raptor-lake-s/13th-generation-core-processors-datasheet-volume-1-of-2/005/pci-express-support/
1612 * https://gitlab.freedesktop.org/drm/amd/-/issues/2663
1613 */
1614static bool amdgpu_device_pcie_dynamic_switching_supported(struct amdgpu_device *adev)
1615{
1616#if IS_ENABLED(CONFIG_X86)
1617 struct cpuinfo_x86 *c = &cpu_data(0);
1618
1619 /* eGPU change speeds based on USB4 fabric conditions */
1620 if (dev_is_removable(adev->dev))
1621 return true;
1622
1623 if (c->x86_vendor == X86_VENDOR_INTEL)
1624 return false;
1625#endif
1626 return true;
1627}
1628
1629/**
1630 * amdgpu_device_should_use_aspm - check if the device should program ASPM
1631 *
1632 * @adev: amdgpu_device pointer
1633 *
1634 * Confirm whether the module parameter and pcie bridge agree that ASPM should
1635 * be set for this device.
1636 *
1637 * Returns true if it should be used or false if not.
1638 */
1639bool amdgpu_device_should_use_aspm(struct amdgpu_device *adev)
1640{
1641 switch (amdgpu_aspm) {
1642 case -1:
1643 break;
1644 case 0:
1645 return false;
1646 case 1:
1647 return true;
1648 default:
1649 return false;
1650 }
1651 if (adev->flags & AMD_IS_APU)
1652 return false;
1653 if (!(adev->pm.pp_feature & PP_PCIE_DPM_MASK))
1654 return false;
1655 return pcie_aspm_enabled(adev->pdev);
1656}
1657
1658/* if we get transitioned to only one device, take VGA back */
1659/**
1660 * amdgpu_device_vga_set_decode - enable/disable vga decode
1661 *
1662 * @pdev: PCI device pointer
1663 * @state: enable/disable vga decode
1664 *
1665 * Enable/disable vga decode (all asics).
1666 * Returns VGA resource flags.
1667 */
1668static unsigned int amdgpu_device_vga_set_decode(struct pci_dev *pdev,
1669 bool state)
1670{
1671 struct amdgpu_device *adev = drm_to_adev(pci_get_drvdata(pdev));
1672
1673 amdgpu_asic_set_vga_state(adev, state);
1674 if (state)
1675 return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
1676 VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
1677 else
1678 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
1679}
1680
1681/**
1682 * amdgpu_device_check_block_size - validate the vm block size
1683 *
1684 * @adev: amdgpu_device pointer
1685 *
1686 * Validates the vm block size specified via module parameter.
1687 * The vm block size defines number of bits in page table versus page directory,
1688 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
1689 * page table and the remaining bits are in the page directory.
1690 */
1691static void amdgpu_device_check_block_size(struct amdgpu_device *adev)
1692{
1693 /* defines number of bits in page table versus page directory,
1694 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
1695 * page table and the remaining bits are in the page directory
1696 */
1697 if (amdgpu_vm_block_size == -1)
1698 return;
1699
1700 if (amdgpu_vm_block_size < 9) {
1701 dev_warn(adev->dev, "VM page table size (%d) too small\n",
1702 amdgpu_vm_block_size);
1703 amdgpu_vm_block_size = -1;
1704 }
1705}
1706
1707/**
1708 * amdgpu_device_check_vm_size - validate the vm size
1709 *
1710 * @adev: amdgpu_device pointer
1711 *
1712 * Validates the vm size in GB specified via module parameter.
1713 * The VM size is the size of the GPU virtual memory space in GB.
1714 */
1715static void amdgpu_device_check_vm_size(struct amdgpu_device *adev)
1716{
1717 /* no need to check the default value */
1718 if (amdgpu_vm_size == -1)
1719 return;
1720
1721 if (amdgpu_vm_size < 1) {
1722 dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
1723 amdgpu_vm_size);
1724 amdgpu_vm_size = -1;
1725 }
1726}
1727
1728static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev)
1729{
1730 struct sysinfo si;
1731 bool is_os_64 = (sizeof(void *) == 8);
1732 uint64_t total_memory;
1733 uint64_t dram_size_seven_GB = 0x1B8000000;
1734 uint64_t dram_size_three_GB = 0xB8000000;
1735
1736 if (amdgpu_smu_memory_pool_size == 0)
1737 return;
1738
1739 if (!is_os_64) {
1740 DRM_WARN("Not 64-bit OS, feature not supported\n");
1741 goto def_value;
1742 }
1743 si_meminfo(&si);
1744 total_memory = (uint64_t)si.totalram * si.mem_unit;
1745
1746 if ((amdgpu_smu_memory_pool_size == 1) ||
1747 (amdgpu_smu_memory_pool_size == 2)) {
1748 if (total_memory < dram_size_three_GB)
1749 goto def_value1;
1750 } else if ((amdgpu_smu_memory_pool_size == 4) ||
1751 (amdgpu_smu_memory_pool_size == 8)) {
1752 if (total_memory < dram_size_seven_GB)
1753 goto def_value1;
1754 } else {
1755 DRM_WARN("Smu memory pool size not supported\n");
1756 goto def_value;
1757 }
1758 adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28;
1759
1760 return;
1761
1762def_value1:
1763 DRM_WARN("No enough system memory\n");
1764def_value:
1765 adev->pm.smu_prv_buffer_size = 0;
1766}
1767
1768static int amdgpu_device_init_apu_flags(struct amdgpu_device *adev)
1769{
1770 if (!(adev->flags & AMD_IS_APU) ||
1771 adev->asic_type < CHIP_RAVEN)
1772 return 0;
1773
1774 switch (adev->asic_type) {
1775 case CHIP_RAVEN:
1776 if (adev->pdev->device == 0x15dd)
1777 adev->apu_flags |= AMD_APU_IS_RAVEN;
1778 if (adev->pdev->device == 0x15d8)
1779 adev->apu_flags |= AMD_APU_IS_PICASSO;
1780 break;
1781 case CHIP_RENOIR:
1782 if ((adev->pdev->device == 0x1636) ||
1783 (adev->pdev->device == 0x164c))
1784 adev->apu_flags |= AMD_APU_IS_RENOIR;
1785 else
1786 adev->apu_flags |= AMD_APU_IS_GREEN_SARDINE;
1787 break;
1788 case CHIP_VANGOGH:
1789 adev->apu_flags |= AMD_APU_IS_VANGOGH;
1790 break;
1791 case CHIP_YELLOW_CARP:
1792 break;
1793 case CHIP_CYAN_SKILLFISH:
1794 if ((adev->pdev->device == 0x13FE) ||
1795 (adev->pdev->device == 0x143F))
1796 adev->apu_flags |= AMD_APU_IS_CYAN_SKILLFISH2;
1797 break;
1798 default:
1799 break;
1800 }
1801
1802 return 0;
1803}
1804
1805/**
1806 * amdgpu_device_check_arguments - validate module params
1807 *
1808 * @adev: amdgpu_device pointer
1809 *
1810 * Validates certain module parameters and updates
1811 * the associated values used by the driver (all asics).
1812 */
1813static int amdgpu_device_check_arguments(struct amdgpu_device *adev)
1814{
1815 if (amdgpu_sched_jobs < 4) {
1816 dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
1817 amdgpu_sched_jobs);
1818 amdgpu_sched_jobs = 4;
1819 } else if (!is_power_of_2(amdgpu_sched_jobs)) {
1820 dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
1821 amdgpu_sched_jobs);
1822 amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
1823 }
1824
1825 if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) {
1826 /* gart size must be greater or equal to 32M */
1827 dev_warn(adev->dev, "gart size (%d) too small\n",
1828 amdgpu_gart_size);
1829 amdgpu_gart_size = -1;
1830 }
1831
1832 if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) {
1833 /* gtt size must be greater or equal to 32M */
1834 dev_warn(adev->dev, "gtt size (%d) too small\n",
1835 amdgpu_gtt_size);
1836 amdgpu_gtt_size = -1;
1837 }
1838
1839 /* valid range is between 4 and 9 inclusive */
1840 if (amdgpu_vm_fragment_size != -1 &&
1841 (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) {
1842 dev_warn(adev->dev, "valid range is between 4 and 9\n");
1843 amdgpu_vm_fragment_size = -1;
1844 }
1845
1846 if (amdgpu_sched_hw_submission < 2) {
1847 dev_warn(adev->dev, "sched hw submission jobs (%d) must be at least 2\n",
1848 amdgpu_sched_hw_submission);
1849 amdgpu_sched_hw_submission = 2;
1850 } else if (!is_power_of_2(amdgpu_sched_hw_submission)) {
1851 dev_warn(adev->dev, "sched hw submission jobs (%d) must be a power of 2\n",
1852 amdgpu_sched_hw_submission);
1853 amdgpu_sched_hw_submission = roundup_pow_of_two(amdgpu_sched_hw_submission);
1854 }
1855
1856 if (amdgpu_reset_method < -1 || amdgpu_reset_method > 4) {
1857 dev_warn(adev->dev, "invalid option for reset method, reverting to default\n");
1858 amdgpu_reset_method = -1;
1859 }
1860
1861 amdgpu_device_check_smu_prv_buffer_size(adev);
1862
1863 amdgpu_device_check_vm_size(adev);
1864
1865 amdgpu_device_check_block_size(adev);
1866
1867 adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type);
1868
1869 return 0;
1870}
1871
1872/**
1873 * amdgpu_switcheroo_set_state - set switcheroo state
1874 *
1875 * @pdev: pci dev pointer
1876 * @state: vga_switcheroo state
1877 *
1878 * Callback for the switcheroo driver. Suspends or resumes
1879 * the asics before or after it is powered up using ACPI methods.
1880 */
1881static void amdgpu_switcheroo_set_state(struct pci_dev *pdev,
1882 enum vga_switcheroo_state state)
1883{
1884 struct drm_device *dev = pci_get_drvdata(pdev);
1885 int r;
1886
1887 if (amdgpu_device_supports_px(dev) && state == VGA_SWITCHEROO_OFF)
1888 return;
1889
1890 if (state == VGA_SWITCHEROO_ON) {
1891 pr_info("switched on\n");
1892 /* don't suspend or resume card normally */
1893 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1894
1895 pci_set_power_state(pdev, PCI_D0);
1896 amdgpu_device_load_pci_state(pdev);
1897 r = pci_enable_device(pdev);
1898 if (r)
1899 DRM_WARN("pci_enable_device failed (%d)\n", r);
1900 amdgpu_device_resume(dev, true);
1901
1902 dev->switch_power_state = DRM_SWITCH_POWER_ON;
1903 } else {
1904 pr_info("switched off\n");
1905 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1906 amdgpu_device_prepare(dev);
1907 amdgpu_device_suspend(dev, true);
1908 amdgpu_device_cache_pci_state(pdev);
1909 /* Shut down the device */
1910 pci_disable_device(pdev);
1911 pci_set_power_state(pdev, PCI_D3cold);
1912 dev->switch_power_state = DRM_SWITCH_POWER_OFF;
1913 }
1914}
1915
1916/**
1917 * amdgpu_switcheroo_can_switch - see if switcheroo state can change
1918 *
1919 * @pdev: pci dev pointer
1920 *
1921 * Callback for the switcheroo driver. Check of the switcheroo
1922 * state can be changed.
1923 * Returns true if the state can be changed, false if not.
1924 */
1925static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
1926{
1927 struct drm_device *dev = pci_get_drvdata(pdev);
1928
1929 /*
1930 * FIXME: open_count is protected by drm_global_mutex but that would lead to
1931 * locking inversion with the driver load path. And the access here is
1932 * completely racy anyway. So don't bother with locking for now.
1933 */
1934 return atomic_read(&dev->open_count) == 0;
1935}
1936
1937static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
1938 .set_gpu_state = amdgpu_switcheroo_set_state,
1939 .reprobe = NULL,
1940 .can_switch = amdgpu_switcheroo_can_switch,
1941};
1942
1943/**
1944 * amdgpu_device_ip_set_clockgating_state - set the CG state
1945 *
1946 * @dev: amdgpu_device pointer
1947 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1948 * @state: clockgating state (gate or ungate)
1949 *
1950 * Sets the requested clockgating state for all instances of
1951 * the hardware IP specified.
1952 * Returns the error code from the last instance.
1953 */
1954int amdgpu_device_ip_set_clockgating_state(void *dev,
1955 enum amd_ip_block_type block_type,
1956 enum amd_clockgating_state state)
1957{
1958 struct amdgpu_device *adev = dev;
1959 int i, r = 0;
1960
1961 for (i = 0; i < adev->num_ip_blocks; i++) {
1962 if (!adev->ip_blocks[i].status.valid)
1963 continue;
1964 if (adev->ip_blocks[i].version->type != block_type)
1965 continue;
1966 if (!adev->ip_blocks[i].version->funcs->set_clockgating_state)
1967 continue;
1968 r = adev->ip_blocks[i].version->funcs->set_clockgating_state(
1969 (void *)adev, state);
1970 if (r)
1971 DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",
1972 adev->ip_blocks[i].version->funcs->name, r);
1973 }
1974 return r;
1975}
1976
1977/**
1978 * amdgpu_device_ip_set_powergating_state - set the PG state
1979 *
1980 * @dev: amdgpu_device pointer
1981 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1982 * @state: powergating state (gate or ungate)
1983 *
1984 * Sets the requested powergating state for all instances of
1985 * the hardware IP specified.
1986 * Returns the error code from the last instance.
1987 */
1988int amdgpu_device_ip_set_powergating_state(void *dev,
1989 enum amd_ip_block_type block_type,
1990 enum amd_powergating_state state)
1991{
1992 struct amdgpu_device *adev = dev;
1993 int i, r = 0;
1994
1995 for (i = 0; i < adev->num_ip_blocks; i++) {
1996 if (!adev->ip_blocks[i].status.valid)
1997 continue;
1998 if (adev->ip_blocks[i].version->type != block_type)
1999 continue;
2000 if (!adev->ip_blocks[i].version->funcs->set_powergating_state)
2001 continue;
2002 r = adev->ip_blocks[i].version->funcs->set_powergating_state(
2003 (void *)adev, state);
2004 if (r)
2005 DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",
2006 adev->ip_blocks[i].version->funcs->name, r);
2007 }
2008 return r;
2009}
2010
2011/**
2012 * amdgpu_device_ip_get_clockgating_state - get the CG state
2013 *
2014 * @adev: amdgpu_device pointer
2015 * @flags: clockgating feature flags
2016 *
2017 * Walks the list of IPs on the device and updates the clockgating
2018 * flags for each IP.
2019 * Updates @flags with the feature flags for each hardware IP where
2020 * clockgating is enabled.
2021 */
2022void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev,
2023 u64 *flags)
2024{
2025 int i;
2026
2027 for (i = 0; i < adev->num_ip_blocks; i++) {
2028 if (!adev->ip_blocks[i].status.valid)
2029 continue;
2030 if (adev->ip_blocks[i].version->funcs->get_clockgating_state)
2031 adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags);
2032 }
2033}
2034
2035/**
2036 * amdgpu_device_ip_wait_for_idle - wait for idle
2037 *
2038 * @adev: amdgpu_device pointer
2039 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
2040 *
2041 * Waits for the request hardware IP to be idle.
2042 * Returns 0 for success or a negative error code on failure.
2043 */
2044int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev,
2045 enum amd_ip_block_type block_type)
2046{
2047 int i, r;
2048
2049 for (i = 0; i < adev->num_ip_blocks; i++) {
2050 if (!adev->ip_blocks[i].status.valid)
2051 continue;
2052 if (adev->ip_blocks[i].version->type == block_type) {
2053 r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
2054 if (r)
2055 return r;
2056 break;
2057 }
2058 }
2059 return 0;
2060
2061}
2062
2063/**
2064 * amdgpu_device_ip_is_idle - is the hardware IP idle
2065 *
2066 * @adev: amdgpu_device pointer
2067 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
2068 *
2069 * Check if the hardware IP is idle or not.
2070 * Returns true if it the IP is idle, false if not.
2071 */
2072bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev,
2073 enum amd_ip_block_type block_type)
2074{
2075 int i;
2076
2077 for (i = 0; i < adev->num_ip_blocks; i++) {
2078 if (!adev->ip_blocks[i].status.valid)
2079 continue;
2080 if (adev->ip_blocks[i].version->type == block_type)
2081 return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
2082 }
2083 return true;
2084
2085}
2086
2087/**
2088 * amdgpu_device_ip_get_ip_block - get a hw IP pointer
2089 *
2090 * @adev: amdgpu_device pointer
2091 * @type: Type of hardware IP (SMU, GFX, UVD, etc.)
2092 *
2093 * Returns a pointer to the hardware IP block structure
2094 * if it exists for the asic, otherwise NULL.
2095 */
2096struct amdgpu_ip_block *
2097amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,
2098 enum amd_ip_block_type type)
2099{
2100 int i;
2101
2102 for (i = 0; i < adev->num_ip_blocks; i++)
2103 if (adev->ip_blocks[i].version->type == type)
2104 return &adev->ip_blocks[i];
2105
2106 return NULL;
2107}
2108
2109/**
2110 * amdgpu_device_ip_block_version_cmp
2111 *
2112 * @adev: amdgpu_device pointer
2113 * @type: enum amd_ip_block_type
2114 * @major: major version
2115 * @minor: minor version
2116 *
2117 * return 0 if equal or greater
2118 * return 1 if smaller or the ip_block doesn't exist
2119 */
2120int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev,
2121 enum amd_ip_block_type type,
2122 u32 major, u32 minor)
2123{
2124 struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type);
2125
2126 if (ip_block && ((ip_block->version->major > major) ||
2127 ((ip_block->version->major == major) &&
2128 (ip_block->version->minor >= minor))))
2129 return 0;
2130
2131 return 1;
2132}
2133
2134/**
2135 * amdgpu_device_ip_block_add
2136 *
2137 * @adev: amdgpu_device pointer
2138 * @ip_block_version: pointer to the IP to add
2139 *
2140 * Adds the IP block driver information to the collection of IPs
2141 * on the asic.
2142 */
2143int amdgpu_device_ip_block_add(struct amdgpu_device *adev,
2144 const struct amdgpu_ip_block_version *ip_block_version)
2145{
2146 if (!ip_block_version)
2147 return -EINVAL;
2148
2149 switch (ip_block_version->type) {
2150 case AMD_IP_BLOCK_TYPE_VCN:
2151 if (adev->harvest_ip_mask & AMD_HARVEST_IP_VCN_MASK)
2152 return 0;
2153 break;
2154 case AMD_IP_BLOCK_TYPE_JPEG:
2155 if (adev->harvest_ip_mask & AMD_HARVEST_IP_JPEG_MASK)
2156 return 0;
2157 break;
2158 default:
2159 break;
2160 }
2161
2162 DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks,
2163 ip_block_version->funcs->name);
2164
2165 adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
2166
2167 return 0;
2168}
2169
2170/**
2171 * amdgpu_device_enable_virtual_display - enable virtual display feature
2172 *
2173 * @adev: amdgpu_device pointer
2174 *
2175 * Enabled the virtual display feature if the user has enabled it via
2176 * the module parameter virtual_display. This feature provides a virtual
2177 * display hardware on headless boards or in virtualized environments.
2178 * This function parses and validates the configuration string specified by
2179 * the user and configues the virtual display configuration (number of
2180 * virtual connectors, crtcs, etc.) specified.
2181 */
2182static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
2183{
2184 adev->enable_virtual_display = false;
2185
2186 if (amdgpu_virtual_display) {
2187 const char *pci_address_name = pci_name(adev->pdev);
2188 char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
2189
2190 pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
2191 pciaddstr_tmp = pciaddstr;
2192 while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
2193 pciaddname = strsep(&pciaddname_tmp, ",");
2194 if (!strcmp("all", pciaddname)
2195 || !strcmp(pci_address_name, pciaddname)) {
2196 long num_crtc;
2197 int res = -1;
2198
2199 adev->enable_virtual_display = true;
2200
2201 if (pciaddname_tmp)
2202 res = kstrtol(pciaddname_tmp, 10,
2203 &num_crtc);
2204
2205 if (!res) {
2206 if (num_crtc < 1)
2207 num_crtc = 1;
2208 if (num_crtc > 6)
2209 num_crtc = 6;
2210 adev->mode_info.num_crtc = num_crtc;
2211 } else {
2212 adev->mode_info.num_crtc = 1;
2213 }
2214 break;
2215 }
2216 }
2217
2218 DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
2219 amdgpu_virtual_display, pci_address_name,
2220 adev->enable_virtual_display, adev->mode_info.num_crtc);
2221
2222 kfree(pciaddstr);
2223 }
2224}
2225
2226void amdgpu_device_set_sriov_virtual_display(struct amdgpu_device *adev)
2227{
2228 if (amdgpu_sriov_vf(adev) && !adev->enable_virtual_display) {
2229 adev->mode_info.num_crtc = 1;
2230 adev->enable_virtual_display = true;
2231 DRM_INFO("virtual_display:%d, num_crtc:%d\n",
2232 adev->enable_virtual_display, adev->mode_info.num_crtc);
2233 }
2234}
2235
2236/**
2237 * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware
2238 *
2239 * @adev: amdgpu_device pointer
2240 *
2241 * Parses the asic configuration parameters specified in the gpu info
2242 * firmware and makes them availale to the driver for use in configuring
2243 * the asic.
2244 * Returns 0 on success, -EINVAL on failure.
2245 */
2246static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev)
2247{
2248 const char *chip_name;
2249 char fw_name[40];
2250 int err;
2251 const struct gpu_info_firmware_header_v1_0 *hdr;
2252
2253 adev->firmware.gpu_info_fw = NULL;
2254
2255 if (adev->mman.discovery_bin)
2256 return 0;
2257
2258 switch (adev->asic_type) {
2259 default:
2260 return 0;
2261 case CHIP_VEGA10:
2262 chip_name = "vega10";
2263 break;
2264 case CHIP_VEGA12:
2265 chip_name = "vega12";
2266 break;
2267 case CHIP_RAVEN:
2268 if (adev->apu_flags & AMD_APU_IS_RAVEN2)
2269 chip_name = "raven2";
2270 else if (adev->apu_flags & AMD_APU_IS_PICASSO)
2271 chip_name = "picasso";
2272 else
2273 chip_name = "raven";
2274 break;
2275 case CHIP_ARCTURUS:
2276 chip_name = "arcturus";
2277 break;
2278 case CHIP_NAVI12:
2279 chip_name = "navi12";
2280 break;
2281 }
2282
2283 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name);
2284 err = amdgpu_ucode_request(adev, &adev->firmware.gpu_info_fw, fw_name);
2285 if (err) {
2286 dev_err(adev->dev,
2287 "Failed to get gpu_info firmware \"%s\"\n",
2288 fw_name);
2289 goto out;
2290 }
2291
2292 hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data;
2293 amdgpu_ucode_print_gpu_info_hdr(&hdr->header);
2294
2295 switch (hdr->version_major) {
2296 case 1:
2297 {
2298 const struct gpu_info_firmware_v1_0 *gpu_info_fw =
2299 (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data +
2300 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
2301
2302 /*
2303 * Should be droped when DAL no longer needs it.
2304 */
2305 if (adev->asic_type == CHIP_NAVI12)
2306 goto parse_soc_bounding_box;
2307
2308 adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se);
2309 adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh);
2310 adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se);
2311 adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se);
2312 adev->gfx.config.max_texture_channel_caches =
2313 le32_to_cpu(gpu_info_fw->gc_num_tccs);
2314 adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs);
2315 adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds);
2316 adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth);
2317 adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth);
2318 adev->gfx.config.double_offchip_lds_buf =
2319 le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer);
2320 adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size);
2321 adev->gfx.cu_info.max_waves_per_simd =
2322 le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd);
2323 adev->gfx.cu_info.max_scratch_slots_per_cu =
2324 le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu);
2325 adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size);
2326 if (hdr->version_minor >= 1) {
2327 const struct gpu_info_firmware_v1_1 *gpu_info_fw =
2328 (const struct gpu_info_firmware_v1_1 *)(adev->firmware.gpu_info_fw->data +
2329 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
2330 adev->gfx.config.num_sc_per_sh =
2331 le32_to_cpu(gpu_info_fw->num_sc_per_sh);
2332 adev->gfx.config.num_packer_per_sc =
2333 le32_to_cpu(gpu_info_fw->num_packer_per_sc);
2334 }
2335
2336parse_soc_bounding_box:
2337 /*
2338 * soc bounding box info is not integrated in disocovery table,
2339 * we always need to parse it from gpu info firmware if needed.
2340 */
2341 if (hdr->version_minor == 2) {
2342 const struct gpu_info_firmware_v1_2 *gpu_info_fw =
2343 (const struct gpu_info_firmware_v1_2 *)(adev->firmware.gpu_info_fw->data +
2344 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
2345 adev->dm.soc_bounding_box = &gpu_info_fw->soc_bounding_box;
2346 }
2347 break;
2348 }
2349 default:
2350 dev_err(adev->dev,
2351 "Unsupported gpu_info table %d\n", hdr->header.ucode_version);
2352 err = -EINVAL;
2353 goto out;
2354 }
2355out:
2356 return err;
2357}
2358
2359/**
2360 * amdgpu_device_ip_early_init - run early init for hardware IPs
2361 *
2362 * @adev: amdgpu_device pointer
2363 *
2364 * Early initialization pass for hardware IPs. The hardware IPs that make
2365 * up each asic are discovered each IP's early_init callback is run. This
2366 * is the first stage in initializing the asic.
2367 * Returns 0 on success, negative error code on failure.
2368 */
2369static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
2370{
2371 struct pci_dev *parent;
2372 int i, r;
2373 bool total;
2374
2375 amdgpu_device_enable_virtual_display(adev);
2376
2377 if (amdgpu_sriov_vf(adev)) {
2378 r = amdgpu_virt_request_full_gpu(adev, true);
2379 if (r)
2380 return r;
2381 }
2382
2383 switch (adev->asic_type) {
2384#ifdef CONFIG_DRM_AMDGPU_SI
2385 case CHIP_VERDE:
2386 case CHIP_TAHITI:
2387 case CHIP_PITCAIRN:
2388 case CHIP_OLAND:
2389 case CHIP_HAINAN:
2390 adev->family = AMDGPU_FAMILY_SI;
2391 r = si_set_ip_blocks(adev);
2392 if (r)
2393 return r;
2394 break;
2395#endif
2396#ifdef CONFIG_DRM_AMDGPU_CIK
2397 case CHIP_BONAIRE:
2398 case CHIP_HAWAII:
2399 case CHIP_KAVERI:
2400 case CHIP_KABINI:
2401 case CHIP_MULLINS:
2402 if (adev->flags & AMD_IS_APU)
2403 adev->family = AMDGPU_FAMILY_KV;
2404 else
2405 adev->family = AMDGPU_FAMILY_CI;
2406
2407 r = cik_set_ip_blocks(adev);
2408 if (r)
2409 return r;
2410 break;
2411#endif
2412 case CHIP_TOPAZ:
2413 case CHIP_TONGA:
2414 case CHIP_FIJI:
2415 case CHIP_POLARIS10:
2416 case CHIP_POLARIS11:
2417 case CHIP_POLARIS12:
2418 case CHIP_VEGAM:
2419 case CHIP_CARRIZO:
2420 case CHIP_STONEY:
2421 if (adev->flags & AMD_IS_APU)
2422 adev->family = AMDGPU_FAMILY_CZ;
2423 else
2424 adev->family = AMDGPU_FAMILY_VI;
2425
2426 r = vi_set_ip_blocks(adev);
2427 if (r)
2428 return r;
2429 break;
2430 default:
2431 r = amdgpu_discovery_set_ip_blocks(adev);
2432 if (r)
2433 return r;
2434 break;
2435 }
2436
2437 if (amdgpu_has_atpx() &&
2438 (amdgpu_is_atpx_hybrid() ||
2439 amdgpu_has_atpx_dgpu_power_cntl()) &&
2440 ((adev->flags & AMD_IS_APU) == 0) &&
2441 !dev_is_removable(&adev->pdev->dev))
2442 adev->flags |= AMD_IS_PX;
2443
2444 if (!(adev->flags & AMD_IS_APU)) {
2445 parent = pcie_find_root_port(adev->pdev);
2446 adev->has_pr3 = parent ? pci_pr3_present(parent) : false;
2447 }
2448
2449
2450 adev->pm.pp_feature = amdgpu_pp_feature_mask;
2451 if (amdgpu_sriov_vf(adev) || sched_policy == KFD_SCHED_POLICY_NO_HWS)
2452 adev->pm.pp_feature &= ~PP_GFXOFF_MASK;
2453 if (amdgpu_sriov_vf(adev) && adev->asic_type == CHIP_SIENNA_CICHLID)
2454 adev->pm.pp_feature &= ~PP_OVERDRIVE_MASK;
2455 if (!amdgpu_device_pcie_dynamic_switching_supported(adev))
2456 adev->pm.pp_feature &= ~PP_PCIE_DPM_MASK;
2457
2458 total = true;
2459 for (i = 0; i < adev->num_ip_blocks; i++) {
2460 if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
2461 DRM_WARN("disabled ip block: %d <%s>\n",
2462 i, adev->ip_blocks[i].version->funcs->name);
2463 adev->ip_blocks[i].status.valid = false;
2464 } else {
2465 if (adev->ip_blocks[i].version->funcs->early_init) {
2466 r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
2467 if (r == -ENOENT) {
2468 adev->ip_blocks[i].status.valid = false;
2469 } else if (r) {
2470 DRM_ERROR("early_init of IP block <%s> failed %d\n",
2471 adev->ip_blocks[i].version->funcs->name, r);
2472 total = false;
2473 } else {
2474 adev->ip_blocks[i].status.valid = true;
2475 }
2476 } else {
2477 adev->ip_blocks[i].status.valid = true;
2478 }
2479 }
2480 /* get the vbios after the asic_funcs are set up */
2481 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) {
2482 r = amdgpu_device_parse_gpu_info_fw(adev);
2483 if (r)
2484 return r;
2485
2486 /* Read BIOS */
2487 if (amdgpu_device_read_bios(adev)) {
2488 if (!amdgpu_get_bios(adev))
2489 return -EINVAL;
2490
2491 r = amdgpu_atombios_init(adev);
2492 if (r) {
2493 dev_err(adev->dev, "amdgpu_atombios_init failed\n");
2494 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0);
2495 return r;
2496 }
2497 }
2498
2499 /*get pf2vf msg info at it's earliest time*/
2500 if (amdgpu_sriov_vf(adev))
2501 amdgpu_virt_init_data_exchange(adev);
2502
2503 }
2504 }
2505 if (!total)
2506 return -ENODEV;
2507
2508 amdgpu_amdkfd_device_probe(adev);
2509 adev->cg_flags &= amdgpu_cg_mask;
2510 adev->pg_flags &= amdgpu_pg_mask;
2511
2512 return 0;
2513}
2514
2515static int amdgpu_device_ip_hw_init_phase1(struct amdgpu_device *adev)
2516{
2517 int i, r;
2518
2519 for (i = 0; i < adev->num_ip_blocks; i++) {
2520 if (!adev->ip_blocks[i].status.sw)
2521 continue;
2522 if (adev->ip_blocks[i].status.hw)
2523 continue;
2524 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2525 (amdgpu_sriov_vf(adev) && (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)) ||
2526 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
2527 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
2528 if (r) {
2529 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
2530 adev->ip_blocks[i].version->funcs->name, r);
2531 return r;
2532 }
2533 adev->ip_blocks[i].status.hw = true;
2534 }
2535 }
2536
2537 return 0;
2538}
2539
2540static int amdgpu_device_ip_hw_init_phase2(struct amdgpu_device *adev)
2541{
2542 int i, r;
2543
2544 for (i = 0; i < adev->num_ip_blocks; i++) {
2545 if (!adev->ip_blocks[i].status.sw)
2546 continue;
2547 if (adev->ip_blocks[i].status.hw)
2548 continue;
2549 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
2550 if (r) {
2551 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
2552 adev->ip_blocks[i].version->funcs->name, r);
2553 return r;
2554 }
2555 adev->ip_blocks[i].status.hw = true;
2556 }
2557
2558 return 0;
2559}
2560
2561static int amdgpu_device_fw_loading(struct amdgpu_device *adev)
2562{
2563 int r = 0;
2564 int i;
2565 uint32_t smu_version;
2566
2567 if (adev->asic_type >= CHIP_VEGA10) {
2568 for (i = 0; i < adev->num_ip_blocks; i++) {
2569 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_PSP)
2570 continue;
2571
2572 if (!adev->ip_blocks[i].status.sw)
2573 continue;
2574
2575 /* no need to do the fw loading again if already done*/
2576 if (adev->ip_blocks[i].status.hw == true)
2577 break;
2578
2579 if (amdgpu_in_reset(adev) || adev->in_suspend) {
2580 r = adev->ip_blocks[i].version->funcs->resume(adev);
2581 if (r) {
2582 DRM_ERROR("resume of IP block <%s> failed %d\n",
2583 adev->ip_blocks[i].version->funcs->name, r);
2584 return r;
2585 }
2586 } else {
2587 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
2588 if (r) {
2589 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
2590 adev->ip_blocks[i].version->funcs->name, r);
2591 return r;
2592 }
2593 }
2594
2595 adev->ip_blocks[i].status.hw = true;
2596 break;
2597 }
2598 }
2599
2600 if (!amdgpu_sriov_vf(adev) || adev->asic_type == CHIP_TONGA)
2601 r = amdgpu_pm_load_smu_firmware(adev, &smu_version);
2602
2603 return r;
2604}
2605
2606static int amdgpu_device_init_schedulers(struct amdgpu_device *adev)
2607{
2608 long timeout;
2609 int r, i;
2610
2611 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
2612 struct amdgpu_ring *ring = adev->rings[i];
2613
2614 /* No need to setup the GPU scheduler for rings that don't need it */
2615 if (!ring || ring->no_scheduler)
2616 continue;
2617
2618 switch (ring->funcs->type) {
2619 case AMDGPU_RING_TYPE_GFX:
2620 timeout = adev->gfx_timeout;
2621 break;
2622 case AMDGPU_RING_TYPE_COMPUTE:
2623 timeout = adev->compute_timeout;
2624 break;
2625 case AMDGPU_RING_TYPE_SDMA:
2626 timeout = adev->sdma_timeout;
2627 break;
2628 default:
2629 timeout = adev->video_timeout;
2630 break;
2631 }
2632
2633 r = drm_sched_init(&ring->sched, &amdgpu_sched_ops, NULL,
2634 DRM_SCHED_PRIORITY_COUNT,
2635 ring->num_hw_submission, 0,
2636 timeout, adev->reset_domain->wq,
2637 ring->sched_score, ring->name,
2638 adev->dev);
2639 if (r) {
2640 DRM_ERROR("Failed to create scheduler on ring %s.\n",
2641 ring->name);
2642 return r;
2643 }
2644 r = amdgpu_uvd_entity_init(adev, ring);
2645 if (r) {
2646 DRM_ERROR("Failed to create UVD scheduling entity on ring %s.\n",
2647 ring->name);
2648 return r;
2649 }
2650 r = amdgpu_vce_entity_init(adev, ring);
2651 if (r) {
2652 DRM_ERROR("Failed to create VCE scheduling entity on ring %s.\n",
2653 ring->name);
2654 return r;
2655 }
2656 }
2657
2658 amdgpu_xcp_update_partition_sched_list(adev);
2659
2660 return 0;
2661}
2662
2663
2664/**
2665 * amdgpu_device_ip_init - run init for hardware IPs
2666 *
2667 * @adev: amdgpu_device pointer
2668 *
2669 * Main initialization pass for hardware IPs. The list of all the hardware
2670 * IPs that make up the asic is walked and the sw_init and hw_init callbacks
2671 * are run. sw_init initializes the software state associated with each IP
2672 * and hw_init initializes the hardware associated with each IP.
2673 * Returns 0 on success, negative error code on failure.
2674 */
2675static int amdgpu_device_ip_init(struct amdgpu_device *adev)
2676{
2677 int i, r;
2678
2679 r = amdgpu_ras_init(adev);
2680 if (r)
2681 return r;
2682
2683 for (i = 0; i < adev->num_ip_blocks; i++) {
2684 if (!adev->ip_blocks[i].status.valid)
2685 continue;
2686 r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
2687 if (r) {
2688 DRM_ERROR("sw_init of IP block <%s> failed %d\n",
2689 adev->ip_blocks[i].version->funcs->name, r);
2690 goto init_failed;
2691 }
2692 adev->ip_blocks[i].status.sw = true;
2693
2694 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) {
2695 /* need to do common hw init early so everything is set up for gmc */
2696 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
2697 if (r) {
2698 DRM_ERROR("hw_init %d failed %d\n", i, r);
2699 goto init_failed;
2700 }
2701 adev->ip_blocks[i].status.hw = true;
2702 } else if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
2703 /* need to do gmc hw init early so we can allocate gpu mem */
2704 /* Try to reserve bad pages early */
2705 if (amdgpu_sriov_vf(adev))
2706 amdgpu_virt_exchange_data(adev);
2707
2708 r = amdgpu_device_mem_scratch_init(adev);
2709 if (r) {
2710 DRM_ERROR("amdgpu_mem_scratch_init failed %d\n", r);
2711 goto init_failed;
2712 }
2713 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
2714 if (r) {
2715 DRM_ERROR("hw_init %d failed %d\n", i, r);
2716 goto init_failed;
2717 }
2718 r = amdgpu_device_wb_init(adev);
2719 if (r) {
2720 DRM_ERROR("amdgpu_device_wb_init failed %d\n", r);
2721 goto init_failed;
2722 }
2723 adev->ip_blocks[i].status.hw = true;
2724
2725 /* right after GMC hw init, we create CSA */
2726 if (adev->gfx.mcbp) {
2727 r = amdgpu_allocate_static_csa(adev, &adev->virt.csa_obj,
2728 AMDGPU_GEM_DOMAIN_VRAM |
2729 AMDGPU_GEM_DOMAIN_GTT,
2730 AMDGPU_CSA_SIZE);
2731 if (r) {
2732 DRM_ERROR("allocate CSA failed %d\n", r);
2733 goto init_failed;
2734 }
2735 }
2736
2737 r = amdgpu_seq64_init(adev);
2738 if (r) {
2739 DRM_ERROR("allocate seq64 failed %d\n", r);
2740 goto init_failed;
2741 }
2742 }
2743 }
2744
2745 if (amdgpu_sriov_vf(adev))
2746 amdgpu_virt_init_data_exchange(adev);
2747
2748 r = amdgpu_ib_pool_init(adev);
2749 if (r) {
2750 dev_err(adev->dev, "IB initialization failed (%d).\n", r);
2751 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r);
2752 goto init_failed;
2753 }
2754
2755 r = amdgpu_ucode_create_bo(adev); /* create ucode bo when sw_init complete*/
2756 if (r)
2757 goto init_failed;
2758
2759 r = amdgpu_device_ip_hw_init_phase1(adev);
2760 if (r)
2761 goto init_failed;
2762
2763 r = amdgpu_device_fw_loading(adev);
2764 if (r)
2765 goto init_failed;
2766
2767 r = amdgpu_device_ip_hw_init_phase2(adev);
2768 if (r)
2769 goto init_failed;
2770
2771 /*
2772 * retired pages will be loaded from eeprom and reserved here,
2773 * it should be called after amdgpu_device_ip_hw_init_phase2 since
2774 * for some ASICs the RAS EEPROM code relies on SMU fully functioning
2775 * for I2C communication which only true at this point.
2776 *
2777 * amdgpu_ras_recovery_init may fail, but the upper only cares the
2778 * failure from bad gpu situation and stop amdgpu init process
2779 * accordingly. For other failed cases, it will still release all
2780 * the resource and print error message, rather than returning one
2781 * negative value to upper level.
2782 *
2783 * Note: theoretically, this should be called before all vram allocations
2784 * to protect retired page from abusing
2785 */
2786 r = amdgpu_ras_recovery_init(adev);
2787 if (r)
2788 goto init_failed;
2789
2790 /**
2791 * In case of XGMI grab extra reference for reset domain for this device
2792 */
2793 if (adev->gmc.xgmi.num_physical_nodes > 1) {
2794 if (amdgpu_xgmi_add_device(adev) == 0) {
2795 if (!amdgpu_sriov_vf(adev)) {
2796 struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev);
2797
2798 if (WARN_ON(!hive)) {
2799 r = -ENOENT;
2800 goto init_failed;
2801 }
2802
2803 if (!hive->reset_domain ||
2804 !amdgpu_reset_get_reset_domain(hive->reset_domain)) {
2805 r = -ENOENT;
2806 amdgpu_put_xgmi_hive(hive);
2807 goto init_failed;
2808 }
2809
2810 /* Drop the early temporary reset domain we created for device */
2811 amdgpu_reset_put_reset_domain(adev->reset_domain);
2812 adev->reset_domain = hive->reset_domain;
2813 amdgpu_put_xgmi_hive(hive);
2814 }
2815 }
2816 }
2817
2818 r = amdgpu_device_init_schedulers(adev);
2819 if (r)
2820 goto init_failed;
2821
2822 if (adev->mman.buffer_funcs_ring->sched.ready)
2823 amdgpu_ttm_set_buffer_funcs_status(adev, true);
2824
2825 /* Don't init kfd if whole hive need to be reset during init */
2826 if (!adev->gmc.xgmi.pending_reset) {
2827 kgd2kfd_init_zone_device(adev);
2828 amdgpu_amdkfd_device_init(adev);
2829 }
2830
2831 amdgpu_fru_get_product_info(adev);
2832
2833init_failed:
2834
2835 return r;
2836}
2837
2838/**
2839 * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer
2840 *
2841 * @adev: amdgpu_device pointer
2842 *
2843 * Writes a reset magic value to the gart pointer in VRAM. The driver calls
2844 * this function before a GPU reset. If the value is retained after a
2845 * GPU reset, VRAM has not been lost. Some GPU resets may destry VRAM contents.
2846 */
2847static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev)
2848{
2849 memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
2850}
2851
2852/**
2853 * amdgpu_device_check_vram_lost - check if vram is valid
2854 *
2855 * @adev: amdgpu_device pointer
2856 *
2857 * Checks the reset magic value written to the gart pointer in VRAM.
2858 * The driver calls this after a GPU reset to see if the contents of
2859 * VRAM is lost or now.
2860 * returns true if vram is lost, false if not.
2861 */
2862static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev)
2863{
2864 if (memcmp(adev->gart.ptr, adev->reset_magic,
2865 AMDGPU_RESET_MAGIC_NUM))
2866 return true;
2867
2868 if (!amdgpu_in_reset(adev))
2869 return false;
2870
2871 /*
2872 * For all ASICs with baco/mode1 reset, the VRAM is
2873 * always assumed to be lost.
2874 */
2875 switch (amdgpu_asic_reset_method(adev)) {
2876 case AMD_RESET_METHOD_BACO:
2877 case AMD_RESET_METHOD_MODE1:
2878 return true;
2879 default:
2880 return false;
2881 }
2882}
2883
2884/**
2885 * amdgpu_device_set_cg_state - set clockgating for amdgpu device
2886 *
2887 * @adev: amdgpu_device pointer
2888 * @state: clockgating state (gate or ungate)
2889 *
2890 * The list of all the hardware IPs that make up the asic is walked and the
2891 * set_clockgating_state callbacks are run.
2892 * Late initialization pass enabling clockgating for hardware IPs.
2893 * Fini or suspend, pass disabling clockgating for hardware IPs.
2894 * Returns 0 on success, negative error code on failure.
2895 */
2896
2897int amdgpu_device_set_cg_state(struct amdgpu_device *adev,
2898 enum amd_clockgating_state state)
2899{
2900 int i, j, r;
2901
2902 if (amdgpu_emu_mode == 1)
2903 return 0;
2904
2905 for (j = 0; j < adev->num_ip_blocks; j++) {
2906 i = state == AMD_CG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
2907 if (!adev->ip_blocks[i].status.late_initialized)
2908 continue;
2909 /* skip CG for GFX, SDMA on S0ix */
2910 if (adev->in_s0ix &&
2911 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX ||
2912 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SDMA))
2913 continue;
2914 /* skip CG for VCE/UVD, it's handled specially */
2915 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
2916 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
2917 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
2918 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG &&
2919 adev->ip_blocks[i].version->funcs->set_clockgating_state) {
2920 /* enable clockgating to save power */
2921 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
2922 state);
2923 if (r) {
2924 DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
2925 adev->ip_blocks[i].version->funcs->name, r);
2926 return r;
2927 }
2928 }
2929 }
2930
2931 return 0;
2932}
2933
2934int amdgpu_device_set_pg_state(struct amdgpu_device *adev,
2935 enum amd_powergating_state state)
2936{
2937 int i, j, r;
2938
2939 if (amdgpu_emu_mode == 1)
2940 return 0;
2941
2942 for (j = 0; j < adev->num_ip_blocks; j++) {
2943 i = state == AMD_PG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
2944 if (!adev->ip_blocks[i].status.late_initialized)
2945 continue;
2946 /* skip PG for GFX, SDMA on S0ix */
2947 if (adev->in_s0ix &&
2948 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX ||
2949 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SDMA))
2950 continue;
2951 /* skip CG for VCE/UVD, it's handled specially */
2952 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
2953 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
2954 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
2955 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG &&
2956 adev->ip_blocks[i].version->funcs->set_powergating_state) {
2957 /* enable powergating to save power */
2958 r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev,
2959 state);
2960 if (r) {
2961 DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n",
2962 adev->ip_blocks[i].version->funcs->name, r);
2963 return r;
2964 }
2965 }
2966 }
2967 return 0;
2968}
2969
2970static int amdgpu_device_enable_mgpu_fan_boost(void)
2971{
2972 struct amdgpu_gpu_instance *gpu_ins;
2973 struct amdgpu_device *adev;
2974 int i, ret = 0;
2975
2976 mutex_lock(&mgpu_info.mutex);
2977
2978 /*
2979 * MGPU fan boost feature should be enabled
2980 * only when there are two or more dGPUs in
2981 * the system
2982 */
2983 if (mgpu_info.num_dgpu < 2)
2984 goto out;
2985
2986 for (i = 0; i < mgpu_info.num_dgpu; i++) {
2987 gpu_ins = &(mgpu_info.gpu_ins[i]);
2988 adev = gpu_ins->adev;
2989 if (!(adev->flags & AMD_IS_APU) &&
2990 !gpu_ins->mgpu_fan_enabled) {
2991 ret = amdgpu_dpm_enable_mgpu_fan_boost(adev);
2992 if (ret)
2993 break;
2994
2995 gpu_ins->mgpu_fan_enabled = 1;
2996 }
2997 }
2998
2999out:
3000 mutex_unlock(&mgpu_info.mutex);
3001
3002 return ret;
3003}
3004
3005/**
3006 * amdgpu_device_ip_late_init - run late init for hardware IPs
3007 *
3008 * @adev: amdgpu_device pointer
3009 *
3010 * Late initialization pass for hardware IPs. The list of all the hardware
3011 * IPs that make up the asic is walked and the late_init callbacks are run.
3012 * late_init covers any special initialization that an IP requires
3013 * after all of the have been initialized or something that needs to happen
3014 * late in the init process.
3015 * Returns 0 on success, negative error code on failure.
3016 */
3017static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
3018{
3019 struct amdgpu_gpu_instance *gpu_instance;
3020 int i = 0, r;
3021
3022 for (i = 0; i < adev->num_ip_blocks; i++) {
3023 if (!adev->ip_blocks[i].status.hw)
3024 continue;
3025 if (adev->ip_blocks[i].version->funcs->late_init) {
3026 r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
3027 if (r) {
3028 DRM_ERROR("late_init of IP block <%s> failed %d\n",
3029 adev->ip_blocks[i].version->funcs->name, r);
3030 return r;
3031 }
3032 }
3033 adev->ip_blocks[i].status.late_initialized = true;
3034 }
3035
3036 r = amdgpu_ras_late_init(adev);
3037 if (r) {
3038 DRM_ERROR("amdgpu_ras_late_init failed %d", r);
3039 return r;
3040 }
3041
3042 amdgpu_ras_set_error_query_ready(adev, true);
3043
3044 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE);
3045 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE);
3046
3047 amdgpu_device_fill_reset_magic(adev);
3048
3049 r = amdgpu_device_enable_mgpu_fan_boost();
3050 if (r)
3051 DRM_ERROR("enable mgpu fan boost failed (%d).\n", r);
3052
3053 /* For passthrough configuration on arcturus and aldebaran, enable special handling SBR */
3054 if (amdgpu_passthrough(adev) &&
3055 ((adev->asic_type == CHIP_ARCTURUS && adev->gmc.xgmi.num_physical_nodes > 1) ||
3056 adev->asic_type == CHIP_ALDEBARAN))
3057 amdgpu_dpm_handle_passthrough_sbr(adev, true);
3058
3059 if (adev->gmc.xgmi.num_physical_nodes > 1) {
3060 mutex_lock(&mgpu_info.mutex);
3061
3062 /*
3063 * Reset device p-state to low as this was booted with high.
3064 *
3065 * This should be performed only after all devices from the same
3066 * hive get initialized.
3067 *
3068 * However, it's unknown how many device in the hive in advance.
3069 * As this is counted one by one during devices initializations.
3070 *
3071 * So, we wait for all XGMI interlinked devices initialized.
3072 * This may bring some delays as those devices may come from
3073 * different hives. But that should be OK.
3074 */
3075 if (mgpu_info.num_dgpu == adev->gmc.xgmi.num_physical_nodes) {
3076 for (i = 0; i < mgpu_info.num_gpu; i++) {
3077 gpu_instance = &(mgpu_info.gpu_ins[i]);
3078 if (gpu_instance->adev->flags & AMD_IS_APU)
3079 continue;
3080
3081 r = amdgpu_xgmi_set_pstate(gpu_instance->adev,
3082 AMDGPU_XGMI_PSTATE_MIN);
3083 if (r) {
3084 DRM_ERROR("pstate setting failed (%d).\n", r);
3085 break;
3086 }
3087 }
3088 }
3089
3090 mutex_unlock(&mgpu_info.mutex);
3091 }
3092
3093 return 0;
3094}
3095
3096/**
3097 * amdgpu_device_smu_fini_early - smu hw_fini wrapper
3098 *
3099 * @adev: amdgpu_device pointer
3100 *
3101 * For ASICs need to disable SMC first
3102 */
3103static void amdgpu_device_smu_fini_early(struct amdgpu_device *adev)
3104{
3105 int i, r;
3106
3107 if (amdgpu_ip_version(adev, GC_HWIP, 0) > IP_VERSION(9, 0, 0))
3108 return;
3109
3110 for (i = 0; i < adev->num_ip_blocks; i++) {
3111 if (!adev->ip_blocks[i].status.hw)
3112 continue;
3113 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
3114 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
3115 /* XXX handle errors */
3116 if (r) {
3117 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
3118 adev->ip_blocks[i].version->funcs->name, r);
3119 }
3120 adev->ip_blocks[i].status.hw = false;
3121 break;
3122 }
3123 }
3124}
3125
3126static int amdgpu_device_ip_fini_early(struct amdgpu_device *adev)
3127{
3128 int i, r;
3129
3130 for (i = 0; i < adev->num_ip_blocks; i++) {
3131 if (!adev->ip_blocks[i].version->funcs->early_fini)
3132 continue;
3133
3134 r = adev->ip_blocks[i].version->funcs->early_fini((void *)adev);
3135 if (r) {
3136 DRM_DEBUG("early_fini of IP block <%s> failed %d\n",
3137 adev->ip_blocks[i].version->funcs->name, r);
3138 }
3139 }
3140
3141 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
3142 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
3143
3144 amdgpu_amdkfd_suspend(adev, false);
3145
3146 /* Workaroud for ASICs need to disable SMC first */
3147 amdgpu_device_smu_fini_early(adev);
3148
3149 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
3150 if (!adev->ip_blocks[i].status.hw)
3151 continue;
3152
3153 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
3154 /* XXX handle errors */
3155 if (r) {
3156 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
3157 adev->ip_blocks[i].version->funcs->name, r);
3158 }
3159
3160 adev->ip_blocks[i].status.hw = false;
3161 }
3162
3163 if (amdgpu_sriov_vf(adev)) {
3164 if (amdgpu_virt_release_full_gpu(adev, false))
3165 DRM_ERROR("failed to release exclusive mode on fini\n");
3166 }
3167
3168 return 0;
3169}
3170
3171/**
3172 * amdgpu_device_ip_fini - run fini for hardware IPs
3173 *
3174 * @adev: amdgpu_device pointer
3175 *
3176 * Main teardown pass for hardware IPs. The list of all the hardware
3177 * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks
3178 * are run. hw_fini tears down the hardware associated with each IP
3179 * and sw_fini tears down any software state associated with each IP.
3180 * Returns 0 on success, negative error code on failure.
3181 */
3182static int amdgpu_device_ip_fini(struct amdgpu_device *adev)
3183{
3184 int i, r;
3185
3186 if (amdgpu_sriov_vf(adev) && adev->virt.ras_init_done)
3187 amdgpu_virt_release_ras_err_handler_data(adev);
3188
3189 if (adev->gmc.xgmi.num_physical_nodes > 1)
3190 amdgpu_xgmi_remove_device(adev);
3191
3192 amdgpu_amdkfd_device_fini_sw(adev);
3193
3194 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
3195 if (!adev->ip_blocks[i].status.sw)
3196 continue;
3197
3198 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
3199 amdgpu_ucode_free_bo(adev);
3200 amdgpu_free_static_csa(&adev->virt.csa_obj);
3201 amdgpu_device_wb_fini(adev);
3202 amdgpu_device_mem_scratch_fini(adev);
3203 amdgpu_ib_pool_fini(adev);
3204 amdgpu_seq64_fini(adev);
3205 }
3206
3207 r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
3208 /* XXX handle errors */
3209 if (r) {
3210 DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
3211 adev->ip_blocks[i].version->funcs->name, r);
3212 }
3213 adev->ip_blocks[i].status.sw = false;
3214 adev->ip_blocks[i].status.valid = false;
3215 }
3216
3217 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
3218 if (!adev->ip_blocks[i].status.late_initialized)
3219 continue;
3220 if (adev->ip_blocks[i].version->funcs->late_fini)
3221 adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
3222 adev->ip_blocks[i].status.late_initialized = false;
3223 }
3224
3225 amdgpu_ras_fini(adev);
3226
3227 return 0;
3228}
3229
3230/**
3231 * amdgpu_device_delayed_init_work_handler - work handler for IB tests
3232 *
3233 * @work: work_struct.
3234 */
3235static void amdgpu_device_delayed_init_work_handler(struct work_struct *work)
3236{
3237 struct amdgpu_device *adev =
3238 container_of(work, struct amdgpu_device, delayed_init_work.work);
3239 int r;
3240
3241 r = amdgpu_ib_ring_tests(adev);
3242 if (r)
3243 DRM_ERROR("ib ring test failed (%d).\n", r);
3244}
3245
3246static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work)
3247{
3248 struct amdgpu_device *adev =
3249 container_of(work, struct amdgpu_device, gfx.gfx_off_delay_work.work);
3250
3251 WARN_ON_ONCE(adev->gfx.gfx_off_state);
3252 WARN_ON_ONCE(adev->gfx.gfx_off_req_count);
3253
3254 if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, true))
3255 adev->gfx.gfx_off_state = true;
3256}
3257
3258/**
3259 * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1)
3260 *
3261 * @adev: amdgpu_device pointer
3262 *
3263 * Main suspend function for hardware IPs. The list of all the hardware
3264 * IPs that make up the asic is walked, clockgating is disabled and the
3265 * suspend callbacks are run. suspend puts the hardware and software state
3266 * in each IP into a state suitable for suspend.
3267 * Returns 0 on success, negative error code on failure.
3268 */
3269static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev)
3270{
3271 int i, r;
3272
3273 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
3274 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
3275
3276 /*
3277 * Per PMFW team's suggestion, driver needs to handle gfxoff
3278 * and df cstate features disablement for gpu reset(e.g. Mode1Reset)
3279 * scenario. Add the missing df cstate disablement here.
3280 */
3281 if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_DISALLOW))
3282 dev_warn(adev->dev, "Failed to disallow df cstate");
3283
3284 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
3285 if (!adev->ip_blocks[i].status.valid)
3286 continue;
3287
3288 /* displays are handled separately */
3289 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_DCE)
3290 continue;
3291
3292 /* XXX handle errors */
3293 r = adev->ip_blocks[i].version->funcs->suspend(adev);
3294 /* XXX handle errors */
3295 if (r) {
3296 DRM_ERROR("suspend of IP block <%s> failed %d\n",
3297 adev->ip_blocks[i].version->funcs->name, r);
3298 return r;
3299 }
3300
3301 adev->ip_blocks[i].status.hw = false;
3302 }
3303
3304 return 0;
3305}
3306
3307/**
3308 * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2)
3309 *
3310 * @adev: amdgpu_device pointer
3311 *
3312 * Main suspend function for hardware IPs. The list of all the hardware
3313 * IPs that make up the asic is walked, clockgating is disabled and the
3314 * suspend callbacks are run. suspend puts the hardware and software state
3315 * in each IP into a state suitable for suspend.
3316 * Returns 0 on success, negative error code on failure.
3317 */
3318static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev)
3319{
3320 int i, r;
3321
3322 if (adev->in_s0ix)
3323 amdgpu_dpm_gfx_state_change(adev, sGpuChangeState_D3Entry);
3324
3325 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
3326 if (!adev->ip_blocks[i].status.valid)
3327 continue;
3328 /* displays are handled in phase1 */
3329 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE)
3330 continue;
3331 /* PSP lost connection when err_event_athub occurs */
3332 if (amdgpu_ras_intr_triggered() &&
3333 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
3334 adev->ip_blocks[i].status.hw = false;
3335 continue;
3336 }
3337
3338 /* skip unnecessary suspend if we do not initialize them yet */
3339 if (adev->gmc.xgmi.pending_reset &&
3340 !(adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
3341 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC ||
3342 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
3343 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH)) {
3344 adev->ip_blocks[i].status.hw = false;
3345 continue;
3346 }
3347
3348 /* skip suspend of gfx/mes and psp for S0ix
3349 * gfx is in gfxoff state, so on resume it will exit gfxoff just
3350 * like at runtime. PSP is also part of the always on hardware
3351 * so no need to suspend it.
3352 */
3353 if (adev->in_s0ix &&
3354 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP ||
3355 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX ||
3356 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_MES))
3357 continue;
3358
3359 /* SDMA 5.x+ is part of GFX power domain so it's covered by GFXOFF */
3360 if (adev->in_s0ix &&
3361 (amdgpu_ip_version(adev, SDMA0_HWIP, 0) >=
3362 IP_VERSION(5, 0, 0)) &&
3363 (adev->ip_blocks[i].version->type ==
3364 AMD_IP_BLOCK_TYPE_SDMA))
3365 continue;
3366
3367 /* Once swPSP provides the IMU, RLC FW binaries to TOS during cold-boot.
3368 * These are in TMR, hence are expected to be reused by PSP-TOS to reload
3369 * from this location and RLC Autoload automatically also gets loaded
3370 * from here based on PMFW -> PSP message during re-init sequence.
3371 * Therefore, the psp suspend & resume should be skipped to avoid destroy
3372 * the TMR and reload FWs again for IMU enabled APU ASICs.
3373 */
3374 if (amdgpu_in_reset(adev) &&
3375 (adev->flags & AMD_IS_APU) && adev->gfx.imu.funcs &&
3376 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)
3377 continue;
3378
3379 /* XXX handle errors */
3380 r = adev->ip_blocks[i].version->funcs->suspend(adev);
3381 /* XXX handle errors */
3382 if (r) {
3383 DRM_ERROR("suspend of IP block <%s> failed %d\n",
3384 adev->ip_blocks[i].version->funcs->name, r);
3385 }
3386 adev->ip_blocks[i].status.hw = false;
3387 /* handle putting the SMC in the appropriate state */
3388 if (!amdgpu_sriov_vf(adev)) {
3389 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
3390 r = amdgpu_dpm_set_mp1_state(adev, adev->mp1_state);
3391 if (r) {
3392 DRM_ERROR("SMC failed to set mp1 state %d, %d\n",
3393 adev->mp1_state, r);
3394 return r;
3395 }
3396 }
3397 }
3398 }
3399
3400 return 0;
3401}
3402
3403/**
3404 * amdgpu_device_ip_suspend - run suspend for hardware IPs
3405 *
3406 * @adev: amdgpu_device pointer
3407 *
3408 * Main suspend function for hardware IPs. The list of all the hardware
3409 * IPs that make up the asic is walked, clockgating is disabled and the
3410 * suspend callbacks are run. suspend puts the hardware and software state
3411 * in each IP into a state suitable for suspend.
3412 * Returns 0 on success, negative error code on failure.
3413 */
3414int amdgpu_device_ip_suspend(struct amdgpu_device *adev)
3415{
3416 int r;
3417
3418 if (amdgpu_sriov_vf(adev)) {
3419 amdgpu_virt_fini_data_exchange(adev);
3420 amdgpu_virt_request_full_gpu(adev, false);
3421 }
3422
3423 amdgpu_ttm_set_buffer_funcs_status(adev, false);
3424
3425 r = amdgpu_device_ip_suspend_phase1(adev);
3426 if (r)
3427 return r;
3428 r = amdgpu_device_ip_suspend_phase2(adev);
3429
3430 if (amdgpu_sriov_vf(adev))
3431 amdgpu_virt_release_full_gpu(adev, false);
3432
3433 return r;
3434}
3435
3436static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
3437{
3438 int i, r;
3439
3440 static enum amd_ip_block_type ip_order[] = {
3441 AMD_IP_BLOCK_TYPE_COMMON,
3442 AMD_IP_BLOCK_TYPE_GMC,
3443 AMD_IP_BLOCK_TYPE_PSP,
3444 AMD_IP_BLOCK_TYPE_IH,
3445 };
3446
3447 for (i = 0; i < adev->num_ip_blocks; i++) {
3448 int j;
3449 struct amdgpu_ip_block *block;
3450
3451 block = &adev->ip_blocks[i];
3452 block->status.hw = false;
3453
3454 for (j = 0; j < ARRAY_SIZE(ip_order); j++) {
3455
3456 if (block->version->type != ip_order[j] ||
3457 !block->status.valid)
3458 continue;
3459
3460 r = block->version->funcs->hw_init(adev);
3461 DRM_INFO("RE-INIT-early: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
3462 if (r)
3463 return r;
3464 block->status.hw = true;
3465 }
3466 }
3467
3468 return 0;
3469}
3470
3471static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
3472{
3473 int i, r;
3474
3475 static enum amd_ip_block_type ip_order[] = {
3476 AMD_IP_BLOCK_TYPE_SMC,
3477 AMD_IP_BLOCK_TYPE_DCE,
3478 AMD_IP_BLOCK_TYPE_GFX,
3479 AMD_IP_BLOCK_TYPE_SDMA,
3480 AMD_IP_BLOCK_TYPE_MES,
3481 AMD_IP_BLOCK_TYPE_UVD,
3482 AMD_IP_BLOCK_TYPE_VCE,
3483 AMD_IP_BLOCK_TYPE_VCN,
3484 AMD_IP_BLOCK_TYPE_JPEG
3485 };
3486
3487 for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
3488 int j;
3489 struct amdgpu_ip_block *block;
3490
3491 for (j = 0; j < adev->num_ip_blocks; j++) {
3492 block = &adev->ip_blocks[j];
3493
3494 if (block->version->type != ip_order[i] ||
3495 !block->status.valid ||
3496 block->status.hw)
3497 continue;
3498
3499 if (block->version->type == AMD_IP_BLOCK_TYPE_SMC)
3500 r = block->version->funcs->resume(adev);
3501 else
3502 r = block->version->funcs->hw_init(adev);
3503
3504 DRM_INFO("RE-INIT-late: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
3505 if (r)
3506 return r;
3507 block->status.hw = true;
3508 }
3509 }
3510
3511 return 0;
3512}
3513
3514/**
3515 * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs
3516 *
3517 * @adev: amdgpu_device pointer
3518 *
3519 * First resume function for hardware IPs. The list of all the hardware
3520 * IPs that make up the asic is walked and the resume callbacks are run for
3521 * COMMON, GMC, and IH. resume puts the hardware into a functional state
3522 * after a suspend and updates the software state as necessary. This
3523 * function is also used for restoring the GPU after a GPU reset.
3524 * Returns 0 on success, negative error code on failure.
3525 */
3526static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)
3527{
3528 int i, r;
3529
3530 for (i = 0; i < adev->num_ip_blocks; i++) {
3531 if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
3532 continue;
3533 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
3534 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
3535 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH ||
3536 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP && amdgpu_sriov_vf(adev))) {
3537
3538 r = adev->ip_blocks[i].version->funcs->resume(adev);
3539 if (r) {
3540 DRM_ERROR("resume of IP block <%s> failed %d\n",
3541 adev->ip_blocks[i].version->funcs->name, r);
3542 return r;
3543 }
3544 adev->ip_blocks[i].status.hw = true;
3545 }
3546 }
3547
3548 return 0;
3549}
3550
3551/**
3552 * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs
3553 *
3554 * @adev: amdgpu_device pointer
3555 *
3556 * First resume function for hardware IPs. The list of all the hardware
3557 * IPs that make up the asic is walked and the resume callbacks are run for
3558 * all blocks except COMMON, GMC, and IH. resume puts the hardware into a
3559 * functional state after a suspend and updates the software state as
3560 * necessary. This function is also used for restoring the GPU after a GPU
3561 * reset.
3562 * Returns 0 on success, negative error code on failure.
3563 */
3564static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)
3565{
3566 int i, r;
3567
3568 for (i = 0; i < adev->num_ip_blocks; i++) {
3569 if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
3570 continue;
3571 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
3572 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
3573 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH ||
3574 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)
3575 continue;
3576 r = adev->ip_blocks[i].version->funcs->resume(adev);
3577 if (r) {
3578 DRM_ERROR("resume of IP block <%s> failed %d\n",
3579 adev->ip_blocks[i].version->funcs->name, r);
3580 return r;
3581 }
3582 adev->ip_blocks[i].status.hw = true;
3583 }
3584
3585 return 0;
3586}
3587
3588/**
3589 * amdgpu_device_ip_resume - run resume for hardware IPs
3590 *
3591 * @adev: amdgpu_device pointer
3592 *
3593 * Main resume function for hardware IPs. The hardware IPs
3594 * are split into two resume functions because they are
3595 * also used in recovering from a GPU reset and some additional
3596 * steps need to be take between them. In this case (S3/S4) they are
3597 * run sequentially.
3598 * Returns 0 on success, negative error code on failure.
3599 */
3600static int amdgpu_device_ip_resume(struct amdgpu_device *adev)
3601{
3602 int r;
3603
3604 r = amdgpu_device_ip_resume_phase1(adev);
3605 if (r)
3606 return r;
3607
3608 r = amdgpu_device_fw_loading(adev);
3609 if (r)
3610 return r;
3611
3612 r = amdgpu_device_ip_resume_phase2(adev);
3613
3614 if (adev->mman.buffer_funcs_ring->sched.ready)
3615 amdgpu_ttm_set_buffer_funcs_status(adev, true);
3616
3617 return r;
3618}
3619
3620/**
3621 * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV
3622 *
3623 * @adev: amdgpu_device pointer
3624 *
3625 * Query the VBIOS data tables to determine if the board supports SR-IOV.
3626 */
3627static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
3628{
3629 if (amdgpu_sriov_vf(adev)) {
3630 if (adev->is_atom_fw) {
3631 if (amdgpu_atomfirmware_gpu_virtualization_supported(adev))
3632 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
3633 } else {
3634 if (amdgpu_atombios_has_gpu_virtualization_table(adev))
3635 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
3636 }
3637
3638 if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
3639 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
3640 }
3641}
3642
3643/**
3644 * amdgpu_device_asic_has_dc_support - determine if DC supports the asic
3645 *
3646 * @asic_type: AMD asic type
3647 *
3648 * Check if there is DC (new modesetting infrastructre) support for an asic.
3649 * returns true if DC has support, false if not.
3650 */
3651bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
3652{
3653 switch (asic_type) {
3654#ifdef CONFIG_DRM_AMDGPU_SI
3655 case CHIP_HAINAN:
3656#endif
3657 case CHIP_TOPAZ:
3658 /* chips with no display hardware */
3659 return false;
3660#if defined(CONFIG_DRM_AMD_DC)
3661 case CHIP_TAHITI:
3662 case CHIP_PITCAIRN:
3663 case CHIP_VERDE:
3664 case CHIP_OLAND:
3665 /*
3666 * We have systems in the wild with these ASICs that require
3667 * LVDS and VGA support which is not supported with DC.
3668 *
3669 * Fallback to the non-DC driver here by default so as not to
3670 * cause regressions.
3671 */
3672#if defined(CONFIG_DRM_AMD_DC_SI)
3673 return amdgpu_dc > 0;
3674#else
3675 return false;
3676#endif
3677 case CHIP_BONAIRE:
3678 case CHIP_KAVERI:
3679 case CHIP_KABINI:
3680 case CHIP_MULLINS:
3681 /*
3682 * We have systems in the wild with these ASICs that require
3683 * VGA support which is not supported with DC.
3684 *
3685 * Fallback to the non-DC driver here by default so as not to
3686 * cause regressions.
3687 */
3688 return amdgpu_dc > 0;
3689 default:
3690 return amdgpu_dc != 0;
3691#else
3692 default:
3693 if (amdgpu_dc > 0)
3694 DRM_INFO_ONCE("Display Core has been requested via kernel parameter but isn't supported by ASIC, ignoring\n");
3695 return false;
3696#endif
3697 }
3698}
3699
3700/**
3701 * amdgpu_device_has_dc_support - check if dc is supported
3702 *
3703 * @adev: amdgpu_device pointer
3704 *
3705 * Returns true for supported, false for not supported
3706 */
3707bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
3708{
3709 if (adev->enable_virtual_display ||
3710 (adev->harvest_ip_mask & AMD_HARVEST_IP_DMU_MASK))
3711 return false;
3712
3713 return amdgpu_device_asic_has_dc_support(adev->asic_type);
3714}
3715
3716static void amdgpu_device_xgmi_reset_func(struct work_struct *__work)
3717{
3718 struct amdgpu_device *adev =
3719 container_of(__work, struct amdgpu_device, xgmi_reset_work);
3720 struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev);
3721
3722 /* It's a bug to not have a hive within this function */
3723 if (WARN_ON(!hive))
3724 return;
3725
3726 /*
3727 * Use task barrier to synchronize all xgmi reset works across the
3728 * hive. task_barrier_enter and task_barrier_exit will block
3729 * until all the threads running the xgmi reset works reach
3730 * those points. task_barrier_full will do both blocks.
3731 */
3732 if (amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) {
3733
3734 task_barrier_enter(&hive->tb);
3735 adev->asic_reset_res = amdgpu_device_baco_enter(adev_to_drm(adev));
3736
3737 if (adev->asic_reset_res)
3738 goto fail;
3739
3740 task_barrier_exit(&hive->tb);
3741 adev->asic_reset_res = amdgpu_device_baco_exit(adev_to_drm(adev));
3742
3743 if (adev->asic_reset_res)
3744 goto fail;
3745
3746 amdgpu_ras_reset_error_count(adev, AMDGPU_RAS_BLOCK__MMHUB);
3747 } else {
3748
3749 task_barrier_full(&hive->tb);
3750 adev->asic_reset_res = amdgpu_asic_reset(adev);
3751 }
3752
3753fail:
3754 if (adev->asic_reset_res)
3755 DRM_WARN("ASIC reset failed with error, %d for drm dev, %s",
3756 adev->asic_reset_res, adev_to_drm(adev)->unique);
3757 amdgpu_put_xgmi_hive(hive);
3758}
3759
3760static int amdgpu_device_get_job_timeout_settings(struct amdgpu_device *adev)
3761{
3762 char *input = amdgpu_lockup_timeout;
3763 char *timeout_setting = NULL;
3764 int index = 0;
3765 long timeout;
3766 int ret = 0;
3767
3768 /*
3769 * By default timeout for non compute jobs is 10000
3770 * and 60000 for compute jobs.
3771 * In SR-IOV or passthrough mode, timeout for compute
3772 * jobs are 60000 by default.
3773 */
3774 adev->gfx_timeout = msecs_to_jiffies(10000);
3775 adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout;
3776 if (amdgpu_sriov_vf(adev))
3777 adev->compute_timeout = amdgpu_sriov_is_pp_one_vf(adev) ?
3778 msecs_to_jiffies(60000) : msecs_to_jiffies(10000);
3779 else
3780 adev->compute_timeout = msecs_to_jiffies(60000);
3781
3782 if (strnlen(input, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) {
3783 while ((timeout_setting = strsep(&input, ",")) &&
3784 strnlen(timeout_setting, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) {
3785 ret = kstrtol(timeout_setting, 0, &timeout);
3786 if (ret)
3787 return ret;
3788
3789 if (timeout == 0) {
3790 index++;
3791 continue;
3792 } else if (timeout < 0) {
3793 timeout = MAX_SCHEDULE_TIMEOUT;
3794 dev_warn(adev->dev, "lockup timeout disabled");
3795 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
3796 } else {
3797 timeout = msecs_to_jiffies(timeout);
3798 }
3799
3800 switch (index++) {
3801 case 0:
3802 adev->gfx_timeout = timeout;
3803 break;
3804 case 1:
3805 adev->compute_timeout = timeout;
3806 break;
3807 case 2:
3808 adev->sdma_timeout = timeout;
3809 break;
3810 case 3:
3811 adev->video_timeout = timeout;
3812 break;
3813 default:
3814 break;
3815 }
3816 }
3817 /*
3818 * There is only one value specified and
3819 * it should apply to all non-compute jobs.
3820 */
3821 if (index == 1) {
3822 adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout;
3823 if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev))
3824 adev->compute_timeout = adev->gfx_timeout;
3825 }
3826 }
3827
3828 return ret;
3829}
3830
3831/**
3832 * amdgpu_device_check_iommu_direct_map - check if RAM direct mapped to GPU
3833 *
3834 * @adev: amdgpu_device pointer
3835 *
3836 * RAM direct mapped to GPU if IOMMU is not enabled or is pass through mode
3837 */
3838static void amdgpu_device_check_iommu_direct_map(struct amdgpu_device *adev)
3839{
3840 struct iommu_domain *domain;
3841
3842 domain = iommu_get_domain_for_dev(adev->dev);
3843 if (!domain || domain->type == IOMMU_DOMAIN_IDENTITY)
3844 adev->ram_is_direct_mapped = true;
3845}
3846
3847static const struct attribute *amdgpu_dev_attributes[] = {
3848 &dev_attr_pcie_replay_count.attr,
3849 NULL
3850};
3851
3852static void amdgpu_device_set_mcbp(struct amdgpu_device *adev)
3853{
3854 if (amdgpu_mcbp == 1)
3855 adev->gfx.mcbp = true;
3856 else if (amdgpu_mcbp == 0)
3857 adev->gfx.mcbp = false;
3858
3859 if (amdgpu_sriov_vf(adev))
3860 adev->gfx.mcbp = true;
3861
3862 if (adev->gfx.mcbp)
3863 DRM_INFO("MCBP is enabled\n");
3864}
3865
3866/**
3867 * amdgpu_device_init - initialize the driver
3868 *
3869 * @adev: amdgpu_device pointer
3870 * @flags: driver flags
3871 *
3872 * Initializes the driver info and hw (all asics).
3873 * Returns 0 for success or an error on failure.
3874 * Called at driver startup.
3875 */
3876int amdgpu_device_init(struct amdgpu_device *adev,
3877 uint32_t flags)
3878{
3879 struct drm_device *ddev = adev_to_drm(adev);
3880 struct pci_dev *pdev = adev->pdev;
3881 int r, i;
3882 bool px = false;
3883 u32 max_MBps;
3884 int tmp;
3885
3886 adev->shutdown = false;
3887 adev->flags = flags;
3888
3889 if (amdgpu_force_asic_type >= 0 && amdgpu_force_asic_type < CHIP_LAST)
3890 adev->asic_type = amdgpu_force_asic_type;
3891 else
3892 adev->asic_type = flags & AMD_ASIC_MASK;
3893
3894 adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
3895 if (amdgpu_emu_mode == 1)
3896 adev->usec_timeout *= 10;
3897 adev->gmc.gart_size = 512 * 1024 * 1024;
3898 adev->accel_working = false;
3899 adev->num_rings = 0;
3900 RCU_INIT_POINTER(adev->gang_submit, dma_fence_get_stub());
3901 adev->mman.buffer_funcs = NULL;
3902 adev->mman.buffer_funcs_ring = NULL;
3903 adev->vm_manager.vm_pte_funcs = NULL;
3904 adev->vm_manager.vm_pte_num_scheds = 0;
3905 adev->gmc.gmc_funcs = NULL;
3906 adev->harvest_ip_mask = 0x0;
3907 adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
3908 bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
3909
3910 adev->smc_rreg = &amdgpu_invalid_rreg;
3911 adev->smc_wreg = &amdgpu_invalid_wreg;
3912 adev->pcie_rreg = &amdgpu_invalid_rreg;
3913 adev->pcie_wreg = &amdgpu_invalid_wreg;
3914 adev->pcie_rreg_ext = &amdgpu_invalid_rreg_ext;
3915 adev->pcie_wreg_ext = &amdgpu_invalid_wreg_ext;
3916 adev->pciep_rreg = &amdgpu_invalid_rreg;
3917 adev->pciep_wreg = &amdgpu_invalid_wreg;
3918 adev->pcie_rreg64 = &amdgpu_invalid_rreg64;
3919 adev->pcie_wreg64 = &amdgpu_invalid_wreg64;
3920 adev->pcie_rreg64_ext = &amdgpu_invalid_rreg64_ext;
3921 adev->pcie_wreg64_ext = &amdgpu_invalid_wreg64_ext;
3922 adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
3923 adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
3924 adev->didt_rreg = &amdgpu_invalid_rreg;
3925 adev->didt_wreg = &amdgpu_invalid_wreg;
3926 adev->gc_cac_rreg = &amdgpu_invalid_rreg;
3927 adev->gc_cac_wreg = &amdgpu_invalid_wreg;
3928 adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
3929 adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
3930
3931 DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
3932 amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
3933 pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
3934
3935 /* mutex initialization are all done here so we
3936 * can recall function without having locking issues
3937 */
3938 mutex_init(&adev->firmware.mutex);
3939 mutex_init(&adev->pm.mutex);
3940 mutex_init(&adev->gfx.gpu_clock_mutex);
3941 mutex_init(&adev->srbm_mutex);
3942 mutex_init(&adev->gfx.pipe_reserve_mutex);
3943 mutex_init(&adev->gfx.gfx_off_mutex);
3944 mutex_init(&adev->gfx.partition_mutex);
3945 mutex_init(&adev->grbm_idx_mutex);
3946 mutex_init(&adev->mn_lock);
3947 mutex_init(&adev->virt.vf_errors.lock);
3948 hash_init(adev->mn_hash);
3949 mutex_init(&adev->psp.mutex);
3950 mutex_init(&adev->notifier_lock);
3951 mutex_init(&adev->pm.stable_pstate_ctx_lock);
3952 mutex_init(&adev->benchmark_mutex);
3953
3954 amdgpu_device_init_apu_flags(adev);
3955
3956 r = amdgpu_device_check_arguments(adev);
3957 if (r)
3958 return r;
3959
3960 spin_lock_init(&adev->mmio_idx_lock);
3961 spin_lock_init(&adev->smc_idx_lock);
3962 spin_lock_init(&adev->pcie_idx_lock);
3963 spin_lock_init(&adev->uvd_ctx_idx_lock);
3964 spin_lock_init(&adev->didt_idx_lock);
3965 spin_lock_init(&adev->gc_cac_idx_lock);
3966 spin_lock_init(&adev->se_cac_idx_lock);
3967 spin_lock_init(&adev->audio_endpt_idx_lock);
3968 spin_lock_init(&adev->mm_stats.lock);
3969
3970 INIT_LIST_HEAD(&adev->shadow_list);
3971 mutex_init(&adev->shadow_list_lock);
3972
3973 INIT_LIST_HEAD(&adev->reset_list);
3974
3975 INIT_LIST_HEAD(&adev->ras_list);
3976
3977 INIT_LIST_HEAD(&adev->pm.od_kobj_list);
3978
3979 INIT_DELAYED_WORK(&adev->delayed_init_work,
3980 amdgpu_device_delayed_init_work_handler);
3981 INIT_DELAYED_WORK(&adev->gfx.gfx_off_delay_work,
3982 amdgpu_device_delay_enable_gfx_off);
3983
3984 INIT_WORK(&adev->xgmi_reset_work, amdgpu_device_xgmi_reset_func);
3985
3986 adev->gfx.gfx_off_req_count = 1;
3987 adev->gfx.gfx_off_residency = 0;
3988 adev->gfx.gfx_off_entrycount = 0;
3989 adev->pm.ac_power = power_supply_is_system_supplied() > 0;
3990
3991 atomic_set(&adev->throttling_logging_enabled, 1);
3992 /*
3993 * If throttling continues, logging will be performed every minute
3994 * to avoid log flooding. "-1" is subtracted since the thermal
3995 * throttling interrupt comes every second. Thus, the total logging
3996 * interval is 59 seconds(retelimited printk interval) + 1(waiting
3997 * for throttling interrupt) = 60 seconds.
3998 */
3999 ratelimit_state_init(&adev->throttling_logging_rs, (60 - 1) * HZ, 1);
4000 ratelimit_set_flags(&adev->throttling_logging_rs, RATELIMIT_MSG_ON_RELEASE);
4001
4002 /* Registers mapping */
4003 /* TODO: block userspace mapping of io register */
4004 if (adev->asic_type >= CHIP_BONAIRE) {
4005 adev->rmmio_base = pci_resource_start(adev->pdev, 5);
4006 adev->rmmio_size = pci_resource_len(adev->pdev, 5);
4007 } else {
4008 adev->rmmio_base = pci_resource_start(adev->pdev, 2);
4009 adev->rmmio_size = pci_resource_len(adev->pdev, 2);
4010 }
4011
4012 for (i = 0; i < AMD_IP_BLOCK_TYPE_NUM; i++)
4013 atomic_set(&adev->pm.pwr_state[i], POWER_STATE_UNKNOWN);
4014
4015 adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
4016 if (!adev->rmmio)
4017 return -ENOMEM;
4018
4019 DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
4020 DRM_INFO("register mmio size: %u\n", (unsigned int)adev->rmmio_size);
4021
4022 /*
4023 * Reset domain needs to be present early, before XGMI hive discovered
4024 * (if any) and intitialized to use reset sem and in_gpu reset flag
4025 * early on during init and before calling to RREG32.
4026 */
4027 adev->reset_domain = amdgpu_reset_create_reset_domain(SINGLE_DEVICE, "amdgpu-reset-dev");
4028 if (!adev->reset_domain)
4029 return -ENOMEM;
4030
4031 /* detect hw virtualization here */
4032 amdgpu_detect_virtualization(adev);
4033
4034 amdgpu_device_get_pcie_info(adev);
4035
4036 r = amdgpu_device_get_job_timeout_settings(adev);
4037 if (r) {
4038 dev_err(adev->dev, "invalid lockup_timeout parameter syntax\n");
4039 return r;
4040 }
4041
4042 /* early init functions */
4043 r = amdgpu_device_ip_early_init(adev);
4044 if (r)
4045 return r;
4046
4047 amdgpu_device_set_mcbp(adev);
4048
4049 /* Get rid of things like offb */
4050 r = drm_aperture_remove_conflicting_pci_framebuffers(adev->pdev, &amdgpu_kms_driver);
4051 if (r)
4052 return r;
4053
4054 /* Enable TMZ based on IP_VERSION */
4055 amdgpu_gmc_tmz_set(adev);
4056
4057 amdgpu_gmc_noretry_set(adev);
4058 /* Need to get xgmi info early to decide the reset behavior*/
4059 if (adev->gmc.xgmi.supported) {
4060 r = adev->gfxhub.funcs->get_xgmi_info(adev);
4061 if (r)
4062 return r;
4063 }
4064
4065 /* enable PCIE atomic ops */
4066 if (amdgpu_sriov_vf(adev)) {
4067 if (adev->virt.fw_reserve.p_pf2vf)
4068 adev->have_atomics_support = ((struct amd_sriov_msg_pf2vf_info *)
4069 adev->virt.fw_reserve.p_pf2vf)->pcie_atomic_ops_support_flags ==
4070 (PCI_EXP_DEVCAP2_ATOMIC_COMP32 | PCI_EXP_DEVCAP2_ATOMIC_COMP64);
4071 /* APUs w/ gfx9 onwards doesn't reply on PCIe atomics, rather it is a
4072 * internal path natively support atomics, set have_atomics_support to true.
4073 */
4074 } else if ((adev->flags & AMD_IS_APU) &&
4075 (amdgpu_ip_version(adev, GC_HWIP, 0) >
4076 IP_VERSION(9, 0, 0))) {
4077 adev->have_atomics_support = true;
4078 } else {
4079 adev->have_atomics_support =
4080 !pci_enable_atomic_ops_to_root(adev->pdev,
4081 PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
4082 PCI_EXP_DEVCAP2_ATOMIC_COMP64);
4083 }
4084
4085 if (!adev->have_atomics_support)
4086 dev_info(adev->dev, "PCIE atomic ops is not supported\n");
4087
4088 /* doorbell bar mapping and doorbell index init*/
4089 amdgpu_doorbell_init(adev);
4090
4091 if (amdgpu_emu_mode == 1) {
4092 /* post the asic on emulation mode */
4093 emu_soc_asic_init(adev);
4094 goto fence_driver_init;
4095 }
4096
4097 amdgpu_reset_init(adev);
4098
4099 /* detect if we are with an SRIOV vbios */
4100 if (adev->bios)
4101 amdgpu_device_detect_sriov_bios(adev);
4102
4103 /* check if we need to reset the asic
4104 * E.g., driver was not cleanly unloaded previously, etc.
4105 */
4106 if (!amdgpu_sriov_vf(adev) && amdgpu_asic_need_reset_on_init(adev)) {
4107 if (adev->gmc.xgmi.num_physical_nodes) {
4108 dev_info(adev->dev, "Pending hive reset.\n");
4109 adev->gmc.xgmi.pending_reset = true;
4110 /* Only need to init necessary block for SMU to handle the reset */
4111 for (i = 0; i < adev->num_ip_blocks; i++) {
4112 if (!adev->ip_blocks[i].status.valid)
4113 continue;
4114 if (!(adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
4115 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
4116 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH ||
4117 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC)) {
4118 DRM_DEBUG("IP %s disabled for hw_init.\n",
4119 adev->ip_blocks[i].version->funcs->name);
4120 adev->ip_blocks[i].status.hw = true;
4121 }
4122 }
4123 } else {
4124 tmp = amdgpu_reset_method;
4125 /* It should do a default reset when loading or reloading the driver,
4126 * regardless of the module parameter reset_method.
4127 */
4128 amdgpu_reset_method = AMD_RESET_METHOD_NONE;
4129 r = amdgpu_asic_reset(adev);
4130 amdgpu_reset_method = tmp;
4131 if (r) {
4132 dev_err(adev->dev, "asic reset on init failed\n");
4133 goto failed;
4134 }
4135 }
4136 }
4137
4138 /* Post card if necessary */
4139 if (amdgpu_device_need_post(adev)) {
4140 if (!adev->bios) {
4141 dev_err(adev->dev, "no vBIOS found\n");
4142 r = -EINVAL;
4143 goto failed;
4144 }
4145 DRM_INFO("GPU posting now...\n");
4146 r = amdgpu_device_asic_init(adev);
4147 if (r) {
4148 dev_err(adev->dev, "gpu post error!\n");
4149 goto failed;
4150 }
4151 }
4152
4153 if (adev->bios) {
4154 if (adev->is_atom_fw) {
4155 /* Initialize clocks */
4156 r = amdgpu_atomfirmware_get_clock_info(adev);
4157 if (r) {
4158 dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
4159 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
4160 goto failed;
4161 }
4162 } else {
4163 /* Initialize clocks */
4164 r = amdgpu_atombios_get_clock_info(adev);
4165 if (r) {
4166 dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
4167 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
4168 goto failed;
4169 }
4170 /* init i2c buses */
4171 if (!amdgpu_device_has_dc_support(adev))
4172 amdgpu_atombios_i2c_init(adev);
4173 }
4174 }
4175
4176fence_driver_init:
4177 /* Fence driver */
4178 r = amdgpu_fence_driver_sw_init(adev);
4179 if (r) {
4180 dev_err(adev->dev, "amdgpu_fence_driver_sw_init failed\n");
4181 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
4182 goto failed;
4183 }
4184
4185 /* init the mode config */
4186 drm_mode_config_init(adev_to_drm(adev));
4187
4188 r = amdgpu_device_ip_init(adev);
4189 if (r) {
4190 dev_err(adev->dev, "amdgpu_device_ip_init failed\n");
4191 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
4192 goto release_ras_con;
4193 }
4194
4195 amdgpu_fence_driver_hw_init(adev);
4196
4197 dev_info(adev->dev,
4198 "SE %d, SH per SE %d, CU per SH %d, active_cu_number %d\n",
4199 adev->gfx.config.max_shader_engines,
4200 adev->gfx.config.max_sh_per_se,
4201 adev->gfx.config.max_cu_per_sh,
4202 adev->gfx.cu_info.number);
4203
4204 adev->accel_working = true;
4205
4206 amdgpu_vm_check_compute_bug(adev);
4207
4208 /* Initialize the buffer migration limit. */
4209 if (amdgpu_moverate >= 0)
4210 max_MBps = amdgpu_moverate;
4211 else
4212 max_MBps = 8; /* Allow 8 MB/s. */
4213 /* Get a log2 for easy divisions. */
4214 adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
4215
4216 /*
4217 * Register gpu instance before amdgpu_device_enable_mgpu_fan_boost.
4218 * Otherwise the mgpu fan boost feature will be skipped due to the
4219 * gpu instance is counted less.
4220 */
4221 amdgpu_register_gpu_instance(adev);
4222
4223 /* enable clockgating, etc. after ib tests, etc. since some blocks require
4224 * explicit gating rather than handling it automatically.
4225 */
4226 if (!adev->gmc.xgmi.pending_reset) {
4227 r = amdgpu_device_ip_late_init(adev);
4228 if (r) {
4229 dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n");
4230 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
4231 goto release_ras_con;
4232 }
4233 /* must succeed. */
4234 amdgpu_ras_resume(adev);
4235 queue_delayed_work(system_wq, &adev->delayed_init_work,
4236 msecs_to_jiffies(AMDGPU_RESUME_MS));
4237 }
4238
4239 if (amdgpu_sriov_vf(adev)) {
4240 amdgpu_virt_release_full_gpu(adev, true);
4241 flush_delayed_work(&adev->delayed_init_work);
4242 }
4243
4244 /*
4245 * Place those sysfs registering after `late_init`. As some of those
4246 * operations performed in `late_init` might affect the sysfs
4247 * interfaces creating.
4248 */
4249 r = amdgpu_atombios_sysfs_init(adev);
4250 if (r)
4251 drm_err(&adev->ddev,
4252 "registering atombios sysfs failed (%d).\n", r);
4253
4254 r = amdgpu_pm_sysfs_init(adev);
4255 if (r)
4256 DRM_ERROR("registering pm sysfs failed (%d).\n", r);
4257
4258 r = amdgpu_ucode_sysfs_init(adev);
4259 if (r) {
4260 adev->ucode_sysfs_en = false;
4261 DRM_ERROR("Creating firmware sysfs failed (%d).\n", r);
4262 } else
4263 adev->ucode_sysfs_en = true;
4264
4265 r = sysfs_create_files(&adev->dev->kobj, amdgpu_dev_attributes);
4266 if (r)
4267 dev_err(adev->dev, "Could not create amdgpu device attr\n");
4268
4269 r = devm_device_add_group(adev->dev, &amdgpu_board_attrs_group);
4270 if (r)
4271 dev_err(adev->dev,
4272 "Could not create amdgpu board attributes\n");
4273
4274 amdgpu_fru_sysfs_init(adev);
4275 amdgpu_reg_state_sysfs_init(adev);
4276
4277 if (IS_ENABLED(CONFIG_PERF_EVENTS))
4278 r = amdgpu_pmu_init(adev);
4279 if (r)
4280 dev_err(adev->dev, "amdgpu_pmu_init failed\n");
4281
4282 /* Have stored pci confspace at hand for restore in sudden PCI error */
4283 if (amdgpu_device_cache_pci_state(adev->pdev))
4284 pci_restore_state(pdev);
4285
4286 /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
4287 /* this will fail for cards that aren't VGA class devices, just
4288 * ignore it
4289 */
4290 if ((adev->pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA)
4291 vga_client_register(adev->pdev, amdgpu_device_vga_set_decode);
4292
4293 px = amdgpu_device_supports_px(ddev);
4294
4295 if (px || (!dev_is_removable(&adev->pdev->dev) &&
4296 apple_gmux_detect(NULL, NULL)))
4297 vga_switcheroo_register_client(adev->pdev,
4298 &amdgpu_switcheroo_ops, px);
4299
4300 if (px)
4301 vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
4302
4303 if (adev->gmc.xgmi.pending_reset)
4304 queue_delayed_work(system_wq, &mgpu_info.delayed_reset_work,
4305 msecs_to_jiffies(AMDGPU_RESUME_MS));
4306
4307 amdgpu_device_check_iommu_direct_map(adev);
4308
4309 return 0;
4310
4311release_ras_con:
4312 if (amdgpu_sriov_vf(adev))
4313 amdgpu_virt_release_full_gpu(adev, true);
4314
4315 /* failed in exclusive mode due to timeout */
4316 if (amdgpu_sriov_vf(adev) &&
4317 !amdgpu_sriov_runtime(adev) &&
4318 amdgpu_virt_mmio_blocked(adev) &&
4319 !amdgpu_virt_wait_reset(adev)) {
4320 dev_err(adev->dev, "VF exclusive mode timeout\n");
4321 /* Don't send request since VF is inactive. */
4322 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
4323 adev->virt.ops = NULL;
4324 r = -EAGAIN;
4325 }
4326 amdgpu_release_ras_context(adev);
4327
4328failed:
4329 amdgpu_vf_error_trans_all(adev);
4330
4331 return r;
4332}
4333
4334static void amdgpu_device_unmap_mmio(struct amdgpu_device *adev)
4335{
4336
4337 /* Clear all CPU mappings pointing to this device */
4338 unmap_mapping_range(adev->ddev.anon_inode->i_mapping, 0, 0, 1);
4339
4340 /* Unmap all mapped bars - Doorbell, registers and VRAM */
4341 amdgpu_doorbell_fini(adev);
4342
4343 iounmap(adev->rmmio);
4344 adev->rmmio = NULL;
4345 if (adev->mman.aper_base_kaddr)
4346 iounmap(adev->mman.aper_base_kaddr);
4347 adev->mman.aper_base_kaddr = NULL;
4348
4349 /* Memory manager related */
4350 if (!adev->gmc.xgmi.connected_to_cpu && !adev->gmc.is_app_apu) {
4351 arch_phys_wc_del(adev->gmc.vram_mtrr);
4352 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
4353 }
4354}
4355
4356/**
4357 * amdgpu_device_fini_hw - tear down the driver
4358 *
4359 * @adev: amdgpu_device pointer
4360 *
4361 * Tear down the driver info (all asics).
4362 * Called at driver shutdown.
4363 */
4364void amdgpu_device_fini_hw(struct amdgpu_device *adev)
4365{
4366 dev_info(adev->dev, "amdgpu: finishing device.\n");
4367 flush_delayed_work(&adev->delayed_init_work);
4368 adev->shutdown = true;
4369
4370 /* make sure IB test finished before entering exclusive mode
4371 * to avoid preemption on IB test
4372 */
4373 if (amdgpu_sriov_vf(adev)) {
4374 amdgpu_virt_request_full_gpu(adev, false);
4375 amdgpu_virt_fini_data_exchange(adev);
4376 }
4377
4378 /* disable all interrupts */
4379 amdgpu_irq_disable_all(adev);
4380 if (adev->mode_info.mode_config_initialized) {
4381 if (!drm_drv_uses_atomic_modeset(adev_to_drm(adev)))
4382 drm_helper_force_disable_all(adev_to_drm(adev));
4383 else
4384 drm_atomic_helper_shutdown(adev_to_drm(adev));
4385 }
4386 amdgpu_fence_driver_hw_fini(adev);
4387
4388 if (adev->mman.initialized)
4389 drain_workqueue(adev->mman.bdev.wq);
4390
4391 if (adev->pm.sysfs_initialized)
4392 amdgpu_pm_sysfs_fini(adev);
4393 if (adev->ucode_sysfs_en)
4394 amdgpu_ucode_sysfs_fini(adev);
4395 sysfs_remove_files(&adev->dev->kobj, amdgpu_dev_attributes);
4396 amdgpu_fru_sysfs_fini(adev);
4397
4398 amdgpu_reg_state_sysfs_fini(adev);
4399
4400 /* disable ras feature must before hw fini */
4401 amdgpu_ras_pre_fini(adev);
4402
4403 amdgpu_ttm_set_buffer_funcs_status(adev, false);
4404
4405 amdgpu_device_ip_fini_early(adev);
4406
4407 amdgpu_irq_fini_hw(adev);
4408
4409 if (adev->mman.initialized)
4410 ttm_device_clear_dma_mappings(&adev->mman.bdev);
4411
4412 amdgpu_gart_dummy_page_fini(adev);
4413
4414 if (drm_dev_is_unplugged(adev_to_drm(adev)))
4415 amdgpu_device_unmap_mmio(adev);
4416
4417}
4418
4419void amdgpu_device_fini_sw(struct amdgpu_device *adev)
4420{
4421 int idx;
4422 bool px;
4423
4424 amdgpu_fence_driver_sw_fini(adev);
4425 amdgpu_device_ip_fini(adev);
4426 amdgpu_ucode_release(&adev->firmware.gpu_info_fw);
4427 adev->accel_working = false;
4428 dma_fence_put(rcu_dereference_protected(adev->gang_submit, true));
4429
4430 amdgpu_reset_fini(adev);
4431
4432 /* free i2c buses */
4433 if (!amdgpu_device_has_dc_support(adev))
4434 amdgpu_i2c_fini(adev);
4435
4436 if (amdgpu_emu_mode != 1)
4437 amdgpu_atombios_fini(adev);
4438
4439 kfree(adev->bios);
4440 adev->bios = NULL;
4441
4442 kfree(adev->fru_info);
4443 adev->fru_info = NULL;
4444
4445 px = amdgpu_device_supports_px(adev_to_drm(adev));
4446
4447 if (px || (!dev_is_removable(&adev->pdev->dev) &&
4448 apple_gmux_detect(NULL, NULL)))
4449 vga_switcheroo_unregister_client(adev->pdev);
4450
4451 if (px)
4452 vga_switcheroo_fini_domain_pm_ops(adev->dev);
4453
4454 if ((adev->pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA)
4455 vga_client_unregister(adev->pdev);
4456
4457 if (drm_dev_enter(adev_to_drm(adev), &idx)) {
4458
4459 iounmap(adev->rmmio);
4460 adev->rmmio = NULL;
4461 amdgpu_doorbell_fini(adev);
4462 drm_dev_exit(idx);
4463 }
4464
4465 if (IS_ENABLED(CONFIG_PERF_EVENTS))
4466 amdgpu_pmu_fini(adev);
4467 if (adev->mman.discovery_bin)
4468 amdgpu_discovery_fini(adev);
4469
4470 amdgpu_reset_put_reset_domain(adev->reset_domain);
4471 adev->reset_domain = NULL;
4472
4473 kfree(adev->pci_state);
4474
4475}
4476
4477/**
4478 * amdgpu_device_evict_resources - evict device resources
4479 * @adev: amdgpu device object
4480 *
4481 * Evicts all ttm device resources(vram BOs, gart table) from the lru list
4482 * of the vram memory type. Mainly used for evicting device resources
4483 * at suspend time.
4484 *
4485 */
4486static int amdgpu_device_evict_resources(struct amdgpu_device *adev)
4487{
4488 int ret;
4489
4490 /* No need to evict vram on APUs for suspend to ram or s2idle */
4491 if ((adev->in_s3 || adev->in_s0ix) && (adev->flags & AMD_IS_APU))
4492 return 0;
4493
4494 ret = amdgpu_ttm_evict_resources(adev, TTM_PL_VRAM);
4495 if (ret)
4496 DRM_WARN("evicting device resources failed\n");
4497 return ret;
4498}
4499
4500/*
4501 * Suspend & resume.
4502 */
4503/**
4504 * amdgpu_device_prepare - prepare for device suspend
4505 *
4506 * @dev: drm dev pointer
4507 *
4508 * Prepare to put the hw in the suspend state (all asics).
4509 * Returns 0 for success or an error on failure.
4510 * Called at driver suspend.
4511 */
4512int amdgpu_device_prepare(struct drm_device *dev)
4513{
4514 struct amdgpu_device *adev = drm_to_adev(dev);
4515 int i, r;
4516
4517 amdgpu_choose_low_power_state(adev);
4518
4519 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
4520 return 0;
4521
4522 /* Evict the majority of BOs before starting suspend sequence */
4523 r = amdgpu_device_evict_resources(adev);
4524 if (r)
4525 goto unprepare;
4526
4527 for (i = 0; i < adev->num_ip_blocks; i++) {
4528 if (!adev->ip_blocks[i].status.valid)
4529 continue;
4530 if (!adev->ip_blocks[i].version->funcs->prepare_suspend)
4531 continue;
4532 r = adev->ip_blocks[i].version->funcs->prepare_suspend((void *)adev);
4533 if (r)
4534 goto unprepare;
4535 }
4536
4537 return 0;
4538
4539unprepare:
4540 adev->in_s0ix = adev->in_s3 = false;
4541
4542 return r;
4543}
4544
4545/**
4546 * amdgpu_device_suspend - initiate device suspend
4547 *
4548 * @dev: drm dev pointer
4549 * @fbcon : notify the fbdev of suspend
4550 *
4551 * Puts the hw in the suspend state (all asics).
4552 * Returns 0 for success or an error on failure.
4553 * Called at driver suspend.
4554 */
4555int amdgpu_device_suspend(struct drm_device *dev, bool fbcon)
4556{
4557 struct amdgpu_device *adev = drm_to_adev(dev);
4558 int r = 0;
4559
4560 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
4561 return 0;
4562
4563 adev->in_suspend = true;
4564
4565 if (amdgpu_sriov_vf(adev)) {
4566 amdgpu_virt_fini_data_exchange(adev);
4567 r = amdgpu_virt_request_full_gpu(adev, false);
4568 if (r)
4569 return r;
4570 }
4571
4572 if (amdgpu_acpi_smart_shift_update(dev, AMDGPU_SS_DEV_D3))
4573 DRM_WARN("smart shift update failed\n");
4574
4575 if (fbcon)
4576 drm_fb_helper_set_suspend_unlocked(adev_to_drm(adev)->fb_helper, true);
4577
4578 cancel_delayed_work_sync(&adev->delayed_init_work);
4579
4580 amdgpu_ras_suspend(adev);
4581
4582 amdgpu_device_ip_suspend_phase1(adev);
4583
4584 if (!adev->in_s0ix)
4585 amdgpu_amdkfd_suspend(adev, adev->in_runpm);
4586
4587 r = amdgpu_device_evict_resources(adev);
4588 if (r)
4589 return r;
4590
4591 amdgpu_ttm_set_buffer_funcs_status(adev, false);
4592
4593 amdgpu_fence_driver_hw_fini(adev);
4594
4595 amdgpu_device_ip_suspend_phase2(adev);
4596
4597 if (amdgpu_sriov_vf(adev))
4598 amdgpu_virt_release_full_gpu(adev, false);
4599
4600 r = amdgpu_dpm_notify_rlc_state(adev, false);
4601 if (r)
4602 return r;
4603
4604 return 0;
4605}
4606
4607/**
4608 * amdgpu_device_resume - initiate device resume
4609 *
4610 * @dev: drm dev pointer
4611 * @fbcon : notify the fbdev of resume
4612 *
4613 * Bring the hw back to operating state (all asics).
4614 * Returns 0 for success or an error on failure.
4615 * Called at driver resume.
4616 */
4617int amdgpu_device_resume(struct drm_device *dev, bool fbcon)
4618{
4619 struct amdgpu_device *adev = drm_to_adev(dev);
4620 int r = 0;
4621
4622 if (amdgpu_sriov_vf(adev)) {
4623 r = amdgpu_virt_request_full_gpu(adev, true);
4624 if (r)
4625 return r;
4626 }
4627
4628 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
4629 return 0;
4630
4631 if (adev->in_s0ix)
4632 amdgpu_dpm_gfx_state_change(adev, sGpuChangeState_D0Entry);
4633
4634 /* post card */
4635 if (amdgpu_device_need_post(adev)) {
4636 r = amdgpu_device_asic_init(adev);
4637 if (r)
4638 dev_err(adev->dev, "amdgpu asic init failed\n");
4639 }
4640
4641 r = amdgpu_device_ip_resume(adev);
4642
4643 if (r) {
4644 dev_err(adev->dev, "amdgpu_device_ip_resume failed (%d).\n", r);
4645 goto exit;
4646 }
4647 amdgpu_fence_driver_hw_init(adev);
4648
4649 if (!adev->in_s0ix) {
4650 r = amdgpu_amdkfd_resume(adev, adev->in_runpm);
4651 if (r)
4652 goto exit;
4653 }
4654
4655 r = amdgpu_device_ip_late_init(adev);
4656 if (r)
4657 goto exit;
4658
4659 queue_delayed_work(system_wq, &adev->delayed_init_work,
4660 msecs_to_jiffies(AMDGPU_RESUME_MS));
4661exit:
4662 if (amdgpu_sriov_vf(adev)) {
4663 amdgpu_virt_init_data_exchange(adev);
4664 amdgpu_virt_release_full_gpu(adev, true);
4665 }
4666
4667 if (r)
4668 return r;
4669
4670 /* Make sure IB tests flushed */
4671 flush_delayed_work(&adev->delayed_init_work);
4672
4673 if (fbcon)
4674 drm_fb_helper_set_suspend_unlocked(adev_to_drm(adev)->fb_helper, false);
4675
4676 amdgpu_ras_resume(adev);
4677
4678 if (adev->mode_info.num_crtc) {
4679 /*
4680 * Most of the connector probing functions try to acquire runtime pm
4681 * refs to ensure that the GPU is powered on when connector polling is
4682 * performed. Since we're calling this from a runtime PM callback,
4683 * trying to acquire rpm refs will cause us to deadlock.
4684 *
4685 * Since we're guaranteed to be holding the rpm lock, it's safe to
4686 * temporarily disable the rpm helpers so this doesn't deadlock us.
4687 */
4688#ifdef CONFIG_PM
4689 dev->dev->power.disable_depth++;
4690#endif
4691 if (!adev->dc_enabled)
4692 drm_helper_hpd_irq_event(dev);
4693 else
4694 drm_kms_helper_hotplug_event(dev);
4695#ifdef CONFIG_PM
4696 dev->dev->power.disable_depth--;
4697#endif
4698 }
4699 adev->in_suspend = false;
4700
4701 if (adev->enable_mes)
4702 amdgpu_mes_self_test(adev);
4703
4704 if (amdgpu_acpi_smart_shift_update(dev, AMDGPU_SS_DEV_D0))
4705 DRM_WARN("smart shift update failed\n");
4706
4707 return 0;
4708}
4709
4710/**
4711 * amdgpu_device_ip_check_soft_reset - did soft reset succeed
4712 *
4713 * @adev: amdgpu_device pointer
4714 *
4715 * The list of all the hardware IPs that make up the asic is walked and
4716 * the check_soft_reset callbacks are run. check_soft_reset determines
4717 * if the asic is still hung or not.
4718 * Returns true if any of the IPs are still in a hung state, false if not.
4719 */
4720static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)
4721{
4722 int i;
4723 bool asic_hang = false;
4724
4725 if (amdgpu_sriov_vf(adev))
4726 return true;
4727
4728 if (amdgpu_asic_need_full_reset(adev))
4729 return true;
4730
4731 for (i = 0; i < adev->num_ip_blocks; i++) {
4732 if (!adev->ip_blocks[i].status.valid)
4733 continue;
4734 if (adev->ip_blocks[i].version->funcs->check_soft_reset)
4735 adev->ip_blocks[i].status.hang =
4736 adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
4737 if (adev->ip_blocks[i].status.hang) {
4738 dev_info(adev->dev, "IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
4739 asic_hang = true;
4740 }
4741 }
4742 return asic_hang;
4743}
4744
4745/**
4746 * amdgpu_device_ip_pre_soft_reset - prepare for soft reset
4747 *
4748 * @adev: amdgpu_device pointer
4749 *
4750 * The list of all the hardware IPs that make up the asic is walked and the
4751 * pre_soft_reset callbacks are run if the block is hung. pre_soft_reset
4752 * handles any IP specific hardware or software state changes that are
4753 * necessary for a soft reset to succeed.
4754 * Returns 0 on success, negative error code on failure.
4755 */
4756static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)
4757{
4758 int i, r = 0;
4759
4760 for (i = 0; i < adev->num_ip_blocks; i++) {
4761 if (!adev->ip_blocks[i].status.valid)
4762 continue;
4763 if (adev->ip_blocks[i].status.hang &&
4764 adev->ip_blocks[i].version->funcs->pre_soft_reset) {
4765 r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
4766 if (r)
4767 return r;
4768 }
4769 }
4770
4771 return 0;
4772}
4773
4774/**
4775 * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed
4776 *
4777 * @adev: amdgpu_device pointer
4778 *
4779 * Some hardware IPs cannot be soft reset. If they are hung, a full gpu
4780 * reset is necessary to recover.
4781 * Returns true if a full asic reset is required, false if not.
4782 */
4783static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)
4784{
4785 int i;
4786
4787 if (amdgpu_asic_need_full_reset(adev))
4788 return true;
4789
4790 for (i = 0; i < adev->num_ip_blocks; i++) {
4791 if (!adev->ip_blocks[i].status.valid)
4792 continue;
4793 if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
4794 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
4795 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
4796 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
4797 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
4798 if (adev->ip_blocks[i].status.hang) {
4799 dev_info(adev->dev, "Some block need full reset!\n");
4800 return true;
4801 }
4802 }
4803 }
4804 return false;
4805}
4806
4807/**
4808 * amdgpu_device_ip_soft_reset - do a soft reset
4809 *
4810 * @adev: amdgpu_device pointer
4811 *
4812 * The list of all the hardware IPs that make up the asic is walked and the
4813 * soft_reset callbacks are run if the block is hung. soft_reset handles any
4814 * IP specific hardware or software state changes that are necessary to soft
4815 * reset the IP.
4816 * Returns 0 on success, negative error code on failure.
4817 */
4818static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)
4819{
4820 int i, r = 0;
4821
4822 for (i = 0; i < adev->num_ip_blocks; i++) {
4823 if (!adev->ip_blocks[i].status.valid)
4824 continue;
4825 if (adev->ip_blocks[i].status.hang &&
4826 adev->ip_blocks[i].version->funcs->soft_reset) {
4827 r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
4828 if (r)
4829 return r;
4830 }
4831 }
4832
4833 return 0;
4834}
4835
4836/**
4837 * amdgpu_device_ip_post_soft_reset - clean up from soft reset
4838 *
4839 * @adev: amdgpu_device pointer
4840 *
4841 * The list of all the hardware IPs that make up the asic is walked and the
4842 * post_soft_reset callbacks are run if the asic was hung. post_soft_reset
4843 * handles any IP specific hardware or software state changes that are
4844 * necessary after the IP has been soft reset.
4845 * Returns 0 on success, negative error code on failure.
4846 */
4847static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)
4848{
4849 int i, r = 0;
4850
4851 for (i = 0; i < adev->num_ip_blocks; i++) {
4852 if (!adev->ip_blocks[i].status.valid)
4853 continue;
4854 if (adev->ip_blocks[i].status.hang &&
4855 adev->ip_blocks[i].version->funcs->post_soft_reset)
4856 r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
4857 if (r)
4858 return r;
4859 }
4860
4861 return 0;
4862}
4863
4864/**
4865 * amdgpu_device_recover_vram - Recover some VRAM contents
4866 *
4867 * @adev: amdgpu_device pointer
4868 *
4869 * Restores the contents of VRAM buffers from the shadows in GTT. Used to
4870 * restore things like GPUVM page tables after a GPU reset where
4871 * the contents of VRAM might be lost.
4872 *
4873 * Returns:
4874 * 0 on success, negative error code on failure.
4875 */
4876static int amdgpu_device_recover_vram(struct amdgpu_device *adev)
4877{
4878 struct dma_fence *fence = NULL, *next = NULL;
4879 struct amdgpu_bo *shadow;
4880 struct amdgpu_bo_vm *vmbo;
4881 long r = 1, tmo;
4882
4883 if (amdgpu_sriov_runtime(adev))
4884 tmo = msecs_to_jiffies(8000);
4885 else
4886 tmo = msecs_to_jiffies(100);
4887
4888 dev_info(adev->dev, "recover vram bo from shadow start\n");
4889 mutex_lock(&adev->shadow_list_lock);
4890 list_for_each_entry(vmbo, &adev->shadow_list, shadow_list) {
4891 /* If vm is compute context or adev is APU, shadow will be NULL */
4892 if (!vmbo->shadow)
4893 continue;
4894 shadow = vmbo->shadow;
4895
4896 /* No need to recover an evicted BO */
4897 if (shadow->tbo.resource->mem_type != TTM_PL_TT ||
4898 shadow->tbo.resource->start == AMDGPU_BO_INVALID_OFFSET ||
4899 shadow->parent->tbo.resource->mem_type != TTM_PL_VRAM)
4900 continue;
4901
4902 r = amdgpu_bo_restore_shadow(shadow, &next);
4903 if (r)
4904 break;
4905
4906 if (fence) {
4907 tmo = dma_fence_wait_timeout(fence, false, tmo);
4908 dma_fence_put(fence);
4909 fence = next;
4910 if (tmo == 0) {
4911 r = -ETIMEDOUT;
4912 break;
4913 } else if (tmo < 0) {
4914 r = tmo;
4915 break;
4916 }
4917 } else {
4918 fence = next;
4919 }
4920 }
4921 mutex_unlock(&adev->shadow_list_lock);
4922
4923 if (fence)
4924 tmo = dma_fence_wait_timeout(fence, false, tmo);
4925 dma_fence_put(fence);
4926
4927 if (r < 0 || tmo <= 0) {
4928 dev_err(adev->dev, "recover vram bo from shadow failed, r is %ld, tmo is %ld\n", r, tmo);
4929 return -EIO;
4930 }
4931
4932 dev_info(adev->dev, "recover vram bo from shadow done\n");
4933 return 0;
4934}
4935
4936
4937/**
4938 * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
4939 *
4940 * @adev: amdgpu_device pointer
4941 * @from_hypervisor: request from hypervisor
4942 *
4943 * do VF FLR and reinitialize Asic
4944 * return 0 means succeeded otherwise failed
4945 */
4946static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
4947 bool from_hypervisor)
4948{
4949 int r;
4950 struct amdgpu_hive_info *hive = NULL;
4951 int retry_limit = 0;
4952
4953retry:
4954 amdgpu_amdkfd_pre_reset(adev);
4955
4956 if (from_hypervisor)
4957 r = amdgpu_virt_request_full_gpu(adev, true);
4958 else
4959 r = amdgpu_virt_reset_gpu(adev);
4960 if (r)
4961 return r;
4962 amdgpu_irq_gpu_reset_resume_helper(adev);
4963
4964 /* some sw clean up VF needs to do before recover */
4965 amdgpu_virt_post_reset(adev);
4966
4967 /* Resume IP prior to SMC */
4968 r = amdgpu_device_ip_reinit_early_sriov(adev);
4969 if (r)
4970 goto error;
4971
4972 amdgpu_virt_init_data_exchange(adev);
4973
4974 r = amdgpu_device_fw_loading(adev);
4975 if (r)
4976 return r;
4977
4978 /* now we are okay to resume SMC/CP/SDMA */
4979 r = amdgpu_device_ip_reinit_late_sriov(adev);
4980 if (r)
4981 goto error;
4982
4983 hive = amdgpu_get_xgmi_hive(adev);
4984 /* Update PSP FW topology after reset */
4985 if (hive && adev->gmc.xgmi.num_physical_nodes > 1)
4986 r = amdgpu_xgmi_update_topology(hive, adev);
4987
4988 if (hive)
4989 amdgpu_put_xgmi_hive(hive);
4990
4991 if (!r) {
4992 r = amdgpu_ib_ring_tests(adev);
4993
4994 amdgpu_amdkfd_post_reset(adev);
4995 }
4996
4997error:
4998 if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
4999 amdgpu_inc_vram_lost(adev);
5000 r = amdgpu_device_recover_vram(adev);
5001 }
5002 amdgpu_virt_release_full_gpu(adev, true);
5003
5004 if (AMDGPU_RETRY_SRIOV_RESET(r)) {
5005 if (retry_limit < AMDGPU_MAX_RETRY_LIMIT) {
5006 retry_limit++;
5007 goto retry;
5008 } else
5009 DRM_ERROR("GPU reset retry is beyond the retry limit\n");
5010 }
5011
5012 return r;
5013}
5014
5015/**
5016 * amdgpu_device_has_job_running - check if there is any job in mirror list
5017 *
5018 * @adev: amdgpu_device pointer
5019 *
5020 * check if there is any job in mirror list
5021 */
5022bool amdgpu_device_has_job_running(struct amdgpu_device *adev)
5023{
5024 int i;
5025 struct drm_sched_job *job;
5026
5027 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
5028 struct amdgpu_ring *ring = adev->rings[i];
5029
5030 if (!amdgpu_ring_sched_ready(ring))
5031 continue;
5032
5033 spin_lock(&ring->sched.job_list_lock);
5034 job = list_first_entry_or_null(&ring->sched.pending_list,
5035 struct drm_sched_job, list);
5036 spin_unlock(&ring->sched.job_list_lock);
5037 if (job)
5038 return true;
5039 }
5040 return false;
5041}
5042
5043/**
5044 * amdgpu_device_should_recover_gpu - check if we should try GPU recovery
5045 *
5046 * @adev: amdgpu_device pointer
5047 *
5048 * Check amdgpu_gpu_recovery and SRIOV status to see if we should try to recover
5049 * a hung GPU.
5050 */
5051bool amdgpu_device_should_recover_gpu(struct amdgpu_device *adev)
5052{
5053
5054 if (amdgpu_gpu_recovery == 0)
5055 goto disabled;
5056
5057 /* Skip soft reset check in fatal error mode */
5058 if (!amdgpu_ras_is_poison_mode_supported(adev))
5059 return true;
5060
5061 if (amdgpu_sriov_vf(adev))
5062 return true;
5063
5064 if (amdgpu_gpu_recovery == -1) {
5065 switch (adev->asic_type) {
5066#ifdef CONFIG_DRM_AMDGPU_SI
5067 case CHIP_VERDE:
5068 case CHIP_TAHITI:
5069 case CHIP_PITCAIRN:
5070 case CHIP_OLAND:
5071 case CHIP_HAINAN:
5072#endif
5073#ifdef CONFIG_DRM_AMDGPU_CIK
5074 case CHIP_KAVERI:
5075 case CHIP_KABINI:
5076 case CHIP_MULLINS:
5077#endif
5078 case CHIP_CARRIZO:
5079 case CHIP_STONEY:
5080 case CHIP_CYAN_SKILLFISH:
5081 goto disabled;
5082 default:
5083 break;
5084 }
5085 }
5086
5087 return true;
5088
5089disabled:
5090 dev_info(adev->dev, "GPU recovery disabled.\n");
5091 return false;
5092}
5093
5094int amdgpu_device_mode1_reset(struct amdgpu_device *adev)
5095{
5096 u32 i;
5097 int ret = 0;
5098
5099 amdgpu_atombios_scratch_regs_engine_hung(adev, true);
5100
5101 dev_info(adev->dev, "GPU mode1 reset\n");
5102
5103 /* disable BM */
5104 pci_clear_master(adev->pdev);
5105
5106 amdgpu_device_cache_pci_state(adev->pdev);
5107
5108 if (amdgpu_dpm_is_mode1_reset_supported(adev)) {
5109 dev_info(adev->dev, "GPU smu mode1 reset\n");
5110 ret = amdgpu_dpm_mode1_reset(adev);
5111 } else {
5112 dev_info(adev->dev, "GPU psp mode1 reset\n");
5113 ret = psp_gpu_reset(adev);
5114 }
5115
5116 if (ret)
5117 goto mode1_reset_failed;
5118
5119 amdgpu_device_load_pci_state(adev->pdev);
5120 ret = amdgpu_psp_wait_for_bootloader(adev);
5121 if (ret)
5122 goto mode1_reset_failed;
5123
5124 /* wait for asic to come out of reset */
5125 for (i = 0; i < adev->usec_timeout; i++) {
5126 u32 memsize = adev->nbio.funcs->get_memsize(adev);
5127
5128 if (memsize != 0xffffffff)
5129 break;
5130 udelay(1);
5131 }
5132
5133 if (i >= adev->usec_timeout) {
5134 ret = -ETIMEDOUT;
5135 goto mode1_reset_failed;
5136 }
5137
5138 amdgpu_atombios_scratch_regs_engine_hung(adev, false);
5139
5140 return 0;
5141
5142mode1_reset_failed:
5143 dev_err(adev->dev, "GPU mode1 reset failed\n");
5144 return ret;
5145}
5146
5147int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,
5148 struct amdgpu_reset_context *reset_context)
5149{
5150 int i, r = 0;
5151 struct amdgpu_job *job = NULL;
5152 bool need_full_reset =
5153 test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
5154
5155 if (reset_context->reset_req_dev == adev)
5156 job = reset_context->job;
5157
5158 if (amdgpu_sriov_vf(adev)) {
5159 /* stop the data exchange thread */
5160 amdgpu_virt_fini_data_exchange(adev);
5161 }
5162
5163 amdgpu_fence_driver_isr_toggle(adev, true);
5164
5165 /* block all schedulers and reset given job's ring */
5166 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
5167 struct amdgpu_ring *ring = adev->rings[i];
5168
5169 if (!amdgpu_ring_sched_ready(ring))
5170 continue;
5171
5172 /* Clear job fence from fence drv to avoid force_completion
5173 * leave NULL and vm flush fence in fence drv
5174 */
5175 amdgpu_fence_driver_clear_job_fences(ring);
5176
5177 /* after all hw jobs are reset, hw fence is meaningless, so force_completion */
5178 amdgpu_fence_driver_force_completion(ring);
5179 }
5180
5181 amdgpu_fence_driver_isr_toggle(adev, false);
5182
5183 if (job && job->vm)
5184 drm_sched_increase_karma(&job->base);
5185
5186 r = amdgpu_reset_prepare_hwcontext(adev, reset_context);
5187 /* If reset handler not implemented, continue; otherwise return */
5188 if (r == -EOPNOTSUPP)
5189 r = 0;
5190 else
5191 return r;
5192
5193 /* Don't suspend on bare metal if we are not going to HW reset the ASIC */
5194 if (!amdgpu_sriov_vf(adev)) {
5195
5196 if (!need_full_reset)
5197 need_full_reset = amdgpu_device_ip_need_full_reset(adev);
5198
5199 if (!need_full_reset && amdgpu_gpu_recovery &&
5200 amdgpu_device_ip_check_soft_reset(adev)) {
5201 amdgpu_device_ip_pre_soft_reset(adev);
5202 r = amdgpu_device_ip_soft_reset(adev);
5203 amdgpu_device_ip_post_soft_reset(adev);
5204 if (r || amdgpu_device_ip_check_soft_reset(adev)) {
5205 dev_info(adev->dev, "soft reset failed, will fallback to full reset!\n");
5206 need_full_reset = true;
5207 }
5208 }
5209
5210 if (need_full_reset)
5211 r = amdgpu_device_ip_suspend(adev);
5212 if (need_full_reset)
5213 set_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
5214 else
5215 clear_bit(AMDGPU_NEED_FULL_RESET,
5216 &reset_context->flags);
5217 }
5218
5219 return r;
5220}
5221
5222static int amdgpu_reset_reg_dumps(struct amdgpu_device *adev)
5223{
5224 int i;
5225
5226 lockdep_assert_held(&adev->reset_domain->sem);
5227
5228 for (i = 0; i < adev->reset_info.num_regs; i++) {
5229 adev->reset_info.reset_dump_reg_value[i] =
5230 RREG32(adev->reset_info.reset_dump_reg_list[i]);
5231
5232 trace_amdgpu_reset_reg_dumps(adev->reset_info.reset_dump_reg_list[i],
5233 adev->reset_info.reset_dump_reg_value[i]);
5234 }
5235
5236 return 0;
5237}
5238
5239int amdgpu_do_asic_reset(struct list_head *device_list_handle,
5240 struct amdgpu_reset_context *reset_context)
5241{
5242 struct amdgpu_device *tmp_adev = NULL;
5243 bool need_full_reset, skip_hw_reset, vram_lost = false;
5244 int r = 0;
5245
5246 /* Try reset handler method first */
5247 tmp_adev = list_first_entry(device_list_handle, struct amdgpu_device,
5248 reset_list);
5249 amdgpu_reset_reg_dumps(tmp_adev);
5250
5251 reset_context->reset_device_list = device_list_handle;
5252 r = amdgpu_reset_perform_reset(tmp_adev, reset_context);
5253 /* If reset handler not implemented, continue; otherwise return */
5254 if (r == -EOPNOTSUPP)
5255 r = 0;
5256 else
5257 return r;
5258
5259 /* Reset handler not implemented, use the default method */
5260 need_full_reset =
5261 test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
5262 skip_hw_reset = test_bit(AMDGPU_SKIP_HW_RESET, &reset_context->flags);
5263
5264 /*
5265 * ASIC reset has to be done on all XGMI hive nodes ASAP
5266 * to allow proper links negotiation in FW (within 1 sec)
5267 */
5268 if (!skip_hw_reset && need_full_reset) {
5269 list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5270 /* For XGMI run all resets in parallel to speed up the process */
5271 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
5272 tmp_adev->gmc.xgmi.pending_reset = false;
5273 if (!queue_work(system_unbound_wq, &tmp_adev->xgmi_reset_work))
5274 r = -EALREADY;
5275 } else
5276 r = amdgpu_asic_reset(tmp_adev);
5277
5278 if (r) {
5279 dev_err(tmp_adev->dev, "ASIC reset failed with error, %d for drm dev, %s",
5280 r, adev_to_drm(tmp_adev)->unique);
5281 goto out;
5282 }
5283 }
5284
5285 /* For XGMI wait for all resets to complete before proceed */
5286 if (!r) {
5287 list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5288 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
5289 flush_work(&tmp_adev->xgmi_reset_work);
5290 r = tmp_adev->asic_reset_res;
5291 if (r)
5292 break;
5293 }
5294 }
5295 }
5296 }
5297
5298 if (!r && amdgpu_ras_intr_triggered()) {
5299 list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5300 amdgpu_ras_reset_error_count(tmp_adev, AMDGPU_RAS_BLOCK__MMHUB);
5301 }
5302
5303 amdgpu_ras_intr_cleared();
5304 }
5305
5306 list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5307 if (need_full_reset) {
5308 /* post card */
5309 r = amdgpu_device_asic_init(tmp_adev);
5310 if (r) {
5311 dev_warn(tmp_adev->dev, "asic atom init failed!");
5312 } else {
5313 dev_info(tmp_adev->dev, "GPU reset succeeded, trying to resume\n");
5314
5315 r = amdgpu_device_ip_resume_phase1(tmp_adev);
5316 if (r)
5317 goto out;
5318
5319 vram_lost = amdgpu_device_check_vram_lost(tmp_adev);
5320
5321 amdgpu_coredump(tmp_adev, vram_lost, reset_context);
5322
5323 if (vram_lost) {
5324 DRM_INFO("VRAM is lost due to GPU reset!\n");
5325 amdgpu_inc_vram_lost(tmp_adev);
5326 }
5327
5328 r = amdgpu_device_fw_loading(tmp_adev);
5329 if (r)
5330 return r;
5331
5332 r = amdgpu_xcp_restore_partition_mode(
5333 tmp_adev->xcp_mgr);
5334 if (r)
5335 goto out;
5336
5337 r = amdgpu_device_ip_resume_phase2(tmp_adev);
5338 if (r)
5339 goto out;
5340
5341 if (tmp_adev->mman.buffer_funcs_ring->sched.ready)
5342 amdgpu_ttm_set_buffer_funcs_status(tmp_adev, true);
5343
5344 if (vram_lost)
5345 amdgpu_device_fill_reset_magic(tmp_adev);
5346
5347 /*
5348 * Add this ASIC as tracked as reset was already
5349 * complete successfully.
5350 */
5351 amdgpu_register_gpu_instance(tmp_adev);
5352
5353 if (!reset_context->hive &&
5354 tmp_adev->gmc.xgmi.num_physical_nodes > 1)
5355 amdgpu_xgmi_add_device(tmp_adev);
5356
5357 r = amdgpu_device_ip_late_init(tmp_adev);
5358 if (r)
5359 goto out;
5360
5361 drm_fb_helper_set_suspend_unlocked(adev_to_drm(tmp_adev)->fb_helper, false);
5362
5363 /*
5364 * The GPU enters bad state once faulty pages
5365 * by ECC has reached the threshold, and ras
5366 * recovery is scheduled next. So add one check
5367 * here to break recovery if it indeed exceeds
5368 * bad page threshold, and remind user to
5369 * retire this GPU or setting one bigger
5370 * bad_page_threshold value to fix this once
5371 * probing driver again.
5372 */
5373 if (!amdgpu_ras_eeprom_check_err_threshold(tmp_adev)) {
5374 /* must succeed. */
5375 amdgpu_ras_resume(tmp_adev);
5376 } else {
5377 r = -EINVAL;
5378 goto out;
5379 }
5380
5381 /* Update PSP FW topology after reset */
5382 if (reset_context->hive &&
5383 tmp_adev->gmc.xgmi.num_physical_nodes > 1)
5384 r = amdgpu_xgmi_update_topology(
5385 reset_context->hive, tmp_adev);
5386 }
5387 }
5388
5389out:
5390 if (!r) {
5391 amdgpu_irq_gpu_reset_resume_helper(tmp_adev);
5392 r = amdgpu_ib_ring_tests(tmp_adev);
5393 if (r) {
5394 dev_err(tmp_adev->dev, "ib ring test failed (%d).\n", r);
5395 need_full_reset = true;
5396 r = -EAGAIN;
5397 goto end;
5398 }
5399 }
5400
5401 if (!r)
5402 r = amdgpu_device_recover_vram(tmp_adev);
5403 else
5404 tmp_adev->asic_reset_res = r;
5405 }
5406
5407end:
5408 if (need_full_reset)
5409 set_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
5410 else
5411 clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
5412 return r;
5413}
5414
5415static void amdgpu_device_set_mp1_state(struct amdgpu_device *adev)
5416{
5417
5418 switch (amdgpu_asic_reset_method(adev)) {
5419 case AMD_RESET_METHOD_MODE1:
5420 adev->mp1_state = PP_MP1_STATE_SHUTDOWN;
5421 break;
5422 case AMD_RESET_METHOD_MODE2:
5423 adev->mp1_state = PP_MP1_STATE_RESET;
5424 break;
5425 default:
5426 adev->mp1_state = PP_MP1_STATE_NONE;
5427 break;
5428 }
5429}
5430
5431static void amdgpu_device_unset_mp1_state(struct amdgpu_device *adev)
5432{
5433 amdgpu_vf_error_trans_all(adev);
5434 adev->mp1_state = PP_MP1_STATE_NONE;
5435}
5436
5437static void amdgpu_device_resume_display_audio(struct amdgpu_device *adev)
5438{
5439 struct pci_dev *p = NULL;
5440
5441 p = pci_get_domain_bus_and_slot(pci_domain_nr(adev->pdev->bus),
5442 adev->pdev->bus->number, 1);
5443 if (p) {
5444 pm_runtime_enable(&(p->dev));
5445 pm_runtime_resume(&(p->dev));
5446 }
5447
5448 pci_dev_put(p);
5449}
5450
5451static int amdgpu_device_suspend_display_audio(struct amdgpu_device *adev)
5452{
5453 enum amd_reset_method reset_method;
5454 struct pci_dev *p = NULL;
5455 u64 expires;
5456
5457 /*
5458 * For now, only BACO and mode1 reset are confirmed
5459 * to suffer the audio issue without proper suspended.
5460 */
5461 reset_method = amdgpu_asic_reset_method(adev);
5462 if ((reset_method != AMD_RESET_METHOD_BACO) &&
5463 (reset_method != AMD_RESET_METHOD_MODE1))
5464 return -EINVAL;
5465
5466 p = pci_get_domain_bus_and_slot(pci_domain_nr(adev->pdev->bus),
5467 adev->pdev->bus->number, 1);
5468 if (!p)
5469 return -ENODEV;
5470
5471 expires = pm_runtime_autosuspend_expiration(&(p->dev));
5472 if (!expires)
5473 /*
5474 * If we cannot get the audio device autosuspend delay,
5475 * a fixed 4S interval will be used. Considering 3S is
5476 * the audio controller default autosuspend delay setting.
5477 * 4S used here is guaranteed to cover that.
5478 */
5479 expires = ktime_get_mono_fast_ns() + NSEC_PER_SEC * 4ULL;
5480
5481 while (!pm_runtime_status_suspended(&(p->dev))) {
5482 if (!pm_runtime_suspend(&(p->dev)))
5483 break;
5484
5485 if (expires < ktime_get_mono_fast_ns()) {
5486 dev_warn(adev->dev, "failed to suspend display audio\n");
5487 pci_dev_put(p);
5488 /* TODO: abort the succeeding gpu reset? */
5489 return -ETIMEDOUT;
5490 }
5491 }
5492
5493 pm_runtime_disable(&(p->dev));
5494
5495 pci_dev_put(p);
5496 return 0;
5497}
5498
5499static inline void amdgpu_device_stop_pending_resets(struct amdgpu_device *adev)
5500{
5501 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
5502
5503#if defined(CONFIG_DEBUG_FS)
5504 if (!amdgpu_sriov_vf(adev))
5505 cancel_work(&adev->reset_work);
5506#endif
5507
5508 if (adev->kfd.dev)
5509 cancel_work(&adev->kfd.reset_work);
5510
5511 if (amdgpu_sriov_vf(adev))
5512 cancel_work(&adev->virt.flr_work);
5513
5514 if (con && adev->ras_enabled)
5515 cancel_work(&con->recovery_work);
5516
5517}
5518
5519/**
5520 * amdgpu_device_gpu_recover - reset the asic and recover scheduler
5521 *
5522 * @adev: amdgpu_device pointer
5523 * @job: which job trigger hang
5524 * @reset_context: amdgpu reset context pointer
5525 *
5526 * Attempt to reset the GPU if it has hung (all asics).
5527 * Attempt to do soft-reset or full-reset and reinitialize Asic
5528 * Returns 0 for success or an error on failure.
5529 */
5530
5531int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
5532 struct amdgpu_job *job,
5533 struct amdgpu_reset_context *reset_context)
5534{
5535 struct list_head device_list, *device_list_handle = NULL;
5536 bool job_signaled = false;
5537 struct amdgpu_hive_info *hive = NULL;
5538 struct amdgpu_device *tmp_adev = NULL;
5539 int i, r = 0;
5540 bool need_emergency_restart = false;
5541 bool audio_suspended = false;
5542
5543 /*
5544 * Special case: RAS triggered and full reset isn't supported
5545 */
5546 need_emergency_restart = amdgpu_ras_need_emergency_restart(adev);
5547
5548 /*
5549 * Flush RAM to disk so that after reboot
5550 * the user can read log and see why the system rebooted.
5551 */
5552 if (need_emergency_restart && amdgpu_ras_get_context(adev) &&
5553 amdgpu_ras_get_context(adev)->reboot) {
5554 DRM_WARN("Emergency reboot.");
5555
5556 ksys_sync_helper();
5557 emergency_restart();
5558 }
5559
5560 dev_info(adev->dev, "GPU %s begin!\n",
5561 need_emergency_restart ? "jobs stop":"reset");
5562
5563 if (!amdgpu_sriov_vf(adev))
5564 hive = amdgpu_get_xgmi_hive(adev);
5565 if (hive)
5566 mutex_lock(&hive->hive_lock);
5567
5568 reset_context->job = job;
5569 reset_context->hive = hive;
5570 /*
5571 * Build list of devices to reset.
5572 * In case we are in XGMI hive mode, resort the device list
5573 * to put adev in the 1st position.
5574 */
5575 INIT_LIST_HEAD(&device_list);
5576 if (!amdgpu_sriov_vf(adev) && (adev->gmc.xgmi.num_physical_nodes > 1)) {
5577 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
5578 list_add_tail(&tmp_adev->reset_list, &device_list);
5579 if (adev->shutdown)
5580 tmp_adev->shutdown = true;
5581 }
5582 if (!list_is_first(&adev->reset_list, &device_list))
5583 list_rotate_to_front(&adev->reset_list, &device_list);
5584 device_list_handle = &device_list;
5585 } else {
5586 list_add_tail(&adev->reset_list, &device_list);
5587 device_list_handle = &device_list;
5588 }
5589
5590 /* We need to lock reset domain only once both for XGMI and single device */
5591 tmp_adev = list_first_entry(device_list_handle, struct amdgpu_device,
5592 reset_list);
5593 amdgpu_device_lock_reset_domain(tmp_adev->reset_domain);
5594
5595 /* block all schedulers and reset given job's ring */
5596 list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5597
5598 amdgpu_device_set_mp1_state(tmp_adev);
5599
5600 /*
5601 * Try to put the audio codec into suspend state
5602 * before gpu reset started.
5603 *
5604 * Due to the power domain of the graphics device
5605 * is shared with AZ power domain. Without this,
5606 * we may change the audio hardware from behind
5607 * the audio driver's back. That will trigger
5608 * some audio codec errors.
5609 */
5610 if (!amdgpu_device_suspend_display_audio(tmp_adev))
5611 audio_suspended = true;
5612
5613 amdgpu_ras_set_error_query_ready(tmp_adev, false);
5614
5615 cancel_delayed_work_sync(&tmp_adev->delayed_init_work);
5616
5617 if (!amdgpu_sriov_vf(tmp_adev))
5618 amdgpu_amdkfd_pre_reset(tmp_adev);
5619
5620 /*
5621 * Mark these ASICs to be reseted as untracked first
5622 * And add them back after reset completed
5623 */
5624 amdgpu_unregister_gpu_instance(tmp_adev);
5625
5626 drm_fb_helper_set_suspend_unlocked(adev_to_drm(tmp_adev)->fb_helper, true);
5627
5628 /* disable ras on ALL IPs */
5629 if (!need_emergency_restart &&
5630 amdgpu_device_ip_need_full_reset(tmp_adev))
5631 amdgpu_ras_suspend(tmp_adev);
5632
5633 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
5634 struct amdgpu_ring *ring = tmp_adev->rings[i];
5635
5636 if (!amdgpu_ring_sched_ready(ring))
5637 continue;
5638
5639 drm_sched_stop(&ring->sched, job ? &job->base : NULL);
5640
5641 if (need_emergency_restart)
5642 amdgpu_job_stop_all_jobs_on_sched(&ring->sched);
5643 }
5644 atomic_inc(&tmp_adev->gpu_reset_counter);
5645 }
5646
5647 if (need_emergency_restart)
5648 goto skip_sched_resume;
5649
5650 /*
5651 * Must check guilty signal here since after this point all old
5652 * HW fences are force signaled.
5653 *
5654 * job->base holds a reference to parent fence
5655 */
5656 if (job && dma_fence_is_signaled(&job->hw_fence)) {
5657 job_signaled = true;
5658 dev_info(adev->dev, "Guilty job already signaled, skipping HW reset");
5659 goto skip_hw_reset;
5660 }
5661
5662retry: /* Rest of adevs pre asic reset from XGMI hive. */
5663 list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5664 r = amdgpu_device_pre_asic_reset(tmp_adev, reset_context);
5665 /*TODO Should we stop ?*/
5666 if (r) {
5667 dev_err(tmp_adev->dev, "GPU pre asic reset failed with err, %d for drm dev, %s ",
5668 r, adev_to_drm(tmp_adev)->unique);
5669 tmp_adev->asic_reset_res = r;
5670 }
5671
5672 /*
5673 * Drop all pending non scheduler resets. Scheduler resets
5674 * were already dropped during drm_sched_stop
5675 */
5676 amdgpu_device_stop_pending_resets(tmp_adev);
5677 }
5678
5679 /* Actual ASIC resets if needed.*/
5680 /* Host driver will handle XGMI hive reset for SRIOV */
5681 if (amdgpu_sriov_vf(adev)) {
5682 r = amdgpu_device_reset_sriov(adev, job ? false : true);
5683 if (r)
5684 adev->asic_reset_res = r;
5685
5686 /* Aldebaran and gfx_11_0_3 support ras in SRIOV, so need resume ras during reset */
5687 if (amdgpu_ip_version(adev, GC_HWIP, 0) ==
5688 IP_VERSION(9, 4, 2) ||
5689 amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(11, 0, 3))
5690 amdgpu_ras_resume(adev);
5691 } else {
5692 r = amdgpu_do_asic_reset(device_list_handle, reset_context);
5693 if (r && r == -EAGAIN)
5694 goto retry;
5695 }
5696
5697skip_hw_reset:
5698
5699 /* Post ASIC reset for all devs .*/
5700 list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5701
5702 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
5703 struct amdgpu_ring *ring = tmp_adev->rings[i];
5704
5705 if (!amdgpu_ring_sched_ready(ring))
5706 continue;
5707
5708 drm_sched_start(&ring->sched, true);
5709 }
5710
5711 if (!drm_drv_uses_atomic_modeset(adev_to_drm(tmp_adev)) && !job_signaled)
5712 drm_helper_resume_force_mode(adev_to_drm(tmp_adev));
5713
5714 if (tmp_adev->asic_reset_res)
5715 r = tmp_adev->asic_reset_res;
5716
5717 tmp_adev->asic_reset_res = 0;
5718
5719 if (r) {
5720 /* bad news, how to tell it to userspace ? */
5721 dev_info(tmp_adev->dev, "GPU reset(%d) failed\n", atomic_read(&tmp_adev->gpu_reset_counter));
5722 amdgpu_vf_error_put(tmp_adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r);
5723 } else {
5724 dev_info(tmp_adev->dev, "GPU reset(%d) succeeded!\n", atomic_read(&tmp_adev->gpu_reset_counter));
5725 if (amdgpu_acpi_smart_shift_update(adev_to_drm(tmp_adev), AMDGPU_SS_DEV_D0))
5726 DRM_WARN("smart shift update failed\n");
5727 }
5728 }
5729
5730skip_sched_resume:
5731 list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5732 /* unlock kfd: SRIOV would do it separately */
5733 if (!need_emergency_restart && !amdgpu_sriov_vf(tmp_adev))
5734 amdgpu_amdkfd_post_reset(tmp_adev);
5735
5736 /* kfd_post_reset will do nothing if kfd device is not initialized,
5737 * need to bring up kfd here if it's not be initialized before
5738 */
5739 if (!adev->kfd.init_complete)
5740 amdgpu_amdkfd_device_init(adev);
5741
5742 if (audio_suspended)
5743 amdgpu_device_resume_display_audio(tmp_adev);
5744
5745 amdgpu_device_unset_mp1_state(tmp_adev);
5746
5747 amdgpu_ras_set_error_query_ready(tmp_adev, true);
5748 }
5749
5750 tmp_adev = list_first_entry(device_list_handle, struct amdgpu_device,
5751 reset_list);
5752 amdgpu_device_unlock_reset_domain(tmp_adev->reset_domain);
5753
5754 if (hive) {
5755 mutex_unlock(&hive->hive_lock);
5756 amdgpu_put_xgmi_hive(hive);
5757 }
5758
5759 if (r)
5760 dev_info(adev->dev, "GPU reset end with ret = %d\n", r);
5761
5762 atomic_set(&adev->reset_domain->reset_res, r);
5763 return r;
5764}
5765
5766/**
5767 * amdgpu_device_partner_bandwidth - find the bandwidth of appropriate partner
5768 *
5769 * @adev: amdgpu_device pointer
5770 * @speed: pointer to the speed of the link
5771 * @width: pointer to the width of the link
5772 *
5773 * Evaluate the hierarchy to find the speed and bandwidth capabilities of the
5774 * first physical partner to an AMD dGPU.
5775 * This will exclude any virtual switches and links.
5776 */
5777static void amdgpu_device_partner_bandwidth(struct amdgpu_device *adev,
5778 enum pci_bus_speed *speed,
5779 enum pcie_link_width *width)
5780{
5781 struct pci_dev *parent = adev->pdev;
5782
5783 if (!speed || !width)
5784 return;
5785
5786 *speed = PCI_SPEED_UNKNOWN;
5787 *width = PCIE_LNK_WIDTH_UNKNOWN;
5788
5789 while ((parent = pci_upstream_bridge(parent))) {
5790 /* skip upstream/downstream switches internal to dGPU*/
5791 if (parent->vendor == PCI_VENDOR_ID_ATI)
5792 continue;
5793 *speed = pcie_get_speed_cap(parent);
5794 *width = pcie_get_width_cap(parent);
5795 break;
5796 }
5797}
5798
5799/**
5800 * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot
5801 *
5802 * @adev: amdgpu_device pointer
5803 *
5804 * Fetchs and stores in the driver the PCIE capabilities (gen speed
5805 * and lanes) of the slot the device is in. Handles APUs and
5806 * virtualized environments where PCIE config space may not be available.
5807 */
5808static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev)
5809{
5810 struct pci_dev *pdev;
5811 enum pci_bus_speed speed_cap, platform_speed_cap;
5812 enum pcie_link_width platform_link_width;
5813
5814 if (amdgpu_pcie_gen_cap)
5815 adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
5816
5817 if (amdgpu_pcie_lane_cap)
5818 adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
5819
5820 /* covers APUs as well */
5821 if (pci_is_root_bus(adev->pdev->bus) && !amdgpu_passthrough(adev)) {
5822 if (adev->pm.pcie_gen_mask == 0)
5823 adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
5824 if (adev->pm.pcie_mlw_mask == 0)
5825 adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
5826 return;
5827 }
5828
5829 if (adev->pm.pcie_gen_mask && adev->pm.pcie_mlw_mask)
5830 return;
5831
5832 amdgpu_device_partner_bandwidth(adev, &platform_speed_cap,
5833 &platform_link_width);
5834
5835 if (adev->pm.pcie_gen_mask == 0) {
5836 /* asic caps */
5837 pdev = adev->pdev;
5838 speed_cap = pcie_get_speed_cap(pdev);
5839 if (speed_cap == PCI_SPEED_UNKNOWN) {
5840 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5841 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5842 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
5843 } else {
5844 if (speed_cap == PCIE_SPEED_32_0GT)
5845 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5846 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5847 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
5848 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4 |
5849 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN5);
5850 else if (speed_cap == PCIE_SPEED_16_0GT)
5851 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5852 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5853 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
5854 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4);
5855 else if (speed_cap == PCIE_SPEED_8_0GT)
5856 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5857 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5858 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
5859 else if (speed_cap == PCIE_SPEED_5_0GT)
5860 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5861 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2);
5862 else
5863 adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1;
5864 }
5865 /* platform caps */
5866 if (platform_speed_cap == PCI_SPEED_UNKNOWN) {
5867 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5868 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
5869 } else {
5870 if (platform_speed_cap == PCIE_SPEED_32_0GT)
5871 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5872 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5873 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
5874 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4 |
5875 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN5);
5876 else if (platform_speed_cap == PCIE_SPEED_16_0GT)
5877 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5878 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5879 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
5880 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4);
5881 else if (platform_speed_cap == PCIE_SPEED_8_0GT)
5882 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5883 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5884 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3);
5885 else if (platform_speed_cap == PCIE_SPEED_5_0GT)
5886 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5887 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
5888 else
5889 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
5890
5891 }
5892 }
5893 if (adev->pm.pcie_mlw_mask == 0) {
5894 if (platform_link_width == PCIE_LNK_WIDTH_UNKNOWN) {
5895 adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK;
5896 } else {
5897 switch (platform_link_width) {
5898 case PCIE_LNK_X32:
5899 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
5900 CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
5901 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
5902 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
5903 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5904 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5905 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5906 break;
5907 case PCIE_LNK_X16:
5908 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
5909 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
5910 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
5911 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5912 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5913 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5914 break;
5915 case PCIE_LNK_X12:
5916 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
5917 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
5918 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5919 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5920 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5921 break;
5922 case PCIE_LNK_X8:
5923 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
5924 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5925 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5926 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5927 break;
5928 case PCIE_LNK_X4:
5929 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5930 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5931 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5932 break;
5933 case PCIE_LNK_X2:
5934 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5935 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5936 break;
5937 case PCIE_LNK_X1:
5938 adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;
5939 break;
5940 default:
5941 break;
5942 }
5943 }
5944 }
5945}
5946
5947/**
5948 * amdgpu_device_is_peer_accessible - Check peer access through PCIe BAR
5949 *
5950 * @adev: amdgpu_device pointer
5951 * @peer_adev: amdgpu_device pointer for peer device trying to access @adev
5952 *
5953 * Return true if @peer_adev can access (DMA) @adev through the PCIe
5954 * BAR, i.e. @adev is "large BAR" and the BAR matches the DMA mask of
5955 * @peer_adev.
5956 */
5957bool amdgpu_device_is_peer_accessible(struct amdgpu_device *adev,
5958 struct amdgpu_device *peer_adev)
5959{
5960#ifdef CONFIG_HSA_AMD_P2P
5961 uint64_t address_mask = peer_adev->dev->dma_mask ?
5962 ~*peer_adev->dev->dma_mask : ~((1ULL << 32) - 1);
5963 resource_size_t aper_limit =
5964 adev->gmc.aper_base + adev->gmc.aper_size - 1;
5965 bool p2p_access =
5966 !adev->gmc.xgmi.connected_to_cpu &&
5967 !(pci_p2pdma_distance(adev->pdev, peer_adev->dev, false) < 0);
5968
5969 return pcie_p2p && p2p_access && (adev->gmc.visible_vram_size &&
5970 adev->gmc.real_vram_size == adev->gmc.visible_vram_size &&
5971 !(adev->gmc.aper_base & address_mask ||
5972 aper_limit & address_mask));
5973#else
5974 return false;
5975#endif
5976}
5977
5978int amdgpu_device_baco_enter(struct drm_device *dev)
5979{
5980 struct amdgpu_device *adev = drm_to_adev(dev);
5981 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
5982
5983 if (!amdgpu_device_supports_baco(dev))
5984 return -ENOTSUPP;
5985
5986 if (ras && adev->ras_enabled &&
5987 adev->nbio.funcs->enable_doorbell_interrupt)
5988 adev->nbio.funcs->enable_doorbell_interrupt(adev, false);
5989
5990 return amdgpu_dpm_baco_enter(adev);
5991}
5992
5993int amdgpu_device_baco_exit(struct drm_device *dev)
5994{
5995 struct amdgpu_device *adev = drm_to_adev(dev);
5996 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
5997 int ret = 0;
5998
5999 if (!amdgpu_device_supports_baco(dev))
6000 return -ENOTSUPP;
6001
6002 ret = amdgpu_dpm_baco_exit(adev);
6003 if (ret)
6004 return ret;
6005
6006 if (ras && adev->ras_enabled &&
6007 adev->nbio.funcs->enable_doorbell_interrupt)
6008 adev->nbio.funcs->enable_doorbell_interrupt(adev, true);
6009
6010 if (amdgpu_passthrough(adev) &&
6011 adev->nbio.funcs->clear_doorbell_interrupt)
6012 adev->nbio.funcs->clear_doorbell_interrupt(adev);
6013
6014 return 0;
6015}
6016
6017/**
6018 * amdgpu_pci_error_detected - Called when a PCI error is detected.
6019 * @pdev: PCI device struct
6020 * @state: PCI channel state
6021 *
6022 * Description: Called when a PCI error is detected.
6023 *
6024 * Return: PCI_ERS_RESULT_NEED_RESET or PCI_ERS_RESULT_DISCONNECT.
6025 */
6026pci_ers_result_t amdgpu_pci_error_detected(struct pci_dev *pdev, pci_channel_state_t state)
6027{
6028 struct drm_device *dev = pci_get_drvdata(pdev);
6029 struct amdgpu_device *adev = drm_to_adev(dev);
6030 int i;
6031
6032 DRM_INFO("PCI error: detected callback, state(%d)!!\n", state);
6033
6034 if (adev->gmc.xgmi.num_physical_nodes > 1) {
6035 DRM_WARN("No support for XGMI hive yet...");
6036 return PCI_ERS_RESULT_DISCONNECT;
6037 }
6038
6039 adev->pci_channel_state = state;
6040
6041 switch (state) {
6042 case pci_channel_io_normal:
6043 return PCI_ERS_RESULT_CAN_RECOVER;
6044 /* Fatal error, prepare for slot reset */
6045 case pci_channel_io_frozen:
6046 /*
6047 * Locking adev->reset_domain->sem will prevent any external access
6048 * to GPU during PCI error recovery
6049 */
6050 amdgpu_device_lock_reset_domain(adev->reset_domain);
6051 amdgpu_device_set_mp1_state(adev);
6052
6053 /*
6054 * Block any work scheduling as we do for regular GPU reset
6055 * for the duration of the recovery
6056 */
6057 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
6058 struct amdgpu_ring *ring = adev->rings[i];
6059
6060 if (!amdgpu_ring_sched_ready(ring))
6061 continue;
6062
6063 drm_sched_stop(&ring->sched, NULL);
6064 }
6065 atomic_inc(&adev->gpu_reset_counter);
6066 return PCI_ERS_RESULT_NEED_RESET;
6067 case pci_channel_io_perm_failure:
6068 /* Permanent error, prepare for device removal */
6069 return PCI_ERS_RESULT_DISCONNECT;
6070 }
6071
6072 return PCI_ERS_RESULT_NEED_RESET;
6073}
6074
6075/**
6076 * amdgpu_pci_mmio_enabled - Enable MMIO and dump debug registers
6077 * @pdev: pointer to PCI device
6078 */
6079pci_ers_result_t amdgpu_pci_mmio_enabled(struct pci_dev *pdev)
6080{
6081
6082 DRM_INFO("PCI error: mmio enabled callback!!\n");
6083
6084 /* TODO - dump whatever for debugging purposes */
6085
6086 /* This called only if amdgpu_pci_error_detected returns
6087 * PCI_ERS_RESULT_CAN_RECOVER. Read/write to the device still
6088 * works, no need to reset slot.
6089 */
6090
6091 return PCI_ERS_RESULT_RECOVERED;
6092}
6093
6094/**
6095 * amdgpu_pci_slot_reset - Called when PCI slot has been reset.
6096 * @pdev: PCI device struct
6097 *
6098 * Description: This routine is called by the pci error recovery
6099 * code after the PCI slot has been reset, just before we
6100 * should resume normal operations.
6101 */
6102pci_ers_result_t amdgpu_pci_slot_reset(struct pci_dev *pdev)
6103{
6104 struct drm_device *dev = pci_get_drvdata(pdev);
6105 struct amdgpu_device *adev = drm_to_adev(dev);
6106 int r, i;
6107 struct amdgpu_reset_context reset_context;
6108 u32 memsize;
6109 struct list_head device_list;
6110
6111 DRM_INFO("PCI error: slot reset callback!!\n");
6112
6113 memset(&reset_context, 0, sizeof(reset_context));
6114
6115 INIT_LIST_HEAD(&device_list);
6116 list_add_tail(&adev->reset_list, &device_list);
6117
6118 /* wait for asic to come out of reset */
6119 msleep(500);
6120
6121 /* Restore PCI confspace */
6122 amdgpu_device_load_pci_state(pdev);
6123
6124 /* confirm ASIC came out of reset */
6125 for (i = 0; i < adev->usec_timeout; i++) {
6126 memsize = amdgpu_asic_get_config_memsize(adev);
6127
6128 if (memsize != 0xffffffff)
6129 break;
6130 udelay(1);
6131 }
6132 if (memsize == 0xffffffff) {
6133 r = -ETIME;
6134 goto out;
6135 }
6136
6137 reset_context.method = AMD_RESET_METHOD_NONE;
6138 reset_context.reset_req_dev = adev;
6139 set_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
6140 set_bit(AMDGPU_SKIP_HW_RESET, &reset_context.flags);
6141
6142 adev->no_hw_access = true;
6143 r = amdgpu_device_pre_asic_reset(adev, &reset_context);
6144 adev->no_hw_access = false;
6145 if (r)
6146 goto out;
6147
6148 r = amdgpu_do_asic_reset(&device_list, &reset_context);
6149
6150out:
6151 if (!r) {
6152 if (amdgpu_device_cache_pci_state(adev->pdev))
6153 pci_restore_state(adev->pdev);
6154
6155 DRM_INFO("PCIe error recovery succeeded\n");
6156 } else {
6157 DRM_ERROR("PCIe error recovery failed, err:%d", r);
6158 amdgpu_device_unset_mp1_state(adev);
6159 amdgpu_device_unlock_reset_domain(adev->reset_domain);
6160 }
6161
6162 return r ? PCI_ERS_RESULT_DISCONNECT : PCI_ERS_RESULT_RECOVERED;
6163}
6164
6165/**
6166 * amdgpu_pci_resume() - resume normal ops after PCI reset
6167 * @pdev: pointer to PCI device
6168 *
6169 * Called when the error recovery driver tells us that its
6170 * OK to resume normal operation.
6171 */
6172void amdgpu_pci_resume(struct pci_dev *pdev)
6173{
6174 struct drm_device *dev = pci_get_drvdata(pdev);
6175 struct amdgpu_device *adev = drm_to_adev(dev);
6176 int i;
6177
6178
6179 DRM_INFO("PCI error: resume callback!!\n");
6180
6181 /* Only continue execution for the case of pci_channel_io_frozen */
6182 if (adev->pci_channel_state != pci_channel_io_frozen)
6183 return;
6184
6185 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
6186 struct amdgpu_ring *ring = adev->rings[i];
6187
6188 if (!amdgpu_ring_sched_ready(ring))
6189 continue;
6190
6191 drm_sched_start(&ring->sched, true);
6192 }
6193
6194 amdgpu_device_unset_mp1_state(adev);
6195 amdgpu_device_unlock_reset_domain(adev->reset_domain);
6196}
6197
6198bool amdgpu_device_cache_pci_state(struct pci_dev *pdev)
6199{
6200 struct drm_device *dev = pci_get_drvdata(pdev);
6201 struct amdgpu_device *adev = drm_to_adev(dev);
6202 int r;
6203
6204 r = pci_save_state(pdev);
6205 if (!r) {
6206 kfree(adev->pci_state);
6207
6208 adev->pci_state = pci_store_saved_state(pdev);
6209
6210 if (!adev->pci_state) {
6211 DRM_ERROR("Failed to store PCI saved state");
6212 return false;
6213 }
6214 } else {
6215 DRM_WARN("Failed to save PCI state, err:%d\n", r);
6216 return false;
6217 }
6218
6219 return true;
6220}
6221
6222bool amdgpu_device_load_pci_state(struct pci_dev *pdev)
6223{
6224 struct drm_device *dev = pci_get_drvdata(pdev);
6225 struct amdgpu_device *adev = drm_to_adev(dev);
6226 int r;
6227
6228 if (!adev->pci_state)
6229 return false;
6230
6231 r = pci_load_saved_state(pdev, adev->pci_state);
6232
6233 if (!r) {
6234 pci_restore_state(pdev);
6235 } else {
6236 DRM_WARN("Failed to load PCI state, err:%d\n", r);
6237 return false;
6238 }
6239
6240 return true;
6241}
6242
6243void amdgpu_device_flush_hdp(struct amdgpu_device *adev,
6244 struct amdgpu_ring *ring)
6245{
6246#ifdef CONFIG_X86_64
6247 if ((adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev))
6248 return;
6249#endif
6250 if (adev->gmc.xgmi.connected_to_cpu)
6251 return;
6252
6253 if (ring && ring->funcs->emit_hdp_flush)
6254 amdgpu_ring_emit_hdp_flush(ring);
6255 else
6256 amdgpu_asic_flush_hdp(adev, ring);
6257}
6258
6259void amdgpu_device_invalidate_hdp(struct amdgpu_device *adev,
6260 struct amdgpu_ring *ring)
6261{
6262#ifdef CONFIG_X86_64
6263 if ((adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev))
6264 return;
6265#endif
6266 if (adev->gmc.xgmi.connected_to_cpu)
6267 return;
6268
6269 amdgpu_asic_invalidate_hdp(adev, ring);
6270}
6271
6272int amdgpu_in_reset(struct amdgpu_device *adev)
6273{
6274 return atomic_read(&adev->reset_domain->in_gpu_reset);
6275}
6276
6277/**
6278 * amdgpu_device_halt() - bring hardware to some kind of halt state
6279 *
6280 * @adev: amdgpu_device pointer
6281 *
6282 * Bring hardware to some kind of halt state so that no one can touch it
6283 * any more. It will help to maintain error context when error occurred.
6284 * Compare to a simple hang, the system will keep stable at least for SSH
6285 * access. Then it should be trivial to inspect the hardware state and
6286 * see what's going on. Implemented as following:
6287 *
6288 * 1. drm_dev_unplug() makes device inaccessible to user space(IOCTLs, etc),
6289 * clears all CPU mappings to device, disallows remappings through page faults
6290 * 2. amdgpu_irq_disable_all() disables all interrupts
6291 * 3. amdgpu_fence_driver_hw_fini() signals all HW fences
6292 * 4. set adev->no_hw_access to avoid potential crashes after setp 5
6293 * 5. amdgpu_device_unmap_mmio() clears all MMIO mappings
6294 * 6. pci_disable_device() and pci_wait_for_pending_transaction()
6295 * flush any in flight DMA operations
6296 */
6297void amdgpu_device_halt(struct amdgpu_device *adev)
6298{
6299 struct pci_dev *pdev = adev->pdev;
6300 struct drm_device *ddev = adev_to_drm(adev);
6301
6302 amdgpu_xcp_dev_unplug(adev);
6303 drm_dev_unplug(ddev);
6304
6305 amdgpu_irq_disable_all(adev);
6306
6307 amdgpu_fence_driver_hw_fini(adev);
6308
6309 adev->no_hw_access = true;
6310
6311 amdgpu_device_unmap_mmio(adev);
6312
6313 pci_disable_device(pdev);
6314 pci_wait_for_pending_transaction(pdev);
6315}
6316
6317u32 amdgpu_device_pcie_port_rreg(struct amdgpu_device *adev,
6318 u32 reg)
6319{
6320 unsigned long flags, address, data;
6321 u32 r;
6322
6323 address = adev->nbio.funcs->get_pcie_port_index_offset(adev);
6324 data = adev->nbio.funcs->get_pcie_port_data_offset(adev);
6325
6326 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
6327 WREG32(address, reg * 4);
6328 (void)RREG32(address);
6329 r = RREG32(data);
6330 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
6331 return r;
6332}
6333
6334void amdgpu_device_pcie_port_wreg(struct amdgpu_device *adev,
6335 u32 reg, u32 v)
6336{
6337 unsigned long flags, address, data;
6338
6339 address = adev->nbio.funcs->get_pcie_port_index_offset(adev);
6340 data = adev->nbio.funcs->get_pcie_port_data_offset(adev);
6341
6342 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
6343 WREG32(address, reg * 4);
6344 (void)RREG32(address);
6345 WREG32(data, v);
6346 (void)RREG32(data);
6347 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
6348}
6349
6350/**
6351 * amdgpu_device_switch_gang - switch to a new gang
6352 * @adev: amdgpu_device pointer
6353 * @gang: the gang to switch to
6354 *
6355 * Try to switch to a new gang.
6356 * Returns: NULL if we switched to the new gang or a reference to the current
6357 * gang leader.
6358 */
6359struct dma_fence *amdgpu_device_switch_gang(struct amdgpu_device *adev,
6360 struct dma_fence *gang)
6361{
6362 struct dma_fence *old = NULL;
6363
6364 do {
6365 dma_fence_put(old);
6366 rcu_read_lock();
6367 old = dma_fence_get_rcu_safe(&adev->gang_submit);
6368 rcu_read_unlock();
6369
6370 if (old == gang)
6371 break;
6372
6373 if (!dma_fence_is_signaled(old))
6374 return old;
6375
6376 } while (cmpxchg((struct dma_fence __force **)&adev->gang_submit,
6377 old, gang) != old);
6378
6379 dma_fence_put(old);
6380 return NULL;
6381}
6382
6383bool amdgpu_device_has_display_hardware(struct amdgpu_device *adev)
6384{
6385 switch (adev->asic_type) {
6386#ifdef CONFIG_DRM_AMDGPU_SI
6387 case CHIP_HAINAN:
6388#endif
6389 case CHIP_TOPAZ:
6390 /* chips with no display hardware */
6391 return false;
6392#ifdef CONFIG_DRM_AMDGPU_SI
6393 case CHIP_TAHITI:
6394 case CHIP_PITCAIRN:
6395 case CHIP_VERDE:
6396 case CHIP_OLAND:
6397#endif
6398#ifdef CONFIG_DRM_AMDGPU_CIK
6399 case CHIP_BONAIRE:
6400 case CHIP_HAWAII:
6401 case CHIP_KAVERI:
6402 case CHIP_KABINI:
6403 case CHIP_MULLINS:
6404#endif
6405 case CHIP_TONGA:
6406 case CHIP_FIJI:
6407 case CHIP_POLARIS10:
6408 case CHIP_POLARIS11:
6409 case CHIP_POLARIS12:
6410 case CHIP_VEGAM:
6411 case CHIP_CARRIZO:
6412 case CHIP_STONEY:
6413 /* chips with display hardware */
6414 return true;
6415 default:
6416 /* IP discovery */
6417 if (!amdgpu_ip_version(adev, DCE_HWIP, 0) ||
6418 (adev->harvest_ip_mask & AMD_HARVEST_IP_DMU_MASK))
6419 return false;
6420 return true;
6421 }
6422}
6423
6424uint32_t amdgpu_device_wait_on_rreg(struct amdgpu_device *adev,
6425 uint32_t inst, uint32_t reg_addr, char reg_name[],
6426 uint32_t expected_value, uint32_t mask)
6427{
6428 uint32_t ret = 0;
6429 uint32_t old_ = 0;
6430 uint32_t tmp_ = RREG32(reg_addr);
6431 uint32_t loop = adev->usec_timeout;
6432
6433 while ((tmp_ & (mask)) != (expected_value)) {
6434 if (old_ != tmp_) {
6435 loop = adev->usec_timeout;
6436 old_ = tmp_;
6437 } else
6438 udelay(1);
6439 tmp_ = RREG32(reg_addr);
6440 loop--;
6441 if (!loop) {
6442 DRM_WARN("Register(%d) [%s] failed to reach value 0x%08x != 0x%08xn",
6443 inst, reg_name, (uint32_t)expected_value,
6444 (uint32_t)(tmp_ & (mask)));
6445 ret = -ETIMEDOUT;
6446 break;
6447 }
6448 }
6449 return ret;
6450}
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 */
28#include <linux/power_supply.h>
29#include <linux/kthread.h>
30#include <linux/module.h>
31#include <linux/console.h>
32#include <linux/slab.h>
33
34#include <drm/drm_atomic_helper.h>
35#include <drm/drm_probe_helper.h>
36#include <drm/amdgpu_drm.h>
37#include <linux/vgaarb.h>
38#include <linux/vga_switcheroo.h>
39#include <linux/efi.h>
40#include "amdgpu.h"
41#include "amdgpu_trace.h"
42#include "amdgpu_i2c.h"
43#include "atom.h"
44#include "amdgpu_atombios.h"
45#include "amdgpu_atomfirmware.h"
46#include "amd_pcie.h"
47#ifdef CONFIG_DRM_AMDGPU_SI
48#include "si.h"
49#endif
50#ifdef CONFIG_DRM_AMDGPU_CIK
51#include "cik.h"
52#endif
53#include "vi.h"
54#include "soc15.h"
55#include "nv.h"
56#include "bif/bif_4_1_d.h"
57#include <linux/pci.h>
58#include <linux/firmware.h>
59#include "amdgpu_vf_error.h"
60
61#include "amdgpu_amdkfd.h"
62#include "amdgpu_pm.h"
63
64#include "amdgpu_xgmi.h"
65#include "amdgpu_ras.h"
66#include "amdgpu_pmu.h"
67
68MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin");
69MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin");
70MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin");
71MODULE_FIRMWARE("amdgpu/picasso_gpu_info.bin");
72MODULE_FIRMWARE("amdgpu/raven2_gpu_info.bin");
73MODULE_FIRMWARE("amdgpu/arcturus_gpu_info.bin");
74MODULE_FIRMWARE("amdgpu/renoir_gpu_info.bin");
75MODULE_FIRMWARE("amdgpu/navi10_gpu_info.bin");
76MODULE_FIRMWARE("amdgpu/navi14_gpu_info.bin");
77MODULE_FIRMWARE("amdgpu/navi12_gpu_info.bin");
78
79#define AMDGPU_RESUME_MS 2000
80
81static const char *amdgpu_asic_name[] = {
82 "TAHITI",
83 "PITCAIRN",
84 "VERDE",
85 "OLAND",
86 "HAINAN",
87 "BONAIRE",
88 "KAVERI",
89 "KABINI",
90 "HAWAII",
91 "MULLINS",
92 "TOPAZ",
93 "TONGA",
94 "FIJI",
95 "CARRIZO",
96 "STONEY",
97 "POLARIS10",
98 "POLARIS11",
99 "POLARIS12",
100 "VEGAM",
101 "VEGA10",
102 "VEGA12",
103 "VEGA20",
104 "RAVEN",
105 "ARCTURUS",
106 "RENOIR",
107 "NAVI10",
108 "NAVI14",
109 "NAVI12",
110 "LAST",
111};
112
113/**
114 * DOC: pcie_replay_count
115 *
116 * The amdgpu driver provides a sysfs API for reporting the total number
117 * of PCIe replays (NAKs)
118 * The file pcie_replay_count is used for this and returns the total
119 * number of replays as a sum of the NAKs generated and NAKs received
120 */
121
122static ssize_t amdgpu_device_get_pcie_replay_count(struct device *dev,
123 struct device_attribute *attr, char *buf)
124{
125 struct drm_device *ddev = dev_get_drvdata(dev);
126 struct amdgpu_device *adev = ddev->dev_private;
127 uint64_t cnt = amdgpu_asic_get_pcie_replay_count(adev);
128
129 return snprintf(buf, PAGE_SIZE, "%llu\n", cnt);
130}
131
132static DEVICE_ATTR(pcie_replay_count, S_IRUGO,
133 amdgpu_device_get_pcie_replay_count, NULL);
134
135static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev);
136
137/**
138 * amdgpu_device_is_px - Is the device is a dGPU with HG/PX power control
139 *
140 * @dev: drm_device pointer
141 *
142 * Returns true if the device is a dGPU with HG/PX power control,
143 * otherwise return false.
144 */
145bool amdgpu_device_is_px(struct drm_device *dev)
146{
147 struct amdgpu_device *adev = dev->dev_private;
148
149 if (adev->flags & AMD_IS_PX)
150 return true;
151 return false;
152}
153
154/*
155 * MMIO register access helper functions.
156 */
157/**
158 * amdgpu_mm_rreg - read a memory mapped IO register
159 *
160 * @adev: amdgpu_device pointer
161 * @reg: dword aligned register offset
162 * @acc_flags: access flags which require special behavior
163 *
164 * Returns the 32 bit value from the offset specified.
165 */
166uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg,
167 uint32_t acc_flags)
168{
169 uint32_t ret;
170
171 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
172 return amdgpu_virt_kiq_rreg(adev, reg);
173
174 if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
175 ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
176 else {
177 unsigned long flags;
178
179 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
180 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
181 ret = readl(((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
182 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
183 }
184 trace_amdgpu_mm_rreg(adev->pdev->device, reg, ret);
185 return ret;
186}
187
188/*
189 * MMIO register read with bytes helper functions
190 * @offset:bytes offset from MMIO start
191 *
192*/
193
194/**
195 * amdgpu_mm_rreg8 - read a memory mapped IO register
196 *
197 * @adev: amdgpu_device pointer
198 * @offset: byte aligned register offset
199 *
200 * Returns the 8 bit value from the offset specified.
201 */
202uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset) {
203 if (offset < adev->rmmio_size)
204 return (readb(adev->rmmio + offset));
205 BUG();
206}
207
208/*
209 * MMIO register write with bytes helper functions
210 * @offset:bytes offset from MMIO start
211 * @value: the value want to be written to the register
212 *
213*/
214/**
215 * amdgpu_mm_wreg8 - read a memory mapped IO register
216 *
217 * @adev: amdgpu_device pointer
218 * @offset: byte aligned register offset
219 * @value: 8 bit value to write
220 *
221 * Writes the value specified to the offset specified.
222 */
223void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value) {
224 if (offset < adev->rmmio_size)
225 writeb(value, adev->rmmio + offset);
226 else
227 BUG();
228}
229
230/**
231 * amdgpu_mm_wreg - write to a memory mapped IO register
232 *
233 * @adev: amdgpu_device pointer
234 * @reg: dword aligned register offset
235 * @v: 32 bit value to write to the register
236 * @acc_flags: access flags which require special behavior
237 *
238 * Writes the value specified to the offset specified.
239 */
240void amdgpu_mm_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v,
241 uint32_t acc_flags)
242{
243 trace_amdgpu_mm_wreg(adev->pdev->device, reg, v);
244
245 if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
246 adev->last_mm_index = v;
247 }
248
249 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
250 return amdgpu_virt_kiq_wreg(adev, reg, v);
251
252 if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
253 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
254 else {
255 unsigned long flags;
256
257 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
258 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
259 writel(v, ((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
260 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
261 }
262
263 if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
264 udelay(500);
265 }
266}
267
268/**
269 * amdgpu_io_rreg - read an IO register
270 *
271 * @adev: amdgpu_device pointer
272 * @reg: dword aligned register offset
273 *
274 * Returns the 32 bit value from the offset specified.
275 */
276u32 amdgpu_io_rreg(struct amdgpu_device *adev, u32 reg)
277{
278 if ((reg * 4) < adev->rio_mem_size)
279 return ioread32(adev->rio_mem + (reg * 4));
280 else {
281 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
282 return ioread32(adev->rio_mem + (mmMM_DATA * 4));
283 }
284}
285
286/**
287 * amdgpu_io_wreg - write to an IO register
288 *
289 * @adev: amdgpu_device pointer
290 * @reg: dword aligned register offset
291 * @v: 32 bit value to write to the register
292 *
293 * Writes the value specified to the offset specified.
294 */
295void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32 v)
296{
297 if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
298 adev->last_mm_index = v;
299 }
300
301 if ((reg * 4) < adev->rio_mem_size)
302 iowrite32(v, adev->rio_mem + (reg * 4));
303 else {
304 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
305 iowrite32(v, adev->rio_mem + (mmMM_DATA * 4));
306 }
307
308 if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
309 udelay(500);
310 }
311}
312
313/**
314 * amdgpu_mm_rdoorbell - read a doorbell dword
315 *
316 * @adev: amdgpu_device pointer
317 * @index: doorbell index
318 *
319 * Returns the value in the doorbell aperture at the
320 * requested doorbell index (CIK).
321 */
322u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
323{
324 if (index < adev->doorbell.num_doorbells) {
325 return readl(adev->doorbell.ptr + index);
326 } else {
327 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
328 return 0;
329 }
330}
331
332/**
333 * amdgpu_mm_wdoorbell - write a doorbell dword
334 *
335 * @adev: amdgpu_device pointer
336 * @index: doorbell index
337 * @v: value to write
338 *
339 * Writes @v to the doorbell aperture at the
340 * requested doorbell index (CIK).
341 */
342void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v)
343{
344 if (index < adev->doorbell.num_doorbells) {
345 writel(v, adev->doorbell.ptr + index);
346 } else {
347 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
348 }
349}
350
351/**
352 * amdgpu_mm_rdoorbell64 - read a doorbell Qword
353 *
354 * @adev: amdgpu_device pointer
355 * @index: doorbell index
356 *
357 * Returns the value in the doorbell aperture at the
358 * requested doorbell index (VEGA10+).
359 */
360u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index)
361{
362 if (index < adev->doorbell.num_doorbells) {
363 return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index));
364 } else {
365 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
366 return 0;
367 }
368}
369
370/**
371 * amdgpu_mm_wdoorbell64 - write a doorbell Qword
372 *
373 * @adev: amdgpu_device pointer
374 * @index: doorbell index
375 * @v: value to write
376 *
377 * Writes @v to the doorbell aperture at the
378 * requested doorbell index (VEGA10+).
379 */
380void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v)
381{
382 if (index < adev->doorbell.num_doorbells) {
383 atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v);
384 } else {
385 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
386 }
387}
388
389/**
390 * amdgpu_invalid_rreg - dummy reg read function
391 *
392 * @adev: amdgpu device pointer
393 * @reg: offset of register
394 *
395 * Dummy register read function. Used for register blocks
396 * that certain asics don't have (all asics).
397 * Returns the value in the register.
398 */
399static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
400{
401 DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
402 BUG();
403 return 0;
404}
405
406/**
407 * amdgpu_invalid_wreg - dummy reg write function
408 *
409 * @adev: amdgpu device pointer
410 * @reg: offset of register
411 * @v: value to write to the register
412 *
413 * Dummy register read function. Used for register blocks
414 * that certain asics don't have (all asics).
415 */
416static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
417{
418 DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
419 reg, v);
420 BUG();
421}
422
423/**
424 * amdgpu_invalid_rreg64 - dummy 64 bit reg read function
425 *
426 * @adev: amdgpu device pointer
427 * @reg: offset of register
428 *
429 * Dummy register read function. Used for register blocks
430 * that certain asics don't have (all asics).
431 * Returns the value in the register.
432 */
433static uint64_t amdgpu_invalid_rreg64(struct amdgpu_device *adev, uint32_t reg)
434{
435 DRM_ERROR("Invalid callback to read 64 bit register 0x%04X\n", reg);
436 BUG();
437 return 0;
438}
439
440/**
441 * amdgpu_invalid_wreg64 - dummy reg write function
442 *
443 * @adev: amdgpu device pointer
444 * @reg: offset of register
445 * @v: value to write to the register
446 *
447 * Dummy register read function. Used for register blocks
448 * that certain asics don't have (all asics).
449 */
450static void amdgpu_invalid_wreg64(struct amdgpu_device *adev, uint32_t reg, uint64_t v)
451{
452 DRM_ERROR("Invalid callback to write 64 bit register 0x%04X with 0x%08llX\n",
453 reg, v);
454 BUG();
455}
456
457/**
458 * amdgpu_block_invalid_rreg - dummy reg read function
459 *
460 * @adev: amdgpu device pointer
461 * @block: offset of instance
462 * @reg: offset of register
463 *
464 * Dummy register read function. Used for register blocks
465 * that certain asics don't have (all asics).
466 * Returns the value in the register.
467 */
468static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
469 uint32_t block, uint32_t reg)
470{
471 DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
472 reg, block);
473 BUG();
474 return 0;
475}
476
477/**
478 * amdgpu_block_invalid_wreg - dummy reg write function
479 *
480 * @adev: amdgpu device pointer
481 * @block: offset of instance
482 * @reg: offset of register
483 * @v: value to write to the register
484 *
485 * Dummy register read function. Used for register blocks
486 * that certain asics don't have (all asics).
487 */
488static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
489 uint32_t block,
490 uint32_t reg, uint32_t v)
491{
492 DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
493 reg, block, v);
494 BUG();
495}
496
497/**
498 * amdgpu_device_vram_scratch_init - allocate the VRAM scratch page
499 *
500 * @adev: amdgpu device pointer
501 *
502 * Allocates a scratch page of VRAM for use by various things in the
503 * driver.
504 */
505static int amdgpu_device_vram_scratch_init(struct amdgpu_device *adev)
506{
507 return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE,
508 PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
509 &adev->vram_scratch.robj,
510 &adev->vram_scratch.gpu_addr,
511 (void **)&adev->vram_scratch.ptr);
512}
513
514/**
515 * amdgpu_device_vram_scratch_fini - Free the VRAM scratch page
516 *
517 * @adev: amdgpu device pointer
518 *
519 * Frees the VRAM scratch page.
520 */
521static void amdgpu_device_vram_scratch_fini(struct amdgpu_device *adev)
522{
523 amdgpu_bo_free_kernel(&adev->vram_scratch.robj, NULL, NULL);
524}
525
526/**
527 * amdgpu_device_program_register_sequence - program an array of registers.
528 *
529 * @adev: amdgpu_device pointer
530 * @registers: pointer to the register array
531 * @array_size: size of the register array
532 *
533 * Programs an array or registers with and and or masks.
534 * This is a helper for setting golden registers.
535 */
536void amdgpu_device_program_register_sequence(struct amdgpu_device *adev,
537 const u32 *registers,
538 const u32 array_size)
539{
540 u32 tmp, reg, and_mask, or_mask;
541 int i;
542
543 if (array_size % 3)
544 return;
545
546 for (i = 0; i < array_size; i +=3) {
547 reg = registers[i + 0];
548 and_mask = registers[i + 1];
549 or_mask = registers[i + 2];
550
551 if (and_mask == 0xffffffff) {
552 tmp = or_mask;
553 } else {
554 tmp = RREG32(reg);
555 tmp &= ~and_mask;
556 if (adev->family >= AMDGPU_FAMILY_AI)
557 tmp |= (or_mask & and_mask);
558 else
559 tmp |= or_mask;
560 }
561 WREG32(reg, tmp);
562 }
563}
564
565/**
566 * amdgpu_device_pci_config_reset - reset the GPU
567 *
568 * @adev: amdgpu_device pointer
569 *
570 * Resets the GPU using the pci config reset sequence.
571 * Only applicable to asics prior to vega10.
572 */
573void amdgpu_device_pci_config_reset(struct amdgpu_device *adev)
574{
575 pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
576}
577
578/*
579 * GPU doorbell aperture helpers function.
580 */
581/**
582 * amdgpu_device_doorbell_init - Init doorbell driver information.
583 *
584 * @adev: amdgpu_device pointer
585 *
586 * Init doorbell driver information (CIK)
587 * Returns 0 on success, error on failure.
588 */
589static int amdgpu_device_doorbell_init(struct amdgpu_device *adev)
590{
591
592 /* No doorbell on SI hardware generation */
593 if (adev->asic_type < CHIP_BONAIRE) {
594 adev->doorbell.base = 0;
595 adev->doorbell.size = 0;
596 adev->doorbell.num_doorbells = 0;
597 adev->doorbell.ptr = NULL;
598 return 0;
599 }
600
601 if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET)
602 return -EINVAL;
603
604 amdgpu_asic_init_doorbell_index(adev);
605
606 /* doorbell bar mapping */
607 adev->doorbell.base = pci_resource_start(adev->pdev, 2);
608 adev->doorbell.size = pci_resource_len(adev->pdev, 2);
609
610 adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32),
611 adev->doorbell_index.max_assignment+1);
612 if (adev->doorbell.num_doorbells == 0)
613 return -EINVAL;
614
615 /* For Vega, reserve and map two pages on doorbell BAR since SDMA
616 * paging queue doorbell use the second page. The
617 * AMDGPU_DOORBELL64_MAX_ASSIGNMENT definition assumes all the
618 * doorbells are in the first page. So with paging queue enabled,
619 * the max num_doorbells should + 1 page (0x400 in dword)
620 */
621 if (adev->asic_type >= CHIP_VEGA10)
622 adev->doorbell.num_doorbells += 0x400;
623
624 adev->doorbell.ptr = ioremap(adev->doorbell.base,
625 adev->doorbell.num_doorbells *
626 sizeof(u32));
627 if (adev->doorbell.ptr == NULL)
628 return -ENOMEM;
629
630 return 0;
631}
632
633/**
634 * amdgpu_device_doorbell_fini - Tear down doorbell driver information.
635 *
636 * @adev: amdgpu_device pointer
637 *
638 * Tear down doorbell driver information (CIK)
639 */
640static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev)
641{
642 iounmap(adev->doorbell.ptr);
643 adev->doorbell.ptr = NULL;
644}
645
646
647
648/*
649 * amdgpu_device_wb_*()
650 * Writeback is the method by which the GPU updates special pages in memory
651 * with the status of certain GPU events (fences, ring pointers,etc.).
652 */
653
654/**
655 * amdgpu_device_wb_fini - Disable Writeback and free memory
656 *
657 * @adev: amdgpu_device pointer
658 *
659 * Disables Writeback and frees the Writeback memory (all asics).
660 * Used at driver shutdown.
661 */
662static void amdgpu_device_wb_fini(struct amdgpu_device *adev)
663{
664 if (adev->wb.wb_obj) {
665 amdgpu_bo_free_kernel(&adev->wb.wb_obj,
666 &adev->wb.gpu_addr,
667 (void **)&adev->wb.wb);
668 adev->wb.wb_obj = NULL;
669 }
670}
671
672/**
673 * amdgpu_device_wb_init- Init Writeback driver info and allocate memory
674 *
675 * @adev: amdgpu_device pointer
676 *
677 * Initializes writeback and allocates writeback memory (all asics).
678 * Used at driver startup.
679 * Returns 0 on success or an -error on failure.
680 */
681static int amdgpu_device_wb_init(struct amdgpu_device *adev)
682{
683 int r;
684
685 if (adev->wb.wb_obj == NULL) {
686 /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */
687 r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8,
688 PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
689 &adev->wb.wb_obj, &adev->wb.gpu_addr,
690 (void **)&adev->wb.wb);
691 if (r) {
692 dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
693 return r;
694 }
695
696 adev->wb.num_wb = AMDGPU_MAX_WB;
697 memset(&adev->wb.used, 0, sizeof(adev->wb.used));
698
699 /* clear wb memory */
700 memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8);
701 }
702
703 return 0;
704}
705
706/**
707 * amdgpu_device_wb_get - Allocate a wb entry
708 *
709 * @adev: amdgpu_device pointer
710 * @wb: wb index
711 *
712 * Allocate a wb slot for use by the driver (all asics).
713 * Returns 0 on success or -EINVAL on failure.
714 */
715int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb)
716{
717 unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
718
719 if (offset < adev->wb.num_wb) {
720 __set_bit(offset, adev->wb.used);
721 *wb = offset << 3; /* convert to dw offset */
722 return 0;
723 } else {
724 return -EINVAL;
725 }
726}
727
728/**
729 * amdgpu_device_wb_free - Free a wb entry
730 *
731 * @adev: amdgpu_device pointer
732 * @wb: wb index
733 *
734 * Free a wb slot allocated for use by the driver (all asics)
735 */
736void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb)
737{
738 wb >>= 3;
739 if (wb < adev->wb.num_wb)
740 __clear_bit(wb, adev->wb.used);
741}
742
743/**
744 * amdgpu_device_resize_fb_bar - try to resize FB BAR
745 *
746 * @adev: amdgpu_device pointer
747 *
748 * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not
749 * to fail, but if any of the BARs is not accessible after the size we abort
750 * driver loading by returning -ENODEV.
751 */
752int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev)
753{
754 u64 space_needed = roundup_pow_of_two(adev->gmc.real_vram_size);
755 u32 rbar_size = order_base_2(((space_needed >> 20) | 1)) - 1;
756 struct pci_bus *root;
757 struct resource *res;
758 unsigned i;
759 u16 cmd;
760 int r;
761
762 /* Bypass for VF */
763 if (amdgpu_sriov_vf(adev))
764 return 0;
765
766 /* Check if the root BUS has 64bit memory resources */
767 root = adev->pdev->bus;
768 while (root->parent)
769 root = root->parent;
770
771 pci_bus_for_each_resource(root, res, i) {
772 if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) &&
773 res->start > 0x100000000ull)
774 break;
775 }
776
777 /* Trying to resize is pointless without a root hub window above 4GB */
778 if (!res)
779 return 0;
780
781 /* Disable memory decoding while we change the BAR addresses and size */
782 pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd);
783 pci_write_config_word(adev->pdev, PCI_COMMAND,
784 cmd & ~PCI_COMMAND_MEMORY);
785
786 /* Free the VRAM and doorbell BAR, we most likely need to move both. */
787 amdgpu_device_doorbell_fini(adev);
788 if (adev->asic_type >= CHIP_BONAIRE)
789 pci_release_resource(adev->pdev, 2);
790
791 pci_release_resource(adev->pdev, 0);
792
793 r = pci_resize_resource(adev->pdev, 0, rbar_size);
794 if (r == -ENOSPC)
795 DRM_INFO("Not enough PCI address space for a large BAR.");
796 else if (r && r != -ENOTSUPP)
797 DRM_ERROR("Problem resizing BAR0 (%d).", r);
798
799 pci_assign_unassigned_bus_resources(adev->pdev->bus);
800
801 /* When the doorbell or fb BAR isn't available we have no chance of
802 * using the device.
803 */
804 r = amdgpu_device_doorbell_init(adev);
805 if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET))
806 return -ENODEV;
807
808 pci_write_config_word(adev->pdev, PCI_COMMAND, cmd);
809
810 return 0;
811}
812
813/*
814 * GPU helpers function.
815 */
816/**
817 * amdgpu_device_need_post - check if the hw need post or not
818 *
819 * @adev: amdgpu_device pointer
820 *
821 * Check if the asic has been initialized (all asics) at driver startup
822 * or post is needed if hw reset is performed.
823 * Returns true if need or false if not.
824 */
825bool amdgpu_device_need_post(struct amdgpu_device *adev)
826{
827 uint32_t reg;
828
829 if (amdgpu_sriov_vf(adev))
830 return false;
831
832 if (amdgpu_passthrough(adev)) {
833 /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
834 * some old smc fw still need driver do vPost otherwise gpu hang, while
835 * those smc fw version above 22.15 doesn't have this flaw, so we force
836 * vpost executed for smc version below 22.15
837 */
838 if (adev->asic_type == CHIP_FIJI) {
839 int err;
840 uint32_t fw_ver;
841 err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
842 /* force vPost if error occured */
843 if (err)
844 return true;
845
846 fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
847 if (fw_ver < 0x00160e00)
848 return true;
849 }
850 }
851
852 if (adev->has_hw_reset) {
853 adev->has_hw_reset = false;
854 return true;
855 }
856
857 /* bios scratch used on CIK+ */
858 if (adev->asic_type >= CHIP_BONAIRE)
859 return amdgpu_atombios_scratch_need_asic_init(adev);
860
861 /* check MEM_SIZE for older asics */
862 reg = amdgpu_asic_get_config_memsize(adev);
863
864 if ((reg != 0) && (reg != 0xffffffff))
865 return false;
866
867 return true;
868}
869
870/* if we get transitioned to only one device, take VGA back */
871/**
872 * amdgpu_device_vga_set_decode - enable/disable vga decode
873 *
874 * @cookie: amdgpu_device pointer
875 * @state: enable/disable vga decode
876 *
877 * Enable/disable vga decode (all asics).
878 * Returns VGA resource flags.
879 */
880static unsigned int amdgpu_device_vga_set_decode(void *cookie, bool state)
881{
882 struct amdgpu_device *adev = cookie;
883 amdgpu_asic_set_vga_state(adev, state);
884 if (state)
885 return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
886 VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
887 else
888 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
889}
890
891/**
892 * amdgpu_device_check_block_size - validate the vm block size
893 *
894 * @adev: amdgpu_device pointer
895 *
896 * Validates the vm block size specified via module parameter.
897 * The vm block size defines number of bits in page table versus page directory,
898 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
899 * page table and the remaining bits are in the page directory.
900 */
901static void amdgpu_device_check_block_size(struct amdgpu_device *adev)
902{
903 /* defines number of bits in page table versus page directory,
904 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
905 * page table and the remaining bits are in the page directory */
906 if (amdgpu_vm_block_size == -1)
907 return;
908
909 if (amdgpu_vm_block_size < 9) {
910 dev_warn(adev->dev, "VM page table size (%d) too small\n",
911 amdgpu_vm_block_size);
912 amdgpu_vm_block_size = -1;
913 }
914}
915
916/**
917 * amdgpu_device_check_vm_size - validate the vm size
918 *
919 * @adev: amdgpu_device pointer
920 *
921 * Validates the vm size in GB specified via module parameter.
922 * The VM size is the size of the GPU virtual memory space in GB.
923 */
924static void amdgpu_device_check_vm_size(struct amdgpu_device *adev)
925{
926 /* no need to check the default value */
927 if (amdgpu_vm_size == -1)
928 return;
929
930 if (amdgpu_vm_size < 1) {
931 dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
932 amdgpu_vm_size);
933 amdgpu_vm_size = -1;
934 }
935}
936
937static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev)
938{
939 struct sysinfo si;
940 bool is_os_64 = (sizeof(void *) == 8) ? true : false;
941 uint64_t total_memory;
942 uint64_t dram_size_seven_GB = 0x1B8000000;
943 uint64_t dram_size_three_GB = 0xB8000000;
944
945 if (amdgpu_smu_memory_pool_size == 0)
946 return;
947
948 if (!is_os_64) {
949 DRM_WARN("Not 64-bit OS, feature not supported\n");
950 goto def_value;
951 }
952 si_meminfo(&si);
953 total_memory = (uint64_t)si.totalram * si.mem_unit;
954
955 if ((amdgpu_smu_memory_pool_size == 1) ||
956 (amdgpu_smu_memory_pool_size == 2)) {
957 if (total_memory < dram_size_three_GB)
958 goto def_value1;
959 } else if ((amdgpu_smu_memory_pool_size == 4) ||
960 (amdgpu_smu_memory_pool_size == 8)) {
961 if (total_memory < dram_size_seven_GB)
962 goto def_value1;
963 } else {
964 DRM_WARN("Smu memory pool size not supported\n");
965 goto def_value;
966 }
967 adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28;
968
969 return;
970
971def_value1:
972 DRM_WARN("No enough system memory\n");
973def_value:
974 adev->pm.smu_prv_buffer_size = 0;
975}
976
977/**
978 * amdgpu_device_check_arguments - validate module params
979 *
980 * @adev: amdgpu_device pointer
981 *
982 * Validates certain module parameters and updates
983 * the associated values used by the driver (all asics).
984 */
985static int amdgpu_device_check_arguments(struct amdgpu_device *adev)
986{
987 int ret = 0;
988
989 if (amdgpu_sched_jobs < 4) {
990 dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
991 amdgpu_sched_jobs);
992 amdgpu_sched_jobs = 4;
993 } else if (!is_power_of_2(amdgpu_sched_jobs)){
994 dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
995 amdgpu_sched_jobs);
996 amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
997 }
998
999 if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) {
1000 /* gart size must be greater or equal to 32M */
1001 dev_warn(adev->dev, "gart size (%d) too small\n",
1002 amdgpu_gart_size);
1003 amdgpu_gart_size = -1;
1004 }
1005
1006 if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) {
1007 /* gtt size must be greater or equal to 32M */
1008 dev_warn(adev->dev, "gtt size (%d) too small\n",
1009 amdgpu_gtt_size);
1010 amdgpu_gtt_size = -1;
1011 }
1012
1013 /* valid range is between 4 and 9 inclusive */
1014 if (amdgpu_vm_fragment_size != -1 &&
1015 (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) {
1016 dev_warn(adev->dev, "valid range is between 4 and 9\n");
1017 amdgpu_vm_fragment_size = -1;
1018 }
1019
1020 amdgpu_device_check_smu_prv_buffer_size(adev);
1021
1022 amdgpu_device_check_vm_size(adev);
1023
1024 amdgpu_device_check_block_size(adev);
1025
1026 ret = amdgpu_device_get_job_timeout_settings(adev);
1027 if (ret) {
1028 dev_err(adev->dev, "invalid lockup_timeout parameter syntax\n");
1029 return ret;
1030 }
1031
1032 adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type);
1033
1034 return ret;
1035}
1036
1037/**
1038 * amdgpu_switcheroo_set_state - set switcheroo state
1039 *
1040 * @pdev: pci dev pointer
1041 * @state: vga_switcheroo state
1042 *
1043 * Callback for the switcheroo driver. Suspends or resumes the
1044 * the asics before or after it is powered up using ACPI methods.
1045 */
1046static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, enum vga_switcheroo_state state)
1047{
1048 struct drm_device *dev = pci_get_drvdata(pdev);
1049
1050 if (amdgpu_device_is_px(dev) && state == VGA_SWITCHEROO_OFF)
1051 return;
1052
1053 if (state == VGA_SWITCHEROO_ON) {
1054 pr_info("amdgpu: switched on\n");
1055 /* don't suspend or resume card normally */
1056 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1057
1058 amdgpu_device_resume(dev, true, true);
1059
1060 dev->switch_power_state = DRM_SWITCH_POWER_ON;
1061 drm_kms_helper_poll_enable(dev);
1062 } else {
1063 pr_info("amdgpu: switched off\n");
1064 drm_kms_helper_poll_disable(dev);
1065 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1066 amdgpu_device_suspend(dev, true, true);
1067 dev->switch_power_state = DRM_SWITCH_POWER_OFF;
1068 }
1069}
1070
1071/**
1072 * amdgpu_switcheroo_can_switch - see if switcheroo state can change
1073 *
1074 * @pdev: pci dev pointer
1075 *
1076 * Callback for the switcheroo driver. Check of the switcheroo
1077 * state can be changed.
1078 * Returns true if the state can be changed, false if not.
1079 */
1080static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
1081{
1082 struct drm_device *dev = pci_get_drvdata(pdev);
1083
1084 /*
1085 * FIXME: open_count is protected by drm_global_mutex but that would lead to
1086 * locking inversion with the driver load path. And the access here is
1087 * completely racy anyway. So don't bother with locking for now.
1088 */
1089 return dev->open_count == 0;
1090}
1091
1092static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
1093 .set_gpu_state = amdgpu_switcheroo_set_state,
1094 .reprobe = NULL,
1095 .can_switch = amdgpu_switcheroo_can_switch,
1096};
1097
1098/**
1099 * amdgpu_device_ip_set_clockgating_state - set the CG state
1100 *
1101 * @dev: amdgpu_device pointer
1102 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1103 * @state: clockgating state (gate or ungate)
1104 *
1105 * Sets the requested clockgating state for all instances of
1106 * the hardware IP specified.
1107 * Returns the error code from the last instance.
1108 */
1109int amdgpu_device_ip_set_clockgating_state(void *dev,
1110 enum amd_ip_block_type block_type,
1111 enum amd_clockgating_state state)
1112{
1113 struct amdgpu_device *adev = dev;
1114 int i, r = 0;
1115
1116 for (i = 0; i < adev->num_ip_blocks; i++) {
1117 if (!adev->ip_blocks[i].status.valid)
1118 continue;
1119 if (adev->ip_blocks[i].version->type != block_type)
1120 continue;
1121 if (!adev->ip_blocks[i].version->funcs->set_clockgating_state)
1122 continue;
1123 r = adev->ip_blocks[i].version->funcs->set_clockgating_state(
1124 (void *)adev, state);
1125 if (r)
1126 DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",
1127 adev->ip_blocks[i].version->funcs->name, r);
1128 }
1129 return r;
1130}
1131
1132/**
1133 * amdgpu_device_ip_set_powergating_state - set the PG state
1134 *
1135 * @dev: amdgpu_device pointer
1136 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1137 * @state: powergating state (gate or ungate)
1138 *
1139 * Sets the requested powergating state for all instances of
1140 * the hardware IP specified.
1141 * Returns the error code from the last instance.
1142 */
1143int amdgpu_device_ip_set_powergating_state(void *dev,
1144 enum amd_ip_block_type block_type,
1145 enum amd_powergating_state state)
1146{
1147 struct amdgpu_device *adev = dev;
1148 int i, r = 0;
1149
1150 for (i = 0; i < adev->num_ip_blocks; i++) {
1151 if (!adev->ip_blocks[i].status.valid)
1152 continue;
1153 if (adev->ip_blocks[i].version->type != block_type)
1154 continue;
1155 if (!adev->ip_blocks[i].version->funcs->set_powergating_state)
1156 continue;
1157 r = adev->ip_blocks[i].version->funcs->set_powergating_state(
1158 (void *)adev, state);
1159 if (r)
1160 DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",
1161 adev->ip_blocks[i].version->funcs->name, r);
1162 }
1163 return r;
1164}
1165
1166/**
1167 * amdgpu_device_ip_get_clockgating_state - get the CG state
1168 *
1169 * @adev: amdgpu_device pointer
1170 * @flags: clockgating feature flags
1171 *
1172 * Walks the list of IPs on the device and updates the clockgating
1173 * flags for each IP.
1174 * Updates @flags with the feature flags for each hardware IP where
1175 * clockgating is enabled.
1176 */
1177void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev,
1178 u32 *flags)
1179{
1180 int i;
1181
1182 for (i = 0; i < adev->num_ip_blocks; i++) {
1183 if (!adev->ip_blocks[i].status.valid)
1184 continue;
1185 if (adev->ip_blocks[i].version->funcs->get_clockgating_state)
1186 adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags);
1187 }
1188}
1189
1190/**
1191 * amdgpu_device_ip_wait_for_idle - wait for idle
1192 *
1193 * @adev: amdgpu_device pointer
1194 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1195 *
1196 * Waits for the request hardware IP to be idle.
1197 * Returns 0 for success or a negative error code on failure.
1198 */
1199int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev,
1200 enum amd_ip_block_type block_type)
1201{
1202 int i, r;
1203
1204 for (i = 0; i < adev->num_ip_blocks; i++) {
1205 if (!adev->ip_blocks[i].status.valid)
1206 continue;
1207 if (adev->ip_blocks[i].version->type == block_type) {
1208 r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
1209 if (r)
1210 return r;
1211 break;
1212 }
1213 }
1214 return 0;
1215
1216}
1217
1218/**
1219 * amdgpu_device_ip_is_idle - is the hardware IP idle
1220 *
1221 * @adev: amdgpu_device pointer
1222 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1223 *
1224 * Check if the hardware IP is idle or not.
1225 * Returns true if it the IP is idle, false if not.
1226 */
1227bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev,
1228 enum amd_ip_block_type block_type)
1229{
1230 int i;
1231
1232 for (i = 0; i < adev->num_ip_blocks; i++) {
1233 if (!adev->ip_blocks[i].status.valid)
1234 continue;
1235 if (adev->ip_blocks[i].version->type == block_type)
1236 return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
1237 }
1238 return true;
1239
1240}
1241
1242/**
1243 * amdgpu_device_ip_get_ip_block - get a hw IP pointer
1244 *
1245 * @adev: amdgpu_device pointer
1246 * @type: Type of hardware IP (SMU, GFX, UVD, etc.)
1247 *
1248 * Returns a pointer to the hardware IP block structure
1249 * if it exists for the asic, otherwise NULL.
1250 */
1251struct amdgpu_ip_block *
1252amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,
1253 enum amd_ip_block_type type)
1254{
1255 int i;
1256
1257 for (i = 0; i < adev->num_ip_blocks; i++)
1258 if (adev->ip_blocks[i].version->type == type)
1259 return &adev->ip_blocks[i];
1260
1261 return NULL;
1262}
1263
1264/**
1265 * amdgpu_device_ip_block_version_cmp
1266 *
1267 * @adev: amdgpu_device pointer
1268 * @type: enum amd_ip_block_type
1269 * @major: major version
1270 * @minor: minor version
1271 *
1272 * return 0 if equal or greater
1273 * return 1 if smaller or the ip_block doesn't exist
1274 */
1275int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev,
1276 enum amd_ip_block_type type,
1277 u32 major, u32 minor)
1278{
1279 struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type);
1280
1281 if (ip_block && ((ip_block->version->major > major) ||
1282 ((ip_block->version->major == major) &&
1283 (ip_block->version->minor >= minor))))
1284 return 0;
1285
1286 return 1;
1287}
1288
1289/**
1290 * amdgpu_device_ip_block_add
1291 *
1292 * @adev: amdgpu_device pointer
1293 * @ip_block_version: pointer to the IP to add
1294 *
1295 * Adds the IP block driver information to the collection of IPs
1296 * on the asic.
1297 */
1298int amdgpu_device_ip_block_add(struct amdgpu_device *adev,
1299 const struct amdgpu_ip_block_version *ip_block_version)
1300{
1301 if (!ip_block_version)
1302 return -EINVAL;
1303
1304 DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks,
1305 ip_block_version->funcs->name);
1306
1307 adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
1308
1309 return 0;
1310}
1311
1312/**
1313 * amdgpu_device_enable_virtual_display - enable virtual display feature
1314 *
1315 * @adev: amdgpu_device pointer
1316 *
1317 * Enabled the virtual display feature if the user has enabled it via
1318 * the module parameter virtual_display. This feature provides a virtual
1319 * display hardware on headless boards or in virtualized environments.
1320 * This function parses and validates the configuration string specified by
1321 * the user and configues the virtual display configuration (number of
1322 * virtual connectors, crtcs, etc.) specified.
1323 */
1324static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
1325{
1326 adev->enable_virtual_display = false;
1327
1328 if (amdgpu_virtual_display) {
1329 struct drm_device *ddev = adev->ddev;
1330 const char *pci_address_name = pci_name(ddev->pdev);
1331 char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
1332
1333 pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
1334 pciaddstr_tmp = pciaddstr;
1335 while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
1336 pciaddname = strsep(&pciaddname_tmp, ",");
1337 if (!strcmp("all", pciaddname)
1338 || !strcmp(pci_address_name, pciaddname)) {
1339 long num_crtc;
1340 int res = -1;
1341
1342 adev->enable_virtual_display = true;
1343
1344 if (pciaddname_tmp)
1345 res = kstrtol(pciaddname_tmp, 10,
1346 &num_crtc);
1347
1348 if (!res) {
1349 if (num_crtc < 1)
1350 num_crtc = 1;
1351 if (num_crtc > 6)
1352 num_crtc = 6;
1353 adev->mode_info.num_crtc = num_crtc;
1354 } else {
1355 adev->mode_info.num_crtc = 1;
1356 }
1357 break;
1358 }
1359 }
1360
1361 DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
1362 amdgpu_virtual_display, pci_address_name,
1363 adev->enable_virtual_display, adev->mode_info.num_crtc);
1364
1365 kfree(pciaddstr);
1366 }
1367}
1368
1369/**
1370 * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware
1371 *
1372 * @adev: amdgpu_device pointer
1373 *
1374 * Parses the asic configuration parameters specified in the gpu info
1375 * firmware and makes them availale to the driver for use in configuring
1376 * the asic.
1377 * Returns 0 on success, -EINVAL on failure.
1378 */
1379static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev)
1380{
1381 const char *chip_name;
1382 char fw_name[30];
1383 int err;
1384 const struct gpu_info_firmware_header_v1_0 *hdr;
1385
1386 adev->firmware.gpu_info_fw = NULL;
1387
1388 switch (adev->asic_type) {
1389 case CHIP_TOPAZ:
1390 case CHIP_TONGA:
1391 case CHIP_FIJI:
1392 case CHIP_POLARIS10:
1393 case CHIP_POLARIS11:
1394 case CHIP_POLARIS12:
1395 case CHIP_VEGAM:
1396 case CHIP_CARRIZO:
1397 case CHIP_STONEY:
1398#ifdef CONFIG_DRM_AMDGPU_SI
1399 case CHIP_VERDE:
1400 case CHIP_TAHITI:
1401 case CHIP_PITCAIRN:
1402 case CHIP_OLAND:
1403 case CHIP_HAINAN:
1404#endif
1405#ifdef CONFIG_DRM_AMDGPU_CIK
1406 case CHIP_BONAIRE:
1407 case CHIP_HAWAII:
1408 case CHIP_KAVERI:
1409 case CHIP_KABINI:
1410 case CHIP_MULLINS:
1411#endif
1412 case CHIP_VEGA20:
1413 default:
1414 return 0;
1415 case CHIP_VEGA10:
1416 chip_name = "vega10";
1417 break;
1418 case CHIP_VEGA12:
1419 chip_name = "vega12";
1420 break;
1421 case CHIP_RAVEN:
1422 if (adev->rev_id >= 8)
1423 chip_name = "raven2";
1424 else if (adev->pdev->device == 0x15d8)
1425 chip_name = "picasso";
1426 else
1427 chip_name = "raven";
1428 break;
1429 case CHIP_ARCTURUS:
1430 chip_name = "arcturus";
1431 break;
1432 case CHIP_RENOIR:
1433 chip_name = "renoir";
1434 break;
1435 case CHIP_NAVI10:
1436 chip_name = "navi10";
1437 break;
1438 case CHIP_NAVI14:
1439 chip_name = "navi14";
1440 break;
1441 case CHIP_NAVI12:
1442 chip_name = "navi12";
1443 break;
1444 }
1445
1446 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name);
1447 err = request_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev);
1448 if (err) {
1449 dev_err(adev->dev,
1450 "Failed to load gpu_info firmware \"%s\"\n",
1451 fw_name);
1452 goto out;
1453 }
1454 err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw);
1455 if (err) {
1456 dev_err(adev->dev,
1457 "Failed to validate gpu_info firmware \"%s\"\n",
1458 fw_name);
1459 goto out;
1460 }
1461
1462 hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data;
1463 amdgpu_ucode_print_gpu_info_hdr(&hdr->header);
1464
1465 switch (hdr->version_major) {
1466 case 1:
1467 {
1468 const struct gpu_info_firmware_v1_0 *gpu_info_fw =
1469 (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data +
1470 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1471
1472 adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se);
1473 adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh);
1474 adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se);
1475 adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se);
1476 adev->gfx.config.max_texture_channel_caches =
1477 le32_to_cpu(gpu_info_fw->gc_num_tccs);
1478 adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs);
1479 adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds);
1480 adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth);
1481 adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth);
1482 adev->gfx.config.double_offchip_lds_buf =
1483 le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer);
1484 adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size);
1485 adev->gfx.cu_info.max_waves_per_simd =
1486 le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd);
1487 adev->gfx.cu_info.max_scratch_slots_per_cu =
1488 le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu);
1489 adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size);
1490 if (hdr->version_minor >= 1) {
1491 const struct gpu_info_firmware_v1_1 *gpu_info_fw =
1492 (const struct gpu_info_firmware_v1_1 *)(adev->firmware.gpu_info_fw->data +
1493 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1494 adev->gfx.config.num_sc_per_sh =
1495 le32_to_cpu(gpu_info_fw->num_sc_per_sh);
1496 adev->gfx.config.num_packer_per_sc =
1497 le32_to_cpu(gpu_info_fw->num_packer_per_sc);
1498 }
1499#ifdef CONFIG_DRM_AMD_DC_DCN2_0
1500 if (hdr->version_minor == 2) {
1501 const struct gpu_info_firmware_v1_2 *gpu_info_fw =
1502 (const struct gpu_info_firmware_v1_2 *)(adev->firmware.gpu_info_fw->data +
1503 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1504 adev->dm.soc_bounding_box = &gpu_info_fw->soc_bounding_box;
1505 }
1506#endif
1507 break;
1508 }
1509 default:
1510 dev_err(adev->dev,
1511 "Unsupported gpu_info table %d\n", hdr->header.ucode_version);
1512 err = -EINVAL;
1513 goto out;
1514 }
1515out:
1516 return err;
1517}
1518
1519/**
1520 * amdgpu_device_ip_early_init - run early init for hardware IPs
1521 *
1522 * @adev: amdgpu_device pointer
1523 *
1524 * Early initialization pass for hardware IPs. The hardware IPs that make
1525 * up each asic are discovered each IP's early_init callback is run. This
1526 * is the first stage in initializing the asic.
1527 * Returns 0 on success, negative error code on failure.
1528 */
1529static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
1530{
1531 int i, r;
1532
1533 amdgpu_device_enable_virtual_display(adev);
1534
1535 switch (adev->asic_type) {
1536 case CHIP_TOPAZ:
1537 case CHIP_TONGA:
1538 case CHIP_FIJI:
1539 case CHIP_POLARIS10:
1540 case CHIP_POLARIS11:
1541 case CHIP_POLARIS12:
1542 case CHIP_VEGAM:
1543 case CHIP_CARRIZO:
1544 case CHIP_STONEY:
1545 if (adev->asic_type == CHIP_CARRIZO || adev->asic_type == CHIP_STONEY)
1546 adev->family = AMDGPU_FAMILY_CZ;
1547 else
1548 adev->family = AMDGPU_FAMILY_VI;
1549
1550 r = vi_set_ip_blocks(adev);
1551 if (r)
1552 return r;
1553 break;
1554#ifdef CONFIG_DRM_AMDGPU_SI
1555 case CHIP_VERDE:
1556 case CHIP_TAHITI:
1557 case CHIP_PITCAIRN:
1558 case CHIP_OLAND:
1559 case CHIP_HAINAN:
1560 adev->family = AMDGPU_FAMILY_SI;
1561 r = si_set_ip_blocks(adev);
1562 if (r)
1563 return r;
1564 break;
1565#endif
1566#ifdef CONFIG_DRM_AMDGPU_CIK
1567 case CHIP_BONAIRE:
1568 case CHIP_HAWAII:
1569 case CHIP_KAVERI:
1570 case CHIP_KABINI:
1571 case CHIP_MULLINS:
1572 if ((adev->asic_type == CHIP_BONAIRE) || (adev->asic_type == CHIP_HAWAII))
1573 adev->family = AMDGPU_FAMILY_CI;
1574 else
1575 adev->family = AMDGPU_FAMILY_KV;
1576
1577 r = cik_set_ip_blocks(adev);
1578 if (r)
1579 return r;
1580 break;
1581#endif
1582 case CHIP_VEGA10:
1583 case CHIP_VEGA12:
1584 case CHIP_VEGA20:
1585 case CHIP_RAVEN:
1586 case CHIP_ARCTURUS:
1587 case CHIP_RENOIR:
1588 if (adev->asic_type == CHIP_RAVEN ||
1589 adev->asic_type == CHIP_RENOIR)
1590 adev->family = AMDGPU_FAMILY_RV;
1591 else
1592 adev->family = AMDGPU_FAMILY_AI;
1593
1594 r = soc15_set_ip_blocks(adev);
1595 if (r)
1596 return r;
1597 break;
1598 case CHIP_NAVI10:
1599 case CHIP_NAVI14:
1600 case CHIP_NAVI12:
1601 adev->family = AMDGPU_FAMILY_NV;
1602
1603 r = nv_set_ip_blocks(adev);
1604 if (r)
1605 return r;
1606 break;
1607 default:
1608 /* FIXME: not supported yet */
1609 return -EINVAL;
1610 }
1611
1612 r = amdgpu_device_parse_gpu_info_fw(adev);
1613 if (r)
1614 return r;
1615
1616 amdgpu_amdkfd_device_probe(adev);
1617
1618 if (amdgpu_sriov_vf(adev)) {
1619 r = amdgpu_virt_request_full_gpu(adev, true);
1620 if (r)
1621 return -EAGAIN;
1622 }
1623
1624 adev->pm.pp_feature = amdgpu_pp_feature_mask;
1625 if (amdgpu_sriov_vf(adev))
1626 adev->pm.pp_feature &= ~PP_GFXOFF_MASK;
1627
1628 for (i = 0; i < adev->num_ip_blocks; i++) {
1629 if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
1630 DRM_ERROR("disabled ip block: %d <%s>\n",
1631 i, adev->ip_blocks[i].version->funcs->name);
1632 adev->ip_blocks[i].status.valid = false;
1633 } else {
1634 if (adev->ip_blocks[i].version->funcs->early_init) {
1635 r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
1636 if (r == -ENOENT) {
1637 adev->ip_blocks[i].status.valid = false;
1638 } else if (r) {
1639 DRM_ERROR("early_init of IP block <%s> failed %d\n",
1640 adev->ip_blocks[i].version->funcs->name, r);
1641 return r;
1642 } else {
1643 adev->ip_blocks[i].status.valid = true;
1644 }
1645 } else {
1646 adev->ip_blocks[i].status.valid = true;
1647 }
1648 }
1649 /* get the vbios after the asic_funcs are set up */
1650 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) {
1651 /* Read BIOS */
1652 if (!amdgpu_get_bios(adev))
1653 return -EINVAL;
1654
1655 r = amdgpu_atombios_init(adev);
1656 if (r) {
1657 dev_err(adev->dev, "amdgpu_atombios_init failed\n");
1658 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0);
1659 return r;
1660 }
1661 }
1662 }
1663
1664 adev->cg_flags &= amdgpu_cg_mask;
1665 adev->pg_flags &= amdgpu_pg_mask;
1666
1667 return 0;
1668}
1669
1670static int amdgpu_device_ip_hw_init_phase1(struct amdgpu_device *adev)
1671{
1672 int i, r;
1673
1674 for (i = 0; i < adev->num_ip_blocks; i++) {
1675 if (!adev->ip_blocks[i].status.sw)
1676 continue;
1677 if (adev->ip_blocks[i].status.hw)
1678 continue;
1679 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
1680 (amdgpu_sriov_vf(adev) && (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)) ||
1681 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
1682 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
1683 if (r) {
1684 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1685 adev->ip_blocks[i].version->funcs->name, r);
1686 return r;
1687 }
1688 adev->ip_blocks[i].status.hw = true;
1689 }
1690 }
1691
1692 return 0;
1693}
1694
1695static int amdgpu_device_ip_hw_init_phase2(struct amdgpu_device *adev)
1696{
1697 int i, r;
1698
1699 for (i = 0; i < adev->num_ip_blocks; i++) {
1700 if (!adev->ip_blocks[i].status.sw)
1701 continue;
1702 if (adev->ip_blocks[i].status.hw)
1703 continue;
1704 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
1705 if (r) {
1706 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1707 adev->ip_blocks[i].version->funcs->name, r);
1708 return r;
1709 }
1710 adev->ip_blocks[i].status.hw = true;
1711 }
1712
1713 return 0;
1714}
1715
1716static int amdgpu_device_fw_loading(struct amdgpu_device *adev)
1717{
1718 int r = 0;
1719 int i;
1720 uint32_t smu_version;
1721
1722 if (adev->asic_type >= CHIP_VEGA10) {
1723 for (i = 0; i < adev->num_ip_blocks; i++) {
1724 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_PSP)
1725 continue;
1726
1727 /* no need to do the fw loading again if already done*/
1728 if (adev->ip_blocks[i].status.hw == true)
1729 break;
1730
1731 if (adev->in_gpu_reset || adev->in_suspend) {
1732 r = adev->ip_blocks[i].version->funcs->resume(adev);
1733 if (r) {
1734 DRM_ERROR("resume of IP block <%s> failed %d\n",
1735 adev->ip_blocks[i].version->funcs->name, r);
1736 return r;
1737 }
1738 } else {
1739 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
1740 if (r) {
1741 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1742 adev->ip_blocks[i].version->funcs->name, r);
1743 return r;
1744 }
1745 }
1746
1747 adev->ip_blocks[i].status.hw = true;
1748 break;
1749 }
1750 }
1751
1752 r = amdgpu_pm_load_smu_firmware(adev, &smu_version);
1753
1754 return r;
1755}
1756
1757/**
1758 * amdgpu_device_ip_init - run init for hardware IPs
1759 *
1760 * @adev: amdgpu_device pointer
1761 *
1762 * Main initialization pass for hardware IPs. The list of all the hardware
1763 * IPs that make up the asic is walked and the sw_init and hw_init callbacks
1764 * are run. sw_init initializes the software state associated with each IP
1765 * and hw_init initializes the hardware associated with each IP.
1766 * Returns 0 on success, negative error code on failure.
1767 */
1768static int amdgpu_device_ip_init(struct amdgpu_device *adev)
1769{
1770 int i, r;
1771
1772 r = amdgpu_ras_init(adev);
1773 if (r)
1774 return r;
1775
1776 for (i = 0; i < adev->num_ip_blocks; i++) {
1777 if (!adev->ip_blocks[i].status.valid)
1778 continue;
1779 r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
1780 if (r) {
1781 DRM_ERROR("sw_init of IP block <%s> failed %d\n",
1782 adev->ip_blocks[i].version->funcs->name, r);
1783 goto init_failed;
1784 }
1785 adev->ip_blocks[i].status.sw = true;
1786
1787 /* need to do gmc hw init early so we can allocate gpu mem */
1788 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1789 r = amdgpu_device_vram_scratch_init(adev);
1790 if (r) {
1791 DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r);
1792 goto init_failed;
1793 }
1794 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1795 if (r) {
1796 DRM_ERROR("hw_init %d failed %d\n", i, r);
1797 goto init_failed;
1798 }
1799 r = amdgpu_device_wb_init(adev);
1800 if (r) {
1801 DRM_ERROR("amdgpu_device_wb_init failed %d\n", r);
1802 goto init_failed;
1803 }
1804 adev->ip_blocks[i].status.hw = true;
1805
1806 /* right after GMC hw init, we create CSA */
1807 if (amdgpu_mcbp || amdgpu_sriov_vf(adev)) {
1808 r = amdgpu_allocate_static_csa(adev, &adev->virt.csa_obj,
1809 AMDGPU_GEM_DOMAIN_VRAM,
1810 AMDGPU_CSA_SIZE);
1811 if (r) {
1812 DRM_ERROR("allocate CSA failed %d\n", r);
1813 goto init_failed;
1814 }
1815 }
1816 }
1817 }
1818
1819 r = amdgpu_ib_pool_init(adev);
1820 if (r) {
1821 dev_err(adev->dev, "IB initialization failed (%d).\n", r);
1822 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r);
1823 goto init_failed;
1824 }
1825
1826 r = amdgpu_ucode_create_bo(adev); /* create ucode bo when sw_init complete*/
1827 if (r)
1828 goto init_failed;
1829
1830 r = amdgpu_device_ip_hw_init_phase1(adev);
1831 if (r)
1832 goto init_failed;
1833
1834 r = amdgpu_device_fw_loading(adev);
1835 if (r)
1836 goto init_failed;
1837
1838 r = amdgpu_device_ip_hw_init_phase2(adev);
1839 if (r)
1840 goto init_failed;
1841
1842 if (adev->gmc.xgmi.num_physical_nodes > 1)
1843 amdgpu_xgmi_add_device(adev);
1844 amdgpu_amdkfd_device_init(adev);
1845
1846init_failed:
1847 if (amdgpu_sriov_vf(adev)) {
1848 if (!r)
1849 amdgpu_virt_init_data_exchange(adev);
1850 amdgpu_virt_release_full_gpu(adev, true);
1851 }
1852
1853 return r;
1854}
1855
1856/**
1857 * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer
1858 *
1859 * @adev: amdgpu_device pointer
1860 *
1861 * Writes a reset magic value to the gart pointer in VRAM. The driver calls
1862 * this function before a GPU reset. If the value is retained after a
1863 * GPU reset, VRAM has not been lost. Some GPU resets may destry VRAM contents.
1864 */
1865static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev)
1866{
1867 memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
1868}
1869
1870/**
1871 * amdgpu_device_check_vram_lost - check if vram is valid
1872 *
1873 * @adev: amdgpu_device pointer
1874 *
1875 * Checks the reset magic value written to the gart pointer in VRAM.
1876 * The driver calls this after a GPU reset to see if the contents of
1877 * VRAM is lost or now.
1878 * returns true if vram is lost, false if not.
1879 */
1880static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev)
1881{
1882 return !!memcmp(adev->gart.ptr, adev->reset_magic,
1883 AMDGPU_RESET_MAGIC_NUM);
1884}
1885
1886/**
1887 * amdgpu_device_set_cg_state - set clockgating for amdgpu device
1888 *
1889 * @adev: amdgpu_device pointer
1890 *
1891 * The list of all the hardware IPs that make up the asic is walked and the
1892 * set_clockgating_state callbacks are run.
1893 * Late initialization pass enabling clockgating for hardware IPs.
1894 * Fini or suspend, pass disabling clockgating for hardware IPs.
1895 * Returns 0 on success, negative error code on failure.
1896 */
1897
1898static int amdgpu_device_set_cg_state(struct amdgpu_device *adev,
1899 enum amd_clockgating_state state)
1900{
1901 int i, j, r;
1902
1903 if (amdgpu_emu_mode == 1)
1904 return 0;
1905
1906 for (j = 0; j < adev->num_ip_blocks; j++) {
1907 i = state == AMD_CG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
1908 if (!adev->ip_blocks[i].status.late_initialized)
1909 continue;
1910 /* skip CG for VCE/UVD, it's handled specially */
1911 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1912 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1913 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
1914 adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1915 /* enable clockgating to save power */
1916 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1917 state);
1918 if (r) {
1919 DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
1920 adev->ip_blocks[i].version->funcs->name, r);
1921 return r;
1922 }
1923 }
1924 }
1925
1926 return 0;
1927}
1928
1929static int amdgpu_device_set_pg_state(struct amdgpu_device *adev, enum amd_powergating_state state)
1930{
1931 int i, j, r;
1932
1933 if (amdgpu_emu_mode == 1)
1934 return 0;
1935
1936 for (j = 0; j < adev->num_ip_blocks; j++) {
1937 i = state == AMD_PG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
1938 if (!adev->ip_blocks[i].status.late_initialized)
1939 continue;
1940 /* skip CG for VCE/UVD, it's handled specially */
1941 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1942 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1943 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
1944 adev->ip_blocks[i].version->funcs->set_powergating_state) {
1945 /* enable powergating to save power */
1946 r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev,
1947 state);
1948 if (r) {
1949 DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n",
1950 adev->ip_blocks[i].version->funcs->name, r);
1951 return r;
1952 }
1953 }
1954 }
1955 return 0;
1956}
1957
1958static int amdgpu_device_enable_mgpu_fan_boost(void)
1959{
1960 struct amdgpu_gpu_instance *gpu_ins;
1961 struct amdgpu_device *adev;
1962 int i, ret = 0;
1963
1964 mutex_lock(&mgpu_info.mutex);
1965
1966 /*
1967 * MGPU fan boost feature should be enabled
1968 * only when there are two or more dGPUs in
1969 * the system
1970 */
1971 if (mgpu_info.num_dgpu < 2)
1972 goto out;
1973
1974 for (i = 0; i < mgpu_info.num_dgpu; i++) {
1975 gpu_ins = &(mgpu_info.gpu_ins[i]);
1976 adev = gpu_ins->adev;
1977 if (!(adev->flags & AMD_IS_APU) &&
1978 !gpu_ins->mgpu_fan_enabled &&
1979 adev->powerplay.pp_funcs &&
1980 adev->powerplay.pp_funcs->enable_mgpu_fan_boost) {
1981 ret = amdgpu_dpm_enable_mgpu_fan_boost(adev);
1982 if (ret)
1983 break;
1984
1985 gpu_ins->mgpu_fan_enabled = 1;
1986 }
1987 }
1988
1989out:
1990 mutex_unlock(&mgpu_info.mutex);
1991
1992 return ret;
1993}
1994
1995/**
1996 * amdgpu_device_ip_late_init - run late init for hardware IPs
1997 *
1998 * @adev: amdgpu_device pointer
1999 *
2000 * Late initialization pass for hardware IPs. The list of all the hardware
2001 * IPs that make up the asic is walked and the late_init callbacks are run.
2002 * late_init covers any special initialization that an IP requires
2003 * after all of the have been initialized or something that needs to happen
2004 * late in the init process.
2005 * Returns 0 on success, negative error code on failure.
2006 */
2007static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
2008{
2009 int i = 0, r;
2010
2011 for (i = 0; i < adev->num_ip_blocks; i++) {
2012 if (!adev->ip_blocks[i].status.hw)
2013 continue;
2014 if (adev->ip_blocks[i].version->funcs->late_init) {
2015 r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
2016 if (r) {
2017 DRM_ERROR("late_init of IP block <%s> failed %d\n",
2018 adev->ip_blocks[i].version->funcs->name, r);
2019 return r;
2020 }
2021 }
2022 adev->ip_blocks[i].status.late_initialized = true;
2023 }
2024
2025 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE);
2026 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE);
2027
2028 amdgpu_device_fill_reset_magic(adev);
2029
2030 r = amdgpu_device_enable_mgpu_fan_boost();
2031 if (r)
2032 DRM_ERROR("enable mgpu fan boost failed (%d).\n", r);
2033
2034 /* set to low pstate by default */
2035 amdgpu_xgmi_set_pstate(adev, 0);
2036
2037 return 0;
2038}
2039
2040/**
2041 * amdgpu_device_ip_fini - run fini for hardware IPs
2042 *
2043 * @adev: amdgpu_device pointer
2044 *
2045 * Main teardown pass for hardware IPs. The list of all the hardware
2046 * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks
2047 * are run. hw_fini tears down the hardware associated with each IP
2048 * and sw_fini tears down any software state associated with each IP.
2049 * Returns 0 on success, negative error code on failure.
2050 */
2051static int amdgpu_device_ip_fini(struct amdgpu_device *adev)
2052{
2053 int i, r;
2054
2055 amdgpu_ras_pre_fini(adev);
2056
2057 if (adev->gmc.xgmi.num_physical_nodes > 1)
2058 amdgpu_xgmi_remove_device(adev);
2059
2060 amdgpu_amdkfd_device_fini(adev);
2061
2062 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
2063 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
2064
2065 /* need to disable SMC first */
2066 for (i = 0; i < adev->num_ip_blocks; i++) {
2067 if (!adev->ip_blocks[i].status.hw)
2068 continue;
2069 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
2070 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
2071 /* XXX handle errors */
2072 if (r) {
2073 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
2074 adev->ip_blocks[i].version->funcs->name, r);
2075 }
2076 adev->ip_blocks[i].status.hw = false;
2077 break;
2078 }
2079 }
2080
2081 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2082 if (!adev->ip_blocks[i].status.hw)
2083 continue;
2084
2085 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
2086 /* XXX handle errors */
2087 if (r) {
2088 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
2089 adev->ip_blocks[i].version->funcs->name, r);
2090 }
2091
2092 adev->ip_blocks[i].status.hw = false;
2093 }
2094
2095
2096 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2097 if (!adev->ip_blocks[i].status.sw)
2098 continue;
2099
2100 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
2101 amdgpu_ucode_free_bo(adev);
2102 amdgpu_free_static_csa(&adev->virt.csa_obj);
2103 amdgpu_device_wb_fini(adev);
2104 amdgpu_device_vram_scratch_fini(adev);
2105 amdgpu_ib_pool_fini(adev);
2106 }
2107
2108 r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
2109 /* XXX handle errors */
2110 if (r) {
2111 DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
2112 adev->ip_blocks[i].version->funcs->name, r);
2113 }
2114 adev->ip_blocks[i].status.sw = false;
2115 adev->ip_blocks[i].status.valid = false;
2116 }
2117
2118 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2119 if (!adev->ip_blocks[i].status.late_initialized)
2120 continue;
2121 if (adev->ip_blocks[i].version->funcs->late_fini)
2122 adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
2123 adev->ip_blocks[i].status.late_initialized = false;
2124 }
2125
2126 amdgpu_ras_fini(adev);
2127
2128 if (amdgpu_sriov_vf(adev))
2129 if (amdgpu_virt_release_full_gpu(adev, false))
2130 DRM_ERROR("failed to release exclusive mode on fini\n");
2131
2132 return 0;
2133}
2134
2135/**
2136 * amdgpu_device_delayed_init_work_handler - work handler for IB tests
2137 *
2138 * @work: work_struct.
2139 */
2140static void amdgpu_device_delayed_init_work_handler(struct work_struct *work)
2141{
2142 struct amdgpu_device *adev =
2143 container_of(work, struct amdgpu_device, delayed_init_work.work);
2144 int r;
2145
2146 r = amdgpu_ib_ring_tests(adev);
2147 if (r)
2148 DRM_ERROR("ib ring test failed (%d).\n", r);
2149}
2150
2151static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work)
2152{
2153 struct amdgpu_device *adev =
2154 container_of(work, struct amdgpu_device, gfx.gfx_off_delay_work.work);
2155
2156 mutex_lock(&adev->gfx.gfx_off_mutex);
2157 if (!adev->gfx.gfx_off_state && !adev->gfx.gfx_off_req_count) {
2158 if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, true))
2159 adev->gfx.gfx_off_state = true;
2160 }
2161 mutex_unlock(&adev->gfx.gfx_off_mutex);
2162}
2163
2164/**
2165 * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1)
2166 *
2167 * @adev: amdgpu_device pointer
2168 *
2169 * Main suspend function for hardware IPs. The list of all the hardware
2170 * IPs that make up the asic is walked, clockgating is disabled and the
2171 * suspend callbacks are run. suspend puts the hardware and software state
2172 * in each IP into a state suitable for suspend.
2173 * Returns 0 on success, negative error code on failure.
2174 */
2175static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev)
2176{
2177 int i, r;
2178
2179 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
2180 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
2181
2182 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2183 if (!adev->ip_blocks[i].status.valid)
2184 continue;
2185 /* displays are handled separately */
2186 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) {
2187 /* XXX handle errors */
2188 r = adev->ip_blocks[i].version->funcs->suspend(adev);
2189 /* XXX handle errors */
2190 if (r) {
2191 DRM_ERROR("suspend of IP block <%s> failed %d\n",
2192 adev->ip_blocks[i].version->funcs->name, r);
2193 return r;
2194 }
2195 adev->ip_blocks[i].status.hw = false;
2196 }
2197 }
2198
2199 return 0;
2200}
2201
2202/**
2203 * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2)
2204 *
2205 * @adev: amdgpu_device pointer
2206 *
2207 * Main suspend function for hardware IPs. The list of all the hardware
2208 * IPs that make up the asic is walked, clockgating is disabled and the
2209 * suspend callbacks are run. suspend puts the hardware and software state
2210 * in each IP into a state suitable for suspend.
2211 * Returns 0 on success, negative error code on failure.
2212 */
2213static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev)
2214{
2215 int i, r;
2216
2217 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2218 if (!adev->ip_blocks[i].status.valid)
2219 continue;
2220 /* displays are handled in phase1 */
2221 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE)
2222 continue;
2223 /* XXX handle errors */
2224 r = adev->ip_blocks[i].version->funcs->suspend(adev);
2225 /* XXX handle errors */
2226 if (r) {
2227 DRM_ERROR("suspend of IP block <%s> failed %d\n",
2228 adev->ip_blocks[i].version->funcs->name, r);
2229 }
2230 adev->ip_blocks[i].status.hw = false;
2231 /* handle putting the SMC in the appropriate state */
2232 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
2233 if (is_support_sw_smu(adev)) {
2234 /* todo */
2235 } else if (adev->powerplay.pp_funcs &&
2236 adev->powerplay.pp_funcs->set_mp1_state) {
2237 r = adev->powerplay.pp_funcs->set_mp1_state(
2238 adev->powerplay.pp_handle,
2239 adev->mp1_state);
2240 if (r) {
2241 DRM_ERROR("SMC failed to set mp1 state %d, %d\n",
2242 adev->mp1_state, r);
2243 return r;
2244 }
2245 }
2246 }
2247
2248 adev->ip_blocks[i].status.hw = false;
2249 }
2250
2251 return 0;
2252}
2253
2254/**
2255 * amdgpu_device_ip_suspend - run suspend for hardware IPs
2256 *
2257 * @adev: amdgpu_device pointer
2258 *
2259 * Main suspend function for hardware IPs. The list of all the hardware
2260 * IPs that make up the asic is walked, clockgating is disabled and the
2261 * suspend callbacks are run. suspend puts the hardware and software state
2262 * in each IP into a state suitable for suspend.
2263 * Returns 0 on success, negative error code on failure.
2264 */
2265int amdgpu_device_ip_suspend(struct amdgpu_device *adev)
2266{
2267 int r;
2268
2269 if (amdgpu_sriov_vf(adev))
2270 amdgpu_virt_request_full_gpu(adev, false);
2271
2272 r = amdgpu_device_ip_suspend_phase1(adev);
2273 if (r)
2274 return r;
2275 r = amdgpu_device_ip_suspend_phase2(adev);
2276
2277 if (amdgpu_sriov_vf(adev))
2278 amdgpu_virt_release_full_gpu(adev, false);
2279
2280 return r;
2281}
2282
2283static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
2284{
2285 int i, r;
2286
2287 static enum amd_ip_block_type ip_order[] = {
2288 AMD_IP_BLOCK_TYPE_GMC,
2289 AMD_IP_BLOCK_TYPE_COMMON,
2290 AMD_IP_BLOCK_TYPE_PSP,
2291 AMD_IP_BLOCK_TYPE_IH,
2292 };
2293
2294 for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
2295 int j;
2296 struct amdgpu_ip_block *block;
2297
2298 for (j = 0; j < adev->num_ip_blocks; j++) {
2299 block = &adev->ip_blocks[j];
2300
2301 block->status.hw = false;
2302 if (block->version->type != ip_order[i] ||
2303 !block->status.valid)
2304 continue;
2305
2306 r = block->version->funcs->hw_init(adev);
2307 DRM_INFO("RE-INIT-early: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
2308 if (r)
2309 return r;
2310 block->status.hw = true;
2311 }
2312 }
2313
2314 return 0;
2315}
2316
2317static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
2318{
2319 int i, r;
2320
2321 static enum amd_ip_block_type ip_order[] = {
2322 AMD_IP_BLOCK_TYPE_SMC,
2323 AMD_IP_BLOCK_TYPE_DCE,
2324 AMD_IP_BLOCK_TYPE_GFX,
2325 AMD_IP_BLOCK_TYPE_SDMA,
2326 AMD_IP_BLOCK_TYPE_UVD,
2327 AMD_IP_BLOCK_TYPE_VCE
2328 };
2329
2330 for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
2331 int j;
2332 struct amdgpu_ip_block *block;
2333
2334 for (j = 0; j < adev->num_ip_blocks; j++) {
2335 block = &adev->ip_blocks[j];
2336
2337 if (block->version->type != ip_order[i] ||
2338 !block->status.valid ||
2339 block->status.hw)
2340 continue;
2341
2342 r = block->version->funcs->hw_init(adev);
2343 DRM_INFO("RE-INIT-late: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
2344 if (r)
2345 return r;
2346 block->status.hw = true;
2347 }
2348 }
2349
2350 return 0;
2351}
2352
2353/**
2354 * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs
2355 *
2356 * @adev: amdgpu_device pointer
2357 *
2358 * First resume function for hardware IPs. The list of all the hardware
2359 * IPs that make up the asic is walked and the resume callbacks are run for
2360 * COMMON, GMC, and IH. resume puts the hardware into a functional state
2361 * after a suspend and updates the software state as necessary. This
2362 * function is also used for restoring the GPU after a GPU reset.
2363 * Returns 0 on success, negative error code on failure.
2364 */
2365static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)
2366{
2367 int i, r;
2368
2369 for (i = 0; i < adev->num_ip_blocks; i++) {
2370 if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
2371 continue;
2372 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2373 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2374 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
2375
2376 r = adev->ip_blocks[i].version->funcs->resume(adev);
2377 if (r) {
2378 DRM_ERROR("resume of IP block <%s> failed %d\n",
2379 adev->ip_blocks[i].version->funcs->name, r);
2380 return r;
2381 }
2382 adev->ip_blocks[i].status.hw = true;
2383 }
2384 }
2385
2386 return 0;
2387}
2388
2389/**
2390 * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs
2391 *
2392 * @adev: amdgpu_device pointer
2393 *
2394 * First resume function for hardware IPs. The list of all the hardware
2395 * IPs that make up the asic is walked and the resume callbacks are run for
2396 * all blocks except COMMON, GMC, and IH. resume puts the hardware into a
2397 * functional state after a suspend and updates the software state as
2398 * necessary. This function is also used for restoring the GPU after a GPU
2399 * reset.
2400 * Returns 0 on success, negative error code on failure.
2401 */
2402static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)
2403{
2404 int i, r;
2405
2406 for (i = 0; i < adev->num_ip_blocks; i++) {
2407 if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
2408 continue;
2409 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2410 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2411 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH ||
2412 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)
2413 continue;
2414 r = adev->ip_blocks[i].version->funcs->resume(adev);
2415 if (r) {
2416 DRM_ERROR("resume of IP block <%s> failed %d\n",
2417 adev->ip_blocks[i].version->funcs->name, r);
2418 return r;
2419 }
2420 adev->ip_blocks[i].status.hw = true;
2421 }
2422
2423 return 0;
2424}
2425
2426/**
2427 * amdgpu_device_ip_resume - run resume for hardware IPs
2428 *
2429 * @adev: amdgpu_device pointer
2430 *
2431 * Main resume function for hardware IPs. The hardware IPs
2432 * are split into two resume functions because they are
2433 * are also used in in recovering from a GPU reset and some additional
2434 * steps need to be take between them. In this case (S3/S4) they are
2435 * run sequentially.
2436 * Returns 0 on success, negative error code on failure.
2437 */
2438static int amdgpu_device_ip_resume(struct amdgpu_device *adev)
2439{
2440 int r;
2441
2442 r = amdgpu_device_ip_resume_phase1(adev);
2443 if (r)
2444 return r;
2445
2446 r = amdgpu_device_fw_loading(adev);
2447 if (r)
2448 return r;
2449
2450 r = amdgpu_device_ip_resume_phase2(adev);
2451
2452 return r;
2453}
2454
2455/**
2456 * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV
2457 *
2458 * @adev: amdgpu_device pointer
2459 *
2460 * Query the VBIOS data tables to determine if the board supports SR-IOV.
2461 */
2462static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
2463{
2464 if (amdgpu_sriov_vf(adev)) {
2465 if (adev->is_atom_fw) {
2466 if (amdgpu_atomfirmware_gpu_supports_virtualization(adev))
2467 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2468 } else {
2469 if (amdgpu_atombios_has_gpu_virtualization_table(adev))
2470 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2471 }
2472
2473 if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
2474 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
2475 }
2476}
2477
2478/**
2479 * amdgpu_device_asic_has_dc_support - determine if DC supports the asic
2480 *
2481 * @asic_type: AMD asic type
2482 *
2483 * Check if there is DC (new modesetting infrastructre) support for an asic.
2484 * returns true if DC has support, false if not.
2485 */
2486bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
2487{
2488 switch (asic_type) {
2489#if defined(CONFIG_DRM_AMD_DC)
2490 case CHIP_BONAIRE:
2491 case CHIP_KAVERI:
2492 case CHIP_KABINI:
2493 case CHIP_MULLINS:
2494 /*
2495 * We have systems in the wild with these ASICs that require
2496 * LVDS and VGA support which is not supported with DC.
2497 *
2498 * Fallback to the non-DC driver here by default so as not to
2499 * cause regressions.
2500 */
2501 return amdgpu_dc > 0;
2502 case CHIP_HAWAII:
2503 case CHIP_CARRIZO:
2504 case CHIP_STONEY:
2505 case CHIP_POLARIS10:
2506 case CHIP_POLARIS11:
2507 case CHIP_POLARIS12:
2508 case CHIP_VEGAM:
2509 case CHIP_TONGA:
2510 case CHIP_FIJI:
2511 case CHIP_VEGA10:
2512 case CHIP_VEGA12:
2513 case CHIP_VEGA20:
2514#if defined(CONFIG_DRM_AMD_DC_DCN1_0)
2515 case CHIP_RAVEN:
2516#endif
2517#if defined(CONFIG_DRM_AMD_DC_DCN2_0)
2518 case CHIP_NAVI10:
2519 case CHIP_NAVI14:
2520 case CHIP_NAVI12:
2521#endif
2522#if defined(CONFIG_DRM_AMD_DC_DCN2_1)
2523 case CHIP_RENOIR:
2524#endif
2525 return amdgpu_dc != 0;
2526#endif
2527 default:
2528 return false;
2529 }
2530}
2531
2532/**
2533 * amdgpu_device_has_dc_support - check if dc is supported
2534 *
2535 * @adev: amdgpu_device_pointer
2536 *
2537 * Returns true for supported, false for not supported
2538 */
2539bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
2540{
2541 if (amdgpu_sriov_vf(adev))
2542 return false;
2543
2544 return amdgpu_device_asic_has_dc_support(adev->asic_type);
2545}
2546
2547
2548static void amdgpu_device_xgmi_reset_func(struct work_struct *__work)
2549{
2550 struct amdgpu_device *adev =
2551 container_of(__work, struct amdgpu_device, xgmi_reset_work);
2552
2553 adev->asic_reset_res = amdgpu_asic_reset(adev);
2554 if (adev->asic_reset_res)
2555 DRM_WARN("ASIC reset failed with error, %d for drm dev, %s",
2556 adev->asic_reset_res, adev->ddev->unique);
2557}
2558
2559
2560/**
2561 * amdgpu_device_init - initialize the driver
2562 *
2563 * @adev: amdgpu_device pointer
2564 * @ddev: drm dev pointer
2565 * @pdev: pci dev pointer
2566 * @flags: driver flags
2567 *
2568 * Initializes the driver info and hw (all asics).
2569 * Returns 0 for success or an error on failure.
2570 * Called at driver startup.
2571 */
2572int amdgpu_device_init(struct amdgpu_device *adev,
2573 struct drm_device *ddev,
2574 struct pci_dev *pdev,
2575 uint32_t flags)
2576{
2577 int r, i;
2578 bool runtime = false;
2579 u32 max_MBps;
2580
2581 adev->shutdown = false;
2582 adev->dev = &pdev->dev;
2583 adev->ddev = ddev;
2584 adev->pdev = pdev;
2585 adev->flags = flags;
2586 adev->asic_type = flags & AMD_ASIC_MASK;
2587 adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
2588 if (amdgpu_emu_mode == 1)
2589 adev->usec_timeout *= 2;
2590 adev->gmc.gart_size = 512 * 1024 * 1024;
2591 adev->accel_working = false;
2592 adev->num_rings = 0;
2593 adev->mman.buffer_funcs = NULL;
2594 adev->mman.buffer_funcs_ring = NULL;
2595 adev->vm_manager.vm_pte_funcs = NULL;
2596 adev->vm_manager.vm_pte_num_rqs = 0;
2597 adev->gmc.gmc_funcs = NULL;
2598 adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
2599 bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
2600
2601 adev->smc_rreg = &amdgpu_invalid_rreg;
2602 adev->smc_wreg = &amdgpu_invalid_wreg;
2603 adev->pcie_rreg = &amdgpu_invalid_rreg;
2604 adev->pcie_wreg = &amdgpu_invalid_wreg;
2605 adev->pciep_rreg = &amdgpu_invalid_rreg;
2606 adev->pciep_wreg = &amdgpu_invalid_wreg;
2607 adev->pcie_rreg64 = &amdgpu_invalid_rreg64;
2608 adev->pcie_wreg64 = &amdgpu_invalid_wreg64;
2609 adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
2610 adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
2611 adev->didt_rreg = &amdgpu_invalid_rreg;
2612 adev->didt_wreg = &amdgpu_invalid_wreg;
2613 adev->gc_cac_rreg = &amdgpu_invalid_rreg;
2614 adev->gc_cac_wreg = &amdgpu_invalid_wreg;
2615 adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
2616 adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
2617
2618 DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
2619 amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
2620 pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
2621
2622 /* mutex initialization are all done here so we
2623 * can recall function without having locking issues */
2624 atomic_set(&adev->irq.ih.lock, 0);
2625 mutex_init(&adev->firmware.mutex);
2626 mutex_init(&adev->pm.mutex);
2627 mutex_init(&adev->gfx.gpu_clock_mutex);
2628 mutex_init(&adev->srbm_mutex);
2629 mutex_init(&adev->gfx.pipe_reserve_mutex);
2630 mutex_init(&adev->gfx.gfx_off_mutex);
2631 mutex_init(&adev->grbm_idx_mutex);
2632 mutex_init(&adev->mn_lock);
2633 mutex_init(&adev->virt.vf_errors.lock);
2634 hash_init(adev->mn_hash);
2635 mutex_init(&adev->lock_reset);
2636 mutex_init(&adev->virt.dpm_mutex);
2637 mutex_init(&adev->psp.mutex);
2638
2639 r = amdgpu_device_check_arguments(adev);
2640 if (r)
2641 return r;
2642
2643 spin_lock_init(&adev->mmio_idx_lock);
2644 spin_lock_init(&adev->smc_idx_lock);
2645 spin_lock_init(&adev->pcie_idx_lock);
2646 spin_lock_init(&adev->uvd_ctx_idx_lock);
2647 spin_lock_init(&adev->didt_idx_lock);
2648 spin_lock_init(&adev->gc_cac_idx_lock);
2649 spin_lock_init(&adev->se_cac_idx_lock);
2650 spin_lock_init(&adev->audio_endpt_idx_lock);
2651 spin_lock_init(&adev->mm_stats.lock);
2652
2653 INIT_LIST_HEAD(&adev->shadow_list);
2654 mutex_init(&adev->shadow_list_lock);
2655
2656 INIT_LIST_HEAD(&adev->ring_lru_list);
2657 spin_lock_init(&adev->ring_lru_list_lock);
2658
2659 INIT_DELAYED_WORK(&adev->delayed_init_work,
2660 amdgpu_device_delayed_init_work_handler);
2661 INIT_DELAYED_WORK(&adev->gfx.gfx_off_delay_work,
2662 amdgpu_device_delay_enable_gfx_off);
2663
2664 INIT_WORK(&adev->xgmi_reset_work, amdgpu_device_xgmi_reset_func);
2665
2666 adev->gfx.gfx_off_req_count = 1;
2667 adev->pm.ac_power = power_supply_is_system_supplied() > 0 ? true : false;
2668
2669 /* Registers mapping */
2670 /* TODO: block userspace mapping of io register */
2671 if (adev->asic_type >= CHIP_BONAIRE) {
2672 adev->rmmio_base = pci_resource_start(adev->pdev, 5);
2673 adev->rmmio_size = pci_resource_len(adev->pdev, 5);
2674 } else {
2675 adev->rmmio_base = pci_resource_start(adev->pdev, 2);
2676 adev->rmmio_size = pci_resource_len(adev->pdev, 2);
2677 }
2678
2679 adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
2680 if (adev->rmmio == NULL) {
2681 return -ENOMEM;
2682 }
2683 DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
2684 DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
2685
2686 /* io port mapping */
2687 for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
2688 if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) {
2689 adev->rio_mem_size = pci_resource_len(adev->pdev, i);
2690 adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size);
2691 break;
2692 }
2693 }
2694 if (adev->rio_mem == NULL)
2695 DRM_INFO("PCI I/O BAR is not found.\n");
2696
2697 /* enable PCIE atomic ops */
2698 r = pci_enable_atomic_ops_to_root(adev->pdev,
2699 PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
2700 PCI_EXP_DEVCAP2_ATOMIC_COMP64);
2701 if (r) {
2702 adev->have_atomics_support = false;
2703 DRM_INFO("PCIE atomic ops is not supported\n");
2704 } else {
2705 adev->have_atomics_support = true;
2706 }
2707
2708 amdgpu_device_get_pcie_info(adev);
2709
2710 if (amdgpu_mcbp)
2711 DRM_INFO("MCBP is enabled\n");
2712
2713 if (amdgpu_mes && adev->asic_type >= CHIP_NAVI10)
2714 adev->enable_mes = true;
2715
2716 if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10) {
2717 r = amdgpu_discovery_init(adev);
2718 if (r) {
2719 dev_err(adev->dev, "amdgpu_discovery_init failed\n");
2720 return r;
2721 }
2722 }
2723
2724 /* early init functions */
2725 r = amdgpu_device_ip_early_init(adev);
2726 if (r)
2727 return r;
2728
2729 /* doorbell bar mapping and doorbell index init*/
2730 amdgpu_device_doorbell_init(adev);
2731
2732 /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
2733 /* this will fail for cards that aren't VGA class devices, just
2734 * ignore it */
2735 vga_client_register(adev->pdev, adev, NULL, amdgpu_device_vga_set_decode);
2736
2737 if (amdgpu_device_is_px(ddev))
2738 runtime = true;
2739 if (!pci_is_thunderbolt_attached(adev->pdev))
2740 vga_switcheroo_register_client(adev->pdev,
2741 &amdgpu_switcheroo_ops, runtime);
2742 if (runtime)
2743 vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
2744
2745 if (amdgpu_emu_mode == 1) {
2746 /* post the asic on emulation mode */
2747 emu_soc_asic_init(adev);
2748 goto fence_driver_init;
2749 }
2750
2751 /* detect if we are with an SRIOV vbios */
2752 amdgpu_device_detect_sriov_bios(adev);
2753
2754 /* check if we need to reset the asic
2755 * E.g., driver was not cleanly unloaded previously, etc.
2756 */
2757 if (!amdgpu_sriov_vf(adev) && amdgpu_asic_need_reset_on_init(adev)) {
2758 r = amdgpu_asic_reset(adev);
2759 if (r) {
2760 dev_err(adev->dev, "asic reset on init failed\n");
2761 goto failed;
2762 }
2763 }
2764
2765 /* Post card if necessary */
2766 if (amdgpu_device_need_post(adev)) {
2767 if (!adev->bios) {
2768 dev_err(adev->dev, "no vBIOS found\n");
2769 r = -EINVAL;
2770 goto failed;
2771 }
2772 DRM_INFO("GPU posting now...\n");
2773 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2774 if (r) {
2775 dev_err(adev->dev, "gpu post error!\n");
2776 goto failed;
2777 }
2778 }
2779
2780 if (adev->is_atom_fw) {
2781 /* Initialize clocks */
2782 r = amdgpu_atomfirmware_get_clock_info(adev);
2783 if (r) {
2784 dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
2785 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
2786 goto failed;
2787 }
2788 } else {
2789 /* Initialize clocks */
2790 r = amdgpu_atombios_get_clock_info(adev);
2791 if (r) {
2792 dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
2793 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
2794 goto failed;
2795 }
2796 /* init i2c buses */
2797 if (!amdgpu_device_has_dc_support(adev))
2798 amdgpu_atombios_i2c_init(adev);
2799 }
2800
2801fence_driver_init:
2802 /* Fence driver */
2803 r = amdgpu_fence_driver_init(adev);
2804 if (r) {
2805 dev_err(adev->dev, "amdgpu_fence_driver_init failed\n");
2806 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
2807 goto failed;
2808 }
2809
2810 /* init the mode config */
2811 drm_mode_config_init(adev->ddev);
2812
2813 r = amdgpu_device_ip_init(adev);
2814 if (r) {
2815 /* failed in exclusive mode due to timeout */
2816 if (amdgpu_sriov_vf(adev) &&
2817 !amdgpu_sriov_runtime(adev) &&
2818 amdgpu_virt_mmio_blocked(adev) &&
2819 !amdgpu_virt_wait_reset(adev)) {
2820 dev_err(adev->dev, "VF exclusive mode timeout\n");
2821 /* Don't send request since VF is inactive. */
2822 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
2823 adev->virt.ops = NULL;
2824 r = -EAGAIN;
2825 goto failed;
2826 }
2827 dev_err(adev->dev, "amdgpu_device_ip_init failed\n");
2828 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
2829 if (amdgpu_virt_request_full_gpu(adev, false))
2830 amdgpu_virt_release_full_gpu(adev, false);
2831 goto failed;
2832 }
2833
2834 adev->accel_working = true;
2835
2836 amdgpu_vm_check_compute_bug(adev);
2837
2838 /* Initialize the buffer migration limit. */
2839 if (amdgpu_moverate >= 0)
2840 max_MBps = amdgpu_moverate;
2841 else
2842 max_MBps = 8; /* Allow 8 MB/s. */
2843 /* Get a log2 for easy divisions. */
2844 adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
2845
2846 amdgpu_fbdev_init(adev);
2847
2848 if (amdgpu_sriov_vf(adev) && amdgim_is_hwperf(adev))
2849 amdgpu_pm_virt_sysfs_init(adev);
2850
2851 r = amdgpu_pm_sysfs_init(adev);
2852 if (r)
2853 DRM_ERROR("registering pm debugfs failed (%d).\n", r);
2854
2855 r = amdgpu_ucode_sysfs_init(adev);
2856 if (r)
2857 DRM_ERROR("Creating firmware sysfs failed (%d).\n", r);
2858
2859 r = amdgpu_debugfs_gem_init(adev);
2860 if (r)
2861 DRM_ERROR("registering gem debugfs failed (%d).\n", r);
2862
2863 r = amdgpu_debugfs_regs_init(adev);
2864 if (r)
2865 DRM_ERROR("registering register debugfs failed (%d).\n", r);
2866
2867 r = amdgpu_debugfs_firmware_init(adev);
2868 if (r)
2869 DRM_ERROR("registering firmware debugfs failed (%d).\n", r);
2870
2871 r = amdgpu_debugfs_init(adev);
2872 if (r)
2873 DRM_ERROR("Creating debugfs files failed (%d).\n", r);
2874
2875 if ((amdgpu_testing & 1)) {
2876 if (adev->accel_working)
2877 amdgpu_test_moves(adev);
2878 else
2879 DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n");
2880 }
2881 if (amdgpu_benchmarking) {
2882 if (adev->accel_working)
2883 amdgpu_benchmark(adev, amdgpu_benchmarking);
2884 else
2885 DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n");
2886 }
2887
2888 /*
2889 * Register gpu instance before amdgpu_device_enable_mgpu_fan_boost.
2890 * Otherwise the mgpu fan boost feature will be skipped due to the
2891 * gpu instance is counted less.
2892 */
2893 amdgpu_register_gpu_instance(adev);
2894
2895 /* enable clockgating, etc. after ib tests, etc. since some blocks require
2896 * explicit gating rather than handling it automatically.
2897 */
2898 r = amdgpu_device_ip_late_init(adev);
2899 if (r) {
2900 dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n");
2901 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
2902 goto failed;
2903 }
2904
2905 /* must succeed. */
2906 amdgpu_ras_resume(adev);
2907
2908 queue_delayed_work(system_wq, &adev->delayed_init_work,
2909 msecs_to_jiffies(AMDGPU_RESUME_MS));
2910
2911 r = device_create_file(adev->dev, &dev_attr_pcie_replay_count);
2912 if (r) {
2913 dev_err(adev->dev, "Could not create pcie_replay_count");
2914 return r;
2915 }
2916
2917 if (IS_ENABLED(CONFIG_PERF_EVENTS))
2918 r = amdgpu_pmu_init(adev);
2919 if (r)
2920 dev_err(adev->dev, "amdgpu_pmu_init failed\n");
2921
2922 return 0;
2923
2924failed:
2925 amdgpu_vf_error_trans_all(adev);
2926 if (runtime)
2927 vga_switcheroo_fini_domain_pm_ops(adev->dev);
2928
2929 return r;
2930}
2931
2932/**
2933 * amdgpu_device_fini - tear down the driver
2934 *
2935 * @adev: amdgpu_device pointer
2936 *
2937 * Tear down the driver info (all asics).
2938 * Called at driver shutdown.
2939 */
2940void amdgpu_device_fini(struct amdgpu_device *adev)
2941{
2942 int r;
2943
2944 DRM_INFO("amdgpu: finishing device.\n");
2945 adev->shutdown = true;
2946 /* disable all interrupts */
2947 amdgpu_irq_disable_all(adev);
2948 if (adev->mode_info.mode_config_initialized){
2949 if (!amdgpu_device_has_dc_support(adev))
2950 drm_helper_force_disable_all(adev->ddev);
2951 else
2952 drm_atomic_helper_shutdown(adev->ddev);
2953 }
2954 amdgpu_fence_driver_fini(adev);
2955 amdgpu_pm_sysfs_fini(adev);
2956 amdgpu_fbdev_fini(adev);
2957 r = amdgpu_device_ip_fini(adev);
2958 if (adev->firmware.gpu_info_fw) {
2959 release_firmware(adev->firmware.gpu_info_fw);
2960 adev->firmware.gpu_info_fw = NULL;
2961 }
2962 adev->accel_working = false;
2963 cancel_delayed_work_sync(&adev->delayed_init_work);
2964 /* free i2c buses */
2965 if (!amdgpu_device_has_dc_support(adev))
2966 amdgpu_i2c_fini(adev);
2967
2968 if (amdgpu_emu_mode != 1)
2969 amdgpu_atombios_fini(adev);
2970
2971 kfree(adev->bios);
2972 adev->bios = NULL;
2973 if (!pci_is_thunderbolt_attached(adev->pdev))
2974 vga_switcheroo_unregister_client(adev->pdev);
2975 if (adev->flags & AMD_IS_PX)
2976 vga_switcheroo_fini_domain_pm_ops(adev->dev);
2977 vga_client_register(adev->pdev, NULL, NULL, NULL);
2978 if (adev->rio_mem)
2979 pci_iounmap(adev->pdev, adev->rio_mem);
2980 adev->rio_mem = NULL;
2981 iounmap(adev->rmmio);
2982 adev->rmmio = NULL;
2983 amdgpu_device_doorbell_fini(adev);
2984 if (amdgpu_sriov_vf(adev) && amdgim_is_hwperf(adev))
2985 amdgpu_pm_virt_sysfs_fini(adev);
2986
2987 amdgpu_debugfs_regs_cleanup(adev);
2988 device_remove_file(adev->dev, &dev_attr_pcie_replay_count);
2989 amdgpu_ucode_sysfs_fini(adev);
2990 if (IS_ENABLED(CONFIG_PERF_EVENTS))
2991 amdgpu_pmu_fini(adev);
2992 amdgpu_debugfs_preempt_cleanup(adev);
2993 if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10)
2994 amdgpu_discovery_fini(adev);
2995}
2996
2997
2998/*
2999 * Suspend & resume.
3000 */
3001/**
3002 * amdgpu_device_suspend - initiate device suspend
3003 *
3004 * @dev: drm dev pointer
3005 * @suspend: suspend state
3006 * @fbcon : notify the fbdev of suspend
3007 *
3008 * Puts the hw in the suspend state (all asics).
3009 * Returns 0 for success or an error on failure.
3010 * Called at driver suspend.
3011 */
3012int amdgpu_device_suspend(struct drm_device *dev, bool suspend, bool fbcon)
3013{
3014 struct amdgpu_device *adev;
3015 struct drm_crtc *crtc;
3016 struct drm_connector *connector;
3017 int r;
3018
3019 if (dev == NULL || dev->dev_private == NULL) {
3020 return -ENODEV;
3021 }
3022
3023 adev = dev->dev_private;
3024
3025 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
3026 return 0;
3027
3028 adev->in_suspend = true;
3029 drm_kms_helper_poll_disable(dev);
3030
3031 if (fbcon)
3032 amdgpu_fbdev_set_suspend(adev, 1);
3033
3034 cancel_delayed_work_sync(&adev->delayed_init_work);
3035
3036 if (!amdgpu_device_has_dc_support(adev)) {
3037 /* turn off display hw */
3038 drm_modeset_lock_all(dev);
3039 list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
3040 drm_helper_connector_dpms(connector, DRM_MODE_DPMS_OFF);
3041 }
3042 drm_modeset_unlock_all(dev);
3043 /* unpin the front buffers and cursors */
3044 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
3045 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
3046 struct drm_framebuffer *fb = crtc->primary->fb;
3047 struct amdgpu_bo *robj;
3048
3049 if (amdgpu_crtc->cursor_bo && !adev->enable_virtual_display) {
3050 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
3051 r = amdgpu_bo_reserve(aobj, true);
3052 if (r == 0) {
3053 amdgpu_bo_unpin(aobj);
3054 amdgpu_bo_unreserve(aobj);
3055 }
3056 }
3057
3058 if (fb == NULL || fb->obj[0] == NULL) {
3059 continue;
3060 }
3061 robj = gem_to_amdgpu_bo(fb->obj[0]);
3062 /* don't unpin kernel fb objects */
3063 if (!amdgpu_fbdev_robj_is_fb(adev, robj)) {
3064 r = amdgpu_bo_reserve(robj, true);
3065 if (r == 0) {
3066 amdgpu_bo_unpin(robj);
3067 amdgpu_bo_unreserve(robj);
3068 }
3069 }
3070 }
3071 }
3072
3073 amdgpu_amdkfd_suspend(adev);
3074
3075 amdgpu_ras_suspend(adev);
3076
3077 r = amdgpu_device_ip_suspend_phase1(adev);
3078
3079 /* evict vram memory */
3080 amdgpu_bo_evict_vram(adev);
3081
3082 amdgpu_fence_driver_suspend(adev);
3083
3084 r = amdgpu_device_ip_suspend_phase2(adev);
3085
3086 /* evict remaining vram memory
3087 * This second call to evict vram is to evict the gart page table
3088 * using the CPU.
3089 */
3090 amdgpu_bo_evict_vram(adev);
3091
3092 pci_save_state(dev->pdev);
3093 if (suspend) {
3094 /* Shut down the device */
3095 pci_disable_device(dev->pdev);
3096 pci_set_power_state(dev->pdev, PCI_D3hot);
3097 } else {
3098 r = amdgpu_asic_reset(adev);
3099 if (r)
3100 DRM_ERROR("amdgpu asic reset failed\n");
3101 }
3102
3103 return 0;
3104}
3105
3106/**
3107 * amdgpu_device_resume - initiate device resume
3108 *
3109 * @dev: drm dev pointer
3110 * @resume: resume state
3111 * @fbcon : notify the fbdev of resume
3112 *
3113 * Bring the hw back to operating state (all asics).
3114 * Returns 0 for success or an error on failure.
3115 * Called at driver resume.
3116 */
3117int amdgpu_device_resume(struct drm_device *dev, bool resume, bool fbcon)
3118{
3119 struct drm_connector *connector;
3120 struct amdgpu_device *adev = dev->dev_private;
3121 struct drm_crtc *crtc;
3122 int r = 0;
3123
3124 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
3125 return 0;
3126
3127 if (resume) {
3128 pci_set_power_state(dev->pdev, PCI_D0);
3129 pci_restore_state(dev->pdev);
3130 r = pci_enable_device(dev->pdev);
3131 if (r)
3132 return r;
3133 }
3134
3135 /* post card */
3136 if (amdgpu_device_need_post(adev)) {
3137 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
3138 if (r)
3139 DRM_ERROR("amdgpu asic init failed\n");
3140 }
3141
3142 r = amdgpu_device_ip_resume(adev);
3143 if (r) {
3144 DRM_ERROR("amdgpu_device_ip_resume failed (%d).\n", r);
3145 return r;
3146 }
3147 amdgpu_fence_driver_resume(adev);
3148
3149
3150 r = amdgpu_device_ip_late_init(adev);
3151 if (r)
3152 return r;
3153
3154 queue_delayed_work(system_wq, &adev->delayed_init_work,
3155 msecs_to_jiffies(AMDGPU_RESUME_MS));
3156
3157 if (!amdgpu_device_has_dc_support(adev)) {
3158 /* pin cursors */
3159 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
3160 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
3161
3162 if (amdgpu_crtc->cursor_bo && !adev->enable_virtual_display) {
3163 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
3164 r = amdgpu_bo_reserve(aobj, true);
3165 if (r == 0) {
3166 r = amdgpu_bo_pin(aobj, AMDGPU_GEM_DOMAIN_VRAM);
3167 if (r != 0)
3168 DRM_ERROR("Failed to pin cursor BO (%d)\n", r);
3169 amdgpu_crtc->cursor_addr = amdgpu_bo_gpu_offset(aobj);
3170 amdgpu_bo_unreserve(aobj);
3171 }
3172 }
3173 }
3174 }
3175 r = amdgpu_amdkfd_resume(adev);
3176 if (r)
3177 return r;
3178
3179 /* Make sure IB tests flushed */
3180 flush_delayed_work(&adev->delayed_init_work);
3181
3182 /* blat the mode back in */
3183 if (fbcon) {
3184 if (!amdgpu_device_has_dc_support(adev)) {
3185 /* pre DCE11 */
3186 drm_helper_resume_force_mode(dev);
3187
3188 /* turn on display hw */
3189 drm_modeset_lock_all(dev);
3190 list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
3191 drm_helper_connector_dpms(connector, DRM_MODE_DPMS_ON);
3192 }
3193 drm_modeset_unlock_all(dev);
3194 }
3195 amdgpu_fbdev_set_suspend(adev, 0);
3196 }
3197
3198 drm_kms_helper_poll_enable(dev);
3199
3200 amdgpu_ras_resume(adev);
3201
3202 /*
3203 * Most of the connector probing functions try to acquire runtime pm
3204 * refs to ensure that the GPU is powered on when connector polling is
3205 * performed. Since we're calling this from a runtime PM callback,
3206 * trying to acquire rpm refs will cause us to deadlock.
3207 *
3208 * Since we're guaranteed to be holding the rpm lock, it's safe to
3209 * temporarily disable the rpm helpers so this doesn't deadlock us.
3210 */
3211#ifdef CONFIG_PM
3212 dev->dev->power.disable_depth++;
3213#endif
3214 if (!amdgpu_device_has_dc_support(adev))
3215 drm_helper_hpd_irq_event(dev);
3216 else
3217 drm_kms_helper_hotplug_event(dev);
3218#ifdef CONFIG_PM
3219 dev->dev->power.disable_depth--;
3220#endif
3221 adev->in_suspend = false;
3222
3223 return 0;
3224}
3225
3226/**
3227 * amdgpu_device_ip_check_soft_reset - did soft reset succeed
3228 *
3229 * @adev: amdgpu_device pointer
3230 *
3231 * The list of all the hardware IPs that make up the asic is walked and
3232 * the check_soft_reset callbacks are run. check_soft_reset determines
3233 * if the asic is still hung or not.
3234 * Returns true if any of the IPs are still in a hung state, false if not.
3235 */
3236static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)
3237{
3238 int i;
3239 bool asic_hang = false;
3240
3241 if (amdgpu_sriov_vf(adev))
3242 return true;
3243
3244 if (amdgpu_asic_need_full_reset(adev))
3245 return true;
3246
3247 for (i = 0; i < adev->num_ip_blocks; i++) {
3248 if (!adev->ip_blocks[i].status.valid)
3249 continue;
3250 if (adev->ip_blocks[i].version->funcs->check_soft_reset)
3251 adev->ip_blocks[i].status.hang =
3252 adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
3253 if (adev->ip_blocks[i].status.hang) {
3254 DRM_INFO("IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
3255 asic_hang = true;
3256 }
3257 }
3258 return asic_hang;
3259}
3260
3261/**
3262 * amdgpu_device_ip_pre_soft_reset - prepare for soft reset
3263 *
3264 * @adev: amdgpu_device pointer
3265 *
3266 * The list of all the hardware IPs that make up the asic is walked and the
3267 * pre_soft_reset callbacks are run if the block is hung. pre_soft_reset
3268 * handles any IP specific hardware or software state changes that are
3269 * necessary for a soft reset to succeed.
3270 * Returns 0 on success, negative error code on failure.
3271 */
3272static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)
3273{
3274 int i, r = 0;
3275
3276 for (i = 0; i < adev->num_ip_blocks; i++) {
3277 if (!adev->ip_blocks[i].status.valid)
3278 continue;
3279 if (adev->ip_blocks[i].status.hang &&
3280 adev->ip_blocks[i].version->funcs->pre_soft_reset) {
3281 r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
3282 if (r)
3283 return r;
3284 }
3285 }
3286
3287 return 0;
3288}
3289
3290/**
3291 * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed
3292 *
3293 * @adev: amdgpu_device pointer
3294 *
3295 * Some hardware IPs cannot be soft reset. If they are hung, a full gpu
3296 * reset is necessary to recover.
3297 * Returns true if a full asic reset is required, false if not.
3298 */
3299static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)
3300{
3301 int i;
3302
3303 if (amdgpu_asic_need_full_reset(adev))
3304 return true;
3305
3306 for (i = 0; i < adev->num_ip_blocks; i++) {
3307 if (!adev->ip_blocks[i].status.valid)
3308 continue;
3309 if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
3310 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
3311 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
3312 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
3313 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
3314 if (adev->ip_blocks[i].status.hang) {
3315 DRM_INFO("Some block need full reset!\n");
3316 return true;
3317 }
3318 }
3319 }
3320 return false;
3321}
3322
3323/**
3324 * amdgpu_device_ip_soft_reset - do a soft reset
3325 *
3326 * @adev: amdgpu_device pointer
3327 *
3328 * The list of all the hardware IPs that make up the asic is walked and the
3329 * soft_reset callbacks are run if the block is hung. soft_reset handles any
3330 * IP specific hardware or software state changes that are necessary to soft
3331 * reset the IP.
3332 * Returns 0 on success, negative error code on failure.
3333 */
3334static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)
3335{
3336 int i, r = 0;
3337
3338 for (i = 0; i < adev->num_ip_blocks; i++) {
3339 if (!adev->ip_blocks[i].status.valid)
3340 continue;
3341 if (adev->ip_blocks[i].status.hang &&
3342 adev->ip_blocks[i].version->funcs->soft_reset) {
3343 r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
3344 if (r)
3345 return r;
3346 }
3347 }
3348
3349 return 0;
3350}
3351
3352/**
3353 * amdgpu_device_ip_post_soft_reset - clean up from soft reset
3354 *
3355 * @adev: amdgpu_device pointer
3356 *
3357 * The list of all the hardware IPs that make up the asic is walked and the
3358 * post_soft_reset callbacks are run if the asic was hung. post_soft_reset
3359 * handles any IP specific hardware or software state changes that are
3360 * necessary after the IP has been soft reset.
3361 * Returns 0 on success, negative error code on failure.
3362 */
3363static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)
3364{
3365 int i, r = 0;
3366
3367 for (i = 0; i < adev->num_ip_blocks; i++) {
3368 if (!adev->ip_blocks[i].status.valid)
3369 continue;
3370 if (adev->ip_blocks[i].status.hang &&
3371 adev->ip_blocks[i].version->funcs->post_soft_reset)
3372 r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
3373 if (r)
3374 return r;
3375 }
3376
3377 return 0;
3378}
3379
3380/**
3381 * amdgpu_device_recover_vram - Recover some VRAM contents
3382 *
3383 * @adev: amdgpu_device pointer
3384 *
3385 * Restores the contents of VRAM buffers from the shadows in GTT. Used to
3386 * restore things like GPUVM page tables after a GPU reset where
3387 * the contents of VRAM might be lost.
3388 *
3389 * Returns:
3390 * 0 on success, negative error code on failure.
3391 */
3392static int amdgpu_device_recover_vram(struct amdgpu_device *adev)
3393{
3394 struct dma_fence *fence = NULL, *next = NULL;
3395 struct amdgpu_bo *shadow;
3396 long r = 1, tmo;
3397
3398 if (amdgpu_sriov_runtime(adev))
3399 tmo = msecs_to_jiffies(8000);
3400 else
3401 tmo = msecs_to_jiffies(100);
3402
3403 DRM_INFO("recover vram bo from shadow start\n");
3404 mutex_lock(&adev->shadow_list_lock);
3405 list_for_each_entry(shadow, &adev->shadow_list, shadow_list) {
3406
3407 /* No need to recover an evicted BO */
3408 if (shadow->tbo.mem.mem_type != TTM_PL_TT ||
3409 shadow->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET ||
3410 shadow->parent->tbo.mem.mem_type != TTM_PL_VRAM)
3411 continue;
3412
3413 r = amdgpu_bo_restore_shadow(shadow, &next);
3414 if (r)
3415 break;
3416
3417 if (fence) {
3418 tmo = dma_fence_wait_timeout(fence, false, tmo);
3419 dma_fence_put(fence);
3420 fence = next;
3421 if (tmo == 0) {
3422 r = -ETIMEDOUT;
3423 break;
3424 } else if (tmo < 0) {
3425 r = tmo;
3426 break;
3427 }
3428 } else {
3429 fence = next;
3430 }
3431 }
3432 mutex_unlock(&adev->shadow_list_lock);
3433
3434 if (fence)
3435 tmo = dma_fence_wait_timeout(fence, false, tmo);
3436 dma_fence_put(fence);
3437
3438 if (r < 0 || tmo <= 0) {
3439 DRM_ERROR("recover vram bo from shadow failed, r is %ld, tmo is %ld\n", r, tmo);
3440 return -EIO;
3441 }
3442
3443 DRM_INFO("recover vram bo from shadow done\n");
3444 return 0;
3445}
3446
3447
3448/**
3449 * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
3450 *
3451 * @adev: amdgpu device pointer
3452 * @from_hypervisor: request from hypervisor
3453 *
3454 * do VF FLR and reinitialize Asic
3455 * return 0 means succeeded otherwise failed
3456 */
3457static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
3458 bool from_hypervisor)
3459{
3460 int r;
3461
3462 if (from_hypervisor)
3463 r = amdgpu_virt_request_full_gpu(adev, true);
3464 else
3465 r = amdgpu_virt_reset_gpu(adev);
3466 if (r)
3467 return r;
3468
3469 amdgpu_amdkfd_pre_reset(adev);
3470
3471 /* Resume IP prior to SMC */
3472 r = amdgpu_device_ip_reinit_early_sriov(adev);
3473 if (r)
3474 goto error;
3475
3476 /* we need recover gart prior to run SMC/CP/SDMA resume */
3477 amdgpu_gtt_mgr_recover(&adev->mman.bdev.man[TTM_PL_TT]);
3478
3479 r = amdgpu_device_fw_loading(adev);
3480 if (r)
3481 return r;
3482
3483 /* now we are okay to resume SMC/CP/SDMA */
3484 r = amdgpu_device_ip_reinit_late_sriov(adev);
3485 if (r)
3486 goto error;
3487
3488 amdgpu_irq_gpu_reset_resume_helper(adev);
3489 r = amdgpu_ib_ring_tests(adev);
3490 amdgpu_amdkfd_post_reset(adev);
3491
3492error:
3493 amdgpu_virt_init_data_exchange(adev);
3494 amdgpu_virt_release_full_gpu(adev, true);
3495 if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
3496 amdgpu_inc_vram_lost(adev);
3497 r = amdgpu_device_recover_vram(adev);
3498 }
3499
3500 return r;
3501}
3502
3503/**
3504 * amdgpu_device_should_recover_gpu - check if we should try GPU recovery
3505 *
3506 * @adev: amdgpu device pointer
3507 *
3508 * Check amdgpu_gpu_recovery and SRIOV status to see if we should try to recover
3509 * a hung GPU.
3510 */
3511bool amdgpu_device_should_recover_gpu(struct amdgpu_device *adev)
3512{
3513 if (!amdgpu_device_ip_check_soft_reset(adev)) {
3514 DRM_INFO("Timeout, but no hardware hang detected.\n");
3515 return false;
3516 }
3517
3518 if (amdgpu_gpu_recovery == 0)
3519 goto disabled;
3520
3521 if (amdgpu_sriov_vf(adev))
3522 return true;
3523
3524 if (amdgpu_gpu_recovery == -1) {
3525 switch (adev->asic_type) {
3526 case CHIP_BONAIRE:
3527 case CHIP_HAWAII:
3528 case CHIP_TOPAZ:
3529 case CHIP_TONGA:
3530 case CHIP_FIJI:
3531 case CHIP_POLARIS10:
3532 case CHIP_POLARIS11:
3533 case CHIP_POLARIS12:
3534 case CHIP_VEGAM:
3535 case CHIP_VEGA20:
3536 case CHIP_VEGA10:
3537 case CHIP_VEGA12:
3538 case CHIP_RAVEN:
3539 break;
3540 default:
3541 goto disabled;
3542 }
3543 }
3544
3545 return true;
3546
3547disabled:
3548 DRM_INFO("GPU recovery disabled.\n");
3549 return false;
3550}
3551
3552
3553static int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,
3554 struct amdgpu_job *job,
3555 bool *need_full_reset_arg)
3556{
3557 int i, r = 0;
3558 bool need_full_reset = *need_full_reset_arg;
3559
3560 /* block all schedulers and reset given job's ring */
3561 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3562 struct amdgpu_ring *ring = adev->rings[i];
3563
3564 if (!ring || !ring->sched.thread)
3565 continue;
3566
3567 /* after all hw jobs are reset, hw fence is meaningless, so force_completion */
3568 amdgpu_fence_driver_force_completion(ring);
3569 }
3570
3571 if(job)
3572 drm_sched_increase_karma(&job->base);
3573
3574 /* Don't suspend on bare metal if we are not going to HW reset the ASIC */
3575 if (!amdgpu_sriov_vf(adev)) {
3576
3577 if (!need_full_reset)
3578 need_full_reset = amdgpu_device_ip_need_full_reset(adev);
3579
3580 if (!need_full_reset) {
3581 amdgpu_device_ip_pre_soft_reset(adev);
3582 r = amdgpu_device_ip_soft_reset(adev);
3583 amdgpu_device_ip_post_soft_reset(adev);
3584 if (r || amdgpu_device_ip_check_soft_reset(adev)) {
3585 DRM_INFO("soft reset failed, will fallback to full reset!\n");
3586 need_full_reset = true;
3587 }
3588 }
3589
3590 if (need_full_reset)
3591 r = amdgpu_device_ip_suspend(adev);
3592
3593 *need_full_reset_arg = need_full_reset;
3594 }
3595
3596 return r;
3597}
3598
3599static int amdgpu_do_asic_reset(struct amdgpu_hive_info *hive,
3600 struct list_head *device_list_handle,
3601 bool *need_full_reset_arg)
3602{
3603 struct amdgpu_device *tmp_adev = NULL;
3604 bool need_full_reset = *need_full_reset_arg, vram_lost = false;
3605 int r = 0;
3606
3607 /*
3608 * ASIC reset has to be done on all HGMI hive nodes ASAP
3609 * to allow proper links negotiation in FW (within 1 sec)
3610 */
3611 if (need_full_reset) {
3612 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
3613 /* For XGMI run all resets in parallel to speed up the process */
3614 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
3615 if (!queue_work(system_highpri_wq, &tmp_adev->xgmi_reset_work))
3616 r = -EALREADY;
3617 } else
3618 r = amdgpu_asic_reset(tmp_adev);
3619
3620 if (r) {
3621 DRM_ERROR("ASIC reset failed with error, %d for drm dev, %s",
3622 r, tmp_adev->ddev->unique);
3623 break;
3624 }
3625 }
3626
3627 /* For XGMI wait for all PSP resets to complete before proceed */
3628 if (!r) {
3629 list_for_each_entry(tmp_adev, device_list_handle,
3630 gmc.xgmi.head) {
3631 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
3632 flush_work(&tmp_adev->xgmi_reset_work);
3633 r = tmp_adev->asic_reset_res;
3634 if (r)
3635 break;
3636 }
3637 }
3638
3639 list_for_each_entry(tmp_adev, device_list_handle,
3640 gmc.xgmi.head) {
3641 amdgpu_ras_reserve_bad_pages(tmp_adev);
3642 }
3643 }
3644 }
3645
3646
3647 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
3648 if (need_full_reset) {
3649 /* post card */
3650 if (amdgpu_atom_asic_init(tmp_adev->mode_info.atom_context))
3651 DRM_WARN("asic atom init failed!");
3652
3653 if (!r) {
3654 dev_info(tmp_adev->dev, "GPU reset succeeded, trying to resume\n");
3655 r = amdgpu_device_ip_resume_phase1(tmp_adev);
3656 if (r)
3657 goto out;
3658
3659 vram_lost = amdgpu_device_check_vram_lost(tmp_adev);
3660 if (vram_lost) {
3661 DRM_INFO("VRAM is lost due to GPU reset!\n");
3662 amdgpu_inc_vram_lost(tmp_adev);
3663 }
3664
3665 r = amdgpu_gtt_mgr_recover(
3666 &tmp_adev->mman.bdev.man[TTM_PL_TT]);
3667 if (r)
3668 goto out;
3669
3670 r = amdgpu_device_fw_loading(tmp_adev);
3671 if (r)
3672 return r;
3673
3674 r = amdgpu_device_ip_resume_phase2(tmp_adev);
3675 if (r)
3676 goto out;
3677
3678 if (vram_lost)
3679 amdgpu_device_fill_reset_magic(tmp_adev);
3680
3681 /*
3682 * Add this ASIC as tracked as reset was already
3683 * complete successfully.
3684 */
3685 amdgpu_register_gpu_instance(tmp_adev);
3686
3687 r = amdgpu_device_ip_late_init(tmp_adev);
3688 if (r)
3689 goto out;
3690
3691 /* must succeed. */
3692 amdgpu_ras_resume(tmp_adev);
3693
3694 /* Update PSP FW topology after reset */
3695 if (hive && tmp_adev->gmc.xgmi.num_physical_nodes > 1)
3696 r = amdgpu_xgmi_update_topology(hive, tmp_adev);
3697 }
3698 }
3699
3700
3701out:
3702 if (!r) {
3703 amdgpu_irq_gpu_reset_resume_helper(tmp_adev);
3704 r = amdgpu_ib_ring_tests(tmp_adev);
3705 if (r) {
3706 dev_err(tmp_adev->dev, "ib ring test failed (%d).\n", r);
3707 r = amdgpu_device_ip_suspend(tmp_adev);
3708 need_full_reset = true;
3709 r = -EAGAIN;
3710 goto end;
3711 }
3712 }
3713
3714 if (!r)
3715 r = amdgpu_device_recover_vram(tmp_adev);
3716 else
3717 tmp_adev->asic_reset_res = r;
3718 }
3719
3720end:
3721 *need_full_reset_arg = need_full_reset;
3722 return r;
3723}
3724
3725static bool amdgpu_device_lock_adev(struct amdgpu_device *adev, bool trylock)
3726{
3727 if (trylock) {
3728 if (!mutex_trylock(&adev->lock_reset))
3729 return false;
3730 } else
3731 mutex_lock(&adev->lock_reset);
3732
3733 atomic_inc(&adev->gpu_reset_counter);
3734 adev->in_gpu_reset = 1;
3735 switch (amdgpu_asic_reset_method(adev)) {
3736 case AMD_RESET_METHOD_MODE1:
3737 adev->mp1_state = PP_MP1_STATE_SHUTDOWN;
3738 break;
3739 case AMD_RESET_METHOD_MODE2:
3740 adev->mp1_state = PP_MP1_STATE_RESET;
3741 break;
3742 default:
3743 adev->mp1_state = PP_MP1_STATE_NONE;
3744 break;
3745 }
3746 /* Block kfd: SRIOV would do it separately */
3747 if (!amdgpu_sriov_vf(adev))
3748 amdgpu_amdkfd_pre_reset(adev);
3749
3750 return true;
3751}
3752
3753static void amdgpu_device_unlock_adev(struct amdgpu_device *adev)
3754{
3755 /*unlock kfd: SRIOV would do it separately */
3756 if (!amdgpu_sriov_vf(adev))
3757 amdgpu_amdkfd_post_reset(adev);
3758 amdgpu_vf_error_trans_all(adev);
3759 adev->mp1_state = PP_MP1_STATE_NONE;
3760 adev->in_gpu_reset = 0;
3761 mutex_unlock(&adev->lock_reset);
3762}
3763
3764
3765/**
3766 * amdgpu_device_gpu_recover - reset the asic and recover scheduler
3767 *
3768 * @adev: amdgpu device pointer
3769 * @job: which job trigger hang
3770 *
3771 * Attempt to reset the GPU if it has hung (all asics).
3772 * Attempt to do soft-reset or full-reset and reinitialize Asic
3773 * Returns 0 for success or an error on failure.
3774 */
3775
3776int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
3777 struct amdgpu_job *job)
3778{
3779 struct list_head device_list, *device_list_handle = NULL;
3780 bool need_full_reset, job_signaled;
3781 struct amdgpu_hive_info *hive = NULL;
3782 struct amdgpu_device *tmp_adev = NULL;
3783 int i, r = 0;
3784
3785 need_full_reset = job_signaled = false;
3786 INIT_LIST_HEAD(&device_list);
3787
3788 dev_info(adev->dev, "GPU reset begin!\n");
3789
3790 cancel_delayed_work_sync(&adev->delayed_init_work);
3791
3792 hive = amdgpu_get_xgmi_hive(adev, false);
3793
3794 /*
3795 * Here we trylock to avoid chain of resets executing from
3796 * either trigger by jobs on different adevs in XGMI hive or jobs on
3797 * different schedulers for same device while this TO handler is running.
3798 * We always reset all schedulers for device and all devices for XGMI
3799 * hive so that should take care of them too.
3800 */
3801
3802 if (hive && !mutex_trylock(&hive->reset_lock)) {
3803 DRM_INFO("Bailing on TDR for s_job:%llx, hive: %llx as another already in progress",
3804 job ? job->base.id : -1, hive->hive_id);
3805 return 0;
3806 }
3807
3808 /* Start with adev pre asic reset first for soft reset check.*/
3809 if (!amdgpu_device_lock_adev(adev, !hive)) {
3810 DRM_INFO("Bailing on TDR for s_job:%llx, as another already in progress",
3811 job ? job->base.id : -1);
3812 return 0;
3813 }
3814
3815 /* Build list of devices to reset */
3816 if (adev->gmc.xgmi.num_physical_nodes > 1) {
3817 if (!hive) {
3818 amdgpu_device_unlock_adev(adev);
3819 return -ENODEV;
3820 }
3821
3822 /*
3823 * In case we are in XGMI hive mode device reset is done for all the
3824 * nodes in the hive to retrain all XGMI links and hence the reset
3825 * sequence is executed in loop on all nodes.
3826 */
3827 device_list_handle = &hive->device_list;
3828 } else {
3829 list_add_tail(&adev->gmc.xgmi.head, &device_list);
3830 device_list_handle = &device_list;
3831 }
3832
3833 /*
3834 * Mark these ASICs to be reseted as untracked first
3835 * And add them back after reset completed
3836 */
3837 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head)
3838 amdgpu_unregister_gpu_instance(tmp_adev);
3839
3840 /* block all schedulers and reset given job's ring */
3841 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
3842 /* disable ras on ALL IPs */
3843 if (amdgpu_device_ip_need_full_reset(tmp_adev))
3844 amdgpu_ras_suspend(tmp_adev);
3845
3846 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3847 struct amdgpu_ring *ring = tmp_adev->rings[i];
3848
3849 if (!ring || !ring->sched.thread)
3850 continue;
3851
3852 drm_sched_stop(&ring->sched, job ? &job->base : NULL);
3853 }
3854 }
3855
3856
3857 /*
3858 * Must check guilty signal here since after this point all old
3859 * HW fences are force signaled.
3860 *
3861 * job->base holds a reference to parent fence
3862 */
3863 if (job && job->base.s_fence->parent &&
3864 dma_fence_is_signaled(job->base.s_fence->parent))
3865 job_signaled = true;
3866
3867 if (!amdgpu_device_ip_need_full_reset(adev))
3868 device_list_handle = &device_list;
3869
3870 if (job_signaled) {
3871 dev_info(adev->dev, "Guilty job already signaled, skipping HW reset");
3872 goto skip_hw_reset;
3873 }
3874
3875
3876 /* Guilty job will be freed after this*/
3877 r = amdgpu_device_pre_asic_reset(adev, job, &need_full_reset);
3878 if (r) {
3879 /*TODO Should we stop ?*/
3880 DRM_ERROR("GPU pre asic reset failed with err, %d for drm dev, %s ",
3881 r, adev->ddev->unique);
3882 adev->asic_reset_res = r;
3883 }
3884
3885retry: /* Rest of adevs pre asic reset from XGMI hive. */
3886 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
3887
3888 if (tmp_adev == adev)
3889 continue;
3890
3891 amdgpu_device_lock_adev(tmp_adev, false);
3892 r = amdgpu_device_pre_asic_reset(tmp_adev,
3893 NULL,
3894 &need_full_reset);
3895 /*TODO Should we stop ?*/
3896 if (r) {
3897 DRM_ERROR("GPU pre asic reset failed with err, %d for drm dev, %s ",
3898 r, tmp_adev->ddev->unique);
3899 tmp_adev->asic_reset_res = r;
3900 }
3901 }
3902
3903 /* Actual ASIC resets if needed.*/
3904 /* TODO Implement XGMI hive reset logic for SRIOV */
3905 if (amdgpu_sriov_vf(adev)) {
3906 r = amdgpu_device_reset_sriov(adev, job ? false : true);
3907 if (r)
3908 adev->asic_reset_res = r;
3909 } else {
3910 r = amdgpu_do_asic_reset(hive, device_list_handle, &need_full_reset);
3911 if (r && r == -EAGAIN)
3912 goto retry;
3913 }
3914
3915skip_hw_reset:
3916
3917 /* Post ASIC reset for all devs .*/
3918 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
3919 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3920 struct amdgpu_ring *ring = tmp_adev->rings[i];
3921
3922 if (!ring || !ring->sched.thread)
3923 continue;
3924
3925 /* No point to resubmit jobs if we didn't HW reset*/
3926 if (!tmp_adev->asic_reset_res && !job_signaled)
3927 drm_sched_resubmit_jobs(&ring->sched);
3928
3929 drm_sched_start(&ring->sched, !tmp_adev->asic_reset_res);
3930 }
3931
3932 if (!amdgpu_device_has_dc_support(tmp_adev) && !job_signaled) {
3933 drm_helper_resume_force_mode(tmp_adev->ddev);
3934 }
3935
3936 tmp_adev->asic_reset_res = 0;
3937
3938 if (r) {
3939 /* bad news, how to tell it to userspace ? */
3940 dev_info(tmp_adev->dev, "GPU reset(%d) failed\n", atomic_read(&adev->gpu_reset_counter));
3941 amdgpu_vf_error_put(tmp_adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r);
3942 } else {
3943 dev_info(tmp_adev->dev, "GPU reset(%d) succeeded!\n", atomic_read(&adev->gpu_reset_counter));
3944 }
3945
3946 amdgpu_device_unlock_adev(tmp_adev);
3947 }
3948
3949 if (hive)
3950 mutex_unlock(&hive->reset_lock);
3951
3952 if (r)
3953 dev_info(adev->dev, "GPU reset end with ret = %d\n", r);
3954 return r;
3955}
3956
3957/**
3958 * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot
3959 *
3960 * @adev: amdgpu_device pointer
3961 *
3962 * Fetchs and stores in the driver the PCIE capabilities (gen speed
3963 * and lanes) of the slot the device is in. Handles APUs and
3964 * virtualized environments where PCIE config space may not be available.
3965 */
3966static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev)
3967{
3968 struct pci_dev *pdev;
3969 enum pci_bus_speed speed_cap, platform_speed_cap;
3970 enum pcie_link_width platform_link_width;
3971
3972 if (amdgpu_pcie_gen_cap)
3973 adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
3974
3975 if (amdgpu_pcie_lane_cap)
3976 adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
3977
3978 /* covers APUs as well */
3979 if (pci_is_root_bus(adev->pdev->bus)) {
3980 if (adev->pm.pcie_gen_mask == 0)
3981 adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
3982 if (adev->pm.pcie_mlw_mask == 0)
3983 adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
3984 return;
3985 }
3986
3987 if (adev->pm.pcie_gen_mask && adev->pm.pcie_mlw_mask)
3988 return;
3989
3990 pcie_bandwidth_available(adev->pdev, NULL,
3991 &platform_speed_cap, &platform_link_width);
3992
3993 if (adev->pm.pcie_gen_mask == 0) {
3994 /* asic caps */
3995 pdev = adev->pdev;
3996 speed_cap = pcie_get_speed_cap(pdev);
3997 if (speed_cap == PCI_SPEED_UNKNOWN) {
3998 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3999 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4000 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
4001 } else {
4002 if (speed_cap == PCIE_SPEED_16_0GT)
4003 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4004 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4005 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
4006 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4);
4007 else if (speed_cap == PCIE_SPEED_8_0GT)
4008 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4009 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4010 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
4011 else if (speed_cap == PCIE_SPEED_5_0GT)
4012 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4013 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2);
4014 else
4015 adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1;
4016 }
4017 /* platform caps */
4018 if (platform_speed_cap == PCI_SPEED_UNKNOWN) {
4019 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4020 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
4021 } else {
4022 if (platform_speed_cap == PCIE_SPEED_16_0GT)
4023 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4024 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4025 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
4026 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4);
4027 else if (platform_speed_cap == PCIE_SPEED_8_0GT)
4028 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4029 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4030 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3);
4031 else if (platform_speed_cap == PCIE_SPEED_5_0GT)
4032 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4033 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
4034 else
4035 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
4036
4037 }
4038 }
4039 if (adev->pm.pcie_mlw_mask == 0) {
4040 if (platform_link_width == PCIE_LNK_WIDTH_UNKNOWN) {
4041 adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK;
4042 } else {
4043 switch (platform_link_width) {
4044 case PCIE_LNK_X32:
4045 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
4046 CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
4047 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
4048 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4049 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4050 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4051 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4052 break;
4053 case PCIE_LNK_X16:
4054 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
4055 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
4056 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4057 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4058 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4059 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4060 break;
4061 case PCIE_LNK_X12:
4062 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
4063 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4064 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4065 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4066 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4067 break;
4068 case PCIE_LNK_X8:
4069 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4070 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4071 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4072 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4073 break;
4074 case PCIE_LNK_X4:
4075 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4076 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4077 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4078 break;
4079 case PCIE_LNK_X2:
4080 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4081 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4082 break;
4083 case PCIE_LNK_X1:
4084 adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;
4085 break;
4086 default:
4087 break;
4088 }
4089 }
4090 }
4091}
4092