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 | |
77 | #include <linux/suspend.h> |
78 | #include <drm/task_barrier.h> |
79 | #include <linux/pm_runtime.h> |
80 | |
81 | #include <drm/drm_drv.h> |
82 | |
83 | #if IS_ENABLED(CONFIG_X86) |
84 | #include <asm/intel-family.h> |
85 | #endif |
86 | |
87 | MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin" ); |
88 | MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin" ); |
89 | MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin" ); |
90 | MODULE_FIRMWARE("amdgpu/picasso_gpu_info.bin" ); |
91 | MODULE_FIRMWARE("amdgpu/raven2_gpu_info.bin" ); |
92 | MODULE_FIRMWARE("amdgpu/arcturus_gpu_info.bin" ); |
93 | MODULE_FIRMWARE("amdgpu/navi12_gpu_info.bin" ); |
94 | |
95 | #define AMDGPU_RESUME_MS 2000 |
96 | #define AMDGPU_MAX_RETRY_LIMIT 2 |
97 | #define AMDGPU_RETRY_SRIOV_RESET(r) ((r) == -EBUSY || (r) == -ETIMEDOUT || (r) == -EINVAL) |
98 | |
99 | static const struct drm_driver amdgpu_kms_driver; |
100 | |
101 | const char *amdgpu_asic_name[] = { |
102 | "TAHITI" , |
103 | "PITCAIRN" , |
104 | "VERDE" , |
105 | "OLAND" , |
106 | "HAINAN" , |
107 | "BONAIRE" , |
108 | "KAVERI" , |
109 | "KABINI" , |
110 | "HAWAII" , |
111 | "MULLINS" , |
112 | "TOPAZ" , |
113 | "TONGA" , |
114 | "FIJI" , |
115 | "CARRIZO" , |
116 | "STONEY" , |
117 | "POLARIS10" , |
118 | "POLARIS11" , |
119 | "POLARIS12" , |
120 | "VEGAM" , |
121 | "VEGA10" , |
122 | "VEGA12" , |
123 | "VEGA20" , |
124 | "RAVEN" , |
125 | "ARCTURUS" , |
126 | "RENOIR" , |
127 | "ALDEBARAN" , |
128 | "NAVI10" , |
129 | "CYAN_SKILLFISH" , |
130 | "NAVI14" , |
131 | "NAVI12" , |
132 | "SIENNA_CICHLID" , |
133 | "NAVY_FLOUNDER" , |
134 | "VANGOGH" , |
135 | "DIMGREY_CAVEFISH" , |
136 | "BEIGE_GOBY" , |
137 | "YELLOW_CARP" , |
138 | "IP DISCOVERY" , |
139 | "LAST" , |
140 | }; |
141 | |
142 | /** |
143 | * DOC: pcie_replay_count |
144 | * |
145 | * The amdgpu driver provides a sysfs API for reporting the total number |
146 | * of PCIe replays (NAKs) |
147 | * The file pcie_replay_count is used for this and returns the total |
148 | * number of replays as a sum of the NAKs generated and NAKs received |
149 | */ |
150 | |
151 | static ssize_t amdgpu_device_get_pcie_replay_count(struct device *dev, |
152 | struct device_attribute *attr, char *buf) |
153 | { |
154 | struct drm_device *ddev = dev_get_drvdata(dev); |
155 | struct amdgpu_device *adev = drm_to_adev(ddev); |
156 | uint64_t cnt = amdgpu_asic_get_pcie_replay_count(adev); |
157 | |
158 | return sysfs_emit(buf, fmt: "%llu\n" , cnt); |
159 | } |
160 | |
161 | static DEVICE_ATTR(pcie_replay_count, 0444, |
162 | amdgpu_device_get_pcie_replay_count, NULL); |
163 | |
164 | /** |
165 | * DOC: board_info |
166 | * |
167 | * The amdgpu driver provides a sysfs API for giving board related information. |
168 | * It provides the form factor information in the format |
169 | * |
170 | * type : form factor |
171 | * |
172 | * Possible form factor values |
173 | * |
174 | * - "cem" - PCIE CEM card |
175 | * - "oam" - Open Compute Accelerator Module |
176 | * - "unknown" - Not known |
177 | * |
178 | */ |
179 | |
180 | static ssize_t amdgpu_device_get_board_info(struct device *dev, |
181 | struct device_attribute *attr, |
182 | char *buf) |
183 | { |
184 | struct drm_device *ddev = dev_get_drvdata(dev); |
185 | struct amdgpu_device *adev = drm_to_adev(ddev); |
186 | enum amdgpu_pkg_type pkg_type = AMDGPU_PKG_TYPE_CEM; |
187 | const char *pkg; |
188 | |
189 | if (adev->smuio.funcs && adev->smuio.funcs->get_pkg_type) |
190 | pkg_type = adev->smuio.funcs->get_pkg_type(adev); |
191 | |
192 | switch (pkg_type) { |
193 | case AMDGPU_PKG_TYPE_CEM: |
194 | pkg = "cem" ; |
195 | break; |
196 | case AMDGPU_PKG_TYPE_OAM: |
197 | pkg = "oam" ; |
198 | break; |
199 | default: |
200 | pkg = "unknown" ; |
201 | break; |
202 | } |
203 | |
204 | return sysfs_emit(buf, fmt: "%s : %s\n" , "type" , pkg); |
205 | } |
206 | |
207 | static DEVICE_ATTR(board_info, 0444, amdgpu_device_get_board_info, NULL); |
208 | |
209 | static struct attribute *amdgpu_board_attrs[] = { |
210 | &dev_attr_board_info.attr, |
211 | NULL, |
212 | }; |
213 | |
214 | static umode_t amdgpu_board_attrs_is_visible(struct kobject *kobj, |
215 | struct attribute *attr, int n) |
216 | { |
217 | struct device *dev = kobj_to_dev(kobj); |
218 | struct drm_device *ddev = dev_get_drvdata(dev); |
219 | struct amdgpu_device *adev = drm_to_adev(ddev); |
220 | |
221 | if (adev->flags & AMD_IS_APU) |
222 | return 0; |
223 | |
224 | return attr->mode; |
225 | } |
226 | |
227 | static const struct attribute_group amdgpu_board_attrs_group = { |
228 | .attrs = amdgpu_board_attrs, |
229 | .is_visible = amdgpu_board_attrs_is_visible |
230 | }; |
231 | |
232 | static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev); |
233 | |
234 | |
235 | /** |
236 | * amdgpu_device_supports_px - Is the device a dGPU with ATPX power control |
237 | * |
238 | * @dev: drm_device pointer |
239 | * |
240 | * Returns true if the device is a dGPU with ATPX power control, |
241 | * otherwise return false. |
242 | */ |
243 | bool amdgpu_device_supports_px(struct drm_device *dev) |
244 | { |
245 | struct amdgpu_device *adev = drm_to_adev(ddev: dev); |
246 | |
247 | if ((adev->flags & AMD_IS_PX) && !amdgpu_is_atpx_hybrid()) |
248 | return true; |
249 | return false; |
250 | } |
251 | |
252 | /** |
253 | * amdgpu_device_supports_boco - Is the device a dGPU with ACPI power resources |
254 | * |
255 | * @dev: drm_device pointer |
256 | * |
257 | * Returns true if the device is a dGPU with ACPI power control, |
258 | * otherwise return false. |
259 | */ |
260 | bool amdgpu_device_supports_boco(struct drm_device *dev) |
261 | { |
262 | struct amdgpu_device *adev = drm_to_adev(ddev: dev); |
263 | |
264 | if (adev->has_pr3 || |
265 | ((adev->flags & AMD_IS_PX) && amdgpu_is_atpx_hybrid())) |
266 | return true; |
267 | return false; |
268 | } |
269 | |
270 | /** |
271 | * amdgpu_device_supports_baco - Does the device support BACO |
272 | * |
273 | * @dev: drm_device pointer |
274 | * |
275 | * Returns true if the device supporte BACO, |
276 | * otherwise return false. |
277 | */ |
278 | bool amdgpu_device_supports_baco(struct drm_device *dev) |
279 | { |
280 | struct amdgpu_device *adev = drm_to_adev(ddev: dev); |
281 | |
282 | return amdgpu_asic_supports_baco(adev); |
283 | } |
284 | |
285 | /** |
286 | * amdgpu_device_supports_smart_shift - Is the device dGPU with |
287 | * smart shift support |
288 | * |
289 | * @dev: drm_device pointer |
290 | * |
291 | * Returns true if the device is a dGPU with Smart Shift support, |
292 | * otherwise returns false. |
293 | */ |
294 | bool amdgpu_device_supports_smart_shift(struct drm_device *dev) |
295 | { |
296 | return (amdgpu_device_supports_boco(dev) && |
297 | amdgpu_acpi_is_power_shift_control_supported()); |
298 | } |
299 | |
300 | /* |
301 | * VRAM access helper functions |
302 | */ |
303 | |
304 | /** |
305 | * amdgpu_device_mm_access - access vram by MM_INDEX/MM_DATA |
306 | * |
307 | * @adev: amdgpu_device pointer |
308 | * @pos: offset of the buffer in vram |
309 | * @buf: virtual address of the buffer in system memory |
310 | * @size: read/write size, sizeof(@buf) must > @size |
311 | * @write: true - write to vram, otherwise - read from vram |
312 | */ |
313 | void amdgpu_device_mm_access(struct amdgpu_device *adev, loff_t pos, |
314 | void *buf, size_t size, bool write) |
315 | { |
316 | unsigned long flags; |
317 | uint32_t hi = ~0, tmp = 0; |
318 | uint32_t *data = buf; |
319 | uint64_t last; |
320 | int idx; |
321 | |
322 | if (!drm_dev_enter(dev: adev_to_drm(adev), idx: &idx)) |
323 | return; |
324 | |
325 | BUG_ON(!IS_ALIGNED(pos, 4) || !IS_ALIGNED(size, 4)); |
326 | |
327 | spin_lock_irqsave(&adev->mmio_idx_lock, flags); |
328 | for (last = pos + size; pos < last; pos += 4) { |
329 | tmp = pos >> 31; |
330 | |
331 | WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)pos) | 0x80000000); |
332 | if (tmp != hi) { |
333 | WREG32_NO_KIQ(mmMM_INDEX_HI, tmp); |
334 | hi = tmp; |
335 | } |
336 | if (write) |
337 | WREG32_NO_KIQ(mmMM_DATA, *data++); |
338 | else |
339 | *data++ = RREG32_NO_KIQ(mmMM_DATA); |
340 | } |
341 | |
342 | spin_unlock_irqrestore(lock: &adev->mmio_idx_lock, flags); |
343 | drm_dev_exit(idx); |
344 | } |
345 | |
346 | /** |
347 | * amdgpu_device_aper_access - access vram by vram aperature |
348 | * |
349 | * @adev: amdgpu_device pointer |
350 | * @pos: offset of the buffer in vram |
351 | * @buf: virtual address of the buffer in system memory |
352 | * @size: read/write size, sizeof(@buf) must > @size |
353 | * @write: true - write to vram, otherwise - read from vram |
354 | * |
355 | * The return value means how many bytes have been transferred. |
356 | */ |
357 | size_t amdgpu_device_aper_access(struct amdgpu_device *adev, loff_t pos, |
358 | void *buf, size_t size, bool write) |
359 | { |
360 | #ifdef CONFIG_64BIT |
361 | void __iomem *addr; |
362 | size_t count = 0; |
363 | uint64_t last; |
364 | |
365 | if (!adev->mman.aper_base_kaddr) |
366 | return 0; |
367 | |
368 | last = min(pos + size, adev->gmc.visible_vram_size); |
369 | if (last > pos) { |
370 | addr = adev->mman.aper_base_kaddr + pos; |
371 | count = last - pos; |
372 | |
373 | if (write) { |
374 | memcpy_toio(addr, buf, count); |
375 | /* Make sure HDP write cache flush happens without any reordering |
376 | * after the system memory contents are sent over PCIe device |
377 | */ |
378 | mb(); |
379 | amdgpu_device_flush_hdp(adev, NULL); |
380 | } else { |
381 | amdgpu_device_invalidate_hdp(adev, NULL); |
382 | /* Make sure HDP read cache is invalidated before issuing a read |
383 | * to the PCIe device |
384 | */ |
385 | mb(); |
386 | memcpy_fromio(buf, addr, count); |
387 | } |
388 | |
389 | } |
390 | |
391 | return count; |
392 | #else |
393 | return 0; |
394 | #endif |
395 | } |
396 | |
397 | /** |
398 | * amdgpu_device_vram_access - read/write a buffer in vram |
399 | * |
400 | * @adev: amdgpu_device pointer |
401 | * @pos: offset of the buffer in vram |
402 | * @buf: virtual address of the buffer in system memory |
403 | * @size: read/write size, sizeof(@buf) must > @size |
404 | * @write: true - write to vram, otherwise - read from vram |
405 | */ |
406 | void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos, |
407 | void *buf, size_t size, bool write) |
408 | { |
409 | size_t count; |
410 | |
411 | /* try to using vram apreature to access vram first */ |
412 | count = amdgpu_device_aper_access(adev, pos, buf, size, write); |
413 | size -= count; |
414 | if (size) { |
415 | /* using MM to access rest vram */ |
416 | pos += count; |
417 | buf += count; |
418 | amdgpu_device_mm_access(adev, pos, buf, size, write); |
419 | } |
420 | } |
421 | |
422 | /* |
423 | * register access helper functions. |
424 | */ |
425 | |
426 | /* Check if hw access should be skipped because of hotplug or device error */ |
427 | bool amdgpu_device_skip_hw_access(struct amdgpu_device *adev) |
428 | { |
429 | if (adev->no_hw_access) |
430 | return true; |
431 | |
432 | #ifdef CONFIG_LOCKDEP |
433 | /* |
434 | * This is a bit complicated to understand, so worth a comment. What we assert |
435 | * here is that the GPU reset is not running on another thread in parallel. |
436 | * |
437 | * For this we trylock the read side of the reset semaphore, if that succeeds |
438 | * we know that the reset is not running in paralell. |
439 | * |
440 | * If the trylock fails we assert that we are either already holding the read |
441 | * side of the lock or are the reset thread itself and hold the write side of |
442 | * the lock. |
443 | */ |
444 | if (in_task()) { |
445 | if (down_read_trylock(sem: &adev->reset_domain->sem)) |
446 | up_read(sem: &adev->reset_domain->sem); |
447 | else |
448 | lockdep_assert_held(&adev->reset_domain->sem); |
449 | } |
450 | #endif |
451 | return false; |
452 | } |
453 | |
454 | /** |
455 | * amdgpu_device_rreg - read a memory mapped IO or indirect register |
456 | * |
457 | * @adev: amdgpu_device pointer |
458 | * @reg: dword aligned register offset |
459 | * @acc_flags: access flags which require special behavior |
460 | * |
461 | * Returns the 32 bit value from the offset specified. |
462 | */ |
463 | uint32_t amdgpu_device_rreg(struct amdgpu_device *adev, |
464 | uint32_t reg, uint32_t acc_flags) |
465 | { |
466 | uint32_t ret; |
467 | |
468 | if (amdgpu_device_skip_hw_access(adev)) |
469 | return 0; |
470 | |
471 | if ((reg * 4) < adev->rmmio_size) { |
472 | if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && |
473 | amdgpu_sriov_runtime(adev) && |
474 | down_read_trylock(sem: &adev->reset_domain->sem)) { |
475 | ret = amdgpu_kiq_rreg(adev, reg); |
476 | up_read(sem: &adev->reset_domain->sem); |
477 | } else { |
478 | ret = readl(addr: ((void __iomem *)adev->rmmio) + (reg * 4)); |
479 | } |
480 | } else { |
481 | ret = adev->pcie_rreg(adev, reg * 4); |
482 | } |
483 | |
484 | trace_amdgpu_device_rreg(did: adev->pdev->device, reg, value: ret); |
485 | |
486 | return ret; |
487 | } |
488 | |
489 | /* |
490 | * MMIO register read with bytes helper functions |
491 | * @offset:bytes offset from MMIO start |
492 | */ |
493 | |
494 | /** |
495 | * amdgpu_mm_rreg8 - read a memory mapped IO register |
496 | * |
497 | * @adev: amdgpu_device pointer |
498 | * @offset: byte aligned register offset |
499 | * |
500 | * Returns the 8 bit value from the offset specified. |
501 | */ |
502 | uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset) |
503 | { |
504 | if (amdgpu_device_skip_hw_access(adev)) |
505 | return 0; |
506 | |
507 | if (offset < adev->rmmio_size) |
508 | return (readb(addr: adev->rmmio + offset)); |
509 | BUG(); |
510 | } |
511 | |
512 | /* |
513 | * MMIO register write with bytes helper functions |
514 | * @offset:bytes offset from MMIO start |
515 | * @value: the value want to be written to the register |
516 | */ |
517 | |
518 | /** |
519 | * amdgpu_mm_wreg8 - read a memory mapped IO register |
520 | * |
521 | * @adev: amdgpu_device pointer |
522 | * @offset: byte aligned register offset |
523 | * @value: 8 bit value to write |
524 | * |
525 | * Writes the value specified to the offset specified. |
526 | */ |
527 | void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value) |
528 | { |
529 | if (amdgpu_device_skip_hw_access(adev)) |
530 | return; |
531 | |
532 | if (offset < adev->rmmio_size) |
533 | writeb(val: value, addr: adev->rmmio + offset); |
534 | else |
535 | BUG(); |
536 | } |
537 | |
538 | /** |
539 | * amdgpu_device_wreg - write to a memory mapped IO or indirect register |
540 | * |
541 | * @adev: amdgpu_device pointer |
542 | * @reg: dword aligned register offset |
543 | * @v: 32 bit value to write to the register |
544 | * @acc_flags: access flags which require special behavior |
545 | * |
546 | * Writes the value specified to the offset specified. |
547 | */ |
548 | void amdgpu_device_wreg(struct amdgpu_device *adev, |
549 | uint32_t reg, uint32_t v, |
550 | uint32_t acc_flags) |
551 | { |
552 | if (amdgpu_device_skip_hw_access(adev)) |
553 | return; |
554 | |
555 | if ((reg * 4) < adev->rmmio_size) { |
556 | if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && |
557 | amdgpu_sriov_runtime(adev) && |
558 | down_read_trylock(sem: &adev->reset_domain->sem)) { |
559 | amdgpu_kiq_wreg(adev, reg, v); |
560 | up_read(sem: &adev->reset_domain->sem); |
561 | } else { |
562 | writel(val: v, addr: ((void __iomem *)adev->rmmio) + (reg * 4)); |
563 | } |
564 | } else { |
565 | adev->pcie_wreg(adev, reg * 4, v); |
566 | } |
567 | |
568 | trace_amdgpu_device_wreg(did: adev->pdev->device, reg, value: v); |
569 | } |
570 | |
571 | /** |
572 | * amdgpu_mm_wreg_mmio_rlc - write register either with direct/indirect mmio or with RLC path if in range |
573 | * |
574 | * @adev: amdgpu_device pointer |
575 | * @reg: mmio/rlc register |
576 | * @v: value to write |
577 | * @xcc_id: xcc accelerated compute core id |
578 | * |
579 | * this function is invoked only for the debugfs register access |
580 | */ |
581 | void amdgpu_mm_wreg_mmio_rlc(struct amdgpu_device *adev, |
582 | uint32_t reg, uint32_t v, |
583 | uint32_t xcc_id) |
584 | { |
585 | if (amdgpu_device_skip_hw_access(adev)) |
586 | return; |
587 | |
588 | if (amdgpu_sriov_fullaccess(adev) && |
589 | adev->gfx.rlc.funcs && |
590 | adev->gfx.rlc.funcs->is_rlcg_access_range) { |
591 | if (adev->gfx.rlc.funcs->is_rlcg_access_range(adev, reg)) |
592 | return amdgpu_sriov_wreg(adev, offset: reg, value: v, acc_flags: 0, hwip: 0, xcc_id); |
593 | } else if ((reg * 4) >= adev->rmmio_size) { |
594 | adev->pcie_wreg(adev, reg * 4, v); |
595 | } else { |
596 | writel(val: v, addr: ((void __iomem *)adev->rmmio) + (reg * 4)); |
597 | } |
598 | } |
599 | |
600 | /** |
601 | * amdgpu_device_indirect_rreg - read an indirect register |
602 | * |
603 | * @adev: amdgpu_device pointer |
604 | * @reg_addr: indirect register address to read from |
605 | * |
606 | * Returns the value of indirect register @reg_addr |
607 | */ |
608 | u32 amdgpu_device_indirect_rreg(struct amdgpu_device *adev, |
609 | u32 reg_addr) |
610 | { |
611 | unsigned long flags, pcie_index, pcie_data; |
612 | void __iomem *pcie_index_offset; |
613 | void __iomem *pcie_data_offset; |
614 | u32 r; |
615 | |
616 | pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev); |
617 | pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev); |
618 | |
619 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); |
620 | pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4; |
621 | pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4; |
622 | |
623 | writel(val: reg_addr, addr: pcie_index_offset); |
624 | readl(addr: pcie_index_offset); |
625 | r = readl(addr: pcie_data_offset); |
626 | spin_unlock_irqrestore(lock: &adev->pcie_idx_lock, flags); |
627 | |
628 | return r; |
629 | } |
630 | |
631 | u32 amdgpu_device_indirect_rreg_ext(struct amdgpu_device *adev, |
632 | u64 reg_addr) |
633 | { |
634 | unsigned long flags, pcie_index, pcie_index_hi, pcie_data; |
635 | u32 r; |
636 | void __iomem *pcie_index_offset; |
637 | void __iomem *pcie_index_hi_offset; |
638 | void __iomem *pcie_data_offset; |
639 | |
640 | pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev); |
641 | pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev); |
642 | if ((reg_addr >> 32) && (adev->nbio.funcs->get_pcie_index_hi_offset)) |
643 | pcie_index_hi = adev->nbio.funcs->get_pcie_index_hi_offset(adev); |
644 | else |
645 | pcie_index_hi = 0; |
646 | |
647 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); |
648 | pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4; |
649 | pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4; |
650 | if (pcie_index_hi != 0) |
651 | pcie_index_hi_offset = (void __iomem *)adev->rmmio + |
652 | pcie_index_hi * 4; |
653 | |
654 | writel(val: reg_addr, addr: pcie_index_offset); |
655 | readl(addr: pcie_index_offset); |
656 | if (pcie_index_hi != 0) { |
657 | writel(val: (reg_addr >> 32) & 0xff, addr: pcie_index_hi_offset); |
658 | readl(addr: pcie_index_hi_offset); |
659 | } |
660 | r = readl(addr: pcie_data_offset); |
661 | |
662 | /* clear the high bits */ |
663 | if (pcie_index_hi != 0) { |
664 | writel(val: 0, addr: pcie_index_hi_offset); |
665 | readl(addr: pcie_index_hi_offset); |
666 | } |
667 | |
668 | spin_unlock_irqrestore(lock: &adev->pcie_idx_lock, flags); |
669 | |
670 | return r; |
671 | } |
672 | |
673 | /** |
674 | * amdgpu_device_indirect_rreg64 - read a 64bits indirect register |
675 | * |
676 | * @adev: amdgpu_device pointer |
677 | * @reg_addr: indirect register address to read from |
678 | * |
679 | * Returns the value of indirect register @reg_addr |
680 | */ |
681 | u64 amdgpu_device_indirect_rreg64(struct amdgpu_device *adev, |
682 | u32 reg_addr) |
683 | { |
684 | unsigned long flags, pcie_index, pcie_data; |
685 | void __iomem *pcie_index_offset; |
686 | void __iomem *pcie_data_offset; |
687 | u64 r; |
688 | |
689 | pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev); |
690 | pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev); |
691 | |
692 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); |
693 | pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4; |
694 | pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4; |
695 | |
696 | /* read low 32 bits */ |
697 | writel(val: reg_addr, addr: pcie_index_offset); |
698 | readl(addr: pcie_index_offset); |
699 | r = readl(addr: pcie_data_offset); |
700 | /* read high 32 bits */ |
701 | writel(val: reg_addr + 4, addr: pcie_index_offset); |
702 | readl(addr: pcie_index_offset); |
703 | r |= ((u64)readl(addr: pcie_data_offset) << 32); |
704 | spin_unlock_irqrestore(lock: &adev->pcie_idx_lock, flags); |
705 | |
706 | return r; |
707 | } |
708 | |
709 | u64 amdgpu_device_indirect_rreg64_ext(struct amdgpu_device *adev, |
710 | u64 reg_addr) |
711 | { |
712 | unsigned long flags, pcie_index, pcie_data; |
713 | unsigned long pcie_index_hi = 0; |
714 | void __iomem *pcie_index_offset; |
715 | void __iomem *pcie_index_hi_offset; |
716 | void __iomem *pcie_data_offset; |
717 | u64 r; |
718 | |
719 | pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev); |
720 | pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev); |
721 | if ((reg_addr >> 32) && (adev->nbio.funcs->get_pcie_index_hi_offset)) |
722 | pcie_index_hi = adev->nbio.funcs->get_pcie_index_hi_offset(adev); |
723 | |
724 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); |
725 | pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4; |
726 | pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4; |
727 | if (pcie_index_hi != 0) |
728 | pcie_index_hi_offset = (void __iomem *)adev->rmmio + |
729 | pcie_index_hi * 4; |
730 | |
731 | /* read low 32 bits */ |
732 | writel(val: reg_addr, addr: pcie_index_offset); |
733 | readl(addr: pcie_index_offset); |
734 | if (pcie_index_hi != 0) { |
735 | writel(val: (reg_addr >> 32) & 0xff, addr: pcie_index_hi_offset); |
736 | readl(addr: pcie_index_hi_offset); |
737 | } |
738 | r = readl(addr: pcie_data_offset); |
739 | /* read high 32 bits */ |
740 | writel(val: reg_addr + 4, addr: pcie_index_offset); |
741 | readl(addr: pcie_index_offset); |
742 | if (pcie_index_hi != 0) { |
743 | writel(val: (reg_addr >> 32) & 0xff, addr: pcie_index_hi_offset); |
744 | readl(addr: pcie_index_hi_offset); |
745 | } |
746 | r |= ((u64)readl(addr: pcie_data_offset) << 32); |
747 | |
748 | /* clear the high bits */ |
749 | if (pcie_index_hi != 0) { |
750 | writel(val: 0, addr: pcie_index_hi_offset); |
751 | readl(addr: pcie_index_hi_offset); |
752 | } |
753 | |
754 | spin_unlock_irqrestore(lock: &adev->pcie_idx_lock, flags); |
755 | |
756 | return r; |
757 | } |
758 | |
759 | /** |
760 | * amdgpu_device_indirect_wreg - write an indirect register address |
761 | * |
762 | * @adev: amdgpu_device pointer |
763 | * @reg_addr: indirect register offset |
764 | * @reg_data: indirect register data |
765 | * |
766 | */ |
767 | void amdgpu_device_indirect_wreg(struct amdgpu_device *adev, |
768 | u32 reg_addr, u32 reg_data) |
769 | { |
770 | unsigned long flags, pcie_index, pcie_data; |
771 | void __iomem *pcie_index_offset; |
772 | void __iomem *pcie_data_offset; |
773 | |
774 | pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev); |
775 | pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev); |
776 | |
777 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); |
778 | pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4; |
779 | pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4; |
780 | |
781 | writel(val: reg_addr, addr: pcie_index_offset); |
782 | readl(addr: pcie_index_offset); |
783 | writel(val: reg_data, addr: pcie_data_offset); |
784 | readl(addr: pcie_data_offset); |
785 | spin_unlock_irqrestore(lock: &adev->pcie_idx_lock, flags); |
786 | } |
787 | |
788 | void amdgpu_device_indirect_wreg_ext(struct amdgpu_device *adev, |
789 | u64 reg_addr, u32 reg_data) |
790 | { |
791 | unsigned long flags, pcie_index, pcie_index_hi, pcie_data; |
792 | void __iomem *pcie_index_offset; |
793 | void __iomem *pcie_index_hi_offset; |
794 | void __iomem *pcie_data_offset; |
795 | |
796 | pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev); |
797 | pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev); |
798 | if ((reg_addr >> 32) && (adev->nbio.funcs->get_pcie_index_hi_offset)) |
799 | pcie_index_hi = adev->nbio.funcs->get_pcie_index_hi_offset(adev); |
800 | else |
801 | pcie_index_hi = 0; |
802 | |
803 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); |
804 | pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4; |
805 | pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4; |
806 | if (pcie_index_hi != 0) |
807 | pcie_index_hi_offset = (void __iomem *)adev->rmmio + |
808 | pcie_index_hi * 4; |
809 | |
810 | writel(val: reg_addr, addr: pcie_index_offset); |
811 | readl(addr: pcie_index_offset); |
812 | if (pcie_index_hi != 0) { |
813 | writel(val: (reg_addr >> 32) & 0xff, addr: pcie_index_hi_offset); |
814 | readl(addr: pcie_index_hi_offset); |
815 | } |
816 | writel(val: reg_data, addr: pcie_data_offset); |
817 | readl(addr: pcie_data_offset); |
818 | |
819 | /* clear the high bits */ |
820 | if (pcie_index_hi != 0) { |
821 | writel(val: 0, addr: pcie_index_hi_offset); |
822 | readl(addr: pcie_index_hi_offset); |
823 | } |
824 | |
825 | spin_unlock_irqrestore(lock: &adev->pcie_idx_lock, flags); |
826 | } |
827 | |
828 | /** |
829 | * amdgpu_device_indirect_wreg64 - write a 64bits indirect register address |
830 | * |
831 | * @adev: amdgpu_device pointer |
832 | * @reg_addr: indirect register offset |
833 | * @reg_data: indirect register data |
834 | * |
835 | */ |
836 | void amdgpu_device_indirect_wreg64(struct amdgpu_device *adev, |
837 | u32 reg_addr, u64 reg_data) |
838 | { |
839 | unsigned long flags, pcie_index, pcie_data; |
840 | void __iomem *pcie_index_offset; |
841 | void __iomem *pcie_data_offset; |
842 | |
843 | pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev); |
844 | pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev); |
845 | |
846 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); |
847 | pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4; |
848 | pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4; |
849 | |
850 | /* write low 32 bits */ |
851 | writel(val: reg_addr, addr: pcie_index_offset); |
852 | readl(addr: pcie_index_offset); |
853 | writel(val: (u32)(reg_data & 0xffffffffULL), addr: pcie_data_offset); |
854 | readl(addr: pcie_data_offset); |
855 | /* write high 32 bits */ |
856 | writel(val: reg_addr + 4, addr: pcie_index_offset); |
857 | readl(addr: pcie_index_offset); |
858 | writel(val: (u32)(reg_data >> 32), addr: pcie_data_offset); |
859 | readl(addr: pcie_data_offset); |
860 | spin_unlock_irqrestore(lock: &adev->pcie_idx_lock, flags); |
861 | } |
862 | |
863 | void amdgpu_device_indirect_wreg64_ext(struct amdgpu_device *adev, |
864 | u64 reg_addr, u64 reg_data) |
865 | { |
866 | unsigned long flags, pcie_index, pcie_data; |
867 | unsigned long pcie_index_hi = 0; |
868 | void __iomem *pcie_index_offset; |
869 | void __iomem *pcie_index_hi_offset; |
870 | void __iomem *pcie_data_offset; |
871 | |
872 | pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev); |
873 | pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev); |
874 | if ((reg_addr >> 32) && (adev->nbio.funcs->get_pcie_index_hi_offset)) |
875 | pcie_index_hi = adev->nbio.funcs->get_pcie_index_hi_offset(adev); |
876 | |
877 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); |
878 | pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4; |
879 | pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4; |
880 | if (pcie_index_hi != 0) |
881 | pcie_index_hi_offset = (void __iomem *)adev->rmmio + |
882 | pcie_index_hi * 4; |
883 | |
884 | /* write low 32 bits */ |
885 | writel(val: reg_addr, addr: pcie_index_offset); |
886 | readl(addr: pcie_index_offset); |
887 | if (pcie_index_hi != 0) { |
888 | writel(val: (reg_addr >> 32) & 0xff, addr: pcie_index_hi_offset); |
889 | readl(addr: pcie_index_hi_offset); |
890 | } |
891 | writel(val: (u32)(reg_data & 0xffffffffULL), addr: pcie_data_offset); |
892 | readl(addr: pcie_data_offset); |
893 | /* write high 32 bits */ |
894 | writel(val: reg_addr + 4, addr: pcie_index_offset); |
895 | readl(addr: pcie_index_offset); |
896 | if (pcie_index_hi != 0) { |
897 | writel(val: (reg_addr >> 32) & 0xff, addr: pcie_index_hi_offset); |
898 | readl(addr: pcie_index_hi_offset); |
899 | } |
900 | writel(val: (u32)(reg_data >> 32), addr: pcie_data_offset); |
901 | readl(addr: pcie_data_offset); |
902 | |
903 | /* clear the high bits */ |
904 | if (pcie_index_hi != 0) { |
905 | writel(val: 0, addr: pcie_index_hi_offset); |
906 | readl(addr: pcie_index_hi_offset); |
907 | } |
908 | |
909 | spin_unlock_irqrestore(lock: &adev->pcie_idx_lock, flags); |
910 | } |
911 | |
912 | /** |
913 | * amdgpu_device_get_rev_id - query device rev_id |
914 | * |
915 | * @adev: amdgpu_device pointer |
916 | * |
917 | * Return device rev_id |
918 | */ |
919 | u32 amdgpu_device_get_rev_id(struct amdgpu_device *adev) |
920 | { |
921 | return adev->nbio.funcs->get_rev_id(adev); |
922 | } |
923 | |
924 | /** |
925 | * amdgpu_invalid_rreg - dummy reg read function |
926 | * |
927 | * @adev: amdgpu_device pointer |
928 | * @reg: offset of register |
929 | * |
930 | * Dummy register read function. Used for register blocks |
931 | * that certain asics don't have (all asics). |
932 | * Returns the value in the register. |
933 | */ |
934 | static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg) |
935 | { |
936 | DRM_ERROR("Invalid callback to read register 0x%04X\n" , reg); |
937 | BUG(); |
938 | return 0; |
939 | } |
940 | |
941 | static uint32_t amdgpu_invalid_rreg_ext(struct amdgpu_device *adev, uint64_t reg) |
942 | { |
943 | DRM_ERROR("Invalid callback to read register 0x%llX\n" , reg); |
944 | BUG(); |
945 | return 0; |
946 | } |
947 | |
948 | /** |
949 | * amdgpu_invalid_wreg - dummy reg write function |
950 | * |
951 | * @adev: amdgpu_device pointer |
952 | * @reg: offset of register |
953 | * @v: value to write to the register |
954 | * |
955 | * Dummy register read function. Used for register blocks |
956 | * that certain asics don't have (all asics). |
957 | */ |
958 | static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v) |
959 | { |
960 | DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n" , |
961 | reg, v); |
962 | BUG(); |
963 | } |
964 | |
965 | static void amdgpu_invalid_wreg_ext(struct amdgpu_device *adev, uint64_t reg, uint32_t v) |
966 | { |
967 | DRM_ERROR("Invalid callback to write register 0x%llX with 0x%08X\n" , |
968 | reg, v); |
969 | BUG(); |
970 | } |
971 | |
972 | /** |
973 | * amdgpu_invalid_rreg64 - dummy 64 bit reg read function |
974 | * |
975 | * @adev: amdgpu_device pointer |
976 | * @reg: offset of register |
977 | * |
978 | * Dummy register read function. Used for register blocks |
979 | * that certain asics don't have (all asics). |
980 | * Returns the value in the register. |
981 | */ |
982 | static uint64_t amdgpu_invalid_rreg64(struct amdgpu_device *adev, uint32_t reg) |
983 | { |
984 | DRM_ERROR("Invalid callback to read 64 bit register 0x%04X\n" , reg); |
985 | BUG(); |
986 | return 0; |
987 | } |
988 | |
989 | static uint64_t amdgpu_invalid_rreg64_ext(struct amdgpu_device *adev, uint64_t reg) |
990 | { |
991 | DRM_ERROR("Invalid callback to read register 0x%llX\n" , reg); |
992 | BUG(); |
993 | return 0; |
994 | } |
995 | |
996 | /** |
997 | * amdgpu_invalid_wreg64 - dummy reg write function |
998 | * |
999 | * @adev: amdgpu_device pointer |
1000 | * @reg: offset of register |
1001 | * @v: value to write to the register |
1002 | * |
1003 | * Dummy register read function. Used for register blocks |
1004 | * that certain asics don't have (all asics). |
1005 | */ |
1006 | static void amdgpu_invalid_wreg64(struct amdgpu_device *adev, uint32_t reg, uint64_t v) |
1007 | { |
1008 | DRM_ERROR("Invalid callback to write 64 bit register 0x%04X with 0x%08llX\n" , |
1009 | reg, v); |
1010 | BUG(); |
1011 | } |
1012 | |
1013 | static void amdgpu_invalid_wreg64_ext(struct amdgpu_device *adev, uint64_t reg, uint64_t v) |
1014 | { |
1015 | DRM_ERROR("Invalid callback to write 64 bit register 0x%llX with 0x%08llX\n" , |
1016 | reg, v); |
1017 | BUG(); |
1018 | } |
1019 | |
1020 | /** |
1021 | * amdgpu_block_invalid_rreg - dummy reg read function |
1022 | * |
1023 | * @adev: amdgpu_device pointer |
1024 | * @block: offset of instance |
1025 | * @reg: offset of register |
1026 | * |
1027 | * Dummy register read function. Used for register blocks |
1028 | * that certain asics don't have (all asics). |
1029 | * Returns the value in the register. |
1030 | */ |
1031 | static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev, |
1032 | uint32_t block, uint32_t reg) |
1033 | { |
1034 | DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n" , |
1035 | reg, block); |
1036 | BUG(); |
1037 | return 0; |
1038 | } |
1039 | |
1040 | /** |
1041 | * amdgpu_block_invalid_wreg - dummy reg write function |
1042 | * |
1043 | * @adev: amdgpu_device pointer |
1044 | * @block: offset of instance |
1045 | * @reg: offset of register |
1046 | * @v: value to write to the register |
1047 | * |
1048 | * Dummy register read function. Used for register blocks |
1049 | * that certain asics don't have (all asics). |
1050 | */ |
1051 | static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev, |
1052 | uint32_t block, |
1053 | uint32_t reg, uint32_t v) |
1054 | { |
1055 | DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n" , |
1056 | reg, block, v); |
1057 | BUG(); |
1058 | } |
1059 | |
1060 | /** |
1061 | * amdgpu_device_asic_init - Wrapper for atom asic_init |
1062 | * |
1063 | * @adev: amdgpu_device pointer |
1064 | * |
1065 | * Does any asic specific work and then calls atom asic init. |
1066 | */ |
1067 | static int amdgpu_device_asic_init(struct amdgpu_device *adev) |
1068 | { |
1069 | int ret; |
1070 | |
1071 | amdgpu_asic_pre_asic_init(adev); |
1072 | |
1073 | if (amdgpu_ip_version(adev, ip: GC_HWIP, inst: 0) == IP_VERSION(9, 4, 3) || |
1074 | amdgpu_ip_version(adev, ip: GC_HWIP, inst: 0) >= IP_VERSION(11, 0, 0)) { |
1075 | amdgpu_psp_wait_for_bootloader(adev); |
1076 | ret = amdgpu_atomfirmware_asic_init(adev, fb_reset: true); |
1077 | /* TODO: check the return val and stop device initialization if boot fails */ |
1078 | amdgpu_psp_query_boot_status(adev); |
1079 | return ret; |
1080 | } else { |
1081 | return amdgpu_atom_asic_init(ctx: adev->mode_info.atom_context); |
1082 | } |
1083 | |
1084 | return 0; |
1085 | } |
1086 | |
1087 | /** |
1088 | * amdgpu_device_mem_scratch_init - allocate the VRAM scratch page |
1089 | * |
1090 | * @adev: amdgpu_device pointer |
1091 | * |
1092 | * Allocates a scratch page of VRAM for use by various things in the |
1093 | * driver. |
1094 | */ |
1095 | static int amdgpu_device_mem_scratch_init(struct amdgpu_device *adev) |
1096 | { |
1097 | return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE, PAGE_SIZE, |
1098 | AMDGPU_GEM_DOMAIN_VRAM | |
1099 | AMDGPU_GEM_DOMAIN_GTT, |
1100 | bo_ptr: &adev->mem_scratch.robj, |
1101 | gpu_addr: &adev->mem_scratch.gpu_addr, |
1102 | cpu_addr: (void **)&adev->mem_scratch.ptr); |
1103 | } |
1104 | |
1105 | /** |
1106 | * amdgpu_device_mem_scratch_fini - Free the VRAM scratch page |
1107 | * |
1108 | * @adev: amdgpu_device pointer |
1109 | * |
1110 | * Frees the VRAM scratch page. |
1111 | */ |
1112 | static void amdgpu_device_mem_scratch_fini(struct amdgpu_device *adev) |
1113 | { |
1114 | amdgpu_bo_free_kernel(bo: &adev->mem_scratch.robj, NULL, NULL); |
1115 | } |
1116 | |
1117 | /** |
1118 | * amdgpu_device_program_register_sequence - program an array of registers. |
1119 | * |
1120 | * @adev: amdgpu_device pointer |
1121 | * @registers: pointer to the register array |
1122 | * @array_size: size of the register array |
1123 | * |
1124 | * Programs an array or registers with and or masks. |
1125 | * This is a helper for setting golden registers. |
1126 | */ |
1127 | void amdgpu_device_program_register_sequence(struct amdgpu_device *adev, |
1128 | const u32 *registers, |
1129 | const u32 array_size) |
1130 | { |
1131 | u32 tmp, reg, and_mask, or_mask; |
1132 | int i; |
1133 | |
1134 | if (array_size % 3) |
1135 | return; |
1136 | |
1137 | for (i = 0; i < array_size; i += 3) { |
1138 | reg = registers[i + 0]; |
1139 | and_mask = registers[i + 1]; |
1140 | or_mask = registers[i + 2]; |
1141 | |
1142 | if (and_mask == 0xffffffff) { |
1143 | tmp = or_mask; |
1144 | } else { |
1145 | tmp = RREG32(reg); |
1146 | tmp &= ~and_mask; |
1147 | if (adev->family >= AMDGPU_FAMILY_AI) |
1148 | tmp |= (or_mask & and_mask); |
1149 | else |
1150 | tmp |= or_mask; |
1151 | } |
1152 | WREG32(reg, tmp); |
1153 | } |
1154 | } |
1155 | |
1156 | /** |
1157 | * amdgpu_device_pci_config_reset - reset the GPU |
1158 | * |
1159 | * @adev: amdgpu_device pointer |
1160 | * |
1161 | * Resets the GPU using the pci config reset sequence. |
1162 | * Only applicable to asics prior to vega10. |
1163 | */ |
1164 | void amdgpu_device_pci_config_reset(struct amdgpu_device *adev) |
1165 | { |
1166 | pci_write_config_dword(dev: adev->pdev, where: 0x7c, AMDGPU_ASIC_RESET_DATA); |
1167 | } |
1168 | |
1169 | /** |
1170 | * amdgpu_device_pci_reset - reset the GPU using generic PCI means |
1171 | * |
1172 | * @adev: amdgpu_device pointer |
1173 | * |
1174 | * Resets the GPU using generic pci reset interfaces (FLR, SBR, etc.). |
1175 | */ |
1176 | int amdgpu_device_pci_reset(struct amdgpu_device *adev) |
1177 | { |
1178 | return pci_reset_function(dev: adev->pdev); |
1179 | } |
1180 | |
1181 | /* |
1182 | * amdgpu_device_wb_*() |
1183 | * Writeback is the method by which the GPU updates special pages in memory |
1184 | * with the status of certain GPU events (fences, ring pointers,etc.). |
1185 | */ |
1186 | |
1187 | /** |
1188 | * amdgpu_device_wb_fini - Disable Writeback and free memory |
1189 | * |
1190 | * @adev: amdgpu_device pointer |
1191 | * |
1192 | * Disables Writeback and frees the Writeback memory (all asics). |
1193 | * Used at driver shutdown. |
1194 | */ |
1195 | static void amdgpu_device_wb_fini(struct amdgpu_device *adev) |
1196 | { |
1197 | if (adev->wb.wb_obj) { |
1198 | amdgpu_bo_free_kernel(bo: &adev->wb.wb_obj, |
1199 | gpu_addr: &adev->wb.gpu_addr, |
1200 | cpu_addr: (void **)&adev->wb.wb); |
1201 | adev->wb.wb_obj = NULL; |
1202 | } |
1203 | } |
1204 | |
1205 | /** |
1206 | * amdgpu_device_wb_init - Init Writeback driver info and allocate memory |
1207 | * |
1208 | * @adev: amdgpu_device pointer |
1209 | * |
1210 | * Initializes writeback and allocates writeback memory (all asics). |
1211 | * Used at driver startup. |
1212 | * Returns 0 on success or an -error on failure. |
1213 | */ |
1214 | static int amdgpu_device_wb_init(struct amdgpu_device *adev) |
1215 | { |
1216 | int r; |
1217 | |
1218 | if (adev->wb.wb_obj == NULL) { |
1219 | /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */ |
1220 | r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8, |
1221 | PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT, |
1222 | bo_ptr: &adev->wb.wb_obj, gpu_addr: &adev->wb.gpu_addr, |
1223 | cpu_addr: (void **)&adev->wb.wb); |
1224 | if (r) { |
1225 | dev_warn(adev->dev, "(%d) create WB bo failed\n" , r); |
1226 | return r; |
1227 | } |
1228 | |
1229 | adev->wb.num_wb = AMDGPU_MAX_WB; |
1230 | memset(&adev->wb.used, 0, sizeof(adev->wb.used)); |
1231 | |
1232 | /* clear wb memory */ |
1233 | memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8); |
1234 | } |
1235 | |
1236 | return 0; |
1237 | } |
1238 | |
1239 | /** |
1240 | * amdgpu_device_wb_get - Allocate a wb entry |
1241 | * |
1242 | * @adev: amdgpu_device pointer |
1243 | * @wb: wb index |
1244 | * |
1245 | * Allocate a wb slot for use by the driver (all asics). |
1246 | * Returns 0 on success or -EINVAL on failure. |
1247 | */ |
1248 | int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb) |
1249 | { |
1250 | unsigned long offset = find_first_zero_bit(addr: adev->wb.used, size: adev->wb.num_wb); |
1251 | |
1252 | if (offset < adev->wb.num_wb) { |
1253 | __set_bit(offset, adev->wb.used); |
1254 | *wb = offset << 3; /* convert to dw offset */ |
1255 | return 0; |
1256 | } else { |
1257 | return -EINVAL; |
1258 | } |
1259 | } |
1260 | |
1261 | /** |
1262 | * amdgpu_device_wb_free - Free a wb entry |
1263 | * |
1264 | * @adev: amdgpu_device pointer |
1265 | * @wb: wb index |
1266 | * |
1267 | * Free a wb slot allocated for use by the driver (all asics) |
1268 | */ |
1269 | void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb) |
1270 | { |
1271 | wb >>= 3; |
1272 | if (wb < adev->wb.num_wb) |
1273 | __clear_bit(wb, adev->wb.used); |
1274 | } |
1275 | |
1276 | /** |
1277 | * amdgpu_device_resize_fb_bar - try to resize FB BAR |
1278 | * |
1279 | * @adev: amdgpu_device pointer |
1280 | * |
1281 | * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not |
1282 | * to fail, but if any of the BARs is not accessible after the size we abort |
1283 | * driver loading by returning -ENODEV. |
1284 | */ |
1285 | int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev) |
1286 | { |
1287 | int rbar_size = pci_rebar_bytes_to_size(bytes: adev->gmc.real_vram_size); |
1288 | struct pci_bus *root; |
1289 | struct resource *res; |
1290 | unsigned int i; |
1291 | u16 cmd; |
1292 | int r; |
1293 | |
1294 | if (!IS_ENABLED(CONFIG_PHYS_ADDR_T_64BIT)) |
1295 | return 0; |
1296 | |
1297 | /* Bypass for VF */ |
1298 | if (amdgpu_sriov_vf(adev)) |
1299 | return 0; |
1300 | |
1301 | /* skip if the bios has already enabled large BAR */ |
1302 | if (adev->gmc.real_vram_size && |
1303 | (pci_resource_len(adev->pdev, 0) >= adev->gmc.real_vram_size)) |
1304 | return 0; |
1305 | |
1306 | /* Check if the root BUS has 64bit memory resources */ |
1307 | root = adev->pdev->bus; |
1308 | while (root->parent) |
1309 | root = root->parent; |
1310 | |
1311 | pci_bus_for_each_resource(root, res, i) { |
1312 | if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) && |
1313 | res->start > 0x100000000ull) |
1314 | break; |
1315 | } |
1316 | |
1317 | /* Trying to resize is pointless without a root hub window above 4GB */ |
1318 | if (!res) |
1319 | return 0; |
1320 | |
1321 | /* Limit the BAR size to what is available */ |
1322 | rbar_size = min(fls(pci_rebar_get_possible_sizes(adev->pdev, 0)) - 1, |
1323 | rbar_size); |
1324 | |
1325 | /* Disable memory decoding while we change the BAR addresses and size */ |
1326 | pci_read_config_word(dev: adev->pdev, PCI_COMMAND, val: &cmd); |
1327 | pci_write_config_word(dev: adev->pdev, PCI_COMMAND, |
1328 | val: cmd & ~PCI_COMMAND_MEMORY); |
1329 | |
1330 | /* Free the VRAM and doorbell BAR, we most likely need to move both. */ |
1331 | amdgpu_doorbell_fini(adev); |
1332 | if (adev->asic_type >= CHIP_BONAIRE) |
1333 | pci_release_resource(dev: adev->pdev, resno: 2); |
1334 | |
1335 | pci_release_resource(dev: adev->pdev, resno: 0); |
1336 | |
1337 | r = pci_resize_resource(dev: adev->pdev, i: 0, size: rbar_size); |
1338 | if (r == -ENOSPC) |
1339 | DRM_INFO("Not enough PCI address space for a large BAR." ); |
1340 | else if (r && r != -ENOTSUPP) |
1341 | DRM_ERROR("Problem resizing BAR0 (%d)." , r); |
1342 | |
1343 | pci_assign_unassigned_bus_resources(bus: adev->pdev->bus); |
1344 | |
1345 | /* When the doorbell or fb BAR isn't available we have no chance of |
1346 | * using the device. |
1347 | */ |
1348 | r = amdgpu_doorbell_init(adev); |
1349 | if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET)) |
1350 | return -ENODEV; |
1351 | |
1352 | pci_write_config_word(dev: adev->pdev, PCI_COMMAND, val: cmd); |
1353 | |
1354 | return 0; |
1355 | } |
1356 | |
1357 | static bool amdgpu_device_read_bios(struct amdgpu_device *adev) |
1358 | { |
1359 | if (hweight32(adev->aid_mask) && (adev->flags & AMD_IS_APU)) |
1360 | return false; |
1361 | |
1362 | return true; |
1363 | } |
1364 | |
1365 | /* |
1366 | * GPU helpers function. |
1367 | */ |
1368 | /** |
1369 | * amdgpu_device_need_post - check if the hw need post or not |
1370 | * |
1371 | * @adev: amdgpu_device pointer |
1372 | * |
1373 | * Check if the asic has been initialized (all asics) at driver startup |
1374 | * or post is needed if hw reset is performed. |
1375 | * Returns true if need or false if not. |
1376 | */ |
1377 | bool amdgpu_device_need_post(struct amdgpu_device *adev) |
1378 | { |
1379 | uint32_t reg; |
1380 | |
1381 | if (amdgpu_sriov_vf(adev)) |
1382 | return false; |
1383 | |
1384 | if (!amdgpu_device_read_bios(adev)) |
1385 | return false; |
1386 | |
1387 | if (amdgpu_passthrough(adev)) { |
1388 | /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot |
1389 | * some old smc fw still need driver do vPost otherwise gpu hang, while |
1390 | * those smc fw version above 22.15 doesn't have this flaw, so we force |
1391 | * vpost executed for smc version below 22.15 |
1392 | */ |
1393 | if (adev->asic_type == CHIP_FIJI) { |
1394 | int err; |
1395 | uint32_t fw_ver; |
1396 | |
1397 | err = request_firmware(fw: &adev->pm.fw, name: "amdgpu/fiji_smc.bin" , device: adev->dev); |
1398 | /* force vPost if error occured */ |
1399 | if (err) |
1400 | return true; |
1401 | |
1402 | fw_ver = *((uint32_t *)adev->pm.fw->data + 69); |
1403 | if (fw_ver < 0x00160e00) |
1404 | return true; |
1405 | } |
1406 | } |
1407 | |
1408 | /* Don't post if we need to reset whole hive on init */ |
1409 | if (adev->gmc.xgmi.pending_reset) |
1410 | return false; |
1411 | |
1412 | if (adev->has_hw_reset) { |
1413 | adev->has_hw_reset = false; |
1414 | return true; |
1415 | } |
1416 | |
1417 | /* bios scratch used on CIK+ */ |
1418 | if (adev->asic_type >= CHIP_BONAIRE) |
1419 | return amdgpu_atombios_scratch_need_asic_init(adev); |
1420 | |
1421 | /* check MEM_SIZE for older asics */ |
1422 | reg = amdgpu_asic_get_config_memsize(adev); |
1423 | |
1424 | if ((reg != 0) && (reg != 0xffffffff)) |
1425 | return false; |
1426 | |
1427 | return true; |
1428 | } |
1429 | |
1430 | /* |
1431 | * Check whether seamless boot is supported. |
1432 | * |
1433 | * So far we only support seamless boot on DCE 3.0 or later. |
1434 | * If users report that it works on older ASICS as well, we may |
1435 | * loosen this. |
1436 | */ |
1437 | bool amdgpu_device_seamless_boot_supported(struct amdgpu_device *adev) |
1438 | { |
1439 | switch (amdgpu_seamless) { |
1440 | case -1: |
1441 | break; |
1442 | case 1: |
1443 | return true; |
1444 | case 0: |
1445 | return false; |
1446 | default: |
1447 | DRM_ERROR("Invalid value for amdgpu.seamless: %d\n" , |
1448 | amdgpu_seamless); |
1449 | return false; |
1450 | } |
1451 | |
1452 | if (!(adev->flags & AMD_IS_APU)) |
1453 | return false; |
1454 | |
1455 | if (adev->mman.keep_stolen_vga_memory) |
1456 | return false; |
1457 | |
1458 | return adev->ip_versions[DCE_HWIP][0] >= IP_VERSION(3, 0, 0); |
1459 | } |
1460 | |
1461 | /* |
1462 | * Intel hosts such as Rocket Lake, Alder Lake, Raptor Lake and Sapphire Rapids |
1463 | * don't support dynamic speed switching. Until we have confirmation from Intel |
1464 | * that a specific host supports it, it's safer that we keep it disabled for all. |
1465 | * |
1466 | * 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/ |
1467 | * https://gitlab.freedesktop.org/drm/amd/-/issues/2663 |
1468 | */ |
1469 | static bool amdgpu_device_pcie_dynamic_switching_supported(void) |
1470 | { |
1471 | #if IS_ENABLED(CONFIG_X86) |
1472 | struct cpuinfo_x86 *c = &cpu_data(0); |
1473 | |
1474 | if (c->x86_vendor == X86_VENDOR_INTEL) |
1475 | return false; |
1476 | #endif |
1477 | return true; |
1478 | } |
1479 | |
1480 | /** |
1481 | * amdgpu_device_should_use_aspm - check if the device should program ASPM |
1482 | * |
1483 | * @adev: amdgpu_device pointer |
1484 | * |
1485 | * Confirm whether the module parameter and pcie bridge agree that ASPM should |
1486 | * be set for this device. |
1487 | * |
1488 | * Returns true if it should be used or false if not. |
1489 | */ |
1490 | bool amdgpu_device_should_use_aspm(struct amdgpu_device *adev) |
1491 | { |
1492 | switch (amdgpu_aspm) { |
1493 | case -1: |
1494 | break; |
1495 | case 0: |
1496 | return false; |
1497 | case 1: |
1498 | return true; |
1499 | default: |
1500 | return false; |
1501 | } |
1502 | if (adev->flags & AMD_IS_APU) |
1503 | return false; |
1504 | if (!(adev->pm.pp_feature & PP_PCIE_DPM_MASK)) |
1505 | return false; |
1506 | return pcie_aspm_enabled(pdev: adev->pdev); |
1507 | } |
1508 | |
1509 | /* if we get transitioned to only one device, take VGA back */ |
1510 | /** |
1511 | * amdgpu_device_vga_set_decode - enable/disable vga decode |
1512 | * |
1513 | * @pdev: PCI device pointer |
1514 | * @state: enable/disable vga decode |
1515 | * |
1516 | * Enable/disable vga decode (all asics). |
1517 | * Returns VGA resource flags. |
1518 | */ |
1519 | static unsigned int amdgpu_device_vga_set_decode(struct pci_dev *pdev, |
1520 | bool state) |
1521 | { |
1522 | struct amdgpu_device *adev = drm_to_adev(ddev: pci_get_drvdata(pdev)); |
1523 | |
1524 | amdgpu_asic_set_vga_state(adev, state); |
1525 | if (state) |
1526 | return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM | |
1527 | VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM; |
1528 | else |
1529 | return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM; |
1530 | } |
1531 | |
1532 | /** |
1533 | * amdgpu_device_check_block_size - validate the vm block size |
1534 | * |
1535 | * @adev: amdgpu_device pointer |
1536 | * |
1537 | * Validates the vm block size specified via module parameter. |
1538 | * The vm block size defines number of bits in page table versus page directory, |
1539 | * a page is 4KB so we have 12 bits offset, minimum 9 bits in the |
1540 | * page table and the remaining bits are in the page directory. |
1541 | */ |
1542 | static void amdgpu_device_check_block_size(struct amdgpu_device *adev) |
1543 | { |
1544 | /* defines number of bits in page table versus page directory, |
1545 | * a page is 4KB so we have 12 bits offset, minimum 9 bits in the |
1546 | * page table and the remaining bits are in the page directory |
1547 | */ |
1548 | if (amdgpu_vm_block_size == -1) |
1549 | return; |
1550 | |
1551 | if (amdgpu_vm_block_size < 9) { |
1552 | dev_warn(adev->dev, "VM page table size (%d) too small\n" , |
1553 | amdgpu_vm_block_size); |
1554 | amdgpu_vm_block_size = -1; |
1555 | } |
1556 | } |
1557 | |
1558 | /** |
1559 | * amdgpu_device_check_vm_size - validate the vm size |
1560 | * |
1561 | * @adev: amdgpu_device pointer |
1562 | * |
1563 | * Validates the vm size in GB specified via module parameter. |
1564 | * The VM size is the size of the GPU virtual memory space in GB. |
1565 | */ |
1566 | static void amdgpu_device_check_vm_size(struct amdgpu_device *adev) |
1567 | { |
1568 | /* no need to check the default value */ |
1569 | if (amdgpu_vm_size == -1) |
1570 | return; |
1571 | |
1572 | if (amdgpu_vm_size < 1) { |
1573 | dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n" , |
1574 | amdgpu_vm_size); |
1575 | amdgpu_vm_size = -1; |
1576 | } |
1577 | } |
1578 | |
1579 | static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev) |
1580 | { |
1581 | struct sysinfo si; |
1582 | bool is_os_64 = (sizeof(void *) == 8); |
1583 | uint64_t total_memory; |
1584 | uint64_t dram_size_seven_GB = 0x1B8000000; |
1585 | uint64_t dram_size_three_GB = 0xB8000000; |
1586 | |
1587 | if (amdgpu_smu_memory_pool_size == 0) |
1588 | return; |
1589 | |
1590 | if (!is_os_64) { |
1591 | DRM_WARN("Not 64-bit OS, feature not supported\n" ); |
1592 | goto def_value; |
1593 | } |
1594 | si_meminfo(val: &si); |
1595 | total_memory = (uint64_t)si.totalram * si.mem_unit; |
1596 | |
1597 | if ((amdgpu_smu_memory_pool_size == 1) || |
1598 | (amdgpu_smu_memory_pool_size == 2)) { |
1599 | if (total_memory < dram_size_three_GB) |
1600 | goto def_value1; |
1601 | } else if ((amdgpu_smu_memory_pool_size == 4) || |
1602 | (amdgpu_smu_memory_pool_size == 8)) { |
1603 | if (total_memory < dram_size_seven_GB) |
1604 | goto def_value1; |
1605 | } else { |
1606 | DRM_WARN("Smu memory pool size not supported\n" ); |
1607 | goto def_value; |
1608 | } |
1609 | adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28; |
1610 | |
1611 | return; |
1612 | |
1613 | def_value1: |
1614 | DRM_WARN("No enough system memory\n" ); |
1615 | def_value: |
1616 | adev->pm.smu_prv_buffer_size = 0; |
1617 | } |
1618 | |
1619 | static int amdgpu_device_init_apu_flags(struct amdgpu_device *adev) |
1620 | { |
1621 | if (!(adev->flags & AMD_IS_APU) || |
1622 | adev->asic_type < CHIP_RAVEN) |
1623 | return 0; |
1624 | |
1625 | switch (adev->asic_type) { |
1626 | case CHIP_RAVEN: |
1627 | if (adev->pdev->device == 0x15dd) |
1628 | adev->apu_flags |= AMD_APU_IS_RAVEN; |
1629 | if (adev->pdev->device == 0x15d8) |
1630 | adev->apu_flags |= AMD_APU_IS_PICASSO; |
1631 | break; |
1632 | case CHIP_RENOIR: |
1633 | if ((adev->pdev->device == 0x1636) || |
1634 | (adev->pdev->device == 0x164c)) |
1635 | adev->apu_flags |= AMD_APU_IS_RENOIR; |
1636 | else |
1637 | adev->apu_flags |= AMD_APU_IS_GREEN_SARDINE; |
1638 | break; |
1639 | case CHIP_VANGOGH: |
1640 | adev->apu_flags |= AMD_APU_IS_VANGOGH; |
1641 | break; |
1642 | case CHIP_YELLOW_CARP: |
1643 | break; |
1644 | case CHIP_CYAN_SKILLFISH: |
1645 | if ((adev->pdev->device == 0x13FE) || |
1646 | (adev->pdev->device == 0x143F)) |
1647 | adev->apu_flags |= AMD_APU_IS_CYAN_SKILLFISH2; |
1648 | break; |
1649 | default: |
1650 | break; |
1651 | } |
1652 | |
1653 | return 0; |
1654 | } |
1655 | |
1656 | /** |
1657 | * amdgpu_device_check_arguments - validate module params |
1658 | * |
1659 | * @adev: amdgpu_device pointer |
1660 | * |
1661 | * Validates certain module parameters and updates |
1662 | * the associated values used by the driver (all asics). |
1663 | */ |
1664 | static int amdgpu_device_check_arguments(struct amdgpu_device *adev) |
1665 | { |
1666 | if (amdgpu_sched_jobs < 4) { |
1667 | dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n" , |
1668 | amdgpu_sched_jobs); |
1669 | amdgpu_sched_jobs = 4; |
1670 | } else if (!is_power_of_2(n: amdgpu_sched_jobs)) { |
1671 | dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n" , |
1672 | amdgpu_sched_jobs); |
1673 | amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs); |
1674 | } |
1675 | |
1676 | if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) { |
1677 | /* gart size must be greater or equal to 32M */ |
1678 | dev_warn(adev->dev, "gart size (%d) too small\n" , |
1679 | amdgpu_gart_size); |
1680 | amdgpu_gart_size = -1; |
1681 | } |
1682 | |
1683 | if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) { |
1684 | /* gtt size must be greater or equal to 32M */ |
1685 | dev_warn(adev->dev, "gtt size (%d) too small\n" , |
1686 | amdgpu_gtt_size); |
1687 | amdgpu_gtt_size = -1; |
1688 | } |
1689 | |
1690 | /* valid range is between 4 and 9 inclusive */ |
1691 | if (amdgpu_vm_fragment_size != -1 && |
1692 | (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) { |
1693 | dev_warn(adev->dev, "valid range is between 4 and 9\n" ); |
1694 | amdgpu_vm_fragment_size = -1; |
1695 | } |
1696 | |
1697 | if (amdgpu_sched_hw_submission < 2) { |
1698 | dev_warn(adev->dev, "sched hw submission jobs (%d) must be at least 2\n" , |
1699 | amdgpu_sched_hw_submission); |
1700 | amdgpu_sched_hw_submission = 2; |
1701 | } else if (!is_power_of_2(n: amdgpu_sched_hw_submission)) { |
1702 | dev_warn(adev->dev, "sched hw submission jobs (%d) must be a power of 2\n" , |
1703 | amdgpu_sched_hw_submission); |
1704 | amdgpu_sched_hw_submission = roundup_pow_of_two(amdgpu_sched_hw_submission); |
1705 | } |
1706 | |
1707 | if (amdgpu_reset_method < -1 || amdgpu_reset_method > 4) { |
1708 | dev_warn(adev->dev, "invalid option for reset method, reverting to default\n" ); |
1709 | amdgpu_reset_method = -1; |
1710 | } |
1711 | |
1712 | amdgpu_device_check_smu_prv_buffer_size(adev); |
1713 | |
1714 | amdgpu_device_check_vm_size(adev); |
1715 | |
1716 | amdgpu_device_check_block_size(adev); |
1717 | |
1718 | adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, load_type: amdgpu_fw_load_type); |
1719 | |
1720 | return 0; |
1721 | } |
1722 | |
1723 | /** |
1724 | * amdgpu_switcheroo_set_state - set switcheroo state |
1725 | * |
1726 | * @pdev: pci dev pointer |
1727 | * @state: vga_switcheroo state |
1728 | * |
1729 | * Callback for the switcheroo driver. Suspends or resumes |
1730 | * the asics before or after it is powered up using ACPI methods. |
1731 | */ |
1732 | static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, |
1733 | enum vga_switcheroo_state state) |
1734 | { |
1735 | struct drm_device *dev = pci_get_drvdata(pdev); |
1736 | int r; |
1737 | |
1738 | if (amdgpu_device_supports_px(dev) && state == VGA_SWITCHEROO_OFF) |
1739 | return; |
1740 | |
1741 | if (state == VGA_SWITCHEROO_ON) { |
1742 | pr_info("switched on\n" ); |
1743 | /* don't suspend or resume card normally */ |
1744 | dev->switch_power_state = DRM_SWITCH_POWER_CHANGING; |
1745 | |
1746 | pci_set_power_state(dev: pdev, PCI_D0); |
1747 | amdgpu_device_load_pci_state(pdev); |
1748 | r = pci_enable_device(dev: pdev); |
1749 | if (r) |
1750 | DRM_WARN("pci_enable_device failed (%d)\n" , r); |
1751 | amdgpu_device_resume(dev, fbcon: true); |
1752 | |
1753 | dev->switch_power_state = DRM_SWITCH_POWER_ON; |
1754 | } else { |
1755 | pr_info("switched off\n" ); |
1756 | dev->switch_power_state = DRM_SWITCH_POWER_CHANGING; |
1757 | amdgpu_device_prepare(dev); |
1758 | amdgpu_device_suspend(dev, fbcon: true); |
1759 | amdgpu_device_cache_pci_state(pdev); |
1760 | /* Shut down the device */ |
1761 | pci_disable_device(dev: pdev); |
1762 | pci_set_power_state(dev: pdev, PCI_D3cold); |
1763 | dev->switch_power_state = DRM_SWITCH_POWER_OFF; |
1764 | } |
1765 | } |
1766 | |
1767 | /** |
1768 | * amdgpu_switcheroo_can_switch - see if switcheroo state can change |
1769 | * |
1770 | * @pdev: pci dev pointer |
1771 | * |
1772 | * Callback for the switcheroo driver. Check of the switcheroo |
1773 | * state can be changed. |
1774 | * Returns true if the state can be changed, false if not. |
1775 | */ |
1776 | static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev) |
1777 | { |
1778 | struct drm_device *dev = pci_get_drvdata(pdev); |
1779 | |
1780 | /* |
1781 | * FIXME: open_count is protected by drm_global_mutex but that would lead to |
1782 | * locking inversion with the driver load path. And the access here is |
1783 | * completely racy anyway. So don't bother with locking for now. |
1784 | */ |
1785 | return atomic_read(v: &dev->open_count) == 0; |
1786 | } |
1787 | |
1788 | static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = { |
1789 | .set_gpu_state = amdgpu_switcheroo_set_state, |
1790 | .reprobe = NULL, |
1791 | .can_switch = amdgpu_switcheroo_can_switch, |
1792 | }; |
1793 | |
1794 | /** |
1795 | * amdgpu_device_ip_set_clockgating_state - set the CG state |
1796 | * |
1797 | * @dev: amdgpu_device pointer |
1798 | * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) |
1799 | * @state: clockgating state (gate or ungate) |
1800 | * |
1801 | * Sets the requested clockgating state for all instances of |
1802 | * the hardware IP specified. |
1803 | * Returns the error code from the last instance. |
1804 | */ |
1805 | int amdgpu_device_ip_set_clockgating_state(void *dev, |
1806 | enum amd_ip_block_type block_type, |
1807 | enum amd_clockgating_state state) |
1808 | { |
1809 | struct amdgpu_device *adev = dev; |
1810 | int i, r = 0; |
1811 | |
1812 | for (i = 0; i < adev->num_ip_blocks; i++) { |
1813 | if (!adev->ip_blocks[i].status.valid) |
1814 | continue; |
1815 | if (adev->ip_blocks[i].version->type != block_type) |
1816 | continue; |
1817 | if (!adev->ip_blocks[i].version->funcs->set_clockgating_state) |
1818 | continue; |
1819 | r = adev->ip_blocks[i].version->funcs->set_clockgating_state( |
1820 | (void *)adev, state); |
1821 | if (r) |
1822 | DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n" , |
1823 | adev->ip_blocks[i].version->funcs->name, r); |
1824 | } |
1825 | return r; |
1826 | } |
1827 | |
1828 | /** |
1829 | * amdgpu_device_ip_set_powergating_state - set the PG state |
1830 | * |
1831 | * @dev: amdgpu_device pointer |
1832 | * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) |
1833 | * @state: powergating state (gate or ungate) |
1834 | * |
1835 | * Sets the requested powergating state for all instances of |
1836 | * the hardware IP specified. |
1837 | * Returns the error code from the last instance. |
1838 | */ |
1839 | int amdgpu_device_ip_set_powergating_state(void *dev, |
1840 | enum amd_ip_block_type block_type, |
1841 | enum amd_powergating_state state) |
1842 | { |
1843 | struct amdgpu_device *adev = dev; |
1844 | int i, r = 0; |
1845 | |
1846 | for (i = 0; i < adev->num_ip_blocks; i++) { |
1847 | if (!adev->ip_blocks[i].status.valid) |
1848 | continue; |
1849 | if (adev->ip_blocks[i].version->type != block_type) |
1850 | continue; |
1851 | if (!adev->ip_blocks[i].version->funcs->set_powergating_state) |
1852 | continue; |
1853 | r = adev->ip_blocks[i].version->funcs->set_powergating_state( |
1854 | (void *)adev, state); |
1855 | if (r) |
1856 | DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n" , |
1857 | adev->ip_blocks[i].version->funcs->name, r); |
1858 | } |
1859 | return r; |
1860 | } |
1861 | |
1862 | /** |
1863 | * amdgpu_device_ip_get_clockgating_state - get the CG state |
1864 | * |
1865 | * @adev: amdgpu_device pointer |
1866 | * @flags: clockgating feature flags |
1867 | * |
1868 | * Walks the list of IPs on the device and updates the clockgating |
1869 | * flags for each IP. |
1870 | * Updates @flags with the feature flags for each hardware IP where |
1871 | * clockgating is enabled. |
1872 | */ |
1873 | void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev, |
1874 | u64 *flags) |
1875 | { |
1876 | int i; |
1877 | |
1878 | for (i = 0; i < adev->num_ip_blocks; i++) { |
1879 | if (!adev->ip_blocks[i].status.valid) |
1880 | continue; |
1881 | if (adev->ip_blocks[i].version->funcs->get_clockgating_state) |
1882 | adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags); |
1883 | } |
1884 | } |
1885 | |
1886 | /** |
1887 | * amdgpu_device_ip_wait_for_idle - wait for idle |
1888 | * |
1889 | * @adev: amdgpu_device pointer |
1890 | * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) |
1891 | * |
1892 | * Waits for the request hardware IP to be idle. |
1893 | * Returns 0 for success or a negative error code on failure. |
1894 | */ |
1895 | int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev, |
1896 | enum amd_ip_block_type block_type) |
1897 | { |
1898 | int i, r; |
1899 | |
1900 | for (i = 0; i < adev->num_ip_blocks; i++) { |
1901 | if (!adev->ip_blocks[i].status.valid) |
1902 | continue; |
1903 | if (adev->ip_blocks[i].version->type == block_type) { |
1904 | r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev); |
1905 | if (r) |
1906 | return r; |
1907 | break; |
1908 | } |
1909 | } |
1910 | return 0; |
1911 | |
1912 | } |
1913 | |
1914 | /** |
1915 | * amdgpu_device_ip_is_idle - is the hardware IP idle |
1916 | * |
1917 | * @adev: amdgpu_device pointer |
1918 | * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) |
1919 | * |
1920 | * Check if the hardware IP is idle or not. |
1921 | * Returns true if it the IP is idle, false if not. |
1922 | */ |
1923 | bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev, |
1924 | enum amd_ip_block_type block_type) |
1925 | { |
1926 | int i; |
1927 | |
1928 | for (i = 0; i < adev->num_ip_blocks; i++) { |
1929 | if (!adev->ip_blocks[i].status.valid) |
1930 | continue; |
1931 | if (adev->ip_blocks[i].version->type == block_type) |
1932 | return adev->ip_blocks[i].version->funcs->is_idle((void *)adev); |
1933 | } |
1934 | return true; |
1935 | |
1936 | } |
1937 | |
1938 | /** |
1939 | * amdgpu_device_ip_get_ip_block - get a hw IP pointer |
1940 | * |
1941 | * @adev: amdgpu_device pointer |
1942 | * @type: Type of hardware IP (SMU, GFX, UVD, etc.) |
1943 | * |
1944 | * Returns a pointer to the hardware IP block structure |
1945 | * if it exists for the asic, otherwise NULL. |
1946 | */ |
1947 | struct amdgpu_ip_block * |
1948 | amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev, |
1949 | enum amd_ip_block_type type) |
1950 | { |
1951 | int i; |
1952 | |
1953 | for (i = 0; i < adev->num_ip_blocks; i++) |
1954 | if (adev->ip_blocks[i].version->type == type) |
1955 | return &adev->ip_blocks[i]; |
1956 | |
1957 | return NULL; |
1958 | } |
1959 | |
1960 | /** |
1961 | * amdgpu_device_ip_block_version_cmp |
1962 | * |
1963 | * @adev: amdgpu_device pointer |
1964 | * @type: enum amd_ip_block_type |
1965 | * @major: major version |
1966 | * @minor: minor version |
1967 | * |
1968 | * return 0 if equal or greater |
1969 | * return 1 if smaller or the ip_block doesn't exist |
1970 | */ |
1971 | int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev, |
1972 | enum amd_ip_block_type type, |
1973 | u32 major, u32 minor) |
1974 | { |
1975 | struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type); |
1976 | |
1977 | if (ip_block && ((ip_block->version->major > major) || |
1978 | ((ip_block->version->major == major) && |
1979 | (ip_block->version->minor >= minor)))) |
1980 | return 0; |
1981 | |
1982 | return 1; |
1983 | } |
1984 | |
1985 | /** |
1986 | * amdgpu_device_ip_block_add |
1987 | * |
1988 | * @adev: amdgpu_device pointer |
1989 | * @ip_block_version: pointer to the IP to add |
1990 | * |
1991 | * Adds the IP block driver information to the collection of IPs |
1992 | * on the asic. |
1993 | */ |
1994 | int amdgpu_device_ip_block_add(struct amdgpu_device *adev, |
1995 | const struct amdgpu_ip_block_version *ip_block_version) |
1996 | { |
1997 | if (!ip_block_version) |
1998 | return -EINVAL; |
1999 | |
2000 | switch (ip_block_version->type) { |
2001 | case AMD_IP_BLOCK_TYPE_VCN: |
2002 | if (adev->harvest_ip_mask & AMD_HARVEST_IP_VCN_MASK) |
2003 | return 0; |
2004 | break; |
2005 | case AMD_IP_BLOCK_TYPE_JPEG: |
2006 | if (adev->harvest_ip_mask & AMD_HARVEST_IP_JPEG_MASK) |
2007 | return 0; |
2008 | break; |
2009 | default: |
2010 | break; |
2011 | } |
2012 | |
2013 | DRM_INFO("add ip block number %d <%s>\n" , adev->num_ip_blocks, |
2014 | ip_block_version->funcs->name); |
2015 | |
2016 | adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version; |
2017 | |
2018 | return 0; |
2019 | } |
2020 | |
2021 | /** |
2022 | * amdgpu_device_enable_virtual_display - enable virtual display feature |
2023 | * |
2024 | * @adev: amdgpu_device pointer |
2025 | * |
2026 | * Enabled the virtual display feature if the user has enabled it via |
2027 | * the module parameter virtual_display. This feature provides a virtual |
2028 | * display hardware on headless boards or in virtualized environments. |
2029 | * This function parses and validates the configuration string specified by |
2030 | * the user and configues the virtual display configuration (number of |
2031 | * virtual connectors, crtcs, etc.) specified. |
2032 | */ |
2033 | static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev) |
2034 | { |
2035 | adev->enable_virtual_display = false; |
2036 | |
2037 | if (amdgpu_virtual_display) { |
2038 | const char *pci_address_name = pci_name(pdev: adev->pdev); |
2039 | char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname; |
2040 | |
2041 | pciaddstr = kstrdup(s: amdgpu_virtual_display, GFP_KERNEL); |
2042 | pciaddstr_tmp = pciaddstr; |
2043 | while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";" ))) { |
2044 | pciaddname = strsep(&pciaddname_tmp, "," ); |
2045 | if (!strcmp("all" , pciaddname) |
2046 | || !strcmp(pci_address_name, pciaddname)) { |
2047 | long num_crtc; |
2048 | int res = -1; |
2049 | |
2050 | adev->enable_virtual_display = true; |
2051 | |
2052 | if (pciaddname_tmp) |
2053 | res = kstrtol(s: pciaddname_tmp, base: 10, |
2054 | res: &num_crtc); |
2055 | |
2056 | if (!res) { |
2057 | if (num_crtc < 1) |
2058 | num_crtc = 1; |
2059 | if (num_crtc > 6) |
2060 | num_crtc = 6; |
2061 | adev->mode_info.num_crtc = num_crtc; |
2062 | } else { |
2063 | adev->mode_info.num_crtc = 1; |
2064 | } |
2065 | break; |
2066 | } |
2067 | } |
2068 | |
2069 | DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n" , |
2070 | amdgpu_virtual_display, pci_address_name, |
2071 | adev->enable_virtual_display, adev->mode_info.num_crtc); |
2072 | |
2073 | kfree(objp: pciaddstr); |
2074 | } |
2075 | } |
2076 | |
2077 | void amdgpu_device_set_sriov_virtual_display(struct amdgpu_device *adev) |
2078 | { |
2079 | if (amdgpu_sriov_vf(adev) && !adev->enable_virtual_display) { |
2080 | adev->mode_info.num_crtc = 1; |
2081 | adev->enable_virtual_display = true; |
2082 | DRM_INFO("virtual_display:%d, num_crtc:%d\n" , |
2083 | adev->enable_virtual_display, adev->mode_info.num_crtc); |
2084 | } |
2085 | } |
2086 | |
2087 | /** |
2088 | * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware |
2089 | * |
2090 | * @adev: amdgpu_device pointer |
2091 | * |
2092 | * Parses the asic configuration parameters specified in the gpu info |
2093 | * firmware and makes them availale to the driver for use in configuring |
2094 | * the asic. |
2095 | * Returns 0 on success, -EINVAL on failure. |
2096 | */ |
2097 | static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev) |
2098 | { |
2099 | const char *chip_name; |
2100 | char fw_name[40]; |
2101 | int err; |
2102 | const struct gpu_info_firmware_header_v1_0 *hdr; |
2103 | |
2104 | adev->firmware.gpu_info_fw = NULL; |
2105 | |
2106 | if (adev->mman.discovery_bin) { |
2107 | /* |
2108 | * FIXME: The bounding box is still needed by Navi12, so |
2109 | * temporarily read it from gpu_info firmware. Should be dropped |
2110 | * when DAL no longer needs it. |
2111 | */ |
2112 | if (adev->asic_type != CHIP_NAVI12) |
2113 | return 0; |
2114 | } |
2115 | |
2116 | switch (adev->asic_type) { |
2117 | default: |
2118 | return 0; |
2119 | case CHIP_VEGA10: |
2120 | chip_name = "vega10" ; |
2121 | break; |
2122 | case CHIP_VEGA12: |
2123 | chip_name = "vega12" ; |
2124 | break; |
2125 | case CHIP_RAVEN: |
2126 | if (adev->apu_flags & AMD_APU_IS_RAVEN2) |
2127 | chip_name = "raven2" ; |
2128 | else if (adev->apu_flags & AMD_APU_IS_PICASSO) |
2129 | chip_name = "picasso" ; |
2130 | else |
2131 | chip_name = "raven" ; |
2132 | break; |
2133 | case CHIP_ARCTURUS: |
2134 | chip_name = "arcturus" ; |
2135 | break; |
2136 | case CHIP_NAVI12: |
2137 | chip_name = "navi12" ; |
2138 | break; |
2139 | } |
2140 | |
2141 | snprintf(buf: fw_name, size: sizeof(fw_name), fmt: "amdgpu/%s_gpu_info.bin" , chip_name); |
2142 | err = amdgpu_ucode_request(adev, fw: &adev->firmware.gpu_info_fw, fw_name); |
2143 | if (err) { |
2144 | dev_err(adev->dev, |
2145 | "Failed to get gpu_info firmware \"%s\"\n" , |
2146 | fw_name); |
2147 | goto out; |
2148 | } |
2149 | |
2150 | hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data; |
2151 | amdgpu_ucode_print_gpu_info_hdr(hdr: &hdr->header); |
2152 | |
2153 | switch (hdr->version_major) { |
2154 | case 1: |
2155 | { |
2156 | const struct gpu_info_firmware_v1_0 *gpu_info_fw = |
2157 | (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data + |
2158 | le32_to_cpu(hdr->header.ucode_array_offset_bytes)); |
2159 | |
2160 | /* |
2161 | * Should be droped when DAL no longer needs it. |
2162 | */ |
2163 | if (adev->asic_type == CHIP_NAVI12) |
2164 | goto parse_soc_bounding_box; |
2165 | |
2166 | adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se); |
2167 | adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh); |
2168 | adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se); |
2169 | adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se); |
2170 | adev->gfx.config.max_texture_channel_caches = |
2171 | le32_to_cpu(gpu_info_fw->gc_num_tccs); |
2172 | adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs); |
2173 | adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds); |
2174 | adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth); |
2175 | adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth); |
2176 | adev->gfx.config.double_offchip_lds_buf = |
2177 | le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer); |
2178 | adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size); |
2179 | adev->gfx.cu_info.max_waves_per_simd = |
2180 | le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd); |
2181 | adev->gfx.cu_info.max_scratch_slots_per_cu = |
2182 | le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu); |
2183 | adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size); |
2184 | if (hdr->version_minor >= 1) { |
2185 | const struct gpu_info_firmware_v1_1 *gpu_info_fw = |
2186 | (const struct gpu_info_firmware_v1_1 *)(adev->firmware.gpu_info_fw->data + |
2187 | le32_to_cpu(hdr->header.ucode_array_offset_bytes)); |
2188 | adev->gfx.config.num_sc_per_sh = |
2189 | le32_to_cpu(gpu_info_fw->num_sc_per_sh); |
2190 | adev->gfx.config.num_packer_per_sc = |
2191 | le32_to_cpu(gpu_info_fw->num_packer_per_sc); |
2192 | } |
2193 | |
2194 | parse_soc_bounding_box: |
2195 | /* |
2196 | * soc bounding box info is not integrated in disocovery table, |
2197 | * we always need to parse it from gpu info firmware if needed. |
2198 | */ |
2199 | if (hdr->version_minor == 2) { |
2200 | const struct gpu_info_firmware_v1_2 *gpu_info_fw = |
2201 | (const struct gpu_info_firmware_v1_2 *)(adev->firmware.gpu_info_fw->data + |
2202 | le32_to_cpu(hdr->header.ucode_array_offset_bytes)); |
2203 | adev->dm.soc_bounding_box = &gpu_info_fw->soc_bounding_box; |
2204 | } |
2205 | break; |
2206 | } |
2207 | default: |
2208 | dev_err(adev->dev, |
2209 | "Unsupported gpu_info table %d\n" , hdr->header.ucode_version); |
2210 | err = -EINVAL; |
2211 | goto out; |
2212 | } |
2213 | out: |
2214 | return err; |
2215 | } |
2216 | |
2217 | /** |
2218 | * amdgpu_device_ip_early_init - run early init for hardware IPs |
2219 | * |
2220 | * @adev: amdgpu_device pointer |
2221 | * |
2222 | * Early initialization pass for hardware IPs. The hardware IPs that make |
2223 | * up each asic are discovered each IP's early_init callback is run. This |
2224 | * is the first stage in initializing the asic. |
2225 | * Returns 0 on success, negative error code on failure. |
2226 | */ |
2227 | static int amdgpu_device_ip_early_init(struct amdgpu_device *adev) |
2228 | { |
2229 | struct pci_dev *parent; |
2230 | int i, r; |
2231 | bool total; |
2232 | |
2233 | amdgpu_device_enable_virtual_display(adev); |
2234 | |
2235 | if (amdgpu_sriov_vf(adev)) { |
2236 | r = amdgpu_virt_request_full_gpu(adev, init: true); |
2237 | if (r) |
2238 | return r; |
2239 | } |
2240 | |
2241 | switch (adev->asic_type) { |
2242 | #ifdef CONFIG_DRM_AMDGPU_SI |
2243 | case CHIP_VERDE: |
2244 | case CHIP_TAHITI: |
2245 | case CHIP_PITCAIRN: |
2246 | case CHIP_OLAND: |
2247 | case CHIP_HAINAN: |
2248 | adev->family = AMDGPU_FAMILY_SI; |
2249 | r = si_set_ip_blocks(adev); |
2250 | if (r) |
2251 | return r; |
2252 | break; |
2253 | #endif |
2254 | #ifdef CONFIG_DRM_AMDGPU_CIK |
2255 | case CHIP_BONAIRE: |
2256 | case CHIP_HAWAII: |
2257 | case CHIP_KAVERI: |
2258 | case CHIP_KABINI: |
2259 | case CHIP_MULLINS: |
2260 | if (adev->flags & AMD_IS_APU) |
2261 | adev->family = AMDGPU_FAMILY_KV; |
2262 | else |
2263 | adev->family = AMDGPU_FAMILY_CI; |
2264 | |
2265 | r = cik_set_ip_blocks(adev); |
2266 | if (r) |
2267 | return r; |
2268 | break; |
2269 | #endif |
2270 | case CHIP_TOPAZ: |
2271 | case CHIP_TONGA: |
2272 | case CHIP_FIJI: |
2273 | case CHIP_POLARIS10: |
2274 | case CHIP_POLARIS11: |
2275 | case CHIP_POLARIS12: |
2276 | case CHIP_VEGAM: |
2277 | case CHIP_CARRIZO: |
2278 | case CHIP_STONEY: |
2279 | if (adev->flags & AMD_IS_APU) |
2280 | adev->family = AMDGPU_FAMILY_CZ; |
2281 | else |
2282 | adev->family = AMDGPU_FAMILY_VI; |
2283 | |
2284 | r = vi_set_ip_blocks(adev); |
2285 | if (r) |
2286 | return r; |
2287 | break; |
2288 | default: |
2289 | r = amdgpu_discovery_set_ip_blocks(adev); |
2290 | if (r) |
2291 | return r; |
2292 | break; |
2293 | } |
2294 | |
2295 | if (amdgpu_has_atpx() && |
2296 | (amdgpu_is_atpx_hybrid() || |
2297 | amdgpu_has_atpx_dgpu_power_cntl()) && |
2298 | ((adev->flags & AMD_IS_APU) == 0) && |
2299 | !dev_is_removable(dev: &adev->pdev->dev)) |
2300 | adev->flags |= AMD_IS_PX; |
2301 | |
2302 | if (!(adev->flags & AMD_IS_APU)) { |
2303 | parent = pcie_find_root_port(dev: adev->pdev); |
2304 | adev->has_pr3 = parent ? pci_pr3_present(pdev: parent) : false; |
2305 | } |
2306 | |
2307 | |
2308 | adev->pm.pp_feature = amdgpu_pp_feature_mask; |
2309 | if (amdgpu_sriov_vf(adev) || sched_policy == KFD_SCHED_POLICY_NO_HWS) |
2310 | adev->pm.pp_feature &= ~PP_GFXOFF_MASK; |
2311 | if (amdgpu_sriov_vf(adev) && adev->asic_type == CHIP_SIENNA_CICHLID) |
2312 | adev->pm.pp_feature &= ~PP_OVERDRIVE_MASK; |
2313 | if (!amdgpu_device_pcie_dynamic_switching_supported()) |
2314 | adev->pm.pp_feature &= ~PP_PCIE_DPM_MASK; |
2315 | |
2316 | total = true; |
2317 | for (i = 0; i < adev->num_ip_blocks; i++) { |
2318 | if ((amdgpu_ip_block_mask & (1 << i)) == 0) { |
2319 | DRM_WARN("disabled ip block: %d <%s>\n" , |
2320 | i, adev->ip_blocks[i].version->funcs->name); |
2321 | adev->ip_blocks[i].status.valid = false; |
2322 | } else { |
2323 | if (adev->ip_blocks[i].version->funcs->early_init) { |
2324 | r = adev->ip_blocks[i].version->funcs->early_init((void *)adev); |
2325 | if (r == -ENOENT) { |
2326 | adev->ip_blocks[i].status.valid = false; |
2327 | } else if (r) { |
2328 | DRM_ERROR("early_init of IP block <%s> failed %d\n" , |
2329 | adev->ip_blocks[i].version->funcs->name, r); |
2330 | total = false; |
2331 | } else { |
2332 | adev->ip_blocks[i].status.valid = true; |
2333 | } |
2334 | } else { |
2335 | adev->ip_blocks[i].status.valid = true; |
2336 | } |
2337 | } |
2338 | /* get the vbios after the asic_funcs are set up */ |
2339 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) { |
2340 | r = amdgpu_device_parse_gpu_info_fw(adev); |
2341 | if (r) |
2342 | return r; |
2343 | |
2344 | /* Read BIOS */ |
2345 | if (amdgpu_device_read_bios(adev)) { |
2346 | if (!amdgpu_get_bios(adev)) |
2347 | return -EINVAL; |
2348 | |
2349 | r = amdgpu_atombios_init(adev); |
2350 | if (r) { |
2351 | dev_err(adev->dev, "amdgpu_atombios_init failed\n" ); |
2352 | amdgpu_vf_error_put(adev, sub_error_code: AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, error_flags: 0, error_data: 0); |
2353 | return r; |
2354 | } |
2355 | } |
2356 | |
2357 | /*get pf2vf msg info at it's earliest time*/ |
2358 | if (amdgpu_sriov_vf(adev)) |
2359 | amdgpu_virt_init_data_exchange(adev); |
2360 | |
2361 | } |
2362 | } |
2363 | if (!total) |
2364 | return -ENODEV; |
2365 | |
2366 | amdgpu_amdkfd_device_probe(adev); |
2367 | adev->cg_flags &= amdgpu_cg_mask; |
2368 | adev->pg_flags &= amdgpu_pg_mask; |
2369 | |
2370 | return 0; |
2371 | } |
2372 | |
2373 | static int amdgpu_device_ip_hw_init_phase1(struct amdgpu_device *adev) |
2374 | { |
2375 | int i, r; |
2376 | |
2377 | for (i = 0; i < adev->num_ip_blocks; i++) { |
2378 | if (!adev->ip_blocks[i].status.sw) |
2379 | continue; |
2380 | if (adev->ip_blocks[i].status.hw) |
2381 | continue; |
2382 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || |
2383 | (amdgpu_sriov_vf(adev) && (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)) || |
2384 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) { |
2385 | r = adev->ip_blocks[i].version->funcs->hw_init(adev); |
2386 | if (r) { |
2387 | DRM_ERROR("hw_init of IP block <%s> failed %d\n" , |
2388 | adev->ip_blocks[i].version->funcs->name, r); |
2389 | return r; |
2390 | } |
2391 | adev->ip_blocks[i].status.hw = true; |
2392 | } |
2393 | } |
2394 | |
2395 | return 0; |
2396 | } |
2397 | |
2398 | static int amdgpu_device_ip_hw_init_phase2(struct amdgpu_device *adev) |
2399 | { |
2400 | int i, r; |
2401 | |
2402 | for (i = 0; i < adev->num_ip_blocks; i++) { |
2403 | if (!adev->ip_blocks[i].status.sw) |
2404 | continue; |
2405 | if (adev->ip_blocks[i].status.hw) |
2406 | continue; |
2407 | r = adev->ip_blocks[i].version->funcs->hw_init(adev); |
2408 | if (r) { |
2409 | DRM_ERROR("hw_init of IP block <%s> failed %d\n" , |
2410 | adev->ip_blocks[i].version->funcs->name, r); |
2411 | return r; |
2412 | } |
2413 | adev->ip_blocks[i].status.hw = true; |
2414 | } |
2415 | |
2416 | return 0; |
2417 | } |
2418 | |
2419 | static int amdgpu_device_fw_loading(struct amdgpu_device *adev) |
2420 | { |
2421 | int r = 0; |
2422 | int i; |
2423 | uint32_t smu_version; |
2424 | |
2425 | if (adev->asic_type >= CHIP_VEGA10) { |
2426 | for (i = 0; i < adev->num_ip_blocks; i++) { |
2427 | if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_PSP) |
2428 | continue; |
2429 | |
2430 | if (!adev->ip_blocks[i].status.sw) |
2431 | continue; |
2432 | |
2433 | /* no need to do the fw loading again if already done*/ |
2434 | if (adev->ip_blocks[i].status.hw == true) |
2435 | break; |
2436 | |
2437 | if (amdgpu_in_reset(adev) || adev->in_suspend) { |
2438 | r = adev->ip_blocks[i].version->funcs->resume(adev); |
2439 | if (r) { |
2440 | DRM_ERROR("resume of IP block <%s> failed %d\n" , |
2441 | adev->ip_blocks[i].version->funcs->name, r); |
2442 | return r; |
2443 | } |
2444 | } else { |
2445 | r = adev->ip_blocks[i].version->funcs->hw_init(adev); |
2446 | if (r) { |
2447 | DRM_ERROR("hw_init of IP block <%s> failed %d\n" , |
2448 | adev->ip_blocks[i].version->funcs->name, r); |
2449 | return r; |
2450 | } |
2451 | } |
2452 | |
2453 | adev->ip_blocks[i].status.hw = true; |
2454 | break; |
2455 | } |
2456 | } |
2457 | |
2458 | if (!amdgpu_sriov_vf(adev) || adev->asic_type == CHIP_TONGA) |
2459 | r = amdgpu_pm_load_smu_firmware(adev, smu_version: &smu_version); |
2460 | |
2461 | return r; |
2462 | } |
2463 | |
2464 | static int amdgpu_device_init_schedulers(struct amdgpu_device *adev) |
2465 | { |
2466 | long timeout; |
2467 | int r, i; |
2468 | |
2469 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { |
2470 | struct amdgpu_ring *ring = adev->rings[i]; |
2471 | |
2472 | /* No need to setup the GPU scheduler for rings that don't need it */ |
2473 | if (!ring || ring->no_scheduler) |
2474 | continue; |
2475 | |
2476 | switch (ring->funcs->type) { |
2477 | case AMDGPU_RING_TYPE_GFX: |
2478 | timeout = adev->gfx_timeout; |
2479 | break; |
2480 | case AMDGPU_RING_TYPE_COMPUTE: |
2481 | timeout = adev->compute_timeout; |
2482 | break; |
2483 | case AMDGPU_RING_TYPE_SDMA: |
2484 | timeout = adev->sdma_timeout; |
2485 | break; |
2486 | default: |
2487 | timeout = adev->video_timeout; |
2488 | break; |
2489 | } |
2490 | |
2491 | r = drm_sched_init(sched: &ring->sched, ops: &amdgpu_sched_ops, |
2492 | num_rqs: DRM_SCHED_PRIORITY_COUNT, |
2493 | hw_submission: ring->num_hw_submission, hang_limit: 0, |
2494 | timeout, timeout_wq: adev->reset_domain->wq, |
2495 | score: ring->sched_score, name: ring->name, |
2496 | dev: adev->dev); |
2497 | if (r) { |
2498 | DRM_ERROR("Failed to create scheduler on ring %s.\n" , |
2499 | ring->name); |
2500 | return r; |
2501 | } |
2502 | } |
2503 | |
2504 | amdgpu_xcp_update_partition_sched_list(adev); |
2505 | |
2506 | return 0; |
2507 | } |
2508 | |
2509 | |
2510 | /** |
2511 | * amdgpu_device_ip_init - run init for hardware IPs |
2512 | * |
2513 | * @adev: amdgpu_device pointer |
2514 | * |
2515 | * Main initialization pass for hardware IPs. The list of all the hardware |
2516 | * IPs that make up the asic is walked and the sw_init and hw_init callbacks |
2517 | * are run. sw_init initializes the software state associated with each IP |
2518 | * and hw_init initializes the hardware associated with each IP. |
2519 | * Returns 0 on success, negative error code on failure. |
2520 | */ |
2521 | static int amdgpu_device_ip_init(struct amdgpu_device *adev) |
2522 | { |
2523 | int i, r; |
2524 | |
2525 | r = amdgpu_ras_init(adev); |
2526 | if (r) |
2527 | return r; |
2528 | |
2529 | for (i = 0; i < adev->num_ip_blocks; i++) { |
2530 | if (!adev->ip_blocks[i].status.valid) |
2531 | continue; |
2532 | r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev); |
2533 | if (r) { |
2534 | DRM_ERROR("sw_init of IP block <%s> failed %d\n" , |
2535 | adev->ip_blocks[i].version->funcs->name, r); |
2536 | goto init_failed; |
2537 | } |
2538 | adev->ip_blocks[i].status.sw = true; |
2539 | |
2540 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) { |
2541 | /* need to do common hw init early so everything is set up for gmc */ |
2542 | r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev); |
2543 | if (r) { |
2544 | DRM_ERROR("hw_init %d failed %d\n" , i, r); |
2545 | goto init_failed; |
2546 | } |
2547 | adev->ip_blocks[i].status.hw = true; |
2548 | } else if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) { |
2549 | /* need to do gmc hw init early so we can allocate gpu mem */ |
2550 | /* Try to reserve bad pages early */ |
2551 | if (amdgpu_sriov_vf(adev)) |
2552 | amdgpu_virt_exchange_data(adev); |
2553 | |
2554 | r = amdgpu_device_mem_scratch_init(adev); |
2555 | if (r) { |
2556 | DRM_ERROR("amdgpu_mem_scratch_init failed %d\n" , r); |
2557 | goto init_failed; |
2558 | } |
2559 | r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev); |
2560 | if (r) { |
2561 | DRM_ERROR("hw_init %d failed %d\n" , i, r); |
2562 | goto init_failed; |
2563 | } |
2564 | r = amdgpu_device_wb_init(adev); |
2565 | if (r) { |
2566 | DRM_ERROR("amdgpu_device_wb_init failed %d\n" , r); |
2567 | goto init_failed; |
2568 | } |
2569 | adev->ip_blocks[i].status.hw = true; |
2570 | |
2571 | /* right after GMC hw init, we create CSA */ |
2572 | if (adev->gfx.mcbp) { |
2573 | r = amdgpu_allocate_static_csa(adev, bo: &adev->virt.csa_obj, |
2574 | AMDGPU_GEM_DOMAIN_VRAM | |
2575 | AMDGPU_GEM_DOMAIN_GTT, |
2576 | AMDGPU_CSA_SIZE); |
2577 | if (r) { |
2578 | DRM_ERROR("allocate CSA failed %d\n" , r); |
2579 | goto init_failed; |
2580 | } |
2581 | } |
2582 | } |
2583 | } |
2584 | |
2585 | if (amdgpu_sriov_vf(adev)) |
2586 | amdgpu_virt_init_data_exchange(adev); |
2587 | |
2588 | r = amdgpu_ib_pool_init(adev); |
2589 | if (r) { |
2590 | dev_err(adev->dev, "IB initialization failed (%d).\n" , r); |
2591 | amdgpu_vf_error_put(adev, sub_error_code: AMDGIM_ERROR_VF_IB_INIT_FAIL, error_flags: 0, error_data: r); |
2592 | goto init_failed; |
2593 | } |
2594 | |
2595 | r = amdgpu_ucode_create_bo(adev); /* create ucode bo when sw_init complete*/ |
2596 | if (r) |
2597 | goto init_failed; |
2598 | |
2599 | r = amdgpu_device_ip_hw_init_phase1(adev); |
2600 | if (r) |
2601 | goto init_failed; |
2602 | |
2603 | r = amdgpu_device_fw_loading(adev); |
2604 | if (r) |
2605 | goto init_failed; |
2606 | |
2607 | r = amdgpu_device_ip_hw_init_phase2(adev); |
2608 | if (r) |
2609 | goto init_failed; |
2610 | |
2611 | /* |
2612 | * retired pages will be loaded from eeprom and reserved here, |
2613 | * it should be called after amdgpu_device_ip_hw_init_phase2 since |
2614 | * for some ASICs the RAS EEPROM code relies on SMU fully functioning |
2615 | * for I2C communication which only true at this point. |
2616 | * |
2617 | * amdgpu_ras_recovery_init may fail, but the upper only cares the |
2618 | * failure from bad gpu situation and stop amdgpu init process |
2619 | * accordingly. For other failed cases, it will still release all |
2620 | * the resource and print error message, rather than returning one |
2621 | * negative value to upper level. |
2622 | * |
2623 | * Note: theoretically, this should be called before all vram allocations |
2624 | * to protect retired page from abusing |
2625 | */ |
2626 | r = amdgpu_ras_recovery_init(adev); |
2627 | if (r) |
2628 | goto init_failed; |
2629 | |
2630 | /** |
2631 | * In case of XGMI grab extra reference for reset domain for this device |
2632 | */ |
2633 | if (adev->gmc.xgmi.num_physical_nodes > 1) { |
2634 | if (amdgpu_xgmi_add_device(adev) == 0) { |
2635 | if (!amdgpu_sriov_vf(adev)) { |
2636 | struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev); |
2637 | |
2638 | if (WARN_ON(!hive)) { |
2639 | r = -ENOENT; |
2640 | goto init_failed; |
2641 | } |
2642 | |
2643 | if (!hive->reset_domain || |
2644 | !amdgpu_reset_get_reset_domain(domain: hive->reset_domain)) { |
2645 | r = -ENOENT; |
2646 | amdgpu_put_xgmi_hive(hive); |
2647 | goto init_failed; |
2648 | } |
2649 | |
2650 | /* Drop the early temporary reset domain we created for device */ |
2651 | amdgpu_reset_put_reset_domain(domain: adev->reset_domain); |
2652 | adev->reset_domain = hive->reset_domain; |
2653 | amdgpu_put_xgmi_hive(hive); |
2654 | } |
2655 | } |
2656 | } |
2657 | |
2658 | r = amdgpu_device_init_schedulers(adev); |
2659 | if (r) |
2660 | goto init_failed; |
2661 | |
2662 | if (adev->mman.buffer_funcs_ring->sched.ready) |
2663 | amdgpu_ttm_set_buffer_funcs_status(adev, enable: true); |
2664 | |
2665 | /* Don't init kfd if whole hive need to be reset during init */ |
2666 | if (!adev->gmc.xgmi.pending_reset) { |
2667 | kgd2kfd_init_zone_device(adev); |
2668 | amdgpu_amdkfd_device_init(adev); |
2669 | } |
2670 | |
2671 | amdgpu_fru_get_product_info(adev); |
2672 | |
2673 | init_failed: |
2674 | |
2675 | return r; |
2676 | } |
2677 | |
2678 | /** |
2679 | * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer |
2680 | * |
2681 | * @adev: amdgpu_device pointer |
2682 | * |
2683 | * Writes a reset magic value to the gart pointer in VRAM. The driver calls |
2684 | * this function before a GPU reset. If the value is retained after a |
2685 | * GPU reset, VRAM has not been lost. Some GPU resets may destry VRAM contents. |
2686 | */ |
2687 | static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev) |
2688 | { |
2689 | memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM); |
2690 | } |
2691 | |
2692 | /** |
2693 | * amdgpu_device_check_vram_lost - check if vram is valid |
2694 | * |
2695 | * @adev: amdgpu_device pointer |
2696 | * |
2697 | * Checks the reset magic value written to the gart pointer in VRAM. |
2698 | * The driver calls this after a GPU reset to see if the contents of |
2699 | * VRAM is lost or now. |
2700 | * returns true if vram is lost, false if not. |
2701 | */ |
2702 | static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev) |
2703 | { |
2704 | if (memcmp(p: adev->gart.ptr, q: adev->reset_magic, |
2705 | AMDGPU_RESET_MAGIC_NUM)) |
2706 | return true; |
2707 | |
2708 | if (!amdgpu_in_reset(adev)) |
2709 | return false; |
2710 | |
2711 | /* |
2712 | * For all ASICs with baco/mode1 reset, the VRAM is |
2713 | * always assumed to be lost. |
2714 | */ |
2715 | switch (amdgpu_asic_reset_method(adev)) { |
2716 | case AMD_RESET_METHOD_BACO: |
2717 | case AMD_RESET_METHOD_MODE1: |
2718 | return true; |
2719 | default: |
2720 | return false; |
2721 | } |
2722 | } |
2723 | |
2724 | /** |
2725 | * amdgpu_device_set_cg_state - set clockgating for amdgpu device |
2726 | * |
2727 | * @adev: amdgpu_device pointer |
2728 | * @state: clockgating state (gate or ungate) |
2729 | * |
2730 | * The list of all the hardware IPs that make up the asic is walked and the |
2731 | * set_clockgating_state callbacks are run. |
2732 | * Late initialization pass enabling clockgating for hardware IPs. |
2733 | * Fini or suspend, pass disabling clockgating for hardware IPs. |
2734 | * Returns 0 on success, negative error code on failure. |
2735 | */ |
2736 | |
2737 | int amdgpu_device_set_cg_state(struct amdgpu_device *adev, |
2738 | enum amd_clockgating_state state) |
2739 | { |
2740 | int i, j, r; |
2741 | |
2742 | if (amdgpu_emu_mode == 1) |
2743 | return 0; |
2744 | |
2745 | for (j = 0; j < adev->num_ip_blocks; j++) { |
2746 | i = state == AMD_CG_STATE_GATE ? j : adev->num_ip_blocks - j - 1; |
2747 | if (!adev->ip_blocks[i].status.late_initialized) |
2748 | continue; |
2749 | /* skip CG for GFX, SDMA on S0ix */ |
2750 | if (adev->in_s0ix && |
2751 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX || |
2752 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SDMA)) |
2753 | continue; |
2754 | /* skip CG for VCE/UVD, it's handled specially */ |
2755 | if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD && |
2756 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE && |
2757 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN && |
2758 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG && |
2759 | adev->ip_blocks[i].version->funcs->set_clockgating_state) { |
2760 | /* enable clockgating to save power */ |
2761 | r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev, |
2762 | state); |
2763 | if (r) { |
2764 | DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n" , |
2765 | adev->ip_blocks[i].version->funcs->name, r); |
2766 | return r; |
2767 | } |
2768 | } |
2769 | } |
2770 | |
2771 | return 0; |
2772 | } |
2773 | |
2774 | int amdgpu_device_set_pg_state(struct amdgpu_device *adev, |
2775 | enum amd_powergating_state state) |
2776 | { |
2777 | int i, j, r; |
2778 | |
2779 | if (amdgpu_emu_mode == 1) |
2780 | return 0; |
2781 | |
2782 | for (j = 0; j < adev->num_ip_blocks; j++) { |
2783 | i = state == AMD_PG_STATE_GATE ? j : adev->num_ip_blocks - j - 1; |
2784 | if (!adev->ip_blocks[i].status.late_initialized) |
2785 | continue; |
2786 | /* skip PG for GFX, SDMA on S0ix */ |
2787 | if (adev->in_s0ix && |
2788 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX || |
2789 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SDMA)) |
2790 | continue; |
2791 | /* skip CG for VCE/UVD, it's handled specially */ |
2792 | if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD && |
2793 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE && |
2794 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN && |
2795 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG && |
2796 | adev->ip_blocks[i].version->funcs->set_powergating_state) { |
2797 | /* enable powergating to save power */ |
2798 | r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev, |
2799 | state); |
2800 | if (r) { |
2801 | DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n" , |
2802 | adev->ip_blocks[i].version->funcs->name, r); |
2803 | return r; |
2804 | } |
2805 | } |
2806 | } |
2807 | return 0; |
2808 | } |
2809 | |
2810 | static int amdgpu_device_enable_mgpu_fan_boost(void) |
2811 | { |
2812 | struct amdgpu_gpu_instance *gpu_ins; |
2813 | struct amdgpu_device *adev; |
2814 | int i, ret = 0; |
2815 | |
2816 | mutex_lock(&mgpu_info.mutex); |
2817 | |
2818 | /* |
2819 | * MGPU fan boost feature should be enabled |
2820 | * only when there are two or more dGPUs in |
2821 | * the system |
2822 | */ |
2823 | if (mgpu_info.num_dgpu < 2) |
2824 | goto out; |
2825 | |
2826 | for (i = 0; i < mgpu_info.num_dgpu; i++) { |
2827 | gpu_ins = &(mgpu_info.gpu_ins[i]); |
2828 | adev = gpu_ins->adev; |
2829 | if (!(adev->flags & AMD_IS_APU) && |
2830 | !gpu_ins->mgpu_fan_enabled) { |
2831 | ret = amdgpu_dpm_enable_mgpu_fan_boost(adev); |
2832 | if (ret) |
2833 | break; |
2834 | |
2835 | gpu_ins->mgpu_fan_enabled = 1; |
2836 | } |
2837 | } |
2838 | |
2839 | out: |
2840 | mutex_unlock(lock: &mgpu_info.mutex); |
2841 | |
2842 | return ret; |
2843 | } |
2844 | |
2845 | /** |
2846 | * amdgpu_device_ip_late_init - run late init for hardware IPs |
2847 | * |
2848 | * @adev: amdgpu_device pointer |
2849 | * |
2850 | * Late initialization pass for hardware IPs. The list of all the hardware |
2851 | * IPs that make up the asic is walked and the late_init callbacks are run. |
2852 | * late_init covers any special initialization that an IP requires |
2853 | * after all of the have been initialized or something that needs to happen |
2854 | * late in the init process. |
2855 | * Returns 0 on success, negative error code on failure. |
2856 | */ |
2857 | static int amdgpu_device_ip_late_init(struct amdgpu_device *adev) |
2858 | { |
2859 | struct amdgpu_gpu_instance *gpu_instance; |
2860 | int i = 0, r; |
2861 | |
2862 | for (i = 0; i < adev->num_ip_blocks; i++) { |
2863 | if (!adev->ip_blocks[i].status.hw) |
2864 | continue; |
2865 | if (adev->ip_blocks[i].version->funcs->late_init) { |
2866 | r = adev->ip_blocks[i].version->funcs->late_init((void *)adev); |
2867 | if (r) { |
2868 | DRM_ERROR("late_init of IP block <%s> failed %d\n" , |
2869 | adev->ip_blocks[i].version->funcs->name, r); |
2870 | return r; |
2871 | } |
2872 | } |
2873 | adev->ip_blocks[i].status.late_initialized = true; |
2874 | } |
2875 | |
2876 | r = amdgpu_ras_late_init(adev); |
2877 | if (r) { |
2878 | DRM_ERROR("amdgpu_ras_late_init failed %d" , r); |
2879 | return r; |
2880 | } |
2881 | |
2882 | amdgpu_ras_set_error_query_ready(adev, ready: true); |
2883 | |
2884 | amdgpu_device_set_cg_state(adev, state: AMD_CG_STATE_GATE); |
2885 | amdgpu_device_set_pg_state(adev, state: AMD_PG_STATE_GATE); |
2886 | |
2887 | amdgpu_device_fill_reset_magic(adev); |
2888 | |
2889 | r = amdgpu_device_enable_mgpu_fan_boost(); |
2890 | if (r) |
2891 | DRM_ERROR("enable mgpu fan boost failed (%d).\n" , r); |
2892 | |
2893 | /* For passthrough configuration on arcturus and aldebaran, enable special handling SBR */ |
2894 | if (amdgpu_passthrough(adev) && |
2895 | ((adev->asic_type == CHIP_ARCTURUS && adev->gmc.xgmi.num_physical_nodes > 1) || |
2896 | adev->asic_type == CHIP_ALDEBARAN)) |
2897 | amdgpu_dpm_handle_passthrough_sbr(adev, enable: true); |
2898 | |
2899 | if (adev->gmc.xgmi.num_physical_nodes > 1) { |
2900 | mutex_lock(&mgpu_info.mutex); |
2901 | |
2902 | /* |
2903 | * Reset device p-state to low as this was booted with high. |
2904 | * |
2905 | * This should be performed only after all devices from the same |
2906 | * hive get initialized. |
2907 | * |
2908 | * However, it's unknown how many device in the hive in advance. |
2909 | * As this is counted one by one during devices initializations. |
2910 | * |
2911 | * So, we wait for all XGMI interlinked devices initialized. |
2912 | * This may bring some delays as those devices may come from |
2913 | * different hives. But that should be OK. |
2914 | */ |
2915 | if (mgpu_info.num_dgpu == adev->gmc.xgmi.num_physical_nodes) { |
2916 | for (i = 0; i < mgpu_info.num_gpu; i++) { |
2917 | gpu_instance = &(mgpu_info.gpu_ins[i]); |
2918 | if (gpu_instance->adev->flags & AMD_IS_APU) |
2919 | continue; |
2920 | |
2921 | r = amdgpu_xgmi_set_pstate(adev: gpu_instance->adev, |
2922 | pstate: AMDGPU_XGMI_PSTATE_MIN); |
2923 | if (r) { |
2924 | DRM_ERROR("pstate setting failed (%d).\n" , r); |
2925 | break; |
2926 | } |
2927 | } |
2928 | } |
2929 | |
2930 | mutex_unlock(lock: &mgpu_info.mutex); |
2931 | } |
2932 | |
2933 | return 0; |
2934 | } |
2935 | |
2936 | /** |
2937 | * amdgpu_device_smu_fini_early - smu hw_fini wrapper |
2938 | * |
2939 | * @adev: amdgpu_device pointer |
2940 | * |
2941 | * For ASICs need to disable SMC first |
2942 | */ |
2943 | static void amdgpu_device_smu_fini_early(struct amdgpu_device *adev) |
2944 | { |
2945 | int i, r; |
2946 | |
2947 | if (amdgpu_ip_version(adev, ip: GC_HWIP, inst: 0) > IP_VERSION(9, 0, 0)) |
2948 | return; |
2949 | |
2950 | for (i = 0; i < adev->num_ip_blocks; i++) { |
2951 | if (!adev->ip_blocks[i].status.hw) |
2952 | continue; |
2953 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) { |
2954 | r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev); |
2955 | /* XXX handle errors */ |
2956 | if (r) { |
2957 | DRM_DEBUG("hw_fini of IP block <%s> failed %d\n" , |
2958 | adev->ip_blocks[i].version->funcs->name, r); |
2959 | } |
2960 | adev->ip_blocks[i].status.hw = false; |
2961 | break; |
2962 | } |
2963 | } |
2964 | } |
2965 | |
2966 | static int amdgpu_device_ip_fini_early(struct amdgpu_device *adev) |
2967 | { |
2968 | int i, r; |
2969 | |
2970 | for (i = 0; i < adev->num_ip_blocks; i++) { |
2971 | if (!adev->ip_blocks[i].version->funcs->early_fini) |
2972 | continue; |
2973 | |
2974 | r = adev->ip_blocks[i].version->funcs->early_fini((void *)adev); |
2975 | if (r) { |
2976 | DRM_DEBUG("early_fini of IP block <%s> failed %d\n" , |
2977 | adev->ip_blocks[i].version->funcs->name, r); |
2978 | } |
2979 | } |
2980 | |
2981 | amdgpu_device_set_pg_state(adev, state: AMD_PG_STATE_UNGATE); |
2982 | amdgpu_device_set_cg_state(adev, state: AMD_CG_STATE_UNGATE); |
2983 | |
2984 | amdgpu_amdkfd_suspend(adev, run_pm: false); |
2985 | |
2986 | /* Workaroud for ASICs need to disable SMC first */ |
2987 | amdgpu_device_smu_fini_early(adev); |
2988 | |
2989 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { |
2990 | if (!adev->ip_blocks[i].status.hw) |
2991 | continue; |
2992 | |
2993 | r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev); |
2994 | /* XXX handle errors */ |
2995 | if (r) { |
2996 | DRM_DEBUG("hw_fini of IP block <%s> failed %d\n" , |
2997 | adev->ip_blocks[i].version->funcs->name, r); |
2998 | } |
2999 | |
3000 | adev->ip_blocks[i].status.hw = false; |
3001 | } |
3002 | |
3003 | if (amdgpu_sriov_vf(adev)) { |
3004 | if (amdgpu_virt_release_full_gpu(adev, init: false)) |
3005 | DRM_ERROR("failed to release exclusive mode on fini\n" ); |
3006 | } |
3007 | |
3008 | return 0; |
3009 | } |
3010 | |
3011 | /** |
3012 | * amdgpu_device_ip_fini - run fini for hardware IPs |
3013 | * |
3014 | * @adev: amdgpu_device pointer |
3015 | * |
3016 | * Main teardown pass for hardware IPs. The list of all the hardware |
3017 | * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks |
3018 | * are run. hw_fini tears down the hardware associated with each IP |
3019 | * and sw_fini tears down any software state associated with each IP. |
3020 | * Returns 0 on success, negative error code on failure. |
3021 | */ |
3022 | static int amdgpu_device_ip_fini(struct amdgpu_device *adev) |
3023 | { |
3024 | int i, r; |
3025 | |
3026 | if (amdgpu_sriov_vf(adev) && adev->virt.ras_init_done) |
3027 | amdgpu_virt_release_ras_err_handler_data(adev); |
3028 | |
3029 | if (adev->gmc.xgmi.num_physical_nodes > 1) |
3030 | amdgpu_xgmi_remove_device(adev); |
3031 | |
3032 | amdgpu_amdkfd_device_fini_sw(adev); |
3033 | |
3034 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { |
3035 | if (!adev->ip_blocks[i].status.sw) |
3036 | continue; |
3037 | |
3038 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) { |
3039 | amdgpu_ucode_free_bo(adev); |
3040 | amdgpu_free_static_csa(bo: &adev->virt.csa_obj); |
3041 | amdgpu_device_wb_fini(adev); |
3042 | amdgpu_device_mem_scratch_fini(adev); |
3043 | amdgpu_ib_pool_fini(adev); |
3044 | } |
3045 | |
3046 | r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev); |
3047 | /* XXX handle errors */ |
3048 | if (r) { |
3049 | DRM_DEBUG("sw_fini of IP block <%s> failed %d\n" , |
3050 | adev->ip_blocks[i].version->funcs->name, r); |
3051 | } |
3052 | adev->ip_blocks[i].status.sw = false; |
3053 | adev->ip_blocks[i].status.valid = false; |
3054 | } |
3055 | |
3056 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { |
3057 | if (!adev->ip_blocks[i].status.late_initialized) |
3058 | continue; |
3059 | if (adev->ip_blocks[i].version->funcs->late_fini) |
3060 | adev->ip_blocks[i].version->funcs->late_fini((void *)adev); |
3061 | adev->ip_blocks[i].status.late_initialized = false; |
3062 | } |
3063 | |
3064 | amdgpu_ras_fini(adev); |
3065 | |
3066 | return 0; |
3067 | } |
3068 | |
3069 | /** |
3070 | * amdgpu_device_delayed_init_work_handler - work handler for IB tests |
3071 | * |
3072 | * @work: work_struct. |
3073 | */ |
3074 | static void amdgpu_device_delayed_init_work_handler(struct work_struct *work) |
3075 | { |
3076 | struct amdgpu_device *adev = |
3077 | container_of(work, struct amdgpu_device, delayed_init_work.work); |
3078 | int r; |
3079 | |
3080 | r = amdgpu_ib_ring_tests(adev); |
3081 | if (r) |
3082 | DRM_ERROR("ib ring test failed (%d).\n" , r); |
3083 | } |
3084 | |
3085 | static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work) |
3086 | { |
3087 | struct amdgpu_device *adev = |
3088 | container_of(work, struct amdgpu_device, gfx.gfx_off_delay_work.work); |
3089 | |
3090 | WARN_ON_ONCE(adev->gfx.gfx_off_state); |
3091 | WARN_ON_ONCE(adev->gfx.gfx_off_req_count); |
3092 | |
3093 | if (!amdgpu_dpm_set_powergating_by_smu(adev, block_type: AMD_IP_BLOCK_TYPE_GFX, gate: true)) |
3094 | adev->gfx.gfx_off_state = true; |
3095 | } |
3096 | |
3097 | /** |
3098 | * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1) |
3099 | * |
3100 | * @adev: amdgpu_device pointer |
3101 | * |
3102 | * Main suspend function for hardware IPs. The list of all the hardware |
3103 | * IPs that make up the asic is walked, clockgating is disabled and the |
3104 | * suspend callbacks are run. suspend puts the hardware and software state |
3105 | * in each IP into a state suitable for suspend. |
3106 | * Returns 0 on success, negative error code on failure. |
3107 | */ |
3108 | static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev) |
3109 | { |
3110 | int i, r; |
3111 | |
3112 | amdgpu_device_set_pg_state(adev, state: AMD_PG_STATE_UNGATE); |
3113 | amdgpu_device_set_cg_state(adev, state: AMD_CG_STATE_UNGATE); |
3114 | |
3115 | /* |
3116 | * Per PMFW team's suggestion, driver needs to handle gfxoff |
3117 | * and df cstate features disablement for gpu reset(e.g. Mode1Reset) |
3118 | * scenario. Add the missing df cstate disablement here. |
3119 | */ |
3120 | if (amdgpu_dpm_set_df_cstate(adev, cstate: DF_CSTATE_DISALLOW)) |
3121 | dev_warn(adev->dev, "Failed to disallow df cstate" ); |
3122 | |
3123 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { |
3124 | if (!adev->ip_blocks[i].status.valid) |
3125 | continue; |
3126 | |
3127 | /* displays are handled separately */ |
3128 | if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_DCE) |
3129 | continue; |
3130 | |
3131 | /* XXX handle errors */ |
3132 | r = adev->ip_blocks[i].version->funcs->suspend(adev); |
3133 | /* XXX handle errors */ |
3134 | if (r) { |
3135 | DRM_ERROR("suspend of IP block <%s> failed %d\n" , |
3136 | adev->ip_blocks[i].version->funcs->name, r); |
3137 | return r; |
3138 | } |
3139 | |
3140 | adev->ip_blocks[i].status.hw = false; |
3141 | } |
3142 | |
3143 | return 0; |
3144 | } |
3145 | |
3146 | /** |
3147 | * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2) |
3148 | * |
3149 | * @adev: amdgpu_device pointer |
3150 | * |
3151 | * Main suspend function for hardware IPs. The list of all the hardware |
3152 | * IPs that make up the asic is walked, clockgating is disabled and the |
3153 | * suspend callbacks are run. suspend puts the hardware and software state |
3154 | * in each IP into a state suitable for suspend. |
3155 | * Returns 0 on success, negative error code on failure. |
3156 | */ |
3157 | static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev) |
3158 | { |
3159 | int i, r; |
3160 | |
3161 | if (adev->in_s0ix) |
3162 | amdgpu_dpm_gfx_state_change(adev, state: sGpuChangeState_D3Entry); |
3163 | |
3164 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { |
3165 | if (!adev->ip_blocks[i].status.valid) |
3166 | continue; |
3167 | /* displays are handled in phase1 */ |
3168 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) |
3169 | continue; |
3170 | /* PSP lost connection when err_event_athub occurs */ |
3171 | if (amdgpu_ras_intr_triggered() && |
3172 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) { |
3173 | adev->ip_blocks[i].status.hw = false; |
3174 | continue; |
3175 | } |
3176 | |
3177 | /* skip unnecessary suspend if we do not initialize them yet */ |
3178 | if (adev->gmc.xgmi.pending_reset && |
3179 | !(adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC || |
3180 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC || |
3181 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || |
3182 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH)) { |
3183 | adev->ip_blocks[i].status.hw = false; |
3184 | continue; |
3185 | } |
3186 | |
3187 | /* skip suspend of gfx/mes and psp for S0ix |
3188 | * gfx is in gfxoff state, so on resume it will exit gfxoff just |
3189 | * like at runtime. PSP is also part of the always on hardware |
3190 | * so no need to suspend it. |
3191 | */ |
3192 | if (adev->in_s0ix && |
3193 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP || |
3194 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX || |
3195 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_MES)) |
3196 | continue; |
3197 | |
3198 | /* SDMA 5.x+ is part of GFX power domain so it's covered by GFXOFF */ |
3199 | if (adev->in_s0ix && |
3200 | (amdgpu_ip_version(adev, ip: SDMA0_HWIP, inst: 0) >= |
3201 | IP_VERSION(5, 0, 0)) && |
3202 | (adev->ip_blocks[i].version->type == |
3203 | AMD_IP_BLOCK_TYPE_SDMA)) |
3204 | continue; |
3205 | |
3206 | /* Once swPSP provides the IMU, RLC FW binaries to TOS during cold-boot. |
3207 | * These are in TMR, hence are expected to be reused by PSP-TOS to reload |
3208 | * from this location and RLC Autoload automatically also gets loaded |
3209 | * from here based on PMFW -> PSP message during re-init sequence. |
3210 | * Therefore, the psp suspend & resume should be skipped to avoid destroy |
3211 | * the TMR and reload FWs again for IMU enabled APU ASICs. |
3212 | */ |
3213 | if (amdgpu_in_reset(adev) && |
3214 | (adev->flags & AMD_IS_APU) && adev->gfx.imu.funcs && |
3215 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) |
3216 | continue; |
3217 | |
3218 | /* XXX handle errors */ |
3219 | r = adev->ip_blocks[i].version->funcs->suspend(adev); |
3220 | /* XXX handle errors */ |
3221 | if (r) { |
3222 | DRM_ERROR("suspend of IP block <%s> failed %d\n" , |
3223 | adev->ip_blocks[i].version->funcs->name, r); |
3224 | } |
3225 | adev->ip_blocks[i].status.hw = false; |
3226 | /* handle putting the SMC in the appropriate state */ |
3227 | if (!amdgpu_sriov_vf(adev)) { |
3228 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) { |
3229 | r = amdgpu_dpm_set_mp1_state(adev, mp1_state: adev->mp1_state); |
3230 | if (r) { |
3231 | DRM_ERROR("SMC failed to set mp1 state %d, %d\n" , |
3232 | adev->mp1_state, r); |
3233 | return r; |
3234 | } |
3235 | } |
3236 | } |
3237 | } |
3238 | |
3239 | return 0; |
3240 | } |
3241 | |
3242 | /** |
3243 | * amdgpu_device_ip_suspend - run suspend for hardware IPs |
3244 | * |
3245 | * @adev: amdgpu_device pointer |
3246 | * |
3247 | * Main suspend function for hardware IPs. The list of all the hardware |
3248 | * IPs that make up the asic is walked, clockgating is disabled and the |
3249 | * suspend callbacks are run. suspend puts the hardware and software state |
3250 | * in each IP into a state suitable for suspend. |
3251 | * Returns 0 on success, negative error code on failure. |
3252 | */ |
3253 | int amdgpu_device_ip_suspend(struct amdgpu_device *adev) |
3254 | { |
3255 | int r; |
3256 | |
3257 | if (amdgpu_sriov_vf(adev)) { |
3258 | amdgpu_virt_fini_data_exchange(adev); |
3259 | amdgpu_virt_request_full_gpu(adev, init: false); |
3260 | } |
3261 | |
3262 | amdgpu_ttm_set_buffer_funcs_status(adev, enable: false); |
3263 | |
3264 | r = amdgpu_device_ip_suspend_phase1(adev); |
3265 | if (r) |
3266 | return r; |
3267 | r = amdgpu_device_ip_suspend_phase2(adev); |
3268 | |
3269 | if (amdgpu_sriov_vf(adev)) |
3270 | amdgpu_virt_release_full_gpu(adev, init: false); |
3271 | |
3272 | return r; |
3273 | } |
3274 | |
3275 | static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev) |
3276 | { |
3277 | int i, r; |
3278 | |
3279 | static enum amd_ip_block_type ip_order[] = { |
3280 | AMD_IP_BLOCK_TYPE_COMMON, |
3281 | AMD_IP_BLOCK_TYPE_GMC, |
3282 | AMD_IP_BLOCK_TYPE_PSP, |
3283 | AMD_IP_BLOCK_TYPE_IH, |
3284 | }; |
3285 | |
3286 | for (i = 0; i < adev->num_ip_blocks; i++) { |
3287 | int j; |
3288 | struct amdgpu_ip_block *block; |
3289 | |
3290 | block = &adev->ip_blocks[i]; |
3291 | block->status.hw = false; |
3292 | |
3293 | for (j = 0; j < ARRAY_SIZE(ip_order); j++) { |
3294 | |
3295 | if (block->version->type != ip_order[j] || |
3296 | !block->status.valid) |
3297 | continue; |
3298 | |
3299 | r = block->version->funcs->hw_init(adev); |
3300 | DRM_INFO("RE-INIT-early: %s %s\n" , block->version->funcs->name, r?"failed" :"succeeded" ); |
3301 | if (r) |
3302 | return r; |
3303 | block->status.hw = true; |
3304 | } |
3305 | } |
3306 | |
3307 | return 0; |
3308 | } |
3309 | |
3310 | static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev) |
3311 | { |
3312 | int i, r; |
3313 | |
3314 | static enum amd_ip_block_type ip_order[] = { |
3315 | AMD_IP_BLOCK_TYPE_SMC, |
3316 | AMD_IP_BLOCK_TYPE_DCE, |
3317 | AMD_IP_BLOCK_TYPE_GFX, |
3318 | AMD_IP_BLOCK_TYPE_SDMA, |
3319 | AMD_IP_BLOCK_TYPE_MES, |
3320 | AMD_IP_BLOCK_TYPE_UVD, |
3321 | AMD_IP_BLOCK_TYPE_VCE, |
3322 | AMD_IP_BLOCK_TYPE_VCN, |
3323 | AMD_IP_BLOCK_TYPE_JPEG |
3324 | }; |
3325 | |
3326 | for (i = 0; i < ARRAY_SIZE(ip_order); i++) { |
3327 | int j; |
3328 | struct amdgpu_ip_block *block; |
3329 | |
3330 | for (j = 0; j < adev->num_ip_blocks; j++) { |
3331 | block = &adev->ip_blocks[j]; |
3332 | |
3333 | if (block->version->type != ip_order[i] || |
3334 | !block->status.valid || |
3335 | block->status.hw) |
3336 | continue; |
3337 | |
3338 | if (block->version->type == AMD_IP_BLOCK_TYPE_SMC) |
3339 | r = block->version->funcs->resume(adev); |
3340 | else |
3341 | r = block->version->funcs->hw_init(adev); |
3342 | |
3343 | DRM_INFO("RE-INIT-late: %s %s\n" , block->version->funcs->name, r?"failed" :"succeeded" ); |
3344 | if (r) |
3345 | return r; |
3346 | block->status.hw = true; |
3347 | } |
3348 | } |
3349 | |
3350 | return 0; |
3351 | } |
3352 | |
3353 | /** |
3354 | * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs |
3355 | * |
3356 | * @adev: amdgpu_device pointer |
3357 | * |
3358 | * First resume function for hardware IPs. The list of all the hardware |
3359 | * IPs that make up the asic is walked and the resume callbacks are run for |
3360 | * COMMON, GMC, and IH. resume puts the hardware into a functional state |
3361 | * after a suspend and updates the software state as necessary. This |
3362 | * function is also used for restoring the GPU after a GPU reset. |
3363 | * Returns 0 on success, negative error code on failure. |
3364 | */ |
3365 | static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev) |
3366 | { |
3367 | int i, r; |
3368 | |
3369 | for (i = 0; i < adev->num_ip_blocks; i++) { |
3370 | if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw) |
3371 | continue; |
3372 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || |
3373 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC || |
3374 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH || |
3375 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP && amdgpu_sriov_vf(adev))) { |
3376 | |
3377 | r = adev->ip_blocks[i].version->funcs->resume(adev); |
3378 | if (r) { |
3379 | DRM_ERROR("resume of IP block <%s> failed %d\n" , |
3380 | adev->ip_blocks[i].version->funcs->name, r); |
3381 | return r; |
3382 | } |
3383 | adev->ip_blocks[i].status.hw = true; |
3384 | } |
3385 | } |
3386 | |
3387 | return 0; |
3388 | } |
3389 | |
3390 | /** |
3391 | * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs |
3392 | * |
3393 | * @adev: amdgpu_device pointer |
3394 | * |
3395 | * First resume function for hardware IPs. The list of all the hardware |
3396 | * IPs that make up the asic is walked and the resume callbacks are run for |
3397 | * all blocks except COMMON, GMC, and IH. resume puts the hardware into a |
3398 | * functional state after a suspend and updates the software state as |
3399 | * necessary. This function is also used for restoring the GPU after a GPU |
3400 | * reset. |
3401 | * Returns 0 on success, negative error code on failure. |
3402 | */ |
3403 | static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev) |
3404 | { |
3405 | int i, r; |
3406 | |
3407 | for (i = 0; i < adev->num_ip_blocks; i++) { |
3408 | if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw) |
3409 | continue; |
3410 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || |
3411 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC || |
3412 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH || |
3413 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) |
3414 | continue; |
3415 | r = adev->ip_blocks[i].version->funcs->resume(adev); |
3416 | if (r) { |
3417 | DRM_ERROR("resume of IP block <%s> failed %d\n" , |
3418 | adev->ip_blocks[i].version->funcs->name, r); |
3419 | return r; |
3420 | } |
3421 | adev->ip_blocks[i].status.hw = true; |
3422 | } |
3423 | |
3424 | return 0; |
3425 | } |
3426 | |
3427 | /** |
3428 | * amdgpu_device_ip_resume - run resume for hardware IPs |
3429 | * |
3430 | * @adev: amdgpu_device pointer |
3431 | * |
3432 | * Main resume function for hardware IPs. The hardware IPs |
3433 | * are split into two resume functions because they are |
3434 | * also used in recovering from a GPU reset and some additional |
3435 | * steps need to be take between them. In this case (S3/S4) they are |
3436 | * run sequentially. |
3437 | * Returns 0 on success, negative error code on failure. |
3438 | */ |
3439 | static int amdgpu_device_ip_resume(struct amdgpu_device *adev) |
3440 | { |
3441 | int r; |
3442 | |
3443 | r = amdgpu_device_ip_resume_phase1(adev); |
3444 | if (r) |
3445 | return r; |
3446 | |
3447 | r = amdgpu_device_fw_loading(adev); |
3448 | if (r) |
3449 | return r; |
3450 | |
3451 | r = amdgpu_device_ip_resume_phase2(adev); |
3452 | |
3453 | if (adev->mman.buffer_funcs_ring->sched.ready) |
3454 | amdgpu_ttm_set_buffer_funcs_status(adev, enable: true); |
3455 | |
3456 | return r; |
3457 | } |
3458 | |
3459 | /** |
3460 | * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV |
3461 | * |
3462 | * @adev: amdgpu_device pointer |
3463 | * |
3464 | * Query the VBIOS data tables to determine if the board supports SR-IOV. |
3465 | */ |
3466 | static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev) |
3467 | { |
3468 | if (amdgpu_sriov_vf(adev)) { |
3469 | if (adev->is_atom_fw) { |
3470 | if (amdgpu_atomfirmware_gpu_virtualization_supported(adev)) |
3471 | adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS; |
3472 | } else { |
3473 | if (amdgpu_atombios_has_gpu_virtualization_table(adev)) |
3474 | adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS; |
3475 | } |
3476 | |
3477 | if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS)) |
3478 | amdgpu_vf_error_put(adev, sub_error_code: AMDGIM_ERROR_VF_NO_VBIOS, error_flags: 0, error_data: 0); |
3479 | } |
3480 | } |
3481 | |
3482 | /** |
3483 | * amdgpu_device_asic_has_dc_support - determine if DC supports the asic |
3484 | * |
3485 | * @asic_type: AMD asic type |
3486 | * |
3487 | * Check if there is DC (new modesetting infrastructre) support for an asic. |
3488 | * returns true if DC has support, false if not. |
3489 | */ |
3490 | bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type) |
3491 | { |
3492 | switch (asic_type) { |
3493 | #ifdef CONFIG_DRM_AMDGPU_SI |
3494 | case CHIP_HAINAN: |
3495 | #endif |
3496 | case CHIP_TOPAZ: |
3497 | /* chips with no display hardware */ |
3498 | return false; |
3499 | #if defined(CONFIG_DRM_AMD_DC) |
3500 | case CHIP_TAHITI: |
3501 | case CHIP_PITCAIRN: |
3502 | case CHIP_VERDE: |
3503 | case CHIP_OLAND: |
3504 | /* |
3505 | * We have systems in the wild with these ASICs that require |
3506 | * LVDS and VGA support which is not supported with DC. |
3507 | * |
3508 | * Fallback to the non-DC driver here by default so as not to |
3509 | * cause regressions. |
3510 | */ |
3511 | #if defined(CONFIG_DRM_AMD_DC_SI) |
3512 | return amdgpu_dc > 0; |
3513 | #else |
3514 | return false; |
3515 | #endif |
3516 | case CHIP_BONAIRE: |
3517 | case CHIP_KAVERI: |
3518 | case CHIP_KABINI: |
3519 | case CHIP_MULLINS: |
3520 | /* |
3521 | * We have systems in the wild with these ASICs that require |
3522 | * VGA support which is not supported with DC. |
3523 | * |
3524 | * Fallback to the non-DC driver here by default so as not to |
3525 | * cause regressions. |
3526 | */ |
3527 | return amdgpu_dc > 0; |
3528 | default: |
3529 | return amdgpu_dc != 0; |
3530 | #else |
3531 | default: |
3532 | if (amdgpu_dc > 0) |
3533 | DRM_INFO_ONCE("Display Core has been requested via kernel parameter but isn't supported by ASIC, ignoring\n" ); |
3534 | return false; |
3535 | #endif |
3536 | } |
3537 | } |
3538 | |
3539 | /** |
3540 | * amdgpu_device_has_dc_support - check if dc is supported |
3541 | * |
3542 | * @adev: amdgpu_device pointer |
3543 | * |
3544 | * Returns true for supported, false for not supported |
3545 | */ |
3546 | bool amdgpu_device_has_dc_support(struct amdgpu_device *adev) |
3547 | { |
3548 | if (adev->enable_virtual_display || |
3549 | (adev->harvest_ip_mask & AMD_HARVEST_IP_DMU_MASK)) |
3550 | return false; |
3551 | |
3552 | return amdgpu_device_asic_has_dc_support(asic_type: adev->asic_type); |
3553 | } |
3554 | |
3555 | static void amdgpu_device_xgmi_reset_func(struct work_struct *__work) |
3556 | { |
3557 | struct amdgpu_device *adev = |
3558 | container_of(__work, struct amdgpu_device, xgmi_reset_work); |
3559 | struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev); |
3560 | |
3561 | /* It's a bug to not have a hive within this function */ |
3562 | if (WARN_ON(!hive)) |
3563 | return; |
3564 | |
3565 | /* |
3566 | * Use task barrier to synchronize all xgmi reset works across the |
3567 | * hive. task_barrier_enter and task_barrier_exit will block |
3568 | * until all the threads running the xgmi reset works reach |
3569 | * those points. task_barrier_full will do both blocks. |
3570 | */ |
3571 | if (amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) { |
3572 | |
3573 | task_barrier_enter(tb: &hive->tb); |
3574 | adev->asic_reset_res = amdgpu_device_baco_enter(dev: adev_to_drm(adev)); |
3575 | |
3576 | if (adev->asic_reset_res) |
3577 | goto fail; |
3578 | |
3579 | task_barrier_exit(tb: &hive->tb); |
3580 | adev->asic_reset_res = amdgpu_device_baco_exit(dev: adev_to_drm(adev)); |
3581 | |
3582 | if (adev->asic_reset_res) |
3583 | goto fail; |
3584 | |
3585 | amdgpu_ras_reset_error_count(adev, block: AMDGPU_RAS_BLOCK__MMHUB); |
3586 | } else { |
3587 | |
3588 | task_barrier_full(tb: &hive->tb); |
3589 | adev->asic_reset_res = amdgpu_asic_reset(adev); |
3590 | } |
3591 | |
3592 | fail: |
3593 | if (adev->asic_reset_res) |
3594 | DRM_WARN("ASIC reset failed with error, %d for drm dev, %s" , |
3595 | adev->asic_reset_res, adev_to_drm(adev)->unique); |
3596 | amdgpu_put_xgmi_hive(hive); |
3597 | } |
3598 | |
3599 | static int amdgpu_device_get_job_timeout_settings(struct amdgpu_device *adev) |
3600 | { |
3601 | char *input = amdgpu_lockup_timeout; |
3602 | char *timeout_setting = NULL; |
3603 | int index = 0; |
3604 | long timeout; |
3605 | int ret = 0; |
3606 | |
3607 | /* |
3608 | * By default timeout for non compute jobs is 10000 |
3609 | * and 60000 for compute jobs. |
3610 | * In SR-IOV or passthrough mode, timeout for compute |
3611 | * jobs are 60000 by default. |
3612 | */ |
3613 | adev->gfx_timeout = msecs_to_jiffies(m: 10000); |
3614 | adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout; |
3615 | if (amdgpu_sriov_vf(adev)) |
3616 | adev->compute_timeout = amdgpu_sriov_is_pp_one_vf(adev) ? |
3617 | msecs_to_jiffies(m: 60000) : msecs_to_jiffies(m: 10000); |
3618 | else |
3619 | adev->compute_timeout = msecs_to_jiffies(m: 60000); |
3620 | |
3621 | if (strnlen(p: input, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) { |
3622 | while ((timeout_setting = strsep(&input, "," )) && |
3623 | strnlen(p: timeout_setting, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) { |
3624 | ret = kstrtol(s: timeout_setting, base: 0, res: &timeout); |
3625 | if (ret) |
3626 | return ret; |
3627 | |
3628 | if (timeout == 0) { |
3629 | index++; |
3630 | continue; |
3631 | } else if (timeout < 0) { |
3632 | timeout = MAX_SCHEDULE_TIMEOUT; |
3633 | dev_warn(adev->dev, "lockup timeout disabled" ); |
3634 | add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK); |
3635 | } else { |
3636 | timeout = msecs_to_jiffies(m: timeout); |
3637 | } |
3638 | |
3639 | switch (index++) { |
3640 | case 0: |
3641 | adev->gfx_timeout = timeout; |
3642 | break; |
3643 | case 1: |
3644 | adev->compute_timeout = timeout; |
3645 | break; |
3646 | case 2: |
3647 | adev->sdma_timeout = timeout; |
3648 | break; |
3649 | case 3: |
3650 | adev->video_timeout = timeout; |
3651 | break; |
3652 | default: |
3653 | break; |
3654 | } |
3655 | } |
3656 | /* |
3657 | * There is only one value specified and |
3658 | * it should apply to all non-compute jobs. |
3659 | */ |
3660 | if (index == 1) { |
3661 | adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout; |
3662 | if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev)) |
3663 | adev->compute_timeout = adev->gfx_timeout; |
3664 | } |
3665 | } |
3666 | |
3667 | return ret; |
3668 | } |
3669 | |
3670 | /** |
3671 | * amdgpu_device_check_iommu_direct_map - check if RAM direct mapped to GPU |
3672 | * |
3673 | * @adev: amdgpu_device pointer |
3674 | * |
3675 | * RAM direct mapped to GPU if IOMMU is not enabled or is pass through mode |
3676 | */ |
3677 | static void amdgpu_device_check_iommu_direct_map(struct amdgpu_device *adev) |
3678 | { |
3679 | struct iommu_domain *domain; |
3680 | |
3681 | domain = iommu_get_domain_for_dev(dev: adev->dev); |
3682 | if (!domain || domain->type == IOMMU_DOMAIN_IDENTITY) |
3683 | adev->ram_is_direct_mapped = true; |
3684 | } |
3685 | |
3686 | static const struct attribute *amdgpu_dev_attributes[] = { |
3687 | &dev_attr_pcie_replay_count.attr, |
3688 | NULL |
3689 | }; |
3690 | |
3691 | static void amdgpu_device_set_mcbp(struct amdgpu_device *adev) |
3692 | { |
3693 | if (amdgpu_mcbp == 1) |
3694 | adev->gfx.mcbp = true; |
3695 | else if (amdgpu_mcbp == 0) |
3696 | adev->gfx.mcbp = false; |
3697 | else if ((amdgpu_ip_version(adev, ip: GC_HWIP, inst: 0) >= IP_VERSION(9, 0, 0)) && |
3698 | (amdgpu_ip_version(adev, ip: GC_HWIP, inst: 0) < IP_VERSION(10, 0, 0)) && |
3699 | adev->gfx.num_gfx_rings) |
3700 | adev->gfx.mcbp = true; |
3701 | |
3702 | if (amdgpu_sriov_vf(adev)) |
3703 | adev->gfx.mcbp = true; |
3704 | |
3705 | if (adev->gfx.mcbp) |
3706 | DRM_INFO("MCBP is enabled\n" ); |
3707 | } |
3708 | |
3709 | /** |
3710 | * amdgpu_device_init - initialize the driver |
3711 | * |
3712 | * @adev: amdgpu_device pointer |
3713 | * @flags: driver flags |
3714 | * |
3715 | * Initializes the driver info and hw (all asics). |
3716 | * Returns 0 for success or an error on failure. |
3717 | * Called at driver startup. |
3718 | */ |
3719 | int amdgpu_device_init(struct amdgpu_device *adev, |
3720 | uint32_t flags) |
3721 | { |
3722 | struct drm_device *ddev = adev_to_drm(adev); |
3723 | struct pci_dev *pdev = adev->pdev; |
3724 | int r, i; |
3725 | bool px = false; |
3726 | u32 max_MBps; |
3727 | int tmp; |
3728 | |
3729 | adev->shutdown = false; |
3730 | adev->flags = flags; |
3731 | |
3732 | if (amdgpu_force_asic_type >= 0 && amdgpu_force_asic_type < CHIP_LAST) |
3733 | adev->asic_type = amdgpu_force_asic_type; |
3734 | else |
3735 | adev->asic_type = flags & AMD_ASIC_MASK; |
3736 | |
3737 | adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT; |
3738 | if (amdgpu_emu_mode == 1) |
3739 | adev->usec_timeout *= 10; |
3740 | adev->gmc.gart_size = 512 * 1024 * 1024; |
3741 | adev->accel_working = false; |
3742 | adev->num_rings = 0; |
3743 | RCU_INIT_POINTER(adev->gang_submit, dma_fence_get_stub()); |
3744 | adev->mman.buffer_funcs = NULL; |
3745 | adev->mman.buffer_funcs_ring = NULL; |
3746 | adev->vm_manager.vm_pte_funcs = NULL; |
3747 | adev->vm_manager.vm_pte_num_scheds = 0; |
3748 | adev->gmc.gmc_funcs = NULL; |
3749 | adev->harvest_ip_mask = 0x0; |
3750 | adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS); |
3751 | bitmap_zero(dst: adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES); |
3752 | |
3753 | adev->smc_rreg = &amdgpu_invalid_rreg; |
3754 | adev->smc_wreg = &amdgpu_invalid_wreg; |
3755 | adev->pcie_rreg = &amdgpu_invalid_rreg; |
3756 | adev->pcie_wreg = &amdgpu_invalid_wreg; |
3757 | adev->pcie_rreg_ext = &amdgpu_invalid_rreg_ext; |
3758 | adev->pcie_wreg_ext = &amdgpu_invalid_wreg_ext; |
3759 | adev->pciep_rreg = &amdgpu_invalid_rreg; |
3760 | adev->pciep_wreg = &amdgpu_invalid_wreg; |
3761 | adev->pcie_rreg64 = &amdgpu_invalid_rreg64; |
3762 | adev->pcie_wreg64 = &amdgpu_invalid_wreg64; |
3763 | adev->pcie_rreg64_ext = &amdgpu_invalid_rreg64_ext; |
3764 | adev->pcie_wreg64_ext = &amdgpu_invalid_wreg64_ext; |
3765 | adev->uvd_ctx_rreg = &amdgpu_invalid_rreg; |
3766 | adev->uvd_ctx_wreg = &amdgpu_invalid_wreg; |
3767 | adev->didt_rreg = &amdgpu_invalid_rreg; |
3768 | adev->didt_wreg = &amdgpu_invalid_wreg; |
3769 | adev->gc_cac_rreg = &amdgpu_invalid_rreg; |
3770 | adev->gc_cac_wreg = &amdgpu_invalid_wreg; |
3771 | adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg; |
3772 | adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg; |
3773 | |
3774 | DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n" , |
3775 | amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device, |
3776 | pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision); |
3777 | |
3778 | /* mutex initialization are all done here so we |
3779 | * can recall function without having locking issues |
3780 | */ |
3781 | mutex_init(&adev->firmware.mutex); |
3782 | mutex_init(&adev->pm.mutex); |
3783 | mutex_init(&adev->gfx.gpu_clock_mutex); |
3784 | mutex_init(&adev->srbm_mutex); |
3785 | mutex_init(&adev->gfx.pipe_reserve_mutex); |
3786 | mutex_init(&adev->gfx.gfx_off_mutex); |
3787 | mutex_init(&adev->gfx.partition_mutex); |
3788 | mutex_init(&adev->grbm_idx_mutex); |
3789 | mutex_init(&adev->mn_lock); |
3790 | mutex_init(&adev->virt.vf_errors.lock); |
3791 | hash_init(adev->mn_hash); |
3792 | mutex_init(&adev->psp.mutex); |
3793 | mutex_init(&adev->notifier_lock); |
3794 | mutex_init(&adev->pm.stable_pstate_ctx_lock); |
3795 | mutex_init(&adev->benchmark_mutex); |
3796 | |
3797 | amdgpu_device_init_apu_flags(adev); |
3798 | |
3799 | r = amdgpu_device_check_arguments(adev); |
3800 | if (r) |
3801 | return r; |
3802 | |
3803 | spin_lock_init(&adev->mmio_idx_lock); |
3804 | spin_lock_init(&adev->smc_idx_lock); |
3805 | spin_lock_init(&adev->pcie_idx_lock); |
3806 | spin_lock_init(&adev->uvd_ctx_idx_lock); |
3807 | spin_lock_init(&adev->didt_idx_lock); |
3808 | spin_lock_init(&adev->gc_cac_idx_lock); |
3809 | spin_lock_init(&adev->se_cac_idx_lock); |
3810 | spin_lock_init(&adev->audio_endpt_idx_lock); |
3811 | spin_lock_init(&adev->mm_stats.lock); |
3812 | |
3813 | INIT_LIST_HEAD(list: &adev->shadow_list); |
3814 | mutex_init(&adev->shadow_list_lock); |
3815 | |
3816 | INIT_LIST_HEAD(list: &adev->reset_list); |
3817 | |
3818 | INIT_LIST_HEAD(list: &adev->ras_list); |
3819 | |
3820 | INIT_LIST_HEAD(list: &adev->pm.od_kobj_list); |
3821 | |
3822 | INIT_DELAYED_WORK(&adev->delayed_init_work, |
3823 | amdgpu_device_delayed_init_work_handler); |
3824 | INIT_DELAYED_WORK(&adev->gfx.gfx_off_delay_work, |
3825 | amdgpu_device_delay_enable_gfx_off); |
3826 | |
3827 | INIT_WORK(&adev->xgmi_reset_work, amdgpu_device_xgmi_reset_func); |
3828 | |
3829 | adev->gfx.gfx_off_req_count = 1; |
3830 | adev->gfx.gfx_off_residency = 0; |
3831 | adev->gfx.gfx_off_entrycount = 0; |
3832 | adev->pm.ac_power = power_supply_is_system_supplied() > 0; |
3833 | |
3834 | atomic_set(v: &adev->throttling_logging_enabled, i: 1); |
3835 | /* |
3836 | * If throttling continues, logging will be performed every minute |
3837 | * to avoid log flooding. "-1" is subtracted since the thermal |
3838 | * throttling interrupt comes every second. Thus, the total logging |
3839 | * interval is 59 seconds(retelimited printk interval) + 1(waiting |
3840 | * for throttling interrupt) = 60 seconds. |
3841 | */ |
3842 | ratelimit_state_init(rs: &adev->throttling_logging_rs, interval: (60 - 1) * HZ, burst: 1); |
3843 | ratelimit_set_flags(rs: &adev->throttling_logging_rs, RATELIMIT_MSG_ON_RELEASE); |
3844 | |
3845 | /* Registers mapping */ |
3846 | /* TODO: block userspace mapping of io register */ |
3847 | if (adev->asic_type >= CHIP_BONAIRE) { |
3848 | adev->rmmio_base = pci_resource_start(adev->pdev, 5); |
3849 | adev->rmmio_size = pci_resource_len(adev->pdev, 5); |
3850 | } else { |
3851 | adev->rmmio_base = pci_resource_start(adev->pdev, 2); |
3852 | adev->rmmio_size = pci_resource_len(adev->pdev, 2); |
3853 | } |
3854 | |
3855 | for (i = 0; i < AMD_IP_BLOCK_TYPE_NUM; i++) |
3856 | atomic_set(v: &adev->pm.pwr_state[i], i: POWER_STATE_UNKNOWN); |
3857 | |
3858 | adev->rmmio = ioremap(offset: adev->rmmio_base, size: adev->rmmio_size); |
3859 | if (!adev->rmmio) |
3860 | return -ENOMEM; |
3861 | |
3862 | DRM_INFO("register mmio base: 0x%08X\n" , (uint32_t)adev->rmmio_base); |
3863 | DRM_INFO("register mmio size: %u\n" , (unsigned int)adev->rmmio_size); |
3864 | |
3865 | /* |
3866 | * Reset domain needs to be present early, before XGMI hive discovered |
3867 | * (if any) and intitialized to use reset sem and in_gpu reset flag |
3868 | * early on during init and before calling to RREG32. |
3869 | */ |
3870 | adev->reset_domain = amdgpu_reset_create_reset_domain(type: SINGLE_DEVICE, wq_name: "amdgpu-reset-dev" ); |
3871 | if (!adev->reset_domain) |
3872 | return -ENOMEM; |
3873 | |
3874 | /* detect hw virtualization here */ |
3875 | amdgpu_detect_virtualization(adev); |
3876 | |
3877 | amdgpu_device_get_pcie_info(adev); |
3878 | |
3879 | r = amdgpu_device_get_job_timeout_settings(adev); |
3880 | if (r) { |
3881 | dev_err(adev->dev, "invalid lockup_timeout parameter syntax\n" ); |
3882 | return r; |
3883 | } |
3884 | |
3885 | /* early init functions */ |
3886 | r = amdgpu_device_ip_early_init(adev); |
3887 | if (r) |
3888 | return r; |
3889 | |
3890 | amdgpu_device_set_mcbp(adev); |
3891 | |
3892 | /* Get rid of things like offb */ |
3893 | r = drm_aperture_remove_conflicting_pci_framebuffers(pdev: adev->pdev, req_driver: &amdgpu_kms_driver); |
3894 | if (r) |
3895 | return r; |
3896 | |
3897 | /* Enable TMZ based on IP_VERSION */ |
3898 | amdgpu_gmc_tmz_set(adev); |
3899 | |
3900 | amdgpu_gmc_noretry_set(adev); |
3901 | /* Need to get xgmi info early to decide the reset behavior*/ |
3902 | if (adev->gmc.xgmi.supported) { |
3903 | r = adev->gfxhub.funcs->get_xgmi_info(adev); |
3904 | if (r) |
3905 | return r; |
3906 | } |
3907 | |
3908 | /* enable PCIE atomic ops */ |
3909 | if (amdgpu_sriov_vf(adev)) { |
3910 | if (adev->virt.fw_reserve.p_pf2vf) |
3911 | adev->have_atomics_support = ((struct amd_sriov_msg_pf2vf_info *) |
3912 | adev->virt.fw_reserve.p_pf2vf)->pcie_atomic_ops_support_flags == |
3913 | (PCI_EXP_DEVCAP2_ATOMIC_COMP32 | PCI_EXP_DEVCAP2_ATOMIC_COMP64); |
3914 | /* APUs w/ gfx9 onwards doesn't reply on PCIe atomics, rather it is a |
3915 | * internal path natively support atomics, set have_atomics_support to true. |
3916 | */ |
3917 | } else if ((adev->flags & AMD_IS_APU) && |
3918 | (amdgpu_ip_version(adev, ip: GC_HWIP, inst: 0) > |
3919 | IP_VERSION(9, 0, 0))) { |
3920 | adev->have_atomics_support = true; |
3921 | } else { |
3922 | adev->have_atomics_support = |
3923 | !pci_enable_atomic_ops_to_root(dev: adev->pdev, |
3924 | PCI_EXP_DEVCAP2_ATOMIC_COMP32 | |
3925 | PCI_EXP_DEVCAP2_ATOMIC_COMP64); |
3926 | } |
3927 | |
3928 | if (!adev->have_atomics_support) |
3929 | dev_info(adev->dev, "PCIE atomic ops is not supported\n" ); |
3930 | |
3931 | /* doorbell bar mapping and doorbell index init*/ |
3932 | amdgpu_doorbell_init(adev); |
3933 | |
3934 | if (amdgpu_emu_mode == 1) { |
3935 | /* post the asic on emulation mode */ |
3936 | emu_soc_asic_init(adev); |
3937 | goto fence_driver_init; |
3938 | } |
3939 | |
3940 | amdgpu_reset_init(adev); |
3941 | |
3942 | /* detect if we are with an SRIOV vbios */ |
3943 | if (adev->bios) |
3944 | amdgpu_device_detect_sriov_bios(adev); |
3945 | |
3946 | /* check if we need to reset the asic |
3947 | * E.g., driver was not cleanly unloaded previously, etc. |
3948 | */ |
3949 | if (!amdgpu_sriov_vf(adev) && amdgpu_asic_need_reset_on_init(adev)) { |
3950 | if (adev->gmc.xgmi.num_physical_nodes) { |
3951 | dev_info(adev->dev, "Pending hive reset.\n" ); |
3952 | adev->gmc.xgmi.pending_reset = true; |
3953 | /* Only need to init necessary block for SMU to handle the reset */ |
3954 | for (i = 0; i < adev->num_ip_blocks; i++) { |
3955 | if (!adev->ip_blocks[i].status.valid) |
3956 | continue; |
3957 | if (!(adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC || |
3958 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || |
3959 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH || |
3960 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC)) { |
3961 | DRM_DEBUG("IP %s disabled for hw_init.\n" , |
3962 | adev->ip_blocks[i].version->funcs->name); |
3963 | adev->ip_blocks[i].status.hw = true; |
3964 | } |
3965 | } |
3966 | } else { |
3967 | switch (amdgpu_ip_version(adev, ip: MP1_HWIP, inst: 0)) { |
3968 | case IP_VERSION(13, 0, 0): |
3969 | case IP_VERSION(13, 0, 7): |
3970 | case IP_VERSION(13, 0, 10): |
3971 | r = psp_gpu_reset(adev); |
3972 | break; |
3973 | default: |
3974 | tmp = amdgpu_reset_method; |
3975 | /* It should do a default reset when loading or reloading the driver, |
3976 | * regardless of the module parameter reset_method. |
3977 | */ |
3978 | amdgpu_reset_method = AMD_RESET_METHOD_NONE; |
3979 | r = amdgpu_asic_reset(adev); |
3980 | amdgpu_reset_method = tmp; |
3981 | break; |
3982 | } |
3983 | |
3984 | if (r) { |
3985 | dev_err(adev->dev, "asic reset on init failed\n" ); |
3986 | goto failed; |
3987 | } |
3988 | } |
3989 | } |
3990 | |
3991 | /* Post card if necessary */ |
3992 | if (amdgpu_device_need_post(adev)) { |
3993 | if (!adev->bios) { |
3994 | dev_err(adev->dev, "no vBIOS found\n" ); |
3995 | r = -EINVAL; |
3996 | goto failed; |
3997 | } |
3998 | DRM_INFO("GPU posting now...\n" ); |
3999 | r = amdgpu_device_asic_init(adev); |
4000 | if (r) { |
4001 | dev_err(adev->dev, "gpu post error!\n" ); |
4002 | goto failed; |
4003 | } |
4004 | } |
4005 | |
4006 | if (adev->bios) { |
4007 | if (adev->is_atom_fw) { |
4008 | /* Initialize clocks */ |
4009 | r = amdgpu_atomfirmware_get_clock_info(adev); |
4010 | if (r) { |
4011 | dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n" ); |
4012 | amdgpu_vf_error_put(adev, sub_error_code: AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, error_flags: 0, error_data: 0); |
4013 | goto failed; |
4014 | } |
4015 | } else { |
4016 | /* Initialize clocks */ |
4017 | r = amdgpu_atombios_get_clock_info(adev); |
4018 | if (r) { |
4019 | dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n" ); |
4020 | amdgpu_vf_error_put(adev, sub_error_code: AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, error_flags: 0, error_data: 0); |
4021 | goto failed; |
4022 | } |
4023 | /* init i2c buses */ |
4024 | if (!amdgpu_device_has_dc_support(adev)) |
4025 | amdgpu_atombios_i2c_init(adev); |
4026 | } |
4027 | } |
4028 | |
4029 | fence_driver_init: |
4030 | /* Fence driver */ |
4031 | r = amdgpu_fence_driver_sw_init(adev); |
4032 | if (r) { |
4033 | dev_err(adev->dev, "amdgpu_fence_driver_sw_init failed\n" ); |
4034 | amdgpu_vf_error_put(adev, sub_error_code: AMDGIM_ERROR_VF_FENCE_INIT_FAIL, error_flags: 0, error_data: 0); |
4035 | goto failed; |
4036 | } |
4037 | |
4038 | /* init the mode config */ |
4039 | drm_mode_config_init(dev: adev_to_drm(adev)); |
4040 | |
4041 | r = amdgpu_device_ip_init(adev); |
4042 | if (r) { |
4043 | dev_err(adev->dev, "amdgpu_device_ip_init failed\n" ); |
4044 | amdgpu_vf_error_put(adev, sub_error_code: AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, error_flags: 0, error_data: 0); |
4045 | goto release_ras_con; |
4046 | } |
4047 | |
4048 | amdgpu_fence_driver_hw_init(adev); |
4049 | |
4050 | dev_info(adev->dev, |
4051 | "SE %d, SH per SE %d, CU per SH %d, active_cu_number %d\n" , |
4052 | adev->gfx.config.max_shader_engines, |
4053 | adev->gfx.config.max_sh_per_se, |
4054 | adev->gfx.config.max_cu_per_sh, |
4055 | adev->gfx.cu_info.number); |
4056 | |
4057 | adev->accel_working = true; |
4058 | |
4059 | amdgpu_vm_check_compute_bug(adev); |
4060 | |
4061 | /* Initialize the buffer migration limit. */ |
4062 | if (amdgpu_moverate >= 0) |
4063 | max_MBps = amdgpu_moverate; |
4064 | else |
4065 | max_MBps = 8; /* Allow 8 MB/s. */ |
4066 | /* Get a log2 for easy divisions. */ |
4067 | adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps)); |
4068 | |
4069 | /* |
4070 | * Register gpu instance before amdgpu_device_enable_mgpu_fan_boost. |
4071 | * Otherwise the mgpu fan boost feature will be skipped due to the |
4072 | * gpu instance is counted less. |
4073 | */ |
4074 | amdgpu_register_gpu_instance(adev); |
4075 | |
4076 | /* enable clockgating, etc. after ib tests, etc. since some blocks require |
4077 | * explicit gating rather than handling it automatically. |
4078 | */ |
4079 | if (!adev->gmc.xgmi.pending_reset) { |
4080 | r = amdgpu_device_ip_late_init(adev); |
4081 | if (r) { |
4082 | dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n" ); |
4083 | amdgpu_vf_error_put(adev, sub_error_code: AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, error_flags: 0, error_data: r); |
4084 | goto release_ras_con; |
4085 | } |
4086 | /* must succeed. */ |
4087 | amdgpu_ras_resume(adev); |
4088 | queue_delayed_work(wq: system_wq, dwork: &adev->delayed_init_work, |
4089 | delay: msecs_to_jiffies(AMDGPU_RESUME_MS)); |
4090 | } |
4091 | |
4092 | if (amdgpu_sriov_vf(adev)) { |
4093 | amdgpu_virt_release_full_gpu(adev, init: true); |
4094 | flush_delayed_work(dwork: &adev->delayed_init_work); |
4095 | } |
4096 | |
4097 | /* |
4098 | * Place those sysfs registering after `late_init`. As some of those |
4099 | * operations performed in `late_init` might affect the sysfs |
4100 | * interfaces creating. |
4101 | */ |
4102 | r = amdgpu_atombios_sysfs_init(adev); |
4103 | if (r) |
4104 | drm_err(&adev->ddev, |
4105 | "registering atombios sysfs failed (%d).\n" , r); |
4106 | |
4107 | r = amdgpu_pm_sysfs_init(adev); |
4108 | if (r) |
4109 | DRM_ERROR("registering pm sysfs failed (%d).\n" , r); |
4110 | |
4111 | r = amdgpu_ucode_sysfs_init(adev); |
4112 | if (r) { |
4113 | adev->ucode_sysfs_en = false; |
4114 | DRM_ERROR("Creating firmware sysfs failed (%d).\n" , r); |
4115 | } else |
4116 | adev->ucode_sysfs_en = true; |
4117 | |
4118 | r = sysfs_create_files(kobj: &adev->dev->kobj, attr: amdgpu_dev_attributes); |
4119 | if (r) |
4120 | dev_err(adev->dev, "Could not create amdgpu device attr\n" ); |
4121 | |
4122 | r = devm_device_add_group(dev: adev->dev, grp: &amdgpu_board_attrs_group); |
4123 | if (r) |
4124 | dev_err(adev->dev, |
4125 | "Could not create amdgpu board attributes\n" ); |
4126 | |
4127 | amdgpu_fru_sysfs_init(adev); |
4128 | |
4129 | if (IS_ENABLED(CONFIG_PERF_EVENTS)) |
4130 | r = amdgpu_pmu_init(adev); |
4131 | if (r) |
4132 | dev_err(adev->dev, "amdgpu_pmu_init failed\n" ); |
4133 | |
4134 | /* Have stored pci confspace at hand for restore in sudden PCI error */ |
4135 | if (amdgpu_device_cache_pci_state(pdev: adev->pdev)) |
4136 | pci_restore_state(dev: pdev); |
4137 | |
4138 | /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */ |
4139 | /* this will fail for cards that aren't VGA class devices, just |
4140 | * ignore it |
4141 | */ |
4142 | if ((adev->pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA) |
4143 | vga_client_register(pdev: adev->pdev, set_decode: amdgpu_device_vga_set_decode); |
4144 | |
4145 | px = amdgpu_device_supports_px(dev: ddev); |
4146 | |
4147 | if (px || (!dev_is_removable(dev: &adev->pdev->dev) && |
4148 | apple_gmux_detect(NULL, NULL))) |
4149 | vga_switcheroo_register_client(dev: adev->pdev, |
4150 | ops: &amdgpu_switcheroo_ops, driver_power_control: px); |
4151 | |
4152 | if (px) |
4153 | vga_switcheroo_init_domain_pm_ops(dev: adev->dev, domain: &adev->vga_pm_domain); |
4154 | |
4155 | if (adev->gmc.xgmi.pending_reset) |
4156 | queue_delayed_work(wq: system_wq, dwork: &mgpu_info.delayed_reset_work, |
4157 | delay: msecs_to_jiffies(AMDGPU_RESUME_MS)); |
4158 | |
4159 | amdgpu_device_check_iommu_direct_map(adev); |
4160 | |
4161 | return 0; |
4162 | |
4163 | release_ras_con: |
4164 | if (amdgpu_sriov_vf(adev)) |
4165 | amdgpu_virt_release_full_gpu(adev, init: true); |
4166 | |
4167 | /* failed in exclusive mode due to timeout */ |
4168 | if (amdgpu_sriov_vf(adev) && |
4169 | !amdgpu_sriov_runtime(adev) && |
4170 | amdgpu_virt_mmio_blocked(adev) && |
4171 | !amdgpu_virt_wait_reset(adev)) { |
4172 | dev_err(adev->dev, "VF exclusive mode timeout\n" ); |
4173 | /* Don't send request since VF is inactive. */ |
4174 | adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME; |
4175 | adev->virt.ops = NULL; |
4176 | r = -EAGAIN; |
4177 | } |
4178 | amdgpu_release_ras_context(adev); |
4179 | |
4180 | failed: |
4181 | amdgpu_vf_error_trans_all(adev); |
4182 | |
4183 | return r; |
4184 | } |
4185 | |
4186 | static void amdgpu_device_unmap_mmio(struct amdgpu_device *adev) |
4187 | { |
4188 | |
4189 | /* Clear all CPU mappings pointing to this device */ |
4190 | unmap_mapping_range(mapping: adev->ddev.anon_inode->i_mapping, holebegin: 0, holelen: 0, even_cows: 1); |
4191 | |
4192 | /* Unmap all mapped bars - Doorbell, registers and VRAM */ |
4193 | amdgpu_doorbell_fini(adev); |
4194 | |
4195 | iounmap(addr: adev->rmmio); |
4196 | adev->rmmio = NULL; |
4197 | if (adev->mman.aper_base_kaddr) |
4198 | iounmap(addr: adev->mman.aper_base_kaddr); |
4199 | adev->mman.aper_base_kaddr = NULL; |
4200 | |
4201 | /* Memory manager related */ |
4202 | if (!adev->gmc.xgmi.connected_to_cpu && !adev->gmc.is_app_apu) { |
4203 | arch_phys_wc_del(handle: adev->gmc.vram_mtrr); |
4204 | arch_io_free_memtype_wc(start: adev->gmc.aper_base, size: adev->gmc.aper_size); |
4205 | } |
4206 | } |
4207 | |
4208 | /** |
4209 | * amdgpu_device_fini_hw - tear down the driver |
4210 | * |
4211 | * @adev: amdgpu_device pointer |
4212 | * |
4213 | * Tear down the driver info (all asics). |
4214 | * Called at driver shutdown. |
4215 | */ |
4216 | void amdgpu_device_fini_hw(struct amdgpu_device *adev) |
4217 | { |
4218 | dev_info(adev->dev, "amdgpu: finishing device.\n" ); |
4219 | flush_delayed_work(dwork: &adev->delayed_init_work); |
4220 | adev->shutdown = true; |
4221 | |
4222 | /* make sure IB test finished before entering exclusive mode |
4223 | * to avoid preemption on IB test |
4224 | */ |
4225 | if (amdgpu_sriov_vf(adev)) { |
4226 | amdgpu_virt_request_full_gpu(adev, init: false); |
4227 | amdgpu_virt_fini_data_exchange(adev); |
4228 | } |
4229 | |
4230 | /* disable all interrupts */ |
4231 | amdgpu_irq_disable_all(adev); |
4232 | if (adev->mode_info.mode_config_initialized) { |
4233 | if (!drm_drv_uses_atomic_modeset(dev: adev_to_drm(adev))) |
4234 | drm_helper_force_disable_all(dev: adev_to_drm(adev)); |
4235 | else |
4236 | drm_atomic_helper_shutdown(dev: adev_to_drm(adev)); |
4237 | } |
4238 | amdgpu_fence_driver_hw_fini(adev); |
4239 | |
4240 | if (adev->mman.initialized) |
4241 | drain_workqueue(wq: adev->mman.bdev.wq); |
4242 | |
4243 | if (adev->pm.sysfs_initialized) |
4244 | amdgpu_pm_sysfs_fini(adev); |
4245 | if (adev->ucode_sysfs_en) |
4246 | amdgpu_ucode_sysfs_fini(adev); |
4247 | sysfs_remove_files(kobj: &adev->dev->kobj, attr: amdgpu_dev_attributes); |
4248 | amdgpu_fru_sysfs_fini(adev); |
4249 | |
4250 | /* disable ras feature must before hw fini */ |
4251 | amdgpu_ras_pre_fini(adev); |
4252 | |
4253 | amdgpu_ttm_set_buffer_funcs_status(adev, enable: false); |
4254 | |
4255 | amdgpu_device_ip_fini_early(adev); |
4256 | |
4257 | amdgpu_irq_fini_hw(adev); |
4258 | |
4259 | if (adev->mman.initialized) |
4260 | ttm_device_clear_dma_mappings(bdev: &adev->mman.bdev); |
4261 | |
4262 | amdgpu_gart_dummy_page_fini(adev); |
4263 | |
4264 | if (drm_dev_is_unplugged(dev: adev_to_drm(adev))) |
4265 | amdgpu_device_unmap_mmio(adev); |
4266 | |
4267 | } |
4268 | |
4269 | void amdgpu_device_fini_sw(struct amdgpu_device *adev) |
4270 | { |
4271 | int idx; |
4272 | bool px; |
4273 | |
4274 | amdgpu_fence_driver_sw_fini(adev); |
4275 | amdgpu_device_ip_fini(adev); |
4276 | amdgpu_ucode_release(fw: &adev->firmware.gpu_info_fw); |
4277 | adev->accel_working = false; |
4278 | dma_fence_put(rcu_dereference_protected(adev->gang_submit, true)); |
4279 | |
4280 | amdgpu_reset_fini(adev); |
4281 | |
4282 | /* free i2c buses */ |
4283 | if (!amdgpu_device_has_dc_support(adev)) |
4284 | amdgpu_i2c_fini(adev); |
4285 | |
4286 | if (amdgpu_emu_mode != 1) |
4287 | amdgpu_atombios_fini(adev); |
4288 | |
4289 | kfree(objp: adev->bios); |
4290 | adev->bios = NULL; |
4291 | |
4292 | kfree(objp: adev->fru_info); |
4293 | adev->fru_info = NULL; |
4294 | |
4295 | px = amdgpu_device_supports_px(dev: adev_to_drm(adev)); |
4296 | |
4297 | if (px || (!dev_is_removable(dev: &adev->pdev->dev) && |
4298 | apple_gmux_detect(NULL, NULL))) |
4299 | vga_switcheroo_unregister_client(dev: adev->pdev); |
4300 | |
4301 | if (px) |
4302 | vga_switcheroo_fini_domain_pm_ops(dev: adev->dev); |
4303 | |
4304 | if ((adev->pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA) |
4305 | vga_client_unregister(pdev: adev->pdev); |
4306 | |
4307 | if (drm_dev_enter(dev: adev_to_drm(adev), idx: &idx)) { |
4308 | |
4309 | iounmap(addr: adev->rmmio); |
4310 | adev->rmmio = NULL; |
4311 | amdgpu_doorbell_fini(adev); |
4312 | drm_dev_exit(idx); |
4313 | } |
4314 | |
4315 | if (IS_ENABLED(CONFIG_PERF_EVENTS)) |
4316 | amdgpu_pmu_fini(adev); |
4317 | if (adev->mman.discovery_bin) |
4318 | amdgpu_discovery_fini(adev); |
4319 | |
4320 | amdgpu_reset_put_reset_domain(domain: adev->reset_domain); |
4321 | adev->reset_domain = NULL; |
4322 | |
4323 | kfree(objp: adev->pci_state); |
4324 | |
4325 | } |
4326 | |
4327 | /** |
4328 | * amdgpu_device_evict_resources - evict device resources |
4329 | * @adev: amdgpu device object |
4330 | * |
4331 | * Evicts all ttm device resources(vram BOs, gart table) from the lru list |
4332 | * of the vram memory type. Mainly used for evicting device resources |
4333 | * at suspend time. |
4334 | * |
4335 | */ |
4336 | static int amdgpu_device_evict_resources(struct amdgpu_device *adev) |
4337 | { |
4338 | int ret; |
4339 | |
4340 | /* No need to evict vram on APUs for suspend to ram or s2idle */ |
4341 | if ((adev->in_s3 || adev->in_s0ix) && (adev->flags & AMD_IS_APU)) |
4342 | return 0; |
4343 | |
4344 | ret = amdgpu_ttm_evict_resources(adev, TTM_PL_VRAM); |
4345 | if (ret) |
4346 | DRM_WARN("evicting device resources failed\n" ); |
4347 | return ret; |
4348 | } |
4349 | |
4350 | /* |
4351 | * Suspend & resume. |
4352 | */ |
4353 | /** |
4354 | * amdgpu_device_prepare - prepare for device suspend |
4355 | * |
4356 | * @dev: drm dev pointer |
4357 | * |
4358 | * Prepare to put the hw in the suspend state (all asics). |
4359 | * Returns 0 for success or an error on failure. |
4360 | * Called at driver suspend. |
4361 | */ |
4362 | int amdgpu_device_prepare(struct drm_device *dev) |
4363 | { |
4364 | struct amdgpu_device *adev = drm_to_adev(ddev: dev); |
4365 | int i, r; |
4366 | |
4367 | if (dev->switch_power_state == DRM_SWITCH_POWER_OFF) |
4368 | return 0; |
4369 | |
4370 | /* Evict the majority of BOs before starting suspend sequence */ |
4371 | r = amdgpu_device_evict_resources(adev); |
4372 | if (r) |
4373 | return r; |
4374 | |
4375 | for (i = 0; i < adev->num_ip_blocks; i++) { |
4376 | if (!adev->ip_blocks[i].status.valid) |
4377 | continue; |
4378 | if (!adev->ip_blocks[i].version->funcs->prepare_suspend) |
4379 | continue; |
4380 | r = adev->ip_blocks[i].version->funcs->prepare_suspend((void *)adev); |
4381 | if (r) |
4382 | return r; |
4383 | } |
4384 | |
4385 | return 0; |
4386 | } |
4387 | |
4388 | /** |
4389 | * amdgpu_device_suspend - initiate device suspend |
4390 | * |
4391 | * @dev: drm dev pointer |
4392 | * @fbcon : notify the fbdev of suspend |
4393 | * |
4394 | * Puts the hw in the suspend state (all asics). |
4395 | * Returns 0 for success or an error on failure. |
4396 | * Called at driver suspend. |
4397 | */ |
4398 | int amdgpu_device_suspend(struct drm_device *dev, bool fbcon) |
4399 | { |
4400 | struct amdgpu_device *adev = drm_to_adev(ddev: dev); |
4401 | int r = 0; |
4402 | |
4403 | if (dev->switch_power_state == DRM_SWITCH_POWER_OFF) |
4404 | return 0; |
4405 | |
4406 | adev->in_suspend = true; |
4407 | |
4408 | if (amdgpu_sriov_vf(adev)) { |
4409 | amdgpu_virt_fini_data_exchange(adev); |
4410 | r = amdgpu_virt_request_full_gpu(adev, init: false); |
4411 | if (r) |
4412 | return r; |
4413 | } |
4414 | |
4415 | if (amdgpu_acpi_smart_shift_update(dev, ss_state: AMDGPU_SS_DEV_D3)) |
4416 | DRM_WARN("smart shift update failed\n" ); |
4417 | |
4418 | if (fbcon) |
4419 | drm_fb_helper_set_suspend_unlocked(fb_helper: adev_to_drm(adev)->fb_helper, suspend: true); |
4420 | |
4421 | cancel_delayed_work_sync(dwork: &adev->delayed_init_work); |
4422 | flush_delayed_work(dwork: &adev->gfx.gfx_off_delay_work); |
4423 | |
4424 | amdgpu_ras_suspend(adev); |
4425 | |
4426 | amdgpu_ttm_set_buffer_funcs_status(adev, enable: false); |
4427 | |
4428 | amdgpu_device_ip_suspend_phase1(adev); |
4429 | |
4430 | if (!adev->in_s0ix) |
4431 | amdgpu_amdkfd_suspend(adev, run_pm: adev->in_runpm); |
4432 | |
4433 | r = amdgpu_device_evict_resources(adev); |
4434 | if (r) |
4435 | return r; |
4436 | |
4437 | amdgpu_fence_driver_hw_fini(adev); |
4438 | |
4439 | amdgpu_device_ip_suspend_phase2(adev); |
4440 | |
4441 | if (amdgpu_sriov_vf(adev)) |
4442 | amdgpu_virt_release_full_gpu(adev, init: false); |
4443 | |
4444 | return 0; |
4445 | } |
4446 | |
4447 | /** |
4448 | * amdgpu_device_resume - initiate device resume |
4449 | * |
4450 | * @dev: drm dev pointer |
4451 | * @fbcon : notify the fbdev of resume |
4452 | * |
4453 | * Bring the hw back to operating state (all asics). |
4454 | * Returns 0 for success or an error on failure. |
4455 | * Called at driver resume. |
4456 | */ |
4457 | int amdgpu_device_resume(struct drm_device *dev, bool fbcon) |
4458 | { |
4459 | struct amdgpu_device *adev = drm_to_adev(ddev: dev); |
4460 | int r = 0; |
4461 | |
4462 | if (amdgpu_sriov_vf(adev)) { |
4463 | r = amdgpu_virt_request_full_gpu(adev, init: true); |
4464 | if (r) |
4465 | return r; |
4466 | } |
4467 | |
4468 | if (dev->switch_power_state == DRM_SWITCH_POWER_OFF) |
4469 | return 0; |
4470 | |
4471 | if (adev->in_s0ix) |
4472 | amdgpu_dpm_gfx_state_change(adev, state: sGpuChangeState_D0Entry); |
4473 | |
4474 | /* post card */ |
4475 | if (amdgpu_device_need_post(adev)) { |
4476 | r = amdgpu_device_asic_init(adev); |
4477 | if (r) |
4478 | dev_err(adev->dev, "amdgpu asic init failed\n" ); |
4479 | } |
4480 | |
4481 | r = amdgpu_device_ip_resume(adev); |
4482 | |
4483 | if (r) { |
4484 | dev_err(adev->dev, "amdgpu_device_ip_resume failed (%d).\n" , r); |
4485 | goto exit; |
4486 | } |
4487 | amdgpu_fence_driver_hw_init(adev); |
4488 | |
4489 | r = amdgpu_device_ip_late_init(adev); |
4490 | if (r) |
4491 | goto exit; |
4492 | |
4493 | queue_delayed_work(wq: system_wq, dwork: &adev->delayed_init_work, |
4494 | delay: msecs_to_jiffies(AMDGPU_RESUME_MS)); |
4495 | |
4496 | if (!adev->in_s0ix) { |
4497 | r = amdgpu_amdkfd_resume(adev, run_pm: adev->in_runpm); |
4498 | if (r) |
4499 | goto exit; |
4500 | } |
4501 | |
4502 | exit: |
4503 | if (amdgpu_sriov_vf(adev)) { |
4504 | amdgpu_virt_init_data_exchange(adev); |
4505 | amdgpu_virt_release_full_gpu(adev, init: true); |
4506 | } |
4507 | |
4508 | if (r) |
4509 | return r; |
4510 | |
4511 | /* Make sure IB tests flushed */ |
4512 | flush_delayed_work(dwork: &adev->delayed_init_work); |
4513 | |
4514 | if (fbcon) |
4515 | drm_fb_helper_set_suspend_unlocked(fb_helper: adev_to_drm(adev)->fb_helper, suspend: false); |
4516 | |
4517 | amdgpu_ras_resume(adev); |
4518 | |
4519 | if (adev->mode_info.num_crtc) { |
4520 | /* |
4521 | * Most of the connector probing functions try to acquire runtime pm |
4522 | * refs to ensure that the GPU is powered on when connector polling is |
4523 | * performed. Since we're calling this from a runtime PM callback, |
4524 | * trying to acquire rpm refs will cause us to deadlock. |
4525 | * |
4526 | * Since we're guaranteed to be holding the rpm lock, it's safe to |
4527 | * temporarily disable the rpm helpers so this doesn't deadlock us. |
4528 | */ |
4529 | #ifdef CONFIG_PM |
4530 | dev->dev->power.disable_depth++; |
4531 | #endif |
4532 | if (!adev->dc_enabled) |
4533 | drm_helper_hpd_irq_event(dev); |
4534 | else |
4535 | drm_kms_helper_hotplug_event(dev); |
4536 | #ifdef CONFIG_PM |
4537 | dev->dev->power.disable_depth--; |
4538 | #endif |
4539 | } |
4540 | adev->in_suspend = false; |
4541 | |
4542 | if (adev->enable_mes) |
4543 | amdgpu_mes_self_test(adev); |
4544 | |
4545 | if (amdgpu_acpi_smart_shift_update(dev, ss_state: AMDGPU_SS_DEV_D0)) |
4546 | DRM_WARN("smart shift update failed\n" ); |
4547 | |
4548 | return 0; |
4549 | } |
4550 | |
4551 | /** |
4552 | * amdgpu_device_ip_check_soft_reset - did soft reset succeed |
4553 | * |
4554 | * @adev: amdgpu_device pointer |
4555 | * |
4556 | * The list of all the hardware IPs that make up the asic is walked and |
4557 | * the check_soft_reset callbacks are run. check_soft_reset determines |
4558 | * if the asic is still hung or not. |
4559 | * Returns true if any of the IPs are still in a hung state, false if not. |
4560 | */ |
4561 | static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev) |
4562 | { |
4563 | int i; |
4564 | bool asic_hang = false; |
4565 | |
4566 | if (amdgpu_sriov_vf(adev)) |
4567 | return true; |
4568 | |
4569 | if (amdgpu_asic_need_full_reset(adev)) |
4570 | return true; |
4571 | |
4572 | for (i = 0; i < adev->num_ip_blocks; i++) { |
4573 | if (!adev->ip_blocks[i].status.valid) |
4574 | continue; |
4575 | if (adev->ip_blocks[i].version->funcs->check_soft_reset) |
4576 | adev->ip_blocks[i].status.hang = |
4577 | adev->ip_blocks[i].version->funcs->check_soft_reset(adev); |
4578 | if (adev->ip_blocks[i].status.hang) { |
4579 | dev_info(adev->dev, "IP block:%s is hung!\n" , adev->ip_blocks[i].version->funcs->name); |
4580 | asic_hang = true; |
4581 | } |
4582 | } |
4583 | return asic_hang; |
4584 | } |
4585 | |
4586 | /** |
4587 | * amdgpu_device_ip_pre_soft_reset - prepare for soft reset |
4588 | * |
4589 | * @adev: amdgpu_device pointer |
4590 | * |
4591 | * The list of all the hardware IPs that make up the asic is walked and the |
4592 | * pre_soft_reset callbacks are run if the block is hung. pre_soft_reset |
4593 | * handles any IP specific hardware or software state changes that are |
4594 | * necessary for a soft reset to succeed. |
4595 | * Returns 0 on success, negative error code on failure. |
4596 | */ |
4597 | static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev) |
4598 | { |
4599 | int i, r = 0; |
4600 | |
4601 | for (i = 0; i < adev->num_ip_blocks; i++) { |
4602 | if (!adev->ip_blocks[i].status.valid) |
4603 | continue; |
4604 | if (adev->ip_blocks[i].status.hang && |
4605 | adev->ip_blocks[i].version->funcs->pre_soft_reset) { |
4606 | r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev); |
4607 | if (r) |
4608 | return r; |
4609 | } |
4610 | } |
4611 | |
4612 | return 0; |
4613 | } |
4614 | |
4615 | /** |
4616 | * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed |
4617 | * |
4618 | * @adev: amdgpu_device pointer |
4619 | * |
4620 | * Some hardware IPs cannot be soft reset. If they are hung, a full gpu |
4621 | * reset is necessary to recover. |
4622 | * Returns true if a full asic reset is required, false if not. |
4623 | */ |
4624 | static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev) |
4625 | { |
4626 | int i; |
4627 | |
4628 | if (amdgpu_asic_need_full_reset(adev)) |
4629 | return true; |
4630 | |
4631 | for (i = 0; i < adev->num_ip_blocks; i++) { |
4632 | if (!adev->ip_blocks[i].status.valid) |
4633 | continue; |
4634 | if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) || |
4635 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) || |
4636 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) || |
4637 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) || |
4638 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) { |
4639 | if (adev->ip_blocks[i].status.hang) { |
4640 | dev_info(adev->dev, "Some block need full reset!\n" ); |
4641 | return true; |
4642 | } |
4643 | } |
4644 | } |
4645 | return false; |
4646 | } |
4647 | |
4648 | /** |
4649 | * amdgpu_device_ip_soft_reset - do a soft reset |
4650 | * |
4651 | * @adev: amdgpu_device pointer |
4652 | * |
4653 | * The list of all the hardware IPs that make up the asic is walked and the |
4654 | * soft_reset callbacks are run if the block is hung. soft_reset handles any |
4655 | * IP specific hardware or software state changes that are necessary to soft |
4656 | * reset the IP. |
4657 | * Returns 0 on success, negative error code on failure. |
4658 | */ |
4659 | static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev) |
4660 | { |
4661 | int i, r = 0; |
4662 | |
4663 | for (i = 0; i < adev->num_ip_blocks; i++) { |
4664 | if (!adev->ip_blocks[i].status.valid) |
4665 | continue; |
4666 | if (adev->ip_blocks[i].status.hang && |
4667 | adev->ip_blocks[i].version->funcs->soft_reset) { |
4668 | r = adev->ip_blocks[i].version->funcs->soft_reset(adev); |
4669 | if (r) |
4670 | return r; |
4671 | } |
4672 | } |
4673 | |
4674 | return 0; |
4675 | } |
4676 | |
4677 | /** |
4678 | * amdgpu_device_ip_post_soft_reset - clean up from soft reset |
4679 | * |
4680 | * @adev: amdgpu_device pointer |
4681 | * |
4682 | * The list of all the hardware IPs that make up the asic is walked and the |
4683 | * post_soft_reset callbacks are run if the asic was hung. post_soft_reset |
4684 | * handles any IP specific hardware or software state changes that are |
4685 | * necessary after the IP has been soft reset. |
4686 | * Returns 0 on success, negative error code on failure. |
4687 | */ |
4688 | static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev) |
4689 | { |
4690 | int i, r = 0; |
4691 | |
4692 | for (i = 0; i < adev->num_ip_blocks; i++) { |
4693 | if (!adev->ip_blocks[i].status.valid) |
4694 | continue; |
4695 | if (adev->ip_blocks[i].status.hang && |
4696 | adev->ip_blocks[i].version->funcs->post_soft_reset) |
4697 | r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev); |
4698 | if (r) |
4699 | return r; |
4700 | } |
4701 | |
4702 | return 0; |
4703 | } |
4704 | |
4705 | /** |
4706 | * amdgpu_device_recover_vram - Recover some VRAM contents |
4707 | * |
4708 | * @adev: amdgpu_device pointer |
4709 | * |
4710 | * Restores the contents of VRAM buffers from the shadows in GTT. Used to |
4711 | * restore things like GPUVM page tables after a GPU reset where |
4712 | * the contents of VRAM might be lost. |
4713 | * |
4714 | * Returns: |
4715 | * 0 on success, negative error code on failure. |
4716 | */ |
4717 | static int amdgpu_device_recover_vram(struct amdgpu_device *adev) |
4718 | { |
4719 | struct dma_fence *fence = NULL, *next = NULL; |
4720 | struct amdgpu_bo *shadow; |
4721 | struct amdgpu_bo_vm *vmbo; |
4722 | long r = 1, tmo; |
4723 | |
4724 | if (amdgpu_sriov_runtime(adev)) |
4725 | tmo = msecs_to_jiffies(m: 8000); |
4726 | else |
4727 | tmo = msecs_to_jiffies(m: 100); |
4728 | |
4729 | dev_info(adev->dev, "recover vram bo from shadow start\n" ); |
4730 | mutex_lock(&adev->shadow_list_lock); |
4731 | list_for_each_entry(vmbo, &adev->shadow_list, shadow_list) { |
4732 | /* If vm is compute context or adev is APU, shadow will be NULL */ |
4733 | if (!vmbo->shadow) |
4734 | continue; |
4735 | shadow = vmbo->shadow; |
4736 | |
4737 | /* No need to recover an evicted BO */ |
4738 | if (shadow->tbo.resource->mem_type != TTM_PL_TT || |
4739 | shadow->tbo.resource->start == AMDGPU_BO_INVALID_OFFSET || |
4740 | shadow->parent->tbo.resource->mem_type != TTM_PL_VRAM) |
4741 | continue; |
4742 | |
4743 | r = amdgpu_bo_restore_shadow(shadow, fence: &next); |
4744 | if (r) |
4745 | break; |
4746 | |
4747 | if (fence) { |
4748 | tmo = dma_fence_wait_timeout(fence, intr: false, timeout: tmo); |
4749 | dma_fence_put(fence); |
4750 | fence = next; |
4751 | if (tmo == 0) { |
4752 | r = -ETIMEDOUT; |
4753 | break; |
4754 | } else if (tmo < 0) { |
4755 | r = tmo; |
4756 | break; |
4757 | } |
4758 | } else { |
4759 | fence = next; |
4760 | } |
4761 | } |
4762 | mutex_unlock(lock: &adev->shadow_list_lock); |
4763 | |
4764 | if (fence) |
4765 | tmo = dma_fence_wait_timeout(fence, intr: false, timeout: tmo); |
4766 | dma_fence_put(fence); |
4767 | |
4768 | if (r < 0 || tmo <= 0) { |
4769 | dev_err(adev->dev, "recover vram bo from shadow failed, r is %ld, tmo is %ld\n" , r, tmo); |
4770 | return -EIO; |
4771 | } |
4772 | |
4773 | dev_info(adev->dev, "recover vram bo from shadow done\n" ); |
4774 | return 0; |
4775 | } |
4776 | |
4777 | |
4778 | /** |
4779 | * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf |
4780 | * |
4781 | * @adev: amdgpu_device pointer |
4782 | * @from_hypervisor: request from hypervisor |
4783 | * |
4784 | * do VF FLR and reinitialize Asic |
4785 | * return 0 means succeeded otherwise failed |
4786 | */ |
4787 | static int amdgpu_device_reset_sriov(struct amdgpu_device *adev, |
4788 | bool from_hypervisor) |
4789 | { |
4790 | int r; |
4791 | struct amdgpu_hive_info *hive = NULL; |
4792 | int retry_limit = 0; |
4793 | |
4794 | retry: |
4795 | amdgpu_amdkfd_pre_reset(adev); |
4796 | |
4797 | if (from_hypervisor) |
4798 | r = amdgpu_virt_request_full_gpu(adev, init: true); |
4799 | else |
4800 | r = amdgpu_virt_reset_gpu(adev); |
4801 | if (r) |
4802 | return r; |
4803 | amdgpu_irq_gpu_reset_resume_helper(adev); |
4804 | |
4805 | /* some sw clean up VF needs to do before recover */ |
4806 | amdgpu_virt_post_reset(adev); |
4807 | |
4808 | /* Resume IP prior to SMC */ |
4809 | r = amdgpu_device_ip_reinit_early_sriov(adev); |
4810 | if (r) |
4811 | goto error; |
4812 | |
4813 | amdgpu_virt_init_data_exchange(adev); |
4814 | |
4815 | r = amdgpu_device_fw_loading(adev); |
4816 | if (r) |
4817 | return r; |
4818 | |
4819 | /* now we are okay to resume SMC/CP/SDMA */ |
4820 | r = amdgpu_device_ip_reinit_late_sriov(adev); |
4821 | if (r) |
4822 | goto error; |
4823 | |
4824 | hive = amdgpu_get_xgmi_hive(adev); |
4825 | /* Update PSP FW topology after reset */ |
4826 | if (hive && adev->gmc.xgmi.num_physical_nodes > 1) |
4827 | r = amdgpu_xgmi_update_topology(hive, adev); |
4828 | |
4829 | if (hive) |
4830 | amdgpu_put_xgmi_hive(hive); |
4831 | |
4832 | if (!r) { |
4833 | r = amdgpu_ib_ring_tests(adev); |
4834 | |
4835 | amdgpu_amdkfd_post_reset(adev); |
4836 | } |
4837 | |
4838 | error: |
4839 | if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) { |
4840 | amdgpu_inc_vram_lost(adev); |
4841 | r = amdgpu_device_recover_vram(adev); |
4842 | } |
4843 | amdgpu_virt_release_full_gpu(adev, init: true); |
4844 | |
4845 | if (AMDGPU_RETRY_SRIOV_RESET(r)) { |
4846 | if (retry_limit < AMDGPU_MAX_RETRY_LIMIT) { |
4847 | retry_limit++; |
4848 | goto retry; |
4849 | } else |
4850 | DRM_ERROR("GPU reset retry is beyond the retry limit\n" ); |
4851 | } |
4852 | |
4853 | return r; |
4854 | } |
4855 | |
4856 | /** |
4857 | * amdgpu_device_has_job_running - check if there is any job in mirror list |
4858 | * |
4859 | * @adev: amdgpu_device pointer |
4860 | * |
4861 | * check if there is any job in mirror list |
4862 | */ |
4863 | bool amdgpu_device_has_job_running(struct amdgpu_device *adev) |
4864 | { |
4865 | int i; |
4866 | struct drm_sched_job *job; |
4867 | |
4868 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { |
4869 | struct amdgpu_ring *ring = adev->rings[i]; |
4870 | |
4871 | if (!ring || !ring->sched.thread) |
4872 | continue; |
4873 | |
4874 | spin_lock(lock: &ring->sched.job_list_lock); |
4875 | job = list_first_entry_or_null(&ring->sched.pending_list, |
4876 | struct drm_sched_job, list); |
4877 | spin_unlock(lock: &ring->sched.job_list_lock); |
4878 | if (job) |
4879 | return true; |
4880 | } |
4881 | return false; |
4882 | } |
4883 | |
4884 | /** |
4885 | * amdgpu_device_should_recover_gpu - check if we should try GPU recovery |
4886 | * |
4887 | * @adev: amdgpu_device pointer |
4888 | * |
4889 | * Check amdgpu_gpu_recovery and SRIOV status to see if we should try to recover |
4890 | * a hung GPU. |
4891 | */ |
4892 | bool amdgpu_device_should_recover_gpu(struct amdgpu_device *adev) |
4893 | { |
4894 | |
4895 | if (amdgpu_gpu_recovery == 0) |
4896 | goto disabled; |
4897 | |
4898 | /* Skip soft reset check in fatal error mode */ |
4899 | if (!amdgpu_ras_is_poison_mode_supported(adev)) |
4900 | return true; |
4901 | |
4902 | if (amdgpu_sriov_vf(adev)) |
4903 | return true; |
4904 | |
4905 | if (amdgpu_gpu_recovery == -1) { |
4906 | switch (adev->asic_type) { |
4907 | #ifdef CONFIG_DRM_AMDGPU_SI |
4908 | case CHIP_VERDE: |
4909 | case CHIP_TAHITI: |
4910 | case CHIP_PITCAIRN: |
4911 | case CHIP_OLAND: |
4912 | case CHIP_HAINAN: |
4913 | #endif |
4914 | #ifdef CONFIG_DRM_AMDGPU_CIK |
4915 | case CHIP_KAVERI: |
4916 | case CHIP_KABINI: |
4917 | case CHIP_MULLINS: |
4918 | #endif |
4919 | case CHIP_CARRIZO: |
4920 | case CHIP_STONEY: |
4921 | case CHIP_CYAN_SKILLFISH: |
4922 | goto disabled; |
4923 | default: |
4924 | break; |
4925 | } |
4926 | } |
4927 | |
4928 | return true; |
4929 | |
4930 | disabled: |
4931 | dev_info(adev->dev, "GPU recovery disabled.\n" ); |
4932 | return false; |
4933 | } |
4934 | |
4935 | int amdgpu_device_mode1_reset(struct amdgpu_device *adev) |
4936 | { |
4937 | u32 i; |
4938 | int ret = 0; |
4939 | |
4940 | amdgpu_atombios_scratch_regs_engine_hung(adev, hung: true); |
4941 | |
4942 | dev_info(adev->dev, "GPU mode1 reset\n" ); |
4943 | |
4944 | /* disable BM */ |
4945 | pci_clear_master(dev: adev->pdev); |
4946 | |
4947 | amdgpu_device_cache_pci_state(pdev: adev->pdev); |
4948 | |
4949 | if (amdgpu_dpm_is_mode1_reset_supported(adev)) { |
4950 | dev_info(adev->dev, "GPU smu mode1 reset\n" ); |
4951 | ret = amdgpu_dpm_mode1_reset(adev); |
4952 | } else { |
4953 | dev_info(adev->dev, "GPU psp mode1 reset\n" ); |
4954 | ret = psp_gpu_reset(adev); |
4955 | } |
4956 | |
4957 | if (ret) |
4958 | goto mode1_reset_failed; |
4959 | |
4960 | amdgpu_device_load_pci_state(pdev: adev->pdev); |
4961 | ret = amdgpu_psp_wait_for_bootloader(adev); |
4962 | if (ret) |
4963 | goto mode1_reset_failed; |
4964 | |
4965 | /* wait for asic to come out of reset */ |
4966 | for (i = 0; i < adev->usec_timeout; i++) { |
4967 | u32 memsize = adev->nbio.funcs->get_memsize(adev); |
4968 | |
4969 | if (memsize != 0xffffffff) |
4970 | break; |
4971 | udelay(1); |
4972 | } |
4973 | |
4974 | if (i >= adev->usec_timeout) { |
4975 | ret = -ETIMEDOUT; |
4976 | goto mode1_reset_failed; |
4977 | } |
4978 | |
4979 | amdgpu_atombios_scratch_regs_engine_hung(adev, hung: false); |
4980 | |
4981 | return 0; |
4982 | |
4983 | mode1_reset_failed: |
4984 | dev_err(adev->dev, "GPU mode1 reset failed\n" ); |
4985 | return ret; |
4986 | } |
4987 | |
4988 | int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, |
4989 | struct amdgpu_reset_context *reset_context) |
4990 | { |
4991 | int i, r = 0; |
4992 | struct amdgpu_job *job = NULL; |
4993 | bool need_full_reset = |
4994 | test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); |
4995 | |
4996 | if (reset_context->reset_req_dev == adev) |
4997 | job = reset_context->job; |
4998 | |
4999 | if (amdgpu_sriov_vf(adev)) { |
5000 | /* stop the data exchange thread */ |
5001 | amdgpu_virt_fini_data_exchange(adev); |
5002 | } |
5003 | |
5004 | amdgpu_fence_driver_isr_toggle(adev, stop: true); |
5005 | |
5006 | /* block all schedulers and reset given job's ring */ |
5007 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { |
5008 | struct amdgpu_ring *ring = adev->rings[i]; |
5009 | |
5010 | if (!ring || !ring->sched.thread) |
5011 | continue; |
5012 | |
5013 | /* Clear job fence from fence drv to avoid force_completion |
5014 | * leave NULL and vm flush fence in fence drv |
5015 | */ |
5016 | amdgpu_fence_driver_clear_job_fences(ring); |
5017 | |
5018 | /* after all hw jobs are reset, hw fence is meaningless, so force_completion */ |
5019 | amdgpu_fence_driver_force_completion(ring); |
5020 | } |
5021 | |
5022 | amdgpu_fence_driver_isr_toggle(adev, stop: false); |
5023 | |
5024 | if (job && job->vm) |
5025 | drm_sched_increase_karma(bad: &job->base); |
5026 | |
5027 | r = amdgpu_reset_prepare_hwcontext(adev, reset_context); |
5028 | /* If reset handler not implemented, continue; otherwise return */ |
5029 | if (r == -EOPNOTSUPP) |
5030 | r = 0; |
5031 | else |
5032 | return r; |
5033 | |
5034 | /* Don't suspend on bare metal if we are not going to HW reset the ASIC */ |
5035 | if (!amdgpu_sriov_vf(adev)) { |
5036 | |
5037 | if (!need_full_reset) |
5038 | need_full_reset = amdgpu_device_ip_need_full_reset(adev); |
5039 | |
5040 | if (!need_full_reset && amdgpu_gpu_recovery && |
5041 | amdgpu_device_ip_check_soft_reset(adev)) { |
5042 | amdgpu_device_ip_pre_soft_reset(adev); |
5043 | r = amdgpu_device_ip_soft_reset(adev); |
5044 | amdgpu_device_ip_post_soft_reset(adev); |
5045 | if (r || amdgpu_device_ip_check_soft_reset(adev)) { |
5046 | dev_info(adev->dev, "soft reset failed, will fallback to full reset!\n" ); |
5047 | need_full_reset = true; |
5048 | } |
5049 | } |
5050 | |
5051 | if (need_full_reset) |
5052 | r = amdgpu_device_ip_suspend(adev); |
5053 | if (need_full_reset) |
5054 | set_bit(nr: AMDGPU_NEED_FULL_RESET, addr: &reset_context->flags); |
5055 | else |
5056 | clear_bit(nr: AMDGPU_NEED_FULL_RESET, |
5057 | addr: &reset_context->flags); |
5058 | } |
5059 | |
5060 | return r; |
5061 | } |
5062 | |
5063 | static int amdgpu_reset_reg_dumps(struct amdgpu_device *adev) |
5064 | { |
5065 | int i; |
5066 | |
5067 | lockdep_assert_held(&adev->reset_domain->sem); |
5068 | |
5069 | for (i = 0; i < adev->reset_info.num_regs; i++) { |
5070 | adev->reset_info.reset_dump_reg_value[i] = |
5071 | RREG32(adev->reset_info.reset_dump_reg_list[i]); |
5072 | |
5073 | trace_amdgpu_reset_reg_dumps(address: adev->reset_info.reset_dump_reg_list[i], |
5074 | value: adev->reset_info.reset_dump_reg_value[i]); |
5075 | } |
5076 | |
5077 | return 0; |
5078 | } |
5079 | |
5080 | int amdgpu_do_asic_reset(struct list_head *device_list_handle, |
5081 | struct amdgpu_reset_context *reset_context) |
5082 | { |
5083 | struct amdgpu_device *tmp_adev = NULL; |
5084 | bool need_full_reset, skip_hw_reset, vram_lost = false; |
5085 | int r = 0; |
5086 | bool gpu_reset_for_dev_remove = 0; |
5087 | |
5088 | /* Try reset handler method first */ |
5089 | tmp_adev = list_first_entry(device_list_handle, struct amdgpu_device, |
5090 | reset_list); |
5091 | amdgpu_reset_reg_dumps(adev: tmp_adev); |
5092 | |
5093 | reset_context->reset_device_list = device_list_handle; |
5094 | r = amdgpu_reset_perform_reset(adev: tmp_adev, reset_context); |
5095 | /* If reset handler not implemented, continue; otherwise return */ |
5096 | if (r == -EOPNOTSUPP) |
5097 | r = 0; |
5098 | else |
5099 | return r; |
5100 | |
5101 | /* Reset handler not implemented, use the default method */ |
5102 | need_full_reset = |
5103 | test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); |
5104 | skip_hw_reset = test_bit(AMDGPU_SKIP_HW_RESET, &reset_context->flags); |
5105 | |
5106 | gpu_reset_for_dev_remove = |
5107 | test_bit(AMDGPU_RESET_FOR_DEVICE_REMOVE, &reset_context->flags) && |
5108 | test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); |
5109 | |
5110 | /* |
5111 | * ASIC reset has to be done on all XGMI hive nodes ASAP |
5112 | * to allow proper links negotiation in FW (within 1 sec) |
5113 | */ |
5114 | if (!skip_hw_reset && need_full_reset) { |
5115 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
5116 | /* For XGMI run all resets in parallel to speed up the process */ |
5117 | if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) { |
5118 | tmp_adev->gmc.xgmi.pending_reset = false; |
5119 | if (!queue_work(wq: system_unbound_wq, work: &tmp_adev->xgmi_reset_work)) |
5120 | r = -EALREADY; |
5121 | } else |
5122 | r = amdgpu_asic_reset(tmp_adev); |
5123 | |
5124 | if (r) { |
5125 | dev_err(tmp_adev->dev, "ASIC reset failed with error, %d for drm dev, %s" , |
5126 | r, adev_to_drm(tmp_adev)->unique); |
5127 | goto out; |
5128 | } |
5129 | } |
5130 | |
5131 | /* For XGMI wait for all resets to complete before proceed */ |
5132 | if (!r) { |
5133 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
5134 | if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) { |
5135 | flush_work(work: &tmp_adev->xgmi_reset_work); |
5136 | r = tmp_adev->asic_reset_res; |
5137 | if (r) |
5138 | break; |
5139 | } |
5140 | } |
5141 | } |
5142 | } |
5143 | |
5144 | if (!r && amdgpu_ras_intr_triggered()) { |
5145 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
5146 | amdgpu_ras_reset_error_count(adev: tmp_adev, block: AMDGPU_RAS_BLOCK__MMHUB); |
5147 | } |
5148 | |
5149 | amdgpu_ras_intr_cleared(); |
5150 | } |
5151 | |
5152 | /* Since the mode1 reset affects base ip blocks, the |
5153 | * phase1 ip blocks need to be resumed. Otherwise there |
5154 | * will be a BIOS signature error and the psp bootloader |
5155 | * can't load kdb on the next amdgpu install. |
5156 | */ |
5157 | if (gpu_reset_for_dev_remove) { |
5158 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) |
5159 | amdgpu_device_ip_resume_phase1(adev: tmp_adev); |
5160 | |
5161 | goto end; |
5162 | } |
5163 | |
5164 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
5165 | if (need_full_reset) { |
5166 | /* post card */ |
5167 | r = amdgpu_device_asic_init(adev: tmp_adev); |
5168 | if (r) { |
5169 | dev_warn(tmp_adev->dev, "asic atom init failed!" ); |
5170 | } else { |
5171 | dev_info(tmp_adev->dev, "GPU reset succeeded, trying to resume\n" ); |
5172 | |
5173 | r = amdgpu_device_ip_resume_phase1(adev: tmp_adev); |
5174 | if (r) |
5175 | goto out; |
5176 | |
5177 | vram_lost = amdgpu_device_check_vram_lost(adev: tmp_adev); |
5178 | |
5179 | amdgpu_coredump(adev: tmp_adev, vram_lost, reset_context); |
5180 | |
5181 | if (vram_lost) { |
5182 | DRM_INFO("VRAM is lost due to GPU reset!\n" ); |
5183 | amdgpu_inc_vram_lost(tmp_adev); |
5184 | } |
5185 | |
5186 | r = amdgpu_device_fw_loading(adev: tmp_adev); |
5187 | if (r) |
5188 | return r; |
5189 | |
5190 | r = amdgpu_xcp_restore_partition_mode( |
5191 | xcp_mgr: tmp_adev->xcp_mgr); |
5192 | if (r) |
5193 | goto out; |
5194 | |
5195 | r = amdgpu_device_ip_resume_phase2(adev: tmp_adev); |
5196 | if (r) |
5197 | goto out; |
5198 | |
5199 | if (tmp_adev->mman.buffer_funcs_ring->sched.ready) |
5200 | amdgpu_ttm_set_buffer_funcs_status(adev: tmp_adev, enable: true); |
5201 | |
5202 | if (vram_lost) |
5203 | amdgpu_device_fill_reset_magic(adev: tmp_adev); |
5204 | |
5205 | /* |
5206 | * Add this ASIC as tracked as reset was already |
5207 | * complete successfully. |
5208 | */ |
5209 | amdgpu_register_gpu_instance(adev: tmp_adev); |
5210 | |
5211 | if (!reset_context->hive && |
5212 | tmp_adev->gmc.xgmi.num_physical_nodes > 1) |
5213 | amdgpu_xgmi_add_device(adev: tmp_adev); |
5214 | |
5215 | r = amdgpu_device_ip_late_init(adev: tmp_adev); |
5216 | if (r) |
5217 | goto out; |
5218 | |
5219 | drm_fb_helper_set_suspend_unlocked(fb_helper: adev_to_drm(adev: tmp_adev)->fb_helper, suspend: false); |
5220 | |
5221 | /* |
5222 | * The GPU enters bad state once faulty pages |
5223 | * by ECC has reached the threshold, and ras |
5224 | * recovery is scheduled next. So add one check |
5225 | * here to break recovery if it indeed exceeds |
5226 | * bad page threshold, and remind user to |
5227 | * retire this GPU or setting one bigger |
5228 | * bad_page_threshold value to fix this once |
5229 | * probing driver again. |
5230 | */ |
5231 | if (!amdgpu_ras_eeprom_check_err_threshold(adev: tmp_adev)) { |
5232 | /* must succeed. */ |
5233 | amdgpu_ras_resume(adev: tmp_adev); |
5234 | } else { |
5235 | r = -EINVAL; |
5236 | goto out; |
5237 | } |
5238 | |
5239 | /* Update PSP FW topology after reset */ |
5240 | if (reset_context->hive && |
5241 | tmp_adev->gmc.xgmi.num_physical_nodes > 1) |
5242 | r = amdgpu_xgmi_update_topology( |
5243 | hive: reset_context->hive, adev: tmp_adev); |
5244 | } |
5245 | } |
5246 | |
5247 | out: |
5248 | if (!r) { |
5249 | amdgpu_irq_gpu_reset_resume_helper(adev: tmp_adev); |
5250 | r = amdgpu_ib_ring_tests(adev: tmp_adev); |
5251 | if (r) { |
5252 | dev_err(tmp_adev->dev, "ib ring test failed (%d).\n" , r); |
5253 | need_full_reset = true; |
5254 | r = -EAGAIN; |
5255 | goto end; |
5256 | } |
5257 | } |
5258 | |
5259 | if (!r) |
5260 | r = amdgpu_device_recover_vram(adev: tmp_adev); |
5261 | else |
5262 | tmp_adev->asic_reset_res = r; |
5263 | } |
5264 | |
5265 | end: |
5266 | if (need_full_reset) |
5267 | set_bit(nr: AMDGPU_NEED_FULL_RESET, addr: &reset_context->flags); |
5268 | else |
5269 | clear_bit(nr: AMDGPU_NEED_FULL_RESET, addr: &reset_context->flags); |
5270 | return r; |
5271 | } |
5272 | |
5273 | static void amdgpu_device_set_mp1_state(struct amdgpu_device *adev) |
5274 | { |
5275 | |
5276 | switch (amdgpu_asic_reset_method(adev)) { |
5277 | case AMD_RESET_METHOD_MODE1: |
5278 | adev->mp1_state = PP_MP1_STATE_SHUTDOWN; |
5279 | break; |
5280 | case AMD_RESET_METHOD_MODE2: |
5281 | adev->mp1_state = PP_MP1_STATE_RESET; |
5282 | break; |
5283 | default: |
5284 | adev->mp1_state = PP_MP1_STATE_NONE; |
5285 | break; |
5286 | } |
5287 | } |
5288 | |
5289 | static void amdgpu_device_unset_mp1_state(struct amdgpu_device *adev) |
5290 | { |
5291 | amdgpu_vf_error_trans_all(adev); |
5292 | adev->mp1_state = PP_MP1_STATE_NONE; |
5293 | } |
5294 | |
5295 | static void amdgpu_device_resume_display_audio(struct amdgpu_device *adev) |
5296 | { |
5297 | struct pci_dev *p = NULL; |
5298 | |
5299 | p = pci_get_domain_bus_and_slot(domain: pci_domain_nr(bus: adev->pdev->bus), |
5300 | bus: adev->pdev->bus->number, devfn: 1); |
5301 | if (p) { |
5302 | pm_runtime_enable(dev: &(p->dev)); |
5303 | pm_runtime_resume(dev: &(p->dev)); |
5304 | } |
5305 | |
5306 | pci_dev_put(dev: p); |
5307 | } |
5308 | |
5309 | static int amdgpu_device_suspend_display_audio(struct amdgpu_device *adev) |
5310 | { |
5311 | enum amd_reset_method reset_method; |
5312 | struct pci_dev *p = NULL; |
5313 | u64 expires; |
5314 | |
5315 | /* |
5316 | * For now, only BACO and mode1 reset are confirmed |
5317 | * to suffer the audio issue without proper suspended. |
5318 | */ |
5319 | reset_method = amdgpu_asic_reset_method(adev); |
5320 | if ((reset_method != AMD_RESET_METHOD_BACO) && |
5321 | (reset_method != AMD_RESET_METHOD_MODE1)) |
5322 | return -EINVAL; |
5323 | |
5324 | p = pci_get_domain_bus_and_slot(domain: pci_domain_nr(bus: adev->pdev->bus), |
5325 | bus: adev->pdev->bus->number, devfn: 1); |
5326 | if (!p) |
5327 | return -ENODEV; |
5328 | |
5329 | expires = pm_runtime_autosuspend_expiration(dev: &(p->dev)); |
5330 | if (!expires) |
5331 | /* |
5332 | * If we cannot get the audio device autosuspend delay, |
5333 | * a fixed 4S interval will be used. Considering 3S is |
5334 | * the audio controller default autosuspend delay setting. |
5335 | * 4S used here is guaranteed to cover that. |
5336 | */ |
5337 | expires = ktime_get_mono_fast_ns() + NSEC_PER_SEC * 4ULL; |
5338 | |
5339 | while (!pm_runtime_status_suspended(dev: &(p->dev))) { |
5340 | if (!pm_runtime_suspend(dev: &(p->dev))) |
5341 | break; |
5342 | |
5343 | if (expires < ktime_get_mono_fast_ns()) { |
5344 | dev_warn(adev->dev, "failed to suspend display audio\n" ); |
5345 | pci_dev_put(dev: p); |
5346 | /* TODO: abort the succeeding gpu reset? */ |
5347 | return -ETIMEDOUT; |
5348 | } |
5349 | } |
5350 | |
5351 | pm_runtime_disable(dev: &(p->dev)); |
5352 | |
5353 | pci_dev_put(dev: p); |
5354 | return 0; |
5355 | } |
5356 | |
5357 | static inline void amdgpu_device_stop_pending_resets(struct amdgpu_device *adev) |
5358 | { |
5359 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); |
5360 | |
5361 | #if defined(CONFIG_DEBUG_FS) |
5362 | if (!amdgpu_sriov_vf(adev)) |
5363 | cancel_work(work: &adev->reset_work); |
5364 | #endif |
5365 | |
5366 | if (adev->kfd.dev) |
5367 | cancel_work(work: &adev->kfd.reset_work); |
5368 | |
5369 | if (amdgpu_sriov_vf(adev)) |
5370 | cancel_work(work: &adev->virt.flr_work); |
5371 | |
5372 | if (con && adev->ras_enabled) |
5373 | cancel_work(work: &con->recovery_work); |
5374 | |
5375 | } |
5376 | |
5377 | /** |
5378 | * amdgpu_device_gpu_recover - reset the asic and recover scheduler |
5379 | * |
5380 | * @adev: amdgpu_device pointer |
5381 | * @job: which job trigger hang |
5382 | * @reset_context: amdgpu reset context pointer |
5383 | * |
5384 | * Attempt to reset the GPU if it has hung (all asics). |
5385 | * Attempt to do soft-reset or full-reset and reinitialize Asic |
5386 | * Returns 0 for success or an error on failure. |
5387 | */ |
5388 | |
5389 | int amdgpu_device_gpu_recover(struct amdgpu_device *adev, |
5390 | struct amdgpu_job *job, |
5391 | struct amdgpu_reset_context *reset_context) |
5392 | { |
5393 | struct list_head device_list, *device_list_handle = NULL; |
5394 | bool job_signaled = false; |
5395 | struct amdgpu_hive_info *hive = NULL; |
5396 | struct amdgpu_device *tmp_adev = NULL; |
5397 | int i, r = 0; |
5398 | bool need_emergency_restart = false; |
5399 | bool audio_suspended = false; |
5400 | bool gpu_reset_for_dev_remove = false; |
5401 | |
5402 | gpu_reset_for_dev_remove = |
5403 | test_bit(AMDGPU_RESET_FOR_DEVICE_REMOVE, &reset_context->flags) && |
5404 | test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); |
5405 | |
5406 | /* |
5407 | * Special case: RAS triggered and full reset isn't supported |
5408 | */ |
5409 | need_emergency_restart = amdgpu_ras_need_emergency_restart(adev); |
5410 | |
5411 | /* |
5412 | * Flush RAM to disk so that after reboot |
5413 | * the user can read log and see why the system rebooted. |
5414 | */ |
5415 | if (need_emergency_restart && amdgpu_ras_get_context(adev) && |
5416 | amdgpu_ras_get_context(adev)->reboot) { |
5417 | DRM_WARN("Emergency reboot." ); |
5418 | |
5419 | ksys_sync_helper(); |
5420 | emergency_restart(); |
5421 | } |
5422 | |
5423 | dev_info(adev->dev, "GPU %s begin!\n" , |
5424 | need_emergency_restart ? "jobs stop" :"reset" ); |
5425 | |
5426 | if (!amdgpu_sriov_vf(adev)) |
5427 | hive = amdgpu_get_xgmi_hive(adev); |
5428 | if (hive) |
5429 | mutex_lock(&hive->hive_lock); |
5430 | |
5431 | reset_context->job = job; |
5432 | reset_context->hive = hive; |
5433 | /* |
5434 | * Build list of devices to reset. |
5435 | * In case we are in XGMI hive mode, resort the device list |
5436 | * to put adev in the 1st position. |
5437 | */ |
5438 | INIT_LIST_HEAD(list: &device_list); |
5439 | if (!amdgpu_sriov_vf(adev) && (adev->gmc.xgmi.num_physical_nodes > 1)) { |
5440 | list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { |
5441 | list_add_tail(new: &tmp_adev->reset_list, head: &device_list); |
5442 | if (gpu_reset_for_dev_remove && adev->shutdown) |
5443 | tmp_adev->shutdown = true; |
5444 | } |
5445 | if (!list_is_first(list: &adev->reset_list, head: &device_list)) |
5446 | list_rotate_to_front(list: &adev->reset_list, head: &device_list); |
5447 | device_list_handle = &device_list; |
5448 | } else { |
5449 | list_add_tail(new: &adev->reset_list, head: &device_list); |
5450 | device_list_handle = &device_list; |
5451 | } |
5452 | |
5453 | /* We need to lock reset domain only once both for XGMI and single device */ |
5454 | tmp_adev = list_first_entry(device_list_handle, struct amdgpu_device, |
5455 | reset_list); |
5456 | amdgpu_device_lock_reset_domain(reset_domain: tmp_adev->reset_domain); |
5457 | |
5458 | /* block all schedulers and reset given job's ring */ |
5459 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
5460 | |
5461 | amdgpu_device_set_mp1_state(adev: tmp_adev); |
5462 | |
5463 | /* |
5464 | * Try to put the audio codec into suspend state |
5465 | * before gpu reset started. |
5466 | * |
5467 | * Due to the power domain of the graphics device |
5468 | * is shared with AZ power domain. Without this, |
5469 | * we may change the audio hardware from behind |
5470 | * the audio driver's back. That will trigger |
5471 | * some audio codec errors. |
5472 | */ |
5473 | if (!amdgpu_device_suspend_display_audio(adev: tmp_adev)) |
5474 | audio_suspended = true; |
5475 | |
5476 | amdgpu_ras_set_error_query_ready(adev: tmp_adev, ready: false); |
5477 | |
5478 | cancel_delayed_work_sync(dwork: &tmp_adev->delayed_init_work); |
5479 | |
5480 | if (!amdgpu_sriov_vf(tmp_adev)) |
5481 | amdgpu_amdkfd_pre_reset(adev: tmp_adev); |
5482 | |
5483 | /* |
5484 | * Mark these ASICs to be reseted as untracked first |
5485 | * And add them back after reset completed |
5486 | */ |
5487 | amdgpu_unregister_gpu_instance(adev: tmp_adev); |
5488 | |
5489 | drm_fb_helper_set_suspend_unlocked(fb_helper: adev_to_drm(adev: tmp_adev)->fb_helper, suspend: true); |
5490 | |
5491 | /* disable ras on ALL IPs */ |
5492 | if (!need_emergency_restart && |
5493 | amdgpu_device_ip_need_full_reset(adev: tmp_adev)) |
5494 | amdgpu_ras_suspend(adev: tmp_adev); |
5495 | |
5496 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { |
5497 | struct amdgpu_ring *ring = tmp_adev->rings[i]; |
5498 | |
5499 | if (!ring || !ring->sched.thread) |
5500 | continue; |
5501 | |
5502 | drm_sched_stop(sched: &ring->sched, bad: job ? &job->base : NULL); |
5503 | |
5504 | if (need_emergency_restart) |
5505 | amdgpu_job_stop_all_jobs_on_sched(sched: &ring->sched); |
5506 | } |
5507 | atomic_inc(v: &tmp_adev->gpu_reset_counter); |
5508 | } |
5509 | |
5510 | if (need_emergency_restart) |
5511 | goto skip_sched_resume; |
5512 | |
5513 | /* |
5514 | * Must check guilty signal here since after this point all old |
5515 | * HW fences are force signaled. |
5516 | * |
5517 | * job->base holds a reference to parent fence |
5518 | */ |
5519 | if (job && dma_fence_is_signaled(fence: &job->hw_fence)) { |
5520 | job_signaled = true; |
5521 | dev_info(adev->dev, "Guilty job already signaled, skipping HW reset" ); |
5522 | goto skip_hw_reset; |
5523 | } |
5524 | |
5525 | retry: /* Rest of adevs pre asic reset from XGMI hive. */ |
5526 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
5527 | if (gpu_reset_for_dev_remove) { |
5528 | /* Workaroud for ASICs need to disable SMC first */ |
5529 | amdgpu_device_smu_fini_early(adev: tmp_adev); |
5530 | } |
5531 | r = amdgpu_device_pre_asic_reset(adev: tmp_adev, reset_context); |
5532 | /*TODO Should we stop ?*/ |
5533 | if (r) { |
5534 | dev_err(tmp_adev->dev, "GPU pre asic reset failed with err, %d for drm dev, %s " , |
5535 | r, adev_to_drm(tmp_adev)->unique); |
5536 | tmp_adev->asic_reset_res = r; |
5537 | } |
5538 | |
5539 | /* |
5540 | * Drop all pending non scheduler resets. Scheduler resets |
5541 | * were already dropped during drm_sched_stop |
5542 | */ |
5543 | amdgpu_device_stop_pending_resets(adev: tmp_adev); |
5544 | } |
5545 | |
5546 | /* Actual ASIC resets if needed.*/ |
5547 | /* Host driver will handle XGMI hive reset for SRIOV */ |
5548 | if (amdgpu_sriov_vf(adev)) { |
5549 | r = amdgpu_device_reset_sriov(adev, from_hypervisor: job ? false : true); |
5550 | if (r) |
5551 | adev->asic_reset_res = r; |
5552 | |
5553 | /* Aldebaran and gfx_11_0_3 support ras in SRIOV, so need resume ras during reset */ |
5554 | if (amdgpu_ip_version(adev, ip: GC_HWIP, inst: 0) == |
5555 | IP_VERSION(9, 4, 2) || |
5556 | amdgpu_ip_version(adev, ip: GC_HWIP, inst: 0) == IP_VERSION(11, 0, 3)) |
5557 | amdgpu_ras_resume(adev); |
5558 | } else { |
5559 | r = amdgpu_do_asic_reset(device_list_handle, reset_context); |
5560 | if (r && r == -EAGAIN) |
5561 | goto retry; |
5562 | |
5563 | if (!r && gpu_reset_for_dev_remove) |
5564 | goto recover_end; |
5565 | } |
5566 | |
5567 | skip_hw_reset: |
5568 | |
5569 | /* Post ASIC reset for all devs .*/ |
5570 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
5571 | |
5572 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { |
5573 | struct amdgpu_ring *ring = tmp_adev->rings[i]; |
5574 | |
5575 | if (!ring || !ring->sched.thread) |
5576 | continue; |
5577 | |
5578 | drm_sched_start(sched: &ring->sched, full_recovery: true); |
5579 | } |
5580 | |
5581 | if (!drm_drv_uses_atomic_modeset(dev: adev_to_drm(adev: tmp_adev)) && !job_signaled) |
5582 | drm_helper_resume_force_mode(dev: adev_to_drm(adev: tmp_adev)); |
5583 | |
5584 | if (tmp_adev->asic_reset_res) |
5585 | r = tmp_adev->asic_reset_res; |
5586 | |
5587 | tmp_adev->asic_reset_res = 0; |
5588 | |
5589 | if (r) { |
5590 | /* bad news, how to tell it to userspace ? */ |
5591 | dev_info(tmp_adev->dev, "GPU reset(%d) failed\n" , atomic_read(&tmp_adev->gpu_reset_counter)); |
5592 | amdgpu_vf_error_put(adev: tmp_adev, sub_error_code: AMDGIM_ERROR_VF_GPU_RESET_FAIL, error_flags: 0, error_data: r); |
5593 | } else { |
5594 | dev_info(tmp_adev->dev, "GPU reset(%d) succeeded!\n" , atomic_read(&tmp_adev->gpu_reset_counter)); |
5595 | if (amdgpu_acpi_smart_shift_update(dev: adev_to_drm(adev: tmp_adev), ss_state: AMDGPU_SS_DEV_D0)) |
5596 | DRM_WARN("smart shift update failed\n" ); |
5597 | } |
5598 | } |
5599 | |
5600 | skip_sched_resume: |
5601 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
5602 | /* unlock kfd: SRIOV would do it separately */ |
5603 | if (!need_emergency_restart && !amdgpu_sriov_vf(tmp_adev)) |
5604 | amdgpu_amdkfd_post_reset(adev: tmp_adev); |
5605 | |
5606 | /* kfd_post_reset will do nothing if kfd device is not initialized, |
5607 | * need to bring up kfd here if it's not be initialized before |
5608 | */ |
5609 | if (!adev->kfd.init_complete) |
5610 | amdgpu_amdkfd_device_init(adev); |
5611 | |
5612 | if (audio_suspended) |
5613 | amdgpu_device_resume_display_audio(adev: tmp_adev); |
5614 | |
5615 | amdgpu_device_unset_mp1_state(adev: tmp_adev); |
5616 | |
5617 | amdgpu_ras_set_error_query_ready(adev: tmp_adev, ready: true); |
5618 | } |
5619 | |
5620 | recover_end: |
5621 | tmp_adev = list_first_entry(device_list_handle, struct amdgpu_device, |
5622 | reset_list); |
5623 | amdgpu_device_unlock_reset_domain(reset_domain: tmp_adev->reset_domain); |
5624 | |
5625 | if (hive) { |
5626 | mutex_unlock(lock: &hive->hive_lock); |
5627 | amdgpu_put_xgmi_hive(hive); |
5628 | } |
5629 | |
5630 | if (r) |
5631 | dev_info(adev->dev, "GPU reset end with ret = %d\n" , r); |
5632 | |
5633 | atomic_set(v: &adev->reset_domain->reset_res, i: r); |
5634 | return r; |
5635 | } |
5636 | |
5637 | /** |
5638 | * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot |
5639 | * |
5640 | * @adev: amdgpu_device pointer |
5641 | * |
5642 | * Fetchs and stores in the driver the PCIE capabilities (gen speed |
5643 | * and lanes) of the slot the device is in. Handles APUs and |
5644 | * virtualized environments where PCIE config space may not be available. |
5645 | */ |
5646 | static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev) |
5647 | { |
5648 | struct pci_dev *pdev; |
5649 | enum pci_bus_speed speed_cap, platform_speed_cap; |
5650 | enum pcie_link_width platform_link_width; |
5651 | |
5652 | if (amdgpu_pcie_gen_cap) |
5653 | adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap; |
5654 | |
5655 | if (amdgpu_pcie_lane_cap) |
5656 | adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap; |
5657 | |
5658 | /* covers APUs as well */ |
5659 | if (pci_is_root_bus(pbus: adev->pdev->bus) && !amdgpu_passthrough(adev)) { |
5660 | if (adev->pm.pcie_gen_mask == 0) |
5661 | adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK; |
5662 | if (adev->pm.pcie_mlw_mask == 0) |
5663 | adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK; |
5664 | return; |
5665 | } |
5666 | |
5667 | if (adev->pm.pcie_gen_mask && adev->pm.pcie_mlw_mask) |
5668 | return; |
5669 | |
5670 | pcie_bandwidth_available(dev: adev->pdev, NULL, |
5671 | speed: &platform_speed_cap, width: &platform_link_width); |
5672 | |
5673 | if (adev->pm.pcie_gen_mask == 0) { |
5674 | /* asic caps */ |
5675 | pdev = adev->pdev; |
5676 | speed_cap = pcie_get_speed_cap(dev: pdev); |
5677 | if (speed_cap == PCI_SPEED_UNKNOWN) { |
5678 | adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 | |
5679 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 | |
5680 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3); |
5681 | } else { |
5682 | if (speed_cap == PCIE_SPEED_32_0GT) |
5683 | adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 | |
5684 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 | |
5685 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 | |
5686 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4 | |
5687 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN5); |
5688 | else if (speed_cap == PCIE_SPEED_16_0GT) |
5689 | adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 | |
5690 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 | |
5691 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 | |
5692 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4); |
5693 | else if (speed_cap == PCIE_SPEED_8_0GT) |
5694 | adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 | |
5695 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 | |
5696 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3); |
5697 | else if (speed_cap == PCIE_SPEED_5_0GT) |
5698 | adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 | |
5699 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2); |
5700 | else |
5701 | adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1; |
5702 | } |
5703 | /* platform caps */ |
5704 | if (platform_speed_cap == PCI_SPEED_UNKNOWN) { |
5705 | adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 | |
5706 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2); |
5707 | } else { |
5708 | if (platform_speed_cap == PCIE_SPEED_32_0GT) |
5709 | adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 | |
5710 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 | |
5711 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 | |
5712 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4 | |
5713 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN5); |
5714 | else if (platform_speed_cap == PCIE_SPEED_16_0GT) |
5715 | adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 | |
5716 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 | |
5717 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 | |
5718 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4); |
5719 | else if (platform_speed_cap == PCIE_SPEED_8_0GT) |
5720 | adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 | |
5721 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 | |
5722 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3); |
5723 | else if (platform_speed_cap == PCIE_SPEED_5_0GT) |
5724 | adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 | |
5725 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2); |
5726 | else |
5727 | adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1; |
5728 | |
5729 | } |
5730 | } |
5731 | if (adev->pm.pcie_mlw_mask == 0) { |
5732 | if (platform_link_width == PCIE_LNK_WIDTH_UNKNOWN) { |
5733 | adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK; |
5734 | } else { |
5735 | switch (platform_link_width) { |
5736 | case PCIE_LNK_X32: |
5737 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 | |
5738 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 | |
5739 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 | |
5740 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 | |
5741 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | |
5742 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | |
5743 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); |
5744 | break; |
5745 | case PCIE_LNK_X16: |
5746 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 | |
5747 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 | |
5748 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 | |
5749 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | |
5750 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | |
5751 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); |
5752 | break; |
5753 | case PCIE_LNK_X12: |
5754 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 | |
5755 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 | |
5756 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | |
5757 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | |
5758 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); |
5759 | break; |
5760 | case PCIE_LNK_X8: |
5761 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 | |
5762 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | |
5763 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | |
5764 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); |
5765 | break; |
5766 | case PCIE_LNK_X4: |
5767 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | |
5768 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | |
5769 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); |
5770 | break; |
5771 | case PCIE_LNK_X2: |
5772 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | |
5773 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); |
5774 | break; |
5775 | case PCIE_LNK_X1: |
5776 | adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1; |
5777 | break; |
5778 | default: |
5779 | break; |
5780 | } |
5781 | } |
5782 | } |
5783 | } |
5784 | |
5785 | /** |
5786 | * amdgpu_device_is_peer_accessible - Check peer access through PCIe BAR |
5787 | * |
5788 | * @adev: amdgpu_device pointer |
5789 | * @peer_adev: amdgpu_device pointer for peer device trying to access @adev |
5790 | * |
5791 | * Return true if @peer_adev can access (DMA) @adev through the PCIe |
5792 | * BAR, i.e. @adev is "large BAR" and the BAR matches the DMA mask of |
5793 | * @peer_adev. |
5794 | */ |
5795 | bool amdgpu_device_is_peer_accessible(struct amdgpu_device *adev, |
5796 | struct amdgpu_device *peer_adev) |
5797 | { |
5798 | #ifdef CONFIG_HSA_AMD_P2P |
5799 | uint64_t address_mask = peer_adev->dev->dma_mask ? |
5800 | ~*peer_adev->dev->dma_mask : ~((1ULL << 32) - 1); |
5801 | resource_size_t aper_limit = |
5802 | adev->gmc.aper_base + adev->gmc.aper_size - 1; |
5803 | bool p2p_access = |
5804 | !adev->gmc.xgmi.connected_to_cpu && |
5805 | !(pci_p2pdma_distance(provider: adev->pdev, client: peer_adev->dev, verbose: false) < 0); |
5806 | |
5807 | return pcie_p2p && p2p_access && (adev->gmc.visible_vram_size && |
5808 | adev->gmc.real_vram_size == adev->gmc.visible_vram_size && |
5809 | !(adev->gmc.aper_base & address_mask || |
5810 | aper_limit & address_mask)); |
5811 | #else |
5812 | return false; |
5813 | #endif |
5814 | } |
5815 | |
5816 | int amdgpu_device_baco_enter(struct drm_device *dev) |
5817 | { |
5818 | struct amdgpu_device *adev = drm_to_adev(ddev: dev); |
5819 | struct amdgpu_ras *ras = amdgpu_ras_get_context(adev); |
5820 | |
5821 | if (!amdgpu_device_supports_baco(dev)) |
5822 | return -ENOTSUPP; |
5823 | |
5824 | if (ras && adev->ras_enabled && |
5825 | adev->nbio.funcs->enable_doorbell_interrupt) |
5826 | adev->nbio.funcs->enable_doorbell_interrupt(adev, false); |
5827 | |
5828 | return amdgpu_dpm_baco_enter(adev); |
5829 | } |
5830 | |
5831 | int amdgpu_device_baco_exit(struct drm_device *dev) |
5832 | { |
5833 | struct amdgpu_device *adev = drm_to_adev(ddev: dev); |
5834 | struct amdgpu_ras *ras = amdgpu_ras_get_context(adev); |
5835 | int ret = 0; |
5836 | |
5837 | if (!amdgpu_device_supports_baco(dev)) |
5838 | return -ENOTSUPP; |
5839 | |
5840 | ret = amdgpu_dpm_baco_exit(adev); |
5841 | if (ret) |
5842 | return ret; |
5843 | |
5844 | if (ras && adev->ras_enabled && |
5845 | adev->nbio.funcs->enable_doorbell_interrupt) |
5846 | adev->nbio.funcs->enable_doorbell_interrupt(adev, true); |
5847 | |
5848 | if (amdgpu_passthrough(adev) && |
5849 | adev->nbio.funcs->clear_doorbell_interrupt) |
5850 | adev->nbio.funcs->clear_doorbell_interrupt(adev); |
5851 | |
5852 | return 0; |
5853 | } |
5854 | |
5855 | /** |
5856 | * amdgpu_pci_error_detected - Called when a PCI error is detected. |
5857 | * @pdev: PCI device struct |
5858 | * @state: PCI channel state |
5859 | * |
5860 | * Description: Called when a PCI error is detected. |
5861 | * |
5862 | * Return: PCI_ERS_RESULT_NEED_RESET or PCI_ERS_RESULT_DISCONNECT. |
5863 | */ |
5864 | pci_ers_result_t amdgpu_pci_error_detected(struct pci_dev *pdev, pci_channel_state_t state) |
5865 | { |
5866 | struct drm_device *dev = pci_get_drvdata(pdev); |
5867 | struct amdgpu_device *adev = drm_to_adev(ddev: dev); |
5868 | int i; |
5869 | |
5870 | DRM_INFO("PCI error: detected callback, state(%d)!!\n" , state); |
5871 | |
5872 | if (adev->gmc.xgmi.num_physical_nodes > 1) { |
5873 | DRM_WARN("No support for XGMI hive yet..." ); |
5874 | return PCI_ERS_RESULT_DISCONNECT; |
5875 | } |
5876 | |
5877 | adev->pci_channel_state = state; |
5878 | |
5879 | switch (state) { |
5880 | case pci_channel_io_normal: |
5881 | return PCI_ERS_RESULT_CAN_RECOVER; |
5882 | /* Fatal error, prepare for slot reset */ |
5883 | case pci_channel_io_frozen: |
5884 | /* |
5885 | * Locking adev->reset_domain->sem will prevent any external access |
5886 | * to GPU during PCI error recovery |
5887 | */ |
5888 | amdgpu_device_lock_reset_domain(reset_domain: adev->reset_domain); |
5889 | amdgpu_device_set_mp1_state(adev); |
5890 | |
5891 | /* |
5892 | * Block any work scheduling as we do for regular GPU reset |
5893 | * for the duration of the recovery |
5894 | */ |
5895 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { |
5896 | struct amdgpu_ring *ring = adev->rings[i]; |
5897 | |
5898 | if (!ring || !ring->sched.thread) |
5899 | continue; |
5900 | |
5901 | drm_sched_stop(sched: &ring->sched, NULL); |
5902 | } |
5903 | atomic_inc(v: &adev->gpu_reset_counter); |
5904 | return PCI_ERS_RESULT_NEED_RESET; |
5905 | case pci_channel_io_perm_failure: |
5906 | /* Permanent error, prepare for device removal */ |
5907 | return PCI_ERS_RESULT_DISCONNECT; |
5908 | } |
5909 | |
5910 | return PCI_ERS_RESULT_NEED_RESET; |
5911 | } |
5912 | |
5913 | /** |
5914 | * amdgpu_pci_mmio_enabled - Enable MMIO and dump debug registers |
5915 | * @pdev: pointer to PCI device |
5916 | */ |
5917 | pci_ers_result_t amdgpu_pci_mmio_enabled(struct pci_dev *pdev) |
5918 | { |
5919 | |
5920 | DRM_INFO("PCI error: mmio enabled callback!!\n" ); |
5921 | |
5922 | /* TODO - dump whatever for debugging purposes */ |
5923 | |
5924 | /* This called only if amdgpu_pci_error_detected returns |
5925 | * PCI_ERS_RESULT_CAN_RECOVER. Read/write to the device still |
5926 | * works, no need to reset slot. |
5927 | */ |
5928 | |
5929 | return PCI_ERS_RESULT_RECOVERED; |
5930 | } |
5931 | |
5932 | /** |
5933 | * amdgpu_pci_slot_reset - Called when PCI slot has been reset. |
5934 | * @pdev: PCI device struct |
5935 | * |
5936 | * Description: This routine is called by the pci error recovery |
5937 | * code after the PCI slot has been reset, just before we |
5938 | * should resume normal operations. |
5939 | */ |
5940 | pci_ers_result_t amdgpu_pci_slot_reset(struct pci_dev *pdev) |
5941 | { |
5942 | struct drm_device *dev = pci_get_drvdata(pdev); |
5943 | struct amdgpu_device *adev = drm_to_adev(ddev: dev); |
5944 | int r, i; |
5945 | struct amdgpu_reset_context reset_context; |
5946 | u32 memsize; |
5947 | struct list_head device_list; |
5948 | |
5949 | DRM_INFO("PCI error: slot reset callback!!\n" ); |
5950 | |
5951 | memset(&reset_context, 0, sizeof(reset_context)); |
5952 | |
5953 | INIT_LIST_HEAD(list: &device_list); |
5954 | list_add_tail(new: &adev->reset_list, head: &device_list); |
5955 | |
5956 | /* wait for asic to come out of reset */ |
5957 | msleep(msecs: 500); |
5958 | |
5959 | /* Restore PCI confspace */ |
5960 | amdgpu_device_load_pci_state(pdev); |
5961 | |
5962 | /* confirm ASIC came out of reset */ |
5963 | for (i = 0; i < adev->usec_timeout; i++) { |
5964 | memsize = amdgpu_asic_get_config_memsize(adev); |
5965 | |
5966 | if (memsize != 0xffffffff) |
5967 | break; |
5968 | udelay(1); |
5969 | } |
5970 | if (memsize == 0xffffffff) { |
5971 | r = -ETIME; |
5972 | goto out; |
5973 | } |
5974 | |
5975 | reset_context.method = AMD_RESET_METHOD_NONE; |
5976 | reset_context.reset_req_dev = adev; |
5977 | set_bit(nr: AMDGPU_NEED_FULL_RESET, addr: &reset_context.flags); |
5978 | set_bit(nr: AMDGPU_SKIP_HW_RESET, addr: &reset_context.flags); |
5979 | |
5980 | adev->no_hw_access = true; |
5981 | r = amdgpu_device_pre_asic_reset(adev, reset_context: &reset_context); |
5982 | adev->no_hw_access = false; |
5983 | if (r) |
5984 | goto out; |
5985 | |
5986 | r = amdgpu_do_asic_reset(device_list_handle: &device_list, reset_context: &reset_context); |
5987 | |
5988 | out: |
5989 | if (!r) { |
5990 | if (amdgpu_device_cache_pci_state(pdev: adev->pdev)) |
5991 | pci_restore_state(dev: adev->pdev); |
5992 | |
5993 | DRM_INFO("PCIe error recovery succeeded\n" ); |
5994 | } else { |
5995 | DRM_ERROR("PCIe error recovery failed, err:%d" , r); |
5996 | amdgpu_device_unset_mp1_state(adev); |
5997 | amdgpu_device_unlock_reset_domain(reset_domain: adev->reset_domain); |
5998 | } |
5999 | |
6000 | return r ? PCI_ERS_RESULT_DISCONNECT : PCI_ERS_RESULT_RECOVERED; |
6001 | } |
6002 | |
6003 | /** |
6004 | * amdgpu_pci_resume() - resume normal ops after PCI reset |
6005 | * @pdev: pointer to PCI device |
6006 | * |
6007 | * Called when the error recovery driver tells us that its |
6008 | * OK to resume normal operation. |
6009 | */ |
6010 | void amdgpu_pci_resume(struct pci_dev *pdev) |
6011 | { |
6012 | struct drm_device *dev = pci_get_drvdata(pdev); |
6013 | struct amdgpu_device *adev = drm_to_adev(ddev: dev); |
6014 | int i; |
6015 | |
6016 | |
6017 | DRM_INFO("PCI error: resume callback!!\n" ); |
6018 | |
6019 | /* Only continue execution for the case of pci_channel_io_frozen */ |
6020 | if (adev->pci_channel_state != pci_channel_io_frozen) |
6021 | return; |
6022 | |
6023 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { |
6024 | struct amdgpu_ring *ring = adev->rings[i]; |
6025 | |
6026 | if (!ring || !ring->sched.thread) |
6027 | continue; |
6028 | |
6029 | drm_sched_start(sched: &ring->sched, full_recovery: true); |
6030 | } |
6031 | |
6032 | amdgpu_device_unset_mp1_state(adev); |
6033 | amdgpu_device_unlock_reset_domain(reset_domain: adev->reset_domain); |
6034 | } |
6035 | |
6036 | bool amdgpu_device_cache_pci_state(struct pci_dev *pdev) |
6037 | { |
6038 | struct drm_device *dev = pci_get_drvdata(pdev); |
6039 | struct amdgpu_device *adev = drm_to_adev(ddev: dev); |
6040 | int r; |
6041 | |
6042 | r = pci_save_state(dev: pdev); |
6043 | if (!r) { |
6044 | kfree(objp: adev->pci_state); |
6045 | |
6046 | adev->pci_state = pci_store_saved_state(dev: pdev); |
6047 | |
6048 | if (!adev->pci_state) { |
6049 | DRM_ERROR("Failed to store PCI saved state" ); |
6050 | return false; |
6051 | } |
6052 | } else { |
6053 | DRM_WARN("Failed to save PCI state, err:%d\n" , r); |
6054 | return false; |
6055 | } |
6056 | |
6057 | return true; |
6058 | } |
6059 | |
6060 | bool amdgpu_device_load_pci_state(struct pci_dev *pdev) |
6061 | { |
6062 | struct drm_device *dev = pci_get_drvdata(pdev); |
6063 | struct amdgpu_device *adev = drm_to_adev(ddev: dev); |
6064 | int r; |
6065 | |
6066 | if (!adev->pci_state) |
6067 | return false; |
6068 | |
6069 | r = pci_load_saved_state(dev: pdev, state: adev->pci_state); |
6070 | |
6071 | if (!r) { |
6072 | pci_restore_state(dev: pdev); |
6073 | } else { |
6074 | DRM_WARN("Failed to load PCI state, err:%d\n" , r); |
6075 | return false; |
6076 | } |
6077 | |
6078 | return true; |
6079 | } |
6080 | |
6081 | void amdgpu_device_flush_hdp(struct amdgpu_device *adev, |
6082 | struct amdgpu_ring *ring) |
6083 | { |
6084 | #ifdef CONFIG_X86_64 |
6085 | if ((adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev)) |
6086 | return; |
6087 | #endif |
6088 | if (adev->gmc.xgmi.connected_to_cpu) |
6089 | return; |
6090 | |
6091 | if (ring && ring->funcs->emit_hdp_flush) |
6092 | amdgpu_ring_emit_hdp_flush(ring); |
6093 | else |
6094 | amdgpu_asic_flush_hdp(adev, ring); |
6095 | } |
6096 | |
6097 | void amdgpu_device_invalidate_hdp(struct amdgpu_device *adev, |
6098 | struct amdgpu_ring *ring) |
6099 | { |
6100 | #ifdef CONFIG_X86_64 |
6101 | if ((adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev)) |
6102 | return; |
6103 | #endif |
6104 | if (adev->gmc.xgmi.connected_to_cpu) |
6105 | return; |
6106 | |
6107 | amdgpu_asic_invalidate_hdp(adev, ring); |
6108 | } |
6109 | |
6110 | int amdgpu_in_reset(struct amdgpu_device *adev) |
6111 | { |
6112 | return atomic_read(v: &adev->reset_domain->in_gpu_reset); |
6113 | } |
6114 | |
6115 | /** |
6116 | * amdgpu_device_halt() - bring hardware to some kind of halt state |
6117 | * |
6118 | * @adev: amdgpu_device pointer |
6119 | * |
6120 | * Bring hardware to some kind of halt state so that no one can touch it |
6121 | * any more. It will help to maintain error context when error occurred. |
6122 | * Compare to a simple hang, the system will keep stable at least for SSH |
6123 | * access. Then it should be trivial to inspect the hardware state and |
6124 | * see what's going on. Implemented as following: |
6125 | * |
6126 | * 1. drm_dev_unplug() makes device inaccessible to user space(IOCTLs, etc), |
6127 | * clears all CPU mappings to device, disallows remappings through page faults |
6128 | * 2. amdgpu_irq_disable_all() disables all interrupts |
6129 | * 3. amdgpu_fence_driver_hw_fini() signals all HW fences |
6130 | * 4. set adev->no_hw_access to avoid potential crashes after setp 5 |
6131 | * 5. amdgpu_device_unmap_mmio() clears all MMIO mappings |
6132 | * 6. pci_disable_device() and pci_wait_for_pending_transaction() |
6133 | * flush any in flight DMA operations |
6134 | */ |
6135 | void amdgpu_device_halt(struct amdgpu_device *adev) |
6136 | { |
6137 | struct pci_dev *pdev = adev->pdev; |
6138 | struct drm_device *ddev = adev_to_drm(adev); |
6139 | |
6140 | amdgpu_xcp_dev_unplug(adev); |
6141 | drm_dev_unplug(dev: ddev); |
6142 | |
6143 | amdgpu_irq_disable_all(adev); |
6144 | |
6145 | amdgpu_fence_driver_hw_fini(adev); |
6146 | |
6147 | adev->no_hw_access = true; |
6148 | |
6149 | amdgpu_device_unmap_mmio(adev); |
6150 | |
6151 | pci_disable_device(dev: pdev); |
6152 | pci_wait_for_pending_transaction(dev: pdev); |
6153 | } |
6154 | |
6155 | u32 amdgpu_device_pcie_port_rreg(struct amdgpu_device *adev, |
6156 | u32 reg) |
6157 | { |
6158 | unsigned long flags, address, data; |
6159 | u32 r; |
6160 | |
6161 | address = adev->nbio.funcs->get_pcie_port_index_offset(adev); |
6162 | data = adev->nbio.funcs->get_pcie_port_data_offset(adev); |
6163 | |
6164 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); |
6165 | WREG32(address, reg * 4); |
6166 | (void)RREG32(address); |
6167 | r = RREG32(data); |
6168 | spin_unlock_irqrestore(lock: &adev->pcie_idx_lock, flags); |
6169 | return r; |
6170 | } |
6171 | |
6172 | void amdgpu_device_pcie_port_wreg(struct amdgpu_device *adev, |
6173 | u32 reg, u32 v) |
6174 | { |
6175 | unsigned long flags, address, data; |
6176 | |
6177 | address = adev->nbio.funcs->get_pcie_port_index_offset(adev); |
6178 | data = adev->nbio.funcs->get_pcie_port_data_offset(adev); |
6179 | |
6180 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); |
6181 | WREG32(address, reg * 4); |
6182 | (void)RREG32(address); |
6183 | WREG32(data, v); |
6184 | (void)RREG32(data); |
6185 | spin_unlock_irqrestore(lock: &adev->pcie_idx_lock, flags); |
6186 | } |
6187 | |
6188 | /** |
6189 | * amdgpu_device_switch_gang - switch to a new gang |
6190 | * @adev: amdgpu_device pointer |
6191 | * @gang: the gang to switch to |
6192 | * |
6193 | * Try to switch to a new gang. |
6194 | * Returns: NULL if we switched to the new gang or a reference to the current |
6195 | * gang leader. |
6196 | */ |
6197 | struct dma_fence *amdgpu_device_switch_gang(struct amdgpu_device *adev, |
6198 | struct dma_fence *gang) |
6199 | { |
6200 | struct dma_fence *old = NULL; |
6201 | |
6202 | do { |
6203 | dma_fence_put(fence: old); |
6204 | rcu_read_lock(); |
6205 | old = dma_fence_get_rcu_safe(fencep: &adev->gang_submit); |
6206 | rcu_read_unlock(); |
6207 | |
6208 | if (old == gang) |
6209 | break; |
6210 | |
6211 | if (!dma_fence_is_signaled(fence: old)) |
6212 | return old; |
6213 | |
6214 | } while (cmpxchg((struct dma_fence __force **)&adev->gang_submit, |
6215 | old, gang) != old); |
6216 | |
6217 | dma_fence_put(fence: old); |
6218 | return NULL; |
6219 | } |
6220 | |
6221 | bool amdgpu_device_has_display_hardware(struct amdgpu_device *adev) |
6222 | { |
6223 | switch (adev->asic_type) { |
6224 | #ifdef CONFIG_DRM_AMDGPU_SI |
6225 | case CHIP_HAINAN: |
6226 | #endif |
6227 | case CHIP_TOPAZ: |
6228 | /* chips with no display hardware */ |
6229 | return false; |
6230 | #ifdef CONFIG_DRM_AMDGPU_SI |
6231 | case CHIP_TAHITI: |
6232 | case CHIP_PITCAIRN: |
6233 | case CHIP_VERDE: |
6234 | case CHIP_OLAND: |
6235 | #endif |
6236 | #ifdef CONFIG_DRM_AMDGPU_CIK |
6237 | case CHIP_BONAIRE: |
6238 | case CHIP_HAWAII: |
6239 | case CHIP_KAVERI: |
6240 | case CHIP_KABINI: |
6241 | case CHIP_MULLINS: |
6242 | #endif |
6243 | case CHIP_TONGA: |
6244 | case CHIP_FIJI: |
6245 | case CHIP_POLARIS10: |
6246 | case CHIP_POLARIS11: |
6247 | case CHIP_POLARIS12: |
6248 | case CHIP_VEGAM: |
6249 | case CHIP_CARRIZO: |
6250 | case CHIP_STONEY: |
6251 | /* chips with display hardware */ |
6252 | return true; |
6253 | default: |
6254 | /* IP discovery */ |
6255 | if (!amdgpu_ip_version(adev, ip: DCE_HWIP, inst: 0) || |
6256 | (adev->harvest_ip_mask & AMD_HARVEST_IP_DMU_MASK)) |
6257 | return false; |
6258 | return true; |
6259 | } |
6260 | } |
6261 | |
6262 | uint32_t amdgpu_device_wait_on_rreg(struct amdgpu_device *adev, |
6263 | uint32_t inst, uint32_t reg_addr, char reg_name[], |
6264 | uint32_t expected_value, uint32_t mask) |
6265 | { |
6266 | uint32_t ret = 0; |
6267 | uint32_t old_ = 0; |
6268 | uint32_t tmp_ = RREG32(reg_addr); |
6269 | uint32_t loop = adev->usec_timeout; |
6270 | |
6271 | while ((tmp_ & (mask)) != (expected_value)) { |
6272 | if (old_ != tmp_) { |
6273 | loop = adev->usec_timeout; |
6274 | old_ = tmp_; |
6275 | } else |
6276 | udelay(1); |
6277 | tmp_ = RREG32(reg_addr); |
6278 | loop--; |
6279 | if (!loop) { |
6280 | DRM_WARN("Register(%d) [%s] failed to reach value 0x%08x != 0x%08xn" , |
6281 | inst, reg_name, (uint32_t)expected_value, |
6282 | (uint32_t)(tmp_ & (mask))); |
6283 | ret = -ETIMEDOUT; |
6284 | break; |
6285 | } |
6286 | } |
6287 | return ret; |
6288 | } |
6289 | |