Commit | Line | Data |
---|---|---|
d38ceaf9 AD |
1 | /* |
2 | * Copyright 2008 Advanced Micro Devices, Inc. | |
3 | * Copyright 2008 Red Hat Inc. | |
4 | * Copyright 2009 Jerome Glisse. | |
5 | * | |
6 | * Permission is hereby granted, free of charge, to any person obtaining a | |
7 | * copy of this software and associated documentation files (the "Software"), | |
8 | * to deal in the Software without restriction, including without limitation | |
9 | * the rights to use, copy, modify, merge, publish, distribute, sublicense, | |
10 | * and/or sell copies of the Software, and to permit persons to whom the | |
11 | * Software is furnished to do so, subject to the following conditions: | |
12 | * | |
13 | * The above copyright notice and this permission notice shall be included in | |
14 | * all copies or substantial portions of the Software. | |
15 | * | |
16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR | |
17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, | |
18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL | |
19 | * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR | |
20 | * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, | |
21 | * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR | |
22 | * OTHER DEALINGS IN THE SOFTWARE. | |
23 | * | |
24 | * Authors: Dave Airlie | |
25 | * Alex Deucher | |
26 | * Jerome Glisse | |
27 | */ | |
b1ddf548 | 28 | #include <linux/power_supply.h> |
0875dc9e | 29 | #include <linux/kthread.h> |
fdf2f6c5 | 30 | #include <linux/module.h> |
d38ceaf9 AD |
31 | #include <linux/console.h> |
32 | #include <linux/slab.h> | |
4a74c38c | 33 | #include <linux/iommu.h> |
901e2be2 | 34 | #include <linux/pci.h> |
3d8785f6 SA |
35 | #include <linux/devcoredump.h> |
36 | #include <generated/utsrelease.h> | |
08a2fd23 | 37 | #include <linux/pci-p2pdma.h> |
d37a3929 | 38 | #include <linux/apple-gmux.h> |
fdf2f6c5 | 39 | |
b7cdb41e | 40 | #include <drm/drm_aperture.h> |
4562236b | 41 | #include <drm/drm_atomic_helper.h> |
973ad627 | 42 | #include <drm/drm_crtc_helper.h> |
45b64fd9 | 43 | #include <drm/drm_fb_helper.h> |
fcd70cd3 | 44 | #include <drm/drm_probe_helper.h> |
d38ceaf9 AD |
45 | #include <drm/amdgpu_drm.h> |
46 | #include <linux/vgaarb.h> | |
47 | #include <linux/vga_switcheroo.h> | |
48 | #include <linux/efi.h> | |
49 | #include "amdgpu.h" | |
f4b373f4 | 50 | #include "amdgpu_trace.h" |
d38ceaf9 AD |
51 | #include "amdgpu_i2c.h" |
52 | #include "atom.h" | |
53 | #include "amdgpu_atombios.h" | |
a5bde2f9 | 54 | #include "amdgpu_atomfirmware.h" |
d0dd7f0c | 55 | #include "amd_pcie.h" |
33f34802 KW |
56 | #ifdef CONFIG_DRM_AMDGPU_SI |
57 | #include "si.h" | |
58 | #endif | |
a2e73f56 AD |
59 | #ifdef CONFIG_DRM_AMDGPU_CIK |
60 | #include "cik.h" | |
61 | #endif | |
aaa36a97 | 62 | #include "vi.h" |
460826e6 | 63 | #include "soc15.h" |
0a5b8c7b | 64 | #include "nv.h" |
d38ceaf9 | 65 | #include "bif/bif_4_1_d.h" |
bec86378 | 66 | #include <linux/firmware.h> |
89041940 | 67 | #include "amdgpu_vf_error.h" |
d38ceaf9 | 68 | |
ba997709 | 69 | #include "amdgpu_amdkfd.h" |
d2f52ac8 | 70 | #include "amdgpu_pm.h" |
d38ceaf9 | 71 | |
5183411b | 72 | #include "amdgpu_xgmi.h" |
c030f2e4 | 73 | #include "amdgpu_ras.h" |
9c7c85f7 | 74 | #include "amdgpu_pmu.h" |
bd607166 | 75 | #include "amdgpu_fru_eeprom.h" |
04442bf7 | 76 | #include "amdgpu_reset.h" |
5183411b | 77 | |
d5ea093e | 78 | #include <linux/suspend.h> |
c6a6e2db | 79 | #include <drm/task_barrier.h> |
3f12acc8 | 80 | #include <linux/pm_runtime.h> |
d5ea093e | 81 | |
f89f8c6b AG |
82 | #include <drm/drm_drv.h> |
83 | ||
3ad5dcfe KHF |
84 | #if IS_ENABLED(CONFIG_X86) |
85 | #include <asm/intel-family.h> | |
86 | #endif | |
87 | ||
e2a75f88 | 88 | MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin"); |
3f76dced | 89 | MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin"); |
2d2e5e7e | 90 | MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin"); |
ad5a67a7 | 91 | MODULE_FIRMWARE("amdgpu/picasso_gpu_info.bin"); |
54c4d17e | 92 | MODULE_FIRMWARE("amdgpu/raven2_gpu_info.bin"); |
65e60f6e | 93 | MODULE_FIRMWARE("amdgpu/arcturus_gpu_info.bin"); |
42b325e5 | 94 | MODULE_FIRMWARE("amdgpu/navi12_gpu_info.bin"); |
e2a75f88 | 95 | |
2dc80b00 | 96 | #define AMDGPU_RESUME_MS 2000 |
7258fa31 SK |
97 | #define AMDGPU_MAX_RETRY_LIMIT 2 |
98 | #define AMDGPU_RETRY_SRIOV_RESET(r) ((r) == -EBUSY || (r) == -ETIMEDOUT || (r) == -EINVAL) | |
2dc80b00 | 99 | |
b7cdb41e ML |
100 | static const struct drm_driver amdgpu_kms_driver; |
101 | ||
050091ab | 102 | const char *amdgpu_asic_name[] = { |
da69c161 KW |
103 | "TAHITI", |
104 | "PITCAIRN", | |
105 | "VERDE", | |
106 | "OLAND", | |
107 | "HAINAN", | |
d38ceaf9 AD |
108 | "BONAIRE", |
109 | "KAVERI", | |
110 | "KABINI", | |
111 | "HAWAII", | |
112 | "MULLINS", | |
113 | "TOPAZ", | |
114 | "TONGA", | |
48299f95 | 115 | "FIJI", |
d38ceaf9 | 116 | "CARRIZO", |
139f4917 | 117 | "STONEY", |
2cc0c0b5 FC |
118 | "POLARIS10", |
119 | "POLARIS11", | |
c4642a47 | 120 | "POLARIS12", |
48ff108d | 121 | "VEGAM", |
d4196f01 | 122 | "VEGA10", |
8fab806a | 123 | "VEGA12", |
956fcddc | 124 | "VEGA20", |
2ca8a5d2 | 125 | "RAVEN", |
d6c3b24e | 126 | "ARCTURUS", |
1eee4228 | 127 | "RENOIR", |
d46b417a | 128 | "ALDEBARAN", |
852a6626 | 129 | "NAVI10", |
d0f56dc2 | 130 | "CYAN_SKILLFISH", |
87dbad02 | 131 | "NAVI14", |
9802f5d7 | 132 | "NAVI12", |
ccaf72d3 | 133 | "SIENNA_CICHLID", |
ddd8fbe7 | 134 | "NAVY_FLOUNDER", |
4f1e9a76 | 135 | "VANGOGH", |
a2468e04 | 136 | "DIMGREY_CAVEFISH", |
6f169591 | 137 | "BEIGE_GOBY", |
ee9236b7 | 138 | "YELLOW_CARP", |
3ae695d6 | 139 | "IP DISCOVERY", |
d38ceaf9 AD |
140 | "LAST", |
141 | }; | |
142 | ||
dcea6e65 KR |
143 | /** |
144 | * DOC: pcie_replay_count | |
145 | * | |
146 | * The amdgpu driver provides a sysfs API for reporting the total number | |
147 | * of PCIe replays (NAKs) | |
148 | * The file pcie_replay_count is used for this and returns the total | |
149 | * number of replays as a sum of the NAKs generated and NAKs received | |
150 | */ | |
151 | ||
152 | static ssize_t amdgpu_device_get_pcie_replay_count(struct device *dev, | |
153 | struct device_attribute *attr, char *buf) | |
154 | { | |
155 | struct drm_device *ddev = dev_get_drvdata(dev); | |
1348969a | 156 | struct amdgpu_device *adev = drm_to_adev(ddev); |
dcea6e65 KR |
157 | uint64_t cnt = amdgpu_asic_get_pcie_replay_count(adev); |
158 | ||
36000c7a | 159 | return sysfs_emit(buf, "%llu\n", cnt); |
dcea6e65 KR |
160 | } |
161 | ||
b8920e1e | 162 | static DEVICE_ATTR(pcie_replay_count, 0444, |
dcea6e65 KR |
163 | amdgpu_device_get_pcie_replay_count, NULL); |
164 | ||
5494d864 AD |
165 | static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev); |
166 | ||
bd607166 | 167 | |
fd496ca8 | 168 | /** |
b98c6299 | 169 | * amdgpu_device_supports_px - Is the device a dGPU with ATPX power control |
fd496ca8 AD |
170 | * |
171 | * @dev: drm_device pointer | |
172 | * | |
b98c6299 | 173 | * Returns true if the device is a dGPU with ATPX power control, |
fd496ca8 AD |
174 | * otherwise return false. |
175 | */ | |
b98c6299 | 176 | bool amdgpu_device_supports_px(struct drm_device *dev) |
fd496ca8 AD |
177 | { |
178 | struct amdgpu_device *adev = drm_to_adev(dev); | |
179 | ||
b98c6299 | 180 | if ((adev->flags & AMD_IS_PX) && !amdgpu_is_atpx_hybrid()) |
fd496ca8 AD |
181 | return true; |
182 | return false; | |
183 | } | |
184 | ||
e3ecdffa | 185 | /** |
0330b848 | 186 | * amdgpu_device_supports_boco - Is the device a dGPU with ACPI power resources |
e3ecdffa AD |
187 | * |
188 | * @dev: drm_device pointer | |
189 | * | |
b98c6299 | 190 | * Returns true if the device is a dGPU with ACPI power control, |
e3ecdffa AD |
191 | * otherwise return false. |
192 | */ | |
31af062a | 193 | bool amdgpu_device_supports_boco(struct drm_device *dev) |
d38ceaf9 | 194 | { |
1348969a | 195 | struct amdgpu_device *adev = drm_to_adev(dev); |
d38ceaf9 | 196 | |
b98c6299 AD |
197 | if (adev->has_pr3 || |
198 | ((adev->flags & AMD_IS_PX) && amdgpu_is_atpx_hybrid())) | |
d38ceaf9 AD |
199 | return true; |
200 | return false; | |
201 | } | |
202 | ||
a69cba42 AD |
203 | /** |
204 | * amdgpu_device_supports_baco - Does the device support BACO | |
205 | * | |
206 | * @dev: drm_device pointer | |
207 | * | |
208 | * Returns true if the device supporte BACO, | |
209 | * otherwise return false. | |
210 | */ | |
211 | bool amdgpu_device_supports_baco(struct drm_device *dev) | |
212 | { | |
1348969a | 213 | struct amdgpu_device *adev = drm_to_adev(dev); |
a69cba42 AD |
214 | |
215 | return amdgpu_asic_supports_baco(adev); | |
216 | } | |
217 | ||
3fa8f89d S |
218 | /** |
219 | * amdgpu_device_supports_smart_shift - Is the device dGPU with | |
220 | * smart shift support | |
221 | * | |
222 | * @dev: drm_device pointer | |
223 | * | |
224 | * Returns true if the device is a dGPU with Smart Shift support, | |
225 | * otherwise returns false. | |
226 | */ | |
227 | bool amdgpu_device_supports_smart_shift(struct drm_device *dev) | |
228 | { | |
229 | return (amdgpu_device_supports_boco(dev) && | |
230 | amdgpu_acpi_is_power_shift_control_supported()); | |
231 | } | |
232 | ||
6e3cd2a9 MCC |
233 | /* |
234 | * VRAM access helper functions | |
235 | */ | |
236 | ||
e35e2b11 | 237 | /** |
048af66b | 238 | * amdgpu_device_mm_access - access vram by MM_INDEX/MM_DATA |
e35e2b11 TY |
239 | * |
240 | * @adev: amdgpu_device pointer | |
241 | * @pos: offset of the buffer in vram | |
242 | * @buf: virtual address of the buffer in system memory | |
243 | * @size: read/write size, sizeof(@buf) must > @size | |
244 | * @write: true - write to vram, otherwise - read from vram | |
245 | */ | |
048af66b KW |
246 | void amdgpu_device_mm_access(struct amdgpu_device *adev, loff_t pos, |
247 | void *buf, size_t size, bool write) | |
e35e2b11 | 248 | { |
e35e2b11 | 249 | unsigned long flags; |
048af66b KW |
250 | uint32_t hi = ~0, tmp = 0; |
251 | uint32_t *data = buf; | |
ce05ac56 | 252 | uint64_t last; |
f89f8c6b | 253 | int idx; |
ce05ac56 | 254 | |
c58a863b | 255 | if (!drm_dev_enter(adev_to_drm(adev), &idx)) |
f89f8c6b | 256 | return; |
9d11eb0d | 257 | |
048af66b KW |
258 | BUG_ON(!IS_ALIGNED(pos, 4) || !IS_ALIGNED(size, 4)); |
259 | ||
260 | spin_lock_irqsave(&adev->mmio_idx_lock, flags); | |
261 | for (last = pos + size; pos < last; pos += 4) { | |
262 | tmp = pos >> 31; | |
263 | ||
264 | WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)pos) | 0x80000000); | |
265 | if (tmp != hi) { | |
266 | WREG32_NO_KIQ(mmMM_INDEX_HI, tmp); | |
267 | hi = tmp; | |
268 | } | |
269 | if (write) | |
270 | WREG32_NO_KIQ(mmMM_DATA, *data++); | |
271 | else | |
272 | *data++ = RREG32_NO_KIQ(mmMM_DATA); | |
273 | } | |
274 | ||
275 | spin_unlock_irqrestore(&adev->mmio_idx_lock, flags); | |
276 | drm_dev_exit(idx); | |
277 | } | |
278 | ||
279 | /** | |
bbe04dec | 280 | * amdgpu_device_aper_access - access vram by vram aperature |
048af66b KW |
281 | * |
282 | * @adev: amdgpu_device pointer | |
283 | * @pos: offset of the buffer in vram | |
284 | * @buf: virtual address of the buffer in system memory | |
285 | * @size: read/write size, sizeof(@buf) must > @size | |
286 | * @write: true - write to vram, otherwise - read from vram | |
287 | * | |
288 | * The return value means how many bytes have been transferred. | |
289 | */ | |
290 | size_t amdgpu_device_aper_access(struct amdgpu_device *adev, loff_t pos, | |
291 | void *buf, size_t size, bool write) | |
292 | { | |
9d11eb0d | 293 | #ifdef CONFIG_64BIT |
048af66b KW |
294 | void __iomem *addr; |
295 | size_t count = 0; | |
296 | uint64_t last; | |
297 | ||
298 | if (!adev->mman.aper_base_kaddr) | |
299 | return 0; | |
300 | ||
9d11eb0d CK |
301 | last = min(pos + size, adev->gmc.visible_vram_size); |
302 | if (last > pos) { | |
048af66b KW |
303 | addr = adev->mman.aper_base_kaddr + pos; |
304 | count = last - pos; | |
9d11eb0d CK |
305 | |
306 | if (write) { | |
307 | memcpy_toio(addr, buf, count); | |
4c452b5c SS |
308 | /* Make sure HDP write cache flush happens without any reordering |
309 | * after the system memory contents are sent over PCIe device | |
310 | */ | |
9d11eb0d | 311 | mb(); |
810085dd | 312 | amdgpu_device_flush_hdp(adev, NULL); |
9d11eb0d | 313 | } else { |
810085dd | 314 | amdgpu_device_invalidate_hdp(adev, NULL); |
4c452b5c SS |
315 | /* Make sure HDP read cache is invalidated before issuing a read |
316 | * to the PCIe device | |
317 | */ | |
9d11eb0d CK |
318 | mb(); |
319 | memcpy_fromio(buf, addr, count); | |
320 | } | |
321 | ||
9d11eb0d | 322 | } |
048af66b KW |
323 | |
324 | return count; | |
325 | #else | |
326 | return 0; | |
9d11eb0d | 327 | #endif |
048af66b | 328 | } |
9d11eb0d | 329 | |
048af66b KW |
330 | /** |
331 | * amdgpu_device_vram_access - read/write a buffer in vram | |
332 | * | |
333 | * @adev: amdgpu_device pointer | |
334 | * @pos: offset of the buffer in vram | |
335 | * @buf: virtual address of the buffer in system memory | |
336 | * @size: read/write size, sizeof(@buf) must > @size | |
337 | * @write: true - write to vram, otherwise - read from vram | |
338 | */ | |
339 | void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos, | |
340 | void *buf, size_t size, bool write) | |
341 | { | |
342 | size_t count; | |
e35e2b11 | 343 | |
048af66b KW |
344 | /* try to using vram apreature to access vram first */ |
345 | count = amdgpu_device_aper_access(adev, pos, buf, size, write); | |
346 | size -= count; | |
347 | if (size) { | |
348 | /* using MM to access rest vram */ | |
349 | pos += count; | |
350 | buf += count; | |
351 | amdgpu_device_mm_access(adev, pos, buf, size, write); | |
e35e2b11 TY |
352 | } |
353 | } | |
354 | ||
d38ceaf9 | 355 | /* |
f7ee1874 | 356 | * register access helper functions. |
d38ceaf9 | 357 | */ |
56b53c0b DL |
358 | |
359 | /* Check if hw access should be skipped because of hotplug or device error */ | |
360 | bool amdgpu_device_skip_hw_access(struct amdgpu_device *adev) | |
361 | { | |
7afefb81 | 362 | if (adev->no_hw_access) |
56b53c0b DL |
363 | return true; |
364 | ||
365 | #ifdef CONFIG_LOCKDEP | |
366 | /* | |
367 | * This is a bit complicated to understand, so worth a comment. What we assert | |
368 | * here is that the GPU reset is not running on another thread in parallel. | |
369 | * | |
370 | * For this we trylock the read side of the reset semaphore, if that succeeds | |
371 | * we know that the reset is not running in paralell. | |
372 | * | |
373 | * If the trylock fails we assert that we are either already holding the read | |
374 | * side of the lock or are the reset thread itself and hold the write side of | |
375 | * the lock. | |
376 | */ | |
377 | if (in_task()) { | |
d0fb18b5 AG |
378 | if (down_read_trylock(&adev->reset_domain->sem)) |
379 | up_read(&adev->reset_domain->sem); | |
56b53c0b | 380 | else |
d0fb18b5 | 381 | lockdep_assert_held(&adev->reset_domain->sem); |
56b53c0b DL |
382 | } |
383 | #endif | |
384 | return false; | |
385 | } | |
386 | ||
e3ecdffa | 387 | /** |
f7ee1874 | 388 | * amdgpu_device_rreg - read a memory mapped IO or indirect register |
e3ecdffa AD |
389 | * |
390 | * @adev: amdgpu_device pointer | |
391 | * @reg: dword aligned register offset | |
392 | * @acc_flags: access flags which require special behavior | |
393 | * | |
394 | * Returns the 32 bit value from the offset specified. | |
395 | */ | |
f7ee1874 HZ |
396 | uint32_t amdgpu_device_rreg(struct amdgpu_device *adev, |
397 | uint32_t reg, uint32_t acc_flags) | |
d38ceaf9 | 398 | { |
f4b373f4 TSD |
399 | uint32_t ret; |
400 | ||
56b53c0b | 401 | if (amdgpu_device_skip_hw_access(adev)) |
bf36b52e AG |
402 | return 0; |
403 | ||
f7ee1874 HZ |
404 | if ((reg * 4) < adev->rmmio_size) { |
405 | if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && | |
406 | amdgpu_sriov_runtime(adev) && | |
d0fb18b5 | 407 | down_read_trylock(&adev->reset_domain->sem)) { |
f7ee1874 | 408 | ret = amdgpu_kiq_rreg(adev, reg); |
d0fb18b5 | 409 | up_read(&adev->reset_domain->sem); |
f7ee1874 HZ |
410 | } else { |
411 | ret = readl(((void __iomem *)adev->rmmio) + (reg * 4)); | |
412 | } | |
413 | } else { | |
414 | ret = adev->pcie_rreg(adev, reg * 4); | |
81202807 | 415 | } |
bc992ba5 | 416 | |
f7ee1874 | 417 | trace_amdgpu_device_rreg(adev->pdev->device, reg, ret); |
e78b579d | 418 | |
f4b373f4 | 419 | return ret; |
d38ceaf9 AD |
420 | } |
421 | ||
421a2a30 ML |
422 | /* |
423 | * MMIO register read with bytes helper functions | |
424 | * @offset:bytes offset from MMIO start | |
b8920e1e | 425 | */ |
421a2a30 | 426 | |
e3ecdffa AD |
427 | /** |
428 | * amdgpu_mm_rreg8 - read a memory mapped IO register | |
429 | * | |
430 | * @adev: amdgpu_device pointer | |
431 | * @offset: byte aligned register offset | |
432 | * | |
433 | * Returns the 8 bit value from the offset specified. | |
434 | */ | |
7cbbc745 AG |
435 | uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset) |
436 | { | |
56b53c0b | 437 | if (amdgpu_device_skip_hw_access(adev)) |
bf36b52e AG |
438 | return 0; |
439 | ||
421a2a30 ML |
440 | if (offset < adev->rmmio_size) |
441 | return (readb(adev->rmmio + offset)); | |
442 | BUG(); | |
443 | } | |
444 | ||
445 | /* | |
446 | * MMIO register write with bytes helper functions | |
447 | * @offset:bytes offset from MMIO start | |
448 | * @value: the value want to be written to the register | |
b8920e1e SS |
449 | */ |
450 | ||
e3ecdffa AD |
451 | /** |
452 | * amdgpu_mm_wreg8 - read a memory mapped IO register | |
453 | * | |
454 | * @adev: amdgpu_device pointer | |
455 | * @offset: byte aligned register offset | |
456 | * @value: 8 bit value to write | |
457 | * | |
458 | * Writes the value specified to the offset specified. | |
459 | */ | |
7cbbc745 AG |
460 | void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value) |
461 | { | |
56b53c0b | 462 | if (amdgpu_device_skip_hw_access(adev)) |
bf36b52e AG |
463 | return; |
464 | ||
421a2a30 ML |
465 | if (offset < adev->rmmio_size) |
466 | writeb(value, adev->rmmio + offset); | |
467 | else | |
468 | BUG(); | |
469 | } | |
470 | ||
e3ecdffa | 471 | /** |
f7ee1874 | 472 | * amdgpu_device_wreg - write to a memory mapped IO or indirect register |
e3ecdffa AD |
473 | * |
474 | * @adev: amdgpu_device pointer | |
475 | * @reg: dword aligned register offset | |
476 | * @v: 32 bit value to write to the register | |
477 | * @acc_flags: access flags which require special behavior | |
478 | * | |
479 | * Writes the value specified to the offset specified. | |
480 | */ | |
f7ee1874 HZ |
481 | void amdgpu_device_wreg(struct amdgpu_device *adev, |
482 | uint32_t reg, uint32_t v, | |
483 | uint32_t acc_flags) | |
d38ceaf9 | 484 | { |
56b53c0b | 485 | if (amdgpu_device_skip_hw_access(adev)) |
bf36b52e AG |
486 | return; |
487 | ||
f7ee1874 HZ |
488 | if ((reg * 4) < adev->rmmio_size) { |
489 | if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && | |
490 | amdgpu_sriov_runtime(adev) && | |
d0fb18b5 | 491 | down_read_trylock(&adev->reset_domain->sem)) { |
f7ee1874 | 492 | amdgpu_kiq_wreg(adev, reg, v); |
d0fb18b5 | 493 | up_read(&adev->reset_domain->sem); |
f7ee1874 HZ |
494 | } else { |
495 | writel(v, ((void __iomem *)adev->rmmio) + (reg * 4)); | |
496 | } | |
497 | } else { | |
498 | adev->pcie_wreg(adev, reg * 4, v); | |
81202807 | 499 | } |
bc992ba5 | 500 | |
f7ee1874 | 501 | trace_amdgpu_device_wreg(adev->pdev->device, reg, v); |
2e0cc4d4 | 502 | } |
d38ceaf9 | 503 | |
03f2abb0 | 504 | /** |
4cc9f86f | 505 | * amdgpu_mm_wreg_mmio_rlc - write register either with direct/indirect mmio or with RLC path if in range |
2e0cc4d4 | 506 | * |
71579346 RB |
507 | * @adev: amdgpu_device pointer |
508 | * @reg: mmio/rlc register | |
509 | * @v: value to write | |
510 | * | |
511 | * this function is invoked only for the debugfs register access | |
03f2abb0 | 512 | */ |
f7ee1874 | 513 | void amdgpu_mm_wreg_mmio_rlc(struct amdgpu_device *adev, |
8ed49dd1 VL |
514 | uint32_t reg, uint32_t v, |
515 | uint32_t xcc_id) | |
2e0cc4d4 | 516 | { |
56b53c0b | 517 | if (amdgpu_device_skip_hw_access(adev)) |
bf36b52e AG |
518 | return; |
519 | ||
2e0cc4d4 | 520 | if (amdgpu_sriov_fullaccess(adev) && |
f7ee1874 HZ |
521 | adev->gfx.rlc.funcs && |
522 | adev->gfx.rlc.funcs->is_rlcg_access_range) { | |
2e0cc4d4 | 523 | if (adev->gfx.rlc.funcs->is_rlcg_access_range(adev, reg)) |
8ed49dd1 | 524 | return amdgpu_sriov_wreg(adev, reg, v, 0, 0, xcc_id); |
4cc9f86f TSD |
525 | } else if ((reg * 4) >= adev->rmmio_size) { |
526 | adev->pcie_wreg(adev, reg * 4, v); | |
f7ee1874 HZ |
527 | } else { |
528 | writel(v, ((void __iomem *)adev->rmmio) + (reg * 4)); | |
47ed4e1c | 529 | } |
d38ceaf9 AD |
530 | } |
531 | ||
1bba3683 HZ |
532 | /** |
533 | * amdgpu_device_indirect_rreg - read an indirect register | |
534 | * | |
535 | * @adev: amdgpu_device pointer | |
22f453fb | 536 | * @reg_addr: indirect register address to read from |
1bba3683 HZ |
537 | * |
538 | * Returns the value of indirect register @reg_addr | |
539 | */ | |
540 | u32 amdgpu_device_indirect_rreg(struct amdgpu_device *adev, | |
1bba3683 HZ |
541 | u32 reg_addr) |
542 | { | |
65ba96e9 | 543 | unsigned long flags, pcie_index, pcie_data; |
1bba3683 HZ |
544 | void __iomem *pcie_index_offset; |
545 | void __iomem *pcie_data_offset; | |
65ba96e9 HZ |
546 | u32 r; |
547 | ||
548 | pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev); | |
549 | pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev); | |
1bba3683 HZ |
550 | |
551 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); | |
552 | pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4; | |
553 | pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4; | |
554 | ||
555 | writel(reg_addr, pcie_index_offset); | |
556 | readl(pcie_index_offset); | |
557 | r = readl(pcie_data_offset); | |
558 | spin_unlock_irqrestore(&adev->pcie_idx_lock, flags); | |
559 | ||
560 | return r; | |
561 | } | |
562 | ||
0c552ed3 LM |
563 | u32 amdgpu_device_indirect_rreg_ext(struct amdgpu_device *adev, |
564 | u64 reg_addr) | |
565 | { | |
566 | unsigned long flags, pcie_index, pcie_index_hi, pcie_data; | |
567 | u32 r; | |
568 | void __iomem *pcie_index_offset; | |
569 | void __iomem *pcie_index_hi_offset; | |
570 | void __iomem *pcie_data_offset; | |
571 | ||
572 | pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev); | |
573 | pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev); | |
574 | if (adev->nbio.funcs->get_pcie_index_hi_offset) | |
575 | pcie_index_hi = adev->nbio.funcs->get_pcie_index_hi_offset(adev); | |
576 | else | |
577 | pcie_index_hi = 0; | |
578 | ||
579 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); | |
580 | pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4; | |
581 | pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4; | |
582 | if (pcie_index_hi != 0) | |
583 | pcie_index_hi_offset = (void __iomem *)adev->rmmio + | |
584 | pcie_index_hi * 4; | |
585 | ||
586 | writel(reg_addr, pcie_index_offset); | |
587 | readl(pcie_index_offset); | |
588 | if (pcie_index_hi != 0) { | |
589 | writel((reg_addr >> 32) & 0xff, pcie_index_hi_offset); | |
590 | readl(pcie_index_hi_offset); | |
591 | } | |
592 | r = readl(pcie_data_offset); | |
593 | ||
594 | /* clear the high bits */ | |
595 | if (pcie_index_hi != 0) { | |
596 | writel(0, pcie_index_hi_offset); | |
597 | readl(pcie_index_hi_offset); | |
598 | } | |
599 | ||
600 | spin_unlock_irqrestore(&adev->pcie_idx_lock, flags); | |
601 | ||
602 | return r; | |
603 | } | |
604 | ||
1bba3683 HZ |
605 | /** |
606 | * amdgpu_device_indirect_rreg64 - read a 64bits indirect register | |
607 | * | |
608 | * @adev: amdgpu_device pointer | |
22f453fb | 609 | * @reg_addr: indirect register address to read from |
1bba3683 HZ |
610 | * |
611 | * Returns the value of indirect register @reg_addr | |
612 | */ | |
613 | u64 amdgpu_device_indirect_rreg64(struct amdgpu_device *adev, | |
1bba3683 HZ |
614 | u32 reg_addr) |
615 | { | |
65ba96e9 | 616 | unsigned long flags, pcie_index, pcie_data; |
1bba3683 HZ |
617 | void __iomem *pcie_index_offset; |
618 | void __iomem *pcie_data_offset; | |
65ba96e9 HZ |
619 | u64 r; |
620 | ||
621 | pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev); | |
622 | pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev); | |
1bba3683 HZ |
623 | |
624 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); | |
625 | pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4; | |
626 | pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4; | |
627 | ||
628 | /* read low 32 bits */ | |
629 | writel(reg_addr, pcie_index_offset); | |
630 | readl(pcie_index_offset); | |
631 | r = readl(pcie_data_offset); | |
632 | /* read high 32 bits */ | |
633 | writel(reg_addr + 4, pcie_index_offset); | |
634 | readl(pcie_index_offset); | |
635 | r |= ((u64)readl(pcie_data_offset) << 32); | |
636 | spin_unlock_irqrestore(&adev->pcie_idx_lock, flags); | |
637 | ||
638 | return r; | |
639 | } | |
640 | ||
641 | /** | |
642 | * amdgpu_device_indirect_wreg - write an indirect register address | |
643 | * | |
644 | * @adev: amdgpu_device pointer | |
1bba3683 HZ |
645 | * @reg_addr: indirect register offset |
646 | * @reg_data: indirect register data | |
647 | * | |
648 | */ | |
649 | void amdgpu_device_indirect_wreg(struct amdgpu_device *adev, | |
1bba3683 HZ |
650 | u32 reg_addr, u32 reg_data) |
651 | { | |
65ba96e9 | 652 | unsigned long flags, pcie_index, pcie_data; |
1bba3683 HZ |
653 | void __iomem *pcie_index_offset; |
654 | void __iomem *pcie_data_offset; | |
655 | ||
65ba96e9 HZ |
656 | pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev); |
657 | pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev); | |
658 | ||
1bba3683 HZ |
659 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); |
660 | pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4; | |
661 | pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4; | |
662 | ||
663 | writel(reg_addr, pcie_index_offset); | |
664 | readl(pcie_index_offset); | |
665 | writel(reg_data, pcie_data_offset); | |
666 | readl(pcie_data_offset); | |
667 | spin_unlock_irqrestore(&adev->pcie_idx_lock, flags); | |
668 | } | |
669 | ||
0c552ed3 LM |
670 | void amdgpu_device_indirect_wreg_ext(struct amdgpu_device *adev, |
671 | u64 reg_addr, u32 reg_data) | |
672 | { | |
673 | unsigned long flags, pcie_index, pcie_index_hi, pcie_data; | |
674 | void __iomem *pcie_index_offset; | |
675 | void __iomem *pcie_index_hi_offset; | |
676 | void __iomem *pcie_data_offset; | |
677 | ||
678 | pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev); | |
679 | pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev); | |
680 | if (adev->nbio.funcs->get_pcie_index_hi_offset) | |
681 | pcie_index_hi = adev->nbio.funcs->get_pcie_index_hi_offset(adev); | |
682 | else | |
683 | pcie_index_hi = 0; | |
684 | ||
685 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); | |
686 | pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4; | |
687 | pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4; | |
688 | if (pcie_index_hi != 0) | |
689 | pcie_index_hi_offset = (void __iomem *)adev->rmmio + | |
690 | pcie_index_hi * 4; | |
691 | ||
692 | writel(reg_addr, pcie_index_offset); | |
693 | readl(pcie_index_offset); | |
694 | if (pcie_index_hi != 0) { | |
695 | writel((reg_addr >> 32) & 0xff, pcie_index_hi_offset); | |
696 | readl(pcie_index_hi_offset); | |
697 | } | |
698 | writel(reg_data, pcie_data_offset); | |
699 | readl(pcie_data_offset); | |
700 | ||
701 | /* clear the high bits */ | |
702 | if (pcie_index_hi != 0) { | |
703 | writel(0, pcie_index_hi_offset); | |
704 | readl(pcie_index_hi_offset); | |
705 | } | |
706 | ||
707 | spin_unlock_irqrestore(&adev->pcie_idx_lock, flags); | |
708 | } | |
709 | ||
1bba3683 HZ |
710 | /** |
711 | * amdgpu_device_indirect_wreg64 - write a 64bits indirect register address | |
712 | * | |
713 | * @adev: amdgpu_device pointer | |
1bba3683 HZ |
714 | * @reg_addr: indirect register offset |
715 | * @reg_data: indirect register data | |
716 | * | |
717 | */ | |
718 | void amdgpu_device_indirect_wreg64(struct amdgpu_device *adev, | |
1bba3683 HZ |
719 | u32 reg_addr, u64 reg_data) |
720 | { | |
65ba96e9 | 721 | unsigned long flags, pcie_index, pcie_data; |
1bba3683 HZ |
722 | void __iomem *pcie_index_offset; |
723 | void __iomem *pcie_data_offset; | |
724 | ||
65ba96e9 HZ |
725 | pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev); |
726 | pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev); | |
727 | ||
1bba3683 HZ |
728 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); |
729 | pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4; | |
730 | pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4; | |
731 | ||
732 | /* write low 32 bits */ | |
733 | writel(reg_addr, pcie_index_offset); | |
734 | readl(pcie_index_offset); | |
735 | writel((u32)(reg_data & 0xffffffffULL), pcie_data_offset); | |
736 | readl(pcie_data_offset); | |
737 | /* write high 32 bits */ | |
738 | writel(reg_addr + 4, pcie_index_offset); | |
739 | readl(pcie_index_offset); | |
740 | writel((u32)(reg_data >> 32), pcie_data_offset); | |
741 | readl(pcie_data_offset); | |
742 | spin_unlock_irqrestore(&adev->pcie_idx_lock, flags); | |
743 | } | |
744 | ||
dabc114e HZ |
745 | /** |
746 | * amdgpu_device_get_rev_id - query device rev_id | |
747 | * | |
748 | * @adev: amdgpu_device pointer | |
749 | * | |
750 | * Return device rev_id | |
751 | */ | |
752 | u32 amdgpu_device_get_rev_id(struct amdgpu_device *adev) | |
753 | { | |
754 | return adev->nbio.funcs->get_rev_id(adev); | |
755 | } | |
756 | ||
d38ceaf9 AD |
757 | /** |
758 | * amdgpu_invalid_rreg - dummy reg read function | |
759 | * | |
982a820b | 760 | * @adev: amdgpu_device pointer |
d38ceaf9 AD |
761 | * @reg: offset of register |
762 | * | |
763 | * Dummy register read function. Used for register blocks | |
764 | * that certain asics don't have (all asics). | |
765 | * Returns the value in the register. | |
766 | */ | |
767 | static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg) | |
768 | { | |
769 | DRM_ERROR("Invalid callback to read register 0x%04X\n", reg); | |
770 | BUG(); | |
771 | return 0; | |
772 | } | |
773 | ||
0c552ed3 LM |
774 | static uint32_t amdgpu_invalid_rreg_ext(struct amdgpu_device *adev, uint64_t reg) |
775 | { | |
776 | DRM_ERROR("Invalid callback to read register 0x%llX\n", reg); | |
777 | BUG(); | |
778 | return 0; | |
779 | } | |
780 | ||
d38ceaf9 AD |
781 | /** |
782 | * amdgpu_invalid_wreg - dummy reg write function | |
783 | * | |
982a820b | 784 | * @adev: amdgpu_device pointer |
d38ceaf9 AD |
785 | * @reg: offset of register |
786 | * @v: value to write to the register | |
787 | * | |
788 | * Dummy register read function. Used for register blocks | |
789 | * that certain asics don't have (all asics). | |
790 | */ | |
791 | static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v) | |
792 | { | |
793 | DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n", | |
794 | reg, v); | |
795 | BUG(); | |
796 | } | |
797 | ||
0c552ed3 LM |
798 | static void amdgpu_invalid_wreg_ext(struct amdgpu_device *adev, uint64_t reg, uint32_t v) |
799 | { | |
800 | DRM_ERROR("Invalid callback to write register 0x%llX with 0x%08X\n", | |
801 | reg, v); | |
802 | BUG(); | |
803 | } | |
804 | ||
4fa1c6a6 TZ |
805 | /** |
806 | * amdgpu_invalid_rreg64 - dummy 64 bit reg read function | |
807 | * | |
982a820b | 808 | * @adev: amdgpu_device pointer |
4fa1c6a6 TZ |
809 | * @reg: offset of register |
810 | * | |
811 | * Dummy register read function. Used for register blocks | |
812 | * that certain asics don't have (all asics). | |
813 | * Returns the value in the register. | |
814 | */ | |
815 | static uint64_t amdgpu_invalid_rreg64(struct amdgpu_device *adev, uint32_t reg) | |
816 | { | |
817 | DRM_ERROR("Invalid callback to read 64 bit register 0x%04X\n", reg); | |
818 | BUG(); | |
819 | return 0; | |
820 | } | |
821 | ||
822 | /** | |
823 | * amdgpu_invalid_wreg64 - dummy reg write function | |
824 | * | |
982a820b | 825 | * @adev: amdgpu_device pointer |
4fa1c6a6 TZ |
826 | * @reg: offset of register |
827 | * @v: value to write to the register | |
828 | * | |
829 | * Dummy register read function. Used for register blocks | |
830 | * that certain asics don't have (all asics). | |
831 | */ | |
832 | static void amdgpu_invalid_wreg64(struct amdgpu_device *adev, uint32_t reg, uint64_t v) | |
833 | { | |
834 | DRM_ERROR("Invalid callback to write 64 bit register 0x%04X with 0x%08llX\n", | |
835 | reg, v); | |
836 | BUG(); | |
837 | } | |
838 | ||
d38ceaf9 AD |
839 | /** |
840 | * amdgpu_block_invalid_rreg - dummy reg read function | |
841 | * | |
982a820b | 842 | * @adev: amdgpu_device pointer |
d38ceaf9 AD |
843 | * @block: offset of instance |
844 | * @reg: offset of register | |
845 | * | |
846 | * Dummy register read function. Used for register blocks | |
847 | * that certain asics don't have (all asics). | |
848 | * Returns the value in the register. | |
849 | */ | |
850 | static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev, | |
851 | uint32_t block, uint32_t reg) | |
852 | { | |
853 | DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n", | |
854 | reg, block); | |
855 | BUG(); | |
856 | return 0; | |
857 | } | |
858 | ||
859 | /** | |
860 | * amdgpu_block_invalid_wreg - dummy reg write function | |
861 | * | |
982a820b | 862 | * @adev: amdgpu_device pointer |
d38ceaf9 AD |
863 | * @block: offset of instance |
864 | * @reg: offset of register | |
865 | * @v: value to write to the register | |
866 | * | |
867 | * Dummy register read function. Used for register blocks | |
868 | * that certain asics don't have (all asics). | |
869 | */ | |
870 | static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev, | |
871 | uint32_t block, | |
872 | uint32_t reg, uint32_t v) | |
873 | { | |
874 | DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n", | |
875 | reg, block, v); | |
876 | BUG(); | |
877 | } | |
878 | ||
4d2997ab AD |
879 | /** |
880 | * amdgpu_device_asic_init - Wrapper for atom asic_init | |
881 | * | |
982a820b | 882 | * @adev: amdgpu_device pointer |
4d2997ab AD |
883 | * |
884 | * Does any asic specific work and then calls atom asic init. | |
885 | */ | |
886 | static int amdgpu_device_asic_init(struct amdgpu_device *adev) | |
887 | { | |
7656168a LL |
888 | int ret; |
889 | ||
4d2997ab AD |
890 | amdgpu_asic_pre_asic_init(adev); |
891 | ||
5db392a0 | 892 | if (adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 3) || |
7656168a LL |
893 | adev->ip_versions[GC_HWIP][0] >= IP_VERSION(11, 0, 0)) { |
894 | amdgpu_psp_wait_for_bootloader(adev); | |
895 | ret = amdgpu_atomfirmware_asic_init(adev, true); | |
896 | return ret; | |
897 | } else { | |
85d1bcc6 | 898 | return amdgpu_atom_asic_init(adev->mode_info.atom_context); |
7656168a LL |
899 | } |
900 | ||
901 | return 0; | |
4d2997ab AD |
902 | } |
903 | ||
e3ecdffa | 904 | /** |
7ccfd79f | 905 | * amdgpu_device_mem_scratch_init - allocate the VRAM scratch page |
e3ecdffa | 906 | * |
982a820b | 907 | * @adev: amdgpu_device pointer |
e3ecdffa AD |
908 | * |
909 | * Allocates a scratch page of VRAM for use by various things in the | |
910 | * driver. | |
911 | */ | |
7ccfd79f | 912 | static int amdgpu_device_mem_scratch_init(struct amdgpu_device *adev) |
d38ceaf9 | 913 | { |
7ccfd79f CK |
914 | return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE, PAGE_SIZE, |
915 | AMDGPU_GEM_DOMAIN_VRAM | | |
916 | AMDGPU_GEM_DOMAIN_GTT, | |
917 | &adev->mem_scratch.robj, | |
918 | &adev->mem_scratch.gpu_addr, | |
919 | (void **)&adev->mem_scratch.ptr); | |
d38ceaf9 AD |
920 | } |
921 | ||
e3ecdffa | 922 | /** |
7ccfd79f | 923 | * amdgpu_device_mem_scratch_fini - Free the VRAM scratch page |
e3ecdffa | 924 | * |
982a820b | 925 | * @adev: amdgpu_device pointer |
e3ecdffa AD |
926 | * |
927 | * Frees the VRAM scratch page. | |
928 | */ | |
7ccfd79f | 929 | static void amdgpu_device_mem_scratch_fini(struct amdgpu_device *adev) |
d38ceaf9 | 930 | { |
7ccfd79f | 931 | amdgpu_bo_free_kernel(&adev->mem_scratch.robj, NULL, NULL); |
d38ceaf9 AD |
932 | } |
933 | ||
934 | /** | |
9c3f2b54 | 935 | * amdgpu_device_program_register_sequence - program an array of registers. |
d38ceaf9 AD |
936 | * |
937 | * @adev: amdgpu_device pointer | |
938 | * @registers: pointer to the register array | |
939 | * @array_size: size of the register array | |
940 | * | |
b8920e1e | 941 | * Programs an array or registers with and or masks. |
d38ceaf9 AD |
942 | * This is a helper for setting golden registers. |
943 | */ | |
9c3f2b54 AD |
944 | void amdgpu_device_program_register_sequence(struct amdgpu_device *adev, |
945 | const u32 *registers, | |
946 | const u32 array_size) | |
d38ceaf9 AD |
947 | { |
948 | u32 tmp, reg, and_mask, or_mask; | |
949 | int i; | |
950 | ||
951 | if (array_size % 3) | |
952 | return; | |
953 | ||
47fc644f | 954 | for (i = 0; i < array_size; i += 3) { |
d38ceaf9 AD |
955 | reg = registers[i + 0]; |
956 | and_mask = registers[i + 1]; | |
957 | or_mask = registers[i + 2]; | |
958 | ||
959 | if (and_mask == 0xffffffff) { | |
960 | tmp = or_mask; | |
961 | } else { | |
962 | tmp = RREG32(reg); | |
963 | tmp &= ~and_mask; | |
e0d07657 HZ |
964 | if (adev->family >= AMDGPU_FAMILY_AI) |
965 | tmp |= (or_mask & and_mask); | |
966 | else | |
967 | tmp |= or_mask; | |
d38ceaf9 AD |
968 | } |
969 | WREG32(reg, tmp); | |
970 | } | |
971 | } | |
972 | ||
e3ecdffa AD |
973 | /** |
974 | * amdgpu_device_pci_config_reset - reset the GPU | |
975 | * | |
976 | * @adev: amdgpu_device pointer | |
977 | * | |
978 | * Resets the GPU using the pci config reset sequence. | |
979 | * Only applicable to asics prior to vega10. | |
980 | */ | |
8111c387 | 981 | void amdgpu_device_pci_config_reset(struct amdgpu_device *adev) |
d38ceaf9 AD |
982 | { |
983 | pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA); | |
984 | } | |
985 | ||
af484df8 AD |
986 | /** |
987 | * amdgpu_device_pci_reset - reset the GPU using generic PCI means | |
988 | * | |
989 | * @adev: amdgpu_device pointer | |
990 | * | |
991 | * Resets the GPU using generic pci reset interfaces (FLR, SBR, etc.). | |
992 | */ | |
993 | int amdgpu_device_pci_reset(struct amdgpu_device *adev) | |
994 | { | |
995 | return pci_reset_function(adev->pdev); | |
996 | } | |
997 | ||
d38ceaf9 | 998 | /* |
06ec9070 | 999 | * amdgpu_device_wb_*() |
455a7bc2 | 1000 | * Writeback is the method by which the GPU updates special pages in memory |
ea81a173 | 1001 | * with the status of certain GPU events (fences, ring pointers,etc.). |
d38ceaf9 AD |
1002 | */ |
1003 | ||
1004 | /** | |
06ec9070 | 1005 | * amdgpu_device_wb_fini - Disable Writeback and free memory |
d38ceaf9 AD |
1006 | * |
1007 | * @adev: amdgpu_device pointer | |
1008 | * | |
1009 | * Disables Writeback and frees the Writeback memory (all asics). | |
1010 | * Used at driver shutdown. | |
1011 | */ | |
06ec9070 | 1012 | static void amdgpu_device_wb_fini(struct amdgpu_device *adev) |
d38ceaf9 AD |
1013 | { |
1014 | if (adev->wb.wb_obj) { | |
a76ed485 AD |
1015 | amdgpu_bo_free_kernel(&adev->wb.wb_obj, |
1016 | &adev->wb.gpu_addr, | |
1017 | (void **)&adev->wb.wb); | |
d38ceaf9 AD |
1018 | adev->wb.wb_obj = NULL; |
1019 | } | |
1020 | } | |
1021 | ||
1022 | /** | |
03f2abb0 | 1023 | * amdgpu_device_wb_init - Init Writeback driver info and allocate memory |
d38ceaf9 AD |
1024 | * |
1025 | * @adev: amdgpu_device pointer | |
1026 | * | |
455a7bc2 | 1027 | * Initializes writeback and allocates writeback memory (all asics). |
d38ceaf9 AD |
1028 | * Used at driver startup. |
1029 | * Returns 0 on success or an -error on failure. | |
1030 | */ | |
06ec9070 | 1031 | static int amdgpu_device_wb_init(struct amdgpu_device *adev) |
d38ceaf9 AD |
1032 | { |
1033 | int r; | |
1034 | ||
1035 | if (adev->wb.wb_obj == NULL) { | |
97407b63 AD |
1036 | /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */ |
1037 | r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8, | |
a76ed485 AD |
1038 | PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT, |
1039 | &adev->wb.wb_obj, &adev->wb.gpu_addr, | |
1040 | (void **)&adev->wb.wb); | |
d38ceaf9 AD |
1041 | if (r) { |
1042 | dev_warn(adev->dev, "(%d) create WB bo failed\n", r); | |
1043 | return r; | |
1044 | } | |
d38ceaf9 AD |
1045 | |
1046 | adev->wb.num_wb = AMDGPU_MAX_WB; | |
1047 | memset(&adev->wb.used, 0, sizeof(adev->wb.used)); | |
1048 | ||
1049 | /* clear wb memory */ | |
73469585 | 1050 | memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8); |
d38ceaf9 AD |
1051 | } |
1052 | ||
1053 | return 0; | |
1054 | } | |
1055 | ||
1056 | /** | |
131b4b36 | 1057 | * amdgpu_device_wb_get - Allocate a wb entry |
d38ceaf9 AD |
1058 | * |
1059 | * @adev: amdgpu_device pointer | |
1060 | * @wb: wb index | |
1061 | * | |
1062 | * Allocate a wb slot for use by the driver (all asics). | |
1063 | * Returns 0 on success or -EINVAL on failure. | |
1064 | */ | |
131b4b36 | 1065 | int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb) |
d38ceaf9 AD |
1066 | { |
1067 | unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb); | |
d38ceaf9 | 1068 | |
97407b63 | 1069 | if (offset < adev->wb.num_wb) { |
7014285a | 1070 | __set_bit(offset, adev->wb.used); |
63ae07ca | 1071 | *wb = offset << 3; /* convert to dw offset */ |
0915fdbc ML |
1072 | return 0; |
1073 | } else { | |
1074 | return -EINVAL; | |
1075 | } | |
1076 | } | |
1077 | ||
d38ceaf9 | 1078 | /** |
131b4b36 | 1079 | * amdgpu_device_wb_free - Free a wb entry |
d38ceaf9 AD |
1080 | * |
1081 | * @adev: amdgpu_device pointer | |
1082 | * @wb: wb index | |
1083 | * | |
1084 | * Free a wb slot allocated for use by the driver (all asics) | |
1085 | */ | |
131b4b36 | 1086 | void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb) |
d38ceaf9 | 1087 | { |
73469585 | 1088 | wb >>= 3; |
d38ceaf9 | 1089 | if (wb < adev->wb.num_wb) |
73469585 | 1090 | __clear_bit(wb, adev->wb.used); |
d38ceaf9 AD |
1091 | } |
1092 | ||
d6895ad3 CK |
1093 | /** |
1094 | * amdgpu_device_resize_fb_bar - try to resize FB BAR | |
1095 | * | |
1096 | * @adev: amdgpu_device pointer | |
1097 | * | |
1098 | * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not | |
1099 | * to fail, but if any of the BARs is not accessible after the size we abort | |
1100 | * driver loading by returning -ENODEV. | |
1101 | */ | |
1102 | int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev) | |
1103 | { | |
453f617a | 1104 | int rbar_size = pci_rebar_bytes_to_size(adev->gmc.real_vram_size); |
31b8adab CK |
1105 | struct pci_bus *root; |
1106 | struct resource *res; | |
b8920e1e | 1107 | unsigned int i; |
d6895ad3 CK |
1108 | u16 cmd; |
1109 | int r; | |
1110 | ||
822130b5 AB |
1111 | if (!IS_ENABLED(CONFIG_PHYS_ADDR_T_64BIT)) |
1112 | return 0; | |
1113 | ||
0c03b912 | 1114 | /* Bypass for VF */ |
1115 | if (amdgpu_sriov_vf(adev)) | |
1116 | return 0; | |
1117 | ||
b7221f2b AD |
1118 | /* skip if the bios has already enabled large BAR */ |
1119 | if (adev->gmc.real_vram_size && | |
1120 | (pci_resource_len(adev->pdev, 0) >= adev->gmc.real_vram_size)) | |
1121 | return 0; | |
1122 | ||
31b8adab CK |
1123 | /* Check if the root BUS has 64bit memory resources */ |
1124 | root = adev->pdev->bus; | |
1125 | while (root->parent) | |
1126 | root = root->parent; | |
1127 | ||
1128 | pci_bus_for_each_resource(root, res, i) { | |
0ebb7c54 | 1129 | if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) && |
31b8adab CK |
1130 | res->start > 0x100000000ull) |
1131 | break; | |
1132 | } | |
1133 | ||
1134 | /* Trying to resize is pointless without a root hub window above 4GB */ | |
1135 | if (!res) | |
1136 | return 0; | |
1137 | ||
453f617a ND |
1138 | /* Limit the BAR size to what is available */ |
1139 | rbar_size = min(fls(pci_rebar_get_possible_sizes(adev->pdev, 0)) - 1, | |
1140 | rbar_size); | |
1141 | ||
d6895ad3 CK |
1142 | /* Disable memory decoding while we change the BAR addresses and size */ |
1143 | pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd); | |
1144 | pci_write_config_word(adev->pdev, PCI_COMMAND, | |
1145 | cmd & ~PCI_COMMAND_MEMORY); | |
1146 | ||
1147 | /* Free the VRAM and doorbell BAR, we most likely need to move both. */ | |
43c064db | 1148 | amdgpu_doorbell_fini(adev); |
d6895ad3 CK |
1149 | if (adev->asic_type >= CHIP_BONAIRE) |
1150 | pci_release_resource(adev->pdev, 2); | |
1151 | ||
1152 | pci_release_resource(adev->pdev, 0); | |
1153 | ||
1154 | r = pci_resize_resource(adev->pdev, 0, rbar_size); | |
1155 | if (r == -ENOSPC) | |
1156 | DRM_INFO("Not enough PCI address space for a large BAR."); | |
1157 | else if (r && r != -ENOTSUPP) | |
1158 | DRM_ERROR("Problem resizing BAR0 (%d).", r); | |
1159 | ||
1160 | pci_assign_unassigned_bus_resources(adev->pdev->bus); | |
1161 | ||
1162 | /* When the doorbell or fb BAR isn't available we have no chance of | |
1163 | * using the device. | |
1164 | */ | |
43c064db | 1165 | r = amdgpu_doorbell_init(adev); |
d6895ad3 CK |
1166 | if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET)) |
1167 | return -ENODEV; | |
1168 | ||
1169 | pci_write_config_word(adev->pdev, PCI_COMMAND, cmd); | |
1170 | ||
1171 | return 0; | |
1172 | } | |
a05502e5 | 1173 | |
9535a86a SZ |
1174 | static bool amdgpu_device_read_bios(struct amdgpu_device *adev) |
1175 | { | |
b8920e1e | 1176 | if (hweight32(adev->aid_mask) && (adev->flags & AMD_IS_APU)) |
9535a86a | 1177 | return false; |
9535a86a SZ |
1178 | |
1179 | return true; | |
1180 | } | |
1181 | ||
d38ceaf9 AD |
1182 | /* |
1183 | * GPU helpers function. | |
1184 | */ | |
1185 | /** | |
39c640c0 | 1186 | * amdgpu_device_need_post - check if the hw need post or not |
d38ceaf9 AD |
1187 | * |
1188 | * @adev: amdgpu_device pointer | |
1189 | * | |
c836fec5 JQ |
1190 | * Check if the asic has been initialized (all asics) at driver startup |
1191 | * or post is needed if hw reset is performed. | |
1192 | * Returns true if need or false if not. | |
d38ceaf9 | 1193 | */ |
39c640c0 | 1194 | bool amdgpu_device_need_post(struct amdgpu_device *adev) |
d38ceaf9 AD |
1195 | { |
1196 | uint32_t reg; | |
1197 | ||
bec86378 ML |
1198 | if (amdgpu_sriov_vf(adev)) |
1199 | return false; | |
1200 | ||
9535a86a SZ |
1201 | if (!amdgpu_device_read_bios(adev)) |
1202 | return false; | |
1203 | ||
bec86378 | 1204 | if (amdgpu_passthrough(adev)) { |
1da2c326 ML |
1205 | /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot |
1206 | * some old smc fw still need driver do vPost otherwise gpu hang, while | |
1207 | * those smc fw version above 22.15 doesn't have this flaw, so we force | |
1208 | * vpost executed for smc version below 22.15 | |
bec86378 ML |
1209 | */ |
1210 | if (adev->asic_type == CHIP_FIJI) { | |
1211 | int err; | |
1212 | uint32_t fw_ver; | |
b8920e1e | 1213 | |
bec86378 ML |
1214 | err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev); |
1215 | /* force vPost if error occured */ | |
1216 | if (err) | |
1217 | return true; | |
1218 | ||
1219 | fw_ver = *((uint32_t *)adev->pm.fw->data + 69); | |
1da2c326 ML |
1220 | if (fw_ver < 0x00160e00) |
1221 | return true; | |
bec86378 | 1222 | } |
bec86378 | 1223 | } |
91fe77eb | 1224 | |
e3c1b071 | 1225 | /* Don't post if we need to reset whole hive on init */ |
1226 | if (adev->gmc.xgmi.pending_reset) | |
1227 | return false; | |
1228 | ||
91fe77eb | 1229 | if (adev->has_hw_reset) { |
1230 | adev->has_hw_reset = false; | |
1231 | return true; | |
1232 | } | |
1233 | ||
1234 | /* bios scratch used on CIK+ */ | |
1235 | if (adev->asic_type >= CHIP_BONAIRE) | |
1236 | return amdgpu_atombios_scratch_need_asic_init(adev); | |
1237 | ||
1238 | /* check MEM_SIZE for older asics */ | |
1239 | reg = amdgpu_asic_get_config_memsize(adev); | |
1240 | ||
1241 | if ((reg != 0) && (reg != 0xffffffff)) | |
1242 | return false; | |
1243 | ||
1244 | return true; | |
bec86378 ML |
1245 | } |
1246 | ||
5d1eb4c4 ML |
1247 | /* |
1248 | * Intel hosts such as Raptor Lake and Sapphire Rapids don't support dynamic | |
1249 | * speed switching. Until we have confirmation from Intel that a specific host | |
1250 | * supports it, it's safer that we keep it disabled for all. | |
1251 | * | |
1252 | * 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/ | |
1253 | * https://gitlab.freedesktop.org/drm/amd/-/issues/2663 | |
1254 | */ | |
1255 | bool amdgpu_device_pcie_dynamic_switching_supported(void) | |
1256 | { | |
1257 | #if IS_ENABLED(CONFIG_X86) | |
1258 | struct cpuinfo_x86 *c = &cpu_data(0); | |
1259 | ||
1260 | if (c->x86_vendor == X86_VENDOR_INTEL) | |
1261 | return false; | |
1262 | #endif | |
1263 | return true; | |
1264 | } | |
1265 | ||
0ab5d711 ML |
1266 | /** |
1267 | * amdgpu_device_should_use_aspm - check if the device should program ASPM | |
1268 | * | |
1269 | * @adev: amdgpu_device pointer | |
1270 | * | |
1271 | * Confirm whether the module parameter and pcie bridge agree that ASPM should | |
1272 | * be set for this device. | |
1273 | * | |
1274 | * Returns true if it should be used or false if not. | |
1275 | */ | |
1276 | bool amdgpu_device_should_use_aspm(struct amdgpu_device *adev) | |
1277 | { | |
1278 | switch (amdgpu_aspm) { | |
1279 | case -1: | |
1280 | break; | |
1281 | case 0: | |
1282 | return false; | |
1283 | case 1: | |
1284 | return true; | |
1285 | default: | |
1286 | return false; | |
1287 | } | |
1288 | return pcie_aspm_enabled(adev->pdev); | |
1289 | } | |
1290 | ||
3ad5dcfe KHF |
1291 | bool amdgpu_device_aspm_support_quirk(void) |
1292 | { | |
1293 | #if IS_ENABLED(CONFIG_X86) | |
1294 | struct cpuinfo_x86 *c = &cpu_data(0); | |
1295 | ||
1296 | return !(c->x86 == 6 && c->x86_model == INTEL_FAM6_ALDERLAKE); | |
1297 | #else | |
1298 | return true; | |
1299 | #endif | |
1300 | } | |
1301 | ||
d38ceaf9 AD |
1302 | /* if we get transitioned to only one device, take VGA back */ |
1303 | /** | |
06ec9070 | 1304 | * amdgpu_device_vga_set_decode - enable/disable vga decode |
d38ceaf9 | 1305 | * |
bf44e8ce | 1306 | * @pdev: PCI device pointer |
d38ceaf9 AD |
1307 | * @state: enable/disable vga decode |
1308 | * | |
1309 | * Enable/disable vga decode (all asics). | |
1310 | * Returns VGA resource flags. | |
1311 | */ | |
bf44e8ce CH |
1312 | static unsigned int amdgpu_device_vga_set_decode(struct pci_dev *pdev, |
1313 | bool state) | |
d38ceaf9 | 1314 | { |
bf44e8ce | 1315 | struct amdgpu_device *adev = drm_to_adev(pci_get_drvdata(pdev)); |
b8920e1e | 1316 | |
d38ceaf9 AD |
1317 | amdgpu_asic_set_vga_state(adev, state); |
1318 | if (state) | |
1319 | return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM | | |
1320 | VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM; | |
1321 | else | |
1322 | return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM; | |
1323 | } | |
1324 | ||
e3ecdffa AD |
1325 | /** |
1326 | * amdgpu_device_check_block_size - validate the vm block size | |
1327 | * | |
1328 | * @adev: amdgpu_device pointer | |
1329 | * | |
1330 | * Validates the vm block size specified via module parameter. | |
1331 | * The vm block size defines number of bits in page table versus page directory, | |
1332 | * a page is 4KB so we have 12 bits offset, minimum 9 bits in the | |
1333 | * page table and the remaining bits are in the page directory. | |
1334 | */ | |
06ec9070 | 1335 | static void amdgpu_device_check_block_size(struct amdgpu_device *adev) |
a1adf8be CZ |
1336 | { |
1337 | /* defines number of bits in page table versus page directory, | |
1338 | * a page is 4KB so we have 12 bits offset, minimum 9 bits in the | |
b8920e1e SS |
1339 | * page table and the remaining bits are in the page directory |
1340 | */ | |
bab4fee7 JZ |
1341 | if (amdgpu_vm_block_size == -1) |
1342 | return; | |
a1adf8be | 1343 | |
bab4fee7 | 1344 | if (amdgpu_vm_block_size < 9) { |
a1adf8be CZ |
1345 | dev_warn(adev->dev, "VM page table size (%d) too small\n", |
1346 | amdgpu_vm_block_size); | |
97489129 | 1347 | amdgpu_vm_block_size = -1; |
a1adf8be | 1348 | } |
a1adf8be CZ |
1349 | } |
1350 | ||
e3ecdffa AD |
1351 | /** |
1352 | * amdgpu_device_check_vm_size - validate the vm size | |
1353 | * | |
1354 | * @adev: amdgpu_device pointer | |
1355 | * | |
1356 | * Validates the vm size in GB specified via module parameter. | |
1357 | * The VM size is the size of the GPU virtual memory space in GB. | |
1358 | */ | |
06ec9070 | 1359 | static void amdgpu_device_check_vm_size(struct amdgpu_device *adev) |
83ca145d | 1360 | { |
64dab074 AD |
1361 | /* no need to check the default value */ |
1362 | if (amdgpu_vm_size == -1) | |
1363 | return; | |
1364 | ||
83ca145d ZJ |
1365 | if (amdgpu_vm_size < 1) { |
1366 | dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n", | |
1367 | amdgpu_vm_size); | |
f3368128 | 1368 | amdgpu_vm_size = -1; |
83ca145d | 1369 | } |
83ca145d ZJ |
1370 | } |
1371 | ||
7951e376 RZ |
1372 | static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev) |
1373 | { | |
1374 | struct sysinfo si; | |
a9d4fe2f | 1375 | bool is_os_64 = (sizeof(void *) == 8); |
7951e376 RZ |
1376 | uint64_t total_memory; |
1377 | uint64_t dram_size_seven_GB = 0x1B8000000; | |
1378 | uint64_t dram_size_three_GB = 0xB8000000; | |
1379 | ||
1380 | if (amdgpu_smu_memory_pool_size == 0) | |
1381 | return; | |
1382 | ||
1383 | if (!is_os_64) { | |
1384 | DRM_WARN("Not 64-bit OS, feature not supported\n"); | |
1385 | goto def_value; | |
1386 | } | |
1387 | si_meminfo(&si); | |
1388 | total_memory = (uint64_t)si.totalram * si.mem_unit; | |
1389 | ||
1390 | if ((amdgpu_smu_memory_pool_size == 1) || | |
1391 | (amdgpu_smu_memory_pool_size == 2)) { | |
1392 | if (total_memory < dram_size_three_GB) | |
1393 | goto def_value1; | |
1394 | } else if ((amdgpu_smu_memory_pool_size == 4) || | |
1395 | (amdgpu_smu_memory_pool_size == 8)) { | |
1396 | if (total_memory < dram_size_seven_GB) | |
1397 | goto def_value1; | |
1398 | } else { | |
1399 | DRM_WARN("Smu memory pool size not supported\n"); | |
1400 | goto def_value; | |
1401 | } | |
1402 | adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28; | |
1403 | ||
1404 | return; | |
1405 | ||
1406 | def_value1: | |
1407 | DRM_WARN("No enough system memory\n"); | |
1408 | def_value: | |
1409 | adev->pm.smu_prv_buffer_size = 0; | |
1410 | } | |
1411 | ||
9f6a7857 HR |
1412 | static int amdgpu_device_init_apu_flags(struct amdgpu_device *adev) |
1413 | { | |
1414 | if (!(adev->flags & AMD_IS_APU) || | |
1415 | adev->asic_type < CHIP_RAVEN) | |
1416 | return 0; | |
1417 | ||
1418 | switch (adev->asic_type) { | |
1419 | case CHIP_RAVEN: | |
1420 | if (adev->pdev->device == 0x15dd) | |
1421 | adev->apu_flags |= AMD_APU_IS_RAVEN; | |
1422 | if (adev->pdev->device == 0x15d8) | |
1423 | adev->apu_flags |= AMD_APU_IS_PICASSO; | |
1424 | break; | |
1425 | case CHIP_RENOIR: | |
1426 | if ((adev->pdev->device == 0x1636) || | |
1427 | (adev->pdev->device == 0x164c)) | |
1428 | adev->apu_flags |= AMD_APU_IS_RENOIR; | |
1429 | else | |
1430 | adev->apu_flags |= AMD_APU_IS_GREEN_SARDINE; | |
1431 | break; | |
1432 | case CHIP_VANGOGH: | |
1433 | adev->apu_flags |= AMD_APU_IS_VANGOGH; | |
1434 | break; | |
1435 | case CHIP_YELLOW_CARP: | |
1436 | break; | |
d0f56dc2 | 1437 | case CHIP_CYAN_SKILLFISH: |
dfcc3e8c AD |
1438 | if ((adev->pdev->device == 0x13FE) || |
1439 | (adev->pdev->device == 0x143F)) | |
d0f56dc2 TZ |
1440 | adev->apu_flags |= AMD_APU_IS_CYAN_SKILLFISH2; |
1441 | break; | |
9f6a7857 | 1442 | default: |
4eaf21b7 | 1443 | break; |
9f6a7857 HR |
1444 | } |
1445 | ||
1446 | return 0; | |
1447 | } | |
1448 | ||
d38ceaf9 | 1449 | /** |
06ec9070 | 1450 | * amdgpu_device_check_arguments - validate module params |
d38ceaf9 AD |
1451 | * |
1452 | * @adev: amdgpu_device pointer | |
1453 | * | |
1454 | * Validates certain module parameters and updates | |
1455 | * the associated values used by the driver (all asics). | |
1456 | */ | |
912dfc84 | 1457 | static int amdgpu_device_check_arguments(struct amdgpu_device *adev) |
d38ceaf9 | 1458 | { |
5b011235 CZ |
1459 | if (amdgpu_sched_jobs < 4) { |
1460 | dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n", | |
1461 | amdgpu_sched_jobs); | |
1462 | amdgpu_sched_jobs = 4; | |
47fc644f | 1463 | } else if (!is_power_of_2(amdgpu_sched_jobs)) { |
5b011235 CZ |
1464 | dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n", |
1465 | amdgpu_sched_jobs); | |
1466 | amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs); | |
1467 | } | |
d38ceaf9 | 1468 | |
83e74db6 | 1469 | if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) { |
f9321cc4 CK |
1470 | /* gart size must be greater or equal to 32M */ |
1471 | dev_warn(adev->dev, "gart size (%d) too small\n", | |
1472 | amdgpu_gart_size); | |
83e74db6 | 1473 | amdgpu_gart_size = -1; |
d38ceaf9 AD |
1474 | } |
1475 | ||
36d38372 | 1476 | if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) { |
c4e1a13a | 1477 | /* gtt size must be greater or equal to 32M */ |
36d38372 CK |
1478 | dev_warn(adev->dev, "gtt size (%d) too small\n", |
1479 | amdgpu_gtt_size); | |
1480 | amdgpu_gtt_size = -1; | |
d38ceaf9 AD |
1481 | } |
1482 | ||
d07f14be RH |
1483 | /* valid range is between 4 and 9 inclusive */ |
1484 | if (amdgpu_vm_fragment_size != -1 && | |
1485 | (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) { | |
1486 | dev_warn(adev->dev, "valid range is between 4 and 9\n"); | |
1487 | amdgpu_vm_fragment_size = -1; | |
1488 | } | |
1489 | ||
5d5bd5e3 KW |
1490 | if (amdgpu_sched_hw_submission < 2) { |
1491 | dev_warn(adev->dev, "sched hw submission jobs (%d) must be at least 2\n", | |
1492 | amdgpu_sched_hw_submission); | |
1493 | amdgpu_sched_hw_submission = 2; | |
1494 | } else if (!is_power_of_2(amdgpu_sched_hw_submission)) { | |
1495 | dev_warn(adev->dev, "sched hw submission jobs (%d) must be a power of 2\n", | |
1496 | amdgpu_sched_hw_submission); | |
1497 | amdgpu_sched_hw_submission = roundup_pow_of_two(amdgpu_sched_hw_submission); | |
1498 | } | |
1499 | ||
2656fd23 AG |
1500 | if (amdgpu_reset_method < -1 || amdgpu_reset_method > 4) { |
1501 | dev_warn(adev->dev, "invalid option for reset method, reverting to default\n"); | |
1502 | amdgpu_reset_method = -1; | |
1503 | } | |
1504 | ||
7951e376 RZ |
1505 | amdgpu_device_check_smu_prv_buffer_size(adev); |
1506 | ||
06ec9070 | 1507 | amdgpu_device_check_vm_size(adev); |
d38ceaf9 | 1508 | |
06ec9070 | 1509 | amdgpu_device_check_block_size(adev); |
6a7f76e7 | 1510 | |
19aede77 | 1511 | adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type); |
912dfc84 | 1512 | |
e3c00faa | 1513 | return 0; |
d38ceaf9 AD |
1514 | } |
1515 | ||
1516 | /** | |
1517 | * amdgpu_switcheroo_set_state - set switcheroo state | |
1518 | * | |
1519 | * @pdev: pci dev pointer | |
1694467b | 1520 | * @state: vga_switcheroo state |
d38ceaf9 | 1521 | * |
12024b17 | 1522 | * Callback for the switcheroo driver. Suspends or resumes |
d38ceaf9 AD |
1523 | * the asics before or after it is powered up using ACPI methods. |
1524 | */ | |
8aba21b7 LT |
1525 | static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, |
1526 | enum vga_switcheroo_state state) | |
d38ceaf9 AD |
1527 | { |
1528 | struct drm_device *dev = pci_get_drvdata(pdev); | |
de185019 | 1529 | int r; |
d38ceaf9 | 1530 | |
b98c6299 | 1531 | if (amdgpu_device_supports_px(dev) && state == VGA_SWITCHEROO_OFF) |
d38ceaf9 AD |
1532 | return; |
1533 | ||
1534 | if (state == VGA_SWITCHEROO_ON) { | |
dd4fa6c1 | 1535 | pr_info("switched on\n"); |
d38ceaf9 AD |
1536 | /* don't suspend or resume card normally */ |
1537 | dev->switch_power_state = DRM_SWITCH_POWER_CHANGING; | |
1538 | ||
8f66090b TZ |
1539 | pci_set_power_state(pdev, PCI_D0); |
1540 | amdgpu_device_load_pci_state(pdev); | |
1541 | r = pci_enable_device(pdev); | |
de185019 AD |
1542 | if (r) |
1543 | DRM_WARN("pci_enable_device failed (%d)\n", r); | |
1544 | amdgpu_device_resume(dev, true); | |
d38ceaf9 | 1545 | |
d38ceaf9 | 1546 | dev->switch_power_state = DRM_SWITCH_POWER_ON; |
d38ceaf9 | 1547 | } else { |
dd4fa6c1 | 1548 | pr_info("switched off\n"); |
d38ceaf9 | 1549 | dev->switch_power_state = DRM_SWITCH_POWER_CHANGING; |
de185019 | 1550 | amdgpu_device_suspend(dev, true); |
8f66090b | 1551 | amdgpu_device_cache_pci_state(pdev); |
de185019 | 1552 | /* Shut down the device */ |
8f66090b TZ |
1553 | pci_disable_device(pdev); |
1554 | pci_set_power_state(pdev, PCI_D3cold); | |
d38ceaf9 AD |
1555 | dev->switch_power_state = DRM_SWITCH_POWER_OFF; |
1556 | } | |
1557 | } | |
1558 | ||
1559 | /** | |
1560 | * amdgpu_switcheroo_can_switch - see if switcheroo state can change | |
1561 | * | |
1562 | * @pdev: pci dev pointer | |
1563 | * | |
1564 | * Callback for the switcheroo driver. Check of the switcheroo | |
1565 | * state can be changed. | |
1566 | * Returns true if the state can be changed, false if not. | |
1567 | */ | |
1568 | static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev) | |
1569 | { | |
1570 | struct drm_device *dev = pci_get_drvdata(pdev); | |
1571 | ||
b8920e1e | 1572 | /* |
d38ceaf9 AD |
1573 | * FIXME: open_count is protected by drm_global_mutex but that would lead to |
1574 | * locking inversion with the driver load path. And the access here is | |
1575 | * completely racy anyway. So don't bother with locking for now. | |
1576 | */ | |
7e13ad89 | 1577 | return atomic_read(&dev->open_count) == 0; |
d38ceaf9 AD |
1578 | } |
1579 | ||
1580 | static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = { | |
1581 | .set_gpu_state = amdgpu_switcheroo_set_state, | |
1582 | .reprobe = NULL, | |
1583 | .can_switch = amdgpu_switcheroo_can_switch, | |
1584 | }; | |
1585 | ||
e3ecdffa AD |
1586 | /** |
1587 | * amdgpu_device_ip_set_clockgating_state - set the CG state | |
1588 | * | |
87e3f136 | 1589 | * @dev: amdgpu_device pointer |
e3ecdffa AD |
1590 | * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) |
1591 | * @state: clockgating state (gate or ungate) | |
1592 | * | |
1593 | * Sets the requested clockgating state for all instances of | |
1594 | * the hardware IP specified. | |
1595 | * Returns the error code from the last instance. | |
1596 | */ | |
43fa561f | 1597 | int amdgpu_device_ip_set_clockgating_state(void *dev, |
2990a1fc AD |
1598 | enum amd_ip_block_type block_type, |
1599 | enum amd_clockgating_state state) | |
d38ceaf9 | 1600 | { |
43fa561f | 1601 | struct amdgpu_device *adev = dev; |
d38ceaf9 AD |
1602 | int i, r = 0; |
1603 | ||
1604 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 1605 | if (!adev->ip_blocks[i].status.valid) |
9ecbe7f5 | 1606 | continue; |
c722865a RZ |
1607 | if (adev->ip_blocks[i].version->type != block_type) |
1608 | continue; | |
1609 | if (!adev->ip_blocks[i].version->funcs->set_clockgating_state) | |
1610 | continue; | |
1611 | r = adev->ip_blocks[i].version->funcs->set_clockgating_state( | |
1612 | (void *)adev, state); | |
1613 | if (r) | |
1614 | DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n", | |
1615 | adev->ip_blocks[i].version->funcs->name, r); | |
d38ceaf9 AD |
1616 | } |
1617 | return r; | |
1618 | } | |
1619 | ||
e3ecdffa AD |
1620 | /** |
1621 | * amdgpu_device_ip_set_powergating_state - set the PG state | |
1622 | * | |
87e3f136 | 1623 | * @dev: amdgpu_device pointer |
e3ecdffa AD |
1624 | * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) |
1625 | * @state: powergating state (gate or ungate) | |
1626 | * | |
1627 | * Sets the requested powergating state for all instances of | |
1628 | * the hardware IP specified. | |
1629 | * Returns the error code from the last instance. | |
1630 | */ | |
43fa561f | 1631 | int amdgpu_device_ip_set_powergating_state(void *dev, |
2990a1fc AD |
1632 | enum amd_ip_block_type block_type, |
1633 | enum amd_powergating_state state) | |
d38ceaf9 | 1634 | { |
43fa561f | 1635 | struct amdgpu_device *adev = dev; |
d38ceaf9 AD |
1636 | int i, r = 0; |
1637 | ||
1638 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 1639 | if (!adev->ip_blocks[i].status.valid) |
9ecbe7f5 | 1640 | continue; |
c722865a RZ |
1641 | if (adev->ip_blocks[i].version->type != block_type) |
1642 | continue; | |
1643 | if (!adev->ip_blocks[i].version->funcs->set_powergating_state) | |
1644 | continue; | |
1645 | r = adev->ip_blocks[i].version->funcs->set_powergating_state( | |
1646 | (void *)adev, state); | |
1647 | if (r) | |
1648 | DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n", | |
1649 | adev->ip_blocks[i].version->funcs->name, r); | |
d38ceaf9 AD |
1650 | } |
1651 | return r; | |
1652 | } | |
1653 | ||
e3ecdffa AD |
1654 | /** |
1655 | * amdgpu_device_ip_get_clockgating_state - get the CG state | |
1656 | * | |
1657 | * @adev: amdgpu_device pointer | |
1658 | * @flags: clockgating feature flags | |
1659 | * | |
1660 | * Walks the list of IPs on the device and updates the clockgating | |
1661 | * flags for each IP. | |
1662 | * Updates @flags with the feature flags for each hardware IP where | |
1663 | * clockgating is enabled. | |
1664 | */ | |
2990a1fc | 1665 | void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev, |
25faeddc | 1666 | u64 *flags) |
6cb2d4e4 HR |
1667 | { |
1668 | int i; | |
1669 | ||
1670 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
1671 | if (!adev->ip_blocks[i].status.valid) | |
1672 | continue; | |
1673 | if (adev->ip_blocks[i].version->funcs->get_clockgating_state) | |
1674 | adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags); | |
1675 | } | |
1676 | } | |
1677 | ||
e3ecdffa AD |
1678 | /** |
1679 | * amdgpu_device_ip_wait_for_idle - wait for idle | |
1680 | * | |
1681 | * @adev: amdgpu_device pointer | |
1682 | * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) | |
1683 | * | |
1684 | * Waits for the request hardware IP to be idle. | |
1685 | * Returns 0 for success or a negative error code on failure. | |
1686 | */ | |
2990a1fc AD |
1687 | int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev, |
1688 | enum amd_ip_block_type block_type) | |
5dbbb60b AD |
1689 | { |
1690 | int i, r; | |
1691 | ||
1692 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 1693 | if (!adev->ip_blocks[i].status.valid) |
9ecbe7f5 | 1694 | continue; |
a1255107 AD |
1695 | if (adev->ip_blocks[i].version->type == block_type) { |
1696 | r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev); | |
5dbbb60b AD |
1697 | if (r) |
1698 | return r; | |
1699 | break; | |
1700 | } | |
1701 | } | |
1702 | return 0; | |
1703 | ||
1704 | } | |
1705 | ||
e3ecdffa AD |
1706 | /** |
1707 | * amdgpu_device_ip_is_idle - is the hardware IP idle | |
1708 | * | |
1709 | * @adev: amdgpu_device pointer | |
1710 | * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) | |
1711 | * | |
1712 | * Check if the hardware IP is idle or not. | |
1713 | * Returns true if it the IP is idle, false if not. | |
1714 | */ | |
2990a1fc AD |
1715 | bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev, |
1716 | enum amd_ip_block_type block_type) | |
5dbbb60b AD |
1717 | { |
1718 | int i; | |
1719 | ||
1720 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 1721 | if (!adev->ip_blocks[i].status.valid) |
9ecbe7f5 | 1722 | continue; |
a1255107 AD |
1723 | if (adev->ip_blocks[i].version->type == block_type) |
1724 | return adev->ip_blocks[i].version->funcs->is_idle((void *)adev); | |
5dbbb60b AD |
1725 | } |
1726 | return true; | |
1727 | ||
1728 | } | |
1729 | ||
e3ecdffa AD |
1730 | /** |
1731 | * amdgpu_device_ip_get_ip_block - get a hw IP pointer | |
1732 | * | |
1733 | * @adev: amdgpu_device pointer | |
87e3f136 | 1734 | * @type: Type of hardware IP (SMU, GFX, UVD, etc.) |
e3ecdffa AD |
1735 | * |
1736 | * Returns a pointer to the hardware IP block structure | |
1737 | * if it exists for the asic, otherwise NULL. | |
1738 | */ | |
2990a1fc AD |
1739 | struct amdgpu_ip_block * |
1740 | amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev, | |
1741 | enum amd_ip_block_type type) | |
d38ceaf9 AD |
1742 | { |
1743 | int i; | |
1744 | ||
1745 | for (i = 0; i < adev->num_ip_blocks; i++) | |
a1255107 | 1746 | if (adev->ip_blocks[i].version->type == type) |
d38ceaf9 AD |
1747 | return &adev->ip_blocks[i]; |
1748 | ||
1749 | return NULL; | |
1750 | } | |
1751 | ||
1752 | /** | |
2990a1fc | 1753 | * amdgpu_device_ip_block_version_cmp |
d38ceaf9 AD |
1754 | * |
1755 | * @adev: amdgpu_device pointer | |
5fc3aeeb | 1756 | * @type: enum amd_ip_block_type |
d38ceaf9 AD |
1757 | * @major: major version |
1758 | * @minor: minor version | |
1759 | * | |
1760 | * return 0 if equal or greater | |
1761 | * return 1 if smaller or the ip_block doesn't exist | |
1762 | */ | |
2990a1fc AD |
1763 | int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev, |
1764 | enum amd_ip_block_type type, | |
1765 | u32 major, u32 minor) | |
d38ceaf9 | 1766 | { |
2990a1fc | 1767 | struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type); |
d38ceaf9 | 1768 | |
a1255107 AD |
1769 | if (ip_block && ((ip_block->version->major > major) || |
1770 | ((ip_block->version->major == major) && | |
1771 | (ip_block->version->minor >= minor)))) | |
d38ceaf9 AD |
1772 | return 0; |
1773 | ||
1774 | return 1; | |
1775 | } | |
1776 | ||
a1255107 | 1777 | /** |
2990a1fc | 1778 | * amdgpu_device_ip_block_add |
a1255107 AD |
1779 | * |
1780 | * @adev: amdgpu_device pointer | |
1781 | * @ip_block_version: pointer to the IP to add | |
1782 | * | |
1783 | * Adds the IP block driver information to the collection of IPs | |
1784 | * on the asic. | |
1785 | */ | |
2990a1fc AD |
1786 | int amdgpu_device_ip_block_add(struct amdgpu_device *adev, |
1787 | const struct amdgpu_ip_block_version *ip_block_version) | |
a1255107 AD |
1788 | { |
1789 | if (!ip_block_version) | |
1790 | return -EINVAL; | |
1791 | ||
7bd939d0 LG |
1792 | switch (ip_block_version->type) { |
1793 | case AMD_IP_BLOCK_TYPE_VCN: | |
1794 | if (adev->harvest_ip_mask & AMD_HARVEST_IP_VCN_MASK) | |
1795 | return 0; | |
1796 | break; | |
1797 | case AMD_IP_BLOCK_TYPE_JPEG: | |
1798 | if (adev->harvest_ip_mask & AMD_HARVEST_IP_JPEG_MASK) | |
1799 | return 0; | |
1800 | break; | |
1801 | default: | |
1802 | break; | |
1803 | } | |
1804 | ||
e966a725 | 1805 | DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks, |
a0bae357 HR |
1806 | ip_block_version->funcs->name); |
1807 | ||
a1255107 AD |
1808 | adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version; |
1809 | ||
1810 | return 0; | |
1811 | } | |
1812 | ||
e3ecdffa AD |
1813 | /** |
1814 | * amdgpu_device_enable_virtual_display - enable virtual display feature | |
1815 | * | |
1816 | * @adev: amdgpu_device pointer | |
1817 | * | |
1818 | * Enabled the virtual display feature if the user has enabled it via | |
1819 | * the module parameter virtual_display. This feature provides a virtual | |
1820 | * display hardware on headless boards or in virtualized environments. | |
1821 | * This function parses and validates the configuration string specified by | |
1822 | * the user and configues the virtual display configuration (number of | |
1823 | * virtual connectors, crtcs, etc.) specified. | |
1824 | */ | |
483ef985 | 1825 | static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev) |
9accf2fd ED |
1826 | { |
1827 | adev->enable_virtual_display = false; | |
1828 | ||
1829 | if (amdgpu_virtual_display) { | |
8f66090b | 1830 | const char *pci_address_name = pci_name(adev->pdev); |
0f66356d | 1831 | char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname; |
9accf2fd ED |
1832 | |
1833 | pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL); | |
1834 | pciaddstr_tmp = pciaddstr; | |
0f66356d ED |
1835 | while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) { |
1836 | pciaddname = strsep(&pciaddname_tmp, ","); | |
967de2a9 YT |
1837 | if (!strcmp("all", pciaddname) |
1838 | || !strcmp(pci_address_name, pciaddname)) { | |
0f66356d ED |
1839 | long num_crtc; |
1840 | int res = -1; | |
1841 | ||
9accf2fd | 1842 | adev->enable_virtual_display = true; |
0f66356d ED |
1843 | |
1844 | if (pciaddname_tmp) | |
1845 | res = kstrtol(pciaddname_tmp, 10, | |
1846 | &num_crtc); | |
1847 | ||
1848 | if (!res) { | |
1849 | if (num_crtc < 1) | |
1850 | num_crtc = 1; | |
1851 | if (num_crtc > 6) | |
1852 | num_crtc = 6; | |
1853 | adev->mode_info.num_crtc = num_crtc; | |
1854 | } else { | |
1855 | adev->mode_info.num_crtc = 1; | |
1856 | } | |
9accf2fd ED |
1857 | break; |
1858 | } | |
1859 | } | |
1860 | ||
0f66356d ED |
1861 | DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n", |
1862 | amdgpu_virtual_display, pci_address_name, | |
1863 | adev->enable_virtual_display, adev->mode_info.num_crtc); | |
9accf2fd ED |
1864 | |
1865 | kfree(pciaddstr); | |
1866 | } | |
1867 | } | |
1868 | ||
25263da3 AD |
1869 | void amdgpu_device_set_sriov_virtual_display(struct amdgpu_device *adev) |
1870 | { | |
1871 | if (amdgpu_sriov_vf(adev) && !adev->enable_virtual_display) { | |
1872 | adev->mode_info.num_crtc = 1; | |
1873 | adev->enable_virtual_display = true; | |
1874 | DRM_INFO("virtual_display:%d, num_crtc:%d\n", | |
1875 | adev->enable_virtual_display, adev->mode_info.num_crtc); | |
1876 | } | |
1877 | } | |
1878 | ||
e3ecdffa AD |
1879 | /** |
1880 | * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware | |
1881 | * | |
1882 | * @adev: amdgpu_device pointer | |
1883 | * | |
1884 | * Parses the asic configuration parameters specified in the gpu info | |
1885 | * firmware and makes them availale to the driver for use in configuring | |
1886 | * the asic. | |
1887 | * Returns 0 on success, -EINVAL on failure. | |
1888 | */ | |
e2a75f88 AD |
1889 | static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev) |
1890 | { | |
e2a75f88 | 1891 | const char *chip_name; |
c0a43457 | 1892 | char fw_name[40]; |
e2a75f88 AD |
1893 | int err; |
1894 | const struct gpu_info_firmware_header_v1_0 *hdr; | |
1895 | ||
ab4fe3e1 HR |
1896 | adev->firmware.gpu_info_fw = NULL; |
1897 | ||
72de33f8 | 1898 | if (adev->mman.discovery_bin) { |
cc375d8c TY |
1899 | /* |
1900 | * FIXME: The bounding box is still needed by Navi12, so | |
e24d0e91 | 1901 | * temporarily read it from gpu_info firmware. Should be dropped |
cc375d8c TY |
1902 | * when DAL no longer needs it. |
1903 | */ | |
1904 | if (adev->asic_type != CHIP_NAVI12) | |
1905 | return 0; | |
258620d0 AD |
1906 | } |
1907 | ||
e2a75f88 | 1908 | switch (adev->asic_type) { |
e2a75f88 AD |
1909 | default: |
1910 | return 0; | |
1911 | case CHIP_VEGA10: | |
1912 | chip_name = "vega10"; | |
1913 | break; | |
3f76dced AD |
1914 | case CHIP_VEGA12: |
1915 | chip_name = "vega12"; | |
1916 | break; | |
2d2e5e7e | 1917 | case CHIP_RAVEN: |
54f78a76 | 1918 | if (adev->apu_flags & AMD_APU_IS_RAVEN2) |
54c4d17e | 1919 | chip_name = "raven2"; |
54f78a76 | 1920 | else if (adev->apu_flags & AMD_APU_IS_PICASSO) |
741deade | 1921 | chip_name = "picasso"; |
54c4d17e FX |
1922 | else |
1923 | chip_name = "raven"; | |
2d2e5e7e | 1924 | break; |
65e60f6e LM |
1925 | case CHIP_ARCTURUS: |
1926 | chip_name = "arcturus"; | |
1927 | break; | |
42b325e5 XY |
1928 | case CHIP_NAVI12: |
1929 | chip_name = "navi12"; | |
1930 | break; | |
e2a75f88 AD |
1931 | } |
1932 | ||
1933 | snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name); | |
b31d3063 | 1934 | err = amdgpu_ucode_request(adev, &adev->firmware.gpu_info_fw, fw_name); |
e2a75f88 AD |
1935 | if (err) { |
1936 | dev_err(adev->dev, | |
b31d3063 | 1937 | "Failed to get gpu_info firmware \"%s\"\n", |
e2a75f88 AD |
1938 | fw_name); |
1939 | goto out; | |
1940 | } | |
1941 | ||
ab4fe3e1 | 1942 | hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data; |
e2a75f88 AD |
1943 | amdgpu_ucode_print_gpu_info_hdr(&hdr->header); |
1944 | ||
1945 | switch (hdr->version_major) { | |
1946 | case 1: | |
1947 | { | |
1948 | const struct gpu_info_firmware_v1_0 *gpu_info_fw = | |
ab4fe3e1 | 1949 | (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data + |
e2a75f88 AD |
1950 | le32_to_cpu(hdr->header.ucode_array_offset_bytes)); |
1951 | ||
cc375d8c TY |
1952 | /* |
1953 | * Should be droped when DAL no longer needs it. | |
1954 | */ | |
1955 | if (adev->asic_type == CHIP_NAVI12) | |
ec51d3fa XY |
1956 | goto parse_soc_bounding_box; |
1957 | ||
b5ab16bf AD |
1958 | adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se); |
1959 | adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh); | |
1960 | adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se); | |
1961 | adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se); | |
e2a75f88 | 1962 | adev->gfx.config.max_texture_channel_caches = |
b5ab16bf AD |
1963 | le32_to_cpu(gpu_info_fw->gc_num_tccs); |
1964 | adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs); | |
1965 | adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds); | |
1966 | adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth); | |
1967 | adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth); | |
e2a75f88 | 1968 | adev->gfx.config.double_offchip_lds_buf = |
b5ab16bf AD |
1969 | le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer); |
1970 | adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size); | |
51fd0370 HZ |
1971 | adev->gfx.cu_info.max_waves_per_simd = |
1972 | le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd); | |
1973 | adev->gfx.cu_info.max_scratch_slots_per_cu = | |
1974 | le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu); | |
1975 | adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size); | |
48321c3d | 1976 | if (hdr->version_minor >= 1) { |
35c2e910 HZ |
1977 | const struct gpu_info_firmware_v1_1 *gpu_info_fw = |
1978 | (const struct gpu_info_firmware_v1_1 *)(adev->firmware.gpu_info_fw->data + | |
1979 | le32_to_cpu(hdr->header.ucode_array_offset_bytes)); | |
1980 | adev->gfx.config.num_sc_per_sh = | |
1981 | le32_to_cpu(gpu_info_fw->num_sc_per_sh); | |
1982 | adev->gfx.config.num_packer_per_sc = | |
1983 | le32_to_cpu(gpu_info_fw->num_packer_per_sc); | |
1984 | } | |
ec51d3fa XY |
1985 | |
1986 | parse_soc_bounding_box: | |
ec51d3fa XY |
1987 | /* |
1988 | * soc bounding box info is not integrated in disocovery table, | |
258620d0 | 1989 | * we always need to parse it from gpu info firmware if needed. |
ec51d3fa | 1990 | */ |
48321c3d HW |
1991 | if (hdr->version_minor == 2) { |
1992 | const struct gpu_info_firmware_v1_2 *gpu_info_fw = | |
1993 | (const struct gpu_info_firmware_v1_2 *)(adev->firmware.gpu_info_fw->data + | |
1994 | le32_to_cpu(hdr->header.ucode_array_offset_bytes)); | |
1995 | adev->dm.soc_bounding_box = &gpu_info_fw->soc_bounding_box; | |
1996 | } | |
e2a75f88 AD |
1997 | break; |
1998 | } | |
1999 | default: | |
2000 | dev_err(adev->dev, | |
2001 | "Unsupported gpu_info table %d\n", hdr->header.ucode_version); | |
2002 | err = -EINVAL; | |
2003 | goto out; | |
2004 | } | |
2005 | out: | |
e2a75f88 AD |
2006 | return err; |
2007 | } | |
2008 | ||
e3ecdffa AD |
2009 | /** |
2010 | * amdgpu_device_ip_early_init - run early init for hardware IPs | |
2011 | * | |
2012 | * @adev: amdgpu_device pointer | |
2013 | * | |
2014 | * Early initialization pass for hardware IPs. The hardware IPs that make | |
2015 | * up each asic are discovered each IP's early_init callback is run. This | |
2016 | * is the first stage in initializing the asic. | |
2017 | * Returns 0 on success, negative error code on failure. | |
2018 | */ | |
06ec9070 | 2019 | static int amdgpu_device_ip_early_init(struct amdgpu_device *adev) |
d38ceaf9 | 2020 | { |
901e2be2 AD |
2021 | struct drm_device *dev = adev_to_drm(adev); |
2022 | struct pci_dev *parent; | |
aaa36a97 | 2023 | int i, r; |
ced69502 | 2024 | bool total; |
d38ceaf9 | 2025 | |
483ef985 | 2026 | amdgpu_device_enable_virtual_display(adev); |
a6be7570 | 2027 | |
00a979f3 | 2028 | if (amdgpu_sriov_vf(adev)) { |
00a979f3 | 2029 | r = amdgpu_virt_request_full_gpu(adev, true); |
aaa36a97 AD |
2030 | if (r) |
2031 | return r; | |
00a979f3 WS |
2032 | } |
2033 | ||
d38ceaf9 | 2034 | switch (adev->asic_type) { |
33f34802 KW |
2035 | #ifdef CONFIG_DRM_AMDGPU_SI |
2036 | case CHIP_VERDE: | |
2037 | case CHIP_TAHITI: | |
2038 | case CHIP_PITCAIRN: | |
2039 | case CHIP_OLAND: | |
2040 | case CHIP_HAINAN: | |
295d0daf | 2041 | adev->family = AMDGPU_FAMILY_SI; |
33f34802 KW |
2042 | r = si_set_ip_blocks(adev); |
2043 | if (r) | |
2044 | return r; | |
2045 | break; | |
2046 | #endif | |
a2e73f56 AD |
2047 | #ifdef CONFIG_DRM_AMDGPU_CIK |
2048 | case CHIP_BONAIRE: | |
2049 | case CHIP_HAWAII: | |
2050 | case CHIP_KAVERI: | |
2051 | case CHIP_KABINI: | |
2052 | case CHIP_MULLINS: | |
e1ad2d53 | 2053 | if (adev->flags & AMD_IS_APU) |
a2e73f56 | 2054 | adev->family = AMDGPU_FAMILY_KV; |
e1ad2d53 AD |
2055 | else |
2056 | adev->family = AMDGPU_FAMILY_CI; | |
a2e73f56 AD |
2057 | |
2058 | r = cik_set_ip_blocks(adev); | |
2059 | if (r) | |
2060 | return r; | |
2061 | break; | |
2062 | #endif | |
da87c30b AD |
2063 | case CHIP_TOPAZ: |
2064 | case CHIP_TONGA: | |
2065 | case CHIP_FIJI: | |
2066 | case CHIP_POLARIS10: | |
2067 | case CHIP_POLARIS11: | |
2068 | case CHIP_POLARIS12: | |
2069 | case CHIP_VEGAM: | |
2070 | case CHIP_CARRIZO: | |
2071 | case CHIP_STONEY: | |
2072 | if (adev->flags & AMD_IS_APU) | |
2073 | adev->family = AMDGPU_FAMILY_CZ; | |
2074 | else | |
2075 | adev->family = AMDGPU_FAMILY_VI; | |
2076 | ||
2077 | r = vi_set_ip_blocks(adev); | |
2078 | if (r) | |
2079 | return r; | |
2080 | break; | |
d38ceaf9 | 2081 | default: |
63352b7f AD |
2082 | r = amdgpu_discovery_set_ip_blocks(adev); |
2083 | if (r) | |
2084 | return r; | |
2085 | break; | |
d38ceaf9 AD |
2086 | } |
2087 | ||
901e2be2 AD |
2088 | if (amdgpu_has_atpx() && |
2089 | (amdgpu_is_atpx_hybrid() || | |
2090 | amdgpu_has_atpx_dgpu_power_cntl()) && | |
2091 | ((adev->flags & AMD_IS_APU) == 0) && | |
2092 | !pci_is_thunderbolt_attached(to_pci_dev(dev->dev))) | |
2093 | adev->flags |= AMD_IS_PX; | |
2094 | ||
85ac2021 AD |
2095 | if (!(adev->flags & AMD_IS_APU)) { |
2096 | parent = pci_upstream_bridge(adev->pdev); | |
2097 | adev->has_pr3 = parent ? pci_pr3_present(parent) : false; | |
2098 | } | |
901e2be2 | 2099 | |
1884734a | 2100 | |
3b94fb10 | 2101 | adev->pm.pp_feature = amdgpu_pp_feature_mask; |
a35ad98b | 2102 | if (amdgpu_sriov_vf(adev) || sched_policy == KFD_SCHED_POLICY_NO_HWS) |
00544006 | 2103 | adev->pm.pp_feature &= ~PP_GFXOFF_MASK; |
4215a119 HC |
2104 | if (amdgpu_sriov_vf(adev) && adev->asic_type == CHIP_SIENNA_CICHLID) |
2105 | adev->pm.pp_feature &= ~PP_OVERDRIVE_MASK; | |
00f54b97 | 2106 | |
ced69502 | 2107 | total = true; |
d38ceaf9 AD |
2108 | for (i = 0; i < adev->num_ip_blocks; i++) { |
2109 | if ((amdgpu_ip_block_mask & (1 << i)) == 0) { | |
0c451baf | 2110 | DRM_WARN("disabled ip block: %d <%s>\n", |
ed8cf00c | 2111 | i, adev->ip_blocks[i].version->funcs->name); |
a1255107 | 2112 | adev->ip_blocks[i].status.valid = false; |
d38ceaf9 | 2113 | } else { |
a1255107 AD |
2114 | if (adev->ip_blocks[i].version->funcs->early_init) { |
2115 | r = adev->ip_blocks[i].version->funcs->early_init((void *)adev); | |
2c1a2784 | 2116 | if (r == -ENOENT) { |
a1255107 | 2117 | adev->ip_blocks[i].status.valid = false; |
2c1a2784 | 2118 | } else if (r) { |
a1255107 AD |
2119 | DRM_ERROR("early_init of IP block <%s> failed %d\n", |
2120 | adev->ip_blocks[i].version->funcs->name, r); | |
ced69502 | 2121 | total = false; |
2c1a2784 | 2122 | } else { |
a1255107 | 2123 | adev->ip_blocks[i].status.valid = true; |
2c1a2784 | 2124 | } |
974e6b64 | 2125 | } else { |
a1255107 | 2126 | adev->ip_blocks[i].status.valid = true; |
d38ceaf9 | 2127 | } |
d38ceaf9 | 2128 | } |
21a249ca AD |
2129 | /* get the vbios after the asic_funcs are set up */ |
2130 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) { | |
6e29c227 AD |
2131 | r = amdgpu_device_parse_gpu_info_fw(adev); |
2132 | if (r) | |
2133 | return r; | |
2134 | ||
21a249ca | 2135 | /* Read BIOS */ |
9535a86a SZ |
2136 | if (amdgpu_device_read_bios(adev)) { |
2137 | if (!amdgpu_get_bios(adev)) | |
2138 | return -EINVAL; | |
21a249ca | 2139 | |
9535a86a SZ |
2140 | r = amdgpu_atombios_init(adev); |
2141 | if (r) { | |
2142 | dev_err(adev->dev, "amdgpu_atombios_init failed\n"); | |
2143 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0); | |
2144 | return r; | |
2145 | } | |
21a249ca | 2146 | } |
77eabc6f PJZ |
2147 | |
2148 | /*get pf2vf msg info at it's earliest time*/ | |
2149 | if (amdgpu_sriov_vf(adev)) | |
2150 | amdgpu_virt_init_data_exchange(adev); | |
2151 | ||
21a249ca | 2152 | } |
d38ceaf9 | 2153 | } |
ced69502 ML |
2154 | if (!total) |
2155 | return -ENODEV; | |
d38ceaf9 | 2156 | |
00fa4035 | 2157 | amdgpu_amdkfd_device_probe(adev); |
395d1fb9 NH |
2158 | adev->cg_flags &= amdgpu_cg_mask; |
2159 | adev->pg_flags &= amdgpu_pg_mask; | |
2160 | ||
d38ceaf9 AD |
2161 | return 0; |
2162 | } | |
2163 | ||
0a4f2520 RZ |
2164 | static int amdgpu_device_ip_hw_init_phase1(struct amdgpu_device *adev) |
2165 | { | |
2166 | int i, r; | |
2167 | ||
2168 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
2169 | if (!adev->ip_blocks[i].status.sw) | |
2170 | continue; | |
2171 | if (adev->ip_blocks[i].status.hw) | |
2172 | continue; | |
2173 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || | |
2d11fd3f | 2174 | (amdgpu_sriov_vf(adev) && (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)) || |
0a4f2520 RZ |
2175 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) { |
2176 | r = adev->ip_blocks[i].version->funcs->hw_init(adev); | |
2177 | if (r) { | |
2178 | DRM_ERROR("hw_init of IP block <%s> failed %d\n", | |
2179 | adev->ip_blocks[i].version->funcs->name, r); | |
2180 | return r; | |
2181 | } | |
2182 | adev->ip_blocks[i].status.hw = true; | |
2183 | } | |
2184 | } | |
2185 | ||
2186 | return 0; | |
2187 | } | |
2188 | ||
2189 | static int amdgpu_device_ip_hw_init_phase2(struct amdgpu_device *adev) | |
2190 | { | |
2191 | int i, r; | |
2192 | ||
2193 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
2194 | if (!adev->ip_blocks[i].status.sw) | |
2195 | continue; | |
2196 | if (adev->ip_blocks[i].status.hw) | |
2197 | continue; | |
2198 | r = adev->ip_blocks[i].version->funcs->hw_init(adev); | |
2199 | if (r) { | |
2200 | DRM_ERROR("hw_init of IP block <%s> failed %d\n", | |
2201 | adev->ip_blocks[i].version->funcs->name, r); | |
2202 | return r; | |
2203 | } | |
2204 | adev->ip_blocks[i].status.hw = true; | |
2205 | } | |
2206 | ||
2207 | return 0; | |
2208 | } | |
2209 | ||
7a3e0bb2 RZ |
2210 | static int amdgpu_device_fw_loading(struct amdgpu_device *adev) |
2211 | { | |
2212 | int r = 0; | |
2213 | int i; | |
80f41f84 | 2214 | uint32_t smu_version; |
7a3e0bb2 RZ |
2215 | |
2216 | if (adev->asic_type >= CHIP_VEGA10) { | |
2217 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
482f0e53 ML |
2218 | if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_PSP) |
2219 | continue; | |
2220 | ||
e3c1b071 | 2221 | if (!adev->ip_blocks[i].status.sw) |
2222 | continue; | |
2223 | ||
482f0e53 ML |
2224 | /* no need to do the fw loading again if already done*/ |
2225 | if (adev->ip_blocks[i].status.hw == true) | |
2226 | break; | |
2227 | ||
53b3f8f4 | 2228 | if (amdgpu_in_reset(adev) || adev->in_suspend) { |
482f0e53 ML |
2229 | r = adev->ip_blocks[i].version->funcs->resume(adev); |
2230 | if (r) { | |
2231 | DRM_ERROR("resume of IP block <%s> failed %d\n", | |
7a3e0bb2 | 2232 | adev->ip_blocks[i].version->funcs->name, r); |
482f0e53 ML |
2233 | return r; |
2234 | } | |
2235 | } else { | |
2236 | r = adev->ip_blocks[i].version->funcs->hw_init(adev); | |
2237 | if (r) { | |
2238 | DRM_ERROR("hw_init of IP block <%s> failed %d\n", | |
2239 | adev->ip_blocks[i].version->funcs->name, r); | |
2240 | return r; | |
7a3e0bb2 | 2241 | } |
7a3e0bb2 | 2242 | } |
482f0e53 ML |
2243 | |
2244 | adev->ip_blocks[i].status.hw = true; | |
2245 | break; | |
7a3e0bb2 RZ |
2246 | } |
2247 | } | |
482f0e53 | 2248 | |
8973d9ec ED |
2249 | if (!amdgpu_sriov_vf(adev) || adev->asic_type == CHIP_TONGA) |
2250 | r = amdgpu_pm_load_smu_firmware(adev, &smu_version); | |
7a3e0bb2 | 2251 | |
80f41f84 | 2252 | return r; |
7a3e0bb2 RZ |
2253 | } |
2254 | ||
5fd8518d AG |
2255 | static int amdgpu_device_init_schedulers(struct amdgpu_device *adev) |
2256 | { | |
2257 | long timeout; | |
2258 | int r, i; | |
2259 | ||
2260 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { | |
2261 | struct amdgpu_ring *ring = adev->rings[i]; | |
2262 | ||
2263 | /* No need to setup the GPU scheduler for rings that don't need it */ | |
2264 | if (!ring || ring->no_scheduler) | |
2265 | continue; | |
2266 | ||
2267 | switch (ring->funcs->type) { | |
2268 | case AMDGPU_RING_TYPE_GFX: | |
2269 | timeout = adev->gfx_timeout; | |
2270 | break; | |
2271 | case AMDGPU_RING_TYPE_COMPUTE: | |
2272 | timeout = adev->compute_timeout; | |
2273 | break; | |
2274 | case AMDGPU_RING_TYPE_SDMA: | |
2275 | timeout = adev->sdma_timeout; | |
2276 | break; | |
2277 | default: | |
2278 | timeout = adev->video_timeout; | |
2279 | break; | |
2280 | } | |
2281 | ||
2282 | r = drm_sched_init(&ring->sched, &amdgpu_sched_ops, | |
56e44960 | 2283 | DRM_SCHED_PRIORITY_COUNT, |
11f25c84 | 2284 | ring->num_hw_submission, 0, |
8ab62eda JG |
2285 | timeout, adev->reset_domain->wq, |
2286 | ring->sched_score, ring->name, | |
2287 | adev->dev); | |
5fd8518d AG |
2288 | if (r) { |
2289 | DRM_ERROR("Failed to create scheduler on ring %s.\n", | |
2290 | ring->name); | |
2291 | return r; | |
2292 | } | |
2293 | } | |
2294 | ||
d425c6f4 JZ |
2295 | amdgpu_xcp_update_partition_sched_list(adev); |
2296 | ||
5fd8518d AG |
2297 | return 0; |
2298 | } | |
2299 | ||
2300 | ||
e3ecdffa AD |
2301 | /** |
2302 | * amdgpu_device_ip_init - run init for hardware IPs | |
2303 | * | |
2304 | * @adev: amdgpu_device pointer | |
2305 | * | |
2306 | * Main initialization pass for hardware IPs. The list of all the hardware | |
2307 | * IPs that make up the asic is walked and the sw_init and hw_init callbacks | |
2308 | * are run. sw_init initializes the software state associated with each IP | |
2309 | * and hw_init initializes the hardware associated with each IP. | |
2310 | * Returns 0 on success, negative error code on failure. | |
2311 | */ | |
06ec9070 | 2312 | static int amdgpu_device_ip_init(struct amdgpu_device *adev) |
d38ceaf9 AD |
2313 | { |
2314 | int i, r; | |
2315 | ||
c030f2e4 | 2316 | r = amdgpu_ras_init(adev); |
2317 | if (r) | |
2318 | return r; | |
2319 | ||
d38ceaf9 | 2320 | for (i = 0; i < adev->num_ip_blocks; i++) { |
a1255107 | 2321 | if (!adev->ip_blocks[i].status.valid) |
d38ceaf9 | 2322 | continue; |
a1255107 | 2323 | r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev); |
2c1a2784 | 2324 | if (r) { |
a1255107 AD |
2325 | DRM_ERROR("sw_init of IP block <%s> failed %d\n", |
2326 | adev->ip_blocks[i].version->funcs->name, r); | |
72d3f592 | 2327 | goto init_failed; |
2c1a2784 | 2328 | } |
a1255107 | 2329 | adev->ip_blocks[i].status.sw = true; |
bfca0289 | 2330 | |
c1c39032 AD |
2331 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) { |
2332 | /* need to do common hw init early so everything is set up for gmc */ | |
2333 | r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev); | |
2334 | if (r) { | |
2335 | DRM_ERROR("hw_init %d failed %d\n", i, r); | |
2336 | goto init_failed; | |
2337 | } | |
2338 | adev->ip_blocks[i].status.hw = true; | |
2339 | } else if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) { | |
2340 | /* need to do gmc hw init early so we can allocate gpu mem */ | |
892deb48 VS |
2341 | /* Try to reserve bad pages early */ |
2342 | if (amdgpu_sriov_vf(adev)) | |
2343 | amdgpu_virt_exchange_data(adev); | |
2344 | ||
7ccfd79f | 2345 | r = amdgpu_device_mem_scratch_init(adev); |
2c1a2784 | 2346 | if (r) { |
7ccfd79f | 2347 | DRM_ERROR("amdgpu_mem_scratch_init failed %d\n", r); |
72d3f592 | 2348 | goto init_failed; |
2c1a2784 | 2349 | } |
a1255107 | 2350 | r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev); |
2c1a2784 AD |
2351 | if (r) { |
2352 | DRM_ERROR("hw_init %d failed %d\n", i, r); | |
72d3f592 | 2353 | goto init_failed; |
2c1a2784 | 2354 | } |
06ec9070 | 2355 | r = amdgpu_device_wb_init(adev); |
2c1a2784 | 2356 | if (r) { |
06ec9070 | 2357 | DRM_ERROR("amdgpu_device_wb_init failed %d\n", r); |
72d3f592 | 2358 | goto init_failed; |
2c1a2784 | 2359 | } |
a1255107 | 2360 | adev->ip_blocks[i].status.hw = true; |
2493664f ML |
2361 | |
2362 | /* right after GMC hw init, we create CSA */ | |
02ff519e | 2363 | if (adev->gfx.mcbp) { |
1e256e27 | 2364 | r = amdgpu_allocate_static_csa(adev, &adev->virt.csa_obj, |
58ab2c08 CK |
2365 | AMDGPU_GEM_DOMAIN_VRAM | |
2366 | AMDGPU_GEM_DOMAIN_GTT, | |
2367 | AMDGPU_CSA_SIZE); | |
2493664f ML |
2368 | if (r) { |
2369 | DRM_ERROR("allocate CSA failed %d\n", r); | |
72d3f592 | 2370 | goto init_failed; |
2493664f ML |
2371 | } |
2372 | } | |
d38ceaf9 AD |
2373 | } |
2374 | } | |
2375 | ||
c9ffa427 | 2376 | if (amdgpu_sriov_vf(adev)) |
22c16d25 | 2377 | amdgpu_virt_init_data_exchange(adev); |
c9ffa427 | 2378 | |
533aed27 AG |
2379 | r = amdgpu_ib_pool_init(adev); |
2380 | if (r) { | |
2381 | dev_err(adev->dev, "IB initialization failed (%d).\n", r); | |
2382 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r); | |
2383 | goto init_failed; | |
2384 | } | |
2385 | ||
c8963ea4 RZ |
2386 | r = amdgpu_ucode_create_bo(adev); /* create ucode bo when sw_init complete*/ |
2387 | if (r) | |
72d3f592 | 2388 | goto init_failed; |
0a4f2520 RZ |
2389 | |
2390 | r = amdgpu_device_ip_hw_init_phase1(adev); | |
2391 | if (r) | |
72d3f592 | 2392 | goto init_failed; |
0a4f2520 | 2393 | |
7a3e0bb2 RZ |
2394 | r = amdgpu_device_fw_loading(adev); |
2395 | if (r) | |
72d3f592 | 2396 | goto init_failed; |
7a3e0bb2 | 2397 | |
0a4f2520 RZ |
2398 | r = amdgpu_device_ip_hw_init_phase2(adev); |
2399 | if (r) | |
72d3f592 | 2400 | goto init_failed; |
d38ceaf9 | 2401 | |
121a2bc6 AG |
2402 | /* |
2403 | * retired pages will be loaded from eeprom and reserved here, | |
2404 | * it should be called after amdgpu_device_ip_hw_init_phase2 since | |
2405 | * for some ASICs the RAS EEPROM code relies on SMU fully functioning | |
2406 | * for I2C communication which only true at this point. | |
b82e65a9 GC |
2407 | * |
2408 | * amdgpu_ras_recovery_init may fail, but the upper only cares the | |
2409 | * failure from bad gpu situation and stop amdgpu init process | |
2410 | * accordingly. For other failed cases, it will still release all | |
2411 | * the resource and print error message, rather than returning one | |
2412 | * negative value to upper level. | |
121a2bc6 AG |
2413 | * |
2414 | * Note: theoretically, this should be called before all vram allocations | |
2415 | * to protect retired page from abusing | |
2416 | */ | |
b82e65a9 GC |
2417 | r = amdgpu_ras_recovery_init(adev); |
2418 | if (r) | |
2419 | goto init_failed; | |
121a2bc6 | 2420 | |
cfbb6b00 AG |
2421 | /** |
2422 | * In case of XGMI grab extra reference for reset domain for this device | |
2423 | */ | |
a4c63caf | 2424 | if (adev->gmc.xgmi.num_physical_nodes > 1) { |
cfbb6b00 | 2425 | if (amdgpu_xgmi_add_device(adev) == 0) { |
46c67660 | 2426 | if (!amdgpu_sriov_vf(adev)) { |
2efc30f0 VC |
2427 | struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev); |
2428 | ||
dfd0287b LH |
2429 | if (WARN_ON(!hive)) { |
2430 | r = -ENOENT; | |
2431 | goto init_failed; | |
2432 | } | |
2433 | ||
46c67660 | 2434 | if (!hive->reset_domain || |
2435 | !amdgpu_reset_get_reset_domain(hive->reset_domain)) { | |
2436 | r = -ENOENT; | |
2437 | amdgpu_put_xgmi_hive(hive); | |
2438 | goto init_failed; | |
2439 | } | |
2440 | ||
2441 | /* Drop the early temporary reset domain we created for device */ | |
2442 | amdgpu_reset_put_reset_domain(adev->reset_domain); | |
2443 | adev->reset_domain = hive->reset_domain; | |
9dfa4860 | 2444 | amdgpu_put_xgmi_hive(hive); |
cfbb6b00 | 2445 | } |
a4c63caf AG |
2446 | } |
2447 | } | |
2448 | ||
5fd8518d AG |
2449 | r = amdgpu_device_init_schedulers(adev); |
2450 | if (r) | |
2451 | goto init_failed; | |
e3c1b071 | 2452 | |
2453 | /* Don't init kfd if whole hive need to be reset during init */ | |
84b4dd3f PY |
2454 | if (!adev->gmc.xgmi.pending_reset) { |
2455 | kgd2kfd_init_zone_device(adev); | |
e3c1b071 | 2456 | amdgpu_amdkfd_device_init(adev); |
84b4dd3f | 2457 | } |
c6332b97 | 2458 | |
bd607166 KR |
2459 | amdgpu_fru_get_product_info(adev); |
2460 | ||
72d3f592 | 2461 | init_failed: |
c6332b97 | 2462 | |
72d3f592 | 2463 | return r; |
d38ceaf9 AD |
2464 | } |
2465 | ||
e3ecdffa AD |
2466 | /** |
2467 | * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer | |
2468 | * | |
2469 | * @adev: amdgpu_device pointer | |
2470 | * | |
2471 | * Writes a reset magic value to the gart pointer in VRAM. The driver calls | |
2472 | * this function before a GPU reset. If the value is retained after a | |
2473 | * GPU reset, VRAM has not been lost. Some GPU resets may destry VRAM contents. | |
2474 | */ | |
06ec9070 | 2475 | static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev) |
0c49e0b8 CZ |
2476 | { |
2477 | memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM); | |
2478 | } | |
2479 | ||
e3ecdffa AD |
2480 | /** |
2481 | * amdgpu_device_check_vram_lost - check if vram is valid | |
2482 | * | |
2483 | * @adev: amdgpu_device pointer | |
2484 | * | |
2485 | * Checks the reset magic value written to the gart pointer in VRAM. | |
2486 | * The driver calls this after a GPU reset to see if the contents of | |
2487 | * VRAM is lost or now. | |
2488 | * returns true if vram is lost, false if not. | |
2489 | */ | |
06ec9070 | 2490 | static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev) |
0c49e0b8 | 2491 | { |
dadce777 EQ |
2492 | if (memcmp(adev->gart.ptr, adev->reset_magic, |
2493 | AMDGPU_RESET_MAGIC_NUM)) | |
2494 | return true; | |
2495 | ||
53b3f8f4 | 2496 | if (!amdgpu_in_reset(adev)) |
dadce777 EQ |
2497 | return false; |
2498 | ||
2499 | /* | |
2500 | * For all ASICs with baco/mode1 reset, the VRAM is | |
2501 | * always assumed to be lost. | |
2502 | */ | |
2503 | switch (amdgpu_asic_reset_method(adev)) { | |
2504 | case AMD_RESET_METHOD_BACO: | |
2505 | case AMD_RESET_METHOD_MODE1: | |
2506 | return true; | |
2507 | default: | |
2508 | return false; | |
2509 | } | |
0c49e0b8 CZ |
2510 | } |
2511 | ||
e3ecdffa | 2512 | /** |
1112a46b | 2513 | * amdgpu_device_set_cg_state - set clockgating for amdgpu device |
e3ecdffa AD |
2514 | * |
2515 | * @adev: amdgpu_device pointer | |
b8b72130 | 2516 | * @state: clockgating state (gate or ungate) |
e3ecdffa | 2517 | * |
e3ecdffa | 2518 | * The list of all the hardware IPs that make up the asic is walked and the |
1112a46b RZ |
2519 | * set_clockgating_state callbacks are run. |
2520 | * Late initialization pass enabling clockgating for hardware IPs. | |
2521 | * Fini or suspend, pass disabling clockgating for hardware IPs. | |
e3ecdffa AD |
2522 | * Returns 0 on success, negative error code on failure. |
2523 | */ | |
fdd34271 | 2524 | |
5d89bb2d LL |
2525 | int amdgpu_device_set_cg_state(struct amdgpu_device *adev, |
2526 | enum amd_clockgating_state state) | |
d38ceaf9 | 2527 | { |
1112a46b | 2528 | int i, j, r; |
d38ceaf9 | 2529 | |
4a2ba394 SL |
2530 | if (amdgpu_emu_mode == 1) |
2531 | return 0; | |
2532 | ||
1112a46b RZ |
2533 | for (j = 0; j < adev->num_ip_blocks; j++) { |
2534 | i = state == AMD_CG_STATE_GATE ? j : adev->num_ip_blocks - j - 1; | |
a2d31dc3 | 2535 | if (!adev->ip_blocks[i].status.late_initialized) |
d38ceaf9 | 2536 | continue; |
47198eb7 | 2537 | /* skip CG for GFX, SDMA on S0ix */ |
5d70a549 | 2538 | if (adev->in_s0ix && |
47198eb7 AD |
2539 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX || |
2540 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SDMA)) | |
5d70a549 | 2541 | continue; |
4a446d55 | 2542 | /* skip CG for VCE/UVD, it's handled specially */ |
a1255107 | 2543 | if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD && |
57716327 | 2544 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE && |
34319b32 | 2545 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN && |
52f2e779 | 2546 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG && |
57716327 | 2547 | adev->ip_blocks[i].version->funcs->set_clockgating_state) { |
4a446d55 | 2548 | /* enable clockgating to save power */ |
a1255107 | 2549 | r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev, |
1112a46b | 2550 | state); |
4a446d55 AD |
2551 | if (r) { |
2552 | DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n", | |
a1255107 | 2553 | adev->ip_blocks[i].version->funcs->name, r); |
4a446d55 AD |
2554 | return r; |
2555 | } | |
b0b00ff1 | 2556 | } |
d38ceaf9 | 2557 | } |
06b18f61 | 2558 | |
c9f96fd5 RZ |
2559 | return 0; |
2560 | } | |
2561 | ||
5d89bb2d LL |
2562 | int amdgpu_device_set_pg_state(struct amdgpu_device *adev, |
2563 | enum amd_powergating_state state) | |
c9f96fd5 | 2564 | { |
1112a46b | 2565 | int i, j, r; |
06b18f61 | 2566 | |
c9f96fd5 RZ |
2567 | if (amdgpu_emu_mode == 1) |
2568 | return 0; | |
2569 | ||
1112a46b RZ |
2570 | for (j = 0; j < adev->num_ip_blocks; j++) { |
2571 | i = state == AMD_PG_STATE_GATE ? j : adev->num_ip_blocks - j - 1; | |
a2d31dc3 | 2572 | if (!adev->ip_blocks[i].status.late_initialized) |
c9f96fd5 | 2573 | continue; |
47198eb7 | 2574 | /* skip PG for GFX, SDMA on S0ix */ |
5d70a549 | 2575 | if (adev->in_s0ix && |
47198eb7 AD |
2576 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX || |
2577 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SDMA)) | |
5d70a549 | 2578 | continue; |
c9f96fd5 RZ |
2579 | /* skip CG for VCE/UVD, it's handled specially */ |
2580 | if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD && | |
2581 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE && | |
2582 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN && | |
52f2e779 | 2583 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG && |
c9f96fd5 RZ |
2584 | adev->ip_blocks[i].version->funcs->set_powergating_state) { |
2585 | /* enable powergating to save power */ | |
2586 | r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev, | |
1112a46b | 2587 | state); |
c9f96fd5 RZ |
2588 | if (r) { |
2589 | DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n", | |
2590 | adev->ip_blocks[i].version->funcs->name, r); | |
2591 | return r; | |
2592 | } | |
2593 | } | |
2594 | } | |
2dc80b00 S |
2595 | return 0; |
2596 | } | |
2597 | ||
beff74bc AD |
2598 | static int amdgpu_device_enable_mgpu_fan_boost(void) |
2599 | { | |
2600 | struct amdgpu_gpu_instance *gpu_ins; | |
2601 | struct amdgpu_device *adev; | |
2602 | int i, ret = 0; | |
2603 | ||
2604 | mutex_lock(&mgpu_info.mutex); | |
2605 | ||
2606 | /* | |
2607 | * MGPU fan boost feature should be enabled | |
2608 | * only when there are two or more dGPUs in | |
2609 | * the system | |
2610 | */ | |
2611 | if (mgpu_info.num_dgpu < 2) | |
2612 | goto out; | |
2613 | ||
2614 | for (i = 0; i < mgpu_info.num_dgpu; i++) { | |
2615 | gpu_ins = &(mgpu_info.gpu_ins[i]); | |
2616 | adev = gpu_ins->adev; | |
2617 | if (!(adev->flags & AMD_IS_APU) && | |
f10bb940 | 2618 | !gpu_ins->mgpu_fan_enabled) { |
beff74bc AD |
2619 | ret = amdgpu_dpm_enable_mgpu_fan_boost(adev); |
2620 | if (ret) | |
2621 | break; | |
2622 | ||
2623 | gpu_ins->mgpu_fan_enabled = 1; | |
2624 | } | |
2625 | } | |
2626 | ||
2627 | out: | |
2628 | mutex_unlock(&mgpu_info.mutex); | |
2629 | ||
2630 | return ret; | |
2631 | } | |
2632 | ||
e3ecdffa AD |
2633 | /** |
2634 | * amdgpu_device_ip_late_init - run late init for hardware IPs | |
2635 | * | |
2636 | * @adev: amdgpu_device pointer | |
2637 | * | |
2638 | * Late initialization pass for hardware IPs. The list of all the hardware | |
2639 | * IPs that make up the asic is walked and the late_init callbacks are run. | |
2640 | * late_init covers any special initialization that an IP requires | |
2641 | * after all of the have been initialized or something that needs to happen | |
2642 | * late in the init process. | |
2643 | * Returns 0 on success, negative error code on failure. | |
2644 | */ | |
06ec9070 | 2645 | static int amdgpu_device_ip_late_init(struct amdgpu_device *adev) |
2dc80b00 | 2646 | { |
60599a03 | 2647 | struct amdgpu_gpu_instance *gpu_instance; |
2dc80b00 S |
2648 | int i = 0, r; |
2649 | ||
2650 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
73f847db | 2651 | if (!adev->ip_blocks[i].status.hw) |
2dc80b00 S |
2652 | continue; |
2653 | if (adev->ip_blocks[i].version->funcs->late_init) { | |
2654 | r = adev->ip_blocks[i].version->funcs->late_init((void *)adev); | |
2655 | if (r) { | |
2656 | DRM_ERROR("late_init of IP block <%s> failed %d\n", | |
2657 | adev->ip_blocks[i].version->funcs->name, r); | |
2658 | return r; | |
2659 | } | |
2dc80b00 | 2660 | } |
73f847db | 2661 | adev->ip_blocks[i].status.late_initialized = true; |
2dc80b00 S |
2662 | } |
2663 | ||
867e24ca | 2664 | r = amdgpu_ras_late_init(adev); |
2665 | if (r) { | |
2666 | DRM_ERROR("amdgpu_ras_late_init failed %d", r); | |
2667 | return r; | |
2668 | } | |
2669 | ||
a891d239 DL |
2670 | amdgpu_ras_set_error_query_ready(adev, true); |
2671 | ||
1112a46b RZ |
2672 | amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE); |
2673 | amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE); | |
916ac57f | 2674 | |
06ec9070 | 2675 | amdgpu_device_fill_reset_magic(adev); |
d38ceaf9 | 2676 | |
beff74bc AD |
2677 | r = amdgpu_device_enable_mgpu_fan_boost(); |
2678 | if (r) | |
2679 | DRM_ERROR("enable mgpu fan boost failed (%d).\n", r); | |
2680 | ||
4da8b639 | 2681 | /* For passthrough configuration on arcturus and aldebaran, enable special handling SBR */ |
47fc644f SS |
2682 | if (amdgpu_passthrough(adev) && |
2683 | ((adev->asic_type == CHIP_ARCTURUS && adev->gmc.xgmi.num_physical_nodes > 1) || | |
2684 | adev->asic_type == CHIP_ALDEBARAN)) | |
bc143d8b | 2685 | amdgpu_dpm_handle_passthrough_sbr(adev, true); |
60599a03 EQ |
2686 | |
2687 | if (adev->gmc.xgmi.num_physical_nodes > 1) { | |
2688 | mutex_lock(&mgpu_info.mutex); | |
2689 | ||
2690 | /* | |
2691 | * Reset device p-state to low as this was booted with high. | |
2692 | * | |
2693 | * This should be performed only after all devices from the same | |
2694 | * hive get initialized. | |
2695 | * | |
2696 | * However, it's unknown how many device in the hive in advance. | |
2697 | * As this is counted one by one during devices initializations. | |
2698 | * | |
2699 | * So, we wait for all XGMI interlinked devices initialized. | |
2700 | * This may bring some delays as those devices may come from | |
2701 | * different hives. But that should be OK. | |
2702 | */ | |
2703 | if (mgpu_info.num_dgpu == adev->gmc.xgmi.num_physical_nodes) { | |
2704 | for (i = 0; i < mgpu_info.num_gpu; i++) { | |
2705 | gpu_instance = &(mgpu_info.gpu_ins[i]); | |
2706 | if (gpu_instance->adev->flags & AMD_IS_APU) | |
2707 | continue; | |
2708 | ||
d84a430d JK |
2709 | r = amdgpu_xgmi_set_pstate(gpu_instance->adev, |
2710 | AMDGPU_XGMI_PSTATE_MIN); | |
60599a03 EQ |
2711 | if (r) { |
2712 | DRM_ERROR("pstate setting failed (%d).\n", r); | |
2713 | break; | |
2714 | } | |
2715 | } | |
2716 | } | |
2717 | ||
2718 | mutex_unlock(&mgpu_info.mutex); | |
2719 | } | |
2720 | ||
d38ceaf9 AD |
2721 | return 0; |
2722 | } | |
2723 | ||
613aa3ea LY |
2724 | /** |
2725 | * amdgpu_device_smu_fini_early - smu hw_fini wrapper | |
2726 | * | |
2727 | * @adev: amdgpu_device pointer | |
2728 | * | |
2729 | * For ASICs need to disable SMC first | |
2730 | */ | |
2731 | static void amdgpu_device_smu_fini_early(struct amdgpu_device *adev) | |
2732 | { | |
2733 | int i, r; | |
2734 | ||
2735 | if (adev->ip_versions[GC_HWIP][0] > IP_VERSION(9, 0, 0)) | |
2736 | return; | |
2737 | ||
2738 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
2739 | if (!adev->ip_blocks[i].status.hw) | |
2740 | continue; | |
2741 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) { | |
2742 | r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev); | |
2743 | /* XXX handle errors */ | |
2744 | if (r) { | |
2745 | DRM_DEBUG("hw_fini of IP block <%s> failed %d\n", | |
2746 | adev->ip_blocks[i].version->funcs->name, r); | |
2747 | } | |
2748 | adev->ip_blocks[i].status.hw = false; | |
2749 | break; | |
2750 | } | |
2751 | } | |
2752 | } | |
2753 | ||
e9669fb7 | 2754 | static int amdgpu_device_ip_fini_early(struct amdgpu_device *adev) |
d38ceaf9 AD |
2755 | { |
2756 | int i, r; | |
2757 | ||
e9669fb7 AG |
2758 | for (i = 0; i < adev->num_ip_blocks; i++) { |
2759 | if (!adev->ip_blocks[i].version->funcs->early_fini) | |
2760 | continue; | |
5278a159 | 2761 | |
e9669fb7 AG |
2762 | r = adev->ip_blocks[i].version->funcs->early_fini((void *)adev); |
2763 | if (r) { | |
2764 | DRM_DEBUG("early_fini of IP block <%s> failed %d\n", | |
2765 | adev->ip_blocks[i].version->funcs->name, r); | |
2766 | } | |
2767 | } | |
c030f2e4 | 2768 | |
05df1f01 | 2769 | amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE); |
fdd34271 RZ |
2770 | amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE); |
2771 | ||
7270e895 TY |
2772 | amdgpu_amdkfd_suspend(adev, false); |
2773 | ||
613aa3ea LY |
2774 | /* Workaroud for ASICs need to disable SMC first */ |
2775 | amdgpu_device_smu_fini_early(adev); | |
3e96dbfd | 2776 | |
d38ceaf9 | 2777 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { |
a1255107 | 2778 | if (!adev->ip_blocks[i].status.hw) |
d38ceaf9 | 2779 | continue; |
8201a67a | 2780 | |
a1255107 | 2781 | r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev); |
d38ceaf9 | 2782 | /* XXX handle errors */ |
2c1a2784 | 2783 | if (r) { |
a1255107 AD |
2784 | DRM_DEBUG("hw_fini of IP block <%s> failed %d\n", |
2785 | adev->ip_blocks[i].version->funcs->name, r); | |
2c1a2784 | 2786 | } |
8201a67a | 2787 | |
a1255107 | 2788 | adev->ip_blocks[i].status.hw = false; |
d38ceaf9 AD |
2789 | } |
2790 | ||
6effad8a GC |
2791 | if (amdgpu_sriov_vf(adev)) { |
2792 | if (amdgpu_virt_release_full_gpu(adev, false)) | |
2793 | DRM_ERROR("failed to release exclusive mode on fini\n"); | |
2794 | } | |
2795 | ||
e9669fb7 AG |
2796 | return 0; |
2797 | } | |
2798 | ||
2799 | /** | |
2800 | * amdgpu_device_ip_fini - run fini for hardware IPs | |
2801 | * | |
2802 | * @adev: amdgpu_device pointer | |
2803 | * | |
2804 | * Main teardown pass for hardware IPs. The list of all the hardware | |
2805 | * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks | |
2806 | * are run. hw_fini tears down the hardware associated with each IP | |
2807 | * and sw_fini tears down any software state associated with each IP. | |
2808 | * Returns 0 on success, negative error code on failure. | |
2809 | */ | |
2810 | static int amdgpu_device_ip_fini(struct amdgpu_device *adev) | |
2811 | { | |
2812 | int i, r; | |
2813 | ||
2814 | if (amdgpu_sriov_vf(adev) && adev->virt.ras_init_done) | |
2815 | amdgpu_virt_release_ras_err_handler_data(adev); | |
2816 | ||
e9669fb7 AG |
2817 | if (adev->gmc.xgmi.num_physical_nodes > 1) |
2818 | amdgpu_xgmi_remove_device(adev); | |
2819 | ||
c004d44e | 2820 | amdgpu_amdkfd_device_fini_sw(adev); |
9950cda2 | 2821 | |
d38ceaf9 | 2822 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { |
a1255107 | 2823 | if (!adev->ip_blocks[i].status.sw) |
d38ceaf9 | 2824 | continue; |
c12aba3a ML |
2825 | |
2826 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) { | |
c8963ea4 | 2827 | amdgpu_ucode_free_bo(adev); |
1e256e27 | 2828 | amdgpu_free_static_csa(&adev->virt.csa_obj); |
c12aba3a | 2829 | amdgpu_device_wb_fini(adev); |
7ccfd79f | 2830 | amdgpu_device_mem_scratch_fini(adev); |
533aed27 | 2831 | amdgpu_ib_pool_fini(adev); |
c12aba3a ML |
2832 | } |
2833 | ||
a1255107 | 2834 | r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev); |
d38ceaf9 | 2835 | /* XXX handle errors */ |
2c1a2784 | 2836 | if (r) { |
a1255107 AD |
2837 | DRM_DEBUG("sw_fini of IP block <%s> failed %d\n", |
2838 | adev->ip_blocks[i].version->funcs->name, r); | |
2c1a2784 | 2839 | } |
a1255107 AD |
2840 | adev->ip_blocks[i].status.sw = false; |
2841 | adev->ip_blocks[i].status.valid = false; | |
d38ceaf9 AD |
2842 | } |
2843 | ||
a6dcfd9c | 2844 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { |
a1255107 | 2845 | if (!adev->ip_blocks[i].status.late_initialized) |
8a2eef1d | 2846 | continue; |
a1255107 AD |
2847 | if (adev->ip_blocks[i].version->funcs->late_fini) |
2848 | adev->ip_blocks[i].version->funcs->late_fini((void *)adev); | |
2849 | adev->ip_blocks[i].status.late_initialized = false; | |
a6dcfd9c ML |
2850 | } |
2851 | ||
c030f2e4 | 2852 | amdgpu_ras_fini(adev); |
2853 | ||
d38ceaf9 AD |
2854 | return 0; |
2855 | } | |
2856 | ||
e3ecdffa | 2857 | /** |
beff74bc | 2858 | * amdgpu_device_delayed_init_work_handler - work handler for IB tests |
e3ecdffa | 2859 | * |
1112a46b | 2860 | * @work: work_struct. |
e3ecdffa | 2861 | */ |
beff74bc | 2862 | static void amdgpu_device_delayed_init_work_handler(struct work_struct *work) |
2dc80b00 S |
2863 | { |
2864 | struct amdgpu_device *adev = | |
beff74bc | 2865 | container_of(work, struct amdgpu_device, delayed_init_work.work); |
916ac57f RZ |
2866 | int r; |
2867 | ||
2868 | r = amdgpu_ib_ring_tests(adev); | |
2869 | if (r) | |
2870 | DRM_ERROR("ib ring test failed (%d).\n", r); | |
2dc80b00 S |
2871 | } |
2872 | ||
1e317b99 RZ |
2873 | static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work) |
2874 | { | |
2875 | struct amdgpu_device *adev = | |
2876 | container_of(work, struct amdgpu_device, gfx.gfx_off_delay_work.work); | |
2877 | ||
90a92662 MD |
2878 | WARN_ON_ONCE(adev->gfx.gfx_off_state); |
2879 | WARN_ON_ONCE(adev->gfx.gfx_off_req_count); | |
2880 | ||
2881 | if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, true)) | |
2882 | adev->gfx.gfx_off_state = true; | |
1e317b99 RZ |
2883 | } |
2884 | ||
e3ecdffa | 2885 | /** |
e7854a03 | 2886 | * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1) |
e3ecdffa AD |
2887 | * |
2888 | * @adev: amdgpu_device pointer | |
2889 | * | |
2890 | * Main suspend function for hardware IPs. The list of all the hardware | |
2891 | * IPs that make up the asic is walked, clockgating is disabled and the | |
2892 | * suspend callbacks are run. suspend puts the hardware and software state | |
2893 | * in each IP into a state suitable for suspend. | |
2894 | * Returns 0 on success, negative error code on failure. | |
2895 | */ | |
e7854a03 AD |
2896 | static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev) |
2897 | { | |
2898 | int i, r; | |
2899 | ||
50ec83f0 AD |
2900 | amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE); |
2901 | amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE); | |
05df1f01 | 2902 | |
b31d6ada EQ |
2903 | /* |
2904 | * Per PMFW team's suggestion, driver needs to handle gfxoff | |
2905 | * and df cstate features disablement for gpu reset(e.g. Mode1Reset) | |
2906 | * scenario. Add the missing df cstate disablement here. | |
2907 | */ | |
2908 | if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_DISALLOW)) | |
2909 | dev_warn(adev->dev, "Failed to disallow df cstate"); | |
2910 | ||
e7854a03 AD |
2911 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { |
2912 | if (!adev->ip_blocks[i].status.valid) | |
2913 | continue; | |
2b9f7848 | 2914 | |
e7854a03 | 2915 | /* displays are handled separately */ |
2b9f7848 ND |
2916 | if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_DCE) |
2917 | continue; | |
2918 | ||
2919 | /* XXX handle errors */ | |
2920 | r = adev->ip_blocks[i].version->funcs->suspend(adev); | |
2921 | /* XXX handle errors */ | |
2922 | if (r) { | |
2923 | DRM_ERROR("suspend of IP block <%s> failed %d\n", | |
2924 | adev->ip_blocks[i].version->funcs->name, r); | |
2925 | return r; | |
e7854a03 | 2926 | } |
2b9f7848 ND |
2927 | |
2928 | adev->ip_blocks[i].status.hw = false; | |
e7854a03 AD |
2929 | } |
2930 | ||
e7854a03 AD |
2931 | return 0; |
2932 | } | |
2933 | ||
2934 | /** | |
2935 | * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2) | |
2936 | * | |
2937 | * @adev: amdgpu_device pointer | |
2938 | * | |
2939 | * Main suspend function for hardware IPs. The list of all the hardware | |
2940 | * IPs that make up the asic is walked, clockgating is disabled and the | |
2941 | * suspend callbacks are run. suspend puts the hardware and software state | |
2942 | * in each IP into a state suitable for suspend. | |
2943 | * Returns 0 on success, negative error code on failure. | |
2944 | */ | |
2945 | static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev) | |
d38ceaf9 AD |
2946 | { |
2947 | int i, r; | |
2948 | ||
557f42a2 | 2949 | if (adev->in_s0ix) |
bc143d8b | 2950 | amdgpu_dpm_gfx_state_change(adev, sGpuChangeState_D3Entry); |
34416931 | 2951 | |
d38ceaf9 | 2952 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { |
a1255107 | 2953 | if (!adev->ip_blocks[i].status.valid) |
d38ceaf9 | 2954 | continue; |
e7854a03 AD |
2955 | /* displays are handled in phase1 */ |
2956 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) | |
2957 | continue; | |
bff77e86 LM |
2958 | /* PSP lost connection when err_event_athub occurs */ |
2959 | if (amdgpu_ras_intr_triggered() && | |
2960 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) { | |
2961 | adev->ip_blocks[i].status.hw = false; | |
2962 | continue; | |
2963 | } | |
e3c1b071 | 2964 | |
2965 | /* skip unnecessary suspend if we do not initialize them yet */ | |
2966 | if (adev->gmc.xgmi.pending_reset && | |
2967 | !(adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC || | |
2968 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC || | |
2969 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || | |
2970 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH)) { | |
2971 | adev->ip_blocks[i].status.hw = false; | |
2972 | continue; | |
2973 | } | |
557f42a2 | 2974 | |
afa6646b | 2975 | /* skip suspend of gfx/mes and psp for S0ix |
32ff160d AD |
2976 | * gfx is in gfxoff state, so on resume it will exit gfxoff just |
2977 | * like at runtime. PSP is also part of the always on hardware | |
2978 | * so no need to suspend it. | |
2979 | */ | |
557f42a2 | 2980 | if (adev->in_s0ix && |
32ff160d | 2981 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP || |
afa6646b AD |
2982 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX || |
2983 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_MES)) | |
557f42a2 AD |
2984 | continue; |
2985 | ||
2a7798ea AD |
2986 | /* SDMA 5.x+ is part of GFX power domain so it's covered by GFXOFF */ |
2987 | if (adev->in_s0ix && | |
2988 | (adev->ip_versions[SDMA0_HWIP][0] >= IP_VERSION(5, 0, 0)) && | |
2989 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SDMA)) | |
2990 | continue; | |
2991 | ||
e11c7750 TH |
2992 | /* Once swPSP provides the IMU, RLC FW binaries to TOS during cold-boot. |
2993 | * These are in TMR, hence are expected to be reused by PSP-TOS to reload | |
2994 | * from this location and RLC Autoload automatically also gets loaded | |
2995 | * from here based on PMFW -> PSP message during re-init sequence. | |
2996 | * Therefore, the psp suspend & resume should be skipped to avoid destroy | |
2997 | * the TMR and reload FWs again for IMU enabled APU ASICs. | |
2998 | */ | |
2999 | if (amdgpu_in_reset(adev) && | |
3000 | (adev->flags & AMD_IS_APU) && adev->gfx.imu.funcs && | |
3001 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) | |
3002 | continue; | |
3003 | ||
d38ceaf9 | 3004 | /* XXX handle errors */ |
a1255107 | 3005 | r = adev->ip_blocks[i].version->funcs->suspend(adev); |
d38ceaf9 | 3006 | /* XXX handle errors */ |
2c1a2784 | 3007 | if (r) { |
a1255107 AD |
3008 | DRM_ERROR("suspend of IP block <%s> failed %d\n", |
3009 | adev->ip_blocks[i].version->funcs->name, r); | |
2c1a2784 | 3010 | } |
876923fb | 3011 | adev->ip_blocks[i].status.hw = false; |
a3a09142 | 3012 | /* handle putting the SMC in the appropriate state */ |
47fc644f | 3013 | if (!amdgpu_sriov_vf(adev)) { |
86b93fd6 JZ |
3014 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) { |
3015 | r = amdgpu_dpm_set_mp1_state(adev, adev->mp1_state); | |
3016 | if (r) { | |
3017 | DRM_ERROR("SMC failed to set mp1 state %d, %d\n", | |
3018 | adev->mp1_state, r); | |
3019 | return r; | |
3020 | } | |
a3a09142 AD |
3021 | } |
3022 | } | |
d38ceaf9 AD |
3023 | } |
3024 | ||
3025 | return 0; | |
3026 | } | |
3027 | ||
e7854a03 AD |
3028 | /** |
3029 | * amdgpu_device_ip_suspend - run suspend for hardware IPs | |
3030 | * | |
3031 | * @adev: amdgpu_device pointer | |
3032 | * | |
3033 | * Main suspend function for hardware IPs. The list of all the hardware | |
3034 | * IPs that make up the asic is walked, clockgating is disabled and the | |
3035 | * suspend callbacks are run. suspend puts the hardware and software state | |
3036 | * in each IP into a state suitable for suspend. | |
3037 | * Returns 0 on success, negative error code on failure. | |
3038 | */ | |
3039 | int amdgpu_device_ip_suspend(struct amdgpu_device *adev) | |
3040 | { | |
3041 | int r; | |
3042 | ||
3c73683c JC |
3043 | if (amdgpu_sriov_vf(adev)) { |
3044 | amdgpu_virt_fini_data_exchange(adev); | |
e7819644 | 3045 | amdgpu_virt_request_full_gpu(adev, false); |
3c73683c | 3046 | } |
e7819644 | 3047 | |
e7854a03 AD |
3048 | r = amdgpu_device_ip_suspend_phase1(adev); |
3049 | if (r) | |
3050 | return r; | |
3051 | r = amdgpu_device_ip_suspend_phase2(adev); | |
3052 | ||
e7819644 YT |
3053 | if (amdgpu_sriov_vf(adev)) |
3054 | amdgpu_virt_release_full_gpu(adev, false); | |
3055 | ||
e7854a03 AD |
3056 | return r; |
3057 | } | |
3058 | ||
06ec9070 | 3059 | static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev) |
a90ad3c2 ML |
3060 | { |
3061 | int i, r; | |
3062 | ||
2cb681b6 | 3063 | static enum amd_ip_block_type ip_order[] = { |
2cb681b6 | 3064 | AMD_IP_BLOCK_TYPE_COMMON, |
c1c39032 | 3065 | AMD_IP_BLOCK_TYPE_GMC, |
39186aef | 3066 | AMD_IP_BLOCK_TYPE_PSP, |
2cb681b6 ML |
3067 | AMD_IP_BLOCK_TYPE_IH, |
3068 | }; | |
a90ad3c2 | 3069 | |
95ea3dbc | 3070 | for (i = 0; i < adev->num_ip_blocks; i++) { |
2cb681b6 ML |
3071 | int j; |
3072 | struct amdgpu_ip_block *block; | |
a90ad3c2 | 3073 | |
4cd2a96d J |
3074 | block = &adev->ip_blocks[i]; |
3075 | block->status.hw = false; | |
2cb681b6 | 3076 | |
4cd2a96d | 3077 | for (j = 0; j < ARRAY_SIZE(ip_order); j++) { |
2cb681b6 | 3078 | |
4cd2a96d | 3079 | if (block->version->type != ip_order[j] || |
2cb681b6 ML |
3080 | !block->status.valid) |
3081 | continue; | |
3082 | ||
3083 | r = block->version->funcs->hw_init(adev); | |
0aaeefcc | 3084 | DRM_INFO("RE-INIT-early: %s %s\n", block->version->funcs->name, r?"failed":"succeeded"); |
c41d1cf6 ML |
3085 | if (r) |
3086 | return r; | |
482f0e53 | 3087 | block->status.hw = true; |
a90ad3c2 ML |
3088 | } |
3089 | } | |
3090 | ||
3091 | return 0; | |
3092 | } | |
3093 | ||
06ec9070 | 3094 | static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev) |
a90ad3c2 ML |
3095 | { |
3096 | int i, r; | |
3097 | ||
2cb681b6 ML |
3098 | static enum amd_ip_block_type ip_order[] = { |
3099 | AMD_IP_BLOCK_TYPE_SMC, | |
3100 | AMD_IP_BLOCK_TYPE_DCE, | |
3101 | AMD_IP_BLOCK_TYPE_GFX, | |
3102 | AMD_IP_BLOCK_TYPE_SDMA, | |
ec64350d | 3103 | AMD_IP_BLOCK_TYPE_MES, |
257deb8c | 3104 | AMD_IP_BLOCK_TYPE_UVD, |
d83c7a07 | 3105 | AMD_IP_BLOCK_TYPE_VCE, |
d2cdc014 YZ |
3106 | AMD_IP_BLOCK_TYPE_VCN, |
3107 | AMD_IP_BLOCK_TYPE_JPEG | |
2cb681b6 | 3108 | }; |
a90ad3c2 | 3109 | |
2cb681b6 ML |
3110 | for (i = 0; i < ARRAY_SIZE(ip_order); i++) { |
3111 | int j; | |
3112 | struct amdgpu_ip_block *block; | |
a90ad3c2 | 3113 | |
2cb681b6 ML |
3114 | for (j = 0; j < adev->num_ip_blocks; j++) { |
3115 | block = &adev->ip_blocks[j]; | |
3116 | ||
3117 | if (block->version->type != ip_order[i] || | |
482f0e53 ML |
3118 | !block->status.valid || |
3119 | block->status.hw) | |
2cb681b6 ML |
3120 | continue; |
3121 | ||
895bd048 JZ |
3122 | if (block->version->type == AMD_IP_BLOCK_TYPE_SMC) |
3123 | r = block->version->funcs->resume(adev); | |
3124 | else | |
3125 | r = block->version->funcs->hw_init(adev); | |
3126 | ||
0aaeefcc | 3127 | DRM_INFO("RE-INIT-late: %s %s\n", block->version->funcs->name, r?"failed":"succeeded"); |
c41d1cf6 ML |
3128 | if (r) |
3129 | return r; | |
482f0e53 | 3130 | block->status.hw = true; |
a90ad3c2 ML |
3131 | } |
3132 | } | |
3133 | ||
3134 | return 0; | |
3135 | } | |
3136 | ||
e3ecdffa AD |
3137 | /** |
3138 | * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs | |
3139 | * | |
3140 | * @adev: amdgpu_device pointer | |
3141 | * | |
3142 | * First resume function for hardware IPs. The list of all the hardware | |
3143 | * IPs that make up the asic is walked and the resume callbacks are run for | |
3144 | * COMMON, GMC, and IH. resume puts the hardware into a functional state | |
3145 | * after a suspend and updates the software state as necessary. This | |
3146 | * function is also used for restoring the GPU after a GPU reset. | |
3147 | * Returns 0 on success, negative error code on failure. | |
3148 | */ | |
06ec9070 | 3149 | static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev) |
d38ceaf9 AD |
3150 | { |
3151 | int i, r; | |
3152 | ||
a90ad3c2 | 3153 | for (i = 0; i < adev->num_ip_blocks; i++) { |
482f0e53 | 3154 | if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw) |
a90ad3c2 | 3155 | continue; |
a90ad3c2 | 3156 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || |
e3ecdffa | 3157 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC || |
d7274ec7 BZ |
3158 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH || |
3159 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP && amdgpu_sriov_vf(adev))) { | |
482f0e53 | 3160 | |
fcf0649f CZ |
3161 | r = adev->ip_blocks[i].version->funcs->resume(adev); |
3162 | if (r) { | |
3163 | DRM_ERROR("resume of IP block <%s> failed %d\n", | |
3164 | adev->ip_blocks[i].version->funcs->name, r); | |
3165 | return r; | |
3166 | } | |
482f0e53 | 3167 | adev->ip_blocks[i].status.hw = true; |
a90ad3c2 ML |
3168 | } |
3169 | } | |
3170 | ||
3171 | return 0; | |
3172 | } | |
3173 | ||
e3ecdffa AD |
3174 | /** |
3175 | * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs | |
3176 | * | |
3177 | * @adev: amdgpu_device pointer | |
3178 | * | |
3179 | * First resume function for hardware IPs. The list of all the hardware | |
3180 | * IPs that make up the asic is walked and the resume callbacks are run for | |
3181 | * all blocks except COMMON, GMC, and IH. resume puts the hardware into a | |
3182 | * functional state after a suspend and updates the software state as | |
3183 | * necessary. This function is also used for restoring the GPU after a GPU | |
3184 | * reset. | |
3185 | * Returns 0 on success, negative error code on failure. | |
3186 | */ | |
06ec9070 | 3187 | static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev) |
d38ceaf9 AD |
3188 | { |
3189 | int i, r; | |
3190 | ||
3191 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
482f0e53 | 3192 | if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw) |
d38ceaf9 | 3193 | continue; |
fcf0649f | 3194 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || |
e3ecdffa | 3195 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC || |
7a3e0bb2 RZ |
3196 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH || |
3197 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) | |
fcf0649f | 3198 | continue; |
a1255107 | 3199 | r = adev->ip_blocks[i].version->funcs->resume(adev); |
2c1a2784 | 3200 | if (r) { |
a1255107 AD |
3201 | DRM_ERROR("resume of IP block <%s> failed %d\n", |
3202 | adev->ip_blocks[i].version->funcs->name, r); | |
d38ceaf9 | 3203 | return r; |
2c1a2784 | 3204 | } |
482f0e53 | 3205 | adev->ip_blocks[i].status.hw = true; |
d38ceaf9 AD |
3206 | } |
3207 | ||
3208 | return 0; | |
3209 | } | |
3210 | ||
e3ecdffa AD |
3211 | /** |
3212 | * amdgpu_device_ip_resume - run resume for hardware IPs | |
3213 | * | |
3214 | * @adev: amdgpu_device pointer | |
3215 | * | |
3216 | * Main resume function for hardware IPs. The hardware IPs | |
3217 | * are split into two resume functions because they are | |
b8920e1e | 3218 | * also used in recovering from a GPU reset and some additional |
e3ecdffa AD |
3219 | * steps need to be take between them. In this case (S3/S4) they are |
3220 | * run sequentially. | |
3221 | * Returns 0 on success, negative error code on failure. | |
3222 | */ | |
06ec9070 | 3223 | static int amdgpu_device_ip_resume(struct amdgpu_device *adev) |
fcf0649f CZ |
3224 | { |
3225 | int r; | |
3226 | ||
06ec9070 | 3227 | r = amdgpu_device_ip_resume_phase1(adev); |
fcf0649f CZ |
3228 | if (r) |
3229 | return r; | |
7a3e0bb2 RZ |
3230 | |
3231 | r = amdgpu_device_fw_loading(adev); | |
3232 | if (r) | |
3233 | return r; | |
3234 | ||
06ec9070 | 3235 | r = amdgpu_device_ip_resume_phase2(adev); |
fcf0649f CZ |
3236 | |
3237 | return r; | |
3238 | } | |
3239 | ||
e3ecdffa AD |
3240 | /** |
3241 | * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV | |
3242 | * | |
3243 | * @adev: amdgpu_device pointer | |
3244 | * | |
3245 | * Query the VBIOS data tables to determine if the board supports SR-IOV. | |
3246 | */ | |
4e99a44e | 3247 | static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev) |
048765ad | 3248 | { |
6867e1b5 ML |
3249 | if (amdgpu_sriov_vf(adev)) { |
3250 | if (adev->is_atom_fw) { | |
58ff791a | 3251 | if (amdgpu_atomfirmware_gpu_virtualization_supported(adev)) |
6867e1b5 ML |
3252 | adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS; |
3253 | } else { | |
3254 | if (amdgpu_atombios_has_gpu_virtualization_table(adev)) | |
3255 | adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS; | |
3256 | } | |
3257 | ||
3258 | if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS)) | |
3259 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0); | |
a5bde2f9 | 3260 | } |
048765ad AR |
3261 | } |
3262 | ||
e3ecdffa AD |
3263 | /** |
3264 | * amdgpu_device_asic_has_dc_support - determine if DC supports the asic | |
3265 | * | |
3266 | * @asic_type: AMD asic type | |
3267 | * | |
3268 | * Check if there is DC (new modesetting infrastructre) support for an asic. | |
3269 | * returns true if DC has support, false if not. | |
3270 | */ | |
4562236b HW |
3271 | bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type) |
3272 | { | |
3273 | switch (asic_type) { | |
0637d417 AD |
3274 | #ifdef CONFIG_DRM_AMDGPU_SI |
3275 | case CHIP_HAINAN: | |
3276 | #endif | |
3277 | case CHIP_TOPAZ: | |
3278 | /* chips with no display hardware */ | |
3279 | return false; | |
4562236b | 3280 | #if defined(CONFIG_DRM_AMD_DC) |
64200c46 MR |
3281 | case CHIP_TAHITI: |
3282 | case CHIP_PITCAIRN: | |
3283 | case CHIP_VERDE: | |
3284 | case CHIP_OLAND: | |
2d32ffd6 AD |
3285 | /* |
3286 | * We have systems in the wild with these ASICs that require | |
3287 | * LVDS and VGA support which is not supported with DC. | |
3288 | * | |
3289 | * Fallback to the non-DC driver here by default so as not to | |
3290 | * cause regressions. | |
3291 | */ | |
3292 | #if defined(CONFIG_DRM_AMD_DC_SI) | |
3293 | return amdgpu_dc > 0; | |
3294 | #else | |
3295 | return false; | |
64200c46 | 3296 | #endif |
4562236b | 3297 | case CHIP_BONAIRE: |
0d6fbccb | 3298 | case CHIP_KAVERI: |
367e6687 AD |
3299 | case CHIP_KABINI: |
3300 | case CHIP_MULLINS: | |
d9fda248 HW |
3301 | /* |
3302 | * We have systems in the wild with these ASICs that require | |
b5a0168e | 3303 | * VGA support which is not supported with DC. |
d9fda248 HW |
3304 | * |
3305 | * Fallback to the non-DC driver here by default so as not to | |
3306 | * cause regressions. | |
3307 | */ | |
3308 | return amdgpu_dc > 0; | |
f7f12b25 | 3309 | default: |
fd187853 | 3310 | return amdgpu_dc != 0; |
f7f12b25 | 3311 | #else |
4562236b | 3312 | default: |
93b09a9a | 3313 | if (amdgpu_dc > 0) |
b8920e1e | 3314 | DRM_INFO_ONCE("Display Core has been requested via kernel parameter but isn't supported by ASIC, ignoring\n"); |
4562236b | 3315 | return false; |
f7f12b25 | 3316 | #endif |
4562236b HW |
3317 | } |
3318 | } | |
3319 | ||
3320 | /** | |
3321 | * amdgpu_device_has_dc_support - check if dc is supported | |
3322 | * | |
982a820b | 3323 | * @adev: amdgpu_device pointer |
4562236b HW |
3324 | * |
3325 | * Returns true for supported, false for not supported | |
3326 | */ | |
3327 | bool amdgpu_device_has_dc_support(struct amdgpu_device *adev) | |
3328 | { | |
25263da3 | 3329 | if (adev->enable_virtual_display || |
abaf210c | 3330 | (adev->harvest_ip_mask & AMD_HARVEST_IP_DMU_MASK)) |
2555039d XY |
3331 | return false; |
3332 | ||
4562236b HW |
3333 | return amdgpu_device_asic_has_dc_support(adev->asic_type); |
3334 | } | |
3335 | ||
d4535e2c AG |
3336 | static void amdgpu_device_xgmi_reset_func(struct work_struct *__work) |
3337 | { | |
3338 | struct amdgpu_device *adev = | |
3339 | container_of(__work, struct amdgpu_device, xgmi_reset_work); | |
d95e8e97 | 3340 | struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev); |
d4535e2c | 3341 | |
c6a6e2db AG |
3342 | /* It's a bug to not have a hive within this function */ |
3343 | if (WARN_ON(!hive)) | |
3344 | return; | |
3345 | ||
3346 | /* | |
3347 | * Use task barrier to synchronize all xgmi reset works across the | |
3348 | * hive. task_barrier_enter and task_barrier_exit will block | |
3349 | * until all the threads running the xgmi reset works reach | |
3350 | * those points. task_barrier_full will do both blocks. | |
3351 | */ | |
3352 | if (amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) { | |
3353 | ||
3354 | task_barrier_enter(&hive->tb); | |
4a580877 | 3355 | adev->asic_reset_res = amdgpu_device_baco_enter(adev_to_drm(adev)); |
c6a6e2db AG |
3356 | |
3357 | if (adev->asic_reset_res) | |
3358 | goto fail; | |
3359 | ||
3360 | task_barrier_exit(&hive->tb); | |
4a580877 | 3361 | adev->asic_reset_res = amdgpu_device_baco_exit(adev_to_drm(adev)); |
c6a6e2db AG |
3362 | |
3363 | if (adev->asic_reset_res) | |
3364 | goto fail; | |
43c4d576 | 3365 | |
5e67bba3 | 3366 | if (adev->mmhub.ras && adev->mmhub.ras->ras_block.hw_ops && |
3367 | adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count) | |
3368 | adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count(adev); | |
c6a6e2db AG |
3369 | } else { |
3370 | ||
3371 | task_barrier_full(&hive->tb); | |
3372 | adev->asic_reset_res = amdgpu_asic_reset(adev); | |
3373 | } | |
ce316fa5 | 3374 | |
c6a6e2db | 3375 | fail: |
d4535e2c | 3376 | if (adev->asic_reset_res) |
fed184e9 | 3377 | DRM_WARN("ASIC reset failed with error, %d for drm dev, %s", |
4a580877 | 3378 | adev->asic_reset_res, adev_to_drm(adev)->unique); |
d95e8e97 | 3379 | amdgpu_put_xgmi_hive(hive); |
d4535e2c AG |
3380 | } |
3381 | ||
71f98027 AD |
3382 | static int amdgpu_device_get_job_timeout_settings(struct amdgpu_device *adev) |
3383 | { | |
3384 | char *input = amdgpu_lockup_timeout; | |
3385 | char *timeout_setting = NULL; | |
3386 | int index = 0; | |
3387 | long timeout; | |
3388 | int ret = 0; | |
3389 | ||
3390 | /* | |
67387dfe AD |
3391 | * By default timeout for non compute jobs is 10000 |
3392 | * and 60000 for compute jobs. | |
71f98027 | 3393 | * In SR-IOV or passthrough mode, timeout for compute |
b7b2a316 | 3394 | * jobs are 60000 by default. |
71f98027 AD |
3395 | */ |
3396 | adev->gfx_timeout = msecs_to_jiffies(10000); | |
3397 | adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout; | |
9882e278 ED |
3398 | if (amdgpu_sriov_vf(adev)) |
3399 | adev->compute_timeout = amdgpu_sriov_is_pp_one_vf(adev) ? | |
3400 | msecs_to_jiffies(60000) : msecs_to_jiffies(10000); | |
71f98027 | 3401 | else |
67387dfe | 3402 | adev->compute_timeout = msecs_to_jiffies(60000); |
71f98027 | 3403 | |
f440ff44 | 3404 | if (strnlen(input, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) { |
71f98027 | 3405 | while ((timeout_setting = strsep(&input, ",")) && |
f440ff44 | 3406 | strnlen(timeout_setting, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) { |
71f98027 AD |
3407 | ret = kstrtol(timeout_setting, 0, &timeout); |
3408 | if (ret) | |
3409 | return ret; | |
3410 | ||
3411 | if (timeout == 0) { | |
3412 | index++; | |
3413 | continue; | |
3414 | } else if (timeout < 0) { | |
3415 | timeout = MAX_SCHEDULE_TIMEOUT; | |
127aedf9 CK |
3416 | dev_warn(adev->dev, "lockup timeout disabled"); |
3417 | add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK); | |
71f98027 AD |
3418 | } else { |
3419 | timeout = msecs_to_jiffies(timeout); | |
3420 | } | |
3421 | ||
3422 | switch (index++) { | |
3423 | case 0: | |
3424 | adev->gfx_timeout = timeout; | |
3425 | break; | |
3426 | case 1: | |
3427 | adev->compute_timeout = timeout; | |
3428 | break; | |
3429 | case 2: | |
3430 | adev->sdma_timeout = timeout; | |
3431 | break; | |
3432 | case 3: | |
3433 | adev->video_timeout = timeout; | |
3434 | break; | |
3435 | default: | |
3436 | break; | |
3437 | } | |
3438 | } | |
3439 | /* | |
3440 | * There is only one value specified and | |
3441 | * it should apply to all non-compute jobs. | |
3442 | */ | |
bcccee89 | 3443 | if (index == 1) { |
71f98027 | 3444 | adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout; |
bcccee89 ED |
3445 | if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev)) |
3446 | adev->compute_timeout = adev->gfx_timeout; | |
3447 | } | |
71f98027 AD |
3448 | } |
3449 | ||
3450 | return ret; | |
3451 | } | |
d4535e2c | 3452 | |
4a74c38c PY |
3453 | /** |
3454 | * amdgpu_device_check_iommu_direct_map - check if RAM direct mapped to GPU | |
3455 | * | |
3456 | * @adev: amdgpu_device pointer | |
3457 | * | |
3458 | * RAM direct mapped to GPU if IOMMU is not enabled or is pass through mode | |
3459 | */ | |
3460 | static void amdgpu_device_check_iommu_direct_map(struct amdgpu_device *adev) | |
3461 | { | |
3462 | struct iommu_domain *domain; | |
3463 | ||
3464 | domain = iommu_get_domain_for_dev(adev->dev); | |
3465 | if (!domain || domain->type == IOMMU_DOMAIN_IDENTITY) | |
3466 | adev->ram_is_direct_mapped = true; | |
3467 | } | |
3468 | ||
77f3a5cd | 3469 | static const struct attribute *amdgpu_dev_attributes[] = { |
77f3a5cd ND |
3470 | &dev_attr_pcie_replay_count.attr, |
3471 | NULL | |
3472 | }; | |
3473 | ||
02ff519e AD |
3474 | static void amdgpu_device_set_mcbp(struct amdgpu_device *adev) |
3475 | { | |
3476 | if (amdgpu_mcbp == 1) | |
3477 | adev->gfx.mcbp = true; | |
1e9e15dc JZ |
3478 | else if (amdgpu_mcbp == 0) |
3479 | adev->gfx.mcbp = false; | |
3480 | else if ((adev->ip_versions[GC_HWIP][0] >= IP_VERSION(9, 0, 0)) && | |
3481 | (adev->ip_versions[GC_HWIP][0] < IP_VERSION(10, 0, 0)) && | |
3482 | adev->gfx.num_gfx_rings) | |
50a7c876 AD |
3483 | adev->gfx.mcbp = true; |
3484 | ||
02ff519e AD |
3485 | if (amdgpu_sriov_vf(adev)) |
3486 | adev->gfx.mcbp = true; | |
3487 | ||
3488 | if (adev->gfx.mcbp) | |
3489 | DRM_INFO("MCBP is enabled\n"); | |
3490 | } | |
3491 | ||
d38ceaf9 AD |
3492 | /** |
3493 | * amdgpu_device_init - initialize the driver | |
3494 | * | |
3495 | * @adev: amdgpu_device pointer | |
d38ceaf9 AD |
3496 | * @flags: driver flags |
3497 | * | |
3498 | * Initializes the driver info and hw (all asics). | |
3499 | * Returns 0 for success or an error on failure. | |
3500 | * Called at driver startup. | |
3501 | */ | |
3502 | int amdgpu_device_init(struct amdgpu_device *adev, | |
d38ceaf9 AD |
3503 | uint32_t flags) |
3504 | { | |
8aba21b7 LT |
3505 | struct drm_device *ddev = adev_to_drm(adev); |
3506 | struct pci_dev *pdev = adev->pdev; | |
d38ceaf9 | 3507 | int r, i; |
b98c6299 | 3508 | bool px = false; |
95844d20 | 3509 | u32 max_MBps; |
59e9fff1 | 3510 | int tmp; |
d38ceaf9 AD |
3511 | |
3512 | adev->shutdown = false; | |
d38ceaf9 | 3513 | adev->flags = flags; |
4e66d7d2 YZ |
3514 | |
3515 | if (amdgpu_force_asic_type >= 0 && amdgpu_force_asic_type < CHIP_LAST) | |
3516 | adev->asic_type = amdgpu_force_asic_type; | |
3517 | else | |
3518 | adev->asic_type = flags & AMD_ASIC_MASK; | |
3519 | ||
d38ceaf9 | 3520 | adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT; |
593aa2d2 | 3521 | if (amdgpu_emu_mode == 1) |
8bdab6bb | 3522 | adev->usec_timeout *= 10; |
770d13b1 | 3523 | adev->gmc.gart_size = 512 * 1024 * 1024; |
d38ceaf9 AD |
3524 | adev->accel_working = false; |
3525 | adev->num_rings = 0; | |
68ce8b24 | 3526 | RCU_INIT_POINTER(adev->gang_submit, dma_fence_get_stub()); |
d38ceaf9 AD |
3527 | adev->mman.buffer_funcs = NULL; |
3528 | adev->mman.buffer_funcs_ring = NULL; | |
3529 | adev->vm_manager.vm_pte_funcs = NULL; | |
0c88b430 | 3530 | adev->vm_manager.vm_pte_num_scheds = 0; |
132f34e4 | 3531 | adev->gmc.gmc_funcs = NULL; |
7bd939d0 | 3532 | adev->harvest_ip_mask = 0x0; |
f54d1867 | 3533 | adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS); |
b8866c26 | 3534 | bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES); |
d38ceaf9 AD |
3535 | |
3536 | adev->smc_rreg = &amdgpu_invalid_rreg; | |
3537 | adev->smc_wreg = &amdgpu_invalid_wreg; | |
3538 | adev->pcie_rreg = &amdgpu_invalid_rreg; | |
3539 | adev->pcie_wreg = &amdgpu_invalid_wreg; | |
0c552ed3 LM |
3540 | adev->pcie_rreg_ext = &amdgpu_invalid_rreg_ext; |
3541 | adev->pcie_wreg_ext = &amdgpu_invalid_wreg_ext; | |
36b9a952 HR |
3542 | adev->pciep_rreg = &amdgpu_invalid_rreg; |
3543 | adev->pciep_wreg = &amdgpu_invalid_wreg; | |
4fa1c6a6 TZ |
3544 | adev->pcie_rreg64 = &amdgpu_invalid_rreg64; |
3545 | adev->pcie_wreg64 = &amdgpu_invalid_wreg64; | |
d38ceaf9 AD |
3546 | adev->uvd_ctx_rreg = &amdgpu_invalid_rreg; |
3547 | adev->uvd_ctx_wreg = &amdgpu_invalid_wreg; | |
3548 | adev->didt_rreg = &amdgpu_invalid_rreg; | |
3549 | adev->didt_wreg = &amdgpu_invalid_wreg; | |
ccdbb20a RZ |
3550 | adev->gc_cac_rreg = &amdgpu_invalid_rreg; |
3551 | adev->gc_cac_wreg = &amdgpu_invalid_wreg; | |
d38ceaf9 AD |
3552 | adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg; |
3553 | adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg; | |
3554 | ||
3e39ab90 AD |
3555 | DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n", |
3556 | amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device, | |
3557 | pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision); | |
d38ceaf9 AD |
3558 | |
3559 | /* mutex initialization are all done here so we | |
b8920e1e SS |
3560 | * can recall function without having locking issues |
3561 | */ | |
0e5ca0d1 | 3562 | mutex_init(&adev->firmware.mutex); |
d38ceaf9 AD |
3563 | mutex_init(&adev->pm.mutex); |
3564 | mutex_init(&adev->gfx.gpu_clock_mutex); | |
3565 | mutex_init(&adev->srbm_mutex); | |
b8866c26 | 3566 | mutex_init(&adev->gfx.pipe_reserve_mutex); |
d23ee13f | 3567 | mutex_init(&adev->gfx.gfx_off_mutex); |
98a54e88 | 3568 | mutex_init(&adev->gfx.partition_mutex); |
d38ceaf9 | 3569 | mutex_init(&adev->grbm_idx_mutex); |
d38ceaf9 | 3570 | mutex_init(&adev->mn_lock); |
e23b74aa | 3571 | mutex_init(&adev->virt.vf_errors.lock); |
d38ceaf9 | 3572 | hash_init(adev->mn_hash); |
32eaeae0 | 3573 | mutex_init(&adev->psp.mutex); |
bd052211 | 3574 | mutex_init(&adev->notifier_lock); |
8cda7a4f | 3575 | mutex_init(&adev->pm.stable_pstate_ctx_lock); |
f113cc32 | 3576 | mutex_init(&adev->benchmark_mutex); |
d38ceaf9 | 3577 | |
ab3b9de6 | 3578 | amdgpu_device_init_apu_flags(adev); |
9f6a7857 | 3579 | |
912dfc84 EQ |
3580 | r = amdgpu_device_check_arguments(adev); |
3581 | if (r) | |
3582 | return r; | |
d38ceaf9 | 3583 | |
d38ceaf9 AD |
3584 | spin_lock_init(&adev->mmio_idx_lock); |
3585 | spin_lock_init(&adev->smc_idx_lock); | |
3586 | spin_lock_init(&adev->pcie_idx_lock); | |
3587 | spin_lock_init(&adev->uvd_ctx_idx_lock); | |
3588 | spin_lock_init(&adev->didt_idx_lock); | |
ccdbb20a | 3589 | spin_lock_init(&adev->gc_cac_idx_lock); |
16abb5d2 | 3590 | spin_lock_init(&adev->se_cac_idx_lock); |
d38ceaf9 | 3591 | spin_lock_init(&adev->audio_endpt_idx_lock); |
95844d20 | 3592 | spin_lock_init(&adev->mm_stats.lock); |
d38ceaf9 | 3593 | |
0c4e7fa5 CZ |
3594 | INIT_LIST_HEAD(&adev->shadow_list); |
3595 | mutex_init(&adev->shadow_list_lock); | |
3596 | ||
655ce9cb | 3597 | INIT_LIST_HEAD(&adev->reset_list); |
3598 | ||
6492e1b0 | 3599 | INIT_LIST_HEAD(&adev->ras_list); |
3600 | ||
beff74bc AD |
3601 | INIT_DELAYED_WORK(&adev->delayed_init_work, |
3602 | amdgpu_device_delayed_init_work_handler); | |
1e317b99 RZ |
3603 | INIT_DELAYED_WORK(&adev->gfx.gfx_off_delay_work, |
3604 | amdgpu_device_delay_enable_gfx_off); | |
2dc80b00 | 3605 | |
d4535e2c AG |
3606 | INIT_WORK(&adev->xgmi_reset_work, amdgpu_device_xgmi_reset_func); |
3607 | ||
d23ee13f | 3608 | adev->gfx.gfx_off_req_count = 1; |
0ad7347a AA |
3609 | adev->gfx.gfx_off_residency = 0; |
3610 | adev->gfx.gfx_off_entrycount = 0; | |
b6e79d9a | 3611 | adev->pm.ac_power = power_supply_is_system_supplied() > 0; |
b1ddf548 | 3612 | |
b265bdbd EQ |
3613 | atomic_set(&adev->throttling_logging_enabled, 1); |
3614 | /* | |
3615 | * If throttling continues, logging will be performed every minute | |
3616 | * to avoid log flooding. "-1" is subtracted since the thermal | |
3617 | * throttling interrupt comes every second. Thus, the total logging | |
3618 | * interval is 59 seconds(retelimited printk interval) + 1(waiting | |
3619 | * for throttling interrupt) = 60 seconds. | |
3620 | */ | |
3621 | ratelimit_state_init(&adev->throttling_logging_rs, (60 - 1) * HZ, 1); | |
3622 | ratelimit_set_flags(&adev->throttling_logging_rs, RATELIMIT_MSG_ON_RELEASE); | |
3623 | ||
0fa49558 AX |
3624 | /* Registers mapping */ |
3625 | /* TODO: block userspace mapping of io register */ | |
da69c161 KW |
3626 | if (adev->asic_type >= CHIP_BONAIRE) { |
3627 | adev->rmmio_base = pci_resource_start(adev->pdev, 5); | |
3628 | adev->rmmio_size = pci_resource_len(adev->pdev, 5); | |
3629 | } else { | |
3630 | adev->rmmio_base = pci_resource_start(adev->pdev, 2); | |
3631 | adev->rmmio_size = pci_resource_len(adev->pdev, 2); | |
3632 | } | |
d38ceaf9 | 3633 | |
6c08e0ef EQ |
3634 | for (i = 0; i < AMD_IP_BLOCK_TYPE_NUM; i++) |
3635 | atomic_set(&adev->pm.pwr_state[i], POWER_STATE_UNKNOWN); | |
3636 | ||
d38ceaf9 | 3637 | adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size); |
b8920e1e | 3638 | if (!adev->rmmio) |
d38ceaf9 | 3639 | return -ENOMEM; |
b8920e1e | 3640 | |
d38ceaf9 | 3641 | DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base); |
b8920e1e | 3642 | DRM_INFO("register mmio size: %u\n", (unsigned int)adev->rmmio_size); |
d38ceaf9 | 3643 | |
436afdfa PY |
3644 | /* |
3645 | * Reset domain needs to be present early, before XGMI hive discovered | |
3646 | * (if any) and intitialized to use reset sem and in_gpu reset flag | |
3647 | * early on during init and before calling to RREG32. | |
3648 | */ | |
3649 | adev->reset_domain = amdgpu_reset_create_reset_domain(SINGLE_DEVICE, "amdgpu-reset-dev"); | |
3650 | if (!adev->reset_domain) | |
3651 | return -ENOMEM; | |
3652 | ||
3aa0115d ML |
3653 | /* detect hw virtualization here */ |
3654 | amdgpu_detect_virtualization(adev); | |
3655 | ||
04e85958 TL |
3656 | amdgpu_device_get_pcie_info(adev); |
3657 | ||
dffa11b4 ML |
3658 | r = amdgpu_device_get_job_timeout_settings(adev); |
3659 | if (r) { | |
3660 | dev_err(adev->dev, "invalid lockup_timeout parameter syntax\n"); | |
4ef87d8f | 3661 | return r; |
a190d1c7 XY |
3662 | } |
3663 | ||
d38ceaf9 | 3664 | /* early init functions */ |
06ec9070 | 3665 | r = amdgpu_device_ip_early_init(adev); |
d38ceaf9 | 3666 | if (r) |
4ef87d8f | 3667 | return r; |
d38ceaf9 | 3668 | |
02ff519e AD |
3669 | amdgpu_device_set_mcbp(adev); |
3670 | ||
b7cdb41e ML |
3671 | /* Get rid of things like offb */ |
3672 | r = drm_aperture_remove_conflicting_pci_framebuffers(adev->pdev, &amdgpu_kms_driver); | |
3673 | if (r) | |
3674 | return r; | |
3675 | ||
4d33e704 SK |
3676 | /* Enable TMZ based on IP_VERSION */ |
3677 | amdgpu_gmc_tmz_set(adev); | |
3678 | ||
957b0787 | 3679 | amdgpu_gmc_noretry_set(adev); |
4a0165f0 VS |
3680 | /* Need to get xgmi info early to decide the reset behavior*/ |
3681 | if (adev->gmc.xgmi.supported) { | |
3682 | r = adev->gfxhub.funcs->get_xgmi_info(adev); | |
3683 | if (r) | |
3684 | return r; | |
3685 | } | |
3686 | ||
8e6d0b69 | 3687 | /* enable PCIE atomic ops */ |
b4520bfd GW |
3688 | if (amdgpu_sriov_vf(adev)) { |
3689 | if (adev->virt.fw_reserve.p_pf2vf) | |
3690 | adev->have_atomics_support = ((struct amd_sriov_msg_pf2vf_info *) | |
3691 | adev->virt.fw_reserve.p_pf2vf)->pcie_atomic_ops_support_flags == | |
3692 | (PCI_EXP_DEVCAP2_ATOMIC_COMP32 | PCI_EXP_DEVCAP2_ATOMIC_COMP64); | |
0e768043 YZ |
3693 | /* APUs w/ gfx9 onwards doesn't reply on PCIe atomics, rather it is a |
3694 | * internal path natively support atomics, set have_atomics_support to true. | |
3695 | */ | |
b4520bfd GW |
3696 | } else if ((adev->flags & AMD_IS_APU) && |
3697 | (adev->ip_versions[GC_HWIP][0] > IP_VERSION(9, 0, 0))) { | |
0e768043 | 3698 | adev->have_atomics_support = true; |
b4520bfd | 3699 | } else { |
8e6d0b69 | 3700 | adev->have_atomics_support = |
3701 | !pci_enable_atomic_ops_to_root(adev->pdev, | |
3702 | PCI_EXP_DEVCAP2_ATOMIC_COMP32 | | |
3703 | PCI_EXP_DEVCAP2_ATOMIC_COMP64); | |
b4520bfd GW |
3704 | } |
3705 | ||
8e6d0b69 | 3706 | if (!adev->have_atomics_support) |
3707 | dev_info(adev->dev, "PCIE atomic ops is not supported\n"); | |
3708 | ||
6585661d | 3709 | /* doorbell bar mapping and doorbell index init*/ |
43c064db | 3710 | amdgpu_doorbell_init(adev); |
6585661d | 3711 | |
9475a943 SL |
3712 | if (amdgpu_emu_mode == 1) { |
3713 | /* post the asic on emulation mode */ | |
3714 | emu_soc_asic_init(adev); | |
bfca0289 | 3715 | goto fence_driver_init; |
9475a943 | 3716 | } |
bfca0289 | 3717 | |
04442bf7 LL |
3718 | amdgpu_reset_init(adev); |
3719 | ||
4e99a44e | 3720 | /* detect if we are with an SRIOV vbios */ |
b4520bfd GW |
3721 | if (adev->bios) |
3722 | amdgpu_device_detect_sriov_bios(adev); | |
048765ad | 3723 | |
95e8e59e AD |
3724 | /* check if we need to reset the asic |
3725 | * E.g., driver was not cleanly unloaded previously, etc. | |
3726 | */ | |
f14899fd | 3727 | if (!amdgpu_sriov_vf(adev) && amdgpu_asic_need_reset_on_init(adev)) { |
e3c1b071 | 3728 | if (adev->gmc.xgmi.num_physical_nodes) { |
3729 | dev_info(adev->dev, "Pending hive reset.\n"); | |
3730 | adev->gmc.xgmi.pending_reset = true; | |
3731 | /* Only need to init necessary block for SMU to handle the reset */ | |
3732 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
3733 | if (!adev->ip_blocks[i].status.valid) | |
3734 | continue; | |
3735 | if (!(adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC || | |
3736 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || | |
3737 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH || | |
3738 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC)) { | |
751f43e7 | 3739 | DRM_DEBUG("IP %s disabled for hw_init.\n", |
e3c1b071 | 3740 | adev->ip_blocks[i].version->funcs->name); |
3741 | adev->ip_blocks[i].status.hw = true; | |
3742 | } | |
3743 | } | |
3744 | } else { | |
59e9fff1 | 3745 | tmp = amdgpu_reset_method; |
3746 | /* It should do a default reset when loading or reloading the driver, | |
3747 | * regardless of the module parameter reset_method. | |
3748 | */ | |
3749 | amdgpu_reset_method = AMD_RESET_METHOD_NONE; | |
e3c1b071 | 3750 | r = amdgpu_asic_reset(adev); |
59e9fff1 | 3751 | amdgpu_reset_method = tmp; |
e3c1b071 | 3752 | if (r) { |
3753 | dev_err(adev->dev, "asic reset on init failed\n"); | |
3754 | goto failed; | |
3755 | } | |
95e8e59e AD |
3756 | } |
3757 | } | |
3758 | ||
d38ceaf9 | 3759 | /* Post card if necessary */ |
39c640c0 | 3760 | if (amdgpu_device_need_post(adev)) { |
d38ceaf9 | 3761 | if (!adev->bios) { |
bec86378 | 3762 | dev_err(adev->dev, "no vBIOS found\n"); |
83ba126a AD |
3763 | r = -EINVAL; |
3764 | goto failed; | |
d38ceaf9 | 3765 | } |
bec86378 | 3766 | DRM_INFO("GPU posting now...\n"); |
4d2997ab | 3767 | r = amdgpu_device_asic_init(adev); |
4e99a44e ML |
3768 | if (r) { |
3769 | dev_err(adev->dev, "gpu post error!\n"); | |
3770 | goto failed; | |
3771 | } | |
d38ceaf9 AD |
3772 | } |
3773 | ||
9535a86a SZ |
3774 | if (adev->bios) { |
3775 | if (adev->is_atom_fw) { | |
3776 | /* Initialize clocks */ | |
3777 | r = amdgpu_atomfirmware_get_clock_info(adev); | |
3778 | if (r) { | |
3779 | dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n"); | |
3780 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0); | |
3781 | goto failed; | |
3782 | } | |
3783 | } else { | |
3784 | /* Initialize clocks */ | |
3785 | r = amdgpu_atombios_get_clock_info(adev); | |
3786 | if (r) { | |
3787 | dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n"); | |
3788 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0); | |
3789 | goto failed; | |
3790 | } | |
3791 | /* init i2c buses */ | |
3792 | if (!amdgpu_device_has_dc_support(adev)) | |
3793 | amdgpu_atombios_i2c_init(adev); | |
a5bde2f9 | 3794 | } |
2c1a2784 | 3795 | } |
d38ceaf9 | 3796 | |
bfca0289 | 3797 | fence_driver_init: |
d38ceaf9 | 3798 | /* Fence driver */ |
067f44c8 | 3799 | r = amdgpu_fence_driver_sw_init(adev); |
2c1a2784 | 3800 | if (r) { |
067f44c8 | 3801 | dev_err(adev->dev, "amdgpu_fence_driver_sw_init failed\n"); |
e23b74aa | 3802 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0); |
83ba126a | 3803 | goto failed; |
2c1a2784 | 3804 | } |
d38ceaf9 AD |
3805 | |
3806 | /* init the mode config */ | |
4a580877 | 3807 | drm_mode_config_init(adev_to_drm(adev)); |
d38ceaf9 | 3808 | |
06ec9070 | 3809 | r = amdgpu_device_ip_init(adev); |
d38ceaf9 | 3810 | if (r) { |
06ec9070 | 3811 | dev_err(adev->dev, "amdgpu_device_ip_init failed\n"); |
e23b74aa | 3812 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0); |
970fd197 | 3813 | goto release_ras_con; |
d38ceaf9 AD |
3814 | } |
3815 | ||
8d35a259 LG |
3816 | amdgpu_fence_driver_hw_init(adev); |
3817 | ||
d69b8971 YZ |
3818 | dev_info(adev->dev, |
3819 | "SE %d, SH per SE %d, CU per SH %d, active_cu_number %d\n", | |
d7f72fe4 YZ |
3820 | adev->gfx.config.max_shader_engines, |
3821 | adev->gfx.config.max_sh_per_se, | |
3822 | adev->gfx.config.max_cu_per_sh, | |
3823 | adev->gfx.cu_info.number); | |
3824 | ||
d38ceaf9 AD |
3825 | adev->accel_working = true; |
3826 | ||
e59c0205 AX |
3827 | amdgpu_vm_check_compute_bug(adev); |
3828 | ||
95844d20 MO |
3829 | /* Initialize the buffer migration limit. */ |
3830 | if (amdgpu_moverate >= 0) | |
3831 | max_MBps = amdgpu_moverate; | |
3832 | else | |
3833 | max_MBps = 8; /* Allow 8 MB/s. */ | |
3834 | /* Get a log2 for easy divisions. */ | |
3835 | adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps)); | |
3836 | ||
184d8384 LL |
3837 | r = amdgpu_atombios_sysfs_init(adev); |
3838 | if (r) | |
3839 | drm_err(&adev->ddev, | |
3840 | "registering atombios sysfs failed (%d).\n", r); | |
3841 | ||
d2f52ac8 | 3842 | r = amdgpu_pm_sysfs_init(adev); |
53e9d836 GC |
3843 | if (r) |
3844 | DRM_ERROR("registering pm sysfs failed (%d).\n", r); | |
d2f52ac8 | 3845 | |
5bb23532 | 3846 | r = amdgpu_ucode_sysfs_init(adev); |
7c868b59 YT |
3847 | if (r) { |
3848 | adev->ucode_sysfs_en = false; | |
5bb23532 | 3849 | DRM_ERROR("Creating firmware sysfs failed (%d).\n", r); |
7c868b59 YT |
3850 | } else |
3851 | adev->ucode_sysfs_en = true; | |
5bb23532 | 3852 | |
b0adca4d EQ |
3853 | /* |
3854 | * Register gpu instance before amdgpu_device_enable_mgpu_fan_boost. | |
3855 | * Otherwise the mgpu fan boost feature will be skipped due to the | |
3856 | * gpu instance is counted less. | |
3857 | */ | |
3858 | amdgpu_register_gpu_instance(adev); | |
3859 | ||
d38ceaf9 AD |
3860 | /* enable clockgating, etc. after ib tests, etc. since some blocks require |
3861 | * explicit gating rather than handling it automatically. | |
3862 | */ | |
e3c1b071 | 3863 | if (!adev->gmc.xgmi.pending_reset) { |
3864 | r = amdgpu_device_ip_late_init(adev); | |
3865 | if (r) { | |
3866 | dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n"); | |
3867 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r); | |
970fd197 | 3868 | goto release_ras_con; |
e3c1b071 | 3869 | } |
3870 | /* must succeed. */ | |
3871 | amdgpu_ras_resume(adev); | |
3872 | queue_delayed_work(system_wq, &adev->delayed_init_work, | |
3873 | msecs_to_jiffies(AMDGPU_RESUME_MS)); | |
2c1a2784 | 3874 | } |
d38ceaf9 | 3875 | |
38eecbe0 CL |
3876 | if (amdgpu_sriov_vf(adev)) { |
3877 | amdgpu_virt_release_full_gpu(adev, true); | |
2c738637 | 3878 | flush_delayed_work(&adev->delayed_init_work); |
38eecbe0 | 3879 | } |
2c738637 | 3880 | |
77f3a5cd | 3881 | r = sysfs_create_files(&adev->dev->kobj, amdgpu_dev_attributes); |
5aea5327 | 3882 | if (r) |
77f3a5cd | 3883 | dev_err(adev->dev, "Could not create amdgpu device attr\n"); |
bd607166 | 3884 | |
7957ec80 LL |
3885 | amdgpu_fru_sysfs_init(adev); |
3886 | ||
d155bef0 AB |
3887 | if (IS_ENABLED(CONFIG_PERF_EVENTS)) |
3888 | r = amdgpu_pmu_init(adev); | |
9c7c85f7 JK |
3889 | if (r) |
3890 | dev_err(adev->dev, "amdgpu_pmu_init failed\n"); | |
3891 | ||
c1dd4aa6 AG |
3892 | /* Have stored pci confspace at hand for restore in sudden PCI error */ |
3893 | if (amdgpu_device_cache_pci_state(adev->pdev)) | |
3894 | pci_restore_state(pdev); | |
3895 | ||
8c3dd61c KHF |
3896 | /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */ |
3897 | /* this will fail for cards that aren't VGA class devices, just | |
b8920e1e SS |
3898 | * ignore it |
3899 | */ | |
8c3dd61c | 3900 | if ((adev->pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA) |
bf44e8ce | 3901 | vga_client_register(adev->pdev, amdgpu_device_vga_set_decode); |
8c3dd61c | 3902 | |
d37a3929 OC |
3903 | px = amdgpu_device_supports_px(ddev); |
3904 | ||
3905 | if (px || (!pci_is_thunderbolt_attached(adev->pdev) && | |
3906 | apple_gmux_detect(NULL, NULL))) | |
8c3dd61c KHF |
3907 | vga_switcheroo_register_client(adev->pdev, |
3908 | &amdgpu_switcheroo_ops, px); | |
d37a3929 OC |
3909 | |
3910 | if (px) | |
8c3dd61c | 3911 | vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain); |
8c3dd61c | 3912 | |
e3c1b071 | 3913 | if (adev->gmc.xgmi.pending_reset) |
3914 | queue_delayed_work(system_wq, &mgpu_info.delayed_reset_work, | |
3915 | msecs_to_jiffies(AMDGPU_RESUME_MS)); | |
3916 | ||
4a74c38c PY |
3917 | amdgpu_device_check_iommu_direct_map(adev); |
3918 | ||
d38ceaf9 | 3919 | return 0; |
83ba126a | 3920 | |
970fd197 | 3921 | release_ras_con: |
38eecbe0 CL |
3922 | if (amdgpu_sriov_vf(adev)) |
3923 | amdgpu_virt_release_full_gpu(adev, true); | |
3924 | ||
3925 | /* failed in exclusive mode due to timeout */ | |
3926 | if (amdgpu_sriov_vf(adev) && | |
3927 | !amdgpu_sriov_runtime(adev) && | |
3928 | amdgpu_virt_mmio_blocked(adev) && | |
3929 | !amdgpu_virt_wait_reset(adev)) { | |
3930 | dev_err(adev->dev, "VF exclusive mode timeout\n"); | |
3931 | /* Don't send request since VF is inactive. */ | |
3932 | adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME; | |
3933 | adev->virt.ops = NULL; | |
3934 | r = -EAGAIN; | |
3935 | } | |
970fd197 SY |
3936 | amdgpu_release_ras_context(adev); |
3937 | ||
83ba126a | 3938 | failed: |
89041940 | 3939 | amdgpu_vf_error_trans_all(adev); |
8840a387 | 3940 | |
83ba126a | 3941 | return r; |
d38ceaf9 AD |
3942 | } |
3943 | ||
07775fc1 AG |
3944 | static void amdgpu_device_unmap_mmio(struct amdgpu_device *adev) |
3945 | { | |
62d5f9f7 | 3946 | |
07775fc1 AG |
3947 | /* Clear all CPU mappings pointing to this device */ |
3948 | unmap_mapping_range(adev->ddev.anon_inode->i_mapping, 0, 0, 1); | |
3949 | ||
3950 | /* Unmap all mapped bars - Doorbell, registers and VRAM */ | |
43c064db | 3951 | amdgpu_doorbell_fini(adev); |
07775fc1 AG |
3952 | |
3953 | iounmap(adev->rmmio); | |
3954 | adev->rmmio = NULL; | |
3955 | if (adev->mman.aper_base_kaddr) | |
3956 | iounmap(adev->mman.aper_base_kaddr); | |
3957 | adev->mman.aper_base_kaddr = NULL; | |
3958 | ||
3959 | /* Memory manager related */ | |
a0ba1279 | 3960 | if (!adev->gmc.xgmi.connected_to_cpu && !adev->gmc.is_app_apu) { |
07775fc1 AG |
3961 | arch_phys_wc_del(adev->gmc.vram_mtrr); |
3962 | arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size); | |
3963 | } | |
3964 | } | |
3965 | ||
d38ceaf9 | 3966 | /** |
bbe04dec | 3967 | * amdgpu_device_fini_hw - tear down the driver |
d38ceaf9 AD |
3968 | * |
3969 | * @adev: amdgpu_device pointer | |
3970 | * | |
3971 | * Tear down the driver info (all asics). | |
3972 | * Called at driver shutdown. | |
3973 | */ | |
72c8c97b | 3974 | void amdgpu_device_fini_hw(struct amdgpu_device *adev) |
d38ceaf9 | 3975 | { |
aac89168 | 3976 | dev_info(adev->dev, "amdgpu: finishing device.\n"); |
9f875167 | 3977 | flush_delayed_work(&adev->delayed_init_work); |
d0d13fe8 | 3978 | adev->shutdown = true; |
9f875167 | 3979 | |
752c683d ML |
3980 | /* make sure IB test finished before entering exclusive mode |
3981 | * to avoid preemption on IB test | |
b8920e1e | 3982 | */ |
519b8b76 | 3983 | if (amdgpu_sriov_vf(adev)) { |
752c683d | 3984 | amdgpu_virt_request_full_gpu(adev, false); |
519b8b76 BZ |
3985 | amdgpu_virt_fini_data_exchange(adev); |
3986 | } | |
752c683d | 3987 | |
e5b03032 ML |
3988 | /* disable all interrupts */ |
3989 | amdgpu_irq_disable_all(adev); | |
47fc644f | 3990 | if (adev->mode_info.mode_config_initialized) { |
1053b9c9 | 3991 | if (!drm_drv_uses_atomic_modeset(adev_to_drm(adev))) |
4a580877 | 3992 | drm_helper_force_disable_all(adev_to_drm(adev)); |
ff97cba8 | 3993 | else |
4a580877 | 3994 | drm_atomic_helper_shutdown(adev_to_drm(adev)); |
ff97cba8 | 3995 | } |
8d35a259 | 3996 | amdgpu_fence_driver_hw_fini(adev); |
72c8c97b | 3997 | |
cd3a8a59 | 3998 | if (adev->mman.initialized) |
9bff18d1 | 3999 | drain_workqueue(adev->mman.bdev.wq); |
98f56188 | 4000 | |
53e9d836 | 4001 | if (adev->pm.sysfs_initialized) |
7c868b59 | 4002 | amdgpu_pm_sysfs_fini(adev); |
72c8c97b AG |
4003 | if (adev->ucode_sysfs_en) |
4004 | amdgpu_ucode_sysfs_fini(adev); | |
4005 | sysfs_remove_files(&adev->dev->kobj, amdgpu_dev_attributes); | |
7957ec80 | 4006 | amdgpu_fru_sysfs_fini(adev); |
72c8c97b | 4007 | |
232d1d43 SY |
4008 | /* disable ras feature must before hw fini */ |
4009 | amdgpu_ras_pre_fini(adev); | |
4010 | ||
e9669fb7 | 4011 | amdgpu_device_ip_fini_early(adev); |
d10d0daa | 4012 | |
a3848df6 YW |
4013 | amdgpu_irq_fini_hw(adev); |
4014 | ||
b6fd6e0f SK |
4015 | if (adev->mman.initialized) |
4016 | ttm_device_clear_dma_mappings(&adev->mman.bdev); | |
894c6890 | 4017 | |
d10d0daa | 4018 | amdgpu_gart_dummy_page_fini(adev); |
07775fc1 | 4019 | |
39934d3e VP |
4020 | if (drm_dev_is_unplugged(adev_to_drm(adev))) |
4021 | amdgpu_device_unmap_mmio(adev); | |
87172e89 | 4022 | |
72c8c97b AG |
4023 | } |
4024 | ||
4025 | void amdgpu_device_fini_sw(struct amdgpu_device *adev) | |
4026 | { | |
62d5f9f7 | 4027 | int idx; |
d37a3929 | 4028 | bool px; |
62d5f9f7 | 4029 | |
8d35a259 | 4030 | amdgpu_fence_driver_sw_fini(adev); |
a5c5d8d5 | 4031 | amdgpu_device_ip_fini(adev); |
b31d3063 | 4032 | amdgpu_ucode_release(&adev->firmware.gpu_info_fw); |
d38ceaf9 | 4033 | adev->accel_working = false; |
68ce8b24 | 4034 | dma_fence_put(rcu_dereference_protected(adev->gang_submit, true)); |
04442bf7 LL |
4035 | |
4036 | amdgpu_reset_fini(adev); | |
4037 | ||
d38ceaf9 | 4038 | /* free i2c buses */ |
4562236b HW |
4039 | if (!amdgpu_device_has_dc_support(adev)) |
4040 | amdgpu_i2c_fini(adev); | |
bfca0289 SL |
4041 | |
4042 | if (amdgpu_emu_mode != 1) | |
4043 | amdgpu_atombios_fini(adev); | |
4044 | ||
d38ceaf9 AD |
4045 | kfree(adev->bios); |
4046 | adev->bios = NULL; | |
d37a3929 OC |
4047 | |
4048 | px = amdgpu_device_supports_px(adev_to_drm(adev)); | |
4049 | ||
4050 | if (px || (!pci_is_thunderbolt_attached(adev->pdev) && | |
4051 | apple_gmux_detect(NULL, NULL))) | |
84c8b22e | 4052 | vga_switcheroo_unregister_client(adev->pdev); |
d37a3929 OC |
4053 | |
4054 | if (px) | |
83ba126a | 4055 | vga_switcheroo_fini_domain_pm_ops(adev->dev); |
d37a3929 | 4056 | |
38d6be81 | 4057 | if ((adev->pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA) |
b8779475 | 4058 | vga_client_unregister(adev->pdev); |
e9bc1bf7 | 4059 | |
62d5f9f7 LS |
4060 | if (drm_dev_enter(adev_to_drm(adev), &idx)) { |
4061 | ||
4062 | iounmap(adev->rmmio); | |
4063 | adev->rmmio = NULL; | |
43c064db | 4064 | amdgpu_doorbell_fini(adev); |
62d5f9f7 LS |
4065 | drm_dev_exit(idx); |
4066 | } | |
4067 | ||
d155bef0 AB |
4068 | if (IS_ENABLED(CONFIG_PERF_EVENTS)) |
4069 | amdgpu_pmu_fini(adev); | |
72de33f8 | 4070 | if (adev->mman.discovery_bin) |
a190d1c7 | 4071 | amdgpu_discovery_fini(adev); |
72c8c97b | 4072 | |
cfbb6b00 AG |
4073 | amdgpu_reset_put_reset_domain(adev->reset_domain); |
4074 | adev->reset_domain = NULL; | |
4075 | ||
72c8c97b AG |
4076 | kfree(adev->pci_state); |
4077 | ||
d38ceaf9 AD |
4078 | } |
4079 | ||
58144d28 ND |
4080 | /** |
4081 | * amdgpu_device_evict_resources - evict device resources | |
4082 | * @adev: amdgpu device object | |
4083 | * | |
4084 | * Evicts all ttm device resources(vram BOs, gart table) from the lru list | |
4085 | * of the vram memory type. Mainly used for evicting device resources | |
4086 | * at suspend time. | |
4087 | * | |
4088 | */ | |
7863c155 | 4089 | static int amdgpu_device_evict_resources(struct amdgpu_device *adev) |
58144d28 | 4090 | { |
7863c155 ML |
4091 | int ret; |
4092 | ||
e53d9665 ML |
4093 | /* No need to evict vram on APUs for suspend to ram or s2idle */ |
4094 | if ((adev->in_s3 || adev->in_s0ix) && (adev->flags & AMD_IS_APU)) | |
7863c155 | 4095 | return 0; |
58144d28 | 4096 | |
7863c155 ML |
4097 | ret = amdgpu_ttm_evict_resources(adev, TTM_PL_VRAM); |
4098 | if (ret) | |
58144d28 | 4099 | DRM_WARN("evicting device resources failed\n"); |
7863c155 | 4100 | return ret; |
58144d28 | 4101 | } |
d38ceaf9 AD |
4102 | |
4103 | /* | |
4104 | * Suspend & resume. | |
4105 | */ | |
4106 | /** | |
810ddc3a | 4107 | * amdgpu_device_suspend - initiate device suspend |
d38ceaf9 | 4108 | * |
87e3f136 | 4109 | * @dev: drm dev pointer |
87e3f136 | 4110 | * @fbcon : notify the fbdev of suspend |
d38ceaf9 AD |
4111 | * |
4112 | * Puts the hw in the suspend state (all asics). | |
4113 | * Returns 0 for success or an error on failure. | |
4114 | * Called at driver suspend. | |
4115 | */ | |
de185019 | 4116 | int amdgpu_device_suspend(struct drm_device *dev, bool fbcon) |
d38ceaf9 | 4117 | { |
a2e15b0e | 4118 | struct amdgpu_device *adev = drm_to_adev(dev); |
d7274ec7 | 4119 | int r = 0; |
d38ceaf9 | 4120 | |
d38ceaf9 AD |
4121 | if (dev->switch_power_state == DRM_SWITCH_POWER_OFF) |
4122 | return 0; | |
4123 | ||
44779b43 | 4124 | adev->in_suspend = true; |
3fa8f89d | 4125 | |
47ea2076 SF |
4126 | /* Evict the majority of BOs before grabbing the full access */ |
4127 | r = amdgpu_device_evict_resources(adev); | |
4128 | if (r) | |
4129 | return r; | |
4130 | ||
d7274ec7 BZ |
4131 | if (amdgpu_sriov_vf(adev)) { |
4132 | amdgpu_virt_fini_data_exchange(adev); | |
4133 | r = amdgpu_virt_request_full_gpu(adev, false); | |
4134 | if (r) | |
4135 | return r; | |
4136 | } | |
4137 | ||
3fa8f89d S |
4138 | if (amdgpu_acpi_smart_shift_update(dev, AMDGPU_SS_DEV_D3)) |
4139 | DRM_WARN("smart shift update failed\n"); | |
4140 | ||
5f818173 | 4141 | if (fbcon) |
087451f3 | 4142 | drm_fb_helper_set_suspend_unlocked(adev_to_drm(adev)->fb_helper, true); |
5f818173 | 4143 | |
beff74bc | 4144 | cancel_delayed_work_sync(&adev->delayed_init_work); |
0dee7263 | 4145 | flush_delayed_work(&adev->gfx.gfx_off_delay_work); |
a5459475 | 4146 | |
5e6932fe | 4147 | amdgpu_ras_suspend(adev); |
4148 | ||
2196927b | 4149 | amdgpu_device_ip_suspend_phase1(adev); |
fe1053b7 | 4150 | |
c004d44e | 4151 | if (!adev->in_s0ix) |
5d3a2d95 | 4152 | amdgpu_amdkfd_suspend(adev, adev->in_runpm); |
94fa5660 | 4153 | |
7863c155 ML |
4154 | r = amdgpu_device_evict_resources(adev); |
4155 | if (r) | |
4156 | return r; | |
d38ceaf9 | 4157 | |
8d35a259 | 4158 | amdgpu_fence_driver_hw_fini(adev); |
d38ceaf9 | 4159 | |
2196927b | 4160 | amdgpu_device_ip_suspend_phase2(adev); |
d38ceaf9 | 4161 | |
d7274ec7 BZ |
4162 | if (amdgpu_sriov_vf(adev)) |
4163 | amdgpu_virt_release_full_gpu(adev, false); | |
4164 | ||
d38ceaf9 AD |
4165 | return 0; |
4166 | } | |
4167 | ||
4168 | /** | |
810ddc3a | 4169 | * amdgpu_device_resume - initiate device resume |
d38ceaf9 | 4170 | * |
87e3f136 | 4171 | * @dev: drm dev pointer |
87e3f136 | 4172 | * @fbcon : notify the fbdev of resume |
d38ceaf9 AD |
4173 | * |
4174 | * Bring the hw back to operating state (all asics). | |
4175 | * Returns 0 for success or an error on failure. | |
4176 | * Called at driver resume. | |
4177 | */ | |
de185019 | 4178 | int amdgpu_device_resume(struct drm_device *dev, bool fbcon) |
d38ceaf9 | 4179 | { |
1348969a | 4180 | struct amdgpu_device *adev = drm_to_adev(dev); |
03161a6e | 4181 | int r = 0; |
d38ceaf9 | 4182 | |
d7274ec7 BZ |
4183 | if (amdgpu_sriov_vf(adev)) { |
4184 | r = amdgpu_virt_request_full_gpu(adev, true); | |
4185 | if (r) | |
4186 | return r; | |
4187 | } | |
4188 | ||
d38ceaf9 AD |
4189 | if (dev->switch_power_state == DRM_SWITCH_POWER_OFF) |
4190 | return 0; | |
4191 | ||
62498733 | 4192 | if (adev->in_s0ix) |
bc143d8b | 4193 | amdgpu_dpm_gfx_state_change(adev, sGpuChangeState_D0Entry); |
628c36d7 | 4194 | |
d38ceaf9 | 4195 | /* post card */ |
39c640c0 | 4196 | if (amdgpu_device_need_post(adev)) { |
4d2997ab | 4197 | r = amdgpu_device_asic_init(adev); |
74b0b157 | 4198 | if (r) |
aac89168 | 4199 | dev_err(adev->dev, "amdgpu asic init failed\n"); |
74b0b157 | 4200 | } |
d38ceaf9 | 4201 | |
06ec9070 | 4202 | r = amdgpu_device_ip_resume(adev); |
d7274ec7 | 4203 | |
e6707218 | 4204 | if (r) { |
aac89168 | 4205 | dev_err(adev->dev, "amdgpu_device_ip_resume failed (%d).\n", r); |
3c22c1ea | 4206 | goto exit; |
e6707218 | 4207 | } |
8d35a259 | 4208 | amdgpu_fence_driver_hw_init(adev); |
5ceb54c6 | 4209 | |
06ec9070 | 4210 | r = amdgpu_device_ip_late_init(adev); |
03161a6e | 4211 | if (r) |
3c22c1ea | 4212 | goto exit; |
d38ceaf9 | 4213 | |
beff74bc AD |
4214 | queue_delayed_work(system_wq, &adev->delayed_init_work, |
4215 | msecs_to_jiffies(AMDGPU_RESUME_MS)); | |
4216 | ||
c004d44e | 4217 | if (!adev->in_s0ix) { |
5d3a2d95 AD |
4218 | r = amdgpu_amdkfd_resume(adev, adev->in_runpm); |
4219 | if (r) | |
3c22c1ea | 4220 | goto exit; |
5d3a2d95 | 4221 | } |
756e6880 | 4222 | |
3c22c1ea SF |
4223 | exit: |
4224 | if (amdgpu_sriov_vf(adev)) { | |
4225 | amdgpu_virt_init_data_exchange(adev); | |
4226 | amdgpu_virt_release_full_gpu(adev, true); | |
4227 | } | |
4228 | ||
4229 | if (r) | |
4230 | return r; | |
4231 | ||
96a5d8d4 | 4232 | /* Make sure IB tests flushed */ |
beff74bc | 4233 | flush_delayed_work(&adev->delayed_init_work); |
96a5d8d4 | 4234 | |
a2e15b0e | 4235 | if (fbcon) |
087451f3 | 4236 | drm_fb_helper_set_suspend_unlocked(adev_to_drm(adev)->fb_helper, false); |
d38ceaf9 | 4237 | |
5e6932fe | 4238 | amdgpu_ras_resume(adev); |
4239 | ||
d09ef243 AD |
4240 | if (adev->mode_info.num_crtc) { |
4241 | /* | |
4242 | * Most of the connector probing functions try to acquire runtime pm | |
4243 | * refs to ensure that the GPU is powered on when connector polling is | |
4244 | * performed. Since we're calling this from a runtime PM callback, | |
4245 | * trying to acquire rpm refs will cause us to deadlock. | |
4246 | * | |
4247 | * Since we're guaranteed to be holding the rpm lock, it's safe to | |
4248 | * temporarily disable the rpm helpers so this doesn't deadlock us. | |
4249 | */ | |
23a1a9e5 | 4250 | #ifdef CONFIG_PM |
d09ef243 | 4251 | dev->dev->power.disable_depth++; |
23a1a9e5 | 4252 | #endif |
d09ef243 AD |
4253 | if (!adev->dc_enabled) |
4254 | drm_helper_hpd_irq_event(dev); | |
4255 | else | |
4256 | drm_kms_helper_hotplug_event(dev); | |
23a1a9e5 | 4257 | #ifdef CONFIG_PM |
d09ef243 | 4258 | dev->dev->power.disable_depth--; |
23a1a9e5 | 4259 | #endif |
d09ef243 | 4260 | } |
44779b43 RZ |
4261 | adev->in_suspend = false; |
4262 | ||
dc907c9d JX |
4263 | if (adev->enable_mes) |
4264 | amdgpu_mes_self_test(adev); | |
4265 | ||
3fa8f89d S |
4266 | if (amdgpu_acpi_smart_shift_update(dev, AMDGPU_SS_DEV_D0)) |
4267 | DRM_WARN("smart shift update failed\n"); | |
4268 | ||
4d3b9ae5 | 4269 | return 0; |
d38ceaf9 AD |
4270 | } |
4271 | ||
e3ecdffa AD |
4272 | /** |
4273 | * amdgpu_device_ip_check_soft_reset - did soft reset succeed | |
4274 | * | |
4275 | * @adev: amdgpu_device pointer | |
4276 | * | |
4277 | * The list of all the hardware IPs that make up the asic is walked and | |
4278 | * the check_soft_reset callbacks are run. check_soft_reset determines | |
4279 | * if the asic is still hung or not. | |
4280 | * Returns true if any of the IPs are still in a hung state, false if not. | |
4281 | */ | |
06ec9070 | 4282 | static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev) |
63fbf42f CZ |
4283 | { |
4284 | int i; | |
4285 | bool asic_hang = false; | |
4286 | ||
f993d628 ML |
4287 | if (amdgpu_sriov_vf(adev)) |
4288 | return true; | |
4289 | ||
8bc04c29 AD |
4290 | if (amdgpu_asic_need_full_reset(adev)) |
4291 | return true; | |
4292 | ||
63fbf42f | 4293 | for (i = 0; i < adev->num_ip_blocks; i++) { |
a1255107 | 4294 | if (!adev->ip_blocks[i].status.valid) |
63fbf42f | 4295 | continue; |
a1255107 AD |
4296 | if (adev->ip_blocks[i].version->funcs->check_soft_reset) |
4297 | adev->ip_blocks[i].status.hang = | |
4298 | adev->ip_blocks[i].version->funcs->check_soft_reset(adev); | |
4299 | if (adev->ip_blocks[i].status.hang) { | |
aac89168 | 4300 | dev_info(adev->dev, "IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name); |
63fbf42f CZ |
4301 | asic_hang = true; |
4302 | } | |
4303 | } | |
4304 | return asic_hang; | |
4305 | } | |
4306 | ||
e3ecdffa AD |
4307 | /** |
4308 | * amdgpu_device_ip_pre_soft_reset - prepare for soft reset | |
4309 | * | |
4310 | * @adev: amdgpu_device pointer | |
4311 | * | |
4312 | * The list of all the hardware IPs that make up the asic is walked and the | |
4313 | * pre_soft_reset callbacks are run if the block is hung. pre_soft_reset | |
4314 | * handles any IP specific hardware or software state changes that are | |
4315 | * necessary for a soft reset to succeed. | |
4316 | * Returns 0 on success, negative error code on failure. | |
4317 | */ | |
06ec9070 | 4318 | static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev) |
d31a501e CZ |
4319 | { |
4320 | int i, r = 0; | |
4321 | ||
4322 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 4323 | if (!adev->ip_blocks[i].status.valid) |
d31a501e | 4324 | continue; |
a1255107 AD |
4325 | if (adev->ip_blocks[i].status.hang && |
4326 | adev->ip_blocks[i].version->funcs->pre_soft_reset) { | |
4327 | r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev); | |
d31a501e CZ |
4328 | if (r) |
4329 | return r; | |
4330 | } | |
4331 | } | |
4332 | ||
4333 | return 0; | |
4334 | } | |
4335 | ||
e3ecdffa AD |
4336 | /** |
4337 | * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed | |
4338 | * | |
4339 | * @adev: amdgpu_device pointer | |
4340 | * | |
4341 | * Some hardware IPs cannot be soft reset. If they are hung, a full gpu | |
4342 | * reset is necessary to recover. | |
4343 | * Returns true if a full asic reset is required, false if not. | |
4344 | */ | |
06ec9070 | 4345 | static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev) |
35d782fe | 4346 | { |
da146d3b AD |
4347 | int i; |
4348 | ||
8bc04c29 AD |
4349 | if (amdgpu_asic_need_full_reset(adev)) |
4350 | return true; | |
4351 | ||
da146d3b | 4352 | for (i = 0; i < adev->num_ip_blocks; i++) { |
a1255107 | 4353 | if (!adev->ip_blocks[i].status.valid) |
da146d3b | 4354 | continue; |
a1255107 AD |
4355 | if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) || |
4356 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) || | |
4357 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) || | |
98512bb8 KW |
4358 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) || |
4359 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) { | |
a1255107 | 4360 | if (adev->ip_blocks[i].status.hang) { |
aac89168 | 4361 | dev_info(adev->dev, "Some block need full reset!\n"); |
da146d3b AD |
4362 | return true; |
4363 | } | |
4364 | } | |
35d782fe CZ |
4365 | } |
4366 | return false; | |
4367 | } | |
4368 | ||
e3ecdffa AD |
4369 | /** |
4370 | * amdgpu_device_ip_soft_reset - do a soft reset | |
4371 | * | |
4372 | * @adev: amdgpu_device pointer | |
4373 | * | |
4374 | * The list of all the hardware IPs that make up the asic is walked and the | |
4375 | * soft_reset callbacks are run if the block is hung. soft_reset handles any | |
4376 | * IP specific hardware or software state changes that are necessary to soft | |
4377 | * reset the IP. | |
4378 | * Returns 0 on success, negative error code on failure. | |
4379 | */ | |
06ec9070 | 4380 | static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev) |
35d782fe CZ |
4381 | { |
4382 | int i, r = 0; | |
4383 | ||
4384 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 4385 | if (!adev->ip_blocks[i].status.valid) |
35d782fe | 4386 | continue; |
a1255107 AD |
4387 | if (adev->ip_blocks[i].status.hang && |
4388 | adev->ip_blocks[i].version->funcs->soft_reset) { | |
4389 | r = adev->ip_blocks[i].version->funcs->soft_reset(adev); | |
35d782fe CZ |
4390 | if (r) |
4391 | return r; | |
4392 | } | |
4393 | } | |
4394 | ||
4395 | return 0; | |
4396 | } | |
4397 | ||
e3ecdffa AD |
4398 | /** |
4399 | * amdgpu_device_ip_post_soft_reset - clean up from soft reset | |
4400 | * | |
4401 | * @adev: amdgpu_device pointer | |
4402 | * | |
4403 | * The list of all the hardware IPs that make up the asic is walked and the | |
4404 | * post_soft_reset callbacks are run if the asic was hung. post_soft_reset | |
4405 | * handles any IP specific hardware or software state changes that are | |
4406 | * necessary after the IP has been soft reset. | |
4407 | * Returns 0 on success, negative error code on failure. | |
4408 | */ | |
06ec9070 | 4409 | static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev) |
35d782fe CZ |
4410 | { |
4411 | int i, r = 0; | |
4412 | ||
4413 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 4414 | if (!adev->ip_blocks[i].status.valid) |
35d782fe | 4415 | continue; |
a1255107 AD |
4416 | if (adev->ip_blocks[i].status.hang && |
4417 | adev->ip_blocks[i].version->funcs->post_soft_reset) | |
4418 | r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev); | |
35d782fe CZ |
4419 | if (r) |
4420 | return r; | |
4421 | } | |
4422 | ||
4423 | return 0; | |
4424 | } | |
4425 | ||
e3ecdffa | 4426 | /** |
c33adbc7 | 4427 | * amdgpu_device_recover_vram - Recover some VRAM contents |
e3ecdffa AD |
4428 | * |
4429 | * @adev: amdgpu_device pointer | |
4430 | * | |
4431 | * Restores the contents of VRAM buffers from the shadows in GTT. Used to | |
4432 | * restore things like GPUVM page tables after a GPU reset where | |
4433 | * the contents of VRAM might be lost. | |
403009bf CK |
4434 | * |
4435 | * Returns: | |
4436 | * 0 on success, negative error code on failure. | |
e3ecdffa | 4437 | */ |
c33adbc7 | 4438 | static int amdgpu_device_recover_vram(struct amdgpu_device *adev) |
c41d1cf6 | 4439 | { |
c41d1cf6 | 4440 | struct dma_fence *fence = NULL, *next = NULL; |
403009bf | 4441 | struct amdgpu_bo *shadow; |
e18aaea7 | 4442 | struct amdgpu_bo_vm *vmbo; |
403009bf | 4443 | long r = 1, tmo; |
c41d1cf6 ML |
4444 | |
4445 | if (amdgpu_sriov_runtime(adev)) | |
b045d3af | 4446 | tmo = msecs_to_jiffies(8000); |
c41d1cf6 ML |
4447 | else |
4448 | tmo = msecs_to_jiffies(100); | |
4449 | ||
aac89168 | 4450 | dev_info(adev->dev, "recover vram bo from shadow start\n"); |
c41d1cf6 | 4451 | mutex_lock(&adev->shadow_list_lock); |
e18aaea7 | 4452 | list_for_each_entry(vmbo, &adev->shadow_list, shadow_list) { |
4994d1f0 LC |
4453 | /* If vm is compute context or adev is APU, shadow will be NULL */ |
4454 | if (!vmbo->shadow) | |
4455 | continue; | |
4456 | shadow = vmbo->shadow; | |
4457 | ||
403009bf | 4458 | /* No need to recover an evicted BO */ |
d3116756 CK |
4459 | if (shadow->tbo.resource->mem_type != TTM_PL_TT || |
4460 | shadow->tbo.resource->start == AMDGPU_BO_INVALID_OFFSET || | |
4461 | shadow->parent->tbo.resource->mem_type != TTM_PL_VRAM) | |
403009bf CK |
4462 | continue; |
4463 | ||
4464 | r = amdgpu_bo_restore_shadow(shadow, &next); | |
4465 | if (r) | |
4466 | break; | |
4467 | ||
c41d1cf6 | 4468 | if (fence) { |
1712fb1a | 4469 | tmo = dma_fence_wait_timeout(fence, false, tmo); |
403009bf CK |
4470 | dma_fence_put(fence); |
4471 | fence = next; | |
1712fb1a | 4472 | if (tmo == 0) { |
4473 | r = -ETIMEDOUT; | |
c41d1cf6 | 4474 | break; |
1712fb1a | 4475 | } else if (tmo < 0) { |
4476 | r = tmo; | |
4477 | break; | |
4478 | } | |
403009bf CK |
4479 | } else { |
4480 | fence = next; | |
c41d1cf6 | 4481 | } |
c41d1cf6 ML |
4482 | } |
4483 | mutex_unlock(&adev->shadow_list_lock); | |
4484 | ||
403009bf CK |
4485 | if (fence) |
4486 | tmo = dma_fence_wait_timeout(fence, false, tmo); | |
c41d1cf6 ML |
4487 | dma_fence_put(fence); |
4488 | ||
1712fb1a | 4489 | if (r < 0 || tmo <= 0) { |
aac89168 | 4490 | dev_err(adev->dev, "recover vram bo from shadow failed, r is %ld, tmo is %ld\n", r, tmo); |
403009bf CK |
4491 | return -EIO; |
4492 | } | |
c41d1cf6 | 4493 | |
aac89168 | 4494 | dev_info(adev->dev, "recover vram bo from shadow done\n"); |
403009bf | 4495 | return 0; |
c41d1cf6 ML |
4496 | } |
4497 | ||
a90ad3c2 | 4498 | |
e3ecdffa | 4499 | /** |
06ec9070 | 4500 | * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf |
5740682e | 4501 | * |
982a820b | 4502 | * @adev: amdgpu_device pointer |
87e3f136 | 4503 | * @from_hypervisor: request from hypervisor |
5740682e ML |
4504 | * |
4505 | * do VF FLR and reinitialize Asic | |
3f48c681 | 4506 | * return 0 means succeeded otherwise failed |
e3ecdffa AD |
4507 | */ |
4508 | static int amdgpu_device_reset_sriov(struct amdgpu_device *adev, | |
4509 | bool from_hypervisor) | |
5740682e ML |
4510 | { |
4511 | int r; | |
a5f67c93 | 4512 | struct amdgpu_hive_info *hive = NULL; |
7258fa31 | 4513 | int retry_limit = 0; |
5740682e | 4514 | |
7258fa31 | 4515 | retry: |
c004d44e | 4516 | amdgpu_amdkfd_pre_reset(adev); |
428890a3 | 4517 | |
5740682e ML |
4518 | if (from_hypervisor) |
4519 | r = amdgpu_virt_request_full_gpu(adev, true); | |
4520 | else | |
4521 | r = amdgpu_virt_reset_gpu(adev); | |
4522 | if (r) | |
4523 | return r; | |
f734b213 | 4524 | amdgpu_irq_gpu_reset_resume_helper(adev); |
a90ad3c2 | 4525 | |
83f24a8f HC |
4526 | /* some sw clean up VF needs to do before recover */ |
4527 | amdgpu_virt_post_reset(adev); | |
4528 | ||
a90ad3c2 | 4529 | /* Resume IP prior to SMC */ |
06ec9070 | 4530 | r = amdgpu_device_ip_reinit_early_sriov(adev); |
5740682e ML |
4531 | if (r) |
4532 | goto error; | |
a90ad3c2 | 4533 | |
c9ffa427 | 4534 | amdgpu_virt_init_data_exchange(adev); |
a90ad3c2 | 4535 | |
7a3e0bb2 RZ |
4536 | r = amdgpu_device_fw_loading(adev); |
4537 | if (r) | |
4538 | return r; | |
4539 | ||
a90ad3c2 | 4540 | /* now we are okay to resume SMC/CP/SDMA */ |
06ec9070 | 4541 | r = amdgpu_device_ip_reinit_late_sriov(adev); |
5740682e ML |
4542 | if (r) |
4543 | goto error; | |
a90ad3c2 | 4544 | |
a5f67c93 ZL |
4545 | hive = amdgpu_get_xgmi_hive(adev); |
4546 | /* Update PSP FW topology after reset */ | |
4547 | if (hive && adev->gmc.xgmi.num_physical_nodes > 1) | |
4548 | r = amdgpu_xgmi_update_topology(hive, adev); | |
4549 | ||
4550 | if (hive) | |
4551 | amdgpu_put_xgmi_hive(hive); | |
4552 | ||
4553 | if (!r) { | |
a5f67c93 | 4554 | r = amdgpu_ib_ring_tests(adev); |
9c12f5cd | 4555 | |
c004d44e | 4556 | amdgpu_amdkfd_post_reset(adev); |
a5f67c93 | 4557 | } |
a90ad3c2 | 4558 | |
abc34253 | 4559 | error: |
c41d1cf6 | 4560 | if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) { |
e3526257 | 4561 | amdgpu_inc_vram_lost(adev); |
c33adbc7 | 4562 | r = amdgpu_device_recover_vram(adev); |
a90ad3c2 | 4563 | } |
437f3e0b | 4564 | amdgpu_virt_release_full_gpu(adev, true); |
a90ad3c2 | 4565 | |
7258fa31 SK |
4566 | if (AMDGPU_RETRY_SRIOV_RESET(r)) { |
4567 | if (retry_limit < AMDGPU_MAX_RETRY_LIMIT) { | |
4568 | retry_limit++; | |
4569 | goto retry; | |
4570 | } else | |
4571 | DRM_ERROR("GPU reset retry is beyond the retry limit\n"); | |
4572 | } | |
4573 | ||
a90ad3c2 ML |
4574 | return r; |
4575 | } | |
4576 | ||
9a1cddd6 | 4577 | /** |
4578 | * amdgpu_device_has_job_running - check if there is any job in mirror list | |
4579 | * | |
982a820b | 4580 | * @adev: amdgpu_device pointer |
9a1cddd6 | 4581 | * |
4582 | * check if there is any job in mirror list | |
4583 | */ | |
4584 | bool amdgpu_device_has_job_running(struct amdgpu_device *adev) | |
4585 | { | |
4586 | int i; | |
4587 | struct drm_sched_job *job; | |
4588 | ||
4589 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { | |
4590 | struct amdgpu_ring *ring = adev->rings[i]; | |
4591 | ||
4592 | if (!ring || !ring->sched.thread) | |
4593 | continue; | |
4594 | ||
4595 | spin_lock(&ring->sched.job_list_lock); | |
6efa4b46 LT |
4596 | job = list_first_entry_or_null(&ring->sched.pending_list, |
4597 | struct drm_sched_job, list); | |
9a1cddd6 | 4598 | spin_unlock(&ring->sched.job_list_lock); |
4599 | if (job) | |
4600 | return true; | |
4601 | } | |
4602 | return false; | |
4603 | } | |
4604 | ||
12938fad CK |
4605 | /** |
4606 | * amdgpu_device_should_recover_gpu - check if we should try GPU recovery | |
4607 | * | |
982a820b | 4608 | * @adev: amdgpu_device pointer |
12938fad CK |
4609 | * |
4610 | * Check amdgpu_gpu_recovery and SRIOV status to see if we should try to recover | |
4611 | * a hung GPU. | |
4612 | */ | |
4613 | bool amdgpu_device_should_recover_gpu(struct amdgpu_device *adev) | |
4614 | { | |
12938fad | 4615 | |
3ba7b418 AG |
4616 | if (amdgpu_gpu_recovery == 0) |
4617 | goto disabled; | |
4618 | ||
1a11a65d YC |
4619 | /* Skip soft reset check in fatal error mode */ |
4620 | if (!amdgpu_ras_is_poison_mode_supported(adev)) | |
4621 | return true; | |
4622 | ||
3ba7b418 AG |
4623 | if (amdgpu_sriov_vf(adev)) |
4624 | return true; | |
4625 | ||
4626 | if (amdgpu_gpu_recovery == -1) { | |
4627 | switch (adev->asic_type) { | |
b3523c45 AD |
4628 | #ifdef CONFIG_DRM_AMDGPU_SI |
4629 | case CHIP_VERDE: | |
4630 | case CHIP_TAHITI: | |
4631 | case CHIP_PITCAIRN: | |
4632 | case CHIP_OLAND: | |
4633 | case CHIP_HAINAN: | |
4634 | #endif | |
4635 | #ifdef CONFIG_DRM_AMDGPU_CIK | |
4636 | case CHIP_KAVERI: | |
4637 | case CHIP_KABINI: | |
4638 | case CHIP_MULLINS: | |
4639 | #endif | |
4640 | case CHIP_CARRIZO: | |
4641 | case CHIP_STONEY: | |
4642 | case CHIP_CYAN_SKILLFISH: | |
3ba7b418 | 4643 | goto disabled; |
b3523c45 AD |
4644 | default: |
4645 | break; | |
3ba7b418 | 4646 | } |
12938fad CK |
4647 | } |
4648 | ||
4649 | return true; | |
3ba7b418 AG |
4650 | |
4651 | disabled: | |
aac89168 | 4652 | dev_info(adev->dev, "GPU recovery disabled.\n"); |
3ba7b418 | 4653 | return false; |
12938fad CK |
4654 | } |
4655 | ||
5c03e584 FX |
4656 | int amdgpu_device_mode1_reset(struct amdgpu_device *adev) |
4657 | { | |
47fc644f SS |
4658 | u32 i; |
4659 | int ret = 0; | |
5c03e584 | 4660 | |
47fc644f | 4661 | amdgpu_atombios_scratch_regs_engine_hung(adev, true); |
5c03e584 | 4662 | |
47fc644f | 4663 | dev_info(adev->dev, "GPU mode1 reset\n"); |
5c03e584 | 4664 | |
47fc644f SS |
4665 | /* disable BM */ |
4666 | pci_clear_master(adev->pdev); | |
5c03e584 | 4667 | |
47fc644f | 4668 | amdgpu_device_cache_pci_state(adev->pdev); |
5c03e584 | 4669 | |
47fc644f SS |
4670 | if (amdgpu_dpm_is_mode1_reset_supported(adev)) { |
4671 | dev_info(adev->dev, "GPU smu mode1 reset\n"); | |
4672 | ret = amdgpu_dpm_mode1_reset(adev); | |
4673 | } else { | |
4674 | dev_info(adev->dev, "GPU psp mode1 reset\n"); | |
4675 | ret = psp_gpu_reset(adev); | |
4676 | } | |
5c03e584 | 4677 | |
47fc644f | 4678 | if (ret) |
7d442437 | 4679 | goto mode1_reset_failed; |
5c03e584 | 4680 | |
47fc644f | 4681 | amdgpu_device_load_pci_state(adev->pdev); |
7656168a LL |
4682 | ret = amdgpu_psp_wait_for_bootloader(adev); |
4683 | if (ret) | |
7d442437 | 4684 | goto mode1_reset_failed; |
5c03e584 | 4685 | |
47fc644f SS |
4686 | /* wait for asic to come out of reset */ |
4687 | for (i = 0; i < adev->usec_timeout; i++) { | |
4688 | u32 memsize = adev->nbio.funcs->get_memsize(adev); | |
5c03e584 | 4689 | |
47fc644f SS |
4690 | if (memsize != 0xffffffff) |
4691 | break; | |
4692 | udelay(1); | |
4693 | } | |
5c03e584 | 4694 | |
7d442437 HZ |
4695 | if (i >= adev->usec_timeout) { |
4696 | ret = -ETIMEDOUT; | |
4697 | goto mode1_reset_failed; | |
4698 | } | |
4699 | ||
47fc644f | 4700 | amdgpu_atombios_scratch_regs_engine_hung(adev, false); |
7656168a | 4701 | |
7d442437 HZ |
4702 | return 0; |
4703 | ||
4704 | mode1_reset_failed: | |
4705 | dev_err(adev->dev, "GPU mode1 reset failed\n"); | |
47fc644f | 4706 | return ret; |
5c03e584 | 4707 | } |
5c6dd71e | 4708 | |
e3c1b071 | 4709 | int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, |
04442bf7 | 4710 | struct amdgpu_reset_context *reset_context) |
26bc5340 | 4711 | { |
5c1e6fa4 | 4712 | int i, r = 0; |
04442bf7 LL |
4713 | struct amdgpu_job *job = NULL; |
4714 | bool need_full_reset = | |
4715 | test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); | |
4716 | ||
4717 | if (reset_context->reset_req_dev == adev) | |
4718 | job = reset_context->job; | |
71182665 | 4719 | |
b602ca5f TZ |
4720 | if (amdgpu_sriov_vf(adev)) { |
4721 | /* stop the data exchange thread */ | |
4722 | amdgpu_virt_fini_data_exchange(adev); | |
4723 | } | |
4724 | ||
9e225fb9 AG |
4725 | amdgpu_fence_driver_isr_toggle(adev, true); |
4726 | ||
71182665 | 4727 | /* block all schedulers and reset given job's ring */ |
0875dc9e CZ |
4728 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { |
4729 | struct amdgpu_ring *ring = adev->rings[i]; | |
4730 | ||
51687759 | 4731 | if (!ring || !ring->sched.thread) |
0875dc9e | 4732 | continue; |
5740682e | 4733 | |
b8920e1e SS |
4734 | /* Clear job fence from fence drv to avoid force_completion |
4735 | * leave NULL and vm flush fence in fence drv | |
4736 | */ | |
5c1e6fa4 | 4737 | amdgpu_fence_driver_clear_job_fences(ring); |
c530b02f | 4738 | |
2f9d4084 ML |
4739 | /* after all hw jobs are reset, hw fence is meaningless, so force_completion */ |
4740 | amdgpu_fence_driver_force_completion(ring); | |
0875dc9e | 4741 | } |
d38ceaf9 | 4742 | |
9e225fb9 AG |
4743 | amdgpu_fence_driver_isr_toggle(adev, false); |
4744 | ||
ff99849b | 4745 | if (job && job->vm) |
222b5f04 AG |
4746 | drm_sched_increase_karma(&job->base); |
4747 | ||
04442bf7 | 4748 | r = amdgpu_reset_prepare_hwcontext(adev, reset_context); |
404b277b | 4749 | /* If reset handler not implemented, continue; otherwise return */ |
b8920e1e | 4750 | if (r == -EOPNOTSUPP) |
404b277b LL |
4751 | r = 0; |
4752 | else | |
04442bf7 LL |
4753 | return r; |
4754 | ||
1d721ed6 | 4755 | /* Don't suspend on bare metal if we are not going to HW reset the ASIC */ |
26bc5340 AG |
4756 | if (!amdgpu_sriov_vf(adev)) { |
4757 | ||
4758 | if (!need_full_reset) | |
4759 | need_full_reset = amdgpu_device_ip_need_full_reset(adev); | |
4760 | ||
360cd081 LG |
4761 | if (!need_full_reset && amdgpu_gpu_recovery && |
4762 | amdgpu_device_ip_check_soft_reset(adev)) { | |
26bc5340 AG |
4763 | amdgpu_device_ip_pre_soft_reset(adev); |
4764 | r = amdgpu_device_ip_soft_reset(adev); | |
4765 | amdgpu_device_ip_post_soft_reset(adev); | |
4766 | if (r || amdgpu_device_ip_check_soft_reset(adev)) { | |
aac89168 | 4767 | dev_info(adev->dev, "soft reset failed, will fallback to full reset!\n"); |
26bc5340 AG |
4768 | need_full_reset = true; |
4769 | } | |
4770 | } | |
4771 | ||
4772 | if (need_full_reset) | |
4773 | r = amdgpu_device_ip_suspend(adev); | |
04442bf7 LL |
4774 | if (need_full_reset) |
4775 | set_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); | |
4776 | else | |
4777 | clear_bit(AMDGPU_NEED_FULL_RESET, | |
4778 | &reset_context->flags); | |
26bc5340 AG |
4779 | } |
4780 | ||
4781 | return r; | |
4782 | } | |
4783 | ||
15fd09a0 SA |
4784 | static int amdgpu_reset_reg_dumps(struct amdgpu_device *adev) |
4785 | { | |
15fd09a0 SA |
4786 | int i; |
4787 | ||
38a15ad9 | 4788 | lockdep_assert_held(&adev->reset_domain->sem); |
15fd09a0 SA |
4789 | |
4790 | for (i = 0; i < adev->num_regs; i++) { | |
651d7ee6 SA |
4791 | adev->reset_dump_reg_value[i] = RREG32(adev->reset_dump_reg_list[i]); |
4792 | trace_amdgpu_reset_reg_dumps(adev->reset_dump_reg_list[i], | |
4793 | adev->reset_dump_reg_value[i]); | |
15fd09a0 SA |
4794 | } |
4795 | ||
4796 | return 0; | |
4797 | } | |
4798 | ||
3d8785f6 SA |
4799 | #ifdef CONFIG_DEV_COREDUMP |
4800 | static ssize_t amdgpu_devcoredump_read(char *buffer, loff_t offset, | |
4801 | size_t count, void *data, size_t datalen) | |
4802 | { | |
4803 | struct drm_printer p; | |
4804 | struct amdgpu_device *adev = data; | |
4805 | struct drm_print_iterator iter; | |
4806 | int i; | |
4807 | ||
4808 | iter.data = buffer; | |
4809 | iter.offset = 0; | |
4810 | iter.start = offset; | |
4811 | iter.remain = count; | |
4812 | ||
4813 | p = drm_coredump_printer(&iter); | |
4814 | ||
4815 | drm_printf(&p, "**** AMDGPU Device Coredump ****\n"); | |
4816 | drm_printf(&p, "kernel: " UTS_RELEASE "\n"); | |
4817 | drm_printf(&p, "module: " KBUILD_MODNAME "\n"); | |
4818 | drm_printf(&p, "time: %lld.%09ld\n", adev->reset_time.tv_sec, adev->reset_time.tv_nsec); | |
4819 | if (adev->reset_task_info.pid) | |
4820 | drm_printf(&p, "process_name: %s PID: %d\n", | |
4821 | adev->reset_task_info.process_name, | |
4822 | adev->reset_task_info.pid); | |
4823 | ||
4824 | if (adev->reset_vram_lost) | |
4825 | drm_printf(&p, "VRAM is lost due to GPU reset!\n"); | |
4826 | if (adev->num_regs) { | |
4827 | drm_printf(&p, "AMDGPU register dumps:\nOffset: Value:\n"); | |
4828 | ||
4829 | for (i = 0; i < adev->num_regs; i++) | |
4830 | drm_printf(&p, "0x%08x: 0x%08x\n", | |
4831 | adev->reset_dump_reg_list[i], | |
4832 | adev->reset_dump_reg_value[i]); | |
4833 | } | |
4834 | ||
4835 | return count - iter.remain; | |
4836 | } | |
4837 | ||
4838 | static void amdgpu_devcoredump_free(void *data) | |
4839 | { | |
4840 | } | |
4841 | ||
4842 | static void amdgpu_reset_capture_coredumpm(struct amdgpu_device *adev) | |
4843 | { | |
4844 | struct drm_device *dev = adev_to_drm(adev); | |
4845 | ||
4846 | ktime_get_ts64(&adev->reset_time); | |
6d1b3455 | 4847 | dev_coredumpm(dev->dev, THIS_MODULE, adev, 0, GFP_NOWAIT, |
3d8785f6 SA |
4848 | amdgpu_devcoredump_read, amdgpu_devcoredump_free); |
4849 | } | |
4850 | #endif | |
4851 | ||
04442bf7 LL |
4852 | int amdgpu_do_asic_reset(struct list_head *device_list_handle, |
4853 | struct amdgpu_reset_context *reset_context) | |
26bc5340 AG |
4854 | { |
4855 | struct amdgpu_device *tmp_adev = NULL; | |
04442bf7 | 4856 | bool need_full_reset, skip_hw_reset, vram_lost = false; |
26bc5340 | 4857 | int r = 0; |
f5c7e779 | 4858 | bool gpu_reset_for_dev_remove = 0; |
26bc5340 | 4859 | |
04442bf7 LL |
4860 | /* Try reset handler method first */ |
4861 | tmp_adev = list_first_entry(device_list_handle, struct amdgpu_device, | |
4862 | reset_list); | |
15fd09a0 | 4863 | amdgpu_reset_reg_dumps(tmp_adev); |
0a83bb35 LL |
4864 | |
4865 | reset_context->reset_device_list = device_list_handle; | |
04442bf7 | 4866 | r = amdgpu_reset_perform_reset(tmp_adev, reset_context); |
404b277b | 4867 | /* If reset handler not implemented, continue; otherwise return */ |
b8920e1e | 4868 | if (r == -EOPNOTSUPP) |
404b277b LL |
4869 | r = 0; |
4870 | else | |
04442bf7 LL |
4871 | return r; |
4872 | ||
4873 | /* Reset handler not implemented, use the default method */ | |
4874 | need_full_reset = | |
4875 | test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); | |
4876 | skip_hw_reset = test_bit(AMDGPU_SKIP_HW_RESET, &reset_context->flags); | |
4877 | ||
f5c7e779 YC |
4878 | gpu_reset_for_dev_remove = |
4879 | test_bit(AMDGPU_RESET_FOR_DEVICE_REMOVE, &reset_context->flags) && | |
4880 | test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); | |
4881 | ||
26bc5340 | 4882 | /* |
655ce9cb | 4883 | * ASIC reset has to be done on all XGMI hive nodes ASAP |
26bc5340 AG |
4884 | * to allow proper links negotiation in FW (within 1 sec) |
4885 | */ | |
7ac71382 | 4886 | if (!skip_hw_reset && need_full_reset) { |
655ce9cb | 4887 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
041a62bc | 4888 | /* For XGMI run all resets in parallel to speed up the process */ |
d4535e2c | 4889 | if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) { |
e3c1b071 | 4890 | tmp_adev->gmc.xgmi.pending_reset = false; |
c96cf282 | 4891 | if (!queue_work(system_unbound_wq, &tmp_adev->xgmi_reset_work)) |
d4535e2c AG |
4892 | r = -EALREADY; |
4893 | } else | |
4894 | r = amdgpu_asic_reset(tmp_adev); | |
d4535e2c | 4895 | |
041a62bc | 4896 | if (r) { |
aac89168 | 4897 | dev_err(tmp_adev->dev, "ASIC reset failed with error, %d for drm dev, %s", |
4a580877 | 4898 | r, adev_to_drm(tmp_adev)->unique); |
041a62bc | 4899 | break; |
ce316fa5 LM |
4900 | } |
4901 | } | |
4902 | ||
041a62bc AG |
4903 | /* For XGMI wait for all resets to complete before proceed */ |
4904 | if (!r) { | |
655ce9cb | 4905 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
ce316fa5 LM |
4906 | if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) { |
4907 | flush_work(&tmp_adev->xgmi_reset_work); | |
4908 | r = tmp_adev->asic_reset_res; | |
4909 | if (r) | |
4910 | break; | |
ce316fa5 LM |
4911 | } |
4912 | } | |
4913 | } | |
ce316fa5 | 4914 | } |
26bc5340 | 4915 | |
43c4d576 | 4916 | if (!r && amdgpu_ras_intr_triggered()) { |
655ce9cb | 4917 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
5e67bba3 | 4918 | if (tmp_adev->mmhub.ras && tmp_adev->mmhub.ras->ras_block.hw_ops && |
4919 | tmp_adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count) | |
4920 | tmp_adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count(tmp_adev); | |
43c4d576 JC |
4921 | } |
4922 | ||
00eaa571 | 4923 | amdgpu_ras_intr_cleared(); |
43c4d576 | 4924 | } |
00eaa571 | 4925 | |
f5c7e779 YC |
4926 | /* Since the mode1 reset affects base ip blocks, the |
4927 | * phase1 ip blocks need to be resumed. Otherwise there | |
4928 | * will be a BIOS signature error and the psp bootloader | |
4929 | * can't load kdb on the next amdgpu install. | |
4930 | */ | |
4931 | if (gpu_reset_for_dev_remove) { | |
4932 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) | |
4933 | amdgpu_device_ip_resume_phase1(tmp_adev); | |
4934 | ||
4935 | goto end; | |
4936 | } | |
4937 | ||
655ce9cb | 4938 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
26bc5340 AG |
4939 | if (need_full_reset) { |
4940 | /* post card */ | |
e3c1b071 | 4941 | r = amdgpu_device_asic_init(tmp_adev); |
4942 | if (r) { | |
aac89168 | 4943 | dev_warn(tmp_adev->dev, "asic atom init failed!"); |
e3c1b071 | 4944 | } else { |
26bc5340 | 4945 | dev_info(tmp_adev->dev, "GPU reset succeeded, trying to resume\n"); |
9cec53c1 | 4946 | |
26bc5340 AG |
4947 | r = amdgpu_device_ip_resume_phase1(tmp_adev); |
4948 | if (r) | |
4949 | goto out; | |
4950 | ||
4951 | vram_lost = amdgpu_device_check_vram_lost(tmp_adev); | |
3d8785f6 SA |
4952 | #ifdef CONFIG_DEV_COREDUMP |
4953 | tmp_adev->reset_vram_lost = vram_lost; | |
4954 | memset(&tmp_adev->reset_task_info, 0, | |
4955 | sizeof(tmp_adev->reset_task_info)); | |
4956 | if (reset_context->job && reset_context->job->vm) | |
4957 | tmp_adev->reset_task_info = | |
4958 | reset_context->job->vm->task_info; | |
4959 | amdgpu_reset_capture_coredumpm(tmp_adev); | |
4960 | #endif | |
26bc5340 | 4961 | if (vram_lost) { |
77e7f829 | 4962 | DRM_INFO("VRAM is lost due to GPU reset!\n"); |
e3526257 | 4963 | amdgpu_inc_vram_lost(tmp_adev); |
26bc5340 AG |
4964 | } |
4965 | ||
26bc5340 AG |
4966 | r = amdgpu_device_fw_loading(tmp_adev); |
4967 | if (r) | |
4968 | return r; | |
4969 | ||
4970 | r = amdgpu_device_ip_resume_phase2(tmp_adev); | |
4971 | if (r) | |
4972 | goto out; | |
4973 | ||
4974 | if (vram_lost) | |
4975 | amdgpu_device_fill_reset_magic(tmp_adev); | |
4976 | ||
fdafb359 EQ |
4977 | /* |
4978 | * Add this ASIC as tracked as reset was already | |
4979 | * complete successfully. | |
4980 | */ | |
4981 | amdgpu_register_gpu_instance(tmp_adev); | |
4982 | ||
04442bf7 LL |
4983 | if (!reset_context->hive && |
4984 | tmp_adev->gmc.xgmi.num_physical_nodes > 1) | |
e3c1b071 | 4985 | amdgpu_xgmi_add_device(tmp_adev); |
4986 | ||
7c04ca50 | 4987 | r = amdgpu_device_ip_late_init(tmp_adev); |
4988 | if (r) | |
4989 | goto out; | |
4990 | ||
087451f3 | 4991 | drm_fb_helper_set_suspend_unlocked(adev_to_drm(tmp_adev)->fb_helper, false); |
565d1941 | 4992 | |
e8fbaf03 GC |
4993 | /* |
4994 | * The GPU enters bad state once faulty pages | |
4995 | * by ECC has reached the threshold, and ras | |
4996 | * recovery is scheduled next. So add one check | |
4997 | * here to break recovery if it indeed exceeds | |
4998 | * bad page threshold, and remind user to | |
4999 | * retire this GPU or setting one bigger | |
5000 | * bad_page_threshold value to fix this once | |
5001 | * probing driver again. | |
5002 | */ | |
11003c68 | 5003 | if (!amdgpu_ras_eeprom_check_err_threshold(tmp_adev)) { |
e8fbaf03 GC |
5004 | /* must succeed. */ |
5005 | amdgpu_ras_resume(tmp_adev); | |
5006 | } else { | |
5007 | r = -EINVAL; | |
5008 | goto out; | |
5009 | } | |
e79a04d5 | 5010 | |
26bc5340 | 5011 | /* Update PSP FW topology after reset */ |
04442bf7 LL |
5012 | if (reset_context->hive && |
5013 | tmp_adev->gmc.xgmi.num_physical_nodes > 1) | |
5014 | r = amdgpu_xgmi_update_topology( | |
5015 | reset_context->hive, tmp_adev); | |
26bc5340 AG |
5016 | } |
5017 | } | |
5018 | ||
26bc5340 AG |
5019 | out: |
5020 | if (!r) { | |
5021 | amdgpu_irq_gpu_reset_resume_helper(tmp_adev); | |
5022 | r = amdgpu_ib_ring_tests(tmp_adev); | |
5023 | if (r) { | |
5024 | dev_err(tmp_adev->dev, "ib ring test failed (%d).\n", r); | |
26bc5340 AG |
5025 | need_full_reset = true; |
5026 | r = -EAGAIN; | |
5027 | goto end; | |
5028 | } | |
5029 | } | |
5030 | ||
5031 | if (!r) | |
5032 | r = amdgpu_device_recover_vram(tmp_adev); | |
5033 | else | |
5034 | tmp_adev->asic_reset_res = r; | |
5035 | } | |
5036 | ||
5037 | end: | |
04442bf7 LL |
5038 | if (need_full_reset) |
5039 | set_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); | |
5040 | else | |
5041 | clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); | |
26bc5340 AG |
5042 | return r; |
5043 | } | |
5044 | ||
e923be99 | 5045 | static void amdgpu_device_set_mp1_state(struct amdgpu_device *adev) |
26bc5340 | 5046 | { |
5740682e | 5047 | |
a3a09142 AD |
5048 | switch (amdgpu_asic_reset_method(adev)) { |
5049 | case AMD_RESET_METHOD_MODE1: | |
5050 | adev->mp1_state = PP_MP1_STATE_SHUTDOWN; | |
5051 | break; | |
5052 | case AMD_RESET_METHOD_MODE2: | |
5053 | adev->mp1_state = PP_MP1_STATE_RESET; | |
5054 | break; | |
5055 | default: | |
5056 | adev->mp1_state = PP_MP1_STATE_NONE; | |
5057 | break; | |
5058 | } | |
26bc5340 | 5059 | } |
d38ceaf9 | 5060 | |
e923be99 | 5061 | static void amdgpu_device_unset_mp1_state(struct amdgpu_device *adev) |
26bc5340 | 5062 | { |
89041940 | 5063 | amdgpu_vf_error_trans_all(adev); |
a3a09142 | 5064 | adev->mp1_state = PP_MP1_STATE_NONE; |
91fb309d HC |
5065 | } |
5066 | ||
3f12acc8 EQ |
5067 | static void amdgpu_device_resume_display_audio(struct amdgpu_device *adev) |
5068 | { | |
5069 | struct pci_dev *p = NULL; | |
5070 | ||
5071 | p = pci_get_domain_bus_and_slot(pci_domain_nr(adev->pdev->bus), | |
5072 | adev->pdev->bus->number, 1); | |
5073 | if (p) { | |
5074 | pm_runtime_enable(&(p->dev)); | |
5075 | pm_runtime_resume(&(p->dev)); | |
5076 | } | |
b85e285e YY |
5077 | |
5078 | pci_dev_put(p); | |
3f12acc8 EQ |
5079 | } |
5080 | ||
5081 | static int amdgpu_device_suspend_display_audio(struct amdgpu_device *adev) | |
5082 | { | |
5083 | enum amd_reset_method reset_method; | |
5084 | struct pci_dev *p = NULL; | |
5085 | u64 expires; | |
5086 | ||
5087 | /* | |
5088 | * For now, only BACO and mode1 reset are confirmed | |
5089 | * to suffer the audio issue without proper suspended. | |
5090 | */ | |
5091 | reset_method = amdgpu_asic_reset_method(adev); | |
5092 | if ((reset_method != AMD_RESET_METHOD_BACO) && | |
5093 | (reset_method != AMD_RESET_METHOD_MODE1)) | |
5094 | return -EINVAL; | |
5095 | ||
5096 | p = pci_get_domain_bus_and_slot(pci_domain_nr(adev->pdev->bus), | |
5097 | adev->pdev->bus->number, 1); | |
5098 | if (!p) | |
5099 | return -ENODEV; | |
5100 | ||
5101 | expires = pm_runtime_autosuspend_expiration(&(p->dev)); | |
5102 | if (!expires) | |
5103 | /* | |
5104 | * If we cannot get the audio device autosuspend delay, | |
5105 | * a fixed 4S interval will be used. Considering 3S is | |
5106 | * the audio controller default autosuspend delay setting. | |
5107 | * 4S used here is guaranteed to cover that. | |
5108 | */ | |
54b7feb9 | 5109 | expires = ktime_get_mono_fast_ns() + NSEC_PER_SEC * 4ULL; |
3f12acc8 EQ |
5110 | |
5111 | while (!pm_runtime_status_suspended(&(p->dev))) { | |
5112 | if (!pm_runtime_suspend(&(p->dev))) | |
5113 | break; | |
5114 | ||
5115 | if (expires < ktime_get_mono_fast_ns()) { | |
5116 | dev_warn(adev->dev, "failed to suspend display audio\n"); | |
b85e285e | 5117 | pci_dev_put(p); |
3f12acc8 EQ |
5118 | /* TODO: abort the succeeding gpu reset? */ |
5119 | return -ETIMEDOUT; | |
5120 | } | |
5121 | } | |
5122 | ||
5123 | pm_runtime_disable(&(p->dev)); | |
5124 | ||
b85e285e | 5125 | pci_dev_put(p); |
3f12acc8 EQ |
5126 | return 0; |
5127 | } | |
5128 | ||
d193b12b | 5129 | static inline void amdgpu_device_stop_pending_resets(struct amdgpu_device *adev) |
247c7b0d AG |
5130 | { |
5131 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
5132 | ||
5133 | #if defined(CONFIG_DEBUG_FS) | |
5134 | if (!amdgpu_sriov_vf(adev)) | |
5135 | cancel_work(&adev->reset_work); | |
5136 | #endif | |
5137 | ||
5138 | if (adev->kfd.dev) | |
5139 | cancel_work(&adev->kfd.reset_work); | |
5140 | ||
5141 | if (amdgpu_sriov_vf(adev)) | |
5142 | cancel_work(&adev->virt.flr_work); | |
5143 | ||
5144 | if (con && adev->ras_enabled) | |
5145 | cancel_work(&con->recovery_work); | |
5146 | ||
5147 | } | |
5148 | ||
26bc5340 | 5149 | /** |
6e9c65f7 | 5150 | * amdgpu_device_gpu_recover - reset the asic and recover scheduler |
26bc5340 | 5151 | * |
982a820b | 5152 | * @adev: amdgpu_device pointer |
26bc5340 | 5153 | * @job: which job trigger hang |
80bd2de1 | 5154 | * @reset_context: amdgpu reset context pointer |
26bc5340 AG |
5155 | * |
5156 | * Attempt to reset the GPU if it has hung (all asics). | |
5157 | * Attempt to do soft-reset or full-reset and reinitialize Asic | |
5158 | * Returns 0 for success or an error on failure. | |
5159 | */ | |
5160 | ||
cf727044 | 5161 | int amdgpu_device_gpu_recover(struct amdgpu_device *adev, |
f1549c09 LG |
5162 | struct amdgpu_job *job, |
5163 | struct amdgpu_reset_context *reset_context) | |
26bc5340 | 5164 | { |
1d721ed6 | 5165 | struct list_head device_list, *device_list_handle = NULL; |
7dd8c205 | 5166 | bool job_signaled = false; |
26bc5340 | 5167 | struct amdgpu_hive_info *hive = NULL; |
26bc5340 | 5168 | struct amdgpu_device *tmp_adev = NULL; |
1d721ed6 | 5169 | int i, r = 0; |
bb5c7235 | 5170 | bool need_emergency_restart = false; |
3f12acc8 | 5171 | bool audio_suspended = false; |
f5c7e779 YC |
5172 | bool gpu_reset_for_dev_remove = false; |
5173 | ||
5174 | gpu_reset_for_dev_remove = | |
5175 | test_bit(AMDGPU_RESET_FOR_DEVICE_REMOVE, &reset_context->flags) && | |
5176 | test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); | |
26bc5340 | 5177 | |
6e3cd2a9 | 5178 | /* |
bb5c7235 WS |
5179 | * Special case: RAS triggered and full reset isn't supported |
5180 | */ | |
5181 | need_emergency_restart = amdgpu_ras_need_emergency_restart(adev); | |
5182 | ||
d5ea093e AG |
5183 | /* |
5184 | * Flush RAM to disk so that after reboot | |
5185 | * the user can read log and see why the system rebooted. | |
5186 | */ | |
bb5c7235 | 5187 | if (need_emergency_restart && amdgpu_ras_get_context(adev)->reboot) { |
d5ea093e AG |
5188 | DRM_WARN("Emergency reboot."); |
5189 | ||
5190 | ksys_sync_helper(); | |
5191 | emergency_restart(); | |
5192 | } | |
5193 | ||
b823821f | 5194 | dev_info(adev->dev, "GPU %s begin!\n", |
bb5c7235 | 5195 | need_emergency_restart ? "jobs stop":"reset"); |
26bc5340 | 5196 | |
175ac6ec ZL |
5197 | if (!amdgpu_sriov_vf(adev)) |
5198 | hive = amdgpu_get_xgmi_hive(adev); | |
681260df | 5199 | if (hive) |
53b3f8f4 | 5200 | mutex_lock(&hive->hive_lock); |
26bc5340 | 5201 | |
f1549c09 LG |
5202 | reset_context->job = job; |
5203 | reset_context->hive = hive; | |
9e94d22c EQ |
5204 | /* |
5205 | * Build list of devices to reset. | |
5206 | * In case we are in XGMI hive mode, resort the device list | |
5207 | * to put adev in the 1st position. | |
5208 | */ | |
5209 | INIT_LIST_HEAD(&device_list); | |
175ac6ec | 5210 | if (!amdgpu_sriov_vf(adev) && (adev->gmc.xgmi.num_physical_nodes > 1)) { |
83d29a5f | 5211 | list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { |
655ce9cb | 5212 | list_add_tail(&tmp_adev->reset_list, &device_list); |
83d29a5f YC |
5213 | if (gpu_reset_for_dev_remove && adev->shutdown) |
5214 | tmp_adev->shutdown = true; | |
5215 | } | |
655ce9cb | 5216 | if (!list_is_first(&adev->reset_list, &device_list)) |
5217 | list_rotate_to_front(&adev->reset_list, &device_list); | |
5218 | device_list_handle = &device_list; | |
26bc5340 | 5219 | } else { |
655ce9cb | 5220 | list_add_tail(&adev->reset_list, &device_list); |
26bc5340 AG |
5221 | device_list_handle = &device_list; |
5222 | } | |
5223 | ||
e923be99 AG |
5224 | /* We need to lock reset domain only once both for XGMI and single device */ |
5225 | tmp_adev = list_first_entry(device_list_handle, struct amdgpu_device, | |
5226 | reset_list); | |
3675c2f2 | 5227 | amdgpu_device_lock_reset_domain(tmp_adev->reset_domain); |
e923be99 | 5228 | |
1d721ed6 | 5229 | /* block all schedulers and reset given job's ring */ |
655ce9cb | 5230 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
f287a3c5 | 5231 | |
e923be99 | 5232 | amdgpu_device_set_mp1_state(tmp_adev); |
f287a3c5 | 5233 | |
3f12acc8 EQ |
5234 | /* |
5235 | * Try to put the audio codec into suspend state | |
5236 | * before gpu reset started. | |
5237 | * | |
5238 | * Due to the power domain of the graphics device | |
5239 | * is shared with AZ power domain. Without this, | |
5240 | * we may change the audio hardware from behind | |
5241 | * the audio driver's back. That will trigger | |
5242 | * some audio codec errors. | |
5243 | */ | |
5244 | if (!amdgpu_device_suspend_display_audio(tmp_adev)) | |
5245 | audio_suspended = true; | |
5246 | ||
9e94d22c EQ |
5247 | amdgpu_ras_set_error_query_ready(tmp_adev, false); |
5248 | ||
52fb44cf EQ |
5249 | cancel_delayed_work_sync(&tmp_adev->delayed_init_work); |
5250 | ||
c004d44e | 5251 | if (!amdgpu_sriov_vf(tmp_adev)) |
428890a3 | 5252 | amdgpu_amdkfd_pre_reset(tmp_adev); |
9e94d22c | 5253 | |
12ffa55d AG |
5254 | /* |
5255 | * Mark these ASICs to be reseted as untracked first | |
5256 | * And add them back after reset completed | |
5257 | */ | |
5258 | amdgpu_unregister_gpu_instance(tmp_adev); | |
5259 | ||
163d4cd2 | 5260 | drm_fb_helper_set_suspend_unlocked(adev_to_drm(tmp_adev)->fb_helper, true); |
565d1941 | 5261 | |
f1c1314b | 5262 | /* disable ras on ALL IPs */ |
bb5c7235 | 5263 | if (!need_emergency_restart && |
b823821f | 5264 | amdgpu_device_ip_need_full_reset(tmp_adev)) |
f1c1314b | 5265 | amdgpu_ras_suspend(tmp_adev); |
5266 | ||
1d721ed6 AG |
5267 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { |
5268 | struct amdgpu_ring *ring = tmp_adev->rings[i]; | |
5269 | ||
5270 | if (!ring || !ring->sched.thread) | |
5271 | continue; | |
5272 | ||
0b2d2c2e | 5273 | drm_sched_stop(&ring->sched, job ? &job->base : NULL); |
7c6e68c7 | 5274 | |
bb5c7235 | 5275 | if (need_emergency_restart) |
7c6e68c7 | 5276 | amdgpu_job_stop_all_jobs_on_sched(&ring->sched); |
1d721ed6 | 5277 | } |
8f8c80f4 | 5278 | atomic_inc(&tmp_adev->gpu_reset_counter); |
1d721ed6 AG |
5279 | } |
5280 | ||
bb5c7235 | 5281 | if (need_emergency_restart) |
7c6e68c7 AG |
5282 | goto skip_sched_resume; |
5283 | ||
1d721ed6 AG |
5284 | /* |
5285 | * Must check guilty signal here since after this point all old | |
5286 | * HW fences are force signaled. | |
5287 | * | |
5288 | * job->base holds a reference to parent fence | |
5289 | */ | |
f6a3f660 | 5290 | if (job && dma_fence_is_signaled(&job->hw_fence)) { |
1d721ed6 | 5291 | job_signaled = true; |
1d721ed6 AG |
5292 | dev_info(adev->dev, "Guilty job already signaled, skipping HW reset"); |
5293 | goto skip_hw_reset; | |
5294 | } | |
5295 | ||
26bc5340 | 5296 | retry: /* Rest of adevs pre asic reset from XGMI hive. */ |
655ce9cb | 5297 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
f5c7e779 YC |
5298 | if (gpu_reset_for_dev_remove) { |
5299 | /* Workaroud for ASICs need to disable SMC first */ | |
5300 | amdgpu_device_smu_fini_early(tmp_adev); | |
5301 | } | |
f1549c09 | 5302 | r = amdgpu_device_pre_asic_reset(tmp_adev, reset_context); |
26bc5340 AG |
5303 | /*TODO Should we stop ?*/ |
5304 | if (r) { | |
aac89168 | 5305 | dev_err(tmp_adev->dev, "GPU pre asic reset failed with err, %d for drm dev, %s ", |
4a580877 | 5306 | r, adev_to_drm(tmp_adev)->unique); |
26bc5340 AG |
5307 | tmp_adev->asic_reset_res = r; |
5308 | } | |
247c7b0d AG |
5309 | |
5310 | /* | |
5311 | * Drop all pending non scheduler resets. Scheduler resets | |
5312 | * were already dropped during drm_sched_stop | |
5313 | */ | |
d193b12b | 5314 | amdgpu_device_stop_pending_resets(tmp_adev); |
26bc5340 AG |
5315 | } |
5316 | ||
5317 | /* Actual ASIC resets if needed.*/ | |
4f30d920 | 5318 | /* Host driver will handle XGMI hive reset for SRIOV */ |
26bc5340 AG |
5319 | if (amdgpu_sriov_vf(adev)) { |
5320 | r = amdgpu_device_reset_sriov(adev, job ? false : true); | |
5321 | if (r) | |
5322 | adev->asic_reset_res = r; | |
950d6425 | 5323 | |
28606c4e YC |
5324 | /* Aldebaran and gfx_11_0_3 support ras in SRIOV, so need resume ras during reset */ |
5325 | if (adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 2) || | |
5326 | adev->ip_versions[GC_HWIP][0] == IP_VERSION(11, 0, 3)) | |
950d6425 | 5327 | amdgpu_ras_resume(adev); |
26bc5340 | 5328 | } else { |
f1549c09 | 5329 | r = amdgpu_do_asic_reset(device_list_handle, reset_context); |
b98a1648 | 5330 | if (r && r == -EAGAIN) |
26bc5340 | 5331 | goto retry; |
f5c7e779 YC |
5332 | |
5333 | if (!r && gpu_reset_for_dev_remove) | |
5334 | goto recover_end; | |
26bc5340 AG |
5335 | } |
5336 | ||
1d721ed6 AG |
5337 | skip_hw_reset: |
5338 | ||
26bc5340 | 5339 | /* Post ASIC reset for all devs .*/ |
655ce9cb | 5340 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
7c6e68c7 | 5341 | |
1d721ed6 AG |
5342 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { |
5343 | struct amdgpu_ring *ring = tmp_adev->rings[i]; | |
5344 | ||
5345 | if (!ring || !ring->sched.thread) | |
5346 | continue; | |
5347 | ||
6868a2c4 | 5348 | drm_sched_start(&ring->sched, true); |
1d721ed6 AG |
5349 | } |
5350 | ||
693073a0 | 5351 | if (adev->enable_mes && adev->ip_versions[GC_HWIP][0] != IP_VERSION(11, 0, 3)) |
ed67f729 JX |
5352 | amdgpu_mes_self_test(tmp_adev); |
5353 | ||
b8920e1e | 5354 | if (!drm_drv_uses_atomic_modeset(adev_to_drm(tmp_adev)) && !job_signaled) |
4a580877 | 5355 | drm_helper_resume_force_mode(adev_to_drm(tmp_adev)); |
1d721ed6 | 5356 | |
7258fa31 SK |
5357 | if (tmp_adev->asic_reset_res) |
5358 | r = tmp_adev->asic_reset_res; | |
5359 | ||
1d721ed6 | 5360 | tmp_adev->asic_reset_res = 0; |
26bc5340 AG |
5361 | |
5362 | if (r) { | |
5363 | /* bad news, how to tell it to userspace ? */ | |
12ffa55d | 5364 | dev_info(tmp_adev->dev, "GPU reset(%d) failed\n", atomic_read(&tmp_adev->gpu_reset_counter)); |
26bc5340 AG |
5365 | amdgpu_vf_error_put(tmp_adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r); |
5366 | } else { | |
12ffa55d | 5367 | dev_info(tmp_adev->dev, "GPU reset(%d) succeeded!\n", atomic_read(&tmp_adev->gpu_reset_counter)); |
3fa8f89d S |
5368 | if (amdgpu_acpi_smart_shift_update(adev_to_drm(tmp_adev), AMDGPU_SS_DEV_D0)) |
5369 | DRM_WARN("smart shift update failed\n"); | |
26bc5340 | 5370 | } |
7c6e68c7 | 5371 | } |
26bc5340 | 5372 | |
7c6e68c7 | 5373 | skip_sched_resume: |
655ce9cb | 5374 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
428890a3 | 5375 | /* unlock kfd: SRIOV would do it separately */ |
c004d44e | 5376 | if (!need_emergency_restart && !amdgpu_sriov_vf(tmp_adev)) |
428890a3 | 5377 | amdgpu_amdkfd_post_reset(tmp_adev); |
8e2712e7 | 5378 | |
5379 | /* kfd_post_reset will do nothing if kfd device is not initialized, | |
5380 | * need to bring up kfd here if it's not be initialized before | |
5381 | */ | |
5382 | if (!adev->kfd.init_complete) | |
5383 | amdgpu_amdkfd_device_init(adev); | |
5384 | ||
3f12acc8 EQ |
5385 | if (audio_suspended) |
5386 | amdgpu_device_resume_display_audio(tmp_adev); | |
e923be99 AG |
5387 | |
5388 | amdgpu_device_unset_mp1_state(tmp_adev); | |
d293470e YC |
5389 | |
5390 | amdgpu_ras_set_error_query_ready(tmp_adev, true); | |
26bc5340 AG |
5391 | } |
5392 | ||
f5c7e779 | 5393 | recover_end: |
e923be99 AG |
5394 | tmp_adev = list_first_entry(device_list_handle, struct amdgpu_device, |
5395 | reset_list); | |
5396 | amdgpu_device_unlock_reset_domain(tmp_adev->reset_domain); | |
5397 | ||
9e94d22c | 5398 | if (hive) { |
9e94d22c | 5399 | mutex_unlock(&hive->hive_lock); |
d95e8e97 | 5400 | amdgpu_put_xgmi_hive(hive); |
9e94d22c | 5401 | } |
26bc5340 | 5402 | |
f287a3c5 | 5403 | if (r) |
26bc5340 | 5404 | dev_info(adev->dev, "GPU reset end with ret = %d\n", r); |
ab9a0b1f AG |
5405 | |
5406 | atomic_set(&adev->reset_domain->reset_res, r); | |
d38ceaf9 AD |
5407 | return r; |
5408 | } | |
5409 | ||
e3ecdffa AD |
5410 | /** |
5411 | * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot | |
5412 | * | |
5413 | * @adev: amdgpu_device pointer | |
5414 | * | |
5415 | * Fetchs and stores in the driver the PCIE capabilities (gen speed | |
5416 | * and lanes) of the slot the device is in. Handles APUs and | |
5417 | * virtualized environments where PCIE config space may not be available. | |
5418 | */ | |
5494d864 | 5419 | static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev) |
d0dd7f0c | 5420 | { |
5d9a6330 | 5421 | struct pci_dev *pdev; |
c5313457 HK |
5422 | enum pci_bus_speed speed_cap, platform_speed_cap; |
5423 | enum pcie_link_width platform_link_width; | |
d0dd7f0c | 5424 | |
cd474ba0 AD |
5425 | if (amdgpu_pcie_gen_cap) |
5426 | adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap; | |
d0dd7f0c | 5427 | |
cd474ba0 AD |
5428 | if (amdgpu_pcie_lane_cap) |
5429 | adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap; | |
d0dd7f0c | 5430 | |
cd474ba0 | 5431 | /* covers APUs as well */ |
04e85958 | 5432 | if (pci_is_root_bus(adev->pdev->bus) && !amdgpu_passthrough(adev)) { |
cd474ba0 AD |
5433 | if (adev->pm.pcie_gen_mask == 0) |
5434 | adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK; | |
5435 | if (adev->pm.pcie_mlw_mask == 0) | |
5436 | adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK; | |
d0dd7f0c | 5437 | return; |
cd474ba0 | 5438 | } |
d0dd7f0c | 5439 | |
c5313457 HK |
5440 | if (adev->pm.pcie_gen_mask && adev->pm.pcie_mlw_mask) |
5441 | return; | |
5442 | ||
dbaa922b AD |
5443 | pcie_bandwidth_available(adev->pdev, NULL, |
5444 | &platform_speed_cap, &platform_link_width); | |
c5313457 | 5445 | |
cd474ba0 | 5446 | if (adev->pm.pcie_gen_mask == 0) { |
5d9a6330 AD |
5447 | /* asic caps */ |
5448 | pdev = adev->pdev; | |
5449 | speed_cap = pcie_get_speed_cap(pdev); | |
5450 | if (speed_cap == PCI_SPEED_UNKNOWN) { | |
5451 | adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 | | |
cd474ba0 AD |
5452 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 | |
5453 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3); | |
cd474ba0 | 5454 | } else { |
2b3a1f51 FX |
5455 | if (speed_cap == PCIE_SPEED_32_0GT) |
5456 | adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 | | |
5457 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 | | |
5458 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 | | |
5459 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4 | | |
5460 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN5); | |
5461 | else if (speed_cap == PCIE_SPEED_16_0GT) | |
5d9a6330 AD |
5462 | adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 | |
5463 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 | | |
5464 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 | | |
5465 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4); | |
5466 | else if (speed_cap == PCIE_SPEED_8_0GT) | |
5467 | adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 | | |
5468 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 | | |
5469 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3); | |
5470 | else if (speed_cap == PCIE_SPEED_5_0GT) | |
5471 | adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 | | |
5472 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2); | |
5473 | else | |
5474 | adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1; | |
5475 | } | |
5476 | /* platform caps */ | |
c5313457 | 5477 | if (platform_speed_cap == PCI_SPEED_UNKNOWN) { |
5d9a6330 AD |
5478 | adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 | |
5479 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2); | |
5480 | } else { | |
2b3a1f51 FX |
5481 | if (platform_speed_cap == PCIE_SPEED_32_0GT) |
5482 | adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 | | |
5483 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 | | |
5484 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 | | |
5485 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4 | | |
5486 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN5); | |
5487 | else if (platform_speed_cap == PCIE_SPEED_16_0GT) | |
5d9a6330 AD |
5488 | adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 | |
5489 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 | | |
5490 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 | | |
5491 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4); | |
c5313457 | 5492 | else if (platform_speed_cap == PCIE_SPEED_8_0GT) |
5d9a6330 AD |
5493 | adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 | |
5494 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 | | |
5495 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3); | |
c5313457 | 5496 | else if (platform_speed_cap == PCIE_SPEED_5_0GT) |
5d9a6330 AD |
5497 | adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 | |
5498 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2); | |
5499 | else | |
5500 | adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1; | |
5501 | ||
cd474ba0 AD |
5502 | } |
5503 | } | |
5504 | if (adev->pm.pcie_mlw_mask == 0) { | |
c5313457 | 5505 | if (platform_link_width == PCIE_LNK_WIDTH_UNKNOWN) { |
5d9a6330 AD |
5506 | adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK; |
5507 | } else { | |
c5313457 | 5508 | switch (platform_link_width) { |
5d9a6330 | 5509 | case PCIE_LNK_X32: |
cd474ba0 AD |
5510 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 | |
5511 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 | | |
5512 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 | | |
5513 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 | | |
5514 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | | |
5515 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | | |
5516 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); | |
5517 | break; | |
5d9a6330 | 5518 | case PCIE_LNK_X16: |
cd474ba0 AD |
5519 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 | |
5520 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 | | |
5521 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 | | |
5522 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | | |
5523 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | | |
5524 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); | |
5525 | break; | |
5d9a6330 | 5526 | case PCIE_LNK_X12: |
cd474ba0 AD |
5527 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 | |
5528 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 | | |
5529 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | | |
5530 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | | |
5531 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); | |
5532 | break; | |
5d9a6330 | 5533 | case PCIE_LNK_X8: |
cd474ba0 AD |
5534 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 | |
5535 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | | |
5536 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | | |
5537 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); | |
5538 | break; | |
5d9a6330 | 5539 | case PCIE_LNK_X4: |
cd474ba0 AD |
5540 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | |
5541 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | | |
5542 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); | |
5543 | break; | |
5d9a6330 | 5544 | case PCIE_LNK_X2: |
cd474ba0 AD |
5545 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | |
5546 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); | |
5547 | break; | |
5d9a6330 | 5548 | case PCIE_LNK_X1: |
cd474ba0 AD |
5549 | adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1; |
5550 | break; | |
5551 | default: | |
5552 | break; | |
5553 | } | |
d0dd7f0c AD |
5554 | } |
5555 | } | |
5556 | } | |
d38ceaf9 | 5557 | |
08a2fd23 RE |
5558 | /** |
5559 | * amdgpu_device_is_peer_accessible - Check peer access through PCIe BAR | |
5560 | * | |
5561 | * @adev: amdgpu_device pointer | |
5562 | * @peer_adev: amdgpu_device pointer for peer device trying to access @adev | |
5563 | * | |
5564 | * Return true if @peer_adev can access (DMA) @adev through the PCIe | |
5565 | * BAR, i.e. @adev is "large BAR" and the BAR matches the DMA mask of | |
5566 | * @peer_adev. | |
5567 | */ | |
5568 | bool amdgpu_device_is_peer_accessible(struct amdgpu_device *adev, | |
5569 | struct amdgpu_device *peer_adev) | |
5570 | { | |
5571 | #ifdef CONFIG_HSA_AMD_P2P | |
5572 | uint64_t address_mask = peer_adev->dev->dma_mask ? | |
5573 | ~*peer_adev->dev->dma_mask : ~((1ULL << 32) - 1); | |
5574 | resource_size_t aper_limit = | |
5575 | adev->gmc.aper_base + adev->gmc.aper_size - 1; | |
bb66ecbf LL |
5576 | bool p2p_access = |
5577 | !adev->gmc.xgmi.connected_to_cpu && | |
5578 | !(pci_p2pdma_distance(adev->pdev, peer_adev->dev, false) < 0); | |
08a2fd23 RE |
5579 | |
5580 | return pcie_p2p && p2p_access && (adev->gmc.visible_vram_size && | |
5581 | adev->gmc.real_vram_size == adev->gmc.visible_vram_size && | |
5582 | !(adev->gmc.aper_base & address_mask || | |
5583 | aper_limit & address_mask)); | |
5584 | #else | |
5585 | return false; | |
5586 | #endif | |
5587 | } | |
5588 | ||
361dbd01 AD |
5589 | int amdgpu_device_baco_enter(struct drm_device *dev) |
5590 | { | |
1348969a | 5591 | struct amdgpu_device *adev = drm_to_adev(dev); |
7a22677b | 5592 | struct amdgpu_ras *ras = amdgpu_ras_get_context(adev); |
361dbd01 | 5593 | |
6ab68650 | 5594 | if (!amdgpu_device_supports_baco(dev)) |
361dbd01 AD |
5595 | return -ENOTSUPP; |
5596 | ||
8ab0d6f0 | 5597 | if (ras && adev->ras_enabled && |
acdae216 | 5598 | adev->nbio.funcs->enable_doorbell_interrupt) |
7a22677b LM |
5599 | adev->nbio.funcs->enable_doorbell_interrupt(adev, false); |
5600 | ||
9530273e | 5601 | return amdgpu_dpm_baco_enter(adev); |
361dbd01 AD |
5602 | } |
5603 | ||
5604 | int amdgpu_device_baco_exit(struct drm_device *dev) | |
5605 | { | |
1348969a | 5606 | struct amdgpu_device *adev = drm_to_adev(dev); |
7a22677b | 5607 | struct amdgpu_ras *ras = amdgpu_ras_get_context(adev); |
9530273e | 5608 | int ret = 0; |
361dbd01 | 5609 | |
6ab68650 | 5610 | if (!amdgpu_device_supports_baco(dev)) |
361dbd01 AD |
5611 | return -ENOTSUPP; |
5612 | ||
9530273e EQ |
5613 | ret = amdgpu_dpm_baco_exit(adev); |
5614 | if (ret) | |
5615 | return ret; | |
7a22677b | 5616 | |
8ab0d6f0 | 5617 | if (ras && adev->ras_enabled && |
acdae216 | 5618 | adev->nbio.funcs->enable_doorbell_interrupt) |
7a22677b LM |
5619 | adev->nbio.funcs->enable_doorbell_interrupt(adev, true); |
5620 | ||
1bece222 CL |
5621 | if (amdgpu_passthrough(adev) && |
5622 | adev->nbio.funcs->clear_doorbell_interrupt) | |
5623 | adev->nbio.funcs->clear_doorbell_interrupt(adev); | |
5624 | ||
7a22677b | 5625 | return 0; |
361dbd01 | 5626 | } |
c9a6b82f AG |
5627 | |
5628 | /** | |
5629 | * amdgpu_pci_error_detected - Called when a PCI error is detected. | |
5630 | * @pdev: PCI device struct | |
5631 | * @state: PCI channel state | |
5632 | * | |
5633 | * Description: Called when a PCI error is detected. | |
5634 | * | |
5635 | * Return: PCI_ERS_RESULT_NEED_RESET or PCI_ERS_RESULT_DISCONNECT. | |
5636 | */ | |
5637 | pci_ers_result_t amdgpu_pci_error_detected(struct pci_dev *pdev, pci_channel_state_t state) | |
5638 | { | |
5639 | struct drm_device *dev = pci_get_drvdata(pdev); | |
5640 | struct amdgpu_device *adev = drm_to_adev(dev); | |
acd89fca | 5641 | int i; |
c9a6b82f AG |
5642 | |
5643 | DRM_INFO("PCI error: detected callback, state(%d)!!\n", state); | |
5644 | ||
6894305c AG |
5645 | if (adev->gmc.xgmi.num_physical_nodes > 1) { |
5646 | DRM_WARN("No support for XGMI hive yet..."); | |
5647 | return PCI_ERS_RESULT_DISCONNECT; | |
5648 | } | |
5649 | ||
e17e27f9 GC |
5650 | adev->pci_channel_state = state; |
5651 | ||
c9a6b82f AG |
5652 | switch (state) { |
5653 | case pci_channel_io_normal: | |
5654 | return PCI_ERS_RESULT_CAN_RECOVER; | |
acd89fca | 5655 | /* Fatal error, prepare for slot reset */ |
8a11d283 TZ |
5656 | case pci_channel_io_frozen: |
5657 | /* | |
d0fb18b5 | 5658 | * Locking adev->reset_domain->sem will prevent any external access |
acd89fca AG |
5659 | * to GPU during PCI error recovery |
5660 | */ | |
3675c2f2 | 5661 | amdgpu_device_lock_reset_domain(adev->reset_domain); |
e923be99 | 5662 | amdgpu_device_set_mp1_state(adev); |
acd89fca AG |
5663 | |
5664 | /* | |
5665 | * Block any work scheduling as we do for regular GPU reset | |
5666 | * for the duration of the recovery | |
5667 | */ | |
5668 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { | |
5669 | struct amdgpu_ring *ring = adev->rings[i]; | |
5670 | ||
5671 | if (!ring || !ring->sched.thread) | |
5672 | continue; | |
5673 | ||
5674 | drm_sched_stop(&ring->sched, NULL); | |
5675 | } | |
8f8c80f4 | 5676 | atomic_inc(&adev->gpu_reset_counter); |
c9a6b82f AG |
5677 | return PCI_ERS_RESULT_NEED_RESET; |
5678 | case pci_channel_io_perm_failure: | |
5679 | /* Permanent error, prepare for device removal */ | |
5680 | return PCI_ERS_RESULT_DISCONNECT; | |
5681 | } | |
5682 | ||
5683 | return PCI_ERS_RESULT_NEED_RESET; | |
5684 | } | |
5685 | ||
5686 | /** | |
5687 | * amdgpu_pci_mmio_enabled - Enable MMIO and dump debug registers | |
5688 | * @pdev: pointer to PCI device | |
5689 | */ | |
5690 | pci_ers_result_t amdgpu_pci_mmio_enabled(struct pci_dev *pdev) | |
5691 | { | |
5692 | ||
5693 | DRM_INFO("PCI error: mmio enabled callback!!\n"); | |
5694 | ||
5695 | /* TODO - dump whatever for debugging purposes */ | |
5696 | ||
5697 | /* This called only if amdgpu_pci_error_detected returns | |
5698 | * PCI_ERS_RESULT_CAN_RECOVER. Read/write to the device still | |
5699 | * works, no need to reset slot. | |
5700 | */ | |
5701 | ||
5702 | return PCI_ERS_RESULT_RECOVERED; | |
5703 | } | |
5704 | ||
5705 | /** | |
5706 | * amdgpu_pci_slot_reset - Called when PCI slot has been reset. | |
5707 | * @pdev: PCI device struct | |
5708 | * | |
5709 | * Description: This routine is called by the pci error recovery | |
5710 | * code after the PCI slot has been reset, just before we | |
5711 | * should resume normal operations. | |
5712 | */ | |
5713 | pci_ers_result_t amdgpu_pci_slot_reset(struct pci_dev *pdev) | |
5714 | { | |
5715 | struct drm_device *dev = pci_get_drvdata(pdev); | |
5716 | struct amdgpu_device *adev = drm_to_adev(dev); | |
362c7b91 | 5717 | int r, i; |
04442bf7 | 5718 | struct amdgpu_reset_context reset_context; |
362c7b91 | 5719 | u32 memsize; |
7ac71382 | 5720 | struct list_head device_list; |
c9a6b82f AG |
5721 | |
5722 | DRM_INFO("PCI error: slot reset callback!!\n"); | |
5723 | ||
04442bf7 LL |
5724 | memset(&reset_context, 0, sizeof(reset_context)); |
5725 | ||
7ac71382 | 5726 | INIT_LIST_HEAD(&device_list); |
655ce9cb | 5727 | list_add_tail(&adev->reset_list, &device_list); |
7ac71382 | 5728 | |
362c7b91 AG |
5729 | /* wait for asic to come out of reset */ |
5730 | msleep(500); | |
5731 | ||
7ac71382 | 5732 | /* Restore PCI confspace */ |
c1dd4aa6 | 5733 | amdgpu_device_load_pci_state(pdev); |
c9a6b82f | 5734 | |
362c7b91 AG |
5735 | /* confirm ASIC came out of reset */ |
5736 | for (i = 0; i < adev->usec_timeout; i++) { | |
5737 | memsize = amdgpu_asic_get_config_memsize(adev); | |
5738 | ||
5739 | if (memsize != 0xffffffff) | |
5740 | break; | |
5741 | udelay(1); | |
5742 | } | |
5743 | if (memsize == 0xffffffff) { | |
5744 | r = -ETIME; | |
5745 | goto out; | |
5746 | } | |
5747 | ||
04442bf7 LL |
5748 | reset_context.method = AMD_RESET_METHOD_NONE; |
5749 | reset_context.reset_req_dev = adev; | |
5750 | set_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags); | |
5751 | set_bit(AMDGPU_SKIP_HW_RESET, &reset_context.flags); | |
5752 | ||
7afefb81 | 5753 | adev->no_hw_access = true; |
04442bf7 | 5754 | r = amdgpu_device_pre_asic_reset(adev, &reset_context); |
7afefb81 | 5755 | adev->no_hw_access = false; |
c9a6b82f AG |
5756 | if (r) |
5757 | goto out; | |
5758 | ||
04442bf7 | 5759 | r = amdgpu_do_asic_reset(&device_list, &reset_context); |
c9a6b82f AG |
5760 | |
5761 | out: | |
c9a6b82f | 5762 | if (!r) { |
c1dd4aa6 AG |
5763 | if (amdgpu_device_cache_pci_state(adev->pdev)) |
5764 | pci_restore_state(adev->pdev); | |
5765 | ||
c9a6b82f AG |
5766 | DRM_INFO("PCIe error recovery succeeded\n"); |
5767 | } else { | |
5768 | DRM_ERROR("PCIe error recovery failed, err:%d", r); | |
e923be99 AG |
5769 | amdgpu_device_unset_mp1_state(adev); |
5770 | amdgpu_device_unlock_reset_domain(adev->reset_domain); | |
c9a6b82f AG |
5771 | } |
5772 | ||
5773 | return r ? PCI_ERS_RESULT_DISCONNECT : PCI_ERS_RESULT_RECOVERED; | |
5774 | } | |
5775 | ||
5776 | /** | |
5777 | * amdgpu_pci_resume() - resume normal ops after PCI reset | |
5778 | * @pdev: pointer to PCI device | |
5779 | * | |
5780 | * Called when the error recovery driver tells us that its | |
505199a3 | 5781 | * OK to resume normal operation. |
c9a6b82f AG |
5782 | */ |
5783 | void amdgpu_pci_resume(struct pci_dev *pdev) | |
5784 | { | |
5785 | struct drm_device *dev = pci_get_drvdata(pdev); | |
5786 | struct amdgpu_device *adev = drm_to_adev(dev); | |
acd89fca | 5787 | int i; |
c9a6b82f | 5788 | |
c9a6b82f AG |
5789 | |
5790 | DRM_INFO("PCI error: resume callback!!\n"); | |
acd89fca | 5791 | |
e17e27f9 GC |
5792 | /* Only continue execution for the case of pci_channel_io_frozen */ |
5793 | if (adev->pci_channel_state != pci_channel_io_frozen) | |
5794 | return; | |
5795 | ||
acd89fca AG |
5796 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { |
5797 | struct amdgpu_ring *ring = adev->rings[i]; | |
5798 | ||
5799 | if (!ring || !ring->sched.thread) | |
5800 | continue; | |
5801 | ||
acd89fca AG |
5802 | drm_sched_start(&ring->sched, true); |
5803 | } | |
5804 | ||
e923be99 AG |
5805 | amdgpu_device_unset_mp1_state(adev); |
5806 | amdgpu_device_unlock_reset_domain(adev->reset_domain); | |
c9a6b82f | 5807 | } |
c1dd4aa6 AG |
5808 | |
5809 | bool amdgpu_device_cache_pci_state(struct pci_dev *pdev) | |
5810 | { | |
5811 | struct drm_device *dev = pci_get_drvdata(pdev); | |
5812 | struct amdgpu_device *adev = drm_to_adev(dev); | |
5813 | int r; | |
5814 | ||
5815 | r = pci_save_state(pdev); | |
5816 | if (!r) { | |
5817 | kfree(adev->pci_state); | |
5818 | ||
5819 | adev->pci_state = pci_store_saved_state(pdev); | |
5820 | ||
5821 | if (!adev->pci_state) { | |
5822 | DRM_ERROR("Failed to store PCI saved state"); | |
5823 | return false; | |
5824 | } | |
5825 | } else { | |
5826 | DRM_WARN("Failed to save PCI state, err:%d\n", r); | |
5827 | return false; | |
5828 | } | |
5829 | ||
5830 | return true; | |
5831 | } | |
5832 | ||
5833 | bool amdgpu_device_load_pci_state(struct pci_dev *pdev) | |
5834 | { | |
5835 | struct drm_device *dev = pci_get_drvdata(pdev); | |
5836 | struct amdgpu_device *adev = drm_to_adev(dev); | |
5837 | int r; | |
5838 | ||
5839 | if (!adev->pci_state) | |
5840 | return false; | |
5841 | ||
5842 | r = pci_load_saved_state(pdev, adev->pci_state); | |
5843 | ||
5844 | if (!r) { | |
5845 | pci_restore_state(pdev); | |
5846 | } else { | |
5847 | DRM_WARN("Failed to load PCI state, err:%d\n", r); | |
5848 | return false; | |
5849 | } | |
5850 | ||
5851 | return true; | |
5852 | } | |
5853 | ||
810085dd EH |
5854 | void amdgpu_device_flush_hdp(struct amdgpu_device *adev, |
5855 | struct amdgpu_ring *ring) | |
5856 | { | |
5857 | #ifdef CONFIG_X86_64 | |
b818a5d3 | 5858 | if ((adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev)) |
810085dd EH |
5859 | return; |
5860 | #endif | |
5861 | if (adev->gmc.xgmi.connected_to_cpu) | |
5862 | return; | |
5863 | ||
5864 | if (ring && ring->funcs->emit_hdp_flush) | |
5865 | amdgpu_ring_emit_hdp_flush(ring); | |
5866 | else | |
5867 | amdgpu_asic_flush_hdp(adev, ring); | |
5868 | } | |
c1dd4aa6 | 5869 | |
810085dd EH |
5870 | void amdgpu_device_invalidate_hdp(struct amdgpu_device *adev, |
5871 | struct amdgpu_ring *ring) | |
5872 | { | |
5873 | #ifdef CONFIG_X86_64 | |
b818a5d3 | 5874 | if ((adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev)) |
810085dd EH |
5875 | return; |
5876 | #endif | |
5877 | if (adev->gmc.xgmi.connected_to_cpu) | |
5878 | return; | |
c1dd4aa6 | 5879 | |
810085dd EH |
5880 | amdgpu_asic_invalidate_hdp(adev, ring); |
5881 | } | |
34f3a4a9 | 5882 | |
89a7a870 AG |
5883 | int amdgpu_in_reset(struct amdgpu_device *adev) |
5884 | { | |
5885 | return atomic_read(&adev->reset_domain->in_gpu_reset); | |
53a17b6b TZ |
5886 | } |
5887 | ||
34f3a4a9 LY |
5888 | /** |
5889 | * amdgpu_device_halt() - bring hardware to some kind of halt state | |
5890 | * | |
5891 | * @adev: amdgpu_device pointer | |
5892 | * | |
5893 | * Bring hardware to some kind of halt state so that no one can touch it | |
5894 | * any more. It will help to maintain error context when error occurred. | |
5895 | * Compare to a simple hang, the system will keep stable at least for SSH | |
5896 | * access. Then it should be trivial to inspect the hardware state and | |
5897 | * see what's going on. Implemented as following: | |
5898 | * | |
5899 | * 1. drm_dev_unplug() makes device inaccessible to user space(IOCTLs, etc), | |
5900 | * clears all CPU mappings to device, disallows remappings through page faults | |
5901 | * 2. amdgpu_irq_disable_all() disables all interrupts | |
5902 | * 3. amdgpu_fence_driver_hw_fini() signals all HW fences | |
5903 | * 4. set adev->no_hw_access to avoid potential crashes after setp 5 | |
5904 | * 5. amdgpu_device_unmap_mmio() clears all MMIO mappings | |
5905 | * 6. pci_disable_device() and pci_wait_for_pending_transaction() | |
5906 | * flush any in flight DMA operations | |
5907 | */ | |
5908 | void amdgpu_device_halt(struct amdgpu_device *adev) | |
5909 | { | |
5910 | struct pci_dev *pdev = adev->pdev; | |
e0f943b4 | 5911 | struct drm_device *ddev = adev_to_drm(adev); |
34f3a4a9 | 5912 | |
2c1c7ba4 | 5913 | amdgpu_xcp_dev_unplug(adev); |
34f3a4a9 LY |
5914 | drm_dev_unplug(ddev); |
5915 | ||
5916 | amdgpu_irq_disable_all(adev); | |
5917 | ||
5918 | amdgpu_fence_driver_hw_fini(adev); | |
5919 | ||
5920 | adev->no_hw_access = true; | |
5921 | ||
5922 | amdgpu_device_unmap_mmio(adev); | |
5923 | ||
5924 | pci_disable_device(pdev); | |
5925 | pci_wait_for_pending_transaction(pdev); | |
5926 | } | |
86700a40 XD |
5927 | |
5928 | u32 amdgpu_device_pcie_port_rreg(struct amdgpu_device *adev, | |
5929 | u32 reg) | |
5930 | { | |
5931 | unsigned long flags, address, data; | |
5932 | u32 r; | |
5933 | ||
5934 | address = adev->nbio.funcs->get_pcie_port_index_offset(adev); | |
5935 | data = adev->nbio.funcs->get_pcie_port_data_offset(adev); | |
5936 | ||
5937 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); | |
5938 | WREG32(address, reg * 4); | |
5939 | (void)RREG32(address); | |
5940 | r = RREG32(data); | |
5941 | spin_unlock_irqrestore(&adev->pcie_idx_lock, flags); | |
5942 | return r; | |
5943 | } | |
5944 | ||
5945 | void amdgpu_device_pcie_port_wreg(struct amdgpu_device *adev, | |
5946 | u32 reg, u32 v) | |
5947 | { | |
5948 | unsigned long flags, address, data; | |
5949 | ||
5950 | address = adev->nbio.funcs->get_pcie_port_index_offset(adev); | |
5951 | data = adev->nbio.funcs->get_pcie_port_data_offset(adev); | |
5952 | ||
5953 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); | |
5954 | WREG32(address, reg * 4); | |
5955 | (void)RREG32(address); | |
5956 | WREG32(data, v); | |
5957 | (void)RREG32(data); | |
5958 | spin_unlock_irqrestore(&adev->pcie_idx_lock, flags); | |
5959 | } | |
68ce8b24 CK |
5960 | |
5961 | /** | |
5962 | * amdgpu_device_switch_gang - switch to a new gang | |
5963 | * @adev: amdgpu_device pointer | |
5964 | * @gang: the gang to switch to | |
5965 | * | |
5966 | * Try to switch to a new gang. | |
5967 | * Returns: NULL if we switched to the new gang or a reference to the current | |
5968 | * gang leader. | |
5969 | */ | |
5970 | struct dma_fence *amdgpu_device_switch_gang(struct amdgpu_device *adev, | |
5971 | struct dma_fence *gang) | |
5972 | { | |
5973 | struct dma_fence *old = NULL; | |
5974 | ||
5975 | do { | |
5976 | dma_fence_put(old); | |
5977 | rcu_read_lock(); | |
5978 | old = dma_fence_get_rcu_safe(&adev->gang_submit); | |
5979 | rcu_read_unlock(); | |
5980 | ||
5981 | if (old == gang) | |
5982 | break; | |
5983 | ||
5984 | if (!dma_fence_is_signaled(old)) | |
5985 | return old; | |
5986 | ||
5987 | } while (cmpxchg((struct dma_fence __force **)&adev->gang_submit, | |
5988 | old, gang) != old); | |
5989 | ||
5990 | dma_fence_put(old); | |
5991 | return NULL; | |
5992 | } | |
220c8cc8 AD |
5993 | |
5994 | bool amdgpu_device_has_display_hardware(struct amdgpu_device *adev) | |
5995 | { | |
5996 | switch (adev->asic_type) { | |
5997 | #ifdef CONFIG_DRM_AMDGPU_SI | |
5998 | case CHIP_HAINAN: | |
5999 | #endif | |
6000 | case CHIP_TOPAZ: | |
6001 | /* chips with no display hardware */ | |
6002 | return false; | |
6003 | #ifdef CONFIG_DRM_AMDGPU_SI | |
6004 | case CHIP_TAHITI: | |
6005 | case CHIP_PITCAIRN: | |
6006 | case CHIP_VERDE: | |
6007 | case CHIP_OLAND: | |
6008 | #endif | |
6009 | #ifdef CONFIG_DRM_AMDGPU_CIK | |
6010 | case CHIP_BONAIRE: | |
6011 | case CHIP_HAWAII: | |
6012 | case CHIP_KAVERI: | |
6013 | case CHIP_KABINI: | |
6014 | case CHIP_MULLINS: | |
6015 | #endif | |
6016 | case CHIP_TONGA: | |
6017 | case CHIP_FIJI: | |
6018 | case CHIP_POLARIS10: | |
6019 | case CHIP_POLARIS11: | |
6020 | case CHIP_POLARIS12: | |
6021 | case CHIP_VEGAM: | |
6022 | case CHIP_CARRIZO: | |
6023 | case CHIP_STONEY: | |
6024 | /* chips with display hardware */ | |
6025 | return true; | |
6026 | default: | |
6027 | /* IP discovery */ | |
6028 | if (!adev->ip_versions[DCE_HWIP][0] || | |
6029 | (adev->harvest_ip_mask & AMD_HARVEST_IP_DMU_MASK)) | |
6030 | return false; | |
6031 | return true; | |
6032 | } | |
6033 | } | |
81283fee JZ |
6034 | |
6035 | uint32_t amdgpu_device_wait_on_rreg(struct amdgpu_device *adev, | |
6036 | uint32_t inst, uint32_t reg_addr, char reg_name[], | |
6037 | uint32_t expected_value, uint32_t mask) | |
6038 | { | |
6039 | uint32_t ret = 0; | |
6040 | uint32_t old_ = 0; | |
6041 | uint32_t tmp_ = RREG32(reg_addr); | |
6042 | uint32_t loop = adev->usec_timeout; | |
6043 | ||
6044 | while ((tmp_ & (mask)) != (expected_value)) { | |
6045 | if (old_ != tmp_) { | |
6046 | loop = adev->usec_timeout; | |
6047 | old_ = tmp_; | |
6048 | } else | |
6049 | udelay(1); | |
6050 | tmp_ = RREG32(reg_addr); | |
6051 | loop--; | |
6052 | if (!loop) { | |
6053 | DRM_WARN("Register(%d) [%s] failed to reach value 0x%08x != 0x%08xn", | |
6054 | inst, reg_name, (uint32_t)expected_value, | |
6055 | (uint32_t)(tmp_ & (mask))); | |
6056 | ret = -ETIMEDOUT; | |
6057 | break; | |
6058 | } | |
6059 | } | |
6060 | return ret; | |
6061 | } |