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 | ||
e2a75f88 | 84 | MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin"); |
3f76dced | 85 | MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin"); |
2d2e5e7e | 86 | MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin"); |
ad5a67a7 | 87 | MODULE_FIRMWARE("amdgpu/picasso_gpu_info.bin"); |
54c4d17e | 88 | MODULE_FIRMWARE("amdgpu/raven2_gpu_info.bin"); |
65e60f6e | 89 | MODULE_FIRMWARE("amdgpu/arcturus_gpu_info.bin"); |
42b325e5 | 90 | MODULE_FIRMWARE("amdgpu/navi12_gpu_info.bin"); |
e2a75f88 | 91 | |
2dc80b00 | 92 | #define AMDGPU_RESUME_MS 2000 |
7258fa31 SK |
93 | #define AMDGPU_MAX_RETRY_LIMIT 2 |
94 | #define AMDGPU_RETRY_SRIOV_RESET(r) ((r) == -EBUSY || (r) == -ETIMEDOUT || (r) == -EINVAL) | |
2dc80b00 | 95 | |
b7cdb41e ML |
96 | static const struct drm_driver amdgpu_kms_driver; |
97 | ||
050091ab | 98 | const char *amdgpu_asic_name[] = { |
da69c161 KW |
99 | "TAHITI", |
100 | "PITCAIRN", | |
101 | "VERDE", | |
102 | "OLAND", | |
103 | "HAINAN", | |
d38ceaf9 AD |
104 | "BONAIRE", |
105 | "KAVERI", | |
106 | "KABINI", | |
107 | "HAWAII", | |
108 | "MULLINS", | |
109 | "TOPAZ", | |
110 | "TONGA", | |
48299f95 | 111 | "FIJI", |
d38ceaf9 | 112 | "CARRIZO", |
139f4917 | 113 | "STONEY", |
2cc0c0b5 FC |
114 | "POLARIS10", |
115 | "POLARIS11", | |
c4642a47 | 116 | "POLARIS12", |
48ff108d | 117 | "VEGAM", |
d4196f01 | 118 | "VEGA10", |
8fab806a | 119 | "VEGA12", |
956fcddc | 120 | "VEGA20", |
2ca8a5d2 | 121 | "RAVEN", |
d6c3b24e | 122 | "ARCTURUS", |
1eee4228 | 123 | "RENOIR", |
d46b417a | 124 | "ALDEBARAN", |
852a6626 | 125 | "NAVI10", |
d0f56dc2 | 126 | "CYAN_SKILLFISH", |
87dbad02 | 127 | "NAVI14", |
9802f5d7 | 128 | "NAVI12", |
ccaf72d3 | 129 | "SIENNA_CICHLID", |
ddd8fbe7 | 130 | "NAVY_FLOUNDER", |
4f1e9a76 | 131 | "VANGOGH", |
a2468e04 | 132 | "DIMGREY_CAVEFISH", |
6f169591 | 133 | "BEIGE_GOBY", |
ee9236b7 | 134 | "YELLOW_CARP", |
3ae695d6 | 135 | "IP DISCOVERY", |
d38ceaf9 AD |
136 | "LAST", |
137 | }; | |
138 | ||
dcea6e65 KR |
139 | /** |
140 | * DOC: pcie_replay_count | |
141 | * | |
142 | * The amdgpu driver provides a sysfs API for reporting the total number | |
143 | * of PCIe replays (NAKs) | |
144 | * The file pcie_replay_count is used for this and returns the total | |
145 | * number of replays as a sum of the NAKs generated and NAKs received | |
146 | */ | |
147 | ||
148 | static ssize_t amdgpu_device_get_pcie_replay_count(struct device *dev, | |
149 | struct device_attribute *attr, char *buf) | |
150 | { | |
151 | struct drm_device *ddev = dev_get_drvdata(dev); | |
1348969a | 152 | struct amdgpu_device *adev = drm_to_adev(ddev); |
dcea6e65 KR |
153 | uint64_t cnt = amdgpu_asic_get_pcie_replay_count(adev); |
154 | ||
36000c7a | 155 | return sysfs_emit(buf, "%llu\n", cnt); |
dcea6e65 KR |
156 | } |
157 | ||
158 | static DEVICE_ATTR(pcie_replay_count, S_IRUGO, | |
159 | amdgpu_device_get_pcie_replay_count, NULL); | |
160 | ||
5494d864 AD |
161 | static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev); |
162 | ||
bd607166 KR |
163 | /** |
164 | * DOC: product_name | |
165 | * | |
166 | * The amdgpu driver provides a sysfs API for reporting the product name | |
167 | * for the device | |
2c496a6c | 168 | * The file product_name is used for this and returns the product name |
bd607166 KR |
169 | * as returned from the FRU. |
170 | * NOTE: This is only available for certain server cards | |
171 | */ | |
172 | ||
173 | static ssize_t amdgpu_device_get_product_name(struct device *dev, | |
174 | struct device_attribute *attr, char *buf) | |
175 | { | |
176 | struct drm_device *ddev = dev_get_drvdata(dev); | |
1348969a | 177 | struct amdgpu_device *adev = drm_to_adev(ddev); |
bd607166 | 178 | |
36000c7a | 179 | return sysfs_emit(buf, "%s\n", adev->product_name); |
bd607166 KR |
180 | } |
181 | ||
182 | static DEVICE_ATTR(product_name, S_IRUGO, | |
183 | amdgpu_device_get_product_name, NULL); | |
184 | ||
185 | /** | |
186 | * DOC: product_number | |
187 | * | |
188 | * The amdgpu driver provides a sysfs API for reporting the part number | |
189 | * for the device | |
2c496a6c | 190 | * The file product_number is used for this and returns the part number |
bd607166 KR |
191 | * as returned from the FRU. |
192 | * NOTE: This is only available for certain server cards | |
193 | */ | |
194 | ||
195 | static ssize_t amdgpu_device_get_product_number(struct device *dev, | |
196 | struct device_attribute *attr, char *buf) | |
197 | { | |
198 | struct drm_device *ddev = dev_get_drvdata(dev); | |
1348969a | 199 | struct amdgpu_device *adev = drm_to_adev(ddev); |
bd607166 | 200 | |
36000c7a | 201 | return sysfs_emit(buf, "%s\n", adev->product_number); |
bd607166 KR |
202 | } |
203 | ||
204 | static DEVICE_ATTR(product_number, S_IRUGO, | |
205 | amdgpu_device_get_product_number, NULL); | |
206 | ||
207 | /** | |
208 | * DOC: serial_number | |
209 | * | |
210 | * The amdgpu driver provides a sysfs API for reporting the serial number | |
211 | * for the device | |
212 | * The file serial_number is used for this and returns the serial number | |
213 | * as returned from the FRU. | |
214 | * NOTE: This is only available for certain server cards | |
215 | */ | |
216 | ||
217 | static ssize_t amdgpu_device_get_serial_number(struct device *dev, | |
218 | struct device_attribute *attr, char *buf) | |
219 | { | |
220 | struct drm_device *ddev = dev_get_drvdata(dev); | |
1348969a | 221 | struct amdgpu_device *adev = drm_to_adev(ddev); |
bd607166 | 222 | |
36000c7a | 223 | return sysfs_emit(buf, "%s\n", adev->serial); |
bd607166 KR |
224 | } |
225 | ||
226 | static DEVICE_ATTR(serial_number, S_IRUGO, | |
227 | amdgpu_device_get_serial_number, NULL); | |
228 | ||
fd496ca8 | 229 | /** |
b98c6299 | 230 | * amdgpu_device_supports_px - Is the device a dGPU with ATPX power control |
fd496ca8 AD |
231 | * |
232 | * @dev: drm_device pointer | |
233 | * | |
b98c6299 | 234 | * Returns true if the device is a dGPU with ATPX power control, |
fd496ca8 AD |
235 | * otherwise return false. |
236 | */ | |
b98c6299 | 237 | bool amdgpu_device_supports_px(struct drm_device *dev) |
fd496ca8 AD |
238 | { |
239 | struct amdgpu_device *adev = drm_to_adev(dev); | |
240 | ||
b98c6299 | 241 | if ((adev->flags & AMD_IS_PX) && !amdgpu_is_atpx_hybrid()) |
fd496ca8 AD |
242 | return true; |
243 | return false; | |
244 | } | |
245 | ||
e3ecdffa | 246 | /** |
0330b848 | 247 | * amdgpu_device_supports_boco - Is the device a dGPU with ACPI power resources |
e3ecdffa AD |
248 | * |
249 | * @dev: drm_device pointer | |
250 | * | |
b98c6299 | 251 | * Returns true if the device is a dGPU with ACPI power control, |
e3ecdffa AD |
252 | * otherwise return false. |
253 | */ | |
31af062a | 254 | bool amdgpu_device_supports_boco(struct drm_device *dev) |
d38ceaf9 | 255 | { |
1348969a | 256 | struct amdgpu_device *adev = drm_to_adev(dev); |
d38ceaf9 | 257 | |
b98c6299 AD |
258 | if (adev->has_pr3 || |
259 | ((adev->flags & AMD_IS_PX) && amdgpu_is_atpx_hybrid())) | |
d38ceaf9 AD |
260 | return true; |
261 | return false; | |
262 | } | |
263 | ||
a69cba42 AD |
264 | /** |
265 | * amdgpu_device_supports_baco - Does the device support BACO | |
266 | * | |
267 | * @dev: drm_device pointer | |
268 | * | |
269 | * Returns true if the device supporte BACO, | |
270 | * otherwise return false. | |
271 | */ | |
272 | bool amdgpu_device_supports_baco(struct drm_device *dev) | |
273 | { | |
1348969a | 274 | struct amdgpu_device *adev = drm_to_adev(dev); |
a69cba42 AD |
275 | |
276 | return amdgpu_asic_supports_baco(adev); | |
277 | } | |
278 | ||
3fa8f89d S |
279 | /** |
280 | * amdgpu_device_supports_smart_shift - Is the device dGPU with | |
281 | * smart shift support | |
282 | * | |
283 | * @dev: drm_device pointer | |
284 | * | |
285 | * Returns true if the device is a dGPU with Smart Shift support, | |
286 | * otherwise returns false. | |
287 | */ | |
288 | bool amdgpu_device_supports_smart_shift(struct drm_device *dev) | |
289 | { | |
290 | return (amdgpu_device_supports_boco(dev) && | |
291 | amdgpu_acpi_is_power_shift_control_supported()); | |
292 | } | |
293 | ||
6e3cd2a9 MCC |
294 | /* |
295 | * VRAM access helper functions | |
296 | */ | |
297 | ||
e35e2b11 | 298 | /** |
048af66b | 299 | * amdgpu_device_mm_access - access vram by MM_INDEX/MM_DATA |
e35e2b11 TY |
300 | * |
301 | * @adev: amdgpu_device pointer | |
302 | * @pos: offset of the buffer in vram | |
303 | * @buf: virtual address of the buffer in system memory | |
304 | * @size: read/write size, sizeof(@buf) must > @size | |
305 | * @write: true - write to vram, otherwise - read from vram | |
306 | */ | |
048af66b KW |
307 | void amdgpu_device_mm_access(struct amdgpu_device *adev, loff_t pos, |
308 | void *buf, size_t size, bool write) | |
e35e2b11 | 309 | { |
e35e2b11 | 310 | unsigned long flags; |
048af66b KW |
311 | uint32_t hi = ~0, tmp = 0; |
312 | uint32_t *data = buf; | |
ce05ac56 | 313 | uint64_t last; |
f89f8c6b | 314 | int idx; |
ce05ac56 | 315 | |
c58a863b | 316 | if (!drm_dev_enter(adev_to_drm(adev), &idx)) |
f89f8c6b | 317 | return; |
9d11eb0d | 318 | |
048af66b KW |
319 | BUG_ON(!IS_ALIGNED(pos, 4) || !IS_ALIGNED(size, 4)); |
320 | ||
321 | spin_lock_irqsave(&adev->mmio_idx_lock, flags); | |
322 | for (last = pos + size; pos < last; pos += 4) { | |
323 | tmp = pos >> 31; | |
324 | ||
325 | WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)pos) | 0x80000000); | |
326 | if (tmp != hi) { | |
327 | WREG32_NO_KIQ(mmMM_INDEX_HI, tmp); | |
328 | hi = tmp; | |
329 | } | |
330 | if (write) | |
331 | WREG32_NO_KIQ(mmMM_DATA, *data++); | |
332 | else | |
333 | *data++ = RREG32_NO_KIQ(mmMM_DATA); | |
334 | } | |
335 | ||
336 | spin_unlock_irqrestore(&adev->mmio_idx_lock, flags); | |
337 | drm_dev_exit(idx); | |
338 | } | |
339 | ||
340 | /** | |
bbe04dec | 341 | * amdgpu_device_aper_access - access vram by vram aperature |
048af66b KW |
342 | * |
343 | * @adev: amdgpu_device pointer | |
344 | * @pos: offset of the buffer in vram | |
345 | * @buf: virtual address of the buffer in system memory | |
346 | * @size: read/write size, sizeof(@buf) must > @size | |
347 | * @write: true - write to vram, otherwise - read from vram | |
348 | * | |
349 | * The return value means how many bytes have been transferred. | |
350 | */ | |
351 | size_t amdgpu_device_aper_access(struct amdgpu_device *adev, loff_t pos, | |
352 | void *buf, size_t size, bool write) | |
353 | { | |
9d11eb0d | 354 | #ifdef CONFIG_64BIT |
048af66b KW |
355 | void __iomem *addr; |
356 | size_t count = 0; | |
357 | uint64_t last; | |
358 | ||
359 | if (!adev->mman.aper_base_kaddr) | |
360 | return 0; | |
361 | ||
9d11eb0d CK |
362 | last = min(pos + size, adev->gmc.visible_vram_size); |
363 | if (last > pos) { | |
048af66b KW |
364 | addr = adev->mman.aper_base_kaddr + pos; |
365 | count = last - pos; | |
9d11eb0d CK |
366 | |
367 | if (write) { | |
368 | memcpy_toio(addr, buf, count); | |
369 | mb(); | |
810085dd | 370 | amdgpu_device_flush_hdp(adev, NULL); |
9d11eb0d | 371 | } else { |
810085dd | 372 | amdgpu_device_invalidate_hdp(adev, NULL); |
9d11eb0d CK |
373 | mb(); |
374 | memcpy_fromio(buf, addr, count); | |
375 | } | |
376 | ||
9d11eb0d | 377 | } |
048af66b KW |
378 | |
379 | return count; | |
380 | #else | |
381 | return 0; | |
9d11eb0d | 382 | #endif |
048af66b | 383 | } |
9d11eb0d | 384 | |
048af66b KW |
385 | /** |
386 | * amdgpu_device_vram_access - read/write a buffer in vram | |
387 | * | |
388 | * @adev: amdgpu_device pointer | |
389 | * @pos: offset of the buffer in vram | |
390 | * @buf: virtual address of the buffer in system memory | |
391 | * @size: read/write size, sizeof(@buf) must > @size | |
392 | * @write: true - write to vram, otherwise - read from vram | |
393 | */ | |
394 | void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos, | |
395 | void *buf, size_t size, bool write) | |
396 | { | |
397 | size_t count; | |
e35e2b11 | 398 | |
048af66b KW |
399 | /* try to using vram apreature to access vram first */ |
400 | count = amdgpu_device_aper_access(adev, pos, buf, size, write); | |
401 | size -= count; | |
402 | if (size) { | |
403 | /* using MM to access rest vram */ | |
404 | pos += count; | |
405 | buf += count; | |
406 | amdgpu_device_mm_access(adev, pos, buf, size, write); | |
e35e2b11 TY |
407 | } |
408 | } | |
409 | ||
d38ceaf9 | 410 | /* |
f7ee1874 | 411 | * register access helper functions. |
d38ceaf9 | 412 | */ |
56b53c0b DL |
413 | |
414 | /* Check if hw access should be skipped because of hotplug or device error */ | |
415 | bool amdgpu_device_skip_hw_access(struct amdgpu_device *adev) | |
416 | { | |
7afefb81 | 417 | if (adev->no_hw_access) |
56b53c0b DL |
418 | return true; |
419 | ||
420 | #ifdef CONFIG_LOCKDEP | |
421 | /* | |
422 | * This is a bit complicated to understand, so worth a comment. What we assert | |
423 | * here is that the GPU reset is not running on another thread in parallel. | |
424 | * | |
425 | * For this we trylock the read side of the reset semaphore, if that succeeds | |
426 | * we know that the reset is not running in paralell. | |
427 | * | |
428 | * If the trylock fails we assert that we are either already holding the read | |
429 | * side of the lock or are the reset thread itself and hold the write side of | |
430 | * the lock. | |
431 | */ | |
432 | if (in_task()) { | |
d0fb18b5 AG |
433 | if (down_read_trylock(&adev->reset_domain->sem)) |
434 | up_read(&adev->reset_domain->sem); | |
56b53c0b | 435 | else |
d0fb18b5 | 436 | lockdep_assert_held(&adev->reset_domain->sem); |
56b53c0b DL |
437 | } |
438 | #endif | |
439 | return false; | |
440 | } | |
441 | ||
e3ecdffa | 442 | /** |
f7ee1874 | 443 | * amdgpu_device_rreg - read a memory mapped IO or indirect register |
e3ecdffa AD |
444 | * |
445 | * @adev: amdgpu_device pointer | |
446 | * @reg: dword aligned register offset | |
447 | * @acc_flags: access flags which require special behavior | |
448 | * | |
449 | * Returns the 32 bit value from the offset specified. | |
450 | */ | |
f7ee1874 HZ |
451 | uint32_t amdgpu_device_rreg(struct amdgpu_device *adev, |
452 | uint32_t reg, uint32_t acc_flags) | |
d38ceaf9 | 453 | { |
f4b373f4 TSD |
454 | uint32_t ret; |
455 | ||
56b53c0b | 456 | if (amdgpu_device_skip_hw_access(adev)) |
bf36b52e AG |
457 | return 0; |
458 | ||
f7ee1874 HZ |
459 | if ((reg * 4) < adev->rmmio_size) { |
460 | if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && | |
461 | amdgpu_sriov_runtime(adev) && | |
d0fb18b5 | 462 | down_read_trylock(&adev->reset_domain->sem)) { |
f7ee1874 | 463 | ret = amdgpu_kiq_rreg(adev, reg); |
d0fb18b5 | 464 | up_read(&adev->reset_domain->sem); |
f7ee1874 HZ |
465 | } else { |
466 | ret = readl(((void __iomem *)adev->rmmio) + (reg * 4)); | |
467 | } | |
468 | } else { | |
469 | ret = adev->pcie_rreg(adev, reg * 4); | |
81202807 | 470 | } |
bc992ba5 | 471 | |
f7ee1874 | 472 | trace_amdgpu_device_rreg(adev->pdev->device, reg, ret); |
e78b579d | 473 | |
f4b373f4 | 474 | return ret; |
d38ceaf9 AD |
475 | } |
476 | ||
421a2a30 ML |
477 | /* |
478 | * MMIO register read with bytes helper functions | |
479 | * @offset:bytes offset from MMIO start | |
480 | * | |
481 | */ | |
482 | ||
e3ecdffa AD |
483 | /** |
484 | * amdgpu_mm_rreg8 - read a memory mapped IO register | |
485 | * | |
486 | * @adev: amdgpu_device pointer | |
487 | * @offset: byte aligned register offset | |
488 | * | |
489 | * Returns the 8 bit value from the offset specified. | |
490 | */ | |
7cbbc745 AG |
491 | uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset) |
492 | { | |
56b53c0b | 493 | if (amdgpu_device_skip_hw_access(adev)) |
bf36b52e AG |
494 | return 0; |
495 | ||
421a2a30 ML |
496 | if (offset < adev->rmmio_size) |
497 | return (readb(adev->rmmio + offset)); | |
498 | BUG(); | |
499 | } | |
500 | ||
501 | /* | |
502 | * MMIO register write with bytes helper functions | |
503 | * @offset:bytes offset from MMIO start | |
504 | * @value: the value want to be written to the register | |
505 | * | |
506 | */ | |
e3ecdffa AD |
507 | /** |
508 | * amdgpu_mm_wreg8 - read a memory mapped IO register | |
509 | * | |
510 | * @adev: amdgpu_device pointer | |
511 | * @offset: byte aligned register offset | |
512 | * @value: 8 bit value to write | |
513 | * | |
514 | * Writes the value specified to the offset specified. | |
515 | */ | |
7cbbc745 AG |
516 | void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value) |
517 | { | |
56b53c0b | 518 | if (amdgpu_device_skip_hw_access(adev)) |
bf36b52e AG |
519 | return; |
520 | ||
421a2a30 ML |
521 | if (offset < adev->rmmio_size) |
522 | writeb(value, adev->rmmio + offset); | |
523 | else | |
524 | BUG(); | |
525 | } | |
526 | ||
e3ecdffa | 527 | /** |
f7ee1874 | 528 | * amdgpu_device_wreg - write to a memory mapped IO or indirect register |
e3ecdffa AD |
529 | * |
530 | * @adev: amdgpu_device pointer | |
531 | * @reg: dword aligned register offset | |
532 | * @v: 32 bit value to write to the register | |
533 | * @acc_flags: access flags which require special behavior | |
534 | * | |
535 | * Writes the value specified to the offset specified. | |
536 | */ | |
f7ee1874 HZ |
537 | void amdgpu_device_wreg(struct amdgpu_device *adev, |
538 | uint32_t reg, uint32_t v, | |
539 | uint32_t acc_flags) | |
d38ceaf9 | 540 | { |
56b53c0b | 541 | if (amdgpu_device_skip_hw_access(adev)) |
bf36b52e AG |
542 | return; |
543 | ||
f7ee1874 HZ |
544 | if ((reg * 4) < adev->rmmio_size) { |
545 | if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && | |
546 | amdgpu_sriov_runtime(adev) && | |
d0fb18b5 | 547 | down_read_trylock(&adev->reset_domain->sem)) { |
f7ee1874 | 548 | amdgpu_kiq_wreg(adev, reg, v); |
d0fb18b5 | 549 | up_read(&adev->reset_domain->sem); |
f7ee1874 HZ |
550 | } else { |
551 | writel(v, ((void __iomem *)adev->rmmio) + (reg * 4)); | |
552 | } | |
553 | } else { | |
554 | adev->pcie_wreg(adev, reg * 4, v); | |
81202807 | 555 | } |
bc992ba5 | 556 | |
f7ee1874 | 557 | trace_amdgpu_device_wreg(adev->pdev->device, reg, v); |
2e0cc4d4 | 558 | } |
d38ceaf9 | 559 | |
03f2abb0 | 560 | /** |
4cc9f86f | 561 | * amdgpu_mm_wreg_mmio_rlc - write register either with direct/indirect mmio or with RLC path if in range |
2e0cc4d4 | 562 | * |
71579346 RB |
563 | * @adev: amdgpu_device pointer |
564 | * @reg: mmio/rlc register | |
565 | * @v: value to write | |
566 | * | |
567 | * this function is invoked only for the debugfs register access | |
03f2abb0 | 568 | */ |
f7ee1874 HZ |
569 | void amdgpu_mm_wreg_mmio_rlc(struct amdgpu_device *adev, |
570 | uint32_t reg, uint32_t v) | |
2e0cc4d4 | 571 | { |
56b53c0b | 572 | if (amdgpu_device_skip_hw_access(adev)) |
bf36b52e AG |
573 | return; |
574 | ||
2e0cc4d4 | 575 | if (amdgpu_sriov_fullaccess(adev) && |
f7ee1874 HZ |
576 | adev->gfx.rlc.funcs && |
577 | adev->gfx.rlc.funcs->is_rlcg_access_range) { | |
2e0cc4d4 | 578 | if (adev->gfx.rlc.funcs->is_rlcg_access_range(adev, reg)) |
1b2dc99e | 579 | return amdgpu_sriov_wreg(adev, reg, v, 0, 0); |
4cc9f86f TSD |
580 | } else if ((reg * 4) >= adev->rmmio_size) { |
581 | adev->pcie_wreg(adev, reg * 4, v); | |
f7ee1874 HZ |
582 | } else { |
583 | writel(v, ((void __iomem *)adev->rmmio) + (reg * 4)); | |
47ed4e1c | 584 | } |
d38ceaf9 AD |
585 | } |
586 | ||
d38ceaf9 AD |
587 | /** |
588 | * amdgpu_mm_rdoorbell - read a doorbell dword | |
589 | * | |
590 | * @adev: amdgpu_device pointer | |
591 | * @index: doorbell index | |
592 | * | |
593 | * Returns the value in the doorbell aperture at the | |
594 | * requested doorbell index (CIK). | |
595 | */ | |
596 | u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index) | |
597 | { | |
56b53c0b | 598 | if (amdgpu_device_skip_hw_access(adev)) |
bf36b52e AG |
599 | return 0; |
600 | ||
d38ceaf9 AD |
601 | if (index < adev->doorbell.num_doorbells) { |
602 | return readl(adev->doorbell.ptr + index); | |
603 | } else { | |
604 | DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index); | |
605 | return 0; | |
606 | } | |
607 | } | |
608 | ||
609 | /** | |
610 | * amdgpu_mm_wdoorbell - write a doorbell dword | |
611 | * | |
612 | * @adev: amdgpu_device pointer | |
613 | * @index: doorbell index | |
614 | * @v: value to write | |
615 | * | |
616 | * Writes @v to the doorbell aperture at the | |
617 | * requested doorbell index (CIK). | |
618 | */ | |
619 | void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v) | |
620 | { | |
56b53c0b | 621 | if (amdgpu_device_skip_hw_access(adev)) |
bf36b52e AG |
622 | return; |
623 | ||
d38ceaf9 AD |
624 | if (index < adev->doorbell.num_doorbells) { |
625 | writel(v, adev->doorbell.ptr + index); | |
626 | } else { | |
627 | DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index); | |
628 | } | |
629 | } | |
630 | ||
832be404 KW |
631 | /** |
632 | * amdgpu_mm_rdoorbell64 - read a doorbell Qword | |
633 | * | |
634 | * @adev: amdgpu_device pointer | |
635 | * @index: doorbell index | |
636 | * | |
637 | * Returns the value in the doorbell aperture at the | |
638 | * requested doorbell index (VEGA10+). | |
639 | */ | |
640 | u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index) | |
641 | { | |
56b53c0b | 642 | if (amdgpu_device_skip_hw_access(adev)) |
bf36b52e AG |
643 | return 0; |
644 | ||
832be404 KW |
645 | if (index < adev->doorbell.num_doorbells) { |
646 | return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index)); | |
647 | } else { | |
648 | DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index); | |
649 | return 0; | |
650 | } | |
651 | } | |
652 | ||
653 | /** | |
654 | * amdgpu_mm_wdoorbell64 - write a doorbell Qword | |
655 | * | |
656 | * @adev: amdgpu_device pointer | |
657 | * @index: doorbell index | |
658 | * @v: value to write | |
659 | * | |
660 | * Writes @v to the doorbell aperture at the | |
661 | * requested doorbell index (VEGA10+). | |
662 | */ | |
663 | void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v) | |
664 | { | |
56b53c0b | 665 | if (amdgpu_device_skip_hw_access(adev)) |
bf36b52e AG |
666 | return; |
667 | ||
832be404 KW |
668 | if (index < adev->doorbell.num_doorbells) { |
669 | atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v); | |
670 | } else { | |
671 | DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index); | |
672 | } | |
673 | } | |
674 | ||
1bba3683 HZ |
675 | /** |
676 | * amdgpu_device_indirect_rreg - read an indirect register | |
677 | * | |
678 | * @adev: amdgpu_device pointer | |
679 | * @pcie_index: mmio register offset | |
680 | * @pcie_data: mmio register offset | |
22f453fb | 681 | * @reg_addr: indirect register address to read from |
1bba3683 HZ |
682 | * |
683 | * Returns the value of indirect register @reg_addr | |
684 | */ | |
685 | u32 amdgpu_device_indirect_rreg(struct amdgpu_device *adev, | |
686 | u32 pcie_index, u32 pcie_data, | |
687 | u32 reg_addr) | |
688 | { | |
689 | unsigned long flags; | |
690 | u32 r; | |
691 | void __iomem *pcie_index_offset; | |
692 | void __iomem *pcie_data_offset; | |
693 | ||
694 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); | |
695 | pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4; | |
696 | pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4; | |
697 | ||
698 | writel(reg_addr, pcie_index_offset); | |
699 | readl(pcie_index_offset); | |
700 | r = readl(pcie_data_offset); | |
701 | spin_unlock_irqrestore(&adev->pcie_idx_lock, flags); | |
702 | ||
703 | return r; | |
704 | } | |
705 | ||
706 | /** | |
707 | * amdgpu_device_indirect_rreg64 - read a 64bits indirect register | |
708 | * | |
709 | * @adev: amdgpu_device pointer | |
710 | * @pcie_index: mmio register offset | |
711 | * @pcie_data: mmio register offset | |
22f453fb | 712 | * @reg_addr: indirect register address to read from |
1bba3683 HZ |
713 | * |
714 | * Returns the value of indirect register @reg_addr | |
715 | */ | |
716 | u64 amdgpu_device_indirect_rreg64(struct amdgpu_device *adev, | |
717 | u32 pcie_index, u32 pcie_data, | |
718 | u32 reg_addr) | |
719 | { | |
720 | unsigned long flags; | |
721 | u64 r; | |
722 | void __iomem *pcie_index_offset; | |
723 | void __iomem *pcie_data_offset; | |
724 | ||
725 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); | |
726 | pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4; | |
727 | pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4; | |
728 | ||
729 | /* read low 32 bits */ | |
730 | writel(reg_addr, pcie_index_offset); | |
731 | readl(pcie_index_offset); | |
732 | r = readl(pcie_data_offset); | |
733 | /* read high 32 bits */ | |
734 | writel(reg_addr + 4, pcie_index_offset); | |
735 | readl(pcie_index_offset); | |
736 | r |= ((u64)readl(pcie_data_offset) << 32); | |
737 | spin_unlock_irqrestore(&adev->pcie_idx_lock, flags); | |
738 | ||
739 | return r; | |
740 | } | |
741 | ||
742 | /** | |
743 | * amdgpu_device_indirect_wreg - write an indirect register address | |
744 | * | |
745 | * @adev: amdgpu_device pointer | |
746 | * @pcie_index: mmio register offset | |
747 | * @pcie_data: mmio register offset | |
748 | * @reg_addr: indirect register offset | |
749 | * @reg_data: indirect register data | |
750 | * | |
751 | */ | |
752 | void amdgpu_device_indirect_wreg(struct amdgpu_device *adev, | |
753 | u32 pcie_index, u32 pcie_data, | |
754 | u32 reg_addr, u32 reg_data) | |
755 | { | |
756 | unsigned long flags; | |
757 | void __iomem *pcie_index_offset; | |
758 | void __iomem *pcie_data_offset; | |
759 | ||
760 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); | |
761 | pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4; | |
762 | pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4; | |
763 | ||
764 | writel(reg_addr, pcie_index_offset); | |
765 | readl(pcie_index_offset); | |
766 | writel(reg_data, pcie_data_offset); | |
767 | readl(pcie_data_offset); | |
768 | spin_unlock_irqrestore(&adev->pcie_idx_lock, flags); | |
769 | } | |
770 | ||
771 | /** | |
772 | * amdgpu_device_indirect_wreg64 - write a 64bits indirect register address | |
773 | * | |
774 | * @adev: amdgpu_device pointer | |
775 | * @pcie_index: mmio register offset | |
776 | * @pcie_data: mmio register offset | |
777 | * @reg_addr: indirect register offset | |
778 | * @reg_data: indirect register data | |
779 | * | |
780 | */ | |
781 | void amdgpu_device_indirect_wreg64(struct amdgpu_device *adev, | |
782 | u32 pcie_index, u32 pcie_data, | |
783 | u32 reg_addr, u64 reg_data) | |
784 | { | |
785 | unsigned long flags; | |
786 | void __iomem *pcie_index_offset; | |
787 | void __iomem *pcie_data_offset; | |
788 | ||
789 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); | |
790 | pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4; | |
791 | pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4; | |
792 | ||
793 | /* write low 32 bits */ | |
794 | writel(reg_addr, pcie_index_offset); | |
795 | readl(pcie_index_offset); | |
796 | writel((u32)(reg_data & 0xffffffffULL), pcie_data_offset); | |
797 | readl(pcie_data_offset); | |
798 | /* write high 32 bits */ | |
799 | writel(reg_addr + 4, pcie_index_offset); | |
800 | readl(pcie_index_offset); | |
801 | writel((u32)(reg_data >> 32), pcie_data_offset); | |
802 | readl(pcie_data_offset); | |
803 | spin_unlock_irqrestore(&adev->pcie_idx_lock, flags); | |
804 | } | |
805 | ||
d38ceaf9 AD |
806 | /** |
807 | * amdgpu_invalid_rreg - dummy reg read function | |
808 | * | |
982a820b | 809 | * @adev: amdgpu_device pointer |
d38ceaf9 AD |
810 | * @reg: offset of register |
811 | * | |
812 | * Dummy register read function. Used for register blocks | |
813 | * that certain asics don't have (all asics). | |
814 | * Returns the value in the register. | |
815 | */ | |
816 | static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg) | |
817 | { | |
818 | DRM_ERROR("Invalid callback to read register 0x%04X\n", reg); | |
819 | BUG(); | |
820 | return 0; | |
821 | } | |
822 | ||
823 | /** | |
824 | * amdgpu_invalid_wreg - dummy reg write function | |
825 | * | |
982a820b | 826 | * @adev: amdgpu_device pointer |
d38ceaf9 AD |
827 | * @reg: offset of register |
828 | * @v: value to write to the register | |
829 | * | |
830 | * Dummy register read function. Used for register blocks | |
831 | * that certain asics don't have (all asics). | |
832 | */ | |
833 | static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v) | |
834 | { | |
835 | DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n", | |
836 | reg, v); | |
837 | BUG(); | |
838 | } | |
839 | ||
4fa1c6a6 TZ |
840 | /** |
841 | * amdgpu_invalid_rreg64 - dummy 64 bit reg read function | |
842 | * | |
982a820b | 843 | * @adev: amdgpu_device pointer |
4fa1c6a6 TZ |
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 uint64_t amdgpu_invalid_rreg64(struct amdgpu_device *adev, uint32_t reg) | |
851 | { | |
852 | DRM_ERROR("Invalid callback to read 64 bit register 0x%04X\n", reg); | |
853 | BUG(); | |
854 | return 0; | |
855 | } | |
856 | ||
857 | /** | |
858 | * amdgpu_invalid_wreg64 - dummy reg write function | |
859 | * | |
982a820b | 860 | * @adev: amdgpu_device pointer |
4fa1c6a6 TZ |
861 | * @reg: offset of register |
862 | * @v: value to write to the register | |
863 | * | |
864 | * Dummy register read function. Used for register blocks | |
865 | * that certain asics don't have (all asics). | |
866 | */ | |
867 | static void amdgpu_invalid_wreg64(struct amdgpu_device *adev, uint32_t reg, uint64_t v) | |
868 | { | |
869 | DRM_ERROR("Invalid callback to write 64 bit register 0x%04X with 0x%08llX\n", | |
870 | reg, v); | |
871 | BUG(); | |
872 | } | |
873 | ||
d38ceaf9 AD |
874 | /** |
875 | * amdgpu_block_invalid_rreg - dummy reg read function | |
876 | * | |
982a820b | 877 | * @adev: amdgpu_device pointer |
d38ceaf9 AD |
878 | * @block: offset of instance |
879 | * @reg: offset of register | |
880 | * | |
881 | * Dummy register read function. Used for register blocks | |
882 | * that certain asics don't have (all asics). | |
883 | * Returns the value in the register. | |
884 | */ | |
885 | static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev, | |
886 | uint32_t block, uint32_t reg) | |
887 | { | |
888 | DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n", | |
889 | reg, block); | |
890 | BUG(); | |
891 | return 0; | |
892 | } | |
893 | ||
894 | /** | |
895 | * amdgpu_block_invalid_wreg - dummy reg write function | |
896 | * | |
982a820b | 897 | * @adev: amdgpu_device pointer |
d38ceaf9 AD |
898 | * @block: offset of instance |
899 | * @reg: offset of register | |
900 | * @v: value to write to the register | |
901 | * | |
902 | * Dummy register read function. Used for register blocks | |
903 | * that certain asics don't have (all asics). | |
904 | */ | |
905 | static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev, | |
906 | uint32_t block, | |
907 | uint32_t reg, uint32_t v) | |
908 | { | |
909 | DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n", | |
910 | reg, block, v); | |
911 | BUG(); | |
912 | } | |
913 | ||
4d2997ab AD |
914 | /** |
915 | * amdgpu_device_asic_init - Wrapper for atom asic_init | |
916 | * | |
982a820b | 917 | * @adev: amdgpu_device pointer |
4d2997ab AD |
918 | * |
919 | * Does any asic specific work and then calls atom asic init. | |
920 | */ | |
921 | static int amdgpu_device_asic_init(struct amdgpu_device *adev) | |
922 | { | |
923 | amdgpu_asic_pre_asic_init(adev); | |
924 | ||
85d1bcc6 HZ |
925 | if (adev->ip_versions[GC_HWIP][0] >= IP_VERSION(11, 0, 0)) |
926 | return amdgpu_atomfirmware_asic_init(adev, true); | |
927 | else | |
928 | return amdgpu_atom_asic_init(adev->mode_info.atom_context); | |
4d2997ab AD |
929 | } |
930 | ||
e3ecdffa | 931 | /** |
7ccfd79f | 932 | * amdgpu_device_mem_scratch_init - allocate the VRAM scratch page |
e3ecdffa | 933 | * |
982a820b | 934 | * @adev: amdgpu_device pointer |
e3ecdffa AD |
935 | * |
936 | * Allocates a scratch page of VRAM for use by various things in the | |
937 | * driver. | |
938 | */ | |
7ccfd79f | 939 | static int amdgpu_device_mem_scratch_init(struct amdgpu_device *adev) |
d38ceaf9 | 940 | { |
7ccfd79f CK |
941 | return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE, PAGE_SIZE, |
942 | AMDGPU_GEM_DOMAIN_VRAM | | |
943 | AMDGPU_GEM_DOMAIN_GTT, | |
944 | &adev->mem_scratch.robj, | |
945 | &adev->mem_scratch.gpu_addr, | |
946 | (void **)&adev->mem_scratch.ptr); | |
d38ceaf9 AD |
947 | } |
948 | ||
e3ecdffa | 949 | /** |
7ccfd79f | 950 | * amdgpu_device_mem_scratch_fini - Free the VRAM scratch page |
e3ecdffa | 951 | * |
982a820b | 952 | * @adev: amdgpu_device pointer |
e3ecdffa AD |
953 | * |
954 | * Frees the VRAM scratch page. | |
955 | */ | |
7ccfd79f | 956 | static void amdgpu_device_mem_scratch_fini(struct amdgpu_device *adev) |
d38ceaf9 | 957 | { |
7ccfd79f | 958 | amdgpu_bo_free_kernel(&adev->mem_scratch.robj, NULL, NULL); |
d38ceaf9 AD |
959 | } |
960 | ||
961 | /** | |
9c3f2b54 | 962 | * amdgpu_device_program_register_sequence - program an array of registers. |
d38ceaf9 AD |
963 | * |
964 | * @adev: amdgpu_device pointer | |
965 | * @registers: pointer to the register array | |
966 | * @array_size: size of the register array | |
967 | * | |
968 | * Programs an array or registers with and and or masks. | |
969 | * This is a helper for setting golden registers. | |
970 | */ | |
9c3f2b54 AD |
971 | void amdgpu_device_program_register_sequence(struct amdgpu_device *adev, |
972 | const u32 *registers, | |
973 | const u32 array_size) | |
d38ceaf9 AD |
974 | { |
975 | u32 tmp, reg, and_mask, or_mask; | |
976 | int i; | |
977 | ||
978 | if (array_size % 3) | |
979 | return; | |
980 | ||
981 | for (i = 0; i < array_size; i +=3) { | |
982 | reg = registers[i + 0]; | |
983 | and_mask = registers[i + 1]; | |
984 | or_mask = registers[i + 2]; | |
985 | ||
986 | if (and_mask == 0xffffffff) { | |
987 | tmp = or_mask; | |
988 | } else { | |
989 | tmp = RREG32(reg); | |
990 | tmp &= ~and_mask; | |
e0d07657 HZ |
991 | if (adev->family >= AMDGPU_FAMILY_AI) |
992 | tmp |= (or_mask & and_mask); | |
993 | else | |
994 | tmp |= or_mask; | |
d38ceaf9 AD |
995 | } |
996 | WREG32(reg, tmp); | |
997 | } | |
998 | } | |
999 | ||
e3ecdffa AD |
1000 | /** |
1001 | * amdgpu_device_pci_config_reset - reset the GPU | |
1002 | * | |
1003 | * @adev: amdgpu_device pointer | |
1004 | * | |
1005 | * Resets the GPU using the pci config reset sequence. | |
1006 | * Only applicable to asics prior to vega10. | |
1007 | */ | |
8111c387 | 1008 | void amdgpu_device_pci_config_reset(struct amdgpu_device *adev) |
d38ceaf9 AD |
1009 | { |
1010 | pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA); | |
1011 | } | |
1012 | ||
af484df8 AD |
1013 | /** |
1014 | * amdgpu_device_pci_reset - reset the GPU using generic PCI means | |
1015 | * | |
1016 | * @adev: amdgpu_device pointer | |
1017 | * | |
1018 | * Resets the GPU using generic pci reset interfaces (FLR, SBR, etc.). | |
1019 | */ | |
1020 | int amdgpu_device_pci_reset(struct amdgpu_device *adev) | |
1021 | { | |
1022 | return pci_reset_function(adev->pdev); | |
1023 | } | |
1024 | ||
d38ceaf9 AD |
1025 | /* |
1026 | * GPU doorbell aperture helpers function. | |
1027 | */ | |
1028 | /** | |
06ec9070 | 1029 | * amdgpu_device_doorbell_init - Init doorbell driver information. |
d38ceaf9 AD |
1030 | * |
1031 | * @adev: amdgpu_device pointer | |
1032 | * | |
1033 | * Init doorbell driver information (CIK) | |
1034 | * Returns 0 on success, error on failure. | |
1035 | */ | |
06ec9070 | 1036 | static int amdgpu_device_doorbell_init(struct amdgpu_device *adev) |
d38ceaf9 | 1037 | { |
6585661d | 1038 | |
705e519e CK |
1039 | /* No doorbell on SI hardware generation */ |
1040 | if (adev->asic_type < CHIP_BONAIRE) { | |
1041 | adev->doorbell.base = 0; | |
1042 | adev->doorbell.size = 0; | |
1043 | adev->doorbell.num_doorbells = 0; | |
1044 | adev->doorbell.ptr = NULL; | |
1045 | return 0; | |
1046 | } | |
1047 | ||
d6895ad3 CK |
1048 | if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET) |
1049 | return -EINVAL; | |
1050 | ||
22357775 AD |
1051 | amdgpu_asic_init_doorbell_index(adev); |
1052 | ||
d38ceaf9 AD |
1053 | /* doorbell bar mapping */ |
1054 | adev->doorbell.base = pci_resource_start(adev->pdev, 2); | |
1055 | adev->doorbell.size = pci_resource_len(adev->pdev, 2); | |
1056 | ||
de33a329 JX |
1057 | if (adev->enable_mes) { |
1058 | adev->doorbell.num_doorbells = | |
1059 | adev->doorbell.size / sizeof(u32); | |
1060 | } else { | |
1061 | adev->doorbell.num_doorbells = | |
1062 | min_t(u32, adev->doorbell.size / sizeof(u32), | |
1063 | adev->doorbell_index.max_assignment+1); | |
1064 | if (adev->doorbell.num_doorbells == 0) | |
1065 | return -EINVAL; | |
1066 | ||
1067 | /* For Vega, reserve and map two pages on doorbell BAR since SDMA | |
1068 | * paging queue doorbell use the second page. The | |
1069 | * AMDGPU_DOORBELL64_MAX_ASSIGNMENT definition assumes all the | |
1070 | * doorbells are in the first page. So with paging queue enabled, | |
1071 | * the max num_doorbells should + 1 page (0x400 in dword) | |
1072 | */ | |
1073 | if (adev->asic_type >= CHIP_VEGA10) | |
1074 | adev->doorbell.num_doorbells += 0x400; | |
1075 | } | |
ec3db8a6 | 1076 | |
8972e5d2 CK |
1077 | adev->doorbell.ptr = ioremap(adev->doorbell.base, |
1078 | adev->doorbell.num_doorbells * | |
1079 | sizeof(u32)); | |
1080 | if (adev->doorbell.ptr == NULL) | |
d38ceaf9 | 1081 | return -ENOMEM; |
d38ceaf9 AD |
1082 | |
1083 | return 0; | |
1084 | } | |
1085 | ||
1086 | /** | |
06ec9070 | 1087 | * amdgpu_device_doorbell_fini - Tear down doorbell driver information. |
d38ceaf9 AD |
1088 | * |
1089 | * @adev: amdgpu_device pointer | |
1090 | * | |
1091 | * Tear down doorbell driver information (CIK) | |
1092 | */ | |
06ec9070 | 1093 | static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev) |
d38ceaf9 AD |
1094 | { |
1095 | iounmap(adev->doorbell.ptr); | |
1096 | adev->doorbell.ptr = NULL; | |
1097 | } | |
1098 | ||
22cb0164 | 1099 | |
d38ceaf9 AD |
1100 | |
1101 | /* | |
06ec9070 | 1102 | * amdgpu_device_wb_*() |
455a7bc2 | 1103 | * Writeback is the method by which the GPU updates special pages in memory |
ea81a173 | 1104 | * with the status of certain GPU events (fences, ring pointers,etc.). |
d38ceaf9 AD |
1105 | */ |
1106 | ||
1107 | /** | |
06ec9070 | 1108 | * amdgpu_device_wb_fini - Disable Writeback and free memory |
d38ceaf9 AD |
1109 | * |
1110 | * @adev: amdgpu_device pointer | |
1111 | * | |
1112 | * Disables Writeback and frees the Writeback memory (all asics). | |
1113 | * Used at driver shutdown. | |
1114 | */ | |
06ec9070 | 1115 | static void amdgpu_device_wb_fini(struct amdgpu_device *adev) |
d38ceaf9 AD |
1116 | { |
1117 | if (adev->wb.wb_obj) { | |
a76ed485 AD |
1118 | amdgpu_bo_free_kernel(&adev->wb.wb_obj, |
1119 | &adev->wb.gpu_addr, | |
1120 | (void **)&adev->wb.wb); | |
d38ceaf9 AD |
1121 | adev->wb.wb_obj = NULL; |
1122 | } | |
1123 | } | |
1124 | ||
1125 | /** | |
03f2abb0 | 1126 | * amdgpu_device_wb_init - Init Writeback driver info and allocate memory |
d38ceaf9 AD |
1127 | * |
1128 | * @adev: amdgpu_device pointer | |
1129 | * | |
455a7bc2 | 1130 | * Initializes writeback and allocates writeback memory (all asics). |
d38ceaf9 AD |
1131 | * Used at driver startup. |
1132 | * Returns 0 on success or an -error on failure. | |
1133 | */ | |
06ec9070 | 1134 | static int amdgpu_device_wb_init(struct amdgpu_device *adev) |
d38ceaf9 AD |
1135 | { |
1136 | int r; | |
1137 | ||
1138 | if (adev->wb.wb_obj == NULL) { | |
97407b63 AD |
1139 | /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */ |
1140 | r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8, | |
a76ed485 AD |
1141 | PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT, |
1142 | &adev->wb.wb_obj, &adev->wb.gpu_addr, | |
1143 | (void **)&adev->wb.wb); | |
d38ceaf9 AD |
1144 | if (r) { |
1145 | dev_warn(adev->dev, "(%d) create WB bo failed\n", r); | |
1146 | return r; | |
1147 | } | |
d38ceaf9 AD |
1148 | |
1149 | adev->wb.num_wb = AMDGPU_MAX_WB; | |
1150 | memset(&adev->wb.used, 0, sizeof(adev->wb.used)); | |
1151 | ||
1152 | /* clear wb memory */ | |
73469585 | 1153 | memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8); |
d38ceaf9 AD |
1154 | } |
1155 | ||
1156 | return 0; | |
1157 | } | |
1158 | ||
1159 | /** | |
131b4b36 | 1160 | * amdgpu_device_wb_get - Allocate a wb entry |
d38ceaf9 AD |
1161 | * |
1162 | * @adev: amdgpu_device pointer | |
1163 | * @wb: wb index | |
1164 | * | |
1165 | * Allocate a wb slot for use by the driver (all asics). | |
1166 | * Returns 0 on success or -EINVAL on failure. | |
1167 | */ | |
131b4b36 | 1168 | int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb) |
d38ceaf9 AD |
1169 | { |
1170 | unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb); | |
d38ceaf9 | 1171 | |
97407b63 | 1172 | if (offset < adev->wb.num_wb) { |
7014285a | 1173 | __set_bit(offset, adev->wb.used); |
63ae07ca | 1174 | *wb = offset << 3; /* convert to dw offset */ |
0915fdbc ML |
1175 | return 0; |
1176 | } else { | |
1177 | return -EINVAL; | |
1178 | } | |
1179 | } | |
1180 | ||
d38ceaf9 | 1181 | /** |
131b4b36 | 1182 | * amdgpu_device_wb_free - Free a wb entry |
d38ceaf9 AD |
1183 | * |
1184 | * @adev: amdgpu_device pointer | |
1185 | * @wb: wb index | |
1186 | * | |
1187 | * Free a wb slot allocated for use by the driver (all asics) | |
1188 | */ | |
131b4b36 | 1189 | void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb) |
d38ceaf9 | 1190 | { |
73469585 | 1191 | wb >>= 3; |
d38ceaf9 | 1192 | if (wb < adev->wb.num_wb) |
73469585 | 1193 | __clear_bit(wb, adev->wb.used); |
d38ceaf9 AD |
1194 | } |
1195 | ||
d6895ad3 CK |
1196 | /** |
1197 | * amdgpu_device_resize_fb_bar - try to resize FB BAR | |
1198 | * | |
1199 | * @adev: amdgpu_device pointer | |
1200 | * | |
1201 | * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not | |
1202 | * to fail, but if any of the BARs is not accessible after the size we abort | |
1203 | * driver loading by returning -ENODEV. | |
1204 | */ | |
1205 | int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev) | |
1206 | { | |
453f617a | 1207 | int rbar_size = pci_rebar_bytes_to_size(adev->gmc.real_vram_size); |
31b8adab CK |
1208 | struct pci_bus *root; |
1209 | struct resource *res; | |
1210 | unsigned i; | |
d6895ad3 CK |
1211 | u16 cmd; |
1212 | int r; | |
1213 | ||
0c03b912 | 1214 | /* Bypass for VF */ |
1215 | if (amdgpu_sriov_vf(adev)) | |
1216 | return 0; | |
1217 | ||
b7221f2b AD |
1218 | /* skip if the bios has already enabled large BAR */ |
1219 | if (adev->gmc.real_vram_size && | |
1220 | (pci_resource_len(adev->pdev, 0) >= adev->gmc.real_vram_size)) | |
1221 | return 0; | |
1222 | ||
31b8adab CK |
1223 | /* Check if the root BUS has 64bit memory resources */ |
1224 | root = adev->pdev->bus; | |
1225 | while (root->parent) | |
1226 | root = root->parent; | |
1227 | ||
1228 | pci_bus_for_each_resource(root, res, i) { | |
0ebb7c54 | 1229 | if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) && |
31b8adab CK |
1230 | res->start > 0x100000000ull) |
1231 | break; | |
1232 | } | |
1233 | ||
1234 | /* Trying to resize is pointless without a root hub window above 4GB */ | |
1235 | if (!res) | |
1236 | return 0; | |
1237 | ||
453f617a ND |
1238 | /* Limit the BAR size to what is available */ |
1239 | rbar_size = min(fls(pci_rebar_get_possible_sizes(adev->pdev, 0)) - 1, | |
1240 | rbar_size); | |
1241 | ||
d6895ad3 CK |
1242 | /* Disable memory decoding while we change the BAR addresses and size */ |
1243 | pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd); | |
1244 | pci_write_config_word(adev->pdev, PCI_COMMAND, | |
1245 | cmd & ~PCI_COMMAND_MEMORY); | |
1246 | ||
1247 | /* Free the VRAM and doorbell BAR, we most likely need to move both. */ | |
06ec9070 | 1248 | amdgpu_device_doorbell_fini(adev); |
d6895ad3 CK |
1249 | if (adev->asic_type >= CHIP_BONAIRE) |
1250 | pci_release_resource(adev->pdev, 2); | |
1251 | ||
1252 | pci_release_resource(adev->pdev, 0); | |
1253 | ||
1254 | r = pci_resize_resource(adev->pdev, 0, rbar_size); | |
1255 | if (r == -ENOSPC) | |
1256 | DRM_INFO("Not enough PCI address space for a large BAR."); | |
1257 | else if (r && r != -ENOTSUPP) | |
1258 | DRM_ERROR("Problem resizing BAR0 (%d).", r); | |
1259 | ||
1260 | pci_assign_unassigned_bus_resources(adev->pdev->bus); | |
1261 | ||
1262 | /* When the doorbell or fb BAR isn't available we have no chance of | |
1263 | * using the device. | |
1264 | */ | |
06ec9070 | 1265 | r = amdgpu_device_doorbell_init(adev); |
d6895ad3 CK |
1266 | if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET)) |
1267 | return -ENODEV; | |
1268 | ||
1269 | pci_write_config_word(adev->pdev, PCI_COMMAND, cmd); | |
1270 | ||
1271 | return 0; | |
1272 | } | |
a05502e5 | 1273 | |
d38ceaf9 AD |
1274 | /* |
1275 | * GPU helpers function. | |
1276 | */ | |
1277 | /** | |
39c640c0 | 1278 | * amdgpu_device_need_post - check if the hw need post or not |
d38ceaf9 AD |
1279 | * |
1280 | * @adev: amdgpu_device pointer | |
1281 | * | |
c836fec5 JQ |
1282 | * Check if the asic has been initialized (all asics) at driver startup |
1283 | * or post is needed if hw reset is performed. | |
1284 | * Returns true if need or false if not. | |
d38ceaf9 | 1285 | */ |
39c640c0 | 1286 | bool amdgpu_device_need_post(struct amdgpu_device *adev) |
d38ceaf9 AD |
1287 | { |
1288 | uint32_t reg; | |
1289 | ||
bec86378 ML |
1290 | if (amdgpu_sriov_vf(adev)) |
1291 | return false; | |
1292 | ||
1293 | if (amdgpu_passthrough(adev)) { | |
1da2c326 ML |
1294 | /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot |
1295 | * some old smc fw still need driver do vPost otherwise gpu hang, while | |
1296 | * those smc fw version above 22.15 doesn't have this flaw, so we force | |
1297 | * vpost executed for smc version below 22.15 | |
bec86378 ML |
1298 | */ |
1299 | if (adev->asic_type == CHIP_FIJI) { | |
1300 | int err; | |
1301 | uint32_t fw_ver; | |
1302 | err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev); | |
1303 | /* force vPost if error occured */ | |
1304 | if (err) | |
1305 | return true; | |
1306 | ||
1307 | fw_ver = *((uint32_t *)adev->pm.fw->data + 69); | |
1da2c326 ML |
1308 | if (fw_ver < 0x00160e00) |
1309 | return true; | |
bec86378 | 1310 | } |
bec86378 | 1311 | } |
91fe77eb | 1312 | |
e3c1b071 | 1313 | /* Don't post if we need to reset whole hive on init */ |
1314 | if (adev->gmc.xgmi.pending_reset) | |
1315 | return false; | |
1316 | ||
91fe77eb | 1317 | if (adev->has_hw_reset) { |
1318 | adev->has_hw_reset = false; | |
1319 | return true; | |
1320 | } | |
1321 | ||
1322 | /* bios scratch used on CIK+ */ | |
1323 | if (adev->asic_type >= CHIP_BONAIRE) | |
1324 | return amdgpu_atombios_scratch_need_asic_init(adev); | |
1325 | ||
1326 | /* check MEM_SIZE for older asics */ | |
1327 | reg = amdgpu_asic_get_config_memsize(adev); | |
1328 | ||
1329 | if ((reg != 0) && (reg != 0xffffffff)) | |
1330 | return false; | |
1331 | ||
1332 | return true; | |
bec86378 ML |
1333 | } |
1334 | ||
0ab5d711 ML |
1335 | /** |
1336 | * amdgpu_device_should_use_aspm - check if the device should program ASPM | |
1337 | * | |
1338 | * @adev: amdgpu_device pointer | |
1339 | * | |
1340 | * Confirm whether the module parameter and pcie bridge agree that ASPM should | |
1341 | * be set for this device. | |
1342 | * | |
1343 | * Returns true if it should be used or false if not. | |
1344 | */ | |
1345 | bool amdgpu_device_should_use_aspm(struct amdgpu_device *adev) | |
1346 | { | |
1347 | switch (amdgpu_aspm) { | |
1348 | case -1: | |
1349 | break; | |
1350 | case 0: | |
1351 | return false; | |
1352 | case 1: | |
1353 | return true; | |
1354 | default: | |
1355 | return false; | |
1356 | } | |
1357 | return pcie_aspm_enabled(adev->pdev); | |
1358 | } | |
1359 | ||
d38ceaf9 AD |
1360 | /* if we get transitioned to only one device, take VGA back */ |
1361 | /** | |
06ec9070 | 1362 | * amdgpu_device_vga_set_decode - enable/disable vga decode |
d38ceaf9 | 1363 | * |
bf44e8ce | 1364 | * @pdev: PCI device pointer |
d38ceaf9 AD |
1365 | * @state: enable/disable vga decode |
1366 | * | |
1367 | * Enable/disable vga decode (all asics). | |
1368 | * Returns VGA resource flags. | |
1369 | */ | |
bf44e8ce CH |
1370 | static unsigned int amdgpu_device_vga_set_decode(struct pci_dev *pdev, |
1371 | bool state) | |
d38ceaf9 | 1372 | { |
bf44e8ce | 1373 | struct amdgpu_device *adev = drm_to_adev(pci_get_drvdata(pdev)); |
d38ceaf9 AD |
1374 | amdgpu_asic_set_vga_state(adev, state); |
1375 | if (state) | |
1376 | return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM | | |
1377 | VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM; | |
1378 | else | |
1379 | return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM; | |
1380 | } | |
1381 | ||
e3ecdffa AD |
1382 | /** |
1383 | * amdgpu_device_check_block_size - validate the vm block size | |
1384 | * | |
1385 | * @adev: amdgpu_device pointer | |
1386 | * | |
1387 | * Validates the vm block size specified via module parameter. | |
1388 | * The vm block size defines number of bits in page table versus page directory, | |
1389 | * a page is 4KB so we have 12 bits offset, minimum 9 bits in the | |
1390 | * page table and the remaining bits are in the page directory. | |
1391 | */ | |
06ec9070 | 1392 | static void amdgpu_device_check_block_size(struct amdgpu_device *adev) |
a1adf8be CZ |
1393 | { |
1394 | /* defines number of bits in page table versus page directory, | |
1395 | * a page is 4KB so we have 12 bits offset, minimum 9 bits in the | |
1396 | * page table and the remaining bits are in the page directory */ | |
bab4fee7 JZ |
1397 | if (amdgpu_vm_block_size == -1) |
1398 | return; | |
a1adf8be | 1399 | |
bab4fee7 | 1400 | if (amdgpu_vm_block_size < 9) { |
a1adf8be CZ |
1401 | dev_warn(adev->dev, "VM page table size (%d) too small\n", |
1402 | amdgpu_vm_block_size); | |
97489129 | 1403 | amdgpu_vm_block_size = -1; |
a1adf8be | 1404 | } |
a1adf8be CZ |
1405 | } |
1406 | ||
e3ecdffa AD |
1407 | /** |
1408 | * amdgpu_device_check_vm_size - validate the vm size | |
1409 | * | |
1410 | * @adev: amdgpu_device pointer | |
1411 | * | |
1412 | * Validates the vm size in GB specified via module parameter. | |
1413 | * The VM size is the size of the GPU virtual memory space in GB. | |
1414 | */ | |
06ec9070 | 1415 | static void amdgpu_device_check_vm_size(struct amdgpu_device *adev) |
83ca145d | 1416 | { |
64dab074 AD |
1417 | /* no need to check the default value */ |
1418 | if (amdgpu_vm_size == -1) | |
1419 | return; | |
1420 | ||
83ca145d ZJ |
1421 | if (amdgpu_vm_size < 1) { |
1422 | dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n", | |
1423 | amdgpu_vm_size); | |
f3368128 | 1424 | amdgpu_vm_size = -1; |
83ca145d | 1425 | } |
83ca145d ZJ |
1426 | } |
1427 | ||
7951e376 RZ |
1428 | static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev) |
1429 | { | |
1430 | struct sysinfo si; | |
a9d4fe2f | 1431 | bool is_os_64 = (sizeof(void *) == 8); |
7951e376 RZ |
1432 | uint64_t total_memory; |
1433 | uint64_t dram_size_seven_GB = 0x1B8000000; | |
1434 | uint64_t dram_size_three_GB = 0xB8000000; | |
1435 | ||
1436 | if (amdgpu_smu_memory_pool_size == 0) | |
1437 | return; | |
1438 | ||
1439 | if (!is_os_64) { | |
1440 | DRM_WARN("Not 64-bit OS, feature not supported\n"); | |
1441 | goto def_value; | |
1442 | } | |
1443 | si_meminfo(&si); | |
1444 | total_memory = (uint64_t)si.totalram * si.mem_unit; | |
1445 | ||
1446 | if ((amdgpu_smu_memory_pool_size == 1) || | |
1447 | (amdgpu_smu_memory_pool_size == 2)) { | |
1448 | if (total_memory < dram_size_three_GB) | |
1449 | goto def_value1; | |
1450 | } else if ((amdgpu_smu_memory_pool_size == 4) || | |
1451 | (amdgpu_smu_memory_pool_size == 8)) { | |
1452 | if (total_memory < dram_size_seven_GB) | |
1453 | goto def_value1; | |
1454 | } else { | |
1455 | DRM_WARN("Smu memory pool size not supported\n"); | |
1456 | goto def_value; | |
1457 | } | |
1458 | adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28; | |
1459 | ||
1460 | return; | |
1461 | ||
1462 | def_value1: | |
1463 | DRM_WARN("No enough system memory\n"); | |
1464 | def_value: | |
1465 | adev->pm.smu_prv_buffer_size = 0; | |
1466 | } | |
1467 | ||
9f6a7857 HR |
1468 | static int amdgpu_device_init_apu_flags(struct amdgpu_device *adev) |
1469 | { | |
1470 | if (!(adev->flags & AMD_IS_APU) || | |
1471 | adev->asic_type < CHIP_RAVEN) | |
1472 | return 0; | |
1473 | ||
1474 | switch (adev->asic_type) { | |
1475 | case CHIP_RAVEN: | |
1476 | if (adev->pdev->device == 0x15dd) | |
1477 | adev->apu_flags |= AMD_APU_IS_RAVEN; | |
1478 | if (adev->pdev->device == 0x15d8) | |
1479 | adev->apu_flags |= AMD_APU_IS_PICASSO; | |
1480 | break; | |
1481 | case CHIP_RENOIR: | |
1482 | if ((adev->pdev->device == 0x1636) || | |
1483 | (adev->pdev->device == 0x164c)) | |
1484 | adev->apu_flags |= AMD_APU_IS_RENOIR; | |
1485 | else | |
1486 | adev->apu_flags |= AMD_APU_IS_GREEN_SARDINE; | |
1487 | break; | |
1488 | case CHIP_VANGOGH: | |
1489 | adev->apu_flags |= AMD_APU_IS_VANGOGH; | |
1490 | break; | |
1491 | case CHIP_YELLOW_CARP: | |
1492 | break; | |
d0f56dc2 | 1493 | case CHIP_CYAN_SKILLFISH: |
dfcc3e8c AD |
1494 | if ((adev->pdev->device == 0x13FE) || |
1495 | (adev->pdev->device == 0x143F)) | |
d0f56dc2 TZ |
1496 | adev->apu_flags |= AMD_APU_IS_CYAN_SKILLFISH2; |
1497 | break; | |
9f6a7857 | 1498 | default: |
4eaf21b7 | 1499 | break; |
9f6a7857 HR |
1500 | } |
1501 | ||
1502 | return 0; | |
1503 | } | |
1504 | ||
d38ceaf9 | 1505 | /** |
06ec9070 | 1506 | * amdgpu_device_check_arguments - validate module params |
d38ceaf9 AD |
1507 | * |
1508 | * @adev: amdgpu_device pointer | |
1509 | * | |
1510 | * Validates certain module parameters and updates | |
1511 | * the associated values used by the driver (all asics). | |
1512 | */ | |
912dfc84 | 1513 | static int amdgpu_device_check_arguments(struct amdgpu_device *adev) |
d38ceaf9 | 1514 | { |
5b011235 CZ |
1515 | if (amdgpu_sched_jobs < 4) { |
1516 | dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n", | |
1517 | amdgpu_sched_jobs); | |
1518 | amdgpu_sched_jobs = 4; | |
76117507 | 1519 | } else if (!is_power_of_2(amdgpu_sched_jobs)){ |
5b011235 CZ |
1520 | dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n", |
1521 | amdgpu_sched_jobs); | |
1522 | amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs); | |
1523 | } | |
d38ceaf9 | 1524 | |
83e74db6 | 1525 | if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) { |
f9321cc4 CK |
1526 | /* gart size must be greater or equal to 32M */ |
1527 | dev_warn(adev->dev, "gart size (%d) too small\n", | |
1528 | amdgpu_gart_size); | |
83e74db6 | 1529 | amdgpu_gart_size = -1; |
d38ceaf9 AD |
1530 | } |
1531 | ||
36d38372 | 1532 | if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) { |
c4e1a13a | 1533 | /* gtt size must be greater or equal to 32M */ |
36d38372 CK |
1534 | dev_warn(adev->dev, "gtt size (%d) too small\n", |
1535 | amdgpu_gtt_size); | |
1536 | amdgpu_gtt_size = -1; | |
d38ceaf9 AD |
1537 | } |
1538 | ||
d07f14be RH |
1539 | /* valid range is between 4 and 9 inclusive */ |
1540 | if (amdgpu_vm_fragment_size != -1 && | |
1541 | (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) { | |
1542 | dev_warn(adev->dev, "valid range is between 4 and 9\n"); | |
1543 | amdgpu_vm_fragment_size = -1; | |
1544 | } | |
1545 | ||
5d5bd5e3 KW |
1546 | if (amdgpu_sched_hw_submission < 2) { |
1547 | dev_warn(adev->dev, "sched hw submission jobs (%d) must be at least 2\n", | |
1548 | amdgpu_sched_hw_submission); | |
1549 | amdgpu_sched_hw_submission = 2; | |
1550 | } else if (!is_power_of_2(amdgpu_sched_hw_submission)) { | |
1551 | dev_warn(adev->dev, "sched hw submission jobs (%d) must be a power of 2\n", | |
1552 | amdgpu_sched_hw_submission); | |
1553 | amdgpu_sched_hw_submission = roundup_pow_of_two(amdgpu_sched_hw_submission); | |
1554 | } | |
1555 | ||
2656fd23 AG |
1556 | if (amdgpu_reset_method < -1 || amdgpu_reset_method > 4) { |
1557 | dev_warn(adev->dev, "invalid option for reset method, reverting to default\n"); | |
1558 | amdgpu_reset_method = -1; | |
1559 | } | |
1560 | ||
7951e376 RZ |
1561 | amdgpu_device_check_smu_prv_buffer_size(adev); |
1562 | ||
06ec9070 | 1563 | amdgpu_device_check_vm_size(adev); |
d38ceaf9 | 1564 | |
06ec9070 | 1565 | amdgpu_device_check_block_size(adev); |
6a7f76e7 | 1566 | |
19aede77 | 1567 | adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type); |
912dfc84 | 1568 | |
e3c00faa | 1569 | return 0; |
d38ceaf9 AD |
1570 | } |
1571 | ||
1572 | /** | |
1573 | * amdgpu_switcheroo_set_state - set switcheroo state | |
1574 | * | |
1575 | * @pdev: pci dev pointer | |
1694467b | 1576 | * @state: vga_switcheroo state |
d38ceaf9 | 1577 | * |
12024b17 | 1578 | * Callback for the switcheroo driver. Suspends or resumes |
d38ceaf9 AD |
1579 | * the asics before or after it is powered up using ACPI methods. |
1580 | */ | |
8aba21b7 LT |
1581 | static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, |
1582 | enum vga_switcheroo_state state) | |
d38ceaf9 AD |
1583 | { |
1584 | struct drm_device *dev = pci_get_drvdata(pdev); | |
de185019 | 1585 | int r; |
d38ceaf9 | 1586 | |
b98c6299 | 1587 | if (amdgpu_device_supports_px(dev) && state == VGA_SWITCHEROO_OFF) |
d38ceaf9 AD |
1588 | return; |
1589 | ||
1590 | if (state == VGA_SWITCHEROO_ON) { | |
dd4fa6c1 | 1591 | pr_info("switched on\n"); |
d38ceaf9 AD |
1592 | /* don't suspend or resume card normally */ |
1593 | dev->switch_power_state = DRM_SWITCH_POWER_CHANGING; | |
1594 | ||
8f66090b TZ |
1595 | pci_set_power_state(pdev, PCI_D0); |
1596 | amdgpu_device_load_pci_state(pdev); | |
1597 | r = pci_enable_device(pdev); | |
de185019 AD |
1598 | if (r) |
1599 | DRM_WARN("pci_enable_device failed (%d)\n", r); | |
1600 | amdgpu_device_resume(dev, true); | |
d38ceaf9 | 1601 | |
d38ceaf9 | 1602 | dev->switch_power_state = DRM_SWITCH_POWER_ON; |
d38ceaf9 | 1603 | } else { |
dd4fa6c1 | 1604 | pr_info("switched off\n"); |
d38ceaf9 | 1605 | dev->switch_power_state = DRM_SWITCH_POWER_CHANGING; |
de185019 | 1606 | amdgpu_device_suspend(dev, true); |
8f66090b | 1607 | amdgpu_device_cache_pci_state(pdev); |
de185019 | 1608 | /* Shut down the device */ |
8f66090b TZ |
1609 | pci_disable_device(pdev); |
1610 | pci_set_power_state(pdev, PCI_D3cold); | |
d38ceaf9 AD |
1611 | dev->switch_power_state = DRM_SWITCH_POWER_OFF; |
1612 | } | |
1613 | } | |
1614 | ||
1615 | /** | |
1616 | * amdgpu_switcheroo_can_switch - see if switcheroo state can change | |
1617 | * | |
1618 | * @pdev: pci dev pointer | |
1619 | * | |
1620 | * Callback for the switcheroo driver. Check of the switcheroo | |
1621 | * state can be changed. | |
1622 | * Returns true if the state can be changed, false if not. | |
1623 | */ | |
1624 | static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev) | |
1625 | { | |
1626 | struct drm_device *dev = pci_get_drvdata(pdev); | |
1627 | ||
1628 | /* | |
1629 | * FIXME: open_count is protected by drm_global_mutex but that would lead to | |
1630 | * locking inversion with the driver load path. And the access here is | |
1631 | * completely racy anyway. So don't bother with locking for now. | |
1632 | */ | |
7e13ad89 | 1633 | return atomic_read(&dev->open_count) == 0; |
d38ceaf9 AD |
1634 | } |
1635 | ||
1636 | static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = { | |
1637 | .set_gpu_state = amdgpu_switcheroo_set_state, | |
1638 | .reprobe = NULL, | |
1639 | .can_switch = amdgpu_switcheroo_can_switch, | |
1640 | }; | |
1641 | ||
e3ecdffa AD |
1642 | /** |
1643 | * amdgpu_device_ip_set_clockgating_state - set the CG state | |
1644 | * | |
87e3f136 | 1645 | * @dev: amdgpu_device pointer |
e3ecdffa AD |
1646 | * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) |
1647 | * @state: clockgating state (gate or ungate) | |
1648 | * | |
1649 | * Sets the requested clockgating state for all instances of | |
1650 | * the hardware IP specified. | |
1651 | * Returns the error code from the last instance. | |
1652 | */ | |
43fa561f | 1653 | int amdgpu_device_ip_set_clockgating_state(void *dev, |
2990a1fc AD |
1654 | enum amd_ip_block_type block_type, |
1655 | enum amd_clockgating_state state) | |
d38ceaf9 | 1656 | { |
43fa561f | 1657 | struct amdgpu_device *adev = dev; |
d38ceaf9 AD |
1658 | int i, r = 0; |
1659 | ||
1660 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 1661 | if (!adev->ip_blocks[i].status.valid) |
9ecbe7f5 | 1662 | continue; |
c722865a RZ |
1663 | if (adev->ip_blocks[i].version->type != block_type) |
1664 | continue; | |
1665 | if (!adev->ip_blocks[i].version->funcs->set_clockgating_state) | |
1666 | continue; | |
1667 | r = adev->ip_blocks[i].version->funcs->set_clockgating_state( | |
1668 | (void *)adev, state); | |
1669 | if (r) | |
1670 | DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n", | |
1671 | adev->ip_blocks[i].version->funcs->name, r); | |
d38ceaf9 AD |
1672 | } |
1673 | return r; | |
1674 | } | |
1675 | ||
e3ecdffa AD |
1676 | /** |
1677 | * amdgpu_device_ip_set_powergating_state - set the PG state | |
1678 | * | |
87e3f136 | 1679 | * @dev: amdgpu_device pointer |
e3ecdffa AD |
1680 | * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) |
1681 | * @state: powergating state (gate or ungate) | |
1682 | * | |
1683 | * Sets the requested powergating state for all instances of | |
1684 | * the hardware IP specified. | |
1685 | * Returns the error code from the last instance. | |
1686 | */ | |
43fa561f | 1687 | int amdgpu_device_ip_set_powergating_state(void *dev, |
2990a1fc AD |
1688 | enum amd_ip_block_type block_type, |
1689 | enum amd_powergating_state state) | |
d38ceaf9 | 1690 | { |
43fa561f | 1691 | struct amdgpu_device *adev = dev; |
d38ceaf9 AD |
1692 | int i, r = 0; |
1693 | ||
1694 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 1695 | if (!adev->ip_blocks[i].status.valid) |
9ecbe7f5 | 1696 | continue; |
c722865a RZ |
1697 | if (adev->ip_blocks[i].version->type != block_type) |
1698 | continue; | |
1699 | if (!adev->ip_blocks[i].version->funcs->set_powergating_state) | |
1700 | continue; | |
1701 | r = adev->ip_blocks[i].version->funcs->set_powergating_state( | |
1702 | (void *)adev, state); | |
1703 | if (r) | |
1704 | DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n", | |
1705 | adev->ip_blocks[i].version->funcs->name, r); | |
d38ceaf9 AD |
1706 | } |
1707 | return r; | |
1708 | } | |
1709 | ||
e3ecdffa AD |
1710 | /** |
1711 | * amdgpu_device_ip_get_clockgating_state - get the CG state | |
1712 | * | |
1713 | * @adev: amdgpu_device pointer | |
1714 | * @flags: clockgating feature flags | |
1715 | * | |
1716 | * Walks the list of IPs on the device and updates the clockgating | |
1717 | * flags for each IP. | |
1718 | * Updates @flags with the feature flags for each hardware IP where | |
1719 | * clockgating is enabled. | |
1720 | */ | |
2990a1fc | 1721 | void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev, |
25faeddc | 1722 | u64 *flags) |
6cb2d4e4 HR |
1723 | { |
1724 | int i; | |
1725 | ||
1726 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
1727 | if (!adev->ip_blocks[i].status.valid) | |
1728 | continue; | |
1729 | if (adev->ip_blocks[i].version->funcs->get_clockgating_state) | |
1730 | adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags); | |
1731 | } | |
1732 | } | |
1733 | ||
e3ecdffa AD |
1734 | /** |
1735 | * amdgpu_device_ip_wait_for_idle - wait for idle | |
1736 | * | |
1737 | * @adev: amdgpu_device pointer | |
1738 | * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) | |
1739 | * | |
1740 | * Waits for the request hardware IP to be idle. | |
1741 | * Returns 0 for success or a negative error code on failure. | |
1742 | */ | |
2990a1fc AD |
1743 | int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev, |
1744 | enum amd_ip_block_type block_type) | |
5dbbb60b AD |
1745 | { |
1746 | int i, r; | |
1747 | ||
1748 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 1749 | if (!adev->ip_blocks[i].status.valid) |
9ecbe7f5 | 1750 | continue; |
a1255107 AD |
1751 | if (adev->ip_blocks[i].version->type == block_type) { |
1752 | r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev); | |
5dbbb60b AD |
1753 | if (r) |
1754 | return r; | |
1755 | break; | |
1756 | } | |
1757 | } | |
1758 | return 0; | |
1759 | ||
1760 | } | |
1761 | ||
e3ecdffa AD |
1762 | /** |
1763 | * amdgpu_device_ip_is_idle - is the hardware IP idle | |
1764 | * | |
1765 | * @adev: amdgpu_device pointer | |
1766 | * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) | |
1767 | * | |
1768 | * Check if the hardware IP is idle or not. | |
1769 | * Returns true if it the IP is idle, false if not. | |
1770 | */ | |
2990a1fc AD |
1771 | bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev, |
1772 | enum amd_ip_block_type block_type) | |
5dbbb60b AD |
1773 | { |
1774 | int i; | |
1775 | ||
1776 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 1777 | if (!adev->ip_blocks[i].status.valid) |
9ecbe7f5 | 1778 | continue; |
a1255107 AD |
1779 | if (adev->ip_blocks[i].version->type == block_type) |
1780 | return adev->ip_blocks[i].version->funcs->is_idle((void *)adev); | |
5dbbb60b AD |
1781 | } |
1782 | return true; | |
1783 | ||
1784 | } | |
1785 | ||
e3ecdffa AD |
1786 | /** |
1787 | * amdgpu_device_ip_get_ip_block - get a hw IP pointer | |
1788 | * | |
1789 | * @adev: amdgpu_device pointer | |
87e3f136 | 1790 | * @type: Type of hardware IP (SMU, GFX, UVD, etc.) |
e3ecdffa AD |
1791 | * |
1792 | * Returns a pointer to the hardware IP block structure | |
1793 | * if it exists for the asic, otherwise NULL. | |
1794 | */ | |
2990a1fc AD |
1795 | struct amdgpu_ip_block * |
1796 | amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev, | |
1797 | enum amd_ip_block_type type) | |
d38ceaf9 AD |
1798 | { |
1799 | int i; | |
1800 | ||
1801 | for (i = 0; i < adev->num_ip_blocks; i++) | |
a1255107 | 1802 | if (adev->ip_blocks[i].version->type == type) |
d38ceaf9 AD |
1803 | return &adev->ip_blocks[i]; |
1804 | ||
1805 | return NULL; | |
1806 | } | |
1807 | ||
1808 | /** | |
2990a1fc | 1809 | * amdgpu_device_ip_block_version_cmp |
d38ceaf9 AD |
1810 | * |
1811 | * @adev: amdgpu_device pointer | |
5fc3aeeb | 1812 | * @type: enum amd_ip_block_type |
d38ceaf9 AD |
1813 | * @major: major version |
1814 | * @minor: minor version | |
1815 | * | |
1816 | * return 0 if equal or greater | |
1817 | * return 1 if smaller or the ip_block doesn't exist | |
1818 | */ | |
2990a1fc AD |
1819 | int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev, |
1820 | enum amd_ip_block_type type, | |
1821 | u32 major, u32 minor) | |
d38ceaf9 | 1822 | { |
2990a1fc | 1823 | struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type); |
d38ceaf9 | 1824 | |
a1255107 AD |
1825 | if (ip_block && ((ip_block->version->major > major) || |
1826 | ((ip_block->version->major == major) && | |
1827 | (ip_block->version->minor >= minor)))) | |
d38ceaf9 AD |
1828 | return 0; |
1829 | ||
1830 | return 1; | |
1831 | } | |
1832 | ||
a1255107 | 1833 | /** |
2990a1fc | 1834 | * amdgpu_device_ip_block_add |
a1255107 AD |
1835 | * |
1836 | * @adev: amdgpu_device pointer | |
1837 | * @ip_block_version: pointer to the IP to add | |
1838 | * | |
1839 | * Adds the IP block driver information to the collection of IPs | |
1840 | * on the asic. | |
1841 | */ | |
2990a1fc AD |
1842 | int amdgpu_device_ip_block_add(struct amdgpu_device *adev, |
1843 | const struct amdgpu_ip_block_version *ip_block_version) | |
a1255107 AD |
1844 | { |
1845 | if (!ip_block_version) | |
1846 | return -EINVAL; | |
1847 | ||
7bd939d0 LG |
1848 | switch (ip_block_version->type) { |
1849 | case AMD_IP_BLOCK_TYPE_VCN: | |
1850 | if (adev->harvest_ip_mask & AMD_HARVEST_IP_VCN_MASK) | |
1851 | return 0; | |
1852 | break; | |
1853 | case AMD_IP_BLOCK_TYPE_JPEG: | |
1854 | if (adev->harvest_ip_mask & AMD_HARVEST_IP_JPEG_MASK) | |
1855 | return 0; | |
1856 | break; | |
1857 | default: | |
1858 | break; | |
1859 | } | |
1860 | ||
e966a725 | 1861 | DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks, |
a0bae357 HR |
1862 | ip_block_version->funcs->name); |
1863 | ||
a1255107 AD |
1864 | adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version; |
1865 | ||
1866 | return 0; | |
1867 | } | |
1868 | ||
e3ecdffa AD |
1869 | /** |
1870 | * amdgpu_device_enable_virtual_display - enable virtual display feature | |
1871 | * | |
1872 | * @adev: amdgpu_device pointer | |
1873 | * | |
1874 | * Enabled the virtual display feature if the user has enabled it via | |
1875 | * the module parameter virtual_display. This feature provides a virtual | |
1876 | * display hardware on headless boards or in virtualized environments. | |
1877 | * This function parses and validates the configuration string specified by | |
1878 | * the user and configues the virtual display configuration (number of | |
1879 | * virtual connectors, crtcs, etc.) specified. | |
1880 | */ | |
483ef985 | 1881 | static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev) |
9accf2fd ED |
1882 | { |
1883 | adev->enable_virtual_display = false; | |
1884 | ||
1885 | if (amdgpu_virtual_display) { | |
8f66090b | 1886 | const char *pci_address_name = pci_name(adev->pdev); |
0f66356d | 1887 | char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname; |
9accf2fd ED |
1888 | |
1889 | pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL); | |
1890 | pciaddstr_tmp = pciaddstr; | |
0f66356d ED |
1891 | while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) { |
1892 | pciaddname = strsep(&pciaddname_tmp, ","); | |
967de2a9 YT |
1893 | if (!strcmp("all", pciaddname) |
1894 | || !strcmp(pci_address_name, pciaddname)) { | |
0f66356d ED |
1895 | long num_crtc; |
1896 | int res = -1; | |
1897 | ||
9accf2fd | 1898 | adev->enable_virtual_display = true; |
0f66356d ED |
1899 | |
1900 | if (pciaddname_tmp) | |
1901 | res = kstrtol(pciaddname_tmp, 10, | |
1902 | &num_crtc); | |
1903 | ||
1904 | if (!res) { | |
1905 | if (num_crtc < 1) | |
1906 | num_crtc = 1; | |
1907 | if (num_crtc > 6) | |
1908 | num_crtc = 6; | |
1909 | adev->mode_info.num_crtc = num_crtc; | |
1910 | } else { | |
1911 | adev->mode_info.num_crtc = 1; | |
1912 | } | |
9accf2fd ED |
1913 | break; |
1914 | } | |
1915 | } | |
1916 | ||
0f66356d ED |
1917 | DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n", |
1918 | amdgpu_virtual_display, pci_address_name, | |
1919 | adev->enable_virtual_display, adev->mode_info.num_crtc); | |
9accf2fd ED |
1920 | |
1921 | kfree(pciaddstr); | |
1922 | } | |
1923 | } | |
1924 | ||
25263da3 AD |
1925 | void amdgpu_device_set_sriov_virtual_display(struct amdgpu_device *adev) |
1926 | { | |
1927 | if (amdgpu_sriov_vf(adev) && !adev->enable_virtual_display) { | |
1928 | adev->mode_info.num_crtc = 1; | |
1929 | adev->enable_virtual_display = true; | |
1930 | DRM_INFO("virtual_display:%d, num_crtc:%d\n", | |
1931 | adev->enable_virtual_display, adev->mode_info.num_crtc); | |
1932 | } | |
1933 | } | |
1934 | ||
e3ecdffa AD |
1935 | /** |
1936 | * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware | |
1937 | * | |
1938 | * @adev: amdgpu_device pointer | |
1939 | * | |
1940 | * Parses the asic configuration parameters specified in the gpu info | |
1941 | * firmware and makes them availale to the driver for use in configuring | |
1942 | * the asic. | |
1943 | * Returns 0 on success, -EINVAL on failure. | |
1944 | */ | |
e2a75f88 AD |
1945 | static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev) |
1946 | { | |
e2a75f88 | 1947 | const char *chip_name; |
c0a43457 | 1948 | char fw_name[40]; |
e2a75f88 AD |
1949 | int err; |
1950 | const struct gpu_info_firmware_header_v1_0 *hdr; | |
1951 | ||
ab4fe3e1 HR |
1952 | adev->firmware.gpu_info_fw = NULL; |
1953 | ||
72de33f8 | 1954 | if (adev->mman.discovery_bin) { |
cc375d8c TY |
1955 | /* |
1956 | * FIXME: The bounding box is still needed by Navi12, so | |
e24d0e91 | 1957 | * temporarily read it from gpu_info firmware. Should be dropped |
cc375d8c TY |
1958 | * when DAL no longer needs it. |
1959 | */ | |
1960 | if (adev->asic_type != CHIP_NAVI12) | |
1961 | return 0; | |
258620d0 AD |
1962 | } |
1963 | ||
e2a75f88 | 1964 | switch (adev->asic_type) { |
e2a75f88 AD |
1965 | default: |
1966 | return 0; | |
1967 | case CHIP_VEGA10: | |
1968 | chip_name = "vega10"; | |
1969 | break; | |
3f76dced AD |
1970 | case CHIP_VEGA12: |
1971 | chip_name = "vega12"; | |
1972 | break; | |
2d2e5e7e | 1973 | case CHIP_RAVEN: |
54f78a76 | 1974 | if (adev->apu_flags & AMD_APU_IS_RAVEN2) |
54c4d17e | 1975 | chip_name = "raven2"; |
54f78a76 | 1976 | else if (adev->apu_flags & AMD_APU_IS_PICASSO) |
741deade | 1977 | chip_name = "picasso"; |
54c4d17e FX |
1978 | else |
1979 | chip_name = "raven"; | |
2d2e5e7e | 1980 | break; |
65e60f6e LM |
1981 | case CHIP_ARCTURUS: |
1982 | chip_name = "arcturus"; | |
1983 | break; | |
42b325e5 XY |
1984 | case CHIP_NAVI12: |
1985 | chip_name = "navi12"; | |
1986 | break; | |
e2a75f88 AD |
1987 | } |
1988 | ||
1989 | snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name); | |
b31d3063 | 1990 | err = amdgpu_ucode_request(adev, &adev->firmware.gpu_info_fw, fw_name); |
e2a75f88 AD |
1991 | if (err) { |
1992 | dev_err(adev->dev, | |
b31d3063 | 1993 | "Failed to get gpu_info firmware \"%s\"\n", |
e2a75f88 AD |
1994 | fw_name); |
1995 | goto out; | |
1996 | } | |
1997 | ||
ab4fe3e1 | 1998 | hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data; |
e2a75f88 AD |
1999 | amdgpu_ucode_print_gpu_info_hdr(&hdr->header); |
2000 | ||
2001 | switch (hdr->version_major) { | |
2002 | case 1: | |
2003 | { | |
2004 | const struct gpu_info_firmware_v1_0 *gpu_info_fw = | |
ab4fe3e1 | 2005 | (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data + |
e2a75f88 AD |
2006 | le32_to_cpu(hdr->header.ucode_array_offset_bytes)); |
2007 | ||
cc375d8c TY |
2008 | /* |
2009 | * Should be droped when DAL no longer needs it. | |
2010 | */ | |
2011 | if (adev->asic_type == CHIP_NAVI12) | |
ec51d3fa XY |
2012 | goto parse_soc_bounding_box; |
2013 | ||
b5ab16bf AD |
2014 | adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se); |
2015 | adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh); | |
2016 | adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se); | |
2017 | adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se); | |
e2a75f88 | 2018 | adev->gfx.config.max_texture_channel_caches = |
b5ab16bf AD |
2019 | le32_to_cpu(gpu_info_fw->gc_num_tccs); |
2020 | adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs); | |
2021 | adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds); | |
2022 | adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth); | |
2023 | adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth); | |
e2a75f88 | 2024 | adev->gfx.config.double_offchip_lds_buf = |
b5ab16bf AD |
2025 | le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer); |
2026 | adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size); | |
51fd0370 HZ |
2027 | adev->gfx.cu_info.max_waves_per_simd = |
2028 | le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd); | |
2029 | adev->gfx.cu_info.max_scratch_slots_per_cu = | |
2030 | le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu); | |
2031 | adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size); | |
48321c3d | 2032 | if (hdr->version_minor >= 1) { |
35c2e910 HZ |
2033 | const struct gpu_info_firmware_v1_1 *gpu_info_fw = |
2034 | (const struct gpu_info_firmware_v1_1 *)(adev->firmware.gpu_info_fw->data + | |
2035 | le32_to_cpu(hdr->header.ucode_array_offset_bytes)); | |
2036 | adev->gfx.config.num_sc_per_sh = | |
2037 | le32_to_cpu(gpu_info_fw->num_sc_per_sh); | |
2038 | adev->gfx.config.num_packer_per_sc = | |
2039 | le32_to_cpu(gpu_info_fw->num_packer_per_sc); | |
2040 | } | |
ec51d3fa XY |
2041 | |
2042 | parse_soc_bounding_box: | |
ec51d3fa XY |
2043 | /* |
2044 | * soc bounding box info is not integrated in disocovery table, | |
258620d0 | 2045 | * we always need to parse it from gpu info firmware if needed. |
ec51d3fa | 2046 | */ |
48321c3d HW |
2047 | if (hdr->version_minor == 2) { |
2048 | const struct gpu_info_firmware_v1_2 *gpu_info_fw = | |
2049 | (const struct gpu_info_firmware_v1_2 *)(adev->firmware.gpu_info_fw->data + | |
2050 | le32_to_cpu(hdr->header.ucode_array_offset_bytes)); | |
2051 | adev->dm.soc_bounding_box = &gpu_info_fw->soc_bounding_box; | |
2052 | } | |
e2a75f88 AD |
2053 | break; |
2054 | } | |
2055 | default: | |
2056 | dev_err(adev->dev, | |
2057 | "Unsupported gpu_info table %d\n", hdr->header.ucode_version); | |
2058 | err = -EINVAL; | |
2059 | goto out; | |
2060 | } | |
2061 | out: | |
e2a75f88 AD |
2062 | return err; |
2063 | } | |
2064 | ||
e3ecdffa AD |
2065 | /** |
2066 | * amdgpu_device_ip_early_init - run early init for hardware IPs | |
2067 | * | |
2068 | * @adev: amdgpu_device pointer | |
2069 | * | |
2070 | * Early initialization pass for hardware IPs. The hardware IPs that make | |
2071 | * up each asic are discovered each IP's early_init callback is run. This | |
2072 | * is the first stage in initializing the asic. | |
2073 | * Returns 0 on success, negative error code on failure. | |
2074 | */ | |
06ec9070 | 2075 | static int amdgpu_device_ip_early_init(struct amdgpu_device *adev) |
d38ceaf9 | 2076 | { |
901e2be2 AD |
2077 | struct drm_device *dev = adev_to_drm(adev); |
2078 | struct pci_dev *parent; | |
aaa36a97 | 2079 | int i, r; |
ced69502 | 2080 | bool total; |
d38ceaf9 | 2081 | |
483ef985 | 2082 | amdgpu_device_enable_virtual_display(adev); |
a6be7570 | 2083 | |
00a979f3 | 2084 | if (amdgpu_sriov_vf(adev)) { |
00a979f3 | 2085 | r = amdgpu_virt_request_full_gpu(adev, true); |
aaa36a97 AD |
2086 | if (r) |
2087 | return r; | |
00a979f3 WS |
2088 | } |
2089 | ||
d38ceaf9 | 2090 | switch (adev->asic_type) { |
33f34802 KW |
2091 | #ifdef CONFIG_DRM_AMDGPU_SI |
2092 | case CHIP_VERDE: | |
2093 | case CHIP_TAHITI: | |
2094 | case CHIP_PITCAIRN: | |
2095 | case CHIP_OLAND: | |
2096 | case CHIP_HAINAN: | |
295d0daf | 2097 | adev->family = AMDGPU_FAMILY_SI; |
33f34802 KW |
2098 | r = si_set_ip_blocks(adev); |
2099 | if (r) | |
2100 | return r; | |
2101 | break; | |
2102 | #endif | |
a2e73f56 AD |
2103 | #ifdef CONFIG_DRM_AMDGPU_CIK |
2104 | case CHIP_BONAIRE: | |
2105 | case CHIP_HAWAII: | |
2106 | case CHIP_KAVERI: | |
2107 | case CHIP_KABINI: | |
2108 | case CHIP_MULLINS: | |
e1ad2d53 | 2109 | if (adev->flags & AMD_IS_APU) |
a2e73f56 | 2110 | adev->family = AMDGPU_FAMILY_KV; |
e1ad2d53 AD |
2111 | else |
2112 | adev->family = AMDGPU_FAMILY_CI; | |
a2e73f56 AD |
2113 | |
2114 | r = cik_set_ip_blocks(adev); | |
2115 | if (r) | |
2116 | return r; | |
2117 | break; | |
2118 | #endif | |
da87c30b AD |
2119 | case CHIP_TOPAZ: |
2120 | case CHIP_TONGA: | |
2121 | case CHIP_FIJI: | |
2122 | case CHIP_POLARIS10: | |
2123 | case CHIP_POLARIS11: | |
2124 | case CHIP_POLARIS12: | |
2125 | case CHIP_VEGAM: | |
2126 | case CHIP_CARRIZO: | |
2127 | case CHIP_STONEY: | |
2128 | if (adev->flags & AMD_IS_APU) | |
2129 | adev->family = AMDGPU_FAMILY_CZ; | |
2130 | else | |
2131 | adev->family = AMDGPU_FAMILY_VI; | |
2132 | ||
2133 | r = vi_set_ip_blocks(adev); | |
2134 | if (r) | |
2135 | return r; | |
2136 | break; | |
d38ceaf9 | 2137 | default: |
63352b7f AD |
2138 | r = amdgpu_discovery_set_ip_blocks(adev); |
2139 | if (r) | |
2140 | return r; | |
2141 | break; | |
d38ceaf9 AD |
2142 | } |
2143 | ||
901e2be2 AD |
2144 | if (amdgpu_has_atpx() && |
2145 | (amdgpu_is_atpx_hybrid() || | |
2146 | amdgpu_has_atpx_dgpu_power_cntl()) && | |
2147 | ((adev->flags & AMD_IS_APU) == 0) && | |
2148 | !pci_is_thunderbolt_attached(to_pci_dev(dev->dev))) | |
2149 | adev->flags |= AMD_IS_PX; | |
2150 | ||
85ac2021 AD |
2151 | if (!(adev->flags & AMD_IS_APU)) { |
2152 | parent = pci_upstream_bridge(adev->pdev); | |
2153 | adev->has_pr3 = parent ? pci_pr3_present(parent) : false; | |
2154 | } | |
901e2be2 | 2155 | |
c004d44e | 2156 | amdgpu_amdkfd_device_probe(adev); |
1884734a | 2157 | |
3b94fb10 | 2158 | adev->pm.pp_feature = amdgpu_pp_feature_mask; |
a35ad98b | 2159 | if (amdgpu_sriov_vf(adev) || sched_policy == KFD_SCHED_POLICY_NO_HWS) |
00544006 | 2160 | adev->pm.pp_feature &= ~PP_GFXOFF_MASK; |
4215a119 HC |
2161 | if (amdgpu_sriov_vf(adev) && adev->asic_type == CHIP_SIENNA_CICHLID) |
2162 | adev->pm.pp_feature &= ~PP_OVERDRIVE_MASK; | |
00f54b97 | 2163 | |
ced69502 | 2164 | total = true; |
d38ceaf9 AD |
2165 | for (i = 0; i < adev->num_ip_blocks; i++) { |
2166 | if ((amdgpu_ip_block_mask & (1 << i)) == 0) { | |
ed8cf00c HR |
2167 | DRM_ERROR("disabled ip block: %d <%s>\n", |
2168 | i, adev->ip_blocks[i].version->funcs->name); | |
a1255107 | 2169 | adev->ip_blocks[i].status.valid = false; |
d38ceaf9 | 2170 | } else { |
a1255107 AD |
2171 | if (adev->ip_blocks[i].version->funcs->early_init) { |
2172 | r = adev->ip_blocks[i].version->funcs->early_init((void *)adev); | |
2c1a2784 | 2173 | if (r == -ENOENT) { |
a1255107 | 2174 | adev->ip_blocks[i].status.valid = false; |
2c1a2784 | 2175 | } else if (r) { |
a1255107 AD |
2176 | DRM_ERROR("early_init of IP block <%s> failed %d\n", |
2177 | adev->ip_blocks[i].version->funcs->name, r); | |
ced69502 | 2178 | total = false; |
2c1a2784 | 2179 | } else { |
a1255107 | 2180 | adev->ip_blocks[i].status.valid = true; |
2c1a2784 | 2181 | } |
974e6b64 | 2182 | } else { |
a1255107 | 2183 | adev->ip_blocks[i].status.valid = true; |
d38ceaf9 | 2184 | } |
d38ceaf9 | 2185 | } |
21a249ca AD |
2186 | /* get the vbios after the asic_funcs are set up */ |
2187 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) { | |
6e29c227 AD |
2188 | r = amdgpu_device_parse_gpu_info_fw(adev); |
2189 | if (r) | |
2190 | return r; | |
2191 | ||
21a249ca AD |
2192 | /* Read BIOS */ |
2193 | if (!amdgpu_get_bios(adev)) | |
2194 | return -EINVAL; | |
2195 | ||
2196 | r = amdgpu_atombios_init(adev); | |
2197 | if (r) { | |
2198 | dev_err(adev->dev, "amdgpu_atombios_init failed\n"); | |
2199 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0); | |
2200 | return r; | |
2201 | } | |
77eabc6f PJZ |
2202 | |
2203 | /*get pf2vf msg info at it's earliest time*/ | |
2204 | if (amdgpu_sriov_vf(adev)) | |
2205 | amdgpu_virt_init_data_exchange(adev); | |
2206 | ||
21a249ca | 2207 | } |
d38ceaf9 | 2208 | } |
ced69502 ML |
2209 | if (!total) |
2210 | return -ENODEV; | |
d38ceaf9 | 2211 | |
395d1fb9 NH |
2212 | adev->cg_flags &= amdgpu_cg_mask; |
2213 | adev->pg_flags &= amdgpu_pg_mask; | |
2214 | ||
d38ceaf9 AD |
2215 | return 0; |
2216 | } | |
2217 | ||
0a4f2520 RZ |
2218 | static int amdgpu_device_ip_hw_init_phase1(struct amdgpu_device *adev) |
2219 | { | |
2220 | int i, r; | |
2221 | ||
2222 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
2223 | if (!adev->ip_blocks[i].status.sw) | |
2224 | continue; | |
2225 | if (adev->ip_blocks[i].status.hw) | |
2226 | continue; | |
2227 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || | |
2d11fd3f | 2228 | (amdgpu_sriov_vf(adev) && (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)) || |
0a4f2520 RZ |
2229 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) { |
2230 | r = adev->ip_blocks[i].version->funcs->hw_init(adev); | |
2231 | if (r) { | |
2232 | DRM_ERROR("hw_init of IP block <%s> failed %d\n", | |
2233 | adev->ip_blocks[i].version->funcs->name, r); | |
2234 | return r; | |
2235 | } | |
2236 | adev->ip_blocks[i].status.hw = true; | |
2237 | } | |
2238 | } | |
2239 | ||
2240 | return 0; | |
2241 | } | |
2242 | ||
2243 | static int amdgpu_device_ip_hw_init_phase2(struct amdgpu_device *adev) | |
2244 | { | |
2245 | int i, r; | |
2246 | ||
2247 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
2248 | if (!adev->ip_blocks[i].status.sw) | |
2249 | continue; | |
2250 | if (adev->ip_blocks[i].status.hw) | |
2251 | continue; | |
2252 | r = adev->ip_blocks[i].version->funcs->hw_init(adev); | |
2253 | if (r) { | |
2254 | DRM_ERROR("hw_init of IP block <%s> failed %d\n", | |
2255 | adev->ip_blocks[i].version->funcs->name, r); | |
2256 | return r; | |
2257 | } | |
2258 | adev->ip_blocks[i].status.hw = true; | |
2259 | } | |
2260 | ||
2261 | return 0; | |
2262 | } | |
2263 | ||
7a3e0bb2 RZ |
2264 | static int amdgpu_device_fw_loading(struct amdgpu_device *adev) |
2265 | { | |
2266 | int r = 0; | |
2267 | int i; | |
80f41f84 | 2268 | uint32_t smu_version; |
7a3e0bb2 RZ |
2269 | |
2270 | if (adev->asic_type >= CHIP_VEGA10) { | |
2271 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
482f0e53 ML |
2272 | if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_PSP) |
2273 | continue; | |
2274 | ||
e3c1b071 | 2275 | if (!adev->ip_blocks[i].status.sw) |
2276 | continue; | |
2277 | ||
482f0e53 ML |
2278 | /* no need to do the fw loading again if already done*/ |
2279 | if (adev->ip_blocks[i].status.hw == true) | |
2280 | break; | |
2281 | ||
53b3f8f4 | 2282 | if (amdgpu_in_reset(adev) || adev->in_suspend) { |
482f0e53 ML |
2283 | r = adev->ip_blocks[i].version->funcs->resume(adev); |
2284 | if (r) { | |
2285 | DRM_ERROR("resume of IP block <%s> failed %d\n", | |
7a3e0bb2 | 2286 | adev->ip_blocks[i].version->funcs->name, r); |
482f0e53 ML |
2287 | return r; |
2288 | } | |
2289 | } else { | |
2290 | r = adev->ip_blocks[i].version->funcs->hw_init(adev); | |
2291 | if (r) { | |
2292 | DRM_ERROR("hw_init of IP block <%s> failed %d\n", | |
2293 | adev->ip_blocks[i].version->funcs->name, r); | |
2294 | return r; | |
7a3e0bb2 | 2295 | } |
7a3e0bb2 | 2296 | } |
482f0e53 ML |
2297 | |
2298 | adev->ip_blocks[i].status.hw = true; | |
2299 | break; | |
7a3e0bb2 RZ |
2300 | } |
2301 | } | |
482f0e53 | 2302 | |
8973d9ec ED |
2303 | if (!amdgpu_sriov_vf(adev) || adev->asic_type == CHIP_TONGA) |
2304 | r = amdgpu_pm_load_smu_firmware(adev, &smu_version); | |
7a3e0bb2 | 2305 | |
80f41f84 | 2306 | return r; |
7a3e0bb2 RZ |
2307 | } |
2308 | ||
5fd8518d AG |
2309 | static int amdgpu_device_init_schedulers(struct amdgpu_device *adev) |
2310 | { | |
2311 | long timeout; | |
2312 | int r, i; | |
2313 | ||
2314 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { | |
2315 | struct amdgpu_ring *ring = adev->rings[i]; | |
2316 | ||
2317 | /* No need to setup the GPU scheduler for rings that don't need it */ | |
2318 | if (!ring || ring->no_scheduler) | |
2319 | continue; | |
2320 | ||
2321 | switch (ring->funcs->type) { | |
2322 | case AMDGPU_RING_TYPE_GFX: | |
2323 | timeout = adev->gfx_timeout; | |
2324 | break; | |
2325 | case AMDGPU_RING_TYPE_COMPUTE: | |
2326 | timeout = adev->compute_timeout; | |
2327 | break; | |
2328 | case AMDGPU_RING_TYPE_SDMA: | |
2329 | timeout = adev->sdma_timeout; | |
2330 | break; | |
2331 | default: | |
2332 | timeout = adev->video_timeout; | |
2333 | break; | |
2334 | } | |
2335 | ||
2336 | r = drm_sched_init(&ring->sched, &amdgpu_sched_ops, | |
2337 | ring->num_hw_submission, amdgpu_job_hang_limit, | |
8ab62eda JG |
2338 | timeout, adev->reset_domain->wq, |
2339 | ring->sched_score, ring->name, | |
2340 | adev->dev); | |
5fd8518d AG |
2341 | if (r) { |
2342 | DRM_ERROR("Failed to create scheduler on ring %s.\n", | |
2343 | ring->name); | |
2344 | return r; | |
2345 | } | |
2346 | } | |
2347 | ||
2348 | return 0; | |
2349 | } | |
2350 | ||
2351 | ||
e3ecdffa AD |
2352 | /** |
2353 | * amdgpu_device_ip_init - run init for hardware IPs | |
2354 | * | |
2355 | * @adev: amdgpu_device pointer | |
2356 | * | |
2357 | * Main initialization pass for hardware IPs. The list of all the hardware | |
2358 | * IPs that make up the asic is walked and the sw_init and hw_init callbacks | |
2359 | * are run. sw_init initializes the software state associated with each IP | |
2360 | * and hw_init initializes the hardware associated with each IP. | |
2361 | * Returns 0 on success, negative error code on failure. | |
2362 | */ | |
06ec9070 | 2363 | static int amdgpu_device_ip_init(struct amdgpu_device *adev) |
d38ceaf9 AD |
2364 | { |
2365 | int i, r; | |
2366 | ||
c030f2e4 | 2367 | r = amdgpu_ras_init(adev); |
2368 | if (r) | |
2369 | return r; | |
2370 | ||
d38ceaf9 | 2371 | for (i = 0; i < adev->num_ip_blocks; i++) { |
a1255107 | 2372 | if (!adev->ip_blocks[i].status.valid) |
d38ceaf9 | 2373 | continue; |
a1255107 | 2374 | r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev); |
2c1a2784 | 2375 | if (r) { |
a1255107 AD |
2376 | DRM_ERROR("sw_init of IP block <%s> failed %d\n", |
2377 | adev->ip_blocks[i].version->funcs->name, r); | |
72d3f592 | 2378 | goto init_failed; |
2c1a2784 | 2379 | } |
a1255107 | 2380 | adev->ip_blocks[i].status.sw = true; |
bfca0289 | 2381 | |
c1c39032 AD |
2382 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) { |
2383 | /* need to do common hw init early so everything is set up for gmc */ | |
2384 | r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev); | |
2385 | if (r) { | |
2386 | DRM_ERROR("hw_init %d failed %d\n", i, r); | |
2387 | goto init_failed; | |
2388 | } | |
2389 | adev->ip_blocks[i].status.hw = true; | |
2390 | } else if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) { | |
2391 | /* need to do gmc hw init early so we can allocate gpu mem */ | |
892deb48 VS |
2392 | /* Try to reserve bad pages early */ |
2393 | if (amdgpu_sriov_vf(adev)) | |
2394 | amdgpu_virt_exchange_data(adev); | |
2395 | ||
7ccfd79f | 2396 | r = amdgpu_device_mem_scratch_init(adev); |
2c1a2784 | 2397 | if (r) { |
7ccfd79f | 2398 | DRM_ERROR("amdgpu_mem_scratch_init failed %d\n", r); |
72d3f592 | 2399 | goto init_failed; |
2c1a2784 | 2400 | } |
a1255107 | 2401 | r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev); |
2c1a2784 AD |
2402 | if (r) { |
2403 | DRM_ERROR("hw_init %d failed %d\n", i, r); | |
72d3f592 | 2404 | goto init_failed; |
2c1a2784 | 2405 | } |
06ec9070 | 2406 | r = amdgpu_device_wb_init(adev); |
2c1a2784 | 2407 | if (r) { |
06ec9070 | 2408 | DRM_ERROR("amdgpu_device_wb_init failed %d\n", r); |
72d3f592 | 2409 | goto init_failed; |
2c1a2784 | 2410 | } |
a1255107 | 2411 | adev->ip_blocks[i].status.hw = true; |
2493664f ML |
2412 | |
2413 | /* right after GMC hw init, we create CSA */ | |
8a1fbb4a | 2414 | if (amdgpu_mcbp) { |
1e256e27 | 2415 | r = amdgpu_allocate_static_csa(adev, &adev->virt.csa_obj, |
58ab2c08 CK |
2416 | AMDGPU_GEM_DOMAIN_VRAM | |
2417 | AMDGPU_GEM_DOMAIN_GTT, | |
2418 | AMDGPU_CSA_SIZE); | |
2493664f ML |
2419 | if (r) { |
2420 | DRM_ERROR("allocate CSA failed %d\n", r); | |
72d3f592 | 2421 | goto init_failed; |
2493664f ML |
2422 | } |
2423 | } | |
d38ceaf9 AD |
2424 | } |
2425 | } | |
2426 | ||
c9ffa427 | 2427 | if (amdgpu_sriov_vf(adev)) |
22c16d25 | 2428 | amdgpu_virt_init_data_exchange(adev); |
c9ffa427 | 2429 | |
533aed27 AG |
2430 | r = amdgpu_ib_pool_init(adev); |
2431 | if (r) { | |
2432 | dev_err(adev->dev, "IB initialization failed (%d).\n", r); | |
2433 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r); | |
2434 | goto init_failed; | |
2435 | } | |
2436 | ||
c8963ea4 RZ |
2437 | r = amdgpu_ucode_create_bo(adev); /* create ucode bo when sw_init complete*/ |
2438 | if (r) | |
72d3f592 | 2439 | goto init_failed; |
0a4f2520 RZ |
2440 | |
2441 | r = amdgpu_device_ip_hw_init_phase1(adev); | |
2442 | if (r) | |
72d3f592 | 2443 | goto init_failed; |
0a4f2520 | 2444 | |
7a3e0bb2 RZ |
2445 | r = amdgpu_device_fw_loading(adev); |
2446 | if (r) | |
72d3f592 | 2447 | goto init_failed; |
7a3e0bb2 | 2448 | |
0a4f2520 RZ |
2449 | r = amdgpu_device_ip_hw_init_phase2(adev); |
2450 | if (r) | |
72d3f592 | 2451 | goto init_failed; |
d38ceaf9 | 2452 | |
121a2bc6 AG |
2453 | /* |
2454 | * retired pages will be loaded from eeprom and reserved here, | |
2455 | * it should be called after amdgpu_device_ip_hw_init_phase2 since | |
2456 | * for some ASICs the RAS EEPROM code relies on SMU fully functioning | |
2457 | * for I2C communication which only true at this point. | |
b82e65a9 GC |
2458 | * |
2459 | * amdgpu_ras_recovery_init may fail, but the upper only cares the | |
2460 | * failure from bad gpu situation and stop amdgpu init process | |
2461 | * accordingly. For other failed cases, it will still release all | |
2462 | * the resource and print error message, rather than returning one | |
2463 | * negative value to upper level. | |
121a2bc6 AG |
2464 | * |
2465 | * Note: theoretically, this should be called before all vram allocations | |
2466 | * to protect retired page from abusing | |
2467 | */ | |
b82e65a9 GC |
2468 | r = amdgpu_ras_recovery_init(adev); |
2469 | if (r) | |
2470 | goto init_failed; | |
121a2bc6 | 2471 | |
cfbb6b00 AG |
2472 | /** |
2473 | * In case of XGMI grab extra reference for reset domain for this device | |
2474 | */ | |
a4c63caf | 2475 | if (adev->gmc.xgmi.num_physical_nodes > 1) { |
cfbb6b00 | 2476 | if (amdgpu_xgmi_add_device(adev) == 0) { |
46c67660 | 2477 | if (!amdgpu_sriov_vf(adev)) { |
2efc30f0 VC |
2478 | struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev); |
2479 | ||
dfd0287b LH |
2480 | if (WARN_ON(!hive)) { |
2481 | r = -ENOENT; | |
2482 | goto init_failed; | |
2483 | } | |
2484 | ||
46c67660 | 2485 | if (!hive->reset_domain || |
2486 | !amdgpu_reset_get_reset_domain(hive->reset_domain)) { | |
2487 | r = -ENOENT; | |
2488 | amdgpu_put_xgmi_hive(hive); | |
2489 | goto init_failed; | |
2490 | } | |
2491 | ||
2492 | /* Drop the early temporary reset domain we created for device */ | |
2493 | amdgpu_reset_put_reset_domain(adev->reset_domain); | |
2494 | adev->reset_domain = hive->reset_domain; | |
9dfa4860 | 2495 | amdgpu_put_xgmi_hive(hive); |
cfbb6b00 | 2496 | } |
a4c63caf AG |
2497 | } |
2498 | } | |
2499 | ||
5fd8518d AG |
2500 | r = amdgpu_device_init_schedulers(adev); |
2501 | if (r) | |
2502 | goto init_failed; | |
e3c1b071 | 2503 | |
2504 | /* Don't init kfd if whole hive need to be reset during init */ | |
c004d44e | 2505 | if (!adev->gmc.xgmi.pending_reset) |
e3c1b071 | 2506 | amdgpu_amdkfd_device_init(adev); |
c6332b97 | 2507 | |
bd607166 KR |
2508 | amdgpu_fru_get_product_info(adev); |
2509 | ||
72d3f592 | 2510 | init_failed: |
c9ffa427 | 2511 | if (amdgpu_sriov_vf(adev)) |
c6332b97 | 2512 | amdgpu_virt_release_full_gpu(adev, true); |
2513 | ||
72d3f592 | 2514 | return r; |
d38ceaf9 AD |
2515 | } |
2516 | ||
e3ecdffa AD |
2517 | /** |
2518 | * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer | |
2519 | * | |
2520 | * @adev: amdgpu_device pointer | |
2521 | * | |
2522 | * Writes a reset magic value to the gart pointer in VRAM. The driver calls | |
2523 | * this function before a GPU reset. If the value is retained after a | |
2524 | * GPU reset, VRAM has not been lost. Some GPU resets may destry VRAM contents. | |
2525 | */ | |
06ec9070 | 2526 | static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev) |
0c49e0b8 CZ |
2527 | { |
2528 | memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM); | |
2529 | } | |
2530 | ||
e3ecdffa AD |
2531 | /** |
2532 | * amdgpu_device_check_vram_lost - check if vram is valid | |
2533 | * | |
2534 | * @adev: amdgpu_device pointer | |
2535 | * | |
2536 | * Checks the reset magic value written to the gart pointer in VRAM. | |
2537 | * The driver calls this after a GPU reset to see if the contents of | |
2538 | * VRAM is lost or now. | |
2539 | * returns true if vram is lost, false if not. | |
2540 | */ | |
06ec9070 | 2541 | static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev) |
0c49e0b8 | 2542 | { |
dadce777 EQ |
2543 | if (memcmp(adev->gart.ptr, adev->reset_magic, |
2544 | AMDGPU_RESET_MAGIC_NUM)) | |
2545 | return true; | |
2546 | ||
53b3f8f4 | 2547 | if (!amdgpu_in_reset(adev)) |
dadce777 EQ |
2548 | return false; |
2549 | ||
2550 | /* | |
2551 | * For all ASICs with baco/mode1 reset, the VRAM is | |
2552 | * always assumed to be lost. | |
2553 | */ | |
2554 | switch (amdgpu_asic_reset_method(adev)) { | |
2555 | case AMD_RESET_METHOD_BACO: | |
2556 | case AMD_RESET_METHOD_MODE1: | |
2557 | return true; | |
2558 | default: | |
2559 | return false; | |
2560 | } | |
0c49e0b8 CZ |
2561 | } |
2562 | ||
e3ecdffa | 2563 | /** |
1112a46b | 2564 | * amdgpu_device_set_cg_state - set clockgating for amdgpu device |
e3ecdffa AD |
2565 | * |
2566 | * @adev: amdgpu_device pointer | |
b8b72130 | 2567 | * @state: clockgating state (gate or ungate) |
e3ecdffa | 2568 | * |
e3ecdffa | 2569 | * The list of all the hardware IPs that make up the asic is walked and the |
1112a46b RZ |
2570 | * set_clockgating_state callbacks are run. |
2571 | * Late initialization pass enabling clockgating for hardware IPs. | |
2572 | * Fini or suspend, pass disabling clockgating for hardware IPs. | |
e3ecdffa AD |
2573 | * Returns 0 on success, negative error code on failure. |
2574 | */ | |
fdd34271 | 2575 | |
5d89bb2d LL |
2576 | int amdgpu_device_set_cg_state(struct amdgpu_device *adev, |
2577 | enum amd_clockgating_state state) | |
d38ceaf9 | 2578 | { |
1112a46b | 2579 | int i, j, r; |
d38ceaf9 | 2580 | |
4a2ba394 SL |
2581 | if (amdgpu_emu_mode == 1) |
2582 | return 0; | |
2583 | ||
1112a46b RZ |
2584 | for (j = 0; j < adev->num_ip_blocks; j++) { |
2585 | i = state == AMD_CG_STATE_GATE ? j : adev->num_ip_blocks - j - 1; | |
a2d31dc3 | 2586 | if (!adev->ip_blocks[i].status.late_initialized) |
d38ceaf9 | 2587 | continue; |
47198eb7 | 2588 | /* skip CG for GFX, SDMA on S0ix */ |
5d70a549 | 2589 | if (adev->in_s0ix && |
47198eb7 AD |
2590 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX || |
2591 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SDMA)) | |
5d70a549 | 2592 | continue; |
4a446d55 | 2593 | /* skip CG for VCE/UVD, it's handled specially */ |
a1255107 | 2594 | if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD && |
57716327 | 2595 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE && |
34319b32 | 2596 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN && |
52f2e779 | 2597 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG && |
57716327 | 2598 | adev->ip_blocks[i].version->funcs->set_clockgating_state) { |
4a446d55 | 2599 | /* enable clockgating to save power */ |
a1255107 | 2600 | r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev, |
1112a46b | 2601 | state); |
4a446d55 AD |
2602 | if (r) { |
2603 | DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n", | |
a1255107 | 2604 | adev->ip_blocks[i].version->funcs->name, r); |
4a446d55 AD |
2605 | return r; |
2606 | } | |
b0b00ff1 | 2607 | } |
d38ceaf9 | 2608 | } |
06b18f61 | 2609 | |
c9f96fd5 RZ |
2610 | return 0; |
2611 | } | |
2612 | ||
5d89bb2d LL |
2613 | int amdgpu_device_set_pg_state(struct amdgpu_device *adev, |
2614 | enum amd_powergating_state state) | |
c9f96fd5 | 2615 | { |
1112a46b | 2616 | int i, j, r; |
06b18f61 | 2617 | |
c9f96fd5 RZ |
2618 | if (amdgpu_emu_mode == 1) |
2619 | return 0; | |
2620 | ||
1112a46b RZ |
2621 | for (j = 0; j < adev->num_ip_blocks; j++) { |
2622 | i = state == AMD_PG_STATE_GATE ? j : adev->num_ip_blocks - j - 1; | |
a2d31dc3 | 2623 | if (!adev->ip_blocks[i].status.late_initialized) |
c9f96fd5 | 2624 | continue; |
47198eb7 | 2625 | /* skip PG for GFX, SDMA on S0ix */ |
5d70a549 | 2626 | if (adev->in_s0ix && |
47198eb7 AD |
2627 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX || |
2628 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SDMA)) | |
5d70a549 | 2629 | continue; |
c9f96fd5 RZ |
2630 | /* skip CG for VCE/UVD, it's handled specially */ |
2631 | if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD && | |
2632 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE && | |
2633 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN && | |
52f2e779 | 2634 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG && |
c9f96fd5 RZ |
2635 | adev->ip_blocks[i].version->funcs->set_powergating_state) { |
2636 | /* enable powergating to save power */ | |
2637 | r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev, | |
1112a46b | 2638 | state); |
c9f96fd5 RZ |
2639 | if (r) { |
2640 | DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n", | |
2641 | adev->ip_blocks[i].version->funcs->name, r); | |
2642 | return r; | |
2643 | } | |
2644 | } | |
2645 | } | |
2dc80b00 S |
2646 | return 0; |
2647 | } | |
2648 | ||
beff74bc AD |
2649 | static int amdgpu_device_enable_mgpu_fan_boost(void) |
2650 | { | |
2651 | struct amdgpu_gpu_instance *gpu_ins; | |
2652 | struct amdgpu_device *adev; | |
2653 | int i, ret = 0; | |
2654 | ||
2655 | mutex_lock(&mgpu_info.mutex); | |
2656 | ||
2657 | /* | |
2658 | * MGPU fan boost feature should be enabled | |
2659 | * only when there are two or more dGPUs in | |
2660 | * the system | |
2661 | */ | |
2662 | if (mgpu_info.num_dgpu < 2) | |
2663 | goto out; | |
2664 | ||
2665 | for (i = 0; i < mgpu_info.num_dgpu; i++) { | |
2666 | gpu_ins = &(mgpu_info.gpu_ins[i]); | |
2667 | adev = gpu_ins->adev; | |
2668 | if (!(adev->flags & AMD_IS_APU) && | |
f10bb940 | 2669 | !gpu_ins->mgpu_fan_enabled) { |
beff74bc AD |
2670 | ret = amdgpu_dpm_enable_mgpu_fan_boost(adev); |
2671 | if (ret) | |
2672 | break; | |
2673 | ||
2674 | gpu_ins->mgpu_fan_enabled = 1; | |
2675 | } | |
2676 | } | |
2677 | ||
2678 | out: | |
2679 | mutex_unlock(&mgpu_info.mutex); | |
2680 | ||
2681 | return ret; | |
2682 | } | |
2683 | ||
e3ecdffa AD |
2684 | /** |
2685 | * amdgpu_device_ip_late_init - run late init for hardware IPs | |
2686 | * | |
2687 | * @adev: amdgpu_device pointer | |
2688 | * | |
2689 | * Late initialization pass for hardware IPs. The list of all the hardware | |
2690 | * IPs that make up the asic is walked and the late_init callbacks are run. | |
2691 | * late_init covers any special initialization that an IP requires | |
2692 | * after all of the have been initialized or something that needs to happen | |
2693 | * late in the init process. | |
2694 | * Returns 0 on success, negative error code on failure. | |
2695 | */ | |
06ec9070 | 2696 | static int amdgpu_device_ip_late_init(struct amdgpu_device *adev) |
2dc80b00 | 2697 | { |
60599a03 | 2698 | struct amdgpu_gpu_instance *gpu_instance; |
2dc80b00 S |
2699 | int i = 0, r; |
2700 | ||
2701 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
73f847db | 2702 | if (!adev->ip_blocks[i].status.hw) |
2dc80b00 S |
2703 | continue; |
2704 | if (adev->ip_blocks[i].version->funcs->late_init) { | |
2705 | r = adev->ip_blocks[i].version->funcs->late_init((void *)adev); | |
2706 | if (r) { | |
2707 | DRM_ERROR("late_init of IP block <%s> failed %d\n", | |
2708 | adev->ip_blocks[i].version->funcs->name, r); | |
2709 | return r; | |
2710 | } | |
2dc80b00 | 2711 | } |
73f847db | 2712 | adev->ip_blocks[i].status.late_initialized = true; |
2dc80b00 S |
2713 | } |
2714 | ||
867e24ca | 2715 | r = amdgpu_ras_late_init(adev); |
2716 | if (r) { | |
2717 | DRM_ERROR("amdgpu_ras_late_init failed %d", r); | |
2718 | return r; | |
2719 | } | |
2720 | ||
a891d239 DL |
2721 | amdgpu_ras_set_error_query_ready(adev, true); |
2722 | ||
1112a46b RZ |
2723 | amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE); |
2724 | amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE); | |
916ac57f | 2725 | |
06ec9070 | 2726 | amdgpu_device_fill_reset_magic(adev); |
d38ceaf9 | 2727 | |
beff74bc AD |
2728 | r = amdgpu_device_enable_mgpu_fan_boost(); |
2729 | if (r) | |
2730 | DRM_ERROR("enable mgpu fan boost failed (%d).\n", r); | |
2731 | ||
4da8b639 | 2732 | /* For passthrough configuration on arcturus and aldebaran, enable special handling SBR */ |
2733 | if (amdgpu_passthrough(adev) && ((adev->asic_type == CHIP_ARCTURUS && adev->gmc.xgmi.num_physical_nodes > 1)|| | |
2734 | adev->asic_type == CHIP_ALDEBARAN )) | |
bc143d8b | 2735 | amdgpu_dpm_handle_passthrough_sbr(adev, true); |
60599a03 EQ |
2736 | |
2737 | if (adev->gmc.xgmi.num_physical_nodes > 1) { | |
2738 | mutex_lock(&mgpu_info.mutex); | |
2739 | ||
2740 | /* | |
2741 | * Reset device p-state to low as this was booted with high. | |
2742 | * | |
2743 | * This should be performed only after all devices from the same | |
2744 | * hive get initialized. | |
2745 | * | |
2746 | * However, it's unknown how many device in the hive in advance. | |
2747 | * As this is counted one by one during devices initializations. | |
2748 | * | |
2749 | * So, we wait for all XGMI interlinked devices initialized. | |
2750 | * This may bring some delays as those devices may come from | |
2751 | * different hives. But that should be OK. | |
2752 | */ | |
2753 | if (mgpu_info.num_dgpu == adev->gmc.xgmi.num_physical_nodes) { | |
2754 | for (i = 0; i < mgpu_info.num_gpu; i++) { | |
2755 | gpu_instance = &(mgpu_info.gpu_ins[i]); | |
2756 | if (gpu_instance->adev->flags & AMD_IS_APU) | |
2757 | continue; | |
2758 | ||
d84a430d JK |
2759 | r = amdgpu_xgmi_set_pstate(gpu_instance->adev, |
2760 | AMDGPU_XGMI_PSTATE_MIN); | |
60599a03 EQ |
2761 | if (r) { |
2762 | DRM_ERROR("pstate setting failed (%d).\n", r); | |
2763 | break; | |
2764 | } | |
2765 | } | |
2766 | } | |
2767 | ||
2768 | mutex_unlock(&mgpu_info.mutex); | |
2769 | } | |
2770 | ||
d38ceaf9 AD |
2771 | return 0; |
2772 | } | |
2773 | ||
613aa3ea LY |
2774 | /** |
2775 | * amdgpu_device_smu_fini_early - smu hw_fini wrapper | |
2776 | * | |
2777 | * @adev: amdgpu_device pointer | |
2778 | * | |
2779 | * For ASICs need to disable SMC first | |
2780 | */ | |
2781 | static void amdgpu_device_smu_fini_early(struct amdgpu_device *adev) | |
2782 | { | |
2783 | int i, r; | |
2784 | ||
2785 | if (adev->ip_versions[GC_HWIP][0] > IP_VERSION(9, 0, 0)) | |
2786 | return; | |
2787 | ||
2788 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
2789 | if (!adev->ip_blocks[i].status.hw) | |
2790 | continue; | |
2791 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) { | |
2792 | r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev); | |
2793 | /* XXX handle errors */ | |
2794 | if (r) { | |
2795 | DRM_DEBUG("hw_fini of IP block <%s> failed %d\n", | |
2796 | adev->ip_blocks[i].version->funcs->name, r); | |
2797 | } | |
2798 | adev->ip_blocks[i].status.hw = false; | |
2799 | break; | |
2800 | } | |
2801 | } | |
2802 | } | |
2803 | ||
e9669fb7 | 2804 | static int amdgpu_device_ip_fini_early(struct amdgpu_device *adev) |
d38ceaf9 AD |
2805 | { |
2806 | int i, r; | |
2807 | ||
e9669fb7 AG |
2808 | for (i = 0; i < adev->num_ip_blocks; i++) { |
2809 | if (!adev->ip_blocks[i].version->funcs->early_fini) | |
2810 | continue; | |
5278a159 | 2811 | |
e9669fb7 AG |
2812 | r = adev->ip_blocks[i].version->funcs->early_fini((void *)adev); |
2813 | if (r) { | |
2814 | DRM_DEBUG("early_fini of IP block <%s> failed %d\n", | |
2815 | adev->ip_blocks[i].version->funcs->name, r); | |
2816 | } | |
2817 | } | |
c030f2e4 | 2818 | |
05df1f01 | 2819 | amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE); |
fdd34271 RZ |
2820 | amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE); |
2821 | ||
7270e895 TY |
2822 | amdgpu_amdkfd_suspend(adev, false); |
2823 | ||
613aa3ea LY |
2824 | /* Workaroud for ASICs need to disable SMC first */ |
2825 | amdgpu_device_smu_fini_early(adev); | |
3e96dbfd | 2826 | |
d38ceaf9 | 2827 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { |
a1255107 | 2828 | if (!adev->ip_blocks[i].status.hw) |
d38ceaf9 | 2829 | continue; |
8201a67a | 2830 | |
a1255107 | 2831 | r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev); |
d38ceaf9 | 2832 | /* XXX handle errors */ |
2c1a2784 | 2833 | if (r) { |
a1255107 AD |
2834 | DRM_DEBUG("hw_fini of IP block <%s> failed %d\n", |
2835 | adev->ip_blocks[i].version->funcs->name, r); | |
2c1a2784 | 2836 | } |
8201a67a | 2837 | |
a1255107 | 2838 | adev->ip_blocks[i].status.hw = false; |
d38ceaf9 AD |
2839 | } |
2840 | ||
6effad8a GC |
2841 | if (amdgpu_sriov_vf(adev)) { |
2842 | if (amdgpu_virt_release_full_gpu(adev, false)) | |
2843 | DRM_ERROR("failed to release exclusive mode on fini\n"); | |
2844 | } | |
2845 | ||
e9669fb7 AG |
2846 | return 0; |
2847 | } | |
2848 | ||
2849 | /** | |
2850 | * amdgpu_device_ip_fini - run fini for hardware IPs | |
2851 | * | |
2852 | * @adev: amdgpu_device pointer | |
2853 | * | |
2854 | * Main teardown pass for hardware IPs. The list of all the hardware | |
2855 | * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks | |
2856 | * are run. hw_fini tears down the hardware associated with each IP | |
2857 | * and sw_fini tears down any software state associated with each IP. | |
2858 | * Returns 0 on success, negative error code on failure. | |
2859 | */ | |
2860 | static int amdgpu_device_ip_fini(struct amdgpu_device *adev) | |
2861 | { | |
2862 | int i, r; | |
2863 | ||
2864 | if (amdgpu_sriov_vf(adev) && adev->virt.ras_init_done) | |
2865 | amdgpu_virt_release_ras_err_handler_data(adev); | |
2866 | ||
e9669fb7 AG |
2867 | if (adev->gmc.xgmi.num_physical_nodes > 1) |
2868 | amdgpu_xgmi_remove_device(adev); | |
2869 | ||
c004d44e | 2870 | amdgpu_amdkfd_device_fini_sw(adev); |
9950cda2 | 2871 | |
d38ceaf9 | 2872 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { |
a1255107 | 2873 | if (!adev->ip_blocks[i].status.sw) |
d38ceaf9 | 2874 | continue; |
c12aba3a ML |
2875 | |
2876 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) { | |
c8963ea4 | 2877 | amdgpu_ucode_free_bo(adev); |
1e256e27 | 2878 | amdgpu_free_static_csa(&adev->virt.csa_obj); |
c12aba3a | 2879 | amdgpu_device_wb_fini(adev); |
7ccfd79f | 2880 | amdgpu_device_mem_scratch_fini(adev); |
533aed27 | 2881 | amdgpu_ib_pool_fini(adev); |
c12aba3a ML |
2882 | } |
2883 | ||
a1255107 | 2884 | r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev); |
d38ceaf9 | 2885 | /* XXX handle errors */ |
2c1a2784 | 2886 | if (r) { |
a1255107 AD |
2887 | DRM_DEBUG("sw_fini of IP block <%s> failed %d\n", |
2888 | adev->ip_blocks[i].version->funcs->name, r); | |
2c1a2784 | 2889 | } |
a1255107 AD |
2890 | adev->ip_blocks[i].status.sw = false; |
2891 | adev->ip_blocks[i].status.valid = false; | |
d38ceaf9 AD |
2892 | } |
2893 | ||
a6dcfd9c | 2894 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { |
a1255107 | 2895 | if (!adev->ip_blocks[i].status.late_initialized) |
8a2eef1d | 2896 | continue; |
a1255107 AD |
2897 | if (adev->ip_blocks[i].version->funcs->late_fini) |
2898 | adev->ip_blocks[i].version->funcs->late_fini((void *)adev); | |
2899 | adev->ip_blocks[i].status.late_initialized = false; | |
a6dcfd9c ML |
2900 | } |
2901 | ||
c030f2e4 | 2902 | amdgpu_ras_fini(adev); |
2903 | ||
d38ceaf9 AD |
2904 | return 0; |
2905 | } | |
2906 | ||
e3ecdffa | 2907 | /** |
beff74bc | 2908 | * amdgpu_device_delayed_init_work_handler - work handler for IB tests |
e3ecdffa | 2909 | * |
1112a46b | 2910 | * @work: work_struct. |
e3ecdffa | 2911 | */ |
beff74bc | 2912 | static void amdgpu_device_delayed_init_work_handler(struct work_struct *work) |
2dc80b00 S |
2913 | { |
2914 | struct amdgpu_device *adev = | |
beff74bc | 2915 | container_of(work, struct amdgpu_device, delayed_init_work.work); |
916ac57f RZ |
2916 | int r; |
2917 | ||
2918 | r = amdgpu_ib_ring_tests(adev); | |
2919 | if (r) | |
2920 | DRM_ERROR("ib ring test failed (%d).\n", r); | |
2dc80b00 S |
2921 | } |
2922 | ||
1e317b99 RZ |
2923 | static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work) |
2924 | { | |
2925 | struct amdgpu_device *adev = | |
2926 | container_of(work, struct amdgpu_device, gfx.gfx_off_delay_work.work); | |
2927 | ||
90a92662 MD |
2928 | WARN_ON_ONCE(adev->gfx.gfx_off_state); |
2929 | WARN_ON_ONCE(adev->gfx.gfx_off_req_count); | |
2930 | ||
2931 | if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, true)) | |
2932 | adev->gfx.gfx_off_state = true; | |
1e317b99 RZ |
2933 | } |
2934 | ||
e3ecdffa | 2935 | /** |
e7854a03 | 2936 | * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1) |
e3ecdffa AD |
2937 | * |
2938 | * @adev: amdgpu_device pointer | |
2939 | * | |
2940 | * Main suspend function for hardware IPs. The list of all the hardware | |
2941 | * IPs that make up the asic is walked, clockgating is disabled and the | |
2942 | * suspend callbacks are run. suspend puts the hardware and software state | |
2943 | * in each IP into a state suitable for suspend. | |
2944 | * Returns 0 on success, negative error code on failure. | |
2945 | */ | |
e7854a03 AD |
2946 | static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev) |
2947 | { | |
2948 | int i, r; | |
2949 | ||
50ec83f0 AD |
2950 | amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE); |
2951 | amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE); | |
05df1f01 | 2952 | |
b31d6ada EQ |
2953 | /* |
2954 | * Per PMFW team's suggestion, driver needs to handle gfxoff | |
2955 | * and df cstate features disablement for gpu reset(e.g. Mode1Reset) | |
2956 | * scenario. Add the missing df cstate disablement here. | |
2957 | */ | |
2958 | if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_DISALLOW)) | |
2959 | dev_warn(adev->dev, "Failed to disallow df cstate"); | |
2960 | ||
e7854a03 AD |
2961 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { |
2962 | if (!adev->ip_blocks[i].status.valid) | |
2963 | continue; | |
2b9f7848 | 2964 | |
e7854a03 | 2965 | /* displays are handled separately */ |
2b9f7848 ND |
2966 | if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_DCE) |
2967 | continue; | |
2968 | ||
2969 | /* XXX handle errors */ | |
2970 | r = adev->ip_blocks[i].version->funcs->suspend(adev); | |
2971 | /* XXX handle errors */ | |
2972 | if (r) { | |
2973 | DRM_ERROR("suspend of IP block <%s> failed %d\n", | |
2974 | adev->ip_blocks[i].version->funcs->name, r); | |
2975 | return r; | |
e7854a03 | 2976 | } |
2b9f7848 ND |
2977 | |
2978 | adev->ip_blocks[i].status.hw = false; | |
e7854a03 AD |
2979 | } |
2980 | ||
e7854a03 AD |
2981 | return 0; |
2982 | } | |
2983 | ||
2984 | /** | |
2985 | * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2) | |
2986 | * | |
2987 | * @adev: amdgpu_device pointer | |
2988 | * | |
2989 | * Main suspend function for hardware IPs. The list of all the hardware | |
2990 | * IPs that make up the asic is walked, clockgating is disabled and the | |
2991 | * suspend callbacks are run. suspend puts the hardware and software state | |
2992 | * in each IP into a state suitable for suspend. | |
2993 | * Returns 0 on success, negative error code on failure. | |
2994 | */ | |
2995 | static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev) | |
d38ceaf9 AD |
2996 | { |
2997 | int i, r; | |
2998 | ||
557f42a2 | 2999 | if (adev->in_s0ix) |
bc143d8b | 3000 | amdgpu_dpm_gfx_state_change(adev, sGpuChangeState_D3Entry); |
34416931 | 3001 | |
d38ceaf9 | 3002 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { |
a1255107 | 3003 | if (!adev->ip_blocks[i].status.valid) |
d38ceaf9 | 3004 | continue; |
e7854a03 AD |
3005 | /* displays are handled in phase1 */ |
3006 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) | |
3007 | continue; | |
bff77e86 LM |
3008 | /* PSP lost connection when err_event_athub occurs */ |
3009 | if (amdgpu_ras_intr_triggered() && | |
3010 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) { | |
3011 | adev->ip_blocks[i].status.hw = false; | |
3012 | continue; | |
3013 | } | |
e3c1b071 | 3014 | |
3015 | /* skip unnecessary suspend if we do not initialize them yet */ | |
3016 | if (adev->gmc.xgmi.pending_reset && | |
3017 | !(adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC || | |
3018 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC || | |
3019 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || | |
3020 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH)) { | |
3021 | adev->ip_blocks[i].status.hw = false; | |
3022 | continue; | |
3023 | } | |
557f42a2 | 3024 | |
afa6646b | 3025 | /* skip suspend of gfx/mes and psp for S0ix |
32ff160d AD |
3026 | * gfx is in gfxoff state, so on resume it will exit gfxoff just |
3027 | * like at runtime. PSP is also part of the always on hardware | |
3028 | * so no need to suspend it. | |
3029 | */ | |
557f42a2 | 3030 | if (adev->in_s0ix && |
32ff160d | 3031 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP || |
afa6646b AD |
3032 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX || |
3033 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_MES)) | |
557f42a2 AD |
3034 | continue; |
3035 | ||
2a7798ea AD |
3036 | /* SDMA 5.x+ is part of GFX power domain so it's covered by GFXOFF */ |
3037 | if (adev->in_s0ix && | |
3038 | (adev->ip_versions[SDMA0_HWIP][0] >= IP_VERSION(5, 0, 0)) && | |
3039 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SDMA)) | |
3040 | continue; | |
3041 | ||
e11c7750 TH |
3042 | /* Once swPSP provides the IMU, RLC FW binaries to TOS during cold-boot. |
3043 | * These are in TMR, hence are expected to be reused by PSP-TOS to reload | |
3044 | * from this location and RLC Autoload automatically also gets loaded | |
3045 | * from here based on PMFW -> PSP message during re-init sequence. | |
3046 | * Therefore, the psp suspend & resume should be skipped to avoid destroy | |
3047 | * the TMR and reload FWs again for IMU enabled APU ASICs. | |
3048 | */ | |
3049 | if (amdgpu_in_reset(adev) && | |
3050 | (adev->flags & AMD_IS_APU) && adev->gfx.imu.funcs && | |
3051 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) | |
3052 | continue; | |
3053 | ||
d38ceaf9 | 3054 | /* XXX handle errors */ |
a1255107 | 3055 | r = adev->ip_blocks[i].version->funcs->suspend(adev); |
d38ceaf9 | 3056 | /* XXX handle errors */ |
2c1a2784 | 3057 | if (r) { |
a1255107 AD |
3058 | DRM_ERROR("suspend of IP block <%s> failed %d\n", |
3059 | adev->ip_blocks[i].version->funcs->name, r); | |
2c1a2784 | 3060 | } |
876923fb | 3061 | adev->ip_blocks[i].status.hw = false; |
a3a09142 | 3062 | /* handle putting the SMC in the appropriate state */ |
86b93fd6 JZ |
3063 | if(!amdgpu_sriov_vf(adev)){ |
3064 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) { | |
3065 | r = amdgpu_dpm_set_mp1_state(adev, adev->mp1_state); | |
3066 | if (r) { | |
3067 | DRM_ERROR("SMC failed to set mp1 state %d, %d\n", | |
3068 | adev->mp1_state, r); | |
3069 | return r; | |
3070 | } | |
a3a09142 AD |
3071 | } |
3072 | } | |
d38ceaf9 AD |
3073 | } |
3074 | ||
3075 | return 0; | |
3076 | } | |
3077 | ||
e7854a03 AD |
3078 | /** |
3079 | * amdgpu_device_ip_suspend - run suspend for hardware IPs | |
3080 | * | |
3081 | * @adev: amdgpu_device pointer | |
3082 | * | |
3083 | * Main suspend function for hardware IPs. The list of all the hardware | |
3084 | * IPs that make up the asic is walked, clockgating is disabled and the | |
3085 | * suspend callbacks are run. suspend puts the hardware and software state | |
3086 | * in each IP into a state suitable for suspend. | |
3087 | * Returns 0 on success, negative error code on failure. | |
3088 | */ | |
3089 | int amdgpu_device_ip_suspend(struct amdgpu_device *adev) | |
3090 | { | |
3091 | int r; | |
3092 | ||
3c73683c JC |
3093 | if (amdgpu_sriov_vf(adev)) { |
3094 | amdgpu_virt_fini_data_exchange(adev); | |
e7819644 | 3095 | amdgpu_virt_request_full_gpu(adev, false); |
3c73683c | 3096 | } |
e7819644 | 3097 | |
e7854a03 AD |
3098 | r = amdgpu_device_ip_suspend_phase1(adev); |
3099 | if (r) | |
3100 | return r; | |
3101 | r = amdgpu_device_ip_suspend_phase2(adev); | |
3102 | ||
e7819644 YT |
3103 | if (amdgpu_sriov_vf(adev)) |
3104 | amdgpu_virt_release_full_gpu(adev, false); | |
3105 | ||
e7854a03 AD |
3106 | return r; |
3107 | } | |
3108 | ||
06ec9070 | 3109 | static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev) |
a90ad3c2 ML |
3110 | { |
3111 | int i, r; | |
3112 | ||
2cb681b6 | 3113 | static enum amd_ip_block_type ip_order[] = { |
2cb681b6 | 3114 | AMD_IP_BLOCK_TYPE_COMMON, |
c1c39032 | 3115 | AMD_IP_BLOCK_TYPE_GMC, |
39186aef | 3116 | AMD_IP_BLOCK_TYPE_PSP, |
2cb681b6 ML |
3117 | AMD_IP_BLOCK_TYPE_IH, |
3118 | }; | |
a90ad3c2 | 3119 | |
95ea3dbc | 3120 | for (i = 0; i < adev->num_ip_blocks; i++) { |
2cb681b6 ML |
3121 | int j; |
3122 | struct amdgpu_ip_block *block; | |
a90ad3c2 | 3123 | |
4cd2a96d J |
3124 | block = &adev->ip_blocks[i]; |
3125 | block->status.hw = false; | |
2cb681b6 | 3126 | |
4cd2a96d | 3127 | for (j = 0; j < ARRAY_SIZE(ip_order); j++) { |
2cb681b6 | 3128 | |
4cd2a96d | 3129 | if (block->version->type != ip_order[j] || |
2cb681b6 ML |
3130 | !block->status.valid) |
3131 | continue; | |
3132 | ||
3133 | r = block->version->funcs->hw_init(adev); | |
0aaeefcc | 3134 | DRM_INFO("RE-INIT-early: %s %s\n", block->version->funcs->name, r?"failed":"succeeded"); |
c41d1cf6 ML |
3135 | if (r) |
3136 | return r; | |
482f0e53 | 3137 | block->status.hw = true; |
a90ad3c2 ML |
3138 | } |
3139 | } | |
3140 | ||
3141 | return 0; | |
3142 | } | |
3143 | ||
06ec9070 | 3144 | static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev) |
a90ad3c2 ML |
3145 | { |
3146 | int i, r; | |
3147 | ||
2cb681b6 ML |
3148 | static enum amd_ip_block_type ip_order[] = { |
3149 | AMD_IP_BLOCK_TYPE_SMC, | |
3150 | AMD_IP_BLOCK_TYPE_DCE, | |
3151 | AMD_IP_BLOCK_TYPE_GFX, | |
3152 | AMD_IP_BLOCK_TYPE_SDMA, | |
257deb8c | 3153 | AMD_IP_BLOCK_TYPE_UVD, |
d83c7a07 JJ |
3154 | AMD_IP_BLOCK_TYPE_VCE, |
3155 | AMD_IP_BLOCK_TYPE_VCN | |
2cb681b6 | 3156 | }; |
a90ad3c2 | 3157 | |
2cb681b6 ML |
3158 | for (i = 0; i < ARRAY_SIZE(ip_order); i++) { |
3159 | int j; | |
3160 | struct amdgpu_ip_block *block; | |
a90ad3c2 | 3161 | |
2cb681b6 ML |
3162 | for (j = 0; j < adev->num_ip_blocks; j++) { |
3163 | block = &adev->ip_blocks[j]; | |
3164 | ||
3165 | if (block->version->type != ip_order[i] || | |
482f0e53 ML |
3166 | !block->status.valid || |
3167 | block->status.hw) | |
2cb681b6 ML |
3168 | continue; |
3169 | ||
895bd048 JZ |
3170 | if (block->version->type == AMD_IP_BLOCK_TYPE_SMC) |
3171 | r = block->version->funcs->resume(adev); | |
3172 | else | |
3173 | r = block->version->funcs->hw_init(adev); | |
3174 | ||
0aaeefcc | 3175 | DRM_INFO("RE-INIT-late: %s %s\n", block->version->funcs->name, r?"failed":"succeeded"); |
c41d1cf6 ML |
3176 | if (r) |
3177 | return r; | |
482f0e53 | 3178 | block->status.hw = true; |
a90ad3c2 ML |
3179 | } |
3180 | } | |
3181 | ||
3182 | return 0; | |
3183 | } | |
3184 | ||
e3ecdffa AD |
3185 | /** |
3186 | * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs | |
3187 | * | |
3188 | * @adev: amdgpu_device pointer | |
3189 | * | |
3190 | * First resume function for hardware IPs. The list of all the hardware | |
3191 | * IPs that make up the asic is walked and the resume callbacks are run for | |
3192 | * COMMON, GMC, and IH. resume puts the hardware into a functional state | |
3193 | * after a suspend and updates the software state as necessary. This | |
3194 | * function is also used for restoring the GPU after a GPU reset. | |
3195 | * Returns 0 on success, negative error code on failure. | |
3196 | */ | |
06ec9070 | 3197 | static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev) |
d38ceaf9 AD |
3198 | { |
3199 | int i, r; | |
3200 | ||
a90ad3c2 | 3201 | for (i = 0; i < adev->num_ip_blocks; i++) { |
482f0e53 | 3202 | if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw) |
a90ad3c2 | 3203 | continue; |
a90ad3c2 | 3204 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || |
e3ecdffa | 3205 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC || |
d7274ec7 BZ |
3206 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH || |
3207 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP && amdgpu_sriov_vf(adev))) { | |
482f0e53 | 3208 | |
fcf0649f CZ |
3209 | r = adev->ip_blocks[i].version->funcs->resume(adev); |
3210 | if (r) { | |
3211 | DRM_ERROR("resume of IP block <%s> failed %d\n", | |
3212 | adev->ip_blocks[i].version->funcs->name, r); | |
3213 | return r; | |
3214 | } | |
482f0e53 | 3215 | adev->ip_blocks[i].status.hw = true; |
a90ad3c2 ML |
3216 | } |
3217 | } | |
3218 | ||
3219 | return 0; | |
3220 | } | |
3221 | ||
e3ecdffa AD |
3222 | /** |
3223 | * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs | |
3224 | * | |
3225 | * @adev: amdgpu_device pointer | |
3226 | * | |
3227 | * First resume function for hardware IPs. The list of all the hardware | |
3228 | * IPs that make up the asic is walked and the resume callbacks are run for | |
3229 | * all blocks except COMMON, GMC, and IH. resume puts the hardware into a | |
3230 | * functional state after a suspend and updates the software state as | |
3231 | * necessary. This function is also used for restoring the GPU after a GPU | |
3232 | * reset. | |
3233 | * Returns 0 on success, negative error code on failure. | |
3234 | */ | |
06ec9070 | 3235 | static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev) |
d38ceaf9 AD |
3236 | { |
3237 | int i, r; | |
3238 | ||
3239 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
482f0e53 | 3240 | if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw) |
d38ceaf9 | 3241 | continue; |
fcf0649f | 3242 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || |
e3ecdffa | 3243 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC || |
7a3e0bb2 RZ |
3244 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH || |
3245 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) | |
fcf0649f | 3246 | continue; |
a1255107 | 3247 | r = adev->ip_blocks[i].version->funcs->resume(adev); |
2c1a2784 | 3248 | if (r) { |
a1255107 AD |
3249 | DRM_ERROR("resume of IP block <%s> failed %d\n", |
3250 | adev->ip_blocks[i].version->funcs->name, r); | |
d38ceaf9 | 3251 | return r; |
2c1a2784 | 3252 | } |
482f0e53 | 3253 | adev->ip_blocks[i].status.hw = true; |
d38ceaf9 AD |
3254 | } |
3255 | ||
3256 | return 0; | |
3257 | } | |
3258 | ||
e3ecdffa AD |
3259 | /** |
3260 | * amdgpu_device_ip_resume - run resume for hardware IPs | |
3261 | * | |
3262 | * @adev: amdgpu_device pointer | |
3263 | * | |
3264 | * Main resume function for hardware IPs. The hardware IPs | |
3265 | * are split into two resume functions because they are | |
3266 | * are also used in in recovering from a GPU reset and some additional | |
3267 | * steps need to be take between them. In this case (S3/S4) they are | |
3268 | * run sequentially. | |
3269 | * Returns 0 on success, negative error code on failure. | |
3270 | */ | |
06ec9070 | 3271 | static int amdgpu_device_ip_resume(struct amdgpu_device *adev) |
fcf0649f CZ |
3272 | { |
3273 | int r; | |
3274 | ||
9cec53c1 JZ |
3275 | r = amdgpu_amdkfd_resume_iommu(adev); |
3276 | if (r) | |
3277 | return r; | |
3278 | ||
06ec9070 | 3279 | r = amdgpu_device_ip_resume_phase1(adev); |
fcf0649f CZ |
3280 | if (r) |
3281 | return r; | |
7a3e0bb2 RZ |
3282 | |
3283 | r = amdgpu_device_fw_loading(adev); | |
3284 | if (r) | |
3285 | return r; | |
3286 | ||
06ec9070 | 3287 | r = amdgpu_device_ip_resume_phase2(adev); |
fcf0649f CZ |
3288 | |
3289 | return r; | |
3290 | } | |
3291 | ||
e3ecdffa AD |
3292 | /** |
3293 | * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV | |
3294 | * | |
3295 | * @adev: amdgpu_device pointer | |
3296 | * | |
3297 | * Query the VBIOS data tables to determine if the board supports SR-IOV. | |
3298 | */ | |
4e99a44e | 3299 | static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev) |
048765ad | 3300 | { |
6867e1b5 ML |
3301 | if (amdgpu_sriov_vf(adev)) { |
3302 | if (adev->is_atom_fw) { | |
58ff791a | 3303 | if (amdgpu_atomfirmware_gpu_virtualization_supported(adev)) |
6867e1b5 ML |
3304 | adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS; |
3305 | } else { | |
3306 | if (amdgpu_atombios_has_gpu_virtualization_table(adev)) | |
3307 | adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS; | |
3308 | } | |
3309 | ||
3310 | if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS)) | |
3311 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0); | |
a5bde2f9 | 3312 | } |
048765ad AR |
3313 | } |
3314 | ||
e3ecdffa AD |
3315 | /** |
3316 | * amdgpu_device_asic_has_dc_support - determine if DC supports the asic | |
3317 | * | |
3318 | * @asic_type: AMD asic type | |
3319 | * | |
3320 | * Check if there is DC (new modesetting infrastructre) support for an asic. | |
3321 | * returns true if DC has support, false if not. | |
3322 | */ | |
4562236b HW |
3323 | bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type) |
3324 | { | |
3325 | switch (asic_type) { | |
0637d417 AD |
3326 | #ifdef CONFIG_DRM_AMDGPU_SI |
3327 | case CHIP_HAINAN: | |
3328 | #endif | |
3329 | case CHIP_TOPAZ: | |
3330 | /* chips with no display hardware */ | |
3331 | return false; | |
4562236b | 3332 | #if defined(CONFIG_DRM_AMD_DC) |
64200c46 MR |
3333 | case CHIP_TAHITI: |
3334 | case CHIP_PITCAIRN: | |
3335 | case CHIP_VERDE: | |
3336 | case CHIP_OLAND: | |
2d32ffd6 AD |
3337 | /* |
3338 | * We have systems in the wild with these ASICs that require | |
3339 | * LVDS and VGA support which is not supported with DC. | |
3340 | * | |
3341 | * Fallback to the non-DC driver here by default so as not to | |
3342 | * cause regressions. | |
3343 | */ | |
3344 | #if defined(CONFIG_DRM_AMD_DC_SI) | |
3345 | return amdgpu_dc > 0; | |
3346 | #else | |
3347 | return false; | |
64200c46 | 3348 | #endif |
4562236b | 3349 | case CHIP_BONAIRE: |
0d6fbccb | 3350 | case CHIP_KAVERI: |
367e6687 AD |
3351 | case CHIP_KABINI: |
3352 | case CHIP_MULLINS: | |
d9fda248 HW |
3353 | /* |
3354 | * We have systems in the wild with these ASICs that require | |
b5a0168e | 3355 | * VGA support which is not supported with DC. |
d9fda248 HW |
3356 | * |
3357 | * Fallback to the non-DC driver here by default so as not to | |
3358 | * cause regressions. | |
3359 | */ | |
3360 | return amdgpu_dc > 0; | |
f7f12b25 | 3361 | default: |
fd187853 | 3362 | return amdgpu_dc != 0; |
f7f12b25 | 3363 | #else |
4562236b | 3364 | default: |
93b09a9a | 3365 | if (amdgpu_dc > 0) |
044a48f4 | 3366 | DRM_INFO_ONCE("Display Core has been requested via kernel parameter " |
93b09a9a | 3367 | "but isn't supported by ASIC, ignoring\n"); |
4562236b | 3368 | return false; |
f7f12b25 | 3369 | #endif |
4562236b HW |
3370 | } |
3371 | } | |
3372 | ||
3373 | /** | |
3374 | * amdgpu_device_has_dc_support - check if dc is supported | |
3375 | * | |
982a820b | 3376 | * @adev: amdgpu_device pointer |
4562236b HW |
3377 | * |
3378 | * Returns true for supported, false for not supported | |
3379 | */ | |
3380 | bool amdgpu_device_has_dc_support(struct amdgpu_device *adev) | |
3381 | { | |
25263da3 | 3382 | if (adev->enable_virtual_display || |
abaf210c | 3383 | (adev->harvest_ip_mask & AMD_HARVEST_IP_DMU_MASK)) |
2555039d XY |
3384 | return false; |
3385 | ||
4562236b HW |
3386 | return amdgpu_device_asic_has_dc_support(adev->asic_type); |
3387 | } | |
3388 | ||
d4535e2c AG |
3389 | static void amdgpu_device_xgmi_reset_func(struct work_struct *__work) |
3390 | { | |
3391 | struct amdgpu_device *adev = | |
3392 | container_of(__work, struct amdgpu_device, xgmi_reset_work); | |
d95e8e97 | 3393 | struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev); |
d4535e2c | 3394 | |
c6a6e2db AG |
3395 | /* It's a bug to not have a hive within this function */ |
3396 | if (WARN_ON(!hive)) | |
3397 | return; | |
3398 | ||
3399 | /* | |
3400 | * Use task barrier to synchronize all xgmi reset works across the | |
3401 | * hive. task_barrier_enter and task_barrier_exit will block | |
3402 | * until all the threads running the xgmi reset works reach | |
3403 | * those points. task_barrier_full will do both blocks. | |
3404 | */ | |
3405 | if (amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) { | |
3406 | ||
3407 | task_barrier_enter(&hive->tb); | |
4a580877 | 3408 | adev->asic_reset_res = amdgpu_device_baco_enter(adev_to_drm(adev)); |
c6a6e2db AG |
3409 | |
3410 | if (adev->asic_reset_res) | |
3411 | goto fail; | |
3412 | ||
3413 | task_barrier_exit(&hive->tb); | |
4a580877 | 3414 | adev->asic_reset_res = amdgpu_device_baco_exit(adev_to_drm(adev)); |
c6a6e2db AG |
3415 | |
3416 | if (adev->asic_reset_res) | |
3417 | goto fail; | |
43c4d576 | 3418 | |
5e67bba3 | 3419 | if (adev->mmhub.ras && adev->mmhub.ras->ras_block.hw_ops && |
3420 | adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count) | |
3421 | adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count(adev); | |
c6a6e2db AG |
3422 | } else { |
3423 | ||
3424 | task_barrier_full(&hive->tb); | |
3425 | adev->asic_reset_res = amdgpu_asic_reset(adev); | |
3426 | } | |
ce316fa5 | 3427 | |
c6a6e2db | 3428 | fail: |
d4535e2c | 3429 | if (adev->asic_reset_res) |
fed184e9 | 3430 | DRM_WARN("ASIC reset failed with error, %d for drm dev, %s", |
4a580877 | 3431 | adev->asic_reset_res, adev_to_drm(adev)->unique); |
d95e8e97 | 3432 | amdgpu_put_xgmi_hive(hive); |
d4535e2c AG |
3433 | } |
3434 | ||
71f98027 AD |
3435 | static int amdgpu_device_get_job_timeout_settings(struct amdgpu_device *adev) |
3436 | { | |
3437 | char *input = amdgpu_lockup_timeout; | |
3438 | char *timeout_setting = NULL; | |
3439 | int index = 0; | |
3440 | long timeout; | |
3441 | int ret = 0; | |
3442 | ||
3443 | /* | |
67387dfe AD |
3444 | * By default timeout for non compute jobs is 10000 |
3445 | * and 60000 for compute jobs. | |
71f98027 | 3446 | * In SR-IOV or passthrough mode, timeout for compute |
b7b2a316 | 3447 | * jobs are 60000 by default. |
71f98027 AD |
3448 | */ |
3449 | adev->gfx_timeout = msecs_to_jiffies(10000); | |
3450 | adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout; | |
9882e278 ED |
3451 | if (amdgpu_sriov_vf(adev)) |
3452 | adev->compute_timeout = amdgpu_sriov_is_pp_one_vf(adev) ? | |
3453 | msecs_to_jiffies(60000) : msecs_to_jiffies(10000); | |
71f98027 | 3454 | else |
67387dfe | 3455 | adev->compute_timeout = msecs_to_jiffies(60000); |
71f98027 | 3456 | |
f440ff44 | 3457 | if (strnlen(input, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) { |
71f98027 | 3458 | while ((timeout_setting = strsep(&input, ",")) && |
f440ff44 | 3459 | strnlen(timeout_setting, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) { |
71f98027 AD |
3460 | ret = kstrtol(timeout_setting, 0, &timeout); |
3461 | if (ret) | |
3462 | return ret; | |
3463 | ||
3464 | if (timeout == 0) { | |
3465 | index++; | |
3466 | continue; | |
3467 | } else if (timeout < 0) { | |
3468 | timeout = MAX_SCHEDULE_TIMEOUT; | |
127aedf9 CK |
3469 | dev_warn(adev->dev, "lockup timeout disabled"); |
3470 | add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK); | |
71f98027 AD |
3471 | } else { |
3472 | timeout = msecs_to_jiffies(timeout); | |
3473 | } | |
3474 | ||
3475 | switch (index++) { | |
3476 | case 0: | |
3477 | adev->gfx_timeout = timeout; | |
3478 | break; | |
3479 | case 1: | |
3480 | adev->compute_timeout = timeout; | |
3481 | break; | |
3482 | case 2: | |
3483 | adev->sdma_timeout = timeout; | |
3484 | break; | |
3485 | case 3: | |
3486 | adev->video_timeout = timeout; | |
3487 | break; | |
3488 | default: | |
3489 | break; | |
3490 | } | |
3491 | } | |
3492 | /* | |
3493 | * There is only one value specified and | |
3494 | * it should apply to all non-compute jobs. | |
3495 | */ | |
bcccee89 | 3496 | if (index == 1) { |
71f98027 | 3497 | adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout; |
bcccee89 ED |
3498 | if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev)) |
3499 | adev->compute_timeout = adev->gfx_timeout; | |
3500 | } | |
71f98027 AD |
3501 | } |
3502 | ||
3503 | return ret; | |
3504 | } | |
d4535e2c | 3505 | |
4a74c38c PY |
3506 | /** |
3507 | * amdgpu_device_check_iommu_direct_map - check if RAM direct mapped to GPU | |
3508 | * | |
3509 | * @adev: amdgpu_device pointer | |
3510 | * | |
3511 | * RAM direct mapped to GPU if IOMMU is not enabled or is pass through mode | |
3512 | */ | |
3513 | static void amdgpu_device_check_iommu_direct_map(struct amdgpu_device *adev) | |
3514 | { | |
3515 | struct iommu_domain *domain; | |
3516 | ||
3517 | domain = iommu_get_domain_for_dev(adev->dev); | |
3518 | if (!domain || domain->type == IOMMU_DOMAIN_IDENTITY) | |
3519 | adev->ram_is_direct_mapped = true; | |
3520 | } | |
3521 | ||
77f3a5cd ND |
3522 | static const struct attribute *amdgpu_dev_attributes[] = { |
3523 | &dev_attr_product_name.attr, | |
3524 | &dev_attr_product_number.attr, | |
3525 | &dev_attr_serial_number.attr, | |
3526 | &dev_attr_pcie_replay_count.attr, | |
3527 | NULL | |
3528 | }; | |
3529 | ||
d38ceaf9 AD |
3530 | /** |
3531 | * amdgpu_device_init - initialize the driver | |
3532 | * | |
3533 | * @adev: amdgpu_device pointer | |
d38ceaf9 AD |
3534 | * @flags: driver flags |
3535 | * | |
3536 | * Initializes the driver info and hw (all asics). | |
3537 | * Returns 0 for success or an error on failure. | |
3538 | * Called at driver startup. | |
3539 | */ | |
3540 | int amdgpu_device_init(struct amdgpu_device *adev, | |
d38ceaf9 AD |
3541 | uint32_t flags) |
3542 | { | |
8aba21b7 LT |
3543 | struct drm_device *ddev = adev_to_drm(adev); |
3544 | struct pci_dev *pdev = adev->pdev; | |
d38ceaf9 | 3545 | int r, i; |
b98c6299 | 3546 | bool px = false; |
95844d20 | 3547 | u32 max_MBps; |
d38ceaf9 AD |
3548 | |
3549 | adev->shutdown = false; | |
d38ceaf9 | 3550 | adev->flags = flags; |
4e66d7d2 YZ |
3551 | |
3552 | if (amdgpu_force_asic_type >= 0 && amdgpu_force_asic_type < CHIP_LAST) | |
3553 | adev->asic_type = amdgpu_force_asic_type; | |
3554 | else | |
3555 | adev->asic_type = flags & AMD_ASIC_MASK; | |
3556 | ||
d38ceaf9 | 3557 | adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT; |
593aa2d2 | 3558 | if (amdgpu_emu_mode == 1) |
8bdab6bb | 3559 | adev->usec_timeout *= 10; |
770d13b1 | 3560 | adev->gmc.gart_size = 512 * 1024 * 1024; |
d38ceaf9 AD |
3561 | adev->accel_working = false; |
3562 | adev->num_rings = 0; | |
68ce8b24 | 3563 | RCU_INIT_POINTER(adev->gang_submit, dma_fence_get_stub()); |
d38ceaf9 AD |
3564 | adev->mman.buffer_funcs = NULL; |
3565 | adev->mman.buffer_funcs_ring = NULL; | |
3566 | adev->vm_manager.vm_pte_funcs = NULL; | |
0c88b430 | 3567 | adev->vm_manager.vm_pte_num_scheds = 0; |
132f34e4 | 3568 | adev->gmc.gmc_funcs = NULL; |
7bd939d0 | 3569 | adev->harvest_ip_mask = 0x0; |
f54d1867 | 3570 | adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS); |
b8866c26 | 3571 | bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES); |
d38ceaf9 AD |
3572 | |
3573 | adev->smc_rreg = &amdgpu_invalid_rreg; | |
3574 | adev->smc_wreg = &amdgpu_invalid_wreg; | |
3575 | adev->pcie_rreg = &amdgpu_invalid_rreg; | |
3576 | adev->pcie_wreg = &amdgpu_invalid_wreg; | |
36b9a952 HR |
3577 | adev->pciep_rreg = &amdgpu_invalid_rreg; |
3578 | adev->pciep_wreg = &amdgpu_invalid_wreg; | |
4fa1c6a6 TZ |
3579 | adev->pcie_rreg64 = &amdgpu_invalid_rreg64; |
3580 | adev->pcie_wreg64 = &amdgpu_invalid_wreg64; | |
d38ceaf9 AD |
3581 | adev->uvd_ctx_rreg = &amdgpu_invalid_rreg; |
3582 | adev->uvd_ctx_wreg = &amdgpu_invalid_wreg; | |
3583 | adev->didt_rreg = &amdgpu_invalid_rreg; | |
3584 | adev->didt_wreg = &amdgpu_invalid_wreg; | |
ccdbb20a RZ |
3585 | adev->gc_cac_rreg = &amdgpu_invalid_rreg; |
3586 | adev->gc_cac_wreg = &amdgpu_invalid_wreg; | |
d38ceaf9 AD |
3587 | adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg; |
3588 | adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg; | |
3589 | ||
3e39ab90 AD |
3590 | DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n", |
3591 | amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device, | |
3592 | pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision); | |
d38ceaf9 AD |
3593 | |
3594 | /* mutex initialization are all done here so we | |
3595 | * can recall function without having locking issues */ | |
0e5ca0d1 | 3596 | mutex_init(&adev->firmware.mutex); |
d38ceaf9 AD |
3597 | mutex_init(&adev->pm.mutex); |
3598 | mutex_init(&adev->gfx.gpu_clock_mutex); | |
3599 | mutex_init(&adev->srbm_mutex); | |
b8866c26 | 3600 | mutex_init(&adev->gfx.pipe_reserve_mutex); |
d23ee13f | 3601 | mutex_init(&adev->gfx.gfx_off_mutex); |
d38ceaf9 | 3602 | mutex_init(&adev->grbm_idx_mutex); |
d38ceaf9 | 3603 | mutex_init(&adev->mn_lock); |
e23b74aa | 3604 | mutex_init(&adev->virt.vf_errors.lock); |
d38ceaf9 | 3605 | hash_init(adev->mn_hash); |
32eaeae0 | 3606 | mutex_init(&adev->psp.mutex); |
bd052211 | 3607 | mutex_init(&adev->notifier_lock); |
8cda7a4f | 3608 | mutex_init(&adev->pm.stable_pstate_ctx_lock); |
f113cc32 | 3609 | mutex_init(&adev->benchmark_mutex); |
d38ceaf9 | 3610 | |
ab3b9de6 | 3611 | amdgpu_device_init_apu_flags(adev); |
9f6a7857 | 3612 | |
912dfc84 EQ |
3613 | r = amdgpu_device_check_arguments(adev); |
3614 | if (r) | |
3615 | return r; | |
d38ceaf9 | 3616 | |
d38ceaf9 AD |
3617 | spin_lock_init(&adev->mmio_idx_lock); |
3618 | spin_lock_init(&adev->smc_idx_lock); | |
3619 | spin_lock_init(&adev->pcie_idx_lock); | |
3620 | spin_lock_init(&adev->uvd_ctx_idx_lock); | |
3621 | spin_lock_init(&adev->didt_idx_lock); | |
ccdbb20a | 3622 | spin_lock_init(&adev->gc_cac_idx_lock); |
16abb5d2 | 3623 | spin_lock_init(&adev->se_cac_idx_lock); |
d38ceaf9 | 3624 | spin_lock_init(&adev->audio_endpt_idx_lock); |
95844d20 | 3625 | spin_lock_init(&adev->mm_stats.lock); |
d38ceaf9 | 3626 | |
0c4e7fa5 CZ |
3627 | INIT_LIST_HEAD(&adev->shadow_list); |
3628 | mutex_init(&adev->shadow_list_lock); | |
3629 | ||
655ce9cb | 3630 | INIT_LIST_HEAD(&adev->reset_list); |
3631 | ||
6492e1b0 | 3632 | INIT_LIST_HEAD(&adev->ras_list); |
3633 | ||
beff74bc AD |
3634 | INIT_DELAYED_WORK(&adev->delayed_init_work, |
3635 | amdgpu_device_delayed_init_work_handler); | |
1e317b99 RZ |
3636 | INIT_DELAYED_WORK(&adev->gfx.gfx_off_delay_work, |
3637 | amdgpu_device_delay_enable_gfx_off); | |
2dc80b00 | 3638 | |
d4535e2c AG |
3639 | INIT_WORK(&adev->xgmi_reset_work, amdgpu_device_xgmi_reset_func); |
3640 | ||
d23ee13f | 3641 | adev->gfx.gfx_off_req_count = 1; |
0ad7347a AA |
3642 | adev->gfx.gfx_off_residency = 0; |
3643 | adev->gfx.gfx_off_entrycount = 0; | |
b6e79d9a | 3644 | adev->pm.ac_power = power_supply_is_system_supplied() > 0; |
b1ddf548 | 3645 | |
b265bdbd EQ |
3646 | atomic_set(&adev->throttling_logging_enabled, 1); |
3647 | /* | |
3648 | * If throttling continues, logging will be performed every minute | |
3649 | * to avoid log flooding. "-1" is subtracted since the thermal | |
3650 | * throttling interrupt comes every second. Thus, the total logging | |
3651 | * interval is 59 seconds(retelimited printk interval) + 1(waiting | |
3652 | * for throttling interrupt) = 60 seconds. | |
3653 | */ | |
3654 | ratelimit_state_init(&adev->throttling_logging_rs, (60 - 1) * HZ, 1); | |
3655 | ratelimit_set_flags(&adev->throttling_logging_rs, RATELIMIT_MSG_ON_RELEASE); | |
3656 | ||
0fa49558 AX |
3657 | /* Registers mapping */ |
3658 | /* TODO: block userspace mapping of io register */ | |
da69c161 KW |
3659 | if (adev->asic_type >= CHIP_BONAIRE) { |
3660 | adev->rmmio_base = pci_resource_start(adev->pdev, 5); | |
3661 | adev->rmmio_size = pci_resource_len(adev->pdev, 5); | |
3662 | } else { | |
3663 | adev->rmmio_base = pci_resource_start(adev->pdev, 2); | |
3664 | adev->rmmio_size = pci_resource_len(adev->pdev, 2); | |
3665 | } | |
d38ceaf9 | 3666 | |
6c08e0ef EQ |
3667 | for (i = 0; i < AMD_IP_BLOCK_TYPE_NUM; i++) |
3668 | atomic_set(&adev->pm.pwr_state[i], POWER_STATE_UNKNOWN); | |
3669 | ||
d38ceaf9 AD |
3670 | adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size); |
3671 | if (adev->rmmio == NULL) { | |
3672 | return -ENOMEM; | |
3673 | } | |
3674 | DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base); | |
3675 | DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size); | |
3676 | ||
5494d864 AD |
3677 | amdgpu_device_get_pcie_info(adev); |
3678 | ||
b239c017 JX |
3679 | if (amdgpu_mcbp) |
3680 | DRM_INFO("MCBP is enabled\n"); | |
3681 | ||
436afdfa PY |
3682 | /* |
3683 | * Reset domain needs to be present early, before XGMI hive discovered | |
3684 | * (if any) and intitialized to use reset sem and in_gpu reset flag | |
3685 | * early on during init and before calling to RREG32. | |
3686 | */ | |
3687 | adev->reset_domain = amdgpu_reset_create_reset_domain(SINGLE_DEVICE, "amdgpu-reset-dev"); | |
3688 | if (!adev->reset_domain) | |
3689 | return -ENOMEM; | |
3690 | ||
3aa0115d ML |
3691 | /* detect hw virtualization here */ |
3692 | amdgpu_detect_virtualization(adev); | |
3693 | ||
dffa11b4 ML |
3694 | r = amdgpu_device_get_job_timeout_settings(adev); |
3695 | if (r) { | |
3696 | dev_err(adev->dev, "invalid lockup_timeout parameter syntax\n"); | |
4ef87d8f | 3697 | return r; |
a190d1c7 XY |
3698 | } |
3699 | ||
d38ceaf9 | 3700 | /* early init functions */ |
06ec9070 | 3701 | r = amdgpu_device_ip_early_init(adev); |
d38ceaf9 | 3702 | if (r) |
4ef87d8f | 3703 | return r; |
d38ceaf9 | 3704 | |
b7cdb41e ML |
3705 | /* Get rid of things like offb */ |
3706 | r = drm_aperture_remove_conflicting_pci_framebuffers(adev->pdev, &amdgpu_kms_driver); | |
3707 | if (r) | |
3708 | return r; | |
3709 | ||
4d33e704 SK |
3710 | /* Enable TMZ based on IP_VERSION */ |
3711 | amdgpu_gmc_tmz_set(adev); | |
3712 | ||
957b0787 | 3713 | amdgpu_gmc_noretry_set(adev); |
4a0165f0 VS |
3714 | /* Need to get xgmi info early to decide the reset behavior*/ |
3715 | if (adev->gmc.xgmi.supported) { | |
3716 | r = adev->gfxhub.funcs->get_xgmi_info(adev); | |
3717 | if (r) | |
3718 | return r; | |
3719 | } | |
3720 | ||
8e6d0b69 | 3721 | /* enable PCIE atomic ops */ |
3722 | if (amdgpu_sriov_vf(adev)) | |
3723 | adev->have_atomics_support = ((struct amd_sriov_msg_pf2vf_info *) | |
e15c9d06 | 3724 | adev->virt.fw_reserve.p_pf2vf)->pcie_atomic_ops_support_flags == |
8e6d0b69 | 3725 | (PCI_EXP_DEVCAP2_ATOMIC_COMP32 | PCI_EXP_DEVCAP2_ATOMIC_COMP64); |
3726 | else | |
3727 | adev->have_atomics_support = | |
3728 | !pci_enable_atomic_ops_to_root(adev->pdev, | |
3729 | PCI_EXP_DEVCAP2_ATOMIC_COMP32 | | |
3730 | PCI_EXP_DEVCAP2_ATOMIC_COMP64); | |
3731 | if (!adev->have_atomics_support) | |
3732 | dev_info(adev->dev, "PCIE atomic ops is not supported\n"); | |
3733 | ||
6585661d OZ |
3734 | /* doorbell bar mapping and doorbell index init*/ |
3735 | amdgpu_device_doorbell_init(adev); | |
3736 | ||
9475a943 SL |
3737 | if (amdgpu_emu_mode == 1) { |
3738 | /* post the asic on emulation mode */ | |
3739 | emu_soc_asic_init(adev); | |
bfca0289 | 3740 | goto fence_driver_init; |
9475a943 | 3741 | } |
bfca0289 | 3742 | |
04442bf7 LL |
3743 | amdgpu_reset_init(adev); |
3744 | ||
4e99a44e ML |
3745 | /* detect if we are with an SRIOV vbios */ |
3746 | amdgpu_device_detect_sriov_bios(adev); | |
048765ad | 3747 | |
95e8e59e AD |
3748 | /* check if we need to reset the asic |
3749 | * E.g., driver was not cleanly unloaded previously, etc. | |
3750 | */ | |
f14899fd | 3751 | if (!amdgpu_sriov_vf(adev) && amdgpu_asic_need_reset_on_init(adev)) { |
e3c1b071 | 3752 | if (adev->gmc.xgmi.num_physical_nodes) { |
3753 | dev_info(adev->dev, "Pending hive reset.\n"); | |
3754 | adev->gmc.xgmi.pending_reset = true; | |
3755 | /* Only need to init necessary block for SMU to handle the reset */ | |
3756 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
3757 | if (!adev->ip_blocks[i].status.valid) | |
3758 | continue; | |
3759 | if (!(adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC || | |
3760 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || | |
3761 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH || | |
3762 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC)) { | |
751f43e7 | 3763 | DRM_DEBUG("IP %s disabled for hw_init.\n", |
e3c1b071 | 3764 | adev->ip_blocks[i].version->funcs->name); |
3765 | adev->ip_blocks[i].status.hw = true; | |
3766 | } | |
3767 | } | |
3768 | } else { | |
3769 | r = amdgpu_asic_reset(adev); | |
3770 | if (r) { | |
3771 | dev_err(adev->dev, "asic reset on init failed\n"); | |
3772 | goto failed; | |
3773 | } | |
95e8e59e AD |
3774 | } |
3775 | } | |
3776 | ||
d38ceaf9 | 3777 | /* Post card if necessary */ |
39c640c0 | 3778 | if (amdgpu_device_need_post(adev)) { |
d38ceaf9 | 3779 | if (!adev->bios) { |
bec86378 | 3780 | dev_err(adev->dev, "no vBIOS found\n"); |
83ba126a AD |
3781 | r = -EINVAL; |
3782 | goto failed; | |
d38ceaf9 | 3783 | } |
bec86378 | 3784 | DRM_INFO("GPU posting now...\n"); |
4d2997ab | 3785 | r = amdgpu_device_asic_init(adev); |
4e99a44e ML |
3786 | if (r) { |
3787 | dev_err(adev->dev, "gpu post error!\n"); | |
3788 | goto failed; | |
3789 | } | |
d38ceaf9 AD |
3790 | } |
3791 | ||
88b64e95 AD |
3792 | if (adev->is_atom_fw) { |
3793 | /* Initialize clocks */ | |
3794 | r = amdgpu_atomfirmware_get_clock_info(adev); | |
3795 | if (r) { | |
3796 | dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n"); | |
e23b74aa | 3797 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0); |
88b64e95 AD |
3798 | goto failed; |
3799 | } | |
3800 | } else { | |
a5bde2f9 AD |
3801 | /* Initialize clocks */ |
3802 | r = amdgpu_atombios_get_clock_info(adev); | |
3803 | if (r) { | |
3804 | dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n"); | |
e23b74aa | 3805 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0); |
89041940 | 3806 | goto failed; |
a5bde2f9 AD |
3807 | } |
3808 | /* init i2c buses */ | |
4562236b HW |
3809 | if (!amdgpu_device_has_dc_support(adev)) |
3810 | amdgpu_atombios_i2c_init(adev); | |
2c1a2784 | 3811 | } |
d38ceaf9 | 3812 | |
bfca0289 | 3813 | fence_driver_init: |
d38ceaf9 | 3814 | /* Fence driver */ |
067f44c8 | 3815 | r = amdgpu_fence_driver_sw_init(adev); |
2c1a2784 | 3816 | if (r) { |
067f44c8 | 3817 | dev_err(adev->dev, "amdgpu_fence_driver_sw_init failed\n"); |
e23b74aa | 3818 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0); |
83ba126a | 3819 | goto failed; |
2c1a2784 | 3820 | } |
d38ceaf9 AD |
3821 | |
3822 | /* init the mode config */ | |
4a580877 | 3823 | drm_mode_config_init(adev_to_drm(adev)); |
d38ceaf9 | 3824 | |
06ec9070 | 3825 | r = amdgpu_device_ip_init(adev); |
d38ceaf9 | 3826 | if (r) { |
8840a387 | 3827 | /* failed in exclusive mode due to timeout */ |
3828 | if (amdgpu_sriov_vf(adev) && | |
3829 | !amdgpu_sriov_runtime(adev) && | |
3830 | amdgpu_virt_mmio_blocked(adev) && | |
3831 | !amdgpu_virt_wait_reset(adev)) { | |
3832 | dev_err(adev->dev, "VF exclusive mode timeout\n"); | |
1daee8b4 PD |
3833 | /* Don't send request since VF is inactive. */ |
3834 | adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME; | |
3835 | adev->virt.ops = NULL; | |
8840a387 | 3836 | r = -EAGAIN; |
970fd197 | 3837 | goto release_ras_con; |
8840a387 | 3838 | } |
06ec9070 | 3839 | dev_err(adev->dev, "amdgpu_device_ip_init failed\n"); |
e23b74aa | 3840 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0); |
970fd197 | 3841 | goto release_ras_con; |
d38ceaf9 AD |
3842 | } |
3843 | ||
8d35a259 LG |
3844 | amdgpu_fence_driver_hw_init(adev); |
3845 | ||
d69b8971 YZ |
3846 | dev_info(adev->dev, |
3847 | "SE %d, SH per SE %d, CU per SH %d, active_cu_number %d\n", | |
d7f72fe4 YZ |
3848 | adev->gfx.config.max_shader_engines, |
3849 | adev->gfx.config.max_sh_per_se, | |
3850 | adev->gfx.config.max_cu_per_sh, | |
3851 | adev->gfx.cu_info.number); | |
3852 | ||
d38ceaf9 AD |
3853 | adev->accel_working = true; |
3854 | ||
e59c0205 AX |
3855 | amdgpu_vm_check_compute_bug(adev); |
3856 | ||
95844d20 MO |
3857 | /* Initialize the buffer migration limit. */ |
3858 | if (amdgpu_moverate >= 0) | |
3859 | max_MBps = amdgpu_moverate; | |
3860 | else | |
3861 | max_MBps = 8; /* Allow 8 MB/s. */ | |
3862 | /* Get a log2 for easy divisions. */ | |
3863 | adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps)); | |
3864 | ||
d2f52ac8 | 3865 | r = amdgpu_pm_sysfs_init(adev); |
7c868b59 YT |
3866 | if (r) { |
3867 | adev->pm_sysfs_en = false; | |
d2f52ac8 | 3868 | DRM_ERROR("registering pm debugfs failed (%d).\n", r); |
7c868b59 YT |
3869 | } else |
3870 | adev->pm_sysfs_en = true; | |
d2f52ac8 | 3871 | |
5bb23532 | 3872 | r = amdgpu_ucode_sysfs_init(adev); |
7c868b59 YT |
3873 | if (r) { |
3874 | adev->ucode_sysfs_en = false; | |
5bb23532 | 3875 | DRM_ERROR("Creating firmware sysfs failed (%d).\n", r); |
7c868b59 YT |
3876 | } else |
3877 | adev->ucode_sysfs_en = true; | |
5bb23532 | 3878 | |
8424f2cc LG |
3879 | r = amdgpu_psp_sysfs_init(adev); |
3880 | if (r) { | |
3881 | adev->psp_sysfs_en = false; | |
3882 | if (!amdgpu_sriov_vf(adev)) | |
3883 | DRM_ERROR("Creating psp sysfs failed\n"); | |
3884 | } else | |
3885 | adev->psp_sysfs_en = true; | |
3886 | ||
b0adca4d EQ |
3887 | /* |
3888 | * Register gpu instance before amdgpu_device_enable_mgpu_fan_boost. | |
3889 | * Otherwise the mgpu fan boost feature will be skipped due to the | |
3890 | * gpu instance is counted less. | |
3891 | */ | |
3892 | amdgpu_register_gpu_instance(adev); | |
3893 | ||
d38ceaf9 AD |
3894 | /* enable clockgating, etc. after ib tests, etc. since some blocks require |
3895 | * explicit gating rather than handling it automatically. | |
3896 | */ | |
e3c1b071 | 3897 | if (!adev->gmc.xgmi.pending_reset) { |
3898 | r = amdgpu_device_ip_late_init(adev); | |
3899 | if (r) { | |
3900 | dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n"); | |
3901 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r); | |
970fd197 | 3902 | goto release_ras_con; |
e3c1b071 | 3903 | } |
3904 | /* must succeed. */ | |
3905 | amdgpu_ras_resume(adev); | |
3906 | queue_delayed_work(system_wq, &adev->delayed_init_work, | |
3907 | msecs_to_jiffies(AMDGPU_RESUME_MS)); | |
2c1a2784 | 3908 | } |
d38ceaf9 | 3909 | |
2c738637 ML |
3910 | if (amdgpu_sriov_vf(adev)) |
3911 | flush_delayed_work(&adev->delayed_init_work); | |
3912 | ||
77f3a5cd | 3913 | r = sysfs_create_files(&adev->dev->kobj, amdgpu_dev_attributes); |
5aea5327 | 3914 | if (r) |
77f3a5cd | 3915 | dev_err(adev->dev, "Could not create amdgpu device attr\n"); |
bd607166 | 3916 | |
d155bef0 AB |
3917 | if (IS_ENABLED(CONFIG_PERF_EVENTS)) |
3918 | r = amdgpu_pmu_init(adev); | |
9c7c85f7 JK |
3919 | if (r) |
3920 | dev_err(adev->dev, "amdgpu_pmu_init failed\n"); | |
3921 | ||
c1dd4aa6 AG |
3922 | /* Have stored pci confspace at hand for restore in sudden PCI error */ |
3923 | if (amdgpu_device_cache_pci_state(adev->pdev)) | |
3924 | pci_restore_state(pdev); | |
3925 | ||
8c3dd61c KHF |
3926 | /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */ |
3927 | /* this will fail for cards that aren't VGA class devices, just | |
3928 | * ignore it */ | |
3929 | if ((adev->pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA) | |
bf44e8ce | 3930 | vga_client_register(adev->pdev, amdgpu_device_vga_set_decode); |
8c3dd61c | 3931 | |
d37a3929 OC |
3932 | px = amdgpu_device_supports_px(ddev); |
3933 | ||
3934 | if (px || (!pci_is_thunderbolt_attached(adev->pdev) && | |
3935 | apple_gmux_detect(NULL, NULL))) | |
8c3dd61c KHF |
3936 | vga_switcheroo_register_client(adev->pdev, |
3937 | &amdgpu_switcheroo_ops, px); | |
d37a3929 OC |
3938 | |
3939 | if (px) | |
8c3dd61c | 3940 | vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain); |
8c3dd61c | 3941 | |
e3c1b071 | 3942 | if (adev->gmc.xgmi.pending_reset) |
3943 | queue_delayed_work(system_wq, &mgpu_info.delayed_reset_work, | |
3944 | msecs_to_jiffies(AMDGPU_RESUME_MS)); | |
3945 | ||
4a74c38c PY |
3946 | amdgpu_device_check_iommu_direct_map(adev); |
3947 | ||
d38ceaf9 | 3948 | return 0; |
83ba126a | 3949 | |
970fd197 SY |
3950 | release_ras_con: |
3951 | amdgpu_release_ras_context(adev); | |
3952 | ||
83ba126a | 3953 | failed: |
89041940 | 3954 | amdgpu_vf_error_trans_all(adev); |
8840a387 | 3955 | |
83ba126a | 3956 | return r; |
d38ceaf9 AD |
3957 | } |
3958 | ||
07775fc1 AG |
3959 | static void amdgpu_device_unmap_mmio(struct amdgpu_device *adev) |
3960 | { | |
62d5f9f7 | 3961 | |
07775fc1 AG |
3962 | /* Clear all CPU mappings pointing to this device */ |
3963 | unmap_mapping_range(adev->ddev.anon_inode->i_mapping, 0, 0, 1); | |
3964 | ||
3965 | /* Unmap all mapped bars - Doorbell, registers and VRAM */ | |
3966 | amdgpu_device_doorbell_fini(adev); | |
3967 | ||
3968 | iounmap(adev->rmmio); | |
3969 | adev->rmmio = NULL; | |
3970 | if (adev->mman.aper_base_kaddr) | |
3971 | iounmap(adev->mman.aper_base_kaddr); | |
3972 | adev->mman.aper_base_kaddr = NULL; | |
3973 | ||
3974 | /* Memory manager related */ | |
3975 | if (!adev->gmc.xgmi.connected_to_cpu) { | |
3976 | arch_phys_wc_del(adev->gmc.vram_mtrr); | |
3977 | arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size); | |
3978 | } | |
3979 | } | |
3980 | ||
d38ceaf9 | 3981 | /** |
bbe04dec | 3982 | * amdgpu_device_fini_hw - tear down the driver |
d38ceaf9 AD |
3983 | * |
3984 | * @adev: amdgpu_device pointer | |
3985 | * | |
3986 | * Tear down the driver info (all asics). | |
3987 | * Called at driver shutdown. | |
3988 | */ | |
72c8c97b | 3989 | void amdgpu_device_fini_hw(struct amdgpu_device *adev) |
d38ceaf9 | 3990 | { |
aac89168 | 3991 | dev_info(adev->dev, "amdgpu: finishing device.\n"); |
9f875167 | 3992 | flush_delayed_work(&adev->delayed_init_work); |
d0d13fe8 | 3993 | adev->shutdown = true; |
9f875167 | 3994 | |
752c683d ML |
3995 | /* make sure IB test finished before entering exclusive mode |
3996 | * to avoid preemption on IB test | |
3997 | * */ | |
519b8b76 | 3998 | if (amdgpu_sriov_vf(adev)) { |
752c683d | 3999 | amdgpu_virt_request_full_gpu(adev, false); |
519b8b76 BZ |
4000 | amdgpu_virt_fini_data_exchange(adev); |
4001 | } | |
752c683d | 4002 | |
e5b03032 ML |
4003 | /* disable all interrupts */ |
4004 | amdgpu_irq_disable_all(adev); | |
ff97cba8 | 4005 | if (adev->mode_info.mode_config_initialized){ |
1053b9c9 | 4006 | if (!drm_drv_uses_atomic_modeset(adev_to_drm(adev))) |
4a580877 | 4007 | drm_helper_force_disable_all(adev_to_drm(adev)); |
ff97cba8 | 4008 | else |
4a580877 | 4009 | drm_atomic_helper_shutdown(adev_to_drm(adev)); |
ff97cba8 | 4010 | } |
8d35a259 | 4011 | amdgpu_fence_driver_hw_fini(adev); |
72c8c97b | 4012 | |
cd3a8a59 | 4013 | if (adev->mman.initialized) |
9bff18d1 | 4014 | drain_workqueue(adev->mman.bdev.wq); |
98f56188 | 4015 | |
7c868b59 YT |
4016 | if (adev->pm_sysfs_en) |
4017 | amdgpu_pm_sysfs_fini(adev); | |
72c8c97b AG |
4018 | if (adev->ucode_sysfs_en) |
4019 | amdgpu_ucode_sysfs_fini(adev); | |
8424f2cc LG |
4020 | if (adev->psp_sysfs_en) |
4021 | amdgpu_psp_sysfs_fini(adev); | |
72c8c97b AG |
4022 | sysfs_remove_files(&adev->dev->kobj, amdgpu_dev_attributes); |
4023 | ||
232d1d43 SY |
4024 | /* disable ras feature must before hw fini */ |
4025 | amdgpu_ras_pre_fini(adev); | |
4026 | ||
e9669fb7 | 4027 | amdgpu_device_ip_fini_early(adev); |
d10d0daa | 4028 | |
a3848df6 YW |
4029 | amdgpu_irq_fini_hw(adev); |
4030 | ||
b6fd6e0f SK |
4031 | if (adev->mman.initialized) |
4032 | ttm_device_clear_dma_mappings(&adev->mman.bdev); | |
894c6890 | 4033 | |
d10d0daa | 4034 | amdgpu_gart_dummy_page_fini(adev); |
07775fc1 | 4035 | |
39934d3e VP |
4036 | if (drm_dev_is_unplugged(adev_to_drm(adev))) |
4037 | amdgpu_device_unmap_mmio(adev); | |
87172e89 | 4038 | |
72c8c97b AG |
4039 | } |
4040 | ||
4041 | void amdgpu_device_fini_sw(struct amdgpu_device *adev) | |
4042 | { | |
62d5f9f7 | 4043 | int idx; |
d37a3929 | 4044 | bool px; |
62d5f9f7 | 4045 | |
8d35a259 | 4046 | amdgpu_fence_driver_sw_fini(adev); |
a5c5d8d5 | 4047 | amdgpu_device_ip_fini(adev); |
b31d3063 | 4048 | amdgpu_ucode_release(&adev->firmware.gpu_info_fw); |
d38ceaf9 | 4049 | adev->accel_working = false; |
68ce8b24 | 4050 | dma_fence_put(rcu_dereference_protected(adev->gang_submit, true)); |
04442bf7 LL |
4051 | |
4052 | amdgpu_reset_fini(adev); | |
4053 | ||
d38ceaf9 | 4054 | /* free i2c buses */ |
4562236b HW |
4055 | if (!amdgpu_device_has_dc_support(adev)) |
4056 | amdgpu_i2c_fini(adev); | |
bfca0289 SL |
4057 | |
4058 | if (amdgpu_emu_mode != 1) | |
4059 | amdgpu_atombios_fini(adev); | |
4060 | ||
d38ceaf9 AD |
4061 | kfree(adev->bios); |
4062 | adev->bios = NULL; | |
d37a3929 OC |
4063 | |
4064 | px = amdgpu_device_supports_px(adev_to_drm(adev)); | |
4065 | ||
4066 | if (px || (!pci_is_thunderbolt_attached(adev->pdev) && | |
4067 | apple_gmux_detect(NULL, NULL))) | |
84c8b22e | 4068 | vga_switcheroo_unregister_client(adev->pdev); |
d37a3929 OC |
4069 | |
4070 | if (px) | |
83ba126a | 4071 | vga_switcheroo_fini_domain_pm_ops(adev->dev); |
d37a3929 | 4072 | |
38d6be81 | 4073 | if ((adev->pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA) |
b8779475 | 4074 | vga_client_unregister(adev->pdev); |
e9bc1bf7 | 4075 | |
62d5f9f7 LS |
4076 | if (drm_dev_enter(adev_to_drm(adev), &idx)) { |
4077 | ||
4078 | iounmap(adev->rmmio); | |
4079 | adev->rmmio = NULL; | |
4080 | amdgpu_device_doorbell_fini(adev); | |
4081 | drm_dev_exit(idx); | |
4082 | } | |
4083 | ||
d155bef0 AB |
4084 | if (IS_ENABLED(CONFIG_PERF_EVENTS)) |
4085 | amdgpu_pmu_fini(adev); | |
72de33f8 | 4086 | if (adev->mman.discovery_bin) |
a190d1c7 | 4087 | amdgpu_discovery_fini(adev); |
72c8c97b | 4088 | |
cfbb6b00 AG |
4089 | amdgpu_reset_put_reset_domain(adev->reset_domain); |
4090 | adev->reset_domain = NULL; | |
4091 | ||
72c8c97b AG |
4092 | kfree(adev->pci_state); |
4093 | ||
d38ceaf9 AD |
4094 | } |
4095 | ||
58144d28 ND |
4096 | /** |
4097 | * amdgpu_device_evict_resources - evict device resources | |
4098 | * @adev: amdgpu device object | |
4099 | * | |
4100 | * Evicts all ttm device resources(vram BOs, gart table) from the lru list | |
4101 | * of the vram memory type. Mainly used for evicting device resources | |
4102 | * at suspend time. | |
4103 | * | |
4104 | */ | |
7863c155 | 4105 | static int amdgpu_device_evict_resources(struct amdgpu_device *adev) |
58144d28 | 4106 | { |
7863c155 ML |
4107 | int ret; |
4108 | ||
e53d9665 ML |
4109 | /* No need to evict vram on APUs for suspend to ram or s2idle */ |
4110 | if ((adev->in_s3 || adev->in_s0ix) && (adev->flags & AMD_IS_APU)) | |
7863c155 | 4111 | return 0; |
58144d28 | 4112 | |
7863c155 ML |
4113 | ret = amdgpu_ttm_evict_resources(adev, TTM_PL_VRAM); |
4114 | if (ret) | |
58144d28 | 4115 | DRM_WARN("evicting device resources failed\n"); |
7863c155 | 4116 | return ret; |
58144d28 | 4117 | } |
d38ceaf9 AD |
4118 | |
4119 | /* | |
4120 | * Suspend & resume. | |
4121 | */ | |
4122 | /** | |
810ddc3a | 4123 | * amdgpu_device_suspend - initiate device suspend |
d38ceaf9 | 4124 | * |
87e3f136 | 4125 | * @dev: drm dev pointer |
87e3f136 | 4126 | * @fbcon : notify the fbdev of suspend |
d38ceaf9 AD |
4127 | * |
4128 | * Puts the hw in the suspend state (all asics). | |
4129 | * Returns 0 for success or an error on failure. | |
4130 | * Called at driver suspend. | |
4131 | */ | |
de185019 | 4132 | int amdgpu_device_suspend(struct drm_device *dev, bool fbcon) |
d38ceaf9 | 4133 | { |
a2e15b0e | 4134 | struct amdgpu_device *adev = drm_to_adev(dev); |
d7274ec7 | 4135 | int r = 0; |
d38ceaf9 | 4136 | |
d38ceaf9 AD |
4137 | if (dev->switch_power_state == DRM_SWITCH_POWER_OFF) |
4138 | return 0; | |
4139 | ||
44779b43 | 4140 | adev->in_suspend = true; |
3fa8f89d | 4141 | |
47ea2076 SF |
4142 | /* Evict the majority of BOs before grabbing the full access */ |
4143 | r = amdgpu_device_evict_resources(adev); | |
4144 | if (r) | |
4145 | return r; | |
4146 | ||
d7274ec7 BZ |
4147 | if (amdgpu_sriov_vf(adev)) { |
4148 | amdgpu_virt_fini_data_exchange(adev); | |
4149 | r = amdgpu_virt_request_full_gpu(adev, false); | |
4150 | if (r) | |
4151 | return r; | |
4152 | } | |
4153 | ||
3fa8f89d S |
4154 | if (amdgpu_acpi_smart_shift_update(dev, AMDGPU_SS_DEV_D3)) |
4155 | DRM_WARN("smart shift update failed\n"); | |
4156 | ||
d38ceaf9 AD |
4157 | drm_kms_helper_poll_disable(dev); |
4158 | ||
5f818173 | 4159 | if (fbcon) |
087451f3 | 4160 | drm_fb_helper_set_suspend_unlocked(adev_to_drm(adev)->fb_helper, true); |
5f818173 | 4161 | |
beff74bc | 4162 | cancel_delayed_work_sync(&adev->delayed_init_work); |
a5459475 | 4163 | |
5e6932fe | 4164 | amdgpu_ras_suspend(adev); |
4165 | ||
2196927b | 4166 | amdgpu_device_ip_suspend_phase1(adev); |
fe1053b7 | 4167 | |
c004d44e | 4168 | if (!adev->in_s0ix) |
5d3a2d95 | 4169 | amdgpu_amdkfd_suspend(adev, adev->in_runpm); |
94fa5660 | 4170 | |
7863c155 ML |
4171 | r = amdgpu_device_evict_resources(adev); |
4172 | if (r) | |
4173 | return r; | |
d38ceaf9 | 4174 | |
8d35a259 | 4175 | amdgpu_fence_driver_hw_fini(adev); |
d38ceaf9 | 4176 | |
2196927b | 4177 | amdgpu_device_ip_suspend_phase2(adev); |
d38ceaf9 | 4178 | |
d7274ec7 BZ |
4179 | if (amdgpu_sriov_vf(adev)) |
4180 | amdgpu_virt_release_full_gpu(adev, false); | |
4181 | ||
d38ceaf9 AD |
4182 | return 0; |
4183 | } | |
4184 | ||
4185 | /** | |
810ddc3a | 4186 | * amdgpu_device_resume - initiate device resume |
d38ceaf9 | 4187 | * |
87e3f136 | 4188 | * @dev: drm dev pointer |
87e3f136 | 4189 | * @fbcon : notify the fbdev of resume |
d38ceaf9 AD |
4190 | * |
4191 | * Bring the hw back to operating state (all asics). | |
4192 | * Returns 0 for success or an error on failure. | |
4193 | * Called at driver resume. | |
4194 | */ | |
de185019 | 4195 | int amdgpu_device_resume(struct drm_device *dev, bool fbcon) |
d38ceaf9 | 4196 | { |
1348969a | 4197 | struct amdgpu_device *adev = drm_to_adev(dev); |
03161a6e | 4198 | int r = 0; |
d38ceaf9 | 4199 | |
d7274ec7 BZ |
4200 | if (amdgpu_sriov_vf(adev)) { |
4201 | r = amdgpu_virt_request_full_gpu(adev, true); | |
4202 | if (r) | |
4203 | return r; | |
4204 | } | |
4205 | ||
d38ceaf9 AD |
4206 | if (dev->switch_power_state == DRM_SWITCH_POWER_OFF) |
4207 | return 0; | |
4208 | ||
62498733 | 4209 | if (adev->in_s0ix) |
bc143d8b | 4210 | amdgpu_dpm_gfx_state_change(adev, sGpuChangeState_D0Entry); |
628c36d7 | 4211 | |
d38ceaf9 | 4212 | /* post card */ |
39c640c0 | 4213 | if (amdgpu_device_need_post(adev)) { |
4d2997ab | 4214 | r = amdgpu_device_asic_init(adev); |
74b0b157 | 4215 | if (r) |
aac89168 | 4216 | dev_err(adev->dev, "amdgpu asic init failed\n"); |
74b0b157 | 4217 | } |
d38ceaf9 | 4218 | |
06ec9070 | 4219 | r = amdgpu_device_ip_resume(adev); |
d7274ec7 | 4220 | |
e6707218 | 4221 | if (r) { |
aac89168 | 4222 | dev_err(adev->dev, "amdgpu_device_ip_resume failed (%d).\n", r); |
3c22c1ea | 4223 | goto exit; |
e6707218 | 4224 | } |
8d35a259 | 4225 | amdgpu_fence_driver_hw_init(adev); |
5ceb54c6 | 4226 | |
06ec9070 | 4227 | r = amdgpu_device_ip_late_init(adev); |
03161a6e | 4228 | if (r) |
3c22c1ea | 4229 | goto exit; |
d38ceaf9 | 4230 | |
beff74bc AD |
4231 | queue_delayed_work(system_wq, &adev->delayed_init_work, |
4232 | msecs_to_jiffies(AMDGPU_RESUME_MS)); | |
4233 | ||
c004d44e | 4234 | if (!adev->in_s0ix) { |
5d3a2d95 AD |
4235 | r = amdgpu_amdkfd_resume(adev, adev->in_runpm); |
4236 | if (r) | |
3c22c1ea | 4237 | goto exit; |
5d3a2d95 | 4238 | } |
756e6880 | 4239 | |
3c22c1ea SF |
4240 | exit: |
4241 | if (amdgpu_sriov_vf(adev)) { | |
4242 | amdgpu_virt_init_data_exchange(adev); | |
4243 | amdgpu_virt_release_full_gpu(adev, true); | |
4244 | } | |
4245 | ||
4246 | if (r) | |
4247 | return r; | |
4248 | ||
96a5d8d4 | 4249 | /* Make sure IB tests flushed */ |
beff74bc | 4250 | flush_delayed_work(&adev->delayed_init_work); |
96a5d8d4 | 4251 | |
a2e15b0e | 4252 | if (fbcon) |
087451f3 | 4253 | drm_fb_helper_set_suspend_unlocked(adev_to_drm(adev)->fb_helper, false); |
d38ceaf9 AD |
4254 | |
4255 | drm_kms_helper_poll_enable(dev); | |
23a1a9e5 | 4256 | |
5e6932fe | 4257 | amdgpu_ras_resume(adev); |
4258 | ||
d09ef243 AD |
4259 | if (adev->mode_info.num_crtc) { |
4260 | /* | |
4261 | * Most of the connector probing functions try to acquire runtime pm | |
4262 | * refs to ensure that the GPU is powered on when connector polling is | |
4263 | * performed. Since we're calling this from a runtime PM callback, | |
4264 | * trying to acquire rpm refs will cause us to deadlock. | |
4265 | * | |
4266 | * Since we're guaranteed to be holding the rpm lock, it's safe to | |
4267 | * temporarily disable the rpm helpers so this doesn't deadlock us. | |
4268 | */ | |
23a1a9e5 | 4269 | #ifdef CONFIG_PM |
d09ef243 | 4270 | dev->dev->power.disable_depth++; |
23a1a9e5 | 4271 | #endif |
d09ef243 AD |
4272 | if (!adev->dc_enabled) |
4273 | drm_helper_hpd_irq_event(dev); | |
4274 | else | |
4275 | drm_kms_helper_hotplug_event(dev); | |
23a1a9e5 | 4276 | #ifdef CONFIG_PM |
d09ef243 | 4277 | dev->dev->power.disable_depth--; |
23a1a9e5 | 4278 | #endif |
d09ef243 | 4279 | } |
44779b43 RZ |
4280 | adev->in_suspend = false; |
4281 | ||
dc907c9d JX |
4282 | if (adev->enable_mes) |
4283 | amdgpu_mes_self_test(adev); | |
4284 | ||
3fa8f89d S |
4285 | if (amdgpu_acpi_smart_shift_update(dev, AMDGPU_SS_DEV_D0)) |
4286 | DRM_WARN("smart shift update failed\n"); | |
4287 | ||
4d3b9ae5 | 4288 | return 0; |
d38ceaf9 AD |
4289 | } |
4290 | ||
e3ecdffa AD |
4291 | /** |
4292 | * amdgpu_device_ip_check_soft_reset - did soft reset succeed | |
4293 | * | |
4294 | * @adev: amdgpu_device pointer | |
4295 | * | |
4296 | * The list of all the hardware IPs that make up the asic is walked and | |
4297 | * the check_soft_reset callbacks are run. check_soft_reset determines | |
4298 | * if the asic is still hung or not. | |
4299 | * Returns true if any of the IPs are still in a hung state, false if not. | |
4300 | */ | |
06ec9070 | 4301 | static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev) |
63fbf42f CZ |
4302 | { |
4303 | int i; | |
4304 | bool asic_hang = false; | |
4305 | ||
f993d628 ML |
4306 | if (amdgpu_sriov_vf(adev)) |
4307 | return true; | |
4308 | ||
8bc04c29 AD |
4309 | if (amdgpu_asic_need_full_reset(adev)) |
4310 | return true; | |
4311 | ||
63fbf42f | 4312 | for (i = 0; i < adev->num_ip_blocks; i++) { |
a1255107 | 4313 | if (!adev->ip_blocks[i].status.valid) |
63fbf42f | 4314 | continue; |
a1255107 AD |
4315 | if (adev->ip_blocks[i].version->funcs->check_soft_reset) |
4316 | adev->ip_blocks[i].status.hang = | |
4317 | adev->ip_blocks[i].version->funcs->check_soft_reset(adev); | |
4318 | if (adev->ip_blocks[i].status.hang) { | |
aac89168 | 4319 | dev_info(adev->dev, "IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name); |
63fbf42f CZ |
4320 | asic_hang = true; |
4321 | } | |
4322 | } | |
4323 | return asic_hang; | |
4324 | } | |
4325 | ||
e3ecdffa AD |
4326 | /** |
4327 | * amdgpu_device_ip_pre_soft_reset - prepare for soft reset | |
4328 | * | |
4329 | * @adev: amdgpu_device pointer | |
4330 | * | |
4331 | * The list of all the hardware IPs that make up the asic is walked and the | |
4332 | * pre_soft_reset callbacks are run if the block is hung. pre_soft_reset | |
4333 | * handles any IP specific hardware or software state changes that are | |
4334 | * necessary for a soft reset to succeed. | |
4335 | * Returns 0 on success, negative error code on failure. | |
4336 | */ | |
06ec9070 | 4337 | static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev) |
d31a501e CZ |
4338 | { |
4339 | int i, r = 0; | |
4340 | ||
4341 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 4342 | if (!adev->ip_blocks[i].status.valid) |
d31a501e | 4343 | continue; |
a1255107 AD |
4344 | if (adev->ip_blocks[i].status.hang && |
4345 | adev->ip_blocks[i].version->funcs->pre_soft_reset) { | |
4346 | r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev); | |
d31a501e CZ |
4347 | if (r) |
4348 | return r; | |
4349 | } | |
4350 | } | |
4351 | ||
4352 | return 0; | |
4353 | } | |
4354 | ||
e3ecdffa AD |
4355 | /** |
4356 | * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed | |
4357 | * | |
4358 | * @adev: amdgpu_device pointer | |
4359 | * | |
4360 | * Some hardware IPs cannot be soft reset. If they are hung, a full gpu | |
4361 | * reset is necessary to recover. | |
4362 | * Returns true if a full asic reset is required, false if not. | |
4363 | */ | |
06ec9070 | 4364 | static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev) |
35d782fe | 4365 | { |
da146d3b AD |
4366 | int i; |
4367 | ||
8bc04c29 AD |
4368 | if (amdgpu_asic_need_full_reset(adev)) |
4369 | return true; | |
4370 | ||
da146d3b | 4371 | for (i = 0; i < adev->num_ip_blocks; i++) { |
a1255107 | 4372 | if (!adev->ip_blocks[i].status.valid) |
da146d3b | 4373 | continue; |
a1255107 AD |
4374 | if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) || |
4375 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) || | |
4376 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) || | |
98512bb8 KW |
4377 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) || |
4378 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) { | |
a1255107 | 4379 | if (adev->ip_blocks[i].status.hang) { |
aac89168 | 4380 | dev_info(adev->dev, "Some block need full reset!\n"); |
da146d3b AD |
4381 | return true; |
4382 | } | |
4383 | } | |
35d782fe CZ |
4384 | } |
4385 | return false; | |
4386 | } | |
4387 | ||
e3ecdffa AD |
4388 | /** |
4389 | * amdgpu_device_ip_soft_reset - do a soft reset | |
4390 | * | |
4391 | * @adev: amdgpu_device pointer | |
4392 | * | |
4393 | * The list of all the hardware IPs that make up the asic is walked and the | |
4394 | * soft_reset callbacks are run if the block is hung. soft_reset handles any | |
4395 | * IP specific hardware or software state changes that are necessary to soft | |
4396 | * reset the IP. | |
4397 | * Returns 0 on success, negative error code on failure. | |
4398 | */ | |
06ec9070 | 4399 | static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev) |
35d782fe CZ |
4400 | { |
4401 | int i, r = 0; | |
4402 | ||
4403 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 4404 | if (!adev->ip_blocks[i].status.valid) |
35d782fe | 4405 | continue; |
a1255107 AD |
4406 | if (adev->ip_blocks[i].status.hang && |
4407 | adev->ip_blocks[i].version->funcs->soft_reset) { | |
4408 | r = adev->ip_blocks[i].version->funcs->soft_reset(adev); | |
35d782fe CZ |
4409 | if (r) |
4410 | return r; | |
4411 | } | |
4412 | } | |
4413 | ||
4414 | return 0; | |
4415 | } | |
4416 | ||
e3ecdffa AD |
4417 | /** |
4418 | * amdgpu_device_ip_post_soft_reset - clean up from soft reset | |
4419 | * | |
4420 | * @adev: amdgpu_device pointer | |
4421 | * | |
4422 | * The list of all the hardware IPs that make up the asic is walked and the | |
4423 | * post_soft_reset callbacks are run if the asic was hung. post_soft_reset | |
4424 | * handles any IP specific hardware or software state changes that are | |
4425 | * necessary after the IP has been soft reset. | |
4426 | * Returns 0 on success, negative error code on failure. | |
4427 | */ | |
06ec9070 | 4428 | static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev) |
35d782fe CZ |
4429 | { |
4430 | int i, r = 0; | |
4431 | ||
4432 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 4433 | if (!adev->ip_blocks[i].status.valid) |
35d782fe | 4434 | continue; |
a1255107 AD |
4435 | if (adev->ip_blocks[i].status.hang && |
4436 | adev->ip_blocks[i].version->funcs->post_soft_reset) | |
4437 | r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev); | |
35d782fe CZ |
4438 | if (r) |
4439 | return r; | |
4440 | } | |
4441 | ||
4442 | return 0; | |
4443 | } | |
4444 | ||
e3ecdffa | 4445 | /** |
c33adbc7 | 4446 | * amdgpu_device_recover_vram - Recover some VRAM contents |
e3ecdffa AD |
4447 | * |
4448 | * @adev: amdgpu_device pointer | |
4449 | * | |
4450 | * Restores the contents of VRAM buffers from the shadows in GTT. Used to | |
4451 | * restore things like GPUVM page tables after a GPU reset where | |
4452 | * the contents of VRAM might be lost. | |
403009bf CK |
4453 | * |
4454 | * Returns: | |
4455 | * 0 on success, negative error code on failure. | |
e3ecdffa | 4456 | */ |
c33adbc7 | 4457 | static int amdgpu_device_recover_vram(struct amdgpu_device *adev) |
c41d1cf6 | 4458 | { |
c41d1cf6 | 4459 | struct dma_fence *fence = NULL, *next = NULL; |
403009bf | 4460 | struct amdgpu_bo *shadow; |
e18aaea7 | 4461 | struct amdgpu_bo_vm *vmbo; |
403009bf | 4462 | long r = 1, tmo; |
c41d1cf6 ML |
4463 | |
4464 | if (amdgpu_sriov_runtime(adev)) | |
b045d3af | 4465 | tmo = msecs_to_jiffies(8000); |
c41d1cf6 ML |
4466 | else |
4467 | tmo = msecs_to_jiffies(100); | |
4468 | ||
aac89168 | 4469 | dev_info(adev->dev, "recover vram bo from shadow start\n"); |
c41d1cf6 | 4470 | mutex_lock(&adev->shadow_list_lock); |
e18aaea7 ND |
4471 | list_for_each_entry(vmbo, &adev->shadow_list, shadow_list) { |
4472 | shadow = &vmbo->bo; | |
403009bf | 4473 | /* No need to recover an evicted BO */ |
d3116756 CK |
4474 | if (shadow->tbo.resource->mem_type != TTM_PL_TT || |
4475 | shadow->tbo.resource->start == AMDGPU_BO_INVALID_OFFSET || | |
4476 | shadow->parent->tbo.resource->mem_type != TTM_PL_VRAM) | |
403009bf CK |
4477 | continue; |
4478 | ||
4479 | r = amdgpu_bo_restore_shadow(shadow, &next); | |
4480 | if (r) | |
4481 | break; | |
4482 | ||
c41d1cf6 | 4483 | if (fence) { |
1712fb1a | 4484 | tmo = dma_fence_wait_timeout(fence, false, tmo); |
403009bf CK |
4485 | dma_fence_put(fence); |
4486 | fence = next; | |
1712fb1a | 4487 | if (tmo == 0) { |
4488 | r = -ETIMEDOUT; | |
c41d1cf6 | 4489 | break; |
1712fb1a | 4490 | } else if (tmo < 0) { |
4491 | r = tmo; | |
4492 | break; | |
4493 | } | |
403009bf CK |
4494 | } else { |
4495 | fence = next; | |
c41d1cf6 | 4496 | } |
c41d1cf6 ML |
4497 | } |
4498 | mutex_unlock(&adev->shadow_list_lock); | |
4499 | ||
403009bf CK |
4500 | if (fence) |
4501 | tmo = dma_fence_wait_timeout(fence, false, tmo); | |
c41d1cf6 ML |
4502 | dma_fence_put(fence); |
4503 | ||
1712fb1a | 4504 | if (r < 0 || tmo <= 0) { |
aac89168 | 4505 | dev_err(adev->dev, "recover vram bo from shadow failed, r is %ld, tmo is %ld\n", r, tmo); |
403009bf CK |
4506 | return -EIO; |
4507 | } | |
c41d1cf6 | 4508 | |
aac89168 | 4509 | dev_info(adev->dev, "recover vram bo from shadow done\n"); |
403009bf | 4510 | return 0; |
c41d1cf6 ML |
4511 | } |
4512 | ||
a90ad3c2 | 4513 | |
e3ecdffa | 4514 | /** |
06ec9070 | 4515 | * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf |
5740682e | 4516 | * |
982a820b | 4517 | * @adev: amdgpu_device pointer |
87e3f136 | 4518 | * @from_hypervisor: request from hypervisor |
5740682e ML |
4519 | * |
4520 | * do VF FLR and reinitialize Asic | |
3f48c681 | 4521 | * return 0 means succeeded otherwise failed |
e3ecdffa AD |
4522 | */ |
4523 | static int amdgpu_device_reset_sriov(struct amdgpu_device *adev, | |
4524 | bool from_hypervisor) | |
5740682e ML |
4525 | { |
4526 | int r; | |
a5f67c93 | 4527 | struct amdgpu_hive_info *hive = NULL; |
7258fa31 | 4528 | int retry_limit = 0; |
5740682e | 4529 | |
7258fa31 | 4530 | retry: |
c004d44e | 4531 | amdgpu_amdkfd_pre_reset(adev); |
428890a3 | 4532 | |
5740682e ML |
4533 | if (from_hypervisor) |
4534 | r = amdgpu_virt_request_full_gpu(adev, true); | |
4535 | else | |
4536 | r = amdgpu_virt_reset_gpu(adev); | |
4537 | if (r) | |
4538 | return r; | |
a90ad3c2 ML |
4539 | |
4540 | /* Resume IP prior to SMC */ | |
06ec9070 | 4541 | r = amdgpu_device_ip_reinit_early_sriov(adev); |
5740682e ML |
4542 | if (r) |
4543 | goto error; | |
a90ad3c2 | 4544 | |
c9ffa427 | 4545 | amdgpu_virt_init_data_exchange(adev); |
a90ad3c2 | 4546 | |
7a3e0bb2 RZ |
4547 | r = amdgpu_device_fw_loading(adev); |
4548 | if (r) | |
4549 | return r; | |
4550 | ||
a90ad3c2 | 4551 | /* now we are okay to resume SMC/CP/SDMA */ |
06ec9070 | 4552 | r = amdgpu_device_ip_reinit_late_sriov(adev); |
5740682e ML |
4553 | if (r) |
4554 | goto error; | |
a90ad3c2 | 4555 | |
a5f67c93 ZL |
4556 | hive = amdgpu_get_xgmi_hive(adev); |
4557 | /* Update PSP FW topology after reset */ | |
4558 | if (hive && adev->gmc.xgmi.num_physical_nodes > 1) | |
4559 | r = amdgpu_xgmi_update_topology(hive, adev); | |
4560 | ||
4561 | if (hive) | |
4562 | amdgpu_put_xgmi_hive(hive); | |
4563 | ||
4564 | if (!r) { | |
4565 | amdgpu_irq_gpu_reset_resume_helper(adev); | |
4566 | r = amdgpu_ib_ring_tests(adev); | |
9c12f5cd | 4567 | |
c004d44e | 4568 | amdgpu_amdkfd_post_reset(adev); |
a5f67c93 | 4569 | } |
a90ad3c2 | 4570 | |
abc34253 | 4571 | error: |
c41d1cf6 | 4572 | if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) { |
e3526257 | 4573 | amdgpu_inc_vram_lost(adev); |
c33adbc7 | 4574 | r = amdgpu_device_recover_vram(adev); |
a90ad3c2 | 4575 | } |
437f3e0b | 4576 | amdgpu_virt_release_full_gpu(adev, true); |
a90ad3c2 | 4577 | |
7258fa31 SK |
4578 | if (AMDGPU_RETRY_SRIOV_RESET(r)) { |
4579 | if (retry_limit < AMDGPU_MAX_RETRY_LIMIT) { | |
4580 | retry_limit++; | |
4581 | goto retry; | |
4582 | } else | |
4583 | DRM_ERROR("GPU reset retry is beyond the retry limit\n"); | |
4584 | } | |
4585 | ||
a90ad3c2 ML |
4586 | return r; |
4587 | } | |
4588 | ||
9a1cddd6 | 4589 | /** |
4590 | * amdgpu_device_has_job_running - check if there is any job in mirror list | |
4591 | * | |
982a820b | 4592 | * @adev: amdgpu_device pointer |
9a1cddd6 | 4593 | * |
4594 | * check if there is any job in mirror list | |
4595 | */ | |
4596 | bool amdgpu_device_has_job_running(struct amdgpu_device *adev) | |
4597 | { | |
4598 | int i; | |
4599 | struct drm_sched_job *job; | |
4600 | ||
4601 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { | |
4602 | struct amdgpu_ring *ring = adev->rings[i]; | |
4603 | ||
4604 | if (!ring || !ring->sched.thread) | |
4605 | continue; | |
4606 | ||
4607 | spin_lock(&ring->sched.job_list_lock); | |
6efa4b46 LT |
4608 | job = list_first_entry_or_null(&ring->sched.pending_list, |
4609 | struct drm_sched_job, list); | |
9a1cddd6 | 4610 | spin_unlock(&ring->sched.job_list_lock); |
4611 | if (job) | |
4612 | return true; | |
4613 | } | |
4614 | return false; | |
4615 | } | |
4616 | ||
12938fad CK |
4617 | /** |
4618 | * amdgpu_device_should_recover_gpu - check if we should try GPU recovery | |
4619 | * | |
982a820b | 4620 | * @adev: amdgpu_device pointer |
12938fad CK |
4621 | * |
4622 | * Check amdgpu_gpu_recovery and SRIOV status to see if we should try to recover | |
4623 | * a hung GPU. | |
4624 | */ | |
4625 | bool amdgpu_device_should_recover_gpu(struct amdgpu_device *adev) | |
4626 | { | |
12938fad | 4627 | |
3ba7b418 AG |
4628 | if (amdgpu_gpu_recovery == 0) |
4629 | goto disabled; | |
4630 | ||
1a11a65d YC |
4631 | /* Skip soft reset check in fatal error mode */ |
4632 | if (!amdgpu_ras_is_poison_mode_supported(adev)) | |
4633 | return true; | |
4634 | ||
3ba7b418 AG |
4635 | if (amdgpu_sriov_vf(adev)) |
4636 | return true; | |
4637 | ||
4638 | if (amdgpu_gpu_recovery == -1) { | |
4639 | switch (adev->asic_type) { | |
b3523c45 AD |
4640 | #ifdef CONFIG_DRM_AMDGPU_SI |
4641 | case CHIP_VERDE: | |
4642 | case CHIP_TAHITI: | |
4643 | case CHIP_PITCAIRN: | |
4644 | case CHIP_OLAND: | |
4645 | case CHIP_HAINAN: | |
4646 | #endif | |
4647 | #ifdef CONFIG_DRM_AMDGPU_CIK | |
4648 | case CHIP_KAVERI: | |
4649 | case CHIP_KABINI: | |
4650 | case CHIP_MULLINS: | |
4651 | #endif | |
4652 | case CHIP_CARRIZO: | |
4653 | case CHIP_STONEY: | |
4654 | case CHIP_CYAN_SKILLFISH: | |
3ba7b418 | 4655 | goto disabled; |
b3523c45 AD |
4656 | default: |
4657 | break; | |
3ba7b418 | 4658 | } |
12938fad CK |
4659 | } |
4660 | ||
4661 | return true; | |
3ba7b418 AG |
4662 | |
4663 | disabled: | |
aac89168 | 4664 | dev_info(adev->dev, "GPU recovery disabled.\n"); |
3ba7b418 | 4665 | return false; |
12938fad CK |
4666 | } |
4667 | ||
5c03e584 FX |
4668 | int amdgpu_device_mode1_reset(struct amdgpu_device *adev) |
4669 | { | |
4670 | u32 i; | |
4671 | int ret = 0; | |
4672 | ||
4673 | amdgpu_atombios_scratch_regs_engine_hung(adev, true); | |
4674 | ||
4675 | dev_info(adev->dev, "GPU mode1 reset\n"); | |
4676 | ||
4677 | /* disable BM */ | |
4678 | pci_clear_master(adev->pdev); | |
4679 | ||
4680 | amdgpu_device_cache_pci_state(adev->pdev); | |
4681 | ||
4682 | if (amdgpu_dpm_is_mode1_reset_supported(adev)) { | |
4683 | dev_info(adev->dev, "GPU smu mode1 reset\n"); | |
4684 | ret = amdgpu_dpm_mode1_reset(adev); | |
4685 | } else { | |
4686 | dev_info(adev->dev, "GPU psp mode1 reset\n"); | |
4687 | ret = psp_gpu_reset(adev); | |
4688 | } | |
4689 | ||
4690 | if (ret) | |
4691 | dev_err(adev->dev, "GPU mode1 reset failed\n"); | |
4692 | ||
4693 | amdgpu_device_load_pci_state(adev->pdev); | |
4694 | ||
4695 | /* wait for asic to come out of reset */ | |
4696 | for (i = 0; i < adev->usec_timeout; i++) { | |
4697 | u32 memsize = adev->nbio.funcs->get_memsize(adev); | |
4698 | ||
4699 | if (memsize != 0xffffffff) | |
4700 | break; | |
4701 | udelay(1); | |
4702 | } | |
4703 | ||
4704 | amdgpu_atombios_scratch_regs_engine_hung(adev, false); | |
4705 | return ret; | |
4706 | } | |
5c6dd71e | 4707 | |
e3c1b071 | 4708 | int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, |
04442bf7 | 4709 | struct amdgpu_reset_context *reset_context) |
26bc5340 | 4710 | { |
5c1e6fa4 | 4711 | int i, r = 0; |
04442bf7 LL |
4712 | struct amdgpu_job *job = NULL; |
4713 | bool need_full_reset = | |
4714 | test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); | |
4715 | ||
4716 | if (reset_context->reset_req_dev == adev) | |
4717 | job = reset_context->job; | |
71182665 | 4718 | |
b602ca5f TZ |
4719 | if (amdgpu_sriov_vf(adev)) { |
4720 | /* stop the data exchange thread */ | |
4721 | amdgpu_virt_fini_data_exchange(adev); | |
4722 | } | |
4723 | ||
9e225fb9 AG |
4724 | amdgpu_fence_driver_isr_toggle(adev, true); |
4725 | ||
71182665 | 4726 | /* block all schedulers and reset given job's ring */ |
0875dc9e CZ |
4727 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { |
4728 | struct amdgpu_ring *ring = adev->rings[i]; | |
4729 | ||
51687759 | 4730 | if (!ring || !ring->sched.thread) |
0875dc9e | 4731 | continue; |
5740682e | 4732 | |
c530b02f JZ |
4733 | /*clear job fence from fence drv to avoid force_completion |
4734 | *leave NULL and vm flush fence in fence drv */ | |
5c1e6fa4 | 4735 | amdgpu_fence_driver_clear_job_fences(ring); |
c530b02f | 4736 | |
2f9d4084 ML |
4737 | /* after all hw jobs are reset, hw fence is meaningless, so force_completion */ |
4738 | amdgpu_fence_driver_force_completion(ring); | |
0875dc9e | 4739 | } |
d38ceaf9 | 4740 | |
9e225fb9 AG |
4741 | amdgpu_fence_driver_isr_toggle(adev, false); |
4742 | ||
ff99849b | 4743 | if (job && job->vm) |
222b5f04 AG |
4744 | drm_sched_increase_karma(&job->base); |
4745 | ||
04442bf7 | 4746 | r = amdgpu_reset_prepare_hwcontext(adev, reset_context); |
404b277b LL |
4747 | /* If reset handler not implemented, continue; otherwise return */ |
4748 | if (r == -ENOSYS) | |
4749 | r = 0; | |
4750 | else | |
04442bf7 LL |
4751 | return r; |
4752 | ||
1d721ed6 | 4753 | /* Don't suspend on bare metal if we are not going to HW reset the ASIC */ |
26bc5340 AG |
4754 | if (!amdgpu_sriov_vf(adev)) { |
4755 | ||
4756 | if (!need_full_reset) | |
4757 | need_full_reset = amdgpu_device_ip_need_full_reset(adev); | |
4758 | ||
360cd081 LG |
4759 | if (!need_full_reset && amdgpu_gpu_recovery && |
4760 | amdgpu_device_ip_check_soft_reset(adev)) { | |
26bc5340 AG |
4761 | amdgpu_device_ip_pre_soft_reset(adev); |
4762 | r = amdgpu_device_ip_soft_reset(adev); | |
4763 | amdgpu_device_ip_post_soft_reset(adev); | |
4764 | if (r || amdgpu_device_ip_check_soft_reset(adev)) { | |
aac89168 | 4765 | dev_info(adev->dev, "soft reset failed, will fallback to full reset!\n"); |
26bc5340 AG |
4766 | need_full_reset = true; |
4767 | } | |
4768 | } | |
4769 | ||
4770 | if (need_full_reset) | |
4771 | r = amdgpu_device_ip_suspend(adev); | |
04442bf7 LL |
4772 | if (need_full_reset) |
4773 | set_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); | |
4774 | else | |
4775 | clear_bit(AMDGPU_NEED_FULL_RESET, | |
4776 | &reset_context->flags); | |
26bc5340 AG |
4777 | } |
4778 | ||
4779 | return r; | |
4780 | } | |
4781 | ||
15fd09a0 SA |
4782 | static int amdgpu_reset_reg_dumps(struct amdgpu_device *adev) |
4783 | { | |
15fd09a0 SA |
4784 | int i; |
4785 | ||
38a15ad9 | 4786 | lockdep_assert_held(&adev->reset_domain->sem); |
15fd09a0 SA |
4787 | |
4788 | for (i = 0; i < adev->num_regs; i++) { | |
651d7ee6 SA |
4789 | adev->reset_dump_reg_value[i] = RREG32(adev->reset_dump_reg_list[i]); |
4790 | trace_amdgpu_reset_reg_dumps(adev->reset_dump_reg_list[i], | |
4791 | adev->reset_dump_reg_value[i]); | |
15fd09a0 SA |
4792 | } |
4793 | ||
4794 | return 0; | |
4795 | } | |
4796 | ||
3d8785f6 SA |
4797 | #ifdef CONFIG_DEV_COREDUMP |
4798 | static ssize_t amdgpu_devcoredump_read(char *buffer, loff_t offset, | |
4799 | size_t count, void *data, size_t datalen) | |
4800 | { | |
4801 | struct drm_printer p; | |
4802 | struct amdgpu_device *adev = data; | |
4803 | struct drm_print_iterator iter; | |
4804 | int i; | |
4805 | ||
4806 | iter.data = buffer; | |
4807 | iter.offset = 0; | |
4808 | iter.start = offset; | |
4809 | iter.remain = count; | |
4810 | ||
4811 | p = drm_coredump_printer(&iter); | |
4812 | ||
4813 | drm_printf(&p, "**** AMDGPU Device Coredump ****\n"); | |
4814 | drm_printf(&p, "kernel: " UTS_RELEASE "\n"); | |
4815 | drm_printf(&p, "module: " KBUILD_MODNAME "\n"); | |
4816 | drm_printf(&p, "time: %lld.%09ld\n", adev->reset_time.tv_sec, adev->reset_time.tv_nsec); | |
4817 | if (adev->reset_task_info.pid) | |
4818 | drm_printf(&p, "process_name: %s PID: %d\n", | |
4819 | adev->reset_task_info.process_name, | |
4820 | adev->reset_task_info.pid); | |
4821 | ||
4822 | if (adev->reset_vram_lost) | |
4823 | drm_printf(&p, "VRAM is lost due to GPU reset!\n"); | |
4824 | if (adev->num_regs) { | |
4825 | drm_printf(&p, "AMDGPU register dumps:\nOffset: Value:\n"); | |
4826 | ||
4827 | for (i = 0; i < adev->num_regs; i++) | |
4828 | drm_printf(&p, "0x%08x: 0x%08x\n", | |
4829 | adev->reset_dump_reg_list[i], | |
4830 | adev->reset_dump_reg_value[i]); | |
4831 | } | |
4832 | ||
4833 | return count - iter.remain; | |
4834 | } | |
4835 | ||
4836 | static void amdgpu_devcoredump_free(void *data) | |
4837 | { | |
4838 | } | |
4839 | ||
4840 | static void amdgpu_reset_capture_coredumpm(struct amdgpu_device *adev) | |
4841 | { | |
4842 | struct drm_device *dev = adev_to_drm(adev); | |
4843 | ||
4844 | ktime_get_ts64(&adev->reset_time); | |
4845 | dev_coredumpm(dev->dev, THIS_MODULE, adev, 0, GFP_KERNEL, | |
4846 | amdgpu_devcoredump_read, amdgpu_devcoredump_free); | |
4847 | } | |
4848 | #endif | |
4849 | ||
04442bf7 LL |
4850 | int amdgpu_do_asic_reset(struct list_head *device_list_handle, |
4851 | struct amdgpu_reset_context *reset_context) | |
26bc5340 AG |
4852 | { |
4853 | struct amdgpu_device *tmp_adev = NULL; | |
04442bf7 | 4854 | bool need_full_reset, skip_hw_reset, vram_lost = false; |
26bc5340 | 4855 | int r = 0; |
f5c7e779 | 4856 | bool gpu_reset_for_dev_remove = 0; |
26bc5340 | 4857 | |
04442bf7 LL |
4858 | /* Try reset handler method first */ |
4859 | tmp_adev = list_first_entry(device_list_handle, struct amdgpu_device, | |
4860 | reset_list); | |
15fd09a0 | 4861 | amdgpu_reset_reg_dumps(tmp_adev); |
0a83bb35 LL |
4862 | |
4863 | reset_context->reset_device_list = device_list_handle; | |
04442bf7 | 4864 | r = amdgpu_reset_perform_reset(tmp_adev, reset_context); |
404b277b LL |
4865 | /* If reset handler not implemented, continue; otherwise return */ |
4866 | if (r == -ENOSYS) | |
4867 | r = 0; | |
4868 | else | |
04442bf7 LL |
4869 | return r; |
4870 | ||
4871 | /* Reset handler not implemented, use the default method */ | |
4872 | need_full_reset = | |
4873 | test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); | |
4874 | skip_hw_reset = test_bit(AMDGPU_SKIP_HW_RESET, &reset_context->flags); | |
4875 | ||
f5c7e779 YC |
4876 | gpu_reset_for_dev_remove = |
4877 | test_bit(AMDGPU_RESET_FOR_DEVICE_REMOVE, &reset_context->flags) && | |
4878 | test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); | |
4879 | ||
26bc5340 | 4880 | /* |
655ce9cb | 4881 | * ASIC reset has to be done on all XGMI hive nodes ASAP |
26bc5340 AG |
4882 | * to allow proper links negotiation in FW (within 1 sec) |
4883 | */ | |
7ac71382 | 4884 | if (!skip_hw_reset && need_full_reset) { |
655ce9cb | 4885 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
041a62bc | 4886 | /* For XGMI run all resets in parallel to speed up the process */ |
d4535e2c | 4887 | if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) { |
e3c1b071 | 4888 | tmp_adev->gmc.xgmi.pending_reset = false; |
c96cf282 | 4889 | if (!queue_work(system_unbound_wq, &tmp_adev->xgmi_reset_work)) |
d4535e2c AG |
4890 | r = -EALREADY; |
4891 | } else | |
4892 | r = amdgpu_asic_reset(tmp_adev); | |
d4535e2c | 4893 | |
041a62bc | 4894 | if (r) { |
aac89168 | 4895 | dev_err(tmp_adev->dev, "ASIC reset failed with error, %d for drm dev, %s", |
4a580877 | 4896 | r, adev_to_drm(tmp_adev)->unique); |
041a62bc | 4897 | break; |
ce316fa5 LM |
4898 | } |
4899 | } | |
4900 | ||
041a62bc AG |
4901 | /* For XGMI wait for all resets to complete before proceed */ |
4902 | if (!r) { | |
655ce9cb | 4903 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
ce316fa5 LM |
4904 | if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) { |
4905 | flush_work(&tmp_adev->xgmi_reset_work); | |
4906 | r = tmp_adev->asic_reset_res; | |
4907 | if (r) | |
4908 | break; | |
ce316fa5 LM |
4909 | } |
4910 | } | |
4911 | } | |
ce316fa5 | 4912 | } |
26bc5340 | 4913 | |
43c4d576 | 4914 | if (!r && amdgpu_ras_intr_triggered()) { |
655ce9cb | 4915 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
5e67bba3 | 4916 | if (tmp_adev->mmhub.ras && tmp_adev->mmhub.ras->ras_block.hw_ops && |
4917 | tmp_adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count) | |
4918 | tmp_adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count(tmp_adev); | |
43c4d576 JC |
4919 | } |
4920 | ||
00eaa571 | 4921 | amdgpu_ras_intr_cleared(); |
43c4d576 | 4922 | } |
00eaa571 | 4923 | |
f5c7e779 YC |
4924 | /* Since the mode1 reset affects base ip blocks, the |
4925 | * phase1 ip blocks need to be resumed. Otherwise there | |
4926 | * will be a BIOS signature error and the psp bootloader | |
4927 | * can't load kdb on the next amdgpu install. | |
4928 | */ | |
4929 | if (gpu_reset_for_dev_remove) { | |
4930 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) | |
4931 | amdgpu_device_ip_resume_phase1(tmp_adev); | |
4932 | ||
4933 | goto end; | |
4934 | } | |
4935 | ||
655ce9cb | 4936 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
26bc5340 AG |
4937 | if (need_full_reset) { |
4938 | /* post card */ | |
e3c1b071 | 4939 | r = amdgpu_device_asic_init(tmp_adev); |
4940 | if (r) { | |
aac89168 | 4941 | dev_warn(tmp_adev->dev, "asic atom init failed!"); |
e3c1b071 | 4942 | } else { |
26bc5340 | 4943 | dev_info(tmp_adev->dev, "GPU reset succeeded, trying to resume\n"); |
9cec53c1 JZ |
4944 | r = amdgpu_amdkfd_resume_iommu(tmp_adev); |
4945 | if (r) | |
4946 | goto out; | |
4947 | ||
26bc5340 AG |
4948 | r = amdgpu_device_ip_resume_phase1(tmp_adev); |
4949 | if (r) | |
4950 | goto out; | |
4951 | ||
4952 | vram_lost = amdgpu_device_check_vram_lost(tmp_adev); | |
3d8785f6 SA |
4953 | #ifdef CONFIG_DEV_COREDUMP |
4954 | tmp_adev->reset_vram_lost = vram_lost; | |
4955 | memset(&tmp_adev->reset_task_info, 0, | |
4956 | sizeof(tmp_adev->reset_task_info)); | |
4957 | if (reset_context->job && reset_context->job->vm) | |
4958 | tmp_adev->reset_task_info = | |
4959 | reset_context->job->vm->task_info; | |
4960 | amdgpu_reset_capture_coredumpm(tmp_adev); | |
4961 | #endif | |
26bc5340 | 4962 | if (vram_lost) { |
77e7f829 | 4963 | DRM_INFO("VRAM is lost due to GPU reset!\n"); |
e3526257 | 4964 | amdgpu_inc_vram_lost(tmp_adev); |
26bc5340 AG |
4965 | } |
4966 | ||
26bc5340 AG |
4967 | r = amdgpu_device_fw_loading(tmp_adev); |
4968 | if (r) | |
4969 | return r; | |
4970 | ||
4971 | r = amdgpu_device_ip_resume_phase2(tmp_adev); | |
4972 | if (r) | |
4973 | goto out; | |
4974 | ||
4975 | if (vram_lost) | |
4976 | amdgpu_device_fill_reset_magic(tmp_adev); | |
4977 | ||
fdafb359 EQ |
4978 | /* |
4979 | * Add this ASIC as tracked as reset was already | |
4980 | * complete successfully. | |
4981 | */ | |
4982 | amdgpu_register_gpu_instance(tmp_adev); | |
4983 | ||
04442bf7 LL |
4984 | if (!reset_context->hive && |
4985 | tmp_adev->gmc.xgmi.num_physical_nodes > 1) | |
e3c1b071 | 4986 | amdgpu_xgmi_add_device(tmp_adev); |
4987 | ||
7c04ca50 | 4988 | r = amdgpu_device_ip_late_init(tmp_adev); |
4989 | if (r) | |
4990 | goto out; | |
4991 | ||
087451f3 | 4992 | drm_fb_helper_set_suspend_unlocked(adev_to_drm(tmp_adev)->fb_helper, false); |
565d1941 | 4993 | |
e8fbaf03 GC |
4994 | /* |
4995 | * The GPU enters bad state once faulty pages | |
4996 | * by ECC has reached the threshold, and ras | |
4997 | * recovery is scheduled next. So add one check | |
4998 | * here to break recovery if it indeed exceeds | |
4999 | * bad page threshold, and remind user to | |
5000 | * retire this GPU or setting one bigger | |
5001 | * bad_page_threshold value to fix this once | |
5002 | * probing driver again. | |
5003 | */ | |
11003c68 | 5004 | if (!amdgpu_ras_eeprom_check_err_threshold(tmp_adev)) { |
e8fbaf03 GC |
5005 | /* must succeed. */ |
5006 | amdgpu_ras_resume(tmp_adev); | |
5007 | } else { | |
5008 | r = -EINVAL; | |
5009 | goto out; | |
5010 | } | |
e79a04d5 | 5011 | |
26bc5340 | 5012 | /* Update PSP FW topology after reset */ |
04442bf7 LL |
5013 | if (reset_context->hive && |
5014 | tmp_adev->gmc.xgmi.num_physical_nodes > 1) | |
5015 | r = amdgpu_xgmi_update_topology( | |
5016 | reset_context->hive, tmp_adev); | |
26bc5340 AG |
5017 | } |
5018 | } | |
5019 | ||
26bc5340 AG |
5020 | out: |
5021 | if (!r) { | |
5022 | amdgpu_irq_gpu_reset_resume_helper(tmp_adev); | |
5023 | r = amdgpu_ib_ring_tests(tmp_adev); | |
5024 | if (r) { | |
5025 | dev_err(tmp_adev->dev, "ib ring test failed (%d).\n", r); | |
26bc5340 AG |
5026 | need_full_reset = true; |
5027 | r = -EAGAIN; | |
5028 | goto end; | |
5029 | } | |
5030 | } | |
5031 | ||
5032 | if (!r) | |
5033 | r = amdgpu_device_recover_vram(tmp_adev); | |
5034 | else | |
5035 | tmp_adev->asic_reset_res = r; | |
5036 | } | |
5037 | ||
5038 | end: | |
04442bf7 LL |
5039 | if (need_full_reset) |
5040 | set_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); | |
5041 | else | |
5042 | clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); | |
26bc5340 AG |
5043 | return r; |
5044 | } | |
5045 | ||
e923be99 | 5046 | static void amdgpu_device_set_mp1_state(struct amdgpu_device *adev) |
26bc5340 | 5047 | { |
5740682e | 5048 | |
a3a09142 AD |
5049 | switch (amdgpu_asic_reset_method(adev)) { |
5050 | case AMD_RESET_METHOD_MODE1: | |
5051 | adev->mp1_state = PP_MP1_STATE_SHUTDOWN; | |
5052 | break; | |
5053 | case AMD_RESET_METHOD_MODE2: | |
5054 | adev->mp1_state = PP_MP1_STATE_RESET; | |
5055 | break; | |
5056 | default: | |
5057 | adev->mp1_state = PP_MP1_STATE_NONE; | |
5058 | break; | |
5059 | } | |
26bc5340 | 5060 | } |
d38ceaf9 | 5061 | |
e923be99 | 5062 | static void amdgpu_device_unset_mp1_state(struct amdgpu_device *adev) |
26bc5340 | 5063 | { |
89041940 | 5064 | amdgpu_vf_error_trans_all(adev); |
a3a09142 | 5065 | adev->mp1_state = PP_MP1_STATE_NONE; |
91fb309d HC |
5066 | } |
5067 | ||
3f12acc8 EQ |
5068 | static void amdgpu_device_resume_display_audio(struct amdgpu_device *adev) |
5069 | { | |
5070 | struct pci_dev *p = NULL; | |
5071 | ||
5072 | p = pci_get_domain_bus_and_slot(pci_domain_nr(adev->pdev->bus), | |
5073 | adev->pdev->bus->number, 1); | |
5074 | if (p) { | |
5075 | pm_runtime_enable(&(p->dev)); | |
5076 | pm_runtime_resume(&(p->dev)); | |
5077 | } | |
b85e285e YY |
5078 | |
5079 | pci_dev_put(p); | |
3f12acc8 EQ |
5080 | } |
5081 | ||
5082 | static int amdgpu_device_suspend_display_audio(struct amdgpu_device *adev) | |
5083 | { | |
5084 | enum amd_reset_method reset_method; | |
5085 | struct pci_dev *p = NULL; | |
5086 | u64 expires; | |
5087 | ||
5088 | /* | |
5089 | * For now, only BACO and mode1 reset are confirmed | |
5090 | * to suffer the audio issue without proper suspended. | |
5091 | */ | |
5092 | reset_method = amdgpu_asic_reset_method(adev); | |
5093 | if ((reset_method != AMD_RESET_METHOD_BACO) && | |
5094 | (reset_method != AMD_RESET_METHOD_MODE1)) | |
5095 | return -EINVAL; | |
5096 | ||
5097 | p = pci_get_domain_bus_and_slot(pci_domain_nr(adev->pdev->bus), | |
5098 | adev->pdev->bus->number, 1); | |
5099 | if (!p) | |
5100 | return -ENODEV; | |
5101 | ||
5102 | expires = pm_runtime_autosuspend_expiration(&(p->dev)); | |
5103 | if (!expires) | |
5104 | /* | |
5105 | * If we cannot get the audio device autosuspend delay, | |
5106 | * a fixed 4S interval will be used. Considering 3S is | |
5107 | * the audio controller default autosuspend delay setting. | |
5108 | * 4S used here is guaranteed to cover that. | |
5109 | */ | |
54b7feb9 | 5110 | expires = ktime_get_mono_fast_ns() + NSEC_PER_SEC * 4ULL; |
3f12acc8 EQ |
5111 | |
5112 | while (!pm_runtime_status_suspended(&(p->dev))) { | |
5113 | if (!pm_runtime_suspend(&(p->dev))) | |
5114 | break; | |
5115 | ||
5116 | if (expires < ktime_get_mono_fast_ns()) { | |
5117 | dev_warn(adev->dev, "failed to suspend display audio\n"); | |
b85e285e | 5118 | pci_dev_put(p); |
3f12acc8 EQ |
5119 | /* TODO: abort the succeeding gpu reset? */ |
5120 | return -ETIMEDOUT; | |
5121 | } | |
5122 | } | |
5123 | ||
5124 | pm_runtime_disable(&(p->dev)); | |
5125 | ||
b85e285e | 5126 | pci_dev_put(p); |
3f12acc8 EQ |
5127 | return 0; |
5128 | } | |
5129 | ||
d193b12b | 5130 | static inline void amdgpu_device_stop_pending_resets(struct amdgpu_device *adev) |
247c7b0d AG |
5131 | { |
5132 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
5133 | ||
5134 | #if defined(CONFIG_DEBUG_FS) | |
5135 | if (!amdgpu_sriov_vf(adev)) | |
5136 | cancel_work(&adev->reset_work); | |
5137 | #endif | |
5138 | ||
5139 | if (adev->kfd.dev) | |
5140 | cancel_work(&adev->kfd.reset_work); | |
5141 | ||
5142 | if (amdgpu_sriov_vf(adev)) | |
5143 | cancel_work(&adev->virt.flr_work); | |
5144 | ||
5145 | if (con && adev->ras_enabled) | |
5146 | cancel_work(&con->recovery_work); | |
5147 | ||
5148 | } | |
5149 | ||
26bc5340 | 5150 | /** |
6e9c65f7 | 5151 | * amdgpu_device_gpu_recover - reset the asic and recover scheduler |
26bc5340 | 5152 | * |
982a820b | 5153 | * @adev: amdgpu_device pointer |
26bc5340 AG |
5154 | * @job: which job trigger hang |
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 SY |
5323 | |
5324 | /* Aldebaran supports ras in SRIOV, so need resume ras during reset */ | |
5325 | if (adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 2)) | |
5326 | amdgpu_ras_resume(adev); | |
26bc5340 | 5327 | } else { |
f1549c09 | 5328 | r = amdgpu_do_asic_reset(device_list_handle, reset_context); |
b98a1648 | 5329 | if (r && r == -EAGAIN) |
26bc5340 | 5330 | goto retry; |
f5c7e779 YC |
5331 | |
5332 | if (!r && gpu_reset_for_dev_remove) | |
5333 | goto recover_end; | |
26bc5340 AG |
5334 | } |
5335 | ||
1d721ed6 AG |
5336 | skip_hw_reset: |
5337 | ||
26bc5340 | 5338 | /* Post ASIC reset for all devs .*/ |
655ce9cb | 5339 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
7c6e68c7 | 5340 | |
1d721ed6 AG |
5341 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { |
5342 | struct amdgpu_ring *ring = tmp_adev->rings[i]; | |
5343 | ||
5344 | if (!ring || !ring->sched.thread) | |
5345 | continue; | |
5346 | ||
6868a2c4 | 5347 | drm_sched_start(&ring->sched, true); |
1d721ed6 AG |
5348 | } |
5349 | ||
693073a0 | 5350 | if (adev->enable_mes && adev->ip_versions[GC_HWIP][0] != IP_VERSION(11, 0, 3)) |
ed67f729 JX |
5351 | amdgpu_mes_self_test(tmp_adev); |
5352 | ||
1053b9c9 | 5353 | if (!drm_drv_uses_atomic_modeset(adev_to_drm(tmp_adev)) && !job_signaled) { |
4a580877 | 5354 | drm_helper_resume_force_mode(adev_to_drm(tmp_adev)); |
1d721ed6 AG |
5355 | } |
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 AD |
5431 | /* covers APUs as well */ |
5432 | if (pci_is_root_bus(adev->pdev->bus)) { | |
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 | |
4a580877 | 5594 | if (!amdgpu_device_supports_baco(adev_to_drm(adev))) |
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 | |
4a580877 | 5610 | if (!amdgpu_device_supports_baco(adev_to_drm(adev))) |
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 LY |
5912 | |
5913 | drm_dev_unplug(ddev); | |
5914 | ||
5915 | amdgpu_irq_disable_all(adev); | |
5916 | ||
5917 | amdgpu_fence_driver_hw_fini(adev); | |
5918 | ||
5919 | adev->no_hw_access = true; | |
5920 | ||
5921 | amdgpu_device_unmap_mmio(adev); | |
5922 | ||
5923 | pci_disable_device(pdev); | |
5924 | pci_wait_for_pending_transaction(pdev); | |
5925 | } | |
86700a40 XD |
5926 | |
5927 | u32 amdgpu_device_pcie_port_rreg(struct amdgpu_device *adev, | |
5928 | u32 reg) | |
5929 | { | |
5930 | unsigned long flags, address, data; | |
5931 | u32 r; | |
5932 | ||
5933 | address = adev->nbio.funcs->get_pcie_port_index_offset(adev); | |
5934 | data = adev->nbio.funcs->get_pcie_port_data_offset(adev); | |
5935 | ||
5936 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); | |
5937 | WREG32(address, reg * 4); | |
5938 | (void)RREG32(address); | |
5939 | r = RREG32(data); | |
5940 | spin_unlock_irqrestore(&adev->pcie_idx_lock, flags); | |
5941 | return r; | |
5942 | } | |
5943 | ||
5944 | void amdgpu_device_pcie_port_wreg(struct amdgpu_device *adev, | |
5945 | u32 reg, u32 v) | |
5946 | { | |
5947 | unsigned long flags, address, data; | |
5948 | ||
5949 | address = adev->nbio.funcs->get_pcie_port_index_offset(adev); | |
5950 | data = adev->nbio.funcs->get_pcie_port_data_offset(adev); | |
5951 | ||
5952 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); | |
5953 | WREG32(address, reg * 4); | |
5954 | (void)RREG32(address); | |
5955 | WREG32(data, v); | |
5956 | (void)RREG32(data); | |
5957 | spin_unlock_irqrestore(&adev->pcie_idx_lock, flags); | |
5958 | } | |
68ce8b24 CK |
5959 | |
5960 | /** | |
5961 | * amdgpu_device_switch_gang - switch to a new gang | |
5962 | * @adev: amdgpu_device pointer | |
5963 | * @gang: the gang to switch to | |
5964 | * | |
5965 | * Try to switch to a new gang. | |
5966 | * Returns: NULL if we switched to the new gang or a reference to the current | |
5967 | * gang leader. | |
5968 | */ | |
5969 | struct dma_fence *amdgpu_device_switch_gang(struct amdgpu_device *adev, | |
5970 | struct dma_fence *gang) | |
5971 | { | |
5972 | struct dma_fence *old = NULL; | |
5973 | ||
5974 | do { | |
5975 | dma_fence_put(old); | |
5976 | rcu_read_lock(); | |
5977 | old = dma_fence_get_rcu_safe(&adev->gang_submit); | |
5978 | rcu_read_unlock(); | |
5979 | ||
5980 | if (old == gang) | |
5981 | break; | |
5982 | ||
5983 | if (!dma_fence_is_signaled(old)) | |
5984 | return old; | |
5985 | ||
5986 | } while (cmpxchg((struct dma_fence __force **)&adev->gang_submit, | |
5987 | old, gang) != old); | |
5988 | ||
5989 | dma_fence_put(old); | |
5990 | return NULL; | |
5991 | } | |
220c8cc8 AD |
5992 | |
5993 | bool amdgpu_device_has_display_hardware(struct amdgpu_device *adev) | |
5994 | { | |
5995 | switch (adev->asic_type) { | |
5996 | #ifdef CONFIG_DRM_AMDGPU_SI | |
5997 | case CHIP_HAINAN: | |
5998 | #endif | |
5999 | case CHIP_TOPAZ: | |
6000 | /* chips with no display hardware */ | |
6001 | return false; | |
6002 | #ifdef CONFIG_DRM_AMDGPU_SI | |
6003 | case CHIP_TAHITI: | |
6004 | case CHIP_PITCAIRN: | |
6005 | case CHIP_VERDE: | |
6006 | case CHIP_OLAND: | |
6007 | #endif | |
6008 | #ifdef CONFIG_DRM_AMDGPU_CIK | |
6009 | case CHIP_BONAIRE: | |
6010 | case CHIP_HAWAII: | |
6011 | case CHIP_KAVERI: | |
6012 | case CHIP_KABINI: | |
6013 | case CHIP_MULLINS: | |
6014 | #endif | |
6015 | case CHIP_TONGA: | |
6016 | case CHIP_FIJI: | |
6017 | case CHIP_POLARIS10: | |
6018 | case CHIP_POLARIS11: | |
6019 | case CHIP_POLARIS12: | |
6020 | case CHIP_VEGAM: | |
6021 | case CHIP_CARRIZO: | |
6022 | case CHIP_STONEY: | |
6023 | /* chips with display hardware */ | |
6024 | return true; | |
6025 | default: | |
6026 | /* IP discovery */ | |
6027 | if (!adev->ip_versions[DCE_HWIP][0] || | |
6028 | (adev->harvest_ip_mask & AMD_HARVEST_IP_DMU_MASK)) | |
6029 | return false; | |
6030 | return true; | |
6031 | } | |
6032 | } |