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> |
fdf2f6c5 | 35 | |
4562236b | 36 | #include <drm/drm_atomic_helper.h> |
fcd70cd3 | 37 | #include <drm/drm_probe_helper.h> |
d38ceaf9 AD |
38 | #include <drm/amdgpu_drm.h> |
39 | #include <linux/vgaarb.h> | |
40 | #include <linux/vga_switcheroo.h> | |
41 | #include <linux/efi.h> | |
42 | #include "amdgpu.h" | |
f4b373f4 | 43 | #include "amdgpu_trace.h" |
d38ceaf9 AD |
44 | #include "amdgpu_i2c.h" |
45 | #include "atom.h" | |
46 | #include "amdgpu_atombios.h" | |
a5bde2f9 | 47 | #include "amdgpu_atomfirmware.h" |
d0dd7f0c | 48 | #include "amd_pcie.h" |
33f34802 KW |
49 | #ifdef CONFIG_DRM_AMDGPU_SI |
50 | #include "si.h" | |
51 | #endif | |
a2e73f56 AD |
52 | #ifdef CONFIG_DRM_AMDGPU_CIK |
53 | #include "cik.h" | |
54 | #endif | |
aaa36a97 | 55 | #include "vi.h" |
460826e6 | 56 | #include "soc15.h" |
0a5b8c7b | 57 | #include "nv.h" |
d38ceaf9 | 58 | #include "bif/bif_4_1_d.h" |
bec86378 | 59 | #include <linux/firmware.h> |
89041940 | 60 | #include "amdgpu_vf_error.h" |
d38ceaf9 | 61 | |
ba997709 | 62 | #include "amdgpu_amdkfd.h" |
d2f52ac8 | 63 | #include "amdgpu_pm.h" |
d38ceaf9 | 64 | |
5183411b | 65 | #include "amdgpu_xgmi.h" |
c030f2e4 | 66 | #include "amdgpu_ras.h" |
9c7c85f7 | 67 | #include "amdgpu_pmu.h" |
bd607166 | 68 | #include "amdgpu_fru_eeprom.h" |
04442bf7 | 69 | #include "amdgpu_reset.h" |
5183411b | 70 | |
d5ea093e | 71 | #include <linux/suspend.h> |
c6a6e2db | 72 | #include <drm/task_barrier.h> |
3f12acc8 | 73 | #include <linux/pm_runtime.h> |
d5ea093e | 74 | |
f89f8c6b AG |
75 | #include <drm/drm_drv.h> |
76 | ||
e2a75f88 | 77 | MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin"); |
3f76dced | 78 | MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin"); |
2d2e5e7e | 79 | MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin"); |
ad5a67a7 | 80 | MODULE_FIRMWARE("amdgpu/picasso_gpu_info.bin"); |
54c4d17e | 81 | MODULE_FIRMWARE("amdgpu/raven2_gpu_info.bin"); |
65e60f6e | 82 | MODULE_FIRMWARE("amdgpu/arcturus_gpu_info.bin"); |
b51a26a0 | 83 | MODULE_FIRMWARE("amdgpu/renoir_gpu_info.bin"); |
23c6268e | 84 | MODULE_FIRMWARE("amdgpu/navi10_gpu_info.bin"); |
ed42cfe1 | 85 | MODULE_FIRMWARE("amdgpu/navi14_gpu_info.bin"); |
42b325e5 | 86 | MODULE_FIRMWARE("amdgpu/navi12_gpu_info.bin"); |
4e52a9f8 | 87 | MODULE_FIRMWARE("amdgpu/vangogh_gpu_info.bin"); |
8bf84f60 | 88 | MODULE_FIRMWARE("amdgpu/yellow_carp_gpu_info.bin"); |
e2a75f88 | 89 | |
2dc80b00 | 90 | #define AMDGPU_RESUME_MS 2000 |
7258fa31 SK |
91 | #define AMDGPU_MAX_RETRY_LIMIT 2 |
92 | #define AMDGPU_RETRY_SRIOV_RESET(r) ((r) == -EBUSY || (r) == -ETIMEDOUT || (r) == -EINVAL) | |
2dc80b00 | 93 | |
050091ab | 94 | const char *amdgpu_asic_name[] = { |
da69c161 KW |
95 | "TAHITI", |
96 | "PITCAIRN", | |
97 | "VERDE", | |
98 | "OLAND", | |
99 | "HAINAN", | |
d38ceaf9 AD |
100 | "BONAIRE", |
101 | "KAVERI", | |
102 | "KABINI", | |
103 | "HAWAII", | |
104 | "MULLINS", | |
105 | "TOPAZ", | |
106 | "TONGA", | |
48299f95 | 107 | "FIJI", |
d38ceaf9 | 108 | "CARRIZO", |
139f4917 | 109 | "STONEY", |
2cc0c0b5 FC |
110 | "POLARIS10", |
111 | "POLARIS11", | |
c4642a47 | 112 | "POLARIS12", |
48ff108d | 113 | "VEGAM", |
d4196f01 | 114 | "VEGA10", |
8fab806a | 115 | "VEGA12", |
956fcddc | 116 | "VEGA20", |
2ca8a5d2 | 117 | "RAVEN", |
d6c3b24e | 118 | "ARCTURUS", |
1eee4228 | 119 | "RENOIR", |
d46b417a | 120 | "ALDEBARAN", |
852a6626 | 121 | "NAVI10", |
d0f56dc2 | 122 | "CYAN_SKILLFISH", |
87dbad02 | 123 | "NAVI14", |
9802f5d7 | 124 | "NAVI12", |
ccaf72d3 | 125 | "SIENNA_CICHLID", |
ddd8fbe7 | 126 | "NAVY_FLOUNDER", |
4f1e9a76 | 127 | "VANGOGH", |
a2468e04 | 128 | "DIMGREY_CAVEFISH", |
6f169591 | 129 | "BEIGE_GOBY", |
ee9236b7 | 130 | "YELLOW_CARP", |
3ae695d6 | 131 | "IP DISCOVERY", |
d38ceaf9 AD |
132 | "LAST", |
133 | }; | |
134 | ||
dcea6e65 KR |
135 | /** |
136 | * DOC: pcie_replay_count | |
137 | * | |
138 | * The amdgpu driver provides a sysfs API for reporting the total number | |
139 | * of PCIe replays (NAKs) | |
140 | * The file pcie_replay_count is used for this and returns the total | |
141 | * number of replays as a sum of the NAKs generated and NAKs received | |
142 | */ | |
143 | ||
144 | static ssize_t amdgpu_device_get_pcie_replay_count(struct device *dev, | |
145 | struct device_attribute *attr, char *buf) | |
146 | { | |
147 | struct drm_device *ddev = dev_get_drvdata(dev); | |
1348969a | 148 | struct amdgpu_device *adev = drm_to_adev(ddev); |
dcea6e65 KR |
149 | uint64_t cnt = amdgpu_asic_get_pcie_replay_count(adev); |
150 | ||
36000c7a | 151 | return sysfs_emit(buf, "%llu\n", cnt); |
dcea6e65 KR |
152 | } |
153 | ||
154 | static DEVICE_ATTR(pcie_replay_count, S_IRUGO, | |
155 | amdgpu_device_get_pcie_replay_count, NULL); | |
156 | ||
5494d864 AD |
157 | static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev); |
158 | ||
bd607166 KR |
159 | /** |
160 | * DOC: product_name | |
161 | * | |
162 | * The amdgpu driver provides a sysfs API for reporting the product name | |
163 | * for the device | |
164 | * The file serial_number is used for this and returns the product name | |
165 | * as returned from the FRU. | |
166 | * NOTE: This is only available for certain server cards | |
167 | */ | |
168 | ||
169 | static ssize_t amdgpu_device_get_product_name(struct device *dev, | |
170 | struct device_attribute *attr, char *buf) | |
171 | { | |
172 | struct drm_device *ddev = dev_get_drvdata(dev); | |
1348969a | 173 | struct amdgpu_device *adev = drm_to_adev(ddev); |
bd607166 | 174 | |
36000c7a | 175 | return sysfs_emit(buf, "%s\n", adev->product_name); |
bd607166 KR |
176 | } |
177 | ||
178 | static DEVICE_ATTR(product_name, S_IRUGO, | |
179 | amdgpu_device_get_product_name, NULL); | |
180 | ||
181 | /** | |
182 | * DOC: product_number | |
183 | * | |
184 | * The amdgpu driver provides a sysfs API for reporting the part number | |
185 | * for the device | |
186 | * The file serial_number is used for this and returns the part number | |
187 | * as returned from the FRU. | |
188 | * NOTE: This is only available for certain server cards | |
189 | */ | |
190 | ||
191 | static ssize_t amdgpu_device_get_product_number(struct device *dev, | |
192 | struct device_attribute *attr, char *buf) | |
193 | { | |
194 | struct drm_device *ddev = dev_get_drvdata(dev); | |
1348969a | 195 | struct amdgpu_device *adev = drm_to_adev(ddev); |
bd607166 | 196 | |
36000c7a | 197 | return sysfs_emit(buf, "%s\n", adev->product_number); |
bd607166 KR |
198 | } |
199 | ||
200 | static DEVICE_ATTR(product_number, S_IRUGO, | |
201 | amdgpu_device_get_product_number, NULL); | |
202 | ||
203 | /** | |
204 | * DOC: serial_number | |
205 | * | |
206 | * The amdgpu driver provides a sysfs API for reporting the serial number | |
207 | * for the device | |
208 | * The file serial_number is used for this and returns the serial number | |
209 | * as returned from the FRU. | |
210 | * NOTE: This is only available for certain server cards | |
211 | */ | |
212 | ||
213 | static ssize_t amdgpu_device_get_serial_number(struct device *dev, | |
214 | struct device_attribute *attr, char *buf) | |
215 | { | |
216 | struct drm_device *ddev = dev_get_drvdata(dev); | |
1348969a | 217 | struct amdgpu_device *adev = drm_to_adev(ddev); |
bd607166 | 218 | |
36000c7a | 219 | return sysfs_emit(buf, "%s\n", adev->serial); |
bd607166 KR |
220 | } |
221 | ||
222 | static DEVICE_ATTR(serial_number, S_IRUGO, | |
223 | amdgpu_device_get_serial_number, NULL); | |
224 | ||
fd496ca8 | 225 | /** |
b98c6299 | 226 | * amdgpu_device_supports_px - Is the device a dGPU with ATPX power control |
fd496ca8 AD |
227 | * |
228 | * @dev: drm_device pointer | |
229 | * | |
b98c6299 | 230 | * Returns true if the device is a dGPU with ATPX power control, |
fd496ca8 AD |
231 | * otherwise return false. |
232 | */ | |
b98c6299 | 233 | bool amdgpu_device_supports_px(struct drm_device *dev) |
fd496ca8 AD |
234 | { |
235 | struct amdgpu_device *adev = drm_to_adev(dev); | |
236 | ||
b98c6299 | 237 | if ((adev->flags & AMD_IS_PX) && !amdgpu_is_atpx_hybrid()) |
fd496ca8 AD |
238 | return true; |
239 | return false; | |
240 | } | |
241 | ||
e3ecdffa | 242 | /** |
0330b848 | 243 | * amdgpu_device_supports_boco - Is the device a dGPU with ACPI power resources |
e3ecdffa AD |
244 | * |
245 | * @dev: drm_device pointer | |
246 | * | |
b98c6299 | 247 | * Returns true if the device is a dGPU with ACPI power control, |
e3ecdffa AD |
248 | * otherwise return false. |
249 | */ | |
31af062a | 250 | bool amdgpu_device_supports_boco(struct drm_device *dev) |
d38ceaf9 | 251 | { |
1348969a | 252 | struct amdgpu_device *adev = drm_to_adev(dev); |
d38ceaf9 | 253 | |
b98c6299 AD |
254 | if (adev->has_pr3 || |
255 | ((adev->flags & AMD_IS_PX) && amdgpu_is_atpx_hybrid())) | |
d38ceaf9 AD |
256 | return true; |
257 | return false; | |
258 | } | |
259 | ||
a69cba42 AD |
260 | /** |
261 | * amdgpu_device_supports_baco - Does the device support BACO | |
262 | * | |
263 | * @dev: drm_device pointer | |
264 | * | |
265 | * Returns true if the device supporte BACO, | |
266 | * otherwise return false. | |
267 | */ | |
268 | bool amdgpu_device_supports_baco(struct drm_device *dev) | |
269 | { | |
1348969a | 270 | struct amdgpu_device *adev = drm_to_adev(dev); |
a69cba42 AD |
271 | |
272 | return amdgpu_asic_supports_baco(adev); | |
273 | } | |
274 | ||
3fa8f89d S |
275 | /** |
276 | * amdgpu_device_supports_smart_shift - Is the device dGPU with | |
277 | * smart shift support | |
278 | * | |
279 | * @dev: drm_device pointer | |
280 | * | |
281 | * Returns true if the device is a dGPU with Smart Shift support, | |
282 | * otherwise returns false. | |
283 | */ | |
284 | bool amdgpu_device_supports_smart_shift(struct drm_device *dev) | |
285 | { | |
286 | return (amdgpu_device_supports_boco(dev) && | |
287 | amdgpu_acpi_is_power_shift_control_supported()); | |
288 | } | |
289 | ||
6e3cd2a9 MCC |
290 | /* |
291 | * VRAM access helper functions | |
292 | */ | |
293 | ||
e35e2b11 | 294 | /** |
048af66b | 295 | * amdgpu_device_mm_access - access vram by MM_INDEX/MM_DATA |
e35e2b11 TY |
296 | * |
297 | * @adev: amdgpu_device pointer | |
298 | * @pos: offset of the buffer in vram | |
299 | * @buf: virtual address of the buffer in system memory | |
300 | * @size: read/write size, sizeof(@buf) must > @size | |
301 | * @write: true - write to vram, otherwise - read from vram | |
302 | */ | |
048af66b KW |
303 | void amdgpu_device_mm_access(struct amdgpu_device *adev, loff_t pos, |
304 | void *buf, size_t size, bool write) | |
e35e2b11 | 305 | { |
e35e2b11 | 306 | unsigned long flags; |
048af66b KW |
307 | uint32_t hi = ~0, tmp = 0; |
308 | uint32_t *data = buf; | |
ce05ac56 | 309 | uint64_t last; |
f89f8c6b | 310 | int idx; |
ce05ac56 | 311 | |
c58a863b | 312 | if (!drm_dev_enter(adev_to_drm(adev), &idx)) |
f89f8c6b | 313 | return; |
9d11eb0d | 314 | |
048af66b KW |
315 | BUG_ON(!IS_ALIGNED(pos, 4) || !IS_ALIGNED(size, 4)); |
316 | ||
317 | spin_lock_irqsave(&adev->mmio_idx_lock, flags); | |
318 | for (last = pos + size; pos < last; pos += 4) { | |
319 | tmp = pos >> 31; | |
320 | ||
321 | WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)pos) | 0x80000000); | |
322 | if (tmp != hi) { | |
323 | WREG32_NO_KIQ(mmMM_INDEX_HI, tmp); | |
324 | hi = tmp; | |
325 | } | |
326 | if (write) | |
327 | WREG32_NO_KIQ(mmMM_DATA, *data++); | |
328 | else | |
329 | *data++ = RREG32_NO_KIQ(mmMM_DATA); | |
330 | } | |
331 | ||
332 | spin_unlock_irqrestore(&adev->mmio_idx_lock, flags); | |
333 | drm_dev_exit(idx); | |
334 | } | |
335 | ||
336 | /** | |
bbe04dec | 337 | * amdgpu_device_aper_access - access vram by vram aperature |
048af66b KW |
338 | * |
339 | * @adev: amdgpu_device pointer | |
340 | * @pos: offset of the buffer in vram | |
341 | * @buf: virtual address of the buffer in system memory | |
342 | * @size: read/write size, sizeof(@buf) must > @size | |
343 | * @write: true - write to vram, otherwise - read from vram | |
344 | * | |
345 | * The return value means how many bytes have been transferred. | |
346 | */ | |
347 | size_t amdgpu_device_aper_access(struct amdgpu_device *adev, loff_t pos, | |
348 | void *buf, size_t size, bool write) | |
349 | { | |
9d11eb0d | 350 | #ifdef CONFIG_64BIT |
048af66b KW |
351 | void __iomem *addr; |
352 | size_t count = 0; | |
353 | uint64_t last; | |
354 | ||
355 | if (!adev->mman.aper_base_kaddr) | |
356 | return 0; | |
357 | ||
9d11eb0d CK |
358 | last = min(pos + size, adev->gmc.visible_vram_size); |
359 | if (last > pos) { | |
048af66b KW |
360 | addr = adev->mman.aper_base_kaddr + pos; |
361 | count = last - pos; | |
9d11eb0d CK |
362 | |
363 | if (write) { | |
364 | memcpy_toio(addr, buf, count); | |
365 | mb(); | |
810085dd | 366 | amdgpu_device_flush_hdp(adev, NULL); |
9d11eb0d | 367 | } else { |
810085dd | 368 | amdgpu_device_invalidate_hdp(adev, NULL); |
9d11eb0d CK |
369 | mb(); |
370 | memcpy_fromio(buf, addr, count); | |
371 | } | |
372 | ||
9d11eb0d | 373 | } |
048af66b KW |
374 | |
375 | return count; | |
376 | #else | |
377 | return 0; | |
9d11eb0d | 378 | #endif |
048af66b | 379 | } |
9d11eb0d | 380 | |
048af66b KW |
381 | /** |
382 | * amdgpu_device_vram_access - read/write a buffer in vram | |
383 | * | |
384 | * @adev: amdgpu_device pointer | |
385 | * @pos: offset of the buffer in vram | |
386 | * @buf: virtual address of the buffer in system memory | |
387 | * @size: read/write size, sizeof(@buf) must > @size | |
388 | * @write: true - write to vram, otherwise - read from vram | |
389 | */ | |
390 | void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos, | |
391 | void *buf, size_t size, bool write) | |
392 | { | |
393 | size_t count; | |
e35e2b11 | 394 | |
048af66b KW |
395 | /* try to using vram apreature to access vram first */ |
396 | count = amdgpu_device_aper_access(adev, pos, buf, size, write); | |
397 | size -= count; | |
398 | if (size) { | |
399 | /* using MM to access rest vram */ | |
400 | pos += count; | |
401 | buf += count; | |
402 | amdgpu_device_mm_access(adev, pos, buf, size, write); | |
e35e2b11 TY |
403 | } |
404 | } | |
405 | ||
d38ceaf9 | 406 | /* |
f7ee1874 | 407 | * register access helper functions. |
d38ceaf9 | 408 | */ |
56b53c0b DL |
409 | |
410 | /* Check if hw access should be skipped because of hotplug or device error */ | |
411 | bool amdgpu_device_skip_hw_access(struct amdgpu_device *adev) | |
412 | { | |
7afefb81 | 413 | if (adev->no_hw_access) |
56b53c0b DL |
414 | return true; |
415 | ||
416 | #ifdef CONFIG_LOCKDEP | |
417 | /* | |
418 | * This is a bit complicated to understand, so worth a comment. What we assert | |
419 | * here is that the GPU reset is not running on another thread in parallel. | |
420 | * | |
421 | * For this we trylock the read side of the reset semaphore, if that succeeds | |
422 | * we know that the reset is not running in paralell. | |
423 | * | |
424 | * If the trylock fails we assert that we are either already holding the read | |
425 | * side of the lock or are the reset thread itself and hold the write side of | |
426 | * the lock. | |
427 | */ | |
428 | if (in_task()) { | |
429 | if (down_read_trylock(&adev->reset_sem)) | |
430 | up_read(&adev->reset_sem); | |
431 | else | |
432 | lockdep_assert_held(&adev->reset_sem); | |
433 | } | |
434 | #endif | |
435 | return false; | |
436 | } | |
437 | ||
e3ecdffa | 438 | /** |
f7ee1874 | 439 | * amdgpu_device_rreg - read a memory mapped IO or indirect register |
e3ecdffa AD |
440 | * |
441 | * @adev: amdgpu_device pointer | |
442 | * @reg: dword aligned register offset | |
443 | * @acc_flags: access flags which require special behavior | |
444 | * | |
445 | * Returns the 32 bit value from the offset specified. | |
446 | */ | |
f7ee1874 HZ |
447 | uint32_t amdgpu_device_rreg(struct amdgpu_device *adev, |
448 | uint32_t reg, uint32_t acc_flags) | |
d38ceaf9 | 449 | { |
f4b373f4 TSD |
450 | uint32_t ret; |
451 | ||
56b53c0b | 452 | if (amdgpu_device_skip_hw_access(adev)) |
bf36b52e AG |
453 | return 0; |
454 | ||
f7ee1874 HZ |
455 | if ((reg * 4) < adev->rmmio_size) { |
456 | if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && | |
457 | amdgpu_sriov_runtime(adev) && | |
458 | down_read_trylock(&adev->reset_sem)) { | |
459 | ret = amdgpu_kiq_rreg(adev, reg); | |
460 | up_read(&adev->reset_sem); | |
461 | } else { | |
462 | ret = readl(((void __iomem *)adev->rmmio) + (reg * 4)); | |
463 | } | |
464 | } else { | |
465 | ret = adev->pcie_rreg(adev, reg * 4); | |
81202807 | 466 | } |
bc992ba5 | 467 | |
f7ee1874 | 468 | trace_amdgpu_device_rreg(adev->pdev->device, reg, ret); |
e78b579d | 469 | |
f4b373f4 | 470 | return ret; |
d38ceaf9 AD |
471 | } |
472 | ||
421a2a30 ML |
473 | /* |
474 | * MMIO register read with bytes helper functions | |
475 | * @offset:bytes offset from MMIO start | |
476 | * | |
477 | */ | |
478 | ||
e3ecdffa AD |
479 | /** |
480 | * amdgpu_mm_rreg8 - read a memory mapped IO register | |
481 | * | |
482 | * @adev: amdgpu_device pointer | |
483 | * @offset: byte aligned register offset | |
484 | * | |
485 | * Returns the 8 bit value from the offset specified. | |
486 | */ | |
7cbbc745 AG |
487 | uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset) |
488 | { | |
56b53c0b | 489 | if (amdgpu_device_skip_hw_access(adev)) |
bf36b52e AG |
490 | return 0; |
491 | ||
421a2a30 ML |
492 | if (offset < adev->rmmio_size) |
493 | return (readb(adev->rmmio + offset)); | |
494 | BUG(); | |
495 | } | |
496 | ||
497 | /* | |
498 | * MMIO register write with bytes helper functions | |
499 | * @offset:bytes offset from MMIO start | |
500 | * @value: the value want to be written to the register | |
501 | * | |
502 | */ | |
e3ecdffa AD |
503 | /** |
504 | * amdgpu_mm_wreg8 - read a memory mapped IO register | |
505 | * | |
506 | * @adev: amdgpu_device pointer | |
507 | * @offset: byte aligned register offset | |
508 | * @value: 8 bit value to write | |
509 | * | |
510 | * Writes the value specified to the offset specified. | |
511 | */ | |
7cbbc745 AG |
512 | void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value) |
513 | { | |
56b53c0b | 514 | if (amdgpu_device_skip_hw_access(adev)) |
bf36b52e AG |
515 | return; |
516 | ||
421a2a30 ML |
517 | if (offset < adev->rmmio_size) |
518 | writeb(value, adev->rmmio + offset); | |
519 | else | |
520 | BUG(); | |
521 | } | |
522 | ||
e3ecdffa | 523 | /** |
f7ee1874 | 524 | * amdgpu_device_wreg - write to a memory mapped IO or indirect register |
e3ecdffa AD |
525 | * |
526 | * @adev: amdgpu_device pointer | |
527 | * @reg: dword aligned register offset | |
528 | * @v: 32 bit value to write to the register | |
529 | * @acc_flags: access flags which require special behavior | |
530 | * | |
531 | * Writes the value specified to the offset specified. | |
532 | */ | |
f7ee1874 HZ |
533 | void amdgpu_device_wreg(struct amdgpu_device *adev, |
534 | uint32_t reg, uint32_t v, | |
535 | uint32_t acc_flags) | |
d38ceaf9 | 536 | { |
56b53c0b | 537 | if (amdgpu_device_skip_hw_access(adev)) |
bf36b52e AG |
538 | return; |
539 | ||
f7ee1874 HZ |
540 | if ((reg * 4) < adev->rmmio_size) { |
541 | if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && | |
542 | amdgpu_sriov_runtime(adev) && | |
543 | down_read_trylock(&adev->reset_sem)) { | |
544 | amdgpu_kiq_wreg(adev, reg, v); | |
545 | up_read(&adev->reset_sem); | |
546 | } else { | |
547 | writel(v, ((void __iomem *)adev->rmmio) + (reg * 4)); | |
548 | } | |
549 | } else { | |
550 | adev->pcie_wreg(adev, reg * 4, v); | |
81202807 | 551 | } |
bc992ba5 | 552 | |
f7ee1874 | 553 | trace_amdgpu_device_wreg(adev->pdev->device, reg, v); |
2e0cc4d4 | 554 | } |
d38ceaf9 | 555 | |
03f2abb0 | 556 | /** |
4cc9f86f | 557 | * amdgpu_mm_wreg_mmio_rlc - write register either with direct/indirect mmio or with RLC path if in range |
2e0cc4d4 | 558 | * |
71579346 RB |
559 | * @adev: amdgpu_device pointer |
560 | * @reg: mmio/rlc register | |
561 | * @v: value to write | |
562 | * | |
563 | * this function is invoked only for the debugfs register access | |
03f2abb0 | 564 | */ |
f7ee1874 HZ |
565 | void amdgpu_mm_wreg_mmio_rlc(struct amdgpu_device *adev, |
566 | uint32_t reg, uint32_t v) | |
2e0cc4d4 | 567 | { |
56b53c0b | 568 | if (amdgpu_device_skip_hw_access(adev)) |
bf36b52e AG |
569 | return; |
570 | ||
2e0cc4d4 | 571 | if (amdgpu_sriov_fullaccess(adev) && |
f7ee1874 HZ |
572 | adev->gfx.rlc.funcs && |
573 | adev->gfx.rlc.funcs->is_rlcg_access_range) { | |
2e0cc4d4 | 574 | if (adev->gfx.rlc.funcs->is_rlcg_access_range(adev, reg)) |
1b2dc99e | 575 | return amdgpu_sriov_wreg(adev, reg, v, 0, 0); |
4cc9f86f TSD |
576 | } else if ((reg * 4) >= adev->rmmio_size) { |
577 | adev->pcie_wreg(adev, reg * 4, v); | |
f7ee1874 HZ |
578 | } else { |
579 | writel(v, ((void __iomem *)adev->rmmio) + (reg * 4)); | |
47ed4e1c | 580 | } |
d38ceaf9 AD |
581 | } |
582 | ||
d38ceaf9 AD |
583 | /** |
584 | * amdgpu_mm_rdoorbell - read a doorbell dword | |
585 | * | |
586 | * @adev: amdgpu_device pointer | |
587 | * @index: doorbell index | |
588 | * | |
589 | * Returns the value in the doorbell aperture at the | |
590 | * requested doorbell index (CIK). | |
591 | */ | |
592 | u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index) | |
593 | { | |
56b53c0b | 594 | if (amdgpu_device_skip_hw_access(adev)) |
bf36b52e AG |
595 | return 0; |
596 | ||
d38ceaf9 AD |
597 | if (index < adev->doorbell.num_doorbells) { |
598 | return readl(adev->doorbell.ptr + index); | |
599 | } else { | |
600 | DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index); | |
601 | return 0; | |
602 | } | |
603 | } | |
604 | ||
605 | /** | |
606 | * amdgpu_mm_wdoorbell - write a doorbell dword | |
607 | * | |
608 | * @adev: amdgpu_device pointer | |
609 | * @index: doorbell index | |
610 | * @v: value to write | |
611 | * | |
612 | * Writes @v to the doorbell aperture at the | |
613 | * requested doorbell index (CIK). | |
614 | */ | |
615 | void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v) | |
616 | { | |
56b53c0b | 617 | if (amdgpu_device_skip_hw_access(adev)) |
bf36b52e AG |
618 | return; |
619 | ||
d38ceaf9 AD |
620 | if (index < adev->doorbell.num_doorbells) { |
621 | writel(v, adev->doorbell.ptr + index); | |
622 | } else { | |
623 | DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index); | |
624 | } | |
625 | } | |
626 | ||
832be404 KW |
627 | /** |
628 | * amdgpu_mm_rdoorbell64 - read a doorbell Qword | |
629 | * | |
630 | * @adev: amdgpu_device pointer | |
631 | * @index: doorbell index | |
632 | * | |
633 | * Returns the value in the doorbell aperture at the | |
634 | * requested doorbell index (VEGA10+). | |
635 | */ | |
636 | u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index) | |
637 | { | |
56b53c0b | 638 | if (amdgpu_device_skip_hw_access(adev)) |
bf36b52e AG |
639 | return 0; |
640 | ||
832be404 KW |
641 | if (index < adev->doorbell.num_doorbells) { |
642 | return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index)); | |
643 | } else { | |
644 | DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index); | |
645 | return 0; | |
646 | } | |
647 | } | |
648 | ||
649 | /** | |
650 | * amdgpu_mm_wdoorbell64 - write a doorbell Qword | |
651 | * | |
652 | * @adev: amdgpu_device pointer | |
653 | * @index: doorbell index | |
654 | * @v: value to write | |
655 | * | |
656 | * Writes @v to the doorbell aperture at the | |
657 | * requested doorbell index (VEGA10+). | |
658 | */ | |
659 | void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v) | |
660 | { | |
56b53c0b | 661 | if (amdgpu_device_skip_hw_access(adev)) |
bf36b52e AG |
662 | return; |
663 | ||
832be404 KW |
664 | if (index < adev->doorbell.num_doorbells) { |
665 | atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v); | |
666 | } else { | |
667 | DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index); | |
668 | } | |
669 | } | |
670 | ||
1bba3683 HZ |
671 | /** |
672 | * amdgpu_device_indirect_rreg - read an indirect register | |
673 | * | |
674 | * @adev: amdgpu_device pointer | |
675 | * @pcie_index: mmio register offset | |
676 | * @pcie_data: mmio register offset | |
22f453fb | 677 | * @reg_addr: indirect register address to read from |
1bba3683 HZ |
678 | * |
679 | * Returns the value of indirect register @reg_addr | |
680 | */ | |
681 | u32 amdgpu_device_indirect_rreg(struct amdgpu_device *adev, | |
682 | u32 pcie_index, u32 pcie_data, | |
683 | u32 reg_addr) | |
684 | { | |
685 | unsigned long flags; | |
686 | u32 r; | |
687 | void __iomem *pcie_index_offset; | |
688 | void __iomem *pcie_data_offset; | |
689 | ||
690 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); | |
691 | pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4; | |
692 | pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4; | |
693 | ||
694 | writel(reg_addr, pcie_index_offset); | |
695 | readl(pcie_index_offset); | |
696 | r = readl(pcie_data_offset); | |
697 | spin_unlock_irqrestore(&adev->pcie_idx_lock, flags); | |
698 | ||
699 | return r; | |
700 | } | |
701 | ||
702 | /** | |
703 | * amdgpu_device_indirect_rreg64 - read a 64bits indirect register | |
704 | * | |
705 | * @adev: amdgpu_device pointer | |
706 | * @pcie_index: mmio register offset | |
707 | * @pcie_data: mmio register offset | |
22f453fb | 708 | * @reg_addr: indirect register address to read from |
1bba3683 HZ |
709 | * |
710 | * Returns the value of indirect register @reg_addr | |
711 | */ | |
712 | u64 amdgpu_device_indirect_rreg64(struct amdgpu_device *adev, | |
713 | u32 pcie_index, u32 pcie_data, | |
714 | u32 reg_addr) | |
715 | { | |
716 | unsigned long flags; | |
717 | u64 r; | |
718 | void __iomem *pcie_index_offset; | |
719 | void __iomem *pcie_data_offset; | |
720 | ||
721 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); | |
722 | pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4; | |
723 | pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4; | |
724 | ||
725 | /* read low 32 bits */ | |
726 | writel(reg_addr, pcie_index_offset); | |
727 | readl(pcie_index_offset); | |
728 | r = readl(pcie_data_offset); | |
729 | /* read high 32 bits */ | |
730 | writel(reg_addr + 4, pcie_index_offset); | |
731 | readl(pcie_index_offset); | |
732 | r |= ((u64)readl(pcie_data_offset) << 32); | |
733 | spin_unlock_irqrestore(&adev->pcie_idx_lock, flags); | |
734 | ||
735 | return r; | |
736 | } | |
737 | ||
738 | /** | |
739 | * amdgpu_device_indirect_wreg - write an indirect register address | |
740 | * | |
741 | * @adev: amdgpu_device pointer | |
742 | * @pcie_index: mmio register offset | |
743 | * @pcie_data: mmio register offset | |
744 | * @reg_addr: indirect register offset | |
745 | * @reg_data: indirect register data | |
746 | * | |
747 | */ | |
748 | void amdgpu_device_indirect_wreg(struct amdgpu_device *adev, | |
749 | u32 pcie_index, u32 pcie_data, | |
750 | u32 reg_addr, u32 reg_data) | |
751 | { | |
752 | unsigned long flags; | |
753 | void __iomem *pcie_index_offset; | |
754 | void __iomem *pcie_data_offset; | |
755 | ||
756 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); | |
757 | pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4; | |
758 | pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4; | |
759 | ||
760 | writel(reg_addr, pcie_index_offset); | |
761 | readl(pcie_index_offset); | |
762 | writel(reg_data, pcie_data_offset); | |
763 | readl(pcie_data_offset); | |
764 | spin_unlock_irqrestore(&adev->pcie_idx_lock, flags); | |
765 | } | |
766 | ||
767 | /** | |
768 | * amdgpu_device_indirect_wreg64 - write a 64bits indirect register address | |
769 | * | |
770 | * @adev: amdgpu_device pointer | |
771 | * @pcie_index: mmio register offset | |
772 | * @pcie_data: mmio register offset | |
773 | * @reg_addr: indirect register offset | |
774 | * @reg_data: indirect register data | |
775 | * | |
776 | */ | |
777 | void amdgpu_device_indirect_wreg64(struct amdgpu_device *adev, | |
778 | u32 pcie_index, u32 pcie_data, | |
779 | u32 reg_addr, u64 reg_data) | |
780 | { | |
781 | unsigned long flags; | |
782 | void __iomem *pcie_index_offset; | |
783 | void __iomem *pcie_data_offset; | |
784 | ||
785 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); | |
786 | pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4; | |
787 | pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4; | |
788 | ||
789 | /* write low 32 bits */ | |
790 | writel(reg_addr, pcie_index_offset); | |
791 | readl(pcie_index_offset); | |
792 | writel((u32)(reg_data & 0xffffffffULL), pcie_data_offset); | |
793 | readl(pcie_data_offset); | |
794 | /* write high 32 bits */ | |
795 | writel(reg_addr + 4, pcie_index_offset); | |
796 | readl(pcie_index_offset); | |
797 | writel((u32)(reg_data >> 32), pcie_data_offset); | |
798 | readl(pcie_data_offset); | |
799 | spin_unlock_irqrestore(&adev->pcie_idx_lock, flags); | |
800 | } | |
801 | ||
d38ceaf9 AD |
802 | /** |
803 | * amdgpu_invalid_rreg - dummy reg read function | |
804 | * | |
982a820b | 805 | * @adev: amdgpu_device pointer |
d38ceaf9 AD |
806 | * @reg: offset of register |
807 | * | |
808 | * Dummy register read function. Used for register blocks | |
809 | * that certain asics don't have (all asics). | |
810 | * Returns the value in the register. | |
811 | */ | |
812 | static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg) | |
813 | { | |
814 | DRM_ERROR("Invalid callback to read register 0x%04X\n", reg); | |
815 | BUG(); | |
816 | return 0; | |
817 | } | |
818 | ||
819 | /** | |
820 | * amdgpu_invalid_wreg - dummy reg write function | |
821 | * | |
982a820b | 822 | * @adev: amdgpu_device pointer |
d38ceaf9 AD |
823 | * @reg: offset of register |
824 | * @v: value to write to the register | |
825 | * | |
826 | * Dummy register read function. Used for register blocks | |
827 | * that certain asics don't have (all asics). | |
828 | */ | |
829 | static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v) | |
830 | { | |
831 | DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n", | |
832 | reg, v); | |
833 | BUG(); | |
834 | } | |
835 | ||
4fa1c6a6 TZ |
836 | /** |
837 | * amdgpu_invalid_rreg64 - dummy 64 bit reg read function | |
838 | * | |
982a820b | 839 | * @adev: amdgpu_device pointer |
4fa1c6a6 TZ |
840 | * @reg: offset of register |
841 | * | |
842 | * Dummy register read function. Used for register blocks | |
843 | * that certain asics don't have (all asics). | |
844 | * Returns the value in the register. | |
845 | */ | |
846 | static uint64_t amdgpu_invalid_rreg64(struct amdgpu_device *adev, uint32_t reg) | |
847 | { | |
848 | DRM_ERROR("Invalid callback to read 64 bit register 0x%04X\n", reg); | |
849 | BUG(); | |
850 | return 0; | |
851 | } | |
852 | ||
853 | /** | |
854 | * amdgpu_invalid_wreg64 - dummy reg write function | |
855 | * | |
982a820b | 856 | * @adev: amdgpu_device pointer |
4fa1c6a6 TZ |
857 | * @reg: offset of register |
858 | * @v: value to write to the register | |
859 | * | |
860 | * Dummy register read function. Used for register blocks | |
861 | * that certain asics don't have (all asics). | |
862 | */ | |
863 | static void amdgpu_invalid_wreg64(struct amdgpu_device *adev, uint32_t reg, uint64_t v) | |
864 | { | |
865 | DRM_ERROR("Invalid callback to write 64 bit register 0x%04X with 0x%08llX\n", | |
866 | reg, v); | |
867 | BUG(); | |
868 | } | |
869 | ||
d38ceaf9 AD |
870 | /** |
871 | * amdgpu_block_invalid_rreg - dummy reg read function | |
872 | * | |
982a820b | 873 | * @adev: amdgpu_device pointer |
d38ceaf9 AD |
874 | * @block: offset of instance |
875 | * @reg: offset of register | |
876 | * | |
877 | * Dummy register read function. Used for register blocks | |
878 | * that certain asics don't have (all asics). | |
879 | * Returns the value in the register. | |
880 | */ | |
881 | static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev, | |
882 | uint32_t block, uint32_t reg) | |
883 | { | |
884 | DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n", | |
885 | reg, block); | |
886 | BUG(); | |
887 | return 0; | |
888 | } | |
889 | ||
890 | /** | |
891 | * amdgpu_block_invalid_wreg - dummy reg write function | |
892 | * | |
982a820b | 893 | * @adev: amdgpu_device pointer |
d38ceaf9 AD |
894 | * @block: offset of instance |
895 | * @reg: offset of register | |
896 | * @v: value to write to the register | |
897 | * | |
898 | * Dummy register read function. Used for register blocks | |
899 | * that certain asics don't have (all asics). | |
900 | */ | |
901 | static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev, | |
902 | uint32_t block, | |
903 | uint32_t reg, uint32_t v) | |
904 | { | |
905 | DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n", | |
906 | reg, block, v); | |
907 | BUG(); | |
908 | } | |
909 | ||
4d2997ab AD |
910 | /** |
911 | * amdgpu_device_asic_init - Wrapper for atom asic_init | |
912 | * | |
982a820b | 913 | * @adev: amdgpu_device pointer |
4d2997ab AD |
914 | * |
915 | * Does any asic specific work and then calls atom asic init. | |
916 | */ | |
917 | static int amdgpu_device_asic_init(struct amdgpu_device *adev) | |
918 | { | |
919 | amdgpu_asic_pre_asic_init(adev); | |
920 | ||
921 | return amdgpu_atom_asic_init(adev->mode_info.atom_context); | |
922 | } | |
923 | ||
e3ecdffa AD |
924 | /** |
925 | * amdgpu_device_vram_scratch_init - allocate the VRAM scratch page | |
926 | * | |
982a820b | 927 | * @adev: amdgpu_device pointer |
e3ecdffa AD |
928 | * |
929 | * Allocates a scratch page of VRAM for use by various things in the | |
930 | * driver. | |
931 | */ | |
06ec9070 | 932 | static int amdgpu_device_vram_scratch_init(struct amdgpu_device *adev) |
d38ceaf9 | 933 | { |
a4a02777 CK |
934 | return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE, |
935 | PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM, | |
936 | &adev->vram_scratch.robj, | |
937 | &adev->vram_scratch.gpu_addr, | |
938 | (void **)&adev->vram_scratch.ptr); | |
d38ceaf9 AD |
939 | } |
940 | ||
e3ecdffa AD |
941 | /** |
942 | * amdgpu_device_vram_scratch_fini - Free the VRAM scratch page | |
943 | * | |
982a820b | 944 | * @adev: amdgpu_device pointer |
e3ecdffa AD |
945 | * |
946 | * Frees the VRAM scratch page. | |
947 | */ | |
06ec9070 | 948 | static void amdgpu_device_vram_scratch_fini(struct amdgpu_device *adev) |
d38ceaf9 | 949 | { |
078af1a3 | 950 | amdgpu_bo_free_kernel(&adev->vram_scratch.robj, NULL, NULL); |
d38ceaf9 AD |
951 | } |
952 | ||
953 | /** | |
9c3f2b54 | 954 | * amdgpu_device_program_register_sequence - program an array of registers. |
d38ceaf9 AD |
955 | * |
956 | * @adev: amdgpu_device pointer | |
957 | * @registers: pointer to the register array | |
958 | * @array_size: size of the register array | |
959 | * | |
960 | * Programs an array or registers with and and or masks. | |
961 | * This is a helper for setting golden registers. | |
962 | */ | |
9c3f2b54 AD |
963 | void amdgpu_device_program_register_sequence(struct amdgpu_device *adev, |
964 | const u32 *registers, | |
965 | const u32 array_size) | |
d38ceaf9 AD |
966 | { |
967 | u32 tmp, reg, and_mask, or_mask; | |
968 | int i; | |
969 | ||
970 | if (array_size % 3) | |
971 | return; | |
972 | ||
973 | for (i = 0; i < array_size; i +=3) { | |
974 | reg = registers[i + 0]; | |
975 | and_mask = registers[i + 1]; | |
976 | or_mask = registers[i + 2]; | |
977 | ||
978 | if (and_mask == 0xffffffff) { | |
979 | tmp = or_mask; | |
980 | } else { | |
981 | tmp = RREG32(reg); | |
982 | tmp &= ~and_mask; | |
e0d07657 HZ |
983 | if (adev->family >= AMDGPU_FAMILY_AI) |
984 | tmp |= (or_mask & and_mask); | |
985 | else | |
986 | tmp |= or_mask; | |
d38ceaf9 AD |
987 | } |
988 | WREG32(reg, tmp); | |
989 | } | |
990 | } | |
991 | ||
e3ecdffa AD |
992 | /** |
993 | * amdgpu_device_pci_config_reset - reset the GPU | |
994 | * | |
995 | * @adev: amdgpu_device pointer | |
996 | * | |
997 | * Resets the GPU using the pci config reset sequence. | |
998 | * Only applicable to asics prior to vega10. | |
999 | */ | |
8111c387 | 1000 | void amdgpu_device_pci_config_reset(struct amdgpu_device *adev) |
d38ceaf9 AD |
1001 | { |
1002 | pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA); | |
1003 | } | |
1004 | ||
af484df8 AD |
1005 | /** |
1006 | * amdgpu_device_pci_reset - reset the GPU using generic PCI means | |
1007 | * | |
1008 | * @adev: amdgpu_device pointer | |
1009 | * | |
1010 | * Resets the GPU using generic pci reset interfaces (FLR, SBR, etc.). | |
1011 | */ | |
1012 | int amdgpu_device_pci_reset(struct amdgpu_device *adev) | |
1013 | { | |
1014 | return pci_reset_function(adev->pdev); | |
1015 | } | |
1016 | ||
d38ceaf9 AD |
1017 | /* |
1018 | * GPU doorbell aperture helpers function. | |
1019 | */ | |
1020 | /** | |
06ec9070 | 1021 | * amdgpu_device_doorbell_init - Init doorbell driver information. |
d38ceaf9 AD |
1022 | * |
1023 | * @adev: amdgpu_device pointer | |
1024 | * | |
1025 | * Init doorbell driver information (CIK) | |
1026 | * Returns 0 on success, error on failure. | |
1027 | */ | |
06ec9070 | 1028 | static int amdgpu_device_doorbell_init(struct amdgpu_device *adev) |
d38ceaf9 | 1029 | { |
6585661d | 1030 | |
705e519e CK |
1031 | /* No doorbell on SI hardware generation */ |
1032 | if (adev->asic_type < CHIP_BONAIRE) { | |
1033 | adev->doorbell.base = 0; | |
1034 | adev->doorbell.size = 0; | |
1035 | adev->doorbell.num_doorbells = 0; | |
1036 | adev->doorbell.ptr = NULL; | |
1037 | return 0; | |
1038 | } | |
1039 | ||
d6895ad3 CK |
1040 | if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET) |
1041 | return -EINVAL; | |
1042 | ||
22357775 AD |
1043 | amdgpu_asic_init_doorbell_index(adev); |
1044 | ||
d38ceaf9 AD |
1045 | /* doorbell bar mapping */ |
1046 | adev->doorbell.base = pci_resource_start(adev->pdev, 2); | |
1047 | adev->doorbell.size = pci_resource_len(adev->pdev, 2); | |
1048 | ||
edf600da | 1049 | adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32), |
9564f192 | 1050 | adev->doorbell_index.max_assignment+1); |
d38ceaf9 AD |
1051 | if (adev->doorbell.num_doorbells == 0) |
1052 | return -EINVAL; | |
1053 | ||
ec3db8a6 | 1054 | /* For Vega, reserve and map two pages on doorbell BAR since SDMA |
88dc26e4 OZ |
1055 | * paging queue doorbell use the second page. The |
1056 | * AMDGPU_DOORBELL64_MAX_ASSIGNMENT definition assumes all the | |
1057 | * doorbells are in the first page. So with paging queue enabled, | |
1058 | * the max num_doorbells should + 1 page (0x400 in dword) | |
ec3db8a6 PY |
1059 | */ |
1060 | if (adev->asic_type >= CHIP_VEGA10) | |
88dc26e4 | 1061 | adev->doorbell.num_doorbells += 0x400; |
ec3db8a6 | 1062 | |
8972e5d2 CK |
1063 | adev->doorbell.ptr = ioremap(adev->doorbell.base, |
1064 | adev->doorbell.num_doorbells * | |
1065 | sizeof(u32)); | |
1066 | if (adev->doorbell.ptr == NULL) | |
d38ceaf9 | 1067 | return -ENOMEM; |
d38ceaf9 AD |
1068 | |
1069 | return 0; | |
1070 | } | |
1071 | ||
1072 | /** | |
06ec9070 | 1073 | * amdgpu_device_doorbell_fini - Tear down doorbell driver information. |
d38ceaf9 AD |
1074 | * |
1075 | * @adev: amdgpu_device pointer | |
1076 | * | |
1077 | * Tear down doorbell driver information (CIK) | |
1078 | */ | |
06ec9070 | 1079 | static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev) |
d38ceaf9 AD |
1080 | { |
1081 | iounmap(adev->doorbell.ptr); | |
1082 | adev->doorbell.ptr = NULL; | |
1083 | } | |
1084 | ||
22cb0164 | 1085 | |
d38ceaf9 AD |
1086 | |
1087 | /* | |
06ec9070 | 1088 | * amdgpu_device_wb_*() |
455a7bc2 | 1089 | * Writeback is the method by which the GPU updates special pages in memory |
ea81a173 | 1090 | * with the status of certain GPU events (fences, ring pointers,etc.). |
d38ceaf9 AD |
1091 | */ |
1092 | ||
1093 | /** | |
06ec9070 | 1094 | * amdgpu_device_wb_fini - Disable Writeback and free memory |
d38ceaf9 AD |
1095 | * |
1096 | * @adev: amdgpu_device pointer | |
1097 | * | |
1098 | * Disables Writeback and frees the Writeback memory (all asics). | |
1099 | * Used at driver shutdown. | |
1100 | */ | |
06ec9070 | 1101 | static void amdgpu_device_wb_fini(struct amdgpu_device *adev) |
d38ceaf9 AD |
1102 | { |
1103 | if (adev->wb.wb_obj) { | |
a76ed485 AD |
1104 | amdgpu_bo_free_kernel(&adev->wb.wb_obj, |
1105 | &adev->wb.gpu_addr, | |
1106 | (void **)&adev->wb.wb); | |
d38ceaf9 AD |
1107 | adev->wb.wb_obj = NULL; |
1108 | } | |
1109 | } | |
1110 | ||
1111 | /** | |
03f2abb0 | 1112 | * amdgpu_device_wb_init - Init Writeback driver info and allocate memory |
d38ceaf9 AD |
1113 | * |
1114 | * @adev: amdgpu_device pointer | |
1115 | * | |
455a7bc2 | 1116 | * Initializes writeback and allocates writeback memory (all asics). |
d38ceaf9 AD |
1117 | * Used at driver startup. |
1118 | * Returns 0 on success or an -error on failure. | |
1119 | */ | |
06ec9070 | 1120 | static int amdgpu_device_wb_init(struct amdgpu_device *adev) |
d38ceaf9 AD |
1121 | { |
1122 | int r; | |
1123 | ||
1124 | if (adev->wb.wb_obj == NULL) { | |
97407b63 AD |
1125 | /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */ |
1126 | r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8, | |
a76ed485 AD |
1127 | PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT, |
1128 | &adev->wb.wb_obj, &adev->wb.gpu_addr, | |
1129 | (void **)&adev->wb.wb); | |
d38ceaf9 AD |
1130 | if (r) { |
1131 | dev_warn(adev->dev, "(%d) create WB bo failed\n", r); | |
1132 | return r; | |
1133 | } | |
d38ceaf9 AD |
1134 | |
1135 | adev->wb.num_wb = AMDGPU_MAX_WB; | |
1136 | memset(&adev->wb.used, 0, sizeof(adev->wb.used)); | |
1137 | ||
1138 | /* clear wb memory */ | |
73469585 | 1139 | memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8); |
d38ceaf9 AD |
1140 | } |
1141 | ||
1142 | return 0; | |
1143 | } | |
1144 | ||
1145 | /** | |
131b4b36 | 1146 | * amdgpu_device_wb_get - Allocate a wb entry |
d38ceaf9 AD |
1147 | * |
1148 | * @adev: amdgpu_device pointer | |
1149 | * @wb: wb index | |
1150 | * | |
1151 | * Allocate a wb slot for use by the driver (all asics). | |
1152 | * Returns 0 on success or -EINVAL on failure. | |
1153 | */ | |
131b4b36 | 1154 | int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb) |
d38ceaf9 AD |
1155 | { |
1156 | unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb); | |
d38ceaf9 | 1157 | |
97407b63 | 1158 | if (offset < adev->wb.num_wb) { |
7014285a | 1159 | __set_bit(offset, adev->wb.used); |
63ae07ca | 1160 | *wb = offset << 3; /* convert to dw offset */ |
0915fdbc ML |
1161 | return 0; |
1162 | } else { | |
1163 | return -EINVAL; | |
1164 | } | |
1165 | } | |
1166 | ||
d38ceaf9 | 1167 | /** |
131b4b36 | 1168 | * amdgpu_device_wb_free - Free a wb entry |
d38ceaf9 AD |
1169 | * |
1170 | * @adev: amdgpu_device pointer | |
1171 | * @wb: wb index | |
1172 | * | |
1173 | * Free a wb slot allocated for use by the driver (all asics) | |
1174 | */ | |
131b4b36 | 1175 | void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb) |
d38ceaf9 | 1176 | { |
73469585 | 1177 | wb >>= 3; |
d38ceaf9 | 1178 | if (wb < adev->wb.num_wb) |
73469585 | 1179 | __clear_bit(wb, adev->wb.used); |
d38ceaf9 AD |
1180 | } |
1181 | ||
d6895ad3 CK |
1182 | /** |
1183 | * amdgpu_device_resize_fb_bar - try to resize FB BAR | |
1184 | * | |
1185 | * @adev: amdgpu_device pointer | |
1186 | * | |
1187 | * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not | |
1188 | * to fail, but if any of the BARs is not accessible after the size we abort | |
1189 | * driver loading by returning -ENODEV. | |
1190 | */ | |
1191 | int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev) | |
1192 | { | |
453f617a | 1193 | int rbar_size = pci_rebar_bytes_to_size(adev->gmc.real_vram_size); |
31b8adab CK |
1194 | struct pci_bus *root; |
1195 | struct resource *res; | |
1196 | unsigned i; | |
d6895ad3 CK |
1197 | u16 cmd; |
1198 | int r; | |
1199 | ||
0c03b912 | 1200 | /* Bypass for VF */ |
1201 | if (amdgpu_sriov_vf(adev)) | |
1202 | return 0; | |
1203 | ||
b7221f2b AD |
1204 | /* skip if the bios has already enabled large BAR */ |
1205 | if (adev->gmc.real_vram_size && | |
1206 | (pci_resource_len(adev->pdev, 0) >= adev->gmc.real_vram_size)) | |
1207 | return 0; | |
1208 | ||
31b8adab CK |
1209 | /* Check if the root BUS has 64bit memory resources */ |
1210 | root = adev->pdev->bus; | |
1211 | while (root->parent) | |
1212 | root = root->parent; | |
1213 | ||
1214 | pci_bus_for_each_resource(root, res, i) { | |
0ebb7c54 | 1215 | if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) && |
31b8adab CK |
1216 | res->start > 0x100000000ull) |
1217 | break; | |
1218 | } | |
1219 | ||
1220 | /* Trying to resize is pointless without a root hub window above 4GB */ | |
1221 | if (!res) | |
1222 | return 0; | |
1223 | ||
453f617a ND |
1224 | /* Limit the BAR size to what is available */ |
1225 | rbar_size = min(fls(pci_rebar_get_possible_sizes(adev->pdev, 0)) - 1, | |
1226 | rbar_size); | |
1227 | ||
d6895ad3 CK |
1228 | /* Disable memory decoding while we change the BAR addresses and size */ |
1229 | pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd); | |
1230 | pci_write_config_word(adev->pdev, PCI_COMMAND, | |
1231 | cmd & ~PCI_COMMAND_MEMORY); | |
1232 | ||
1233 | /* Free the VRAM and doorbell BAR, we most likely need to move both. */ | |
06ec9070 | 1234 | amdgpu_device_doorbell_fini(adev); |
d6895ad3 CK |
1235 | if (adev->asic_type >= CHIP_BONAIRE) |
1236 | pci_release_resource(adev->pdev, 2); | |
1237 | ||
1238 | pci_release_resource(adev->pdev, 0); | |
1239 | ||
1240 | r = pci_resize_resource(adev->pdev, 0, rbar_size); | |
1241 | if (r == -ENOSPC) | |
1242 | DRM_INFO("Not enough PCI address space for a large BAR."); | |
1243 | else if (r && r != -ENOTSUPP) | |
1244 | DRM_ERROR("Problem resizing BAR0 (%d).", r); | |
1245 | ||
1246 | pci_assign_unassigned_bus_resources(adev->pdev->bus); | |
1247 | ||
1248 | /* When the doorbell or fb BAR isn't available we have no chance of | |
1249 | * using the device. | |
1250 | */ | |
06ec9070 | 1251 | r = amdgpu_device_doorbell_init(adev); |
d6895ad3 CK |
1252 | if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET)) |
1253 | return -ENODEV; | |
1254 | ||
1255 | pci_write_config_word(adev->pdev, PCI_COMMAND, cmd); | |
1256 | ||
1257 | return 0; | |
1258 | } | |
a05502e5 | 1259 | |
d38ceaf9 AD |
1260 | /* |
1261 | * GPU helpers function. | |
1262 | */ | |
1263 | /** | |
39c640c0 | 1264 | * amdgpu_device_need_post - check if the hw need post or not |
d38ceaf9 AD |
1265 | * |
1266 | * @adev: amdgpu_device pointer | |
1267 | * | |
c836fec5 JQ |
1268 | * Check if the asic has been initialized (all asics) at driver startup |
1269 | * or post is needed if hw reset is performed. | |
1270 | * Returns true if need or false if not. | |
d38ceaf9 | 1271 | */ |
39c640c0 | 1272 | bool amdgpu_device_need_post(struct amdgpu_device *adev) |
d38ceaf9 AD |
1273 | { |
1274 | uint32_t reg; | |
1275 | ||
bec86378 ML |
1276 | if (amdgpu_sriov_vf(adev)) |
1277 | return false; | |
1278 | ||
1279 | if (amdgpu_passthrough(adev)) { | |
1da2c326 ML |
1280 | /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot |
1281 | * some old smc fw still need driver do vPost otherwise gpu hang, while | |
1282 | * those smc fw version above 22.15 doesn't have this flaw, so we force | |
1283 | * vpost executed for smc version below 22.15 | |
bec86378 ML |
1284 | */ |
1285 | if (adev->asic_type == CHIP_FIJI) { | |
1286 | int err; | |
1287 | uint32_t fw_ver; | |
1288 | err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev); | |
1289 | /* force vPost if error occured */ | |
1290 | if (err) | |
1291 | return true; | |
1292 | ||
1293 | fw_ver = *((uint32_t *)adev->pm.fw->data + 69); | |
1da2c326 ML |
1294 | if (fw_ver < 0x00160e00) |
1295 | return true; | |
bec86378 | 1296 | } |
bec86378 | 1297 | } |
91fe77eb | 1298 | |
e3c1b071 | 1299 | /* Don't post if we need to reset whole hive on init */ |
1300 | if (adev->gmc.xgmi.pending_reset) | |
1301 | return false; | |
1302 | ||
91fe77eb | 1303 | if (adev->has_hw_reset) { |
1304 | adev->has_hw_reset = false; | |
1305 | return true; | |
1306 | } | |
1307 | ||
1308 | /* bios scratch used on CIK+ */ | |
1309 | if (adev->asic_type >= CHIP_BONAIRE) | |
1310 | return amdgpu_atombios_scratch_need_asic_init(adev); | |
1311 | ||
1312 | /* check MEM_SIZE for older asics */ | |
1313 | reg = amdgpu_asic_get_config_memsize(adev); | |
1314 | ||
1315 | if ((reg != 0) && (reg != 0xffffffff)) | |
1316 | return false; | |
1317 | ||
1318 | return true; | |
bec86378 ML |
1319 | } |
1320 | ||
d38ceaf9 AD |
1321 | /* if we get transitioned to only one device, take VGA back */ |
1322 | /** | |
06ec9070 | 1323 | * amdgpu_device_vga_set_decode - enable/disable vga decode |
d38ceaf9 | 1324 | * |
bf44e8ce | 1325 | * @pdev: PCI device pointer |
d38ceaf9 AD |
1326 | * @state: enable/disable vga decode |
1327 | * | |
1328 | * Enable/disable vga decode (all asics). | |
1329 | * Returns VGA resource flags. | |
1330 | */ | |
bf44e8ce CH |
1331 | static unsigned int amdgpu_device_vga_set_decode(struct pci_dev *pdev, |
1332 | bool state) | |
d38ceaf9 | 1333 | { |
bf44e8ce | 1334 | struct amdgpu_device *adev = drm_to_adev(pci_get_drvdata(pdev)); |
d38ceaf9 AD |
1335 | amdgpu_asic_set_vga_state(adev, state); |
1336 | if (state) | |
1337 | return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM | | |
1338 | VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM; | |
1339 | else | |
1340 | return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM; | |
1341 | } | |
1342 | ||
e3ecdffa AD |
1343 | /** |
1344 | * amdgpu_device_check_block_size - validate the vm block size | |
1345 | * | |
1346 | * @adev: amdgpu_device pointer | |
1347 | * | |
1348 | * Validates the vm block size specified via module parameter. | |
1349 | * The vm block size defines number of bits in page table versus page directory, | |
1350 | * a page is 4KB so we have 12 bits offset, minimum 9 bits in the | |
1351 | * page table and the remaining bits are in the page directory. | |
1352 | */ | |
06ec9070 | 1353 | static void amdgpu_device_check_block_size(struct amdgpu_device *adev) |
a1adf8be CZ |
1354 | { |
1355 | /* defines number of bits in page table versus page directory, | |
1356 | * a page is 4KB so we have 12 bits offset, minimum 9 bits in the | |
1357 | * page table and the remaining bits are in the page directory */ | |
bab4fee7 JZ |
1358 | if (amdgpu_vm_block_size == -1) |
1359 | return; | |
a1adf8be | 1360 | |
bab4fee7 | 1361 | if (amdgpu_vm_block_size < 9) { |
a1adf8be CZ |
1362 | dev_warn(adev->dev, "VM page table size (%d) too small\n", |
1363 | amdgpu_vm_block_size); | |
97489129 | 1364 | amdgpu_vm_block_size = -1; |
a1adf8be | 1365 | } |
a1adf8be CZ |
1366 | } |
1367 | ||
e3ecdffa AD |
1368 | /** |
1369 | * amdgpu_device_check_vm_size - validate the vm size | |
1370 | * | |
1371 | * @adev: amdgpu_device pointer | |
1372 | * | |
1373 | * Validates the vm size in GB specified via module parameter. | |
1374 | * The VM size is the size of the GPU virtual memory space in GB. | |
1375 | */ | |
06ec9070 | 1376 | static void amdgpu_device_check_vm_size(struct amdgpu_device *adev) |
83ca145d | 1377 | { |
64dab074 AD |
1378 | /* no need to check the default value */ |
1379 | if (amdgpu_vm_size == -1) | |
1380 | return; | |
1381 | ||
83ca145d ZJ |
1382 | if (amdgpu_vm_size < 1) { |
1383 | dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n", | |
1384 | amdgpu_vm_size); | |
f3368128 | 1385 | amdgpu_vm_size = -1; |
83ca145d | 1386 | } |
83ca145d ZJ |
1387 | } |
1388 | ||
7951e376 RZ |
1389 | static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev) |
1390 | { | |
1391 | struct sysinfo si; | |
a9d4fe2f | 1392 | bool is_os_64 = (sizeof(void *) == 8); |
7951e376 RZ |
1393 | uint64_t total_memory; |
1394 | uint64_t dram_size_seven_GB = 0x1B8000000; | |
1395 | uint64_t dram_size_three_GB = 0xB8000000; | |
1396 | ||
1397 | if (amdgpu_smu_memory_pool_size == 0) | |
1398 | return; | |
1399 | ||
1400 | if (!is_os_64) { | |
1401 | DRM_WARN("Not 64-bit OS, feature not supported\n"); | |
1402 | goto def_value; | |
1403 | } | |
1404 | si_meminfo(&si); | |
1405 | total_memory = (uint64_t)si.totalram * si.mem_unit; | |
1406 | ||
1407 | if ((amdgpu_smu_memory_pool_size == 1) || | |
1408 | (amdgpu_smu_memory_pool_size == 2)) { | |
1409 | if (total_memory < dram_size_three_GB) | |
1410 | goto def_value1; | |
1411 | } else if ((amdgpu_smu_memory_pool_size == 4) || | |
1412 | (amdgpu_smu_memory_pool_size == 8)) { | |
1413 | if (total_memory < dram_size_seven_GB) | |
1414 | goto def_value1; | |
1415 | } else { | |
1416 | DRM_WARN("Smu memory pool size not supported\n"); | |
1417 | goto def_value; | |
1418 | } | |
1419 | adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28; | |
1420 | ||
1421 | return; | |
1422 | ||
1423 | def_value1: | |
1424 | DRM_WARN("No enough system memory\n"); | |
1425 | def_value: | |
1426 | adev->pm.smu_prv_buffer_size = 0; | |
1427 | } | |
1428 | ||
9f6a7857 HR |
1429 | static int amdgpu_device_init_apu_flags(struct amdgpu_device *adev) |
1430 | { | |
1431 | if (!(adev->flags & AMD_IS_APU) || | |
1432 | adev->asic_type < CHIP_RAVEN) | |
1433 | return 0; | |
1434 | ||
1435 | switch (adev->asic_type) { | |
1436 | case CHIP_RAVEN: | |
1437 | if (adev->pdev->device == 0x15dd) | |
1438 | adev->apu_flags |= AMD_APU_IS_RAVEN; | |
1439 | if (adev->pdev->device == 0x15d8) | |
1440 | adev->apu_flags |= AMD_APU_IS_PICASSO; | |
1441 | break; | |
1442 | case CHIP_RENOIR: | |
1443 | if ((adev->pdev->device == 0x1636) || | |
1444 | (adev->pdev->device == 0x164c)) | |
1445 | adev->apu_flags |= AMD_APU_IS_RENOIR; | |
1446 | else | |
1447 | adev->apu_flags |= AMD_APU_IS_GREEN_SARDINE; | |
1448 | break; | |
1449 | case CHIP_VANGOGH: | |
1450 | adev->apu_flags |= AMD_APU_IS_VANGOGH; | |
1451 | break; | |
1452 | case CHIP_YELLOW_CARP: | |
1453 | break; | |
d0f56dc2 TZ |
1454 | case CHIP_CYAN_SKILLFISH: |
1455 | if (adev->pdev->device == 0x13FE) | |
1456 | adev->apu_flags |= AMD_APU_IS_CYAN_SKILLFISH2; | |
1457 | break; | |
9f6a7857 | 1458 | default: |
4eaf21b7 | 1459 | break; |
9f6a7857 HR |
1460 | } |
1461 | ||
1462 | return 0; | |
1463 | } | |
1464 | ||
d38ceaf9 | 1465 | /** |
06ec9070 | 1466 | * amdgpu_device_check_arguments - validate module params |
d38ceaf9 AD |
1467 | * |
1468 | * @adev: amdgpu_device pointer | |
1469 | * | |
1470 | * Validates certain module parameters and updates | |
1471 | * the associated values used by the driver (all asics). | |
1472 | */ | |
912dfc84 | 1473 | static int amdgpu_device_check_arguments(struct amdgpu_device *adev) |
d38ceaf9 | 1474 | { |
5b011235 CZ |
1475 | if (amdgpu_sched_jobs < 4) { |
1476 | dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n", | |
1477 | amdgpu_sched_jobs); | |
1478 | amdgpu_sched_jobs = 4; | |
76117507 | 1479 | } else if (!is_power_of_2(amdgpu_sched_jobs)){ |
5b011235 CZ |
1480 | dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n", |
1481 | amdgpu_sched_jobs); | |
1482 | amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs); | |
1483 | } | |
d38ceaf9 | 1484 | |
83e74db6 | 1485 | if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) { |
f9321cc4 CK |
1486 | /* gart size must be greater or equal to 32M */ |
1487 | dev_warn(adev->dev, "gart size (%d) too small\n", | |
1488 | amdgpu_gart_size); | |
83e74db6 | 1489 | amdgpu_gart_size = -1; |
d38ceaf9 AD |
1490 | } |
1491 | ||
36d38372 | 1492 | if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) { |
c4e1a13a | 1493 | /* gtt size must be greater or equal to 32M */ |
36d38372 CK |
1494 | dev_warn(adev->dev, "gtt size (%d) too small\n", |
1495 | amdgpu_gtt_size); | |
1496 | amdgpu_gtt_size = -1; | |
d38ceaf9 AD |
1497 | } |
1498 | ||
d07f14be RH |
1499 | /* valid range is between 4 and 9 inclusive */ |
1500 | if (amdgpu_vm_fragment_size != -1 && | |
1501 | (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) { | |
1502 | dev_warn(adev->dev, "valid range is between 4 and 9\n"); | |
1503 | amdgpu_vm_fragment_size = -1; | |
1504 | } | |
1505 | ||
5d5bd5e3 KW |
1506 | if (amdgpu_sched_hw_submission < 2) { |
1507 | dev_warn(adev->dev, "sched hw submission jobs (%d) must be at least 2\n", | |
1508 | amdgpu_sched_hw_submission); | |
1509 | amdgpu_sched_hw_submission = 2; | |
1510 | } else if (!is_power_of_2(amdgpu_sched_hw_submission)) { | |
1511 | dev_warn(adev->dev, "sched hw submission jobs (%d) must be a power of 2\n", | |
1512 | amdgpu_sched_hw_submission); | |
1513 | amdgpu_sched_hw_submission = roundup_pow_of_two(amdgpu_sched_hw_submission); | |
1514 | } | |
1515 | ||
7951e376 RZ |
1516 | amdgpu_device_check_smu_prv_buffer_size(adev); |
1517 | ||
06ec9070 | 1518 | amdgpu_device_check_vm_size(adev); |
d38ceaf9 | 1519 | |
06ec9070 | 1520 | amdgpu_device_check_block_size(adev); |
6a7f76e7 | 1521 | |
19aede77 | 1522 | adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type); |
912dfc84 | 1523 | |
c6252390 | 1524 | amdgpu_gmc_tmz_set(adev); |
01a8dcec | 1525 | |
9b498efa AD |
1526 | amdgpu_gmc_noretry_set(adev); |
1527 | ||
e3c00faa | 1528 | return 0; |
d38ceaf9 AD |
1529 | } |
1530 | ||
1531 | /** | |
1532 | * amdgpu_switcheroo_set_state - set switcheroo state | |
1533 | * | |
1534 | * @pdev: pci dev pointer | |
1694467b | 1535 | * @state: vga_switcheroo state |
d38ceaf9 AD |
1536 | * |
1537 | * Callback for the switcheroo driver. Suspends or resumes the | |
1538 | * the asics before or after it is powered up using ACPI methods. | |
1539 | */ | |
8aba21b7 LT |
1540 | static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, |
1541 | enum vga_switcheroo_state state) | |
d38ceaf9 AD |
1542 | { |
1543 | struct drm_device *dev = pci_get_drvdata(pdev); | |
de185019 | 1544 | int r; |
d38ceaf9 | 1545 | |
b98c6299 | 1546 | if (amdgpu_device_supports_px(dev) && state == VGA_SWITCHEROO_OFF) |
d38ceaf9 AD |
1547 | return; |
1548 | ||
1549 | if (state == VGA_SWITCHEROO_ON) { | |
dd4fa6c1 | 1550 | pr_info("switched on\n"); |
d38ceaf9 AD |
1551 | /* don't suspend or resume card normally */ |
1552 | dev->switch_power_state = DRM_SWITCH_POWER_CHANGING; | |
1553 | ||
8f66090b TZ |
1554 | pci_set_power_state(pdev, PCI_D0); |
1555 | amdgpu_device_load_pci_state(pdev); | |
1556 | r = pci_enable_device(pdev); | |
de185019 AD |
1557 | if (r) |
1558 | DRM_WARN("pci_enable_device failed (%d)\n", r); | |
1559 | amdgpu_device_resume(dev, true); | |
d38ceaf9 | 1560 | |
d38ceaf9 | 1561 | dev->switch_power_state = DRM_SWITCH_POWER_ON; |
d38ceaf9 | 1562 | } else { |
dd4fa6c1 | 1563 | pr_info("switched off\n"); |
d38ceaf9 | 1564 | dev->switch_power_state = DRM_SWITCH_POWER_CHANGING; |
de185019 | 1565 | amdgpu_device_suspend(dev, true); |
8f66090b | 1566 | amdgpu_device_cache_pci_state(pdev); |
de185019 | 1567 | /* Shut down the device */ |
8f66090b TZ |
1568 | pci_disable_device(pdev); |
1569 | pci_set_power_state(pdev, PCI_D3cold); | |
d38ceaf9 AD |
1570 | dev->switch_power_state = DRM_SWITCH_POWER_OFF; |
1571 | } | |
1572 | } | |
1573 | ||
1574 | /** | |
1575 | * amdgpu_switcheroo_can_switch - see if switcheroo state can change | |
1576 | * | |
1577 | * @pdev: pci dev pointer | |
1578 | * | |
1579 | * Callback for the switcheroo driver. Check of the switcheroo | |
1580 | * state can be changed. | |
1581 | * Returns true if the state can be changed, false if not. | |
1582 | */ | |
1583 | static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev) | |
1584 | { | |
1585 | struct drm_device *dev = pci_get_drvdata(pdev); | |
1586 | ||
1587 | /* | |
1588 | * FIXME: open_count is protected by drm_global_mutex but that would lead to | |
1589 | * locking inversion with the driver load path. And the access here is | |
1590 | * completely racy anyway. So don't bother with locking for now. | |
1591 | */ | |
7e13ad89 | 1592 | return atomic_read(&dev->open_count) == 0; |
d38ceaf9 AD |
1593 | } |
1594 | ||
1595 | static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = { | |
1596 | .set_gpu_state = amdgpu_switcheroo_set_state, | |
1597 | .reprobe = NULL, | |
1598 | .can_switch = amdgpu_switcheroo_can_switch, | |
1599 | }; | |
1600 | ||
e3ecdffa AD |
1601 | /** |
1602 | * amdgpu_device_ip_set_clockgating_state - set the CG state | |
1603 | * | |
87e3f136 | 1604 | * @dev: amdgpu_device pointer |
e3ecdffa AD |
1605 | * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) |
1606 | * @state: clockgating state (gate or ungate) | |
1607 | * | |
1608 | * Sets the requested clockgating state for all instances of | |
1609 | * the hardware IP specified. | |
1610 | * Returns the error code from the last instance. | |
1611 | */ | |
43fa561f | 1612 | int amdgpu_device_ip_set_clockgating_state(void *dev, |
2990a1fc AD |
1613 | enum amd_ip_block_type block_type, |
1614 | enum amd_clockgating_state state) | |
d38ceaf9 | 1615 | { |
43fa561f | 1616 | struct amdgpu_device *adev = dev; |
d38ceaf9 AD |
1617 | int i, r = 0; |
1618 | ||
1619 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 1620 | if (!adev->ip_blocks[i].status.valid) |
9ecbe7f5 | 1621 | continue; |
c722865a RZ |
1622 | if (adev->ip_blocks[i].version->type != block_type) |
1623 | continue; | |
1624 | if (!adev->ip_blocks[i].version->funcs->set_clockgating_state) | |
1625 | continue; | |
1626 | r = adev->ip_blocks[i].version->funcs->set_clockgating_state( | |
1627 | (void *)adev, state); | |
1628 | if (r) | |
1629 | DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n", | |
1630 | adev->ip_blocks[i].version->funcs->name, r); | |
d38ceaf9 AD |
1631 | } |
1632 | return r; | |
1633 | } | |
1634 | ||
e3ecdffa AD |
1635 | /** |
1636 | * amdgpu_device_ip_set_powergating_state - set the PG state | |
1637 | * | |
87e3f136 | 1638 | * @dev: amdgpu_device pointer |
e3ecdffa AD |
1639 | * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) |
1640 | * @state: powergating state (gate or ungate) | |
1641 | * | |
1642 | * Sets the requested powergating state for all instances of | |
1643 | * the hardware IP specified. | |
1644 | * Returns the error code from the last instance. | |
1645 | */ | |
43fa561f | 1646 | int amdgpu_device_ip_set_powergating_state(void *dev, |
2990a1fc AD |
1647 | enum amd_ip_block_type block_type, |
1648 | enum amd_powergating_state state) | |
d38ceaf9 | 1649 | { |
43fa561f | 1650 | struct amdgpu_device *adev = dev; |
d38ceaf9 AD |
1651 | int i, r = 0; |
1652 | ||
1653 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 1654 | if (!adev->ip_blocks[i].status.valid) |
9ecbe7f5 | 1655 | continue; |
c722865a RZ |
1656 | if (adev->ip_blocks[i].version->type != block_type) |
1657 | continue; | |
1658 | if (!adev->ip_blocks[i].version->funcs->set_powergating_state) | |
1659 | continue; | |
1660 | r = adev->ip_blocks[i].version->funcs->set_powergating_state( | |
1661 | (void *)adev, state); | |
1662 | if (r) | |
1663 | DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n", | |
1664 | adev->ip_blocks[i].version->funcs->name, r); | |
d38ceaf9 AD |
1665 | } |
1666 | return r; | |
1667 | } | |
1668 | ||
e3ecdffa AD |
1669 | /** |
1670 | * amdgpu_device_ip_get_clockgating_state - get the CG state | |
1671 | * | |
1672 | * @adev: amdgpu_device pointer | |
1673 | * @flags: clockgating feature flags | |
1674 | * | |
1675 | * Walks the list of IPs on the device and updates the clockgating | |
1676 | * flags for each IP. | |
1677 | * Updates @flags with the feature flags for each hardware IP where | |
1678 | * clockgating is enabled. | |
1679 | */ | |
2990a1fc AD |
1680 | void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev, |
1681 | u32 *flags) | |
6cb2d4e4 HR |
1682 | { |
1683 | int i; | |
1684 | ||
1685 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
1686 | if (!adev->ip_blocks[i].status.valid) | |
1687 | continue; | |
1688 | if (adev->ip_blocks[i].version->funcs->get_clockgating_state) | |
1689 | adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags); | |
1690 | } | |
1691 | } | |
1692 | ||
e3ecdffa AD |
1693 | /** |
1694 | * amdgpu_device_ip_wait_for_idle - wait for idle | |
1695 | * | |
1696 | * @adev: amdgpu_device pointer | |
1697 | * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) | |
1698 | * | |
1699 | * Waits for the request hardware IP to be idle. | |
1700 | * Returns 0 for success or a negative error code on failure. | |
1701 | */ | |
2990a1fc AD |
1702 | int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev, |
1703 | enum amd_ip_block_type block_type) | |
5dbbb60b AD |
1704 | { |
1705 | int i, r; | |
1706 | ||
1707 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 1708 | if (!adev->ip_blocks[i].status.valid) |
9ecbe7f5 | 1709 | continue; |
a1255107 AD |
1710 | if (adev->ip_blocks[i].version->type == block_type) { |
1711 | r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev); | |
5dbbb60b AD |
1712 | if (r) |
1713 | return r; | |
1714 | break; | |
1715 | } | |
1716 | } | |
1717 | return 0; | |
1718 | ||
1719 | } | |
1720 | ||
e3ecdffa AD |
1721 | /** |
1722 | * amdgpu_device_ip_is_idle - is the hardware IP idle | |
1723 | * | |
1724 | * @adev: amdgpu_device pointer | |
1725 | * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) | |
1726 | * | |
1727 | * Check if the hardware IP is idle or not. | |
1728 | * Returns true if it the IP is idle, false if not. | |
1729 | */ | |
2990a1fc AD |
1730 | bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev, |
1731 | enum amd_ip_block_type block_type) | |
5dbbb60b AD |
1732 | { |
1733 | int i; | |
1734 | ||
1735 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 1736 | if (!adev->ip_blocks[i].status.valid) |
9ecbe7f5 | 1737 | continue; |
a1255107 AD |
1738 | if (adev->ip_blocks[i].version->type == block_type) |
1739 | return adev->ip_blocks[i].version->funcs->is_idle((void *)adev); | |
5dbbb60b AD |
1740 | } |
1741 | return true; | |
1742 | ||
1743 | } | |
1744 | ||
e3ecdffa AD |
1745 | /** |
1746 | * amdgpu_device_ip_get_ip_block - get a hw IP pointer | |
1747 | * | |
1748 | * @adev: amdgpu_device pointer | |
87e3f136 | 1749 | * @type: Type of hardware IP (SMU, GFX, UVD, etc.) |
e3ecdffa AD |
1750 | * |
1751 | * Returns a pointer to the hardware IP block structure | |
1752 | * if it exists for the asic, otherwise NULL. | |
1753 | */ | |
2990a1fc AD |
1754 | struct amdgpu_ip_block * |
1755 | amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev, | |
1756 | enum amd_ip_block_type type) | |
d38ceaf9 AD |
1757 | { |
1758 | int i; | |
1759 | ||
1760 | for (i = 0; i < adev->num_ip_blocks; i++) | |
a1255107 | 1761 | if (adev->ip_blocks[i].version->type == type) |
d38ceaf9 AD |
1762 | return &adev->ip_blocks[i]; |
1763 | ||
1764 | return NULL; | |
1765 | } | |
1766 | ||
1767 | /** | |
2990a1fc | 1768 | * amdgpu_device_ip_block_version_cmp |
d38ceaf9 AD |
1769 | * |
1770 | * @adev: amdgpu_device pointer | |
5fc3aeeb | 1771 | * @type: enum amd_ip_block_type |
d38ceaf9 AD |
1772 | * @major: major version |
1773 | * @minor: minor version | |
1774 | * | |
1775 | * return 0 if equal or greater | |
1776 | * return 1 if smaller or the ip_block doesn't exist | |
1777 | */ | |
2990a1fc AD |
1778 | int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev, |
1779 | enum amd_ip_block_type type, | |
1780 | u32 major, u32 minor) | |
d38ceaf9 | 1781 | { |
2990a1fc | 1782 | struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type); |
d38ceaf9 | 1783 | |
a1255107 AD |
1784 | if (ip_block && ((ip_block->version->major > major) || |
1785 | ((ip_block->version->major == major) && | |
1786 | (ip_block->version->minor >= minor)))) | |
d38ceaf9 AD |
1787 | return 0; |
1788 | ||
1789 | return 1; | |
1790 | } | |
1791 | ||
a1255107 | 1792 | /** |
2990a1fc | 1793 | * amdgpu_device_ip_block_add |
a1255107 AD |
1794 | * |
1795 | * @adev: amdgpu_device pointer | |
1796 | * @ip_block_version: pointer to the IP to add | |
1797 | * | |
1798 | * Adds the IP block driver information to the collection of IPs | |
1799 | * on the asic. | |
1800 | */ | |
2990a1fc AD |
1801 | int amdgpu_device_ip_block_add(struct amdgpu_device *adev, |
1802 | const struct amdgpu_ip_block_version *ip_block_version) | |
a1255107 AD |
1803 | { |
1804 | if (!ip_block_version) | |
1805 | return -EINVAL; | |
1806 | ||
7bd939d0 LG |
1807 | switch (ip_block_version->type) { |
1808 | case AMD_IP_BLOCK_TYPE_VCN: | |
1809 | if (adev->harvest_ip_mask & AMD_HARVEST_IP_VCN_MASK) | |
1810 | return 0; | |
1811 | break; | |
1812 | case AMD_IP_BLOCK_TYPE_JPEG: | |
1813 | if (adev->harvest_ip_mask & AMD_HARVEST_IP_JPEG_MASK) | |
1814 | return 0; | |
1815 | break; | |
1816 | default: | |
1817 | break; | |
1818 | } | |
1819 | ||
e966a725 | 1820 | DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks, |
a0bae357 HR |
1821 | ip_block_version->funcs->name); |
1822 | ||
a1255107 AD |
1823 | adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version; |
1824 | ||
1825 | return 0; | |
1826 | } | |
1827 | ||
e3ecdffa AD |
1828 | /** |
1829 | * amdgpu_device_enable_virtual_display - enable virtual display feature | |
1830 | * | |
1831 | * @adev: amdgpu_device pointer | |
1832 | * | |
1833 | * Enabled the virtual display feature if the user has enabled it via | |
1834 | * the module parameter virtual_display. This feature provides a virtual | |
1835 | * display hardware on headless boards or in virtualized environments. | |
1836 | * This function parses and validates the configuration string specified by | |
1837 | * the user and configues the virtual display configuration (number of | |
1838 | * virtual connectors, crtcs, etc.) specified. | |
1839 | */ | |
483ef985 | 1840 | static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev) |
9accf2fd ED |
1841 | { |
1842 | adev->enable_virtual_display = false; | |
1843 | ||
1844 | if (amdgpu_virtual_display) { | |
8f66090b | 1845 | const char *pci_address_name = pci_name(adev->pdev); |
0f66356d | 1846 | char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname; |
9accf2fd ED |
1847 | |
1848 | pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL); | |
1849 | pciaddstr_tmp = pciaddstr; | |
0f66356d ED |
1850 | while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) { |
1851 | pciaddname = strsep(&pciaddname_tmp, ","); | |
967de2a9 YT |
1852 | if (!strcmp("all", pciaddname) |
1853 | || !strcmp(pci_address_name, pciaddname)) { | |
0f66356d ED |
1854 | long num_crtc; |
1855 | int res = -1; | |
1856 | ||
9accf2fd | 1857 | adev->enable_virtual_display = true; |
0f66356d ED |
1858 | |
1859 | if (pciaddname_tmp) | |
1860 | res = kstrtol(pciaddname_tmp, 10, | |
1861 | &num_crtc); | |
1862 | ||
1863 | if (!res) { | |
1864 | if (num_crtc < 1) | |
1865 | num_crtc = 1; | |
1866 | if (num_crtc > 6) | |
1867 | num_crtc = 6; | |
1868 | adev->mode_info.num_crtc = num_crtc; | |
1869 | } else { | |
1870 | adev->mode_info.num_crtc = 1; | |
1871 | } | |
9accf2fd ED |
1872 | break; |
1873 | } | |
1874 | } | |
1875 | ||
0f66356d ED |
1876 | DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n", |
1877 | amdgpu_virtual_display, pci_address_name, | |
1878 | adev->enable_virtual_display, adev->mode_info.num_crtc); | |
9accf2fd ED |
1879 | |
1880 | kfree(pciaddstr); | |
1881 | } | |
1882 | } | |
1883 | ||
e3ecdffa AD |
1884 | /** |
1885 | * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware | |
1886 | * | |
1887 | * @adev: amdgpu_device pointer | |
1888 | * | |
1889 | * Parses the asic configuration parameters specified in the gpu info | |
1890 | * firmware and makes them availale to the driver for use in configuring | |
1891 | * the asic. | |
1892 | * Returns 0 on success, -EINVAL on failure. | |
1893 | */ | |
e2a75f88 AD |
1894 | static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev) |
1895 | { | |
e2a75f88 | 1896 | const char *chip_name; |
c0a43457 | 1897 | char fw_name[40]; |
e2a75f88 AD |
1898 | int err; |
1899 | const struct gpu_info_firmware_header_v1_0 *hdr; | |
1900 | ||
ab4fe3e1 HR |
1901 | adev->firmware.gpu_info_fw = NULL; |
1902 | ||
72de33f8 | 1903 | if (adev->mman.discovery_bin) { |
258620d0 | 1904 | amdgpu_discovery_get_gfx_info(adev); |
cc375d8c TY |
1905 | |
1906 | /* | |
1907 | * FIXME: The bounding box is still needed by Navi12, so | |
1908 | * temporarily read it from gpu_info firmware. Should be droped | |
1909 | * when DAL no longer needs it. | |
1910 | */ | |
1911 | if (adev->asic_type != CHIP_NAVI12) | |
1912 | return 0; | |
258620d0 AD |
1913 | } |
1914 | ||
e2a75f88 | 1915 | switch (adev->asic_type) { |
e2a75f88 AD |
1916 | #ifdef CONFIG_DRM_AMDGPU_SI |
1917 | case CHIP_VERDE: | |
1918 | case CHIP_TAHITI: | |
1919 | case CHIP_PITCAIRN: | |
1920 | case CHIP_OLAND: | |
1921 | case CHIP_HAINAN: | |
1922 | #endif | |
1923 | #ifdef CONFIG_DRM_AMDGPU_CIK | |
1924 | case CHIP_BONAIRE: | |
1925 | case CHIP_HAWAII: | |
1926 | case CHIP_KAVERI: | |
1927 | case CHIP_KABINI: | |
1928 | case CHIP_MULLINS: | |
1929 | #endif | |
da87c30b AD |
1930 | case CHIP_TOPAZ: |
1931 | case CHIP_TONGA: | |
1932 | case CHIP_FIJI: | |
1933 | case CHIP_POLARIS10: | |
1934 | case CHIP_POLARIS11: | |
1935 | case CHIP_POLARIS12: | |
1936 | case CHIP_VEGAM: | |
1937 | case CHIP_CARRIZO: | |
1938 | case CHIP_STONEY: | |
27c0bc71 | 1939 | case CHIP_VEGA20: |
44b3253a | 1940 | case CHIP_ALDEBARAN: |
84d244a3 JC |
1941 | case CHIP_SIENNA_CICHLID: |
1942 | case CHIP_NAVY_FLOUNDER: | |
eac88a5f | 1943 | case CHIP_DIMGREY_CAVEFISH: |
0e5f4b09 | 1944 | case CHIP_BEIGE_GOBY: |
e2a75f88 AD |
1945 | default: |
1946 | return 0; | |
1947 | case CHIP_VEGA10: | |
1948 | chip_name = "vega10"; | |
1949 | break; | |
3f76dced AD |
1950 | case CHIP_VEGA12: |
1951 | chip_name = "vega12"; | |
1952 | break; | |
2d2e5e7e | 1953 | case CHIP_RAVEN: |
54f78a76 | 1954 | if (adev->apu_flags & AMD_APU_IS_RAVEN2) |
54c4d17e | 1955 | chip_name = "raven2"; |
54f78a76 | 1956 | else if (adev->apu_flags & AMD_APU_IS_PICASSO) |
741deade | 1957 | chip_name = "picasso"; |
54c4d17e FX |
1958 | else |
1959 | chip_name = "raven"; | |
2d2e5e7e | 1960 | break; |
65e60f6e LM |
1961 | case CHIP_ARCTURUS: |
1962 | chip_name = "arcturus"; | |
1963 | break; | |
b51a26a0 | 1964 | case CHIP_RENOIR: |
2e62f0b5 PL |
1965 | if (adev->apu_flags & AMD_APU_IS_RENOIR) |
1966 | chip_name = "renoir"; | |
1967 | else | |
1968 | chip_name = "green_sardine"; | |
b51a26a0 | 1969 | break; |
23c6268e HR |
1970 | case CHIP_NAVI10: |
1971 | chip_name = "navi10"; | |
1972 | break; | |
ed42cfe1 XY |
1973 | case CHIP_NAVI14: |
1974 | chip_name = "navi14"; | |
1975 | break; | |
42b325e5 XY |
1976 | case CHIP_NAVI12: |
1977 | chip_name = "navi12"; | |
1978 | break; | |
4e52a9f8 HR |
1979 | case CHIP_VANGOGH: |
1980 | chip_name = "vangogh"; | |
1981 | break; | |
8bf84f60 AL |
1982 | case CHIP_YELLOW_CARP: |
1983 | chip_name = "yellow_carp"; | |
1984 | break; | |
e2a75f88 AD |
1985 | } |
1986 | ||
1987 | snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name); | |
ab4fe3e1 | 1988 | err = request_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev); |
e2a75f88 AD |
1989 | if (err) { |
1990 | dev_err(adev->dev, | |
1991 | "Failed to load gpu_info firmware \"%s\"\n", | |
1992 | fw_name); | |
1993 | goto out; | |
1994 | } | |
ab4fe3e1 | 1995 | err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw); |
e2a75f88 AD |
1996 | if (err) { |
1997 | dev_err(adev->dev, | |
1998 | "Failed to validate gpu_info firmware \"%s\"\n", | |
1999 | fw_name); | |
2000 | goto out; | |
2001 | } | |
2002 | ||
ab4fe3e1 | 2003 | hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data; |
e2a75f88 AD |
2004 | amdgpu_ucode_print_gpu_info_hdr(&hdr->header); |
2005 | ||
2006 | switch (hdr->version_major) { | |
2007 | case 1: | |
2008 | { | |
2009 | const struct gpu_info_firmware_v1_0 *gpu_info_fw = | |
ab4fe3e1 | 2010 | (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data + |
e2a75f88 AD |
2011 | le32_to_cpu(hdr->header.ucode_array_offset_bytes)); |
2012 | ||
cc375d8c TY |
2013 | /* |
2014 | * Should be droped when DAL no longer needs it. | |
2015 | */ | |
2016 | if (adev->asic_type == CHIP_NAVI12) | |
ec51d3fa XY |
2017 | goto parse_soc_bounding_box; |
2018 | ||
b5ab16bf AD |
2019 | adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se); |
2020 | adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh); | |
2021 | adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se); | |
2022 | adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se); | |
e2a75f88 | 2023 | adev->gfx.config.max_texture_channel_caches = |
b5ab16bf AD |
2024 | le32_to_cpu(gpu_info_fw->gc_num_tccs); |
2025 | adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs); | |
2026 | adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds); | |
2027 | adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth); | |
2028 | adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth); | |
e2a75f88 | 2029 | adev->gfx.config.double_offchip_lds_buf = |
b5ab16bf AD |
2030 | le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer); |
2031 | adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size); | |
51fd0370 HZ |
2032 | adev->gfx.cu_info.max_waves_per_simd = |
2033 | le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd); | |
2034 | adev->gfx.cu_info.max_scratch_slots_per_cu = | |
2035 | le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu); | |
2036 | adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size); | |
48321c3d | 2037 | if (hdr->version_minor >= 1) { |
35c2e910 HZ |
2038 | const struct gpu_info_firmware_v1_1 *gpu_info_fw = |
2039 | (const struct gpu_info_firmware_v1_1 *)(adev->firmware.gpu_info_fw->data + | |
2040 | le32_to_cpu(hdr->header.ucode_array_offset_bytes)); | |
2041 | adev->gfx.config.num_sc_per_sh = | |
2042 | le32_to_cpu(gpu_info_fw->num_sc_per_sh); | |
2043 | adev->gfx.config.num_packer_per_sc = | |
2044 | le32_to_cpu(gpu_info_fw->num_packer_per_sc); | |
2045 | } | |
ec51d3fa XY |
2046 | |
2047 | parse_soc_bounding_box: | |
ec51d3fa XY |
2048 | /* |
2049 | * soc bounding box info is not integrated in disocovery table, | |
258620d0 | 2050 | * we always need to parse it from gpu info firmware if needed. |
ec51d3fa | 2051 | */ |
48321c3d HW |
2052 | if (hdr->version_minor == 2) { |
2053 | const struct gpu_info_firmware_v1_2 *gpu_info_fw = | |
2054 | (const struct gpu_info_firmware_v1_2 *)(adev->firmware.gpu_info_fw->data + | |
2055 | le32_to_cpu(hdr->header.ucode_array_offset_bytes)); | |
2056 | adev->dm.soc_bounding_box = &gpu_info_fw->soc_bounding_box; | |
2057 | } | |
e2a75f88 AD |
2058 | break; |
2059 | } | |
2060 | default: | |
2061 | dev_err(adev->dev, | |
2062 | "Unsupported gpu_info table %d\n", hdr->header.ucode_version); | |
2063 | err = -EINVAL; | |
2064 | goto out; | |
2065 | } | |
2066 | out: | |
e2a75f88 AD |
2067 | return err; |
2068 | } | |
2069 | ||
e3ecdffa AD |
2070 | /** |
2071 | * amdgpu_device_ip_early_init - run early init for hardware IPs | |
2072 | * | |
2073 | * @adev: amdgpu_device pointer | |
2074 | * | |
2075 | * Early initialization pass for hardware IPs. The hardware IPs that make | |
2076 | * up each asic are discovered each IP's early_init callback is run. This | |
2077 | * is the first stage in initializing the asic. | |
2078 | * Returns 0 on success, negative error code on failure. | |
2079 | */ | |
06ec9070 | 2080 | static int amdgpu_device_ip_early_init(struct amdgpu_device *adev) |
d38ceaf9 | 2081 | { |
901e2be2 AD |
2082 | struct drm_device *dev = adev_to_drm(adev); |
2083 | struct pci_dev *parent; | |
aaa36a97 | 2084 | int i, r; |
d38ceaf9 | 2085 | |
483ef985 | 2086 | amdgpu_device_enable_virtual_display(adev); |
a6be7570 | 2087 | |
00a979f3 | 2088 | if (amdgpu_sriov_vf(adev)) { |
00a979f3 | 2089 | r = amdgpu_virt_request_full_gpu(adev, true); |
aaa36a97 AD |
2090 | if (r) |
2091 | return r; | |
00a979f3 WS |
2092 | } |
2093 | ||
d38ceaf9 | 2094 | switch (adev->asic_type) { |
33f34802 KW |
2095 | #ifdef CONFIG_DRM_AMDGPU_SI |
2096 | case CHIP_VERDE: | |
2097 | case CHIP_TAHITI: | |
2098 | case CHIP_PITCAIRN: | |
2099 | case CHIP_OLAND: | |
2100 | case CHIP_HAINAN: | |
295d0daf | 2101 | adev->family = AMDGPU_FAMILY_SI; |
33f34802 KW |
2102 | r = si_set_ip_blocks(adev); |
2103 | if (r) | |
2104 | return r; | |
2105 | break; | |
2106 | #endif | |
a2e73f56 AD |
2107 | #ifdef CONFIG_DRM_AMDGPU_CIK |
2108 | case CHIP_BONAIRE: | |
2109 | case CHIP_HAWAII: | |
2110 | case CHIP_KAVERI: | |
2111 | case CHIP_KABINI: | |
2112 | case CHIP_MULLINS: | |
e1ad2d53 | 2113 | if (adev->flags & AMD_IS_APU) |
a2e73f56 | 2114 | adev->family = AMDGPU_FAMILY_KV; |
e1ad2d53 AD |
2115 | else |
2116 | adev->family = AMDGPU_FAMILY_CI; | |
a2e73f56 AD |
2117 | |
2118 | r = cik_set_ip_blocks(adev); | |
2119 | if (r) | |
2120 | return r; | |
2121 | break; | |
2122 | #endif | |
da87c30b AD |
2123 | case CHIP_TOPAZ: |
2124 | case CHIP_TONGA: | |
2125 | case CHIP_FIJI: | |
2126 | case CHIP_POLARIS10: | |
2127 | case CHIP_POLARIS11: | |
2128 | case CHIP_POLARIS12: | |
2129 | case CHIP_VEGAM: | |
2130 | case CHIP_CARRIZO: | |
2131 | case CHIP_STONEY: | |
2132 | if (adev->flags & AMD_IS_APU) | |
2133 | adev->family = AMDGPU_FAMILY_CZ; | |
2134 | else | |
2135 | adev->family = AMDGPU_FAMILY_VI; | |
2136 | ||
2137 | r = vi_set_ip_blocks(adev); | |
2138 | if (r) | |
2139 | return r; | |
2140 | break; | |
d38ceaf9 | 2141 | default: |
63352b7f AD |
2142 | r = amdgpu_discovery_set_ip_blocks(adev); |
2143 | if (r) | |
2144 | return r; | |
2145 | break; | |
d38ceaf9 AD |
2146 | } |
2147 | ||
901e2be2 AD |
2148 | if (amdgpu_has_atpx() && |
2149 | (amdgpu_is_atpx_hybrid() || | |
2150 | amdgpu_has_atpx_dgpu_power_cntl()) && | |
2151 | ((adev->flags & AMD_IS_APU) == 0) && | |
2152 | !pci_is_thunderbolt_attached(to_pci_dev(dev->dev))) | |
2153 | adev->flags |= AMD_IS_PX; | |
2154 | ||
2155 | parent = pci_upstream_bridge(adev->pdev); | |
2156 | adev->has_pr3 = parent ? pci_pr3_present(parent) : false; | |
2157 | ||
1884734a | 2158 | amdgpu_amdkfd_device_probe(adev); |
2159 | ||
3b94fb10 | 2160 | adev->pm.pp_feature = amdgpu_pp_feature_mask; |
a35ad98b | 2161 | if (amdgpu_sriov_vf(adev) || sched_policy == KFD_SCHED_POLICY_NO_HWS) |
00544006 | 2162 | adev->pm.pp_feature &= ~PP_GFXOFF_MASK; |
4215a119 HC |
2163 | if (amdgpu_sriov_vf(adev) && adev->asic_type == CHIP_SIENNA_CICHLID) |
2164 | adev->pm.pp_feature &= ~PP_OVERDRIVE_MASK; | |
00f54b97 | 2165 | |
d38ceaf9 AD |
2166 | for (i = 0; i < adev->num_ip_blocks; i++) { |
2167 | if ((amdgpu_ip_block_mask & (1 << i)) == 0) { | |
ed8cf00c HR |
2168 | DRM_ERROR("disabled ip block: %d <%s>\n", |
2169 | i, adev->ip_blocks[i].version->funcs->name); | |
a1255107 | 2170 | adev->ip_blocks[i].status.valid = false; |
d38ceaf9 | 2171 | } else { |
a1255107 AD |
2172 | if (adev->ip_blocks[i].version->funcs->early_init) { |
2173 | r = adev->ip_blocks[i].version->funcs->early_init((void *)adev); | |
2c1a2784 | 2174 | if (r == -ENOENT) { |
a1255107 | 2175 | adev->ip_blocks[i].status.valid = false; |
2c1a2784 | 2176 | } else if (r) { |
a1255107 AD |
2177 | DRM_ERROR("early_init of IP block <%s> failed %d\n", |
2178 | adev->ip_blocks[i].version->funcs->name, r); | |
d38ceaf9 | 2179 | return r; |
2c1a2784 | 2180 | } else { |
a1255107 | 2181 | adev->ip_blocks[i].status.valid = true; |
2c1a2784 | 2182 | } |
974e6b64 | 2183 | } else { |
a1255107 | 2184 | adev->ip_blocks[i].status.valid = true; |
d38ceaf9 | 2185 | } |
d38ceaf9 | 2186 | } |
21a249ca AD |
2187 | /* get the vbios after the asic_funcs are set up */ |
2188 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) { | |
6e29c227 AD |
2189 | r = amdgpu_device_parse_gpu_info_fw(adev); |
2190 | if (r) | |
2191 | return r; | |
2192 | ||
21a249ca AD |
2193 | /* Read BIOS */ |
2194 | if (!amdgpu_get_bios(adev)) | |
2195 | return -EINVAL; | |
2196 | ||
2197 | r = amdgpu_atombios_init(adev); | |
2198 | if (r) { | |
2199 | dev_err(adev->dev, "amdgpu_atombios_init failed\n"); | |
2200 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0); | |
2201 | return r; | |
2202 | } | |
77eabc6f PJZ |
2203 | |
2204 | /*get pf2vf msg info at it's earliest time*/ | |
2205 | if (amdgpu_sriov_vf(adev)) | |
2206 | amdgpu_virt_init_data_exchange(adev); | |
2207 | ||
21a249ca | 2208 | } |
d38ceaf9 AD |
2209 | } |
2210 | ||
395d1fb9 NH |
2211 | adev->cg_flags &= amdgpu_cg_mask; |
2212 | adev->pg_flags &= amdgpu_pg_mask; | |
2213 | ||
d38ceaf9 AD |
2214 | return 0; |
2215 | } | |
2216 | ||
0a4f2520 RZ |
2217 | static int amdgpu_device_ip_hw_init_phase1(struct amdgpu_device *adev) |
2218 | { | |
2219 | int i, r; | |
2220 | ||
2221 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
2222 | if (!adev->ip_blocks[i].status.sw) | |
2223 | continue; | |
2224 | if (adev->ip_blocks[i].status.hw) | |
2225 | continue; | |
2226 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || | |
2d11fd3f | 2227 | (amdgpu_sriov_vf(adev) && (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)) || |
0a4f2520 RZ |
2228 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) { |
2229 | r = adev->ip_blocks[i].version->funcs->hw_init(adev); | |
2230 | if (r) { | |
2231 | DRM_ERROR("hw_init of IP block <%s> failed %d\n", | |
2232 | adev->ip_blocks[i].version->funcs->name, r); | |
2233 | return r; | |
2234 | } | |
2235 | adev->ip_blocks[i].status.hw = true; | |
2236 | } | |
2237 | } | |
2238 | ||
2239 | return 0; | |
2240 | } | |
2241 | ||
2242 | static int amdgpu_device_ip_hw_init_phase2(struct amdgpu_device *adev) | |
2243 | { | |
2244 | int i, r; | |
2245 | ||
2246 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
2247 | if (!adev->ip_blocks[i].status.sw) | |
2248 | continue; | |
2249 | if (adev->ip_blocks[i].status.hw) | |
2250 | continue; | |
2251 | r = adev->ip_blocks[i].version->funcs->hw_init(adev); | |
2252 | if (r) { | |
2253 | DRM_ERROR("hw_init of IP block <%s> failed %d\n", | |
2254 | adev->ip_blocks[i].version->funcs->name, r); | |
2255 | return r; | |
2256 | } | |
2257 | adev->ip_blocks[i].status.hw = true; | |
2258 | } | |
2259 | ||
2260 | return 0; | |
2261 | } | |
2262 | ||
7a3e0bb2 RZ |
2263 | static int amdgpu_device_fw_loading(struct amdgpu_device *adev) |
2264 | { | |
2265 | int r = 0; | |
2266 | int i; | |
80f41f84 | 2267 | uint32_t smu_version; |
7a3e0bb2 RZ |
2268 | |
2269 | if (adev->asic_type >= CHIP_VEGA10) { | |
2270 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
482f0e53 ML |
2271 | if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_PSP) |
2272 | continue; | |
2273 | ||
e3c1b071 | 2274 | if (!adev->ip_blocks[i].status.sw) |
2275 | continue; | |
2276 | ||
482f0e53 ML |
2277 | /* no need to do the fw loading again if already done*/ |
2278 | if (adev->ip_blocks[i].status.hw == true) | |
2279 | break; | |
2280 | ||
53b3f8f4 | 2281 | if (amdgpu_in_reset(adev) || adev->in_suspend) { |
482f0e53 ML |
2282 | r = adev->ip_blocks[i].version->funcs->resume(adev); |
2283 | if (r) { | |
2284 | DRM_ERROR("resume of IP block <%s> failed %d\n", | |
7a3e0bb2 | 2285 | adev->ip_blocks[i].version->funcs->name, r); |
482f0e53 ML |
2286 | return r; |
2287 | } | |
2288 | } else { | |
2289 | r = adev->ip_blocks[i].version->funcs->hw_init(adev); | |
2290 | if (r) { | |
2291 | DRM_ERROR("hw_init of IP block <%s> failed %d\n", | |
2292 | adev->ip_blocks[i].version->funcs->name, r); | |
2293 | return r; | |
7a3e0bb2 | 2294 | } |
7a3e0bb2 | 2295 | } |
482f0e53 ML |
2296 | |
2297 | adev->ip_blocks[i].status.hw = true; | |
2298 | break; | |
7a3e0bb2 RZ |
2299 | } |
2300 | } | |
482f0e53 | 2301 | |
8973d9ec ED |
2302 | if (!amdgpu_sriov_vf(adev) || adev->asic_type == CHIP_TONGA) |
2303 | r = amdgpu_pm_load_smu_firmware(adev, &smu_version); | |
7a3e0bb2 | 2304 | |
80f41f84 | 2305 | return r; |
7a3e0bb2 RZ |
2306 | } |
2307 | ||
e3ecdffa AD |
2308 | /** |
2309 | * amdgpu_device_ip_init - run init for hardware IPs | |
2310 | * | |
2311 | * @adev: amdgpu_device pointer | |
2312 | * | |
2313 | * Main initialization pass for hardware IPs. The list of all the hardware | |
2314 | * IPs that make up the asic is walked and the sw_init and hw_init callbacks | |
2315 | * are run. sw_init initializes the software state associated with each IP | |
2316 | * and hw_init initializes the hardware associated with each IP. | |
2317 | * Returns 0 on success, negative error code on failure. | |
2318 | */ | |
06ec9070 | 2319 | static int amdgpu_device_ip_init(struct amdgpu_device *adev) |
d38ceaf9 AD |
2320 | { |
2321 | int i, r; | |
2322 | ||
c030f2e4 | 2323 | r = amdgpu_ras_init(adev); |
2324 | if (r) | |
2325 | return r; | |
2326 | ||
d38ceaf9 | 2327 | for (i = 0; i < adev->num_ip_blocks; i++) { |
a1255107 | 2328 | if (!adev->ip_blocks[i].status.valid) |
d38ceaf9 | 2329 | continue; |
a1255107 | 2330 | r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev); |
2c1a2784 | 2331 | if (r) { |
a1255107 AD |
2332 | DRM_ERROR("sw_init of IP block <%s> failed %d\n", |
2333 | adev->ip_blocks[i].version->funcs->name, r); | |
72d3f592 | 2334 | goto init_failed; |
2c1a2784 | 2335 | } |
a1255107 | 2336 | adev->ip_blocks[i].status.sw = true; |
bfca0289 | 2337 | |
d38ceaf9 | 2338 | /* need to do gmc hw init early so we can allocate gpu mem */ |
a1255107 | 2339 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) { |
892deb48 VS |
2340 | /* Try to reserve bad pages early */ |
2341 | if (amdgpu_sriov_vf(adev)) | |
2342 | amdgpu_virt_exchange_data(adev); | |
2343 | ||
06ec9070 | 2344 | r = amdgpu_device_vram_scratch_init(adev); |
2c1a2784 AD |
2345 | if (r) { |
2346 | DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r); | |
72d3f592 | 2347 | goto init_failed; |
2c1a2784 | 2348 | } |
a1255107 | 2349 | r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev); |
2c1a2784 AD |
2350 | if (r) { |
2351 | DRM_ERROR("hw_init %d failed %d\n", i, r); | |
72d3f592 | 2352 | goto init_failed; |
2c1a2784 | 2353 | } |
06ec9070 | 2354 | r = amdgpu_device_wb_init(adev); |
2c1a2784 | 2355 | if (r) { |
06ec9070 | 2356 | DRM_ERROR("amdgpu_device_wb_init failed %d\n", r); |
72d3f592 | 2357 | goto init_failed; |
2c1a2784 | 2358 | } |
a1255107 | 2359 | adev->ip_blocks[i].status.hw = true; |
2493664f ML |
2360 | |
2361 | /* right after GMC hw init, we create CSA */ | |
f92d5c61 | 2362 | if (amdgpu_mcbp || amdgpu_sriov_vf(adev)) { |
1e256e27 RZ |
2363 | r = amdgpu_allocate_static_csa(adev, &adev->virt.csa_obj, |
2364 | AMDGPU_GEM_DOMAIN_VRAM, | |
2365 | AMDGPU_CSA_SIZE); | |
2493664f ML |
2366 | if (r) { |
2367 | DRM_ERROR("allocate CSA failed %d\n", r); | |
72d3f592 | 2368 | goto init_failed; |
2493664f ML |
2369 | } |
2370 | } | |
d38ceaf9 AD |
2371 | } |
2372 | } | |
2373 | ||
c9ffa427 | 2374 | if (amdgpu_sriov_vf(adev)) |
22c16d25 | 2375 | amdgpu_virt_init_data_exchange(adev); |
c9ffa427 | 2376 | |
533aed27 AG |
2377 | r = amdgpu_ib_pool_init(adev); |
2378 | if (r) { | |
2379 | dev_err(adev->dev, "IB initialization failed (%d).\n", r); | |
2380 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r); | |
2381 | goto init_failed; | |
2382 | } | |
2383 | ||
c8963ea4 RZ |
2384 | r = amdgpu_ucode_create_bo(adev); /* create ucode bo when sw_init complete*/ |
2385 | if (r) | |
72d3f592 | 2386 | goto init_failed; |
0a4f2520 RZ |
2387 | |
2388 | r = amdgpu_device_ip_hw_init_phase1(adev); | |
2389 | if (r) | |
72d3f592 | 2390 | goto init_failed; |
0a4f2520 | 2391 | |
7a3e0bb2 RZ |
2392 | r = amdgpu_device_fw_loading(adev); |
2393 | if (r) | |
72d3f592 | 2394 | goto init_failed; |
7a3e0bb2 | 2395 | |
0a4f2520 RZ |
2396 | r = amdgpu_device_ip_hw_init_phase2(adev); |
2397 | if (r) | |
72d3f592 | 2398 | goto init_failed; |
d38ceaf9 | 2399 | |
121a2bc6 AG |
2400 | /* |
2401 | * retired pages will be loaded from eeprom and reserved here, | |
2402 | * it should be called after amdgpu_device_ip_hw_init_phase2 since | |
2403 | * for some ASICs the RAS EEPROM code relies on SMU fully functioning | |
2404 | * for I2C communication which only true at this point. | |
b82e65a9 GC |
2405 | * |
2406 | * amdgpu_ras_recovery_init may fail, but the upper only cares the | |
2407 | * failure from bad gpu situation and stop amdgpu init process | |
2408 | * accordingly. For other failed cases, it will still release all | |
2409 | * the resource and print error message, rather than returning one | |
2410 | * negative value to upper level. | |
121a2bc6 AG |
2411 | * |
2412 | * Note: theoretically, this should be called before all vram allocations | |
2413 | * to protect retired page from abusing | |
2414 | */ | |
b82e65a9 GC |
2415 | r = amdgpu_ras_recovery_init(adev); |
2416 | if (r) | |
2417 | goto init_failed; | |
121a2bc6 | 2418 | |
3e2e2ab5 HZ |
2419 | if (adev->gmc.xgmi.num_physical_nodes > 1) |
2420 | amdgpu_xgmi_add_device(adev); | |
e3c1b071 | 2421 | |
2422 | /* Don't init kfd if whole hive need to be reset during init */ | |
2423 | if (!adev->gmc.xgmi.pending_reset) | |
2424 | amdgpu_amdkfd_device_init(adev); | |
c6332b97 | 2425 | |
bd607166 KR |
2426 | amdgpu_fru_get_product_info(adev); |
2427 | ||
72d3f592 | 2428 | init_failed: |
c9ffa427 | 2429 | if (amdgpu_sriov_vf(adev)) |
c6332b97 | 2430 | amdgpu_virt_release_full_gpu(adev, true); |
2431 | ||
72d3f592 | 2432 | return r; |
d38ceaf9 AD |
2433 | } |
2434 | ||
e3ecdffa AD |
2435 | /** |
2436 | * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer | |
2437 | * | |
2438 | * @adev: amdgpu_device pointer | |
2439 | * | |
2440 | * Writes a reset magic value to the gart pointer in VRAM. The driver calls | |
2441 | * this function before a GPU reset. If the value is retained after a | |
2442 | * GPU reset, VRAM has not been lost. Some GPU resets may destry VRAM contents. | |
2443 | */ | |
06ec9070 | 2444 | static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev) |
0c49e0b8 CZ |
2445 | { |
2446 | memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM); | |
2447 | } | |
2448 | ||
e3ecdffa AD |
2449 | /** |
2450 | * amdgpu_device_check_vram_lost - check if vram is valid | |
2451 | * | |
2452 | * @adev: amdgpu_device pointer | |
2453 | * | |
2454 | * Checks the reset magic value written to the gart pointer in VRAM. | |
2455 | * The driver calls this after a GPU reset to see if the contents of | |
2456 | * VRAM is lost or now. | |
2457 | * returns true if vram is lost, false if not. | |
2458 | */ | |
06ec9070 | 2459 | static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev) |
0c49e0b8 | 2460 | { |
dadce777 EQ |
2461 | if (memcmp(adev->gart.ptr, adev->reset_magic, |
2462 | AMDGPU_RESET_MAGIC_NUM)) | |
2463 | return true; | |
2464 | ||
53b3f8f4 | 2465 | if (!amdgpu_in_reset(adev)) |
dadce777 EQ |
2466 | return false; |
2467 | ||
2468 | /* | |
2469 | * For all ASICs with baco/mode1 reset, the VRAM is | |
2470 | * always assumed to be lost. | |
2471 | */ | |
2472 | switch (amdgpu_asic_reset_method(adev)) { | |
2473 | case AMD_RESET_METHOD_BACO: | |
2474 | case AMD_RESET_METHOD_MODE1: | |
2475 | return true; | |
2476 | default: | |
2477 | return false; | |
2478 | } | |
0c49e0b8 CZ |
2479 | } |
2480 | ||
e3ecdffa | 2481 | /** |
1112a46b | 2482 | * amdgpu_device_set_cg_state - set clockgating for amdgpu device |
e3ecdffa AD |
2483 | * |
2484 | * @adev: amdgpu_device pointer | |
b8b72130 | 2485 | * @state: clockgating state (gate or ungate) |
e3ecdffa | 2486 | * |
e3ecdffa | 2487 | * The list of all the hardware IPs that make up the asic is walked and the |
1112a46b RZ |
2488 | * set_clockgating_state callbacks are run. |
2489 | * Late initialization pass enabling clockgating for hardware IPs. | |
2490 | * Fini or suspend, pass disabling clockgating for hardware IPs. | |
e3ecdffa AD |
2491 | * Returns 0 on success, negative error code on failure. |
2492 | */ | |
fdd34271 | 2493 | |
5d89bb2d LL |
2494 | int amdgpu_device_set_cg_state(struct amdgpu_device *adev, |
2495 | enum amd_clockgating_state state) | |
d38ceaf9 | 2496 | { |
1112a46b | 2497 | int i, j, r; |
d38ceaf9 | 2498 | |
4a2ba394 SL |
2499 | if (amdgpu_emu_mode == 1) |
2500 | return 0; | |
2501 | ||
1112a46b RZ |
2502 | for (j = 0; j < adev->num_ip_blocks; j++) { |
2503 | i = state == AMD_CG_STATE_GATE ? j : adev->num_ip_blocks - j - 1; | |
a2d31dc3 | 2504 | if (!adev->ip_blocks[i].status.late_initialized) |
d38ceaf9 | 2505 | continue; |
5d70a549 PV |
2506 | /* skip CG for GFX on S0ix */ |
2507 | if (adev->in_s0ix && | |
2508 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX) | |
2509 | continue; | |
4a446d55 | 2510 | /* skip CG for VCE/UVD, it's handled specially */ |
a1255107 | 2511 | if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD && |
57716327 | 2512 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE && |
34319b32 | 2513 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN && |
52f2e779 | 2514 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG && |
57716327 | 2515 | adev->ip_blocks[i].version->funcs->set_clockgating_state) { |
4a446d55 | 2516 | /* enable clockgating to save power */ |
a1255107 | 2517 | r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev, |
1112a46b | 2518 | state); |
4a446d55 AD |
2519 | if (r) { |
2520 | DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n", | |
a1255107 | 2521 | adev->ip_blocks[i].version->funcs->name, r); |
4a446d55 AD |
2522 | return r; |
2523 | } | |
b0b00ff1 | 2524 | } |
d38ceaf9 | 2525 | } |
06b18f61 | 2526 | |
c9f96fd5 RZ |
2527 | return 0; |
2528 | } | |
2529 | ||
5d89bb2d LL |
2530 | int amdgpu_device_set_pg_state(struct amdgpu_device *adev, |
2531 | enum amd_powergating_state state) | |
c9f96fd5 | 2532 | { |
1112a46b | 2533 | int i, j, r; |
06b18f61 | 2534 | |
c9f96fd5 RZ |
2535 | if (amdgpu_emu_mode == 1) |
2536 | return 0; | |
2537 | ||
1112a46b RZ |
2538 | for (j = 0; j < adev->num_ip_blocks; j++) { |
2539 | i = state == AMD_PG_STATE_GATE ? j : adev->num_ip_blocks - j - 1; | |
a2d31dc3 | 2540 | if (!adev->ip_blocks[i].status.late_initialized) |
c9f96fd5 | 2541 | continue; |
5d70a549 PV |
2542 | /* skip PG for GFX on S0ix */ |
2543 | if (adev->in_s0ix && | |
2544 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX) | |
2545 | continue; | |
c9f96fd5 RZ |
2546 | /* skip CG for VCE/UVD, it's handled specially */ |
2547 | if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD && | |
2548 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE && | |
2549 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN && | |
52f2e779 | 2550 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG && |
c9f96fd5 RZ |
2551 | adev->ip_blocks[i].version->funcs->set_powergating_state) { |
2552 | /* enable powergating to save power */ | |
2553 | r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev, | |
1112a46b | 2554 | state); |
c9f96fd5 RZ |
2555 | if (r) { |
2556 | DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n", | |
2557 | adev->ip_blocks[i].version->funcs->name, r); | |
2558 | return r; | |
2559 | } | |
2560 | } | |
2561 | } | |
2dc80b00 S |
2562 | return 0; |
2563 | } | |
2564 | ||
beff74bc AD |
2565 | static int amdgpu_device_enable_mgpu_fan_boost(void) |
2566 | { | |
2567 | struct amdgpu_gpu_instance *gpu_ins; | |
2568 | struct amdgpu_device *adev; | |
2569 | int i, ret = 0; | |
2570 | ||
2571 | mutex_lock(&mgpu_info.mutex); | |
2572 | ||
2573 | /* | |
2574 | * MGPU fan boost feature should be enabled | |
2575 | * only when there are two or more dGPUs in | |
2576 | * the system | |
2577 | */ | |
2578 | if (mgpu_info.num_dgpu < 2) | |
2579 | goto out; | |
2580 | ||
2581 | for (i = 0; i < mgpu_info.num_dgpu; i++) { | |
2582 | gpu_ins = &(mgpu_info.gpu_ins[i]); | |
2583 | adev = gpu_ins->adev; | |
2584 | if (!(adev->flags & AMD_IS_APU) && | |
f10bb940 | 2585 | !gpu_ins->mgpu_fan_enabled) { |
beff74bc AD |
2586 | ret = amdgpu_dpm_enable_mgpu_fan_boost(adev); |
2587 | if (ret) | |
2588 | break; | |
2589 | ||
2590 | gpu_ins->mgpu_fan_enabled = 1; | |
2591 | } | |
2592 | } | |
2593 | ||
2594 | out: | |
2595 | mutex_unlock(&mgpu_info.mutex); | |
2596 | ||
2597 | return ret; | |
2598 | } | |
2599 | ||
e3ecdffa AD |
2600 | /** |
2601 | * amdgpu_device_ip_late_init - run late init for hardware IPs | |
2602 | * | |
2603 | * @adev: amdgpu_device pointer | |
2604 | * | |
2605 | * Late initialization pass for hardware IPs. The list of all the hardware | |
2606 | * IPs that make up the asic is walked and the late_init callbacks are run. | |
2607 | * late_init covers any special initialization that an IP requires | |
2608 | * after all of the have been initialized or something that needs to happen | |
2609 | * late in the init process. | |
2610 | * Returns 0 on success, negative error code on failure. | |
2611 | */ | |
06ec9070 | 2612 | static int amdgpu_device_ip_late_init(struct amdgpu_device *adev) |
2dc80b00 | 2613 | { |
60599a03 | 2614 | struct amdgpu_gpu_instance *gpu_instance; |
2dc80b00 S |
2615 | int i = 0, r; |
2616 | ||
2617 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
73f847db | 2618 | if (!adev->ip_blocks[i].status.hw) |
2dc80b00 S |
2619 | continue; |
2620 | if (adev->ip_blocks[i].version->funcs->late_init) { | |
2621 | r = adev->ip_blocks[i].version->funcs->late_init((void *)adev); | |
2622 | if (r) { | |
2623 | DRM_ERROR("late_init of IP block <%s> failed %d\n", | |
2624 | adev->ip_blocks[i].version->funcs->name, r); | |
2625 | return r; | |
2626 | } | |
2dc80b00 | 2627 | } |
73f847db | 2628 | adev->ip_blocks[i].status.late_initialized = true; |
2dc80b00 S |
2629 | } |
2630 | ||
a891d239 DL |
2631 | amdgpu_ras_set_error_query_ready(adev, true); |
2632 | ||
1112a46b RZ |
2633 | amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE); |
2634 | amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE); | |
916ac57f | 2635 | |
06ec9070 | 2636 | amdgpu_device_fill_reset_magic(adev); |
d38ceaf9 | 2637 | |
beff74bc AD |
2638 | r = amdgpu_device_enable_mgpu_fan_boost(); |
2639 | if (r) | |
2640 | DRM_ERROR("enable mgpu fan boost failed (%d).\n", r); | |
2641 | ||
4da8b639 | 2642 | /* For passthrough configuration on arcturus and aldebaran, enable special handling SBR */ |
2643 | if (amdgpu_passthrough(adev) && ((adev->asic_type == CHIP_ARCTURUS && adev->gmc.xgmi.num_physical_nodes > 1)|| | |
2644 | adev->asic_type == CHIP_ALDEBARAN )) | |
bc143d8b | 2645 | amdgpu_dpm_handle_passthrough_sbr(adev, true); |
60599a03 EQ |
2646 | |
2647 | if (adev->gmc.xgmi.num_physical_nodes > 1) { | |
2648 | mutex_lock(&mgpu_info.mutex); | |
2649 | ||
2650 | /* | |
2651 | * Reset device p-state to low as this was booted with high. | |
2652 | * | |
2653 | * This should be performed only after all devices from the same | |
2654 | * hive get initialized. | |
2655 | * | |
2656 | * However, it's unknown how many device in the hive in advance. | |
2657 | * As this is counted one by one during devices initializations. | |
2658 | * | |
2659 | * So, we wait for all XGMI interlinked devices initialized. | |
2660 | * This may bring some delays as those devices may come from | |
2661 | * different hives. But that should be OK. | |
2662 | */ | |
2663 | if (mgpu_info.num_dgpu == adev->gmc.xgmi.num_physical_nodes) { | |
2664 | for (i = 0; i < mgpu_info.num_gpu; i++) { | |
2665 | gpu_instance = &(mgpu_info.gpu_ins[i]); | |
2666 | if (gpu_instance->adev->flags & AMD_IS_APU) | |
2667 | continue; | |
2668 | ||
d84a430d JK |
2669 | r = amdgpu_xgmi_set_pstate(gpu_instance->adev, |
2670 | AMDGPU_XGMI_PSTATE_MIN); | |
60599a03 EQ |
2671 | if (r) { |
2672 | DRM_ERROR("pstate setting failed (%d).\n", r); | |
2673 | break; | |
2674 | } | |
2675 | } | |
2676 | } | |
2677 | ||
2678 | mutex_unlock(&mgpu_info.mutex); | |
2679 | } | |
2680 | ||
d38ceaf9 AD |
2681 | return 0; |
2682 | } | |
2683 | ||
613aa3ea LY |
2684 | /** |
2685 | * amdgpu_device_smu_fini_early - smu hw_fini wrapper | |
2686 | * | |
2687 | * @adev: amdgpu_device pointer | |
2688 | * | |
2689 | * For ASICs need to disable SMC first | |
2690 | */ | |
2691 | static void amdgpu_device_smu_fini_early(struct amdgpu_device *adev) | |
2692 | { | |
2693 | int i, r; | |
2694 | ||
2695 | if (adev->ip_versions[GC_HWIP][0] > IP_VERSION(9, 0, 0)) | |
2696 | return; | |
2697 | ||
2698 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
2699 | if (!adev->ip_blocks[i].status.hw) | |
2700 | continue; | |
2701 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) { | |
2702 | r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev); | |
2703 | /* XXX handle errors */ | |
2704 | if (r) { | |
2705 | DRM_DEBUG("hw_fini of IP block <%s> failed %d\n", | |
2706 | adev->ip_blocks[i].version->funcs->name, r); | |
2707 | } | |
2708 | adev->ip_blocks[i].status.hw = false; | |
2709 | break; | |
2710 | } | |
2711 | } | |
2712 | } | |
2713 | ||
e9669fb7 | 2714 | static int amdgpu_device_ip_fini_early(struct amdgpu_device *adev) |
d38ceaf9 AD |
2715 | { |
2716 | int i, r; | |
2717 | ||
e9669fb7 AG |
2718 | for (i = 0; i < adev->num_ip_blocks; i++) { |
2719 | if (!adev->ip_blocks[i].version->funcs->early_fini) | |
2720 | continue; | |
5278a159 | 2721 | |
e9669fb7 AG |
2722 | r = adev->ip_blocks[i].version->funcs->early_fini((void *)adev); |
2723 | if (r) { | |
2724 | DRM_DEBUG("early_fini of IP block <%s> failed %d\n", | |
2725 | adev->ip_blocks[i].version->funcs->name, r); | |
2726 | } | |
2727 | } | |
c030f2e4 | 2728 | |
05df1f01 | 2729 | amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE); |
fdd34271 RZ |
2730 | amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE); |
2731 | ||
7270e895 TY |
2732 | amdgpu_amdkfd_suspend(adev, false); |
2733 | ||
613aa3ea LY |
2734 | /* Workaroud for ASICs need to disable SMC first */ |
2735 | amdgpu_device_smu_fini_early(adev); | |
3e96dbfd | 2736 | |
d38ceaf9 | 2737 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { |
a1255107 | 2738 | if (!adev->ip_blocks[i].status.hw) |
d38ceaf9 | 2739 | continue; |
8201a67a | 2740 | |
a1255107 | 2741 | r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev); |
d38ceaf9 | 2742 | /* XXX handle errors */ |
2c1a2784 | 2743 | if (r) { |
a1255107 AD |
2744 | DRM_DEBUG("hw_fini of IP block <%s> failed %d\n", |
2745 | adev->ip_blocks[i].version->funcs->name, r); | |
2c1a2784 | 2746 | } |
8201a67a | 2747 | |
a1255107 | 2748 | adev->ip_blocks[i].status.hw = false; |
d38ceaf9 AD |
2749 | } |
2750 | ||
6effad8a GC |
2751 | if (amdgpu_sriov_vf(adev)) { |
2752 | if (amdgpu_virt_release_full_gpu(adev, false)) | |
2753 | DRM_ERROR("failed to release exclusive mode on fini\n"); | |
2754 | } | |
2755 | ||
e9669fb7 AG |
2756 | return 0; |
2757 | } | |
2758 | ||
2759 | /** | |
2760 | * amdgpu_device_ip_fini - run fini for hardware IPs | |
2761 | * | |
2762 | * @adev: amdgpu_device pointer | |
2763 | * | |
2764 | * Main teardown pass for hardware IPs. The list of all the hardware | |
2765 | * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks | |
2766 | * are run. hw_fini tears down the hardware associated with each IP | |
2767 | * and sw_fini tears down any software state associated with each IP. | |
2768 | * Returns 0 on success, negative error code on failure. | |
2769 | */ | |
2770 | static int amdgpu_device_ip_fini(struct amdgpu_device *adev) | |
2771 | { | |
2772 | int i, r; | |
2773 | ||
2774 | if (amdgpu_sriov_vf(adev) && adev->virt.ras_init_done) | |
2775 | amdgpu_virt_release_ras_err_handler_data(adev); | |
2776 | ||
e9669fb7 AG |
2777 | if (adev->gmc.xgmi.num_physical_nodes > 1) |
2778 | amdgpu_xgmi_remove_device(adev); | |
2779 | ||
2780 | amdgpu_amdkfd_device_fini_sw(adev); | |
9950cda2 | 2781 | |
d38ceaf9 | 2782 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { |
a1255107 | 2783 | if (!adev->ip_blocks[i].status.sw) |
d38ceaf9 | 2784 | continue; |
c12aba3a ML |
2785 | |
2786 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) { | |
c8963ea4 | 2787 | amdgpu_ucode_free_bo(adev); |
1e256e27 | 2788 | amdgpu_free_static_csa(&adev->virt.csa_obj); |
c12aba3a ML |
2789 | amdgpu_device_wb_fini(adev); |
2790 | amdgpu_device_vram_scratch_fini(adev); | |
533aed27 | 2791 | amdgpu_ib_pool_fini(adev); |
c12aba3a ML |
2792 | } |
2793 | ||
a1255107 | 2794 | r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev); |
d38ceaf9 | 2795 | /* XXX handle errors */ |
2c1a2784 | 2796 | if (r) { |
a1255107 AD |
2797 | DRM_DEBUG("sw_fini of IP block <%s> failed %d\n", |
2798 | adev->ip_blocks[i].version->funcs->name, r); | |
2c1a2784 | 2799 | } |
a1255107 AD |
2800 | adev->ip_blocks[i].status.sw = false; |
2801 | adev->ip_blocks[i].status.valid = false; | |
d38ceaf9 AD |
2802 | } |
2803 | ||
a6dcfd9c | 2804 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { |
a1255107 | 2805 | if (!adev->ip_blocks[i].status.late_initialized) |
8a2eef1d | 2806 | continue; |
a1255107 AD |
2807 | if (adev->ip_blocks[i].version->funcs->late_fini) |
2808 | adev->ip_blocks[i].version->funcs->late_fini((void *)adev); | |
2809 | adev->ip_blocks[i].status.late_initialized = false; | |
a6dcfd9c ML |
2810 | } |
2811 | ||
c030f2e4 | 2812 | amdgpu_ras_fini(adev); |
2813 | ||
d38ceaf9 AD |
2814 | return 0; |
2815 | } | |
2816 | ||
e3ecdffa | 2817 | /** |
beff74bc | 2818 | * amdgpu_device_delayed_init_work_handler - work handler for IB tests |
e3ecdffa | 2819 | * |
1112a46b | 2820 | * @work: work_struct. |
e3ecdffa | 2821 | */ |
beff74bc | 2822 | static void amdgpu_device_delayed_init_work_handler(struct work_struct *work) |
2dc80b00 S |
2823 | { |
2824 | struct amdgpu_device *adev = | |
beff74bc | 2825 | container_of(work, struct amdgpu_device, delayed_init_work.work); |
916ac57f RZ |
2826 | int r; |
2827 | ||
2828 | r = amdgpu_ib_ring_tests(adev); | |
2829 | if (r) | |
2830 | DRM_ERROR("ib ring test failed (%d).\n", r); | |
2dc80b00 S |
2831 | } |
2832 | ||
1e317b99 RZ |
2833 | static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work) |
2834 | { | |
2835 | struct amdgpu_device *adev = | |
2836 | container_of(work, struct amdgpu_device, gfx.gfx_off_delay_work.work); | |
2837 | ||
90a92662 MD |
2838 | WARN_ON_ONCE(adev->gfx.gfx_off_state); |
2839 | WARN_ON_ONCE(adev->gfx.gfx_off_req_count); | |
2840 | ||
2841 | if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, true)) | |
2842 | adev->gfx.gfx_off_state = true; | |
1e317b99 RZ |
2843 | } |
2844 | ||
e3ecdffa | 2845 | /** |
e7854a03 | 2846 | * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1) |
e3ecdffa AD |
2847 | * |
2848 | * @adev: amdgpu_device pointer | |
2849 | * | |
2850 | * Main suspend function for hardware IPs. The list of all the hardware | |
2851 | * IPs that make up the asic is walked, clockgating is disabled and the | |
2852 | * suspend callbacks are run. suspend puts the hardware and software state | |
2853 | * in each IP into a state suitable for suspend. | |
2854 | * Returns 0 on success, negative error code on failure. | |
2855 | */ | |
e7854a03 AD |
2856 | static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev) |
2857 | { | |
2858 | int i, r; | |
2859 | ||
50ec83f0 AD |
2860 | amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE); |
2861 | amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE); | |
05df1f01 | 2862 | |
e7854a03 AD |
2863 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { |
2864 | if (!adev->ip_blocks[i].status.valid) | |
2865 | continue; | |
2b9f7848 | 2866 | |
e7854a03 | 2867 | /* displays are handled separately */ |
2b9f7848 ND |
2868 | if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_DCE) |
2869 | continue; | |
2870 | ||
2871 | /* XXX handle errors */ | |
2872 | r = adev->ip_blocks[i].version->funcs->suspend(adev); | |
2873 | /* XXX handle errors */ | |
2874 | if (r) { | |
2875 | DRM_ERROR("suspend of IP block <%s> failed %d\n", | |
2876 | adev->ip_blocks[i].version->funcs->name, r); | |
2877 | return r; | |
e7854a03 | 2878 | } |
2b9f7848 ND |
2879 | |
2880 | adev->ip_blocks[i].status.hw = false; | |
e7854a03 AD |
2881 | } |
2882 | ||
e7854a03 AD |
2883 | return 0; |
2884 | } | |
2885 | ||
2886 | /** | |
2887 | * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2) | |
2888 | * | |
2889 | * @adev: amdgpu_device pointer | |
2890 | * | |
2891 | * Main suspend function for hardware IPs. The list of all the hardware | |
2892 | * IPs that make up the asic is walked, clockgating is disabled and the | |
2893 | * suspend callbacks are run. suspend puts the hardware and software state | |
2894 | * in each IP into a state suitable for suspend. | |
2895 | * Returns 0 on success, negative error code on failure. | |
2896 | */ | |
2897 | static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev) | |
d38ceaf9 AD |
2898 | { |
2899 | int i, r; | |
2900 | ||
557f42a2 | 2901 | if (adev->in_s0ix) |
bc143d8b | 2902 | amdgpu_dpm_gfx_state_change(adev, sGpuChangeState_D3Entry); |
34416931 | 2903 | |
d38ceaf9 | 2904 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { |
a1255107 | 2905 | if (!adev->ip_blocks[i].status.valid) |
d38ceaf9 | 2906 | continue; |
e7854a03 AD |
2907 | /* displays are handled in phase1 */ |
2908 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) | |
2909 | continue; | |
bff77e86 LM |
2910 | /* PSP lost connection when err_event_athub occurs */ |
2911 | if (amdgpu_ras_intr_triggered() && | |
2912 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) { | |
2913 | adev->ip_blocks[i].status.hw = false; | |
2914 | continue; | |
2915 | } | |
e3c1b071 | 2916 | |
2917 | /* skip unnecessary suspend if we do not initialize them yet */ | |
2918 | if (adev->gmc.xgmi.pending_reset && | |
2919 | !(adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC || | |
2920 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC || | |
2921 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || | |
2922 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH)) { | |
2923 | adev->ip_blocks[i].status.hw = false; | |
2924 | continue; | |
2925 | } | |
557f42a2 | 2926 | |
32ff160d AD |
2927 | /* skip suspend of gfx and psp for S0ix |
2928 | * gfx is in gfxoff state, so on resume it will exit gfxoff just | |
2929 | * like at runtime. PSP is also part of the always on hardware | |
2930 | * so no need to suspend it. | |
2931 | */ | |
557f42a2 | 2932 | if (adev->in_s0ix && |
32ff160d AD |
2933 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP || |
2934 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX)) | |
557f42a2 AD |
2935 | continue; |
2936 | ||
d38ceaf9 | 2937 | /* XXX handle errors */ |
a1255107 | 2938 | r = adev->ip_blocks[i].version->funcs->suspend(adev); |
d38ceaf9 | 2939 | /* XXX handle errors */ |
2c1a2784 | 2940 | if (r) { |
a1255107 AD |
2941 | DRM_ERROR("suspend of IP block <%s> failed %d\n", |
2942 | adev->ip_blocks[i].version->funcs->name, r); | |
2c1a2784 | 2943 | } |
876923fb | 2944 | adev->ip_blocks[i].status.hw = false; |
a3a09142 | 2945 | /* handle putting the SMC in the appropriate state */ |
86b93fd6 JZ |
2946 | if(!amdgpu_sriov_vf(adev)){ |
2947 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) { | |
2948 | r = amdgpu_dpm_set_mp1_state(adev, adev->mp1_state); | |
2949 | if (r) { | |
2950 | DRM_ERROR("SMC failed to set mp1 state %d, %d\n", | |
2951 | adev->mp1_state, r); | |
2952 | return r; | |
2953 | } | |
a3a09142 AD |
2954 | } |
2955 | } | |
d38ceaf9 AD |
2956 | } |
2957 | ||
2958 | return 0; | |
2959 | } | |
2960 | ||
e7854a03 AD |
2961 | /** |
2962 | * amdgpu_device_ip_suspend - run suspend for hardware IPs | |
2963 | * | |
2964 | * @adev: amdgpu_device pointer | |
2965 | * | |
2966 | * Main suspend function for hardware IPs. The list of all the hardware | |
2967 | * IPs that make up the asic is walked, clockgating is disabled and the | |
2968 | * suspend callbacks are run. suspend puts the hardware and software state | |
2969 | * in each IP into a state suitable for suspend. | |
2970 | * Returns 0 on success, negative error code on failure. | |
2971 | */ | |
2972 | int amdgpu_device_ip_suspend(struct amdgpu_device *adev) | |
2973 | { | |
2974 | int r; | |
2975 | ||
3c73683c JC |
2976 | if (amdgpu_sriov_vf(adev)) { |
2977 | amdgpu_virt_fini_data_exchange(adev); | |
e7819644 | 2978 | amdgpu_virt_request_full_gpu(adev, false); |
3c73683c | 2979 | } |
e7819644 | 2980 | |
e7854a03 AD |
2981 | r = amdgpu_device_ip_suspend_phase1(adev); |
2982 | if (r) | |
2983 | return r; | |
2984 | r = amdgpu_device_ip_suspend_phase2(adev); | |
2985 | ||
e7819644 YT |
2986 | if (amdgpu_sriov_vf(adev)) |
2987 | amdgpu_virt_release_full_gpu(adev, false); | |
2988 | ||
e7854a03 AD |
2989 | return r; |
2990 | } | |
2991 | ||
06ec9070 | 2992 | static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev) |
a90ad3c2 ML |
2993 | { |
2994 | int i, r; | |
2995 | ||
2cb681b6 ML |
2996 | static enum amd_ip_block_type ip_order[] = { |
2997 | AMD_IP_BLOCK_TYPE_GMC, | |
2998 | AMD_IP_BLOCK_TYPE_COMMON, | |
39186aef | 2999 | AMD_IP_BLOCK_TYPE_PSP, |
2cb681b6 ML |
3000 | AMD_IP_BLOCK_TYPE_IH, |
3001 | }; | |
a90ad3c2 | 3002 | |
95ea3dbc | 3003 | for (i = 0; i < adev->num_ip_blocks; i++) { |
2cb681b6 ML |
3004 | int j; |
3005 | struct amdgpu_ip_block *block; | |
a90ad3c2 | 3006 | |
4cd2a96d J |
3007 | block = &adev->ip_blocks[i]; |
3008 | block->status.hw = false; | |
2cb681b6 | 3009 | |
4cd2a96d | 3010 | for (j = 0; j < ARRAY_SIZE(ip_order); j++) { |
2cb681b6 | 3011 | |
4cd2a96d | 3012 | if (block->version->type != ip_order[j] || |
2cb681b6 ML |
3013 | !block->status.valid) |
3014 | continue; | |
3015 | ||
3016 | r = block->version->funcs->hw_init(adev); | |
0aaeefcc | 3017 | DRM_INFO("RE-INIT-early: %s %s\n", block->version->funcs->name, r?"failed":"succeeded"); |
c41d1cf6 ML |
3018 | if (r) |
3019 | return r; | |
482f0e53 | 3020 | block->status.hw = true; |
a90ad3c2 ML |
3021 | } |
3022 | } | |
3023 | ||
3024 | return 0; | |
3025 | } | |
3026 | ||
06ec9070 | 3027 | static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev) |
a90ad3c2 ML |
3028 | { |
3029 | int i, r; | |
3030 | ||
2cb681b6 ML |
3031 | static enum amd_ip_block_type ip_order[] = { |
3032 | AMD_IP_BLOCK_TYPE_SMC, | |
3033 | AMD_IP_BLOCK_TYPE_DCE, | |
3034 | AMD_IP_BLOCK_TYPE_GFX, | |
3035 | AMD_IP_BLOCK_TYPE_SDMA, | |
257deb8c | 3036 | AMD_IP_BLOCK_TYPE_UVD, |
d83c7a07 JJ |
3037 | AMD_IP_BLOCK_TYPE_VCE, |
3038 | AMD_IP_BLOCK_TYPE_VCN | |
2cb681b6 | 3039 | }; |
a90ad3c2 | 3040 | |
2cb681b6 ML |
3041 | for (i = 0; i < ARRAY_SIZE(ip_order); i++) { |
3042 | int j; | |
3043 | struct amdgpu_ip_block *block; | |
a90ad3c2 | 3044 | |
2cb681b6 ML |
3045 | for (j = 0; j < adev->num_ip_blocks; j++) { |
3046 | block = &adev->ip_blocks[j]; | |
3047 | ||
3048 | if (block->version->type != ip_order[i] || | |
482f0e53 ML |
3049 | !block->status.valid || |
3050 | block->status.hw) | |
2cb681b6 ML |
3051 | continue; |
3052 | ||
895bd048 JZ |
3053 | if (block->version->type == AMD_IP_BLOCK_TYPE_SMC) |
3054 | r = block->version->funcs->resume(adev); | |
3055 | else | |
3056 | r = block->version->funcs->hw_init(adev); | |
3057 | ||
0aaeefcc | 3058 | DRM_INFO("RE-INIT-late: %s %s\n", block->version->funcs->name, r?"failed":"succeeded"); |
c41d1cf6 ML |
3059 | if (r) |
3060 | return r; | |
482f0e53 | 3061 | block->status.hw = true; |
a90ad3c2 ML |
3062 | } |
3063 | } | |
3064 | ||
3065 | return 0; | |
3066 | } | |
3067 | ||
e3ecdffa AD |
3068 | /** |
3069 | * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs | |
3070 | * | |
3071 | * @adev: amdgpu_device pointer | |
3072 | * | |
3073 | * First resume function for hardware IPs. The list of all the hardware | |
3074 | * IPs that make up the asic is walked and the resume callbacks are run for | |
3075 | * COMMON, GMC, and IH. resume puts the hardware into a functional state | |
3076 | * after a suspend and updates the software state as necessary. This | |
3077 | * function is also used for restoring the GPU after a GPU reset. | |
3078 | * Returns 0 on success, negative error code on failure. | |
3079 | */ | |
06ec9070 | 3080 | static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev) |
d38ceaf9 AD |
3081 | { |
3082 | int i, r; | |
3083 | ||
a90ad3c2 | 3084 | for (i = 0; i < adev->num_ip_blocks; i++) { |
482f0e53 | 3085 | if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw) |
a90ad3c2 | 3086 | continue; |
a90ad3c2 | 3087 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || |
e3ecdffa AD |
3088 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC || |
3089 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) { | |
482f0e53 | 3090 | |
fcf0649f CZ |
3091 | r = adev->ip_blocks[i].version->funcs->resume(adev); |
3092 | if (r) { | |
3093 | DRM_ERROR("resume of IP block <%s> failed %d\n", | |
3094 | adev->ip_blocks[i].version->funcs->name, r); | |
3095 | return r; | |
3096 | } | |
482f0e53 | 3097 | adev->ip_blocks[i].status.hw = true; |
a90ad3c2 ML |
3098 | } |
3099 | } | |
3100 | ||
3101 | return 0; | |
3102 | } | |
3103 | ||
e3ecdffa AD |
3104 | /** |
3105 | * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs | |
3106 | * | |
3107 | * @adev: amdgpu_device pointer | |
3108 | * | |
3109 | * First resume function for hardware IPs. The list of all the hardware | |
3110 | * IPs that make up the asic is walked and the resume callbacks are run for | |
3111 | * all blocks except COMMON, GMC, and IH. resume puts the hardware into a | |
3112 | * functional state after a suspend and updates the software state as | |
3113 | * necessary. This function is also used for restoring the GPU after a GPU | |
3114 | * reset. | |
3115 | * Returns 0 on success, negative error code on failure. | |
3116 | */ | |
06ec9070 | 3117 | static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev) |
d38ceaf9 AD |
3118 | { |
3119 | int i, r; | |
3120 | ||
3121 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
482f0e53 | 3122 | if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw) |
d38ceaf9 | 3123 | continue; |
fcf0649f | 3124 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || |
e3ecdffa | 3125 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC || |
7a3e0bb2 RZ |
3126 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH || |
3127 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) | |
fcf0649f | 3128 | continue; |
a1255107 | 3129 | r = adev->ip_blocks[i].version->funcs->resume(adev); |
2c1a2784 | 3130 | if (r) { |
a1255107 AD |
3131 | DRM_ERROR("resume of IP block <%s> failed %d\n", |
3132 | adev->ip_blocks[i].version->funcs->name, r); | |
d38ceaf9 | 3133 | return r; |
2c1a2784 | 3134 | } |
482f0e53 | 3135 | adev->ip_blocks[i].status.hw = true; |
d38ceaf9 AD |
3136 | } |
3137 | ||
3138 | return 0; | |
3139 | } | |
3140 | ||
e3ecdffa AD |
3141 | /** |
3142 | * amdgpu_device_ip_resume - run resume for hardware IPs | |
3143 | * | |
3144 | * @adev: amdgpu_device pointer | |
3145 | * | |
3146 | * Main resume function for hardware IPs. The hardware IPs | |
3147 | * are split into two resume functions because they are | |
3148 | * are also used in in recovering from a GPU reset and some additional | |
3149 | * steps need to be take between them. In this case (S3/S4) they are | |
3150 | * run sequentially. | |
3151 | * Returns 0 on success, negative error code on failure. | |
3152 | */ | |
06ec9070 | 3153 | static int amdgpu_device_ip_resume(struct amdgpu_device *adev) |
fcf0649f CZ |
3154 | { |
3155 | int r; | |
3156 | ||
9cec53c1 JZ |
3157 | r = amdgpu_amdkfd_resume_iommu(adev); |
3158 | if (r) | |
3159 | return r; | |
3160 | ||
06ec9070 | 3161 | r = amdgpu_device_ip_resume_phase1(adev); |
fcf0649f CZ |
3162 | if (r) |
3163 | return r; | |
7a3e0bb2 RZ |
3164 | |
3165 | r = amdgpu_device_fw_loading(adev); | |
3166 | if (r) | |
3167 | return r; | |
3168 | ||
06ec9070 | 3169 | r = amdgpu_device_ip_resume_phase2(adev); |
fcf0649f CZ |
3170 | |
3171 | return r; | |
3172 | } | |
3173 | ||
e3ecdffa AD |
3174 | /** |
3175 | * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV | |
3176 | * | |
3177 | * @adev: amdgpu_device pointer | |
3178 | * | |
3179 | * Query the VBIOS data tables to determine if the board supports SR-IOV. | |
3180 | */ | |
4e99a44e | 3181 | static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev) |
048765ad | 3182 | { |
6867e1b5 ML |
3183 | if (amdgpu_sriov_vf(adev)) { |
3184 | if (adev->is_atom_fw) { | |
58ff791a | 3185 | if (amdgpu_atomfirmware_gpu_virtualization_supported(adev)) |
6867e1b5 ML |
3186 | adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS; |
3187 | } else { | |
3188 | if (amdgpu_atombios_has_gpu_virtualization_table(adev)) | |
3189 | adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS; | |
3190 | } | |
3191 | ||
3192 | if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS)) | |
3193 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0); | |
a5bde2f9 | 3194 | } |
048765ad AR |
3195 | } |
3196 | ||
e3ecdffa AD |
3197 | /** |
3198 | * amdgpu_device_asic_has_dc_support - determine if DC supports the asic | |
3199 | * | |
3200 | * @asic_type: AMD asic type | |
3201 | * | |
3202 | * Check if there is DC (new modesetting infrastructre) support for an asic. | |
3203 | * returns true if DC has support, false if not. | |
3204 | */ | |
4562236b HW |
3205 | bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type) |
3206 | { | |
3207 | switch (asic_type) { | |
0637d417 AD |
3208 | #ifdef CONFIG_DRM_AMDGPU_SI |
3209 | case CHIP_HAINAN: | |
3210 | #endif | |
3211 | case CHIP_TOPAZ: | |
3212 | /* chips with no display hardware */ | |
3213 | return false; | |
4562236b | 3214 | #if defined(CONFIG_DRM_AMD_DC) |
64200c46 MR |
3215 | case CHIP_TAHITI: |
3216 | case CHIP_PITCAIRN: | |
3217 | case CHIP_VERDE: | |
3218 | case CHIP_OLAND: | |
2d32ffd6 AD |
3219 | /* |
3220 | * We have systems in the wild with these ASICs that require | |
3221 | * LVDS and VGA support which is not supported with DC. | |
3222 | * | |
3223 | * Fallback to the non-DC driver here by default so as not to | |
3224 | * cause regressions. | |
3225 | */ | |
3226 | #if defined(CONFIG_DRM_AMD_DC_SI) | |
3227 | return amdgpu_dc > 0; | |
3228 | #else | |
3229 | return false; | |
64200c46 | 3230 | #endif |
4562236b | 3231 | case CHIP_BONAIRE: |
0d6fbccb | 3232 | case CHIP_KAVERI: |
367e6687 AD |
3233 | case CHIP_KABINI: |
3234 | case CHIP_MULLINS: | |
d9fda248 HW |
3235 | /* |
3236 | * We have systems in the wild with these ASICs that require | |
3237 | * LVDS and VGA support which is not supported with DC. | |
3238 | * | |
3239 | * Fallback to the non-DC driver here by default so as not to | |
3240 | * cause regressions. | |
3241 | */ | |
3242 | return amdgpu_dc > 0; | |
3243 | case CHIP_HAWAII: | |
4562236b HW |
3244 | case CHIP_CARRIZO: |
3245 | case CHIP_STONEY: | |
4562236b | 3246 | case CHIP_POLARIS10: |
675fd32b | 3247 | case CHIP_POLARIS11: |
2c8ad2d5 | 3248 | case CHIP_POLARIS12: |
675fd32b | 3249 | case CHIP_VEGAM: |
4562236b HW |
3250 | case CHIP_TONGA: |
3251 | case CHIP_FIJI: | |
42f8ffa1 | 3252 | case CHIP_VEGA10: |
dca7b401 | 3253 | case CHIP_VEGA12: |
c6034aa2 | 3254 | case CHIP_VEGA20: |
b86a1aa3 | 3255 | #if defined(CONFIG_DRM_AMD_DC_DCN) |
fd187853 | 3256 | case CHIP_RAVEN: |
b4f199c7 | 3257 | case CHIP_NAVI10: |
8fceceb6 | 3258 | case CHIP_NAVI14: |
078655d9 | 3259 | case CHIP_NAVI12: |
e1c14c43 | 3260 | case CHIP_RENOIR: |
3f68c01b | 3261 | case CHIP_CYAN_SKILLFISH: |
81d9bfb8 | 3262 | case CHIP_SIENNA_CICHLID: |
a6c5308f | 3263 | case CHIP_NAVY_FLOUNDER: |
7cc656e2 | 3264 | case CHIP_DIMGREY_CAVEFISH: |
ddaed58b | 3265 | case CHIP_BEIGE_GOBY: |
84b934bc | 3266 | case CHIP_VANGOGH: |
c8b73f7f | 3267 | case CHIP_YELLOW_CARP: |
42f8ffa1 | 3268 | #endif |
f7f12b25 | 3269 | default: |
fd187853 | 3270 | return amdgpu_dc != 0; |
f7f12b25 | 3271 | #else |
4562236b | 3272 | default: |
93b09a9a | 3273 | if (amdgpu_dc > 0) |
044a48f4 | 3274 | DRM_INFO_ONCE("Display Core has been requested via kernel parameter " |
93b09a9a | 3275 | "but isn't supported by ASIC, ignoring\n"); |
4562236b | 3276 | return false; |
f7f12b25 | 3277 | #endif |
4562236b HW |
3278 | } |
3279 | } | |
3280 | ||
3281 | /** | |
3282 | * amdgpu_device_has_dc_support - check if dc is supported | |
3283 | * | |
982a820b | 3284 | * @adev: amdgpu_device pointer |
4562236b HW |
3285 | * |
3286 | * Returns true for supported, false for not supported | |
3287 | */ | |
3288 | bool amdgpu_device_has_dc_support(struct amdgpu_device *adev) | |
3289 | { | |
abaf210c AS |
3290 | if (amdgpu_sriov_vf(adev) || |
3291 | adev->enable_virtual_display || | |
3292 | (adev->harvest_ip_mask & AMD_HARVEST_IP_DMU_MASK)) | |
2555039d XY |
3293 | return false; |
3294 | ||
4562236b HW |
3295 | return amdgpu_device_asic_has_dc_support(adev->asic_type); |
3296 | } | |
3297 | ||
d4535e2c AG |
3298 | static void amdgpu_device_xgmi_reset_func(struct work_struct *__work) |
3299 | { | |
3300 | struct amdgpu_device *adev = | |
3301 | container_of(__work, struct amdgpu_device, xgmi_reset_work); | |
d95e8e97 | 3302 | struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev); |
d4535e2c | 3303 | |
c6a6e2db AG |
3304 | /* It's a bug to not have a hive within this function */ |
3305 | if (WARN_ON(!hive)) | |
3306 | return; | |
3307 | ||
3308 | /* | |
3309 | * Use task barrier to synchronize all xgmi reset works across the | |
3310 | * hive. task_barrier_enter and task_barrier_exit will block | |
3311 | * until all the threads running the xgmi reset works reach | |
3312 | * those points. task_barrier_full will do both blocks. | |
3313 | */ | |
3314 | if (amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) { | |
3315 | ||
3316 | task_barrier_enter(&hive->tb); | |
4a580877 | 3317 | adev->asic_reset_res = amdgpu_device_baco_enter(adev_to_drm(adev)); |
c6a6e2db AG |
3318 | |
3319 | if (adev->asic_reset_res) | |
3320 | goto fail; | |
3321 | ||
3322 | task_barrier_exit(&hive->tb); | |
4a580877 | 3323 | adev->asic_reset_res = amdgpu_device_baco_exit(adev_to_drm(adev)); |
c6a6e2db AG |
3324 | |
3325 | if (adev->asic_reset_res) | |
3326 | goto fail; | |
43c4d576 | 3327 | |
5e67bba3 | 3328 | if (adev->mmhub.ras && adev->mmhub.ras->ras_block.hw_ops && |
3329 | adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count) | |
3330 | adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count(adev); | |
c6a6e2db AG |
3331 | } else { |
3332 | ||
3333 | task_barrier_full(&hive->tb); | |
3334 | adev->asic_reset_res = amdgpu_asic_reset(adev); | |
3335 | } | |
ce316fa5 | 3336 | |
c6a6e2db | 3337 | fail: |
d4535e2c | 3338 | if (adev->asic_reset_res) |
fed184e9 | 3339 | DRM_WARN("ASIC reset failed with error, %d for drm dev, %s", |
4a580877 | 3340 | adev->asic_reset_res, adev_to_drm(adev)->unique); |
d95e8e97 | 3341 | amdgpu_put_xgmi_hive(hive); |
d4535e2c AG |
3342 | } |
3343 | ||
71f98027 AD |
3344 | static int amdgpu_device_get_job_timeout_settings(struct amdgpu_device *adev) |
3345 | { | |
3346 | char *input = amdgpu_lockup_timeout; | |
3347 | char *timeout_setting = NULL; | |
3348 | int index = 0; | |
3349 | long timeout; | |
3350 | int ret = 0; | |
3351 | ||
3352 | /* | |
67387dfe AD |
3353 | * By default timeout for non compute jobs is 10000 |
3354 | * and 60000 for compute jobs. | |
71f98027 | 3355 | * In SR-IOV or passthrough mode, timeout for compute |
b7b2a316 | 3356 | * jobs are 60000 by default. |
71f98027 AD |
3357 | */ |
3358 | adev->gfx_timeout = msecs_to_jiffies(10000); | |
3359 | adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout; | |
9882e278 ED |
3360 | if (amdgpu_sriov_vf(adev)) |
3361 | adev->compute_timeout = amdgpu_sriov_is_pp_one_vf(adev) ? | |
3362 | msecs_to_jiffies(60000) : msecs_to_jiffies(10000); | |
71f98027 | 3363 | else |
67387dfe | 3364 | adev->compute_timeout = msecs_to_jiffies(60000); |
71f98027 | 3365 | |
f440ff44 | 3366 | if (strnlen(input, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) { |
71f98027 | 3367 | while ((timeout_setting = strsep(&input, ",")) && |
f440ff44 | 3368 | strnlen(timeout_setting, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) { |
71f98027 AD |
3369 | ret = kstrtol(timeout_setting, 0, &timeout); |
3370 | if (ret) | |
3371 | return ret; | |
3372 | ||
3373 | if (timeout == 0) { | |
3374 | index++; | |
3375 | continue; | |
3376 | } else if (timeout < 0) { | |
3377 | timeout = MAX_SCHEDULE_TIMEOUT; | |
127aedf9 CK |
3378 | dev_warn(adev->dev, "lockup timeout disabled"); |
3379 | add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK); | |
71f98027 AD |
3380 | } else { |
3381 | timeout = msecs_to_jiffies(timeout); | |
3382 | } | |
3383 | ||
3384 | switch (index++) { | |
3385 | case 0: | |
3386 | adev->gfx_timeout = timeout; | |
3387 | break; | |
3388 | case 1: | |
3389 | adev->compute_timeout = timeout; | |
3390 | break; | |
3391 | case 2: | |
3392 | adev->sdma_timeout = timeout; | |
3393 | break; | |
3394 | case 3: | |
3395 | adev->video_timeout = timeout; | |
3396 | break; | |
3397 | default: | |
3398 | break; | |
3399 | } | |
3400 | } | |
3401 | /* | |
3402 | * There is only one value specified and | |
3403 | * it should apply to all non-compute jobs. | |
3404 | */ | |
bcccee89 | 3405 | if (index == 1) { |
71f98027 | 3406 | adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout; |
bcccee89 ED |
3407 | if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev)) |
3408 | adev->compute_timeout = adev->gfx_timeout; | |
3409 | } | |
71f98027 AD |
3410 | } |
3411 | ||
3412 | return ret; | |
3413 | } | |
d4535e2c | 3414 | |
4a74c38c PY |
3415 | /** |
3416 | * amdgpu_device_check_iommu_direct_map - check if RAM direct mapped to GPU | |
3417 | * | |
3418 | * @adev: amdgpu_device pointer | |
3419 | * | |
3420 | * RAM direct mapped to GPU if IOMMU is not enabled or is pass through mode | |
3421 | */ | |
3422 | static void amdgpu_device_check_iommu_direct_map(struct amdgpu_device *adev) | |
3423 | { | |
3424 | struct iommu_domain *domain; | |
3425 | ||
3426 | domain = iommu_get_domain_for_dev(adev->dev); | |
3427 | if (!domain || domain->type == IOMMU_DOMAIN_IDENTITY) | |
3428 | adev->ram_is_direct_mapped = true; | |
3429 | } | |
3430 | ||
77f3a5cd ND |
3431 | static const struct attribute *amdgpu_dev_attributes[] = { |
3432 | &dev_attr_product_name.attr, | |
3433 | &dev_attr_product_number.attr, | |
3434 | &dev_attr_serial_number.attr, | |
3435 | &dev_attr_pcie_replay_count.attr, | |
3436 | NULL | |
3437 | }; | |
3438 | ||
d38ceaf9 AD |
3439 | /** |
3440 | * amdgpu_device_init - initialize the driver | |
3441 | * | |
3442 | * @adev: amdgpu_device pointer | |
d38ceaf9 AD |
3443 | * @flags: driver flags |
3444 | * | |
3445 | * Initializes the driver info and hw (all asics). | |
3446 | * Returns 0 for success or an error on failure. | |
3447 | * Called at driver startup. | |
3448 | */ | |
3449 | int amdgpu_device_init(struct amdgpu_device *adev, | |
d38ceaf9 AD |
3450 | uint32_t flags) |
3451 | { | |
8aba21b7 LT |
3452 | struct drm_device *ddev = adev_to_drm(adev); |
3453 | struct pci_dev *pdev = adev->pdev; | |
d38ceaf9 | 3454 | int r, i; |
b98c6299 | 3455 | bool px = false; |
95844d20 | 3456 | u32 max_MBps; |
d38ceaf9 AD |
3457 | |
3458 | adev->shutdown = false; | |
d38ceaf9 | 3459 | adev->flags = flags; |
4e66d7d2 YZ |
3460 | |
3461 | if (amdgpu_force_asic_type >= 0 && amdgpu_force_asic_type < CHIP_LAST) | |
3462 | adev->asic_type = amdgpu_force_asic_type; | |
3463 | else | |
3464 | adev->asic_type = flags & AMD_ASIC_MASK; | |
3465 | ||
d38ceaf9 | 3466 | adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT; |
593aa2d2 | 3467 | if (amdgpu_emu_mode == 1) |
8bdab6bb | 3468 | adev->usec_timeout *= 10; |
770d13b1 | 3469 | adev->gmc.gart_size = 512 * 1024 * 1024; |
d38ceaf9 AD |
3470 | adev->accel_working = false; |
3471 | adev->num_rings = 0; | |
3472 | adev->mman.buffer_funcs = NULL; | |
3473 | adev->mman.buffer_funcs_ring = NULL; | |
3474 | adev->vm_manager.vm_pte_funcs = NULL; | |
0c88b430 | 3475 | adev->vm_manager.vm_pte_num_scheds = 0; |
132f34e4 | 3476 | adev->gmc.gmc_funcs = NULL; |
7bd939d0 | 3477 | adev->harvest_ip_mask = 0x0; |
f54d1867 | 3478 | adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS); |
b8866c26 | 3479 | bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES); |
d38ceaf9 AD |
3480 | |
3481 | adev->smc_rreg = &amdgpu_invalid_rreg; | |
3482 | adev->smc_wreg = &amdgpu_invalid_wreg; | |
3483 | adev->pcie_rreg = &amdgpu_invalid_rreg; | |
3484 | adev->pcie_wreg = &amdgpu_invalid_wreg; | |
36b9a952 HR |
3485 | adev->pciep_rreg = &amdgpu_invalid_rreg; |
3486 | adev->pciep_wreg = &amdgpu_invalid_wreg; | |
4fa1c6a6 TZ |
3487 | adev->pcie_rreg64 = &amdgpu_invalid_rreg64; |
3488 | adev->pcie_wreg64 = &amdgpu_invalid_wreg64; | |
d38ceaf9 AD |
3489 | adev->uvd_ctx_rreg = &amdgpu_invalid_rreg; |
3490 | adev->uvd_ctx_wreg = &amdgpu_invalid_wreg; | |
3491 | adev->didt_rreg = &amdgpu_invalid_rreg; | |
3492 | adev->didt_wreg = &amdgpu_invalid_wreg; | |
ccdbb20a RZ |
3493 | adev->gc_cac_rreg = &amdgpu_invalid_rreg; |
3494 | adev->gc_cac_wreg = &amdgpu_invalid_wreg; | |
d38ceaf9 AD |
3495 | adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg; |
3496 | adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg; | |
3497 | ||
3e39ab90 AD |
3498 | DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n", |
3499 | amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device, | |
3500 | pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision); | |
d38ceaf9 AD |
3501 | |
3502 | /* mutex initialization are all done here so we | |
3503 | * can recall function without having locking issues */ | |
0e5ca0d1 | 3504 | mutex_init(&adev->firmware.mutex); |
d38ceaf9 AD |
3505 | mutex_init(&adev->pm.mutex); |
3506 | mutex_init(&adev->gfx.gpu_clock_mutex); | |
3507 | mutex_init(&adev->srbm_mutex); | |
b8866c26 | 3508 | mutex_init(&adev->gfx.pipe_reserve_mutex); |
d23ee13f | 3509 | mutex_init(&adev->gfx.gfx_off_mutex); |
d38ceaf9 | 3510 | mutex_init(&adev->grbm_idx_mutex); |
d38ceaf9 | 3511 | mutex_init(&adev->mn_lock); |
e23b74aa | 3512 | mutex_init(&adev->virt.vf_errors.lock); |
d38ceaf9 | 3513 | hash_init(adev->mn_hash); |
53b3f8f4 | 3514 | atomic_set(&adev->in_gpu_reset, 0); |
6049db43 | 3515 | init_rwsem(&adev->reset_sem); |
32eaeae0 | 3516 | mutex_init(&adev->psp.mutex); |
bd052211 | 3517 | mutex_init(&adev->notifier_lock); |
8cda7a4f | 3518 | mutex_init(&adev->pm.stable_pstate_ctx_lock); |
d38ceaf9 | 3519 | |
ab3b9de6 | 3520 | amdgpu_device_init_apu_flags(adev); |
9f6a7857 | 3521 | |
912dfc84 EQ |
3522 | r = amdgpu_device_check_arguments(adev); |
3523 | if (r) | |
3524 | return r; | |
d38ceaf9 | 3525 | |
d38ceaf9 AD |
3526 | spin_lock_init(&adev->mmio_idx_lock); |
3527 | spin_lock_init(&adev->smc_idx_lock); | |
3528 | spin_lock_init(&adev->pcie_idx_lock); | |
3529 | spin_lock_init(&adev->uvd_ctx_idx_lock); | |
3530 | spin_lock_init(&adev->didt_idx_lock); | |
ccdbb20a | 3531 | spin_lock_init(&adev->gc_cac_idx_lock); |
16abb5d2 | 3532 | spin_lock_init(&adev->se_cac_idx_lock); |
d38ceaf9 | 3533 | spin_lock_init(&adev->audio_endpt_idx_lock); |
95844d20 | 3534 | spin_lock_init(&adev->mm_stats.lock); |
d38ceaf9 | 3535 | |
0c4e7fa5 CZ |
3536 | INIT_LIST_HEAD(&adev->shadow_list); |
3537 | mutex_init(&adev->shadow_list_lock); | |
3538 | ||
655ce9cb | 3539 | INIT_LIST_HEAD(&adev->reset_list); |
3540 | ||
6492e1b0 | 3541 | INIT_LIST_HEAD(&adev->ras_list); |
3542 | ||
beff74bc AD |
3543 | INIT_DELAYED_WORK(&adev->delayed_init_work, |
3544 | amdgpu_device_delayed_init_work_handler); | |
1e317b99 RZ |
3545 | INIT_DELAYED_WORK(&adev->gfx.gfx_off_delay_work, |
3546 | amdgpu_device_delay_enable_gfx_off); | |
2dc80b00 | 3547 | |
d4535e2c AG |
3548 | INIT_WORK(&adev->xgmi_reset_work, amdgpu_device_xgmi_reset_func); |
3549 | ||
d23ee13f | 3550 | adev->gfx.gfx_off_req_count = 1; |
b6e79d9a | 3551 | adev->pm.ac_power = power_supply_is_system_supplied() > 0; |
b1ddf548 | 3552 | |
b265bdbd EQ |
3553 | atomic_set(&adev->throttling_logging_enabled, 1); |
3554 | /* | |
3555 | * If throttling continues, logging will be performed every minute | |
3556 | * to avoid log flooding. "-1" is subtracted since the thermal | |
3557 | * throttling interrupt comes every second. Thus, the total logging | |
3558 | * interval is 59 seconds(retelimited printk interval) + 1(waiting | |
3559 | * for throttling interrupt) = 60 seconds. | |
3560 | */ | |
3561 | ratelimit_state_init(&adev->throttling_logging_rs, (60 - 1) * HZ, 1); | |
3562 | ratelimit_set_flags(&adev->throttling_logging_rs, RATELIMIT_MSG_ON_RELEASE); | |
3563 | ||
0fa49558 AX |
3564 | /* Registers mapping */ |
3565 | /* TODO: block userspace mapping of io register */ | |
da69c161 KW |
3566 | if (adev->asic_type >= CHIP_BONAIRE) { |
3567 | adev->rmmio_base = pci_resource_start(adev->pdev, 5); | |
3568 | adev->rmmio_size = pci_resource_len(adev->pdev, 5); | |
3569 | } else { | |
3570 | adev->rmmio_base = pci_resource_start(adev->pdev, 2); | |
3571 | adev->rmmio_size = pci_resource_len(adev->pdev, 2); | |
3572 | } | |
d38ceaf9 | 3573 | |
6c08e0ef EQ |
3574 | for (i = 0; i < AMD_IP_BLOCK_TYPE_NUM; i++) |
3575 | atomic_set(&adev->pm.pwr_state[i], POWER_STATE_UNKNOWN); | |
3576 | ||
d38ceaf9 AD |
3577 | adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size); |
3578 | if (adev->rmmio == NULL) { | |
3579 | return -ENOMEM; | |
3580 | } | |
3581 | DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base); | |
3582 | DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size); | |
3583 | ||
5494d864 AD |
3584 | amdgpu_device_get_pcie_info(adev); |
3585 | ||
b239c017 JX |
3586 | if (amdgpu_mcbp) |
3587 | DRM_INFO("MCBP is enabled\n"); | |
3588 | ||
5f84cc63 JX |
3589 | if (amdgpu_mes && adev->asic_type >= CHIP_NAVI10) |
3590 | adev->enable_mes = true; | |
3591 | ||
3aa0115d ML |
3592 | /* detect hw virtualization here */ |
3593 | amdgpu_detect_virtualization(adev); | |
3594 | ||
dffa11b4 ML |
3595 | r = amdgpu_device_get_job_timeout_settings(adev); |
3596 | if (r) { | |
3597 | dev_err(adev->dev, "invalid lockup_timeout parameter syntax\n"); | |
4ef87d8f | 3598 | return r; |
a190d1c7 XY |
3599 | } |
3600 | ||
d38ceaf9 | 3601 | /* early init functions */ |
06ec9070 | 3602 | r = amdgpu_device_ip_early_init(adev); |
d38ceaf9 | 3603 | if (r) |
4ef87d8f | 3604 | return r; |
d38ceaf9 | 3605 | |
4a0165f0 VS |
3606 | /* Need to get xgmi info early to decide the reset behavior*/ |
3607 | if (adev->gmc.xgmi.supported) { | |
3608 | r = adev->gfxhub.funcs->get_xgmi_info(adev); | |
3609 | if (r) | |
3610 | return r; | |
3611 | } | |
3612 | ||
8e6d0b69 | 3613 | /* enable PCIE atomic ops */ |
3614 | if (amdgpu_sriov_vf(adev)) | |
3615 | adev->have_atomics_support = ((struct amd_sriov_msg_pf2vf_info *) | |
3616 | adev->virt.fw_reserve.p_pf2vf)->pcie_atomic_ops_enabled_flags == | |
3617 | (PCI_EXP_DEVCAP2_ATOMIC_COMP32 | PCI_EXP_DEVCAP2_ATOMIC_COMP64); | |
3618 | else | |
3619 | adev->have_atomics_support = | |
3620 | !pci_enable_atomic_ops_to_root(adev->pdev, | |
3621 | PCI_EXP_DEVCAP2_ATOMIC_COMP32 | | |
3622 | PCI_EXP_DEVCAP2_ATOMIC_COMP64); | |
3623 | if (!adev->have_atomics_support) | |
3624 | dev_info(adev->dev, "PCIE atomic ops is not supported\n"); | |
3625 | ||
6585661d OZ |
3626 | /* doorbell bar mapping and doorbell index init*/ |
3627 | amdgpu_device_doorbell_init(adev); | |
3628 | ||
9475a943 SL |
3629 | if (amdgpu_emu_mode == 1) { |
3630 | /* post the asic on emulation mode */ | |
3631 | emu_soc_asic_init(adev); | |
bfca0289 | 3632 | goto fence_driver_init; |
9475a943 | 3633 | } |
bfca0289 | 3634 | |
04442bf7 LL |
3635 | amdgpu_reset_init(adev); |
3636 | ||
4e99a44e ML |
3637 | /* detect if we are with an SRIOV vbios */ |
3638 | amdgpu_device_detect_sriov_bios(adev); | |
048765ad | 3639 | |
95e8e59e AD |
3640 | /* check if we need to reset the asic |
3641 | * E.g., driver was not cleanly unloaded previously, etc. | |
3642 | */ | |
f14899fd | 3643 | if (!amdgpu_sriov_vf(adev) && amdgpu_asic_need_reset_on_init(adev)) { |
e3c1b071 | 3644 | if (adev->gmc.xgmi.num_physical_nodes) { |
3645 | dev_info(adev->dev, "Pending hive reset.\n"); | |
3646 | adev->gmc.xgmi.pending_reset = true; | |
3647 | /* Only need to init necessary block for SMU to handle the reset */ | |
3648 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
3649 | if (!adev->ip_blocks[i].status.valid) | |
3650 | continue; | |
3651 | if (!(adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC || | |
3652 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || | |
3653 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH || | |
3654 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC)) { | |
751f43e7 | 3655 | DRM_DEBUG("IP %s disabled for hw_init.\n", |
e3c1b071 | 3656 | adev->ip_blocks[i].version->funcs->name); |
3657 | adev->ip_blocks[i].status.hw = true; | |
3658 | } | |
3659 | } | |
3660 | } else { | |
3661 | r = amdgpu_asic_reset(adev); | |
3662 | if (r) { | |
3663 | dev_err(adev->dev, "asic reset on init failed\n"); | |
3664 | goto failed; | |
3665 | } | |
95e8e59e AD |
3666 | } |
3667 | } | |
3668 | ||
8f66090b | 3669 | pci_enable_pcie_error_reporting(adev->pdev); |
c9a6b82f | 3670 | |
d38ceaf9 | 3671 | /* Post card if necessary */ |
39c640c0 | 3672 | if (amdgpu_device_need_post(adev)) { |
d38ceaf9 | 3673 | if (!adev->bios) { |
bec86378 | 3674 | dev_err(adev->dev, "no vBIOS found\n"); |
83ba126a AD |
3675 | r = -EINVAL; |
3676 | goto failed; | |
d38ceaf9 | 3677 | } |
bec86378 | 3678 | DRM_INFO("GPU posting now...\n"); |
4d2997ab | 3679 | r = amdgpu_device_asic_init(adev); |
4e99a44e ML |
3680 | if (r) { |
3681 | dev_err(adev->dev, "gpu post error!\n"); | |
3682 | goto failed; | |
3683 | } | |
d38ceaf9 AD |
3684 | } |
3685 | ||
88b64e95 AD |
3686 | if (adev->is_atom_fw) { |
3687 | /* Initialize clocks */ | |
3688 | r = amdgpu_atomfirmware_get_clock_info(adev); | |
3689 | if (r) { | |
3690 | dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n"); | |
e23b74aa | 3691 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0); |
88b64e95 AD |
3692 | goto failed; |
3693 | } | |
3694 | } else { | |
a5bde2f9 AD |
3695 | /* Initialize clocks */ |
3696 | r = amdgpu_atombios_get_clock_info(adev); | |
3697 | if (r) { | |
3698 | dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n"); | |
e23b74aa | 3699 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0); |
89041940 | 3700 | goto failed; |
a5bde2f9 AD |
3701 | } |
3702 | /* init i2c buses */ | |
4562236b HW |
3703 | if (!amdgpu_device_has_dc_support(adev)) |
3704 | amdgpu_atombios_i2c_init(adev); | |
2c1a2784 | 3705 | } |
d38ceaf9 | 3706 | |
bfca0289 | 3707 | fence_driver_init: |
d38ceaf9 | 3708 | /* Fence driver */ |
067f44c8 | 3709 | r = amdgpu_fence_driver_sw_init(adev); |
2c1a2784 | 3710 | if (r) { |
067f44c8 | 3711 | dev_err(adev->dev, "amdgpu_fence_driver_sw_init failed\n"); |
e23b74aa | 3712 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0); |
83ba126a | 3713 | goto failed; |
2c1a2784 | 3714 | } |
d38ceaf9 AD |
3715 | |
3716 | /* init the mode config */ | |
4a580877 | 3717 | drm_mode_config_init(adev_to_drm(adev)); |
d38ceaf9 | 3718 | |
06ec9070 | 3719 | r = amdgpu_device_ip_init(adev); |
d38ceaf9 | 3720 | if (r) { |
8840a387 | 3721 | /* failed in exclusive mode due to timeout */ |
3722 | if (amdgpu_sriov_vf(adev) && | |
3723 | !amdgpu_sriov_runtime(adev) && | |
3724 | amdgpu_virt_mmio_blocked(adev) && | |
3725 | !amdgpu_virt_wait_reset(adev)) { | |
3726 | dev_err(adev->dev, "VF exclusive mode timeout\n"); | |
1daee8b4 PD |
3727 | /* Don't send request since VF is inactive. */ |
3728 | adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME; | |
3729 | adev->virt.ops = NULL; | |
8840a387 | 3730 | r = -EAGAIN; |
970fd197 | 3731 | goto release_ras_con; |
8840a387 | 3732 | } |
06ec9070 | 3733 | dev_err(adev->dev, "amdgpu_device_ip_init failed\n"); |
e23b74aa | 3734 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0); |
970fd197 | 3735 | goto release_ras_con; |
d38ceaf9 AD |
3736 | } |
3737 | ||
8d35a259 LG |
3738 | amdgpu_fence_driver_hw_init(adev); |
3739 | ||
d69b8971 YZ |
3740 | dev_info(adev->dev, |
3741 | "SE %d, SH per SE %d, CU per SH %d, active_cu_number %d\n", | |
d7f72fe4 YZ |
3742 | adev->gfx.config.max_shader_engines, |
3743 | adev->gfx.config.max_sh_per_se, | |
3744 | adev->gfx.config.max_cu_per_sh, | |
3745 | adev->gfx.cu_info.number); | |
3746 | ||
d38ceaf9 AD |
3747 | adev->accel_working = true; |
3748 | ||
e59c0205 AX |
3749 | amdgpu_vm_check_compute_bug(adev); |
3750 | ||
95844d20 MO |
3751 | /* Initialize the buffer migration limit. */ |
3752 | if (amdgpu_moverate >= 0) | |
3753 | max_MBps = amdgpu_moverate; | |
3754 | else | |
3755 | max_MBps = 8; /* Allow 8 MB/s. */ | |
3756 | /* Get a log2 for easy divisions. */ | |
3757 | adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps)); | |
3758 | ||
d2f52ac8 | 3759 | r = amdgpu_pm_sysfs_init(adev); |
7c868b59 YT |
3760 | if (r) { |
3761 | adev->pm_sysfs_en = false; | |
d2f52ac8 | 3762 | DRM_ERROR("registering pm debugfs failed (%d).\n", r); |
7c868b59 YT |
3763 | } else |
3764 | adev->pm_sysfs_en = true; | |
d2f52ac8 | 3765 | |
5bb23532 | 3766 | r = amdgpu_ucode_sysfs_init(adev); |
7c868b59 YT |
3767 | if (r) { |
3768 | adev->ucode_sysfs_en = false; | |
5bb23532 | 3769 | DRM_ERROR("Creating firmware sysfs failed (%d).\n", r); |
7c868b59 YT |
3770 | } else |
3771 | adev->ucode_sysfs_en = true; | |
5bb23532 | 3772 | |
d38ceaf9 AD |
3773 | if ((amdgpu_testing & 1)) { |
3774 | if (adev->accel_working) | |
3775 | amdgpu_test_moves(adev); | |
3776 | else | |
3777 | DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n"); | |
3778 | } | |
d38ceaf9 AD |
3779 | if (amdgpu_benchmarking) { |
3780 | if (adev->accel_working) | |
3781 | amdgpu_benchmark(adev, amdgpu_benchmarking); | |
3782 | else | |
3783 | DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n"); | |
3784 | } | |
3785 | ||
b0adca4d EQ |
3786 | /* |
3787 | * Register gpu instance before amdgpu_device_enable_mgpu_fan_boost. | |
3788 | * Otherwise the mgpu fan boost feature will be skipped due to the | |
3789 | * gpu instance is counted less. | |
3790 | */ | |
3791 | amdgpu_register_gpu_instance(adev); | |
3792 | ||
d38ceaf9 AD |
3793 | /* enable clockgating, etc. after ib tests, etc. since some blocks require |
3794 | * explicit gating rather than handling it automatically. | |
3795 | */ | |
e3c1b071 | 3796 | if (!adev->gmc.xgmi.pending_reset) { |
3797 | r = amdgpu_device_ip_late_init(adev); | |
3798 | if (r) { | |
3799 | dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n"); | |
3800 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r); | |
970fd197 | 3801 | goto release_ras_con; |
e3c1b071 | 3802 | } |
3803 | /* must succeed. */ | |
3804 | amdgpu_ras_resume(adev); | |
3805 | queue_delayed_work(system_wq, &adev->delayed_init_work, | |
3806 | msecs_to_jiffies(AMDGPU_RESUME_MS)); | |
2c1a2784 | 3807 | } |
d38ceaf9 | 3808 | |
2c738637 ML |
3809 | if (amdgpu_sriov_vf(adev)) |
3810 | flush_delayed_work(&adev->delayed_init_work); | |
3811 | ||
77f3a5cd | 3812 | r = sysfs_create_files(&adev->dev->kobj, amdgpu_dev_attributes); |
5aea5327 | 3813 | if (r) |
77f3a5cd | 3814 | dev_err(adev->dev, "Could not create amdgpu device attr\n"); |
bd607166 | 3815 | |
d155bef0 AB |
3816 | if (IS_ENABLED(CONFIG_PERF_EVENTS)) |
3817 | r = amdgpu_pmu_init(adev); | |
9c7c85f7 JK |
3818 | if (r) |
3819 | dev_err(adev->dev, "amdgpu_pmu_init failed\n"); | |
3820 | ||
c1dd4aa6 AG |
3821 | /* Have stored pci confspace at hand for restore in sudden PCI error */ |
3822 | if (amdgpu_device_cache_pci_state(adev->pdev)) | |
3823 | pci_restore_state(pdev); | |
3824 | ||
8c3dd61c KHF |
3825 | /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */ |
3826 | /* this will fail for cards that aren't VGA class devices, just | |
3827 | * ignore it */ | |
3828 | if ((adev->pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA) | |
bf44e8ce | 3829 | vga_client_register(adev->pdev, amdgpu_device_vga_set_decode); |
8c3dd61c KHF |
3830 | |
3831 | if (amdgpu_device_supports_px(ddev)) { | |
3832 | px = true; | |
3833 | vga_switcheroo_register_client(adev->pdev, | |
3834 | &amdgpu_switcheroo_ops, px); | |
3835 | vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain); | |
3836 | } | |
3837 | ||
e3c1b071 | 3838 | if (adev->gmc.xgmi.pending_reset) |
3839 | queue_delayed_work(system_wq, &mgpu_info.delayed_reset_work, | |
3840 | msecs_to_jiffies(AMDGPU_RESUME_MS)); | |
3841 | ||
4a74c38c PY |
3842 | amdgpu_device_check_iommu_direct_map(adev); |
3843 | ||
d38ceaf9 | 3844 | return 0; |
83ba126a | 3845 | |
970fd197 SY |
3846 | release_ras_con: |
3847 | amdgpu_release_ras_context(adev); | |
3848 | ||
83ba126a | 3849 | failed: |
89041940 | 3850 | amdgpu_vf_error_trans_all(adev); |
8840a387 | 3851 | |
83ba126a | 3852 | return r; |
d38ceaf9 AD |
3853 | } |
3854 | ||
07775fc1 AG |
3855 | static void amdgpu_device_unmap_mmio(struct amdgpu_device *adev) |
3856 | { | |
62d5f9f7 | 3857 | |
07775fc1 AG |
3858 | /* Clear all CPU mappings pointing to this device */ |
3859 | unmap_mapping_range(adev->ddev.anon_inode->i_mapping, 0, 0, 1); | |
3860 | ||
3861 | /* Unmap all mapped bars - Doorbell, registers and VRAM */ | |
3862 | amdgpu_device_doorbell_fini(adev); | |
3863 | ||
3864 | iounmap(adev->rmmio); | |
3865 | adev->rmmio = NULL; | |
3866 | if (adev->mman.aper_base_kaddr) | |
3867 | iounmap(adev->mman.aper_base_kaddr); | |
3868 | adev->mman.aper_base_kaddr = NULL; | |
3869 | ||
3870 | /* Memory manager related */ | |
3871 | if (!adev->gmc.xgmi.connected_to_cpu) { | |
3872 | arch_phys_wc_del(adev->gmc.vram_mtrr); | |
3873 | arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size); | |
3874 | } | |
3875 | } | |
3876 | ||
d38ceaf9 | 3877 | /** |
bbe04dec | 3878 | * amdgpu_device_fini_hw - tear down the driver |
d38ceaf9 AD |
3879 | * |
3880 | * @adev: amdgpu_device pointer | |
3881 | * | |
3882 | * Tear down the driver info (all asics). | |
3883 | * Called at driver shutdown. | |
3884 | */ | |
72c8c97b | 3885 | void amdgpu_device_fini_hw(struct amdgpu_device *adev) |
d38ceaf9 | 3886 | { |
aac89168 | 3887 | dev_info(adev->dev, "amdgpu: finishing device.\n"); |
9f875167 | 3888 | flush_delayed_work(&adev->delayed_init_work); |
691191a2 YW |
3889 | if (adev->mman.initialized) { |
3890 | flush_delayed_work(&adev->mman.bdev.wq); | |
e78b3197 | 3891 | ttm_bo_lock_delayed_workqueue(&adev->mman.bdev); |
691191a2 | 3892 | } |
d0d13fe8 | 3893 | adev->shutdown = true; |
9f875167 | 3894 | |
752c683d ML |
3895 | /* make sure IB test finished before entering exclusive mode |
3896 | * to avoid preemption on IB test | |
3897 | * */ | |
519b8b76 | 3898 | if (amdgpu_sriov_vf(adev)) { |
752c683d | 3899 | amdgpu_virt_request_full_gpu(adev, false); |
519b8b76 BZ |
3900 | amdgpu_virt_fini_data_exchange(adev); |
3901 | } | |
752c683d | 3902 | |
e5b03032 ML |
3903 | /* disable all interrupts */ |
3904 | amdgpu_irq_disable_all(adev); | |
ff97cba8 | 3905 | if (adev->mode_info.mode_config_initialized){ |
1053b9c9 | 3906 | if (!drm_drv_uses_atomic_modeset(adev_to_drm(adev))) |
4a580877 | 3907 | drm_helper_force_disable_all(adev_to_drm(adev)); |
ff97cba8 | 3908 | else |
4a580877 | 3909 | drm_atomic_helper_shutdown(adev_to_drm(adev)); |
ff97cba8 | 3910 | } |
8d35a259 | 3911 | amdgpu_fence_driver_hw_fini(adev); |
72c8c97b | 3912 | |
7c868b59 YT |
3913 | if (adev->pm_sysfs_en) |
3914 | amdgpu_pm_sysfs_fini(adev); | |
72c8c97b AG |
3915 | if (adev->ucode_sysfs_en) |
3916 | amdgpu_ucode_sysfs_fini(adev); | |
3917 | sysfs_remove_files(&adev->dev->kobj, amdgpu_dev_attributes); | |
3918 | ||
232d1d43 SY |
3919 | /* disable ras feature must before hw fini */ |
3920 | amdgpu_ras_pre_fini(adev); | |
3921 | ||
e9669fb7 | 3922 | amdgpu_device_ip_fini_early(adev); |
d10d0daa | 3923 | |
a3848df6 YW |
3924 | amdgpu_irq_fini_hw(adev); |
3925 | ||
b6fd6e0f SK |
3926 | if (adev->mman.initialized) |
3927 | ttm_device_clear_dma_mappings(&adev->mman.bdev); | |
894c6890 | 3928 | |
d10d0daa | 3929 | amdgpu_gart_dummy_page_fini(adev); |
07775fc1 | 3930 | |
87172e89 LS |
3931 | if (drm_dev_is_unplugged(adev_to_drm(adev))) |
3932 | amdgpu_device_unmap_mmio(adev); | |
3933 | ||
72c8c97b AG |
3934 | } |
3935 | ||
3936 | void amdgpu_device_fini_sw(struct amdgpu_device *adev) | |
3937 | { | |
62d5f9f7 LS |
3938 | int idx; |
3939 | ||
8d35a259 | 3940 | amdgpu_fence_driver_sw_fini(adev); |
a5c5d8d5 | 3941 | amdgpu_device_ip_fini(adev); |
75e1658e ND |
3942 | release_firmware(adev->firmware.gpu_info_fw); |
3943 | adev->firmware.gpu_info_fw = NULL; | |
d38ceaf9 | 3944 | adev->accel_working = false; |
04442bf7 LL |
3945 | |
3946 | amdgpu_reset_fini(adev); | |
3947 | ||
d38ceaf9 | 3948 | /* free i2c buses */ |
4562236b HW |
3949 | if (!amdgpu_device_has_dc_support(adev)) |
3950 | amdgpu_i2c_fini(adev); | |
bfca0289 SL |
3951 | |
3952 | if (amdgpu_emu_mode != 1) | |
3953 | amdgpu_atombios_fini(adev); | |
3954 | ||
d38ceaf9 AD |
3955 | kfree(adev->bios); |
3956 | adev->bios = NULL; | |
b98c6299 | 3957 | if (amdgpu_device_supports_px(adev_to_drm(adev))) { |
84c8b22e | 3958 | vga_switcheroo_unregister_client(adev->pdev); |
83ba126a | 3959 | vga_switcheroo_fini_domain_pm_ops(adev->dev); |
b98c6299 | 3960 | } |
38d6be81 | 3961 | if ((adev->pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA) |
b8779475 | 3962 | vga_client_unregister(adev->pdev); |
e9bc1bf7 | 3963 | |
62d5f9f7 LS |
3964 | if (drm_dev_enter(adev_to_drm(adev), &idx)) { |
3965 | ||
3966 | iounmap(adev->rmmio); | |
3967 | adev->rmmio = NULL; | |
3968 | amdgpu_device_doorbell_fini(adev); | |
3969 | drm_dev_exit(idx); | |
3970 | } | |
3971 | ||
d155bef0 AB |
3972 | if (IS_ENABLED(CONFIG_PERF_EVENTS)) |
3973 | amdgpu_pmu_fini(adev); | |
72de33f8 | 3974 | if (adev->mman.discovery_bin) |
a190d1c7 | 3975 | amdgpu_discovery_fini(adev); |
72c8c97b AG |
3976 | |
3977 | kfree(adev->pci_state); | |
3978 | ||
d38ceaf9 AD |
3979 | } |
3980 | ||
58144d28 ND |
3981 | /** |
3982 | * amdgpu_device_evict_resources - evict device resources | |
3983 | * @adev: amdgpu device object | |
3984 | * | |
3985 | * Evicts all ttm device resources(vram BOs, gart table) from the lru list | |
3986 | * of the vram memory type. Mainly used for evicting device resources | |
3987 | * at suspend time. | |
3988 | * | |
3989 | */ | |
3990 | static void amdgpu_device_evict_resources(struct amdgpu_device *adev) | |
3991 | { | |
e53d9665 ML |
3992 | /* No need to evict vram on APUs for suspend to ram or s2idle */ |
3993 | if ((adev->in_s3 || adev->in_s0ix) && (adev->flags & AMD_IS_APU)) | |
58144d28 ND |
3994 | return; |
3995 | ||
3996 | if (amdgpu_ttm_evict_resources(adev, TTM_PL_VRAM)) | |
3997 | DRM_WARN("evicting device resources failed\n"); | |
3998 | ||
3999 | } | |
d38ceaf9 AD |
4000 | |
4001 | /* | |
4002 | * Suspend & resume. | |
4003 | */ | |
4004 | /** | |
810ddc3a | 4005 | * amdgpu_device_suspend - initiate device suspend |
d38ceaf9 | 4006 | * |
87e3f136 | 4007 | * @dev: drm dev pointer |
87e3f136 | 4008 | * @fbcon : notify the fbdev of suspend |
d38ceaf9 AD |
4009 | * |
4010 | * Puts the hw in the suspend state (all asics). | |
4011 | * Returns 0 for success or an error on failure. | |
4012 | * Called at driver suspend. | |
4013 | */ | |
de185019 | 4014 | int amdgpu_device_suspend(struct drm_device *dev, bool fbcon) |
d38ceaf9 | 4015 | { |
a2e15b0e | 4016 | struct amdgpu_device *adev = drm_to_adev(dev); |
d38ceaf9 | 4017 | |
d38ceaf9 AD |
4018 | if (dev->switch_power_state == DRM_SWITCH_POWER_OFF) |
4019 | return 0; | |
4020 | ||
44779b43 | 4021 | adev->in_suspend = true; |
3fa8f89d S |
4022 | |
4023 | if (amdgpu_acpi_smart_shift_update(dev, AMDGPU_SS_DEV_D3)) | |
4024 | DRM_WARN("smart shift update failed\n"); | |
4025 | ||
d38ceaf9 AD |
4026 | drm_kms_helper_poll_disable(dev); |
4027 | ||
5f818173 | 4028 | if (fbcon) |
087451f3 | 4029 | drm_fb_helper_set_suspend_unlocked(adev_to_drm(adev)->fb_helper, true); |
5f818173 | 4030 | |
beff74bc | 4031 | cancel_delayed_work_sync(&adev->delayed_init_work); |
a5459475 | 4032 | |
5e6932fe | 4033 | amdgpu_ras_suspend(adev); |
4034 | ||
2196927b | 4035 | amdgpu_device_ip_suspend_phase1(adev); |
fe1053b7 | 4036 | |
5d3a2d95 AD |
4037 | if (!adev->in_s0ix) |
4038 | amdgpu_amdkfd_suspend(adev, adev->in_runpm); | |
94fa5660 | 4039 | |
58144d28 | 4040 | amdgpu_device_evict_resources(adev); |
d38ceaf9 | 4041 | |
8d35a259 | 4042 | amdgpu_fence_driver_hw_fini(adev); |
d38ceaf9 | 4043 | |
2196927b | 4044 | amdgpu_device_ip_suspend_phase2(adev); |
d38ceaf9 | 4045 | |
d38ceaf9 AD |
4046 | return 0; |
4047 | } | |
4048 | ||
4049 | /** | |
810ddc3a | 4050 | * amdgpu_device_resume - initiate device resume |
d38ceaf9 | 4051 | * |
87e3f136 | 4052 | * @dev: drm dev pointer |
87e3f136 | 4053 | * @fbcon : notify the fbdev of resume |
d38ceaf9 AD |
4054 | * |
4055 | * Bring the hw back to operating state (all asics). | |
4056 | * Returns 0 for success or an error on failure. | |
4057 | * Called at driver resume. | |
4058 | */ | |
de185019 | 4059 | int amdgpu_device_resume(struct drm_device *dev, bool fbcon) |
d38ceaf9 | 4060 | { |
1348969a | 4061 | struct amdgpu_device *adev = drm_to_adev(dev); |
03161a6e | 4062 | int r = 0; |
d38ceaf9 AD |
4063 | |
4064 | if (dev->switch_power_state == DRM_SWITCH_POWER_OFF) | |
4065 | return 0; | |
4066 | ||
62498733 | 4067 | if (adev->in_s0ix) |
bc143d8b | 4068 | amdgpu_dpm_gfx_state_change(adev, sGpuChangeState_D0Entry); |
628c36d7 | 4069 | |
d38ceaf9 | 4070 | /* post card */ |
39c640c0 | 4071 | if (amdgpu_device_need_post(adev)) { |
4d2997ab | 4072 | r = amdgpu_device_asic_init(adev); |
74b0b157 | 4073 | if (r) |
aac89168 | 4074 | dev_err(adev->dev, "amdgpu asic init failed\n"); |
74b0b157 | 4075 | } |
d38ceaf9 | 4076 | |
06ec9070 | 4077 | r = amdgpu_device_ip_resume(adev); |
e6707218 | 4078 | if (r) { |
aac89168 | 4079 | dev_err(adev->dev, "amdgpu_device_ip_resume failed (%d).\n", r); |
4d3b9ae5 | 4080 | return r; |
e6707218 | 4081 | } |
8d35a259 | 4082 | amdgpu_fence_driver_hw_init(adev); |
5ceb54c6 | 4083 | |
06ec9070 | 4084 | r = amdgpu_device_ip_late_init(adev); |
03161a6e | 4085 | if (r) |
4d3b9ae5 | 4086 | return r; |
d38ceaf9 | 4087 | |
beff74bc AD |
4088 | queue_delayed_work(system_wq, &adev->delayed_init_work, |
4089 | msecs_to_jiffies(AMDGPU_RESUME_MS)); | |
4090 | ||
5d3a2d95 AD |
4091 | if (!adev->in_s0ix) { |
4092 | r = amdgpu_amdkfd_resume(adev, adev->in_runpm); | |
4093 | if (r) | |
4094 | return r; | |
4095 | } | |
756e6880 | 4096 | |
96a5d8d4 | 4097 | /* Make sure IB tests flushed */ |
beff74bc | 4098 | flush_delayed_work(&adev->delayed_init_work); |
96a5d8d4 | 4099 | |
a2e15b0e | 4100 | if (fbcon) |
087451f3 | 4101 | drm_fb_helper_set_suspend_unlocked(adev_to_drm(adev)->fb_helper, false); |
d38ceaf9 AD |
4102 | |
4103 | drm_kms_helper_poll_enable(dev); | |
23a1a9e5 | 4104 | |
5e6932fe | 4105 | amdgpu_ras_resume(adev); |
4106 | ||
23a1a9e5 L |
4107 | /* |
4108 | * Most of the connector probing functions try to acquire runtime pm | |
4109 | * refs to ensure that the GPU is powered on when connector polling is | |
4110 | * performed. Since we're calling this from a runtime PM callback, | |
4111 | * trying to acquire rpm refs will cause us to deadlock. | |
4112 | * | |
4113 | * Since we're guaranteed to be holding the rpm lock, it's safe to | |
4114 | * temporarily disable the rpm helpers so this doesn't deadlock us. | |
4115 | */ | |
4116 | #ifdef CONFIG_PM | |
4117 | dev->dev->power.disable_depth++; | |
4118 | #endif | |
4562236b HW |
4119 | if (!amdgpu_device_has_dc_support(adev)) |
4120 | drm_helper_hpd_irq_event(dev); | |
4121 | else | |
4122 | drm_kms_helper_hotplug_event(dev); | |
23a1a9e5 L |
4123 | #ifdef CONFIG_PM |
4124 | dev->dev->power.disable_depth--; | |
4125 | #endif | |
44779b43 RZ |
4126 | adev->in_suspend = false; |
4127 | ||
3fa8f89d S |
4128 | if (amdgpu_acpi_smart_shift_update(dev, AMDGPU_SS_DEV_D0)) |
4129 | DRM_WARN("smart shift update failed\n"); | |
4130 | ||
4d3b9ae5 | 4131 | return 0; |
d38ceaf9 AD |
4132 | } |
4133 | ||
e3ecdffa AD |
4134 | /** |
4135 | * amdgpu_device_ip_check_soft_reset - did soft reset succeed | |
4136 | * | |
4137 | * @adev: amdgpu_device pointer | |
4138 | * | |
4139 | * The list of all the hardware IPs that make up the asic is walked and | |
4140 | * the check_soft_reset callbacks are run. check_soft_reset determines | |
4141 | * if the asic is still hung or not. | |
4142 | * Returns true if any of the IPs are still in a hung state, false if not. | |
4143 | */ | |
06ec9070 | 4144 | static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev) |
63fbf42f CZ |
4145 | { |
4146 | int i; | |
4147 | bool asic_hang = false; | |
4148 | ||
f993d628 ML |
4149 | if (amdgpu_sriov_vf(adev)) |
4150 | return true; | |
4151 | ||
8bc04c29 AD |
4152 | if (amdgpu_asic_need_full_reset(adev)) |
4153 | return true; | |
4154 | ||
63fbf42f | 4155 | for (i = 0; i < adev->num_ip_blocks; i++) { |
a1255107 | 4156 | if (!adev->ip_blocks[i].status.valid) |
63fbf42f | 4157 | continue; |
a1255107 AD |
4158 | if (adev->ip_blocks[i].version->funcs->check_soft_reset) |
4159 | adev->ip_blocks[i].status.hang = | |
4160 | adev->ip_blocks[i].version->funcs->check_soft_reset(adev); | |
4161 | if (adev->ip_blocks[i].status.hang) { | |
aac89168 | 4162 | dev_info(adev->dev, "IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name); |
63fbf42f CZ |
4163 | asic_hang = true; |
4164 | } | |
4165 | } | |
4166 | return asic_hang; | |
4167 | } | |
4168 | ||
e3ecdffa AD |
4169 | /** |
4170 | * amdgpu_device_ip_pre_soft_reset - prepare for soft reset | |
4171 | * | |
4172 | * @adev: amdgpu_device pointer | |
4173 | * | |
4174 | * The list of all the hardware IPs that make up the asic is walked and the | |
4175 | * pre_soft_reset callbacks are run if the block is hung. pre_soft_reset | |
4176 | * handles any IP specific hardware or software state changes that are | |
4177 | * necessary for a soft reset to succeed. | |
4178 | * Returns 0 on success, negative error code on failure. | |
4179 | */ | |
06ec9070 | 4180 | static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev) |
d31a501e CZ |
4181 | { |
4182 | int i, r = 0; | |
4183 | ||
4184 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 4185 | if (!adev->ip_blocks[i].status.valid) |
d31a501e | 4186 | continue; |
a1255107 AD |
4187 | if (adev->ip_blocks[i].status.hang && |
4188 | adev->ip_blocks[i].version->funcs->pre_soft_reset) { | |
4189 | r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev); | |
d31a501e CZ |
4190 | if (r) |
4191 | return r; | |
4192 | } | |
4193 | } | |
4194 | ||
4195 | return 0; | |
4196 | } | |
4197 | ||
e3ecdffa AD |
4198 | /** |
4199 | * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed | |
4200 | * | |
4201 | * @adev: amdgpu_device pointer | |
4202 | * | |
4203 | * Some hardware IPs cannot be soft reset. If they are hung, a full gpu | |
4204 | * reset is necessary to recover. | |
4205 | * Returns true if a full asic reset is required, false if not. | |
4206 | */ | |
06ec9070 | 4207 | static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev) |
35d782fe | 4208 | { |
da146d3b AD |
4209 | int i; |
4210 | ||
8bc04c29 AD |
4211 | if (amdgpu_asic_need_full_reset(adev)) |
4212 | return true; | |
4213 | ||
da146d3b | 4214 | for (i = 0; i < adev->num_ip_blocks; i++) { |
a1255107 | 4215 | if (!adev->ip_blocks[i].status.valid) |
da146d3b | 4216 | continue; |
a1255107 AD |
4217 | if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) || |
4218 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) || | |
4219 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) || | |
98512bb8 KW |
4220 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) || |
4221 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) { | |
a1255107 | 4222 | if (adev->ip_blocks[i].status.hang) { |
aac89168 | 4223 | dev_info(adev->dev, "Some block need full reset!\n"); |
da146d3b AD |
4224 | return true; |
4225 | } | |
4226 | } | |
35d782fe CZ |
4227 | } |
4228 | return false; | |
4229 | } | |
4230 | ||
e3ecdffa AD |
4231 | /** |
4232 | * amdgpu_device_ip_soft_reset - do a soft reset | |
4233 | * | |
4234 | * @adev: amdgpu_device pointer | |
4235 | * | |
4236 | * The list of all the hardware IPs that make up the asic is walked and the | |
4237 | * soft_reset callbacks are run if the block is hung. soft_reset handles any | |
4238 | * IP specific hardware or software state changes that are necessary to soft | |
4239 | * reset the IP. | |
4240 | * Returns 0 on success, negative error code on failure. | |
4241 | */ | |
06ec9070 | 4242 | static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev) |
35d782fe CZ |
4243 | { |
4244 | int i, r = 0; | |
4245 | ||
4246 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 4247 | if (!adev->ip_blocks[i].status.valid) |
35d782fe | 4248 | continue; |
a1255107 AD |
4249 | if (adev->ip_blocks[i].status.hang && |
4250 | adev->ip_blocks[i].version->funcs->soft_reset) { | |
4251 | r = adev->ip_blocks[i].version->funcs->soft_reset(adev); | |
35d782fe CZ |
4252 | if (r) |
4253 | return r; | |
4254 | } | |
4255 | } | |
4256 | ||
4257 | return 0; | |
4258 | } | |
4259 | ||
e3ecdffa AD |
4260 | /** |
4261 | * amdgpu_device_ip_post_soft_reset - clean up from soft reset | |
4262 | * | |
4263 | * @adev: amdgpu_device pointer | |
4264 | * | |
4265 | * The list of all the hardware IPs that make up the asic is walked and the | |
4266 | * post_soft_reset callbacks are run if the asic was hung. post_soft_reset | |
4267 | * handles any IP specific hardware or software state changes that are | |
4268 | * necessary after the IP has been soft reset. | |
4269 | * Returns 0 on success, negative error code on failure. | |
4270 | */ | |
06ec9070 | 4271 | static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev) |
35d782fe CZ |
4272 | { |
4273 | int i, r = 0; | |
4274 | ||
4275 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 4276 | if (!adev->ip_blocks[i].status.valid) |
35d782fe | 4277 | continue; |
a1255107 AD |
4278 | if (adev->ip_blocks[i].status.hang && |
4279 | adev->ip_blocks[i].version->funcs->post_soft_reset) | |
4280 | r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev); | |
35d782fe CZ |
4281 | if (r) |
4282 | return r; | |
4283 | } | |
4284 | ||
4285 | return 0; | |
4286 | } | |
4287 | ||
e3ecdffa | 4288 | /** |
c33adbc7 | 4289 | * amdgpu_device_recover_vram - Recover some VRAM contents |
e3ecdffa AD |
4290 | * |
4291 | * @adev: amdgpu_device pointer | |
4292 | * | |
4293 | * Restores the contents of VRAM buffers from the shadows in GTT. Used to | |
4294 | * restore things like GPUVM page tables after a GPU reset where | |
4295 | * the contents of VRAM might be lost. | |
403009bf CK |
4296 | * |
4297 | * Returns: | |
4298 | * 0 on success, negative error code on failure. | |
e3ecdffa | 4299 | */ |
c33adbc7 | 4300 | static int amdgpu_device_recover_vram(struct amdgpu_device *adev) |
c41d1cf6 | 4301 | { |
c41d1cf6 | 4302 | struct dma_fence *fence = NULL, *next = NULL; |
403009bf | 4303 | struct amdgpu_bo *shadow; |
e18aaea7 | 4304 | struct amdgpu_bo_vm *vmbo; |
403009bf | 4305 | long r = 1, tmo; |
c41d1cf6 ML |
4306 | |
4307 | if (amdgpu_sriov_runtime(adev)) | |
b045d3af | 4308 | tmo = msecs_to_jiffies(8000); |
c41d1cf6 ML |
4309 | else |
4310 | tmo = msecs_to_jiffies(100); | |
4311 | ||
aac89168 | 4312 | dev_info(adev->dev, "recover vram bo from shadow start\n"); |
c41d1cf6 | 4313 | mutex_lock(&adev->shadow_list_lock); |
e18aaea7 ND |
4314 | list_for_each_entry(vmbo, &adev->shadow_list, shadow_list) { |
4315 | shadow = &vmbo->bo; | |
403009bf | 4316 | /* No need to recover an evicted BO */ |
d3116756 CK |
4317 | if (shadow->tbo.resource->mem_type != TTM_PL_TT || |
4318 | shadow->tbo.resource->start == AMDGPU_BO_INVALID_OFFSET || | |
4319 | shadow->parent->tbo.resource->mem_type != TTM_PL_VRAM) | |
403009bf CK |
4320 | continue; |
4321 | ||
4322 | r = amdgpu_bo_restore_shadow(shadow, &next); | |
4323 | if (r) | |
4324 | break; | |
4325 | ||
c41d1cf6 | 4326 | if (fence) { |
1712fb1a | 4327 | tmo = dma_fence_wait_timeout(fence, false, tmo); |
403009bf CK |
4328 | dma_fence_put(fence); |
4329 | fence = next; | |
1712fb1a | 4330 | if (tmo == 0) { |
4331 | r = -ETIMEDOUT; | |
c41d1cf6 | 4332 | break; |
1712fb1a | 4333 | } else if (tmo < 0) { |
4334 | r = tmo; | |
4335 | break; | |
4336 | } | |
403009bf CK |
4337 | } else { |
4338 | fence = next; | |
c41d1cf6 | 4339 | } |
c41d1cf6 ML |
4340 | } |
4341 | mutex_unlock(&adev->shadow_list_lock); | |
4342 | ||
403009bf CK |
4343 | if (fence) |
4344 | tmo = dma_fence_wait_timeout(fence, false, tmo); | |
c41d1cf6 ML |
4345 | dma_fence_put(fence); |
4346 | ||
1712fb1a | 4347 | if (r < 0 || tmo <= 0) { |
aac89168 | 4348 | dev_err(adev->dev, "recover vram bo from shadow failed, r is %ld, tmo is %ld\n", r, tmo); |
403009bf CK |
4349 | return -EIO; |
4350 | } | |
c41d1cf6 | 4351 | |
aac89168 | 4352 | dev_info(adev->dev, "recover vram bo from shadow done\n"); |
403009bf | 4353 | return 0; |
c41d1cf6 ML |
4354 | } |
4355 | ||
a90ad3c2 | 4356 | |
e3ecdffa | 4357 | /** |
06ec9070 | 4358 | * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf |
5740682e | 4359 | * |
982a820b | 4360 | * @adev: amdgpu_device pointer |
87e3f136 | 4361 | * @from_hypervisor: request from hypervisor |
5740682e ML |
4362 | * |
4363 | * do VF FLR and reinitialize Asic | |
3f48c681 | 4364 | * return 0 means succeeded otherwise failed |
e3ecdffa AD |
4365 | */ |
4366 | static int amdgpu_device_reset_sriov(struct amdgpu_device *adev, | |
4367 | bool from_hypervisor) | |
5740682e ML |
4368 | { |
4369 | int r; | |
a5f67c93 | 4370 | struct amdgpu_hive_info *hive = NULL; |
7258fa31 | 4371 | int retry_limit = 0; |
5740682e | 4372 | |
7258fa31 | 4373 | retry: |
992110d7 | 4374 | amdgpu_amdkfd_pre_reset(adev); |
5740682e | 4375 | |
428890a3 | 4376 | amdgpu_amdkfd_pre_reset(adev); |
4377 | ||
5740682e ML |
4378 | if (from_hypervisor) |
4379 | r = amdgpu_virt_request_full_gpu(adev, true); | |
4380 | else | |
4381 | r = amdgpu_virt_reset_gpu(adev); | |
4382 | if (r) | |
4383 | return r; | |
a90ad3c2 ML |
4384 | |
4385 | /* Resume IP prior to SMC */ | |
06ec9070 | 4386 | r = amdgpu_device_ip_reinit_early_sriov(adev); |
5740682e ML |
4387 | if (r) |
4388 | goto error; | |
a90ad3c2 | 4389 | |
c9ffa427 | 4390 | amdgpu_virt_init_data_exchange(adev); |
a90ad3c2 | 4391 | |
7a3e0bb2 RZ |
4392 | r = amdgpu_device_fw_loading(adev); |
4393 | if (r) | |
4394 | return r; | |
4395 | ||
a90ad3c2 | 4396 | /* now we are okay to resume SMC/CP/SDMA */ |
06ec9070 | 4397 | r = amdgpu_device_ip_reinit_late_sriov(adev); |
5740682e ML |
4398 | if (r) |
4399 | goto error; | |
a90ad3c2 | 4400 | |
a5f67c93 ZL |
4401 | hive = amdgpu_get_xgmi_hive(adev); |
4402 | /* Update PSP FW topology after reset */ | |
4403 | if (hive && adev->gmc.xgmi.num_physical_nodes > 1) | |
4404 | r = amdgpu_xgmi_update_topology(hive, adev); | |
4405 | ||
4406 | if (hive) | |
4407 | amdgpu_put_xgmi_hive(hive); | |
4408 | ||
4409 | if (!r) { | |
4410 | amdgpu_irq_gpu_reset_resume_helper(adev); | |
4411 | r = amdgpu_ib_ring_tests(adev); | |
4412 | amdgpu_amdkfd_post_reset(adev); | |
4413 | } | |
a90ad3c2 | 4414 | |
abc34253 | 4415 | error: |
c41d1cf6 | 4416 | if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) { |
e3526257 | 4417 | amdgpu_inc_vram_lost(adev); |
c33adbc7 | 4418 | r = amdgpu_device_recover_vram(adev); |
a90ad3c2 | 4419 | } |
437f3e0b | 4420 | amdgpu_virt_release_full_gpu(adev, true); |
a90ad3c2 | 4421 | |
7258fa31 SK |
4422 | if (AMDGPU_RETRY_SRIOV_RESET(r)) { |
4423 | if (retry_limit < AMDGPU_MAX_RETRY_LIMIT) { | |
4424 | retry_limit++; | |
4425 | goto retry; | |
4426 | } else | |
4427 | DRM_ERROR("GPU reset retry is beyond the retry limit\n"); | |
4428 | } | |
4429 | ||
a90ad3c2 ML |
4430 | return r; |
4431 | } | |
4432 | ||
9a1cddd6 | 4433 | /** |
4434 | * amdgpu_device_has_job_running - check if there is any job in mirror list | |
4435 | * | |
982a820b | 4436 | * @adev: amdgpu_device pointer |
9a1cddd6 | 4437 | * |
4438 | * check if there is any job in mirror list | |
4439 | */ | |
4440 | bool amdgpu_device_has_job_running(struct amdgpu_device *adev) | |
4441 | { | |
4442 | int i; | |
4443 | struct drm_sched_job *job; | |
4444 | ||
4445 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { | |
4446 | struct amdgpu_ring *ring = adev->rings[i]; | |
4447 | ||
4448 | if (!ring || !ring->sched.thread) | |
4449 | continue; | |
4450 | ||
4451 | spin_lock(&ring->sched.job_list_lock); | |
6efa4b46 LT |
4452 | job = list_first_entry_or_null(&ring->sched.pending_list, |
4453 | struct drm_sched_job, list); | |
9a1cddd6 | 4454 | spin_unlock(&ring->sched.job_list_lock); |
4455 | if (job) | |
4456 | return true; | |
4457 | } | |
4458 | return false; | |
4459 | } | |
4460 | ||
12938fad CK |
4461 | /** |
4462 | * amdgpu_device_should_recover_gpu - check if we should try GPU recovery | |
4463 | * | |
982a820b | 4464 | * @adev: amdgpu_device pointer |
12938fad CK |
4465 | * |
4466 | * Check amdgpu_gpu_recovery and SRIOV status to see if we should try to recover | |
4467 | * a hung GPU. | |
4468 | */ | |
4469 | bool amdgpu_device_should_recover_gpu(struct amdgpu_device *adev) | |
4470 | { | |
4471 | if (!amdgpu_device_ip_check_soft_reset(adev)) { | |
aac89168 | 4472 | dev_info(adev->dev, "Timeout, but no hardware hang detected.\n"); |
12938fad CK |
4473 | return false; |
4474 | } | |
4475 | ||
3ba7b418 AG |
4476 | if (amdgpu_gpu_recovery == 0) |
4477 | goto disabled; | |
4478 | ||
4479 | if (amdgpu_sriov_vf(adev)) | |
4480 | return true; | |
4481 | ||
4482 | if (amdgpu_gpu_recovery == -1) { | |
4483 | switch (adev->asic_type) { | |
b3523c45 AD |
4484 | #ifdef CONFIG_DRM_AMDGPU_SI |
4485 | case CHIP_VERDE: | |
4486 | case CHIP_TAHITI: | |
4487 | case CHIP_PITCAIRN: | |
4488 | case CHIP_OLAND: | |
4489 | case CHIP_HAINAN: | |
4490 | #endif | |
4491 | #ifdef CONFIG_DRM_AMDGPU_CIK | |
4492 | case CHIP_KAVERI: | |
4493 | case CHIP_KABINI: | |
4494 | case CHIP_MULLINS: | |
4495 | #endif | |
4496 | case CHIP_CARRIZO: | |
4497 | case CHIP_STONEY: | |
4498 | case CHIP_CYAN_SKILLFISH: | |
3ba7b418 | 4499 | goto disabled; |
b3523c45 AD |
4500 | default: |
4501 | break; | |
3ba7b418 | 4502 | } |
12938fad CK |
4503 | } |
4504 | ||
4505 | return true; | |
3ba7b418 AG |
4506 | |
4507 | disabled: | |
aac89168 | 4508 | dev_info(adev->dev, "GPU recovery disabled.\n"); |
3ba7b418 | 4509 | return false; |
12938fad CK |
4510 | } |
4511 | ||
5c03e584 FX |
4512 | int amdgpu_device_mode1_reset(struct amdgpu_device *adev) |
4513 | { | |
4514 | u32 i; | |
4515 | int ret = 0; | |
4516 | ||
4517 | amdgpu_atombios_scratch_regs_engine_hung(adev, true); | |
4518 | ||
4519 | dev_info(adev->dev, "GPU mode1 reset\n"); | |
4520 | ||
4521 | /* disable BM */ | |
4522 | pci_clear_master(adev->pdev); | |
4523 | ||
4524 | amdgpu_device_cache_pci_state(adev->pdev); | |
4525 | ||
4526 | if (amdgpu_dpm_is_mode1_reset_supported(adev)) { | |
4527 | dev_info(adev->dev, "GPU smu mode1 reset\n"); | |
4528 | ret = amdgpu_dpm_mode1_reset(adev); | |
4529 | } else { | |
4530 | dev_info(adev->dev, "GPU psp mode1 reset\n"); | |
4531 | ret = psp_gpu_reset(adev); | |
4532 | } | |
4533 | ||
4534 | if (ret) | |
4535 | dev_err(adev->dev, "GPU mode1 reset failed\n"); | |
4536 | ||
4537 | amdgpu_device_load_pci_state(adev->pdev); | |
4538 | ||
4539 | /* wait for asic to come out of reset */ | |
4540 | for (i = 0; i < adev->usec_timeout; i++) { | |
4541 | u32 memsize = adev->nbio.funcs->get_memsize(adev); | |
4542 | ||
4543 | if (memsize != 0xffffffff) | |
4544 | break; | |
4545 | udelay(1); | |
4546 | } | |
4547 | ||
4548 | amdgpu_atombios_scratch_regs_engine_hung(adev, false); | |
4549 | return ret; | |
4550 | } | |
5c6dd71e | 4551 | |
e3c1b071 | 4552 | int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, |
04442bf7 | 4553 | struct amdgpu_reset_context *reset_context) |
26bc5340 | 4554 | { |
5c1e6fa4 | 4555 | int i, r = 0; |
04442bf7 LL |
4556 | struct amdgpu_job *job = NULL; |
4557 | bool need_full_reset = | |
4558 | test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); | |
4559 | ||
4560 | if (reset_context->reset_req_dev == adev) | |
4561 | job = reset_context->job; | |
71182665 | 4562 | |
b602ca5f TZ |
4563 | if (amdgpu_sriov_vf(adev)) { |
4564 | /* stop the data exchange thread */ | |
4565 | amdgpu_virt_fini_data_exchange(adev); | |
4566 | } | |
4567 | ||
71182665 | 4568 | /* block all schedulers and reset given job's ring */ |
0875dc9e CZ |
4569 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { |
4570 | struct amdgpu_ring *ring = adev->rings[i]; | |
4571 | ||
51687759 | 4572 | if (!ring || !ring->sched.thread) |
0875dc9e | 4573 | continue; |
5740682e | 4574 | |
c530b02f JZ |
4575 | /*clear job fence from fence drv to avoid force_completion |
4576 | *leave NULL and vm flush fence in fence drv */ | |
5c1e6fa4 | 4577 | amdgpu_fence_driver_clear_job_fences(ring); |
c530b02f | 4578 | |
2f9d4084 ML |
4579 | /* after all hw jobs are reset, hw fence is meaningless, so force_completion */ |
4580 | amdgpu_fence_driver_force_completion(ring); | |
0875dc9e | 4581 | } |
d38ceaf9 | 4582 | |
ff99849b | 4583 | if (job && job->vm) |
222b5f04 AG |
4584 | drm_sched_increase_karma(&job->base); |
4585 | ||
04442bf7 | 4586 | r = amdgpu_reset_prepare_hwcontext(adev, reset_context); |
404b277b LL |
4587 | /* If reset handler not implemented, continue; otherwise return */ |
4588 | if (r == -ENOSYS) | |
4589 | r = 0; | |
4590 | else | |
04442bf7 LL |
4591 | return r; |
4592 | ||
1d721ed6 | 4593 | /* Don't suspend on bare metal if we are not going to HW reset the ASIC */ |
26bc5340 AG |
4594 | if (!amdgpu_sriov_vf(adev)) { |
4595 | ||
4596 | if (!need_full_reset) | |
4597 | need_full_reset = amdgpu_device_ip_need_full_reset(adev); | |
4598 | ||
4599 | if (!need_full_reset) { | |
4600 | amdgpu_device_ip_pre_soft_reset(adev); | |
4601 | r = amdgpu_device_ip_soft_reset(adev); | |
4602 | amdgpu_device_ip_post_soft_reset(adev); | |
4603 | if (r || amdgpu_device_ip_check_soft_reset(adev)) { | |
aac89168 | 4604 | dev_info(adev->dev, "soft reset failed, will fallback to full reset!\n"); |
26bc5340 AG |
4605 | need_full_reset = true; |
4606 | } | |
4607 | } | |
4608 | ||
4609 | if (need_full_reset) | |
4610 | r = amdgpu_device_ip_suspend(adev); | |
04442bf7 LL |
4611 | if (need_full_reset) |
4612 | set_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); | |
4613 | else | |
4614 | clear_bit(AMDGPU_NEED_FULL_RESET, | |
4615 | &reset_context->flags); | |
26bc5340 AG |
4616 | } |
4617 | ||
4618 | return r; | |
4619 | } | |
4620 | ||
04442bf7 LL |
4621 | int amdgpu_do_asic_reset(struct list_head *device_list_handle, |
4622 | struct amdgpu_reset_context *reset_context) | |
26bc5340 AG |
4623 | { |
4624 | struct amdgpu_device *tmp_adev = NULL; | |
04442bf7 | 4625 | bool need_full_reset, skip_hw_reset, vram_lost = false; |
26bc5340 AG |
4626 | int r = 0; |
4627 | ||
04442bf7 LL |
4628 | /* Try reset handler method first */ |
4629 | tmp_adev = list_first_entry(device_list_handle, struct amdgpu_device, | |
4630 | reset_list); | |
4631 | r = amdgpu_reset_perform_reset(tmp_adev, reset_context); | |
404b277b LL |
4632 | /* If reset handler not implemented, continue; otherwise return */ |
4633 | if (r == -ENOSYS) | |
4634 | r = 0; | |
4635 | else | |
04442bf7 LL |
4636 | return r; |
4637 | ||
4638 | /* Reset handler not implemented, use the default method */ | |
4639 | need_full_reset = | |
4640 | test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); | |
4641 | skip_hw_reset = test_bit(AMDGPU_SKIP_HW_RESET, &reset_context->flags); | |
4642 | ||
26bc5340 | 4643 | /* |
655ce9cb | 4644 | * ASIC reset has to be done on all XGMI hive nodes ASAP |
26bc5340 AG |
4645 | * to allow proper links negotiation in FW (within 1 sec) |
4646 | */ | |
7ac71382 | 4647 | if (!skip_hw_reset && need_full_reset) { |
655ce9cb | 4648 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
041a62bc | 4649 | /* For XGMI run all resets in parallel to speed up the process */ |
d4535e2c | 4650 | if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) { |
e3c1b071 | 4651 | tmp_adev->gmc.xgmi.pending_reset = false; |
c96cf282 | 4652 | if (!queue_work(system_unbound_wq, &tmp_adev->xgmi_reset_work)) |
d4535e2c AG |
4653 | r = -EALREADY; |
4654 | } else | |
4655 | r = amdgpu_asic_reset(tmp_adev); | |
d4535e2c | 4656 | |
041a62bc | 4657 | if (r) { |
aac89168 | 4658 | dev_err(tmp_adev->dev, "ASIC reset failed with error, %d for drm dev, %s", |
4a580877 | 4659 | r, adev_to_drm(tmp_adev)->unique); |
041a62bc | 4660 | break; |
ce316fa5 LM |
4661 | } |
4662 | } | |
4663 | ||
041a62bc AG |
4664 | /* For XGMI wait for all resets to complete before proceed */ |
4665 | if (!r) { | |
655ce9cb | 4666 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
ce316fa5 LM |
4667 | if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) { |
4668 | flush_work(&tmp_adev->xgmi_reset_work); | |
4669 | r = tmp_adev->asic_reset_res; | |
4670 | if (r) | |
4671 | break; | |
ce316fa5 LM |
4672 | } |
4673 | } | |
4674 | } | |
ce316fa5 | 4675 | } |
26bc5340 | 4676 | |
43c4d576 | 4677 | if (!r && amdgpu_ras_intr_triggered()) { |
655ce9cb | 4678 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
5e67bba3 | 4679 | if (tmp_adev->mmhub.ras && tmp_adev->mmhub.ras->ras_block.hw_ops && |
4680 | tmp_adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count) | |
4681 | tmp_adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count(tmp_adev); | |
43c4d576 JC |
4682 | } |
4683 | ||
00eaa571 | 4684 | amdgpu_ras_intr_cleared(); |
43c4d576 | 4685 | } |
00eaa571 | 4686 | |
655ce9cb | 4687 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
26bc5340 AG |
4688 | if (need_full_reset) { |
4689 | /* post card */ | |
e3c1b071 | 4690 | r = amdgpu_device_asic_init(tmp_adev); |
4691 | if (r) { | |
aac89168 | 4692 | dev_warn(tmp_adev->dev, "asic atom init failed!"); |
e3c1b071 | 4693 | } else { |
26bc5340 | 4694 | dev_info(tmp_adev->dev, "GPU reset succeeded, trying to resume\n"); |
9cec53c1 JZ |
4695 | r = amdgpu_amdkfd_resume_iommu(tmp_adev); |
4696 | if (r) | |
4697 | goto out; | |
4698 | ||
26bc5340 AG |
4699 | r = amdgpu_device_ip_resume_phase1(tmp_adev); |
4700 | if (r) | |
4701 | goto out; | |
4702 | ||
4703 | vram_lost = amdgpu_device_check_vram_lost(tmp_adev); | |
4704 | if (vram_lost) { | |
77e7f829 | 4705 | DRM_INFO("VRAM is lost due to GPU reset!\n"); |
e3526257 | 4706 | amdgpu_inc_vram_lost(tmp_adev); |
26bc5340 AG |
4707 | } |
4708 | ||
26bc5340 AG |
4709 | r = amdgpu_device_fw_loading(tmp_adev); |
4710 | if (r) | |
4711 | return r; | |
4712 | ||
4713 | r = amdgpu_device_ip_resume_phase2(tmp_adev); | |
4714 | if (r) | |
4715 | goto out; | |
4716 | ||
4717 | if (vram_lost) | |
4718 | amdgpu_device_fill_reset_magic(tmp_adev); | |
4719 | ||
fdafb359 EQ |
4720 | /* |
4721 | * Add this ASIC as tracked as reset was already | |
4722 | * complete successfully. | |
4723 | */ | |
4724 | amdgpu_register_gpu_instance(tmp_adev); | |
4725 | ||
04442bf7 LL |
4726 | if (!reset_context->hive && |
4727 | tmp_adev->gmc.xgmi.num_physical_nodes > 1) | |
e3c1b071 | 4728 | amdgpu_xgmi_add_device(tmp_adev); |
4729 | ||
7c04ca50 | 4730 | r = amdgpu_device_ip_late_init(tmp_adev); |
4731 | if (r) | |
4732 | goto out; | |
4733 | ||
087451f3 | 4734 | drm_fb_helper_set_suspend_unlocked(adev_to_drm(tmp_adev)->fb_helper, false); |
565d1941 | 4735 | |
e8fbaf03 GC |
4736 | /* |
4737 | * The GPU enters bad state once faulty pages | |
4738 | * by ECC has reached the threshold, and ras | |
4739 | * recovery is scheduled next. So add one check | |
4740 | * here to break recovery if it indeed exceeds | |
4741 | * bad page threshold, and remind user to | |
4742 | * retire this GPU or setting one bigger | |
4743 | * bad_page_threshold value to fix this once | |
4744 | * probing driver again. | |
4745 | */ | |
11003c68 | 4746 | if (!amdgpu_ras_eeprom_check_err_threshold(tmp_adev)) { |
e8fbaf03 GC |
4747 | /* must succeed. */ |
4748 | amdgpu_ras_resume(tmp_adev); | |
4749 | } else { | |
4750 | r = -EINVAL; | |
4751 | goto out; | |
4752 | } | |
e79a04d5 | 4753 | |
26bc5340 | 4754 | /* Update PSP FW topology after reset */ |
04442bf7 LL |
4755 | if (reset_context->hive && |
4756 | tmp_adev->gmc.xgmi.num_physical_nodes > 1) | |
4757 | r = amdgpu_xgmi_update_topology( | |
4758 | reset_context->hive, tmp_adev); | |
26bc5340 AG |
4759 | } |
4760 | } | |
4761 | ||
26bc5340 AG |
4762 | out: |
4763 | if (!r) { | |
4764 | amdgpu_irq_gpu_reset_resume_helper(tmp_adev); | |
4765 | r = amdgpu_ib_ring_tests(tmp_adev); | |
4766 | if (r) { | |
4767 | dev_err(tmp_adev->dev, "ib ring test failed (%d).\n", r); | |
26bc5340 AG |
4768 | need_full_reset = true; |
4769 | r = -EAGAIN; | |
4770 | goto end; | |
4771 | } | |
4772 | } | |
4773 | ||
4774 | if (!r) | |
4775 | r = amdgpu_device_recover_vram(tmp_adev); | |
4776 | else | |
4777 | tmp_adev->asic_reset_res = r; | |
4778 | } | |
4779 | ||
4780 | end: | |
04442bf7 LL |
4781 | if (need_full_reset) |
4782 | set_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); | |
4783 | else | |
4784 | clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); | |
26bc5340 AG |
4785 | return r; |
4786 | } | |
4787 | ||
08ebb485 DL |
4788 | static bool amdgpu_device_lock_adev(struct amdgpu_device *adev, |
4789 | struct amdgpu_hive_info *hive) | |
26bc5340 | 4790 | { |
53b3f8f4 DL |
4791 | if (atomic_cmpxchg(&adev->in_gpu_reset, 0, 1) != 0) |
4792 | return false; | |
4793 | ||
08ebb485 DL |
4794 | if (hive) { |
4795 | down_write_nest_lock(&adev->reset_sem, &hive->hive_lock); | |
4796 | } else { | |
4797 | down_write(&adev->reset_sem); | |
4798 | } | |
5740682e | 4799 | |
a3a09142 AD |
4800 | switch (amdgpu_asic_reset_method(adev)) { |
4801 | case AMD_RESET_METHOD_MODE1: | |
4802 | adev->mp1_state = PP_MP1_STATE_SHUTDOWN; | |
4803 | break; | |
4804 | case AMD_RESET_METHOD_MODE2: | |
4805 | adev->mp1_state = PP_MP1_STATE_RESET; | |
4806 | break; | |
4807 | default: | |
4808 | adev->mp1_state = PP_MP1_STATE_NONE; | |
4809 | break; | |
4810 | } | |
1d721ed6 AG |
4811 | |
4812 | return true; | |
26bc5340 | 4813 | } |
d38ceaf9 | 4814 | |
26bc5340 AG |
4815 | static void amdgpu_device_unlock_adev(struct amdgpu_device *adev) |
4816 | { | |
89041940 | 4817 | amdgpu_vf_error_trans_all(adev); |
a3a09142 | 4818 | adev->mp1_state = PP_MP1_STATE_NONE; |
53b3f8f4 | 4819 | atomic_set(&adev->in_gpu_reset, 0); |
6049db43 | 4820 | up_write(&adev->reset_sem); |
26bc5340 AG |
4821 | } |
4822 | ||
91fb309d HC |
4823 | /* |
4824 | * to lockup a list of amdgpu devices in a hive safely, if not a hive | |
4825 | * with multiple nodes, it will be similar as amdgpu_device_lock_adev. | |
4826 | * | |
4827 | * unlock won't require roll back. | |
4828 | */ | |
4829 | static int amdgpu_device_lock_hive_adev(struct amdgpu_device *adev, struct amdgpu_hive_info *hive) | |
4830 | { | |
4831 | struct amdgpu_device *tmp_adev = NULL; | |
4832 | ||
175ac6ec | 4833 | if (!amdgpu_sriov_vf(adev) && (adev->gmc.xgmi.num_physical_nodes > 1)) { |
91fb309d HC |
4834 | if (!hive) { |
4835 | dev_err(adev->dev, "Hive is NULL while device has multiple xgmi nodes"); | |
4836 | return -ENODEV; | |
4837 | } | |
4838 | list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { | |
4839 | if (!amdgpu_device_lock_adev(tmp_adev, hive)) | |
4840 | goto roll_back; | |
4841 | } | |
4842 | } else if (!amdgpu_device_lock_adev(adev, hive)) | |
4843 | return -EAGAIN; | |
4844 | ||
4845 | return 0; | |
4846 | roll_back: | |
4847 | if (!list_is_first(&tmp_adev->gmc.xgmi.head, &hive->device_list)) { | |
4848 | /* | |
4849 | * if the lockup iteration break in the middle of a hive, | |
4850 | * it may means there may has a race issue, | |
4851 | * or a hive device locked up independently. | |
4852 | * we may be in trouble and may not, so will try to roll back | |
4853 | * the lock and give out a warnning. | |
4854 | */ | |
4855 | dev_warn(tmp_adev->dev, "Hive lock iteration broke in the middle. Rolling back to unlock"); | |
4856 | list_for_each_entry_continue_reverse(tmp_adev, &hive->device_list, gmc.xgmi.head) { | |
4857 | amdgpu_device_unlock_adev(tmp_adev); | |
4858 | } | |
4859 | } | |
4860 | return -EAGAIN; | |
4861 | } | |
4862 | ||
3f12acc8 EQ |
4863 | static void amdgpu_device_resume_display_audio(struct amdgpu_device *adev) |
4864 | { | |
4865 | struct pci_dev *p = NULL; | |
4866 | ||
4867 | p = pci_get_domain_bus_and_slot(pci_domain_nr(adev->pdev->bus), | |
4868 | adev->pdev->bus->number, 1); | |
4869 | if (p) { | |
4870 | pm_runtime_enable(&(p->dev)); | |
4871 | pm_runtime_resume(&(p->dev)); | |
4872 | } | |
4873 | } | |
4874 | ||
4875 | static int amdgpu_device_suspend_display_audio(struct amdgpu_device *adev) | |
4876 | { | |
4877 | enum amd_reset_method reset_method; | |
4878 | struct pci_dev *p = NULL; | |
4879 | u64 expires; | |
4880 | ||
4881 | /* | |
4882 | * For now, only BACO and mode1 reset are confirmed | |
4883 | * to suffer the audio issue without proper suspended. | |
4884 | */ | |
4885 | reset_method = amdgpu_asic_reset_method(adev); | |
4886 | if ((reset_method != AMD_RESET_METHOD_BACO) && | |
4887 | (reset_method != AMD_RESET_METHOD_MODE1)) | |
4888 | return -EINVAL; | |
4889 | ||
4890 | p = pci_get_domain_bus_and_slot(pci_domain_nr(adev->pdev->bus), | |
4891 | adev->pdev->bus->number, 1); | |
4892 | if (!p) | |
4893 | return -ENODEV; | |
4894 | ||
4895 | expires = pm_runtime_autosuspend_expiration(&(p->dev)); | |
4896 | if (!expires) | |
4897 | /* | |
4898 | * If we cannot get the audio device autosuspend delay, | |
4899 | * a fixed 4S interval will be used. Considering 3S is | |
4900 | * the audio controller default autosuspend delay setting. | |
4901 | * 4S used here is guaranteed to cover that. | |
4902 | */ | |
54b7feb9 | 4903 | expires = ktime_get_mono_fast_ns() + NSEC_PER_SEC * 4ULL; |
3f12acc8 EQ |
4904 | |
4905 | while (!pm_runtime_status_suspended(&(p->dev))) { | |
4906 | if (!pm_runtime_suspend(&(p->dev))) | |
4907 | break; | |
4908 | ||
4909 | if (expires < ktime_get_mono_fast_ns()) { | |
4910 | dev_warn(adev->dev, "failed to suspend display audio\n"); | |
4911 | /* TODO: abort the succeeding gpu reset? */ | |
4912 | return -ETIMEDOUT; | |
4913 | } | |
4914 | } | |
4915 | ||
4916 | pm_runtime_disable(&(p->dev)); | |
4917 | ||
4918 | return 0; | |
4919 | } | |
4920 | ||
9d8d96be | 4921 | static void amdgpu_device_recheck_guilty_jobs( |
04442bf7 LL |
4922 | struct amdgpu_device *adev, struct list_head *device_list_handle, |
4923 | struct amdgpu_reset_context *reset_context) | |
e6c6338f JZ |
4924 | { |
4925 | int i, r = 0; | |
4926 | ||
4927 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { | |
4928 | struct amdgpu_ring *ring = adev->rings[i]; | |
4929 | int ret = 0; | |
4930 | struct drm_sched_job *s_job; | |
4931 | ||
4932 | if (!ring || !ring->sched.thread) | |
4933 | continue; | |
4934 | ||
4935 | s_job = list_first_entry_or_null(&ring->sched.pending_list, | |
4936 | struct drm_sched_job, list); | |
4937 | if (s_job == NULL) | |
4938 | continue; | |
4939 | ||
4940 | /* clear job's guilty and depend the folowing step to decide the real one */ | |
4941 | drm_sched_reset_karma(s_job); | |
38d4e463 JC |
4942 | /* for the real bad job, it will be resubmitted twice, adding a dma_fence_get |
4943 | * to make sure fence is balanced */ | |
4944 | dma_fence_get(s_job->s_fence->parent); | |
e6c6338f JZ |
4945 | drm_sched_resubmit_jobs_ext(&ring->sched, 1); |
4946 | ||
4947 | ret = dma_fence_wait_timeout(s_job->s_fence->parent, false, ring->sched.timeout); | |
4948 | if (ret == 0) { /* timeout */ | |
4949 | DRM_ERROR("Found the real bad job! ring:%s, job_id:%llx\n", | |
4950 | ring->sched.name, s_job->id); | |
4951 | ||
4952 | /* set guilty */ | |
4953 | drm_sched_increase_karma(s_job); | |
4954 | retry: | |
4955 | /* do hw reset */ | |
4956 | if (amdgpu_sriov_vf(adev)) { | |
4957 | amdgpu_virt_fini_data_exchange(adev); | |
4958 | r = amdgpu_device_reset_sriov(adev, false); | |
4959 | if (r) | |
4960 | adev->asic_reset_res = r; | |
4961 | } else { | |
04442bf7 LL |
4962 | clear_bit(AMDGPU_SKIP_HW_RESET, |
4963 | &reset_context->flags); | |
4964 | r = amdgpu_do_asic_reset(device_list_handle, | |
4965 | reset_context); | |
e6c6338f JZ |
4966 | if (r && r == -EAGAIN) |
4967 | goto retry; | |
4968 | } | |
4969 | ||
4970 | /* | |
4971 | * add reset counter so that the following | |
4972 | * resubmitted job could flush vmid | |
4973 | */ | |
4974 | atomic_inc(&adev->gpu_reset_counter); | |
4975 | continue; | |
4976 | } | |
4977 | ||
4978 | /* got the hw fence, signal finished fence */ | |
4979 | atomic_dec(ring->sched.score); | |
38d4e463 | 4980 | dma_fence_put(s_job->s_fence->parent); |
e6c6338f JZ |
4981 | dma_fence_get(&s_job->s_fence->finished); |
4982 | dma_fence_signal(&s_job->s_fence->finished); | |
4983 | dma_fence_put(&s_job->s_fence->finished); | |
4984 | ||
4985 | /* remove node from list and free the job */ | |
4986 | spin_lock(&ring->sched.job_list_lock); | |
4987 | list_del_init(&s_job->list); | |
4988 | spin_unlock(&ring->sched.job_list_lock); | |
4989 | ring->sched.ops->free_job(s_job); | |
4990 | } | |
4991 | } | |
4992 | ||
26bc5340 AG |
4993 | /** |
4994 | * amdgpu_device_gpu_recover - reset the asic and recover scheduler | |
4995 | * | |
982a820b | 4996 | * @adev: amdgpu_device pointer |
26bc5340 AG |
4997 | * @job: which job trigger hang |
4998 | * | |
4999 | * Attempt to reset the GPU if it has hung (all asics). | |
5000 | * Attempt to do soft-reset or full-reset and reinitialize Asic | |
5001 | * Returns 0 for success or an error on failure. | |
5002 | */ | |
5003 | ||
5004 | int amdgpu_device_gpu_recover(struct amdgpu_device *adev, | |
5005 | struct amdgpu_job *job) | |
5006 | { | |
1d721ed6 | 5007 | struct list_head device_list, *device_list_handle = NULL; |
7dd8c205 | 5008 | bool job_signaled = false; |
26bc5340 | 5009 | struct amdgpu_hive_info *hive = NULL; |
26bc5340 | 5010 | struct amdgpu_device *tmp_adev = NULL; |
1d721ed6 | 5011 | int i, r = 0; |
bb5c7235 | 5012 | bool need_emergency_restart = false; |
3f12acc8 | 5013 | bool audio_suspended = false; |
e6c6338f | 5014 | int tmp_vram_lost_counter; |
04442bf7 LL |
5015 | struct amdgpu_reset_context reset_context; |
5016 | ||
5017 | memset(&reset_context, 0, sizeof(reset_context)); | |
26bc5340 | 5018 | |
6e3cd2a9 | 5019 | /* |
bb5c7235 WS |
5020 | * Special case: RAS triggered and full reset isn't supported |
5021 | */ | |
5022 | need_emergency_restart = amdgpu_ras_need_emergency_restart(adev); | |
5023 | ||
d5ea093e AG |
5024 | /* |
5025 | * Flush RAM to disk so that after reboot | |
5026 | * the user can read log and see why the system rebooted. | |
5027 | */ | |
bb5c7235 | 5028 | if (need_emergency_restart && amdgpu_ras_get_context(adev)->reboot) { |
d5ea093e AG |
5029 | DRM_WARN("Emergency reboot."); |
5030 | ||
5031 | ksys_sync_helper(); | |
5032 | emergency_restart(); | |
5033 | } | |
5034 | ||
b823821f | 5035 | dev_info(adev->dev, "GPU %s begin!\n", |
bb5c7235 | 5036 | need_emergency_restart ? "jobs stop":"reset"); |
26bc5340 AG |
5037 | |
5038 | /* | |
1d721ed6 AG |
5039 | * Here we trylock to avoid chain of resets executing from |
5040 | * either trigger by jobs on different adevs in XGMI hive or jobs on | |
5041 | * different schedulers for same device while this TO handler is running. | |
5042 | * We always reset all schedulers for device and all devices for XGMI | |
5043 | * hive so that should take care of them too. | |
26bc5340 | 5044 | */ |
175ac6ec ZL |
5045 | if (!amdgpu_sriov_vf(adev)) |
5046 | hive = amdgpu_get_xgmi_hive(adev); | |
53b3f8f4 DL |
5047 | if (hive) { |
5048 | if (atomic_cmpxchg(&hive->in_reset, 0, 1) != 0) { | |
5049 | DRM_INFO("Bailing on TDR for s_job:%llx, hive: %llx as another already in progress", | |
5050 | job ? job->base.id : -1, hive->hive_id); | |
d95e8e97 | 5051 | amdgpu_put_xgmi_hive(hive); |
ff99849b | 5052 | if (job && job->vm) |
91fb309d | 5053 | drm_sched_increase_karma(&job->base); |
53b3f8f4 DL |
5054 | return 0; |
5055 | } | |
5056 | mutex_lock(&hive->hive_lock); | |
1d721ed6 | 5057 | } |
26bc5340 | 5058 | |
04442bf7 LL |
5059 | reset_context.method = AMD_RESET_METHOD_NONE; |
5060 | reset_context.reset_req_dev = adev; | |
5061 | reset_context.job = job; | |
5062 | reset_context.hive = hive; | |
5063 | clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags); | |
5064 | ||
91fb309d HC |
5065 | /* |
5066 | * lock the device before we try to operate the linked list | |
5067 | * if didn't get the device lock, don't touch the linked list since | |
5068 | * others may iterating it. | |
5069 | */ | |
5070 | r = amdgpu_device_lock_hive_adev(adev, hive); | |
5071 | if (r) { | |
5072 | dev_info(adev->dev, "Bailing on TDR for s_job:%llx, as another already in progress", | |
5073 | job ? job->base.id : -1); | |
5074 | ||
5075 | /* even we skipped this reset, still need to set the job to guilty */ | |
ff99849b | 5076 | if (job && job->vm) |
91fb309d HC |
5077 | drm_sched_increase_karma(&job->base); |
5078 | goto skip_recovery; | |
5079 | } | |
5080 | ||
9e94d22c EQ |
5081 | /* |
5082 | * Build list of devices to reset. | |
5083 | * In case we are in XGMI hive mode, resort the device list | |
5084 | * to put adev in the 1st position. | |
5085 | */ | |
5086 | INIT_LIST_HEAD(&device_list); | |
175ac6ec | 5087 | if (!amdgpu_sriov_vf(adev) && (adev->gmc.xgmi.num_physical_nodes > 1)) { |
655ce9cb | 5088 | list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) |
5089 | list_add_tail(&tmp_adev->reset_list, &device_list); | |
5090 | if (!list_is_first(&adev->reset_list, &device_list)) | |
5091 | list_rotate_to_front(&adev->reset_list, &device_list); | |
5092 | device_list_handle = &device_list; | |
26bc5340 | 5093 | } else { |
655ce9cb | 5094 | list_add_tail(&adev->reset_list, &device_list); |
26bc5340 AG |
5095 | device_list_handle = &device_list; |
5096 | } | |
5097 | ||
1d721ed6 | 5098 | /* block all schedulers and reset given job's ring */ |
655ce9cb | 5099 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
3f12acc8 EQ |
5100 | /* |
5101 | * Try to put the audio codec into suspend state | |
5102 | * before gpu reset started. | |
5103 | * | |
5104 | * Due to the power domain of the graphics device | |
5105 | * is shared with AZ power domain. Without this, | |
5106 | * we may change the audio hardware from behind | |
5107 | * the audio driver's back. That will trigger | |
5108 | * some audio codec errors. | |
5109 | */ | |
5110 | if (!amdgpu_device_suspend_display_audio(tmp_adev)) | |
5111 | audio_suspended = true; | |
5112 | ||
9e94d22c EQ |
5113 | amdgpu_ras_set_error_query_ready(tmp_adev, false); |
5114 | ||
52fb44cf EQ |
5115 | cancel_delayed_work_sync(&tmp_adev->delayed_init_work); |
5116 | ||
428890a3 | 5117 | if (!amdgpu_sriov_vf(tmp_adev)) |
5118 | amdgpu_amdkfd_pre_reset(tmp_adev); | |
9e94d22c | 5119 | |
12ffa55d AG |
5120 | /* |
5121 | * Mark these ASICs to be reseted as untracked first | |
5122 | * And add them back after reset completed | |
5123 | */ | |
5124 | amdgpu_unregister_gpu_instance(tmp_adev); | |
5125 | ||
087451f3 | 5126 | drm_fb_helper_set_suspend_unlocked(adev_to_drm(adev)->fb_helper, true); |
565d1941 | 5127 | |
f1c1314b | 5128 | /* disable ras on ALL IPs */ |
bb5c7235 | 5129 | if (!need_emergency_restart && |
b823821f | 5130 | amdgpu_device_ip_need_full_reset(tmp_adev)) |
f1c1314b | 5131 | amdgpu_ras_suspend(tmp_adev); |
5132 | ||
1d721ed6 AG |
5133 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { |
5134 | struct amdgpu_ring *ring = tmp_adev->rings[i]; | |
5135 | ||
5136 | if (!ring || !ring->sched.thread) | |
5137 | continue; | |
5138 | ||
0b2d2c2e | 5139 | drm_sched_stop(&ring->sched, job ? &job->base : NULL); |
7c6e68c7 | 5140 | |
bb5c7235 | 5141 | if (need_emergency_restart) |
7c6e68c7 | 5142 | amdgpu_job_stop_all_jobs_on_sched(&ring->sched); |
1d721ed6 | 5143 | } |
8f8c80f4 | 5144 | atomic_inc(&tmp_adev->gpu_reset_counter); |
1d721ed6 AG |
5145 | } |
5146 | ||
bb5c7235 | 5147 | if (need_emergency_restart) |
7c6e68c7 AG |
5148 | goto skip_sched_resume; |
5149 | ||
1d721ed6 AG |
5150 | /* |
5151 | * Must check guilty signal here since after this point all old | |
5152 | * HW fences are force signaled. | |
5153 | * | |
5154 | * job->base holds a reference to parent fence | |
5155 | */ | |
5156 | if (job && job->base.s_fence->parent && | |
7dd8c205 | 5157 | dma_fence_is_signaled(job->base.s_fence->parent)) { |
1d721ed6 | 5158 | job_signaled = true; |
1d721ed6 AG |
5159 | dev_info(adev->dev, "Guilty job already signaled, skipping HW reset"); |
5160 | goto skip_hw_reset; | |
5161 | } | |
5162 | ||
26bc5340 | 5163 | retry: /* Rest of adevs pre asic reset from XGMI hive. */ |
655ce9cb | 5164 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
04442bf7 | 5165 | r = amdgpu_device_pre_asic_reset(tmp_adev, &reset_context); |
26bc5340 AG |
5166 | /*TODO Should we stop ?*/ |
5167 | if (r) { | |
aac89168 | 5168 | dev_err(tmp_adev->dev, "GPU pre asic reset failed with err, %d for drm dev, %s ", |
4a580877 | 5169 | r, adev_to_drm(tmp_adev)->unique); |
26bc5340 AG |
5170 | tmp_adev->asic_reset_res = r; |
5171 | } | |
5172 | } | |
5173 | ||
e6c6338f | 5174 | tmp_vram_lost_counter = atomic_read(&((adev)->vram_lost_counter)); |
26bc5340 | 5175 | /* Actual ASIC resets if needed.*/ |
4f30d920 | 5176 | /* Host driver will handle XGMI hive reset for SRIOV */ |
26bc5340 AG |
5177 | if (amdgpu_sriov_vf(adev)) { |
5178 | r = amdgpu_device_reset_sriov(adev, job ? false : true); | |
5179 | if (r) | |
5180 | adev->asic_reset_res = r; | |
5181 | } else { | |
04442bf7 | 5182 | r = amdgpu_do_asic_reset(device_list_handle, &reset_context); |
26bc5340 AG |
5183 | if (r && r == -EAGAIN) |
5184 | goto retry; | |
5185 | } | |
5186 | ||
1d721ed6 AG |
5187 | skip_hw_reset: |
5188 | ||
26bc5340 | 5189 | /* Post ASIC reset for all devs .*/ |
655ce9cb | 5190 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
7c6e68c7 | 5191 | |
e6c6338f JZ |
5192 | /* |
5193 | * Sometimes a later bad compute job can block a good gfx job as gfx | |
5194 | * and compute ring share internal GC HW mutually. We add an additional | |
5195 | * guilty jobs recheck step to find the real guilty job, it synchronously | |
5196 | * submits and pends for the first job being signaled. If it gets timeout, | |
5197 | * we identify it as a real guilty job. | |
5198 | */ | |
5199 | if (amdgpu_gpu_recovery == 2 && | |
5200 | !(tmp_vram_lost_counter < atomic_read(&adev->vram_lost_counter))) | |
04442bf7 LL |
5201 | amdgpu_device_recheck_guilty_jobs( |
5202 | tmp_adev, device_list_handle, &reset_context); | |
e6c6338f | 5203 | |
1d721ed6 AG |
5204 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { |
5205 | struct amdgpu_ring *ring = tmp_adev->rings[i]; | |
5206 | ||
5207 | if (!ring || !ring->sched.thread) | |
5208 | continue; | |
5209 | ||
5210 | /* No point to resubmit jobs if we didn't HW reset*/ | |
5211 | if (!tmp_adev->asic_reset_res && !job_signaled) | |
5212 | drm_sched_resubmit_jobs(&ring->sched); | |
5213 | ||
5214 | drm_sched_start(&ring->sched, !tmp_adev->asic_reset_res); | |
5215 | } | |
5216 | ||
1053b9c9 | 5217 | if (!drm_drv_uses_atomic_modeset(adev_to_drm(tmp_adev)) && !job_signaled) { |
4a580877 | 5218 | drm_helper_resume_force_mode(adev_to_drm(tmp_adev)); |
1d721ed6 AG |
5219 | } |
5220 | ||
7258fa31 SK |
5221 | if (tmp_adev->asic_reset_res) |
5222 | r = tmp_adev->asic_reset_res; | |
5223 | ||
1d721ed6 | 5224 | tmp_adev->asic_reset_res = 0; |
26bc5340 AG |
5225 | |
5226 | if (r) { | |
5227 | /* bad news, how to tell it to userspace ? */ | |
12ffa55d | 5228 | dev_info(tmp_adev->dev, "GPU reset(%d) failed\n", atomic_read(&tmp_adev->gpu_reset_counter)); |
26bc5340 AG |
5229 | amdgpu_vf_error_put(tmp_adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r); |
5230 | } else { | |
12ffa55d | 5231 | dev_info(tmp_adev->dev, "GPU reset(%d) succeeded!\n", atomic_read(&tmp_adev->gpu_reset_counter)); |
3fa8f89d S |
5232 | if (amdgpu_acpi_smart_shift_update(adev_to_drm(tmp_adev), AMDGPU_SS_DEV_D0)) |
5233 | DRM_WARN("smart shift update failed\n"); | |
26bc5340 | 5234 | } |
7c6e68c7 | 5235 | } |
26bc5340 | 5236 | |
7c6e68c7 | 5237 | skip_sched_resume: |
655ce9cb | 5238 | list_for_each_entry(tmp_adev, device_list_handle, reset_list) { |
428890a3 | 5239 | /* unlock kfd: SRIOV would do it separately */ |
5240 | if (!need_emergency_restart && !amdgpu_sriov_vf(tmp_adev)) | |
5241 | amdgpu_amdkfd_post_reset(tmp_adev); | |
8e2712e7 | 5242 | |
5243 | /* kfd_post_reset will do nothing if kfd device is not initialized, | |
5244 | * need to bring up kfd here if it's not be initialized before | |
5245 | */ | |
5246 | if (!adev->kfd.init_complete) | |
5247 | amdgpu_amdkfd_device_init(adev); | |
5248 | ||
3f12acc8 EQ |
5249 | if (audio_suspended) |
5250 | amdgpu_device_resume_display_audio(tmp_adev); | |
26bc5340 AG |
5251 | amdgpu_device_unlock_adev(tmp_adev); |
5252 | } | |
5253 | ||
cbfd17f7 | 5254 | skip_recovery: |
9e94d22c | 5255 | if (hive) { |
53b3f8f4 | 5256 | atomic_set(&hive->in_reset, 0); |
9e94d22c | 5257 | mutex_unlock(&hive->hive_lock); |
d95e8e97 | 5258 | amdgpu_put_xgmi_hive(hive); |
9e94d22c | 5259 | } |
26bc5340 | 5260 | |
91fb309d | 5261 | if (r && r != -EAGAIN) |
26bc5340 | 5262 | dev_info(adev->dev, "GPU reset end with ret = %d\n", r); |
d38ceaf9 AD |
5263 | return r; |
5264 | } | |
5265 | ||
e3ecdffa AD |
5266 | /** |
5267 | * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot | |
5268 | * | |
5269 | * @adev: amdgpu_device pointer | |
5270 | * | |
5271 | * Fetchs and stores in the driver the PCIE capabilities (gen speed | |
5272 | * and lanes) of the slot the device is in. Handles APUs and | |
5273 | * virtualized environments where PCIE config space may not be available. | |
5274 | */ | |
5494d864 | 5275 | static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev) |
d0dd7f0c | 5276 | { |
5d9a6330 | 5277 | struct pci_dev *pdev; |
c5313457 HK |
5278 | enum pci_bus_speed speed_cap, platform_speed_cap; |
5279 | enum pcie_link_width platform_link_width; | |
d0dd7f0c | 5280 | |
cd474ba0 AD |
5281 | if (amdgpu_pcie_gen_cap) |
5282 | adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap; | |
d0dd7f0c | 5283 | |
cd474ba0 AD |
5284 | if (amdgpu_pcie_lane_cap) |
5285 | adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap; | |
d0dd7f0c | 5286 | |
cd474ba0 AD |
5287 | /* covers APUs as well */ |
5288 | if (pci_is_root_bus(adev->pdev->bus)) { | |
5289 | if (adev->pm.pcie_gen_mask == 0) | |
5290 | adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK; | |
5291 | if (adev->pm.pcie_mlw_mask == 0) | |
5292 | adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK; | |
d0dd7f0c | 5293 | return; |
cd474ba0 | 5294 | } |
d0dd7f0c | 5295 | |
c5313457 HK |
5296 | if (adev->pm.pcie_gen_mask && adev->pm.pcie_mlw_mask) |
5297 | return; | |
5298 | ||
dbaa922b AD |
5299 | pcie_bandwidth_available(adev->pdev, NULL, |
5300 | &platform_speed_cap, &platform_link_width); | |
c5313457 | 5301 | |
cd474ba0 | 5302 | if (adev->pm.pcie_gen_mask == 0) { |
5d9a6330 AD |
5303 | /* asic caps */ |
5304 | pdev = adev->pdev; | |
5305 | speed_cap = pcie_get_speed_cap(pdev); | |
5306 | if (speed_cap == PCI_SPEED_UNKNOWN) { | |
5307 | adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 | | |
cd474ba0 AD |
5308 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 | |
5309 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3); | |
cd474ba0 | 5310 | } else { |
2b3a1f51 FX |
5311 | if (speed_cap == PCIE_SPEED_32_0GT) |
5312 | adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 | | |
5313 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 | | |
5314 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 | | |
5315 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4 | | |
5316 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN5); | |
5317 | else if (speed_cap == PCIE_SPEED_16_0GT) | |
5d9a6330 AD |
5318 | adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 | |
5319 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 | | |
5320 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 | | |
5321 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4); | |
5322 | else if (speed_cap == PCIE_SPEED_8_0GT) | |
5323 | adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 | | |
5324 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 | | |
5325 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3); | |
5326 | else if (speed_cap == PCIE_SPEED_5_0GT) | |
5327 | adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 | | |
5328 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2); | |
5329 | else | |
5330 | adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1; | |
5331 | } | |
5332 | /* platform caps */ | |
c5313457 | 5333 | if (platform_speed_cap == PCI_SPEED_UNKNOWN) { |
5d9a6330 AD |
5334 | adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 | |
5335 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2); | |
5336 | } else { | |
2b3a1f51 FX |
5337 | if (platform_speed_cap == PCIE_SPEED_32_0GT) |
5338 | adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 | | |
5339 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 | | |
5340 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 | | |
5341 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4 | | |
5342 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN5); | |
5343 | else if (platform_speed_cap == PCIE_SPEED_16_0GT) | |
5d9a6330 AD |
5344 | adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 | |
5345 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 | | |
5346 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 | | |
5347 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4); | |
c5313457 | 5348 | else if (platform_speed_cap == PCIE_SPEED_8_0GT) |
5d9a6330 AD |
5349 | adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 | |
5350 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 | | |
5351 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3); | |
c5313457 | 5352 | else if (platform_speed_cap == PCIE_SPEED_5_0GT) |
5d9a6330 AD |
5353 | adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 | |
5354 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2); | |
5355 | else | |
5356 | adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1; | |
5357 | ||
cd474ba0 AD |
5358 | } |
5359 | } | |
5360 | if (adev->pm.pcie_mlw_mask == 0) { | |
c5313457 | 5361 | if (platform_link_width == PCIE_LNK_WIDTH_UNKNOWN) { |
5d9a6330 AD |
5362 | adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK; |
5363 | } else { | |
c5313457 | 5364 | switch (platform_link_width) { |
5d9a6330 | 5365 | case PCIE_LNK_X32: |
cd474ba0 AD |
5366 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 | |
5367 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 | | |
5368 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 | | |
5369 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 | | |
5370 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | | |
5371 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | | |
5372 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); | |
5373 | break; | |
5d9a6330 | 5374 | case PCIE_LNK_X16: |
cd474ba0 AD |
5375 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 | |
5376 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 | | |
5377 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 | | |
5378 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | | |
5379 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | | |
5380 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); | |
5381 | break; | |
5d9a6330 | 5382 | case PCIE_LNK_X12: |
cd474ba0 AD |
5383 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 | |
5384 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 | | |
5385 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | | |
5386 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | | |
5387 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); | |
5388 | break; | |
5d9a6330 | 5389 | case PCIE_LNK_X8: |
cd474ba0 AD |
5390 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 | |
5391 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | | |
5392 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | | |
5393 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); | |
5394 | break; | |
5d9a6330 | 5395 | case PCIE_LNK_X4: |
cd474ba0 AD |
5396 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | |
5397 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | | |
5398 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); | |
5399 | break; | |
5d9a6330 | 5400 | case PCIE_LNK_X2: |
cd474ba0 AD |
5401 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | |
5402 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); | |
5403 | break; | |
5d9a6330 | 5404 | case PCIE_LNK_X1: |
cd474ba0 AD |
5405 | adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1; |
5406 | break; | |
5407 | default: | |
5408 | break; | |
5409 | } | |
d0dd7f0c AD |
5410 | } |
5411 | } | |
5412 | } | |
d38ceaf9 | 5413 | |
361dbd01 AD |
5414 | int amdgpu_device_baco_enter(struct drm_device *dev) |
5415 | { | |
1348969a | 5416 | struct amdgpu_device *adev = drm_to_adev(dev); |
7a22677b | 5417 | struct amdgpu_ras *ras = amdgpu_ras_get_context(adev); |
361dbd01 | 5418 | |
4a580877 | 5419 | if (!amdgpu_device_supports_baco(adev_to_drm(adev))) |
361dbd01 AD |
5420 | return -ENOTSUPP; |
5421 | ||
8ab0d6f0 | 5422 | if (ras && adev->ras_enabled && |
acdae216 | 5423 | adev->nbio.funcs->enable_doorbell_interrupt) |
7a22677b LM |
5424 | adev->nbio.funcs->enable_doorbell_interrupt(adev, false); |
5425 | ||
9530273e | 5426 | return amdgpu_dpm_baco_enter(adev); |
361dbd01 AD |
5427 | } |
5428 | ||
5429 | int amdgpu_device_baco_exit(struct drm_device *dev) | |
5430 | { | |
1348969a | 5431 | struct amdgpu_device *adev = drm_to_adev(dev); |
7a22677b | 5432 | struct amdgpu_ras *ras = amdgpu_ras_get_context(adev); |
9530273e | 5433 | int ret = 0; |
361dbd01 | 5434 | |
4a580877 | 5435 | if (!amdgpu_device_supports_baco(adev_to_drm(adev))) |
361dbd01 AD |
5436 | return -ENOTSUPP; |
5437 | ||
9530273e EQ |
5438 | ret = amdgpu_dpm_baco_exit(adev); |
5439 | if (ret) | |
5440 | return ret; | |
7a22677b | 5441 | |
8ab0d6f0 | 5442 | if (ras && adev->ras_enabled && |
acdae216 | 5443 | adev->nbio.funcs->enable_doorbell_interrupt) |
7a22677b LM |
5444 | adev->nbio.funcs->enable_doorbell_interrupt(adev, true); |
5445 | ||
1bece222 CL |
5446 | if (amdgpu_passthrough(adev) && |
5447 | adev->nbio.funcs->clear_doorbell_interrupt) | |
5448 | adev->nbio.funcs->clear_doorbell_interrupt(adev); | |
5449 | ||
7a22677b | 5450 | return 0; |
361dbd01 | 5451 | } |
c9a6b82f | 5452 | |
acd89fca AG |
5453 | static void amdgpu_cancel_all_tdr(struct amdgpu_device *adev) |
5454 | { | |
5455 | int i; | |
5456 | ||
5457 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { | |
5458 | struct amdgpu_ring *ring = adev->rings[i]; | |
5459 | ||
5460 | if (!ring || !ring->sched.thread) | |
5461 | continue; | |
5462 | ||
5463 | cancel_delayed_work_sync(&ring->sched.work_tdr); | |
5464 | } | |
5465 | } | |
5466 | ||
c9a6b82f AG |
5467 | /** |
5468 | * amdgpu_pci_error_detected - Called when a PCI error is detected. | |
5469 | * @pdev: PCI device struct | |
5470 | * @state: PCI channel state | |
5471 | * | |
5472 | * Description: Called when a PCI error is detected. | |
5473 | * | |
5474 | * Return: PCI_ERS_RESULT_NEED_RESET or PCI_ERS_RESULT_DISCONNECT. | |
5475 | */ | |
5476 | pci_ers_result_t amdgpu_pci_error_detected(struct pci_dev *pdev, pci_channel_state_t state) | |
5477 | { | |
5478 | struct drm_device *dev = pci_get_drvdata(pdev); | |
5479 | struct amdgpu_device *adev = drm_to_adev(dev); | |
acd89fca | 5480 | int i; |
c9a6b82f AG |
5481 | |
5482 | DRM_INFO("PCI error: detected callback, state(%d)!!\n", state); | |
5483 | ||
6894305c AG |
5484 | if (adev->gmc.xgmi.num_physical_nodes > 1) { |
5485 | DRM_WARN("No support for XGMI hive yet..."); | |
5486 | return PCI_ERS_RESULT_DISCONNECT; | |
5487 | } | |
5488 | ||
e17e27f9 GC |
5489 | adev->pci_channel_state = state; |
5490 | ||
c9a6b82f AG |
5491 | switch (state) { |
5492 | case pci_channel_io_normal: | |
5493 | return PCI_ERS_RESULT_CAN_RECOVER; | |
acd89fca | 5494 | /* Fatal error, prepare for slot reset */ |
8a11d283 TZ |
5495 | case pci_channel_io_frozen: |
5496 | /* | |
acd89fca AG |
5497 | * Cancel and wait for all TDRs in progress if failing to |
5498 | * set adev->in_gpu_reset in amdgpu_device_lock_adev | |
5499 | * | |
5500 | * Locking adev->reset_sem will prevent any external access | |
5501 | * to GPU during PCI error recovery | |
5502 | */ | |
5503 | while (!amdgpu_device_lock_adev(adev, NULL)) | |
5504 | amdgpu_cancel_all_tdr(adev); | |
5505 | ||
5506 | /* | |
5507 | * Block any work scheduling as we do for regular GPU reset | |
5508 | * for the duration of the recovery | |
5509 | */ | |
5510 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { | |
5511 | struct amdgpu_ring *ring = adev->rings[i]; | |
5512 | ||
5513 | if (!ring || !ring->sched.thread) | |
5514 | continue; | |
5515 | ||
5516 | drm_sched_stop(&ring->sched, NULL); | |
5517 | } | |
8f8c80f4 | 5518 | atomic_inc(&adev->gpu_reset_counter); |
c9a6b82f AG |
5519 | return PCI_ERS_RESULT_NEED_RESET; |
5520 | case pci_channel_io_perm_failure: | |
5521 | /* Permanent error, prepare for device removal */ | |
5522 | return PCI_ERS_RESULT_DISCONNECT; | |
5523 | } | |
5524 | ||
5525 | return PCI_ERS_RESULT_NEED_RESET; | |
5526 | } | |
5527 | ||
5528 | /** | |
5529 | * amdgpu_pci_mmio_enabled - Enable MMIO and dump debug registers | |
5530 | * @pdev: pointer to PCI device | |
5531 | */ | |
5532 | pci_ers_result_t amdgpu_pci_mmio_enabled(struct pci_dev *pdev) | |
5533 | { | |
5534 | ||
5535 | DRM_INFO("PCI error: mmio enabled callback!!\n"); | |
5536 | ||
5537 | /* TODO - dump whatever for debugging purposes */ | |
5538 | ||
5539 | /* This called only if amdgpu_pci_error_detected returns | |
5540 | * PCI_ERS_RESULT_CAN_RECOVER. Read/write to the device still | |
5541 | * works, no need to reset slot. | |
5542 | */ | |
5543 | ||
5544 | return PCI_ERS_RESULT_RECOVERED; | |
5545 | } | |
5546 | ||
5547 | /** | |
5548 | * amdgpu_pci_slot_reset - Called when PCI slot has been reset. | |
5549 | * @pdev: PCI device struct | |
5550 | * | |
5551 | * Description: This routine is called by the pci error recovery | |
5552 | * code after the PCI slot has been reset, just before we | |
5553 | * should resume normal operations. | |
5554 | */ | |
5555 | pci_ers_result_t amdgpu_pci_slot_reset(struct pci_dev *pdev) | |
5556 | { | |
5557 | struct drm_device *dev = pci_get_drvdata(pdev); | |
5558 | struct amdgpu_device *adev = drm_to_adev(dev); | |
362c7b91 | 5559 | int r, i; |
04442bf7 | 5560 | struct amdgpu_reset_context reset_context; |
362c7b91 | 5561 | u32 memsize; |
7ac71382 | 5562 | struct list_head device_list; |
c9a6b82f AG |
5563 | |
5564 | DRM_INFO("PCI error: slot reset callback!!\n"); | |
5565 | ||
04442bf7 LL |
5566 | memset(&reset_context, 0, sizeof(reset_context)); |
5567 | ||
7ac71382 | 5568 | INIT_LIST_HEAD(&device_list); |
655ce9cb | 5569 | list_add_tail(&adev->reset_list, &device_list); |
7ac71382 | 5570 | |
362c7b91 AG |
5571 | /* wait for asic to come out of reset */ |
5572 | msleep(500); | |
5573 | ||
7ac71382 | 5574 | /* Restore PCI confspace */ |
c1dd4aa6 | 5575 | amdgpu_device_load_pci_state(pdev); |
c9a6b82f | 5576 | |
362c7b91 AG |
5577 | /* confirm ASIC came out of reset */ |
5578 | for (i = 0; i < adev->usec_timeout; i++) { | |
5579 | memsize = amdgpu_asic_get_config_memsize(adev); | |
5580 | ||
5581 | if (memsize != 0xffffffff) | |
5582 | break; | |
5583 | udelay(1); | |
5584 | } | |
5585 | if (memsize == 0xffffffff) { | |
5586 | r = -ETIME; | |
5587 | goto out; | |
5588 | } | |
5589 | ||
04442bf7 LL |
5590 | reset_context.method = AMD_RESET_METHOD_NONE; |
5591 | reset_context.reset_req_dev = adev; | |
5592 | set_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags); | |
5593 | set_bit(AMDGPU_SKIP_HW_RESET, &reset_context.flags); | |
5594 | ||
7afefb81 | 5595 | adev->no_hw_access = true; |
04442bf7 | 5596 | r = amdgpu_device_pre_asic_reset(adev, &reset_context); |
7afefb81 | 5597 | adev->no_hw_access = false; |
c9a6b82f AG |
5598 | if (r) |
5599 | goto out; | |
5600 | ||
04442bf7 | 5601 | r = amdgpu_do_asic_reset(&device_list, &reset_context); |
c9a6b82f AG |
5602 | |
5603 | out: | |
c9a6b82f | 5604 | if (!r) { |
c1dd4aa6 AG |
5605 | if (amdgpu_device_cache_pci_state(adev->pdev)) |
5606 | pci_restore_state(adev->pdev); | |
5607 | ||
c9a6b82f AG |
5608 | DRM_INFO("PCIe error recovery succeeded\n"); |
5609 | } else { | |
5610 | DRM_ERROR("PCIe error recovery failed, err:%d", r); | |
5611 | amdgpu_device_unlock_adev(adev); | |
5612 | } | |
5613 | ||
5614 | return r ? PCI_ERS_RESULT_DISCONNECT : PCI_ERS_RESULT_RECOVERED; | |
5615 | } | |
5616 | ||
5617 | /** | |
5618 | * amdgpu_pci_resume() - resume normal ops after PCI reset | |
5619 | * @pdev: pointer to PCI device | |
5620 | * | |
5621 | * Called when the error recovery driver tells us that its | |
505199a3 | 5622 | * OK to resume normal operation. |
c9a6b82f AG |
5623 | */ |
5624 | void amdgpu_pci_resume(struct pci_dev *pdev) | |
5625 | { | |
5626 | struct drm_device *dev = pci_get_drvdata(pdev); | |
5627 | struct amdgpu_device *adev = drm_to_adev(dev); | |
acd89fca | 5628 | int i; |
c9a6b82f | 5629 | |
c9a6b82f AG |
5630 | |
5631 | DRM_INFO("PCI error: resume callback!!\n"); | |
acd89fca | 5632 | |
e17e27f9 GC |
5633 | /* Only continue execution for the case of pci_channel_io_frozen */ |
5634 | if (adev->pci_channel_state != pci_channel_io_frozen) | |
5635 | return; | |
5636 | ||
acd89fca AG |
5637 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { |
5638 | struct amdgpu_ring *ring = adev->rings[i]; | |
5639 | ||
5640 | if (!ring || !ring->sched.thread) | |
5641 | continue; | |
5642 | ||
5643 | ||
5644 | drm_sched_resubmit_jobs(&ring->sched); | |
5645 | drm_sched_start(&ring->sched, true); | |
5646 | } | |
5647 | ||
5648 | amdgpu_device_unlock_adev(adev); | |
c9a6b82f | 5649 | } |
c1dd4aa6 AG |
5650 | |
5651 | bool amdgpu_device_cache_pci_state(struct pci_dev *pdev) | |
5652 | { | |
5653 | struct drm_device *dev = pci_get_drvdata(pdev); | |
5654 | struct amdgpu_device *adev = drm_to_adev(dev); | |
5655 | int r; | |
5656 | ||
5657 | r = pci_save_state(pdev); | |
5658 | if (!r) { | |
5659 | kfree(adev->pci_state); | |
5660 | ||
5661 | adev->pci_state = pci_store_saved_state(pdev); | |
5662 | ||
5663 | if (!adev->pci_state) { | |
5664 | DRM_ERROR("Failed to store PCI saved state"); | |
5665 | return false; | |
5666 | } | |
5667 | } else { | |
5668 | DRM_WARN("Failed to save PCI state, err:%d\n", r); | |
5669 | return false; | |
5670 | } | |
5671 | ||
5672 | return true; | |
5673 | } | |
5674 | ||
5675 | bool amdgpu_device_load_pci_state(struct pci_dev *pdev) | |
5676 | { | |
5677 | struct drm_device *dev = pci_get_drvdata(pdev); | |
5678 | struct amdgpu_device *adev = drm_to_adev(dev); | |
5679 | int r; | |
5680 | ||
5681 | if (!adev->pci_state) | |
5682 | return false; | |
5683 | ||
5684 | r = pci_load_saved_state(pdev, adev->pci_state); | |
5685 | ||
5686 | if (!r) { | |
5687 | pci_restore_state(pdev); | |
5688 | } else { | |
5689 | DRM_WARN("Failed to load PCI state, err:%d\n", r); | |
5690 | return false; | |
5691 | } | |
5692 | ||
5693 | return true; | |
5694 | } | |
5695 | ||
810085dd EH |
5696 | void amdgpu_device_flush_hdp(struct amdgpu_device *adev, |
5697 | struct amdgpu_ring *ring) | |
5698 | { | |
5699 | #ifdef CONFIG_X86_64 | |
5700 | if (adev->flags & AMD_IS_APU) | |
5701 | return; | |
5702 | #endif | |
5703 | if (adev->gmc.xgmi.connected_to_cpu) | |
5704 | return; | |
5705 | ||
5706 | if (ring && ring->funcs->emit_hdp_flush) | |
5707 | amdgpu_ring_emit_hdp_flush(ring); | |
5708 | else | |
5709 | amdgpu_asic_flush_hdp(adev, ring); | |
5710 | } | |
c1dd4aa6 | 5711 | |
810085dd EH |
5712 | void amdgpu_device_invalidate_hdp(struct amdgpu_device *adev, |
5713 | struct amdgpu_ring *ring) | |
5714 | { | |
5715 | #ifdef CONFIG_X86_64 | |
5716 | if (adev->flags & AMD_IS_APU) | |
5717 | return; | |
5718 | #endif | |
5719 | if (adev->gmc.xgmi.connected_to_cpu) | |
5720 | return; | |
c1dd4aa6 | 5721 | |
810085dd EH |
5722 | amdgpu_asic_invalidate_hdp(adev, ring); |
5723 | } | |
34f3a4a9 LY |
5724 | |
5725 | /** | |
5726 | * amdgpu_device_halt() - bring hardware to some kind of halt state | |
5727 | * | |
5728 | * @adev: amdgpu_device pointer | |
5729 | * | |
5730 | * Bring hardware to some kind of halt state so that no one can touch it | |
5731 | * any more. It will help to maintain error context when error occurred. | |
5732 | * Compare to a simple hang, the system will keep stable at least for SSH | |
5733 | * access. Then it should be trivial to inspect the hardware state and | |
5734 | * see what's going on. Implemented as following: | |
5735 | * | |
5736 | * 1. drm_dev_unplug() makes device inaccessible to user space(IOCTLs, etc), | |
5737 | * clears all CPU mappings to device, disallows remappings through page faults | |
5738 | * 2. amdgpu_irq_disable_all() disables all interrupts | |
5739 | * 3. amdgpu_fence_driver_hw_fini() signals all HW fences | |
5740 | * 4. set adev->no_hw_access to avoid potential crashes after setp 5 | |
5741 | * 5. amdgpu_device_unmap_mmio() clears all MMIO mappings | |
5742 | * 6. pci_disable_device() and pci_wait_for_pending_transaction() | |
5743 | * flush any in flight DMA operations | |
5744 | */ | |
5745 | void amdgpu_device_halt(struct amdgpu_device *adev) | |
5746 | { | |
5747 | struct pci_dev *pdev = adev->pdev; | |
e0f943b4 | 5748 | struct drm_device *ddev = adev_to_drm(adev); |
34f3a4a9 LY |
5749 | |
5750 | drm_dev_unplug(ddev); | |
5751 | ||
5752 | amdgpu_irq_disable_all(adev); | |
5753 | ||
5754 | amdgpu_fence_driver_hw_fini(adev); | |
5755 | ||
5756 | adev->no_hw_access = true; | |
5757 | ||
5758 | amdgpu_device_unmap_mmio(adev); | |
5759 | ||
5760 | pci_disable_device(pdev); | |
5761 | pci_wait_for_pending_transaction(pdev); | |
5762 | } | |
86700a40 XD |
5763 | |
5764 | u32 amdgpu_device_pcie_port_rreg(struct amdgpu_device *adev, | |
5765 | u32 reg) | |
5766 | { | |
5767 | unsigned long flags, address, data; | |
5768 | u32 r; | |
5769 | ||
5770 | address = adev->nbio.funcs->get_pcie_port_index_offset(adev); | |
5771 | data = adev->nbio.funcs->get_pcie_port_data_offset(adev); | |
5772 | ||
5773 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); | |
5774 | WREG32(address, reg * 4); | |
5775 | (void)RREG32(address); | |
5776 | r = RREG32(data); | |
5777 | spin_unlock_irqrestore(&adev->pcie_idx_lock, flags); | |
5778 | return r; | |
5779 | } | |
5780 | ||
5781 | void amdgpu_device_pcie_port_wreg(struct amdgpu_device *adev, | |
5782 | u32 reg, u32 v) | |
5783 | { | |
5784 | unsigned long flags, address, data; | |
5785 | ||
5786 | address = adev->nbio.funcs->get_pcie_port_index_offset(adev); | |
5787 | data = adev->nbio.funcs->get_pcie_port_data_offset(adev); | |
5788 | ||
5789 | spin_lock_irqsave(&adev->pcie_idx_lock, flags); | |
5790 | WREG32(address, reg * 4); | |
5791 | (void)RREG32(address); | |
5792 | WREG32(data, v); | |
5793 | (void)RREG32(data); | |
5794 | spin_unlock_irqrestore(&adev->pcie_idx_lock, flags); | |
5795 | } |