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> | |
fdf2f6c5 | 33 | |
4562236b | 34 | #include <drm/drm_atomic_helper.h> |
fcd70cd3 | 35 | #include <drm/drm_probe_helper.h> |
d38ceaf9 AD |
36 | #include <drm/amdgpu_drm.h> |
37 | #include <linux/vgaarb.h> | |
38 | #include <linux/vga_switcheroo.h> | |
39 | #include <linux/efi.h> | |
40 | #include "amdgpu.h" | |
f4b373f4 | 41 | #include "amdgpu_trace.h" |
d38ceaf9 AD |
42 | #include "amdgpu_i2c.h" |
43 | #include "atom.h" | |
44 | #include "amdgpu_atombios.h" | |
a5bde2f9 | 45 | #include "amdgpu_atomfirmware.h" |
d0dd7f0c | 46 | #include "amd_pcie.h" |
33f34802 KW |
47 | #ifdef CONFIG_DRM_AMDGPU_SI |
48 | #include "si.h" | |
49 | #endif | |
a2e73f56 AD |
50 | #ifdef CONFIG_DRM_AMDGPU_CIK |
51 | #include "cik.h" | |
52 | #endif | |
aaa36a97 | 53 | #include "vi.h" |
460826e6 | 54 | #include "soc15.h" |
0a5b8c7b | 55 | #include "nv.h" |
d38ceaf9 | 56 | #include "bif/bif_4_1_d.h" |
9accf2fd | 57 | #include <linux/pci.h> |
bec86378 | 58 | #include <linux/firmware.h> |
89041940 | 59 | #include "amdgpu_vf_error.h" |
d38ceaf9 | 60 | |
ba997709 | 61 | #include "amdgpu_amdkfd.h" |
d2f52ac8 | 62 | #include "amdgpu_pm.h" |
d38ceaf9 | 63 | |
5183411b | 64 | #include "amdgpu_xgmi.h" |
c030f2e4 | 65 | #include "amdgpu_ras.h" |
9c7c85f7 | 66 | #include "amdgpu_pmu.h" |
5183411b | 67 | |
d5ea093e | 68 | #include <linux/suspend.h> |
c6a6e2db | 69 | #include <drm/task_barrier.h> |
d5ea093e | 70 | |
e2a75f88 | 71 | MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin"); |
3f76dced | 72 | MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin"); |
2d2e5e7e | 73 | MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin"); |
ad5a67a7 | 74 | MODULE_FIRMWARE("amdgpu/picasso_gpu_info.bin"); |
54c4d17e | 75 | MODULE_FIRMWARE("amdgpu/raven2_gpu_info.bin"); |
65e60f6e | 76 | MODULE_FIRMWARE("amdgpu/arcturus_gpu_info.bin"); |
b51a26a0 | 77 | MODULE_FIRMWARE("amdgpu/renoir_gpu_info.bin"); |
23c6268e | 78 | MODULE_FIRMWARE("amdgpu/navi10_gpu_info.bin"); |
ed42cfe1 | 79 | MODULE_FIRMWARE("amdgpu/navi14_gpu_info.bin"); |
42b325e5 | 80 | MODULE_FIRMWARE("amdgpu/navi12_gpu_info.bin"); |
e2a75f88 | 81 | |
2dc80b00 S |
82 | #define AMDGPU_RESUME_MS 2000 |
83 | ||
050091ab | 84 | const char *amdgpu_asic_name[] = { |
da69c161 KW |
85 | "TAHITI", |
86 | "PITCAIRN", | |
87 | "VERDE", | |
88 | "OLAND", | |
89 | "HAINAN", | |
d38ceaf9 AD |
90 | "BONAIRE", |
91 | "KAVERI", | |
92 | "KABINI", | |
93 | "HAWAII", | |
94 | "MULLINS", | |
95 | "TOPAZ", | |
96 | "TONGA", | |
48299f95 | 97 | "FIJI", |
d38ceaf9 | 98 | "CARRIZO", |
139f4917 | 99 | "STONEY", |
2cc0c0b5 FC |
100 | "POLARIS10", |
101 | "POLARIS11", | |
c4642a47 | 102 | "POLARIS12", |
48ff108d | 103 | "VEGAM", |
d4196f01 | 104 | "VEGA10", |
8fab806a | 105 | "VEGA12", |
956fcddc | 106 | "VEGA20", |
2ca8a5d2 | 107 | "RAVEN", |
d6c3b24e | 108 | "ARCTURUS", |
1eee4228 | 109 | "RENOIR", |
852a6626 | 110 | "NAVI10", |
87dbad02 | 111 | "NAVI14", |
9802f5d7 | 112 | "NAVI12", |
d38ceaf9 AD |
113 | "LAST", |
114 | }; | |
115 | ||
dcea6e65 KR |
116 | /** |
117 | * DOC: pcie_replay_count | |
118 | * | |
119 | * The amdgpu driver provides a sysfs API for reporting the total number | |
120 | * of PCIe replays (NAKs) | |
121 | * The file pcie_replay_count is used for this and returns the total | |
122 | * number of replays as a sum of the NAKs generated and NAKs received | |
123 | */ | |
124 | ||
125 | static ssize_t amdgpu_device_get_pcie_replay_count(struct device *dev, | |
126 | struct device_attribute *attr, char *buf) | |
127 | { | |
128 | struct drm_device *ddev = dev_get_drvdata(dev); | |
129 | struct amdgpu_device *adev = ddev->dev_private; | |
130 | uint64_t cnt = amdgpu_asic_get_pcie_replay_count(adev); | |
131 | ||
132 | return snprintf(buf, PAGE_SIZE, "%llu\n", cnt); | |
133 | } | |
134 | ||
135 | static DEVICE_ATTR(pcie_replay_count, S_IRUGO, | |
136 | amdgpu_device_get_pcie_replay_count, NULL); | |
137 | ||
5494d864 AD |
138 | static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev); |
139 | ||
e3ecdffa | 140 | /** |
31af062a | 141 | * amdgpu_device_supports_boco - Is the device a dGPU with HG/PX power control |
e3ecdffa AD |
142 | * |
143 | * @dev: drm_device pointer | |
144 | * | |
145 | * Returns true if the device is a dGPU with HG/PX power control, | |
146 | * otherwise return false. | |
147 | */ | |
31af062a | 148 | bool amdgpu_device_supports_boco(struct drm_device *dev) |
d38ceaf9 AD |
149 | { |
150 | struct amdgpu_device *adev = dev->dev_private; | |
151 | ||
2f7d10b3 | 152 | if (adev->flags & AMD_IS_PX) |
d38ceaf9 AD |
153 | return true; |
154 | return false; | |
155 | } | |
156 | ||
a69cba42 AD |
157 | /** |
158 | * amdgpu_device_supports_baco - Does the device support BACO | |
159 | * | |
160 | * @dev: drm_device pointer | |
161 | * | |
162 | * Returns true if the device supporte BACO, | |
163 | * otherwise return false. | |
164 | */ | |
165 | bool amdgpu_device_supports_baco(struct drm_device *dev) | |
166 | { | |
167 | struct amdgpu_device *adev = dev->dev_private; | |
168 | ||
169 | return amdgpu_asic_supports_baco(adev); | |
170 | } | |
171 | ||
e35e2b11 TY |
172 | /** |
173 | * VRAM access helper functions. | |
174 | * | |
175 | * amdgpu_device_vram_access - read/write a buffer in vram | |
176 | * | |
177 | * @adev: amdgpu_device pointer | |
178 | * @pos: offset of the buffer in vram | |
179 | * @buf: virtual address of the buffer in system memory | |
180 | * @size: read/write size, sizeof(@buf) must > @size | |
181 | * @write: true - write to vram, otherwise - read from vram | |
182 | */ | |
183 | void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos, | |
184 | uint32_t *buf, size_t size, bool write) | |
185 | { | |
e35e2b11 | 186 | unsigned long flags; |
ce05ac56 CK |
187 | uint32_t hi = ~0; |
188 | uint64_t last; | |
189 | ||
c12b84d6 CK |
190 | |
191 | #ifdef CONFIG_64BIT | |
192 | last = min(pos + size, adev->gmc.visible_vram_size); | |
193 | if (last > pos) { | |
194 | void __iomem *addr = adev->mman.aper_base_kaddr + pos; | |
195 | size_t count = last - pos; | |
196 | ||
197 | if (write) { | |
198 | memcpy_toio(addr, buf, count); | |
199 | mb(); | |
200 | amdgpu_asic_flush_hdp(adev, NULL); | |
201 | } else { | |
202 | amdgpu_asic_invalidate_hdp(adev, NULL); | |
203 | mb(); | |
204 | memcpy_fromio(buf, addr, count); | |
205 | } | |
206 | ||
207 | if (count == size) | |
208 | return; | |
209 | ||
210 | pos += count; | |
211 | buf += count / 4; | |
212 | size -= count; | |
213 | } | |
214 | #endif | |
215 | ||
ce05ac56 CK |
216 | spin_lock_irqsave(&adev->mmio_idx_lock, flags); |
217 | for (last = pos + size; pos < last; pos += 4) { | |
218 | uint32_t tmp = pos >> 31; | |
e35e2b11 | 219 | |
e35e2b11 | 220 | WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)pos) | 0x80000000); |
ce05ac56 CK |
221 | if (tmp != hi) { |
222 | WREG32_NO_KIQ(mmMM_INDEX_HI, tmp); | |
223 | hi = tmp; | |
224 | } | |
e35e2b11 TY |
225 | if (write) |
226 | WREG32_NO_KIQ(mmMM_DATA, *buf++); | |
227 | else | |
228 | *buf++ = RREG32_NO_KIQ(mmMM_DATA); | |
e35e2b11 | 229 | } |
ce05ac56 | 230 | spin_unlock_irqrestore(&adev->mmio_idx_lock, flags); |
e35e2b11 TY |
231 | } |
232 | ||
d38ceaf9 AD |
233 | /* |
234 | * MMIO register access helper functions. | |
235 | */ | |
e3ecdffa AD |
236 | /** |
237 | * amdgpu_mm_rreg - read a memory mapped IO register | |
238 | * | |
239 | * @adev: amdgpu_device pointer | |
240 | * @reg: dword aligned register offset | |
241 | * @acc_flags: access flags which require special behavior | |
242 | * | |
243 | * Returns the 32 bit value from the offset specified. | |
244 | */ | |
d38ceaf9 | 245 | uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg, |
15d72fd7 | 246 | uint32_t acc_flags) |
d38ceaf9 | 247 | { |
f4b373f4 TSD |
248 | uint32_t ret; |
249 | ||
c68dbcd8 | 250 | if ((acc_flags & AMDGPU_REGS_KIQ) || (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))) |
d33a99c4 | 251 | return amdgpu_kiq_rreg(adev, reg); |
bc992ba5 | 252 | |
15d72fd7 | 253 | if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX)) |
f4b373f4 | 254 | ret = readl(((void __iomem *)adev->rmmio) + (reg * 4)); |
d38ceaf9 AD |
255 | else { |
256 | unsigned long flags; | |
d38ceaf9 AD |
257 | |
258 | spin_lock_irqsave(&adev->mmio_idx_lock, flags); | |
259 | writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4)); | |
260 | ret = readl(((void __iomem *)adev->rmmio) + (mmMM_DATA * 4)); | |
261 | spin_unlock_irqrestore(&adev->mmio_idx_lock, flags); | |
d38ceaf9 | 262 | } |
f4b373f4 TSD |
263 | trace_amdgpu_mm_rreg(adev->pdev->device, reg, ret); |
264 | return ret; | |
d38ceaf9 AD |
265 | } |
266 | ||
421a2a30 ML |
267 | /* |
268 | * MMIO register read with bytes helper functions | |
269 | * @offset:bytes offset from MMIO start | |
270 | * | |
271 | */ | |
272 | ||
e3ecdffa AD |
273 | /** |
274 | * amdgpu_mm_rreg8 - read a memory mapped IO register | |
275 | * | |
276 | * @adev: amdgpu_device pointer | |
277 | * @offset: byte aligned register offset | |
278 | * | |
279 | * Returns the 8 bit value from the offset specified. | |
280 | */ | |
421a2a30 ML |
281 | uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset) { |
282 | if (offset < adev->rmmio_size) | |
283 | return (readb(adev->rmmio + offset)); | |
284 | BUG(); | |
285 | } | |
286 | ||
287 | /* | |
288 | * MMIO register write with bytes helper functions | |
289 | * @offset:bytes offset from MMIO start | |
290 | * @value: the value want to be written to the register | |
291 | * | |
292 | */ | |
e3ecdffa AD |
293 | /** |
294 | * amdgpu_mm_wreg8 - read a memory mapped IO register | |
295 | * | |
296 | * @adev: amdgpu_device pointer | |
297 | * @offset: byte aligned register offset | |
298 | * @value: 8 bit value to write | |
299 | * | |
300 | * Writes the value specified to the offset specified. | |
301 | */ | |
421a2a30 ML |
302 | void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value) { |
303 | if (offset < adev->rmmio_size) | |
304 | writeb(value, adev->rmmio + offset); | |
305 | else | |
306 | BUG(); | |
307 | } | |
308 | ||
2e0cc4d4 ML |
309 | void static inline amdgpu_mm_wreg_mmio(struct amdgpu_device *adev, uint32_t reg, uint32_t v, uint32_t acc_flags) |
310 | { | |
311 | trace_amdgpu_mm_wreg(adev->pdev->device, reg, v); | |
312 | ||
313 | if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX)) | |
314 | writel(v, ((void __iomem *)adev->rmmio) + (reg * 4)); | |
315 | else { | |
316 | unsigned long flags; | |
317 | ||
318 | spin_lock_irqsave(&adev->mmio_idx_lock, flags); | |
319 | writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4)); | |
320 | writel(v, ((void __iomem *)adev->rmmio) + (mmMM_DATA * 4)); | |
321 | spin_unlock_irqrestore(&adev->mmio_idx_lock, flags); | |
322 | } | |
323 | ||
324 | if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) { | |
325 | udelay(500); | |
326 | } | |
327 | } | |
328 | ||
e3ecdffa AD |
329 | /** |
330 | * amdgpu_mm_wreg - write to a memory mapped IO register | |
331 | * | |
332 | * @adev: amdgpu_device pointer | |
333 | * @reg: dword aligned register offset | |
334 | * @v: 32 bit value to write to the register | |
335 | * @acc_flags: access flags which require special behavior | |
336 | * | |
337 | * Writes the value specified to the offset specified. | |
338 | */ | |
d38ceaf9 | 339 | void amdgpu_mm_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v, |
15d72fd7 | 340 | uint32_t acc_flags) |
d38ceaf9 | 341 | { |
47ed4e1c KW |
342 | if (adev->asic_type >= CHIP_VEGA10 && reg == 0) { |
343 | adev->last_mm_index = v; | |
344 | } | |
345 | ||
c68dbcd8 | 346 | if ((acc_flags & AMDGPU_REGS_KIQ) || (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))) |
d33a99c4 | 347 | return amdgpu_kiq_wreg(adev, reg, v); |
bc992ba5 | 348 | |
2e0cc4d4 ML |
349 | amdgpu_mm_wreg_mmio(adev, reg, v, acc_flags); |
350 | } | |
d38ceaf9 | 351 | |
2e0cc4d4 ML |
352 | /* |
353 | * amdgpu_mm_wreg_mmio_rlc - write register either with mmio or with RLC path if in range | |
354 | * | |
355 | * this function is invoked only the debugfs register access | |
356 | * */ | |
357 | void amdgpu_mm_wreg_mmio_rlc(struct amdgpu_device *adev, uint32_t reg, uint32_t v, | |
358 | uint32_t acc_flags) | |
359 | { | |
360 | if (amdgpu_sriov_fullaccess(adev) && | |
361 | adev->gfx.rlc.funcs && | |
362 | adev->gfx.rlc.funcs->is_rlcg_access_range) { | |
47ed4e1c | 363 | |
2e0cc4d4 ML |
364 | if (adev->gfx.rlc.funcs->is_rlcg_access_range(adev, reg)) |
365 | return adev->gfx.rlc.funcs->rlcg_wreg(adev, reg, v); | |
47ed4e1c | 366 | } |
2e0cc4d4 ML |
367 | |
368 | amdgpu_mm_wreg_mmio(adev, reg, v, acc_flags); | |
d38ceaf9 AD |
369 | } |
370 | ||
e3ecdffa AD |
371 | /** |
372 | * amdgpu_io_rreg - read an IO register | |
373 | * | |
374 | * @adev: amdgpu_device pointer | |
375 | * @reg: dword aligned register offset | |
376 | * | |
377 | * Returns the 32 bit value from the offset specified. | |
378 | */ | |
d38ceaf9 AD |
379 | u32 amdgpu_io_rreg(struct amdgpu_device *adev, u32 reg) |
380 | { | |
381 | if ((reg * 4) < adev->rio_mem_size) | |
382 | return ioread32(adev->rio_mem + (reg * 4)); | |
383 | else { | |
384 | iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4)); | |
385 | return ioread32(adev->rio_mem + (mmMM_DATA * 4)); | |
386 | } | |
387 | } | |
388 | ||
e3ecdffa AD |
389 | /** |
390 | * amdgpu_io_wreg - write to an IO register | |
391 | * | |
392 | * @adev: amdgpu_device pointer | |
393 | * @reg: dword aligned register offset | |
394 | * @v: 32 bit value to write to the register | |
395 | * | |
396 | * Writes the value specified to the offset specified. | |
397 | */ | |
d38ceaf9 AD |
398 | void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32 v) |
399 | { | |
47ed4e1c KW |
400 | if (adev->asic_type >= CHIP_VEGA10 && reg == 0) { |
401 | adev->last_mm_index = v; | |
402 | } | |
d38ceaf9 AD |
403 | |
404 | if ((reg * 4) < adev->rio_mem_size) | |
405 | iowrite32(v, adev->rio_mem + (reg * 4)); | |
406 | else { | |
407 | iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4)); | |
408 | iowrite32(v, adev->rio_mem + (mmMM_DATA * 4)); | |
409 | } | |
47ed4e1c KW |
410 | |
411 | if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) { | |
412 | udelay(500); | |
413 | } | |
d38ceaf9 AD |
414 | } |
415 | ||
416 | /** | |
417 | * amdgpu_mm_rdoorbell - read a doorbell dword | |
418 | * | |
419 | * @adev: amdgpu_device pointer | |
420 | * @index: doorbell index | |
421 | * | |
422 | * Returns the value in the doorbell aperture at the | |
423 | * requested doorbell index (CIK). | |
424 | */ | |
425 | u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index) | |
426 | { | |
427 | if (index < adev->doorbell.num_doorbells) { | |
428 | return readl(adev->doorbell.ptr + index); | |
429 | } else { | |
430 | DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index); | |
431 | return 0; | |
432 | } | |
433 | } | |
434 | ||
435 | /** | |
436 | * amdgpu_mm_wdoorbell - write a doorbell dword | |
437 | * | |
438 | * @adev: amdgpu_device pointer | |
439 | * @index: doorbell index | |
440 | * @v: value to write | |
441 | * | |
442 | * Writes @v to the doorbell aperture at the | |
443 | * requested doorbell index (CIK). | |
444 | */ | |
445 | void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v) | |
446 | { | |
447 | if (index < adev->doorbell.num_doorbells) { | |
448 | writel(v, adev->doorbell.ptr + index); | |
449 | } else { | |
450 | DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index); | |
451 | } | |
452 | } | |
453 | ||
832be404 KW |
454 | /** |
455 | * amdgpu_mm_rdoorbell64 - read a doorbell Qword | |
456 | * | |
457 | * @adev: amdgpu_device pointer | |
458 | * @index: doorbell index | |
459 | * | |
460 | * Returns the value in the doorbell aperture at the | |
461 | * requested doorbell index (VEGA10+). | |
462 | */ | |
463 | u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index) | |
464 | { | |
465 | if (index < adev->doorbell.num_doorbells) { | |
466 | return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index)); | |
467 | } else { | |
468 | DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index); | |
469 | return 0; | |
470 | } | |
471 | } | |
472 | ||
473 | /** | |
474 | * amdgpu_mm_wdoorbell64 - write a doorbell Qword | |
475 | * | |
476 | * @adev: amdgpu_device pointer | |
477 | * @index: doorbell index | |
478 | * @v: value to write | |
479 | * | |
480 | * Writes @v to the doorbell aperture at the | |
481 | * requested doorbell index (VEGA10+). | |
482 | */ | |
483 | void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v) | |
484 | { | |
485 | if (index < adev->doorbell.num_doorbells) { | |
486 | atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v); | |
487 | } else { | |
488 | DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index); | |
489 | } | |
490 | } | |
491 | ||
d38ceaf9 AD |
492 | /** |
493 | * amdgpu_invalid_rreg - dummy reg read function | |
494 | * | |
495 | * @adev: amdgpu device pointer | |
496 | * @reg: offset of register | |
497 | * | |
498 | * Dummy register read function. Used for register blocks | |
499 | * that certain asics don't have (all asics). | |
500 | * Returns the value in the register. | |
501 | */ | |
502 | static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg) | |
503 | { | |
504 | DRM_ERROR("Invalid callback to read register 0x%04X\n", reg); | |
505 | BUG(); | |
506 | return 0; | |
507 | } | |
508 | ||
509 | /** | |
510 | * amdgpu_invalid_wreg - dummy reg write function | |
511 | * | |
512 | * @adev: amdgpu device pointer | |
513 | * @reg: offset of register | |
514 | * @v: value to write to the register | |
515 | * | |
516 | * Dummy register read function. Used for register blocks | |
517 | * that certain asics don't have (all asics). | |
518 | */ | |
519 | static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v) | |
520 | { | |
521 | DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n", | |
522 | reg, v); | |
523 | BUG(); | |
524 | } | |
525 | ||
4fa1c6a6 TZ |
526 | /** |
527 | * amdgpu_invalid_rreg64 - dummy 64 bit reg read function | |
528 | * | |
529 | * @adev: amdgpu device pointer | |
530 | * @reg: offset of register | |
531 | * | |
532 | * Dummy register read function. Used for register blocks | |
533 | * that certain asics don't have (all asics). | |
534 | * Returns the value in the register. | |
535 | */ | |
536 | static uint64_t amdgpu_invalid_rreg64(struct amdgpu_device *adev, uint32_t reg) | |
537 | { | |
538 | DRM_ERROR("Invalid callback to read 64 bit register 0x%04X\n", reg); | |
539 | BUG(); | |
540 | return 0; | |
541 | } | |
542 | ||
543 | /** | |
544 | * amdgpu_invalid_wreg64 - dummy reg write function | |
545 | * | |
546 | * @adev: amdgpu device pointer | |
547 | * @reg: offset of register | |
548 | * @v: value to write to the register | |
549 | * | |
550 | * Dummy register read function. Used for register blocks | |
551 | * that certain asics don't have (all asics). | |
552 | */ | |
553 | static void amdgpu_invalid_wreg64(struct amdgpu_device *adev, uint32_t reg, uint64_t v) | |
554 | { | |
555 | DRM_ERROR("Invalid callback to write 64 bit register 0x%04X with 0x%08llX\n", | |
556 | reg, v); | |
557 | BUG(); | |
558 | } | |
559 | ||
d38ceaf9 AD |
560 | /** |
561 | * amdgpu_block_invalid_rreg - dummy reg read function | |
562 | * | |
563 | * @adev: amdgpu device pointer | |
564 | * @block: offset of instance | |
565 | * @reg: offset of register | |
566 | * | |
567 | * Dummy register read function. Used for register blocks | |
568 | * that certain asics don't have (all asics). | |
569 | * Returns the value in the register. | |
570 | */ | |
571 | static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev, | |
572 | uint32_t block, uint32_t reg) | |
573 | { | |
574 | DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n", | |
575 | reg, block); | |
576 | BUG(); | |
577 | return 0; | |
578 | } | |
579 | ||
580 | /** | |
581 | * amdgpu_block_invalid_wreg - dummy reg write function | |
582 | * | |
583 | * @adev: amdgpu device pointer | |
584 | * @block: offset of instance | |
585 | * @reg: offset of register | |
586 | * @v: value to write to the register | |
587 | * | |
588 | * Dummy register read function. Used for register blocks | |
589 | * that certain asics don't have (all asics). | |
590 | */ | |
591 | static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev, | |
592 | uint32_t block, | |
593 | uint32_t reg, uint32_t v) | |
594 | { | |
595 | DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n", | |
596 | reg, block, v); | |
597 | BUG(); | |
598 | } | |
599 | ||
e3ecdffa AD |
600 | /** |
601 | * amdgpu_device_vram_scratch_init - allocate the VRAM scratch page | |
602 | * | |
603 | * @adev: amdgpu device pointer | |
604 | * | |
605 | * Allocates a scratch page of VRAM for use by various things in the | |
606 | * driver. | |
607 | */ | |
06ec9070 | 608 | static int amdgpu_device_vram_scratch_init(struct amdgpu_device *adev) |
d38ceaf9 | 609 | { |
a4a02777 CK |
610 | return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE, |
611 | PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM, | |
612 | &adev->vram_scratch.robj, | |
613 | &adev->vram_scratch.gpu_addr, | |
614 | (void **)&adev->vram_scratch.ptr); | |
d38ceaf9 AD |
615 | } |
616 | ||
e3ecdffa AD |
617 | /** |
618 | * amdgpu_device_vram_scratch_fini - Free the VRAM scratch page | |
619 | * | |
620 | * @adev: amdgpu device pointer | |
621 | * | |
622 | * Frees the VRAM scratch page. | |
623 | */ | |
06ec9070 | 624 | static void amdgpu_device_vram_scratch_fini(struct amdgpu_device *adev) |
d38ceaf9 | 625 | { |
078af1a3 | 626 | amdgpu_bo_free_kernel(&adev->vram_scratch.robj, NULL, NULL); |
d38ceaf9 AD |
627 | } |
628 | ||
629 | /** | |
9c3f2b54 | 630 | * amdgpu_device_program_register_sequence - program an array of registers. |
d38ceaf9 AD |
631 | * |
632 | * @adev: amdgpu_device pointer | |
633 | * @registers: pointer to the register array | |
634 | * @array_size: size of the register array | |
635 | * | |
636 | * Programs an array or registers with and and or masks. | |
637 | * This is a helper for setting golden registers. | |
638 | */ | |
9c3f2b54 AD |
639 | void amdgpu_device_program_register_sequence(struct amdgpu_device *adev, |
640 | const u32 *registers, | |
641 | const u32 array_size) | |
d38ceaf9 AD |
642 | { |
643 | u32 tmp, reg, and_mask, or_mask; | |
644 | int i; | |
645 | ||
646 | if (array_size % 3) | |
647 | return; | |
648 | ||
649 | for (i = 0; i < array_size; i +=3) { | |
650 | reg = registers[i + 0]; | |
651 | and_mask = registers[i + 1]; | |
652 | or_mask = registers[i + 2]; | |
653 | ||
654 | if (and_mask == 0xffffffff) { | |
655 | tmp = or_mask; | |
656 | } else { | |
657 | tmp = RREG32(reg); | |
658 | tmp &= ~and_mask; | |
e0d07657 HZ |
659 | if (adev->family >= AMDGPU_FAMILY_AI) |
660 | tmp |= (or_mask & and_mask); | |
661 | else | |
662 | tmp |= or_mask; | |
d38ceaf9 AD |
663 | } |
664 | WREG32(reg, tmp); | |
665 | } | |
666 | } | |
667 | ||
e3ecdffa AD |
668 | /** |
669 | * amdgpu_device_pci_config_reset - reset the GPU | |
670 | * | |
671 | * @adev: amdgpu_device pointer | |
672 | * | |
673 | * Resets the GPU using the pci config reset sequence. | |
674 | * Only applicable to asics prior to vega10. | |
675 | */ | |
8111c387 | 676 | void amdgpu_device_pci_config_reset(struct amdgpu_device *adev) |
d38ceaf9 AD |
677 | { |
678 | pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA); | |
679 | } | |
680 | ||
681 | /* | |
682 | * GPU doorbell aperture helpers function. | |
683 | */ | |
684 | /** | |
06ec9070 | 685 | * amdgpu_device_doorbell_init - Init doorbell driver information. |
d38ceaf9 AD |
686 | * |
687 | * @adev: amdgpu_device pointer | |
688 | * | |
689 | * Init doorbell driver information (CIK) | |
690 | * Returns 0 on success, error on failure. | |
691 | */ | |
06ec9070 | 692 | static int amdgpu_device_doorbell_init(struct amdgpu_device *adev) |
d38ceaf9 | 693 | { |
6585661d | 694 | |
705e519e CK |
695 | /* No doorbell on SI hardware generation */ |
696 | if (adev->asic_type < CHIP_BONAIRE) { | |
697 | adev->doorbell.base = 0; | |
698 | adev->doorbell.size = 0; | |
699 | adev->doorbell.num_doorbells = 0; | |
700 | adev->doorbell.ptr = NULL; | |
701 | return 0; | |
702 | } | |
703 | ||
d6895ad3 CK |
704 | if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET) |
705 | return -EINVAL; | |
706 | ||
22357775 AD |
707 | amdgpu_asic_init_doorbell_index(adev); |
708 | ||
d38ceaf9 AD |
709 | /* doorbell bar mapping */ |
710 | adev->doorbell.base = pci_resource_start(adev->pdev, 2); | |
711 | adev->doorbell.size = pci_resource_len(adev->pdev, 2); | |
712 | ||
edf600da | 713 | adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32), |
9564f192 | 714 | adev->doorbell_index.max_assignment+1); |
d38ceaf9 AD |
715 | if (adev->doorbell.num_doorbells == 0) |
716 | return -EINVAL; | |
717 | ||
ec3db8a6 | 718 | /* For Vega, reserve and map two pages on doorbell BAR since SDMA |
88dc26e4 OZ |
719 | * paging queue doorbell use the second page. The |
720 | * AMDGPU_DOORBELL64_MAX_ASSIGNMENT definition assumes all the | |
721 | * doorbells are in the first page. So with paging queue enabled, | |
722 | * the max num_doorbells should + 1 page (0x400 in dword) | |
ec3db8a6 PY |
723 | */ |
724 | if (adev->asic_type >= CHIP_VEGA10) | |
88dc26e4 | 725 | adev->doorbell.num_doorbells += 0x400; |
ec3db8a6 | 726 | |
8972e5d2 CK |
727 | adev->doorbell.ptr = ioremap(adev->doorbell.base, |
728 | adev->doorbell.num_doorbells * | |
729 | sizeof(u32)); | |
730 | if (adev->doorbell.ptr == NULL) | |
d38ceaf9 | 731 | return -ENOMEM; |
d38ceaf9 AD |
732 | |
733 | return 0; | |
734 | } | |
735 | ||
736 | /** | |
06ec9070 | 737 | * amdgpu_device_doorbell_fini - Tear down doorbell driver information. |
d38ceaf9 AD |
738 | * |
739 | * @adev: amdgpu_device pointer | |
740 | * | |
741 | * Tear down doorbell driver information (CIK) | |
742 | */ | |
06ec9070 | 743 | static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev) |
d38ceaf9 AD |
744 | { |
745 | iounmap(adev->doorbell.ptr); | |
746 | adev->doorbell.ptr = NULL; | |
747 | } | |
748 | ||
22cb0164 | 749 | |
d38ceaf9 AD |
750 | |
751 | /* | |
06ec9070 | 752 | * amdgpu_device_wb_*() |
455a7bc2 | 753 | * Writeback is the method by which the GPU updates special pages in memory |
ea81a173 | 754 | * with the status of certain GPU events (fences, ring pointers,etc.). |
d38ceaf9 AD |
755 | */ |
756 | ||
757 | /** | |
06ec9070 | 758 | * amdgpu_device_wb_fini - Disable Writeback and free memory |
d38ceaf9 AD |
759 | * |
760 | * @adev: amdgpu_device pointer | |
761 | * | |
762 | * Disables Writeback and frees the Writeback memory (all asics). | |
763 | * Used at driver shutdown. | |
764 | */ | |
06ec9070 | 765 | static void amdgpu_device_wb_fini(struct amdgpu_device *adev) |
d38ceaf9 AD |
766 | { |
767 | if (adev->wb.wb_obj) { | |
a76ed485 AD |
768 | amdgpu_bo_free_kernel(&adev->wb.wb_obj, |
769 | &adev->wb.gpu_addr, | |
770 | (void **)&adev->wb.wb); | |
d38ceaf9 AD |
771 | adev->wb.wb_obj = NULL; |
772 | } | |
773 | } | |
774 | ||
775 | /** | |
06ec9070 | 776 | * amdgpu_device_wb_init- Init Writeback driver info and allocate memory |
d38ceaf9 AD |
777 | * |
778 | * @adev: amdgpu_device pointer | |
779 | * | |
455a7bc2 | 780 | * Initializes writeback and allocates writeback memory (all asics). |
d38ceaf9 AD |
781 | * Used at driver startup. |
782 | * Returns 0 on success or an -error on failure. | |
783 | */ | |
06ec9070 | 784 | static int amdgpu_device_wb_init(struct amdgpu_device *adev) |
d38ceaf9 AD |
785 | { |
786 | int r; | |
787 | ||
788 | if (adev->wb.wb_obj == NULL) { | |
97407b63 AD |
789 | /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */ |
790 | r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8, | |
a76ed485 AD |
791 | PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT, |
792 | &adev->wb.wb_obj, &adev->wb.gpu_addr, | |
793 | (void **)&adev->wb.wb); | |
d38ceaf9 AD |
794 | if (r) { |
795 | dev_warn(adev->dev, "(%d) create WB bo failed\n", r); | |
796 | return r; | |
797 | } | |
d38ceaf9 AD |
798 | |
799 | adev->wb.num_wb = AMDGPU_MAX_WB; | |
800 | memset(&adev->wb.used, 0, sizeof(adev->wb.used)); | |
801 | ||
802 | /* clear wb memory */ | |
73469585 | 803 | memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8); |
d38ceaf9 AD |
804 | } |
805 | ||
806 | return 0; | |
807 | } | |
808 | ||
809 | /** | |
131b4b36 | 810 | * amdgpu_device_wb_get - Allocate a wb entry |
d38ceaf9 AD |
811 | * |
812 | * @adev: amdgpu_device pointer | |
813 | * @wb: wb index | |
814 | * | |
815 | * Allocate a wb slot for use by the driver (all asics). | |
816 | * Returns 0 on success or -EINVAL on failure. | |
817 | */ | |
131b4b36 | 818 | int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb) |
d38ceaf9 AD |
819 | { |
820 | unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb); | |
d38ceaf9 | 821 | |
97407b63 | 822 | if (offset < adev->wb.num_wb) { |
7014285a | 823 | __set_bit(offset, adev->wb.used); |
63ae07ca | 824 | *wb = offset << 3; /* convert to dw offset */ |
0915fdbc ML |
825 | return 0; |
826 | } else { | |
827 | return -EINVAL; | |
828 | } | |
829 | } | |
830 | ||
d38ceaf9 | 831 | /** |
131b4b36 | 832 | * amdgpu_device_wb_free - Free a wb entry |
d38ceaf9 AD |
833 | * |
834 | * @adev: amdgpu_device pointer | |
835 | * @wb: wb index | |
836 | * | |
837 | * Free a wb slot allocated for use by the driver (all asics) | |
838 | */ | |
131b4b36 | 839 | void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb) |
d38ceaf9 | 840 | { |
73469585 | 841 | wb >>= 3; |
d38ceaf9 | 842 | if (wb < adev->wb.num_wb) |
73469585 | 843 | __clear_bit(wb, adev->wb.used); |
d38ceaf9 AD |
844 | } |
845 | ||
d6895ad3 CK |
846 | /** |
847 | * amdgpu_device_resize_fb_bar - try to resize FB BAR | |
848 | * | |
849 | * @adev: amdgpu_device pointer | |
850 | * | |
851 | * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not | |
852 | * to fail, but if any of the BARs is not accessible after the size we abort | |
853 | * driver loading by returning -ENODEV. | |
854 | */ | |
855 | int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev) | |
856 | { | |
770d13b1 | 857 | u64 space_needed = roundup_pow_of_two(adev->gmc.real_vram_size); |
d6895ad3 | 858 | u32 rbar_size = order_base_2(((space_needed >> 20) | 1)) - 1; |
31b8adab CK |
859 | struct pci_bus *root; |
860 | struct resource *res; | |
861 | unsigned i; | |
d6895ad3 CK |
862 | u16 cmd; |
863 | int r; | |
864 | ||
0c03b912 | 865 | /* Bypass for VF */ |
866 | if (amdgpu_sriov_vf(adev)) | |
867 | return 0; | |
868 | ||
31b8adab CK |
869 | /* Check if the root BUS has 64bit memory resources */ |
870 | root = adev->pdev->bus; | |
871 | while (root->parent) | |
872 | root = root->parent; | |
873 | ||
874 | pci_bus_for_each_resource(root, res, i) { | |
0ebb7c54 | 875 | if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) && |
31b8adab CK |
876 | res->start > 0x100000000ull) |
877 | break; | |
878 | } | |
879 | ||
880 | /* Trying to resize is pointless without a root hub window above 4GB */ | |
881 | if (!res) | |
882 | return 0; | |
883 | ||
d6895ad3 CK |
884 | /* Disable memory decoding while we change the BAR addresses and size */ |
885 | pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd); | |
886 | pci_write_config_word(adev->pdev, PCI_COMMAND, | |
887 | cmd & ~PCI_COMMAND_MEMORY); | |
888 | ||
889 | /* Free the VRAM and doorbell BAR, we most likely need to move both. */ | |
06ec9070 | 890 | amdgpu_device_doorbell_fini(adev); |
d6895ad3 CK |
891 | if (adev->asic_type >= CHIP_BONAIRE) |
892 | pci_release_resource(adev->pdev, 2); | |
893 | ||
894 | pci_release_resource(adev->pdev, 0); | |
895 | ||
896 | r = pci_resize_resource(adev->pdev, 0, rbar_size); | |
897 | if (r == -ENOSPC) | |
898 | DRM_INFO("Not enough PCI address space for a large BAR."); | |
899 | else if (r && r != -ENOTSUPP) | |
900 | DRM_ERROR("Problem resizing BAR0 (%d).", r); | |
901 | ||
902 | pci_assign_unassigned_bus_resources(adev->pdev->bus); | |
903 | ||
904 | /* When the doorbell or fb BAR isn't available we have no chance of | |
905 | * using the device. | |
906 | */ | |
06ec9070 | 907 | r = amdgpu_device_doorbell_init(adev); |
d6895ad3 CK |
908 | if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET)) |
909 | return -ENODEV; | |
910 | ||
911 | pci_write_config_word(adev->pdev, PCI_COMMAND, cmd); | |
912 | ||
913 | return 0; | |
914 | } | |
a05502e5 | 915 | |
d38ceaf9 AD |
916 | /* |
917 | * GPU helpers function. | |
918 | */ | |
919 | /** | |
39c640c0 | 920 | * amdgpu_device_need_post - check if the hw need post or not |
d38ceaf9 AD |
921 | * |
922 | * @adev: amdgpu_device pointer | |
923 | * | |
c836fec5 JQ |
924 | * Check if the asic has been initialized (all asics) at driver startup |
925 | * or post is needed if hw reset is performed. | |
926 | * Returns true if need or false if not. | |
d38ceaf9 | 927 | */ |
39c640c0 | 928 | bool amdgpu_device_need_post(struct amdgpu_device *adev) |
d38ceaf9 AD |
929 | { |
930 | uint32_t reg; | |
931 | ||
bec86378 ML |
932 | if (amdgpu_sriov_vf(adev)) |
933 | return false; | |
934 | ||
935 | if (amdgpu_passthrough(adev)) { | |
1da2c326 ML |
936 | /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot |
937 | * some old smc fw still need driver do vPost otherwise gpu hang, while | |
938 | * those smc fw version above 22.15 doesn't have this flaw, so we force | |
939 | * vpost executed for smc version below 22.15 | |
bec86378 ML |
940 | */ |
941 | if (adev->asic_type == CHIP_FIJI) { | |
942 | int err; | |
943 | uint32_t fw_ver; | |
944 | err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev); | |
945 | /* force vPost if error occured */ | |
946 | if (err) | |
947 | return true; | |
948 | ||
949 | fw_ver = *((uint32_t *)adev->pm.fw->data + 69); | |
1da2c326 ML |
950 | if (fw_ver < 0x00160e00) |
951 | return true; | |
bec86378 | 952 | } |
bec86378 | 953 | } |
91fe77eb | 954 | |
955 | if (adev->has_hw_reset) { | |
956 | adev->has_hw_reset = false; | |
957 | return true; | |
958 | } | |
959 | ||
960 | /* bios scratch used on CIK+ */ | |
961 | if (adev->asic_type >= CHIP_BONAIRE) | |
962 | return amdgpu_atombios_scratch_need_asic_init(adev); | |
963 | ||
964 | /* check MEM_SIZE for older asics */ | |
965 | reg = amdgpu_asic_get_config_memsize(adev); | |
966 | ||
967 | if ((reg != 0) && (reg != 0xffffffff)) | |
968 | return false; | |
969 | ||
970 | return true; | |
bec86378 ML |
971 | } |
972 | ||
d38ceaf9 AD |
973 | /* if we get transitioned to only one device, take VGA back */ |
974 | /** | |
06ec9070 | 975 | * amdgpu_device_vga_set_decode - enable/disable vga decode |
d38ceaf9 AD |
976 | * |
977 | * @cookie: amdgpu_device pointer | |
978 | * @state: enable/disable vga decode | |
979 | * | |
980 | * Enable/disable vga decode (all asics). | |
981 | * Returns VGA resource flags. | |
982 | */ | |
06ec9070 | 983 | static unsigned int amdgpu_device_vga_set_decode(void *cookie, bool state) |
d38ceaf9 AD |
984 | { |
985 | struct amdgpu_device *adev = cookie; | |
986 | amdgpu_asic_set_vga_state(adev, state); | |
987 | if (state) | |
988 | return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM | | |
989 | VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM; | |
990 | else | |
991 | return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM; | |
992 | } | |
993 | ||
e3ecdffa AD |
994 | /** |
995 | * amdgpu_device_check_block_size - validate the vm block size | |
996 | * | |
997 | * @adev: amdgpu_device pointer | |
998 | * | |
999 | * Validates the vm block size specified via module parameter. | |
1000 | * The vm block size defines number of bits in page table versus page directory, | |
1001 | * a page is 4KB so we have 12 bits offset, minimum 9 bits in the | |
1002 | * page table and the remaining bits are in the page directory. | |
1003 | */ | |
06ec9070 | 1004 | static void amdgpu_device_check_block_size(struct amdgpu_device *adev) |
a1adf8be CZ |
1005 | { |
1006 | /* defines number of bits in page table versus page directory, | |
1007 | * a page is 4KB so we have 12 bits offset, minimum 9 bits in the | |
1008 | * page table and the remaining bits are in the page directory */ | |
bab4fee7 JZ |
1009 | if (amdgpu_vm_block_size == -1) |
1010 | return; | |
a1adf8be | 1011 | |
bab4fee7 | 1012 | if (amdgpu_vm_block_size < 9) { |
a1adf8be CZ |
1013 | dev_warn(adev->dev, "VM page table size (%d) too small\n", |
1014 | amdgpu_vm_block_size); | |
97489129 | 1015 | amdgpu_vm_block_size = -1; |
a1adf8be | 1016 | } |
a1adf8be CZ |
1017 | } |
1018 | ||
e3ecdffa AD |
1019 | /** |
1020 | * amdgpu_device_check_vm_size - validate the vm size | |
1021 | * | |
1022 | * @adev: amdgpu_device pointer | |
1023 | * | |
1024 | * Validates the vm size in GB specified via module parameter. | |
1025 | * The VM size is the size of the GPU virtual memory space in GB. | |
1026 | */ | |
06ec9070 | 1027 | static void amdgpu_device_check_vm_size(struct amdgpu_device *adev) |
83ca145d | 1028 | { |
64dab074 AD |
1029 | /* no need to check the default value */ |
1030 | if (amdgpu_vm_size == -1) | |
1031 | return; | |
1032 | ||
83ca145d ZJ |
1033 | if (amdgpu_vm_size < 1) { |
1034 | dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n", | |
1035 | amdgpu_vm_size); | |
f3368128 | 1036 | amdgpu_vm_size = -1; |
83ca145d | 1037 | } |
83ca145d ZJ |
1038 | } |
1039 | ||
7951e376 RZ |
1040 | static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev) |
1041 | { | |
1042 | struct sysinfo si; | |
a9d4fe2f | 1043 | bool is_os_64 = (sizeof(void *) == 8); |
7951e376 RZ |
1044 | uint64_t total_memory; |
1045 | uint64_t dram_size_seven_GB = 0x1B8000000; | |
1046 | uint64_t dram_size_three_GB = 0xB8000000; | |
1047 | ||
1048 | if (amdgpu_smu_memory_pool_size == 0) | |
1049 | return; | |
1050 | ||
1051 | if (!is_os_64) { | |
1052 | DRM_WARN("Not 64-bit OS, feature not supported\n"); | |
1053 | goto def_value; | |
1054 | } | |
1055 | si_meminfo(&si); | |
1056 | total_memory = (uint64_t)si.totalram * si.mem_unit; | |
1057 | ||
1058 | if ((amdgpu_smu_memory_pool_size == 1) || | |
1059 | (amdgpu_smu_memory_pool_size == 2)) { | |
1060 | if (total_memory < dram_size_three_GB) | |
1061 | goto def_value1; | |
1062 | } else if ((amdgpu_smu_memory_pool_size == 4) || | |
1063 | (amdgpu_smu_memory_pool_size == 8)) { | |
1064 | if (total_memory < dram_size_seven_GB) | |
1065 | goto def_value1; | |
1066 | } else { | |
1067 | DRM_WARN("Smu memory pool size not supported\n"); | |
1068 | goto def_value; | |
1069 | } | |
1070 | adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28; | |
1071 | ||
1072 | return; | |
1073 | ||
1074 | def_value1: | |
1075 | DRM_WARN("No enough system memory\n"); | |
1076 | def_value: | |
1077 | adev->pm.smu_prv_buffer_size = 0; | |
1078 | } | |
1079 | ||
d38ceaf9 | 1080 | /** |
06ec9070 | 1081 | * amdgpu_device_check_arguments - validate module params |
d38ceaf9 AD |
1082 | * |
1083 | * @adev: amdgpu_device pointer | |
1084 | * | |
1085 | * Validates certain module parameters and updates | |
1086 | * the associated values used by the driver (all asics). | |
1087 | */ | |
912dfc84 | 1088 | static int amdgpu_device_check_arguments(struct amdgpu_device *adev) |
d38ceaf9 | 1089 | { |
5b011235 CZ |
1090 | if (amdgpu_sched_jobs < 4) { |
1091 | dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n", | |
1092 | amdgpu_sched_jobs); | |
1093 | amdgpu_sched_jobs = 4; | |
76117507 | 1094 | } else if (!is_power_of_2(amdgpu_sched_jobs)){ |
5b011235 CZ |
1095 | dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n", |
1096 | amdgpu_sched_jobs); | |
1097 | amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs); | |
1098 | } | |
d38ceaf9 | 1099 | |
83e74db6 | 1100 | if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) { |
f9321cc4 CK |
1101 | /* gart size must be greater or equal to 32M */ |
1102 | dev_warn(adev->dev, "gart size (%d) too small\n", | |
1103 | amdgpu_gart_size); | |
83e74db6 | 1104 | amdgpu_gart_size = -1; |
d38ceaf9 AD |
1105 | } |
1106 | ||
36d38372 | 1107 | if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) { |
c4e1a13a | 1108 | /* gtt size must be greater or equal to 32M */ |
36d38372 CK |
1109 | dev_warn(adev->dev, "gtt size (%d) too small\n", |
1110 | amdgpu_gtt_size); | |
1111 | amdgpu_gtt_size = -1; | |
d38ceaf9 AD |
1112 | } |
1113 | ||
d07f14be RH |
1114 | /* valid range is between 4 and 9 inclusive */ |
1115 | if (amdgpu_vm_fragment_size != -1 && | |
1116 | (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) { | |
1117 | dev_warn(adev->dev, "valid range is between 4 and 9\n"); | |
1118 | amdgpu_vm_fragment_size = -1; | |
1119 | } | |
1120 | ||
7951e376 RZ |
1121 | amdgpu_device_check_smu_prv_buffer_size(adev); |
1122 | ||
06ec9070 | 1123 | amdgpu_device_check_vm_size(adev); |
d38ceaf9 | 1124 | |
06ec9070 | 1125 | amdgpu_device_check_block_size(adev); |
6a7f76e7 | 1126 | |
19aede77 | 1127 | adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type); |
912dfc84 | 1128 | |
e3c00faa | 1129 | return 0; |
d38ceaf9 AD |
1130 | } |
1131 | ||
1132 | /** | |
1133 | * amdgpu_switcheroo_set_state - set switcheroo state | |
1134 | * | |
1135 | * @pdev: pci dev pointer | |
1694467b | 1136 | * @state: vga_switcheroo state |
d38ceaf9 AD |
1137 | * |
1138 | * Callback for the switcheroo driver. Suspends or resumes the | |
1139 | * the asics before or after it is powered up using ACPI methods. | |
1140 | */ | |
1141 | static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, enum vga_switcheroo_state state) | |
1142 | { | |
1143 | struct drm_device *dev = pci_get_drvdata(pdev); | |
de185019 | 1144 | int r; |
d38ceaf9 | 1145 | |
31af062a | 1146 | if (amdgpu_device_supports_boco(dev) && state == VGA_SWITCHEROO_OFF) |
d38ceaf9 AD |
1147 | return; |
1148 | ||
1149 | if (state == VGA_SWITCHEROO_ON) { | |
7ca85295 | 1150 | pr_info("amdgpu: switched on\n"); |
d38ceaf9 AD |
1151 | /* don't suspend or resume card normally */ |
1152 | dev->switch_power_state = DRM_SWITCH_POWER_CHANGING; | |
1153 | ||
de185019 AD |
1154 | pci_set_power_state(dev->pdev, PCI_D0); |
1155 | pci_restore_state(dev->pdev); | |
1156 | r = pci_enable_device(dev->pdev); | |
1157 | if (r) | |
1158 | DRM_WARN("pci_enable_device failed (%d)\n", r); | |
1159 | amdgpu_device_resume(dev, true); | |
d38ceaf9 | 1160 | |
d38ceaf9 AD |
1161 | dev->switch_power_state = DRM_SWITCH_POWER_ON; |
1162 | drm_kms_helper_poll_enable(dev); | |
1163 | } else { | |
7ca85295 | 1164 | pr_info("amdgpu: switched off\n"); |
d38ceaf9 AD |
1165 | drm_kms_helper_poll_disable(dev); |
1166 | dev->switch_power_state = DRM_SWITCH_POWER_CHANGING; | |
de185019 AD |
1167 | amdgpu_device_suspend(dev, true); |
1168 | pci_save_state(dev->pdev); | |
1169 | /* Shut down the device */ | |
1170 | pci_disable_device(dev->pdev); | |
1171 | pci_set_power_state(dev->pdev, PCI_D3cold); | |
d38ceaf9 AD |
1172 | dev->switch_power_state = DRM_SWITCH_POWER_OFF; |
1173 | } | |
1174 | } | |
1175 | ||
1176 | /** | |
1177 | * amdgpu_switcheroo_can_switch - see if switcheroo state can change | |
1178 | * | |
1179 | * @pdev: pci dev pointer | |
1180 | * | |
1181 | * Callback for the switcheroo driver. Check of the switcheroo | |
1182 | * state can be changed. | |
1183 | * Returns true if the state can be changed, false if not. | |
1184 | */ | |
1185 | static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev) | |
1186 | { | |
1187 | struct drm_device *dev = pci_get_drvdata(pdev); | |
1188 | ||
1189 | /* | |
1190 | * FIXME: open_count is protected by drm_global_mutex but that would lead to | |
1191 | * locking inversion with the driver load path. And the access here is | |
1192 | * completely racy anyway. So don't bother with locking for now. | |
1193 | */ | |
7e13ad89 | 1194 | return atomic_read(&dev->open_count) == 0; |
d38ceaf9 AD |
1195 | } |
1196 | ||
1197 | static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = { | |
1198 | .set_gpu_state = amdgpu_switcheroo_set_state, | |
1199 | .reprobe = NULL, | |
1200 | .can_switch = amdgpu_switcheroo_can_switch, | |
1201 | }; | |
1202 | ||
e3ecdffa AD |
1203 | /** |
1204 | * amdgpu_device_ip_set_clockgating_state - set the CG state | |
1205 | * | |
87e3f136 | 1206 | * @dev: amdgpu_device pointer |
e3ecdffa AD |
1207 | * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) |
1208 | * @state: clockgating state (gate or ungate) | |
1209 | * | |
1210 | * Sets the requested clockgating state for all instances of | |
1211 | * the hardware IP specified. | |
1212 | * Returns the error code from the last instance. | |
1213 | */ | |
43fa561f | 1214 | int amdgpu_device_ip_set_clockgating_state(void *dev, |
2990a1fc AD |
1215 | enum amd_ip_block_type block_type, |
1216 | enum amd_clockgating_state state) | |
d38ceaf9 | 1217 | { |
43fa561f | 1218 | struct amdgpu_device *adev = dev; |
d38ceaf9 AD |
1219 | int i, r = 0; |
1220 | ||
1221 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 1222 | if (!adev->ip_blocks[i].status.valid) |
9ecbe7f5 | 1223 | continue; |
c722865a RZ |
1224 | if (adev->ip_blocks[i].version->type != block_type) |
1225 | continue; | |
1226 | if (!adev->ip_blocks[i].version->funcs->set_clockgating_state) | |
1227 | continue; | |
1228 | r = adev->ip_blocks[i].version->funcs->set_clockgating_state( | |
1229 | (void *)adev, state); | |
1230 | if (r) | |
1231 | DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n", | |
1232 | adev->ip_blocks[i].version->funcs->name, r); | |
d38ceaf9 AD |
1233 | } |
1234 | return r; | |
1235 | } | |
1236 | ||
e3ecdffa AD |
1237 | /** |
1238 | * amdgpu_device_ip_set_powergating_state - set the PG state | |
1239 | * | |
87e3f136 | 1240 | * @dev: amdgpu_device pointer |
e3ecdffa AD |
1241 | * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) |
1242 | * @state: powergating state (gate or ungate) | |
1243 | * | |
1244 | * Sets the requested powergating state for all instances of | |
1245 | * the hardware IP specified. | |
1246 | * Returns the error code from the last instance. | |
1247 | */ | |
43fa561f | 1248 | int amdgpu_device_ip_set_powergating_state(void *dev, |
2990a1fc AD |
1249 | enum amd_ip_block_type block_type, |
1250 | enum amd_powergating_state state) | |
d38ceaf9 | 1251 | { |
43fa561f | 1252 | struct amdgpu_device *adev = dev; |
d38ceaf9 AD |
1253 | int i, r = 0; |
1254 | ||
1255 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 1256 | if (!adev->ip_blocks[i].status.valid) |
9ecbe7f5 | 1257 | continue; |
c722865a RZ |
1258 | if (adev->ip_blocks[i].version->type != block_type) |
1259 | continue; | |
1260 | if (!adev->ip_blocks[i].version->funcs->set_powergating_state) | |
1261 | continue; | |
1262 | r = adev->ip_blocks[i].version->funcs->set_powergating_state( | |
1263 | (void *)adev, state); | |
1264 | if (r) | |
1265 | DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n", | |
1266 | adev->ip_blocks[i].version->funcs->name, r); | |
d38ceaf9 AD |
1267 | } |
1268 | return r; | |
1269 | } | |
1270 | ||
e3ecdffa AD |
1271 | /** |
1272 | * amdgpu_device_ip_get_clockgating_state - get the CG state | |
1273 | * | |
1274 | * @adev: amdgpu_device pointer | |
1275 | * @flags: clockgating feature flags | |
1276 | * | |
1277 | * Walks the list of IPs on the device and updates the clockgating | |
1278 | * flags for each IP. | |
1279 | * Updates @flags with the feature flags for each hardware IP where | |
1280 | * clockgating is enabled. | |
1281 | */ | |
2990a1fc AD |
1282 | void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev, |
1283 | u32 *flags) | |
6cb2d4e4 HR |
1284 | { |
1285 | int i; | |
1286 | ||
1287 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
1288 | if (!adev->ip_blocks[i].status.valid) | |
1289 | continue; | |
1290 | if (adev->ip_blocks[i].version->funcs->get_clockgating_state) | |
1291 | adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags); | |
1292 | } | |
1293 | } | |
1294 | ||
e3ecdffa AD |
1295 | /** |
1296 | * amdgpu_device_ip_wait_for_idle - wait for idle | |
1297 | * | |
1298 | * @adev: amdgpu_device pointer | |
1299 | * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) | |
1300 | * | |
1301 | * Waits for the request hardware IP to be idle. | |
1302 | * Returns 0 for success or a negative error code on failure. | |
1303 | */ | |
2990a1fc AD |
1304 | int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev, |
1305 | enum amd_ip_block_type block_type) | |
5dbbb60b AD |
1306 | { |
1307 | int i, r; | |
1308 | ||
1309 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 1310 | if (!adev->ip_blocks[i].status.valid) |
9ecbe7f5 | 1311 | continue; |
a1255107 AD |
1312 | if (adev->ip_blocks[i].version->type == block_type) { |
1313 | r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev); | |
5dbbb60b AD |
1314 | if (r) |
1315 | return r; | |
1316 | break; | |
1317 | } | |
1318 | } | |
1319 | return 0; | |
1320 | ||
1321 | } | |
1322 | ||
e3ecdffa AD |
1323 | /** |
1324 | * amdgpu_device_ip_is_idle - is the hardware IP idle | |
1325 | * | |
1326 | * @adev: amdgpu_device pointer | |
1327 | * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) | |
1328 | * | |
1329 | * Check if the hardware IP is idle or not. | |
1330 | * Returns true if it the IP is idle, false if not. | |
1331 | */ | |
2990a1fc AD |
1332 | bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev, |
1333 | enum amd_ip_block_type block_type) | |
5dbbb60b AD |
1334 | { |
1335 | int i; | |
1336 | ||
1337 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 1338 | if (!adev->ip_blocks[i].status.valid) |
9ecbe7f5 | 1339 | continue; |
a1255107 AD |
1340 | if (adev->ip_blocks[i].version->type == block_type) |
1341 | return adev->ip_blocks[i].version->funcs->is_idle((void *)adev); | |
5dbbb60b AD |
1342 | } |
1343 | return true; | |
1344 | ||
1345 | } | |
1346 | ||
e3ecdffa AD |
1347 | /** |
1348 | * amdgpu_device_ip_get_ip_block - get a hw IP pointer | |
1349 | * | |
1350 | * @adev: amdgpu_device pointer | |
87e3f136 | 1351 | * @type: Type of hardware IP (SMU, GFX, UVD, etc.) |
e3ecdffa AD |
1352 | * |
1353 | * Returns a pointer to the hardware IP block structure | |
1354 | * if it exists for the asic, otherwise NULL. | |
1355 | */ | |
2990a1fc AD |
1356 | struct amdgpu_ip_block * |
1357 | amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev, | |
1358 | enum amd_ip_block_type type) | |
d38ceaf9 AD |
1359 | { |
1360 | int i; | |
1361 | ||
1362 | for (i = 0; i < adev->num_ip_blocks; i++) | |
a1255107 | 1363 | if (adev->ip_blocks[i].version->type == type) |
d38ceaf9 AD |
1364 | return &adev->ip_blocks[i]; |
1365 | ||
1366 | return NULL; | |
1367 | } | |
1368 | ||
1369 | /** | |
2990a1fc | 1370 | * amdgpu_device_ip_block_version_cmp |
d38ceaf9 AD |
1371 | * |
1372 | * @adev: amdgpu_device pointer | |
5fc3aeeb | 1373 | * @type: enum amd_ip_block_type |
d38ceaf9 AD |
1374 | * @major: major version |
1375 | * @minor: minor version | |
1376 | * | |
1377 | * return 0 if equal or greater | |
1378 | * return 1 if smaller or the ip_block doesn't exist | |
1379 | */ | |
2990a1fc AD |
1380 | int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev, |
1381 | enum amd_ip_block_type type, | |
1382 | u32 major, u32 minor) | |
d38ceaf9 | 1383 | { |
2990a1fc | 1384 | struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type); |
d38ceaf9 | 1385 | |
a1255107 AD |
1386 | if (ip_block && ((ip_block->version->major > major) || |
1387 | ((ip_block->version->major == major) && | |
1388 | (ip_block->version->minor >= minor)))) | |
d38ceaf9 AD |
1389 | return 0; |
1390 | ||
1391 | return 1; | |
1392 | } | |
1393 | ||
a1255107 | 1394 | /** |
2990a1fc | 1395 | * amdgpu_device_ip_block_add |
a1255107 AD |
1396 | * |
1397 | * @adev: amdgpu_device pointer | |
1398 | * @ip_block_version: pointer to the IP to add | |
1399 | * | |
1400 | * Adds the IP block driver information to the collection of IPs | |
1401 | * on the asic. | |
1402 | */ | |
2990a1fc AD |
1403 | int amdgpu_device_ip_block_add(struct amdgpu_device *adev, |
1404 | const struct amdgpu_ip_block_version *ip_block_version) | |
a1255107 AD |
1405 | { |
1406 | if (!ip_block_version) | |
1407 | return -EINVAL; | |
1408 | ||
e966a725 | 1409 | DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks, |
a0bae357 HR |
1410 | ip_block_version->funcs->name); |
1411 | ||
a1255107 AD |
1412 | adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version; |
1413 | ||
1414 | return 0; | |
1415 | } | |
1416 | ||
e3ecdffa AD |
1417 | /** |
1418 | * amdgpu_device_enable_virtual_display - enable virtual display feature | |
1419 | * | |
1420 | * @adev: amdgpu_device pointer | |
1421 | * | |
1422 | * Enabled the virtual display feature if the user has enabled it via | |
1423 | * the module parameter virtual_display. This feature provides a virtual | |
1424 | * display hardware on headless boards or in virtualized environments. | |
1425 | * This function parses and validates the configuration string specified by | |
1426 | * the user and configues the virtual display configuration (number of | |
1427 | * virtual connectors, crtcs, etc.) specified. | |
1428 | */ | |
483ef985 | 1429 | static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev) |
9accf2fd ED |
1430 | { |
1431 | adev->enable_virtual_display = false; | |
1432 | ||
1433 | if (amdgpu_virtual_display) { | |
1434 | struct drm_device *ddev = adev->ddev; | |
1435 | const char *pci_address_name = pci_name(ddev->pdev); | |
0f66356d | 1436 | char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname; |
9accf2fd ED |
1437 | |
1438 | pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL); | |
1439 | pciaddstr_tmp = pciaddstr; | |
0f66356d ED |
1440 | while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) { |
1441 | pciaddname = strsep(&pciaddname_tmp, ","); | |
967de2a9 YT |
1442 | if (!strcmp("all", pciaddname) |
1443 | || !strcmp(pci_address_name, pciaddname)) { | |
0f66356d ED |
1444 | long num_crtc; |
1445 | int res = -1; | |
1446 | ||
9accf2fd | 1447 | adev->enable_virtual_display = true; |
0f66356d ED |
1448 | |
1449 | if (pciaddname_tmp) | |
1450 | res = kstrtol(pciaddname_tmp, 10, | |
1451 | &num_crtc); | |
1452 | ||
1453 | if (!res) { | |
1454 | if (num_crtc < 1) | |
1455 | num_crtc = 1; | |
1456 | if (num_crtc > 6) | |
1457 | num_crtc = 6; | |
1458 | adev->mode_info.num_crtc = num_crtc; | |
1459 | } else { | |
1460 | adev->mode_info.num_crtc = 1; | |
1461 | } | |
9accf2fd ED |
1462 | break; |
1463 | } | |
1464 | } | |
1465 | ||
0f66356d ED |
1466 | DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n", |
1467 | amdgpu_virtual_display, pci_address_name, | |
1468 | adev->enable_virtual_display, adev->mode_info.num_crtc); | |
9accf2fd ED |
1469 | |
1470 | kfree(pciaddstr); | |
1471 | } | |
1472 | } | |
1473 | ||
e3ecdffa AD |
1474 | /** |
1475 | * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware | |
1476 | * | |
1477 | * @adev: amdgpu_device pointer | |
1478 | * | |
1479 | * Parses the asic configuration parameters specified in the gpu info | |
1480 | * firmware and makes them availale to the driver for use in configuring | |
1481 | * the asic. | |
1482 | * Returns 0 on success, -EINVAL on failure. | |
1483 | */ | |
e2a75f88 AD |
1484 | static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev) |
1485 | { | |
e2a75f88 AD |
1486 | const char *chip_name; |
1487 | char fw_name[30]; | |
1488 | int err; | |
1489 | const struct gpu_info_firmware_header_v1_0 *hdr; | |
1490 | ||
ab4fe3e1 HR |
1491 | adev->firmware.gpu_info_fw = NULL; |
1492 | ||
e2a75f88 AD |
1493 | switch (adev->asic_type) { |
1494 | case CHIP_TOPAZ: | |
1495 | case CHIP_TONGA: | |
1496 | case CHIP_FIJI: | |
e2a75f88 | 1497 | case CHIP_POLARIS10: |
cc07f18d | 1498 | case CHIP_POLARIS11: |
e2a75f88 | 1499 | case CHIP_POLARIS12: |
cc07f18d | 1500 | case CHIP_VEGAM: |
e2a75f88 AD |
1501 | case CHIP_CARRIZO: |
1502 | case CHIP_STONEY: | |
1503 | #ifdef CONFIG_DRM_AMDGPU_SI | |
1504 | case CHIP_VERDE: | |
1505 | case CHIP_TAHITI: | |
1506 | case CHIP_PITCAIRN: | |
1507 | case CHIP_OLAND: | |
1508 | case CHIP_HAINAN: | |
1509 | #endif | |
1510 | #ifdef CONFIG_DRM_AMDGPU_CIK | |
1511 | case CHIP_BONAIRE: | |
1512 | case CHIP_HAWAII: | |
1513 | case CHIP_KAVERI: | |
1514 | case CHIP_KABINI: | |
1515 | case CHIP_MULLINS: | |
1516 | #endif | |
27c0bc71 | 1517 | case CHIP_VEGA20: |
e2a75f88 AD |
1518 | default: |
1519 | return 0; | |
1520 | case CHIP_VEGA10: | |
1521 | chip_name = "vega10"; | |
1522 | break; | |
3f76dced AD |
1523 | case CHIP_VEGA12: |
1524 | chip_name = "vega12"; | |
1525 | break; | |
2d2e5e7e | 1526 | case CHIP_RAVEN: |
54c4d17e FX |
1527 | if (adev->rev_id >= 8) |
1528 | chip_name = "raven2"; | |
741deade AD |
1529 | else if (adev->pdev->device == 0x15d8) |
1530 | chip_name = "picasso"; | |
54c4d17e FX |
1531 | else |
1532 | chip_name = "raven"; | |
2d2e5e7e | 1533 | break; |
65e60f6e LM |
1534 | case CHIP_ARCTURUS: |
1535 | chip_name = "arcturus"; | |
1536 | break; | |
b51a26a0 HR |
1537 | case CHIP_RENOIR: |
1538 | chip_name = "renoir"; | |
1539 | break; | |
23c6268e HR |
1540 | case CHIP_NAVI10: |
1541 | chip_name = "navi10"; | |
1542 | break; | |
ed42cfe1 XY |
1543 | case CHIP_NAVI14: |
1544 | chip_name = "navi14"; | |
1545 | break; | |
42b325e5 XY |
1546 | case CHIP_NAVI12: |
1547 | chip_name = "navi12"; | |
1548 | break; | |
e2a75f88 AD |
1549 | } |
1550 | ||
1551 | snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name); | |
ab4fe3e1 | 1552 | err = request_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev); |
e2a75f88 AD |
1553 | if (err) { |
1554 | dev_err(adev->dev, | |
1555 | "Failed to load gpu_info firmware \"%s\"\n", | |
1556 | fw_name); | |
1557 | goto out; | |
1558 | } | |
ab4fe3e1 | 1559 | err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw); |
e2a75f88 AD |
1560 | if (err) { |
1561 | dev_err(adev->dev, | |
1562 | "Failed to validate gpu_info firmware \"%s\"\n", | |
1563 | fw_name); | |
1564 | goto out; | |
1565 | } | |
1566 | ||
ab4fe3e1 | 1567 | hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data; |
e2a75f88 AD |
1568 | amdgpu_ucode_print_gpu_info_hdr(&hdr->header); |
1569 | ||
1570 | switch (hdr->version_major) { | |
1571 | case 1: | |
1572 | { | |
1573 | const struct gpu_info_firmware_v1_0 *gpu_info_fw = | |
ab4fe3e1 | 1574 | (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data + |
e2a75f88 AD |
1575 | le32_to_cpu(hdr->header.ucode_array_offset_bytes)); |
1576 | ||
ec51d3fa XY |
1577 | if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10) |
1578 | goto parse_soc_bounding_box; | |
1579 | ||
b5ab16bf AD |
1580 | adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se); |
1581 | adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh); | |
1582 | adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se); | |
1583 | adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se); | |
e2a75f88 | 1584 | adev->gfx.config.max_texture_channel_caches = |
b5ab16bf AD |
1585 | le32_to_cpu(gpu_info_fw->gc_num_tccs); |
1586 | adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs); | |
1587 | adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds); | |
1588 | adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth); | |
1589 | adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth); | |
e2a75f88 | 1590 | adev->gfx.config.double_offchip_lds_buf = |
b5ab16bf AD |
1591 | le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer); |
1592 | adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size); | |
51fd0370 HZ |
1593 | adev->gfx.cu_info.max_waves_per_simd = |
1594 | le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd); | |
1595 | adev->gfx.cu_info.max_scratch_slots_per_cu = | |
1596 | le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu); | |
1597 | adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size); | |
48321c3d | 1598 | if (hdr->version_minor >= 1) { |
35c2e910 HZ |
1599 | const struct gpu_info_firmware_v1_1 *gpu_info_fw = |
1600 | (const struct gpu_info_firmware_v1_1 *)(adev->firmware.gpu_info_fw->data + | |
1601 | le32_to_cpu(hdr->header.ucode_array_offset_bytes)); | |
1602 | adev->gfx.config.num_sc_per_sh = | |
1603 | le32_to_cpu(gpu_info_fw->num_sc_per_sh); | |
1604 | adev->gfx.config.num_packer_per_sc = | |
1605 | le32_to_cpu(gpu_info_fw->num_packer_per_sc); | |
1606 | } | |
ec51d3fa XY |
1607 | |
1608 | parse_soc_bounding_box: | |
ec51d3fa XY |
1609 | /* |
1610 | * soc bounding box info is not integrated in disocovery table, | |
1611 | * we always need to parse it from gpu info firmware. | |
1612 | */ | |
48321c3d HW |
1613 | if (hdr->version_minor == 2) { |
1614 | const struct gpu_info_firmware_v1_2 *gpu_info_fw = | |
1615 | (const struct gpu_info_firmware_v1_2 *)(adev->firmware.gpu_info_fw->data + | |
1616 | le32_to_cpu(hdr->header.ucode_array_offset_bytes)); | |
1617 | adev->dm.soc_bounding_box = &gpu_info_fw->soc_bounding_box; | |
1618 | } | |
e2a75f88 AD |
1619 | break; |
1620 | } | |
1621 | default: | |
1622 | dev_err(adev->dev, | |
1623 | "Unsupported gpu_info table %d\n", hdr->header.ucode_version); | |
1624 | err = -EINVAL; | |
1625 | goto out; | |
1626 | } | |
1627 | out: | |
e2a75f88 AD |
1628 | return err; |
1629 | } | |
1630 | ||
e3ecdffa AD |
1631 | /** |
1632 | * amdgpu_device_ip_early_init - run early init for hardware IPs | |
1633 | * | |
1634 | * @adev: amdgpu_device pointer | |
1635 | * | |
1636 | * Early initialization pass for hardware IPs. The hardware IPs that make | |
1637 | * up each asic are discovered each IP's early_init callback is run. This | |
1638 | * is the first stage in initializing the asic. | |
1639 | * Returns 0 on success, negative error code on failure. | |
1640 | */ | |
06ec9070 | 1641 | static int amdgpu_device_ip_early_init(struct amdgpu_device *adev) |
d38ceaf9 | 1642 | { |
aaa36a97 | 1643 | int i, r; |
d38ceaf9 | 1644 | |
483ef985 | 1645 | amdgpu_device_enable_virtual_display(adev); |
a6be7570 | 1646 | |
d38ceaf9 | 1647 | switch (adev->asic_type) { |
aaa36a97 AD |
1648 | case CHIP_TOPAZ: |
1649 | case CHIP_TONGA: | |
48299f95 | 1650 | case CHIP_FIJI: |
2cc0c0b5 | 1651 | case CHIP_POLARIS10: |
32cc7e53 | 1652 | case CHIP_POLARIS11: |
c4642a47 | 1653 | case CHIP_POLARIS12: |
32cc7e53 | 1654 | case CHIP_VEGAM: |
aaa36a97 | 1655 | case CHIP_CARRIZO: |
39bb0c92 SL |
1656 | case CHIP_STONEY: |
1657 | if (adev->asic_type == CHIP_CARRIZO || adev->asic_type == CHIP_STONEY) | |
aaa36a97 AD |
1658 | adev->family = AMDGPU_FAMILY_CZ; |
1659 | else | |
1660 | adev->family = AMDGPU_FAMILY_VI; | |
1661 | ||
1662 | r = vi_set_ip_blocks(adev); | |
1663 | if (r) | |
1664 | return r; | |
1665 | break; | |
33f34802 KW |
1666 | #ifdef CONFIG_DRM_AMDGPU_SI |
1667 | case CHIP_VERDE: | |
1668 | case CHIP_TAHITI: | |
1669 | case CHIP_PITCAIRN: | |
1670 | case CHIP_OLAND: | |
1671 | case CHIP_HAINAN: | |
295d0daf | 1672 | adev->family = AMDGPU_FAMILY_SI; |
33f34802 KW |
1673 | r = si_set_ip_blocks(adev); |
1674 | if (r) | |
1675 | return r; | |
1676 | break; | |
1677 | #endif | |
a2e73f56 AD |
1678 | #ifdef CONFIG_DRM_AMDGPU_CIK |
1679 | case CHIP_BONAIRE: | |
1680 | case CHIP_HAWAII: | |
1681 | case CHIP_KAVERI: | |
1682 | case CHIP_KABINI: | |
1683 | case CHIP_MULLINS: | |
1684 | if ((adev->asic_type == CHIP_BONAIRE) || (adev->asic_type == CHIP_HAWAII)) | |
1685 | adev->family = AMDGPU_FAMILY_CI; | |
1686 | else | |
1687 | adev->family = AMDGPU_FAMILY_KV; | |
1688 | ||
1689 | r = cik_set_ip_blocks(adev); | |
1690 | if (r) | |
1691 | return r; | |
1692 | break; | |
1693 | #endif | |
e48a3cd9 AD |
1694 | case CHIP_VEGA10: |
1695 | case CHIP_VEGA12: | |
e4bd8170 | 1696 | case CHIP_VEGA20: |
e48a3cd9 | 1697 | case CHIP_RAVEN: |
61cf44c1 | 1698 | case CHIP_ARCTURUS: |
b51a26a0 HR |
1699 | case CHIP_RENOIR: |
1700 | if (adev->asic_type == CHIP_RAVEN || | |
1701 | adev->asic_type == CHIP_RENOIR) | |
2ca8a5d2 CZ |
1702 | adev->family = AMDGPU_FAMILY_RV; |
1703 | else | |
1704 | adev->family = AMDGPU_FAMILY_AI; | |
460826e6 KW |
1705 | |
1706 | r = soc15_set_ip_blocks(adev); | |
1707 | if (r) | |
1708 | return r; | |
1709 | break; | |
0a5b8c7b | 1710 | case CHIP_NAVI10: |
7ecb5cd4 | 1711 | case CHIP_NAVI14: |
4808cf9c | 1712 | case CHIP_NAVI12: |
0a5b8c7b HR |
1713 | adev->family = AMDGPU_FAMILY_NV; |
1714 | ||
1715 | r = nv_set_ip_blocks(adev); | |
1716 | if (r) | |
1717 | return r; | |
1718 | break; | |
d38ceaf9 AD |
1719 | default: |
1720 | /* FIXME: not supported yet */ | |
1721 | return -EINVAL; | |
1722 | } | |
1723 | ||
e2a75f88 AD |
1724 | r = amdgpu_device_parse_gpu_info_fw(adev); |
1725 | if (r) | |
1726 | return r; | |
1727 | ||
ec51d3fa XY |
1728 | if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10) |
1729 | amdgpu_discovery_get_gfx_info(adev); | |
1730 | ||
1884734a | 1731 | amdgpu_amdkfd_device_probe(adev); |
1732 | ||
3149d9da XY |
1733 | if (amdgpu_sriov_vf(adev)) { |
1734 | r = amdgpu_virt_request_full_gpu(adev, true); | |
1735 | if (r) | |
5ffa61c1 | 1736 | return -EAGAIN; |
3149d9da XY |
1737 | } |
1738 | ||
3b94fb10 | 1739 | adev->pm.pp_feature = amdgpu_pp_feature_mask; |
a35ad98b | 1740 | if (amdgpu_sriov_vf(adev) || sched_policy == KFD_SCHED_POLICY_NO_HWS) |
00544006 | 1741 | adev->pm.pp_feature &= ~PP_GFXOFF_MASK; |
00f54b97 | 1742 | |
d38ceaf9 AD |
1743 | for (i = 0; i < adev->num_ip_blocks; i++) { |
1744 | if ((amdgpu_ip_block_mask & (1 << i)) == 0) { | |
ed8cf00c HR |
1745 | DRM_ERROR("disabled ip block: %d <%s>\n", |
1746 | i, adev->ip_blocks[i].version->funcs->name); | |
a1255107 | 1747 | adev->ip_blocks[i].status.valid = false; |
d38ceaf9 | 1748 | } else { |
a1255107 AD |
1749 | if (adev->ip_blocks[i].version->funcs->early_init) { |
1750 | r = adev->ip_blocks[i].version->funcs->early_init((void *)adev); | |
2c1a2784 | 1751 | if (r == -ENOENT) { |
a1255107 | 1752 | adev->ip_blocks[i].status.valid = false; |
2c1a2784 | 1753 | } else if (r) { |
a1255107 AD |
1754 | DRM_ERROR("early_init of IP block <%s> failed %d\n", |
1755 | adev->ip_blocks[i].version->funcs->name, r); | |
d38ceaf9 | 1756 | return r; |
2c1a2784 | 1757 | } else { |
a1255107 | 1758 | adev->ip_blocks[i].status.valid = true; |
2c1a2784 | 1759 | } |
974e6b64 | 1760 | } else { |
a1255107 | 1761 | adev->ip_blocks[i].status.valid = true; |
d38ceaf9 | 1762 | } |
d38ceaf9 | 1763 | } |
21a249ca AD |
1764 | /* get the vbios after the asic_funcs are set up */ |
1765 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) { | |
1766 | /* Read BIOS */ | |
1767 | if (!amdgpu_get_bios(adev)) | |
1768 | return -EINVAL; | |
1769 | ||
1770 | r = amdgpu_atombios_init(adev); | |
1771 | if (r) { | |
1772 | dev_err(adev->dev, "amdgpu_atombios_init failed\n"); | |
1773 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0); | |
1774 | return r; | |
1775 | } | |
1776 | } | |
d38ceaf9 AD |
1777 | } |
1778 | ||
395d1fb9 NH |
1779 | adev->cg_flags &= amdgpu_cg_mask; |
1780 | adev->pg_flags &= amdgpu_pg_mask; | |
1781 | ||
d38ceaf9 AD |
1782 | return 0; |
1783 | } | |
1784 | ||
0a4f2520 RZ |
1785 | static int amdgpu_device_ip_hw_init_phase1(struct amdgpu_device *adev) |
1786 | { | |
1787 | int i, r; | |
1788 | ||
1789 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
1790 | if (!adev->ip_blocks[i].status.sw) | |
1791 | continue; | |
1792 | if (adev->ip_blocks[i].status.hw) | |
1793 | continue; | |
1794 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || | |
2d11fd3f | 1795 | (amdgpu_sriov_vf(adev) && (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)) || |
0a4f2520 RZ |
1796 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) { |
1797 | r = adev->ip_blocks[i].version->funcs->hw_init(adev); | |
1798 | if (r) { | |
1799 | DRM_ERROR("hw_init of IP block <%s> failed %d\n", | |
1800 | adev->ip_blocks[i].version->funcs->name, r); | |
1801 | return r; | |
1802 | } | |
1803 | adev->ip_blocks[i].status.hw = true; | |
1804 | } | |
1805 | } | |
1806 | ||
1807 | return 0; | |
1808 | } | |
1809 | ||
1810 | static int amdgpu_device_ip_hw_init_phase2(struct amdgpu_device *adev) | |
1811 | { | |
1812 | int i, r; | |
1813 | ||
1814 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
1815 | if (!adev->ip_blocks[i].status.sw) | |
1816 | continue; | |
1817 | if (adev->ip_blocks[i].status.hw) | |
1818 | continue; | |
1819 | r = adev->ip_blocks[i].version->funcs->hw_init(adev); | |
1820 | if (r) { | |
1821 | DRM_ERROR("hw_init of IP block <%s> failed %d\n", | |
1822 | adev->ip_blocks[i].version->funcs->name, r); | |
1823 | return r; | |
1824 | } | |
1825 | adev->ip_blocks[i].status.hw = true; | |
1826 | } | |
1827 | ||
1828 | return 0; | |
1829 | } | |
1830 | ||
7a3e0bb2 RZ |
1831 | static int amdgpu_device_fw_loading(struct amdgpu_device *adev) |
1832 | { | |
1833 | int r = 0; | |
1834 | int i; | |
80f41f84 | 1835 | uint32_t smu_version; |
7a3e0bb2 RZ |
1836 | |
1837 | if (adev->asic_type >= CHIP_VEGA10) { | |
1838 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
482f0e53 ML |
1839 | if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_PSP) |
1840 | continue; | |
1841 | ||
1842 | /* no need to do the fw loading again if already done*/ | |
1843 | if (adev->ip_blocks[i].status.hw == true) | |
1844 | break; | |
1845 | ||
1846 | if (adev->in_gpu_reset || adev->in_suspend) { | |
1847 | r = adev->ip_blocks[i].version->funcs->resume(adev); | |
1848 | if (r) { | |
1849 | DRM_ERROR("resume of IP block <%s> failed %d\n", | |
7a3e0bb2 | 1850 | adev->ip_blocks[i].version->funcs->name, r); |
482f0e53 ML |
1851 | return r; |
1852 | } | |
1853 | } else { | |
1854 | r = adev->ip_blocks[i].version->funcs->hw_init(adev); | |
1855 | if (r) { | |
1856 | DRM_ERROR("hw_init of IP block <%s> failed %d\n", | |
1857 | adev->ip_blocks[i].version->funcs->name, r); | |
1858 | return r; | |
7a3e0bb2 | 1859 | } |
7a3e0bb2 | 1860 | } |
482f0e53 ML |
1861 | |
1862 | adev->ip_blocks[i].status.hw = true; | |
1863 | break; | |
7a3e0bb2 RZ |
1864 | } |
1865 | } | |
482f0e53 | 1866 | |
8973d9ec ED |
1867 | if (!amdgpu_sriov_vf(adev) || adev->asic_type == CHIP_TONGA) |
1868 | r = amdgpu_pm_load_smu_firmware(adev, &smu_version); | |
7a3e0bb2 | 1869 | |
80f41f84 | 1870 | return r; |
7a3e0bb2 RZ |
1871 | } |
1872 | ||
e3ecdffa AD |
1873 | /** |
1874 | * amdgpu_device_ip_init - run init for hardware IPs | |
1875 | * | |
1876 | * @adev: amdgpu_device pointer | |
1877 | * | |
1878 | * Main initialization pass for hardware IPs. The list of all the hardware | |
1879 | * IPs that make up the asic is walked and the sw_init and hw_init callbacks | |
1880 | * are run. sw_init initializes the software state associated with each IP | |
1881 | * and hw_init initializes the hardware associated with each IP. | |
1882 | * Returns 0 on success, negative error code on failure. | |
1883 | */ | |
06ec9070 | 1884 | static int amdgpu_device_ip_init(struct amdgpu_device *adev) |
d38ceaf9 AD |
1885 | { |
1886 | int i, r; | |
1887 | ||
c030f2e4 | 1888 | r = amdgpu_ras_init(adev); |
1889 | if (r) | |
1890 | return r; | |
1891 | ||
d38ceaf9 | 1892 | for (i = 0; i < adev->num_ip_blocks; i++) { |
a1255107 | 1893 | if (!adev->ip_blocks[i].status.valid) |
d38ceaf9 | 1894 | continue; |
a1255107 | 1895 | r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev); |
2c1a2784 | 1896 | if (r) { |
a1255107 AD |
1897 | DRM_ERROR("sw_init of IP block <%s> failed %d\n", |
1898 | adev->ip_blocks[i].version->funcs->name, r); | |
72d3f592 | 1899 | goto init_failed; |
2c1a2784 | 1900 | } |
a1255107 | 1901 | adev->ip_blocks[i].status.sw = true; |
bfca0289 | 1902 | |
d38ceaf9 | 1903 | /* need to do gmc hw init early so we can allocate gpu mem */ |
a1255107 | 1904 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) { |
06ec9070 | 1905 | r = amdgpu_device_vram_scratch_init(adev); |
2c1a2784 AD |
1906 | if (r) { |
1907 | DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r); | |
72d3f592 | 1908 | goto init_failed; |
2c1a2784 | 1909 | } |
a1255107 | 1910 | r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev); |
2c1a2784 AD |
1911 | if (r) { |
1912 | DRM_ERROR("hw_init %d failed %d\n", i, r); | |
72d3f592 | 1913 | goto init_failed; |
2c1a2784 | 1914 | } |
06ec9070 | 1915 | r = amdgpu_device_wb_init(adev); |
2c1a2784 | 1916 | if (r) { |
06ec9070 | 1917 | DRM_ERROR("amdgpu_device_wb_init failed %d\n", r); |
72d3f592 | 1918 | goto init_failed; |
2c1a2784 | 1919 | } |
a1255107 | 1920 | adev->ip_blocks[i].status.hw = true; |
2493664f ML |
1921 | |
1922 | /* right after GMC hw init, we create CSA */ | |
f92d5c61 | 1923 | if (amdgpu_mcbp || amdgpu_sriov_vf(adev)) { |
1e256e27 RZ |
1924 | r = amdgpu_allocate_static_csa(adev, &adev->virt.csa_obj, |
1925 | AMDGPU_GEM_DOMAIN_VRAM, | |
1926 | AMDGPU_CSA_SIZE); | |
2493664f ML |
1927 | if (r) { |
1928 | DRM_ERROR("allocate CSA failed %d\n", r); | |
72d3f592 | 1929 | goto init_failed; |
2493664f ML |
1930 | } |
1931 | } | |
d38ceaf9 AD |
1932 | } |
1933 | } | |
1934 | ||
c9ffa427 YT |
1935 | if (amdgpu_sriov_vf(adev)) |
1936 | amdgpu_virt_init_data_exchange(adev); | |
1937 | ||
533aed27 AG |
1938 | r = amdgpu_ib_pool_init(adev); |
1939 | if (r) { | |
1940 | dev_err(adev->dev, "IB initialization failed (%d).\n", r); | |
1941 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r); | |
1942 | goto init_failed; | |
1943 | } | |
1944 | ||
c8963ea4 RZ |
1945 | r = amdgpu_ucode_create_bo(adev); /* create ucode bo when sw_init complete*/ |
1946 | if (r) | |
72d3f592 | 1947 | goto init_failed; |
0a4f2520 RZ |
1948 | |
1949 | r = amdgpu_device_ip_hw_init_phase1(adev); | |
1950 | if (r) | |
72d3f592 | 1951 | goto init_failed; |
0a4f2520 | 1952 | |
7a3e0bb2 RZ |
1953 | r = amdgpu_device_fw_loading(adev); |
1954 | if (r) | |
72d3f592 | 1955 | goto init_failed; |
7a3e0bb2 | 1956 | |
0a4f2520 RZ |
1957 | r = amdgpu_device_ip_hw_init_phase2(adev); |
1958 | if (r) | |
72d3f592 | 1959 | goto init_failed; |
d38ceaf9 | 1960 | |
121a2bc6 AG |
1961 | /* |
1962 | * retired pages will be loaded from eeprom and reserved here, | |
1963 | * it should be called after amdgpu_device_ip_hw_init_phase2 since | |
1964 | * for some ASICs the RAS EEPROM code relies on SMU fully functioning | |
1965 | * for I2C communication which only true at this point. | |
1966 | * recovery_init may fail, but it can free all resources allocated by | |
1967 | * itself and its failure should not stop amdgpu init process. | |
1968 | * | |
1969 | * Note: theoretically, this should be called before all vram allocations | |
1970 | * to protect retired page from abusing | |
1971 | */ | |
1972 | amdgpu_ras_recovery_init(adev); | |
1973 | ||
3e2e2ab5 HZ |
1974 | if (adev->gmc.xgmi.num_physical_nodes > 1) |
1975 | amdgpu_xgmi_add_device(adev); | |
1884734a | 1976 | amdgpu_amdkfd_device_init(adev); |
c6332b97 | 1977 | |
72d3f592 | 1978 | init_failed: |
c9ffa427 | 1979 | if (amdgpu_sriov_vf(adev)) |
c6332b97 | 1980 | amdgpu_virt_release_full_gpu(adev, true); |
1981 | ||
72d3f592 | 1982 | return r; |
d38ceaf9 AD |
1983 | } |
1984 | ||
e3ecdffa AD |
1985 | /** |
1986 | * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer | |
1987 | * | |
1988 | * @adev: amdgpu_device pointer | |
1989 | * | |
1990 | * Writes a reset magic value to the gart pointer in VRAM. The driver calls | |
1991 | * this function before a GPU reset. If the value is retained after a | |
1992 | * GPU reset, VRAM has not been lost. Some GPU resets may destry VRAM contents. | |
1993 | */ | |
06ec9070 | 1994 | static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev) |
0c49e0b8 CZ |
1995 | { |
1996 | memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM); | |
1997 | } | |
1998 | ||
e3ecdffa AD |
1999 | /** |
2000 | * amdgpu_device_check_vram_lost - check if vram is valid | |
2001 | * | |
2002 | * @adev: amdgpu_device pointer | |
2003 | * | |
2004 | * Checks the reset magic value written to the gart pointer in VRAM. | |
2005 | * The driver calls this after a GPU reset to see if the contents of | |
2006 | * VRAM is lost or now. | |
2007 | * returns true if vram is lost, false if not. | |
2008 | */ | |
06ec9070 | 2009 | static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev) |
0c49e0b8 CZ |
2010 | { |
2011 | return !!memcmp(adev->gart.ptr, adev->reset_magic, | |
2012 | AMDGPU_RESET_MAGIC_NUM); | |
2013 | } | |
2014 | ||
e3ecdffa | 2015 | /** |
1112a46b | 2016 | * amdgpu_device_set_cg_state - set clockgating for amdgpu device |
e3ecdffa AD |
2017 | * |
2018 | * @adev: amdgpu_device pointer | |
b8b72130 | 2019 | * @state: clockgating state (gate or ungate) |
e3ecdffa | 2020 | * |
e3ecdffa | 2021 | * The list of all the hardware IPs that make up the asic is walked and the |
1112a46b RZ |
2022 | * set_clockgating_state callbacks are run. |
2023 | * Late initialization pass enabling clockgating for hardware IPs. | |
2024 | * Fini or suspend, pass disabling clockgating for hardware IPs. | |
e3ecdffa AD |
2025 | * Returns 0 on success, negative error code on failure. |
2026 | */ | |
fdd34271 | 2027 | |
1112a46b RZ |
2028 | static int amdgpu_device_set_cg_state(struct amdgpu_device *adev, |
2029 | enum amd_clockgating_state state) | |
d38ceaf9 | 2030 | { |
1112a46b | 2031 | int i, j, r; |
d38ceaf9 | 2032 | |
4a2ba394 SL |
2033 | if (amdgpu_emu_mode == 1) |
2034 | return 0; | |
2035 | ||
1112a46b RZ |
2036 | for (j = 0; j < adev->num_ip_blocks; j++) { |
2037 | i = state == AMD_CG_STATE_GATE ? j : adev->num_ip_blocks - j - 1; | |
a2d31dc3 | 2038 | if (!adev->ip_blocks[i].status.late_initialized) |
d38ceaf9 | 2039 | continue; |
4a446d55 | 2040 | /* skip CG for VCE/UVD, it's handled specially */ |
a1255107 | 2041 | if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD && |
57716327 | 2042 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE && |
34319b32 | 2043 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN && |
52f2e779 | 2044 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG && |
57716327 | 2045 | adev->ip_blocks[i].version->funcs->set_clockgating_state) { |
4a446d55 | 2046 | /* enable clockgating to save power */ |
a1255107 | 2047 | r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev, |
1112a46b | 2048 | state); |
4a446d55 AD |
2049 | if (r) { |
2050 | DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n", | |
a1255107 | 2051 | adev->ip_blocks[i].version->funcs->name, r); |
4a446d55 AD |
2052 | return r; |
2053 | } | |
b0b00ff1 | 2054 | } |
d38ceaf9 | 2055 | } |
06b18f61 | 2056 | |
c9f96fd5 RZ |
2057 | return 0; |
2058 | } | |
2059 | ||
1112a46b | 2060 | static int amdgpu_device_set_pg_state(struct amdgpu_device *adev, enum amd_powergating_state state) |
c9f96fd5 | 2061 | { |
1112a46b | 2062 | int i, j, r; |
06b18f61 | 2063 | |
c9f96fd5 RZ |
2064 | if (amdgpu_emu_mode == 1) |
2065 | return 0; | |
2066 | ||
1112a46b RZ |
2067 | for (j = 0; j < adev->num_ip_blocks; j++) { |
2068 | i = state == AMD_PG_STATE_GATE ? j : adev->num_ip_blocks - j - 1; | |
a2d31dc3 | 2069 | if (!adev->ip_blocks[i].status.late_initialized) |
c9f96fd5 RZ |
2070 | continue; |
2071 | /* skip CG for VCE/UVD, it's handled specially */ | |
2072 | if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD && | |
2073 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE && | |
2074 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN && | |
52f2e779 | 2075 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG && |
c9f96fd5 RZ |
2076 | adev->ip_blocks[i].version->funcs->set_powergating_state) { |
2077 | /* enable powergating to save power */ | |
2078 | r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev, | |
1112a46b | 2079 | state); |
c9f96fd5 RZ |
2080 | if (r) { |
2081 | DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n", | |
2082 | adev->ip_blocks[i].version->funcs->name, r); | |
2083 | return r; | |
2084 | } | |
2085 | } | |
2086 | } | |
2dc80b00 S |
2087 | return 0; |
2088 | } | |
2089 | ||
beff74bc AD |
2090 | static int amdgpu_device_enable_mgpu_fan_boost(void) |
2091 | { | |
2092 | struct amdgpu_gpu_instance *gpu_ins; | |
2093 | struct amdgpu_device *adev; | |
2094 | int i, ret = 0; | |
2095 | ||
2096 | mutex_lock(&mgpu_info.mutex); | |
2097 | ||
2098 | /* | |
2099 | * MGPU fan boost feature should be enabled | |
2100 | * only when there are two or more dGPUs in | |
2101 | * the system | |
2102 | */ | |
2103 | if (mgpu_info.num_dgpu < 2) | |
2104 | goto out; | |
2105 | ||
2106 | for (i = 0; i < mgpu_info.num_dgpu; i++) { | |
2107 | gpu_ins = &(mgpu_info.gpu_ins[i]); | |
2108 | adev = gpu_ins->adev; | |
2109 | if (!(adev->flags & AMD_IS_APU) && | |
2110 | !gpu_ins->mgpu_fan_enabled && | |
2111 | adev->powerplay.pp_funcs && | |
2112 | adev->powerplay.pp_funcs->enable_mgpu_fan_boost) { | |
2113 | ret = amdgpu_dpm_enable_mgpu_fan_boost(adev); | |
2114 | if (ret) | |
2115 | break; | |
2116 | ||
2117 | gpu_ins->mgpu_fan_enabled = 1; | |
2118 | } | |
2119 | } | |
2120 | ||
2121 | out: | |
2122 | mutex_unlock(&mgpu_info.mutex); | |
2123 | ||
2124 | return ret; | |
2125 | } | |
2126 | ||
e3ecdffa AD |
2127 | /** |
2128 | * amdgpu_device_ip_late_init - run late init for hardware IPs | |
2129 | * | |
2130 | * @adev: amdgpu_device pointer | |
2131 | * | |
2132 | * Late initialization pass for hardware IPs. The list of all the hardware | |
2133 | * IPs that make up the asic is walked and the late_init callbacks are run. | |
2134 | * late_init covers any special initialization that an IP requires | |
2135 | * after all of the have been initialized or something that needs to happen | |
2136 | * late in the init process. | |
2137 | * Returns 0 on success, negative error code on failure. | |
2138 | */ | |
06ec9070 | 2139 | static int amdgpu_device_ip_late_init(struct amdgpu_device *adev) |
2dc80b00 | 2140 | { |
60599a03 | 2141 | struct amdgpu_gpu_instance *gpu_instance; |
2dc80b00 S |
2142 | int i = 0, r; |
2143 | ||
2144 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
73f847db | 2145 | if (!adev->ip_blocks[i].status.hw) |
2dc80b00 S |
2146 | continue; |
2147 | if (adev->ip_blocks[i].version->funcs->late_init) { | |
2148 | r = adev->ip_blocks[i].version->funcs->late_init((void *)adev); | |
2149 | if (r) { | |
2150 | DRM_ERROR("late_init of IP block <%s> failed %d\n", | |
2151 | adev->ip_blocks[i].version->funcs->name, r); | |
2152 | return r; | |
2153 | } | |
2dc80b00 | 2154 | } |
73f847db | 2155 | adev->ip_blocks[i].status.late_initialized = true; |
2dc80b00 S |
2156 | } |
2157 | ||
1112a46b RZ |
2158 | amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE); |
2159 | amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE); | |
916ac57f | 2160 | |
06ec9070 | 2161 | amdgpu_device_fill_reset_magic(adev); |
d38ceaf9 | 2162 | |
beff74bc AD |
2163 | r = amdgpu_device_enable_mgpu_fan_boost(); |
2164 | if (r) | |
2165 | DRM_ERROR("enable mgpu fan boost failed (%d).\n", r); | |
2166 | ||
60599a03 EQ |
2167 | |
2168 | if (adev->gmc.xgmi.num_physical_nodes > 1) { | |
2169 | mutex_lock(&mgpu_info.mutex); | |
2170 | ||
2171 | /* | |
2172 | * Reset device p-state to low as this was booted with high. | |
2173 | * | |
2174 | * This should be performed only after all devices from the same | |
2175 | * hive get initialized. | |
2176 | * | |
2177 | * However, it's unknown how many device in the hive in advance. | |
2178 | * As this is counted one by one during devices initializations. | |
2179 | * | |
2180 | * So, we wait for all XGMI interlinked devices initialized. | |
2181 | * This may bring some delays as those devices may come from | |
2182 | * different hives. But that should be OK. | |
2183 | */ | |
2184 | if (mgpu_info.num_dgpu == adev->gmc.xgmi.num_physical_nodes) { | |
2185 | for (i = 0; i < mgpu_info.num_gpu; i++) { | |
2186 | gpu_instance = &(mgpu_info.gpu_ins[i]); | |
2187 | if (gpu_instance->adev->flags & AMD_IS_APU) | |
2188 | continue; | |
2189 | ||
2190 | r = amdgpu_xgmi_set_pstate(gpu_instance->adev, 0); | |
2191 | if (r) { | |
2192 | DRM_ERROR("pstate setting failed (%d).\n", r); | |
2193 | break; | |
2194 | } | |
2195 | } | |
2196 | } | |
2197 | ||
2198 | mutex_unlock(&mgpu_info.mutex); | |
2199 | } | |
2200 | ||
d38ceaf9 AD |
2201 | return 0; |
2202 | } | |
2203 | ||
e3ecdffa AD |
2204 | /** |
2205 | * amdgpu_device_ip_fini - run fini for hardware IPs | |
2206 | * | |
2207 | * @adev: amdgpu_device pointer | |
2208 | * | |
2209 | * Main teardown pass for hardware IPs. The list of all the hardware | |
2210 | * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks | |
2211 | * are run. hw_fini tears down the hardware associated with each IP | |
2212 | * and sw_fini tears down any software state associated with each IP. | |
2213 | * Returns 0 on success, negative error code on failure. | |
2214 | */ | |
06ec9070 | 2215 | static int amdgpu_device_ip_fini(struct amdgpu_device *adev) |
d38ceaf9 AD |
2216 | { |
2217 | int i, r; | |
2218 | ||
c030f2e4 | 2219 | amdgpu_ras_pre_fini(adev); |
2220 | ||
a82400b5 AG |
2221 | if (adev->gmc.xgmi.num_physical_nodes > 1) |
2222 | amdgpu_xgmi_remove_device(adev); | |
2223 | ||
1884734a | 2224 | amdgpu_amdkfd_device_fini(adev); |
05df1f01 RZ |
2225 | |
2226 | amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE); | |
fdd34271 RZ |
2227 | amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE); |
2228 | ||
3e96dbfd AD |
2229 | /* need to disable SMC first */ |
2230 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 2231 | if (!adev->ip_blocks[i].status.hw) |
3e96dbfd | 2232 | continue; |
fdd34271 | 2233 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) { |
a1255107 | 2234 | r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev); |
3e96dbfd AD |
2235 | /* XXX handle errors */ |
2236 | if (r) { | |
2237 | DRM_DEBUG("hw_fini of IP block <%s> failed %d\n", | |
a1255107 | 2238 | adev->ip_blocks[i].version->funcs->name, r); |
3e96dbfd | 2239 | } |
a1255107 | 2240 | adev->ip_blocks[i].status.hw = false; |
3e96dbfd AD |
2241 | break; |
2242 | } | |
2243 | } | |
2244 | ||
d38ceaf9 | 2245 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { |
a1255107 | 2246 | if (!adev->ip_blocks[i].status.hw) |
d38ceaf9 | 2247 | continue; |
8201a67a | 2248 | |
a1255107 | 2249 | r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev); |
d38ceaf9 | 2250 | /* XXX handle errors */ |
2c1a2784 | 2251 | if (r) { |
a1255107 AD |
2252 | DRM_DEBUG("hw_fini of IP block <%s> failed %d\n", |
2253 | adev->ip_blocks[i].version->funcs->name, r); | |
2c1a2784 | 2254 | } |
8201a67a | 2255 | |
a1255107 | 2256 | adev->ip_blocks[i].status.hw = false; |
d38ceaf9 AD |
2257 | } |
2258 | ||
9950cda2 | 2259 | |
d38ceaf9 | 2260 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { |
a1255107 | 2261 | if (!adev->ip_blocks[i].status.sw) |
d38ceaf9 | 2262 | continue; |
c12aba3a ML |
2263 | |
2264 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) { | |
c8963ea4 | 2265 | amdgpu_ucode_free_bo(adev); |
1e256e27 | 2266 | amdgpu_free_static_csa(&adev->virt.csa_obj); |
c12aba3a ML |
2267 | amdgpu_device_wb_fini(adev); |
2268 | amdgpu_device_vram_scratch_fini(adev); | |
533aed27 | 2269 | amdgpu_ib_pool_fini(adev); |
c12aba3a ML |
2270 | } |
2271 | ||
a1255107 | 2272 | r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev); |
d38ceaf9 | 2273 | /* XXX handle errors */ |
2c1a2784 | 2274 | if (r) { |
a1255107 AD |
2275 | DRM_DEBUG("sw_fini of IP block <%s> failed %d\n", |
2276 | adev->ip_blocks[i].version->funcs->name, r); | |
2c1a2784 | 2277 | } |
a1255107 AD |
2278 | adev->ip_blocks[i].status.sw = false; |
2279 | adev->ip_blocks[i].status.valid = false; | |
d38ceaf9 AD |
2280 | } |
2281 | ||
a6dcfd9c | 2282 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { |
a1255107 | 2283 | if (!adev->ip_blocks[i].status.late_initialized) |
8a2eef1d | 2284 | continue; |
a1255107 AD |
2285 | if (adev->ip_blocks[i].version->funcs->late_fini) |
2286 | adev->ip_blocks[i].version->funcs->late_fini((void *)adev); | |
2287 | adev->ip_blocks[i].status.late_initialized = false; | |
a6dcfd9c ML |
2288 | } |
2289 | ||
c030f2e4 | 2290 | amdgpu_ras_fini(adev); |
2291 | ||
030308fc | 2292 | if (amdgpu_sriov_vf(adev)) |
24136135 ML |
2293 | if (amdgpu_virt_release_full_gpu(adev, false)) |
2294 | DRM_ERROR("failed to release exclusive mode on fini\n"); | |
2493664f | 2295 | |
d38ceaf9 AD |
2296 | return 0; |
2297 | } | |
2298 | ||
e3ecdffa | 2299 | /** |
beff74bc | 2300 | * amdgpu_device_delayed_init_work_handler - work handler for IB tests |
e3ecdffa | 2301 | * |
1112a46b | 2302 | * @work: work_struct. |
e3ecdffa | 2303 | */ |
beff74bc | 2304 | static void amdgpu_device_delayed_init_work_handler(struct work_struct *work) |
2dc80b00 S |
2305 | { |
2306 | struct amdgpu_device *adev = | |
beff74bc | 2307 | container_of(work, struct amdgpu_device, delayed_init_work.work); |
916ac57f RZ |
2308 | int r; |
2309 | ||
2310 | r = amdgpu_ib_ring_tests(adev); | |
2311 | if (r) | |
2312 | DRM_ERROR("ib ring test failed (%d).\n", r); | |
2dc80b00 S |
2313 | } |
2314 | ||
1e317b99 RZ |
2315 | static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work) |
2316 | { | |
2317 | struct amdgpu_device *adev = | |
2318 | container_of(work, struct amdgpu_device, gfx.gfx_off_delay_work.work); | |
2319 | ||
2320 | mutex_lock(&adev->gfx.gfx_off_mutex); | |
2321 | if (!adev->gfx.gfx_off_state && !adev->gfx.gfx_off_req_count) { | |
2322 | if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, true)) | |
2323 | adev->gfx.gfx_off_state = true; | |
2324 | } | |
2325 | mutex_unlock(&adev->gfx.gfx_off_mutex); | |
2326 | } | |
2327 | ||
e3ecdffa | 2328 | /** |
e7854a03 | 2329 | * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1) |
e3ecdffa AD |
2330 | * |
2331 | * @adev: amdgpu_device pointer | |
2332 | * | |
2333 | * Main suspend function for hardware IPs. The list of all the hardware | |
2334 | * IPs that make up the asic is walked, clockgating is disabled and the | |
2335 | * suspend callbacks are run. suspend puts the hardware and software state | |
2336 | * in each IP into a state suitable for suspend. | |
2337 | * Returns 0 on success, negative error code on failure. | |
2338 | */ | |
e7854a03 AD |
2339 | static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev) |
2340 | { | |
2341 | int i, r; | |
2342 | ||
05df1f01 | 2343 | amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE); |
fdd34271 | 2344 | amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE); |
05df1f01 | 2345 | |
e7854a03 AD |
2346 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { |
2347 | if (!adev->ip_blocks[i].status.valid) | |
2348 | continue; | |
2349 | /* displays are handled separately */ | |
2350 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) { | |
e7854a03 AD |
2351 | /* XXX handle errors */ |
2352 | r = adev->ip_blocks[i].version->funcs->suspend(adev); | |
2353 | /* XXX handle errors */ | |
2354 | if (r) { | |
2355 | DRM_ERROR("suspend of IP block <%s> failed %d\n", | |
2356 | adev->ip_blocks[i].version->funcs->name, r); | |
482f0e53 | 2357 | return r; |
e7854a03 | 2358 | } |
482f0e53 | 2359 | adev->ip_blocks[i].status.hw = false; |
e7854a03 AD |
2360 | } |
2361 | } | |
2362 | ||
e7854a03 AD |
2363 | return 0; |
2364 | } | |
2365 | ||
2366 | /** | |
2367 | * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2) | |
2368 | * | |
2369 | * @adev: amdgpu_device pointer | |
2370 | * | |
2371 | * Main suspend function for hardware IPs. The list of all the hardware | |
2372 | * IPs that make up the asic is walked, clockgating is disabled and the | |
2373 | * suspend callbacks are run. suspend puts the hardware and software state | |
2374 | * in each IP into a state suitable for suspend. | |
2375 | * Returns 0 on success, negative error code on failure. | |
2376 | */ | |
2377 | static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev) | |
d38ceaf9 AD |
2378 | { |
2379 | int i, r; | |
2380 | ||
2381 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { | |
a1255107 | 2382 | if (!adev->ip_blocks[i].status.valid) |
d38ceaf9 | 2383 | continue; |
e7854a03 AD |
2384 | /* displays are handled in phase1 */ |
2385 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) | |
2386 | continue; | |
bff77e86 LM |
2387 | /* PSP lost connection when err_event_athub occurs */ |
2388 | if (amdgpu_ras_intr_triggered() && | |
2389 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) { | |
2390 | adev->ip_blocks[i].status.hw = false; | |
2391 | continue; | |
2392 | } | |
d38ceaf9 | 2393 | /* XXX handle errors */ |
a1255107 | 2394 | r = adev->ip_blocks[i].version->funcs->suspend(adev); |
d38ceaf9 | 2395 | /* XXX handle errors */ |
2c1a2784 | 2396 | if (r) { |
a1255107 AD |
2397 | DRM_ERROR("suspend of IP block <%s> failed %d\n", |
2398 | adev->ip_blocks[i].version->funcs->name, r); | |
2c1a2784 | 2399 | } |
876923fb | 2400 | adev->ip_blocks[i].status.hw = false; |
a3a09142 | 2401 | /* handle putting the SMC in the appropriate state */ |
86b93fd6 JZ |
2402 | if(!amdgpu_sriov_vf(adev)){ |
2403 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) { | |
2404 | r = amdgpu_dpm_set_mp1_state(adev, adev->mp1_state); | |
2405 | if (r) { | |
2406 | DRM_ERROR("SMC failed to set mp1 state %d, %d\n", | |
2407 | adev->mp1_state, r); | |
2408 | return r; | |
2409 | } | |
a3a09142 AD |
2410 | } |
2411 | } | |
b5507c7e | 2412 | adev->ip_blocks[i].status.hw = false; |
d38ceaf9 AD |
2413 | } |
2414 | ||
2415 | return 0; | |
2416 | } | |
2417 | ||
e7854a03 AD |
2418 | /** |
2419 | * amdgpu_device_ip_suspend - run suspend for hardware IPs | |
2420 | * | |
2421 | * @adev: amdgpu_device pointer | |
2422 | * | |
2423 | * Main suspend function for hardware IPs. The list of all the hardware | |
2424 | * IPs that make up the asic is walked, clockgating is disabled and the | |
2425 | * suspend callbacks are run. suspend puts the hardware and software state | |
2426 | * in each IP into a state suitable for suspend. | |
2427 | * Returns 0 on success, negative error code on failure. | |
2428 | */ | |
2429 | int amdgpu_device_ip_suspend(struct amdgpu_device *adev) | |
2430 | { | |
2431 | int r; | |
2432 | ||
e7819644 YT |
2433 | if (amdgpu_sriov_vf(adev)) |
2434 | amdgpu_virt_request_full_gpu(adev, false); | |
2435 | ||
e7854a03 AD |
2436 | r = amdgpu_device_ip_suspend_phase1(adev); |
2437 | if (r) | |
2438 | return r; | |
2439 | r = amdgpu_device_ip_suspend_phase2(adev); | |
2440 | ||
e7819644 YT |
2441 | if (amdgpu_sriov_vf(adev)) |
2442 | amdgpu_virt_release_full_gpu(adev, false); | |
2443 | ||
e7854a03 AD |
2444 | return r; |
2445 | } | |
2446 | ||
06ec9070 | 2447 | static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev) |
a90ad3c2 ML |
2448 | { |
2449 | int i, r; | |
2450 | ||
2cb681b6 ML |
2451 | static enum amd_ip_block_type ip_order[] = { |
2452 | AMD_IP_BLOCK_TYPE_GMC, | |
2453 | AMD_IP_BLOCK_TYPE_COMMON, | |
39186aef | 2454 | AMD_IP_BLOCK_TYPE_PSP, |
2cb681b6 ML |
2455 | AMD_IP_BLOCK_TYPE_IH, |
2456 | }; | |
a90ad3c2 | 2457 | |
2cb681b6 ML |
2458 | for (i = 0; i < ARRAY_SIZE(ip_order); i++) { |
2459 | int j; | |
2460 | struct amdgpu_ip_block *block; | |
a90ad3c2 | 2461 | |
2cb681b6 ML |
2462 | for (j = 0; j < adev->num_ip_blocks; j++) { |
2463 | block = &adev->ip_blocks[j]; | |
2464 | ||
482f0e53 | 2465 | block->status.hw = false; |
2cb681b6 ML |
2466 | if (block->version->type != ip_order[i] || |
2467 | !block->status.valid) | |
2468 | continue; | |
2469 | ||
2470 | r = block->version->funcs->hw_init(adev); | |
0aaeefcc | 2471 | DRM_INFO("RE-INIT-early: %s %s\n", block->version->funcs->name, r?"failed":"succeeded"); |
c41d1cf6 ML |
2472 | if (r) |
2473 | return r; | |
482f0e53 | 2474 | block->status.hw = true; |
a90ad3c2 ML |
2475 | } |
2476 | } | |
2477 | ||
2478 | return 0; | |
2479 | } | |
2480 | ||
06ec9070 | 2481 | static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev) |
a90ad3c2 ML |
2482 | { |
2483 | int i, r; | |
2484 | ||
2cb681b6 ML |
2485 | static enum amd_ip_block_type ip_order[] = { |
2486 | AMD_IP_BLOCK_TYPE_SMC, | |
2487 | AMD_IP_BLOCK_TYPE_DCE, | |
2488 | AMD_IP_BLOCK_TYPE_GFX, | |
2489 | AMD_IP_BLOCK_TYPE_SDMA, | |
257deb8c | 2490 | AMD_IP_BLOCK_TYPE_UVD, |
d83c7a07 JJ |
2491 | AMD_IP_BLOCK_TYPE_VCE, |
2492 | AMD_IP_BLOCK_TYPE_VCN | |
2cb681b6 | 2493 | }; |
a90ad3c2 | 2494 | |
2cb681b6 ML |
2495 | for (i = 0; i < ARRAY_SIZE(ip_order); i++) { |
2496 | int j; | |
2497 | struct amdgpu_ip_block *block; | |
a90ad3c2 | 2498 | |
2cb681b6 ML |
2499 | for (j = 0; j < adev->num_ip_blocks; j++) { |
2500 | block = &adev->ip_blocks[j]; | |
2501 | ||
2502 | if (block->version->type != ip_order[i] || | |
482f0e53 ML |
2503 | !block->status.valid || |
2504 | block->status.hw) | |
2cb681b6 ML |
2505 | continue; |
2506 | ||
895bd048 JZ |
2507 | if (block->version->type == AMD_IP_BLOCK_TYPE_SMC) |
2508 | r = block->version->funcs->resume(adev); | |
2509 | else | |
2510 | r = block->version->funcs->hw_init(adev); | |
2511 | ||
0aaeefcc | 2512 | DRM_INFO("RE-INIT-late: %s %s\n", block->version->funcs->name, r?"failed":"succeeded"); |
c41d1cf6 ML |
2513 | if (r) |
2514 | return r; | |
482f0e53 | 2515 | block->status.hw = true; |
a90ad3c2 ML |
2516 | } |
2517 | } | |
2518 | ||
2519 | return 0; | |
2520 | } | |
2521 | ||
e3ecdffa AD |
2522 | /** |
2523 | * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs | |
2524 | * | |
2525 | * @adev: amdgpu_device pointer | |
2526 | * | |
2527 | * First resume function for hardware IPs. The list of all the hardware | |
2528 | * IPs that make up the asic is walked and the resume callbacks are run for | |
2529 | * COMMON, GMC, and IH. resume puts the hardware into a functional state | |
2530 | * after a suspend and updates the software state as necessary. This | |
2531 | * function is also used for restoring the GPU after a GPU reset. | |
2532 | * Returns 0 on success, negative error code on failure. | |
2533 | */ | |
06ec9070 | 2534 | static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev) |
d38ceaf9 AD |
2535 | { |
2536 | int i, r; | |
2537 | ||
a90ad3c2 | 2538 | for (i = 0; i < adev->num_ip_blocks; i++) { |
482f0e53 | 2539 | if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw) |
a90ad3c2 | 2540 | continue; |
a90ad3c2 | 2541 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || |
e3ecdffa AD |
2542 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC || |
2543 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) { | |
482f0e53 | 2544 | |
fcf0649f CZ |
2545 | r = adev->ip_blocks[i].version->funcs->resume(adev); |
2546 | if (r) { | |
2547 | DRM_ERROR("resume of IP block <%s> failed %d\n", | |
2548 | adev->ip_blocks[i].version->funcs->name, r); | |
2549 | return r; | |
2550 | } | |
482f0e53 | 2551 | adev->ip_blocks[i].status.hw = true; |
a90ad3c2 ML |
2552 | } |
2553 | } | |
2554 | ||
2555 | return 0; | |
2556 | } | |
2557 | ||
e3ecdffa AD |
2558 | /** |
2559 | * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs | |
2560 | * | |
2561 | * @adev: amdgpu_device pointer | |
2562 | * | |
2563 | * First resume function for hardware IPs. The list of all the hardware | |
2564 | * IPs that make up the asic is walked and the resume callbacks are run for | |
2565 | * all blocks except COMMON, GMC, and IH. resume puts the hardware into a | |
2566 | * functional state after a suspend and updates the software state as | |
2567 | * necessary. This function is also used for restoring the GPU after a GPU | |
2568 | * reset. | |
2569 | * Returns 0 on success, negative error code on failure. | |
2570 | */ | |
06ec9070 | 2571 | static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev) |
d38ceaf9 AD |
2572 | { |
2573 | int i, r; | |
2574 | ||
2575 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
482f0e53 | 2576 | if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw) |
d38ceaf9 | 2577 | continue; |
fcf0649f | 2578 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || |
e3ecdffa | 2579 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC || |
7a3e0bb2 RZ |
2580 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH || |
2581 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) | |
fcf0649f | 2582 | continue; |
a1255107 | 2583 | r = adev->ip_blocks[i].version->funcs->resume(adev); |
2c1a2784 | 2584 | if (r) { |
a1255107 AD |
2585 | DRM_ERROR("resume of IP block <%s> failed %d\n", |
2586 | adev->ip_blocks[i].version->funcs->name, r); | |
d38ceaf9 | 2587 | return r; |
2c1a2784 | 2588 | } |
482f0e53 | 2589 | adev->ip_blocks[i].status.hw = true; |
d38ceaf9 AD |
2590 | } |
2591 | ||
2592 | return 0; | |
2593 | } | |
2594 | ||
e3ecdffa AD |
2595 | /** |
2596 | * amdgpu_device_ip_resume - run resume for hardware IPs | |
2597 | * | |
2598 | * @adev: amdgpu_device pointer | |
2599 | * | |
2600 | * Main resume function for hardware IPs. The hardware IPs | |
2601 | * are split into two resume functions because they are | |
2602 | * are also used in in recovering from a GPU reset and some additional | |
2603 | * steps need to be take between them. In this case (S3/S4) they are | |
2604 | * run sequentially. | |
2605 | * Returns 0 on success, negative error code on failure. | |
2606 | */ | |
06ec9070 | 2607 | static int amdgpu_device_ip_resume(struct amdgpu_device *adev) |
fcf0649f CZ |
2608 | { |
2609 | int r; | |
2610 | ||
06ec9070 | 2611 | r = amdgpu_device_ip_resume_phase1(adev); |
fcf0649f CZ |
2612 | if (r) |
2613 | return r; | |
7a3e0bb2 RZ |
2614 | |
2615 | r = amdgpu_device_fw_loading(adev); | |
2616 | if (r) | |
2617 | return r; | |
2618 | ||
06ec9070 | 2619 | r = amdgpu_device_ip_resume_phase2(adev); |
fcf0649f CZ |
2620 | |
2621 | return r; | |
2622 | } | |
2623 | ||
e3ecdffa AD |
2624 | /** |
2625 | * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV | |
2626 | * | |
2627 | * @adev: amdgpu_device pointer | |
2628 | * | |
2629 | * Query the VBIOS data tables to determine if the board supports SR-IOV. | |
2630 | */ | |
4e99a44e | 2631 | static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev) |
048765ad | 2632 | { |
6867e1b5 ML |
2633 | if (amdgpu_sriov_vf(adev)) { |
2634 | if (adev->is_atom_fw) { | |
2635 | if (amdgpu_atomfirmware_gpu_supports_virtualization(adev)) | |
2636 | adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS; | |
2637 | } else { | |
2638 | if (amdgpu_atombios_has_gpu_virtualization_table(adev)) | |
2639 | adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS; | |
2640 | } | |
2641 | ||
2642 | if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS)) | |
2643 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0); | |
a5bde2f9 | 2644 | } |
048765ad AR |
2645 | } |
2646 | ||
e3ecdffa AD |
2647 | /** |
2648 | * amdgpu_device_asic_has_dc_support - determine if DC supports the asic | |
2649 | * | |
2650 | * @asic_type: AMD asic type | |
2651 | * | |
2652 | * Check if there is DC (new modesetting infrastructre) support for an asic. | |
2653 | * returns true if DC has support, false if not. | |
2654 | */ | |
4562236b HW |
2655 | bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type) |
2656 | { | |
2657 | switch (asic_type) { | |
2658 | #if defined(CONFIG_DRM_AMD_DC) | |
2659 | case CHIP_BONAIRE: | |
0d6fbccb | 2660 | case CHIP_KAVERI: |
367e6687 AD |
2661 | case CHIP_KABINI: |
2662 | case CHIP_MULLINS: | |
d9fda248 HW |
2663 | /* |
2664 | * We have systems in the wild with these ASICs that require | |
2665 | * LVDS and VGA support which is not supported with DC. | |
2666 | * | |
2667 | * Fallback to the non-DC driver here by default so as not to | |
2668 | * cause regressions. | |
2669 | */ | |
2670 | return amdgpu_dc > 0; | |
2671 | case CHIP_HAWAII: | |
4562236b HW |
2672 | case CHIP_CARRIZO: |
2673 | case CHIP_STONEY: | |
4562236b | 2674 | case CHIP_POLARIS10: |
675fd32b | 2675 | case CHIP_POLARIS11: |
2c8ad2d5 | 2676 | case CHIP_POLARIS12: |
675fd32b | 2677 | case CHIP_VEGAM: |
4562236b HW |
2678 | case CHIP_TONGA: |
2679 | case CHIP_FIJI: | |
42f8ffa1 | 2680 | case CHIP_VEGA10: |
dca7b401 | 2681 | case CHIP_VEGA12: |
c6034aa2 | 2682 | case CHIP_VEGA20: |
b86a1aa3 | 2683 | #if defined(CONFIG_DRM_AMD_DC_DCN) |
fd187853 | 2684 | case CHIP_RAVEN: |
b4f199c7 | 2685 | case CHIP_NAVI10: |
8fceceb6 | 2686 | case CHIP_NAVI14: |
078655d9 | 2687 | case CHIP_NAVI12: |
e1c14c43 | 2688 | case CHIP_RENOIR: |
42f8ffa1 | 2689 | #endif |
fd187853 | 2690 | return amdgpu_dc != 0; |
4562236b HW |
2691 | #endif |
2692 | default: | |
93b09a9a SS |
2693 | if (amdgpu_dc > 0) |
2694 | DRM_INFO("Display Core has been requested via kernel parameter " | |
2695 | "but isn't supported by ASIC, ignoring\n"); | |
4562236b HW |
2696 | return false; |
2697 | } | |
2698 | } | |
2699 | ||
2700 | /** | |
2701 | * amdgpu_device_has_dc_support - check if dc is supported | |
2702 | * | |
2703 | * @adev: amdgpu_device_pointer | |
2704 | * | |
2705 | * Returns true for supported, false for not supported | |
2706 | */ | |
2707 | bool amdgpu_device_has_dc_support(struct amdgpu_device *adev) | |
2708 | { | |
2555039d XY |
2709 | if (amdgpu_sriov_vf(adev)) |
2710 | return false; | |
2711 | ||
4562236b HW |
2712 | return amdgpu_device_asic_has_dc_support(adev->asic_type); |
2713 | } | |
2714 | ||
d4535e2c AG |
2715 | |
2716 | static void amdgpu_device_xgmi_reset_func(struct work_struct *__work) | |
2717 | { | |
2718 | struct amdgpu_device *adev = | |
2719 | container_of(__work, struct amdgpu_device, xgmi_reset_work); | |
c6a6e2db | 2720 | struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev, 0); |
d4535e2c | 2721 | |
c6a6e2db AG |
2722 | /* It's a bug to not have a hive within this function */ |
2723 | if (WARN_ON(!hive)) | |
2724 | return; | |
2725 | ||
2726 | /* | |
2727 | * Use task barrier to synchronize all xgmi reset works across the | |
2728 | * hive. task_barrier_enter and task_barrier_exit will block | |
2729 | * until all the threads running the xgmi reset works reach | |
2730 | * those points. task_barrier_full will do both blocks. | |
2731 | */ | |
2732 | if (amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) { | |
2733 | ||
2734 | task_barrier_enter(&hive->tb); | |
2735 | adev->asic_reset_res = amdgpu_device_baco_enter(adev->ddev); | |
2736 | ||
2737 | if (adev->asic_reset_res) | |
2738 | goto fail; | |
2739 | ||
2740 | task_barrier_exit(&hive->tb); | |
2741 | adev->asic_reset_res = amdgpu_device_baco_exit(adev->ddev); | |
2742 | ||
2743 | if (adev->asic_reset_res) | |
2744 | goto fail; | |
2745 | } else { | |
2746 | ||
2747 | task_barrier_full(&hive->tb); | |
2748 | adev->asic_reset_res = amdgpu_asic_reset(adev); | |
2749 | } | |
ce316fa5 | 2750 | |
c6a6e2db | 2751 | fail: |
d4535e2c | 2752 | if (adev->asic_reset_res) |
fed184e9 | 2753 | DRM_WARN("ASIC reset failed with error, %d for drm dev, %s", |
d4535e2c AG |
2754 | adev->asic_reset_res, adev->ddev->unique); |
2755 | } | |
2756 | ||
71f98027 AD |
2757 | static int amdgpu_device_get_job_timeout_settings(struct amdgpu_device *adev) |
2758 | { | |
2759 | char *input = amdgpu_lockup_timeout; | |
2760 | char *timeout_setting = NULL; | |
2761 | int index = 0; | |
2762 | long timeout; | |
2763 | int ret = 0; | |
2764 | ||
2765 | /* | |
2766 | * By default timeout for non compute jobs is 10000. | |
2767 | * And there is no timeout enforced on compute jobs. | |
2768 | * In SR-IOV or passthrough mode, timeout for compute | |
2769 | * jobs are 10000 by default. | |
2770 | */ | |
2771 | adev->gfx_timeout = msecs_to_jiffies(10000); | |
2772 | adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout; | |
2773 | if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev)) | |
2774 | adev->compute_timeout = adev->gfx_timeout; | |
2775 | else | |
2776 | adev->compute_timeout = MAX_SCHEDULE_TIMEOUT; | |
2777 | ||
f440ff44 | 2778 | if (strnlen(input, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) { |
71f98027 | 2779 | while ((timeout_setting = strsep(&input, ",")) && |
f440ff44 | 2780 | strnlen(timeout_setting, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) { |
71f98027 AD |
2781 | ret = kstrtol(timeout_setting, 0, &timeout); |
2782 | if (ret) | |
2783 | return ret; | |
2784 | ||
2785 | if (timeout == 0) { | |
2786 | index++; | |
2787 | continue; | |
2788 | } else if (timeout < 0) { | |
2789 | timeout = MAX_SCHEDULE_TIMEOUT; | |
2790 | } else { | |
2791 | timeout = msecs_to_jiffies(timeout); | |
2792 | } | |
2793 | ||
2794 | switch (index++) { | |
2795 | case 0: | |
2796 | adev->gfx_timeout = timeout; | |
2797 | break; | |
2798 | case 1: | |
2799 | adev->compute_timeout = timeout; | |
2800 | break; | |
2801 | case 2: | |
2802 | adev->sdma_timeout = timeout; | |
2803 | break; | |
2804 | case 3: | |
2805 | adev->video_timeout = timeout; | |
2806 | break; | |
2807 | default: | |
2808 | break; | |
2809 | } | |
2810 | } | |
2811 | /* | |
2812 | * There is only one value specified and | |
2813 | * it should apply to all non-compute jobs. | |
2814 | */ | |
bcccee89 | 2815 | if (index == 1) { |
71f98027 | 2816 | adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout; |
bcccee89 ED |
2817 | if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev)) |
2818 | adev->compute_timeout = adev->gfx_timeout; | |
2819 | } | |
71f98027 AD |
2820 | } |
2821 | ||
2822 | return ret; | |
2823 | } | |
d4535e2c | 2824 | |
d38ceaf9 AD |
2825 | /** |
2826 | * amdgpu_device_init - initialize the driver | |
2827 | * | |
2828 | * @adev: amdgpu_device pointer | |
87e3f136 | 2829 | * @ddev: drm dev pointer |
d38ceaf9 AD |
2830 | * @pdev: pci dev pointer |
2831 | * @flags: driver flags | |
2832 | * | |
2833 | * Initializes the driver info and hw (all asics). | |
2834 | * Returns 0 for success or an error on failure. | |
2835 | * Called at driver startup. | |
2836 | */ | |
2837 | int amdgpu_device_init(struct amdgpu_device *adev, | |
2838 | struct drm_device *ddev, | |
2839 | struct pci_dev *pdev, | |
2840 | uint32_t flags) | |
2841 | { | |
2842 | int r, i; | |
3840c5bc | 2843 | bool boco = false; |
95844d20 | 2844 | u32 max_MBps; |
d38ceaf9 AD |
2845 | |
2846 | adev->shutdown = false; | |
2847 | adev->dev = &pdev->dev; | |
2848 | adev->ddev = ddev; | |
2849 | adev->pdev = pdev; | |
2850 | adev->flags = flags; | |
4e66d7d2 YZ |
2851 | |
2852 | if (amdgpu_force_asic_type >= 0 && amdgpu_force_asic_type < CHIP_LAST) | |
2853 | adev->asic_type = amdgpu_force_asic_type; | |
2854 | else | |
2855 | adev->asic_type = flags & AMD_ASIC_MASK; | |
2856 | ||
d38ceaf9 | 2857 | adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT; |
593aa2d2 | 2858 | if (amdgpu_emu_mode == 1) |
8bdab6bb | 2859 | adev->usec_timeout *= 10; |
770d13b1 | 2860 | adev->gmc.gart_size = 512 * 1024 * 1024; |
d38ceaf9 AD |
2861 | adev->accel_working = false; |
2862 | adev->num_rings = 0; | |
2863 | adev->mman.buffer_funcs = NULL; | |
2864 | adev->mman.buffer_funcs_ring = NULL; | |
2865 | adev->vm_manager.vm_pte_funcs = NULL; | |
0c88b430 | 2866 | adev->vm_manager.vm_pte_num_scheds = 0; |
132f34e4 | 2867 | adev->gmc.gmc_funcs = NULL; |
f54d1867 | 2868 | adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS); |
b8866c26 | 2869 | bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES); |
d38ceaf9 AD |
2870 | |
2871 | adev->smc_rreg = &amdgpu_invalid_rreg; | |
2872 | adev->smc_wreg = &amdgpu_invalid_wreg; | |
2873 | adev->pcie_rreg = &amdgpu_invalid_rreg; | |
2874 | adev->pcie_wreg = &amdgpu_invalid_wreg; | |
36b9a952 HR |
2875 | adev->pciep_rreg = &amdgpu_invalid_rreg; |
2876 | adev->pciep_wreg = &amdgpu_invalid_wreg; | |
4fa1c6a6 TZ |
2877 | adev->pcie_rreg64 = &amdgpu_invalid_rreg64; |
2878 | adev->pcie_wreg64 = &amdgpu_invalid_wreg64; | |
d38ceaf9 AD |
2879 | adev->uvd_ctx_rreg = &amdgpu_invalid_rreg; |
2880 | adev->uvd_ctx_wreg = &amdgpu_invalid_wreg; | |
2881 | adev->didt_rreg = &amdgpu_invalid_rreg; | |
2882 | adev->didt_wreg = &amdgpu_invalid_wreg; | |
ccdbb20a RZ |
2883 | adev->gc_cac_rreg = &amdgpu_invalid_rreg; |
2884 | adev->gc_cac_wreg = &amdgpu_invalid_wreg; | |
d38ceaf9 AD |
2885 | adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg; |
2886 | adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg; | |
2887 | ||
3e39ab90 AD |
2888 | DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n", |
2889 | amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device, | |
2890 | pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision); | |
d38ceaf9 AD |
2891 | |
2892 | /* mutex initialization are all done here so we | |
2893 | * can recall function without having locking issues */ | |
d38ceaf9 | 2894 | atomic_set(&adev->irq.ih.lock, 0); |
0e5ca0d1 | 2895 | mutex_init(&adev->firmware.mutex); |
d38ceaf9 AD |
2896 | mutex_init(&adev->pm.mutex); |
2897 | mutex_init(&adev->gfx.gpu_clock_mutex); | |
2898 | mutex_init(&adev->srbm_mutex); | |
b8866c26 | 2899 | mutex_init(&adev->gfx.pipe_reserve_mutex); |
d23ee13f | 2900 | mutex_init(&adev->gfx.gfx_off_mutex); |
d38ceaf9 | 2901 | mutex_init(&adev->grbm_idx_mutex); |
d38ceaf9 | 2902 | mutex_init(&adev->mn_lock); |
e23b74aa | 2903 | mutex_init(&adev->virt.vf_errors.lock); |
d38ceaf9 | 2904 | hash_init(adev->mn_hash); |
13a752e3 | 2905 | mutex_init(&adev->lock_reset); |
32eaeae0 | 2906 | mutex_init(&adev->psp.mutex); |
bd052211 | 2907 | mutex_init(&adev->notifier_lock); |
d38ceaf9 | 2908 | |
912dfc84 EQ |
2909 | r = amdgpu_device_check_arguments(adev); |
2910 | if (r) | |
2911 | return r; | |
d38ceaf9 | 2912 | |
d38ceaf9 AD |
2913 | spin_lock_init(&adev->mmio_idx_lock); |
2914 | spin_lock_init(&adev->smc_idx_lock); | |
2915 | spin_lock_init(&adev->pcie_idx_lock); | |
2916 | spin_lock_init(&adev->uvd_ctx_idx_lock); | |
2917 | spin_lock_init(&adev->didt_idx_lock); | |
ccdbb20a | 2918 | spin_lock_init(&adev->gc_cac_idx_lock); |
16abb5d2 | 2919 | spin_lock_init(&adev->se_cac_idx_lock); |
d38ceaf9 | 2920 | spin_lock_init(&adev->audio_endpt_idx_lock); |
95844d20 | 2921 | spin_lock_init(&adev->mm_stats.lock); |
d38ceaf9 | 2922 | |
0c4e7fa5 CZ |
2923 | INIT_LIST_HEAD(&adev->shadow_list); |
2924 | mutex_init(&adev->shadow_list_lock); | |
2925 | ||
795f2813 AR |
2926 | INIT_LIST_HEAD(&adev->ring_lru_list); |
2927 | spin_lock_init(&adev->ring_lru_list_lock); | |
2928 | ||
beff74bc AD |
2929 | INIT_DELAYED_WORK(&adev->delayed_init_work, |
2930 | amdgpu_device_delayed_init_work_handler); | |
1e317b99 RZ |
2931 | INIT_DELAYED_WORK(&adev->gfx.gfx_off_delay_work, |
2932 | amdgpu_device_delay_enable_gfx_off); | |
2dc80b00 | 2933 | |
d4535e2c AG |
2934 | INIT_WORK(&adev->xgmi_reset_work, amdgpu_device_xgmi_reset_func); |
2935 | ||
d23ee13f | 2936 | adev->gfx.gfx_off_req_count = 1; |
b1ddf548 RZ |
2937 | adev->pm.ac_power = power_supply_is_system_supplied() > 0 ? true : false; |
2938 | ||
0fa49558 AX |
2939 | /* Registers mapping */ |
2940 | /* TODO: block userspace mapping of io register */ | |
da69c161 KW |
2941 | if (adev->asic_type >= CHIP_BONAIRE) { |
2942 | adev->rmmio_base = pci_resource_start(adev->pdev, 5); | |
2943 | adev->rmmio_size = pci_resource_len(adev->pdev, 5); | |
2944 | } else { | |
2945 | adev->rmmio_base = pci_resource_start(adev->pdev, 2); | |
2946 | adev->rmmio_size = pci_resource_len(adev->pdev, 2); | |
2947 | } | |
d38ceaf9 | 2948 | |
d38ceaf9 AD |
2949 | adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size); |
2950 | if (adev->rmmio == NULL) { | |
2951 | return -ENOMEM; | |
2952 | } | |
2953 | DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base); | |
2954 | DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size); | |
2955 | ||
d38ceaf9 AD |
2956 | /* io port mapping */ |
2957 | for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) { | |
2958 | if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) { | |
2959 | adev->rio_mem_size = pci_resource_len(adev->pdev, i); | |
2960 | adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size); | |
2961 | break; | |
2962 | } | |
2963 | } | |
2964 | if (adev->rio_mem == NULL) | |
b64a18c5 | 2965 | DRM_INFO("PCI I/O BAR is not found.\n"); |
d38ceaf9 | 2966 | |
b2109d8e JX |
2967 | /* enable PCIE atomic ops */ |
2968 | r = pci_enable_atomic_ops_to_root(adev->pdev, | |
2969 | PCI_EXP_DEVCAP2_ATOMIC_COMP32 | | |
2970 | PCI_EXP_DEVCAP2_ATOMIC_COMP64); | |
2971 | if (r) { | |
2972 | adev->have_atomics_support = false; | |
2973 | DRM_INFO("PCIE atomic ops is not supported\n"); | |
2974 | } else { | |
2975 | adev->have_atomics_support = true; | |
2976 | } | |
2977 | ||
5494d864 AD |
2978 | amdgpu_device_get_pcie_info(adev); |
2979 | ||
b239c017 JX |
2980 | if (amdgpu_mcbp) |
2981 | DRM_INFO("MCBP is enabled\n"); | |
2982 | ||
5f84cc63 JX |
2983 | if (amdgpu_mes && adev->asic_type >= CHIP_NAVI10) |
2984 | adev->enable_mes = true; | |
2985 | ||
f54eeab4 | 2986 | if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10) { |
a190d1c7 XY |
2987 | r = amdgpu_discovery_init(adev); |
2988 | if (r) { | |
2989 | dev_err(adev->dev, "amdgpu_discovery_init failed\n"); | |
2990 | return r; | |
2991 | } | |
2992 | } | |
2993 | ||
d38ceaf9 | 2994 | /* early init functions */ |
06ec9070 | 2995 | r = amdgpu_device_ip_early_init(adev); |
d38ceaf9 AD |
2996 | if (r) |
2997 | return r; | |
2998 | ||
df99ac0f JZ |
2999 | r = amdgpu_device_get_job_timeout_settings(adev); |
3000 | if (r) { | |
3001 | dev_err(adev->dev, "invalid lockup_timeout parameter syntax\n"); | |
3002 | return r; | |
3003 | } | |
3004 | ||
6585661d OZ |
3005 | /* doorbell bar mapping and doorbell index init*/ |
3006 | amdgpu_device_doorbell_init(adev); | |
3007 | ||
d38ceaf9 AD |
3008 | /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */ |
3009 | /* this will fail for cards that aren't VGA class devices, just | |
3010 | * ignore it */ | |
06ec9070 | 3011 | vga_client_register(adev->pdev, adev, NULL, amdgpu_device_vga_set_decode); |
d38ceaf9 | 3012 | |
31af062a | 3013 | if (amdgpu_device_supports_boco(ddev)) |
3840c5bc AD |
3014 | boco = true; |
3015 | if (amdgpu_has_atpx() && | |
3016 | (amdgpu_is_atpx_hybrid() || | |
3017 | amdgpu_has_atpx_dgpu_power_cntl()) && | |
3018 | !pci_is_thunderbolt_attached(adev->pdev)) | |
84c8b22e | 3019 | vga_switcheroo_register_client(adev->pdev, |
3840c5bc AD |
3020 | &amdgpu_switcheroo_ops, boco); |
3021 | if (boco) | |
d38ceaf9 AD |
3022 | vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain); |
3023 | ||
9475a943 SL |
3024 | if (amdgpu_emu_mode == 1) { |
3025 | /* post the asic on emulation mode */ | |
3026 | emu_soc_asic_init(adev); | |
bfca0289 | 3027 | goto fence_driver_init; |
9475a943 | 3028 | } |
bfca0289 | 3029 | |
4e99a44e ML |
3030 | /* detect if we are with an SRIOV vbios */ |
3031 | amdgpu_device_detect_sriov_bios(adev); | |
048765ad | 3032 | |
95e8e59e AD |
3033 | /* check if we need to reset the asic |
3034 | * E.g., driver was not cleanly unloaded previously, etc. | |
3035 | */ | |
f14899fd | 3036 | if (!amdgpu_sriov_vf(adev) && amdgpu_asic_need_reset_on_init(adev)) { |
95e8e59e AD |
3037 | r = amdgpu_asic_reset(adev); |
3038 | if (r) { | |
3039 | dev_err(adev->dev, "asic reset on init failed\n"); | |
3040 | goto failed; | |
3041 | } | |
3042 | } | |
3043 | ||
d38ceaf9 | 3044 | /* Post card if necessary */ |
39c640c0 | 3045 | if (amdgpu_device_need_post(adev)) { |
d38ceaf9 | 3046 | if (!adev->bios) { |
bec86378 | 3047 | dev_err(adev->dev, "no vBIOS found\n"); |
83ba126a AD |
3048 | r = -EINVAL; |
3049 | goto failed; | |
d38ceaf9 | 3050 | } |
bec86378 | 3051 | DRM_INFO("GPU posting now...\n"); |
4e99a44e ML |
3052 | r = amdgpu_atom_asic_init(adev->mode_info.atom_context); |
3053 | if (r) { | |
3054 | dev_err(adev->dev, "gpu post error!\n"); | |
3055 | goto failed; | |
3056 | } | |
d38ceaf9 AD |
3057 | } |
3058 | ||
88b64e95 AD |
3059 | if (adev->is_atom_fw) { |
3060 | /* Initialize clocks */ | |
3061 | r = amdgpu_atomfirmware_get_clock_info(adev); | |
3062 | if (r) { | |
3063 | dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n"); | |
e23b74aa | 3064 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0); |
88b64e95 AD |
3065 | goto failed; |
3066 | } | |
3067 | } else { | |
a5bde2f9 AD |
3068 | /* Initialize clocks */ |
3069 | r = amdgpu_atombios_get_clock_info(adev); | |
3070 | if (r) { | |
3071 | dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n"); | |
e23b74aa | 3072 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0); |
89041940 | 3073 | goto failed; |
a5bde2f9 AD |
3074 | } |
3075 | /* init i2c buses */ | |
4562236b HW |
3076 | if (!amdgpu_device_has_dc_support(adev)) |
3077 | amdgpu_atombios_i2c_init(adev); | |
2c1a2784 | 3078 | } |
d38ceaf9 | 3079 | |
bfca0289 | 3080 | fence_driver_init: |
d38ceaf9 AD |
3081 | /* Fence driver */ |
3082 | r = amdgpu_fence_driver_init(adev); | |
2c1a2784 AD |
3083 | if (r) { |
3084 | dev_err(adev->dev, "amdgpu_fence_driver_init failed\n"); | |
e23b74aa | 3085 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0); |
83ba126a | 3086 | goto failed; |
2c1a2784 | 3087 | } |
d38ceaf9 AD |
3088 | |
3089 | /* init the mode config */ | |
3090 | drm_mode_config_init(adev->ddev); | |
3091 | ||
06ec9070 | 3092 | r = amdgpu_device_ip_init(adev); |
d38ceaf9 | 3093 | if (r) { |
8840a387 | 3094 | /* failed in exclusive mode due to timeout */ |
3095 | if (amdgpu_sriov_vf(adev) && | |
3096 | !amdgpu_sriov_runtime(adev) && | |
3097 | amdgpu_virt_mmio_blocked(adev) && | |
3098 | !amdgpu_virt_wait_reset(adev)) { | |
3099 | dev_err(adev->dev, "VF exclusive mode timeout\n"); | |
1daee8b4 PD |
3100 | /* Don't send request since VF is inactive. */ |
3101 | adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME; | |
3102 | adev->virt.ops = NULL; | |
8840a387 | 3103 | r = -EAGAIN; |
3104 | goto failed; | |
3105 | } | |
06ec9070 | 3106 | dev_err(adev->dev, "amdgpu_device_ip_init failed\n"); |
e23b74aa | 3107 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0); |
83ba126a | 3108 | goto failed; |
d38ceaf9 AD |
3109 | } |
3110 | ||
d7f72fe4 YZ |
3111 | DRM_DEBUG("SE %d, SH per SE %d, CU per SH %d, active_cu_number %d\n", |
3112 | adev->gfx.config.max_shader_engines, | |
3113 | adev->gfx.config.max_sh_per_se, | |
3114 | adev->gfx.config.max_cu_per_sh, | |
3115 | adev->gfx.cu_info.number); | |
3116 | ||
f880799d ND |
3117 | amdgpu_ctx_init_sched(adev); |
3118 | ||
d38ceaf9 AD |
3119 | adev->accel_working = true; |
3120 | ||
e59c0205 AX |
3121 | amdgpu_vm_check_compute_bug(adev); |
3122 | ||
95844d20 MO |
3123 | /* Initialize the buffer migration limit. */ |
3124 | if (amdgpu_moverate >= 0) | |
3125 | max_MBps = amdgpu_moverate; | |
3126 | else | |
3127 | max_MBps = 8; /* Allow 8 MB/s. */ | |
3128 | /* Get a log2 for easy divisions. */ | |
3129 | adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps)); | |
3130 | ||
9bc92b9c ML |
3131 | amdgpu_fbdev_init(adev); |
3132 | ||
d2f52ac8 | 3133 | r = amdgpu_pm_sysfs_init(adev); |
7c868b59 YT |
3134 | if (r) { |
3135 | adev->pm_sysfs_en = false; | |
d2f52ac8 | 3136 | DRM_ERROR("registering pm debugfs failed (%d).\n", r); |
7c868b59 YT |
3137 | } else |
3138 | adev->pm_sysfs_en = true; | |
d2f52ac8 | 3139 | |
5bb23532 | 3140 | r = amdgpu_ucode_sysfs_init(adev); |
7c868b59 YT |
3141 | if (r) { |
3142 | adev->ucode_sysfs_en = false; | |
5bb23532 | 3143 | DRM_ERROR("Creating firmware sysfs failed (%d).\n", r); |
7c868b59 YT |
3144 | } else |
3145 | adev->ucode_sysfs_en = true; | |
5bb23532 | 3146 | |
d38ceaf9 AD |
3147 | if ((amdgpu_testing & 1)) { |
3148 | if (adev->accel_working) | |
3149 | amdgpu_test_moves(adev); | |
3150 | else | |
3151 | DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n"); | |
3152 | } | |
d38ceaf9 AD |
3153 | if (amdgpu_benchmarking) { |
3154 | if (adev->accel_working) | |
3155 | amdgpu_benchmark(adev, amdgpu_benchmarking); | |
3156 | else | |
3157 | DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n"); | |
3158 | } | |
3159 | ||
b0adca4d EQ |
3160 | /* |
3161 | * Register gpu instance before amdgpu_device_enable_mgpu_fan_boost. | |
3162 | * Otherwise the mgpu fan boost feature will be skipped due to the | |
3163 | * gpu instance is counted less. | |
3164 | */ | |
3165 | amdgpu_register_gpu_instance(adev); | |
3166 | ||
d38ceaf9 AD |
3167 | /* enable clockgating, etc. after ib tests, etc. since some blocks require |
3168 | * explicit gating rather than handling it automatically. | |
3169 | */ | |
06ec9070 | 3170 | r = amdgpu_device_ip_late_init(adev); |
2c1a2784 | 3171 | if (r) { |
06ec9070 | 3172 | dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n"); |
e23b74aa | 3173 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r); |
83ba126a | 3174 | goto failed; |
2c1a2784 | 3175 | } |
d38ceaf9 | 3176 | |
108c6a63 | 3177 | /* must succeed. */ |
511fdbc3 | 3178 | amdgpu_ras_resume(adev); |
108c6a63 | 3179 | |
beff74bc AD |
3180 | queue_delayed_work(system_wq, &adev->delayed_init_work, |
3181 | msecs_to_jiffies(AMDGPU_RESUME_MS)); | |
3182 | ||
dcea6e65 KR |
3183 | r = device_create_file(adev->dev, &dev_attr_pcie_replay_count); |
3184 | if (r) { | |
3185 | dev_err(adev->dev, "Could not create pcie_replay_count"); | |
3186 | return r; | |
3187 | } | |
108c6a63 | 3188 | |
d155bef0 AB |
3189 | if (IS_ENABLED(CONFIG_PERF_EVENTS)) |
3190 | r = amdgpu_pmu_init(adev); | |
9c7c85f7 JK |
3191 | if (r) |
3192 | dev_err(adev->dev, "amdgpu_pmu_init failed\n"); | |
3193 | ||
d38ceaf9 | 3194 | return 0; |
83ba126a AD |
3195 | |
3196 | failed: | |
89041940 | 3197 | amdgpu_vf_error_trans_all(adev); |
3840c5bc | 3198 | if (boco) |
83ba126a | 3199 | vga_switcheroo_fini_domain_pm_ops(adev->dev); |
8840a387 | 3200 | |
83ba126a | 3201 | return r; |
d38ceaf9 AD |
3202 | } |
3203 | ||
d38ceaf9 AD |
3204 | /** |
3205 | * amdgpu_device_fini - tear down the driver | |
3206 | * | |
3207 | * @adev: amdgpu_device pointer | |
3208 | * | |
3209 | * Tear down the driver info (all asics). | |
3210 | * Called at driver shutdown. | |
3211 | */ | |
3212 | void amdgpu_device_fini(struct amdgpu_device *adev) | |
3213 | { | |
3214 | int r; | |
3215 | ||
3216 | DRM_INFO("amdgpu: finishing device.\n"); | |
9f875167 | 3217 | flush_delayed_work(&adev->delayed_init_work); |
d0d13fe8 | 3218 | adev->shutdown = true; |
9f875167 | 3219 | |
752c683d ML |
3220 | /* make sure IB test finished before entering exclusive mode |
3221 | * to avoid preemption on IB test | |
3222 | * */ | |
3223 | if (amdgpu_sriov_vf(adev)) | |
3224 | amdgpu_virt_request_full_gpu(adev, false); | |
3225 | ||
e5b03032 ML |
3226 | /* disable all interrupts */ |
3227 | amdgpu_irq_disable_all(adev); | |
ff97cba8 ML |
3228 | if (adev->mode_info.mode_config_initialized){ |
3229 | if (!amdgpu_device_has_dc_support(adev)) | |
c2d88e06 | 3230 | drm_helper_force_disable_all(adev->ddev); |
ff97cba8 ML |
3231 | else |
3232 | drm_atomic_helper_shutdown(adev->ddev); | |
3233 | } | |
d38ceaf9 | 3234 | amdgpu_fence_driver_fini(adev); |
7c868b59 YT |
3235 | if (adev->pm_sysfs_en) |
3236 | amdgpu_pm_sysfs_fini(adev); | |
d38ceaf9 | 3237 | amdgpu_fbdev_fini(adev); |
06ec9070 | 3238 | r = amdgpu_device_ip_fini(adev); |
ab4fe3e1 HR |
3239 | if (adev->firmware.gpu_info_fw) { |
3240 | release_firmware(adev->firmware.gpu_info_fw); | |
3241 | adev->firmware.gpu_info_fw = NULL; | |
3242 | } | |
d38ceaf9 AD |
3243 | adev->accel_working = false; |
3244 | /* free i2c buses */ | |
4562236b HW |
3245 | if (!amdgpu_device_has_dc_support(adev)) |
3246 | amdgpu_i2c_fini(adev); | |
bfca0289 SL |
3247 | |
3248 | if (amdgpu_emu_mode != 1) | |
3249 | amdgpu_atombios_fini(adev); | |
3250 | ||
d38ceaf9 AD |
3251 | kfree(adev->bios); |
3252 | adev->bios = NULL; | |
3840c5bc AD |
3253 | if (amdgpu_has_atpx() && |
3254 | (amdgpu_is_atpx_hybrid() || | |
3255 | amdgpu_has_atpx_dgpu_power_cntl()) && | |
3256 | !pci_is_thunderbolt_attached(adev->pdev)) | |
84c8b22e | 3257 | vga_switcheroo_unregister_client(adev->pdev); |
3840c5bc | 3258 | if (amdgpu_device_supports_boco(adev->ddev)) |
83ba126a | 3259 | vga_switcheroo_fini_domain_pm_ops(adev->dev); |
d38ceaf9 AD |
3260 | vga_client_register(adev->pdev, NULL, NULL, NULL); |
3261 | if (adev->rio_mem) | |
3262 | pci_iounmap(adev->pdev, adev->rio_mem); | |
3263 | adev->rio_mem = NULL; | |
3264 | iounmap(adev->rmmio); | |
3265 | adev->rmmio = NULL; | |
06ec9070 | 3266 | amdgpu_device_doorbell_fini(adev); |
e9bc1bf7 | 3267 | |
dcea6e65 | 3268 | device_remove_file(adev->dev, &dev_attr_pcie_replay_count); |
7c868b59 YT |
3269 | if (adev->ucode_sysfs_en) |
3270 | amdgpu_ucode_sysfs_fini(adev); | |
d155bef0 AB |
3271 | if (IS_ENABLED(CONFIG_PERF_EVENTS)) |
3272 | amdgpu_pmu_fini(adev); | |
f54eeab4 | 3273 | if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10) |
a190d1c7 | 3274 | amdgpu_discovery_fini(adev); |
d38ceaf9 AD |
3275 | } |
3276 | ||
3277 | ||
3278 | /* | |
3279 | * Suspend & resume. | |
3280 | */ | |
3281 | /** | |
810ddc3a | 3282 | * amdgpu_device_suspend - initiate device suspend |
d38ceaf9 | 3283 | * |
87e3f136 DP |
3284 | * @dev: drm dev pointer |
3285 | * @suspend: suspend state | |
3286 | * @fbcon : notify the fbdev of suspend | |
d38ceaf9 AD |
3287 | * |
3288 | * Puts the hw in the suspend state (all asics). | |
3289 | * Returns 0 for success or an error on failure. | |
3290 | * Called at driver suspend. | |
3291 | */ | |
de185019 | 3292 | int amdgpu_device_suspend(struct drm_device *dev, bool fbcon) |
d38ceaf9 AD |
3293 | { |
3294 | struct amdgpu_device *adev; | |
3295 | struct drm_crtc *crtc; | |
3296 | struct drm_connector *connector; | |
f8d2d39e | 3297 | struct drm_connector_list_iter iter; |
5ceb54c6 | 3298 | int r; |
d38ceaf9 AD |
3299 | |
3300 | if (dev == NULL || dev->dev_private == NULL) { | |
3301 | return -ENODEV; | |
3302 | } | |
3303 | ||
3304 | adev = dev->dev_private; | |
3305 | ||
3306 | if (dev->switch_power_state == DRM_SWITCH_POWER_OFF) | |
3307 | return 0; | |
3308 | ||
44779b43 | 3309 | adev->in_suspend = true; |
d38ceaf9 AD |
3310 | drm_kms_helper_poll_disable(dev); |
3311 | ||
5f818173 S |
3312 | if (fbcon) |
3313 | amdgpu_fbdev_set_suspend(adev, 1); | |
3314 | ||
beff74bc | 3315 | cancel_delayed_work_sync(&adev->delayed_init_work); |
a5459475 | 3316 | |
4562236b HW |
3317 | if (!amdgpu_device_has_dc_support(adev)) { |
3318 | /* turn off display hw */ | |
3319 | drm_modeset_lock_all(dev); | |
f8d2d39e LP |
3320 | drm_connector_list_iter_begin(dev, &iter); |
3321 | drm_for_each_connector_iter(connector, &iter) | |
3322 | drm_helper_connector_dpms(connector, | |
3323 | DRM_MODE_DPMS_OFF); | |
3324 | drm_connector_list_iter_end(&iter); | |
4562236b | 3325 | drm_modeset_unlock_all(dev); |
fe1053b7 AD |
3326 | /* unpin the front buffers and cursors */ |
3327 | list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) { | |
3328 | struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc); | |
3329 | struct drm_framebuffer *fb = crtc->primary->fb; | |
3330 | struct amdgpu_bo *robj; | |
3331 | ||
91334223 | 3332 | if (amdgpu_crtc->cursor_bo && !adev->enable_virtual_display) { |
fe1053b7 AD |
3333 | struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo); |
3334 | r = amdgpu_bo_reserve(aobj, true); | |
3335 | if (r == 0) { | |
3336 | amdgpu_bo_unpin(aobj); | |
3337 | amdgpu_bo_unreserve(aobj); | |
3338 | } | |
756e6880 | 3339 | } |
756e6880 | 3340 | |
fe1053b7 AD |
3341 | if (fb == NULL || fb->obj[0] == NULL) { |
3342 | continue; | |
3343 | } | |
3344 | robj = gem_to_amdgpu_bo(fb->obj[0]); | |
3345 | /* don't unpin kernel fb objects */ | |
3346 | if (!amdgpu_fbdev_robj_is_fb(adev, robj)) { | |
3347 | r = amdgpu_bo_reserve(robj, true); | |
3348 | if (r == 0) { | |
3349 | amdgpu_bo_unpin(robj); | |
3350 | amdgpu_bo_unreserve(robj); | |
3351 | } | |
d38ceaf9 AD |
3352 | } |
3353 | } | |
3354 | } | |
fe1053b7 | 3355 | |
9593f4d6 | 3356 | amdgpu_amdkfd_suspend(adev, !fbcon); |
fe1053b7 | 3357 | |
5e6932fe | 3358 | amdgpu_ras_suspend(adev); |
3359 | ||
fe1053b7 AD |
3360 | r = amdgpu_device_ip_suspend_phase1(adev); |
3361 | ||
d38ceaf9 AD |
3362 | /* evict vram memory */ |
3363 | amdgpu_bo_evict_vram(adev); | |
3364 | ||
5ceb54c6 | 3365 | amdgpu_fence_driver_suspend(adev); |
d38ceaf9 | 3366 | |
fe1053b7 | 3367 | r = amdgpu_device_ip_suspend_phase2(adev); |
d38ceaf9 | 3368 | |
a0a71e49 AD |
3369 | /* evict remaining vram memory |
3370 | * This second call to evict vram is to evict the gart page table | |
3371 | * using the CPU. | |
3372 | */ | |
d38ceaf9 AD |
3373 | amdgpu_bo_evict_vram(adev); |
3374 | ||
d38ceaf9 AD |
3375 | return 0; |
3376 | } | |
3377 | ||
3378 | /** | |
810ddc3a | 3379 | * amdgpu_device_resume - initiate device resume |
d38ceaf9 | 3380 | * |
87e3f136 DP |
3381 | * @dev: drm dev pointer |
3382 | * @resume: resume state | |
3383 | * @fbcon : notify the fbdev of resume | |
d38ceaf9 AD |
3384 | * |
3385 | * Bring the hw back to operating state (all asics). | |
3386 | * Returns 0 for success or an error on failure. | |
3387 | * Called at driver resume. | |
3388 | */ | |
de185019 | 3389 | int amdgpu_device_resume(struct drm_device *dev, bool fbcon) |
d38ceaf9 AD |
3390 | { |
3391 | struct drm_connector *connector; | |
f8d2d39e | 3392 | struct drm_connector_list_iter iter; |
d38ceaf9 | 3393 | struct amdgpu_device *adev = dev->dev_private; |
756e6880 | 3394 | struct drm_crtc *crtc; |
03161a6e | 3395 | int r = 0; |
d38ceaf9 AD |
3396 | |
3397 | if (dev->switch_power_state == DRM_SWITCH_POWER_OFF) | |
3398 | return 0; | |
3399 | ||
d38ceaf9 | 3400 | /* post card */ |
39c640c0 | 3401 | if (amdgpu_device_need_post(adev)) { |
74b0b157 | 3402 | r = amdgpu_atom_asic_init(adev->mode_info.atom_context); |
3403 | if (r) | |
3404 | DRM_ERROR("amdgpu asic init failed\n"); | |
3405 | } | |
d38ceaf9 | 3406 | |
06ec9070 | 3407 | r = amdgpu_device_ip_resume(adev); |
e6707218 | 3408 | if (r) { |
06ec9070 | 3409 | DRM_ERROR("amdgpu_device_ip_resume failed (%d).\n", r); |
4d3b9ae5 | 3410 | return r; |
e6707218 | 3411 | } |
5ceb54c6 AD |
3412 | amdgpu_fence_driver_resume(adev); |
3413 | ||
d38ceaf9 | 3414 | |
06ec9070 | 3415 | r = amdgpu_device_ip_late_init(adev); |
03161a6e | 3416 | if (r) |
4d3b9ae5 | 3417 | return r; |
d38ceaf9 | 3418 | |
beff74bc AD |
3419 | queue_delayed_work(system_wq, &adev->delayed_init_work, |
3420 | msecs_to_jiffies(AMDGPU_RESUME_MS)); | |
3421 | ||
fe1053b7 AD |
3422 | if (!amdgpu_device_has_dc_support(adev)) { |
3423 | /* pin cursors */ | |
3424 | list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) { | |
3425 | struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc); | |
3426 | ||
91334223 | 3427 | if (amdgpu_crtc->cursor_bo && !adev->enable_virtual_display) { |
fe1053b7 AD |
3428 | struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo); |
3429 | r = amdgpu_bo_reserve(aobj, true); | |
3430 | if (r == 0) { | |
3431 | r = amdgpu_bo_pin(aobj, AMDGPU_GEM_DOMAIN_VRAM); | |
3432 | if (r != 0) | |
3433 | DRM_ERROR("Failed to pin cursor BO (%d)\n", r); | |
3434 | amdgpu_crtc->cursor_addr = amdgpu_bo_gpu_offset(aobj); | |
3435 | amdgpu_bo_unreserve(aobj); | |
3436 | } | |
756e6880 AD |
3437 | } |
3438 | } | |
3439 | } | |
9593f4d6 | 3440 | r = amdgpu_amdkfd_resume(adev, !fbcon); |
ba997709 YZ |
3441 | if (r) |
3442 | return r; | |
756e6880 | 3443 | |
96a5d8d4 | 3444 | /* Make sure IB tests flushed */ |
beff74bc | 3445 | flush_delayed_work(&adev->delayed_init_work); |
96a5d8d4 | 3446 | |
d38ceaf9 AD |
3447 | /* blat the mode back in */ |
3448 | if (fbcon) { | |
4562236b HW |
3449 | if (!amdgpu_device_has_dc_support(adev)) { |
3450 | /* pre DCE11 */ | |
3451 | drm_helper_resume_force_mode(dev); | |
3452 | ||
3453 | /* turn on display hw */ | |
3454 | drm_modeset_lock_all(dev); | |
f8d2d39e LP |
3455 | |
3456 | drm_connector_list_iter_begin(dev, &iter); | |
3457 | drm_for_each_connector_iter(connector, &iter) | |
3458 | drm_helper_connector_dpms(connector, | |
3459 | DRM_MODE_DPMS_ON); | |
3460 | drm_connector_list_iter_end(&iter); | |
3461 | ||
4562236b | 3462 | drm_modeset_unlock_all(dev); |
d38ceaf9 | 3463 | } |
4d3b9ae5 | 3464 | amdgpu_fbdev_set_suspend(adev, 0); |
d38ceaf9 AD |
3465 | } |
3466 | ||
3467 | drm_kms_helper_poll_enable(dev); | |
23a1a9e5 | 3468 | |
5e6932fe | 3469 | amdgpu_ras_resume(adev); |
3470 | ||
23a1a9e5 L |
3471 | /* |
3472 | * Most of the connector probing functions try to acquire runtime pm | |
3473 | * refs to ensure that the GPU is powered on when connector polling is | |
3474 | * performed. Since we're calling this from a runtime PM callback, | |
3475 | * trying to acquire rpm refs will cause us to deadlock. | |
3476 | * | |
3477 | * Since we're guaranteed to be holding the rpm lock, it's safe to | |
3478 | * temporarily disable the rpm helpers so this doesn't deadlock us. | |
3479 | */ | |
3480 | #ifdef CONFIG_PM | |
3481 | dev->dev->power.disable_depth++; | |
3482 | #endif | |
4562236b HW |
3483 | if (!amdgpu_device_has_dc_support(adev)) |
3484 | drm_helper_hpd_irq_event(dev); | |
3485 | else | |
3486 | drm_kms_helper_hotplug_event(dev); | |
23a1a9e5 L |
3487 | #ifdef CONFIG_PM |
3488 | dev->dev->power.disable_depth--; | |
3489 | #endif | |
44779b43 RZ |
3490 | adev->in_suspend = false; |
3491 | ||
4d3b9ae5 | 3492 | return 0; |
d38ceaf9 AD |
3493 | } |
3494 | ||
e3ecdffa AD |
3495 | /** |
3496 | * amdgpu_device_ip_check_soft_reset - did soft reset succeed | |
3497 | * | |
3498 | * @adev: amdgpu_device pointer | |
3499 | * | |
3500 | * The list of all the hardware IPs that make up the asic is walked and | |
3501 | * the check_soft_reset callbacks are run. check_soft_reset determines | |
3502 | * if the asic is still hung or not. | |
3503 | * Returns true if any of the IPs are still in a hung state, false if not. | |
3504 | */ | |
06ec9070 | 3505 | static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev) |
63fbf42f CZ |
3506 | { |
3507 | int i; | |
3508 | bool asic_hang = false; | |
3509 | ||
f993d628 ML |
3510 | if (amdgpu_sriov_vf(adev)) |
3511 | return true; | |
3512 | ||
8bc04c29 AD |
3513 | if (amdgpu_asic_need_full_reset(adev)) |
3514 | return true; | |
3515 | ||
63fbf42f | 3516 | for (i = 0; i < adev->num_ip_blocks; i++) { |
a1255107 | 3517 | if (!adev->ip_blocks[i].status.valid) |
63fbf42f | 3518 | continue; |
a1255107 AD |
3519 | if (adev->ip_blocks[i].version->funcs->check_soft_reset) |
3520 | adev->ip_blocks[i].status.hang = | |
3521 | adev->ip_blocks[i].version->funcs->check_soft_reset(adev); | |
3522 | if (adev->ip_blocks[i].status.hang) { | |
3523 | DRM_INFO("IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name); | |
63fbf42f CZ |
3524 | asic_hang = true; |
3525 | } | |
3526 | } | |
3527 | return asic_hang; | |
3528 | } | |
3529 | ||
e3ecdffa AD |
3530 | /** |
3531 | * amdgpu_device_ip_pre_soft_reset - prepare for soft reset | |
3532 | * | |
3533 | * @adev: amdgpu_device pointer | |
3534 | * | |
3535 | * The list of all the hardware IPs that make up the asic is walked and the | |
3536 | * pre_soft_reset callbacks are run if the block is hung. pre_soft_reset | |
3537 | * handles any IP specific hardware or software state changes that are | |
3538 | * necessary for a soft reset to succeed. | |
3539 | * Returns 0 on success, negative error code on failure. | |
3540 | */ | |
06ec9070 | 3541 | static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev) |
d31a501e CZ |
3542 | { |
3543 | int i, r = 0; | |
3544 | ||
3545 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 3546 | if (!adev->ip_blocks[i].status.valid) |
d31a501e | 3547 | continue; |
a1255107 AD |
3548 | if (adev->ip_blocks[i].status.hang && |
3549 | adev->ip_blocks[i].version->funcs->pre_soft_reset) { | |
3550 | r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev); | |
d31a501e CZ |
3551 | if (r) |
3552 | return r; | |
3553 | } | |
3554 | } | |
3555 | ||
3556 | return 0; | |
3557 | } | |
3558 | ||
e3ecdffa AD |
3559 | /** |
3560 | * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed | |
3561 | * | |
3562 | * @adev: amdgpu_device pointer | |
3563 | * | |
3564 | * Some hardware IPs cannot be soft reset. If they are hung, a full gpu | |
3565 | * reset is necessary to recover. | |
3566 | * Returns true if a full asic reset is required, false if not. | |
3567 | */ | |
06ec9070 | 3568 | static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev) |
35d782fe | 3569 | { |
da146d3b AD |
3570 | int i; |
3571 | ||
8bc04c29 AD |
3572 | if (amdgpu_asic_need_full_reset(adev)) |
3573 | return true; | |
3574 | ||
da146d3b | 3575 | for (i = 0; i < adev->num_ip_blocks; i++) { |
a1255107 | 3576 | if (!adev->ip_blocks[i].status.valid) |
da146d3b | 3577 | continue; |
a1255107 AD |
3578 | if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) || |
3579 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) || | |
3580 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) || | |
98512bb8 KW |
3581 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) || |
3582 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) { | |
a1255107 | 3583 | if (adev->ip_blocks[i].status.hang) { |
da146d3b AD |
3584 | DRM_INFO("Some block need full reset!\n"); |
3585 | return true; | |
3586 | } | |
3587 | } | |
35d782fe CZ |
3588 | } |
3589 | return false; | |
3590 | } | |
3591 | ||
e3ecdffa AD |
3592 | /** |
3593 | * amdgpu_device_ip_soft_reset - do a soft reset | |
3594 | * | |
3595 | * @adev: amdgpu_device pointer | |
3596 | * | |
3597 | * The list of all the hardware IPs that make up the asic is walked and the | |
3598 | * soft_reset callbacks are run if the block is hung. soft_reset handles any | |
3599 | * IP specific hardware or software state changes that are necessary to soft | |
3600 | * reset the IP. | |
3601 | * Returns 0 on success, negative error code on failure. | |
3602 | */ | |
06ec9070 | 3603 | static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev) |
35d782fe CZ |
3604 | { |
3605 | int i, r = 0; | |
3606 | ||
3607 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 3608 | if (!adev->ip_blocks[i].status.valid) |
35d782fe | 3609 | continue; |
a1255107 AD |
3610 | if (adev->ip_blocks[i].status.hang && |
3611 | adev->ip_blocks[i].version->funcs->soft_reset) { | |
3612 | r = adev->ip_blocks[i].version->funcs->soft_reset(adev); | |
35d782fe CZ |
3613 | if (r) |
3614 | return r; | |
3615 | } | |
3616 | } | |
3617 | ||
3618 | return 0; | |
3619 | } | |
3620 | ||
e3ecdffa AD |
3621 | /** |
3622 | * amdgpu_device_ip_post_soft_reset - clean up from soft reset | |
3623 | * | |
3624 | * @adev: amdgpu_device pointer | |
3625 | * | |
3626 | * The list of all the hardware IPs that make up the asic is walked and the | |
3627 | * post_soft_reset callbacks are run if the asic was hung. post_soft_reset | |
3628 | * handles any IP specific hardware or software state changes that are | |
3629 | * necessary after the IP has been soft reset. | |
3630 | * Returns 0 on success, negative error code on failure. | |
3631 | */ | |
06ec9070 | 3632 | static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev) |
35d782fe CZ |
3633 | { |
3634 | int i, r = 0; | |
3635 | ||
3636 | for (i = 0; i < adev->num_ip_blocks; i++) { | |
a1255107 | 3637 | if (!adev->ip_blocks[i].status.valid) |
35d782fe | 3638 | continue; |
a1255107 AD |
3639 | if (adev->ip_blocks[i].status.hang && |
3640 | adev->ip_blocks[i].version->funcs->post_soft_reset) | |
3641 | r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev); | |
35d782fe CZ |
3642 | if (r) |
3643 | return r; | |
3644 | } | |
3645 | ||
3646 | return 0; | |
3647 | } | |
3648 | ||
e3ecdffa | 3649 | /** |
c33adbc7 | 3650 | * amdgpu_device_recover_vram - Recover some VRAM contents |
e3ecdffa AD |
3651 | * |
3652 | * @adev: amdgpu_device pointer | |
3653 | * | |
3654 | * Restores the contents of VRAM buffers from the shadows in GTT. Used to | |
3655 | * restore things like GPUVM page tables after a GPU reset where | |
3656 | * the contents of VRAM might be lost. | |
403009bf CK |
3657 | * |
3658 | * Returns: | |
3659 | * 0 on success, negative error code on failure. | |
e3ecdffa | 3660 | */ |
c33adbc7 | 3661 | static int amdgpu_device_recover_vram(struct amdgpu_device *adev) |
c41d1cf6 | 3662 | { |
c41d1cf6 | 3663 | struct dma_fence *fence = NULL, *next = NULL; |
403009bf CK |
3664 | struct amdgpu_bo *shadow; |
3665 | long r = 1, tmo; | |
c41d1cf6 ML |
3666 | |
3667 | if (amdgpu_sriov_runtime(adev)) | |
b045d3af | 3668 | tmo = msecs_to_jiffies(8000); |
c41d1cf6 ML |
3669 | else |
3670 | tmo = msecs_to_jiffies(100); | |
3671 | ||
3672 | DRM_INFO("recover vram bo from shadow start\n"); | |
3673 | mutex_lock(&adev->shadow_list_lock); | |
403009bf CK |
3674 | list_for_each_entry(shadow, &adev->shadow_list, shadow_list) { |
3675 | ||
3676 | /* No need to recover an evicted BO */ | |
3677 | if (shadow->tbo.mem.mem_type != TTM_PL_TT || | |
b575f10d | 3678 | shadow->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET || |
403009bf CK |
3679 | shadow->parent->tbo.mem.mem_type != TTM_PL_VRAM) |
3680 | continue; | |
3681 | ||
3682 | r = amdgpu_bo_restore_shadow(shadow, &next); | |
3683 | if (r) | |
3684 | break; | |
3685 | ||
c41d1cf6 | 3686 | if (fence) { |
1712fb1a | 3687 | tmo = dma_fence_wait_timeout(fence, false, tmo); |
403009bf CK |
3688 | dma_fence_put(fence); |
3689 | fence = next; | |
1712fb1a | 3690 | if (tmo == 0) { |
3691 | r = -ETIMEDOUT; | |
c41d1cf6 | 3692 | break; |
1712fb1a | 3693 | } else if (tmo < 0) { |
3694 | r = tmo; | |
3695 | break; | |
3696 | } | |
403009bf CK |
3697 | } else { |
3698 | fence = next; | |
c41d1cf6 | 3699 | } |
c41d1cf6 ML |
3700 | } |
3701 | mutex_unlock(&adev->shadow_list_lock); | |
3702 | ||
403009bf CK |
3703 | if (fence) |
3704 | tmo = dma_fence_wait_timeout(fence, false, tmo); | |
c41d1cf6 ML |
3705 | dma_fence_put(fence); |
3706 | ||
1712fb1a | 3707 | if (r < 0 || tmo <= 0) { |
3708 | DRM_ERROR("recover vram bo from shadow failed, r is %ld, tmo is %ld\n", r, tmo); | |
403009bf CK |
3709 | return -EIO; |
3710 | } | |
c41d1cf6 | 3711 | |
403009bf CK |
3712 | DRM_INFO("recover vram bo from shadow done\n"); |
3713 | return 0; | |
c41d1cf6 ML |
3714 | } |
3715 | ||
a90ad3c2 | 3716 | |
e3ecdffa | 3717 | /** |
06ec9070 | 3718 | * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf |
5740682e ML |
3719 | * |
3720 | * @adev: amdgpu device pointer | |
87e3f136 | 3721 | * @from_hypervisor: request from hypervisor |
5740682e ML |
3722 | * |
3723 | * do VF FLR and reinitialize Asic | |
3f48c681 | 3724 | * return 0 means succeeded otherwise failed |
e3ecdffa AD |
3725 | */ |
3726 | static int amdgpu_device_reset_sriov(struct amdgpu_device *adev, | |
3727 | bool from_hypervisor) | |
5740682e ML |
3728 | { |
3729 | int r; | |
3730 | ||
3731 | if (from_hypervisor) | |
3732 | r = amdgpu_virt_request_full_gpu(adev, true); | |
3733 | else | |
3734 | r = amdgpu_virt_reset_gpu(adev); | |
3735 | if (r) | |
3736 | return r; | |
a90ad3c2 ML |
3737 | |
3738 | /* Resume IP prior to SMC */ | |
06ec9070 | 3739 | r = amdgpu_device_ip_reinit_early_sriov(adev); |
5740682e ML |
3740 | if (r) |
3741 | goto error; | |
a90ad3c2 | 3742 | |
c9ffa427 | 3743 | amdgpu_virt_init_data_exchange(adev); |
a90ad3c2 | 3744 | /* we need recover gart prior to run SMC/CP/SDMA resume */ |
c1c7ce8f | 3745 | amdgpu_gtt_mgr_recover(&adev->mman.bdev.man[TTM_PL_TT]); |
a90ad3c2 | 3746 | |
7a3e0bb2 RZ |
3747 | r = amdgpu_device_fw_loading(adev); |
3748 | if (r) | |
3749 | return r; | |
3750 | ||
a90ad3c2 | 3751 | /* now we are okay to resume SMC/CP/SDMA */ |
06ec9070 | 3752 | r = amdgpu_device_ip_reinit_late_sriov(adev); |
5740682e ML |
3753 | if (r) |
3754 | goto error; | |
a90ad3c2 ML |
3755 | |
3756 | amdgpu_irq_gpu_reset_resume_helper(adev); | |
5740682e | 3757 | r = amdgpu_ib_ring_tests(adev); |
f81e8d53 | 3758 | amdgpu_amdkfd_post_reset(adev); |
a90ad3c2 | 3759 | |
abc34253 ED |
3760 | error: |
3761 | amdgpu_virt_release_full_gpu(adev, true); | |
c41d1cf6 | 3762 | if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) { |
e3526257 | 3763 | amdgpu_inc_vram_lost(adev); |
c33adbc7 | 3764 | r = amdgpu_device_recover_vram(adev); |
a90ad3c2 ML |
3765 | } |
3766 | ||
3767 | return r; | |
3768 | } | |
3769 | ||
12938fad CK |
3770 | /** |
3771 | * amdgpu_device_should_recover_gpu - check if we should try GPU recovery | |
3772 | * | |
3773 | * @adev: amdgpu device pointer | |
3774 | * | |
3775 | * Check amdgpu_gpu_recovery and SRIOV status to see if we should try to recover | |
3776 | * a hung GPU. | |
3777 | */ | |
3778 | bool amdgpu_device_should_recover_gpu(struct amdgpu_device *adev) | |
3779 | { | |
3780 | if (!amdgpu_device_ip_check_soft_reset(adev)) { | |
3781 | DRM_INFO("Timeout, but no hardware hang detected.\n"); | |
3782 | return false; | |
3783 | } | |
3784 | ||
3ba7b418 AG |
3785 | if (amdgpu_gpu_recovery == 0) |
3786 | goto disabled; | |
3787 | ||
3788 | if (amdgpu_sriov_vf(adev)) | |
3789 | return true; | |
3790 | ||
3791 | if (amdgpu_gpu_recovery == -1) { | |
3792 | switch (adev->asic_type) { | |
fc42d47c AG |
3793 | case CHIP_BONAIRE: |
3794 | case CHIP_HAWAII: | |
3ba7b418 AG |
3795 | case CHIP_TOPAZ: |
3796 | case CHIP_TONGA: | |
3797 | case CHIP_FIJI: | |
3798 | case CHIP_POLARIS10: | |
3799 | case CHIP_POLARIS11: | |
3800 | case CHIP_POLARIS12: | |
3801 | case CHIP_VEGAM: | |
3802 | case CHIP_VEGA20: | |
3803 | case CHIP_VEGA10: | |
3804 | case CHIP_VEGA12: | |
c43b849f | 3805 | case CHIP_RAVEN: |
e9d4cf91 | 3806 | case CHIP_ARCTURUS: |
2cb44fb0 | 3807 | case CHIP_RENOIR: |
658c6639 AD |
3808 | case CHIP_NAVI10: |
3809 | case CHIP_NAVI14: | |
3810 | case CHIP_NAVI12: | |
3ba7b418 AG |
3811 | break; |
3812 | default: | |
3813 | goto disabled; | |
3814 | } | |
12938fad CK |
3815 | } |
3816 | ||
3817 | return true; | |
3ba7b418 AG |
3818 | |
3819 | disabled: | |
3820 | DRM_INFO("GPU recovery disabled.\n"); | |
3821 | return false; | |
12938fad CK |
3822 | } |
3823 | ||
5c6dd71e | 3824 | |
26bc5340 AG |
3825 | static int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, |
3826 | struct amdgpu_job *job, | |
3827 | bool *need_full_reset_arg) | |
3828 | { | |
3829 | int i, r = 0; | |
3830 | bool need_full_reset = *need_full_reset_arg; | |
71182665 | 3831 | |
71182665 | 3832 | /* block all schedulers and reset given job's ring */ |
0875dc9e CZ |
3833 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { |
3834 | struct amdgpu_ring *ring = adev->rings[i]; | |
3835 | ||
51687759 | 3836 | if (!ring || !ring->sched.thread) |
0875dc9e | 3837 | continue; |
5740682e | 3838 | |
2f9d4084 ML |
3839 | /* after all hw jobs are reset, hw fence is meaningless, so force_completion */ |
3840 | amdgpu_fence_driver_force_completion(ring); | |
0875dc9e | 3841 | } |
d38ceaf9 | 3842 | |
222b5f04 AG |
3843 | if(job) |
3844 | drm_sched_increase_karma(&job->base); | |
3845 | ||
1d721ed6 | 3846 | /* Don't suspend on bare metal if we are not going to HW reset the ASIC */ |
26bc5340 AG |
3847 | if (!amdgpu_sriov_vf(adev)) { |
3848 | ||
3849 | if (!need_full_reset) | |
3850 | need_full_reset = amdgpu_device_ip_need_full_reset(adev); | |
3851 | ||
3852 | if (!need_full_reset) { | |
3853 | amdgpu_device_ip_pre_soft_reset(adev); | |
3854 | r = amdgpu_device_ip_soft_reset(adev); | |
3855 | amdgpu_device_ip_post_soft_reset(adev); | |
3856 | if (r || amdgpu_device_ip_check_soft_reset(adev)) { | |
3857 | DRM_INFO("soft reset failed, will fallback to full reset!\n"); | |
3858 | need_full_reset = true; | |
3859 | } | |
3860 | } | |
3861 | ||
3862 | if (need_full_reset) | |
3863 | r = amdgpu_device_ip_suspend(adev); | |
3864 | ||
3865 | *need_full_reset_arg = need_full_reset; | |
3866 | } | |
3867 | ||
3868 | return r; | |
3869 | } | |
3870 | ||
041a62bc | 3871 | static int amdgpu_do_asic_reset(struct amdgpu_hive_info *hive, |
26bc5340 AG |
3872 | struct list_head *device_list_handle, |
3873 | bool *need_full_reset_arg) | |
3874 | { | |
3875 | struct amdgpu_device *tmp_adev = NULL; | |
3876 | bool need_full_reset = *need_full_reset_arg, vram_lost = false; | |
3877 | int r = 0; | |
3878 | ||
3879 | /* | |
3880 | * ASIC reset has to be done on all HGMI hive nodes ASAP | |
3881 | * to allow proper links negotiation in FW (within 1 sec) | |
3882 | */ | |
3883 | if (need_full_reset) { | |
3884 | list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) { | |
041a62bc | 3885 | /* For XGMI run all resets in parallel to speed up the process */ |
d4535e2c | 3886 | if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) { |
c96cf282 | 3887 | if (!queue_work(system_unbound_wq, &tmp_adev->xgmi_reset_work)) |
d4535e2c AG |
3888 | r = -EALREADY; |
3889 | } else | |
3890 | r = amdgpu_asic_reset(tmp_adev); | |
d4535e2c | 3891 | |
041a62bc AG |
3892 | if (r) { |
3893 | DRM_ERROR("ASIC reset failed with error, %d for drm dev, %s", | |
3894 | r, tmp_adev->ddev->unique); | |
3895 | break; | |
ce316fa5 LM |
3896 | } |
3897 | } | |
3898 | ||
041a62bc AG |
3899 | /* For XGMI wait for all resets to complete before proceed */ |
3900 | if (!r) { | |
ce316fa5 LM |
3901 | list_for_each_entry(tmp_adev, device_list_handle, |
3902 | gmc.xgmi.head) { | |
3903 | if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) { | |
3904 | flush_work(&tmp_adev->xgmi_reset_work); | |
3905 | r = tmp_adev->asic_reset_res; | |
3906 | if (r) | |
3907 | break; | |
ce316fa5 LM |
3908 | } |
3909 | } | |
3910 | } | |
ce316fa5 | 3911 | } |
26bc5340 | 3912 | |
00eaa571 LM |
3913 | if (!r && amdgpu_ras_intr_triggered()) |
3914 | amdgpu_ras_intr_cleared(); | |
3915 | ||
26bc5340 AG |
3916 | list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) { |
3917 | if (need_full_reset) { | |
3918 | /* post card */ | |
3919 | if (amdgpu_atom_asic_init(tmp_adev->mode_info.atom_context)) | |
3920 | DRM_WARN("asic atom init failed!"); | |
3921 | ||
3922 | if (!r) { | |
3923 | dev_info(tmp_adev->dev, "GPU reset succeeded, trying to resume\n"); | |
3924 | r = amdgpu_device_ip_resume_phase1(tmp_adev); | |
3925 | if (r) | |
3926 | goto out; | |
3927 | ||
3928 | vram_lost = amdgpu_device_check_vram_lost(tmp_adev); | |
3929 | if (vram_lost) { | |
77e7f829 | 3930 | DRM_INFO("VRAM is lost due to GPU reset!\n"); |
e3526257 | 3931 | amdgpu_inc_vram_lost(tmp_adev); |
26bc5340 AG |
3932 | } |
3933 | ||
3934 | r = amdgpu_gtt_mgr_recover( | |
3935 | &tmp_adev->mman.bdev.man[TTM_PL_TT]); | |
3936 | if (r) | |
3937 | goto out; | |
3938 | ||
3939 | r = amdgpu_device_fw_loading(tmp_adev); | |
3940 | if (r) | |
3941 | return r; | |
3942 | ||
3943 | r = amdgpu_device_ip_resume_phase2(tmp_adev); | |
3944 | if (r) | |
3945 | goto out; | |
3946 | ||
3947 | if (vram_lost) | |
3948 | amdgpu_device_fill_reset_magic(tmp_adev); | |
3949 | ||
fdafb359 EQ |
3950 | /* |
3951 | * Add this ASIC as tracked as reset was already | |
3952 | * complete successfully. | |
3953 | */ | |
3954 | amdgpu_register_gpu_instance(tmp_adev); | |
3955 | ||
7c04ca50 | 3956 | r = amdgpu_device_ip_late_init(tmp_adev); |
3957 | if (r) | |
3958 | goto out; | |
3959 | ||
565d1941 EQ |
3960 | amdgpu_fbdev_set_suspend(tmp_adev, 0); |
3961 | ||
e79a04d5 | 3962 | /* must succeed. */ |
511fdbc3 | 3963 | amdgpu_ras_resume(tmp_adev); |
e79a04d5 | 3964 | |
26bc5340 AG |
3965 | /* Update PSP FW topology after reset */ |
3966 | if (hive && tmp_adev->gmc.xgmi.num_physical_nodes > 1) | |
3967 | r = amdgpu_xgmi_update_topology(hive, tmp_adev); | |
3968 | } | |
3969 | } | |
3970 | ||
3971 | ||
3972 | out: | |
3973 | if (!r) { | |
3974 | amdgpu_irq_gpu_reset_resume_helper(tmp_adev); | |
3975 | r = amdgpu_ib_ring_tests(tmp_adev); | |
3976 | if (r) { | |
3977 | dev_err(tmp_adev->dev, "ib ring test failed (%d).\n", r); | |
3978 | r = amdgpu_device_ip_suspend(tmp_adev); | |
3979 | need_full_reset = true; | |
3980 | r = -EAGAIN; | |
3981 | goto end; | |
3982 | } | |
3983 | } | |
3984 | ||
3985 | if (!r) | |
3986 | r = amdgpu_device_recover_vram(tmp_adev); | |
3987 | else | |
3988 | tmp_adev->asic_reset_res = r; | |
3989 | } | |
3990 | ||
3991 | end: | |
3992 | *need_full_reset_arg = need_full_reset; | |
3993 | return r; | |
3994 | } | |
3995 | ||
1d721ed6 | 3996 | static bool amdgpu_device_lock_adev(struct amdgpu_device *adev, bool trylock) |
26bc5340 | 3997 | { |
1d721ed6 AG |
3998 | if (trylock) { |
3999 | if (!mutex_trylock(&adev->lock_reset)) | |
4000 | return false; | |
4001 | } else | |
4002 | mutex_lock(&adev->lock_reset); | |
5740682e | 4003 | |
26bc5340 | 4004 | atomic_inc(&adev->gpu_reset_counter); |
2a9b90ae | 4005 | adev->in_gpu_reset = true; |
a3a09142 AD |
4006 | switch (amdgpu_asic_reset_method(adev)) { |
4007 | case AMD_RESET_METHOD_MODE1: | |
4008 | adev->mp1_state = PP_MP1_STATE_SHUTDOWN; | |
4009 | break; | |
4010 | case AMD_RESET_METHOD_MODE2: | |
4011 | adev->mp1_state = PP_MP1_STATE_RESET; | |
4012 | break; | |
4013 | default: | |
4014 | adev->mp1_state = PP_MP1_STATE_NONE; | |
4015 | break; | |
4016 | } | |
1d721ed6 AG |
4017 | |
4018 | return true; | |
26bc5340 | 4019 | } |
d38ceaf9 | 4020 | |
26bc5340 AG |
4021 | static void amdgpu_device_unlock_adev(struct amdgpu_device *adev) |
4022 | { | |
89041940 | 4023 | amdgpu_vf_error_trans_all(adev); |
a3a09142 | 4024 | adev->mp1_state = PP_MP1_STATE_NONE; |
2a9b90ae | 4025 | adev->in_gpu_reset = false; |
13a752e3 | 4026 | mutex_unlock(&adev->lock_reset); |
26bc5340 AG |
4027 | } |
4028 | ||
26bc5340 AG |
4029 | /** |
4030 | * amdgpu_device_gpu_recover - reset the asic and recover scheduler | |
4031 | * | |
4032 | * @adev: amdgpu device pointer | |
4033 | * @job: which job trigger hang | |
4034 | * | |
4035 | * Attempt to reset the GPU if it has hung (all asics). | |
4036 | * Attempt to do soft-reset or full-reset and reinitialize Asic | |
4037 | * Returns 0 for success or an error on failure. | |
4038 | */ | |
4039 | ||
4040 | int amdgpu_device_gpu_recover(struct amdgpu_device *adev, | |
4041 | struct amdgpu_job *job) | |
4042 | { | |
1d721ed6 AG |
4043 | struct list_head device_list, *device_list_handle = NULL; |
4044 | bool need_full_reset, job_signaled; | |
26bc5340 | 4045 | struct amdgpu_hive_info *hive = NULL; |
26bc5340 | 4046 | struct amdgpu_device *tmp_adev = NULL; |
1d721ed6 | 4047 | int i, r = 0; |
7c6e68c7 | 4048 | bool in_ras_intr = amdgpu_ras_intr_triggered(); |
b823821f LM |
4049 | bool use_baco = |
4050 | (amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) ? | |
4051 | true : false; | |
26bc5340 | 4052 | |
d5ea093e AG |
4053 | /* |
4054 | * Flush RAM to disk so that after reboot | |
4055 | * the user can read log and see why the system rebooted. | |
4056 | */ | |
b823821f | 4057 | if (in_ras_intr && !use_baco && amdgpu_ras_get_context(adev)->reboot) { |
d5ea093e AG |
4058 | |
4059 | DRM_WARN("Emergency reboot."); | |
4060 | ||
4061 | ksys_sync_helper(); | |
4062 | emergency_restart(); | |
4063 | } | |
4064 | ||
1d721ed6 | 4065 | need_full_reset = job_signaled = false; |
26bc5340 AG |
4066 | INIT_LIST_HEAD(&device_list); |
4067 | ||
b823821f LM |
4068 | dev_info(adev->dev, "GPU %s begin!\n", |
4069 | (in_ras_intr && !use_baco) ? "jobs stop":"reset"); | |
26bc5340 | 4070 | |
beff74bc | 4071 | cancel_delayed_work_sync(&adev->delayed_init_work); |
c53e4db7 | 4072 | |
1d721ed6 AG |
4073 | hive = amdgpu_get_xgmi_hive(adev, false); |
4074 | ||
26bc5340 | 4075 | /* |
1d721ed6 AG |
4076 | * Here we trylock to avoid chain of resets executing from |
4077 | * either trigger by jobs on different adevs in XGMI hive or jobs on | |
4078 | * different schedulers for same device while this TO handler is running. | |
4079 | * We always reset all schedulers for device and all devices for XGMI | |
4080 | * hive so that should take care of them too. | |
26bc5340 | 4081 | */ |
1d721ed6 AG |
4082 | |
4083 | if (hive && !mutex_trylock(&hive->reset_lock)) { | |
4084 | DRM_INFO("Bailing on TDR for s_job:%llx, hive: %llx as another already in progress", | |
0b2d2c2e | 4085 | job ? job->base.id : -1, hive->hive_id); |
26bc5340 | 4086 | return 0; |
1d721ed6 | 4087 | } |
26bc5340 AG |
4088 | |
4089 | /* Start with adev pre asic reset first for soft reset check.*/ | |
1d721ed6 AG |
4090 | if (!amdgpu_device_lock_adev(adev, !hive)) { |
4091 | DRM_INFO("Bailing on TDR for s_job:%llx, as another already in progress", | |
0b2d2c2e | 4092 | job ? job->base.id : -1); |
1d721ed6 | 4093 | return 0; |
26bc5340 AG |
4094 | } |
4095 | ||
7c6e68c7 AG |
4096 | /* Block kfd: SRIOV would do it separately */ |
4097 | if (!amdgpu_sriov_vf(adev)) | |
4098 | amdgpu_amdkfd_pre_reset(adev); | |
4099 | ||
26bc5340 | 4100 | /* Build list of devices to reset */ |
1d721ed6 | 4101 | if (adev->gmc.xgmi.num_physical_nodes > 1) { |
26bc5340 | 4102 | if (!hive) { |
7c6e68c7 AG |
4103 | /*unlock kfd: SRIOV would do it separately */ |
4104 | if (!amdgpu_sriov_vf(adev)) | |
4105 | amdgpu_amdkfd_post_reset(adev); | |
26bc5340 AG |
4106 | amdgpu_device_unlock_adev(adev); |
4107 | return -ENODEV; | |
4108 | } | |
4109 | ||
4110 | /* | |
4111 | * In case we are in XGMI hive mode device reset is done for all the | |
4112 | * nodes in the hive to retrain all XGMI links and hence the reset | |
4113 | * sequence is executed in loop on all nodes. | |
4114 | */ | |
4115 | device_list_handle = &hive->device_list; | |
4116 | } else { | |
4117 | list_add_tail(&adev->gmc.xgmi.head, &device_list); | |
4118 | device_list_handle = &device_list; | |
4119 | } | |
4120 | ||
1d721ed6 AG |
4121 | /* block all schedulers and reset given job's ring */ |
4122 | list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) { | |
7c6e68c7 | 4123 | if (tmp_adev != adev) { |
12ffa55d | 4124 | amdgpu_device_lock_adev(tmp_adev, false); |
7c6e68c7 AG |
4125 | if (!amdgpu_sriov_vf(tmp_adev)) |
4126 | amdgpu_amdkfd_pre_reset(tmp_adev); | |
4127 | } | |
4128 | ||
12ffa55d AG |
4129 | /* |
4130 | * Mark these ASICs to be reseted as untracked first | |
4131 | * And add them back after reset completed | |
4132 | */ | |
4133 | amdgpu_unregister_gpu_instance(tmp_adev); | |
4134 | ||
565d1941 EQ |
4135 | amdgpu_fbdev_set_suspend(adev, 1); |
4136 | ||
f1c1314b | 4137 | /* disable ras on ALL IPs */ |
b823821f LM |
4138 | if (!(in_ras_intr && !use_baco) && |
4139 | amdgpu_device_ip_need_full_reset(tmp_adev)) | |
f1c1314b | 4140 | amdgpu_ras_suspend(tmp_adev); |
4141 | ||
1d721ed6 AG |
4142 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { |
4143 | struct amdgpu_ring *ring = tmp_adev->rings[i]; | |
4144 | ||
4145 | if (!ring || !ring->sched.thread) | |
4146 | continue; | |
4147 | ||
0b2d2c2e | 4148 | drm_sched_stop(&ring->sched, job ? &job->base : NULL); |
7c6e68c7 | 4149 | |
b823821f | 4150 | if (in_ras_intr && !use_baco) |
7c6e68c7 | 4151 | amdgpu_job_stop_all_jobs_on_sched(&ring->sched); |
1d721ed6 AG |
4152 | } |
4153 | } | |
4154 | ||
4155 | ||
b823821f | 4156 | if (in_ras_intr && !use_baco) |
7c6e68c7 AG |
4157 | goto skip_sched_resume; |
4158 | ||
1d721ed6 AG |
4159 | /* |
4160 | * Must check guilty signal here since after this point all old | |
4161 | * HW fences are force signaled. | |
4162 | * | |
4163 | * job->base holds a reference to parent fence | |
4164 | */ | |
4165 | if (job && job->base.s_fence->parent && | |
4166 | dma_fence_is_signaled(job->base.s_fence->parent)) | |
4167 | job_signaled = true; | |
4168 | ||
1d721ed6 AG |
4169 | if (job_signaled) { |
4170 | dev_info(adev->dev, "Guilty job already signaled, skipping HW reset"); | |
4171 | goto skip_hw_reset; | |
4172 | } | |
4173 | ||
4174 | ||
4175 | /* Guilty job will be freed after this*/ | |
0b2d2c2e | 4176 | r = amdgpu_device_pre_asic_reset(adev, job, &need_full_reset); |
1d721ed6 AG |
4177 | if (r) { |
4178 | /*TODO Should we stop ?*/ | |
4179 | DRM_ERROR("GPU pre asic reset failed with err, %d for drm dev, %s ", | |
4180 | r, adev->ddev->unique); | |
4181 | adev->asic_reset_res = r; | |
4182 | } | |
4183 | ||
26bc5340 AG |
4184 | retry: /* Rest of adevs pre asic reset from XGMI hive. */ |
4185 | list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) { | |
4186 | ||
4187 | if (tmp_adev == adev) | |
4188 | continue; | |
4189 | ||
26bc5340 AG |
4190 | r = amdgpu_device_pre_asic_reset(tmp_adev, |
4191 | NULL, | |
4192 | &need_full_reset); | |
4193 | /*TODO Should we stop ?*/ | |
4194 | if (r) { | |
4195 | DRM_ERROR("GPU pre asic reset failed with err, %d for drm dev, %s ", | |
4196 | r, tmp_adev->ddev->unique); | |
4197 | tmp_adev->asic_reset_res = r; | |
4198 | } | |
4199 | } | |
4200 | ||
4201 | /* Actual ASIC resets if needed.*/ | |
4202 | /* TODO Implement XGMI hive reset logic for SRIOV */ | |
4203 | if (amdgpu_sriov_vf(adev)) { | |
4204 | r = amdgpu_device_reset_sriov(adev, job ? false : true); | |
4205 | if (r) | |
4206 | adev->asic_reset_res = r; | |
4207 | } else { | |
041a62bc | 4208 | r = amdgpu_do_asic_reset(hive, device_list_handle, &need_full_reset); |
26bc5340 AG |
4209 | if (r && r == -EAGAIN) |
4210 | goto retry; | |
4211 | } | |
4212 | ||
1d721ed6 AG |
4213 | skip_hw_reset: |
4214 | ||
26bc5340 AG |
4215 | /* Post ASIC reset for all devs .*/ |
4216 | list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) { | |
7c6e68c7 | 4217 | |
1d721ed6 AG |
4218 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { |
4219 | struct amdgpu_ring *ring = tmp_adev->rings[i]; | |
4220 | ||
4221 | if (!ring || !ring->sched.thread) | |
4222 | continue; | |
4223 | ||
4224 | /* No point to resubmit jobs if we didn't HW reset*/ | |
4225 | if (!tmp_adev->asic_reset_res && !job_signaled) | |
4226 | drm_sched_resubmit_jobs(&ring->sched); | |
4227 | ||
4228 | drm_sched_start(&ring->sched, !tmp_adev->asic_reset_res); | |
4229 | } | |
4230 | ||
4231 | if (!amdgpu_device_has_dc_support(tmp_adev) && !job_signaled) { | |
4232 | drm_helper_resume_force_mode(tmp_adev->ddev); | |
4233 | } | |
4234 | ||
4235 | tmp_adev->asic_reset_res = 0; | |
26bc5340 AG |
4236 | |
4237 | if (r) { | |
4238 | /* bad news, how to tell it to userspace ? */ | |
12ffa55d | 4239 | dev_info(tmp_adev->dev, "GPU reset(%d) failed\n", atomic_read(&tmp_adev->gpu_reset_counter)); |
26bc5340 AG |
4240 | amdgpu_vf_error_put(tmp_adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r); |
4241 | } else { | |
12ffa55d | 4242 | dev_info(tmp_adev->dev, "GPU reset(%d) succeeded!\n", atomic_read(&tmp_adev->gpu_reset_counter)); |
26bc5340 | 4243 | } |
7c6e68c7 | 4244 | } |
26bc5340 | 4245 | |
7c6e68c7 AG |
4246 | skip_sched_resume: |
4247 | list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) { | |
4248 | /*unlock kfd: SRIOV would do it separately */ | |
b823821f | 4249 | if (!(in_ras_intr && !use_baco) && !amdgpu_sriov_vf(tmp_adev)) |
7c6e68c7 | 4250 | amdgpu_amdkfd_post_reset(tmp_adev); |
26bc5340 AG |
4251 | amdgpu_device_unlock_adev(tmp_adev); |
4252 | } | |
4253 | ||
1d721ed6 | 4254 | if (hive) |
22d6575b | 4255 | mutex_unlock(&hive->reset_lock); |
26bc5340 AG |
4256 | |
4257 | if (r) | |
4258 | dev_info(adev->dev, "GPU reset end with ret = %d\n", r); | |
d38ceaf9 AD |
4259 | return r; |
4260 | } | |
4261 | ||
e3ecdffa AD |
4262 | /** |
4263 | * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot | |
4264 | * | |
4265 | * @adev: amdgpu_device pointer | |
4266 | * | |
4267 | * Fetchs and stores in the driver the PCIE capabilities (gen speed | |
4268 | * and lanes) of the slot the device is in. Handles APUs and | |
4269 | * virtualized environments where PCIE config space may not be available. | |
4270 | */ | |
5494d864 | 4271 | static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev) |
d0dd7f0c | 4272 | { |
5d9a6330 | 4273 | struct pci_dev *pdev; |
c5313457 HK |
4274 | enum pci_bus_speed speed_cap, platform_speed_cap; |
4275 | enum pcie_link_width platform_link_width; | |
d0dd7f0c | 4276 | |
cd474ba0 AD |
4277 | if (amdgpu_pcie_gen_cap) |
4278 | adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap; | |
d0dd7f0c | 4279 | |
cd474ba0 AD |
4280 | if (amdgpu_pcie_lane_cap) |
4281 | adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap; | |
d0dd7f0c | 4282 | |
cd474ba0 AD |
4283 | /* covers APUs as well */ |
4284 | if (pci_is_root_bus(adev->pdev->bus)) { | |
4285 | if (adev->pm.pcie_gen_mask == 0) | |
4286 | adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK; | |
4287 | if (adev->pm.pcie_mlw_mask == 0) | |
4288 | adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK; | |
d0dd7f0c | 4289 | return; |
cd474ba0 | 4290 | } |
d0dd7f0c | 4291 | |
c5313457 HK |
4292 | if (adev->pm.pcie_gen_mask && adev->pm.pcie_mlw_mask) |
4293 | return; | |
4294 | ||
dbaa922b AD |
4295 | pcie_bandwidth_available(adev->pdev, NULL, |
4296 | &platform_speed_cap, &platform_link_width); | |
c5313457 | 4297 | |
cd474ba0 | 4298 | if (adev->pm.pcie_gen_mask == 0) { |
5d9a6330 AD |
4299 | /* asic caps */ |
4300 | pdev = adev->pdev; | |
4301 | speed_cap = pcie_get_speed_cap(pdev); | |
4302 | if (speed_cap == PCI_SPEED_UNKNOWN) { | |
4303 | adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 | | |
cd474ba0 AD |
4304 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 | |
4305 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3); | |
cd474ba0 | 4306 | } else { |
5d9a6330 AD |
4307 | if (speed_cap == PCIE_SPEED_16_0GT) |
4308 | adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 | | |
4309 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 | | |
4310 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 | | |
4311 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4); | |
4312 | else if (speed_cap == PCIE_SPEED_8_0GT) | |
4313 | adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 | | |
4314 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 | | |
4315 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3); | |
4316 | else if (speed_cap == PCIE_SPEED_5_0GT) | |
4317 | adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 | | |
4318 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2); | |
4319 | else | |
4320 | adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1; | |
4321 | } | |
4322 | /* platform caps */ | |
c5313457 | 4323 | if (platform_speed_cap == PCI_SPEED_UNKNOWN) { |
5d9a6330 AD |
4324 | adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 | |
4325 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2); | |
4326 | } else { | |
c5313457 | 4327 | if (platform_speed_cap == PCIE_SPEED_16_0GT) |
5d9a6330 AD |
4328 | adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 | |
4329 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 | | |
4330 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 | | |
4331 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4); | |
c5313457 | 4332 | else if (platform_speed_cap == PCIE_SPEED_8_0GT) |
5d9a6330 AD |
4333 | adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 | |
4334 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 | | |
4335 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3); | |
c5313457 | 4336 | else if (platform_speed_cap == PCIE_SPEED_5_0GT) |
5d9a6330 AD |
4337 | adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 | |
4338 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2); | |
4339 | else | |
4340 | adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1; | |
4341 | ||
cd474ba0 AD |
4342 | } |
4343 | } | |
4344 | if (adev->pm.pcie_mlw_mask == 0) { | |
c5313457 | 4345 | if (platform_link_width == PCIE_LNK_WIDTH_UNKNOWN) { |
5d9a6330 AD |
4346 | adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK; |
4347 | } else { | |
c5313457 | 4348 | switch (platform_link_width) { |
5d9a6330 | 4349 | case PCIE_LNK_X32: |
cd474ba0 AD |
4350 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 | |
4351 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 | | |
4352 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 | | |
4353 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 | | |
4354 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | | |
4355 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | | |
4356 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); | |
4357 | break; | |
5d9a6330 | 4358 | case PCIE_LNK_X16: |
cd474ba0 AD |
4359 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 | |
4360 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 | | |
4361 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 | | |
4362 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | | |
4363 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | | |
4364 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); | |
4365 | break; | |
5d9a6330 | 4366 | case PCIE_LNK_X12: |
cd474ba0 AD |
4367 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 | |
4368 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 | | |
4369 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | | |
4370 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | | |
4371 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); | |
4372 | break; | |
5d9a6330 | 4373 | case PCIE_LNK_X8: |
cd474ba0 AD |
4374 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 | |
4375 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | | |
4376 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | | |
4377 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); | |
4378 | break; | |
5d9a6330 | 4379 | case PCIE_LNK_X4: |
cd474ba0 AD |
4380 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | |
4381 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | | |
4382 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); | |
4383 | break; | |
5d9a6330 | 4384 | case PCIE_LNK_X2: |
cd474ba0 AD |
4385 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | |
4386 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); | |
4387 | break; | |
5d9a6330 | 4388 | case PCIE_LNK_X1: |
cd474ba0 AD |
4389 | adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1; |
4390 | break; | |
4391 | default: | |
4392 | break; | |
4393 | } | |
d0dd7f0c AD |
4394 | } |
4395 | } | |
4396 | } | |
d38ceaf9 | 4397 | |
361dbd01 AD |
4398 | int amdgpu_device_baco_enter(struct drm_device *dev) |
4399 | { | |
4400 | struct amdgpu_device *adev = dev->dev_private; | |
7a22677b | 4401 | struct amdgpu_ras *ras = amdgpu_ras_get_context(adev); |
361dbd01 AD |
4402 | |
4403 | if (!amdgpu_device_supports_baco(adev->ddev)) | |
4404 | return -ENOTSUPP; | |
4405 | ||
7a22677b LM |
4406 | if (ras && ras->supported) |
4407 | adev->nbio.funcs->enable_doorbell_interrupt(adev, false); | |
4408 | ||
9530273e | 4409 | return amdgpu_dpm_baco_enter(adev); |
361dbd01 AD |
4410 | } |
4411 | ||
4412 | int amdgpu_device_baco_exit(struct drm_device *dev) | |
4413 | { | |
4414 | struct amdgpu_device *adev = dev->dev_private; | |
7a22677b | 4415 | struct amdgpu_ras *ras = amdgpu_ras_get_context(adev); |
9530273e | 4416 | int ret = 0; |
361dbd01 AD |
4417 | |
4418 | if (!amdgpu_device_supports_baco(adev->ddev)) | |
4419 | return -ENOTSUPP; | |
4420 | ||
9530273e EQ |
4421 | ret = amdgpu_dpm_baco_exit(adev); |
4422 | if (ret) | |
4423 | return ret; | |
7a22677b LM |
4424 | |
4425 | if (ras && ras->supported) | |
4426 | adev->nbio.funcs->enable_doorbell_interrupt(adev, true); | |
4427 | ||
4428 | return 0; | |
361dbd01 | 4429 | } |