xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c (revision 6396bb221514d2876fd6dc0aa2a1f240d99b37bb)
1 /*
2  * Copyright 2008 Advanced Micro Devices, Inc.
3  * Copyright 2008 Red Hat Inc.
4  * Copyright 2009 Jerome Glisse.
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a
7  * copy of this software and associated documentation files (the "Software"),
8  * to deal in the Software without restriction, including without limitation
9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10  * and/or sell copies of the Software, and to permit persons to whom the
11  * Software is furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in
14  * all copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22  * OTHER DEALINGS IN THE SOFTWARE.
23  *
24  * Authors: Dave Airlie
25  *          Alex Deucher
26  *          Jerome Glisse
27  */
28 #include <linux/kthread.h>
29 #include <linux/console.h>
30 #include <linux/slab.h>
31 #include <drm/drmP.h>
32 #include <drm/drm_crtc_helper.h>
33 #include <drm/drm_atomic_helper.h>
34 #include <drm/amdgpu_drm.h>
35 #include <linux/vgaarb.h>
36 #include <linux/vga_switcheroo.h>
37 #include <linux/efi.h>
38 #include "amdgpu.h"
39 #include "amdgpu_trace.h"
40 #include "amdgpu_i2c.h"
41 #include "atom.h"
42 #include "amdgpu_atombios.h"
43 #include "amdgpu_atomfirmware.h"
44 #include "amd_pcie.h"
45 #ifdef CONFIG_DRM_AMDGPU_SI
46 #include "si.h"
47 #endif
48 #ifdef CONFIG_DRM_AMDGPU_CIK
49 #include "cik.h"
50 #endif
51 #include "vi.h"
52 #include "soc15.h"
53 #include "bif/bif_4_1_d.h"
54 #include <linux/pci.h>
55 #include <linux/firmware.h>
56 #include "amdgpu_vf_error.h"
57 
58 #include "amdgpu_amdkfd.h"
59 #include "amdgpu_pm.h"
60 
61 MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin");
62 MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin");
63 MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin");
64 
65 #define AMDGPU_RESUME_MS		2000
66 
67 static const char *amdgpu_asic_name[] = {
68 	"TAHITI",
69 	"PITCAIRN",
70 	"VERDE",
71 	"OLAND",
72 	"HAINAN",
73 	"BONAIRE",
74 	"KAVERI",
75 	"KABINI",
76 	"HAWAII",
77 	"MULLINS",
78 	"TOPAZ",
79 	"TONGA",
80 	"FIJI",
81 	"CARRIZO",
82 	"STONEY",
83 	"POLARIS10",
84 	"POLARIS11",
85 	"POLARIS12",
86 	"VEGAM",
87 	"VEGA10",
88 	"VEGA12",
89 	"VEGA20",
90 	"RAVEN",
91 	"LAST",
92 };
93 
94 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev);
95 
96 /**
97  * amdgpu_device_is_px - Is the device is a dGPU with HG/PX power control
98  *
99  * @dev: drm_device pointer
100  *
101  * Returns true if the device is a dGPU with HG/PX power control,
102  * otherwise return false.
103  */
104 bool amdgpu_device_is_px(struct drm_device *dev)
105 {
106 	struct amdgpu_device *adev = dev->dev_private;
107 
108 	if (adev->flags & AMD_IS_PX)
109 		return true;
110 	return false;
111 }
112 
113 /*
114  * MMIO register access helper functions.
115  */
116 /**
117  * amdgpu_mm_rreg - read a memory mapped IO register
118  *
119  * @adev: amdgpu_device pointer
120  * @reg: dword aligned register offset
121  * @acc_flags: access flags which require special behavior
122  *
123  * Returns the 32 bit value from the offset specified.
124  */
125 uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg,
126 			uint32_t acc_flags)
127 {
128 	uint32_t ret;
129 
130 	if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
131 		return amdgpu_virt_kiq_rreg(adev, reg);
132 
133 	if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
134 		ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
135 	else {
136 		unsigned long flags;
137 
138 		spin_lock_irqsave(&adev->mmio_idx_lock, flags);
139 		writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
140 		ret = readl(((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
141 		spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
142 	}
143 	trace_amdgpu_mm_rreg(adev->pdev->device, reg, ret);
144 	return ret;
145 }
146 
147 /*
148  * MMIO register read with bytes helper functions
149  * @offset:bytes offset from MMIO start
150  *
151 */
152 
153 /**
154  * amdgpu_mm_rreg8 - read a memory mapped IO register
155  *
156  * @adev: amdgpu_device pointer
157  * @offset: byte aligned register offset
158  *
159  * Returns the 8 bit value from the offset specified.
160  */
161 uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset) {
162 	if (offset < adev->rmmio_size)
163 		return (readb(adev->rmmio + offset));
164 	BUG();
165 }
166 
167 /*
168  * MMIO register write with bytes helper functions
169  * @offset:bytes offset from MMIO start
170  * @value: the value want to be written to the register
171  *
172 */
173 /**
174  * amdgpu_mm_wreg8 - read a memory mapped IO register
175  *
176  * @adev: amdgpu_device pointer
177  * @offset: byte aligned register offset
178  * @value: 8 bit value to write
179  *
180  * Writes the value specified to the offset specified.
181  */
182 void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value) {
183 	if (offset < adev->rmmio_size)
184 		writeb(value, adev->rmmio + offset);
185 	else
186 		BUG();
187 }
188 
189 /**
190  * amdgpu_mm_wreg - write to a memory mapped IO register
191  *
192  * @adev: amdgpu_device pointer
193  * @reg: dword aligned register offset
194  * @v: 32 bit value to write to the register
195  * @acc_flags: access flags which require special behavior
196  *
197  * Writes the value specified to the offset specified.
198  */
199 void amdgpu_mm_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v,
200 		    uint32_t acc_flags)
201 {
202 	trace_amdgpu_mm_wreg(adev->pdev->device, reg, v);
203 
204 	if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
205 		adev->last_mm_index = v;
206 	}
207 
208 	if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
209 		return amdgpu_virt_kiq_wreg(adev, reg, v);
210 
211 	if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
212 		writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
213 	else {
214 		unsigned long flags;
215 
216 		spin_lock_irqsave(&adev->mmio_idx_lock, flags);
217 		writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
218 		writel(v, ((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
219 		spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
220 	}
221 
222 	if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
223 		udelay(500);
224 	}
225 }
226 
227 /**
228  * amdgpu_io_rreg - read an IO register
229  *
230  * @adev: amdgpu_device pointer
231  * @reg: dword aligned register offset
232  *
233  * Returns the 32 bit value from the offset specified.
234  */
235 u32 amdgpu_io_rreg(struct amdgpu_device *adev, u32 reg)
236 {
237 	if ((reg * 4) < adev->rio_mem_size)
238 		return ioread32(adev->rio_mem + (reg * 4));
239 	else {
240 		iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
241 		return ioread32(adev->rio_mem + (mmMM_DATA * 4));
242 	}
243 }
244 
245 /**
246  * amdgpu_io_wreg - write to an IO register
247  *
248  * @adev: amdgpu_device pointer
249  * @reg: dword aligned register offset
250  * @v: 32 bit value to write to the register
251  *
252  * Writes the value specified to the offset specified.
253  */
254 void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32 v)
255 {
256 	if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
257 		adev->last_mm_index = v;
258 	}
259 
260 	if ((reg * 4) < adev->rio_mem_size)
261 		iowrite32(v, adev->rio_mem + (reg * 4));
262 	else {
263 		iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
264 		iowrite32(v, adev->rio_mem + (mmMM_DATA * 4));
265 	}
266 
267 	if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
268 		udelay(500);
269 	}
270 }
271 
272 /**
273  * amdgpu_mm_rdoorbell - read a doorbell dword
274  *
275  * @adev: amdgpu_device pointer
276  * @index: doorbell index
277  *
278  * Returns the value in the doorbell aperture at the
279  * requested doorbell index (CIK).
280  */
281 u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
282 {
283 	if (index < adev->doorbell.num_doorbells) {
284 		return readl(adev->doorbell.ptr + index);
285 	} else {
286 		DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
287 		return 0;
288 	}
289 }
290 
291 /**
292  * amdgpu_mm_wdoorbell - write a doorbell dword
293  *
294  * @adev: amdgpu_device pointer
295  * @index: doorbell index
296  * @v: value to write
297  *
298  * Writes @v to the doorbell aperture at the
299  * requested doorbell index (CIK).
300  */
301 void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v)
302 {
303 	if (index < adev->doorbell.num_doorbells) {
304 		writel(v, adev->doorbell.ptr + index);
305 	} else {
306 		DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
307 	}
308 }
309 
310 /**
311  * amdgpu_mm_rdoorbell64 - read a doorbell Qword
312  *
313  * @adev: amdgpu_device pointer
314  * @index: doorbell index
315  *
316  * Returns the value in the doorbell aperture at the
317  * requested doorbell index (VEGA10+).
318  */
319 u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index)
320 {
321 	if (index < adev->doorbell.num_doorbells) {
322 		return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index));
323 	} else {
324 		DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
325 		return 0;
326 	}
327 }
328 
329 /**
330  * amdgpu_mm_wdoorbell64 - write a doorbell Qword
331  *
332  * @adev: amdgpu_device pointer
333  * @index: doorbell index
334  * @v: value to write
335  *
336  * Writes @v to the doorbell aperture at the
337  * requested doorbell index (VEGA10+).
338  */
339 void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v)
340 {
341 	if (index < adev->doorbell.num_doorbells) {
342 		atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v);
343 	} else {
344 		DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
345 	}
346 }
347 
348 /**
349  * amdgpu_invalid_rreg - dummy reg read function
350  *
351  * @adev: amdgpu device pointer
352  * @reg: offset of register
353  *
354  * Dummy register read function.  Used for register blocks
355  * that certain asics don't have (all asics).
356  * Returns the value in the register.
357  */
358 static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
359 {
360 	DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
361 	BUG();
362 	return 0;
363 }
364 
365 /**
366  * amdgpu_invalid_wreg - dummy reg write function
367  *
368  * @adev: amdgpu device pointer
369  * @reg: offset of register
370  * @v: value to write to the register
371  *
372  * Dummy register read function.  Used for register blocks
373  * that certain asics don't have (all asics).
374  */
375 static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
376 {
377 	DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
378 		  reg, v);
379 	BUG();
380 }
381 
382 /**
383  * amdgpu_block_invalid_rreg - dummy reg read function
384  *
385  * @adev: amdgpu device pointer
386  * @block: offset of instance
387  * @reg: offset of register
388  *
389  * Dummy register read function.  Used for register blocks
390  * that certain asics don't have (all asics).
391  * Returns the value in the register.
392  */
393 static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
394 					  uint32_t block, uint32_t reg)
395 {
396 	DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
397 		  reg, block);
398 	BUG();
399 	return 0;
400 }
401 
402 /**
403  * amdgpu_block_invalid_wreg - dummy reg write function
404  *
405  * @adev: amdgpu device pointer
406  * @block: offset of instance
407  * @reg: offset of register
408  * @v: value to write to the register
409  *
410  * Dummy register read function.  Used for register blocks
411  * that certain asics don't have (all asics).
412  */
413 static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
414 				      uint32_t block,
415 				      uint32_t reg, uint32_t v)
416 {
417 	DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
418 		  reg, block, v);
419 	BUG();
420 }
421 
422 /**
423  * amdgpu_device_vram_scratch_init - allocate the VRAM scratch page
424  *
425  * @adev: amdgpu device pointer
426  *
427  * Allocates a scratch page of VRAM for use by various things in the
428  * driver.
429  */
430 static int amdgpu_device_vram_scratch_init(struct amdgpu_device *adev)
431 {
432 	return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE,
433 				       PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
434 				       &adev->vram_scratch.robj,
435 				       &adev->vram_scratch.gpu_addr,
436 				       (void **)&adev->vram_scratch.ptr);
437 }
438 
439 /**
440  * amdgpu_device_vram_scratch_fini - Free the VRAM scratch page
441  *
442  * @adev: amdgpu device pointer
443  *
444  * Frees the VRAM scratch page.
445  */
446 static void amdgpu_device_vram_scratch_fini(struct amdgpu_device *adev)
447 {
448 	amdgpu_bo_free_kernel(&adev->vram_scratch.robj, NULL, NULL);
449 }
450 
451 /**
452  * amdgpu_device_program_register_sequence - program an array of registers.
453  *
454  * @adev: amdgpu_device pointer
455  * @registers: pointer to the register array
456  * @array_size: size of the register array
457  *
458  * Programs an array or registers with and and or masks.
459  * This is a helper for setting golden registers.
460  */
461 void amdgpu_device_program_register_sequence(struct amdgpu_device *adev,
462 					     const u32 *registers,
463 					     const u32 array_size)
464 {
465 	u32 tmp, reg, and_mask, or_mask;
466 	int i;
467 
468 	if (array_size % 3)
469 		return;
470 
471 	for (i = 0; i < array_size; i +=3) {
472 		reg = registers[i + 0];
473 		and_mask = registers[i + 1];
474 		or_mask = registers[i + 2];
475 
476 		if (and_mask == 0xffffffff) {
477 			tmp = or_mask;
478 		} else {
479 			tmp = RREG32(reg);
480 			tmp &= ~and_mask;
481 			tmp |= or_mask;
482 		}
483 		WREG32(reg, tmp);
484 	}
485 }
486 
487 /**
488  * amdgpu_device_pci_config_reset - reset the GPU
489  *
490  * @adev: amdgpu_device pointer
491  *
492  * Resets the GPU using the pci config reset sequence.
493  * Only applicable to asics prior to vega10.
494  */
495 void amdgpu_device_pci_config_reset(struct amdgpu_device *adev)
496 {
497 	pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
498 }
499 
500 /*
501  * GPU doorbell aperture helpers function.
502  */
503 /**
504  * amdgpu_device_doorbell_init - Init doorbell driver information.
505  *
506  * @adev: amdgpu_device pointer
507  *
508  * Init doorbell driver information (CIK)
509  * Returns 0 on success, error on failure.
510  */
511 static int amdgpu_device_doorbell_init(struct amdgpu_device *adev)
512 {
513 	/* No doorbell on SI hardware generation */
514 	if (adev->asic_type < CHIP_BONAIRE) {
515 		adev->doorbell.base = 0;
516 		adev->doorbell.size = 0;
517 		adev->doorbell.num_doorbells = 0;
518 		adev->doorbell.ptr = NULL;
519 		return 0;
520 	}
521 
522 	if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET)
523 		return -EINVAL;
524 
525 	/* doorbell bar mapping */
526 	adev->doorbell.base = pci_resource_start(adev->pdev, 2);
527 	adev->doorbell.size = pci_resource_len(adev->pdev, 2);
528 
529 	adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32),
530 					     AMDGPU_DOORBELL_MAX_ASSIGNMENT+1);
531 	if (adev->doorbell.num_doorbells == 0)
532 		return -EINVAL;
533 
534 	adev->doorbell.ptr = ioremap(adev->doorbell.base,
535 				     adev->doorbell.num_doorbells *
536 				     sizeof(u32));
537 	if (adev->doorbell.ptr == NULL)
538 		return -ENOMEM;
539 
540 	return 0;
541 }
542 
543 /**
544  * amdgpu_device_doorbell_fini - Tear down doorbell driver information.
545  *
546  * @adev: amdgpu_device pointer
547  *
548  * Tear down doorbell driver information (CIK)
549  */
550 static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev)
551 {
552 	iounmap(adev->doorbell.ptr);
553 	adev->doorbell.ptr = NULL;
554 }
555 
556 
557 
558 /*
559  * amdgpu_device_wb_*()
560  * Writeback is the method by which the GPU updates special pages in memory
561  * with the status of certain GPU events (fences, ring pointers,etc.).
562  */
563 
564 /**
565  * amdgpu_device_wb_fini - Disable Writeback and free memory
566  *
567  * @adev: amdgpu_device pointer
568  *
569  * Disables Writeback and frees the Writeback memory (all asics).
570  * Used at driver shutdown.
571  */
572 static void amdgpu_device_wb_fini(struct amdgpu_device *adev)
573 {
574 	if (adev->wb.wb_obj) {
575 		amdgpu_bo_free_kernel(&adev->wb.wb_obj,
576 				      &adev->wb.gpu_addr,
577 				      (void **)&adev->wb.wb);
578 		adev->wb.wb_obj = NULL;
579 	}
580 }
581 
582 /**
583  * amdgpu_device_wb_init- Init Writeback driver info and allocate memory
584  *
585  * @adev: amdgpu_device pointer
586  *
587  * Initializes writeback and allocates writeback memory (all asics).
588  * Used at driver startup.
589  * Returns 0 on success or an -error on failure.
590  */
591 static int amdgpu_device_wb_init(struct amdgpu_device *adev)
592 {
593 	int r;
594 
595 	if (adev->wb.wb_obj == NULL) {
596 		/* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */
597 		r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8,
598 					    PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
599 					    &adev->wb.wb_obj, &adev->wb.gpu_addr,
600 					    (void **)&adev->wb.wb);
601 		if (r) {
602 			dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
603 			return r;
604 		}
605 
606 		adev->wb.num_wb = AMDGPU_MAX_WB;
607 		memset(&adev->wb.used, 0, sizeof(adev->wb.used));
608 
609 		/* clear wb memory */
610 		memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8);
611 	}
612 
613 	return 0;
614 }
615 
616 /**
617  * amdgpu_device_wb_get - Allocate a wb entry
618  *
619  * @adev: amdgpu_device pointer
620  * @wb: wb index
621  *
622  * Allocate a wb slot for use by the driver (all asics).
623  * Returns 0 on success or -EINVAL on failure.
624  */
625 int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb)
626 {
627 	unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
628 
629 	if (offset < adev->wb.num_wb) {
630 		__set_bit(offset, adev->wb.used);
631 		*wb = offset << 3; /* convert to dw offset */
632 		return 0;
633 	} else {
634 		return -EINVAL;
635 	}
636 }
637 
638 /**
639  * amdgpu_device_wb_free - Free a wb entry
640  *
641  * @adev: amdgpu_device pointer
642  * @wb: wb index
643  *
644  * Free a wb slot allocated for use by the driver (all asics)
645  */
646 void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb)
647 {
648 	wb >>= 3;
649 	if (wb < adev->wb.num_wb)
650 		__clear_bit(wb, adev->wb.used);
651 }
652 
653 /**
654  * amdgpu_device_vram_location - try to find VRAM location
655  *
656  * @adev: amdgpu device structure holding all necessary informations
657  * @mc: memory controller structure holding memory informations
658  * @base: base address at which to put VRAM
659  *
660  * Function will try to place VRAM at base address provided
661  * as parameter.
662  */
663 void amdgpu_device_vram_location(struct amdgpu_device *adev,
664 				 struct amdgpu_gmc *mc, u64 base)
665 {
666 	uint64_t limit = (uint64_t)amdgpu_vram_limit << 20;
667 
668 	mc->vram_start = base;
669 	mc->vram_end = mc->vram_start + mc->mc_vram_size - 1;
670 	if (limit && limit < mc->real_vram_size)
671 		mc->real_vram_size = limit;
672 	dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
673 			mc->mc_vram_size >> 20, mc->vram_start,
674 			mc->vram_end, mc->real_vram_size >> 20);
675 }
676 
677 /**
678  * amdgpu_device_gart_location - try to find GTT location
679  *
680  * @adev: amdgpu device structure holding all necessary informations
681  * @mc: memory controller structure holding memory informations
682  *
683  * Function will place try to place GTT before or after VRAM.
684  *
685  * If GTT size is bigger than space left then we ajust GTT size.
686  * Thus function will never fails.
687  *
688  * FIXME: when reducing GTT size align new size on power of 2.
689  */
690 void amdgpu_device_gart_location(struct amdgpu_device *adev,
691 				 struct amdgpu_gmc *mc)
692 {
693 	u64 size_af, size_bf;
694 
695 	mc->gart_size += adev->pm.smu_prv_buffer_size;
696 
697 	size_af = adev->gmc.mc_mask - mc->vram_end;
698 	size_bf = mc->vram_start;
699 	if (size_bf > size_af) {
700 		if (mc->gart_size > size_bf) {
701 			dev_warn(adev->dev, "limiting GTT\n");
702 			mc->gart_size = size_bf;
703 		}
704 		mc->gart_start = 0;
705 	} else {
706 		if (mc->gart_size > size_af) {
707 			dev_warn(adev->dev, "limiting GTT\n");
708 			mc->gart_size = size_af;
709 		}
710 		/* VCE doesn't like it when BOs cross a 4GB segment, so align
711 		 * the GART base on a 4GB boundary as well.
712 		 */
713 		mc->gart_start = ALIGN(mc->vram_end + 1, 0x100000000ULL);
714 	}
715 	mc->gart_end = mc->gart_start + mc->gart_size - 1;
716 	dev_info(adev->dev, "GTT: %lluM 0x%016llX - 0x%016llX\n",
717 			mc->gart_size >> 20, mc->gart_start, mc->gart_end);
718 }
719 
720 /**
721  * amdgpu_device_resize_fb_bar - try to resize FB BAR
722  *
723  * @adev: amdgpu_device pointer
724  *
725  * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not
726  * to fail, but if any of the BARs is not accessible after the size we abort
727  * driver loading by returning -ENODEV.
728  */
729 int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev)
730 {
731 	u64 space_needed = roundup_pow_of_two(adev->gmc.real_vram_size);
732 	u32 rbar_size = order_base_2(((space_needed >> 20) | 1)) - 1;
733 	struct pci_bus *root;
734 	struct resource *res;
735 	unsigned i;
736 	u16 cmd;
737 	int r;
738 
739 	/* Bypass for VF */
740 	if (amdgpu_sriov_vf(adev))
741 		return 0;
742 
743 	/* Check if the root BUS has 64bit memory resources */
744 	root = adev->pdev->bus;
745 	while (root->parent)
746 		root = root->parent;
747 
748 	pci_bus_for_each_resource(root, res, i) {
749 		if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) &&
750 		    res->start > 0x100000000ull)
751 			break;
752 	}
753 
754 	/* Trying to resize is pointless without a root hub window above 4GB */
755 	if (!res)
756 		return 0;
757 
758 	/* Disable memory decoding while we change the BAR addresses and size */
759 	pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd);
760 	pci_write_config_word(adev->pdev, PCI_COMMAND,
761 			      cmd & ~PCI_COMMAND_MEMORY);
762 
763 	/* Free the VRAM and doorbell BAR, we most likely need to move both. */
764 	amdgpu_device_doorbell_fini(adev);
765 	if (adev->asic_type >= CHIP_BONAIRE)
766 		pci_release_resource(adev->pdev, 2);
767 
768 	pci_release_resource(adev->pdev, 0);
769 
770 	r = pci_resize_resource(adev->pdev, 0, rbar_size);
771 	if (r == -ENOSPC)
772 		DRM_INFO("Not enough PCI address space for a large BAR.");
773 	else if (r && r != -ENOTSUPP)
774 		DRM_ERROR("Problem resizing BAR0 (%d).", r);
775 
776 	pci_assign_unassigned_bus_resources(adev->pdev->bus);
777 
778 	/* When the doorbell or fb BAR isn't available we have no chance of
779 	 * using the device.
780 	 */
781 	r = amdgpu_device_doorbell_init(adev);
782 	if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET))
783 		return -ENODEV;
784 
785 	pci_write_config_word(adev->pdev, PCI_COMMAND, cmd);
786 
787 	return 0;
788 }
789 
790 /*
791  * GPU helpers function.
792  */
793 /**
794  * amdgpu_device_need_post - check if the hw need post or not
795  *
796  * @adev: amdgpu_device pointer
797  *
798  * Check if the asic has been initialized (all asics) at driver startup
799  * or post is needed if  hw reset is performed.
800  * Returns true if need or false if not.
801  */
802 bool amdgpu_device_need_post(struct amdgpu_device *adev)
803 {
804 	uint32_t reg;
805 
806 	if (amdgpu_sriov_vf(adev))
807 		return false;
808 
809 	if (amdgpu_passthrough(adev)) {
810 		/* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
811 		 * some old smc fw still need driver do vPost otherwise gpu hang, while
812 		 * those smc fw version above 22.15 doesn't have this flaw, so we force
813 		 * vpost executed for smc version below 22.15
814 		 */
815 		if (adev->asic_type == CHIP_FIJI) {
816 			int err;
817 			uint32_t fw_ver;
818 			err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
819 			/* force vPost if error occured */
820 			if (err)
821 				return true;
822 
823 			fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
824 			if (fw_ver < 0x00160e00)
825 				return true;
826 		}
827 	}
828 
829 	if (adev->has_hw_reset) {
830 		adev->has_hw_reset = false;
831 		return true;
832 	}
833 
834 	/* bios scratch used on CIK+ */
835 	if (adev->asic_type >= CHIP_BONAIRE)
836 		return amdgpu_atombios_scratch_need_asic_init(adev);
837 
838 	/* check MEM_SIZE for older asics */
839 	reg = amdgpu_asic_get_config_memsize(adev);
840 
841 	if ((reg != 0) && (reg != 0xffffffff))
842 		return false;
843 
844 	return true;
845 }
846 
847 /* if we get transitioned to only one device, take VGA back */
848 /**
849  * amdgpu_device_vga_set_decode - enable/disable vga decode
850  *
851  * @cookie: amdgpu_device pointer
852  * @state: enable/disable vga decode
853  *
854  * Enable/disable vga decode (all asics).
855  * Returns VGA resource flags.
856  */
857 static unsigned int amdgpu_device_vga_set_decode(void *cookie, bool state)
858 {
859 	struct amdgpu_device *adev = cookie;
860 	amdgpu_asic_set_vga_state(adev, state);
861 	if (state)
862 		return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
863 		       VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
864 	else
865 		return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
866 }
867 
868 /**
869  * amdgpu_device_check_block_size - validate the vm block size
870  *
871  * @adev: amdgpu_device pointer
872  *
873  * Validates the vm block size specified via module parameter.
874  * The vm block size defines number of bits in page table versus page directory,
875  * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
876  * page table and the remaining bits are in the page directory.
877  */
878 static void amdgpu_device_check_block_size(struct amdgpu_device *adev)
879 {
880 	/* defines number of bits in page table versus page directory,
881 	 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
882 	 * page table and the remaining bits are in the page directory */
883 	if (amdgpu_vm_block_size == -1)
884 		return;
885 
886 	if (amdgpu_vm_block_size < 9) {
887 		dev_warn(adev->dev, "VM page table size (%d) too small\n",
888 			 amdgpu_vm_block_size);
889 		amdgpu_vm_block_size = -1;
890 	}
891 }
892 
893 /**
894  * amdgpu_device_check_vm_size - validate the vm size
895  *
896  * @adev: amdgpu_device pointer
897  *
898  * Validates the vm size in GB specified via module parameter.
899  * The VM size is the size of the GPU virtual memory space in GB.
900  */
901 static void amdgpu_device_check_vm_size(struct amdgpu_device *adev)
902 {
903 	/* no need to check the default value */
904 	if (amdgpu_vm_size == -1)
905 		return;
906 
907 	if (amdgpu_vm_size < 1) {
908 		dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
909 			 amdgpu_vm_size);
910 		amdgpu_vm_size = -1;
911 	}
912 }
913 
914 static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev)
915 {
916 	struct sysinfo si;
917 	bool is_os_64 = (sizeof(void *) == 8) ? true : false;
918 	uint64_t total_memory;
919 	uint64_t dram_size_seven_GB = 0x1B8000000;
920 	uint64_t dram_size_three_GB = 0xB8000000;
921 
922 	if (amdgpu_smu_memory_pool_size == 0)
923 		return;
924 
925 	if (!is_os_64) {
926 		DRM_WARN("Not 64-bit OS, feature not supported\n");
927 		goto def_value;
928 	}
929 	si_meminfo(&si);
930 	total_memory = (uint64_t)si.totalram * si.mem_unit;
931 
932 	if ((amdgpu_smu_memory_pool_size == 1) ||
933 		(amdgpu_smu_memory_pool_size == 2)) {
934 		if (total_memory < dram_size_three_GB)
935 			goto def_value1;
936 	} else if ((amdgpu_smu_memory_pool_size == 4) ||
937 		(amdgpu_smu_memory_pool_size == 8)) {
938 		if (total_memory < dram_size_seven_GB)
939 			goto def_value1;
940 	} else {
941 		DRM_WARN("Smu memory pool size not supported\n");
942 		goto def_value;
943 	}
944 	adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28;
945 
946 	return;
947 
948 def_value1:
949 	DRM_WARN("No enough system memory\n");
950 def_value:
951 	adev->pm.smu_prv_buffer_size = 0;
952 }
953 
954 /**
955  * amdgpu_device_check_arguments - validate module params
956  *
957  * @adev: amdgpu_device pointer
958  *
959  * Validates certain module parameters and updates
960  * the associated values used by the driver (all asics).
961  */
962 static void amdgpu_device_check_arguments(struct amdgpu_device *adev)
963 {
964 	if (amdgpu_sched_jobs < 4) {
965 		dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
966 			 amdgpu_sched_jobs);
967 		amdgpu_sched_jobs = 4;
968 	} else if (!is_power_of_2(amdgpu_sched_jobs)){
969 		dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
970 			 amdgpu_sched_jobs);
971 		amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
972 	}
973 
974 	if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) {
975 		/* gart size must be greater or equal to 32M */
976 		dev_warn(adev->dev, "gart size (%d) too small\n",
977 			 amdgpu_gart_size);
978 		amdgpu_gart_size = -1;
979 	}
980 
981 	if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) {
982 		/* gtt size must be greater or equal to 32M */
983 		dev_warn(adev->dev, "gtt size (%d) too small\n",
984 				 amdgpu_gtt_size);
985 		amdgpu_gtt_size = -1;
986 	}
987 
988 	/* valid range is between 4 and 9 inclusive */
989 	if (amdgpu_vm_fragment_size != -1 &&
990 	    (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) {
991 		dev_warn(adev->dev, "valid range is between 4 and 9\n");
992 		amdgpu_vm_fragment_size = -1;
993 	}
994 
995 	amdgpu_device_check_smu_prv_buffer_size(adev);
996 
997 	amdgpu_device_check_vm_size(adev);
998 
999 	amdgpu_device_check_block_size(adev);
1000 
1001 	if (amdgpu_vram_page_split != -1 && (amdgpu_vram_page_split < 16 ||
1002 	    !is_power_of_2(amdgpu_vram_page_split))) {
1003 		dev_warn(adev->dev, "invalid VRAM page split (%d)\n",
1004 			 amdgpu_vram_page_split);
1005 		amdgpu_vram_page_split = 1024;
1006 	}
1007 
1008 	if (amdgpu_lockup_timeout == 0) {
1009 		dev_warn(adev->dev, "lockup_timeout msut be > 0, adjusting to 10000\n");
1010 		amdgpu_lockup_timeout = 10000;
1011 	}
1012 
1013 	adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type);
1014 }
1015 
1016 /**
1017  * amdgpu_switcheroo_set_state - set switcheroo state
1018  *
1019  * @pdev: pci dev pointer
1020  * @state: vga_switcheroo state
1021  *
1022  * Callback for the switcheroo driver.  Suspends or resumes the
1023  * the asics before or after it is powered up using ACPI methods.
1024  */
1025 static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, enum vga_switcheroo_state state)
1026 {
1027 	struct drm_device *dev = pci_get_drvdata(pdev);
1028 
1029 	if (amdgpu_device_is_px(dev) && state == VGA_SWITCHEROO_OFF)
1030 		return;
1031 
1032 	if (state == VGA_SWITCHEROO_ON) {
1033 		pr_info("amdgpu: switched on\n");
1034 		/* don't suspend or resume card normally */
1035 		dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1036 
1037 		amdgpu_device_resume(dev, true, true);
1038 
1039 		dev->switch_power_state = DRM_SWITCH_POWER_ON;
1040 		drm_kms_helper_poll_enable(dev);
1041 	} else {
1042 		pr_info("amdgpu: switched off\n");
1043 		drm_kms_helper_poll_disable(dev);
1044 		dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1045 		amdgpu_device_suspend(dev, true, true);
1046 		dev->switch_power_state = DRM_SWITCH_POWER_OFF;
1047 	}
1048 }
1049 
1050 /**
1051  * amdgpu_switcheroo_can_switch - see if switcheroo state can change
1052  *
1053  * @pdev: pci dev pointer
1054  *
1055  * Callback for the switcheroo driver.  Check of the switcheroo
1056  * state can be changed.
1057  * Returns true if the state can be changed, false if not.
1058  */
1059 static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
1060 {
1061 	struct drm_device *dev = pci_get_drvdata(pdev);
1062 
1063 	/*
1064 	* FIXME: open_count is protected by drm_global_mutex but that would lead to
1065 	* locking inversion with the driver load path. And the access here is
1066 	* completely racy anyway. So don't bother with locking for now.
1067 	*/
1068 	return dev->open_count == 0;
1069 }
1070 
1071 static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
1072 	.set_gpu_state = amdgpu_switcheroo_set_state,
1073 	.reprobe = NULL,
1074 	.can_switch = amdgpu_switcheroo_can_switch,
1075 };
1076 
1077 /**
1078  * amdgpu_device_ip_set_clockgating_state - set the CG state
1079  *
1080  * @adev: amdgpu_device pointer
1081  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1082  * @state: clockgating state (gate or ungate)
1083  *
1084  * Sets the requested clockgating state for all instances of
1085  * the hardware IP specified.
1086  * Returns the error code from the last instance.
1087  */
1088 int amdgpu_device_ip_set_clockgating_state(void *dev,
1089 					   enum amd_ip_block_type block_type,
1090 					   enum amd_clockgating_state state)
1091 {
1092 	struct amdgpu_device *adev = dev;
1093 	int i, r = 0;
1094 
1095 	for (i = 0; i < adev->num_ip_blocks; i++) {
1096 		if (!adev->ip_blocks[i].status.valid)
1097 			continue;
1098 		if (adev->ip_blocks[i].version->type != block_type)
1099 			continue;
1100 		if (!adev->ip_blocks[i].version->funcs->set_clockgating_state)
1101 			continue;
1102 		r = adev->ip_blocks[i].version->funcs->set_clockgating_state(
1103 			(void *)adev, state);
1104 		if (r)
1105 			DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",
1106 				  adev->ip_blocks[i].version->funcs->name, r);
1107 	}
1108 	return r;
1109 }
1110 
1111 /**
1112  * amdgpu_device_ip_set_powergating_state - set the PG state
1113  *
1114  * @adev: amdgpu_device pointer
1115  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1116  * @state: powergating state (gate or ungate)
1117  *
1118  * Sets the requested powergating state for all instances of
1119  * the hardware IP specified.
1120  * Returns the error code from the last instance.
1121  */
1122 int amdgpu_device_ip_set_powergating_state(void *dev,
1123 					   enum amd_ip_block_type block_type,
1124 					   enum amd_powergating_state state)
1125 {
1126 	struct amdgpu_device *adev = dev;
1127 	int i, r = 0;
1128 
1129 	for (i = 0; i < adev->num_ip_blocks; i++) {
1130 		if (!adev->ip_blocks[i].status.valid)
1131 			continue;
1132 		if (adev->ip_blocks[i].version->type != block_type)
1133 			continue;
1134 		if (!adev->ip_blocks[i].version->funcs->set_powergating_state)
1135 			continue;
1136 		r = adev->ip_blocks[i].version->funcs->set_powergating_state(
1137 			(void *)adev, state);
1138 		if (r)
1139 			DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",
1140 				  adev->ip_blocks[i].version->funcs->name, r);
1141 	}
1142 	return r;
1143 }
1144 
1145 /**
1146  * amdgpu_device_ip_get_clockgating_state - get the CG state
1147  *
1148  * @adev: amdgpu_device pointer
1149  * @flags: clockgating feature flags
1150  *
1151  * Walks the list of IPs on the device and updates the clockgating
1152  * flags for each IP.
1153  * Updates @flags with the feature flags for each hardware IP where
1154  * clockgating is enabled.
1155  */
1156 void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev,
1157 					    u32 *flags)
1158 {
1159 	int i;
1160 
1161 	for (i = 0; i < adev->num_ip_blocks; i++) {
1162 		if (!adev->ip_blocks[i].status.valid)
1163 			continue;
1164 		if (adev->ip_blocks[i].version->funcs->get_clockgating_state)
1165 			adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags);
1166 	}
1167 }
1168 
1169 /**
1170  * amdgpu_device_ip_wait_for_idle - wait for idle
1171  *
1172  * @adev: amdgpu_device pointer
1173  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1174  *
1175  * Waits for the request hardware IP to be idle.
1176  * Returns 0 for success or a negative error code on failure.
1177  */
1178 int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev,
1179 				   enum amd_ip_block_type block_type)
1180 {
1181 	int i, r;
1182 
1183 	for (i = 0; i < adev->num_ip_blocks; i++) {
1184 		if (!adev->ip_blocks[i].status.valid)
1185 			continue;
1186 		if (adev->ip_blocks[i].version->type == block_type) {
1187 			r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
1188 			if (r)
1189 				return r;
1190 			break;
1191 		}
1192 	}
1193 	return 0;
1194 
1195 }
1196 
1197 /**
1198  * amdgpu_device_ip_is_idle - is the hardware IP idle
1199  *
1200  * @adev: amdgpu_device pointer
1201  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1202  *
1203  * Check if the hardware IP is idle or not.
1204  * Returns true if it the IP is idle, false if not.
1205  */
1206 bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev,
1207 			      enum amd_ip_block_type block_type)
1208 {
1209 	int i;
1210 
1211 	for (i = 0; i < adev->num_ip_blocks; i++) {
1212 		if (!adev->ip_blocks[i].status.valid)
1213 			continue;
1214 		if (adev->ip_blocks[i].version->type == block_type)
1215 			return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
1216 	}
1217 	return true;
1218 
1219 }
1220 
1221 /**
1222  * amdgpu_device_ip_get_ip_block - get a hw IP pointer
1223  *
1224  * @adev: amdgpu_device pointer
1225  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1226  *
1227  * Returns a pointer to the hardware IP block structure
1228  * if it exists for the asic, otherwise NULL.
1229  */
1230 struct amdgpu_ip_block *
1231 amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,
1232 			      enum amd_ip_block_type type)
1233 {
1234 	int i;
1235 
1236 	for (i = 0; i < adev->num_ip_blocks; i++)
1237 		if (adev->ip_blocks[i].version->type == type)
1238 			return &adev->ip_blocks[i];
1239 
1240 	return NULL;
1241 }
1242 
1243 /**
1244  * amdgpu_device_ip_block_version_cmp
1245  *
1246  * @adev: amdgpu_device pointer
1247  * @type: enum amd_ip_block_type
1248  * @major: major version
1249  * @minor: minor version
1250  *
1251  * return 0 if equal or greater
1252  * return 1 if smaller or the ip_block doesn't exist
1253  */
1254 int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev,
1255 				       enum amd_ip_block_type type,
1256 				       u32 major, u32 minor)
1257 {
1258 	struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type);
1259 
1260 	if (ip_block && ((ip_block->version->major > major) ||
1261 			((ip_block->version->major == major) &&
1262 			(ip_block->version->minor >= minor))))
1263 		return 0;
1264 
1265 	return 1;
1266 }
1267 
1268 /**
1269  * amdgpu_device_ip_block_add
1270  *
1271  * @adev: amdgpu_device pointer
1272  * @ip_block_version: pointer to the IP to add
1273  *
1274  * Adds the IP block driver information to the collection of IPs
1275  * on the asic.
1276  */
1277 int amdgpu_device_ip_block_add(struct amdgpu_device *adev,
1278 			       const struct amdgpu_ip_block_version *ip_block_version)
1279 {
1280 	if (!ip_block_version)
1281 		return -EINVAL;
1282 
1283 	DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks,
1284 		  ip_block_version->funcs->name);
1285 
1286 	adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
1287 
1288 	return 0;
1289 }
1290 
1291 /**
1292  * amdgpu_device_enable_virtual_display - enable virtual display feature
1293  *
1294  * @adev: amdgpu_device pointer
1295  *
1296  * Enabled the virtual display feature if the user has enabled it via
1297  * the module parameter virtual_display.  This feature provides a virtual
1298  * display hardware on headless boards or in virtualized environments.
1299  * This function parses and validates the configuration string specified by
1300  * the user and configues the virtual display configuration (number of
1301  * virtual connectors, crtcs, etc.) specified.
1302  */
1303 static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
1304 {
1305 	adev->enable_virtual_display = false;
1306 
1307 	if (amdgpu_virtual_display) {
1308 		struct drm_device *ddev = adev->ddev;
1309 		const char *pci_address_name = pci_name(ddev->pdev);
1310 		char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
1311 
1312 		pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
1313 		pciaddstr_tmp = pciaddstr;
1314 		while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
1315 			pciaddname = strsep(&pciaddname_tmp, ",");
1316 			if (!strcmp("all", pciaddname)
1317 			    || !strcmp(pci_address_name, pciaddname)) {
1318 				long num_crtc;
1319 				int res = -1;
1320 
1321 				adev->enable_virtual_display = true;
1322 
1323 				if (pciaddname_tmp)
1324 					res = kstrtol(pciaddname_tmp, 10,
1325 						      &num_crtc);
1326 
1327 				if (!res) {
1328 					if (num_crtc < 1)
1329 						num_crtc = 1;
1330 					if (num_crtc > 6)
1331 						num_crtc = 6;
1332 					adev->mode_info.num_crtc = num_crtc;
1333 				} else {
1334 					adev->mode_info.num_crtc = 1;
1335 				}
1336 				break;
1337 			}
1338 		}
1339 
1340 		DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
1341 			 amdgpu_virtual_display, pci_address_name,
1342 			 adev->enable_virtual_display, adev->mode_info.num_crtc);
1343 
1344 		kfree(pciaddstr);
1345 	}
1346 }
1347 
1348 /**
1349  * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware
1350  *
1351  * @adev: amdgpu_device pointer
1352  *
1353  * Parses the asic configuration parameters specified in the gpu info
1354  * firmware and makes them availale to the driver for use in configuring
1355  * the asic.
1356  * Returns 0 on success, -EINVAL on failure.
1357  */
1358 static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev)
1359 {
1360 	const char *chip_name;
1361 	char fw_name[30];
1362 	int err;
1363 	const struct gpu_info_firmware_header_v1_0 *hdr;
1364 
1365 	adev->firmware.gpu_info_fw = NULL;
1366 
1367 	switch (adev->asic_type) {
1368 	case CHIP_TOPAZ:
1369 	case CHIP_TONGA:
1370 	case CHIP_FIJI:
1371 	case CHIP_POLARIS10:
1372 	case CHIP_POLARIS11:
1373 	case CHIP_POLARIS12:
1374 	case CHIP_VEGAM:
1375 	case CHIP_CARRIZO:
1376 	case CHIP_STONEY:
1377 #ifdef CONFIG_DRM_AMDGPU_SI
1378 	case CHIP_VERDE:
1379 	case CHIP_TAHITI:
1380 	case CHIP_PITCAIRN:
1381 	case CHIP_OLAND:
1382 	case CHIP_HAINAN:
1383 #endif
1384 #ifdef CONFIG_DRM_AMDGPU_CIK
1385 	case CHIP_BONAIRE:
1386 	case CHIP_HAWAII:
1387 	case CHIP_KAVERI:
1388 	case CHIP_KABINI:
1389 	case CHIP_MULLINS:
1390 #endif
1391 	case CHIP_VEGA20:
1392 	default:
1393 		return 0;
1394 	case CHIP_VEGA10:
1395 		chip_name = "vega10";
1396 		break;
1397 	case CHIP_VEGA12:
1398 		chip_name = "vega12";
1399 		break;
1400 	case CHIP_RAVEN:
1401 		chip_name = "raven";
1402 		break;
1403 	}
1404 
1405 	snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name);
1406 	err = request_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev);
1407 	if (err) {
1408 		dev_err(adev->dev,
1409 			"Failed to load gpu_info firmware \"%s\"\n",
1410 			fw_name);
1411 		goto out;
1412 	}
1413 	err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw);
1414 	if (err) {
1415 		dev_err(adev->dev,
1416 			"Failed to validate gpu_info firmware \"%s\"\n",
1417 			fw_name);
1418 		goto out;
1419 	}
1420 
1421 	hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data;
1422 	amdgpu_ucode_print_gpu_info_hdr(&hdr->header);
1423 
1424 	switch (hdr->version_major) {
1425 	case 1:
1426 	{
1427 		const struct gpu_info_firmware_v1_0 *gpu_info_fw =
1428 			(const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data +
1429 								le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1430 
1431 		adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se);
1432 		adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh);
1433 		adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se);
1434 		adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se);
1435 		adev->gfx.config.max_texture_channel_caches =
1436 			le32_to_cpu(gpu_info_fw->gc_num_tccs);
1437 		adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs);
1438 		adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds);
1439 		adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth);
1440 		adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth);
1441 		adev->gfx.config.double_offchip_lds_buf =
1442 			le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer);
1443 		adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size);
1444 		adev->gfx.cu_info.max_waves_per_simd =
1445 			le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd);
1446 		adev->gfx.cu_info.max_scratch_slots_per_cu =
1447 			le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu);
1448 		adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size);
1449 		break;
1450 	}
1451 	default:
1452 		dev_err(adev->dev,
1453 			"Unsupported gpu_info table %d\n", hdr->header.ucode_version);
1454 		err = -EINVAL;
1455 		goto out;
1456 	}
1457 out:
1458 	return err;
1459 }
1460 
1461 /**
1462  * amdgpu_device_ip_early_init - run early init for hardware IPs
1463  *
1464  * @adev: amdgpu_device pointer
1465  *
1466  * Early initialization pass for hardware IPs.  The hardware IPs that make
1467  * up each asic are discovered each IP's early_init callback is run.  This
1468  * is the first stage in initializing the asic.
1469  * Returns 0 on success, negative error code on failure.
1470  */
1471 static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
1472 {
1473 	int i, r;
1474 
1475 	amdgpu_device_enable_virtual_display(adev);
1476 
1477 	switch (adev->asic_type) {
1478 	case CHIP_TOPAZ:
1479 	case CHIP_TONGA:
1480 	case CHIP_FIJI:
1481 	case CHIP_POLARIS10:
1482 	case CHIP_POLARIS11:
1483 	case CHIP_POLARIS12:
1484 	case CHIP_VEGAM:
1485 	case CHIP_CARRIZO:
1486 	case CHIP_STONEY:
1487 		if (adev->asic_type == CHIP_CARRIZO || adev->asic_type == CHIP_STONEY)
1488 			adev->family = AMDGPU_FAMILY_CZ;
1489 		else
1490 			adev->family = AMDGPU_FAMILY_VI;
1491 
1492 		r = vi_set_ip_blocks(adev);
1493 		if (r)
1494 			return r;
1495 		break;
1496 #ifdef CONFIG_DRM_AMDGPU_SI
1497 	case CHIP_VERDE:
1498 	case CHIP_TAHITI:
1499 	case CHIP_PITCAIRN:
1500 	case CHIP_OLAND:
1501 	case CHIP_HAINAN:
1502 		adev->family = AMDGPU_FAMILY_SI;
1503 		r = si_set_ip_blocks(adev);
1504 		if (r)
1505 			return r;
1506 		break;
1507 #endif
1508 #ifdef CONFIG_DRM_AMDGPU_CIK
1509 	case CHIP_BONAIRE:
1510 	case CHIP_HAWAII:
1511 	case CHIP_KAVERI:
1512 	case CHIP_KABINI:
1513 	case CHIP_MULLINS:
1514 		if ((adev->asic_type == CHIP_BONAIRE) || (adev->asic_type == CHIP_HAWAII))
1515 			adev->family = AMDGPU_FAMILY_CI;
1516 		else
1517 			adev->family = AMDGPU_FAMILY_KV;
1518 
1519 		r = cik_set_ip_blocks(adev);
1520 		if (r)
1521 			return r;
1522 		break;
1523 #endif
1524 	case CHIP_VEGA10:
1525 	case CHIP_VEGA12:
1526 	case CHIP_VEGA20:
1527 	case CHIP_RAVEN:
1528 		if (adev->asic_type == CHIP_RAVEN)
1529 			adev->family = AMDGPU_FAMILY_RV;
1530 		else
1531 			adev->family = AMDGPU_FAMILY_AI;
1532 
1533 		r = soc15_set_ip_blocks(adev);
1534 		if (r)
1535 			return r;
1536 		break;
1537 	default:
1538 		/* FIXME: not supported yet */
1539 		return -EINVAL;
1540 	}
1541 
1542 	r = amdgpu_device_parse_gpu_info_fw(adev);
1543 	if (r)
1544 		return r;
1545 
1546 	amdgpu_amdkfd_device_probe(adev);
1547 
1548 	if (amdgpu_sriov_vf(adev)) {
1549 		r = amdgpu_virt_request_full_gpu(adev, true);
1550 		if (r)
1551 			return -EAGAIN;
1552 	}
1553 
1554 	adev->powerplay.pp_feature = amdgpu_pp_feature_mask;
1555 
1556 	for (i = 0; i < adev->num_ip_blocks; i++) {
1557 		if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
1558 			DRM_ERROR("disabled ip block: %d <%s>\n",
1559 				  i, adev->ip_blocks[i].version->funcs->name);
1560 			adev->ip_blocks[i].status.valid = false;
1561 		} else {
1562 			if (adev->ip_blocks[i].version->funcs->early_init) {
1563 				r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
1564 				if (r == -ENOENT) {
1565 					adev->ip_blocks[i].status.valid = false;
1566 				} else if (r) {
1567 					DRM_ERROR("early_init of IP block <%s> failed %d\n",
1568 						  adev->ip_blocks[i].version->funcs->name, r);
1569 					return r;
1570 				} else {
1571 					adev->ip_blocks[i].status.valid = true;
1572 				}
1573 			} else {
1574 				adev->ip_blocks[i].status.valid = true;
1575 			}
1576 		}
1577 	}
1578 
1579 	adev->cg_flags &= amdgpu_cg_mask;
1580 	adev->pg_flags &= amdgpu_pg_mask;
1581 
1582 	return 0;
1583 }
1584 
1585 /**
1586  * amdgpu_device_ip_init - run init for hardware IPs
1587  *
1588  * @adev: amdgpu_device pointer
1589  *
1590  * Main initialization pass for hardware IPs.  The list of all the hardware
1591  * IPs that make up the asic is walked and the sw_init and hw_init callbacks
1592  * are run.  sw_init initializes the software state associated with each IP
1593  * and hw_init initializes the hardware associated with each IP.
1594  * Returns 0 on success, negative error code on failure.
1595  */
1596 static int amdgpu_device_ip_init(struct amdgpu_device *adev)
1597 {
1598 	int i, r;
1599 
1600 	for (i = 0; i < adev->num_ip_blocks; i++) {
1601 		if (!adev->ip_blocks[i].status.valid)
1602 			continue;
1603 		r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
1604 		if (r) {
1605 			DRM_ERROR("sw_init of IP block <%s> failed %d\n",
1606 				  adev->ip_blocks[i].version->funcs->name, r);
1607 			return r;
1608 		}
1609 		adev->ip_blocks[i].status.sw = true;
1610 
1611 		/* need to do gmc hw init early so we can allocate gpu mem */
1612 		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1613 			r = amdgpu_device_vram_scratch_init(adev);
1614 			if (r) {
1615 				DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r);
1616 				return r;
1617 			}
1618 			r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1619 			if (r) {
1620 				DRM_ERROR("hw_init %d failed %d\n", i, r);
1621 				return r;
1622 			}
1623 			r = amdgpu_device_wb_init(adev);
1624 			if (r) {
1625 				DRM_ERROR("amdgpu_device_wb_init failed %d\n", r);
1626 				return r;
1627 			}
1628 			adev->ip_blocks[i].status.hw = true;
1629 
1630 			/* right after GMC hw init, we create CSA */
1631 			if (amdgpu_sriov_vf(adev)) {
1632 				r = amdgpu_allocate_static_csa(adev);
1633 				if (r) {
1634 					DRM_ERROR("allocate CSA failed %d\n", r);
1635 					return r;
1636 				}
1637 			}
1638 		}
1639 	}
1640 
1641 	for (i = 0; i < adev->num_ip_blocks; i++) {
1642 		if (!adev->ip_blocks[i].status.sw)
1643 			continue;
1644 		if (adev->ip_blocks[i].status.hw)
1645 			continue;
1646 		r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1647 		if (r) {
1648 			DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1649 				  adev->ip_blocks[i].version->funcs->name, r);
1650 			return r;
1651 		}
1652 		adev->ip_blocks[i].status.hw = true;
1653 	}
1654 
1655 	amdgpu_amdkfd_device_init(adev);
1656 
1657 	if (amdgpu_sriov_vf(adev))
1658 		amdgpu_virt_release_full_gpu(adev, true);
1659 
1660 	return 0;
1661 }
1662 
1663 /**
1664  * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer
1665  *
1666  * @adev: amdgpu_device pointer
1667  *
1668  * Writes a reset magic value to the gart pointer in VRAM.  The driver calls
1669  * this function before a GPU reset.  If the value is retained after a
1670  * GPU reset, VRAM has not been lost.  Some GPU resets may destry VRAM contents.
1671  */
1672 static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev)
1673 {
1674 	memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
1675 }
1676 
1677 /**
1678  * amdgpu_device_check_vram_lost - check if vram is valid
1679  *
1680  * @adev: amdgpu_device pointer
1681  *
1682  * Checks the reset magic value written to the gart pointer in VRAM.
1683  * The driver calls this after a GPU reset to see if the contents of
1684  * VRAM is lost or now.
1685  * returns true if vram is lost, false if not.
1686  */
1687 static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev)
1688 {
1689 	return !!memcmp(adev->gart.ptr, adev->reset_magic,
1690 			AMDGPU_RESET_MAGIC_NUM);
1691 }
1692 
1693 /**
1694  * amdgpu_device_ip_late_set_cg_state - late init for clockgating
1695  *
1696  * @adev: amdgpu_device pointer
1697  *
1698  * Late initialization pass enabling clockgating for hardware IPs.
1699  * The list of all the hardware IPs that make up the asic is walked and the
1700  * set_clockgating_state callbacks are run.  This stage is run late
1701  * in the init process.
1702  * Returns 0 on success, negative error code on failure.
1703  */
1704 static int amdgpu_device_ip_late_set_cg_state(struct amdgpu_device *adev)
1705 {
1706 	int i = 0, r;
1707 
1708 	if (amdgpu_emu_mode == 1)
1709 		return 0;
1710 
1711 	r = amdgpu_ib_ring_tests(adev);
1712 	if (r)
1713 		DRM_ERROR("ib ring test failed (%d).\n", r);
1714 
1715 	for (i = 0; i < adev->num_ip_blocks; i++) {
1716 		if (!adev->ip_blocks[i].status.valid)
1717 			continue;
1718 		/* skip CG for VCE/UVD, it's handled specially */
1719 		if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1720 		    adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1721 		    adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
1722 		    adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1723 			/* enable clockgating to save power */
1724 			r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1725 										     AMD_CG_STATE_GATE);
1726 			if (r) {
1727 				DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
1728 					  adev->ip_blocks[i].version->funcs->name, r);
1729 				return r;
1730 			}
1731 		}
1732 	}
1733 	return 0;
1734 }
1735 
1736 /**
1737  * amdgpu_device_ip_late_init - run late init for hardware IPs
1738  *
1739  * @adev: amdgpu_device pointer
1740  *
1741  * Late initialization pass for hardware IPs.  The list of all the hardware
1742  * IPs that make up the asic is walked and the late_init callbacks are run.
1743  * late_init covers any special initialization that an IP requires
1744  * after all of the have been initialized or something that needs to happen
1745  * late in the init process.
1746  * Returns 0 on success, negative error code on failure.
1747  */
1748 static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
1749 {
1750 	int i = 0, r;
1751 
1752 	for (i = 0; i < adev->num_ip_blocks; i++) {
1753 		if (!adev->ip_blocks[i].status.valid)
1754 			continue;
1755 		if (adev->ip_blocks[i].version->funcs->late_init) {
1756 			r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
1757 			if (r) {
1758 				DRM_ERROR("late_init of IP block <%s> failed %d\n",
1759 					  adev->ip_blocks[i].version->funcs->name, r);
1760 				return r;
1761 			}
1762 			adev->ip_blocks[i].status.late_initialized = true;
1763 		}
1764 	}
1765 
1766 	queue_delayed_work(system_wq, &adev->late_init_work,
1767 			   msecs_to_jiffies(AMDGPU_RESUME_MS));
1768 
1769 	amdgpu_device_fill_reset_magic(adev);
1770 
1771 	return 0;
1772 }
1773 
1774 /**
1775  * amdgpu_device_ip_fini - run fini for hardware IPs
1776  *
1777  * @adev: amdgpu_device pointer
1778  *
1779  * Main teardown pass for hardware IPs.  The list of all the hardware
1780  * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks
1781  * are run.  hw_fini tears down the hardware associated with each IP
1782  * and sw_fini tears down any software state associated with each IP.
1783  * Returns 0 on success, negative error code on failure.
1784  */
1785 static int amdgpu_device_ip_fini(struct amdgpu_device *adev)
1786 {
1787 	int i, r;
1788 
1789 	amdgpu_amdkfd_device_fini(adev);
1790 	/* need to disable SMC first */
1791 	for (i = 0; i < adev->num_ip_blocks; i++) {
1792 		if (!adev->ip_blocks[i].status.hw)
1793 			continue;
1794 		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC &&
1795 			adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1796 			/* ungate blocks before hw fini so that we can shutdown the blocks safely */
1797 			r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1798 										     AMD_CG_STATE_UNGATE);
1799 			if (r) {
1800 				DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1801 					  adev->ip_blocks[i].version->funcs->name, r);
1802 				return r;
1803 			}
1804 			r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
1805 			/* XXX handle errors */
1806 			if (r) {
1807 				DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
1808 					  adev->ip_blocks[i].version->funcs->name, r);
1809 			}
1810 			adev->ip_blocks[i].status.hw = false;
1811 			break;
1812 		}
1813 	}
1814 
1815 	for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1816 		if (!adev->ip_blocks[i].status.hw)
1817 			continue;
1818 
1819 		if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1820 			adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1821 			adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
1822 			adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1823 			/* ungate blocks before hw fini so that we can shutdown the blocks safely */
1824 			r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1825 										     AMD_CG_STATE_UNGATE);
1826 			if (r) {
1827 				DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1828 					  adev->ip_blocks[i].version->funcs->name, r);
1829 				return r;
1830 			}
1831 		}
1832 
1833 		r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
1834 		/* XXX handle errors */
1835 		if (r) {
1836 			DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
1837 				  adev->ip_blocks[i].version->funcs->name, r);
1838 		}
1839 
1840 		adev->ip_blocks[i].status.hw = false;
1841 	}
1842 
1843 
1844 	for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1845 		if (!adev->ip_blocks[i].status.sw)
1846 			continue;
1847 
1848 		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1849 			amdgpu_free_static_csa(adev);
1850 			amdgpu_device_wb_fini(adev);
1851 			amdgpu_device_vram_scratch_fini(adev);
1852 		}
1853 
1854 		r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
1855 		/* XXX handle errors */
1856 		if (r) {
1857 			DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
1858 				  adev->ip_blocks[i].version->funcs->name, r);
1859 		}
1860 		adev->ip_blocks[i].status.sw = false;
1861 		adev->ip_blocks[i].status.valid = false;
1862 	}
1863 
1864 	for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1865 		if (!adev->ip_blocks[i].status.late_initialized)
1866 			continue;
1867 		if (adev->ip_blocks[i].version->funcs->late_fini)
1868 			adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
1869 		adev->ip_blocks[i].status.late_initialized = false;
1870 	}
1871 
1872 	if (amdgpu_sriov_vf(adev))
1873 		if (amdgpu_virt_release_full_gpu(adev, false))
1874 			DRM_ERROR("failed to release exclusive mode on fini\n");
1875 
1876 	return 0;
1877 }
1878 
1879 /**
1880  * amdgpu_device_ip_late_init_func_handler - work handler for clockgating
1881  *
1882  * @work: work_struct
1883  *
1884  * Work handler for amdgpu_device_ip_late_set_cg_state.  We put the
1885  * clockgating setup into a worker thread to speed up driver init and
1886  * resume from suspend.
1887  */
1888 static void amdgpu_device_ip_late_init_func_handler(struct work_struct *work)
1889 {
1890 	struct amdgpu_device *adev =
1891 		container_of(work, struct amdgpu_device, late_init_work.work);
1892 	amdgpu_device_ip_late_set_cg_state(adev);
1893 }
1894 
1895 /**
1896  * amdgpu_device_ip_suspend - run suspend for hardware IPs
1897  *
1898  * @adev: amdgpu_device pointer
1899  *
1900  * Main suspend function for hardware IPs.  The list of all the hardware
1901  * IPs that make up the asic is walked, clockgating is disabled and the
1902  * suspend callbacks are run.  suspend puts the hardware and software state
1903  * in each IP into a state suitable for suspend.
1904  * Returns 0 on success, negative error code on failure.
1905  */
1906 int amdgpu_device_ip_suspend(struct amdgpu_device *adev)
1907 {
1908 	int i, r;
1909 
1910 	if (amdgpu_sriov_vf(adev))
1911 		amdgpu_virt_request_full_gpu(adev, false);
1912 
1913 	/* ungate SMC block powergating */
1914 	if (adev->powerplay.pp_feature & PP_GFXOFF_MASK)
1915 		amdgpu_device_ip_set_powergating_state(adev,
1916 						       AMD_IP_BLOCK_TYPE_SMC,
1917 						       AMD_CG_STATE_UNGATE);
1918 
1919 	/* ungate SMC block first */
1920 	r = amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_SMC,
1921 						   AMD_CG_STATE_UNGATE);
1922 	if (r) {
1923 		DRM_ERROR("set_clockgating_state(ungate) SMC failed %d\n", r);
1924 	}
1925 
1926 	for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1927 		if (!adev->ip_blocks[i].status.valid)
1928 			continue;
1929 		/* ungate blocks so that suspend can properly shut them down */
1930 		if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_SMC &&
1931 			adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1932 			r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1933 										     AMD_CG_STATE_UNGATE);
1934 			if (r) {
1935 				DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1936 					  adev->ip_blocks[i].version->funcs->name, r);
1937 			}
1938 		}
1939 		/* XXX handle errors */
1940 		r = adev->ip_blocks[i].version->funcs->suspend(adev);
1941 		/* XXX handle errors */
1942 		if (r) {
1943 			DRM_ERROR("suspend of IP block <%s> failed %d\n",
1944 				  adev->ip_blocks[i].version->funcs->name, r);
1945 		}
1946 	}
1947 
1948 	if (amdgpu_sriov_vf(adev))
1949 		amdgpu_virt_release_full_gpu(adev, false);
1950 
1951 	return 0;
1952 }
1953 
1954 static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
1955 {
1956 	int i, r;
1957 
1958 	static enum amd_ip_block_type ip_order[] = {
1959 		AMD_IP_BLOCK_TYPE_GMC,
1960 		AMD_IP_BLOCK_TYPE_COMMON,
1961 		AMD_IP_BLOCK_TYPE_IH,
1962 	};
1963 
1964 	for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
1965 		int j;
1966 		struct amdgpu_ip_block *block;
1967 
1968 		for (j = 0; j < adev->num_ip_blocks; j++) {
1969 			block = &adev->ip_blocks[j];
1970 
1971 			if (block->version->type != ip_order[i] ||
1972 				!block->status.valid)
1973 				continue;
1974 
1975 			r = block->version->funcs->hw_init(adev);
1976 			DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"successed");
1977 			if (r)
1978 				return r;
1979 		}
1980 	}
1981 
1982 	return 0;
1983 }
1984 
1985 static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
1986 {
1987 	int i, r;
1988 
1989 	static enum amd_ip_block_type ip_order[] = {
1990 		AMD_IP_BLOCK_TYPE_SMC,
1991 		AMD_IP_BLOCK_TYPE_PSP,
1992 		AMD_IP_BLOCK_TYPE_DCE,
1993 		AMD_IP_BLOCK_TYPE_GFX,
1994 		AMD_IP_BLOCK_TYPE_SDMA,
1995 		AMD_IP_BLOCK_TYPE_UVD,
1996 		AMD_IP_BLOCK_TYPE_VCE
1997 	};
1998 
1999 	for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
2000 		int j;
2001 		struct amdgpu_ip_block *block;
2002 
2003 		for (j = 0; j < adev->num_ip_blocks; j++) {
2004 			block = &adev->ip_blocks[j];
2005 
2006 			if (block->version->type != ip_order[i] ||
2007 				!block->status.valid)
2008 				continue;
2009 
2010 			r = block->version->funcs->hw_init(adev);
2011 			DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"successed");
2012 			if (r)
2013 				return r;
2014 		}
2015 	}
2016 
2017 	return 0;
2018 }
2019 
2020 /**
2021  * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs
2022  *
2023  * @adev: amdgpu_device pointer
2024  *
2025  * First resume function for hardware IPs.  The list of all the hardware
2026  * IPs that make up the asic is walked and the resume callbacks are run for
2027  * COMMON, GMC, and IH.  resume puts the hardware into a functional state
2028  * after a suspend and updates the software state as necessary.  This
2029  * function is also used for restoring the GPU after a GPU reset.
2030  * Returns 0 on success, negative error code on failure.
2031  */
2032 static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)
2033 {
2034 	int i, r;
2035 
2036 	for (i = 0; i < adev->num_ip_blocks; i++) {
2037 		if (!adev->ip_blocks[i].status.valid)
2038 			continue;
2039 		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2040 		    adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2041 		    adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
2042 			r = adev->ip_blocks[i].version->funcs->resume(adev);
2043 			if (r) {
2044 				DRM_ERROR("resume of IP block <%s> failed %d\n",
2045 					  adev->ip_blocks[i].version->funcs->name, r);
2046 				return r;
2047 			}
2048 		}
2049 	}
2050 
2051 	return 0;
2052 }
2053 
2054 /**
2055  * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs
2056  *
2057  * @adev: amdgpu_device pointer
2058  *
2059  * First resume function for hardware IPs.  The list of all the hardware
2060  * IPs that make up the asic is walked and the resume callbacks are run for
2061  * all blocks except COMMON, GMC, and IH.  resume puts the hardware into a
2062  * functional state after a suspend and updates the software state as
2063  * necessary.  This function is also used for restoring the GPU after a GPU
2064  * reset.
2065  * Returns 0 on success, negative error code on failure.
2066  */
2067 static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)
2068 {
2069 	int i, r;
2070 
2071 	for (i = 0; i < adev->num_ip_blocks; i++) {
2072 		if (!adev->ip_blocks[i].status.valid)
2073 			continue;
2074 		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2075 		    adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2076 		    adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH)
2077 			continue;
2078 		r = adev->ip_blocks[i].version->funcs->resume(adev);
2079 		if (r) {
2080 			DRM_ERROR("resume of IP block <%s> failed %d\n",
2081 				  adev->ip_blocks[i].version->funcs->name, r);
2082 			return r;
2083 		}
2084 	}
2085 
2086 	return 0;
2087 }
2088 
2089 /**
2090  * amdgpu_device_ip_resume - run resume for hardware IPs
2091  *
2092  * @adev: amdgpu_device pointer
2093  *
2094  * Main resume function for hardware IPs.  The hardware IPs
2095  * are split into two resume functions because they are
2096  * are also used in in recovering from a GPU reset and some additional
2097  * steps need to be take between them.  In this case (S3/S4) they are
2098  * run sequentially.
2099  * Returns 0 on success, negative error code on failure.
2100  */
2101 static int amdgpu_device_ip_resume(struct amdgpu_device *adev)
2102 {
2103 	int r;
2104 
2105 	r = amdgpu_device_ip_resume_phase1(adev);
2106 	if (r)
2107 		return r;
2108 	r = amdgpu_device_ip_resume_phase2(adev);
2109 
2110 	return r;
2111 }
2112 
2113 /**
2114  * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV
2115  *
2116  * @adev: amdgpu_device pointer
2117  *
2118  * Query the VBIOS data tables to determine if the board supports SR-IOV.
2119  */
2120 static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
2121 {
2122 	if (amdgpu_sriov_vf(adev)) {
2123 		if (adev->is_atom_fw) {
2124 			if (amdgpu_atomfirmware_gpu_supports_virtualization(adev))
2125 				adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2126 		} else {
2127 			if (amdgpu_atombios_has_gpu_virtualization_table(adev))
2128 				adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2129 		}
2130 
2131 		if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
2132 			amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
2133 	}
2134 }
2135 
2136 /**
2137  * amdgpu_device_asic_has_dc_support - determine if DC supports the asic
2138  *
2139  * @asic_type: AMD asic type
2140  *
2141  * Check if there is DC (new modesetting infrastructre) support for an asic.
2142  * returns true if DC has support, false if not.
2143  */
2144 bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
2145 {
2146 	switch (asic_type) {
2147 #if defined(CONFIG_DRM_AMD_DC)
2148 	case CHIP_BONAIRE:
2149 	case CHIP_HAWAII:
2150 	case CHIP_KAVERI:
2151 	case CHIP_KABINI:
2152 	case CHIP_MULLINS:
2153 	case CHIP_CARRIZO:
2154 	case CHIP_STONEY:
2155 	case CHIP_POLARIS10:
2156 	case CHIP_POLARIS11:
2157 	case CHIP_POLARIS12:
2158 	case CHIP_VEGAM:
2159 	case CHIP_TONGA:
2160 	case CHIP_FIJI:
2161 	case CHIP_VEGA10:
2162 	case CHIP_VEGA12:
2163 	case CHIP_VEGA20:
2164 #if defined(CONFIG_DRM_AMD_DC_DCN1_0)
2165 	case CHIP_RAVEN:
2166 #endif
2167 		return amdgpu_dc != 0;
2168 #endif
2169 	default:
2170 		return false;
2171 	}
2172 }
2173 
2174 /**
2175  * amdgpu_device_has_dc_support - check if dc is supported
2176  *
2177  * @adev: amdgpu_device_pointer
2178  *
2179  * Returns true for supported, false for not supported
2180  */
2181 bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
2182 {
2183 	if (amdgpu_sriov_vf(adev))
2184 		return false;
2185 
2186 	return amdgpu_device_asic_has_dc_support(adev->asic_type);
2187 }
2188 
2189 /**
2190  * amdgpu_device_init - initialize the driver
2191  *
2192  * @adev: amdgpu_device pointer
2193  * @pdev: drm dev pointer
2194  * @pdev: pci dev pointer
2195  * @flags: driver flags
2196  *
2197  * Initializes the driver info and hw (all asics).
2198  * Returns 0 for success or an error on failure.
2199  * Called at driver startup.
2200  */
2201 int amdgpu_device_init(struct amdgpu_device *adev,
2202 		       struct drm_device *ddev,
2203 		       struct pci_dev *pdev,
2204 		       uint32_t flags)
2205 {
2206 	int r, i;
2207 	bool runtime = false;
2208 	u32 max_MBps;
2209 
2210 	adev->shutdown = false;
2211 	adev->dev = &pdev->dev;
2212 	adev->ddev = ddev;
2213 	adev->pdev = pdev;
2214 	adev->flags = flags;
2215 	adev->asic_type = flags & AMD_ASIC_MASK;
2216 	adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
2217 	if (amdgpu_emu_mode == 1)
2218 		adev->usec_timeout *= 2;
2219 	adev->gmc.gart_size = 512 * 1024 * 1024;
2220 	adev->accel_working = false;
2221 	adev->num_rings = 0;
2222 	adev->mman.buffer_funcs = NULL;
2223 	adev->mman.buffer_funcs_ring = NULL;
2224 	adev->vm_manager.vm_pte_funcs = NULL;
2225 	adev->vm_manager.vm_pte_num_rings = 0;
2226 	adev->gmc.gmc_funcs = NULL;
2227 	adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
2228 	bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
2229 
2230 	adev->smc_rreg = &amdgpu_invalid_rreg;
2231 	adev->smc_wreg = &amdgpu_invalid_wreg;
2232 	adev->pcie_rreg = &amdgpu_invalid_rreg;
2233 	adev->pcie_wreg = &amdgpu_invalid_wreg;
2234 	adev->pciep_rreg = &amdgpu_invalid_rreg;
2235 	adev->pciep_wreg = &amdgpu_invalid_wreg;
2236 	adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
2237 	adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
2238 	adev->didt_rreg = &amdgpu_invalid_rreg;
2239 	adev->didt_wreg = &amdgpu_invalid_wreg;
2240 	adev->gc_cac_rreg = &amdgpu_invalid_rreg;
2241 	adev->gc_cac_wreg = &amdgpu_invalid_wreg;
2242 	adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
2243 	adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
2244 
2245 	DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
2246 		 amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
2247 		 pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
2248 
2249 	/* mutex initialization are all done here so we
2250 	 * can recall function without having locking issues */
2251 	atomic_set(&adev->irq.ih.lock, 0);
2252 	mutex_init(&adev->firmware.mutex);
2253 	mutex_init(&adev->pm.mutex);
2254 	mutex_init(&adev->gfx.gpu_clock_mutex);
2255 	mutex_init(&adev->srbm_mutex);
2256 	mutex_init(&adev->gfx.pipe_reserve_mutex);
2257 	mutex_init(&adev->grbm_idx_mutex);
2258 	mutex_init(&adev->mn_lock);
2259 	mutex_init(&adev->virt.vf_errors.lock);
2260 	hash_init(adev->mn_hash);
2261 	mutex_init(&adev->lock_reset);
2262 
2263 	amdgpu_device_check_arguments(adev);
2264 
2265 	spin_lock_init(&adev->mmio_idx_lock);
2266 	spin_lock_init(&adev->smc_idx_lock);
2267 	spin_lock_init(&adev->pcie_idx_lock);
2268 	spin_lock_init(&adev->uvd_ctx_idx_lock);
2269 	spin_lock_init(&adev->didt_idx_lock);
2270 	spin_lock_init(&adev->gc_cac_idx_lock);
2271 	spin_lock_init(&adev->se_cac_idx_lock);
2272 	spin_lock_init(&adev->audio_endpt_idx_lock);
2273 	spin_lock_init(&adev->mm_stats.lock);
2274 
2275 	INIT_LIST_HEAD(&adev->shadow_list);
2276 	mutex_init(&adev->shadow_list_lock);
2277 
2278 	INIT_LIST_HEAD(&adev->ring_lru_list);
2279 	spin_lock_init(&adev->ring_lru_list_lock);
2280 
2281 	INIT_DELAYED_WORK(&adev->late_init_work,
2282 			  amdgpu_device_ip_late_init_func_handler);
2283 
2284 	/* Registers mapping */
2285 	/* TODO: block userspace mapping of io register */
2286 	if (adev->asic_type >= CHIP_BONAIRE) {
2287 		adev->rmmio_base = pci_resource_start(adev->pdev, 5);
2288 		adev->rmmio_size = pci_resource_len(adev->pdev, 5);
2289 	} else {
2290 		adev->rmmio_base = pci_resource_start(adev->pdev, 2);
2291 		adev->rmmio_size = pci_resource_len(adev->pdev, 2);
2292 	}
2293 
2294 	adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
2295 	if (adev->rmmio == NULL) {
2296 		return -ENOMEM;
2297 	}
2298 	DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
2299 	DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
2300 
2301 	/* doorbell bar mapping */
2302 	amdgpu_device_doorbell_init(adev);
2303 
2304 	/* io port mapping */
2305 	for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
2306 		if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) {
2307 			adev->rio_mem_size = pci_resource_len(adev->pdev, i);
2308 			adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size);
2309 			break;
2310 		}
2311 	}
2312 	if (adev->rio_mem == NULL)
2313 		DRM_INFO("PCI I/O BAR is not found.\n");
2314 
2315 	amdgpu_device_get_pcie_info(adev);
2316 
2317 	/* early init functions */
2318 	r = amdgpu_device_ip_early_init(adev);
2319 	if (r)
2320 		return r;
2321 
2322 	/* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
2323 	/* this will fail for cards that aren't VGA class devices, just
2324 	 * ignore it */
2325 	vga_client_register(adev->pdev, adev, NULL, amdgpu_device_vga_set_decode);
2326 
2327 	if (amdgpu_device_is_px(ddev))
2328 		runtime = true;
2329 	if (!pci_is_thunderbolt_attached(adev->pdev))
2330 		vga_switcheroo_register_client(adev->pdev,
2331 					       &amdgpu_switcheroo_ops, runtime);
2332 	if (runtime)
2333 		vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
2334 
2335 	if (amdgpu_emu_mode == 1) {
2336 		/* post the asic on emulation mode */
2337 		emu_soc_asic_init(adev);
2338 		goto fence_driver_init;
2339 	}
2340 
2341 	/* Read BIOS */
2342 	if (!amdgpu_get_bios(adev)) {
2343 		r = -EINVAL;
2344 		goto failed;
2345 	}
2346 
2347 	r = amdgpu_atombios_init(adev);
2348 	if (r) {
2349 		dev_err(adev->dev, "amdgpu_atombios_init failed\n");
2350 		amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0);
2351 		goto failed;
2352 	}
2353 
2354 	/* detect if we are with an SRIOV vbios */
2355 	amdgpu_device_detect_sriov_bios(adev);
2356 
2357 	/* Post card if necessary */
2358 	if (amdgpu_device_need_post(adev)) {
2359 		if (!adev->bios) {
2360 			dev_err(adev->dev, "no vBIOS found\n");
2361 			r = -EINVAL;
2362 			goto failed;
2363 		}
2364 		DRM_INFO("GPU posting now...\n");
2365 		r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2366 		if (r) {
2367 			dev_err(adev->dev, "gpu post error!\n");
2368 			goto failed;
2369 		}
2370 	}
2371 
2372 	if (adev->is_atom_fw) {
2373 		/* Initialize clocks */
2374 		r = amdgpu_atomfirmware_get_clock_info(adev);
2375 		if (r) {
2376 			dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
2377 			amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
2378 			goto failed;
2379 		}
2380 	} else {
2381 		/* Initialize clocks */
2382 		r = amdgpu_atombios_get_clock_info(adev);
2383 		if (r) {
2384 			dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
2385 			amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
2386 			goto failed;
2387 		}
2388 		/* init i2c buses */
2389 		if (!amdgpu_device_has_dc_support(adev))
2390 			amdgpu_atombios_i2c_init(adev);
2391 	}
2392 
2393 fence_driver_init:
2394 	/* Fence driver */
2395 	r = amdgpu_fence_driver_init(adev);
2396 	if (r) {
2397 		dev_err(adev->dev, "amdgpu_fence_driver_init failed\n");
2398 		amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
2399 		goto failed;
2400 	}
2401 
2402 	/* init the mode config */
2403 	drm_mode_config_init(adev->ddev);
2404 
2405 	r = amdgpu_device_ip_init(adev);
2406 	if (r) {
2407 		/* failed in exclusive mode due to timeout */
2408 		if (amdgpu_sriov_vf(adev) &&
2409 		    !amdgpu_sriov_runtime(adev) &&
2410 		    amdgpu_virt_mmio_blocked(adev) &&
2411 		    !amdgpu_virt_wait_reset(adev)) {
2412 			dev_err(adev->dev, "VF exclusive mode timeout\n");
2413 			/* Don't send request since VF is inactive. */
2414 			adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
2415 			adev->virt.ops = NULL;
2416 			r = -EAGAIN;
2417 			goto failed;
2418 		}
2419 		dev_err(adev->dev, "amdgpu_device_ip_init failed\n");
2420 		amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
2421 		goto failed;
2422 	}
2423 
2424 	adev->accel_working = true;
2425 
2426 	amdgpu_vm_check_compute_bug(adev);
2427 
2428 	/* Initialize the buffer migration limit. */
2429 	if (amdgpu_moverate >= 0)
2430 		max_MBps = amdgpu_moverate;
2431 	else
2432 		max_MBps = 8; /* Allow 8 MB/s. */
2433 	/* Get a log2 for easy divisions. */
2434 	adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
2435 
2436 	r = amdgpu_ib_pool_init(adev);
2437 	if (r) {
2438 		dev_err(adev->dev, "IB initialization failed (%d).\n", r);
2439 		amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r);
2440 		goto failed;
2441 	}
2442 
2443 	if (amdgpu_sriov_vf(adev))
2444 		amdgpu_virt_init_data_exchange(adev);
2445 
2446 	amdgpu_fbdev_init(adev);
2447 
2448 	r = amdgpu_pm_sysfs_init(adev);
2449 	if (r)
2450 		DRM_ERROR("registering pm debugfs failed (%d).\n", r);
2451 
2452 	r = amdgpu_debugfs_gem_init(adev);
2453 	if (r)
2454 		DRM_ERROR("registering gem debugfs failed (%d).\n", r);
2455 
2456 	r = amdgpu_debugfs_regs_init(adev);
2457 	if (r)
2458 		DRM_ERROR("registering register debugfs failed (%d).\n", r);
2459 
2460 	r = amdgpu_debugfs_firmware_init(adev);
2461 	if (r)
2462 		DRM_ERROR("registering firmware debugfs failed (%d).\n", r);
2463 
2464 	r = amdgpu_debugfs_init(adev);
2465 	if (r)
2466 		DRM_ERROR("Creating debugfs files failed (%d).\n", r);
2467 
2468 	if ((amdgpu_testing & 1)) {
2469 		if (adev->accel_working)
2470 			amdgpu_test_moves(adev);
2471 		else
2472 			DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n");
2473 	}
2474 	if (amdgpu_benchmarking) {
2475 		if (adev->accel_working)
2476 			amdgpu_benchmark(adev, amdgpu_benchmarking);
2477 		else
2478 			DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n");
2479 	}
2480 
2481 	/* enable clockgating, etc. after ib tests, etc. since some blocks require
2482 	 * explicit gating rather than handling it automatically.
2483 	 */
2484 	r = amdgpu_device_ip_late_init(adev);
2485 	if (r) {
2486 		dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n");
2487 		amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
2488 		goto failed;
2489 	}
2490 
2491 	return 0;
2492 
2493 failed:
2494 	amdgpu_vf_error_trans_all(adev);
2495 	if (runtime)
2496 		vga_switcheroo_fini_domain_pm_ops(adev->dev);
2497 
2498 	return r;
2499 }
2500 
2501 /**
2502  * amdgpu_device_fini - tear down the driver
2503  *
2504  * @adev: amdgpu_device pointer
2505  *
2506  * Tear down the driver info (all asics).
2507  * Called at driver shutdown.
2508  */
2509 void amdgpu_device_fini(struct amdgpu_device *adev)
2510 {
2511 	int r;
2512 
2513 	DRM_INFO("amdgpu: finishing device.\n");
2514 	adev->shutdown = true;
2515 	/* disable all interrupts */
2516 	amdgpu_irq_disable_all(adev);
2517 	if (adev->mode_info.mode_config_initialized){
2518 		if (!amdgpu_device_has_dc_support(adev))
2519 			drm_crtc_force_disable_all(adev->ddev);
2520 		else
2521 			drm_atomic_helper_shutdown(adev->ddev);
2522 	}
2523 	amdgpu_ib_pool_fini(adev);
2524 	amdgpu_fence_driver_fini(adev);
2525 	amdgpu_pm_sysfs_fini(adev);
2526 	amdgpu_fbdev_fini(adev);
2527 	r = amdgpu_device_ip_fini(adev);
2528 	if (adev->firmware.gpu_info_fw) {
2529 		release_firmware(adev->firmware.gpu_info_fw);
2530 		adev->firmware.gpu_info_fw = NULL;
2531 	}
2532 	adev->accel_working = false;
2533 	cancel_delayed_work_sync(&adev->late_init_work);
2534 	/* free i2c buses */
2535 	if (!amdgpu_device_has_dc_support(adev))
2536 		amdgpu_i2c_fini(adev);
2537 
2538 	if (amdgpu_emu_mode != 1)
2539 		amdgpu_atombios_fini(adev);
2540 
2541 	kfree(adev->bios);
2542 	adev->bios = NULL;
2543 	if (!pci_is_thunderbolt_attached(adev->pdev))
2544 		vga_switcheroo_unregister_client(adev->pdev);
2545 	if (adev->flags & AMD_IS_PX)
2546 		vga_switcheroo_fini_domain_pm_ops(adev->dev);
2547 	vga_client_register(adev->pdev, NULL, NULL, NULL);
2548 	if (adev->rio_mem)
2549 		pci_iounmap(adev->pdev, adev->rio_mem);
2550 	adev->rio_mem = NULL;
2551 	iounmap(adev->rmmio);
2552 	adev->rmmio = NULL;
2553 	amdgpu_device_doorbell_fini(adev);
2554 	amdgpu_debugfs_regs_cleanup(adev);
2555 }
2556 
2557 
2558 /*
2559  * Suspend & resume.
2560  */
2561 /**
2562  * amdgpu_device_suspend - initiate device suspend
2563  *
2564  * @pdev: drm dev pointer
2565  * @state: suspend state
2566  *
2567  * Puts the hw in the suspend state (all asics).
2568  * Returns 0 for success or an error on failure.
2569  * Called at driver suspend.
2570  */
2571 int amdgpu_device_suspend(struct drm_device *dev, bool suspend, bool fbcon)
2572 {
2573 	struct amdgpu_device *adev;
2574 	struct drm_crtc *crtc;
2575 	struct drm_connector *connector;
2576 	int r;
2577 
2578 	if (dev == NULL || dev->dev_private == NULL) {
2579 		return -ENODEV;
2580 	}
2581 
2582 	adev = dev->dev_private;
2583 
2584 	if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
2585 		return 0;
2586 
2587 	drm_kms_helper_poll_disable(dev);
2588 
2589 	if (!amdgpu_device_has_dc_support(adev)) {
2590 		/* turn off display hw */
2591 		drm_modeset_lock_all(dev);
2592 		list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2593 			drm_helper_connector_dpms(connector, DRM_MODE_DPMS_OFF);
2594 		}
2595 		drm_modeset_unlock_all(dev);
2596 	}
2597 
2598 	amdgpu_amdkfd_suspend(adev);
2599 
2600 	/* unpin the front buffers and cursors */
2601 	list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2602 		struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2603 		struct drm_framebuffer *fb = crtc->primary->fb;
2604 		struct amdgpu_bo *robj;
2605 
2606 		if (amdgpu_crtc->cursor_bo) {
2607 			struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2608 			r = amdgpu_bo_reserve(aobj, true);
2609 			if (r == 0) {
2610 				amdgpu_bo_unpin(aobj);
2611 				amdgpu_bo_unreserve(aobj);
2612 			}
2613 		}
2614 
2615 		if (fb == NULL || fb->obj[0] == NULL) {
2616 			continue;
2617 		}
2618 		robj = gem_to_amdgpu_bo(fb->obj[0]);
2619 		/* don't unpin kernel fb objects */
2620 		if (!amdgpu_fbdev_robj_is_fb(adev, robj)) {
2621 			r = amdgpu_bo_reserve(robj, true);
2622 			if (r == 0) {
2623 				amdgpu_bo_unpin(robj);
2624 				amdgpu_bo_unreserve(robj);
2625 			}
2626 		}
2627 	}
2628 	/* evict vram memory */
2629 	amdgpu_bo_evict_vram(adev);
2630 
2631 	amdgpu_fence_driver_suspend(adev);
2632 
2633 	r = amdgpu_device_ip_suspend(adev);
2634 
2635 	/* evict remaining vram memory
2636 	 * This second call to evict vram is to evict the gart page table
2637 	 * using the CPU.
2638 	 */
2639 	amdgpu_bo_evict_vram(adev);
2640 
2641 	pci_save_state(dev->pdev);
2642 	if (suspend) {
2643 		/* Shut down the device */
2644 		pci_disable_device(dev->pdev);
2645 		pci_set_power_state(dev->pdev, PCI_D3hot);
2646 	} else {
2647 		r = amdgpu_asic_reset(adev);
2648 		if (r)
2649 			DRM_ERROR("amdgpu asic reset failed\n");
2650 	}
2651 
2652 	if (fbcon) {
2653 		console_lock();
2654 		amdgpu_fbdev_set_suspend(adev, 1);
2655 		console_unlock();
2656 	}
2657 	return 0;
2658 }
2659 
2660 /**
2661  * amdgpu_device_resume - initiate device resume
2662  *
2663  * @pdev: drm dev pointer
2664  *
2665  * Bring the hw back to operating state (all asics).
2666  * Returns 0 for success or an error on failure.
2667  * Called at driver resume.
2668  */
2669 int amdgpu_device_resume(struct drm_device *dev, bool resume, bool fbcon)
2670 {
2671 	struct drm_connector *connector;
2672 	struct amdgpu_device *adev = dev->dev_private;
2673 	struct drm_crtc *crtc;
2674 	int r = 0;
2675 
2676 	if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
2677 		return 0;
2678 
2679 	if (fbcon)
2680 		console_lock();
2681 
2682 	if (resume) {
2683 		pci_set_power_state(dev->pdev, PCI_D0);
2684 		pci_restore_state(dev->pdev);
2685 		r = pci_enable_device(dev->pdev);
2686 		if (r)
2687 			goto unlock;
2688 	}
2689 
2690 	/* post card */
2691 	if (amdgpu_device_need_post(adev)) {
2692 		r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2693 		if (r)
2694 			DRM_ERROR("amdgpu asic init failed\n");
2695 	}
2696 
2697 	r = amdgpu_device_ip_resume(adev);
2698 	if (r) {
2699 		DRM_ERROR("amdgpu_device_ip_resume failed (%d).\n", r);
2700 		goto unlock;
2701 	}
2702 	amdgpu_fence_driver_resume(adev);
2703 
2704 
2705 	r = amdgpu_device_ip_late_init(adev);
2706 	if (r)
2707 		goto unlock;
2708 
2709 	/* pin cursors */
2710 	list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2711 		struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2712 
2713 		if (amdgpu_crtc->cursor_bo) {
2714 			struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2715 			r = amdgpu_bo_reserve(aobj, true);
2716 			if (r == 0) {
2717 				r = amdgpu_bo_pin(aobj,
2718 						  AMDGPU_GEM_DOMAIN_VRAM,
2719 						  &amdgpu_crtc->cursor_addr);
2720 				if (r != 0)
2721 					DRM_ERROR("Failed to pin cursor BO (%d)\n", r);
2722 				amdgpu_bo_unreserve(aobj);
2723 			}
2724 		}
2725 	}
2726 	r = amdgpu_amdkfd_resume(adev);
2727 	if (r)
2728 		return r;
2729 
2730 	/* blat the mode back in */
2731 	if (fbcon) {
2732 		if (!amdgpu_device_has_dc_support(adev)) {
2733 			/* pre DCE11 */
2734 			drm_helper_resume_force_mode(dev);
2735 
2736 			/* turn on display hw */
2737 			drm_modeset_lock_all(dev);
2738 			list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2739 				drm_helper_connector_dpms(connector, DRM_MODE_DPMS_ON);
2740 			}
2741 			drm_modeset_unlock_all(dev);
2742 		}
2743 	}
2744 
2745 	drm_kms_helper_poll_enable(dev);
2746 
2747 	/*
2748 	 * Most of the connector probing functions try to acquire runtime pm
2749 	 * refs to ensure that the GPU is powered on when connector polling is
2750 	 * performed. Since we're calling this from a runtime PM callback,
2751 	 * trying to acquire rpm refs will cause us to deadlock.
2752 	 *
2753 	 * Since we're guaranteed to be holding the rpm lock, it's safe to
2754 	 * temporarily disable the rpm helpers so this doesn't deadlock us.
2755 	 */
2756 #ifdef CONFIG_PM
2757 	dev->dev->power.disable_depth++;
2758 #endif
2759 	if (!amdgpu_device_has_dc_support(adev))
2760 		drm_helper_hpd_irq_event(dev);
2761 	else
2762 		drm_kms_helper_hotplug_event(dev);
2763 #ifdef CONFIG_PM
2764 	dev->dev->power.disable_depth--;
2765 #endif
2766 
2767 	if (fbcon)
2768 		amdgpu_fbdev_set_suspend(adev, 0);
2769 
2770 unlock:
2771 	if (fbcon)
2772 		console_unlock();
2773 
2774 	return r;
2775 }
2776 
2777 /**
2778  * amdgpu_device_ip_check_soft_reset - did soft reset succeed
2779  *
2780  * @adev: amdgpu_device pointer
2781  *
2782  * The list of all the hardware IPs that make up the asic is walked and
2783  * the check_soft_reset callbacks are run.  check_soft_reset determines
2784  * if the asic is still hung or not.
2785  * Returns true if any of the IPs are still in a hung state, false if not.
2786  */
2787 static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)
2788 {
2789 	int i;
2790 	bool asic_hang = false;
2791 
2792 	if (amdgpu_sriov_vf(adev))
2793 		return true;
2794 
2795 	if (amdgpu_asic_need_full_reset(adev))
2796 		return true;
2797 
2798 	for (i = 0; i < adev->num_ip_blocks; i++) {
2799 		if (!adev->ip_blocks[i].status.valid)
2800 			continue;
2801 		if (adev->ip_blocks[i].version->funcs->check_soft_reset)
2802 			adev->ip_blocks[i].status.hang =
2803 				adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
2804 		if (adev->ip_blocks[i].status.hang) {
2805 			DRM_INFO("IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
2806 			asic_hang = true;
2807 		}
2808 	}
2809 	return asic_hang;
2810 }
2811 
2812 /**
2813  * amdgpu_device_ip_pre_soft_reset - prepare for soft reset
2814  *
2815  * @adev: amdgpu_device pointer
2816  *
2817  * The list of all the hardware IPs that make up the asic is walked and the
2818  * pre_soft_reset callbacks are run if the block is hung.  pre_soft_reset
2819  * handles any IP specific hardware or software state changes that are
2820  * necessary for a soft reset to succeed.
2821  * Returns 0 on success, negative error code on failure.
2822  */
2823 static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)
2824 {
2825 	int i, r = 0;
2826 
2827 	for (i = 0; i < adev->num_ip_blocks; i++) {
2828 		if (!adev->ip_blocks[i].status.valid)
2829 			continue;
2830 		if (adev->ip_blocks[i].status.hang &&
2831 		    adev->ip_blocks[i].version->funcs->pre_soft_reset) {
2832 			r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
2833 			if (r)
2834 				return r;
2835 		}
2836 	}
2837 
2838 	return 0;
2839 }
2840 
2841 /**
2842  * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed
2843  *
2844  * @adev: amdgpu_device pointer
2845  *
2846  * Some hardware IPs cannot be soft reset.  If they are hung, a full gpu
2847  * reset is necessary to recover.
2848  * Returns true if a full asic reset is required, false if not.
2849  */
2850 static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)
2851 {
2852 	int i;
2853 
2854 	if (amdgpu_asic_need_full_reset(adev))
2855 		return true;
2856 
2857 	for (i = 0; i < adev->num_ip_blocks; i++) {
2858 		if (!adev->ip_blocks[i].status.valid)
2859 			continue;
2860 		if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
2861 		    (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
2862 		    (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
2863 		    (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
2864 		     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
2865 			if (adev->ip_blocks[i].status.hang) {
2866 				DRM_INFO("Some block need full reset!\n");
2867 				return true;
2868 			}
2869 		}
2870 	}
2871 	return false;
2872 }
2873 
2874 /**
2875  * amdgpu_device_ip_soft_reset - do a soft reset
2876  *
2877  * @adev: amdgpu_device pointer
2878  *
2879  * The list of all the hardware IPs that make up the asic is walked and the
2880  * soft_reset callbacks are run if the block is hung.  soft_reset handles any
2881  * IP specific hardware or software state changes that are necessary to soft
2882  * reset the IP.
2883  * Returns 0 on success, negative error code on failure.
2884  */
2885 static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)
2886 {
2887 	int i, r = 0;
2888 
2889 	for (i = 0; i < adev->num_ip_blocks; i++) {
2890 		if (!adev->ip_blocks[i].status.valid)
2891 			continue;
2892 		if (adev->ip_blocks[i].status.hang &&
2893 		    adev->ip_blocks[i].version->funcs->soft_reset) {
2894 			r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
2895 			if (r)
2896 				return r;
2897 		}
2898 	}
2899 
2900 	return 0;
2901 }
2902 
2903 /**
2904  * amdgpu_device_ip_post_soft_reset - clean up from soft reset
2905  *
2906  * @adev: amdgpu_device pointer
2907  *
2908  * The list of all the hardware IPs that make up the asic is walked and the
2909  * post_soft_reset callbacks are run if the asic was hung.  post_soft_reset
2910  * handles any IP specific hardware or software state changes that are
2911  * necessary after the IP has been soft reset.
2912  * Returns 0 on success, negative error code on failure.
2913  */
2914 static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)
2915 {
2916 	int i, r = 0;
2917 
2918 	for (i = 0; i < adev->num_ip_blocks; i++) {
2919 		if (!adev->ip_blocks[i].status.valid)
2920 			continue;
2921 		if (adev->ip_blocks[i].status.hang &&
2922 		    adev->ip_blocks[i].version->funcs->post_soft_reset)
2923 			r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
2924 		if (r)
2925 			return r;
2926 	}
2927 
2928 	return 0;
2929 }
2930 
2931 /**
2932  * amdgpu_device_recover_vram_from_shadow - restore shadowed VRAM buffers
2933  *
2934  * @adev: amdgpu_device pointer
2935  * @ring: amdgpu_ring for the engine handling the buffer operations
2936  * @bo: amdgpu_bo buffer whose shadow is being restored
2937  * @fence: dma_fence associated with the operation
2938  *
2939  * Restores the VRAM buffer contents from the shadow in GTT.  Used to
2940  * restore things like GPUVM page tables after a GPU reset where
2941  * the contents of VRAM might be lost.
2942  * Returns 0 on success, negative error code on failure.
2943  */
2944 static int amdgpu_device_recover_vram_from_shadow(struct amdgpu_device *adev,
2945 						  struct amdgpu_ring *ring,
2946 						  struct amdgpu_bo *bo,
2947 						  struct dma_fence **fence)
2948 {
2949 	uint32_t domain;
2950 	int r;
2951 
2952 	if (!bo->shadow)
2953 		return 0;
2954 
2955 	r = amdgpu_bo_reserve(bo, true);
2956 	if (r)
2957 		return r;
2958 	domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
2959 	/* if bo has been evicted, then no need to recover */
2960 	if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
2961 		r = amdgpu_bo_validate(bo->shadow);
2962 		if (r) {
2963 			DRM_ERROR("bo validate failed!\n");
2964 			goto err;
2965 		}
2966 
2967 		r = amdgpu_bo_restore_from_shadow(adev, ring, bo,
2968 						 NULL, fence, true);
2969 		if (r) {
2970 			DRM_ERROR("recover page table failed!\n");
2971 			goto err;
2972 		}
2973 	}
2974 err:
2975 	amdgpu_bo_unreserve(bo);
2976 	return r;
2977 }
2978 
2979 /**
2980  * amdgpu_device_handle_vram_lost - Handle the loss of VRAM contents
2981  *
2982  * @adev: amdgpu_device pointer
2983  *
2984  * Restores the contents of VRAM buffers from the shadows in GTT.  Used to
2985  * restore things like GPUVM page tables after a GPU reset where
2986  * the contents of VRAM might be lost.
2987  * Returns 0 on success, 1 on failure.
2988  */
2989 static int amdgpu_device_handle_vram_lost(struct amdgpu_device *adev)
2990 {
2991 	struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
2992 	struct amdgpu_bo *bo, *tmp;
2993 	struct dma_fence *fence = NULL, *next = NULL;
2994 	long r = 1;
2995 	int i = 0;
2996 	long tmo;
2997 
2998 	if (amdgpu_sriov_runtime(adev))
2999 		tmo = msecs_to_jiffies(amdgpu_lockup_timeout);
3000 	else
3001 		tmo = msecs_to_jiffies(100);
3002 
3003 	DRM_INFO("recover vram bo from shadow start\n");
3004 	mutex_lock(&adev->shadow_list_lock);
3005 	list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) {
3006 		next = NULL;
3007 		amdgpu_device_recover_vram_from_shadow(adev, ring, bo, &next);
3008 		if (fence) {
3009 			r = dma_fence_wait_timeout(fence, false, tmo);
3010 			if (r == 0)
3011 				pr_err("wait fence %p[%d] timeout\n", fence, i);
3012 			else if (r < 0)
3013 				pr_err("wait fence %p[%d] interrupted\n", fence, i);
3014 			if (r < 1) {
3015 				dma_fence_put(fence);
3016 				fence = next;
3017 				break;
3018 			}
3019 			i++;
3020 		}
3021 
3022 		dma_fence_put(fence);
3023 		fence = next;
3024 	}
3025 	mutex_unlock(&adev->shadow_list_lock);
3026 
3027 	if (fence) {
3028 		r = dma_fence_wait_timeout(fence, false, tmo);
3029 		if (r == 0)
3030 			pr_err("wait fence %p[%d] timeout\n", fence, i);
3031 		else if (r < 0)
3032 			pr_err("wait fence %p[%d] interrupted\n", fence, i);
3033 
3034 	}
3035 	dma_fence_put(fence);
3036 
3037 	if (r > 0)
3038 		DRM_INFO("recover vram bo from shadow done\n");
3039 	else
3040 		DRM_ERROR("recover vram bo from shadow failed\n");
3041 
3042 	return (r > 0) ? 0 : 1;
3043 }
3044 
3045 /**
3046  * amdgpu_device_reset - reset ASIC/GPU for bare-metal or passthrough
3047  *
3048  * @adev: amdgpu device pointer
3049  *
3050  * attempt to do soft-reset or full-reset and reinitialize Asic
3051  * return 0 means successed otherwise failed
3052  */
3053 static int amdgpu_device_reset(struct amdgpu_device *adev)
3054 {
3055 	bool need_full_reset, vram_lost = 0;
3056 	int r;
3057 
3058 	need_full_reset = amdgpu_device_ip_need_full_reset(adev);
3059 
3060 	if (!need_full_reset) {
3061 		amdgpu_device_ip_pre_soft_reset(adev);
3062 		r = amdgpu_device_ip_soft_reset(adev);
3063 		amdgpu_device_ip_post_soft_reset(adev);
3064 		if (r || amdgpu_device_ip_check_soft_reset(adev)) {
3065 			DRM_INFO("soft reset failed, will fallback to full reset!\n");
3066 			need_full_reset = true;
3067 		}
3068 	}
3069 
3070 	if (need_full_reset) {
3071 		r = amdgpu_device_ip_suspend(adev);
3072 
3073 retry:
3074 		r = amdgpu_asic_reset(adev);
3075 		/* post card */
3076 		amdgpu_atom_asic_init(adev->mode_info.atom_context);
3077 
3078 		if (!r) {
3079 			dev_info(adev->dev, "GPU reset succeeded, trying to resume\n");
3080 			r = amdgpu_device_ip_resume_phase1(adev);
3081 			if (r)
3082 				goto out;
3083 
3084 			vram_lost = amdgpu_device_check_vram_lost(adev);
3085 			if (vram_lost) {
3086 				DRM_ERROR("VRAM is lost!\n");
3087 				atomic_inc(&adev->vram_lost_counter);
3088 			}
3089 
3090 			r = amdgpu_gtt_mgr_recover(
3091 				&adev->mman.bdev.man[TTM_PL_TT]);
3092 			if (r)
3093 				goto out;
3094 
3095 			r = amdgpu_device_ip_resume_phase2(adev);
3096 			if (r)
3097 				goto out;
3098 
3099 			if (vram_lost)
3100 				amdgpu_device_fill_reset_magic(adev);
3101 		}
3102 	}
3103 
3104 out:
3105 	if (!r) {
3106 		amdgpu_irq_gpu_reset_resume_helper(adev);
3107 		r = amdgpu_ib_ring_tests(adev);
3108 		if (r) {
3109 			dev_err(adev->dev, "ib ring test failed (%d).\n", r);
3110 			r = amdgpu_device_ip_suspend(adev);
3111 			need_full_reset = true;
3112 			goto retry;
3113 		}
3114 	}
3115 
3116 	if (!r && ((need_full_reset && !(adev->flags & AMD_IS_APU)) || vram_lost))
3117 		r = amdgpu_device_handle_vram_lost(adev);
3118 
3119 	return r;
3120 }
3121 
3122 /**
3123  * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
3124  *
3125  * @adev: amdgpu device pointer
3126  *
3127  * do VF FLR and reinitialize Asic
3128  * return 0 means successed otherwise failed
3129  */
3130 static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
3131 				     bool from_hypervisor)
3132 {
3133 	int r;
3134 
3135 	if (from_hypervisor)
3136 		r = amdgpu_virt_request_full_gpu(adev, true);
3137 	else
3138 		r = amdgpu_virt_reset_gpu(adev);
3139 	if (r)
3140 		return r;
3141 
3142 	/* Resume IP prior to SMC */
3143 	r = amdgpu_device_ip_reinit_early_sriov(adev);
3144 	if (r)
3145 		goto error;
3146 
3147 	/* we need recover gart prior to run SMC/CP/SDMA resume */
3148 	amdgpu_gtt_mgr_recover(&adev->mman.bdev.man[TTM_PL_TT]);
3149 
3150 	/* now we are okay to resume SMC/CP/SDMA */
3151 	r = amdgpu_device_ip_reinit_late_sriov(adev);
3152 	if (r)
3153 		goto error;
3154 
3155 	amdgpu_irq_gpu_reset_resume_helper(adev);
3156 	r = amdgpu_ib_ring_tests(adev);
3157 
3158 error:
3159 	amdgpu_virt_release_full_gpu(adev, true);
3160 	if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
3161 		atomic_inc(&adev->vram_lost_counter);
3162 		r = amdgpu_device_handle_vram_lost(adev);
3163 	}
3164 
3165 	return r;
3166 }
3167 
3168 /**
3169  * amdgpu_device_gpu_recover - reset the asic and recover scheduler
3170  *
3171  * @adev: amdgpu device pointer
3172  * @job: which job trigger hang
3173  * @force forces reset regardless of amdgpu_gpu_recovery
3174  *
3175  * Attempt to reset the GPU if it has hung (all asics).
3176  * Returns 0 for success or an error on failure.
3177  */
3178 int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
3179 			      struct amdgpu_job *job, bool force)
3180 {
3181 	int i, r, resched;
3182 
3183 	if (!force && !amdgpu_device_ip_check_soft_reset(adev)) {
3184 		DRM_INFO("No hardware hang detected. Did some blocks stall?\n");
3185 		return 0;
3186 	}
3187 
3188 	if (!force && (amdgpu_gpu_recovery == 0 ||
3189 			(amdgpu_gpu_recovery == -1  && !amdgpu_sriov_vf(adev)))) {
3190 		DRM_INFO("GPU recovery disabled.\n");
3191 		return 0;
3192 	}
3193 
3194 	dev_info(adev->dev, "GPU reset begin!\n");
3195 
3196 	mutex_lock(&adev->lock_reset);
3197 	atomic_inc(&adev->gpu_reset_counter);
3198 	adev->in_gpu_reset = 1;
3199 
3200 	/* block TTM */
3201 	resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
3202 
3203 	/* block all schedulers and reset given job's ring */
3204 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3205 		struct amdgpu_ring *ring = adev->rings[i];
3206 
3207 		if (!ring || !ring->sched.thread)
3208 			continue;
3209 
3210 		kthread_park(ring->sched.thread);
3211 
3212 		if (job && job->ring->idx != i)
3213 			continue;
3214 
3215 		drm_sched_hw_job_reset(&ring->sched, &job->base);
3216 
3217 		/* after all hw jobs are reset, hw fence is meaningless, so force_completion */
3218 		amdgpu_fence_driver_force_completion(ring);
3219 	}
3220 
3221 	if (amdgpu_sriov_vf(adev))
3222 		r = amdgpu_device_reset_sriov(adev, job ? false : true);
3223 	else
3224 		r = amdgpu_device_reset(adev);
3225 
3226 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3227 		struct amdgpu_ring *ring = adev->rings[i];
3228 
3229 		if (!ring || !ring->sched.thread)
3230 			continue;
3231 
3232 		/* only need recovery sched of the given job's ring
3233 		 * or all rings (in the case @job is NULL)
3234 		 * after above amdgpu_reset accomplished
3235 		 */
3236 		if ((!job || job->ring->idx == i) && !r)
3237 			drm_sched_job_recovery(&ring->sched);
3238 
3239 		kthread_unpark(ring->sched.thread);
3240 	}
3241 
3242 	if (!amdgpu_device_has_dc_support(adev)) {
3243 		drm_helper_resume_force_mode(adev->ddev);
3244 	}
3245 
3246 	ttm_bo_unlock_delayed_workqueue(&adev->mman.bdev, resched);
3247 
3248 	if (r) {
3249 		/* bad news, how to tell it to userspace ? */
3250 		dev_info(adev->dev, "GPU reset(%d) failed\n", atomic_read(&adev->gpu_reset_counter));
3251 		amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r);
3252 	} else {
3253 		dev_info(adev->dev, "GPU reset(%d) successed!\n",atomic_read(&adev->gpu_reset_counter));
3254 	}
3255 
3256 	amdgpu_vf_error_trans_all(adev);
3257 	adev->in_gpu_reset = 0;
3258 	mutex_unlock(&adev->lock_reset);
3259 	return r;
3260 }
3261 
3262 /**
3263  * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot
3264  *
3265  * @adev: amdgpu_device pointer
3266  *
3267  * Fetchs and stores in the driver the PCIE capabilities (gen speed
3268  * and lanes) of the slot the device is in. Handles APUs and
3269  * virtualized environments where PCIE config space may not be available.
3270  */
3271 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev)
3272 {
3273 	u32 mask;
3274 	int ret;
3275 
3276 	if (amdgpu_pcie_gen_cap)
3277 		adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
3278 
3279 	if (amdgpu_pcie_lane_cap)
3280 		adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
3281 
3282 	/* covers APUs as well */
3283 	if (pci_is_root_bus(adev->pdev->bus)) {
3284 		if (adev->pm.pcie_gen_mask == 0)
3285 			adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
3286 		if (adev->pm.pcie_mlw_mask == 0)
3287 			adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
3288 		return;
3289 	}
3290 
3291 	if (adev->pm.pcie_gen_mask == 0) {
3292 		ret = drm_pcie_get_speed_cap_mask(adev->ddev, &mask);
3293 		if (!ret) {
3294 			adev->pm.pcie_gen_mask = (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3295 						  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3296 						  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
3297 
3298 			if (mask & DRM_PCIE_SPEED_25)
3299 				adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
3300 			if (mask & DRM_PCIE_SPEED_50)
3301 				adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2;
3302 			if (mask & DRM_PCIE_SPEED_80)
3303 				adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3;
3304 		} else {
3305 			adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
3306 		}
3307 	}
3308 	if (adev->pm.pcie_mlw_mask == 0) {
3309 		ret = drm_pcie_get_max_link_width(adev->ddev, &mask);
3310 		if (!ret) {
3311 			switch (mask) {
3312 			case 32:
3313 				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
3314 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
3315 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3316 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3317 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3318 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3319 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3320 				break;
3321 			case 16:
3322 				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
3323 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3324 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3325 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3326 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3327 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3328 				break;
3329 			case 12:
3330 				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3331 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3332 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3333 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3334 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3335 				break;
3336 			case 8:
3337 				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3338 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3339 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3340 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3341 				break;
3342 			case 4:
3343 				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3344 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3345 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3346 				break;
3347 			case 2:
3348 				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3349 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3350 				break;
3351 			case 1:
3352 				adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;
3353 				break;
3354 			default:
3355 				break;
3356 			}
3357 		} else {
3358 			adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
3359 		}
3360 	}
3361 }
3362 
3363