xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c (revision 812dd0202379a242f764f75aa30abfbef4398ba4)
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/raven_gpu_info.bin");
63 
64 #define AMDGPU_RESUME_MS		2000
65 
66 static const char *amdgpu_asic_name[] = {
67 	"TAHITI",
68 	"PITCAIRN",
69 	"VERDE",
70 	"OLAND",
71 	"HAINAN",
72 	"BONAIRE",
73 	"KAVERI",
74 	"KABINI",
75 	"HAWAII",
76 	"MULLINS",
77 	"TOPAZ",
78 	"TONGA",
79 	"FIJI",
80 	"CARRIZO",
81 	"STONEY",
82 	"POLARIS10",
83 	"POLARIS11",
84 	"POLARIS12",
85 	"VEGA10",
86 	"RAVEN",
87 	"LAST",
88 };
89 
90 bool amdgpu_device_is_px(struct drm_device *dev)
91 {
92 	struct amdgpu_device *adev = dev->dev_private;
93 
94 	if (adev->flags & AMD_IS_PX)
95 		return true;
96 	return false;
97 }
98 
99 /*
100  * MMIO register access helper functions.
101  */
102 uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg,
103 			uint32_t acc_flags)
104 {
105 	uint32_t ret;
106 
107 	if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
108 		return amdgpu_virt_kiq_rreg(adev, reg);
109 
110 	if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
111 		ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
112 	else {
113 		unsigned long flags;
114 
115 		spin_lock_irqsave(&adev->mmio_idx_lock, flags);
116 		writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
117 		ret = readl(((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
118 		spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
119 	}
120 	trace_amdgpu_mm_rreg(adev->pdev->device, reg, ret);
121 	return ret;
122 }
123 
124 void amdgpu_mm_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v,
125 		    uint32_t acc_flags)
126 {
127 	trace_amdgpu_mm_wreg(adev->pdev->device, reg, v);
128 
129 	if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
130 		adev->last_mm_index = v;
131 	}
132 
133 	if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
134 		return amdgpu_virt_kiq_wreg(adev, reg, v);
135 
136 	if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
137 		writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
138 	else {
139 		unsigned long flags;
140 
141 		spin_lock_irqsave(&adev->mmio_idx_lock, flags);
142 		writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
143 		writel(v, ((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
144 		spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
145 	}
146 
147 	if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
148 		udelay(500);
149 	}
150 }
151 
152 u32 amdgpu_io_rreg(struct amdgpu_device *adev, u32 reg)
153 {
154 	if ((reg * 4) < adev->rio_mem_size)
155 		return ioread32(adev->rio_mem + (reg * 4));
156 	else {
157 		iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
158 		return ioread32(adev->rio_mem + (mmMM_DATA * 4));
159 	}
160 }
161 
162 void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32 v)
163 {
164 	if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
165 		adev->last_mm_index = v;
166 	}
167 
168 	if ((reg * 4) < adev->rio_mem_size)
169 		iowrite32(v, adev->rio_mem + (reg * 4));
170 	else {
171 		iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
172 		iowrite32(v, adev->rio_mem + (mmMM_DATA * 4));
173 	}
174 
175 	if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
176 		udelay(500);
177 	}
178 }
179 
180 /**
181  * amdgpu_mm_rdoorbell - read a doorbell dword
182  *
183  * @adev: amdgpu_device pointer
184  * @index: doorbell index
185  *
186  * Returns the value in the doorbell aperture at the
187  * requested doorbell index (CIK).
188  */
189 u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
190 {
191 	if (index < adev->doorbell.num_doorbells) {
192 		return readl(adev->doorbell.ptr + index);
193 	} else {
194 		DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
195 		return 0;
196 	}
197 }
198 
199 /**
200  * amdgpu_mm_wdoorbell - write a doorbell dword
201  *
202  * @adev: amdgpu_device pointer
203  * @index: doorbell index
204  * @v: value to write
205  *
206  * Writes @v to the doorbell aperture at the
207  * requested doorbell index (CIK).
208  */
209 void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v)
210 {
211 	if (index < adev->doorbell.num_doorbells) {
212 		writel(v, adev->doorbell.ptr + index);
213 	} else {
214 		DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
215 	}
216 }
217 
218 /**
219  * amdgpu_mm_rdoorbell64 - read a doorbell Qword
220  *
221  * @adev: amdgpu_device pointer
222  * @index: doorbell index
223  *
224  * Returns the value in the doorbell aperture at the
225  * requested doorbell index (VEGA10+).
226  */
227 u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index)
228 {
229 	if (index < adev->doorbell.num_doorbells) {
230 		return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index));
231 	} else {
232 		DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
233 		return 0;
234 	}
235 }
236 
237 /**
238  * amdgpu_mm_wdoorbell64 - write a doorbell Qword
239  *
240  * @adev: amdgpu_device pointer
241  * @index: doorbell index
242  * @v: value to write
243  *
244  * Writes @v to the doorbell aperture at the
245  * requested doorbell index (VEGA10+).
246  */
247 void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v)
248 {
249 	if (index < adev->doorbell.num_doorbells) {
250 		atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v);
251 	} else {
252 		DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
253 	}
254 }
255 
256 /**
257  * amdgpu_invalid_rreg - dummy reg read function
258  *
259  * @adev: amdgpu device pointer
260  * @reg: offset of register
261  *
262  * Dummy register read function.  Used for register blocks
263  * that certain asics don't have (all asics).
264  * Returns the value in the register.
265  */
266 static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
267 {
268 	DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
269 	BUG();
270 	return 0;
271 }
272 
273 /**
274  * amdgpu_invalid_wreg - dummy reg write function
275  *
276  * @adev: amdgpu device pointer
277  * @reg: offset of register
278  * @v: value to write to the register
279  *
280  * Dummy register read function.  Used for register blocks
281  * that certain asics don't have (all asics).
282  */
283 static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
284 {
285 	DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
286 		  reg, v);
287 	BUG();
288 }
289 
290 /**
291  * amdgpu_block_invalid_rreg - dummy reg read function
292  *
293  * @adev: amdgpu device pointer
294  * @block: offset of instance
295  * @reg: offset of register
296  *
297  * Dummy register read function.  Used for register blocks
298  * that certain asics don't have (all asics).
299  * Returns the value in the register.
300  */
301 static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
302 					  uint32_t block, uint32_t reg)
303 {
304 	DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
305 		  reg, block);
306 	BUG();
307 	return 0;
308 }
309 
310 /**
311  * amdgpu_block_invalid_wreg - dummy reg write function
312  *
313  * @adev: amdgpu device pointer
314  * @block: offset of instance
315  * @reg: offset of register
316  * @v: value to write to the register
317  *
318  * Dummy register read function.  Used for register blocks
319  * that certain asics don't have (all asics).
320  */
321 static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
322 				      uint32_t block,
323 				      uint32_t reg, uint32_t v)
324 {
325 	DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
326 		  reg, block, v);
327 	BUG();
328 }
329 
330 static int amdgpu_device_vram_scratch_init(struct amdgpu_device *adev)
331 {
332 	return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE,
333 				       PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
334 				       &adev->vram_scratch.robj,
335 				       &adev->vram_scratch.gpu_addr,
336 				       (void **)&adev->vram_scratch.ptr);
337 }
338 
339 static void amdgpu_device_vram_scratch_fini(struct amdgpu_device *adev)
340 {
341 	amdgpu_bo_free_kernel(&adev->vram_scratch.robj, NULL, NULL);
342 }
343 
344 /**
345  * amdgpu_device_program_register_sequence - program an array of registers.
346  *
347  * @adev: amdgpu_device pointer
348  * @registers: pointer to the register array
349  * @array_size: size of the register array
350  *
351  * Programs an array or registers with and and or masks.
352  * This is a helper for setting golden registers.
353  */
354 void amdgpu_device_program_register_sequence(struct amdgpu_device *adev,
355 					     const u32 *registers,
356 					     const u32 array_size)
357 {
358 	u32 tmp, reg, and_mask, or_mask;
359 	int i;
360 
361 	if (array_size % 3)
362 		return;
363 
364 	for (i = 0; i < array_size; i +=3) {
365 		reg = registers[i + 0];
366 		and_mask = registers[i + 1];
367 		or_mask = registers[i + 2];
368 
369 		if (and_mask == 0xffffffff) {
370 			tmp = or_mask;
371 		} else {
372 			tmp = RREG32(reg);
373 			tmp &= ~and_mask;
374 			tmp |= or_mask;
375 		}
376 		WREG32(reg, tmp);
377 	}
378 }
379 
380 void amdgpu_device_pci_config_reset(struct amdgpu_device *adev)
381 {
382 	pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
383 }
384 
385 /*
386  * GPU doorbell aperture helpers function.
387  */
388 /**
389  * amdgpu_device_doorbell_init - Init doorbell driver information.
390  *
391  * @adev: amdgpu_device pointer
392  *
393  * Init doorbell driver information (CIK)
394  * Returns 0 on success, error on failure.
395  */
396 static int amdgpu_device_doorbell_init(struct amdgpu_device *adev)
397 {
398 	/* No doorbell on SI hardware generation */
399 	if (adev->asic_type < CHIP_BONAIRE) {
400 		adev->doorbell.base = 0;
401 		adev->doorbell.size = 0;
402 		adev->doorbell.num_doorbells = 0;
403 		adev->doorbell.ptr = NULL;
404 		return 0;
405 	}
406 
407 	if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET)
408 		return -EINVAL;
409 
410 	/* doorbell bar mapping */
411 	adev->doorbell.base = pci_resource_start(adev->pdev, 2);
412 	adev->doorbell.size = pci_resource_len(adev->pdev, 2);
413 
414 	adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32),
415 					     AMDGPU_DOORBELL_MAX_ASSIGNMENT+1);
416 	if (adev->doorbell.num_doorbells == 0)
417 		return -EINVAL;
418 
419 	adev->doorbell.ptr = ioremap(adev->doorbell.base,
420 				     adev->doorbell.num_doorbells *
421 				     sizeof(u32));
422 	if (adev->doorbell.ptr == NULL)
423 		return -ENOMEM;
424 
425 	return 0;
426 }
427 
428 /**
429  * amdgpu_device_doorbell_fini - Tear down doorbell driver information.
430  *
431  * @adev: amdgpu_device pointer
432  *
433  * Tear down doorbell driver information (CIK)
434  */
435 static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev)
436 {
437 	iounmap(adev->doorbell.ptr);
438 	adev->doorbell.ptr = NULL;
439 }
440 
441 
442 
443 /*
444  * amdgpu_device_wb_*()
445  * Writeback is the method by which the GPU updates special pages in memory
446  * with the status of certain GPU events (fences, ring pointers,etc.).
447  */
448 
449 /**
450  * amdgpu_device_wb_fini - Disable Writeback and free memory
451  *
452  * @adev: amdgpu_device pointer
453  *
454  * Disables Writeback and frees the Writeback memory (all asics).
455  * Used at driver shutdown.
456  */
457 static void amdgpu_device_wb_fini(struct amdgpu_device *adev)
458 {
459 	if (adev->wb.wb_obj) {
460 		amdgpu_bo_free_kernel(&adev->wb.wb_obj,
461 				      &adev->wb.gpu_addr,
462 				      (void **)&adev->wb.wb);
463 		adev->wb.wb_obj = NULL;
464 	}
465 }
466 
467 /**
468  * amdgpu_device_wb_init- Init Writeback driver info and allocate memory
469  *
470  * @adev: amdgpu_device pointer
471  *
472  * Initializes writeback and allocates writeback memory (all asics).
473  * Used at driver startup.
474  * Returns 0 on success or an -error on failure.
475  */
476 static int amdgpu_device_wb_init(struct amdgpu_device *adev)
477 {
478 	int r;
479 
480 	if (adev->wb.wb_obj == NULL) {
481 		/* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */
482 		r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8,
483 					    PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
484 					    &adev->wb.wb_obj, &adev->wb.gpu_addr,
485 					    (void **)&adev->wb.wb);
486 		if (r) {
487 			dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
488 			return r;
489 		}
490 
491 		adev->wb.num_wb = AMDGPU_MAX_WB;
492 		memset(&adev->wb.used, 0, sizeof(adev->wb.used));
493 
494 		/* clear wb memory */
495 		memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8);
496 	}
497 
498 	return 0;
499 }
500 
501 /**
502  * amdgpu_device_wb_get - Allocate a wb entry
503  *
504  * @adev: amdgpu_device pointer
505  * @wb: wb index
506  *
507  * Allocate a wb slot for use by the driver (all asics).
508  * Returns 0 on success or -EINVAL on failure.
509  */
510 int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb)
511 {
512 	unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
513 
514 	if (offset < adev->wb.num_wb) {
515 		__set_bit(offset, adev->wb.used);
516 		*wb = offset << 3; /* convert to dw offset */
517 		return 0;
518 	} else {
519 		return -EINVAL;
520 	}
521 }
522 
523 /**
524  * amdgpu_device_wb_free - Free a wb entry
525  *
526  * @adev: amdgpu_device pointer
527  * @wb: wb index
528  *
529  * Free a wb slot allocated for use by the driver (all asics)
530  */
531 void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb)
532 {
533 	wb >>= 3;
534 	if (wb < adev->wb.num_wb)
535 		__clear_bit(wb, adev->wb.used);
536 }
537 
538 /**
539  * amdgpu_device_vram_location - try to find VRAM location
540  * @adev: amdgpu device structure holding all necessary informations
541  * @mc: memory controller structure holding memory informations
542  * @base: base address at which to put VRAM
543  *
544  * Function will try to place VRAM at base address provided
545  * as parameter.
546  */
547 void amdgpu_device_vram_location(struct amdgpu_device *adev,
548 				 struct amdgpu_mc *mc, u64 base)
549 {
550 	uint64_t limit = (uint64_t)amdgpu_vram_limit << 20;
551 
552 	mc->vram_start = base;
553 	mc->vram_end = mc->vram_start + mc->mc_vram_size - 1;
554 	if (limit && limit < mc->real_vram_size)
555 		mc->real_vram_size = limit;
556 	dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
557 			mc->mc_vram_size >> 20, mc->vram_start,
558 			mc->vram_end, mc->real_vram_size >> 20);
559 }
560 
561 /**
562  * amdgpu_device_gart_location - try to find GTT location
563  * @adev: amdgpu device structure holding all necessary informations
564  * @mc: memory controller structure holding memory informations
565  *
566  * Function will place try to place GTT before or after VRAM.
567  *
568  * If GTT size is bigger than space left then we ajust GTT size.
569  * Thus function will never fails.
570  *
571  * FIXME: when reducing GTT size align new size on power of 2.
572  */
573 void amdgpu_device_gart_location(struct amdgpu_device *adev,
574 				 struct amdgpu_mc *mc)
575 {
576 	u64 size_af, size_bf;
577 
578 	size_af = adev->mc.mc_mask - mc->vram_end;
579 	size_bf = mc->vram_start;
580 	if (size_bf > size_af) {
581 		if (mc->gart_size > size_bf) {
582 			dev_warn(adev->dev, "limiting GTT\n");
583 			mc->gart_size = size_bf;
584 		}
585 		mc->gart_start = 0;
586 	} else {
587 		if (mc->gart_size > size_af) {
588 			dev_warn(adev->dev, "limiting GTT\n");
589 			mc->gart_size = size_af;
590 		}
591 		/* VCE doesn't like it when BOs cross a 4GB segment, so align
592 		 * the GART base on a 4GB boundary as well.
593 		 */
594 		mc->gart_start = ALIGN(mc->vram_end + 1, 0x100000000ULL);
595 	}
596 	mc->gart_end = mc->gart_start + mc->gart_size - 1;
597 	dev_info(adev->dev, "GTT: %lluM 0x%016llX - 0x%016llX\n",
598 			mc->gart_size >> 20, mc->gart_start, mc->gart_end);
599 }
600 
601 /**
602  * amdgpu_device_resize_fb_bar - try to resize FB BAR
603  *
604  * @adev: amdgpu_device pointer
605  *
606  * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not
607  * to fail, but if any of the BARs is not accessible after the size we abort
608  * driver loading by returning -ENODEV.
609  */
610 int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev)
611 {
612 	u64 space_needed = roundup_pow_of_two(adev->mc.real_vram_size);
613 	u32 rbar_size = order_base_2(((space_needed >> 20) | 1)) - 1;
614 	struct pci_bus *root;
615 	struct resource *res;
616 	unsigned i;
617 	u16 cmd;
618 	int r;
619 
620 	/* Bypass for VF */
621 	if (amdgpu_sriov_vf(adev))
622 		return 0;
623 
624 	/* Check if the root BUS has 64bit memory resources */
625 	root = adev->pdev->bus;
626 	while (root->parent)
627 		root = root->parent;
628 
629 	pci_bus_for_each_resource(root, res, i) {
630 		if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) &&
631 		    res->start > 0x100000000ull)
632 			break;
633 	}
634 
635 	/* Trying to resize is pointless without a root hub window above 4GB */
636 	if (!res)
637 		return 0;
638 
639 	/* Disable memory decoding while we change the BAR addresses and size */
640 	pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd);
641 	pci_write_config_word(adev->pdev, PCI_COMMAND,
642 			      cmd & ~PCI_COMMAND_MEMORY);
643 
644 	/* Free the VRAM and doorbell BAR, we most likely need to move both. */
645 	amdgpu_device_doorbell_fini(adev);
646 	if (adev->asic_type >= CHIP_BONAIRE)
647 		pci_release_resource(adev->pdev, 2);
648 
649 	pci_release_resource(adev->pdev, 0);
650 
651 	r = pci_resize_resource(adev->pdev, 0, rbar_size);
652 	if (r == -ENOSPC)
653 		DRM_INFO("Not enough PCI address space for a large BAR.");
654 	else if (r && r != -ENOTSUPP)
655 		DRM_ERROR("Problem resizing BAR0 (%d).", r);
656 
657 	pci_assign_unassigned_bus_resources(adev->pdev->bus);
658 
659 	/* When the doorbell or fb BAR isn't available we have no chance of
660 	 * using the device.
661 	 */
662 	r = amdgpu_device_doorbell_init(adev);
663 	if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET))
664 		return -ENODEV;
665 
666 	pci_write_config_word(adev->pdev, PCI_COMMAND, cmd);
667 
668 	return 0;
669 }
670 
671 /*
672  * GPU helpers function.
673  */
674 /**
675  * amdgpu_device_need_post - check if the hw need post or not
676  *
677  * @adev: amdgpu_device pointer
678  *
679  * Check if the asic has been initialized (all asics) at driver startup
680  * or post is needed if  hw reset is performed.
681  * Returns true if need or false if not.
682  */
683 bool amdgpu_device_need_post(struct amdgpu_device *adev)
684 {
685 	uint32_t reg;
686 
687 	if (amdgpu_sriov_vf(adev))
688 		return false;
689 
690 	if (amdgpu_passthrough(adev)) {
691 		/* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
692 		 * some old smc fw still need driver do vPost otherwise gpu hang, while
693 		 * those smc fw version above 22.15 doesn't have this flaw, so we force
694 		 * vpost executed for smc version below 22.15
695 		 */
696 		if (adev->asic_type == CHIP_FIJI) {
697 			int err;
698 			uint32_t fw_ver;
699 			err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
700 			/* force vPost if error occured */
701 			if (err)
702 				return true;
703 
704 			fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
705 			if (fw_ver < 0x00160e00)
706 				return true;
707 		}
708 	}
709 
710 	if (adev->has_hw_reset) {
711 		adev->has_hw_reset = false;
712 		return true;
713 	}
714 
715 	/* bios scratch used on CIK+ */
716 	if (adev->asic_type >= CHIP_BONAIRE)
717 		return amdgpu_atombios_scratch_need_asic_init(adev);
718 
719 	/* check MEM_SIZE for older asics */
720 	reg = amdgpu_asic_get_config_memsize(adev);
721 
722 	if ((reg != 0) && (reg != 0xffffffff))
723 		return false;
724 
725 	return true;
726 }
727 
728 /* if we get transitioned to only one device, take VGA back */
729 /**
730  * amdgpu_device_vga_set_decode - enable/disable vga decode
731  *
732  * @cookie: amdgpu_device pointer
733  * @state: enable/disable vga decode
734  *
735  * Enable/disable vga decode (all asics).
736  * Returns VGA resource flags.
737  */
738 static unsigned int amdgpu_device_vga_set_decode(void *cookie, bool state)
739 {
740 	struct amdgpu_device *adev = cookie;
741 	amdgpu_asic_set_vga_state(adev, state);
742 	if (state)
743 		return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
744 		       VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
745 	else
746 		return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
747 }
748 
749 static void amdgpu_device_check_block_size(struct amdgpu_device *adev)
750 {
751 	/* defines number of bits in page table versus page directory,
752 	 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
753 	 * page table and the remaining bits are in the page directory */
754 	if (amdgpu_vm_block_size == -1)
755 		return;
756 
757 	if (amdgpu_vm_block_size < 9) {
758 		dev_warn(adev->dev, "VM page table size (%d) too small\n",
759 			 amdgpu_vm_block_size);
760 		amdgpu_vm_block_size = -1;
761 	}
762 }
763 
764 static void amdgpu_device_check_vm_size(struct amdgpu_device *adev)
765 {
766 	/* no need to check the default value */
767 	if (amdgpu_vm_size == -1)
768 		return;
769 
770 	if (amdgpu_vm_size < 1) {
771 		dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
772 			 amdgpu_vm_size);
773 		amdgpu_vm_size = -1;
774 	}
775 }
776 
777 /**
778  * amdgpu_device_check_arguments - validate module params
779  *
780  * @adev: amdgpu_device pointer
781  *
782  * Validates certain module parameters and updates
783  * the associated values used by the driver (all asics).
784  */
785 static void amdgpu_device_check_arguments(struct amdgpu_device *adev)
786 {
787 	if (amdgpu_sched_jobs < 4) {
788 		dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
789 			 amdgpu_sched_jobs);
790 		amdgpu_sched_jobs = 4;
791 	} else if (!is_power_of_2(amdgpu_sched_jobs)){
792 		dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
793 			 amdgpu_sched_jobs);
794 		amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
795 	}
796 
797 	if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) {
798 		/* gart size must be greater or equal to 32M */
799 		dev_warn(adev->dev, "gart size (%d) too small\n",
800 			 amdgpu_gart_size);
801 		amdgpu_gart_size = -1;
802 	}
803 
804 	if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) {
805 		/* gtt size must be greater or equal to 32M */
806 		dev_warn(adev->dev, "gtt size (%d) too small\n",
807 				 amdgpu_gtt_size);
808 		amdgpu_gtt_size = -1;
809 	}
810 
811 	/* valid range is between 4 and 9 inclusive */
812 	if (amdgpu_vm_fragment_size != -1 &&
813 	    (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) {
814 		dev_warn(adev->dev, "valid range is between 4 and 9\n");
815 		amdgpu_vm_fragment_size = -1;
816 	}
817 
818 	amdgpu_device_check_vm_size(adev);
819 
820 	amdgpu_device_check_block_size(adev);
821 
822 	if (amdgpu_vram_page_split != -1 && (amdgpu_vram_page_split < 16 ||
823 	    !is_power_of_2(amdgpu_vram_page_split))) {
824 		dev_warn(adev->dev, "invalid VRAM page split (%d)\n",
825 			 amdgpu_vram_page_split);
826 		amdgpu_vram_page_split = 1024;
827 	}
828 
829 	if (amdgpu_lockup_timeout == 0) {
830 		dev_warn(adev->dev, "lockup_timeout msut be > 0, adjusting to 10000\n");
831 		amdgpu_lockup_timeout = 10000;
832 	}
833 }
834 
835 /**
836  * amdgpu_switcheroo_set_state - set switcheroo state
837  *
838  * @pdev: pci dev pointer
839  * @state: vga_switcheroo state
840  *
841  * Callback for the switcheroo driver.  Suspends or resumes the
842  * the asics before or after it is powered up using ACPI methods.
843  */
844 static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, enum vga_switcheroo_state state)
845 {
846 	struct drm_device *dev = pci_get_drvdata(pdev);
847 
848 	if (amdgpu_device_is_px(dev) && state == VGA_SWITCHEROO_OFF)
849 		return;
850 
851 	if (state == VGA_SWITCHEROO_ON) {
852 		pr_info("amdgpu: switched on\n");
853 		/* don't suspend or resume card normally */
854 		dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
855 
856 		amdgpu_device_resume(dev, true, true);
857 
858 		dev->switch_power_state = DRM_SWITCH_POWER_ON;
859 		drm_kms_helper_poll_enable(dev);
860 	} else {
861 		pr_info("amdgpu: switched off\n");
862 		drm_kms_helper_poll_disable(dev);
863 		dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
864 		amdgpu_device_suspend(dev, true, true);
865 		dev->switch_power_state = DRM_SWITCH_POWER_OFF;
866 	}
867 }
868 
869 /**
870  * amdgpu_switcheroo_can_switch - see if switcheroo state can change
871  *
872  * @pdev: pci dev pointer
873  *
874  * Callback for the switcheroo driver.  Check of the switcheroo
875  * state can be changed.
876  * Returns true if the state can be changed, false if not.
877  */
878 static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
879 {
880 	struct drm_device *dev = pci_get_drvdata(pdev);
881 
882 	/*
883 	* FIXME: open_count is protected by drm_global_mutex but that would lead to
884 	* locking inversion with the driver load path. And the access here is
885 	* completely racy anyway. So don't bother with locking for now.
886 	*/
887 	return dev->open_count == 0;
888 }
889 
890 static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
891 	.set_gpu_state = amdgpu_switcheroo_set_state,
892 	.reprobe = NULL,
893 	.can_switch = amdgpu_switcheroo_can_switch,
894 };
895 
896 int amdgpu_device_ip_set_clockgating_state(struct amdgpu_device *adev,
897 					   enum amd_ip_block_type block_type,
898 					   enum amd_clockgating_state state)
899 {
900 	int i, r = 0;
901 
902 	for (i = 0; i < adev->num_ip_blocks; i++) {
903 		if (!adev->ip_blocks[i].status.valid)
904 			continue;
905 		if (adev->ip_blocks[i].version->type != block_type)
906 			continue;
907 		if (!adev->ip_blocks[i].version->funcs->set_clockgating_state)
908 			continue;
909 		r = adev->ip_blocks[i].version->funcs->set_clockgating_state(
910 			(void *)adev, state);
911 		if (r)
912 			DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",
913 				  adev->ip_blocks[i].version->funcs->name, r);
914 	}
915 	return r;
916 }
917 
918 int amdgpu_device_ip_set_powergating_state(struct amdgpu_device *adev,
919 					   enum amd_ip_block_type block_type,
920 					   enum amd_powergating_state state)
921 {
922 	int i, r = 0;
923 
924 	for (i = 0; i < adev->num_ip_blocks; i++) {
925 		if (!adev->ip_blocks[i].status.valid)
926 			continue;
927 		if (adev->ip_blocks[i].version->type != block_type)
928 			continue;
929 		if (!adev->ip_blocks[i].version->funcs->set_powergating_state)
930 			continue;
931 		r = adev->ip_blocks[i].version->funcs->set_powergating_state(
932 			(void *)adev, state);
933 		if (r)
934 			DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",
935 				  adev->ip_blocks[i].version->funcs->name, r);
936 	}
937 	return r;
938 }
939 
940 void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev,
941 					    u32 *flags)
942 {
943 	int i;
944 
945 	for (i = 0; i < adev->num_ip_blocks; i++) {
946 		if (!adev->ip_blocks[i].status.valid)
947 			continue;
948 		if (adev->ip_blocks[i].version->funcs->get_clockgating_state)
949 			adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags);
950 	}
951 }
952 
953 int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev,
954 				   enum amd_ip_block_type block_type)
955 {
956 	int i, r;
957 
958 	for (i = 0; i < adev->num_ip_blocks; i++) {
959 		if (!adev->ip_blocks[i].status.valid)
960 			continue;
961 		if (adev->ip_blocks[i].version->type == block_type) {
962 			r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
963 			if (r)
964 				return r;
965 			break;
966 		}
967 	}
968 	return 0;
969 
970 }
971 
972 bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev,
973 			      enum amd_ip_block_type block_type)
974 {
975 	int i;
976 
977 	for (i = 0; i < adev->num_ip_blocks; i++) {
978 		if (!adev->ip_blocks[i].status.valid)
979 			continue;
980 		if (adev->ip_blocks[i].version->type == block_type)
981 			return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
982 	}
983 	return true;
984 
985 }
986 
987 struct amdgpu_ip_block *
988 amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,
989 			      enum amd_ip_block_type type)
990 {
991 	int i;
992 
993 	for (i = 0; i < adev->num_ip_blocks; i++)
994 		if (adev->ip_blocks[i].version->type == type)
995 			return &adev->ip_blocks[i];
996 
997 	return NULL;
998 }
999 
1000 /**
1001  * amdgpu_device_ip_block_version_cmp
1002  *
1003  * @adev: amdgpu_device pointer
1004  * @type: enum amd_ip_block_type
1005  * @major: major version
1006  * @minor: minor version
1007  *
1008  * return 0 if equal or greater
1009  * return 1 if smaller or the ip_block doesn't exist
1010  */
1011 int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev,
1012 				       enum amd_ip_block_type type,
1013 				       u32 major, u32 minor)
1014 {
1015 	struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type);
1016 
1017 	if (ip_block && ((ip_block->version->major > major) ||
1018 			((ip_block->version->major == major) &&
1019 			(ip_block->version->minor >= minor))))
1020 		return 0;
1021 
1022 	return 1;
1023 }
1024 
1025 /**
1026  * amdgpu_device_ip_block_add
1027  *
1028  * @adev: amdgpu_device pointer
1029  * @ip_block_version: pointer to the IP to add
1030  *
1031  * Adds the IP block driver information to the collection of IPs
1032  * on the asic.
1033  */
1034 int amdgpu_device_ip_block_add(struct amdgpu_device *adev,
1035 			       const struct amdgpu_ip_block_version *ip_block_version)
1036 {
1037 	if (!ip_block_version)
1038 		return -EINVAL;
1039 
1040 	DRM_DEBUG("add ip block number %d <%s>\n", adev->num_ip_blocks,
1041 		  ip_block_version->funcs->name);
1042 
1043 	adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
1044 
1045 	return 0;
1046 }
1047 
1048 static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
1049 {
1050 	adev->enable_virtual_display = false;
1051 
1052 	if (amdgpu_virtual_display) {
1053 		struct drm_device *ddev = adev->ddev;
1054 		const char *pci_address_name = pci_name(ddev->pdev);
1055 		char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
1056 
1057 		pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
1058 		pciaddstr_tmp = pciaddstr;
1059 		while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
1060 			pciaddname = strsep(&pciaddname_tmp, ",");
1061 			if (!strcmp("all", pciaddname)
1062 			    || !strcmp(pci_address_name, pciaddname)) {
1063 				long num_crtc;
1064 				int res = -1;
1065 
1066 				adev->enable_virtual_display = true;
1067 
1068 				if (pciaddname_tmp)
1069 					res = kstrtol(pciaddname_tmp, 10,
1070 						      &num_crtc);
1071 
1072 				if (!res) {
1073 					if (num_crtc < 1)
1074 						num_crtc = 1;
1075 					if (num_crtc > 6)
1076 						num_crtc = 6;
1077 					adev->mode_info.num_crtc = num_crtc;
1078 				} else {
1079 					adev->mode_info.num_crtc = 1;
1080 				}
1081 				break;
1082 			}
1083 		}
1084 
1085 		DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
1086 			 amdgpu_virtual_display, pci_address_name,
1087 			 adev->enable_virtual_display, adev->mode_info.num_crtc);
1088 
1089 		kfree(pciaddstr);
1090 	}
1091 }
1092 
1093 static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev)
1094 {
1095 	const char *chip_name;
1096 	char fw_name[30];
1097 	int err;
1098 	const struct gpu_info_firmware_header_v1_0 *hdr;
1099 
1100 	adev->firmware.gpu_info_fw = NULL;
1101 
1102 	switch (adev->asic_type) {
1103 	case CHIP_TOPAZ:
1104 	case CHIP_TONGA:
1105 	case CHIP_FIJI:
1106 	case CHIP_POLARIS11:
1107 	case CHIP_POLARIS10:
1108 	case CHIP_POLARIS12:
1109 	case CHIP_CARRIZO:
1110 	case CHIP_STONEY:
1111 #ifdef CONFIG_DRM_AMDGPU_SI
1112 	case CHIP_VERDE:
1113 	case CHIP_TAHITI:
1114 	case CHIP_PITCAIRN:
1115 	case CHIP_OLAND:
1116 	case CHIP_HAINAN:
1117 #endif
1118 #ifdef CONFIG_DRM_AMDGPU_CIK
1119 	case CHIP_BONAIRE:
1120 	case CHIP_HAWAII:
1121 	case CHIP_KAVERI:
1122 	case CHIP_KABINI:
1123 	case CHIP_MULLINS:
1124 #endif
1125 	default:
1126 		return 0;
1127 	case CHIP_VEGA10:
1128 		chip_name = "vega10";
1129 		break;
1130 	case CHIP_RAVEN:
1131 		chip_name = "raven";
1132 		break;
1133 	}
1134 
1135 	snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name);
1136 	err = request_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev);
1137 	if (err) {
1138 		dev_err(adev->dev,
1139 			"Failed to load gpu_info firmware \"%s\"\n",
1140 			fw_name);
1141 		goto out;
1142 	}
1143 	err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw);
1144 	if (err) {
1145 		dev_err(adev->dev,
1146 			"Failed to validate gpu_info firmware \"%s\"\n",
1147 			fw_name);
1148 		goto out;
1149 	}
1150 
1151 	hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data;
1152 	amdgpu_ucode_print_gpu_info_hdr(&hdr->header);
1153 
1154 	switch (hdr->version_major) {
1155 	case 1:
1156 	{
1157 		const struct gpu_info_firmware_v1_0 *gpu_info_fw =
1158 			(const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data +
1159 								le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1160 
1161 		adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se);
1162 		adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh);
1163 		adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se);
1164 		adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se);
1165 		adev->gfx.config.max_texture_channel_caches =
1166 			le32_to_cpu(gpu_info_fw->gc_num_tccs);
1167 		adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs);
1168 		adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds);
1169 		adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth);
1170 		adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth);
1171 		adev->gfx.config.double_offchip_lds_buf =
1172 			le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer);
1173 		adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size);
1174 		adev->gfx.cu_info.max_waves_per_simd =
1175 			le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd);
1176 		adev->gfx.cu_info.max_scratch_slots_per_cu =
1177 			le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu);
1178 		adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size);
1179 		break;
1180 	}
1181 	default:
1182 		dev_err(adev->dev,
1183 			"Unsupported gpu_info table %d\n", hdr->header.ucode_version);
1184 		err = -EINVAL;
1185 		goto out;
1186 	}
1187 out:
1188 	return err;
1189 }
1190 
1191 static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
1192 {
1193 	int i, r;
1194 
1195 	amdgpu_device_enable_virtual_display(adev);
1196 
1197 	switch (adev->asic_type) {
1198 	case CHIP_TOPAZ:
1199 	case CHIP_TONGA:
1200 	case CHIP_FIJI:
1201 	case CHIP_POLARIS11:
1202 	case CHIP_POLARIS10:
1203 	case CHIP_POLARIS12:
1204 	case CHIP_CARRIZO:
1205 	case CHIP_STONEY:
1206 		if (adev->asic_type == CHIP_CARRIZO || adev->asic_type == CHIP_STONEY)
1207 			adev->family = AMDGPU_FAMILY_CZ;
1208 		else
1209 			adev->family = AMDGPU_FAMILY_VI;
1210 
1211 		r = vi_set_ip_blocks(adev);
1212 		if (r)
1213 			return r;
1214 		break;
1215 #ifdef CONFIG_DRM_AMDGPU_SI
1216 	case CHIP_VERDE:
1217 	case CHIP_TAHITI:
1218 	case CHIP_PITCAIRN:
1219 	case CHIP_OLAND:
1220 	case CHIP_HAINAN:
1221 		adev->family = AMDGPU_FAMILY_SI;
1222 		r = si_set_ip_blocks(adev);
1223 		if (r)
1224 			return r;
1225 		break;
1226 #endif
1227 #ifdef CONFIG_DRM_AMDGPU_CIK
1228 	case CHIP_BONAIRE:
1229 	case CHIP_HAWAII:
1230 	case CHIP_KAVERI:
1231 	case CHIP_KABINI:
1232 	case CHIP_MULLINS:
1233 		if ((adev->asic_type == CHIP_BONAIRE) || (adev->asic_type == CHIP_HAWAII))
1234 			adev->family = AMDGPU_FAMILY_CI;
1235 		else
1236 			adev->family = AMDGPU_FAMILY_KV;
1237 
1238 		r = cik_set_ip_blocks(adev);
1239 		if (r)
1240 			return r;
1241 		break;
1242 #endif
1243 	case  CHIP_VEGA10:
1244 	case  CHIP_RAVEN:
1245 		if (adev->asic_type == CHIP_RAVEN)
1246 			adev->family = AMDGPU_FAMILY_RV;
1247 		else
1248 			adev->family = AMDGPU_FAMILY_AI;
1249 
1250 		r = soc15_set_ip_blocks(adev);
1251 		if (r)
1252 			return r;
1253 		break;
1254 	default:
1255 		/* FIXME: not supported yet */
1256 		return -EINVAL;
1257 	}
1258 
1259 	r = amdgpu_device_parse_gpu_info_fw(adev);
1260 	if (r)
1261 		return r;
1262 
1263 	amdgpu_amdkfd_device_probe(adev);
1264 
1265 	if (amdgpu_sriov_vf(adev)) {
1266 		r = amdgpu_virt_request_full_gpu(adev, true);
1267 		if (r)
1268 			return -EAGAIN;
1269 	}
1270 
1271 	for (i = 0; i < adev->num_ip_blocks; i++) {
1272 		if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
1273 			DRM_ERROR("disabled ip block: %d <%s>\n",
1274 				  i, adev->ip_blocks[i].version->funcs->name);
1275 			adev->ip_blocks[i].status.valid = false;
1276 		} else {
1277 			if (adev->ip_blocks[i].version->funcs->early_init) {
1278 				r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
1279 				if (r == -ENOENT) {
1280 					adev->ip_blocks[i].status.valid = false;
1281 				} else if (r) {
1282 					DRM_ERROR("early_init of IP block <%s> failed %d\n",
1283 						  adev->ip_blocks[i].version->funcs->name, r);
1284 					return r;
1285 				} else {
1286 					adev->ip_blocks[i].status.valid = true;
1287 				}
1288 			} else {
1289 				adev->ip_blocks[i].status.valid = true;
1290 			}
1291 		}
1292 	}
1293 
1294 	adev->cg_flags &= amdgpu_cg_mask;
1295 	adev->pg_flags &= amdgpu_pg_mask;
1296 
1297 	return 0;
1298 }
1299 
1300 static int amdgpu_device_ip_init(struct amdgpu_device *adev)
1301 {
1302 	int i, r;
1303 
1304 	for (i = 0; i < adev->num_ip_blocks; i++) {
1305 		if (!adev->ip_blocks[i].status.valid)
1306 			continue;
1307 		r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
1308 		if (r) {
1309 			DRM_ERROR("sw_init of IP block <%s> failed %d\n",
1310 				  adev->ip_blocks[i].version->funcs->name, r);
1311 			return r;
1312 		}
1313 		adev->ip_blocks[i].status.sw = true;
1314 		/* need to do gmc hw init early so we can allocate gpu mem */
1315 		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1316 			r = amdgpu_device_vram_scratch_init(adev);
1317 			if (r) {
1318 				DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r);
1319 				return r;
1320 			}
1321 			r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1322 			if (r) {
1323 				DRM_ERROR("hw_init %d failed %d\n", i, r);
1324 				return r;
1325 			}
1326 			r = amdgpu_device_wb_init(adev);
1327 			if (r) {
1328 				DRM_ERROR("amdgpu_device_wb_init failed %d\n", r);
1329 				return r;
1330 			}
1331 			adev->ip_blocks[i].status.hw = true;
1332 
1333 			/* right after GMC hw init, we create CSA */
1334 			if (amdgpu_sriov_vf(adev)) {
1335 				r = amdgpu_allocate_static_csa(adev);
1336 				if (r) {
1337 					DRM_ERROR("allocate CSA failed %d\n", r);
1338 					return r;
1339 				}
1340 			}
1341 		}
1342 	}
1343 
1344 	for (i = 0; i < adev->num_ip_blocks; i++) {
1345 		if (!adev->ip_blocks[i].status.sw)
1346 			continue;
1347 		/* gmc hw init is done early */
1348 		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC)
1349 			continue;
1350 		r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1351 		if (r) {
1352 			DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1353 				  adev->ip_blocks[i].version->funcs->name, r);
1354 			return r;
1355 		}
1356 		adev->ip_blocks[i].status.hw = true;
1357 	}
1358 
1359 	amdgpu_amdkfd_device_init(adev);
1360 
1361 	if (amdgpu_sriov_vf(adev))
1362 		amdgpu_virt_release_full_gpu(adev, true);
1363 
1364 	return 0;
1365 }
1366 
1367 static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev)
1368 {
1369 	memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
1370 }
1371 
1372 static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev)
1373 {
1374 	return !!memcmp(adev->gart.ptr, adev->reset_magic,
1375 			AMDGPU_RESET_MAGIC_NUM);
1376 }
1377 
1378 static int amdgpu_device_ip_late_set_cg_state(struct amdgpu_device *adev)
1379 {
1380 	int i = 0, r;
1381 
1382 	for (i = 0; i < adev->num_ip_blocks; i++) {
1383 		if (!adev->ip_blocks[i].status.valid)
1384 			continue;
1385 		/* skip CG for VCE/UVD, it's handled specially */
1386 		if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1387 		    adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE) {
1388 			/* enable clockgating to save power */
1389 			r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1390 										     AMD_CG_STATE_GATE);
1391 			if (r) {
1392 				DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
1393 					  adev->ip_blocks[i].version->funcs->name, r);
1394 				return r;
1395 			}
1396 		}
1397 	}
1398 	return 0;
1399 }
1400 
1401 static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
1402 {
1403 	int i = 0, r;
1404 
1405 	for (i = 0; i < adev->num_ip_blocks; i++) {
1406 		if (!adev->ip_blocks[i].status.valid)
1407 			continue;
1408 		if (adev->ip_blocks[i].version->funcs->late_init) {
1409 			r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
1410 			if (r) {
1411 				DRM_ERROR("late_init of IP block <%s> failed %d\n",
1412 					  adev->ip_blocks[i].version->funcs->name, r);
1413 				return r;
1414 			}
1415 			adev->ip_blocks[i].status.late_initialized = true;
1416 		}
1417 	}
1418 
1419 	mod_delayed_work(system_wq, &adev->late_init_work,
1420 			msecs_to_jiffies(AMDGPU_RESUME_MS));
1421 
1422 	amdgpu_device_fill_reset_magic(adev);
1423 
1424 	return 0;
1425 }
1426 
1427 static int amdgpu_device_ip_fini(struct amdgpu_device *adev)
1428 {
1429 	int i, r;
1430 
1431 	amdgpu_amdkfd_device_fini(adev);
1432 	/* need to disable SMC first */
1433 	for (i = 0; i < adev->num_ip_blocks; i++) {
1434 		if (!adev->ip_blocks[i].status.hw)
1435 			continue;
1436 		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
1437 			/* ungate blocks before hw fini so that we can shutdown the blocks safely */
1438 			r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1439 										     AMD_CG_STATE_UNGATE);
1440 			if (r) {
1441 				DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1442 					  adev->ip_blocks[i].version->funcs->name, r);
1443 				return r;
1444 			}
1445 			r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
1446 			/* XXX handle errors */
1447 			if (r) {
1448 				DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
1449 					  adev->ip_blocks[i].version->funcs->name, r);
1450 			}
1451 			adev->ip_blocks[i].status.hw = false;
1452 			break;
1453 		}
1454 	}
1455 
1456 	for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1457 		if (!adev->ip_blocks[i].status.hw)
1458 			continue;
1459 
1460 		if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1461 			adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE) {
1462 			/* ungate blocks before hw fini so that we can shutdown the blocks safely */
1463 			r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1464 										     AMD_CG_STATE_UNGATE);
1465 			if (r) {
1466 				DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1467 					  adev->ip_blocks[i].version->funcs->name, r);
1468 				return r;
1469 			}
1470 		}
1471 
1472 		r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
1473 		/* XXX handle errors */
1474 		if (r) {
1475 			DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
1476 				  adev->ip_blocks[i].version->funcs->name, r);
1477 		}
1478 
1479 		adev->ip_blocks[i].status.hw = false;
1480 	}
1481 
1482 	for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1483 		if (!adev->ip_blocks[i].status.sw)
1484 			continue;
1485 
1486 		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1487 			amdgpu_free_static_csa(adev);
1488 			amdgpu_device_wb_fini(adev);
1489 			amdgpu_device_vram_scratch_fini(adev);
1490 		}
1491 
1492 		r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
1493 		/* XXX handle errors */
1494 		if (r) {
1495 			DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
1496 				  adev->ip_blocks[i].version->funcs->name, r);
1497 		}
1498 		adev->ip_blocks[i].status.sw = false;
1499 		adev->ip_blocks[i].status.valid = false;
1500 	}
1501 
1502 	for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1503 		if (!adev->ip_blocks[i].status.late_initialized)
1504 			continue;
1505 		if (adev->ip_blocks[i].version->funcs->late_fini)
1506 			adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
1507 		adev->ip_blocks[i].status.late_initialized = false;
1508 	}
1509 
1510 	if (amdgpu_sriov_vf(adev))
1511 		if (amdgpu_virt_release_full_gpu(adev, false))
1512 			DRM_ERROR("failed to release exclusive mode on fini\n");
1513 
1514 	return 0;
1515 }
1516 
1517 static void amdgpu_device_ip_late_init_func_handler(struct work_struct *work)
1518 {
1519 	struct amdgpu_device *adev =
1520 		container_of(work, struct amdgpu_device, late_init_work.work);
1521 	amdgpu_device_ip_late_set_cg_state(adev);
1522 }
1523 
1524 int amdgpu_device_ip_suspend(struct amdgpu_device *adev)
1525 {
1526 	int i, r;
1527 
1528 	if (amdgpu_sriov_vf(adev))
1529 		amdgpu_virt_request_full_gpu(adev, false);
1530 
1531 	/* ungate SMC block first */
1532 	r = amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_SMC,
1533 						   AMD_CG_STATE_UNGATE);
1534 	if (r) {
1535 		DRM_ERROR("set_clockgating_state(ungate) SMC failed %d\n", r);
1536 	}
1537 
1538 	for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1539 		if (!adev->ip_blocks[i].status.valid)
1540 			continue;
1541 		/* ungate blocks so that suspend can properly shut them down */
1542 		if (i != AMD_IP_BLOCK_TYPE_SMC) {
1543 			r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1544 										     AMD_CG_STATE_UNGATE);
1545 			if (r) {
1546 				DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1547 					  adev->ip_blocks[i].version->funcs->name, r);
1548 			}
1549 		}
1550 		/* XXX handle errors */
1551 		r = adev->ip_blocks[i].version->funcs->suspend(adev);
1552 		/* XXX handle errors */
1553 		if (r) {
1554 			DRM_ERROR("suspend of IP block <%s> failed %d\n",
1555 				  adev->ip_blocks[i].version->funcs->name, r);
1556 		}
1557 	}
1558 
1559 	if (amdgpu_sriov_vf(adev))
1560 		amdgpu_virt_release_full_gpu(adev, false);
1561 
1562 	return 0;
1563 }
1564 
1565 static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
1566 {
1567 	int i, r;
1568 
1569 	static enum amd_ip_block_type ip_order[] = {
1570 		AMD_IP_BLOCK_TYPE_GMC,
1571 		AMD_IP_BLOCK_TYPE_COMMON,
1572 		AMD_IP_BLOCK_TYPE_IH,
1573 	};
1574 
1575 	for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
1576 		int j;
1577 		struct amdgpu_ip_block *block;
1578 
1579 		for (j = 0; j < adev->num_ip_blocks; j++) {
1580 			block = &adev->ip_blocks[j];
1581 
1582 			if (block->version->type != ip_order[i] ||
1583 				!block->status.valid)
1584 				continue;
1585 
1586 			r = block->version->funcs->hw_init(adev);
1587 			DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"successed");
1588 		}
1589 	}
1590 
1591 	return 0;
1592 }
1593 
1594 static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
1595 {
1596 	int i, r;
1597 
1598 	static enum amd_ip_block_type ip_order[] = {
1599 		AMD_IP_BLOCK_TYPE_SMC,
1600 		AMD_IP_BLOCK_TYPE_PSP,
1601 		AMD_IP_BLOCK_TYPE_DCE,
1602 		AMD_IP_BLOCK_TYPE_GFX,
1603 		AMD_IP_BLOCK_TYPE_SDMA,
1604 		AMD_IP_BLOCK_TYPE_UVD,
1605 		AMD_IP_BLOCK_TYPE_VCE
1606 	};
1607 
1608 	for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
1609 		int j;
1610 		struct amdgpu_ip_block *block;
1611 
1612 		for (j = 0; j < adev->num_ip_blocks; j++) {
1613 			block = &adev->ip_blocks[j];
1614 
1615 			if (block->version->type != ip_order[i] ||
1616 				!block->status.valid)
1617 				continue;
1618 
1619 			r = block->version->funcs->hw_init(adev);
1620 			DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"successed");
1621 		}
1622 	}
1623 
1624 	return 0;
1625 }
1626 
1627 static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)
1628 {
1629 	int i, r;
1630 
1631 	for (i = 0; i < adev->num_ip_blocks; i++) {
1632 		if (!adev->ip_blocks[i].status.valid)
1633 			continue;
1634 		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
1635 				adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
1636 				adev->ip_blocks[i].version->type ==
1637 				AMD_IP_BLOCK_TYPE_IH) {
1638 			r = adev->ip_blocks[i].version->funcs->resume(adev);
1639 			if (r) {
1640 				DRM_ERROR("resume of IP block <%s> failed %d\n",
1641 					  adev->ip_blocks[i].version->funcs->name, r);
1642 				return r;
1643 			}
1644 		}
1645 	}
1646 
1647 	return 0;
1648 }
1649 
1650 static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)
1651 {
1652 	int i, r;
1653 
1654 	for (i = 0; i < adev->num_ip_blocks; i++) {
1655 		if (!adev->ip_blocks[i].status.valid)
1656 			continue;
1657 		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
1658 				adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
1659 				adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH )
1660 			continue;
1661 		r = adev->ip_blocks[i].version->funcs->resume(adev);
1662 		if (r) {
1663 			DRM_ERROR("resume of IP block <%s> failed %d\n",
1664 				  adev->ip_blocks[i].version->funcs->name, r);
1665 			return r;
1666 		}
1667 	}
1668 
1669 	return 0;
1670 }
1671 
1672 static int amdgpu_device_ip_resume(struct amdgpu_device *adev)
1673 {
1674 	int r;
1675 
1676 	r = amdgpu_device_ip_resume_phase1(adev);
1677 	if (r)
1678 		return r;
1679 	r = amdgpu_device_ip_resume_phase2(adev);
1680 
1681 	return r;
1682 }
1683 
1684 static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
1685 {
1686 	if (amdgpu_sriov_vf(adev)) {
1687 		if (adev->is_atom_fw) {
1688 			if (amdgpu_atomfirmware_gpu_supports_virtualization(adev))
1689 				adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
1690 		} else {
1691 			if (amdgpu_atombios_has_gpu_virtualization_table(adev))
1692 				adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
1693 		}
1694 
1695 		if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
1696 			amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
1697 	}
1698 }
1699 
1700 bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
1701 {
1702 	switch (asic_type) {
1703 #if defined(CONFIG_DRM_AMD_DC)
1704 	case CHIP_BONAIRE:
1705 	case CHIP_HAWAII:
1706 	case CHIP_KAVERI:
1707 	case CHIP_CARRIZO:
1708 	case CHIP_STONEY:
1709 	case CHIP_POLARIS11:
1710 	case CHIP_POLARIS10:
1711 	case CHIP_POLARIS12:
1712 	case CHIP_TONGA:
1713 	case CHIP_FIJI:
1714 #if defined(CONFIG_DRM_AMD_DC_PRE_VEGA)
1715 		return amdgpu_dc != 0;
1716 #endif
1717 	case CHIP_KABINI:
1718 	case CHIP_MULLINS:
1719 		return amdgpu_dc > 0;
1720 	case CHIP_VEGA10:
1721 #if defined(CONFIG_DRM_AMD_DC_DCN1_0)
1722 	case CHIP_RAVEN:
1723 #endif
1724 		return amdgpu_dc != 0;
1725 #endif
1726 	default:
1727 		return false;
1728 	}
1729 }
1730 
1731 /**
1732  * amdgpu_device_has_dc_support - check if dc is supported
1733  *
1734  * @adev: amdgpu_device_pointer
1735  *
1736  * Returns true for supported, false for not supported
1737  */
1738 bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
1739 {
1740 	if (amdgpu_sriov_vf(adev))
1741 		return false;
1742 
1743 	return amdgpu_device_asic_has_dc_support(adev->asic_type);
1744 }
1745 
1746 /**
1747  * amdgpu_device_init - initialize the driver
1748  *
1749  * @adev: amdgpu_device pointer
1750  * @pdev: drm dev pointer
1751  * @pdev: pci dev pointer
1752  * @flags: driver flags
1753  *
1754  * Initializes the driver info and hw (all asics).
1755  * Returns 0 for success or an error on failure.
1756  * Called at driver startup.
1757  */
1758 int amdgpu_device_init(struct amdgpu_device *adev,
1759 		       struct drm_device *ddev,
1760 		       struct pci_dev *pdev,
1761 		       uint32_t flags)
1762 {
1763 	int r, i;
1764 	bool runtime = false;
1765 	u32 max_MBps;
1766 
1767 	adev->shutdown = false;
1768 	adev->dev = &pdev->dev;
1769 	adev->ddev = ddev;
1770 	adev->pdev = pdev;
1771 	adev->flags = flags;
1772 	adev->asic_type = flags & AMD_ASIC_MASK;
1773 	adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
1774 	adev->mc.gart_size = 512 * 1024 * 1024;
1775 	adev->accel_working = false;
1776 	adev->num_rings = 0;
1777 	adev->mman.buffer_funcs = NULL;
1778 	adev->mman.buffer_funcs_ring = NULL;
1779 	adev->vm_manager.vm_pte_funcs = NULL;
1780 	adev->vm_manager.vm_pte_num_rings = 0;
1781 	adev->gart.gart_funcs = NULL;
1782 	adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
1783 	bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
1784 
1785 	adev->smc_rreg = &amdgpu_invalid_rreg;
1786 	adev->smc_wreg = &amdgpu_invalid_wreg;
1787 	adev->pcie_rreg = &amdgpu_invalid_rreg;
1788 	adev->pcie_wreg = &amdgpu_invalid_wreg;
1789 	adev->pciep_rreg = &amdgpu_invalid_rreg;
1790 	adev->pciep_wreg = &amdgpu_invalid_wreg;
1791 	adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
1792 	adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
1793 	adev->didt_rreg = &amdgpu_invalid_rreg;
1794 	adev->didt_wreg = &amdgpu_invalid_wreg;
1795 	adev->gc_cac_rreg = &amdgpu_invalid_rreg;
1796 	adev->gc_cac_wreg = &amdgpu_invalid_wreg;
1797 	adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
1798 	adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
1799 
1800 	DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
1801 		 amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
1802 		 pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
1803 
1804 	/* mutex initialization are all done here so we
1805 	 * can recall function without having locking issues */
1806 	atomic_set(&adev->irq.ih.lock, 0);
1807 	mutex_init(&adev->firmware.mutex);
1808 	mutex_init(&adev->pm.mutex);
1809 	mutex_init(&adev->gfx.gpu_clock_mutex);
1810 	mutex_init(&adev->srbm_mutex);
1811 	mutex_init(&adev->gfx.pipe_reserve_mutex);
1812 	mutex_init(&adev->grbm_idx_mutex);
1813 	mutex_init(&adev->mn_lock);
1814 	mutex_init(&adev->virt.vf_errors.lock);
1815 	hash_init(adev->mn_hash);
1816 	mutex_init(&adev->lock_reset);
1817 
1818 	amdgpu_device_check_arguments(adev);
1819 
1820 	spin_lock_init(&adev->mmio_idx_lock);
1821 	spin_lock_init(&adev->smc_idx_lock);
1822 	spin_lock_init(&adev->pcie_idx_lock);
1823 	spin_lock_init(&adev->uvd_ctx_idx_lock);
1824 	spin_lock_init(&adev->didt_idx_lock);
1825 	spin_lock_init(&adev->gc_cac_idx_lock);
1826 	spin_lock_init(&adev->se_cac_idx_lock);
1827 	spin_lock_init(&adev->audio_endpt_idx_lock);
1828 	spin_lock_init(&adev->mm_stats.lock);
1829 
1830 	INIT_LIST_HEAD(&adev->shadow_list);
1831 	mutex_init(&adev->shadow_list_lock);
1832 
1833 	INIT_LIST_HEAD(&adev->ring_lru_list);
1834 	spin_lock_init(&adev->ring_lru_list_lock);
1835 
1836 	INIT_DELAYED_WORK(&adev->late_init_work,
1837 			  amdgpu_device_ip_late_init_func_handler);
1838 
1839 	/* Registers mapping */
1840 	/* TODO: block userspace mapping of io register */
1841 	if (adev->asic_type >= CHIP_BONAIRE) {
1842 		adev->rmmio_base = pci_resource_start(adev->pdev, 5);
1843 		adev->rmmio_size = pci_resource_len(adev->pdev, 5);
1844 	} else {
1845 		adev->rmmio_base = pci_resource_start(adev->pdev, 2);
1846 		adev->rmmio_size = pci_resource_len(adev->pdev, 2);
1847 	}
1848 
1849 	adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
1850 	if (adev->rmmio == NULL) {
1851 		return -ENOMEM;
1852 	}
1853 	DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
1854 	DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
1855 
1856 	/* doorbell bar mapping */
1857 	amdgpu_device_doorbell_init(adev);
1858 
1859 	/* io port mapping */
1860 	for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
1861 		if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) {
1862 			adev->rio_mem_size = pci_resource_len(adev->pdev, i);
1863 			adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size);
1864 			break;
1865 		}
1866 	}
1867 	if (adev->rio_mem == NULL)
1868 		DRM_INFO("PCI I/O BAR is not found.\n");
1869 
1870 	/* early init functions */
1871 	r = amdgpu_device_ip_early_init(adev);
1872 	if (r)
1873 		return r;
1874 
1875 	/* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
1876 	/* this will fail for cards that aren't VGA class devices, just
1877 	 * ignore it */
1878 	vga_client_register(adev->pdev, adev, NULL, amdgpu_device_vga_set_decode);
1879 
1880 	if (amdgpu_device_is_px(ddev))
1881 		runtime = true;
1882 	if (!pci_is_thunderbolt_attached(adev->pdev))
1883 		vga_switcheroo_register_client(adev->pdev,
1884 					       &amdgpu_switcheroo_ops, runtime);
1885 	if (runtime)
1886 		vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
1887 
1888 	/* Read BIOS */
1889 	if (!amdgpu_get_bios(adev)) {
1890 		r = -EINVAL;
1891 		goto failed;
1892 	}
1893 
1894 	r = amdgpu_atombios_init(adev);
1895 	if (r) {
1896 		dev_err(adev->dev, "amdgpu_atombios_init failed\n");
1897 		amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0);
1898 		goto failed;
1899 	}
1900 
1901 	/* detect if we are with an SRIOV vbios */
1902 	amdgpu_device_detect_sriov_bios(adev);
1903 
1904 	/* Post card if necessary */
1905 	if (amdgpu_device_need_post(adev)) {
1906 		if (!adev->bios) {
1907 			dev_err(adev->dev, "no vBIOS found\n");
1908 			r = -EINVAL;
1909 			goto failed;
1910 		}
1911 		DRM_INFO("GPU posting now...\n");
1912 		r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
1913 		if (r) {
1914 			dev_err(adev->dev, "gpu post error!\n");
1915 			goto failed;
1916 		}
1917 	}
1918 
1919 	if (adev->is_atom_fw) {
1920 		/* Initialize clocks */
1921 		r = amdgpu_atomfirmware_get_clock_info(adev);
1922 		if (r) {
1923 			dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
1924 			amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
1925 			goto failed;
1926 		}
1927 	} else {
1928 		/* Initialize clocks */
1929 		r = amdgpu_atombios_get_clock_info(adev);
1930 		if (r) {
1931 			dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
1932 			amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
1933 			goto failed;
1934 		}
1935 		/* init i2c buses */
1936 		if (!amdgpu_device_has_dc_support(adev))
1937 			amdgpu_atombios_i2c_init(adev);
1938 	}
1939 
1940 	/* Fence driver */
1941 	r = amdgpu_fence_driver_init(adev);
1942 	if (r) {
1943 		dev_err(adev->dev, "amdgpu_fence_driver_init failed\n");
1944 		amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
1945 		goto failed;
1946 	}
1947 
1948 	/* init the mode config */
1949 	drm_mode_config_init(adev->ddev);
1950 
1951 	r = amdgpu_device_ip_init(adev);
1952 	if (r) {
1953 		/* failed in exclusive mode due to timeout */
1954 		if (amdgpu_sriov_vf(adev) &&
1955 		    !amdgpu_sriov_runtime(adev) &&
1956 		    amdgpu_virt_mmio_blocked(adev) &&
1957 		    !amdgpu_virt_wait_reset(adev)) {
1958 			dev_err(adev->dev, "VF exclusive mode timeout\n");
1959 			/* Don't send request since VF is inactive. */
1960 			adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
1961 			adev->virt.ops = NULL;
1962 			r = -EAGAIN;
1963 			goto failed;
1964 		}
1965 		dev_err(adev->dev, "amdgpu_device_ip_init failed\n");
1966 		amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
1967 		amdgpu_device_ip_fini(adev);
1968 		goto failed;
1969 	}
1970 
1971 	adev->accel_working = true;
1972 
1973 	amdgpu_vm_check_compute_bug(adev);
1974 
1975 	/* Initialize the buffer migration limit. */
1976 	if (amdgpu_moverate >= 0)
1977 		max_MBps = amdgpu_moverate;
1978 	else
1979 		max_MBps = 8; /* Allow 8 MB/s. */
1980 	/* Get a log2 for easy divisions. */
1981 	adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
1982 
1983 	r = amdgpu_ib_pool_init(adev);
1984 	if (r) {
1985 		dev_err(adev->dev, "IB initialization failed (%d).\n", r);
1986 		amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r);
1987 		goto failed;
1988 	}
1989 
1990 	r = amdgpu_ib_ring_tests(adev);
1991 	if (r)
1992 		DRM_ERROR("ib ring test failed (%d).\n", r);
1993 
1994 	if (amdgpu_sriov_vf(adev))
1995 		amdgpu_virt_init_data_exchange(adev);
1996 
1997 	amdgpu_fbdev_init(adev);
1998 
1999 	r = amdgpu_pm_sysfs_init(adev);
2000 	if (r)
2001 		DRM_ERROR("registering pm debugfs failed (%d).\n", r);
2002 
2003 	r = amdgpu_debugfs_gem_init(adev);
2004 	if (r)
2005 		DRM_ERROR("registering gem debugfs failed (%d).\n", r);
2006 
2007 	r = amdgpu_debugfs_regs_init(adev);
2008 	if (r)
2009 		DRM_ERROR("registering register debugfs failed (%d).\n", r);
2010 
2011 	r = amdgpu_debugfs_firmware_init(adev);
2012 	if (r)
2013 		DRM_ERROR("registering firmware debugfs failed (%d).\n", r);
2014 
2015 	r = amdgpu_debugfs_init(adev);
2016 	if (r)
2017 		DRM_ERROR("Creating debugfs files failed (%d).\n", r);
2018 
2019 	if ((amdgpu_testing & 1)) {
2020 		if (adev->accel_working)
2021 			amdgpu_test_moves(adev);
2022 		else
2023 			DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n");
2024 	}
2025 	if (amdgpu_benchmarking) {
2026 		if (adev->accel_working)
2027 			amdgpu_benchmark(adev, amdgpu_benchmarking);
2028 		else
2029 			DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n");
2030 	}
2031 
2032 	/* enable clockgating, etc. after ib tests, etc. since some blocks require
2033 	 * explicit gating rather than handling it automatically.
2034 	 */
2035 	r = amdgpu_device_ip_late_init(adev);
2036 	if (r) {
2037 		dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n");
2038 		amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
2039 		goto failed;
2040 	}
2041 
2042 	return 0;
2043 
2044 failed:
2045 	amdgpu_vf_error_trans_all(adev);
2046 	if (runtime)
2047 		vga_switcheroo_fini_domain_pm_ops(adev->dev);
2048 
2049 	return r;
2050 }
2051 
2052 /**
2053  * amdgpu_device_fini - tear down the driver
2054  *
2055  * @adev: amdgpu_device pointer
2056  *
2057  * Tear down the driver info (all asics).
2058  * Called at driver shutdown.
2059  */
2060 void amdgpu_device_fini(struct amdgpu_device *adev)
2061 {
2062 	int r;
2063 
2064 	DRM_INFO("amdgpu: finishing device.\n");
2065 	adev->shutdown = true;
2066 	if (adev->mode_info.mode_config_initialized)
2067 		drm_crtc_force_disable_all(adev->ddev);
2068 
2069 	amdgpu_ib_pool_fini(adev);
2070 	amdgpu_fence_driver_fini(adev);
2071 	amdgpu_fbdev_fini(adev);
2072 	r = amdgpu_device_ip_fini(adev);
2073 	if (adev->firmware.gpu_info_fw) {
2074 		release_firmware(adev->firmware.gpu_info_fw);
2075 		adev->firmware.gpu_info_fw = NULL;
2076 	}
2077 	adev->accel_working = false;
2078 	cancel_delayed_work_sync(&adev->late_init_work);
2079 	/* free i2c buses */
2080 	if (!amdgpu_device_has_dc_support(adev))
2081 		amdgpu_i2c_fini(adev);
2082 	amdgpu_atombios_fini(adev);
2083 	kfree(adev->bios);
2084 	adev->bios = NULL;
2085 	if (!pci_is_thunderbolt_attached(adev->pdev))
2086 		vga_switcheroo_unregister_client(adev->pdev);
2087 	if (adev->flags & AMD_IS_PX)
2088 		vga_switcheroo_fini_domain_pm_ops(adev->dev);
2089 	vga_client_register(adev->pdev, NULL, NULL, NULL);
2090 	if (adev->rio_mem)
2091 		pci_iounmap(adev->pdev, adev->rio_mem);
2092 	adev->rio_mem = NULL;
2093 	iounmap(adev->rmmio);
2094 	adev->rmmio = NULL;
2095 	amdgpu_device_doorbell_fini(adev);
2096 	amdgpu_pm_sysfs_fini(adev);
2097 	amdgpu_debugfs_regs_cleanup(adev);
2098 }
2099 
2100 
2101 /*
2102  * Suspend & resume.
2103  */
2104 /**
2105  * amdgpu_device_suspend - initiate device suspend
2106  *
2107  * @pdev: drm dev pointer
2108  * @state: suspend state
2109  *
2110  * Puts the hw in the suspend state (all asics).
2111  * Returns 0 for success or an error on failure.
2112  * Called at driver suspend.
2113  */
2114 int amdgpu_device_suspend(struct drm_device *dev, bool suspend, bool fbcon)
2115 {
2116 	struct amdgpu_device *adev;
2117 	struct drm_crtc *crtc;
2118 	struct drm_connector *connector;
2119 	int r;
2120 
2121 	if (dev == NULL || dev->dev_private == NULL) {
2122 		return -ENODEV;
2123 	}
2124 
2125 	adev = dev->dev_private;
2126 
2127 	if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
2128 		return 0;
2129 
2130 	drm_kms_helper_poll_disable(dev);
2131 
2132 	if (!amdgpu_device_has_dc_support(adev)) {
2133 		/* turn off display hw */
2134 		drm_modeset_lock_all(dev);
2135 		list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2136 			drm_helper_connector_dpms(connector, DRM_MODE_DPMS_OFF);
2137 		}
2138 		drm_modeset_unlock_all(dev);
2139 	}
2140 
2141 	amdgpu_amdkfd_suspend(adev);
2142 
2143 	/* unpin the front buffers and cursors */
2144 	list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2145 		struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2146 		struct amdgpu_framebuffer *rfb = to_amdgpu_framebuffer(crtc->primary->fb);
2147 		struct amdgpu_bo *robj;
2148 
2149 		if (amdgpu_crtc->cursor_bo) {
2150 			struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2151 			r = amdgpu_bo_reserve(aobj, true);
2152 			if (r == 0) {
2153 				amdgpu_bo_unpin(aobj);
2154 				amdgpu_bo_unreserve(aobj);
2155 			}
2156 		}
2157 
2158 		if (rfb == NULL || rfb->obj == NULL) {
2159 			continue;
2160 		}
2161 		robj = gem_to_amdgpu_bo(rfb->obj);
2162 		/* don't unpin kernel fb objects */
2163 		if (!amdgpu_fbdev_robj_is_fb(adev, robj)) {
2164 			r = amdgpu_bo_reserve(robj, true);
2165 			if (r == 0) {
2166 				amdgpu_bo_unpin(robj);
2167 				amdgpu_bo_unreserve(robj);
2168 			}
2169 		}
2170 	}
2171 	/* evict vram memory */
2172 	amdgpu_bo_evict_vram(adev);
2173 
2174 	amdgpu_fence_driver_suspend(adev);
2175 
2176 	r = amdgpu_device_ip_suspend(adev);
2177 
2178 	/* evict remaining vram memory
2179 	 * This second call to evict vram is to evict the gart page table
2180 	 * using the CPU.
2181 	 */
2182 	amdgpu_bo_evict_vram(adev);
2183 
2184 	pci_save_state(dev->pdev);
2185 	if (suspend) {
2186 		/* Shut down the device */
2187 		pci_disable_device(dev->pdev);
2188 		pci_set_power_state(dev->pdev, PCI_D3hot);
2189 	} else {
2190 		r = amdgpu_asic_reset(adev);
2191 		if (r)
2192 			DRM_ERROR("amdgpu asic reset failed\n");
2193 	}
2194 
2195 	if (fbcon) {
2196 		console_lock();
2197 		amdgpu_fbdev_set_suspend(adev, 1);
2198 		console_unlock();
2199 	}
2200 	return 0;
2201 }
2202 
2203 /**
2204  * amdgpu_device_resume - initiate device resume
2205  *
2206  * @pdev: drm dev pointer
2207  *
2208  * Bring the hw back to operating state (all asics).
2209  * Returns 0 for success or an error on failure.
2210  * Called at driver resume.
2211  */
2212 int amdgpu_device_resume(struct drm_device *dev, bool resume, bool fbcon)
2213 {
2214 	struct drm_connector *connector;
2215 	struct amdgpu_device *adev = dev->dev_private;
2216 	struct drm_crtc *crtc;
2217 	int r = 0;
2218 
2219 	if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
2220 		return 0;
2221 
2222 	if (fbcon)
2223 		console_lock();
2224 
2225 	if (resume) {
2226 		pci_set_power_state(dev->pdev, PCI_D0);
2227 		pci_restore_state(dev->pdev);
2228 		r = pci_enable_device(dev->pdev);
2229 		if (r)
2230 			goto unlock;
2231 	}
2232 
2233 	/* post card */
2234 	if (amdgpu_device_need_post(adev)) {
2235 		r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2236 		if (r)
2237 			DRM_ERROR("amdgpu asic init failed\n");
2238 	}
2239 
2240 	r = amdgpu_device_ip_resume(adev);
2241 	if (r) {
2242 		DRM_ERROR("amdgpu_device_ip_resume failed (%d).\n", r);
2243 		goto unlock;
2244 	}
2245 	amdgpu_fence_driver_resume(adev);
2246 
2247 	if (resume) {
2248 		r = amdgpu_ib_ring_tests(adev);
2249 		if (r)
2250 			DRM_ERROR("ib ring test failed (%d).\n", r);
2251 	}
2252 
2253 	r = amdgpu_device_ip_late_init(adev);
2254 	if (r)
2255 		goto unlock;
2256 
2257 	/* pin cursors */
2258 	list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2259 		struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2260 
2261 		if (amdgpu_crtc->cursor_bo) {
2262 			struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2263 			r = amdgpu_bo_reserve(aobj, true);
2264 			if (r == 0) {
2265 				r = amdgpu_bo_pin(aobj,
2266 						  AMDGPU_GEM_DOMAIN_VRAM,
2267 						  &amdgpu_crtc->cursor_addr);
2268 				if (r != 0)
2269 					DRM_ERROR("Failed to pin cursor BO (%d)\n", r);
2270 				amdgpu_bo_unreserve(aobj);
2271 			}
2272 		}
2273 	}
2274 	r = amdgpu_amdkfd_resume(adev);
2275 	if (r)
2276 		return r;
2277 
2278 	/* blat the mode back in */
2279 	if (fbcon) {
2280 		if (!amdgpu_device_has_dc_support(adev)) {
2281 			/* pre DCE11 */
2282 			drm_helper_resume_force_mode(dev);
2283 
2284 			/* turn on display hw */
2285 			drm_modeset_lock_all(dev);
2286 			list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2287 				drm_helper_connector_dpms(connector, DRM_MODE_DPMS_ON);
2288 			}
2289 			drm_modeset_unlock_all(dev);
2290 		}
2291 	}
2292 
2293 	drm_kms_helper_poll_enable(dev);
2294 
2295 	/*
2296 	 * Most of the connector probing functions try to acquire runtime pm
2297 	 * refs to ensure that the GPU is powered on when connector polling is
2298 	 * performed. Since we're calling this from a runtime PM callback,
2299 	 * trying to acquire rpm refs will cause us to deadlock.
2300 	 *
2301 	 * Since we're guaranteed to be holding the rpm lock, it's safe to
2302 	 * temporarily disable the rpm helpers so this doesn't deadlock us.
2303 	 */
2304 #ifdef CONFIG_PM
2305 	dev->dev->power.disable_depth++;
2306 #endif
2307 	if (!amdgpu_device_has_dc_support(adev))
2308 		drm_helper_hpd_irq_event(dev);
2309 	else
2310 		drm_kms_helper_hotplug_event(dev);
2311 #ifdef CONFIG_PM
2312 	dev->dev->power.disable_depth--;
2313 #endif
2314 
2315 	if (fbcon)
2316 		amdgpu_fbdev_set_suspend(adev, 0);
2317 
2318 unlock:
2319 	if (fbcon)
2320 		console_unlock();
2321 
2322 	return r;
2323 }
2324 
2325 static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)
2326 {
2327 	int i;
2328 	bool asic_hang = false;
2329 
2330 	if (amdgpu_sriov_vf(adev))
2331 		return true;
2332 
2333 	for (i = 0; i < adev->num_ip_blocks; i++) {
2334 		if (!adev->ip_blocks[i].status.valid)
2335 			continue;
2336 		if (adev->ip_blocks[i].version->funcs->check_soft_reset)
2337 			adev->ip_blocks[i].status.hang =
2338 				adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
2339 		if (adev->ip_blocks[i].status.hang) {
2340 			DRM_INFO("IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
2341 			asic_hang = true;
2342 		}
2343 	}
2344 	return asic_hang;
2345 }
2346 
2347 static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)
2348 {
2349 	int i, r = 0;
2350 
2351 	for (i = 0; i < adev->num_ip_blocks; i++) {
2352 		if (!adev->ip_blocks[i].status.valid)
2353 			continue;
2354 		if (adev->ip_blocks[i].status.hang &&
2355 		    adev->ip_blocks[i].version->funcs->pre_soft_reset) {
2356 			r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
2357 			if (r)
2358 				return r;
2359 		}
2360 	}
2361 
2362 	return 0;
2363 }
2364 
2365 static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)
2366 {
2367 	int i;
2368 
2369 	for (i = 0; i < adev->num_ip_blocks; i++) {
2370 		if (!adev->ip_blocks[i].status.valid)
2371 			continue;
2372 		if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
2373 		    (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
2374 		    (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
2375 		    (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
2376 		     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
2377 			if (adev->ip_blocks[i].status.hang) {
2378 				DRM_INFO("Some block need full reset!\n");
2379 				return true;
2380 			}
2381 		}
2382 	}
2383 	return false;
2384 }
2385 
2386 static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)
2387 {
2388 	int i, r = 0;
2389 
2390 	for (i = 0; i < adev->num_ip_blocks; i++) {
2391 		if (!adev->ip_blocks[i].status.valid)
2392 			continue;
2393 		if (adev->ip_blocks[i].status.hang &&
2394 		    adev->ip_blocks[i].version->funcs->soft_reset) {
2395 			r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
2396 			if (r)
2397 				return r;
2398 		}
2399 	}
2400 
2401 	return 0;
2402 }
2403 
2404 static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)
2405 {
2406 	int i, r = 0;
2407 
2408 	for (i = 0; i < adev->num_ip_blocks; i++) {
2409 		if (!adev->ip_blocks[i].status.valid)
2410 			continue;
2411 		if (adev->ip_blocks[i].status.hang &&
2412 		    adev->ip_blocks[i].version->funcs->post_soft_reset)
2413 			r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
2414 		if (r)
2415 			return r;
2416 	}
2417 
2418 	return 0;
2419 }
2420 
2421 static int amdgpu_device_recover_vram_from_shadow(struct amdgpu_device *adev,
2422 						  struct amdgpu_ring *ring,
2423 						  struct amdgpu_bo *bo,
2424 						  struct dma_fence **fence)
2425 {
2426 	uint32_t domain;
2427 	int r;
2428 
2429 	if (!bo->shadow)
2430 		return 0;
2431 
2432 	r = amdgpu_bo_reserve(bo, true);
2433 	if (r)
2434 		return r;
2435 	domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
2436 	/* if bo has been evicted, then no need to recover */
2437 	if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
2438 		r = amdgpu_bo_validate(bo->shadow);
2439 		if (r) {
2440 			DRM_ERROR("bo validate failed!\n");
2441 			goto err;
2442 		}
2443 
2444 		r = amdgpu_bo_restore_from_shadow(adev, ring, bo,
2445 						 NULL, fence, true);
2446 		if (r) {
2447 			DRM_ERROR("recover page table failed!\n");
2448 			goto err;
2449 		}
2450 	}
2451 err:
2452 	amdgpu_bo_unreserve(bo);
2453 	return r;
2454 }
2455 
2456 /*
2457  * amdgpu_device_reset - reset ASIC/GPU for bare-metal or passthrough
2458  *
2459  * @adev: amdgpu device pointer
2460  * @reset_flags: output param tells caller the reset result
2461  *
2462  * attempt to do soft-reset or full-reset and reinitialize Asic
2463  * return 0 means successed otherwise failed
2464 */
2465 static int amdgpu_device_reset(struct amdgpu_device *adev,
2466 			       uint64_t* reset_flags)
2467 {
2468 	bool need_full_reset, vram_lost = 0;
2469 	int r;
2470 
2471 	need_full_reset = amdgpu_device_ip_need_full_reset(adev);
2472 
2473 	if (!need_full_reset) {
2474 		amdgpu_device_ip_pre_soft_reset(adev);
2475 		r = amdgpu_device_ip_soft_reset(adev);
2476 		amdgpu_device_ip_post_soft_reset(adev);
2477 		if (r || amdgpu_device_ip_check_soft_reset(adev)) {
2478 			DRM_INFO("soft reset failed, will fallback to full reset!\n");
2479 			need_full_reset = true;
2480 		}
2481 
2482 	}
2483 
2484 	if (need_full_reset) {
2485 		r = amdgpu_device_ip_suspend(adev);
2486 
2487 retry:
2488 		r = amdgpu_asic_reset(adev);
2489 		/* post card */
2490 		amdgpu_atom_asic_init(adev->mode_info.atom_context);
2491 
2492 		if (!r) {
2493 			dev_info(adev->dev, "GPU reset succeeded, trying to resume\n");
2494 			r = amdgpu_device_ip_resume_phase1(adev);
2495 			if (r)
2496 				goto out;
2497 
2498 			vram_lost = amdgpu_device_check_vram_lost(adev);
2499 			if (vram_lost) {
2500 				DRM_ERROR("VRAM is lost!\n");
2501 				atomic_inc(&adev->vram_lost_counter);
2502 			}
2503 
2504 			r = amdgpu_gtt_mgr_recover(
2505 				&adev->mman.bdev.man[TTM_PL_TT]);
2506 			if (r)
2507 				goto out;
2508 
2509 			r = amdgpu_device_ip_resume_phase2(adev);
2510 			if (r)
2511 				goto out;
2512 
2513 			if (vram_lost)
2514 				amdgpu_device_fill_reset_magic(adev);
2515 		}
2516 	}
2517 
2518 out:
2519 	if (!r) {
2520 		amdgpu_irq_gpu_reset_resume_helper(adev);
2521 		r = amdgpu_ib_ring_tests(adev);
2522 		if (r) {
2523 			dev_err(adev->dev, "ib ring test failed (%d).\n", r);
2524 			r = amdgpu_device_ip_suspend(adev);
2525 			need_full_reset = true;
2526 			goto retry;
2527 		}
2528 	}
2529 
2530 	if (reset_flags) {
2531 		if (vram_lost)
2532 			(*reset_flags) |= AMDGPU_RESET_INFO_VRAM_LOST;
2533 
2534 		if (need_full_reset)
2535 			(*reset_flags) |= AMDGPU_RESET_INFO_FULLRESET;
2536 	}
2537 
2538 	return r;
2539 }
2540 
2541 /*
2542  * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
2543  *
2544  * @adev: amdgpu device pointer
2545  * @reset_flags: output param tells caller the reset result
2546  *
2547  * do VF FLR and reinitialize Asic
2548  * return 0 means successed otherwise failed
2549 */
2550 static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
2551 				     uint64_t *reset_flags,
2552 				     bool from_hypervisor)
2553 {
2554 	int r;
2555 
2556 	if (from_hypervisor)
2557 		r = amdgpu_virt_request_full_gpu(adev, true);
2558 	else
2559 		r = amdgpu_virt_reset_gpu(adev);
2560 	if (r)
2561 		return r;
2562 
2563 	/* Resume IP prior to SMC */
2564 	r = amdgpu_device_ip_reinit_early_sriov(adev);
2565 	if (r)
2566 		goto error;
2567 
2568 	/* we need recover gart prior to run SMC/CP/SDMA resume */
2569 	amdgpu_gtt_mgr_recover(&adev->mman.bdev.man[TTM_PL_TT]);
2570 
2571 	/* now we are okay to resume SMC/CP/SDMA */
2572 	r = amdgpu_device_ip_reinit_late_sriov(adev);
2573 	if (r)
2574 		goto error;
2575 
2576 	amdgpu_irq_gpu_reset_resume_helper(adev);
2577 	r = amdgpu_ib_ring_tests(adev);
2578 	if (r)
2579 		dev_err(adev->dev, "[GPU_RESET] ib ring test failed (%d).\n", r);
2580 
2581 error:
2582 	/* release full control of GPU after ib test */
2583 	amdgpu_virt_release_full_gpu(adev, true);
2584 
2585 	if (reset_flags) {
2586 		if (adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
2587 			(*reset_flags) |= AMDGPU_RESET_INFO_VRAM_LOST;
2588 			atomic_inc(&adev->vram_lost_counter);
2589 		}
2590 
2591 		/* VF FLR or hotlink reset is always full-reset */
2592 		(*reset_flags) |= AMDGPU_RESET_INFO_FULLRESET;
2593 	}
2594 
2595 	return r;
2596 }
2597 
2598 /**
2599  * amdgpu_device_gpu_recover - reset the asic and recover scheduler
2600  *
2601  * @adev: amdgpu device pointer
2602  * @job: which job trigger hang
2603  * @force forces reset regardless of amdgpu_gpu_recovery
2604  *
2605  * Attempt to reset the GPU if it has hung (all asics).
2606  * Returns 0 for success or an error on failure.
2607  */
2608 int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
2609 			      struct amdgpu_job *job, bool force)
2610 {
2611 	struct drm_atomic_state *state = NULL;
2612 	uint64_t reset_flags = 0;
2613 	int i, r, resched;
2614 
2615 	if (!force && !amdgpu_device_ip_check_soft_reset(adev)) {
2616 		DRM_INFO("No hardware hang detected. Did some blocks stall?\n");
2617 		return 0;
2618 	}
2619 
2620 	if (!force && (amdgpu_gpu_recovery == 0 ||
2621 			(amdgpu_gpu_recovery == -1  && !amdgpu_sriov_vf(adev)))) {
2622 		DRM_INFO("GPU recovery disabled.\n");
2623 		return 0;
2624 	}
2625 
2626 	dev_info(adev->dev, "GPU reset begin!\n");
2627 
2628 	mutex_lock(&adev->lock_reset);
2629 	atomic_inc(&adev->gpu_reset_counter);
2630 	adev->in_gpu_reset = 1;
2631 
2632 	/* block TTM */
2633 	resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
2634 	/* store modesetting */
2635 	if (amdgpu_device_has_dc_support(adev))
2636 		state = drm_atomic_helper_suspend(adev->ddev);
2637 
2638 	/* block scheduler */
2639 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
2640 		struct amdgpu_ring *ring = adev->rings[i];
2641 
2642 		if (!ring || !ring->sched.thread)
2643 			continue;
2644 
2645 		/* only focus on the ring hit timeout if &job not NULL */
2646 		if (job && job->ring->idx != i)
2647 			continue;
2648 
2649 		kthread_park(ring->sched.thread);
2650 		drm_sched_hw_job_reset(&ring->sched, &job->base);
2651 
2652 		/* after all hw jobs are reset, hw fence is meaningless, so force_completion */
2653 		amdgpu_fence_driver_force_completion(ring);
2654 	}
2655 
2656 	if (amdgpu_sriov_vf(adev))
2657 		r = amdgpu_device_reset_sriov(adev, &reset_flags, job ? false : true);
2658 	else
2659 		r = amdgpu_device_reset(adev, &reset_flags);
2660 
2661 	if (!r) {
2662 		if (((reset_flags & AMDGPU_RESET_INFO_FULLRESET) && !(adev->flags & AMD_IS_APU)) ||
2663 			(reset_flags & AMDGPU_RESET_INFO_VRAM_LOST)) {
2664 			struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
2665 			struct amdgpu_bo *bo, *tmp;
2666 			struct dma_fence *fence = NULL, *next = NULL;
2667 
2668 			DRM_INFO("recover vram bo from shadow\n");
2669 			mutex_lock(&adev->shadow_list_lock);
2670 			list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) {
2671 				next = NULL;
2672 				amdgpu_device_recover_vram_from_shadow(adev, ring, bo, &next);
2673 				if (fence) {
2674 					r = dma_fence_wait(fence, false);
2675 					if (r) {
2676 						WARN(r, "recovery from shadow isn't completed\n");
2677 						break;
2678 					}
2679 				}
2680 
2681 				dma_fence_put(fence);
2682 				fence = next;
2683 			}
2684 			mutex_unlock(&adev->shadow_list_lock);
2685 			if (fence) {
2686 				r = dma_fence_wait(fence, false);
2687 				if (r)
2688 					WARN(r, "recovery from shadow isn't completed\n");
2689 			}
2690 			dma_fence_put(fence);
2691 		}
2692 
2693 		for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
2694 			struct amdgpu_ring *ring = adev->rings[i];
2695 
2696 			if (!ring || !ring->sched.thread)
2697 				continue;
2698 
2699 			/* only focus on the ring hit timeout if &job not NULL */
2700 			if (job && job->ring->idx != i)
2701 				continue;
2702 
2703 			drm_sched_job_recovery(&ring->sched);
2704 			kthread_unpark(ring->sched.thread);
2705 		}
2706 	} else {
2707 		for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
2708 			struct amdgpu_ring *ring = adev->rings[i];
2709 
2710 			if (!ring || !ring->sched.thread)
2711 				continue;
2712 
2713 			/* only focus on the ring hit timeout if &job not NULL */
2714 			if (job && job->ring->idx != i)
2715 				continue;
2716 
2717 			kthread_unpark(adev->rings[i]->sched.thread);
2718 		}
2719 	}
2720 
2721 	if (amdgpu_device_has_dc_support(adev)) {
2722 		if (drm_atomic_helper_resume(adev->ddev, state))
2723 			dev_info(adev->dev, "drm resume failed:%d\n", r);
2724 	} else {
2725 		drm_helper_resume_force_mode(adev->ddev);
2726 	}
2727 
2728 	ttm_bo_unlock_delayed_workqueue(&adev->mman.bdev, resched);
2729 
2730 	if (r) {
2731 		/* bad news, how to tell it to userspace ? */
2732 		dev_info(adev->dev, "GPU reset(%d) failed\n", atomic_read(&adev->gpu_reset_counter));
2733 		amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r);
2734 	} else {
2735 		dev_info(adev->dev, "GPU reset(%d) successed!\n",atomic_read(&adev->gpu_reset_counter));
2736 	}
2737 
2738 	amdgpu_vf_error_trans_all(adev);
2739 	adev->in_gpu_reset = 0;
2740 	mutex_unlock(&adev->lock_reset);
2741 	return r;
2742 }
2743 
2744 void amdgpu_device_get_pcie_info(struct amdgpu_device *adev)
2745 {
2746 	u32 mask;
2747 	int ret;
2748 
2749 	if (amdgpu_pcie_gen_cap)
2750 		adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
2751 
2752 	if (amdgpu_pcie_lane_cap)
2753 		adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
2754 
2755 	/* covers APUs as well */
2756 	if (pci_is_root_bus(adev->pdev->bus)) {
2757 		if (adev->pm.pcie_gen_mask == 0)
2758 			adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
2759 		if (adev->pm.pcie_mlw_mask == 0)
2760 			adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
2761 		return;
2762 	}
2763 
2764 	if (adev->pm.pcie_gen_mask == 0) {
2765 		ret = drm_pcie_get_speed_cap_mask(adev->ddev, &mask);
2766 		if (!ret) {
2767 			adev->pm.pcie_gen_mask = (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
2768 						  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
2769 						  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
2770 
2771 			if (mask & DRM_PCIE_SPEED_25)
2772 				adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
2773 			if (mask & DRM_PCIE_SPEED_50)
2774 				adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2;
2775 			if (mask & DRM_PCIE_SPEED_80)
2776 				adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3;
2777 		} else {
2778 			adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
2779 		}
2780 	}
2781 	if (adev->pm.pcie_mlw_mask == 0) {
2782 		ret = drm_pcie_get_max_link_width(adev->ddev, &mask);
2783 		if (!ret) {
2784 			switch (mask) {
2785 			case 32:
2786 				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
2787 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
2788 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
2789 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
2790 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
2791 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2792 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2793 				break;
2794 			case 16:
2795 				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
2796 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
2797 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
2798 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
2799 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2800 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2801 				break;
2802 			case 12:
2803 				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
2804 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
2805 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
2806 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2807 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2808 				break;
2809 			case 8:
2810 				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
2811 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
2812 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2813 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2814 				break;
2815 			case 4:
2816 				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
2817 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2818 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2819 				break;
2820 			case 2:
2821 				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2822 							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2823 				break;
2824 			case 1:
2825 				adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;
2826 				break;
2827 			default:
2828 				break;
2829 			}
2830 		} else {
2831 			adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
2832 		}
2833 	}
2834 }
2835 
2836