]> Git Repo - linux.git/blob - drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
Merge tag 'gvt-next-2018-09-04' of https://github.com/intel/gvt-linux into drm-intel...
[linux.git] / drivers / gpu / drm / amd / amdgpu / amdgpu_device.c
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/power_supply.h>
29 #include <linux/kthread.h>
30 #include <linux/console.h>
31 #include <linux/slab.h>
32 #include <drm/drmP.h>
33 #include <drm/drm_crtc_helper.h>
34 #include <drm/drm_atomic_helper.h>
35 #include <drm/amdgpu_drm.h>
36 #include <linux/vgaarb.h>
37 #include <linux/vga_switcheroo.h>
38 #include <linux/efi.h>
39 #include "amdgpu.h"
40 #include "amdgpu_trace.h"
41 #include "amdgpu_i2c.h"
42 #include "atom.h"
43 #include "amdgpu_atombios.h"
44 #include "amdgpu_atomfirmware.h"
45 #include "amd_pcie.h"
46 #ifdef CONFIG_DRM_AMDGPU_SI
47 #include "si.h"
48 #endif
49 #ifdef CONFIG_DRM_AMDGPU_CIK
50 #include "cik.h"
51 #endif
52 #include "vi.h"
53 #include "soc15.h"
54 #include "bif/bif_4_1_d.h"
55 #include <linux/pci.h>
56 #include <linux/firmware.h>
57 #include "amdgpu_vf_error.h"
58
59 #include "amdgpu_amdkfd.h"
60 #include "amdgpu_pm.h"
61
62 MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin");
63 MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin");
64 MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin");
65
66 #define AMDGPU_RESUME_MS                2000
67
68 static const char *amdgpu_asic_name[] = {
69         "TAHITI",
70         "PITCAIRN",
71         "VERDE",
72         "OLAND",
73         "HAINAN",
74         "BONAIRE",
75         "KAVERI",
76         "KABINI",
77         "HAWAII",
78         "MULLINS",
79         "TOPAZ",
80         "TONGA",
81         "FIJI",
82         "CARRIZO",
83         "STONEY",
84         "POLARIS10",
85         "POLARIS11",
86         "POLARIS12",
87         "VEGAM",
88         "VEGA10",
89         "VEGA12",
90         "VEGA20",
91         "RAVEN",
92         "LAST",
93 };
94
95 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev);
96
97 /**
98  * amdgpu_device_is_px - Is the device is a dGPU with HG/PX power control
99  *
100  * @dev: drm_device pointer
101  *
102  * Returns true if the device is a dGPU with HG/PX power control,
103  * otherwise return false.
104  */
105 bool amdgpu_device_is_px(struct drm_device *dev)
106 {
107         struct amdgpu_device *adev = dev->dev_private;
108
109         if (adev->flags & AMD_IS_PX)
110                 return true;
111         return false;
112 }
113
114 /*
115  * MMIO register access helper functions.
116  */
117 /**
118  * amdgpu_mm_rreg - read a memory mapped IO register
119  *
120  * @adev: amdgpu_device pointer
121  * @reg: dword aligned register offset
122  * @acc_flags: access flags which require special behavior
123  *
124  * Returns the 32 bit value from the offset specified.
125  */
126 uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg,
127                         uint32_t acc_flags)
128 {
129         uint32_t ret;
130
131         if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
132                 return amdgpu_virt_kiq_rreg(adev, reg);
133
134         if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
135                 ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
136         else {
137                 unsigned long flags;
138
139                 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
140                 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
141                 ret = readl(((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
142                 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
143         }
144         trace_amdgpu_mm_rreg(adev->pdev->device, reg, ret);
145         return ret;
146 }
147
148 /*
149  * MMIO register read with bytes helper functions
150  * @offset:bytes offset from MMIO start
151  *
152 */
153
154 /**
155  * amdgpu_mm_rreg8 - read a memory mapped IO register
156  *
157  * @adev: amdgpu_device pointer
158  * @offset: byte aligned register offset
159  *
160  * Returns the 8 bit value from the offset specified.
161  */
162 uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset) {
163         if (offset < adev->rmmio_size)
164                 return (readb(adev->rmmio + offset));
165         BUG();
166 }
167
168 /*
169  * MMIO register write with bytes helper functions
170  * @offset:bytes offset from MMIO start
171  * @value: the value want to be written to the register
172  *
173 */
174 /**
175  * amdgpu_mm_wreg8 - read a memory mapped IO register
176  *
177  * @adev: amdgpu_device pointer
178  * @offset: byte aligned register offset
179  * @value: 8 bit value to write
180  *
181  * Writes the value specified to the offset specified.
182  */
183 void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value) {
184         if (offset < adev->rmmio_size)
185                 writeb(value, adev->rmmio + offset);
186         else
187                 BUG();
188 }
189
190 /**
191  * amdgpu_mm_wreg - write to a memory mapped IO register
192  *
193  * @adev: amdgpu_device pointer
194  * @reg: dword aligned register offset
195  * @v: 32 bit value to write to the register
196  * @acc_flags: access flags which require special behavior
197  *
198  * Writes the value specified to the offset specified.
199  */
200 void amdgpu_mm_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v,
201                     uint32_t acc_flags)
202 {
203         trace_amdgpu_mm_wreg(adev->pdev->device, reg, v);
204
205         if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
206                 adev->last_mm_index = v;
207         }
208
209         if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
210                 return amdgpu_virt_kiq_wreg(adev, reg, v);
211
212         if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
213                 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
214         else {
215                 unsigned long flags;
216
217                 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
218                 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
219                 writel(v, ((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
220                 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
221         }
222
223         if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
224                 udelay(500);
225         }
226 }
227
228 /**
229  * amdgpu_io_rreg - read an IO register
230  *
231  * @adev: amdgpu_device pointer
232  * @reg: dword aligned register offset
233  *
234  * Returns the 32 bit value from the offset specified.
235  */
236 u32 amdgpu_io_rreg(struct amdgpu_device *adev, u32 reg)
237 {
238         if ((reg * 4) < adev->rio_mem_size)
239                 return ioread32(adev->rio_mem + (reg * 4));
240         else {
241                 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
242                 return ioread32(adev->rio_mem + (mmMM_DATA * 4));
243         }
244 }
245
246 /**
247  * amdgpu_io_wreg - write to an IO register
248  *
249  * @adev: amdgpu_device pointer
250  * @reg: dword aligned register offset
251  * @v: 32 bit value to write to the register
252  *
253  * Writes the value specified to the offset specified.
254  */
255 void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32 v)
256 {
257         if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
258                 adev->last_mm_index = v;
259         }
260
261         if ((reg * 4) < adev->rio_mem_size)
262                 iowrite32(v, adev->rio_mem + (reg * 4));
263         else {
264                 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
265                 iowrite32(v, adev->rio_mem + (mmMM_DATA * 4));
266         }
267
268         if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
269                 udelay(500);
270         }
271 }
272
273 /**
274  * amdgpu_mm_rdoorbell - read a doorbell dword
275  *
276  * @adev: amdgpu_device pointer
277  * @index: doorbell index
278  *
279  * Returns the value in the doorbell aperture at the
280  * requested doorbell index (CIK).
281  */
282 u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
283 {
284         if (index < adev->doorbell.num_doorbells) {
285                 return readl(adev->doorbell.ptr + index);
286         } else {
287                 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
288                 return 0;
289         }
290 }
291
292 /**
293  * amdgpu_mm_wdoorbell - write a doorbell dword
294  *
295  * @adev: amdgpu_device pointer
296  * @index: doorbell index
297  * @v: value to write
298  *
299  * Writes @v to the doorbell aperture at the
300  * requested doorbell index (CIK).
301  */
302 void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v)
303 {
304         if (index < adev->doorbell.num_doorbells) {
305                 writel(v, adev->doorbell.ptr + index);
306         } else {
307                 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
308         }
309 }
310
311 /**
312  * amdgpu_mm_rdoorbell64 - read a doorbell Qword
313  *
314  * @adev: amdgpu_device pointer
315  * @index: doorbell index
316  *
317  * Returns the value in the doorbell aperture at the
318  * requested doorbell index (VEGA10+).
319  */
320 u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index)
321 {
322         if (index < adev->doorbell.num_doorbells) {
323                 return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index));
324         } else {
325                 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
326                 return 0;
327         }
328 }
329
330 /**
331  * amdgpu_mm_wdoorbell64 - write a doorbell Qword
332  *
333  * @adev: amdgpu_device pointer
334  * @index: doorbell index
335  * @v: value to write
336  *
337  * Writes @v to the doorbell aperture at the
338  * requested doorbell index (VEGA10+).
339  */
340 void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v)
341 {
342         if (index < adev->doorbell.num_doorbells) {
343                 atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v);
344         } else {
345                 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
346         }
347 }
348
349 /**
350  * amdgpu_invalid_rreg - dummy reg read function
351  *
352  * @adev: amdgpu device pointer
353  * @reg: offset of register
354  *
355  * Dummy register read function.  Used for register blocks
356  * that certain asics don't have (all asics).
357  * Returns the value in the register.
358  */
359 static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
360 {
361         DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
362         BUG();
363         return 0;
364 }
365
366 /**
367  * amdgpu_invalid_wreg - dummy reg write function
368  *
369  * @adev: amdgpu device pointer
370  * @reg: offset of register
371  * @v: value to write to the register
372  *
373  * Dummy register read function.  Used for register blocks
374  * that certain asics don't have (all asics).
375  */
376 static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
377 {
378         DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
379                   reg, v);
380         BUG();
381 }
382
383 /**
384  * amdgpu_block_invalid_rreg - dummy reg read function
385  *
386  * @adev: amdgpu device pointer
387  * @block: offset of instance
388  * @reg: offset of register
389  *
390  * Dummy register read function.  Used for register blocks
391  * that certain asics don't have (all asics).
392  * Returns the value in the register.
393  */
394 static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
395                                           uint32_t block, uint32_t reg)
396 {
397         DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
398                   reg, block);
399         BUG();
400         return 0;
401 }
402
403 /**
404  * amdgpu_block_invalid_wreg - dummy reg write function
405  *
406  * @adev: amdgpu device pointer
407  * @block: offset of instance
408  * @reg: offset of register
409  * @v: value to write to the register
410  *
411  * Dummy register read function.  Used for register blocks
412  * that certain asics don't have (all asics).
413  */
414 static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
415                                       uint32_t block,
416                                       uint32_t reg, uint32_t v)
417 {
418         DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
419                   reg, block, v);
420         BUG();
421 }
422
423 /**
424  * amdgpu_device_vram_scratch_init - allocate the VRAM scratch page
425  *
426  * @adev: amdgpu device pointer
427  *
428  * Allocates a scratch page of VRAM for use by various things in the
429  * driver.
430  */
431 static int amdgpu_device_vram_scratch_init(struct amdgpu_device *adev)
432 {
433         return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE,
434                                        PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
435                                        &adev->vram_scratch.robj,
436                                        &adev->vram_scratch.gpu_addr,
437                                        (void **)&adev->vram_scratch.ptr);
438 }
439
440 /**
441  * amdgpu_device_vram_scratch_fini - Free the VRAM scratch page
442  *
443  * @adev: amdgpu device pointer
444  *
445  * Frees the VRAM scratch page.
446  */
447 static void amdgpu_device_vram_scratch_fini(struct amdgpu_device *adev)
448 {
449         amdgpu_bo_free_kernel(&adev->vram_scratch.robj, NULL, NULL);
450 }
451
452 /**
453  * amdgpu_device_program_register_sequence - program an array of registers.
454  *
455  * @adev: amdgpu_device pointer
456  * @registers: pointer to the register array
457  * @array_size: size of the register array
458  *
459  * Programs an array or registers with and and or masks.
460  * This is a helper for setting golden registers.
461  */
462 void amdgpu_device_program_register_sequence(struct amdgpu_device *adev,
463                                              const u32 *registers,
464                                              const u32 array_size)
465 {
466         u32 tmp, reg, and_mask, or_mask;
467         int i;
468
469         if (array_size % 3)
470                 return;
471
472         for (i = 0; i < array_size; i +=3) {
473                 reg = registers[i + 0];
474                 and_mask = registers[i + 1];
475                 or_mask = registers[i + 2];
476
477                 if (and_mask == 0xffffffff) {
478                         tmp = or_mask;
479                 } else {
480                         tmp = RREG32(reg);
481                         tmp &= ~and_mask;
482                         tmp |= or_mask;
483                 }
484                 WREG32(reg, tmp);
485         }
486 }
487
488 /**
489  * amdgpu_device_pci_config_reset - reset the GPU
490  *
491  * @adev: amdgpu_device pointer
492  *
493  * Resets the GPU using the pci config reset sequence.
494  * Only applicable to asics prior to vega10.
495  */
496 void amdgpu_device_pci_config_reset(struct amdgpu_device *adev)
497 {
498         pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
499 }
500
501 /*
502  * GPU doorbell aperture helpers function.
503  */
504 /**
505  * amdgpu_device_doorbell_init - Init doorbell driver information.
506  *
507  * @adev: amdgpu_device pointer
508  *
509  * Init doorbell driver information (CIK)
510  * Returns 0 on success, error on failure.
511  */
512 static int amdgpu_device_doorbell_init(struct amdgpu_device *adev)
513 {
514         /* No doorbell on SI hardware generation */
515         if (adev->asic_type < CHIP_BONAIRE) {
516                 adev->doorbell.base = 0;
517                 adev->doorbell.size = 0;
518                 adev->doorbell.num_doorbells = 0;
519                 adev->doorbell.ptr = NULL;
520                 return 0;
521         }
522
523         if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET)
524                 return -EINVAL;
525
526         /* doorbell bar mapping */
527         adev->doorbell.base = pci_resource_start(adev->pdev, 2);
528         adev->doorbell.size = pci_resource_len(adev->pdev, 2);
529
530         adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32),
531                                              AMDGPU_DOORBELL_MAX_ASSIGNMENT+1);
532         if (adev->doorbell.num_doorbells == 0)
533                 return -EINVAL;
534
535         adev->doorbell.ptr = ioremap(adev->doorbell.base,
536                                      adev->doorbell.num_doorbells *
537                                      sizeof(u32));
538         if (adev->doorbell.ptr == NULL)
539                 return -ENOMEM;
540
541         return 0;
542 }
543
544 /**
545  * amdgpu_device_doorbell_fini - Tear down doorbell driver information.
546  *
547  * @adev: amdgpu_device pointer
548  *
549  * Tear down doorbell driver information (CIK)
550  */
551 static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev)
552 {
553         iounmap(adev->doorbell.ptr);
554         adev->doorbell.ptr = NULL;
555 }
556
557
558
559 /*
560  * amdgpu_device_wb_*()
561  * Writeback is the method by which the GPU updates special pages in memory
562  * with the status of certain GPU events (fences, ring pointers,etc.).
563  */
564
565 /**
566  * amdgpu_device_wb_fini - Disable Writeback and free memory
567  *
568  * @adev: amdgpu_device pointer
569  *
570  * Disables Writeback and frees the Writeback memory (all asics).
571  * Used at driver shutdown.
572  */
573 static void amdgpu_device_wb_fini(struct amdgpu_device *adev)
574 {
575         if (adev->wb.wb_obj) {
576                 amdgpu_bo_free_kernel(&adev->wb.wb_obj,
577                                       &adev->wb.gpu_addr,
578                                       (void **)&adev->wb.wb);
579                 adev->wb.wb_obj = NULL;
580         }
581 }
582
583 /**
584  * amdgpu_device_wb_init- Init Writeback driver info and allocate memory
585  *
586  * @adev: amdgpu_device pointer
587  *
588  * Initializes writeback and allocates writeback memory (all asics).
589  * Used at driver startup.
590  * Returns 0 on success or an -error on failure.
591  */
592 static int amdgpu_device_wb_init(struct amdgpu_device *adev)
593 {
594         int r;
595
596         if (adev->wb.wb_obj == NULL) {
597                 /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */
598                 r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8,
599                                             PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
600                                             &adev->wb.wb_obj, &adev->wb.gpu_addr,
601                                             (void **)&adev->wb.wb);
602                 if (r) {
603                         dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
604                         return r;
605                 }
606
607                 adev->wb.num_wb = AMDGPU_MAX_WB;
608                 memset(&adev->wb.used, 0, sizeof(adev->wb.used));
609
610                 /* clear wb memory */
611                 memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8);
612         }
613
614         return 0;
615 }
616
617 /**
618  * amdgpu_device_wb_get - Allocate a wb entry
619  *
620  * @adev: amdgpu_device pointer
621  * @wb: wb index
622  *
623  * Allocate a wb slot for use by the driver (all asics).
624  * Returns 0 on success or -EINVAL on failure.
625  */
626 int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb)
627 {
628         unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
629
630         if (offset < adev->wb.num_wb) {
631                 __set_bit(offset, adev->wb.used);
632                 *wb = offset << 3; /* convert to dw offset */
633                 return 0;
634         } else {
635                 return -EINVAL;
636         }
637 }
638
639 /**
640  * amdgpu_device_wb_free - Free a wb entry
641  *
642  * @adev: amdgpu_device pointer
643  * @wb: wb index
644  *
645  * Free a wb slot allocated for use by the driver (all asics)
646  */
647 void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb)
648 {
649         wb >>= 3;
650         if (wb < adev->wb.num_wb)
651                 __clear_bit(wb, adev->wb.used);
652 }
653
654 /**
655  * amdgpu_device_vram_location - try to find VRAM location
656  *
657  * @adev: amdgpu device structure holding all necessary informations
658  * @mc: memory controller structure holding memory informations
659  * @base: base address at which to put VRAM
660  *
661  * Function will try to place VRAM at base address provided
662  * as parameter.
663  */
664 void amdgpu_device_vram_location(struct amdgpu_device *adev,
665                                  struct amdgpu_gmc *mc, u64 base)
666 {
667         uint64_t limit = (uint64_t)amdgpu_vram_limit << 20;
668
669         mc->vram_start = base;
670         mc->vram_end = mc->vram_start + mc->mc_vram_size - 1;
671         if (limit && limit < mc->real_vram_size)
672                 mc->real_vram_size = limit;
673         dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
674                         mc->mc_vram_size >> 20, mc->vram_start,
675                         mc->vram_end, mc->real_vram_size >> 20);
676 }
677
678 /**
679  * amdgpu_device_gart_location - try to find GART location
680  *
681  * @adev: amdgpu device structure holding all necessary informations
682  * @mc: memory controller structure holding memory informations
683  *
684  * Function will place try to place GART before or after VRAM.
685  *
686  * If GART size is bigger than space left then we ajust GART size.
687  * Thus function will never fails.
688  */
689 void amdgpu_device_gart_location(struct amdgpu_device *adev,
690                                  struct amdgpu_gmc *mc)
691 {
692         u64 size_af, size_bf;
693
694         mc->gart_size += adev->pm.smu_prv_buffer_size;
695
696         size_af = adev->gmc.mc_mask - mc->vram_end;
697         size_bf = mc->vram_start;
698         if (size_bf > size_af) {
699                 if (mc->gart_size > size_bf) {
700                         dev_warn(adev->dev, "limiting GART\n");
701                         mc->gart_size = size_bf;
702                 }
703                 mc->gart_start = 0;
704         } else {
705                 if (mc->gart_size > size_af) {
706                         dev_warn(adev->dev, "limiting GART\n");
707                         mc->gart_size = size_af;
708                 }
709                 /* VCE doesn't like it when BOs cross a 4GB segment, so align
710                  * the GART base on a 4GB boundary as well.
711                  */
712                 mc->gart_start = ALIGN(mc->vram_end + 1, 0x100000000ULL);
713         }
714         mc->gart_end = mc->gart_start + mc->gart_size - 1;
715         dev_info(adev->dev, "GART: %lluM 0x%016llX - 0x%016llX\n",
716                         mc->gart_size >> 20, mc->gart_start, mc->gart_end);
717 }
718
719 /**
720  * amdgpu_device_resize_fb_bar - try to resize FB BAR
721  *
722  * @adev: amdgpu_device pointer
723  *
724  * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not
725  * to fail, but if any of the BARs is not accessible after the size we abort
726  * driver loading by returning -ENODEV.
727  */
728 int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev)
729 {
730         u64 space_needed = roundup_pow_of_two(adev->gmc.real_vram_size);
731         u32 rbar_size = order_base_2(((space_needed >> 20) | 1)) - 1;
732         struct pci_bus *root;
733         struct resource *res;
734         unsigned i;
735         u16 cmd;
736         int r;
737
738         /* Bypass for VF */
739         if (amdgpu_sriov_vf(adev))
740                 return 0;
741
742         /* Check if the root BUS has 64bit memory resources */
743         root = adev->pdev->bus;
744         while (root->parent)
745                 root = root->parent;
746
747         pci_bus_for_each_resource(root, res, i) {
748                 if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) &&
749                     res->start > 0x100000000ull)
750                         break;
751         }
752
753         /* Trying to resize is pointless without a root hub window above 4GB */
754         if (!res)
755                 return 0;
756
757         /* Disable memory decoding while we change the BAR addresses and size */
758         pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd);
759         pci_write_config_word(adev->pdev, PCI_COMMAND,
760                               cmd & ~PCI_COMMAND_MEMORY);
761
762         /* Free the VRAM and doorbell BAR, we most likely need to move both. */
763         amdgpu_device_doorbell_fini(adev);
764         if (adev->asic_type >= CHIP_BONAIRE)
765                 pci_release_resource(adev->pdev, 2);
766
767         pci_release_resource(adev->pdev, 0);
768
769         r = pci_resize_resource(adev->pdev, 0, rbar_size);
770         if (r == -ENOSPC)
771                 DRM_INFO("Not enough PCI address space for a large BAR.");
772         else if (r && r != -ENOTSUPP)
773                 DRM_ERROR("Problem resizing BAR0 (%d).", r);
774
775         pci_assign_unassigned_bus_resources(adev->pdev->bus);
776
777         /* When the doorbell or fb BAR isn't available we have no chance of
778          * using the device.
779          */
780         r = amdgpu_device_doorbell_init(adev);
781         if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET))
782                 return -ENODEV;
783
784         pci_write_config_word(adev->pdev, PCI_COMMAND, cmd);
785
786         return 0;
787 }
788
789 /*
790  * GPU helpers function.
791  */
792 /**
793  * amdgpu_device_need_post - check if the hw need post or not
794  *
795  * @adev: amdgpu_device pointer
796  *
797  * Check if the asic has been initialized (all asics) at driver startup
798  * or post is needed if  hw reset is performed.
799  * Returns true if need or false if not.
800  */
801 bool amdgpu_device_need_post(struct amdgpu_device *adev)
802 {
803         uint32_t reg;
804
805         if (amdgpu_sriov_vf(adev))
806                 return false;
807
808         if (amdgpu_passthrough(adev)) {
809                 /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
810                  * some old smc fw still need driver do vPost otherwise gpu hang, while
811                  * those smc fw version above 22.15 doesn't have this flaw, so we force
812                  * vpost executed for smc version below 22.15
813                  */
814                 if (adev->asic_type == CHIP_FIJI) {
815                         int err;
816                         uint32_t fw_ver;
817                         err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
818                         /* force vPost if error occured */
819                         if (err)
820                                 return true;
821
822                         fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
823                         if (fw_ver < 0x00160e00)
824                                 return true;
825                 }
826         }
827
828         if (adev->has_hw_reset) {
829                 adev->has_hw_reset = false;
830                 return true;
831         }
832
833         /* bios scratch used on CIK+ */
834         if (adev->asic_type >= CHIP_BONAIRE)
835                 return amdgpu_atombios_scratch_need_asic_init(adev);
836
837         /* check MEM_SIZE for older asics */
838         reg = amdgpu_asic_get_config_memsize(adev);
839
840         if ((reg != 0) && (reg != 0xffffffff))
841                 return false;
842
843         return true;
844 }
845
846 /* if we get transitioned to only one device, take VGA back */
847 /**
848  * amdgpu_device_vga_set_decode - enable/disable vga decode
849  *
850  * @cookie: amdgpu_device pointer
851  * @state: enable/disable vga decode
852  *
853  * Enable/disable vga decode (all asics).
854  * Returns VGA resource flags.
855  */
856 static unsigned int amdgpu_device_vga_set_decode(void *cookie, bool state)
857 {
858         struct amdgpu_device *adev = cookie;
859         amdgpu_asic_set_vga_state(adev, state);
860         if (state)
861                 return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
862                        VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
863         else
864                 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
865 }
866
867 /**
868  * amdgpu_device_check_block_size - validate the vm block size
869  *
870  * @adev: amdgpu_device pointer
871  *
872  * Validates the vm block size specified via module parameter.
873  * The vm block size defines number of bits in page table versus page directory,
874  * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
875  * page table and the remaining bits are in the page directory.
876  */
877 static void amdgpu_device_check_block_size(struct amdgpu_device *adev)
878 {
879         /* defines number of bits in page table versus page directory,
880          * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
881          * page table and the remaining bits are in the page directory */
882         if (amdgpu_vm_block_size == -1)
883                 return;
884
885         if (amdgpu_vm_block_size < 9) {
886                 dev_warn(adev->dev, "VM page table size (%d) too small\n",
887                          amdgpu_vm_block_size);
888                 amdgpu_vm_block_size = -1;
889         }
890 }
891
892 /**
893  * amdgpu_device_check_vm_size - validate the vm size
894  *
895  * @adev: amdgpu_device pointer
896  *
897  * Validates the vm size in GB specified via module parameter.
898  * The VM size is the size of the GPU virtual memory space in GB.
899  */
900 static void amdgpu_device_check_vm_size(struct amdgpu_device *adev)
901 {
902         /* no need to check the default value */
903         if (amdgpu_vm_size == -1)
904                 return;
905
906         if (amdgpu_vm_size < 1) {
907                 dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
908                          amdgpu_vm_size);
909                 amdgpu_vm_size = -1;
910         }
911 }
912
913 static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev)
914 {
915         struct sysinfo si;
916         bool is_os_64 = (sizeof(void *) == 8) ? true : false;
917         uint64_t total_memory;
918         uint64_t dram_size_seven_GB = 0x1B8000000;
919         uint64_t dram_size_three_GB = 0xB8000000;
920
921         if (amdgpu_smu_memory_pool_size == 0)
922                 return;
923
924         if (!is_os_64) {
925                 DRM_WARN("Not 64-bit OS, feature not supported\n");
926                 goto def_value;
927         }
928         si_meminfo(&si);
929         total_memory = (uint64_t)si.totalram * si.mem_unit;
930
931         if ((amdgpu_smu_memory_pool_size == 1) ||
932                 (amdgpu_smu_memory_pool_size == 2)) {
933                 if (total_memory < dram_size_three_GB)
934                         goto def_value1;
935         } else if ((amdgpu_smu_memory_pool_size == 4) ||
936                 (amdgpu_smu_memory_pool_size == 8)) {
937                 if (total_memory < dram_size_seven_GB)
938                         goto def_value1;
939         } else {
940                 DRM_WARN("Smu memory pool size not supported\n");
941                 goto def_value;
942         }
943         adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28;
944
945         return;
946
947 def_value1:
948         DRM_WARN("No enough system memory\n");
949 def_value:
950         adev->pm.smu_prv_buffer_size = 0;
951 }
952
953 /**
954  * amdgpu_device_check_arguments - validate module params
955  *
956  * @adev: amdgpu_device pointer
957  *
958  * Validates certain module parameters and updates
959  * the associated values used by the driver (all asics).
960  */
961 static void amdgpu_device_check_arguments(struct amdgpu_device *adev)
962 {
963         if (amdgpu_sched_jobs < 4) {
964                 dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
965                          amdgpu_sched_jobs);
966                 amdgpu_sched_jobs = 4;
967         } else if (!is_power_of_2(amdgpu_sched_jobs)){
968                 dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
969                          amdgpu_sched_jobs);
970                 amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
971         }
972
973         if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) {
974                 /* gart size must be greater or equal to 32M */
975                 dev_warn(adev->dev, "gart size (%d) too small\n",
976                          amdgpu_gart_size);
977                 amdgpu_gart_size = -1;
978         }
979
980         if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) {
981                 /* gtt size must be greater or equal to 32M */
982                 dev_warn(adev->dev, "gtt size (%d) too small\n",
983                                  amdgpu_gtt_size);
984                 amdgpu_gtt_size = -1;
985         }
986
987         /* valid range is between 4 and 9 inclusive */
988         if (amdgpu_vm_fragment_size != -1 &&
989             (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) {
990                 dev_warn(adev->dev, "valid range is between 4 and 9\n");
991                 amdgpu_vm_fragment_size = -1;
992         }
993
994         amdgpu_device_check_smu_prv_buffer_size(adev);
995
996         amdgpu_device_check_vm_size(adev);
997
998         amdgpu_device_check_block_size(adev);
999
1000         if (amdgpu_vram_page_split != -1 && (amdgpu_vram_page_split < 16 ||
1001             !is_power_of_2(amdgpu_vram_page_split))) {
1002                 dev_warn(adev->dev, "invalid VRAM page split (%d)\n",
1003                          amdgpu_vram_page_split);
1004                 amdgpu_vram_page_split = 1024;
1005         }
1006
1007         if (amdgpu_lockup_timeout == 0) {
1008                 dev_warn(adev->dev, "lockup_timeout msut be > 0, adjusting to 10000\n");
1009                 amdgpu_lockup_timeout = 10000;
1010         }
1011
1012         adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type);
1013 }
1014
1015 /**
1016  * amdgpu_switcheroo_set_state - set switcheroo state
1017  *
1018  * @pdev: pci dev pointer
1019  * @state: vga_switcheroo state
1020  *
1021  * Callback for the switcheroo driver.  Suspends or resumes the
1022  * the asics before or after it is powered up using ACPI methods.
1023  */
1024 static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, enum vga_switcheroo_state state)
1025 {
1026         struct drm_device *dev = pci_get_drvdata(pdev);
1027
1028         if (amdgpu_device_is_px(dev) && state == VGA_SWITCHEROO_OFF)
1029                 return;
1030
1031         if (state == VGA_SWITCHEROO_ON) {
1032                 pr_info("amdgpu: switched on\n");
1033                 /* don't suspend or resume card normally */
1034                 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1035
1036                 amdgpu_device_resume(dev, true, true);
1037
1038                 dev->switch_power_state = DRM_SWITCH_POWER_ON;
1039                 drm_kms_helper_poll_enable(dev);
1040         } else {
1041                 pr_info("amdgpu: switched off\n");
1042                 drm_kms_helper_poll_disable(dev);
1043                 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1044                 amdgpu_device_suspend(dev, true, true);
1045                 dev->switch_power_state = DRM_SWITCH_POWER_OFF;
1046         }
1047 }
1048
1049 /**
1050  * amdgpu_switcheroo_can_switch - see if switcheroo state can change
1051  *
1052  * @pdev: pci dev pointer
1053  *
1054  * Callback for the switcheroo driver.  Check of the switcheroo
1055  * state can be changed.
1056  * Returns true if the state can be changed, false if not.
1057  */
1058 static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
1059 {
1060         struct drm_device *dev = pci_get_drvdata(pdev);
1061
1062         /*
1063         * FIXME: open_count is protected by drm_global_mutex but that would lead to
1064         * locking inversion with the driver load path. And the access here is
1065         * completely racy anyway. So don't bother with locking for now.
1066         */
1067         return dev->open_count == 0;
1068 }
1069
1070 static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
1071         .set_gpu_state = amdgpu_switcheroo_set_state,
1072         .reprobe = NULL,
1073         .can_switch = amdgpu_switcheroo_can_switch,
1074 };
1075
1076 /**
1077  * amdgpu_device_ip_set_clockgating_state - set the CG state
1078  *
1079  * @dev: amdgpu_device pointer
1080  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1081  * @state: clockgating state (gate or ungate)
1082  *
1083  * Sets the requested clockgating state for all instances of
1084  * the hardware IP specified.
1085  * Returns the error code from the last instance.
1086  */
1087 int amdgpu_device_ip_set_clockgating_state(void *dev,
1088                                            enum amd_ip_block_type block_type,
1089                                            enum amd_clockgating_state state)
1090 {
1091         struct amdgpu_device *adev = dev;
1092         int i, r = 0;
1093
1094         for (i = 0; i < adev->num_ip_blocks; i++) {
1095                 if (!adev->ip_blocks[i].status.valid)
1096                         continue;
1097                 if (adev->ip_blocks[i].version->type != block_type)
1098                         continue;
1099                 if (!adev->ip_blocks[i].version->funcs->set_clockgating_state)
1100                         continue;
1101                 r = adev->ip_blocks[i].version->funcs->set_clockgating_state(
1102                         (void *)adev, state);
1103                 if (r)
1104                         DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",
1105                                   adev->ip_blocks[i].version->funcs->name, r);
1106         }
1107         return r;
1108 }
1109
1110 /**
1111  * amdgpu_device_ip_set_powergating_state - set the PG state
1112  *
1113  * @dev: amdgpu_device pointer
1114  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1115  * @state: powergating state (gate or ungate)
1116  *
1117  * Sets the requested powergating state for all instances of
1118  * the hardware IP specified.
1119  * Returns the error code from the last instance.
1120  */
1121 int amdgpu_device_ip_set_powergating_state(void *dev,
1122                                            enum amd_ip_block_type block_type,
1123                                            enum amd_powergating_state state)
1124 {
1125         struct amdgpu_device *adev = dev;
1126         int i, r = 0;
1127
1128         for (i = 0; i < adev->num_ip_blocks; i++) {
1129                 if (!adev->ip_blocks[i].status.valid)
1130                         continue;
1131                 if (adev->ip_blocks[i].version->type != block_type)
1132                         continue;
1133                 if (!adev->ip_blocks[i].version->funcs->set_powergating_state)
1134                         continue;
1135                 r = adev->ip_blocks[i].version->funcs->set_powergating_state(
1136                         (void *)adev, state);
1137                 if (r)
1138                         DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",
1139                                   adev->ip_blocks[i].version->funcs->name, r);
1140         }
1141         return r;
1142 }
1143
1144 /**
1145  * amdgpu_device_ip_get_clockgating_state - get the CG state
1146  *
1147  * @adev: amdgpu_device pointer
1148  * @flags: clockgating feature flags
1149  *
1150  * Walks the list of IPs on the device and updates the clockgating
1151  * flags for each IP.
1152  * Updates @flags with the feature flags for each hardware IP where
1153  * clockgating is enabled.
1154  */
1155 void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev,
1156                                             u32 *flags)
1157 {
1158         int i;
1159
1160         for (i = 0; i < adev->num_ip_blocks; i++) {
1161                 if (!adev->ip_blocks[i].status.valid)
1162                         continue;
1163                 if (adev->ip_blocks[i].version->funcs->get_clockgating_state)
1164                         adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags);
1165         }
1166 }
1167
1168 /**
1169  * amdgpu_device_ip_wait_for_idle - wait for idle
1170  *
1171  * @adev: amdgpu_device pointer
1172  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1173  *
1174  * Waits for the request hardware IP to be idle.
1175  * Returns 0 for success or a negative error code on failure.
1176  */
1177 int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev,
1178                                    enum amd_ip_block_type block_type)
1179 {
1180         int i, r;
1181
1182         for (i = 0; i < adev->num_ip_blocks; i++) {
1183                 if (!adev->ip_blocks[i].status.valid)
1184                         continue;
1185                 if (adev->ip_blocks[i].version->type == block_type) {
1186                         r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
1187                         if (r)
1188                                 return r;
1189                         break;
1190                 }
1191         }
1192         return 0;
1193
1194 }
1195
1196 /**
1197  * amdgpu_device_ip_is_idle - is the hardware IP idle
1198  *
1199  * @adev: amdgpu_device pointer
1200  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1201  *
1202  * Check if the hardware IP is idle or not.
1203  * Returns true if it the IP is idle, false if not.
1204  */
1205 bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev,
1206                               enum amd_ip_block_type block_type)
1207 {
1208         int i;
1209
1210         for (i = 0; i < adev->num_ip_blocks; i++) {
1211                 if (!adev->ip_blocks[i].status.valid)
1212                         continue;
1213                 if (adev->ip_blocks[i].version->type == block_type)
1214                         return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
1215         }
1216         return true;
1217
1218 }
1219
1220 /**
1221  * amdgpu_device_ip_get_ip_block - get a hw IP pointer
1222  *
1223  * @adev: amdgpu_device pointer
1224  * @type: Type of hardware IP (SMU, GFX, UVD, etc.)
1225  *
1226  * Returns a pointer to the hardware IP block structure
1227  * if it exists for the asic, otherwise NULL.
1228  */
1229 struct amdgpu_ip_block *
1230 amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,
1231                               enum amd_ip_block_type type)
1232 {
1233         int i;
1234
1235         for (i = 0; i < adev->num_ip_blocks; i++)
1236                 if (adev->ip_blocks[i].version->type == type)
1237                         return &adev->ip_blocks[i];
1238
1239         return NULL;
1240 }
1241
1242 /**
1243  * amdgpu_device_ip_block_version_cmp
1244  *
1245  * @adev: amdgpu_device pointer
1246  * @type: enum amd_ip_block_type
1247  * @major: major version
1248  * @minor: minor version
1249  *
1250  * return 0 if equal or greater
1251  * return 1 if smaller or the ip_block doesn't exist
1252  */
1253 int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev,
1254                                        enum amd_ip_block_type type,
1255                                        u32 major, u32 minor)
1256 {
1257         struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type);
1258
1259         if (ip_block && ((ip_block->version->major > major) ||
1260                         ((ip_block->version->major == major) &&
1261                         (ip_block->version->minor >= minor))))
1262                 return 0;
1263
1264         return 1;
1265 }
1266
1267 /**
1268  * amdgpu_device_ip_block_add
1269  *
1270  * @adev: amdgpu_device pointer
1271  * @ip_block_version: pointer to the IP to add
1272  *
1273  * Adds the IP block driver information to the collection of IPs
1274  * on the asic.
1275  */
1276 int amdgpu_device_ip_block_add(struct amdgpu_device *adev,
1277                                const struct amdgpu_ip_block_version *ip_block_version)
1278 {
1279         if (!ip_block_version)
1280                 return -EINVAL;
1281
1282         DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks,
1283                   ip_block_version->funcs->name);
1284
1285         adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
1286
1287         return 0;
1288 }
1289
1290 /**
1291  * amdgpu_device_enable_virtual_display - enable virtual display feature
1292  *
1293  * @adev: amdgpu_device pointer
1294  *
1295  * Enabled the virtual display feature if the user has enabled it via
1296  * the module parameter virtual_display.  This feature provides a virtual
1297  * display hardware on headless boards or in virtualized environments.
1298  * This function parses and validates the configuration string specified by
1299  * the user and configues the virtual display configuration (number of
1300  * virtual connectors, crtcs, etc.) specified.
1301  */
1302 static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
1303 {
1304         adev->enable_virtual_display = false;
1305
1306         if (amdgpu_virtual_display) {
1307                 struct drm_device *ddev = adev->ddev;
1308                 const char *pci_address_name = pci_name(ddev->pdev);
1309                 char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
1310
1311                 pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
1312                 pciaddstr_tmp = pciaddstr;
1313                 while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
1314                         pciaddname = strsep(&pciaddname_tmp, ",");
1315                         if (!strcmp("all", pciaddname)
1316                             || !strcmp(pci_address_name, pciaddname)) {
1317                                 long num_crtc;
1318                                 int res = -1;
1319
1320                                 adev->enable_virtual_display = true;
1321
1322                                 if (pciaddname_tmp)
1323                                         res = kstrtol(pciaddname_tmp, 10,
1324                                                       &num_crtc);
1325
1326                                 if (!res) {
1327                                         if (num_crtc < 1)
1328                                                 num_crtc = 1;
1329                                         if (num_crtc > 6)
1330                                                 num_crtc = 6;
1331                                         adev->mode_info.num_crtc = num_crtc;
1332                                 } else {
1333                                         adev->mode_info.num_crtc = 1;
1334                                 }
1335                                 break;
1336                         }
1337                 }
1338
1339                 DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
1340                          amdgpu_virtual_display, pci_address_name,
1341                          adev->enable_virtual_display, adev->mode_info.num_crtc);
1342
1343                 kfree(pciaddstr);
1344         }
1345 }
1346
1347 /**
1348  * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware
1349  *
1350  * @adev: amdgpu_device pointer
1351  *
1352  * Parses the asic configuration parameters specified in the gpu info
1353  * firmware and makes them availale to the driver for use in configuring
1354  * the asic.
1355  * Returns 0 on success, -EINVAL on failure.
1356  */
1357 static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev)
1358 {
1359         const char *chip_name;
1360         char fw_name[30];
1361         int err;
1362         const struct gpu_info_firmware_header_v1_0 *hdr;
1363
1364         adev->firmware.gpu_info_fw = NULL;
1365
1366         switch (adev->asic_type) {
1367         case CHIP_TOPAZ:
1368         case CHIP_TONGA:
1369         case CHIP_FIJI:
1370         case CHIP_POLARIS10:
1371         case CHIP_POLARIS11:
1372         case CHIP_POLARIS12:
1373         case CHIP_VEGAM:
1374         case CHIP_CARRIZO:
1375         case CHIP_STONEY:
1376 #ifdef CONFIG_DRM_AMDGPU_SI
1377         case CHIP_VERDE:
1378         case CHIP_TAHITI:
1379         case CHIP_PITCAIRN:
1380         case CHIP_OLAND:
1381         case CHIP_HAINAN:
1382 #endif
1383 #ifdef CONFIG_DRM_AMDGPU_CIK
1384         case CHIP_BONAIRE:
1385         case CHIP_HAWAII:
1386         case CHIP_KAVERI:
1387         case CHIP_KABINI:
1388         case CHIP_MULLINS:
1389 #endif
1390         case CHIP_VEGA20:
1391         default:
1392                 return 0;
1393         case CHIP_VEGA10:
1394                 chip_name = "vega10";
1395                 break;
1396         case CHIP_VEGA12:
1397                 chip_name = "vega12";
1398                 break;
1399         case CHIP_RAVEN:
1400                 chip_name = "raven";
1401                 break;
1402         }
1403
1404         snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name);
1405         err = request_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev);
1406         if (err) {
1407                 dev_err(adev->dev,
1408                         "Failed to load gpu_info firmware \"%s\"\n",
1409                         fw_name);
1410                 goto out;
1411         }
1412         err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw);
1413         if (err) {
1414                 dev_err(adev->dev,
1415                         "Failed to validate gpu_info firmware \"%s\"\n",
1416                         fw_name);
1417                 goto out;
1418         }
1419
1420         hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data;
1421         amdgpu_ucode_print_gpu_info_hdr(&hdr->header);
1422
1423         switch (hdr->version_major) {
1424         case 1:
1425         {
1426                 const struct gpu_info_firmware_v1_0 *gpu_info_fw =
1427                         (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data +
1428                                                                 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1429
1430                 adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se);
1431                 adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh);
1432                 adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se);
1433                 adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se);
1434                 adev->gfx.config.max_texture_channel_caches =
1435                         le32_to_cpu(gpu_info_fw->gc_num_tccs);
1436                 adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs);
1437                 adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds);
1438                 adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth);
1439                 adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth);
1440                 adev->gfx.config.double_offchip_lds_buf =
1441                         le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer);
1442                 adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size);
1443                 adev->gfx.cu_info.max_waves_per_simd =
1444                         le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd);
1445                 adev->gfx.cu_info.max_scratch_slots_per_cu =
1446                         le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu);
1447                 adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size);
1448                 break;
1449         }
1450         default:
1451                 dev_err(adev->dev,
1452                         "Unsupported gpu_info table %d\n", hdr->header.ucode_version);
1453                 err = -EINVAL;
1454                 goto out;
1455         }
1456 out:
1457         return err;
1458 }
1459
1460 /**
1461  * amdgpu_device_ip_early_init - run early init for hardware IPs
1462  *
1463  * @adev: amdgpu_device pointer
1464  *
1465  * Early initialization pass for hardware IPs.  The hardware IPs that make
1466  * up each asic are discovered each IP's early_init callback is run.  This
1467  * is the first stage in initializing the asic.
1468  * Returns 0 on success, negative error code on failure.
1469  */
1470 static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
1471 {
1472         int i, r;
1473
1474         amdgpu_device_enable_virtual_display(adev);
1475
1476         switch (adev->asic_type) {
1477         case CHIP_TOPAZ:
1478         case CHIP_TONGA:
1479         case CHIP_FIJI:
1480         case CHIP_POLARIS10:
1481         case CHIP_POLARIS11:
1482         case CHIP_POLARIS12:
1483         case CHIP_VEGAM:
1484         case CHIP_CARRIZO:
1485         case CHIP_STONEY:
1486                 if (adev->asic_type == CHIP_CARRIZO || adev->asic_type == CHIP_STONEY)
1487                         adev->family = AMDGPU_FAMILY_CZ;
1488                 else
1489                         adev->family = AMDGPU_FAMILY_VI;
1490
1491                 r = vi_set_ip_blocks(adev);
1492                 if (r)
1493                         return r;
1494                 break;
1495 #ifdef CONFIG_DRM_AMDGPU_SI
1496         case CHIP_VERDE:
1497         case CHIP_TAHITI:
1498         case CHIP_PITCAIRN:
1499         case CHIP_OLAND:
1500         case CHIP_HAINAN:
1501                 adev->family = AMDGPU_FAMILY_SI;
1502                 r = si_set_ip_blocks(adev);
1503                 if (r)
1504                         return r;
1505                 break;
1506 #endif
1507 #ifdef CONFIG_DRM_AMDGPU_CIK
1508         case CHIP_BONAIRE:
1509         case CHIP_HAWAII:
1510         case CHIP_KAVERI:
1511         case CHIP_KABINI:
1512         case CHIP_MULLINS:
1513                 if ((adev->asic_type == CHIP_BONAIRE) || (adev->asic_type == CHIP_HAWAII))
1514                         adev->family = AMDGPU_FAMILY_CI;
1515                 else
1516                         adev->family = AMDGPU_FAMILY_KV;
1517
1518                 r = cik_set_ip_blocks(adev);
1519                 if (r)
1520                         return r;
1521                 break;
1522 #endif
1523         case CHIP_VEGA10:
1524         case CHIP_VEGA12:
1525         case CHIP_VEGA20:
1526         case CHIP_RAVEN:
1527                 if (adev->asic_type == CHIP_RAVEN)
1528                         adev->family = AMDGPU_FAMILY_RV;
1529                 else
1530                         adev->family = AMDGPU_FAMILY_AI;
1531
1532                 r = soc15_set_ip_blocks(adev);
1533                 if (r)
1534                         return r;
1535                 break;
1536         default:
1537                 /* FIXME: not supported yet */
1538                 return -EINVAL;
1539         }
1540
1541         r = amdgpu_device_parse_gpu_info_fw(adev);
1542         if (r)
1543                 return r;
1544
1545         amdgpu_amdkfd_device_probe(adev);
1546
1547         if (amdgpu_sriov_vf(adev)) {
1548                 r = amdgpu_virt_request_full_gpu(adev, true);
1549                 if (r)
1550                         return -EAGAIN;
1551         }
1552
1553         adev->powerplay.pp_feature = amdgpu_pp_feature_mask;
1554
1555         for (i = 0; i < adev->num_ip_blocks; i++) {
1556                 if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
1557                         DRM_ERROR("disabled ip block: %d <%s>\n",
1558                                   i, adev->ip_blocks[i].version->funcs->name);
1559                         adev->ip_blocks[i].status.valid = false;
1560                 } else {
1561                         if (adev->ip_blocks[i].version->funcs->early_init) {
1562                                 r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
1563                                 if (r == -ENOENT) {
1564                                         adev->ip_blocks[i].status.valid = false;
1565                                 } else if (r) {
1566                                         DRM_ERROR("early_init of IP block <%s> failed %d\n",
1567                                                   adev->ip_blocks[i].version->funcs->name, r);
1568                                         return r;
1569                                 } else {
1570                                         adev->ip_blocks[i].status.valid = true;
1571                                 }
1572                         } else {
1573                                 adev->ip_blocks[i].status.valid = true;
1574                         }
1575                 }
1576         }
1577
1578         adev->cg_flags &= amdgpu_cg_mask;
1579         adev->pg_flags &= amdgpu_pg_mask;
1580
1581         return 0;
1582 }
1583
1584 /**
1585  * amdgpu_device_ip_init - run init for hardware IPs
1586  *
1587  * @adev: amdgpu_device pointer
1588  *
1589  * Main initialization pass for hardware IPs.  The list of all the hardware
1590  * IPs that make up the asic is walked and the sw_init and hw_init callbacks
1591  * are run.  sw_init initializes the software state associated with each IP
1592  * and hw_init initializes the hardware associated with each IP.
1593  * Returns 0 on success, negative error code on failure.
1594  */
1595 static int amdgpu_device_ip_init(struct amdgpu_device *adev)
1596 {
1597         int i, r;
1598
1599         for (i = 0; i < adev->num_ip_blocks; i++) {
1600                 if (!adev->ip_blocks[i].status.valid)
1601                         continue;
1602                 r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
1603                 if (r) {
1604                         DRM_ERROR("sw_init of IP block <%s> failed %d\n",
1605                                   adev->ip_blocks[i].version->funcs->name, r);
1606                         return r;
1607                 }
1608                 adev->ip_blocks[i].status.sw = true;
1609
1610                 /* need to do gmc hw init early so we can allocate gpu mem */
1611                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1612                         r = amdgpu_device_vram_scratch_init(adev);
1613                         if (r) {
1614                                 DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r);
1615                                 return r;
1616                         }
1617                         r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1618                         if (r) {
1619                                 DRM_ERROR("hw_init %d failed %d\n", i, r);
1620                                 return r;
1621                         }
1622                         r = amdgpu_device_wb_init(adev);
1623                         if (r) {
1624                                 DRM_ERROR("amdgpu_device_wb_init failed %d\n", r);
1625                                 return r;
1626                         }
1627                         adev->ip_blocks[i].status.hw = true;
1628
1629                         /* right after GMC hw init, we create CSA */
1630                         if (amdgpu_sriov_vf(adev)) {
1631                                 r = amdgpu_allocate_static_csa(adev);
1632                                 if (r) {
1633                                         DRM_ERROR("allocate CSA failed %d\n", r);
1634                                         return r;
1635                                 }
1636                         }
1637                 }
1638         }
1639
1640         for (i = 0; i < adev->num_ip_blocks; i++) {
1641                 if (!adev->ip_blocks[i].status.sw)
1642                         continue;
1643                 if (adev->ip_blocks[i].status.hw)
1644                         continue;
1645                 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1646                 if (r) {
1647                         DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1648                                   adev->ip_blocks[i].version->funcs->name, r);
1649                         return r;
1650                 }
1651                 adev->ip_blocks[i].status.hw = true;
1652         }
1653
1654         amdgpu_amdkfd_device_init(adev);
1655
1656         if (amdgpu_sriov_vf(adev))
1657                 amdgpu_virt_release_full_gpu(adev, true);
1658
1659         return 0;
1660 }
1661
1662 /**
1663  * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer
1664  *
1665  * @adev: amdgpu_device pointer
1666  *
1667  * Writes a reset magic value to the gart pointer in VRAM.  The driver calls
1668  * this function before a GPU reset.  If the value is retained after a
1669  * GPU reset, VRAM has not been lost.  Some GPU resets may destry VRAM contents.
1670  */
1671 static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev)
1672 {
1673         memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
1674 }
1675
1676 /**
1677  * amdgpu_device_check_vram_lost - check if vram is valid
1678  *
1679  * @adev: amdgpu_device pointer
1680  *
1681  * Checks the reset magic value written to the gart pointer in VRAM.
1682  * The driver calls this after a GPU reset to see if the contents of
1683  * VRAM is lost or now.
1684  * returns true if vram is lost, false if not.
1685  */
1686 static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev)
1687 {
1688         return !!memcmp(adev->gart.ptr, adev->reset_magic,
1689                         AMDGPU_RESET_MAGIC_NUM);
1690 }
1691
1692 /**
1693  * amdgpu_device_ip_late_set_cg_state - late init for clockgating
1694  *
1695  * @adev: amdgpu_device pointer
1696  *
1697  * Late initialization pass enabling clockgating for hardware IPs.
1698  * The list of all the hardware IPs that make up the asic is walked and the
1699  * set_clockgating_state callbacks are run.  This stage is run late
1700  * in the init process.
1701  * Returns 0 on success, negative error code on failure.
1702  */
1703 static int amdgpu_device_ip_late_set_cg_state(struct amdgpu_device *adev)
1704 {
1705         int i = 0, r;
1706
1707         if (amdgpu_emu_mode == 1)
1708                 return 0;
1709
1710         for (i = 0; i < adev->num_ip_blocks; i++) {
1711                 if (!adev->ip_blocks[i].status.valid)
1712                         continue;
1713                 /* skip CG for VCE/UVD, it's handled specially */
1714                 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1715                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1716                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
1717                     adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1718                         /* enable clockgating to save power */
1719                         r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1720                                                                                      AMD_CG_STATE_GATE);
1721                         if (r) {
1722                                 DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
1723                                           adev->ip_blocks[i].version->funcs->name, r);
1724                                 return r;
1725                         }
1726                 }
1727         }
1728
1729         return 0;
1730 }
1731
1732 static int amdgpu_device_ip_late_set_pg_state(struct amdgpu_device *adev)
1733 {
1734         int i = 0, r;
1735
1736         if (amdgpu_emu_mode == 1)
1737                 return 0;
1738
1739         for (i = 0; i < adev->num_ip_blocks; i++) {
1740                 if (!adev->ip_blocks[i].status.valid)
1741                         continue;
1742                 /* skip CG for VCE/UVD, it's handled specially */
1743                 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1744                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1745                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
1746                     adev->ip_blocks[i].version->funcs->set_powergating_state) {
1747                         /* enable powergating to save power */
1748                         r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev,
1749                                                                                      AMD_PG_STATE_GATE);
1750                         if (r) {
1751                                 DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n",
1752                                           adev->ip_blocks[i].version->funcs->name, r);
1753                                 return r;
1754                         }
1755                 }
1756         }
1757         return 0;
1758 }
1759
1760 /**
1761  * amdgpu_device_ip_late_init - run late init for hardware IPs
1762  *
1763  * @adev: amdgpu_device pointer
1764  *
1765  * Late initialization pass for hardware IPs.  The list of all the hardware
1766  * IPs that make up the asic is walked and the late_init callbacks are run.
1767  * late_init covers any special initialization that an IP requires
1768  * after all of the have been initialized or something that needs to happen
1769  * late in the init process.
1770  * Returns 0 on success, negative error code on failure.
1771  */
1772 static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
1773 {
1774         int i = 0, r;
1775
1776         for (i = 0; i < adev->num_ip_blocks; i++) {
1777                 if (!adev->ip_blocks[i].status.valid)
1778                         continue;
1779                 if (adev->ip_blocks[i].version->funcs->late_init) {
1780                         r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
1781                         if (r) {
1782                                 DRM_ERROR("late_init of IP block <%s> failed %d\n",
1783                                           adev->ip_blocks[i].version->funcs->name, r);
1784                                 return r;
1785                         }
1786                         adev->ip_blocks[i].status.late_initialized = true;
1787                 }
1788         }
1789
1790         amdgpu_device_ip_late_set_cg_state(adev);
1791         amdgpu_device_ip_late_set_pg_state(adev);
1792
1793         queue_delayed_work(system_wq, &adev->late_init_work,
1794                            msecs_to_jiffies(AMDGPU_RESUME_MS));
1795
1796         amdgpu_device_fill_reset_magic(adev);
1797
1798         return 0;
1799 }
1800
1801 /**
1802  * amdgpu_device_ip_fini - run fini for hardware IPs
1803  *
1804  * @adev: amdgpu_device pointer
1805  *
1806  * Main teardown pass for hardware IPs.  The list of all the hardware
1807  * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks
1808  * are run.  hw_fini tears down the hardware associated with each IP
1809  * and sw_fini tears down any software state associated with each IP.
1810  * Returns 0 on success, negative error code on failure.
1811  */
1812 static int amdgpu_device_ip_fini(struct amdgpu_device *adev)
1813 {
1814         int i, r;
1815
1816         amdgpu_amdkfd_device_fini(adev);
1817         /* need to disable SMC first */
1818         for (i = 0; i < adev->num_ip_blocks; i++) {
1819                 if (!adev->ip_blocks[i].status.hw)
1820                         continue;
1821                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC &&
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                         if (adev->powerplay.pp_funcs->set_powergating_by_smu)
1832                                 amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false);
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                         adev->ip_blocks[i].status.hw = false;
1840                         break;
1841                 }
1842         }
1843
1844         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1845                 if (!adev->ip_blocks[i].status.hw)
1846                         continue;
1847
1848                 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1849                         adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1850                         adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
1851                         adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1852                         /* ungate blocks before hw fini so that we can shutdown the blocks safely */
1853                         r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1854                                                                                      AMD_CG_STATE_UNGATE);
1855                         if (r) {
1856                                 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1857                                           adev->ip_blocks[i].version->funcs->name, r);
1858                                 return r;
1859                         }
1860                 }
1861
1862                 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
1863                 /* XXX handle errors */
1864                 if (r) {
1865                         DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
1866                                   adev->ip_blocks[i].version->funcs->name, r);
1867                 }
1868
1869                 adev->ip_blocks[i].status.hw = false;
1870         }
1871
1872
1873         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1874                 if (!adev->ip_blocks[i].status.sw)
1875                         continue;
1876
1877                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1878                         amdgpu_free_static_csa(adev);
1879                         amdgpu_device_wb_fini(adev);
1880                         amdgpu_device_vram_scratch_fini(adev);
1881                 }
1882
1883                 r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
1884                 /* XXX handle errors */
1885                 if (r) {
1886                         DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
1887                                   adev->ip_blocks[i].version->funcs->name, r);
1888                 }
1889                 adev->ip_blocks[i].status.sw = false;
1890                 adev->ip_blocks[i].status.valid = false;
1891         }
1892
1893         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1894                 if (!adev->ip_blocks[i].status.late_initialized)
1895                         continue;
1896                 if (adev->ip_blocks[i].version->funcs->late_fini)
1897                         adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
1898                 adev->ip_blocks[i].status.late_initialized = false;
1899         }
1900
1901         if (amdgpu_sriov_vf(adev))
1902                 if (amdgpu_virt_release_full_gpu(adev, false))
1903                         DRM_ERROR("failed to release exclusive mode on fini\n");
1904
1905         return 0;
1906 }
1907
1908 /**
1909  * amdgpu_device_ip_late_init_func_handler - work handler for clockgating
1910  *
1911  * @work: work_struct
1912  *
1913  * Work handler for amdgpu_device_ip_late_set_cg_state.  We put the
1914  * clockgating setup into a worker thread to speed up driver init and
1915  * resume from suspend.
1916  */
1917 static void amdgpu_device_ip_late_init_func_handler(struct work_struct *work)
1918 {
1919         struct amdgpu_device *adev =
1920                 container_of(work, struct amdgpu_device, late_init_work.work);
1921         int r;
1922
1923         r = amdgpu_ib_ring_tests(adev);
1924         if (r)
1925                 DRM_ERROR("ib ring test failed (%d).\n", r);
1926 }
1927
1928 /**
1929  * amdgpu_device_ip_suspend - run suspend for hardware IPs
1930  *
1931  * @adev: amdgpu_device pointer
1932  *
1933  * Main suspend function for hardware IPs.  The list of all the hardware
1934  * IPs that make up the asic is walked, clockgating is disabled and the
1935  * suspend callbacks are run.  suspend puts the hardware and software state
1936  * in each IP into a state suitable for suspend.
1937  * Returns 0 on success, negative error code on failure.
1938  */
1939 int amdgpu_device_ip_suspend(struct amdgpu_device *adev)
1940 {
1941         int i, r;
1942
1943         if (amdgpu_sriov_vf(adev))
1944                 amdgpu_virt_request_full_gpu(adev, false);
1945
1946         /* ungate SMC block first */
1947         r = amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_SMC,
1948                                                    AMD_CG_STATE_UNGATE);
1949         if (r) {
1950                 DRM_ERROR("set_clockgating_state(ungate) SMC failed %d\n", r);
1951         }
1952
1953         /* call smu to disable gfx off feature first when suspend */
1954         if (adev->powerplay.pp_funcs->set_powergating_by_smu)
1955                 amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false);
1956
1957         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1958                 if (!adev->ip_blocks[i].status.valid)
1959                         continue;
1960                 /* ungate blocks so that suspend can properly shut them down */
1961                 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_SMC &&
1962                         adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1963                         r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1964                                                                                      AMD_CG_STATE_UNGATE);
1965                         if (r) {
1966                                 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1967                                           adev->ip_blocks[i].version->funcs->name, r);
1968                         }
1969                 }
1970                 /* XXX handle errors */
1971                 r = adev->ip_blocks[i].version->funcs->suspend(adev);
1972                 /* XXX handle errors */
1973                 if (r) {
1974                         DRM_ERROR("suspend of IP block <%s> failed %d\n",
1975                                   adev->ip_blocks[i].version->funcs->name, r);
1976                 }
1977         }
1978
1979         if (amdgpu_sriov_vf(adev))
1980                 amdgpu_virt_release_full_gpu(adev, false);
1981
1982         return 0;
1983 }
1984
1985 static int amdgpu_device_ip_reinit_early_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_GMC,
1991                 AMD_IP_BLOCK_TYPE_COMMON,
1992                 AMD_IP_BLOCK_TYPE_IH,
1993         };
1994
1995         for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
1996                 int j;
1997                 struct amdgpu_ip_block *block;
1998
1999                 for (j = 0; j < adev->num_ip_blocks; j++) {
2000                         block = &adev->ip_blocks[j];
2001
2002                         if (block->version->type != ip_order[i] ||
2003                                 !block->status.valid)
2004                                 continue;
2005
2006                         r = block->version->funcs->hw_init(adev);
2007                         DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"successed");
2008                         if (r)
2009                                 return r;
2010                 }
2011         }
2012
2013         return 0;
2014 }
2015
2016 static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
2017 {
2018         int i, r;
2019
2020         static enum amd_ip_block_type ip_order[] = {
2021                 AMD_IP_BLOCK_TYPE_SMC,
2022                 AMD_IP_BLOCK_TYPE_PSP,
2023                 AMD_IP_BLOCK_TYPE_DCE,
2024                 AMD_IP_BLOCK_TYPE_GFX,
2025                 AMD_IP_BLOCK_TYPE_SDMA,
2026                 AMD_IP_BLOCK_TYPE_UVD,
2027                 AMD_IP_BLOCK_TYPE_VCE
2028         };
2029
2030         for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
2031                 int j;
2032                 struct amdgpu_ip_block *block;
2033
2034                 for (j = 0; j < adev->num_ip_blocks; j++) {
2035                         block = &adev->ip_blocks[j];
2036
2037                         if (block->version->type != ip_order[i] ||
2038                                 !block->status.valid)
2039                                 continue;
2040
2041                         r = block->version->funcs->hw_init(adev);
2042                         DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"successed");
2043                         if (r)
2044                                 return r;
2045                 }
2046         }
2047
2048         return 0;
2049 }
2050
2051 /**
2052  * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs
2053  *
2054  * @adev: amdgpu_device pointer
2055  *
2056  * First resume function for hardware IPs.  The list of all the hardware
2057  * IPs that make up the asic is walked and the resume callbacks are run for
2058  * COMMON, GMC, and IH.  resume puts the hardware into a functional state
2059  * after a suspend and updates the software state as necessary.  This
2060  * function is also used for restoring the GPU after a GPU reset.
2061  * Returns 0 on success, negative error code on failure.
2062  */
2063 static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)
2064 {
2065         int i, r;
2066
2067         for (i = 0; i < adev->num_ip_blocks; i++) {
2068                 if (!adev->ip_blocks[i].status.valid)
2069                         continue;
2070                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2071                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2072                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
2073                         r = adev->ip_blocks[i].version->funcs->resume(adev);
2074                         if (r) {
2075                                 DRM_ERROR("resume of IP block <%s> failed %d\n",
2076                                           adev->ip_blocks[i].version->funcs->name, r);
2077                                 return r;
2078                         }
2079                 }
2080         }
2081
2082         return 0;
2083 }
2084
2085 /**
2086  * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs
2087  *
2088  * @adev: amdgpu_device pointer
2089  *
2090  * First resume function for hardware IPs.  The list of all the hardware
2091  * IPs that make up the asic is walked and the resume callbacks are run for
2092  * all blocks except COMMON, GMC, and IH.  resume puts the hardware into a
2093  * functional state after a suspend and updates the software state as
2094  * necessary.  This function is also used for restoring the GPU after a GPU
2095  * reset.
2096  * Returns 0 on success, negative error code on failure.
2097  */
2098 static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)
2099 {
2100         int i, r;
2101
2102         for (i = 0; i < adev->num_ip_blocks; i++) {
2103                 if (!adev->ip_blocks[i].status.valid)
2104                         continue;
2105                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2106                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2107                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH)
2108                         continue;
2109                 r = adev->ip_blocks[i].version->funcs->resume(adev);
2110                 if (r) {
2111                         DRM_ERROR("resume of IP block <%s> failed %d\n",
2112                                   adev->ip_blocks[i].version->funcs->name, r);
2113                         return r;
2114                 }
2115         }
2116
2117         return 0;
2118 }
2119
2120 /**
2121  * amdgpu_device_ip_resume - run resume for hardware IPs
2122  *
2123  * @adev: amdgpu_device pointer
2124  *
2125  * Main resume function for hardware IPs.  The hardware IPs
2126  * are split into two resume functions because they are
2127  * are also used in in recovering from a GPU reset and some additional
2128  * steps need to be take between them.  In this case (S3/S4) they are
2129  * run sequentially.
2130  * Returns 0 on success, negative error code on failure.
2131  */
2132 static int amdgpu_device_ip_resume(struct amdgpu_device *adev)
2133 {
2134         int r;
2135
2136         r = amdgpu_device_ip_resume_phase1(adev);
2137         if (r)
2138                 return r;
2139         r = amdgpu_device_ip_resume_phase2(adev);
2140
2141         return r;
2142 }
2143
2144 /**
2145  * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV
2146  *
2147  * @adev: amdgpu_device pointer
2148  *
2149  * Query the VBIOS data tables to determine if the board supports SR-IOV.
2150  */
2151 static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
2152 {
2153         if (amdgpu_sriov_vf(adev)) {
2154                 if (adev->is_atom_fw) {
2155                         if (amdgpu_atomfirmware_gpu_supports_virtualization(adev))
2156                                 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2157                 } else {
2158                         if (amdgpu_atombios_has_gpu_virtualization_table(adev))
2159                                 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2160                 }
2161
2162                 if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
2163                         amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
2164         }
2165 }
2166
2167 /**
2168  * amdgpu_device_asic_has_dc_support - determine if DC supports the asic
2169  *
2170  * @asic_type: AMD asic type
2171  *
2172  * Check if there is DC (new modesetting infrastructre) support for an asic.
2173  * returns true if DC has support, false if not.
2174  */
2175 bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
2176 {
2177         switch (asic_type) {
2178 #if defined(CONFIG_DRM_AMD_DC)
2179         case CHIP_BONAIRE:
2180         case CHIP_KAVERI:
2181         case CHIP_KABINI:
2182         case CHIP_MULLINS:
2183                 /*
2184                  * We have systems in the wild with these ASICs that require
2185                  * LVDS and VGA support which is not supported with DC.
2186                  *
2187                  * Fallback to the non-DC driver here by default so as not to
2188                  * cause regressions.
2189                  */
2190                 return amdgpu_dc > 0;
2191         case CHIP_HAWAII:
2192         case CHIP_CARRIZO:
2193         case CHIP_STONEY:
2194         case CHIP_POLARIS10:
2195         case CHIP_POLARIS11:
2196         case CHIP_POLARIS12:
2197         case CHIP_VEGAM:
2198         case CHIP_TONGA:
2199         case CHIP_FIJI:
2200         case CHIP_VEGA10:
2201         case CHIP_VEGA12:
2202         case CHIP_VEGA20:
2203 #ifdef CONFIG_X86
2204         case CHIP_RAVEN:
2205 #endif
2206                 return amdgpu_dc != 0;
2207 #endif
2208         default:
2209                 return false;
2210         }
2211 }
2212
2213 /**
2214  * amdgpu_device_has_dc_support - check if dc is supported
2215  *
2216  * @adev: amdgpu_device_pointer
2217  *
2218  * Returns true for supported, false for not supported
2219  */
2220 bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
2221 {
2222         if (amdgpu_sriov_vf(adev))
2223                 return false;
2224
2225         return amdgpu_device_asic_has_dc_support(adev->asic_type);
2226 }
2227
2228 /**
2229  * amdgpu_device_init - initialize the driver
2230  *
2231  * @adev: amdgpu_device pointer
2232  * @ddev: drm dev pointer
2233  * @pdev: pci dev pointer
2234  * @flags: driver flags
2235  *
2236  * Initializes the driver info and hw (all asics).
2237  * Returns 0 for success or an error on failure.
2238  * Called at driver startup.
2239  */
2240 int amdgpu_device_init(struct amdgpu_device *adev,
2241                        struct drm_device *ddev,
2242                        struct pci_dev *pdev,
2243                        uint32_t flags)
2244 {
2245         int r, i;
2246         bool runtime = false;
2247         u32 max_MBps;
2248
2249         adev->shutdown = false;
2250         adev->dev = &pdev->dev;
2251         adev->ddev = ddev;
2252         adev->pdev = pdev;
2253         adev->flags = flags;
2254         adev->asic_type = flags & AMD_ASIC_MASK;
2255         adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
2256         if (amdgpu_emu_mode == 1)
2257                 adev->usec_timeout *= 2;
2258         adev->gmc.gart_size = 512 * 1024 * 1024;
2259         adev->accel_working = false;
2260         adev->num_rings = 0;
2261         adev->mman.buffer_funcs = NULL;
2262         adev->mman.buffer_funcs_ring = NULL;
2263         adev->vm_manager.vm_pte_funcs = NULL;
2264         adev->vm_manager.vm_pte_num_rings = 0;
2265         adev->gmc.gmc_funcs = NULL;
2266         adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
2267         bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
2268
2269         adev->smc_rreg = &amdgpu_invalid_rreg;
2270         adev->smc_wreg = &amdgpu_invalid_wreg;
2271         adev->pcie_rreg = &amdgpu_invalid_rreg;
2272         adev->pcie_wreg = &amdgpu_invalid_wreg;
2273         adev->pciep_rreg = &amdgpu_invalid_rreg;
2274         adev->pciep_wreg = &amdgpu_invalid_wreg;
2275         adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
2276         adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
2277         adev->didt_rreg = &amdgpu_invalid_rreg;
2278         adev->didt_wreg = &amdgpu_invalid_wreg;
2279         adev->gc_cac_rreg = &amdgpu_invalid_rreg;
2280         adev->gc_cac_wreg = &amdgpu_invalid_wreg;
2281         adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
2282         adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
2283
2284         DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
2285                  amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
2286                  pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
2287
2288         /* mutex initialization are all done here so we
2289          * can recall function without having locking issues */
2290         atomic_set(&adev->irq.ih.lock, 0);
2291         mutex_init(&adev->firmware.mutex);
2292         mutex_init(&adev->pm.mutex);
2293         mutex_init(&adev->gfx.gpu_clock_mutex);
2294         mutex_init(&adev->srbm_mutex);
2295         mutex_init(&adev->gfx.pipe_reserve_mutex);
2296         mutex_init(&adev->grbm_idx_mutex);
2297         mutex_init(&adev->mn_lock);
2298         mutex_init(&adev->virt.vf_errors.lock);
2299         hash_init(adev->mn_hash);
2300         mutex_init(&adev->lock_reset);
2301
2302         amdgpu_device_check_arguments(adev);
2303
2304         spin_lock_init(&adev->mmio_idx_lock);
2305         spin_lock_init(&adev->smc_idx_lock);
2306         spin_lock_init(&adev->pcie_idx_lock);
2307         spin_lock_init(&adev->uvd_ctx_idx_lock);
2308         spin_lock_init(&adev->didt_idx_lock);
2309         spin_lock_init(&adev->gc_cac_idx_lock);
2310         spin_lock_init(&adev->se_cac_idx_lock);
2311         spin_lock_init(&adev->audio_endpt_idx_lock);
2312         spin_lock_init(&adev->mm_stats.lock);
2313
2314         INIT_LIST_HEAD(&adev->shadow_list);
2315         mutex_init(&adev->shadow_list_lock);
2316
2317         INIT_LIST_HEAD(&adev->ring_lru_list);
2318         spin_lock_init(&adev->ring_lru_list_lock);
2319
2320         INIT_DELAYED_WORK(&adev->late_init_work,
2321                           amdgpu_device_ip_late_init_func_handler);
2322
2323         adev->pm.ac_power = power_supply_is_system_supplied() > 0 ? true : false;
2324
2325         /* Registers mapping */
2326         /* TODO: block userspace mapping of io register */
2327         if (adev->asic_type >= CHIP_BONAIRE) {
2328                 adev->rmmio_base = pci_resource_start(adev->pdev, 5);
2329                 adev->rmmio_size = pci_resource_len(adev->pdev, 5);
2330         } else {
2331                 adev->rmmio_base = pci_resource_start(adev->pdev, 2);
2332                 adev->rmmio_size = pci_resource_len(adev->pdev, 2);
2333         }
2334
2335         adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
2336         if (adev->rmmio == NULL) {
2337                 return -ENOMEM;
2338         }
2339         DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
2340         DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
2341
2342         /* doorbell bar mapping */
2343         amdgpu_device_doorbell_init(adev);
2344
2345         /* io port mapping */
2346         for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
2347                 if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) {
2348                         adev->rio_mem_size = pci_resource_len(adev->pdev, i);
2349                         adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size);
2350                         break;
2351                 }
2352         }
2353         if (adev->rio_mem == NULL)
2354                 DRM_INFO("PCI I/O BAR is not found.\n");
2355
2356         amdgpu_device_get_pcie_info(adev);
2357
2358         /* early init functions */
2359         r = amdgpu_device_ip_early_init(adev);
2360         if (r)
2361                 return r;
2362
2363         /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
2364         /* this will fail for cards that aren't VGA class devices, just
2365          * ignore it */
2366         vga_client_register(adev->pdev, adev, NULL, amdgpu_device_vga_set_decode);
2367
2368         if (amdgpu_device_is_px(ddev))
2369                 runtime = true;
2370         if (!pci_is_thunderbolt_attached(adev->pdev))
2371                 vga_switcheroo_register_client(adev->pdev,
2372                                                &amdgpu_switcheroo_ops, runtime);
2373         if (runtime)
2374                 vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
2375
2376         if (amdgpu_emu_mode == 1) {
2377                 /* post the asic on emulation mode */
2378                 emu_soc_asic_init(adev);
2379                 goto fence_driver_init;
2380         }
2381
2382         /* Read BIOS */
2383         if (!amdgpu_get_bios(adev)) {
2384                 r = -EINVAL;
2385                 goto failed;
2386         }
2387
2388         r = amdgpu_atombios_init(adev);
2389         if (r) {
2390                 dev_err(adev->dev, "amdgpu_atombios_init failed\n");
2391                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0);
2392                 goto failed;
2393         }
2394
2395         /* detect if we are with an SRIOV vbios */
2396         amdgpu_device_detect_sriov_bios(adev);
2397
2398         /* Post card if necessary */
2399         if (amdgpu_device_need_post(adev)) {
2400                 if (!adev->bios) {
2401                         dev_err(adev->dev, "no vBIOS found\n");
2402                         r = -EINVAL;
2403                         goto failed;
2404                 }
2405                 DRM_INFO("GPU posting now...\n");
2406                 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2407                 if (r) {
2408                         dev_err(adev->dev, "gpu post error!\n");
2409                         goto failed;
2410                 }
2411         }
2412
2413         if (adev->is_atom_fw) {
2414                 /* Initialize clocks */
2415                 r = amdgpu_atomfirmware_get_clock_info(adev);
2416                 if (r) {
2417                         dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
2418                         amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
2419                         goto failed;
2420                 }
2421         } else {
2422                 /* Initialize clocks */
2423                 r = amdgpu_atombios_get_clock_info(adev);
2424                 if (r) {
2425                         dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
2426                         amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
2427                         goto failed;
2428                 }
2429                 /* init i2c buses */
2430                 if (!amdgpu_device_has_dc_support(adev))
2431                         amdgpu_atombios_i2c_init(adev);
2432         }
2433
2434 fence_driver_init:
2435         /* Fence driver */
2436         r = amdgpu_fence_driver_init(adev);
2437         if (r) {
2438                 dev_err(adev->dev, "amdgpu_fence_driver_init failed\n");
2439                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
2440                 goto failed;
2441         }
2442
2443         /* init the mode config */
2444         drm_mode_config_init(adev->ddev);
2445
2446         r = amdgpu_device_ip_init(adev);
2447         if (r) {
2448                 /* failed in exclusive mode due to timeout */
2449                 if (amdgpu_sriov_vf(adev) &&
2450                     !amdgpu_sriov_runtime(adev) &&
2451                     amdgpu_virt_mmio_blocked(adev) &&
2452                     !amdgpu_virt_wait_reset(adev)) {
2453                         dev_err(adev->dev, "VF exclusive mode timeout\n");
2454                         /* Don't send request since VF is inactive. */
2455                         adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
2456                         adev->virt.ops = NULL;
2457                         r = -EAGAIN;
2458                         goto failed;
2459                 }
2460                 dev_err(adev->dev, "amdgpu_device_ip_init failed\n");
2461                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
2462                 goto failed;
2463         }
2464
2465         adev->accel_working = true;
2466
2467         amdgpu_vm_check_compute_bug(adev);
2468
2469         /* Initialize the buffer migration limit. */
2470         if (amdgpu_moverate >= 0)
2471                 max_MBps = amdgpu_moverate;
2472         else
2473                 max_MBps = 8; /* Allow 8 MB/s. */
2474         /* Get a log2 for easy divisions. */
2475         adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
2476
2477         r = amdgpu_ib_pool_init(adev);
2478         if (r) {
2479                 dev_err(adev->dev, "IB initialization failed (%d).\n", r);
2480                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r);
2481                 goto failed;
2482         }
2483
2484         if (amdgpu_sriov_vf(adev))
2485                 amdgpu_virt_init_data_exchange(adev);
2486
2487         amdgpu_fbdev_init(adev);
2488
2489         r = amdgpu_pm_sysfs_init(adev);
2490         if (r)
2491                 DRM_ERROR("registering pm debugfs failed (%d).\n", r);
2492
2493         r = amdgpu_debugfs_gem_init(adev);
2494         if (r)
2495                 DRM_ERROR("registering gem debugfs failed (%d).\n", r);
2496
2497         r = amdgpu_debugfs_regs_init(adev);
2498         if (r)
2499                 DRM_ERROR("registering register debugfs failed (%d).\n", r);
2500
2501         r = amdgpu_debugfs_firmware_init(adev);
2502         if (r)
2503                 DRM_ERROR("registering firmware debugfs failed (%d).\n", r);
2504
2505         r = amdgpu_debugfs_init(adev);
2506         if (r)
2507                 DRM_ERROR("Creating debugfs files failed (%d).\n", r);
2508
2509         if ((amdgpu_testing & 1)) {
2510                 if (adev->accel_working)
2511                         amdgpu_test_moves(adev);
2512                 else
2513                         DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n");
2514         }
2515         if (amdgpu_benchmarking) {
2516                 if (adev->accel_working)
2517                         amdgpu_benchmark(adev, amdgpu_benchmarking);
2518                 else
2519                         DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n");
2520         }
2521
2522         /* enable clockgating, etc. after ib tests, etc. since some blocks require
2523          * explicit gating rather than handling it automatically.
2524          */
2525         r = amdgpu_device_ip_late_init(adev);
2526         if (r) {
2527                 dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n");
2528                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
2529                 goto failed;
2530         }
2531
2532         return 0;
2533
2534 failed:
2535         amdgpu_vf_error_trans_all(adev);
2536         if (runtime)
2537                 vga_switcheroo_fini_domain_pm_ops(adev->dev);
2538
2539         return r;
2540 }
2541
2542 /**
2543  * amdgpu_device_fini - tear down the driver
2544  *
2545  * @adev: amdgpu_device pointer
2546  *
2547  * Tear down the driver info (all asics).
2548  * Called at driver shutdown.
2549  */
2550 void amdgpu_device_fini(struct amdgpu_device *adev)
2551 {
2552         int r;
2553
2554         DRM_INFO("amdgpu: finishing device.\n");
2555         adev->shutdown = true;
2556         /* disable all interrupts */
2557         amdgpu_irq_disable_all(adev);
2558         if (adev->mode_info.mode_config_initialized){
2559                 if (!amdgpu_device_has_dc_support(adev))
2560                         drm_crtc_force_disable_all(adev->ddev);
2561                 else
2562                         drm_atomic_helper_shutdown(adev->ddev);
2563         }
2564         amdgpu_ib_pool_fini(adev);
2565         amdgpu_fence_driver_fini(adev);
2566         amdgpu_pm_sysfs_fini(adev);
2567         amdgpu_fbdev_fini(adev);
2568         r = amdgpu_device_ip_fini(adev);
2569         if (adev->firmware.gpu_info_fw) {
2570                 release_firmware(adev->firmware.gpu_info_fw);
2571                 adev->firmware.gpu_info_fw = NULL;
2572         }
2573         adev->accel_working = false;
2574         cancel_delayed_work_sync(&adev->late_init_work);
2575         /* free i2c buses */
2576         if (!amdgpu_device_has_dc_support(adev))
2577                 amdgpu_i2c_fini(adev);
2578
2579         if (amdgpu_emu_mode != 1)
2580                 amdgpu_atombios_fini(adev);
2581
2582         kfree(adev->bios);
2583         adev->bios = NULL;
2584         if (!pci_is_thunderbolt_attached(adev->pdev))
2585                 vga_switcheroo_unregister_client(adev->pdev);
2586         if (adev->flags & AMD_IS_PX)
2587                 vga_switcheroo_fini_domain_pm_ops(adev->dev);
2588         vga_client_register(adev->pdev, NULL, NULL, NULL);
2589         if (adev->rio_mem)
2590                 pci_iounmap(adev->pdev, adev->rio_mem);
2591         adev->rio_mem = NULL;
2592         iounmap(adev->rmmio);
2593         adev->rmmio = NULL;
2594         amdgpu_device_doorbell_fini(adev);
2595         amdgpu_debugfs_regs_cleanup(adev);
2596 }
2597
2598
2599 /*
2600  * Suspend & resume.
2601  */
2602 /**
2603  * amdgpu_device_suspend - initiate device suspend
2604  *
2605  * @dev: drm dev pointer
2606  * @suspend: suspend state
2607  * @fbcon : notify the fbdev of suspend
2608  *
2609  * Puts the hw in the suspend state (all asics).
2610  * Returns 0 for success or an error on failure.
2611  * Called at driver suspend.
2612  */
2613 int amdgpu_device_suspend(struct drm_device *dev, bool suspend, bool fbcon)
2614 {
2615         struct amdgpu_device *adev;
2616         struct drm_crtc *crtc;
2617         struct drm_connector *connector;
2618         int r;
2619
2620         if (dev == NULL || dev->dev_private == NULL) {
2621                 return -ENODEV;
2622         }
2623
2624         adev = dev->dev_private;
2625
2626         if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
2627                 return 0;
2628
2629         drm_kms_helper_poll_disable(dev);
2630
2631         if (!amdgpu_device_has_dc_support(adev)) {
2632                 /* turn off display hw */
2633                 drm_modeset_lock_all(dev);
2634                 list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2635                         drm_helper_connector_dpms(connector, DRM_MODE_DPMS_OFF);
2636                 }
2637                 drm_modeset_unlock_all(dev);
2638         }
2639
2640         amdgpu_amdkfd_suspend(adev);
2641
2642         /* unpin the front buffers and cursors */
2643         list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2644                 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2645                 struct drm_framebuffer *fb = crtc->primary->fb;
2646                 struct amdgpu_bo *robj;
2647
2648                 if (amdgpu_crtc->cursor_bo) {
2649                         struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2650                         r = amdgpu_bo_reserve(aobj, true);
2651                         if (r == 0) {
2652                                 amdgpu_bo_unpin(aobj);
2653                                 amdgpu_bo_unreserve(aobj);
2654                         }
2655                 }
2656
2657                 if (fb == NULL || fb->obj[0] == NULL) {
2658                         continue;
2659                 }
2660                 robj = gem_to_amdgpu_bo(fb->obj[0]);
2661                 /* don't unpin kernel fb objects */
2662                 if (!amdgpu_fbdev_robj_is_fb(adev, robj)) {
2663                         r = amdgpu_bo_reserve(robj, true);
2664                         if (r == 0) {
2665                                 amdgpu_bo_unpin(robj);
2666                                 amdgpu_bo_unreserve(robj);
2667                         }
2668                 }
2669         }
2670         /* evict vram memory */
2671         amdgpu_bo_evict_vram(adev);
2672
2673         amdgpu_fence_driver_suspend(adev);
2674
2675         r = amdgpu_device_ip_suspend(adev);
2676
2677         /* evict remaining vram memory
2678          * This second call to evict vram is to evict the gart page table
2679          * using the CPU.
2680          */
2681         amdgpu_bo_evict_vram(adev);
2682
2683         pci_save_state(dev->pdev);
2684         if (suspend) {
2685                 /* Shut down the device */
2686                 pci_disable_device(dev->pdev);
2687                 pci_set_power_state(dev->pdev, PCI_D3hot);
2688         } else {
2689                 r = amdgpu_asic_reset(adev);
2690                 if (r)
2691                         DRM_ERROR("amdgpu asic reset failed\n");
2692         }
2693
2694         if (fbcon) {
2695                 console_lock();
2696                 amdgpu_fbdev_set_suspend(adev, 1);
2697                 console_unlock();
2698         }
2699         return 0;
2700 }
2701
2702 /**
2703  * amdgpu_device_resume - initiate device resume
2704  *
2705  * @dev: drm dev pointer
2706  * @resume: resume state
2707  * @fbcon : notify the fbdev of resume
2708  *
2709  * Bring the hw back to operating state (all asics).
2710  * Returns 0 for success or an error on failure.
2711  * Called at driver resume.
2712  */
2713 int amdgpu_device_resume(struct drm_device *dev, bool resume, bool fbcon)
2714 {
2715         struct drm_connector *connector;
2716         struct amdgpu_device *adev = dev->dev_private;
2717         struct drm_crtc *crtc;
2718         int r = 0;
2719
2720         if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
2721                 return 0;
2722
2723         if (fbcon)
2724                 console_lock();
2725
2726         if (resume) {
2727                 pci_set_power_state(dev->pdev, PCI_D0);
2728                 pci_restore_state(dev->pdev);
2729                 r = pci_enable_device(dev->pdev);
2730                 if (r)
2731                         goto unlock;
2732         }
2733
2734         /* post card */
2735         if (amdgpu_device_need_post(adev)) {
2736                 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2737                 if (r)
2738                         DRM_ERROR("amdgpu asic init failed\n");
2739         }
2740
2741         r = amdgpu_device_ip_resume(adev);
2742         if (r) {
2743                 DRM_ERROR("amdgpu_device_ip_resume failed (%d).\n", r);
2744                 goto unlock;
2745         }
2746         amdgpu_fence_driver_resume(adev);
2747
2748
2749         r = amdgpu_device_ip_late_init(adev);
2750         if (r)
2751                 goto unlock;
2752
2753         /* pin cursors */
2754         list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2755                 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2756
2757                 if (amdgpu_crtc->cursor_bo) {
2758                         struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2759                         r = amdgpu_bo_reserve(aobj, true);
2760                         if (r == 0) {
2761                                 r = amdgpu_bo_pin(aobj, AMDGPU_GEM_DOMAIN_VRAM);
2762                                 if (r != 0)
2763                                         DRM_ERROR("Failed to pin cursor BO (%d)\n", r);
2764                                 amdgpu_crtc->cursor_addr = amdgpu_bo_gpu_offset(aobj);
2765                                 amdgpu_bo_unreserve(aobj);
2766                         }
2767                 }
2768         }
2769         r = amdgpu_amdkfd_resume(adev);
2770         if (r)
2771                 return r;
2772
2773         /* blat the mode back in */
2774         if (fbcon) {
2775                 if (!amdgpu_device_has_dc_support(adev)) {
2776                         /* pre DCE11 */
2777                         drm_helper_resume_force_mode(dev);
2778
2779                         /* turn on display hw */
2780                         drm_modeset_lock_all(dev);
2781                         list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2782                                 drm_helper_connector_dpms(connector, DRM_MODE_DPMS_ON);
2783                         }
2784                         drm_modeset_unlock_all(dev);
2785                 }
2786         }
2787
2788         drm_kms_helper_poll_enable(dev);
2789
2790         /*
2791          * Most of the connector probing functions try to acquire runtime pm
2792          * refs to ensure that the GPU is powered on when connector polling is
2793          * performed. Since we're calling this from a runtime PM callback,
2794          * trying to acquire rpm refs will cause us to deadlock.
2795          *
2796          * Since we're guaranteed to be holding the rpm lock, it's safe to
2797          * temporarily disable the rpm helpers so this doesn't deadlock us.
2798          */
2799 #ifdef CONFIG_PM
2800         dev->dev->power.disable_depth++;
2801 #endif
2802         if (!amdgpu_device_has_dc_support(adev))
2803                 drm_helper_hpd_irq_event(dev);
2804         else
2805                 drm_kms_helper_hotplug_event(dev);
2806 #ifdef CONFIG_PM
2807         dev->dev->power.disable_depth--;
2808 #endif
2809
2810         if (fbcon)
2811                 amdgpu_fbdev_set_suspend(adev, 0);
2812
2813 unlock:
2814         if (fbcon)
2815                 console_unlock();
2816
2817         return r;
2818 }
2819
2820 /**
2821  * amdgpu_device_ip_check_soft_reset - did soft reset succeed
2822  *
2823  * @adev: amdgpu_device pointer
2824  *
2825  * The list of all the hardware IPs that make up the asic is walked and
2826  * the check_soft_reset callbacks are run.  check_soft_reset determines
2827  * if the asic is still hung or not.
2828  * Returns true if any of the IPs are still in a hung state, false if not.
2829  */
2830 static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)
2831 {
2832         int i;
2833         bool asic_hang = false;
2834
2835         if (amdgpu_sriov_vf(adev))
2836                 return true;
2837
2838         if (amdgpu_asic_need_full_reset(adev))
2839                 return true;
2840
2841         for (i = 0; i < adev->num_ip_blocks; i++) {
2842                 if (!adev->ip_blocks[i].status.valid)
2843                         continue;
2844                 if (adev->ip_blocks[i].version->funcs->check_soft_reset)
2845                         adev->ip_blocks[i].status.hang =
2846                                 adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
2847                 if (adev->ip_blocks[i].status.hang) {
2848                         DRM_INFO("IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
2849                         asic_hang = true;
2850                 }
2851         }
2852         return asic_hang;
2853 }
2854
2855 /**
2856  * amdgpu_device_ip_pre_soft_reset - prepare for soft reset
2857  *
2858  * @adev: amdgpu_device pointer
2859  *
2860  * The list of all the hardware IPs that make up the asic is walked and the
2861  * pre_soft_reset callbacks are run if the block is hung.  pre_soft_reset
2862  * handles any IP specific hardware or software state changes that are
2863  * necessary for a soft reset to succeed.
2864  * Returns 0 on success, negative error code on failure.
2865  */
2866 static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)
2867 {
2868         int i, r = 0;
2869
2870         for (i = 0; i < adev->num_ip_blocks; i++) {
2871                 if (!adev->ip_blocks[i].status.valid)
2872                         continue;
2873                 if (adev->ip_blocks[i].status.hang &&
2874                     adev->ip_blocks[i].version->funcs->pre_soft_reset) {
2875                         r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
2876                         if (r)
2877                                 return r;
2878                 }
2879         }
2880
2881         return 0;
2882 }
2883
2884 /**
2885  * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed
2886  *
2887  * @adev: amdgpu_device pointer
2888  *
2889  * Some hardware IPs cannot be soft reset.  If they are hung, a full gpu
2890  * reset is necessary to recover.
2891  * Returns true if a full asic reset is required, false if not.
2892  */
2893 static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)
2894 {
2895         int i;
2896
2897         if (amdgpu_asic_need_full_reset(adev))
2898                 return true;
2899
2900         for (i = 0; i < adev->num_ip_blocks; i++) {
2901                 if (!adev->ip_blocks[i].status.valid)
2902                         continue;
2903                 if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
2904                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
2905                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
2906                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
2907                      adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
2908                         if (adev->ip_blocks[i].status.hang) {
2909                                 DRM_INFO("Some block need full reset!\n");
2910                                 return true;
2911                         }
2912                 }
2913         }
2914         return false;
2915 }
2916
2917 /**
2918  * amdgpu_device_ip_soft_reset - do a soft reset
2919  *
2920  * @adev: amdgpu_device pointer
2921  *
2922  * The list of all the hardware IPs that make up the asic is walked and the
2923  * soft_reset callbacks are run if the block is hung.  soft_reset handles any
2924  * IP specific hardware or software state changes that are necessary to soft
2925  * reset the IP.
2926  * Returns 0 on success, negative error code on failure.
2927  */
2928 static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)
2929 {
2930         int i, r = 0;
2931
2932         for (i = 0; i < adev->num_ip_blocks; i++) {
2933                 if (!adev->ip_blocks[i].status.valid)
2934                         continue;
2935                 if (adev->ip_blocks[i].status.hang &&
2936                     adev->ip_blocks[i].version->funcs->soft_reset) {
2937                         r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
2938                         if (r)
2939                                 return r;
2940                 }
2941         }
2942
2943         return 0;
2944 }
2945
2946 /**
2947  * amdgpu_device_ip_post_soft_reset - clean up from soft reset
2948  *
2949  * @adev: amdgpu_device pointer
2950  *
2951  * The list of all the hardware IPs that make up the asic is walked and the
2952  * post_soft_reset callbacks are run if the asic was hung.  post_soft_reset
2953  * handles any IP specific hardware or software state changes that are
2954  * necessary after the IP has been soft reset.
2955  * Returns 0 on success, negative error code on failure.
2956  */
2957 static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)
2958 {
2959         int i, r = 0;
2960
2961         for (i = 0; i < adev->num_ip_blocks; i++) {
2962                 if (!adev->ip_blocks[i].status.valid)
2963                         continue;
2964                 if (adev->ip_blocks[i].status.hang &&
2965                     adev->ip_blocks[i].version->funcs->post_soft_reset)
2966                         r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
2967                 if (r)
2968                         return r;
2969         }
2970
2971         return 0;
2972 }
2973
2974 /**
2975  * amdgpu_device_recover_vram_from_shadow - restore shadowed VRAM buffers
2976  *
2977  * @adev: amdgpu_device pointer
2978  * @ring: amdgpu_ring for the engine handling the buffer operations
2979  * @bo: amdgpu_bo buffer whose shadow is being restored
2980  * @fence: dma_fence associated with the operation
2981  *
2982  * Restores the VRAM buffer contents from the shadow in GTT.  Used to
2983  * restore things like GPUVM page tables after a GPU reset where
2984  * the contents of VRAM might be lost.
2985  * Returns 0 on success, negative error code on failure.
2986  */
2987 static int amdgpu_device_recover_vram_from_shadow(struct amdgpu_device *adev,
2988                                                   struct amdgpu_ring *ring,
2989                                                   struct amdgpu_bo *bo,
2990                                                   struct dma_fence **fence)
2991 {
2992         uint32_t domain;
2993         int r;
2994
2995         if (!bo->shadow)
2996                 return 0;
2997
2998         r = amdgpu_bo_reserve(bo, true);
2999         if (r)
3000                 return r;
3001         domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
3002         /* if bo has been evicted, then no need to recover */
3003         if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
3004                 r = amdgpu_bo_validate(bo->shadow);
3005                 if (r) {
3006                         DRM_ERROR("bo validate failed!\n");
3007                         goto err;
3008                 }
3009
3010                 r = amdgpu_bo_restore_from_shadow(adev, ring, bo,
3011                                                  NULL, fence, true);
3012                 if (r) {
3013                         DRM_ERROR("recover page table failed!\n");
3014                         goto err;
3015                 }
3016         }
3017 err:
3018         amdgpu_bo_unreserve(bo);
3019         return r;
3020 }
3021
3022 /**
3023  * amdgpu_device_handle_vram_lost - Handle the loss of VRAM contents
3024  *
3025  * @adev: amdgpu_device pointer
3026  *
3027  * Restores the contents of VRAM buffers from the shadows in GTT.  Used to
3028  * restore things like GPUVM page tables after a GPU reset where
3029  * the contents of VRAM might be lost.
3030  * Returns 0 on success, 1 on failure.
3031  */
3032 static int amdgpu_device_handle_vram_lost(struct amdgpu_device *adev)
3033 {
3034         struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
3035         struct amdgpu_bo *bo, *tmp;
3036         struct dma_fence *fence = NULL, *next = NULL;
3037         long r = 1;
3038         int i = 0;
3039         long tmo;
3040
3041         if (amdgpu_sriov_runtime(adev))
3042                 tmo = msecs_to_jiffies(amdgpu_lockup_timeout);
3043         else
3044                 tmo = msecs_to_jiffies(100);
3045
3046         DRM_INFO("recover vram bo from shadow start\n");
3047         mutex_lock(&adev->shadow_list_lock);
3048         list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) {
3049                 next = NULL;
3050                 amdgpu_device_recover_vram_from_shadow(adev, ring, bo, &next);
3051                 if (fence) {
3052                         r = dma_fence_wait_timeout(fence, false, tmo);
3053                         if (r == 0)
3054                                 pr_err("wait fence %p[%d] timeout\n", fence, i);
3055                         else if (r < 0)
3056                                 pr_err("wait fence %p[%d] interrupted\n", fence, i);
3057                         if (r < 1) {
3058                                 dma_fence_put(fence);
3059                                 fence = next;
3060                                 break;
3061                         }
3062                         i++;
3063                 }
3064
3065                 dma_fence_put(fence);
3066                 fence = next;
3067         }
3068         mutex_unlock(&adev->shadow_list_lock);
3069
3070         if (fence) {
3071                 r = dma_fence_wait_timeout(fence, false, tmo);
3072                 if (r == 0)
3073                         pr_err("wait fence %p[%d] timeout\n", fence, i);
3074                 else if (r < 0)
3075                         pr_err("wait fence %p[%d] interrupted\n", fence, i);
3076
3077         }
3078         dma_fence_put(fence);
3079
3080         if (r > 0)
3081                 DRM_INFO("recover vram bo from shadow done\n");
3082         else
3083                 DRM_ERROR("recover vram bo from shadow failed\n");
3084
3085         return (r > 0) ? 0 : 1;
3086 }
3087
3088 /**
3089  * amdgpu_device_reset - reset ASIC/GPU for bare-metal or passthrough
3090  *
3091  * @adev: amdgpu device pointer
3092  *
3093  * attempt to do soft-reset or full-reset and reinitialize Asic
3094  * return 0 means successed otherwise failed
3095  */
3096 static int amdgpu_device_reset(struct amdgpu_device *adev)
3097 {
3098         bool need_full_reset, vram_lost = 0;
3099         int r;
3100
3101         need_full_reset = amdgpu_device_ip_need_full_reset(adev);
3102
3103         if (!need_full_reset) {
3104                 amdgpu_device_ip_pre_soft_reset(adev);
3105                 r = amdgpu_device_ip_soft_reset(adev);
3106                 amdgpu_device_ip_post_soft_reset(adev);
3107                 if (r || amdgpu_device_ip_check_soft_reset(adev)) {
3108                         DRM_INFO("soft reset failed, will fallback to full reset!\n");
3109                         need_full_reset = true;
3110                 }
3111         }
3112
3113         if (need_full_reset) {
3114                 r = amdgpu_device_ip_suspend(adev);
3115
3116 retry:
3117                 r = amdgpu_asic_reset(adev);
3118                 /* post card */
3119                 amdgpu_atom_asic_init(adev->mode_info.atom_context);
3120
3121                 if (!r) {
3122                         dev_info(adev->dev, "GPU reset succeeded, trying to resume\n");
3123                         r = amdgpu_device_ip_resume_phase1(adev);
3124                         if (r)
3125                                 goto out;
3126
3127                         vram_lost = amdgpu_device_check_vram_lost(adev);
3128                         if (vram_lost) {
3129                                 DRM_ERROR("VRAM is lost!\n");
3130                                 atomic_inc(&adev->vram_lost_counter);
3131                         }
3132
3133                         r = amdgpu_gtt_mgr_recover(
3134                                 &adev->mman.bdev.man[TTM_PL_TT]);
3135                         if (r)
3136                                 goto out;
3137
3138                         r = amdgpu_device_ip_resume_phase2(adev);
3139                         if (r)
3140                                 goto out;
3141
3142                         if (vram_lost)
3143                                 amdgpu_device_fill_reset_magic(adev);
3144                 }
3145         }
3146
3147 out:
3148         if (!r) {
3149                 amdgpu_irq_gpu_reset_resume_helper(adev);
3150                 r = amdgpu_ib_ring_tests(adev);
3151                 if (r) {
3152                         dev_err(adev->dev, "ib ring test failed (%d).\n", r);
3153                         r = amdgpu_device_ip_suspend(adev);
3154                         need_full_reset = true;
3155                         goto retry;
3156                 }
3157         }
3158
3159         if (!r && ((need_full_reset && !(adev->flags & AMD_IS_APU)) || vram_lost))
3160                 r = amdgpu_device_handle_vram_lost(adev);
3161
3162         return r;
3163 }
3164
3165 /**
3166  * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
3167  *
3168  * @adev: amdgpu device pointer
3169  * @from_hypervisor: request from hypervisor
3170  *
3171  * do VF FLR and reinitialize Asic
3172  * return 0 means successed otherwise failed
3173  */
3174 static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
3175                                      bool from_hypervisor)
3176 {
3177         int r;
3178
3179         if (from_hypervisor)
3180                 r = amdgpu_virt_request_full_gpu(adev, true);
3181         else
3182                 r = amdgpu_virt_reset_gpu(adev);
3183         if (r)
3184                 return r;
3185
3186         /* Resume IP prior to SMC */
3187         r = amdgpu_device_ip_reinit_early_sriov(adev);
3188         if (r)
3189                 goto error;
3190
3191         /* we need recover gart prior to run SMC/CP/SDMA resume */
3192         amdgpu_gtt_mgr_recover(&adev->mman.bdev.man[TTM_PL_TT]);
3193
3194         /* now we are okay to resume SMC/CP/SDMA */
3195         r = amdgpu_device_ip_reinit_late_sriov(adev);
3196         if (r)
3197                 goto error;
3198
3199         amdgpu_irq_gpu_reset_resume_helper(adev);
3200         r = amdgpu_ib_ring_tests(adev);
3201
3202 error:
3203         amdgpu_virt_release_full_gpu(adev, true);
3204         if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
3205                 atomic_inc(&adev->vram_lost_counter);
3206                 r = amdgpu_device_handle_vram_lost(adev);
3207         }
3208
3209         return r;
3210 }
3211
3212 /**
3213  * amdgpu_device_gpu_recover - reset the asic and recover scheduler
3214  *
3215  * @adev: amdgpu device pointer
3216  * @job: which job trigger hang
3217  * @force: forces reset regardless of amdgpu_gpu_recovery
3218  *
3219  * Attempt to reset the GPU if it has hung (all asics).
3220  * Returns 0 for success or an error on failure.
3221  */
3222 int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
3223                               struct amdgpu_job *job, bool force)
3224 {
3225         int i, r, resched;
3226
3227         if (!force && !amdgpu_device_ip_check_soft_reset(adev)) {
3228                 DRM_INFO("No hardware hang detected. Did some blocks stall?\n");
3229                 return 0;
3230         }
3231
3232         if (!force && (amdgpu_gpu_recovery == 0 ||
3233                         (amdgpu_gpu_recovery == -1  && !amdgpu_sriov_vf(adev)))) {
3234                 DRM_INFO("GPU recovery disabled.\n");
3235                 return 0;
3236         }
3237
3238         dev_info(adev->dev, "GPU reset begin!\n");
3239
3240         mutex_lock(&adev->lock_reset);
3241         atomic_inc(&adev->gpu_reset_counter);
3242         adev->in_gpu_reset = 1;
3243
3244         /* block TTM */
3245         resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
3246
3247         /* block all schedulers and reset given job's ring */
3248         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3249                 struct amdgpu_ring *ring = adev->rings[i];
3250
3251                 if (!ring || !ring->sched.thread)
3252                         continue;
3253
3254                 kthread_park(ring->sched.thread);
3255
3256                 if (job && job->base.sched == &ring->sched)
3257                         continue;
3258
3259                 drm_sched_hw_job_reset(&ring->sched, &job->base);
3260
3261                 /* after all hw jobs are reset, hw fence is meaningless, so force_completion */
3262                 amdgpu_fence_driver_force_completion(ring);
3263         }
3264
3265         if (amdgpu_sriov_vf(adev))
3266                 r = amdgpu_device_reset_sriov(adev, job ? false : true);
3267         else
3268                 r = amdgpu_device_reset(adev);
3269
3270         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3271                 struct amdgpu_ring *ring = adev->rings[i];
3272
3273                 if (!ring || !ring->sched.thread)
3274                         continue;
3275
3276                 /* only need recovery sched of the given job's ring
3277                  * or all rings (in the case @job is NULL)
3278                  * after above amdgpu_reset accomplished
3279                  */
3280                 if ((!job || job->base.sched == &ring->sched) && !r)
3281                         drm_sched_job_recovery(&ring->sched);
3282
3283                 kthread_unpark(ring->sched.thread);
3284         }
3285
3286         if (!amdgpu_device_has_dc_support(adev)) {
3287                 drm_helper_resume_force_mode(adev->ddev);
3288         }
3289
3290         ttm_bo_unlock_delayed_workqueue(&adev->mman.bdev, resched);
3291
3292         if (r) {
3293                 /* bad news, how to tell it to userspace ? */
3294                 dev_info(adev->dev, "GPU reset(%d) failed\n", atomic_read(&adev->gpu_reset_counter));
3295                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r);
3296         } else {
3297                 dev_info(adev->dev, "GPU reset(%d) successed!\n",atomic_read(&adev->gpu_reset_counter));
3298         }
3299
3300         amdgpu_vf_error_trans_all(adev);
3301         adev->in_gpu_reset = 0;
3302         mutex_unlock(&adev->lock_reset);
3303         return r;
3304 }
3305
3306 /**
3307  * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot
3308  *
3309  * @adev: amdgpu_device pointer
3310  *
3311  * Fetchs and stores in the driver the PCIE capabilities (gen speed
3312  * and lanes) of the slot the device is in. Handles APUs and
3313  * virtualized environments where PCIE config space may not be available.
3314  */
3315 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev)
3316 {
3317         struct pci_dev *pdev;
3318         enum pci_bus_speed speed_cap;
3319         enum pcie_link_width link_width;
3320
3321         if (amdgpu_pcie_gen_cap)
3322                 adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
3323
3324         if (amdgpu_pcie_lane_cap)
3325                 adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
3326
3327         /* covers APUs as well */
3328         if (pci_is_root_bus(adev->pdev->bus)) {
3329                 if (adev->pm.pcie_gen_mask == 0)
3330                         adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
3331                 if (adev->pm.pcie_mlw_mask == 0)
3332                         adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
3333                 return;
3334         }
3335
3336         if (adev->pm.pcie_gen_mask == 0) {
3337                 /* asic caps */
3338                 pdev = adev->pdev;
3339                 speed_cap = pcie_get_speed_cap(pdev);
3340                 if (speed_cap == PCI_SPEED_UNKNOWN) {
3341                         adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3342                                                   CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3343                                                   CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
3344                 } else {
3345                         if (speed_cap == PCIE_SPEED_16_0GT)
3346                                 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3347                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3348                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
3349                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4);
3350                         else if (speed_cap == PCIE_SPEED_8_0GT)
3351                                 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3352                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3353                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
3354                         else if (speed_cap == PCIE_SPEED_5_0GT)
3355                                 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3356                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2);
3357                         else
3358                                 adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1;
3359                 }
3360                 /* platform caps */
3361                 pdev = adev->ddev->pdev->bus->self;
3362                 speed_cap = pcie_get_speed_cap(pdev);
3363                 if (speed_cap == PCI_SPEED_UNKNOWN) {
3364                         adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3365                                                    CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
3366                 } else {
3367                         if (speed_cap == PCIE_SPEED_16_0GT)
3368                                 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3369                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3370                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
3371                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4);
3372                         else if (speed_cap == PCIE_SPEED_8_0GT)
3373                                 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3374                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3375                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3);
3376                         else if (speed_cap == PCIE_SPEED_5_0GT)
3377                                 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3378                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
3379                         else
3380                                 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
3381
3382                 }
3383         }
3384         if (adev->pm.pcie_mlw_mask == 0) {
3385                 pdev = adev->ddev->pdev->bus->self;
3386                 link_width = pcie_get_width_cap(pdev);
3387                 if (link_width == PCIE_LNK_WIDTH_UNKNOWN) {
3388                         adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK;
3389                 } else {
3390                         switch (link_width) {
3391                         case PCIE_LNK_X32:
3392                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
3393                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
3394                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3395                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3396                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3397                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3398                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3399                                 break;
3400                         case PCIE_LNK_X16:
3401                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
3402                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3403                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3404                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3405                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3406                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3407                                 break;
3408                         case PCIE_LNK_X12:
3409                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3410                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3411                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3412                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3413                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3414                                 break;
3415                         case PCIE_LNK_X8:
3416                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3417                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3418                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3419                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3420                                 break;
3421                         case PCIE_LNK_X4:
3422                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3423                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3424                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3425                                 break;
3426                         case PCIE_LNK_X2:
3427                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3428                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3429                                 break;
3430                         case PCIE_LNK_X1:
3431                                 adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;
3432                                 break;
3433                         default:
3434                                 break;
3435                         }
3436                 }
3437         }
3438 }
3439
This page took 0.245477 seconds and 4 git commands to generate.