]> Git Repo - linux.git/blob - drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
drm/amdgpu: move amdgpu_device_(vram|gtt)_location
[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_resize_fb_bar - try to resize FB BAR
656  *
657  * @adev: amdgpu_device pointer
658  *
659  * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not
660  * to fail, but if any of the BARs is not accessible after the size we abort
661  * driver loading by returning -ENODEV.
662  */
663 int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev)
664 {
665         u64 space_needed = roundup_pow_of_two(adev->gmc.real_vram_size);
666         u32 rbar_size = order_base_2(((space_needed >> 20) | 1)) - 1;
667         struct pci_bus *root;
668         struct resource *res;
669         unsigned i;
670         u16 cmd;
671         int r;
672
673         /* Bypass for VF */
674         if (amdgpu_sriov_vf(adev))
675                 return 0;
676
677         /* Check if the root BUS has 64bit memory resources */
678         root = adev->pdev->bus;
679         while (root->parent)
680                 root = root->parent;
681
682         pci_bus_for_each_resource(root, res, i) {
683                 if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) &&
684                     res->start > 0x100000000ull)
685                         break;
686         }
687
688         /* Trying to resize is pointless without a root hub window above 4GB */
689         if (!res)
690                 return 0;
691
692         /* Disable memory decoding while we change the BAR addresses and size */
693         pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd);
694         pci_write_config_word(adev->pdev, PCI_COMMAND,
695                               cmd & ~PCI_COMMAND_MEMORY);
696
697         /* Free the VRAM and doorbell BAR, we most likely need to move both. */
698         amdgpu_device_doorbell_fini(adev);
699         if (adev->asic_type >= CHIP_BONAIRE)
700                 pci_release_resource(adev->pdev, 2);
701
702         pci_release_resource(adev->pdev, 0);
703
704         r = pci_resize_resource(adev->pdev, 0, rbar_size);
705         if (r == -ENOSPC)
706                 DRM_INFO("Not enough PCI address space for a large BAR.");
707         else if (r && r != -ENOTSUPP)
708                 DRM_ERROR("Problem resizing BAR0 (%d).", r);
709
710         pci_assign_unassigned_bus_resources(adev->pdev->bus);
711
712         /* When the doorbell or fb BAR isn't available we have no chance of
713          * using the device.
714          */
715         r = amdgpu_device_doorbell_init(adev);
716         if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET))
717                 return -ENODEV;
718
719         pci_write_config_word(adev->pdev, PCI_COMMAND, cmd);
720
721         return 0;
722 }
723
724 /*
725  * GPU helpers function.
726  */
727 /**
728  * amdgpu_device_need_post - check if the hw need post or not
729  *
730  * @adev: amdgpu_device pointer
731  *
732  * Check if the asic has been initialized (all asics) at driver startup
733  * or post is needed if  hw reset is performed.
734  * Returns true if need or false if not.
735  */
736 bool amdgpu_device_need_post(struct amdgpu_device *adev)
737 {
738         uint32_t reg;
739
740         if (amdgpu_sriov_vf(adev))
741                 return false;
742
743         if (amdgpu_passthrough(adev)) {
744                 /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
745                  * some old smc fw still need driver do vPost otherwise gpu hang, while
746                  * those smc fw version above 22.15 doesn't have this flaw, so we force
747                  * vpost executed for smc version below 22.15
748                  */
749                 if (adev->asic_type == CHIP_FIJI) {
750                         int err;
751                         uint32_t fw_ver;
752                         err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
753                         /* force vPost if error occured */
754                         if (err)
755                                 return true;
756
757                         fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
758                         if (fw_ver < 0x00160e00)
759                                 return true;
760                 }
761         }
762
763         if (adev->has_hw_reset) {
764                 adev->has_hw_reset = false;
765                 return true;
766         }
767
768         /* bios scratch used on CIK+ */
769         if (adev->asic_type >= CHIP_BONAIRE)
770                 return amdgpu_atombios_scratch_need_asic_init(adev);
771
772         /* check MEM_SIZE for older asics */
773         reg = amdgpu_asic_get_config_memsize(adev);
774
775         if ((reg != 0) && (reg != 0xffffffff))
776                 return false;
777
778         return true;
779 }
780
781 /* if we get transitioned to only one device, take VGA back */
782 /**
783  * amdgpu_device_vga_set_decode - enable/disable vga decode
784  *
785  * @cookie: amdgpu_device pointer
786  * @state: enable/disable vga decode
787  *
788  * Enable/disable vga decode (all asics).
789  * Returns VGA resource flags.
790  */
791 static unsigned int amdgpu_device_vga_set_decode(void *cookie, bool state)
792 {
793         struct amdgpu_device *adev = cookie;
794         amdgpu_asic_set_vga_state(adev, state);
795         if (state)
796                 return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
797                        VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
798         else
799                 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
800 }
801
802 /**
803  * amdgpu_device_check_block_size - validate the vm block size
804  *
805  * @adev: amdgpu_device pointer
806  *
807  * Validates the vm block size specified via module parameter.
808  * The vm block size defines number of bits in page table versus page directory,
809  * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
810  * page table and the remaining bits are in the page directory.
811  */
812 static void amdgpu_device_check_block_size(struct amdgpu_device *adev)
813 {
814         /* defines number of bits in page table versus page directory,
815          * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
816          * page table and the remaining bits are in the page directory */
817         if (amdgpu_vm_block_size == -1)
818                 return;
819
820         if (amdgpu_vm_block_size < 9) {
821                 dev_warn(adev->dev, "VM page table size (%d) too small\n",
822                          amdgpu_vm_block_size);
823                 amdgpu_vm_block_size = -1;
824         }
825 }
826
827 /**
828  * amdgpu_device_check_vm_size - validate the vm size
829  *
830  * @adev: amdgpu_device pointer
831  *
832  * Validates the vm size in GB specified via module parameter.
833  * The VM size is the size of the GPU virtual memory space in GB.
834  */
835 static void amdgpu_device_check_vm_size(struct amdgpu_device *adev)
836 {
837         /* no need to check the default value */
838         if (amdgpu_vm_size == -1)
839                 return;
840
841         if (amdgpu_vm_size < 1) {
842                 dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
843                          amdgpu_vm_size);
844                 amdgpu_vm_size = -1;
845         }
846 }
847
848 static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev)
849 {
850         struct sysinfo si;
851         bool is_os_64 = (sizeof(void *) == 8) ? true : false;
852         uint64_t total_memory;
853         uint64_t dram_size_seven_GB = 0x1B8000000;
854         uint64_t dram_size_three_GB = 0xB8000000;
855
856         if (amdgpu_smu_memory_pool_size == 0)
857                 return;
858
859         if (!is_os_64) {
860                 DRM_WARN("Not 64-bit OS, feature not supported\n");
861                 goto def_value;
862         }
863         si_meminfo(&si);
864         total_memory = (uint64_t)si.totalram * si.mem_unit;
865
866         if ((amdgpu_smu_memory_pool_size == 1) ||
867                 (amdgpu_smu_memory_pool_size == 2)) {
868                 if (total_memory < dram_size_three_GB)
869                         goto def_value1;
870         } else if ((amdgpu_smu_memory_pool_size == 4) ||
871                 (amdgpu_smu_memory_pool_size == 8)) {
872                 if (total_memory < dram_size_seven_GB)
873                         goto def_value1;
874         } else {
875                 DRM_WARN("Smu memory pool size not supported\n");
876                 goto def_value;
877         }
878         adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28;
879
880         return;
881
882 def_value1:
883         DRM_WARN("No enough system memory\n");
884 def_value:
885         adev->pm.smu_prv_buffer_size = 0;
886 }
887
888 /**
889  * amdgpu_device_check_arguments - validate module params
890  *
891  * @adev: amdgpu_device pointer
892  *
893  * Validates certain module parameters and updates
894  * the associated values used by the driver (all asics).
895  */
896 static void amdgpu_device_check_arguments(struct amdgpu_device *adev)
897 {
898         if (amdgpu_sched_jobs < 4) {
899                 dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
900                          amdgpu_sched_jobs);
901                 amdgpu_sched_jobs = 4;
902         } else if (!is_power_of_2(amdgpu_sched_jobs)){
903                 dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
904                          amdgpu_sched_jobs);
905                 amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
906         }
907
908         if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) {
909                 /* gart size must be greater or equal to 32M */
910                 dev_warn(adev->dev, "gart size (%d) too small\n",
911                          amdgpu_gart_size);
912                 amdgpu_gart_size = -1;
913         }
914
915         if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) {
916                 /* gtt size must be greater or equal to 32M */
917                 dev_warn(adev->dev, "gtt size (%d) too small\n",
918                                  amdgpu_gtt_size);
919                 amdgpu_gtt_size = -1;
920         }
921
922         /* valid range is between 4 and 9 inclusive */
923         if (amdgpu_vm_fragment_size != -1 &&
924             (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) {
925                 dev_warn(adev->dev, "valid range is between 4 and 9\n");
926                 amdgpu_vm_fragment_size = -1;
927         }
928
929         amdgpu_device_check_smu_prv_buffer_size(adev);
930
931         amdgpu_device_check_vm_size(adev);
932
933         amdgpu_device_check_block_size(adev);
934
935         if (amdgpu_vram_page_split != -1 && (amdgpu_vram_page_split < 16 ||
936             !is_power_of_2(amdgpu_vram_page_split))) {
937                 dev_warn(adev->dev, "invalid VRAM page split (%d)\n",
938                          amdgpu_vram_page_split);
939                 amdgpu_vram_page_split = 1024;
940         }
941
942         if (amdgpu_lockup_timeout == 0) {
943                 dev_warn(adev->dev, "lockup_timeout msut be > 0, adjusting to 10000\n");
944                 amdgpu_lockup_timeout = 10000;
945         }
946
947         adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type);
948 }
949
950 /**
951  * amdgpu_switcheroo_set_state - set switcheroo state
952  *
953  * @pdev: pci dev pointer
954  * @state: vga_switcheroo state
955  *
956  * Callback for the switcheroo driver.  Suspends or resumes the
957  * the asics before or after it is powered up using ACPI methods.
958  */
959 static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, enum vga_switcheroo_state state)
960 {
961         struct drm_device *dev = pci_get_drvdata(pdev);
962
963         if (amdgpu_device_is_px(dev) && state == VGA_SWITCHEROO_OFF)
964                 return;
965
966         if (state == VGA_SWITCHEROO_ON) {
967                 pr_info("amdgpu: switched on\n");
968                 /* don't suspend or resume card normally */
969                 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
970
971                 amdgpu_device_resume(dev, true, true);
972
973                 dev->switch_power_state = DRM_SWITCH_POWER_ON;
974                 drm_kms_helper_poll_enable(dev);
975         } else {
976                 pr_info("amdgpu: switched off\n");
977                 drm_kms_helper_poll_disable(dev);
978                 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
979                 amdgpu_device_suspend(dev, true, true);
980                 dev->switch_power_state = DRM_SWITCH_POWER_OFF;
981         }
982 }
983
984 /**
985  * amdgpu_switcheroo_can_switch - see if switcheroo state can change
986  *
987  * @pdev: pci dev pointer
988  *
989  * Callback for the switcheroo driver.  Check of the switcheroo
990  * state can be changed.
991  * Returns true if the state can be changed, false if not.
992  */
993 static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
994 {
995         struct drm_device *dev = pci_get_drvdata(pdev);
996
997         /*
998         * FIXME: open_count is protected by drm_global_mutex but that would lead to
999         * locking inversion with the driver load path. And the access here is
1000         * completely racy anyway. So don't bother with locking for now.
1001         */
1002         return dev->open_count == 0;
1003 }
1004
1005 static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
1006         .set_gpu_state = amdgpu_switcheroo_set_state,
1007         .reprobe = NULL,
1008         .can_switch = amdgpu_switcheroo_can_switch,
1009 };
1010
1011 /**
1012  * amdgpu_device_ip_set_clockgating_state - set the CG state
1013  *
1014  * @dev: amdgpu_device pointer
1015  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1016  * @state: clockgating state (gate or ungate)
1017  *
1018  * Sets the requested clockgating state for all instances of
1019  * the hardware IP specified.
1020  * Returns the error code from the last instance.
1021  */
1022 int amdgpu_device_ip_set_clockgating_state(void *dev,
1023                                            enum amd_ip_block_type block_type,
1024                                            enum amd_clockgating_state state)
1025 {
1026         struct amdgpu_device *adev = dev;
1027         int i, r = 0;
1028
1029         for (i = 0; i < adev->num_ip_blocks; i++) {
1030                 if (!adev->ip_blocks[i].status.valid)
1031                         continue;
1032                 if (adev->ip_blocks[i].version->type != block_type)
1033                         continue;
1034                 if (!adev->ip_blocks[i].version->funcs->set_clockgating_state)
1035                         continue;
1036                 r = adev->ip_blocks[i].version->funcs->set_clockgating_state(
1037                         (void *)adev, state);
1038                 if (r)
1039                         DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",
1040                                   adev->ip_blocks[i].version->funcs->name, r);
1041         }
1042         return r;
1043 }
1044
1045 /**
1046  * amdgpu_device_ip_set_powergating_state - set the PG state
1047  *
1048  * @dev: amdgpu_device pointer
1049  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1050  * @state: powergating state (gate or ungate)
1051  *
1052  * Sets the requested powergating state for all instances of
1053  * the hardware IP specified.
1054  * Returns the error code from the last instance.
1055  */
1056 int amdgpu_device_ip_set_powergating_state(void *dev,
1057                                            enum amd_ip_block_type block_type,
1058                                            enum amd_powergating_state state)
1059 {
1060         struct amdgpu_device *adev = dev;
1061         int i, r = 0;
1062
1063         for (i = 0; i < adev->num_ip_blocks; i++) {
1064                 if (!adev->ip_blocks[i].status.valid)
1065                         continue;
1066                 if (adev->ip_blocks[i].version->type != block_type)
1067                         continue;
1068                 if (!adev->ip_blocks[i].version->funcs->set_powergating_state)
1069                         continue;
1070                 r = adev->ip_blocks[i].version->funcs->set_powergating_state(
1071                         (void *)adev, state);
1072                 if (r)
1073                         DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",
1074                                   adev->ip_blocks[i].version->funcs->name, r);
1075         }
1076         return r;
1077 }
1078
1079 /**
1080  * amdgpu_device_ip_get_clockgating_state - get the CG state
1081  *
1082  * @adev: amdgpu_device pointer
1083  * @flags: clockgating feature flags
1084  *
1085  * Walks the list of IPs on the device and updates the clockgating
1086  * flags for each IP.
1087  * Updates @flags with the feature flags for each hardware IP where
1088  * clockgating is enabled.
1089  */
1090 void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev,
1091                                             u32 *flags)
1092 {
1093         int i;
1094
1095         for (i = 0; i < adev->num_ip_blocks; i++) {
1096                 if (!adev->ip_blocks[i].status.valid)
1097                         continue;
1098                 if (adev->ip_blocks[i].version->funcs->get_clockgating_state)
1099                         adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags);
1100         }
1101 }
1102
1103 /**
1104  * amdgpu_device_ip_wait_for_idle - wait for idle
1105  *
1106  * @adev: amdgpu_device pointer
1107  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1108  *
1109  * Waits for the request hardware IP to be idle.
1110  * Returns 0 for success or a negative error code on failure.
1111  */
1112 int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev,
1113                                    enum amd_ip_block_type block_type)
1114 {
1115         int i, r;
1116
1117         for (i = 0; i < adev->num_ip_blocks; i++) {
1118                 if (!adev->ip_blocks[i].status.valid)
1119                         continue;
1120                 if (adev->ip_blocks[i].version->type == block_type) {
1121                         r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
1122                         if (r)
1123                                 return r;
1124                         break;
1125                 }
1126         }
1127         return 0;
1128
1129 }
1130
1131 /**
1132  * amdgpu_device_ip_is_idle - is the hardware IP idle
1133  *
1134  * @adev: amdgpu_device pointer
1135  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1136  *
1137  * Check if the hardware IP is idle or not.
1138  * Returns true if it the IP is idle, false if not.
1139  */
1140 bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev,
1141                               enum amd_ip_block_type block_type)
1142 {
1143         int i;
1144
1145         for (i = 0; i < adev->num_ip_blocks; i++) {
1146                 if (!adev->ip_blocks[i].status.valid)
1147                         continue;
1148                 if (adev->ip_blocks[i].version->type == block_type)
1149                         return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
1150         }
1151         return true;
1152
1153 }
1154
1155 /**
1156  * amdgpu_device_ip_get_ip_block - get a hw IP pointer
1157  *
1158  * @adev: amdgpu_device pointer
1159  * @type: Type of hardware IP (SMU, GFX, UVD, etc.)
1160  *
1161  * Returns a pointer to the hardware IP block structure
1162  * if it exists for the asic, otherwise NULL.
1163  */
1164 struct amdgpu_ip_block *
1165 amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,
1166                               enum amd_ip_block_type type)
1167 {
1168         int i;
1169
1170         for (i = 0; i < adev->num_ip_blocks; i++)
1171                 if (adev->ip_blocks[i].version->type == type)
1172                         return &adev->ip_blocks[i];
1173
1174         return NULL;
1175 }
1176
1177 /**
1178  * amdgpu_device_ip_block_version_cmp
1179  *
1180  * @adev: amdgpu_device pointer
1181  * @type: enum amd_ip_block_type
1182  * @major: major version
1183  * @minor: minor version
1184  *
1185  * return 0 if equal or greater
1186  * return 1 if smaller or the ip_block doesn't exist
1187  */
1188 int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev,
1189                                        enum amd_ip_block_type type,
1190                                        u32 major, u32 minor)
1191 {
1192         struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type);
1193
1194         if (ip_block && ((ip_block->version->major > major) ||
1195                         ((ip_block->version->major == major) &&
1196                         (ip_block->version->minor >= minor))))
1197                 return 0;
1198
1199         return 1;
1200 }
1201
1202 /**
1203  * amdgpu_device_ip_block_add
1204  *
1205  * @adev: amdgpu_device pointer
1206  * @ip_block_version: pointer to the IP to add
1207  *
1208  * Adds the IP block driver information to the collection of IPs
1209  * on the asic.
1210  */
1211 int amdgpu_device_ip_block_add(struct amdgpu_device *adev,
1212                                const struct amdgpu_ip_block_version *ip_block_version)
1213 {
1214         if (!ip_block_version)
1215                 return -EINVAL;
1216
1217         DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks,
1218                   ip_block_version->funcs->name);
1219
1220         adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
1221
1222         return 0;
1223 }
1224
1225 /**
1226  * amdgpu_device_enable_virtual_display - enable virtual display feature
1227  *
1228  * @adev: amdgpu_device pointer
1229  *
1230  * Enabled the virtual display feature if the user has enabled it via
1231  * the module parameter virtual_display.  This feature provides a virtual
1232  * display hardware on headless boards or in virtualized environments.
1233  * This function parses and validates the configuration string specified by
1234  * the user and configues the virtual display configuration (number of
1235  * virtual connectors, crtcs, etc.) specified.
1236  */
1237 static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
1238 {
1239         adev->enable_virtual_display = false;
1240
1241         if (amdgpu_virtual_display) {
1242                 struct drm_device *ddev = adev->ddev;
1243                 const char *pci_address_name = pci_name(ddev->pdev);
1244                 char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
1245
1246                 pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
1247                 pciaddstr_tmp = pciaddstr;
1248                 while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
1249                         pciaddname = strsep(&pciaddname_tmp, ",");
1250                         if (!strcmp("all", pciaddname)
1251                             || !strcmp(pci_address_name, pciaddname)) {
1252                                 long num_crtc;
1253                                 int res = -1;
1254
1255                                 adev->enable_virtual_display = true;
1256
1257                                 if (pciaddname_tmp)
1258                                         res = kstrtol(pciaddname_tmp, 10,
1259                                                       &num_crtc);
1260
1261                                 if (!res) {
1262                                         if (num_crtc < 1)
1263                                                 num_crtc = 1;
1264                                         if (num_crtc > 6)
1265                                                 num_crtc = 6;
1266                                         adev->mode_info.num_crtc = num_crtc;
1267                                 } else {
1268                                         adev->mode_info.num_crtc = 1;
1269                                 }
1270                                 break;
1271                         }
1272                 }
1273
1274                 DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
1275                          amdgpu_virtual_display, pci_address_name,
1276                          adev->enable_virtual_display, adev->mode_info.num_crtc);
1277
1278                 kfree(pciaddstr);
1279         }
1280 }
1281
1282 /**
1283  * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware
1284  *
1285  * @adev: amdgpu_device pointer
1286  *
1287  * Parses the asic configuration parameters specified in the gpu info
1288  * firmware and makes them availale to the driver for use in configuring
1289  * the asic.
1290  * Returns 0 on success, -EINVAL on failure.
1291  */
1292 static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev)
1293 {
1294         const char *chip_name;
1295         char fw_name[30];
1296         int err;
1297         const struct gpu_info_firmware_header_v1_0 *hdr;
1298
1299         adev->firmware.gpu_info_fw = NULL;
1300
1301         switch (adev->asic_type) {
1302         case CHIP_TOPAZ:
1303         case CHIP_TONGA:
1304         case CHIP_FIJI:
1305         case CHIP_POLARIS10:
1306         case CHIP_POLARIS11:
1307         case CHIP_POLARIS12:
1308         case CHIP_VEGAM:
1309         case CHIP_CARRIZO:
1310         case CHIP_STONEY:
1311 #ifdef CONFIG_DRM_AMDGPU_SI
1312         case CHIP_VERDE:
1313         case CHIP_TAHITI:
1314         case CHIP_PITCAIRN:
1315         case CHIP_OLAND:
1316         case CHIP_HAINAN:
1317 #endif
1318 #ifdef CONFIG_DRM_AMDGPU_CIK
1319         case CHIP_BONAIRE:
1320         case CHIP_HAWAII:
1321         case CHIP_KAVERI:
1322         case CHIP_KABINI:
1323         case CHIP_MULLINS:
1324 #endif
1325         case CHIP_VEGA20:
1326         default:
1327                 return 0;
1328         case CHIP_VEGA10:
1329                 chip_name = "vega10";
1330                 break;
1331         case CHIP_VEGA12:
1332                 chip_name = "vega12";
1333                 break;
1334         case CHIP_RAVEN:
1335                 chip_name = "raven";
1336                 break;
1337         }
1338
1339         snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name);
1340         err = request_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev);
1341         if (err) {
1342                 dev_err(adev->dev,
1343                         "Failed to load gpu_info firmware \"%s\"\n",
1344                         fw_name);
1345                 goto out;
1346         }
1347         err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw);
1348         if (err) {
1349                 dev_err(adev->dev,
1350                         "Failed to validate gpu_info firmware \"%s\"\n",
1351                         fw_name);
1352                 goto out;
1353         }
1354
1355         hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data;
1356         amdgpu_ucode_print_gpu_info_hdr(&hdr->header);
1357
1358         switch (hdr->version_major) {
1359         case 1:
1360         {
1361                 const struct gpu_info_firmware_v1_0 *gpu_info_fw =
1362                         (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data +
1363                                                                 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1364
1365                 adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se);
1366                 adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh);
1367                 adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se);
1368                 adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se);
1369                 adev->gfx.config.max_texture_channel_caches =
1370                         le32_to_cpu(gpu_info_fw->gc_num_tccs);
1371                 adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs);
1372                 adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds);
1373                 adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth);
1374                 adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth);
1375                 adev->gfx.config.double_offchip_lds_buf =
1376                         le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer);
1377                 adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size);
1378                 adev->gfx.cu_info.max_waves_per_simd =
1379                         le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd);
1380                 adev->gfx.cu_info.max_scratch_slots_per_cu =
1381                         le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu);
1382                 adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size);
1383                 break;
1384         }
1385         default:
1386                 dev_err(adev->dev,
1387                         "Unsupported gpu_info table %d\n", hdr->header.ucode_version);
1388                 err = -EINVAL;
1389                 goto out;
1390         }
1391 out:
1392         return err;
1393 }
1394
1395 /**
1396  * amdgpu_device_ip_early_init - run early init for hardware IPs
1397  *
1398  * @adev: amdgpu_device pointer
1399  *
1400  * Early initialization pass for hardware IPs.  The hardware IPs that make
1401  * up each asic are discovered each IP's early_init callback is run.  This
1402  * is the first stage in initializing the asic.
1403  * Returns 0 on success, negative error code on failure.
1404  */
1405 static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
1406 {
1407         int i, r;
1408
1409         amdgpu_device_enable_virtual_display(adev);
1410
1411         switch (adev->asic_type) {
1412         case CHIP_TOPAZ:
1413         case CHIP_TONGA:
1414         case CHIP_FIJI:
1415         case CHIP_POLARIS10:
1416         case CHIP_POLARIS11:
1417         case CHIP_POLARIS12:
1418         case CHIP_VEGAM:
1419         case CHIP_CARRIZO:
1420         case CHIP_STONEY:
1421                 if (adev->asic_type == CHIP_CARRIZO || adev->asic_type == CHIP_STONEY)
1422                         adev->family = AMDGPU_FAMILY_CZ;
1423                 else
1424                         adev->family = AMDGPU_FAMILY_VI;
1425
1426                 r = vi_set_ip_blocks(adev);
1427                 if (r)
1428                         return r;
1429                 break;
1430 #ifdef CONFIG_DRM_AMDGPU_SI
1431         case CHIP_VERDE:
1432         case CHIP_TAHITI:
1433         case CHIP_PITCAIRN:
1434         case CHIP_OLAND:
1435         case CHIP_HAINAN:
1436                 adev->family = AMDGPU_FAMILY_SI;
1437                 r = si_set_ip_blocks(adev);
1438                 if (r)
1439                         return r;
1440                 break;
1441 #endif
1442 #ifdef CONFIG_DRM_AMDGPU_CIK
1443         case CHIP_BONAIRE:
1444         case CHIP_HAWAII:
1445         case CHIP_KAVERI:
1446         case CHIP_KABINI:
1447         case CHIP_MULLINS:
1448                 if ((adev->asic_type == CHIP_BONAIRE) || (adev->asic_type == CHIP_HAWAII))
1449                         adev->family = AMDGPU_FAMILY_CI;
1450                 else
1451                         adev->family = AMDGPU_FAMILY_KV;
1452
1453                 r = cik_set_ip_blocks(adev);
1454                 if (r)
1455                         return r;
1456                 break;
1457 #endif
1458         case CHIP_VEGA10:
1459         case CHIP_VEGA12:
1460         case CHIP_VEGA20:
1461         case CHIP_RAVEN:
1462                 if (adev->asic_type == CHIP_RAVEN)
1463                         adev->family = AMDGPU_FAMILY_RV;
1464                 else
1465                         adev->family = AMDGPU_FAMILY_AI;
1466
1467                 r = soc15_set_ip_blocks(adev);
1468                 if (r)
1469                         return r;
1470                 break;
1471         default:
1472                 /* FIXME: not supported yet */
1473                 return -EINVAL;
1474         }
1475
1476         r = amdgpu_device_parse_gpu_info_fw(adev);
1477         if (r)
1478                 return r;
1479
1480         amdgpu_amdkfd_device_probe(adev);
1481
1482         if (amdgpu_sriov_vf(adev)) {
1483                 r = amdgpu_virt_request_full_gpu(adev, true);
1484                 if (r)
1485                         return -EAGAIN;
1486         }
1487
1488         adev->powerplay.pp_feature = amdgpu_pp_feature_mask;
1489
1490         for (i = 0; i < adev->num_ip_blocks; i++) {
1491                 if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
1492                         DRM_ERROR("disabled ip block: %d <%s>\n",
1493                                   i, adev->ip_blocks[i].version->funcs->name);
1494                         adev->ip_blocks[i].status.valid = false;
1495                 } else {
1496                         if (adev->ip_blocks[i].version->funcs->early_init) {
1497                                 r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
1498                                 if (r == -ENOENT) {
1499                                         adev->ip_blocks[i].status.valid = false;
1500                                 } else if (r) {
1501                                         DRM_ERROR("early_init of IP block <%s> failed %d\n",
1502                                                   adev->ip_blocks[i].version->funcs->name, r);
1503                                         return r;
1504                                 } else {
1505                                         adev->ip_blocks[i].status.valid = true;
1506                                 }
1507                         } else {
1508                                 adev->ip_blocks[i].status.valid = true;
1509                         }
1510                 }
1511         }
1512
1513         adev->cg_flags &= amdgpu_cg_mask;
1514         adev->pg_flags &= amdgpu_pg_mask;
1515
1516         return 0;
1517 }
1518
1519 /**
1520  * amdgpu_device_ip_init - run init for hardware IPs
1521  *
1522  * @adev: amdgpu_device pointer
1523  *
1524  * Main initialization pass for hardware IPs.  The list of all the hardware
1525  * IPs that make up the asic is walked and the sw_init and hw_init callbacks
1526  * are run.  sw_init initializes the software state associated with each IP
1527  * and hw_init initializes the hardware associated with each IP.
1528  * Returns 0 on success, negative error code on failure.
1529  */
1530 static int amdgpu_device_ip_init(struct amdgpu_device *adev)
1531 {
1532         int i, r;
1533
1534         for (i = 0; i < adev->num_ip_blocks; i++) {
1535                 if (!adev->ip_blocks[i].status.valid)
1536                         continue;
1537                 r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
1538                 if (r) {
1539                         DRM_ERROR("sw_init of IP block <%s> failed %d\n",
1540                                   adev->ip_blocks[i].version->funcs->name, r);
1541                         return r;
1542                 }
1543                 adev->ip_blocks[i].status.sw = true;
1544
1545                 /* need to do gmc hw init early so we can allocate gpu mem */
1546                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1547                         r = amdgpu_device_vram_scratch_init(adev);
1548                         if (r) {
1549                                 DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r);
1550                                 return r;
1551                         }
1552                         r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1553                         if (r) {
1554                                 DRM_ERROR("hw_init %d failed %d\n", i, r);
1555                                 return r;
1556                         }
1557                         r = amdgpu_device_wb_init(adev);
1558                         if (r) {
1559                                 DRM_ERROR("amdgpu_device_wb_init failed %d\n", r);
1560                                 return r;
1561                         }
1562                         adev->ip_blocks[i].status.hw = true;
1563
1564                         /* right after GMC hw init, we create CSA */
1565                         if (amdgpu_sriov_vf(adev)) {
1566                                 r = amdgpu_allocate_static_csa(adev);
1567                                 if (r) {
1568                                         DRM_ERROR("allocate CSA failed %d\n", r);
1569                                         return r;
1570                                 }
1571                         }
1572                 }
1573         }
1574
1575         for (i = 0; i < adev->num_ip_blocks; i++) {
1576                 if (!adev->ip_blocks[i].status.sw)
1577                         continue;
1578                 if (adev->ip_blocks[i].status.hw)
1579                         continue;
1580                 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1581                 if (r) {
1582                         DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1583                                   adev->ip_blocks[i].version->funcs->name, r);
1584                         return r;
1585                 }
1586                 adev->ip_blocks[i].status.hw = true;
1587         }
1588
1589         amdgpu_amdkfd_device_init(adev);
1590
1591         if (amdgpu_sriov_vf(adev))
1592                 amdgpu_virt_release_full_gpu(adev, true);
1593
1594         return 0;
1595 }
1596
1597 /**
1598  * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer
1599  *
1600  * @adev: amdgpu_device pointer
1601  *
1602  * Writes a reset magic value to the gart pointer in VRAM.  The driver calls
1603  * this function before a GPU reset.  If the value is retained after a
1604  * GPU reset, VRAM has not been lost.  Some GPU resets may destry VRAM contents.
1605  */
1606 static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev)
1607 {
1608         memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
1609 }
1610
1611 /**
1612  * amdgpu_device_check_vram_lost - check if vram is valid
1613  *
1614  * @adev: amdgpu_device pointer
1615  *
1616  * Checks the reset magic value written to the gart pointer in VRAM.
1617  * The driver calls this after a GPU reset to see if the contents of
1618  * VRAM is lost or now.
1619  * returns true if vram is lost, false if not.
1620  */
1621 static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev)
1622 {
1623         return !!memcmp(adev->gart.ptr, adev->reset_magic,
1624                         AMDGPU_RESET_MAGIC_NUM);
1625 }
1626
1627 /**
1628  * amdgpu_device_set_cg_state - set clockgating for amdgpu device
1629  *
1630  * @adev: amdgpu_device pointer
1631  *
1632  * The list of all the hardware IPs that make up the asic is walked and the
1633  * set_clockgating_state callbacks are run.
1634  * Late initialization pass enabling clockgating for hardware IPs.
1635  * Fini or suspend, pass disabling clockgating for hardware IPs.
1636  * Returns 0 on success, negative error code on failure.
1637  */
1638
1639 static int amdgpu_device_set_cg_state(struct amdgpu_device *adev,
1640                                                 enum amd_clockgating_state state)
1641 {
1642         int i, j, r;
1643
1644         if (amdgpu_emu_mode == 1)
1645                 return 0;
1646
1647         for (j = 0; j < adev->num_ip_blocks; j++) {
1648                 i = state == AMD_CG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
1649                 if (!adev->ip_blocks[i].status.valid)
1650                         continue;
1651                 /* skip CG for VCE/UVD, it's handled specially */
1652                 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1653                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1654                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
1655                     adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1656                         /* enable clockgating to save power */
1657                         r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1658                                                                                      state);
1659                         if (r) {
1660                                 DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
1661                                           adev->ip_blocks[i].version->funcs->name, r);
1662                                 return r;
1663                         }
1664                 }
1665         }
1666
1667         return 0;
1668 }
1669
1670 static int amdgpu_device_set_pg_state(struct amdgpu_device *adev, enum amd_powergating_state state)
1671 {
1672         int i, j, r;
1673
1674         if (amdgpu_emu_mode == 1)
1675                 return 0;
1676
1677         for (j = 0; j < adev->num_ip_blocks; j++) {
1678                 i = state == AMD_PG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
1679                 if (!adev->ip_blocks[i].status.valid)
1680                         continue;
1681                 /* skip CG for VCE/UVD, it's handled specially */
1682                 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1683                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1684                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
1685                     adev->ip_blocks[i].version->funcs->set_powergating_state) {
1686                         /* enable powergating to save power */
1687                         r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev,
1688                                                                                         state);
1689                         if (r) {
1690                                 DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n",
1691                                           adev->ip_blocks[i].version->funcs->name, r);
1692                                 return r;
1693                         }
1694                 }
1695         }
1696         return 0;
1697 }
1698
1699 /**
1700  * amdgpu_device_ip_late_init - run late init for hardware IPs
1701  *
1702  * @adev: amdgpu_device pointer
1703  *
1704  * Late initialization pass for hardware IPs.  The list of all the hardware
1705  * IPs that make up the asic is walked and the late_init callbacks are run.
1706  * late_init covers any special initialization that an IP requires
1707  * after all of the have been initialized or something that needs to happen
1708  * late in the init process.
1709  * Returns 0 on success, negative error code on failure.
1710  */
1711 static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
1712 {
1713         int i = 0, r;
1714
1715         for (i = 0; i < adev->num_ip_blocks; i++) {
1716                 if (!adev->ip_blocks[i].status.valid)
1717                         continue;
1718                 if (adev->ip_blocks[i].version->funcs->late_init) {
1719                         r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
1720                         if (r) {
1721                                 DRM_ERROR("late_init of IP block <%s> failed %d\n",
1722                                           adev->ip_blocks[i].version->funcs->name, r);
1723                                 return r;
1724                         }
1725                         adev->ip_blocks[i].status.late_initialized = true;
1726                 }
1727         }
1728
1729         amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE);
1730         amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE);
1731
1732         queue_delayed_work(system_wq, &adev->late_init_work,
1733                            msecs_to_jiffies(AMDGPU_RESUME_MS));
1734
1735         amdgpu_device_fill_reset_magic(adev);
1736
1737         return 0;
1738 }
1739
1740 /**
1741  * amdgpu_device_ip_fini - run fini for hardware IPs
1742  *
1743  * @adev: amdgpu_device pointer
1744  *
1745  * Main teardown pass for hardware IPs.  The list of all the hardware
1746  * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks
1747  * are run.  hw_fini tears down the hardware associated with each IP
1748  * and sw_fini tears down any software state associated with each IP.
1749  * Returns 0 on success, negative error code on failure.
1750  */
1751 static int amdgpu_device_ip_fini(struct amdgpu_device *adev)
1752 {
1753         int i, r;
1754
1755         amdgpu_amdkfd_device_fini(adev);
1756
1757         amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
1758         amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
1759
1760         /* need to disable SMC first */
1761         for (i = 0; i < adev->num_ip_blocks; i++) {
1762                 if (!adev->ip_blocks[i].status.hw)
1763                         continue;
1764                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
1765                         r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
1766                         /* XXX handle errors */
1767                         if (r) {
1768                                 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
1769                                           adev->ip_blocks[i].version->funcs->name, r);
1770                         }
1771                         adev->ip_blocks[i].status.hw = false;
1772                         break;
1773                 }
1774         }
1775
1776         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1777                 if (!adev->ip_blocks[i].status.hw)
1778                         continue;
1779
1780                 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
1781                 /* XXX handle errors */
1782                 if (r) {
1783                         DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
1784                                   adev->ip_blocks[i].version->funcs->name, r);
1785                 }
1786
1787                 adev->ip_blocks[i].status.hw = false;
1788         }
1789
1790
1791         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1792                 if (!adev->ip_blocks[i].status.sw)
1793                         continue;
1794
1795                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1796                         amdgpu_free_static_csa(adev);
1797                         amdgpu_device_wb_fini(adev);
1798                         amdgpu_device_vram_scratch_fini(adev);
1799                 }
1800
1801                 r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
1802                 /* XXX handle errors */
1803                 if (r) {
1804                         DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
1805                                   adev->ip_blocks[i].version->funcs->name, r);
1806                 }
1807                 adev->ip_blocks[i].status.sw = false;
1808                 adev->ip_blocks[i].status.valid = false;
1809         }
1810
1811         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1812                 if (!adev->ip_blocks[i].status.late_initialized)
1813                         continue;
1814                 if (adev->ip_blocks[i].version->funcs->late_fini)
1815                         adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
1816                 adev->ip_blocks[i].status.late_initialized = false;
1817         }
1818
1819         if (amdgpu_sriov_vf(adev))
1820                 if (amdgpu_virt_release_full_gpu(adev, false))
1821                         DRM_ERROR("failed to release exclusive mode on fini\n");
1822
1823         return 0;
1824 }
1825
1826 /**
1827  * amdgpu_device_ip_late_init_func_handler - work handler for ib test
1828  *
1829  * @work: work_struct.
1830  */
1831 static void amdgpu_device_ip_late_init_func_handler(struct work_struct *work)
1832 {
1833         struct amdgpu_device *adev =
1834                 container_of(work, struct amdgpu_device, late_init_work.work);
1835         int r;
1836
1837         r = amdgpu_ib_ring_tests(adev);
1838         if (r)
1839                 DRM_ERROR("ib ring test failed (%d).\n", r);
1840 }
1841
1842 static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work)
1843 {
1844         struct amdgpu_device *adev =
1845                 container_of(work, struct amdgpu_device, gfx.gfx_off_delay_work.work);
1846
1847         mutex_lock(&adev->gfx.gfx_off_mutex);
1848         if (!adev->gfx.gfx_off_state && !adev->gfx.gfx_off_req_count) {
1849                 if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, true))
1850                         adev->gfx.gfx_off_state = true;
1851         }
1852         mutex_unlock(&adev->gfx.gfx_off_mutex);
1853 }
1854
1855 /**
1856  * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1)
1857  *
1858  * @adev: amdgpu_device pointer
1859  *
1860  * Main suspend function for hardware IPs.  The list of all the hardware
1861  * IPs that make up the asic is walked, clockgating is disabled and the
1862  * suspend callbacks are run.  suspend puts the hardware and software state
1863  * in each IP into a state suitable for suspend.
1864  * Returns 0 on success, negative error code on failure.
1865  */
1866 static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev)
1867 {
1868         int i, r;
1869
1870         amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
1871         amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
1872
1873         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1874                 if (!adev->ip_blocks[i].status.valid)
1875                         continue;
1876                 /* displays are handled separately */
1877                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) {
1878                         /* XXX handle errors */
1879                         r = adev->ip_blocks[i].version->funcs->suspend(adev);
1880                         /* XXX handle errors */
1881                         if (r) {
1882                                 DRM_ERROR("suspend of IP block <%s> failed %d\n",
1883                                           adev->ip_blocks[i].version->funcs->name, r);
1884                         }
1885                 }
1886         }
1887
1888         return 0;
1889 }
1890
1891 /**
1892  * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2)
1893  *
1894  * @adev: amdgpu_device pointer
1895  *
1896  * Main suspend function for hardware IPs.  The list of all the hardware
1897  * IPs that make up the asic is walked, clockgating is disabled and the
1898  * suspend callbacks are run.  suspend puts the hardware and software state
1899  * in each IP into a state suitable for suspend.
1900  * Returns 0 on success, negative error code on failure.
1901  */
1902 static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev)
1903 {
1904         int i, r;
1905
1906         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1907                 if (!adev->ip_blocks[i].status.valid)
1908                         continue;
1909                 /* displays are handled in phase1 */
1910                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE)
1911                         continue;
1912                 /* XXX handle errors */
1913                 r = adev->ip_blocks[i].version->funcs->suspend(adev);
1914                 /* XXX handle errors */
1915                 if (r) {
1916                         DRM_ERROR("suspend of IP block <%s> failed %d\n",
1917                                   adev->ip_blocks[i].version->funcs->name, r);
1918                 }
1919         }
1920
1921         return 0;
1922 }
1923
1924 /**
1925  * amdgpu_device_ip_suspend - run suspend for hardware IPs
1926  *
1927  * @adev: amdgpu_device pointer
1928  *
1929  * Main suspend function for hardware IPs.  The list of all the hardware
1930  * IPs that make up the asic is walked, clockgating is disabled and the
1931  * suspend callbacks are run.  suspend puts the hardware and software state
1932  * in each IP into a state suitable for suspend.
1933  * Returns 0 on success, negative error code on failure.
1934  */
1935 int amdgpu_device_ip_suspend(struct amdgpu_device *adev)
1936 {
1937         int r;
1938
1939         if (amdgpu_sriov_vf(adev))
1940                 amdgpu_virt_request_full_gpu(adev, false);
1941
1942         r = amdgpu_device_ip_suspend_phase1(adev);
1943         if (r)
1944                 return r;
1945         r = amdgpu_device_ip_suspend_phase2(adev);
1946
1947         if (amdgpu_sriov_vf(adev))
1948                 amdgpu_virt_release_full_gpu(adev, false);
1949
1950         return r;
1951 }
1952
1953 static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
1954 {
1955         int i, r;
1956
1957         static enum amd_ip_block_type ip_order[] = {
1958                 AMD_IP_BLOCK_TYPE_GMC,
1959                 AMD_IP_BLOCK_TYPE_COMMON,
1960                 AMD_IP_BLOCK_TYPE_IH,
1961         };
1962
1963         for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
1964                 int j;
1965                 struct amdgpu_ip_block *block;
1966
1967                 for (j = 0; j < adev->num_ip_blocks; j++) {
1968                         block = &adev->ip_blocks[j];
1969
1970                         if (block->version->type != ip_order[i] ||
1971                                 !block->status.valid)
1972                                 continue;
1973
1974                         r = block->version->funcs->hw_init(adev);
1975                         DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
1976                         if (r)
1977                                 return r;
1978                 }
1979         }
1980
1981         return 0;
1982 }
1983
1984 static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
1985 {
1986         int i, r;
1987
1988         static enum amd_ip_block_type ip_order[] = {
1989                 AMD_IP_BLOCK_TYPE_SMC,
1990                 AMD_IP_BLOCK_TYPE_PSP,
1991                 AMD_IP_BLOCK_TYPE_DCE,
1992                 AMD_IP_BLOCK_TYPE_GFX,
1993                 AMD_IP_BLOCK_TYPE_SDMA,
1994                 AMD_IP_BLOCK_TYPE_UVD,
1995                 AMD_IP_BLOCK_TYPE_VCE
1996         };
1997
1998         for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
1999                 int j;
2000                 struct amdgpu_ip_block *block;
2001
2002                 for (j = 0; j < adev->num_ip_blocks; j++) {
2003                         block = &adev->ip_blocks[j];
2004
2005                         if (block->version->type != ip_order[i] ||
2006                                 !block->status.valid)
2007                                 continue;
2008
2009                         r = block->version->funcs->hw_init(adev);
2010                         DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
2011                         if (r)
2012                                 return r;
2013                 }
2014         }
2015
2016         return 0;
2017 }
2018
2019 /**
2020  * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs
2021  *
2022  * @adev: amdgpu_device pointer
2023  *
2024  * First resume function for hardware IPs.  The list of all the hardware
2025  * IPs that make up the asic is walked and the resume callbacks are run for
2026  * COMMON, GMC, and IH.  resume puts the hardware into a functional state
2027  * after a suspend and updates the software state as necessary.  This
2028  * function is also used for restoring the GPU after a GPU reset.
2029  * Returns 0 on success, negative error code on failure.
2030  */
2031 static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)
2032 {
2033         int i, r;
2034
2035         for (i = 0; i < adev->num_ip_blocks; i++) {
2036                 if (!adev->ip_blocks[i].status.valid)
2037                         continue;
2038                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2039                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2040                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
2041                         r = adev->ip_blocks[i].version->funcs->resume(adev);
2042                         if (r) {
2043                                 DRM_ERROR("resume of IP block <%s> failed %d\n",
2044                                           adev->ip_blocks[i].version->funcs->name, r);
2045                                 return r;
2046                         }
2047                 }
2048         }
2049
2050         return 0;
2051 }
2052
2053 /**
2054  * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs
2055  *
2056  * @adev: amdgpu_device pointer
2057  *
2058  * First resume function for hardware IPs.  The list of all the hardware
2059  * IPs that make up the asic is walked and the resume callbacks are run for
2060  * all blocks except COMMON, GMC, and IH.  resume puts the hardware into a
2061  * functional state after a suspend and updates the software state as
2062  * necessary.  This function is also used for restoring the GPU after a GPU
2063  * reset.
2064  * Returns 0 on success, negative error code on failure.
2065  */
2066 static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)
2067 {
2068         int i, r;
2069
2070         for (i = 0; i < adev->num_ip_blocks; i++) {
2071                 if (!adev->ip_blocks[i].status.valid)
2072                         continue;
2073                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2074                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2075                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH)
2076                         continue;
2077                 r = adev->ip_blocks[i].version->funcs->resume(adev);
2078                 if (r) {
2079                         DRM_ERROR("resume of IP block <%s> failed %d\n",
2080                                   adev->ip_blocks[i].version->funcs->name, r);
2081                         return r;
2082                 }
2083         }
2084
2085         return 0;
2086 }
2087
2088 /**
2089  * amdgpu_device_ip_resume - run resume for hardware IPs
2090  *
2091  * @adev: amdgpu_device pointer
2092  *
2093  * Main resume function for hardware IPs.  The hardware IPs
2094  * are split into two resume functions because they are
2095  * are also used in in recovering from a GPU reset and some additional
2096  * steps need to be take between them.  In this case (S3/S4) they are
2097  * run sequentially.
2098  * Returns 0 on success, negative error code on failure.
2099  */
2100 static int amdgpu_device_ip_resume(struct amdgpu_device *adev)
2101 {
2102         int r;
2103
2104         r = amdgpu_device_ip_resume_phase1(adev);
2105         if (r)
2106                 return r;
2107         r = amdgpu_device_ip_resume_phase2(adev);
2108
2109         return r;
2110 }
2111
2112 /**
2113  * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV
2114  *
2115  * @adev: amdgpu_device pointer
2116  *
2117  * Query the VBIOS data tables to determine if the board supports SR-IOV.
2118  */
2119 static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
2120 {
2121         if (amdgpu_sriov_vf(adev)) {
2122                 if (adev->is_atom_fw) {
2123                         if (amdgpu_atomfirmware_gpu_supports_virtualization(adev))
2124                                 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2125                 } else {
2126                         if (amdgpu_atombios_has_gpu_virtualization_table(adev))
2127                                 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2128                 }
2129
2130                 if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
2131                         amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
2132         }
2133 }
2134
2135 /**
2136  * amdgpu_device_asic_has_dc_support - determine if DC supports the asic
2137  *
2138  * @asic_type: AMD asic type
2139  *
2140  * Check if there is DC (new modesetting infrastructre) support for an asic.
2141  * returns true if DC has support, false if not.
2142  */
2143 bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
2144 {
2145         switch (asic_type) {
2146 #if defined(CONFIG_DRM_AMD_DC)
2147         case CHIP_BONAIRE:
2148         case CHIP_KAVERI:
2149         case CHIP_KABINI:
2150         case CHIP_MULLINS:
2151                 /*
2152                  * We have systems in the wild with these ASICs that require
2153                  * LVDS and VGA support which is not supported with DC.
2154                  *
2155                  * Fallback to the non-DC driver here by default so as not to
2156                  * cause regressions.
2157                  */
2158                 return amdgpu_dc > 0;
2159         case CHIP_HAWAII:
2160         case CHIP_CARRIZO:
2161         case CHIP_STONEY:
2162         case CHIP_POLARIS10:
2163         case CHIP_POLARIS11:
2164         case CHIP_POLARIS12:
2165         case CHIP_VEGAM:
2166         case CHIP_TONGA:
2167         case CHIP_FIJI:
2168         case CHIP_VEGA10:
2169         case CHIP_VEGA12:
2170         case CHIP_VEGA20:
2171 #if defined(CONFIG_DRM_AMD_DC_DCN1_0)
2172         case CHIP_RAVEN:
2173 #endif
2174                 return amdgpu_dc != 0;
2175 #endif
2176         default:
2177                 return false;
2178         }
2179 }
2180
2181 /**
2182  * amdgpu_device_has_dc_support - check if dc is supported
2183  *
2184  * @adev: amdgpu_device_pointer
2185  *
2186  * Returns true for supported, false for not supported
2187  */
2188 bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
2189 {
2190         if (amdgpu_sriov_vf(adev))
2191                 return false;
2192
2193         return amdgpu_device_asic_has_dc_support(adev->asic_type);
2194 }
2195
2196 /**
2197  * amdgpu_device_init - initialize the driver
2198  *
2199  * @adev: amdgpu_device pointer
2200  * @ddev: drm dev pointer
2201  * @pdev: pci dev pointer
2202  * @flags: driver flags
2203  *
2204  * Initializes the driver info and hw (all asics).
2205  * Returns 0 for success or an error on failure.
2206  * Called at driver startup.
2207  */
2208 int amdgpu_device_init(struct amdgpu_device *adev,
2209                        struct drm_device *ddev,
2210                        struct pci_dev *pdev,
2211                        uint32_t flags)
2212 {
2213         int r, i;
2214         bool runtime = false;
2215         u32 max_MBps;
2216
2217         adev->shutdown = false;
2218         adev->dev = &pdev->dev;
2219         adev->ddev = ddev;
2220         adev->pdev = pdev;
2221         adev->flags = flags;
2222         adev->asic_type = flags & AMD_ASIC_MASK;
2223         adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
2224         if (amdgpu_emu_mode == 1)
2225                 adev->usec_timeout *= 2;
2226         adev->gmc.gart_size = 512 * 1024 * 1024;
2227         adev->accel_working = false;
2228         adev->num_rings = 0;
2229         adev->mman.buffer_funcs = NULL;
2230         adev->mman.buffer_funcs_ring = NULL;
2231         adev->vm_manager.vm_pte_funcs = NULL;
2232         adev->vm_manager.vm_pte_num_rqs = 0;
2233         adev->gmc.gmc_funcs = NULL;
2234         adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
2235         bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
2236
2237         adev->smc_rreg = &amdgpu_invalid_rreg;
2238         adev->smc_wreg = &amdgpu_invalid_wreg;
2239         adev->pcie_rreg = &amdgpu_invalid_rreg;
2240         adev->pcie_wreg = &amdgpu_invalid_wreg;
2241         adev->pciep_rreg = &amdgpu_invalid_rreg;
2242         adev->pciep_wreg = &amdgpu_invalid_wreg;
2243         adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
2244         adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
2245         adev->didt_rreg = &amdgpu_invalid_rreg;
2246         adev->didt_wreg = &amdgpu_invalid_wreg;
2247         adev->gc_cac_rreg = &amdgpu_invalid_rreg;
2248         adev->gc_cac_wreg = &amdgpu_invalid_wreg;
2249         adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
2250         adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
2251
2252         DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
2253                  amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
2254                  pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
2255
2256         /* mutex initialization are all done here so we
2257          * can recall function without having locking issues */
2258         atomic_set(&adev->irq.ih.lock, 0);
2259         mutex_init(&adev->firmware.mutex);
2260         mutex_init(&adev->pm.mutex);
2261         mutex_init(&adev->gfx.gpu_clock_mutex);
2262         mutex_init(&adev->srbm_mutex);
2263         mutex_init(&adev->gfx.pipe_reserve_mutex);
2264         mutex_init(&adev->gfx.gfx_off_mutex);
2265         mutex_init(&adev->grbm_idx_mutex);
2266         mutex_init(&adev->mn_lock);
2267         mutex_init(&adev->virt.vf_errors.lock);
2268         hash_init(adev->mn_hash);
2269         mutex_init(&adev->lock_reset);
2270
2271         amdgpu_device_check_arguments(adev);
2272
2273         spin_lock_init(&adev->mmio_idx_lock);
2274         spin_lock_init(&adev->smc_idx_lock);
2275         spin_lock_init(&adev->pcie_idx_lock);
2276         spin_lock_init(&adev->uvd_ctx_idx_lock);
2277         spin_lock_init(&adev->didt_idx_lock);
2278         spin_lock_init(&adev->gc_cac_idx_lock);
2279         spin_lock_init(&adev->se_cac_idx_lock);
2280         spin_lock_init(&adev->audio_endpt_idx_lock);
2281         spin_lock_init(&adev->mm_stats.lock);
2282
2283         INIT_LIST_HEAD(&adev->shadow_list);
2284         mutex_init(&adev->shadow_list_lock);
2285
2286         INIT_LIST_HEAD(&adev->ring_lru_list);
2287         spin_lock_init(&adev->ring_lru_list_lock);
2288
2289         INIT_DELAYED_WORK(&adev->late_init_work,
2290                           amdgpu_device_ip_late_init_func_handler);
2291         INIT_DELAYED_WORK(&adev->gfx.gfx_off_delay_work,
2292                           amdgpu_device_delay_enable_gfx_off);
2293
2294         adev->gfx.gfx_off_req_count = 1;
2295         adev->pm.ac_power = power_supply_is_system_supplied() > 0 ? true : false;
2296
2297         /* Registers mapping */
2298         /* TODO: block userspace mapping of io register */
2299         if (adev->asic_type >= CHIP_BONAIRE) {
2300                 adev->rmmio_base = pci_resource_start(adev->pdev, 5);
2301                 adev->rmmio_size = pci_resource_len(adev->pdev, 5);
2302         } else {
2303                 adev->rmmio_base = pci_resource_start(adev->pdev, 2);
2304                 adev->rmmio_size = pci_resource_len(adev->pdev, 2);
2305         }
2306
2307         adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
2308         if (adev->rmmio == NULL) {
2309                 return -ENOMEM;
2310         }
2311         DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
2312         DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
2313
2314         /* doorbell bar mapping */
2315         amdgpu_device_doorbell_init(adev);
2316
2317         /* io port mapping */
2318         for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
2319                 if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) {
2320                         adev->rio_mem_size = pci_resource_len(adev->pdev, i);
2321                         adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size);
2322                         break;
2323                 }
2324         }
2325         if (adev->rio_mem == NULL)
2326                 DRM_INFO("PCI I/O BAR is not found.\n");
2327
2328         amdgpu_device_get_pcie_info(adev);
2329
2330         /* early init functions */
2331         r = amdgpu_device_ip_early_init(adev);
2332         if (r)
2333                 return r;
2334
2335         /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
2336         /* this will fail for cards that aren't VGA class devices, just
2337          * ignore it */
2338         vga_client_register(adev->pdev, adev, NULL, amdgpu_device_vga_set_decode);
2339
2340         if (amdgpu_device_is_px(ddev))
2341                 runtime = true;
2342         if (!pci_is_thunderbolt_attached(adev->pdev))
2343                 vga_switcheroo_register_client(adev->pdev,
2344                                                &amdgpu_switcheroo_ops, runtime);
2345         if (runtime)
2346                 vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
2347
2348         if (amdgpu_emu_mode == 1) {
2349                 /* post the asic on emulation mode */
2350                 emu_soc_asic_init(adev);
2351                 goto fence_driver_init;
2352         }
2353
2354         /* Read BIOS */
2355         if (!amdgpu_get_bios(adev)) {
2356                 r = -EINVAL;
2357                 goto failed;
2358         }
2359
2360         r = amdgpu_atombios_init(adev);
2361         if (r) {
2362                 dev_err(adev->dev, "amdgpu_atombios_init failed\n");
2363                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0);
2364                 goto failed;
2365         }
2366
2367         /* detect if we are with an SRIOV vbios */
2368         amdgpu_device_detect_sriov_bios(adev);
2369
2370         /* Post card if necessary */
2371         if (amdgpu_device_need_post(adev)) {
2372                 if (!adev->bios) {
2373                         dev_err(adev->dev, "no vBIOS found\n");
2374                         r = -EINVAL;
2375                         goto failed;
2376                 }
2377                 DRM_INFO("GPU posting now...\n");
2378                 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2379                 if (r) {
2380                         dev_err(adev->dev, "gpu post error!\n");
2381                         goto failed;
2382                 }
2383         }
2384
2385         if (adev->is_atom_fw) {
2386                 /* Initialize clocks */
2387                 r = amdgpu_atomfirmware_get_clock_info(adev);
2388                 if (r) {
2389                         dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
2390                         amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
2391                         goto failed;
2392                 }
2393         } else {
2394                 /* Initialize clocks */
2395                 r = amdgpu_atombios_get_clock_info(adev);
2396                 if (r) {
2397                         dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
2398                         amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
2399                         goto failed;
2400                 }
2401                 /* init i2c buses */
2402                 if (!amdgpu_device_has_dc_support(adev))
2403                         amdgpu_atombios_i2c_init(adev);
2404         }
2405
2406 fence_driver_init:
2407         /* Fence driver */
2408         r = amdgpu_fence_driver_init(adev);
2409         if (r) {
2410                 dev_err(adev->dev, "amdgpu_fence_driver_init failed\n");
2411                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
2412                 goto failed;
2413         }
2414
2415         /* init the mode config */
2416         drm_mode_config_init(adev->ddev);
2417
2418         r = amdgpu_device_ip_init(adev);
2419         if (r) {
2420                 /* failed in exclusive mode due to timeout */
2421                 if (amdgpu_sriov_vf(adev) &&
2422                     !amdgpu_sriov_runtime(adev) &&
2423                     amdgpu_virt_mmio_blocked(adev) &&
2424                     !amdgpu_virt_wait_reset(adev)) {
2425                         dev_err(adev->dev, "VF exclusive mode timeout\n");
2426                         /* Don't send request since VF is inactive. */
2427                         adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
2428                         adev->virt.ops = NULL;
2429                         r = -EAGAIN;
2430                         goto failed;
2431                 }
2432                 dev_err(adev->dev, "amdgpu_device_ip_init failed\n");
2433                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
2434                 goto failed;
2435         }
2436
2437         adev->accel_working = true;
2438
2439         amdgpu_vm_check_compute_bug(adev);
2440
2441         /* Initialize the buffer migration limit. */
2442         if (amdgpu_moverate >= 0)
2443                 max_MBps = amdgpu_moverate;
2444         else
2445                 max_MBps = 8; /* Allow 8 MB/s. */
2446         /* Get a log2 for easy divisions. */
2447         adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
2448
2449         r = amdgpu_ib_pool_init(adev);
2450         if (r) {
2451                 dev_err(adev->dev, "IB initialization failed (%d).\n", r);
2452                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r);
2453                 goto failed;
2454         }
2455
2456         if (amdgpu_sriov_vf(adev))
2457                 amdgpu_virt_init_data_exchange(adev);
2458
2459         amdgpu_fbdev_init(adev);
2460
2461         r = amdgpu_pm_sysfs_init(adev);
2462         if (r)
2463                 DRM_ERROR("registering pm debugfs failed (%d).\n", r);
2464
2465         r = amdgpu_debugfs_gem_init(adev);
2466         if (r)
2467                 DRM_ERROR("registering gem debugfs failed (%d).\n", r);
2468
2469         r = amdgpu_debugfs_regs_init(adev);
2470         if (r)
2471                 DRM_ERROR("registering register debugfs failed (%d).\n", r);
2472
2473         r = amdgpu_debugfs_firmware_init(adev);
2474         if (r)
2475                 DRM_ERROR("registering firmware debugfs failed (%d).\n", r);
2476
2477         r = amdgpu_debugfs_init(adev);
2478         if (r)
2479                 DRM_ERROR("Creating debugfs files failed (%d).\n", r);
2480
2481         if ((amdgpu_testing & 1)) {
2482                 if (adev->accel_working)
2483                         amdgpu_test_moves(adev);
2484                 else
2485                         DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n");
2486         }
2487         if (amdgpu_benchmarking) {
2488                 if (adev->accel_working)
2489                         amdgpu_benchmark(adev, amdgpu_benchmarking);
2490                 else
2491                         DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n");
2492         }
2493
2494         /* enable clockgating, etc. after ib tests, etc. since some blocks require
2495          * explicit gating rather than handling it automatically.
2496          */
2497         r = amdgpu_device_ip_late_init(adev);
2498         if (r) {
2499                 dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n");
2500                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
2501                 goto failed;
2502         }
2503
2504         return 0;
2505
2506 failed:
2507         amdgpu_vf_error_trans_all(adev);
2508         if (runtime)
2509                 vga_switcheroo_fini_domain_pm_ops(adev->dev);
2510
2511         return r;
2512 }
2513
2514 /**
2515  * amdgpu_device_fini - tear down the driver
2516  *
2517  * @adev: amdgpu_device pointer
2518  *
2519  * Tear down the driver info (all asics).
2520  * Called at driver shutdown.
2521  */
2522 void amdgpu_device_fini(struct amdgpu_device *adev)
2523 {
2524         int r;
2525
2526         DRM_INFO("amdgpu: finishing device.\n");
2527         adev->shutdown = true;
2528         /* disable all interrupts */
2529         amdgpu_irq_disable_all(adev);
2530         if (adev->mode_info.mode_config_initialized){
2531                 if (!amdgpu_device_has_dc_support(adev))
2532                         drm_crtc_force_disable_all(adev->ddev);
2533                 else
2534                         drm_atomic_helper_shutdown(adev->ddev);
2535         }
2536         amdgpu_ib_pool_fini(adev);
2537         amdgpu_fence_driver_fini(adev);
2538         amdgpu_pm_sysfs_fini(adev);
2539         amdgpu_fbdev_fini(adev);
2540         r = amdgpu_device_ip_fini(adev);
2541         if (adev->firmware.gpu_info_fw) {
2542                 release_firmware(adev->firmware.gpu_info_fw);
2543                 adev->firmware.gpu_info_fw = NULL;
2544         }
2545         adev->accel_working = false;
2546         cancel_delayed_work_sync(&adev->late_init_work);
2547         /* free i2c buses */
2548         if (!amdgpu_device_has_dc_support(adev))
2549                 amdgpu_i2c_fini(adev);
2550
2551         if (amdgpu_emu_mode != 1)
2552                 amdgpu_atombios_fini(adev);
2553
2554         kfree(adev->bios);
2555         adev->bios = NULL;
2556         if (!pci_is_thunderbolt_attached(adev->pdev))
2557                 vga_switcheroo_unregister_client(adev->pdev);
2558         if (adev->flags & AMD_IS_PX)
2559                 vga_switcheroo_fini_domain_pm_ops(adev->dev);
2560         vga_client_register(adev->pdev, NULL, NULL, NULL);
2561         if (adev->rio_mem)
2562                 pci_iounmap(adev->pdev, adev->rio_mem);
2563         adev->rio_mem = NULL;
2564         iounmap(adev->rmmio);
2565         adev->rmmio = NULL;
2566         amdgpu_device_doorbell_fini(adev);
2567         amdgpu_debugfs_regs_cleanup(adev);
2568 }
2569
2570
2571 /*
2572  * Suspend & resume.
2573  */
2574 /**
2575  * amdgpu_device_suspend - initiate device suspend
2576  *
2577  * @dev: drm dev pointer
2578  * @suspend: suspend state
2579  * @fbcon : notify the fbdev of suspend
2580  *
2581  * Puts the hw in the suspend state (all asics).
2582  * Returns 0 for success or an error on failure.
2583  * Called at driver suspend.
2584  */
2585 int amdgpu_device_suspend(struct drm_device *dev, bool suspend, bool fbcon)
2586 {
2587         struct amdgpu_device *adev;
2588         struct drm_crtc *crtc;
2589         struct drm_connector *connector;
2590         int r;
2591
2592         if (dev == NULL || dev->dev_private == NULL) {
2593                 return -ENODEV;
2594         }
2595
2596         adev = dev->dev_private;
2597
2598         if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
2599                 return 0;
2600
2601         drm_kms_helper_poll_disable(dev);
2602
2603         if (fbcon)
2604                 amdgpu_fbdev_set_suspend(adev, 1);
2605
2606         cancel_delayed_work_sync(&adev->late_init_work);
2607
2608         if (!amdgpu_device_has_dc_support(adev)) {
2609                 /* turn off display hw */
2610                 drm_modeset_lock_all(dev);
2611                 list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2612                         drm_helper_connector_dpms(connector, DRM_MODE_DPMS_OFF);
2613                 }
2614                 drm_modeset_unlock_all(dev);
2615                         /* unpin the front buffers and cursors */
2616                 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2617                         struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2618                         struct drm_framebuffer *fb = crtc->primary->fb;
2619                         struct amdgpu_bo *robj;
2620
2621                         if (amdgpu_crtc->cursor_bo) {
2622                                 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2623                                 r = amdgpu_bo_reserve(aobj, true);
2624                                 if (r == 0) {
2625                                         amdgpu_bo_unpin(aobj);
2626                                         amdgpu_bo_unreserve(aobj);
2627                                 }
2628                         }
2629
2630                         if (fb == NULL || fb->obj[0] == NULL) {
2631                                 continue;
2632                         }
2633                         robj = gem_to_amdgpu_bo(fb->obj[0]);
2634                         /* don't unpin kernel fb objects */
2635                         if (!amdgpu_fbdev_robj_is_fb(adev, robj)) {
2636                                 r = amdgpu_bo_reserve(robj, true);
2637                                 if (r == 0) {
2638                                         amdgpu_bo_unpin(robj);
2639                                         amdgpu_bo_unreserve(robj);
2640                                 }
2641                         }
2642                 }
2643         }
2644
2645         amdgpu_amdkfd_suspend(adev);
2646
2647         r = amdgpu_device_ip_suspend_phase1(adev);
2648
2649         /* evict vram memory */
2650         amdgpu_bo_evict_vram(adev);
2651
2652         amdgpu_fence_driver_suspend(adev);
2653
2654         r = amdgpu_device_ip_suspend_phase2(adev);
2655
2656         /* evict remaining vram memory
2657          * This second call to evict vram is to evict the gart page table
2658          * using the CPU.
2659          */
2660         amdgpu_bo_evict_vram(adev);
2661
2662         pci_save_state(dev->pdev);
2663         if (suspend) {
2664                 /* Shut down the device */
2665                 pci_disable_device(dev->pdev);
2666                 pci_set_power_state(dev->pdev, PCI_D3hot);
2667         } else {
2668                 r = amdgpu_asic_reset(adev);
2669                 if (r)
2670                         DRM_ERROR("amdgpu asic reset failed\n");
2671         }
2672
2673         return 0;
2674 }
2675
2676 /**
2677  * amdgpu_device_resume - initiate device resume
2678  *
2679  * @dev: drm dev pointer
2680  * @resume: resume state
2681  * @fbcon : notify the fbdev of resume
2682  *
2683  * Bring the hw back to operating state (all asics).
2684  * Returns 0 for success or an error on failure.
2685  * Called at driver resume.
2686  */
2687 int amdgpu_device_resume(struct drm_device *dev, bool resume, bool fbcon)
2688 {
2689         struct drm_connector *connector;
2690         struct amdgpu_device *adev = dev->dev_private;
2691         struct drm_crtc *crtc;
2692         int r = 0;
2693
2694         if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
2695                 return 0;
2696
2697         if (resume) {
2698                 pci_set_power_state(dev->pdev, PCI_D0);
2699                 pci_restore_state(dev->pdev);
2700                 r = pci_enable_device(dev->pdev);
2701                 if (r)
2702                         return r;
2703         }
2704
2705         /* post card */
2706         if (amdgpu_device_need_post(adev)) {
2707                 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2708                 if (r)
2709                         DRM_ERROR("amdgpu asic init failed\n");
2710         }
2711
2712         r = amdgpu_device_ip_resume(adev);
2713         if (r) {
2714                 DRM_ERROR("amdgpu_device_ip_resume failed (%d).\n", r);
2715                 return r;
2716         }
2717         amdgpu_fence_driver_resume(adev);
2718
2719
2720         r = amdgpu_device_ip_late_init(adev);
2721         if (r)
2722                 return r;
2723
2724         if (!amdgpu_device_has_dc_support(adev)) {
2725                 /* pin cursors */
2726                 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2727                         struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2728
2729                         if (amdgpu_crtc->cursor_bo) {
2730                                 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2731                                 r = amdgpu_bo_reserve(aobj, true);
2732                                 if (r == 0) {
2733                                         r = amdgpu_bo_pin(aobj, AMDGPU_GEM_DOMAIN_VRAM);
2734                                         if (r != 0)
2735                                                 DRM_ERROR("Failed to pin cursor BO (%d)\n", r);
2736                                         amdgpu_crtc->cursor_addr = amdgpu_bo_gpu_offset(aobj);
2737                                         amdgpu_bo_unreserve(aobj);
2738                                 }
2739                         }
2740                 }
2741         }
2742         r = amdgpu_amdkfd_resume(adev);
2743         if (r)
2744                 return r;
2745
2746         /* Make sure IB tests flushed */
2747         flush_delayed_work(&adev->late_init_work);
2748
2749         /* blat the mode back in */
2750         if (fbcon) {
2751                 if (!amdgpu_device_has_dc_support(adev)) {
2752                         /* pre DCE11 */
2753                         drm_helper_resume_force_mode(dev);
2754
2755                         /* turn on display hw */
2756                         drm_modeset_lock_all(dev);
2757                         list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2758                                 drm_helper_connector_dpms(connector, DRM_MODE_DPMS_ON);
2759                         }
2760                         drm_modeset_unlock_all(dev);
2761                 }
2762                 amdgpu_fbdev_set_suspend(adev, 0);
2763         }
2764
2765         drm_kms_helper_poll_enable(dev);
2766
2767         /*
2768          * Most of the connector probing functions try to acquire runtime pm
2769          * refs to ensure that the GPU is powered on when connector polling is
2770          * performed. Since we're calling this from a runtime PM callback,
2771          * trying to acquire rpm refs will cause us to deadlock.
2772          *
2773          * Since we're guaranteed to be holding the rpm lock, it's safe to
2774          * temporarily disable the rpm helpers so this doesn't deadlock us.
2775          */
2776 #ifdef CONFIG_PM
2777         dev->dev->power.disable_depth++;
2778 #endif
2779         if (!amdgpu_device_has_dc_support(adev))
2780                 drm_helper_hpd_irq_event(dev);
2781         else
2782                 drm_kms_helper_hotplug_event(dev);
2783 #ifdef CONFIG_PM
2784         dev->dev->power.disable_depth--;
2785 #endif
2786         return 0;
2787 }
2788
2789 /**
2790  * amdgpu_device_ip_check_soft_reset - did soft reset succeed
2791  *
2792  * @adev: amdgpu_device pointer
2793  *
2794  * The list of all the hardware IPs that make up the asic is walked and
2795  * the check_soft_reset callbacks are run.  check_soft_reset determines
2796  * if the asic is still hung or not.
2797  * Returns true if any of the IPs are still in a hung state, false if not.
2798  */
2799 static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)
2800 {
2801         int i;
2802         bool asic_hang = false;
2803
2804         if (amdgpu_sriov_vf(adev))
2805                 return true;
2806
2807         if (amdgpu_asic_need_full_reset(adev))
2808                 return true;
2809
2810         for (i = 0; i < adev->num_ip_blocks; i++) {
2811                 if (!adev->ip_blocks[i].status.valid)
2812                         continue;
2813                 if (adev->ip_blocks[i].version->funcs->check_soft_reset)
2814                         adev->ip_blocks[i].status.hang =
2815                                 adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
2816                 if (adev->ip_blocks[i].status.hang) {
2817                         DRM_INFO("IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
2818                         asic_hang = true;
2819                 }
2820         }
2821         return asic_hang;
2822 }
2823
2824 /**
2825  * amdgpu_device_ip_pre_soft_reset - prepare for soft reset
2826  *
2827  * @adev: amdgpu_device pointer
2828  *
2829  * The list of all the hardware IPs that make up the asic is walked and the
2830  * pre_soft_reset callbacks are run if the block is hung.  pre_soft_reset
2831  * handles any IP specific hardware or software state changes that are
2832  * necessary for a soft reset to succeed.
2833  * Returns 0 on success, negative error code on failure.
2834  */
2835 static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)
2836 {
2837         int i, r = 0;
2838
2839         for (i = 0; i < adev->num_ip_blocks; i++) {
2840                 if (!adev->ip_blocks[i].status.valid)
2841                         continue;
2842                 if (adev->ip_blocks[i].status.hang &&
2843                     adev->ip_blocks[i].version->funcs->pre_soft_reset) {
2844                         r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
2845                         if (r)
2846                                 return r;
2847                 }
2848         }
2849
2850         return 0;
2851 }
2852
2853 /**
2854  * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed
2855  *
2856  * @adev: amdgpu_device pointer
2857  *
2858  * Some hardware IPs cannot be soft reset.  If they are hung, a full gpu
2859  * reset is necessary to recover.
2860  * Returns true if a full asic reset is required, false if not.
2861  */
2862 static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)
2863 {
2864         int i;
2865
2866         if (amdgpu_asic_need_full_reset(adev))
2867                 return true;
2868
2869         for (i = 0; i < adev->num_ip_blocks; i++) {
2870                 if (!adev->ip_blocks[i].status.valid)
2871                         continue;
2872                 if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
2873                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
2874                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
2875                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
2876                      adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
2877                         if (adev->ip_blocks[i].status.hang) {
2878                                 DRM_INFO("Some block need full reset!\n");
2879                                 return true;
2880                         }
2881                 }
2882         }
2883         return false;
2884 }
2885
2886 /**
2887  * amdgpu_device_ip_soft_reset - do a soft reset
2888  *
2889  * @adev: amdgpu_device pointer
2890  *
2891  * The list of all the hardware IPs that make up the asic is walked and the
2892  * soft_reset callbacks are run if the block is hung.  soft_reset handles any
2893  * IP specific hardware or software state changes that are necessary to soft
2894  * reset the IP.
2895  * Returns 0 on success, negative error code on failure.
2896  */
2897 static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)
2898 {
2899         int i, r = 0;
2900
2901         for (i = 0; i < adev->num_ip_blocks; i++) {
2902                 if (!adev->ip_blocks[i].status.valid)
2903                         continue;
2904                 if (adev->ip_blocks[i].status.hang &&
2905                     adev->ip_blocks[i].version->funcs->soft_reset) {
2906                         r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
2907                         if (r)
2908                                 return r;
2909                 }
2910         }
2911
2912         return 0;
2913 }
2914
2915 /**
2916  * amdgpu_device_ip_post_soft_reset - clean up from soft reset
2917  *
2918  * @adev: amdgpu_device pointer
2919  *
2920  * The list of all the hardware IPs that make up the asic is walked and the
2921  * post_soft_reset callbacks are run if the asic was hung.  post_soft_reset
2922  * handles any IP specific hardware or software state changes that are
2923  * necessary after the IP has been soft reset.
2924  * Returns 0 on success, negative error code on failure.
2925  */
2926 static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)
2927 {
2928         int i, r = 0;
2929
2930         for (i = 0; i < adev->num_ip_blocks; i++) {
2931                 if (!adev->ip_blocks[i].status.valid)
2932                         continue;
2933                 if (adev->ip_blocks[i].status.hang &&
2934                     adev->ip_blocks[i].version->funcs->post_soft_reset)
2935                         r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
2936                 if (r)
2937                         return r;
2938         }
2939
2940         return 0;
2941 }
2942
2943 /**
2944  * amdgpu_device_recover_vram_from_shadow - restore shadowed VRAM buffers
2945  *
2946  * @adev: amdgpu_device pointer
2947  * @ring: amdgpu_ring for the engine handling the buffer operations
2948  * @bo: amdgpu_bo buffer whose shadow is being restored
2949  * @fence: dma_fence associated with the operation
2950  *
2951  * Restores the VRAM buffer contents from the shadow in GTT.  Used to
2952  * restore things like GPUVM page tables after a GPU reset where
2953  * the contents of VRAM might be lost.
2954  * Returns 0 on success, negative error code on failure.
2955  */
2956 static int amdgpu_device_recover_vram_from_shadow(struct amdgpu_device *adev,
2957                                                   struct amdgpu_ring *ring,
2958                                                   struct amdgpu_bo *bo,
2959                                                   struct dma_fence **fence)
2960 {
2961         uint32_t domain;
2962         int r;
2963
2964         if (!bo->shadow)
2965                 return 0;
2966
2967         r = amdgpu_bo_reserve(bo, true);
2968         if (r)
2969                 return r;
2970         domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
2971         /* if bo has been evicted, then no need to recover */
2972         if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
2973                 r = amdgpu_bo_validate(bo->shadow);
2974                 if (r) {
2975                         DRM_ERROR("bo validate failed!\n");
2976                         goto err;
2977                 }
2978
2979                 r = amdgpu_bo_restore_from_shadow(adev, ring, bo,
2980                                                  NULL, fence, true);
2981                 if (r) {
2982                         DRM_ERROR("recover page table failed!\n");
2983                         goto err;
2984                 }
2985         }
2986 err:
2987         amdgpu_bo_unreserve(bo);
2988         return r;
2989 }
2990
2991 /**
2992  * amdgpu_device_handle_vram_lost - Handle the loss of VRAM contents
2993  *
2994  * @adev: amdgpu_device pointer
2995  *
2996  * Restores the contents of VRAM buffers from the shadows in GTT.  Used to
2997  * restore things like GPUVM page tables after a GPU reset where
2998  * the contents of VRAM might be lost.
2999  * Returns 0 on success, 1 on failure.
3000  */
3001 static int amdgpu_device_handle_vram_lost(struct amdgpu_device *adev)
3002 {
3003         struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
3004         struct amdgpu_bo *bo, *tmp;
3005         struct dma_fence *fence = NULL, *next = NULL;
3006         long r = 1;
3007         int i = 0;
3008         long tmo;
3009
3010         if (amdgpu_sriov_runtime(adev))
3011                 tmo = msecs_to_jiffies(8000);
3012         else
3013                 tmo = msecs_to_jiffies(100);
3014
3015         DRM_INFO("recover vram bo from shadow start\n");
3016         mutex_lock(&adev->shadow_list_lock);
3017         list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) {
3018                 next = NULL;
3019                 amdgpu_device_recover_vram_from_shadow(adev, ring, bo, &next);
3020                 if (fence) {
3021                         r = dma_fence_wait_timeout(fence, false, tmo);
3022                         if (r == 0)
3023                                 pr_err("wait fence %p[%d] timeout\n", fence, i);
3024                         else if (r < 0)
3025                                 pr_err("wait fence %p[%d] interrupted\n", fence, i);
3026                         if (r < 1) {
3027                                 dma_fence_put(fence);
3028                                 fence = next;
3029                                 break;
3030                         }
3031                         i++;
3032                 }
3033
3034                 dma_fence_put(fence);
3035                 fence = next;
3036         }
3037         mutex_unlock(&adev->shadow_list_lock);
3038
3039         if (fence) {
3040                 r = dma_fence_wait_timeout(fence, false, tmo);
3041                 if (r == 0)
3042                         pr_err("wait fence %p[%d] timeout\n", fence, i);
3043                 else if (r < 0)
3044                         pr_err("wait fence %p[%d] interrupted\n", fence, i);
3045
3046         }
3047         dma_fence_put(fence);
3048
3049         if (r > 0)
3050                 DRM_INFO("recover vram bo from shadow done\n");
3051         else
3052                 DRM_ERROR("recover vram bo from shadow failed\n");
3053
3054         return (r > 0) ? 0 : 1;
3055 }
3056
3057 /**
3058  * amdgpu_device_reset - reset ASIC/GPU for bare-metal or passthrough
3059  *
3060  * @adev: amdgpu device pointer
3061  *
3062  * attempt to do soft-reset or full-reset and reinitialize Asic
3063  * return 0 means succeeded otherwise failed
3064  */
3065 static int amdgpu_device_reset(struct amdgpu_device *adev)
3066 {
3067         bool need_full_reset, vram_lost = 0;
3068         int r;
3069
3070         need_full_reset = amdgpu_device_ip_need_full_reset(adev);
3071
3072         if (!need_full_reset) {
3073                 amdgpu_device_ip_pre_soft_reset(adev);
3074                 r = amdgpu_device_ip_soft_reset(adev);
3075                 amdgpu_device_ip_post_soft_reset(adev);
3076                 if (r || amdgpu_device_ip_check_soft_reset(adev)) {
3077                         DRM_INFO("soft reset failed, will fallback to full reset!\n");
3078                         need_full_reset = true;
3079                 }
3080         }
3081
3082         if (need_full_reset) {
3083                 r = amdgpu_device_ip_suspend(adev);
3084
3085 retry:
3086                 r = amdgpu_asic_reset(adev);
3087                 /* post card */
3088                 amdgpu_atom_asic_init(adev->mode_info.atom_context);
3089
3090                 if (!r) {
3091                         dev_info(adev->dev, "GPU reset succeeded, trying to resume\n");
3092                         r = amdgpu_device_ip_resume_phase1(adev);
3093                         if (r)
3094                                 goto out;
3095
3096                         vram_lost = amdgpu_device_check_vram_lost(adev);
3097                         if (vram_lost) {
3098                                 DRM_ERROR("VRAM is lost!\n");
3099                                 atomic_inc(&adev->vram_lost_counter);
3100                         }
3101
3102                         r = amdgpu_gtt_mgr_recover(
3103                                 &adev->mman.bdev.man[TTM_PL_TT]);
3104                         if (r)
3105                                 goto out;
3106
3107                         r = amdgpu_device_ip_resume_phase2(adev);
3108                         if (r)
3109                                 goto out;
3110
3111                         if (vram_lost)
3112                                 amdgpu_device_fill_reset_magic(adev);
3113                 }
3114         }
3115
3116 out:
3117         if (!r) {
3118                 amdgpu_irq_gpu_reset_resume_helper(adev);
3119                 r = amdgpu_ib_ring_tests(adev);
3120                 if (r) {
3121                         dev_err(adev->dev, "ib ring test failed (%d).\n", r);
3122                         r = amdgpu_device_ip_suspend(adev);
3123                         need_full_reset = true;
3124                         goto retry;
3125                 }
3126         }
3127
3128         if (!r && ((need_full_reset && !(adev->flags & AMD_IS_APU)) || vram_lost))
3129                 r = amdgpu_device_handle_vram_lost(adev);
3130
3131         return r;
3132 }
3133
3134 /**
3135  * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
3136  *
3137  * @adev: amdgpu device pointer
3138  * @from_hypervisor: request from hypervisor
3139  *
3140  * do VF FLR and reinitialize Asic
3141  * return 0 means succeeded otherwise failed
3142  */
3143 static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
3144                                      bool from_hypervisor)
3145 {
3146         int r;
3147
3148         if (from_hypervisor)
3149                 r = amdgpu_virt_request_full_gpu(adev, true);
3150         else
3151                 r = amdgpu_virt_reset_gpu(adev);
3152         if (r)
3153                 return r;
3154
3155         /* Resume IP prior to SMC */
3156         r = amdgpu_device_ip_reinit_early_sriov(adev);
3157         if (r)
3158                 goto error;
3159
3160         /* we need recover gart prior to run SMC/CP/SDMA resume */
3161         amdgpu_gtt_mgr_recover(&adev->mman.bdev.man[TTM_PL_TT]);
3162
3163         /* now we are okay to resume SMC/CP/SDMA */
3164         r = amdgpu_device_ip_reinit_late_sriov(adev);
3165         if (r)
3166                 goto error;
3167
3168         amdgpu_irq_gpu_reset_resume_helper(adev);
3169         r = amdgpu_ib_ring_tests(adev);
3170
3171 error:
3172         amdgpu_virt_release_full_gpu(adev, true);
3173         if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
3174                 atomic_inc(&adev->vram_lost_counter);
3175                 r = amdgpu_device_handle_vram_lost(adev);
3176         }
3177
3178         return r;
3179 }
3180
3181 /**
3182  * amdgpu_device_should_recover_gpu - check if we should try GPU recovery
3183  *
3184  * @adev: amdgpu device pointer
3185  *
3186  * Check amdgpu_gpu_recovery and SRIOV status to see if we should try to recover
3187  * a hung GPU.
3188  */
3189 bool amdgpu_device_should_recover_gpu(struct amdgpu_device *adev)
3190 {
3191         if (!amdgpu_device_ip_check_soft_reset(adev)) {
3192                 DRM_INFO("Timeout, but no hardware hang detected.\n");
3193                 return false;
3194         }
3195
3196         if (amdgpu_gpu_recovery == 0 || (amdgpu_gpu_recovery == -1  &&
3197                                          !amdgpu_sriov_vf(adev))) {
3198                 DRM_INFO("GPU recovery disabled.\n");
3199                 return false;
3200         }
3201
3202         return true;
3203 }
3204
3205 /**
3206  * amdgpu_device_gpu_recover - reset the asic and recover scheduler
3207  *
3208  * @adev: amdgpu device pointer
3209  * @job: which job trigger hang
3210  *
3211  * Attempt to reset the GPU if it has hung (all asics).
3212  * Returns 0 for success or an error on failure.
3213  */
3214 int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
3215                               struct amdgpu_job *job)
3216 {
3217         int i, r, resched;
3218
3219         dev_info(adev->dev, "GPU reset begin!\n");
3220
3221         mutex_lock(&adev->lock_reset);
3222         atomic_inc(&adev->gpu_reset_counter);
3223         adev->in_gpu_reset = 1;
3224
3225         /* Block kfd */
3226         amdgpu_amdkfd_pre_reset(adev);
3227
3228         /* block TTM */
3229         resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
3230
3231         /* block all schedulers and reset given job's ring */
3232         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3233                 struct amdgpu_ring *ring = adev->rings[i];
3234
3235                 if (!ring || !ring->sched.thread)
3236                         continue;
3237
3238                 kthread_park(ring->sched.thread);
3239
3240                 if (job && job->base.sched == &ring->sched)
3241                         continue;
3242
3243                 drm_sched_hw_job_reset(&ring->sched, job ? &job->base : NULL);
3244
3245                 /* after all hw jobs are reset, hw fence is meaningless, so force_completion */
3246                 amdgpu_fence_driver_force_completion(ring);
3247         }
3248
3249         if (amdgpu_sriov_vf(adev))
3250                 r = amdgpu_device_reset_sriov(adev, job ? false : true);
3251         else
3252                 r = amdgpu_device_reset(adev);
3253
3254         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3255                 struct amdgpu_ring *ring = adev->rings[i];
3256
3257                 if (!ring || !ring->sched.thread)
3258                         continue;
3259
3260                 /* only need recovery sched of the given job's ring
3261                  * or all rings (in the case @job is NULL)
3262                  * after above amdgpu_reset accomplished
3263                  */
3264                 if ((!job || job->base.sched == &ring->sched) && !r)
3265                         drm_sched_job_recovery(&ring->sched);
3266
3267                 kthread_unpark(ring->sched.thread);
3268         }
3269
3270         if (!amdgpu_device_has_dc_support(adev)) {
3271                 drm_helper_resume_force_mode(adev->ddev);
3272         }
3273
3274         ttm_bo_unlock_delayed_workqueue(&adev->mman.bdev, resched);
3275
3276         if (r) {
3277                 /* bad news, how to tell it to userspace ? */
3278                 dev_info(adev->dev, "GPU reset(%d) failed\n", atomic_read(&adev->gpu_reset_counter));
3279                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r);
3280         } else {
3281                 dev_info(adev->dev, "GPU reset(%d) succeeded!\n",atomic_read(&adev->gpu_reset_counter));
3282         }
3283
3284         /*unlock kfd */
3285         amdgpu_amdkfd_post_reset(adev);
3286         amdgpu_vf_error_trans_all(adev);
3287         adev->in_gpu_reset = 0;
3288         mutex_unlock(&adev->lock_reset);
3289         return r;
3290 }
3291
3292 /**
3293  * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot
3294  *
3295  * @adev: amdgpu_device pointer
3296  *
3297  * Fetchs and stores in the driver the PCIE capabilities (gen speed
3298  * and lanes) of the slot the device is in. Handles APUs and
3299  * virtualized environments where PCIE config space may not be available.
3300  */
3301 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev)
3302 {
3303         struct pci_dev *pdev;
3304         enum pci_bus_speed speed_cap;
3305         enum pcie_link_width link_width;
3306
3307         if (amdgpu_pcie_gen_cap)
3308                 adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
3309
3310         if (amdgpu_pcie_lane_cap)
3311                 adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
3312
3313         /* covers APUs as well */
3314         if (pci_is_root_bus(adev->pdev->bus)) {
3315                 if (adev->pm.pcie_gen_mask == 0)
3316                         adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
3317                 if (adev->pm.pcie_mlw_mask == 0)
3318                         adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
3319                 return;
3320         }
3321
3322         if (adev->pm.pcie_gen_mask == 0) {
3323                 /* asic caps */
3324                 pdev = adev->pdev;
3325                 speed_cap = pcie_get_speed_cap(pdev);
3326                 if (speed_cap == PCI_SPEED_UNKNOWN) {
3327                         adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3328                                                   CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3329                                                   CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
3330                 } else {
3331                         if (speed_cap == PCIE_SPEED_16_0GT)
3332                                 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3333                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3334                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
3335                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4);
3336                         else if (speed_cap == PCIE_SPEED_8_0GT)
3337                                 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3338                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3339                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
3340                         else if (speed_cap == PCIE_SPEED_5_0GT)
3341                                 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3342                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2);
3343                         else
3344                                 adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1;
3345                 }
3346                 /* platform caps */
3347                 pdev = adev->ddev->pdev->bus->self;
3348                 speed_cap = pcie_get_speed_cap(pdev);
3349                 if (speed_cap == PCI_SPEED_UNKNOWN) {
3350                         adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3351                                                    CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
3352                 } else {
3353                         if (speed_cap == PCIE_SPEED_16_0GT)
3354                                 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3355                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3356                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
3357                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4);
3358                         else if (speed_cap == PCIE_SPEED_8_0GT)
3359                                 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3360                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3361                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3);
3362                         else if (speed_cap == PCIE_SPEED_5_0GT)
3363                                 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3364                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
3365                         else
3366                                 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
3367
3368                 }
3369         }
3370         if (adev->pm.pcie_mlw_mask == 0) {
3371                 pdev = adev->ddev->pdev->bus->self;
3372                 link_width = pcie_get_width_cap(pdev);
3373                 if (link_width == PCIE_LNK_WIDTH_UNKNOWN) {
3374                         adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK;
3375                 } else {
3376                         switch (link_width) {
3377                         case PCIE_LNK_X32:
3378                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
3379                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
3380                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3381                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3382                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3383                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3384                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3385                                 break;
3386                         case PCIE_LNK_X16:
3387                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
3388                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3389                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3390                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3391                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3392                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3393                                 break;
3394                         case PCIE_LNK_X12:
3395                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3396                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3397                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3398                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3399                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3400                                 break;
3401                         case PCIE_LNK_X8:
3402                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3403                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3404                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3405                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3406                                 break;
3407                         case PCIE_LNK_X4:
3408                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3409                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3410                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3411                                 break;
3412                         case PCIE_LNK_X2:
3413                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3414                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3415                                 break;
3416                         case PCIE_LNK_X1:
3417                                 adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;
3418                                 break;
3419                         default:
3420                                 break;
3421                         }
3422                 }
3423         }
3424 }
3425
This page took 0.244089 seconds and 4 git commands to generate.