drm/amdgpu: Call amdgpu_ucode_fini_bo in amd_powerplay.c
[linux-2.6-block.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/kthread.h>
29 #include <linux/console.h>
30 #include <linux/slab.h>
31 #include <drm/drmP.h>
32 #include <drm/drm_crtc_helper.h>
33 #include <drm/drm_atomic_helper.h>
34 #include <drm/amdgpu_drm.h>
35 #include <linux/vgaarb.h>
36 #include <linux/vga_switcheroo.h>
37 #include <linux/efi.h>
38 #include "amdgpu.h"
39 #include "amdgpu_trace.h"
40 #include "amdgpu_i2c.h"
41 #include "atom.h"
42 #include "amdgpu_atombios.h"
43 #include "amdgpu_atomfirmware.h"
44 #include "amd_pcie.h"
45 #ifdef CONFIG_DRM_AMDGPU_SI
46 #include "si.h"
47 #endif
48 #ifdef CONFIG_DRM_AMDGPU_CIK
49 #include "cik.h"
50 #endif
51 #include "vi.h"
52 #include "soc15.h"
53 #include "bif/bif_4_1_d.h"
54 #include <linux/pci.h>
55 #include <linux/firmware.h>
56 #include "amdgpu_vf_error.h"
57
58 #include "amdgpu_amdkfd.h"
59 #include "amdgpu_pm.h"
60
61 MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin");
62 MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin");
63
64 #define AMDGPU_RESUME_MS                2000
65
66 static const char *amdgpu_asic_name[] = {
67         "TAHITI",
68         "PITCAIRN",
69         "VERDE",
70         "OLAND",
71         "HAINAN",
72         "BONAIRE",
73         "KAVERI",
74         "KABINI",
75         "HAWAII",
76         "MULLINS",
77         "TOPAZ",
78         "TONGA",
79         "FIJI",
80         "CARRIZO",
81         "STONEY",
82         "POLARIS10",
83         "POLARIS11",
84         "POLARIS12",
85         "VEGA10",
86         "RAVEN",
87         "LAST",
88 };
89
90 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev);
91
92 bool amdgpu_device_is_px(struct drm_device *dev)
93 {
94         struct amdgpu_device *adev = dev->dev_private;
95
96         if (adev->flags & AMD_IS_PX)
97                 return true;
98         return false;
99 }
100
101 /*
102  * MMIO register access helper functions.
103  */
104 uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg,
105                         uint32_t acc_flags)
106 {
107         uint32_t ret;
108
109         if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
110                 return amdgpu_virt_kiq_rreg(adev, reg);
111
112         if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
113                 ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
114         else {
115                 unsigned long flags;
116
117                 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
118                 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
119                 ret = readl(((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
120                 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
121         }
122         trace_amdgpu_mm_rreg(adev->pdev->device, reg, ret);
123         return ret;
124 }
125
126 /*
127  * MMIO register read with bytes helper functions
128  * @offset:bytes offset from MMIO start
129  *
130 */
131
132 uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset) {
133         if (offset < adev->rmmio_size)
134                 return (readb(adev->rmmio + offset));
135         BUG();
136 }
137
138 /*
139  * MMIO register write with bytes helper functions
140  * @offset:bytes offset from MMIO start
141  * @value: the value want to be written to the register
142  *
143 */
144 void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value) {
145         if (offset < adev->rmmio_size)
146                 writeb(value, adev->rmmio + offset);
147         else
148                 BUG();
149 }
150
151
152 void amdgpu_mm_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v,
153                     uint32_t acc_flags)
154 {
155         trace_amdgpu_mm_wreg(adev->pdev->device, reg, v);
156
157         if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
158                 adev->last_mm_index = v;
159         }
160
161         if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
162                 return amdgpu_virt_kiq_wreg(adev, reg, v);
163
164         if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
165                 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
166         else {
167                 unsigned long flags;
168
169                 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
170                 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
171                 writel(v, ((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
172                 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
173         }
174
175         if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
176                 udelay(500);
177         }
178 }
179
180 u32 amdgpu_io_rreg(struct amdgpu_device *adev, u32 reg)
181 {
182         if ((reg * 4) < adev->rio_mem_size)
183                 return ioread32(adev->rio_mem + (reg * 4));
184         else {
185                 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
186                 return ioread32(adev->rio_mem + (mmMM_DATA * 4));
187         }
188 }
189
190 void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32 v)
191 {
192         if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
193                 adev->last_mm_index = v;
194         }
195
196         if ((reg * 4) < adev->rio_mem_size)
197                 iowrite32(v, adev->rio_mem + (reg * 4));
198         else {
199                 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
200                 iowrite32(v, adev->rio_mem + (mmMM_DATA * 4));
201         }
202
203         if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
204                 udelay(500);
205         }
206 }
207
208 /**
209  * amdgpu_mm_rdoorbell - read a doorbell dword
210  *
211  * @adev: amdgpu_device pointer
212  * @index: doorbell index
213  *
214  * Returns the value in the doorbell aperture at the
215  * requested doorbell index (CIK).
216  */
217 u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
218 {
219         if (index < adev->doorbell.num_doorbells) {
220                 return readl(adev->doorbell.ptr + index);
221         } else {
222                 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
223                 return 0;
224         }
225 }
226
227 /**
228  * amdgpu_mm_wdoorbell - write a doorbell dword
229  *
230  * @adev: amdgpu_device pointer
231  * @index: doorbell index
232  * @v: value to write
233  *
234  * Writes @v to the doorbell aperture at the
235  * requested doorbell index (CIK).
236  */
237 void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v)
238 {
239         if (index < adev->doorbell.num_doorbells) {
240                 writel(v, adev->doorbell.ptr + index);
241         } else {
242                 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
243         }
244 }
245
246 /**
247  * amdgpu_mm_rdoorbell64 - read a doorbell Qword
248  *
249  * @adev: amdgpu_device pointer
250  * @index: doorbell index
251  *
252  * Returns the value in the doorbell aperture at the
253  * requested doorbell index (VEGA10+).
254  */
255 u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index)
256 {
257         if (index < adev->doorbell.num_doorbells) {
258                 return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index));
259         } else {
260                 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
261                 return 0;
262         }
263 }
264
265 /**
266  * amdgpu_mm_wdoorbell64 - write a doorbell Qword
267  *
268  * @adev: amdgpu_device pointer
269  * @index: doorbell index
270  * @v: value to write
271  *
272  * Writes @v to the doorbell aperture at the
273  * requested doorbell index (VEGA10+).
274  */
275 void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v)
276 {
277         if (index < adev->doorbell.num_doorbells) {
278                 atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v);
279         } else {
280                 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
281         }
282 }
283
284 /**
285  * amdgpu_invalid_rreg - dummy reg read function
286  *
287  * @adev: amdgpu device pointer
288  * @reg: offset of register
289  *
290  * Dummy register read function.  Used for register blocks
291  * that certain asics don't have (all asics).
292  * Returns the value in the register.
293  */
294 static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
295 {
296         DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
297         BUG();
298         return 0;
299 }
300
301 /**
302  * amdgpu_invalid_wreg - dummy reg write function
303  *
304  * @adev: amdgpu device pointer
305  * @reg: offset of register
306  * @v: value to write to the register
307  *
308  * Dummy register read function.  Used for register blocks
309  * that certain asics don't have (all asics).
310  */
311 static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
312 {
313         DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
314                   reg, v);
315         BUG();
316 }
317
318 /**
319  * amdgpu_block_invalid_rreg - dummy reg read function
320  *
321  * @adev: amdgpu device pointer
322  * @block: offset of instance
323  * @reg: offset of register
324  *
325  * Dummy register read function.  Used for register blocks
326  * that certain asics don't have (all asics).
327  * Returns the value in the register.
328  */
329 static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
330                                           uint32_t block, uint32_t reg)
331 {
332         DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
333                   reg, block);
334         BUG();
335         return 0;
336 }
337
338 /**
339  * amdgpu_block_invalid_wreg - dummy reg write function
340  *
341  * @adev: amdgpu device pointer
342  * @block: offset of instance
343  * @reg: offset of register
344  * @v: value to write to the register
345  *
346  * Dummy register read function.  Used for register blocks
347  * that certain asics don't have (all asics).
348  */
349 static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
350                                       uint32_t block,
351                                       uint32_t reg, uint32_t v)
352 {
353         DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
354                   reg, block, v);
355         BUG();
356 }
357
358 static int amdgpu_device_vram_scratch_init(struct amdgpu_device *adev)
359 {
360         return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE,
361                                        PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
362                                        &adev->vram_scratch.robj,
363                                        &adev->vram_scratch.gpu_addr,
364                                        (void **)&adev->vram_scratch.ptr);
365 }
366
367 static void amdgpu_device_vram_scratch_fini(struct amdgpu_device *adev)
368 {
369         amdgpu_bo_free_kernel(&adev->vram_scratch.robj, NULL, NULL);
370 }
371
372 /**
373  * amdgpu_device_program_register_sequence - program an array of registers.
374  *
375  * @adev: amdgpu_device pointer
376  * @registers: pointer to the register array
377  * @array_size: size of the register array
378  *
379  * Programs an array or registers with and and or masks.
380  * This is a helper for setting golden registers.
381  */
382 void amdgpu_device_program_register_sequence(struct amdgpu_device *adev,
383                                              const u32 *registers,
384                                              const u32 array_size)
385 {
386         u32 tmp, reg, and_mask, or_mask;
387         int i;
388
389         if (array_size % 3)
390                 return;
391
392         for (i = 0; i < array_size; i +=3) {
393                 reg = registers[i + 0];
394                 and_mask = registers[i + 1];
395                 or_mask = registers[i + 2];
396
397                 if (and_mask == 0xffffffff) {
398                         tmp = or_mask;
399                 } else {
400                         tmp = RREG32(reg);
401                         tmp &= ~and_mask;
402                         tmp |= or_mask;
403                 }
404                 WREG32(reg, tmp);
405         }
406 }
407
408 void amdgpu_device_pci_config_reset(struct amdgpu_device *adev)
409 {
410         pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
411 }
412
413 /*
414  * GPU doorbell aperture helpers function.
415  */
416 /**
417  * amdgpu_device_doorbell_init - Init doorbell driver information.
418  *
419  * @adev: amdgpu_device pointer
420  *
421  * Init doorbell driver information (CIK)
422  * Returns 0 on success, error on failure.
423  */
424 static int amdgpu_device_doorbell_init(struct amdgpu_device *adev)
425 {
426         /* No doorbell on SI hardware generation */
427         if (adev->asic_type < CHIP_BONAIRE) {
428                 adev->doorbell.base = 0;
429                 adev->doorbell.size = 0;
430                 adev->doorbell.num_doorbells = 0;
431                 adev->doorbell.ptr = NULL;
432                 return 0;
433         }
434
435         if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET)
436                 return -EINVAL;
437
438         /* doorbell bar mapping */
439         adev->doorbell.base = pci_resource_start(adev->pdev, 2);
440         adev->doorbell.size = pci_resource_len(adev->pdev, 2);
441
442         adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32),
443                                              AMDGPU_DOORBELL_MAX_ASSIGNMENT+1);
444         if (adev->doorbell.num_doorbells == 0)
445                 return -EINVAL;
446
447         adev->doorbell.ptr = ioremap(adev->doorbell.base,
448                                      adev->doorbell.num_doorbells *
449                                      sizeof(u32));
450         if (adev->doorbell.ptr == NULL)
451                 return -ENOMEM;
452
453         return 0;
454 }
455
456 /**
457  * amdgpu_device_doorbell_fini - Tear down doorbell driver information.
458  *
459  * @adev: amdgpu_device pointer
460  *
461  * Tear down doorbell driver information (CIK)
462  */
463 static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev)
464 {
465         iounmap(adev->doorbell.ptr);
466         adev->doorbell.ptr = NULL;
467 }
468
469
470
471 /*
472  * amdgpu_device_wb_*()
473  * Writeback is the method by which the GPU updates special pages in memory
474  * with the status of certain GPU events (fences, ring pointers,etc.).
475  */
476
477 /**
478  * amdgpu_device_wb_fini - Disable Writeback and free memory
479  *
480  * @adev: amdgpu_device pointer
481  *
482  * Disables Writeback and frees the Writeback memory (all asics).
483  * Used at driver shutdown.
484  */
485 static void amdgpu_device_wb_fini(struct amdgpu_device *adev)
486 {
487         if (adev->wb.wb_obj) {
488                 amdgpu_bo_free_kernel(&adev->wb.wb_obj,
489                                       &adev->wb.gpu_addr,
490                                       (void **)&adev->wb.wb);
491                 adev->wb.wb_obj = NULL;
492         }
493 }
494
495 /**
496  * amdgpu_device_wb_init- Init Writeback driver info and allocate memory
497  *
498  * @adev: amdgpu_device pointer
499  *
500  * Initializes writeback and allocates writeback memory (all asics).
501  * Used at driver startup.
502  * Returns 0 on success or an -error on failure.
503  */
504 static int amdgpu_device_wb_init(struct amdgpu_device *adev)
505 {
506         int r;
507
508         if (adev->wb.wb_obj == NULL) {
509                 /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */
510                 r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8,
511                                             PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
512                                             &adev->wb.wb_obj, &adev->wb.gpu_addr,
513                                             (void **)&adev->wb.wb);
514                 if (r) {
515                         dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
516                         return r;
517                 }
518
519                 adev->wb.num_wb = AMDGPU_MAX_WB;
520                 memset(&adev->wb.used, 0, sizeof(adev->wb.used));
521
522                 /* clear wb memory */
523                 memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8);
524         }
525
526         return 0;
527 }
528
529 /**
530  * amdgpu_device_wb_get - Allocate a wb entry
531  *
532  * @adev: amdgpu_device pointer
533  * @wb: wb index
534  *
535  * Allocate a wb slot for use by the driver (all asics).
536  * Returns 0 on success or -EINVAL on failure.
537  */
538 int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb)
539 {
540         unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
541
542         if (offset < adev->wb.num_wb) {
543                 __set_bit(offset, adev->wb.used);
544                 *wb = offset << 3; /* convert to dw offset */
545                 return 0;
546         } else {
547                 return -EINVAL;
548         }
549 }
550
551 /**
552  * amdgpu_device_wb_free - Free a wb entry
553  *
554  * @adev: amdgpu_device pointer
555  * @wb: wb index
556  *
557  * Free a wb slot allocated for use by the driver (all asics)
558  */
559 void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb)
560 {
561         wb >>= 3;
562         if (wb < adev->wb.num_wb)
563                 __clear_bit(wb, adev->wb.used);
564 }
565
566 /**
567  * amdgpu_device_vram_location - try to find VRAM location
568  * @adev: amdgpu device structure holding all necessary informations
569  * @mc: memory controller structure holding memory informations
570  * @base: base address at which to put VRAM
571  *
572  * Function will try to place VRAM at base address provided
573  * as parameter.
574  */
575 void amdgpu_device_vram_location(struct amdgpu_device *adev,
576                                  struct amdgpu_gmc *mc, u64 base)
577 {
578         uint64_t limit = (uint64_t)amdgpu_vram_limit << 20;
579
580         mc->vram_start = base;
581         mc->vram_end = mc->vram_start + mc->mc_vram_size - 1;
582         if (limit && limit < mc->real_vram_size)
583                 mc->real_vram_size = limit;
584         dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
585                         mc->mc_vram_size >> 20, mc->vram_start,
586                         mc->vram_end, mc->real_vram_size >> 20);
587 }
588
589 /**
590  * amdgpu_device_gart_location - try to find GTT location
591  * @adev: amdgpu device structure holding all necessary informations
592  * @mc: memory controller structure holding memory informations
593  *
594  * Function will place try to place GTT before or after VRAM.
595  *
596  * If GTT size is bigger than space left then we ajust GTT size.
597  * Thus function will never fails.
598  *
599  * FIXME: when reducing GTT size align new size on power of 2.
600  */
601 void amdgpu_device_gart_location(struct amdgpu_device *adev,
602                                  struct amdgpu_gmc *mc)
603 {
604         u64 size_af, size_bf;
605
606         size_af = adev->gmc.mc_mask - mc->vram_end;
607         size_bf = mc->vram_start;
608         if (size_bf > size_af) {
609                 if (mc->gart_size > size_bf) {
610                         dev_warn(adev->dev, "limiting GTT\n");
611                         mc->gart_size = size_bf;
612                 }
613                 mc->gart_start = 0;
614         } else {
615                 if (mc->gart_size > size_af) {
616                         dev_warn(adev->dev, "limiting GTT\n");
617                         mc->gart_size = size_af;
618                 }
619                 /* VCE doesn't like it when BOs cross a 4GB segment, so align
620                  * the GART base on a 4GB boundary as well.
621                  */
622                 mc->gart_start = ALIGN(mc->vram_end + 1, 0x100000000ULL);
623         }
624         mc->gart_end = mc->gart_start + mc->gart_size - 1;
625         dev_info(adev->dev, "GTT: %lluM 0x%016llX - 0x%016llX\n",
626                         mc->gart_size >> 20, mc->gart_start, mc->gart_end);
627 }
628
629 /**
630  * amdgpu_device_resize_fb_bar - try to resize FB BAR
631  *
632  * @adev: amdgpu_device pointer
633  *
634  * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not
635  * to fail, but if any of the BARs is not accessible after the size we abort
636  * driver loading by returning -ENODEV.
637  */
638 int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev)
639 {
640         u64 space_needed = roundup_pow_of_two(adev->gmc.real_vram_size);
641         u32 rbar_size = order_base_2(((space_needed >> 20) | 1)) - 1;
642         struct pci_bus *root;
643         struct resource *res;
644         unsigned i;
645         u16 cmd;
646         int r;
647
648         /* Bypass for VF */
649         if (amdgpu_sriov_vf(adev))
650                 return 0;
651
652         /* Check if the root BUS has 64bit memory resources */
653         root = adev->pdev->bus;
654         while (root->parent)
655                 root = root->parent;
656
657         pci_bus_for_each_resource(root, res, i) {
658                 if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) &&
659                     res->start > 0x100000000ull)
660                         break;
661         }
662
663         /* Trying to resize is pointless without a root hub window above 4GB */
664         if (!res)
665                 return 0;
666
667         /* Disable memory decoding while we change the BAR addresses and size */
668         pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd);
669         pci_write_config_word(adev->pdev, PCI_COMMAND,
670                               cmd & ~PCI_COMMAND_MEMORY);
671
672         /* Free the VRAM and doorbell BAR, we most likely need to move both. */
673         amdgpu_device_doorbell_fini(adev);
674         if (adev->asic_type >= CHIP_BONAIRE)
675                 pci_release_resource(adev->pdev, 2);
676
677         pci_release_resource(adev->pdev, 0);
678
679         r = pci_resize_resource(adev->pdev, 0, rbar_size);
680         if (r == -ENOSPC)
681                 DRM_INFO("Not enough PCI address space for a large BAR.");
682         else if (r && r != -ENOTSUPP)
683                 DRM_ERROR("Problem resizing BAR0 (%d).", r);
684
685         pci_assign_unassigned_bus_resources(adev->pdev->bus);
686
687         /* When the doorbell or fb BAR isn't available we have no chance of
688          * using the device.
689          */
690         r = amdgpu_device_doorbell_init(adev);
691         if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET))
692                 return -ENODEV;
693
694         pci_write_config_word(adev->pdev, PCI_COMMAND, cmd);
695
696         return 0;
697 }
698
699 /*
700  * GPU helpers function.
701  */
702 /**
703  * amdgpu_device_need_post - check if the hw need post or not
704  *
705  * @adev: amdgpu_device pointer
706  *
707  * Check if the asic has been initialized (all asics) at driver startup
708  * or post is needed if  hw reset is performed.
709  * Returns true if need or false if not.
710  */
711 bool amdgpu_device_need_post(struct amdgpu_device *adev)
712 {
713         uint32_t reg;
714
715         if (amdgpu_sriov_vf(adev))
716                 return false;
717
718         if (amdgpu_passthrough(adev)) {
719                 /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
720                  * some old smc fw still need driver do vPost otherwise gpu hang, while
721                  * those smc fw version above 22.15 doesn't have this flaw, so we force
722                  * vpost executed for smc version below 22.15
723                  */
724                 if (adev->asic_type == CHIP_FIJI) {
725                         int err;
726                         uint32_t fw_ver;
727                         err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
728                         /* force vPost if error occured */
729                         if (err)
730                                 return true;
731
732                         fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
733                         if (fw_ver < 0x00160e00)
734                                 return true;
735                 }
736         }
737
738         if (adev->has_hw_reset) {
739                 adev->has_hw_reset = false;
740                 return true;
741         }
742
743         /* bios scratch used on CIK+ */
744         if (adev->asic_type >= CHIP_BONAIRE)
745                 return amdgpu_atombios_scratch_need_asic_init(adev);
746
747         /* check MEM_SIZE for older asics */
748         reg = amdgpu_asic_get_config_memsize(adev);
749
750         if ((reg != 0) && (reg != 0xffffffff))
751                 return false;
752
753         return true;
754 }
755
756 /* if we get transitioned to only one device, take VGA back */
757 /**
758  * amdgpu_device_vga_set_decode - enable/disable vga decode
759  *
760  * @cookie: amdgpu_device pointer
761  * @state: enable/disable vga decode
762  *
763  * Enable/disable vga decode (all asics).
764  * Returns VGA resource flags.
765  */
766 static unsigned int amdgpu_device_vga_set_decode(void *cookie, bool state)
767 {
768         struct amdgpu_device *adev = cookie;
769         amdgpu_asic_set_vga_state(adev, state);
770         if (state)
771                 return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
772                        VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
773         else
774                 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
775 }
776
777 static void amdgpu_device_check_block_size(struct amdgpu_device *adev)
778 {
779         /* defines number of bits in page table versus page directory,
780          * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
781          * page table and the remaining bits are in the page directory */
782         if (amdgpu_vm_block_size == -1)
783                 return;
784
785         if (amdgpu_vm_block_size < 9) {
786                 dev_warn(adev->dev, "VM page table size (%d) too small\n",
787                          amdgpu_vm_block_size);
788                 amdgpu_vm_block_size = -1;
789         }
790 }
791
792 static void amdgpu_device_check_vm_size(struct amdgpu_device *adev)
793 {
794         /* no need to check the default value */
795         if (amdgpu_vm_size == -1)
796                 return;
797
798         if (amdgpu_vm_size < 1) {
799                 dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
800                          amdgpu_vm_size);
801                 amdgpu_vm_size = -1;
802         }
803 }
804
805 /**
806  * amdgpu_device_check_arguments - validate module params
807  *
808  * @adev: amdgpu_device pointer
809  *
810  * Validates certain module parameters and updates
811  * the associated values used by the driver (all asics).
812  */
813 static void amdgpu_device_check_arguments(struct amdgpu_device *adev)
814 {
815         if (amdgpu_sched_jobs < 4) {
816                 dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
817                          amdgpu_sched_jobs);
818                 amdgpu_sched_jobs = 4;
819         } else if (!is_power_of_2(amdgpu_sched_jobs)){
820                 dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
821                          amdgpu_sched_jobs);
822                 amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
823         }
824
825         if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) {
826                 /* gart size must be greater or equal to 32M */
827                 dev_warn(adev->dev, "gart size (%d) too small\n",
828                          amdgpu_gart_size);
829                 amdgpu_gart_size = -1;
830         }
831
832         if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) {
833                 /* gtt size must be greater or equal to 32M */
834                 dev_warn(adev->dev, "gtt size (%d) too small\n",
835                                  amdgpu_gtt_size);
836                 amdgpu_gtt_size = -1;
837         }
838
839         /* valid range is between 4 and 9 inclusive */
840         if (amdgpu_vm_fragment_size != -1 &&
841             (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) {
842                 dev_warn(adev->dev, "valid range is between 4 and 9\n");
843                 amdgpu_vm_fragment_size = -1;
844         }
845
846         amdgpu_device_check_vm_size(adev);
847
848         amdgpu_device_check_block_size(adev);
849
850         if (amdgpu_vram_page_split != -1 && (amdgpu_vram_page_split < 16 ||
851             !is_power_of_2(amdgpu_vram_page_split))) {
852                 dev_warn(adev->dev, "invalid VRAM page split (%d)\n",
853                          amdgpu_vram_page_split);
854                 amdgpu_vram_page_split = 1024;
855         }
856
857         if (amdgpu_lockup_timeout == 0) {
858                 dev_warn(adev->dev, "lockup_timeout msut be > 0, adjusting to 10000\n");
859                 amdgpu_lockup_timeout = 10000;
860         }
861
862         adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type);
863 }
864
865 /**
866  * amdgpu_switcheroo_set_state - set switcheroo state
867  *
868  * @pdev: pci dev pointer
869  * @state: vga_switcheroo state
870  *
871  * Callback for the switcheroo driver.  Suspends or resumes the
872  * the asics before or after it is powered up using ACPI methods.
873  */
874 static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, enum vga_switcheroo_state state)
875 {
876         struct drm_device *dev = pci_get_drvdata(pdev);
877
878         if (amdgpu_device_is_px(dev) && state == VGA_SWITCHEROO_OFF)
879                 return;
880
881         if (state == VGA_SWITCHEROO_ON) {
882                 pr_info("amdgpu: switched on\n");
883                 /* don't suspend or resume card normally */
884                 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
885
886                 amdgpu_device_resume(dev, true, true);
887
888                 dev->switch_power_state = DRM_SWITCH_POWER_ON;
889                 drm_kms_helper_poll_enable(dev);
890         } else {
891                 pr_info("amdgpu: switched off\n");
892                 drm_kms_helper_poll_disable(dev);
893                 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
894                 amdgpu_device_suspend(dev, true, true);
895                 dev->switch_power_state = DRM_SWITCH_POWER_OFF;
896         }
897 }
898
899 /**
900  * amdgpu_switcheroo_can_switch - see if switcheroo state can change
901  *
902  * @pdev: pci dev pointer
903  *
904  * Callback for the switcheroo driver.  Check of the switcheroo
905  * state can be changed.
906  * Returns true if the state can be changed, false if not.
907  */
908 static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
909 {
910         struct drm_device *dev = pci_get_drvdata(pdev);
911
912         /*
913         * FIXME: open_count is protected by drm_global_mutex but that would lead to
914         * locking inversion with the driver load path. And the access here is
915         * completely racy anyway. So don't bother with locking for now.
916         */
917         return dev->open_count == 0;
918 }
919
920 static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
921         .set_gpu_state = amdgpu_switcheroo_set_state,
922         .reprobe = NULL,
923         .can_switch = amdgpu_switcheroo_can_switch,
924 };
925
926 int amdgpu_device_ip_set_clockgating_state(struct amdgpu_device *adev,
927                                            enum amd_ip_block_type block_type,
928                                            enum amd_clockgating_state state)
929 {
930         int i, r = 0;
931
932         for (i = 0; i < adev->num_ip_blocks; i++) {
933                 if (!adev->ip_blocks[i].status.valid)
934                         continue;
935                 if (adev->ip_blocks[i].version->type != block_type)
936                         continue;
937                 if (!adev->ip_blocks[i].version->funcs->set_clockgating_state)
938                         continue;
939                 r = adev->ip_blocks[i].version->funcs->set_clockgating_state(
940                         (void *)adev, state);
941                 if (r)
942                         DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",
943                                   adev->ip_blocks[i].version->funcs->name, r);
944         }
945         return r;
946 }
947
948 int amdgpu_device_ip_set_powergating_state(struct amdgpu_device *adev,
949                                            enum amd_ip_block_type block_type,
950                                            enum amd_powergating_state state)
951 {
952         int i, r = 0;
953
954         for (i = 0; i < adev->num_ip_blocks; i++) {
955                 if (!adev->ip_blocks[i].status.valid)
956                         continue;
957                 if (adev->ip_blocks[i].version->type != block_type)
958                         continue;
959                 if (!adev->ip_blocks[i].version->funcs->set_powergating_state)
960                         continue;
961                 r = adev->ip_blocks[i].version->funcs->set_powergating_state(
962                         (void *)adev, state);
963                 if (r)
964                         DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",
965                                   adev->ip_blocks[i].version->funcs->name, r);
966         }
967         return r;
968 }
969
970 void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev,
971                                             u32 *flags)
972 {
973         int i;
974
975         for (i = 0; i < adev->num_ip_blocks; i++) {
976                 if (!adev->ip_blocks[i].status.valid)
977                         continue;
978                 if (adev->ip_blocks[i].version->funcs->get_clockgating_state)
979                         adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags);
980         }
981 }
982
983 int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev,
984                                    enum amd_ip_block_type block_type)
985 {
986         int i, r;
987
988         for (i = 0; i < adev->num_ip_blocks; i++) {
989                 if (!adev->ip_blocks[i].status.valid)
990                         continue;
991                 if (adev->ip_blocks[i].version->type == block_type) {
992                         r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
993                         if (r)
994                                 return r;
995                         break;
996                 }
997         }
998         return 0;
999
1000 }
1001
1002 bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev,
1003                               enum amd_ip_block_type block_type)
1004 {
1005         int i;
1006
1007         for (i = 0; i < adev->num_ip_blocks; i++) {
1008                 if (!adev->ip_blocks[i].status.valid)
1009                         continue;
1010                 if (adev->ip_blocks[i].version->type == block_type)
1011                         return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
1012         }
1013         return true;
1014
1015 }
1016
1017 struct amdgpu_ip_block *
1018 amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,
1019                               enum amd_ip_block_type type)
1020 {
1021         int i;
1022
1023         for (i = 0; i < adev->num_ip_blocks; i++)
1024                 if (adev->ip_blocks[i].version->type == type)
1025                         return &adev->ip_blocks[i];
1026
1027         return NULL;
1028 }
1029
1030 /**
1031  * amdgpu_device_ip_block_version_cmp
1032  *
1033  * @adev: amdgpu_device pointer
1034  * @type: enum amd_ip_block_type
1035  * @major: major version
1036  * @minor: minor version
1037  *
1038  * return 0 if equal or greater
1039  * return 1 if smaller or the ip_block doesn't exist
1040  */
1041 int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev,
1042                                        enum amd_ip_block_type type,
1043                                        u32 major, u32 minor)
1044 {
1045         struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type);
1046
1047         if (ip_block && ((ip_block->version->major > major) ||
1048                         ((ip_block->version->major == major) &&
1049                         (ip_block->version->minor >= minor))))
1050                 return 0;
1051
1052         return 1;
1053 }
1054
1055 /**
1056  * amdgpu_device_ip_block_add
1057  *
1058  * @adev: amdgpu_device pointer
1059  * @ip_block_version: pointer to the IP to add
1060  *
1061  * Adds the IP block driver information to the collection of IPs
1062  * on the asic.
1063  */
1064 int amdgpu_device_ip_block_add(struct amdgpu_device *adev,
1065                                const struct amdgpu_ip_block_version *ip_block_version)
1066 {
1067         if (!ip_block_version)
1068                 return -EINVAL;
1069
1070         DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks,
1071                   ip_block_version->funcs->name);
1072
1073         adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
1074
1075         return 0;
1076 }
1077
1078 static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
1079 {
1080         adev->enable_virtual_display = false;
1081
1082         if (amdgpu_virtual_display) {
1083                 struct drm_device *ddev = adev->ddev;
1084                 const char *pci_address_name = pci_name(ddev->pdev);
1085                 char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
1086
1087                 pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
1088                 pciaddstr_tmp = pciaddstr;
1089                 while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
1090                         pciaddname = strsep(&pciaddname_tmp, ",");
1091                         if (!strcmp("all", pciaddname)
1092                             || !strcmp(pci_address_name, pciaddname)) {
1093                                 long num_crtc;
1094                                 int res = -1;
1095
1096                                 adev->enable_virtual_display = true;
1097
1098                                 if (pciaddname_tmp)
1099                                         res = kstrtol(pciaddname_tmp, 10,
1100                                                       &num_crtc);
1101
1102                                 if (!res) {
1103                                         if (num_crtc < 1)
1104                                                 num_crtc = 1;
1105                                         if (num_crtc > 6)
1106                                                 num_crtc = 6;
1107                                         adev->mode_info.num_crtc = num_crtc;
1108                                 } else {
1109                                         adev->mode_info.num_crtc = 1;
1110                                 }
1111                                 break;
1112                         }
1113                 }
1114
1115                 DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
1116                          amdgpu_virtual_display, pci_address_name,
1117                          adev->enable_virtual_display, adev->mode_info.num_crtc);
1118
1119                 kfree(pciaddstr);
1120         }
1121 }
1122
1123 static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev)
1124 {
1125         const char *chip_name;
1126         char fw_name[30];
1127         int err;
1128         const struct gpu_info_firmware_header_v1_0 *hdr;
1129
1130         adev->firmware.gpu_info_fw = NULL;
1131
1132         switch (adev->asic_type) {
1133         case CHIP_TOPAZ:
1134         case CHIP_TONGA:
1135         case CHIP_FIJI:
1136         case CHIP_POLARIS11:
1137         case CHIP_POLARIS10:
1138         case CHIP_POLARIS12:
1139         case CHIP_CARRIZO:
1140         case CHIP_STONEY:
1141 #ifdef CONFIG_DRM_AMDGPU_SI
1142         case CHIP_VERDE:
1143         case CHIP_TAHITI:
1144         case CHIP_PITCAIRN:
1145         case CHIP_OLAND:
1146         case CHIP_HAINAN:
1147 #endif
1148 #ifdef CONFIG_DRM_AMDGPU_CIK
1149         case CHIP_BONAIRE:
1150         case CHIP_HAWAII:
1151         case CHIP_KAVERI:
1152         case CHIP_KABINI:
1153         case CHIP_MULLINS:
1154 #endif
1155         default:
1156                 return 0;
1157         case CHIP_VEGA10:
1158                 chip_name = "vega10";
1159                 break;
1160         case CHIP_RAVEN:
1161                 chip_name = "raven";
1162                 break;
1163         }
1164
1165         snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name);
1166         err = request_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev);
1167         if (err) {
1168                 dev_err(adev->dev,
1169                         "Failed to load gpu_info firmware \"%s\"\n",
1170                         fw_name);
1171                 goto out;
1172         }
1173         err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw);
1174         if (err) {
1175                 dev_err(adev->dev,
1176                         "Failed to validate gpu_info firmware \"%s\"\n",
1177                         fw_name);
1178                 goto out;
1179         }
1180
1181         hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data;
1182         amdgpu_ucode_print_gpu_info_hdr(&hdr->header);
1183
1184         switch (hdr->version_major) {
1185         case 1:
1186         {
1187                 const struct gpu_info_firmware_v1_0 *gpu_info_fw =
1188                         (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data +
1189                                                                 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1190
1191                 adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se);
1192                 adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh);
1193                 adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se);
1194                 adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se);
1195                 adev->gfx.config.max_texture_channel_caches =
1196                         le32_to_cpu(gpu_info_fw->gc_num_tccs);
1197                 adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs);
1198                 adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds);
1199                 adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth);
1200                 adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth);
1201                 adev->gfx.config.double_offchip_lds_buf =
1202                         le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer);
1203                 adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size);
1204                 adev->gfx.cu_info.max_waves_per_simd =
1205                         le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd);
1206                 adev->gfx.cu_info.max_scratch_slots_per_cu =
1207                         le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu);
1208                 adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size);
1209                 break;
1210         }
1211         default:
1212                 dev_err(adev->dev,
1213                         "Unsupported gpu_info table %d\n", hdr->header.ucode_version);
1214                 err = -EINVAL;
1215                 goto out;
1216         }
1217 out:
1218         return err;
1219 }
1220
1221 static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
1222 {
1223         int i, r;
1224
1225         amdgpu_device_enable_virtual_display(adev);
1226
1227         switch (adev->asic_type) {
1228         case CHIP_TOPAZ:
1229         case CHIP_TONGA:
1230         case CHIP_FIJI:
1231         case CHIP_POLARIS11:
1232         case CHIP_POLARIS10:
1233         case CHIP_POLARIS12:
1234         case CHIP_CARRIZO:
1235         case CHIP_STONEY:
1236                 if (adev->asic_type == CHIP_CARRIZO || adev->asic_type == CHIP_STONEY)
1237                         adev->family = AMDGPU_FAMILY_CZ;
1238                 else
1239                         adev->family = AMDGPU_FAMILY_VI;
1240
1241                 r = vi_set_ip_blocks(adev);
1242                 if (r)
1243                         return r;
1244                 break;
1245 #ifdef CONFIG_DRM_AMDGPU_SI
1246         case CHIP_VERDE:
1247         case CHIP_TAHITI:
1248         case CHIP_PITCAIRN:
1249         case CHIP_OLAND:
1250         case CHIP_HAINAN:
1251                 adev->family = AMDGPU_FAMILY_SI;
1252                 r = si_set_ip_blocks(adev);
1253                 if (r)
1254                         return r;
1255                 break;
1256 #endif
1257 #ifdef CONFIG_DRM_AMDGPU_CIK
1258         case CHIP_BONAIRE:
1259         case CHIP_HAWAII:
1260         case CHIP_KAVERI:
1261         case CHIP_KABINI:
1262         case CHIP_MULLINS:
1263                 if ((adev->asic_type == CHIP_BONAIRE) || (adev->asic_type == CHIP_HAWAII))
1264                         adev->family = AMDGPU_FAMILY_CI;
1265                 else
1266                         adev->family = AMDGPU_FAMILY_KV;
1267
1268                 r = cik_set_ip_blocks(adev);
1269                 if (r)
1270                         return r;
1271                 break;
1272 #endif
1273         case  CHIP_VEGA10:
1274         case  CHIP_RAVEN:
1275                 if (adev->asic_type == CHIP_RAVEN)
1276                         adev->family = AMDGPU_FAMILY_RV;
1277                 else
1278                         adev->family = AMDGPU_FAMILY_AI;
1279
1280                 r = soc15_set_ip_blocks(adev);
1281                 if (r)
1282                         return r;
1283                 break;
1284         default:
1285                 /* FIXME: not supported yet */
1286                 return -EINVAL;
1287         }
1288
1289         r = amdgpu_device_parse_gpu_info_fw(adev);
1290         if (r)
1291                 return r;
1292
1293         amdgpu_amdkfd_device_probe(adev);
1294
1295         if (amdgpu_sriov_vf(adev)) {
1296                 r = amdgpu_virt_request_full_gpu(adev, true);
1297                 if (r)
1298                         return -EAGAIN;
1299         }
1300
1301         for (i = 0; i < adev->num_ip_blocks; i++) {
1302                 if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
1303                         DRM_ERROR("disabled ip block: %d <%s>\n",
1304                                   i, adev->ip_blocks[i].version->funcs->name);
1305                         adev->ip_blocks[i].status.valid = false;
1306                 } else {
1307                         if (adev->ip_blocks[i].version->funcs->early_init) {
1308                                 r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
1309                                 if (r == -ENOENT) {
1310                                         adev->ip_blocks[i].status.valid = false;
1311                                 } else if (r) {
1312                                         DRM_ERROR("early_init of IP block <%s> failed %d\n",
1313                                                   adev->ip_blocks[i].version->funcs->name, r);
1314                                         return r;
1315                                 } else {
1316                                         adev->ip_blocks[i].status.valid = true;
1317                                 }
1318                         } else {
1319                                 adev->ip_blocks[i].status.valid = true;
1320                         }
1321                 }
1322         }
1323
1324         adev->cg_flags &= amdgpu_cg_mask;
1325         adev->pg_flags &= amdgpu_pg_mask;
1326
1327         return 0;
1328 }
1329
1330 static int amdgpu_device_ip_init(struct amdgpu_device *adev)
1331 {
1332         int i, r;
1333
1334         for (i = 0; i < adev->num_ip_blocks; i++) {
1335                 if (!adev->ip_blocks[i].status.valid)
1336                         continue;
1337                 r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
1338                 if (r) {
1339                         DRM_ERROR("sw_init of IP block <%s> failed %d\n",
1340                                   adev->ip_blocks[i].version->funcs->name, r);
1341                         return r;
1342                 }
1343                 adev->ip_blocks[i].status.sw = true;
1344
1345                 /* need to do gmc hw init early so we can allocate gpu mem */
1346                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1347                         r = amdgpu_device_vram_scratch_init(adev);
1348                         if (r) {
1349                                 DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r);
1350                                 return r;
1351                         }
1352                         r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1353                         if (r) {
1354                                 DRM_ERROR("hw_init %d failed %d\n", i, r);
1355                                 return r;
1356                         }
1357                         r = amdgpu_device_wb_init(adev);
1358                         if (r) {
1359                                 DRM_ERROR("amdgpu_device_wb_init failed %d\n", r);
1360                                 return r;
1361                         }
1362                         adev->ip_blocks[i].status.hw = true;
1363
1364                         /* right after GMC hw init, we create CSA */
1365                         if (amdgpu_sriov_vf(adev)) {
1366                                 r = amdgpu_allocate_static_csa(adev);
1367                                 if (r) {
1368                                         DRM_ERROR("allocate CSA failed %d\n", r);
1369                                         return r;
1370                                 }
1371                         }
1372                 }
1373         }
1374
1375         for (i = 0; i < adev->num_ip_blocks; i++) {
1376                 if (!adev->ip_blocks[i].status.sw)
1377                         continue;
1378                 if (adev->ip_blocks[i].status.hw)
1379                         continue;
1380                 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1381                 if (r) {
1382                         DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1383                                   adev->ip_blocks[i].version->funcs->name, r);
1384                         return r;
1385                 }
1386                 adev->ip_blocks[i].status.hw = true;
1387         }
1388
1389         amdgpu_amdkfd_device_init(adev);
1390
1391         if (amdgpu_sriov_vf(adev))
1392                 amdgpu_virt_release_full_gpu(adev, true);
1393
1394         return 0;
1395 }
1396
1397 static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev)
1398 {
1399         memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
1400 }
1401
1402 static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev)
1403 {
1404         return !!memcmp(adev->gart.ptr, adev->reset_magic,
1405                         AMDGPU_RESET_MAGIC_NUM);
1406 }
1407
1408 static int amdgpu_device_ip_late_set_cg_state(struct amdgpu_device *adev)
1409 {
1410         int i = 0, r;
1411
1412         if (amdgpu_emu_mode == 1)
1413                 return 0;
1414
1415         for (i = 0; i < adev->num_ip_blocks; i++) {
1416                 if (!adev->ip_blocks[i].status.valid)
1417                         continue;
1418                 /* skip CG for VCE/UVD, it's handled specially */
1419                 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1420                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1421                     adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1422                         /* enable clockgating to save power */
1423                         r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1424                                                                                      AMD_CG_STATE_GATE);
1425                         if (r) {
1426                                 DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
1427                                           adev->ip_blocks[i].version->funcs->name, r);
1428                                 return r;
1429                         }
1430                 }
1431         }
1432         return 0;
1433 }
1434
1435 static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
1436 {
1437         int i = 0, r;
1438
1439         for (i = 0; i < adev->num_ip_blocks; i++) {
1440                 if (!adev->ip_blocks[i].status.valid)
1441                         continue;
1442                 if (adev->ip_blocks[i].version->funcs->late_init) {
1443                         r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
1444                         if (r) {
1445                                 DRM_ERROR("late_init of IP block <%s> failed %d\n",
1446                                           adev->ip_blocks[i].version->funcs->name, r);
1447                                 return r;
1448                         }
1449                         adev->ip_blocks[i].status.late_initialized = true;
1450                 }
1451         }
1452
1453         mod_delayed_work(system_wq, &adev->late_init_work,
1454                         msecs_to_jiffies(AMDGPU_RESUME_MS));
1455
1456         amdgpu_device_fill_reset_magic(adev);
1457
1458         return 0;
1459 }
1460
1461 static int amdgpu_device_ip_fini(struct amdgpu_device *adev)
1462 {
1463         int i, r;
1464
1465         amdgpu_amdkfd_device_fini(adev);
1466         /* need to disable SMC first */
1467         for (i = 0; i < adev->num_ip_blocks; i++) {
1468                 if (!adev->ip_blocks[i].status.hw)
1469                         continue;
1470                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC &&
1471                         adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1472                         /* ungate blocks before hw fini so that we can shutdown the blocks safely */
1473                         r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1474                                                                                      AMD_CG_STATE_UNGATE);
1475                         if (r) {
1476                                 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1477                                           adev->ip_blocks[i].version->funcs->name, r);
1478                                 return r;
1479                         }
1480                         r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
1481                         /* XXX handle errors */
1482                         if (r) {
1483                                 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
1484                                           adev->ip_blocks[i].version->funcs->name, r);
1485                         }
1486                         adev->ip_blocks[i].status.hw = false;
1487                         break;
1488                 }
1489         }
1490
1491         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1492                 if (!adev->ip_blocks[i].status.hw)
1493                         continue;
1494
1495                 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1496                         adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE) {
1497                         /* ungate blocks before hw fini so that we can shutdown the blocks safely */
1498                         r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1499                                                                                      AMD_CG_STATE_UNGATE);
1500                         if (r) {
1501                                 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1502                                           adev->ip_blocks[i].version->funcs->name, r);
1503                                 return r;
1504                         }
1505                 }
1506
1507                 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
1508                 /* XXX handle errors */
1509                 if (r) {
1510                         DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
1511                                   adev->ip_blocks[i].version->funcs->name, r);
1512                 }
1513
1514                 adev->ip_blocks[i].status.hw = false;
1515         }
1516
1517         /* disable all interrupts */
1518         amdgpu_irq_disable_all(adev);
1519
1520         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1521                 if (!adev->ip_blocks[i].status.sw)
1522                         continue;
1523
1524                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1525                         amdgpu_free_static_csa(adev);
1526                         amdgpu_device_wb_fini(adev);
1527                         amdgpu_device_vram_scratch_fini(adev);
1528                 }
1529
1530                 r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
1531                 /* XXX handle errors */
1532                 if (r) {
1533                         DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
1534                                   adev->ip_blocks[i].version->funcs->name, r);
1535                 }
1536                 adev->ip_blocks[i].status.sw = false;
1537                 adev->ip_blocks[i].status.valid = false;
1538         }
1539
1540         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1541                 if (!adev->ip_blocks[i].status.late_initialized)
1542                         continue;
1543                 if (adev->ip_blocks[i].version->funcs->late_fini)
1544                         adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
1545                 adev->ip_blocks[i].status.late_initialized = false;
1546         }
1547
1548         if (amdgpu_sriov_vf(adev))
1549                 if (amdgpu_virt_release_full_gpu(adev, false))
1550                         DRM_ERROR("failed to release exclusive mode on fini\n");
1551
1552         return 0;
1553 }
1554
1555 static void amdgpu_device_ip_late_init_func_handler(struct work_struct *work)
1556 {
1557         struct amdgpu_device *adev =
1558                 container_of(work, struct amdgpu_device, late_init_work.work);
1559         amdgpu_device_ip_late_set_cg_state(adev);
1560 }
1561
1562 int amdgpu_device_ip_suspend(struct amdgpu_device *adev)
1563 {
1564         int i, r;
1565
1566         if (amdgpu_sriov_vf(adev))
1567                 amdgpu_virt_request_full_gpu(adev, false);
1568
1569         /* ungate SMC block first */
1570         r = amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_SMC,
1571                                                    AMD_CG_STATE_UNGATE);
1572         if (r) {
1573                 DRM_ERROR("set_clockgating_state(ungate) SMC failed %d\n", r);
1574         }
1575
1576         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1577                 if (!adev->ip_blocks[i].status.valid)
1578                         continue;
1579                 /* ungate blocks so that suspend can properly shut them down */
1580                 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_SMC &&
1581                         adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1582                         r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1583                                                                                      AMD_CG_STATE_UNGATE);
1584                         if (r) {
1585                                 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1586                                           adev->ip_blocks[i].version->funcs->name, r);
1587                         }
1588                 }
1589                 /* XXX handle errors */
1590                 r = adev->ip_blocks[i].version->funcs->suspend(adev);
1591                 /* XXX handle errors */
1592                 if (r) {
1593                         DRM_ERROR("suspend of IP block <%s> failed %d\n",
1594                                   adev->ip_blocks[i].version->funcs->name, r);
1595                 }
1596         }
1597
1598         if (amdgpu_sriov_vf(adev))
1599                 amdgpu_virt_release_full_gpu(adev, false);
1600
1601         return 0;
1602 }
1603
1604 static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
1605 {
1606         int i, r;
1607
1608         static enum amd_ip_block_type ip_order[] = {
1609                 AMD_IP_BLOCK_TYPE_GMC,
1610                 AMD_IP_BLOCK_TYPE_COMMON,
1611                 AMD_IP_BLOCK_TYPE_IH,
1612         };
1613
1614         for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
1615                 int j;
1616                 struct amdgpu_ip_block *block;
1617
1618                 for (j = 0; j < adev->num_ip_blocks; j++) {
1619                         block = &adev->ip_blocks[j];
1620
1621                         if (block->version->type != ip_order[i] ||
1622                                 !block->status.valid)
1623                                 continue;
1624
1625                         r = block->version->funcs->hw_init(adev);
1626                         DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"successed");
1627                         if (r)
1628                                 return r;
1629                 }
1630         }
1631
1632         return 0;
1633 }
1634
1635 static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
1636 {
1637         int i, r;
1638
1639         static enum amd_ip_block_type ip_order[] = {
1640                 AMD_IP_BLOCK_TYPE_SMC,
1641                 AMD_IP_BLOCK_TYPE_PSP,
1642                 AMD_IP_BLOCK_TYPE_DCE,
1643                 AMD_IP_BLOCK_TYPE_GFX,
1644                 AMD_IP_BLOCK_TYPE_SDMA,
1645                 AMD_IP_BLOCK_TYPE_UVD,
1646                 AMD_IP_BLOCK_TYPE_VCE
1647         };
1648
1649         for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
1650                 int j;
1651                 struct amdgpu_ip_block *block;
1652
1653                 for (j = 0; j < adev->num_ip_blocks; j++) {
1654                         block = &adev->ip_blocks[j];
1655
1656                         if (block->version->type != ip_order[i] ||
1657                                 !block->status.valid)
1658                                 continue;
1659
1660                         r = block->version->funcs->hw_init(adev);
1661                         DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"successed");
1662                         if (r)
1663                                 return r;
1664                 }
1665         }
1666
1667         return 0;
1668 }
1669
1670 static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)
1671 {
1672         int i, r;
1673
1674         for (i = 0; i < adev->num_ip_blocks; i++) {
1675                 if (!adev->ip_blocks[i].status.valid)
1676                         continue;
1677                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
1678                                 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
1679                                 adev->ip_blocks[i].version->type ==
1680                                 AMD_IP_BLOCK_TYPE_IH) {
1681                         r = adev->ip_blocks[i].version->funcs->resume(adev);
1682                         if (r) {
1683                                 DRM_ERROR("resume of IP block <%s> failed %d\n",
1684                                           adev->ip_blocks[i].version->funcs->name, r);
1685                                 return r;
1686                         }
1687                 }
1688         }
1689
1690         return 0;
1691 }
1692
1693 static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)
1694 {
1695         int i, r;
1696
1697         for (i = 0; i < adev->num_ip_blocks; i++) {
1698                 if (!adev->ip_blocks[i].status.valid)
1699                         continue;
1700                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
1701                                 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
1702                                 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH )
1703                         continue;
1704                 r = adev->ip_blocks[i].version->funcs->resume(adev);
1705                 if (r) {
1706                         DRM_ERROR("resume of IP block <%s> failed %d\n",
1707                                   adev->ip_blocks[i].version->funcs->name, r);
1708                         return r;
1709                 }
1710         }
1711
1712         return 0;
1713 }
1714
1715 static int amdgpu_device_ip_resume(struct amdgpu_device *adev)
1716 {
1717         int r;
1718
1719         r = amdgpu_device_ip_resume_phase1(adev);
1720         if (r)
1721                 return r;
1722         r = amdgpu_device_ip_resume_phase2(adev);
1723
1724         return r;
1725 }
1726
1727 static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
1728 {
1729         if (amdgpu_sriov_vf(adev)) {
1730                 if (adev->is_atom_fw) {
1731                         if (amdgpu_atomfirmware_gpu_supports_virtualization(adev))
1732                                 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
1733                 } else {
1734                         if (amdgpu_atombios_has_gpu_virtualization_table(adev))
1735                                 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
1736                 }
1737
1738                 if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
1739                         amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
1740         }
1741 }
1742
1743 bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
1744 {
1745         switch (asic_type) {
1746 #if defined(CONFIG_DRM_AMD_DC)
1747         case CHIP_BONAIRE:
1748         case CHIP_HAWAII:
1749         case CHIP_KAVERI:
1750         case CHIP_KABINI:
1751         case CHIP_MULLINS:
1752         case CHIP_CARRIZO:
1753         case CHIP_STONEY:
1754         case CHIP_POLARIS11:
1755         case CHIP_POLARIS10:
1756         case CHIP_POLARIS12:
1757         case CHIP_TONGA:
1758         case CHIP_FIJI:
1759 #if defined(CONFIG_DRM_AMD_DC_PRE_VEGA)
1760                 return amdgpu_dc != 0;
1761 #endif
1762         case CHIP_VEGA10:
1763 #if defined(CONFIG_DRM_AMD_DC_DCN1_0)
1764         case CHIP_RAVEN:
1765 #endif
1766                 return amdgpu_dc != 0;
1767 #endif
1768         default:
1769                 return false;
1770         }
1771 }
1772
1773 /**
1774  * amdgpu_device_has_dc_support - check if dc is supported
1775  *
1776  * @adev: amdgpu_device_pointer
1777  *
1778  * Returns true for supported, false for not supported
1779  */
1780 bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
1781 {
1782         if (amdgpu_sriov_vf(adev))
1783                 return false;
1784
1785         return amdgpu_device_asic_has_dc_support(adev->asic_type);
1786 }
1787
1788 /**
1789  * amdgpu_device_init - initialize the driver
1790  *
1791  * @adev: amdgpu_device pointer
1792  * @pdev: drm dev pointer
1793  * @pdev: pci dev pointer
1794  * @flags: driver flags
1795  *
1796  * Initializes the driver info and hw (all asics).
1797  * Returns 0 for success or an error on failure.
1798  * Called at driver startup.
1799  */
1800 int amdgpu_device_init(struct amdgpu_device *adev,
1801                        struct drm_device *ddev,
1802                        struct pci_dev *pdev,
1803                        uint32_t flags)
1804 {
1805         int r, i;
1806         bool runtime = false;
1807         u32 max_MBps;
1808
1809         adev->shutdown = false;
1810         adev->dev = &pdev->dev;
1811         adev->ddev = ddev;
1812         adev->pdev = pdev;
1813         adev->flags = flags;
1814         adev->asic_type = flags & AMD_ASIC_MASK;
1815         adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
1816         if (amdgpu_emu_mode == 1)
1817                 adev->usec_timeout *= 2;
1818         adev->gmc.gart_size = 512 * 1024 * 1024;
1819         adev->accel_working = false;
1820         adev->num_rings = 0;
1821         adev->mman.buffer_funcs = NULL;
1822         adev->mman.buffer_funcs_ring = NULL;
1823         adev->vm_manager.vm_pte_funcs = NULL;
1824         adev->vm_manager.vm_pte_num_rings = 0;
1825         adev->gmc.gmc_funcs = NULL;
1826         adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
1827         bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
1828
1829         adev->smc_rreg = &amdgpu_invalid_rreg;
1830         adev->smc_wreg = &amdgpu_invalid_wreg;
1831         adev->pcie_rreg = &amdgpu_invalid_rreg;
1832         adev->pcie_wreg = &amdgpu_invalid_wreg;
1833         adev->pciep_rreg = &amdgpu_invalid_rreg;
1834         adev->pciep_wreg = &amdgpu_invalid_wreg;
1835         adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
1836         adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
1837         adev->didt_rreg = &amdgpu_invalid_rreg;
1838         adev->didt_wreg = &amdgpu_invalid_wreg;
1839         adev->gc_cac_rreg = &amdgpu_invalid_rreg;
1840         adev->gc_cac_wreg = &amdgpu_invalid_wreg;
1841         adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
1842         adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
1843
1844         DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
1845                  amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
1846                  pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
1847
1848         /* mutex initialization are all done here so we
1849          * can recall function without having locking issues */
1850         atomic_set(&adev->irq.ih.lock, 0);
1851         mutex_init(&adev->firmware.mutex);
1852         mutex_init(&adev->pm.mutex);
1853         mutex_init(&adev->gfx.gpu_clock_mutex);
1854         mutex_init(&adev->srbm_mutex);
1855         mutex_init(&adev->gfx.pipe_reserve_mutex);
1856         mutex_init(&adev->grbm_idx_mutex);
1857         mutex_init(&adev->mn_lock);
1858         mutex_init(&adev->virt.vf_errors.lock);
1859         hash_init(adev->mn_hash);
1860         mutex_init(&adev->lock_reset);
1861
1862         amdgpu_device_check_arguments(adev);
1863
1864         spin_lock_init(&adev->mmio_idx_lock);
1865         spin_lock_init(&adev->smc_idx_lock);
1866         spin_lock_init(&adev->pcie_idx_lock);
1867         spin_lock_init(&adev->uvd_ctx_idx_lock);
1868         spin_lock_init(&adev->didt_idx_lock);
1869         spin_lock_init(&adev->gc_cac_idx_lock);
1870         spin_lock_init(&adev->se_cac_idx_lock);
1871         spin_lock_init(&adev->audio_endpt_idx_lock);
1872         spin_lock_init(&adev->mm_stats.lock);
1873
1874         INIT_LIST_HEAD(&adev->shadow_list);
1875         mutex_init(&adev->shadow_list_lock);
1876
1877         INIT_LIST_HEAD(&adev->ring_lru_list);
1878         spin_lock_init(&adev->ring_lru_list_lock);
1879
1880         INIT_DELAYED_WORK(&adev->late_init_work,
1881                           amdgpu_device_ip_late_init_func_handler);
1882
1883         /* Registers mapping */
1884         /* TODO: block userspace mapping of io register */
1885         if (adev->asic_type >= CHIP_BONAIRE) {
1886                 adev->rmmio_base = pci_resource_start(adev->pdev, 5);
1887                 adev->rmmio_size = pci_resource_len(adev->pdev, 5);
1888         } else {
1889                 adev->rmmio_base = pci_resource_start(adev->pdev, 2);
1890                 adev->rmmio_size = pci_resource_len(adev->pdev, 2);
1891         }
1892
1893         adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
1894         if (adev->rmmio == NULL) {
1895                 return -ENOMEM;
1896         }
1897         DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
1898         DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
1899
1900         /* doorbell bar mapping */
1901         amdgpu_device_doorbell_init(adev);
1902
1903         /* io port mapping */
1904         for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
1905                 if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) {
1906                         adev->rio_mem_size = pci_resource_len(adev->pdev, i);
1907                         adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size);
1908                         break;
1909                 }
1910         }
1911         if (adev->rio_mem == NULL)
1912                 DRM_INFO("PCI I/O BAR is not found.\n");
1913
1914         amdgpu_device_get_pcie_info(adev);
1915
1916         /* early init functions */
1917         r = amdgpu_device_ip_early_init(adev);
1918         if (r)
1919                 return r;
1920
1921         /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
1922         /* this will fail for cards that aren't VGA class devices, just
1923          * ignore it */
1924         vga_client_register(adev->pdev, adev, NULL, amdgpu_device_vga_set_decode);
1925
1926         if (amdgpu_device_is_px(ddev))
1927                 runtime = true;
1928         if (!pci_is_thunderbolt_attached(adev->pdev))
1929                 vga_switcheroo_register_client(adev->pdev,
1930                                                &amdgpu_switcheroo_ops, runtime);
1931         if (runtime)
1932                 vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
1933
1934         if (amdgpu_emu_mode == 1) {
1935                 /* post the asic on emulation mode */
1936                 emu_soc_asic_init(adev);
1937                 goto fence_driver_init;
1938         }
1939
1940         /* Read BIOS */
1941         if (!amdgpu_get_bios(adev)) {
1942                 r = -EINVAL;
1943                 goto failed;
1944         }
1945
1946         r = amdgpu_atombios_init(adev);
1947         if (r) {
1948                 dev_err(adev->dev, "amdgpu_atombios_init failed\n");
1949                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0);
1950                 goto failed;
1951         }
1952
1953         /* detect if we are with an SRIOV vbios */
1954         amdgpu_device_detect_sriov_bios(adev);
1955
1956         /* Post card if necessary */
1957         if (amdgpu_device_need_post(adev)) {
1958                 if (!adev->bios) {
1959                         dev_err(adev->dev, "no vBIOS found\n");
1960                         r = -EINVAL;
1961                         goto failed;
1962                 }
1963                 DRM_INFO("GPU posting now...\n");
1964                 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
1965                 if (r) {
1966                         dev_err(adev->dev, "gpu post error!\n");
1967                         goto failed;
1968                 }
1969         }
1970
1971         if (adev->is_atom_fw) {
1972                 /* Initialize clocks */
1973                 r = amdgpu_atomfirmware_get_clock_info(adev);
1974                 if (r) {
1975                         dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
1976                         amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
1977                         goto failed;
1978                 }
1979         } else {
1980                 /* Initialize clocks */
1981                 r = amdgpu_atombios_get_clock_info(adev);
1982                 if (r) {
1983                         dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
1984                         amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
1985                         goto failed;
1986                 }
1987                 /* init i2c buses */
1988                 if (!amdgpu_device_has_dc_support(adev))
1989                         amdgpu_atombios_i2c_init(adev);
1990         }
1991
1992 fence_driver_init:
1993         /* Fence driver */
1994         r = amdgpu_fence_driver_init(adev);
1995         if (r) {
1996                 dev_err(adev->dev, "amdgpu_fence_driver_init failed\n");
1997                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
1998                 goto failed;
1999         }
2000
2001         /* init the mode config */
2002         drm_mode_config_init(adev->ddev);
2003
2004         r = amdgpu_device_ip_init(adev);
2005         if (r) {
2006                 /* failed in exclusive mode due to timeout */
2007                 if (amdgpu_sriov_vf(adev) &&
2008                     !amdgpu_sriov_runtime(adev) &&
2009                     amdgpu_virt_mmio_blocked(adev) &&
2010                     !amdgpu_virt_wait_reset(adev)) {
2011                         dev_err(adev->dev, "VF exclusive mode timeout\n");
2012                         /* Don't send request since VF is inactive. */
2013                         adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
2014                         adev->virt.ops = NULL;
2015                         r = -EAGAIN;
2016                         goto failed;
2017                 }
2018                 dev_err(adev->dev, "amdgpu_device_ip_init failed\n");
2019                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
2020                 amdgpu_device_ip_fini(adev);
2021                 goto failed;
2022         }
2023
2024         adev->accel_working = true;
2025
2026         amdgpu_vm_check_compute_bug(adev);
2027
2028         /* Initialize the buffer migration limit. */
2029         if (amdgpu_moverate >= 0)
2030                 max_MBps = amdgpu_moverate;
2031         else
2032                 max_MBps = 8; /* Allow 8 MB/s. */
2033         /* Get a log2 for easy divisions. */
2034         adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
2035
2036         r = amdgpu_ib_pool_init(adev);
2037         if (r) {
2038                 dev_err(adev->dev, "IB initialization failed (%d).\n", r);
2039                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r);
2040                 goto failed;
2041         }
2042
2043         r = amdgpu_ib_ring_tests(adev);
2044         if (r)
2045                 DRM_ERROR("ib ring test failed (%d).\n", r);
2046
2047         if (amdgpu_sriov_vf(adev))
2048                 amdgpu_virt_init_data_exchange(adev);
2049
2050         amdgpu_fbdev_init(adev);
2051
2052         r = amdgpu_pm_sysfs_init(adev);
2053         if (r)
2054                 DRM_ERROR("registering pm debugfs failed (%d).\n", r);
2055
2056         r = amdgpu_debugfs_gem_init(adev);
2057         if (r)
2058                 DRM_ERROR("registering gem debugfs failed (%d).\n", r);
2059
2060         r = amdgpu_debugfs_regs_init(adev);
2061         if (r)
2062                 DRM_ERROR("registering register debugfs failed (%d).\n", r);
2063
2064         r = amdgpu_debugfs_firmware_init(adev);
2065         if (r)
2066                 DRM_ERROR("registering firmware debugfs failed (%d).\n", r);
2067
2068         r = amdgpu_debugfs_init(adev);
2069         if (r)
2070                 DRM_ERROR("Creating debugfs files failed (%d).\n", r);
2071
2072         if ((amdgpu_testing & 1)) {
2073                 if (adev->accel_working)
2074                         amdgpu_test_moves(adev);
2075                 else
2076                         DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n");
2077         }
2078         if (amdgpu_benchmarking) {
2079                 if (adev->accel_working)
2080                         amdgpu_benchmark(adev, amdgpu_benchmarking);
2081                 else
2082                         DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n");
2083         }
2084
2085         /* enable clockgating, etc. after ib tests, etc. since some blocks require
2086          * explicit gating rather than handling it automatically.
2087          */
2088         r = amdgpu_device_ip_late_init(adev);
2089         if (r) {
2090                 dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n");
2091                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
2092                 goto failed;
2093         }
2094
2095         return 0;
2096
2097 failed:
2098         amdgpu_vf_error_trans_all(adev);
2099         if (runtime)
2100                 vga_switcheroo_fini_domain_pm_ops(adev->dev);
2101
2102         return r;
2103 }
2104
2105 /**
2106  * amdgpu_device_fini - tear down the driver
2107  *
2108  * @adev: amdgpu_device pointer
2109  *
2110  * Tear down the driver info (all asics).
2111  * Called at driver shutdown.
2112  */
2113 void amdgpu_device_fini(struct amdgpu_device *adev)
2114 {
2115         int r;
2116
2117         DRM_INFO("amdgpu: finishing device.\n");
2118         adev->shutdown = true;
2119         if (adev->mode_info.mode_config_initialized)
2120                 drm_crtc_force_disable_all(adev->ddev);
2121
2122         amdgpu_ib_pool_fini(adev);
2123         amdgpu_fence_driver_fini(adev);
2124         amdgpu_pm_sysfs_fini(adev);
2125         amdgpu_fbdev_fini(adev);
2126         r = amdgpu_device_ip_fini(adev);
2127         if (adev->firmware.gpu_info_fw) {
2128                 release_firmware(adev->firmware.gpu_info_fw);
2129                 adev->firmware.gpu_info_fw = NULL;
2130         }
2131         adev->accel_working = false;
2132         cancel_delayed_work_sync(&adev->late_init_work);
2133         /* free i2c buses */
2134         if (!amdgpu_device_has_dc_support(adev))
2135                 amdgpu_i2c_fini(adev);
2136
2137         if (amdgpu_emu_mode != 1)
2138                 amdgpu_atombios_fini(adev);
2139
2140         kfree(adev->bios);
2141         adev->bios = NULL;
2142         if (!pci_is_thunderbolt_attached(adev->pdev))
2143                 vga_switcheroo_unregister_client(adev->pdev);
2144         if (adev->flags & AMD_IS_PX)
2145                 vga_switcheroo_fini_domain_pm_ops(adev->dev);
2146         vga_client_register(adev->pdev, NULL, NULL, NULL);
2147         if (adev->rio_mem)
2148                 pci_iounmap(adev->pdev, adev->rio_mem);
2149         adev->rio_mem = NULL;
2150         iounmap(adev->rmmio);
2151         adev->rmmio = NULL;
2152         amdgpu_device_doorbell_fini(adev);
2153         amdgpu_debugfs_regs_cleanup(adev);
2154 }
2155
2156
2157 /*
2158  * Suspend & resume.
2159  */
2160 /**
2161  * amdgpu_device_suspend - initiate device suspend
2162  *
2163  * @pdev: drm dev pointer
2164  * @state: suspend state
2165  *
2166  * Puts the hw in the suspend state (all asics).
2167  * Returns 0 for success or an error on failure.
2168  * Called at driver suspend.
2169  */
2170 int amdgpu_device_suspend(struct drm_device *dev, bool suspend, bool fbcon)
2171 {
2172         struct amdgpu_device *adev;
2173         struct drm_crtc *crtc;
2174         struct drm_connector *connector;
2175         int r;
2176
2177         if (dev == NULL || dev->dev_private == NULL) {
2178                 return -ENODEV;
2179         }
2180
2181         adev = dev->dev_private;
2182
2183         if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
2184                 return 0;
2185
2186         drm_kms_helper_poll_disable(dev);
2187
2188         if (!amdgpu_device_has_dc_support(adev)) {
2189                 /* turn off display hw */
2190                 drm_modeset_lock_all(dev);
2191                 list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2192                         drm_helper_connector_dpms(connector, DRM_MODE_DPMS_OFF);
2193                 }
2194                 drm_modeset_unlock_all(dev);
2195         }
2196
2197         amdgpu_amdkfd_suspend(adev);
2198
2199         /* unpin the front buffers and cursors */
2200         list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2201                 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2202                 struct amdgpu_framebuffer *rfb = to_amdgpu_framebuffer(crtc->primary->fb);
2203                 struct amdgpu_bo *robj;
2204
2205                 if (amdgpu_crtc->cursor_bo) {
2206                         struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2207                         r = amdgpu_bo_reserve(aobj, true);
2208                         if (r == 0) {
2209                                 amdgpu_bo_unpin(aobj);
2210                                 amdgpu_bo_unreserve(aobj);
2211                         }
2212                 }
2213
2214                 if (rfb == NULL || rfb->obj == NULL) {
2215                         continue;
2216                 }
2217                 robj = gem_to_amdgpu_bo(rfb->obj);
2218                 /* don't unpin kernel fb objects */
2219                 if (!amdgpu_fbdev_robj_is_fb(adev, robj)) {
2220                         r = amdgpu_bo_reserve(robj, true);
2221                         if (r == 0) {
2222                                 amdgpu_bo_unpin(robj);
2223                                 amdgpu_bo_unreserve(robj);
2224                         }
2225                 }
2226         }
2227         /* evict vram memory */
2228         amdgpu_bo_evict_vram(adev);
2229
2230         amdgpu_fence_driver_suspend(adev);
2231
2232         r = amdgpu_device_ip_suspend(adev);
2233
2234         /* evict remaining vram memory
2235          * This second call to evict vram is to evict the gart page table
2236          * using the CPU.
2237          */
2238         amdgpu_bo_evict_vram(adev);
2239
2240         pci_save_state(dev->pdev);
2241         if (suspend) {
2242                 /* Shut down the device */
2243                 pci_disable_device(dev->pdev);
2244                 pci_set_power_state(dev->pdev, PCI_D3hot);
2245         } else {
2246                 r = amdgpu_asic_reset(adev);
2247                 if (r)
2248                         DRM_ERROR("amdgpu asic reset failed\n");
2249         }
2250
2251         if (fbcon) {
2252                 console_lock();
2253                 amdgpu_fbdev_set_suspend(adev, 1);
2254                 console_unlock();
2255         }
2256         return 0;
2257 }
2258
2259 /**
2260  * amdgpu_device_resume - initiate device resume
2261  *
2262  * @pdev: drm dev pointer
2263  *
2264  * Bring the hw back to operating state (all asics).
2265  * Returns 0 for success or an error on failure.
2266  * Called at driver resume.
2267  */
2268 int amdgpu_device_resume(struct drm_device *dev, bool resume, bool fbcon)
2269 {
2270         struct drm_connector *connector;
2271         struct amdgpu_device *adev = dev->dev_private;
2272         struct drm_crtc *crtc;
2273         int r = 0;
2274
2275         if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
2276                 return 0;
2277
2278         if (fbcon)
2279                 console_lock();
2280
2281         if (resume) {
2282                 pci_set_power_state(dev->pdev, PCI_D0);
2283                 pci_restore_state(dev->pdev);
2284                 r = pci_enable_device(dev->pdev);
2285                 if (r)
2286                         goto unlock;
2287         }
2288
2289         /* post card */
2290         if (amdgpu_device_need_post(adev)) {
2291                 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2292                 if (r)
2293                         DRM_ERROR("amdgpu asic init failed\n");
2294         }
2295
2296         r = amdgpu_device_ip_resume(adev);
2297         if (r) {
2298                 DRM_ERROR("amdgpu_device_ip_resume failed (%d).\n", r);
2299                 goto unlock;
2300         }
2301         amdgpu_fence_driver_resume(adev);
2302
2303         if (resume) {
2304                 r = amdgpu_ib_ring_tests(adev);
2305                 if (r)
2306                         DRM_ERROR("ib ring test failed (%d).\n", r);
2307         }
2308
2309         r = amdgpu_device_ip_late_init(adev);
2310         if (r)
2311                 goto unlock;
2312
2313         /* pin cursors */
2314         list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2315                 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2316
2317                 if (amdgpu_crtc->cursor_bo) {
2318                         struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2319                         r = amdgpu_bo_reserve(aobj, true);
2320                         if (r == 0) {
2321                                 r = amdgpu_bo_pin(aobj,
2322                                                   AMDGPU_GEM_DOMAIN_VRAM,
2323                                                   &amdgpu_crtc->cursor_addr);
2324                                 if (r != 0)
2325                                         DRM_ERROR("Failed to pin cursor BO (%d)\n", r);
2326                                 amdgpu_bo_unreserve(aobj);
2327                         }
2328                 }
2329         }
2330         r = amdgpu_amdkfd_resume(adev);
2331         if (r)
2332                 return r;
2333
2334         /* blat the mode back in */
2335         if (fbcon) {
2336                 if (!amdgpu_device_has_dc_support(adev)) {
2337                         /* pre DCE11 */
2338                         drm_helper_resume_force_mode(dev);
2339
2340                         /* turn on display hw */
2341                         drm_modeset_lock_all(dev);
2342                         list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2343                                 drm_helper_connector_dpms(connector, DRM_MODE_DPMS_ON);
2344                         }
2345                         drm_modeset_unlock_all(dev);
2346                 }
2347         }
2348
2349         drm_kms_helper_poll_enable(dev);
2350
2351         /*
2352          * Most of the connector probing functions try to acquire runtime pm
2353          * refs to ensure that the GPU is powered on when connector polling is
2354          * performed. Since we're calling this from a runtime PM callback,
2355          * trying to acquire rpm refs will cause us to deadlock.
2356          *
2357          * Since we're guaranteed to be holding the rpm lock, it's safe to
2358          * temporarily disable the rpm helpers so this doesn't deadlock us.
2359          */
2360 #ifdef CONFIG_PM
2361         dev->dev->power.disable_depth++;
2362 #endif
2363         if (!amdgpu_device_has_dc_support(adev))
2364                 drm_helper_hpd_irq_event(dev);
2365         else
2366                 drm_kms_helper_hotplug_event(dev);
2367 #ifdef CONFIG_PM
2368         dev->dev->power.disable_depth--;
2369 #endif
2370
2371         if (fbcon)
2372                 amdgpu_fbdev_set_suspend(adev, 0);
2373
2374 unlock:
2375         if (fbcon)
2376                 console_unlock();
2377
2378         return r;
2379 }
2380
2381 static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)
2382 {
2383         int i;
2384         bool asic_hang = false;
2385
2386         if (amdgpu_sriov_vf(adev))
2387                 return true;
2388
2389         for (i = 0; i < adev->num_ip_blocks; i++) {
2390                 if (!adev->ip_blocks[i].status.valid)
2391                         continue;
2392                 if (adev->ip_blocks[i].version->funcs->check_soft_reset)
2393                         adev->ip_blocks[i].status.hang =
2394                                 adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
2395                 if (adev->ip_blocks[i].status.hang) {
2396                         DRM_INFO("IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
2397                         asic_hang = true;
2398                 }
2399         }
2400         return asic_hang;
2401 }
2402
2403 static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)
2404 {
2405         int i, r = 0;
2406
2407         for (i = 0; i < adev->num_ip_blocks; i++) {
2408                 if (!adev->ip_blocks[i].status.valid)
2409                         continue;
2410                 if (adev->ip_blocks[i].status.hang &&
2411                     adev->ip_blocks[i].version->funcs->pre_soft_reset) {
2412                         r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
2413                         if (r)
2414                                 return r;
2415                 }
2416         }
2417
2418         return 0;
2419 }
2420
2421 static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)
2422 {
2423         int i;
2424
2425         for (i = 0; i < adev->num_ip_blocks; i++) {
2426                 if (!adev->ip_blocks[i].status.valid)
2427                         continue;
2428                 if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
2429                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
2430                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
2431                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
2432                      adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
2433                         if (adev->ip_blocks[i].status.hang) {
2434                                 DRM_INFO("Some block need full reset!\n");
2435                                 return true;
2436                         }
2437                 }
2438         }
2439         return false;
2440 }
2441
2442 static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)
2443 {
2444         int i, r = 0;
2445
2446         for (i = 0; i < adev->num_ip_blocks; i++) {
2447                 if (!adev->ip_blocks[i].status.valid)
2448                         continue;
2449                 if (adev->ip_blocks[i].status.hang &&
2450                     adev->ip_blocks[i].version->funcs->soft_reset) {
2451                         r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
2452                         if (r)
2453                                 return r;
2454                 }
2455         }
2456
2457         return 0;
2458 }
2459
2460 static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)
2461 {
2462         int i, r = 0;
2463
2464         for (i = 0; i < adev->num_ip_blocks; i++) {
2465                 if (!adev->ip_blocks[i].status.valid)
2466                         continue;
2467                 if (adev->ip_blocks[i].status.hang &&
2468                     adev->ip_blocks[i].version->funcs->post_soft_reset)
2469                         r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
2470                 if (r)
2471                         return r;
2472         }
2473
2474         return 0;
2475 }
2476
2477 static int amdgpu_device_recover_vram_from_shadow(struct amdgpu_device *adev,
2478                                                   struct amdgpu_ring *ring,
2479                                                   struct amdgpu_bo *bo,
2480                                                   struct dma_fence **fence)
2481 {
2482         uint32_t domain;
2483         int r;
2484
2485         if (!bo->shadow)
2486                 return 0;
2487
2488         r = amdgpu_bo_reserve(bo, true);
2489         if (r)
2490                 return r;
2491         domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
2492         /* if bo has been evicted, then no need to recover */
2493         if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
2494                 r = amdgpu_bo_validate(bo->shadow);
2495                 if (r) {
2496                         DRM_ERROR("bo validate failed!\n");
2497                         goto err;
2498                 }
2499
2500                 r = amdgpu_bo_restore_from_shadow(adev, ring, bo,
2501                                                  NULL, fence, true);
2502                 if (r) {
2503                         DRM_ERROR("recover page table failed!\n");
2504                         goto err;
2505                 }
2506         }
2507 err:
2508         amdgpu_bo_unreserve(bo);
2509         return r;
2510 }
2511
2512 static int amdgpu_device_handle_vram_lost(struct amdgpu_device *adev)
2513 {
2514         struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
2515         struct amdgpu_bo *bo, *tmp;
2516         struct dma_fence *fence = NULL, *next = NULL;
2517         long r = 1;
2518         int i = 0;
2519         long tmo;
2520
2521         if (amdgpu_sriov_runtime(adev))
2522                 tmo = msecs_to_jiffies(amdgpu_lockup_timeout);
2523         else
2524                 tmo = msecs_to_jiffies(100);
2525
2526         DRM_INFO("recover vram bo from shadow start\n");
2527         mutex_lock(&adev->shadow_list_lock);
2528         list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) {
2529                 next = NULL;
2530                 amdgpu_device_recover_vram_from_shadow(adev, ring, bo, &next);
2531                 if (fence) {
2532                         r = dma_fence_wait_timeout(fence, false, tmo);
2533                         if (r == 0)
2534                                 pr_err("wait fence %p[%d] timeout\n", fence, i);
2535                         else if (r < 0)
2536                                 pr_err("wait fence %p[%d] interrupted\n", fence, i);
2537                         if (r < 1) {
2538                                 dma_fence_put(fence);
2539                                 fence = next;
2540                                 break;
2541                         }
2542                         i++;
2543                 }
2544
2545                 dma_fence_put(fence);
2546                 fence = next;
2547         }
2548         mutex_unlock(&adev->shadow_list_lock);
2549
2550         if (fence) {
2551                 r = dma_fence_wait_timeout(fence, false, tmo);
2552                 if (r == 0)
2553                         pr_err("wait fence %p[%d] timeout\n", fence, i);
2554                 else if (r < 0)
2555                         pr_err("wait fence %p[%d] interrupted\n", fence, i);
2556
2557         }
2558         dma_fence_put(fence);
2559
2560         if (r > 0)
2561                 DRM_INFO("recover vram bo from shadow done\n");
2562         else
2563                 DRM_ERROR("recover vram bo from shadow failed\n");
2564
2565         return (r > 0?0:1);
2566 }
2567
2568 /*
2569  * amdgpu_device_reset - reset ASIC/GPU for bare-metal or passthrough
2570  *
2571  * @adev: amdgpu device pointer
2572  *
2573  * attempt to do soft-reset or full-reset and reinitialize Asic
2574  * return 0 means successed otherwise failed
2575 */
2576 static int amdgpu_device_reset(struct amdgpu_device *adev)
2577 {
2578         bool need_full_reset, vram_lost = 0;
2579         int r;
2580
2581         need_full_reset = amdgpu_device_ip_need_full_reset(adev);
2582
2583         if (!need_full_reset) {
2584                 amdgpu_device_ip_pre_soft_reset(adev);
2585                 r = amdgpu_device_ip_soft_reset(adev);
2586                 amdgpu_device_ip_post_soft_reset(adev);
2587                 if (r || amdgpu_device_ip_check_soft_reset(adev)) {
2588                         DRM_INFO("soft reset failed, will fallback to full reset!\n");
2589                         need_full_reset = true;
2590                 }
2591         }
2592
2593         if (need_full_reset) {
2594                 r = amdgpu_device_ip_suspend(adev);
2595
2596 retry:
2597                 r = amdgpu_asic_reset(adev);
2598                 /* post card */
2599                 amdgpu_atom_asic_init(adev->mode_info.atom_context);
2600
2601                 if (!r) {
2602                         dev_info(adev->dev, "GPU reset succeeded, trying to resume\n");
2603                         r = amdgpu_device_ip_resume_phase1(adev);
2604                         if (r)
2605                                 goto out;
2606
2607                         vram_lost = amdgpu_device_check_vram_lost(adev);
2608                         if (vram_lost) {
2609                                 DRM_ERROR("VRAM is lost!\n");
2610                                 atomic_inc(&adev->vram_lost_counter);
2611                         }
2612
2613                         r = amdgpu_gtt_mgr_recover(
2614                                 &adev->mman.bdev.man[TTM_PL_TT]);
2615                         if (r)
2616                                 goto out;
2617
2618                         r = amdgpu_device_ip_resume_phase2(adev);
2619                         if (r)
2620                                 goto out;
2621
2622                         if (vram_lost)
2623                                 amdgpu_device_fill_reset_magic(adev);
2624                 }
2625         }
2626
2627 out:
2628         if (!r) {
2629                 amdgpu_irq_gpu_reset_resume_helper(adev);
2630                 r = amdgpu_ib_ring_tests(adev);
2631                 if (r) {
2632                         dev_err(adev->dev, "ib ring test failed (%d).\n", r);
2633                         r = amdgpu_device_ip_suspend(adev);
2634                         need_full_reset = true;
2635                         goto retry;
2636                 }
2637         }
2638
2639         if (!r && ((need_full_reset && !(adev->flags & AMD_IS_APU)) || vram_lost))
2640                 r = amdgpu_device_handle_vram_lost(adev);
2641
2642         return r;
2643 }
2644
2645 /*
2646  * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
2647  *
2648  * @adev: amdgpu device pointer
2649  *
2650  * do VF FLR and reinitialize Asic
2651  * return 0 means successed otherwise failed
2652 */
2653 static int amdgpu_device_reset_sriov(struct amdgpu_device *adev, bool from_hypervisor)
2654 {
2655         int r;
2656
2657         if (from_hypervisor)
2658                 r = amdgpu_virt_request_full_gpu(adev, true);
2659         else
2660                 r = amdgpu_virt_reset_gpu(adev);
2661         if (r)
2662                 return r;
2663
2664         /* Resume IP prior to SMC */
2665         r = amdgpu_device_ip_reinit_early_sriov(adev);
2666         if (r)
2667                 goto error;
2668
2669         /* we need recover gart prior to run SMC/CP/SDMA resume */
2670         amdgpu_gtt_mgr_recover(&adev->mman.bdev.man[TTM_PL_TT]);
2671
2672         /* now we are okay to resume SMC/CP/SDMA */
2673         r = amdgpu_device_ip_reinit_late_sriov(adev);
2674         amdgpu_virt_release_full_gpu(adev, true);
2675         if (r)
2676                 goto error;
2677
2678         amdgpu_irq_gpu_reset_resume_helper(adev);
2679         r = amdgpu_ib_ring_tests(adev);
2680
2681         if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
2682                 atomic_inc(&adev->vram_lost_counter);
2683                 r = amdgpu_device_handle_vram_lost(adev);
2684         }
2685
2686 error:
2687
2688         return r;
2689 }
2690
2691 /**
2692  * amdgpu_device_gpu_recover - reset the asic and recover scheduler
2693  *
2694  * @adev: amdgpu device pointer
2695  * @job: which job trigger hang
2696  * @force forces reset regardless of amdgpu_gpu_recovery
2697  *
2698  * Attempt to reset the GPU if it has hung (all asics).
2699  * Returns 0 for success or an error on failure.
2700  */
2701 int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
2702                               struct amdgpu_job *job, bool force)
2703 {
2704         struct drm_atomic_state *state = NULL;
2705         int i, r, resched;
2706
2707         if (!force && !amdgpu_device_ip_check_soft_reset(adev)) {
2708                 DRM_INFO("No hardware hang detected. Did some blocks stall?\n");
2709                 return 0;
2710         }
2711
2712         if (!force && (amdgpu_gpu_recovery == 0 ||
2713                         (amdgpu_gpu_recovery == -1  && !amdgpu_sriov_vf(adev)))) {
2714                 DRM_INFO("GPU recovery disabled.\n");
2715                 return 0;
2716         }
2717
2718         dev_info(adev->dev, "GPU reset begin!\n");
2719
2720         mutex_lock(&adev->lock_reset);
2721         atomic_inc(&adev->gpu_reset_counter);
2722         adev->in_gpu_reset = 1;
2723
2724         /* block TTM */
2725         resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
2726
2727         /* store modesetting */
2728         if (amdgpu_device_has_dc_support(adev))
2729                 state = drm_atomic_helper_suspend(adev->ddev);
2730
2731         /* block all schedulers and reset given job's ring */
2732         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
2733                 struct amdgpu_ring *ring = adev->rings[i];
2734
2735                 if (!ring || !ring->sched.thread)
2736                         continue;
2737
2738                 kthread_park(ring->sched.thread);
2739
2740                 if (job && job->ring->idx != i)
2741                         continue;
2742
2743                 drm_sched_hw_job_reset(&ring->sched, &job->base);
2744
2745                 /* after all hw jobs are reset, hw fence is meaningless, so force_completion */
2746                 amdgpu_fence_driver_force_completion(ring);
2747         }
2748
2749         if (amdgpu_sriov_vf(adev))
2750                 r = amdgpu_device_reset_sriov(adev, job ? false : true);
2751         else
2752                 r = amdgpu_device_reset(adev);
2753
2754         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
2755                 struct amdgpu_ring *ring = adev->rings[i];
2756
2757                 if (!ring || !ring->sched.thread)
2758                         continue;
2759
2760                 /* only need recovery sched of the given job's ring
2761                  * or all rings (in the case @job is NULL)
2762                  * after above amdgpu_reset accomplished
2763                  */
2764                 if ((!job || job->ring->idx == i) && !r)
2765                         drm_sched_job_recovery(&ring->sched);
2766
2767                 kthread_unpark(ring->sched.thread);
2768         }
2769
2770         if (amdgpu_device_has_dc_support(adev)) {
2771                 if (drm_atomic_helper_resume(adev->ddev, state))
2772                         dev_info(adev->dev, "drm resume failed:%d\n", r);
2773         } else {
2774                 drm_helper_resume_force_mode(adev->ddev);
2775         }
2776
2777         ttm_bo_unlock_delayed_workqueue(&adev->mman.bdev, resched);
2778
2779         if (r) {
2780                 /* bad news, how to tell it to userspace ? */
2781                 dev_info(adev->dev, "GPU reset(%d) failed\n", atomic_read(&adev->gpu_reset_counter));
2782                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r);
2783         } else {
2784                 dev_info(adev->dev, "GPU reset(%d) successed!\n",atomic_read(&adev->gpu_reset_counter));
2785         }
2786
2787         amdgpu_vf_error_trans_all(adev);
2788         adev->in_gpu_reset = 0;
2789         mutex_unlock(&adev->lock_reset);
2790         return r;
2791 }
2792
2793 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev)
2794 {
2795         u32 mask;
2796         int ret;
2797
2798         if (amdgpu_pcie_gen_cap)
2799                 adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
2800
2801         if (amdgpu_pcie_lane_cap)
2802                 adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
2803
2804         /* covers APUs as well */
2805         if (pci_is_root_bus(adev->pdev->bus)) {
2806                 if (adev->pm.pcie_gen_mask == 0)
2807                         adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
2808                 if (adev->pm.pcie_mlw_mask == 0)
2809                         adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
2810                 return;
2811         }
2812
2813         if (adev->pm.pcie_gen_mask == 0) {
2814                 ret = drm_pcie_get_speed_cap_mask(adev->ddev, &mask);
2815                 if (!ret) {
2816                         adev->pm.pcie_gen_mask = (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
2817                                                   CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
2818                                                   CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
2819
2820                         if (mask & DRM_PCIE_SPEED_25)
2821                                 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
2822                         if (mask & DRM_PCIE_SPEED_50)
2823                                 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2;
2824                         if (mask & DRM_PCIE_SPEED_80)
2825                                 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3;
2826                 } else {
2827                         adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
2828                 }
2829         }
2830         if (adev->pm.pcie_mlw_mask == 0) {
2831                 ret = drm_pcie_get_max_link_width(adev->ddev, &mask);
2832                 if (!ret) {
2833                         switch (mask) {
2834                         case 32:
2835                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
2836                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
2837                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
2838                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
2839                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
2840                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2841                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2842                                 break;
2843                         case 16:
2844                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
2845                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
2846                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
2847                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
2848                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2849                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2850                                 break;
2851                         case 12:
2852                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
2853                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
2854                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
2855                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2856                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2857                                 break;
2858                         case 8:
2859                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
2860                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
2861                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2862                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2863                                 break;
2864                         case 4:
2865                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
2866                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2867                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2868                                 break;
2869                         case 2:
2870                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2871                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2872                                 break;
2873                         case 1:
2874                                 adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;
2875                                 break;
2876                         default:
2877                                 break;
2878                         }
2879                 } else {
2880                         adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
2881                 }
2882         }
2883 }
2884