2 * Copyright 2008 Advanced Micro Devices, Inc.
3 * Copyright 2008 Red Hat Inc.
4 * Copyright 2009 Jerome Glisse.
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:
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
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.
24 * Authors: Dave Airlie
28 #include <linux/power_supply.h>
29 #include <linux/kthread.h>
30 #include <linux/module.h>
31 #include <linux/console.h>
32 #include <linux/slab.h>
34 #include <drm/drm_atomic_helper.h>
35 #include <drm/drm_probe_helper.h>
36 #include <drm/amdgpu_drm.h>
37 #include <linux/vgaarb.h>
38 #include <linux/vga_switcheroo.h>
39 #include <linux/efi.h>
41 #include "amdgpu_trace.h"
42 #include "amdgpu_i2c.h"
44 #include "amdgpu_atombios.h"
45 #include "amdgpu_atomfirmware.h"
47 #ifdef CONFIG_DRM_AMDGPU_SI
50 #ifdef CONFIG_DRM_AMDGPU_CIK
56 #include "bif/bif_4_1_d.h"
57 #include <linux/pci.h>
58 #include <linux/firmware.h>
59 #include "amdgpu_vf_error.h"
61 #include "amdgpu_amdkfd.h"
62 #include "amdgpu_pm.h"
64 #include "amdgpu_xgmi.h"
65 #include "amdgpu_ras.h"
66 #include "amdgpu_pmu.h"
67 #include "amdgpu_fru_eeprom.h"
69 #include <linux/suspend.h>
70 #include <drm/task_barrier.h>
71 #include <linux/pm_runtime.h>
73 MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin");
74 MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin");
75 MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin");
76 MODULE_FIRMWARE("amdgpu/picasso_gpu_info.bin");
77 MODULE_FIRMWARE("amdgpu/raven2_gpu_info.bin");
78 MODULE_FIRMWARE("amdgpu/arcturus_gpu_info.bin");
79 MODULE_FIRMWARE("amdgpu/renoir_gpu_info.bin");
80 MODULE_FIRMWARE("amdgpu/navi10_gpu_info.bin");
81 MODULE_FIRMWARE("amdgpu/navi14_gpu_info.bin");
82 MODULE_FIRMWARE("amdgpu/navi12_gpu_info.bin");
83 MODULE_FIRMWARE("amdgpu/sienna_cichlid_gpu_info.bin");
84 MODULE_FIRMWARE("amdgpu/navy_flounder_gpu_info.bin");
86 #define AMDGPU_RESUME_MS 2000
88 const char *amdgpu_asic_name[] = {
123 * DOC: pcie_replay_count
125 * The amdgpu driver provides a sysfs API for reporting the total number
126 * of PCIe replays (NAKs)
127 * The file pcie_replay_count is used for this and returns the total
128 * number of replays as a sum of the NAKs generated and NAKs received
131 static ssize_t amdgpu_device_get_pcie_replay_count(struct device *dev,
132 struct device_attribute *attr, char *buf)
134 struct drm_device *ddev = dev_get_drvdata(dev);
135 struct amdgpu_device *adev = drm_to_adev(ddev);
136 uint64_t cnt = amdgpu_asic_get_pcie_replay_count(adev);
138 return snprintf(buf, PAGE_SIZE, "%llu\n", cnt);
141 static DEVICE_ATTR(pcie_replay_count, S_IRUGO,
142 amdgpu_device_get_pcie_replay_count, NULL);
144 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev);
149 * The amdgpu driver provides a sysfs API for reporting the product name
151 * The file serial_number is used for this and returns the product name
152 * as returned from the FRU.
153 * NOTE: This is only available for certain server cards
156 static ssize_t amdgpu_device_get_product_name(struct device *dev,
157 struct device_attribute *attr, char *buf)
159 struct drm_device *ddev = dev_get_drvdata(dev);
160 struct amdgpu_device *adev = drm_to_adev(ddev);
162 return snprintf(buf, PAGE_SIZE, "%s\n", adev->product_name);
165 static DEVICE_ATTR(product_name, S_IRUGO,
166 amdgpu_device_get_product_name, NULL);
169 * DOC: product_number
171 * The amdgpu driver provides a sysfs API for reporting the part number
173 * The file serial_number is used for this and returns the part number
174 * as returned from the FRU.
175 * NOTE: This is only available for certain server cards
178 static ssize_t amdgpu_device_get_product_number(struct device *dev,
179 struct device_attribute *attr, char *buf)
181 struct drm_device *ddev = dev_get_drvdata(dev);
182 struct amdgpu_device *adev = drm_to_adev(ddev);
184 return snprintf(buf, PAGE_SIZE, "%s\n", adev->product_number);
187 static DEVICE_ATTR(product_number, S_IRUGO,
188 amdgpu_device_get_product_number, NULL);
193 * The amdgpu driver provides a sysfs API for reporting the serial number
195 * The file serial_number is used for this and returns the serial number
196 * as returned from the FRU.
197 * NOTE: This is only available for certain server cards
200 static ssize_t amdgpu_device_get_serial_number(struct device *dev,
201 struct device_attribute *attr, char *buf)
203 struct drm_device *ddev = dev_get_drvdata(dev);
204 struct amdgpu_device *adev = drm_to_adev(ddev);
206 return snprintf(buf, PAGE_SIZE, "%s\n", adev->serial);
209 static DEVICE_ATTR(serial_number, S_IRUGO,
210 amdgpu_device_get_serial_number, NULL);
213 * amdgpu_device_supports_boco - Is the device a dGPU with HG/PX power control
215 * @dev: drm_device pointer
217 * Returns true if the device is a dGPU with HG/PX power control,
218 * otherwise return false.
220 bool amdgpu_device_supports_boco(struct drm_device *dev)
222 struct amdgpu_device *adev = drm_to_adev(dev);
224 if (adev->flags & AMD_IS_PX)
230 * amdgpu_device_supports_baco - Does the device support BACO
232 * @dev: drm_device pointer
234 * Returns true if the device supporte BACO,
235 * otherwise return false.
237 bool amdgpu_device_supports_baco(struct drm_device *dev)
239 struct amdgpu_device *adev = drm_to_adev(dev);
241 return amdgpu_asic_supports_baco(adev);
245 * VRAM access helper functions.
247 * amdgpu_device_vram_access - read/write a buffer in vram
249 * @adev: amdgpu_device pointer
250 * @pos: offset of the buffer in vram
251 * @buf: virtual address of the buffer in system memory
252 * @size: read/write size, sizeof(@buf) must > @size
253 * @write: true - write to vram, otherwise - read from vram
255 void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos,
256 uint32_t *buf, size_t size, bool write)
264 last = min(pos + size, adev->gmc.visible_vram_size);
266 void __iomem *addr = adev->mman.aper_base_kaddr + pos;
267 size_t count = last - pos;
270 memcpy_toio(addr, buf, count);
272 amdgpu_asic_flush_hdp(adev, NULL);
274 amdgpu_asic_invalidate_hdp(adev, NULL);
276 memcpy_fromio(buf, addr, count);
288 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
289 for (last = pos + size; pos < last; pos += 4) {
290 uint32_t tmp = pos >> 31;
292 WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)pos) | 0x80000000);
294 WREG32_NO_KIQ(mmMM_INDEX_HI, tmp);
298 WREG32_NO_KIQ(mmMM_DATA, *buf++);
300 *buf++ = RREG32_NO_KIQ(mmMM_DATA);
302 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
306 * MMIO register access helper functions.
309 * amdgpu_mm_rreg - read a memory mapped IO register
311 * @adev: amdgpu_device pointer
312 * @reg: dword aligned register offset
313 * @acc_flags: access flags which require special behavior
315 * Returns the 32 bit value from the offset specified.
317 uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg,
322 if (adev->in_pci_err_recovery)
325 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev) &&
326 down_read_trylock(&adev->reset_sem)) {
327 ret = amdgpu_kiq_rreg(adev, reg);
328 up_read(&adev->reset_sem);
332 if ((reg * 4) < adev->rmmio_size)
333 ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
337 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
338 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
339 ret = readl(((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
340 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
343 trace_amdgpu_mm_rreg(adev->pdev->device, reg, ret);
348 * MMIO register read with bytes helper functions
349 * @offset:bytes offset from MMIO start
354 * amdgpu_mm_rreg8 - read a memory mapped IO register
356 * @adev: amdgpu_device pointer
357 * @offset: byte aligned register offset
359 * Returns the 8 bit value from the offset specified.
361 uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset) {
362 if (adev->in_pci_err_recovery)
365 if (offset < adev->rmmio_size)
366 return (readb(adev->rmmio + offset));
371 * MMIO register write with bytes helper functions
372 * @offset:bytes offset from MMIO start
373 * @value: the value want to be written to the register
377 * amdgpu_mm_wreg8 - read a memory mapped IO register
379 * @adev: amdgpu_device pointer
380 * @offset: byte aligned register offset
381 * @value: 8 bit value to write
383 * Writes the value specified to the offset specified.
385 void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value) {
386 if (adev->in_pci_err_recovery)
389 if (offset < adev->rmmio_size)
390 writeb(value, adev->rmmio + offset);
395 static inline void amdgpu_mm_wreg_mmio(struct amdgpu_device *adev,
396 uint32_t reg, uint32_t v,
399 if (adev->in_pci_err_recovery)
402 trace_amdgpu_mm_wreg(adev->pdev->device, reg, v);
404 if ((reg * 4) < adev->rmmio_size)
405 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
409 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
410 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
411 writel(v, ((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
412 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
417 * amdgpu_mm_wreg - write to a memory mapped IO register
419 * @adev: amdgpu_device pointer
420 * @reg: dword aligned register offset
421 * @v: 32 bit value to write to the register
422 * @acc_flags: access flags which require special behavior
424 * Writes the value specified to the offset specified.
426 void amdgpu_mm_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v,
429 if (adev->in_pci_err_recovery)
432 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev) &&
433 down_read_trylock(&adev->reset_sem)) {
434 amdgpu_kiq_wreg(adev, reg, v);
435 up_read(&adev->reset_sem);
439 amdgpu_mm_wreg_mmio(adev, reg, v, acc_flags);
443 * amdgpu_mm_wreg_mmio_rlc - write register either with mmio or with RLC path if in range
445 * this function is invoked only the debugfs register access
447 void amdgpu_mm_wreg_mmio_rlc(struct amdgpu_device *adev, uint32_t reg, uint32_t v,
450 if (adev->in_pci_err_recovery)
453 if (amdgpu_sriov_fullaccess(adev) &&
454 adev->gfx.rlc.funcs &&
455 adev->gfx.rlc.funcs->is_rlcg_access_range) {
457 if (adev->gfx.rlc.funcs->is_rlcg_access_range(adev, reg))
458 return adev->gfx.rlc.funcs->rlcg_wreg(adev, reg, v);
461 amdgpu_mm_wreg_mmio(adev, reg, v, acc_flags);
465 * amdgpu_io_rreg - read an IO register
467 * @adev: amdgpu_device pointer
468 * @reg: dword aligned register offset
470 * Returns the 32 bit value from the offset specified.
472 u32 amdgpu_io_rreg(struct amdgpu_device *adev, u32 reg)
474 if (adev->in_pci_err_recovery)
477 if ((reg * 4) < adev->rio_mem_size)
478 return ioread32(adev->rio_mem + (reg * 4));
480 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
481 return ioread32(adev->rio_mem + (mmMM_DATA * 4));
486 * amdgpu_io_wreg - write to an IO register
488 * @adev: amdgpu_device pointer
489 * @reg: dword aligned register offset
490 * @v: 32 bit value to write to the register
492 * Writes the value specified to the offset specified.
494 void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32 v)
496 if (adev->in_pci_err_recovery)
499 if ((reg * 4) < adev->rio_mem_size)
500 iowrite32(v, adev->rio_mem + (reg * 4));
502 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
503 iowrite32(v, adev->rio_mem + (mmMM_DATA * 4));
508 * amdgpu_mm_rdoorbell - read a doorbell dword
510 * @adev: amdgpu_device pointer
511 * @index: doorbell index
513 * Returns the value in the doorbell aperture at the
514 * requested doorbell index (CIK).
516 u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
518 if (adev->in_pci_err_recovery)
521 if (index < adev->doorbell.num_doorbells) {
522 return readl(adev->doorbell.ptr + index);
524 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
530 * amdgpu_mm_wdoorbell - write a doorbell dword
532 * @adev: amdgpu_device pointer
533 * @index: doorbell index
536 * Writes @v to the doorbell aperture at the
537 * requested doorbell index (CIK).
539 void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v)
541 if (adev->in_pci_err_recovery)
544 if (index < adev->doorbell.num_doorbells) {
545 writel(v, adev->doorbell.ptr + index);
547 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
552 * amdgpu_mm_rdoorbell64 - read a doorbell Qword
554 * @adev: amdgpu_device pointer
555 * @index: doorbell index
557 * Returns the value in the doorbell aperture at the
558 * requested doorbell index (VEGA10+).
560 u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index)
562 if (adev->in_pci_err_recovery)
565 if (index < adev->doorbell.num_doorbells) {
566 return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index));
568 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
574 * amdgpu_mm_wdoorbell64 - write a doorbell Qword
576 * @adev: amdgpu_device pointer
577 * @index: doorbell index
580 * Writes @v to the doorbell aperture at the
581 * requested doorbell index (VEGA10+).
583 void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v)
585 if (adev->in_pci_err_recovery)
588 if (index < adev->doorbell.num_doorbells) {
589 atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v);
591 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
596 * amdgpu_invalid_rreg - dummy reg read function
598 * @adev: amdgpu device pointer
599 * @reg: offset of register
601 * Dummy register read function. Used for register blocks
602 * that certain asics don't have (all asics).
603 * Returns the value in the register.
605 static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
607 DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
613 * amdgpu_invalid_wreg - dummy reg write function
615 * @adev: amdgpu device pointer
616 * @reg: offset of register
617 * @v: value to write to the register
619 * Dummy register read function. Used for register blocks
620 * that certain asics don't have (all asics).
622 static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
624 DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
630 * amdgpu_invalid_rreg64 - dummy 64 bit reg read function
632 * @adev: amdgpu device pointer
633 * @reg: offset of register
635 * Dummy register read function. Used for register blocks
636 * that certain asics don't have (all asics).
637 * Returns the value in the register.
639 static uint64_t amdgpu_invalid_rreg64(struct amdgpu_device *adev, uint32_t reg)
641 DRM_ERROR("Invalid callback to read 64 bit register 0x%04X\n", reg);
647 * amdgpu_invalid_wreg64 - dummy reg write function
649 * @adev: amdgpu device pointer
650 * @reg: offset of register
651 * @v: value to write to the register
653 * Dummy register read function. Used for register blocks
654 * that certain asics don't have (all asics).
656 static void amdgpu_invalid_wreg64(struct amdgpu_device *adev, uint32_t reg, uint64_t v)
658 DRM_ERROR("Invalid callback to write 64 bit register 0x%04X with 0x%08llX\n",
664 * amdgpu_block_invalid_rreg - dummy reg read function
666 * @adev: amdgpu device pointer
667 * @block: offset of instance
668 * @reg: offset of register
670 * Dummy register read function. Used for register blocks
671 * that certain asics don't have (all asics).
672 * Returns the value in the register.
674 static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
675 uint32_t block, uint32_t reg)
677 DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
684 * amdgpu_block_invalid_wreg - dummy reg write function
686 * @adev: amdgpu device pointer
687 * @block: offset of instance
688 * @reg: offset of register
689 * @v: value to write to the register
691 * Dummy register read function. Used for register blocks
692 * that certain asics don't have (all asics).
694 static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
696 uint32_t reg, uint32_t v)
698 DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
704 * amdgpu_device_asic_init - Wrapper for atom asic_init
706 * @dev: drm_device pointer
708 * Does any asic specific work and then calls atom asic init.
710 static int amdgpu_device_asic_init(struct amdgpu_device *adev)
712 amdgpu_asic_pre_asic_init(adev);
714 return amdgpu_atom_asic_init(adev->mode_info.atom_context);
718 * amdgpu_device_vram_scratch_init - allocate the VRAM scratch page
720 * @adev: amdgpu device pointer
722 * Allocates a scratch page of VRAM for use by various things in the
725 static int amdgpu_device_vram_scratch_init(struct amdgpu_device *adev)
727 return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE,
728 PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
729 &adev->vram_scratch.robj,
730 &adev->vram_scratch.gpu_addr,
731 (void **)&adev->vram_scratch.ptr);
735 * amdgpu_device_vram_scratch_fini - Free the VRAM scratch page
737 * @adev: amdgpu device pointer
739 * Frees the VRAM scratch page.
741 static void amdgpu_device_vram_scratch_fini(struct amdgpu_device *adev)
743 amdgpu_bo_free_kernel(&adev->vram_scratch.robj, NULL, NULL);
747 * amdgpu_device_program_register_sequence - program an array of registers.
749 * @adev: amdgpu_device pointer
750 * @registers: pointer to the register array
751 * @array_size: size of the register array
753 * Programs an array or registers with and and or masks.
754 * This is a helper for setting golden registers.
756 void amdgpu_device_program_register_sequence(struct amdgpu_device *adev,
757 const u32 *registers,
758 const u32 array_size)
760 u32 tmp, reg, and_mask, or_mask;
766 for (i = 0; i < array_size; i +=3) {
767 reg = registers[i + 0];
768 and_mask = registers[i + 1];
769 or_mask = registers[i + 2];
771 if (and_mask == 0xffffffff) {
776 if (adev->family >= AMDGPU_FAMILY_AI)
777 tmp |= (or_mask & and_mask);
786 * amdgpu_device_pci_config_reset - reset the GPU
788 * @adev: amdgpu_device pointer
790 * Resets the GPU using the pci config reset sequence.
791 * Only applicable to asics prior to vega10.
793 void amdgpu_device_pci_config_reset(struct amdgpu_device *adev)
795 pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
799 * GPU doorbell aperture helpers function.
802 * amdgpu_device_doorbell_init - Init doorbell driver information.
804 * @adev: amdgpu_device pointer
806 * Init doorbell driver information (CIK)
807 * Returns 0 on success, error on failure.
809 static int amdgpu_device_doorbell_init(struct amdgpu_device *adev)
812 /* No doorbell on SI hardware generation */
813 if (adev->asic_type < CHIP_BONAIRE) {
814 adev->doorbell.base = 0;
815 adev->doorbell.size = 0;
816 adev->doorbell.num_doorbells = 0;
817 adev->doorbell.ptr = NULL;
821 if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET)
824 amdgpu_asic_init_doorbell_index(adev);
826 /* doorbell bar mapping */
827 adev->doorbell.base = pci_resource_start(adev->pdev, 2);
828 adev->doorbell.size = pci_resource_len(adev->pdev, 2);
830 adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32),
831 adev->doorbell_index.max_assignment+1);
832 if (adev->doorbell.num_doorbells == 0)
835 /* For Vega, reserve and map two pages on doorbell BAR since SDMA
836 * paging queue doorbell use the second page. The
837 * AMDGPU_DOORBELL64_MAX_ASSIGNMENT definition assumes all the
838 * doorbells are in the first page. So with paging queue enabled,
839 * the max num_doorbells should + 1 page (0x400 in dword)
841 if (adev->asic_type >= CHIP_VEGA10)
842 adev->doorbell.num_doorbells += 0x400;
844 adev->doorbell.ptr = ioremap(adev->doorbell.base,
845 adev->doorbell.num_doorbells *
847 if (adev->doorbell.ptr == NULL)
854 * amdgpu_device_doorbell_fini - Tear down doorbell driver information.
856 * @adev: amdgpu_device pointer
858 * Tear down doorbell driver information (CIK)
860 static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev)
862 iounmap(adev->doorbell.ptr);
863 adev->doorbell.ptr = NULL;
869 * amdgpu_device_wb_*()
870 * Writeback is the method by which the GPU updates special pages in memory
871 * with the status of certain GPU events (fences, ring pointers,etc.).
875 * amdgpu_device_wb_fini - Disable Writeback and free memory
877 * @adev: amdgpu_device pointer
879 * Disables Writeback and frees the Writeback memory (all asics).
880 * Used at driver shutdown.
882 static void amdgpu_device_wb_fini(struct amdgpu_device *adev)
884 if (adev->wb.wb_obj) {
885 amdgpu_bo_free_kernel(&adev->wb.wb_obj,
887 (void **)&adev->wb.wb);
888 adev->wb.wb_obj = NULL;
893 * amdgpu_device_wb_init- Init Writeback driver info and allocate memory
895 * @adev: amdgpu_device pointer
897 * Initializes writeback and allocates writeback memory (all asics).
898 * Used at driver startup.
899 * Returns 0 on success or an -error on failure.
901 static int amdgpu_device_wb_init(struct amdgpu_device *adev)
905 if (adev->wb.wb_obj == NULL) {
906 /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */
907 r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8,
908 PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
909 &adev->wb.wb_obj, &adev->wb.gpu_addr,
910 (void **)&adev->wb.wb);
912 dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
916 adev->wb.num_wb = AMDGPU_MAX_WB;
917 memset(&adev->wb.used, 0, sizeof(adev->wb.used));
919 /* clear wb memory */
920 memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8);
927 * amdgpu_device_wb_get - Allocate a wb entry
929 * @adev: amdgpu_device pointer
932 * Allocate a wb slot for use by the driver (all asics).
933 * Returns 0 on success or -EINVAL on failure.
935 int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb)
937 unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
939 if (offset < adev->wb.num_wb) {
940 __set_bit(offset, adev->wb.used);
941 *wb = offset << 3; /* convert to dw offset */
949 * amdgpu_device_wb_free - Free a wb entry
951 * @adev: amdgpu_device pointer
954 * Free a wb slot allocated for use by the driver (all asics)
956 void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb)
959 if (wb < adev->wb.num_wb)
960 __clear_bit(wb, adev->wb.used);
964 * amdgpu_device_resize_fb_bar - try to resize FB BAR
966 * @adev: amdgpu_device pointer
968 * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not
969 * to fail, but if any of the BARs is not accessible after the size we abort
970 * driver loading by returning -ENODEV.
972 int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev)
974 u64 space_needed = roundup_pow_of_two(adev->gmc.real_vram_size);
975 u32 rbar_size = order_base_2(((space_needed >> 20) | 1)) - 1;
976 struct pci_bus *root;
977 struct resource *res;
983 if (amdgpu_sriov_vf(adev))
986 /* skip if the bios has already enabled large BAR */
987 if (adev->gmc.real_vram_size &&
988 (pci_resource_len(adev->pdev, 0) >= adev->gmc.real_vram_size))
991 /* Check if the root BUS has 64bit memory resources */
992 root = adev->pdev->bus;
996 pci_bus_for_each_resource(root, res, i) {
997 if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) &&
998 res->start > 0x100000000ull)
1002 /* Trying to resize is pointless without a root hub window above 4GB */
1006 /* Disable memory decoding while we change the BAR addresses and size */
1007 pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd);
1008 pci_write_config_word(adev->pdev, PCI_COMMAND,
1009 cmd & ~PCI_COMMAND_MEMORY);
1011 /* Free the VRAM and doorbell BAR, we most likely need to move both. */
1012 amdgpu_device_doorbell_fini(adev);
1013 if (adev->asic_type >= CHIP_BONAIRE)
1014 pci_release_resource(adev->pdev, 2);
1016 pci_release_resource(adev->pdev, 0);
1018 r = pci_resize_resource(adev->pdev, 0, rbar_size);
1020 DRM_INFO("Not enough PCI address space for a large BAR.");
1021 else if (r && r != -ENOTSUPP)
1022 DRM_ERROR("Problem resizing BAR0 (%d).", r);
1024 pci_assign_unassigned_bus_resources(adev->pdev->bus);
1026 /* When the doorbell or fb BAR isn't available we have no chance of
1029 r = amdgpu_device_doorbell_init(adev);
1030 if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET))
1033 pci_write_config_word(adev->pdev, PCI_COMMAND, cmd);
1039 * GPU helpers function.
1042 * amdgpu_device_need_post - check if the hw need post or not
1044 * @adev: amdgpu_device pointer
1046 * Check if the asic has been initialized (all asics) at driver startup
1047 * or post is needed if hw reset is performed.
1048 * Returns true if need or false if not.
1050 bool amdgpu_device_need_post(struct amdgpu_device *adev)
1054 if (amdgpu_sriov_vf(adev))
1057 if (amdgpu_passthrough(adev)) {
1058 /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
1059 * some old smc fw still need driver do vPost otherwise gpu hang, while
1060 * those smc fw version above 22.15 doesn't have this flaw, so we force
1061 * vpost executed for smc version below 22.15
1063 if (adev->asic_type == CHIP_FIJI) {
1066 err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
1067 /* force vPost if error occured */
1071 fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
1072 if (fw_ver < 0x00160e00)
1077 if (adev->has_hw_reset) {
1078 adev->has_hw_reset = false;
1082 /* bios scratch used on CIK+ */
1083 if (adev->asic_type >= CHIP_BONAIRE)
1084 return amdgpu_atombios_scratch_need_asic_init(adev);
1086 /* check MEM_SIZE for older asics */
1087 reg = amdgpu_asic_get_config_memsize(adev);
1089 if ((reg != 0) && (reg != 0xffffffff))
1095 /* if we get transitioned to only one device, take VGA back */
1097 * amdgpu_device_vga_set_decode - enable/disable vga decode
1099 * @cookie: amdgpu_device pointer
1100 * @state: enable/disable vga decode
1102 * Enable/disable vga decode (all asics).
1103 * Returns VGA resource flags.
1105 static unsigned int amdgpu_device_vga_set_decode(void *cookie, bool state)
1107 struct amdgpu_device *adev = cookie;
1108 amdgpu_asic_set_vga_state(adev, state);
1110 return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
1111 VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
1113 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
1117 * amdgpu_device_check_block_size - validate the vm block size
1119 * @adev: amdgpu_device pointer
1121 * Validates the vm block size specified via module parameter.
1122 * The vm block size defines number of bits in page table versus page directory,
1123 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
1124 * page table and the remaining bits are in the page directory.
1126 static void amdgpu_device_check_block_size(struct amdgpu_device *adev)
1128 /* defines number of bits in page table versus page directory,
1129 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
1130 * page table and the remaining bits are in the page directory */
1131 if (amdgpu_vm_block_size == -1)
1134 if (amdgpu_vm_block_size < 9) {
1135 dev_warn(adev->dev, "VM page table size (%d) too small\n",
1136 amdgpu_vm_block_size);
1137 amdgpu_vm_block_size = -1;
1142 * amdgpu_device_check_vm_size - validate the vm size
1144 * @adev: amdgpu_device pointer
1146 * Validates the vm size in GB specified via module parameter.
1147 * The VM size is the size of the GPU virtual memory space in GB.
1149 static void amdgpu_device_check_vm_size(struct amdgpu_device *adev)
1151 /* no need to check the default value */
1152 if (amdgpu_vm_size == -1)
1155 if (amdgpu_vm_size < 1) {
1156 dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
1158 amdgpu_vm_size = -1;
1162 static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev)
1165 bool is_os_64 = (sizeof(void *) == 8);
1166 uint64_t total_memory;
1167 uint64_t dram_size_seven_GB = 0x1B8000000;
1168 uint64_t dram_size_three_GB = 0xB8000000;
1170 if (amdgpu_smu_memory_pool_size == 0)
1174 DRM_WARN("Not 64-bit OS, feature not supported\n");
1178 total_memory = (uint64_t)si.totalram * si.mem_unit;
1180 if ((amdgpu_smu_memory_pool_size == 1) ||
1181 (amdgpu_smu_memory_pool_size == 2)) {
1182 if (total_memory < dram_size_three_GB)
1184 } else if ((amdgpu_smu_memory_pool_size == 4) ||
1185 (amdgpu_smu_memory_pool_size == 8)) {
1186 if (total_memory < dram_size_seven_GB)
1189 DRM_WARN("Smu memory pool size not supported\n");
1192 adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28;
1197 DRM_WARN("No enough system memory\n");
1199 adev->pm.smu_prv_buffer_size = 0;
1203 * amdgpu_device_check_arguments - validate module params
1205 * @adev: amdgpu_device pointer
1207 * Validates certain module parameters and updates
1208 * the associated values used by the driver (all asics).
1210 static int amdgpu_device_check_arguments(struct amdgpu_device *adev)
1212 if (amdgpu_sched_jobs < 4) {
1213 dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
1215 amdgpu_sched_jobs = 4;
1216 } else if (!is_power_of_2(amdgpu_sched_jobs)){
1217 dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
1219 amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
1222 if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) {
1223 /* gart size must be greater or equal to 32M */
1224 dev_warn(adev->dev, "gart size (%d) too small\n",
1226 amdgpu_gart_size = -1;
1229 if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) {
1230 /* gtt size must be greater or equal to 32M */
1231 dev_warn(adev->dev, "gtt size (%d) too small\n",
1233 amdgpu_gtt_size = -1;
1236 /* valid range is between 4 and 9 inclusive */
1237 if (amdgpu_vm_fragment_size != -1 &&
1238 (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) {
1239 dev_warn(adev->dev, "valid range is between 4 and 9\n");
1240 amdgpu_vm_fragment_size = -1;
1243 if (amdgpu_sched_hw_submission < 2) {
1244 dev_warn(adev->dev, "sched hw submission jobs (%d) must be at least 2\n",
1245 amdgpu_sched_hw_submission);
1246 amdgpu_sched_hw_submission = 2;
1247 } else if (!is_power_of_2(amdgpu_sched_hw_submission)) {
1248 dev_warn(adev->dev, "sched hw submission jobs (%d) must be a power of 2\n",
1249 amdgpu_sched_hw_submission);
1250 amdgpu_sched_hw_submission = roundup_pow_of_two(amdgpu_sched_hw_submission);
1253 amdgpu_device_check_smu_prv_buffer_size(adev);
1255 amdgpu_device_check_vm_size(adev);
1257 amdgpu_device_check_block_size(adev);
1259 adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type);
1261 amdgpu_gmc_tmz_set(adev);
1263 if (amdgpu_num_kcq > 8 || amdgpu_num_kcq < 0) {
1265 dev_warn(adev->dev, "set kernel compute queue number to 8 due to invalid parameter provided by user\n");
1272 * amdgpu_switcheroo_set_state - set switcheroo state
1274 * @pdev: pci dev pointer
1275 * @state: vga_switcheroo state
1277 * Callback for the switcheroo driver. Suspends or resumes the
1278 * the asics before or after it is powered up using ACPI methods.
1280 static void amdgpu_switcheroo_set_state(struct pci_dev *pdev,
1281 enum vga_switcheroo_state state)
1283 struct drm_device *dev = pci_get_drvdata(pdev);
1286 if (amdgpu_device_supports_boco(dev) && state == VGA_SWITCHEROO_OFF)
1289 if (state == VGA_SWITCHEROO_ON) {
1290 pr_info("switched on\n");
1291 /* don't suspend or resume card normally */
1292 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1294 pci_set_power_state(dev->pdev, PCI_D0);
1295 pci_restore_state(dev->pdev);
1296 r = pci_enable_device(dev->pdev);
1298 DRM_WARN("pci_enable_device failed (%d)\n", r);
1299 amdgpu_device_resume(dev, true);
1301 dev->switch_power_state = DRM_SWITCH_POWER_ON;
1302 drm_kms_helper_poll_enable(dev);
1304 pr_info("switched off\n");
1305 drm_kms_helper_poll_disable(dev);
1306 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1307 amdgpu_device_suspend(dev, true);
1308 pci_save_state(dev->pdev);
1309 /* Shut down the device */
1310 pci_disable_device(dev->pdev);
1311 pci_set_power_state(dev->pdev, PCI_D3cold);
1312 dev->switch_power_state = DRM_SWITCH_POWER_OFF;
1317 * amdgpu_switcheroo_can_switch - see if switcheroo state can change
1319 * @pdev: pci dev pointer
1321 * Callback for the switcheroo driver. Check of the switcheroo
1322 * state can be changed.
1323 * Returns true if the state can be changed, false if not.
1325 static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
1327 struct drm_device *dev = pci_get_drvdata(pdev);
1330 * FIXME: open_count is protected by drm_global_mutex but that would lead to
1331 * locking inversion with the driver load path. And the access here is
1332 * completely racy anyway. So don't bother with locking for now.
1334 return atomic_read(&dev->open_count) == 0;
1337 static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
1338 .set_gpu_state = amdgpu_switcheroo_set_state,
1340 .can_switch = amdgpu_switcheroo_can_switch,
1344 * amdgpu_device_ip_set_clockgating_state - set the CG state
1346 * @dev: amdgpu_device pointer
1347 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1348 * @state: clockgating state (gate or ungate)
1350 * Sets the requested clockgating state for all instances of
1351 * the hardware IP specified.
1352 * Returns the error code from the last instance.
1354 int amdgpu_device_ip_set_clockgating_state(void *dev,
1355 enum amd_ip_block_type block_type,
1356 enum amd_clockgating_state state)
1358 struct amdgpu_device *adev = dev;
1361 for (i = 0; i < adev->num_ip_blocks; i++) {
1362 if (!adev->ip_blocks[i].status.valid)
1364 if (adev->ip_blocks[i].version->type != block_type)
1366 if (!adev->ip_blocks[i].version->funcs->set_clockgating_state)
1368 r = adev->ip_blocks[i].version->funcs->set_clockgating_state(
1369 (void *)adev, state);
1371 DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",
1372 adev->ip_blocks[i].version->funcs->name, r);
1378 * amdgpu_device_ip_set_powergating_state - set the PG state
1380 * @dev: amdgpu_device pointer
1381 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1382 * @state: powergating state (gate or ungate)
1384 * Sets the requested powergating state for all instances of
1385 * the hardware IP specified.
1386 * Returns the error code from the last instance.
1388 int amdgpu_device_ip_set_powergating_state(void *dev,
1389 enum amd_ip_block_type block_type,
1390 enum amd_powergating_state state)
1392 struct amdgpu_device *adev = dev;
1395 for (i = 0; i < adev->num_ip_blocks; i++) {
1396 if (!adev->ip_blocks[i].status.valid)
1398 if (adev->ip_blocks[i].version->type != block_type)
1400 if (!adev->ip_blocks[i].version->funcs->set_powergating_state)
1402 r = adev->ip_blocks[i].version->funcs->set_powergating_state(
1403 (void *)adev, state);
1405 DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",
1406 adev->ip_blocks[i].version->funcs->name, r);
1412 * amdgpu_device_ip_get_clockgating_state - get the CG state
1414 * @adev: amdgpu_device pointer
1415 * @flags: clockgating feature flags
1417 * Walks the list of IPs on the device and updates the clockgating
1418 * flags for each IP.
1419 * Updates @flags with the feature flags for each hardware IP where
1420 * clockgating is enabled.
1422 void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev,
1427 for (i = 0; i < adev->num_ip_blocks; i++) {
1428 if (!adev->ip_blocks[i].status.valid)
1430 if (adev->ip_blocks[i].version->funcs->get_clockgating_state)
1431 adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags);
1436 * amdgpu_device_ip_wait_for_idle - wait for idle
1438 * @adev: amdgpu_device pointer
1439 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1441 * Waits for the request hardware IP to be idle.
1442 * Returns 0 for success or a negative error code on failure.
1444 int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev,
1445 enum amd_ip_block_type block_type)
1449 for (i = 0; i < adev->num_ip_blocks; i++) {
1450 if (!adev->ip_blocks[i].status.valid)
1452 if (adev->ip_blocks[i].version->type == block_type) {
1453 r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
1464 * amdgpu_device_ip_is_idle - is the hardware IP idle
1466 * @adev: amdgpu_device pointer
1467 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1469 * Check if the hardware IP is idle or not.
1470 * Returns true if it the IP is idle, false if not.
1472 bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev,
1473 enum amd_ip_block_type block_type)
1477 for (i = 0; i < adev->num_ip_blocks; i++) {
1478 if (!adev->ip_blocks[i].status.valid)
1480 if (adev->ip_blocks[i].version->type == block_type)
1481 return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
1488 * amdgpu_device_ip_get_ip_block - get a hw IP pointer
1490 * @adev: amdgpu_device pointer
1491 * @type: Type of hardware IP (SMU, GFX, UVD, etc.)
1493 * Returns a pointer to the hardware IP block structure
1494 * if it exists for the asic, otherwise NULL.
1496 struct amdgpu_ip_block *
1497 amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,
1498 enum amd_ip_block_type type)
1502 for (i = 0; i < adev->num_ip_blocks; i++)
1503 if (adev->ip_blocks[i].version->type == type)
1504 return &adev->ip_blocks[i];
1510 * amdgpu_device_ip_block_version_cmp
1512 * @adev: amdgpu_device pointer
1513 * @type: enum amd_ip_block_type
1514 * @major: major version
1515 * @minor: minor version
1517 * return 0 if equal or greater
1518 * return 1 if smaller or the ip_block doesn't exist
1520 int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev,
1521 enum amd_ip_block_type type,
1522 u32 major, u32 minor)
1524 struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type);
1526 if (ip_block && ((ip_block->version->major > major) ||
1527 ((ip_block->version->major == major) &&
1528 (ip_block->version->minor >= minor))))
1535 * amdgpu_device_ip_block_add
1537 * @adev: amdgpu_device pointer
1538 * @ip_block_version: pointer to the IP to add
1540 * Adds the IP block driver information to the collection of IPs
1543 int amdgpu_device_ip_block_add(struct amdgpu_device *adev,
1544 const struct amdgpu_ip_block_version *ip_block_version)
1546 if (!ip_block_version)
1549 DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks,
1550 ip_block_version->funcs->name);
1552 adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
1558 * amdgpu_device_enable_virtual_display - enable virtual display feature
1560 * @adev: amdgpu_device pointer
1562 * Enabled the virtual display feature if the user has enabled it via
1563 * the module parameter virtual_display. This feature provides a virtual
1564 * display hardware on headless boards or in virtualized environments.
1565 * This function parses and validates the configuration string specified by
1566 * the user and configues the virtual display configuration (number of
1567 * virtual connectors, crtcs, etc.) specified.
1569 static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
1571 adev->enable_virtual_display = false;
1573 if (amdgpu_virtual_display) {
1574 struct drm_device *ddev = adev_to_drm(adev);
1575 const char *pci_address_name = pci_name(ddev->pdev);
1576 char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
1578 pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
1579 pciaddstr_tmp = pciaddstr;
1580 while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
1581 pciaddname = strsep(&pciaddname_tmp, ",");
1582 if (!strcmp("all", pciaddname)
1583 || !strcmp(pci_address_name, pciaddname)) {
1587 adev->enable_virtual_display = true;
1590 res = kstrtol(pciaddname_tmp, 10,
1598 adev->mode_info.num_crtc = num_crtc;
1600 adev->mode_info.num_crtc = 1;
1606 DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
1607 amdgpu_virtual_display, pci_address_name,
1608 adev->enable_virtual_display, adev->mode_info.num_crtc);
1615 * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware
1617 * @adev: amdgpu_device pointer
1619 * Parses the asic configuration parameters specified in the gpu info
1620 * firmware and makes them availale to the driver for use in configuring
1622 * Returns 0 on success, -EINVAL on failure.
1624 static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev)
1626 const char *chip_name;
1629 const struct gpu_info_firmware_header_v1_0 *hdr;
1631 adev->firmware.gpu_info_fw = NULL;
1633 if (adev->mman.discovery_bin) {
1634 amdgpu_discovery_get_gfx_info(adev);
1637 * FIXME: The bounding box is still needed by Navi12, so
1638 * temporarily read it from gpu_info firmware. Should be droped
1639 * when DAL no longer needs it.
1641 if (adev->asic_type != CHIP_NAVI12)
1645 switch (adev->asic_type) {
1646 #ifdef CONFIG_DRM_AMDGPU_SI
1653 #ifdef CONFIG_DRM_AMDGPU_CIK
1663 case CHIP_POLARIS10:
1664 case CHIP_POLARIS11:
1665 case CHIP_POLARIS12:
1673 chip_name = "vega10";
1676 chip_name = "vega12";
1679 if (adev->apu_flags & AMD_APU_IS_RAVEN2)
1680 chip_name = "raven2";
1681 else if (adev->apu_flags & AMD_APU_IS_PICASSO)
1682 chip_name = "picasso";
1684 chip_name = "raven";
1687 chip_name = "arcturus";
1690 chip_name = "renoir";
1693 chip_name = "navi10";
1696 chip_name = "navi14";
1699 chip_name = "navi12";
1701 case CHIP_SIENNA_CICHLID:
1702 chip_name = "sienna_cichlid";
1704 case CHIP_NAVY_FLOUNDER:
1705 chip_name = "navy_flounder";
1709 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name);
1710 err = request_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev);
1713 "Failed to load gpu_info firmware \"%s\"\n",
1717 err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw);
1720 "Failed to validate gpu_info firmware \"%s\"\n",
1725 hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data;
1726 amdgpu_ucode_print_gpu_info_hdr(&hdr->header);
1728 switch (hdr->version_major) {
1731 const struct gpu_info_firmware_v1_0 *gpu_info_fw =
1732 (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data +
1733 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1736 * Should be droped when DAL no longer needs it.
1738 if (adev->asic_type == CHIP_NAVI12)
1739 goto parse_soc_bounding_box;
1741 adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se);
1742 adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh);
1743 adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se);
1744 adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se);
1745 adev->gfx.config.max_texture_channel_caches =
1746 le32_to_cpu(gpu_info_fw->gc_num_tccs);
1747 adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs);
1748 adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds);
1749 adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth);
1750 adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth);
1751 adev->gfx.config.double_offchip_lds_buf =
1752 le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer);
1753 adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size);
1754 adev->gfx.cu_info.max_waves_per_simd =
1755 le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd);
1756 adev->gfx.cu_info.max_scratch_slots_per_cu =
1757 le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu);
1758 adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size);
1759 if (hdr->version_minor >= 1) {
1760 const struct gpu_info_firmware_v1_1 *gpu_info_fw =
1761 (const struct gpu_info_firmware_v1_1 *)(adev->firmware.gpu_info_fw->data +
1762 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1763 adev->gfx.config.num_sc_per_sh =
1764 le32_to_cpu(gpu_info_fw->num_sc_per_sh);
1765 adev->gfx.config.num_packer_per_sc =
1766 le32_to_cpu(gpu_info_fw->num_packer_per_sc);
1769 parse_soc_bounding_box:
1771 * soc bounding box info is not integrated in disocovery table,
1772 * we always need to parse it from gpu info firmware if needed.
1774 if (hdr->version_minor == 2) {
1775 const struct gpu_info_firmware_v1_2 *gpu_info_fw =
1776 (const struct gpu_info_firmware_v1_2 *)(adev->firmware.gpu_info_fw->data +
1777 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1778 adev->dm.soc_bounding_box = &gpu_info_fw->soc_bounding_box;
1784 "Unsupported gpu_info table %d\n", hdr->header.ucode_version);
1793 * amdgpu_device_ip_early_init - run early init for hardware IPs
1795 * @adev: amdgpu_device pointer
1797 * Early initialization pass for hardware IPs. The hardware IPs that make
1798 * up each asic are discovered each IP's early_init callback is run. This
1799 * is the first stage in initializing the asic.
1800 * Returns 0 on success, negative error code on failure.
1802 static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
1806 amdgpu_device_enable_virtual_display(adev);
1808 if (amdgpu_sriov_vf(adev)) {
1809 r = amdgpu_virt_request_full_gpu(adev, true);
1814 switch (adev->asic_type) {
1815 #ifdef CONFIG_DRM_AMDGPU_SI
1821 adev->family = AMDGPU_FAMILY_SI;
1822 r = si_set_ip_blocks(adev);
1827 #ifdef CONFIG_DRM_AMDGPU_CIK
1833 if (adev->flags & AMD_IS_APU)
1834 adev->family = AMDGPU_FAMILY_KV;
1836 adev->family = AMDGPU_FAMILY_CI;
1838 r = cik_set_ip_blocks(adev);
1846 case CHIP_POLARIS10:
1847 case CHIP_POLARIS11:
1848 case CHIP_POLARIS12:
1852 if (adev->flags & AMD_IS_APU)
1853 adev->family = AMDGPU_FAMILY_CZ;
1855 adev->family = AMDGPU_FAMILY_VI;
1857 r = vi_set_ip_blocks(adev);
1867 if (adev->flags & AMD_IS_APU)
1868 adev->family = AMDGPU_FAMILY_RV;
1870 adev->family = AMDGPU_FAMILY_AI;
1872 r = soc15_set_ip_blocks(adev);
1879 case CHIP_SIENNA_CICHLID:
1880 case CHIP_NAVY_FLOUNDER:
1881 adev->family = AMDGPU_FAMILY_NV;
1883 r = nv_set_ip_blocks(adev);
1888 /* FIXME: not supported yet */
1892 amdgpu_amdkfd_device_probe(adev);
1894 adev->pm.pp_feature = amdgpu_pp_feature_mask;
1895 if (amdgpu_sriov_vf(adev) || sched_policy == KFD_SCHED_POLICY_NO_HWS)
1896 adev->pm.pp_feature &= ~PP_GFXOFF_MASK;
1898 for (i = 0; i < adev->num_ip_blocks; i++) {
1899 if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
1900 DRM_ERROR("disabled ip block: %d <%s>\n",
1901 i, adev->ip_blocks[i].version->funcs->name);
1902 adev->ip_blocks[i].status.valid = false;
1904 if (adev->ip_blocks[i].version->funcs->early_init) {
1905 r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
1907 adev->ip_blocks[i].status.valid = false;
1909 DRM_ERROR("early_init of IP block <%s> failed %d\n",
1910 adev->ip_blocks[i].version->funcs->name, r);
1913 adev->ip_blocks[i].status.valid = true;
1916 adev->ip_blocks[i].status.valid = true;
1919 /* get the vbios after the asic_funcs are set up */
1920 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) {
1921 r = amdgpu_device_parse_gpu_info_fw(adev);
1926 if (!amdgpu_get_bios(adev))
1929 r = amdgpu_atombios_init(adev);
1931 dev_err(adev->dev, "amdgpu_atombios_init failed\n");
1932 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0);
1938 adev->cg_flags &= amdgpu_cg_mask;
1939 adev->pg_flags &= amdgpu_pg_mask;
1944 static int amdgpu_device_ip_hw_init_phase1(struct amdgpu_device *adev)
1948 for (i = 0; i < adev->num_ip_blocks; i++) {
1949 if (!adev->ip_blocks[i].status.sw)
1951 if (adev->ip_blocks[i].status.hw)
1953 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
1954 (amdgpu_sriov_vf(adev) && (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)) ||
1955 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
1956 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
1958 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1959 adev->ip_blocks[i].version->funcs->name, r);
1962 adev->ip_blocks[i].status.hw = true;
1969 static int amdgpu_device_ip_hw_init_phase2(struct amdgpu_device *adev)
1973 for (i = 0; i < adev->num_ip_blocks; i++) {
1974 if (!adev->ip_blocks[i].status.sw)
1976 if (adev->ip_blocks[i].status.hw)
1978 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
1980 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1981 adev->ip_blocks[i].version->funcs->name, r);
1984 adev->ip_blocks[i].status.hw = true;
1990 static int amdgpu_device_fw_loading(struct amdgpu_device *adev)
1994 uint32_t smu_version;
1996 if (adev->asic_type >= CHIP_VEGA10) {
1997 for (i = 0; i < adev->num_ip_blocks; i++) {
1998 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_PSP)
2001 /* no need to do the fw loading again if already done*/
2002 if (adev->ip_blocks[i].status.hw == true)
2005 if (amdgpu_in_reset(adev) || adev->in_suspend) {
2006 r = adev->ip_blocks[i].version->funcs->resume(adev);
2008 DRM_ERROR("resume of IP block <%s> failed %d\n",
2009 adev->ip_blocks[i].version->funcs->name, r);
2013 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
2015 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
2016 adev->ip_blocks[i].version->funcs->name, r);
2021 adev->ip_blocks[i].status.hw = true;
2026 if (!amdgpu_sriov_vf(adev) || adev->asic_type == CHIP_TONGA)
2027 r = amdgpu_pm_load_smu_firmware(adev, &smu_version);
2033 * amdgpu_device_ip_init - run init for hardware IPs
2035 * @adev: amdgpu_device pointer
2037 * Main initialization pass for hardware IPs. The list of all the hardware
2038 * IPs that make up the asic is walked and the sw_init and hw_init callbacks
2039 * are run. sw_init initializes the software state associated with each IP
2040 * and hw_init initializes the hardware associated with each IP.
2041 * Returns 0 on success, negative error code on failure.
2043 static int amdgpu_device_ip_init(struct amdgpu_device *adev)
2047 r = amdgpu_ras_init(adev);
2051 for (i = 0; i < adev->num_ip_blocks; i++) {
2052 if (!adev->ip_blocks[i].status.valid)
2054 r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
2056 DRM_ERROR("sw_init of IP block <%s> failed %d\n",
2057 adev->ip_blocks[i].version->funcs->name, r);
2060 adev->ip_blocks[i].status.sw = true;
2062 /* need to do gmc hw init early so we can allocate gpu mem */
2063 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
2064 r = amdgpu_device_vram_scratch_init(adev);
2066 DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r);
2069 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
2071 DRM_ERROR("hw_init %d failed %d\n", i, r);
2074 r = amdgpu_device_wb_init(adev);
2076 DRM_ERROR("amdgpu_device_wb_init failed %d\n", r);
2079 adev->ip_blocks[i].status.hw = true;
2081 /* right after GMC hw init, we create CSA */
2082 if (amdgpu_mcbp || amdgpu_sriov_vf(adev)) {
2083 r = amdgpu_allocate_static_csa(adev, &adev->virt.csa_obj,
2084 AMDGPU_GEM_DOMAIN_VRAM,
2087 DRM_ERROR("allocate CSA failed %d\n", r);
2094 if (amdgpu_sriov_vf(adev))
2095 amdgpu_virt_init_data_exchange(adev);
2097 r = amdgpu_ib_pool_init(adev);
2099 dev_err(adev->dev, "IB initialization failed (%d).\n", r);
2100 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r);
2104 r = amdgpu_ucode_create_bo(adev); /* create ucode bo when sw_init complete*/
2108 r = amdgpu_device_ip_hw_init_phase1(adev);
2112 r = amdgpu_device_fw_loading(adev);
2116 r = amdgpu_device_ip_hw_init_phase2(adev);
2121 * retired pages will be loaded from eeprom and reserved here,
2122 * it should be called after amdgpu_device_ip_hw_init_phase2 since
2123 * for some ASICs the RAS EEPROM code relies on SMU fully functioning
2124 * for I2C communication which only true at this point.
2126 * amdgpu_ras_recovery_init may fail, but the upper only cares the
2127 * failure from bad gpu situation and stop amdgpu init process
2128 * accordingly. For other failed cases, it will still release all
2129 * the resource and print error message, rather than returning one
2130 * negative value to upper level.
2132 * Note: theoretically, this should be called before all vram allocations
2133 * to protect retired page from abusing
2135 r = amdgpu_ras_recovery_init(adev);
2139 if (adev->gmc.xgmi.num_physical_nodes > 1)
2140 amdgpu_xgmi_add_device(adev);
2141 amdgpu_amdkfd_device_init(adev);
2143 amdgpu_fru_get_product_info(adev);
2146 if (amdgpu_sriov_vf(adev))
2147 amdgpu_virt_release_full_gpu(adev, true);
2153 * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer
2155 * @adev: amdgpu_device pointer
2157 * Writes a reset magic value to the gart pointer in VRAM. The driver calls
2158 * this function before a GPU reset. If the value is retained after a
2159 * GPU reset, VRAM has not been lost. Some GPU resets may destry VRAM contents.
2161 static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev)
2163 memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
2167 * amdgpu_device_check_vram_lost - check if vram is valid
2169 * @adev: amdgpu_device pointer
2171 * Checks the reset magic value written to the gart pointer in VRAM.
2172 * The driver calls this after a GPU reset to see if the contents of
2173 * VRAM is lost or now.
2174 * returns true if vram is lost, false if not.
2176 static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev)
2178 if (memcmp(adev->gart.ptr, adev->reset_magic,
2179 AMDGPU_RESET_MAGIC_NUM))
2182 if (!amdgpu_in_reset(adev))
2186 * For all ASICs with baco/mode1 reset, the VRAM is
2187 * always assumed to be lost.
2189 switch (amdgpu_asic_reset_method(adev)) {
2190 case AMD_RESET_METHOD_BACO:
2191 case AMD_RESET_METHOD_MODE1:
2199 * amdgpu_device_set_cg_state - set clockgating for amdgpu device
2201 * @adev: amdgpu_device pointer
2202 * @state: clockgating state (gate or ungate)
2204 * The list of all the hardware IPs that make up the asic is walked and the
2205 * set_clockgating_state callbacks are run.
2206 * Late initialization pass enabling clockgating for hardware IPs.
2207 * Fini or suspend, pass disabling clockgating for hardware IPs.
2208 * Returns 0 on success, negative error code on failure.
2211 static int amdgpu_device_set_cg_state(struct amdgpu_device *adev,
2212 enum amd_clockgating_state state)
2216 if (amdgpu_emu_mode == 1)
2219 for (j = 0; j < adev->num_ip_blocks; j++) {
2220 i = state == AMD_CG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
2221 if (!adev->ip_blocks[i].status.late_initialized)
2223 /* skip CG for VCE/UVD, it's handled specially */
2224 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
2225 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
2226 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
2227 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG &&
2228 adev->ip_blocks[i].version->funcs->set_clockgating_state) {
2229 /* enable clockgating to save power */
2230 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
2233 DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
2234 adev->ip_blocks[i].version->funcs->name, r);
2243 static int amdgpu_device_set_pg_state(struct amdgpu_device *adev, enum amd_powergating_state state)
2247 if (amdgpu_emu_mode == 1)
2250 for (j = 0; j < adev->num_ip_blocks; j++) {
2251 i = state == AMD_PG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
2252 if (!adev->ip_blocks[i].status.late_initialized)
2254 /* skip CG for VCE/UVD, it's handled specially */
2255 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
2256 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
2257 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
2258 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG &&
2259 adev->ip_blocks[i].version->funcs->set_powergating_state) {
2260 /* enable powergating to save power */
2261 r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev,
2264 DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n",
2265 adev->ip_blocks[i].version->funcs->name, r);
2273 static int amdgpu_device_enable_mgpu_fan_boost(void)
2275 struct amdgpu_gpu_instance *gpu_ins;
2276 struct amdgpu_device *adev;
2279 mutex_lock(&mgpu_info.mutex);
2282 * MGPU fan boost feature should be enabled
2283 * only when there are two or more dGPUs in
2286 if (mgpu_info.num_dgpu < 2)
2289 for (i = 0; i < mgpu_info.num_dgpu; i++) {
2290 gpu_ins = &(mgpu_info.gpu_ins[i]);
2291 adev = gpu_ins->adev;
2292 if (!(adev->flags & AMD_IS_APU) &&
2293 !gpu_ins->mgpu_fan_enabled) {
2294 ret = amdgpu_dpm_enable_mgpu_fan_boost(adev);
2298 gpu_ins->mgpu_fan_enabled = 1;
2303 mutex_unlock(&mgpu_info.mutex);
2309 * amdgpu_device_ip_late_init - run late init for hardware IPs
2311 * @adev: amdgpu_device pointer
2313 * Late initialization pass for hardware IPs. The list of all the hardware
2314 * IPs that make up the asic is walked and the late_init callbacks are run.
2315 * late_init covers any special initialization that an IP requires
2316 * after all of the have been initialized or something that needs to happen
2317 * late in the init process.
2318 * Returns 0 on success, negative error code on failure.
2320 static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
2322 struct amdgpu_gpu_instance *gpu_instance;
2325 for (i = 0; i < adev->num_ip_blocks; i++) {
2326 if (!adev->ip_blocks[i].status.hw)
2328 if (adev->ip_blocks[i].version->funcs->late_init) {
2329 r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
2331 DRM_ERROR("late_init of IP block <%s> failed %d\n",
2332 adev->ip_blocks[i].version->funcs->name, r);
2336 adev->ip_blocks[i].status.late_initialized = true;
2339 amdgpu_ras_set_error_query_ready(adev, true);
2341 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE);
2342 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE);
2344 amdgpu_device_fill_reset_magic(adev);
2346 r = amdgpu_device_enable_mgpu_fan_boost();
2348 DRM_ERROR("enable mgpu fan boost failed (%d).\n", r);
2351 if (adev->gmc.xgmi.num_physical_nodes > 1) {
2352 mutex_lock(&mgpu_info.mutex);
2355 * Reset device p-state to low as this was booted with high.
2357 * This should be performed only after all devices from the same
2358 * hive get initialized.
2360 * However, it's unknown how many device in the hive in advance.
2361 * As this is counted one by one during devices initializations.
2363 * So, we wait for all XGMI interlinked devices initialized.
2364 * This may bring some delays as those devices may come from
2365 * different hives. But that should be OK.
2367 if (mgpu_info.num_dgpu == adev->gmc.xgmi.num_physical_nodes) {
2368 for (i = 0; i < mgpu_info.num_gpu; i++) {
2369 gpu_instance = &(mgpu_info.gpu_ins[i]);
2370 if (gpu_instance->adev->flags & AMD_IS_APU)
2373 r = amdgpu_xgmi_set_pstate(gpu_instance->adev,
2374 AMDGPU_XGMI_PSTATE_MIN);
2376 DRM_ERROR("pstate setting failed (%d).\n", r);
2382 mutex_unlock(&mgpu_info.mutex);
2389 * amdgpu_device_ip_fini - run fini for hardware IPs
2391 * @adev: amdgpu_device pointer
2393 * Main teardown pass for hardware IPs. The list of all the hardware
2394 * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks
2395 * are run. hw_fini tears down the hardware associated with each IP
2396 * and sw_fini tears down any software state associated with each IP.
2397 * Returns 0 on success, negative error code on failure.
2399 static int amdgpu_device_ip_fini(struct amdgpu_device *adev)
2403 if (amdgpu_sriov_vf(adev) && adev->virt.ras_init_done)
2404 amdgpu_virt_release_ras_err_handler_data(adev);
2406 amdgpu_ras_pre_fini(adev);
2408 if (adev->gmc.xgmi.num_physical_nodes > 1)
2409 amdgpu_xgmi_remove_device(adev);
2411 amdgpu_amdkfd_device_fini(adev);
2413 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
2414 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
2416 /* need to disable SMC first */
2417 for (i = 0; i < adev->num_ip_blocks; i++) {
2418 if (!adev->ip_blocks[i].status.hw)
2420 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
2421 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
2422 /* XXX handle errors */
2424 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
2425 adev->ip_blocks[i].version->funcs->name, r);
2427 adev->ip_blocks[i].status.hw = false;
2432 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2433 if (!adev->ip_blocks[i].status.hw)
2436 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
2437 /* XXX handle errors */
2439 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
2440 adev->ip_blocks[i].version->funcs->name, r);
2443 adev->ip_blocks[i].status.hw = false;
2447 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2448 if (!adev->ip_blocks[i].status.sw)
2451 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
2452 amdgpu_ucode_free_bo(adev);
2453 amdgpu_free_static_csa(&adev->virt.csa_obj);
2454 amdgpu_device_wb_fini(adev);
2455 amdgpu_device_vram_scratch_fini(adev);
2456 amdgpu_ib_pool_fini(adev);
2459 r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
2460 /* XXX handle errors */
2462 DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
2463 adev->ip_blocks[i].version->funcs->name, r);
2465 adev->ip_blocks[i].status.sw = false;
2466 adev->ip_blocks[i].status.valid = false;
2469 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2470 if (!adev->ip_blocks[i].status.late_initialized)
2472 if (adev->ip_blocks[i].version->funcs->late_fini)
2473 adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
2474 adev->ip_blocks[i].status.late_initialized = false;
2477 amdgpu_ras_fini(adev);
2479 if (amdgpu_sriov_vf(adev))
2480 if (amdgpu_virt_release_full_gpu(adev, false))
2481 DRM_ERROR("failed to release exclusive mode on fini\n");
2487 * amdgpu_device_delayed_init_work_handler - work handler for IB tests
2489 * @work: work_struct.
2491 static void amdgpu_device_delayed_init_work_handler(struct work_struct *work)
2493 struct amdgpu_device *adev =
2494 container_of(work, struct amdgpu_device, delayed_init_work.work);
2497 r = amdgpu_ib_ring_tests(adev);
2499 DRM_ERROR("ib ring test failed (%d).\n", r);
2502 static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work)
2504 struct amdgpu_device *adev =
2505 container_of(work, struct amdgpu_device, gfx.gfx_off_delay_work.work);
2507 mutex_lock(&adev->gfx.gfx_off_mutex);
2508 if (!adev->gfx.gfx_off_state && !adev->gfx.gfx_off_req_count) {
2509 if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, true))
2510 adev->gfx.gfx_off_state = true;
2512 mutex_unlock(&adev->gfx.gfx_off_mutex);
2516 * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1)
2518 * @adev: amdgpu_device pointer
2520 * Main suspend function for hardware IPs. The list of all the hardware
2521 * IPs that make up the asic is walked, clockgating is disabled and the
2522 * suspend callbacks are run. suspend puts the hardware and software state
2523 * in each IP into a state suitable for suspend.
2524 * Returns 0 on success, negative error code on failure.
2526 static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev)
2530 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
2531 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
2533 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2534 if (!adev->ip_blocks[i].status.valid)
2537 /* displays are handled separately */
2538 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_DCE)
2541 /* XXX handle errors */
2542 r = adev->ip_blocks[i].version->funcs->suspend(adev);
2543 /* XXX handle errors */
2545 DRM_ERROR("suspend of IP block <%s> failed %d\n",
2546 adev->ip_blocks[i].version->funcs->name, r);
2550 adev->ip_blocks[i].status.hw = false;
2557 * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2)
2559 * @adev: amdgpu_device pointer
2561 * Main suspend function for hardware IPs. The list of all the hardware
2562 * IPs that make up the asic is walked, clockgating is disabled and the
2563 * suspend callbacks are run. suspend puts the hardware and software state
2564 * in each IP into a state suitable for suspend.
2565 * Returns 0 on success, negative error code on failure.
2567 static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev)
2571 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2572 if (!adev->ip_blocks[i].status.valid)
2574 /* displays are handled in phase1 */
2575 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE)
2577 /* PSP lost connection when err_event_athub occurs */
2578 if (amdgpu_ras_intr_triggered() &&
2579 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
2580 adev->ip_blocks[i].status.hw = false;
2583 /* XXX handle errors */
2584 r = adev->ip_blocks[i].version->funcs->suspend(adev);
2585 /* XXX handle errors */
2587 DRM_ERROR("suspend of IP block <%s> failed %d\n",
2588 adev->ip_blocks[i].version->funcs->name, r);
2590 adev->ip_blocks[i].status.hw = false;
2591 /* handle putting the SMC in the appropriate state */
2592 if(!amdgpu_sriov_vf(adev)){
2593 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
2594 r = amdgpu_dpm_set_mp1_state(adev, adev->mp1_state);
2596 DRM_ERROR("SMC failed to set mp1 state %d, %d\n",
2597 adev->mp1_state, r);
2602 adev->ip_blocks[i].status.hw = false;
2609 * amdgpu_device_ip_suspend - run suspend for hardware IPs
2611 * @adev: amdgpu_device pointer
2613 * Main suspend function for hardware IPs. The list of all the hardware
2614 * IPs that make up the asic is walked, clockgating is disabled and the
2615 * suspend callbacks are run. suspend puts the hardware and software state
2616 * in each IP into a state suitable for suspend.
2617 * Returns 0 on success, negative error code on failure.
2619 int amdgpu_device_ip_suspend(struct amdgpu_device *adev)
2623 if (amdgpu_sriov_vf(adev))
2624 amdgpu_virt_request_full_gpu(adev, false);
2626 r = amdgpu_device_ip_suspend_phase1(adev);
2629 r = amdgpu_device_ip_suspend_phase2(adev);
2631 if (amdgpu_sriov_vf(adev))
2632 amdgpu_virt_release_full_gpu(adev, false);
2637 static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
2641 static enum amd_ip_block_type ip_order[] = {
2642 AMD_IP_BLOCK_TYPE_GMC,
2643 AMD_IP_BLOCK_TYPE_COMMON,
2644 AMD_IP_BLOCK_TYPE_PSP,
2645 AMD_IP_BLOCK_TYPE_IH,
2648 for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
2650 struct amdgpu_ip_block *block;
2652 block = &adev->ip_blocks[i];
2653 block->status.hw = false;
2655 for (j = 0; j < ARRAY_SIZE(ip_order); j++) {
2657 if (block->version->type != ip_order[j] ||
2658 !block->status.valid)
2661 r = block->version->funcs->hw_init(adev);
2662 DRM_INFO("RE-INIT-early: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
2665 block->status.hw = true;
2672 static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
2676 static enum amd_ip_block_type ip_order[] = {
2677 AMD_IP_BLOCK_TYPE_SMC,
2678 AMD_IP_BLOCK_TYPE_DCE,
2679 AMD_IP_BLOCK_TYPE_GFX,
2680 AMD_IP_BLOCK_TYPE_SDMA,
2681 AMD_IP_BLOCK_TYPE_UVD,
2682 AMD_IP_BLOCK_TYPE_VCE,
2683 AMD_IP_BLOCK_TYPE_VCN
2686 for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
2688 struct amdgpu_ip_block *block;
2690 for (j = 0; j < adev->num_ip_blocks; j++) {
2691 block = &adev->ip_blocks[j];
2693 if (block->version->type != ip_order[i] ||
2694 !block->status.valid ||
2698 if (block->version->type == AMD_IP_BLOCK_TYPE_SMC)
2699 r = block->version->funcs->resume(adev);
2701 r = block->version->funcs->hw_init(adev);
2703 DRM_INFO("RE-INIT-late: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
2706 block->status.hw = true;
2714 * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs
2716 * @adev: amdgpu_device pointer
2718 * First resume function for hardware IPs. The list of all the hardware
2719 * IPs that make up the asic is walked and the resume callbacks are run for
2720 * COMMON, GMC, and IH. resume puts the hardware into a functional state
2721 * after a suspend and updates the software state as necessary. This
2722 * function is also used for restoring the GPU after a GPU reset.
2723 * Returns 0 on success, negative error code on failure.
2725 static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)
2729 for (i = 0; i < adev->num_ip_blocks; i++) {
2730 if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
2732 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2733 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2734 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
2736 r = adev->ip_blocks[i].version->funcs->resume(adev);
2738 DRM_ERROR("resume of IP block <%s> failed %d\n",
2739 adev->ip_blocks[i].version->funcs->name, r);
2742 adev->ip_blocks[i].status.hw = true;
2750 * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs
2752 * @adev: amdgpu_device pointer
2754 * First resume function for hardware IPs. The list of all the hardware
2755 * IPs that make up the asic is walked and the resume callbacks are run for
2756 * all blocks except COMMON, GMC, and IH. resume puts the hardware into a
2757 * functional state after a suspend and updates the software state as
2758 * necessary. This function is also used for restoring the GPU after a GPU
2760 * Returns 0 on success, negative error code on failure.
2762 static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)
2766 for (i = 0; i < adev->num_ip_blocks; i++) {
2767 if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
2769 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2770 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2771 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH ||
2772 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)
2774 r = adev->ip_blocks[i].version->funcs->resume(adev);
2776 DRM_ERROR("resume of IP block <%s> failed %d\n",
2777 adev->ip_blocks[i].version->funcs->name, r);
2780 adev->ip_blocks[i].status.hw = true;
2787 * amdgpu_device_ip_resume - run resume for hardware IPs
2789 * @adev: amdgpu_device pointer
2791 * Main resume function for hardware IPs. The hardware IPs
2792 * are split into two resume functions because they are
2793 * are also used in in recovering from a GPU reset and some additional
2794 * steps need to be take between them. In this case (S3/S4) they are
2796 * Returns 0 on success, negative error code on failure.
2798 static int amdgpu_device_ip_resume(struct amdgpu_device *adev)
2802 r = amdgpu_device_ip_resume_phase1(adev);
2806 r = amdgpu_device_fw_loading(adev);
2810 r = amdgpu_device_ip_resume_phase2(adev);
2816 * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV
2818 * @adev: amdgpu_device pointer
2820 * Query the VBIOS data tables to determine if the board supports SR-IOV.
2822 static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
2824 if (amdgpu_sriov_vf(adev)) {
2825 if (adev->is_atom_fw) {
2826 if (amdgpu_atomfirmware_gpu_supports_virtualization(adev))
2827 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2829 if (amdgpu_atombios_has_gpu_virtualization_table(adev))
2830 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2833 if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
2834 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
2839 * amdgpu_device_asic_has_dc_support - determine if DC supports the asic
2841 * @asic_type: AMD asic type
2843 * Check if there is DC (new modesetting infrastructre) support for an asic.
2844 * returns true if DC has support, false if not.
2846 bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
2848 switch (asic_type) {
2849 #if defined(CONFIG_DRM_AMD_DC)
2850 #if defined(CONFIG_DRM_AMD_DC_SI)
2861 * We have systems in the wild with these ASICs that require
2862 * LVDS and VGA support which is not supported with DC.
2864 * Fallback to the non-DC driver here by default so as not to
2865 * cause regressions.
2867 return amdgpu_dc > 0;
2871 case CHIP_POLARIS10:
2872 case CHIP_POLARIS11:
2873 case CHIP_POLARIS12:
2880 #if defined(CONFIG_DRM_AMD_DC_DCN)
2887 #if defined(CONFIG_DRM_AMD_DC_DCN3_0)
2888 case CHIP_SIENNA_CICHLID:
2889 case CHIP_NAVY_FLOUNDER:
2891 return amdgpu_dc != 0;
2895 DRM_INFO("Display Core has been requested via kernel parameter "
2896 "but isn't supported by ASIC, ignoring\n");
2902 * amdgpu_device_has_dc_support - check if dc is supported
2904 * @adev: amdgpu_device_pointer
2906 * Returns true for supported, false for not supported
2908 bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
2910 if (amdgpu_sriov_vf(adev) || adev->enable_virtual_display)
2913 return amdgpu_device_asic_has_dc_support(adev->asic_type);
2917 static void amdgpu_device_xgmi_reset_func(struct work_struct *__work)
2919 struct amdgpu_device *adev =
2920 container_of(__work, struct amdgpu_device, xgmi_reset_work);
2921 struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev);
2923 /* It's a bug to not have a hive within this function */
2928 * Use task barrier to synchronize all xgmi reset works across the
2929 * hive. task_barrier_enter and task_barrier_exit will block
2930 * until all the threads running the xgmi reset works reach
2931 * those points. task_barrier_full will do both blocks.
2933 if (amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) {
2935 task_barrier_enter(&hive->tb);
2936 adev->asic_reset_res = amdgpu_device_baco_enter(adev_to_drm(adev));
2938 if (adev->asic_reset_res)
2941 task_barrier_exit(&hive->tb);
2942 adev->asic_reset_res = amdgpu_device_baco_exit(adev_to_drm(adev));
2944 if (adev->asic_reset_res)
2947 if (adev->mmhub.funcs && adev->mmhub.funcs->reset_ras_error_count)
2948 adev->mmhub.funcs->reset_ras_error_count(adev);
2951 task_barrier_full(&hive->tb);
2952 adev->asic_reset_res = amdgpu_asic_reset(adev);
2956 if (adev->asic_reset_res)
2957 DRM_WARN("ASIC reset failed with error, %d for drm dev, %s",
2958 adev->asic_reset_res, adev_to_drm(adev)->unique);
2959 amdgpu_put_xgmi_hive(hive);
2962 static int amdgpu_device_get_job_timeout_settings(struct amdgpu_device *adev)
2964 char *input = amdgpu_lockup_timeout;
2965 char *timeout_setting = NULL;
2971 * By default timeout for non compute jobs is 10000.
2972 * And there is no timeout enforced on compute jobs.
2973 * In SR-IOV or passthrough mode, timeout for compute
2974 * jobs are 60000 by default.
2976 adev->gfx_timeout = msecs_to_jiffies(10000);
2977 adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout;
2978 if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev))
2979 adev->compute_timeout = msecs_to_jiffies(60000);
2981 adev->compute_timeout = MAX_SCHEDULE_TIMEOUT;
2983 if (strnlen(input, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) {
2984 while ((timeout_setting = strsep(&input, ",")) &&
2985 strnlen(timeout_setting, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) {
2986 ret = kstrtol(timeout_setting, 0, &timeout);
2993 } else if (timeout < 0) {
2994 timeout = MAX_SCHEDULE_TIMEOUT;
2996 timeout = msecs_to_jiffies(timeout);
3001 adev->gfx_timeout = timeout;
3004 adev->compute_timeout = timeout;
3007 adev->sdma_timeout = timeout;
3010 adev->video_timeout = timeout;
3017 * There is only one value specified and
3018 * it should apply to all non-compute jobs.
3021 adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout;
3022 if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev))
3023 adev->compute_timeout = adev->gfx_timeout;
3030 static const struct attribute *amdgpu_dev_attributes[] = {
3031 &dev_attr_product_name.attr,
3032 &dev_attr_product_number.attr,
3033 &dev_attr_serial_number.attr,
3034 &dev_attr_pcie_replay_count.attr,
3040 * amdgpu_device_init - initialize the driver
3042 * @adev: amdgpu_device pointer
3043 * @flags: driver flags
3045 * Initializes the driver info and hw (all asics).
3046 * Returns 0 for success or an error on failure.
3047 * Called at driver startup.
3049 int amdgpu_device_init(struct amdgpu_device *adev,
3052 struct drm_device *ddev = adev_to_drm(adev);
3053 struct pci_dev *pdev = adev->pdev;
3058 adev->shutdown = false;
3059 adev->flags = flags;
3061 if (amdgpu_force_asic_type >= 0 && amdgpu_force_asic_type < CHIP_LAST)
3062 adev->asic_type = amdgpu_force_asic_type;
3064 adev->asic_type = flags & AMD_ASIC_MASK;
3066 adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
3067 if (amdgpu_emu_mode == 1)
3068 adev->usec_timeout *= 10;
3069 adev->gmc.gart_size = 512 * 1024 * 1024;
3070 adev->accel_working = false;
3071 adev->num_rings = 0;
3072 adev->mman.buffer_funcs = NULL;
3073 adev->mman.buffer_funcs_ring = NULL;
3074 adev->vm_manager.vm_pte_funcs = NULL;
3075 adev->vm_manager.vm_pte_num_scheds = 0;
3076 adev->gmc.gmc_funcs = NULL;
3077 adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
3078 bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
3080 adev->smc_rreg = &amdgpu_invalid_rreg;
3081 adev->smc_wreg = &amdgpu_invalid_wreg;
3082 adev->pcie_rreg = &amdgpu_invalid_rreg;
3083 adev->pcie_wreg = &amdgpu_invalid_wreg;
3084 adev->pciep_rreg = &amdgpu_invalid_rreg;
3085 adev->pciep_wreg = &amdgpu_invalid_wreg;
3086 adev->pcie_rreg64 = &amdgpu_invalid_rreg64;
3087 adev->pcie_wreg64 = &amdgpu_invalid_wreg64;
3088 adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
3089 adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
3090 adev->didt_rreg = &amdgpu_invalid_rreg;
3091 adev->didt_wreg = &amdgpu_invalid_wreg;
3092 adev->gc_cac_rreg = &amdgpu_invalid_rreg;
3093 adev->gc_cac_wreg = &amdgpu_invalid_wreg;
3094 adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
3095 adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
3097 DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
3098 amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
3099 pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
3101 /* mutex initialization are all done here so we
3102 * can recall function without having locking issues */
3103 atomic_set(&adev->irq.ih.lock, 0);
3104 mutex_init(&adev->firmware.mutex);
3105 mutex_init(&adev->pm.mutex);
3106 mutex_init(&adev->gfx.gpu_clock_mutex);
3107 mutex_init(&adev->srbm_mutex);
3108 mutex_init(&adev->gfx.pipe_reserve_mutex);
3109 mutex_init(&adev->gfx.gfx_off_mutex);
3110 mutex_init(&adev->grbm_idx_mutex);
3111 mutex_init(&adev->mn_lock);
3112 mutex_init(&adev->virt.vf_errors.lock);
3113 hash_init(adev->mn_hash);
3114 atomic_set(&adev->in_gpu_reset, 0);
3115 init_rwsem(&adev->reset_sem);
3116 mutex_init(&adev->psp.mutex);
3117 mutex_init(&adev->notifier_lock);
3119 r = amdgpu_device_check_arguments(adev);
3123 spin_lock_init(&adev->mmio_idx_lock);
3124 spin_lock_init(&adev->smc_idx_lock);
3125 spin_lock_init(&adev->pcie_idx_lock);
3126 spin_lock_init(&adev->uvd_ctx_idx_lock);
3127 spin_lock_init(&adev->didt_idx_lock);
3128 spin_lock_init(&adev->gc_cac_idx_lock);
3129 spin_lock_init(&adev->se_cac_idx_lock);
3130 spin_lock_init(&adev->audio_endpt_idx_lock);
3131 spin_lock_init(&adev->mm_stats.lock);
3133 INIT_LIST_HEAD(&adev->shadow_list);
3134 mutex_init(&adev->shadow_list_lock);
3136 INIT_DELAYED_WORK(&adev->delayed_init_work,
3137 amdgpu_device_delayed_init_work_handler);
3138 INIT_DELAYED_WORK(&adev->gfx.gfx_off_delay_work,
3139 amdgpu_device_delay_enable_gfx_off);
3141 INIT_WORK(&adev->xgmi_reset_work, amdgpu_device_xgmi_reset_func);
3143 adev->gfx.gfx_off_req_count = 1;
3144 adev->pm.ac_power = power_supply_is_system_supplied() > 0;
3146 atomic_set(&adev->throttling_logging_enabled, 1);
3148 * If throttling continues, logging will be performed every minute
3149 * to avoid log flooding. "-1" is subtracted since the thermal
3150 * throttling interrupt comes every second. Thus, the total logging
3151 * interval is 59 seconds(retelimited printk interval) + 1(waiting
3152 * for throttling interrupt) = 60 seconds.
3154 ratelimit_state_init(&adev->throttling_logging_rs, (60 - 1) * HZ, 1);
3155 ratelimit_set_flags(&adev->throttling_logging_rs, RATELIMIT_MSG_ON_RELEASE);
3157 /* Registers mapping */
3158 /* TODO: block userspace mapping of io register */
3159 if (adev->asic_type >= CHIP_BONAIRE) {
3160 adev->rmmio_base = pci_resource_start(adev->pdev, 5);
3161 adev->rmmio_size = pci_resource_len(adev->pdev, 5);
3163 adev->rmmio_base = pci_resource_start(adev->pdev, 2);
3164 adev->rmmio_size = pci_resource_len(adev->pdev, 2);
3167 adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
3168 if (adev->rmmio == NULL) {
3171 DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
3172 DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
3174 /* io port mapping */
3175 for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
3176 if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) {
3177 adev->rio_mem_size = pci_resource_len(adev->pdev, i);
3178 adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size);
3182 if (adev->rio_mem == NULL)
3183 DRM_INFO("PCI I/O BAR is not found.\n");
3185 /* enable PCIE atomic ops */
3186 r = pci_enable_atomic_ops_to_root(adev->pdev,
3187 PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
3188 PCI_EXP_DEVCAP2_ATOMIC_COMP64);
3190 adev->have_atomics_support = false;
3191 DRM_INFO("PCIE atomic ops is not supported\n");
3193 adev->have_atomics_support = true;
3196 amdgpu_device_get_pcie_info(adev);
3199 DRM_INFO("MCBP is enabled\n");
3201 if (amdgpu_mes && adev->asic_type >= CHIP_NAVI10)
3202 adev->enable_mes = true;
3204 /* detect hw virtualization here */
3205 amdgpu_detect_virtualization(adev);
3207 r = amdgpu_device_get_job_timeout_settings(adev);
3209 dev_err(adev->dev, "invalid lockup_timeout parameter syntax\n");
3213 /* early init functions */
3214 r = amdgpu_device_ip_early_init(adev);
3218 /* doorbell bar mapping and doorbell index init*/
3219 amdgpu_device_doorbell_init(adev);
3221 /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
3222 /* this will fail for cards that aren't VGA class devices, just
3224 vga_client_register(adev->pdev, adev, NULL, amdgpu_device_vga_set_decode);
3226 if (amdgpu_device_supports_boco(ddev))
3228 if (amdgpu_has_atpx() &&
3229 (amdgpu_is_atpx_hybrid() ||
3230 amdgpu_has_atpx_dgpu_power_cntl()) &&
3231 !pci_is_thunderbolt_attached(adev->pdev))
3232 vga_switcheroo_register_client(adev->pdev,
3233 &amdgpu_switcheroo_ops, boco);
3235 vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
3237 if (amdgpu_emu_mode == 1) {
3238 /* post the asic on emulation mode */
3239 emu_soc_asic_init(adev);
3240 goto fence_driver_init;
3243 /* detect if we are with an SRIOV vbios */
3244 amdgpu_device_detect_sriov_bios(adev);
3246 /* check if we need to reset the asic
3247 * E.g., driver was not cleanly unloaded previously, etc.
3249 if (!amdgpu_sriov_vf(adev) && amdgpu_asic_need_reset_on_init(adev)) {
3250 r = amdgpu_asic_reset(adev);
3252 dev_err(adev->dev, "asic reset on init failed\n");
3257 pci_enable_pcie_error_reporting(adev->ddev.pdev);
3259 /* Post card if necessary */
3260 if (amdgpu_device_need_post(adev)) {
3262 dev_err(adev->dev, "no vBIOS found\n");
3266 DRM_INFO("GPU posting now...\n");
3267 r = amdgpu_device_asic_init(adev);
3269 dev_err(adev->dev, "gpu post error!\n");
3274 if (adev->is_atom_fw) {
3275 /* Initialize clocks */
3276 r = amdgpu_atomfirmware_get_clock_info(adev);
3278 dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
3279 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
3283 /* Initialize clocks */
3284 r = amdgpu_atombios_get_clock_info(adev);
3286 dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
3287 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
3290 /* init i2c buses */
3291 if (!amdgpu_device_has_dc_support(adev))
3292 amdgpu_atombios_i2c_init(adev);
3297 r = amdgpu_fence_driver_init(adev);
3299 dev_err(adev->dev, "amdgpu_fence_driver_init failed\n");
3300 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
3304 /* init the mode config */
3305 drm_mode_config_init(adev_to_drm(adev));
3307 r = amdgpu_device_ip_init(adev);
3309 /* failed in exclusive mode due to timeout */
3310 if (amdgpu_sriov_vf(adev) &&
3311 !amdgpu_sriov_runtime(adev) &&
3312 amdgpu_virt_mmio_blocked(adev) &&
3313 !amdgpu_virt_wait_reset(adev)) {
3314 dev_err(adev->dev, "VF exclusive mode timeout\n");
3315 /* Don't send request since VF is inactive. */
3316 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
3317 adev->virt.ops = NULL;
3321 dev_err(adev->dev, "amdgpu_device_ip_init failed\n");
3322 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
3327 "SE %d, SH per SE %d, CU per SH %d, active_cu_number %d\n",
3328 adev->gfx.config.max_shader_engines,
3329 adev->gfx.config.max_sh_per_se,
3330 adev->gfx.config.max_cu_per_sh,
3331 adev->gfx.cu_info.number);
3333 adev->accel_working = true;
3335 amdgpu_vm_check_compute_bug(adev);
3337 /* Initialize the buffer migration limit. */
3338 if (amdgpu_moverate >= 0)
3339 max_MBps = amdgpu_moverate;
3341 max_MBps = 8; /* Allow 8 MB/s. */
3342 /* Get a log2 for easy divisions. */
3343 adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
3345 amdgpu_fbdev_init(adev);
3347 r = amdgpu_pm_sysfs_init(adev);
3349 adev->pm_sysfs_en = false;
3350 DRM_ERROR("registering pm debugfs failed (%d).\n", r);
3352 adev->pm_sysfs_en = true;
3354 r = amdgpu_ucode_sysfs_init(adev);
3356 adev->ucode_sysfs_en = false;
3357 DRM_ERROR("Creating firmware sysfs failed (%d).\n", r);
3359 adev->ucode_sysfs_en = true;
3361 if ((amdgpu_testing & 1)) {
3362 if (adev->accel_working)
3363 amdgpu_test_moves(adev);
3365 DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n");
3367 if (amdgpu_benchmarking) {
3368 if (adev->accel_working)
3369 amdgpu_benchmark(adev, amdgpu_benchmarking);
3371 DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n");
3375 * Register gpu instance before amdgpu_device_enable_mgpu_fan_boost.
3376 * Otherwise the mgpu fan boost feature will be skipped due to the
3377 * gpu instance is counted less.
3379 amdgpu_register_gpu_instance(adev);
3381 /* enable clockgating, etc. after ib tests, etc. since some blocks require
3382 * explicit gating rather than handling it automatically.
3384 r = amdgpu_device_ip_late_init(adev);
3386 dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n");
3387 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
3392 amdgpu_ras_resume(adev);
3394 queue_delayed_work(system_wq, &adev->delayed_init_work,
3395 msecs_to_jiffies(AMDGPU_RESUME_MS));
3397 if (amdgpu_sriov_vf(adev))
3398 flush_delayed_work(&adev->delayed_init_work);
3400 r = sysfs_create_files(&adev->dev->kobj, amdgpu_dev_attributes);
3402 dev_err(adev->dev, "Could not create amdgpu device attr\n");
3406 if (IS_ENABLED(CONFIG_PERF_EVENTS))
3407 r = amdgpu_pmu_init(adev);
3409 dev_err(adev->dev, "amdgpu_pmu_init failed\n");
3414 amdgpu_vf_error_trans_all(adev);
3416 vga_switcheroo_fini_domain_pm_ops(adev->dev);
3422 * amdgpu_device_fini - tear down the driver
3424 * @adev: amdgpu_device pointer
3426 * Tear down the driver info (all asics).
3427 * Called at driver shutdown.
3429 void amdgpu_device_fini(struct amdgpu_device *adev)
3431 dev_info(adev->dev, "amdgpu: finishing device.\n");
3432 flush_delayed_work(&adev->delayed_init_work);
3433 adev->shutdown = true;
3435 /* make sure IB test finished before entering exclusive mode
3436 * to avoid preemption on IB test
3438 if (amdgpu_sriov_vf(adev))
3439 amdgpu_virt_request_full_gpu(adev, false);
3441 /* disable all interrupts */
3442 amdgpu_irq_disable_all(adev);
3443 if (adev->mode_info.mode_config_initialized){
3444 if (!amdgpu_device_has_dc_support(adev))
3445 drm_helper_force_disable_all(adev_to_drm(adev));
3447 drm_atomic_helper_shutdown(adev_to_drm(adev));
3449 amdgpu_fence_driver_fini(adev);
3450 if (adev->pm_sysfs_en)
3451 amdgpu_pm_sysfs_fini(adev);
3452 amdgpu_fbdev_fini(adev);
3453 amdgpu_device_ip_fini(adev);
3454 release_firmware(adev->firmware.gpu_info_fw);
3455 adev->firmware.gpu_info_fw = NULL;
3456 adev->accel_working = false;
3457 /* free i2c buses */
3458 if (!amdgpu_device_has_dc_support(adev))
3459 amdgpu_i2c_fini(adev);
3461 if (amdgpu_emu_mode != 1)
3462 amdgpu_atombios_fini(adev);
3466 if (amdgpu_has_atpx() &&
3467 (amdgpu_is_atpx_hybrid() ||
3468 amdgpu_has_atpx_dgpu_power_cntl()) &&
3469 !pci_is_thunderbolt_attached(adev->pdev))
3470 vga_switcheroo_unregister_client(adev->pdev);
3471 if (amdgpu_device_supports_boco(adev_to_drm(adev)))
3472 vga_switcheroo_fini_domain_pm_ops(adev->dev);
3473 vga_client_register(adev->pdev, NULL, NULL, NULL);
3475 pci_iounmap(adev->pdev, adev->rio_mem);
3476 adev->rio_mem = NULL;
3477 iounmap(adev->rmmio);
3479 amdgpu_device_doorbell_fini(adev);
3481 if (adev->ucode_sysfs_en)
3482 amdgpu_ucode_sysfs_fini(adev);
3484 sysfs_remove_files(&adev->dev->kobj, amdgpu_dev_attributes);
3485 if (IS_ENABLED(CONFIG_PERF_EVENTS))
3486 amdgpu_pmu_fini(adev);
3487 if (adev->mman.discovery_bin)
3488 amdgpu_discovery_fini(adev);
3496 * amdgpu_device_suspend - initiate device suspend
3498 * @dev: drm dev pointer
3499 * @fbcon : notify the fbdev of suspend
3501 * Puts the hw in the suspend state (all asics).
3502 * Returns 0 for success or an error on failure.
3503 * Called at driver suspend.
3505 int amdgpu_device_suspend(struct drm_device *dev, bool fbcon)
3507 struct amdgpu_device *adev;
3508 struct drm_crtc *crtc;
3509 struct drm_connector *connector;
3510 struct drm_connector_list_iter iter;
3513 adev = drm_to_adev(dev);
3515 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
3518 adev->in_suspend = true;
3519 drm_kms_helper_poll_disable(dev);
3522 amdgpu_fbdev_set_suspend(adev, 1);
3524 cancel_delayed_work_sync(&adev->delayed_init_work);
3526 if (!amdgpu_device_has_dc_support(adev)) {
3527 /* turn off display hw */
3528 drm_modeset_lock_all(dev);
3529 drm_connector_list_iter_begin(dev, &iter);
3530 drm_for_each_connector_iter(connector, &iter)
3531 drm_helper_connector_dpms(connector,
3533 drm_connector_list_iter_end(&iter);
3534 drm_modeset_unlock_all(dev);
3535 /* unpin the front buffers and cursors */
3536 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
3537 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
3538 struct drm_framebuffer *fb = crtc->primary->fb;
3539 struct amdgpu_bo *robj;
3541 if (amdgpu_crtc->cursor_bo && !adev->enable_virtual_display) {
3542 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
3543 r = amdgpu_bo_reserve(aobj, true);
3545 amdgpu_bo_unpin(aobj);
3546 amdgpu_bo_unreserve(aobj);
3550 if (fb == NULL || fb->obj[0] == NULL) {
3553 robj = gem_to_amdgpu_bo(fb->obj[0]);
3554 /* don't unpin kernel fb objects */
3555 if (!amdgpu_fbdev_robj_is_fb(adev, robj)) {
3556 r = amdgpu_bo_reserve(robj, true);
3558 amdgpu_bo_unpin(robj);
3559 amdgpu_bo_unreserve(robj);
3565 amdgpu_ras_suspend(adev);
3567 r = amdgpu_device_ip_suspend_phase1(adev);
3569 amdgpu_amdkfd_suspend(adev, !fbcon);
3571 /* evict vram memory */
3572 amdgpu_bo_evict_vram(adev);
3574 amdgpu_fence_driver_suspend(adev);
3576 r = amdgpu_device_ip_suspend_phase2(adev);
3578 /* evict remaining vram memory
3579 * This second call to evict vram is to evict the gart page table
3582 amdgpu_bo_evict_vram(adev);
3588 * amdgpu_device_resume - initiate device resume
3590 * @dev: drm dev pointer
3591 * @fbcon : notify the fbdev of resume
3593 * Bring the hw back to operating state (all asics).
3594 * Returns 0 for success or an error on failure.
3595 * Called at driver resume.
3597 int amdgpu_device_resume(struct drm_device *dev, bool fbcon)
3599 struct drm_connector *connector;
3600 struct drm_connector_list_iter iter;
3601 struct amdgpu_device *adev = drm_to_adev(dev);
3602 struct drm_crtc *crtc;
3605 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
3609 if (amdgpu_device_need_post(adev)) {
3610 r = amdgpu_device_asic_init(adev);
3612 dev_err(adev->dev, "amdgpu asic init failed\n");
3615 r = amdgpu_device_ip_resume(adev);
3617 dev_err(adev->dev, "amdgpu_device_ip_resume failed (%d).\n", r);
3620 amdgpu_fence_driver_resume(adev);
3623 r = amdgpu_device_ip_late_init(adev);
3627 queue_delayed_work(system_wq, &adev->delayed_init_work,
3628 msecs_to_jiffies(AMDGPU_RESUME_MS));
3630 if (!amdgpu_device_has_dc_support(adev)) {
3632 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
3633 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
3635 if (amdgpu_crtc->cursor_bo && !adev->enable_virtual_display) {
3636 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
3637 r = amdgpu_bo_reserve(aobj, true);
3639 r = amdgpu_bo_pin(aobj, AMDGPU_GEM_DOMAIN_VRAM);
3641 dev_err(adev->dev, "Failed to pin cursor BO (%d)\n", r);
3642 amdgpu_crtc->cursor_addr = amdgpu_bo_gpu_offset(aobj);
3643 amdgpu_bo_unreserve(aobj);
3648 r = amdgpu_amdkfd_resume(adev, !fbcon);
3652 /* Make sure IB tests flushed */
3653 flush_delayed_work(&adev->delayed_init_work);
3655 /* blat the mode back in */
3657 if (!amdgpu_device_has_dc_support(adev)) {
3659 drm_helper_resume_force_mode(dev);
3661 /* turn on display hw */
3662 drm_modeset_lock_all(dev);
3664 drm_connector_list_iter_begin(dev, &iter);
3665 drm_for_each_connector_iter(connector, &iter)
3666 drm_helper_connector_dpms(connector,
3668 drm_connector_list_iter_end(&iter);
3670 drm_modeset_unlock_all(dev);
3672 amdgpu_fbdev_set_suspend(adev, 0);
3675 drm_kms_helper_poll_enable(dev);
3677 amdgpu_ras_resume(adev);
3680 * Most of the connector probing functions try to acquire runtime pm
3681 * refs to ensure that the GPU is powered on when connector polling is
3682 * performed. Since we're calling this from a runtime PM callback,
3683 * trying to acquire rpm refs will cause us to deadlock.
3685 * Since we're guaranteed to be holding the rpm lock, it's safe to
3686 * temporarily disable the rpm helpers so this doesn't deadlock us.
3689 dev->dev->power.disable_depth++;
3691 if (!amdgpu_device_has_dc_support(adev))
3692 drm_helper_hpd_irq_event(dev);
3694 drm_kms_helper_hotplug_event(dev);
3696 dev->dev->power.disable_depth--;
3698 adev->in_suspend = false;
3704 * amdgpu_device_ip_check_soft_reset - did soft reset succeed
3706 * @adev: amdgpu_device pointer
3708 * The list of all the hardware IPs that make up the asic is walked and
3709 * the check_soft_reset callbacks are run. check_soft_reset determines
3710 * if the asic is still hung or not.
3711 * Returns true if any of the IPs are still in a hung state, false if not.
3713 static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)
3716 bool asic_hang = false;
3718 if (amdgpu_sriov_vf(adev))
3721 if (amdgpu_asic_need_full_reset(adev))
3724 for (i = 0; i < adev->num_ip_blocks; i++) {
3725 if (!adev->ip_blocks[i].status.valid)
3727 if (adev->ip_blocks[i].version->funcs->check_soft_reset)
3728 adev->ip_blocks[i].status.hang =
3729 adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
3730 if (adev->ip_blocks[i].status.hang) {
3731 dev_info(adev->dev, "IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
3739 * amdgpu_device_ip_pre_soft_reset - prepare for soft reset
3741 * @adev: amdgpu_device pointer
3743 * The list of all the hardware IPs that make up the asic is walked and the
3744 * pre_soft_reset callbacks are run if the block is hung. pre_soft_reset
3745 * handles any IP specific hardware or software state changes that are
3746 * necessary for a soft reset to succeed.
3747 * Returns 0 on success, negative error code on failure.
3749 static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)
3753 for (i = 0; i < adev->num_ip_blocks; i++) {
3754 if (!adev->ip_blocks[i].status.valid)
3756 if (adev->ip_blocks[i].status.hang &&
3757 adev->ip_blocks[i].version->funcs->pre_soft_reset) {
3758 r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
3768 * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed
3770 * @adev: amdgpu_device pointer
3772 * Some hardware IPs cannot be soft reset. If they are hung, a full gpu
3773 * reset is necessary to recover.
3774 * Returns true if a full asic reset is required, false if not.
3776 static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)
3780 if (amdgpu_asic_need_full_reset(adev))
3783 for (i = 0; i < adev->num_ip_blocks; i++) {
3784 if (!adev->ip_blocks[i].status.valid)
3786 if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
3787 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
3788 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
3789 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
3790 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
3791 if (adev->ip_blocks[i].status.hang) {
3792 dev_info(adev->dev, "Some block need full reset!\n");
3801 * amdgpu_device_ip_soft_reset - do a soft reset
3803 * @adev: amdgpu_device pointer
3805 * The list of all the hardware IPs that make up the asic is walked and the
3806 * soft_reset callbacks are run if the block is hung. soft_reset handles any
3807 * IP specific hardware or software state changes that are necessary to soft
3809 * Returns 0 on success, negative error code on failure.
3811 static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)
3815 for (i = 0; i < adev->num_ip_blocks; i++) {
3816 if (!adev->ip_blocks[i].status.valid)
3818 if (adev->ip_blocks[i].status.hang &&
3819 adev->ip_blocks[i].version->funcs->soft_reset) {
3820 r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
3830 * amdgpu_device_ip_post_soft_reset - clean up from soft reset
3832 * @adev: amdgpu_device pointer
3834 * The list of all the hardware IPs that make up the asic is walked and the
3835 * post_soft_reset callbacks are run if the asic was hung. post_soft_reset
3836 * handles any IP specific hardware or software state changes that are
3837 * necessary after the IP has been soft reset.
3838 * Returns 0 on success, negative error code on failure.
3840 static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)
3844 for (i = 0; i < adev->num_ip_blocks; i++) {
3845 if (!adev->ip_blocks[i].status.valid)
3847 if (adev->ip_blocks[i].status.hang &&
3848 adev->ip_blocks[i].version->funcs->post_soft_reset)
3849 r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
3858 * amdgpu_device_recover_vram - Recover some VRAM contents
3860 * @adev: amdgpu_device pointer
3862 * Restores the contents of VRAM buffers from the shadows in GTT. Used to
3863 * restore things like GPUVM page tables after a GPU reset where
3864 * the contents of VRAM might be lost.
3867 * 0 on success, negative error code on failure.
3869 static int amdgpu_device_recover_vram(struct amdgpu_device *adev)
3871 struct dma_fence *fence = NULL, *next = NULL;
3872 struct amdgpu_bo *shadow;
3875 if (amdgpu_sriov_runtime(adev))
3876 tmo = msecs_to_jiffies(8000);
3878 tmo = msecs_to_jiffies(100);
3880 dev_info(adev->dev, "recover vram bo from shadow start\n");
3881 mutex_lock(&adev->shadow_list_lock);
3882 list_for_each_entry(shadow, &adev->shadow_list, shadow_list) {
3884 /* No need to recover an evicted BO */
3885 if (shadow->tbo.mem.mem_type != TTM_PL_TT ||
3886 shadow->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET ||
3887 shadow->parent->tbo.mem.mem_type != TTM_PL_VRAM)
3890 r = amdgpu_bo_restore_shadow(shadow, &next);
3895 tmo = dma_fence_wait_timeout(fence, false, tmo);
3896 dma_fence_put(fence);
3901 } else if (tmo < 0) {
3909 mutex_unlock(&adev->shadow_list_lock);
3912 tmo = dma_fence_wait_timeout(fence, false, tmo);
3913 dma_fence_put(fence);
3915 if (r < 0 || tmo <= 0) {
3916 dev_err(adev->dev, "recover vram bo from shadow failed, r is %ld, tmo is %ld\n", r, tmo);
3920 dev_info(adev->dev, "recover vram bo from shadow done\n");
3926 * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
3928 * @adev: amdgpu device pointer
3929 * @from_hypervisor: request from hypervisor
3931 * do VF FLR and reinitialize Asic
3932 * return 0 means succeeded otherwise failed
3934 static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
3935 bool from_hypervisor)
3939 if (from_hypervisor)
3940 r = amdgpu_virt_request_full_gpu(adev, true);
3942 r = amdgpu_virt_reset_gpu(adev);
3946 amdgpu_amdkfd_pre_reset(adev);
3948 /* Resume IP prior to SMC */
3949 r = amdgpu_device_ip_reinit_early_sriov(adev);
3953 amdgpu_virt_init_data_exchange(adev);
3954 /* we need recover gart prior to run SMC/CP/SDMA resume */
3955 amdgpu_gtt_mgr_recover(ttm_manager_type(&adev->mman.bdev, TTM_PL_TT));
3957 r = amdgpu_device_fw_loading(adev);
3961 /* now we are okay to resume SMC/CP/SDMA */
3962 r = amdgpu_device_ip_reinit_late_sriov(adev);
3966 amdgpu_irq_gpu_reset_resume_helper(adev);
3967 r = amdgpu_ib_ring_tests(adev);
3968 amdgpu_amdkfd_post_reset(adev);
3971 amdgpu_virt_release_full_gpu(adev, true);
3972 if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
3973 amdgpu_inc_vram_lost(adev);
3974 r = amdgpu_device_recover_vram(adev);
3981 * amdgpu_device_has_job_running - check if there is any job in mirror list
3983 * @adev: amdgpu device pointer
3985 * check if there is any job in mirror list
3987 bool amdgpu_device_has_job_running(struct amdgpu_device *adev)
3990 struct drm_sched_job *job;
3992 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3993 struct amdgpu_ring *ring = adev->rings[i];
3995 if (!ring || !ring->sched.thread)
3998 spin_lock(&ring->sched.job_list_lock);
3999 job = list_first_entry_or_null(&ring->sched.ring_mirror_list,
4000 struct drm_sched_job, node);
4001 spin_unlock(&ring->sched.job_list_lock);
4009 * amdgpu_device_should_recover_gpu - check if we should try GPU recovery
4011 * @adev: amdgpu device pointer
4013 * Check amdgpu_gpu_recovery and SRIOV status to see if we should try to recover
4016 bool amdgpu_device_should_recover_gpu(struct amdgpu_device *adev)
4018 if (!amdgpu_device_ip_check_soft_reset(adev)) {
4019 dev_info(adev->dev, "Timeout, but no hardware hang detected.\n");
4023 if (amdgpu_gpu_recovery == 0)
4026 if (amdgpu_sriov_vf(adev))
4029 if (amdgpu_gpu_recovery == -1) {
4030 switch (adev->asic_type) {
4036 case CHIP_POLARIS10:
4037 case CHIP_POLARIS11:
4038 case CHIP_POLARIS12:
4049 case CHIP_SIENNA_CICHLID:
4059 dev_info(adev->dev, "GPU recovery disabled.\n");
4064 static int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,
4065 struct amdgpu_job *job,
4066 bool *need_full_reset_arg)
4069 bool need_full_reset = *need_full_reset_arg;
4071 amdgpu_debugfs_wait_dump(adev);
4073 /* block all schedulers and reset given job's ring */
4074 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
4075 struct amdgpu_ring *ring = adev->rings[i];
4077 if (!ring || !ring->sched.thread)
4080 /* after all hw jobs are reset, hw fence is meaningless, so force_completion */
4081 amdgpu_fence_driver_force_completion(ring);
4085 drm_sched_increase_karma(&job->base);
4087 /* Don't suspend on bare metal if we are not going to HW reset the ASIC */
4088 if (!amdgpu_sriov_vf(adev)) {
4090 if (!need_full_reset)
4091 need_full_reset = amdgpu_device_ip_need_full_reset(adev);
4093 if (!need_full_reset) {
4094 amdgpu_device_ip_pre_soft_reset(adev);
4095 r = amdgpu_device_ip_soft_reset(adev);
4096 amdgpu_device_ip_post_soft_reset(adev);
4097 if (r || amdgpu_device_ip_check_soft_reset(adev)) {
4098 dev_info(adev->dev, "soft reset failed, will fallback to full reset!\n");
4099 need_full_reset = true;
4103 if (need_full_reset)
4104 r = amdgpu_device_ip_suspend(adev);
4106 *need_full_reset_arg = need_full_reset;
4112 static int amdgpu_do_asic_reset(struct amdgpu_hive_info *hive,
4113 struct list_head *device_list_handle,
4114 bool *need_full_reset_arg)
4116 struct amdgpu_device *tmp_adev = NULL;
4117 bool need_full_reset = *need_full_reset_arg, vram_lost = false;
4121 * ASIC reset has to be done on all HGMI hive nodes ASAP
4122 * to allow proper links negotiation in FW (within 1 sec)
4124 if (need_full_reset) {
4125 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
4126 /* For XGMI run all resets in parallel to speed up the process */
4127 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
4128 if (!queue_work(system_unbound_wq, &tmp_adev->xgmi_reset_work))
4131 r = amdgpu_asic_reset(tmp_adev);
4134 dev_err(tmp_adev->dev, "ASIC reset failed with error, %d for drm dev, %s",
4135 r, adev_to_drm(tmp_adev)->unique);
4140 /* For XGMI wait for all resets to complete before proceed */
4142 list_for_each_entry(tmp_adev, device_list_handle,
4144 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
4145 flush_work(&tmp_adev->xgmi_reset_work);
4146 r = tmp_adev->asic_reset_res;
4154 if (!r && amdgpu_ras_intr_triggered()) {
4155 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
4156 if (tmp_adev->mmhub.funcs &&
4157 tmp_adev->mmhub.funcs->reset_ras_error_count)
4158 tmp_adev->mmhub.funcs->reset_ras_error_count(tmp_adev);
4161 amdgpu_ras_intr_cleared();
4164 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
4165 if (need_full_reset) {
4167 if (amdgpu_device_asic_init(tmp_adev))
4168 dev_warn(tmp_adev->dev, "asic atom init failed!");
4171 dev_info(tmp_adev->dev, "GPU reset succeeded, trying to resume\n");
4172 r = amdgpu_device_ip_resume_phase1(tmp_adev);
4176 vram_lost = amdgpu_device_check_vram_lost(tmp_adev);
4178 DRM_INFO("VRAM is lost due to GPU reset!\n");
4179 amdgpu_inc_vram_lost(tmp_adev);
4182 r = amdgpu_gtt_mgr_recover(ttm_manager_type(&tmp_adev->mman.bdev, TTM_PL_TT));
4186 r = amdgpu_device_fw_loading(tmp_adev);
4190 r = amdgpu_device_ip_resume_phase2(tmp_adev);
4195 amdgpu_device_fill_reset_magic(tmp_adev);
4198 * Add this ASIC as tracked as reset was already
4199 * complete successfully.
4201 amdgpu_register_gpu_instance(tmp_adev);
4203 r = amdgpu_device_ip_late_init(tmp_adev);
4207 amdgpu_fbdev_set_suspend(tmp_adev, 0);
4210 * The GPU enters bad state once faulty pages
4211 * by ECC has reached the threshold, and ras
4212 * recovery is scheduled next. So add one check
4213 * here to break recovery if it indeed exceeds
4214 * bad page threshold, and remind user to
4215 * retire this GPU or setting one bigger
4216 * bad_page_threshold value to fix this once
4217 * probing driver again.
4219 if (!amdgpu_ras_check_err_threshold(tmp_adev)) {
4221 amdgpu_ras_resume(tmp_adev);
4227 /* Update PSP FW topology after reset */
4228 if (hive && tmp_adev->gmc.xgmi.num_physical_nodes > 1)
4229 r = amdgpu_xgmi_update_topology(hive, tmp_adev);
4235 amdgpu_irq_gpu_reset_resume_helper(tmp_adev);
4236 r = amdgpu_ib_ring_tests(tmp_adev);
4238 dev_err(tmp_adev->dev, "ib ring test failed (%d).\n", r);
4239 r = amdgpu_device_ip_suspend(tmp_adev);
4240 need_full_reset = true;
4247 r = amdgpu_device_recover_vram(tmp_adev);
4249 tmp_adev->asic_reset_res = r;
4253 *need_full_reset_arg = need_full_reset;
4257 static bool amdgpu_device_lock_adev(struct amdgpu_device *adev,
4258 struct amdgpu_hive_info *hive)
4260 if (atomic_cmpxchg(&adev->in_gpu_reset, 0, 1) != 0)
4264 down_write_nest_lock(&adev->reset_sem, &hive->hive_lock);
4266 down_write(&adev->reset_sem);
4269 atomic_inc(&adev->gpu_reset_counter);
4270 switch (amdgpu_asic_reset_method(adev)) {
4271 case AMD_RESET_METHOD_MODE1:
4272 adev->mp1_state = PP_MP1_STATE_SHUTDOWN;
4274 case AMD_RESET_METHOD_MODE2:
4275 adev->mp1_state = PP_MP1_STATE_RESET;
4278 adev->mp1_state = PP_MP1_STATE_NONE;
4285 static void amdgpu_device_unlock_adev(struct amdgpu_device *adev)
4287 amdgpu_vf_error_trans_all(adev);
4288 adev->mp1_state = PP_MP1_STATE_NONE;
4289 atomic_set(&adev->in_gpu_reset, 0);
4290 up_write(&adev->reset_sem);
4293 static void amdgpu_device_resume_display_audio(struct amdgpu_device *adev)
4295 struct pci_dev *p = NULL;
4297 p = pci_get_domain_bus_and_slot(pci_domain_nr(adev->pdev->bus),
4298 adev->pdev->bus->number, 1);
4300 pm_runtime_enable(&(p->dev));
4301 pm_runtime_resume(&(p->dev));
4305 static int amdgpu_device_suspend_display_audio(struct amdgpu_device *adev)
4307 enum amd_reset_method reset_method;
4308 struct pci_dev *p = NULL;
4312 * For now, only BACO and mode1 reset are confirmed
4313 * to suffer the audio issue without proper suspended.
4315 reset_method = amdgpu_asic_reset_method(adev);
4316 if ((reset_method != AMD_RESET_METHOD_BACO) &&
4317 (reset_method != AMD_RESET_METHOD_MODE1))
4320 p = pci_get_domain_bus_and_slot(pci_domain_nr(adev->pdev->bus),
4321 adev->pdev->bus->number, 1);
4325 expires = pm_runtime_autosuspend_expiration(&(p->dev));
4328 * If we cannot get the audio device autosuspend delay,
4329 * a fixed 4S interval will be used. Considering 3S is
4330 * the audio controller default autosuspend delay setting.
4331 * 4S used here is guaranteed to cover that.
4333 expires = ktime_get_mono_fast_ns() + NSEC_PER_SEC * 4ULL;
4335 while (!pm_runtime_status_suspended(&(p->dev))) {
4336 if (!pm_runtime_suspend(&(p->dev)))
4339 if (expires < ktime_get_mono_fast_ns()) {
4340 dev_warn(adev->dev, "failed to suspend display audio\n");
4341 /* TODO: abort the succeeding gpu reset? */
4346 pm_runtime_disable(&(p->dev));
4352 * amdgpu_device_gpu_recover - reset the asic and recover scheduler
4354 * @adev: amdgpu device pointer
4355 * @job: which job trigger hang
4357 * Attempt to reset the GPU if it has hung (all asics).
4358 * Attempt to do soft-reset or full-reset and reinitialize Asic
4359 * Returns 0 for success or an error on failure.
4362 int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
4363 struct amdgpu_job *job)
4365 struct list_head device_list, *device_list_handle = NULL;
4366 bool need_full_reset = false;
4367 bool job_signaled = false;
4368 struct amdgpu_hive_info *hive = NULL;
4369 struct amdgpu_device *tmp_adev = NULL;
4371 bool need_emergency_restart = false;
4372 bool audio_suspended = false;
4375 * Special case: RAS triggered and full reset isn't supported
4377 need_emergency_restart = amdgpu_ras_need_emergency_restart(adev);
4380 * Flush RAM to disk so that after reboot
4381 * the user can read log and see why the system rebooted.
4383 if (need_emergency_restart && amdgpu_ras_get_context(adev)->reboot) {
4384 DRM_WARN("Emergency reboot.");
4387 emergency_restart();
4390 dev_info(adev->dev, "GPU %s begin!\n",
4391 need_emergency_restart ? "jobs stop":"reset");
4394 * Here we trylock to avoid chain of resets executing from
4395 * either trigger by jobs on different adevs in XGMI hive or jobs on
4396 * different schedulers for same device while this TO handler is running.
4397 * We always reset all schedulers for device and all devices for XGMI
4398 * hive so that should take care of them too.
4400 hive = amdgpu_get_xgmi_hive(adev);
4402 if (atomic_cmpxchg(&hive->in_reset, 0, 1) != 0) {
4403 DRM_INFO("Bailing on TDR for s_job:%llx, hive: %llx as another already in progress",
4404 job ? job->base.id : -1, hive->hive_id);
4405 amdgpu_put_xgmi_hive(hive);
4408 mutex_lock(&hive->hive_lock);
4412 * Build list of devices to reset.
4413 * In case we are in XGMI hive mode, resort the device list
4414 * to put adev in the 1st position.
4416 INIT_LIST_HEAD(&device_list);
4417 if (adev->gmc.xgmi.num_physical_nodes > 1) {
4420 if (!list_is_first(&adev->gmc.xgmi.head, &hive->device_list))
4421 list_rotate_to_front(&adev->gmc.xgmi.head, &hive->device_list);
4422 device_list_handle = &hive->device_list;
4424 list_add_tail(&adev->gmc.xgmi.head, &device_list);
4425 device_list_handle = &device_list;
4428 /* block all schedulers and reset given job's ring */
4429 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
4430 if (!amdgpu_device_lock_adev(tmp_adev, hive)) {
4431 dev_info(tmp_adev->dev, "Bailing on TDR for s_job:%llx, as another already in progress",
4432 job ? job->base.id : -1);
4438 * Try to put the audio codec into suspend state
4439 * before gpu reset started.
4441 * Due to the power domain of the graphics device
4442 * is shared with AZ power domain. Without this,
4443 * we may change the audio hardware from behind
4444 * the audio driver's back. That will trigger
4445 * some audio codec errors.
4447 if (!amdgpu_device_suspend_display_audio(tmp_adev))
4448 audio_suspended = true;
4450 amdgpu_ras_set_error_query_ready(tmp_adev, false);
4452 cancel_delayed_work_sync(&tmp_adev->delayed_init_work);
4454 if (!amdgpu_sriov_vf(tmp_adev))
4455 amdgpu_amdkfd_pre_reset(tmp_adev);
4458 * Mark these ASICs to be reseted as untracked first
4459 * And add them back after reset completed
4461 amdgpu_unregister_gpu_instance(tmp_adev);
4463 amdgpu_fbdev_set_suspend(tmp_adev, 1);
4465 /* disable ras on ALL IPs */
4466 if (!need_emergency_restart &&
4467 amdgpu_device_ip_need_full_reset(tmp_adev))
4468 amdgpu_ras_suspend(tmp_adev);
4470 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
4471 struct amdgpu_ring *ring = tmp_adev->rings[i];
4473 if (!ring || !ring->sched.thread)
4476 drm_sched_stop(&ring->sched, job ? &job->base : NULL);
4478 if (need_emergency_restart)
4479 amdgpu_job_stop_all_jobs_on_sched(&ring->sched);
4483 if (need_emergency_restart)
4484 goto skip_sched_resume;
4487 * Must check guilty signal here since after this point all old
4488 * HW fences are force signaled.
4490 * job->base holds a reference to parent fence
4492 if (job && job->base.s_fence->parent &&
4493 dma_fence_is_signaled(job->base.s_fence->parent)) {
4494 job_signaled = true;
4495 dev_info(adev->dev, "Guilty job already signaled, skipping HW reset");
4499 retry: /* Rest of adevs pre asic reset from XGMI hive. */
4500 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
4501 r = amdgpu_device_pre_asic_reset(tmp_adev,
4504 /*TODO Should we stop ?*/
4506 dev_err(tmp_adev->dev, "GPU pre asic reset failed with err, %d for drm dev, %s ",
4507 r, adev_to_drm(tmp_adev)->unique);
4508 tmp_adev->asic_reset_res = r;
4512 /* Actual ASIC resets if needed.*/
4513 /* TODO Implement XGMI hive reset logic for SRIOV */
4514 if (amdgpu_sriov_vf(adev)) {
4515 r = amdgpu_device_reset_sriov(adev, job ? false : true);
4517 adev->asic_reset_res = r;
4519 r = amdgpu_do_asic_reset(hive, device_list_handle, &need_full_reset);
4520 if (r && r == -EAGAIN)
4526 /* Post ASIC reset for all devs .*/
4527 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
4529 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
4530 struct amdgpu_ring *ring = tmp_adev->rings[i];
4532 if (!ring || !ring->sched.thread)
4535 /* No point to resubmit jobs if we didn't HW reset*/
4536 if (!tmp_adev->asic_reset_res && !job_signaled)
4537 drm_sched_resubmit_jobs(&ring->sched);
4539 drm_sched_start(&ring->sched, !tmp_adev->asic_reset_res);
4542 if (!amdgpu_device_has_dc_support(tmp_adev) && !job_signaled) {
4543 drm_helper_resume_force_mode(adev_to_drm(tmp_adev));
4546 tmp_adev->asic_reset_res = 0;
4549 /* bad news, how to tell it to userspace ? */
4550 dev_info(tmp_adev->dev, "GPU reset(%d) failed\n", atomic_read(&tmp_adev->gpu_reset_counter));
4551 amdgpu_vf_error_put(tmp_adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r);
4553 dev_info(tmp_adev->dev, "GPU reset(%d) succeeded!\n", atomic_read(&tmp_adev->gpu_reset_counter));
4558 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
4559 /*unlock kfd: SRIOV would do it separately */
4560 if (!need_emergency_restart && !amdgpu_sriov_vf(tmp_adev))
4561 amdgpu_amdkfd_post_reset(tmp_adev);
4562 if (audio_suspended)
4563 amdgpu_device_resume_display_audio(tmp_adev);
4564 amdgpu_device_unlock_adev(tmp_adev);
4569 atomic_set(&hive->in_reset, 0);
4570 mutex_unlock(&hive->hive_lock);
4571 amdgpu_put_xgmi_hive(hive);
4575 dev_info(adev->dev, "GPU reset end with ret = %d\n", r);
4580 * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot
4582 * @adev: amdgpu_device pointer
4584 * Fetchs and stores in the driver the PCIE capabilities (gen speed
4585 * and lanes) of the slot the device is in. Handles APUs and
4586 * virtualized environments where PCIE config space may not be available.
4588 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev)
4590 struct pci_dev *pdev;
4591 enum pci_bus_speed speed_cap, platform_speed_cap;
4592 enum pcie_link_width platform_link_width;
4594 if (amdgpu_pcie_gen_cap)
4595 adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
4597 if (amdgpu_pcie_lane_cap)
4598 adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
4600 /* covers APUs as well */
4601 if (pci_is_root_bus(adev->pdev->bus)) {
4602 if (adev->pm.pcie_gen_mask == 0)
4603 adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
4604 if (adev->pm.pcie_mlw_mask == 0)
4605 adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
4609 if (adev->pm.pcie_gen_mask && adev->pm.pcie_mlw_mask)
4612 pcie_bandwidth_available(adev->pdev, NULL,
4613 &platform_speed_cap, &platform_link_width);
4615 if (adev->pm.pcie_gen_mask == 0) {
4618 speed_cap = pcie_get_speed_cap(pdev);
4619 if (speed_cap == PCI_SPEED_UNKNOWN) {
4620 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4621 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4622 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
4624 if (speed_cap == PCIE_SPEED_16_0GT)
4625 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4626 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4627 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
4628 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4);
4629 else if (speed_cap == PCIE_SPEED_8_0GT)
4630 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4631 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4632 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
4633 else if (speed_cap == PCIE_SPEED_5_0GT)
4634 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4635 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2);
4637 adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1;
4640 if (platform_speed_cap == PCI_SPEED_UNKNOWN) {
4641 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4642 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
4644 if (platform_speed_cap == PCIE_SPEED_16_0GT)
4645 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4646 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4647 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
4648 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4);
4649 else if (platform_speed_cap == PCIE_SPEED_8_0GT)
4650 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4651 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4652 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3);
4653 else if (platform_speed_cap == PCIE_SPEED_5_0GT)
4654 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4655 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
4657 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
4661 if (adev->pm.pcie_mlw_mask == 0) {
4662 if (platform_link_width == PCIE_LNK_WIDTH_UNKNOWN) {
4663 adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK;
4665 switch (platform_link_width) {
4667 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
4668 CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
4669 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
4670 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4671 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4672 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4673 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4676 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
4677 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
4678 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4679 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4680 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4681 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4684 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
4685 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4686 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4687 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4688 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4691 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4692 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4693 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4694 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4697 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4698 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4699 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4702 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4703 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4706 adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;
4715 int amdgpu_device_baco_enter(struct drm_device *dev)
4717 struct amdgpu_device *adev = drm_to_adev(dev);
4718 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
4720 if (!amdgpu_device_supports_baco(adev_to_drm(adev)))
4723 if (ras && ras->supported)
4724 adev->nbio.funcs->enable_doorbell_interrupt(adev, false);
4726 return amdgpu_dpm_baco_enter(adev);
4729 int amdgpu_device_baco_exit(struct drm_device *dev)
4731 struct amdgpu_device *adev = drm_to_adev(dev);
4732 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
4735 if (!amdgpu_device_supports_baco(adev_to_drm(adev)))
4738 ret = amdgpu_dpm_baco_exit(adev);
4742 if (ras && ras->supported)
4743 adev->nbio.funcs->enable_doorbell_interrupt(adev, true);
4748 static void amdgpu_cancel_all_tdr(struct amdgpu_device *adev)
4752 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
4753 struct amdgpu_ring *ring = adev->rings[i];
4755 if (!ring || !ring->sched.thread)
4758 cancel_delayed_work_sync(&ring->sched.work_tdr);
4763 * amdgpu_pci_error_detected - Called when a PCI error is detected.
4764 * @pdev: PCI device struct
4765 * @state: PCI channel state
4767 * Description: Called when a PCI error is detected.
4769 * Return: PCI_ERS_RESULT_NEED_RESET or PCI_ERS_RESULT_DISCONNECT.
4771 pci_ers_result_t amdgpu_pci_error_detected(struct pci_dev *pdev, pci_channel_state_t state)
4773 struct drm_device *dev = pci_get_drvdata(pdev);
4774 struct amdgpu_device *adev = drm_to_adev(dev);
4777 DRM_INFO("PCI error: detected callback, state(%d)!!\n", state);
4780 case pci_channel_io_normal:
4781 return PCI_ERS_RESULT_CAN_RECOVER;
4782 /* Fatal error, prepare for slot reset */
4783 case pci_channel_io_frozen:
4785 * Cancel and wait for all TDRs in progress if failing to
4786 * set adev->in_gpu_reset in amdgpu_device_lock_adev
4788 * Locking adev->reset_sem will prevent any external access
4789 * to GPU during PCI error recovery
4791 while (!amdgpu_device_lock_adev(adev, NULL))
4792 amdgpu_cancel_all_tdr(adev);
4795 * Block any work scheduling as we do for regular GPU reset
4796 * for the duration of the recovery
4798 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
4799 struct amdgpu_ring *ring = adev->rings[i];
4801 if (!ring || !ring->sched.thread)
4804 drm_sched_stop(&ring->sched, NULL);
4806 return PCI_ERS_RESULT_NEED_RESET;
4807 case pci_channel_io_perm_failure:
4808 /* Permanent error, prepare for device removal */
4809 return PCI_ERS_RESULT_DISCONNECT;
4812 return PCI_ERS_RESULT_NEED_RESET;
4816 * amdgpu_pci_mmio_enabled - Enable MMIO and dump debug registers
4817 * @pdev: pointer to PCI device
4819 pci_ers_result_t amdgpu_pci_mmio_enabled(struct pci_dev *pdev)
4822 DRM_INFO("PCI error: mmio enabled callback!!\n");
4824 /* TODO - dump whatever for debugging purposes */
4826 /* This called only if amdgpu_pci_error_detected returns
4827 * PCI_ERS_RESULT_CAN_RECOVER. Read/write to the device still
4828 * works, no need to reset slot.
4831 return PCI_ERS_RESULT_RECOVERED;
4835 * amdgpu_pci_slot_reset - Called when PCI slot has been reset.
4836 * @pdev: PCI device struct
4838 * Description: This routine is called by the pci error recovery
4839 * code after the PCI slot has been reset, just before we
4840 * should resume normal operations.
4842 pci_ers_result_t amdgpu_pci_slot_reset(struct pci_dev *pdev)
4844 struct drm_device *dev = pci_get_drvdata(pdev);
4845 struct amdgpu_device *adev = drm_to_adev(dev);
4850 DRM_INFO("PCI error: slot reset callback!!\n");
4852 /* wait for asic to come out of reset */
4855 pci_restore_state(pdev);
4857 /* confirm ASIC came out of reset */
4858 for (i = 0; i < adev->usec_timeout; i++) {
4859 memsize = amdgpu_asic_get_config_memsize(adev);
4861 if (memsize != 0xffffffff)
4865 if (memsize == 0xffffffff) {
4870 /* TODO Call amdgpu_pre_asic_reset instead */
4871 adev->in_pci_err_recovery = true;
4872 r = amdgpu_device_ip_suspend(adev);
4873 adev->in_pci_err_recovery = false;
4879 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
4883 r = amdgpu_device_ip_resume_phase1(adev);
4887 vram_lost = amdgpu_device_check_vram_lost(adev);
4889 DRM_INFO("VRAM is lost due to GPU reset!\n");
4890 amdgpu_inc_vram_lost(adev);
4893 r = amdgpu_gtt_mgr_recover(
4894 &adev->mman.bdev.man[TTM_PL_TT]);
4898 r = amdgpu_device_fw_loading(adev);
4902 r = amdgpu_device_ip_resume_phase2(adev);
4907 amdgpu_device_fill_reset_magic(adev);
4910 * Add this ASIC as tracked as reset was already
4911 * complete successfully.
4913 amdgpu_register_gpu_instance(adev);
4915 r = amdgpu_device_ip_late_init(adev);
4919 amdgpu_fbdev_set_suspend(adev, 0);
4922 amdgpu_ras_resume(adev);
4925 amdgpu_irq_gpu_reset_resume_helper(adev);
4926 r = amdgpu_ib_ring_tests(adev);
4930 r = amdgpu_device_recover_vram(adev);
4935 DRM_INFO("PCIe error recovery succeeded\n");
4937 DRM_ERROR("PCIe error recovery failed, err:%d", r);
4938 amdgpu_device_unlock_adev(adev);
4941 return r ? PCI_ERS_RESULT_DISCONNECT : PCI_ERS_RESULT_RECOVERED;
4945 * amdgpu_pci_resume() - resume normal ops after PCI reset
4946 * @pdev: pointer to PCI device
4948 * Called when the error recovery driver tells us that its
4949 * OK to resume normal operation. Use completion to allow
4950 * halted scsi ops to resume.
4952 void amdgpu_pci_resume(struct pci_dev *pdev)
4954 struct drm_device *dev = pci_get_drvdata(pdev);
4955 struct amdgpu_device *adev = drm_to_adev(dev);
4959 DRM_INFO("PCI error: resume callback!!\n");
4961 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
4962 struct amdgpu_ring *ring = adev->rings[i];
4964 if (!ring || !ring->sched.thread)
4968 drm_sched_resubmit_jobs(&ring->sched);
4969 drm_sched_start(&ring->sched, true);
4972 amdgpu_device_unlock_adev(adev);