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>
33 #include <linux/iommu.h>
34 #include <linux/pci.h>
36 #include <drm/drm_atomic_helper.h>
37 #include <drm/drm_probe_helper.h>
38 #include <drm/amdgpu_drm.h>
39 #include <linux/vgaarb.h>
40 #include <linux/vga_switcheroo.h>
41 #include <linux/efi.h>
43 #include "amdgpu_trace.h"
44 #include "amdgpu_i2c.h"
46 #include "amdgpu_atombios.h"
47 #include "amdgpu_atomfirmware.h"
49 #ifdef CONFIG_DRM_AMDGPU_SI
52 #ifdef CONFIG_DRM_AMDGPU_CIK
58 #include "bif/bif_4_1_d.h"
59 #include <linux/firmware.h>
60 #include "amdgpu_vf_error.h"
62 #include "amdgpu_amdkfd.h"
63 #include "amdgpu_pm.h"
65 #include "amdgpu_xgmi.h"
66 #include "amdgpu_ras.h"
67 #include "amdgpu_pmu.h"
68 #include "amdgpu_fru_eeprom.h"
69 #include "amdgpu_reset.h"
71 #include <linux/suspend.h>
72 #include <drm/task_barrier.h>
73 #include <linux/pm_runtime.h>
75 #include <drm/drm_drv.h>
77 MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin");
78 MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin");
79 MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin");
80 MODULE_FIRMWARE("amdgpu/picasso_gpu_info.bin");
81 MODULE_FIRMWARE("amdgpu/raven2_gpu_info.bin");
82 MODULE_FIRMWARE("amdgpu/arcturus_gpu_info.bin");
83 MODULE_FIRMWARE("amdgpu/renoir_gpu_info.bin");
84 MODULE_FIRMWARE("amdgpu/navi10_gpu_info.bin");
85 MODULE_FIRMWARE("amdgpu/navi14_gpu_info.bin");
86 MODULE_FIRMWARE("amdgpu/navi12_gpu_info.bin");
87 MODULE_FIRMWARE("amdgpu/vangogh_gpu_info.bin");
88 MODULE_FIRMWARE("amdgpu/yellow_carp_gpu_info.bin");
90 #define AMDGPU_RESUME_MS 2000
91 #define AMDGPU_MAX_RETRY_LIMIT 2
92 #define AMDGPU_RETRY_SRIOV_RESET(r) ((r) == -EBUSY || (r) == -ETIMEDOUT || (r) == -EINVAL)
94 const char *amdgpu_asic_name[] = {
136 * DOC: pcie_replay_count
138 * The amdgpu driver provides a sysfs API for reporting the total number
139 * of PCIe replays (NAKs)
140 * The file pcie_replay_count is used for this and returns the total
141 * number of replays as a sum of the NAKs generated and NAKs received
144 static ssize_t amdgpu_device_get_pcie_replay_count(struct device *dev,
145 struct device_attribute *attr, char *buf)
147 struct drm_device *ddev = dev_get_drvdata(dev);
148 struct amdgpu_device *adev = drm_to_adev(ddev);
149 uint64_t cnt = amdgpu_asic_get_pcie_replay_count(adev);
151 return sysfs_emit(buf, "%llu\n", cnt);
154 static DEVICE_ATTR(pcie_replay_count, S_IRUGO,
155 amdgpu_device_get_pcie_replay_count, NULL);
157 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev);
162 * The amdgpu driver provides a sysfs API for reporting the product name
164 * The file serial_number is used for this and returns the product name
165 * as returned from the FRU.
166 * NOTE: This is only available for certain server cards
169 static ssize_t amdgpu_device_get_product_name(struct device *dev,
170 struct device_attribute *attr, char *buf)
172 struct drm_device *ddev = dev_get_drvdata(dev);
173 struct amdgpu_device *adev = drm_to_adev(ddev);
175 return sysfs_emit(buf, "%s\n", adev->product_name);
178 static DEVICE_ATTR(product_name, S_IRUGO,
179 amdgpu_device_get_product_name, NULL);
182 * DOC: product_number
184 * The amdgpu driver provides a sysfs API for reporting the part number
186 * The file serial_number is used for this and returns the part number
187 * as returned from the FRU.
188 * NOTE: This is only available for certain server cards
191 static ssize_t amdgpu_device_get_product_number(struct device *dev,
192 struct device_attribute *attr, char *buf)
194 struct drm_device *ddev = dev_get_drvdata(dev);
195 struct amdgpu_device *adev = drm_to_adev(ddev);
197 return sysfs_emit(buf, "%s\n", adev->product_number);
200 static DEVICE_ATTR(product_number, S_IRUGO,
201 amdgpu_device_get_product_number, NULL);
206 * The amdgpu driver provides a sysfs API for reporting the serial number
208 * The file serial_number is used for this and returns the serial number
209 * as returned from the FRU.
210 * NOTE: This is only available for certain server cards
213 static ssize_t amdgpu_device_get_serial_number(struct device *dev,
214 struct device_attribute *attr, char *buf)
216 struct drm_device *ddev = dev_get_drvdata(dev);
217 struct amdgpu_device *adev = drm_to_adev(ddev);
219 return sysfs_emit(buf, "%s\n", adev->serial);
222 static DEVICE_ATTR(serial_number, S_IRUGO,
223 amdgpu_device_get_serial_number, NULL);
226 * amdgpu_device_supports_px - Is the device a dGPU with ATPX power control
228 * @dev: drm_device pointer
230 * Returns true if the device is a dGPU with ATPX power control,
231 * otherwise return false.
233 bool amdgpu_device_supports_px(struct drm_device *dev)
235 struct amdgpu_device *adev = drm_to_adev(dev);
237 if ((adev->flags & AMD_IS_PX) && !amdgpu_is_atpx_hybrid())
243 * amdgpu_device_supports_boco - Is the device a dGPU with ACPI power resources
245 * @dev: drm_device pointer
247 * Returns true if the device is a dGPU with ACPI power control,
248 * otherwise return false.
250 bool amdgpu_device_supports_boco(struct drm_device *dev)
252 struct amdgpu_device *adev = drm_to_adev(dev);
255 ((adev->flags & AMD_IS_PX) && amdgpu_is_atpx_hybrid()))
261 * amdgpu_device_supports_baco - Does the device support BACO
263 * @dev: drm_device pointer
265 * Returns true if the device supporte BACO,
266 * otherwise return false.
268 bool amdgpu_device_supports_baco(struct drm_device *dev)
270 struct amdgpu_device *adev = drm_to_adev(dev);
272 return amdgpu_asic_supports_baco(adev);
276 * amdgpu_device_supports_smart_shift - Is the device dGPU with
277 * smart shift support
279 * @dev: drm_device pointer
281 * Returns true if the device is a dGPU with Smart Shift support,
282 * otherwise returns false.
284 bool amdgpu_device_supports_smart_shift(struct drm_device *dev)
286 return (amdgpu_device_supports_boco(dev) &&
287 amdgpu_acpi_is_power_shift_control_supported());
291 * VRAM access helper functions
295 * amdgpu_device_mm_access - access vram by MM_INDEX/MM_DATA
297 * @adev: amdgpu_device pointer
298 * @pos: offset of the buffer in vram
299 * @buf: virtual address of the buffer in system memory
300 * @size: read/write size, sizeof(@buf) must > @size
301 * @write: true - write to vram, otherwise - read from vram
303 void amdgpu_device_mm_access(struct amdgpu_device *adev, loff_t pos,
304 void *buf, size_t size, bool write)
307 uint32_t hi = ~0, tmp = 0;
308 uint32_t *data = buf;
312 if (!drm_dev_enter(adev_to_drm(adev), &idx))
315 BUG_ON(!IS_ALIGNED(pos, 4) || !IS_ALIGNED(size, 4));
317 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
318 for (last = pos + size; pos < last; pos += 4) {
321 WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)pos) | 0x80000000);
323 WREG32_NO_KIQ(mmMM_INDEX_HI, tmp);
327 WREG32_NO_KIQ(mmMM_DATA, *data++);
329 *data++ = RREG32_NO_KIQ(mmMM_DATA);
332 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
337 * amdgpu_device_aper_access - access vram by vram aperature
339 * @adev: amdgpu_device pointer
340 * @pos: offset of the buffer in vram
341 * @buf: virtual address of the buffer in system memory
342 * @size: read/write size, sizeof(@buf) must > @size
343 * @write: true - write to vram, otherwise - read from vram
345 * The return value means how many bytes have been transferred.
347 size_t amdgpu_device_aper_access(struct amdgpu_device *adev, loff_t pos,
348 void *buf, size_t size, bool write)
355 if (!adev->mman.aper_base_kaddr)
358 last = min(pos + size, adev->gmc.visible_vram_size);
360 addr = adev->mman.aper_base_kaddr + pos;
364 memcpy_toio(addr, buf, count);
366 amdgpu_device_flush_hdp(adev, NULL);
368 amdgpu_device_invalidate_hdp(adev, NULL);
370 memcpy_fromio(buf, addr, count);
382 * amdgpu_device_vram_access - read/write a buffer in vram
384 * @adev: amdgpu_device pointer
385 * @pos: offset of the buffer in vram
386 * @buf: virtual address of the buffer in system memory
387 * @size: read/write size, sizeof(@buf) must > @size
388 * @write: true - write to vram, otherwise - read from vram
390 void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos,
391 void *buf, size_t size, bool write)
395 /* try to using vram apreature to access vram first */
396 count = amdgpu_device_aper_access(adev, pos, buf, size, write);
399 /* using MM to access rest vram */
402 amdgpu_device_mm_access(adev, pos, buf, size, write);
407 * register access helper functions.
410 /* Check if hw access should be skipped because of hotplug or device error */
411 bool amdgpu_device_skip_hw_access(struct amdgpu_device *adev)
413 if (adev->no_hw_access)
416 #ifdef CONFIG_LOCKDEP
418 * This is a bit complicated to understand, so worth a comment. What we assert
419 * here is that the GPU reset is not running on another thread in parallel.
421 * For this we trylock the read side of the reset semaphore, if that succeeds
422 * we know that the reset is not running in paralell.
424 * If the trylock fails we assert that we are either already holding the read
425 * side of the lock or are the reset thread itself and hold the write side of
429 if (down_read_trylock(&adev->reset_sem))
430 up_read(&adev->reset_sem);
432 lockdep_assert_held(&adev->reset_sem);
439 * amdgpu_device_rreg - read a memory mapped IO or indirect register
441 * @adev: amdgpu_device pointer
442 * @reg: dword aligned register offset
443 * @acc_flags: access flags which require special behavior
445 * Returns the 32 bit value from the offset specified.
447 uint32_t amdgpu_device_rreg(struct amdgpu_device *adev,
448 uint32_t reg, uint32_t acc_flags)
452 if (amdgpu_device_skip_hw_access(adev))
455 if ((reg * 4) < adev->rmmio_size) {
456 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) &&
457 amdgpu_sriov_runtime(adev) &&
458 down_read_trylock(&adev->reset_sem)) {
459 ret = amdgpu_kiq_rreg(adev, reg);
460 up_read(&adev->reset_sem);
462 ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
465 ret = adev->pcie_rreg(adev, reg * 4);
468 trace_amdgpu_device_rreg(adev->pdev->device, reg, ret);
474 * MMIO register read with bytes helper functions
475 * @offset:bytes offset from MMIO start
480 * amdgpu_mm_rreg8 - read a memory mapped IO register
482 * @adev: amdgpu_device pointer
483 * @offset: byte aligned register offset
485 * Returns the 8 bit value from the offset specified.
487 uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset)
489 if (amdgpu_device_skip_hw_access(adev))
492 if (offset < adev->rmmio_size)
493 return (readb(adev->rmmio + offset));
498 * MMIO register write with bytes helper functions
499 * @offset:bytes offset from MMIO start
500 * @value: the value want to be written to the register
504 * amdgpu_mm_wreg8 - read a memory mapped IO register
506 * @adev: amdgpu_device pointer
507 * @offset: byte aligned register offset
508 * @value: 8 bit value to write
510 * Writes the value specified to the offset specified.
512 void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value)
514 if (amdgpu_device_skip_hw_access(adev))
517 if (offset < adev->rmmio_size)
518 writeb(value, adev->rmmio + offset);
524 * amdgpu_device_wreg - write to a memory mapped IO or indirect register
526 * @adev: amdgpu_device pointer
527 * @reg: dword aligned register offset
528 * @v: 32 bit value to write to the register
529 * @acc_flags: access flags which require special behavior
531 * Writes the value specified to the offset specified.
533 void amdgpu_device_wreg(struct amdgpu_device *adev,
534 uint32_t reg, uint32_t v,
537 if (amdgpu_device_skip_hw_access(adev))
540 if ((reg * 4) < adev->rmmio_size) {
541 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) &&
542 amdgpu_sriov_runtime(adev) &&
543 down_read_trylock(&adev->reset_sem)) {
544 amdgpu_kiq_wreg(adev, reg, v);
545 up_read(&adev->reset_sem);
547 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
550 adev->pcie_wreg(adev, reg * 4, v);
553 trace_amdgpu_device_wreg(adev->pdev->device, reg, v);
557 * amdgpu_mm_wreg_mmio_rlc - write register either with direct/indirect mmio or with RLC path if in range
559 * @adev: amdgpu_device pointer
560 * @reg: mmio/rlc register
563 * this function is invoked only for the debugfs register access
565 void amdgpu_mm_wreg_mmio_rlc(struct amdgpu_device *adev,
566 uint32_t reg, uint32_t v)
568 if (amdgpu_device_skip_hw_access(adev))
571 if (amdgpu_sriov_fullaccess(adev) &&
572 adev->gfx.rlc.funcs &&
573 adev->gfx.rlc.funcs->is_rlcg_access_range) {
574 if (adev->gfx.rlc.funcs->is_rlcg_access_range(adev, reg))
575 return amdgpu_sriov_wreg(adev, reg, v, 0, 0);
576 } else if ((reg * 4) >= adev->rmmio_size) {
577 adev->pcie_wreg(adev, reg * 4, v);
579 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
584 * amdgpu_mm_rdoorbell - read a doorbell dword
586 * @adev: amdgpu_device pointer
587 * @index: doorbell index
589 * Returns the value in the doorbell aperture at the
590 * requested doorbell index (CIK).
592 u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
594 if (amdgpu_device_skip_hw_access(adev))
597 if (index < adev->doorbell.num_doorbells) {
598 return readl(adev->doorbell.ptr + index);
600 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
606 * amdgpu_mm_wdoorbell - write a doorbell dword
608 * @adev: amdgpu_device pointer
609 * @index: doorbell index
612 * Writes @v to the doorbell aperture at the
613 * requested doorbell index (CIK).
615 void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v)
617 if (amdgpu_device_skip_hw_access(adev))
620 if (index < adev->doorbell.num_doorbells) {
621 writel(v, adev->doorbell.ptr + index);
623 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
628 * amdgpu_mm_rdoorbell64 - read a doorbell Qword
630 * @adev: amdgpu_device pointer
631 * @index: doorbell index
633 * Returns the value in the doorbell aperture at the
634 * requested doorbell index (VEGA10+).
636 u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index)
638 if (amdgpu_device_skip_hw_access(adev))
641 if (index < adev->doorbell.num_doorbells) {
642 return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index));
644 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
650 * amdgpu_mm_wdoorbell64 - write a doorbell Qword
652 * @adev: amdgpu_device pointer
653 * @index: doorbell index
656 * Writes @v to the doorbell aperture at the
657 * requested doorbell index (VEGA10+).
659 void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v)
661 if (amdgpu_device_skip_hw_access(adev))
664 if (index < adev->doorbell.num_doorbells) {
665 atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v);
667 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
672 * amdgpu_device_indirect_rreg - read an indirect register
674 * @adev: amdgpu_device pointer
675 * @pcie_index: mmio register offset
676 * @pcie_data: mmio register offset
677 * @reg_addr: indirect register address to read from
679 * Returns the value of indirect register @reg_addr
681 u32 amdgpu_device_indirect_rreg(struct amdgpu_device *adev,
682 u32 pcie_index, u32 pcie_data,
687 void __iomem *pcie_index_offset;
688 void __iomem *pcie_data_offset;
690 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
691 pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
692 pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
694 writel(reg_addr, pcie_index_offset);
695 readl(pcie_index_offset);
696 r = readl(pcie_data_offset);
697 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
703 * amdgpu_device_indirect_rreg64 - read a 64bits indirect register
705 * @adev: amdgpu_device pointer
706 * @pcie_index: mmio register offset
707 * @pcie_data: mmio register offset
708 * @reg_addr: indirect register address to read from
710 * Returns the value of indirect register @reg_addr
712 u64 amdgpu_device_indirect_rreg64(struct amdgpu_device *adev,
713 u32 pcie_index, u32 pcie_data,
718 void __iomem *pcie_index_offset;
719 void __iomem *pcie_data_offset;
721 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
722 pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
723 pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
725 /* read low 32 bits */
726 writel(reg_addr, pcie_index_offset);
727 readl(pcie_index_offset);
728 r = readl(pcie_data_offset);
729 /* read high 32 bits */
730 writel(reg_addr + 4, pcie_index_offset);
731 readl(pcie_index_offset);
732 r |= ((u64)readl(pcie_data_offset) << 32);
733 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
739 * amdgpu_device_indirect_wreg - write an indirect register address
741 * @adev: amdgpu_device pointer
742 * @pcie_index: mmio register offset
743 * @pcie_data: mmio register offset
744 * @reg_addr: indirect register offset
745 * @reg_data: indirect register data
748 void amdgpu_device_indirect_wreg(struct amdgpu_device *adev,
749 u32 pcie_index, u32 pcie_data,
750 u32 reg_addr, u32 reg_data)
753 void __iomem *pcie_index_offset;
754 void __iomem *pcie_data_offset;
756 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
757 pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
758 pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
760 writel(reg_addr, pcie_index_offset);
761 readl(pcie_index_offset);
762 writel(reg_data, pcie_data_offset);
763 readl(pcie_data_offset);
764 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
768 * amdgpu_device_indirect_wreg64 - write a 64bits indirect register address
770 * @adev: amdgpu_device pointer
771 * @pcie_index: mmio register offset
772 * @pcie_data: mmio register offset
773 * @reg_addr: indirect register offset
774 * @reg_data: indirect register data
777 void amdgpu_device_indirect_wreg64(struct amdgpu_device *adev,
778 u32 pcie_index, u32 pcie_data,
779 u32 reg_addr, u64 reg_data)
782 void __iomem *pcie_index_offset;
783 void __iomem *pcie_data_offset;
785 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
786 pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
787 pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
789 /* write low 32 bits */
790 writel(reg_addr, pcie_index_offset);
791 readl(pcie_index_offset);
792 writel((u32)(reg_data & 0xffffffffULL), pcie_data_offset);
793 readl(pcie_data_offset);
794 /* write high 32 bits */
795 writel(reg_addr + 4, pcie_index_offset);
796 readl(pcie_index_offset);
797 writel((u32)(reg_data >> 32), pcie_data_offset);
798 readl(pcie_data_offset);
799 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
803 * amdgpu_invalid_rreg - dummy reg read function
805 * @adev: amdgpu_device pointer
806 * @reg: offset of register
808 * Dummy register read function. Used for register blocks
809 * that certain asics don't have (all asics).
810 * Returns the value in the register.
812 static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
814 DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
820 * amdgpu_invalid_wreg - dummy reg write function
822 * @adev: amdgpu_device pointer
823 * @reg: offset of register
824 * @v: value to write to the register
826 * Dummy register read function. Used for register blocks
827 * that certain asics don't have (all asics).
829 static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
831 DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
837 * amdgpu_invalid_rreg64 - dummy 64 bit reg read function
839 * @adev: amdgpu_device pointer
840 * @reg: offset of register
842 * Dummy register read function. Used for register blocks
843 * that certain asics don't have (all asics).
844 * Returns the value in the register.
846 static uint64_t amdgpu_invalid_rreg64(struct amdgpu_device *adev, uint32_t reg)
848 DRM_ERROR("Invalid callback to read 64 bit register 0x%04X\n", reg);
854 * amdgpu_invalid_wreg64 - dummy reg write function
856 * @adev: amdgpu_device pointer
857 * @reg: offset of register
858 * @v: value to write to the register
860 * Dummy register read function. Used for register blocks
861 * that certain asics don't have (all asics).
863 static void amdgpu_invalid_wreg64(struct amdgpu_device *adev, uint32_t reg, uint64_t v)
865 DRM_ERROR("Invalid callback to write 64 bit register 0x%04X with 0x%08llX\n",
871 * amdgpu_block_invalid_rreg - dummy reg read function
873 * @adev: amdgpu_device pointer
874 * @block: offset of instance
875 * @reg: offset of register
877 * Dummy register read function. Used for register blocks
878 * that certain asics don't have (all asics).
879 * Returns the value in the register.
881 static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
882 uint32_t block, uint32_t reg)
884 DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
891 * amdgpu_block_invalid_wreg - dummy reg write function
893 * @adev: amdgpu_device pointer
894 * @block: offset of instance
895 * @reg: offset of register
896 * @v: value to write to the register
898 * Dummy register read function. Used for register blocks
899 * that certain asics don't have (all asics).
901 static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
903 uint32_t reg, uint32_t v)
905 DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
911 * amdgpu_device_asic_init - Wrapper for atom asic_init
913 * @adev: amdgpu_device pointer
915 * Does any asic specific work and then calls atom asic init.
917 static int amdgpu_device_asic_init(struct amdgpu_device *adev)
919 amdgpu_asic_pre_asic_init(adev);
921 return amdgpu_atom_asic_init(adev->mode_info.atom_context);
925 * amdgpu_device_vram_scratch_init - allocate the VRAM scratch page
927 * @adev: amdgpu_device pointer
929 * Allocates a scratch page of VRAM for use by various things in the
932 static int amdgpu_device_vram_scratch_init(struct amdgpu_device *adev)
934 return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE,
935 PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
936 &adev->vram_scratch.robj,
937 &adev->vram_scratch.gpu_addr,
938 (void **)&adev->vram_scratch.ptr);
942 * amdgpu_device_vram_scratch_fini - Free the VRAM scratch page
944 * @adev: amdgpu_device pointer
946 * Frees the VRAM scratch page.
948 static void amdgpu_device_vram_scratch_fini(struct amdgpu_device *adev)
950 amdgpu_bo_free_kernel(&adev->vram_scratch.robj, NULL, NULL);
954 * amdgpu_device_program_register_sequence - program an array of registers.
956 * @adev: amdgpu_device pointer
957 * @registers: pointer to the register array
958 * @array_size: size of the register array
960 * Programs an array or registers with and and or masks.
961 * This is a helper for setting golden registers.
963 void amdgpu_device_program_register_sequence(struct amdgpu_device *adev,
964 const u32 *registers,
965 const u32 array_size)
967 u32 tmp, reg, and_mask, or_mask;
973 for (i = 0; i < array_size; i +=3) {
974 reg = registers[i + 0];
975 and_mask = registers[i + 1];
976 or_mask = registers[i + 2];
978 if (and_mask == 0xffffffff) {
983 if (adev->family >= AMDGPU_FAMILY_AI)
984 tmp |= (or_mask & and_mask);
993 * amdgpu_device_pci_config_reset - reset the GPU
995 * @adev: amdgpu_device pointer
997 * Resets the GPU using the pci config reset sequence.
998 * Only applicable to asics prior to vega10.
1000 void amdgpu_device_pci_config_reset(struct amdgpu_device *adev)
1002 pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
1006 * amdgpu_device_pci_reset - reset the GPU using generic PCI means
1008 * @adev: amdgpu_device pointer
1010 * Resets the GPU using generic pci reset interfaces (FLR, SBR, etc.).
1012 int amdgpu_device_pci_reset(struct amdgpu_device *adev)
1014 return pci_reset_function(adev->pdev);
1018 * GPU doorbell aperture helpers function.
1021 * amdgpu_device_doorbell_init - Init doorbell driver information.
1023 * @adev: amdgpu_device pointer
1025 * Init doorbell driver information (CIK)
1026 * Returns 0 on success, error on failure.
1028 static int amdgpu_device_doorbell_init(struct amdgpu_device *adev)
1031 /* No doorbell on SI hardware generation */
1032 if (adev->asic_type < CHIP_BONAIRE) {
1033 adev->doorbell.base = 0;
1034 adev->doorbell.size = 0;
1035 adev->doorbell.num_doorbells = 0;
1036 adev->doorbell.ptr = NULL;
1040 if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET)
1043 amdgpu_asic_init_doorbell_index(adev);
1045 /* doorbell bar mapping */
1046 adev->doorbell.base = pci_resource_start(adev->pdev, 2);
1047 adev->doorbell.size = pci_resource_len(adev->pdev, 2);
1049 adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32),
1050 adev->doorbell_index.max_assignment+1);
1051 if (adev->doorbell.num_doorbells == 0)
1054 /* For Vega, reserve and map two pages on doorbell BAR since SDMA
1055 * paging queue doorbell use the second page. The
1056 * AMDGPU_DOORBELL64_MAX_ASSIGNMENT definition assumes all the
1057 * doorbells are in the first page. So with paging queue enabled,
1058 * the max num_doorbells should + 1 page (0x400 in dword)
1060 if (adev->asic_type >= CHIP_VEGA10)
1061 adev->doorbell.num_doorbells += 0x400;
1063 adev->doorbell.ptr = ioremap(adev->doorbell.base,
1064 adev->doorbell.num_doorbells *
1066 if (adev->doorbell.ptr == NULL)
1073 * amdgpu_device_doorbell_fini - Tear down doorbell driver information.
1075 * @adev: amdgpu_device pointer
1077 * Tear down doorbell driver information (CIK)
1079 static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev)
1081 iounmap(adev->doorbell.ptr);
1082 adev->doorbell.ptr = NULL;
1088 * amdgpu_device_wb_*()
1089 * Writeback is the method by which the GPU updates special pages in memory
1090 * with the status of certain GPU events (fences, ring pointers,etc.).
1094 * amdgpu_device_wb_fini - Disable Writeback and free memory
1096 * @adev: amdgpu_device pointer
1098 * Disables Writeback and frees the Writeback memory (all asics).
1099 * Used at driver shutdown.
1101 static void amdgpu_device_wb_fini(struct amdgpu_device *adev)
1103 if (adev->wb.wb_obj) {
1104 amdgpu_bo_free_kernel(&adev->wb.wb_obj,
1106 (void **)&adev->wb.wb);
1107 adev->wb.wb_obj = NULL;
1112 * amdgpu_device_wb_init - Init Writeback driver info and allocate memory
1114 * @adev: amdgpu_device pointer
1116 * Initializes writeback and allocates writeback memory (all asics).
1117 * Used at driver startup.
1118 * Returns 0 on success or an -error on failure.
1120 static int amdgpu_device_wb_init(struct amdgpu_device *adev)
1124 if (adev->wb.wb_obj == NULL) {
1125 /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */
1126 r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8,
1127 PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
1128 &adev->wb.wb_obj, &adev->wb.gpu_addr,
1129 (void **)&adev->wb.wb);
1131 dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
1135 adev->wb.num_wb = AMDGPU_MAX_WB;
1136 memset(&adev->wb.used, 0, sizeof(adev->wb.used));
1138 /* clear wb memory */
1139 memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8);
1146 * amdgpu_device_wb_get - Allocate a wb entry
1148 * @adev: amdgpu_device pointer
1151 * Allocate a wb slot for use by the driver (all asics).
1152 * Returns 0 on success or -EINVAL on failure.
1154 int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb)
1156 unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
1158 if (offset < adev->wb.num_wb) {
1159 __set_bit(offset, adev->wb.used);
1160 *wb = offset << 3; /* convert to dw offset */
1168 * amdgpu_device_wb_free - Free a wb entry
1170 * @adev: amdgpu_device pointer
1173 * Free a wb slot allocated for use by the driver (all asics)
1175 void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb)
1178 if (wb < adev->wb.num_wb)
1179 __clear_bit(wb, adev->wb.used);
1183 * amdgpu_device_resize_fb_bar - try to resize FB BAR
1185 * @adev: amdgpu_device pointer
1187 * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not
1188 * to fail, but if any of the BARs is not accessible after the size we abort
1189 * driver loading by returning -ENODEV.
1191 int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev)
1193 int rbar_size = pci_rebar_bytes_to_size(adev->gmc.real_vram_size);
1194 struct pci_bus *root;
1195 struct resource *res;
1201 if (amdgpu_sriov_vf(adev))
1204 /* skip if the bios has already enabled large BAR */
1205 if (adev->gmc.real_vram_size &&
1206 (pci_resource_len(adev->pdev, 0) >= adev->gmc.real_vram_size))
1209 /* Check if the root BUS has 64bit memory resources */
1210 root = adev->pdev->bus;
1211 while (root->parent)
1212 root = root->parent;
1214 pci_bus_for_each_resource(root, res, i) {
1215 if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) &&
1216 res->start > 0x100000000ull)
1220 /* Trying to resize is pointless without a root hub window above 4GB */
1224 /* Limit the BAR size to what is available */
1225 rbar_size = min(fls(pci_rebar_get_possible_sizes(adev->pdev, 0)) - 1,
1228 /* Disable memory decoding while we change the BAR addresses and size */
1229 pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd);
1230 pci_write_config_word(adev->pdev, PCI_COMMAND,
1231 cmd & ~PCI_COMMAND_MEMORY);
1233 /* Free the VRAM and doorbell BAR, we most likely need to move both. */
1234 amdgpu_device_doorbell_fini(adev);
1235 if (adev->asic_type >= CHIP_BONAIRE)
1236 pci_release_resource(adev->pdev, 2);
1238 pci_release_resource(adev->pdev, 0);
1240 r = pci_resize_resource(adev->pdev, 0, rbar_size);
1242 DRM_INFO("Not enough PCI address space for a large BAR.");
1243 else if (r && r != -ENOTSUPP)
1244 DRM_ERROR("Problem resizing BAR0 (%d).", r);
1246 pci_assign_unassigned_bus_resources(adev->pdev->bus);
1248 /* When the doorbell or fb BAR isn't available we have no chance of
1251 r = amdgpu_device_doorbell_init(adev);
1252 if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET))
1255 pci_write_config_word(adev->pdev, PCI_COMMAND, cmd);
1261 * GPU helpers function.
1264 * amdgpu_device_need_post - check if the hw need post or not
1266 * @adev: amdgpu_device pointer
1268 * Check if the asic has been initialized (all asics) at driver startup
1269 * or post is needed if hw reset is performed.
1270 * Returns true if need or false if not.
1272 bool amdgpu_device_need_post(struct amdgpu_device *adev)
1276 if (amdgpu_sriov_vf(adev))
1279 if (amdgpu_passthrough(adev)) {
1280 /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
1281 * some old smc fw still need driver do vPost otherwise gpu hang, while
1282 * those smc fw version above 22.15 doesn't have this flaw, so we force
1283 * vpost executed for smc version below 22.15
1285 if (adev->asic_type == CHIP_FIJI) {
1288 err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
1289 /* force vPost if error occured */
1293 fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
1294 if (fw_ver < 0x00160e00)
1299 /* Don't post if we need to reset whole hive on init */
1300 if (adev->gmc.xgmi.pending_reset)
1303 if (adev->has_hw_reset) {
1304 adev->has_hw_reset = false;
1308 /* bios scratch used on CIK+ */
1309 if (adev->asic_type >= CHIP_BONAIRE)
1310 return amdgpu_atombios_scratch_need_asic_init(adev);
1312 /* check MEM_SIZE for older asics */
1313 reg = amdgpu_asic_get_config_memsize(adev);
1315 if ((reg != 0) && (reg != 0xffffffff))
1321 /* if we get transitioned to only one device, take VGA back */
1323 * amdgpu_device_vga_set_decode - enable/disable vga decode
1325 * @pdev: PCI device pointer
1326 * @state: enable/disable vga decode
1328 * Enable/disable vga decode (all asics).
1329 * Returns VGA resource flags.
1331 static unsigned int amdgpu_device_vga_set_decode(struct pci_dev *pdev,
1334 struct amdgpu_device *adev = drm_to_adev(pci_get_drvdata(pdev));
1335 amdgpu_asic_set_vga_state(adev, state);
1337 return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
1338 VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
1340 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
1344 * amdgpu_device_check_block_size - validate the vm block size
1346 * @adev: amdgpu_device pointer
1348 * Validates the vm block size specified via module parameter.
1349 * The vm block size defines number of bits in page table versus page directory,
1350 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
1351 * page table and the remaining bits are in the page directory.
1353 static void amdgpu_device_check_block_size(struct amdgpu_device *adev)
1355 /* defines number of bits in page table versus page directory,
1356 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
1357 * page table and the remaining bits are in the page directory */
1358 if (amdgpu_vm_block_size == -1)
1361 if (amdgpu_vm_block_size < 9) {
1362 dev_warn(adev->dev, "VM page table size (%d) too small\n",
1363 amdgpu_vm_block_size);
1364 amdgpu_vm_block_size = -1;
1369 * amdgpu_device_check_vm_size - validate the vm size
1371 * @adev: amdgpu_device pointer
1373 * Validates the vm size in GB specified via module parameter.
1374 * The VM size is the size of the GPU virtual memory space in GB.
1376 static void amdgpu_device_check_vm_size(struct amdgpu_device *adev)
1378 /* no need to check the default value */
1379 if (amdgpu_vm_size == -1)
1382 if (amdgpu_vm_size < 1) {
1383 dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
1385 amdgpu_vm_size = -1;
1389 static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev)
1392 bool is_os_64 = (sizeof(void *) == 8);
1393 uint64_t total_memory;
1394 uint64_t dram_size_seven_GB = 0x1B8000000;
1395 uint64_t dram_size_three_GB = 0xB8000000;
1397 if (amdgpu_smu_memory_pool_size == 0)
1401 DRM_WARN("Not 64-bit OS, feature not supported\n");
1405 total_memory = (uint64_t)si.totalram * si.mem_unit;
1407 if ((amdgpu_smu_memory_pool_size == 1) ||
1408 (amdgpu_smu_memory_pool_size == 2)) {
1409 if (total_memory < dram_size_three_GB)
1411 } else if ((amdgpu_smu_memory_pool_size == 4) ||
1412 (amdgpu_smu_memory_pool_size == 8)) {
1413 if (total_memory < dram_size_seven_GB)
1416 DRM_WARN("Smu memory pool size not supported\n");
1419 adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28;
1424 DRM_WARN("No enough system memory\n");
1426 adev->pm.smu_prv_buffer_size = 0;
1429 static int amdgpu_device_init_apu_flags(struct amdgpu_device *adev)
1431 if (!(adev->flags & AMD_IS_APU) ||
1432 adev->asic_type < CHIP_RAVEN)
1435 switch (adev->asic_type) {
1437 if (adev->pdev->device == 0x15dd)
1438 adev->apu_flags |= AMD_APU_IS_RAVEN;
1439 if (adev->pdev->device == 0x15d8)
1440 adev->apu_flags |= AMD_APU_IS_PICASSO;
1443 if ((adev->pdev->device == 0x1636) ||
1444 (adev->pdev->device == 0x164c))
1445 adev->apu_flags |= AMD_APU_IS_RENOIR;
1447 adev->apu_flags |= AMD_APU_IS_GREEN_SARDINE;
1450 adev->apu_flags |= AMD_APU_IS_VANGOGH;
1452 case CHIP_YELLOW_CARP:
1454 case CHIP_CYAN_SKILLFISH:
1455 if ((adev->pdev->device == 0x13FE) ||
1456 (adev->pdev->device == 0x143F))
1457 adev->apu_flags |= AMD_APU_IS_CYAN_SKILLFISH2;
1467 * amdgpu_device_check_arguments - validate module params
1469 * @adev: amdgpu_device pointer
1471 * Validates certain module parameters and updates
1472 * the associated values used by the driver (all asics).
1474 static int amdgpu_device_check_arguments(struct amdgpu_device *adev)
1476 if (amdgpu_sched_jobs < 4) {
1477 dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
1479 amdgpu_sched_jobs = 4;
1480 } else if (!is_power_of_2(amdgpu_sched_jobs)){
1481 dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
1483 amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
1486 if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) {
1487 /* gart size must be greater or equal to 32M */
1488 dev_warn(adev->dev, "gart size (%d) too small\n",
1490 amdgpu_gart_size = -1;
1493 if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) {
1494 /* gtt size must be greater or equal to 32M */
1495 dev_warn(adev->dev, "gtt size (%d) too small\n",
1497 amdgpu_gtt_size = -1;
1500 /* valid range is between 4 and 9 inclusive */
1501 if (amdgpu_vm_fragment_size != -1 &&
1502 (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) {
1503 dev_warn(adev->dev, "valid range is between 4 and 9\n");
1504 amdgpu_vm_fragment_size = -1;
1507 if (amdgpu_sched_hw_submission < 2) {
1508 dev_warn(adev->dev, "sched hw submission jobs (%d) must be at least 2\n",
1509 amdgpu_sched_hw_submission);
1510 amdgpu_sched_hw_submission = 2;
1511 } else if (!is_power_of_2(amdgpu_sched_hw_submission)) {
1512 dev_warn(adev->dev, "sched hw submission jobs (%d) must be a power of 2\n",
1513 amdgpu_sched_hw_submission);
1514 amdgpu_sched_hw_submission = roundup_pow_of_two(amdgpu_sched_hw_submission);
1517 amdgpu_device_check_smu_prv_buffer_size(adev);
1519 amdgpu_device_check_vm_size(adev);
1521 amdgpu_device_check_block_size(adev);
1523 adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type);
1525 amdgpu_gmc_tmz_set(adev);
1527 amdgpu_gmc_noretry_set(adev);
1533 * amdgpu_switcheroo_set_state - set switcheroo state
1535 * @pdev: pci dev pointer
1536 * @state: vga_switcheroo state
1538 * Callback for the switcheroo driver. Suspends or resumes the
1539 * the asics before or after it is powered up using ACPI methods.
1541 static void amdgpu_switcheroo_set_state(struct pci_dev *pdev,
1542 enum vga_switcheroo_state state)
1544 struct drm_device *dev = pci_get_drvdata(pdev);
1547 if (amdgpu_device_supports_px(dev) && state == VGA_SWITCHEROO_OFF)
1550 if (state == VGA_SWITCHEROO_ON) {
1551 pr_info("switched on\n");
1552 /* don't suspend or resume card normally */
1553 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1555 pci_set_power_state(pdev, PCI_D0);
1556 amdgpu_device_load_pci_state(pdev);
1557 r = pci_enable_device(pdev);
1559 DRM_WARN("pci_enable_device failed (%d)\n", r);
1560 amdgpu_device_resume(dev, true);
1562 dev->switch_power_state = DRM_SWITCH_POWER_ON;
1564 pr_info("switched off\n");
1565 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1566 amdgpu_device_suspend(dev, true);
1567 amdgpu_device_cache_pci_state(pdev);
1568 /* Shut down the device */
1569 pci_disable_device(pdev);
1570 pci_set_power_state(pdev, PCI_D3cold);
1571 dev->switch_power_state = DRM_SWITCH_POWER_OFF;
1576 * amdgpu_switcheroo_can_switch - see if switcheroo state can change
1578 * @pdev: pci dev pointer
1580 * Callback for the switcheroo driver. Check of the switcheroo
1581 * state can be changed.
1582 * Returns true if the state can be changed, false if not.
1584 static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
1586 struct drm_device *dev = pci_get_drvdata(pdev);
1589 * FIXME: open_count is protected by drm_global_mutex but that would lead to
1590 * locking inversion with the driver load path. And the access here is
1591 * completely racy anyway. So don't bother with locking for now.
1593 return atomic_read(&dev->open_count) == 0;
1596 static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
1597 .set_gpu_state = amdgpu_switcheroo_set_state,
1599 .can_switch = amdgpu_switcheroo_can_switch,
1603 * amdgpu_device_ip_set_clockgating_state - set the CG state
1605 * @dev: amdgpu_device pointer
1606 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1607 * @state: clockgating state (gate or ungate)
1609 * Sets the requested clockgating state for all instances of
1610 * the hardware IP specified.
1611 * Returns the error code from the last instance.
1613 int amdgpu_device_ip_set_clockgating_state(void *dev,
1614 enum amd_ip_block_type block_type,
1615 enum amd_clockgating_state state)
1617 struct amdgpu_device *adev = dev;
1620 for (i = 0; i < adev->num_ip_blocks; i++) {
1621 if (!adev->ip_blocks[i].status.valid)
1623 if (adev->ip_blocks[i].version->type != block_type)
1625 if (!adev->ip_blocks[i].version->funcs->set_clockgating_state)
1627 r = adev->ip_blocks[i].version->funcs->set_clockgating_state(
1628 (void *)adev, state);
1630 DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",
1631 adev->ip_blocks[i].version->funcs->name, r);
1637 * amdgpu_device_ip_set_powergating_state - set the PG state
1639 * @dev: amdgpu_device pointer
1640 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1641 * @state: powergating state (gate or ungate)
1643 * Sets the requested powergating state for all instances of
1644 * the hardware IP specified.
1645 * Returns the error code from the last instance.
1647 int amdgpu_device_ip_set_powergating_state(void *dev,
1648 enum amd_ip_block_type block_type,
1649 enum amd_powergating_state state)
1651 struct amdgpu_device *adev = dev;
1654 for (i = 0; i < adev->num_ip_blocks; i++) {
1655 if (!adev->ip_blocks[i].status.valid)
1657 if (adev->ip_blocks[i].version->type != block_type)
1659 if (!adev->ip_blocks[i].version->funcs->set_powergating_state)
1661 r = adev->ip_blocks[i].version->funcs->set_powergating_state(
1662 (void *)adev, state);
1664 DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",
1665 adev->ip_blocks[i].version->funcs->name, r);
1671 * amdgpu_device_ip_get_clockgating_state - get the CG state
1673 * @adev: amdgpu_device pointer
1674 * @flags: clockgating feature flags
1676 * Walks the list of IPs on the device and updates the clockgating
1677 * flags for each IP.
1678 * Updates @flags with the feature flags for each hardware IP where
1679 * clockgating is enabled.
1681 void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev,
1686 for (i = 0; i < adev->num_ip_blocks; i++) {
1687 if (!adev->ip_blocks[i].status.valid)
1689 if (adev->ip_blocks[i].version->funcs->get_clockgating_state)
1690 adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags);
1695 * amdgpu_device_ip_wait_for_idle - wait for idle
1697 * @adev: amdgpu_device pointer
1698 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1700 * Waits for the request hardware IP to be idle.
1701 * Returns 0 for success or a negative error code on failure.
1703 int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev,
1704 enum amd_ip_block_type block_type)
1708 for (i = 0; i < adev->num_ip_blocks; i++) {
1709 if (!adev->ip_blocks[i].status.valid)
1711 if (adev->ip_blocks[i].version->type == block_type) {
1712 r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
1723 * amdgpu_device_ip_is_idle - is the hardware IP idle
1725 * @adev: amdgpu_device pointer
1726 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1728 * Check if the hardware IP is idle or not.
1729 * Returns true if it the IP is idle, false if not.
1731 bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev,
1732 enum amd_ip_block_type block_type)
1736 for (i = 0; i < adev->num_ip_blocks; i++) {
1737 if (!adev->ip_blocks[i].status.valid)
1739 if (adev->ip_blocks[i].version->type == block_type)
1740 return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
1747 * amdgpu_device_ip_get_ip_block - get a hw IP pointer
1749 * @adev: amdgpu_device pointer
1750 * @type: Type of hardware IP (SMU, GFX, UVD, etc.)
1752 * Returns a pointer to the hardware IP block structure
1753 * if it exists for the asic, otherwise NULL.
1755 struct amdgpu_ip_block *
1756 amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,
1757 enum amd_ip_block_type type)
1761 for (i = 0; i < adev->num_ip_blocks; i++)
1762 if (adev->ip_blocks[i].version->type == type)
1763 return &adev->ip_blocks[i];
1769 * amdgpu_device_ip_block_version_cmp
1771 * @adev: amdgpu_device pointer
1772 * @type: enum amd_ip_block_type
1773 * @major: major version
1774 * @minor: minor version
1776 * return 0 if equal or greater
1777 * return 1 if smaller or the ip_block doesn't exist
1779 int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev,
1780 enum amd_ip_block_type type,
1781 u32 major, u32 minor)
1783 struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type);
1785 if (ip_block && ((ip_block->version->major > major) ||
1786 ((ip_block->version->major == major) &&
1787 (ip_block->version->minor >= minor))))
1794 * amdgpu_device_ip_block_add
1796 * @adev: amdgpu_device pointer
1797 * @ip_block_version: pointer to the IP to add
1799 * Adds the IP block driver information to the collection of IPs
1802 int amdgpu_device_ip_block_add(struct amdgpu_device *adev,
1803 const struct amdgpu_ip_block_version *ip_block_version)
1805 if (!ip_block_version)
1808 switch (ip_block_version->type) {
1809 case AMD_IP_BLOCK_TYPE_VCN:
1810 if (adev->harvest_ip_mask & AMD_HARVEST_IP_VCN_MASK)
1813 case AMD_IP_BLOCK_TYPE_JPEG:
1814 if (adev->harvest_ip_mask & AMD_HARVEST_IP_JPEG_MASK)
1821 DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks,
1822 ip_block_version->funcs->name);
1824 adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
1830 * amdgpu_device_enable_virtual_display - enable virtual display feature
1832 * @adev: amdgpu_device pointer
1834 * Enabled the virtual display feature if the user has enabled it via
1835 * the module parameter virtual_display. This feature provides a virtual
1836 * display hardware on headless boards or in virtualized environments.
1837 * This function parses and validates the configuration string specified by
1838 * the user and configues the virtual display configuration (number of
1839 * virtual connectors, crtcs, etc.) specified.
1841 static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
1843 adev->enable_virtual_display = false;
1845 if (amdgpu_virtual_display) {
1846 const char *pci_address_name = pci_name(adev->pdev);
1847 char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
1849 pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
1850 pciaddstr_tmp = pciaddstr;
1851 while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
1852 pciaddname = strsep(&pciaddname_tmp, ",");
1853 if (!strcmp("all", pciaddname)
1854 || !strcmp(pci_address_name, pciaddname)) {
1858 adev->enable_virtual_display = true;
1861 res = kstrtol(pciaddname_tmp, 10,
1869 adev->mode_info.num_crtc = num_crtc;
1871 adev->mode_info.num_crtc = 1;
1877 DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
1878 amdgpu_virtual_display, pci_address_name,
1879 adev->enable_virtual_display, adev->mode_info.num_crtc);
1886 * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware
1888 * @adev: amdgpu_device pointer
1890 * Parses the asic configuration parameters specified in the gpu info
1891 * firmware and makes them availale to the driver for use in configuring
1893 * Returns 0 on success, -EINVAL on failure.
1895 static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev)
1897 const char *chip_name;
1900 const struct gpu_info_firmware_header_v1_0 *hdr;
1902 adev->firmware.gpu_info_fw = NULL;
1904 if (adev->mman.discovery_bin) {
1905 amdgpu_discovery_get_gfx_info(adev);
1908 * FIXME: The bounding box is still needed by Navi12, so
1909 * temporarily read it from gpu_info firmware. Should be droped
1910 * when DAL no longer needs it.
1912 if (adev->asic_type != CHIP_NAVI12)
1916 switch (adev->asic_type) {
1917 #ifdef CONFIG_DRM_AMDGPU_SI
1924 #ifdef CONFIG_DRM_AMDGPU_CIK
1934 case CHIP_POLARIS10:
1935 case CHIP_POLARIS11:
1936 case CHIP_POLARIS12:
1941 case CHIP_ALDEBARAN:
1942 case CHIP_SIENNA_CICHLID:
1943 case CHIP_NAVY_FLOUNDER:
1944 case CHIP_DIMGREY_CAVEFISH:
1945 case CHIP_BEIGE_GOBY:
1949 chip_name = "vega10";
1952 chip_name = "vega12";
1955 if (adev->apu_flags & AMD_APU_IS_RAVEN2)
1956 chip_name = "raven2";
1957 else if (adev->apu_flags & AMD_APU_IS_PICASSO)
1958 chip_name = "picasso";
1960 chip_name = "raven";
1963 chip_name = "arcturus";
1966 if (adev->apu_flags & AMD_APU_IS_RENOIR)
1967 chip_name = "renoir";
1969 chip_name = "green_sardine";
1972 chip_name = "navi10";
1975 chip_name = "navi14";
1978 chip_name = "navi12";
1981 chip_name = "vangogh";
1983 case CHIP_YELLOW_CARP:
1984 chip_name = "yellow_carp";
1988 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name);
1989 err = request_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev);
1992 "Failed to load gpu_info firmware \"%s\"\n",
1996 err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw);
1999 "Failed to validate gpu_info firmware \"%s\"\n",
2004 hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data;
2005 amdgpu_ucode_print_gpu_info_hdr(&hdr->header);
2007 switch (hdr->version_major) {
2010 const struct gpu_info_firmware_v1_0 *gpu_info_fw =
2011 (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data +
2012 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
2015 * Should be droped when DAL no longer needs it.
2017 if (adev->asic_type == CHIP_NAVI12)
2018 goto parse_soc_bounding_box;
2020 adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se);
2021 adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh);
2022 adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se);
2023 adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se);
2024 adev->gfx.config.max_texture_channel_caches =
2025 le32_to_cpu(gpu_info_fw->gc_num_tccs);
2026 adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs);
2027 adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds);
2028 adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth);
2029 adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth);
2030 adev->gfx.config.double_offchip_lds_buf =
2031 le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer);
2032 adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size);
2033 adev->gfx.cu_info.max_waves_per_simd =
2034 le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd);
2035 adev->gfx.cu_info.max_scratch_slots_per_cu =
2036 le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu);
2037 adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size);
2038 if (hdr->version_minor >= 1) {
2039 const struct gpu_info_firmware_v1_1 *gpu_info_fw =
2040 (const struct gpu_info_firmware_v1_1 *)(adev->firmware.gpu_info_fw->data +
2041 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
2042 adev->gfx.config.num_sc_per_sh =
2043 le32_to_cpu(gpu_info_fw->num_sc_per_sh);
2044 adev->gfx.config.num_packer_per_sc =
2045 le32_to_cpu(gpu_info_fw->num_packer_per_sc);
2048 parse_soc_bounding_box:
2050 * soc bounding box info is not integrated in disocovery table,
2051 * we always need to parse it from gpu info firmware if needed.
2053 if (hdr->version_minor == 2) {
2054 const struct gpu_info_firmware_v1_2 *gpu_info_fw =
2055 (const struct gpu_info_firmware_v1_2 *)(adev->firmware.gpu_info_fw->data +
2056 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
2057 adev->dm.soc_bounding_box = &gpu_info_fw->soc_bounding_box;
2063 "Unsupported gpu_info table %d\n", hdr->header.ucode_version);
2072 * amdgpu_device_ip_early_init - run early init for hardware IPs
2074 * @adev: amdgpu_device pointer
2076 * Early initialization pass for hardware IPs. The hardware IPs that make
2077 * up each asic are discovered each IP's early_init callback is run. This
2078 * is the first stage in initializing the asic.
2079 * Returns 0 on success, negative error code on failure.
2081 static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
2083 struct drm_device *dev = adev_to_drm(adev);
2084 struct pci_dev *parent;
2087 amdgpu_device_enable_virtual_display(adev);
2089 if (amdgpu_sriov_vf(adev)) {
2090 r = amdgpu_virt_request_full_gpu(adev, true);
2095 switch (adev->asic_type) {
2096 #ifdef CONFIG_DRM_AMDGPU_SI
2102 adev->family = AMDGPU_FAMILY_SI;
2103 r = si_set_ip_blocks(adev);
2108 #ifdef CONFIG_DRM_AMDGPU_CIK
2114 if (adev->flags & AMD_IS_APU)
2115 adev->family = AMDGPU_FAMILY_KV;
2117 adev->family = AMDGPU_FAMILY_CI;
2119 r = cik_set_ip_blocks(adev);
2127 case CHIP_POLARIS10:
2128 case CHIP_POLARIS11:
2129 case CHIP_POLARIS12:
2133 if (adev->flags & AMD_IS_APU)
2134 adev->family = AMDGPU_FAMILY_CZ;
2136 adev->family = AMDGPU_FAMILY_VI;
2138 r = vi_set_ip_blocks(adev);
2143 r = amdgpu_discovery_set_ip_blocks(adev);
2149 if (amdgpu_has_atpx() &&
2150 (amdgpu_is_atpx_hybrid() ||
2151 amdgpu_has_atpx_dgpu_power_cntl()) &&
2152 ((adev->flags & AMD_IS_APU) == 0) &&
2153 !pci_is_thunderbolt_attached(to_pci_dev(dev->dev)))
2154 adev->flags |= AMD_IS_PX;
2156 parent = pci_upstream_bridge(adev->pdev);
2157 adev->has_pr3 = parent ? pci_pr3_present(parent) : false;
2159 amdgpu_amdkfd_device_probe(adev);
2161 adev->pm.pp_feature = amdgpu_pp_feature_mask;
2162 if (amdgpu_sriov_vf(adev) || sched_policy == KFD_SCHED_POLICY_NO_HWS)
2163 adev->pm.pp_feature &= ~PP_GFXOFF_MASK;
2164 if (amdgpu_sriov_vf(adev) && adev->asic_type == CHIP_SIENNA_CICHLID)
2165 adev->pm.pp_feature &= ~PP_OVERDRIVE_MASK;
2167 for (i = 0; i < adev->num_ip_blocks; i++) {
2168 if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
2169 DRM_ERROR("disabled ip block: %d <%s>\n",
2170 i, adev->ip_blocks[i].version->funcs->name);
2171 adev->ip_blocks[i].status.valid = false;
2173 if (adev->ip_blocks[i].version->funcs->early_init) {
2174 r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
2176 adev->ip_blocks[i].status.valid = false;
2178 DRM_ERROR("early_init of IP block <%s> failed %d\n",
2179 adev->ip_blocks[i].version->funcs->name, r);
2182 adev->ip_blocks[i].status.valid = true;
2185 adev->ip_blocks[i].status.valid = true;
2188 /* get the vbios after the asic_funcs are set up */
2189 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) {
2190 r = amdgpu_device_parse_gpu_info_fw(adev);
2195 if (!amdgpu_get_bios(adev))
2198 r = amdgpu_atombios_init(adev);
2200 dev_err(adev->dev, "amdgpu_atombios_init failed\n");
2201 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0);
2205 /*get pf2vf msg info at it's earliest time*/
2206 if (amdgpu_sriov_vf(adev))
2207 amdgpu_virt_init_data_exchange(adev);
2212 adev->cg_flags &= amdgpu_cg_mask;
2213 adev->pg_flags &= amdgpu_pg_mask;
2218 static int amdgpu_device_ip_hw_init_phase1(struct amdgpu_device *adev)
2222 for (i = 0; i < adev->num_ip_blocks; i++) {
2223 if (!adev->ip_blocks[i].status.sw)
2225 if (adev->ip_blocks[i].status.hw)
2227 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2228 (amdgpu_sriov_vf(adev) && (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)) ||
2229 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
2230 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
2232 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
2233 adev->ip_blocks[i].version->funcs->name, r);
2236 adev->ip_blocks[i].status.hw = true;
2243 static int amdgpu_device_ip_hw_init_phase2(struct amdgpu_device *adev)
2247 for (i = 0; i < adev->num_ip_blocks; i++) {
2248 if (!adev->ip_blocks[i].status.sw)
2250 if (adev->ip_blocks[i].status.hw)
2252 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
2254 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
2255 adev->ip_blocks[i].version->funcs->name, r);
2258 adev->ip_blocks[i].status.hw = true;
2264 static int amdgpu_device_fw_loading(struct amdgpu_device *adev)
2268 uint32_t smu_version;
2270 if (adev->asic_type >= CHIP_VEGA10) {
2271 for (i = 0; i < adev->num_ip_blocks; i++) {
2272 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_PSP)
2275 if (!adev->ip_blocks[i].status.sw)
2278 /* no need to do the fw loading again if already done*/
2279 if (adev->ip_blocks[i].status.hw == true)
2282 if (amdgpu_in_reset(adev) || adev->in_suspend) {
2283 r = adev->ip_blocks[i].version->funcs->resume(adev);
2285 DRM_ERROR("resume of IP block <%s> failed %d\n",
2286 adev->ip_blocks[i].version->funcs->name, r);
2290 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
2292 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
2293 adev->ip_blocks[i].version->funcs->name, r);
2298 adev->ip_blocks[i].status.hw = true;
2303 if (!amdgpu_sriov_vf(adev) || adev->asic_type == CHIP_TONGA)
2304 r = amdgpu_pm_load_smu_firmware(adev, &smu_version);
2310 * amdgpu_device_ip_init - run init for hardware IPs
2312 * @adev: amdgpu_device pointer
2314 * Main initialization pass for hardware IPs. The list of all the hardware
2315 * IPs that make up the asic is walked and the sw_init and hw_init callbacks
2316 * are run. sw_init initializes the software state associated with each IP
2317 * and hw_init initializes the hardware associated with each IP.
2318 * Returns 0 on success, negative error code on failure.
2320 static int amdgpu_device_ip_init(struct amdgpu_device *adev)
2324 r = amdgpu_ras_init(adev);
2328 for (i = 0; i < adev->num_ip_blocks; i++) {
2329 if (!adev->ip_blocks[i].status.valid)
2331 r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
2333 DRM_ERROR("sw_init of IP block <%s> failed %d\n",
2334 adev->ip_blocks[i].version->funcs->name, r);
2337 adev->ip_blocks[i].status.sw = true;
2339 /* need to do gmc hw init early so we can allocate gpu mem */
2340 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
2341 /* Try to reserve bad pages early */
2342 if (amdgpu_sriov_vf(adev))
2343 amdgpu_virt_exchange_data(adev);
2345 r = amdgpu_device_vram_scratch_init(adev);
2347 DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r);
2350 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
2352 DRM_ERROR("hw_init %d failed %d\n", i, r);
2355 r = amdgpu_device_wb_init(adev);
2357 DRM_ERROR("amdgpu_device_wb_init failed %d\n", r);
2360 adev->ip_blocks[i].status.hw = true;
2362 /* right after GMC hw init, we create CSA */
2363 if (amdgpu_mcbp || amdgpu_sriov_vf(adev)) {
2364 r = amdgpu_allocate_static_csa(adev, &adev->virt.csa_obj,
2365 AMDGPU_GEM_DOMAIN_VRAM,
2368 DRM_ERROR("allocate CSA failed %d\n", r);
2375 if (amdgpu_sriov_vf(adev))
2376 amdgpu_virt_init_data_exchange(adev);
2378 r = amdgpu_ib_pool_init(adev);
2380 dev_err(adev->dev, "IB initialization failed (%d).\n", r);
2381 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r);
2385 r = amdgpu_ucode_create_bo(adev); /* create ucode bo when sw_init complete*/
2389 r = amdgpu_device_ip_hw_init_phase1(adev);
2393 r = amdgpu_device_fw_loading(adev);
2397 r = amdgpu_device_ip_hw_init_phase2(adev);
2402 * retired pages will be loaded from eeprom and reserved here,
2403 * it should be called after amdgpu_device_ip_hw_init_phase2 since
2404 * for some ASICs the RAS EEPROM code relies on SMU fully functioning
2405 * for I2C communication which only true at this point.
2407 * amdgpu_ras_recovery_init may fail, but the upper only cares the
2408 * failure from bad gpu situation and stop amdgpu init process
2409 * accordingly. For other failed cases, it will still release all
2410 * the resource and print error message, rather than returning one
2411 * negative value to upper level.
2413 * Note: theoretically, this should be called before all vram allocations
2414 * to protect retired page from abusing
2416 r = amdgpu_ras_recovery_init(adev);
2420 if (adev->gmc.xgmi.num_physical_nodes > 1)
2421 amdgpu_xgmi_add_device(adev);
2423 /* Don't init kfd if whole hive need to be reset during init */
2424 if (!adev->gmc.xgmi.pending_reset)
2425 amdgpu_amdkfd_device_init(adev);
2427 amdgpu_fru_get_product_info(adev);
2430 if (amdgpu_sriov_vf(adev))
2431 amdgpu_virt_release_full_gpu(adev, true);
2437 * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer
2439 * @adev: amdgpu_device pointer
2441 * Writes a reset magic value to the gart pointer in VRAM. The driver calls
2442 * this function before a GPU reset. If the value is retained after a
2443 * GPU reset, VRAM has not been lost. Some GPU resets may destry VRAM contents.
2445 static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev)
2447 memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
2451 * amdgpu_device_check_vram_lost - check if vram is valid
2453 * @adev: amdgpu_device pointer
2455 * Checks the reset magic value written to the gart pointer in VRAM.
2456 * The driver calls this after a GPU reset to see if the contents of
2457 * VRAM is lost or now.
2458 * returns true if vram is lost, false if not.
2460 static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev)
2462 if (memcmp(adev->gart.ptr, adev->reset_magic,
2463 AMDGPU_RESET_MAGIC_NUM))
2466 if (!amdgpu_in_reset(adev))
2470 * For all ASICs with baco/mode1 reset, the VRAM is
2471 * always assumed to be lost.
2473 switch (amdgpu_asic_reset_method(adev)) {
2474 case AMD_RESET_METHOD_BACO:
2475 case AMD_RESET_METHOD_MODE1:
2483 * amdgpu_device_set_cg_state - set clockgating for amdgpu device
2485 * @adev: amdgpu_device pointer
2486 * @state: clockgating state (gate or ungate)
2488 * The list of all the hardware IPs that make up the asic is walked and the
2489 * set_clockgating_state callbacks are run.
2490 * Late initialization pass enabling clockgating for hardware IPs.
2491 * Fini or suspend, pass disabling clockgating for hardware IPs.
2492 * Returns 0 on success, negative error code on failure.
2495 int amdgpu_device_set_cg_state(struct amdgpu_device *adev,
2496 enum amd_clockgating_state state)
2500 if (amdgpu_emu_mode == 1)
2503 for (j = 0; j < adev->num_ip_blocks; j++) {
2504 i = state == AMD_CG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
2505 if (!adev->ip_blocks[i].status.late_initialized)
2507 /* skip CG for GFX on S0ix */
2508 if (adev->in_s0ix &&
2509 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX)
2511 /* skip CG for VCE/UVD, it's handled specially */
2512 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
2513 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
2514 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
2515 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG &&
2516 adev->ip_blocks[i].version->funcs->set_clockgating_state) {
2517 /* enable clockgating to save power */
2518 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
2521 DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
2522 adev->ip_blocks[i].version->funcs->name, r);
2531 int amdgpu_device_set_pg_state(struct amdgpu_device *adev,
2532 enum amd_powergating_state state)
2536 if (amdgpu_emu_mode == 1)
2539 for (j = 0; j < adev->num_ip_blocks; j++) {
2540 i = state == AMD_PG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
2541 if (!adev->ip_blocks[i].status.late_initialized)
2543 /* skip PG for GFX on S0ix */
2544 if (adev->in_s0ix &&
2545 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX)
2547 /* skip CG for VCE/UVD, it's handled specially */
2548 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
2549 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
2550 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
2551 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG &&
2552 adev->ip_blocks[i].version->funcs->set_powergating_state) {
2553 /* enable powergating to save power */
2554 r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev,
2557 DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n",
2558 adev->ip_blocks[i].version->funcs->name, r);
2566 static int amdgpu_device_enable_mgpu_fan_boost(void)
2568 struct amdgpu_gpu_instance *gpu_ins;
2569 struct amdgpu_device *adev;
2572 mutex_lock(&mgpu_info.mutex);
2575 * MGPU fan boost feature should be enabled
2576 * only when there are two or more dGPUs in
2579 if (mgpu_info.num_dgpu < 2)
2582 for (i = 0; i < mgpu_info.num_dgpu; i++) {
2583 gpu_ins = &(mgpu_info.gpu_ins[i]);
2584 adev = gpu_ins->adev;
2585 if (!(adev->flags & AMD_IS_APU) &&
2586 !gpu_ins->mgpu_fan_enabled) {
2587 ret = amdgpu_dpm_enable_mgpu_fan_boost(adev);
2591 gpu_ins->mgpu_fan_enabled = 1;
2596 mutex_unlock(&mgpu_info.mutex);
2602 * amdgpu_device_ip_late_init - run late init for hardware IPs
2604 * @adev: amdgpu_device pointer
2606 * Late initialization pass for hardware IPs. The list of all the hardware
2607 * IPs that make up the asic is walked and the late_init callbacks are run.
2608 * late_init covers any special initialization that an IP requires
2609 * after all of the have been initialized or something that needs to happen
2610 * late in the init process.
2611 * Returns 0 on success, negative error code on failure.
2613 static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
2615 struct amdgpu_gpu_instance *gpu_instance;
2618 for (i = 0; i < adev->num_ip_blocks; i++) {
2619 if (!adev->ip_blocks[i].status.hw)
2621 if (adev->ip_blocks[i].version->funcs->late_init) {
2622 r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
2624 DRM_ERROR("late_init of IP block <%s> failed %d\n",
2625 adev->ip_blocks[i].version->funcs->name, r);
2629 adev->ip_blocks[i].status.late_initialized = true;
2632 amdgpu_ras_set_error_query_ready(adev, true);
2634 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE);
2635 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE);
2637 amdgpu_device_fill_reset_magic(adev);
2639 r = amdgpu_device_enable_mgpu_fan_boost();
2641 DRM_ERROR("enable mgpu fan boost failed (%d).\n", r);
2643 /* For passthrough configuration on arcturus and aldebaran, enable special handling SBR */
2644 if (amdgpu_passthrough(adev) && ((adev->asic_type == CHIP_ARCTURUS && adev->gmc.xgmi.num_physical_nodes > 1)||
2645 adev->asic_type == CHIP_ALDEBARAN ))
2646 amdgpu_dpm_handle_passthrough_sbr(adev, true);
2648 if (adev->gmc.xgmi.num_physical_nodes > 1) {
2649 mutex_lock(&mgpu_info.mutex);
2652 * Reset device p-state to low as this was booted with high.
2654 * This should be performed only after all devices from the same
2655 * hive get initialized.
2657 * However, it's unknown how many device in the hive in advance.
2658 * As this is counted one by one during devices initializations.
2660 * So, we wait for all XGMI interlinked devices initialized.
2661 * This may bring some delays as those devices may come from
2662 * different hives. But that should be OK.
2664 if (mgpu_info.num_dgpu == adev->gmc.xgmi.num_physical_nodes) {
2665 for (i = 0; i < mgpu_info.num_gpu; i++) {
2666 gpu_instance = &(mgpu_info.gpu_ins[i]);
2667 if (gpu_instance->adev->flags & AMD_IS_APU)
2670 r = amdgpu_xgmi_set_pstate(gpu_instance->adev,
2671 AMDGPU_XGMI_PSTATE_MIN);
2673 DRM_ERROR("pstate setting failed (%d).\n", r);
2679 mutex_unlock(&mgpu_info.mutex);
2686 * amdgpu_device_smu_fini_early - smu hw_fini wrapper
2688 * @adev: amdgpu_device pointer
2690 * For ASICs need to disable SMC first
2692 static void amdgpu_device_smu_fini_early(struct amdgpu_device *adev)
2696 if (adev->ip_versions[GC_HWIP][0] > IP_VERSION(9, 0, 0))
2699 for (i = 0; i < adev->num_ip_blocks; i++) {
2700 if (!adev->ip_blocks[i].status.hw)
2702 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
2703 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
2704 /* XXX handle errors */
2706 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
2707 adev->ip_blocks[i].version->funcs->name, r);
2709 adev->ip_blocks[i].status.hw = false;
2715 static int amdgpu_device_ip_fini_early(struct amdgpu_device *adev)
2719 for (i = 0; i < adev->num_ip_blocks; i++) {
2720 if (!adev->ip_blocks[i].version->funcs->early_fini)
2723 r = adev->ip_blocks[i].version->funcs->early_fini((void *)adev);
2725 DRM_DEBUG("early_fini of IP block <%s> failed %d\n",
2726 adev->ip_blocks[i].version->funcs->name, r);
2730 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
2731 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
2733 amdgpu_amdkfd_suspend(adev, false);
2735 /* Workaroud for ASICs need to disable SMC first */
2736 amdgpu_device_smu_fini_early(adev);
2738 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2739 if (!adev->ip_blocks[i].status.hw)
2742 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
2743 /* XXX handle errors */
2745 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
2746 adev->ip_blocks[i].version->funcs->name, r);
2749 adev->ip_blocks[i].status.hw = false;
2752 if (amdgpu_sriov_vf(adev)) {
2753 if (amdgpu_virt_release_full_gpu(adev, false))
2754 DRM_ERROR("failed to release exclusive mode on fini\n");
2761 * amdgpu_device_ip_fini - run fini for hardware IPs
2763 * @adev: amdgpu_device pointer
2765 * Main teardown pass for hardware IPs. The list of all the hardware
2766 * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks
2767 * are run. hw_fini tears down the hardware associated with each IP
2768 * and sw_fini tears down any software state associated with each IP.
2769 * Returns 0 on success, negative error code on failure.
2771 static int amdgpu_device_ip_fini(struct amdgpu_device *adev)
2775 if (amdgpu_sriov_vf(adev) && adev->virt.ras_init_done)
2776 amdgpu_virt_release_ras_err_handler_data(adev);
2778 if (adev->gmc.xgmi.num_physical_nodes > 1)
2779 amdgpu_xgmi_remove_device(adev);
2781 amdgpu_amdkfd_device_fini_sw(adev);
2783 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2784 if (!adev->ip_blocks[i].status.sw)
2787 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
2788 amdgpu_ucode_free_bo(adev);
2789 amdgpu_free_static_csa(&adev->virt.csa_obj);
2790 amdgpu_device_wb_fini(adev);
2791 amdgpu_device_vram_scratch_fini(adev);
2792 amdgpu_ib_pool_fini(adev);
2795 r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
2796 /* XXX handle errors */
2798 DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
2799 adev->ip_blocks[i].version->funcs->name, r);
2801 adev->ip_blocks[i].status.sw = false;
2802 adev->ip_blocks[i].status.valid = false;
2805 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2806 if (!adev->ip_blocks[i].status.late_initialized)
2808 if (adev->ip_blocks[i].version->funcs->late_fini)
2809 adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
2810 adev->ip_blocks[i].status.late_initialized = false;
2813 amdgpu_ras_fini(adev);
2819 * amdgpu_device_delayed_init_work_handler - work handler for IB tests
2821 * @work: work_struct.
2823 static void amdgpu_device_delayed_init_work_handler(struct work_struct *work)
2825 struct amdgpu_device *adev =
2826 container_of(work, struct amdgpu_device, delayed_init_work.work);
2829 r = amdgpu_ib_ring_tests(adev);
2831 DRM_ERROR("ib ring test failed (%d).\n", r);
2834 static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work)
2836 struct amdgpu_device *adev =
2837 container_of(work, struct amdgpu_device, gfx.gfx_off_delay_work.work);
2839 WARN_ON_ONCE(adev->gfx.gfx_off_state);
2840 WARN_ON_ONCE(adev->gfx.gfx_off_req_count);
2842 if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, true))
2843 adev->gfx.gfx_off_state = true;
2847 * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1)
2849 * @adev: amdgpu_device pointer
2851 * Main suspend function for hardware IPs. The list of all the hardware
2852 * IPs that make up the asic is walked, clockgating is disabled and the
2853 * suspend callbacks are run. suspend puts the hardware and software state
2854 * in each IP into a state suitable for suspend.
2855 * Returns 0 on success, negative error code on failure.
2857 static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev)
2861 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
2862 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
2864 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2865 if (!adev->ip_blocks[i].status.valid)
2868 /* displays are handled separately */
2869 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_DCE)
2872 /* XXX handle errors */
2873 r = adev->ip_blocks[i].version->funcs->suspend(adev);
2874 /* XXX handle errors */
2876 DRM_ERROR("suspend of IP block <%s> failed %d\n",
2877 adev->ip_blocks[i].version->funcs->name, r);
2881 adev->ip_blocks[i].status.hw = false;
2888 * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2)
2890 * @adev: amdgpu_device pointer
2892 * Main suspend function for hardware IPs. The list of all the hardware
2893 * IPs that make up the asic is walked, clockgating is disabled and the
2894 * suspend callbacks are run. suspend puts the hardware and software state
2895 * in each IP into a state suitable for suspend.
2896 * Returns 0 on success, negative error code on failure.
2898 static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev)
2903 amdgpu_dpm_gfx_state_change(adev, sGpuChangeState_D3Entry);
2905 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2906 if (!adev->ip_blocks[i].status.valid)
2908 /* displays are handled in phase1 */
2909 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE)
2911 /* PSP lost connection when err_event_athub occurs */
2912 if (amdgpu_ras_intr_triggered() &&
2913 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
2914 adev->ip_blocks[i].status.hw = false;
2918 /* skip unnecessary suspend if we do not initialize them yet */
2919 if (adev->gmc.xgmi.pending_reset &&
2920 !(adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2921 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC ||
2922 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2923 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH)) {
2924 adev->ip_blocks[i].status.hw = false;
2928 /* skip suspend of gfx and psp for S0ix
2929 * gfx is in gfxoff state, so on resume it will exit gfxoff just
2930 * like at runtime. PSP is also part of the always on hardware
2931 * so no need to suspend it.
2933 if (adev->in_s0ix &&
2934 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP ||
2935 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX))
2938 /* XXX handle errors */
2939 r = adev->ip_blocks[i].version->funcs->suspend(adev);
2940 /* XXX handle errors */
2942 DRM_ERROR("suspend of IP block <%s> failed %d\n",
2943 adev->ip_blocks[i].version->funcs->name, r);
2945 adev->ip_blocks[i].status.hw = false;
2946 /* handle putting the SMC in the appropriate state */
2947 if(!amdgpu_sriov_vf(adev)){
2948 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
2949 r = amdgpu_dpm_set_mp1_state(adev, adev->mp1_state);
2951 DRM_ERROR("SMC failed to set mp1 state %d, %d\n",
2952 adev->mp1_state, r);
2963 * amdgpu_device_ip_suspend - run suspend for hardware IPs
2965 * @adev: amdgpu_device pointer
2967 * Main suspend function for hardware IPs. The list of all the hardware
2968 * IPs that make up the asic is walked, clockgating is disabled and the
2969 * suspend callbacks are run. suspend puts the hardware and software state
2970 * in each IP into a state suitable for suspend.
2971 * Returns 0 on success, negative error code on failure.
2973 int amdgpu_device_ip_suspend(struct amdgpu_device *adev)
2977 if (amdgpu_sriov_vf(adev)) {
2978 amdgpu_virt_fini_data_exchange(adev);
2979 amdgpu_virt_request_full_gpu(adev, false);
2982 r = amdgpu_device_ip_suspend_phase1(adev);
2985 r = amdgpu_device_ip_suspend_phase2(adev);
2987 if (amdgpu_sriov_vf(adev))
2988 amdgpu_virt_release_full_gpu(adev, false);
2993 static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
2997 static enum amd_ip_block_type ip_order[] = {
2998 AMD_IP_BLOCK_TYPE_GMC,
2999 AMD_IP_BLOCK_TYPE_COMMON,
3000 AMD_IP_BLOCK_TYPE_PSP,
3001 AMD_IP_BLOCK_TYPE_IH,
3004 for (i = 0; i < adev->num_ip_blocks; i++) {
3006 struct amdgpu_ip_block *block;
3008 block = &adev->ip_blocks[i];
3009 block->status.hw = false;
3011 for (j = 0; j < ARRAY_SIZE(ip_order); j++) {
3013 if (block->version->type != ip_order[j] ||
3014 !block->status.valid)
3017 r = block->version->funcs->hw_init(adev);
3018 DRM_INFO("RE-INIT-early: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
3021 block->status.hw = true;
3028 static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
3032 static enum amd_ip_block_type ip_order[] = {
3033 AMD_IP_BLOCK_TYPE_SMC,
3034 AMD_IP_BLOCK_TYPE_DCE,
3035 AMD_IP_BLOCK_TYPE_GFX,
3036 AMD_IP_BLOCK_TYPE_SDMA,
3037 AMD_IP_BLOCK_TYPE_UVD,
3038 AMD_IP_BLOCK_TYPE_VCE,
3039 AMD_IP_BLOCK_TYPE_VCN
3042 for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
3044 struct amdgpu_ip_block *block;
3046 for (j = 0; j < adev->num_ip_blocks; j++) {
3047 block = &adev->ip_blocks[j];
3049 if (block->version->type != ip_order[i] ||
3050 !block->status.valid ||
3054 if (block->version->type == AMD_IP_BLOCK_TYPE_SMC)
3055 r = block->version->funcs->resume(adev);
3057 r = block->version->funcs->hw_init(adev);
3059 DRM_INFO("RE-INIT-late: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
3062 block->status.hw = true;
3070 * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs
3072 * @adev: amdgpu_device pointer
3074 * First resume function for hardware IPs. The list of all the hardware
3075 * IPs that make up the asic is walked and the resume callbacks are run for
3076 * COMMON, GMC, and IH. resume puts the hardware into a functional state
3077 * after a suspend and updates the software state as necessary. This
3078 * function is also used for restoring the GPU after a GPU reset.
3079 * Returns 0 on success, negative error code on failure.
3081 static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)
3085 for (i = 0; i < adev->num_ip_blocks; i++) {
3086 if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
3088 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
3089 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
3090 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
3092 r = adev->ip_blocks[i].version->funcs->resume(adev);
3094 DRM_ERROR("resume of IP block <%s> failed %d\n",
3095 adev->ip_blocks[i].version->funcs->name, r);
3098 adev->ip_blocks[i].status.hw = true;
3106 * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs
3108 * @adev: amdgpu_device pointer
3110 * First resume function for hardware IPs. The list of all the hardware
3111 * IPs that make up the asic is walked and the resume callbacks are run for
3112 * all blocks except COMMON, GMC, and IH. resume puts the hardware into a
3113 * functional state after a suspend and updates the software state as
3114 * necessary. This function is also used for restoring the GPU after a GPU
3116 * Returns 0 on success, negative error code on failure.
3118 static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)
3122 for (i = 0; i < adev->num_ip_blocks; i++) {
3123 if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
3125 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
3126 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
3127 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH ||
3128 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)
3130 r = adev->ip_blocks[i].version->funcs->resume(adev);
3132 DRM_ERROR("resume of IP block <%s> failed %d\n",
3133 adev->ip_blocks[i].version->funcs->name, r);
3136 adev->ip_blocks[i].status.hw = true;
3143 * amdgpu_device_ip_resume - run resume for hardware IPs
3145 * @adev: amdgpu_device pointer
3147 * Main resume function for hardware IPs. The hardware IPs
3148 * are split into two resume functions because they are
3149 * are also used in in recovering from a GPU reset and some additional
3150 * steps need to be take between them. In this case (S3/S4) they are
3152 * Returns 0 on success, negative error code on failure.
3154 static int amdgpu_device_ip_resume(struct amdgpu_device *adev)
3158 r = amdgpu_amdkfd_resume_iommu(adev);
3162 r = amdgpu_device_ip_resume_phase1(adev);
3166 r = amdgpu_device_fw_loading(adev);
3170 r = amdgpu_device_ip_resume_phase2(adev);
3176 * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV
3178 * @adev: amdgpu_device pointer
3180 * Query the VBIOS data tables to determine if the board supports SR-IOV.
3182 static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
3184 if (amdgpu_sriov_vf(adev)) {
3185 if (adev->is_atom_fw) {
3186 if (amdgpu_atomfirmware_gpu_virtualization_supported(adev))
3187 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
3189 if (amdgpu_atombios_has_gpu_virtualization_table(adev))
3190 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
3193 if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
3194 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
3199 * amdgpu_device_asic_has_dc_support - determine if DC supports the asic
3201 * @asic_type: AMD asic type
3203 * Check if there is DC (new modesetting infrastructre) support for an asic.
3204 * returns true if DC has support, false if not.
3206 bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
3208 switch (asic_type) {
3209 #ifdef CONFIG_DRM_AMDGPU_SI
3213 /* chips with no display hardware */
3215 #if defined(CONFIG_DRM_AMD_DC)
3221 * We have systems in the wild with these ASICs that require
3222 * LVDS and VGA support which is not supported with DC.
3224 * Fallback to the non-DC driver here by default so as not to
3225 * cause regressions.
3227 #if defined(CONFIG_DRM_AMD_DC_SI)
3228 return amdgpu_dc > 0;
3237 * We have systems in the wild with these ASICs that require
3238 * LVDS and VGA support which is not supported with DC.
3240 * Fallback to the non-DC driver here by default so as not to
3241 * cause regressions.
3243 return amdgpu_dc > 0;
3247 case CHIP_POLARIS10:
3248 case CHIP_POLARIS11:
3249 case CHIP_POLARIS12:
3256 #if defined(CONFIG_DRM_AMD_DC_DCN)
3262 case CHIP_CYAN_SKILLFISH:
3263 case CHIP_SIENNA_CICHLID:
3264 case CHIP_NAVY_FLOUNDER:
3265 case CHIP_DIMGREY_CAVEFISH:
3266 case CHIP_BEIGE_GOBY:
3268 case CHIP_YELLOW_CARP:
3271 return amdgpu_dc != 0;
3275 DRM_INFO_ONCE("Display Core has been requested via kernel parameter "
3276 "but isn't supported by ASIC, ignoring\n");
3283 * amdgpu_device_has_dc_support - check if dc is supported
3285 * @adev: amdgpu_device pointer
3287 * Returns true for supported, false for not supported
3289 bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
3291 if (amdgpu_sriov_vf(adev) ||
3292 adev->enable_virtual_display ||
3293 (adev->harvest_ip_mask & AMD_HARVEST_IP_DMU_MASK))
3296 return amdgpu_device_asic_has_dc_support(adev->asic_type);
3299 static void amdgpu_device_xgmi_reset_func(struct work_struct *__work)
3301 struct amdgpu_device *adev =
3302 container_of(__work, struct amdgpu_device, xgmi_reset_work);
3303 struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev);
3305 /* It's a bug to not have a hive within this function */
3310 * Use task barrier to synchronize all xgmi reset works across the
3311 * hive. task_barrier_enter and task_barrier_exit will block
3312 * until all the threads running the xgmi reset works reach
3313 * those points. task_barrier_full will do both blocks.
3315 if (amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) {
3317 task_barrier_enter(&hive->tb);
3318 adev->asic_reset_res = amdgpu_device_baco_enter(adev_to_drm(adev));
3320 if (adev->asic_reset_res)
3323 task_barrier_exit(&hive->tb);
3324 adev->asic_reset_res = amdgpu_device_baco_exit(adev_to_drm(adev));
3326 if (adev->asic_reset_res)
3329 if (adev->mmhub.ras && adev->mmhub.ras->ras_block.hw_ops &&
3330 adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count)
3331 adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count(adev);
3334 task_barrier_full(&hive->tb);
3335 adev->asic_reset_res = amdgpu_asic_reset(adev);
3339 if (adev->asic_reset_res)
3340 DRM_WARN("ASIC reset failed with error, %d for drm dev, %s",
3341 adev->asic_reset_res, adev_to_drm(adev)->unique);
3342 amdgpu_put_xgmi_hive(hive);
3345 static int amdgpu_device_get_job_timeout_settings(struct amdgpu_device *adev)
3347 char *input = amdgpu_lockup_timeout;
3348 char *timeout_setting = NULL;
3354 * By default timeout for non compute jobs is 10000
3355 * and 60000 for compute jobs.
3356 * In SR-IOV or passthrough mode, timeout for compute
3357 * jobs are 60000 by default.
3359 adev->gfx_timeout = msecs_to_jiffies(10000);
3360 adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout;
3361 if (amdgpu_sriov_vf(adev))
3362 adev->compute_timeout = amdgpu_sriov_is_pp_one_vf(adev) ?
3363 msecs_to_jiffies(60000) : msecs_to_jiffies(10000);
3365 adev->compute_timeout = msecs_to_jiffies(60000);
3367 if (strnlen(input, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) {
3368 while ((timeout_setting = strsep(&input, ",")) &&
3369 strnlen(timeout_setting, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) {
3370 ret = kstrtol(timeout_setting, 0, &timeout);
3377 } else if (timeout < 0) {
3378 timeout = MAX_SCHEDULE_TIMEOUT;
3379 dev_warn(adev->dev, "lockup timeout disabled");
3380 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
3382 timeout = msecs_to_jiffies(timeout);
3387 adev->gfx_timeout = timeout;
3390 adev->compute_timeout = timeout;
3393 adev->sdma_timeout = timeout;
3396 adev->video_timeout = timeout;
3403 * There is only one value specified and
3404 * it should apply to all non-compute jobs.
3407 adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout;
3408 if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev))
3409 adev->compute_timeout = adev->gfx_timeout;
3417 * amdgpu_device_check_iommu_direct_map - check if RAM direct mapped to GPU
3419 * @adev: amdgpu_device pointer
3421 * RAM direct mapped to GPU if IOMMU is not enabled or is pass through mode
3423 static void amdgpu_device_check_iommu_direct_map(struct amdgpu_device *adev)
3425 struct iommu_domain *domain;
3427 domain = iommu_get_domain_for_dev(adev->dev);
3428 if (!domain || domain->type == IOMMU_DOMAIN_IDENTITY)
3429 adev->ram_is_direct_mapped = true;
3432 static const struct attribute *amdgpu_dev_attributes[] = {
3433 &dev_attr_product_name.attr,
3434 &dev_attr_product_number.attr,
3435 &dev_attr_serial_number.attr,
3436 &dev_attr_pcie_replay_count.attr,
3441 * amdgpu_device_init - initialize the driver
3443 * @adev: amdgpu_device pointer
3444 * @flags: driver flags
3446 * Initializes the driver info and hw (all asics).
3447 * Returns 0 for success or an error on failure.
3448 * Called at driver startup.
3450 int amdgpu_device_init(struct amdgpu_device *adev,
3453 struct drm_device *ddev = adev_to_drm(adev);
3454 struct pci_dev *pdev = adev->pdev;
3459 adev->shutdown = false;
3460 adev->flags = flags;
3462 if (amdgpu_force_asic_type >= 0 && amdgpu_force_asic_type < CHIP_LAST)
3463 adev->asic_type = amdgpu_force_asic_type;
3465 adev->asic_type = flags & AMD_ASIC_MASK;
3467 adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
3468 if (amdgpu_emu_mode == 1)
3469 adev->usec_timeout *= 10;
3470 adev->gmc.gart_size = 512 * 1024 * 1024;
3471 adev->accel_working = false;
3472 adev->num_rings = 0;
3473 adev->mman.buffer_funcs = NULL;
3474 adev->mman.buffer_funcs_ring = NULL;
3475 adev->vm_manager.vm_pte_funcs = NULL;
3476 adev->vm_manager.vm_pte_num_scheds = 0;
3477 adev->gmc.gmc_funcs = NULL;
3478 adev->harvest_ip_mask = 0x0;
3479 adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
3480 bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
3482 adev->smc_rreg = &amdgpu_invalid_rreg;
3483 adev->smc_wreg = &amdgpu_invalid_wreg;
3484 adev->pcie_rreg = &amdgpu_invalid_rreg;
3485 adev->pcie_wreg = &amdgpu_invalid_wreg;
3486 adev->pciep_rreg = &amdgpu_invalid_rreg;
3487 adev->pciep_wreg = &amdgpu_invalid_wreg;
3488 adev->pcie_rreg64 = &amdgpu_invalid_rreg64;
3489 adev->pcie_wreg64 = &amdgpu_invalid_wreg64;
3490 adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
3491 adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
3492 adev->didt_rreg = &amdgpu_invalid_rreg;
3493 adev->didt_wreg = &amdgpu_invalid_wreg;
3494 adev->gc_cac_rreg = &amdgpu_invalid_rreg;
3495 adev->gc_cac_wreg = &amdgpu_invalid_wreg;
3496 adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
3497 adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
3499 DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
3500 amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
3501 pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
3503 /* mutex initialization are all done here so we
3504 * can recall function without having locking issues */
3505 mutex_init(&adev->firmware.mutex);
3506 mutex_init(&adev->pm.mutex);
3507 mutex_init(&adev->gfx.gpu_clock_mutex);
3508 mutex_init(&adev->srbm_mutex);
3509 mutex_init(&adev->gfx.pipe_reserve_mutex);
3510 mutex_init(&adev->gfx.gfx_off_mutex);
3511 mutex_init(&adev->grbm_idx_mutex);
3512 mutex_init(&adev->mn_lock);
3513 mutex_init(&adev->virt.vf_errors.lock);
3514 hash_init(adev->mn_hash);
3515 atomic_set(&adev->in_gpu_reset, 0);
3516 init_rwsem(&adev->reset_sem);
3517 mutex_init(&adev->psp.mutex);
3518 mutex_init(&adev->notifier_lock);
3519 mutex_init(&adev->pm.stable_pstate_ctx_lock);
3521 amdgpu_device_init_apu_flags(adev);
3523 r = amdgpu_device_check_arguments(adev);
3527 spin_lock_init(&adev->mmio_idx_lock);
3528 spin_lock_init(&adev->smc_idx_lock);
3529 spin_lock_init(&adev->pcie_idx_lock);
3530 spin_lock_init(&adev->uvd_ctx_idx_lock);
3531 spin_lock_init(&adev->didt_idx_lock);
3532 spin_lock_init(&adev->gc_cac_idx_lock);
3533 spin_lock_init(&adev->se_cac_idx_lock);
3534 spin_lock_init(&adev->audio_endpt_idx_lock);
3535 spin_lock_init(&adev->mm_stats.lock);
3537 INIT_LIST_HEAD(&adev->shadow_list);
3538 mutex_init(&adev->shadow_list_lock);
3540 INIT_LIST_HEAD(&adev->reset_list);
3542 INIT_LIST_HEAD(&adev->ras_list);
3544 INIT_DELAYED_WORK(&adev->delayed_init_work,
3545 amdgpu_device_delayed_init_work_handler);
3546 INIT_DELAYED_WORK(&adev->gfx.gfx_off_delay_work,
3547 amdgpu_device_delay_enable_gfx_off);
3549 INIT_WORK(&adev->xgmi_reset_work, amdgpu_device_xgmi_reset_func);
3551 adev->gfx.gfx_off_req_count = 1;
3552 adev->pm.ac_power = power_supply_is_system_supplied() > 0;
3554 atomic_set(&adev->throttling_logging_enabled, 1);
3556 * If throttling continues, logging will be performed every minute
3557 * to avoid log flooding. "-1" is subtracted since the thermal
3558 * throttling interrupt comes every second. Thus, the total logging
3559 * interval is 59 seconds(retelimited printk interval) + 1(waiting
3560 * for throttling interrupt) = 60 seconds.
3562 ratelimit_state_init(&adev->throttling_logging_rs, (60 - 1) * HZ, 1);
3563 ratelimit_set_flags(&adev->throttling_logging_rs, RATELIMIT_MSG_ON_RELEASE);
3565 /* Registers mapping */
3566 /* TODO: block userspace mapping of io register */
3567 if (adev->asic_type >= CHIP_BONAIRE) {
3568 adev->rmmio_base = pci_resource_start(adev->pdev, 5);
3569 adev->rmmio_size = pci_resource_len(adev->pdev, 5);
3571 adev->rmmio_base = pci_resource_start(adev->pdev, 2);
3572 adev->rmmio_size = pci_resource_len(adev->pdev, 2);
3575 for (i = 0; i < AMD_IP_BLOCK_TYPE_NUM; i++)
3576 atomic_set(&adev->pm.pwr_state[i], POWER_STATE_UNKNOWN);
3578 adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
3579 if (adev->rmmio == NULL) {
3582 DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
3583 DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
3585 amdgpu_device_get_pcie_info(adev);
3588 DRM_INFO("MCBP is enabled\n");
3590 if (amdgpu_mes && adev->asic_type >= CHIP_NAVI10)
3591 adev->enable_mes = true;
3593 /* detect hw virtualization here */
3594 amdgpu_detect_virtualization(adev);
3596 r = amdgpu_device_get_job_timeout_settings(adev);
3598 dev_err(adev->dev, "invalid lockup_timeout parameter syntax\n");
3602 /* early init functions */
3603 r = amdgpu_device_ip_early_init(adev);
3607 /* Need to get xgmi info early to decide the reset behavior*/
3608 if (adev->gmc.xgmi.supported) {
3609 r = adev->gfxhub.funcs->get_xgmi_info(adev);
3614 /* enable PCIE atomic ops */
3615 if (amdgpu_sriov_vf(adev))
3616 adev->have_atomics_support = ((struct amd_sriov_msg_pf2vf_info *)
3617 adev->virt.fw_reserve.p_pf2vf)->pcie_atomic_ops_enabled_flags ==
3618 (PCI_EXP_DEVCAP2_ATOMIC_COMP32 | PCI_EXP_DEVCAP2_ATOMIC_COMP64);
3620 adev->have_atomics_support =
3621 !pci_enable_atomic_ops_to_root(adev->pdev,
3622 PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
3623 PCI_EXP_DEVCAP2_ATOMIC_COMP64);
3624 if (!adev->have_atomics_support)
3625 dev_info(adev->dev, "PCIE atomic ops is not supported\n");
3627 /* doorbell bar mapping and doorbell index init*/
3628 amdgpu_device_doorbell_init(adev);
3630 if (amdgpu_emu_mode == 1) {
3631 /* post the asic on emulation mode */
3632 emu_soc_asic_init(adev);
3633 goto fence_driver_init;
3636 amdgpu_reset_init(adev);
3638 /* detect if we are with an SRIOV vbios */
3639 amdgpu_device_detect_sriov_bios(adev);
3641 /* check if we need to reset the asic
3642 * E.g., driver was not cleanly unloaded previously, etc.
3644 if (!amdgpu_sriov_vf(adev) && amdgpu_asic_need_reset_on_init(adev)) {
3645 if (adev->gmc.xgmi.num_physical_nodes) {
3646 dev_info(adev->dev, "Pending hive reset.\n");
3647 adev->gmc.xgmi.pending_reset = true;
3648 /* Only need to init necessary block for SMU to handle the reset */
3649 for (i = 0; i < adev->num_ip_blocks; i++) {
3650 if (!adev->ip_blocks[i].status.valid)
3652 if (!(adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
3653 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
3654 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH ||
3655 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC)) {
3656 DRM_DEBUG("IP %s disabled for hw_init.\n",
3657 adev->ip_blocks[i].version->funcs->name);
3658 adev->ip_blocks[i].status.hw = true;
3662 r = amdgpu_asic_reset(adev);
3664 dev_err(adev->dev, "asic reset on init failed\n");
3670 pci_enable_pcie_error_reporting(adev->pdev);
3672 /* Post card if necessary */
3673 if (amdgpu_device_need_post(adev)) {
3675 dev_err(adev->dev, "no vBIOS found\n");
3679 DRM_INFO("GPU posting now...\n");
3680 r = amdgpu_device_asic_init(adev);
3682 dev_err(adev->dev, "gpu post error!\n");
3687 if (adev->is_atom_fw) {
3688 /* Initialize clocks */
3689 r = amdgpu_atomfirmware_get_clock_info(adev);
3691 dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
3692 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
3696 /* Initialize clocks */
3697 r = amdgpu_atombios_get_clock_info(adev);
3699 dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
3700 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
3703 /* init i2c buses */
3704 if (!amdgpu_device_has_dc_support(adev))
3705 amdgpu_atombios_i2c_init(adev);
3710 r = amdgpu_fence_driver_sw_init(adev);
3712 dev_err(adev->dev, "amdgpu_fence_driver_sw_init failed\n");
3713 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
3717 /* init the mode config */
3718 drm_mode_config_init(adev_to_drm(adev));
3720 r = amdgpu_device_ip_init(adev);
3722 /* failed in exclusive mode due to timeout */
3723 if (amdgpu_sriov_vf(adev) &&
3724 !amdgpu_sriov_runtime(adev) &&
3725 amdgpu_virt_mmio_blocked(adev) &&
3726 !amdgpu_virt_wait_reset(adev)) {
3727 dev_err(adev->dev, "VF exclusive mode timeout\n");
3728 /* Don't send request since VF is inactive. */
3729 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
3730 adev->virt.ops = NULL;
3732 goto release_ras_con;
3734 dev_err(adev->dev, "amdgpu_device_ip_init failed\n");
3735 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
3736 goto release_ras_con;
3739 amdgpu_fence_driver_hw_init(adev);
3742 "SE %d, SH per SE %d, CU per SH %d, active_cu_number %d\n",
3743 adev->gfx.config.max_shader_engines,
3744 adev->gfx.config.max_sh_per_se,
3745 adev->gfx.config.max_cu_per_sh,
3746 adev->gfx.cu_info.number);
3748 adev->accel_working = true;
3750 amdgpu_vm_check_compute_bug(adev);
3752 /* Initialize the buffer migration limit. */
3753 if (amdgpu_moverate >= 0)
3754 max_MBps = amdgpu_moverate;
3756 max_MBps = 8; /* Allow 8 MB/s. */
3757 /* Get a log2 for easy divisions. */
3758 adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
3760 r = amdgpu_pm_sysfs_init(adev);
3762 adev->pm_sysfs_en = false;
3763 DRM_ERROR("registering pm debugfs failed (%d).\n", r);
3765 adev->pm_sysfs_en = true;
3767 r = amdgpu_ucode_sysfs_init(adev);
3769 adev->ucode_sysfs_en = false;
3770 DRM_ERROR("Creating firmware sysfs failed (%d).\n", r);
3772 adev->ucode_sysfs_en = true;
3774 if ((amdgpu_testing & 1)) {
3775 if (adev->accel_working)
3776 amdgpu_test_moves(adev);
3778 DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n");
3780 if (amdgpu_benchmarking) {
3781 if (adev->accel_working)
3782 amdgpu_benchmark(adev, amdgpu_benchmarking);
3784 DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n");
3788 * Register gpu instance before amdgpu_device_enable_mgpu_fan_boost.
3789 * Otherwise the mgpu fan boost feature will be skipped due to the
3790 * gpu instance is counted less.
3792 amdgpu_register_gpu_instance(adev);
3794 /* enable clockgating, etc. after ib tests, etc. since some blocks require
3795 * explicit gating rather than handling it automatically.
3797 if (!adev->gmc.xgmi.pending_reset) {
3798 r = amdgpu_device_ip_late_init(adev);
3800 dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n");
3801 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
3802 goto release_ras_con;
3805 amdgpu_ras_resume(adev);
3806 queue_delayed_work(system_wq, &adev->delayed_init_work,
3807 msecs_to_jiffies(AMDGPU_RESUME_MS));
3810 if (amdgpu_sriov_vf(adev))
3811 flush_delayed_work(&adev->delayed_init_work);
3813 r = sysfs_create_files(&adev->dev->kobj, amdgpu_dev_attributes);
3815 dev_err(adev->dev, "Could not create amdgpu device attr\n");
3817 if (IS_ENABLED(CONFIG_PERF_EVENTS))
3818 r = amdgpu_pmu_init(adev);
3820 dev_err(adev->dev, "amdgpu_pmu_init failed\n");
3822 /* Have stored pci confspace at hand for restore in sudden PCI error */
3823 if (amdgpu_device_cache_pci_state(adev->pdev))
3824 pci_restore_state(pdev);
3826 /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
3827 /* this will fail for cards that aren't VGA class devices, just
3829 if ((adev->pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA)
3830 vga_client_register(adev->pdev, amdgpu_device_vga_set_decode);
3832 if (amdgpu_device_supports_px(ddev)) {
3834 vga_switcheroo_register_client(adev->pdev,
3835 &amdgpu_switcheroo_ops, px);
3836 vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
3839 if (adev->gmc.xgmi.pending_reset)
3840 queue_delayed_work(system_wq, &mgpu_info.delayed_reset_work,
3841 msecs_to_jiffies(AMDGPU_RESUME_MS));
3843 amdgpu_device_check_iommu_direct_map(adev);
3848 amdgpu_release_ras_context(adev);
3851 amdgpu_vf_error_trans_all(adev);
3856 static void amdgpu_device_unmap_mmio(struct amdgpu_device *adev)
3859 /* Clear all CPU mappings pointing to this device */
3860 unmap_mapping_range(adev->ddev.anon_inode->i_mapping, 0, 0, 1);
3862 /* Unmap all mapped bars - Doorbell, registers and VRAM */
3863 amdgpu_device_doorbell_fini(adev);
3865 iounmap(adev->rmmio);
3867 if (adev->mman.aper_base_kaddr)
3868 iounmap(adev->mman.aper_base_kaddr);
3869 adev->mman.aper_base_kaddr = NULL;
3871 /* Memory manager related */
3872 if (!adev->gmc.xgmi.connected_to_cpu) {
3873 arch_phys_wc_del(adev->gmc.vram_mtrr);
3874 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
3879 * amdgpu_device_fini_hw - tear down the driver
3881 * @adev: amdgpu_device pointer
3883 * Tear down the driver info (all asics).
3884 * Called at driver shutdown.
3886 void amdgpu_device_fini_hw(struct amdgpu_device *adev)
3888 dev_info(adev->dev, "amdgpu: finishing device.\n");
3889 flush_delayed_work(&adev->delayed_init_work);
3890 if (adev->mman.initialized) {
3891 flush_delayed_work(&adev->mman.bdev.wq);
3892 ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
3894 adev->shutdown = true;
3896 /* make sure IB test finished before entering exclusive mode
3897 * to avoid preemption on IB test
3899 if (amdgpu_sriov_vf(adev)) {
3900 amdgpu_virt_request_full_gpu(adev, false);
3901 amdgpu_virt_fini_data_exchange(adev);
3904 /* disable all interrupts */
3905 amdgpu_irq_disable_all(adev);
3906 if (adev->mode_info.mode_config_initialized){
3907 if (!drm_drv_uses_atomic_modeset(adev_to_drm(adev)))
3908 drm_helper_force_disable_all(adev_to_drm(adev));
3910 drm_atomic_helper_shutdown(adev_to_drm(adev));
3912 amdgpu_fence_driver_hw_fini(adev);
3914 if (adev->pm_sysfs_en)
3915 amdgpu_pm_sysfs_fini(adev);
3916 if (adev->ucode_sysfs_en)
3917 amdgpu_ucode_sysfs_fini(adev);
3918 sysfs_remove_files(&adev->dev->kobj, amdgpu_dev_attributes);
3920 /* disable ras feature must before hw fini */
3921 amdgpu_ras_pre_fini(adev);
3923 amdgpu_device_ip_fini_early(adev);
3925 amdgpu_irq_fini_hw(adev);
3927 if (adev->mman.initialized)
3928 ttm_device_clear_dma_mappings(&adev->mman.bdev);
3930 amdgpu_gart_dummy_page_fini(adev);
3932 if (drm_dev_is_unplugged(adev_to_drm(adev)))
3933 amdgpu_device_unmap_mmio(adev);
3937 void amdgpu_device_fini_sw(struct amdgpu_device *adev)
3941 amdgpu_fence_driver_sw_fini(adev);
3942 amdgpu_device_ip_fini(adev);
3943 release_firmware(adev->firmware.gpu_info_fw);
3944 adev->firmware.gpu_info_fw = NULL;
3945 adev->accel_working = false;
3947 amdgpu_reset_fini(adev);
3949 /* free i2c buses */
3950 if (!amdgpu_device_has_dc_support(adev))
3951 amdgpu_i2c_fini(adev);
3953 if (amdgpu_emu_mode != 1)
3954 amdgpu_atombios_fini(adev);
3958 if (amdgpu_device_supports_px(adev_to_drm(adev))) {
3959 vga_switcheroo_unregister_client(adev->pdev);
3960 vga_switcheroo_fini_domain_pm_ops(adev->dev);
3962 if ((adev->pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA)
3963 vga_client_unregister(adev->pdev);
3965 if (drm_dev_enter(adev_to_drm(adev), &idx)) {
3967 iounmap(adev->rmmio);
3969 amdgpu_device_doorbell_fini(adev);
3973 if (IS_ENABLED(CONFIG_PERF_EVENTS))
3974 amdgpu_pmu_fini(adev);
3975 if (adev->mman.discovery_bin)
3976 amdgpu_discovery_fini(adev);
3978 kfree(adev->pci_state);
3983 * amdgpu_device_evict_resources - evict device resources
3984 * @adev: amdgpu device object
3986 * Evicts all ttm device resources(vram BOs, gart table) from the lru list
3987 * of the vram memory type. Mainly used for evicting device resources
3991 static void amdgpu_device_evict_resources(struct amdgpu_device *adev)
3993 /* No need to evict vram on APUs for suspend to ram or s2idle */
3994 if ((adev->in_s3 || adev->in_s0ix) && (adev->flags & AMD_IS_APU))
3997 if (amdgpu_ttm_evict_resources(adev, TTM_PL_VRAM))
3998 DRM_WARN("evicting device resources failed\n");
4006 * amdgpu_device_suspend - initiate device suspend
4008 * @dev: drm dev pointer
4009 * @fbcon : notify the fbdev of suspend
4011 * Puts the hw in the suspend state (all asics).
4012 * Returns 0 for success or an error on failure.
4013 * Called at driver suspend.
4015 int amdgpu_device_suspend(struct drm_device *dev, bool fbcon)
4017 struct amdgpu_device *adev = drm_to_adev(dev);
4019 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
4022 adev->in_suspend = true;
4024 if (amdgpu_acpi_smart_shift_update(dev, AMDGPU_SS_DEV_D3))
4025 DRM_WARN("smart shift update failed\n");
4027 drm_kms_helper_poll_disable(dev);
4030 drm_fb_helper_set_suspend_unlocked(adev_to_drm(adev)->fb_helper, true);
4032 cancel_delayed_work_sync(&adev->delayed_init_work);
4034 amdgpu_ras_suspend(adev);
4036 amdgpu_device_ip_suspend_phase1(adev);
4039 amdgpu_amdkfd_suspend(adev, adev->in_runpm);
4041 amdgpu_device_evict_resources(adev);
4043 amdgpu_fence_driver_hw_fini(adev);
4045 amdgpu_device_ip_suspend_phase2(adev);
4051 * amdgpu_device_resume - initiate device resume
4053 * @dev: drm dev pointer
4054 * @fbcon : notify the fbdev of resume
4056 * Bring the hw back to operating state (all asics).
4057 * Returns 0 for success or an error on failure.
4058 * Called at driver resume.
4060 int amdgpu_device_resume(struct drm_device *dev, bool fbcon)
4062 struct amdgpu_device *adev = drm_to_adev(dev);
4065 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
4069 amdgpu_dpm_gfx_state_change(adev, sGpuChangeState_D0Entry);
4072 if (amdgpu_device_need_post(adev)) {
4073 r = amdgpu_device_asic_init(adev);
4075 dev_err(adev->dev, "amdgpu asic init failed\n");
4078 r = amdgpu_device_ip_resume(adev);
4080 dev_err(adev->dev, "amdgpu_device_ip_resume failed (%d).\n", r);
4083 amdgpu_fence_driver_hw_init(adev);
4085 r = amdgpu_device_ip_late_init(adev);
4089 queue_delayed_work(system_wq, &adev->delayed_init_work,
4090 msecs_to_jiffies(AMDGPU_RESUME_MS));
4092 if (!adev->in_s0ix) {
4093 r = amdgpu_amdkfd_resume(adev, adev->in_runpm);
4098 /* Make sure IB tests flushed */
4099 flush_delayed_work(&adev->delayed_init_work);
4102 drm_fb_helper_set_suspend_unlocked(adev_to_drm(adev)->fb_helper, false);
4104 drm_kms_helper_poll_enable(dev);
4106 amdgpu_ras_resume(adev);
4109 * Most of the connector probing functions try to acquire runtime pm
4110 * refs to ensure that the GPU is powered on when connector polling is
4111 * performed. Since we're calling this from a runtime PM callback,
4112 * trying to acquire rpm refs will cause us to deadlock.
4114 * Since we're guaranteed to be holding the rpm lock, it's safe to
4115 * temporarily disable the rpm helpers so this doesn't deadlock us.
4118 dev->dev->power.disable_depth++;
4120 if (!amdgpu_device_has_dc_support(adev))
4121 drm_helper_hpd_irq_event(dev);
4123 drm_kms_helper_hotplug_event(dev);
4125 dev->dev->power.disable_depth--;
4127 adev->in_suspend = false;
4129 if (amdgpu_acpi_smart_shift_update(dev, AMDGPU_SS_DEV_D0))
4130 DRM_WARN("smart shift update failed\n");
4136 * amdgpu_device_ip_check_soft_reset - did soft reset succeed
4138 * @adev: amdgpu_device pointer
4140 * The list of all the hardware IPs that make up the asic is walked and
4141 * the check_soft_reset callbacks are run. check_soft_reset determines
4142 * if the asic is still hung or not.
4143 * Returns true if any of the IPs are still in a hung state, false if not.
4145 static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)
4148 bool asic_hang = false;
4150 if (amdgpu_sriov_vf(adev))
4153 if (amdgpu_asic_need_full_reset(adev))
4156 for (i = 0; i < adev->num_ip_blocks; i++) {
4157 if (!adev->ip_blocks[i].status.valid)
4159 if (adev->ip_blocks[i].version->funcs->check_soft_reset)
4160 adev->ip_blocks[i].status.hang =
4161 adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
4162 if (adev->ip_blocks[i].status.hang) {
4163 dev_info(adev->dev, "IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
4171 * amdgpu_device_ip_pre_soft_reset - prepare for soft reset
4173 * @adev: amdgpu_device pointer
4175 * The list of all the hardware IPs that make up the asic is walked and the
4176 * pre_soft_reset callbacks are run if the block is hung. pre_soft_reset
4177 * handles any IP specific hardware or software state changes that are
4178 * necessary for a soft reset to succeed.
4179 * Returns 0 on success, negative error code on failure.
4181 static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)
4185 for (i = 0; i < adev->num_ip_blocks; i++) {
4186 if (!adev->ip_blocks[i].status.valid)
4188 if (adev->ip_blocks[i].status.hang &&
4189 adev->ip_blocks[i].version->funcs->pre_soft_reset) {
4190 r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
4200 * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed
4202 * @adev: amdgpu_device pointer
4204 * Some hardware IPs cannot be soft reset. If they are hung, a full gpu
4205 * reset is necessary to recover.
4206 * Returns true if a full asic reset is required, false if not.
4208 static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)
4212 if (amdgpu_asic_need_full_reset(adev))
4215 for (i = 0; i < adev->num_ip_blocks; i++) {
4216 if (!adev->ip_blocks[i].status.valid)
4218 if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
4219 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
4220 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
4221 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
4222 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
4223 if (adev->ip_blocks[i].status.hang) {
4224 dev_info(adev->dev, "Some block need full reset!\n");
4233 * amdgpu_device_ip_soft_reset - do a soft reset
4235 * @adev: amdgpu_device pointer
4237 * The list of all the hardware IPs that make up the asic is walked and the
4238 * soft_reset callbacks are run if the block is hung. soft_reset handles any
4239 * IP specific hardware or software state changes that are necessary to soft
4241 * Returns 0 on success, negative error code on failure.
4243 static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)
4247 for (i = 0; i < adev->num_ip_blocks; i++) {
4248 if (!adev->ip_blocks[i].status.valid)
4250 if (adev->ip_blocks[i].status.hang &&
4251 adev->ip_blocks[i].version->funcs->soft_reset) {
4252 r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
4262 * amdgpu_device_ip_post_soft_reset - clean up from soft reset
4264 * @adev: amdgpu_device pointer
4266 * The list of all the hardware IPs that make up the asic is walked and the
4267 * post_soft_reset callbacks are run if the asic was hung. post_soft_reset
4268 * handles any IP specific hardware or software state changes that are
4269 * necessary after the IP has been soft reset.
4270 * Returns 0 on success, negative error code on failure.
4272 static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)
4276 for (i = 0; i < adev->num_ip_blocks; i++) {
4277 if (!adev->ip_blocks[i].status.valid)
4279 if (adev->ip_blocks[i].status.hang &&
4280 adev->ip_blocks[i].version->funcs->post_soft_reset)
4281 r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
4290 * amdgpu_device_recover_vram - Recover some VRAM contents
4292 * @adev: amdgpu_device pointer
4294 * Restores the contents of VRAM buffers from the shadows in GTT. Used to
4295 * restore things like GPUVM page tables after a GPU reset where
4296 * the contents of VRAM might be lost.
4299 * 0 on success, negative error code on failure.
4301 static int amdgpu_device_recover_vram(struct amdgpu_device *adev)
4303 struct dma_fence *fence = NULL, *next = NULL;
4304 struct amdgpu_bo *shadow;
4305 struct amdgpu_bo_vm *vmbo;
4308 if (amdgpu_sriov_runtime(adev))
4309 tmo = msecs_to_jiffies(8000);
4311 tmo = msecs_to_jiffies(100);
4313 dev_info(adev->dev, "recover vram bo from shadow start\n");
4314 mutex_lock(&adev->shadow_list_lock);
4315 list_for_each_entry(vmbo, &adev->shadow_list, shadow_list) {
4317 /* No need to recover an evicted BO */
4318 if (shadow->tbo.resource->mem_type != TTM_PL_TT ||
4319 shadow->tbo.resource->start == AMDGPU_BO_INVALID_OFFSET ||
4320 shadow->parent->tbo.resource->mem_type != TTM_PL_VRAM)
4323 r = amdgpu_bo_restore_shadow(shadow, &next);
4328 tmo = dma_fence_wait_timeout(fence, false, tmo);
4329 dma_fence_put(fence);
4334 } else if (tmo < 0) {
4342 mutex_unlock(&adev->shadow_list_lock);
4345 tmo = dma_fence_wait_timeout(fence, false, tmo);
4346 dma_fence_put(fence);
4348 if (r < 0 || tmo <= 0) {
4349 dev_err(adev->dev, "recover vram bo from shadow failed, r is %ld, tmo is %ld\n", r, tmo);
4353 dev_info(adev->dev, "recover vram bo from shadow done\n");
4359 * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
4361 * @adev: amdgpu_device pointer
4362 * @from_hypervisor: request from hypervisor
4364 * do VF FLR and reinitialize Asic
4365 * return 0 means succeeded otherwise failed
4367 static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
4368 bool from_hypervisor)
4371 struct amdgpu_hive_info *hive = NULL;
4372 int retry_limit = 0;
4375 amdgpu_amdkfd_pre_reset(adev);
4377 amdgpu_amdkfd_pre_reset(adev);
4379 if (from_hypervisor)
4380 r = amdgpu_virt_request_full_gpu(adev, true);
4382 r = amdgpu_virt_reset_gpu(adev);
4386 /* Resume IP prior to SMC */
4387 r = amdgpu_device_ip_reinit_early_sriov(adev);
4391 amdgpu_virt_init_data_exchange(adev);
4393 r = amdgpu_device_fw_loading(adev);
4397 /* now we are okay to resume SMC/CP/SDMA */
4398 r = amdgpu_device_ip_reinit_late_sriov(adev);
4402 hive = amdgpu_get_xgmi_hive(adev);
4403 /* Update PSP FW topology after reset */
4404 if (hive && adev->gmc.xgmi.num_physical_nodes > 1)
4405 r = amdgpu_xgmi_update_topology(hive, adev);
4408 amdgpu_put_xgmi_hive(hive);
4411 amdgpu_irq_gpu_reset_resume_helper(adev);
4412 r = amdgpu_ib_ring_tests(adev);
4413 amdgpu_amdkfd_post_reset(adev);
4417 if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
4418 amdgpu_inc_vram_lost(adev);
4419 r = amdgpu_device_recover_vram(adev);
4421 amdgpu_virt_release_full_gpu(adev, true);
4423 if (AMDGPU_RETRY_SRIOV_RESET(r)) {
4424 if (retry_limit < AMDGPU_MAX_RETRY_LIMIT) {
4428 DRM_ERROR("GPU reset retry is beyond the retry limit\n");
4435 * amdgpu_device_has_job_running - check if there is any job in mirror list
4437 * @adev: amdgpu_device pointer
4439 * check if there is any job in mirror list
4441 bool amdgpu_device_has_job_running(struct amdgpu_device *adev)
4444 struct drm_sched_job *job;
4446 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
4447 struct amdgpu_ring *ring = adev->rings[i];
4449 if (!ring || !ring->sched.thread)
4452 spin_lock(&ring->sched.job_list_lock);
4453 job = list_first_entry_or_null(&ring->sched.pending_list,
4454 struct drm_sched_job, list);
4455 spin_unlock(&ring->sched.job_list_lock);
4463 * amdgpu_device_should_recover_gpu - check if we should try GPU recovery
4465 * @adev: amdgpu_device pointer
4467 * Check amdgpu_gpu_recovery and SRIOV status to see if we should try to recover
4470 bool amdgpu_device_should_recover_gpu(struct amdgpu_device *adev)
4472 if (!amdgpu_device_ip_check_soft_reset(adev)) {
4473 dev_info(adev->dev, "Timeout, but no hardware hang detected.\n");
4477 if (amdgpu_gpu_recovery == 0)
4480 if (amdgpu_sriov_vf(adev))
4483 if (amdgpu_gpu_recovery == -1) {
4484 switch (adev->asic_type) {
4485 #ifdef CONFIG_DRM_AMDGPU_SI
4492 #ifdef CONFIG_DRM_AMDGPU_CIK
4499 case CHIP_CYAN_SKILLFISH:
4509 dev_info(adev->dev, "GPU recovery disabled.\n");
4513 int amdgpu_device_mode1_reset(struct amdgpu_device *adev)
4518 amdgpu_atombios_scratch_regs_engine_hung(adev, true);
4520 dev_info(adev->dev, "GPU mode1 reset\n");
4523 pci_clear_master(adev->pdev);
4525 amdgpu_device_cache_pci_state(adev->pdev);
4527 if (amdgpu_dpm_is_mode1_reset_supported(adev)) {
4528 dev_info(adev->dev, "GPU smu mode1 reset\n");
4529 ret = amdgpu_dpm_mode1_reset(adev);
4531 dev_info(adev->dev, "GPU psp mode1 reset\n");
4532 ret = psp_gpu_reset(adev);
4536 dev_err(adev->dev, "GPU mode1 reset failed\n");
4538 amdgpu_device_load_pci_state(adev->pdev);
4540 /* wait for asic to come out of reset */
4541 for (i = 0; i < adev->usec_timeout; i++) {
4542 u32 memsize = adev->nbio.funcs->get_memsize(adev);
4544 if (memsize != 0xffffffff)
4549 amdgpu_atombios_scratch_regs_engine_hung(adev, false);
4553 int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,
4554 struct amdgpu_reset_context *reset_context)
4557 struct amdgpu_job *job = NULL;
4558 bool need_full_reset =
4559 test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
4561 if (reset_context->reset_req_dev == adev)
4562 job = reset_context->job;
4564 if (amdgpu_sriov_vf(adev)) {
4565 /* stop the data exchange thread */
4566 amdgpu_virt_fini_data_exchange(adev);
4569 /* block all schedulers and reset given job's ring */
4570 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
4571 struct amdgpu_ring *ring = adev->rings[i];
4573 if (!ring || !ring->sched.thread)
4576 /*clear job fence from fence drv to avoid force_completion
4577 *leave NULL and vm flush fence in fence drv */
4578 amdgpu_fence_driver_clear_job_fences(ring);
4580 /* after all hw jobs are reset, hw fence is meaningless, so force_completion */
4581 amdgpu_fence_driver_force_completion(ring);
4585 drm_sched_increase_karma(&job->base);
4587 r = amdgpu_reset_prepare_hwcontext(adev, reset_context);
4588 /* If reset handler not implemented, continue; otherwise return */
4594 /* Don't suspend on bare metal if we are not going to HW reset the ASIC */
4595 if (!amdgpu_sriov_vf(adev)) {
4597 if (!need_full_reset)
4598 need_full_reset = amdgpu_device_ip_need_full_reset(adev);
4600 if (!need_full_reset) {
4601 amdgpu_device_ip_pre_soft_reset(adev);
4602 r = amdgpu_device_ip_soft_reset(adev);
4603 amdgpu_device_ip_post_soft_reset(adev);
4604 if (r || amdgpu_device_ip_check_soft_reset(adev)) {
4605 dev_info(adev->dev, "soft reset failed, will fallback to full reset!\n");
4606 need_full_reset = true;
4610 if (need_full_reset)
4611 r = amdgpu_device_ip_suspend(adev);
4612 if (need_full_reset)
4613 set_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
4615 clear_bit(AMDGPU_NEED_FULL_RESET,
4616 &reset_context->flags);
4622 int amdgpu_do_asic_reset(struct list_head *device_list_handle,
4623 struct amdgpu_reset_context *reset_context)
4625 struct amdgpu_device *tmp_adev = NULL;
4626 bool need_full_reset, skip_hw_reset, vram_lost = false;
4629 /* Try reset handler method first */
4630 tmp_adev = list_first_entry(device_list_handle, struct amdgpu_device,
4632 r = amdgpu_reset_perform_reset(tmp_adev, reset_context);
4633 /* If reset handler not implemented, continue; otherwise return */
4639 /* Reset handler not implemented, use the default method */
4641 test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
4642 skip_hw_reset = test_bit(AMDGPU_SKIP_HW_RESET, &reset_context->flags);
4645 * ASIC reset has to be done on all XGMI hive nodes ASAP
4646 * to allow proper links negotiation in FW (within 1 sec)
4648 if (!skip_hw_reset && need_full_reset) {
4649 list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
4650 /* For XGMI run all resets in parallel to speed up the process */
4651 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
4652 tmp_adev->gmc.xgmi.pending_reset = false;
4653 if (!queue_work(system_unbound_wq, &tmp_adev->xgmi_reset_work))
4656 r = amdgpu_asic_reset(tmp_adev);
4659 dev_err(tmp_adev->dev, "ASIC reset failed with error, %d for drm dev, %s",
4660 r, adev_to_drm(tmp_adev)->unique);
4665 /* For XGMI wait for all resets to complete before proceed */
4667 list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
4668 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
4669 flush_work(&tmp_adev->xgmi_reset_work);
4670 r = tmp_adev->asic_reset_res;
4678 if (!r && amdgpu_ras_intr_triggered()) {
4679 list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
4680 if (tmp_adev->mmhub.ras && tmp_adev->mmhub.ras->ras_block.hw_ops &&
4681 tmp_adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count)
4682 tmp_adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count(tmp_adev);
4685 amdgpu_ras_intr_cleared();
4688 list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
4689 if (need_full_reset) {
4691 r = amdgpu_device_asic_init(tmp_adev);
4693 dev_warn(tmp_adev->dev, "asic atom init failed!");
4695 dev_info(tmp_adev->dev, "GPU reset succeeded, trying to resume\n");
4696 r = amdgpu_amdkfd_resume_iommu(tmp_adev);
4700 r = amdgpu_device_ip_resume_phase1(tmp_adev);
4704 vram_lost = amdgpu_device_check_vram_lost(tmp_adev);
4706 DRM_INFO("VRAM is lost due to GPU reset!\n");
4707 amdgpu_inc_vram_lost(tmp_adev);
4710 r = amdgpu_device_fw_loading(tmp_adev);
4714 r = amdgpu_device_ip_resume_phase2(tmp_adev);
4719 amdgpu_device_fill_reset_magic(tmp_adev);
4722 * Add this ASIC as tracked as reset was already
4723 * complete successfully.
4725 amdgpu_register_gpu_instance(tmp_adev);
4727 if (!reset_context->hive &&
4728 tmp_adev->gmc.xgmi.num_physical_nodes > 1)
4729 amdgpu_xgmi_add_device(tmp_adev);
4731 r = amdgpu_device_ip_late_init(tmp_adev);
4735 drm_fb_helper_set_suspend_unlocked(adev_to_drm(tmp_adev)->fb_helper, false);
4738 * The GPU enters bad state once faulty pages
4739 * by ECC has reached the threshold, and ras
4740 * recovery is scheduled next. So add one check
4741 * here to break recovery if it indeed exceeds
4742 * bad page threshold, and remind user to
4743 * retire this GPU or setting one bigger
4744 * bad_page_threshold value to fix this once
4745 * probing driver again.
4747 if (!amdgpu_ras_eeprom_check_err_threshold(tmp_adev)) {
4749 amdgpu_ras_resume(tmp_adev);
4755 /* Update PSP FW topology after reset */
4756 if (reset_context->hive &&
4757 tmp_adev->gmc.xgmi.num_physical_nodes > 1)
4758 r = amdgpu_xgmi_update_topology(
4759 reset_context->hive, tmp_adev);
4765 amdgpu_irq_gpu_reset_resume_helper(tmp_adev);
4766 r = amdgpu_ib_ring_tests(tmp_adev);
4768 dev_err(tmp_adev->dev, "ib ring test failed (%d).\n", r);
4769 need_full_reset = true;
4776 r = amdgpu_device_recover_vram(tmp_adev);
4778 tmp_adev->asic_reset_res = r;
4782 if (need_full_reset)
4783 set_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
4785 clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
4789 static bool amdgpu_device_lock_adev(struct amdgpu_device *adev,
4790 struct amdgpu_hive_info *hive)
4792 if (atomic_cmpxchg(&adev->in_gpu_reset, 0, 1) != 0)
4796 down_write_nest_lock(&adev->reset_sem, &hive->hive_lock);
4798 down_write(&adev->reset_sem);
4801 switch (amdgpu_asic_reset_method(adev)) {
4802 case AMD_RESET_METHOD_MODE1:
4803 adev->mp1_state = PP_MP1_STATE_SHUTDOWN;
4805 case AMD_RESET_METHOD_MODE2:
4806 adev->mp1_state = PP_MP1_STATE_RESET;
4809 adev->mp1_state = PP_MP1_STATE_NONE;
4816 static void amdgpu_device_unlock_adev(struct amdgpu_device *adev)
4818 amdgpu_vf_error_trans_all(adev);
4819 adev->mp1_state = PP_MP1_STATE_NONE;
4820 atomic_set(&adev->in_gpu_reset, 0);
4821 up_write(&adev->reset_sem);
4825 * to lockup a list of amdgpu devices in a hive safely, if not a hive
4826 * with multiple nodes, it will be similar as amdgpu_device_lock_adev.
4828 * unlock won't require roll back.
4830 static int amdgpu_device_lock_hive_adev(struct amdgpu_device *adev, struct amdgpu_hive_info *hive)
4832 struct amdgpu_device *tmp_adev = NULL;
4834 if (!amdgpu_sriov_vf(adev) && (adev->gmc.xgmi.num_physical_nodes > 1)) {
4836 dev_err(adev->dev, "Hive is NULL while device has multiple xgmi nodes");
4839 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
4840 if (!amdgpu_device_lock_adev(tmp_adev, hive))
4843 } else if (!amdgpu_device_lock_adev(adev, hive))
4848 if (!list_is_first(&tmp_adev->gmc.xgmi.head, &hive->device_list)) {
4850 * if the lockup iteration break in the middle of a hive,
4851 * it may means there may has a race issue,
4852 * or a hive device locked up independently.
4853 * we may be in trouble and may not, so will try to roll back
4854 * the lock and give out a warnning.
4856 dev_warn(tmp_adev->dev, "Hive lock iteration broke in the middle. Rolling back to unlock");
4857 list_for_each_entry_continue_reverse(tmp_adev, &hive->device_list, gmc.xgmi.head) {
4858 amdgpu_device_unlock_adev(tmp_adev);
4864 static void amdgpu_device_resume_display_audio(struct amdgpu_device *adev)
4866 struct pci_dev *p = NULL;
4868 p = pci_get_domain_bus_and_slot(pci_domain_nr(adev->pdev->bus),
4869 adev->pdev->bus->number, 1);
4871 pm_runtime_enable(&(p->dev));
4872 pm_runtime_resume(&(p->dev));
4876 static int amdgpu_device_suspend_display_audio(struct amdgpu_device *adev)
4878 enum amd_reset_method reset_method;
4879 struct pci_dev *p = NULL;
4883 * For now, only BACO and mode1 reset are confirmed
4884 * to suffer the audio issue without proper suspended.
4886 reset_method = amdgpu_asic_reset_method(adev);
4887 if ((reset_method != AMD_RESET_METHOD_BACO) &&
4888 (reset_method != AMD_RESET_METHOD_MODE1))
4891 p = pci_get_domain_bus_and_slot(pci_domain_nr(adev->pdev->bus),
4892 adev->pdev->bus->number, 1);
4896 expires = pm_runtime_autosuspend_expiration(&(p->dev));
4899 * If we cannot get the audio device autosuspend delay,
4900 * a fixed 4S interval will be used. Considering 3S is
4901 * the audio controller default autosuspend delay setting.
4902 * 4S used here is guaranteed to cover that.
4904 expires = ktime_get_mono_fast_ns() + NSEC_PER_SEC * 4ULL;
4906 while (!pm_runtime_status_suspended(&(p->dev))) {
4907 if (!pm_runtime_suspend(&(p->dev)))
4910 if (expires < ktime_get_mono_fast_ns()) {
4911 dev_warn(adev->dev, "failed to suspend display audio\n");
4912 /* TODO: abort the succeeding gpu reset? */
4917 pm_runtime_disable(&(p->dev));
4922 static void amdgpu_device_recheck_guilty_jobs(
4923 struct amdgpu_device *adev, struct list_head *device_list_handle,
4924 struct amdgpu_reset_context *reset_context)
4928 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
4929 struct amdgpu_ring *ring = adev->rings[i];
4931 struct drm_sched_job *s_job;
4933 if (!ring || !ring->sched.thread)
4936 s_job = list_first_entry_or_null(&ring->sched.pending_list,
4937 struct drm_sched_job, list);
4941 /* clear job's guilty and depend the folowing step to decide the real one */
4942 drm_sched_reset_karma(s_job);
4943 /* for the real bad job, it will be resubmitted twice, adding a dma_fence_get
4944 * to make sure fence is balanced */
4945 dma_fence_get(s_job->s_fence->parent);
4946 drm_sched_resubmit_jobs_ext(&ring->sched, 1);
4948 ret = dma_fence_wait_timeout(s_job->s_fence->parent, false, ring->sched.timeout);
4949 if (ret == 0) { /* timeout */
4950 DRM_ERROR("Found the real bad job! ring:%s, job_id:%llx\n",
4951 ring->sched.name, s_job->id);
4954 drm_sched_increase_karma(s_job);
4957 if (amdgpu_sriov_vf(adev)) {
4958 amdgpu_virt_fini_data_exchange(adev);
4959 r = amdgpu_device_reset_sriov(adev, false);
4961 adev->asic_reset_res = r;
4963 clear_bit(AMDGPU_SKIP_HW_RESET,
4964 &reset_context->flags);
4965 r = amdgpu_do_asic_reset(device_list_handle,
4967 if (r && r == -EAGAIN)
4972 * add reset counter so that the following
4973 * resubmitted job could flush vmid
4975 atomic_inc(&adev->gpu_reset_counter);
4979 /* got the hw fence, signal finished fence */
4980 atomic_dec(ring->sched.score);
4981 dma_fence_put(s_job->s_fence->parent);
4982 dma_fence_get(&s_job->s_fence->finished);
4983 dma_fence_signal(&s_job->s_fence->finished);
4984 dma_fence_put(&s_job->s_fence->finished);
4986 /* remove node from list and free the job */
4987 spin_lock(&ring->sched.job_list_lock);
4988 list_del_init(&s_job->list);
4989 spin_unlock(&ring->sched.job_list_lock);
4990 ring->sched.ops->free_job(s_job);
4995 * amdgpu_device_gpu_recover - reset the asic and recover scheduler
4997 * @adev: amdgpu_device pointer
4998 * @job: which job trigger hang
5000 * Attempt to reset the GPU if it has hung (all asics).
5001 * Attempt to do soft-reset or full-reset and reinitialize Asic
5002 * Returns 0 for success or an error on failure.
5005 int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
5006 struct amdgpu_job *job)
5008 struct list_head device_list, *device_list_handle = NULL;
5009 bool job_signaled = false;
5010 struct amdgpu_hive_info *hive = NULL;
5011 struct amdgpu_device *tmp_adev = NULL;
5013 bool need_emergency_restart = false;
5014 bool audio_suspended = false;
5015 int tmp_vram_lost_counter;
5016 struct amdgpu_reset_context reset_context;
5018 memset(&reset_context, 0, sizeof(reset_context));
5021 * Special case: RAS triggered and full reset isn't supported
5023 need_emergency_restart = amdgpu_ras_need_emergency_restart(adev);
5026 * Flush RAM to disk so that after reboot
5027 * the user can read log and see why the system rebooted.
5029 if (need_emergency_restart && amdgpu_ras_get_context(adev)->reboot) {
5030 DRM_WARN("Emergency reboot.");
5033 emergency_restart();
5036 dev_info(adev->dev, "GPU %s begin!\n",
5037 need_emergency_restart ? "jobs stop":"reset");
5040 * Here we trylock to avoid chain of resets executing from
5041 * either trigger by jobs on different adevs in XGMI hive or jobs on
5042 * different schedulers for same device while this TO handler is running.
5043 * We always reset all schedulers for device and all devices for XGMI
5044 * hive so that should take care of them too.
5046 if (!amdgpu_sriov_vf(adev))
5047 hive = amdgpu_get_xgmi_hive(adev);
5049 if (atomic_cmpxchg(&hive->in_reset, 0, 1) != 0) {
5050 DRM_INFO("Bailing on TDR for s_job:%llx, hive: %llx as another already in progress",
5051 job ? job->base.id : -1, hive->hive_id);
5052 amdgpu_put_xgmi_hive(hive);
5054 drm_sched_increase_karma(&job->base);
5057 mutex_lock(&hive->hive_lock);
5060 reset_context.method = AMD_RESET_METHOD_NONE;
5061 reset_context.reset_req_dev = adev;
5062 reset_context.job = job;
5063 reset_context.hive = hive;
5064 clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
5067 * lock the device before we try to operate the linked list
5068 * if didn't get the device lock, don't touch the linked list since
5069 * others may iterating it.
5071 r = amdgpu_device_lock_hive_adev(adev, hive);
5073 dev_info(adev->dev, "Bailing on TDR for s_job:%llx, as another already in progress",
5074 job ? job->base.id : -1);
5076 /* even we skipped this reset, still need to set the job to guilty */
5078 drm_sched_increase_karma(&job->base);
5083 * Build list of devices to reset.
5084 * In case we are in XGMI hive mode, resort the device list
5085 * to put adev in the 1st position.
5087 INIT_LIST_HEAD(&device_list);
5088 if (!amdgpu_sriov_vf(adev) && (adev->gmc.xgmi.num_physical_nodes > 1)) {
5089 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head)
5090 list_add_tail(&tmp_adev->reset_list, &device_list);
5091 if (!list_is_first(&adev->reset_list, &device_list))
5092 list_rotate_to_front(&adev->reset_list, &device_list);
5093 device_list_handle = &device_list;
5095 list_add_tail(&adev->reset_list, &device_list);
5096 device_list_handle = &device_list;
5099 /* block all schedulers and reset given job's ring */
5100 list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5102 * Try to put the audio codec into suspend state
5103 * before gpu reset started.
5105 * Due to the power domain of the graphics device
5106 * is shared with AZ power domain. Without this,
5107 * we may change the audio hardware from behind
5108 * the audio driver's back. That will trigger
5109 * some audio codec errors.
5111 if (!amdgpu_device_suspend_display_audio(tmp_adev))
5112 audio_suspended = true;
5114 amdgpu_ras_set_error_query_ready(tmp_adev, false);
5116 cancel_delayed_work_sync(&tmp_adev->delayed_init_work);
5118 if (!amdgpu_sriov_vf(tmp_adev))
5119 amdgpu_amdkfd_pre_reset(tmp_adev);
5122 * Mark these ASICs to be reseted as untracked first
5123 * And add them back after reset completed
5125 amdgpu_unregister_gpu_instance(tmp_adev);
5127 drm_fb_helper_set_suspend_unlocked(adev_to_drm(adev)->fb_helper, true);
5129 /* disable ras on ALL IPs */
5130 if (!need_emergency_restart &&
5131 amdgpu_device_ip_need_full_reset(tmp_adev))
5132 amdgpu_ras_suspend(tmp_adev);
5134 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
5135 struct amdgpu_ring *ring = tmp_adev->rings[i];
5137 if (!ring || !ring->sched.thread)
5140 drm_sched_stop(&ring->sched, job ? &job->base : NULL);
5142 if (need_emergency_restart)
5143 amdgpu_job_stop_all_jobs_on_sched(&ring->sched);
5145 atomic_inc(&tmp_adev->gpu_reset_counter);
5148 if (need_emergency_restart)
5149 goto skip_sched_resume;
5152 * Must check guilty signal here since after this point all old
5153 * HW fences are force signaled.
5155 * job->base holds a reference to parent fence
5157 if (job && job->base.s_fence->parent &&
5158 dma_fence_is_signaled(job->base.s_fence->parent)) {
5159 job_signaled = true;
5160 dev_info(adev->dev, "Guilty job already signaled, skipping HW reset");
5164 retry: /* Rest of adevs pre asic reset from XGMI hive. */
5165 list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5166 r = amdgpu_device_pre_asic_reset(tmp_adev, &reset_context);
5167 /*TODO Should we stop ?*/
5169 dev_err(tmp_adev->dev, "GPU pre asic reset failed with err, %d for drm dev, %s ",
5170 r, adev_to_drm(tmp_adev)->unique);
5171 tmp_adev->asic_reset_res = r;
5175 tmp_vram_lost_counter = atomic_read(&((adev)->vram_lost_counter));
5176 /* Actual ASIC resets if needed.*/
5177 /* Host driver will handle XGMI hive reset for SRIOV */
5178 if (amdgpu_sriov_vf(adev)) {
5179 r = amdgpu_device_reset_sriov(adev, job ? false : true);
5181 adev->asic_reset_res = r;
5183 r = amdgpu_do_asic_reset(device_list_handle, &reset_context);
5184 if (r && r == -EAGAIN)
5190 /* Post ASIC reset for all devs .*/
5191 list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5194 * Sometimes a later bad compute job can block a good gfx job as gfx
5195 * and compute ring share internal GC HW mutually. We add an additional
5196 * guilty jobs recheck step to find the real guilty job, it synchronously
5197 * submits and pends for the first job being signaled. If it gets timeout,
5198 * we identify it as a real guilty job.
5200 if (amdgpu_gpu_recovery == 2 &&
5201 !(tmp_vram_lost_counter < atomic_read(&adev->vram_lost_counter)))
5202 amdgpu_device_recheck_guilty_jobs(
5203 tmp_adev, device_list_handle, &reset_context);
5205 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
5206 struct amdgpu_ring *ring = tmp_adev->rings[i];
5208 if (!ring || !ring->sched.thread)
5211 /* No point to resubmit jobs if we didn't HW reset*/
5212 if (!tmp_adev->asic_reset_res && !job_signaled)
5213 drm_sched_resubmit_jobs(&ring->sched);
5215 drm_sched_start(&ring->sched, !tmp_adev->asic_reset_res);
5218 if (!drm_drv_uses_atomic_modeset(adev_to_drm(tmp_adev)) && !job_signaled) {
5219 drm_helper_resume_force_mode(adev_to_drm(tmp_adev));
5222 if (tmp_adev->asic_reset_res)
5223 r = tmp_adev->asic_reset_res;
5225 tmp_adev->asic_reset_res = 0;
5228 /* bad news, how to tell it to userspace ? */
5229 dev_info(tmp_adev->dev, "GPU reset(%d) failed\n", atomic_read(&tmp_adev->gpu_reset_counter));
5230 amdgpu_vf_error_put(tmp_adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r);
5232 dev_info(tmp_adev->dev, "GPU reset(%d) succeeded!\n", atomic_read(&tmp_adev->gpu_reset_counter));
5233 if (amdgpu_acpi_smart_shift_update(adev_to_drm(tmp_adev), AMDGPU_SS_DEV_D0))
5234 DRM_WARN("smart shift update failed\n");
5239 list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5240 /* unlock kfd: SRIOV would do it separately */
5241 if (!need_emergency_restart && !amdgpu_sriov_vf(tmp_adev))
5242 amdgpu_amdkfd_post_reset(tmp_adev);
5244 /* kfd_post_reset will do nothing if kfd device is not initialized,
5245 * need to bring up kfd here if it's not be initialized before
5247 if (!adev->kfd.init_complete)
5248 amdgpu_amdkfd_device_init(adev);
5250 if (audio_suspended)
5251 amdgpu_device_resume_display_audio(tmp_adev);
5252 amdgpu_device_unlock_adev(tmp_adev);
5257 atomic_set(&hive->in_reset, 0);
5258 mutex_unlock(&hive->hive_lock);
5259 amdgpu_put_xgmi_hive(hive);
5262 if (r && r != -EAGAIN)
5263 dev_info(adev->dev, "GPU reset end with ret = %d\n", r);
5268 * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot
5270 * @adev: amdgpu_device pointer
5272 * Fetchs and stores in the driver the PCIE capabilities (gen speed
5273 * and lanes) of the slot the device is in. Handles APUs and
5274 * virtualized environments where PCIE config space may not be available.
5276 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev)
5278 struct pci_dev *pdev;
5279 enum pci_bus_speed speed_cap, platform_speed_cap;
5280 enum pcie_link_width platform_link_width;
5282 if (amdgpu_pcie_gen_cap)
5283 adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
5285 if (amdgpu_pcie_lane_cap)
5286 adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
5288 /* covers APUs as well */
5289 if (pci_is_root_bus(adev->pdev->bus)) {
5290 if (adev->pm.pcie_gen_mask == 0)
5291 adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
5292 if (adev->pm.pcie_mlw_mask == 0)
5293 adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
5297 if (adev->pm.pcie_gen_mask && adev->pm.pcie_mlw_mask)
5300 pcie_bandwidth_available(adev->pdev, NULL,
5301 &platform_speed_cap, &platform_link_width);
5303 if (adev->pm.pcie_gen_mask == 0) {
5306 speed_cap = pcie_get_speed_cap(pdev);
5307 if (speed_cap == PCI_SPEED_UNKNOWN) {
5308 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5309 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5310 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
5312 if (speed_cap == PCIE_SPEED_32_0GT)
5313 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5314 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5315 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
5316 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4 |
5317 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN5);
5318 else if (speed_cap == PCIE_SPEED_16_0GT)
5319 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5320 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5321 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
5322 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4);
5323 else if (speed_cap == PCIE_SPEED_8_0GT)
5324 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5325 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5326 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
5327 else if (speed_cap == PCIE_SPEED_5_0GT)
5328 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5329 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2);
5331 adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1;
5334 if (platform_speed_cap == PCI_SPEED_UNKNOWN) {
5335 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5336 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
5338 if (platform_speed_cap == PCIE_SPEED_32_0GT)
5339 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5340 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5341 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
5342 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4 |
5343 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN5);
5344 else if (platform_speed_cap == PCIE_SPEED_16_0GT)
5345 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5346 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5347 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
5348 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4);
5349 else if (platform_speed_cap == PCIE_SPEED_8_0GT)
5350 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5351 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5352 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3);
5353 else if (platform_speed_cap == PCIE_SPEED_5_0GT)
5354 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5355 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
5357 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
5361 if (adev->pm.pcie_mlw_mask == 0) {
5362 if (platform_link_width == PCIE_LNK_WIDTH_UNKNOWN) {
5363 adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK;
5365 switch (platform_link_width) {
5367 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
5368 CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
5369 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
5370 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
5371 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5372 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5373 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5376 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
5377 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
5378 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
5379 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5380 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5381 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5384 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
5385 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
5386 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5387 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5388 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5391 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
5392 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5393 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5394 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5397 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5398 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5399 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5402 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5403 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5406 adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;
5415 int amdgpu_device_baco_enter(struct drm_device *dev)
5417 struct amdgpu_device *adev = drm_to_adev(dev);
5418 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
5420 if (!amdgpu_device_supports_baco(adev_to_drm(adev)))
5423 if (ras && adev->ras_enabled &&
5424 adev->nbio.funcs->enable_doorbell_interrupt)
5425 adev->nbio.funcs->enable_doorbell_interrupt(adev, false);
5427 return amdgpu_dpm_baco_enter(adev);
5430 int amdgpu_device_baco_exit(struct drm_device *dev)
5432 struct amdgpu_device *adev = drm_to_adev(dev);
5433 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
5436 if (!amdgpu_device_supports_baco(adev_to_drm(adev)))
5439 ret = amdgpu_dpm_baco_exit(adev);
5443 if (ras && adev->ras_enabled &&
5444 adev->nbio.funcs->enable_doorbell_interrupt)
5445 adev->nbio.funcs->enable_doorbell_interrupt(adev, true);
5447 if (amdgpu_passthrough(adev) &&
5448 adev->nbio.funcs->clear_doorbell_interrupt)
5449 adev->nbio.funcs->clear_doorbell_interrupt(adev);
5454 static void amdgpu_cancel_all_tdr(struct amdgpu_device *adev)
5458 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
5459 struct amdgpu_ring *ring = adev->rings[i];
5461 if (!ring || !ring->sched.thread)
5464 cancel_delayed_work_sync(&ring->sched.work_tdr);
5469 * amdgpu_pci_error_detected - Called when a PCI error is detected.
5470 * @pdev: PCI device struct
5471 * @state: PCI channel state
5473 * Description: Called when a PCI error is detected.
5475 * Return: PCI_ERS_RESULT_NEED_RESET or PCI_ERS_RESULT_DISCONNECT.
5477 pci_ers_result_t amdgpu_pci_error_detected(struct pci_dev *pdev, pci_channel_state_t state)
5479 struct drm_device *dev = pci_get_drvdata(pdev);
5480 struct amdgpu_device *adev = drm_to_adev(dev);
5483 DRM_INFO("PCI error: detected callback, state(%d)!!\n", state);
5485 if (adev->gmc.xgmi.num_physical_nodes > 1) {
5486 DRM_WARN("No support for XGMI hive yet...");
5487 return PCI_ERS_RESULT_DISCONNECT;
5490 adev->pci_channel_state = state;
5493 case pci_channel_io_normal:
5494 return PCI_ERS_RESULT_CAN_RECOVER;
5495 /* Fatal error, prepare for slot reset */
5496 case pci_channel_io_frozen:
5498 * Cancel and wait for all TDRs in progress if failing to
5499 * set adev->in_gpu_reset in amdgpu_device_lock_adev
5501 * Locking adev->reset_sem will prevent any external access
5502 * to GPU during PCI error recovery
5504 while (!amdgpu_device_lock_adev(adev, NULL))
5505 amdgpu_cancel_all_tdr(adev);
5508 * Block any work scheduling as we do for regular GPU reset
5509 * for the duration of the recovery
5511 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
5512 struct amdgpu_ring *ring = adev->rings[i];
5514 if (!ring || !ring->sched.thread)
5517 drm_sched_stop(&ring->sched, NULL);
5519 atomic_inc(&adev->gpu_reset_counter);
5520 return PCI_ERS_RESULT_NEED_RESET;
5521 case pci_channel_io_perm_failure:
5522 /* Permanent error, prepare for device removal */
5523 return PCI_ERS_RESULT_DISCONNECT;
5526 return PCI_ERS_RESULT_NEED_RESET;
5530 * amdgpu_pci_mmio_enabled - Enable MMIO and dump debug registers
5531 * @pdev: pointer to PCI device
5533 pci_ers_result_t amdgpu_pci_mmio_enabled(struct pci_dev *pdev)
5536 DRM_INFO("PCI error: mmio enabled callback!!\n");
5538 /* TODO - dump whatever for debugging purposes */
5540 /* This called only if amdgpu_pci_error_detected returns
5541 * PCI_ERS_RESULT_CAN_RECOVER. Read/write to the device still
5542 * works, no need to reset slot.
5545 return PCI_ERS_RESULT_RECOVERED;
5549 * amdgpu_pci_slot_reset - Called when PCI slot has been reset.
5550 * @pdev: PCI device struct
5552 * Description: This routine is called by the pci error recovery
5553 * code after the PCI slot has been reset, just before we
5554 * should resume normal operations.
5556 pci_ers_result_t amdgpu_pci_slot_reset(struct pci_dev *pdev)
5558 struct drm_device *dev = pci_get_drvdata(pdev);
5559 struct amdgpu_device *adev = drm_to_adev(dev);
5561 struct amdgpu_reset_context reset_context;
5563 struct list_head device_list;
5565 DRM_INFO("PCI error: slot reset callback!!\n");
5567 memset(&reset_context, 0, sizeof(reset_context));
5569 INIT_LIST_HEAD(&device_list);
5570 list_add_tail(&adev->reset_list, &device_list);
5572 /* wait for asic to come out of reset */
5575 /* Restore PCI confspace */
5576 amdgpu_device_load_pci_state(pdev);
5578 /* confirm ASIC came out of reset */
5579 for (i = 0; i < adev->usec_timeout; i++) {
5580 memsize = amdgpu_asic_get_config_memsize(adev);
5582 if (memsize != 0xffffffff)
5586 if (memsize == 0xffffffff) {
5591 reset_context.method = AMD_RESET_METHOD_NONE;
5592 reset_context.reset_req_dev = adev;
5593 set_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
5594 set_bit(AMDGPU_SKIP_HW_RESET, &reset_context.flags);
5596 adev->no_hw_access = true;
5597 r = amdgpu_device_pre_asic_reset(adev, &reset_context);
5598 adev->no_hw_access = false;
5602 r = amdgpu_do_asic_reset(&device_list, &reset_context);
5606 if (amdgpu_device_cache_pci_state(adev->pdev))
5607 pci_restore_state(adev->pdev);
5609 DRM_INFO("PCIe error recovery succeeded\n");
5611 DRM_ERROR("PCIe error recovery failed, err:%d", r);
5612 amdgpu_device_unlock_adev(adev);
5615 return r ? PCI_ERS_RESULT_DISCONNECT : PCI_ERS_RESULT_RECOVERED;
5619 * amdgpu_pci_resume() - resume normal ops after PCI reset
5620 * @pdev: pointer to PCI device
5622 * Called when the error recovery driver tells us that its
5623 * OK to resume normal operation.
5625 void amdgpu_pci_resume(struct pci_dev *pdev)
5627 struct drm_device *dev = pci_get_drvdata(pdev);
5628 struct amdgpu_device *adev = drm_to_adev(dev);
5632 DRM_INFO("PCI error: resume callback!!\n");
5634 /* Only continue execution for the case of pci_channel_io_frozen */
5635 if (adev->pci_channel_state != pci_channel_io_frozen)
5638 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
5639 struct amdgpu_ring *ring = adev->rings[i];
5641 if (!ring || !ring->sched.thread)
5645 drm_sched_resubmit_jobs(&ring->sched);
5646 drm_sched_start(&ring->sched, true);
5649 amdgpu_device_unlock_adev(adev);
5652 bool amdgpu_device_cache_pci_state(struct pci_dev *pdev)
5654 struct drm_device *dev = pci_get_drvdata(pdev);
5655 struct amdgpu_device *adev = drm_to_adev(dev);
5658 r = pci_save_state(pdev);
5660 kfree(adev->pci_state);
5662 adev->pci_state = pci_store_saved_state(pdev);
5664 if (!adev->pci_state) {
5665 DRM_ERROR("Failed to store PCI saved state");
5669 DRM_WARN("Failed to save PCI state, err:%d\n", r);
5676 bool amdgpu_device_load_pci_state(struct pci_dev *pdev)
5678 struct drm_device *dev = pci_get_drvdata(pdev);
5679 struct amdgpu_device *adev = drm_to_adev(dev);
5682 if (!adev->pci_state)
5685 r = pci_load_saved_state(pdev, adev->pci_state);
5688 pci_restore_state(pdev);
5690 DRM_WARN("Failed to load PCI state, err:%d\n", r);
5697 void amdgpu_device_flush_hdp(struct amdgpu_device *adev,
5698 struct amdgpu_ring *ring)
5700 #ifdef CONFIG_X86_64
5701 if (adev->flags & AMD_IS_APU)
5704 if (adev->gmc.xgmi.connected_to_cpu)
5707 if (ring && ring->funcs->emit_hdp_flush)
5708 amdgpu_ring_emit_hdp_flush(ring);
5710 amdgpu_asic_flush_hdp(adev, ring);
5713 void amdgpu_device_invalidate_hdp(struct amdgpu_device *adev,
5714 struct amdgpu_ring *ring)
5716 #ifdef CONFIG_X86_64
5717 if (adev->flags & AMD_IS_APU)
5720 if (adev->gmc.xgmi.connected_to_cpu)
5723 amdgpu_asic_invalidate_hdp(adev, ring);
5727 * amdgpu_device_halt() - bring hardware to some kind of halt state
5729 * @adev: amdgpu_device pointer
5731 * Bring hardware to some kind of halt state so that no one can touch it
5732 * any more. It will help to maintain error context when error occurred.
5733 * Compare to a simple hang, the system will keep stable at least for SSH
5734 * access. Then it should be trivial to inspect the hardware state and
5735 * see what's going on. Implemented as following:
5737 * 1. drm_dev_unplug() makes device inaccessible to user space(IOCTLs, etc),
5738 * clears all CPU mappings to device, disallows remappings through page faults
5739 * 2. amdgpu_irq_disable_all() disables all interrupts
5740 * 3. amdgpu_fence_driver_hw_fini() signals all HW fences
5741 * 4. set adev->no_hw_access to avoid potential crashes after setp 5
5742 * 5. amdgpu_device_unmap_mmio() clears all MMIO mappings
5743 * 6. pci_disable_device() and pci_wait_for_pending_transaction()
5744 * flush any in flight DMA operations
5746 void amdgpu_device_halt(struct amdgpu_device *adev)
5748 struct pci_dev *pdev = adev->pdev;
5749 struct drm_device *ddev = adev_to_drm(adev);
5751 drm_dev_unplug(ddev);
5753 amdgpu_irq_disable_all(adev);
5755 amdgpu_fence_driver_hw_fini(adev);
5757 adev->no_hw_access = true;
5759 amdgpu_device_unmap_mmio(adev);
5761 pci_disable_device(pdev);
5762 pci_wait_for_pending_transaction(pdev);
5765 u32 amdgpu_device_pcie_port_rreg(struct amdgpu_device *adev,
5768 unsigned long flags, address, data;
5771 address = adev->nbio.funcs->get_pcie_port_index_offset(adev);
5772 data = adev->nbio.funcs->get_pcie_port_data_offset(adev);
5774 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
5775 WREG32(address, reg * 4);
5776 (void)RREG32(address);
5778 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
5782 void amdgpu_device_pcie_port_wreg(struct amdgpu_device *adev,
5785 unsigned long flags, address, data;
5787 address = adev->nbio.funcs->get_pcie_port_index_offset(adev);
5788 data = adev->nbio.funcs->get_pcie_port_data_offset(adev);
5790 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
5791 WREG32(address, reg * 4);
5792 (void)RREG32(address);
5795 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);