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>
35 #include <linux/devcoredump.h>
36 #include <generated/utsrelease.h>
37 #include <linux/pci-p2pdma.h>
38 #include <linux/apple-gmux.h>
40 #include <drm/drm_aperture.h>
41 #include <drm/drm_atomic_helper.h>
42 #include <drm/drm_crtc_helper.h>
43 #include <drm/drm_fb_helper.h>
44 #include <drm/drm_probe_helper.h>
45 #include <drm/amdgpu_drm.h>
46 #include <linux/vgaarb.h>
47 #include <linux/vga_switcheroo.h>
48 #include <linux/efi.h>
50 #include "amdgpu_trace.h"
51 #include "amdgpu_i2c.h"
53 #include "amdgpu_atombios.h"
54 #include "amdgpu_atomfirmware.h"
56 #ifdef CONFIG_DRM_AMDGPU_SI
59 #ifdef CONFIG_DRM_AMDGPU_CIK
65 #include "bif/bif_4_1_d.h"
66 #include <linux/firmware.h>
67 #include "amdgpu_vf_error.h"
69 #include "amdgpu_amdkfd.h"
70 #include "amdgpu_pm.h"
72 #include "amdgpu_xgmi.h"
73 #include "amdgpu_ras.h"
74 #include "amdgpu_pmu.h"
75 #include "amdgpu_fru_eeprom.h"
76 #include "amdgpu_reset.h"
78 #include <linux/suspend.h>
79 #include <drm/task_barrier.h>
80 #include <linux/pm_runtime.h>
82 #include <drm/drm_drv.h>
84 #if IS_ENABLED(CONFIG_X86)
85 #include <asm/intel-family.h>
88 MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin");
89 MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin");
90 MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin");
91 MODULE_FIRMWARE("amdgpu/picasso_gpu_info.bin");
92 MODULE_FIRMWARE("amdgpu/raven2_gpu_info.bin");
93 MODULE_FIRMWARE("amdgpu/arcturus_gpu_info.bin");
94 MODULE_FIRMWARE("amdgpu/navi12_gpu_info.bin");
96 #define AMDGPU_RESUME_MS 2000
97 #define AMDGPU_MAX_RETRY_LIMIT 2
98 #define AMDGPU_RETRY_SRIOV_RESET(r) ((r) == -EBUSY || (r) == -ETIMEDOUT || (r) == -EINVAL)
100 static const struct drm_driver amdgpu_kms_driver;
102 const char *amdgpu_asic_name[] = {
144 * DOC: pcie_replay_count
146 * The amdgpu driver provides a sysfs API for reporting the total number
147 * of PCIe replays (NAKs)
148 * The file pcie_replay_count is used for this and returns the total
149 * number of replays as a sum of the NAKs generated and NAKs received
152 static ssize_t amdgpu_device_get_pcie_replay_count(struct device *dev,
153 struct device_attribute *attr, char *buf)
155 struct drm_device *ddev = dev_get_drvdata(dev);
156 struct amdgpu_device *adev = drm_to_adev(ddev);
157 uint64_t cnt = amdgpu_asic_get_pcie_replay_count(adev);
159 return sysfs_emit(buf, "%llu\n", cnt);
162 static DEVICE_ATTR(pcie_replay_count, S_IRUGO,
163 amdgpu_device_get_pcie_replay_count, NULL);
165 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev);
170 * The amdgpu driver provides a sysfs API for reporting the product name
172 * The file product_name is used for this and returns the product name
173 * as returned from the FRU.
174 * NOTE: This is only available for certain server cards
177 static ssize_t amdgpu_device_get_product_name(struct device *dev,
178 struct device_attribute *attr, char *buf)
180 struct drm_device *ddev = dev_get_drvdata(dev);
181 struct amdgpu_device *adev = drm_to_adev(ddev);
183 return sysfs_emit(buf, "%s\n", adev->product_name);
186 static DEVICE_ATTR(product_name, S_IRUGO,
187 amdgpu_device_get_product_name, NULL);
190 * DOC: product_number
192 * The amdgpu driver provides a sysfs API for reporting the part number
194 * The file product_number is used for this and returns the part number
195 * as returned from the FRU.
196 * NOTE: This is only available for certain server cards
199 static ssize_t amdgpu_device_get_product_number(struct device *dev,
200 struct device_attribute *attr, char *buf)
202 struct drm_device *ddev = dev_get_drvdata(dev);
203 struct amdgpu_device *adev = drm_to_adev(ddev);
205 return sysfs_emit(buf, "%s\n", adev->product_number);
208 static DEVICE_ATTR(product_number, S_IRUGO,
209 amdgpu_device_get_product_number, NULL);
214 * The amdgpu driver provides a sysfs API for reporting the serial number
216 * The file serial_number is used for this and returns the serial number
217 * as returned from the FRU.
218 * NOTE: This is only available for certain server cards
221 static ssize_t amdgpu_device_get_serial_number(struct device *dev,
222 struct device_attribute *attr, char *buf)
224 struct drm_device *ddev = dev_get_drvdata(dev);
225 struct amdgpu_device *adev = drm_to_adev(ddev);
227 return sysfs_emit(buf, "%s\n", adev->serial);
230 static DEVICE_ATTR(serial_number, S_IRUGO,
231 amdgpu_device_get_serial_number, NULL);
234 * amdgpu_device_supports_px - Is the device a dGPU with ATPX power control
236 * @dev: drm_device pointer
238 * Returns true if the device is a dGPU with ATPX power control,
239 * otherwise return false.
241 bool amdgpu_device_supports_px(struct drm_device *dev)
243 struct amdgpu_device *adev = drm_to_adev(dev);
245 if ((adev->flags & AMD_IS_PX) && !amdgpu_is_atpx_hybrid())
251 * amdgpu_device_supports_boco - Is the device a dGPU with ACPI power resources
253 * @dev: drm_device pointer
255 * Returns true if the device is a dGPU with ACPI power control,
256 * otherwise return false.
258 bool amdgpu_device_supports_boco(struct drm_device *dev)
260 struct amdgpu_device *adev = drm_to_adev(dev);
263 ((adev->flags & AMD_IS_PX) && amdgpu_is_atpx_hybrid()))
269 * amdgpu_device_supports_baco - Does the device support BACO
271 * @dev: drm_device pointer
273 * Returns true if the device supporte BACO,
274 * otherwise return false.
276 bool amdgpu_device_supports_baco(struct drm_device *dev)
278 struct amdgpu_device *adev = drm_to_adev(dev);
280 return amdgpu_asic_supports_baco(adev);
284 * amdgpu_device_supports_smart_shift - Is the device dGPU with
285 * smart shift support
287 * @dev: drm_device pointer
289 * Returns true if the device is a dGPU with Smart Shift support,
290 * otherwise returns false.
292 bool amdgpu_device_supports_smart_shift(struct drm_device *dev)
294 return (amdgpu_device_supports_boco(dev) &&
295 amdgpu_acpi_is_power_shift_control_supported());
299 * VRAM access helper functions
303 * amdgpu_device_mm_access - access vram by MM_INDEX/MM_DATA
305 * @adev: amdgpu_device pointer
306 * @pos: offset of the buffer in vram
307 * @buf: virtual address of the buffer in system memory
308 * @size: read/write size, sizeof(@buf) must > @size
309 * @write: true - write to vram, otherwise - read from vram
311 void amdgpu_device_mm_access(struct amdgpu_device *adev, loff_t pos,
312 void *buf, size_t size, bool write)
315 uint32_t hi = ~0, tmp = 0;
316 uint32_t *data = buf;
320 if (!drm_dev_enter(adev_to_drm(adev), &idx))
323 BUG_ON(!IS_ALIGNED(pos, 4) || !IS_ALIGNED(size, 4));
325 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
326 for (last = pos + size; pos < last; pos += 4) {
329 WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)pos) | 0x80000000);
331 WREG32_NO_KIQ(mmMM_INDEX_HI, tmp);
335 WREG32_NO_KIQ(mmMM_DATA, *data++);
337 *data++ = RREG32_NO_KIQ(mmMM_DATA);
340 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
345 * amdgpu_device_aper_access - access vram by vram aperature
347 * @adev: amdgpu_device pointer
348 * @pos: offset of the buffer in vram
349 * @buf: virtual address of the buffer in system memory
350 * @size: read/write size, sizeof(@buf) must > @size
351 * @write: true - write to vram, otherwise - read from vram
353 * The return value means how many bytes have been transferred.
355 size_t amdgpu_device_aper_access(struct amdgpu_device *adev, loff_t pos,
356 void *buf, size_t size, bool write)
363 if (!adev->mman.aper_base_kaddr)
366 last = min(pos + size, adev->gmc.visible_vram_size);
368 addr = adev->mman.aper_base_kaddr + pos;
372 memcpy_toio(addr, buf, count);
374 amdgpu_device_flush_hdp(adev, NULL);
376 amdgpu_device_invalidate_hdp(adev, NULL);
378 memcpy_fromio(buf, addr, count);
390 * amdgpu_device_vram_access - read/write a buffer in vram
392 * @adev: amdgpu_device pointer
393 * @pos: offset of the buffer in vram
394 * @buf: virtual address of the buffer in system memory
395 * @size: read/write size, sizeof(@buf) must > @size
396 * @write: true - write to vram, otherwise - read from vram
398 void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos,
399 void *buf, size_t size, bool write)
403 /* try to using vram apreature to access vram first */
404 count = amdgpu_device_aper_access(adev, pos, buf, size, write);
407 /* using MM to access rest vram */
410 amdgpu_device_mm_access(adev, pos, buf, size, write);
415 * register access helper functions.
418 /* Check if hw access should be skipped because of hotplug or device error */
419 bool amdgpu_device_skip_hw_access(struct amdgpu_device *adev)
421 if (adev->no_hw_access)
424 #ifdef CONFIG_LOCKDEP
426 * This is a bit complicated to understand, so worth a comment. What we assert
427 * here is that the GPU reset is not running on another thread in parallel.
429 * For this we trylock the read side of the reset semaphore, if that succeeds
430 * we know that the reset is not running in paralell.
432 * If the trylock fails we assert that we are either already holding the read
433 * side of the lock or are the reset thread itself and hold the write side of
437 if (down_read_trylock(&adev->reset_domain->sem))
438 up_read(&adev->reset_domain->sem);
440 lockdep_assert_held(&adev->reset_domain->sem);
447 * amdgpu_device_rreg - read a memory mapped IO or indirect register
449 * @adev: amdgpu_device pointer
450 * @reg: dword aligned register offset
451 * @acc_flags: access flags which require special behavior
453 * Returns the 32 bit value from the offset specified.
455 uint32_t amdgpu_device_rreg(struct amdgpu_device *adev,
456 uint32_t reg, uint32_t acc_flags)
460 if (amdgpu_device_skip_hw_access(adev))
463 if ((reg * 4) < adev->rmmio_size) {
464 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) &&
465 amdgpu_sriov_runtime(adev) &&
466 down_read_trylock(&adev->reset_domain->sem)) {
467 ret = amdgpu_kiq_rreg(adev, reg);
468 up_read(&adev->reset_domain->sem);
470 ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
473 ret = adev->pcie_rreg(adev, reg * 4);
476 trace_amdgpu_device_rreg(adev->pdev->device, reg, ret);
482 * MMIO register read with bytes helper functions
483 * @offset:bytes offset from MMIO start
488 * amdgpu_mm_rreg8 - read a memory mapped IO register
490 * @adev: amdgpu_device pointer
491 * @offset: byte aligned register offset
493 * Returns the 8 bit value from the offset specified.
495 uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset)
497 if (amdgpu_device_skip_hw_access(adev))
500 if (offset < adev->rmmio_size)
501 return (readb(adev->rmmio + offset));
506 * MMIO register write with bytes helper functions
507 * @offset:bytes offset from MMIO start
508 * @value: the value want to be written to the register
512 * amdgpu_mm_wreg8 - read a memory mapped IO register
514 * @adev: amdgpu_device pointer
515 * @offset: byte aligned register offset
516 * @value: 8 bit value to write
518 * Writes the value specified to the offset specified.
520 void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value)
522 if (amdgpu_device_skip_hw_access(adev))
525 if (offset < adev->rmmio_size)
526 writeb(value, adev->rmmio + offset);
532 * amdgpu_device_wreg - write to a memory mapped IO or indirect register
534 * @adev: amdgpu_device pointer
535 * @reg: dword aligned register offset
536 * @v: 32 bit value to write to the register
537 * @acc_flags: access flags which require special behavior
539 * Writes the value specified to the offset specified.
541 void amdgpu_device_wreg(struct amdgpu_device *adev,
542 uint32_t reg, uint32_t v,
545 if (amdgpu_device_skip_hw_access(adev))
548 if ((reg * 4) < adev->rmmio_size) {
549 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) &&
550 amdgpu_sriov_runtime(adev) &&
551 down_read_trylock(&adev->reset_domain->sem)) {
552 amdgpu_kiq_wreg(adev, reg, v);
553 up_read(&adev->reset_domain->sem);
555 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
558 adev->pcie_wreg(adev, reg * 4, v);
561 trace_amdgpu_device_wreg(adev->pdev->device, reg, v);
565 * amdgpu_mm_wreg_mmio_rlc - write register either with direct/indirect mmio or with RLC path if in range
567 * @adev: amdgpu_device pointer
568 * @reg: mmio/rlc register
571 * this function is invoked only for the debugfs register access
573 void amdgpu_mm_wreg_mmio_rlc(struct amdgpu_device *adev,
574 uint32_t reg, uint32_t v)
576 if (amdgpu_device_skip_hw_access(adev))
579 if (amdgpu_sriov_fullaccess(adev) &&
580 adev->gfx.rlc.funcs &&
581 adev->gfx.rlc.funcs->is_rlcg_access_range) {
582 if (adev->gfx.rlc.funcs->is_rlcg_access_range(adev, reg))
583 return amdgpu_sriov_wreg(adev, reg, v, 0, 0);
584 } else if ((reg * 4) >= adev->rmmio_size) {
585 adev->pcie_wreg(adev, reg * 4, v);
587 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
592 * amdgpu_mm_rdoorbell - read a doorbell dword
594 * @adev: amdgpu_device pointer
595 * @index: doorbell index
597 * Returns the value in the doorbell aperture at the
598 * requested doorbell index (CIK).
600 u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
602 if (amdgpu_device_skip_hw_access(adev))
605 if (index < adev->doorbell.num_kernel_doorbells) {
606 return readl(adev->doorbell.ptr + index);
608 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
614 * amdgpu_mm_wdoorbell - write a doorbell dword
616 * @adev: amdgpu_device pointer
617 * @index: doorbell index
620 * Writes @v to the doorbell aperture at the
621 * requested doorbell index (CIK).
623 void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v)
625 if (amdgpu_device_skip_hw_access(adev))
628 if (index < adev->doorbell.num_kernel_doorbells) {
629 writel(v, adev->doorbell.ptr + index);
631 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
636 * amdgpu_mm_rdoorbell64 - read a doorbell Qword
638 * @adev: amdgpu_device pointer
639 * @index: doorbell index
641 * Returns the value in the doorbell aperture at the
642 * requested doorbell index (VEGA10+).
644 u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index)
646 if (amdgpu_device_skip_hw_access(adev))
649 if (index < adev->doorbell.num_kernel_doorbells) {
650 return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index));
652 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
658 * amdgpu_mm_wdoorbell64 - write a doorbell Qword
660 * @adev: amdgpu_device pointer
661 * @index: doorbell index
664 * Writes @v to the doorbell aperture at the
665 * requested doorbell index (VEGA10+).
667 void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v)
669 if (amdgpu_device_skip_hw_access(adev))
672 if (index < adev->doorbell.num_kernel_doorbells) {
673 atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v);
675 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
680 * amdgpu_device_indirect_rreg - read an indirect register
682 * @adev: amdgpu_device pointer
683 * @reg_addr: indirect register address to read from
685 * Returns the value of indirect register @reg_addr
687 u32 amdgpu_device_indirect_rreg(struct amdgpu_device *adev,
690 unsigned long flags, pcie_index, pcie_data;
691 void __iomem *pcie_index_offset;
692 void __iomem *pcie_data_offset;
695 pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev);
696 pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev);
698 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
699 pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
700 pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
702 writel(reg_addr, pcie_index_offset);
703 readl(pcie_index_offset);
704 r = readl(pcie_data_offset);
705 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
710 u32 amdgpu_device_indirect_rreg_ext(struct amdgpu_device *adev,
713 unsigned long flags, pcie_index, pcie_index_hi, pcie_data;
715 void __iomem *pcie_index_offset;
716 void __iomem *pcie_index_hi_offset;
717 void __iomem *pcie_data_offset;
719 pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev);
720 pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev);
721 if (adev->nbio.funcs->get_pcie_index_hi_offset)
722 pcie_index_hi = adev->nbio.funcs->get_pcie_index_hi_offset(adev);
726 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
727 pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
728 pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
729 if (pcie_index_hi != 0)
730 pcie_index_hi_offset = (void __iomem *)adev->rmmio +
733 writel(reg_addr, pcie_index_offset);
734 readl(pcie_index_offset);
735 if (pcie_index_hi != 0) {
736 writel((reg_addr >> 32) & 0xff, pcie_index_hi_offset);
737 readl(pcie_index_hi_offset);
739 r = readl(pcie_data_offset);
741 /* clear the high bits */
742 if (pcie_index_hi != 0) {
743 writel(0, pcie_index_hi_offset);
744 readl(pcie_index_hi_offset);
747 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
753 * amdgpu_device_indirect_rreg64 - read a 64bits indirect register
755 * @adev: amdgpu_device pointer
756 * @reg_addr: indirect register address to read from
758 * Returns the value of indirect register @reg_addr
760 u64 amdgpu_device_indirect_rreg64(struct amdgpu_device *adev,
763 unsigned long flags, pcie_index, pcie_data;
764 void __iomem *pcie_index_offset;
765 void __iomem *pcie_data_offset;
768 pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev);
769 pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev);
771 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
772 pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
773 pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
775 /* read low 32 bits */
776 writel(reg_addr, pcie_index_offset);
777 readl(pcie_index_offset);
778 r = readl(pcie_data_offset);
779 /* read high 32 bits */
780 writel(reg_addr + 4, pcie_index_offset);
781 readl(pcie_index_offset);
782 r |= ((u64)readl(pcie_data_offset) << 32);
783 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
789 * amdgpu_device_indirect_wreg - write an indirect register address
791 * @adev: amdgpu_device pointer
792 * @reg_addr: indirect register offset
793 * @reg_data: indirect register data
796 void amdgpu_device_indirect_wreg(struct amdgpu_device *adev,
797 u32 reg_addr, u32 reg_data)
799 unsigned long flags, pcie_index, pcie_data;
800 void __iomem *pcie_index_offset;
801 void __iomem *pcie_data_offset;
803 pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev);
804 pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev);
806 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
807 pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
808 pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
810 writel(reg_addr, pcie_index_offset);
811 readl(pcie_index_offset);
812 writel(reg_data, pcie_data_offset);
813 readl(pcie_data_offset);
814 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
817 void amdgpu_device_indirect_wreg_ext(struct amdgpu_device *adev,
818 u64 reg_addr, u32 reg_data)
820 unsigned long flags, pcie_index, pcie_index_hi, pcie_data;
821 void __iomem *pcie_index_offset;
822 void __iomem *pcie_index_hi_offset;
823 void __iomem *pcie_data_offset;
825 pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev);
826 pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev);
827 if (adev->nbio.funcs->get_pcie_index_hi_offset)
828 pcie_index_hi = adev->nbio.funcs->get_pcie_index_hi_offset(adev);
832 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
833 pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
834 pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
835 if (pcie_index_hi != 0)
836 pcie_index_hi_offset = (void __iomem *)adev->rmmio +
839 writel(reg_addr, pcie_index_offset);
840 readl(pcie_index_offset);
841 if (pcie_index_hi != 0) {
842 writel((reg_addr >> 32) & 0xff, pcie_index_hi_offset);
843 readl(pcie_index_hi_offset);
845 writel(reg_data, pcie_data_offset);
846 readl(pcie_data_offset);
848 /* clear the high bits */
849 if (pcie_index_hi != 0) {
850 writel(0, pcie_index_hi_offset);
851 readl(pcie_index_hi_offset);
854 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
858 * amdgpu_device_indirect_wreg64 - write a 64bits indirect register address
860 * @adev: amdgpu_device pointer
861 * @reg_addr: indirect register offset
862 * @reg_data: indirect register data
865 void amdgpu_device_indirect_wreg64(struct amdgpu_device *adev,
866 u32 reg_addr, u64 reg_data)
868 unsigned long flags, pcie_index, pcie_data;
869 void __iomem *pcie_index_offset;
870 void __iomem *pcie_data_offset;
872 pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev);
873 pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev);
875 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
876 pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
877 pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
879 /* write low 32 bits */
880 writel(reg_addr, pcie_index_offset);
881 readl(pcie_index_offset);
882 writel((u32)(reg_data & 0xffffffffULL), pcie_data_offset);
883 readl(pcie_data_offset);
884 /* write high 32 bits */
885 writel(reg_addr + 4, pcie_index_offset);
886 readl(pcie_index_offset);
887 writel((u32)(reg_data >> 32), pcie_data_offset);
888 readl(pcie_data_offset);
889 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
893 * amdgpu_device_get_rev_id - query device rev_id
895 * @adev: amdgpu_device pointer
897 * Return device rev_id
899 u32 amdgpu_device_get_rev_id(struct amdgpu_device *adev)
901 return adev->nbio.funcs->get_rev_id(adev);
905 * amdgpu_invalid_rreg - dummy reg read function
907 * @adev: amdgpu_device pointer
908 * @reg: offset of register
910 * Dummy register read function. Used for register blocks
911 * that certain asics don't have (all asics).
912 * Returns the value in the register.
914 static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
916 DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
921 static uint32_t amdgpu_invalid_rreg_ext(struct amdgpu_device *adev, uint64_t reg)
923 DRM_ERROR("Invalid callback to read register 0x%llX\n", reg);
929 * amdgpu_invalid_wreg - dummy reg write function
931 * @adev: amdgpu_device pointer
932 * @reg: offset of register
933 * @v: value to write to the register
935 * Dummy register read function. Used for register blocks
936 * that certain asics don't have (all asics).
938 static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
940 DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
945 static void amdgpu_invalid_wreg_ext(struct amdgpu_device *adev, uint64_t reg, uint32_t v)
947 DRM_ERROR("Invalid callback to write register 0x%llX with 0x%08X\n",
953 * amdgpu_invalid_rreg64 - dummy 64 bit reg read function
955 * @adev: amdgpu_device pointer
956 * @reg: offset of register
958 * Dummy register read function. Used for register blocks
959 * that certain asics don't have (all asics).
960 * Returns the value in the register.
962 static uint64_t amdgpu_invalid_rreg64(struct amdgpu_device *adev, uint32_t reg)
964 DRM_ERROR("Invalid callback to read 64 bit register 0x%04X\n", reg);
970 * amdgpu_invalid_wreg64 - dummy reg write function
972 * @adev: amdgpu_device pointer
973 * @reg: offset of register
974 * @v: value to write to the register
976 * Dummy register read function. Used for register blocks
977 * that certain asics don't have (all asics).
979 static void amdgpu_invalid_wreg64(struct amdgpu_device *adev, uint32_t reg, uint64_t v)
981 DRM_ERROR("Invalid callback to write 64 bit register 0x%04X with 0x%08llX\n",
987 * amdgpu_block_invalid_rreg - dummy reg read function
989 * @adev: amdgpu_device pointer
990 * @block: offset of instance
991 * @reg: offset of register
993 * Dummy register read function. Used for register blocks
994 * that certain asics don't have (all asics).
995 * Returns the value in the register.
997 static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
998 uint32_t block, uint32_t reg)
1000 DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
1007 * amdgpu_block_invalid_wreg - dummy reg write function
1009 * @adev: amdgpu_device pointer
1010 * @block: offset of instance
1011 * @reg: offset of register
1012 * @v: value to write to the register
1014 * Dummy register read function. Used for register blocks
1015 * that certain asics don't have (all asics).
1017 static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
1019 uint32_t reg, uint32_t v)
1021 DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
1027 * amdgpu_device_asic_init - Wrapper for atom asic_init
1029 * @adev: amdgpu_device pointer
1031 * Does any asic specific work and then calls atom asic init.
1033 static int amdgpu_device_asic_init(struct amdgpu_device *adev)
1035 amdgpu_asic_pre_asic_init(adev);
1037 if (adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 3) ||
1038 adev->ip_versions[GC_HWIP][0] >= IP_VERSION(11, 0, 0))
1039 return amdgpu_atomfirmware_asic_init(adev, true);
1041 return amdgpu_atom_asic_init(adev->mode_info.atom_context);
1045 * amdgpu_device_mem_scratch_init - allocate the VRAM scratch page
1047 * @adev: amdgpu_device pointer
1049 * Allocates a scratch page of VRAM for use by various things in the
1052 static int amdgpu_device_mem_scratch_init(struct amdgpu_device *adev)
1054 return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE, PAGE_SIZE,
1055 AMDGPU_GEM_DOMAIN_VRAM |
1056 AMDGPU_GEM_DOMAIN_GTT,
1057 &adev->mem_scratch.robj,
1058 &adev->mem_scratch.gpu_addr,
1059 (void **)&adev->mem_scratch.ptr);
1063 * amdgpu_device_mem_scratch_fini - Free the VRAM scratch page
1065 * @adev: amdgpu_device pointer
1067 * Frees the VRAM scratch page.
1069 static void amdgpu_device_mem_scratch_fini(struct amdgpu_device *adev)
1071 amdgpu_bo_free_kernel(&adev->mem_scratch.robj, NULL, NULL);
1075 * amdgpu_device_program_register_sequence - program an array of registers.
1077 * @adev: amdgpu_device pointer
1078 * @registers: pointer to the register array
1079 * @array_size: size of the register array
1081 * Programs an array or registers with and and or masks.
1082 * This is a helper for setting golden registers.
1084 void amdgpu_device_program_register_sequence(struct amdgpu_device *adev,
1085 const u32 *registers,
1086 const u32 array_size)
1088 u32 tmp, reg, and_mask, or_mask;
1094 for (i = 0; i < array_size; i += 3) {
1095 reg = registers[i + 0];
1096 and_mask = registers[i + 1];
1097 or_mask = registers[i + 2];
1099 if (and_mask == 0xffffffff) {
1104 if (adev->family >= AMDGPU_FAMILY_AI)
1105 tmp |= (or_mask & and_mask);
1114 * amdgpu_device_pci_config_reset - reset the GPU
1116 * @adev: amdgpu_device pointer
1118 * Resets the GPU using the pci config reset sequence.
1119 * Only applicable to asics prior to vega10.
1121 void amdgpu_device_pci_config_reset(struct amdgpu_device *adev)
1123 pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
1127 * amdgpu_device_pci_reset - reset the GPU using generic PCI means
1129 * @adev: amdgpu_device pointer
1131 * Resets the GPU using generic pci reset interfaces (FLR, SBR, etc.).
1133 int amdgpu_device_pci_reset(struct amdgpu_device *adev)
1135 return pci_reset_function(adev->pdev);
1139 * GPU doorbell aperture helpers function.
1142 * amdgpu_device_doorbell_init - Init doorbell driver information.
1144 * @adev: amdgpu_device pointer
1146 * Init doorbell driver information (CIK)
1147 * Returns 0 on success, error on failure.
1149 static int amdgpu_device_doorbell_init(struct amdgpu_device *adev)
1152 /* No doorbell on SI hardware generation */
1153 if (adev->asic_type < CHIP_BONAIRE) {
1154 adev->doorbell.base = 0;
1155 adev->doorbell.size = 0;
1156 adev->doorbell.num_kernel_doorbells = 0;
1157 adev->doorbell.ptr = NULL;
1161 if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET)
1164 amdgpu_asic_init_doorbell_index(adev);
1166 /* doorbell bar mapping */
1167 adev->doorbell.base = pci_resource_start(adev->pdev, 2);
1168 adev->doorbell.size = pci_resource_len(adev->pdev, 2);
1170 if (adev->enable_mes) {
1171 adev->doorbell.num_kernel_doorbells =
1172 adev->doorbell.size / sizeof(u32);
1174 adev->doorbell.num_kernel_doorbells =
1175 min_t(u32, adev->doorbell.size / sizeof(u32),
1176 adev->doorbell_index.max_assignment+1);
1177 if (adev->doorbell.num_kernel_doorbells == 0)
1180 /* For Vega, reserve and map two pages on doorbell BAR since SDMA
1181 * paging queue doorbell use the second page. The
1182 * AMDGPU_DOORBELL64_MAX_ASSIGNMENT definition assumes all the
1183 * doorbells are in the first page. So with paging queue enabled,
1184 * the max num_kernel_doorbells should + 1 page (0x400 in dword)
1186 if (adev->ip_versions[SDMA0_HWIP][0] >= IP_VERSION(4, 0, 0) &&
1187 adev->ip_versions[SDMA0_HWIP][0] < IP_VERSION(4, 2, 0))
1188 adev->doorbell.num_kernel_doorbells += 0x400;
1191 adev->doorbell.ptr = ioremap(adev->doorbell.base,
1192 adev->doorbell.num_kernel_doorbells *
1194 if (adev->doorbell.ptr == NULL)
1201 * amdgpu_device_doorbell_fini - Tear down doorbell driver information.
1203 * @adev: amdgpu_device pointer
1205 * Tear down doorbell driver information (CIK)
1207 static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev)
1209 iounmap(adev->doorbell.ptr);
1210 adev->doorbell.ptr = NULL;
1216 * amdgpu_device_wb_*()
1217 * Writeback is the method by which the GPU updates special pages in memory
1218 * with the status of certain GPU events (fences, ring pointers,etc.).
1222 * amdgpu_device_wb_fini - Disable Writeback and free memory
1224 * @adev: amdgpu_device pointer
1226 * Disables Writeback and frees the Writeback memory (all asics).
1227 * Used at driver shutdown.
1229 static void amdgpu_device_wb_fini(struct amdgpu_device *adev)
1231 if (adev->wb.wb_obj) {
1232 amdgpu_bo_free_kernel(&adev->wb.wb_obj,
1234 (void **)&adev->wb.wb);
1235 adev->wb.wb_obj = NULL;
1240 * amdgpu_device_wb_init - Init Writeback driver info and allocate memory
1242 * @adev: amdgpu_device pointer
1244 * Initializes writeback and allocates writeback memory (all asics).
1245 * Used at driver startup.
1246 * Returns 0 on success or an -error on failure.
1248 static int amdgpu_device_wb_init(struct amdgpu_device *adev)
1252 if (adev->wb.wb_obj == NULL) {
1253 /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */
1254 r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8,
1255 PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
1256 &adev->wb.wb_obj, &adev->wb.gpu_addr,
1257 (void **)&adev->wb.wb);
1259 dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
1263 adev->wb.num_wb = AMDGPU_MAX_WB;
1264 memset(&adev->wb.used, 0, sizeof(adev->wb.used));
1266 /* clear wb memory */
1267 memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8);
1274 * amdgpu_device_wb_get - Allocate a wb entry
1276 * @adev: amdgpu_device pointer
1279 * Allocate a wb slot for use by the driver (all asics).
1280 * Returns 0 on success or -EINVAL on failure.
1282 int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb)
1284 unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
1286 if (offset < adev->wb.num_wb) {
1287 __set_bit(offset, adev->wb.used);
1288 *wb = offset << 3; /* convert to dw offset */
1296 * amdgpu_device_wb_free - Free a wb entry
1298 * @adev: amdgpu_device pointer
1301 * Free a wb slot allocated for use by the driver (all asics)
1303 void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb)
1306 if (wb < adev->wb.num_wb)
1307 __clear_bit(wb, adev->wb.used);
1311 * amdgpu_device_resize_fb_bar - try to resize FB BAR
1313 * @adev: amdgpu_device pointer
1315 * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not
1316 * to fail, but if any of the BARs is not accessible after the size we abort
1317 * driver loading by returning -ENODEV.
1319 int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev)
1321 int rbar_size = pci_rebar_bytes_to_size(adev->gmc.real_vram_size);
1322 struct pci_bus *root;
1323 struct resource *res;
1329 if (amdgpu_sriov_vf(adev))
1332 /* skip if the bios has already enabled large BAR */
1333 if (adev->gmc.real_vram_size &&
1334 (pci_resource_len(adev->pdev, 0) >= adev->gmc.real_vram_size))
1337 /* Check if the root BUS has 64bit memory resources */
1338 root = adev->pdev->bus;
1339 while (root->parent)
1340 root = root->parent;
1342 pci_bus_for_each_resource(root, res, i) {
1343 if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) &&
1344 res->start > 0x100000000ull)
1348 /* Trying to resize is pointless without a root hub window above 4GB */
1352 /* Limit the BAR size to what is available */
1353 rbar_size = min(fls(pci_rebar_get_possible_sizes(adev->pdev, 0)) - 1,
1356 /* Disable memory decoding while we change the BAR addresses and size */
1357 pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd);
1358 pci_write_config_word(adev->pdev, PCI_COMMAND,
1359 cmd & ~PCI_COMMAND_MEMORY);
1361 /* Free the VRAM and doorbell BAR, we most likely need to move both. */
1362 amdgpu_device_doorbell_fini(adev);
1363 if (adev->asic_type >= CHIP_BONAIRE)
1364 pci_release_resource(adev->pdev, 2);
1366 pci_release_resource(adev->pdev, 0);
1368 r = pci_resize_resource(adev->pdev, 0, rbar_size);
1370 DRM_INFO("Not enough PCI address space for a large BAR.");
1371 else if (r && r != -ENOTSUPP)
1372 DRM_ERROR("Problem resizing BAR0 (%d).", r);
1374 pci_assign_unassigned_bus_resources(adev->pdev->bus);
1376 /* When the doorbell or fb BAR isn't available we have no chance of
1379 r = amdgpu_device_doorbell_init(adev);
1380 if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET))
1383 pci_write_config_word(adev->pdev, PCI_COMMAND, cmd);
1388 static bool amdgpu_device_read_bios(struct amdgpu_device *adev)
1390 if (hweight32(adev->aid_mask) && (adev->flags & AMD_IS_APU)) {
1398 * GPU helpers function.
1401 * amdgpu_device_need_post - check if the hw need post or not
1403 * @adev: amdgpu_device pointer
1405 * Check if the asic has been initialized (all asics) at driver startup
1406 * or post is needed if hw reset is performed.
1407 * Returns true if need or false if not.
1409 bool amdgpu_device_need_post(struct amdgpu_device *adev)
1413 if (amdgpu_sriov_vf(adev))
1416 if (!amdgpu_device_read_bios(adev))
1419 if (amdgpu_passthrough(adev)) {
1420 /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
1421 * some old smc fw still need driver do vPost otherwise gpu hang, while
1422 * those smc fw version above 22.15 doesn't have this flaw, so we force
1423 * vpost executed for smc version below 22.15
1425 if (adev->asic_type == CHIP_FIJI) {
1428 err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
1429 /* force vPost if error occured */
1433 fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
1434 if (fw_ver < 0x00160e00)
1439 /* Don't post if we need to reset whole hive on init */
1440 if (adev->gmc.xgmi.pending_reset)
1443 if (adev->has_hw_reset) {
1444 adev->has_hw_reset = false;
1448 /* bios scratch used on CIK+ */
1449 if (adev->asic_type >= CHIP_BONAIRE)
1450 return amdgpu_atombios_scratch_need_asic_init(adev);
1452 /* check MEM_SIZE for older asics */
1453 reg = amdgpu_asic_get_config_memsize(adev);
1455 if ((reg != 0) && (reg != 0xffffffff))
1462 * On APUs with >= 64GB white flickering has been observed w/ SG enabled.
1463 * Disable S/G on such systems until we have a proper fix.
1464 * https://gitlab.freedesktop.org/drm/amd/-/issues/2354
1465 * https://gitlab.freedesktop.org/drm/amd/-/issues/2735
1467 bool amdgpu_sg_display_supported(struct amdgpu_device *adev)
1469 switch (amdgpu_sg_display) {
1479 if ((totalram_pages() << (PAGE_SHIFT - 10)) +
1480 (adev->gmc.real_vram_size / 1024) >= 64000000) {
1481 DRM_WARN("Disabling S/G due to >=64GB RAM\n");
1488 * Intel hosts such as Raptor Lake and Sapphire Rapids don't support dynamic
1489 * speed switching. Until we have confirmation from Intel that a specific host
1490 * supports it, it's safer that we keep it disabled for all.
1492 * https://edc.intel.com/content/www/us/en/design/products/platforms/details/raptor-lake-s/13th-generation-core-processors-datasheet-volume-1-of-2/005/pci-express-support/
1493 * https://gitlab.freedesktop.org/drm/amd/-/issues/2663
1495 bool amdgpu_device_pcie_dynamic_switching_supported(void)
1497 #if IS_ENABLED(CONFIG_X86)
1498 struct cpuinfo_x86 *c = &cpu_data(0);
1500 if (c->x86_vendor == X86_VENDOR_INTEL)
1507 * amdgpu_device_should_use_aspm - check if the device should program ASPM
1509 * @adev: amdgpu_device pointer
1511 * Confirm whether the module parameter and pcie bridge agree that ASPM should
1512 * be set for this device.
1514 * Returns true if it should be used or false if not.
1516 bool amdgpu_device_should_use_aspm(struct amdgpu_device *adev)
1518 switch (amdgpu_aspm) {
1528 return pcie_aspm_enabled(adev->pdev);
1531 bool amdgpu_device_aspm_support_quirk(void)
1533 #if IS_ENABLED(CONFIG_X86)
1534 struct cpuinfo_x86 *c = &cpu_data(0);
1536 return !(c->x86 == 6 && c->x86_model == INTEL_FAM6_ALDERLAKE);
1542 /* if we get transitioned to only one device, take VGA back */
1544 * amdgpu_device_vga_set_decode - enable/disable vga decode
1546 * @pdev: PCI device pointer
1547 * @state: enable/disable vga decode
1549 * Enable/disable vga decode (all asics).
1550 * Returns VGA resource flags.
1552 static unsigned int amdgpu_device_vga_set_decode(struct pci_dev *pdev,
1555 struct amdgpu_device *adev = drm_to_adev(pci_get_drvdata(pdev));
1556 amdgpu_asic_set_vga_state(adev, state);
1558 return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
1559 VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
1561 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
1565 * amdgpu_device_check_block_size - validate the vm block size
1567 * @adev: amdgpu_device pointer
1569 * Validates the vm block size specified via module parameter.
1570 * The vm block size defines number of bits in page table versus page directory,
1571 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
1572 * page table and the remaining bits are in the page directory.
1574 static void amdgpu_device_check_block_size(struct amdgpu_device *adev)
1576 /* defines number of bits in page table versus page directory,
1577 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
1578 * page table and the remaining bits are in the page directory */
1579 if (amdgpu_vm_block_size == -1)
1582 if (amdgpu_vm_block_size < 9) {
1583 dev_warn(adev->dev, "VM page table size (%d) too small\n",
1584 amdgpu_vm_block_size);
1585 amdgpu_vm_block_size = -1;
1590 * amdgpu_device_check_vm_size - validate the vm size
1592 * @adev: amdgpu_device pointer
1594 * Validates the vm size in GB specified via module parameter.
1595 * The VM size is the size of the GPU virtual memory space in GB.
1597 static void amdgpu_device_check_vm_size(struct amdgpu_device *adev)
1599 /* no need to check the default value */
1600 if (amdgpu_vm_size == -1)
1603 if (amdgpu_vm_size < 1) {
1604 dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
1606 amdgpu_vm_size = -1;
1610 static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev)
1613 bool is_os_64 = (sizeof(void *) == 8);
1614 uint64_t total_memory;
1615 uint64_t dram_size_seven_GB = 0x1B8000000;
1616 uint64_t dram_size_three_GB = 0xB8000000;
1618 if (amdgpu_smu_memory_pool_size == 0)
1622 DRM_WARN("Not 64-bit OS, feature not supported\n");
1626 total_memory = (uint64_t)si.totalram * si.mem_unit;
1628 if ((amdgpu_smu_memory_pool_size == 1) ||
1629 (amdgpu_smu_memory_pool_size == 2)) {
1630 if (total_memory < dram_size_three_GB)
1632 } else if ((amdgpu_smu_memory_pool_size == 4) ||
1633 (amdgpu_smu_memory_pool_size == 8)) {
1634 if (total_memory < dram_size_seven_GB)
1637 DRM_WARN("Smu memory pool size not supported\n");
1640 adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28;
1645 DRM_WARN("No enough system memory\n");
1647 adev->pm.smu_prv_buffer_size = 0;
1650 static int amdgpu_device_init_apu_flags(struct amdgpu_device *adev)
1652 if (!(adev->flags & AMD_IS_APU) ||
1653 adev->asic_type < CHIP_RAVEN)
1656 switch (adev->asic_type) {
1658 if (adev->pdev->device == 0x15dd)
1659 adev->apu_flags |= AMD_APU_IS_RAVEN;
1660 if (adev->pdev->device == 0x15d8)
1661 adev->apu_flags |= AMD_APU_IS_PICASSO;
1664 if ((adev->pdev->device == 0x1636) ||
1665 (adev->pdev->device == 0x164c))
1666 adev->apu_flags |= AMD_APU_IS_RENOIR;
1668 adev->apu_flags |= AMD_APU_IS_GREEN_SARDINE;
1671 adev->apu_flags |= AMD_APU_IS_VANGOGH;
1673 case CHIP_YELLOW_CARP:
1675 case CHIP_CYAN_SKILLFISH:
1676 if ((adev->pdev->device == 0x13FE) ||
1677 (adev->pdev->device == 0x143F))
1678 adev->apu_flags |= AMD_APU_IS_CYAN_SKILLFISH2;
1688 * amdgpu_device_check_arguments - validate module params
1690 * @adev: amdgpu_device pointer
1692 * Validates certain module parameters and updates
1693 * the associated values used by the driver (all asics).
1695 static int amdgpu_device_check_arguments(struct amdgpu_device *adev)
1697 if (amdgpu_sched_jobs < 4) {
1698 dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
1700 amdgpu_sched_jobs = 4;
1701 } else if (!is_power_of_2(amdgpu_sched_jobs)) {
1702 dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
1704 amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
1707 if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) {
1708 /* gart size must be greater or equal to 32M */
1709 dev_warn(adev->dev, "gart size (%d) too small\n",
1711 amdgpu_gart_size = -1;
1714 if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) {
1715 /* gtt size must be greater or equal to 32M */
1716 dev_warn(adev->dev, "gtt size (%d) too small\n",
1718 amdgpu_gtt_size = -1;
1721 /* valid range is between 4 and 9 inclusive */
1722 if (amdgpu_vm_fragment_size != -1 &&
1723 (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) {
1724 dev_warn(adev->dev, "valid range is between 4 and 9\n");
1725 amdgpu_vm_fragment_size = -1;
1728 if (amdgpu_sched_hw_submission < 2) {
1729 dev_warn(adev->dev, "sched hw submission jobs (%d) must be at least 2\n",
1730 amdgpu_sched_hw_submission);
1731 amdgpu_sched_hw_submission = 2;
1732 } else if (!is_power_of_2(amdgpu_sched_hw_submission)) {
1733 dev_warn(adev->dev, "sched hw submission jobs (%d) must be a power of 2\n",
1734 amdgpu_sched_hw_submission);
1735 amdgpu_sched_hw_submission = roundup_pow_of_two(amdgpu_sched_hw_submission);
1738 if (amdgpu_reset_method < -1 || amdgpu_reset_method > 4) {
1739 dev_warn(adev->dev, "invalid option for reset method, reverting to default\n");
1740 amdgpu_reset_method = -1;
1743 amdgpu_device_check_smu_prv_buffer_size(adev);
1745 amdgpu_device_check_vm_size(adev);
1747 amdgpu_device_check_block_size(adev);
1749 adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type);
1755 * amdgpu_switcheroo_set_state - set switcheroo state
1757 * @pdev: pci dev pointer
1758 * @state: vga_switcheroo state
1760 * Callback for the switcheroo driver. Suspends or resumes
1761 * the asics before or after it is powered up using ACPI methods.
1763 static void amdgpu_switcheroo_set_state(struct pci_dev *pdev,
1764 enum vga_switcheroo_state state)
1766 struct drm_device *dev = pci_get_drvdata(pdev);
1769 if (amdgpu_device_supports_px(dev) && state == VGA_SWITCHEROO_OFF)
1772 if (state == VGA_SWITCHEROO_ON) {
1773 pr_info("switched on\n");
1774 /* don't suspend or resume card normally */
1775 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1777 pci_set_power_state(pdev, PCI_D0);
1778 amdgpu_device_load_pci_state(pdev);
1779 r = pci_enable_device(pdev);
1781 DRM_WARN("pci_enable_device failed (%d)\n", r);
1782 amdgpu_device_resume(dev, true);
1784 dev->switch_power_state = DRM_SWITCH_POWER_ON;
1786 pr_info("switched off\n");
1787 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1788 amdgpu_device_suspend(dev, true);
1789 amdgpu_device_cache_pci_state(pdev);
1790 /* Shut down the device */
1791 pci_disable_device(pdev);
1792 pci_set_power_state(pdev, PCI_D3cold);
1793 dev->switch_power_state = DRM_SWITCH_POWER_OFF;
1798 * amdgpu_switcheroo_can_switch - see if switcheroo state can change
1800 * @pdev: pci dev pointer
1802 * Callback for the switcheroo driver. Check of the switcheroo
1803 * state can be changed.
1804 * Returns true if the state can be changed, false if not.
1806 static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
1808 struct drm_device *dev = pci_get_drvdata(pdev);
1811 * FIXME: open_count is protected by drm_global_mutex but that would lead to
1812 * locking inversion with the driver load path. And the access here is
1813 * completely racy anyway. So don't bother with locking for now.
1815 return atomic_read(&dev->open_count) == 0;
1818 static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
1819 .set_gpu_state = amdgpu_switcheroo_set_state,
1821 .can_switch = amdgpu_switcheroo_can_switch,
1825 * amdgpu_device_ip_set_clockgating_state - set the CG state
1827 * @dev: amdgpu_device pointer
1828 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1829 * @state: clockgating state (gate or ungate)
1831 * Sets the requested clockgating state for all instances of
1832 * the hardware IP specified.
1833 * Returns the error code from the last instance.
1835 int amdgpu_device_ip_set_clockgating_state(void *dev,
1836 enum amd_ip_block_type block_type,
1837 enum amd_clockgating_state state)
1839 struct amdgpu_device *adev = dev;
1842 for (i = 0; i < adev->num_ip_blocks; i++) {
1843 if (!adev->ip_blocks[i].status.valid)
1845 if (adev->ip_blocks[i].version->type != block_type)
1847 if (!adev->ip_blocks[i].version->funcs->set_clockgating_state)
1849 r = adev->ip_blocks[i].version->funcs->set_clockgating_state(
1850 (void *)adev, state);
1852 DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",
1853 adev->ip_blocks[i].version->funcs->name, r);
1859 * amdgpu_device_ip_set_powergating_state - set the PG state
1861 * @dev: amdgpu_device pointer
1862 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1863 * @state: powergating state (gate or ungate)
1865 * Sets the requested powergating state for all instances of
1866 * the hardware IP specified.
1867 * Returns the error code from the last instance.
1869 int amdgpu_device_ip_set_powergating_state(void *dev,
1870 enum amd_ip_block_type block_type,
1871 enum amd_powergating_state state)
1873 struct amdgpu_device *adev = dev;
1876 for (i = 0; i < adev->num_ip_blocks; i++) {
1877 if (!adev->ip_blocks[i].status.valid)
1879 if (adev->ip_blocks[i].version->type != block_type)
1881 if (!adev->ip_blocks[i].version->funcs->set_powergating_state)
1883 r = adev->ip_blocks[i].version->funcs->set_powergating_state(
1884 (void *)adev, state);
1886 DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",
1887 adev->ip_blocks[i].version->funcs->name, r);
1893 * amdgpu_device_ip_get_clockgating_state - get the CG state
1895 * @adev: amdgpu_device pointer
1896 * @flags: clockgating feature flags
1898 * Walks the list of IPs on the device and updates the clockgating
1899 * flags for each IP.
1900 * Updates @flags with the feature flags for each hardware IP where
1901 * clockgating is enabled.
1903 void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev,
1908 for (i = 0; i < adev->num_ip_blocks; i++) {
1909 if (!adev->ip_blocks[i].status.valid)
1911 if (adev->ip_blocks[i].version->funcs->get_clockgating_state)
1912 adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags);
1917 * amdgpu_device_ip_wait_for_idle - wait for idle
1919 * @adev: amdgpu_device pointer
1920 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1922 * Waits for the request hardware IP to be idle.
1923 * Returns 0 for success or a negative error code on failure.
1925 int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev,
1926 enum amd_ip_block_type block_type)
1930 for (i = 0; i < adev->num_ip_blocks; i++) {
1931 if (!adev->ip_blocks[i].status.valid)
1933 if (adev->ip_blocks[i].version->type == block_type) {
1934 r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
1945 * amdgpu_device_ip_is_idle - is the hardware IP idle
1947 * @adev: amdgpu_device pointer
1948 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1950 * Check if the hardware IP is idle or not.
1951 * Returns true if it the IP is idle, false if not.
1953 bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev,
1954 enum amd_ip_block_type block_type)
1958 for (i = 0; i < adev->num_ip_blocks; i++) {
1959 if (!adev->ip_blocks[i].status.valid)
1961 if (adev->ip_blocks[i].version->type == block_type)
1962 return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
1969 * amdgpu_device_ip_get_ip_block - get a hw IP pointer
1971 * @adev: amdgpu_device pointer
1972 * @type: Type of hardware IP (SMU, GFX, UVD, etc.)
1974 * Returns a pointer to the hardware IP block structure
1975 * if it exists for the asic, otherwise NULL.
1977 struct amdgpu_ip_block *
1978 amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,
1979 enum amd_ip_block_type type)
1983 for (i = 0; i < adev->num_ip_blocks; i++)
1984 if (adev->ip_blocks[i].version->type == type)
1985 return &adev->ip_blocks[i];
1991 * amdgpu_device_ip_block_version_cmp
1993 * @adev: amdgpu_device pointer
1994 * @type: enum amd_ip_block_type
1995 * @major: major version
1996 * @minor: minor version
1998 * return 0 if equal or greater
1999 * return 1 if smaller or the ip_block doesn't exist
2001 int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev,
2002 enum amd_ip_block_type type,
2003 u32 major, u32 minor)
2005 struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type);
2007 if (ip_block && ((ip_block->version->major > major) ||
2008 ((ip_block->version->major == major) &&
2009 (ip_block->version->minor >= minor))))
2016 * amdgpu_device_ip_block_add
2018 * @adev: amdgpu_device pointer
2019 * @ip_block_version: pointer to the IP to add
2021 * Adds the IP block driver information to the collection of IPs
2024 int amdgpu_device_ip_block_add(struct amdgpu_device *adev,
2025 const struct amdgpu_ip_block_version *ip_block_version)
2027 if (!ip_block_version)
2030 switch (ip_block_version->type) {
2031 case AMD_IP_BLOCK_TYPE_VCN:
2032 if (adev->harvest_ip_mask & AMD_HARVEST_IP_VCN_MASK)
2035 case AMD_IP_BLOCK_TYPE_JPEG:
2036 if (adev->harvest_ip_mask & AMD_HARVEST_IP_JPEG_MASK)
2043 DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks,
2044 ip_block_version->funcs->name);
2046 adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
2052 * amdgpu_device_enable_virtual_display - enable virtual display feature
2054 * @adev: amdgpu_device pointer
2056 * Enabled the virtual display feature if the user has enabled it via
2057 * the module parameter virtual_display. This feature provides a virtual
2058 * display hardware on headless boards or in virtualized environments.
2059 * This function parses and validates the configuration string specified by
2060 * the user and configues the virtual display configuration (number of
2061 * virtual connectors, crtcs, etc.) specified.
2063 static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
2065 adev->enable_virtual_display = false;
2067 if (amdgpu_virtual_display) {
2068 const char *pci_address_name = pci_name(adev->pdev);
2069 char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
2071 pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
2072 pciaddstr_tmp = pciaddstr;
2073 while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
2074 pciaddname = strsep(&pciaddname_tmp, ",");
2075 if (!strcmp("all", pciaddname)
2076 || !strcmp(pci_address_name, pciaddname)) {
2080 adev->enable_virtual_display = true;
2083 res = kstrtol(pciaddname_tmp, 10,
2091 adev->mode_info.num_crtc = num_crtc;
2093 adev->mode_info.num_crtc = 1;
2099 DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
2100 amdgpu_virtual_display, pci_address_name,
2101 adev->enable_virtual_display, adev->mode_info.num_crtc);
2107 void amdgpu_device_set_sriov_virtual_display(struct amdgpu_device *adev)
2109 if (amdgpu_sriov_vf(adev) && !adev->enable_virtual_display) {
2110 adev->mode_info.num_crtc = 1;
2111 adev->enable_virtual_display = true;
2112 DRM_INFO("virtual_display:%d, num_crtc:%d\n",
2113 adev->enable_virtual_display, adev->mode_info.num_crtc);
2118 * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware
2120 * @adev: amdgpu_device pointer
2122 * Parses the asic configuration parameters specified in the gpu info
2123 * firmware and makes them availale to the driver for use in configuring
2125 * Returns 0 on success, -EINVAL on failure.
2127 static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev)
2129 const char *chip_name;
2132 const struct gpu_info_firmware_header_v1_0 *hdr;
2134 adev->firmware.gpu_info_fw = NULL;
2136 if (adev->mman.discovery_bin) {
2138 * FIXME: The bounding box is still needed by Navi12, so
2139 * temporarily read it from gpu_info firmware. Should be dropped
2140 * when DAL no longer needs it.
2142 if (adev->asic_type != CHIP_NAVI12)
2146 switch (adev->asic_type) {
2150 chip_name = "vega10";
2153 chip_name = "vega12";
2156 if (adev->apu_flags & AMD_APU_IS_RAVEN2)
2157 chip_name = "raven2";
2158 else if (adev->apu_flags & AMD_APU_IS_PICASSO)
2159 chip_name = "picasso";
2161 chip_name = "raven";
2164 chip_name = "arcturus";
2167 chip_name = "navi12";
2171 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name);
2172 err = amdgpu_ucode_request(adev, &adev->firmware.gpu_info_fw, fw_name);
2175 "Failed to get gpu_info firmware \"%s\"\n",
2180 hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data;
2181 amdgpu_ucode_print_gpu_info_hdr(&hdr->header);
2183 switch (hdr->version_major) {
2186 const struct gpu_info_firmware_v1_0 *gpu_info_fw =
2187 (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data +
2188 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
2191 * Should be droped when DAL no longer needs it.
2193 if (adev->asic_type == CHIP_NAVI12)
2194 goto parse_soc_bounding_box;
2196 adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se);
2197 adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh);
2198 adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se);
2199 adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se);
2200 adev->gfx.config.max_texture_channel_caches =
2201 le32_to_cpu(gpu_info_fw->gc_num_tccs);
2202 adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs);
2203 adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds);
2204 adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth);
2205 adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth);
2206 adev->gfx.config.double_offchip_lds_buf =
2207 le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer);
2208 adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size);
2209 adev->gfx.cu_info.max_waves_per_simd =
2210 le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd);
2211 adev->gfx.cu_info.max_scratch_slots_per_cu =
2212 le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu);
2213 adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size);
2214 if (hdr->version_minor >= 1) {
2215 const struct gpu_info_firmware_v1_1 *gpu_info_fw =
2216 (const struct gpu_info_firmware_v1_1 *)(adev->firmware.gpu_info_fw->data +
2217 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
2218 adev->gfx.config.num_sc_per_sh =
2219 le32_to_cpu(gpu_info_fw->num_sc_per_sh);
2220 adev->gfx.config.num_packer_per_sc =
2221 le32_to_cpu(gpu_info_fw->num_packer_per_sc);
2224 parse_soc_bounding_box:
2226 * soc bounding box info is not integrated in disocovery table,
2227 * we always need to parse it from gpu info firmware if needed.
2229 if (hdr->version_minor == 2) {
2230 const struct gpu_info_firmware_v1_2 *gpu_info_fw =
2231 (const struct gpu_info_firmware_v1_2 *)(adev->firmware.gpu_info_fw->data +
2232 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
2233 adev->dm.soc_bounding_box = &gpu_info_fw->soc_bounding_box;
2239 "Unsupported gpu_info table %d\n", hdr->header.ucode_version);
2248 * amdgpu_device_ip_early_init - run early init for hardware IPs
2250 * @adev: amdgpu_device pointer
2252 * Early initialization pass for hardware IPs. The hardware IPs that make
2253 * up each asic are discovered each IP's early_init callback is run. This
2254 * is the first stage in initializing the asic.
2255 * Returns 0 on success, negative error code on failure.
2257 static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
2259 struct drm_device *dev = adev_to_drm(adev);
2260 struct pci_dev *parent;
2264 amdgpu_device_enable_virtual_display(adev);
2266 if (amdgpu_sriov_vf(adev)) {
2267 r = amdgpu_virt_request_full_gpu(adev, true);
2272 switch (adev->asic_type) {
2273 #ifdef CONFIG_DRM_AMDGPU_SI
2279 adev->family = AMDGPU_FAMILY_SI;
2280 r = si_set_ip_blocks(adev);
2285 #ifdef CONFIG_DRM_AMDGPU_CIK
2291 if (adev->flags & AMD_IS_APU)
2292 adev->family = AMDGPU_FAMILY_KV;
2294 adev->family = AMDGPU_FAMILY_CI;
2296 r = cik_set_ip_blocks(adev);
2304 case CHIP_POLARIS10:
2305 case CHIP_POLARIS11:
2306 case CHIP_POLARIS12:
2310 if (adev->flags & AMD_IS_APU)
2311 adev->family = AMDGPU_FAMILY_CZ;
2313 adev->family = AMDGPU_FAMILY_VI;
2315 r = vi_set_ip_blocks(adev);
2320 r = amdgpu_discovery_set_ip_blocks(adev);
2326 if (amdgpu_has_atpx() &&
2327 (amdgpu_is_atpx_hybrid() ||
2328 amdgpu_has_atpx_dgpu_power_cntl()) &&
2329 ((adev->flags & AMD_IS_APU) == 0) &&
2330 !pci_is_thunderbolt_attached(to_pci_dev(dev->dev)))
2331 adev->flags |= AMD_IS_PX;
2333 if (!(adev->flags & AMD_IS_APU)) {
2334 parent = pci_upstream_bridge(adev->pdev);
2335 adev->has_pr3 = parent ? pci_pr3_present(parent) : false;
2339 adev->pm.pp_feature = amdgpu_pp_feature_mask;
2340 if (amdgpu_sriov_vf(adev) || sched_policy == KFD_SCHED_POLICY_NO_HWS)
2341 adev->pm.pp_feature &= ~PP_GFXOFF_MASK;
2342 if (amdgpu_sriov_vf(adev) && adev->asic_type == CHIP_SIENNA_CICHLID)
2343 adev->pm.pp_feature &= ~PP_OVERDRIVE_MASK;
2346 for (i = 0; i < adev->num_ip_blocks; i++) {
2347 if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
2348 DRM_WARN("disabled ip block: %d <%s>\n",
2349 i, adev->ip_blocks[i].version->funcs->name);
2350 adev->ip_blocks[i].status.valid = false;
2352 if (adev->ip_blocks[i].version->funcs->early_init) {
2353 r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
2355 adev->ip_blocks[i].status.valid = false;
2357 DRM_ERROR("early_init of IP block <%s> failed %d\n",
2358 adev->ip_blocks[i].version->funcs->name, r);
2361 adev->ip_blocks[i].status.valid = true;
2364 adev->ip_blocks[i].status.valid = true;
2367 /* get the vbios after the asic_funcs are set up */
2368 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) {
2369 r = amdgpu_device_parse_gpu_info_fw(adev);
2374 if (amdgpu_device_read_bios(adev)) {
2375 if (!amdgpu_get_bios(adev))
2378 r = amdgpu_atombios_init(adev);
2380 dev_err(adev->dev, "amdgpu_atombios_init failed\n");
2381 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0);
2386 /*get pf2vf msg info at it's earliest time*/
2387 if (amdgpu_sriov_vf(adev))
2388 amdgpu_virt_init_data_exchange(adev);
2395 amdgpu_amdkfd_device_probe(adev);
2396 adev->cg_flags &= amdgpu_cg_mask;
2397 adev->pg_flags &= amdgpu_pg_mask;
2402 static int amdgpu_device_ip_hw_init_phase1(struct amdgpu_device *adev)
2406 for (i = 0; i < adev->num_ip_blocks; i++) {
2407 if (!adev->ip_blocks[i].status.sw)
2409 if (adev->ip_blocks[i].status.hw)
2411 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2412 (amdgpu_sriov_vf(adev) && (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)) ||
2413 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
2414 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
2416 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
2417 adev->ip_blocks[i].version->funcs->name, r);
2420 adev->ip_blocks[i].status.hw = true;
2427 static int amdgpu_device_ip_hw_init_phase2(struct amdgpu_device *adev)
2431 for (i = 0; i < adev->num_ip_blocks; i++) {
2432 if (!adev->ip_blocks[i].status.sw)
2434 if (adev->ip_blocks[i].status.hw)
2436 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
2438 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
2439 adev->ip_blocks[i].version->funcs->name, r);
2442 adev->ip_blocks[i].status.hw = true;
2448 static int amdgpu_device_fw_loading(struct amdgpu_device *adev)
2452 uint32_t smu_version;
2454 if (adev->asic_type >= CHIP_VEGA10) {
2455 for (i = 0; i < adev->num_ip_blocks; i++) {
2456 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_PSP)
2459 if (!adev->ip_blocks[i].status.sw)
2462 /* no need to do the fw loading again if already done*/
2463 if (adev->ip_blocks[i].status.hw == true)
2466 if (amdgpu_in_reset(adev) || adev->in_suspend) {
2467 r = adev->ip_blocks[i].version->funcs->resume(adev);
2469 DRM_ERROR("resume of IP block <%s> failed %d\n",
2470 adev->ip_blocks[i].version->funcs->name, r);
2474 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
2476 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
2477 adev->ip_blocks[i].version->funcs->name, r);
2482 adev->ip_blocks[i].status.hw = true;
2487 if (!amdgpu_sriov_vf(adev) || adev->asic_type == CHIP_TONGA)
2488 r = amdgpu_pm_load_smu_firmware(adev, &smu_version);
2493 static int amdgpu_device_init_schedulers(struct amdgpu_device *adev)
2498 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
2499 struct amdgpu_ring *ring = adev->rings[i];
2501 /* No need to setup the GPU scheduler for rings that don't need it */
2502 if (!ring || ring->no_scheduler)
2505 switch (ring->funcs->type) {
2506 case AMDGPU_RING_TYPE_GFX:
2507 timeout = adev->gfx_timeout;
2509 case AMDGPU_RING_TYPE_COMPUTE:
2510 timeout = adev->compute_timeout;
2512 case AMDGPU_RING_TYPE_SDMA:
2513 timeout = adev->sdma_timeout;
2516 timeout = adev->video_timeout;
2520 r = drm_sched_init(&ring->sched, &amdgpu_sched_ops,
2521 ring->num_hw_submission, 0,
2522 timeout, adev->reset_domain->wq,
2523 ring->sched_score, ring->name,
2526 DRM_ERROR("Failed to create scheduler on ring %s.\n",
2532 amdgpu_xcp_update_partition_sched_list(adev);
2539 * amdgpu_device_ip_init - run init for hardware IPs
2541 * @adev: amdgpu_device pointer
2543 * Main initialization pass for hardware IPs. The list of all the hardware
2544 * IPs that make up the asic is walked and the sw_init and hw_init callbacks
2545 * are run. sw_init initializes the software state associated with each IP
2546 * and hw_init initializes the hardware associated with each IP.
2547 * Returns 0 on success, negative error code on failure.
2549 static int amdgpu_device_ip_init(struct amdgpu_device *adev)
2553 r = amdgpu_ras_init(adev);
2557 for (i = 0; i < adev->num_ip_blocks; i++) {
2558 if (!adev->ip_blocks[i].status.valid)
2560 r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
2562 DRM_ERROR("sw_init of IP block <%s> failed %d\n",
2563 adev->ip_blocks[i].version->funcs->name, r);
2566 adev->ip_blocks[i].status.sw = true;
2568 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) {
2569 /* need to do common hw init early so everything is set up for gmc */
2570 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
2572 DRM_ERROR("hw_init %d failed %d\n", i, r);
2575 adev->ip_blocks[i].status.hw = true;
2576 } else if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
2577 /* need to do gmc hw init early so we can allocate gpu mem */
2578 /* Try to reserve bad pages early */
2579 if (amdgpu_sriov_vf(adev))
2580 amdgpu_virt_exchange_data(adev);
2582 r = amdgpu_device_mem_scratch_init(adev);
2584 DRM_ERROR("amdgpu_mem_scratch_init failed %d\n", r);
2587 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
2589 DRM_ERROR("hw_init %d failed %d\n", i, r);
2592 r = amdgpu_device_wb_init(adev);
2594 DRM_ERROR("amdgpu_device_wb_init failed %d\n", r);
2597 adev->ip_blocks[i].status.hw = true;
2599 /* right after GMC hw init, we create CSA */
2600 if (adev->gfx.mcbp) {
2601 r = amdgpu_allocate_static_csa(adev, &adev->virt.csa_obj,
2602 AMDGPU_GEM_DOMAIN_VRAM |
2603 AMDGPU_GEM_DOMAIN_GTT,
2606 DRM_ERROR("allocate CSA failed %d\n", r);
2613 if (amdgpu_sriov_vf(adev))
2614 amdgpu_virt_init_data_exchange(adev);
2616 r = amdgpu_ib_pool_init(adev);
2618 dev_err(adev->dev, "IB initialization failed (%d).\n", r);
2619 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r);
2623 r = amdgpu_ucode_create_bo(adev); /* create ucode bo when sw_init complete*/
2627 r = amdgpu_device_ip_hw_init_phase1(adev);
2631 r = amdgpu_device_fw_loading(adev);
2635 r = amdgpu_device_ip_hw_init_phase2(adev);
2640 * retired pages will be loaded from eeprom and reserved here,
2641 * it should be called after amdgpu_device_ip_hw_init_phase2 since
2642 * for some ASICs the RAS EEPROM code relies on SMU fully functioning
2643 * for I2C communication which only true at this point.
2645 * amdgpu_ras_recovery_init may fail, but the upper only cares the
2646 * failure from bad gpu situation and stop amdgpu init process
2647 * accordingly. For other failed cases, it will still release all
2648 * the resource and print error message, rather than returning one
2649 * negative value to upper level.
2651 * Note: theoretically, this should be called before all vram allocations
2652 * to protect retired page from abusing
2654 r = amdgpu_ras_recovery_init(adev);
2659 * In case of XGMI grab extra reference for reset domain for this device
2661 if (adev->gmc.xgmi.num_physical_nodes > 1) {
2662 if (amdgpu_xgmi_add_device(adev) == 0) {
2663 if (!amdgpu_sriov_vf(adev)) {
2664 struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev);
2666 if (WARN_ON(!hive)) {
2671 if (!hive->reset_domain ||
2672 !amdgpu_reset_get_reset_domain(hive->reset_domain)) {
2674 amdgpu_put_xgmi_hive(hive);
2678 /* Drop the early temporary reset domain we created for device */
2679 amdgpu_reset_put_reset_domain(adev->reset_domain);
2680 adev->reset_domain = hive->reset_domain;
2681 amdgpu_put_xgmi_hive(hive);
2686 r = amdgpu_device_init_schedulers(adev);
2690 /* Don't init kfd if whole hive need to be reset during init */
2691 if (!adev->gmc.xgmi.pending_reset) {
2692 kgd2kfd_init_zone_device(adev);
2693 amdgpu_amdkfd_device_init(adev);
2696 amdgpu_fru_get_product_info(adev);
2704 * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer
2706 * @adev: amdgpu_device pointer
2708 * Writes a reset magic value to the gart pointer in VRAM. The driver calls
2709 * this function before a GPU reset. If the value is retained after a
2710 * GPU reset, VRAM has not been lost. Some GPU resets may destry VRAM contents.
2712 static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev)
2714 memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
2718 * amdgpu_device_check_vram_lost - check if vram is valid
2720 * @adev: amdgpu_device pointer
2722 * Checks the reset magic value written to the gart pointer in VRAM.
2723 * The driver calls this after a GPU reset to see if the contents of
2724 * VRAM is lost or now.
2725 * returns true if vram is lost, false if not.
2727 static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev)
2729 if (memcmp(adev->gart.ptr, adev->reset_magic,
2730 AMDGPU_RESET_MAGIC_NUM))
2733 if (!amdgpu_in_reset(adev))
2737 * For all ASICs with baco/mode1 reset, the VRAM is
2738 * always assumed to be lost.
2740 switch (amdgpu_asic_reset_method(adev)) {
2741 case AMD_RESET_METHOD_BACO:
2742 case AMD_RESET_METHOD_MODE1:
2750 * amdgpu_device_set_cg_state - set clockgating for amdgpu device
2752 * @adev: amdgpu_device pointer
2753 * @state: clockgating state (gate or ungate)
2755 * The list of all the hardware IPs that make up the asic is walked and the
2756 * set_clockgating_state callbacks are run.
2757 * Late initialization pass enabling clockgating for hardware IPs.
2758 * Fini or suspend, pass disabling clockgating for hardware IPs.
2759 * Returns 0 on success, negative error code on failure.
2762 int amdgpu_device_set_cg_state(struct amdgpu_device *adev,
2763 enum amd_clockgating_state state)
2767 if (amdgpu_emu_mode == 1)
2770 for (j = 0; j < adev->num_ip_blocks; j++) {
2771 i = state == AMD_CG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
2772 if (!adev->ip_blocks[i].status.late_initialized)
2774 /* skip CG for GFX, SDMA on S0ix */
2775 if (adev->in_s0ix &&
2776 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX ||
2777 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SDMA))
2779 /* skip CG for VCE/UVD, it's handled specially */
2780 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
2781 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
2782 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
2783 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG &&
2784 adev->ip_blocks[i].version->funcs->set_clockgating_state) {
2785 /* enable clockgating to save power */
2786 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
2789 DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
2790 adev->ip_blocks[i].version->funcs->name, r);
2799 int amdgpu_device_set_pg_state(struct amdgpu_device *adev,
2800 enum amd_powergating_state state)
2804 if (amdgpu_emu_mode == 1)
2807 for (j = 0; j < adev->num_ip_blocks; j++) {
2808 i = state == AMD_PG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
2809 if (!adev->ip_blocks[i].status.late_initialized)
2811 /* skip PG for GFX, SDMA on S0ix */
2812 if (adev->in_s0ix &&
2813 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX ||
2814 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SDMA))
2816 /* skip CG for VCE/UVD, it's handled specially */
2817 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
2818 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
2819 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
2820 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG &&
2821 adev->ip_blocks[i].version->funcs->set_powergating_state) {
2822 /* enable powergating to save power */
2823 r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev,
2826 DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n",
2827 adev->ip_blocks[i].version->funcs->name, r);
2835 static int amdgpu_device_enable_mgpu_fan_boost(void)
2837 struct amdgpu_gpu_instance *gpu_ins;
2838 struct amdgpu_device *adev;
2841 mutex_lock(&mgpu_info.mutex);
2844 * MGPU fan boost feature should be enabled
2845 * only when there are two or more dGPUs in
2848 if (mgpu_info.num_dgpu < 2)
2851 for (i = 0; i < mgpu_info.num_dgpu; i++) {
2852 gpu_ins = &(mgpu_info.gpu_ins[i]);
2853 adev = gpu_ins->adev;
2854 if (!(adev->flags & AMD_IS_APU) &&
2855 !gpu_ins->mgpu_fan_enabled) {
2856 ret = amdgpu_dpm_enable_mgpu_fan_boost(adev);
2860 gpu_ins->mgpu_fan_enabled = 1;
2865 mutex_unlock(&mgpu_info.mutex);
2871 * amdgpu_device_ip_late_init - run late init for hardware IPs
2873 * @adev: amdgpu_device pointer
2875 * Late initialization pass for hardware IPs. The list of all the hardware
2876 * IPs that make up the asic is walked and the late_init callbacks are run.
2877 * late_init covers any special initialization that an IP requires
2878 * after all of the have been initialized or something that needs to happen
2879 * late in the init process.
2880 * Returns 0 on success, negative error code on failure.
2882 static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
2884 struct amdgpu_gpu_instance *gpu_instance;
2887 for (i = 0; i < adev->num_ip_blocks; i++) {
2888 if (!adev->ip_blocks[i].status.hw)
2890 if (adev->ip_blocks[i].version->funcs->late_init) {
2891 r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
2893 DRM_ERROR("late_init of IP block <%s> failed %d\n",
2894 adev->ip_blocks[i].version->funcs->name, r);
2898 adev->ip_blocks[i].status.late_initialized = true;
2901 r = amdgpu_ras_late_init(adev);
2903 DRM_ERROR("amdgpu_ras_late_init failed %d", r);
2907 amdgpu_ras_set_error_query_ready(adev, true);
2909 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE);
2910 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE);
2912 amdgpu_device_fill_reset_magic(adev);
2914 r = amdgpu_device_enable_mgpu_fan_boost();
2916 DRM_ERROR("enable mgpu fan boost failed (%d).\n", r);
2918 /* For passthrough configuration on arcturus and aldebaran, enable special handling SBR */
2919 if (amdgpu_passthrough(adev) &&
2920 ((adev->asic_type == CHIP_ARCTURUS && adev->gmc.xgmi.num_physical_nodes > 1) ||
2921 adev->asic_type == CHIP_ALDEBARAN))
2922 amdgpu_dpm_handle_passthrough_sbr(adev, true);
2924 if (adev->gmc.xgmi.num_physical_nodes > 1) {
2925 mutex_lock(&mgpu_info.mutex);
2928 * Reset device p-state to low as this was booted with high.
2930 * This should be performed only after all devices from the same
2931 * hive get initialized.
2933 * However, it's unknown how many device in the hive in advance.
2934 * As this is counted one by one during devices initializations.
2936 * So, we wait for all XGMI interlinked devices initialized.
2937 * This may bring some delays as those devices may come from
2938 * different hives. But that should be OK.
2940 if (mgpu_info.num_dgpu == adev->gmc.xgmi.num_physical_nodes) {
2941 for (i = 0; i < mgpu_info.num_gpu; i++) {
2942 gpu_instance = &(mgpu_info.gpu_ins[i]);
2943 if (gpu_instance->adev->flags & AMD_IS_APU)
2946 r = amdgpu_xgmi_set_pstate(gpu_instance->adev,
2947 AMDGPU_XGMI_PSTATE_MIN);
2949 DRM_ERROR("pstate setting failed (%d).\n", r);
2955 mutex_unlock(&mgpu_info.mutex);
2962 * amdgpu_device_smu_fini_early - smu hw_fini wrapper
2964 * @adev: amdgpu_device pointer
2966 * For ASICs need to disable SMC first
2968 static void amdgpu_device_smu_fini_early(struct amdgpu_device *adev)
2972 if (adev->ip_versions[GC_HWIP][0] > IP_VERSION(9, 0, 0))
2975 for (i = 0; i < adev->num_ip_blocks; i++) {
2976 if (!adev->ip_blocks[i].status.hw)
2978 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
2979 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
2980 /* XXX handle errors */
2982 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
2983 adev->ip_blocks[i].version->funcs->name, r);
2985 adev->ip_blocks[i].status.hw = false;
2991 static int amdgpu_device_ip_fini_early(struct amdgpu_device *adev)
2995 for (i = 0; i < adev->num_ip_blocks; i++) {
2996 if (!adev->ip_blocks[i].version->funcs->early_fini)
2999 r = adev->ip_blocks[i].version->funcs->early_fini((void *)adev);
3001 DRM_DEBUG("early_fini of IP block <%s> failed %d\n",
3002 adev->ip_blocks[i].version->funcs->name, r);
3006 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
3007 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
3009 amdgpu_amdkfd_suspend(adev, false);
3011 /* Workaroud for ASICs need to disable SMC first */
3012 amdgpu_device_smu_fini_early(adev);
3014 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
3015 if (!adev->ip_blocks[i].status.hw)
3018 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
3019 /* XXX handle errors */
3021 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
3022 adev->ip_blocks[i].version->funcs->name, r);
3025 adev->ip_blocks[i].status.hw = false;
3028 if (amdgpu_sriov_vf(adev)) {
3029 if (amdgpu_virt_release_full_gpu(adev, false))
3030 DRM_ERROR("failed to release exclusive mode on fini\n");
3037 * amdgpu_device_ip_fini - run fini for hardware IPs
3039 * @adev: amdgpu_device pointer
3041 * Main teardown pass for hardware IPs. The list of all the hardware
3042 * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks
3043 * are run. hw_fini tears down the hardware associated with each IP
3044 * and sw_fini tears down any software state associated with each IP.
3045 * Returns 0 on success, negative error code on failure.
3047 static int amdgpu_device_ip_fini(struct amdgpu_device *adev)
3051 if (amdgpu_sriov_vf(adev) && adev->virt.ras_init_done)
3052 amdgpu_virt_release_ras_err_handler_data(adev);
3054 if (adev->gmc.xgmi.num_physical_nodes > 1)
3055 amdgpu_xgmi_remove_device(adev);
3057 amdgpu_amdkfd_device_fini_sw(adev);
3059 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
3060 if (!adev->ip_blocks[i].status.sw)
3063 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
3064 amdgpu_ucode_free_bo(adev);
3065 amdgpu_free_static_csa(&adev->virt.csa_obj);
3066 amdgpu_device_wb_fini(adev);
3067 amdgpu_device_mem_scratch_fini(adev);
3068 amdgpu_ib_pool_fini(adev);
3071 r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
3072 /* XXX handle errors */
3074 DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
3075 adev->ip_blocks[i].version->funcs->name, r);
3077 adev->ip_blocks[i].status.sw = false;
3078 adev->ip_blocks[i].status.valid = false;
3081 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
3082 if (!adev->ip_blocks[i].status.late_initialized)
3084 if (adev->ip_blocks[i].version->funcs->late_fini)
3085 adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
3086 adev->ip_blocks[i].status.late_initialized = false;
3089 amdgpu_ras_fini(adev);
3095 * amdgpu_device_delayed_init_work_handler - work handler for IB tests
3097 * @work: work_struct.
3099 static void amdgpu_device_delayed_init_work_handler(struct work_struct *work)
3101 struct amdgpu_device *adev =
3102 container_of(work, struct amdgpu_device, delayed_init_work.work);
3105 r = amdgpu_ib_ring_tests(adev);
3107 DRM_ERROR("ib ring test failed (%d).\n", r);
3110 static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work)
3112 struct amdgpu_device *adev =
3113 container_of(work, struct amdgpu_device, gfx.gfx_off_delay_work.work);
3115 WARN_ON_ONCE(adev->gfx.gfx_off_state);
3116 WARN_ON_ONCE(adev->gfx.gfx_off_req_count);
3118 if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, true))
3119 adev->gfx.gfx_off_state = true;
3123 * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1)
3125 * @adev: amdgpu_device pointer
3127 * Main suspend function for hardware IPs. The list of all the hardware
3128 * IPs that make up the asic is walked, clockgating is disabled and the
3129 * suspend callbacks are run. suspend puts the hardware and software state
3130 * in each IP into a state suitable for suspend.
3131 * Returns 0 on success, negative error code on failure.
3133 static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev)
3137 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
3138 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
3141 * Per PMFW team's suggestion, driver needs to handle gfxoff
3142 * and df cstate features disablement for gpu reset(e.g. Mode1Reset)
3143 * scenario. Add the missing df cstate disablement here.
3145 if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_DISALLOW))
3146 dev_warn(adev->dev, "Failed to disallow df cstate");
3148 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
3149 if (!adev->ip_blocks[i].status.valid)
3152 /* displays are handled separately */
3153 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_DCE)
3156 /* XXX handle errors */
3157 r = adev->ip_blocks[i].version->funcs->suspend(adev);
3158 /* XXX handle errors */
3160 DRM_ERROR("suspend of IP block <%s> failed %d\n",
3161 adev->ip_blocks[i].version->funcs->name, r);
3165 adev->ip_blocks[i].status.hw = false;
3172 * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2)
3174 * @adev: amdgpu_device pointer
3176 * Main suspend function for hardware IPs. The list of all the hardware
3177 * IPs that make up the asic is walked, clockgating is disabled and the
3178 * suspend callbacks are run. suspend puts the hardware and software state
3179 * in each IP into a state suitable for suspend.
3180 * Returns 0 on success, negative error code on failure.
3182 static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev)
3187 amdgpu_dpm_gfx_state_change(adev, sGpuChangeState_D3Entry);
3189 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
3190 if (!adev->ip_blocks[i].status.valid)
3192 /* displays are handled in phase1 */
3193 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE)
3195 /* PSP lost connection when err_event_athub occurs */
3196 if (amdgpu_ras_intr_triggered() &&
3197 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
3198 adev->ip_blocks[i].status.hw = false;
3202 /* skip unnecessary suspend if we do not initialize them yet */
3203 if (adev->gmc.xgmi.pending_reset &&
3204 !(adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
3205 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC ||
3206 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
3207 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH)) {
3208 adev->ip_blocks[i].status.hw = false;
3212 /* skip suspend of gfx/mes and psp for S0ix
3213 * gfx is in gfxoff state, so on resume it will exit gfxoff just
3214 * like at runtime. PSP is also part of the always on hardware
3215 * so no need to suspend it.
3217 if (adev->in_s0ix &&
3218 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP ||
3219 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX ||
3220 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_MES))
3223 /* SDMA 5.x+ is part of GFX power domain so it's covered by GFXOFF */
3224 if (adev->in_s0ix &&
3225 (adev->ip_versions[SDMA0_HWIP][0] >= IP_VERSION(5, 0, 0)) &&
3226 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SDMA))
3229 /* Once swPSP provides the IMU, RLC FW binaries to TOS during cold-boot.
3230 * These are in TMR, hence are expected to be reused by PSP-TOS to reload
3231 * from this location and RLC Autoload automatically also gets loaded
3232 * from here based on PMFW -> PSP message during re-init sequence.
3233 * Therefore, the psp suspend & resume should be skipped to avoid destroy
3234 * the TMR and reload FWs again for IMU enabled APU ASICs.
3236 if (amdgpu_in_reset(adev) &&
3237 (adev->flags & AMD_IS_APU) && adev->gfx.imu.funcs &&
3238 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)
3241 /* XXX handle errors */
3242 r = adev->ip_blocks[i].version->funcs->suspend(adev);
3243 /* XXX handle errors */
3245 DRM_ERROR("suspend of IP block <%s> failed %d\n",
3246 adev->ip_blocks[i].version->funcs->name, r);
3248 adev->ip_blocks[i].status.hw = false;
3249 /* handle putting the SMC in the appropriate state */
3250 if (!amdgpu_sriov_vf(adev)) {
3251 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
3252 r = amdgpu_dpm_set_mp1_state(adev, adev->mp1_state);
3254 DRM_ERROR("SMC failed to set mp1 state %d, %d\n",
3255 adev->mp1_state, r);
3266 * amdgpu_device_ip_suspend - run suspend for hardware IPs
3268 * @adev: amdgpu_device pointer
3270 * Main suspend function for hardware IPs. The list of all the hardware
3271 * IPs that make up the asic is walked, clockgating is disabled and the
3272 * suspend callbacks are run. suspend puts the hardware and software state
3273 * in each IP into a state suitable for suspend.
3274 * Returns 0 on success, negative error code on failure.
3276 int amdgpu_device_ip_suspend(struct amdgpu_device *adev)
3280 if (amdgpu_sriov_vf(adev)) {
3281 amdgpu_virt_fini_data_exchange(adev);
3282 amdgpu_virt_request_full_gpu(adev, false);
3285 r = amdgpu_device_ip_suspend_phase1(adev);
3288 r = amdgpu_device_ip_suspend_phase2(adev);
3290 if (amdgpu_sriov_vf(adev))
3291 amdgpu_virt_release_full_gpu(adev, false);
3296 static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
3300 static enum amd_ip_block_type ip_order[] = {
3301 AMD_IP_BLOCK_TYPE_COMMON,
3302 AMD_IP_BLOCK_TYPE_GMC,
3303 AMD_IP_BLOCK_TYPE_PSP,
3304 AMD_IP_BLOCK_TYPE_IH,
3307 for (i = 0; i < adev->num_ip_blocks; i++) {
3309 struct amdgpu_ip_block *block;
3311 block = &adev->ip_blocks[i];
3312 block->status.hw = false;
3314 for (j = 0; j < ARRAY_SIZE(ip_order); j++) {
3316 if (block->version->type != ip_order[j] ||
3317 !block->status.valid)
3320 r = block->version->funcs->hw_init(adev);
3321 DRM_INFO("RE-INIT-early: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
3324 block->status.hw = true;
3331 static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
3335 static enum amd_ip_block_type ip_order[] = {
3336 AMD_IP_BLOCK_TYPE_SMC,
3337 AMD_IP_BLOCK_TYPE_DCE,
3338 AMD_IP_BLOCK_TYPE_GFX,
3339 AMD_IP_BLOCK_TYPE_SDMA,
3340 AMD_IP_BLOCK_TYPE_MES,
3341 AMD_IP_BLOCK_TYPE_UVD,
3342 AMD_IP_BLOCK_TYPE_VCE,
3343 AMD_IP_BLOCK_TYPE_VCN,
3344 AMD_IP_BLOCK_TYPE_JPEG
3347 for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
3349 struct amdgpu_ip_block *block;
3351 for (j = 0; j < adev->num_ip_blocks; j++) {
3352 block = &adev->ip_blocks[j];
3354 if (block->version->type != ip_order[i] ||
3355 !block->status.valid ||
3359 if (block->version->type == AMD_IP_BLOCK_TYPE_SMC)
3360 r = block->version->funcs->resume(adev);
3362 r = block->version->funcs->hw_init(adev);
3364 DRM_INFO("RE-INIT-late: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
3367 block->status.hw = true;
3375 * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs
3377 * @adev: amdgpu_device pointer
3379 * First resume function for hardware IPs. The list of all the hardware
3380 * IPs that make up the asic is walked and the resume callbacks are run for
3381 * COMMON, GMC, and IH. resume puts the hardware into a functional state
3382 * after a suspend and updates the software state as necessary. This
3383 * function is also used for restoring the GPU after a GPU reset.
3384 * Returns 0 on success, negative error code on failure.
3386 static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)
3390 for (i = 0; i < adev->num_ip_blocks; i++) {
3391 if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
3393 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
3394 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
3395 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH ||
3396 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP && amdgpu_sriov_vf(adev))) {
3398 r = adev->ip_blocks[i].version->funcs->resume(adev);
3400 DRM_ERROR("resume of IP block <%s> failed %d\n",
3401 adev->ip_blocks[i].version->funcs->name, r);
3404 adev->ip_blocks[i].status.hw = true;
3412 * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs
3414 * @adev: amdgpu_device pointer
3416 * First resume function for hardware IPs. The list of all the hardware
3417 * IPs that make up the asic is walked and the resume callbacks are run for
3418 * all blocks except COMMON, GMC, and IH. resume puts the hardware into a
3419 * functional state after a suspend and updates the software state as
3420 * necessary. This function is also used for restoring the GPU after a GPU
3422 * Returns 0 on success, negative error code on failure.
3424 static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)
3428 for (i = 0; i < adev->num_ip_blocks; i++) {
3429 if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
3431 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
3432 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
3433 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH ||
3434 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)
3436 r = adev->ip_blocks[i].version->funcs->resume(adev);
3438 DRM_ERROR("resume of IP block <%s> failed %d\n",
3439 adev->ip_blocks[i].version->funcs->name, r);
3442 adev->ip_blocks[i].status.hw = true;
3449 * amdgpu_device_ip_resume - run resume for hardware IPs
3451 * @adev: amdgpu_device pointer
3453 * Main resume function for hardware IPs. The hardware IPs
3454 * are split into two resume functions because they are
3455 * are also used in in recovering from a GPU reset and some additional
3456 * steps need to be take between them. In this case (S3/S4) they are
3458 * Returns 0 on success, negative error code on failure.
3460 static int amdgpu_device_ip_resume(struct amdgpu_device *adev)
3464 if (!adev->in_s0ix) {
3465 r = amdgpu_amdkfd_resume_iommu(adev);
3470 r = amdgpu_device_ip_resume_phase1(adev);
3474 r = amdgpu_device_fw_loading(adev);
3478 r = amdgpu_device_ip_resume_phase2(adev);
3484 * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV
3486 * @adev: amdgpu_device pointer
3488 * Query the VBIOS data tables to determine if the board supports SR-IOV.
3490 static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
3492 if (amdgpu_sriov_vf(adev)) {
3493 if (adev->is_atom_fw) {
3494 if (amdgpu_atomfirmware_gpu_virtualization_supported(adev))
3495 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
3497 if (amdgpu_atombios_has_gpu_virtualization_table(adev))
3498 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
3501 if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
3502 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
3507 * amdgpu_device_asic_has_dc_support - determine if DC supports the asic
3509 * @asic_type: AMD asic type
3511 * Check if there is DC (new modesetting infrastructre) support for an asic.
3512 * returns true if DC has support, false if not.
3514 bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
3516 switch (asic_type) {
3517 #ifdef CONFIG_DRM_AMDGPU_SI
3521 /* chips with no display hardware */
3523 #if defined(CONFIG_DRM_AMD_DC)
3529 * We have systems in the wild with these ASICs that require
3530 * LVDS and VGA support which is not supported with DC.
3532 * Fallback to the non-DC driver here by default so as not to
3533 * cause regressions.
3535 #if defined(CONFIG_DRM_AMD_DC_SI)
3536 return amdgpu_dc > 0;
3545 * We have systems in the wild with these ASICs that require
3546 * VGA support which is not supported with DC.
3548 * Fallback to the non-DC driver here by default so as not to
3549 * cause regressions.
3551 return amdgpu_dc > 0;
3553 return amdgpu_dc != 0;
3557 DRM_INFO_ONCE("Display Core has been requested via kernel parameter "
3558 "but isn't supported by ASIC, ignoring\n");
3565 * amdgpu_device_has_dc_support - check if dc is supported
3567 * @adev: amdgpu_device pointer
3569 * Returns true for supported, false for not supported
3571 bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
3573 if (adev->enable_virtual_display ||
3574 (adev->harvest_ip_mask & AMD_HARVEST_IP_DMU_MASK))
3577 return amdgpu_device_asic_has_dc_support(adev->asic_type);
3580 static void amdgpu_device_xgmi_reset_func(struct work_struct *__work)
3582 struct amdgpu_device *adev =
3583 container_of(__work, struct amdgpu_device, xgmi_reset_work);
3584 struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev);
3586 /* It's a bug to not have a hive within this function */
3591 * Use task barrier to synchronize all xgmi reset works across the
3592 * hive. task_barrier_enter and task_barrier_exit will block
3593 * until all the threads running the xgmi reset works reach
3594 * those points. task_barrier_full will do both blocks.
3596 if (amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) {
3598 task_barrier_enter(&hive->tb);
3599 adev->asic_reset_res = amdgpu_device_baco_enter(adev_to_drm(adev));
3601 if (adev->asic_reset_res)
3604 task_barrier_exit(&hive->tb);
3605 adev->asic_reset_res = amdgpu_device_baco_exit(adev_to_drm(adev));
3607 if (adev->asic_reset_res)
3610 if (adev->mmhub.ras && adev->mmhub.ras->ras_block.hw_ops &&
3611 adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count)
3612 adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count(adev);
3615 task_barrier_full(&hive->tb);
3616 adev->asic_reset_res = amdgpu_asic_reset(adev);
3620 if (adev->asic_reset_res)
3621 DRM_WARN("ASIC reset failed with error, %d for drm dev, %s",
3622 adev->asic_reset_res, adev_to_drm(adev)->unique);
3623 amdgpu_put_xgmi_hive(hive);
3626 static int amdgpu_device_get_job_timeout_settings(struct amdgpu_device *adev)
3628 char *input = amdgpu_lockup_timeout;
3629 char *timeout_setting = NULL;
3635 * By default timeout for non compute jobs is 10000
3636 * and 60000 for compute jobs.
3637 * In SR-IOV or passthrough mode, timeout for compute
3638 * jobs are 60000 by default.
3640 adev->gfx_timeout = msecs_to_jiffies(10000);
3641 adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout;
3642 if (amdgpu_sriov_vf(adev))
3643 adev->compute_timeout = amdgpu_sriov_is_pp_one_vf(adev) ?
3644 msecs_to_jiffies(60000) : msecs_to_jiffies(10000);
3646 adev->compute_timeout = msecs_to_jiffies(60000);
3648 if (strnlen(input, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) {
3649 while ((timeout_setting = strsep(&input, ",")) &&
3650 strnlen(timeout_setting, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) {
3651 ret = kstrtol(timeout_setting, 0, &timeout);
3658 } else if (timeout < 0) {
3659 timeout = MAX_SCHEDULE_TIMEOUT;
3660 dev_warn(adev->dev, "lockup timeout disabled");
3661 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
3663 timeout = msecs_to_jiffies(timeout);
3668 adev->gfx_timeout = timeout;
3671 adev->compute_timeout = timeout;
3674 adev->sdma_timeout = timeout;
3677 adev->video_timeout = timeout;
3684 * There is only one value specified and
3685 * it should apply to all non-compute jobs.
3688 adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout;
3689 if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev))
3690 adev->compute_timeout = adev->gfx_timeout;
3698 * amdgpu_device_check_iommu_direct_map - check if RAM direct mapped to GPU
3700 * @adev: amdgpu_device pointer
3702 * RAM direct mapped to GPU if IOMMU is not enabled or is pass through mode
3704 static void amdgpu_device_check_iommu_direct_map(struct amdgpu_device *adev)
3706 struct iommu_domain *domain;
3708 domain = iommu_get_domain_for_dev(adev->dev);
3709 if (!domain || domain->type == IOMMU_DOMAIN_IDENTITY)
3710 adev->ram_is_direct_mapped = true;
3713 static const struct attribute *amdgpu_dev_attributes[] = {
3714 &dev_attr_product_name.attr,
3715 &dev_attr_product_number.attr,
3716 &dev_attr_serial_number.attr,
3717 &dev_attr_pcie_replay_count.attr,
3721 static void amdgpu_device_set_mcbp(struct amdgpu_device *adev)
3723 if (amdgpu_mcbp == 1)
3724 adev->gfx.mcbp = true;
3725 else if (amdgpu_mcbp == 0)
3726 adev->gfx.mcbp = false;
3727 else if ((adev->ip_versions[GC_HWIP][0] >= IP_VERSION(9, 0, 0)) &&
3728 (adev->ip_versions[GC_HWIP][0] < IP_VERSION(10, 0, 0)) &&
3729 adev->gfx.num_gfx_rings)
3730 adev->gfx.mcbp = true;
3732 if (amdgpu_sriov_vf(adev))
3733 adev->gfx.mcbp = true;
3736 DRM_INFO("MCBP is enabled\n");
3740 * amdgpu_device_init - initialize the driver
3742 * @adev: amdgpu_device pointer
3743 * @flags: driver flags
3745 * Initializes the driver info and hw (all asics).
3746 * Returns 0 for success or an error on failure.
3747 * Called at driver startup.
3749 int amdgpu_device_init(struct amdgpu_device *adev,
3752 struct drm_device *ddev = adev_to_drm(adev);
3753 struct pci_dev *pdev = adev->pdev;
3759 adev->shutdown = false;
3760 adev->flags = flags;
3762 if (amdgpu_force_asic_type >= 0 && amdgpu_force_asic_type < CHIP_LAST)
3763 adev->asic_type = amdgpu_force_asic_type;
3765 adev->asic_type = flags & AMD_ASIC_MASK;
3767 adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
3768 if (amdgpu_emu_mode == 1)
3769 adev->usec_timeout *= 10;
3770 adev->gmc.gart_size = 512 * 1024 * 1024;
3771 adev->accel_working = false;
3772 adev->num_rings = 0;
3773 RCU_INIT_POINTER(adev->gang_submit, dma_fence_get_stub());
3774 adev->mman.buffer_funcs = NULL;
3775 adev->mman.buffer_funcs_ring = NULL;
3776 adev->vm_manager.vm_pte_funcs = NULL;
3777 adev->vm_manager.vm_pte_num_scheds = 0;
3778 adev->gmc.gmc_funcs = NULL;
3779 adev->harvest_ip_mask = 0x0;
3780 adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
3781 bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
3783 adev->smc_rreg = &amdgpu_invalid_rreg;
3784 adev->smc_wreg = &amdgpu_invalid_wreg;
3785 adev->pcie_rreg = &amdgpu_invalid_rreg;
3786 adev->pcie_wreg = &amdgpu_invalid_wreg;
3787 adev->pcie_rreg_ext = &amdgpu_invalid_rreg_ext;
3788 adev->pcie_wreg_ext = &amdgpu_invalid_wreg_ext;
3789 adev->pciep_rreg = &amdgpu_invalid_rreg;
3790 adev->pciep_wreg = &amdgpu_invalid_wreg;
3791 adev->pcie_rreg64 = &amdgpu_invalid_rreg64;
3792 adev->pcie_wreg64 = &amdgpu_invalid_wreg64;
3793 adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
3794 adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
3795 adev->didt_rreg = &amdgpu_invalid_rreg;
3796 adev->didt_wreg = &amdgpu_invalid_wreg;
3797 adev->gc_cac_rreg = &amdgpu_invalid_rreg;
3798 adev->gc_cac_wreg = &amdgpu_invalid_wreg;
3799 adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
3800 adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
3802 DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
3803 amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
3804 pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
3806 /* mutex initialization are all done here so we
3807 * can recall function without having locking issues */
3808 mutex_init(&adev->firmware.mutex);
3809 mutex_init(&adev->pm.mutex);
3810 mutex_init(&adev->gfx.gpu_clock_mutex);
3811 mutex_init(&adev->srbm_mutex);
3812 mutex_init(&adev->gfx.pipe_reserve_mutex);
3813 mutex_init(&adev->gfx.gfx_off_mutex);
3814 mutex_init(&adev->gfx.partition_mutex);
3815 mutex_init(&adev->grbm_idx_mutex);
3816 mutex_init(&adev->mn_lock);
3817 mutex_init(&adev->virt.vf_errors.lock);
3818 hash_init(adev->mn_hash);
3819 mutex_init(&adev->psp.mutex);
3820 mutex_init(&adev->notifier_lock);
3821 mutex_init(&adev->pm.stable_pstate_ctx_lock);
3822 mutex_init(&adev->benchmark_mutex);
3824 amdgpu_device_init_apu_flags(adev);
3826 r = amdgpu_device_check_arguments(adev);
3830 spin_lock_init(&adev->mmio_idx_lock);
3831 spin_lock_init(&adev->smc_idx_lock);
3832 spin_lock_init(&adev->pcie_idx_lock);
3833 spin_lock_init(&adev->uvd_ctx_idx_lock);
3834 spin_lock_init(&adev->didt_idx_lock);
3835 spin_lock_init(&adev->gc_cac_idx_lock);
3836 spin_lock_init(&adev->se_cac_idx_lock);
3837 spin_lock_init(&adev->audio_endpt_idx_lock);
3838 spin_lock_init(&adev->mm_stats.lock);
3840 INIT_LIST_HEAD(&adev->shadow_list);
3841 mutex_init(&adev->shadow_list_lock);
3843 INIT_LIST_HEAD(&adev->reset_list);
3845 INIT_LIST_HEAD(&adev->ras_list);
3847 INIT_DELAYED_WORK(&adev->delayed_init_work,
3848 amdgpu_device_delayed_init_work_handler);
3849 INIT_DELAYED_WORK(&adev->gfx.gfx_off_delay_work,
3850 amdgpu_device_delay_enable_gfx_off);
3852 INIT_WORK(&adev->xgmi_reset_work, amdgpu_device_xgmi_reset_func);
3854 adev->gfx.gfx_off_req_count = 1;
3855 adev->gfx.gfx_off_residency = 0;
3856 adev->gfx.gfx_off_entrycount = 0;
3857 adev->pm.ac_power = power_supply_is_system_supplied() > 0;
3859 atomic_set(&adev->throttling_logging_enabled, 1);
3861 * If throttling continues, logging will be performed every minute
3862 * to avoid log flooding. "-1" is subtracted since the thermal
3863 * throttling interrupt comes every second. Thus, the total logging
3864 * interval is 59 seconds(retelimited printk interval) + 1(waiting
3865 * for throttling interrupt) = 60 seconds.
3867 ratelimit_state_init(&adev->throttling_logging_rs, (60 - 1) * HZ, 1);
3868 ratelimit_set_flags(&adev->throttling_logging_rs, RATELIMIT_MSG_ON_RELEASE);
3870 /* Registers mapping */
3871 /* TODO: block userspace mapping of io register */
3872 if (adev->asic_type >= CHIP_BONAIRE) {
3873 adev->rmmio_base = pci_resource_start(adev->pdev, 5);
3874 adev->rmmio_size = pci_resource_len(adev->pdev, 5);
3876 adev->rmmio_base = pci_resource_start(adev->pdev, 2);
3877 adev->rmmio_size = pci_resource_len(adev->pdev, 2);
3880 for (i = 0; i < AMD_IP_BLOCK_TYPE_NUM; i++)
3881 atomic_set(&adev->pm.pwr_state[i], POWER_STATE_UNKNOWN);
3883 adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
3884 if (adev->rmmio == NULL) {
3887 DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
3888 DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
3891 * Reset domain needs to be present early, before XGMI hive discovered
3892 * (if any) and intitialized to use reset sem and in_gpu reset flag
3893 * early on during init and before calling to RREG32.
3895 adev->reset_domain = amdgpu_reset_create_reset_domain(SINGLE_DEVICE, "amdgpu-reset-dev");
3896 if (!adev->reset_domain)
3899 /* detect hw virtualization here */
3900 amdgpu_detect_virtualization(adev);
3902 amdgpu_device_get_pcie_info(adev);
3904 r = amdgpu_device_get_job_timeout_settings(adev);
3906 dev_err(adev->dev, "invalid lockup_timeout parameter syntax\n");
3910 /* early init functions */
3911 r = amdgpu_device_ip_early_init(adev);
3915 amdgpu_device_set_mcbp(adev);
3917 /* Get rid of things like offb */
3918 r = drm_aperture_remove_conflicting_pci_framebuffers(adev->pdev, &amdgpu_kms_driver);
3922 /* Enable TMZ based on IP_VERSION */
3923 amdgpu_gmc_tmz_set(adev);
3925 amdgpu_gmc_noretry_set(adev);
3926 /* Need to get xgmi info early to decide the reset behavior*/
3927 if (adev->gmc.xgmi.supported) {
3928 r = adev->gfxhub.funcs->get_xgmi_info(adev);
3933 /* enable PCIE atomic ops */
3934 if (amdgpu_sriov_vf(adev)) {
3935 if (adev->virt.fw_reserve.p_pf2vf)
3936 adev->have_atomics_support = ((struct amd_sriov_msg_pf2vf_info *)
3937 adev->virt.fw_reserve.p_pf2vf)->pcie_atomic_ops_support_flags ==
3938 (PCI_EXP_DEVCAP2_ATOMIC_COMP32 | PCI_EXP_DEVCAP2_ATOMIC_COMP64);
3939 /* APUs w/ gfx9 onwards doesn't reply on PCIe atomics, rather it is a
3940 * internal path natively support atomics, set have_atomics_support to true.
3942 } else if ((adev->flags & AMD_IS_APU) &&
3943 (adev->ip_versions[GC_HWIP][0] > IP_VERSION(9, 0, 0))) {
3944 adev->have_atomics_support = true;
3946 adev->have_atomics_support =
3947 !pci_enable_atomic_ops_to_root(adev->pdev,
3948 PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
3949 PCI_EXP_DEVCAP2_ATOMIC_COMP64);
3952 if (!adev->have_atomics_support)
3953 dev_info(adev->dev, "PCIE atomic ops is not supported\n");
3955 /* doorbell bar mapping and doorbell index init*/
3956 amdgpu_device_doorbell_init(adev);
3958 if (amdgpu_emu_mode == 1) {
3959 /* post the asic on emulation mode */
3960 emu_soc_asic_init(adev);
3961 goto fence_driver_init;
3964 amdgpu_reset_init(adev);
3966 /* detect if we are with an SRIOV vbios */
3968 amdgpu_device_detect_sriov_bios(adev);
3970 /* check if we need to reset the asic
3971 * E.g., driver was not cleanly unloaded previously, etc.
3973 if (!amdgpu_sriov_vf(adev) && amdgpu_asic_need_reset_on_init(adev)) {
3974 if (adev->gmc.xgmi.num_physical_nodes) {
3975 dev_info(adev->dev, "Pending hive reset.\n");
3976 adev->gmc.xgmi.pending_reset = true;
3977 /* Only need to init necessary block for SMU to handle the reset */
3978 for (i = 0; i < adev->num_ip_blocks; i++) {
3979 if (!adev->ip_blocks[i].status.valid)
3981 if (!(adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
3982 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
3983 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH ||
3984 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC)) {
3985 DRM_DEBUG("IP %s disabled for hw_init.\n",
3986 adev->ip_blocks[i].version->funcs->name);
3987 adev->ip_blocks[i].status.hw = true;
3991 tmp = amdgpu_reset_method;
3992 /* It should do a default reset when loading or reloading the driver,
3993 * regardless of the module parameter reset_method.
3995 amdgpu_reset_method = AMD_RESET_METHOD_NONE;
3996 r = amdgpu_asic_reset(adev);
3997 amdgpu_reset_method = tmp;
3999 dev_err(adev->dev, "asic reset on init failed\n");
4005 /* Post card if necessary */
4006 if (amdgpu_device_need_post(adev)) {
4008 dev_err(adev->dev, "no vBIOS found\n");
4012 DRM_INFO("GPU posting now...\n");
4013 r = amdgpu_device_asic_init(adev);
4015 dev_err(adev->dev, "gpu post error!\n");
4021 if (adev->is_atom_fw) {
4022 /* Initialize clocks */
4023 r = amdgpu_atomfirmware_get_clock_info(adev);
4025 dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
4026 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
4030 /* Initialize clocks */
4031 r = amdgpu_atombios_get_clock_info(adev);
4033 dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
4034 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
4037 /* init i2c buses */
4038 if (!amdgpu_device_has_dc_support(adev))
4039 amdgpu_atombios_i2c_init(adev);
4045 r = amdgpu_fence_driver_sw_init(adev);
4047 dev_err(adev->dev, "amdgpu_fence_driver_sw_init failed\n");
4048 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
4052 /* init the mode config */
4053 drm_mode_config_init(adev_to_drm(adev));
4055 r = amdgpu_device_ip_init(adev);
4057 dev_err(adev->dev, "amdgpu_device_ip_init failed\n");
4058 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
4059 goto release_ras_con;
4062 amdgpu_fence_driver_hw_init(adev);
4065 "SE %d, SH per SE %d, CU per SH %d, active_cu_number %d\n",
4066 adev->gfx.config.max_shader_engines,
4067 adev->gfx.config.max_sh_per_se,
4068 adev->gfx.config.max_cu_per_sh,
4069 adev->gfx.cu_info.number);
4071 adev->accel_working = true;
4073 amdgpu_vm_check_compute_bug(adev);
4075 /* Initialize the buffer migration limit. */
4076 if (amdgpu_moverate >= 0)
4077 max_MBps = amdgpu_moverate;
4079 max_MBps = 8; /* Allow 8 MB/s. */
4080 /* Get a log2 for easy divisions. */
4081 adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
4083 r = amdgpu_atombios_sysfs_init(adev);
4085 drm_err(&adev->ddev,
4086 "registering atombios sysfs failed (%d).\n", r);
4088 r = amdgpu_pm_sysfs_init(adev);
4090 DRM_ERROR("registering pm sysfs failed (%d).\n", r);
4092 r = amdgpu_ucode_sysfs_init(adev);
4094 adev->ucode_sysfs_en = false;
4095 DRM_ERROR("Creating firmware sysfs failed (%d).\n", r);
4097 adev->ucode_sysfs_en = true;
4099 r = amdgpu_psp_sysfs_init(adev);
4101 adev->psp_sysfs_en = false;
4102 if (!amdgpu_sriov_vf(adev))
4103 DRM_ERROR("Creating psp sysfs failed\n");
4105 adev->psp_sysfs_en = true;
4108 * Register gpu instance before amdgpu_device_enable_mgpu_fan_boost.
4109 * Otherwise the mgpu fan boost feature will be skipped due to the
4110 * gpu instance is counted less.
4112 amdgpu_register_gpu_instance(adev);
4114 /* enable clockgating, etc. after ib tests, etc. since some blocks require
4115 * explicit gating rather than handling it automatically.
4117 if (!adev->gmc.xgmi.pending_reset) {
4118 r = amdgpu_device_ip_late_init(adev);
4120 dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n");
4121 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
4122 goto release_ras_con;
4125 amdgpu_ras_resume(adev);
4126 queue_delayed_work(system_wq, &adev->delayed_init_work,
4127 msecs_to_jiffies(AMDGPU_RESUME_MS));
4130 if (amdgpu_sriov_vf(adev)) {
4131 amdgpu_virt_release_full_gpu(adev, true);
4132 flush_delayed_work(&adev->delayed_init_work);
4135 r = sysfs_create_files(&adev->dev->kobj, amdgpu_dev_attributes);
4137 dev_err(adev->dev, "Could not create amdgpu device attr\n");
4139 if (IS_ENABLED(CONFIG_PERF_EVENTS))
4140 r = amdgpu_pmu_init(adev);
4142 dev_err(adev->dev, "amdgpu_pmu_init failed\n");
4144 /* Have stored pci confspace at hand for restore in sudden PCI error */
4145 if (amdgpu_device_cache_pci_state(adev->pdev))
4146 pci_restore_state(pdev);
4148 /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
4149 /* this will fail for cards that aren't VGA class devices, just
4151 if ((adev->pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA)
4152 vga_client_register(adev->pdev, amdgpu_device_vga_set_decode);
4154 px = amdgpu_device_supports_px(ddev);
4156 if (px || (!pci_is_thunderbolt_attached(adev->pdev) &&
4157 apple_gmux_detect(NULL, NULL)))
4158 vga_switcheroo_register_client(adev->pdev,
4159 &amdgpu_switcheroo_ops, px);
4162 vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
4164 if (adev->gmc.xgmi.pending_reset)
4165 queue_delayed_work(system_wq, &mgpu_info.delayed_reset_work,
4166 msecs_to_jiffies(AMDGPU_RESUME_MS));
4168 amdgpu_device_check_iommu_direct_map(adev);
4173 if (amdgpu_sriov_vf(adev))
4174 amdgpu_virt_release_full_gpu(adev, true);
4176 /* failed in exclusive mode due to timeout */
4177 if (amdgpu_sriov_vf(adev) &&
4178 !amdgpu_sriov_runtime(adev) &&
4179 amdgpu_virt_mmio_blocked(adev) &&
4180 !amdgpu_virt_wait_reset(adev)) {
4181 dev_err(adev->dev, "VF exclusive mode timeout\n");
4182 /* Don't send request since VF is inactive. */
4183 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
4184 adev->virt.ops = NULL;
4187 amdgpu_release_ras_context(adev);
4190 amdgpu_vf_error_trans_all(adev);
4195 static void amdgpu_device_unmap_mmio(struct amdgpu_device *adev)
4198 /* Clear all CPU mappings pointing to this device */
4199 unmap_mapping_range(adev->ddev.anon_inode->i_mapping, 0, 0, 1);
4201 /* Unmap all mapped bars - Doorbell, registers and VRAM */
4202 amdgpu_device_doorbell_fini(adev);
4204 iounmap(adev->rmmio);
4206 if (adev->mman.aper_base_kaddr)
4207 iounmap(adev->mman.aper_base_kaddr);
4208 adev->mman.aper_base_kaddr = NULL;
4210 /* Memory manager related */
4211 if (!adev->gmc.xgmi.connected_to_cpu && !adev->gmc.is_app_apu) {
4212 arch_phys_wc_del(adev->gmc.vram_mtrr);
4213 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
4218 * amdgpu_device_fini_hw - tear down the driver
4220 * @adev: amdgpu_device pointer
4222 * Tear down the driver info (all asics).
4223 * Called at driver shutdown.
4225 void amdgpu_device_fini_hw(struct amdgpu_device *adev)
4227 dev_info(adev->dev, "amdgpu: finishing device.\n");
4228 flush_delayed_work(&adev->delayed_init_work);
4229 adev->shutdown = true;
4231 /* make sure IB test finished before entering exclusive mode
4232 * to avoid preemption on IB test
4234 if (amdgpu_sriov_vf(adev)) {
4235 amdgpu_virt_request_full_gpu(adev, false);
4236 amdgpu_virt_fini_data_exchange(adev);
4239 /* disable all interrupts */
4240 amdgpu_irq_disable_all(adev);
4241 if (adev->mode_info.mode_config_initialized) {
4242 if (!drm_drv_uses_atomic_modeset(adev_to_drm(adev)))
4243 drm_helper_force_disable_all(adev_to_drm(adev));
4245 drm_atomic_helper_shutdown(adev_to_drm(adev));
4247 amdgpu_fence_driver_hw_fini(adev);
4249 if (adev->mman.initialized)
4250 drain_workqueue(adev->mman.bdev.wq);
4252 if (adev->pm.sysfs_initialized)
4253 amdgpu_pm_sysfs_fini(adev);
4254 if (adev->ucode_sysfs_en)
4255 amdgpu_ucode_sysfs_fini(adev);
4256 if (adev->psp_sysfs_en)
4257 amdgpu_psp_sysfs_fini(adev);
4258 sysfs_remove_files(&adev->dev->kobj, amdgpu_dev_attributes);
4260 /* disable ras feature must before hw fini */
4261 amdgpu_ras_pre_fini(adev);
4263 amdgpu_device_ip_fini_early(adev);
4265 amdgpu_irq_fini_hw(adev);
4267 if (adev->mman.initialized)
4268 ttm_device_clear_dma_mappings(&adev->mman.bdev);
4270 amdgpu_gart_dummy_page_fini(adev);
4272 if (drm_dev_is_unplugged(adev_to_drm(adev)))
4273 amdgpu_device_unmap_mmio(adev);
4277 void amdgpu_device_fini_sw(struct amdgpu_device *adev)
4282 amdgpu_fence_driver_sw_fini(adev);
4283 amdgpu_device_ip_fini(adev);
4284 amdgpu_ucode_release(&adev->firmware.gpu_info_fw);
4285 adev->accel_working = false;
4286 dma_fence_put(rcu_dereference_protected(adev->gang_submit, true));
4288 amdgpu_reset_fini(adev);
4290 /* free i2c buses */
4291 if (!amdgpu_device_has_dc_support(adev))
4292 amdgpu_i2c_fini(adev);
4294 if (amdgpu_emu_mode != 1)
4295 amdgpu_atombios_fini(adev);
4300 px = amdgpu_device_supports_px(adev_to_drm(adev));
4302 if (px || (!pci_is_thunderbolt_attached(adev->pdev) &&
4303 apple_gmux_detect(NULL, NULL)))
4304 vga_switcheroo_unregister_client(adev->pdev);
4307 vga_switcheroo_fini_domain_pm_ops(adev->dev);
4309 if ((adev->pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA)
4310 vga_client_unregister(adev->pdev);
4312 if (drm_dev_enter(adev_to_drm(adev), &idx)) {
4314 iounmap(adev->rmmio);
4316 amdgpu_device_doorbell_fini(adev);
4320 if (IS_ENABLED(CONFIG_PERF_EVENTS))
4321 amdgpu_pmu_fini(adev);
4322 if (adev->mman.discovery_bin)
4323 amdgpu_discovery_fini(adev);
4325 amdgpu_reset_put_reset_domain(adev->reset_domain);
4326 adev->reset_domain = NULL;
4328 kfree(adev->pci_state);
4333 * amdgpu_device_evict_resources - evict device resources
4334 * @adev: amdgpu device object
4336 * Evicts all ttm device resources(vram BOs, gart table) from the lru list
4337 * of the vram memory type. Mainly used for evicting device resources
4341 static int amdgpu_device_evict_resources(struct amdgpu_device *adev)
4345 /* No need to evict vram on APUs for suspend to ram or s2idle */
4346 if ((adev->in_s3 || adev->in_s0ix) && (adev->flags & AMD_IS_APU))
4349 ret = amdgpu_ttm_evict_resources(adev, TTM_PL_VRAM);
4351 DRM_WARN("evicting device resources failed\n");
4359 * amdgpu_device_suspend - initiate device suspend
4361 * @dev: drm dev pointer
4362 * @fbcon : notify the fbdev of suspend
4364 * Puts the hw in the suspend state (all asics).
4365 * Returns 0 for success or an error on failure.
4366 * Called at driver suspend.
4368 int amdgpu_device_suspend(struct drm_device *dev, bool fbcon)
4370 struct amdgpu_device *adev = drm_to_adev(dev);
4373 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
4376 adev->in_suspend = true;
4378 /* Evict the majority of BOs before grabbing the full access */
4379 r = amdgpu_device_evict_resources(adev);
4383 if (amdgpu_sriov_vf(adev)) {
4384 amdgpu_virt_fini_data_exchange(adev);
4385 r = amdgpu_virt_request_full_gpu(adev, false);
4390 if (amdgpu_acpi_smart_shift_update(dev, AMDGPU_SS_DEV_D3))
4391 DRM_WARN("smart shift update failed\n");
4394 drm_fb_helper_set_suspend_unlocked(adev_to_drm(adev)->fb_helper, true);
4396 cancel_delayed_work_sync(&adev->delayed_init_work);
4397 flush_delayed_work(&adev->gfx.gfx_off_delay_work);
4399 amdgpu_ras_suspend(adev);
4401 amdgpu_device_ip_suspend_phase1(adev);
4404 amdgpu_amdkfd_suspend(adev, adev->in_runpm);
4406 r = amdgpu_device_evict_resources(adev);
4410 amdgpu_fence_driver_hw_fini(adev);
4412 amdgpu_device_ip_suspend_phase2(adev);
4414 if (amdgpu_sriov_vf(adev))
4415 amdgpu_virt_release_full_gpu(adev, false);
4421 * amdgpu_device_resume - initiate device resume
4423 * @dev: drm dev pointer
4424 * @fbcon : notify the fbdev of resume
4426 * Bring the hw back to operating state (all asics).
4427 * Returns 0 for success or an error on failure.
4428 * Called at driver resume.
4430 int amdgpu_device_resume(struct drm_device *dev, bool fbcon)
4432 struct amdgpu_device *adev = drm_to_adev(dev);
4435 if (amdgpu_sriov_vf(adev)) {
4436 r = amdgpu_virt_request_full_gpu(adev, true);
4441 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
4445 amdgpu_dpm_gfx_state_change(adev, sGpuChangeState_D0Entry);
4448 if (amdgpu_device_need_post(adev)) {
4449 r = amdgpu_device_asic_init(adev);
4451 dev_err(adev->dev, "amdgpu asic init failed\n");
4454 r = amdgpu_device_ip_resume(adev);
4457 dev_err(adev->dev, "amdgpu_device_ip_resume failed (%d).\n", r);
4460 amdgpu_fence_driver_hw_init(adev);
4462 r = amdgpu_device_ip_late_init(adev);
4466 queue_delayed_work(system_wq, &adev->delayed_init_work,
4467 msecs_to_jiffies(AMDGPU_RESUME_MS));
4469 if (!adev->in_s0ix) {
4470 r = amdgpu_amdkfd_resume(adev, adev->in_runpm);
4476 if (amdgpu_sriov_vf(adev)) {
4477 amdgpu_virt_init_data_exchange(adev);
4478 amdgpu_virt_release_full_gpu(adev, true);
4484 /* Make sure IB tests flushed */
4485 flush_delayed_work(&adev->delayed_init_work);
4488 drm_fb_helper_set_suspend_unlocked(adev_to_drm(adev)->fb_helper, false);
4490 amdgpu_ras_resume(adev);
4492 if (adev->mode_info.num_crtc) {
4494 * Most of the connector probing functions try to acquire runtime pm
4495 * refs to ensure that the GPU is powered on when connector polling is
4496 * performed. Since we're calling this from a runtime PM callback,
4497 * trying to acquire rpm refs will cause us to deadlock.
4499 * Since we're guaranteed to be holding the rpm lock, it's safe to
4500 * temporarily disable the rpm helpers so this doesn't deadlock us.
4503 dev->dev->power.disable_depth++;
4505 if (!adev->dc_enabled)
4506 drm_helper_hpd_irq_event(dev);
4508 drm_kms_helper_hotplug_event(dev);
4510 dev->dev->power.disable_depth--;
4513 adev->in_suspend = false;
4515 if (adev->enable_mes)
4516 amdgpu_mes_self_test(adev);
4518 if (amdgpu_acpi_smart_shift_update(dev, AMDGPU_SS_DEV_D0))
4519 DRM_WARN("smart shift update failed\n");
4525 * amdgpu_device_ip_check_soft_reset - did soft reset succeed
4527 * @adev: amdgpu_device pointer
4529 * The list of all the hardware IPs that make up the asic is walked and
4530 * the check_soft_reset callbacks are run. check_soft_reset determines
4531 * if the asic is still hung or not.
4532 * Returns true if any of the IPs are still in a hung state, false if not.
4534 static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)
4537 bool asic_hang = false;
4539 if (amdgpu_sriov_vf(adev))
4542 if (amdgpu_asic_need_full_reset(adev))
4545 for (i = 0; i < adev->num_ip_blocks; i++) {
4546 if (!adev->ip_blocks[i].status.valid)
4548 if (adev->ip_blocks[i].version->funcs->check_soft_reset)
4549 adev->ip_blocks[i].status.hang =
4550 adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
4551 if (adev->ip_blocks[i].status.hang) {
4552 dev_info(adev->dev, "IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
4560 * amdgpu_device_ip_pre_soft_reset - prepare for soft reset
4562 * @adev: amdgpu_device pointer
4564 * The list of all the hardware IPs that make up the asic is walked and the
4565 * pre_soft_reset callbacks are run if the block is hung. pre_soft_reset
4566 * handles any IP specific hardware or software state changes that are
4567 * necessary for a soft reset to succeed.
4568 * Returns 0 on success, negative error code on failure.
4570 static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)
4574 for (i = 0; i < adev->num_ip_blocks; i++) {
4575 if (!adev->ip_blocks[i].status.valid)
4577 if (adev->ip_blocks[i].status.hang &&
4578 adev->ip_blocks[i].version->funcs->pre_soft_reset) {
4579 r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
4589 * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed
4591 * @adev: amdgpu_device pointer
4593 * Some hardware IPs cannot be soft reset. If they are hung, a full gpu
4594 * reset is necessary to recover.
4595 * Returns true if a full asic reset is required, false if not.
4597 static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)
4601 if (amdgpu_asic_need_full_reset(adev))
4604 for (i = 0; i < adev->num_ip_blocks; i++) {
4605 if (!adev->ip_blocks[i].status.valid)
4607 if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
4608 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
4609 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
4610 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
4611 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
4612 if (adev->ip_blocks[i].status.hang) {
4613 dev_info(adev->dev, "Some block need full reset!\n");
4622 * amdgpu_device_ip_soft_reset - do a soft reset
4624 * @adev: amdgpu_device pointer
4626 * The list of all the hardware IPs that make up the asic is walked and the
4627 * soft_reset callbacks are run if the block is hung. soft_reset handles any
4628 * IP specific hardware or software state changes that are necessary to soft
4630 * Returns 0 on success, negative error code on failure.
4632 static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)
4636 for (i = 0; i < adev->num_ip_blocks; i++) {
4637 if (!adev->ip_blocks[i].status.valid)
4639 if (adev->ip_blocks[i].status.hang &&
4640 adev->ip_blocks[i].version->funcs->soft_reset) {
4641 r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
4651 * amdgpu_device_ip_post_soft_reset - clean up from soft reset
4653 * @adev: amdgpu_device pointer
4655 * The list of all the hardware IPs that make up the asic is walked and the
4656 * post_soft_reset callbacks are run if the asic was hung. post_soft_reset
4657 * handles any IP specific hardware or software state changes that are
4658 * necessary after the IP has been soft reset.
4659 * Returns 0 on success, negative error code on failure.
4661 static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)
4665 for (i = 0; i < adev->num_ip_blocks; i++) {
4666 if (!adev->ip_blocks[i].status.valid)
4668 if (adev->ip_blocks[i].status.hang &&
4669 adev->ip_blocks[i].version->funcs->post_soft_reset)
4670 r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
4679 * amdgpu_device_recover_vram - Recover some VRAM contents
4681 * @adev: amdgpu_device pointer
4683 * Restores the contents of VRAM buffers from the shadows in GTT. Used to
4684 * restore things like GPUVM page tables after a GPU reset where
4685 * the contents of VRAM might be lost.
4688 * 0 on success, negative error code on failure.
4690 static int amdgpu_device_recover_vram(struct amdgpu_device *adev)
4692 struct dma_fence *fence = NULL, *next = NULL;
4693 struct amdgpu_bo *shadow;
4694 struct amdgpu_bo_vm *vmbo;
4697 if (amdgpu_sriov_runtime(adev))
4698 tmo = msecs_to_jiffies(8000);
4700 tmo = msecs_to_jiffies(100);
4702 dev_info(adev->dev, "recover vram bo from shadow start\n");
4703 mutex_lock(&adev->shadow_list_lock);
4704 list_for_each_entry(vmbo, &adev->shadow_list, shadow_list) {
4705 /* If vm is compute context or adev is APU, shadow will be NULL */
4708 shadow = vmbo->shadow;
4710 /* No need to recover an evicted BO */
4711 if (shadow->tbo.resource->mem_type != TTM_PL_TT ||
4712 shadow->tbo.resource->start == AMDGPU_BO_INVALID_OFFSET ||
4713 shadow->parent->tbo.resource->mem_type != TTM_PL_VRAM)
4716 r = amdgpu_bo_restore_shadow(shadow, &next);
4721 tmo = dma_fence_wait_timeout(fence, false, tmo);
4722 dma_fence_put(fence);
4727 } else if (tmo < 0) {
4735 mutex_unlock(&adev->shadow_list_lock);
4738 tmo = dma_fence_wait_timeout(fence, false, tmo);
4739 dma_fence_put(fence);
4741 if (r < 0 || tmo <= 0) {
4742 dev_err(adev->dev, "recover vram bo from shadow failed, r is %ld, tmo is %ld\n", r, tmo);
4746 dev_info(adev->dev, "recover vram bo from shadow done\n");
4752 * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
4754 * @adev: amdgpu_device pointer
4755 * @from_hypervisor: request from hypervisor
4757 * do VF FLR and reinitialize Asic
4758 * return 0 means succeeded otherwise failed
4760 static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
4761 bool from_hypervisor)
4764 struct amdgpu_hive_info *hive = NULL;
4765 int retry_limit = 0;
4768 amdgpu_amdkfd_pre_reset(adev);
4770 if (from_hypervisor)
4771 r = amdgpu_virt_request_full_gpu(adev, true);
4773 r = amdgpu_virt_reset_gpu(adev);
4777 /* Resume IP prior to SMC */
4778 r = amdgpu_device_ip_reinit_early_sriov(adev);
4782 amdgpu_virt_init_data_exchange(adev);
4784 r = amdgpu_device_fw_loading(adev);
4788 /* now we are okay to resume SMC/CP/SDMA */
4789 r = amdgpu_device_ip_reinit_late_sriov(adev);
4793 hive = amdgpu_get_xgmi_hive(adev);
4794 /* Update PSP FW topology after reset */
4795 if (hive && adev->gmc.xgmi.num_physical_nodes > 1)
4796 r = amdgpu_xgmi_update_topology(hive, adev);
4799 amdgpu_put_xgmi_hive(hive);
4802 amdgpu_irq_gpu_reset_resume_helper(adev);
4803 r = amdgpu_ib_ring_tests(adev);
4805 amdgpu_amdkfd_post_reset(adev);
4809 if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
4810 amdgpu_inc_vram_lost(adev);
4811 r = amdgpu_device_recover_vram(adev);
4813 amdgpu_virt_release_full_gpu(adev, true);
4815 if (AMDGPU_RETRY_SRIOV_RESET(r)) {
4816 if (retry_limit < AMDGPU_MAX_RETRY_LIMIT) {
4820 DRM_ERROR("GPU reset retry is beyond the retry limit\n");
4827 * amdgpu_device_has_job_running - check if there is any job in mirror list
4829 * @adev: amdgpu_device pointer
4831 * check if there is any job in mirror list
4833 bool amdgpu_device_has_job_running(struct amdgpu_device *adev)
4836 struct drm_sched_job *job;
4838 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
4839 struct amdgpu_ring *ring = adev->rings[i];
4841 if (!ring || !ring->sched.thread)
4844 spin_lock(&ring->sched.job_list_lock);
4845 job = list_first_entry_or_null(&ring->sched.pending_list,
4846 struct drm_sched_job, list);
4847 spin_unlock(&ring->sched.job_list_lock);
4855 * amdgpu_device_should_recover_gpu - check if we should try GPU recovery
4857 * @adev: amdgpu_device pointer
4859 * Check amdgpu_gpu_recovery and SRIOV status to see if we should try to recover
4862 bool amdgpu_device_should_recover_gpu(struct amdgpu_device *adev)
4865 if (amdgpu_gpu_recovery == 0)
4868 /* Skip soft reset check in fatal error mode */
4869 if (!amdgpu_ras_is_poison_mode_supported(adev))
4872 if (amdgpu_sriov_vf(adev))
4875 if (amdgpu_gpu_recovery == -1) {
4876 switch (adev->asic_type) {
4877 #ifdef CONFIG_DRM_AMDGPU_SI
4884 #ifdef CONFIG_DRM_AMDGPU_CIK
4891 case CHIP_CYAN_SKILLFISH:
4901 dev_info(adev->dev, "GPU recovery disabled.\n");
4905 int amdgpu_device_mode1_reset(struct amdgpu_device *adev)
4910 amdgpu_atombios_scratch_regs_engine_hung(adev, true);
4912 dev_info(adev->dev, "GPU mode1 reset\n");
4915 pci_clear_master(adev->pdev);
4917 amdgpu_device_cache_pci_state(adev->pdev);
4919 if (amdgpu_dpm_is_mode1_reset_supported(adev)) {
4920 dev_info(adev->dev, "GPU smu mode1 reset\n");
4921 ret = amdgpu_dpm_mode1_reset(adev);
4923 dev_info(adev->dev, "GPU psp mode1 reset\n");
4924 ret = psp_gpu_reset(adev);
4928 dev_err(adev->dev, "GPU mode1 reset failed\n");
4930 amdgpu_device_load_pci_state(adev->pdev);
4932 /* wait for asic to come out of reset */
4933 for (i = 0; i < adev->usec_timeout; i++) {
4934 u32 memsize = adev->nbio.funcs->get_memsize(adev);
4936 if (memsize != 0xffffffff)
4941 amdgpu_atombios_scratch_regs_engine_hung(adev, false);
4945 int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,
4946 struct amdgpu_reset_context *reset_context)
4949 struct amdgpu_job *job = NULL;
4950 bool need_full_reset =
4951 test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
4953 if (reset_context->reset_req_dev == adev)
4954 job = reset_context->job;
4956 if (amdgpu_sriov_vf(adev)) {
4957 /* stop the data exchange thread */
4958 amdgpu_virt_fini_data_exchange(adev);
4961 amdgpu_fence_driver_isr_toggle(adev, true);
4963 /* block all schedulers and reset given job's ring */
4964 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
4965 struct amdgpu_ring *ring = adev->rings[i];
4967 if (!ring || !ring->sched.thread)
4970 /*clear job fence from fence drv to avoid force_completion
4971 *leave NULL and vm flush fence in fence drv */
4972 amdgpu_fence_driver_clear_job_fences(ring);
4974 /* after all hw jobs are reset, hw fence is meaningless, so force_completion */
4975 amdgpu_fence_driver_force_completion(ring);
4978 amdgpu_fence_driver_isr_toggle(adev, false);
4981 drm_sched_increase_karma(&job->base);
4983 r = amdgpu_reset_prepare_hwcontext(adev, reset_context);
4984 /* If reset handler not implemented, continue; otherwise return */
4990 /* Don't suspend on bare metal if we are not going to HW reset the ASIC */
4991 if (!amdgpu_sriov_vf(adev)) {
4993 if (!need_full_reset)
4994 need_full_reset = amdgpu_device_ip_need_full_reset(adev);
4996 if (!need_full_reset && amdgpu_gpu_recovery &&
4997 amdgpu_device_ip_check_soft_reset(adev)) {
4998 amdgpu_device_ip_pre_soft_reset(adev);
4999 r = amdgpu_device_ip_soft_reset(adev);
5000 amdgpu_device_ip_post_soft_reset(adev);
5001 if (r || amdgpu_device_ip_check_soft_reset(adev)) {
5002 dev_info(adev->dev, "soft reset failed, will fallback to full reset!\n");
5003 need_full_reset = true;
5007 if (need_full_reset)
5008 r = amdgpu_device_ip_suspend(adev);
5009 if (need_full_reset)
5010 set_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
5012 clear_bit(AMDGPU_NEED_FULL_RESET,
5013 &reset_context->flags);
5019 static int amdgpu_reset_reg_dumps(struct amdgpu_device *adev)
5023 lockdep_assert_held(&adev->reset_domain->sem);
5025 for (i = 0; i < adev->num_regs; i++) {
5026 adev->reset_dump_reg_value[i] = RREG32(adev->reset_dump_reg_list[i]);
5027 trace_amdgpu_reset_reg_dumps(adev->reset_dump_reg_list[i],
5028 adev->reset_dump_reg_value[i]);
5034 #ifdef CONFIG_DEV_COREDUMP
5035 static ssize_t amdgpu_devcoredump_read(char *buffer, loff_t offset,
5036 size_t count, void *data, size_t datalen)
5038 struct drm_printer p;
5039 struct amdgpu_device *adev = data;
5040 struct drm_print_iterator iter;
5045 iter.start = offset;
5046 iter.remain = count;
5048 p = drm_coredump_printer(&iter);
5050 drm_printf(&p, "**** AMDGPU Device Coredump ****\n");
5051 drm_printf(&p, "kernel: " UTS_RELEASE "\n");
5052 drm_printf(&p, "module: " KBUILD_MODNAME "\n");
5053 drm_printf(&p, "time: %lld.%09ld\n", adev->reset_time.tv_sec, adev->reset_time.tv_nsec);
5054 if (adev->reset_task_info.pid)
5055 drm_printf(&p, "process_name: %s PID: %d\n",
5056 adev->reset_task_info.process_name,
5057 adev->reset_task_info.pid);
5059 if (adev->reset_vram_lost)
5060 drm_printf(&p, "VRAM is lost due to GPU reset!\n");
5061 if (adev->num_regs) {
5062 drm_printf(&p, "AMDGPU register dumps:\nOffset: Value:\n");
5064 for (i = 0; i < adev->num_regs; i++)
5065 drm_printf(&p, "0x%08x: 0x%08x\n",
5066 adev->reset_dump_reg_list[i],
5067 adev->reset_dump_reg_value[i]);
5070 return count - iter.remain;
5073 static void amdgpu_devcoredump_free(void *data)
5077 static void amdgpu_reset_capture_coredumpm(struct amdgpu_device *adev)
5079 struct drm_device *dev = adev_to_drm(adev);
5081 ktime_get_ts64(&adev->reset_time);
5082 dev_coredumpm(dev->dev, THIS_MODULE, adev, 0, GFP_KERNEL,
5083 amdgpu_devcoredump_read, amdgpu_devcoredump_free);
5087 int amdgpu_do_asic_reset(struct list_head *device_list_handle,
5088 struct amdgpu_reset_context *reset_context)
5090 struct amdgpu_device *tmp_adev = NULL;
5091 bool need_full_reset, skip_hw_reset, vram_lost = false;
5093 bool gpu_reset_for_dev_remove = 0;
5095 /* Try reset handler method first */
5096 tmp_adev = list_first_entry(device_list_handle, struct amdgpu_device,
5098 amdgpu_reset_reg_dumps(tmp_adev);
5100 reset_context->reset_device_list = device_list_handle;
5101 r = amdgpu_reset_perform_reset(tmp_adev, reset_context);
5102 /* If reset handler not implemented, continue; otherwise return */
5108 /* Reset handler not implemented, use the default method */
5110 test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
5111 skip_hw_reset = test_bit(AMDGPU_SKIP_HW_RESET, &reset_context->flags);
5113 gpu_reset_for_dev_remove =
5114 test_bit(AMDGPU_RESET_FOR_DEVICE_REMOVE, &reset_context->flags) &&
5115 test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
5118 * ASIC reset has to be done on all XGMI hive nodes ASAP
5119 * to allow proper links negotiation in FW (within 1 sec)
5121 if (!skip_hw_reset && need_full_reset) {
5122 list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5123 /* For XGMI run all resets in parallel to speed up the process */
5124 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
5125 tmp_adev->gmc.xgmi.pending_reset = false;
5126 if (!queue_work(system_unbound_wq, &tmp_adev->xgmi_reset_work))
5129 r = amdgpu_asic_reset(tmp_adev);
5132 dev_err(tmp_adev->dev, "ASIC reset failed with error, %d for drm dev, %s",
5133 r, adev_to_drm(tmp_adev)->unique);
5138 /* For XGMI wait for all resets to complete before proceed */
5140 list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5141 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
5142 flush_work(&tmp_adev->xgmi_reset_work);
5143 r = tmp_adev->asic_reset_res;
5151 if (!r && amdgpu_ras_intr_triggered()) {
5152 list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5153 if (tmp_adev->mmhub.ras && tmp_adev->mmhub.ras->ras_block.hw_ops &&
5154 tmp_adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count)
5155 tmp_adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count(tmp_adev);
5158 amdgpu_ras_intr_cleared();
5161 /* Since the mode1 reset affects base ip blocks, the
5162 * phase1 ip blocks need to be resumed. Otherwise there
5163 * will be a BIOS signature error and the psp bootloader
5164 * can't load kdb on the next amdgpu install.
5166 if (gpu_reset_for_dev_remove) {
5167 list_for_each_entry(tmp_adev, device_list_handle, reset_list)
5168 amdgpu_device_ip_resume_phase1(tmp_adev);
5173 list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5174 if (need_full_reset) {
5176 r = amdgpu_device_asic_init(tmp_adev);
5178 dev_warn(tmp_adev->dev, "asic atom init failed!");
5180 dev_info(tmp_adev->dev, "GPU reset succeeded, trying to resume\n");
5181 r = amdgpu_amdkfd_resume_iommu(tmp_adev);
5185 r = amdgpu_device_ip_resume_phase1(tmp_adev);
5189 vram_lost = amdgpu_device_check_vram_lost(tmp_adev);
5190 #ifdef CONFIG_DEV_COREDUMP
5191 tmp_adev->reset_vram_lost = vram_lost;
5192 memset(&tmp_adev->reset_task_info, 0,
5193 sizeof(tmp_adev->reset_task_info));
5194 if (reset_context->job && reset_context->job->vm)
5195 tmp_adev->reset_task_info =
5196 reset_context->job->vm->task_info;
5197 amdgpu_reset_capture_coredumpm(tmp_adev);
5200 DRM_INFO("VRAM is lost due to GPU reset!\n");
5201 amdgpu_inc_vram_lost(tmp_adev);
5204 r = amdgpu_device_fw_loading(tmp_adev);
5208 r = amdgpu_device_ip_resume_phase2(tmp_adev);
5213 amdgpu_device_fill_reset_magic(tmp_adev);
5216 * Add this ASIC as tracked as reset was already
5217 * complete successfully.
5219 amdgpu_register_gpu_instance(tmp_adev);
5221 if (!reset_context->hive &&
5222 tmp_adev->gmc.xgmi.num_physical_nodes > 1)
5223 amdgpu_xgmi_add_device(tmp_adev);
5225 r = amdgpu_device_ip_late_init(tmp_adev);
5229 drm_fb_helper_set_suspend_unlocked(adev_to_drm(tmp_adev)->fb_helper, false);
5232 * The GPU enters bad state once faulty pages
5233 * by ECC has reached the threshold, and ras
5234 * recovery is scheduled next. So add one check
5235 * here to break recovery if it indeed exceeds
5236 * bad page threshold, and remind user to
5237 * retire this GPU or setting one bigger
5238 * bad_page_threshold value to fix this once
5239 * probing driver again.
5241 if (!amdgpu_ras_eeprom_check_err_threshold(tmp_adev)) {
5243 amdgpu_ras_resume(tmp_adev);
5249 /* Update PSP FW topology after reset */
5250 if (reset_context->hive &&
5251 tmp_adev->gmc.xgmi.num_physical_nodes > 1)
5252 r = amdgpu_xgmi_update_topology(
5253 reset_context->hive, tmp_adev);
5259 amdgpu_irq_gpu_reset_resume_helper(tmp_adev);
5260 r = amdgpu_ib_ring_tests(tmp_adev);
5262 dev_err(tmp_adev->dev, "ib ring test failed (%d).\n", r);
5263 need_full_reset = true;
5270 r = amdgpu_device_recover_vram(tmp_adev);
5272 tmp_adev->asic_reset_res = r;
5276 if (need_full_reset)
5277 set_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
5279 clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
5283 static void amdgpu_device_set_mp1_state(struct amdgpu_device *adev)
5286 switch (amdgpu_asic_reset_method(adev)) {
5287 case AMD_RESET_METHOD_MODE1:
5288 adev->mp1_state = PP_MP1_STATE_SHUTDOWN;
5290 case AMD_RESET_METHOD_MODE2:
5291 adev->mp1_state = PP_MP1_STATE_RESET;
5294 adev->mp1_state = PP_MP1_STATE_NONE;
5299 static void amdgpu_device_unset_mp1_state(struct amdgpu_device *adev)
5301 amdgpu_vf_error_trans_all(adev);
5302 adev->mp1_state = PP_MP1_STATE_NONE;
5305 static void amdgpu_device_resume_display_audio(struct amdgpu_device *adev)
5307 struct pci_dev *p = NULL;
5309 p = pci_get_domain_bus_and_slot(pci_domain_nr(adev->pdev->bus),
5310 adev->pdev->bus->number, 1);
5312 pm_runtime_enable(&(p->dev));
5313 pm_runtime_resume(&(p->dev));
5319 static int amdgpu_device_suspend_display_audio(struct amdgpu_device *adev)
5321 enum amd_reset_method reset_method;
5322 struct pci_dev *p = NULL;
5326 * For now, only BACO and mode1 reset are confirmed
5327 * to suffer the audio issue without proper suspended.
5329 reset_method = amdgpu_asic_reset_method(adev);
5330 if ((reset_method != AMD_RESET_METHOD_BACO) &&
5331 (reset_method != AMD_RESET_METHOD_MODE1))
5334 p = pci_get_domain_bus_and_slot(pci_domain_nr(adev->pdev->bus),
5335 adev->pdev->bus->number, 1);
5339 expires = pm_runtime_autosuspend_expiration(&(p->dev));
5342 * If we cannot get the audio device autosuspend delay,
5343 * a fixed 4S interval will be used. Considering 3S is
5344 * the audio controller default autosuspend delay setting.
5345 * 4S used here is guaranteed to cover that.
5347 expires = ktime_get_mono_fast_ns() + NSEC_PER_SEC * 4ULL;
5349 while (!pm_runtime_status_suspended(&(p->dev))) {
5350 if (!pm_runtime_suspend(&(p->dev)))
5353 if (expires < ktime_get_mono_fast_ns()) {
5354 dev_warn(adev->dev, "failed to suspend display audio\n");
5356 /* TODO: abort the succeeding gpu reset? */
5361 pm_runtime_disable(&(p->dev));
5367 static inline void amdgpu_device_stop_pending_resets(struct amdgpu_device *adev)
5369 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
5371 #if defined(CONFIG_DEBUG_FS)
5372 if (!amdgpu_sriov_vf(adev))
5373 cancel_work(&adev->reset_work);
5377 cancel_work(&adev->kfd.reset_work);
5379 if (amdgpu_sriov_vf(adev))
5380 cancel_work(&adev->virt.flr_work);
5382 if (con && adev->ras_enabled)
5383 cancel_work(&con->recovery_work);
5388 * amdgpu_device_gpu_recover - reset the asic and recover scheduler
5390 * @adev: amdgpu_device pointer
5391 * @job: which job trigger hang
5392 * @reset_context: amdgpu reset context pointer
5394 * Attempt to reset the GPU if it has hung (all asics).
5395 * Attempt to do soft-reset or full-reset and reinitialize Asic
5396 * Returns 0 for success or an error on failure.
5399 int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
5400 struct amdgpu_job *job,
5401 struct amdgpu_reset_context *reset_context)
5403 struct list_head device_list, *device_list_handle = NULL;
5404 bool job_signaled = false;
5405 struct amdgpu_hive_info *hive = NULL;
5406 struct amdgpu_device *tmp_adev = NULL;
5408 bool need_emergency_restart = false;
5409 bool audio_suspended = false;
5410 bool gpu_reset_for_dev_remove = false;
5412 gpu_reset_for_dev_remove =
5413 test_bit(AMDGPU_RESET_FOR_DEVICE_REMOVE, &reset_context->flags) &&
5414 test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
5417 * Special case: RAS triggered and full reset isn't supported
5419 need_emergency_restart = amdgpu_ras_need_emergency_restart(adev);
5422 * Flush RAM to disk so that after reboot
5423 * the user can read log and see why the system rebooted.
5425 if (need_emergency_restart && amdgpu_ras_get_context(adev)->reboot) {
5426 DRM_WARN("Emergency reboot.");
5429 emergency_restart();
5432 dev_info(adev->dev, "GPU %s begin!\n",
5433 need_emergency_restart ? "jobs stop":"reset");
5435 if (!amdgpu_sriov_vf(adev))
5436 hive = amdgpu_get_xgmi_hive(adev);
5438 mutex_lock(&hive->hive_lock);
5440 reset_context->job = job;
5441 reset_context->hive = hive;
5443 * Build list of devices to reset.
5444 * In case we are in XGMI hive mode, resort the device list
5445 * to put adev in the 1st position.
5447 INIT_LIST_HEAD(&device_list);
5448 if (!amdgpu_sriov_vf(adev) && (adev->gmc.xgmi.num_physical_nodes > 1)) {
5449 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
5450 list_add_tail(&tmp_adev->reset_list, &device_list);
5451 if (gpu_reset_for_dev_remove && adev->shutdown)
5452 tmp_adev->shutdown = true;
5454 if (!list_is_first(&adev->reset_list, &device_list))
5455 list_rotate_to_front(&adev->reset_list, &device_list);
5456 device_list_handle = &device_list;
5458 list_add_tail(&adev->reset_list, &device_list);
5459 device_list_handle = &device_list;
5462 /* We need to lock reset domain only once both for XGMI and single device */
5463 tmp_adev = list_first_entry(device_list_handle, struct amdgpu_device,
5465 amdgpu_device_lock_reset_domain(tmp_adev->reset_domain);
5467 /* block all schedulers and reset given job's ring */
5468 list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5470 amdgpu_device_set_mp1_state(tmp_adev);
5473 * Try to put the audio codec into suspend state
5474 * before gpu reset started.
5476 * Due to the power domain of the graphics device
5477 * is shared with AZ power domain. Without this,
5478 * we may change the audio hardware from behind
5479 * the audio driver's back. That will trigger
5480 * some audio codec errors.
5482 if (!amdgpu_device_suspend_display_audio(tmp_adev))
5483 audio_suspended = true;
5485 amdgpu_ras_set_error_query_ready(tmp_adev, false);
5487 cancel_delayed_work_sync(&tmp_adev->delayed_init_work);
5489 if (!amdgpu_sriov_vf(tmp_adev))
5490 amdgpu_amdkfd_pre_reset(tmp_adev);
5493 * Mark these ASICs to be reseted as untracked first
5494 * And add them back after reset completed
5496 amdgpu_unregister_gpu_instance(tmp_adev);
5498 drm_fb_helper_set_suspend_unlocked(adev_to_drm(tmp_adev)->fb_helper, true);
5500 /* disable ras on ALL IPs */
5501 if (!need_emergency_restart &&
5502 amdgpu_device_ip_need_full_reset(tmp_adev))
5503 amdgpu_ras_suspend(tmp_adev);
5505 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
5506 struct amdgpu_ring *ring = tmp_adev->rings[i];
5508 if (!ring || !ring->sched.thread)
5511 drm_sched_stop(&ring->sched, job ? &job->base : NULL);
5513 if (need_emergency_restart)
5514 amdgpu_job_stop_all_jobs_on_sched(&ring->sched);
5516 atomic_inc(&tmp_adev->gpu_reset_counter);
5519 if (need_emergency_restart)
5520 goto skip_sched_resume;
5523 * Must check guilty signal here since after this point all old
5524 * HW fences are force signaled.
5526 * job->base holds a reference to parent fence
5528 if (job && dma_fence_is_signaled(&job->hw_fence)) {
5529 job_signaled = true;
5530 dev_info(adev->dev, "Guilty job already signaled, skipping HW reset");
5534 retry: /* Rest of adevs pre asic reset from XGMI hive. */
5535 list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5536 if (gpu_reset_for_dev_remove) {
5537 /* Workaroud for ASICs need to disable SMC first */
5538 amdgpu_device_smu_fini_early(tmp_adev);
5540 r = amdgpu_device_pre_asic_reset(tmp_adev, reset_context);
5541 /*TODO Should we stop ?*/
5543 dev_err(tmp_adev->dev, "GPU pre asic reset failed with err, %d for drm dev, %s ",
5544 r, adev_to_drm(tmp_adev)->unique);
5545 tmp_adev->asic_reset_res = r;
5549 * Drop all pending non scheduler resets. Scheduler resets
5550 * were already dropped during drm_sched_stop
5552 amdgpu_device_stop_pending_resets(tmp_adev);
5555 /* Actual ASIC resets if needed.*/
5556 /* Host driver will handle XGMI hive reset for SRIOV */
5557 if (amdgpu_sriov_vf(adev)) {
5558 r = amdgpu_device_reset_sriov(adev, job ? false : true);
5560 adev->asic_reset_res = r;
5562 /* Aldebaran and gfx_11_0_3 support ras in SRIOV, so need resume ras during reset */
5563 if (adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 2) ||
5564 adev->ip_versions[GC_HWIP][0] == IP_VERSION(11, 0, 3))
5565 amdgpu_ras_resume(adev);
5567 r = amdgpu_do_asic_reset(device_list_handle, reset_context);
5568 if (r && r == -EAGAIN)
5571 if (!r && gpu_reset_for_dev_remove)
5577 /* Post ASIC reset for all devs .*/
5578 list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5580 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
5581 struct amdgpu_ring *ring = tmp_adev->rings[i];
5583 if (!ring || !ring->sched.thread)
5586 drm_sched_start(&ring->sched, true);
5589 if (adev->enable_mes && adev->ip_versions[GC_HWIP][0] != IP_VERSION(11, 0, 3))
5590 amdgpu_mes_self_test(tmp_adev);
5592 if (!drm_drv_uses_atomic_modeset(adev_to_drm(tmp_adev)) && !job_signaled) {
5593 drm_helper_resume_force_mode(adev_to_drm(tmp_adev));
5596 if (tmp_adev->asic_reset_res)
5597 r = tmp_adev->asic_reset_res;
5599 tmp_adev->asic_reset_res = 0;
5602 /* bad news, how to tell it to userspace ? */
5603 dev_info(tmp_adev->dev, "GPU reset(%d) failed\n", atomic_read(&tmp_adev->gpu_reset_counter));
5604 amdgpu_vf_error_put(tmp_adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r);
5606 dev_info(tmp_adev->dev, "GPU reset(%d) succeeded!\n", atomic_read(&tmp_adev->gpu_reset_counter));
5607 if (amdgpu_acpi_smart_shift_update(adev_to_drm(tmp_adev), AMDGPU_SS_DEV_D0))
5608 DRM_WARN("smart shift update failed\n");
5613 list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5614 /* unlock kfd: SRIOV would do it separately */
5615 if (!need_emergency_restart && !amdgpu_sriov_vf(tmp_adev))
5616 amdgpu_amdkfd_post_reset(tmp_adev);
5618 /* kfd_post_reset will do nothing if kfd device is not initialized,
5619 * need to bring up kfd here if it's not be initialized before
5621 if (!adev->kfd.init_complete)
5622 amdgpu_amdkfd_device_init(adev);
5624 if (audio_suspended)
5625 amdgpu_device_resume_display_audio(tmp_adev);
5627 amdgpu_device_unset_mp1_state(tmp_adev);
5629 amdgpu_ras_set_error_query_ready(tmp_adev, true);
5633 tmp_adev = list_first_entry(device_list_handle, struct amdgpu_device,
5635 amdgpu_device_unlock_reset_domain(tmp_adev->reset_domain);
5638 mutex_unlock(&hive->hive_lock);
5639 amdgpu_put_xgmi_hive(hive);
5643 dev_info(adev->dev, "GPU reset end with ret = %d\n", r);
5645 atomic_set(&adev->reset_domain->reset_res, r);
5650 * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot
5652 * @adev: amdgpu_device pointer
5654 * Fetchs and stores in the driver the PCIE capabilities (gen speed
5655 * and lanes) of the slot the device is in. Handles APUs and
5656 * virtualized environments where PCIE config space may not be available.
5658 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev)
5660 struct pci_dev *pdev;
5661 enum pci_bus_speed speed_cap, platform_speed_cap;
5662 enum pcie_link_width platform_link_width;
5664 if (amdgpu_pcie_gen_cap)
5665 adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
5667 if (amdgpu_pcie_lane_cap)
5668 adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
5670 /* covers APUs as well */
5671 if (pci_is_root_bus(adev->pdev->bus) && !amdgpu_passthrough(adev)) {
5672 if (adev->pm.pcie_gen_mask == 0)
5673 adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
5674 if (adev->pm.pcie_mlw_mask == 0)
5675 adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
5679 if (adev->pm.pcie_gen_mask && adev->pm.pcie_mlw_mask)
5682 pcie_bandwidth_available(adev->pdev, NULL,
5683 &platform_speed_cap, &platform_link_width);
5685 if (adev->pm.pcie_gen_mask == 0) {
5688 speed_cap = pcie_get_speed_cap(pdev);
5689 if (speed_cap == PCI_SPEED_UNKNOWN) {
5690 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5691 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5692 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
5694 if (speed_cap == PCIE_SPEED_32_0GT)
5695 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5696 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5697 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
5698 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4 |
5699 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN5);
5700 else if (speed_cap == PCIE_SPEED_16_0GT)
5701 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5702 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5703 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
5704 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4);
5705 else if (speed_cap == PCIE_SPEED_8_0GT)
5706 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5707 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5708 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
5709 else if (speed_cap == PCIE_SPEED_5_0GT)
5710 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5711 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2);
5713 adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1;
5716 if (platform_speed_cap == PCI_SPEED_UNKNOWN) {
5717 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5718 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
5720 if (platform_speed_cap == PCIE_SPEED_32_0GT)
5721 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5722 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5723 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
5724 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4 |
5725 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN5);
5726 else if (platform_speed_cap == PCIE_SPEED_16_0GT)
5727 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5728 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5729 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
5730 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4);
5731 else if (platform_speed_cap == PCIE_SPEED_8_0GT)
5732 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5733 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5734 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3);
5735 else if (platform_speed_cap == PCIE_SPEED_5_0GT)
5736 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5737 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
5739 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
5743 if (adev->pm.pcie_mlw_mask == 0) {
5744 if (platform_link_width == PCIE_LNK_WIDTH_UNKNOWN) {
5745 adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK;
5747 switch (platform_link_width) {
5749 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
5750 CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
5751 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
5752 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
5753 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5754 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5755 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5758 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
5759 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
5760 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
5761 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5762 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5763 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5766 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
5767 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
5768 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5769 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5770 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5773 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
5774 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5775 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5776 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5779 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5780 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5781 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5784 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5785 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5788 adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;
5798 * amdgpu_device_is_peer_accessible - Check peer access through PCIe BAR
5800 * @adev: amdgpu_device pointer
5801 * @peer_adev: amdgpu_device pointer for peer device trying to access @adev
5803 * Return true if @peer_adev can access (DMA) @adev through the PCIe
5804 * BAR, i.e. @adev is "large BAR" and the BAR matches the DMA mask of
5807 bool amdgpu_device_is_peer_accessible(struct amdgpu_device *adev,
5808 struct amdgpu_device *peer_adev)
5810 #ifdef CONFIG_HSA_AMD_P2P
5811 uint64_t address_mask = peer_adev->dev->dma_mask ?
5812 ~*peer_adev->dev->dma_mask : ~((1ULL << 32) - 1);
5813 resource_size_t aper_limit =
5814 adev->gmc.aper_base + adev->gmc.aper_size - 1;
5816 !adev->gmc.xgmi.connected_to_cpu &&
5817 !(pci_p2pdma_distance(adev->pdev, peer_adev->dev, false) < 0);
5819 return pcie_p2p && p2p_access && (adev->gmc.visible_vram_size &&
5820 adev->gmc.real_vram_size == adev->gmc.visible_vram_size &&
5821 !(adev->gmc.aper_base & address_mask ||
5822 aper_limit & address_mask));
5828 int amdgpu_device_baco_enter(struct drm_device *dev)
5830 struct amdgpu_device *adev = drm_to_adev(dev);
5831 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
5833 if (!amdgpu_device_supports_baco(dev))
5836 if (ras && adev->ras_enabled &&
5837 adev->nbio.funcs->enable_doorbell_interrupt)
5838 adev->nbio.funcs->enable_doorbell_interrupt(adev, false);
5840 return amdgpu_dpm_baco_enter(adev);
5843 int amdgpu_device_baco_exit(struct drm_device *dev)
5845 struct amdgpu_device *adev = drm_to_adev(dev);
5846 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
5849 if (!amdgpu_device_supports_baco(dev))
5852 ret = amdgpu_dpm_baco_exit(adev);
5856 if (ras && adev->ras_enabled &&
5857 adev->nbio.funcs->enable_doorbell_interrupt)
5858 adev->nbio.funcs->enable_doorbell_interrupt(adev, true);
5860 if (amdgpu_passthrough(adev) &&
5861 adev->nbio.funcs->clear_doorbell_interrupt)
5862 adev->nbio.funcs->clear_doorbell_interrupt(adev);
5868 * amdgpu_pci_error_detected - Called when a PCI error is detected.
5869 * @pdev: PCI device struct
5870 * @state: PCI channel state
5872 * Description: Called when a PCI error is detected.
5874 * Return: PCI_ERS_RESULT_NEED_RESET or PCI_ERS_RESULT_DISCONNECT.
5876 pci_ers_result_t amdgpu_pci_error_detected(struct pci_dev *pdev, pci_channel_state_t state)
5878 struct drm_device *dev = pci_get_drvdata(pdev);
5879 struct amdgpu_device *adev = drm_to_adev(dev);
5882 DRM_INFO("PCI error: detected callback, state(%d)!!\n", state);
5884 if (adev->gmc.xgmi.num_physical_nodes > 1) {
5885 DRM_WARN("No support for XGMI hive yet...");
5886 return PCI_ERS_RESULT_DISCONNECT;
5889 adev->pci_channel_state = state;
5892 case pci_channel_io_normal:
5893 return PCI_ERS_RESULT_CAN_RECOVER;
5894 /* Fatal error, prepare for slot reset */
5895 case pci_channel_io_frozen:
5897 * Locking adev->reset_domain->sem will prevent any external access
5898 * to GPU during PCI error recovery
5900 amdgpu_device_lock_reset_domain(adev->reset_domain);
5901 amdgpu_device_set_mp1_state(adev);
5904 * Block any work scheduling as we do for regular GPU reset
5905 * for the duration of the recovery
5907 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
5908 struct amdgpu_ring *ring = adev->rings[i];
5910 if (!ring || !ring->sched.thread)
5913 drm_sched_stop(&ring->sched, NULL);
5915 atomic_inc(&adev->gpu_reset_counter);
5916 return PCI_ERS_RESULT_NEED_RESET;
5917 case pci_channel_io_perm_failure:
5918 /* Permanent error, prepare for device removal */
5919 return PCI_ERS_RESULT_DISCONNECT;
5922 return PCI_ERS_RESULT_NEED_RESET;
5926 * amdgpu_pci_mmio_enabled - Enable MMIO and dump debug registers
5927 * @pdev: pointer to PCI device
5929 pci_ers_result_t amdgpu_pci_mmio_enabled(struct pci_dev *pdev)
5932 DRM_INFO("PCI error: mmio enabled callback!!\n");
5934 /* TODO - dump whatever for debugging purposes */
5936 /* This called only if amdgpu_pci_error_detected returns
5937 * PCI_ERS_RESULT_CAN_RECOVER. Read/write to the device still
5938 * works, no need to reset slot.
5941 return PCI_ERS_RESULT_RECOVERED;
5945 * amdgpu_pci_slot_reset - Called when PCI slot has been reset.
5946 * @pdev: PCI device struct
5948 * Description: This routine is called by the pci error recovery
5949 * code after the PCI slot has been reset, just before we
5950 * should resume normal operations.
5952 pci_ers_result_t amdgpu_pci_slot_reset(struct pci_dev *pdev)
5954 struct drm_device *dev = pci_get_drvdata(pdev);
5955 struct amdgpu_device *adev = drm_to_adev(dev);
5957 struct amdgpu_reset_context reset_context;
5959 struct list_head device_list;
5961 DRM_INFO("PCI error: slot reset callback!!\n");
5963 memset(&reset_context, 0, sizeof(reset_context));
5965 INIT_LIST_HEAD(&device_list);
5966 list_add_tail(&adev->reset_list, &device_list);
5968 /* wait for asic to come out of reset */
5971 /* Restore PCI confspace */
5972 amdgpu_device_load_pci_state(pdev);
5974 /* confirm ASIC came out of reset */
5975 for (i = 0; i < adev->usec_timeout; i++) {
5976 memsize = amdgpu_asic_get_config_memsize(adev);
5978 if (memsize != 0xffffffff)
5982 if (memsize == 0xffffffff) {
5987 reset_context.method = AMD_RESET_METHOD_NONE;
5988 reset_context.reset_req_dev = adev;
5989 set_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
5990 set_bit(AMDGPU_SKIP_HW_RESET, &reset_context.flags);
5992 adev->no_hw_access = true;
5993 r = amdgpu_device_pre_asic_reset(adev, &reset_context);
5994 adev->no_hw_access = false;
5998 r = amdgpu_do_asic_reset(&device_list, &reset_context);
6002 if (amdgpu_device_cache_pci_state(adev->pdev))
6003 pci_restore_state(adev->pdev);
6005 DRM_INFO("PCIe error recovery succeeded\n");
6007 DRM_ERROR("PCIe error recovery failed, err:%d", r);
6008 amdgpu_device_unset_mp1_state(adev);
6009 amdgpu_device_unlock_reset_domain(adev->reset_domain);
6012 return r ? PCI_ERS_RESULT_DISCONNECT : PCI_ERS_RESULT_RECOVERED;
6016 * amdgpu_pci_resume() - resume normal ops after PCI reset
6017 * @pdev: pointer to PCI device
6019 * Called when the error recovery driver tells us that its
6020 * OK to resume normal operation.
6022 void amdgpu_pci_resume(struct pci_dev *pdev)
6024 struct drm_device *dev = pci_get_drvdata(pdev);
6025 struct amdgpu_device *adev = drm_to_adev(dev);
6029 DRM_INFO("PCI error: resume callback!!\n");
6031 /* Only continue execution for the case of pci_channel_io_frozen */
6032 if (adev->pci_channel_state != pci_channel_io_frozen)
6035 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
6036 struct amdgpu_ring *ring = adev->rings[i];
6038 if (!ring || !ring->sched.thread)
6041 drm_sched_start(&ring->sched, true);
6044 amdgpu_device_unset_mp1_state(adev);
6045 amdgpu_device_unlock_reset_domain(adev->reset_domain);
6048 bool amdgpu_device_cache_pci_state(struct pci_dev *pdev)
6050 struct drm_device *dev = pci_get_drvdata(pdev);
6051 struct amdgpu_device *adev = drm_to_adev(dev);
6054 r = pci_save_state(pdev);
6056 kfree(adev->pci_state);
6058 adev->pci_state = pci_store_saved_state(pdev);
6060 if (!adev->pci_state) {
6061 DRM_ERROR("Failed to store PCI saved state");
6065 DRM_WARN("Failed to save PCI state, err:%d\n", r);
6072 bool amdgpu_device_load_pci_state(struct pci_dev *pdev)
6074 struct drm_device *dev = pci_get_drvdata(pdev);
6075 struct amdgpu_device *adev = drm_to_adev(dev);
6078 if (!adev->pci_state)
6081 r = pci_load_saved_state(pdev, adev->pci_state);
6084 pci_restore_state(pdev);
6086 DRM_WARN("Failed to load PCI state, err:%d\n", r);
6093 void amdgpu_device_flush_hdp(struct amdgpu_device *adev,
6094 struct amdgpu_ring *ring)
6096 #ifdef CONFIG_X86_64
6097 if ((adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev))
6100 if (adev->gmc.xgmi.connected_to_cpu)
6103 if (ring && ring->funcs->emit_hdp_flush)
6104 amdgpu_ring_emit_hdp_flush(ring);
6106 amdgpu_asic_flush_hdp(adev, ring);
6109 void amdgpu_device_invalidate_hdp(struct amdgpu_device *adev,
6110 struct amdgpu_ring *ring)
6112 #ifdef CONFIG_X86_64
6113 if ((adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev))
6116 if (adev->gmc.xgmi.connected_to_cpu)
6119 amdgpu_asic_invalidate_hdp(adev, ring);
6122 int amdgpu_in_reset(struct amdgpu_device *adev)
6124 return atomic_read(&adev->reset_domain->in_gpu_reset);
6128 * amdgpu_device_halt() - bring hardware to some kind of halt state
6130 * @adev: amdgpu_device pointer
6132 * Bring hardware to some kind of halt state so that no one can touch it
6133 * any more. It will help to maintain error context when error occurred.
6134 * Compare to a simple hang, the system will keep stable at least for SSH
6135 * access. Then it should be trivial to inspect the hardware state and
6136 * see what's going on. Implemented as following:
6138 * 1. drm_dev_unplug() makes device inaccessible to user space(IOCTLs, etc),
6139 * clears all CPU mappings to device, disallows remappings through page faults
6140 * 2. amdgpu_irq_disable_all() disables all interrupts
6141 * 3. amdgpu_fence_driver_hw_fini() signals all HW fences
6142 * 4. set adev->no_hw_access to avoid potential crashes after setp 5
6143 * 5. amdgpu_device_unmap_mmio() clears all MMIO mappings
6144 * 6. pci_disable_device() and pci_wait_for_pending_transaction()
6145 * flush any in flight DMA operations
6147 void amdgpu_device_halt(struct amdgpu_device *adev)
6149 struct pci_dev *pdev = adev->pdev;
6150 struct drm_device *ddev = adev_to_drm(adev);
6152 amdgpu_xcp_dev_unplug(adev);
6153 drm_dev_unplug(ddev);
6155 amdgpu_irq_disable_all(adev);
6157 amdgpu_fence_driver_hw_fini(adev);
6159 adev->no_hw_access = true;
6161 amdgpu_device_unmap_mmio(adev);
6163 pci_disable_device(pdev);
6164 pci_wait_for_pending_transaction(pdev);
6167 u32 amdgpu_device_pcie_port_rreg(struct amdgpu_device *adev,
6170 unsigned long flags, address, data;
6173 address = adev->nbio.funcs->get_pcie_port_index_offset(adev);
6174 data = adev->nbio.funcs->get_pcie_port_data_offset(adev);
6176 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
6177 WREG32(address, reg * 4);
6178 (void)RREG32(address);
6180 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
6184 void amdgpu_device_pcie_port_wreg(struct amdgpu_device *adev,
6187 unsigned long flags, address, data;
6189 address = adev->nbio.funcs->get_pcie_port_index_offset(adev);
6190 data = adev->nbio.funcs->get_pcie_port_data_offset(adev);
6192 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
6193 WREG32(address, reg * 4);
6194 (void)RREG32(address);
6197 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
6201 * amdgpu_device_switch_gang - switch to a new gang
6202 * @adev: amdgpu_device pointer
6203 * @gang: the gang to switch to
6205 * Try to switch to a new gang.
6206 * Returns: NULL if we switched to the new gang or a reference to the current
6209 struct dma_fence *amdgpu_device_switch_gang(struct amdgpu_device *adev,
6210 struct dma_fence *gang)
6212 struct dma_fence *old = NULL;
6217 old = dma_fence_get_rcu_safe(&adev->gang_submit);
6223 if (!dma_fence_is_signaled(old))
6226 } while (cmpxchg((struct dma_fence __force **)&adev->gang_submit,
6233 bool amdgpu_device_has_display_hardware(struct amdgpu_device *adev)
6235 switch (adev->asic_type) {
6236 #ifdef CONFIG_DRM_AMDGPU_SI
6240 /* chips with no display hardware */
6242 #ifdef CONFIG_DRM_AMDGPU_SI
6248 #ifdef CONFIG_DRM_AMDGPU_CIK
6257 case CHIP_POLARIS10:
6258 case CHIP_POLARIS11:
6259 case CHIP_POLARIS12:
6263 /* chips with display hardware */
6267 if (!adev->ip_versions[DCE_HWIP][0] ||
6268 (adev->harvest_ip_mask & AMD_HARVEST_IP_DMU_MASK))
6274 uint32_t amdgpu_device_wait_on_rreg(struct amdgpu_device *adev,
6275 uint32_t inst, uint32_t reg_addr, char reg_name[],
6276 uint32_t expected_value, uint32_t mask)
6280 uint32_t tmp_ = RREG32(reg_addr);
6281 uint32_t loop = adev->usec_timeout;
6283 while ((tmp_ & (mask)) != (expected_value)) {
6285 loop = adev->usec_timeout;
6289 tmp_ = RREG32(reg_addr);
6292 DRM_WARN("Register(%d) [%s] failed to reach value 0x%08x != 0x%08xn",
6293 inst, reg_name, (uint32_t)expected_value,
6294 (uint32_t)(tmp_ & (mask)));