2 * Copyright 2008 Advanced Micro Devices, Inc.
3 * Copyright 2008 Red Hat Inc.
4 * Copyright 2009 Jerome Glisse.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22 * OTHER DEALINGS IN THE SOFTWARE.
24 * Authors: Dave Airlie
28 #include <linux/power_supply.h>
29 #include <linux/kthread.h>
30 #include <linux/module.h>
31 #include <linux/console.h>
32 #include <linux/slab.h>
34 #include <drm/drm_atomic_helper.h>
35 #include <drm/drm_probe_helper.h>
36 #include <drm/amdgpu_drm.h>
37 #include <linux/vgaarb.h>
38 #include <linux/vga_switcheroo.h>
39 #include <linux/efi.h>
41 #include "amdgpu_trace.h"
42 #include "amdgpu_i2c.h"
44 #include "amdgpu_atombios.h"
45 #include "amdgpu_atomfirmware.h"
47 #ifdef CONFIG_DRM_AMDGPU_SI
50 #ifdef CONFIG_DRM_AMDGPU_CIK
56 #include "bif/bif_4_1_d.h"
57 #include <linux/pci.h>
58 #include <linux/firmware.h>
59 #include "amdgpu_vf_error.h"
61 #include "amdgpu_amdkfd.h"
62 #include "amdgpu_pm.h"
64 #include "amdgpu_xgmi.h"
65 #include "amdgpu_ras.h"
66 #include "amdgpu_pmu.h"
68 #include <linux/suspend.h>
69 #include <drm/task_barrier.h>
71 MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin");
72 MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin");
73 MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin");
74 MODULE_FIRMWARE("amdgpu/picasso_gpu_info.bin");
75 MODULE_FIRMWARE("amdgpu/raven2_gpu_info.bin");
76 MODULE_FIRMWARE("amdgpu/arcturus_gpu_info.bin");
77 MODULE_FIRMWARE("amdgpu/renoir_gpu_info.bin");
78 MODULE_FIRMWARE("amdgpu/navi10_gpu_info.bin");
79 MODULE_FIRMWARE("amdgpu/navi14_gpu_info.bin");
80 MODULE_FIRMWARE("amdgpu/navi12_gpu_info.bin");
82 #define AMDGPU_RESUME_MS 2000
84 const char *amdgpu_asic_name[] = {
117 * DOC: pcie_replay_count
119 * The amdgpu driver provides a sysfs API for reporting the total number
120 * of PCIe replays (NAKs)
121 * The file pcie_replay_count is used for this and returns the total
122 * number of replays as a sum of the NAKs generated and NAKs received
125 static ssize_t amdgpu_device_get_pcie_replay_count(struct device *dev,
126 struct device_attribute *attr, char *buf)
128 struct drm_device *ddev = dev_get_drvdata(dev);
129 struct amdgpu_device *adev = ddev->dev_private;
130 uint64_t cnt = amdgpu_asic_get_pcie_replay_count(adev);
132 return snprintf(buf, PAGE_SIZE, "%llu\n", cnt);
135 static DEVICE_ATTR(pcie_replay_count, S_IRUGO,
136 amdgpu_device_get_pcie_replay_count, NULL);
138 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev);
141 * amdgpu_device_supports_boco - Is the device a dGPU with HG/PX power control
143 * @dev: drm_device pointer
145 * Returns true if the device is a dGPU with HG/PX power control,
146 * otherwise return false.
148 bool amdgpu_device_supports_boco(struct drm_device *dev)
150 struct amdgpu_device *adev = dev->dev_private;
152 if (adev->flags & AMD_IS_PX)
158 * amdgpu_device_supports_baco - Does the device support BACO
160 * @dev: drm_device pointer
162 * Returns true if the device supporte BACO,
163 * otherwise return false.
165 bool amdgpu_device_supports_baco(struct drm_device *dev)
167 struct amdgpu_device *adev = dev->dev_private;
169 return amdgpu_asic_supports_baco(adev);
173 * VRAM access helper functions.
175 * amdgpu_device_vram_access - read/write a buffer in vram
177 * @adev: amdgpu_device pointer
178 * @pos: offset of the buffer in vram
179 * @buf: virtual address of the buffer in system memory
180 * @size: read/write size, sizeof(@buf) must > @size
181 * @write: true - write to vram, otherwise - read from vram
183 void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos,
184 uint32_t *buf, size_t size, bool write)
190 for (last += pos; pos <= last; pos += 4) {
191 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
192 WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)pos) | 0x80000000);
193 WREG32_NO_KIQ(mmMM_INDEX_HI, pos >> 31);
195 WREG32_NO_KIQ(mmMM_DATA, *buf++);
197 *buf++ = RREG32_NO_KIQ(mmMM_DATA);
198 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
203 * MMIO register access helper functions.
206 * amdgpu_mm_rreg - read a memory mapped IO register
208 * @adev: amdgpu_device pointer
209 * @reg: dword aligned register offset
210 * @acc_flags: access flags which require special behavior
212 * Returns the 32 bit value from the offset specified.
214 uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg,
219 if ((acc_flags & AMDGPU_REGS_KIQ) || (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev)))
220 return amdgpu_kiq_rreg(adev, reg);
222 if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
223 ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
227 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
228 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
229 ret = readl(((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
230 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
232 trace_amdgpu_mm_rreg(adev->pdev->device, reg, ret);
237 * MMIO register read with bytes helper functions
238 * @offset:bytes offset from MMIO start
243 * amdgpu_mm_rreg8 - read a memory mapped IO register
245 * @adev: amdgpu_device pointer
246 * @offset: byte aligned register offset
248 * Returns the 8 bit value from the offset specified.
250 uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset) {
251 if (offset < adev->rmmio_size)
252 return (readb(adev->rmmio + offset));
257 * MMIO register write with bytes helper functions
258 * @offset:bytes offset from MMIO start
259 * @value: the value want to be written to the register
263 * amdgpu_mm_wreg8 - read a memory mapped IO register
265 * @adev: amdgpu_device pointer
266 * @offset: byte aligned register offset
267 * @value: 8 bit value to write
269 * Writes the value specified to the offset specified.
271 void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value) {
272 if (offset < adev->rmmio_size)
273 writeb(value, adev->rmmio + offset);
279 * amdgpu_mm_wreg - write to a memory mapped IO register
281 * @adev: amdgpu_device pointer
282 * @reg: dword aligned register offset
283 * @v: 32 bit value to write to the register
284 * @acc_flags: access flags which require special behavior
286 * Writes the value specified to the offset specified.
288 void amdgpu_mm_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v,
291 trace_amdgpu_mm_wreg(adev->pdev->device, reg, v);
293 if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
294 adev->last_mm_index = v;
297 if ((acc_flags & AMDGPU_REGS_KIQ) || (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev)))
298 return amdgpu_kiq_wreg(adev, reg, v);
300 if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
301 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
305 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
306 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
307 writel(v, ((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
308 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
311 if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
317 * amdgpu_io_rreg - read an IO register
319 * @adev: amdgpu_device pointer
320 * @reg: dword aligned register offset
322 * Returns the 32 bit value from the offset specified.
324 u32 amdgpu_io_rreg(struct amdgpu_device *adev, u32 reg)
326 if ((reg * 4) < adev->rio_mem_size)
327 return ioread32(adev->rio_mem + (reg * 4));
329 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
330 return ioread32(adev->rio_mem + (mmMM_DATA * 4));
335 * amdgpu_io_wreg - write to an IO register
337 * @adev: amdgpu_device pointer
338 * @reg: dword aligned register offset
339 * @v: 32 bit value to write to the register
341 * Writes the value specified to the offset specified.
343 void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32 v)
345 if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
346 adev->last_mm_index = v;
349 if ((reg * 4) < adev->rio_mem_size)
350 iowrite32(v, adev->rio_mem + (reg * 4));
352 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
353 iowrite32(v, adev->rio_mem + (mmMM_DATA * 4));
356 if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
362 * amdgpu_mm_rdoorbell - read a doorbell dword
364 * @adev: amdgpu_device pointer
365 * @index: doorbell index
367 * Returns the value in the doorbell aperture at the
368 * requested doorbell index (CIK).
370 u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
372 if (index < adev->doorbell.num_doorbells) {
373 return readl(adev->doorbell.ptr + index);
375 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
381 * amdgpu_mm_wdoorbell - write a doorbell dword
383 * @adev: amdgpu_device pointer
384 * @index: doorbell index
387 * Writes @v to the doorbell aperture at the
388 * requested doorbell index (CIK).
390 void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v)
392 if (index < adev->doorbell.num_doorbells) {
393 writel(v, adev->doorbell.ptr + index);
395 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
400 * amdgpu_mm_rdoorbell64 - read a doorbell Qword
402 * @adev: amdgpu_device pointer
403 * @index: doorbell index
405 * Returns the value in the doorbell aperture at the
406 * requested doorbell index (VEGA10+).
408 u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index)
410 if (index < adev->doorbell.num_doorbells) {
411 return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index));
413 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
419 * amdgpu_mm_wdoorbell64 - write a doorbell Qword
421 * @adev: amdgpu_device pointer
422 * @index: doorbell index
425 * Writes @v to the doorbell aperture at the
426 * requested doorbell index (VEGA10+).
428 void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v)
430 if (index < adev->doorbell.num_doorbells) {
431 atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v);
433 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
438 * amdgpu_invalid_rreg - dummy reg read function
440 * @adev: amdgpu device pointer
441 * @reg: offset of register
443 * Dummy register read function. Used for register blocks
444 * that certain asics don't have (all asics).
445 * Returns the value in the register.
447 static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
449 DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
455 * amdgpu_invalid_wreg - dummy reg write function
457 * @adev: amdgpu device pointer
458 * @reg: offset of register
459 * @v: value to write to the register
461 * Dummy register read function. Used for register blocks
462 * that certain asics don't have (all asics).
464 static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
466 DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
472 * amdgpu_invalid_rreg64 - dummy 64 bit reg read function
474 * @adev: amdgpu device pointer
475 * @reg: offset of register
477 * Dummy register read function. Used for register blocks
478 * that certain asics don't have (all asics).
479 * Returns the value in the register.
481 static uint64_t amdgpu_invalid_rreg64(struct amdgpu_device *adev, uint32_t reg)
483 DRM_ERROR("Invalid callback to read 64 bit register 0x%04X\n", reg);
489 * amdgpu_invalid_wreg64 - dummy reg write function
491 * @adev: amdgpu device pointer
492 * @reg: offset of register
493 * @v: value to write to the register
495 * Dummy register read function. Used for register blocks
496 * that certain asics don't have (all asics).
498 static void amdgpu_invalid_wreg64(struct amdgpu_device *adev, uint32_t reg, uint64_t v)
500 DRM_ERROR("Invalid callback to write 64 bit register 0x%04X with 0x%08llX\n",
506 * amdgpu_block_invalid_rreg - dummy reg read function
508 * @adev: amdgpu device pointer
509 * @block: offset of instance
510 * @reg: offset of register
512 * Dummy register read function. Used for register blocks
513 * that certain asics don't have (all asics).
514 * Returns the value in the register.
516 static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
517 uint32_t block, uint32_t reg)
519 DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
526 * amdgpu_block_invalid_wreg - dummy reg write function
528 * @adev: amdgpu device pointer
529 * @block: offset of instance
530 * @reg: offset of register
531 * @v: value to write to the register
533 * Dummy register read function. Used for register blocks
534 * that certain asics don't have (all asics).
536 static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
538 uint32_t reg, uint32_t v)
540 DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
546 * amdgpu_device_vram_scratch_init - allocate the VRAM scratch page
548 * @adev: amdgpu device pointer
550 * Allocates a scratch page of VRAM for use by various things in the
553 static int amdgpu_device_vram_scratch_init(struct amdgpu_device *adev)
555 return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE,
556 PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
557 &adev->vram_scratch.robj,
558 &adev->vram_scratch.gpu_addr,
559 (void **)&adev->vram_scratch.ptr);
563 * amdgpu_device_vram_scratch_fini - Free the VRAM scratch page
565 * @adev: amdgpu device pointer
567 * Frees the VRAM scratch page.
569 static void amdgpu_device_vram_scratch_fini(struct amdgpu_device *adev)
571 amdgpu_bo_free_kernel(&adev->vram_scratch.robj, NULL, NULL);
575 * amdgpu_device_program_register_sequence - program an array of registers.
577 * @adev: amdgpu_device pointer
578 * @registers: pointer to the register array
579 * @array_size: size of the register array
581 * Programs an array or registers with and and or masks.
582 * This is a helper for setting golden registers.
584 void amdgpu_device_program_register_sequence(struct amdgpu_device *adev,
585 const u32 *registers,
586 const u32 array_size)
588 u32 tmp, reg, and_mask, or_mask;
594 for (i = 0; i < array_size; i +=3) {
595 reg = registers[i + 0];
596 and_mask = registers[i + 1];
597 or_mask = registers[i + 2];
599 if (and_mask == 0xffffffff) {
604 if (adev->family >= AMDGPU_FAMILY_AI)
605 tmp |= (or_mask & and_mask);
614 * amdgpu_device_pci_config_reset - reset the GPU
616 * @adev: amdgpu_device pointer
618 * Resets the GPU using the pci config reset sequence.
619 * Only applicable to asics prior to vega10.
621 void amdgpu_device_pci_config_reset(struct amdgpu_device *adev)
623 pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
627 * GPU doorbell aperture helpers function.
630 * amdgpu_device_doorbell_init - Init doorbell driver information.
632 * @adev: amdgpu_device pointer
634 * Init doorbell driver information (CIK)
635 * Returns 0 on success, error on failure.
637 static int amdgpu_device_doorbell_init(struct amdgpu_device *adev)
640 /* No doorbell on SI hardware generation */
641 if (adev->asic_type < CHIP_BONAIRE) {
642 adev->doorbell.base = 0;
643 adev->doorbell.size = 0;
644 adev->doorbell.num_doorbells = 0;
645 adev->doorbell.ptr = NULL;
649 if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET)
652 amdgpu_asic_init_doorbell_index(adev);
654 /* doorbell bar mapping */
655 adev->doorbell.base = pci_resource_start(adev->pdev, 2);
656 adev->doorbell.size = pci_resource_len(adev->pdev, 2);
658 adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32),
659 adev->doorbell_index.max_assignment+1);
660 if (adev->doorbell.num_doorbells == 0)
663 /* For Vega, reserve and map two pages on doorbell BAR since SDMA
664 * paging queue doorbell use the second page. The
665 * AMDGPU_DOORBELL64_MAX_ASSIGNMENT definition assumes all the
666 * doorbells are in the first page. So with paging queue enabled,
667 * the max num_doorbells should + 1 page (0x400 in dword)
669 if (adev->asic_type >= CHIP_VEGA10)
670 adev->doorbell.num_doorbells += 0x400;
672 adev->doorbell.ptr = ioremap(adev->doorbell.base,
673 adev->doorbell.num_doorbells *
675 if (adev->doorbell.ptr == NULL)
682 * amdgpu_device_doorbell_fini - Tear down doorbell driver information.
684 * @adev: amdgpu_device pointer
686 * Tear down doorbell driver information (CIK)
688 static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev)
690 iounmap(adev->doorbell.ptr);
691 adev->doorbell.ptr = NULL;
697 * amdgpu_device_wb_*()
698 * Writeback is the method by which the GPU updates special pages in memory
699 * with the status of certain GPU events (fences, ring pointers,etc.).
703 * amdgpu_device_wb_fini - Disable Writeback and free memory
705 * @adev: amdgpu_device pointer
707 * Disables Writeback and frees the Writeback memory (all asics).
708 * Used at driver shutdown.
710 static void amdgpu_device_wb_fini(struct amdgpu_device *adev)
712 if (adev->wb.wb_obj) {
713 amdgpu_bo_free_kernel(&adev->wb.wb_obj,
715 (void **)&adev->wb.wb);
716 adev->wb.wb_obj = NULL;
721 * amdgpu_device_wb_init- Init Writeback driver info and allocate memory
723 * @adev: amdgpu_device pointer
725 * Initializes writeback and allocates writeback memory (all asics).
726 * Used at driver startup.
727 * Returns 0 on success or an -error on failure.
729 static int amdgpu_device_wb_init(struct amdgpu_device *adev)
733 if (adev->wb.wb_obj == NULL) {
734 /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */
735 r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8,
736 PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
737 &adev->wb.wb_obj, &adev->wb.gpu_addr,
738 (void **)&adev->wb.wb);
740 dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
744 adev->wb.num_wb = AMDGPU_MAX_WB;
745 memset(&adev->wb.used, 0, sizeof(adev->wb.used));
747 /* clear wb memory */
748 memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8);
755 * amdgpu_device_wb_get - Allocate a wb entry
757 * @adev: amdgpu_device pointer
760 * Allocate a wb slot for use by the driver (all asics).
761 * Returns 0 on success or -EINVAL on failure.
763 int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb)
765 unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
767 if (offset < adev->wb.num_wb) {
768 __set_bit(offset, adev->wb.used);
769 *wb = offset << 3; /* convert to dw offset */
777 * amdgpu_device_wb_free - Free a wb entry
779 * @adev: amdgpu_device pointer
782 * Free a wb slot allocated for use by the driver (all asics)
784 void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb)
787 if (wb < adev->wb.num_wb)
788 __clear_bit(wb, adev->wb.used);
792 * amdgpu_device_resize_fb_bar - try to resize FB BAR
794 * @adev: amdgpu_device pointer
796 * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not
797 * to fail, but if any of the BARs is not accessible after the size we abort
798 * driver loading by returning -ENODEV.
800 int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev)
802 u64 space_needed = roundup_pow_of_two(adev->gmc.real_vram_size);
803 u32 rbar_size = order_base_2(((space_needed >> 20) | 1)) - 1;
804 struct pci_bus *root;
805 struct resource *res;
811 if (amdgpu_sriov_vf(adev))
814 /* Check if the root BUS has 64bit memory resources */
815 root = adev->pdev->bus;
819 pci_bus_for_each_resource(root, res, i) {
820 if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) &&
821 res->start > 0x100000000ull)
825 /* Trying to resize is pointless without a root hub window above 4GB */
829 /* Disable memory decoding while we change the BAR addresses and size */
830 pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd);
831 pci_write_config_word(adev->pdev, PCI_COMMAND,
832 cmd & ~PCI_COMMAND_MEMORY);
834 /* Free the VRAM and doorbell BAR, we most likely need to move both. */
835 amdgpu_device_doorbell_fini(adev);
836 if (adev->asic_type >= CHIP_BONAIRE)
837 pci_release_resource(adev->pdev, 2);
839 pci_release_resource(adev->pdev, 0);
841 r = pci_resize_resource(adev->pdev, 0, rbar_size);
843 DRM_INFO("Not enough PCI address space for a large BAR.");
844 else if (r && r != -ENOTSUPP)
845 DRM_ERROR("Problem resizing BAR0 (%d).", r);
847 pci_assign_unassigned_bus_resources(adev->pdev->bus);
849 /* When the doorbell or fb BAR isn't available we have no chance of
852 r = amdgpu_device_doorbell_init(adev);
853 if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET))
856 pci_write_config_word(adev->pdev, PCI_COMMAND, cmd);
862 * GPU helpers function.
865 * amdgpu_device_need_post - check if the hw need post or not
867 * @adev: amdgpu_device pointer
869 * Check if the asic has been initialized (all asics) at driver startup
870 * or post is needed if hw reset is performed.
871 * Returns true if need or false if not.
873 bool amdgpu_device_need_post(struct amdgpu_device *adev)
877 if (amdgpu_sriov_vf(adev))
880 if (amdgpu_passthrough(adev)) {
881 /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
882 * some old smc fw still need driver do vPost otherwise gpu hang, while
883 * those smc fw version above 22.15 doesn't have this flaw, so we force
884 * vpost executed for smc version below 22.15
886 if (adev->asic_type == CHIP_FIJI) {
889 err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
890 /* force vPost if error occured */
894 fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
895 if (fw_ver < 0x00160e00)
900 if (adev->has_hw_reset) {
901 adev->has_hw_reset = false;
905 /* bios scratch used on CIK+ */
906 if (adev->asic_type >= CHIP_BONAIRE)
907 return amdgpu_atombios_scratch_need_asic_init(adev);
909 /* check MEM_SIZE for older asics */
910 reg = amdgpu_asic_get_config_memsize(adev);
912 if ((reg != 0) && (reg != 0xffffffff))
918 /* if we get transitioned to only one device, take VGA back */
920 * amdgpu_device_vga_set_decode - enable/disable vga decode
922 * @cookie: amdgpu_device pointer
923 * @state: enable/disable vga decode
925 * Enable/disable vga decode (all asics).
926 * Returns VGA resource flags.
928 static unsigned int amdgpu_device_vga_set_decode(void *cookie, bool state)
930 struct amdgpu_device *adev = cookie;
931 amdgpu_asic_set_vga_state(adev, state);
933 return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
934 VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
936 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
940 * amdgpu_device_check_block_size - validate the vm block size
942 * @adev: amdgpu_device pointer
944 * Validates the vm block size specified via module parameter.
945 * The vm block size defines number of bits in page table versus page directory,
946 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
947 * page table and the remaining bits are in the page directory.
949 static void amdgpu_device_check_block_size(struct amdgpu_device *adev)
951 /* defines number of bits in page table versus page directory,
952 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
953 * page table and the remaining bits are in the page directory */
954 if (amdgpu_vm_block_size == -1)
957 if (amdgpu_vm_block_size < 9) {
958 dev_warn(adev->dev, "VM page table size (%d) too small\n",
959 amdgpu_vm_block_size);
960 amdgpu_vm_block_size = -1;
965 * amdgpu_device_check_vm_size - validate the vm size
967 * @adev: amdgpu_device pointer
969 * Validates the vm size in GB specified via module parameter.
970 * The VM size is the size of the GPU virtual memory space in GB.
972 static void amdgpu_device_check_vm_size(struct amdgpu_device *adev)
974 /* no need to check the default value */
975 if (amdgpu_vm_size == -1)
978 if (amdgpu_vm_size < 1) {
979 dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
985 static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev)
988 bool is_os_64 = (sizeof(void *) == 8);
989 uint64_t total_memory;
990 uint64_t dram_size_seven_GB = 0x1B8000000;
991 uint64_t dram_size_three_GB = 0xB8000000;
993 if (amdgpu_smu_memory_pool_size == 0)
997 DRM_WARN("Not 64-bit OS, feature not supported\n");
1001 total_memory = (uint64_t)si.totalram * si.mem_unit;
1003 if ((amdgpu_smu_memory_pool_size == 1) ||
1004 (amdgpu_smu_memory_pool_size == 2)) {
1005 if (total_memory < dram_size_three_GB)
1007 } else if ((amdgpu_smu_memory_pool_size == 4) ||
1008 (amdgpu_smu_memory_pool_size == 8)) {
1009 if (total_memory < dram_size_seven_GB)
1012 DRM_WARN("Smu memory pool size not supported\n");
1015 adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28;
1020 DRM_WARN("No enough system memory\n");
1022 adev->pm.smu_prv_buffer_size = 0;
1026 * amdgpu_device_check_arguments - validate module params
1028 * @adev: amdgpu_device pointer
1030 * Validates certain module parameters and updates
1031 * the associated values used by the driver (all asics).
1033 static int amdgpu_device_check_arguments(struct amdgpu_device *adev)
1035 if (amdgpu_sched_jobs < 4) {
1036 dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
1038 amdgpu_sched_jobs = 4;
1039 } else if (!is_power_of_2(amdgpu_sched_jobs)){
1040 dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
1042 amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
1045 if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) {
1046 /* gart size must be greater or equal to 32M */
1047 dev_warn(adev->dev, "gart size (%d) too small\n",
1049 amdgpu_gart_size = -1;
1052 if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) {
1053 /* gtt size must be greater or equal to 32M */
1054 dev_warn(adev->dev, "gtt size (%d) too small\n",
1056 amdgpu_gtt_size = -1;
1059 /* valid range is between 4 and 9 inclusive */
1060 if (amdgpu_vm_fragment_size != -1 &&
1061 (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) {
1062 dev_warn(adev->dev, "valid range is between 4 and 9\n");
1063 amdgpu_vm_fragment_size = -1;
1066 amdgpu_device_check_smu_prv_buffer_size(adev);
1068 amdgpu_device_check_vm_size(adev);
1070 amdgpu_device_check_block_size(adev);
1072 adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type);
1078 * amdgpu_switcheroo_set_state - set switcheroo state
1080 * @pdev: pci dev pointer
1081 * @state: vga_switcheroo state
1083 * Callback for the switcheroo driver. Suspends or resumes the
1084 * the asics before or after it is powered up using ACPI methods.
1086 static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, enum vga_switcheroo_state state)
1088 struct drm_device *dev = pci_get_drvdata(pdev);
1091 if (amdgpu_device_supports_boco(dev) && state == VGA_SWITCHEROO_OFF)
1094 if (state == VGA_SWITCHEROO_ON) {
1095 pr_info("amdgpu: switched on\n");
1096 /* don't suspend or resume card normally */
1097 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1099 pci_set_power_state(dev->pdev, PCI_D0);
1100 pci_restore_state(dev->pdev);
1101 r = pci_enable_device(dev->pdev);
1103 DRM_WARN("pci_enable_device failed (%d)\n", r);
1104 amdgpu_device_resume(dev, true);
1106 dev->switch_power_state = DRM_SWITCH_POWER_ON;
1107 drm_kms_helper_poll_enable(dev);
1109 pr_info("amdgpu: switched off\n");
1110 drm_kms_helper_poll_disable(dev);
1111 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1112 amdgpu_device_suspend(dev, true);
1113 pci_save_state(dev->pdev);
1114 /* Shut down the device */
1115 pci_disable_device(dev->pdev);
1116 pci_set_power_state(dev->pdev, PCI_D3cold);
1117 dev->switch_power_state = DRM_SWITCH_POWER_OFF;
1122 * amdgpu_switcheroo_can_switch - see if switcheroo state can change
1124 * @pdev: pci dev pointer
1126 * Callback for the switcheroo driver. Check of the switcheroo
1127 * state can be changed.
1128 * Returns true if the state can be changed, false if not.
1130 static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
1132 struct drm_device *dev = pci_get_drvdata(pdev);
1135 * FIXME: open_count is protected by drm_global_mutex but that would lead to
1136 * locking inversion with the driver load path. And the access here is
1137 * completely racy anyway. So don't bother with locking for now.
1139 return dev->open_count == 0;
1142 static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
1143 .set_gpu_state = amdgpu_switcheroo_set_state,
1145 .can_switch = amdgpu_switcheroo_can_switch,
1149 * amdgpu_device_ip_set_clockgating_state - set the CG state
1151 * @dev: amdgpu_device pointer
1152 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1153 * @state: clockgating state (gate or ungate)
1155 * Sets the requested clockgating state for all instances of
1156 * the hardware IP specified.
1157 * Returns the error code from the last instance.
1159 int amdgpu_device_ip_set_clockgating_state(void *dev,
1160 enum amd_ip_block_type block_type,
1161 enum amd_clockgating_state state)
1163 struct amdgpu_device *adev = dev;
1166 for (i = 0; i < adev->num_ip_blocks; i++) {
1167 if (!adev->ip_blocks[i].status.valid)
1169 if (adev->ip_blocks[i].version->type != block_type)
1171 if (!adev->ip_blocks[i].version->funcs->set_clockgating_state)
1173 r = adev->ip_blocks[i].version->funcs->set_clockgating_state(
1174 (void *)adev, state);
1176 DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",
1177 adev->ip_blocks[i].version->funcs->name, r);
1183 * amdgpu_device_ip_set_powergating_state - set the PG state
1185 * @dev: amdgpu_device pointer
1186 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1187 * @state: powergating state (gate or ungate)
1189 * Sets the requested powergating state for all instances of
1190 * the hardware IP specified.
1191 * Returns the error code from the last instance.
1193 int amdgpu_device_ip_set_powergating_state(void *dev,
1194 enum amd_ip_block_type block_type,
1195 enum amd_powergating_state state)
1197 struct amdgpu_device *adev = dev;
1200 for (i = 0; i < adev->num_ip_blocks; i++) {
1201 if (!adev->ip_blocks[i].status.valid)
1203 if (adev->ip_blocks[i].version->type != block_type)
1205 if (!adev->ip_blocks[i].version->funcs->set_powergating_state)
1207 r = adev->ip_blocks[i].version->funcs->set_powergating_state(
1208 (void *)adev, state);
1210 DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",
1211 adev->ip_blocks[i].version->funcs->name, r);
1217 * amdgpu_device_ip_get_clockgating_state - get the CG state
1219 * @adev: amdgpu_device pointer
1220 * @flags: clockgating feature flags
1222 * Walks the list of IPs on the device and updates the clockgating
1223 * flags for each IP.
1224 * Updates @flags with the feature flags for each hardware IP where
1225 * clockgating is enabled.
1227 void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev,
1232 for (i = 0; i < adev->num_ip_blocks; i++) {
1233 if (!adev->ip_blocks[i].status.valid)
1235 if (adev->ip_blocks[i].version->funcs->get_clockgating_state)
1236 adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags);
1241 * amdgpu_device_ip_wait_for_idle - wait for idle
1243 * @adev: amdgpu_device pointer
1244 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1246 * Waits for the request hardware IP to be idle.
1247 * Returns 0 for success or a negative error code on failure.
1249 int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev,
1250 enum amd_ip_block_type block_type)
1254 for (i = 0; i < adev->num_ip_blocks; i++) {
1255 if (!adev->ip_blocks[i].status.valid)
1257 if (adev->ip_blocks[i].version->type == block_type) {
1258 r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
1269 * amdgpu_device_ip_is_idle - is the hardware IP idle
1271 * @adev: amdgpu_device pointer
1272 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1274 * Check if the hardware IP is idle or not.
1275 * Returns true if it the IP is idle, false if not.
1277 bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev,
1278 enum amd_ip_block_type block_type)
1282 for (i = 0; i < adev->num_ip_blocks; i++) {
1283 if (!adev->ip_blocks[i].status.valid)
1285 if (adev->ip_blocks[i].version->type == block_type)
1286 return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
1293 * amdgpu_device_ip_get_ip_block - get a hw IP pointer
1295 * @adev: amdgpu_device pointer
1296 * @type: Type of hardware IP (SMU, GFX, UVD, etc.)
1298 * Returns a pointer to the hardware IP block structure
1299 * if it exists for the asic, otherwise NULL.
1301 struct amdgpu_ip_block *
1302 amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,
1303 enum amd_ip_block_type type)
1307 for (i = 0; i < adev->num_ip_blocks; i++)
1308 if (adev->ip_blocks[i].version->type == type)
1309 return &adev->ip_blocks[i];
1315 * amdgpu_device_ip_block_version_cmp
1317 * @adev: amdgpu_device pointer
1318 * @type: enum amd_ip_block_type
1319 * @major: major version
1320 * @minor: minor version
1322 * return 0 if equal or greater
1323 * return 1 if smaller or the ip_block doesn't exist
1325 int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev,
1326 enum amd_ip_block_type type,
1327 u32 major, u32 minor)
1329 struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type);
1331 if (ip_block && ((ip_block->version->major > major) ||
1332 ((ip_block->version->major == major) &&
1333 (ip_block->version->minor >= minor))))
1340 * amdgpu_device_ip_block_add
1342 * @adev: amdgpu_device pointer
1343 * @ip_block_version: pointer to the IP to add
1345 * Adds the IP block driver information to the collection of IPs
1348 int amdgpu_device_ip_block_add(struct amdgpu_device *adev,
1349 const struct amdgpu_ip_block_version *ip_block_version)
1351 if (!ip_block_version)
1354 DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks,
1355 ip_block_version->funcs->name);
1357 adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
1363 * amdgpu_device_enable_virtual_display - enable virtual display feature
1365 * @adev: amdgpu_device pointer
1367 * Enabled the virtual display feature if the user has enabled it via
1368 * the module parameter virtual_display. This feature provides a virtual
1369 * display hardware on headless boards or in virtualized environments.
1370 * This function parses and validates the configuration string specified by
1371 * the user and configues the virtual display configuration (number of
1372 * virtual connectors, crtcs, etc.) specified.
1374 static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
1376 adev->enable_virtual_display = false;
1378 if (amdgpu_virtual_display) {
1379 struct drm_device *ddev = adev->ddev;
1380 const char *pci_address_name = pci_name(ddev->pdev);
1381 char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
1383 pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
1384 pciaddstr_tmp = pciaddstr;
1385 while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
1386 pciaddname = strsep(&pciaddname_tmp, ",");
1387 if (!strcmp("all", pciaddname)
1388 || !strcmp(pci_address_name, pciaddname)) {
1392 adev->enable_virtual_display = true;
1395 res = kstrtol(pciaddname_tmp, 10,
1403 adev->mode_info.num_crtc = num_crtc;
1405 adev->mode_info.num_crtc = 1;
1411 DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
1412 amdgpu_virtual_display, pci_address_name,
1413 adev->enable_virtual_display, adev->mode_info.num_crtc);
1420 * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware
1422 * @adev: amdgpu_device pointer
1424 * Parses the asic configuration parameters specified in the gpu info
1425 * firmware and makes them availale to the driver for use in configuring
1427 * Returns 0 on success, -EINVAL on failure.
1429 static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev)
1431 const char *chip_name;
1434 const struct gpu_info_firmware_header_v1_0 *hdr;
1436 adev->firmware.gpu_info_fw = NULL;
1438 switch (adev->asic_type) {
1442 case CHIP_POLARIS10:
1443 case CHIP_POLARIS11:
1444 case CHIP_POLARIS12:
1448 #ifdef CONFIG_DRM_AMDGPU_SI
1455 #ifdef CONFIG_DRM_AMDGPU_CIK
1466 chip_name = "vega10";
1469 chip_name = "vega12";
1472 if (adev->rev_id >= 8)
1473 chip_name = "raven2";
1474 else if (adev->pdev->device == 0x15d8)
1475 chip_name = "picasso";
1477 chip_name = "raven";
1480 chip_name = "arcturus";
1483 chip_name = "renoir";
1486 chip_name = "navi10";
1489 chip_name = "navi14";
1492 chip_name = "navi12";
1496 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name);
1497 err = request_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev);
1500 "Failed to load gpu_info firmware \"%s\"\n",
1504 err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw);
1507 "Failed to validate gpu_info firmware \"%s\"\n",
1512 hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data;
1513 amdgpu_ucode_print_gpu_info_hdr(&hdr->header);
1515 switch (hdr->version_major) {
1518 const struct gpu_info_firmware_v1_0 *gpu_info_fw =
1519 (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data +
1520 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1522 if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10)
1523 goto parse_soc_bounding_box;
1525 adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se);
1526 adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh);
1527 adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se);
1528 adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se);
1529 adev->gfx.config.max_texture_channel_caches =
1530 le32_to_cpu(gpu_info_fw->gc_num_tccs);
1531 adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs);
1532 adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds);
1533 adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth);
1534 adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth);
1535 adev->gfx.config.double_offchip_lds_buf =
1536 le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer);
1537 adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size);
1538 adev->gfx.cu_info.max_waves_per_simd =
1539 le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd);
1540 adev->gfx.cu_info.max_scratch_slots_per_cu =
1541 le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu);
1542 adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size);
1543 if (hdr->version_minor >= 1) {
1544 const struct gpu_info_firmware_v1_1 *gpu_info_fw =
1545 (const struct gpu_info_firmware_v1_1 *)(adev->firmware.gpu_info_fw->data +
1546 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1547 adev->gfx.config.num_sc_per_sh =
1548 le32_to_cpu(gpu_info_fw->num_sc_per_sh);
1549 adev->gfx.config.num_packer_per_sc =
1550 le32_to_cpu(gpu_info_fw->num_packer_per_sc);
1553 parse_soc_bounding_box:
1555 * soc bounding box info is not integrated in disocovery table,
1556 * we always need to parse it from gpu info firmware.
1558 if (hdr->version_minor == 2) {
1559 const struct gpu_info_firmware_v1_2 *gpu_info_fw =
1560 (const struct gpu_info_firmware_v1_2 *)(adev->firmware.gpu_info_fw->data +
1561 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1562 adev->dm.soc_bounding_box = &gpu_info_fw->soc_bounding_box;
1568 "Unsupported gpu_info table %d\n", hdr->header.ucode_version);
1577 * amdgpu_device_ip_early_init - run early init for hardware IPs
1579 * @adev: amdgpu_device pointer
1581 * Early initialization pass for hardware IPs. The hardware IPs that make
1582 * up each asic are discovered each IP's early_init callback is run. This
1583 * is the first stage in initializing the asic.
1584 * Returns 0 on success, negative error code on failure.
1586 static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
1590 amdgpu_device_enable_virtual_display(adev);
1592 switch (adev->asic_type) {
1596 case CHIP_POLARIS10:
1597 case CHIP_POLARIS11:
1598 case CHIP_POLARIS12:
1602 if (adev->asic_type == CHIP_CARRIZO || adev->asic_type == CHIP_STONEY)
1603 adev->family = AMDGPU_FAMILY_CZ;
1605 adev->family = AMDGPU_FAMILY_VI;
1607 r = vi_set_ip_blocks(adev);
1611 #ifdef CONFIG_DRM_AMDGPU_SI
1617 adev->family = AMDGPU_FAMILY_SI;
1618 r = si_set_ip_blocks(adev);
1623 #ifdef CONFIG_DRM_AMDGPU_CIK
1629 if ((adev->asic_type == CHIP_BONAIRE) || (adev->asic_type == CHIP_HAWAII))
1630 adev->family = AMDGPU_FAMILY_CI;
1632 adev->family = AMDGPU_FAMILY_KV;
1634 r = cik_set_ip_blocks(adev);
1645 if (adev->asic_type == CHIP_RAVEN ||
1646 adev->asic_type == CHIP_RENOIR)
1647 adev->family = AMDGPU_FAMILY_RV;
1649 adev->family = AMDGPU_FAMILY_AI;
1651 r = soc15_set_ip_blocks(adev);
1658 adev->family = AMDGPU_FAMILY_NV;
1660 r = nv_set_ip_blocks(adev);
1665 /* FIXME: not supported yet */
1669 r = amdgpu_device_parse_gpu_info_fw(adev);
1673 if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10)
1674 amdgpu_discovery_get_gfx_info(adev);
1676 amdgpu_amdkfd_device_probe(adev);
1678 if (amdgpu_sriov_vf(adev)) {
1679 r = amdgpu_virt_request_full_gpu(adev, true);
1684 adev->pm.pp_feature = amdgpu_pp_feature_mask;
1685 if (amdgpu_sriov_vf(adev) || sched_policy == KFD_SCHED_POLICY_NO_HWS)
1686 adev->pm.pp_feature &= ~PP_GFXOFF_MASK;
1688 for (i = 0; i < adev->num_ip_blocks; i++) {
1689 if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
1690 DRM_ERROR("disabled ip block: %d <%s>\n",
1691 i, adev->ip_blocks[i].version->funcs->name);
1692 adev->ip_blocks[i].status.valid = false;
1694 if (adev->ip_blocks[i].version->funcs->early_init) {
1695 r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
1697 adev->ip_blocks[i].status.valid = false;
1699 DRM_ERROR("early_init of IP block <%s> failed %d\n",
1700 adev->ip_blocks[i].version->funcs->name, r);
1703 adev->ip_blocks[i].status.valid = true;
1706 adev->ip_blocks[i].status.valid = true;
1709 /* get the vbios after the asic_funcs are set up */
1710 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) {
1712 if (!amdgpu_get_bios(adev))
1715 r = amdgpu_atombios_init(adev);
1717 dev_err(adev->dev, "amdgpu_atombios_init failed\n");
1718 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0);
1724 adev->cg_flags &= amdgpu_cg_mask;
1725 adev->pg_flags &= amdgpu_pg_mask;
1730 static int amdgpu_device_ip_hw_init_phase1(struct amdgpu_device *adev)
1734 for (i = 0; i < adev->num_ip_blocks; i++) {
1735 if (!adev->ip_blocks[i].status.sw)
1737 if (adev->ip_blocks[i].status.hw)
1739 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
1740 (amdgpu_sriov_vf(adev) && (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)) ||
1741 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
1742 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
1744 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1745 adev->ip_blocks[i].version->funcs->name, r);
1748 adev->ip_blocks[i].status.hw = true;
1755 static int amdgpu_device_ip_hw_init_phase2(struct amdgpu_device *adev)
1759 for (i = 0; i < adev->num_ip_blocks; i++) {
1760 if (!adev->ip_blocks[i].status.sw)
1762 if (adev->ip_blocks[i].status.hw)
1764 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
1766 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1767 adev->ip_blocks[i].version->funcs->name, r);
1770 adev->ip_blocks[i].status.hw = true;
1776 static int amdgpu_device_fw_loading(struct amdgpu_device *adev)
1780 uint32_t smu_version;
1782 if (adev->asic_type >= CHIP_VEGA10) {
1783 for (i = 0; i < adev->num_ip_blocks; i++) {
1784 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_PSP)
1787 /* no need to do the fw loading again if already done*/
1788 if (adev->ip_blocks[i].status.hw == true)
1791 if (adev->in_gpu_reset || adev->in_suspend) {
1792 r = adev->ip_blocks[i].version->funcs->resume(adev);
1794 DRM_ERROR("resume of IP block <%s> failed %d\n",
1795 adev->ip_blocks[i].version->funcs->name, r);
1799 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
1801 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1802 adev->ip_blocks[i].version->funcs->name, r);
1807 adev->ip_blocks[i].status.hw = true;
1812 if (!amdgpu_sriov_vf(adev) || adev->asic_type == CHIP_TONGA)
1813 r = amdgpu_pm_load_smu_firmware(adev, &smu_version);
1819 * amdgpu_device_ip_init - run init for hardware IPs
1821 * @adev: amdgpu_device pointer
1823 * Main initialization pass for hardware IPs. The list of all the hardware
1824 * IPs that make up the asic is walked and the sw_init and hw_init callbacks
1825 * are run. sw_init initializes the software state associated with each IP
1826 * and hw_init initializes the hardware associated with each IP.
1827 * Returns 0 on success, negative error code on failure.
1829 static int amdgpu_device_ip_init(struct amdgpu_device *adev)
1833 r = amdgpu_ras_init(adev);
1837 for (i = 0; i < adev->num_ip_blocks; i++) {
1838 if (!adev->ip_blocks[i].status.valid)
1840 r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
1842 DRM_ERROR("sw_init of IP block <%s> failed %d\n",
1843 adev->ip_blocks[i].version->funcs->name, r);
1846 adev->ip_blocks[i].status.sw = true;
1848 /* need to do gmc hw init early so we can allocate gpu mem */
1849 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1850 r = amdgpu_device_vram_scratch_init(adev);
1852 DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r);
1855 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1857 DRM_ERROR("hw_init %d failed %d\n", i, r);
1860 r = amdgpu_device_wb_init(adev);
1862 DRM_ERROR("amdgpu_device_wb_init failed %d\n", r);
1865 adev->ip_blocks[i].status.hw = true;
1867 /* right after GMC hw init, we create CSA */
1868 if (amdgpu_mcbp || amdgpu_sriov_vf(adev)) {
1869 r = amdgpu_allocate_static_csa(adev, &adev->virt.csa_obj,
1870 AMDGPU_GEM_DOMAIN_VRAM,
1873 DRM_ERROR("allocate CSA failed %d\n", r);
1880 if (amdgpu_sriov_vf(adev))
1881 amdgpu_virt_init_data_exchange(adev);
1883 r = amdgpu_ib_pool_init(adev);
1885 dev_err(adev->dev, "IB initialization failed (%d).\n", r);
1886 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r);
1890 r = amdgpu_ucode_create_bo(adev); /* create ucode bo when sw_init complete*/
1894 r = amdgpu_device_ip_hw_init_phase1(adev);
1898 r = amdgpu_device_fw_loading(adev);
1902 r = amdgpu_device_ip_hw_init_phase2(adev);
1907 * retired pages will be loaded from eeprom and reserved here,
1908 * it should be called after amdgpu_device_ip_hw_init_phase2 since
1909 * for some ASICs the RAS EEPROM code relies on SMU fully functioning
1910 * for I2C communication which only true at this point.
1911 * recovery_init may fail, but it can free all resources allocated by
1912 * itself and its failure should not stop amdgpu init process.
1914 * Note: theoretically, this should be called before all vram allocations
1915 * to protect retired page from abusing
1917 amdgpu_ras_recovery_init(adev);
1919 if (adev->gmc.xgmi.num_physical_nodes > 1)
1920 amdgpu_xgmi_add_device(adev);
1921 amdgpu_amdkfd_device_init(adev);
1924 if (amdgpu_sriov_vf(adev))
1925 amdgpu_virt_release_full_gpu(adev, true);
1931 * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer
1933 * @adev: amdgpu_device pointer
1935 * Writes a reset magic value to the gart pointer in VRAM. The driver calls
1936 * this function before a GPU reset. If the value is retained after a
1937 * GPU reset, VRAM has not been lost. Some GPU resets may destry VRAM contents.
1939 static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev)
1941 memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
1945 * amdgpu_device_check_vram_lost - check if vram is valid
1947 * @adev: amdgpu_device pointer
1949 * Checks the reset magic value written to the gart pointer in VRAM.
1950 * The driver calls this after a GPU reset to see if the contents of
1951 * VRAM is lost or now.
1952 * returns true if vram is lost, false if not.
1954 static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev)
1956 return !!memcmp(adev->gart.ptr, adev->reset_magic,
1957 AMDGPU_RESET_MAGIC_NUM);
1961 * amdgpu_device_set_cg_state - set clockgating for amdgpu device
1963 * @adev: amdgpu_device pointer
1964 * @state: clockgating state (gate or ungate)
1966 * The list of all the hardware IPs that make up the asic is walked and the
1967 * set_clockgating_state callbacks are run.
1968 * Late initialization pass enabling clockgating for hardware IPs.
1969 * Fini or suspend, pass disabling clockgating for hardware IPs.
1970 * Returns 0 on success, negative error code on failure.
1973 static int amdgpu_device_set_cg_state(struct amdgpu_device *adev,
1974 enum amd_clockgating_state state)
1978 if (amdgpu_emu_mode == 1)
1981 for (j = 0; j < adev->num_ip_blocks; j++) {
1982 i = state == AMD_CG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
1983 if (!adev->ip_blocks[i].status.late_initialized)
1985 /* skip CG for VCE/UVD, it's handled specially */
1986 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1987 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1988 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
1989 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG &&
1990 adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1991 /* enable clockgating to save power */
1992 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1995 DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
1996 adev->ip_blocks[i].version->funcs->name, r);
2005 static int amdgpu_device_set_pg_state(struct amdgpu_device *adev, enum amd_powergating_state state)
2009 if (amdgpu_emu_mode == 1)
2012 for (j = 0; j < adev->num_ip_blocks; j++) {
2013 i = state == AMD_PG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
2014 if (!adev->ip_blocks[i].status.late_initialized)
2016 /* skip CG for VCE/UVD, it's handled specially */
2017 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
2018 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
2019 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
2020 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG &&
2021 adev->ip_blocks[i].version->funcs->set_powergating_state) {
2022 /* enable powergating to save power */
2023 r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev,
2026 DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n",
2027 adev->ip_blocks[i].version->funcs->name, r);
2035 static int amdgpu_device_enable_mgpu_fan_boost(void)
2037 struct amdgpu_gpu_instance *gpu_ins;
2038 struct amdgpu_device *adev;
2041 mutex_lock(&mgpu_info.mutex);
2044 * MGPU fan boost feature should be enabled
2045 * only when there are two or more dGPUs in
2048 if (mgpu_info.num_dgpu < 2)
2051 for (i = 0; i < mgpu_info.num_dgpu; i++) {
2052 gpu_ins = &(mgpu_info.gpu_ins[i]);
2053 adev = gpu_ins->adev;
2054 if (!(adev->flags & AMD_IS_APU) &&
2055 !gpu_ins->mgpu_fan_enabled &&
2056 adev->powerplay.pp_funcs &&
2057 adev->powerplay.pp_funcs->enable_mgpu_fan_boost) {
2058 ret = amdgpu_dpm_enable_mgpu_fan_boost(adev);
2062 gpu_ins->mgpu_fan_enabled = 1;
2067 mutex_unlock(&mgpu_info.mutex);
2073 * amdgpu_device_ip_late_init - run late init for hardware IPs
2075 * @adev: amdgpu_device pointer
2077 * Late initialization pass for hardware IPs. The list of all the hardware
2078 * IPs that make up the asic is walked and the late_init callbacks are run.
2079 * late_init covers any special initialization that an IP requires
2080 * after all of the have been initialized or something that needs to happen
2081 * late in the init process.
2082 * Returns 0 on success, negative error code on failure.
2084 static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
2086 struct amdgpu_gpu_instance *gpu_instance;
2089 for (i = 0; i < adev->num_ip_blocks; i++) {
2090 if (!adev->ip_blocks[i].status.hw)
2092 if (adev->ip_blocks[i].version->funcs->late_init) {
2093 r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
2095 DRM_ERROR("late_init of IP block <%s> failed %d\n",
2096 adev->ip_blocks[i].version->funcs->name, r);
2100 adev->ip_blocks[i].status.late_initialized = true;
2103 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE);
2104 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE);
2106 amdgpu_device_fill_reset_magic(adev);
2108 r = amdgpu_device_enable_mgpu_fan_boost();
2110 DRM_ERROR("enable mgpu fan boost failed (%d).\n", r);
2113 if (adev->gmc.xgmi.num_physical_nodes > 1) {
2114 mutex_lock(&mgpu_info.mutex);
2117 * Reset device p-state to low as this was booted with high.
2119 * This should be performed only after all devices from the same
2120 * hive get initialized.
2122 * However, it's unknown how many device in the hive in advance.
2123 * As this is counted one by one during devices initializations.
2125 * So, we wait for all XGMI interlinked devices initialized.
2126 * This may bring some delays as those devices may come from
2127 * different hives. But that should be OK.
2129 if (mgpu_info.num_dgpu == adev->gmc.xgmi.num_physical_nodes) {
2130 for (i = 0; i < mgpu_info.num_gpu; i++) {
2131 gpu_instance = &(mgpu_info.gpu_ins[i]);
2132 if (gpu_instance->adev->flags & AMD_IS_APU)
2135 r = amdgpu_xgmi_set_pstate(gpu_instance->adev, 0);
2137 DRM_ERROR("pstate setting failed (%d).\n", r);
2143 mutex_unlock(&mgpu_info.mutex);
2150 * amdgpu_device_ip_fini - run fini for hardware IPs
2152 * @adev: amdgpu_device pointer
2154 * Main teardown pass for hardware IPs. The list of all the hardware
2155 * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks
2156 * are run. hw_fini tears down the hardware associated with each IP
2157 * and sw_fini tears down any software state associated with each IP.
2158 * Returns 0 on success, negative error code on failure.
2160 static int amdgpu_device_ip_fini(struct amdgpu_device *adev)
2164 amdgpu_ras_pre_fini(adev);
2166 if (adev->gmc.xgmi.num_physical_nodes > 1)
2167 amdgpu_xgmi_remove_device(adev);
2169 amdgpu_amdkfd_device_fini(adev);
2171 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
2172 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
2174 /* need to disable SMC first */
2175 for (i = 0; i < adev->num_ip_blocks; i++) {
2176 if (!adev->ip_blocks[i].status.hw)
2178 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
2179 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
2180 /* XXX handle errors */
2182 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
2183 adev->ip_blocks[i].version->funcs->name, r);
2185 adev->ip_blocks[i].status.hw = false;
2190 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2191 if (!adev->ip_blocks[i].status.hw)
2194 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
2195 /* XXX handle errors */
2197 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
2198 adev->ip_blocks[i].version->funcs->name, r);
2201 adev->ip_blocks[i].status.hw = false;
2205 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2206 if (!adev->ip_blocks[i].status.sw)
2209 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
2210 amdgpu_ucode_free_bo(adev);
2211 amdgpu_free_static_csa(&adev->virt.csa_obj);
2212 amdgpu_device_wb_fini(adev);
2213 amdgpu_device_vram_scratch_fini(adev);
2214 amdgpu_ib_pool_fini(adev);
2217 r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
2218 /* XXX handle errors */
2220 DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
2221 adev->ip_blocks[i].version->funcs->name, r);
2223 adev->ip_blocks[i].status.sw = false;
2224 adev->ip_blocks[i].status.valid = false;
2227 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2228 if (!adev->ip_blocks[i].status.late_initialized)
2230 if (adev->ip_blocks[i].version->funcs->late_fini)
2231 adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
2232 adev->ip_blocks[i].status.late_initialized = false;
2235 amdgpu_ras_fini(adev);
2237 if (amdgpu_sriov_vf(adev))
2238 if (amdgpu_virt_release_full_gpu(adev, false))
2239 DRM_ERROR("failed to release exclusive mode on fini\n");
2245 * amdgpu_device_delayed_init_work_handler - work handler for IB tests
2247 * @work: work_struct.
2249 static void amdgpu_device_delayed_init_work_handler(struct work_struct *work)
2251 struct amdgpu_device *adev =
2252 container_of(work, struct amdgpu_device, delayed_init_work.work);
2255 r = amdgpu_ib_ring_tests(adev);
2257 DRM_ERROR("ib ring test failed (%d).\n", r);
2260 static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work)
2262 struct amdgpu_device *adev =
2263 container_of(work, struct amdgpu_device, gfx.gfx_off_delay_work.work);
2265 mutex_lock(&adev->gfx.gfx_off_mutex);
2266 if (!adev->gfx.gfx_off_state && !adev->gfx.gfx_off_req_count) {
2267 if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, true))
2268 adev->gfx.gfx_off_state = true;
2270 mutex_unlock(&adev->gfx.gfx_off_mutex);
2274 * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1)
2276 * @adev: amdgpu_device pointer
2278 * Main suspend function for hardware IPs. The list of all the hardware
2279 * IPs that make up the asic is walked, clockgating is disabled and the
2280 * suspend callbacks are run. suspend puts the hardware and software state
2281 * in each IP into a state suitable for suspend.
2282 * Returns 0 on success, negative error code on failure.
2284 static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev)
2288 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
2289 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
2291 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2292 if (!adev->ip_blocks[i].status.valid)
2294 /* displays are handled separately */
2295 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) {
2296 /* XXX handle errors */
2297 r = adev->ip_blocks[i].version->funcs->suspend(adev);
2298 /* XXX handle errors */
2300 DRM_ERROR("suspend of IP block <%s> failed %d\n",
2301 adev->ip_blocks[i].version->funcs->name, r);
2304 adev->ip_blocks[i].status.hw = false;
2312 * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2)
2314 * @adev: amdgpu_device pointer
2316 * Main suspend function for hardware IPs. The list of all the hardware
2317 * IPs that make up the asic is walked, clockgating is disabled and the
2318 * suspend callbacks are run. suspend puts the hardware and software state
2319 * in each IP into a state suitable for suspend.
2320 * Returns 0 on success, negative error code on failure.
2322 static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev)
2326 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2327 if (!adev->ip_blocks[i].status.valid)
2329 /* displays are handled in phase1 */
2330 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE)
2332 /* PSP lost connection when err_event_athub occurs */
2333 if (amdgpu_ras_intr_triggered() &&
2334 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
2335 adev->ip_blocks[i].status.hw = false;
2338 /* XXX handle errors */
2339 r = adev->ip_blocks[i].version->funcs->suspend(adev);
2340 /* XXX handle errors */
2342 DRM_ERROR("suspend of IP block <%s> failed %d\n",
2343 adev->ip_blocks[i].version->funcs->name, r);
2345 adev->ip_blocks[i].status.hw = false;
2346 /* handle putting the SMC in the appropriate state */
2347 if(!amdgpu_sriov_vf(adev)){
2348 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
2349 r = amdgpu_dpm_set_mp1_state(adev, adev->mp1_state);
2351 DRM_ERROR("SMC failed to set mp1 state %d, %d\n",
2352 adev->mp1_state, r);
2357 adev->ip_blocks[i].status.hw = false;
2364 * amdgpu_device_ip_suspend - run suspend for hardware IPs
2366 * @adev: amdgpu_device pointer
2368 * Main suspend function for hardware IPs. The list of all the hardware
2369 * IPs that make up the asic is walked, clockgating is disabled and the
2370 * suspend callbacks are run. suspend puts the hardware and software state
2371 * in each IP into a state suitable for suspend.
2372 * Returns 0 on success, negative error code on failure.
2374 int amdgpu_device_ip_suspend(struct amdgpu_device *adev)
2378 if (amdgpu_sriov_vf(adev))
2379 amdgpu_virt_request_full_gpu(adev, false);
2381 r = amdgpu_device_ip_suspend_phase1(adev);
2384 r = amdgpu_device_ip_suspend_phase2(adev);
2386 if (amdgpu_sriov_vf(adev))
2387 amdgpu_virt_release_full_gpu(adev, false);
2392 static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
2396 static enum amd_ip_block_type ip_order[] = {
2397 AMD_IP_BLOCK_TYPE_GMC,
2398 AMD_IP_BLOCK_TYPE_COMMON,
2399 AMD_IP_BLOCK_TYPE_PSP,
2400 AMD_IP_BLOCK_TYPE_IH,
2403 for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
2405 struct amdgpu_ip_block *block;
2407 for (j = 0; j < adev->num_ip_blocks; j++) {
2408 block = &adev->ip_blocks[j];
2410 block->status.hw = false;
2411 if (block->version->type != ip_order[i] ||
2412 !block->status.valid)
2415 r = block->version->funcs->hw_init(adev);
2416 DRM_INFO("RE-INIT-early: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
2419 block->status.hw = true;
2426 static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
2430 static enum amd_ip_block_type ip_order[] = {
2431 AMD_IP_BLOCK_TYPE_SMC,
2432 AMD_IP_BLOCK_TYPE_DCE,
2433 AMD_IP_BLOCK_TYPE_GFX,
2434 AMD_IP_BLOCK_TYPE_SDMA,
2435 AMD_IP_BLOCK_TYPE_UVD,
2436 AMD_IP_BLOCK_TYPE_VCE,
2437 AMD_IP_BLOCK_TYPE_VCN
2440 for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
2442 struct amdgpu_ip_block *block;
2444 for (j = 0; j < adev->num_ip_blocks; j++) {
2445 block = &adev->ip_blocks[j];
2447 if (block->version->type != ip_order[i] ||
2448 !block->status.valid ||
2452 if (block->version->type == AMD_IP_BLOCK_TYPE_SMC)
2453 r = block->version->funcs->resume(adev);
2455 r = block->version->funcs->hw_init(adev);
2457 DRM_INFO("RE-INIT-late: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
2460 block->status.hw = true;
2468 * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs
2470 * @adev: amdgpu_device pointer
2472 * First resume function for hardware IPs. The list of all the hardware
2473 * IPs that make up the asic is walked and the resume callbacks are run for
2474 * COMMON, GMC, and IH. resume puts the hardware into a functional state
2475 * after a suspend and updates the software state as necessary. This
2476 * function is also used for restoring the GPU after a GPU reset.
2477 * Returns 0 on success, negative error code on failure.
2479 static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)
2483 for (i = 0; i < adev->num_ip_blocks; i++) {
2484 if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
2486 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2487 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2488 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
2490 r = adev->ip_blocks[i].version->funcs->resume(adev);
2492 DRM_ERROR("resume of IP block <%s> failed %d\n",
2493 adev->ip_blocks[i].version->funcs->name, r);
2496 adev->ip_blocks[i].status.hw = true;
2504 * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs
2506 * @adev: amdgpu_device pointer
2508 * First resume function for hardware IPs. The list of all the hardware
2509 * IPs that make up the asic is walked and the resume callbacks are run for
2510 * all blocks except COMMON, GMC, and IH. resume puts the hardware into a
2511 * functional state after a suspend and updates the software state as
2512 * necessary. This function is also used for restoring the GPU after a GPU
2514 * Returns 0 on success, negative error code on failure.
2516 static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)
2520 for (i = 0; i < adev->num_ip_blocks; i++) {
2521 if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
2523 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2524 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2525 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH ||
2526 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)
2528 r = adev->ip_blocks[i].version->funcs->resume(adev);
2530 DRM_ERROR("resume of IP block <%s> failed %d\n",
2531 adev->ip_blocks[i].version->funcs->name, r);
2534 adev->ip_blocks[i].status.hw = true;
2541 * amdgpu_device_ip_resume - run resume for hardware IPs
2543 * @adev: amdgpu_device pointer
2545 * Main resume function for hardware IPs. The hardware IPs
2546 * are split into two resume functions because they are
2547 * are also used in in recovering from a GPU reset and some additional
2548 * steps need to be take between them. In this case (S3/S4) they are
2550 * Returns 0 on success, negative error code on failure.
2552 static int amdgpu_device_ip_resume(struct amdgpu_device *adev)
2556 r = amdgpu_device_ip_resume_phase1(adev);
2560 r = amdgpu_device_fw_loading(adev);
2564 r = amdgpu_device_ip_resume_phase2(adev);
2570 * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV
2572 * @adev: amdgpu_device pointer
2574 * Query the VBIOS data tables to determine if the board supports SR-IOV.
2576 static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
2578 if (amdgpu_sriov_vf(adev)) {
2579 if (adev->is_atom_fw) {
2580 if (amdgpu_atomfirmware_gpu_supports_virtualization(adev))
2581 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2583 if (amdgpu_atombios_has_gpu_virtualization_table(adev))
2584 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2587 if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
2588 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
2593 * amdgpu_device_asic_has_dc_support - determine if DC supports the asic
2595 * @asic_type: AMD asic type
2597 * Check if there is DC (new modesetting infrastructre) support for an asic.
2598 * returns true if DC has support, false if not.
2600 bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
2602 switch (asic_type) {
2603 #if defined(CONFIG_DRM_AMD_DC)
2609 * We have systems in the wild with these ASICs that require
2610 * LVDS and VGA support which is not supported with DC.
2612 * Fallback to the non-DC driver here by default so as not to
2613 * cause regressions.
2615 return amdgpu_dc > 0;
2619 case CHIP_POLARIS10:
2620 case CHIP_POLARIS11:
2621 case CHIP_POLARIS12:
2628 #if defined(CONFIG_DRM_AMD_DC_DCN)
2635 return amdgpu_dc != 0;
2639 DRM_INFO("Display Core has been requested via kernel parameter "
2640 "but isn't supported by ASIC, ignoring\n");
2646 * amdgpu_device_has_dc_support - check if dc is supported
2648 * @adev: amdgpu_device_pointer
2650 * Returns true for supported, false for not supported
2652 bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
2654 if (amdgpu_sriov_vf(adev))
2657 return amdgpu_device_asic_has_dc_support(adev->asic_type);
2661 static void amdgpu_device_xgmi_reset_func(struct work_struct *__work)
2663 struct amdgpu_device *adev =
2664 container_of(__work, struct amdgpu_device, xgmi_reset_work);
2665 struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev, 0);
2667 /* It's a bug to not have a hive within this function */
2672 * Use task barrier to synchronize all xgmi reset works across the
2673 * hive. task_barrier_enter and task_barrier_exit will block
2674 * until all the threads running the xgmi reset works reach
2675 * those points. task_barrier_full will do both blocks.
2677 if (amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) {
2679 task_barrier_enter(&hive->tb);
2680 adev->asic_reset_res = amdgpu_device_baco_enter(adev->ddev);
2682 if (adev->asic_reset_res)
2685 task_barrier_exit(&hive->tb);
2686 adev->asic_reset_res = amdgpu_device_baco_exit(adev->ddev);
2688 if (adev->asic_reset_res)
2692 task_barrier_full(&hive->tb);
2693 adev->asic_reset_res = amdgpu_asic_reset(adev);
2697 if (adev->asic_reset_res)
2698 DRM_WARN("ASIC reset failed with error, %d for drm dev, %s",
2699 adev->asic_reset_res, adev->ddev->unique);
2702 static int amdgpu_device_get_job_timeout_settings(struct amdgpu_device *adev)
2704 char *input = amdgpu_lockup_timeout;
2705 char *timeout_setting = NULL;
2711 * By default timeout for non compute jobs is 10000.
2712 * And there is no timeout enforced on compute jobs.
2713 * In SR-IOV or passthrough mode, timeout for compute
2714 * jobs are 10000 by default.
2716 adev->gfx_timeout = msecs_to_jiffies(10000);
2717 adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout;
2718 if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev))
2719 adev->compute_timeout = adev->gfx_timeout;
2721 adev->compute_timeout = MAX_SCHEDULE_TIMEOUT;
2723 if (strnlen(input, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) {
2724 while ((timeout_setting = strsep(&input, ",")) &&
2725 strnlen(timeout_setting, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) {
2726 ret = kstrtol(timeout_setting, 0, &timeout);
2733 } else if (timeout < 0) {
2734 timeout = MAX_SCHEDULE_TIMEOUT;
2736 timeout = msecs_to_jiffies(timeout);
2741 adev->gfx_timeout = timeout;
2744 adev->compute_timeout = timeout;
2747 adev->sdma_timeout = timeout;
2750 adev->video_timeout = timeout;
2757 * There is only one value specified and
2758 * it should apply to all non-compute jobs.
2761 adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout;
2762 if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev))
2763 adev->compute_timeout = adev->gfx_timeout;
2771 * amdgpu_device_init - initialize the driver
2773 * @adev: amdgpu_device pointer
2774 * @ddev: drm dev pointer
2775 * @pdev: pci dev pointer
2776 * @flags: driver flags
2778 * Initializes the driver info and hw (all asics).
2779 * Returns 0 for success or an error on failure.
2780 * Called at driver startup.
2782 int amdgpu_device_init(struct amdgpu_device *adev,
2783 struct drm_device *ddev,
2784 struct pci_dev *pdev,
2791 adev->shutdown = false;
2792 adev->dev = &pdev->dev;
2795 adev->flags = flags;
2797 if (amdgpu_force_asic_type >= 0 && amdgpu_force_asic_type < CHIP_LAST)
2798 adev->asic_type = amdgpu_force_asic_type;
2800 adev->asic_type = flags & AMD_ASIC_MASK;
2802 adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
2803 if (amdgpu_emu_mode == 1)
2804 adev->usec_timeout *= 2;
2805 adev->gmc.gart_size = 512 * 1024 * 1024;
2806 adev->accel_working = false;
2807 adev->num_rings = 0;
2808 adev->mman.buffer_funcs = NULL;
2809 adev->mman.buffer_funcs_ring = NULL;
2810 adev->vm_manager.vm_pte_funcs = NULL;
2811 adev->vm_manager.vm_pte_num_scheds = 0;
2812 adev->gmc.gmc_funcs = NULL;
2813 adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
2814 bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
2816 adev->smc_rreg = &amdgpu_invalid_rreg;
2817 adev->smc_wreg = &amdgpu_invalid_wreg;
2818 adev->pcie_rreg = &amdgpu_invalid_rreg;
2819 adev->pcie_wreg = &amdgpu_invalid_wreg;
2820 adev->pciep_rreg = &amdgpu_invalid_rreg;
2821 adev->pciep_wreg = &amdgpu_invalid_wreg;
2822 adev->pcie_rreg64 = &amdgpu_invalid_rreg64;
2823 adev->pcie_wreg64 = &amdgpu_invalid_wreg64;
2824 adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
2825 adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
2826 adev->didt_rreg = &amdgpu_invalid_rreg;
2827 adev->didt_wreg = &amdgpu_invalid_wreg;
2828 adev->gc_cac_rreg = &amdgpu_invalid_rreg;
2829 adev->gc_cac_wreg = &amdgpu_invalid_wreg;
2830 adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
2831 adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
2833 DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
2834 amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
2835 pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
2837 /* mutex initialization are all done here so we
2838 * can recall function without having locking issues */
2839 atomic_set(&adev->irq.ih.lock, 0);
2840 mutex_init(&adev->firmware.mutex);
2841 mutex_init(&adev->pm.mutex);
2842 mutex_init(&adev->gfx.gpu_clock_mutex);
2843 mutex_init(&adev->srbm_mutex);
2844 mutex_init(&adev->gfx.pipe_reserve_mutex);
2845 mutex_init(&adev->gfx.gfx_off_mutex);
2846 mutex_init(&adev->grbm_idx_mutex);
2847 mutex_init(&adev->mn_lock);
2848 mutex_init(&adev->virt.vf_errors.lock);
2849 hash_init(adev->mn_hash);
2850 mutex_init(&adev->lock_reset);
2851 mutex_init(&adev->psp.mutex);
2852 mutex_init(&adev->notifier_lock);
2854 r = amdgpu_device_check_arguments(adev);
2858 spin_lock_init(&adev->mmio_idx_lock);
2859 spin_lock_init(&adev->smc_idx_lock);
2860 spin_lock_init(&adev->pcie_idx_lock);
2861 spin_lock_init(&adev->uvd_ctx_idx_lock);
2862 spin_lock_init(&adev->didt_idx_lock);
2863 spin_lock_init(&adev->gc_cac_idx_lock);
2864 spin_lock_init(&adev->se_cac_idx_lock);
2865 spin_lock_init(&adev->audio_endpt_idx_lock);
2866 spin_lock_init(&adev->mm_stats.lock);
2868 INIT_LIST_HEAD(&adev->shadow_list);
2869 mutex_init(&adev->shadow_list_lock);
2871 INIT_LIST_HEAD(&adev->ring_lru_list);
2872 spin_lock_init(&adev->ring_lru_list_lock);
2874 INIT_DELAYED_WORK(&adev->delayed_init_work,
2875 amdgpu_device_delayed_init_work_handler);
2876 INIT_DELAYED_WORK(&adev->gfx.gfx_off_delay_work,
2877 amdgpu_device_delay_enable_gfx_off);
2879 INIT_WORK(&adev->xgmi_reset_work, amdgpu_device_xgmi_reset_func);
2881 adev->gfx.gfx_off_req_count = 1;
2882 adev->pm.ac_power = power_supply_is_system_supplied() > 0 ? true : false;
2884 /* Registers mapping */
2885 /* TODO: block userspace mapping of io register */
2886 if (adev->asic_type >= CHIP_BONAIRE) {
2887 adev->rmmio_base = pci_resource_start(adev->pdev, 5);
2888 adev->rmmio_size = pci_resource_len(adev->pdev, 5);
2890 adev->rmmio_base = pci_resource_start(adev->pdev, 2);
2891 adev->rmmio_size = pci_resource_len(adev->pdev, 2);
2894 adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
2895 if (adev->rmmio == NULL) {
2898 DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
2899 DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
2901 /* io port mapping */
2902 for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
2903 if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) {
2904 adev->rio_mem_size = pci_resource_len(adev->pdev, i);
2905 adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size);
2909 if (adev->rio_mem == NULL)
2910 DRM_INFO("PCI I/O BAR is not found.\n");
2912 /* enable PCIE atomic ops */
2913 r = pci_enable_atomic_ops_to_root(adev->pdev,
2914 PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
2915 PCI_EXP_DEVCAP2_ATOMIC_COMP64);
2917 adev->have_atomics_support = false;
2918 DRM_INFO("PCIE atomic ops is not supported\n");
2920 adev->have_atomics_support = true;
2923 amdgpu_device_get_pcie_info(adev);
2926 DRM_INFO("MCBP is enabled\n");
2928 if (amdgpu_mes && adev->asic_type >= CHIP_NAVI10)
2929 adev->enable_mes = true;
2931 if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10) {
2932 r = amdgpu_discovery_init(adev);
2934 dev_err(adev->dev, "amdgpu_discovery_init failed\n");
2939 /* early init functions */
2940 r = amdgpu_device_ip_early_init(adev);
2944 r = amdgpu_device_get_job_timeout_settings(adev);
2946 dev_err(adev->dev, "invalid lockup_timeout parameter syntax\n");
2950 /* doorbell bar mapping and doorbell index init*/
2951 amdgpu_device_doorbell_init(adev);
2953 /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
2954 /* this will fail for cards that aren't VGA class devices, just
2956 vga_client_register(adev->pdev, adev, NULL, amdgpu_device_vga_set_decode);
2958 if (amdgpu_device_supports_boco(ddev))
2960 if (amdgpu_has_atpx() &&
2961 (amdgpu_is_atpx_hybrid() ||
2962 amdgpu_has_atpx_dgpu_power_cntl()) &&
2963 !pci_is_thunderbolt_attached(adev->pdev))
2964 vga_switcheroo_register_client(adev->pdev,
2965 &amdgpu_switcheroo_ops, boco);
2967 vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
2969 if (amdgpu_emu_mode == 1) {
2970 /* post the asic on emulation mode */
2971 emu_soc_asic_init(adev);
2972 goto fence_driver_init;
2975 /* detect if we are with an SRIOV vbios */
2976 amdgpu_device_detect_sriov_bios(adev);
2978 /* check if we need to reset the asic
2979 * E.g., driver was not cleanly unloaded previously, etc.
2981 if (!amdgpu_sriov_vf(adev) && amdgpu_asic_need_reset_on_init(adev)) {
2982 r = amdgpu_asic_reset(adev);
2984 dev_err(adev->dev, "asic reset on init failed\n");
2989 /* Post card if necessary */
2990 if (amdgpu_device_need_post(adev)) {
2992 dev_err(adev->dev, "no vBIOS found\n");
2996 DRM_INFO("GPU posting now...\n");
2997 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2999 dev_err(adev->dev, "gpu post error!\n");
3004 if (adev->is_atom_fw) {
3005 /* Initialize clocks */
3006 r = amdgpu_atomfirmware_get_clock_info(adev);
3008 dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
3009 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
3013 /* Initialize clocks */
3014 r = amdgpu_atombios_get_clock_info(adev);
3016 dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
3017 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
3020 /* init i2c buses */
3021 if (!amdgpu_device_has_dc_support(adev))
3022 amdgpu_atombios_i2c_init(adev);
3027 r = amdgpu_fence_driver_init(adev);
3029 dev_err(adev->dev, "amdgpu_fence_driver_init failed\n");
3030 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
3034 /* init the mode config */
3035 drm_mode_config_init(adev->ddev);
3037 r = amdgpu_device_ip_init(adev);
3039 /* failed in exclusive mode due to timeout */
3040 if (amdgpu_sriov_vf(adev) &&
3041 !amdgpu_sriov_runtime(adev) &&
3042 amdgpu_virt_mmio_blocked(adev) &&
3043 !amdgpu_virt_wait_reset(adev)) {
3044 dev_err(adev->dev, "VF exclusive mode timeout\n");
3045 /* Don't send request since VF is inactive. */
3046 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
3047 adev->virt.ops = NULL;
3051 dev_err(adev->dev, "amdgpu_device_ip_init failed\n");
3052 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
3056 DRM_DEBUG("SE %d, SH per SE %d, CU per SH %d, active_cu_number %d\n",
3057 adev->gfx.config.max_shader_engines,
3058 adev->gfx.config.max_sh_per_se,
3059 adev->gfx.config.max_cu_per_sh,
3060 adev->gfx.cu_info.number);
3062 amdgpu_ctx_init_sched(adev);
3064 adev->accel_working = true;
3066 amdgpu_vm_check_compute_bug(adev);
3068 /* Initialize the buffer migration limit. */
3069 if (amdgpu_moverate >= 0)
3070 max_MBps = amdgpu_moverate;
3072 max_MBps = 8; /* Allow 8 MB/s. */
3073 /* Get a log2 for easy divisions. */
3074 adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
3076 amdgpu_fbdev_init(adev);
3078 r = amdgpu_pm_sysfs_init(adev);
3080 adev->pm_sysfs_en = false;
3081 DRM_ERROR("registering pm debugfs failed (%d).\n", r);
3083 adev->pm_sysfs_en = true;
3085 r = amdgpu_ucode_sysfs_init(adev);
3087 adev->ucode_sysfs_en = false;
3088 DRM_ERROR("Creating firmware sysfs failed (%d).\n", r);
3090 adev->ucode_sysfs_en = true;
3092 r = amdgpu_debugfs_gem_init(adev);
3094 DRM_ERROR("registering gem debugfs failed (%d).\n", r);
3096 r = amdgpu_debugfs_regs_init(adev);
3098 DRM_ERROR("registering register debugfs failed (%d).\n", r);
3100 r = amdgpu_debugfs_firmware_init(adev);
3102 DRM_ERROR("registering firmware debugfs failed (%d).\n", r);
3104 r = amdgpu_debugfs_init(adev);
3106 DRM_ERROR("Creating debugfs files failed (%d).\n", r);
3108 if ((amdgpu_testing & 1)) {
3109 if (adev->accel_working)
3110 amdgpu_test_moves(adev);
3112 DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n");
3114 if (amdgpu_benchmarking) {
3115 if (adev->accel_working)
3116 amdgpu_benchmark(adev, amdgpu_benchmarking);
3118 DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n");
3122 * Register gpu instance before amdgpu_device_enable_mgpu_fan_boost.
3123 * Otherwise the mgpu fan boost feature will be skipped due to the
3124 * gpu instance is counted less.
3126 amdgpu_register_gpu_instance(adev);
3128 /* enable clockgating, etc. after ib tests, etc. since some blocks require
3129 * explicit gating rather than handling it automatically.
3131 r = amdgpu_device_ip_late_init(adev);
3133 dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n");
3134 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
3139 amdgpu_ras_resume(adev);
3141 queue_delayed_work(system_wq, &adev->delayed_init_work,
3142 msecs_to_jiffies(AMDGPU_RESUME_MS));
3144 r = device_create_file(adev->dev, &dev_attr_pcie_replay_count);
3146 dev_err(adev->dev, "Could not create pcie_replay_count");
3150 if (IS_ENABLED(CONFIG_PERF_EVENTS))
3151 r = amdgpu_pmu_init(adev);
3153 dev_err(adev->dev, "amdgpu_pmu_init failed\n");
3158 amdgpu_vf_error_trans_all(adev);
3160 vga_switcheroo_fini_domain_pm_ops(adev->dev);
3166 * amdgpu_device_fini - tear down the driver
3168 * @adev: amdgpu_device pointer
3170 * Tear down the driver info (all asics).
3171 * Called at driver shutdown.
3173 void amdgpu_device_fini(struct amdgpu_device *adev)
3177 DRM_INFO("amdgpu: finishing device.\n");
3178 flush_delayed_work(&adev->delayed_init_work);
3179 adev->shutdown = true;
3181 /* disable all interrupts */
3182 amdgpu_irq_disable_all(adev);
3183 if (adev->mode_info.mode_config_initialized){
3184 if (!amdgpu_device_has_dc_support(adev))
3185 drm_helper_force_disable_all(adev->ddev);
3187 drm_atomic_helper_shutdown(adev->ddev);
3189 amdgpu_fence_driver_fini(adev);
3190 if (adev->pm_sysfs_en)
3191 amdgpu_pm_sysfs_fini(adev);
3192 amdgpu_fbdev_fini(adev);
3193 r = amdgpu_device_ip_fini(adev);
3194 if (adev->firmware.gpu_info_fw) {
3195 release_firmware(adev->firmware.gpu_info_fw);
3196 adev->firmware.gpu_info_fw = NULL;
3198 adev->accel_working = false;
3199 /* free i2c buses */
3200 if (!amdgpu_device_has_dc_support(adev))
3201 amdgpu_i2c_fini(adev);
3203 if (amdgpu_emu_mode != 1)
3204 amdgpu_atombios_fini(adev);
3208 if (amdgpu_has_atpx() &&
3209 (amdgpu_is_atpx_hybrid() ||
3210 amdgpu_has_atpx_dgpu_power_cntl()) &&
3211 !pci_is_thunderbolt_attached(adev->pdev))
3212 vga_switcheroo_unregister_client(adev->pdev);
3213 if (amdgpu_device_supports_boco(adev->ddev))
3214 vga_switcheroo_fini_domain_pm_ops(adev->dev);
3215 vga_client_register(adev->pdev, NULL, NULL, NULL);
3217 pci_iounmap(adev->pdev, adev->rio_mem);
3218 adev->rio_mem = NULL;
3219 iounmap(adev->rmmio);
3221 amdgpu_device_doorbell_fini(adev);
3223 amdgpu_debugfs_regs_cleanup(adev);
3224 device_remove_file(adev->dev, &dev_attr_pcie_replay_count);
3225 if (adev->ucode_sysfs_en)
3226 amdgpu_ucode_sysfs_fini(adev);
3227 if (IS_ENABLED(CONFIG_PERF_EVENTS))
3228 amdgpu_pmu_fini(adev);
3229 amdgpu_debugfs_preempt_cleanup(adev);
3230 if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10)
3231 amdgpu_discovery_fini(adev);
3239 * amdgpu_device_suspend - initiate device suspend
3241 * @dev: drm dev pointer
3242 * @suspend: suspend state
3243 * @fbcon : notify the fbdev of suspend
3245 * Puts the hw in the suspend state (all asics).
3246 * Returns 0 for success or an error on failure.
3247 * Called at driver suspend.
3249 int amdgpu_device_suspend(struct drm_device *dev, bool fbcon)
3251 struct amdgpu_device *adev;
3252 struct drm_crtc *crtc;
3253 struct drm_connector *connector;
3254 struct drm_connector_list_iter iter;
3257 if (dev == NULL || dev->dev_private == NULL) {
3261 adev = dev->dev_private;
3263 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
3266 adev->in_suspend = true;
3267 drm_kms_helper_poll_disable(dev);
3270 amdgpu_fbdev_set_suspend(adev, 1);
3272 cancel_delayed_work_sync(&adev->delayed_init_work);
3274 if (!amdgpu_device_has_dc_support(adev)) {
3275 /* turn off display hw */
3276 drm_modeset_lock_all(dev);
3277 drm_connector_list_iter_begin(dev, &iter);
3278 drm_for_each_connector_iter(connector, &iter)
3279 drm_helper_connector_dpms(connector,
3281 drm_connector_list_iter_end(&iter);
3282 drm_modeset_unlock_all(dev);
3283 /* unpin the front buffers and cursors */
3284 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
3285 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
3286 struct drm_framebuffer *fb = crtc->primary->fb;
3287 struct amdgpu_bo *robj;
3289 if (amdgpu_crtc->cursor_bo && !adev->enable_virtual_display) {
3290 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
3291 r = amdgpu_bo_reserve(aobj, true);
3293 amdgpu_bo_unpin(aobj);
3294 amdgpu_bo_unreserve(aobj);
3298 if (fb == NULL || fb->obj[0] == NULL) {
3301 robj = gem_to_amdgpu_bo(fb->obj[0]);
3302 /* don't unpin kernel fb objects */
3303 if (!amdgpu_fbdev_robj_is_fb(adev, robj)) {
3304 r = amdgpu_bo_reserve(robj, true);
3306 amdgpu_bo_unpin(robj);
3307 amdgpu_bo_unreserve(robj);
3313 amdgpu_amdkfd_suspend(adev);
3315 amdgpu_ras_suspend(adev);
3317 r = amdgpu_device_ip_suspend_phase1(adev);
3319 /* evict vram memory */
3320 amdgpu_bo_evict_vram(adev);
3322 amdgpu_fence_driver_suspend(adev);
3324 r = amdgpu_device_ip_suspend_phase2(adev);
3326 /* evict remaining vram memory
3327 * This second call to evict vram is to evict the gart page table
3330 amdgpu_bo_evict_vram(adev);
3336 * amdgpu_device_resume - initiate device resume
3338 * @dev: drm dev pointer
3339 * @resume: resume state
3340 * @fbcon : notify the fbdev of resume
3342 * Bring the hw back to operating state (all asics).
3343 * Returns 0 for success or an error on failure.
3344 * Called at driver resume.
3346 int amdgpu_device_resume(struct drm_device *dev, bool fbcon)
3348 struct drm_connector *connector;
3349 struct drm_connector_list_iter iter;
3350 struct amdgpu_device *adev = dev->dev_private;
3351 struct drm_crtc *crtc;
3354 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
3358 if (amdgpu_device_need_post(adev)) {
3359 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
3361 DRM_ERROR("amdgpu asic init failed\n");
3364 r = amdgpu_device_ip_resume(adev);
3366 DRM_ERROR("amdgpu_device_ip_resume failed (%d).\n", r);
3369 amdgpu_fence_driver_resume(adev);
3372 r = amdgpu_device_ip_late_init(adev);
3376 queue_delayed_work(system_wq, &adev->delayed_init_work,
3377 msecs_to_jiffies(AMDGPU_RESUME_MS));
3379 if (!amdgpu_device_has_dc_support(adev)) {
3381 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
3382 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
3384 if (amdgpu_crtc->cursor_bo && !adev->enable_virtual_display) {
3385 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
3386 r = amdgpu_bo_reserve(aobj, true);
3388 r = amdgpu_bo_pin(aobj, AMDGPU_GEM_DOMAIN_VRAM);
3390 DRM_ERROR("Failed to pin cursor BO (%d)\n", r);
3391 amdgpu_crtc->cursor_addr = amdgpu_bo_gpu_offset(aobj);
3392 amdgpu_bo_unreserve(aobj);
3397 r = amdgpu_amdkfd_resume(adev);
3401 /* Make sure IB tests flushed */
3402 flush_delayed_work(&adev->delayed_init_work);
3404 /* blat the mode back in */
3406 if (!amdgpu_device_has_dc_support(adev)) {
3408 drm_helper_resume_force_mode(dev);
3410 /* turn on display hw */
3411 drm_modeset_lock_all(dev);
3413 drm_connector_list_iter_begin(dev, &iter);
3414 drm_for_each_connector_iter(connector, &iter)
3415 drm_helper_connector_dpms(connector,
3417 drm_connector_list_iter_end(&iter);
3419 drm_modeset_unlock_all(dev);
3421 amdgpu_fbdev_set_suspend(adev, 0);
3424 drm_kms_helper_poll_enable(dev);
3426 amdgpu_ras_resume(adev);
3429 * Most of the connector probing functions try to acquire runtime pm
3430 * refs to ensure that the GPU is powered on when connector polling is
3431 * performed. Since we're calling this from a runtime PM callback,
3432 * trying to acquire rpm refs will cause us to deadlock.
3434 * Since we're guaranteed to be holding the rpm lock, it's safe to
3435 * temporarily disable the rpm helpers so this doesn't deadlock us.
3438 dev->dev->power.disable_depth++;
3440 if (!amdgpu_device_has_dc_support(adev))
3441 drm_helper_hpd_irq_event(dev);
3443 drm_kms_helper_hotplug_event(dev);
3445 dev->dev->power.disable_depth--;
3447 adev->in_suspend = false;
3453 * amdgpu_device_ip_check_soft_reset - did soft reset succeed
3455 * @adev: amdgpu_device pointer
3457 * The list of all the hardware IPs that make up the asic is walked and
3458 * the check_soft_reset callbacks are run. check_soft_reset determines
3459 * if the asic is still hung or not.
3460 * Returns true if any of the IPs are still in a hung state, false if not.
3462 static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)
3465 bool asic_hang = false;
3467 if (amdgpu_sriov_vf(adev))
3470 if (amdgpu_asic_need_full_reset(adev))
3473 for (i = 0; i < adev->num_ip_blocks; i++) {
3474 if (!adev->ip_blocks[i].status.valid)
3476 if (adev->ip_blocks[i].version->funcs->check_soft_reset)
3477 adev->ip_blocks[i].status.hang =
3478 adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
3479 if (adev->ip_blocks[i].status.hang) {
3480 DRM_INFO("IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
3488 * amdgpu_device_ip_pre_soft_reset - prepare for soft reset
3490 * @adev: amdgpu_device pointer
3492 * The list of all the hardware IPs that make up the asic is walked and the
3493 * pre_soft_reset callbacks are run if the block is hung. pre_soft_reset
3494 * handles any IP specific hardware or software state changes that are
3495 * necessary for a soft reset to succeed.
3496 * Returns 0 on success, negative error code on failure.
3498 static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)
3502 for (i = 0; i < adev->num_ip_blocks; i++) {
3503 if (!adev->ip_blocks[i].status.valid)
3505 if (adev->ip_blocks[i].status.hang &&
3506 adev->ip_blocks[i].version->funcs->pre_soft_reset) {
3507 r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
3517 * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed
3519 * @adev: amdgpu_device pointer
3521 * Some hardware IPs cannot be soft reset. If they are hung, a full gpu
3522 * reset is necessary to recover.
3523 * Returns true if a full asic reset is required, false if not.
3525 static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)
3529 if (amdgpu_asic_need_full_reset(adev))
3532 for (i = 0; i < adev->num_ip_blocks; i++) {
3533 if (!adev->ip_blocks[i].status.valid)
3535 if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
3536 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
3537 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
3538 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
3539 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
3540 if (adev->ip_blocks[i].status.hang) {
3541 DRM_INFO("Some block need full reset!\n");
3550 * amdgpu_device_ip_soft_reset - do a soft reset
3552 * @adev: amdgpu_device pointer
3554 * The list of all the hardware IPs that make up the asic is walked and the
3555 * soft_reset callbacks are run if the block is hung. soft_reset handles any
3556 * IP specific hardware or software state changes that are necessary to soft
3558 * Returns 0 on success, negative error code on failure.
3560 static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)
3564 for (i = 0; i < adev->num_ip_blocks; i++) {
3565 if (!adev->ip_blocks[i].status.valid)
3567 if (adev->ip_blocks[i].status.hang &&
3568 adev->ip_blocks[i].version->funcs->soft_reset) {
3569 r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
3579 * amdgpu_device_ip_post_soft_reset - clean up from soft reset
3581 * @adev: amdgpu_device pointer
3583 * The list of all the hardware IPs that make up the asic is walked and the
3584 * post_soft_reset callbacks are run if the asic was hung. post_soft_reset
3585 * handles any IP specific hardware or software state changes that are
3586 * necessary after the IP has been soft reset.
3587 * Returns 0 on success, negative error code on failure.
3589 static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)
3593 for (i = 0; i < adev->num_ip_blocks; i++) {
3594 if (!adev->ip_blocks[i].status.valid)
3596 if (adev->ip_blocks[i].status.hang &&
3597 adev->ip_blocks[i].version->funcs->post_soft_reset)
3598 r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
3607 * amdgpu_device_recover_vram - Recover some VRAM contents
3609 * @adev: amdgpu_device pointer
3611 * Restores the contents of VRAM buffers from the shadows in GTT. Used to
3612 * restore things like GPUVM page tables after a GPU reset where
3613 * the contents of VRAM might be lost.
3616 * 0 on success, negative error code on failure.
3618 static int amdgpu_device_recover_vram(struct amdgpu_device *adev)
3620 struct dma_fence *fence = NULL, *next = NULL;
3621 struct amdgpu_bo *shadow;
3624 if (amdgpu_sriov_runtime(adev))
3625 tmo = msecs_to_jiffies(8000);
3627 tmo = msecs_to_jiffies(100);
3629 DRM_INFO("recover vram bo from shadow start\n");
3630 mutex_lock(&adev->shadow_list_lock);
3631 list_for_each_entry(shadow, &adev->shadow_list, shadow_list) {
3633 /* No need to recover an evicted BO */
3634 if (shadow->tbo.mem.mem_type != TTM_PL_TT ||
3635 shadow->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET ||
3636 shadow->parent->tbo.mem.mem_type != TTM_PL_VRAM)
3639 r = amdgpu_bo_restore_shadow(shadow, &next);
3644 tmo = dma_fence_wait_timeout(fence, false, tmo);
3645 dma_fence_put(fence);
3650 } else if (tmo < 0) {
3658 mutex_unlock(&adev->shadow_list_lock);
3661 tmo = dma_fence_wait_timeout(fence, false, tmo);
3662 dma_fence_put(fence);
3664 if (r < 0 || tmo <= 0) {
3665 DRM_ERROR("recover vram bo from shadow failed, r is %ld, tmo is %ld\n", r, tmo);
3669 DRM_INFO("recover vram bo from shadow done\n");
3675 * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
3677 * @adev: amdgpu device pointer
3678 * @from_hypervisor: request from hypervisor
3680 * do VF FLR and reinitialize Asic
3681 * return 0 means succeeded otherwise failed
3683 static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
3684 bool from_hypervisor)
3688 if (from_hypervisor)
3689 r = amdgpu_virt_request_full_gpu(adev, true);
3691 r = amdgpu_virt_reset_gpu(adev);
3695 /* Resume IP prior to SMC */
3696 r = amdgpu_device_ip_reinit_early_sriov(adev);
3700 amdgpu_virt_init_data_exchange(adev);
3701 /* we need recover gart prior to run SMC/CP/SDMA resume */
3702 amdgpu_gtt_mgr_recover(&adev->mman.bdev.man[TTM_PL_TT]);
3704 r = amdgpu_device_fw_loading(adev);
3708 /* now we are okay to resume SMC/CP/SDMA */
3709 r = amdgpu_device_ip_reinit_late_sriov(adev);
3713 amdgpu_irq_gpu_reset_resume_helper(adev);
3714 r = amdgpu_ib_ring_tests(adev);
3715 amdgpu_amdkfd_post_reset(adev);
3718 amdgpu_virt_release_full_gpu(adev, true);
3719 if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
3720 amdgpu_inc_vram_lost(adev);
3721 r = amdgpu_device_recover_vram(adev);
3728 * amdgpu_device_should_recover_gpu - check if we should try GPU recovery
3730 * @adev: amdgpu device pointer
3732 * Check amdgpu_gpu_recovery and SRIOV status to see if we should try to recover
3735 bool amdgpu_device_should_recover_gpu(struct amdgpu_device *adev)
3737 if (!amdgpu_device_ip_check_soft_reset(adev)) {
3738 DRM_INFO("Timeout, but no hardware hang detected.\n");
3742 if (amdgpu_gpu_recovery == 0)
3745 if (amdgpu_sriov_vf(adev))
3748 if (amdgpu_gpu_recovery == -1) {
3749 switch (adev->asic_type) {
3755 case CHIP_POLARIS10:
3756 case CHIP_POLARIS11:
3757 case CHIP_POLARIS12:
3777 DRM_INFO("GPU recovery disabled.\n");
3782 static int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,
3783 struct amdgpu_job *job,
3784 bool *need_full_reset_arg)
3787 bool need_full_reset = *need_full_reset_arg;
3789 /* block all schedulers and reset given job's ring */
3790 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3791 struct amdgpu_ring *ring = adev->rings[i];
3793 if (!ring || !ring->sched.thread)
3796 /* after all hw jobs are reset, hw fence is meaningless, so force_completion */
3797 amdgpu_fence_driver_force_completion(ring);
3801 drm_sched_increase_karma(&job->base);
3803 /* Don't suspend on bare metal if we are not going to HW reset the ASIC */
3804 if (!amdgpu_sriov_vf(adev)) {
3806 if (!need_full_reset)
3807 need_full_reset = amdgpu_device_ip_need_full_reset(adev);
3809 if (!need_full_reset) {
3810 amdgpu_device_ip_pre_soft_reset(adev);
3811 r = amdgpu_device_ip_soft_reset(adev);
3812 amdgpu_device_ip_post_soft_reset(adev);
3813 if (r || amdgpu_device_ip_check_soft_reset(adev)) {
3814 DRM_INFO("soft reset failed, will fallback to full reset!\n");
3815 need_full_reset = true;
3819 if (need_full_reset)
3820 r = amdgpu_device_ip_suspend(adev);
3822 *need_full_reset_arg = need_full_reset;
3828 static int amdgpu_do_asic_reset(struct amdgpu_hive_info *hive,
3829 struct list_head *device_list_handle,
3830 bool *need_full_reset_arg)
3832 struct amdgpu_device *tmp_adev = NULL;
3833 bool need_full_reset = *need_full_reset_arg, vram_lost = false;
3837 * ASIC reset has to be done on all HGMI hive nodes ASAP
3838 * to allow proper links negotiation in FW (within 1 sec)
3840 if (need_full_reset) {
3841 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
3842 /* For XGMI run all resets in parallel to speed up the process */
3843 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
3844 if (!queue_work(system_unbound_wq, &tmp_adev->xgmi_reset_work))
3847 r = amdgpu_asic_reset(tmp_adev);
3850 DRM_ERROR("ASIC reset failed with error, %d for drm dev, %s",
3851 r, tmp_adev->ddev->unique);
3856 /* For XGMI wait for all resets to complete before proceed */
3858 list_for_each_entry(tmp_adev, device_list_handle,
3860 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
3861 flush_work(&tmp_adev->xgmi_reset_work);
3862 r = tmp_adev->asic_reset_res;
3870 if (!r && amdgpu_ras_intr_triggered())
3871 amdgpu_ras_intr_cleared();
3873 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
3874 if (need_full_reset) {
3876 if (amdgpu_atom_asic_init(tmp_adev->mode_info.atom_context))
3877 DRM_WARN("asic atom init failed!");
3880 dev_info(tmp_adev->dev, "GPU reset succeeded, trying to resume\n");
3881 r = amdgpu_device_ip_resume_phase1(tmp_adev);
3885 vram_lost = amdgpu_device_check_vram_lost(tmp_adev);
3887 DRM_INFO("VRAM is lost due to GPU reset!\n");
3888 amdgpu_inc_vram_lost(tmp_adev);
3891 r = amdgpu_gtt_mgr_recover(
3892 &tmp_adev->mman.bdev.man[TTM_PL_TT]);
3896 r = amdgpu_device_fw_loading(tmp_adev);
3900 r = amdgpu_device_ip_resume_phase2(tmp_adev);
3905 amdgpu_device_fill_reset_magic(tmp_adev);
3908 * Add this ASIC as tracked as reset was already
3909 * complete successfully.
3911 amdgpu_register_gpu_instance(tmp_adev);
3913 r = amdgpu_device_ip_late_init(tmp_adev);
3918 amdgpu_ras_resume(tmp_adev);
3920 /* Update PSP FW topology after reset */
3921 if (hive && tmp_adev->gmc.xgmi.num_physical_nodes > 1)
3922 r = amdgpu_xgmi_update_topology(hive, tmp_adev);
3929 amdgpu_irq_gpu_reset_resume_helper(tmp_adev);
3930 r = amdgpu_ib_ring_tests(tmp_adev);
3932 dev_err(tmp_adev->dev, "ib ring test failed (%d).\n", r);
3933 r = amdgpu_device_ip_suspend(tmp_adev);
3934 need_full_reset = true;
3941 r = amdgpu_device_recover_vram(tmp_adev);
3943 tmp_adev->asic_reset_res = r;
3947 *need_full_reset_arg = need_full_reset;
3951 static bool amdgpu_device_lock_adev(struct amdgpu_device *adev, bool trylock)
3954 if (!mutex_trylock(&adev->lock_reset))
3957 mutex_lock(&adev->lock_reset);
3959 atomic_inc(&adev->gpu_reset_counter);
3960 adev->in_gpu_reset = true;
3961 switch (amdgpu_asic_reset_method(adev)) {
3962 case AMD_RESET_METHOD_MODE1:
3963 adev->mp1_state = PP_MP1_STATE_SHUTDOWN;
3965 case AMD_RESET_METHOD_MODE2:
3966 adev->mp1_state = PP_MP1_STATE_RESET;
3969 adev->mp1_state = PP_MP1_STATE_NONE;
3976 static void amdgpu_device_unlock_adev(struct amdgpu_device *adev)
3978 amdgpu_vf_error_trans_all(adev);
3979 adev->mp1_state = PP_MP1_STATE_NONE;
3980 adev->in_gpu_reset = false;
3981 mutex_unlock(&adev->lock_reset);
3985 * amdgpu_device_gpu_recover - reset the asic and recover scheduler
3987 * @adev: amdgpu device pointer
3988 * @job: which job trigger hang
3990 * Attempt to reset the GPU if it has hung (all asics).
3991 * Attempt to do soft-reset or full-reset and reinitialize Asic
3992 * Returns 0 for success or an error on failure.
3995 int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
3996 struct amdgpu_job *job)
3998 struct list_head device_list, *device_list_handle = NULL;
3999 bool need_full_reset, job_signaled;
4000 struct amdgpu_hive_info *hive = NULL;
4001 struct amdgpu_device *tmp_adev = NULL;
4003 bool in_ras_intr = amdgpu_ras_intr_triggered();
4005 (amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) ?
4009 * Flush RAM to disk so that after reboot
4010 * the user can read log and see why the system rebooted.
4012 if (in_ras_intr && !use_baco && amdgpu_ras_get_context(adev)->reboot) {
4014 DRM_WARN("Emergency reboot.");
4017 emergency_restart();
4020 need_full_reset = job_signaled = false;
4021 INIT_LIST_HEAD(&device_list);
4023 dev_info(adev->dev, "GPU %s begin!\n",
4024 (in_ras_intr && !use_baco) ? "jobs stop":"reset");
4026 cancel_delayed_work_sync(&adev->delayed_init_work);
4028 hive = amdgpu_get_xgmi_hive(adev, false);
4031 * Here we trylock to avoid chain of resets executing from
4032 * either trigger by jobs on different adevs in XGMI hive or jobs on
4033 * different schedulers for same device while this TO handler is running.
4034 * We always reset all schedulers for device and all devices for XGMI
4035 * hive so that should take care of them too.
4038 if (hive && !mutex_trylock(&hive->reset_lock)) {
4039 DRM_INFO("Bailing on TDR for s_job:%llx, hive: %llx as another already in progress",
4040 job ? job->base.id : -1, hive->hive_id);
4044 /* Start with adev pre asic reset first for soft reset check.*/
4045 if (!amdgpu_device_lock_adev(adev, !hive)) {
4046 DRM_INFO("Bailing on TDR for s_job:%llx, as another already in progress",
4047 job ? job->base.id : -1);
4051 /* Block kfd: SRIOV would do it separately */
4052 if (!amdgpu_sriov_vf(adev))
4053 amdgpu_amdkfd_pre_reset(adev);
4055 /* Build list of devices to reset */
4056 if (adev->gmc.xgmi.num_physical_nodes > 1) {
4058 /*unlock kfd: SRIOV would do it separately */
4059 if (!amdgpu_sriov_vf(adev))
4060 amdgpu_amdkfd_post_reset(adev);
4061 amdgpu_device_unlock_adev(adev);
4066 * In case we are in XGMI hive mode device reset is done for all the
4067 * nodes in the hive to retrain all XGMI links and hence the reset
4068 * sequence is executed in loop on all nodes.
4070 device_list_handle = &hive->device_list;
4072 list_add_tail(&adev->gmc.xgmi.head, &device_list);
4073 device_list_handle = &device_list;
4076 /* block all schedulers and reset given job's ring */
4077 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
4078 if (tmp_adev != adev) {
4079 amdgpu_device_lock_adev(tmp_adev, false);
4080 if (!amdgpu_sriov_vf(tmp_adev))
4081 amdgpu_amdkfd_pre_reset(tmp_adev);
4085 * Mark these ASICs to be reseted as untracked first
4086 * And add them back after reset completed
4088 amdgpu_unregister_gpu_instance(tmp_adev);
4090 /* disable ras on ALL IPs */
4091 if (!(in_ras_intr && !use_baco) &&
4092 amdgpu_device_ip_need_full_reset(tmp_adev))
4093 amdgpu_ras_suspend(tmp_adev);
4095 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
4096 struct amdgpu_ring *ring = tmp_adev->rings[i];
4098 if (!ring || !ring->sched.thread)
4101 drm_sched_stop(&ring->sched, job ? &job->base : NULL);
4103 if (in_ras_intr && !use_baco)
4104 amdgpu_job_stop_all_jobs_on_sched(&ring->sched);
4109 if (in_ras_intr && !use_baco)
4110 goto skip_sched_resume;
4113 * Must check guilty signal here since after this point all old
4114 * HW fences are force signaled.
4116 * job->base holds a reference to parent fence
4118 if (job && job->base.s_fence->parent &&
4119 dma_fence_is_signaled(job->base.s_fence->parent))
4120 job_signaled = true;
4123 dev_info(adev->dev, "Guilty job already signaled, skipping HW reset");
4128 /* Guilty job will be freed after this*/
4129 r = amdgpu_device_pre_asic_reset(adev, job, &need_full_reset);
4131 /*TODO Should we stop ?*/
4132 DRM_ERROR("GPU pre asic reset failed with err, %d for drm dev, %s ",
4133 r, adev->ddev->unique);
4134 adev->asic_reset_res = r;
4137 retry: /* Rest of adevs pre asic reset from XGMI hive. */
4138 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
4140 if (tmp_adev == adev)
4143 r = amdgpu_device_pre_asic_reset(tmp_adev,
4146 /*TODO Should we stop ?*/
4148 DRM_ERROR("GPU pre asic reset failed with err, %d for drm dev, %s ",
4149 r, tmp_adev->ddev->unique);
4150 tmp_adev->asic_reset_res = r;
4154 /* Actual ASIC resets if needed.*/
4155 /* TODO Implement XGMI hive reset logic for SRIOV */
4156 if (amdgpu_sriov_vf(adev)) {
4157 r = amdgpu_device_reset_sriov(adev, job ? false : true);
4159 adev->asic_reset_res = r;
4161 r = amdgpu_do_asic_reset(hive, device_list_handle, &need_full_reset);
4162 if (r && r == -EAGAIN)
4168 /* Post ASIC reset for all devs .*/
4169 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
4171 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
4172 struct amdgpu_ring *ring = tmp_adev->rings[i];
4174 if (!ring || !ring->sched.thread)
4177 /* No point to resubmit jobs if we didn't HW reset*/
4178 if (!tmp_adev->asic_reset_res && !job_signaled)
4179 drm_sched_resubmit_jobs(&ring->sched);
4181 drm_sched_start(&ring->sched, !tmp_adev->asic_reset_res);
4184 if (!amdgpu_device_has_dc_support(tmp_adev) && !job_signaled) {
4185 drm_helper_resume_force_mode(tmp_adev->ddev);
4188 tmp_adev->asic_reset_res = 0;
4191 /* bad news, how to tell it to userspace ? */
4192 dev_info(tmp_adev->dev, "GPU reset(%d) failed\n", atomic_read(&tmp_adev->gpu_reset_counter));
4193 amdgpu_vf_error_put(tmp_adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r);
4195 dev_info(tmp_adev->dev, "GPU reset(%d) succeeded!\n", atomic_read(&tmp_adev->gpu_reset_counter));
4200 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
4201 /*unlock kfd: SRIOV would do it separately */
4202 if (!(in_ras_intr && !use_baco) && !amdgpu_sriov_vf(tmp_adev))
4203 amdgpu_amdkfd_post_reset(tmp_adev);
4204 amdgpu_device_unlock_adev(tmp_adev);
4208 mutex_unlock(&hive->reset_lock);
4211 dev_info(adev->dev, "GPU reset end with ret = %d\n", r);
4216 * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot
4218 * @adev: amdgpu_device pointer
4220 * Fetchs and stores in the driver the PCIE capabilities (gen speed
4221 * and lanes) of the slot the device is in. Handles APUs and
4222 * virtualized environments where PCIE config space may not be available.
4224 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev)
4226 struct pci_dev *pdev;
4227 enum pci_bus_speed speed_cap, platform_speed_cap;
4228 enum pcie_link_width platform_link_width;
4230 if (amdgpu_pcie_gen_cap)
4231 adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
4233 if (amdgpu_pcie_lane_cap)
4234 adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
4236 /* covers APUs as well */
4237 if (pci_is_root_bus(adev->pdev->bus)) {
4238 if (adev->pm.pcie_gen_mask == 0)
4239 adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
4240 if (adev->pm.pcie_mlw_mask == 0)
4241 adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
4245 if (adev->pm.pcie_gen_mask && adev->pm.pcie_mlw_mask)
4248 pcie_bandwidth_available(adev->pdev, NULL,
4249 &platform_speed_cap, &platform_link_width);
4251 if (adev->pm.pcie_gen_mask == 0) {
4254 speed_cap = pcie_get_speed_cap(pdev);
4255 if (speed_cap == PCI_SPEED_UNKNOWN) {
4256 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4257 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4258 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
4260 if (speed_cap == PCIE_SPEED_16_0GT)
4261 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4262 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4263 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
4264 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4);
4265 else if (speed_cap == PCIE_SPEED_8_0GT)
4266 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4267 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4268 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
4269 else if (speed_cap == PCIE_SPEED_5_0GT)
4270 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4271 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2);
4273 adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1;
4276 if (platform_speed_cap == PCI_SPEED_UNKNOWN) {
4277 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4278 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
4280 if (platform_speed_cap == PCIE_SPEED_16_0GT)
4281 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4282 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4283 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
4284 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4);
4285 else if (platform_speed_cap == PCIE_SPEED_8_0GT)
4286 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4287 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4288 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3);
4289 else if (platform_speed_cap == PCIE_SPEED_5_0GT)
4290 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4291 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
4293 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
4297 if (adev->pm.pcie_mlw_mask == 0) {
4298 if (platform_link_width == PCIE_LNK_WIDTH_UNKNOWN) {
4299 adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK;
4301 switch (platform_link_width) {
4303 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
4304 CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
4305 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
4306 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4307 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4308 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4309 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4312 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
4313 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
4314 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4315 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4316 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4317 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4320 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
4321 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4322 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4323 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4324 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4327 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4328 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4329 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4330 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4333 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4334 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4335 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4338 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4339 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4342 adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;
4351 int amdgpu_device_baco_enter(struct drm_device *dev)
4353 struct amdgpu_device *adev = dev->dev_private;
4354 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
4356 if (!amdgpu_device_supports_baco(adev->ddev))
4359 if (ras && ras->supported)
4360 adev->nbio.funcs->enable_doorbell_interrupt(adev, false);
4362 return amdgpu_dpm_baco_enter(adev);
4365 int amdgpu_device_baco_exit(struct drm_device *dev)
4367 struct amdgpu_device *adev = dev->dev_private;
4368 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
4371 if (!amdgpu_device_supports_baco(adev->ddev))
4374 ret = amdgpu_dpm_baco_exit(adev);
4378 if (ras && ras->supported)
4379 adev->nbio.funcs->enable_doorbell_interrupt(adev, true);