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>
70 MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin");
71 MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin");
72 MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin");
73 MODULE_FIRMWARE("amdgpu/picasso_gpu_info.bin");
74 MODULE_FIRMWARE("amdgpu/raven2_gpu_info.bin");
75 MODULE_FIRMWARE("amdgpu/arcturus_gpu_info.bin");
76 MODULE_FIRMWARE("amdgpu/renoir_gpu_info.bin");
77 MODULE_FIRMWARE("amdgpu/navi10_gpu_info.bin");
78 MODULE_FIRMWARE("amdgpu/navi14_gpu_info.bin");
79 MODULE_FIRMWARE("amdgpu/navi12_gpu_info.bin");
81 #define AMDGPU_RESUME_MS 2000
83 const char *amdgpu_asic_name[] = {
116 * DOC: pcie_replay_count
118 * The amdgpu driver provides a sysfs API for reporting the total number
119 * of PCIe replays (NAKs)
120 * The file pcie_replay_count is used for this and returns the total
121 * number of replays as a sum of the NAKs generated and NAKs received
124 static ssize_t amdgpu_device_get_pcie_replay_count(struct device *dev,
125 struct device_attribute *attr, char *buf)
127 struct drm_device *ddev = dev_get_drvdata(dev);
128 struct amdgpu_device *adev = ddev->dev_private;
129 uint64_t cnt = amdgpu_asic_get_pcie_replay_count(adev);
131 return snprintf(buf, PAGE_SIZE, "%llu\n", cnt);
134 static DEVICE_ATTR(pcie_replay_count, S_IRUGO,
135 amdgpu_device_get_pcie_replay_count, NULL);
137 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev);
140 * amdgpu_device_is_px - Is the device is a dGPU with HG/PX power control
142 * @dev: drm_device pointer
144 * Returns true if the device is a dGPU with HG/PX power control,
145 * otherwise return false.
147 bool amdgpu_device_is_px(struct drm_device *dev)
149 struct amdgpu_device *adev = dev->dev_private;
151 if (adev->flags & AMD_IS_PX)
157 * VRAM access helper functions.
159 * amdgpu_device_vram_access - read/write a buffer in vram
161 * @adev: amdgpu_device pointer
162 * @pos: offset of the buffer in vram
163 * @buf: virtual address of the buffer in system memory
164 * @size: read/write size, sizeof(@buf) must > @size
165 * @write: true - write to vram, otherwise - read from vram
167 void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos,
168 uint32_t *buf, size_t size, bool write)
174 for (last += pos; pos <= last; pos += 4) {
175 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
176 WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)pos) | 0x80000000);
177 WREG32_NO_KIQ(mmMM_INDEX_HI, pos >> 31);
179 WREG32_NO_KIQ(mmMM_DATA, *buf++);
181 *buf++ = RREG32_NO_KIQ(mmMM_DATA);
182 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
187 * MMIO register access helper functions.
190 * amdgpu_mm_rreg - read a memory mapped IO register
192 * @adev: amdgpu_device pointer
193 * @reg: dword aligned register offset
194 * @acc_flags: access flags which require special behavior
196 * Returns the 32 bit value from the offset specified.
198 uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg,
203 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
204 return amdgpu_virt_kiq_rreg(adev, reg);
206 if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
207 ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
211 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
212 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
213 ret = readl(((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
214 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
216 trace_amdgpu_mm_rreg(adev->pdev->device, reg, ret);
221 * MMIO register read with bytes helper functions
222 * @offset:bytes offset from MMIO start
227 * amdgpu_mm_rreg8 - read a memory mapped IO register
229 * @adev: amdgpu_device pointer
230 * @offset: byte aligned register offset
232 * Returns the 8 bit value from the offset specified.
234 uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset) {
235 if (offset < adev->rmmio_size)
236 return (readb(adev->rmmio + offset));
241 * MMIO register write with bytes helper functions
242 * @offset:bytes offset from MMIO start
243 * @value: the value want to be written to the register
247 * amdgpu_mm_wreg8 - read a memory mapped IO register
249 * @adev: amdgpu_device pointer
250 * @offset: byte aligned register offset
251 * @value: 8 bit value to write
253 * Writes the value specified to the offset specified.
255 void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value) {
256 if (offset < adev->rmmio_size)
257 writeb(value, adev->rmmio + offset);
263 * amdgpu_mm_wreg - write to a memory mapped IO register
265 * @adev: amdgpu_device pointer
266 * @reg: dword aligned register offset
267 * @v: 32 bit value to write to the register
268 * @acc_flags: access flags which require special behavior
270 * Writes the value specified to the offset specified.
272 void amdgpu_mm_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v,
275 trace_amdgpu_mm_wreg(adev->pdev->device, reg, v);
277 if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
278 adev->last_mm_index = v;
281 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
282 return amdgpu_virt_kiq_wreg(adev, reg, v);
284 if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
285 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
289 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
290 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
291 writel(v, ((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
292 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
295 if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
301 * amdgpu_io_rreg - read an IO register
303 * @adev: amdgpu_device pointer
304 * @reg: dword aligned register offset
306 * Returns the 32 bit value from the offset specified.
308 u32 amdgpu_io_rreg(struct amdgpu_device *adev, u32 reg)
310 if ((reg * 4) < adev->rio_mem_size)
311 return ioread32(adev->rio_mem + (reg * 4));
313 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
314 return ioread32(adev->rio_mem + (mmMM_DATA * 4));
319 * amdgpu_io_wreg - write to an IO register
321 * @adev: amdgpu_device pointer
322 * @reg: dword aligned register offset
323 * @v: 32 bit value to write to the register
325 * Writes the value specified to the offset specified.
327 void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32 v)
329 if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
330 adev->last_mm_index = v;
333 if ((reg * 4) < adev->rio_mem_size)
334 iowrite32(v, adev->rio_mem + (reg * 4));
336 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
337 iowrite32(v, adev->rio_mem + (mmMM_DATA * 4));
340 if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
346 * amdgpu_mm_rdoorbell - read a doorbell dword
348 * @adev: amdgpu_device pointer
349 * @index: doorbell index
351 * Returns the value in the doorbell aperture at the
352 * requested doorbell index (CIK).
354 u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
356 if (index < adev->doorbell.num_doorbells) {
357 return readl(adev->doorbell.ptr + index);
359 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
365 * amdgpu_mm_wdoorbell - write a doorbell dword
367 * @adev: amdgpu_device pointer
368 * @index: doorbell index
371 * Writes @v to the doorbell aperture at the
372 * requested doorbell index (CIK).
374 void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v)
376 if (index < adev->doorbell.num_doorbells) {
377 writel(v, adev->doorbell.ptr + index);
379 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
384 * amdgpu_mm_rdoorbell64 - read a doorbell Qword
386 * @adev: amdgpu_device pointer
387 * @index: doorbell index
389 * Returns the value in the doorbell aperture at the
390 * requested doorbell index (VEGA10+).
392 u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index)
394 if (index < adev->doorbell.num_doorbells) {
395 return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index));
397 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
403 * amdgpu_mm_wdoorbell64 - write a doorbell Qword
405 * @adev: amdgpu_device pointer
406 * @index: doorbell index
409 * Writes @v to the doorbell aperture at the
410 * requested doorbell index (VEGA10+).
412 void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v)
414 if (index < adev->doorbell.num_doorbells) {
415 atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v);
417 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
422 * amdgpu_invalid_rreg - dummy reg read function
424 * @adev: amdgpu device pointer
425 * @reg: offset of register
427 * Dummy register read function. Used for register blocks
428 * that certain asics don't have (all asics).
429 * Returns the value in the register.
431 static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
433 DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
439 * amdgpu_invalid_wreg - dummy reg write function
441 * @adev: amdgpu device pointer
442 * @reg: offset of register
443 * @v: value to write to the register
445 * Dummy register read function. Used for register blocks
446 * that certain asics don't have (all asics).
448 static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
450 DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
456 * amdgpu_invalid_rreg64 - dummy 64 bit reg read function
458 * @adev: amdgpu device pointer
459 * @reg: offset of register
461 * Dummy register read function. Used for register blocks
462 * that certain asics don't have (all asics).
463 * Returns the value in the register.
465 static uint64_t amdgpu_invalid_rreg64(struct amdgpu_device *adev, uint32_t reg)
467 DRM_ERROR("Invalid callback to read 64 bit register 0x%04X\n", reg);
473 * amdgpu_invalid_wreg64 - dummy reg write function
475 * @adev: amdgpu device pointer
476 * @reg: offset of register
477 * @v: value to write to the register
479 * Dummy register read function. Used for register blocks
480 * that certain asics don't have (all asics).
482 static void amdgpu_invalid_wreg64(struct amdgpu_device *adev, uint32_t reg, uint64_t v)
484 DRM_ERROR("Invalid callback to write 64 bit register 0x%04X with 0x%08llX\n",
490 * amdgpu_block_invalid_rreg - dummy reg read function
492 * @adev: amdgpu device pointer
493 * @block: offset of instance
494 * @reg: offset of register
496 * Dummy register read function. Used for register blocks
497 * that certain asics don't have (all asics).
498 * Returns the value in the register.
500 static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
501 uint32_t block, uint32_t reg)
503 DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
510 * amdgpu_block_invalid_wreg - dummy reg write function
512 * @adev: amdgpu device pointer
513 * @block: offset of instance
514 * @reg: offset of register
515 * @v: value to write to the register
517 * Dummy register read function. Used for register blocks
518 * that certain asics don't have (all asics).
520 static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
522 uint32_t reg, uint32_t v)
524 DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
530 * amdgpu_device_vram_scratch_init - allocate the VRAM scratch page
532 * @adev: amdgpu device pointer
534 * Allocates a scratch page of VRAM for use by various things in the
537 static int amdgpu_device_vram_scratch_init(struct amdgpu_device *adev)
539 return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE,
540 PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
541 &adev->vram_scratch.robj,
542 &adev->vram_scratch.gpu_addr,
543 (void **)&adev->vram_scratch.ptr);
547 * amdgpu_device_vram_scratch_fini - Free the VRAM scratch page
549 * @adev: amdgpu device pointer
551 * Frees the VRAM scratch page.
553 static void amdgpu_device_vram_scratch_fini(struct amdgpu_device *adev)
555 amdgpu_bo_free_kernel(&adev->vram_scratch.robj, NULL, NULL);
559 * amdgpu_device_program_register_sequence - program an array of registers.
561 * @adev: amdgpu_device pointer
562 * @registers: pointer to the register array
563 * @array_size: size of the register array
565 * Programs an array or registers with and and or masks.
566 * This is a helper for setting golden registers.
568 void amdgpu_device_program_register_sequence(struct amdgpu_device *adev,
569 const u32 *registers,
570 const u32 array_size)
572 u32 tmp, reg, and_mask, or_mask;
578 for (i = 0; i < array_size; i +=3) {
579 reg = registers[i + 0];
580 and_mask = registers[i + 1];
581 or_mask = registers[i + 2];
583 if (and_mask == 0xffffffff) {
588 if (adev->family >= AMDGPU_FAMILY_AI)
589 tmp |= (or_mask & and_mask);
598 * amdgpu_device_pci_config_reset - reset the GPU
600 * @adev: amdgpu_device pointer
602 * Resets the GPU using the pci config reset sequence.
603 * Only applicable to asics prior to vega10.
605 void amdgpu_device_pci_config_reset(struct amdgpu_device *adev)
607 pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
611 * GPU doorbell aperture helpers function.
614 * amdgpu_device_doorbell_init - Init doorbell driver information.
616 * @adev: amdgpu_device pointer
618 * Init doorbell driver information (CIK)
619 * Returns 0 on success, error on failure.
621 static int amdgpu_device_doorbell_init(struct amdgpu_device *adev)
624 /* No doorbell on SI hardware generation */
625 if (adev->asic_type < CHIP_BONAIRE) {
626 adev->doorbell.base = 0;
627 adev->doorbell.size = 0;
628 adev->doorbell.num_doorbells = 0;
629 adev->doorbell.ptr = NULL;
633 if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET)
636 amdgpu_asic_init_doorbell_index(adev);
638 /* doorbell bar mapping */
639 adev->doorbell.base = pci_resource_start(adev->pdev, 2);
640 adev->doorbell.size = pci_resource_len(adev->pdev, 2);
642 adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32),
643 adev->doorbell_index.max_assignment+1);
644 if (adev->doorbell.num_doorbells == 0)
647 /* For Vega, reserve and map two pages on doorbell BAR since SDMA
648 * paging queue doorbell use the second page. The
649 * AMDGPU_DOORBELL64_MAX_ASSIGNMENT definition assumes all the
650 * doorbells are in the first page. So with paging queue enabled,
651 * the max num_doorbells should + 1 page (0x400 in dword)
653 if (adev->asic_type >= CHIP_VEGA10)
654 adev->doorbell.num_doorbells += 0x400;
656 adev->doorbell.ptr = ioremap(adev->doorbell.base,
657 adev->doorbell.num_doorbells *
659 if (adev->doorbell.ptr == NULL)
666 * amdgpu_device_doorbell_fini - Tear down doorbell driver information.
668 * @adev: amdgpu_device pointer
670 * Tear down doorbell driver information (CIK)
672 static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev)
674 iounmap(adev->doorbell.ptr);
675 adev->doorbell.ptr = NULL;
681 * amdgpu_device_wb_*()
682 * Writeback is the method by which the GPU updates special pages in memory
683 * with the status of certain GPU events (fences, ring pointers,etc.).
687 * amdgpu_device_wb_fini - Disable Writeback and free memory
689 * @adev: amdgpu_device pointer
691 * Disables Writeback and frees the Writeback memory (all asics).
692 * Used at driver shutdown.
694 static void amdgpu_device_wb_fini(struct amdgpu_device *adev)
696 if (adev->wb.wb_obj) {
697 amdgpu_bo_free_kernel(&adev->wb.wb_obj,
699 (void **)&adev->wb.wb);
700 adev->wb.wb_obj = NULL;
705 * amdgpu_device_wb_init- Init Writeback driver info and allocate memory
707 * @adev: amdgpu_device pointer
709 * Initializes writeback and allocates writeback memory (all asics).
710 * Used at driver startup.
711 * Returns 0 on success or an -error on failure.
713 static int amdgpu_device_wb_init(struct amdgpu_device *adev)
717 if (adev->wb.wb_obj == NULL) {
718 /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */
719 r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8,
720 PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
721 &adev->wb.wb_obj, &adev->wb.gpu_addr,
722 (void **)&adev->wb.wb);
724 dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
728 adev->wb.num_wb = AMDGPU_MAX_WB;
729 memset(&adev->wb.used, 0, sizeof(adev->wb.used));
731 /* clear wb memory */
732 memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8);
739 * amdgpu_device_wb_get - Allocate a wb entry
741 * @adev: amdgpu_device pointer
744 * Allocate a wb slot for use by the driver (all asics).
745 * Returns 0 on success or -EINVAL on failure.
747 int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb)
749 unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
751 if (offset < adev->wb.num_wb) {
752 __set_bit(offset, adev->wb.used);
753 *wb = offset << 3; /* convert to dw offset */
761 * amdgpu_device_wb_free - Free a wb entry
763 * @adev: amdgpu_device pointer
766 * Free a wb slot allocated for use by the driver (all asics)
768 void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb)
771 if (wb < adev->wb.num_wb)
772 __clear_bit(wb, adev->wb.used);
776 * amdgpu_device_resize_fb_bar - try to resize FB BAR
778 * @adev: amdgpu_device pointer
780 * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not
781 * to fail, but if any of the BARs is not accessible after the size we abort
782 * driver loading by returning -ENODEV.
784 int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev)
786 u64 space_needed = roundup_pow_of_two(adev->gmc.real_vram_size);
787 u32 rbar_size = order_base_2(((space_needed >> 20) | 1)) - 1;
788 struct pci_bus *root;
789 struct resource *res;
795 if (amdgpu_sriov_vf(adev))
798 /* Check if the root BUS has 64bit memory resources */
799 root = adev->pdev->bus;
803 pci_bus_for_each_resource(root, res, i) {
804 if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) &&
805 res->start > 0x100000000ull)
809 /* Trying to resize is pointless without a root hub window above 4GB */
813 /* Disable memory decoding while we change the BAR addresses and size */
814 pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd);
815 pci_write_config_word(adev->pdev, PCI_COMMAND,
816 cmd & ~PCI_COMMAND_MEMORY);
818 /* Free the VRAM and doorbell BAR, we most likely need to move both. */
819 amdgpu_device_doorbell_fini(adev);
820 if (adev->asic_type >= CHIP_BONAIRE)
821 pci_release_resource(adev->pdev, 2);
823 pci_release_resource(adev->pdev, 0);
825 r = pci_resize_resource(adev->pdev, 0, rbar_size);
827 DRM_INFO("Not enough PCI address space for a large BAR.");
828 else if (r && r != -ENOTSUPP)
829 DRM_ERROR("Problem resizing BAR0 (%d).", r);
831 pci_assign_unassigned_bus_resources(adev->pdev->bus);
833 /* When the doorbell or fb BAR isn't available we have no chance of
836 r = amdgpu_device_doorbell_init(adev);
837 if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET))
840 pci_write_config_word(adev->pdev, PCI_COMMAND, cmd);
846 * GPU helpers function.
849 * amdgpu_device_need_post - check if the hw need post or not
851 * @adev: amdgpu_device pointer
853 * Check if the asic has been initialized (all asics) at driver startup
854 * or post is needed if hw reset is performed.
855 * Returns true if need or false if not.
857 bool amdgpu_device_need_post(struct amdgpu_device *adev)
861 if (amdgpu_sriov_vf(adev))
864 if (amdgpu_passthrough(adev)) {
865 /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
866 * some old smc fw still need driver do vPost otherwise gpu hang, while
867 * those smc fw version above 22.15 doesn't have this flaw, so we force
868 * vpost executed for smc version below 22.15
870 if (adev->asic_type == CHIP_FIJI) {
873 err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
874 /* force vPost if error occured */
878 fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
879 if (fw_ver < 0x00160e00)
884 if (adev->has_hw_reset) {
885 adev->has_hw_reset = false;
889 /* bios scratch used on CIK+ */
890 if (adev->asic_type >= CHIP_BONAIRE)
891 return amdgpu_atombios_scratch_need_asic_init(adev);
893 /* check MEM_SIZE for older asics */
894 reg = amdgpu_asic_get_config_memsize(adev);
896 if ((reg != 0) && (reg != 0xffffffff))
902 /* if we get transitioned to only one device, take VGA back */
904 * amdgpu_device_vga_set_decode - enable/disable vga decode
906 * @cookie: amdgpu_device pointer
907 * @state: enable/disable vga decode
909 * Enable/disable vga decode (all asics).
910 * Returns VGA resource flags.
912 static unsigned int amdgpu_device_vga_set_decode(void *cookie, bool state)
914 struct amdgpu_device *adev = cookie;
915 amdgpu_asic_set_vga_state(adev, state);
917 return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
918 VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
920 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
924 * amdgpu_device_check_block_size - validate the vm block size
926 * @adev: amdgpu_device pointer
928 * Validates the vm block size specified via module parameter.
929 * The vm block size defines number of bits in page table versus page directory,
930 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
931 * page table and the remaining bits are in the page directory.
933 static void amdgpu_device_check_block_size(struct amdgpu_device *adev)
935 /* defines number of bits in page table versus page directory,
936 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
937 * page table and the remaining bits are in the page directory */
938 if (amdgpu_vm_block_size == -1)
941 if (amdgpu_vm_block_size < 9) {
942 dev_warn(adev->dev, "VM page table size (%d) too small\n",
943 amdgpu_vm_block_size);
944 amdgpu_vm_block_size = -1;
949 * amdgpu_device_check_vm_size - validate the vm size
951 * @adev: amdgpu_device pointer
953 * Validates the vm size in GB specified via module parameter.
954 * The VM size is the size of the GPU virtual memory space in GB.
956 static void amdgpu_device_check_vm_size(struct amdgpu_device *adev)
958 /* no need to check the default value */
959 if (amdgpu_vm_size == -1)
962 if (amdgpu_vm_size < 1) {
963 dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
969 static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev)
972 bool is_os_64 = (sizeof(void *) == 8) ? true : false;
973 uint64_t total_memory;
974 uint64_t dram_size_seven_GB = 0x1B8000000;
975 uint64_t dram_size_three_GB = 0xB8000000;
977 if (amdgpu_smu_memory_pool_size == 0)
981 DRM_WARN("Not 64-bit OS, feature not supported\n");
985 total_memory = (uint64_t)si.totalram * si.mem_unit;
987 if ((amdgpu_smu_memory_pool_size == 1) ||
988 (amdgpu_smu_memory_pool_size == 2)) {
989 if (total_memory < dram_size_three_GB)
991 } else if ((amdgpu_smu_memory_pool_size == 4) ||
992 (amdgpu_smu_memory_pool_size == 8)) {
993 if (total_memory < dram_size_seven_GB)
996 DRM_WARN("Smu memory pool size not supported\n");
999 adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28;
1004 DRM_WARN("No enough system memory\n");
1006 adev->pm.smu_prv_buffer_size = 0;
1010 * amdgpu_device_check_arguments - validate module params
1012 * @adev: amdgpu_device pointer
1014 * Validates certain module parameters and updates
1015 * the associated values used by the driver (all asics).
1017 static int amdgpu_device_check_arguments(struct amdgpu_device *adev)
1021 if (amdgpu_sched_jobs < 4) {
1022 dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
1024 amdgpu_sched_jobs = 4;
1025 } else if (!is_power_of_2(amdgpu_sched_jobs)){
1026 dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
1028 amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
1031 if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) {
1032 /* gart size must be greater or equal to 32M */
1033 dev_warn(adev->dev, "gart size (%d) too small\n",
1035 amdgpu_gart_size = -1;
1038 if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) {
1039 /* gtt size must be greater or equal to 32M */
1040 dev_warn(adev->dev, "gtt size (%d) too small\n",
1042 amdgpu_gtt_size = -1;
1045 /* valid range is between 4 and 9 inclusive */
1046 if (amdgpu_vm_fragment_size != -1 &&
1047 (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) {
1048 dev_warn(adev->dev, "valid range is between 4 and 9\n");
1049 amdgpu_vm_fragment_size = -1;
1052 amdgpu_device_check_smu_prv_buffer_size(adev);
1054 amdgpu_device_check_vm_size(adev);
1056 amdgpu_device_check_block_size(adev);
1058 adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type);
1064 * amdgpu_switcheroo_set_state - set switcheroo state
1066 * @pdev: pci dev pointer
1067 * @state: vga_switcheroo state
1069 * Callback for the switcheroo driver. Suspends or resumes the
1070 * the asics before or after it is powered up using ACPI methods.
1072 static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, enum vga_switcheroo_state state)
1074 struct drm_device *dev = pci_get_drvdata(pdev);
1076 if (amdgpu_device_is_px(dev) && state == VGA_SWITCHEROO_OFF)
1079 if (state == VGA_SWITCHEROO_ON) {
1080 pr_info("amdgpu: switched on\n");
1081 /* don't suspend or resume card normally */
1082 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1084 amdgpu_device_resume(dev, true, true);
1086 dev->switch_power_state = DRM_SWITCH_POWER_ON;
1087 drm_kms_helper_poll_enable(dev);
1089 pr_info("amdgpu: switched off\n");
1090 drm_kms_helper_poll_disable(dev);
1091 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1092 amdgpu_device_suspend(dev, true, true);
1093 dev->switch_power_state = DRM_SWITCH_POWER_OFF;
1098 * amdgpu_switcheroo_can_switch - see if switcheroo state can change
1100 * @pdev: pci dev pointer
1102 * Callback for the switcheroo driver. Check of the switcheroo
1103 * state can be changed.
1104 * Returns true if the state can be changed, false if not.
1106 static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
1108 struct drm_device *dev = pci_get_drvdata(pdev);
1111 * FIXME: open_count is protected by drm_global_mutex but that would lead to
1112 * locking inversion with the driver load path. And the access here is
1113 * completely racy anyway. So don't bother with locking for now.
1115 return dev->open_count == 0;
1118 static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
1119 .set_gpu_state = amdgpu_switcheroo_set_state,
1121 .can_switch = amdgpu_switcheroo_can_switch,
1125 * amdgpu_device_ip_set_clockgating_state - set the CG state
1127 * @dev: amdgpu_device pointer
1128 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1129 * @state: clockgating state (gate or ungate)
1131 * Sets the requested clockgating state for all instances of
1132 * the hardware IP specified.
1133 * Returns the error code from the last instance.
1135 int amdgpu_device_ip_set_clockgating_state(void *dev,
1136 enum amd_ip_block_type block_type,
1137 enum amd_clockgating_state state)
1139 struct amdgpu_device *adev = dev;
1142 for (i = 0; i < adev->num_ip_blocks; i++) {
1143 if (!adev->ip_blocks[i].status.valid)
1145 if (adev->ip_blocks[i].version->type != block_type)
1147 if (!adev->ip_blocks[i].version->funcs->set_clockgating_state)
1149 r = adev->ip_blocks[i].version->funcs->set_clockgating_state(
1150 (void *)adev, state);
1152 DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",
1153 adev->ip_blocks[i].version->funcs->name, r);
1159 * amdgpu_device_ip_set_powergating_state - set the PG state
1161 * @dev: amdgpu_device pointer
1162 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1163 * @state: powergating state (gate or ungate)
1165 * Sets the requested powergating state for all instances of
1166 * the hardware IP specified.
1167 * Returns the error code from the last instance.
1169 int amdgpu_device_ip_set_powergating_state(void *dev,
1170 enum amd_ip_block_type block_type,
1171 enum amd_powergating_state state)
1173 struct amdgpu_device *adev = dev;
1176 for (i = 0; i < adev->num_ip_blocks; i++) {
1177 if (!adev->ip_blocks[i].status.valid)
1179 if (adev->ip_blocks[i].version->type != block_type)
1181 if (!adev->ip_blocks[i].version->funcs->set_powergating_state)
1183 r = adev->ip_blocks[i].version->funcs->set_powergating_state(
1184 (void *)adev, state);
1186 DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",
1187 adev->ip_blocks[i].version->funcs->name, r);
1193 * amdgpu_device_ip_get_clockgating_state - get the CG state
1195 * @adev: amdgpu_device pointer
1196 * @flags: clockgating feature flags
1198 * Walks the list of IPs on the device and updates the clockgating
1199 * flags for each IP.
1200 * Updates @flags with the feature flags for each hardware IP where
1201 * clockgating is enabled.
1203 void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev,
1208 for (i = 0; i < adev->num_ip_blocks; i++) {
1209 if (!adev->ip_blocks[i].status.valid)
1211 if (adev->ip_blocks[i].version->funcs->get_clockgating_state)
1212 adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags);
1217 * amdgpu_device_ip_wait_for_idle - wait for idle
1219 * @adev: amdgpu_device pointer
1220 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1222 * Waits for the request hardware IP to be idle.
1223 * Returns 0 for success or a negative error code on failure.
1225 int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev,
1226 enum amd_ip_block_type block_type)
1230 for (i = 0; i < adev->num_ip_blocks; i++) {
1231 if (!adev->ip_blocks[i].status.valid)
1233 if (adev->ip_blocks[i].version->type == block_type) {
1234 r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
1245 * amdgpu_device_ip_is_idle - is the hardware IP idle
1247 * @adev: amdgpu_device pointer
1248 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1250 * Check if the hardware IP is idle or not.
1251 * Returns true if it the IP is idle, false if not.
1253 bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev,
1254 enum amd_ip_block_type block_type)
1258 for (i = 0; i < adev->num_ip_blocks; i++) {
1259 if (!adev->ip_blocks[i].status.valid)
1261 if (adev->ip_blocks[i].version->type == block_type)
1262 return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
1269 * amdgpu_device_ip_get_ip_block - get a hw IP pointer
1271 * @adev: amdgpu_device pointer
1272 * @type: Type of hardware IP (SMU, GFX, UVD, etc.)
1274 * Returns a pointer to the hardware IP block structure
1275 * if it exists for the asic, otherwise NULL.
1277 struct amdgpu_ip_block *
1278 amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,
1279 enum amd_ip_block_type type)
1283 for (i = 0; i < adev->num_ip_blocks; i++)
1284 if (adev->ip_blocks[i].version->type == type)
1285 return &adev->ip_blocks[i];
1291 * amdgpu_device_ip_block_version_cmp
1293 * @adev: amdgpu_device pointer
1294 * @type: enum amd_ip_block_type
1295 * @major: major version
1296 * @minor: minor version
1298 * return 0 if equal or greater
1299 * return 1 if smaller or the ip_block doesn't exist
1301 int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev,
1302 enum amd_ip_block_type type,
1303 u32 major, u32 minor)
1305 struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type);
1307 if (ip_block && ((ip_block->version->major > major) ||
1308 ((ip_block->version->major == major) &&
1309 (ip_block->version->minor >= minor))))
1316 * amdgpu_device_ip_block_add
1318 * @adev: amdgpu_device pointer
1319 * @ip_block_version: pointer to the IP to add
1321 * Adds the IP block driver information to the collection of IPs
1324 int amdgpu_device_ip_block_add(struct amdgpu_device *adev,
1325 const struct amdgpu_ip_block_version *ip_block_version)
1327 if (!ip_block_version)
1330 DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks,
1331 ip_block_version->funcs->name);
1333 adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
1339 * amdgpu_device_enable_virtual_display - enable virtual display feature
1341 * @adev: amdgpu_device pointer
1343 * Enabled the virtual display feature if the user has enabled it via
1344 * the module parameter virtual_display. This feature provides a virtual
1345 * display hardware on headless boards or in virtualized environments.
1346 * This function parses and validates the configuration string specified by
1347 * the user and configues the virtual display configuration (number of
1348 * virtual connectors, crtcs, etc.) specified.
1350 static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
1352 adev->enable_virtual_display = false;
1354 if (amdgpu_virtual_display) {
1355 struct drm_device *ddev = adev->ddev;
1356 const char *pci_address_name = pci_name(ddev->pdev);
1357 char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
1359 pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
1360 pciaddstr_tmp = pciaddstr;
1361 while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
1362 pciaddname = strsep(&pciaddname_tmp, ",");
1363 if (!strcmp("all", pciaddname)
1364 || !strcmp(pci_address_name, pciaddname)) {
1368 adev->enable_virtual_display = true;
1371 res = kstrtol(pciaddname_tmp, 10,
1379 adev->mode_info.num_crtc = num_crtc;
1381 adev->mode_info.num_crtc = 1;
1387 DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
1388 amdgpu_virtual_display, pci_address_name,
1389 adev->enable_virtual_display, adev->mode_info.num_crtc);
1396 * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware
1398 * @adev: amdgpu_device pointer
1400 * Parses the asic configuration parameters specified in the gpu info
1401 * firmware and makes them availale to the driver for use in configuring
1403 * Returns 0 on success, -EINVAL on failure.
1405 static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev)
1407 const char *chip_name;
1410 const struct gpu_info_firmware_header_v1_0 *hdr;
1412 adev->firmware.gpu_info_fw = NULL;
1414 switch (adev->asic_type) {
1418 case CHIP_POLARIS10:
1419 case CHIP_POLARIS11:
1420 case CHIP_POLARIS12:
1424 #ifdef CONFIG_DRM_AMDGPU_SI
1431 #ifdef CONFIG_DRM_AMDGPU_CIK
1442 chip_name = "vega10";
1445 chip_name = "vega12";
1448 if (adev->rev_id >= 8)
1449 chip_name = "raven2";
1450 else if (adev->pdev->device == 0x15d8)
1451 chip_name = "picasso";
1453 chip_name = "raven";
1456 chip_name = "arcturus";
1459 chip_name = "renoir";
1462 chip_name = "navi10";
1465 chip_name = "navi14";
1468 chip_name = "navi12";
1472 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name);
1473 err = request_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev);
1476 "Failed to load gpu_info firmware \"%s\"\n",
1480 err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw);
1483 "Failed to validate gpu_info firmware \"%s\"\n",
1488 hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data;
1489 amdgpu_ucode_print_gpu_info_hdr(&hdr->header);
1491 switch (hdr->version_major) {
1494 const struct gpu_info_firmware_v1_0 *gpu_info_fw =
1495 (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data +
1496 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1498 if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10)
1499 goto parse_soc_bounding_box;
1501 adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se);
1502 adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh);
1503 adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se);
1504 adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se);
1505 adev->gfx.config.max_texture_channel_caches =
1506 le32_to_cpu(gpu_info_fw->gc_num_tccs);
1507 adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs);
1508 adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds);
1509 adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth);
1510 adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth);
1511 adev->gfx.config.double_offchip_lds_buf =
1512 le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer);
1513 adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size);
1514 adev->gfx.cu_info.max_waves_per_simd =
1515 le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd);
1516 adev->gfx.cu_info.max_scratch_slots_per_cu =
1517 le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu);
1518 adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size);
1519 if (hdr->version_minor >= 1) {
1520 const struct gpu_info_firmware_v1_1 *gpu_info_fw =
1521 (const struct gpu_info_firmware_v1_1 *)(adev->firmware.gpu_info_fw->data +
1522 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1523 adev->gfx.config.num_sc_per_sh =
1524 le32_to_cpu(gpu_info_fw->num_sc_per_sh);
1525 adev->gfx.config.num_packer_per_sc =
1526 le32_to_cpu(gpu_info_fw->num_packer_per_sc);
1529 parse_soc_bounding_box:
1531 * soc bounding box info is not integrated in disocovery table,
1532 * we always need to parse it from gpu info firmware.
1534 if (hdr->version_minor == 2) {
1535 const struct gpu_info_firmware_v1_2 *gpu_info_fw =
1536 (const struct gpu_info_firmware_v1_2 *)(adev->firmware.gpu_info_fw->data +
1537 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1538 adev->dm.soc_bounding_box = &gpu_info_fw->soc_bounding_box;
1544 "Unsupported gpu_info table %d\n", hdr->header.ucode_version);
1553 * amdgpu_device_ip_early_init - run early init for hardware IPs
1555 * @adev: amdgpu_device pointer
1557 * Early initialization pass for hardware IPs. The hardware IPs that make
1558 * up each asic are discovered each IP's early_init callback is run. This
1559 * is the first stage in initializing the asic.
1560 * Returns 0 on success, negative error code on failure.
1562 static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
1566 amdgpu_device_enable_virtual_display(adev);
1568 switch (adev->asic_type) {
1572 case CHIP_POLARIS10:
1573 case CHIP_POLARIS11:
1574 case CHIP_POLARIS12:
1578 if (adev->asic_type == CHIP_CARRIZO || adev->asic_type == CHIP_STONEY)
1579 adev->family = AMDGPU_FAMILY_CZ;
1581 adev->family = AMDGPU_FAMILY_VI;
1583 r = vi_set_ip_blocks(adev);
1587 #ifdef CONFIG_DRM_AMDGPU_SI
1593 adev->family = AMDGPU_FAMILY_SI;
1594 r = si_set_ip_blocks(adev);
1599 #ifdef CONFIG_DRM_AMDGPU_CIK
1605 if ((adev->asic_type == CHIP_BONAIRE) || (adev->asic_type == CHIP_HAWAII))
1606 adev->family = AMDGPU_FAMILY_CI;
1608 adev->family = AMDGPU_FAMILY_KV;
1610 r = cik_set_ip_blocks(adev);
1621 if (adev->asic_type == CHIP_RAVEN ||
1622 adev->asic_type == CHIP_RENOIR)
1623 adev->family = AMDGPU_FAMILY_RV;
1625 adev->family = AMDGPU_FAMILY_AI;
1627 r = soc15_set_ip_blocks(adev);
1634 adev->family = AMDGPU_FAMILY_NV;
1636 r = nv_set_ip_blocks(adev);
1641 /* FIXME: not supported yet */
1645 r = amdgpu_device_parse_gpu_info_fw(adev);
1649 if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10)
1650 amdgpu_discovery_get_gfx_info(adev);
1652 amdgpu_amdkfd_device_probe(adev);
1654 if (amdgpu_sriov_vf(adev)) {
1655 r = amdgpu_virt_request_full_gpu(adev, true);
1660 adev->pm.pp_feature = amdgpu_pp_feature_mask;
1661 if (amdgpu_sriov_vf(adev) || sched_policy == KFD_SCHED_POLICY_NO_HWS)
1662 adev->pm.pp_feature &= ~PP_GFXOFF_MASK;
1664 for (i = 0; i < adev->num_ip_blocks; i++) {
1665 if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
1666 DRM_ERROR("disabled ip block: %d <%s>\n",
1667 i, adev->ip_blocks[i].version->funcs->name);
1668 adev->ip_blocks[i].status.valid = false;
1670 if (adev->ip_blocks[i].version->funcs->early_init) {
1671 r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
1673 adev->ip_blocks[i].status.valid = false;
1675 DRM_ERROR("early_init of IP block <%s> failed %d\n",
1676 adev->ip_blocks[i].version->funcs->name, r);
1679 adev->ip_blocks[i].status.valid = true;
1682 adev->ip_blocks[i].status.valid = true;
1685 /* get the vbios after the asic_funcs are set up */
1686 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) {
1688 if (!amdgpu_get_bios(adev))
1691 r = amdgpu_atombios_init(adev);
1693 dev_err(adev->dev, "amdgpu_atombios_init failed\n");
1694 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0);
1700 adev->cg_flags &= amdgpu_cg_mask;
1701 adev->pg_flags &= amdgpu_pg_mask;
1706 static int amdgpu_device_ip_hw_init_phase1(struct amdgpu_device *adev)
1710 for (i = 0; i < adev->num_ip_blocks; i++) {
1711 if (!adev->ip_blocks[i].status.sw)
1713 if (adev->ip_blocks[i].status.hw)
1715 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
1716 (amdgpu_sriov_vf(adev) && (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)) ||
1717 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
1718 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
1720 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1721 adev->ip_blocks[i].version->funcs->name, r);
1724 adev->ip_blocks[i].status.hw = true;
1731 static int amdgpu_device_ip_hw_init_phase2(struct amdgpu_device *adev)
1735 for (i = 0; i < adev->num_ip_blocks; i++) {
1736 if (!adev->ip_blocks[i].status.sw)
1738 if (adev->ip_blocks[i].status.hw)
1740 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
1742 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1743 adev->ip_blocks[i].version->funcs->name, r);
1746 adev->ip_blocks[i].status.hw = true;
1752 static int amdgpu_device_fw_loading(struct amdgpu_device *adev)
1756 uint32_t smu_version;
1758 if (adev->asic_type >= CHIP_VEGA10) {
1759 for (i = 0; i < adev->num_ip_blocks; i++) {
1760 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_PSP)
1763 /* no need to do the fw loading again if already done*/
1764 if (adev->ip_blocks[i].status.hw == true)
1767 if (adev->in_gpu_reset || adev->in_suspend) {
1768 r = adev->ip_blocks[i].version->funcs->resume(adev);
1770 DRM_ERROR("resume of IP block <%s> failed %d\n",
1771 adev->ip_blocks[i].version->funcs->name, r);
1775 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
1777 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1778 adev->ip_blocks[i].version->funcs->name, r);
1783 adev->ip_blocks[i].status.hw = true;
1788 r = amdgpu_pm_load_smu_firmware(adev, &smu_version);
1794 * amdgpu_device_ip_init - run init for hardware IPs
1796 * @adev: amdgpu_device pointer
1798 * Main initialization pass for hardware IPs. The list of all the hardware
1799 * IPs that make up the asic is walked and the sw_init and hw_init callbacks
1800 * are run. sw_init initializes the software state associated with each IP
1801 * and hw_init initializes the hardware associated with each IP.
1802 * Returns 0 on success, negative error code on failure.
1804 static int amdgpu_device_ip_init(struct amdgpu_device *adev)
1808 r = amdgpu_ras_init(adev);
1812 for (i = 0; i < adev->num_ip_blocks; i++) {
1813 if (!adev->ip_blocks[i].status.valid)
1815 r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
1817 DRM_ERROR("sw_init of IP block <%s> failed %d\n",
1818 adev->ip_blocks[i].version->funcs->name, r);
1821 adev->ip_blocks[i].status.sw = true;
1823 /* need to do gmc hw init early so we can allocate gpu mem */
1824 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1825 r = amdgpu_device_vram_scratch_init(adev);
1827 DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r);
1830 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1832 DRM_ERROR("hw_init %d failed %d\n", i, r);
1835 r = amdgpu_device_wb_init(adev);
1837 DRM_ERROR("amdgpu_device_wb_init failed %d\n", r);
1840 adev->ip_blocks[i].status.hw = true;
1842 /* right after GMC hw init, we create CSA */
1843 if (amdgpu_mcbp || amdgpu_sriov_vf(adev)) {
1844 r = amdgpu_allocate_static_csa(adev, &adev->virt.csa_obj,
1845 AMDGPU_GEM_DOMAIN_VRAM,
1848 DRM_ERROR("allocate CSA failed %d\n", r);
1855 r = amdgpu_ib_pool_init(adev);
1857 dev_err(adev->dev, "IB initialization failed (%d).\n", r);
1858 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r);
1862 r = amdgpu_ucode_create_bo(adev); /* create ucode bo when sw_init complete*/
1866 r = amdgpu_device_ip_hw_init_phase1(adev);
1870 r = amdgpu_device_fw_loading(adev);
1874 r = amdgpu_device_ip_hw_init_phase2(adev);
1879 * retired pages will be loaded from eeprom and reserved here,
1880 * it should be called after amdgpu_device_ip_hw_init_phase2 since
1881 * for some ASICs the RAS EEPROM code relies on SMU fully functioning
1882 * for I2C communication which only true at this point.
1883 * recovery_init may fail, but it can free all resources allocated by
1884 * itself and its failure should not stop amdgpu init process.
1886 * Note: theoretically, this should be called before all vram allocations
1887 * to protect retired page from abusing
1889 amdgpu_ras_recovery_init(adev);
1891 if (adev->gmc.xgmi.num_physical_nodes > 1)
1892 amdgpu_xgmi_add_device(adev);
1893 amdgpu_amdkfd_device_init(adev);
1896 if (amdgpu_sriov_vf(adev)) {
1898 amdgpu_virt_init_data_exchange(adev);
1899 amdgpu_virt_release_full_gpu(adev, true);
1906 * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer
1908 * @adev: amdgpu_device pointer
1910 * Writes a reset magic value to the gart pointer in VRAM. The driver calls
1911 * this function before a GPU reset. If the value is retained after a
1912 * GPU reset, VRAM has not been lost. Some GPU resets may destry VRAM contents.
1914 static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev)
1916 memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
1920 * amdgpu_device_check_vram_lost - check if vram is valid
1922 * @adev: amdgpu_device pointer
1924 * Checks the reset magic value written to the gart pointer in VRAM.
1925 * The driver calls this after a GPU reset to see if the contents of
1926 * VRAM is lost or now.
1927 * returns true if vram is lost, false if not.
1929 static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev)
1931 return !!memcmp(adev->gart.ptr, adev->reset_magic,
1932 AMDGPU_RESET_MAGIC_NUM);
1936 * amdgpu_device_set_cg_state - set clockgating for amdgpu device
1938 * @adev: amdgpu_device pointer
1940 * The list of all the hardware IPs that make up the asic is walked and the
1941 * set_clockgating_state callbacks are run.
1942 * Late initialization pass enabling clockgating for hardware IPs.
1943 * Fini or suspend, pass disabling clockgating for hardware IPs.
1944 * Returns 0 on success, negative error code on failure.
1947 static int amdgpu_device_set_cg_state(struct amdgpu_device *adev,
1948 enum amd_clockgating_state state)
1952 if (amdgpu_emu_mode == 1)
1955 for (j = 0; j < adev->num_ip_blocks; j++) {
1956 i = state == AMD_CG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
1957 if (!adev->ip_blocks[i].status.late_initialized)
1959 /* skip CG for VCE/UVD, it's handled specially */
1960 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1961 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1962 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
1963 adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1964 /* enable clockgating to save power */
1965 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1968 DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
1969 adev->ip_blocks[i].version->funcs->name, r);
1978 static int amdgpu_device_set_pg_state(struct amdgpu_device *adev, enum amd_powergating_state state)
1982 if (amdgpu_emu_mode == 1)
1985 for (j = 0; j < adev->num_ip_blocks; j++) {
1986 i = state == AMD_PG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
1987 if (!adev->ip_blocks[i].status.late_initialized)
1989 /* skip CG for VCE/UVD, it's handled specially */
1990 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1991 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1992 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
1993 adev->ip_blocks[i].version->funcs->set_powergating_state) {
1994 /* enable powergating to save power */
1995 r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev,
1998 DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n",
1999 adev->ip_blocks[i].version->funcs->name, r);
2007 static int amdgpu_device_enable_mgpu_fan_boost(void)
2009 struct amdgpu_gpu_instance *gpu_ins;
2010 struct amdgpu_device *adev;
2013 mutex_lock(&mgpu_info.mutex);
2016 * MGPU fan boost feature should be enabled
2017 * only when there are two or more dGPUs in
2020 if (mgpu_info.num_dgpu < 2)
2023 for (i = 0; i < mgpu_info.num_dgpu; i++) {
2024 gpu_ins = &(mgpu_info.gpu_ins[i]);
2025 adev = gpu_ins->adev;
2026 if (!(adev->flags & AMD_IS_APU) &&
2027 !gpu_ins->mgpu_fan_enabled &&
2028 adev->powerplay.pp_funcs &&
2029 adev->powerplay.pp_funcs->enable_mgpu_fan_boost) {
2030 ret = amdgpu_dpm_enable_mgpu_fan_boost(adev);
2034 gpu_ins->mgpu_fan_enabled = 1;
2039 mutex_unlock(&mgpu_info.mutex);
2045 * amdgpu_device_ip_late_init - run late init for hardware IPs
2047 * @adev: amdgpu_device pointer
2049 * Late initialization pass for hardware IPs. The list of all the hardware
2050 * IPs that make up the asic is walked and the late_init callbacks are run.
2051 * late_init covers any special initialization that an IP requires
2052 * after all of the have been initialized or something that needs to happen
2053 * late in the init process.
2054 * Returns 0 on success, negative error code on failure.
2056 static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
2058 struct amdgpu_gpu_instance *gpu_instance;
2061 for (i = 0; i < adev->num_ip_blocks; i++) {
2062 if (!adev->ip_blocks[i].status.hw)
2064 if (adev->ip_blocks[i].version->funcs->late_init) {
2065 r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
2067 DRM_ERROR("late_init of IP block <%s> failed %d\n",
2068 adev->ip_blocks[i].version->funcs->name, r);
2072 adev->ip_blocks[i].status.late_initialized = true;
2075 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE);
2076 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE);
2078 amdgpu_device_fill_reset_magic(adev);
2080 r = amdgpu_device_enable_mgpu_fan_boost();
2082 DRM_ERROR("enable mgpu fan boost failed (%d).\n", r);
2085 if (adev->gmc.xgmi.num_physical_nodes > 1) {
2086 mutex_lock(&mgpu_info.mutex);
2089 * Reset device p-state to low as this was booted with high.
2091 * This should be performed only after all devices from the same
2092 * hive get initialized.
2094 * However, it's unknown how many device in the hive in advance.
2095 * As this is counted one by one during devices initializations.
2097 * So, we wait for all XGMI interlinked devices initialized.
2098 * This may bring some delays as those devices may come from
2099 * different hives. But that should be OK.
2101 if (mgpu_info.num_dgpu == adev->gmc.xgmi.num_physical_nodes) {
2102 for (i = 0; i < mgpu_info.num_gpu; i++) {
2103 gpu_instance = &(mgpu_info.gpu_ins[i]);
2104 if (gpu_instance->adev->flags & AMD_IS_APU)
2107 r = amdgpu_xgmi_set_pstate(gpu_instance->adev, 0);
2109 DRM_ERROR("pstate setting failed (%d).\n", r);
2115 mutex_unlock(&mgpu_info.mutex);
2122 * amdgpu_device_ip_fini - run fini for hardware IPs
2124 * @adev: amdgpu_device pointer
2126 * Main teardown pass for hardware IPs. The list of all the hardware
2127 * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks
2128 * are run. hw_fini tears down the hardware associated with each IP
2129 * and sw_fini tears down any software state associated with each IP.
2130 * Returns 0 on success, negative error code on failure.
2132 static int amdgpu_device_ip_fini(struct amdgpu_device *adev)
2136 amdgpu_ras_pre_fini(adev);
2138 if (adev->gmc.xgmi.num_physical_nodes > 1)
2139 amdgpu_xgmi_remove_device(adev);
2141 amdgpu_amdkfd_device_fini(adev);
2143 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
2144 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
2146 /* need to disable SMC first */
2147 for (i = 0; i < adev->num_ip_blocks; i++) {
2148 if (!adev->ip_blocks[i].status.hw)
2150 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
2151 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
2152 /* XXX handle errors */
2154 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
2155 adev->ip_blocks[i].version->funcs->name, r);
2157 adev->ip_blocks[i].status.hw = false;
2162 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2163 if (!adev->ip_blocks[i].status.hw)
2166 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
2167 /* XXX handle errors */
2169 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
2170 adev->ip_blocks[i].version->funcs->name, r);
2173 adev->ip_blocks[i].status.hw = false;
2177 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2178 if (!adev->ip_blocks[i].status.sw)
2181 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
2182 amdgpu_ucode_free_bo(adev);
2183 amdgpu_free_static_csa(&adev->virt.csa_obj);
2184 amdgpu_device_wb_fini(adev);
2185 amdgpu_device_vram_scratch_fini(adev);
2186 amdgpu_ib_pool_fini(adev);
2189 r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
2190 /* XXX handle errors */
2192 DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
2193 adev->ip_blocks[i].version->funcs->name, r);
2195 adev->ip_blocks[i].status.sw = false;
2196 adev->ip_blocks[i].status.valid = false;
2199 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2200 if (!adev->ip_blocks[i].status.late_initialized)
2202 if (adev->ip_blocks[i].version->funcs->late_fini)
2203 adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
2204 adev->ip_blocks[i].status.late_initialized = false;
2207 amdgpu_ras_fini(adev);
2209 if (amdgpu_sriov_vf(adev))
2210 if (amdgpu_virt_release_full_gpu(adev, false))
2211 DRM_ERROR("failed to release exclusive mode on fini\n");
2217 * amdgpu_device_delayed_init_work_handler - work handler for IB tests
2219 * @work: work_struct.
2221 static void amdgpu_device_delayed_init_work_handler(struct work_struct *work)
2223 struct amdgpu_device *adev =
2224 container_of(work, struct amdgpu_device, delayed_init_work.work);
2227 r = amdgpu_ib_ring_tests(adev);
2229 DRM_ERROR("ib ring test failed (%d).\n", r);
2232 static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work)
2234 struct amdgpu_device *adev =
2235 container_of(work, struct amdgpu_device, gfx.gfx_off_delay_work.work);
2237 mutex_lock(&adev->gfx.gfx_off_mutex);
2238 if (!adev->gfx.gfx_off_state && !adev->gfx.gfx_off_req_count) {
2239 if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, true))
2240 adev->gfx.gfx_off_state = true;
2242 mutex_unlock(&adev->gfx.gfx_off_mutex);
2246 * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1)
2248 * @adev: amdgpu_device pointer
2250 * Main suspend function for hardware IPs. The list of all the hardware
2251 * IPs that make up the asic is walked, clockgating is disabled and the
2252 * suspend callbacks are run. suspend puts the hardware and software state
2253 * in each IP into a state suitable for suspend.
2254 * Returns 0 on success, negative error code on failure.
2256 static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev)
2260 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
2261 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
2263 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2264 if (!adev->ip_blocks[i].status.valid)
2266 /* displays are handled separately */
2267 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) {
2268 /* XXX handle errors */
2269 r = adev->ip_blocks[i].version->funcs->suspend(adev);
2270 /* XXX handle errors */
2272 DRM_ERROR("suspend of IP block <%s> failed %d\n",
2273 adev->ip_blocks[i].version->funcs->name, r);
2276 adev->ip_blocks[i].status.hw = false;
2284 * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2)
2286 * @adev: amdgpu_device pointer
2288 * Main suspend function for hardware IPs. The list of all the hardware
2289 * IPs that make up the asic is walked, clockgating is disabled and the
2290 * suspend callbacks are run. suspend puts the hardware and software state
2291 * in each IP into a state suitable for suspend.
2292 * Returns 0 on success, negative error code on failure.
2294 static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev)
2298 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2299 if (!adev->ip_blocks[i].status.valid)
2301 /* displays are handled in phase1 */
2302 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE)
2304 /* PSP lost connection when err_event_athub occurs */
2305 if (amdgpu_ras_intr_triggered() &&
2306 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
2307 adev->ip_blocks[i].status.hw = false;
2310 /* XXX handle errors */
2311 r = adev->ip_blocks[i].version->funcs->suspend(adev);
2312 /* XXX handle errors */
2314 DRM_ERROR("suspend of IP block <%s> failed %d\n",
2315 adev->ip_blocks[i].version->funcs->name, r);
2317 adev->ip_blocks[i].status.hw = false;
2318 /* handle putting the SMC in the appropriate state */
2319 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
2320 if (is_support_sw_smu(adev)) {
2321 r = smu_set_mp1_state(&adev->smu, adev->mp1_state);
2322 } else if (adev->powerplay.pp_funcs &&
2323 adev->powerplay.pp_funcs->set_mp1_state) {
2324 r = adev->powerplay.pp_funcs->set_mp1_state(
2325 adev->powerplay.pp_handle,
2329 DRM_ERROR("SMC failed to set mp1 state %d, %d\n",
2330 adev->mp1_state, r);
2335 adev->ip_blocks[i].status.hw = false;
2342 * amdgpu_device_ip_suspend - run suspend for hardware IPs
2344 * @adev: amdgpu_device pointer
2346 * Main suspend function for hardware IPs. The list of all the hardware
2347 * IPs that make up the asic is walked, clockgating is disabled and the
2348 * suspend callbacks are run. suspend puts the hardware and software state
2349 * in each IP into a state suitable for suspend.
2350 * Returns 0 on success, negative error code on failure.
2352 int amdgpu_device_ip_suspend(struct amdgpu_device *adev)
2356 if (amdgpu_sriov_vf(adev))
2357 amdgpu_virt_request_full_gpu(adev, false);
2359 r = amdgpu_device_ip_suspend_phase1(adev);
2362 r = amdgpu_device_ip_suspend_phase2(adev);
2364 if (amdgpu_sriov_vf(adev))
2365 amdgpu_virt_release_full_gpu(adev, false);
2370 static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
2374 static enum amd_ip_block_type ip_order[] = {
2375 AMD_IP_BLOCK_TYPE_GMC,
2376 AMD_IP_BLOCK_TYPE_COMMON,
2377 AMD_IP_BLOCK_TYPE_PSP,
2378 AMD_IP_BLOCK_TYPE_IH,
2381 for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
2383 struct amdgpu_ip_block *block;
2385 for (j = 0; j < adev->num_ip_blocks; j++) {
2386 block = &adev->ip_blocks[j];
2388 block->status.hw = false;
2389 if (block->version->type != ip_order[i] ||
2390 !block->status.valid)
2393 r = block->version->funcs->hw_init(adev);
2394 DRM_INFO("RE-INIT-early: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
2397 block->status.hw = true;
2404 static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
2408 static enum amd_ip_block_type ip_order[] = {
2409 AMD_IP_BLOCK_TYPE_SMC,
2410 AMD_IP_BLOCK_TYPE_DCE,
2411 AMD_IP_BLOCK_TYPE_GFX,
2412 AMD_IP_BLOCK_TYPE_SDMA,
2413 AMD_IP_BLOCK_TYPE_UVD,
2414 AMD_IP_BLOCK_TYPE_VCE
2417 for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
2419 struct amdgpu_ip_block *block;
2421 for (j = 0; j < adev->num_ip_blocks; j++) {
2422 block = &adev->ip_blocks[j];
2424 if (block->version->type != ip_order[i] ||
2425 !block->status.valid ||
2429 r = block->version->funcs->hw_init(adev);
2430 DRM_INFO("RE-INIT-late: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
2433 block->status.hw = true;
2441 * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs
2443 * @adev: amdgpu_device pointer
2445 * First resume function for hardware IPs. The list of all the hardware
2446 * IPs that make up the asic is walked and the resume callbacks are run for
2447 * COMMON, GMC, and IH. resume puts the hardware into a functional state
2448 * after a suspend and updates the software state as necessary. This
2449 * function is also used for restoring the GPU after a GPU reset.
2450 * Returns 0 on success, negative error code on failure.
2452 static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)
2456 for (i = 0; i < adev->num_ip_blocks; i++) {
2457 if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
2459 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2460 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2461 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
2463 r = adev->ip_blocks[i].version->funcs->resume(adev);
2465 DRM_ERROR("resume of IP block <%s> failed %d\n",
2466 adev->ip_blocks[i].version->funcs->name, r);
2469 adev->ip_blocks[i].status.hw = true;
2477 * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs
2479 * @adev: amdgpu_device pointer
2481 * First resume function for hardware IPs. The list of all the hardware
2482 * IPs that make up the asic is walked and the resume callbacks are run for
2483 * all blocks except COMMON, GMC, and IH. resume puts the hardware into a
2484 * functional state after a suspend and updates the software state as
2485 * necessary. This function is also used for restoring the GPU after a GPU
2487 * Returns 0 on success, negative error code on failure.
2489 static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)
2493 for (i = 0; i < adev->num_ip_blocks; i++) {
2494 if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
2496 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2497 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2498 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH ||
2499 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)
2501 r = adev->ip_blocks[i].version->funcs->resume(adev);
2503 DRM_ERROR("resume of IP block <%s> failed %d\n",
2504 adev->ip_blocks[i].version->funcs->name, r);
2507 adev->ip_blocks[i].status.hw = true;
2514 * amdgpu_device_ip_resume - run resume for hardware IPs
2516 * @adev: amdgpu_device pointer
2518 * Main resume function for hardware IPs. The hardware IPs
2519 * are split into two resume functions because they are
2520 * are also used in in recovering from a GPU reset and some additional
2521 * steps need to be take between them. In this case (S3/S4) they are
2523 * Returns 0 on success, negative error code on failure.
2525 static int amdgpu_device_ip_resume(struct amdgpu_device *adev)
2529 r = amdgpu_device_ip_resume_phase1(adev);
2533 r = amdgpu_device_fw_loading(adev);
2537 r = amdgpu_device_ip_resume_phase2(adev);
2543 * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV
2545 * @adev: amdgpu_device pointer
2547 * Query the VBIOS data tables to determine if the board supports SR-IOV.
2549 static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
2551 if (amdgpu_sriov_vf(adev)) {
2552 if (adev->is_atom_fw) {
2553 if (amdgpu_atomfirmware_gpu_supports_virtualization(adev))
2554 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2556 if (amdgpu_atombios_has_gpu_virtualization_table(adev))
2557 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2560 if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
2561 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
2566 * amdgpu_device_asic_has_dc_support - determine if DC supports the asic
2568 * @asic_type: AMD asic type
2570 * Check if there is DC (new modesetting infrastructre) support for an asic.
2571 * returns true if DC has support, false if not.
2573 bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
2575 switch (asic_type) {
2576 #if defined(CONFIG_DRM_AMD_DC)
2582 * We have systems in the wild with these ASICs that require
2583 * LVDS and VGA support which is not supported with DC.
2585 * Fallback to the non-DC driver here by default so as not to
2586 * cause regressions.
2588 return amdgpu_dc > 0;
2592 case CHIP_POLARIS10:
2593 case CHIP_POLARIS11:
2594 case CHIP_POLARIS12:
2601 #if defined(CONFIG_DRM_AMD_DC_DCN1_0)
2607 #if defined(CONFIG_DRM_AMD_DC_DCN2_1)
2610 return amdgpu_dc != 0;
2618 * amdgpu_device_has_dc_support - check if dc is supported
2620 * @adev: amdgpu_device_pointer
2622 * Returns true for supported, false for not supported
2624 bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
2626 if (amdgpu_sriov_vf(adev))
2629 return amdgpu_device_asic_has_dc_support(adev->asic_type);
2633 static void amdgpu_device_xgmi_reset_func(struct work_struct *__work)
2635 struct amdgpu_device *adev =
2636 container_of(__work, struct amdgpu_device, xgmi_reset_work);
2638 adev->asic_reset_res = amdgpu_asic_reset(adev);
2639 if (adev->asic_reset_res)
2640 DRM_WARN("ASIC reset failed with error, %d for drm dev, %s",
2641 adev->asic_reset_res, adev->ddev->unique);
2644 static int amdgpu_device_get_job_timeout_settings(struct amdgpu_device *adev)
2646 char *input = amdgpu_lockup_timeout;
2647 char *timeout_setting = NULL;
2653 * By default timeout for non compute jobs is 10000.
2654 * And there is no timeout enforced on compute jobs.
2655 * In SR-IOV or passthrough mode, timeout for compute
2656 * jobs are 10000 by default.
2658 adev->gfx_timeout = msecs_to_jiffies(10000);
2659 adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout;
2660 if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev))
2661 adev->compute_timeout = adev->gfx_timeout;
2663 adev->compute_timeout = MAX_SCHEDULE_TIMEOUT;
2665 if (strnlen(input, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) {
2666 while ((timeout_setting = strsep(&input, ",")) &&
2667 strnlen(timeout_setting, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) {
2668 ret = kstrtol(timeout_setting, 0, &timeout);
2675 } else if (timeout < 0) {
2676 timeout = MAX_SCHEDULE_TIMEOUT;
2678 timeout = msecs_to_jiffies(timeout);
2683 adev->gfx_timeout = timeout;
2686 adev->compute_timeout = timeout;
2689 adev->sdma_timeout = timeout;
2692 adev->video_timeout = timeout;
2699 * There is only one value specified and
2700 * it should apply to all non-compute jobs.
2703 adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout;
2704 if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev))
2705 adev->compute_timeout = adev->gfx_timeout;
2713 * amdgpu_device_init - initialize the driver
2715 * @adev: amdgpu_device pointer
2716 * @ddev: drm dev pointer
2717 * @pdev: pci dev pointer
2718 * @flags: driver flags
2720 * Initializes the driver info and hw (all asics).
2721 * Returns 0 for success or an error on failure.
2722 * Called at driver startup.
2724 int amdgpu_device_init(struct amdgpu_device *adev,
2725 struct drm_device *ddev,
2726 struct pci_dev *pdev,
2730 bool runtime = false;
2733 adev->shutdown = false;
2734 adev->dev = &pdev->dev;
2737 adev->flags = flags;
2739 if (amdgpu_force_asic_type >= 0 && amdgpu_force_asic_type < CHIP_LAST)
2740 adev->asic_type = amdgpu_force_asic_type;
2742 adev->asic_type = flags & AMD_ASIC_MASK;
2744 adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
2745 if (amdgpu_emu_mode == 1)
2746 adev->usec_timeout *= 2;
2747 adev->gmc.gart_size = 512 * 1024 * 1024;
2748 adev->accel_working = false;
2749 adev->num_rings = 0;
2750 adev->mman.buffer_funcs = NULL;
2751 adev->mman.buffer_funcs_ring = NULL;
2752 adev->vm_manager.vm_pte_funcs = NULL;
2753 adev->vm_manager.vm_pte_num_rqs = 0;
2754 adev->gmc.gmc_funcs = NULL;
2755 adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
2756 bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
2758 adev->smc_rreg = &amdgpu_invalid_rreg;
2759 adev->smc_wreg = &amdgpu_invalid_wreg;
2760 adev->pcie_rreg = &amdgpu_invalid_rreg;
2761 adev->pcie_wreg = &amdgpu_invalid_wreg;
2762 adev->pciep_rreg = &amdgpu_invalid_rreg;
2763 adev->pciep_wreg = &amdgpu_invalid_wreg;
2764 adev->pcie_rreg64 = &amdgpu_invalid_rreg64;
2765 adev->pcie_wreg64 = &amdgpu_invalid_wreg64;
2766 adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
2767 adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
2768 adev->didt_rreg = &amdgpu_invalid_rreg;
2769 adev->didt_wreg = &amdgpu_invalid_wreg;
2770 adev->gc_cac_rreg = &amdgpu_invalid_rreg;
2771 adev->gc_cac_wreg = &amdgpu_invalid_wreg;
2772 adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
2773 adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
2775 DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
2776 amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
2777 pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
2779 /* mutex initialization are all done here so we
2780 * can recall function without having locking issues */
2781 atomic_set(&adev->irq.ih.lock, 0);
2782 mutex_init(&adev->firmware.mutex);
2783 mutex_init(&adev->pm.mutex);
2784 mutex_init(&adev->gfx.gpu_clock_mutex);
2785 mutex_init(&adev->srbm_mutex);
2786 mutex_init(&adev->gfx.pipe_reserve_mutex);
2787 mutex_init(&adev->gfx.gfx_off_mutex);
2788 mutex_init(&adev->grbm_idx_mutex);
2789 mutex_init(&adev->mn_lock);
2790 mutex_init(&adev->virt.vf_errors.lock);
2791 hash_init(adev->mn_hash);
2792 mutex_init(&adev->lock_reset);
2793 mutex_init(&adev->virt.dpm_mutex);
2794 mutex_init(&adev->psp.mutex);
2796 r = amdgpu_device_check_arguments(adev);
2800 spin_lock_init(&adev->mmio_idx_lock);
2801 spin_lock_init(&adev->smc_idx_lock);
2802 spin_lock_init(&adev->pcie_idx_lock);
2803 spin_lock_init(&adev->uvd_ctx_idx_lock);
2804 spin_lock_init(&adev->didt_idx_lock);
2805 spin_lock_init(&adev->gc_cac_idx_lock);
2806 spin_lock_init(&adev->se_cac_idx_lock);
2807 spin_lock_init(&adev->audio_endpt_idx_lock);
2808 spin_lock_init(&adev->mm_stats.lock);
2810 INIT_LIST_HEAD(&adev->shadow_list);
2811 mutex_init(&adev->shadow_list_lock);
2813 INIT_LIST_HEAD(&adev->ring_lru_list);
2814 spin_lock_init(&adev->ring_lru_list_lock);
2816 INIT_DELAYED_WORK(&adev->delayed_init_work,
2817 amdgpu_device_delayed_init_work_handler);
2818 INIT_DELAYED_WORK(&adev->gfx.gfx_off_delay_work,
2819 amdgpu_device_delay_enable_gfx_off);
2821 INIT_WORK(&adev->xgmi_reset_work, amdgpu_device_xgmi_reset_func);
2823 adev->gfx.gfx_off_req_count = 1;
2824 adev->pm.ac_power = power_supply_is_system_supplied() > 0 ? true : false;
2826 /* Registers mapping */
2827 /* TODO: block userspace mapping of io register */
2828 if (adev->asic_type >= CHIP_BONAIRE) {
2829 adev->rmmio_base = pci_resource_start(adev->pdev, 5);
2830 adev->rmmio_size = pci_resource_len(adev->pdev, 5);
2832 adev->rmmio_base = pci_resource_start(adev->pdev, 2);
2833 adev->rmmio_size = pci_resource_len(adev->pdev, 2);
2836 adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
2837 if (adev->rmmio == NULL) {
2840 DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
2841 DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
2843 /* io port mapping */
2844 for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
2845 if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) {
2846 adev->rio_mem_size = pci_resource_len(adev->pdev, i);
2847 adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size);
2851 if (adev->rio_mem == NULL)
2852 DRM_INFO("PCI I/O BAR is not found.\n");
2854 /* enable PCIE atomic ops */
2855 r = pci_enable_atomic_ops_to_root(adev->pdev,
2856 PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
2857 PCI_EXP_DEVCAP2_ATOMIC_COMP64);
2859 adev->have_atomics_support = false;
2860 DRM_INFO("PCIE atomic ops is not supported\n");
2862 adev->have_atomics_support = true;
2865 amdgpu_device_get_pcie_info(adev);
2868 DRM_INFO("MCBP is enabled\n");
2870 if (amdgpu_mes && adev->asic_type >= CHIP_NAVI10)
2871 adev->enable_mes = true;
2873 if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10) {
2874 r = amdgpu_discovery_init(adev);
2876 dev_err(adev->dev, "amdgpu_discovery_init failed\n");
2881 /* early init functions */
2882 r = amdgpu_device_ip_early_init(adev);
2886 r = amdgpu_device_get_job_timeout_settings(adev);
2888 dev_err(adev->dev, "invalid lockup_timeout parameter syntax\n");
2892 /* doorbell bar mapping and doorbell index init*/
2893 amdgpu_device_doorbell_init(adev);
2895 /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
2896 /* this will fail for cards that aren't VGA class devices, just
2898 vga_client_register(adev->pdev, adev, NULL, amdgpu_device_vga_set_decode);
2900 if (amdgpu_device_is_px(ddev))
2902 if (!pci_is_thunderbolt_attached(adev->pdev))
2903 vga_switcheroo_register_client(adev->pdev,
2904 &amdgpu_switcheroo_ops, runtime);
2906 vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
2908 if (amdgpu_emu_mode == 1) {
2909 /* post the asic on emulation mode */
2910 emu_soc_asic_init(adev);
2911 goto fence_driver_init;
2914 /* detect if we are with an SRIOV vbios */
2915 amdgpu_device_detect_sriov_bios(adev);
2917 /* check if we need to reset the asic
2918 * E.g., driver was not cleanly unloaded previously, etc.
2920 if (!amdgpu_sriov_vf(adev) && amdgpu_asic_need_reset_on_init(adev)) {
2921 r = amdgpu_asic_reset(adev);
2923 dev_err(adev->dev, "asic reset on init failed\n");
2928 /* Post card if necessary */
2929 if (amdgpu_device_need_post(adev)) {
2931 dev_err(adev->dev, "no vBIOS found\n");
2935 DRM_INFO("GPU posting now...\n");
2936 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2938 dev_err(adev->dev, "gpu post error!\n");
2943 if (adev->is_atom_fw) {
2944 /* Initialize clocks */
2945 r = amdgpu_atomfirmware_get_clock_info(adev);
2947 dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
2948 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
2952 /* Initialize clocks */
2953 r = amdgpu_atombios_get_clock_info(adev);
2955 dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
2956 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
2959 /* init i2c buses */
2960 if (!amdgpu_device_has_dc_support(adev))
2961 amdgpu_atombios_i2c_init(adev);
2966 r = amdgpu_fence_driver_init(adev);
2968 dev_err(adev->dev, "amdgpu_fence_driver_init failed\n");
2969 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
2973 /* init the mode config */
2974 drm_mode_config_init(adev->ddev);
2976 r = amdgpu_device_ip_init(adev);
2978 /* failed in exclusive mode due to timeout */
2979 if (amdgpu_sriov_vf(adev) &&
2980 !amdgpu_sriov_runtime(adev) &&
2981 amdgpu_virt_mmio_blocked(adev) &&
2982 !amdgpu_virt_wait_reset(adev)) {
2983 dev_err(adev->dev, "VF exclusive mode timeout\n");
2984 /* Don't send request since VF is inactive. */
2985 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
2986 adev->virt.ops = NULL;
2990 dev_err(adev->dev, "amdgpu_device_ip_init failed\n");
2991 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
2992 if (amdgpu_virt_request_full_gpu(adev, false))
2993 amdgpu_virt_release_full_gpu(adev, false);
2997 adev->accel_working = true;
2999 amdgpu_vm_check_compute_bug(adev);
3001 /* Initialize the buffer migration limit. */
3002 if (amdgpu_moverate >= 0)
3003 max_MBps = amdgpu_moverate;
3005 max_MBps = 8; /* Allow 8 MB/s. */
3006 /* Get a log2 for easy divisions. */
3007 adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
3009 amdgpu_fbdev_init(adev);
3011 if (amdgpu_sriov_vf(adev) && amdgim_is_hwperf(adev))
3012 amdgpu_pm_virt_sysfs_init(adev);
3014 r = amdgpu_pm_sysfs_init(adev);
3016 DRM_ERROR("registering pm debugfs failed (%d).\n", r);
3018 r = amdgpu_ucode_sysfs_init(adev);
3020 DRM_ERROR("Creating firmware sysfs failed (%d).\n", r);
3022 r = amdgpu_debugfs_gem_init(adev);
3024 DRM_ERROR("registering gem debugfs failed (%d).\n", r);
3026 r = amdgpu_debugfs_regs_init(adev);
3028 DRM_ERROR("registering register debugfs failed (%d).\n", r);
3030 r = amdgpu_debugfs_firmware_init(adev);
3032 DRM_ERROR("registering firmware debugfs failed (%d).\n", r);
3034 r = amdgpu_debugfs_init(adev);
3036 DRM_ERROR("Creating debugfs files failed (%d).\n", r);
3038 if ((amdgpu_testing & 1)) {
3039 if (adev->accel_working)
3040 amdgpu_test_moves(adev);
3042 DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n");
3044 if (amdgpu_benchmarking) {
3045 if (adev->accel_working)
3046 amdgpu_benchmark(adev, amdgpu_benchmarking);
3048 DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n");
3052 * Register gpu instance before amdgpu_device_enable_mgpu_fan_boost.
3053 * Otherwise the mgpu fan boost feature will be skipped due to the
3054 * gpu instance is counted less.
3056 amdgpu_register_gpu_instance(adev);
3058 /* enable clockgating, etc. after ib tests, etc. since some blocks require
3059 * explicit gating rather than handling it automatically.
3061 r = amdgpu_device_ip_late_init(adev);
3063 dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n");
3064 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
3069 amdgpu_ras_resume(adev);
3071 queue_delayed_work(system_wq, &adev->delayed_init_work,
3072 msecs_to_jiffies(AMDGPU_RESUME_MS));
3074 r = device_create_file(adev->dev, &dev_attr_pcie_replay_count);
3076 dev_err(adev->dev, "Could not create pcie_replay_count");
3080 if (IS_ENABLED(CONFIG_PERF_EVENTS))
3081 r = amdgpu_pmu_init(adev);
3083 dev_err(adev->dev, "amdgpu_pmu_init failed\n");
3088 amdgpu_vf_error_trans_all(adev);
3090 vga_switcheroo_fini_domain_pm_ops(adev->dev);
3096 * amdgpu_device_fini - tear down the driver
3098 * @adev: amdgpu_device pointer
3100 * Tear down the driver info (all asics).
3101 * Called at driver shutdown.
3103 void amdgpu_device_fini(struct amdgpu_device *adev)
3107 DRM_INFO("amdgpu: finishing device.\n");
3108 adev->shutdown = true;
3110 flush_delayed_work(&adev->delayed_init_work);
3112 /* disable all interrupts */
3113 amdgpu_irq_disable_all(adev);
3114 if (adev->mode_info.mode_config_initialized){
3115 if (!amdgpu_device_has_dc_support(adev))
3116 drm_helper_force_disable_all(adev->ddev);
3118 drm_atomic_helper_shutdown(adev->ddev);
3120 amdgpu_fence_driver_fini(adev);
3121 amdgpu_pm_sysfs_fini(adev);
3122 amdgpu_fbdev_fini(adev);
3123 r = amdgpu_device_ip_fini(adev);
3124 if (adev->firmware.gpu_info_fw) {
3125 release_firmware(adev->firmware.gpu_info_fw);
3126 adev->firmware.gpu_info_fw = NULL;
3128 adev->accel_working = false;
3129 cancel_delayed_work_sync(&adev->delayed_init_work);
3130 /* free i2c buses */
3131 if (!amdgpu_device_has_dc_support(adev))
3132 amdgpu_i2c_fini(adev);
3134 if (amdgpu_emu_mode != 1)
3135 amdgpu_atombios_fini(adev);
3139 if (!pci_is_thunderbolt_attached(adev->pdev))
3140 vga_switcheroo_unregister_client(adev->pdev);
3141 if (adev->flags & AMD_IS_PX)
3142 vga_switcheroo_fini_domain_pm_ops(adev->dev);
3143 vga_client_register(adev->pdev, NULL, NULL, NULL);
3145 pci_iounmap(adev->pdev, adev->rio_mem);
3146 adev->rio_mem = NULL;
3147 iounmap(adev->rmmio);
3149 amdgpu_device_doorbell_fini(adev);
3150 if (amdgpu_sriov_vf(adev) && amdgim_is_hwperf(adev))
3151 amdgpu_pm_virt_sysfs_fini(adev);
3153 amdgpu_debugfs_regs_cleanup(adev);
3154 device_remove_file(adev->dev, &dev_attr_pcie_replay_count);
3155 amdgpu_ucode_sysfs_fini(adev);
3156 if (IS_ENABLED(CONFIG_PERF_EVENTS))
3157 amdgpu_pmu_fini(adev);
3158 amdgpu_debugfs_preempt_cleanup(adev);
3159 if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10)
3160 amdgpu_discovery_fini(adev);
3168 * amdgpu_device_suspend - initiate device suspend
3170 * @dev: drm dev pointer
3171 * @suspend: suspend state
3172 * @fbcon : notify the fbdev of suspend
3174 * Puts the hw in the suspend state (all asics).
3175 * Returns 0 for success or an error on failure.
3176 * Called at driver suspend.
3178 int amdgpu_device_suspend(struct drm_device *dev, bool suspend, bool fbcon)
3180 struct amdgpu_device *adev;
3181 struct drm_crtc *crtc;
3182 struct drm_connector *connector;
3183 struct drm_connector_list_iter iter;
3186 if (dev == NULL || dev->dev_private == NULL) {
3190 adev = dev->dev_private;
3192 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
3195 adev->in_suspend = true;
3196 drm_kms_helper_poll_disable(dev);
3199 amdgpu_fbdev_set_suspend(adev, 1);
3201 cancel_delayed_work_sync(&adev->delayed_init_work);
3203 if (!amdgpu_device_has_dc_support(adev)) {
3204 /* turn off display hw */
3205 drm_modeset_lock_all(dev);
3206 drm_connector_list_iter_begin(dev, &iter);
3207 drm_for_each_connector_iter(connector, &iter)
3208 drm_helper_connector_dpms(connector,
3210 drm_connector_list_iter_end(&iter);
3211 drm_modeset_unlock_all(dev);
3212 /* unpin the front buffers and cursors */
3213 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
3214 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
3215 struct drm_framebuffer *fb = crtc->primary->fb;
3216 struct amdgpu_bo *robj;
3218 if (amdgpu_crtc->cursor_bo && !adev->enable_virtual_display) {
3219 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
3220 r = amdgpu_bo_reserve(aobj, true);
3222 amdgpu_bo_unpin(aobj);
3223 amdgpu_bo_unreserve(aobj);
3227 if (fb == NULL || fb->obj[0] == NULL) {
3230 robj = gem_to_amdgpu_bo(fb->obj[0]);
3231 /* don't unpin kernel fb objects */
3232 if (!amdgpu_fbdev_robj_is_fb(adev, robj)) {
3233 r = amdgpu_bo_reserve(robj, true);
3235 amdgpu_bo_unpin(robj);
3236 amdgpu_bo_unreserve(robj);
3242 amdgpu_amdkfd_suspend(adev);
3244 amdgpu_ras_suspend(adev);
3246 r = amdgpu_device_ip_suspend_phase1(adev);
3248 /* evict vram memory */
3249 amdgpu_bo_evict_vram(adev);
3251 amdgpu_fence_driver_suspend(adev);
3253 r = amdgpu_device_ip_suspend_phase2(adev);
3255 /* evict remaining vram memory
3256 * This second call to evict vram is to evict the gart page table
3259 amdgpu_bo_evict_vram(adev);
3262 pci_save_state(dev->pdev);
3263 /* Shut down the device */
3264 pci_disable_device(dev->pdev);
3265 pci_set_power_state(dev->pdev, PCI_D3hot);
3272 * amdgpu_device_resume - initiate device resume
3274 * @dev: drm dev pointer
3275 * @resume: resume state
3276 * @fbcon : notify the fbdev of resume
3278 * Bring the hw back to operating state (all asics).
3279 * Returns 0 for success or an error on failure.
3280 * Called at driver resume.
3282 int amdgpu_device_resume(struct drm_device *dev, bool resume, bool fbcon)
3284 struct drm_connector *connector;
3285 struct drm_connector_list_iter iter;
3286 struct amdgpu_device *adev = dev->dev_private;
3287 struct drm_crtc *crtc;
3290 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
3294 pci_set_power_state(dev->pdev, PCI_D0);
3295 pci_restore_state(dev->pdev);
3296 r = pci_enable_device(dev->pdev);
3302 if (amdgpu_device_need_post(adev)) {
3303 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
3305 DRM_ERROR("amdgpu asic init failed\n");
3308 r = amdgpu_device_ip_resume(adev);
3310 DRM_ERROR("amdgpu_device_ip_resume failed (%d).\n", r);
3313 amdgpu_fence_driver_resume(adev);
3316 r = amdgpu_device_ip_late_init(adev);
3320 queue_delayed_work(system_wq, &adev->delayed_init_work,
3321 msecs_to_jiffies(AMDGPU_RESUME_MS));
3323 if (!amdgpu_device_has_dc_support(adev)) {
3325 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
3326 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
3328 if (amdgpu_crtc->cursor_bo && !adev->enable_virtual_display) {
3329 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
3330 r = amdgpu_bo_reserve(aobj, true);
3332 r = amdgpu_bo_pin(aobj, AMDGPU_GEM_DOMAIN_VRAM);
3334 DRM_ERROR("Failed to pin cursor BO (%d)\n", r);
3335 amdgpu_crtc->cursor_addr = amdgpu_bo_gpu_offset(aobj);
3336 amdgpu_bo_unreserve(aobj);
3341 r = amdgpu_amdkfd_resume(adev);
3345 /* Make sure IB tests flushed */
3346 flush_delayed_work(&adev->delayed_init_work);
3348 /* blat the mode back in */
3350 if (!amdgpu_device_has_dc_support(adev)) {
3352 drm_helper_resume_force_mode(dev);
3354 /* turn on display hw */
3355 drm_modeset_lock_all(dev);
3357 drm_connector_list_iter_begin(dev, &iter);
3358 drm_for_each_connector_iter(connector, &iter)
3359 drm_helper_connector_dpms(connector,
3361 drm_connector_list_iter_end(&iter);
3363 drm_modeset_unlock_all(dev);
3365 amdgpu_fbdev_set_suspend(adev, 0);
3368 drm_kms_helper_poll_enable(dev);
3370 amdgpu_ras_resume(adev);
3373 * Most of the connector probing functions try to acquire runtime pm
3374 * refs to ensure that the GPU is powered on when connector polling is
3375 * performed. Since we're calling this from a runtime PM callback,
3376 * trying to acquire rpm refs will cause us to deadlock.
3378 * Since we're guaranteed to be holding the rpm lock, it's safe to
3379 * temporarily disable the rpm helpers so this doesn't deadlock us.
3382 dev->dev->power.disable_depth++;
3384 if (!amdgpu_device_has_dc_support(adev))
3385 drm_helper_hpd_irq_event(dev);
3387 drm_kms_helper_hotplug_event(dev);
3389 dev->dev->power.disable_depth--;
3391 adev->in_suspend = false;
3397 * amdgpu_device_ip_check_soft_reset - did soft reset succeed
3399 * @adev: amdgpu_device pointer
3401 * The list of all the hardware IPs that make up the asic is walked and
3402 * the check_soft_reset callbacks are run. check_soft_reset determines
3403 * if the asic is still hung or not.
3404 * Returns true if any of the IPs are still in a hung state, false if not.
3406 static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)
3409 bool asic_hang = false;
3411 if (amdgpu_sriov_vf(adev))
3414 if (amdgpu_asic_need_full_reset(adev))
3417 for (i = 0; i < adev->num_ip_blocks; i++) {
3418 if (!adev->ip_blocks[i].status.valid)
3420 if (adev->ip_blocks[i].version->funcs->check_soft_reset)
3421 adev->ip_blocks[i].status.hang =
3422 adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
3423 if (adev->ip_blocks[i].status.hang) {
3424 DRM_INFO("IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
3432 * amdgpu_device_ip_pre_soft_reset - prepare for soft reset
3434 * @adev: amdgpu_device pointer
3436 * The list of all the hardware IPs that make up the asic is walked and the
3437 * pre_soft_reset callbacks are run if the block is hung. pre_soft_reset
3438 * handles any IP specific hardware or software state changes that are
3439 * necessary for a soft reset to succeed.
3440 * Returns 0 on success, negative error code on failure.
3442 static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)
3446 for (i = 0; i < adev->num_ip_blocks; i++) {
3447 if (!adev->ip_blocks[i].status.valid)
3449 if (adev->ip_blocks[i].status.hang &&
3450 adev->ip_blocks[i].version->funcs->pre_soft_reset) {
3451 r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
3461 * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed
3463 * @adev: amdgpu_device pointer
3465 * Some hardware IPs cannot be soft reset. If they are hung, a full gpu
3466 * reset is necessary to recover.
3467 * Returns true if a full asic reset is required, false if not.
3469 static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)
3473 if (amdgpu_asic_need_full_reset(adev))
3476 for (i = 0; i < adev->num_ip_blocks; i++) {
3477 if (!adev->ip_blocks[i].status.valid)
3479 if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
3480 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
3481 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
3482 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
3483 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
3484 if (adev->ip_blocks[i].status.hang) {
3485 DRM_INFO("Some block need full reset!\n");
3494 * amdgpu_device_ip_soft_reset - do a soft reset
3496 * @adev: amdgpu_device pointer
3498 * The list of all the hardware IPs that make up the asic is walked and the
3499 * soft_reset callbacks are run if the block is hung. soft_reset handles any
3500 * IP specific hardware or software state changes that are necessary to soft
3502 * Returns 0 on success, negative error code on failure.
3504 static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)
3508 for (i = 0; i < adev->num_ip_blocks; i++) {
3509 if (!adev->ip_blocks[i].status.valid)
3511 if (adev->ip_blocks[i].status.hang &&
3512 adev->ip_blocks[i].version->funcs->soft_reset) {
3513 r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
3523 * amdgpu_device_ip_post_soft_reset - clean up from soft reset
3525 * @adev: amdgpu_device pointer
3527 * The list of all the hardware IPs that make up the asic is walked and the
3528 * post_soft_reset callbacks are run if the asic was hung. post_soft_reset
3529 * handles any IP specific hardware or software state changes that are
3530 * necessary after the IP has been soft reset.
3531 * Returns 0 on success, negative error code on failure.
3533 static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)
3537 for (i = 0; i < adev->num_ip_blocks; i++) {
3538 if (!adev->ip_blocks[i].status.valid)
3540 if (adev->ip_blocks[i].status.hang &&
3541 adev->ip_blocks[i].version->funcs->post_soft_reset)
3542 r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
3551 * amdgpu_device_recover_vram - Recover some VRAM contents
3553 * @adev: amdgpu_device pointer
3555 * Restores the contents of VRAM buffers from the shadows in GTT. Used to
3556 * restore things like GPUVM page tables after a GPU reset where
3557 * the contents of VRAM might be lost.
3560 * 0 on success, negative error code on failure.
3562 static int amdgpu_device_recover_vram(struct amdgpu_device *adev)
3564 struct dma_fence *fence = NULL, *next = NULL;
3565 struct amdgpu_bo *shadow;
3568 if (amdgpu_sriov_runtime(adev))
3569 tmo = msecs_to_jiffies(8000);
3571 tmo = msecs_to_jiffies(100);
3573 DRM_INFO("recover vram bo from shadow start\n");
3574 mutex_lock(&adev->shadow_list_lock);
3575 list_for_each_entry(shadow, &adev->shadow_list, shadow_list) {
3577 /* No need to recover an evicted BO */
3578 if (shadow->tbo.mem.mem_type != TTM_PL_TT ||
3579 shadow->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET ||
3580 shadow->parent->tbo.mem.mem_type != TTM_PL_VRAM)
3583 r = amdgpu_bo_restore_shadow(shadow, &next);
3588 tmo = dma_fence_wait_timeout(fence, false, tmo);
3589 dma_fence_put(fence);
3594 } else if (tmo < 0) {
3602 mutex_unlock(&adev->shadow_list_lock);
3605 tmo = dma_fence_wait_timeout(fence, false, tmo);
3606 dma_fence_put(fence);
3608 if (r < 0 || tmo <= 0) {
3609 DRM_ERROR("recover vram bo from shadow failed, r is %ld, tmo is %ld\n", r, tmo);
3613 DRM_INFO("recover vram bo from shadow done\n");
3619 * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
3621 * @adev: amdgpu device pointer
3622 * @from_hypervisor: request from hypervisor
3624 * do VF FLR and reinitialize Asic
3625 * return 0 means succeeded otherwise failed
3627 static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
3628 bool from_hypervisor)
3632 if (from_hypervisor)
3633 r = amdgpu_virt_request_full_gpu(adev, true);
3635 r = amdgpu_virt_reset_gpu(adev);
3639 amdgpu_amdkfd_pre_reset(adev);
3641 /* Resume IP prior to SMC */
3642 r = amdgpu_device_ip_reinit_early_sriov(adev);
3646 /* we need recover gart prior to run SMC/CP/SDMA resume */
3647 amdgpu_gtt_mgr_recover(&adev->mman.bdev.man[TTM_PL_TT]);
3649 r = amdgpu_device_fw_loading(adev);
3653 /* now we are okay to resume SMC/CP/SDMA */
3654 r = amdgpu_device_ip_reinit_late_sriov(adev);
3658 amdgpu_irq_gpu_reset_resume_helper(adev);
3659 r = amdgpu_ib_ring_tests(adev);
3660 amdgpu_amdkfd_post_reset(adev);
3663 amdgpu_virt_init_data_exchange(adev);
3664 amdgpu_virt_release_full_gpu(adev, true);
3665 if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
3666 amdgpu_inc_vram_lost(adev);
3667 r = amdgpu_device_recover_vram(adev);
3674 * amdgpu_device_should_recover_gpu - check if we should try GPU recovery
3676 * @adev: amdgpu device pointer
3678 * Check amdgpu_gpu_recovery and SRIOV status to see if we should try to recover
3681 bool amdgpu_device_should_recover_gpu(struct amdgpu_device *adev)
3683 if (!amdgpu_device_ip_check_soft_reset(adev)) {
3684 DRM_INFO("Timeout, but no hardware hang detected.\n");
3688 if (amdgpu_gpu_recovery == 0)
3691 if (amdgpu_sriov_vf(adev))
3694 if (amdgpu_gpu_recovery == -1) {
3695 switch (adev->asic_type) {
3701 case CHIP_POLARIS10:
3702 case CHIP_POLARIS11:
3703 case CHIP_POLARIS12:
3718 DRM_INFO("GPU recovery disabled.\n");
3723 static int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,
3724 struct amdgpu_job *job,
3725 bool *need_full_reset_arg)
3728 bool need_full_reset = *need_full_reset_arg;
3730 /* block all schedulers and reset given job's ring */
3731 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3732 struct amdgpu_ring *ring = adev->rings[i];
3734 if (!ring || !ring->sched.thread)
3737 /* after all hw jobs are reset, hw fence is meaningless, so force_completion */
3738 amdgpu_fence_driver_force_completion(ring);
3742 drm_sched_increase_karma(&job->base);
3744 /* Don't suspend on bare metal if we are not going to HW reset the ASIC */
3745 if (!amdgpu_sriov_vf(adev)) {
3747 if (!need_full_reset)
3748 need_full_reset = amdgpu_device_ip_need_full_reset(adev);
3750 if (!need_full_reset) {
3751 amdgpu_device_ip_pre_soft_reset(adev);
3752 r = amdgpu_device_ip_soft_reset(adev);
3753 amdgpu_device_ip_post_soft_reset(adev);
3754 if (r || amdgpu_device_ip_check_soft_reset(adev)) {
3755 DRM_INFO("soft reset failed, will fallback to full reset!\n");
3756 need_full_reset = true;
3760 if (need_full_reset)
3761 r = amdgpu_device_ip_suspend(adev);
3763 *need_full_reset_arg = need_full_reset;
3769 static int amdgpu_do_asic_reset(struct amdgpu_hive_info *hive,
3770 struct list_head *device_list_handle,
3771 bool *need_full_reset_arg)
3773 struct amdgpu_device *tmp_adev = NULL;
3774 bool need_full_reset = *need_full_reset_arg, vram_lost = false;
3778 * ASIC reset has to be done on all HGMI hive nodes ASAP
3779 * to allow proper links negotiation in FW (within 1 sec)
3781 if (need_full_reset) {
3782 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
3783 /* For XGMI run all resets in parallel to speed up the process */
3784 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
3785 if (!queue_work(system_highpri_wq, &tmp_adev->xgmi_reset_work))
3788 r = amdgpu_asic_reset(tmp_adev);
3791 DRM_ERROR("ASIC reset failed with error, %d for drm dev, %s",
3792 r, tmp_adev->ddev->unique);
3797 /* For XGMI wait for all PSP resets to complete before proceed */
3799 list_for_each_entry(tmp_adev, device_list_handle,
3801 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
3802 flush_work(&tmp_adev->xgmi_reset_work);
3803 r = tmp_adev->asic_reset_res;
3812 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
3813 if (need_full_reset) {
3815 if (amdgpu_atom_asic_init(tmp_adev->mode_info.atom_context))
3816 DRM_WARN("asic atom init failed!");
3819 dev_info(tmp_adev->dev, "GPU reset succeeded, trying to resume\n");
3820 r = amdgpu_device_ip_resume_phase1(tmp_adev);
3824 vram_lost = amdgpu_device_check_vram_lost(tmp_adev);
3826 DRM_INFO("VRAM is lost due to GPU reset!\n");
3827 amdgpu_inc_vram_lost(tmp_adev);
3830 r = amdgpu_gtt_mgr_recover(
3831 &tmp_adev->mman.bdev.man[TTM_PL_TT]);
3835 r = amdgpu_device_fw_loading(tmp_adev);
3839 r = amdgpu_device_ip_resume_phase2(tmp_adev);
3844 amdgpu_device_fill_reset_magic(tmp_adev);
3847 * Add this ASIC as tracked as reset was already
3848 * complete successfully.
3850 amdgpu_register_gpu_instance(tmp_adev);
3852 r = amdgpu_device_ip_late_init(tmp_adev);
3857 amdgpu_ras_resume(tmp_adev);
3859 /* Update PSP FW topology after reset */
3860 if (hive && tmp_adev->gmc.xgmi.num_physical_nodes > 1)
3861 r = amdgpu_xgmi_update_topology(hive, tmp_adev);
3868 amdgpu_irq_gpu_reset_resume_helper(tmp_adev);
3869 r = amdgpu_ib_ring_tests(tmp_adev);
3871 dev_err(tmp_adev->dev, "ib ring test failed (%d).\n", r);
3872 r = amdgpu_device_ip_suspend(tmp_adev);
3873 need_full_reset = true;
3880 r = amdgpu_device_recover_vram(tmp_adev);
3882 tmp_adev->asic_reset_res = r;
3886 *need_full_reset_arg = need_full_reset;
3890 static bool amdgpu_device_lock_adev(struct amdgpu_device *adev, bool trylock)
3893 if (!mutex_trylock(&adev->lock_reset))
3896 mutex_lock(&adev->lock_reset);
3898 atomic_inc(&adev->gpu_reset_counter);
3899 adev->in_gpu_reset = 1;
3900 switch (amdgpu_asic_reset_method(adev)) {
3901 case AMD_RESET_METHOD_MODE1:
3902 adev->mp1_state = PP_MP1_STATE_SHUTDOWN;
3904 case AMD_RESET_METHOD_MODE2:
3905 adev->mp1_state = PP_MP1_STATE_RESET;
3908 adev->mp1_state = PP_MP1_STATE_NONE;
3915 static void amdgpu_device_unlock_adev(struct amdgpu_device *adev)
3917 amdgpu_vf_error_trans_all(adev);
3918 adev->mp1_state = PP_MP1_STATE_NONE;
3919 adev->in_gpu_reset = 0;
3920 mutex_unlock(&adev->lock_reset);
3924 * amdgpu_device_gpu_recover - reset the asic and recover scheduler
3926 * @adev: amdgpu device pointer
3927 * @job: which job trigger hang
3929 * Attempt to reset the GPU if it has hung (all asics).
3930 * Attempt to do soft-reset or full-reset and reinitialize Asic
3931 * Returns 0 for success or an error on failure.
3934 int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
3935 struct amdgpu_job *job)
3937 struct list_head device_list, *device_list_handle = NULL;
3938 bool need_full_reset, job_signaled;
3939 struct amdgpu_hive_info *hive = NULL;
3940 struct amdgpu_device *tmp_adev = NULL;
3942 bool in_ras_intr = amdgpu_ras_intr_triggered();
3945 * Flush RAM to disk so that after reboot
3946 * the user can read log and see why the system rebooted.
3948 if (in_ras_intr && amdgpu_ras_get_context(adev)->reboot) {
3950 DRM_WARN("Emergency reboot.");
3953 emergency_restart();
3956 need_full_reset = job_signaled = false;
3957 INIT_LIST_HEAD(&device_list);
3959 dev_info(adev->dev, "GPU %s begin!\n", in_ras_intr ? "jobs stop":"reset");
3961 cancel_delayed_work_sync(&adev->delayed_init_work);
3963 hive = amdgpu_get_xgmi_hive(adev, false);
3966 * Here we trylock to avoid chain of resets executing from
3967 * either trigger by jobs on different adevs in XGMI hive or jobs on
3968 * different schedulers for same device while this TO handler is running.
3969 * We always reset all schedulers for device and all devices for XGMI
3970 * hive so that should take care of them too.
3973 if (hive && !mutex_trylock(&hive->reset_lock)) {
3974 DRM_INFO("Bailing on TDR for s_job:%llx, hive: %llx as another already in progress",
3975 job ? job->base.id : -1, hive->hive_id);
3979 /* Start with adev pre asic reset first for soft reset check.*/
3980 if (!amdgpu_device_lock_adev(adev, !hive)) {
3981 DRM_INFO("Bailing on TDR for s_job:%llx, as another already in progress",
3982 job ? job->base.id : -1);
3986 /* Block kfd: SRIOV would do it separately */
3987 if (!amdgpu_sriov_vf(adev))
3988 amdgpu_amdkfd_pre_reset(adev);
3990 /* Build list of devices to reset */
3991 if (adev->gmc.xgmi.num_physical_nodes > 1) {
3993 /*unlock kfd: SRIOV would do it separately */
3994 if (!amdgpu_sriov_vf(adev))
3995 amdgpu_amdkfd_post_reset(adev);
3996 amdgpu_device_unlock_adev(adev);
4001 * In case we are in XGMI hive mode device reset is done for all the
4002 * nodes in the hive to retrain all XGMI links and hence the reset
4003 * sequence is executed in loop on all nodes.
4005 device_list_handle = &hive->device_list;
4007 list_add_tail(&adev->gmc.xgmi.head, &device_list);
4008 device_list_handle = &device_list;
4011 /* block all schedulers and reset given job's ring */
4012 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
4013 if (tmp_adev != adev) {
4014 amdgpu_device_lock_adev(tmp_adev, false);
4015 if (!amdgpu_sriov_vf(tmp_adev))
4016 amdgpu_amdkfd_pre_reset(tmp_adev);
4020 * Mark these ASICs to be reseted as untracked first
4021 * And add them back after reset completed
4023 amdgpu_unregister_gpu_instance(tmp_adev);
4025 /* disable ras on ALL IPs */
4026 if (!in_ras_intr && amdgpu_device_ip_need_full_reset(tmp_adev))
4027 amdgpu_ras_suspend(tmp_adev);
4029 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
4030 struct amdgpu_ring *ring = tmp_adev->rings[i];
4032 if (!ring || !ring->sched.thread)
4035 drm_sched_stop(&ring->sched, job ? &job->base : NULL);
4038 amdgpu_job_stop_all_jobs_on_sched(&ring->sched);
4044 goto skip_sched_resume;
4047 * Must check guilty signal here since after this point all old
4048 * HW fences are force signaled.
4050 * job->base holds a reference to parent fence
4052 if (job && job->base.s_fence->parent &&
4053 dma_fence_is_signaled(job->base.s_fence->parent))
4054 job_signaled = true;
4057 dev_info(adev->dev, "Guilty job already signaled, skipping HW reset");
4062 /* Guilty job will be freed after this*/
4063 r = amdgpu_device_pre_asic_reset(adev, job, &need_full_reset);
4065 /*TODO Should we stop ?*/
4066 DRM_ERROR("GPU pre asic reset failed with err, %d for drm dev, %s ",
4067 r, adev->ddev->unique);
4068 adev->asic_reset_res = r;
4071 retry: /* Rest of adevs pre asic reset from XGMI hive. */
4072 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
4074 if (tmp_adev == adev)
4077 r = amdgpu_device_pre_asic_reset(tmp_adev,
4080 /*TODO Should we stop ?*/
4082 DRM_ERROR("GPU pre asic reset failed with err, %d for drm dev, %s ",
4083 r, tmp_adev->ddev->unique);
4084 tmp_adev->asic_reset_res = r;
4088 /* Actual ASIC resets if needed.*/
4089 /* TODO Implement XGMI hive reset logic for SRIOV */
4090 if (amdgpu_sriov_vf(adev)) {
4091 r = amdgpu_device_reset_sriov(adev, job ? false : true);
4093 adev->asic_reset_res = r;
4095 r = amdgpu_do_asic_reset(hive, device_list_handle, &need_full_reset);
4096 if (r && r == -EAGAIN)
4102 /* Post ASIC reset for all devs .*/
4103 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
4105 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
4106 struct amdgpu_ring *ring = tmp_adev->rings[i];
4108 if (!ring || !ring->sched.thread)
4111 /* No point to resubmit jobs if we didn't HW reset*/
4112 if (!tmp_adev->asic_reset_res && !job_signaled)
4113 drm_sched_resubmit_jobs(&ring->sched);
4115 drm_sched_start(&ring->sched, !tmp_adev->asic_reset_res);
4118 if (!amdgpu_device_has_dc_support(tmp_adev) && !job_signaled) {
4119 drm_helper_resume_force_mode(tmp_adev->ddev);
4122 tmp_adev->asic_reset_res = 0;
4125 /* bad news, how to tell it to userspace ? */
4126 dev_info(tmp_adev->dev, "GPU reset(%d) failed\n", atomic_read(&tmp_adev->gpu_reset_counter));
4127 amdgpu_vf_error_put(tmp_adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r);
4129 dev_info(tmp_adev->dev, "GPU reset(%d) succeeded!\n", atomic_read(&tmp_adev->gpu_reset_counter));
4134 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
4135 /*unlock kfd: SRIOV would do it separately */
4136 if (!in_ras_intr && !amdgpu_sriov_vf(tmp_adev))
4137 amdgpu_amdkfd_post_reset(tmp_adev);
4138 amdgpu_device_unlock_adev(tmp_adev);
4142 mutex_unlock(&hive->reset_lock);
4145 dev_info(adev->dev, "GPU reset end with ret = %d\n", r);
4150 * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot
4152 * @adev: amdgpu_device pointer
4154 * Fetchs and stores in the driver the PCIE capabilities (gen speed
4155 * and lanes) of the slot the device is in. Handles APUs and
4156 * virtualized environments where PCIE config space may not be available.
4158 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev)
4160 struct pci_dev *pdev;
4161 enum pci_bus_speed speed_cap, platform_speed_cap;
4162 enum pcie_link_width platform_link_width;
4164 if (amdgpu_pcie_gen_cap)
4165 adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
4167 if (amdgpu_pcie_lane_cap)
4168 adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
4170 /* covers APUs as well */
4171 if (pci_is_root_bus(adev->pdev->bus)) {
4172 if (adev->pm.pcie_gen_mask == 0)
4173 adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
4174 if (adev->pm.pcie_mlw_mask == 0)
4175 adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
4179 if (adev->pm.pcie_gen_mask && adev->pm.pcie_mlw_mask)
4182 pcie_bandwidth_available(adev->pdev, NULL,
4183 &platform_speed_cap, &platform_link_width);
4185 if (adev->pm.pcie_gen_mask == 0) {
4188 speed_cap = pcie_get_speed_cap(pdev);
4189 if (speed_cap == PCI_SPEED_UNKNOWN) {
4190 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4191 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4192 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
4194 if (speed_cap == PCIE_SPEED_16_0GT)
4195 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4196 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4197 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
4198 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4);
4199 else if (speed_cap == PCIE_SPEED_8_0GT)
4200 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4201 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4202 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
4203 else if (speed_cap == PCIE_SPEED_5_0GT)
4204 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4205 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2);
4207 adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1;
4210 if (platform_speed_cap == PCI_SPEED_UNKNOWN) {
4211 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4212 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
4214 if (platform_speed_cap == PCIE_SPEED_16_0GT)
4215 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4216 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4217 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
4218 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4);
4219 else if (platform_speed_cap == PCIE_SPEED_8_0GT)
4220 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4221 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4222 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3);
4223 else if (platform_speed_cap == PCIE_SPEED_5_0GT)
4224 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4225 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
4227 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
4231 if (adev->pm.pcie_mlw_mask == 0) {
4232 if (platform_link_width == PCIE_LNK_WIDTH_UNKNOWN) {
4233 adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK;
4235 switch (platform_link_width) {
4237 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
4238 CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
4239 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
4240 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4241 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4242 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4243 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4246 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
4247 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
4248 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4249 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4250 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4251 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4254 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
4255 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4256 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4257 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4258 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4261 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4262 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4263 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4264 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4267 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4268 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4269 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4272 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4273 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4276 adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;