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/console.h>
31 #include <linux/slab.h>
33 #include <drm/drm_crtc_helper.h>
34 #include <drm/drm_atomic_helper.h>
35 #include <drm/amdgpu_drm.h>
36 #include <linux/vgaarb.h>
37 #include <linux/vga_switcheroo.h>
38 #include <linux/efi.h>
40 #include "amdgpu_trace.h"
41 #include "amdgpu_i2c.h"
43 #include "amdgpu_atombios.h"
44 #include "amdgpu_atomfirmware.h"
46 #ifdef CONFIG_DRM_AMDGPU_SI
49 #ifdef CONFIG_DRM_AMDGPU_CIK
54 #include "bif/bif_4_1_d.h"
55 #include <linux/pci.h>
56 #include <linux/firmware.h>
57 #include "amdgpu_vf_error.h"
59 #include "amdgpu_amdkfd.h"
60 #include "amdgpu_pm.h"
62 MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin");
63 MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin");
64 MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin");
65 MODULE_FIRMWARE("amdgpu/picasso_gpu_info.bin");
66 MODULE_FIRMWARE("amdgpu/raven2_gpu_info.bin");
68 #define AMDGPU_RESUME_MS 2000
70 static const char *amdgpu_asic_name[] = {
97 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev);
100 * amdgpu_device_is_px - Is the device is a dGPU with HG/PX power control
102 * @dev: drm_device pointer
104 * Returns true if the device is a dGPU with HG/PX power control,
105 * otherwise return false.
107 bool amdgpu_device_is_px(struct drm_device *dev)
109 struct amdgpu_device *adev = dev->dev_private;
111 if (adev->flags & AMD_IS_PX)
117 * MMIO register access helper functions.
120 * amdgpu_mm_rreg - read a memory mapped IO register
122 * @adev: amdgpu_device pointer
123 * @reg: dword aligned register offset
124 * @acc_flags: access flags which require special behavior
126 * Returns the 32 bit value from the offset specified.
128 uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg,
133 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
134 return amdgpu_virt_kiq_rreg(adev, reg);
136 if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
137 ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
141 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
142 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
143 ret = readl(((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
144 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
146 trace_amdgpu_mm_rreg(adev->pdev->device, reg, ret);
151 * MMIO register read with bytes helper functions
152 * @offset:bytes offset from MMIO start
157 * amdgpu_mm_rreg8 - read a memory mapped IO register
159 * @adev: amdgpu_device pointer
160 * @offset: byte aligned register offset
162 * Returns the 8 bit value from the offset specified.
164 uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset) {
165 if (offset < adev->rmmio_size)
166 return (readb(adev->rmmio + offset));
171 * MMIO register write with bytes helper functions
172 * @offset:bytes offset from MMIO start
173 * @value: the value want to be written to the register
177 * amdgpu_mm_wreg8 - read a memory mapped IO register
179 * @adev: amdgpu_device pointer
180 * @offset: byte aligned register offset
181 * @value: 8 bit value to write
183 * Writes the value specified to the offset specified.
185 void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value) {
186 if (offset < adev->rmmio_size)
187 writeb(value, adev->rmmio + offset);
193 * amdgpu_mm_wreg - write to a memory mapped IO register
195 * @adev: amdgpu_device pointer
196 * @reg: dword aligned register offset
197 * @v: 32 bit value to write to the register
198 * @acc_flags: access flags which require special behavior
200 * Writes the value specified to the offset specified.
202 void amdgpu_mm_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v,
205 trace_amdgpu_mm_wreg(adev->pdev->device, reg, v);
207 if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
208 adev->last_mm_index = v;
211 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
212 return amdgpu_virt_kiq_wreg(adev, reg, v);
214 if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
215 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
219 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
220 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
221 writel(v, ((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
222 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
225 if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
231 * amdgpu_io_rreg - read an IO register
233 * @adev: amdgpu_device pointer
234 * @reg: dword aligned register offset
236 * Returns the 32 bit value from the offset specified.
238 u32 amdgpu_io_rreg(struct amdgpu_device *adev, u32 reg)
240 if ((reg * 4) < adev->rio_mem_size)
241 return ioread32(adev->rio_mem + (reg * 4));
243 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
244 return ioread32(adev->rio_mem + (mmMM_DATA * 4));
249 * amdgpu_io_wreg - write to an IO register
251 * @adev: amdgpu_device pointer
252 * @reg: dword aligned register offset
253 * @v: 32 bit value to write to the register
255 * Writes the value specified to the offset specified.
257 void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32 v)
259 if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
260 adev->last_mm_index = v;
263 if ((reg * 4) < adev->rio_mem_size)
264 iowrite32(v, adev->rio_mem + (reg * 4));
266 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
267 iowrite32(v, adev->rio_mem + (mmMM_DATA * 4));
270 if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
276 * amdgpu_mm_rdoorbell - read a doorbell dword
278 * @adev: amdgpu_device pointer
279 * @index: doorbell index
281 * Returns the value in the doorbell aperture at the
282 * requested doorbell index (CIK).
284 u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
286 if (index < adev->doorbell.num_doorbells) {
287 return readl(adev->doorbell.ptr + index);
289 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
295 * amdgpu_mm_wdoorbell - write a doorbell dword
297 * @adev: amdgpu_device pointer
298 * @index: doorbell index
301 * Writes @v to the doorbell aperture at the
302 * requested doorbell index (CIK).
304 void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v)
306 if (index < adev->doorbell.num_doorbells) {
307 writel(v, adev->doorbell.ptr + index);
309 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
314 * amdgpu_mm_rdoorbell64 - read a doorbell Qword
316 * @adev: amdgpu_device pointer
317 * @index: doorbell index
319 * Returns the value in the doorbell aperture at the
320 * requested doorbell index (VEGA10+).
322 u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index)
324 if (index < adev->doorbell.num_doorbells) {
325 return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index));
327 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
333 * amdgpu_mm_wdoorbell64 - write a doorbell Qword
335 * @adev: amdgpu_device pointer
336 * @index: doorbell index
339 * Writes @v to the doorbell aperture at the
340 * requested doorbell index (VEGA10+).
342 void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v)
344 if (index < adev->doorbell.num_doorbells) {
345 atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v);
347 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
352 * amdgpu_invalid_rreg - dummy reg read function
354 * @adev: amdgpu device pointer
355 * @reg: offset of register
357 * Dummy register read function. Used for register blocks
358 * that certain asics don't have (all asics).
359 * Returns the value in the register.
361 static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
363 DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
369 * amdgpu_invalid_wreg - dummy reg write function
371 * @adev: amdgpu device pointer
372 * @reg: offset of register
373 * @v: value to write to the register
375 * Dummy register read function. Used for register blocks
376 * that certain asics don't have (all asics).
378 static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
380 DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
386 * amdgpu_block_invalid_rreg - dummy reg read function
388 * @adev: amdgpu device pointer
389 * @block: offset of instance
390 * @reg: offset of register
392 * Dummy register read function. Used for register blocks
393 * that certain asics don't have (all asics).
394 * Returns the value in the register.
396 static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
397 uint32_t block, uint32_t reg)
399 DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
406 * amdgpu_block_invalid_wreg - dummy reg write function
408 * @adev: amdgpu device pointer
409 * @block: offset of instance
410 * @reg: offset of register
411 * @v: value to write to the register
413 * Dummy register read function. Used for register blocks
414 * that certain asics don't have (all asics).
416 static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
418 uint32_t reg, uint32_t v)
420 DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
426 * amdgpu_device_vram_scratch_init - allocate the VRAM scratch page
428 * @adev: amdgpu device pointer
430 * Allocates a scratch page of VRAM for use by various things in the
433 static int amdgpu_device_vram_scratch_init(struct amdgpu_device *adev)
435 return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE,
436 PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
437 &adev->vram_scratch.robj,
438 &adev->vram_scratch.gpu_addr,
439 (void **)&adev->vram_scratch.ptr);
443 * amdgpu_device_vram_scratch_fini - Free the VRAM scratch page
445 * @adev: amdgpu device pointer
447 * Frees the VRAM scratch page.
449 static void amdgpu_device_vram_scratch_fini(struct amdgpu_device *adev)
451 amdgpu_bo_free_kernel(&adev->vram_scratch.robj, NULL, NULL);
455 * amdgpu_device_program_register_sequence - program an array of registers.
457 * @adev: amdgpu_device pointer
458 * @registers: pointer to the register array
459 * @array_size: size of the register array
461 * Programs an array or registers with and and or masks.
462 * This is a helper for setting golden registers.
464 void amdgpu_device_program_register_sequence(struct amdgpu_device *adev,
465 const u32 *registers,
466 const u32 array_size)
468 u32 tmp, reg, and_mask, or_mask;
474 for (i = 0; i < array_size; i +=3) {
475 reg = registers[i + 0];
476 and_mask = registers[i + 1];
477 or_mask = registers[i + 2];
479 if (and_mask == 0xffffffff) {
491 * amdgpu_device_pci_config_reset - reset the GPU
493 * @adev: amdgpu_device pointer
495 * Resets the GPU using the pci config reset sequence.
496 * Only applicable to asics prior to vega10.
498 void amdgpu_device_pci_config_reset(struct amdgpu_device *adev)
500 pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
504 * GPU doorbell aperture helpers function.
507 * amdgpu_device_doorbell_init - Init doorbell driver information.
509 * @adev: amdgpu_device pointer
511 * Init doorbell driver information (CIK)
512 * Returns 0 on success, error on failure.
514 static int amdgpu_device_doorbell_init(struct amdgpu_device *adev)
516 /* No doorbell on SI hardware generation */
517 if (adev->asic_type < CHIP_BONAIRE) {
518 adev->doorbell.base = 0;
519 adev->doorbell.size = 0;
520 adev->doorbell.num_doorbells = 0;
521 adev->doorbell.ptr = NULL;
525 if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET)
528 /* doorbell bar mapping */
529 adev->doorbell.base = pci_resource_start(adev->pdev, 2);
530 adev->doorbell.size = pci_resource_len(adev->pdev, 2);
532 adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32),
533 AMDGPU_DOORBELL_MAX_ASSIGNMENT+1);
534 if (adev->doorbell.num_doorbells == 0)
537 adev->doorbell.ptr = ioremap(adev->doorbell.base,
538 adev->doorbell.num_doorbells *
540 if (adev->doorbell.ptr == NULL)
547 * amdgpu_device_doorbell_fini - Tear down doorbell driver information.
549 * @adev: amdgpu_device pointer
551 * Tear down doorbell driver information (CIK)
553 static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev)
555 iounmap(adev->doorbell.ptr);
556 adev->doorbell.ptr = NULL;
562 * amdgpu_device_wb_*()
563 * Writeback is the method by which the GPU updates special pages in memory
564 * with the status of certain GPU events (fences, ring pointers,etc.).
568 * amdgpu_device_wb_fini - Disable Writeback and free memory
570 * @adev: amdgpu_device pointer
572 * Disables Writeback and frees the Writeback memory (all asics).
573 * Used at driver shutdown.
575 static void amdgpu_device_wb_fini(struct amdgpu_device *adev)
577 if (adev->wb.wb_obj) {
578 amdgpu_bo_free_kernel(&adev->wb.wb_obj,
580 (void **)&adev->wb.wb);
581 adev->wb.wb_obj = NULL;
586 * amdgpu_device_wb_init- Init Writeback driver info and allocate memory
588 * @adev: amdgpu_device pointer
590 * Initializes writeback and allocates writeback memory (all asics).
591 * Used at driver startup.
592 * Returns 0 on success or an -error on failure.
594 static int amdgpu_device_wb_init(struct amdgpu_device *adev)
598 if (adev->wb.wb_obj == NULL) {
599 /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */
600 r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8,
601 PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
602 &adev->wb.wb_obj, &adev->wb.gpu_addr,
603 (void **)&adev->wb.wb);
605 dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
609 adev->wb.num_wb = AMDGPU_MAX_WB;
610 memset(&adev->wb.used, 0, sizeof(adev->wb.used));
612 /* clear wb memory */
613 memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8);
620 * amdgpu_device_wb_get - Allocate a wb entry
622 * @adev: amdgpu_device pointer
625 * Allocate a wb slot for use by the driver (all asics).
626 * Returns 0 on success or -EINVAL on failure.
628 int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb)
630 unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
632 if (offset < adev->wb.num_wb) {
633 __set_bit(offset, adev->wb.used);
634 *wb = offset << 3; /* convert to dw offset */
642 * amdgpu_device_wb_free - Free a wb entry
644 * @adev: amdgpu_device pointer
647 * Free a wb slot allocated for use by the driver (all asics)
649 void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb)
652 if (wb < adev->wb.num_wb)
653 __clear_bit(wb, adev->wb.used);
657 * amdgpu_device_resize_fb_bar - try to resize FB BAR
659 * @adev: amdgpu_device pointer
661 * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not
662 * to fail, but if any of the BARs is not accessible after the size we abort
663 * driver loading by returning -ENODEV.
665 int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev)
667 u64 space_needed = roundup_pow_of_two(adev->gmc.real_vram_size);
668 u32 rbar_size = order_base_2(((space_needed >> 20) | 1)) - 1;
669 struct pci_bus *root;
670 struct resource *res;
676 if (amdgpu_sriov_vf(adev))
679 /* Check if the root BUS has 64bit memory resources */
680 root = adev->pdev->bus;
684 pci_bus_for_each_resource(root, res, i) {
685 if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) &&
686 res->start > 0x100000000ull)
690 /* Trying to resize is pointless without a root hub window above 4GB */
694 /* Disable memory decoding while we change the BAR addresses and size */
695 pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd);
696 pci_write_config_word(adev->pdev, PCI_COMMAND,
697 cmd & ~PCI_COMMAND_MEMORY);
699 /* Free the VRAM and doorbell BAR, we most likely need to move both. */
700 amdgpu_device_doorbell_fini(adev);
701 if (adev->asic_type >= CHIP_BONAIRE)
702 pci_release_resource(adev->pdev, 2);
704 pci_release_resource(adev->pdev, 0);
706 r = pci_resize_resource(adev->pdev, 0, rbar_size);
708 DRM_INFO("Not enough PCI address space for a large BAR.");
709 else if (r && r != -ENOTSUPP)
710 DRM_ERROR("Problem resizing BAR0 (%d).", r);
712 pci_assign_unassigned_bus_resources(adev->pdev->bus);
714 /* When the doorbell or fb BAR isn't available we have no chance of
717 r = amdgpu_device_doorbell_init(adev);
718 if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET))
721 pci_write_config_word(adev->pdev, PCI_COMMAND, cmd);
727 * GPU helpers function.
730 * amdgpu_device_need_post - check if the hw need post or not
732 * @adev: amdgpu_device pointer
734 * Check if the asic has been initialized (all asics) at driver startup
735 * or post is needed if hw reset is performed.
736 * Returns true if need or false if not.
738 bool amdgpu_device_need_post(struct amdgpu_device *adev)
742 if (amdgpu_sriov_vf(adev))
745 if (amdgpu_passthrough(adev)) {
746 /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
747 * some old smc fw still need driver do vPost otherwise gpu hang, while
748 * those smc fw version above 22.15 doesn't have this flaw, so we force
749 * vpost executed for smc version below 22.15
751 if (adev->asic_type == CHIP_FIJI) {
754 err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
755 /* force vPost if error occured */
759 fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
760 if (fw_ver < 0x00160e00)
765 if (adev->has_hw_reset) {
766 adev->has_hw_reset = false;
770 /* bios scratch used on CIK+ */
771 if (adev->asic_type >= CHIP_BONAIRE)
772 return amdgpu_atombios_scratch_need_asic_init(adev);
774 /* check MEM_SIZE for older asics */
775 reg = amdgpu_asic_get_config_memsize(adev);
777 if ((reg != 0) && (reg != 0xffffffff))
783 /* if we get transitioned to only one device, take VGA back */
785 * amdgpu_device_vga_set_decode - enable/disable vga decode
787 * @cookie: amdgpu_device pointer
788 * @state: enable/disable vga decode
790 * Enable/disable vga decode (all asics).
791 * Returns VGA resource flags.
793 static unsigned int amdgpu_device_vga_set_decode(void *cookie, bool state)
795 struct amdgpu_device *adev = cookie;
796 amdgpu_asic_set_vga_state(adev, state);
798 return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
799 VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
801 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
805 * amdgpu_device_check_block_size - validate the vm block size
807 * @adev: amdgpu_device pointer
809 * Validates the vm block size specified via module parameter.
810 * The vm block size defines number of bits in page table versus page directory,
811 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
812 * page table and the remaining bits are in the page directory.
814 static void amdgpu_device_check_block_size(struct amdgpu_device *adev)
816 /* defines number of bits in page table versus page directory,
817 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
818 * page table and the remaining bits are in the page directory */
819 if (amdgpu_vm_block_size == -1)
822 if (amdgpu_vm_block_size < 9) {
823 dev_warn(adev->dev, "VM page table size (%d) too small\n",
824 amdgpu_vm_block_size);
825 amdgpu_vm_block_size = -1;
830 * amdgpu_device_check_vm_size - validate the vm size
832 * @adev: amdgpu_device pointer
834 * Validates the vm size in GB specified via module parameter.
835 * The VM size is the size of the GPU virtual memory space in GB.
837 static void amdgpu_device_check_vm_size(struct amdgpu_device *adev)
839 /* no need to check the default value */
840 if (amdgpu_vm_size == -1)
843 if (amdgpu_vm_size < 1) {
844 dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
850 static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev)
853 bool is_os_64 = (sizeof(void *) == 8) ? true : false;
854 uint64_t total_memory;
855 uint64_t dram_size_seven_GB = 0x1B8000000;
856 uint64_t dram_size_three_GB = 0xB8000000;
858 if (amdgpu_smu_memory_pool_size == 0)
862 DRM_WARN("Not 64-bit OS, feature not supported\n");
866 total_memory = (uint64_t)si.totalram * si.mem_unit;
868 if ((amdgpu_smu_memory_pool_size == 1) ||
869 (amdgpu_smu_memory_pool_size == 2)) {
870 if (total_memory < dram_size_three_GB)
872 } else if ((amdgpu_smu_memory_pool_size == 4) ||
873 (amdgpu_smu_memory_pool_size == 8)) {
874 if (total_memory < dram_size_seven_GB)
877 DRM_WARN("Smu memory pool size not supported\n");
880 adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28;
885 DRM_WARN("No enough system memory\n");
887 adev->pm.smu_prv_buffer_size = 0;
891 * amdgpu_device_check_arguments - validate module params
893 * @adev: amdgpu_device pointer
895 * Validates certain module parameters and updates
896 * the associated values used by the driver (all asics).
898 static void amdgpu_device_check_arguments(struct amdgpu_device *adev)
900 if (amdgpu_sched_jobs < 4) {
901 dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
903 amdgpu_sched_jobs = 4;
904 } else if (!is_power_of_2(amdgpu_sched_jobs)){
905 dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
907 amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
910 if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) {
911 /* gart size must be greater or equal to 32M */
912 dev_warn(adev->dev, "gart size (%d) too small\n",
914 amdgpu_gart_size = -1;
917 if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) {
918 /* gtt size must be greater or equal to 32M */
919 dev_warn(adev->dev, "gtt size (%d) too small\n",
921 amdgpu_gtt_size = -1;
924 /* valid range is between 4 and 9 inclusive */
925 if (amdgpu_vm_fragment_size != -1 &&
926 (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) {
927 dev_warn(adev->dev, "valid range is between 4 and 9\n");
928 amdgpu_vm_fragment_size = -1;
931 amdgpu_device_check_smu_prv_buffer_size(adev);
933 amdgpu_device_check_vm_size(adev);
935 amdgpu_device_check_block_size(adev);
937 if (amdgpu_vram_page_split != -1 && (amdgpu_vram_page_split < 16 ||
938 !is_power_of_2(amdgpu_vram_page_split))) {
939 dev_warn(adev->dev, "invalid VRAM page split (%d)\n",
940 amdgpu_vram_page_split);
941 amdgpu_vram_page_split = 1024;
944 if (amdgpu_lockup_timeout == 0) {
945 dev_warn(adev->dev, "lockup_timeout msut be > 0, adjusting to 10000\n");
946 amdgpu_lockup_timeout = 10000;
949 adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type);
953 * amdgpu_switcheroo_set_state - set switcheroo state
955 * @pdev: pci dev pointer
956 * @state: vga_switcheroo state
958 * Callback for the switcheroo driver. Suspends or resumes the
959 * the asics before or after it is powered up using ACPI methods.
961 static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, enum vga_switcheroo_state state)
963 struct drm_device *dev = pci_get_drvdata(pdev);
965 if (amdgpu_device_is_px(dev) && state == VGA_SWITCHEROO_OFF)
968 if (state == VGA_SWITCHEROO_ON) {
969 pr_info("amdgpu: switched on\n");
970 /* don't suspend or resume card normally */
971 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
973 amdgpu_device_resume(dev, true, true);
975 dev->switch_power_state = DRM_SWITCH_POWER_ON;
976 drm_kms_helper_poll_enable(dev);
978 pr_info("amdgpu: switched off\n");
979 drm_kms_helper_poll_disable(dev);
980 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
981 amdgpu_device_suspend(dev, true, true);
982 dev->switch_power_state = DRM_SWITCH_POWER_OFF;
987 * amdgpu_switcheroo_can_switch - see if switcheroo state can change
989 * @pdev: pci dev pointer
991 * Callback for the switcheroo driver. Check of the switcheroo
992 * state can be changed.
993 * Returns true if the state can be changed, false if not.
995 static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
997 struct drm_device *dev = pci_get_drvdata(pdev);
1000 * FIXME: open_count is protected by drm_global_mutex but that would lead to
1001 * locking inversion with the driver load path. And the access here is
1002 * completely racy anyway. So don't bother with locking for now.
1004 return dev->open_count == 0;
1007 static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
1008 .set_gpu_state = amdgpu_switcheroo_set_state,
1010 .can_switch = amdgpu_switcheroo_can_switch,
1014 * amdgpu_device_ip_set_clockgating_state - set the CG state
1016 * @dev: amdgpu_device pointer
1017 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1018 * @state: clockgating state (gate or ungate)
1020 * Sets the requested clockgating state for all instances of
1021 * the hardware IP specified.
1022 * Returns the error code from the last instance.
1024 int amdgpu_device_ip_set_clockgating_state(void *dev,
1025 enum amd_ip_block_type block_type,
1026 enum amd_clockgating_state state)
1028 struct amdgpu_device *adev = dev;
1031 for (i = 0; i < adev->num_ip_blocks; i++) {
1032 if (!adev->ip_blocks[i].status.valid)
1034 if (adev->ip_blocks[i].version->type != block_type)
1036 if (!adev->ip_blocks[i].version->funcs->set_clockgating_state)
1038 r = adev->ip_blocks[i].version->funcs->set_clockgating_state(
1039 (void *)adev, state);
1041 DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",
1042 adev->ip_blocks[i].version->funcs->name, r);
1048 * amdgpu_device_ip_set_powergating_state - set the PG state
1050 * @dev: amdgpu_device pointer
1051 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1052 * @state: powergating state (gate or ungate)
1054 * Sets the requested powergating state for all instances of
1055 * the hardware IP specified.
1056 * Returns the error code from the last instance.
1058 int amdgpu_device_ip_set_powergating_state(void *dev,
1059 enum amd_ip_block_type block_type,
1060 enum amd_powergating_state state)
1062 struct amdgpu_device *adev = dev;
1065 for (i = 0; i < adev->num_ip_blocks; i++) {
1066 if (!adev->ip_blocks[i].status.valid)
1068 if (adev->ip_blocks[i].version->type != block_type)
1070 if (!adev->ip_blocks[i].version->funcs->set_powergating_state)
1072 r = adev->ip_blocks[i].version->funcs->set_powergating_state(
1073 (void *)adev, state);
1075 DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",
1076 adev->ip_blocks[i].version->funcs->name, r);
1082 * amdgpu_device_ip_get_clockgating_state - get the CG state
1084 * @adev: amdgpu_device pointer
1085 * @flags: clockgating feature flags
1087 * Walks the list of IPs on the device and updates the clockgating
1088 * flags for each IP.
1089 * Updates @flags with the feature flags for each hardware IP where
1090 * clockgating is enabled.
1092 void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev,
1097 for (i = 0; i < adev->num_ip_blocks; i++) {
1098 if (!adev->ip_blocks[i].status.valid)
1100 if (adev->ip_blocks[i].version->funcs->get_clockgating_state)
1101 adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags);
1106 * amdgpu_device_ip_wait_for_idle - wait for idle
1108 * @adev: amdgpu_device pointer
1109 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1111 * Waits for the request hardware IP to be idle.
1112 * Returns 0 for success or a negative error code on failure.
1114 int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev,
1115 enum amd_ip_block_type block_type)
1119 for (i = 0; i < adev->num_ip_blocks; i++) {
1120 if (!adev->ip_blocks[i].status.valid)
1122 if (adev->ip_blocks[i].version->type == block_type) {
1123 r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
1134 * amdgpu_device_ip_is_idle - is the hardware IP idle
1136 * @adev: amdgpu_device pointer
1137 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1139 * Check if the hardware IP is idle or not.
1140 * Returns true if it the IP is idle, false if not.
1142 bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev,
1143 enum amd_ip_block_type block_type)
1147 for (i = 0; i < adev->num_ip_blocks; i++) {
1148 if (!adev->ip_blocks[i].status.valid)
1150 if (adev->ip_blocks[i].version->type == block_type)
1151 return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
1158 * amdgpu_device_ip_get_ip_block - get a hw IP pointer
1160 * @adev: amdgpu_device pointer
1161 * @type: Type of hardware IP (SMU, GFX, UVD, etc.)
1163 * Returns a pointer to the hardware IP block structure
1164 * if it exists for the asic, otherwise NULL.
1166 struct amdgpu_ip_block *
1167 amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,
1168 enum amd_ip_block_type type)
1172 for (i = 0; i < adev->num_ip_blocks; i++)
1173 if (adev->ip_blocks[i].version->type == type)
1174 return &adev->ip_blocks[i];
1180 * amdgpu_device_ip_block_version_cmp
1182 * @adev: amdgpu_device pointer
1183 * @type: enum amd_ip_block_type
1184 * @major: major version
1185 * @minor: minor version
1187 * return 0 if equal or greater
1188 * return 1 if smaller or the ip_block doesn't exist
1190 int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev,
1191 enum amd_ip_block_type type,
1192 u32 major, u32 minor)
1194 struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type);
1196 if (ip_block && ((ip_block->version->major > major) ||
1197 ((ip_block->version->major == major) &&
1198 (ip_block->version->minor >= minor))))
1205 * amdgpu_device_ip_block_add
1207 * @adev: amdgpu_device pointer
1208 * @ip_block_version: pointer to the IP to add
1210 * Adds the IP block driver information to the collection of IPs
1213 int amdgpu_device_ip_block_add(struct amdgpu_device *adev,
1214 const struct amdgpu_ip_block_version *ip_block_version)
1216 if (!ip_block_version)
1219 DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks,
1220 ip_block_version->funcs->name);
1222 adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
1228 * amdgpu_device_enable_virtual_display - enable virtual display feature
1230 * @adev: amdgpu_device pointer
1232 * Enabled the virtual display feature if the user has enabled it via
1233 * the module parameter virtual_display. This feature provides a virtual
1234 * display hardware on headless boards or in virtualized environments.
1235 * This function parses and validates the configuration string specified by
1236 * the user and configues the virtual display configuration (number of
1237 * virtual connectors, crtcs, etc.) specified.
1239 static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
1241 adev->enable_virtual_display = false;
1243 if (amdgpu_virtual_display) {
1244 struct drm_device *ddev = adev->ddev;
1245 const char *pci_address_name = pci_name(ddev->pdev);
1246 char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
1248 pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
1249 pciaddstr_tmp = pciaddstr;
1250 while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
1251 pciaddname = strsep(&pciaddname_tmp, ",");
1252 if (!strcmp("all", pciaddname)
1253 || !strcmp(pci_address_name, pciaddname)) {
1257 adev->enable_virtual_display = true;
1260 res = kstrtol(pciaddname_tmp, 10,
1268 adev->mode_info.num_crtc = num_crtc;
1270 adev->mode_info.num_crtc = 1;
1276 DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
1277 amdgpu_virtual_display, pci_address_name,
1278 adev->enable_virtual_display, adev->mode_info.num_crtc);
1285 * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware
1287 * @adev: amdgpu_device pointer
1289 * Parses the asic configuration parameters specified in the gpu info
1290 * firmware and makes them availale to the driver for use in configuring
1292 * Returns 0 on success, -EINVAL on failure.
1294 static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev)
1296 const char *chip_name;
1299 const struct gpu_info_firmware_header_v1_0 *hdr;
1301 adev->firmware.gpu_info_fw = NULL;
1303 switch (adev->asic_type) {
1307 case CHIP_POLARIS10:
1308 case CHIP_POLARIS11:
1309 case CHIP_POLARIS12:
1313 #ifdef CONFIG_DRM_AMDGPU_SI
1320 #ifdef CONFIG_DRM_AMDGPU_CIK
1331 chip_name = "vega10";
1334 chip_name = "vega12";
1337 if (adev->rev_id >= 8)
1338 chip_name = "raven2";
1339 else if (adev->pdev->device == 0x15d8)
1340 chip_name = "picasso";
1342 chip_name = "raven";
1346 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name);
1347 err = request_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev);
1350 "Failed to load gpu_info firmware \"%s\"\n",
1354 err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw);
1357 "Failed to validate gpu_info firmware \"%s\"\n",
1362 hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data;
1363 amdgpu_ucode_print_gpu_info_hdr(&hdr->header);
1365 switch (hdr->version_major) {
1368 const struct gpu_info_firmware_v1_0 *gpu_info_fw =
1369 (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data +
1370 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1372 adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se);
1373 adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh);
1374 adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se);
1375 adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se);
1376 adev->gfx.config.max_texture_channel_caches =
1377 le32_to_cpu(gpu_info_fw->gc_num_tccs);
1378 adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs);
1379 adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds);
1380 adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth);
1381 adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth);
1382 adev->gfx.config.double_offchip_lds_buf =
1383 le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer);
1384 adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size);
1385 adev->gfx.cu_info.max_waves_per_simd =
1386 le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd);
1387 adev->gfx.cu_info.max_scratch_slots_per_cu =
1388 le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu);
1389 adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size);
1394 "Unsupported gpu_info table %d\n", hdr->header.ucode_version);
1403 * amdgpu_device_ip_early_init - run early init for hardware IPs
1405 * @adev: amdgpu_device pointer
1407 * Early initialization pass for hardware IPs. The hardware IPs that make
1408 * up each asic are discovered each IP's early_init callback is run. This
1409 * is the first stage in initializing the asic.
1410 * Returns 0 on success, negative error code on failure.
1412 static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
1416 amdgpu_device_enable_virtual_display(adev);
1418 switch (adev->asic_type) {
1422 case CHIP_POLARIS10:
1423 case CHIP_POLARIS11:
1424 case CHIP_POLARIS12:
1428 if (adev->asic_type == CHIP_CARRIZO || adev->asic_type == CHIP_STONEY)
1429 adev->family = AMDGPU_FAMILY_CZ;
1431 adev->family = AMDGPU_FAMILY_VI;
1433 r = vi_set_ip_blocks(adev);
1437 #ifdef CONFIG_DRM_AMDGPU_SI
1443 adev->family = AMDGPU_FAMILY_SI;
1444 r = si_set_ip_blocks(adev);
1449 #ifdef CONFIG_DRM_AMDGPU_CIK
1455 if ((adev->asic_type == CHIP_BONAIRE) || (adev->asic_type == CHIP_HAWAII))
1456 adev->family = AMDGPU_FAMILY_CI;
1458 adev->family = AMDGPU_FAMILY_KV;
1460 r = cik_set_ip_blocks(adev);
1469 if (adev->asic_type == CHIP_RAVEN)
1470 adev->family = AMDGPU_FAMILY_RV;
1472 adev->family = AMDGPU_FAMILY_AI;
1474 r = soc15_set_ip_blocks(adev);
1479 /* FIXME: not supported yet */
1483 r = amdgpu_device_parse_gpu_info_fw(adev);
1487 amdgpu_amdkfd_device_probe(adev);
1489 if (amdgpu_sriov_vf(adev)) {
1490 r = amdgpu_virt_request_full_gpu(adev, true);
1495 adev->powerplay.pp_feature = amdgpu_pp_feature_mask;
1497 for (i = 0; i < adev->num_ip_blocks; i++) {
1498 if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
1499 DRM_ERROR("disabled ip block: %d <%s>\n",
1500 i, adev->ip_blocks[i].version->funcs->name);
1501 adev->ip_blocks[i].status.valid = false;
1503 if (adev->ip_blocks[i].version->funcs->early_init) {
1504 r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
1506 adev->ip_blocks[i].status.valid = false;
1508 DRM_ERROR("early_init of IP block <%s> failed %d\n",
1509 adev->ip_blocks[i].version->funcs->name, r);
1512 adev->ip_blocks[i].status.valid = true;
1515 adev->ip_blocks[i].status.valid = true;
1520 adev->cg_flags &= amdgpu_cg_mask;
1521 adev->pg_flags &= amdgpu_pg_mask;
1526 static int amdgpu_device_ip_hw_init_phase1(struct amdgpu_device *adev)
1530 for (i = 0; i < adev->num_ip_blocks; i++) {
1531 if (!adev->ip_blocks[i].status.sw)
1533 if (adev->ip_blocks[i].status.hw)
1535 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
1536 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
1537 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
1539 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1540 adev->ip_blocks[i].version->funcs->name, r);
1543 adev->ip_blocks[i].status.hw = true;
1550 static int amdgpu_device_ip_hw_init_phase2(struct amdgpu_device *adev)
1554 for (i = 0; i < adev->num_ip_blocks; i++) {
1555 if (!adev->ip_blocks[i].status.sw)
1557 if (adev->ip_blocks[i].status.hw)
1559 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
1561 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1562 adev->ip_blocks[i].version->funcs->name, r);
1565 adev->ip_blocks[i].status.hw = true;
1571 static int amdgpu_device_fw_loading(struct amdgpu_device *adev)
1576 if (adev->asic_type >= CHIP_VEGA10) {
1577 for (i = 0; i < adev->num_ip_blocks; i++) {
1578 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
1579 if (adev->in_gpu_reset || adev->in_suspend) {
1580 if (amdgpu_sriov_vf(adev) && adev->in_gpu_reset)
1581 break; /* sriov gpu reset, psp need to do hw_init before IH because of hw limit */
1582 r = adev->ip_blocks[i].version->funcs->resume(adev);
1584 DRM_ERROR("resume of IP block <%s> failed %d\n",
1585 adev->ip_blocks[i].version->funcs->name, r);
1589 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
1591 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1592 adev->ip_blocks[i].version->funcs->name, r);
1596 adev->ip_blocks[i].status.hw = true;
1601 if (adev->powerplay.pp_funcs && adev->powerplay.pp_funcs->load_firmware) {
1602 r = adev->powerplay.pp_funcs->load_firmware(adev->powerplay.pp_handle);
1604 pr_err("firmware loading failed\n");
1613 * amdgpu_device_ip_init - run init for hardware IPs
1615 * @adev: amdgpu_device pointer
1617 * Main initialization pass for hardware IPs. The list of all the hardware
1618 * IPs that make up the asic is walked and the sw_init and hw_init callbacks
1619 * are run. sw_init initializes the software state associated with each IP
1620 * and hw_init initializes the hardware associated with each IP.
1621 * Returns 0 on success, negative error code on failure.
1623 static int amdgpu_device_ip_init(struct amdgpu_device *adev)
1627 for (i = 0; i < adev->num_ip_blocks; i++) {
1628 if (!adev->ip_blocks[i].status.valid)
1630 r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
1632 DRM_ERROR("sw_init of IP block <%s> failed %d\n",
1633 adev->ip_blocks[i].version->funcs->name, r);
1636 adev->ip_blocks[i].status.sw = true;
1638 /* need to do gmc hw init early so we can allocate gpu mem */
1639 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1640 r = amdgpu_device_vram_scratch_init(adev);
1642 DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r);
1645 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1647 DRM_ERROR("hw_init %d failed %d\n", i, r);
1650 r = amdgpu_device_wb_init(adev);
1652 DRM_ERROR("amdgpu_device_wb_init failed %d\n", r);
1655 adev->ip_blocks[i].status.hw = true;
1657 /* right after GMC hw init, we create CSA */
1658 if (amdgpu_sriov_vf(adev)) {
1659 r = amdgpu_allocate_static_csa(adev);
1661 DRM_ERROR("allocate CSA failed %d\n", r);
1668 r = amdgpu_ucode_create_bo(adev); /* create ucode bo when sw_init complete*/
1672 r = amdgpu_device_ip_hw_init_phase1(adev);
1676 r = amdgpu_device_fw_loading(adev);
1680 r = amdgpu_device_ip_hw_init_phase2(adev);
1684 amdgpu_xgmi_add_device(adev);
1685 amdgpu_amdkfd_device_init(adev);
1687 if (amdgpu_sriov_vf(adev))
1688 amdgpu_virt_release_full_gpu(adev, true);
1694 * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer
1696 * @adev: amdgpu_device pointer
1698 * Writes a reset magic value to the gart pointer in VRAM. The driver calls
1699 * this function before a GPU reset. If the value is retained after a
1700 * GPU reset, VRAM has not been lost. Some GPU resets may destry VRAM contents.
1702 static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev)
1704 memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
1708 * amdgpu_device_check_vram_lost - check if vram is valid
1710 * @adev: amdgpu_device pointer
1712 * Checks the reset magic value written to the gart pointer in VRAM.
1713 * The driver calls this after a GPU reset to see if the contents of
1714 * VRAM is lost or now.
1715 * returns true if vram is lost, false if not.
1717 static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev)
1719 return !!memcmp(adev->gart.ptr, adev->reset_magic,
1720 AMDGPU_RESET_MAGIC_NUM);
1724 * amdgpu_device_set_cg_state - set clockgating for amdgpu device
1726 * @adev: amdgpu_device pointer
1728 * The list of all the hardware IPs that make up the asic is walked and the
1729 * set_clockgating_state callbacks are run.
1730 * Late initialization pass enabling clockgating for hardware IPs.
1731 * Fini or suspend, pass disabling clockgating for hardware IPs.
1732 * Returns 0 on success, negative error code on failure.
1735 static int amdgpu_device_set_cg_state(struct amdgpu_device *adev,
1736 enum amd_clockgating_state state)
1740 if (amdgpu_emu_mode == 1)
1743 for (j = 0; j < adev->num_ip_blocks; j++) {
1744 i = state == AMD_CG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
1745 if (!adev->ip_blocks[i].status.late_initialized)
1747 /* skip CG for VCE/UVD, it's handled specially */
1748 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1749 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1750 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
1751 adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1752 /* enable clockgating to save power */
1753 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1756 DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
1757 adev->ip_blocks[i].version->funcs->name, r);
1766 static int amdgpu_device_set_pg_state(struct amdgpu_device *adev, enum amd_powergating_state state)
1770 if (amdgpu_emu_mode == 1)
1773 for (j = 0; j < adev->num_ip_blocks; j++) {
1774 i = state == AMD_PG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
1775 if (!adev->ip_blocks[i].status.late_initialized)
1777 /* skip CG for VCE/UVD, it's handled specially */
1778 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1779 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1780 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
1781 adev->ip_blocks[i].version->funcs->set_powergating_state) {
1782 /* enable powergating to save power */
1783 r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev,
1786 DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n",
1787 adev->ip_blocks[i].version->funcs->name, r);
1796 * amdgpu_device_ip_late_init - run late init for hardware IPs
1798 * @adev: amdgpu_device pointer
1800 * Late initialization pass for hardware IPs. The list of all the hardware
1801 * IPs that make up the asic is walked and the late_init callbacks are run.
1802 * late_init covers any special initialization that an IP requires
1803 * after all of the have been initialized or something that needs to happen
1804 * late in the init process.
1805 * Returns 0 on success, negative error code on failure.
1807 static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
1811 for (i = 0; i < adev->num_ip_blocks; i++) {
1812 if (!adev->ip_blocks[i].status.hw)
1814 if (adev->ip_blocks[i].version->funcs->late_init) {
1815 r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
1817 DRM_ERROR("late_init of IP block <%s> failed %d\n",
1818 adev->ip_blocks[i].version->funcs->name, r);
1822 adev->ip_blocks[i].status.late_initialized = true;
1825 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE);
1826 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE);
1828 queue_delayed_work(system_wq, &adev->late_init_work,
1829 msecs_to_jiffies(AMDGPU_RESUME_MS));
1831 amdgpu_device_fill_reset_magic(adev);
1837 * amdgpu_device_ip_fini - run fini for hardware IPs
1839 * @adev: amdgpu_device pointer
1841 * Main teardown pass for hardware IPs. The list of all the hardware
1842 * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks
1843 * are run. hw_fini tears down the hardware associated with each IP
1844 * and sw_fini tears down any software state associated with each IP.
1845 * Returns 0 on success, negative error code on failure.
1847 static int amdgpu_device_ip_fini(struct amdgpu_device *adev)
1851 amdgpu_amdkfd_device_fini(adev);
1853 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
1854 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
1856 /* need to disable SMC first */
1857 for (i = 0; i < adev->num_ip_blocks; i++) {
1858 if (!adev->ip_blocks[i].status.hw)
1860 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
1861 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
1862 /* XXX handle errors */
1864 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
1865 adev->ip_blocks[i].version->funcs->name, r);
1867 adev->ip_blocks[i].status.hw = false;
1872 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1873 if (!adev->ip_blocks[i].status.hw)
1876 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
1877 /* XXX handle errors */
1879 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
1880 adev->ip_blocks[i].version->funcs->name, r);
1883 adev->ip_blocks[i].status.hw = false;
1887 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1888 if (!adev->ip_blocks[i].status.sw)
1891 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1892 amdgpu_ucode_free_bo(adev);
1893 amdgpu_free_static_csa(adev);
1894 amdgpu_device_wb_fini(adev);
1895 amdgpu_device_vram_scratch_fini(adev);
1898 r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
1899 /* XXX handle errors */
1901 DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
1902 adev->ip_blocks[i].version->funcs->name, r);
1904 adev->ip_blocks[i].status.sw = false;
1905 adev->ip_blocks[i].status.valid = false;
1908 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1909 if (!adev->ip_blocks[i].status.late_initialized)
1911 if (adev->ip_blocks[i].version->funcs->late_fini)
1912 adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
1913 adev->ip_blocks[i].status.late_initialized = false;
1916 if (amdgpu_sriov_vf(adev))
1917 if (amdgpu_virt_release_full_gpu(adev, false))
1918 DRM_ERROR("failed to release exclusive mode on fini\n");
1923 static int amdgpu_device_enable_mgpu_fan_boost(void)
1925 struct amdgpu_gpu_instance *gpu_ins;
1926 struct amdgpu_device *adev;
1929 mutex_lock(&mgpu_info.mutex);
1932 * MGPU fan boost feature should be enabled
1933 * only when there are two or more dGPUs in
1936 if (mgpu_info.num_dgpu < 2)
1939 for (i = 0; i < mgpu_info.num_dgpu; i++) {
1940 gpu_ins = &(mgpu_info.gpu_ins[i]);
1941 adev = gpu_ins->adev;
1942 if (!(adev->flags & AMD_IS_APU) &&
1943 !gpu_ins->mgpu_fan_enabled &&
1944 adev->powerplay.pp_funcs &&
1945 adev->powerplay.pp_funcs->enable_mgpu_fan_boost) {
1946 ret = amdgpu_dpm_enable_mgpu_fan_boost(adev);
1950 gpu_ins->mgpu_fan_enabled = 1;
1955 mutex_unlock(&mgpu_info.mutex);
1961 * amdgpu_device_ip_late_init_func_handler - work handler for ib test
1963 * @work: work_struct.
1965 static void amdgpu_device_ip_late_init_func_handler(struct work_struct *work)
1967 struct amdgpu_device *adev =
1968 container_of(work, struct amdgpu_device, late_init_work.work);
1971 r = amdgpu_ib_ring_tests(adev);
1973 DRM_ERROR("ib ring test failed (%d).\n", r);
1975 r = amdgpu_device_enable_mgpu_fan_boost();
1977 DRM_ERROR("enable mgpu fan boost failed (%d).\n", r);
1980 static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work)
1982 struct amdgpu_device *adev =
1983 container_of(work, struct amdgpu_device, gfx.gfx_off_delay_work.work);
1985 mutex_lock(&adev->gfx.gfx_off_mutex);
1986 if (!adev->gfx.gfx_off_state && !adev->gfx.gfx_off_req_count) {
1987 if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, true))
1988 adev->gfx.gfx_off_state = true;
1990 mutex_unlock(&adev->gfx.gfx_off_mutex);
1994 * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1)
1996 * @adev: amdgpu_device pointer
1998 * Main suspend function for hardware IPs. The list of all the hardware
1999 * IPs that make up the asic is walked, clockgating is disabled and the
2000 * suspend callbacks are run. suspend puts the hardware and software state
2001 * in each IP into a state suitable for suspend.
2002 * Returns 0 on success, negative error code on failure.
2004 static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev)
2008 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
2009 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
2011 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2012 if (!adev->ip_blocks[i].status.valid)
2014 /* displays are handled separately */
2015 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) {
2016 /* XXX handle errors */
2017 r = adev->ip_blocks[i].version->funcs->suspend(adev);
2018 /* XXX handle errors */
2020 DRM_ERROR("suspend of IP block <%s> failed %d\n",
2021 adev->ip_blocks[i].version->funcs->name, r);
2030 * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2)
2032 * @adev: amdgpu_device pointer
2034 * Main suspend function for hardware IPs. The list of all the hardware
2035 * IPs that make up the asic is walked, clockgating is disabled and the
2036 * suspend callbacks are run. suspend puts the hardware and software state
2037 * in each IP into a state suitable for suspend.
2038 * Returns 0 on success, negative error code on failure.
2040 static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev)
2044 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2045 if (!adev->ip_blocks[i].status.valid)
2047 /* displays are handled in phase1 */
2048 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE)
2050 /* XXX handle errors */
2051 r = adev->ip_blocks[i].version->funcs->suspend(adev);
2052 /* XXX handle errors */
2054 DRM_ERROR("suspend of IP block <%s> failed %d\n",
2055 adev->ip_blocks[i].version->funcs->name, r);
2063 * amdgpu_device_ip_suspend - run suspend for hardware IPs
2065 * @adev: amdgpu_device pointer
2067 * Main suspend function for hardware IPs. The list of all the hardware
2068 * IPs that make up the asic is walked, clockgating is disabled and the
2069 * suspend callbacks are run. suspend puts the hardware and software state
2070 * in each IP into a state suitable for suspend.
2071 * Returns 0 on success, negative error code on failure.
2073 int amdgpu_device_ip_suspend(struct amdgpu_device *adev)
2077 if (amdgpu_sriov_vf(adev))
2078 amdgpu_virt_request_full_gpu(adev, false);
2080 r = amdgpu_device_ip_suspend_phase1(adev);
2083 r = amdgpu_device_ip_suspend_phase2(adev);
2085 if (amdgpu_sriov_vf(adev))
2086 amdgpu_virt_release_full_gpu(adev, false);
2091 static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
2095 static enum amd_ip_block_type ip_order[] = {
2096 AMD_IP_BLOCK_TYPE_GMC,
2097 AMD_IP_BLOCK_TYPE_COMMON,
2098 AMD_IP_BLOCK_TYPE_PSP,
2099 AMD_IP_BLOCK_TYPE_IH,
2102 for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
2104 struct amdgpu_ip_block *block;
2106 for (j = 0; j < adev->num_ip_blocks; j++) {
2107 block = &adev->ip_blocks[j];
2109 if (block->version->type != ip_order[i] ||
2110 !block->status.valid)
2113 r = block->version->funcs->hw_init(adev);
2114 DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
2123 static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
2127 static enum amd_ip_block_type ip_order[] = {
2128 AMD_IP_BLOCK_TYPE_SMC,
2129 AMD_IP_BLOCK_TYPE_DCE,
2130 AMD_IP_BLOCK_TYPE_GFX,
2131 AMD_IP_BLOCK_TYPE_SDMA,
2132 AMD_IP_BLOCK_TYPE_UVD,
2133 AMD_IP_BLOCK_TYPE_VCE
2136 for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
2138 struct amdgpu_ip_block *block;
2140 for (j = 0; j < adev->num_ip_blocks; j++) {
2141 block = &adev->ip_blocks[j];
2143 if (block->version->type != ip_order[i] ||
2144 !block->status.valid)
2147 r = block->version->funcs->hw_init(adev);
2148 DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
2158 * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs
2160 * @adev: amdgpu_device pointer
2162 * First resume function for hardware IPs. The list of all the hardware
2163 * IPs that make up the asic is walked and the resume callbacks are run for
2164 * COMMON, GMC, and IH. resume puts the hardware into a functional state
2165 * after a suspend and updates the software state as necessary. This
2166 * function is also used for restoring the GPU after a GPU reset.
2167 * Returns 0 on success, negative error code on failure.
2169 static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)
2173 for (i = 0; i < adev->num_ip_blocks; i++) {
2174 if (!adev->ip_blocks[i].status.valid)
2176 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2177 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2178 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
2179 r = adev->ip_blocks[i].version->funcs->resume(adev);
2181 DRM_ERROR("resume of IP block <%s> failed %d\n",
2182 adev->ip_blocks[i].version->funcs->name, r);
2192 * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs
2194 * @adev: amdgpu_device pointer
2196 * First resume function for hardware IPs. The list of all the hardware
2197 * IPs that make up the asic is walked and the resume callbacks are run for
2198 * all blocks except COMMON, GMC, and IH. resume puts the hardware into a
2199 * functional state after a suspend and updates the software state as
2200 * necessary. This function is also used for restoring the GPU after a GPU
2202 * Returns 0 on success, negative error code on failure.
2204 static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)
2208 for (i = 0; i < adev->num_ip_blocks; i++) {
2209 if (!adev->ip_blocks[i].status.valid)
2211 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2212 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2213 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH ||
2214 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)
2216 r = adev->ip_blocks[i].version->funcs->resume(adev);
2218 DRM_ERROR("resume of IP block <%s> failed %d\n",
2219 adev->ip_blocks[i].version->funcs->name, r);
2228 * amdgpu_device_ip_resume - run resume for hardware IPs
2230 * @adev: amdgpu_device pointer
2232 * Main resume function for hardware IPs. The hardware IPs
2233 * are split into two resume functions because they are
2234 * are also used in in recovering from a GPU reset and some additional
2235 * steps need to be take between them. In this case (S3/S4) they are
2237 * Returns 0 on success, negative error code on failure.
2239 static int amdgpu_device_ip_resume(struct amdgpu_device *adev)
2243 r = amdgpu_device_ip_resume_phase1(adev);
2247 r = amdgpu_device_fw_loading(adev);
2251 r = amdgpu_device_ip_resume_phase2(adev);
2257 * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV
2259 * @adev: amdgpu_device pointer
2261 * Query the VBIOS data tables to determine if the board supports SR-IOV.
2263 static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
2265 if (amdgpu_sriov_vf(adev)) {
2266 if (adev->is_atom_fw) {
2267 if (amdgpu_atomfirmware_gpu_supports_virtualization(adev))
2268 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2270 if (amdgpu_atombios_has_gpu_virtualization_table(adev))
2271 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2274 if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
2275 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
2280 * amdgpu_device_asic_has_dc_support - determine if DC supports the asic
2282 * @asic_type: AMD asic type
2284 * Check if there is DC (new modesetting infrastructre) support for an asic.
2285 * returns true if DC has support, false if not.
2287 bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
2289 switch (asic_type) {
2290 #if defined(CONFIG_DRM_AMD_DC)
2296 * We have systems in the wild with these ASICs that require
2297 * LVDS and VGA support which is not supported with DC.
2299 * Fallback to the non-DC driver here by default so as not to
2300 * cause regressions.
2302 return amdgpu_dc > 0;
2306 case CHIP_POLARIS10:
2307 case CHIP_POLARIS11:
2308 case CHIP_POLARIS12:
2315 #if defined(CONFIG_DRM_AMD_DC_DCN1_0)
2318 return amdgpu_dc != 0;
2326 * amdgpu_device_has_dc_support - check if dc is supported
2328 * @adev: amdgpu_device_pointer
2330 * Returns true for supported, false for not supported
2332 bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
2334 if (amdgpu_sriov_vf(adev))
2337 return amdgpu_device_asic_has_dc_support(adev->asic_type);
2341 * amdgpu_device_init - initialize the driver
2343 * @adev: amdgpu_device pointer
2344 * @ddev: drm dev pointer
2345 * @pdev: pci dev pointer
2346 * @flags: driver flags
2348 * Initializes the driver info and hw (all asics).
2349 * Returns 0 for success or an error on failure.
2350 * Called at driver startup.
2352 int amdgpu_device_init(struct amdgpu_device *adev,
2353 struct drm_device *ddev,
2354 struct pci_dev *pdev,
2358 bool runtime = false;
2361 adev->shutdown = false;
2362 adev->dev = &pdev->dev;
2365 adev->flags = flags;
2366 adev->asic_type = flags & AMD_ASIC_MASK;
2367 adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
2368 if (amdgpu_emu_mode == 1)
2369 adev->usec_timeout *= 2;
2370 adev->gmc.gart_size = 512 * 1024 * 1024;
2371 adev->accel_working = false;
2372 adev->num_rings = 0;
2373 adev->mman.buffer_funcs = NULL;
2374 adev->mman.buffer_funcs_ring = NULL;
2375 adev->vm_manager.vm_pte_funcs = NULL;
2376 adev->vm_manager.vm_pte_num_rqs = 0;
2377 adev->gmc.gmc_funcs = NULL;
2378 adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
2379 bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
2381 adev->smc_rreg = &amdgpu_invalid_rreg;
2382 adev->smc_wreg = &amdgpu_invalid_wreg;
2383 adev->pcie_rreg = &amdgpu_invalid_rreg;
2384 adev->pcie_wreg = &amdgpu_invalid_wreg;
2385 adev->pciep_rreg = &amdgpu_invalid_rreg;
2386 adev->pciep_wreg = &amdgpu_invalid_wreg;
2387 adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
2388 adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
2389 adev->didt_rreg = &amdgpu_invalid_rreg;
2390 adev->didt_wreg = &amdgpu_invalid_wreg;
2391 adev->gc_cac_rreg = &amdgpu_invalid_rreg;
2392 adev->gc_cac_wreg = &amdgpu_invalid_wreg;
2393 adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
2394 adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
2396 DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
2397 amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
2398 pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
2400 /* mutex initialization are all done here so we
2401 * can recall function without having locking issues */
2402 atomic_set(&adev->irq.ih.lock, 0);
2403 mutex_init(&adev->firmware.mutex);
2404 mutex_init(&adev->pm.mutex);
2405 mutex_init(&adev->gfx.gpu_clock_mutex);
2406 mutex_init(&adev->srbm_mutex);
2407 mutex_init(&adev->gfx.pipe_reserve_mutex);
2408 mutex_init(&adev->gfx.gfx_off_mutex);
2409 mutex_init(&adev->grbm_idx_mutex);
2410 mutex_init(&adev->mn_lock);
2411 mutex_init(&adev->virt.vf_errors.lock);
2412 hash_init(adev->mn_hash);
2413 mutex_init(&adev->lock_reset);
2415 amdgpu_device_check_arguments(adev);
2417 spin_lock_init(&adev->mmio_idx_lock);
2418 spin_lock_init(&adev->smc_idx_lock);
2419 spin_lock_init(&adev->pcie_idx_lock);
2420 spin_lock_init(&adev->uvd_ctx_idx_lock);
2421 spin_lock_init(&adev->didt_idx_lock);
2422 spin_lock_init(&adev->gc_cac_idx_lock);
2423 spin_lock_init(&adev->se_cac_idx_lock);
2424 spin_lock_init(&adev->audio_endpt_idx_lock);
2425 spin_lock_init(&adev->mm_stats.lock);
2427 INIT_LIST_HEAD(&adev->shadow_list);
2428 mutex_init(&adev->shadow_list_lock);
2430 INIT_LIST_HEAD(&adev->ring_lru_list);
2431 spin_lock_init(&adev->ring_lru_list_lock);
2433 INIT_DELAYED_WORK(&adev->late_init_work,
2434 amdgpu_device_ip_late_init_func_handler);
2435 INIT_DELAYED_WORK(&adev->gfx.gfx_off_delay_work,
2436 amdgpu_device_delay_enable_gfx_off);
2438 adev->gfx.gfx_off_req_count = 1;
2439 adev->pm.ac_power = power_supply_is_system_supplied() > 0 ? true : false;
2441 /* Registers mapping */
2442 /* TODO: block userspace mapping of io register */
2443 if (adev->asic_type >= CHIP_BONAIRE) {
2444 adev->rmmio_base = pci_resource_start(adev->pdev, 5);
2445 adev->rmmio_size = pci_resource_len(adev->pdev, 5);
2447 adev->rmmio_base = pci_resource_start(adev->pdev, 2);
2448 adev->rmmio_size = pci_resource_len(adev->pdev, 2);
2451 adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
2452 if (adev->rmmio == NULL) {
2455 DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
2456 DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
2458 /* doorbell bar mapping */
2459 amdgpu_device_doorbell_init(adev);
2461 /* io port mapping */
2462 for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
2463 if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) {
2464 adev->rio_mem_size = pci_resource_len(adev->pdev, i);
2465 adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size);
2469 if (adev->rio_mem == NULL)
2470 DRM_INFO("PCI I/O BAR is not found.\n");
2472 amdgpu_device_get_pcie_info(adev);
2474 /* early init functions */
2475 r = amdgpu_device_ip_early_init(adev);
2479 /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
2480 /* this will fail for cards that aren't VGA class devices, just
2482 vga_client_register(adev->pdev, adev, NULL, amdgpu_device_vga_set_decode);
2484 if (amdgpu_device_is_px(ddev))
2486 if (!pci_is_thunderbolt_attached(adev->pdev))
2487 vga_switcheroo_register_client(adev->pdev,
2488 &amdgpu_switcheroo_ops, runtime);
2490 vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
2492 if (amdgpu_emu_mode == 1) {
2493 /* post the asic on emulation mode */
2494 emu_soc_asic_init(adev);
2495 goto fence_driver_init;
2499 if (!amdgpu_get_bios(adev)) {
2504 r = amdgpu_atombios_init(adev);
2506 dev_err(adev->dev, "amdgpu_atombios_init failed\n");
2507 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0);
2511 /* detect if we are with an SRIOV vbios */
2512 amdgpu_device_detect_sriov_bios(adev);
2514 /* Post card if necessary */
2515 if (amdgpu_device_need_post(adev)) {
2517 dev_err(adev->dev, "no vBIOS found\n");
2521 DRM_INFO("GPU posting now...\n");
2522 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2524 dev_err(adev->dev, "gpu post error!\n");
2529 if (adev->is_atom_fw) {
2530 /* Initialize clocks */
2531 r = amdgpu_atomfirmware_get_clock_info(adev);
2533 dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
2534 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
2538 /* Initialize clocks */
2539 r = amdgpu_atombios_get_clock_info(adev);
2541 dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
2542 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
2545 /* init i2c buses */
2546 if (!amdgpu_device_has_dc_support(adev))
2547 amdgpu_atombios_i2c_init(adev);
2552 r = amdgpu_fence_driver_init(adev);
2554 dev_err(adev->dev, "amdgpu_fence_driver_init failed\n");
2555 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
2559 /* init the mode config */
2560 drm_mode_config_init(adev->ddev);
2562 r = amdgpu_device_ip_init(adev);
2564 /* failed in exclusive mode due to timeout */
2565 if (amdgpu_sriov_vf(adev) &&
2566 !amdgpu_sriov_runtime(adev) &&
2567 amdgpu_virt_mmio_blocked(adev) &&
2568 !amdgpu_virt_wait_reset(adev)) {
2569 dev_err(adev->dev, "VF exclusive mode timeout\n");
2570 /* Don't send request since VF is inactive. */
2571 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
2572 adev->virt.ops = NULL;
2576 dev_err(adev->dev, "amdgpu_device_ip_init failed\n");
2577 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
2581 adev->accel_working = true;
2583 amdgpu_vm_check_compute_bug(adev);
2585 /* Initialize the buffer migration limit. */
2586 if (amdgpu_moverate >= 0)
2587 max_MBps = amdgpu_moverate;
2589 max_MBps = 8; /* Allow 8 MB/s. */
2590 /* Get a log2 for easy divisions. */
2591 adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
2593 r = amdgpu_ib_pool_init(adev);
2595 dev_err(adev->dev, "IB initialization failed (%d).\n", r);
2596 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r);
2600 if (amdgpu_sriov_vf(adev))
2601 amdgpu_virt_init_data_exchange(adev);
2603 amdgpu_fbdev_init(adev);
2605 r = amdgpu_pm_sysfs_init(adev);
2607 DRM_ERROR("registering pm debugfs failed (%d).\n", r);
2609 r = amdgpu_debugfs_gem_init(adev);
2611 DRM_ERROR("registering gem debugfs failed (%d).\n", r);
2613 r = amdgpu_debugfs_regs_init(adev);
2615 DRM_ERROR("registering register debugfs failed (%d).\n", r);
2617 r = amdgpu_debugfs_firmware_init(adev);
2619 DRM_ERROR("registering firmware debugfs failed (%d).\n", r);
2621 r = amdgpu_debugfs_init(adev);
2623 DRM_ERROR("Creating debugfs files failed (%d).\n", r);
2625 if ((amdgpu_testing & 1)) {
2626 if (adev->accel_working)
2627 amdgpu_test_moves(adev);
2629 DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n");
2631 if (amdgpu_benchmarking) {
2632 if (adev->accel_working)
2633 amdgpu_benchmark(adev, amdgpu_benchmarking);
2635 DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n");
2638 /* enable clockgating, etc. after ib tests, etc. since some blocks require
2639 * explicit gating rather than handling it automatically.
2641 r = amdgpu_device_ip_late_init(adev);
2643 dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n");
2644 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
2651 amdgpu_vf_error_trans_all(adev);
2653 vga_switcheroo_fini_domain_pm_ops(adev->dev);
2659 * amdgpu_device_fini - tear down the driver
2661 * @adev: amdgpu_device pointer
2663 * Tear down the driver info (all asics).
2664 * Called at driver shutdown.
2666 void amdgpu_device_fini(struct amdgpu_device *adev)
2670 DRM_INFO("amdgpu: finishing device.\n");
2671 adev->shutdown = true;
2672 /* disable all interrupts */
2673 amdgpu_irq_disable_all(adev);
2674 if (adev->mode_info.mode_config_initialized){
2675 if (!amdgpu_device_has_dc_support(adev))
2676 drm_crtc_force_disable_all(adev->ddev);
2678 drm_atomic_helper_shutdown(adev->ddev);
2680 amdgpu_ib_pool_fini(adev);
2681 amdgpu_fence_driver_fini(adev);
2682 amdgpu_pm_sysfs_fini(adev);
2683 amdgpu_fbdev_fini(adev);
2684 r = amdgpu_device_ip_fini(adev);
2685 if (adev->firmware.gpu_info_fw) {
2686 release_firmware(adev->firmware.gpu_info_fw);
2687 adev->firmware.gpu_info_fw = NULL;
2689 adev->accel_working = false;
2690 cancel_delayed_work_sync(&adev->late_init_work);
2691 /* free i2c buses */
2692 if (!amdgpu_device_has_dc_support(adev))
2693 amdgpu_i2c_fini(adev);
2695 if (amdgpu_emu_mode != 1)
2696 amdgpu_atombios_fini(adev);
2700 if (!pci_is_thunderbolt_attached(adev->pdev))
2701 vga_switcheroo_unregister_client(adev->pdev);
2702 if (adev->flags & AMD_IS_PX)
2703 vga_switcheroo_fini_domain_pm_ops(adev->dev);
2704 vga_client_register(adev->pdev, NULL, NULL, NULL);
2706 pci_iounmap(adev->pdev, adev->rio_mem);
2707 adev->rio_mem = NULL;
2708 iounmap(adev->rmmio);
2710 amdgpu_device_doorbell_fini(adev);
2711 amdgpu_debugfs_regs_cleanup(adev);
2719 * amdgpu_device_suspend - initiate device suspend
2721 * @dev: drm dev pointer
2722 * @suspend: suspend state
2723 * @fbcon : notify the fbdev of suspend
2725 * Puts the hw in the suspend state (all asics).
2726 * Returns 0 for success or an error on failure.
2727 * Called at driver suspend.
2729 int amdgpu_device_suspend(struct drm_device *dev, bool suspend, bool fbcon)
2731 struct amdgpu_device *adev;
2732 struct drm_crtc *crtc;
2733 struct drm_connector *connector;
2736 if (dev == NULL || dev->dev_private == NULL) {
2740 adev = dev->dev_private;
2742 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
2745 adev->in_suspend = true;
2746 drm_kms_helper_poll_disable(dev);
2749 amdgpu_fbdev_set_suspend(adev, 1);
2751 cancel_delayed_work_sync(&adev->late_init_work);
2753 if (!amdgpu_device_has_dc_support(adev)) {
2754 /* turn off display hw */
2755 drm_modeset_lock_all(dev);
2756 list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2757 drm_helper_connector_dpms(connector, DRM_MODE_DPMS_OFF);
2759 drm_modeset_unlock_all(dev);
2760 /* unpin the front buffers and cursors */
2761 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2762 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2763 struct drm_framebuffer *fb = crtc->primary->fb;
2764 struct amdgpu_bo *robj;
2766 if (amdgpu_crtc->cursor_bo) {
2767 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2768 r = amdgpu_bo_reserve(aobj, true);
2770 amdgpu_bo_unpin(aobj);
2771 amdgpu_bo_unreserve(aobj);
2775 if (fb == NULL || fb->obj[0] == NULL) {
2778 robj = gem_to_amdgpu_bo(fb->obj[0]);
2779 /* don't unpin kernel fb objects */
2780 if (!amdgpu_fbdev_robj_is_fb(adev, robj)) {
2781 r = amdgpu_bo_reserve(robj, true);
2783 amdgpu_bo_unpin(robj);
2784 amdgpu_bo_unreserve(robj);
2790 amdgpu_amdkfd_suspend(adev);
2792 r = amdgpu_device_ip_suspend_phase1(adev);
2794 /* evict vram memory */
2795 amdgpu_bo_evict_vram(adev);
2797 amdgpu_fence_driver_suspend(adev);
2799 r = amdgpu_device_ip_suspend_phase2(adev);
2801 /* evict remaining vram memory
2802 * This second call to evict vram is to evict the gart page table
2805 amdgpu_bo_evict_vram(adev);
2807 pci_save_state(dev->pdev);
2809 /* Shut down the device */
2810 pci_disable_device(dev->pdev);
2811 pci_set_power_state(dev->pdev, PCI_D3hot);
2813 r = amdgpu_asic_reset(adev);
2815 DRM_ERROR("amdgpu asic reset failed\n");
2822 * amdgpu_device_resume - initiate device resume
2824 * @dev: drm dev pointer
2825 * @resume: resume state
2826 * @fbcon : notify the fbdev of resume
2828 * Bring the hw back to operating state (all asics).
2829 * Returns 0 for success or an error on failure.
2830 * Called at driver resume.
2832 int amdgpu_device_resume(struct drm_device *dev, bool resume, bool fbcon)
2834 struct drm_connector *connector;
2835 struct amdgpu_device *adev = dev->dev_private;
2836 struct drm_crtc *crtc;
2839 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
2843 pci_set_power_state(dev->pdev, PCI_D0);
2844 pci_restore_state(dev->pdev);
2845 r = pci_enable_device(dev->pdev);
2851 if (amdgpu_device_need_post(adev)) {
2852 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2854 DRM_ERROR("amdgpu asic init failed\n");
2857 r = amdgpu_device_ip_resume(adev);
2859 DRM_ERROR("amdgpu_device_ip_resume failed (%d).\n", r);
2862 amdgpu_fence_driver_resume(adev);
2865 r = amdgpu_device_ip_late_init(adev);
2869 if (!amdgpu_device_has_dc_support(adev)) {
2871 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2872 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2874 if (amdgpu_crtc->cursor_bo) {
2875 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2876 r = amdgpu_bo_reserve(aobj, true);
2878 r = amdgpu_bo_pin(aobj, AMDGPU_GEM_DOMAIN_VRAM);
2880 DRM_ERROR("Failed to pin cursor BO (%d)\n", r);
2881 amdgpu_crtc->cursor_addr = amdgpu_bo_gpu_offset(aobj);
2882 amdgpu_bo_unreserve(aobj);
2887 r = amdgpu_amdkfd_resume(adev);
2891 /* Make sure IB tests flushed */
2892 flush_delayed_work(&adev->late_init_work);
2894 /* blat the mode back in */
2896 if (!amdgpu_device_has_dc_support(adev)) {
2898 drm_helper_resume_force_mode(dev);
2900 /* turn on display hw */
2901 drm_modeset_lock_all(dev);
2902 list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2903 drm_helper_connector_dpms(connector, DRM_MODE_DPMS_ON);
2905 drm_modeset_unlock_all(dev);
2907 amdgpu_fbdev_set_suspend(adev, 0);
2910 drm_kms_helper_poll_enable(dev);
2913 * Most of the connector probing functions try to acquire runtime pm
2914 * refs to ensure that the GPU is powered on when connector polling is
2915 * performed. Since we're calling this from a runtime PM callback,
2916 * trying to acquire rpm refs will cause us to deadlock.
2918 * Since we're guaranteed to be holding the rpm lock, it's safe to
2919 * temporarily disable the rpm helpers so this doesn't deadlock us.
2922 dev->dev->power.disable_depth++;
2924 if (!amdgpu_device_has_dc_support(adev))
2925 drm_helper_hpd_irq_event(dev);
2927 drm_kms_helper_hotplug_event(dev);
2929 dev->dev->power.disable_depth--;
2931 adev->in_suspend = false;
2937 * amdgpu_device_ip_check_soft_reset - did soft reset succeed
2939 * @adev: amdgpu_device pointer
2941 * The list of all the hardware IPs that make up the asic is walked and
2942 * the check_soft_reset callbacks are run. check_soft_reset determines
2943 * if the asic is still hung or not.
2944 * Returns true if any of the IPs are still in a hung state, false if not.
2946 static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)
2949 bool asic_hang = false;
2951 if (amdgpu_sriov_vf(adev))
2954 if (amdgpu_asic_need_full_reset(adev))
2957 for (i = 0; i < adev->num_ip_blocks; i++) {
2958 if (!adev->ip_blocks[i].status.valid)
2960 if (adev->ip_blocks[i].version->funcs->check_soft_reset)
2961 adev->ip_blocks[i].status.hang =
2962 adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
2963 if (adev->ip_blocks[i].status.hang) {
2964 DRM_INFO("IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
2972 * amdgpu_device_ip_pre_soft_reset - prepare for soft reset
2974 * @adev: amdgpu_device pointer
2976 * The list of all the hardware IPs that make up the asic is walked and the
2977 * pre_soft_reset callbacks are run if the block is hung. pre_soft_reset
2978 * handles any IP specific hardware or software state changes that are
2979 * necessary for a soft reset to succeed.
2980 * Returns 0 on success, negative error code on failure.
2982 static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)
2986 for (i = 0; i < adev->num_ip_blocks; i++) {
2987 if (!adev->ip_blocks[i].status.valid)
2989 if (adev->ip_blocks[i].status.hang &&
2990 adev->ip_blocks[i].version->funcs->pre_soft_reset) {
2991 r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
3001 * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed
3003 * @adev: amdgpu_device pointer
3005 * Some hardware IPs cannot be soft reset. If they are hung, a full gpu
3006 * reset is necessary to recover.
3007 * Returns true if a full asic reset is required, false if not.
3009 static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)
3013 if (amdgpu_asic_need_full_reset(adev))
3016 for (i = 0; i < adev->num_ip_blocks; i++) {
3017 if (!adev->ip_blocks[i].status.valid)
3019 if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
3020 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
3021 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
3022 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
3023 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
3024 if (adev->ip_blocks[i].status.hang) {
3025 DRM_INFO("Some block need full reset!\n");
3034 * amdgpu_device_ip_soft_reset - do a soft reset
3036 * @adev: amdgpu_device pointer
3038 * The list of all the hardware IPs that make up the asic is walked and the
3039 * soft_reset callbacks are run if the block is hung. soft_reset handles any
3040 * IP specific hardware or software state changes that are necessary to soft
3042 * Returns 0 on success, negative error code on failure.
3044 static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)
3048 for (i = 0; i < adev->num_ip_blocks; i++) {
3049 if (!adev->ip_blocks[i].status.valid)
3051 if (adev->ip_blocks[i].status.hang &&
3052 adev->ip_blocks[i].version->funcs->soft_reset) {
3053 r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
3063 * amdgpu_device_ip_post_soft_reset - clean up from soft reset
3065 * @adev: amdgpu_device pointer
3067 * The list of all the hardware IPs that make up the asic is walked and the
3068 * post_soft_reset callbacks are run if the asic was hung. post_soft_reset
3069 * handles any IP specific hardware or software state changes that are
3070 * necessary after the IP has been soft reset.
3071 * Returns 0 on success, negative error code on failure.
3073 static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)
3077 for (i = 0; i < adev->num_ip_blocks; i++) {
3078 if (!adev->ip_blocks[i].status.valid)
3080 if (adev->ip_blocks[i].status.hang &&
3081 adev->ip_blocks[i].version->funcs->post_soft_reset)
3082 r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
3091 * amdgpu_device_recover_vram - Recover some VRAM contents
3093 * @adev: amdgpu_device pointer
3095 * Restores the contents of VRAM buffers from the shadows in GTT. Used to
3096 * restore things like GPUVM page tables after a GPU reset where
3097 * the contents of VRAM might be lost.
3100 * 0 on success, negative error code on failure.
3102 static int amdgpu_device_recover_vram(struct amdgpu_device *adev)
3104 struct dma_fence *fence = NULL, *next = NULL;
3105 struct amdgpu_bo *shadow;
3108 if (amdgpu_sriov_runtime(adev))
3109 tmo = msecs_to_jiffies(8000);
3111 tmo = msecs_to_jiffies(100);
3113 DRM_INFO("recover vram bo from shadow start\n");
3114 mutex_lock(&adev->shadow_list_lock);
3115 list_for_each_entry(shadow, &adev->shadow_list, shadow_list) {
3117 /* No need to recover an evicted BO */
3118 if (shadow->tbo.mem.mem_type != TTM_PL_TT ||
3119 shadow->parent->tbo.mem.mem_type != TTM_PL_VRAM)
3122 r = amdgpu_bo_restore_shadow(shadow, &next);
3127 r = dma_fence_wait_timeout(fence, false, tmo);
3128 dma_fence_put(fence);
3136 mutex_unlock(&adev->shadow_list_lock);
3139 tmo = dma_fence_wait_timeout(fence, false, tmo);
3140 dma_fence_put(fence);
3142 if (r <= 0 || tmo <= 0) {
3143 DRM_ERROR("recover vram bo from shadow failed\n");
3147 DRM_INFO("recover vram bo from shadow done\n");
3152 * amdgpu_device_reset - reset ASIC/GPU for bare-metal or passthrough
3154 * @adev: amdgpu device pointer
3156 * attempt to do soft-reset or full-reset and reinitialize Asic
3157 * return 0 means succeeded otherwise failed
3159 static int amdgpu_device_reset(struct amdgpu_device *adev)
3161 bool need_full_reset, vram_lost = 0;
3164 need_full_reset = amdgpu_device_ip_need_full_reset(adev);
3166 if (!need_full_reset) {
3167 amdgpu_device_ip_pre_soft_reset(adev);
3168 r = amdgpu_device_ip_soft_reset(adev);
3169 amdgpu_device_ip_post_soft_reset(adev);
3170 if (r || amdgpu_device_ip_check_soft_reset(adev)) {
3171 DRM_INFO("soft reset failed, will fallback to full reset!\n");
3172 need_full_reset = true;
3176 if (need_full_reset) {
3177 r = amdgpu_device_ip_suspend(adev);
3180 r = amdgpu_asic_reset(adev);
3182 amdgpu_atom_asic_init(adev->mode_info.atom_context);
3185 dev_info(adev->dev, "GPU reset succeeded, trying to resume\n");
3186 r = amdgpu_device_ip_resume_phase1(adev);
3190 vram_lost = amdgpu_device_check_vram_lost(adev);
3192 DRM_ERROR("VRAM is lost!\n");
3193 atomic_inc(&adev->vram_lost_counter);
3196 r = amdgpu_gtt_mgr_recover(
3197 &adev->mman.bdev.man[TTM_PL_TT]);
3201 r = amdgpu_device_fw_loading(adev);
3205 r = amdgpu_device_ip_resume_phase2(adev);
3210 amdgpu_device_fill_reset_magic(adev);
3216 amdgpu_irq_gpu_reset_resume_helper(adev);
3217 r = amdgpu_ib_ring_tests(adev);
3219 dev_err(adev->dev, "ib ring test failed (%d).\n", r);
3220 r = amdgpu_device_ip_suspend(adev);
3221 need_full_reset = true;
3227 r = amdgpu_device_recover_vram(adev);
3233 * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
3235 * @adev: amdgpu device pointer
3236 * @from_hypervisor: request from hypervisor
3238 * do VF FLR and reinitialize Asic
3239 * return 0 means succeeded otherwise failed
3241 static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
3242 bool from_hypervisor)
3246 if (from_hypervisor)
3247 r = amdgpu_virt_request_full_gpu(adev, true);
3249 r = amdgpu_virt_reset_gpu(adev);
3253 /* Resume IP prior to SMC */
3254 r = amdgpu_device_ip_reinit_early_sriov(adev);
3258 /* we need recover gart prior to run SMC/CP/SDMA resume */
3259 amdgpu_gtt_mgr_recover(&adev->mman.bdev.man[TTM_PL_TT]);
3261 r = amdgpu_device_fw_loading(adev);
3265 /* now we are okay to resume SMC/CP/SDMA */
3266 r = amdgpu_device_ip_reinit_late_sriov(adev);
3270 amdgpu_irq_gpu_reset_resume_helper(adev);
3271 r = amdgpu_ib_ring_tests(adev);
3274 amdgpu_virt_release_full_gpu(adev, true);
3275 if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
3276 atomic_inc(&adev->vram_lost_counter);
3277 r = amdgpu_device_recover_vram(adev);
3284 * amdgpu_device_should_recover_gpu - check if we should try GPU recovery
3286 * @adev: amdgpu device pointer
3288 * Check amdgpu_gpu_recovery and SRIOV status to see if we should try to recover
3291 bool amdgpu_device_should_recover_gpu(struct amdgpu_device *adev)
3293 if (!amdgpu_device_ip_check_soft_reset(adev)) {
3294 DRM_INFO("Timeout, but no hardware hang detected.\n");
3298 if (amdgpu_gpu_recovery == 0 || (amdgpu_gpu_recovery == -1 &&
3299 !amdgpu_sriov_vf(adev))) {
3300 DRM_INFO("GPU recovery disabled.\n");
3308 * amdgpu_device_gpu_recover - reset the asic and recover scheduler
3310 * @adev: amdgpu device pointer
3311 * @job: which job trigger hang
3313 * Attempt to reset the GPU if it has hung (all asics).
3314 * Returns 0 for success or an error on failure.
3316 int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
3317 struct amdgpu_job *job)
3321 dev_info(adev->dev, "GPU reset begin!\n");
3323 mutex_lock(&adev->lock_reset);
3324 atomic_inc(&adev->gpu_reset_counter);
3325 adev->in_gpu_reset = 1;
3328 amdgpu_amdkfd_pre_reset(adev);
3331 resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
3333 /* block all schedulers and reset given job's ring */
3334 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3335 struct amdgpu_ring *ring = adev->rings[i];
3337 if (!ring || !ring->sched.thread)
3340 kthread_park(ring->sched.thread);
3342 if (job && job->base.sched != &ring->sched)
3345 drm_sched_hw_job_reset(&ring->sched, job ? &job->base : NULL);
3347 /* after all hw jobs are reset, hw fence is meaningless, so force_completion */
3348 amdgpu_fence_driver_force_completion(ring);
3351 if (amdgpu_sriov_vf(adev))
3352 r = amdgpu_device_reset_sriov(adev, job ? false : true);
3354 r = amdgpu_device_reset(adev);
3356 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3357 struct amdgpu_ring *ring = adev->rings[i];
3359 if (!ring || !ring->sched.thread)
3362 /* only need recovery sched of the given job's ring
3363 * or all rings (in the case @job is NULL)
3364 * after above amdgpu_reset accomplished
3366 if ((!job || job->base.sched == &ring->sched) && !r)
3367 drm_sched_job_recovery(&ring->sched);
3369 kthread_unpark(ring->sched.thread);
3372 if (!amdgpu_device_has_dc_support(adev)) {
3373 drm_helper_resume_force_mode(adev->ddev);
3376 ttm_bo_unlock_delayed_workqueue(&adev->mman.bdev, resched);
3379 /* bad news, how to tell it to userspace ? */
3380 dev_info(adev->dev, "GPU reset(%d) failed\n", atomic_read(&adev->gpu_reset_counter));
3381 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r);
3383 dev_info(adev->dev, "GPU reset(%d) succeeded!\n",atomic_read(&adev->gpu_reset_counter));
3387 amdgpu_amdkfd_post_reset(adev);
3388 amdgpu_vf_error_trans_all(adev);
3389 adev->in_gpu_reset = 0;
3390 mutex_unlock(&adev->lock_reset);
3395 * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot
3397 * @adev: amdgpu_device pointer
3399 * Fetchs and stores in the driver the PCIE capabilities (gen speed
3400 * and lanes) of the slot the device is in. Handles APUs and
3401 * virtualized environments where PCIE config space may not be available.
3403 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev)
3405 struct pci_dev *pdev;
3406 enum pci_bus_speed speed_cap;
3407 enum pcie_link_width link_width;
3409 if (amdgpu_pcie_gen_cap)
3410 adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
3412 if (amdgpu_pcie_lane_cap)
3413 adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
3415 /* covers APUs as well */
3416 if (pci_is_root_bus(adev->pdev->bus)) {
3417 if (adev->pm.pcie_gen_mask == 0)
3418 adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
3419 if (adev->pm.pcie_mlw_mask == 0)
3420 adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
3424 if (adev->pm.pcie_gen_mask == 0) {
3427 speed_cap = pcie_get_speed_cap(pdev);
3428 if (speed_cap == PCI_SPEED_UNKNOWN) {
3429 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3430 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3431 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
3433 if (speed_cap == PCIE_SPEED_16_0GT)
3434 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3435 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3436 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
3437 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4);
3438 else if (speed_cap == PCIE_SPEED_8_0GT)
3439 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3440 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3441 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
3442 else if (speed_cap == PCIE_SPEED_5_0GT)
3443 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3444 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2);
3446 adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1;
3449 pdev = adev->ddev->pdev->bus->self;
3450 speed_cap = pcie_get_speed_cap(pdev);
3451 if (speed_cap == PCI_SPEED_UNKNOWN) {
3452 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3453 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
3455 if (speed_cap == PCIE_SPEED_16_0GT)
3456 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3457 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3458 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
3459 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4);
3460 else if (speed_cap == PCIE_SPEED_8_0GT)
3461 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3462 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3463 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3);
3464 else if (speed_cap == PCIE_SPEED_5_0GT)
3465 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3466 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
3468 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
3472 if (adev->pm.pcie_mlw_mask == 0) {
3473 pdev = adev->ddev->pdev->bus->self;
3474 link_width = pcie_get_width_cap(pdev);
3475 if (link_width == PCIE_LNK_WIDTH_UNKNOWN) {
3476 adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK;
3478 switch (link_width) {
3480 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
3481 CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
3482 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3483 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3484 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3485 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3486 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3489 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
3490 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3491 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3492 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3493 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3494 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3497 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3498 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3499 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3500 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3501 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3504 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3505 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3506 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3507 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3510 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3511 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3512 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3515 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3516 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3519 adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;