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");
66 #define AMDGPU_RESUME_MS 2000
68 static const char *amdgpu_asic_name[] = {
95 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev);
98 * amdgpu_device_is_px - Is the device is a dGPU with HG/PX power control
100 * @dev: drm_device pointer
102 * Returns true if the device is a dGPU with HG/PX power control,
103 * otherwise return false.
105 bool amdgpu_device_is_px(struct drm_device *dev)
107 struct amdgpu_device *adev = dev->dev_private;
109 if (adev->flags & AMD_IS_PX)
115 * MMIO register access helper functions.
118 * amdgpu_mm_rreg - read a memory mapped IO register
120 * @adev: amdgpu_device pointer
121 * @reg: dword aligned register offset
122 * @acc_flags: access flags which require special behavior
124 * Returns the 32 bit value from the offset specified.
126 uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg,
131 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
132 return amdgpu_virt_kiq_rreg(adev, reg);
134 if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
135 ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
139 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
140 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
141 ret = readl(((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
142 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
144 trace_amdgpu_mm_rreg(adev->pdev->device, reg, ret);
149 * MMIO register read with bytes helper functions
150 * @offset:bytes offset from MMIO start
155 * amdgpu_mm_rreg8 - read a memory mapped IO register
157 * @adev: amdgpu_device pointer
158 * @offset: byte aligned register offset
160 * Returns the 8 bit value from the offset specified.
162 uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset) {
163 if (offset < adev->rmmio_size)
164 return (readb(adev->rmmio + offset));
169 * MMIO register write with bytes helper functions
170 * @offset:bytes offset from MMIO start
171 * @value: the value want to be written to the register
175 * amdgpu_mm_wreg8 - read a memory mapped IO register
177 * @adev: amdgpu_device pointer
178 * @offset: byte aligned register offset
179 * @value: 8 bit value to write
181 * Writes the value specified to the offset specified.
183 void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value) {
184 if (offset < adev->rmmio_size)
185 writeb(value, adev->rmmio + offset);
191 * amdgpu_mm_wreg - write to a memory mapped IO register
193 * @adev: amdgpu_device pointer
194 * @reg: dword aligned register offset
195 * @v: 32 bit value to write to the register
196 * @acc_flags: access flags which require special behavior
198 * Writes the value specified to the offset specified.
200 void amdgpu_mm_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v,
203 trace_amdgpu_mm_wreg(adev->pdev->device, reg, v);
205 if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
206 adev->last_mm_index = v;
209 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
210 return amdgpu_virt_kiq_wreg(adev, reg, v);
212 if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
213 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
217 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
218 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
219 writel(v, ((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
220 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
223 if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
229 * amdgpu_io_rreg - read an IO register
231 * @adev: amdgpu_device pointer
232 * @reg: dword aligned register offset
234 * Returns the 32 bit value from the offset specified.
236 u32 amdgpu_io_rreg(struct amdgpu_device *adev, u32 reg)
238 if ((reg * 4) < adev->rio_mem_size)
239 return ioread32(adev->rio_mem + (reg * 4));
241 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
242 return ioread32(adev->rio_mem + (mmMM_DATA * 4));
247 * amdgpu_io_wreg - write to an IO register
249 * @adev: amdgpu_device pointer
250 * @reg: dword aligned register offset
251 * @v: 32 bit value to write to the register
253 * Writes the value specified to the offset specified.
255 void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32 v)
257 if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
258 adev->last_mm_index = v;
261 if ((reg * 4) < adev->rio_mem_size)
262 iowrite32(v, adev->rio_mem + (reg * 4));
264 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
265 iowrite32(v, adev->rio_mem + (mmMM_DATA * 4));
268 if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
274 * amdgpu_mm_rdoorbell - read a doorbell dword
276 * @adev: amdgpu_device pointer
277 * @index: doorbell index
279 * Returns the value in the doorbell aperture at the
280 * requested doorbell index (CIK).
282 u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
284 if (index < adev->doorbell.num_doorbells) {
285 return readl(adev->doorbell.ptr + index);
287 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
293 * amdgpu_mm_wdoorbell - write a doorbell dword
295 * @adev: amdgpu_device pointer
296 * @index: doorbell index
299 * Writes @v to the doorbell aperture at the
300 * requested doorbell index (CIK).
302 void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v)
304 if (index < adev->doorbell.num_doorbells) {
305 writel(v, adev->doorbell.ptr + index);
307 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
312 * amdgpu_mm_rdoorbell64 - read a doorbell Qword
314 * @adev: amdgpu_device pointer
315 * @index: doorbell index
317 * Returns the value in the doorbell aperture at the
318 * requested doorbell index (VEGA10+).
320 u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index)
322 if (index < adev->doorbell.num_doorbells) {
323 return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index));
325 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
331 * amdgpu_mm_wdoorbell64 - write a doorbell Qword
333 * @adev: amdgpu_device pointer
334 * @index: doorbell index
337 * Writes @v to the doorbell aperture at the
338 * requested doorbell index (VEGA10+).
340 void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v)
342 if (index < adev->doorbell.num_doorbells) {
343 atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v);
345 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
350 * amdgpu_invalid_rreg - dummy reg read function
352 * @adev: amdgpu device pointer
353 * @reg: offset of register
355 * Dummy register read function. Used for register blocks
356 * that certain asics don't have (all asics).
357 * Returns the value in the register.
359 static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
361 DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
367 * amdgpu_invalid_wreg - dummy reg write function
369 * @adev: amdgpu device pointer
370 * @reg: offset of register
371 * @v: value to write to the register
373 * Dummy register read function. Used for register blocks
374 * that certain asics don't have (all asics).
376 static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
378 DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
384 * amdgpu_block_invalid_rreg - dummy reg read function
386 * @adev: amdgpu device pointer
387 * @block: offset of instance
388 * @reg: offset of register
390 * Dummy register read function. Used for register blocks
391 * that certain asics don't have (all asics).
392 * Returns the value in the register.
394 static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
395 uint32_t block, uint32_t reg)
397 DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
404 * amdgpu_block_invalid_wreg - dummy reg write function
406 * @adev: amdgpu device pointer
407 * @block: offset of instance
408 * @reg: offset of register
409 * @v: value to write to the register
411 * Dummy register read function. Used for register blocks
412 * that certain asics don't have (all asics).
414 static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
416 uint32_t reg, uint32_t v)
418 DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
424 * amdgpu_device_vram_scratch_init - allocate the VRAM scratch page
426 * @adev: amdgpu device pointer
428 * Allocates a scratch page of VRAM for use by various things in the
431 static int amdgpu_device_vram_scratch_init(struct amdgpu_device *adev)
433 return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE,
434 PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
435 &adev->vram_scratch.robj,
436 &adev->vram_scratch.gpu_addr,
437 (void **)&adev->vram_scratch.ptr);
441 * amdgpu_device_vram_scratch_fini - Free the VRAM scratch page
443 * @adev: amdgpu device pointer
445 * Frees the VRAM scratch page.
447 static void amdgpu_device_vram_scratch_fini(struct amdgpu_device *adev)
449 amdgpu_bo_free_kernel(&adev->vram_scratch.robj, NULL, NULL);
453 * amdgpu_device_program_register_sequence - program an array of registers.
455 * @adev: amdgpu_device pointer
456 * @registers: pointer to the register array
457 * @array_size: size of the register array
459 * Programs an array or registers with and and or masks.
460 * This is a helper for setting golden registers.
462 void amdgpu_device_program_register_sequence(struct amdgpu_device *adev,
463 const u32 *registers,
464 const u32 array_size)
466 u32 tmp, reg, and_mask, or_mask;
472 for (i = 0; i < array_size; i +=3) {
473 reg = registers[i + 0];
474 and_mask = registers[i + 1];
475 or_mask = registers[i + 2];
477 if (and_mask == 0xffffffff) {
489 * amdgpu_device_pci_config_reset - reset the GPU
491 * @adev: amdgpu_device pointer
493 * Resets the GPU using the pci config reset sequence.
494 * Only applicable to asics prior to vega10.
496 void amdgpu_device_pci_config_reset(struct amdgpu_device *adev)
498 pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
502 * GPU doorbell aperture helpers function.
505 * amdgpu_device_doorbell_init - Init doorbell driver information.
507 * @adev: amdgpu_device pointer
509 * Init doorbell driver information (CIK)
510 * Returns 0 on success, error on failure.
512 static int amdgpu_device_doorbell_init(struct amdgpu_device *adev)
514 /* No doorbell on SI hardware generation */
515 if (adev->asic_type < CHIP_BONAIRE) {
516 adev->doorbell.base = 0;
517 adev->doorbell.size = 0;
518 adev->doorbell.num_doorbells = 0;
519 adev->doorbell.ptr = NULL;
523 if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET)
526 /* doorbell bar mapping */
527 adev->doorbell.base = pci_resource_start(adev->pdev, 2);
528 adev->doorbell.size = pci_resource_len(adev->pdev, 2);
530 adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32),
531 AMDGPU_DOORBELL_MAX_ASSIGNMENT+1);
532 if (adev->doorbell.num_doorbells == 0)
535 adev->doorbell.ptr = ioremap(adev->doorbell.base,
536 adev->doorbell.num_doorbells *
538 if (adev->doorbell.ptr == NULL)
545 * amdgpu_device_doorbell_fini - Tear down doorbell driver information.
547 * @adev: amdgpu_device pointer
549 * Tear down doorbell driver information (CIK)
551 static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev)
553 iounmap(adev->doorbell.ptr);
554 adev->doorbell.ptr = NULL;
560 * amdgpu_device_wb_*()
561 * Writeback is the method by which the GPU updates special pages in memory
562 * with the status of certain GPU events (fences, ring pointers,etc.).
566 * amdgpu_device_wb_fini - Disable Writeback and free memory
568 * @adev: amdgpu_device pointer
570 * Disables Writeback and frees the Writeback memory (all asics).
571 * Used at driver shutdown.
573 static void amdgpu_device_wb_fini(struct amdgpu_device *adev)
575 if (adev->wb.wb_obj) {
576 amdgpu_bo_free_kernel(&adev->wb.wb_obj,
578 (void **)&adev->wb.wb);
579 adev->wb.wb_obj = NULL;
584 * amdgpu_device_wb_init- Init Writeback driver info and allocate memory
586 * @adev: amdgpu_device pointer
588 * Initializes writeback and allocates writeback memory (all asics).
589 * Used at driver startup.
590 * Returns 0 on success or an -error on failure.
592 static int amdgpu_device_wb_init(struct amdgpu_device *adev)
596 if (adev->wb.wb_obj == NULL) {
597 /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */
598 r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8,
599 PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
600 &adev->wb.wb_obj, &adev->wb.gpu_addr,
601 (void **)&adev->wb.wb);
603 dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
607 adev->wb.num_wb = AMDGPU_MAX_WB;
608 memset(&adev->wb.used, 0, sizeof(adev->wb.used));
610 /* clear wb memory */
611 memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8);
618 * amdgpu_device_wb_get - Allocate a wb entry
620 * @adev: amdgpu_device pointer
623 * Allocate a wb slot for use by the driver (all asics).
624 * Returns 0 on success or -EINVAL on failure.
626 int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb)
628 unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
630 if (offset < adev->wb.num_wb) {
631 __set_bit(offset, adev->wb.used);
632 *wb = offset << 3; /* convert to dw offset */
640 * amdgpu_device_wb_free - Free a wb entry
642 * @adev: amdgpu_device pointer
645 * Free a wb slot allocated for use by the driver (all asics)
647 void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb)
650 if (wb < adev->wb.num_wb)
651 __clear_bit(wb, adev->wb.used);
655 * amdgpu_device_resize_fb_bar - try to resize FB BAR
657 * @adev: amdgpu_device pointer
659 * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not
660 * to fail, but if any of the BARs is not accessible after the size we abort
661 * driver loading by returning -ENODEV.
663 int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev)
665 u64 space_needed = roundup_pow_of_two(adev->gmc.real_vram_size);
666 u32 rbar_size = order_base_2(((space_needed >> 20) | 1)) - 1;
667 struct pci_bus *root;
668 struct resource *res;
674 if (amdgpu_sriov_vf(adev))
677 /* Check if the root BUS has 64bit memory resources */
678 root = adev->pdev->bus;
682 pci_bus_for_each_resource(root, res, i) {
683 if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) &&
684 res->start > 0x100000000ull)
688 /* Trying to resize is pointless without a root hub window above 4GB */
692 /* Disable memory decoding while we change the BAR addresses and size */
693 pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd);
694 pci_write_config_word(adev->pdev, PCI_COMMAND,
695 cmd & ~PCI_COMMAND_MEMORY);
697 /* Free the VRAM and doorbell BAR, we most likely need to move both. */
698 amdgpu_device_doorbell_fini(adev);
699 if (adev->asic_type >= CHIP_BONAIRE)
700 pci_release_resource(adev->pdev, 2);
702 pci_release_resource(adev->pdev, 0);
704 r = pci_resize_resource(adev->pdev, 0, rbar_size);
706 DRM_INFO("Not enough PCI address space for a large BAR.");
707 else if (r && r != -ENOTSUPP)
708 DRM_ERROR("Problem resizing BAR0 (%d).", r);
710 pci_assign_unassigned_bus_resources(adev->pdev->bus);
712 /* When the doorbell or fb BAR isn't available we have no chance of
715 r = amdgpu_device_doorbell_init(adev);
716 if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET))
719 pci_write_config_word(adev->pdev, PCI_COMMAND, cmd);
725 * GPU helpers function.
728 * amdgpu_device_need_post - check if the hw need post or not
730 * @adev: amdgpu_device pointer
732 * Check if the asic has been initialized (all asics) at driver startup
733 * or post is needed if hw reset is performed.
734 * Returns true if need or false if not.
736 bool amdgpu_device_need_post(struct amdgpu_device *adev)
740 if (amdgpu_sriov_vf(adev))
743 if (amdgpu_passthrough(adev)) {
744 /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
745 * some old smc fw still need driver do vPost otherwise gpu hang, while
746 * those smc fw version above 22.15 doesn't have this flaw, so we force
747 * vpost executed for smc version below 22.15
749 if (adev->asic_type == CHIP_FIJI) {
752 err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
753 /* force vPost if error occured */
757 fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
758 if (fw_ver < 0x00160e00)
763 if (adev->has_hw_reset) {
764 adev->has_hw_reset = false;
768 /* bios scratch used on CIK+ */
769 if (adev->asic_type >= CHIP_BONAIRE)
770 return amdgpu_atombios_scratch_need_asic_init(adev);
772 /* check MEM_SIZE for older asics */
773 reg = amdgpu_asic_get_config_memsize(adev);
775 if ((reg != 0) && (reg != 0xffffffff))
781 /* if we get transitioned to only one device, take VGA back */
783 * amdgpu_device_vga_set_decode - enable/disable vga decode
785 * @cookie: amdgpu_device pointer
786 * @state: enable/disable vga decode
788 * Enable/disable vga decode (all asics).
789 * Returns VGA resource flags.
791 static unsigned int amdgpu_device_vga_set_decode(void *cookie, bool state)
793 struct amdgpu_device *adev = cookie;
794 amdgpu_asic_set_vga_state(adev, state);
796 return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
797 VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
799 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
803 * amdgpu_device_check_block_size - validate the vm block size
805 * @adev: amdgpu_device pointer
807 * Validates the vm block size specified via module parameter.
808 * The vm block size defines number of bits in page table versus page directory,
809 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
810 * page table and the remaining bits are in the page directory.
812 static void amdgpu_device_check_block_size(struct amdgpu_device *adev)
814 /* defines number of bits in page table versus page directory,
815 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
816 * page table and the remaining bits are in the page directory */
817 if (amdgpu_vm_block_size == -1)
820 if (amdgpu_vm_block_size < 9) {
821 dev_warn(adev->dev, "VM page table size (%d) too small\n",
822 amdgpu_vm_block_size);
823 amdgpu_vm_block_size = -1;
828 * amdgpu_device_check_vm_size - validate the vm size
830 * @adev: amdgpu_device pointer
832 * Validates the vm size in GB specified via module parameter.
833 * The VM size is the size of the GPU virtual memory space in GB.
835 static void amdgpu_device_check_vm_size(struct amdgpu_device *adev)
837 /* no need to check the default value */
838 if (amdgpu_vm_size == -1)
841 if (amdgpu_vm_size < 1) {
842 dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
848 static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev)
851 bool is_os_64 = (sizeof(void *) == 8) ? true : false;
852 uint64_t total_memory;
853 uint64_t dram_size_seven_GB = 0x1B8000000;
854 uint64_t dram_size_three_GB = 0xB8000000;
856 if (amdgpu_smu_memory_pool_size == 0)
860 DRM_WARN("Not 64-bit OS, feature not supported\n");
864 total_memory = (uint64_t)si.totalram * si.mem_unit;
866 if ((amdgpu_smu_memory_pool_size == 1) ||
867 (amdgpu_smu_memory_pool_size == 2)) {
868 if (total_memory < dram_size_three_GB)
870 } else if ((amdgpu_smu_memory_pool_size == 4) ||
871 (amdgpu_smu_memory_pool_size == 8)) {
872 if (total_memory < dram_size_seven_GB)
875 DRM_WARN("Smu memory pool size not supported\n");
878 adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28;
883 DRM_WARN("No enough system memory\n");
885 adev->pm.smu_prv_buffer_size = 0;
889 * amdgpu_device_check_arguments - validate module params
891 * @adev: amdgpu_device pointer
893 * Validates certain module parameters and updates
894 * the associated values used by the driver (all asics).
896 static void amdgpu_device_check_arguments(struct amdgpu_device *adev)
898 if (amdgpu_sched_jobs < 4) {
899 dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
901 amdgpu_sched_jobs = 4;
902 } else if (!is_power_of_2(amdgpu_sched_jobs)){
903 dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
905 amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
908 if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) {
909 /* gart size must be greater or equal to 32M */
910 dev_warn(adev->dev, "gart size (%d) too small\n",
912 amdgpu_gart_size = -1;
915 if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) {
916 /* gtt size must be greater or equal to 32M */
917 dev_warn(adev->dev, "gtt size (%d) too small\n",
919 amdgpu_gtt_size = -1;
922 /* valid range is between 4 and 9 inclusive */
923 if (amdgpu_vm_fragment_size != -1 &&
924 (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) {
925 dev_warn(adev->dev, "valid range is between 4 and 9\n");
926 amdgpu_vm_fragment_size = -1;
929 amdgpu_device_check_smu_prv_buffer_size(adev);
931 amdgpu_device_check_vm_size(adev);
933 amdgpu_device_check_block_size(adev);
935 if (amdgpu_vram_page_split != -1 && (amdgpu_vram_page_split < 16 ||
936 !is_power_of_2(amdgpu_vram_page_split))) {
937 dev_warn(adev->dev, "invalid VRAM page split (%d)\n",
938 amdgpu_vram_page_split);
939 amdgpu_vram_page_split = 1024;
942 if (amdgpu_lockup_timeout == 0) {
943 dev_warn(adev->dev, "lockup_timeout msut be > 0, adjusting to 10000\n");
944 amdgpu_lockup_timeout = 10000;
947 adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type);
951 * amdgpu_switcheroo_set_state - set switcheroo state
953 * @pdev: pci dev pointer
954 * @state: vga_switcheroo state
956 * Callback for the switcheroo driver. Suspends or resumes the
957 * the asics before or after it is powered up using ACPI methods.
959 static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, enum vga_switcheroo_state state)
961 struct drm_device *dev = pci_get_drvdata(pdev);
963 if (amdgpu_device_is_px(dev) && state == VGA_SWITCHEROO_OFF)
966 if (state == VGA_SWITCHEROO_ON) {
967 pr_info("amdgpu: switched on\n");
968 /* don't suspend or resume card normally */
969 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
971 amdgpu_device_resume(dev, true, true);
973 dev->switch_power_state = DRM_SWITCH_POWER_ON;
974 drm_kms_helper_poll_enable(dev);
976 pr_info("amdgpu: switched off\n");
977 drm_kms_helper_poll_disable(dev);
978 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
979 amdgpu_device_suspend(dev, true, true);
980 dev->switch_power_state = DRM_SWITCH_POWER_OFF;
985 * amdgpu_switcheroo_can_switch - see if switcheroo state can change
987 * @pdev: pci dev pointer
989 * Callback for the switcheroo driver. Check of the switcheroo
990 * state can be changed.
991 * Returns true if the state can be changed, false if not.
993 static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
995 struct drm_device *dev = pci_get_drvdata(pdev);
998 * FIXME: open_count is protected by drm_global_mutex but that would lead to
999 * locking inversion with the driver load path. And the access here is
1000 * completely racy anyway. So don't bother with locking for now.
1002 return dev->open_count == 0;
1005 static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
1006 .set_gpu_state = amdgpu_switcheroo_set_state,
1008 .can_switch = amdgpu_switcheroo_can_switch,
1012 * amdgpu_device_ip_set_clockgating_state - set the CG state
1014 * @dev: amdgpu_device pointer
1015 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1016 * @state: clockgating state (gate or ungate)
1018 * Sets the requested clockgating state for all instances of
1019 * the hardware IP specified.
1020 * Returns the error code from the last instance.
1022 int amdgpu_device_ip_set_clockgating_state(void *dev,
1023 enum amd_ip_block_type block_type,
1024 enum amd_clockgating_state state)
1026 struct amdgpu_device *adev = dev;
1029 for (i = 0; i < adev->num_ip_blocks; i++) {
1030 if (!adev->ip_blocks[i].status.valid)
1032 if (adev->ip_blocks[i].version->type != block_type)
1034 if (!adev->ip_blocks[i].version->funcs->set_clockgating_state)
1036 r = adev->ip_blocks[i].version->funcs->set_clockgating_state(
1037 (void *)adev, state);
1039 DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",
1040 adev->ip_blocks[i].version->funcs->name, r);
1046 * amdgpu_device_ip_set_powergating_state - set the PG state
1048 * @dev: amdgpu_device pointer
1049 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1050 * @state: powergating state (gate or ungate)
1052 * Sets the requested powergating state for all instances of
1053 * the hardware IP specified.
1054 * Returns the error code from the last instance.
1056 int amdgpu_device_ip_set_powergating_state(void *dev,
1057 enum amd_ip_block_type block_type,
1058 enum amd_powergating_state state)
1060 struct amdgpu_device *adev = dev;
1063 for (i = 0; i < adev->num_ip_blocks; i++) {
1064 if (!adev->ip_blocks[i].status.valid)
1066 if (adev->ip_blocks[i].version->type != block_type)
1068 if (!adev->ip_blocks[i].version->funcs->set_powergating_state)
1070 r = adev->ip_blocks[i].version->funcs->set_powergating_state(
1071 (void *)adev, state);
1073 DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",
1074 adev->ip_blocks[i].version->funcs->name, r);
1080 * amdgpu_device_ip_get_clockgating_state - get the CG state
1082 * @adev: amdgpu_device pointer
1083 * @flags: clockgating feature flags
1085 * Walks the list of IPs on the device and updates the clockgating
1086 * flags for each IP.
1087 * Updates @flags with the feature flags for each hardware IP where
1088 * clockgating is enabled.
1090 void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev,
1095 for (i = 0; i < adev->num_ip_blocks; i++) {
1096 if (!adev->ip_blocks[i].status.valid)
1098 if (adev->ip_blocks[i].version->funcs->get_clockgating_state)
1099 adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags);
1104 * amdgpu_device_ip_wait_for_idle - wait for idle
1106 * @adev: amdgpu_device pointer
1107 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1109 * Waits for the request hardware IP to be idle.
1110 * Returns 0 for success or a negative error code on failure.
1112 int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev,
1113 enum amd_ip_block_type block_type)
1117 for (i = 0; i < adev->num_ip_blocks; i++) {
1118 if (!adev->ip_blocks[i].status.valid)
1120 if (adev->ip_blocks[i].version->type == block_type) {
1121 r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
1132 * amdgpu_device_ip_is_idle - is the hardware IP idle
1134 * @adev: amdgpu_device pointer
1135 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1137 * Check if the hardware IP is idle or not.
1138 * Returns true if it the IP is idle, false if not.
1140 bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev,
1141 enum amd_ip_block_type block_type)
1145 for (i = 0; i < adev->num_ip_blocks; i++) {
1146 if (!adev->ip_blocks[i].status.valid)
1148 if (adev->ip_blocks[i].version->type == block_type)
1149 return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
1156 * amdgpu_device_ip_get_ip_block - get a hw IP pointer
1158 * @adev: amdgpu_device pointer
1159 * @type: Type of hardware IP (SMU, GFX, UVD, etc.)
1161 * Returns a pointer to the hardware IP block structure
1162 * if it exists for the asic, otherwise NULL.
1164 struct amdgpu_ip_block *
1165 amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,
1166 enum amd_ip_block_type type)
1170 for (i = 0; i < adev->num_ip_blocks; i++)
1171 if (adev->ip_blocks[i].version->type == type)
1172 return &adev->ip_blocks[i];
1178 * amdgpu_device_ip_block_version_cmp
1180 * @adev: amdgpu_device pointer
1181 * @type: enum amd_ip_block_type
1182 * @major: major version
1183 * @minor: minor version
1185 * return 0 if equal or greater
1186 * return 1 if smaller or the ip_block doesn't exist
1188 int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev,
1189 enum amd_ip_block_type type,
1190 u32 major, u32 minor)
1192 struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type);
1194 if (ip_block && ((ip_block->version->major > major) ||
1195 ((ip_block->version->major == major) &&
1196 (ip_block->version->minor >= minor))))
1203 * amdgpu_device_ip_block_add
1205 * @adev: amdgpu_device pointer
1206 * @ip_block_version: pointer to the IP to add
1208 * Adds the IP block driver information to the collection of IPs
1211 int amdgpu_device_ip_block_add(struct amdgpu_device *adev,
1212 const struct amdgpu_ip_block_version *ip_block_version)
1214 if (!ip_block_version)
1217 DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks,
1218 ip_block_version->funcs->name);
1220 adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
1226 * amdgpu_device_enable_virtual_display - enable virtual display feature
1228 * @adev: amdgpu_device pointer
1230 * Enabled the virtual display feature if the user has enabled it via
1231 * the module parameter virtual_display. This feature provides a virtual
1232 * display hardware on headless boards or in virtualized environments.
1233 * This function parses and validates the configuration string specified by
1234 * the user and configues the virtual display configuration (number of
1235 * virtual connectors, crtcs, etc.) specified.
1237 static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
1239 adev->enable_virtual_display = false;
1241 if (amdgpu_virtual_display) {
1242 struct drm_device *ddev = adev->ddev;
1243 const char *pci_address_name = pci_name(ddev->pdev);
1244 char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
1246 pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
1247 pciaddstr_tmp = pciaddstr;
1248 while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
1249 pciaddname = strsep(&pciaddname_tmp, ",");
1250 if (!strcmp("all", pciaddname)
1251 || !strcmp(pci_address_name, pciaddname)) {
1255 adev->enable_virtual_display = true;
1258 res = kstrtol(pciaddname_tmp, 10,
1266 adev->mode_info.num_crtc = num_crtc;
1268 adev->mode_info.num_crtc = 1;
1274 DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
1275 amdgpu_virtual_display, pci_address_name,
1276 adev->enable_virtual_display, adev->mode_info.num_crtc);
1283 * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware
1285 * @adev: amdgpu_device pointer
1287 * Parses the asic configuration parameters specified in the gpu info
1288 * firmware and makes them availale to the driver for use in configuring
1290 * Returns 0 on success, -EINVAL on failure.
1292 static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev)
1294 const char *chip_name;
1297 const struct gpu_info_firmware_header_v1_0 *hdr;
1299 adev->firmware.gpu_info_fw = NULL;
1301 switch (adev->asic_type) {
1305 case CHIP_POLARIS10:
1306 case CHIP_POLARIS11:
1307 case CHIP_POLARIS12:
1311 #ifdef CONFIG_DRM_AMDGPU_SI
1318 #ifdef CONFIG_DRM_AMDGPU_CIK
1329 chip_name = "vega10";
1332 chip_name = "vega12";
1335 chip_name = "raven";
1339 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name);
1340 err = request_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev);
1343 "Failed to load gpu_info firmware \"%s\"\n",
1347 err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw);
1350 "Failed to validate gpu_info firmware \"%s\"\n",
1355 hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data;
1356 amdgpu_ucode_print_gpu_info_hdr(&hdr->header);
1358 switch (hdr->version_major) {
1361 const struct gpu_info_firmware_v1_0 *gpu_info_fw =
1362 (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data +
1363 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1365 adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se);
1366 adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh);
1367 adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se);
1368 adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se);
1369 adev->gfx.config.max_texture_channel_caches =
1370 le32_to_cpu(gpu_info_fw->gc_num_tccs);
1371 adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs);
1372 adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds);
1373 adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth);
1374 adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth);
1375 adev->gfx.config.double_offchip_lds_buf =
1376 le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer);
1377 adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size);
1378 adev->gfx.cu_info.max_waves_per_simd =
1379 le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd);
1380 adev->gfx.cu_info.max_scratch_slots_per_cu =
1381 le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu);
1382 adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size);
1387 "Unsupported gpu_info table %d\n", hdr->header.ucode_version);
1396 * amdgpu_device_ip_early_init - run early init for hardware IPs
1398 * @adev: amdgpu_device pointer
1400 * Early initialization pass for hardware IPs. The hardware IPs that make
1401 * up each asic are discovered each IP's early_init callback is run. This
1402 * is the first stage in initializing the asic.
1403 * Returns 0 on success, negative error code on failure.
1405 static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
1409 amdgpu_device_enable_virtual_display(adev);
1411 switch (adev->asic_type) {
1415 case CHIP_POLARIS10:
1416 case CHIP_POLARIS11:
1417 case CHIP_POLARIS12:
1421 if (adev->asic_type == CHIP_CARRIZO || adev->asic_type == CHIP_STONEY)
1422 adev->family = AMDGPU_FAMILY_CZ;
1424 adev->family = AMDGPU_FAMILY_VI;
1426 r = vi_set_ip_blocks(adev);
1430 #ifdef CONFIG_DRM_AMDGPU_SI
1436 adev->family = AMDGPU_FAMILY_SI;
1437 r = si_set_ip_blocks(adev);
1442 #ifdef CONFIG_DRM_AMDGPU_CIK
1448 if ((adev->asic_type == CHIP_BONAIRE) || (adev->asic_type == CHIP_HAWAII))
1449 adev->family = AMDGPU_FAMILY_CI;
1451 adev->family = AMDGPU_FAMILY_KV;
1453 r = cik_set_ip_blocks(adev);
1462 if (adev->asic_type == CHIP_RAVEN)
1463 adev->family = AMDGPU_FAMILY_RV;
1465 adev->family = AMDGPU_FAMILY_AI;
1467 r = soc15_set_ip_blocks(adev);
1472 /* FIXME: not supported yet */
1476 r = amdgpu_device_parse_gpu_info_fw(adev);
1480 amdgpu_amdkfd_device_probe(adev);
1482 if (amdgpu_sriov_vf(adev)) {
1483 r = amdgpu_virt_request_full_gpu(adev, true);
1488 adev->powerplay.pp_feature = amdgpu_pp_feature_mask;
1490 for (i = 0; i < adev->num_ip_blocks; i++) {
1491 if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
1492 DRM_ERROR("disabled ip block: %d <%s>\n",
1493 i, adev->ip_blocks[i].version->funcs->name);
1494 adev->ip_blocks[i].status.valid = false;
1496 if (adev->ip_blocks[i].version->funcs->early_init) {
1497 r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
1499 adev->ip_blocks[i].status.valid = false;
1501 DRM_ERROR("early_init of IP block <%s> failed %d\n",
1502 adev->ip_blocks[i].version->funcs->name, r);
1505 adev->ip_blocks[i].status.valid = true;
1508 adev->ip_blocks[i].status.valid = true;
1513 adev->cg_flags &= amdgpu_cg_mask;
1514 adev->pg_flags &= amdgpu_pg_mask;
1520 * amdgpu_device_ip_init - run init for hardware IPs
1522 * @adev: amdgpu_device pointer
1524 * Main initialization pass for hardware IPs. The list of all the hardware
1525 * IPs that make up the asic is walked and the sw_init and hw_init callbacks
1526 * are run. sw_init initializes the software state associated with each IP
1527 * and hw_init initializes the hardware associated with each IP.
1528 * Returns 0 on success, negative error code on failure.
1530 static int amdgpu_device_ip_init(struct amdgpu_device *adev)
1534 for (i = 0; i < adev->num_ip_blocks; i++) {
1535 if (!adev->ip_blocks[i].status.valid)
1537 r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
1539 DRM_ERROR("sw_init of IP block <%s> failed %d\n",
1540 adev->ip_blocks[i].version->funcs->name, r);
1543 adev->ip_blocks[i].status.sw = true;
1545 /* need to do gmc hw init early so we can allocate gpu mem */
1546 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1547 r = amdgpu_device_vram_scratch_init(adev);
1549 DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r);
1552 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1554 DRM_ERROR("hw_init %d failed %d\n", i, r);
1557 r = amdgpu_device_wb_init(adev);
1559 DRM_ERROR("amdgpu_device_wb_init failed %d\n", r);
1562 adev->ip_blocks[i].status.hw = true;
1564 /* right after GMC hw init, we create CSA */
1565 if (amdgpu_sriov_vf(adev)) {
1566 r = amdgpu_allocate_static_csa(adev);
1568 DRM_ERROR("allocate CSA failed %d\n", r);
1575 for (i = 0; i < adev->num_ip_blocks; i++) {
1576 if (!adev->ip_blocks[i].status.sw)
1578 if (adev->ip_blocks[i].status.hw)
1580 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1582 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1583 adev->ip_blocks[i].version->funcs->name, r);
1586 adev->ip_blocks[i].status.hw = true;
1589 amdgpu_amdkfd_device_init(adev);
1591 if (amdgpu_sriov_vf(adev))
1592 amdgpu_virt_release_full_gpu(adev, true);
1598 * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer
1600 * @adev: amdgpu_device pointer
1602 * Writes a reset magic value to the gart pointer in VRAM. The driver calls
1603 * this function before a GPU reset. If the value is retained after a
1604 * GPU reset, VRAM has not been lost. Some GPU resets may destry VRAM contents.
1606 static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev)
1608 memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
1612 * amdgpu_device_check_vram_lost - check if vram is valid
1614 * @adev: amdgpu_device pointer
1616 * Checks the reset magic value written to the gart pointer in VRAM.
1617 * The driver calls this after a GPU reset to see if the contents of
1618 * VRAM is lost or now.
1619 * returns true if vram is lost, false if not.
1621 static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev)
1623 return !!memcmp(adev->gart.ptr, adev->reset_magic,
1624 AMDGPU_RESET_MAGIC_NUM);
1628 * amdgpu_device_set_cg_state - set clockgating for amdgpu device
1630 * @adev: amdgpu_device pointer
1632 * The list of all the hardware IPs that make up the asic is walked and the
1633 * set_clockgating_state callbacks are run.
1634 * Late initialization pass enabling clockgating for hardware IPs.
1635 * Fini or suspend, pass disabling clockgating for hardware IPs.
1636 * Returns 0 on success, negative error code on failure.
1639 static int amdgpu_device_set_cg_state(struct amdgpu_device *adev,
1640 enum amd_clockgating_state state)
1644 if (amdgpu_emu_mode == 1)
1647 for (j = 0; j < adev->num_ip_blocks; j++) {
1648 i = state == AMD_CG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
1649 if (!adev->ip_blocks[i].status.valid)
1651 /* skip CG for VCE/UVD, it's handled specially */
1652 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1653 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1654 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
1655 adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1656 /* enable clockgating to save power */
1657 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1660 DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
1661 adev->ip_blocks[i].version->funcs->name, r);
1670 static int amdgpu_device_set_pg_state(struct amdgpu_device *adev, enum amd_powergating_state state)
1674 if (amdgpu_emu_mode == 1)
1677 for (j = 0; j < adev->num_ip_blocks; j++) {
1678 i = state == AMD_PG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
1679 if (!adev->ip_blocks[i].status.valid)
1681 /* skip CG for VCE/UVD, it's handled specially */
1682 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1683 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1684 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
1685 adev->ip_blocks[i].version->funcs->set_powergating_state) {
1686 /* enable powergating to save power */
1687 r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev,
1690 DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n",
1691 adev->ip_blocks[i].version->funcs->name, r);
1700 * amdgpu_device_ip_late_init - run late init for hardware IPs
1702 * @adev: amdgpu_device pointer
1704 * Late initialization pass for hardware IPs. The list of all the hardware
1705 * IPs that make up the asic is walked and the late_init callbacks are run.
1706 * late_init covers any special initialization that an IP requires
1707 * after all of the have been initialized or something that needs to happen
1708 * late in the init process.
1709 * Returns 0 on success, negative error code on failure.
1711 static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
1715 for (i = 0; i < adev->num_ip_blocks; i++) {
1716 if (!adev->ip_blocks[i].status.valid)
1718 if (adev->ip_blocks[i].version->funcs->late_init) {
1719 r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
1721 DRM_ERROR("late_init of IP block <%s> failed %d\n",
1722 adev->ip_blocks[i].version->funcs->name, r);
1725 adev->ip_blocks[i].status.late_initialized = true;
1729 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE);
1730 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE);
1732 queue_delayed_work(system_wq, &adev->late_init_work,
1733 msecs_to_jiffies(AMDGPU_RESUME_MS));
1735 amdgpu_device_fill_reset_magic(adev);
1741 * amdgpu_device_ip_fini - run fini for hardware IPs
1743 * @adev: amdgpu_device pointer
1745 * Main teardown pass for hardware IPs. The list of all the hardware
1746 * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks
1747 * are run. hw_fini tears down the hardware associated with each IP
1748 * and sw_fini tears down any software state associated with each IP.
1749 * Returns 0 on success, negative error code on failure.
1751 static int amdgpu_device_ip_fini(struct amdgpu_device *adev)
1755 amdgpu_amdkfd_device_fini(adev);
1757 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
1758 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
1760 /* need to disable SMC first */
1761 for (i = 0; i < adev->num_ip_blocks; i++) {
1762 if (!adev->ip_blocks[i].status.hw)
1764 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
1765 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
1766 /* XXX handle errors */
1768 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
1769 adev->ip_blocks[i].version->funcs->name, r);
1771 adev->ip_blocks[i].status.hw = false;
1776 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1777 if (!adev->ip_blocks[i].status.hw)
1780 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
1781 /* XXX handle errors */
1783 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
1784 adev->ip_blocks[i].version->funcs->name, r);
1787 adev->ip_blocks[i].status.hw = false;
1791 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1792 if (!adev->ip_blocks[i].status.sw)
1795 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1796 amdgpu_free_static_csa(adev);
1797 amdgpu_device_wb_fini(adev);
1798 amdgpu_device_vram_scratch_fini(adev);
1801 r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
1802 /* XXX handle errors */
1804 DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
1805 adev->ip_blocks[i].version->funcs->name, r);
1807 adev->ip_blocks[i].status.sw = false;
1808 adev->ip_blocks[i].status.valid = false;
1811 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1812 if (!adev->ip_blocks[i].status.late_initialized)
1814 if (adev->ip_blocks[i].version->funcs->late_fini)
1815 adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
1816 adev->ip_blocks[i].status.late_initialized = false;
1819 if (amdgpu_sriov_vf(adev))
1820 if (amdgpu_virt_release_full_gpu(adev, false))
1821 DRM_ERROR("failed to release exclusive mode on fini\n");
1827 * amdgpu_device_ip_late_init_func_handler - work handler for ib test
1829 * @work: work_struct.
1831 static void amdgpu_device_ip_late_init_func_handler(struct work_struct *work)
1833 struct amdgpu_device *adev =
1834 container_of(work, struct amdgpu_device, late_init_work.work);
1837 r = amdgpu_ib_ring_tests(adev);
1839 DRM_ERROR("ib ring test failed (%d).\n", r);
1842 static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work)
1844 struct amdgpu_device *adev =
1845 container_of(work, struct amdgpu_device, gfx.gfx_off_delay_work.work);
1847 mutex_lock(&adev->gfx.gfx_off_mutex);
1848 if (!adev->gfx.gfx_off_state && !adev->gfx.gfx_off_req_count) {
1849 if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, true))
1850 adev->gfx.gfx_off_state = true;
1852 mutex_unlock(&adev->gfx.gfx_off_mutex);
1856 * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1)
1858 * @adev: amdgpu_device pointer
1860 * Main suspend function for hardware IPs. The list of all the hardware
1861 * IPs that make up the asic is walked, clockgating is disabled and the
1862 * suspend callbacks are run. suspend puts the hardware and software state
1863 * in each IP into a state suitable for suspend.
1864 * Returns 0 on success, negative error code on failure.
1866 static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev)
1870 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
1871 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
1873 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1874 if (!adev->ip_blocks[i].status.valid)
1876 /* displays are handled separately */
1877 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) {
1878 /* XXX handle errors */
1879 r = adev->ip_blocks[i].version->funcs->suspend(adev);
1880 /* XXX handle errors */
1882 DRM_ERROR("suspend of IP block <%s> failed %d\n",
1883 adev->ip_blocks[i].version->funcs->name, r);
1892 * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2)
1894 * @adev: amdgpu_device pointer
1896 * Main suspend function for hardware IPs. The list of all the hardware
1897 * IPs that make up the asic is walked, clockgating is disabled and the
1898 * suspend callbacks are run. suspend puts the hardware and software state
1899 * in each IP into a state suitable for suspend.
1900 * Returns 0 on success, negative error code on failure.
1902 static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev)
1906 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1907 if (!adev->ip_blocks[i].status.valid)
1909 /* displays are handled in phase1 */
1910 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE)
1912 /* XXX handle errors */
1913 r = adev->ip_blocks[i].version->funcs->suspend(adev);
1914 /* XXX handle errors */
1916 DRM_ERROR("suspend of IP block <%s> failed %d\n",
1917 adev->ip_blocks[i].version->funcs->name, r);
1925 * amdgpu_device_ip_suspend - run suspend for hardware IPs
1927 * @adev: amdgpu_device pointer
1929 * Main suspend function for hardware IPs. The list of all the hardware
1930 * IPs that make up the asic is walked, clockgating is disabled and the
1931 * suspend callbacks are run. suspend puts the hardware and software state
1932 * in each IP into a state suitable for suspend.
1933 * Returns 0 on success, negative error code on failure.
1935 int amdgpu_device_ip_suspend(struct amdgpu_device *adev)
1939 if (amdgpu_sriov_vf(adev))
1940 amdgpu_virt_request_full_gpu(adev, false);
1942 r = amdgpu_device_ip_suspend_phase1(adev);
1945 r = amdgpu_device_ip_suspend_phase2(adev);
1947 if (amdgpu_sriov_vf(adev))
1948 amdgpu_virt_release_full_gpu(adev, false);
1953 static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
1957 static enum amd_ip_block_type ip_order[] = {
1958 AMD_IP_BLOCK_TYPE_GMC,
1959 AMD_IP_BLOCK_TYPE_COMMON,
1960 AMD_IP_BLOCK_TYPE_IH,
1963 for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
1965 struct amdgpu_ip_block *block;
1967 for (j = 0; j < adev->num_ip_blocks; j++) {
1968 block = &adev->ip_blocks[j];
1970 if (block->version->type != ip_order[i] ||
1971 !block->status.valid)
1974 r = block->version->funcs->hw_init(adev);
1975 DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
1984 static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
1988 static enum amd_ip_block_type ip_order[] = {
1989 AMD_IP_BLOCK_TYPE_SMC,
1990 AMD_IP_BLOCK_TYPE_PSP,
1991 AMD_IP_BLOCK_TYPE_DCE,
1992 AMD_IP_BLOCK_TYPE_GFX,
1993 AMD_IP_BLOCK_TYPE_SDMA,
1994 AMD_IP_BLOCK_TYPE_UVD,
1995 AMD_IP_BLOCK_TYPE_VCE
1998 for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
2000 struct amdgpu_ip_block *block;
2002 for (j = 0; j < adev->num_ip_blocks; j++) {
2003 block = &adev->ip_blocks[j];
2005 if (block->version->type != ip_order[i] ||
2006 !block->status.valid)
2009 r = block->version->funcs->hw_init(adev);
2010 DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
2020 * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs
2022 * @adev: amdgpu_device pointer
2024 * First resume function for hardware IPs. The list of all the hardware
2025 * IPs that make up the asic is walked and the resume callbacks are run for
2026 * COMMON, GMC, and IH. resume puts the hardware into a functional state
2027 * after a suspend and updates the software state as necessary. This
2028 * function is also used for restoring the GPU after a GPU reset.
2029 * Returns 0 on success, negative error code on failure.
2031 static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)
2035 for (i = 0; i < adev->num_ip_blocks; i++) {
2036 if (!adev->ip_blocks[i].status.valid)
2038 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2039 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2040 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
2041 r = adev->ip_blocks[i].version->funcs->resume(adev);
2043 DRM_ERROR("resume of IP block <%s> failed %d\n",
2044 adev->ip_blocks[i].version->funcs->name, r);
2054 * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs
2056 * @adev: amdgpu_device pointer
2058 * First resume function for hardware IPs. The list of all the hardware
2059 * IPs that make up the asic is walked and the resume callbacks are run for
2060 * all blocks except COMMON, GMC, and IH. resume puts the hardware into a
2061 * functional state after a suspend and updates the software state as
2062 * necessary. This function is also used for restoring the GPU after a GPU
2064 * Returns 0 on success, negative error code on failure.
2066 static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)
2070 for (i = 0; i < adev->num_ip_blocks; i++) {
2071 if (!adev->ip_blocks[i].status.valid)
2073 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2074 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2075 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH)
2077 r = adev->ip_blocks[i].version->funcs->resume(adev);
2079 DRM_ERROR("resume of IP block <%s> failed %d\n",
2080 adev->ip_blocks[i].version->funcs->name, r);
2089 * amdgpu_device_ip_resume - run resume for hardware IPs
2091 * @adev: amdgpu_device pointer
2093 * Main resume function for hardware IPs. The hardware IPs
2094 * are split into two resume functions because they are
2095 * are also used in in recovering from a GPU reset and some additional
2096 * steps need to be take between them. In this case (S3/S4) they are
2098 * Returns 0 on success, negative error code on failure.
2100 static int amdgpu_device_ip_resume(struct amdgpu_device *adev)
2104 r = amdgpu_device_ip_resume_phase1(adev);
2107 r = amdgpu_device_ip_resume_phase2(adev);
2113 * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV
2115 * @adev: amdgpu_device pointer
2117 * Query the VBIOS data tables to determine if the board supports SR-IOV.
2119 static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
2121 if (amdgpu_sriov_vf(adev)) {
2122 if (adev->is_atom_fw) {
2123 if (amdgpu_atomfirmware_gpu_supports_virtualization(adev))
2124 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2126 if (amdgpu_atombios_has_gpu_virtualization_table(adev))
2127 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2130 if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
2131 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
2136 * amdgpu_device_asic_has_dc_support - determine if DC supports the asic
2138 * @asic_type: AMD asic type
2140 * Check if there is DC (new modesetting infrastructre) support for an asic.
2141 * returns true if DC has support, false if not.
2143 bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
2145 switch (asic_type) {
2146 #if defined(CONFIG_DRM_AMD_DC)
2152 * We have systems in the wild with these ASICs that require
2153 * LVDS and VGA support which is not supported with DC.
2155 * Fallback to the non-DC driver here by default so as not to
2156 * cause regressions.
2158 return amdgpu_dc > 0;
2162 case CHIP_POLARIS10:
2163 case CHIP_POLARIS11:
2164 case CHIP_POLARIS12:
2171 #if defined(CONFIG_DRM_AMD_DC_DCN1_0)
2174 return amdgpu_dc != 0;
2182 * amdgpu_device_has_dc_support - check if dc is supported
2184 * @adev: amdgpu_device_pointer
2186 * Returns true for supported, false for not supported
2188 bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
2190 if (amdgpu_sriov_vf(adev))
2193 return amdgpu_device_asic_has_dc_support(adev->asic_type);
2197 * amdgpu_device_init - initialize the driver
2199 * @adev: amdgpu_device pointer
2200 * @ddev: drm dev pointer
2201 * @pdev: pci dev pointer
2202 * @flags: driver flags
2204 * Initializes the driver info and hw (all asics).
2205 * Returns 0 for success or an error on failure.
2206 * Called at driver startup.
2208 int amdgpu_device_init(struct amdgpu_device *adev,
2209 struct drm_device *ddev,
2210 struct pci_dev *pdev,
2214 bool runtime = false;
2217 adev->shutdown = false;
2218 adev->dev = &pdev->dev;
2221 adev->flags = flags;
2222 adev->asic_type = flags & AMD_ASIC_MASK;
2223 adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
2224 if (amdgpu_emu_mode == 1)
2225 adev->usec_timeout *= 2;
2226 adev->gmc.gart_size = 512 * 1024 * 1024;
2227 adev->accel_working = false;
2228 adev->num_rings = 0;
2229 adev->mman.buffer_funcs = NULL;
2230 adev->mman.buffer_funcs_ring = NULL;
2231 adev->vm_manager.vm_pte_funcs = NULL;
2232 adev->vm_manager.vm_pte_num_rqs = 0;
2233 adev->gmc.gmc_funcs = NULL;
2234 adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
2235 bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
2237 adev->smc_rreg = &amdgpu_invalid_rreg;
2238 adev->smc_wreg = &amdgpu_invalid_wreg;
2239 adev->pcie_rreg = &amdgpu_invalid_rreg;
2240 adev->pcie_wreg = &amdgpu_invalid_wreg;
2241 adev->pciep_rreg = &amdgpu_invalid_rreg;
2242 adev->pciep_wreg = &amdgpu_invalid_wreg;
2243 adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
2244 adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
2245 adev->didt_rreg = &amdgpu_invalid_rreg;
2246 adev->didt_wreg = &amdgpu_invalid_wreg;
2247 adev->gc_cac_rreg = &amdgpu_invalid_rreg;
2248 adev->gc_cac_wreg = &amdgpu_invalid_wreg;
2249 adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
2250 adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
2252 DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
2253 amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
2254 pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
2256 /* mutex initialization are all done here so we
2257 * can recall function without having locking issues */
2258 atomic_set(&adev->irq.ih.lock, 0);
2259 mutex_init(&adev->firmware.mutex);
2260 mutex_init(&adev->pm.mutex);
2261 mutex_init(&adev->gfx.gpu_clock_mutex);
2262 mutex_init(&adev->srbm_mutex);
2263 mutex_init(&adev->gfx.pipe_reserve_mutex);
2264 mutex_init(&adev->gfx.gfx_off_mutex);
2265 mutex_init(&adev->grbm_idx_mutex);
2266 mutex_init(&adev->mn_lock);
2267 mutex_init(&adev->virt.vf_errors.lock);
2268 hash_init(adev->mn_hash);
2269 mutex_init(&adev->lock_reset);
2271 amdgpu_device_check_arguments(adev);
2273 spin_lock_init(&adev->mmio_idx_lock);
2274 spin_lock_init(&adev->smc_idx_lock);
2275 spin_lock_init(&adev->pcie_idx_lock);
2276 spin_lock_init(&adev->uvd_ctx_idx_lock);
2277 spin_lock_init(&adev->didt_idx_lock);
2278 spin_lock_init(&adev->gc_cac_idx_lock);
2279 spin_lock_init(&adev->se_cac_idx_lock);
2280 spin_lock_init(&adev->audio_endpt_idx_lock);
2281 spin_lock_init(&adev->mm_stats.lock);
2283 INIT_LIST_HEAD(&adev->shadow_list);
2284 mutex_init(&adev->shadow_list_lock);
2286 INIT_LIST_HEAD(&adev->ring_lru_list);
2287 spin_lock_init(&adev->ring_lru_list_lock);
2289 INIT_DELAYED_WORK(&adev->late_init_work,
2290 amdgpu_device_ip_late_init_func_handler);
2291 INIT_DELAYED_WORK(&adev->gfx.gfx_off_delay_work,
2292 amdgpu_device_delay_enable_gfx_off);
2294 adev->gfx.gfx_off_req_count = 1;
2295 adev->pm.ac_power = power_supply_is_system_supplied() > 0 ? true : false;
2297 /* Registers mapping */
2298 /* TODO: block userspace mapping of io register */
2299 if (adev->asic_type >= CHIP_BONAIRE) {
2300 adev->rmmio_base = pci_resource_start(adev->pdev, 5);
2301 adev->rmmio_size = pci_resource_len(adev->pdev, 5);
2303 adev->rmmio_base = pci_resource_start(adev->pdev, 2);
2304 adev->rmmio_size = pci_resource_len(adev->pdev, 2);
2307 adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
2308 if (adev->rmmio == NULL) {
2311 DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
2312 DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
2314 /* doorbell bar mapping */
2315 amdgpu_device_doorbell_init(adev);
2317 /* io port mapping */
2318 for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
2319 if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) {
2320 adev->rio_mem_size = pci_resource_len(adev->pdev, i);
2321 adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size);
2325 if (adev->rio_mem == NULL)
2326 DRM_INFO("PCI I/O BAR is not found.\n");
2328 amdgpu_device_get_pcie_info(adev);
2330 /* early init functions */
2331 r = amdgpu_device_ip_early_init(adev);
2335 /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
2336 /* this will fail for cards that aren't VGA class devices, just
2338 vga_client_register(adev->pdev, adev, NULL, amdgpu_device_vga_set_decode);
2340 if (amdgpu_device_is_px(ddev))
2342 if (!pci_is_thunderbolt_attached(adev->pdev))
2343 vga_switcheroo_register_client(adev->pdev,
2344 &amdgpu_switcheroo_ops, runtime);
2346 vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
2348 if (amdgpu_emu_mode == 1) {
2349 /* post the asic on emulation mode */
2350 emu_soc_asic_init(adev);
2351 goto fence_driver_init;
2355 if (!amdgpu_get_bios(adev)) {
2360 r = amdgpu_atombios_init(adev);
2362 dev_err(adev->dev, "amdgpu_atombios_init failed\n");
2363 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0);
2367 /* detect if we are with an SRIOV vbios */
2368 amdgpu_device_detect_sriov_bios(adev);
2370 /* Post card if necessary */
2371 if (amdgpu_device_need_post(adev)) {
2373 dev_err(adev->dev, "no vBIOS found\n");
2377 DRM_INFO("GPU posting now...\n");
2378 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2380 dev_err(adev->dev, "gpu post error!\n");
2385 if (adev->is_atom_fw) {
2386 /* Initialize clocks */
2387 r = amdgpu_atomfirmware_get_clock_info(adev);
2389 dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
2390 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
2394 /* Initialize clocks */
2395 r = amdgpu_atombios_get_clock_info(adev);
2397 dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
2398 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
2401 /* init i2c buses */
2402 if (!amdgpu_device_has_dc_support(adev))
2403 amdgpu_atombios_i2c_init(adev);
2408 r = amdgpu_fence_driver_init(adev);
2410 dev_err(adev->dev, "amdgpu_fence_driver_init failed\n");
2411 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
2415 /* init the mode config */
2416 drm_mode_config_init(adev->ddev);
2418 r = amdgpu_device_ip_init(adev);
2420 /* failed in exclusive mode due to timeout */
2421 if (amdgpu_sriov_vf(adev) &&
2422 !amdgpu_sriov_runtime(adev) &&
2423 amdgpu_virt_mmio_blocked(adev) &&
2424 !amdgpu_virt_wait_reset(adev)) {
2425 dev_err(adev->dev, "VF exclusive mode timeout\n");
2426 /* Don't send request since VF is inactive. */
2427 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
2428 adev->virt.ops = NULL;
2432 dev_err(adev->dev, "amdgpu_device_ip_init failed\n");
2433 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
2437 adev->accel_working = true;
2439 amdgpu_vm_check_compute_bug(adev);
2441 /* Initialize the buffer migration limit. */
2442 if (amdgpu_moverate >= 0)
2443 max_MBps = amdgpu_moverate;
2445 max_MBps = 8; /* Allow 8 MB/s. */
2446 /* Get a log2 for easy divisions. */
2447 adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
2449 r = amdgpu_ib_pool_init(adev);
2451 dev_err(adev->dev, "IB initialization failed (%d).\n", r);
2452 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r);
2456 if (amdgpu_sriov_vf(adev))
2457 amdgpu_virt_init_data_exchange(adev);
2459 amdgpu_fbdev_init(adev);
2461 r = amdgpu_pm_sysfs_init(adev);
2463 DRM_ERROR("registering pm debugfs failed (%d).\n", r);
2465 r = amdgpu_debugfs_gem_init(adev);
2467 DRM_ERROR("registering gem debugfs failed (%d).\n", r);
2469 r = amdgpu_debugfs_regs_init(adev);
2471 DRM_ERROR("registering register debugfs failed (%d).\n", r);
2473 r = amdgpu_debugfs_firmware_init(adev);
2475 DRM_ERROR("registering firmware debugfs failed (%d).\n", r);
2477 r = amdgpu_debugfs_init(adev);
2479 DRM_ERROR("Creating debugfs files failed (%d).\n", r);
2481 if ((amdgpu_testing & 1)) {
2482 if (adev->accel_working)
2483 amdgpu_test_moves(adev);
2485 DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n");
2487 if (amdgpu_benchmarking) {
2488 if (adev->accel_working)
2489 amdgpu_benchmark(adev, amdgpu_benchmarking);
2491 DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n");
2494 /* enable clockgating, etc. after ib tests, etc. since some blocks require
2495 * explicit gating rather than handling it automatically.
2497 r = amdgpu_device_ip_late_init(adev);
2499 dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n");
2500 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
2507 amdgpu_vf_error_trans_all(adev);
2509 vga_switcheroo_fini_domain_pm_ops(adev->dev);
2515 * amdgpu_device_fini - tear down the driver
2517 * @adev: amdgpu_device pointer
2519 * Tear down the driver info (all asics).
2520 * Called at driver shutdown.
2522 void amdgpu_device_fini(struct amdgpu_device *adev)
2526 DRM_INFO("amdgpu: finishing device.\n");
2527 adev->shutdown = true;
2528 /* disable all interrupts */
2529 amdgpu_irq_disable_all(adev);
2530 if (adev->mode_info.mode_config_initialized){
2531 if (!amdgpu_device_has_dc_support(adev))
2532 drm_crtc_force_disable_all(adev->ddev);
2534 drm_atomic_helper_shutdown(adev->ddev);
2536 amdgpu_ib_pool_fini(adev);
2537 amdgpu_fence_driver_fini(adev);
2538 amdgpu_pm_sysfs_fini(adev);
2539 amdgpu_fbdev_fini(adev);
2540 r = amdgpu_device_ip_fini(adev);
2541 if (adev->firmware.gpu_info_fw) {
2542 release_firmware(adev->firmware.gpu_info_fw);
2543 adev->firmware.gpu_info_fw = NULL;
2545 adev->accel_working = false;
2546 cancel_delayed_work_sync(&adev->late_init_work);
2547 /* free i2c buses */
2548 if (!amdgpu_device_has_dc_support(adev))
2549 amdgpu_i2c_fini(adev);
2551 if (amdgpu_emu_mode != 1)
2552 amdgpu_atombios_fini(adev);
2556 if (!pci_is_thunderbolt_attached(adev->pdev))
2557 vga_switcheroo_unregister_client(adev->pdev);
2558 if (adev->flags & AMD_IS_PX)
2559 vga_switcheroo_fini_domain_pm_ops(adev->dev);
2560 vga_client_register(adev->pdev, NULL, NULL, NULL);
2562 pci_iounmap(adev->pdev, adev->rio_mem);
2563 adev->rio_mem = NULL;
2564 iounmap(adev->rmmio);
2566 amdgpu_device_doorbell_fini(adev);
2567 amdgpu_debugfs_regs_cleanup(adev);
2575 * amdgpu_device_suspend - initiate device suspend
2577 * @dev: drm dev pointer
2578 * @suspend: suspend state
2579 * @fbcon : notify the fbdev of suspend
2581 * Puts the hw in the suspend state (all asics).
2582 * Returns 0 for success or an error on failure.
2583 * Called at driver suspend.
2585 int amdgpu_device_suspend(struct drm_device *dev, bool suspend, bool fbcon)
2587 struct amdgpu_device *adev;
2588 struct drm_crtc *crtc;
2589 struct drm_connector *connector;
2592 if (dev == NULL || dev->dev_private == NULL) {
2596 adev = dev->dev_private;
2598 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
2601 drm_kms_helper_poll_disable(dev);
2604 amdgpu_fbdev_set_suspend(adev, 1);
2606 cancel_delayed_work_sync(&adev->late_init_work);
2608 if (!amdgpu_device_has_dc_support(adev)) {
2609 /* turn off display hw */
2610 drm_modeset_lock_all(dev);
2611 list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2612 drm_helper_connector_dpms(connector, DRM_MODE_DPMS_OFF);
2614 drm_modeset_unlock_all(dev);
2615 /* unpin the front buffers and cursors */
2616 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2617 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2618 struct drm_framebuffer *fb = crtc->primary->fb;
2619 struct amdgpu_bo *robj;
2621 if (amdgpu_crtc->cursor_bo) {
2622 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2623 r = amdgpu_bo_reserve(aobj, true);
2625 amdgpu_bo_unpin(aobj);
2626 amdgpu_bo_unreserve(aobj);
2630 if (fb == NULL || fb->obj[0] == NULL) {
2633 robj = gem_to_amdgpu_bo(fb->obj[0]);
2634 /* don't unpin kernel fb objects */
2635 if (!amdgpu_fbdev_robj_is_fb(adev, robj)) {
2636 r = amdgpu_bo_reserve(robj, true);
2638 amdgpu_bo_unpin(robj);
2639 amdgpu_bo_unreserve(robj);
2645 amdgpu_amdkfd_suspend(adev);
2647 r = amdgpu_device_ip_suspend_phase1(adev);
2649 /* evict vram memory */
2650 amdgpu_bo_evict_vram(adev);
2652 amdgpu_fence_driver_suspend(adev);
2654 r = amdgpu_device_ip_suspend_phase2(adev);
2656 /* evict remaining vram memory
2657 * This second call to evict vram is to evict the gart page table
2660 amdgpu_bo_evict_vram(adev);
2662 pci_save_state(dev->pdev);
2664 /* Shut down the device */
2665 pci_disable_device(dev->pdev);
2666 pci_set_power_state(dev->pdev, PCI_D3hot);
2668 r = amdgpu_asic_reset(adev);
2670 DRM_ERROR("amdgpu asic reset failed\n");
2677 * amdgpu_device_resume - initiate device resume
2679 * @dev: drm dev pointer
2680 * @resume: resume state
2681 * @fbcon : notify the fbdev of resume
2683 * Bring the hw back to operating state (all asics).
2684 * Returns 0 for success or an error on failure.
2685 * Called at driver resume.
2687 int amdgpu_device_resume(struct drm_device *dev, bool resume, bool fbcon)
2689 struct drm_connector *connector;
2690 struct amdgpu_device *adev = dev->dev_private;
2691 struct drm_crtc *crtc;
2694 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
2698 pci_set_power_state(dev->pdev, PCI_D0);
2699 pci_restore_state(dev->pdev);
2700 r = pci_enable_device(dev->pdev);
2706 if (amdgpu_device_need_post(adev)) {
2707 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2709 DRM_ERROR("amdgpu asic init failed\n");
2712 r = amdgpu_device_ip_resume(adev);
2714 DRM_ERROR("amdgpu_device_ip_resume failed (%d).\n", r);
2717 amdgpu_fence_driver_resume(adev);
2720 r = amdgpu_device_ip_late_init(adev);
2724 if (!amdgpu_device_has_dc_support(adev)) {
2726 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2727 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2729 if (amdgpu_crtc->cursor_bo) {
2730 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2731 r = amdgpu_bo_reserve(aobj, true);
2733 r = amdgpu_bo_pin(aobj, AMDGPU_GEM_DOMAIN_VRAM);
2735 DRM_ERROR("Failed to pin cursor BO (%d)\n", r);
2736 amdgpu_crtc->cursor_addr = amdgpu_bo_gpu_offset(aobj);
2737 amdgpu_bo_unreserve(aobj);
2742 r = amdgpu_amdkfd_resume(adev);
2746 /* Make sure IB tests flushed */
2747 flush_delayed_work(&adev->late_init_work);
2749 /* blat the mode back in */
2751 if (!amdgpu_device_has_dc_support(adev)) {
2753 drm_helper_resume_force_mode(dev);
2755 /* turn on display hw */
2756 drm_modeset_lock_all(dev);
2757 list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2758 drm_helper_connector_dpms(connector, DRM_MODE_DPMS_ON);
2760 drm_modeset_unlock_all(dev);
2762 amdgpu_fbdev_set_suspend(adev, 0);
2765 drm_kms_helper_poll_enable(dev);
2768 * Most of the connector probing functions try to acquire runtime pm
2769 * refs to ensure that the GPU is powered on when connector polling is
2770 * performed. Since we're calling this from a runtime PM callback,
2771 * trying to acquire rpm refs will cause us to deadlock.
2773 * Since we're guaranteed to be holding the rpm lock, it's safe to
2774 * temporarily disable the rpm helpers so this doesn't deadlock us.
2777 dev->dev->power.disable_depth++;
2779 if (!amdgpu_device_has_dc_support(adev))
2780 drm_helper_hpd_irq_event(dev);
2782 drm_kms_helper_hotplug_event(dev);
2784 dev->dev->power.disable_depth--;
2790 * amdgpu_device_ip_check_soft_reset - did soft reset succeed
2792 * @adev: amdgpu_device pointer
2794 * The list of all the hardware IPs that make up the asic is walked and
2795 * the check_soft_reset callbacks are run. check_soft_reset determines
2796 * if the asic is still hung or not.
2797 * Returns true if any of the IPs are still in a hung state, false if not.
2799 static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)
2802 bool asic_hang = false;
2804 if (amdgpu_sriov_vf(adev))
2807 if (amdgpu_asic_need_full_reset(adev))
2810 for (i = 0; i < adev->num_ip_blocks; i++) {
2811 if (!adev->ip_blocks[i].status.valid)
2813 if (adev->ip_blocks[i].version->funcs->check_soft_reset)
2814 adev->ip_blocks[i].status.hang =
2815 adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
2816 if (adev->ip_blocks[i].status.hang) {
2817 DRM_INFO("IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
2825 * amdgpu_device_ip_pre_soft_reset - prepare for soft reset
2827 * @adev: amdgpu_device pointer
2829 * The list of all the hardware IPs that make up the asic is walked and the
2830 * pre_soft_reset callbacks are run if the block is hung. pre_soft_reset
2831 * handles any IP specific hardware or software state changes that are
2832 * necessary for a soft reset to succeed.
2833 * Returns 0 on success, negative error code on failure.
2835 static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)
2839 for (i = 0; i < adev->num_ip_blocks; i++) {
2840 if (!adev->ip_blocks[i].status.valid)
2842 if (adev->ip_blocks[i].status.hang &&
2843 adev->ip_blocks[i].version->funcs->pre_soft_reset) {
2844 r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
2854 * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed
2856 * @adev: amdgpu_device pointer
2858 * Some hardware IPs cannot be soft reset. If they are hung, a full gpu
2859 * reset is necessary to recover.
2860 * Returns true if a full asic reset is required, false if not.
2862 static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)
2866 if (amdgpu_asic_need_full_reset(adev))
2869 for (i = 0; i < adev->num_ip_blocks; i++) {
2870 if (!adev->ip_blocks[i].status.valid)
2872 if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
2873 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
2874 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
2875 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
2876 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
2877 if (adev->ip_blocks[i].status.hang) {
2878 DRM_INFO("Some block need full reset!\n");
2887 * amdgpu_device_ip_soft_reset - do a soft reset
2889 * @adev: amdgpu_device pointer
2891 * The list of all the hardware IPs that make up the asic is walked and the
2892 * soft_reset callbacks are run if the block is hung. soft_reset handles any
2893 * IP specific hardware or software state changes that are necessary to soft
2895 * Returns 0 on success, negative error code on failure.
2897 static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)
2901 for (i = 0; i < adev->num_ip_blocks; i++) {
2902 if (!adev->ip_blocks[i].status.valid)
2904 if (adev->ip_blocks[i].status.hang &&
2905 adev->ip_blocks[i].version->funcs->soft_reset) {
2906 r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
2916 * amdgpu_device_ip_post_soft_reset - clean up from soft reset
2918 * @adev: amdgpu_device pointer
2920 * The list of all the hardware IPs that make up the asic is walked and the
2921 * post_soft_reset callbacks are run if the asic was hung. post_soft_reset
2922 * handles any IP specific hardware or software state changes that are
2923 * necessary after the IP has been soft reset.
2924 * Returns 0 on success, negative error code on failure.
2926 static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)
2930 for (i = 0; i < adev->num_ip_blocks; i++) {
2931 if (!adev->ip_blocks[i].status.valid)
2933 if (adev->ip_blocks[i].status.hang &&
2934 adev->ip_blocks[i].version->funcs->post_soft_reset)
2935 r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
2944 * amdgpu_device_recover_vram_from_shadow - restore shadowed VRAM buffers
2946 * @adev: amdgpu_device pointer
2947 * @ring: amdgpu_ring for the engine handling the buffer operations
2948 * @bo: amdgpu_bo buffer whose shadow is being restored
2949 * @fence: dma_fence associated with the operation
2951 * Restores the VRAM buffer contents from the shadow in GTT. Used to
2952 * restore things like GPUVM page tables after a GPU reset where
2953 * the contents of VRAM might be lost.
2954 * Returns 0 on success, negative error code on failure.
2956 static int amdgpu_device_recover_vram_from_shadow(struct amdgpu_device *adev,
2957 struct amdgpu_ring *ring,
2958 struct amdgpu_bo *bo,
2959 struct dma_fence **fence)
2967 r = amdgpu_bo_reserve(bo, true);
2970 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
2971 /* if bo has been evicted, then no need to recover */
2972 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
2973 r = amdgpu_bo_validate(bo->shadow);
2975 DRM_ERROR("bo validate failed!\n");
2979 r = amdgpu_bo_restore_from_shadow(adev, ring, bo,
2982 DRM_ERROR("recover page table failed!\n");
2987 amdgpu_bo_unreserve(bo);
2992 * amdgpu_device_handle_vram_lost - Handle the loss of VRAM contents
2994 * @adev: amdgpu_device pointer
2996 * Restores the contents of VRAM buffers from the shadows in GTT. Used to
2997 * restore things like GPUVM page tables after a GPU reset where
2998 * the contents of VRAM might be lost.
2999 * Returns 0 on success, 1 on failure.
3001 static int amdgpu_device_handle_vram_lost(struct amdgpu_device *adev)
3003 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
3004 struct amdgpu_bo *bo, *tmp;
3005 struct dma_fence *fence = NULL, *next = NULL;
3010 if (amdgpu_sriov_runtime(adev))
3011 tmo = msecs_to_jiffies(8000);
3013 tmo = msecs_to_jiffies(100);
3015 DRM_INFO("recover vram bo from shadow start\n");
3016 mutex_lock(&adev->shadow_list_lock);
3017 list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) {
3019 amdgpu_device_recover_vram_from_shadow(adev, ring, bo, &next);
3021 r = dma_fence_wait_timeout(fence, false, tmo);
3023 pr_err("wait fence %p[%d] timeout\n", fence, i);
3025 pr_err("wait fence %p[%d] interrupted\n", fence, i);
3027 dma_fence_put(fence);
3034 dma_fence_put(fence);
3037 mutex_unlock(&adev->shadow_list_lock);
3040 r = dma_fence_wait_timeout(fence, false, tmo);
3042 pr_err("wait fence %p[%d] timeout\n", fence, i);
3044 pr_err("wait fence %p[%d] interrupted\n", fence, i);
3047 dma_fence_put(fence);
3050 DRM_INFO("recover vram bo from shadow done\n");
3052 DRM_ERROR("recover vram bo from shadow failed\n");
3054 return (r > 0) ? 0 : 1;
3058 * amdgpu_device_reset - reset ASIC/GPU for bare-metal or passthrough
3060 * @adev: amdgpu device pointer
3062 * attempt to do soft-reset or full-reset and reinitialize Asic
3063 * return 0 means succeeded otherwise failed
3065 static int amdgpu_device_reset(struct amdgpu_device *adev)
3067 bool need_full_reset, vram_lost = 0;
3070 need_full_reset = amdgpu_device_ip_need_full_reset(adev);
3072 if (!need_full_reset) {
3073 amdgpu_device_ip_pre_soft_reset(adev);
3074 r = amdgpu_device_ip_soft_reset(adev);
3075 amdgpu_device_ip_post_soft_reset(adev);
3076 if (r || amdgpu_device_ip_check_soft_reset(adev)) {
3077 DRM_INFO("soft reset failed, will fallback to full reset!\n");
3078 need_full_reset = true;
3082 if (need_full_reset) {
3083 r = amdgpu_device_ip_suspend(adev);
3086 r = amdgpu_asic_reset(adev);
3088 amdgpu_atom_asic_init(adev->mode_info.atom_context);
3091 dev_info(adev->dev, "GPU reset succeeded, trying to resume\n");
3092 r = amdgpu_device_ip_resume_phase1(adev);
3096 vram_lost = amdgpu_device_check_vram_lost(adev);
3098 DRM_ERROR("VRAM is lost!\n");
3099 atomic_inc(&adev->vram_lost_counter);
3102 r = amdgpu_gtt_mgr_recover(
3103 &adev->mman.bdev.man[TTM_PL_TT]);
3107 r = amdgpu_device_ip_resume_phase2(adev);
3112 amdgpu_device_fill_reset_magic(adev);
3118 amdgpu_irq_gpu_reset_resume_helper(adev);
3119 r = amdgpu_ib_ring_tests(adev);
3121 dev_err(adev->dev, "ib ring test failed (%d).\n", r);
3122 r = amdgpu_device_ip_suspend(adev);
3123 need_full_reset = true;
3128 if (!r && ((need_full_reset && !(adev->flags & AMD_IS_APU)) || vram_lost))
3129 r = amdgpu_device_handle_vram_lost(adev);
3135 * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
3137 * @adev: amdgpu device pointer
3138 * @from_hypervisor: request from hypervisor
3140 * do VF FLR and reinitialize Asic
3141 * return 0 means succeeded otherwise failed
3143 static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
3144 bool from_hypervisor)
3148 if (from_hypervisor)
3149 r = amdgpu_virt_request_full_gpu(adev, true);
3151 r = amdgpu_virt_reset_gpu(adev);
3155 /* Resume IP prior to SMC */
3156 r = amdgpu_device_ip_reinit_early_sriov(adev);
3160 /* we need recover gart prior to run SMC/CP/SDMA resume */
3161 amdgpu_gtt_mgr_recover(&adev->mman.bdev.man[TTM_PL_TT]);
3163 /* now we are okay to resume SMC/CP/SDMA */
3164 r = amdgpu_device_ip_reinit_late_sriov(adev);
3168 amdgpu_irq_gpu_reset_resume_helper(adev);
3169 r = amdgpu_ib_ring_tests(adev);
3172 amdgpu_virt_release_full_gpu(adev, true);
3173 if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
3174 atomic_inc(&adev->vram_lost_counter);
3175 r = amdgpu_device_handle_vram_lost(adev);
3182 * amdgpu_device_should_recover_gpu - check if we should try GPU recovery
3184 * @adev: amdgpu device pointer
3186 * Check amdgpu_gpu_recovery and SRIOV status to see if we should try to recover
3189 bool amdgpu_device_should_recover_gpu(struct amdgpu_device *adev)
3191 if (!amdgpu_device_ip_check_soft_reset(adev)) {
3192 DRM_INFO("Timeout, but no hardware hang detected.\n");
3196 if (amdgpu_gpu_recovery == 0 || (amdgpu_gpu_recovery == -1 &&
3197 !amdgpu_sriov_vf(adev))) {
3198 DRM_INFO("GPU recovery disabled.\n");
3206 * amdgpu_device_gpu_recover - reset the asic and recover scheduler
3208 * @adev: amdgpu device pointer
3209 * @job: which job trigger hang
3211 * Attempt to reset the GPU if it has hung (all asics).
3212 * Returns 0 for success or an error on failure.
3214 int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
3215 struct amdgpu_job *job)
3219 dev_info(adev->dev, "GPU reset begin!\n");
3221 mutex_lock(&adev->lock_reset);
3222 atomic_inc(&adev->gpu_reset_counter);
3223 adev->in_gpu_reset = 1;
3226 amdgpu_amdkfd_pre_reset(adev);
3229 resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
3231 /* block all schedulers and reset given job's ring */
3232 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3233 struct amdgpu_ring *ring = adev->rings[i];
3235 if (!ring || !ring->sched.thread)
3238 kthread_park(ring->sched.thread);
3240 if (job && job->base.sched == &ring->sched)
3243 drm_sched_hw_job_reset(&ring->sched, job ? &job->base : NULL);
3245 /* after all hw jobs are reset, hw fence is meaningless, so force_completion */
3246 amdgpu_fence_driver_force_completion(ring);
3249 if (amdgpu_sriov_vf(adev))
3250 r = amdgpu_device_reset_sriov(adev, job ? false : true);
3252 r = amdgpu_device_reset(adev);
3254 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3255 struct amdgpu_ring *ring = adev->rings[i];
3257 if (!ring || !ring->sched.thread)
3260 /* only need recovery sched of the given job's ring
3261 * or all rings (in the case @job is NULL)
3262 * after above amdgpu_reset accomplished
3264 if ((!job || job->base.sched == &ring->sched) && !r)
3265 drm_sched_job_recovery(&ring->sched);
3267 kthread_unpark(ring->sched.thread);
3270 if (!amdgpu_device_has_dc_support(adev)) {
3271 drm_helper_resume_force_mode(adev->ddev);
3274 ttm_bo_unlock_delayed_workqueue(&adev->mman.bdev, resched);
3277 /* bad news, how to tell it to userspace ? */
3278 dev_info(adev->dev, "GPU reset(%d) failed\n", atomic_read(&adev->gpu_reset_counter));
3279 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r);
3281 dev_info(adev->dev, "GPU reset(%d) succeeded!\n",atomic_read(&adev->gpu_reset_counter));
3285 amdgpu_amdkfd_post_reset(adev);
3286 amdgpu_vf_error_trans_all(adev);
3287 adev->in_gpu_reset = 0;
3288 mutex_unlock(&adev->lock_reset);
3293 * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot
3295 * @adev: amdgpu_device pointer
3297 * Fetchs and stores in the driver the PCIE capabilities (gen speed
3298 * and lanes) of the slot the device is in. Handles APUs and
3299 * virtualized environments where PCIE config space may not be available.
3301 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev)
3303 struct pci_dev *pdev;
3304 enum pci_bus_speed speed_cap;
3305 enum pcie_link_width link_width;
3307 if (amdgpu_pcie_gen_cap)
3308 adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
3310 if (amdgpu_pcie_lane_cap)
3311 adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
3313 /* covers APUs as well */
3314 if (pci_is_root_bus(adev->pdev->bus)) {
3315 if (adev->pm.pcie_gen_mask == 0)
3316 adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
3317 if (adev->pm.pcie_mlw_mask == 0)
3318 adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
3322 if (adev->pm.pcie_gen_mask == 0) {
3325 speed_cap = pcie_get_speed_cap(pdev);
3326 if (speed_cap == PCI_SPEED_UNKNOWN) {
3327 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3328 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3329 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
3331 if (speed_cap == PCIE_SPEED_16_0GT)
3332 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3333 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3334 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
3335 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4);
3336 else if (speed_cap == PCIE_SPEED_8_0GT)
3337 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3338 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3339 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
3340 else if (speed_cap == PCIE_SPEED_5_0GT)
3341 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3342 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2);
3344 adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1;
3347 pdev = adev->ddev->pdev->bus->self;
3348 speed_cap = pcie_get_speed_cap(pdev);
3349 if (speed_cap == PCI_SPEED_UNKNOWN) {
3350 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3351 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
3353 if (speed_cap == PCIE_SPEED_16_0GT)
3354 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3355 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3356 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
3357 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4);
3358 else if (speed_cap == PCIE_SPEED_8_0GT)
3359 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3360 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3361 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3);
3362 else if (speed_cap == PCIE_SPEED_5_0GT)
3363 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3364 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
3366 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
3370 if (adev->pm.pcie_mlw_mask == 0) {
3371 pdev = adev->ddev->pdev->bus->self;
3372 link_width = pcie_get_width_cap(pdev);
3373 if (link_width == PCIE_LNK_WIDTH_UNKNOWN) {
3374 adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK;
3376 switch (link_width) {
3378 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
3379 CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
3380 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3381 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3382 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3383 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3384 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3387 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
3388 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3389 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3390 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3391 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3392 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3395 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3396 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3397 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3398 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3399 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3402 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3403 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3404 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3405 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3408 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3409 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3410 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3413 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3414 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3417 adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;