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"
64 #define AMDGPU_RESUME_MS 2000
66 static const char *amdgpu_asic_name[] = {
93 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev);
96 * amdgpu_device_is_px - Is the device is a dGPU with HG/PX power control
98 * @dev: drm_device pointer
100 * Returns true if the device is a dGPU with HG/PX power control,
101 * otherwise return false.
103 bool amdgpu_device_is_px(struct drm_device *dev)
105 struct amdgpu_device *adev = dev->dev_private;
107 if (adev->flags & AMD_IS_PX)
113 * MMIO register access helper functions.
116 * amdgpu_mm_rreg - read a memory mapped IO register
118 * @adev: amdgpu_device pointer
119 * @reg: dword aligned register offset
120 * @acc_flags: access flags which require special behavior
122 * Returns the 32 bit value from the offset specified.
124 uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg,
129 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
130 return amdgpu_virt_kiq_rreg(adev, reg);
132 if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
133 ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
137 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
138 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
139 ret = readl(((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
140 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
142 trace_amdgpu_mm_rreg(adev->pdev->device, reg, ret);
147 * MMIO register read with bytes helper functions
148 * @offset:bytes offset from MMIO start
153 * amdgpu_mm_rreg8 - read a memory mapped IO register
155 * @adev: amdgpu_device pointer
156 * @offset: byte aligned register offset
158 * Returns the 8 bit value from the offset specified.
160 uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset) {
161 if (offset < adev->rmmio_size)
162 return (readb(adev->rmmio + offset));
167 * MMIO register write with bytes helper functions
168 * @offset:bytes offset from MMIO start
169 * @value: the value want to be written to the register
173 * amdgpu_mm_wreg8 - read a memory mapped IO register
175 * @adev: amdgpu_device pointer
176 * @offset: byte aligned register offset
177 * @value: 8 bit value to write
179 * Writes the value specified to the offset specified.
181 void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value) {
182 if (offset < adev->rmmio_size)
183 writeb(value, adev->rmmio + offset);
189 * amdgpu_mm_wreg - write to a memory mapped IO register
191 * @adev: amdgpu_device pointer
192 * @reg: dword aligned register offset
193 * @v: 32 bit value to write to the register
194 * @acc_flags: access flags which require special behavior
196 * Writes the value specified to the offset specified.
198 void amdgpu_mm_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v,
201 trace_amdgpu_mm_wreg(adev->pdev->device, reg, v);
203 if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
204 adev->last_mm_index = v;
207 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
208 return amdgpu_virt_kiq_wreg(adev, reg, v);
210 if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
211 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
215 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
216 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
217 writel(v, ((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
218 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
221 if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
227 * amdgpu_io_rreg - read an IO register
229 * @adev: amdgpu_device pointer
230 * @reg: dword aligned register offset
232 * Returns the 32 bit value from the offset specified.
234 u32 amdgpu_io_rreg(struct amdgpu_device *adev, u32 reg)
236 if ((reg * 4) < adev->rio_mem_size)
237 return ioread32(adev->rio_mem + (reg * 4));
239 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
240 return ioread32(adev->rio_mem + (mmMM_DATA * 4));
245 * amdgpu_io_wreg - write to an IO register
247 * @adev: amdgpu_device pointer
248 * @reg: dword aligned register offset
249 * @v: 32 bit value to write to the register
251 * Writes the value specified to the offset specified.
253 void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32 v)
255 if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
256 adev->last_mm_index = v;
259 if ((reg * 4) < adev->rio_mem_size)
260 iowrite32(v, adev->rio_mem + (reg * 4));
262 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
263 iowrite32(v, adev->rio_mem + (mmMM_DATA * 4));
266 if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
272 * amdgpu_mm_rdoorbell - read a doorbell dword
274 * @adev: amdgpu_device pointer
275 * @index: doorbell index
277 * Returns the value in the doorbell aperture at the
278 * requested doorbell index (CIK).
280 u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
282 if (index < adev->doorbell.num_doorbells) {
283 return readl(adev->doorbell.ptr + index);
285 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
291 * amdgpu_mm_wdoorbell - write a doorbell dword
293 * @adev: amdgpu_device pointer
294 * @index: doorbell index
297 * Writes @v to the doorbell aperture at the
298 * requested doorbell index (CIK).
300 void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v)
302 if (index < adev->doorbell.num_doorbells) {
303 writel(v, adev->doorbell.ptr + index);
305 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
310 * amdgpu_mm_rdoorbell64 - read a doorbell Qword
312 * @adev: amdgpu_device pointer
313 * @index: doorbell index
315 * Returns the value in the doorbell aperture at the
316 * requested doorbell index (VEGA10+).
318 u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index)
320 if (index < adev->doorbell.num_doorbells) {
321 return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index));
323 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
329 * amdgpu_mm_wdoorbell64 - write a doorbell Qword
331 * @adev: amdgpu_device pointer
332 * @index: doorbell index
335 * Writes @v to the doorbell aperture at the
336 * requested doorbell index (VEGA10+).
338 void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v)
340 if (index < adev->doorbell.num_doorbells) {
341 atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v);
343 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
348 * amdgpu_invalid_rreg - dummy reg read function
350 * @adev: amdgpu device pointer
351 * @reg: offset of register
353 * Dummy register read function. Used for register blocks
354 * that certain asics don't have (all asics).
355 * Returns the value in the register.
357 static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
359 DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
365 * amdgpu_invalid_wreg - dummy reg write function
367 * @adev: amdgpu device pointer
368 * @reg: offset of register
369 * @v: value to write to the register
371 * Dummy register read function. Used for register blocks
372 * that certain asics don't have (all asics).
374 static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
376 DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
382 * amdgpu_block_invalid_rreg - dummy reg read function
384 * @adev: amdgpu device pointer
385 * @block: offset of instance
386 * @reg: offset of register
388 * Dummy register read function. Used for register blocks
389 * that certain asics don't have (all asics).
390 * Returns the value in the register.
392 static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
393 uint32_t block, uint32_t reg)
395 DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
402 * amdgpu_block_invalid_wreg - dummy reg write function
404 * @adev: amdgpu device pointer
405 * @block: offset of instance
406 * @reg: offset of register
407 * @v: value to write to the register
409 * Dummy register read function. Used for register blocks
410 * that certain asics don't have (all asics).
412 static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
414 uint32_t reg, uint32_t v)
416 DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
422 * amdgpu_device_vram_scratch_init - allocate the VRAM scratch page
424 * @adev: amdgpu device pointer
426 * Allocates a scratch page of VRAM for use by various things in the
429 static int amdgpu_device_vram_scratch_init(struct amdgpu_device *adev)
431 return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE,
432 PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
433 &adev->vram_scratch.robj,
434 &adev->vram_scratch.gpu_addr,
435 (void **)&adev->vram_scratch.ptr);
439 * amdgpu_device_vram_scratch_fini - Free the VRAM scratch page
441 * @adev: amdgpu device pointer
443 * Frees the VRAM scratch page.
445 static void amdgpu_device_vram_scratch_fini(struct amdgpu_device *adev)
447 amdgpu_bo_free_kernel(&adev->vram_scratch.robj, NULL, NULL);
451 * amdgpu_device_program_register_sequence - program an array of registers.
453 * @adev: amdgpu_device pointer
454 * @registers: pointer to the register array
455 * @array_size: size of the register array
457 * Programs an array or registers with and and or masks.
458 * This is a helper for setting golden registers.
460 void amdgpu_device_program_register_sequence(struct amdgpu_device *adev,
461 const u32 *registers,
462 const u32 array_size)
464 u32 tmp, reg, and_mask, or_mask;
470 for (i = 0; i < array_size; i +=3) {
471 reg = registers[i + 0];
472 and_mask = registers[i + 1];
473 or_mask = registers[i + 2];
475 if (and_mask == 0xffffffff) {
487 * amdgpu_device_pci_config_reset - reset the GPU
489 * @adev: amdgpu_device pointer
491 * Resets the GPU using the pci config reset sequence.
492 * Only applicable to asics prior to vega10.
494 void amdgpu_device_pci_config_reset(struct amdgpu_device *adev)
496 pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
500 * GPU doorbell aperture helpers function.
503 * amdgpu_device_doorbell_init - Init doorbell driver information.
505 * @adev: amdgpu_device pointer
507 * Init doorbell driver information (CIK)
508 * Returns 0 on success, error on failure.
510 static int amdgpu_device_doorbell_init(struct amdgpu_device *adev)
512 /* No doorbell on SI hardware generation */
513 if (adev->asic_type < CHIP_BONAIRE) {
514 adev->doorbell.base = 0;
515 adev->doorbell.size = 0;
516 adev->doorbell.num_doorbells = 0;
517 adev->doorbell.ptr = NULL;
521 if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET)
524 /* doorbell bar mapping */
525 adev->doorbell.base = pci_resource_start(adev->pdev, 2);
526 adev->doorbell.size = pci_resource_len(adev->pdev, 2);
528 adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32),
529 AMDGPU_DOORBELL_MAX_ASSIGNMENT+1);
530 if (adev->doorbell.num_doorbells == 0)
533 adev->doorbell.ptr = ioremap(adev->doorbell.base,
534 adev->doorbell.num_doorbells *
536 if (adev->doorbell.ptr == NULL)
543 * amdgpu_device_doorbell_fini - Tear down doorbell driver information.
545 * @adev: amdgpu_device pointer
547 * Tear down doorbell driver information (CIK)
549 static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev)
551 iounmap(adev->doorbell.ptr);
552 adev->doorbell.ptr = NULL;
558 * amdgpu_device_wb_*()
559 * Writeback is the method by which the GPU updates special pages in memory
560 * with the status of certain GPU events (fences, ring pointers,etc.).
564 * amdgpu_device_wb_fini - Disable Writeback and free memory
566 * @adev: amdgpu_device pointer
568 * Disables Writeback and frees the Writeback memory (all asics).
569 * Used at driver shutdown.
571 static void amdgpu_device_wb_fini(struct amdgpu_device *adev)
573 if (adev->wb.wb_obj) {
574 amdgpu_bo_free_kernel(&adev->wb.wb_obj,
576 (void **)&adev->wb.wb);
577 adev->wb.wb_obj = NULL;
582 * amdgpu_device_wb_init- Init Writeback driver info and allocate memory
584 * @adev: amdgpu_device pointer
586 * Initializes writeback and allocates writeback memory (all asics).
587 * Used at driver startup.
588 * Returns 0 on success or an -error on failure.
590 static int amdgpu_device_wb_init(struct amdgpu_device *adev)
594 if (adev->wb.wb_obj == NULL) {
595 /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */
596 r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8,
597 PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
598 &adev->wb.wb_obj, &adev->wb.gpu_addr,
599 (void **)&adev->wb.wb);
601 dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
605 adev->wb.num_wb = AMDGPU_MAX_WB;
606 memset(&adev->wb.used, 0, sizeof(adev->wb.used));
608 /* clear wb memory */
609 memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8);
616 * amdgpu_device_wb_get - Allocate a wb entry
618 * @adev: amdgpu_device pointer
621 * Allocate a wb slot for use by the driver (all asics).
622 * Returns 0 on success or -EINVAL on failure.
624 int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb)
626 unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
628 if (offset < adev->wb.num_wb) {
629 __set_bit(offset, adev->wb.used);
630 *wb = offset << 3; /* convert to dw offset */
638 * amdgpu_device_wb_free - Free a wb entry
640 * @adev: amdgpu_device pointer
643 * Free a wb slot allocated for use by the driver (all asics)
645 void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb)
648 if (wb < adev->wb.num_wb)
649 __clear_bit(wb, adev->wb.used);
653 * amdgpu_device_vram_location - try to find VRAM location
655 * @adev: amdgpu device structure holding all necessary informations
656 * @mc: memory controller structure holding memory informations
657 * @base: base address at which to put VRAM
659 * Function will try to place VRAM at base address provided
662 void amdgpu_device_vram_location(struct amdgpu_device *adev,
663 struct amdgpu_gmc *mc, u64 base)
665 uint64_t limit = (uint64_t)amdgpu_vram_limit << 20;
667 mc->vram_start = base;
668 mc->vram_end = mc->vram_start + mc->mc_vram_size - 1;
669 if (limit && limit < mc->real_vram_size)
670 mc->real_vram_size = limit;
671 dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
672 mc->mc_vram_size >> 20, mc->vram_start,
673 mc->vram_end, mc->real_vram_size >> 20);
677 * amdgpu_device_gart_location - try to find GART location
679 * @adev: amdgpu device structure holding all necessary informations
680 * @mc: memory controller structure holding memory informations
682 * Function will place try to place GART before or after VRAM.
684 * If GART size is bigger than space left then we ajust GART size.
685 * Thus function will never fails.
687 void amdgpu_device_gart_location(struct amdgpu_device *adev,
688 struct amdgpu_gmc *mc)
690 u64 size_af, size_bf;
692 mc->gart_size += adev->pm.smu_prv_buffer_size;
694 size_af = adev->gmc.mc_mask - mc->vram_end;
695 size_bf = mc->vram_start;
696 if (size_bf > size_af) {
697 if (mc->gart_size > size_bf) {
698 dev_warn(adev->dev, "limiting GART\n");
699 mc->gart_size = size_bf;
703 if (mc->gart_size > size_af) {
704 dev_warn(adev->dev, "limiting GART\n");
705 mc->gart_size = size_af;
707 /* VCE doesn't like it when BOs cross a 4GB segment, so align
708 * the GART base on a 4GB boundary as well.
710 mc->gart_start = ALIGN(mc->vram_end + 1, 0x100000000ULL);
712 mc->gart_end = mc->gart_start + mc->gart_size - 1;
713 dev_info(adev->dev, "GART: %lluM 0x%016llX - 0x%016llX\n",
714 mc->gart_size >> 20, mc->gart_start, mc->gart_end);
718 * amdgpu_device_resize_fb_bar - try to resize FB BAR
720 * @adev: amdgpu_device pointer
722 * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not
723 * to fail, but if any of the BARs is not accessible after the size we abort
724 * driver loading by returning -ENODEV.
726 int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev)
728 u64 space_needed = roundup_pow_of_two(adev->gmc.real_vram_size);
729 u32 rbar_size = order_base_2(((space_needed >> 20) | 1)) - 1;
730 struct pci_bus *root;
731 struct resource *res;
737 if (amdgpu_sriov_vf(adev))
740 /* Check if the root BUS has 64bit memory resources */
741 root = adev->pdev->bus;
745 pci_bus_for_each_resource(root, res, i) {
746 if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) &&
747 res->start > 0x100000000ull)
751 /* Trying to resize is pointless without a root hub window above 4GB */
755 /* Disable memory decoding while we change the BAR addresses and size */
756 pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd);
757 pci_write_config_word(adev->pdev, PCI_COMMAND,
758 cmd & ~PCI_COMMAND_MEMORY);
760 /* Free the VRAM and doorbell BAR, we most likely need to move both. */
761 amdgpu_device_doorbell_fini(adev);
762 if (adev->asic_type >= CHIP_BONAIRE)
763 pci_release_resource(adev->pdev, 2);
765 pci_release_resource(adev->pdev, 0);
767 r = pci_resize_resource(adev->pdev, 0, rbar_size);
769 DRM_INFO("Not enough PCI address space for a large BAR.");
770 else if (r && r != -ENOTSUPP)
771 DRM_ERROR("Problem resizing BAR0 (%d).", r);
773 pci_assign_unassigned_bus_resources(adev->pdev->bus);
775 /* When the doorbell or fb BAR isn't available we have no chance of
778 r = amdgpu_device_doorbell_init(adev);
779 if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET))
782 pci_write_config_word(adev->pdev, PCI_COMMAND, cmd);
788 * GPU helpers function.
791 * amdgpu_device_need_post - check if the hw need post or not
793 * @adev: amdgpu_device pointer
795 * Check if the asic has been initialized (all asics) at driver startup
796 * or post is needed if hw reset is performed.
797 * Returns true if need or false if not.
799 bool amdgpu_device_need_post(struct amdgpu_device *adev)
803 if (amdgpu_sriov_vf(adev))
806 if (amdgpu_passthrough(adev)) {
807 /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
808 * some old smc fw still need driver do vPost otherwise gpu hang, while
809 * those smc fw version above 22.15 doesn't have this flaw, so we force
810 * vpost executed for smc version below 22.15
812 if (adev->asic_type == CHIP_FIJI) {
815 err = reject_firmware(&adev->pm.fw, "/*(DEBLOBBED)*/", adev->dev);
816 /* force vPost if error occured */
820 fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
821 if (fw_ver < 0x00160e00)
826 if (adev->has_hw_reset) {
827 adev->has_hw_reset = false;
831 /* bios scratch used on CIK+ */
832 if (adev->asic_type >= CHIP_BONAIRE)
833 return amdgpu_atombios_scratch_need_asic_init(adev);
835 /* check MEM_SIZE for older asics */
836 reg = amdgpu_asic_get_config_memsize(adev);
838 if ((reg != 0) && (reg != 0xffffffff))
844 /* if we get transitioned to only one device, take VGA back */
846 * amdgpu_device_vga_set_decode - enable/disable vga decode
848 * @cookie: amdgpu_device pointer
849 * @state: enable/disable vga decode
851 * Enable/disable vga decode (all asics).
852 * Returns VGA resource flags.
854 static unsigned int amdgpu_device_vga_set_decode(void *cookie, bool state)
856 struct amdgpu_device *adev = cookie;
857 amdgpu_asic_set_vga_state(adev, state);
859 return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
860 VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
862 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
866 * amdgpu_device_check_block_size - validate the vm block size
868 * @adev: amdgpu_device pointer
870 * Validates the vm block size specified via module parameter.
871 * The vm block size defines number of bits in page table versus page directory,
872 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
873 * page table and the remaining bits are in the page directory.
875 static void amdgpu_device_check_block_size(struct amdgpu_device *adev)
877 /* defines number of bits in page table versus page directory,
878 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
879 * page table and the remaining bits are in the page directory */
880 if (amdgpu_vm_block_size == -1)
883 if (amdgpu_vm_block_size < 9) {
884 dev_warn(adev->dev, "VM page table size (%d) too small\n",
885 amdgpu_vm_block_size);
886 amdgpu_vm_block_size = -1;
891 * amdgpu_device_check_vm_size - validate the vm size
893 * @adev: amdgpu_device pointer
895 * Validates the vm size in GB specified via module parameter.
896 * The VM size is the size of the GPU virtual memory space in GB.
898 static void amdgpu_device_check_vm_size(struct amdgpu_device *adev)
900 /* no need to check the default value */
901 if (amdgpu_vm_size == -1)
904 if (amdgpu_vm_size < 1) {
905 dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
911 static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev)
914 bool is_os_64 = (sizeof(void *) == 8) ? true : false;
915 uint64_t total_memory;
916 uint64_t dram_size_seven_GB = 0x1B8000000;
917 uint64_t dram_size_three_GB = 0xB8000000;
919 if (amdgpu_smu_memory_pool_size == 0)
923 DRM_WARN("Not 64-bit OS, feature not supported\n");
927 total_memory = (uint64_t)si.totalram * si.mem_unit;
929 if ((amdgpu_smu_memory_pool_size == 1) ||
930 (amdgpu_smu_memory_pool_size == 2)) {
931 if (total_memory < dram_size_three_GB)
933 } else if ((amdgpu_smu_memory_pool_size == 4) ||
934 (amdgpu_smu_memory_pool_size == 8)) {
935 if (total_memory < dram_size_seven_GB)
938 DRM_WARN("Smu memory pool size not supported\n");
941 adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28;
946 DRM_WARN("No enough system memory\n");
948 adev->pm.smu_prv_buffer_size = 0;
952 * amdgpu_device_check_arguments - validate module params
954 * @adev: amdgpu_device pointer
956 * Validates certain module parameters and updates
957 * the associated values used by the driver (all asics).
959 static void amdgpu_device_check_arguments(struct amdgpu_device *adev)
961 if (amdgpu_sched_jobs < 4) {
962 dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
964 amdgpu_sched_jobs = 4;
965 } else if (!is_power_of_2(amdgpu_sched_jobs)){
966 dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
968 amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
971 if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) {
972 /* gart size must be greater or equal to 32M */
973 dev_warn(adev->dev, "gart size (%d) too small\n",
975 amdgpu_gart_size = -1;
978 if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) {
979 /* gtt size must be greater or equal to 32M */
980 dev_warn(adev->dev, "gtt size (%d) too small\n",
982 amdgpu_gtt_size = -1;
985 /* valid range is between 4 and 9 inclusive */
986 if (amdgpu_vm_fragment_size != -1 &&
987 (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) {
988 dev_warn(adev->dev, "valid range is between 4 and 9\n");
989 amdgpu_vm_fragment_size = -1;
992 amdgpu_device_check_smu_prv_buffer_size(adev);
994 amdgpu_device_check_vm_size(adev);
996 amdgpu_device_check_block_size(adev);
998 if (amdgpu_vram_page_split != -1 && (amdgpu_vram_page_split < 16 ||
999 !is_power_of_2(amdgpu_vram_page_split))) {
1000 dev_warn(adev->dev, "invalid VRAM page split (%d)\n",
1001 amdgpu_vram_page_split);
1002 amdgpu_vram_page_split = 1024;
1005 if (amdgpu_lockup_timeout == 0) {
1006 dev_warn(adev->dev, "lockup_timeout msut be > 0, adjusting to 10000\n");
1007 amdgpu_lockup_timeout = 10000;
1010 adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type);
1014 * amdgpu_switcheroo_set_state - set switcheroo state
1016 * @pdev: pci dev pointer
1017 * @state: vga_switcheroo state
1019 * Callback for the switcheroo driver. Suspends or resumes the
1020 * the asics before or after it is powered up using ACPI methods.
1022 static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, enum vga_switcheroo_state state)
1024 struct drm_device *dev = pci_get_drvdata(pdev);
1026 if (amdgpu_device_is_px(dev) && state == VGA_SWITCHEROO_OFF)
1029 if (state == VGA_SWITCHEROO_ON) {
1030 pr_info("amdgpu: switched on\n");
1031 /* don't suspend or resume card normally */
1032 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1034 amdgpu_device_resume(dev, true, true);
1036 dev->switch_power_state = DRM_SWITCH_POWER_ON;
1037 drm_kms_helper_poll_enable(dev);
1039 pr_info("amdgpu: switched off\n");
1040 drm_kms_helper_poll_disable(dev);
1041 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1042 amdgpu_device_suspend(dev, true, true);
1043 dev->switch_power_state = DRM_SWITCH_POWER_OFF;
1048 * amdgpu_switcheroo_can_switch - see if switcheroo state can change
1050 * @pdev: pci dev pointer
1052 * Callback for the switcheroo driver. Check of the switcheroo
1053 * state can be changed.
1054 * Returns true if the state can be changed, false if not.
1056 static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
1058 struct drm_device *dev = pci_get_drvdata(pdev);
1061 * FIXME: open_count is protected by drm_global_mutex but that would lead to
1062 * locking inversion with the driver load path. And the access here is
1063 * completely racy anyway. So don't bother with locking for now.
1065 return dev->open_count == 0;
1068 static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
1069 .set_gpu_state = amdgpu_switcheroo_set_state,
1071 .can_switch = amdgpu_switcheroo_can_switch,
1075 * amdgpu_device_ip_set_clockgating_state - set the CG state
1077 * @dev: amdgpu_device pointer
1078 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1079 * @state: clockgating state (gate or ungate)
1081 * Sets the requested clockgating state for all instances of
1082 * the hardware IP specified.
1083 * Returns the error code from the last instance.
1085 int amdgpu_device_ip_set_clockgating_state(void *dev,
1086 enum amd_ip_block_type block_type,
1087 enum amd_clockgating_state state)
1089 struct amdgpu_device *adev = dev;
1092 for (i = 0; i < adev->num_ip_blocks; i++) {
1093 if (!adev->ip_blocks[i].status.valid)
1095 if (adev->ip_blocks[i].version->type != block_type)
1097 if (!adev->ip_blocks[i].version->funcs->set_clockgating_state)
1099 r = adev->ip_blocks[i].version->funcs->set_clockgating_state(
1100 (void *)adev, state);
1102 DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",
1103 adev->ip_blocks[i].version->funcs->name, r);
1109 * amdgpu_device_ip_set_powergating_state - set the PG state
1111 * @dev: amdgpu_device pointer
1112 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1113 * @state: powergating state (gate or ungate)
1115 * Sets the requested powergating state for all instances of
1116 * the hardware IP specified.
1117 * Returns the error code from the last instance.
1119 int amdgpu_device_ip_set_powergating_state(void *dev,
1120 enum amd_ip_block_type block_type,
1121 enum amd_powergating_state state)
1123 struct amdgpu_device *adev = dev;
1126 for (i = 0; i < adev->num_ip_blocks; i++) {
1127 if (!adev->ip_blocks[i].status.valid)
1129 if (adev->ip_blocks[i].version->type != block_type)
1131 if (!adev->ip_blocks[i].version->funcs->set_powergating_state)
1133 r = adev->ip_blocks[i].version->funcs->set_powergating_state(
1134 (void *)adev, state);
1136 DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",
1137 adev->ip_blocks[i].version->funcs->name, r);
1143 * amdgpu_device_ip_get_clockgating_state - get the CG state
1145 * @adev: amdgpu_device pointer
1146 * @flags: clockgating feature flags
1148 * Walks the list of IPs on the device and updates the clockgating
1149 * flags for each IP.
1150 * Updates @flags with the feature flags for each hardware IP where
1151 * clockgating is enabled.
1153 void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev,
1158 for (i = 0; i < adev->num_ip_blocks; i++) {
1159 if (!adev->ip_blocks[i].status.valid)
1161 if (adev->ip_blocks[i].version->funcs->get_clockgating_state)
1162 adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags);
1167 * amdgpu_device_ip_wait_for_idle - wait for idle
1169 * @adev: amdgpu_device pointer
1170 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1172 * Waits for the request hardware IP to be idle.
1173 * Returns 0 for success or a negative error code on failure.
1175 int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev,
1176 enum amd_ip_block_type block_type)
1180 for (i = 0; i < adev->num_ip_blocks; i++) {
1181 if (!adev->ip_blocks[i].status.valid)
1183 if (adev->ip_blocks[i].version->type == block_type) {
1184 r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
1195 * amdgpu_device_ip_is_idle - is the hardware IP idle
1197 * @adev: amdgpu_device pointer
1198 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1200 * Check if the hardware IP is idle or not.
1201 * Returns true if it the IP is idle, false if not.
1203 bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev,
1204 enum amd_ip_block_type block_type)
1208 for (i = 0; i < adev->num_ip_blocks; i++) {
1209 if (!adev->ip_blocks[i].status.valid)
1211 if (adev->ip_blocks[i].version->type == block_type)
1212 return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
1219 * amdgpu_device_ip_get_ip_block - get a hw IP pointer
1221 * @adev: amdgpu_device pointer
1222 * @type: Type of hardware IP (SMU, GFX, UVD, etc.)
1224 * Returns a pointer to the hardware IP block structure
1225 * if it exists for the asic, otherwise NULL.
1227 struct amdgpu_ip_block *
1228 amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,
1229 enum amd_ip_block_type type)
1233 for (i = 0; i < adev->num_ip_blocks; i++)
1234 if (adev->ip_blocks[i].version->type == type)
1235 return &adev->ip_blocks[i];
1241 * amdgpu_device_ip_block_version_cmp
1243 * @adev: amdgpu_device pointer
1244 * @type: enum amd_ip_block_type
1245 * @major: major version
1246 * @minor: minor version
1248 * return 0 if equal or greater
1249 * return 1 if smaller or the ip_block doesn't exist
1251 int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev,
1252 enum amd_ip_block_type type,
1253 u32 major, u32 minor)
1255 struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type);
1257 if (ip_block && ((ip_block->version->major > major) ||
1258 ((ip_block->version->major == major) &&
1259 (ip_block->version->minor >= minor))))
1266 * amdgpu_device_ip_block_add
1268 * @adev: amdgpu_device pointer
1269 * @ip_block_version: pointer to the IP to add
1271 * Adds the IP block driver information to the collection of IPs
1274 int amdgpu_device_ip_block_add(struct amdgpu_device *adev,
1275 const struct amdgpu_ip_block_version *ip_block_version)
1277 if (!ip_block_version)
1280 DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks,
1281 ip_block_version->funcs->name);
1283 adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
1289 * amdgpu_device_enable_virtual_display - enable virtual display feature
1291 * @adev: amdgpu_device pointer
1293 * Enabled the virtual display feature if the user has enabled it via
1294 * the module parameter virtual_display. This feature provides a virtual
1295 * display hardware on headless boards or in virtualized environments.
1296 * This function parses and validates the configuration string specified by
1297 * the user and configues the virtual display configuration (number of
1298 * virtual connectors, crtcs, etc.) specified.
1300 static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
1302 adev->enable_virtual_display = false;
1304 if (amdgpu_virtual_display) {
1305 struct drm_device *ddev = adev->ddev;
1306 const char *pci_address_name = pci_name(ddev->pdev);
1307 char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
1309 pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
1310 pciaddstr_tmp = pciaddstr;
1311 while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
1312 pciaddname = strsep(&pciaddname_tmp, ",");
1313 if (!strcmp("all", pciaddname)
1314 || !strcmp(pci_address_name, pciaddname)) {
1318 adev->enable_virtual_display = true;
1321 res = kstrtol(pciaddname_tmp, 10,
1329 adev->mode_info.num_crtc = num_crtc;
1331 adev->mode_info.num_crtc = 1;
1337 DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
1338 amdgpu_virtual_display, pci_address_name,
1339 adev->enable_virtual_display, adev->mode_info.num_crtc);
1346 * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware
1348 * @adev: amdgpu_device pointer
1350 * Parses the asic configuration parameters specified in the gpu info
1351 * firmware and makes them availale to the driver for use in configuring
1353 * Returns 0 on success, -EINVAL on failure.
1355 static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev)
1357 const char *chip_name;
1360 const struct gpu_info_firmware_header_v1_0 *hdr;
1362 adev->firmware.gpu_info_fw = NULL;
1364 switch (adev->asic_type) {
1368 case CHIP_POLARIS10:
1369 case CHIP_POLARIS11:
1370 case CHIP_POLARIS12:
1374 #ifdef CONFIG_DRM_AMDGPU_SI
1381 #ifdef CONFIG_DRM_AMDGPU_CIK
1392 chip_name = "vega10";
1395 chip_name = "vega12";
1398 chip_name = "raven";
1402 snprintf(fw_name, sizeof(fw_name), "/*(DEBLOBBED)*/", chip_name);
1403 err = reject_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev);
1406 "Failed to load gpu_info firmware \"%s\"\n",
1410 err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw);
1413 "Failed to validate gpu_info firmware \"%s\"\n",
1418 hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data;
1419 amdgpu_ucode_print_gpu_info_hdr(&hdr->header);
1421 switch (hdr->version_major) {
1424 const struct gpu_info_firmware_v1_0 *gpu_info_fw =
1425 (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data +
1426 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1428 adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se);
1429 adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh);
1430 adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se);
1431 adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se);
1432 adev->gfx.config.max_texture_channel_caches =
1433 le32_to_cpu(gpu_info_fw->gc_num_tccs);
1434 adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs);
1435 adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds);
1436 adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth);
1437 adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth);
1438 adev->gfx.config.double_offchip_lds_buf =
1439 le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer);
1440 adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size);
1441 adev->gfx.cu_info.max_waves_per_simd =
1442 le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd);
1443 adev->gfx.cu_info.max_scratch_slots_per_cu =
1444 le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu);
1445 adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size);
1450 "Unsupported gpu_info table %d\n", hdr->header.ucode_version);
1459 * amdgpu_device_ip_early_init - run early init for hardware IPs
1461 * @adev: amdgpu_device pointer
1463 * Early initialization pass for hardware IPs. The hardware IPs that make
1464 * up each asic are discovered each IP's early_init callback is run. This
1465 * is the first stage in initializing the asic.
1466 * Returns 0 on success, negative error code on failure.
1468 static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
1472 amdgpu_device_enable_virtual_display(adev);
1474 switch (adev->asic_type) {
1478 case CHIP_POLARIS10:
1479 case CHIP_POLARIS11:
1480 case CHIP_POLARIS12:
1484 if (adev->asic_type == CHIP_CARRIZO || adev->asic_type == CHIP_STONEY)
1485 adev->family = AMDGPU_FAMILY_CZ;
1487 adev->family = AMDGPU_FAMILY_VI;
1489 r = vi_set_ip_blocks(adev);
1493 #ifdef CONFIG_DRM_AMDGPU_SI
1499 adev->family = AMDGPU_FAMILY_SI;
1500 r = si_set_ip_blocks(adev);
1505 #ifdef CONFIG_DRM_AMDGPU_CIK
1511 if ((adev->asic_type == CHIP_BONAIRE) || (adev->asic_type == CHIP_HAWAII))
1512 adev->family = AMDGPU_FAMILY_CI;
1514 adev->family = AMDGPU_FAMILY_KV;
1516 r = cik_set_ip_blocks(adev);
1525 if (adev->asic_type == CHIP_RAVEN)
1526 adev->family = AMDGPU_FAMILY_RV;
1528 adev->family = AMDGPU_FAMILY_AI;
1530 r = soc15_set_ip_blocks(adev);
1535 /* FIXME: not supported yet */
1539 r = amdgpu_device_parse_gpu_info_fw(adev);
1543 amdgpu_amdkfd_device_probe(adev);
1545 if (amdgpu_sriov_vf(adev)) {
1546 r = amdgpu_virt_request_full_gpu(adev, true);
1551 adev->powerplay.pp_feature = amdgpu_pp_feature_mask;
1553 for (i = 0; i < adev->num_ip_blocks; i++) {
1554 if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
1555 DRM_ERROR("disabled ip block: %d <%s>\n",
1556 i, adev->ip_blocks[i].version->funcs->name);
1557 adev->ip_blocks[i].status.valid = false;
1559 if (adev->ip_blocks[i].version->funcs->early_init) {
1560 r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
1562 adev->ip_blocks[i].status.valid = false;
1564 DRM_ERROR("early_init of IP block <%s> failed %d\n",
1565 adev->ip_blocks[i].version->funcs->name, r);
1568 adev->ip_blocks[i].status.valid = true;
1571 adev->ip_blocks[i].status.valid = true;
1576 adev->cg_flags &= amdgpu_cg_mask;
1577 adev->pg_flags &= amdgpu_pg_mask;
1583 * amdgpu_device_ip_init - run init for hardware IPs
1585 * @adev: amdgpu_device pointer
1587 * Main initialization pass for hardware IPs. The list of all the hardware
1588 * IPs that make up the asic is walked and the sw_init and hw_init callbacks
1589 * are run. sw_init initializes the software state associated with each IP
1590 * and hw_init initializes the hardware associated with each IP.
1591 * Returns 0 on success, negative error code on failure.
1593 static int amdgpu_device_ip_init(struct amdgpu_device *adev)
1597 for (i = 0; i < adev->num_ip_blocks; i++) {
1598 if (!adev->ip_blocks[i].status.valid)
1600 r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
1602 DRM_ERROR("sw_init of IP block <%s> failed %d\n",
1603 adev->ip_blocks[i].version->funcs->name, r);
1606 adev->ip_blocks[i].status.sw = true;
1608 /* need to do gmc hw init early so we can allocate gpu mem */
1609 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1610 r = amdgpu_device_vram_scratch_init(adev);
1612 DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r);
1615 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1617 DRM_ERROR("hw_init %d failed %d\n", i, r);
1620 r = amdgpu_device_wb_init(adev);
1622 DRM_ERROR("amdgpu_device_wb_init failed %d\n", r);
1625 adev->ip_blocks[i].status.hw = true;
1627 /* right after GMC hw init, we create CSA */
1628 if (amdgpu_sriov_vf(adev)) {
1629 r = amdgpu_allocate_static_csa(adev);
1631 DRM_ERROR("allocate CSA failed %d\n", r);
1638 for (i = 0; i < adev->num_ip_blocks; i++) {
1639 if (!adev->ip_blocks[i].status.sw)
1641 if (adev->ip_blocks[i].status.hw)
1643 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1645 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1646 adev->ip_blocks[i].version->funcs->name, r);
1649 adev->ip_blocks[i].status.hw = true;
1652 amdgpu_amdkfd_device_init(adev);
1654 if (amdgpu_sriov_vf(adev)) {
1655 amdgpu_virt_init_data_exchange(adev);
1656 amdgpu_virt_release_full_gpu(adev, true);
1663 * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer
1665 * @adev: amdgpu_device pointer
1667 * Writes a reset magic value to the gart pointer in VRAM. The driver calls
1668 * this function before a GPU reset. If the value is retained after a
1669 * GPU reset, VRAM has not been lost. Some GPU resets may destry VRAM contents.
1671 static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev)
1673 memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
1677 * amdgpu_device_check_vram_lost - check if vram is valid
1679 * @adev: amdgpu_device pointer
1681 * Checks the reset magic value written to the gart pointer in VRAM.
1682 * The driver calls this after a GPU reset to see if the contents of
1683 * VRAM is lost or now.
1684 * returns true if vram is lost, false if not.
1686 static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev)
1688 return !!memcmp(adev->gart.ptr, adev->reset_magic,
1689 AMDGPU_RESET_MAGIC_NUM);
1693 * amdgpu_device_ip_late_set_cg_state - late init for clockgating
1695 * @adev: amdgpu_device pointer
1697 * Late initialization pass enabling clockgating for hardware IPs.
1698 * The list of all the hardware IPs that make up the asic is walked and the
1699 * set_clockgating_state callbacks are run. This stage is run late
1700 * in the init process.
1701 * Returns 0 on success, negative error code on failure.
1703 static int amdgpu_device_ip_late_set_cg_state(struct amdgpu_device *adev)
1707 if (amdgpu_emu_mode == 1)
1710 for (i = 0; i < adev->num_ip_blocks; i++) {
1711 if (!adev->ip_blocks[i].status.valid)
1713 /* skip CG for VCE/UVD, it's handled specially */
1714 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1715 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1716 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
1717 adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1718 /* enable clockgating to save power */
1719 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1722 DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
1723 adev->ip_blocks[i].version->funcs->name, r);
1732 static int amdgpu_device_ip_late_set_pg_state(struct amdgpu_device *adev)
1736 if (amdgpu_emu_mode == 1)
1739 for (i = 0; i < adev->num_ip_blocks; i++) {
1740 if (!adev->ip_blocks[i].status.valid)
1742 /* skip CG for VCE/UVD, it's handled specially */
1743 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1744 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1745 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
1746 adev->ip_blocks[i].version->funcs->set_powergating_state) {
1747 /* enable powergating to save power */
1748 r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev,
1751 DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n",
1752 adev->ip_blocks[i].version->funcs->name, r);
1761 * amdgpu_device_ip_late_init - run late init for hardware IPs
1763 * @adev: amdgpu_device pointer
1765 * Late initialization pass for hardware IPs. The list of all the hardware
1766 * IPs that make up the asic is walked and the late_init callbacks are run.
1767 * late_init covers any special initialization that an IP requires
1768 * after all of the have been initialized or something that needs to happen
1769 * late in the init process.
1770 * Returns 0 on success, negative error code on failure.
1772 static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
1776 for (i = 0; i < adev->num_ip_blocks; i++) {
1777 if (!adev->ip_blocks[i].status.valid)
1779 if (adev->ip_blocks[i].version->funcs->late_init) {
1780 r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
1782 DRM_ERROR("late_init of IP block <%s> failed %d\n",
1783 adev->ip_blocks[i].version->funcs->name, r);
1786 adev->ip_blocks[i].status.late_initialized = true;
1790 amdgpu_device_ip_late_set_cg_state(adev);
1791 amdgpu_device_ip_late_set_pg_state(adev);
1793 queue_delayed_work(system_wq, &adev->late_init_work,
1794 msecs_to_jiffies(AMDGPU_RESUME_MS));
1796 amdgpu_device_fill_reset_magic(adev);
1802 * amdgpu_device_ip_fini - run fini for hardware IPs
1804 * @adev: amdgpu_device pointer
1806 * Main teardown pass for hardware IPs. The list of all the hardware
1807 * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks
1808 * are run. hw_fini tears down the hardware associated with each IP
1809 * and sw_fini tears down any software state associated with each IP.
1810 * Returns 0 on success, negative error code on failure.
1812 static int amdgpu_device_ip_fini(struct amdgpu_device *adev)
1816 amdgpu_amdkfd_device_fini(adev);
1817 /* need to disable SMC first */
1818 for (i = 0; i < adev->num_ip_blocks; i++) {
1819 if (!adev->ip_blocks[i].status.hw)
1821 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC &&
1822 adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1823 /* ungate blocks before hw fini so that we can shutdown the blocks safely */
1824 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1825 AMD_CG_STATE_UNGATE);
1827 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1828 adev->ip_blocks[i].version->funcs->name, r);
1831 if (adev->powerplay.pp_funcs->set_powergating_by_smu)
1832 amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false);
1833 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
1834 /* XXX handle errors */
1836 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
1837 adev->ip_blocks[i].version->funcs->name, r);
1839 adev->ip_blocks[i].status.hw = false;
1844 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1845 if (!adev->ip_blocks[i].status.hw)
1848 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1849 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1850 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
1851 adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1852 /* ungate blocks before hw fini so that we can shutdown the blocks safely */
1853 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1854 AMD_CG_STATE_UNGATE);
1856 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1857 adev->ip_blocks[i].version->funcs->name, r);
1862 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
1863 /* XXX handle errors */
1865 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
1866 adev->ip_blocks[i].version->funcs->name, r);
1869 adev->ip_blocks[i].status.hw = false;
1873 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1874 if (!adev->ip_blocks[i].status.sw)
1877 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1878 amdgpu_free_static_csa(adev);
1879 amdgpu_device_wb_fini(adev);
1880 amdgpu_device_vram_scratch_fini(adev);
1883 r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
1884 /* XXX handle errors */
1886 DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
1887 adev->ip_blocks[i].version->funcs->name, r);
1889 adev->ip_blocks[i].status.sw = false;
1890 adev->ip_blocks[i].status.valid = false;
1893 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1894 if (!adev->ip_blocks[i].status.late_initialized)
1896 if (adev->ip_blocks[i].version->funcs->late_fini)
1897 adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
1898 adev->ip_blocks[i].status.late_initialized = false;
1901 if (amdgpu_sriov_vf(adev))
1902 if (amdgpu_virt_release_full_gpu(adev, false))
1903 DRM_ERROR("failed to release exclusive mode on fini\n");
1909 * amdgpu_device_ip_late_init_func_handler - work handler for clockgating
1911 * @work: work_struct
1913 * Work handler for amdgpu_device_ip_late_set_cg_state. We put the
1914 * clockgating setup into a worker thread to speed up driver init and
1915 * resume from suspend.
1917 static void amdgpu_device_ip_late_init_func_handler(struct work_struct *work)
1919 struct amdgpu_device *adev =
1920 container_of(work, struct amdgpu_device, late_init_work.work);
1923 r = amdgpu_ib_ring_tests(adev);
1925 DRM_ERROR("ib ring test failed (%d).\n", r);
1929 * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1)
1931 * @adev: amdgpu_device pointer
1933 * Main suspend function for hardware IPs. The list of all the hardware
1934 * IPs that make up the asic is walked, clockgating is disabled and the
1935 * suspend callbacks are run. suspend puts the hardware and software state
1936 * in each IP into a state suitable for suspend.
1937 * Returns 0 on success, negative error code on failure.
1939 static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev)
1943 if (amdgpu_sriov_vf(adev))
1944 amdgpu_virt_request_full_gpu(adev, false);
1946 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1947 if (!adev->ip_blocks[i].status.valid)
1949 /* displays are handled separately */
1950 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) {
1951 /* ungate blocks so that suspend can properly shut them down */
1952 if (adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1953 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1954 AMD_CG_STATE_UNGATE);
1956 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1957 adev->ip_blocks[i].version->funcs->name, r);
1960 /* XXX handle errors */
1961 r = adev->ip_blocks[i].version->funcs->suspend(adev);
1962 /* XXX handle errors */
1964 DRM_ERROR("suspend of IP block <%s> failed %d\n",
1965 adev->ip_blocks[i].version->funcs->name, r);
1970 if (amdgpu_sriov_vf(adev))
1971 amdgpu_virt_release_full_gpu(adev, false);
1977 * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2)
1979 * @adev: amdgpu_device pointer
1981 * Main suspend function for hardware IPs. The list of all the hardware
1982 * IPs that make up the asic is walked, clockgating is disabled and the
1983 * suspend callbacks are run. suspend puts the hardware and software state
1984 * in each IP into a state suitable for suspend.
1985 * Returns 0 on success, negative error code on failure.
1987 static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev)
1991 if (amdgpu_sriov_vf(adev))
1992 amdgpu_virt_request_full_gpu(adev, false);
1994 /* ungate SMC block first */
1995 r = amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_SMC,
1996 AMD_CG_STATE_UNGATE);
1998 DRM_ERROR("set_clockgating_state(ungate) SMC failed %d\n", r);
2001 /* call smu to disable gfx off feature first when suspend */
2002 if (adev->powerplay.pp_funcs->set_powergating_by_smu)
2003 amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false);
2005 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2006 if (!adev->ip_blocks[i].status.valid)
2008 /* displays are handled in phase1 */
2009 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE)
2011 /* ungate blocks so that suspend can properly shut them down */
2012 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_SMC &&
2013 adev->ip_blocks[i].version->funcs->set_clockgating_state) {
2014 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
2015 AMD_CG_STATE_UNGATE);
2017 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
2018 adev->ip_blocks[i].version->funcs->name, r);
2021 /* XXX handle errors */
2022 r = adev->ip_blocks[i].version->funcs->suspend(adev);
2023 /* XXX handle errors */
2025 DRM_ERROR("suspend of IP block <%s> failed %d\n",
2026 adev->ip_blocks[i].version->funcs->name, r);
2030 if (amdgpu_sriov_vf(adev))
2031 amdgpu_virt_release_full_gpu(adev, false);
2037 * amdgpu_device_ip_suspend - run suspend for hardware IPs
2039 * @adev: amdgpu_device pointer
2041 * Main suspend function for hardware IPs. The list of all the hardware
2042 * IPs that make up the asic is walked, clockgating is disabled and the
2043 * suspend callbacks are run. suspend puts the hardware and software state
2044 * in each IP into a state suitable for suspend.
2045 * Returns 0 on success, negative error code on failure.
2047 int amdgpu_device_ip_suspend(struct amdgpu_device *adev)
2051 r = amdgpu_device_ip_suspend_phase1(adev);
2054 r = amdgpu_device_ip_suspend_phase2(adev);
2059 static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
2063 static enum amd_ip_block_type ip_order[] = {
2064 AMD_IP_BLOCK_TYPE_GMC,
2065 AMD_IP_BLOCK_TYPE_COMMON,
2066 AMD_IP_BLOCK_TYPE_PSP,
2067 AMD_IP_BLOCK_TYPE_IH,
2070 for (i = 0; i < adev->num_ip_blocks; i++) {
2072 struct amdgpu_ip_block *block;
2074 for (j = 0; j < adev->num_ip_blocks; j++) {
2075 block = &adev->ip_blocks[j];
2077 if (block->version->type != ip_order[i] ||
2078 !block->status.valid)
2081 r = block->version->funcs->hw_init(adev);
2082 DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
2091 static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
2095 static enum amd_ip_block_type ip_order[] = {
2096 AMD_IP_BLOCK_TYPE_SMC,
2097 AMD_IP_BLOCK_TYPE_DCE,
2098 AMD_IP_BLOCK_TYPE_GFX,
2099 AMD_IP_BLOCK_TYPE_SDMA,
2100 AMD_IP_BLOCK_TYPE_UVD,
2101 AMD_IP_BLOCK_TYPE_VCE
2104 for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
2106 struct amdgpu_ip_block *block;
2108 for (j = 0; j < adev->num_ip_blocks; j++) {
2109 block = &adev->ip_blocks[j];
2111 if (block->version->type != ip_order[i] ||
2112 !block->status.valid)
2115 r = block->version->funcs->hw_init(adev);
2116 DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
2126 * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs
2128 * @adev: amdgpu_device pointer
2130 * First resume function for hardware IPs. The list of all the hardware
2131 * IPs that make up the asic is walked and the resume callbacks are run for
2132 * COMMON, GMC, and IH. resume puts the hardware into a functional state
2133 * after a suspend and updates the software state as necessary. This
2134 * function is also used for restoring the GPU after a GPU reset.
2135 * Returns 0 on success, negative error code on failure.
2137 static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)
2141 for (i = 0; i < adev->num_ip_blocks; i++) {
2142 if (!adev->ip_blocks[i].status.valid)
2144 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2145 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2146 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
2147 r = adev->ip_blocks[i].version->funcs->resume(adev);
2149 DRM_ERROR("resume of IP block <%s> failed %d\n",
2150 adev->ip_blocks[i].version->funcs->name, r);
2160 * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs
2162 * @adev: amdgpu_device pointer
2164 * First resume function for hardware IPs. The list of all the hardware
2165 * IPs that make up the asic is walked and the resume callbacks are run for
2166 * all blocks except COMMON, GMC, and IH. resume puts the hardware into a
2167 * functional state after a suspend and updates the software state as
2168 * necessary. This function is also used for restoring the GPU after a GPU
2170 * Returns 0 on success, negative error code on failure.
2172 static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)
2176 for (i = 0; i < adev->num_ip_blocks; i++) {
2177 if (!adev->ip_blocks[i].status.valid)
2179 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2180 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2181 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH)
2183 r = adev->ip_blocks[i].version->funcs->resume(adev);
2185 DRM_ERROR("resume of IP block <%s> failed %d\n",
2186 adev->ip_blocks[i].version->funcs->name, r);
2195 * amdgpu_device_ip_resume - run resume for hardware IPs
2197 * @adev: amdgpu_device pointer
2199 * Main resume function for hardware IPs. The hardware IPs
2200 * are split into two resume functions because they are
2201 * are also used in in recovering from a GPU reset and some additional
2202 * steps need to be take between them. In this case (S3/S4) they are
2204 * Returns 0 on success, negative error code on failure.
2206 static int amdgpu_device_ip_resume(struct amdgpu_device *adev)
2210 r = amdgpu_device_ip_resume_phase1(adev);
2213 r = amdgpu_device_ip_resume_phase2(adev);
2219 * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV
2221 * @adev: amdgpu_device pointer
2223 * Query the VBIOS data tables to determine if the board supports SR-IOV.
2225 static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
2227 if (amdgpu_sriov_vf(adev)) {
2228 if (adev->is_atom_fw) {
2229 if (amdgpu_atomfirmware_gpu_supports_virtualization(adev))
2230 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2232 if (amdgpu_atombios_has_gpu_virtualization_table(adev))
2233 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2236 if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
2237 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
2242 * amdgpu_device_asic_has_dc_support - determine if DC supports the asic
2244 * @asic_type: AMD asic type
2246 * Check if there is DC (new modesetting infrastructre) support for an asic.
2247 * returns true if DC has support, false if not.
2249 bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
2251 switch (asic_type) {
2252 #if defined(CONFIG_DRM_AMD_DC)
2258 * We have systems in the wild with these ASICs that require
2259 * LVDS and VGA support which is not supported with DC.
2261 * Fallback to the non-DC driver here by default so as not to
2262 * cause regressions.
2264 return amdgpu_dc > 0;
2268 case CHIP_POLARIS10:
2269 case CHIP_POLARIS11:
2270 case CHIP_POLARIS12:
2277 #if defined(CONFIG_DRM_AMD_DC_DCN1_0)
2280 return amdgpu_dc != 0;
2288 * amdgpu_device_has_dc_support - check if dc is supported
2290 * @adev: amdgpu_device_pointer
2292 * Returns true for supported, false for not supported
2294 bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
2296 if (amdgpu_sriov_vf(adev))
2299 return amdgpu_device_asic_has_dc_support(adev->asic_type);
2303 * amdgpu_device_init - initialize the driver
2305 * @adev: amdgpu_device pointer
2306 * @ddev: drm dev pointer
2307 * @pdev: pci dev pointer
2308 * @flags: driver flags
2310 * Initializes the driver info and hw (all asics).
2311 * Returns 0 for success or an error on failure.
2312 * Called at driver startup.
2314 int amdgpu_device_init(struct amdgpu_device *adev,
2315 struct drm_device *ddev,
2316 struct pci_dev *pdev,
2320 bool runtime = false;
2323 adev->shutdown = false;
2324 adev->dev = &pdev->dev;
2327 adev->flags = flags;
2328 adev->asic_type = flags & AMD_ASIC_MASK;
2329 adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
2330 if (amdgpu_emu_mode == 1)
2331 adev->usec_timeout *= 2;
2332 adev->gmc.gart_size = 512 * 1024 * 1024;
2333 adev->accel_working = false;
2334 adev->num_rings = 0;
2335 adev->mman.buffer_funcs = NULL;
2336 adev->mman.buffer_funcs_ring = NULL;
2337 adev->vm_manager.vm_pte_funcs = NULL;
2338 adev->vm_manager.vm_pte_num_rings = 0;
2339 adev->gmc.gmc_funcs = NULL;
2340 adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
2341 bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
2343 adev->smc_rreg = &amdgpu_invalid_rreg;
2344 adev->smc_wreg = &amdgpu_invalid_wreg;
2345 adev->pcie_rreg = &amdgpu_invalid_rreg;
2346 adev->pcie_wreg = &amdgpu_invalid_wreg;
2347 adev->pciep_rreg = &amdgpu_invalid_rreg;
2348 adev->pciep_wreg = &amdgpu_invalid_wreg;
2349 adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
2350 adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
2351 adev->didt_rreg = &amdgpu_invalid_rreg;
2352 adev->didt_wreg = &amdgpu_invalid_wreg;
2353 adev->gc_cac_rreg = &amdgpu_invalid_rreg;
2354 adev->gc_cac_wreg = &amdgpu_invalid_wreg;
2355 adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
2356 adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
2358 DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
2359 amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
2360 pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
2362 /* mutex initialization are all done here so we
2363 * can recall function without having locking issues */
2364 atomic_set(&adev->irq.ih.lock, 0);
2365 mutex_init(&adev->firmware.mutex);
2366 mutex_init(&adev->pm.mutex);
2367 mutex_init(&adev->gfx.gpu_clock_mutex);
2368 mutex_init(&adev->srbm_mutex);
2369 mutex_init(&adev->gfx.pipe_reserve_mutex);
2370 mutex_init(&adev->grbm_idx_mutex);
2371 mutex_init(&adev->mn_lock);
2372 mutex_init(&adev->virt.vf_errors.lock);
2373 hash_init(adev->mn_hash);
2374 mutex_init(&adev->lock_reset);
2376 amdgpu_device_check_arguments(adev);
2378 spin_lock_init(&adev->mmio_idx_lock);
2379 spin_lock_init(&adev->smc_idx_lock);
2380 spin_lock_init(&adev->pcie_idx_lock);
2381 spin_lock_init(&adev->uvd_ctx_idx_lock);
2382 spin_lock_init(&adev->didt_idx_lock);
2383 spin_lock_init(&adev->gc_cac_idx_lock);
2384 spin_lock_init(&adev->se_cac_idx_lock);
2385 spin_lock_init(&adev->audio_endpt_idx_lock);
2386 spin_lock_init(&adev->mm_stats.lock);
2388 INIT_LIST_HEAD(&adev->shadow_list);
2389 mutex_init(&adev->shadow_list_lock);
2391 INIT_LIST_HEAD(&adev->ring_lru_list);
2392 spin_lock_init(&adev->ring_lru_list_lock);
2394 INIT_DELAYED_WORK(&adev->late_init_work,
2395 amdgpu_device_ip_late_init_func_handler);
2397 adev->pm.ac_power = power_supply_is_system_supplied() > 0 ? true : false;
2399 /* Registers mapping */
2400 /* TODO: block userspace mapping of io register */
2401 if (adev->asic_type >= CHIP_BONAIRE) {
2402 adev->rmmio_base = pci_resource_start(adev->pdev, 5);
2403 adev->rmmio_size = pci_resource_len(adev->pdev, 5);
2405 adev->rmmio_base = pci_resource_start(adev->pdev, 2);
2406 adev->rmmio_size = pci_resource_len(adev->pdev, 2);
2409 adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
2410 if (adev->rmmio == NULL) {
2413 DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
2414 DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
2416 /* doorbell bar mapping */
2417 amdgpu_device_doorbell_init(adev);
2419 /* io port mapping */
2420 for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
2421 if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) {
2422 adev->rio_mem_size = pci_resource_len(adev->pdev, i);
2423 adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size);
2427 if (adev->rio_mem == NULL)
2428 DRM_INFO("PCI I/O BAR is not found.\n");
2430 amdgpu_device_get_pcie_info(adev);
2432 /* early init functions */
2433 r = amdgpu_device_ip_early_init(adev);
2437 /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
2438 /* this will fail for cards that aren't VGA class devices, just
2440 vga_client_register(adev->pdev, adev, NULL, amdgpu_device_vga_set_decode);
2442 if (amdgpu_device_is_px(ddev))
2444 if (!pci_is_thunderbolt_attached(adev->pdev))
2445 vga_switcheroo_register_client(adev->pdev,
2446 &amdgpu_switcheroo_ops, runtime);
2448 vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
2450 if (amdgpu_emu_mode == 1) {
2451 /* post the asic on emulation mode */
2452 emu_soc_asic_init(adev);
2453 goto fence_driver_init;
2457 if (!amdgpu_get_bios(adev)) {
2462 r = amdgpu_atombios_init(adev);
2464 dev_err(adev->dev, "amdgpu_atombios_init failed\n");
2465 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0);
2469 /* detect if we are with an SRIOV vbios */
2470 amdgpu_device_detect_sriov_bios(adev);
2472 /* Post card if necessary */
2473 if (amdgpu_device_need_post(adev)) {
2475 dev_err(adev->dev, "no vBIOS found\n");
2479 DRM_INFO("GPU posting now...\n");
2480 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2482 dev_err(adev->dev, "gpu post error!\n");
2487 if (adev->is_atom_fw) {
2488 /* Initialize clocks */
2489 r = amdgpu_atomfirmware_get_clock_info(adev);
2491 dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
2492 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
2496 /* Initialize clocks */
2497 r = amdgpu_atombios_get_clock_info(adev);
2499 dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
2500 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
2503 /* init i2c buses */
2504 if (!amdgpu_device_has_dc_support(adev))
2505 amdgpu_atombios_i2c_init(adev);
2510 r = amdgpu_fence_driver_init(adev);
2512 dev_err(adev->dev, "amdgpu_fence_driver_init failed\n");
2513 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
2517 /* init the mode config */
2518 drm_mode_config_init(adev->ddev);
2520 r = amdgpu_device_ip_init(adev);
2522 /* failed in exclusive mode due to timeout */
2523 if (amdgpu_sriov_vf(adev) &&
2524 !amdgpu_sriov_runtime(adev) &&
2525 amdgpu_virt_mmio_blocked(adev) &&
2526 !amdgpu_virt_wait_reset(adev)) {
2527 dev_err(adev->dev, "VF exclusive mode timeout\n");
2528 /* Don't send request since VF is inactive. */
2529 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
2530 adev->virt.ops = NULL;
2534 dev_err(adev->dev, "amdgpu_device_ip_init failed\n");
2535 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
2539 adev->accel_working = true;
2541 amdgpu_vm_check_compute_bug(adev);
2543 /* Initialize the buffer migration limit. */
2544 if (amdgpu_moverate >= 0)
2545 max_MBps = amdgpu_moverate;
2547 max_MBps = 8; /* Allow 8 MB/s. */
2548 /* Get a log2 for easy divisions. */
2549 adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
2551 r = amdgpu_ib_pool_init(adev);
2553 dev_err(adev->dev, "IB initialization failed (%d).\n", r);
2554 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r);
2558 amdgpu_fbdev_init(adev);
2560 r = amdgpu_pm_sysfs_init(adev);
2562 DRM_ERROR("registering pm debugfs failed (%d).\n", r);
2564 r = amdgpu_debugfs_gem_init(adev);
2566 DRM_ERROR("registering gem debugfs failed (%d).\n", r);
2568 r = amdgpu_debugfs_regs_init(adev);
2570 DRM_ERROR("registering register debugfs failed (%d).\n", r);
2572 r = amdgpu_debugfs_firmware_init(adev);
2574 DRM_ERROR("registering firmware debugfs failed (%d).\n", r);
2576 r = amdgpu_debugfs_init(adev);
2578 DRM_ERROR("Creating debugfs files failed (%d).\n", r);
2580 if ((amdgpu_testing & 1)) {
2581 if (adev->accel_working)
2582 amdgpu_test_moves(adev);
2584 DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n");
2586 if (amdgpu_benchmarking) {
2587 if (adev->accel_working)
2588 amdgpu_benchmark(adev, amdgpu_benchmarking);
2590 DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n");
2593 /* enable clockgating, etc. after ib tests, etc. since some blocks require
2594 * explicit gating rather than handling it automatically.
2596 r = amdgpu_device_ip_late_init(adev);
2598 dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n");
2599 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
2606 amdgpu_vf_error_trans_all(adev);
2608 vga_switcheroo_fini_domain_pm_ops(adev->dev);
2614 * amdgpu_device_fini - tear down the driver
2616 * @adev: amdgpu_device pointer
2618 * Tear down the driver info (all asics).
2619 * Called at driver shutdown.
2621 void amdgpu_device_fini(struct amdgpu_device *adev)
2625 DRM_INFO("amdgpu: finishing device.\n");
2626 adev->shutdown = true;
2627 /* disable all interrupts */
2628 amdgpu_irq_disable_all(adev);
2629 if (adev->mode_info.mode_config_initialized){
2630 if (!amdgpu_device_has_dc_support(adev))
2631 drm_crtc_force_disable_all(adev->ddev);
2633 drm_atomic_helper_shutdown(adev->ddev);
2635 amdgpu_ib_pool_fini(adev);
2636 amdgpu_fence_driver_fini(adev);
2637 amdgpu_pm_sysfs_fini(adev);
2638 amdgpu_fbdev_fini(adev);
2639 r = amdgpu_device_ip_fini(adev);
2640 if (adev->firmware.gpu_info_fw) {
2641 release_firmware(adev->firmware.gpu_info_fw);
2642 adev->firmware.gpu_info_fw = NULL;
2644 adev->accel_working = false;
2645 cancel_delayed_work_sync(&adev->late_init_work);
2646 /* free i2c buses */
2647 if (!amdgpu_device_has_dc_support(adev))
2648 amdgpu_i2c_fini(adev);
2650 if (amdgpu_emu_mode != 1)
2651 amdgpu_atombios_fini(adev);
2655 if (!pci_is_thunderbolt_attached(adev->pdev))
2656 vga_switcheroo_unregister_client(adev->pdev);
2657 if (adev->flags & AMD_IS_PX)
2658 vga_switcheroo_fini_domain_pm_ops(adev->dev);
2659 vga_client_register(adev->pdev, NULL, NULL, NULL);
2661 pci_iounmap(adev->pdev, adev->rio_mem);
2662 adev->rio_mem = NULL;
2663 iounmap(adev->rmmio);
2665 amdgpu_device_doorbell_fini(adev);
2666 amdgpu_debugfs_regs_cleanup(adev);
2674 * amdgpu_device_suspend - initiate device suspend
2676 * @dev: drm dev pointer
2677 * @suspend: suspend state
2678 * @fbcon : notify the fbdev of suspend
2680 * Puts the hw in the suspend state (all asics).
2681 * Returns 0 for success or an error on failure.
2682 * Called at driver suspend.
2684 int amdgpu_device_suspend(struct drm_device *dev, bool suspend, bool fbcon)
2686 struct amdgpu_device *adev;
2687 struct drm_crtc *crtc;
2688 struct drm_connector *connector;
2691 if (dev == NULL || dev->dev_private == NULL) {
2695 adev = dev->dev_private;
2697 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
2700 drm_kms_helper_poll_disable(dev);
2703 amdgpu_fbdev_set_suspend(adev, 1);
2705 if (!amdgpu_device_has_dc_support(adev)) {
2706 /* turn off display hw */
2707 drm_modeset_lock_all(dev);
2708 list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2709 drm_helper_connector_dpms(connector, DRM_MODE_DPMS_OFF);
2711 drm_modeset_unlock_all(dev);
2712 /* unpin the front buffers and cursors */
2713 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2714 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2715 struct drm_framebuffer *fb = crtc->primary->fb;
2716 struct amdgpu_bo *robj;
2718 if (amdgpu_crtc->cursor_bo) {
2719 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2720 r = amdgpu_bo_reserve(aobj, true);
2722 amdgpu_bo_unpin(aobj);
2723 amdgpu_bo_unreserve(aobj);
2727 if (fb == NULL || fb->obj[0] == NULL) {
2730 robj = gem_to_amdgpu_bo(fb->obj[0]);
2731 /* don't unpin kernel fb objects */
2732 if (!amdgpu_fbdev_robj_is_fb(adev, robj)) {
2733 r = amdgpu_bo_reserve(robj, true);
2735 amdgpu_bo_unpin(robj);
2736 amdgpu_bo_unreserve(robj);
2742 amdgpu_amdkfd_suspend(adev);
2744 r = amdgpu_device_ip_suspend_phase1(adev);
2746 /* evict vram memory */
2747 amdgpu_bo_evict_vram(adev);
2749 amdgpu_fence_driver_suspend(adev);
2751 r = amdgpu_device_ip_suspend_phase2(adev);
2753 /* evict remaining vram memory
2754 * This second call to evict vram is to evict the gart page table
2757 amdgpu_bo_evict_vram(adev);
2759 pci_save_state(dev->pdev);
2761 /* Shut down the device */
2762 pci_disable_device(dev->pdev);
2763 pci_set_power_state(dev->pdev, PCI_D3hot);
2765 r = amdgpu_asic_reset(adev);
2767 DRM_ERROR("amdgpu asic reset failed\n");
2774 * amdgpu_device_resume - initiate device resume
2776 * @dev: drm dev pointer
2777 * @resume: resume state
2778 * @fbcon : notify the fbdev of resume
2780 * Bring the hw back to operating state (all asics).
2781 * Returns 0 for success or an error on failure.
2782 * Called at driver resume.
2784 int amdgpu_device_resume(struct drm_device *dev, bool resume, bool fbcon)
2786 struct drm_connector *connector;
2787 struct amdgpu_device *adev = dev->dev_private;
2788 struct drm_crtc *crtc;
2791 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
2795 pci_set_power_state(dev->pdev, PCI_D0);
2796 pci_restore_state(dev->pdev);
2797 r = pci_enable_device(dev->pdev);
2803 if (amdgpu_device_need_post(adev)) {
2804 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2806 DRM_ERROR("amdgpu asic init failed\n");
2809 r = amdgpu_device_ip_resume(adev);
2811 DRM_ERROR("amdgpu_device_ip_resume failed (%d).\n", r);
2814 amdgpu_fence_driver_resume(adev);
2817 r = amdgpu_device_ip_late_init(adev);
2821 if (!amdgpu_device_has_dc_support(adev)) {
2823 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2824 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2826 if (amdgpu_crtc->cursor_bo) {
2827 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2828 r = amdgpu_bo_reserve(aobj, true);
2830 r = amdgpu_bo_pin(aobj, AMDGPU_GEM_DOMAIN_VRAM);
2832 DRM_ERROR("Failed to pin cursor BO (%d)\n", r);
2833 amdgpu_crtc->cursor_addr = amdgpu_bo_gpu_offset(aobj);
2834 amdgpu_bo_unreserve(aobj);
2839 r = amdgpu_amdkfd_resume(adev);
2843 /* Make sure IB tests flushed */
2844 flush_delayed_work(&adev->late_init_work);
2846 /* blat the mode back in */
2848 if (!amdgpu_device_has_dc_support(adev)) {
2850 drm_helper_resume_force_mode(dev);
2852 /* turn on display hw */
2853 drm_modeset_lock_all(dev);
2854 list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2855 drm_helper_connector_dpms(connector, DRM_MODE_DPMS_ON);
2857 drm_modeset_unlock_all(dev);
2859 amdgpu_fbdev_set_suspend(adev, 0);
2862 drm_kms_helper_poll_enable(dev);
2865 * Most of the connector probing functions try to acquire runtime pm
2866 * refs to ensure that the GPU is powered on when connector polling is
2867 * performed. Since we're calling this from a runtime PM callback,
2868 * trying to acquire rpm refs will cause us to deadlock.
2870 * Since we're guaranteed to be holding the rpm lock, it's safe to
2871 * temporarily disable the rpm helpers so this doesn't deadlock us.
2874 dev->dev->power.disable_depth++;
2876 if (!amdgpu_device_has_dc_support(adev))
2877 drm_helper_hpd_irq_event(dev);
2879 drm_kms_helper_hotplug_event(dev);
2881 dev->dev->power.disable_depth--;
2887 * amdgpu_device_ip_check_soft_reset - did soft reset succeed
2889 * @adev: amdgpu_device pointer
2891 * The list of all the hardware IPs that make up the asic is walked and
2892 * the check_soft_reset callbacks are run. check_soft_reset determines
2893 * if the asic is still hung or not.
2894 * Returns true if any of the IPs are still in a hung state, false if not.
2896 static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)
2899 bool asic_hang = false;
2901 if (amdgpu_sriov_vf(adev))
2904 if (amdgpu_asic_need_full_reset(adev))
2907 for (i = 0; i < adev->num_ip_blocks; i++) {
2908 if (!adev->ip_blocks[i].status.valid)
2910 if (adev->ip_blocks[i].version->funcs->check_soft_reset)
2911 adev->ip_blocks[i].status.hang =
2912 adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
2913 if (adev->ip_blocks[i].status.hang) {
2914 DRM_INFO("IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
2922 * amdgpu_device_ip_pre_soft_reset - prepare for soft reset
2924 * @adev: amdgpu_device pointer
2926 * The list of all the hardware IPs that make up the asic is walked and the
2927 * pre_soft_reset callbacks are run if the block is hung. pre_soft_reset
2928 * handles any IP specific hardware or software state changes that are
2929 * necessary for a soft reset to succeed.
2930 * Returns 0 on success, negative error code on failure.
2932 static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)
2936 for (i = 0; i < adev->num_ip_blocks; i++) {
2937 if (!adev->ip_blocks[i].status.valid)
2939 if (adev->ip_blocks[i].status.hang &&
2940 adev->ip_blocks[i].version->funcs->pre_soft_reset) {
2941 r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
2951 * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed
2953 * @adev: amdgpu_device pointer
2955 * Some hardware IPs cannot be soft reset. If they are hung, a full gpu
2956 * reset is necessary to recover.
2957 * Returns true if a full asic reset is required, false if not.
2959 static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)
2963 if (amdgpu_asic_need_full_reset(adev))
2966 for (i = 0; i < adev->num_ip_blocks; i++) {
2967 if (!adev->ip_blocks[i].status.valid)
2969 if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
2970 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
2971 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
2972 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
2973 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
2974 if (adev->ip_blocks[i].status.hang) {
2975 DRM_INFO("Some block need full reset!\n");
2984 * amdgpu_device_ip_soft_reset - do a soft reset
2986 * @adev: amdgpu_device pointer
2988 * The list of all the hardware IPs that make up the asic is walked and the
2989 * soft_reset callbacks are run if the block is hung. soft_reset handles any
2990 * IP specific hardware or software state changes that are necessary to soft
2992 * Returns 0 on success, negative error code on failure.
2994 static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)
2998 for (i = 0; i < adev->num_ip_blocks; i++) {
2999 if (!adev->ip_blocks[i].status.valid)
3001 if (adev->ip_blocks[i].status.hang &&
3002 adev->ip_blocks[i].version->funcs->soft_reset) {
3003 r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
3013 * amdgpu_device_ip_post_soft_reset - clean up from soft reset
3015 * @adev: amdgpu_device pointer
3017 * The list of all the hardware IPs that make up the asic is walked and the
3018 * post_soft_reset callbacks are run if the asic was hung. post_soft_reset
3019 * handles any IP specific hardware or software state changes that are
3020 * necessary after the IP has been soft reset.
3021 * Returns 0 on success, negative error code on failure.
3023 static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)
3027 for (i = 0; i < adev->num_ip_blocks; i++) {
3028 if (!adev->ip_blocks[i].status.valid)
3030 if (adev->ip_blocks[i].status.hang &&
3031 adev->ip_blocks[i].version->funcs->post_soft_reset)
3032 r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
3041 * amdgpu_device_recover_vram_from_shadow - restore shadowed VRAM buffers
3043 * @adev: amdgpu_device pointer
3044 * @ring: amdgpu_ring for the engine handling the buffer operations
3045 * @bo: amdgpu_bo buffer whose shadow is being restored
3046 * @fence: dma_fence associated with the operation
3048 * Restores the VRAM buffer contents from the shadow in GTT. Used to
3049 * restore things like GPUVM page tables after a GPU reset where
3050 * the contents of VRAM might be lost.
3051 * Returns 0 on success, negative error code on failure.
3053 static int amdgpu_device_recover_vram_from_shadow(struct amdgpu_device *adev,
3054 struct amdgpu_ring *ring,
3055 struct amdgpu_bo *bo,
3056 struct dma_fence **fence)
3064 r = amdgpu_bo_reserve(bo, true);
3067 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
3068 /* if bo has been evicted, then no need to recover */
3069 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
3070 r = amdgpu_bo_validate(bo->shadow);
3072 DRM_ERROR("bo validate failed!\n");
3076 r = amdgpu_bo_restore_from_shadow(adev, ring, bo,
3079 DRM_ERROR("recover page table failed!\n");
3084 amdgpu_bo_unreserve(bo);
3089 * amdgpu_device_handle_vram_lost - Handle the loss of VRAM contents
3091 * @adev: amdgpu_device pointer
3093 * Restores the contents of VRAM buffers from the shadows in GTT. Used to
3094 * restore things like GPUVM page tables after a GPU reset where
3095 * the contents of VRAM might be lost.
3096 * Returns 0 on success, 1 on failure.
3098 static int amdgpu_device_handle_vram_lost(struct amdgpu_device *adev)
3100 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
3101 struct amdgpu_bo *bo, *tmp;
3102 struct dma_fence *fence = NULL, *next = NULL;
3107 if (amdgpu_sriov_runtime(adev))
3108 tmo = msecs_to_jiffies(8000);
3110 tmo = msecs_to_jiffies(100);
3112 DRM_INFO("recover vram bo from shadow start\n");
3113 mutex_lock(&adev->shadow_list_lock);
3114 list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) {
3116 amdgpu_device_recover_vram_from_shadow(adev, ring, bo, &next);
3118 r = dma_fence_wait_timeout(fence, false, tmo);
3120 pr_err("wait fence %p[%d] timeout\n", fence, i);
3122 pr_err("wait fence %p[%d] interrupted\n", fence, i);
3124 dma_fence_put(fence);
3131 dma_fence_put(fence);
3134 mutex_unlock(&adev->shadow_list_lock);
3137 r = dma_fence_wait_timeout(fence, false, tmo);
3139 pr_err("wait fence %p[%d] timeout\n", fence, i);
3141 pr_err("wait fence %p[%d] interrupted\n", fence, i);
3144 dma_fence_put(fence);
3147 DRM_INFO("recover vram bo from shadow done\n");
3149 DRM_ERROR("recover vram bo from shadow failed\n");
3151 return (r > 0) ? 0 : 1;
3155 * amdgpu_device_reset - reset ASIC/GPU for bare-metal or passthrough
3157 * @adev: amdgpu device pointer
3159 * attempt to do soft-reset or full-reset and reinitialize Asic
3160 * return 0 means succeeded otherwise failed
3162 static int amdgpu_device_reset(struct amdgpu_device *adev)
3164 bool need_full_reset, vram_lost = 0;
3167 need_full_reset = amdgpu_device_ip_need_full_reset(adev);
3169 if (!need_full_reset) {
3170 amdgpu_device_ip_pre_soft_reset(adev);
3171 r = amdgpu_device_ip_soft_reset(adev);
3172 amdgpu_device_ip_post_soft_reset(adev);
3173 if (r || amdgpu_device_ip_check_soft_reset(adev)) {
3174 DRM_INFO("soft reset failed, will fallback to full reset!\n");
3175 need_full_reset = true;
3179 if (need_full_reset) {
3180 r = amdgpu_device_ip_suspend(adev);
3183 r = amdgpu_asic_reset(adev);
3185 amdgpu_atom_asic_init(adev->mode_info.atom_context);
3188 dev_info(adev->dev, "GPU reset succeeded, trying to resume\n");
3189 r = amdgpu_device_ip_resume_phase1(adev);
3193 vram_lost = amdgpu_device_check_vram_lost(adev);
3195 DRM_ERROR("VRAM is lost!\n");
3196 atomic_inc(&adev->vram_lost_counter);
3199 r = amdgpu_gtt_mgr_recover(
3200 &adev->mman.bdev.man[TTM_PL_TT]);
3204 r = amdgpu_device_ip_resume_phase2(adev);
3209 amdgpu_device_fill_reset_magic(adev);
3215 amdgpu_irq_gpu_reset_resume_helper(adev);
3216 r = amdgpu_ib_ring_tests(adev);
3218 dev_err(adev->dev, "ib ring test failed (%d).\n", r);
3219 r = amdgpu_device_ip_suspend(adev);
3220 need_full_reset = true;
3225 if (!r && ((need_full_reset && !(adev->flags & AMD_IS_APU)) || vram_lost))
3226 r = amdgpu_device_handle_vram_lost(adev);
3232 * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
3234 * @adev: amdgpu device pointer
3235 * @from_hypervisor: request from hypervisor
3237 * do VF FLR and reinitialize Asic
3238 * return 0 means succeeded otherwise failed
3240 static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
3241 bool from_hypervisor)
3245 if (from_hypervisor)
3246 r = amdgpu_virt_request_full_gpu(adev, true);
3248 r = amdgpu_virt_reset_gpu(adev);
3252 /* Resume IP prior to SMC */
3253 r = amdgpu_device_ip_reinit_early_sriov(adev);
3257 /* we need recover gart prior to run SMC/CP/SDMA resume */
3258 amdgpu_gtt_mgr_recover(&adev->mman.bdev.man[TTM_PL_TT]);
3260 /* now we are okay to resume SMC/CP/SDMA */
3261 r = amdgpu_device_ip_reinit_late_sriov(adev);
3265 amdgpu_irq_gpu_reset_resume_helper(adev);
3266 r = amdgpu_ib_ring_tests(adev);
3269 amdgpu_virt_init_data_exchange(adev);
3270 amdgpu_virt_release_full_gpu(adev, true);
3271 if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
3272 atomic_inc(&adev->vram_lost_counter);
3273 r = amdgpu_device_handle_vram_lost(adev);
3280 * amdgpu_device_gpu_recover - reset the asic and recover scheduler
3282 * @adev: amdgpu device pointer
3283 * @job: which job trigger hang
3284 * @force: forces reset regardless of amdgpu_gpu_recovery
3286 * Attempt to reset the GPU if it has hung (all asics).
3287 * Returns 0 for success or an error on failure.
3289 int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
3290 struct amdgpu_job *job, bool force)
3294 if (!force && !amdgpu_device_ip_check_soft_reset(adev)) {
3295 DRM_INFO("No hardware hang detected. Did some blocks stall?\n");
3299 if (!force && (amdgpu_gpu_recovery == 0 ||
3300 (amdgpu_gpu_recovery == -1 && !amdgpu_sriov_vf(adev)))) {
3301 DRM_INFO("GPU recovery disabled.\n");
3305 dev_info(adev->dev, "GPU reset begin!\n");
3307 mutex_lock(&adev->lock_reset);
3308 atomic_inc(&adev->gpu_reset_counter);
3309 adev->in_gpu_reset = 1;
3312 amdgpu_amdkfd_pre_reset(adev);
3315 resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
3317 /* block all schedulers and reset given job's ring */
3318 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3319 struct amdgpu_ring *ring = adev->rings[i];
3321 if (!ring || !ring->sched.thread)
3324 kthread_park(ring->sched.thread);
3326 if (job && job->base.sched == &ring->sched)
3329 drm_sched_hw_job_reset(&ring->sched, job ? &job->base : NULL);
3331 /* after all hw jobs are reset, hw fence is meaningless, so force_completion */
3332 amdgpu_fence_driver_force_completion(ring);
3335 if (amdgpu_sriov_vf(adev))
3336 r = amdgpu_device_reset_sriov(adev, job ? false : true);
3338 r = amdgpu_device_reset(adev);
3340 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3341 struct amdgpu_ring *ring = adev->rings[i];
3343 if (!ring || !ring->sched.thread)
3346 /* only need recovery sched of the given job's ring
3347 * or all rings (in the case @job is NULL)
3348 * after above amdgpu_reset accomplished
3350 if ((!job || job->base.sched == &ring->sched) && !r)
3351 drm_sched_job_recovery(&ring->sched);
3353 kthread_unpark(ring->sched.thread);
3356 if (!amdgpu_device_has_dc_support(adev)) {
3357 drm_helper_resume_force_mode(adev->ddev);
3360 ttm_bo_unlock_delayed_workqueue(&adev->mman.bdev, resched);
3363 /* bad news, how to tell it to userspace ? */
3364 dev_info(adev->dev, "GPU reset(%d) failed\n", atomic_read(&adev->gpu_reset_counter));
3365 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r);
3367 dev_info(adev->dev, "GPU reset(%d) succeeded!\n",atomic_read(&adev->gpu_reset_counter));
3371 amdgpu_amdkfd_post_reset(adev);
3372 amdgpu_vf_error_trans_all(adev);
3373 adev->in_gpu_reset = 0;
3374 mutex_unlock(&adev->lock_reset);
3379 * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot
3381 * @adev: amdgpu_device pointer
3383 * Fetchs and stores in the driver the PCIE capabilities (gen speed
3384 * and lanes) of the slot the device is in. Handles APUs and
3385 * virtualized environments where PCIE config space may not be available.
3387 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev)
3389 struct pci_dev *pdev;
3390 enum pci_bus_speed speed_cap;
3391 enum pcie_link_width link_width;
3393 if (amdgpu_pcie_gen_cap)
3394 adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
3396 if (amdgpu_pcie_lane_cap)
3397 adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
3399 /* covers APUs as well */
3400 if (pci_is_root_bus(adev->pdev->bus)) {
3401 if (adev->pm.pcie_gen_mask == 0)
3402 adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
3403 if (adev->pm.pcie_mlw_mask == 0)
3404 adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
3408 if (adev->pm.pcie_gen_mask == 0) {
3411 speed_cap = pcie_get_speed_cap(pdev);
3412 if (speed_cap == PCI_SPEED_UNKNOWN) {
3413 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3414 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3415 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
3417 if (speed_cap == PCIE_SPEED_16_0GT)
3418 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3419 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3420 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
3421 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4);
3422 else if (speed_cap == PCIE_SPEED_8_0GT)
3423 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3424 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3425 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
3426 else if (speed_cap == PCIE_SPEED_5_0GT)
3427 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3428 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2);
3430 adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1;
3433 pdev = adev->ddev->pdev->bus->self;
3434 speed_cap = pcie_get_speed_cap(pdev);
3435 if (speed_cap == PCI_SPEED_UNKNOWN) {
3436 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3437 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
3439 if (speed_cap == PCIE_SPEED_16_0GT)
3440 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3441 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3442 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
3443 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4);
3444 else if (speed_cap == PCIE_SPEED_8_0GT)
3445 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3446 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3447 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3);
3448 else if (speed_cap == PCIE_SPEED_5_0GT)
3449 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3450 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
3452 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
3456 if (adev->pm.pcie_mlw_mask == 0) {
3457 pdev = adev->ddev->pdev->bus->self;
3458 link_width = pcie_get_width_cap(pdev);
3459 if (link_width == PCIE_LNK_WIDTH_UNKNOWN) {
3460 adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK;
3462 switch (link_width) {
3464 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
3465 CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
3466 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3467 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3468 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3469 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3470 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3473 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
3474 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3475 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3476 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3477 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3478 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3481 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3482 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3483 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3484 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3485 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3488 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3489 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3490 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3491 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3494 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3495 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3496 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3499 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3500 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3503 adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;