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/kthread.h>
29 #include <linux/console.h>
30 #include <linux/slab.h>
32 #include <drm/drm_crtc_helper.h>
33 #include <drm/drm_atomic_helper.h>
34 #include <drm/amdgpu_drm.h>
35 #include <linux/vgaarb.h>
36 #include <linux/vga_switcheroo.h>
37 #include <linux/efi.h>
39 #include "amdgpu_trace.h"
40 #include "amdgpu_i2c.h"
42 #include "amdgpu_atombios.h"
43 #include "amdgpu_atomfirmware.h"
45 #ifdef CONFIG_DRM_AMDGPU_SI
48 #ifdef CONFIG_DRM_AMDGPU_CIK
53 #include "bif/bif_4_1_d.h"
54 #include <linux/pci.h>
55 #include <linux/firmware.h>
56 #include "amdgpu_vf_error.h"
58 #include "amdgpu_amdkfd.h"
59 #include "amdgpu_pm.h"
61 MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin");
62 MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin");
63 MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin");
65 #define AMDGPU_RESUME_MS 2000
67 static const char *amdgpu_asic_name[] = {
92 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev);
95 * amdgpu_device_is_px - Is the device is a dGPU with HG/PX power control
97 * @dev: drm_device pointer
99 * Returns true if the device is a dGPU with HG/PX power control,
100 * otherwise return false.
102 bool amdgpu_device_is_px(struct drm_device *dev)
104 struct amdgpu_device *adev = dev->dev_private;
106 if (adev->flags & AMD_IS_PX)
112 * MMIO register access helper functions.
115 * amdgpu_mm_rreg - read a memory mapped IO register
117 * @adev: amdgpu_device pointer
118 * @reg: dword aligned register offset
119 * @acc_flags: access flags which require special behavior
121 * Returns the 32 bit value from the offset specified.
123 uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg,
128 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
129 return amdgpu_virt_kiq_rreg(adev, reg);
131 if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
132 ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
136 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
137 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
138 ret = readl(((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
139 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
141 trace_amdgpu_mm_rreg(adev->pdev->device, reg, ret);
146 * MMIO register read with bytes helper functions
147 * @offset:bytes offset from MMIO start
152 * amdgpu_mm_rreg8 - read a memory mapped IO register
154 * @adev: amdgpu_device pointer
155 * @offset: byte aligned register offset
157 * Returns the 8 bit value from the offset specified.
159 uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset) {
160 if (offset < adev->rmmio_size)
161 return (readb(adev->rmmio + offset));
166 * MMIO register write with bytes helper functions
167 * @offset:bytes offset from MMIO start
168 * @value: the value want to be written to the register
172 * amdgpu_mm_wreg8 - read a memory mapped IO register
174 * @adev: amdgpu_device pointer
175 * @offset: byte aligned register offset
176 * @value: 8 bit value to write
178 * Writes the value specified to the offset specified.
180 void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value) {
181 if (offset < adev->rmmio_size)
182 writeb(value, adev->rmmio + offset);
188 * amdgpu_mm_wreg - write to a memory mapped IO register
190 * @adev: amdgpu_device pointer
191 * @reg: dword aligned register offset
192 * @v: 32 bit value to write to the register
193 * @acc_flags: access flags which require special behavior
195 * Writes the value specified to the offset specified.
197 void amdgpu_mm_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v,
200 trace_amdgpu_mm_wreg(adev->pdev->device, reg, v);
202 if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
203 adev->last_mm_index = v;
206 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
207 return amdgpu_virt_kiq_wreg(adev, reg, v);
209 if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
210 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
214 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
215 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
216 writel(v, ((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
217 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
220 if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
226 * amdgpu_io_rreg - read an IO register
228 * @adev: amdgpu_device pointer
229 * @reg: dword aligned register offset
231 * Returns the 32 bit value from the offset specified.
233 u32 amdgpu_io_rreg(struct amdgpu_device *adev, u32 reg)
235 if ((reg * 4) < adev->rio_mem_size)
236 return ioread32(adev->rio_mem + (reg * 4));
238 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
239 return ioread32(adev->rio_mem + (mmMM_DATA * 4));
244 * amdgpu_io_wreg - write to an IO register
246 * @adev: amdgpu_device pointer
247 * @reg: dword aligned register offset
248 * @v: 32 bit value to write to the register
250 * Writes the value specified to the offset specified.
252 void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32 v)
254 if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
255 adev->last_mm_index = v;
258 if ((reg * 4) < adev->rio_mem_size)
259 iowrite32(v, adev->rio_mem + (reg * 4));
261 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
262 iowrite32(v, adev->rio_mem + (mmMM_DATA * 4));
265 if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
271 * amdgpu_mm_rdoorbell - read a doorbell dword
273 * @adev: amdgpu_device pointer
274 * @index: doorbell index
276 * Returns the value in the doorbell aperture at the
277 * requested doorbell index (CIK).
279 u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
281 if (index < adev->doorbell.num_doorbells) {
282 return readl(adev->doorbell.ptr + index);
284 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
290 * amdgpu_mm_wdoorbell - write a doorbell dword
292 * @adev: amdgpu_device pointer
293 * @index: doorbell index
296 * Writes @v to the doorbell aperture at the
297 * requested doorbell index (CIK).
299 void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v)
301 if (index < adev->doorbell.num_doorbells) {
302 writel(v, adev->doorbell.ptr + index);
304 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
309 * amdgpu_mm_rdoorbell64 - read a doorbell Qword
311 * @adev: amdgpu_device pointer
312 * @index: doorbell index
314 * Returns the value in the doorbell aperture at the
315 * requested doorbell index (VEGA10+).
317 u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index)
319 if (index < adev->doorbell.num_doorbells) {
320 return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index));
322 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
328 * amdgpu_mm_wdoorbell64 - write a doorbell Qword
330 * @adev: amdgpu_device pointer
331 * @index: doorbell index
334 * Writes @v to the doorbell aperture at the
335 * requested doorbell index (VEGA10+).
337 void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v)
339 if (index < adev->doorbell.num_doorbells) {
340 atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v);
342 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
347 * amdgpu_invalid_rreg - dummy reg read function
349 * @adev: amdgpu device pointer
350 * @reg: offset of register
352 * Dummy register read function. Used for register blocks
353 * that certain asics don't have (all asics).
354 * Returns the value in the register.
356 static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
358 DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
364 * amdgpu_invalid_wreg - dummy reg write function
366 * @adev: amdgpu device pointer
367 * @reg: offset of register
368 * @v: value to write to the register
370 * Dummy register read function. Used for register blocks
371 * that certain asics don't have (all asics).
373 static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
375 DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
381 * amdgpu_block_invalid_rreg - dummy reg read function
383 * @adev: amdgpu device pointer
384 * @block: offset of instance
385 * @reg: offset of register
387 * Dummy register read function. Used for register blocks
388 * that certain asics don't have (all asics).
389 * Returns the value in the register.
391 static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
392 uint32_t block, uint32_t reg)
394 DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
401 * amdgpu_block_invalid_wreg - dummy reg write function
403 * @adev: amdgpu device pointer
404 * @block: offset of instance
405 * @reg: offset of register
406 * @v: value to write to the register
408 * Dummy register read function. Used for register blocks
409 * that certain asics don't have (all asics).
411 static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
413 uint32_t reg, uint32_t v)
415 DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
421 * amdgpu_device_vram_scratch_init - allocate the VRAM scratch page
423 * @adev: amdgpu device pointer
425 * Allocates a scratch page of VRAM for use by various things in the
428 static int amdgpu_device_vram_scratch_init(struct amdgpu_device *adev)
430 return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE,
431 PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
432 &adev->vram_scratch.robj,
433 &adev->vram_scratch.gpu_addr,
434 (void **)&adev->vram_scratch.ptr);
438 * amdgpu_device_vram_scratch_fini - Free the VRAM scratch page
440 * @adev: amdgpu device pointer
442 * Frees the VRAM scratch page.
444 static void amdgpu_device_vram_scratch_fini(struct amdgpu_device *adev)
446 amdgpu_bo_free_kernel(&adev->vram_scratch.robj, NULL, NULL);
450 * amdgpu_device_program_register_sequence - program an array of registers.
452 * @adev: amdgpu_device pointer
453 * @registers: pointer to the register array
454 * @array_size: size of the register array
456 * Programs an array or registers with and and or masks.
457 * This is a helper for setting golden registers.
459 void amdgpu_device_program_register_sequence(struct amdgpu_device *adev,
460 const u32 *registers,
461 const u32 array_size)
463 u32 tmp, reg, and_mask, or_mask;
469 for (i = 0; i < array_size; i +=3) {
470 reg = registers[i + 0];
471 and_mask = registers[i + 1];
472 or_mask = registers[i + 2];
474 if (and_mask == 0xffffffff) {
486 * amdgpu_device_pci_config_reset - reset the GPU
488 * @adev: amdgpu_device pointer
490 * Resets the GPU using the pci config reset sequence.
491 * Only applicable to asics prior to vega10.
493 void amdgpu_device_pci_config_reset(struct amdgpu_device *adev)
495 pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
499 * GPU doorbell aperture helpers function.
502 * amdgpu_device_doorbell_init - Init doorbell driver information.
504 * @adev: amdgpu_device pointer
506 * Init doorbell driver information (CIK)
507 * Returns 0 on success, error on failure.
509 static int amdgpu_device_doorbell_init(struct amdgpu_device *adev)
511 /* No doorbell on SI hardware generation */
512 if (adev->asic_type < CHIP_BONAIRE) {
513 adev->doorbell.base = 0;
514 adev->doorbell.size = 0;
515 adev->doorbell.num_doorbells = 0;
516 adev->doorbell.ptr = NULL;
520 if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET)
523 /* doorbell bar mapping */
524 adev->doorbell.base = pci_resource_start(adev->pdev, 2);
525 adev->doorbell.size = pci_resource_len(adev->pdev, 2);
527 adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32),
528 AMDGPU_DOORBELL_MAX_ASSIGNMENT+1);
529 if (adev->doorbell.num_doorbells == 0)
532 adev->doorbell.ptr = ioremap(adev->doorbell.base,
533 adev->doorbell.num_doorbells *
535 if (adev->doorbell.ptr == NULL)
542 * amdgpu_device_doorbell_fini - Tear down doorbell driver information.
544 * @adev: amdgpu_device pointer
546 * Tear down doorbell driver information (CIK)
548 static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev)
550 iounmap(adev->doorbell.ptr);
551 adev->doorbell.ptr = NULL;
557 * amdgpu_device_wb_*()
558 * Writeback is the method by which the GPU updates special pages in memory
559 * with the status of certain GPU events (fences, ring pointers,etc.).
563 * amdgpu_device_wb_fini - Disable Writeback and free memory
565 * @adev: amdgpu_device pointer
567 * Disables Writeback and frees the Writeback memory (all asics).
568 * Used at driver shutdown.
570 static void amdgpu_device_wb_fini(struct amdgpu_device *adev)
572 if (adev->wb.wb_obj) {
573 amdgpu_bo_free_kernel(&adev->wb.wb_obj,
575 (void **)&adev->wb.wb);
576 adev->wb.wb_obj = NULL;
581 * amdgpu_device_wb_init- Init Writeback driver info and allocate memory
583 * @adev: amdgpu_device pointer
585 * Initializes writeback and allocates writeback memory (all asics).
586 * Used at driver startup.
587 * Returns 0 on success or an -error on failure.
589 static int amdgpu_device_wb_init(struct amdgpu_device *adev)
593 if (adev->wb.wb_obj == NULL) {
594 /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */
595 r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8,
596 PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
597 &adev->wb.wb_obj, &adev->wb.gpu_addr,
598 (void **)&adev->wb.wb);
600 dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
604 adev->wb.num_wb = AMDGPU_MAX_WB;
605 memset(&adev->wb.used, 0, sizeof(adev->wb.used));
607 /* clear wb memory */
608 memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8);
615 * amdgpu_device_wb_get - Allocate a wb entry
617 * @adev: amdgpu_device pointer
620 * Allocate a wb slot for use by the driver (all asics).
621 * Returns 0 on success or -EINVAL on failure.
623 int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb)
625 unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
627 if (offset < adev->wb.num_wb) {
628 __set_bit(offset, adev->wb.used);
629 *wb = offset << 3; /* convert to dw offset */
637 * amdgpu_device_wb_free - Free a wb entry
639 * @adev: amdgpu_device pointer
642 * Free a wb slot allocated for use by the driver (all asics)
644 void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb)
647 if (wb < adev->wb.num_wb)
648 __clear_bit(wb, adev->wb.used);
652 * amdgpu_device_vram_location - try to find VRAM location
654 * @adev: amdgpu device structure holding all necessary informations
655 * @mc: memory controller structure holding memory informations
656 * @base: base address at which to put VRAM
658 * Function will try to place VRAM at base address provided
661 void amdgpu_device_vram_location(struct amdgpu_device *adev,
662 struct amdgpu_gmc *mc, u64 base)
664 uint64_t limit = (uint64_t)amdgpu_vram_limit << 20;
666 mc->vram_start = base;
667 mc->vram_end = mc->vram_start + mc->mc_vram_size - 1;
668 if (limit && limit < mc->real_vram_size)
669 mc->real_vram_size = limit;
670 dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
671 mc->mc_vram_size >> 20, mc->vram_start,
672 mc->vram_end, mc->real_vram_size >> 20);
676 * amdgpu_device_gart_location - try to find GTT location
678 * @adev: amdgpu device structure holding all necessary informations
679 * @mc: memory controller structure holding memory informations
681 * Function will place try to place GTT before or after VRAM.
683 * If GTT size is bigger than space left then we ajust GTT size.
684 * Thus function will never fails.
686 * FIXME: when reducing GTT size align new size on power of 2.
688 void amdgpu_device_gart_location(struct amdgpu_device *adev,
689 struct amdgpu_gmc *mc)
691 u64 size_af, size_bf;
693 size_af = adev->gmc.mc_mask - mc->vram_end;
694 size_bf = mc->vram_start;
695 if (size_bf > size_af) {
696 if (mc->gart_size > size_bf) {
697 dev_warn(adev->dev, "limiting GTT\n");
698 mc->gart_size = size_bf;
702 if (mc->gart_size > size_af) {
703 dev_warn(adev->dev, "limiting GTT\n");
704 mc->gart_size = size_af;
706 /* VCE doesn't like it when BOs cross a 4GB segment, so align
707 * the GART base on a 4GB boundary as well.
709 mc->gart_start = ALIGN(mc->vram_end + 1, 0x100000000ULL);
711 mc->gart_end = mc->gart_start + mc->gart_size - 1;
712 dev_info(adev->dev, "GTT: %lluM 0x%016llX - 0x%016llX\n",
713 mc->gart_size >> 20, mc->gart_start, mc->gart_end);
717 * amdgpu_device_resize_fb_bar - try to resize FB BAR
719 * @adev: amdgpu_device pointer
721 * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not
722 * to fail, but if any of the BARs is not accessible after the size we abort
723 * driver loading by returning -ENODEV.
725 int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev)
727 u64 space_needed = roundup_pow_of_two(adev->gmc.real_vram_size);
728 u32 rbar_size = order_base_2(((space_needed >> 20) | 1)) - 1;
729 struct pci_bus *root;
730 struct resource *res;
736 if (amdgpu_sriov_vf(adev))
739 /* Check if the root BUS has 64bit memory resources */
740 root = adev->pdev->bus;
744 pci_bus_for_each_resource(root, res, i) {
745 if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) &&
746 res->start > 0x100000000ull)
750 /* Trying to resize is pointless without a root hub window above 4GB */
754 /* Disable memory decoding while we change the BAR addresses and size */
755 pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd);
756 pci_write_config_word(adev->pdev, PCI_COMMAND,
757 cmd & ~PCI_COMMAND_MEMORY);
759 /* Free the VRAM and doorbell BAR, we most likely need to move both. */
760 amdgpu_device_doorbell_fini(adev);
761 if (adev->asic_type >= CHIP_BONAIRE)
762 pci_release_resource(adev->pdev, 2);
764 pci_release_resource(adev->pdev, 0);
766 r = pci_resize_resource(adev->pdev, 0, rbar_size);
768 DRM_INFO("Not enough PCI address space for a large BAR.");
769 else if (r && r != -ENOTSUPP)
770 DRM_ERROR("Problem resizing BAR0 (%d).", r);
772 pci_assign_unassigned_bus_resources(adev->pdev->bus);
774 /* When the doorbell or fb BAR isn't available we have no chance of
777 r = amdgpu_device_doorbell_init(adev);
778 if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET))
781 pci_write_config_word(adev->pdev, PCI_COMMAND, cmd);
787 * GPU helpers function.
790 * amdgpu_device_need_post - check if the hw need post or not
792 * @adev: amdgpu_device pointer
794 * Check if the asic has been initialized (all asics) at driver startup
795 * or post is needed if hw reset is performed.
796 * Returns true if need or false if not.
798 bool amdgpu_device_need_post(struct amdgpu_device *adev)
802 if (amdgpu_sriov_vf(adev))
805 if (amdgpu_passthrough(adev)) {
806 /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
807 * some old smc fw still need driver do vPost otherwise gpu hang, while
808 * those smc fw version above 22.15 doesn't have this flaw, so we force
809 * vpost executed for smc version below 22.15
811 if (adev->asic_type == CHIP_FIJI) {
814 err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
815 /* force vPost if error occured */
819 fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
820 if (fw_ver < 0x00160e00)
825 if (adev->has_hw_reset) {
826 adev->has_hw_reset = false;
830 /* bios scratch used on CIK+ */
831 if (adev->asic_type >= CHIP_BONAIRE)
832 return amdgpu_atombios_scratch_need_asic_init(adev);
834 /* check MEM_SIZE for older asics */
835 reg = amdgpu_asic_get_config_memsize(adev);
837 if ((reg != 0) && (reg != 0xffffffff))
843 /* if we get transitioned to only one device, take VGA back */
845 * amdgpu_device_vga_set_decode - enable/disable vga decode
847 * @cookie: amdgpu_device pointer
848 * @state: enable/disable vga decode
850 * Enable/disable vga decode (all asics).
851 * Returns VGA resource flags.
853 static unsigned int amdgpu_device_vga_set_decode(void *cookie, bool state)
855 struct amdgpu_device *adev = cookie;
856 amdgpu_asic_set_vga_state(adev, state);
858 return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
859 VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
861 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
865 * amdgpu_device_check_block_size - validate the vm block size
867 * @adev: amdgpu_device pointer
869 * Validates the vm block size specified via module parameter.
870 * The vm block size defines number of bits in page table versus page directory,
871 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
872 * page table and the remaining bits are in the page directory.
874 static void amdgpu_device_check_block_size(struct amdgpu_device *adev)
876 /* defines number of bits in page table versus page directory,
877 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
878 * page table and the remaining bits are in the page directory */
879 if (amdgpu_vm_block_size == -1)
882 if (amdgpu_vm_block_size < 9) {
883 dev_warn(adev->dev, "VM page table size (%d) too small\n",
884 amdgpu_vm_block_size);
885 amdgpu_vm_block_size = -1;
890 * amdgpu_device_check_vm_size - validate the vm size
892 * @adev: amdgpu_device pointer
894 * Validates the vm size in GB specified via module parameter.
895 * The VM size is the size of the GPU virtual memory space in GB.
897 static void amdgpu_device_check_vm_size(struct amdgpu_device *adev)
899 /* no need to check the default value */
900 if (amdgpu_vm_size == -1)
903 if (amdgpu_vm_size < 1) {
904 dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
911 * amdgpu_device_check_arguments - validate module params
913 * @adev: amdgpu_device pointer
915 * Validates certain module parameters and updates
916 * the associated values used by the driver (all asics).
918 static void amdgpu_device_check_arguments(struct amdgpu_device *adev)
920 if (amdgpu_sched_jobs < 4) {
921 dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
923 amdgpu_sched_jobs = 4;
924 } else if (!is_power_of_2(amdgpu_sched_jobs)){
925 dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
927 amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
930 if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) {
931 /* gart size must be greater or equal to 32M */
932 dev_warn(adev->dev, "gart size (%d) too small\n",
934 amdgpu_gart_size = -1;
937 if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) {
938 /* gtt size must be greater or equal to 32M */
939 dev_warn(adev->dev, "gtt size (%d) too small\n",
941 amdgpu_gtt_size = -1;
944 /* valid range is between 4 and 9 inclusive */
945 if (amdgpu_vm_fragment_size != -1 &&
946 (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) {
947 dev_warn(adev->dev, "valid range is between 4 and 9\n");
948 amdgpu_vm_fragment_size = -1;
951 amdgpu_device_check_vm_size(adev);
953 amdgpu_device_check_block_size(adev);
955 if (amdgpu_vram_page_split != -1 && (amdgpu_vram_page_split < 16 ||
956 !is_power_of_2(amdgpu_vram_page_split))) {
957 dev_warn(adev->dev, "invalid VRAM page split (%d)\n",
958 amdgpu_vram_page_split);
959 amdgpu_vram_page_split = 1024;
962 if (amdgpu_lockup_timeout == 0) {
963 dev_warn(adev->dev, "lockup_timeout msut be > 0, adjusting to 10000\n");
964 amdgpu_lockup_timeout = 10000;
967 adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type);
971 * amdgpu_switcheroo_set_state - set switcheroo state
973 * @pdev: pci dev pointer
974 * @state: vga_switcheroo state
976 * Callback for the switcheroo driver. Suspends or resumes the
977 * the asics before or after it is powered up using ACPI methods.
979 static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, enum vga_switcheroo_state state)
981 struct drm_device *dev = pci_get_drvdata(pdev);
983 if (amdgpu_device_is_px(dev) && state == VGA_SWITCHEROO_OFF)
986 if (state == VGA_SWITCHEROO_ON) {
987 pr_info("amdgpu: switched on\n");
988 /* don't suspend or resume card normally */
989 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
991 amdgpu_device_resume(dev, true, true);
993 dev->switch_power_state = DRM_SWITCH_POWER_ON;
994 drm_kms_helper_poll_enable(dev);
996 pr_info("amdgpu: switched off\n");
997 drm_kms_helper_poll_disable(dev);
998 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
999 amdgpu_device_suspend(dev, true, true);
1000 dev->switch_power_state = DRM_SWITCH_POWER_OFF;
1005 * amdgpu_switcheroo_can_switch - see if switcheroo state can change
1007 * @pdev: pci dev pointer
1009 * Callback for the switcheroo driver. Check of the switcheroo
1010 * state can be changed.
1011 * Returns true if the state can be changed, false if not.
1013 static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
1015 struct drm_device *dev = pci_get_drvdata(pdev);
1018 * FIXME: open_count is protected by drm_global_mutex but that would lead to
1019 * locking inversion with the driver load path. And the access here is
1020 * completely racy anyway. So don't bother with locking for now.
1022 return dev->open_count == 0;
1025 static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
1026 .set_gpu_state = amdgpu_switcheroo_set_state,
1028 .can_switch = amdgpu_switcheroo_can_switch,
1032 * amdgpu_device_ip_set_clockgating_state - set the CG state
1034 * @adev: amdgpu_device pointer
1035 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1036 * @state: clockgating state (gate or ungate)
1038 * Sets the requested clockgating state for all instances of
1039 * the hardware IP specified.
1040 * Returns the error code from the last instance.
1042 int amdgpu_device_ip_set_clockgating_state(void *dev,
1043 enum amd_ip_block_type block_type,
1044 enum amd_clockgating_state state)
1046 struct amdgpu_device *adev = dev;
1049 for (i = 0; i < adev->num_ip_blocks; i++) {
1050 if (!adev->ip_blocks[i].status.valid)
1052 if (adev->ip_blocks[i].version->type != block_type)
1054 if (!adev->ip_blocks[i].version->funcs->set_clockgating_state)
1056 r = adev->ip_blocks[i].version->funcs->set_clockgating_state(
1057 (void *)adev, state);
1059 DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",
1060 adev->ip_blocks[i].version->funcs->name, r);
1066 * amdgpu_device_ip_set_powergating_state - set the PG state
1068 * @adev: amdgpu_device pointer
1069 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1070 * @state: powergating state (gate or ungate)
1072 * Sets the requested powergating state for all instances of
1073 * the hardware IP specified.
1074 * Returns the error code from the last instance.
1076 int amdgpu_device_ip_set_powergating_state(void *dev,
1077 enum amd_ip_block_type block_type,
1078 enum amd_powergating_state state)
1080 struct amdgpu_device *adev = dev;
1083 for (i = 0; i < adev->num_ip_blocks; i++) {
1084 if (!adev->ip_blocks[i].status.valid)
1086 if (adev->ip_blocks[i].version->type != block_type)
1088 if (!adev->ip_blocks[i].version->funcs->set_powergating_state)
1090 r = adev->ip_blocks[i].version->funcs->set_powergating_state(
1091 (void *)adev, state);
1093 DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",
1094 adev->ip_blocks[i].version->funcs->name, r);
1100 * amdgpu_device_ip_get_clockgating_state - get the CG state
1102 * @adev: amdgpu_device pointer
1103 * @flags: clockgating feature flags
1105 * Walks the list of IPs on the device and updates the clockgating
1106 * flags for each IP.
1107 * Updates @flags with the feature flags for each hardware IP where
1108 * clockgating is enabled.
1110 void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev,
1115 for (i = 0; i < adev->num_ip_blocks; i++) {
1116 if (!adev->ip_blocks[i].status.valid)
1118 if (adev->ip_blocks[i].version->funcs->get_clockgating_state)
1119 adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags);
1124 * amdgpu_device_ip_wait_for_idle - wait for idle
1126 * @adev: amdgpu_device pointer
1127 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1129 * Waits for the request hardware IP to be idle.
1130 * Returns 0 for success or a negative error code on failure.
1132 int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev,
1133 enum amd_ip_block_type block_type)
1137 for (i = 0; i < adev->num_ip_blocks; i++) {
1138 if (!adev->ip_blocks[i].status.valid)
1140 if (adev->ip_blocks[i].version->type == block_type) {
1141 r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
1152 * amdgpu_device_ip_is_idle - is the hardware IP idle
1154 * @adev: amdgpu_device pointer
1155 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1157 * Check if the hardware IP is idle or not.
1158 * Returns true if it the IP is idle, false if not.
1160 bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev,
1161 enum amd_ip_block_type block_type)
1165 for (i = 0; i < adev->num_ip_blocks; i++) {
1166 if (!adev->ip_blocks[i].status.valid)
1168 if (adev->ip_blocks[i].version->type == block_type)
1169 return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
1176 * amdgpu_device_ip_get_ip_block - get a hw IP pointer
1178 * @adev: amdgpu_device pointer
1179 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1181 * Returns a pointer to the hardware IP block structure
1182 * if it exists for the asic, otherwise NULL.
1184 struct amdgpu_ip_block *
1185 amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,
1186 enum amd_ip_block_type type)
1190 for (i = 0; i < adev->num_ip_blocks; i++)
1191 if (adev->ip_blocks[i].version->type == type)
1192 return &adev->ip_blocks[i];
1198 * amdgpu_device_ip_block_version_cmp
1200 * @adev: amdgpu_device pointer
1201 * @type: enum amd_ip_block_type
1202 * @major: major version
1203 * @minor: minor version
1205 * return 0 if equal or greater
1206 * return 1 if smaller or the ip_block doesn't exist
1208 int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev,
1209 enum amd_ip_block_type type,
1210 u32 major, u32 minor)
1212 struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type);
1214 if (ip_block && ((ip_block->version->major > major) ||
1215 ((ip_block->version->major == major) &&
1216 (ip_block->version->minor >= minor))))
1223 * amdgpu_device_ip_block_add
1225 * @adev: amdgpu_device pointer
1226 * @ip_block_version: pointer to the IP to add
1228 * Adds the IP block driver information to the collection of IPs
1231 int amdgpu_device_ip_block_add(struct amdgpu_device *adev,
1232 const struct amdgpu_ip_block_version *ip_block_version)
1234 if (!ip_block_version)
1237 DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks,
1238 ip_block_version->funcs->name);
1240 adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
1246 * amdgpu_device_enable_virtual_display - enable virtual display feature
1248 * @adev: amdgpu_device pointer
1250 * Enabled the virtual display feature if the user has enabled it via
1251 * the module parameter virtual_display. This feature provides a virtual
1252 * display hardware on headless boards or in virtualized environments.
1253 * This function parses and validates the configuration string specified by
1254 * the user and configues the virtual display configuration (number of
1255 * virtual connectors, crtcs, etc.) specified.
1257 static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
1259 adev->enable_virtual_display = false;
1261 if (amdgpu_virtual_display) {
1262 struct drm_device *ddev = adev->ddev;
1263 const char *pci_address_name = pci_name(ddev->pdev);
1264 char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
1266 pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
1267 pciaddstr_tmp = pciaddstr;
1268 while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
1269 pciaddname = strsep(&pciaddname_tmp, ",");
1270 if (!strcmp("all", pciaddname)
1271 || !strcmp(pci_address_name, pciaddname)) {
1275 adev->enable_virtual_display = true;
1278 res = kstrtol(pciaddname_tmp, 10,
1286 adev->mode_info.num_crtc = num_crtc;
1288 adev->mode_info.num_crtc = 1;
1294 DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
1295 amdgpu_virtual_display, pci_address_name,
1296 adev->enable_virtual_display, adev->mode_info.num_crtc);
1303 * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware
1305 * @adev: amdgpu_device pointer
1307 * Parses the asic configuration parameters specified in the gpu info
1308 * firmware and makes them availale to the driver for use in configuring
1310 * Returns 0 on success, -EINVAL on failure.
1312 static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev)
1314 const char *chip_name;
1317 const struct gpu_info_firmware_header_v1_0 *hdr;
1319 adev->firmware.gpu_info_fw = NULL;
1321 switch (adev->asic_type) {
1325 case CHIP_POLARIS11:
1326 case CHIP_POLARIS10:
1327 case CHIP_POLARIS12:
1330 #ifdef CONFIG_DRM_AMDGPU_SI
1337 #ifdef CONFIG_DRM_AMDGPU_CIK
1347 chip_name = "vega10";
1350 chip_name = "vega12";
1353 chip_name = "raven";
1357 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name);
1358 err = request_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev);
1361 "Failed to load gpu_info firmware \"%s\"\n",
1365 err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw);
1368 "Failed to validate gpu_info firmware \"%s\"\n",
1373 hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data;
1374 amdgpu_ucode_print_gpu_info_hdr(&hdr->header);
1376 switch (hdr->version_major) {
1379 const struct gpu_info_firmware_v1_0 *gpu_info_fw =
1380 (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data +
1381 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1383 adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se);
1384 adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh);
1385 adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se);
1386 adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se);
1387 adev->gfx.config.max_texture_channel_caches =
1388 le32_to_cpu(gpu_info_fw->gc_num_tccs);
1389 adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs);
1390 adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds);
1391 adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth);
1392 adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth);
1393 adev->gfx.config.double_offchip_lds_buf =
1394 le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer);
1395 adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size);
1396 adev->gfx.cu_info.max_waves_per_simd =
1397 le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd);
1398 adev->gfx.cu_info.max_scratch_slots_per_cu =
1399 le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu);
1400 adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size);
1405 "Unsupported gpu_info table %d\n", hdr->header.ucode_version);
1414 * amdgpu_device_ip_early_init - run early init for hardware IPs
1416 * @adev: amdgpu_device pointer
1418 * Early initialization pass for hardware IPs. The hardware IPs that make
1419 * up each asic are discovered each IP's early_init callback is run. This
1420 * is the first stage in initializing the asic.
1421 * Returns 0 on success, negative error code on failure.
1423 static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
1427 amdgpu_device_enable_virtual_display(adev);
1429 switch (adev->asic_type) {
1433 case CHIP_POLARIS11:
1434 case CHIP_POLARIS10:
1435 case CHIP_POLARIS12:
1438 if (adev->asic_type == CHIP_CARRIZO || adev->asic_type == CHIP_STONEY)
1439 adev->family = AMDGPU_FAMILY_CZ;
1441 adev->family = AMDGPU_FAMILY_VI;
1443 r = vi_set_ip_blocks(adev);
1447 #ifdef CONFIG_DRM_AMDGPU_SI
1453 adev->family = AMDGPU_FAMILY_SI;
1454 r = si_set_ip_blocks(adev);
1459 #ifdef CONFIG_DRM_AMDGPU_CIK
1465 if ((adev->asic_type == CHIP_BONAIRE) || (adev->asic_type == CHIP_HAWAII))
1466 adev->family = AMDGPU_FAMILY_CI;
1468 adev->family = AMDGPU_FAMILY_KV;
1470 r = cik_set_ip_blocks(adev);
1478 if (adev->asic_type == CHIP_RAVEN)
1479 adev->family = AMDGPU_FAMILY_RV;
1481 adev->family = AMDGPU_FAMILY_AI;
1483 r = soc15_set_ip_blocks(adev);
1488 /* FIXME: not supported yet */
1492 r = amdgpu_device_parse_gpu_info_fw(adev);
1496 amdgpu_amdkfd_device_probe(adev);
1498 if (amdgpu_sriov_vf(adev)) {
1499 r = amdgpu_virt_request_full_gpu(adev, true);
1504 for (i = 0; i < adev->num_ip_blocks; i++) {
1505 if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
1506 DRM_ERROR("disabled ip block: %d <%s>\n",
1507 i, adev->ip_blocks[i].version->funcs->name);
1508 adev->ip_blocks[i].status.valid = false;
1510 if (adev->ip_blocks[i].version->funcs->early_init) {
1511 r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
1513 adev->ip_blocks[i].status.valid = false;
1515 DRM_ERROR("early_init of IP block <%s> failed %d\n",
1516 adev->ip_blocks[i].version->funcs->name, r);
1519 adev->ip_blocks[i].status.valid = true;
1522 adev->ip_blocks[i].status.valid = true;
1527 adev->cg_flags &= amdgpu_cg_mask;
1528 adev->pg_flags &= amdgpu_pg_mask;
1534 * amdgpu_device_ip_init - run init for hardware IPs
1536 * @adev: amdgpu_device pointer
1538 * Main initialization pass for hardware IPs. The list of all the hardware
1539 * IPs that make up the asic is walked and the sw_init and hw_init callbacks
1540 * are run. sw_init initializes the software state associated with each IP
1541 * and hw_init initializes the hardware associated with each IP.
1542 * Returns 0 on success, negative error code on failure.
1544 static int amdgpu_device_ip_init(struct amdgpu_device *adev)
1548 for (i = 0; i < adev->num_ip_blocks; i++) {
1549 if (!adev->ip_blocks[i].status.valid)
1551 r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
1553 DRM_ERROR("sw_init of IP block <%s> failed %d\n",
1554 adev->ip_blocks[i].version->funcs->name, r);
1557 adev->ip_blocks[i].status.sw = true;
1559 /* need to do gmc hw init early so we can allocate gpu mem */
1560 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1561 r = amdgpu_device_vram_scratch_init(adev);
1563 DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r);
1566 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1568 DRM_ERROR("hw_init %d failed %d\n", i, r);
1571 r = amdgpu_device_wb_init(adev);
1573 DRM_ERROR("amdgpu_device_wb_init failed %d\n", r);
1576 adev->ip_blocks[i].status.hw = true;
1578 /* right after GMC hw init, we create CSA */
1579 if (amdgpu_sriov_vf(adev)) {
1580 r = amdgpu_allocate_static_csa(adev);
1582 DRM_ERROR("allocate CSA failed %d\n", r);
1589 for (i = 0; i < adev->num_ip_blocks; i++) {
1590 if (!adev->ip_blocks[i].status.sw)
1592 if (adev->ip_blocks[i].status.hw)
1594 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1596 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1597 adev->ip_blocks[i].version->funcs->name, r);
1600 adev->ip_blocks[i].status.hw = true;
1603 amdgpu_amdkfd_device_init(adev);
1605 if (amdgpu_sriov_vf(adev))
1606 amdgpu_virt_release_full_gpu(adev, true);
1612 * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer
1614 * @adev: amdgpu_device pointer
1616 * Writes a reset magic value to the gart pointer in VRAM. The driver calls
1617 * this function before a GPU reset. If the value is retained after a
1618 * GPU reset, VRAM has not been lost. Some GPU resets may destry VRAM contents.
1620 static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev)
1622 memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
1626 * amdgpu_device_check_vram_lost - check if vram is valid
1628 * @adev: amdgpu_device pointer
1630 * Checks the reset magic value written to the gart pointer in VRAM.
1631 * The driver calls this after a GPU reset to see if the contents of
1632 * VRAM is lost or now.
1633 * returns true if vram is lost, false if not.
1635 static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev)
1637 return !!memcmp(adev->gart.ptr, adev->reset_magic,
1638 AMDGPU_RESET_MAGIC_NUM);
1642 * amdgpu_device_ip_late_set_cg_state - late init for clockgating
1644 * @adev: amdgpu_device pointer
1646 * Late initialization pass enabling clockgating for hardware IPs.
1647 * The list of all the hardware IPs that make up the asic is walked and the
1648 * set_clockgating_state callbacks are run. This stage is run late
1649 * in the init process.
1650 * Returns 0 on success, negative error code on failure.
1652 static int amdgpu_device_ip_late_set_cg_state(struct amdgpu_device *adev)
1656 if (amdgpu_emu_mode == 1)
1659 r = amdgpu_ib_ring_tests(adev);
1661 DRM_ERROR("ib ring test failed (%d).\n", r);
1663 for (i = 0; i < adev->num_ip_blocks; i++) {
1664 if (!adev->ip_blocks[i].status.valid)
1666 /* skip CG for VCE/UVD, it's handled specially */
1667 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1668 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1669 adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1670 /* enable clockgating to save power */
1671 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1674 DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
1675 adev->ip_blocks[i].version->funcs->name, r);
1684 * amdgpu_device_ip_late_init - run late init for hardware IPs
1686 * @adev: amdgpu_device pointer
1688 * Late initialization pass for hardware IPs. The list of all the hardware
1689 * IPs that make up the asic is walked and the late_init callbacks are run.
1690 * late_init covers any special initialization that an IP requires
1691 * after all of the have been initialized or something that needs to happen
1692 * late in the init process.
1693 * Returns 0 on success, negative error code on failure.
1695 static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
1699 for (i = 0; i < adev->num_ip_blocks; i++) {
1700 if (!adev->ip_blocks[i].status.valid)
1702 if (adev->ip_blocks[i].version->funcs->late_init) {
1703 r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
1705 DRM_ERROR("late_init of IP block <%s> failed %d\n",
1706 adev->ip_blocks[i].version->funcs->name, r);
1709 adev->ip_blocks[i].status.late_initialized = true;
1713 queue_delayed_work(system_wq, &adev->late_init_work,
1714 msecs_to_jiffies(AMDGPU_RESUME_MS));
1716 amdgpu_device_fill_reset_magic(adev);
1722 * amdgpu_device_ip_fini - run fini for hardware IPs
1724 * @adev: amdgpu_device pointer
1726 * Main teardown pass for hardware IPs. The list of all the hardware
1727 * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks
1728 * are run. hw_fini tears down the hardware associated with each IP
1729 * and sw_fini tears down any software state associated with each IP.
1730 * Returns 0 on success, negative error code on failure.
1732 static int amdgpu_device_ip_fini(struct amdgpu_device *adev)
1736 amdgpu_amdkfd_device_fini(adev);
1737 /* need to disable SMC first */
1738 for (i = 0; i < adev->num_ip_blocks; i++) {
1739 if (!adev->ip_blocks[i].status.hw)
1741 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC &&
1742 adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1743 /* ungate blocks before hw fini so that we can shutdown the blocks safely */
1744 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1745 AMD_CG_STATE_UNGATE);
1747 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1748 adev->ip_blocks[i].version->funcs->name, r);
1751 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
1752 /* XXX handle errors */
1754 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
1755 adev->ip_blocks[i].version->funcs->name, r);
1757 adev->ip_blocks[i].status.hw = false;
1762 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1763 if (!adev->ip_blocks[i].status.hw)
1766 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1767 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1768 adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1769 /* ungate blocks before hw fini so that we can shutdown the blocks safely */
1770 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1771 AMD_CG_STATE_UNGATE);
1773 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1774 adev->ip_blocks[i].version->funcs->name, r);
1779 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
1780 /* XXX handle errors */
1782 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
1783 adev->ip_blocks[i].version->funcs->name, r);
1786 adev->ip_blocks[i].status.hw = false;
1790 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1791 if (!adev->ip_blocks[i].status.sw)
1794 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1795 amdgpu_free_static_csa(adev);
1796 amdgpu_device_wb_fini(adev);
1797 amdgpu_device_vram_scratch_fini(adev);
1800 r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
1801 /* XXX handle errors */
1803 DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
1804 adev->ip_blocks[i].version->funcs->name, r);
1806 adev->ip_blocks[i].status.sw = false;
1807 adev->ip_blocks[i].status.valid = false;
1810 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1811 if (!adev->ip_blocks[i].status.late_initialized)
1813 if (adev->ip_blocks[i].version->funcs->late_fini)
1814 adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
1815 adev->ip_blocks[i].status.late_initialized = false;
1818 if (amdgpu_sriov_vf(adev))
1819 if (amdgpu_virt_release_full_gpu(adev, false))
1820 DRM_ERROR("failed to release exclusive mode on fini\n");
1826 * amdgpu_device_ip_late_init_func_handler - work handler for clockgating
1828 * @work: work_struct
1830 * Work handler for amdgpu_device_ip_late_set_cg_state. We put the
1831 * clockgating setup into a worker thread to speed up driver init and
1832 * resume from suspend.
1834 static void amdgpu_device_ip_late_init_func_handler(struct work_struct *work)
1836 struct amdgpu_device *adev =
1837 container_of(work, struct amdgpu_device, late_init_work.work);
1838 amdgpu_device_ip_late_set_cg_state(adev);
1842 * amdgpu_device_ip_suspend - run suspend for hardware IPs
1844 * @adev: amdgpu_device pointer
1846 * Main suspend function for hardware IPs. The list of all the hardware
1847 * IPs that make up the asic is walked, clockgating is disabled and the
1848 * suspend callbacks are run. suspend puts the hardware and software state
1849 * in each IP into a state suitable for suspend.
1850 * Returns 0 on success, negative error code on failure.
1852 int amdgpu_device_ip_suspend(struct amdgpu_device *adev)
1856 if (amdgpu_sriov_vf(adev))
1857 amdgpu_virt_request_full_gpu(adev, false);
1859 /* ungate SMC block first */
1860 r = amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_SMC,
1861 AMD_CG_STATE_UNGATE);
1863 DRM_ERROR("set_clockgating_state(ungate) SMC failed %d\n", r);
1866 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1867 if (!adev->ip_blocks[i].status.valid)
1869 /* ungate blocks so that suspend can properly shut them down */
1870 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_SMC &&
1871 adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1872 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1873 AMD_CG_STATE_UNGATE);
1875 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1876 adev->ip_blocks[i].version->funcs->name, r);
1879 /* XXX handle errors */
1880 r = adev->ip_blocks[i].version->funcs->suspend(adev);
1881 /* XXX handle errors */
1883 DRM_ERROR("suspend of IP block <%s> failed %d\n",
1884 adev->ip_blocks[i].version->funcs->name, r);
1888 if (amdgpu_sriov_vf(adev))
1889 amdgpu_virt_release_full_gpu(adev, false);
1894 static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
1898 static enum amd_ip_block_type ip_order[] = {
1899 AMD_IP_BLOCK_TYPE_GMC,
1900 AMD_IP_BLOCK_TYPE_COMMON,
1901 AMD_IP_BLOCK_TYPE_IH,
1904 for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
1906 struct amdgpu_ip_block *block;
1908 for (j = 0; j < adev->num_ip_blocks; j++) {
1909 block = &adev->ip_blocks[j];
1911 if (block->version->type != ip_order[i] ||
1912 !block->status.valid)
1915 r = block->version->funcs->hw_init(adev);
1916 DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"successed");
1925 static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
1929 static enum amd_ip_block_type ip_order[] = {
1930 AMD_IP_BLOCK_TYPE_SMC,
1931 AMD_IP_BLOCK_TYPE_PSP,
1932 AMD_IP_BLOCK_TYPE_DCE,
1933 AMD_IP_BLOCK_TYPE_GFX,
1934 AMD_IP_BLOCK_TYPE_SDMA,
1935 AMD_IP_BLOCK_TYPE_UVD,
1936 AMD_IP_BLOCK_TYPE_VCE
1939 for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
1941 struct amdgpu_ip_block *block;
1943 for (j = 0; j < adev->num_ip_blocks; j++) {
1944 block = &adev->ip_blocks[j];
1946 if (block->version->type != ip_order[i] ||
1947 !block->status.valid)
1950 r = block->version->funcs->hw_init(adev);
1951 DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"successed");
1961 * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs
1963 * @adev: amdgpu_device pointer
1965 * First resume function for hardware IPs. The list of all the hardware
1966 * IPs that make up the asic is walked and the resume callbacks are run for
1967 * COMMON, GMC, and IH. resume puts the hardware into a functional state
1968 * after a suspend and updates the software state as necessary. This
1969 * function is also used for restoring the GPU after a GPU reset.
1970 * Returns 0 on success, negative error code on failure.
1972 static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)
1976 for (i = 0; i < adev->num_ip_blocks; i++) {
1977 if (!adev->ip_blocks[i].status.valid)
1979 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
1980 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
1981 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
1982 r = adev->ip_blocks[i].version->funcs->resume(adev);
1984 DRM_ERROR("resume of IP block <%s> failed %d\n",
1985 adev->ip_blocks[i].version->funcs->name, r);
1995 * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs
1997 * @adev: amdgpu_device pointer
1999 * First resume function for hardware IPs. The list of all the hardware
2000 * IPs that make up the asic is walked and the resume callbacks are run for
2001 * all blocks except COMMON, GMC, and IH. resume puts the hardware into a
2002 * functional state after a suspend and updates the software state as
2003 * necessary. This function is also used for restoring the GPU after a GPU
2005 * Returns 0 on success, negative error code on failure.
2007 static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)
2011 for (i = 0; i < adev->num_ip_blocks; i++) {
2012 if (!adev->ip_blocks[i].status.valid)
2014 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2015 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2016 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH)
2018 r = adev->ip_blocks[i].version->funcs->resume(adev);
2020 DRM_ERROR("resume of IP block <%s> failed %d\n",
2021 adev->ip_blocks[i].version->funcs->name, r);
2030 * amdgpu_device_ip_resume - run resume for hardware IPs
2032 * @adev: amdgpu_device pointer
2034 * Main resume function for hardware IPs. The hardware IPs
2035 * are split into two resume functions because they are
2036 * are also used in in recovering from a GPU reset and some additional
2037 * steps need to be take between them. In this case (S3/S4) they are
2039 * Returns 0 on success, negative error code on failure.
2041 static int amdgpu_device_ip_resume(struct amdgpu_device *adev)
2045 r = amdgpu_device_ip_resume_phase1(adev);
2048 r = amdgpu_device_ip_resume_phase2(adev);
2054 * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV
2056 * @adev: amdgpu_device pointer
2058 * Query the VBIOS data tables to determine if the board supports SR-IOV.
2060 static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
2062 if (amdgpu_sriov_vf(adev)) {
2063 if (adev->is_atom_fw) {
2064 if (amdgpu_atomfirmware_gpu_supports_virtualization(adev))
2065 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2067 if (amdgpu_atombios_has_gpu_virtualization_table(adev))
2068 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2071 if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
2072 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
2077 * amdgpu_device_asic_has_dc_support - determine if DC supports the asic
2079 * @asic_type: AMD asic type
2081 * Check if there is DC (new modesetting infrastructre) support for an asic.
2082 * returns true if DC has support, false if not.
2084 bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
2086 switch (asic_type) {
2087 #if defined(CONFIG_DRM_AMD_DC)
2095 case CHIP_POLARIS11:
2096 case CHIP_POLARIS10:
2097 case CHIP_POLARIS12:
2102 #if defined(CONFIG_DRM_AMD_DC_DCN1_0)
2105 return amdgpu_dc != 0;
2113 * amdgpu_device_has_dc_support - check if dc is supported
2115 * @adev: amdgpu_device_pointer
2117 * Returns true for supported, false for not supported
2119 bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
2121 if (amdgpu_sriov_vf(adev))
2124 return amdgpu_device_asic_has_dc_support(adev->asic_type);
2128 * amdgpu_device_init - initialize the driver
2130 * @adev: amdgpu_device pointer
2131 * @pdev: drm dev pointer
2132 * @pdev: pci dev pointer
2133 * @flags: driver flags
2135 * Initializes the driver info and hw (all asics).
2136 * Returns 0 for success or an error on failure.
2137 * Called at driver startup.
2139 int amdgpu_device_init(struct amdgpu_device *adev,
2140 struct drm_device *ddev,
2141 struct pci_dev *pdev,
2145 bool runtime = false;
2148 adev->shutdown = false;
2149 adev->dev = &pdev->dev;
2152 adev->flags = flags;
2153 adev->asic_type = flags & AMD_ASIC_MASK;
2154 adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
2155 if (amdgpu_emu_mode == 1)
2156 adev->usec_timeout *= 2;
2157 adev->gmc.gart_size = 512 * 1024 * 1024;
2158 adev->accel_working = false;
2159 adev->num_rings = 0;
2160 adev->mman.buffer_funcs = NULL;
2161 adev->mman.buffer_funcs_ring = NULL;
2162 adev->vm_manager.vm_pte_funcs = NULL;
2163 adev->vm_manager.vm_pte_num_rings = 0;
2164 adev->gmc.gmc_funcs = NULL;
2165 adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
2166 bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
2168 adev->smc_rreg = &amdgpu_invalid_rreg;
2169 adev->smc_wreg = &amdgpu_invalid_wreg;
2170 adev->pcie_rreg = &amdgpu_invalid_rreg;
2171 adev->pcie_wreg = &amdgpu_invalid_wreg;
2172 adev->pciep_rreg = &amdgpu_invalid_rreg;
2173 adev->pciep_wreg = &amdgpu_invalid_wreg;
2174 adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
2175 adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
2176 adev->didt_rreg = &amdgpu_invalid_rreg;
2177 adev->didt_wreg = &amdgpu_invalid_wreg;
2178 adev->gc_cac_rreg = &amdgpu_invalid_rreg;
2179 adev->gc_cac_wreg = &amdgpu_invalid_wreg;
2180 adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
2181 adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
2183 DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
2184 amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
2185 pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
2187 /* mutex initialization are all done here so we
2188 * can recall function without having locking issues */
2189 atomic_set(&adev->irq.ih.lock, 0);
2190 mutex_init(&adev->firmware.mutex);
2191 mutex_init(&adev->pm.mutex);
2192 mutex_init(&adev->gfx.gpu_clock_mutex);
2193 mutex_init(&adev->srbm_mutex);
2194 mutex_init(&adev->gfx.pipe_reserve_mutex);
2195 mutex_init(&adev->grbm_idx_mutex);
2196 mutex_init(&adev->mn_lock);
2197 mutex_init(&adev->virt.vf_errors.lock);
2198 hash_init(adev->mn_hash);
2199 mutex_init(&adev->lock_reset);
2201 amdgpu_device_check_arguments(adev);
2203 spin_lock_init(&adev->mmio_idx_lock);
2204 spin_lock_init(&adev->smc_idx_lock);
2205 spin_lock_init(&adev->pcie_idx_lock);
2206 spin_lock_init(&adev->uvd_ctx_idx_lock);
2207 spin_lock_init(&adev->didt_idx_lock);
2208 spin_lock_init(&adev->gc_cac_idx_lock);
2209 spin_lock_init(&adev->se_cac_idx_lock);
2210 spin_lock_init(&adev->audio_endpt_idx_lock);
2211 spin_lock_init(&adev->mm_stats.lock);
2213 INIT_LIST_HEAD(&adev->shadow_list);
2214 mutex_init(&adev->shadow_list_lock);
2216 INIT_LIST_HEAD(&adev->ring_lru_list);
2217 spin_lock_init(&adev->ring_lru_list_lock);
2219 INIT_DELAYED_WORK(&adev->late_init_work,
2220 amdgpu_device_ip_late_init_func_handler);
2222 /* Registers mapping */
2223 /* TODO: block userspace mapping of io register */
2224 if (adev->asic_type >= CHIP_BONAIRE) {
2225 adev->rmmio_base = pci_resource_start(adev->pdev, 5);
2226 adev->rmmio_size = pci_resource_len(adev->pdev, 5);
2228 adev->rmmio_base = pci_resource_start(adev->pdev, 2);
2229 adev->rmmio_size = pci_resource_len(adev->pdev, 2);
2232 adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
2233 if (adev->rmmio == NULL) {
2236 DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
2237 DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
2239 /* doorbell bar mapping */
2240 amdgpu_device_doorbell_init(adev);
2242 /* io port mapping */
2243 for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
2244 if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) {
2245 adev->rio_mem_size = pci_resource_len(adev->pdev, i);
2246 adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size);
2250 if (adev->rio_mem == NULL)
2251 DRM_INFO("PCI I/O BAR is not found.\n");
2253 amdgpu_device_get_pcie_info(adev);
2255 /* early init functions */
2256 r = amdgpu_device_ip_early_init(adev);
2260 /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
2261 /* this will fail for cards that aren't VGA class devices, just
2263 vga_client_register(adev->pdev, adev, NULL, amdgpu_device_vga_set_decode);
2265 if (amdgpu_device_is_px(ddev))
2267 if (!pci_is_thunderbolt_attached(adev->pdev))
2268 vga_switcheroo_register_client(adev->pdev,
2269 &amdgpu_switcheroo_ops, runtime);
2271 vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
2273 if (amdgpu_emu_mode == 1) {
2274 /* post the asic on emulation mode */
2275 emu_soc_asic_init(adev);
2276 goto fence_driver_init;
2280 if (!amdgpu_get_bios(adev)) {
2285 r = amdgpu_atombios_init(adev);
2287 dev_err(adev->dev, "amdgpu_atombios_init failed\n");
2288 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0);
2292 /* detect if we are with an SRIOV vbios */
2293 amdgpu_device_detect_sriov_bios(adev);
2295 /* Post card if necessary */
2296 if (amdgpu_device_need_post(adev)) {
2298 dev_err(adev->dev, "no vBIOS found\n");
2302 DRM_INFO("GPU posting now...\n");
2303 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2305 dev_err(adev->dev, "gpu post error!\n");
2310 if (adev->is_atom_fw) {
2311 /* Initialize clocks */
2312 r = amdgpu_atomfirmware_get_clock_info(adev);
2314 dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
2315 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
2319 /* Initialize clocks */
2320 r = amdgpu_atombios_get_clock_info(adev);
2322 dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
2323 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
2326 /* init i2c buses */
2327 if (!amdgpu_device_has_dc_support(adev))
2328 amdgpu_atombios_i2c_init(adev);
2333 r = amdgpu_fence_driver_init(adev);
2335 dev_err(adev->dev, "amdgpu_fence_driver_init failed\n");
2336 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
2340 /* init the mode config */
2341 drm_mode_config_init(adev->ddev);
2343 r = amdgpu_device_ip_init(adev);
2345 /* failed in exclusive mode due to timeout */
2346 if (amdgpu_sriov_vf(adev) &&
2347 !amdgpu_sriov_runtime(adev) &&
2348 amdgpu_virt_mmio_blocked(adev) &&
2349 !amdgpu_virt_wait_reset(adev)) {
2350 dev_err(adev->dev, "VF exclusive mode timeout\n");
2351 /* Don't send request since VF is inactive. */
2352 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
2353 adev->virt.ops = NULL;
2357 dev_err(adev->dev, "amdgpu_device_ip_init failed\n");
2358 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
2362 adev->accel_working = true;
2364 amdgpu_vm_check_compute_bug(adev);
2366 /* Initialize the buffer migration limit. */
2367 if (amdgpu_moverate >= 0)
2368 max_MBps = amdgpu_moverate;
2370 max_MBps = 8; /* Allow 8 MB/s. */
2371 /* Get a log2 for easy divisions. */
2372 adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
2374 r = amdgpu_ib_pool_init(adev);
2376 dev_err(adev->dev, "IB initialization failed (%d).\n", r);
2377 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r);
2381 if (amdgpu_sriov_vf(adev))
2382 amdgpu_virt_init_data_exchange(adev);
2384 amdgpu_fbdev_init(adev);
2386 r = amdgpu_pm_sysfs_init(adev);
2388 DRM_ERROR("registering pm debugfs failed (%d).\n", r);
2390 r = amdgpu_debugfs_gem_init(adev);
2392 DRM_ERROR("registering gem debugfs failed (%d).\n", r);
2394 r = amdgpu_debugfs_regs_init(adev);
2396 DRM_ERROR("registering register debugfs failed (%d).\n", r);
2398 r = amdgpu_debugfs_firmware_init(adev);
2400 DRM_ERROR("registering firmware debugfs failed (%d).\n", r);
2402 r = amdgpu_debugfs_init(adev);
2404 DRM_ERROR("Creating debugfs files failed (%d).\n", r);
2406 if ((amdgpu_testing & 1)) {
2407 if (adev->accel_working)
2408 amdgpu_test_moves(adev);
2410 DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n");
2412 if (amdgpu_benchmarking) {
2413 if (adev->accel_working)
2414 amdgpu_benchmark(adev, amdgpu_benchmarking);
2416 DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n");
2419 /* enable clockgating, etc. after ib tests, etc. since some blocks require
2420 * explicit gating rather than handling it automatically.
2422 r = amdgpu_device_ip_late_init(adev);
2424 dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n");
2425 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
2432 amdgpu_vf_error_trans_all(adev);
2434 vga_switcheroo_fini_domain_pm_ops(adev->dev);
2440 * amdgpu_device_fini - tear down the driver
2442 * @adev: amdgpu_device pointer
2444 * Tear down the driver info (all asics).
2445 * Called at driver shutdown.
2447 void amdgpu_device_fini(struct amdgpu_device *adev)
2451 DRM_INFO("amdgpu: finishing device.\n");
2452 adev->shutdown = true;
2453 /* disable all interrupts */
2454 amdgpu_irq_disable_all(adev);
2455 if (adev->mode_info.mode_config_initialized){
2456 if (!amdgpu_device_has_dc_support(adev))
2457 drm_crtc_force_disable_all(adev->ddev);
2459 drm_atomic_helper_shutdown(adev->ddev);
2461 amdgpu_ib_pool_fini(adev);
2462 amdgpu_fence_driver_fini(adev);
2463 amdgpu_pm_sysfs_fini(adev);
2464 amdgpu_fbdev_fini(adev);
2465 r = amdgpu_device_ip_fini(adev);
2466 if (adev->firmware.gpu_info_fw) {
2467 release_firmware(adev->firmware.gpu_info_fw);
2468 adev->firmware.gpu_info_fw = NULL;
2470 adev->accel_working = false;
2471 cancel_delayed_work_sync(&adev->late_init_work);
2472 /* free i2c buses */
2473 if (!amdgpu_device_has_dc_support(adev))
2474 amdgpu_i2c_fini(adev);
2476 if (amdgpu_emu_mode != 1)
2477 amdgpu_atombios_fini(adev);
2481 if (!pci_is_thunderbolt_attached(adev->pdev))
2482 vga_switcheroo_unregister_client(adev->pdev);
2483 if (adev->flags & AMD_IS_PX)
2484 vga_switcheroo_fini_domain_pm_ops(adev->dev);
2485 vga_client_register(adev->pdev, NULL, NULL, NULL);
2487 pci_iounmap(adev->pdev, adev->rio_mem);
2488 adev->rio_mem = NULL;
2489 iounmap(adev->rmmio);
2491 amdgpu_device_doorbell_fini(adev);
2492 amdgpu_debugfs_regs_cleanup(adev);
2500 * amdgpu_device_suspend - initiate device suspend
2502 * @pdev: drm dev pointer
2503 * @state: suspend state
2505 * Puts the hw in the suspend state (all asics).
2506 * Returns 0 for success or an error on failure.
2507 * Called at driver suspend.
2509 int amdgpu_device_suspend(struct drm_device *dev, bool suspend, bool fbcon)
2511 struct amdgpu_device *adev;
2512 struct drm_crtc *crtc;
2513 struct drm_connector *connector;
2516 if (dev == NULL || dev->dev_private == NULL) {
2520 adev = dev->dev_private;
2522 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
2525 drm_kms_helper_poll_disable(dev);
2527 if (!amdgpu_device_has_dc_support(adev)) {
2528 /* turn off display hw */
2529 drm_modeset_lock_all(dev);
2530 list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2531 drm_helper_connector_dpms(connector, DRM_MODE_DPMS_OFF);
2533 drm_modeset_unlock_all(dev);
2536 amdgpu_amdkfd_suspend(adev);
2538 /* unpin the front buffers and cursors */
2539 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2540 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2541 struct drm_framebuffer *fb = crtc->primary->fb;
2542 struct amdgpu_bo *robj;
2544 if (amdgpu_crtc->cursor_bo) {
2545 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2546 r = amdgpu_bo_reserve(aobj, true);
2548 amdgpu_bo_unpin(aobj);
2549 amdgpu_bo_unreserve(aobj);
2553 if (fb == NULL || fb->obj[0] == NULL) {
2556 robj = gem_to_amdgpu_bo(fb->obj[0]);
2557 /* don't unpin kernel fb objects */
2558 if (!amdgpu_fbdev_robj_is_fb(adev, robj)) {
2559 r = amdgpu_bo_reserve(robj, true);
2561 amdgpu_bo_unpin(robj);
2562 amdgpu_bo_unreserve(robj);
2566 /* evict vram memory */
2567 amdgpu_bo_evict_vram(adev);
2569 amdgpu_fence_driver_suspend(adev);
2571 r = amdgpu_device_ip_suspend(adev);
2573 /* evict remaining vram memory
2574 * This second call to evict vram is to evict the gart page table
2577 amdgpu_bo_evict_vram(adev);
2579 pci_save_state(dev->pdev);
2581 /* Shut down the device */
2582 pci_disable_device(dev->pdev);
2583 pci_set_power_state(dev->pdev, PCI_D3hot);
2585 r = amdgpu_asic_reset(adev);
2587 DRM_ERROR("amdgpu asic reset failed\n");
2592 amdgpu_fbdev_set_suspend(adev, 1);
2599 * amdgpu_device_resume - initiate device resume
2601 * @pdev: drm dev pointer
2603 * Bring the hw back to operating state (all asics).
2604 * Returns 0 for success or an error on failure.
2605 * Called at driver resume.
2607 int amdgpu_device_resume(struct drm_device *dev, bool resume, bool fbcon)
2609 struct drm_connector *connector;
2610 struct amdgpu_device *adev = dev->dev_private;
2611 struct drm_crtc *crtc;
2614 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
2621 pci_set_power_state(dev->pdev, PCI_D0);
2622 pci_restore_state(dev->pdev);
2623 r = pci_enable_device(dev->pdev);
2629 if (amdgpu_device_need_post(adev)) {
2630 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2632 DRM_ERROR("amdgpu asic init failed\n");
2635 r = amdgpu_device_ip_resume(adev);
2637 DRM_ERROR("amdgpu_device_ip_resume failed (%d).\n", r);
2640 amdgpu_fence_driver_resume(adev);
2643 r = amdgpu_device_ip_late_init(adev);
2648 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2649 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2651 if (amdgpu_crtc->cursor_bo) {
2652 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2653 r = amdgpu_bo_reserve(aobj, true);
2655 r = amdgpu_bo_pin(aobj,
2656 AMDGPU_GEM_DOMAIN_VRAM,
2657 &amdgpu_crtc->cursor_addr);
2659 DRM_ERROR("Failed to pin cursor BO (%d)\n", r);
2660 amdgpu_bo_unreserve(aobj);
2664 r = amdgpu_amdkfd_resume(adev);
2668 /* blat the mode back in */
2670 if (!amdgpu_device_has_dc_support(adev)) {
2672 drm_helper_resume_force_mode(dev);
2674 /* turn on display hw */
2675 drm_modeset_lock_all(dev);
2676 list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2677 drm_helper_connector_dpms(connector, DRM_MODE_DPMS_ON);
2679 drm_modeset_unlock_all(dev);
2683 drm_kms_helper_poll_enable(dev);
2686 * Most of the connector probing functions try to acquire runtime pm
2687 * refs to ensure that the GPU is powered on when connector polling is
2688 * performed. Since we're calling this from a runtime PM callback,
2689 * trying to acquire rpm refs will cause us to deadlock.
2691 * Since we're guaranteed to be holding the rpm lock, it's safe to
2692 * temporarily disable the rpm helpers so this doesn't deadlock us.
2695 dev->dev->power.disable_depth++;
2697 if (!amdgpu_device_has_dc_support(adev))
2698 drm_helper_hpd_irq_event(dev);
2700 drm_kms_helper_hotplug_event(dev);
2702 dev->dev->power.disable_depth--;
2706 amdgpu_fbdev_set_suspend(adev, 0);
2716 * amdgpu_device_ip_check_soft_reset - did soft reset succeed
2718 * @adev: amdgpu_device pointer
2720 * The list of all the hardware IPs that make up the asic is walked and
2721 * the check_soft_reset callbacks are run. check_soft_reset determines
2722 * if the asic is still hung or not.
2723 * Returns true if any of the IPs are still in a hung state, false if not.
2725 static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)
2728 bool asic_hang = false;
2730 if (amdgpu_sriov_vf(adev))
2733 if (amdgpu_asic_need_full_reset(adev))
2736 for (i = 0; i < adev->num_ip_blocks; i++) {
2737 if (!adev->ip_blocks[i].status.valid)
2739 if (adev->ip_blocks[i].version->funcs->check_soft_reset)
2740 adev->ip_blocks[i].status.hang =
2741 adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
2742 if (adev->ip_blocks[i].status.hang) {
2743 DRM_INFO("IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
2751 * amdgpu_device_ip_pre_soft_reset - prepare for soft reset
2753 * @adev: amdgpu_device pointer
2755 * The list of all the hardware IPs that make up the asic is walked and the
2756 * pre_soft_reset callbacks are run if the block is hung. pre_soft_reset
2757 * handles any IP specific hardware or software state changes that are
2758 * necessary for a soft reset to succeed.
2759 * Returns 0 on success, negative error code on failure.
2761 static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)
2765 for (i = 0; i < adev->num_ip_blocks; i++) {
2766 if (!adev->ip_blocks[i].status.valid)
2768 if (adev->ip_blocks[i].status.hang &&
2769 adev->ip_blocks[i].version->funcs->pre_soft_reset) {
2770 r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
2780 * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed
2782 * @adev: amdgpu_device pointer
2784 * Some hardware IPs cannot be soft reset. If they are hung, a full gpu
2785 * reset is necessary to recover.
2786 * Returns true if a full asic reset is required, false if not.
2788 static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)
2792 if (amdgpu_asic_need_full_reset(adev))
2795 for (i = 0; i < adev->num_ip_blocks; i++) {
2796 if (!adev->ip_blocks[i].status.valid)
2798 if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
2799 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
2800 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
2801 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
2802 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
2803 if (adev->ip_blocks[i].status.hang) {
2804 DRM_INFO("Some block need full reset!\n");
2813 * amdgpu_device_ip_soft_reset - do a soft reset
2815 * @adev: amdgpu_device pointer
2817 * The list of all the hardware IPs that make up the asic is walked and the
2818 * soft_reset callbacks are run if the block is hung. soft_reset handles any
2819 * IP specific hardware or software state changes that are necessary to soft
2821 * Returns 0 on success, negative error code on failure.
2823 static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)
2827 for (i = 0; i < adev->num_ip_blocks; i++) {
2828 if (!adev->ip_blocks[i].status.valid)
2830 if (adev->ip_blocks[i].status.hang &&
2831 adev->ip_blocks[i].version->funcs->soft_reset) {
2832 r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
2842 * amdgpu_device_ip_post_soft_reset - clean up from soft reset
2844 * @adev: amdgpu_device pointer
2846 * The list of all the hardware IPs that make up the asic is walked and the
2847 * post_soft_reset callbacks are run if the asic was hung. post_soft_reset
2848 * handles any IP specific hardware or software state changes that are
2849 * necessary after the IP has been soft reset.
2850 * Returns 0 on success, negative error code on failure.
2852 static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)
2856 for (i = 0; i < adev->num_ip_blocks; i++) {
2857 if (!adev->ip_blocks[i].status.valid)
2859 if (adev->ip_blocks[i].status.hang &&
2860 adev->ip_blocks[i].version->funcs->post_soft_reset)
2861 r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
2870 * amdgpu_device_recover_vram_from_shadow - restore shadowed VRAM buffers
2872 * @adev: amdgpu_device pointer
2873 * @ring: amdgpu_ring for the engine handling the buffer operations
2874 * @bo: amdgpu_bo buffer whose shadow is being restored
2875 * @fence: dma_fence associated with the operation
2877 * Restores the VRAM buffer contents from the shadow in GTT. Used to
2878 * restore things like GPUVM page tables after a GPU reset where
2879 * the contents of VRAM might be lost.
2880 * Returns 0 on success, negative error code on failure.
2882 static int amdgpu_device_recover_vram_from_shadow(struct amdgpu_device *adev,
2883 struct amdgpu_ring *ring,
2884 struct amdgpu_bo *bo,
2885 struct dma_fence **fence)
2893 r = amdgpu_bo_reserve(bo, true);
2896 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
2897 /* if bo has been evicted, then no need to recover */
2898 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
2899 r = amdgpu_bo_validate(bo->shadow);
2901 DRM_ERROR("bo validate failed!\n");
2905 r = amdgpu_bo_restore_from_shadow(adev, ring, bo,
2908 DRM_ERROR("recover page table failed!\n");
2913 amdgpu_bo_unreserve(bo);
2918 * amdgpu_device_handle_vram_lost - Handle the loss of VRAM contents
2920 * @adev: amdgpu_device pointer
2922 * Restores the contents of VRAM buffers from the shadows in GTT. Used to
2923 * restore things like GPUVM page tables after a GPU reset where
2924 * the contents of VRAM might be lost.
2925 * Returns 0 on success, 1 on failure.
2927 static int amdgpu_device_handle_vram_lost(struct amdgpu_device *adev)
2929 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
2930 struct amdgpu_bo *bo, *tmp;
2931 struct dma_fence *fence = NULL, *next = NULL;
2936 if (amdgpu_sriov_runtime(adev))
2937 tmo = msecs_to_jiffies(amdgpu_lockup_timeout);
2939 tmo = msecs_to_jiffies(100);
2941 DRM_INFO("recover vram bo from shadow start\n");
2942 mutex_lock(&adev->shadow_list_lock);
2943 list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) {
2945 amdgpu_device_recover_vram_from_shadow(adev, ring, bo, &next);
2947 r = dma_fence_wait_timeout(fence, false, tmo);
2949 pr_err("wait fence %p[%d] timeout\n", fence, i);
2951 pr_err("wait fence %p[%d] interrupted\n", fence, i);
2953 dma_fence_put(fence);
2960 dma_fence_put(fence);
2963 mutex_unlock(&adev->shadow_list_lock);
2966 r = dma_fence_wait_timeout(fence, false, tmo);
2968 pr_err("wait fence %p[%d] timeout\n", fence, i);
2970 pr_err("wait fence %p[%d] interrupted\n", fence, i);
2973 dma_fence_put(fence);
2976 DRM_INFO("recover vram bo from shadow done\n");
2978 DRM_ERROR("recover vram bo from shadow failed\n");
2980 return (r > 0) ? 0 : 1;
2984 * amdgpu_device_reset - reset ASIC/GPU for bare-metal or passthrough
2986 * @adev: amdgpu device pointer
2988 * attempt to do soft-reset or full-reset and reinitialize Asic
2989 * return 0 means successed otherwise failed
2991 static int amdgpu_device_reset(struct amdgpu_device *adev)
2993 bool need_full_reset, vram_lost = 0;
2996 need_full_reset = amdgpu_device_ip_need_full_reset(adev);
2998 if (!need_full_reset) {
2999 amdgpu_device_ip_pre_soft_reset(adev);
3000 r = amdgpu_device_ip_soft_reset(adev);
3001 amdgpu_device_ip_post_soft_reset(adev);
3002 if (r || amdgpu_device_ip_check_soft_reset(adev)) {
3003 DRM_INFO("soft reset failed, will fallback to full reset!\n");
3004 need_full_reset = true;
3008 if (need_full_reset) {
3009 r = amdgpu_device_ip_suspend(adev);
3012 r = amdgpu_asic_reset(adev);
3014 amdgpu_atom_asic_init(adev->mode_info.atom_context);
3017 dev_info(adev->dev, "GPU reset succeeded, trying to resume\n");
3018 r = amdgpu_device_ip_resume_phase1(adev);
3022 vram_lost = amdgpu_device_check_vram_lost(adev);
3024 DRM_ERROR("VRAM is lost!\n");
3025 atomic_inc(&adev->vram_lost_counter);
3028 r = amdgpu_gtt_mgr_recover(
3029 &adev->mman.bdev.man[TTM_PL_TT]);
3033 r = amdgpu_device_ip_resume_phase2(adev);
3038 amdgpu_device_fill_reset_magic(adev);
3044 amdgpu_irq_gpu_reset_resume_helper(adev);
3045 r = amdgpu_ib_ring_tests(adev);
3047 dev_err(adev->dev, "ib ring test failed (%d).\n", r);
3048 r = amdgpu_device_ip_suspend(adev);
3049 need_full_reset = true;
3054 if (!r && ((need_full_reset && !(adev->flags & AMD_IS_APU)) || vram_lost))
3055 r = amdgpu_device_handle_vram_lost(adev);
3061 * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
3063 * @adev: amdgpu device pointer
3065 * do VF FLR and reinitialize Asic
3066 * return 0 means successed otherwise failed
3068 static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
3069 bool from_hypervisor)
3073 if (from_hypervisor)
3074 r = amdgpu_virt_request_full_gpu(adev, true);
3076 r = amdgpu_virt_reset_gpu(adev);
3080 /* Resume IP prior to SMC */
3081 r = amdgpu_device_ip_reinit_early_sriov(adev);
3085 /* we need recover gart prior to run SMC/CP/SDMA resume */
3086 amdgpu_gtt_mgr_recover(&adev->mman.bdev.man[TTM_PL_TT]);
3088 /* now we are okay to resume SMC/CP/SDMA */
3089 r = amdgpu_device_ip_reinit_late_sriov(adev);
3090 amdgpu_virt_release_full_gpu(adev, true);
3094 amdgpu_irq_gpu_reset_resume_helper(adev);
3095 r = amdgpu_ib_ring_tests(adev);
3097 if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
3098 atomic_inc(&adev->vram_lost_counter);
3099 r = amdgpu_device_handle_vram_lost(adev);
3108 * amdgpu_device_gpu_recover - reset the asic and recover scheduler
3110 * @adev: amdgpu device pointer
3111 * @job: which job trigger hang
3112 * @force forces reset regardless of amdgpu_gpu_recovery
3114 * Attempt to reset the GPU if it has hung (all asics).
3115 * Returns 0 for success or an error on failure.
3117 int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
3118 struct amdgpu_job *job, bool force)
3120 struct drm_atomic_state *state = NULL;
3123 if (!force && !amdgpu_device_ip_check_soft_reset(adev)) {
3124 DRM_INFO("No hardware hang detected. Did some blocks stall?\n");
3128 if (!force && (amdgpu_gpu_recovery == 0 ||
3129 (amdgpu_gpu_recovery == -1 && !amdgpu_sriov_vf(adev)))) {
3130 DRM_INFO("GPU recovery disabled.\n");
3134 dev_info(adev->dev, "GPU reset begin!\n");
3136 mutex_lock(&adev->lock_reset);
3137 atomic_inc(&adev->gpu_reset_counter);
3138 adev->in_gpu_reset = 1;
3141 resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
3143 /* store modesetting */
3144 if (amdgpu_device_has_dc_support(adev))
3145 state = drm_atomic_helper_suspend(adev->ddev);
3147 /* block all schedulers and reset given job's ring */
3148 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3149 struct amdgpu_ring *ring = adev->rings[i];
3151 if (!ring || !ring->sched.thread)
3154 kthread_park(ring->sched.thread);
3156 if (job && job->ring->idx != i)
3159 drm_sched_hw_job_reset(&ring->sched, &job->base);
3161 /* after all hw jobs are reset, hw fence is meaningless, so force_completion */
3162 amdgpu_fence_driver_force_completion(ring);
3165 if (amdgpu_sriov_vf(adev))
3166 r = amdgpu_device_reset_sriov(adev, job ? false : true);
3168 r = amdgpu_device_reset(adev);
3170 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3171 struct amdgpu_ring *ring = adev->rings[i];
3173 if (!ring || !ring->sched.thread)
3176 /* only need recovery sched of the given job's ring
3177 * or all rings (in the case @job is NULL)
3178 * after above amdgpu_reset accomplished
3180 if ((!job || job->ring->idx == i) && !r)
3181 drm_sched_job_recovery(&ring->sched);
3183 kthread_unpark(ring->sched.thread);
3186 if (amdgpu_device_has_dc_support(adev)) {
3187 if (drm_atomic_helper_resume(adev->ddev, state))
3188 dev_info(adev->dev, "drm resume failed:%d\n", r);
3190 drm_helper_resume_force_mode(adev->ddev);
3193 ttm_bo_unlock_delayed_workqueue(&adev->mman.bdev, resched);
3196 /* bad news, how to tell it to userspace ? */
3197 dev_info(adev->dev, "GPU reset(%d) failed\n", atomic_read(&adev->gpu_reset_counter));
3198 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r);
3200 dev_info(adev->dev, "GPU reset(%d) successed!\n",atomic_read(&adev->gpu_reset_counter));
3203 amdgpu_vf_error_trans_all(adev);
3204 adev->in_gpu_reset = 0;
3205 mutex_unlock(&adev->lock_reset);
3210 * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot
3212 * @adev: amdgpu_device pointer
3214 * Fetchs and stores in the driver the PCIE capabilities (gen speed
3215 * and lanes) of the slot the device is in. Handles APUs and
3216 * virtualized environments where PCIE config space may not be available.
3218 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev)
3223 if (amdgpu_pcie_gen_cap)
3224 adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
3226 if (amdgpu_pcie_lane_cap)
3227 adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
3229 /* covers APUs as well */
3230 if (pci_is_root_bus(adev->pdev->bus)) {
3231 if (adev->pm.pcie_gen_mask == 0)
3232 adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
3233 if (adev->pm.pcie_mlw_mask == 0)
3234 adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
3238 if (adev->pm.pcie_gen_mask == 0) {
3239 ret = drm_pcie_get_speed_cap_mask(adev->ddev, &mask);
3241 adev->pm.pcie_gen_mask = (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3242 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3243 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
3245 if (mask & DRM_PCIE_SPEED_25)
3246 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
3247 if (mask & DRM_PCIE_SPEED_50)
3248 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2;
3249 if (mask & DRM_PCIE_SPEED_80)
3250 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3;
3252 adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
3255 if (adev->pm.pcie_mlw_mask == 0) {
3256 ret = drm_pcie_get_max_link_width(adev->ddev, &mask);
3260 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
3261 CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
3262 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3263 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3264 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3265 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3266 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3269 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
3270 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3271 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3272 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3273 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3274 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3277 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3278 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3279 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3280 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3281 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3284 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3285 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3286 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3287 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3290 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3291 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3292 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3295 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3296 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3299 adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;
3305 adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;