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>
31 #include <linux/debugfs.h>
33 #include <drm/drm_crtc_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"
44 #ifdef CONFIG_DRM_AMDGPU_SI
47 #ifdef CONFIG_DRM_AMDGPU_CIK
51 #include "bif/bif_4_1_d.h"
52 #include <linux/pci.h>
53 #include <linux/firmware.h>
55 static int amdgpu_debugfs_regs_init(struct amdgpu_device *adev);
56 static void amdgpu_debugfs_regs_cleanup(struct amdgpu_device *adev);
58 static const char *amdgpu_asic_name[] = {
79 bool amdgpu_device_is_px(struct drm_device *dev)
81 struct amdgpu_device *adev = dev->dev_private;
83 if (adev->flags & AMD_IS_PX)
89 * MMIO register access helper functions.
91 uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg,
96 if ((reg * 4) < adev->rmmio_size && !always_indirect)
97 ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
101 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
102 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
103 ret = readl(((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
104 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
106 trace_amdgpu_mm_rreg(adev->pdev->device, reg, ret);
110 void amdgpu_mm_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v,
111 bool always_indirect)
113 trace_amdgpu_mm_wreg(adev->pdev->device, reg, v);
115 if ((reg * 4) < adev->rmmio_size && !always_indirect)
116 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
120 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
121 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
122 writel(v, ((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
123 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
127 u32 amdgpu_io_rreg(struct amdgpu_device *adev, u32 reg)
129 if ((reg * 4) < adev->rio_mem_size)
130 return ioread32(adev->rio_mem + (reg * 4));
132 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
133 return ioread32(adev->rio_mem + (mmMM_DATA * 4));
137 void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32 v)
140 if ((reg * 4) < adev->rio_mem_size)
141 iowrite32(v, adev->rio_mem + (reg * 4));
143 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
144 iowrite32(v, adev->rio_mem + (mmMM_DATA * 4));
149 * amdgpu_mm_rdoorbell - read a doorbell dword
151 * @adev: amdgpu_device pointer
152 * @index: doorbell index
154 * Returns the value in the doorbell aperture at the
155 * requested doorbell index (CIK).
157 u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
159 if (index < adev->doorbell.num_doorbells) {
160 return readl(adev->doorbell.ptr + index);
162 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
168 * amdgpu_mm_wdoorbell - write a doorbell dword
170 * @adev: amdgpu_device pointer
171 * @index: doorbell index
174 * Writes @v to the doorbell aperture at the
175 * requested doorbell index (CIK).
177 void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v)
179 if (index < adev->doorbell.num_doorbells) {
180 writel(v, adev->doorbell.ptr + index);
182 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
187 * amdgpu_invalid_rreg - dummy reg read function
189 * @adev: amdgpu device pointer
190 * @reg: offset of register
192 * Dummy register read function. Used for register blocks
193 * that certain asics don't have (all asics).
194 * Returns the value in the register.
196 static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
198 DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
204 * amdgpu_invalid_wreg - dummy reg write function
206 * @adev: amdgpu device pointer
207 * @reg: offset of register
208 * @v: value to write to the register
210 * Dummy register read function. Used for register blocks
211 * that certain asics don't have (all asics).
213 static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
215 DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
221 * amdgpu_block_invalid_rreg - dummy reg read function
223 * @adev: amdgpu device pointer
224 * @block: offset of instance
225 * @reg: offset of register
227 * Dummy register read function. Used for register blocks
228 * that certain asics don't have (all asics).
229 * Returns the value in the register.
231 static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
232 uint32_t block, uint32_t reg)
234 DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
241 * amdgpu_block_invalid_wreg - dummy reg write function
243 * @adev: amdgpu device pointer
244 * @block: offset of instance
245 * @reg: offset of register
246 * @v: value to write to the register
248 * Dummy register read function. Used for register blocks
249 * that certain asics don't have (all asics).
251 static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
253 uint32_t reg, uint32_t v)
255 DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
260 static int amdgpu_vram_scratch_init(struct amdgpu_device *adev)
264 if (adev->vram_scratch.robj == NULL) {
265 r = amdgpu_bo_create(adev, AMDGPU_GPU_PAGE_SIZE,
266 PAGE_SIZE, true, AMDGPU_GEM_DOMAIN_VRAM,
267 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
268 AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS,
269 NULL, NULL, &adev->vram_scratch.robj);
275 r = amdgpu_bo_reserve(adev->vram_scratch.robj, false);
276 if (unlikely(r != 0))
278 r = amdgpu_bo_pin(adev->vram_scratch.robj,
279 AMDGPU_GEM_DOMAIN_VRAM, &adev->vram_scratch.gpu_addr);
281 amdgpu_bo_unreserve(adev->vram_scratch.robj);
284 r = amdgpu_bo_kmap(adev->vram_scratch.robj,
285 (void **)&adev->vram_scratch.ptr);
287 amdgpu_bo_unpin(adev->vram_scratch.robj);
288 amdgpu_bo_unreserve(adev->vram_scratch.robj);
293 static void amdgpu_vram_scratch_fini(struct amdgpu_device *adev)
297 if (adev->vram_scratch.robj == NULL) {
300 r = amdgpu_bo_reserve(adev->vram_scratch.robj, false);
301 if (likely(r == 0)) {
302 amdgpu_bo_kunmap(adev->vram_scratch.robj);
303 amdgpu_bo_unpin(adev->vram_scratch.robj);
304 amdgpu_bo_unreserve(adev->vram_scratch.robj);
306 amdgpu_bo_unref(&adev->vram_scratch.robj);
310 * amdgpu_program_register_sequence - program an array of registers.
312 * @adev: amdgpu_device pointer
313 * @registers: pointer to the register array
314 * @array_size: size of the register array
316 * Programs an array or registers with and and or masks.
317 * This is a helper for setting golden registers.
319 void amdgpu_program_register_sequence(struct amdgpu_device *adev,
320 const u32 *registers,
321 const u32 array_size)
323 u32 tmp, reg, and_mask, or_mask;
329 for (i = 0; i < array_size; i +=3) {
330 reg = registers[i + 0];
331 and_mask = registers[i + 1];
332 or_mask = registers[i + 2];
334 if (and_mask == 0xffffffff) {
345 void amdgpu_pci_config_reset(struct amdgpu_device *adev)
347 pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
351 * GPU doorbell aperture helpers function.
354 * amdgpu_doorbell_init - Init doorbell driver information.
356 * @adev: amdgpu_device pointer
358 * Init doorbell driver information (CIK)
359 * Returns 0 on success, error on failure.
361 static int amdgpu_doorbell_init(struct amdgpu_device *adev)
363 /* doorbell bar mapping */
364 adev->doorbell.base = pci_resource_start(adev->pdev, 2);
365 adev->doorbell.size = pci_resource_len(adev->pdev, 2);
367 adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32),
368 AMDGPU_DOORBELL_MAX_ASSIGNMENT+1);
369 if (adev->doorbell.num_doorbells == 0)
372 adev->doorbell.ptr = ioremap(adev->doorbell.base, adev->doorbell.num_doorbells * sizeof(u32));
373 if (adev->doorbell.ptr == NULL) {
376 DRM_INFO("doorbell mmio base: 0x%08X\n", (uint32_t)adev->doorbell.base);
377 DRM_INFO("doorbell mmio size: %u\n", (unsigned)adev->doorbell.size);
383 * amdgpu_doorbell_fini - Tear down doorbell driver information.
385 * @adev: amdgpu_device pointer
387 * Tear down doorbell driver information (CIK)
389 static void amdgpu_doorbell_fini(struct amdgpu_device *adev)
391 iounmap(adev->doorbell.ptr);
392 adev->doorbell.ptr = NULL;
396 * amdgpu_doorbell_get_kfd_info - Report doorbell configuration required to
399 * @adev: amdgpu_device pointer
400 * @aperture_base: output returning doorbell aperture base physical address
401 * @aperture_size: output returning doorbell aperture size in bytes
402 * @start_offset: output returning # of doorbell bytes reserved for amdgpu.
404 * amdgpu and amdkfd share the doorbell aperture. amdgpu sets it up,
405 * takes doorbells required for its own rings and reports the setup to amdkfd.
406 * amdgpu reserved doorbells are at the start of the doorbell aperture.
408 void amdgpu_doorbell_get_kfd_info(struct amdgpu_device *adev,
409 phys_addr_t *aperture_base,
410 size_t *aperture_size,
411 size_t *start_offset)
414 * The first num_doorbells are used by amdgpu.
415 * amdkfd takes whatever's left in the aperture.
417 if (adev->doorbell.size > adev->doorbell.num_doorbells * sizeof(u32)) {
418 *aperture_base = adev->doorbell.base;
419 *aperture_size = adev->doorbell.size;
420 *start_offset = adev->doorbell.num_doorbells * sizeof(u32);
430 * Writeback is the the method by which the the GPU updates special pages
431 * in memory with the status of certain GPU events (fences, ring pointers,
436 * amdgpu_wb_fini - Disable Writeback and free memory
438 * @adev: amdgpu_device pointer
440 * Disables Writeback and frees the Writeback memory (all asics).
441 * Used at driver shutdown.
443 static void amdgpu_wb_fini(struct amdgpu_device *adev)
445 if (adev->wb.wb_obj) {
446 amdgpu_bo_free_kernel(&adev->wb.wb_obj,
448 (void **)&adev->wb.wb);
449 adev->wb.wb_obj = NULL;
454 * amdgpu_wb_init- Init Writeback driver info and allocate memory
456 * @adev: amdgpu_device pointer
458 * Disables Writeback and frees the Writeback memory (all asics).
459 * Used at driver startup.
460 * Returns 0 on success or an -error on failure.
462 static int amdgpu_wb_init(struct amdgpu_device *adev)
466 if (adev->wb.wb_obj == NULL) {
467 r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * 4,
468 PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
469 &adev->wb.wb_obj, &adev->wb.gpu_addr,
470 (void **)&adev->wb.wb);
472 dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
476 adev->wb.num_wb = AMDGPU_MAX_WB;
477 memset(&adev->wb.used, 0, sizeof(adev->wb.used));
479 /* clear wb memory */
480 memset((char *)adev->wb.wb, 0, AMDGPU_GPU_PAGE_SIZE);
487 * amdgpu_wb_get - Allocate a wb entry
489 * @adev: amdgpu_device pointer
492 * Allocate a wb slot for use by the driver (all asics).
493 * Returns 0 on success or -EINVAL on failure.
495 int amdgpu_wb_get(struct amdgpu_device *adev, u32 *wb)
497 unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
498 if (offset < adev->wb.num_wb) {
499 __set_bit(offset, adev->wb.used);
508 * amdgpu_wb_free - Free a wb entry
510 * @adev: amdgpu_device pointer
513 * Free a wb slot allocated for use by the driver (all asics)
515 void amdgpu_wb_free(struct amdgpu_device *adev, u32 wb)
517 if (wb < adev->wb.num_wb)
518 __clear_bit(wb, adev->wb.used);
522 * amdgpu_vram_location - try to find VRAM location
523 * @adev: amdgpu device structure holding all necessary informations
524 * @mc: memory controller structure holding memory informations
525 * @base: base address at which to put VRAM
527 * Function will place try to place VRAM at base address provided
528 * as parameter (which is so far either PCI aperture address or
529 * for IGP TOM base address).
531 * If there is not enough space to fit the unvisible VRAM in the 32bits
532 * address space then we limit the VRAM size to the aperture.
534 * Note: We don't explicitly enforce VRAM start to be aligned on VRAM size,
535 * this shouldn't be a problem as we are using the PCI aperture as a reference.
536 * Otherwise this would be needed for rv280, all r3xx, and all r4xx, but
539 * Note: we use mc_vram_size as on some board we need to program the mc to
540 * cover the whole aperture even if VRAM size is inferior to aperture size
541 * Novell bug 204882 + along with lots of ubuntu ones
543 * Note: when limiting vram it's safe to overwritte real_vram_size because
544 * we are not in case where real_vram_size is inferior to mc_vram_size (ie
545 * note afected by bogus hw of Novell bug 204882 + along with lots of ubuntu
548 * Note: IGP TOM addr should be the same as the aperture addr, we don't
549 * explicitly check for that thought.
551 * FIXME: when reducing VRAM size align new size on power of 2.
553 void amdgpu_vram_location(struct amdgpu_device *adev, struct amdgpu_mc *mc, u64 base)
555 uint64_t limit = (uint64_t)amdgpu_vram_limit << 20;
557 mc->vram_start = base;
558 if (mc->mc_vram_size > (adev->mc.mc_mask - base + 1)) {
559 dev_warn(adev->dev, "limiting VRAM to PCI aperture size\n");
560 mc->real_vram_size = mc->aper_size;
561 mc->mc_vram_size = mc->aper_size;
563 mc->vram_end = mc->vram_start + mc->mc_vram_size - 1;
564 if (limit && limit < mc->real_vram_size)
565 mc->real_vram_size = limit;
566 dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
567 mc->mc_vram_size >> 20, mc->vram_start,
568 mc->vram_end, mc->real_vram_size >> 20);
572 * amdgpu_gtt_location - try to find GTT location
573 * @adev: amdgpu device structure holding all necessary informations
574 * @mc: memory controller structure holding memory informations
576 * Function will place try to place GTT before or after VRAM.
578 * If GTT size is bigger than space left then we ajust GTT size.
579 * Thus function will never fails.
581 * FIXME: when reducing GTT size align new size on power of 2.
583 void amdgpu_gtt_location(struct amdgpu_device *adev, struct amdgpu_mc *mc)
585 u64 size_af, size_bf;
587 size_af = ((adev->mc.mc_mask - mc->vram_end) + mc->gtt_base_align) & ~mc->gtt_base_align;
588 size_bf = mc->vram_start & ~mc->gtt_base_align;
589 if (size_bf > size_af) {
590 if (mc->gtt_size > size_bf) {
591 dev_warn(adev->dev, "limiting GTT\n");
592 mc->gtt_size = size_bf;
594 mc->gtt_start = (mc->vram_start & ~mc->gtt_base_align) - mc->gtt_size;
596 if (mc->gtt_size > size_af) {
597 dev_warn(adev->dev, "limiting GTT\n");
598 mc->gtt_size = size_af;
600 mc->gtt_start = (mc->vram_end + 1 + mc->gtt_base_align) & ~mc->gtt_base_align;
602 mc->gtt_end = mc->gtt_start + mc->gtt_size - 1;
603 dev_info(adev->dev, "GTT: %lluM 0x%016llX - 0x%016llX\n",
604 mc->gtt_size >> 20, mc->gtt_start, mc->gtt_end);
608 * GPU helpers function.
611 * amdgpu_card_posted - check if the hw has already been initialized
613 * @adev: amdgpu_device pointer
615 * Check if the asic has been initialized (all asics).
616 * Used at driver startup.
617 * Returns true if initialized or false if not.
619 bool amdgpu_card_posted(struct amdgpu_device *adev)
623 /* then check MEM_SIZE, in case the crtcs are off */
624 reg = RREG32(mmCONFIG_MEMSIZE);
633 static bool amdgpu_vpost_needed(struct amdgpu_device *adev)
635 if (amdgpu_sriov_vf(adev))
638 if (amdgpu_passthrough(adev)) {
639 /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
640 * some old smc fw still need driver do vPost otherwise gpu hang, while
641 * those smc fw version above 22.15 doesn't have this flaw, so we force
642 * vpost executed for smc version below 22.15
644 if (adev->asic_type == CHIP_FIJI) {
647 err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
648 /* force vPost if error occured */
652 fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
653 if (fw_ver < 0x00160e00)
657 return !amdgpu_card_posted(adev);
661 * amdgpu_dummy_page_init - init dummy page used by the driver
663 * @adev: amdgpu_device pointer
665 * Allocate the dummy page used by the driver (all asics).
666 * This dummy page is used by the driver as a filler for gart entries
667 * when pages are taken out of the GART
668 * Returns 0 on sucess, -ENOMEM on failure.
670 int amdgpu_dummy_page_init(struct amdgpu_device *adev)
672 if (adev->dummy_page.page)
674 adev->dummy_page.page = alloc_page(GFP_DMA32 | GFP_KERNEL | __GFP_ZERO);
675 if (adev->dummy_page.page == NULL)
677 adev->dummy_page.addr = pci_map_page(adev->pdev, adev->dummy_page.page,
678 0, PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
679 if (pci_dma_mapping_error(adev->pdev, adev->dummy_page.addr)) {
680 dev_err(&adev->pdev->dev, "Failed to DMA MAP the dummy page\n");
681 __free_page(adev->dummy_page.page);
682 adev->dummy_page.page = NULL;
689 * amdgpu_dummy_page_fini - free dummy page used by the driver
691 * @adev: amdgpu_device pointer
693 * Frees the dummy page used by the driver (all asics).
695 void amdgpu_dummy_page_fini(struct amdgpu_device *adev)
697 if (adev->dummy_page.page == NULL)
699 pci_unmap_page(adev->pdev, adev->dummy_page.addr,
700 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
701 __free_page(adev->dummy_page.page);
702 adev->dummy_page.page = NULL;
706 /* ATOM accessor methods */
708 * ATOM is an interpreted byte code stored in tables in the vbios. The
709 * driver registers callbacks to access registers and the interpreter
710 * in the driver parses the tables and executes then to program specific
711 * actions (set display modes, asic init, etc.). See amdgpu_atombios.c,
712 * atombios.h, and atom.c
716 * cail_pll_read - read PLL register
718 * @info: atom card_info pointer
719 * @reg: PLL register offset
721 * Provides a PLL register accessor for the atom interpreter (r4xx+).
722 * Returns the value of the PLL register.
724 static uint32_t cail_pll_read(struct card_info *info, uint32_t reg)
730 * cail_pll_write - write PLL register
732 * @info: atom card_info pointer
733 * @reg: PLL register offset
734 * @val: value to write to the pll register
736 * Provides a PLL register accessor for the atom interpreter (r4xx+).
738 static void cail_pll_write(struct card_info *info, uint32_t reg, uint32_t val)
744 * cail_mc_read - read MC (Memory Controller) register
746 * @info: atom card_info pointer
747 * @reg: MC register offset
749 * Provides an MC register accessor for the atom interpreter (r4xx+).
750 * Returns the value of the MC register.
752 static uint32_t cail_mc_read(struct card_info *info, uint32_t reg)
758 * cail_mc_write - write MC (Memory Controller) register
760 * @info: atom card_info pointer
761 * @reg: MC register offset
762 * @val: value to write to the pll register
764 * Provides a MC register accessor for the atom interpreter (r4xx+).
766 static void cail_mc_write(struct card_info *info, uint32_t reg, uint32_t val)
772 * cail_reg_write - write MMIO register
774 * @info: atom card_info pointer
775 * @reg: MMIO register offset
776 * @val: value to write to the pll register
778 * Provides a MMIO register accessor for the atom interpreter (r4xx+).
780 static void cail_reg_write(struct card_info *info, uint32_t reg, uint32_t val)
782 struct amdgpu_device *adev = info->dev->dev_private;
788 * cail_reg_read - read MMIO register
790 * @info: atom card_info pointer
791 * @reg: MMIO register offset
793 * Provides an MMIO register accessor for the atom interpreter (r4xx+).
794 * Returns the value of the MMIO register.
796 static uint32_t cail_reg_read(struct card_info *info, uint32_t reg)
798 struct amdgpu_device *adev = info->dev->dev_private;
806 * cail_ioreg_write - write IO register
808 * @info: atom card_info pointer
809 * @reg: IO register offset
810 * @val: value to write to the pll register
812 * Provides a IO register accessor for the atom interpreter (r4xx+).
814 static void cail_ioreg_write(struct card_info *info, uint32_t reg, uint32_t val)
816 struct amdgpu_device *adev = info->dev->dev_private;
822 * cail_ioreg_read - read IO register
824 * @info: atom card_info pointer
825 * @reg: IO register offset
827 * Provides an IO register accessor for the atom interpreter (r4xx+).
828 * Returns the value of the IO register.
830 static uint32_t cail_ioreg_read(struct card_info *info, uint32_t reg)
832 struct amdgpu_device *adev = info->dev->dev_private;
840 * amdgpu_atombios_fini - free the driver info and callbacks for atombios
842 * @adev: amdgpu_device pointer
844 * Frees the driver info and register access callbacks for the ATOM
845 * interpreter (r4xx+).
846 * Called at driver shutdown.
848 static void amdgpu_atombios_fini(struct amdgpu_device *adev)
850 if (adev->mode_info.atom_context) {
851 kfree(adev->mode_info.atom_context->scratch);
852 kfree(adev->mode_info.atom_context->iio);
854 kfree(adev->mode_info.atom_context);
855 adev->mode_info.atom_context = NULL;
856 kfree(adev->mode_info.atom_card_info);
857 adev->mode_info.atom_card_info = NULL;
861 * amdgpu_atombios_init - init the driver info and callbacks for atombios
863 * @adev: amdgpu_device pointer
865 * Initializes the driver info and register access callbacks for the
866 * ATOM interpreter (r4xx+).
867 * Returns 0 on sucess, -ENOMEM on failure.
868 * Called at driver startup.
870 static int amdgpu_atombios_init(struct amdgpu_device *adev)
872 struct card_info *atom_card_info =
873 kzalloc(sizeof(struct card_info), GFP_KERNEL);
878 adev->mode_info.atom_card_info = atom_card_info;
879 atom_card_info->dev = adev->ddev;
880 atom_card_info->reg_read = cail_reg_read;
881 atom_card_info->reg_write = cail_reg_write;
882 /* needed for iio ops */
884 atom_card_info->ioreg_read = cail_ioreg_read;
885 atom_card_info->ioreg_write = cail_ioreg_write;
887 DRM_ERROR("Unable to find PCI I/O BAR; using MMIO for ATOM IIO\n");
888 atom_card_info->ioreg_read = cail_reg_read;
889 atom_card_info->ioreg_write = cail_reg_write;
891 atom_card_info->mc_read = cail_mc_read;
892 atom_card_info->mc_write = cail_mc_write;
893 atom_card_info->pll_read = cail_pll_read;
894 atom_card_info->pll_write = cail_pll_write;
896 adev->mode_info.atom_context = amdgpu_atom_parse(atom_card_info, adev->bios);
897 if (!adev->mode_info.atom_context) {
898 amdgpu_atombios_fini(adev);
902 mutex_init(&adev->mode_info.atom_context->mutex);
903 amdgpu_atombios_scratch_regs_init(adev);
904 amdgpu_atom_allocate_fb_scratch(adev->mode_info.atom_context);
908 /* if we get transitioned to only one device, take VGA back */
910 * amdgpu_vga_set_decode - enable/disable vga decode
912 * @cookie: amdgpu_device pointer
913 * @state: enable/disable vga decode
915 * Enable/disable vga decode (all asics).
916 * Returns VGA resource flags.
918 static unsigned int amdgpu_vga_set_decode(void *cookie, bool state)
920 struct amdgpu_device *adev = cookie;
921 amdgpu_asic_set_vga_state(adev, state);
923 return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
924 VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
926 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
930 * amdgpu_check_pot_argument - check that argument is a power of two
932 * @arg: value to check
934 * Validates that a certain argument is a power of two (all asics).
935 * Returns true if argument is valid.
937 static bool amdgpu_check_pot_argument(int arg)
939 return (arg & (arg - 1)) == 0;
943 * amdgpu_check_arguments - validate module params
945 * @adev: amdgpu_device pointer
947 * Validates certain module parameters and updates
948 * the associated values used by the driver (all asics).
950 static void amdgpu_check_arguments(struct amdgpu_device *adev)
952 if (amdgpu_sched_jobs < 4) {
953 dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
955 amdgpu_sched_jobs = 4;
956 } else if (!amdgpu_check_pot_argument(amdgpu_sched_jobs)){
957 dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
959 amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
962 if (amdgpu_gart_size != -1) {
963 /* gtt size must be greater or equal to 32M */
964 if (amdgpu_gart_size < 32) {
965 dev_warn(adev->dev, "gart size (%d) too small\n",
967 amdgpu_gart_size = -1;
971 if (!amdgpu_check_pot_argument(amdgpu_vm_size)) {
972 dev_warn(adev->dev, "VM size (%d) must be a power of 2\n",
977 if (amdgpu_vm_size < 1) {
978 dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
984 * Max GPUVM size for Cayman, SI and CI are 40 bits.
986 if (amdgpu_vm_size > 1024) {
987 dev_warn(adev->dev, "VM size (%d) too large, max is 1TB\n",
992 /* defines number of bits in page table versus page directory,
993 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
994 * page table and the remaining bits are in the page directory */
995 if (amdgpu_vm_block_size == -1) {
997 /* Total bits covered by PD + PTs */
998 unsigned bits = ilog2(amdgpu_vm_size) + 18;
1000 /* Make sure the PD is 4K in size up to 8GB address space.
1001 Above that split equal between PD and PTs */
1002 if (amdgpu_vm_size <= 8)
1003 amdgpu_vm_block_size = bits - 9;
1005 amdgpu_vm_block_size = (bits + 3) / 2;
1007 } else if (amdgpu_vm_block_size < 9) {
1008 dev_warn(adev->dev, "VM page table size (%d) too small\n",
1009 amdgpu_vm_block_size);
1010 amdgpu_vm_block_size = 9;
1013 if (amdgpu_vm_block_size > 24 ||
1014 (amdgpu_vm_size * 1024) < (1ull << amdgpu_vm_block_size)) {
1015 dev_warn(adev->dev, "VM page table size (%d) too large\n",
1016 amdgpu_vm_block_size);
1017 amdgpu_vm_block_size = 9;
1020 if (amdgpu_vram_page_split != -1 && (amdgpu_vram_page_split < 16 ||
1021 !amdgpu_check_pot_argument(amdgpu_vram_page_split))) {
1022 dev_warn(adev->dev, "invalid VRAM page split (%d)\n",
1023 amdgpu_vram_page_split);
1024 amdgpu_vram_page_split = 1024;
1029 * amdgpu_switcheroo_set_state - set switcheroo state
1031 * @pdev: pci dev pointer
1032 * @state: vga_switcheroo state
1034 * Callback for the switcheroo driver. Suspends or resumes the
1035 * the asics before or after it is powered up using ACPI methods.
1037 static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, enum vga_switcheroo_state state)
1039 struct drm_device *dev = pci_get_drvdata(pdev);
1041 if (amdgpu_device_is_px(dev) && state == VGA_SWITCHEROO_OFF)
1044 if (state == VGA_SWITCHEROO_ON) {
1045 unsigned d3_delay = dev->pdev->d3_delay;
1047 printk(KERN_INFO "amdgpu: switched on\n");
1048 /* don't suspend or resume card normally */
1049 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1051 amdgpu_device_resume(dev, true, true);
1053 dev->pdev->d3_delay = d3_delay;
1055 dev->switch_power_state = DRM_SWITCH_POWER_ON;
1056 drm_kms_helper_poll_enable(dev);
1058 printk(KERN_INFO "amdgpu: switched off\n");
1059 drm_kms_helper_poll_disable(dev);
1060 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1061 amdgpu_device_suspend(dev, true, true);
1062 dev->switch_power_state = DRM_SWITCH_POWER_OFF;
1067 * amdgpu_switcheroo_can_switch - see if switcheroo state can change
1069 * @pdev: pci dev pointer
1071 * Callback for the switcheroo driver. Check of the switcheroo
1072 * state can be changed.
1073 * Returns true if the state can be changed, false if not.
1075 static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
1077 struct drm_device *dev = pci_get_drvdata(pdev);
1080 * FIXME: open_count is protected by drm_global_mutex but that would lead to
1081 * locking inversion with the driver load path. And the access here is
1082 * completely racy anyway. So don't bother with locking for now.
1084 return dev->open_count == 0;
1087 static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
1088 .set_gpu_state = amdgpu_switcheroo_set_state,
1090 .can_switch = amdgpu_switcheroo_can_switch,
1093 int amdgpu_set_clockgating_state(struct amdgpu_device *adev,
1094 enum amd_ip_block_type block_type,
1095 enum amd_clockgating_state state)
1099 for (i = 0; i < adev->num_ip_blocks; i++) {
1100 if (!adev->ip_blocks[i].status.valid)
1102 if (adev->ip_blocks[i].version->type == block_type) {
1103 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1113 int amdgpu_set_powergating_state(struct amdgpu_device *adev,
1114 enum amd_ip_block_type block_type,
1115 enum amd_powergating_state state)
1119 for (i = 0; i < adev->num_ip_blocks; i++) {
1120 if (!adev->ip_blocks[i].status.valid)
1122 if (adev->ip_blocks[i].version->type == block_type) {
1123 r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev,
1133 int amdgpu_wait_for_idle(struct amdgpu_device *adev,
1134 enum amd_ip_block_type block_type)
1138 for (i = 0; i < adev->num_ip_blocks; i++) {
1139 if (!adev->ip_blocks[i].status.valid)
1141 if (adev->ip_blocks[i].version->type == block_type) {
1142 r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
1152 bool amdgpu_is_idle(struct amdgpu_device *adev,
1153 enum amd_ip_block_type block_type)
1157 for (i = 0; i < adev->num_ip_blocks; i++) {
1158 if (!adev->ip_blocks[i].status.valid)
1160 if (adev->ip_blocks[i].version->type == block_type)
1161 return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
1167 struct amdgpu_ip_block * amdgpu_get_ip_block(struct amdgpu_device *adev,
1168 enum amd_ip_block_type type)
1172 for (i = 0; i < adev->num_ip_blocks; i++)
1173 if (adev->ip_blocks[i].version->type == type)
1174 return &adev->ip_blocks[i];
1180 * amdgpu_ip_block_version_cmp
1182 * @adev: amdgpu_device pointer
1183 * @type: enum amd_ip_block_type
1184 * @major: major version
1185 * @minor: minor version
1187 * return 0 if equal or greater
1188 * return 1 if smaller or the ip_block doesn't exist
1190 int amdgpu_ip_block_version_cmp(struct amdgpu_device *adev,
1191 enum amd_ip_block_type type,
1192 u32 major, u32 minor)
1194 struct amdgpu_ip_block *ip_block = amdgpu_get_ip_block(adev, type);
1196 if (ip_block && ((ip_block->version->major > major) ||
1197 ((ip_block->version->major == major) &&
1198 (ip_block->version->minor >= minor))))
1205 * amdgpu_ip_block_add
1207 * @adev: amdgpu_device pointer
1208 * @ip_block_version: pointer to the IP to add
1210 * Adds the IP block driver information to the collection of IPs
1213 int amdgpu_ip_block_add(struct amdgpu_device *adev,
1214 const struct amdgpu_ip_block_version *ip_block_version)
1216 if (!ip_block_version)
1219 adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
1224 static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
1226 adev->enable_virtual_display = false;
1228 if (amdgpu_virtual_display) {
1229 struct drm_device *ddev = adev->ddev;
1230 const char *pci_address_name = pci_name(ddev->pdev);
1231 char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
1233 pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
1234 pciaddstr_tmp = pciaddstr;
1235 while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
1236 pciaddname = strsep(&pciaddname_tmp, ",");
1237 if (!strcmp(pci_address_name, pciaddname)) {
1241 adev->enable_virtual_display = true;
1244 res = kstrtol(pciaddname_tmp, 10,
1252 adev->mode_info.num_crtc = num_crtc;
1254 adev->mode_info.num_crtc = 1;
1260 DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
1261 amdgpu_virtual_display, pci_address_name,
1262 adev->enable_virtual_display, adev->mode_info.num_crtc);
1268 static int amdgpu_early_init(struct amdgpu_device *adev)
1272 amdgpu_device_enable_virtual_display(adev);
1274 switch (adev->asic_type) {
1278 case CHIP_POLARIS11:
1279 case CHIP_POLARIS10:
1282 if (adev->asic_type == CHIP_CARRIZO || adev->asic_type == CHIP_STONEY)
1283 adev->family = AMDGPU_FAMILY_CZ;
1285 adev->family = AMDGPU_FAMILY_VI;
1287 r = vi_set_ip_blocks(adev);
1291 #ifdef CONFIG_DRM_AMDGPU_SI
1297 adev->family = AMDGPU_FAMILY_SI;
1298 r = si_set_ip_blocks(adev);
1303 #ifdef CONFIG_DRM_AMDGPU_CIK
1309 if ((adev->asic_type == CHIP_BONAIRE) || (adev->asic_type == CHIP_HAWAII))
1310 adev->family = AMDGPU_FAMILY_CI;
1312 adev->family = AMDGPU_FAMILY_KV;
1314 r = cik_set_ip_blocks(adev);
1320 /* FIXME: not supported yet */
1324 for (i = 0; i < adev->num_ip_blocks; i++) {
1325 if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
1326 DRM_ERROR("disabled ip block: %d\n", i);
1327 adev->ip_blocks[i].status.valid = false;
1329 if (adev->ip_blocks[i].version->funcs->early_init) {
1330 r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
1332 adev->ip_blocks[i].status.valid = false;
1334 DRM_ERROR("early_init of IP block <%s> failed %d\n",
1335 adev->ip_blocks[i].version->funcs->name, r);
1338 adev->ip_blocks[i].status.valid = true;
1341 adev->ip_blocks[i].status.valid = true;
1346 adev->cg_flags &= amdgpu_cg_mask;
1347 adev->pg_flags &= amdgpu_pg_mask;
1352 static int amdgpu_init(struct amdgpu_device *adev)
1356 for (i = 0; i < adev->num_ip_blocks; i++) {
1357 if (!adev->ip_blocks[i].status.valid)
1359 r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
1361 DRM_ERROR("sw_init of IP block <%s> failed %d\n",
1362 adev->ip_blocks[i].version->funcs->name, r);
1365 adev->ip_blocks[i].status.sw = true;
1366 /* need to do gmc hw init early so we can allocate gpu mem */
1367 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1368 r = amdgpu_vram_scratch_init(adev);
1370 DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r);
1373 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1375 DRM_ERROR("hw_init %d failed %d\n", i, r);
1378 r = amdgpu_wb_init(adev);
1380 DRM_ERROR("amdgpu_wb_init failed %d\n", r);
1383 adev->ip_blocks[i].status.hw = true;
1387 for (i = 0; i < adev->num_ip_blocks; i++) {
1388 if (!adev->ip_blocks[i].status.sw)
1390 /* gmc hw init is done early */
1391 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC)
1393 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1395 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1396 adev->ip_blocks[i].version->funcs->name, r);
1399 adev->ip_blocks[i].status.hw = true;
1405 static int amdgpu_late_init(struct amdgpu_device *adev)
1409 for (i = 0; i < adev->num_ip_blocks; i++) {
1410 if (!adev->ip_blocks[i].status.valid)
1412 if (adev->ip_blocks[i].version->funcs->late_init) {
1413 r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
1415 DRM_ERROR("late_init of IP block <%s> failed %d\n",
1416 adev->ip_blocks[i].version->funcs->name, r);
1419 adev->ip_blocks[i].status.late_initialized = true;
1421 /* skip CG for VCE/UVD, it's handled specially */
1422 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1423 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE) {
1424 /* enable clockgating to save power */
1425 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1428 DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
1429 adev->ip_blocks[i].version->funcs->name, r);
1438 static int amdgpu_fini(struct amdgpu_device *adev)
1442 /* need to disable SMC first */
1443 for (i = 0; i < adev->num_ip_blocks; i++) {
1444 if (!adev->ip_blocks[i].status.hw)
1446 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
1447 /* ungate blocks before hw fini so that we can shutdown the blocks safely */
1448 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1449 AMD_CG_STATE_UNGATE);
1451 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1452 adev->ip_blocks[i].version->funcs->name, r);
1455 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
1456 /* XXX handle errors */
1458 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
1459 adev->ip_blocks[i].version->funcs->name, r);
1461 adev->ip_blocks[i].status.hw = false;
1466 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1467 if (!adev->ip_blocks[i].status.hw)
1469 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1470 amdgpu_wb_fini(adev);
1471 amdgpu_vram_scratch_fini(adev);
1474 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1475 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE) {
1476 /* ungate blocks before hw fini so that we can shutdown the blocks safely */
1477 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1478 AMD_CG_STATE_UNGATE);
1480 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1481 adev->ip_blocks[i].version->funcs->name, r);
1486 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
1487 /* XXX handle errors */
1489 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
1490 adev->ip_blocks[i].version->funcs->name, r);
1493 adev->ip_blocks[i].status.hw = false;
1496 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1497 if (!adev->ip_blocks[i].status.sw)
1499 r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
1500 /* XXX handle errors */
1502 DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
1503 adev->ip_blocks[i].version->funcs->name, r);
1505 adev->ip_blocks[i].status.sw = false;
1506 adev->ip_blocks[i].status.valid = false;
1509 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1510 if (!adev->ip_blocks[i].status.late_initialized)
1512 if (adev->ip_blocks[i].version->funcs->late_fini)
1513 adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
1514 adev->ip_blocks[i].status.late_initialized = false;
1520 int amdgpu_suspend(struct amdgpu_device *adev)
1524 /* ungate SMC block first */
1525 r = amdgpu_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_SMC,
1526 AMD_CG_STATE_UNGATE);
1528 DRM_ERROR("set_clockgating_state(ungate) SMC failed %d\n",r);
1531 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1532 if (!adev->ip_blocks[i].status.valid)
1534 /* ungate blocks so that suspend can properly shut them down */
1535 if (i != AMD_IP_BLOCK_TYPE_SMC) {
1536 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1537 AMD_CG_STATE_UNGATE);
1539 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1540 adev->ip_blocks[i].version->funcs->name, r);
1543 /* XXX handle errors */
1544 r = adev->ip_blocks[i].version->funcs->suspend(adev);
1545 /* XXX handle errors */
1547 DRM_ERROR("suspend of IP block <%s> failed %d\n",
1548 adev->ip_blocks[i].version->funcs->name, r);
1555 static int amdgpu_resume(struct amdgpu_device *adev)
1559 for (i = 0; i < adev->num_ip_blocks; i++) {
1560 if (!adev->ip_blocks[i].status.valid)
1562 r = adev->ip_blocks[i].version->funcs->resume(adev);
1564 DRM_ERROR("resume of IP block <%s> failed %d\n",
1565 adev->ip_blocks[i].version->funcs->name, r);
1573 static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
1575 if (amdgpu_atombios_has_gpu_virtualization_table(adev))
1576 adev->virtualization.virtual_caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
1580 * amdgpu_device_init - initialize the driver
1582 * @adev: amdgpu_device pointer
1583 * @pdev: drm dev pointer
1584 * @pdev: pci dev pointer
1585 * @flags: driver flags
1587 * Initializes the driver info and hw (all asics).
1588 * Returns 0 for success or an error on failure.
1589 * Called at driver startup.
1591 int amdgpu_device_init(struct amdgpu_device *adev,
1592 struct drm_device *ddev,
1593 struct pci_dev *pdev,
1597 bool runtime = false;
1600 adev->shutdown = false;
1601 adev->dev = &pdev->dev;
1604 adev->flags = flags;
1605 adev->asic_type = flags & AMD_ASIC_MASK;
1606 adev->is_atom_bios = false;
1607 adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
1608 adev->mc.gtt_size = 512 * 1024 * 1024;
1609 adev->accel_working = false;
1610 adev->num_rings = 0;
1611 adev->mman.buffer_funcs = NULL;
1612 adev->mman.buffer_funcs_ring = NULL;
1613 adev->vm_manager.vm_pte_funcs = NULL;
1614 adev->vm_manager.vm_pte_num_rings = 0;
1615 adev->gart.gart_funcs = NULL;
1616 adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
1618 adev->smc_rreg = &amdgpu_invalid_rreg;
1619 adev->smc_wreg = &amdgpu_invalid_wreg;
1620 adev->pcie_rreg = &amdgpu_invalid_rreg;
1621 adev->pcie_wreg = &amdgpu_invalid_wreg;
1622 adev->pciep_rreg = &amdgpu_invalid_rreg;
1623 adev->pciep_wreg = &amdgpu_invalid_wreg;
1624 adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
1625 adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
1626 adev->didt_rreg = &amdgpu_invalid_rreg;
1627 adev->didt_wreg = &amdgpu_invalid_wreg;
1628 adev->gc_cac_rreg = &amdgpu_invalid_rreg;
1629 adev->gc_cac_wreg = &amdgpu_invalid_wreg;
1630 adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
1631 adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
1634 DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
1635 amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
1636 pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
1638 /* mutex initialization are all done here so we
1639 * can recall function without having locking issues */
1640 mutex_init(&adev->vm_manager.lock);
1641 atomic_set(&adev->irq.ih.lock, 0);
1642 mutex_init(&adev->pm.mutex);
1643 mutex_init(&adev->gfx.gpu_clock_mutex);
1644 mutex_init(&adev->srbm_mutex);
1645 mutex_init(&adev->grbm_idx_mutex);
1646 mutex_init(&adev->mn_lock);
1647 hash_init(adev->mn_hash);
1649 amdgpu_check_arguments(adev);
1651 /* Registers mapping */
1652 /* TODO: block userspace mapping of io register */
1653 spin_lock_init(&adev->mmio_idx_lock);
1654 spin_lock_init(&adev->smc_idx_lock);
1655 spin_lock_init(&adev->pcie_idx_lock);
1656 spin_lock_init(&adev->uvd_ctx_idx_lock);
1657 spin_lock_init(&adev->didt_idx_lock);
1658 spin_lock_init(&adev->gc_cac_idx_lock);
1659 spin_lock_init(&adev->audio_endpt_idx_lock);
1660 spin_lock_init(&adev->mm_stats.lock);
1662 INIT_LIST_HEAD(&adev->shadow_list);
1663 mutex_init(&adev->shadow_list_lock);
1665 INIT_LIST_HEAD(&adev->gtt_list);
1666 spin_lock_init(&adev->gtt_list_lock);
1668 if (adev->asic_type >= CHIP_BONAIRE) {
1669 adev->rmmio_base = pci_resource_start(adev->pdev, 5);
1670 adev->rmmio_size = pci_resource_len(adev->pdev, 5);
1672 adev->rmmio_base = pci_resource_start(adev->pdev, 2);
1673 adev->rmmio_size = pci_resource_len(adev->pdev, 2);
1676 adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
1677 if (adev->rmmio == NULL) {
1680 DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
1681 DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
1683 if (adev->asic_type >= CHIP_BONAIRE)
1684 /* doorbell bar mapping */
1685 amdgpu_doorbell_init(adev);
1687 /* io port mapping */
1688 for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
1689 if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) {
1690 adev->rio_mem_size = pci_resource_len(adev->pdev, i);
1691 adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size);
1695 if (adev->rio_mem == NULL)
1696 DRM_ERROR("Unable to find PCI I/O BAR\n");
1698 /* early init functions */
1699 r = amdgpu_early_init(adev);
1703 /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
1704 /* this will fail for cards that aren't VGA class devices, just
1706 vga_client_register(adev->pdev, adev, NULL, amdgpu_vga_set_decode);
1708 if (amdgpu_runtime_pm == 1)
1710 if (amdgpu_device_is_px(ddev))
1712 vga_switcheroo_register_client(adev->pdev, &amdgpu_switcheroo_ops, runtime);
1714 vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
1717 if (!amdgpu_get_bios(adev)) {
1721 /* Must be an ATOMBIOS */
1722 if (!adev->is_atom_bios) {
1723 dev_err(adev->dev, "Expecting atombios for GPU\n");
1727 r = amdgpu_atombios_init(adev);
1729 dev_err(adev->dev, "amdgpu_atombios_init failed\n");
1733 /* detect if we are with an SRIOV vbios */
1734 amdgpu_device_detect_sriov_bios(adev);
1736 /* Post card if necessary */
1737 if (amdgpu_vpost_needed(adev)) {
1739 dev_err(adev->dev, "no vBIOS found\n");
1743 DRM_INFO("GPU posting now...\n");
1744 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
1746 dev_err(adev->dev, "gpu post error!\n");
1750 DRM_INFO("GPU post is not needed\n");
1753 /* Initialize clocks */
1754 r = amdgpu_atombios_get_clock_info(adev);
1756 dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
1759 /* init i2c buses */
1760 amdgpu_atombios_i2c_init(adev);
1763 r = amdgpu_fence_driver_init(adev);
1765 dev_err(adev->dev, "amdgpu_fence_driver_init failed\n");
1769 /* init the mode config */
1770 drm_mode_config_init(adev->ddev);
1772 r = amdgpu_init(adev);
1774 dev_err(adev->dev, "amdgpu_init failed\n");
1779 adev->accel_working = true;
1781 /* Initialize the buffer migration limit. */
1782 if (amdgpu_moverate >= 0)
1783 max_MBps = amdgpu_moverate;
1785 max_MBps = 8; /* Allow 8 MB/s. */
1786 /* Get a log2 for easy divisions. */
1787 adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
1789 amdgpu_fbdev_init(adev);
1791 r = amdgpu_ib_pool_init(adev);
1793 dev_err(adev->dev, "IB initialization failed (%d).\n", r);
1797 r = amdgpu_ib_ring_tests(adev);
1799 DRM_ERROR("ib ring test failed (%d).\n", r);
1801 r = amdgpu_gem_debugfs_init(adev);
1803 DRM_ERROR("registering gem debugfs failed (%d).\n", r);
1806 r = amdgpu_debugfs_regs_init(adev);
1808 DRM_ERROR("registering register debugfs failed (%d).\n", r);
1811 r = amdgpu_debugfs_firmware_init(adev);
1813 DRM_ERROR("registering firmware debugfs failed (%d).\n", r);
1817 if ((amdgpu_testing & 1)) {
1818 if (adev->accel_working)
1819 amdgpu_test_moves(adev);
1821 DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n");
1823 if ((amdgpu_testing & 2)) {
1824 if (adev->accel_working)
1825 amdgpu_test_syncing(adev);
1827 DRM_INFO("amdgpu: acceleration disabled, skipping sync tests\n");
1829 if (amdgpu_benchmarking) {
1830 if (adev->accel_working)
1831 amdgpu_benchmark(adev, amdgpu_benchmarking);
1833 DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n");
1836 /* enable clockgating, etc. after ib tests, etc. since some blocks require
1837 * explicit gating rather than handling it automatically.
1839 r = amdgpu_late_init(adev);
1841 dev_err(adev->dev, "amdgpu_late_init failed\n");
1849 vga_switcheroo_fini_domain_pm_ops(adev->dev);
1853 static void amdgpu_debugfs_remove_files(struct amdgpu_device *adev);
1856 * amdgpu_device_fini - tear down the driver
1858 * @adev: amdgpu_device pointer
1860 * Tear down the driver info (all asics).
1861 * Called at driver shutdown.
1863 void amdgpu_device_fini(struct amdgpu_device *adev)
1867 DRM_INFO("amdgpu: finishing device.\n");
1868 adev->shutdown = true;
1869 drm_crtc_force_disable_all(adev->ddev);
1870 /* evict vram memory */
1871 amdgpu_bo_evict_vram(adev);
1872 amdgpu_ib_pool_fini(adev);
1873 amdgpu_fence_driver_fini(adev);
1874 amdgpu_fbdev_fini(adev);
1875 r = amdgpu_fini(adev);
1876 adev->accel_working = false;
1877 /* free i2c buses */
1878 amdgpu_i2c_fini(adev);
1879 amdgpu_atombios_fini(adev);
1882 vga_switcheroo_unregister_client(adev->pdev);
1883 if (adev->flags & AMD_IS_PX)
1884 vga_switcheroo_fini_domain_pm_ops(adev->dev);
1885 vga_client_register(adev->pdev, NULL, NULL, NULL);
1887 pci_iounmap(adev->pdev, adev->rio_mem);
1888 adev->rio_mem = NULL;
1889 iounmap(adev->rmmio);
1891 if (adev->asic_type >= CHIP_BONAIRE)
1892 amdgpu_doorbell_fini(adev);
1893 amdgpu_debugfs_regs_cleanup(adev);
1894 amdgpu_debugfs_remove_files(adev);
1902 * amdgpu_device_suspend - initiate device suspend
1904 * @pdev: drm dev pointer
1905 * @state: suspend state
1907 * Puts the hw in the suspend state (all asics).
1908 * Returns 0 for success or an error on failure.
1909 * Called at driver suspend.
1911 int amdgpu_device_suspend(struct drm_device *dev, bool suspend, bool fbcon)
1913 struct amdgpu_device *adev;
1914 struct drm_crtc *crtc;
1915 struct drm_connector *connector;
1918 if (dev == NULL || dev->dev_private == NULL) {
1922 adev = dev->dev_private;
1924 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
1927 drm_kms_helper_poll_disable(dev);
1929 /* turn off display hw */
1930 drm_modeset_lock_all(dev);
1931 list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
1932 drm_helper_connector_dpms(connector, DRM_MODE_DPMS_OFF);
1934 drm_modeset_unlock_all(dev);
1936 /* unpin the front buffers and cursors */
1937 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
1938 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
1939 struct amdgpu_framebuffer *rfb = to_amdgpu_framebuffer(crtc->primary->fb);
1940 struct amdgpu_bo *robj;
1942 if (amdgpu_crtc->cursor_bo) {
1943 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
1944 r = amdgpu_bo_reserve(aobj, false);
1946 amdgpu_bo_unpin(aobj);
1947 amdgpu_bo_unreserve(aobj);
1951 if (rfb == NULL || rfb->obj == NULL) {
1954 robj = gem_to_amdgpu_bo(rfb->obj);
1955 /* don't unpin kernel fb objects */
1956 if (!amdgpu_fbdev_robj_is_fb(adev, robj)) {
1957 r = amdgpu_bo_reserve(robj, false);
1959 amdgpu_bo_unpin(robj);
1960 amdgpu_bo_unreserve(robj);
1964 /* evict vram memory */
1965 amdgpu_bo_evict_vram(adev);
1967 amdgpu_fence_driver_suspend(adev);
1969 r = amdgpu_suspend(adev);
1971 /* evict remaining vram memory
1972 * This second call to evict vram is to evict the gart page table
1975 amdgpu_bo_evict_vram(adev);
1977 amdgpu_atombios_scratch_regs_save(adev);
1978 pci_save_state(dev->pdev);
1980 /* Shut down the device */
1981 pci_disable_device(dev->pdev);
1982 pci_set_power_state(dev->pdev, PCI_D3hot);
1984 r = amdgpu_asic_reset(adev);
1986 DRM_ERROR("amdgpu asic reset failed\n");
1991 amdgpu_fbdev_set_suspend(adev, 1);
1998 * amdgpu_device_resume - initiate device resume
2000 * @pdev: drm dev pointer
2002 * Bring the hw back to operating state (all asics).
2003 * Returns 0 for success or an error on failure.
2004 * Called at driver resume.
2006 int amdgpu_device_resume(struct drm_device *dev, bool resume, bool fbcon)
2008 struct drm_connector *connector;
2009 struct amdgpu_device *adev = dev->dev_private;
2010 struct drm_crtc *crtc;
2013 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
2020 pci_set_power_state(dev->pdev, PCI_D0);
2021 pci_restore_state(dev->pdev);
2022 r = pci_enable_device(dev->pdev);
2029 amdgpu_atombios_scratch_regs_restore(adev);
2032 if (!amdgpu_card_posted(adev) || !resume) {
2033 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2035 DRM_ERROR("amdgpu asic init failed\n");
2038 r = amdgpu_resume(adev);
2040 DRM_ERROR("amdgpu_resume failed (%d).\n", r);
2042 amdgpu_fence_driver_resume(adev);
2045 r = amdgpu_ib_ring_tests(adev);
2047 DRM_ERROR("ib ring test failed (%d).\n", r);
2050 r = amdgpu_late_init(adev);
2055 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2056 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2058 if (amdgpu_crtc->cursor_bo) {
2059 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2060 r = amdgpu_bo_reserve(aobj, false);
2062 r = amdgpu_bo_pin(aobj,
2063 AMDGPU_GEM_DOMAIN_VRAM,
2064 &amdgpu_crtc->cursor_addr);
2066 DRM_ERROR("Failed to pin cursor BO (%d)\n", r);
2067 amdgpu_bo_unreserve(aobj);
2072 /* blat the mode back in */
2074 drm_helper_resume_force_mode(dev);
2075 /* turn on display hw */
2076 drm_modeset_lock_all(dev);
2077 list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2078 drm_helper_connector_dpms(connector, DRM_MODE_DPMS_ON);
2080 drm_modeset_unlock_all(dev);
2083 drm_kms_helper_poll_enable(dev);
2086 * Most of the connector probing functions try to acquire runtime pm
2087 * refs to ensure that the GPU is powered on when connector polling is
2088 * performed. Since we're calling this from a runtime PM callback,
2089 * trying to acquire rpm refs will cause us to deadlock.
2091 * Since we're guaranteed to be holding the rpm lock, it's safe to
2092 * temporarily disable the rpm helpers so this doesn't deadlock us.
2095 dev->dev->power.disable_depth++;
2097 drm_helper_hpd_irq_event(dev);
2099 dev->dev->power.disable_depth--;
2103 amdgpu_fbdev_set_suspend(adev, 0);
2110 static bool amdgpu_check_soft_reset(struct amdgpu_device *adev)
2113 bool asic_hang = false;
2115 for (i = 0; i < adev->num_ip_blocks; i++) {
2116 if (!adev->ip_blocks[i].status.valid)
2118 if (adev->ip_blocks[i].version->funcs->check_soft_reset)
2119 adev->ip_blocks[i].status.hang =
2120 adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
2121 if (adev->ip_blocks[i].status.hang) {
2122 DRM_INFO("IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
2129 static int amdgpu_pre_soft_reset(struct amdgpu_device *adev)
2133 for (i = 0; i < adev->num_ip_blocks; i++) {
2134 if (!adev->ip_blocks[i].status.valid)
2136 if (adev->ip_blocks[i].status.hang &&
2137 adev->ip_blocks[i].version->funcs->pre_soft_reset) {
2138 r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
2147 static bool amdgpu_need_full_reset(struct amdgpu_device *adev)
2151 for (i = 0; i < adev->num_ip_blocks; i++) {
2152 if (!adev->ip_blocks[i].status.valid)
2154 if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
2155 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
2156 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
2157 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE)) {
2158 if (adev->ip_blocks[i].status.hang) {
2159 DRM_INFO("Some block need full reset!\n");
2167 static int amdgpu_soft_reset(struct amdgpu_device *adev)
2171 for (i = 0; i < adev->num_ip_blocks; i++) {
2172 if (!adev->ip_blocks[i].status.valid)
2174 if (adev->ip_blocks[i].status.hang &&
2175 adev->ip_blocks[i].version->funcs->soft_reset) {
2176 r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
2185 static int amdgpu_post_soft_reset(struct amdgpu_device *adev)
2189 for (i = 0; i < adev->num_ip_blocks; i++) {
2190 if (!adev->ip_blocks[i].status.valid)
2192 if (adev->ip_blocks[i].status.hang &&
2193 adev->ip_blocks[i].version->funcs->post_soft_reset)
2194 r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
2202 bool amdgpu_need_backup(struct amdgpu_device *adev)
2204 if (adev->flags & AMD_IS_APU)
2207 return amdgpu_lockup_timeout > 0 ? true : false;
2210 static int amdgpu_recover_vram_from_shadow(struct amdgpu_device *adev,
2211 struct amdgpu_ring *ring,
2212 struct amdgpu_bo *bo,
2213 struct dma_fence **fence)
2221 r = amdgpu_bo_reserve(bo, false);
2224 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
2225 /* if bo has been evicted, then no need to recover */
2226 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
2227 r = amdgpu_bo_restore_from_shadow(adev, ring, bo,
2230 DRM_ERROR("recover page table failed!\n");
2235 amdgpu_bo_unreserve(bo);
2240 * amdgpu_gpu_reset - reset the asic
2242 * @adev: amdgpu device pointer
2244 * Attempt the reset the GPU if it has hung (all asics).
2245 * Returns 0 for success or an error on failure.
2247 int amdgpu_gpu_reset(struct amdgpu_device *adev)
2251 bool need_full_reset;
2253 if (!amdgpu_check_soft_reset(adev)) {
2254 DRM_INFO("No hardware hang detected. Did some blocks stall?\n");
2258 atomic_inc(&adev->gpu_reset_counter);
2261 resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
2263 /* block scheduler */
2264 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
2265 struct amdgpu_ring *ring = adev->rings[i];
2269 kthread_park(ring->sched.thread);
2270 amd_sched_hw_job_reset(&ring->sched);
2272 /* after all hw jobs are reset, hw fence is meaningless, so force_completion */
2273 amdgpu_fence_driver_force_completion(adev);
2275 need_full_reset = amdgpu_need_full_reset(adev);
2277 if (!need_full_reset) {
2278 amdgpu_pre_soft_reset(adev);
2279 r = amdgpu_soft_reset(adev);
2280 amdgpu_post_soft_reset(adev);
2281 if (r || amdgpu_check_soft_reset(adev)) {
2282 DRM_INFO("soft reset failed, will fallback to full reset!\n");
2283 need_full_reset = true;
2287 if (need_full_reset) {
2288 r = amdgpu_suspend(adev);
2291 /* Disable fb access */
2292 if (adev->mode_info.num_crtc) {
2293 struct amdgpu_mode_mc_save save;
2294 amdgpu_display_stop_mc_access(adev, &save);
2295 amdgpu_wait_for_idle(adev, AMD_IP_BLOCK_TYPE_GMC);
2297 amdgpu_atombios_scratch_regs_save(adev);
2298 r = amdgpu_asic_reset(adev);
2299 amdgpu_atombios_scratch_regs_restore(adev);
2301 amdgpu_atom_asic_init(adev->mode_info.atom_context);
2304 dev_info(adev->dev, "GPU reset succeeded, trying to resume\n");
2305 r = amdgpu_resume(adev);
2309 amdgpu_irq_gpu_reset_resume_helper(adev);
2310 if (need_full_reset && amdgpu_need_backup(adev)) {
2311 r = amdgpu_ttm_recover_gart(adev);
2313 DRM_ERROR("gart recovery failed!!!\n");
2315 r = amdgpu_ib_ring_tests(adev);
2317 dev_err(adev->dev, "ib ring test failed (%d).\n", r);
2318 r = amdgpu_suspend(adev);
2319 need_full_reset = true;
2323 * recovery vm page tables, since we cannot depend on VRAM is
2324 * consistent after gpu full reset.
2326 if (need_full_reset && amdgpu_need_backup(adev)) {
2327 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
2328 struct amdgpu_bo *bo, *tmp;
2329 struct dma_fence *fence = NULL, *next = NULL;
2331 DRM_INFO("recover vram bo from shadow\n");
2332 mutex_lock(&adev->shadow_list_lock);
2333 list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) {
2334 amdgpu_recover_vram_from_shadow(adev, ring, bo, &next);
2336 r = dma_fence_wait(fence, false);
2338 WARN(r, "recovery from shadow isn't comleted\n");
2343 dma_fence_put(fence);
2346 mutex_unlock(&adev->shadow_list_lock);
2348 r = dma_fence_wait(fence, false);
2350 WARN(r, "recovery from shadow isn't comleted\n");
2352 dma_fence_put(fence);
2354 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
2355 struct amdgpu_ring *ring = adev->rings[i];
2359 amd_sched_job_recovery(&ring->sched);
2360 kthread_unpark(ring->sched.thread);
2363 dev_err(adev->dev, "asic resume failed (%d).\n", r);
2364 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
2365 if (adev->rings[i]) {
2366 kthread_unpark(adev->rings[i]->sched.thread);
2371 drm_helper_resume_force_mode(adev->ddev);
2373 ttm_bo_unlock_delayed_workqueue(&adev->mman.bdev, resched);
2375 /* bad news, how to tell it to userspace ? */
2376 dev_info(adev->dev, "GPU reset failed\n");
2382 void amdgpu_get_pcie_info(struct amdgpu_device *adev)
2387 if (amdgpu_pcie_gen_cap)
2388 adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
2390 if (amdgpu_pcie_lane_cap)
2391 adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
2393 /* covers APUs as well */
2394 if (pci_is_root_bus(adev->pdev->bus)) {
2395 if (adev->pm.pcie_gen_mask == 0)
2396 adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
2397 if (adev->pm.pcie_mlw_mask == 0)
2398 adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
2402 if (adev->pm.pcie_gen_mask == 0) {
2403 ret = drm_pcie_get_speed_cap_mask(adev->ddev, &mask);
2405 adev->pm.pcie_gen_mask = (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
2406 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
2407 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
2409 if (mask & DRM_PCIE_SPEED_25)
2410 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
2411 if (mask & DRM_PCIE_SPEED_50)
2412 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2;
2413 if (mask & DRM_PCIE_SPEED_80)
2414 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3;
2416 adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
2419 if (adev->pm.pcie_mlw_mask == 0) {
2420 ret = drm_pcie_get_max_link_width(adev->ddev, &mask);
2424 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
2425 CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
2426 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
2427 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
2428 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
2429 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2430 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2433 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
2434 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
2435 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
2436 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
2437 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2438 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2441 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
2442 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
2443 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
2444 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2445 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2448 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
2449 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
2450 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2451 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2454 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
2455 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2456 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2459 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2460 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2463 adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;
2469 adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
2477 int amdgpu_debugfs_add_files(struct amdgpu_device *adev,
2478 const struct drm_info_list *files,
2483 for (i = 0; i < adev->debugfs_count; i++) {
2484 if (adev->debugfs[i].files == files) {
2485 /* Already registered */
2490 i = adev->debugfs_count + 1;
2491 if (i > AMDGPU_DEBUGFS_MAX_COMPONENTS) {
2492 DRM_ERROR("Reached maximum number of debugfs components.\n");
2493 DRM_ERROR("Report so we increase "
2494 "AMDGPU_DEBUGFS_MAX_COMPONENTS.\n");
2497 adev->debugfs[adev->debugfs_count].files = files;
2498 adev->debugfs[adev->debugfs_count].num_files = nfiles;
2499 adev->debugfs_count = i;
2500 #if defined(CONFIG_DEBUG_FS)
2501 drm_debugfs_create_files(files, nfiles,
2502 adev->ddev->primary->debugfs_root,
2503 adev->ddev->primary);
2508 static void amdgpu_debugfs_remove_files(struct amdgpu_device *adev)
2510 #if defined(CONFIG_DEBUG_FS)
2513 for (i = 0; i < adev->debugfs_count; i++) {
2514 drm_debugfs_remove_files(adev->debugfs[i].files,
2515 adev->debugfs[i].num_files,
2516 adev->ddev->primary);
2521 #if defined(CONFIG_DEBUG_FS)
2523 static ssize_t amdgpu_debugfs_regs_read(struct file *f, char __user *buf,
2524 size_t size, loff_t *pos)
2526 struct amdgpu_device *adev = file_inode(f)->i_private;
2529 bool pm_pg_lock, use_bank;
2530 unsigned instance_bank, sh_bank, se_bank;
2532 if (size & 0x3 || *pos & 0x3)
2535 /* are we reading registers for which a PG lock is necessary? */
2536 pm_pg_lock = (*pos >> 23) & 1;
2538 if (*pos & (1ULL << 62)) {
2539 se_bank = (*pos >> 24) & 0x3FF;
2540 sh_bank = (*pos >> 34) & 0x3FF;
2541 instance_bank = (*pos >> 44) & 0x3FF;
2543 if (se_bank == 0x3FF)
2544 se_bank = 0xFFFFFFFF;
2545 if (sh_bank == 0x3FF)
2546 sh_bank = 0xFFFFFFFF;
2547 if (instance_bank == 0x3FF)
2548 instance_bank = 0xFFFFFFFF;
2557 if ((sh_bank != 0xFFFFFFFF && sh_bank >= adev->gfx.config.max_sh_per_se) ||
2558 (se_bank != 0xFFFFFFFF && se_bank >= adev->gfx.config.max_shader_engines))
2560 mutex_lock(&adev->grbm_idx_mutex);
2561 amdgpu_gfx_select_se_sh(adev, se_bank,
2562 sh_bank, instance_bank);
2566 mutex_lock(&adev->pm.mutex);
2571 if (*pos > adev->rmmio_size)
2574 value = RREG32(*pos >> 2);
2575 r = put_user(value, (uint32_t *)buf);
2589 amdgpu_gfx_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff);
2590 mutex_unlock(&adev->grbm_idx_mutex);
2594 mutex_unlock(&adev->pm.mutex);
2599 static ssize_t amdgpu_debugfs_regs_write(struct file *f, const char __user *buf,
2600 size_t size, loff_t *pos)
2602 struct amdgpu_device *adev = file_inode(f)->i_private;
2605 bool pm_pg_lock, use_bank;
2606 unsigned instance_bank, sh_bank, se_bank;
2608 if (size & 0x3 || *pos & 0x3)
2611 /* are we reading registers for which a PG lock is necessary? */
2612 pm_pg_lock = (*pos >> 23) & 1;
2614 if (*pos & (1ULL << 62)) {
2615 se_bank = (*pos >> 24) & 0x3FF;
2616 sh_bank = (*pos >> 34) & 0x3FF;
2617 instance_bank = (*pos >> 44) & 0x3FF;
2619 if (se_bank == 0x3FF)
2620 se_bank = 0xFFFFFFFF;
2621 if (sh_bank == 0x3FF)
2622 sh_bank = 0xFFFFFFFF;
2623 if (instance_bank == 0x3FF)
2624 instance_bank = 0xFFFFFFFF;
2633 if ((sh_bank != 0xFFFFFFFF && sh_bank >= adev->gfx.config.max_sh_per_se) ||
2634 (se_bank != 0xFFFFFFFF && se_bank >= adev->gfx.config.max_shader_engines))
2636 mutex_lock(&adev->grbm_idx_mutex);
2637 amdgpu_gfx_select_se_sh(adev, se_bank,
2638 sh_bank, instance_bank);
2642 mutex_lock(&adev->pm.mutex);
2647 if (*pos > adev->rmmio_size)
2650 r = get_user(value, (uint32_t *)buf);
2654 WREG32(*pos >> 2, value);
2663 amdgpu_gfx_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff);
2664 mutex_unlock(&adev->grbm_idx_mutex);
2668 mutex_unlock(&adev->pm.mutex);
2673 static ssize_t amdgpu_debugfs_regs_pcie_read(struct file *f, char __user *buf,
2674 size_t size, loff_t *pos)
2676 struct amdgpu_device *adev = file_inode(f)->i_private;
2680 if (size & 0x3 || *pos & 0x3)
2686 value = RREG32_PCIE(*pos >> 2);
2687 r = put_user(value, (uint32_t *)buf);
2700 static ssize_t amdgpu_debugfs_regs_pcie_write(struct file *f, const char __user *buf,
2701 size_t size, loff_t *pos)
2703 struct amdgpu_device *adev = file_inode(f)->i_private;
2707 if (size & 0x3 || *pos & 0x3)
2713 r = get_user(value, (uint32_t *)buf);
2717 WREG32_PCIE(*pos >> 2, value);
2728 static ssize_t amdgpu_debugfs_regs_didt_read(struct file *f, char __user *buf,
2729 size_t size, loff_t *pos)
2731 struct amdgpu_device *adev = file_inode(f)->i_private;
2735 if (size & 0x3 || *pos & 0x3)
2741 value = RREG32_DIDT(*pos >> 2);
2742 r = put_user(value, (uint32_t *)buf);
2755 static ssize_t amdgpu_debugfs_regs_didt_write(struct file *f, const char __user *buf,
2756 size_t size, loff_t *pos)
2758 struct amdgpu_device *adev = file_inode(f)->i_private;
2762 if (size & 0x3 || *pos & 0x3)
2768 r = get_user(value, (uint32_t *)buf);
2772 WREG32_DIDT(*pos >> 2, value);
2783 static ssize_t amdgpu_debugfs_regs_smc_read(struct file *f, char __user *buf,
2784 size_t size, loff_t *pos)
2786 struct amdgpu_device *adev = file_inode(f)->i_private;
2790 if (size & 0x3 || *pos & 0x3)
2796 value = RREG32_SMC(*pos);
2797 r = put_user(value, (uint32_t *)buf);
2810 static ssize_t amdgpu_debugfs_regs_smc_write(struct file *f, const char __user *buf,
2811 size_t size, loff_t *pos)
2813 struct amdgpu_device *adev = file_inode(f)->i_private;
2817 if (size & 0x3 || *pos & 0x3)
2823 r = get_user(value, (uint32_t *)buf);
2827 WREG32_SMC(*pos, value);
2838 static ssize_t amdgpu_debugfs_gca_config_read(struct file *f, char __user *buf,
2839 size_t size, loff_t *pos)
2841 struct amdgpu_device *adev = file_inode(f)->i_private;
2844 uint32_t *config, no_regs = 0;
2846 if (size & 0x3 || *pos & 0x3)
2849 config = kmalloc_array(256, sizeof(*config), GFP_KERNEL);
2853 /* version, increment each time something is added */
2854 config[no_regs++] = 2;
2855 config[no_regs++] = adev->gfx.config.max_shader_engines;
2856 config[no_regs++] = adev->gfx.config.max_tile_pipes;
2857 config[no_regs++] = adev->gfx.config.max_cu_per_sh;
2858 config[no_regs++] = adev->gfx.config.max_sh_per_se;
2859 config[no_regs++] = adev->gfx.config.max_backends_per_se;
2860 config[no_regs++] = adev->gfx.config.max_texture_channel_caches;
2861 config[no_regs++] = adev->gfx.config.max_gprs;
2862 config[no_regs++] = adev->gfx.config.max_gs_threads;
2863 config[no_regs++] = adev->gfx.config.max_hw_contexts;
2864 config[no_regs++] = adev->gfx.config.sc_prim_fifo_size_frontend;
2865 config[no_regs++] = adev->gfx.config.sc_prim_fifo_size_backend;
2866 config[no_regs++] = adev->gfx.config.sc_hiz_tile_fifo_size;
2867 config[no_regs++] = adev->gfx.config.sc_earlyz_tile_fifo_size;
2868 config[no_regs++] = adev->gfx.config.num_tile_pipes;
2869 config[no_regs++] = adev->gfx.config.backend_enable_mask;
2870 config[no_regs++] = adev->gfx.config.mem_max_burst_length_bytes;
2871 config[no_regs++] = adev->gfx.config.mem_row_size_in_kb;
2872 config[no_regs++] = adev->gfx.config.shader_engine_tile_size;
2873 config[no_regs++] = adev->gfx.config.num_gpus;
2874 config[no_regs++] = adev->gfx.config.multi_gpu_tile_size;
2875 config[no_regs++] = adev->gfx.config.mc_arb_ramcfg;
2876 config[no_regs++] = adev->gfx.config.gb_addr_config;
2877 config[no_regs++] = adev->gfx.config.num_rbs;
2880 config[no_regs++] = adev->rev_id;
2881 config[no_regs++] = adev->pg_flags;
2882 config[no_regs++] = adev->cg_flags;
2885 config[no_regs++] = adev->family;
2886 config[no_regs++] = adev->external_rev_id;
2888 while (size && (*pos < no_regs * 4)) {
2891 value = config[*pos >> 2];
2892 r = put_user(value, (uint32_t *)buf);
2908 static ssize_t amdgpu_debugfs_sensor_read(struct file *f, char __user *buf,
2909 size_t size, loff_t *pos)
2911 struct amdgpu_device *adev = file_inode(f)->i_private;
2915 if (size != 4 || *pos & 0x3)
2918 /* convert offset to sensor number */
2921 if (adev->powerplay.pp_funcs && adev->powerplay.pp_funcs->read_sensor)
2922 r = adev->powerplay.pp_funcs->read_sensor(adev->powerplay.pp_handle, idx, &value);
2927 r = put_user(value, (int32_t *)buf);
2932 static ssize_t amdgpu_debugfs_wave_read(struct file *f, char __user *buf,
2933 size_t size, loff_t *pos)
2935 struct amdgpu_device *adev = f->f_inode->i_private;
2938 uint32_t offset, se, sh, cu, wave, simd, data[32];
2940 if (size & 3 || *pos & 3)
2944 offset = (*pos & 0x7F);
2945 se = ((*pos >> 7) & 0xFF);
2946 sh = ((*pos >> 15) & 0xFF);
2947 cu = ((*pos >> 23) & 0xFF);
2948 wave = ((*pos >> 31) & 0xFF);
2949 simd = ((*pos >> 37) & 0xFF);
2951 /* switch to the specific se/sh/cu */
2952 mutex_lock(&adev->grbm_idx_mutex);
2953 amdgpu_gfx_select_se_sh(adev, se, sh, cu);
2956 if (adev->gfx.funcs->read_wave_data)
2957 adev->gfx.funcs->read_wave_data(adev, simd, wave, data, &x);
2959 amdgpu_gfx_select_se_sh(adev, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
2960 mutex_unlock(&adev->grbm_idx_mutex);
2965 while (size && (offset < x * 4)) {
2968 value = data[offset >> 2];
2969 r = put_user(value, (uint32_t *)buf);
2982 static ssize_t amdgpu_debugfs_gpr_read(struct file *f, char __user *buf,
2983 size_t size, loff_t *pos)
2985 struct amdgpu_device *adev = f->f_inode->i_private;
2988 uint32_t offset, se, sh, cu, wave, simd, thread, bank, *data;
2990 if (size & 3 || *pos & 3)
2994 offset = (*pos & 0xFFF); /* in dwords */
2995 se = ((*pos >> 12) & 0xFF);
2996 sh = ((*pos >> 20) & 0xFF);
2997 cu = ((*pos >> 28) & 0xFF);
2998 wave = ((*pos >> 36) & 0xFF);
2999 simd = ((*pos >> 44) & 0xFF);
3000 thread = ((*pos >> 52) & 0xFF);
3001 bank = ((*pos >> 60) & 1);
3003 data = kmalloc_array(1024, sizeof(*data), GFP_KERNEL);
3007 /* switch to the specific se/sh/cu */
3008 mutex_lock(&adev->grbm_idx_mutex);
3009 amdgpu_gfx_select_se_sh(adev, se, sh, cu);
3012 if (adev->gfx.funcs->read_wave_vgprs)
3013 adev->gfx.funcs->read_wave_vgprs(adev, simd, wave, thread, offset, size>>2, data);
3015 if (adev->gfx.funcs->read_wave_sgprs)
3016 adev->gfx.funcs->read_wave_sgprs(adev, simd, wave, offset, size>>2, data);
3019 amdgpu_gfx_select_se_sh(adev, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
3020 mutex_unlock(&adev->grbm_idx_mutex);
3025 value = data[offset++];
3026 r = put_user(value, (uint32_t *)buf);
3042 static const struct file_operations amdgpu_debugfs_regs_fops = {
3043 .owner = THIS_MODULE,
3044 .read = amdgpu_debugfs_regs_read,
3045 .write = amdgpu_debugfs_regs_write,
3046 .llseek = default_llseek
3048 static const struct file_operations amdgpu_debugfs_regs_didt_fops = {
3049 .owner = THIS_MODULE,
3050 .read = amdgpu_debugfs_regs_didt_read,
3051 .write = amdgpu_debugfs_regs_didt_write,
3052 .llseek = default_llseek
3054 static const struct file_operations amdgpu_debugfs_regs_pcie_fops = {
3055 .owner = THIS_MODULE,
3056 .read = amdgpu_debugfs_regs_pcie_read,
3057 .write = amdgpu_debugfs_regs_pcie_write,
3058 .llseek = default_llseek
3060 static const struct file_operations amdgpu_debugfs_regs_smc_fops = {
3061 .owner = THIS_MODULE,
3062 .read = amdgpu_debugfs_regs_smc_read,
3063 .write = amdgpu_debugfs_regs_smc_write,
3064 .llseek = default_llseek
3067 static const struct file_operations amdgpu_debugfs_gca_config_fops = {
3068 .owner = THIS_MODULE,
3069 .read = amdgpu_debugfs_gca_config_read,
3070 .llseek = default_llseek
3073 static const struct file_operations amdgpu_debugfs_sensors_fops = {
3074 .owner = THIS_MODULE,
3075 .read = amdgpu_debugfs_sensor_read,
3076 .llseek = default_llseek
3079 static const struct file_operations amdgpu_debugfs_wave_fops = {
3080 .owner = THIS_MODULE,
3081 .read = amdgpu_debugfs_wave_read,
3082 .llseek = default_llseek
3084 static const struct file_operations amdgpu_debugfs_gpr_fops = {
3085 .owner = THIS_MODULE,
3086 .read = amdgpu_debugfs_gpr_read,
3087 .llseek = default_llseek
3090 static const struct file_operations *debugfs_regs[] = {
3091 &amdgpu_debugfs_regs_fops,
3092 &amdgpu_debugfs_regs_didt_fops,
3093 &amdgpu_debugfs_regs_pcie_fops,
3094 &amdgpu_debugfs_regs_smc_fops,
3095 &amdgpu_debugfs_gca_config_fops,
3096 &amdgpu_debugfs_sensors_fops,
3097 &amdgpu_debugfs_wave_fops,
3098 &amdgpu_debugfs_gpr_fops,
3101 static const char *debugfs_regs_names[] = {
3106 "amdgpu_gca_config",
3112 static int amdgpu_debugfs_regs_init(struct amdgpu_device *adev)
3114 struct drm_minor *minor = adev->ddev->primary;
3115 struct dentry *ent, *root = minor->debugfs_root;
3118 for (i = 0; i < ARRAY_SIZE(debugfs_regs); i++) {
3119 ent = debugfs_create_file(debugfs_regs_names[i],
3120 S_IFREG | S_IRUGO, root,
3121 adev, debugfs_regs[i]);
3123 for (j = 0; j < i; j++) {
3124 debugfs_remove(adev->debugfs_regs[i]);
3125 adev->debugfs_regs[i] = NULL;
3127 return PTR_ERR(ent);
3131 i_size_write(ent->d_inode, adev->rmmio_size);
3132 adev->debugfs_regs[i] = ent;
3138 static void amdgpu_debugfs_regs_cleanup(struct amdgpu_device *adev)
3142 for (i = 0; i < ARRAY_SIZE(debugfs_regs); i++) {
3143 if (adev->debugfs_regs[i]) {
3144 debugfs_remove(adev->debugfs_regs[i]);
3145 adev->debugfs_regs[i] = NULL;
3150 int amdgpu_debugfs_init(struct drm_minor *minor)
3155 void amdgpu_debugfs_cleanup(struct drm_minor *minor)
3159 static int amdgpu_debugfs_regs_init(struct amdgpu_device *adev)
3163 static void amdgpu_debugfs_regs_cleanup(struct amdgpu_device *adev) { }