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[] = {
80 bool amdgpu_device_is_px(struct drm_device *dev)
82 struct amdgpu_device *adev = dev->dev_private;
84 if (adev->flags & AMD_IS_PX)
90 * MMIO register access helper functions.
92 uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg,
97 if ((reg * 4) < adev->rmmio_size && !always_indirect)
98 ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
102 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
103 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
104 ret = readl(((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
105 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
107 trace_amdgpu_mm_rreg(adev->pdev->device, reg, ret);
111 void amdgpu_mm_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v,
112 bool always_indirect)
114 trace_amdgpu_mm_wreg(adev->pdev->device, reg, v);
116 if ((reg * 4) < adev->rmmio_size && !always_indirect)
117 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
121 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
122 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
123 writel(v, ((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
124 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
128 u32 amdgpu_io_rreg(struct amdgpu_device *adev, u32 reg)
130 if ((reg * 4) < adev->rio_mem_size)
131 return ioread32(adev->rio_mem + (reg * 4));
133 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
134 return ioread32(adev->rio_mem + (mmMM_DATA * 4));
138 void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32 v)
141 if ((reg * 4) < adev->rio_mem_size)
142 iowrite32(v, adev->rio_mem + (reg * 4));
144 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
145 iowrite32(v, adev->rio_mem + (mmMM_DATA * 4));
150 * amdgpu_mm_rdoorbell - read a doorbell dword
152 * @adev: amdgpu_device pointer
153 * @index: doorbell index
155 * Returns the value in the doorbell aperture at the
156 * requested doorbell index (CIK).
158 u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
160 if (index < adev->doorbell.num_doorbells) {
161 return readl(adev->doorbell.ptr + index);
163 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
169 * amdgpu_mm_wdoorbell - write a doorbell dword
171 * @adev: amdgpu_device pointer
172 * @index: doorbell index
175 * Writes @v to the doorbell aperture at the
176 * requested doorbell index (CIK).
178 void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v)
180 if (index < adev->doorbell.num_doorbells) {
181 writel(v, adev->doorbell.ptr + index);
183 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
188 * amdgpu_invalid_rreg - dummy reg read function
190 * @adev: amdgpu device pointer
191 * @reg: offset of register
193 * Dummy register read function. Used for register blocks
194 * that certain asics don't have (all asics).
195 * Returns the value in the register.
197 static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
199 DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
205 * amdgpu_invalid_wreg - dummy reg write function
207 * @adev: amdgpu device pointer
208 * @reg: offset of register
209 * @v: value to write to the register
211 * Dummy register read function. Used for register blocks
212 * that certain asics don't have (all asics).
214 static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
216 DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
222 * amdgpu_block_invalid_rreg - dummy reg read function
224 * @adev: amdgpu device pointer
225 * @block: offset of instance
226 * @reg: offset of register
228 * Dummy register read function. Used for register blocks
229 * that certain asics don't have (all asics).
230 * Returns the value in the register.
232 static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
233 uint32_t block, uint32_t reg)
235 DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
242 * amdgpu_block_invalid_wreg - dummy reg write function
244 * @adev: amdgpu device pointer
245 * @block: offset of instance
246 * @reg: offset of register
247 * @v: value to write to the register
249 * Dummy register read function. Used for register blocks
250 * that certain asics don't have (all asics).
252 static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
254 uint32_t reg, uint32_t v)
256 DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
261 static int amdgpu_vram_scratch_init(struct amdgpu_device *adev)
265 if (adev->vram_scratch.robj == NULL) {
266 r = amdgpu_bo_create(adev, AMDGPU_GPU_PAGE_SIZE,
267 PAGE_SIZE, true, AMDGPU_GEM_DOMAIN_VRAM,
268 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
269 AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS,
270 NULL, NULL, &adev->vram_scratch.robj);
276 r = amdgpu_bo_reserve(adev->vram_scratch.robj, false);
277 if (unlikely(r != 0))
279 r = amdgpu_bo_pin(adev->vram_scratch.robj,
280 AMDGPU_GEM_DOMAIN_VRAM, &adev->vram_scratch.gpu_addr);
282 amdgpu_bo_unreserve(adev->vram_scratch.robj);
285 r = amdgpu_bo_kmap(adev->vram_scratch.robj,
286 (void **)&adev->vram_scratch.ptr);
288 amdgpu_bo_unpin(adev->vram_scratch.robj);
289 amdgpu_bo_unreserve(adev->vram_scratch.robj);
294 static void amdgpu_vram_scratch_fini(struct amdgpu_device *adev)
298 if (adev->vram_scratch.robj == NULL) {
301 r = amdgpu_bo_reserve(adev->vram_scratch.robj, false);
302 if (likely(r == 0)) {
303 amdgpu_bo_kunmap(adev->vram_scratch.robj);
304 amdgpu_bo_unpin(adev->vram_scratch.robj);
305 amdgpu_bo_unreserve(adev->vram_scratch.robj);
307 amdgpu_bo_unref(&adev->vram_scratch.robj);
311 * amdgpu_program_register_sequence - program an array of registers.
313 * @adev: amdgpu_device pointer
314 * @registers: pointer to the register array
315 * @array_size: size of the register array
317 * Programs an array or registers with and and or masks.
318 * This is a helper for setting golden registers.
320 void amdgpu_program_register_sequence(struct amdgpu_device *adev,
321 const u32 *registers,
322 const u32 array_size)
324 u32 tmp, reg, and_mask, or_mask;
330 for (i = 0; i < array_size; i +=3) {
331 reg = registers[i + 0];
332 and_mask = registers[i + 1];
333 or_mask = registers[i + 2];
335 if (and_mask == 0xffffffff) {
346 void amdgpu_pci_config_reset(struct amdgpu_device *adev)
348 pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
352 * GPU doorbell aperture helpers function.
355 * amdgpu_doorbell_init - Init doorbell driver information.
357 * @adev: amdgpu_device pointer
359 * Init doorbell driver information (CIK)
360 * Returns 0 on success, error on failure.
362 static int amdgpu_doorbell_init(struct amdgpu_device *adev)
364 /* doorbell bar mapping */
365 adev->doorbell.base = pci_resource_start(adev->pdev, 2);
366 adev->doorbell.size = pci_resource_len(adev->pdev, 2);
368 adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32),
369 AMDGPU_DOORBELL_MAX_ASSIGNMENT+1);
370 if (adev->doorbell.num_doorbells == 0)
373 adev->doorbell.ptr = ioremap(adev->doorbell.base, adev->doorbell.num_doorbells * sizeof(u32));
374 if (adev->doorbell.ptr == NULL) {
377 DRM_INFO("doorbell mmio base: 0x%08X\n", (uint32_t)adev->doorbell.base);
378 DRM_INFO("doorbell mmio size: %u\n", (unsigned)adev->doorbell.size);
384 * amdgpu_doorbell_fini - Tear down doorbell driver information.
386 * @adev: amdgpu_device pointer
388 * Tear down doorbell driver information (CIK)
390 static void amdgpu_doorbell_fini(struct amdgpu_device *adev)
392 iounmap(adev->doorbell.ptr);
393 adev->doorbell.ptr = NULL;
397 * amdgpu_doorbell_get_kfd_info - Report doorbell configuration required to
400 * @adev: amdgpu_device pointer
401 * @aperture_base: output returning doorbell aperture base physical address
402 * @aperture_size: output returning doorbell aperture size in bytes
403 * @start_offset: output returning # of doorbell bytes reserved for amdgpu.
405 * amdgpu and amdkfd share the doorbell aperture. amdgpu sets it up,
406 * takes doorbells required for its own rings and reports the setup to amdkfd.
407 * amdgpu reserved doorbells are at the start of the doorbell aperture.
409 void amdgpu_doorbell_get_kfd_info(struct amdgpu_device *adev,
410 phys_addr_t *aperture_base,
411 size_t *aperture_size,
412 size_t *start_offset)
415 * The first num_doorbells are used by amdgpu.
416 * amdkfd takes whatever's left in the aperture.
418 if (adev->doorbell.size > adev->doorbell.num_doorbells * sizeof(u32)) {
419 *aperture_base = adev->doorbell.base;
420 *aperture_size = adev->doorbell.size;
421 *start_offset = adev->doorbell.num_doorbells * sizeof(u32);
431 * Writeback is the the method by which the the GPU updates special pages
432 * in memory with the status of certain GPU events (fences, ring pointers,
437 * amdgpu_wb_fini - Disable Writeback and free memory
439 * @adev: amdgpu_device pointer
441 * Disables Writeback and frees the Writeback memory (all asics).
442 * Used at driver shutdown.
444 static void amdgpu_wb_fini(struct amdgpu_device *adev)
446 if (adev->wb.wb_obj) {
447 amdgpu_bo_free_kernel(&adev->wb.wb_obj,
449 (void **)&adev->wb.wb);
450 adev->wb.wb_obj = NULL;
455 * amdgpu_wb_init- Init Writeback driver info and allocate memory
457 * @adev: amdgpu_device pointer
459 * Disables Writeback and frees the Writeback memory (all asics).
460 * Used at driver startup.
461 * Returns 0 on success or an -error on failure.
463 static int amdgpu_wb_init(struct amdgpu_device *adev)
467 if (adev->wb.wb_obj == NULL) {
468 r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * 4,
469 PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
470 &adev->wb.wb_obj, &adev->wb.gpu_addr,
471 (void **)&adev->wb.wb);
473 dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
477 adev->wb.num_wb = AMDGPU_MAX_WB;
478 memset(&adev->wb.used, 0, sizeof(adev->wb.used));
480 /* clear wb memory */
481 memset((char *)adev->wb.wb, 0, AMDGPU_GPU_PAGE_SIZE);
488 * amdgpu_wb_get - Allocate a wb entry
490 * @adev: amdgpu_device pointer
493 * Allocate a wb slot for use by the driver (all asics).
494 * Returns 0 on success or -EINVAL on failure.
496 int amdgpu_wb_get(struct amdgpu_device *adev, u32 *wb)
498 unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
499 if (offset < adev->wb.num_wb) {
500 __set_bit(offset, adev->wb.used);
509 * amdgpu_wb_free - Free a wb entry
511 * @adev: amdgpu_device pointer
514 * Free a wb slot allocated for use by the driver (all asics)
516 void amdgpu_wb_free(struct amdgpu_device *adev, u32 wb)
518 if (wb < adev->wb.num_wb)
519 __clear_bit(wb, adev->wb.used);
523 * amdgpu_vram_location - try to find VRAM location
524 * @adev: amdgpu device structure holding all necessary informations
525 * @mc: memory controller structure holding memory informations
526 * @base: base address at which to put VRAM
528 * Function will place try to place VRAM at base address provided
529 * as parameter (which is so far either PCI aperture address or
530 * for IGP TOM base address).
532 * If there is not enough space to fit the unvisible VRAM in the 32bits
533 * address space then we limit the VRAM size to the aperture.
535 * Note: We don't explicitly enforce VRAM start to be aligned on VRAM size,
536 * this shouldn't be a problem as we are using the PCI aperture as a reference.
537 * Otherwise this would be needed for rv280, all r3xx, and all r4xx, but
540 * Note: we use mc_vram_size as on some board we need to program the mc to
541 * cover the whole aperture even if VRAM size is inferior to aperture size
542 * Novell bug 204882 + along with lots of ubuntu ones
544 * Note: when limiting vram it's safe to overwritte real_vram_size because
545 * we are not in case where real_vram_size is inferior to mc_vram_size (ie
546 * note afected by bogus hw of Novell bug 204882 + along with lots of ubuntu
549 * Note: IGP TOM addr should be the same as the aperture addr, we don't
550 * explicitly check for that thought.
552 * FIXME: when reducing VRAM size align new size on power of 2.
554 void amdgpu_vram_location(struct amdgpu_device *adev, struct amdgpu_mc *mc, u64 base)
556 uint64_t limit = (uint64_t)amdgpu_vram_limit << 20;
558 mc->vram_start = base;
559 if (mc->mc_vram_size > (adev->mc.mc_mask - base + 1)) {
560 dev_warn(adev->dev, "limiting VRAM to PCI aperture size\n");
561 mc->real_vram_size = mc->aper_size;
562 mc->mc_vram_size = mc->aper_size;
564 mc->vram_end = mc->vram_start + mc->mc_vram_size - 1;
565 if (limit && limit < mc->real_vram_size)
566 mc->real_vram_size = limit;
567 dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
568 mc->mc_vram_size >> 20, mc->vram_start,
569 mc->vram_end, mc->real_vram_size >> 20);
573 * amdgpu_gtt_location - try to find GTT location
574 * @adev: amdgpu device structure holding all necessary informations
575 * @mc: memory controller structure holding memory informations
577 * Function will place try to place GTT before or after VRAM.
579 * If GTT size is bigger than space left then we ajust GTT size.
580 * Thus function will never fails.
582 * FIXME: when reducing GTT size align new size on power of 2.
584 void amdgpu_gtt_location(struct amdgpu_device *adev, struct amdgpu_mc *mc)
586 u64 size_af, size_bf;
588 size_af = ((adev->mc.mc_mask - mc->vram_end) + mc->gtt_base_align) & ~mc->gtt_base_align;
589 size_bf = mc->vram_start & ~mc->gtt_base_align;
590 if (size_bf > size_af) {
591 if (mc->gtt_size > size_bf) {
592 dev_warn(adev->dev, "limiting GTT\n");
593 mc->gtt_size = size_bf;
595 mc->gtt_start = (mc->vram_start & ~mc->gtt_base_align) - mc->gtt_size;
597 if (mc->gtt_size > size_af) {
598 dev_warn(adev->dev, "limiting GTT\n");
599 mc->gtt_size = size_af;
601 mc->gtt_start = (mc->vram_end + 1 + mc->gtt_base_align) & ~mc->gtt_base_align;
603 mc->gtt_end = mc->gtt_start + mc->gtt_size - 1;
604 dev_info(adev->dev, "GTT: %lluM 0x%016llX - 0x%016llX\n",
605 mc->gtt_size >> 20, mc->gtt_start, mc->gtt_end);
609 * GPU helpers function.
612 * amdgpu_card_posted - check if the hw has already been initialized
614 * @adev: amdgpu_device pointer
616 * Check if the asic has been initialized (all asics).
617 * Used at driver startup.
618 * Returns true if initialized or false if not.
620 bool amdgpu_card_posted(struct amdgpu_device *adev)
624 /* then check MEM_SIZE, in case the crtcs are off */
625 reg = RREG32(mmCONFIG_MEMSIZE);
634 static bool amdgpu_vpost_needed(struct amdgpu_device *adev)
636 if (amdgpu_sriov_vf(adev))
639 if (amdgpu_passthrough(adev)) {
640 /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
641 * some old smc fw still need driver do vPost otherwise gpu hang, while
642 * those smc fw version above 22.15 doesn't have this flaw, so we force
643 * vpost executed for smc version below 22.15
645 if (adev->asic_type == CHIP_FIJI) {
648 err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
649 /* force vPost if error occured */
653 fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
654 if (fw_ver < 0x00160e00)
658 return !amdgpu_card_posted(adev);
662 * amdgpu_dummy_page_init - init dummy page used by the driver
664 * @adev: amdgpu_device pointer
666 * Allocate the dummy page used by the driver (all asics).
667 * This dummy page is used by the driver as a filler for gart entries
668 * when pages are taken out of the GART
669 * Returns 0 on sucess, -ENOMEM on failure.
671 int amdgpu_dummy_page_init(struct amdgpu_device *adev)
673 if (adev->dummy_page.page)
675 adev->dummy_page.page = alloc_page(GFP_DMA32 | GFP_KERNEL | __GFP_ZERO);
676 if (adev->dummy_page.page == NULL)
678 adev->dummy_page.addr = pci_map_page(adev->pdev, adev->dummy_page.page,
679 0, PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
680 if (pci_dma_mapping_error(adev->pdev, adev->dummy_page.addr)) {
681 dev_err(&adev->pdev->dev, "Failed to DMA MAP the dummy page\n");
682 __free_page(adev->dummy_page.page);
683 adev->dummy_page.page = NULL;
690 * amdgpu_dummy_page_fini - free dummy page used by the driver
692 * @adev: amdgpu_device pointer
694 * Frees the dummy page used by the driver (all asics).
696 void amdgpu_dummy_page_fini(struct amdgpu_device *adev)
698 if (adev->dummy_page.page == NULL)
700 pci_unmap_page(adev->pdev, adev->dummy_page.addr,
701 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
702 __free_page(adev->dummy_page.page);
703 adev->dummy_page.page = NULL;
707 /* ATOM accessor methods */
709 * ATOM is an interpreted byte code stored in tables in the vbios. The
710 * driver registers callbacks to access registers and the interpreter
711 * in the driver parses the tables and executes then to program specific
712 * actions (set display modes, asic init, etc.). See amdgpu_atombios.c,
713 * atombios.h, and atom.c
717 * cail_pll_read - read PLL register
719 * @info: atom card_info pointer
720 * @reg: PLL register offset
722 * Provides a PLL register accessor for the atom interpreter (r4xx+).
723 * Returns the value of the PLL register.
725 static uint32_t cail_pll_read(struct card_info *info, uint32_t reg)
731 * cail_pll_write - write PLL register
733 * @info: atom card_info pointer
734 * @reg: PLL register offset
735 * @val: value to write to the pll register
737 * Provides a PLL register accessor for the atom interpreter (r4xx+).
739 static void cail_pll_write(struct card_info *info, uint32_t reg, uint32_t val)
745 * cail_mc_read - read MC (Memory Controller) register
747 * @info: atom card_info pointer
748 * @reg: MC register offset
750 * Provides an MC register accessor for the atom interpreter (r4xx+).
751 * Returns the value of the MC register.
753 static uint32_t cail_mc_read(struct card_info *info, uint32_t reg)
759 * cail_mc_write - write MC (Memory Controller) register
761 * @info: atom card_info pointer
762 * @reg: MC register offset
763 * @val: value to write to the pll register
765 * Provides a MC register accessor for the atom interpreter (r4xx+).
767 static void cail_mc_write(struct card_info *info, uint32_t reg, uint32_t val)
773 * cail_reg_write - write MMIO register
775 * @info: atom card_info pointer
776 * @reg: MMIO register offset
777 * @val: value to write to the pll register
779 * Provides a MMIO register accessor for the atom interpreter (r4xx+).
781 static void cail_reg_write(struct card_info *info, uint32_t reg, uint32_t val)
783 struct amdgpu_device *adev = info->dev->dev_private;
789 * cail_reg_read - read MMIO register
791 * @info: atom card_info pointer
792 * @reg: MMIO register offset
794 * Provides an MMIO register accessor for the atom interpreter (r4xx+).
795 * Returns the value of the MMIO register.
797 static uint32_t cail_reg_read(struct card_info *info, uint32_t reg)
799 struct amdgpu_device *adev = info->dev->dev_private;
807 * cail_ioreg_write - write IO register
809 * @info: atom card_info pointer
810 * @reg: IO register offset
811 * @val: value to write to the pll register
813 * Provides a IO register accessor for the atom interpreter (r4xx+).
815 static void cail_ioreg_write(struct card_info *info, uint32_t reg, uint32_t val)
817 struct amdgpu_device *adev = info->dev->dev_private;
823 * cail_ioreg_read - read IO register
825 * @info: atom card_info pointer
826 * @reg: IO register offset
828 * Provides an IO register accessor for the atom interpreter (r4xx+).
829 * Returns the value of the IO register.
831 static uint32_t cail_ioreg_read(struct card_info *info, uint32_t reg)
833 struct amdgpu_device *adev = info->dev->dev_private;
841 * amdgpu_atombios_fini - free the driver info and callbacks for atombios
843 * @adev: amdgpu_device pointer
845 * Frees the driver info and register access callbacks for the ATOM
846 * interpreter (r4xx+).
847 * Called at driver shutdown.
849 static void amdgpu_atombios_fini(struct amdgpu_device *adev)
851 if (adev->mode_info.atom_context) {
852 kfree(adev->mode_info.atom_context->scratch);
853 kfree(adev->mode_info.atom_context->iio);
855 kfree(adev->mode_info.atom_context);
856 adev->mode_info.atom_context = NULL;
857 kfree(adev->mode_info.atom_card_info);
858 adev->mode_info.atom_card_info = NULL;
862 * amdgpu_atombios_init - init the driver info and callbacks for atombios
864 * @adev: amdgpu_device pointer
866 * Initializes the driver info and register access callbacks for the
867 * ATOM interpreter (r4xx+).
868 * Returns 0 on sucess, -ENOMEM on failure.
869 * Called at driver startup.
871 static int amdgpu_atombios_init(struct amdgpu_device *adev)
873 struct card_info *atom_card_info =
874 kzalloc(sizeof(struct card_info), GFP_KERNEL);
879 adev->mode_info.atom_card_info = atom_card_info;
880 atom_card_info->dev = adev->ddev;
881 atom_card_info->reg_read = cail_reg_read;
882 atom_card_info->reg_write = cail_reg_write;
883 /* needed for iio ops */
885 atom_card_info->ioreg_read = cail_ioreg_read;
886 atom_card_info->ioreg_write = cail_ioreg_write;
888 DRM_ERROR("Unable to find PCI I/O BAR; using MMIO for ATOM IIO\n");
889 atom_card_info->ioreg_read = cail_reg_read;
890 atom_card_info->ioreg_write = cail_reg_write;
892 atom_card_info->mc_read = cail_mc_read;
893 atom_card_info->mc_write = cail_mc_write;
894 atom_card_info->pll_read = cail_pll_read;
895 atom_card_info->pll_write = cail_pll_write;
897 adev->mode_info.atom_context = amdgpu_atom_parse(atom_card_info, adev->bios);
898 if (!adev->mode_info.atom_context) {
899 amdgpu_atombios_fini(adev);
903 mutex_init(&adev->mode_info.atom_context->mutex);
904 amdgpu_atombios_scratch_regs_init(adev);
905 amdgpu_atom_allocate_fb_scratch(adev->mode_info.atom_context);
909 /* if we get transitioned to only one device, take VGA back */
911 * amdgpu_vga_set_decode - enable/disable vga decode
913 * @cookie: amdgpu_device pointer
914 * @state: enable/disable vga decode
916 * Enable/disable vga decode (all asics).
917 * Returns VGA resource flags.
919 static unsigned int amdgpu_vga_set_decode(void *cookie, bool state)
921 struct amdgpu_device *adev = cookie;
922 amdgpu_asic_set_vga_state(adev, state);
924 return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
925 VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
927 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
931 * amdgpu_check_pot_argument - check that argument is a power of two
933 * @arg: value to check
935 * Validates that a certain argument is a power of two (all asics).
936 * Returns true if argument is valid.
938 static bool amdgpu_check_pot_argument(int arg)
940 return (arg & (arg - 1)) == 0;
944 * amdgpu_check_arguments - validate module params
946 * @adev: amdgpu_device pointer
948 * Validates certain module parameters and updates
949 * the associated values used by the driver (all asics).
951 static void amdgpu_check_arguments(struct amdgpu_device *adev)
953 if (amdgpu_sched_jobs < 4) {
954 dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
956 amdgpu_sched_jobs = 4;
957 } else if (!amdgpu_check_pot_argument(amdgpu_sched_jobs)){
958 dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
960 amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
963 if (amdgpu_gart_size != -1) {
964 /* gtt size must be greater or equal to 32M */
965 if (amdgpu_gart_size < 32) {
966 dev_warn(adev->dev, "gart size (%d) too small\n",
968 amdgpu_gart_size = -1;
972 if (!amdgpu_check_pot_argument(amdgpu_vm_size)) {
973 dev_warn(adev->dev, "VM size (%d) must be a power of 2\n",
978 if (amdgpu_vm_size < 1) {
979 dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
985 * Max GPUVM size for Cayman, SI and CI are 40 bits.
987 if (amdgpu_vm_size > 1024) {
988 dev_warn(adev->dev, "VM size (%d) too large, max is 1TB\n",
993 /* defines number of bits in page table versus page directory,
994 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
995 * page table and the remaining bits are in the page directory */
996 if (amdgpu_vm_block_size == -1) {
998 /* Total bits covered by PD + PTs */
999 unsigned bits = ilog2(amdgpu_vm_size) + 18;
1001 /* Make sure the PD is 4K in size up to 8GB address space.
1002 Above that split equal between PD and PTs */
1003 if (amdgpu_vm_size <= 8)
1004 amdgpu_vm_block_size = bits - 9;
1006 amdgpu_vm_block_size = (bits + 3) / 2;
1008 } else if (amdgpu_vm_block_size < 9) {
1009 dev_warn(adev->dev, "VM page table size (%d) too small\n",
1010 amdgpu_vm_block_size);
1011 amdgpu_vm_block_size = 9;
1014 if (amdgpu_vm_block_size > 24 ||
1015 (amdgpu_vm_size * 1024) < (1ull << amdgpu_vm_block_size)) {
1016 dev_warn(adev->dev, "VM page table size (%d) too large\n",
1017 amdgpu_vm_block_size);
1018 amdgpu_vm_block_size = 9;
1021 if (amdgpu_vram_page_split != -1 && (amdgpu_vram_page_split < 16 ||
1022 !amdgpu_check_pot_argument(amdgpu_vram_page_split))) {
1023 dev_warn(adev->dev, "invalid VRAM page split (%d)\n",
1024 amdgpu_vram_page_split);
1025 amdgpu_vram_page_split = 1024;
1030 * amdgpu_switcheroo_set_state - set switcheroo state
1032 * @pdev: pci dev pointer
1033 * @state: vga_switcheroo state
1035 * Callback for the switcheroo driver. Suspends or resumes the
1036 * the asics before or after it is powered up using ACPI methods.
1038 static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, enum vga_switcheroo_state state)
1040 struct drm_device *dev = pci_get_drvdata(pdev);
1042 if (amdgpu_device_is_px(dev) && state == VGA_SWITCHEROO_OFF)
1045 if (state == VGA_SWITCHEROO_ON) {
1046 unsigned d3_delay = dev->pdev->d3_delay;
1048 printk(KERN_INFO "amdgpu: switched on\n");
1049 /* don't suspend or resume card normally */
1050 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1052 amdgpu_device_resume(dev, true, true);
1054 dev->pdev->d3_delay = d3_delay;
1056 dev->switch_power_state = DRM_SWITCH_POWER_ON;
1057 drm_kms_helper_poll_enable(dev);
1059 printk(KERN_INFO "amdgpu: switched off\n");
1060 drm_kms_helper_poll_disable(dev);
1061 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1062 amdgpu_device_suspend(dev, true, true);
1063 dev->switch_power_state = DRM_SWITCH_POWER_OFF;
1068 * amdgpu_switcheroo_can_switch - see if switcheroo state can change
1070 * @pdev: pci dev pointer
1072 * Callback for the switcheroo driver. Check of the switcheroo
1073 * state can be changed.
1074 * Returns true if the state can be changed, false if not.
1076 static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
1078 struct drm_device *dev = pci_get_drvdata(pdev);
1081 * FIXME: open_count is protected by drm_global_mutex but that would lead to
1082 * locking inversion with the driver load path. And the access here is
1083 * completely racy anyway. So don't bother with locking for now.
1085 return dev->open_count == 0;
1088 static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
1089 .set_gpu_state = amdgpu_switcheroo_set_state,
1091 .can_switch = amdgpu_switcheroo_can_switch,
1094 int amdgpu_set_clockgating_state(struct amdgpu_device *adev,
1095 enum amd_ip_block_type block_type,
1096 enum amd_clockgating_state state)
1100 for (i = 0; i < adev->num_ip_blocks; i++) {
1101 if (!adev->ip_blocks[i].status.valid)
1103 if (adev->ip_blocks[i].version->type == block_type) {
1104 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1114 int amdgpu_set_powergating_state(struct amdgpu_device *adev,
1115 enum amd_ip_block_type block_type,
1116 enum amd_powergating_state state)
1120 for (i = 0; i < adev->num_ip_blocks; i++) {
1121 if (!adev->ip_blocks[i].status.valid)
1123 if (adev->ip_blocks[i].version->type == block_type) {
1124 r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev,
1134 int amdgpu_wait_for_idle(struct amdgpu_device *adev,
1135 enum amd_ip_block_type block_type)
1139 for (i = 0; i < adev->num_ip_blocks; i++) {
1140 if (!adev->ip_blocks[i].status.valid)
1142 if (adev->ip_blocks[i].version->type == block_type) {
1143 r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
1153 bool amdgpu_is_idle(struct amdgpu_device *adev,
1154 enum amd_ip_block_type block_type)
1158 for (i = 0; i < adev->num_ip_blocks; i++) {
1159 if (!adev->ip_blocks[i].status.valid)
1161 if (adev->ip_blocks[i].version->type == block_type)
1162 return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
1168 struct amdgpu_ip_block * amdgpu_get_ip_block(struct amdgpu_device *adev,
1169 enum amd_ip_block_type type)
1173 for (i = 0; i < adev->num_ip_blocks; i++)
1174 if (adev->ip_blocks[i].version->type == type)
1175 return &adev->ip_blocks[i];
1181 * amdgpu_ip_block_version_cmp
1183 * @adev: amdgpu_device pointer
1184 * @type: enum amd_ip_block_type
1185 * @major: major version
1186 * @minor: minor version
1188 * return 0 if equal or greater
1189 * return 1 if smaller or the ip_block doesn't exist
1191 int amdgpu_ip_block_version_cmp(struct amdgpu_device *adev,
1192 enum amd_ip_block_type type,
1193 u32 major, u32 minor)
1195 struct amdgpu_ip_block *ip_block = amdgpu_get_ip_block(adev, type);
1197 if (ip_block && ((ip_block->version->major > major) ||
1198 ((ip_block->version->major == major) &&
1199 (ip_block->version->minor >= minor))))
1206 * amdgpu_ip_block_add
1208 * @adev: amdgpu_device pointer
1209 * @ip_block_version: pointer to the IP to add
1211 * Adds the IP block driver information to the collection of IPs
1214 int amdgpu_ip_block_add(struct amdgpu_device *adev,
1215 const struct amdgpu_ip_block_version *ip_block_version)
1217 if (!ip_block_version)
1220 adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
1225 static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
1227 adev->enable_virtual_display = false;
1229 if (amdgpu_virtual_display) {
1230 struct drm_device *ddev = adev->ddev;
1231 const char *pci_address_name = pci_name(ddev->pdev);
1232 char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
1234 pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
1235 pciaddstr_tmp = pciaddstr;
1236 while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
1237 pciaddname = strsep(&pciaddname_tmp, ",");
1238 if (!strcmp(pci_address_name, pciaddname)) {
1242 adev->enable_virtual_display = true;
1245 res = kstrtol(pciaddname_tmp, 10,
1253 adev->mode_info.num_crtc = num_crtc;
1255 adev->mode_info.num_crtc = 1;
1261 DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
1262 amdgpu_virtual_display, pci_address_name,
1263 adev->enable_virtual_display, adev->mode_info.num_crtc);
1269 static int amdgpu_early_init(struct amdgpu_device *adev)
1273 amdgpu_device_enable_virtual_display(adev);
1275 switch (adev->asic_type) {
1279 case CHIP_POLARIS11:
1280 case CHIP_POLARIS10:
1281 case CHIP_POLARIS12:
1284 if (adev->asic_type == CHIP_CARRIZO || adev->asic_type == CHIP_STONEY)
1285 adev->family = AMDGPU_FAMILY_CZ;
1287 adev->family = AMDGPU_FAMILY_VI;
1289 r = vi_set_ip_blocks(adev);
1293 #ifdef CONFIG_DRM_AMDGPU_SI
1299 adev->family = AMDGPU_FAMILY_SI;
1300 r = si_set_ip_blocks(adev);
1305 #ifdef CONFIG_DRM_AMDGPU_CIK
1311 if ((adev->asic_type == CHIP_BONAIRE) || (adev->asic_type == CHIP_HAWAII))
1312 adev->family = AMDGPU_FAMILY_CI;
1314 adev->family = AMDGPU_FAMILY_KV;
1316 r = cik_set_ip_blocks(adev);
1322 /* FIXME: not supported yet */
1326 for (i = 0; i < adev->num_ip_blocks; i++) {
1327 if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
1328 DRM_ERROR("disabled ip block: %d\n", i);
1329 adev->ip_blocks[i].status.valid = false;
1331 if (adev->ip_blocks[i].version->funcs->early_init) {
1332 r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
1334 adev->ip_blocks[i].status.valid = false;
1336 DRM_ERROR("early_init of IP block <%s> failed %d\n",
1337 adev->ip_blocks[i].version->funcs->name, r);
1340 adev->ip_blocks[i].status.valid = true;
1343 adev->ip_blocks[i].status.valid = true;
1348 adev->cg_flags &= amdgpu_cg_mask;
1349 adev->pg_flags &= amdgpu_pg_mask;
1354 static int amdgpu_init(struct amdgpu_device *adev)
1358 for (i = 0; i < adev->num_ip_blocks; i++) {
1359 if (!adev->ip_blocks[i].status.valid)
1361 r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
1363 DRM_ERROR("sw_init of IP block <%s> failed %d\n",
1364 adev->ip_blocks[i].version->funcs->name, r);
1367 adev->ip_blocks[i].status.sw = true;
1368 /* need to do gmc hw init early so we can allocate gpu mem */
1369 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1370 r = amdgpu_vram_scratch_init(adev);
1372 DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r);
1375 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1377 DRM_ERROR("hw_init %d failed %d\n", i, r);
1380 r = amdgpu_wb_init(adev);
1382 DRM_ERROR("amdgpu_wb_init failed %d\n", r);
1385 adev->ip_blocks[i].status.hw = true;
1389 for (i = 0; i < adev->num_ip_blocks; i++) {
1390 if (!adev->ip_blocks[i].status.sw)
1392 /* gmc hw init is done early */
1393 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC)
1395 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1397 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1398 adev->ip_blocks[i].version->funcs->name, r);
1401 adev->ip_blocks[i].status.hw = true;
1407 static int amdgpu_late_init(struct amdgpu_device *adev)
1411 for (i = 0; i < adev->num_ip_blocks; i++) {
1412 if (!adev->ip_blocks[i].status.valid)
1414 if (adev->ip_blocks[i].version->funcs->late_init) {
1415 r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
1417 DRM_ERROR("late_init of IP block <%s> failed %d\n",
1418 adev->ip_blocks[i].version->funcs->name, r);
1421 adev->ip_blocks[i].status.late_initialized = true;
1423 /* skip CG for VCE/UVD, it's handled specially */
1424 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1425 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE) {
1426 /* enable clockgating to save power */
1427 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1430 DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
1431 adev->ip_blocks[i].version->funcs->name, r);
1440 static int amdgpu_fini(struct amdgpu_device *adev)
1444 /* need to disable SMC first */
1445 for (i = 0; i < adev->num_ip_blocks; i++) {
1446 if (!adev->ip_blocks[i].status.hw)
1448 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
1449 /* ungate blocks before hw fini so that we can shutdown the blocks safely */
1450 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1451 AMD_CG_STATE_UNGATE);
1453 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1454 adev->ip_blocks[i].version->funcs->name, r);
1457 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
1458 /* XXX handle errors */
1460 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
1461 adev->ip_blocks[i].version->funcs->name, r);
1463 adev->ip_blocks[i].status.hw = false;
1468 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1469 if (!adev->ip_blocks[i].status.hw)
1471 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1472 amdgpu_wb_fini(adev);
1473 amdgpu_vram_scratch_fini(adev);
1476 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1477 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE) {
1478 /* ungate blocks before hw fini so that we can shutdown the blocks safely */
1479 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1480 AMD_CG_STATE_UNGATE);
1482 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1483 adev->ip_blocks[i].version->funcs->name, r);
1488 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
1489 /* XXX handle errors */
1491 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
1492 adev->ip_blocks[i].version->funcs->name, r);
1495 adev->ip_blocks[i].status.hw = false;
1498 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1499 if (!adev->ip_blocks[i].status.sw)
1501 r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
1502 /* XXX handle errors */
1504 DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
1505 adev->ip_blocks[i].version->funcs->name, r);
1507 adev->ip_blocks[i].status.sw = false;
1508 adev->ip_blocks[i].status.valid = false;
1511 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1512 if (!adev->ip_blocks[i].status.late_initialized)
1514 if (adev->ip_blocks[i].version->funcs->late_fini)
1515 adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
1516 adev->ip_blocks[i].status.late_initialized = false;
1522 int amdgpu_suspend(struct amdgpu_device *adev)
1526 /* ungate SMC block first */
1527 r = amdgpu_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_SMC,
1528 AMD_CG_STATE_UNGATE);
1530 DRM_ERROR("set_clockgating_state(ungate) SMC failed %d\n",r);
1533 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1534 if (!adev->ip_blocks[i].status.valid)
1536 /* ungate blocks so that suspend can properly shut them down */
1537 if (i != AMD_IP_BLOCK_TYPE_SMC) {
1538 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1539 AMD_CG_STATE_UNGATE);
1541 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1542 adev->ip_blocks[i].version->funcs->name, r);
1545 /* XXX handle errors */
1546 r = adev->ip_blocks[i].version->funcs->suspend(adev);
1547 /* XXX handle errors */
1549 DRM_ERROR("suspend of IP block <%s> failed %d\n",
1550 adev->ip_blocks[i].version->funcs->name, r);
1557 static int amdgpu_resume(struct amdgpu_device *adev)
1561 for (i = 0; i < adev->num_ip_blocks; i++) {
1562 if (!adev->ip_blocks[i].status.valid)
1564 r = adev->ip_blocks[i].version->funcs->resume(adev);
1566 DRM_ERROR("resume of IP block <%s> failed %d\n",
1567 adev->ip_blocks[i].version->funcs->name, r);
1575 static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
1577 if (amdgpu_atombios_has_gpu_virtualization_table(adev))
1578 adev->virtualization.virtual_caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
1582 * amdgpu_device_init - initialize the driver
1584 * @adev: amdgpu_device pointer
1585 * @pdev: drm dev pointer
1586 * @pdev: pci dev pointer
1587 * @flags: driver flags
1589 * Initializes the driver info and hw (all asics).
1590 * Returns 0 for success or an error on failure.
1591 * Called at driver startup.
1593 int amdgpu_device_init(struct amdgpu_device *adev,
1594 struct drm_device *ddev,
1595 struct pci_dev *pdev,
1599 bool runtime = false;
1602 adev->shutdown = false;
1603 adev->dev = &pdev->dev;
1606 adev->flags = flags;
1607 adev->asic_type = flags & AMD_ASIC_MASK;
1608 adev->is_atom_bios = false;
1609 adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
1610 adev->mc.gtt_size = 512 * 1024 * 1024;
1611 adev->accel_working = false;
1612 adev->num_rings = 0;
1613 adev->mman.buffer_funcs = NULL;
1614 adev->mman.buffer_funcs_ring = NULL;
1615 adev->vm_manager.vm_pte_funcs = NULL;
1616 adev->vm_manager.vm_pte_num_rings = 0;
1617 adev->gart.gart_funcs = NULL;
1618 adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
1620 adev->smc_rreg = &amdgpu_invalid_rreg;
1621 adev->smc_wreg = &amdgpu_invalid_wreg;
1622 adev->pcie_rreg = &amdgpu_invalid_rreg;
1623 adev->pcie_wreg = &amdgpu_invalid_wreg;
1624 adev->pciep_rreg = &amdgpu_invalid_rreg;
1625 adev->pciep_wreg = &amdgpu_invalid_wreg;
1626 adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
1627 adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
1628 adev->didt_rreg = &amdgpu_invalid_rreg;
1629 adev->didt_wreg = &amdgpu_invalid_wreg;
1630 adev->gc_cac_rreg = &amdgpu_invalid_rreg;
1631 adev->gc_cac_wreg = &amdgpu_invalid_wreg;
1632 adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
1633 adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
1636 DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
1637 amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
1638 pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
1640 /* mutex initialization are all done here so we
1641 * can recall function without having locking issues */
1642 mutex_init(&adev->vm_manager.lock);
1643 atomic_set(&adev->irq.ih.lock, 0);
1644 mutex_init(&adev->pm.mutex);
1645 mutex_init(&adev->gfx.gpu_clock_mutex);
1646 mutex_init(&adev->srbm_mutex);
1647 mutex_init(&adev->grbm_idx_mutex);
1648 mutex_init(&adev->mn_lock);
1649 hash_init(adev->mn_hash);
1651 amdgpu_check_arguments(adev);
1653 /* Registers mapping */
1654 /* TODO: block userspace mapping of io register */
1655 spin_lock_init(&adev->mmio_idx_lock);
1656 spin_lock_init(&adev->smc_idx_lock);
1657 spin_lock_init(&adev->pcie_idx_lock);
1658 spin_lock_init(&adev->uvd_ctx_idx_lock);
1659 spin_lock_init(&adev->didt_idx_lock);
1660 spin_lock_init(&adev->gc_cac_idx_lock);
1661 spin_lock_init(&adev->audio_endpt_idx_lock);
1662 spin_lock_init(&adev->mm_stats.lock);
1664 INIT_LIST_HEAD(&adev->shadow_list);
1665 mutex_init(&adev->shadow_list_lock);
1667 INIT_LIST_HEAD(&adev->gtt_list);
1668 spin_lock_init(&adev->gtt_list_lock);
1670 if (adev->asic_type >= CHIP_BONAIRE) {
1671 adev->rmmio_base = pci_resource_start(adev->pdev, 5);
1672 adev->rmmio_size = pci_resource_len(adev->pdev, 5);
1674 adev->rmmio_base = pci_resource_start(adev->pdev, 2);
1675 adev->rmmio_size = pci_resource_len(adev->pdev, 2);
1678 adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
1679 if (adev->rmmio == NULL) {
1682 DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
1683 DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
1685 if (adev->asic_type >= CHIP_BONAIRE)
1686 /* doorbell bar mapping */
1687 amdgpu_doorbell_init(adev);
1689 /* io port mapping */
1690 for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
1691 if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) {
1692 adev->rio_mem_size = pci_resource_len(adev->pdev, i);
1693 adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size);
1697 if (adev->rio_mem == NULL)
1698 DRM_ERROR("Unable to find PCI I/O BAR\n");
1700 /* early init functions */
1701 r = amdgpu_early_init(adev);
1705 /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
1706 /* this will fail for cards that aren't VGA class devices, just
1708 vga_client_register(adev->pdev, adev, NULL, amdgpu_vga_set_decode);
1710 if (amdgpu_runtime_pm == 1)
1712 if (amdgpu_device_is_px(ddev))
1714 vga_switcheroo_register_client(adev->pdev, &amdgpu_switcheroo_ops, runtime);
1716 vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
1719 if (!amdgpu_get_bios(adev)) {
1723 /* Must be an ATOMBIOS */
1724 if (!adev->is_atom_bios) {
1725 dev_err(adev->dev, "Expecting atombios for GPU\n");
1729 r = amdgpu_atombios_init(adev);
1731 dev_err(adev->dev, "amdgpu_atombios_init failed\n");
1735 /* detect if we are with an SRIOV vbios */
1736 amdgpu_device_detect_sriov_bios(adev);
1738 /* Post card if necessary */
1739 if (amdgpu_vpost_needed(adev)) {
1741 dev_err(adev->dev, "no vBIOS found\n");
1745 DRM_INFO("GPU posting now...\n");
1746 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
1748 dev_err(adev->dev, "gpu post error!\n");
1752 DRM_INFO("GPU post is not needed\n");
1755 /* Initialize clocks */
1756 r = amdgpu_atombios_get_clock_info(adev);
1758 dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
1761 /* init i2c buses */
1762 amdgpu_atombios_i2c_init(adev);
1765 r = amdgpu_fence_driver_init(adev);
1767 dev_err(adev->dev, "amdgpu_fence_driver_init failed\n");
1771 /* init the mode config */
1772 drm_mode_config_init(adev->ddev);
1774 r = amdgpu_init(adev);
1776 dev_err(adev->dev, "amdgpu_init failed\n");
1781 adev->accel_working = true;
1783 /* Initialize the buffer migration limit. */
1784 if (amdgpu_moverate >= 0)
1785 max_MBps = amdgpu_moverate;
1787 max_MBps = 8; /* Allow 8 MB/s. */
1788 /* Get a log2 for easy divisions. */
1789 adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
1791 amdgpu_fbdev_init(adev);
1793 r = amdgpu_ib_pool_init(adev);
1795 dev_err(adev->dev, "IB initialization failed (%d).\n", r);
1799 r = amdgpu_ib_ring_tests(adev);
1801 DRM_ERROR("ib ring test failed (%d).\n", r);
1803 r = amdgpu_gem_debugfs_init(adev);
1805 DRM_ERROR("registering gem debugfs failed (%d).\n", r);
1808 r = amdgpu_debugfs_regs_init(adev);
1810 DRM_ERROR("registering register debugfs failed (%d).\n", r);
1813 r = amdgpu_debugfs_firmware_init(adev);
1815 DRM_ERROR("registering firmware debugfs failed (%d).\n", r);
1819 if ((amdgpu_testing & 1)) {
1820 if (adev->accel_working)
1821 amdgpu_test_moves(adev);
1823 DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n");
1825 if ((amdgpu_testing & 2)) {
1826 if (adev->accel_working)
1827 amdgpu_test_syncing(adev);
1829 DRM_INFO("amdgpu: acceleration disabled, skipping sync tests\n");
1831 if (amdgpu_benchmarking) {
1832 if (adev->accel_working)
1833 amdgpu_benchmark(adev, amdgpu_benchmarking);
1835 DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n");
1838 /* enable clockgating, etc. after ib tests, etc. since some blocks require
1839 * explicit gating rather than handling it automatically.
1841 r = amdgpu_late_init(adev);
1843 dev_err(adev->dev, "amdgpu_late_init failed\n");
1851 vga_switcheroo_fini_domain_pm_ops(adev->dev);
1855 static void amdgpu_debugfs_remove_files(struct amdgpu_device *adev);
1858 * amdgpu_device_fini - tear down the driver
1860 * @adev: amdgpu_device pointer
1862 * Tear down the driver info (all asics).
1863 * Called at driver shutdown.
1865 void amdgpu_device_fini(struct amdgpu_device *adev)
1869 DRM_INFO("amdgpu: finishing device.\n");
1870 adev->shutdown = true;
1871 drm_crtc_force_disable_all(adev->ddev);
1872 /* evict vram memory */
1873 amdgpu_bo_evict_vram(adev);
1874 amdgpu_ib_pool_fini(adev);
1875 amdgpu_fence_driver_fini(adev);
1876 amdgpu_fbdev_fini(adev);
1877 r = amdgpu_fini(adev);
1878 adev->accel_working = false;
1879 /* free i2c buses */
1880 amdgpu_i2c_fini(adev);
1881 amdgpu_atombios_fini(adev);
1884 vga_switcheroo_unregister_client(adev->pdev);
1885 if (adev->flags & AMD_IS_PX)
1886 vga_switcheroo_fini_domain_pm_ops(adev->dev);
1887 vga_client_register(adev->pdev, NULL, NULL, NULL);
1889 pci_iounmap(adev->pdev, adev->rio_mem);
1890 adev->rio_mem = NULL;
1891 iounmap(adev->rmmio);
1893 if (adev->asic_type >= CHIP_BONAIRE)
1894 amdgpu_doorbell_fini(adev);
1895 amdgpu_debugfs_regs_cleanup(adev);
1896 amdgpu_debugfs_remove_files(adev);
1904 * amdgpu_device_suspend - initiate device suspend
1906 * @pdev: drm dev pointer
1907 * @state: suspend state
1909 * Puts the hw in the suspend state (all asics).
1910 * Returns 0 for success or an error on failure.
1911 * Called at driver suspend.
1913 int amdgpu_device_suspend(struct drm_device *dev, bool suspend, bool fbcon)
1915 struct amdgpu_device *adev;
1916 struct drm_crtc *crtc;
1917 struct drm_connector *connector;
1920 if (dev == NULL || dev->dev_private == NULL) {
1924 adev = dev->dev_private;
1926 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
1929 drm_kms_helper_poll_disable(dev);
1931 /* turn off display hw */
1932 drm_modeset_lock_all(dev);
1933 list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
1934 drm_helper_connector_dpms(connector, DRM_MODE_DPMS_OFF);
1936 drm_modeset_unlock_all(dev);
1938 /* unpin the front buffers and cursors */
1939 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
1940 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
1941 struct amdgpu_framebuffer *rfb = to_amdgpu_framebuffer(crtc->primary->fb);
1942 struct amdgpu_bo *robj;
1944 if (amdgpu_crtc->cursor_bo) {
1945 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
1946 r = amdgpu_bo_reserve(aobj, false);
1948 amdgpu_bo_unpin(aobj);
1949 amdgpu_bo_unreserve(aobj);
1953 if (rfb == NULL || rfb->obj == NULL) {
1956 robj = gem_to_amdgpu_bo(rfb->obj);
1957 /* don't unpin kernel fb objects */
1958 if (!amdgpu_fbdev_robj_is_fb(adev, robj)) {
1959 r = amdgpu_bo_reserve(robj, false);
1961 amdgpu_bo_unpin(robj);
1962 amdgpu_bo_unreserve(robj);
1966 /* evict vram memory */
1967 amdgpu_bo_evict_vram(adev);
1969 amdgpu_fence_driver_suspend(adev);
1971 r = amdgpu_suspend(adev);
1973 /* evict remaining vram memory
1974 * This second call to evict vram is to evict the gart page table
1977 amdgpu_bo_evict_vram(adev);
1979 amdgpu_atombios_scratch_regs_save(adev);
1980 pci_save_state(dev->pdev);
1982 /* Shut down the device */
1983 pci_disable_device(dev->pdev);
1984 pci_set_power_state(dev->pdev, PCI_D3hot);
1986 r = amdgpu_asic_reset(adev);
1988 DRM_ERROR("amdgpu asic reset failed\n");
1993 amdgpu_fbdev_set_suspend(adev, 1);
2000 * amdgpu_device_resume - initiate device resume
2002 * @pdev: drm dev pointer
2004 * Bring the hw back to operating state (all asics).
2005 * Returns 0 for success or an error on failure.
2006 * Called at driver resume.
2008 int amdgpu_device_resume(struct drm_device *dev, bool resume, bool fbcon)
2010 struct drm_connector *connector;
2011 struct amdgpu_device *adev = dev->dev_private;
2012 struct drm_crtc *crtc;
2015 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
2022 pci_set_power_state(dev->pdev, PCI_D0);
2023 pci_restore_state(dev->pdev);
2024 r = pci_enable_device(dev->pdev);
2031 amdgpu_atombios_scratch_regs_restore(adev);
2034 if (!amdgpu_card_posted(adev) || !resume) {
2035 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2037 DRM_ERROR("amdgpu asic init failed\n");
2040 r = amdgpu_resume(adev);
2042 DRM_ERROR("amdgpu_resume failed (%d).\n", r);
2044 amdgpu_fence_driver_resume(adev);
2047 r = amdgpu_ib_ring_tests(adev);
2049 DRM_ERROR("ib ring test failed (%d).\n", r);
2052 r = amdgpu_late_init(adev);
2057 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2058 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2060 if (amdgpu_crtc->cursor_bo) {
2061 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2062 r = amdgpu_bo_reserve(aobj, false);
2064 r = amdgpu_bo_pin(aobj,
2065 AMDGPU_GEM_DOMAIN_VRAM,
2066 &amdgpu_crtc->cursor_addr);
2068 DRM_ERROR("Failed to pin cursor BO (%d)\n", r);
2069 amdgpu_bo_unreserve(aobj);
2074 /* blat the mode back in */
2076 drm_helper_resume_force_mode(dev);
2077 /* turn on display hw */
2078 drm_modeset_lock_all(dev);
2079 list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2080 drm_helper_connector_dpms(connector, DRM_MODE_DPMS_ON);
2082 drm_modeset_unlock_all(dev);
2085 drm_kms_helper_poll_enable(dev);
2088 * Most of the connector probing functions try to acquire runtime pm
2089 * refs to ensure that the GPU is powered on when connector polling is
2090 * performed. Since we're calling this from a runtime PM callback,
2091 * trying to acquire rpm refs will cause us to deadlock.
2093 * Since we're guaranteed to be holding the rpm lock, it's safe to
2094 * temporarily disable the rpm helpers so this doesn't deadlock us.
2097 dev->dev->power.disable_depth++;
2099 drm_helper_hpd_irq_event(dev);
2101 dev->dev->power.disable_depth--;
2105 amdgpu_fbdev_set_suspend(adev, 0);
2112 static bool amdgpu_check_soft_reset(struct amdgpu_device *adev)
2115 bool asic_hang = false;
2117 for (i = 0; i < adev->num_ip_blocks; i++) {
2118 if (!adev->ip_blocks[i].status.valid)
2120 if (adev->ip_blocks[i].version->funcs->check_soft_reset)
2121 adev->ip_blocks[i].status.hang =
2122 adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
2123 if (adev->ip_blocks[i].status.hang) {
2124 DRM_INFO("IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
2131 static int amdgpu_pre_soft_reset(struct amdgpu_device *adev)
2135 for (i = 0; i < adev->num_ip_blocks; i++) {
2136 if (!adev->ip_blocks[i].status.valid)
2138 if (adev->ip_blocks[i].status.hang &&
2139 adev->ip_blocks[i].version->funcs->pre_soft_reset) {
2140 r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
2149 static bool amdgpu_need_full_reset(struct amdgpu_device *adev)
2153 for (i = 0; i < adev->num_ip_blocks; i++) {
2154 if (!adev->ip_blocks[i].status.valid)
2156 if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
2157 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
2158 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
2159 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE)) {
2160 if (adev->ip_blocks[i].status.hang) {
2161 DRM_INFO("Some block need full reset!\n");
2169 static int amdgpu_soft_reset(struct amdgpu_device *adev)
2173 for (i = 0; i < adev->num_ip_blocks; i++) {
2174 if (!adev->ip_blocks[i].status.valid)
2176 if (adev->ip_blocks[i].status.hang &&
2177 adev->ip_blocks[i].version->funcs->soft_reset) {
2178 r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
2187 static int amdgpu_post_soft_reset(struct amdgpu_device *adev)
2191 for (i = 0; i < adev->num_ip_blocks; i++) {
2192 if (!adev->ip_blocks[i].status.valid)
2194 if (adev->ip_blocks[i].status.hang &&
2195 adev->ip_blocks[i].version->funcs->post_soft_reset)
2196 r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
2204 bool amdgpu_need_backup(struct amdgpu_device *adev)
2206 if (adev->flags & AMD_IS_APU)
2209 return amdgpu_lockup_timeout > 0 ? true : false;
2212 static int amdgpu_recover_vram_from_shadow(struct amdgpu_device *adev,
2213 struct amdgpu_ring *ring,
2214 struct amdgpu_bo *bo,
2215 struct dma_fence **fence)
2223 r = amdgpu_bo_reserve(bo, false);
2226 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
2227 /* if bo has been evicted, then no need to recover */
2228 if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
2229 r = amdgpu_bo_restore_from_shadow(adev, ring, bo,
2232 DRM_ERROR("recover page table failed!\n");
2237 amdgpu_bo_unreserve(bo);
2242 * amdgpu_gpu_reset - reset the asic
2244 * @adev: amdgpu device pointer
2246 * Attempt the reset the GPU if it has hung (all asics).
2247 * Returns 0 for success or an error on failure.
2249 int amdgpu_gpu_reset(struct amdgpu_device *adev)
2253 bool need_full_reset;
2255 if (!amdgpu_check_soft_reset(adev)) {
2256 DRM_INFO("No hardware hang detected. Did some blocks stall?\n");
2260 atomic_inc(&adev->gpu_reset_counter);
2263 resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
2265 /* block scheduler */
2266 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
2267 struct amdgpu_ring *ring = adev->rings[i];
2271 kthread_park(ring->sched.thread);
2272 amd_sched_hw_job_reset(&ring->sched);
2274 /* after all hw jobs are reset, hw fence is meaningless, so force_completion */
2275 amdgpu_fence_driver_force_completion(adev);
2277 need_full_reset = amdgpu_need_full_reset(adev);
2279 if (!need_full_reset) {
2280 amdgpu_pre_soft_reset(adev);
2281 r = amdgpu_soft_reset(adev);
2282 amdgpu_post_soft_reset(adev);
2283 if (r || amdgpu_check_soft_reset(adev)) {
2284 DRM_INFO("soft reset failed, will fallback to full reset!\n");
2285 need_full_reset = true;
2289 if (need_full_reset) {
2290 r = amdgpu_suspend(adev);
2293 /* Disable fb access */
2294 if (adev->mode_info.num_crtc) {
2295 struct amdgpu_mode_mc_save save;
2296 amdgpu_display_stop_mc_access(adev, &save);
2297 amdgpu_wait_for_idle(adev, AMD_IP_BLOCK_TYPE_GMC);
2299 amdgpu_atombios_scratch_regs_save(adev);
2300 r = amdgpu_asic_reset(adev);
2301 amdgpu_atombios_scratch_regs_restore(adev);
2303 amdgpu_atom_asic_init(adev->mode_info.atom_context);
2306 dev_info(adev->dev, "GPU reset succeeded, trying to resume\n");
2307 r = amdgpu_resume(adev);
2311 amdgpu_irq_gpu_reset_resume_helper(adev);
2312 if (need_full_reset && amdgpu_need_backup(adev)) {
2313 r = amdgpu_ttm_recover_gart(adev);
2315 DRM_ERROR("gart recovery failed!!!\n");
2317 r = amdgpu_ib_ring_tests(adev);
2319 dev_err(adev->dev, "ib ring test failed (%d).\n", r);
2320 r = amdgpu_suspend(adev);
2321 need_full_reset = true;
2325 * recovery vm page tables, since we cannot depend on VRAM is
2326 * consistent after gpu full reset.
2328 if (need_full_reset && amdgpu_need_backup(adev)) {
2329 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
2330 struct amdgpu_bo *bo, *tmp;
2331 struct dma_fence *fence = NULL, *next = NULL;
2333 DRM_INFO("recover vram bo from shadow\n");
2334 mutex_lock(&adev->shadow_list_lock);
2335 list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) {
2336 amdgpu_recover_vram_from_shadow(adev, ring, bo, &next);
2338 r = dma_fence_wait(fence, false);
2340 WARN(r, "recovery from shadow isn't comleted\n");
2345 dma_fence_put(fence);
2348 mutex_unlock(&adev->shadow_list_lock);
2350 r = dma_fence_wait(fence, false);
2352 WARN(r, "recovery from shadow isn't comleted\n");
2354 dma_fence_put(fence);
2356 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
2357 struct amdgpu_ring *ring = adev->rings[i];
2361 amd_sched_job_recovery(&ring->sched);
2362 kthread_unpark(ring->sched.thread);
2365 dev_err(adev->dev, "asic resume failed (%d).\n", r);
2366 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
2367 if (adev->rings[i]) {
2368 kthread_unpark(adev->rings[i]->sched.thread);
2373 drm_helper_resume_force_mode(adev->ddev);
2375 ttm_bo_unlock_delayed_workqueue(&adev->mman.bdev, resched);
2377 /* bad news, how to tell it to userspace ? */
2378 dev_info(adev->dev, "GPU reset failed\n");
2384 void amdgpu_get_pcie_info(struct amdgpu_device *adev)
2389 if (amdgpu_pcie_gen_cap)
2390 adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
2392 if (amdgpu_pcie_lane_cap)
2393 adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
2395 /* covers APUs as well */
2396 if (pci_is_root_bus(adev->pdev->bus)) {
2397 if (adev->pm.pcie_gen_mask == 0)
2398 adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
2399 if (adev->pm.pcie_mlw_mask == 0)
2400 adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
2404 if (adev->pm.pcie_gen_mask == 0) {
2405 ret = drm_pcie_get_speed_cap_mask(adev->ddev, &mask);
2407 adev->pm.pcie_gen_mask = (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
2408 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
2409 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
2411 if (mask & DRM_PCIE_SPEED_25)
2412 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
2413 if (mask & DRM_PCIE_SPEED_50)
2414 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2;
2415 if (mask & DRM_PCIE_SPEED_80)
2416 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3;
2418 adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
2421 if (adev->pm.pcie_mlw_mask == 0) {
2422 ret = drm_pcie_get_max_link_width(adev->ddev, &mask);
2426 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
2427 CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
2428 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
2429 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
2430 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
2431 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2432 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2435 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
2436 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
2437 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
2438 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
2439 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2440 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2443 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
2444 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
2445 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
2446 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2447 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2450 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
2451 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
2452 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2453 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2456 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
2457 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2458 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2461 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2462 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2465 adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;
2471 adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
2479 int amdgpu_debugfs_add_files(struct amdgpu_device *adev,
2480 const struct drm_info_list *files,
2485 for (i = 0; i < adev->debugfs_count; i++) {
2486 if (adev->debugfs[i].files == files) {
2487 /* Already registered */
2492 i = adev->debugfs_count + 1;
2493 if (i > AMDGPU_DEBUGFS_MAX_COMPONENTS) {
2494 DRM_ERROR("Reached maximum number of debugfs components.\n");
2495 DRM_ERROR("Report so we increase "
2496 "AMDGPU_DEBUGFS_MAX_COMPONENTS.\n");
2499 adev->debugfs[adev->debugfs_count].files = files;
2500 adev->debugfs[adev->debugfs_count].num_files = nfiles;
2501 adev->debugfs_count = i;
2502 #if defined(CONFIG_DEBUG_FS)
2503 drm_debugfs_create_files(files, nfiles,
2504 adev->ddev->primary->debugfs_root,
2505 adev->ddev->primary);
2510 static void amdgpu_debugfs_remove_files(struct amdgpu_device *adev)
2512 #if defined(CONFIG_DEBUG_FS)
2515 for (i = 0; i < adev->debugfs_count; i++) {
2516 drm_debugfs_remove_files(adev->debugfs[i].files,
2517 adev->debugfs[i].num_files,
2518 adev->ddev->primary);
2523 #if defined(CONFIG_DEBUG_FS)
2525 static ssize_t amdgpu_debugfs_regs_read(struct file *f, char __user *buf,
2526 size_t size, loff_t *pos)
2528 struct amdgpu_device *adev = file_inode(f)->i_private;
2531 bool pm_pg_lock, use_bank;
2532 unsigned instance_bank, sh_bank, se_bank;
2534 if (size & 0x3 || *pos & 0x3)
2537 /* are we reading registers for which a PG lock is necessary? */
2538 pm_pg_lock = (*pos >> 23) & 1;
2540 if (*pos & (1ULL << 62)) {
2541 se_bank = (*pos >> 24) & 0x3FF;
2542 sh_bank = (*pos >> 34) & 0x3FF;
2543 instance_bank = (*pos >> 44) & 0x3FF;
2545 if (se_bank == 0x3FF)
2546 se_bank = 0xFFFFFFFF;
2547 if (sh_bank == 0x3FF)
2548 sh_bank = 0xFFFFFFFF;
2549 if (instance_bank == 0x3FF)
2550 instance_bank = 0xFFFFFFFF;
2559 if ((sh_bank != 0xFFFFFFFF && sh_bank >= adev->gfx.config.max_sh_per_se) ||
2560 (se_bank != 0xFFFFFFFF && se_bank >= adev->gfx.config.max_shader_engines))
2562 mutex_lock(&adev->grbm_idx_mutex);
2563 amdgpu_gfx_select_se_sh(adev, se_bank,
2564 sh_bank, instance_bank);
2568 mutex_lock(&adev->pm.mutex);
2573 if (*pos > adev->rmmio_size)
2576 value = RREG32(*pos >> 2);
2577 r = put_user(value, (uint32_t *)buf);
2591 amdgpu_gfx_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff);
2592 mutex_unlock(&adev->grbm_idx_mutex);
2596 mutex_unlock(&adev->pm.mutex);
2601 static ssize_t amdgpu_debugfs_regs_write(struct file *f, const char __user *buf,
2602 size_t size, loff_t *pos)
2604 struct amdgpu_device *adev = file_inode(f)->i_private;
2607 bool pm_pg_lock, use_bank;
2608 unsigned instance_bank, sh_bank, se_bank;
2610 if (size & 0x3 || *pos & 0x3)
2613 /* are we reading registers for which a PG lock is necessary? */
2614 pm_pg_lock = (*pos >> 23) & 1;
2616 if (*pos & (1ULL << 62)) {
2617 se_bank = (*pos >> 24) & 0x3FF;
2618 sh_bank = (*pos >> 34) & 0x3FF;
2619 instance_bank = (*pos >> 44) & 0x3FF;
2621 if (se_bank == 0x3FF)
2622 se_bank = 0xFFFFFFFF;
2623 if (sh_bank == 0x3FF)
2624 sh_bank = 0xFFFFFFFF;
2625 if (instance_bank == 0x3FF)
2626 instance_bank = 0xFFFFFFFF;
2635 if ((sh_bank != 0xFFFFFFFF && sh_bank >= adev->gfx.config.max_sh_per_se) ||
2636 (se_bank != 0xFFFFFFFF && se_bank >= adev->gfx.config.max_shader_engines))
2638 mutex_lock(&adev->grbm_idx_mutex);
2639 amdgpu_gfx_select_se_sh(adev, se_bank,
2640 sh_bank, instance_bank);
2644 mutex_lock(&adev->pm.mutex);
2649 if (*pos > adev->rmmio_size)
2652 r = get_user(value, (uint32_t *)buf);
2656 WREG32(*pos >> 2, value);
2665 amdgpu_gfx_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff);
2666 mutex_unlock(&adev->grbm_idx_mutex);
2670 mutex_unlock(&adev->pm.mutex);
2675 static ssize_t amdgpu_debugfs_regs_pcie_read(struct file *f, char __user *buf,
2676 size_t size, loff_t *pos)
2678 struct amdgpu_device *adev = file_inode(f)->i_private;
2682 if (size & 0x3 || *pos & 0x3)
2688 value = RREG32_PCIE(*pos >> 2);
2689 r = put_user(value, (uint32_t *)buf);
2702 static ssize_t amdgpu_debugfs_regs_pcie_write(struct file *f, const char __user *buf,
2703 size_t size, loff_t *pos)
2705 struct amdgpu_device *adev = file_inode(f)->i_private;
2709 if (size & 0x3 || *pos & 0x3)
2715 r = get_user(value, (uint32_t *)buf);
2719 WREG32_PCIE(*pos >> 2, value);
2730 static ssize_t amdgpu_debugfs_regs_didt_read(struct file *f, char __user *buf,
2731 size_t size, loff_t *pos)
2733 struct amdgpu_device *adev = file_inode(f)->i_private;
2737 if (size & 0x3 || *pos & 0x3)
2743 value = RREG32_DIDT(*pos >> 2);
2744 r = put_user(value, (uint32_t *)buf);
2757 static ssize_t amdgpu_debugfs_regs_didt_write(struct file *f, const char __user *buf,
2758 size_t size, loff_t *pos)
2760 struct amdgpu_device *adev = file_inode(f)->i_private;
2764 if (size & 0x3 || *pos & 0x3)
2770 r = get_user(value, (uint32_t *)buf);
2774 WREG32_DIDT(*pos >> 2, value);
2785 static ssize_t amdgpu_debugfs_regs_smc_read(struct file *f, char __user *buf,
2786 size_t size, loff_t *pos)
2788 struct amdgpu_device *adev = file_inode(f)->i_private;
2792 if (size & 0x3 || *pos & 0x3)
2798 value = RREG32_SMC(*pos);
2799 r = put_user(value, (uint32_t *)buf);
2812 static ssize_t amdgpu_debugfs_regs_smc_write(struct file *f, const char __user *buf,
2813 size_t size, loff_t *pos)
2815 struct amdgpu_device *adev = file_inode(f)->i_private;
2819 if (size & 0x3 || *pos & 0x3)
2825 r = get_user(value, (uint32_t *)buf);
2829 WREG32_SMC(*pos, value);
2840 static ssize_t amdgpu_debugfs_gca_config_read(struct file *f, char __user *buf,
2841 size_t size, loff_t *pos)
2843 struct amdgpu_device *adev = file_inode(f)->i_private;
2846 uint32_t *config, no_regs = 0;
2848 if (size & 0x3 || *pos & 0x3)
2851 config = kmalloc_array(256, sizeof(*config), GFP_KERNEL);
2855 /* version, increment each time something is added */
2856 config[no_regs++] = 2;
2857 config[no_regs++] = adev->gfx.config.max_shader_engines;
2858 config[no_regs++] = adev->gfx.config.max_tile_pipes;
2859 config[no_regs++] = adev->gfx.config.max_cu_per_sh;
2860 config[no_regs++] = adev->gfx.config.max_sh_per_se;
2861 config[no_regs++] = adev->gfx.config.max_backends_per_se;
2862 config[no_regs++] = adev->gfx.config.max_texture_channel_caches;
2863 config[no_regs++] = adev->gfx.config.max_gprs;
2864 config[no_regs++] = adev->gfx.config.max_gs_threads;
2865 config[no_regs++] = adev->gfx.config.max_hw_contexts;
2866 config[no_regs++] = adev->gfx.config.sc_prim_fifo_size_frontend;
2867 config[no_regs++] = adev->gfx.config.sc_prim_fifo_size_backend;
2868 config[no_regs++] = adev->gfx.config.sc_hiz_tile_fifo_size;
2869 config[no_regs++] = adev->gfx.config.sc_earlyz_tile_fifo_size;
2870 config[no_regs++] = adev->gfx.config.num_tile_pipes;
2871 config[no_regs++] = adev->gfx.config.backend_enable_mask;
2872 config[no_regs++] = adev->gfx.config.mem_max_burst_length_bytes;
2873 config[no_regs++] = adev->gfx.config.mem_row_size_in_kb;
2874 config[no_regs++] = adev->gfx.config.shader_engine_tile_size;
2875 config[no_regs++] = adev->gfx.config.num_gpus;
2876 config[no_regs++] = adev->gfx.config.multi_gpu_tile_size;
2877 config[no_regs++] = adev->gfx.config.mc_arb_ramcfg;
2878 config[no_regs++] = adev->gfx.config.gb_addr_config;
2879 config[no_regs++] = adev->gfx.config.num_rbs;
2882 config[no_regs++] = adev->rev_id;
2883 config[no_regs++] = adev->pg_flags;
2884 config[no_regs++] = adev->cg_flags;
2887 config[no_regs++] = adev->family;
2888 config[no_regs++] = adev->external_rev_id;
2890 while (size && (*pos < no_regs * 4)) {
2893 value = config[*pos >> 2];
2894 r = put_user(value, (uint32_t *)buf);
2910 static ssize_t amdgpu_debugfs_sensor_read(struct file *f, char __user *buf,
2911 size_t size, loff_t *pos)
2913 struct amdgpu_device *adev = file_inode(f)->i_private;
2917 if (size != 4 || *pos & 0x3)
2920 /* convert offset to sensor number */
2923 if (adev->powerplay.pp_funcs && adev->powerplay.pp_funcs->read_sensor)
2924 r = adev->powerplay.pp_funcs->read_sensor(adev->powerplay.pp_handle, idx, &value);
2929 r = put_user(value, (int32_t *)buf);
2934 static ssize_t amdgpu_debugfs_wave_read(struct file *f, char __user *buf,
2935 size_t size, loff_t *pos)
2937 struct amdgpu_device *adev = f->f_inode->i_private;
2940 uint32_t offset, se, sh, cu, wave, simd, data[32];
2942 if (size & 3 || *pos & 3)
2946 offset = (*pos & 0x7F);
2947 se = ((*pos >> 7) & 0xFF);
2948 sh = ((*pos >> 15) & 0xFF);
2949 cu = ((*pos >> 23) & 0xFF);
2950 wave = ((*pos >> 31) & 0xFF);
2951 simd = ((*pos >> 37) & 0xFF);
2953 /* switch to the specific se/sh/cu */
2954 mutex_lock(&adev->grbm_idx_mutex);
2955 amdgpu_gfx_select_se_sh(adev, se, sh, cu);
2958 if (adev->gfx.funcs->read_wave_data)
2959 adev->gfx.funcs->read_wave_data(adev, simd, wave, data, &x);
2961 amdgpu_gfx_select_se_sh(adev, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
2962 mutex_unlock(&adev->grbm_idx_mutex);
2967 while (size && (offset < x * 4)) {
2970 value = data[offset >> 2];
2971 r = put_user(value, (uint32_t *)buf);
2984 static ssize_t amdgpu_debugfs_gpr_read(struct file *f, char __user *buf,
2985 size_t size, loff_t *pos)
2987 struct amdgpu_device *adev = f->f_inode->i_private;
2990 uint32_t offset, se, sh, cu, wave, simd, thread, bank, *data;
2992 if (size & 3 || *pos & 3)
2996 offset = (*pos & 0xFFF); /* in dwords */
2997 se = ((*pos >> 12) & 0xFF);
2998 sh = ((*pos >> 20) & 0xFF);
2999 cu = ((*pos >> 28) & 0xFF);
3000 wave = ((*pos >> 36) & 0xFF);
3001 simd = ((*pos >> 44) & 0xFF);
3002 thread = ((*pos >> 52) & 0xFF);
3003 bank = ((*pos >> 60) & 1);
3005 data = kmalloc_array(1024, sizeof(*data), GFP_KERNEL);
3009 /* switch to the specific se/sh/cu */
3010 mutex_lock(&adev->grbm_idx_mutex);
3011 amdgpu_gfx_select_se_sh(adev, se, sh, cu);
3014 if (adev->gfx.funcs->read_wave_vgprs)
3015 adev->gfx.funcs->read_wave_vgprs(adev, simd, wave, thread, offset, size>>2, data);
3017 if (adev->gfx.funcs->read_wave_sgprs)
3018 adev->gfx.funcs->read_wave_sgprs(adev, simd, wave, offset, size>>2, data);
3021 amdgpu_gfx_select_se_sh(adev, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
3022 mutex_unlock(&adev->grbm_idx_mutex);
3027 value = data[offset++];
3028 r = put_user(value, (uint32_t *)buf);
3044 static const struct file_operations amdgpu_debugfs_regs_fops = {
3045 .owner = THIS_MODULE,
3046 .read = amdgpu_debugfs_regs_read,
3047 .write = amdgpu_debugfs_regs_write,
3048 .llseek = default_llseek
3050 static const struct file_operations amdgpu_debugfs_regs_didt_fops = {
3051 .owner = THIS_MODULE,
3052 .read = amdgpu_debugfs_regs_didt_read,
3053 .write = amdgpu_debugfs_regs_didt_write,
3054 .llseek = default_llseek
3056 static const struct file_operations amdgpu_debugfs_regs_pcie_fops = {
3057 .owner = THIS_MODULE,
3058 .read = amdgpu_debugfs_regs_pcie_read,
3059 .write = amdgpu_debugfs_regs_pcie_write,
3060 .llseek = default_llseek
3062 static const struct file_operations amdgpu_debugfs_regs_smc_fops = {
3063 .owner = THIS_MODULE,
3064 .read = amdgpu_debugfs_regs_smc_read,
3065 .write = amdgpu_debugfs_regs_smc_write,
3066 .llseek = default_llseek
3069 static const struct file_operations amdgpu_debugfs_gca_config_fops = {
3070 .owner = THIS_MODULE,
3071 .read = amdgpu_debugfs_gca_config_read,
3072 .llseek = default_llseek
3075 static const struct file_operations amdgpu_debugfs_sensors_fops = {
3076 .owner = THIS_MODULE,
3077 .read = amdgpu_debugfs_sensor_read,
3078 .llseek = default_llseek
3081 static const struct file_operations amdgpu_debugfs_wave_fops = {
3082 .owner = THIS_MODULE,
3083 .read = amdgpu_debugfs_wave_read,
3084 .llseek = default_llseek
3086 static const struct file_operations amdgpu_debugfs_gpr_fops = {
3087 .owner = THIS_MODULE,
3088 .read = amdgpu_debugfs_gpr_read,
3089 .llseek = default_llseek
3092 static const struct file_operations *debugfs_regs[] = {
3093 &amdgpu_debugfs_regs_fops,
3094 &amdgpu_debugfs_regs_didt_fops,
3095 &amdgpu_debugfs_regs_pcie_fops,
3096 &amdgpu_debugfs_regs_smc_fops,
3097 &amdgpu_debugfs_gca_config_fops,
3098 &amdgpu_debugfs_sensors_fops,
3099 &amdgpu_debugfs_wave_fops,
3100 &amdgpu_debugfs_gpr_fops,
3103 static const char *debugfs_regs_names[] = {
3108 "amdgpu_gca_config",
3114 static int amdgpu_debugfs_regs_init(struct amdgpu_device *adev)
3116 struct drm_minor *minor = adev->ddev->primary;
3117 struct dentry *ent, *root = minor->debugfs_root;
3120 for (i = 0; i < ARRAY_SIZE(debugfs_regs); i++) {
3121 ent = debugfs_create_file(debugfs_regs_names[i],
3122 S_IFREG | S_IRUGO, root,
3123 adev, debugfs_regs[i]);
3125 for (j = 0; j < i; j++) {
3126 debugfs_remove(adev->debugfs_regs[i]);
3127 adev->debugfs_regs[i] = NULL;
3129 return PTR_ERR(ent);
3133 i_size_write(ent->d_inode, adev->rmmio_size);
3134 adev->debugfs_regs[i] = ent;
3140 static void amdgpu_debugfs_regs_cleanup(struct amdgpu_device *adev)
3144 for (i = 0; i < ARRAY_SIZE(debugfs_regs); i++) {
3145 if (adev->debugfs_regs[i]) {
3146 debugfs_remove(adev->debugfs_regs[i]);
3147 adev->debugfs_regs[i] = NULL;
3152 int amdgpu_debugfs_init(struct drm_minor *minor)
3157 void amdgpu_debugfs_cleanup(struct drm_minor *minor)
3161 static int amdgpu_debugfs_regs_init(struct amdgpu_device *adev)
3165 static void amdgpu_debugfs_regs_cleanup(struct amdgpu_device *adev) { }