2 * Copyright © 2010 Daniel Vetter
3 * Copyright © 2011-2014 Intel Corporation
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
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 AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
26 #include <linux/seq_file.h>
28 #include <drm/i915_drm.h>
30 #include "i915_vgpu.h"
31 #include "i915_trace.h"
32 #include "intel_drv.h"
34 #include <linux/bitmap.h>
36 #include <sys/mplock2.h>
39 * DOC: Global GTT views
41 * Background and previous state
43 * Historically objects could exists (be bound) in global GTT space only as
44 * singular instances with a view representing all of the object's backing pages
45 * in a linear fashion. This view will be called a normal view.
47 * To support multiple views of the same object, where the number of mapped
48 * pages is not equal to the backing store, or where the layout of the pages
49 * is not linear, concept of a GGTT view was added.
51 * One example of an alternative view is a stereo display driven by a single
52 * image. In this case we would have a framebuffer looking like this
58 * Above would represent a normal GGTT view as normally mapped for GPU or CPU
59 * rendering. In contrast, fed to the display engine would be an alternative
60 * view which could look something like this:
65 * In this example both the size and layout of pages in the alternative view is
66 * different from the normal view.
68 * Implementation and usage
70 * GGTT views are implemented using VMAs and are distinguished via enum
71 * i915_ggtt_view_type and struct i915_ggtt_view.
73 * A new flavour of core GEM functions which work with GGTT bound objects were
74 * added with the _ggtt_ infix, and sometimes with _view postfix to avoid
75 * renaming in large amounts of code. They take the struct i915_ggtt_view
76 * parameter encapsulating all metadata required to implement a view.
78 * As a helper for callers which are only interested in the normal view,
79 * globally const i915_ggtt_view_normal singleton instance exists. All old core
80 * GEM API functions, the ones not taking the view parameter, are operating on,
81 * or with the normal GGTT view.
83 * Code wanting to add or use a new GGTT view needs to:
85 * 1. Add a new enum with a suitable name.
86 * 2. Extend the metadata in the i915_ggtt_view structure if required.
87 * 3. Add support to i915_get_vma_pages().
89 * New views are required to build a scatter-gather table from within the
90 * i915_get_vma_pages function. This table is stored in the vma.ggtt_view and
91 * exists for the lifetime of an VMA.
93 * Core API is designed to have copy semantics which means that passed in
94 * struct i915_ggtt_view does not need to be persistent (left around after
95 * calling the core API functions).
100 i915_get_ggtt_vma_pages(struct i915_vma *vma);
102 const struct i915_ggtt_view i915_ggtt_view_normal = {
103 .type = I915_GGTT_VIEW_NORMAL,
105 const struct i915_ggtt_view i915_ggtt_view_rotated = {
106 .type = I915_GGTT_VIEW_ROTATED,
109 static int sanitize_enable_ppgtt(struct drm_device *dev, int enable_ppgtt)
111 bool has_aliasing_ppgtt;
113 bool has_full_48bit_ppgtt;
115 has_aliasing_ppgtt = INTEL_INFO(dev)->gen >= 6;
116 has_full_ppgtt = INTEL_INFO(dev)->gen >= 7;
117 has_full_48bit_ppgtt = IS_BROADWELL(dev) || INTEL_INFO(dev)->gen >= 9;
119 if (intel_vgpu_active(dev))
120 has_full_ppgtt = false; /* emulation is too hard */
123 * We don't allow disabling PPGTT for gen9+ as it's a requirement for
124 * execlists, the sole mechanism available to submit work.
126 if (INTEL_INFO(dev)->gen < 9 &&
127 (enable_ppgtt == 0 || !has_aliasing_ppgtt))
130 if (enable_ppgtt == 1)
133 if (enable_ppgtt == 2 && has_full_ppgtt)
136 if (enable_ppgtt == 3 && has_full_48bit_ppgtt)
139 #ifdef CONFIG_INTEL_IOMMU
140 /* Disable ppgtt on SNB if VT-d is on. */
141 if (INTEL_INFO(dev)->gen == 6 && intel_iommu_gfx_mapped) {
142 DRM_INFO("Disabling PPGTT because VT-d is on\n");
147 /* Early VLV doesn't have this */
148 if (IS_VALLEYVIEW(dev) && dev->pdev->revision < 0xb) {
149 DRM_DEBUG_DRIVER("disabling PPGTT on pre-B3 step VLV\n");
153 if (INTEL_INFO(dev)->gen >= 8 && i915.enable_execlists)
154 return has_full_48bit_ppgtt ? 3 : 2;
156 return has_aliasing_ppgtt ? 1 : 0;
159 static int ppgtt_bind_vma(struct i915_vma *vma,
160 enum i915_cache_level cache_level,
165 /* Currently applicable only to VLV */
167 pte_flags |= PTE_READ_ONLY;
169 vma->vm->insert_entries(vma->vm, vma->obj->pages, vma->node.start,
170 cache_level, pte_flags);
175 static void ppgtt_unbind_vma(struct i915_vma *vma)
177 vma->vm->clear_range(vma->vm,
183 static gen8_pte_t gen8_pte_encode(dma_addr_t addr,
184 enum i915_cache_level level,
187 gen8_pte_t pte = valid ? _PAGE_PRESENT | _PAGE_RW : 0;
191 case I915_CACHE_NONE:
192 pte |= PPAT_UNCACHED_INDEX;
195 pte |= PPAT_DISPLAY_ELLC_INDEX;
198 pte |= PPAT_CACHED_INDEX;
205 static gen8_pde_t gen8_pde_encode(const dma_addr_t addr,
206 const enum i915_cache_level level)
208 gen8_pde_t pde = _PAGE_PRESENT | _PAGE_RW;
210 if (level != I915_CACHE_NONE)
211 pde |= PPAT_CACHED_PDE_INDEX;
213 pde |= PPAT_UNCACHED_INDEX;
217 #define gen8_pdpe_encode gen8_pde_encode
218 #define gen8_pml4e_encode gen8_pde_encode
220 static gen6_pte_t snb_pte_encode(dma_addr_t addr,
221 enum i915_cache_level level,
222 bool valid, u32 unused)
224 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
225 pte |= GEN6_PTE_ADDR_ENCODE(addr);
228 case I915_CACHE_L3_LLC:
230 pte |= GEN6_PTE_CACHE_LLC;
232 case I915_CACHE_NONE:
233 pte |= GEN6_PTE_UNCACHED;
242 static gen6_pte_t ivb_pte_encode(dma_addr_t addr,
243 enum i915_cache_level level,
244 bool valid, u32 unused)
246 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
247 pte |= GEN6_PTE_ADDR_ENCODE(addr);
250 case I915_CACHE_L3_LLC:
251 pte |= GEN7_PTE_CACHE_L3_LLC;
254 pte |= GEN6_PTE_CACHE_LLC;
256 case I915_CACHE_NONE:
257 pte |= GEN6_PTE_UNCACHED;
266 static gen6_pte_t byt_pte_encode(dma_addr_t addr,
267 enum i915_cache_level level,
268 bool valid, u32 flags)
270 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
271 pte |= GEN6_PTE_ADDR_ENCODE(addr);
273 if (!(flags & PTE_READ_ONLY))
274 pte |= BYT_PTE_WRITEABLE;
276 if (level != I915_CACHE_NONE)
277 pte |= BYT_PTE_SNOOPED_BY_CPU_CACHES;
282 static gen6_pte_t hsw_pte_encode(dma_addr_t addr,
283 enum i915_cache_level level,
284 bool valid, u32 unused)
286 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
287 pte |= HSW_PTE_ADDR_ENCODE(addr);
289 if (level != I915_CACHE_NONE)
290 pte |= HSW_WB_LLC_AGE3;
295 static gen6_pte_t iris_pte_encode(dma_addr_t addr,
296 enum i915_cache_level level,
297 bool valid, u32 unused)
299 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
300 pte |= HSW_PTE_ADDR_ENCODE(addr);
303 case I915_CACHE_NONE:
306 pte |= HSW_WT_ELLC_LLC_AGE3;
309 pte |= HSW_WB_ELLC_LLC_AGE3;
316 static int __setup_page_dma(struct drm_device *dev,
317 struct i915_page_dma *p, gfp_t flags)
319 struct device *device = &dev->pdev->dev;
321 p->page = alloc_page(flags);
325 p->daddr = dma_map_page(device,
326 p->page, 0, 4096, PCI_DMA_BIDIRECTIONAL);
328 if (dma_mapping_error(device, p->daddr)) {
329 __free_page(p->page);
336 static int setup_page_dma(struct drm_device *dev, struct i915_page_dma *p)
338 return __setup_page_dma(dev, p, GFP_KERNEL);
341 static void cleanup_page_dma(struct drm_device *dev, struct i915_page_dma *p)
343 if (WARN_ON(!p->page))
346 dma_unmap_page(&dev->pdev->dev, p->daddr, 4096, PCI_DMA_BIDIRECTIONAL);
347 __free_page(p->page);
348 memset(p, 0, sizeof(*p));
351 static void *kmap_page_dma(struct i915_page_dma *p)
353 return kmap_atomic(p->page);
356 /* We use the flushing unmap only with ppgtt structures:
357 * page directories, page tables and scratch pages.
359 static void kunmap_page_dma(struct drm_device *dev, void *vaddr)
361 /* There are only few exceptions for gen >=6. chv and bxt.
362 * And we are not sure about the latter so play safe for now.
364 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
365 drm_clflush_virt_range(vaddr, PAGE_SIZE);
367 kunmap_atomic(vaddr);
370 #define kmap_px(px) kmap_page_dma(px_base(px))
371 #define kunmap_px(ppgtt, vaddr) kunmap_page_dma((ppgtt)->base.dev, (vaddr))
373 #define setup_px(dev, px) setup_page_dma((dev), px_base(px))
374 #define cleanup_px(dev, px) cleanup_page_dma((dev), px_base(px))
375 #define fill_px(dev, px, v) fill_page_dma((dev), px_base(px), (v))
376 #define fill32_px(dev, px, v) fill_page_dma_32((dev), px_base(px), (v))
378 static void fill_page_dma(struct drm_device *dev, struct i915_page_dma *p,
382 uint64_t * const vaddr = kmap_page_dma(p);
384 for (i = 0; i < 512; i++)
387 kunmap_page_dma(dev, vaddr);
390 static void fill_page_dma_32(struct drm_device *dev, struct i915_page_dma *p,
391 const uint32_t val32)
397 fill_page_dma(dev, p, v);
400 static struct i915_page_scratch *alloc_scratch_page(struct drm_device *dev)
402 struct i915_page_scratch *sp;
405 sp = kzalloc(sizeof(*sp), GFP_KERNEL);
407 return ERR_PTR(-ENOMEM);
409 ret = __setup_page_dma(dev, px_base(sp), GFP_DMA32 | __GFP_ZERO);
415 set_pages_uc(px_page(sp), 1);
420 static void free_scratch_page(struct drm_device *dev,
421 struct i915_page_scratch *sp)
423 set_pages_wb(px_page(sp), 1);
429 static struct i915_page_table *alloc_pt(struct drm_device *dev)
431 struct i915_page_table *pt;
432 const size_t count = INTEL_INFO(dev)->gen >= 8 ?
433 GEN8_PTES : GEN6_PTES;
436 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
438 return ERR_PTR(-ENOMEM);
440 pt->used_ptes = kcalloc(BITS_TO_LONGS(count), sizeof(*pt->used_ptes),
446 ret = setup_px(dev, pt);
453 kfree(pt->used_ptes);
460 static void free_pt(struct drm_device *dev, struct i915_page_table *pt)
463 kfree(pt->used_ptes);
467 static void gen8_initialize_pt(struct i915_address_space *vm,
468 struct i915_page_table *pt)
470 gen8_pte_t scratch_pte;
472 scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
473 I915_CACHE_LLC, true);
475 fill_px(vm->dev, pt, scratch_pte);
478 static void gen6_initialize_pt(struct i915_address_space *vm,
479 struct i915_page_table *pt)
481 gen6_pte_t scratch_pte;
483 WARN_ON(px_dma(vm->scratch_page) == 0);
485 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
486 I915_CACHE_LLC, true, 0);
488 fill32_px(vm->dev, pt, scratch_pte);
491 static struct i915_page_directory *alloc_pd(struct drm_device *dev)
493 struct i915_page_directory *pd;
496 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
498 return ERR_PTR(-ENOMEM);
500 pd->used_pdes = kcalloc(BITS_TO_LONGS(I915_PDES),
501 sizeof(*pd->used_pdes), GFP_KERNEL);
505 ret = setup_px(dev, pd);
512 kfree(pd->used_pdes);
519 static void free_pd(struct drm_device *dev, struct i915_page_directory *pd)
523 kfree(pd->used_pdes);
528 static void gen8_initialize_pd(struct i915_address_space *vm,
529 struct i915_page_directory *pd)
531 gen8_pde_t scratch_pde;
533 scratch_pde = gen8_pde_encode(px_dma(vm->scratch_pt), I915_CACHE_LLC);
535 fill_px(vm->dev, pd, scratch_pde);
538 static int __pdp_init(struct drm_device *dev,
539 struct i915_page_directory_pointer *pdp)
541 size_t pdpes = I915_PDPES_PER_PDP(dev);
543 pdp->used_pdpes = kcalloc(BITS_TO_LONGS(pdpes),
544 sizeof(unsigned long),
546 if (!pdp->used_pdpes)
549 pdp->page_directory = kcalloc(pdpes, sizeof(*pdp->page_directory),
551 if (!pdp->page_directory) {
552 kfree(pdp->used_pdpes);
553 /* the PDP might be the statically allocated top level. Keep it
554 * as clean as possible */
555 pdp->used_pdpes = NULL;
562 static void __pdp_fini(struct i915_page_directory_pointer *pdp)
564 kfree(pdp->used_pdpes);
565 kfree(pdp->page_directory);
566 pdp->page_directory = NULL;
570 i915_page_directory_pointer *alloc_pdp(struct drm_device *dev)
572 struct i915_page_directory_pointer *pdp;
575 WARN_ON(!USES_FULL_48BIT_PPGTT(dev));
577 pdp = kzalloc(sizeof(*pdp), GFP_KERNEL);
579 return ERR_PTR(-ENOMEM);
581 ret = __pdp_init(dev, pdp);
585 ret = setup_px(dev, pdp);
599 static void free_pdp(struct drm_device *dev,
600 struct i915_page_directory_pointer *pdp)
603 if (USES_FULL_48BIT_PPGTT(dev)) {
604 cleanup_px(dev, pdp);
609 static void gen8_initialize_pdp(struct i915_address_space *vm,
610 struct i915_page_directory_pointer *pdp)
612 gen8_ppgtt_pdpe_t scratch_pdpe;
614 scratch_pdpe = gen8_pdpe_encode(px_dma(vm->scratch_pd), I915_CACHE_LLC);
616 fill_px(vm->dev, pdp, scratch_pdpe);
619 static void gen8_initialize_pml4(struct i915_address_space *vm,
620 struct i915_pml4 *pml4)
622 gen8_ppgtt_pml4e_t scratch_pml4e;
624 scratch_pml4e = gen8_pml4e_encode(px_dma(vm->scratch_pdp),
627 fill_px(vm->dev, pml4, scratch_pml4e);
631 gen8_setup_page_directory(struct i915_hw_ppgtt *ppgtt,
632 struct i915_page_directory_pointer *pdp,
633 struct i915_page_directory *pd,
636 gen8_ppgtt_pdpe_t *page_directorypo;
638 if (!USES_FULL_48BIT_PPGTT(ppgtt->base.dev))
641 page_directorypo = kmap_px(pdp);
642 page_directorypo[index] = gen8_pdpe_encode(px_dma(pd), I915_CACHE_LLC);
643 kunmap_px(ppgtt, page_directorypo);
647 gen8_setup_page_directory_pointer(struct i915_hw_ppgtt *ppgtt,
648 struct i915_pml4 *pml4,
649 struct i915_page_directory_pointer *pdp,
652 gen8_ppgtt_pml4e_t *pagemap = kmap_px(pml4);
654 WARN_ON(!USES_FULL_48BIT_PPGTT(ppgtt->base.dev));
655 pagemap[index] = gen8_pml4e_encode(px_dma(pdp), I915_CACHE_LLC);
656 kunmap_px(ppgtt, pagemap);
659 /* Broadwell Page Directory Pointer Descriptors */
660 static int gen8_write_pdp(struct drm_i915_gem_request *req,
664 struct intel_engine_cs *ring = req->ring;
669 ret = intel_ring_begin(req, 6);
673 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
674 intel_ring_emit_reg(ring, GEN8_RING_PDP_UDW(ring, entry));
675 intel_ring_emit(ring, upper_32_bits(addr));
676 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
677 intel_ring_emit_reg(ring, GEN8_RING_PDP_LDW(ring, entry));
678 intel_ring_emit(ring, lower_32_bits(addr));
679 intel_ring_advance(ring);
684 static int gen8_legacy_mm_switch(struct i915_hw_ppgtt *ppgtt,
685 struct drm_i915_gem_request *req)
689 for (i = GEN8_LEGACY_PDPES - 1; i >= 0; i--) {
690 const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i);
692 ret = gen8_write_pdp(req, i, pd_daddr);
700 static int gen8_48b_mm_switch(struct i915_hw_ppgtt *ppgtt,
701 struct drm_i915_gem_request *req)
703 return gen8_write_pdp(req, 0, px_dma(&ppgtt->pml4));
706 static void gen8_ppgtt_clear_pte_range(struct i915_address_space *vm,
707 struct i915_page_directory_pointer *pdp,
710 gen8_pte_t scratch_pte)
712 struct i915_hw_ppgtt *ppgtt =
713 container_of(vm, struct i915_hw_ppgtt, base);
714 gen8_pte_t *pt_vaddr;
715 unsigned pdpe = gen8_pdpe_index(start);
716 unsigned pde = gen8_pde_index(start);
717 unsigned pte = gen8_pte_index(start);
718 unsigned num_entries = length >> PAGE_SHIFT;
719 unsigned last_pte, i;
724 while (num_entries) {
725 struct i915_page_directory *pd;
726 struct i915_page_table *pt;
728 if (WARN_ON(!pdp->page_directory[pdpe]))
731 pd = pdp->page_directory[pdpe];
733 if (WARN_ON(!pd->page_table[pde]))
736 pt = pd->page_table[pde];
738 if (WARN_ON(!px_page(pt)))
741 last_pte = pte + num_entries;
742 if (last_pte > GEN8_PTES)
743 last_pte = GEN8_PTES;
745 pt_vaddr = kmap_px(pt);
747 for (i = pte; i < last_pte; i++) {
748 pt_vaddr[i] = scratch_pte;
752 kunmap_px(ppgtt, pt);
755 if (++pde == I915_PDES) {
756 if (++pdpe == I915_PDPES_PER_PDP(vm->dev))
763 static void gen8_ppgtt_clear_range(struct i915_address_space *vm,
768 struct i915_hw_ppgtt *ppgtt =
769 container_of(vm, struct i915_hw_ppgtt, base);
770 gen8_pte_t scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
771 I915_CACHE_LLC, use_scratch);
773 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
774 gen8_ppgtt_clear_pte_range(vm, &ppgtt->pdp, start, length,
778 struct i915_page_directory_pointer *pdp;
780 gen8_for_each_pml4e(pdp, &ppgtt->pml4, start, length, pml4e) {
781 gen8_ppgtt_clear_pte_range(vm, pdp, start, length,
788 gen8_ppgtt_insert_pte_entries(struct i915_address_space *vm,
789 struct i915_page_directory_pointer *pdp,
790 struct sg_page_iter *sg_iter,
792 enum i915_cache_level cache_level)
794 struct i915_hw_ppgtt *ppgtt =
795 container_of(vm, struct i915_hw_ppgtt, base);
796 gen8_pte_t *pt_vaddr;
797 unsigned pdpe = gen8_pdpe_index(start);
798 unsigned pde = gen8_pde_index(start);
799 unsigned pte = gen8_pte_index(start);
803 while (__sg_page_iter_next(sg_iter)) {
804 if (pt_vaddr == NULL) {
805 struct i915_page_directory *pd = pdp->page_directory[pdpe];
806 struct i915_page_table *pt = pd->page_table[pde];
807 pt_vaddr = kmap_px(pt);
811 gen8_pte_encode(sg_page_iter_dma_address(sg_iter),
813 if (++pte == GEN8_PTES) {
814 kunmap_px(ppgtt, pt_vaddr);
816 if (++pde == I915_PDES) {
817 if (++pdpe == I915_PDPES_PER_PDP(vm->dev))
826 kunmap_px(ppgtt, pt_vaddr);
829 static void gen8_ppgtt_insert_entries(struct i915_address_space *vm,
830 struct sg_table *pages,
832 enum i915_cache_level cache_level,
835 struct i915_hw_ppgtt *ppgtt =
836 container_of(vm, struct i915_hw_ppgtt, base);
837 struct sg_page_iter sg_iter;
839 __sg_page_iter_start(&sg_iter, pages->sgl, sg_nents(pages->sgl), 0);
841 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
842 gen8_ppgtt_insert_pte_entries(vm, &ppgtt->pdp, &sg_iter, start,
845 struct i915_page_directory_pointer *pdp;
847 uint64_t length = (uint64_t)pages->orig_nents << PAGE_SHIFT;
849 gen8_for_each_pml4e(pdp, &ppgtt->pml4, start, length, pml4e) {
850 gen8_ppgtt_insert_pte_entries(vm, pdp, &sg_iter,
856 static void gen8_free_page_tables(struct drm_device *dev,
857 struct i915_page_directory *pd)
864 for_each_set_bit(i, pd->used_pdes, I915_PDES) {
865 if (WARN_ON(!pd->page_table[i]))
868 free_pt(dev, pd->page_table[i]);
869 pd->page_table[i] = NULL;
873 static int gen8_init_scratch(struct i915_address_space *vm)
875 struct drm_device *dev = vm->dev;
877 vm->scratch_page = alloc_scratch_page(dev);
878 if (IS_ERR(vm->scratch_page))
879 return PTR_ERR(vm->scratch_page);
881 vm->scratch_pt = alloc_pt(dev);
882 if (IS_ERR(vm->scratch_pt)) {
883 free_scratch_page(dev, vm->scratch_page);
884 return PTR_ERR(vm->scratch_pt);
887 vm->scratch_pd = alloc_pd(dev);
888 if (IS_ERR(vm->scratch_pd)) {
889 free_pt(dev, vm->scratch_pt);
890 free_scratch_page(dev, vm->scratch_page);
891 return PTR_ERR(vm->scratch_pd);
894 if (USES_FULL_48BIT_PPGTT(dev)) {
895 vm->scratch_pdp = alloc_pdp(dev);
896 if (IS_ERR(vm->scratch_pdp)) {
897 free_pd(dev, vm->scratch_pd);
898 free_pt(dev, vm->scratch_pt);
899 free_scratch_page(dev, vm->scratch_page);
900 return PTR_ERR(vm->scratch_pdp);
904 gen8_initialize_pt(vm, vm->scratch_pt);
905 gen8_initialize_pd(vm, vm->scratch_pd);
906 if (USES_FULL_48BIT_PPGTT(dev))
907 gen8_initialize_pdp(vm, vm->scratch_pdp);
912 static int gen8_ppgtt_notify_vgt(struct i915_hw_ppgtt *ppgtt, bool create)
914 enum vgt_g2v_type msg;
915 struct drm_device *dev = ppgtt->base.dev;
916 struct drm_i915_private *dev_priv = dev->dev_private;
919 if (USES_FULL_48BIT_PPGTT(dev)) {
920 u64 daddr = px_dma(&ppgtt->pml4);
922 I915_WRITE(vgtif_reg(pdp[0].lo), lower_32_bits(daddr));
923 I915_WRITE(vgtif_reg(pdp[0].hi), upper_32_bits(daddr));
925 msg = (create ? VGT_G2V_PPGTT_L4_PAGE_TABLE_CREATE :
926 VGT_G2V_PPGTT_L4_PAGE_TABLE_DESTROY);
928 for (i = 0; i < GEN8_LEGACY_PDPES; i++) {
929 u64 daddr = i915_page_dir_dma_addr(ppgtt, i);
931 I915_WRITE(vgtif_reg(pdp[i].lo), lower_32_bits(daddr));
932 I915_WRITE(vgtif_reg(pdp[i].hi), upper_32_bits(daddr));
935 msg = (create ? VGT_G2V_PPGTT_L3_PAGE_TABLE_CREATE :
936 VGT_G2V_PPGTT_L3_PAGE_TABLE_DESTROY);
939 I915_WRITE(vgtif_reg(g2v_notify), msg);
944 static void gen8_free_scratch(struct i915_address_space *vm)
946 struct drm_device *dev = vm->dev;
948 if (USES_FULL_48BIT_PPGTT(dev))
949 free_pdp(dev, vm->scratch_pdp);
950 free_pd(dev, vm->scratch_pd);
951 free_pt(dev, vm->scratch_pt);
952 free_scratch_page(dev, vm->scratch_page);
955 static void gen8_ppgtt_cleanup_3lvl(struct drm_device *dev,
956 struct i915_page_directory_pointer *pdp)
960 for_each_set_bit(i, pdp->used_pdpes, I915_PDPES_PER_PDP(dev)) {
961 if (WARN_ON(!pdp->page_directory[i]))
964 gen8_free_page_tables(dev, pdp->page_directory[i]);
965 free_pd(dev, pdp->page_directory[i]);
971 static void gen8_ppgtt_cleanup_4lvl(struct i915_hw_ppgtt *ppgtt)
975 for_each_set_bit(i, ppgtt->pml4.used_pml4es, GEN8_PML4ES_PER_PML4) {
976 if (WARN_ON(!ppgtt->pml4.pdps[i]))
979 gen8_ppgtt_cleanup_3lvl(ppgtt->base.dev, ppgtt->pml4.pdps[i]);
982 cleanup_px(ppgtt->base.dev, &ppgtt->pml4);
985 static void gen8_ppgtt_cleanup(struct i915_address_space *vm)
987 struct i915_hw_ppgtt *ppgtt =
988 container_of(vm, struct i915_hw_ppgtt, base);
990 if (intel_vgpu_active(vm->dev))
991 gen8_ppgtt_notify_vgt(ppgtt, false);
993 if (!USES_FULL_48BIT_PPGTT(ppgtt->base.dev))
994 gen8_ppgtt_cleanup_3lvl(ppgtt->base.dev, &ppgtt->pdp);
996 gen8_ppgtt_cleanup_4lvl(ppgtt);
998 gen8_free_scratch(vm);
1002 * gen8_ppgtt_alloc_pagetabs() - Allocate page tables for VA range.
1003 * @vm: Master vm structure.
1004 * @pd: Page directory for this address range.
1005 * @start: Starting virtual address to begin allocations.
1006 * @length: Size of the allocations.
1007 * @new_pts: Bitmap set by function with new allocations. Likely used by the
1008 * caller to free on error.
1010 * Allocate the required number of page tables. Extremely similar to
1011 * gen8_ppgtt_alloc_page_directories(). The main difference is here we are limited by
1012 * the page directory boundary (instead of the page directory pointer). That
1013 * boundary is 1GB virtual. Therefore, unlike gen8_ppgtt_alloc_page_directories(), it is
1014 * possible, and likely that the caller will need to use multiple calls of this
1015 * function to achieve the appropriate allocation.
1017 * Return: 0 if success; negative error code otherwise.
1019 static int gen8_ppgtt_alloc_pagetabs(struct i915_address_space *vm,
1020 struct i915_page_directory *pd,
1023 unsigned long *new_pts)
1025 struct drm_device *dev = vm->dev;
1026 struct i915_page_table *pt;
1029 gen8_for_each_pde(pt, pd, start, length, pde) {
1030 /* Don't reallocate page tables */
1031 if (test_bit(pde, pd->used_pdes)) {
1032 /* Scratch is never allocated this way */
1033 WARN_ON(pt == vm->scratch_pt);
1041 gen8_initialize_pt(vm, pt);
1042 pd->page_table[pde] = pt;
1043 __set_bit(pde, new_pts);
1044 trace_i915_page_table_entry_alloc(vm, pde, start, GEN8_PDE_SHIFT);
1050 for_each_set_bit(pde, new_pts, I915_PDES)
1051 free_pt(dev, pd->page_table[pde]);
1057 * gen8_ppgtt_alloc_page_directories() - Allocate page directories for VA range.
1058 * @vm: Master vm structure.
1059 * @pdp: Page directory pointer for this address range.
1060 * @start: Starting virtual address to begin allocations.
1061 * @length: Size of the allocations.
1062 * @new_pds: Bitmap set by function with new allocations. Likely used by the
1063 * caller to free on error.
1065 * Allocate the required number of page directories starting at the pde index of
1066 * @start, and ending at the pde index @start + @length. This function will skip
1067 * over already allocated page directories within the range, and only allocate
1068 * new ones, setting the appropriate pointer within the pdp as well as the
1069 * correct position in the bitmap @new_pds.
1071 * The function will only allocate the pages within the range for a give page
1072 * directory pointer. In other words, if @start + @length straddles a virtually
1073 * addressed PDP boundary (512GB for 4k pages), there will be more allocations
1074 * required by the caller, This is not currently possible, and the BUG in the
1075 * code will prevent it.
1077 * Return: 0 if success; negative error code otherwise.
1080 gen8_ppgtt_alloc_page_directories(struct i915_address_space *vm,
1081 struct i915_page_directory_pointer *pdp,
1084 unsigned long *new_pds)
1086 struct drm_device *dev = vm->dev;
1087 struct i915_page_directory *pd;
1089 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1091 WARN_ON(!bitmap_empty(new_pds, pdpes));
1093 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1094 if (test_bit(pdpe, pdp->used_pdpes))
1101 gen8_initialize_pd(vm, pd);
1102 pdp->page_directory[pdpe] = pd;
1103 __set_bit(pdpe, new_pds);
1104 trace_i915_page_directory_entry_alloc(vm, pdpe, start, GEN8_PDPE_SHIFT);
1110 for_each_set_bit(pdpe, new_pds, pdpes)
1111 free_pd(dev, pdp->page_directory[pdpe]);
1117 * gen8_ppgtt_alloc_page_dirpointers() - Allocate pdps for VA range.
1118 * @vm: Master vm structure.
1119 * @pml4: Page map level 4 for this address range.
1120 * @start: Starting virtual address to begin allocations.
1121 * @length: Size of the allocations.
1122 * @new_pdps: Bitmap set by function with new allocations. Likely used by the
1123 * caller to free on error.
1125 * Allocate the required number of page directory pointers. Extremely similar to
1126 * gen8_ppgtt_alloc_page_directories() and gen8_ppgtt_alloc_pagetabs().
1127 * The main difference is here we are limited by the pml4 boundary (instead of
1128 * the page directory pointer).
1130 * Return: 0 if success; negative error code otherwise.
1133 gen8_ppgtt_alloc_page_dirpointers(struct i915_address_space *vm,
1134 struct i915_pml4 *pml4,
1137 unsigned long *new_pdps)
1139 struct drm_device *dev = vm->dev;
1140 struct i915_page_directory_pointer *pdp;
1143 WARN_ON(!bitmap_empty(new_pdps, GEN8_PML4ES_PER_PML4));
1145 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1146 if (!test_bit(pml4e, pml4->used_pml4es)) {
1147 pdp = alloc_pdp(dev);
1151 gen8_initialize_pdp(vm, pdp);
1152 pml4->pdps[pml4e] = pdp;
1153 __set_bit(pml4e, new_pdps);
1154 trace_i915_page_directory_pointer_entry_alloc(vm,
1164 for_each_set_bit(pml4e, new_pdps, GEN8_PML4ES_PER_PML4)
1165 free_pdp(dev, pml4->pdps[pml4e]);
1171 free_gen8_temp_bitmaps(unsigned long *new_pds, unsigned long *new_pts)
1177 /* Fills in the page directory bitmap, and the array of page tables bitmap. Both
1178 * of these are based on the number of PDPEs in the system.
1181 int __must_check alloc_gen8_temp_bitmaps(unsigned long **new_pds,
1182 unsigned long **new_pts,
1188 pds = kcalloc(BITS_TO_LONGS(pdpes), sizeof(unsigned long), GFP_TEMPORARY);
1192 pts = kcalloc(pdpes, BITS_TO_LONGS(I915_PDES) * sizeof(unsigned long),
1203 free_gen8_temp_bitmaps(pds, pts);
1207 /* PDE TLBs are a pain to invalidate on GEN8+. When we modify
1208 * the page table structures, we mark them dirty so that
1209 * context switching/execlist queuing code takes extra steps
1210 * to ensure that tlbs are flushed.
1212 static void mark_tlbs_dirty(struct i915_hw_ppgtt *ppgtt)
1214 ppgtt->pd_dirty_rings = INTEL_INFO(ppgtt->base.dev)->ring_mask;
1217 static int gen8_alloc_va_range_3lvl(struct i915_address_space *vm,
1218 struct i915_page_directory_pointer *pdp,
1222 struct i915_hw_ppgtt *ppgtt =
1223 container_of(vm, struct i915_hw_ppgtt, base);
1224 unsigned long *new_page_dirs, *new_page_tables;
1225 struct drm_device *dev = vm->dev;
1226 struct i915_page_directory *pd;
1227 const uint64_t orig_start = start;
1228 const uint64_t orig_length = length;
1230 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1233 /* Wrap is never okay since we can only represent 48b, and we don't
1234 * actually use the other side of the canonical address space.
1236 if (WARN_ON(start + length < start))
1239 if (WARN_ON(start + length > vm->total))
1242 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables, pdpes);
1246 /* Do the allocations first so we can easily bail out */
1247 ret = gen8_ppgtt_alloc_page_directories(vm, pdp, start, length,
1250 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1254 /* For every page directory referenced, allocate page tables */
1255 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1256 ret = gen8_ppgtt_alloc_pagetabs(vm, pd, start, length,
1257 new_page_tables + pdpe * BITS_TO_LONGS(I915_PDES));
1263 length = orig_length;
1265 /* Allocations have completed successfully, so set the bitmaps, and do
1267 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1268 gen8_pde_t *const page_directory = kmap_px(pd);
1269 struct i915_page_table *pt;
1270 uint64_t pd_len = length;
1271 uint64_t pd_start = start;
1274 /* Every pd should be allocated, we just did that above. */
1277 gen8_for_each_pde(pt, pd, pd_start, pd_len, pde) {
1278 /* Same reasoning as pd */
1281 WARN_ON(!gen8_pte_count(pd_start, pd_len));
1283 /* Set our used ptes within the page table */
1284 bitmap_set(pt->used_ptes,
1285 gen8_pte_index(pd_start),
1286 gen8_pte_count(pd_start, pd_len));
1288 /* Our pde is now pointing to the pagetable, pt */
1289 __set_bit(pde, pd->used_pdes);
1291 /* Map the PDE to the page table */
1292 page_directory[pde] = gen8_pde_encode(px_dma(pt),
1294 trace_i915_page_table_entry_map(&ppgtt->base, pde, pt,
1295 gen8_pte_index(start),
1296 gen8_pte_count(start, length),
1299 /* NB: We haven't yet mapped ptes to pages. At this
1300 * point we're still relying on insert_entries() */
1303 kunmap_px(ppgtt, page_directory);
1304 __set_bit(pdpe, pdp->used_pdpes);
1305 gen8_setup_page_directory(ppgtt, pdp, pd, pdpe);
1308 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1309 mark_tlbs_dirty(ppgtt);
1316 for_each_set_bit(temp, new_page_tables + pdpe *
1317 BITS_TO_LONGS(I915_PDES), I915_PDES)
1318 free_pt(dev, pdp->page_directory[pdpe]->page_table[temp]);
1321 for_each_set_bit(pdpe, new_page_dirs, pdpes)
1322 free_pd(dev, pdp->page_directory[pdpe]);
1324 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1325 mark_tlbs_dirty(ppgtt);
1329 static int gen8_alloc_va_range_4lvl(struct i915_address_space *vm,
1330 struct i915_pml4 *pml4,
1334 DECLARE_BITMAP(new_pdps, GEN8_PML4ES_PER_PML4);
1335 struct i915_hw_ppgtt *ppgtt =
1336 container_of(vm, struct i915_hw_ppgtt, base);
1337 struct i915_page_directory_pointer *pdp;
1341 /* Do the pml4 allocations first, so we don't need to track the newly
1342 * allocated tables below the pdp */
1343 bitmap_zero(new_pdps, GEN8_PML4ES_PER_PML4);
1345 /* The pagedirectory and pagetable allocations are done in the shared 3
1346 * and 4 level code. Just allocate the pdps.
1348 ret = gen8_ppgtt_alloc_page_dirpointers(vm, pml4, start, length,
1353 WARN(bitmap_weight(new_pdps, GEN8_PML4ES_PER_PML4) > 2,
1354 "The allocation has spanned more than 512GB. "
1355 "It is highly likely this is incorrect.");
1357 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1360 ret = gen8_alloc_va_range_3lvl(vm, pdp, start, length);
1364 gen8_setup_page_directory_pointer(ppgtt, pml4, pdp, pml4e);
1367 bitmap_or(pml4->used_pml4es, new_pdps, pml4->used_pml4es,
1368 GEN8_PML4ES_PER_PML4);
1373 for_each_set_bit(pml4e, new_pdps, GEN8_PML4ES_PER_PML4)
1374 gen8_ppgtt_cleanup_3lvl(vm->dev, pml4->pdps[pml4e]);
1379 static int gen8_alloc_va_range(struct i915_address_space *vm,
1380 uint64_t start, uint64_t length)
1382 struct i915_hw_ppgtt *ppgtt =
1383 container_of(vm, struct i915_hw_ppgtt, base);
1385 if (USES_FULL_48BIT_PPGTT(vm->dev))
1386 return gen8_alloc_va_range_4lvl(vm, &ppgtt->pml4, start, length);
1388 return gen8_alloc_va_range_3lvl(vm, &ppgtt->pdp, start, length);
1391 static void gen8_dump_pdp(struct i915_page_directory_pointer *pdp,
1392 uint64_t start, uint64_t length,
1393 gen8_pte_t scratch_pte,
1396 struct i915_page_directory *pd;
1399 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1400 struct i915_page_table *pt;
1401 uint64_t pd_len = length;
1402 uint64_t pd_start = start;
1405 if (!test_bit(pdpe, pdp->used_pdpes))
1408 seq_printf(m, "\tPDPE #%d\n", pdpe);
1409 gen8_for_each_pde(pt, pd, pd_start, pd_len, pde) {
1411 gen8_pte_t *pt_vaddr;
1413 if (!test_bit(pde, pd->used_pdes))
1416 pt_vaddr = kmap_px(pt);
1417 for (pte = 0; pte < GEN8_PTES; pte += 4) {
1419 (pdpe << GEN8_PDPE_SHIFT) |
1420 (pde << GEN8_PDE_SHIFT) |
1421 (pte << GEN8_PTE_SHIFT);
1425 for (i = 0; i < 4; i++)
1426 if (pt_vaddr[pte + i] != scratch_pte)
1431 seq_printf(m, "\t\t0x%lx [%03d,%03d,%04d]: =", va, pdpe, pde, pte);
1432 for (i = 0; i < 4; i++) {
1433 if (pt_vaddr[pte + i] != scratch_pte)
1434 seq_printf(m, " %lx", pt_vaddr[pte + i]);
1436 seq_puts(m, " SCRATCH ");
1440 /* don't use kunmap_px, it could trigger
1441 * an unnecessary flush.
1443 kunmap_atomic(pt_vaddr);
1448 static void gen8_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
1450 struct i915_address_space *vm = &ppgtt->base;
1451 uint64_t start = ppgtt->base.start;
1452 uint64_t length = ppgtt->base.total;
1453 gen8_pte_t scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
1454 I915_CACHE_LLC, true);
1456 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
1457 gen8_dump_pdp(&ppgtt->pdp, start, length, scratch_pte, m);
1460 struct i915_pml4 *pml4 = &ppgtt->pml4;
1461 struct i915_page_directory_pointer *pdp;
1463 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1464 if (!test_bit(pml4e, pml4->used_pml4es))
1467 seq_printf(m, " PML4E #%lu\n", pml4e);
1468 gen8_dump_pdp(pdp, start, length, scratch_pte, m);
1473 static int gen8_preallocate_top_level_pdps(struct i915_hw_ppgtt *ppgtt)
1475 unsigned long *new_page_dirs, *new_page_tables;
1476 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1479 /* We allocate temp bitmap for page tables for no gain
1480 * but as this is for init only, lets keep the things simple
1482 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables, pdpes);
1486 /* Allocate for all pdps regardless of how the ppgtt
1489 ret = gen8_ppgtt_alloc_page_directories(&ppgtt->base, &ppgtt->pdp,
1493 *ppgtt->pdp.used_pdpes = *new_page_dirs;
1495 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1501 * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers
1502 * with a net effect resembling a 2-level page table in normal x86 terms. Each
1503 * PDP represents 1GB of memory 4 * 512 * 512 * 4096 = 4GB legacy 32b address
1507 static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
1511 ret = gen8_init_scratch(&ppgtt->base);
1515 ppgtt->base.start = 0;
1516 ppgtt->base.cleanup = gen8_ppgtt_cleanup;
1517 ppgtt->base.allocate_va_range = gen8_alloc_va_range;
1518 ppgtt->base.insert_entries = gen8_ppgtt_insert_entries;
1519 ppgtt->base.clear_range = gen8_ppgtt_clear_range;
1520 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
1521 ppgtt->base.bind_vma = ppgtt_bind_vma;
1522 ppgtt->debug_dump = gen8_dump_ppgtt;
1524 if (USES_FULL_48BIT_PPGTT(ppgtt->base.dev)) {
1525 ret = setup_px(ppgtt->base.dev, &ppgtt->pml4);
1529 gen8_initialize_pml4(&ppgtt->base, &ppgtt->pml4);
1531 ppgtt->base.total = 1ULL << 48;
1532 ppgtt->switch_mm = gen8_48b_mm_switch;
1534 ret = __pdp_init(ppgtt->base.dev, &ppgtt->pdp);
1538 ppgtt->base.total = 1ULL << 32;
1539 ppgtt->switch_mm = gen8_legacy_mm_switch;
1540 trace_i915_page_directory_pointer_entry_alloc(&ppgtt->base,
1544 if (intel_vgpu_active(ppgtt->base.dev)) {
1545 ret = gen8_preallocate_top_level_pdps(ppgtt);
1551 if (intel_vgpu_active(ppgtt->base.dev))
1552 gen8_ppgtt_notify_vgt(ppgtt, true);
1557 gen8_free_scratch(&ppgtt->base);
1561 static void gen6_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
1563 struct i915_address_space *vm = &ppgtt->base;
1564 struct i915_page_table *unused;
1565 gen6_pte_t scratch_pte;
1567 uint32_t pte, pde, temp;
1568 uint32_t start = ppgtt->base.start, length = ppgtt->base.total;
1570 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1571 I915_CACHE_LLC, true, 0);
1573 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde) {
1575 gen6_pte_t *pt_vaddr;
1576 const dma_addr_t pt_addr = px_dma(ppgtt->pd.page_table[pde]);
1577 pd_entry = readl(ppgtt->pd_addr + pde);
1578 expected = (GEN6_PDE_ADDR_ENCODE(pt_addr) | GEN6_PDE_VALID);
1580 if (pd_entry != expected)
1581 seq_printf(m, "\tPDE #%d mismatch: Actual PDE: %x Expected PDE: %x\n",
1585 seq_printf(m, "\tPDE: %x\n", pd_entry);
1587 pt_vaddr = kmap_px(ppgtt->pd.page_table[pde]);
1589 for (pte = 0; pte < GEN6_PTES; pte+=4) {
1591 (pde * PAGE_SIZE * GEN6_PTES) +
1595 for (i = 0; i < 4; i++)
1596 if (pt_vaddr[pte + i] != scratch_pte)
1601 seq_printf(m, "\t\t0x%lx [%03d,%04d]: =", va, pde, pte);
1602 for (i = 0; i < 4; i++) {
1603 if (pt_vaddr[pte + i] != scratch_pte)
1604 seq_printf(m, " %08x", pt_vaddr[pte + i]);
1606 seq_puts(m, " SCRATCH ");
1610 kunmap_px(ppgtt, pt_vaddr);
1614 /* Write pde (index) from the page directory @pd to the page table @pt */
1615 static void gen6_write_pde(struct i915_page_directory *pd,
1616 const int pde, struct i915_page_table *pt)
1618 /* Caller needs to make sure the write completes if necessary */
1619 struct i915_hw_ppgtt *ppgtt =
1620 container_of(pd, struct i915_hw_ppgtt, pd);
1623 pd_entry = GEN6_PDE_ADDR_ENCODE(px_dma(pt));
1624 pd_entry |= GEN6_PDE_VALID;
1626 writel(pd_entry, ppgtt->pd_addr + pde);
1629 /* Write all the page tables found in the ppgtt structure to incrementing page
1631 static void gen6_write_page_range(struct drm_i915_private *dev_priv,
1632 struct i915_page_directory *pd,
1633 uint32_t start, uint32_t length)
1635 struct i915_page_table *pt;
1638 gen6_for_each_pde(pt, pd, start, length, temp, pde)
1639 gen6_write_pde(pd, pde, pt);
1641 /* Make sure write is complete before other code can use this page
1642 * table. Also require for WC mapped PTEs */
1643 readl(dev_priv->gtt.gsm);
1646 static uint32_t get_pd_offset(struct i915_hw_ppgtt *ppgtt)
1648 BUG_ON(ppgtt->pd.base.ggtt_offset & 0x3f);
1650 return (ppgtt->pd.base.ggtt_offset / 64) << 16;
1653 static int hsw_mm_switch(struct i915_hw_ppgtt *ppgtt,
1654 struct drm_i915_gem_request *req)
1656 struct intel_engine_cs *ring = req->ring;
1659 /* NB: TLBs must be flushed and invalidated before a switch */
1660 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1664 ret = intel_ring_begin(req, 6);
1668 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1669 intel_ring_emit_reg(ring, RING_PP_DIR_DCLV(ring));
1670 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1671 intel_ring_emit_reg(ring, RING_PP_DIR_BASE(ring));
1672 intel_ring_emit(ring, get_pd_offset(ppgtt));
1673 intel_ring_emit(ring, MI_NOOP);
1674 intel_ring_advance(ring);
1679 static int vgpu_mm_switch(struct i915_hw_ppgtt *ppgtt,
1680 struct drm_i915_gem_request *req)
1682 struct intel_engine_cs *ring = req->ring;
1683 struct drm_i915_private *dev_priv = to_i915(ppgtt->base.dev);
1685 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
1686 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
1690 static int gen7_mm_switch(struct i915_hw_ppgtt *ppgtt,
1691 struct drm_i915_gem_request *req)
1693 struct intel_engine_cs *ring = req->ring;
1696 /* NB: TLBs must be flushed and invalidated before a switch */
1697 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1701 ret = intel_ring_begin(req, 6);
1705 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1706 intel_ring_emit_reg(ring, RING_PP_DIR_DCLV(ring));
1707 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1708 intel_ring_emit_reg(ring, RING_PP_DIR_BASE(ring));
1709 intel_ring_emit(ring, get_pd_offset(ppgtt));
1710 intel_ring_emit(ring, MI_NOOP);
1711 intel_ring_advance(ring);
1713 /* XXX: RCS is the only one to auto invalidate the TLBs? */
1714 if (ring->id != RCS) {
1715 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1723 static int gen6_mm_switch(struct i915_hw_ppgtt *ppgtt,
1724 struct drm_i915_gem_request *req)
1726 struct intel_engine_cs *ring = req->ring;
1727 struct drm_device *dev = ppgtt->base.dev;
1728 struct drm_i915_private *dev_priv = dev->dev_private;
1731 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
1732 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
1734 POSTING_READ(RING_PP_DIR_DCLV(ring));
1739 static void gen8_ppgtt_enable(struct drm_device *dev)
1741 struct drm_i915_private *dev_priv = dev->dev_private;
1742 struct intel_engine_cs *ring;
1745 for_each_ring(ring, dev_priv, j) {
1746 u32 four_level = USES_FULL_48BIT_PPGTT(dev) ? GEN8_GFX_PPGTT_48B : 0;
1747 I915_WRITE(RING_MODE_GEN7(ring),
1748 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE | four_level));
1752 static void gen7_ppgtt_enable(struct drm_device *dev)
1754 struct drm_i915_private *dev_priv = dev->dev_private;
1755 struct intel_engine_cs *ring;
1756 uint32_t ecochk, ecobits;
1759 ecobits = I915_READ(GAC_ECO_BITS);
1760 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B);
1762 ecochk = I915_READ(GAM_ECOCHK);
1763 if (IS_HASWELL(dev)) {
1764 ecochk |= ECOCHK_PPGTT_WB_HSW;
1766 ecochk |= ECOCHK_PPGTT_LLC_IVB;
1767 ecochk &= ~ECOCHK_PPGTT_GFDT_IVB;
1769 I915_WRITE(GAM_ECOCHK, ecochk);
1771 for_each_ring(ring, dev_priv, i) {
1772 /* GFX_MODE is per-ring on gen7+ */
1773 I915_WRITE(RING_MODE_GEN7(ring),
1774 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1778 static void gen6_ppgtt_enable(struct drm_device *dev)
1780 struct drm_i915_private *dev_priv = dev->dev_private;
1781 uint32_t ecochk, gab_ctl, ecobits;
1783 ecobits = I915_READ(GAC_ECO_BITS);
1784 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_SNB_BIT |
1785 ECOBITS_PPGTT_CACHE64B);
1787 gab_ctl = I915_READ(GAB_CTL);
1788 I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT);
1790 ecochk = I915_READ(GAM_ECOCHK);
1791 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT | ECOCHK_PPGTT_CACHE64B);
1793 I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1796 /* PPGTT support for Sandybdrige/Gen6 and later */
1797 static void gen6_ppgtt_clear_range(struct i915_address_space *vm,
1802 struct i915_hw_ppgtt *ppgtt =
1803 container_of(vm, struct i915_hw_ppgtt, base);
1804 gen6_pte_t *pt_vaddr, scratch_pte;
1805 unsigned first_entry = start >> PAGE_SHIFT;
1806 unsigned num_entries = length >> PAGE_SHIFT;
1807 unsigned act_pt = first_entry / GEN6_PTES;
1808 unsigned first_pte = first_entry % GEN6_PTES;
1809 unsigned last_pte, i;
1811 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1812 I915_CACHE_LLC, true, 0);
1814 while (num_entries) {
1815 last_pte = first_pte + num_entries;
1816 if (last_pte > GEN6_PTES)
1817 last_pte = GEN6_PTES;
1819 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1821 for (i = first_pte; i < last_pte; i++)
1822 pt_vaddr[i] = scratch_pte;
1824 kunmap_px(ppgtt, pt_vaddr);
1826 num_entries -= last_pte - first_pte;
1832 static void gen6_ppgtt_insert_entries(struct i915_address_space *vm,
1833 struct sg_table *pages,
1835 enum i915_cache_level cache_level, u32 flags)
1837 struct i915_hw_ppgtt *ppgtt =
1838 container_of(vm, struct i915_hw_ppgtt, base);
1839 gen6_pte_t *pt_vaddr;
1840 unsigned first_entry = start >> PAGE_SHIFT;
1841 unsigned act_pt = first_entry / GEN6_PTES;
1842 unsigned act_pte = first_entry % GEN6_PTES;
1843 struct sg_page_iter sg_iter;
1846 for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) {
1847 if (pt_vaddr == NULL)
1848 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1851 vm->pte_encode(sg_page_iter_dma_address(&sg_iter),
1852 cache_level, true, flags);
1854 if (++act_pte == GEN6_PTES) {
1855 kunmap_px(ppgtt, pt_vaddr);
1862 kunmap_px(ppgtt, pt_vaddr);
1865 static int gen6_alloc_va_range(struct i915_address_space *vm,
1866 uint64_t start_in, uint64_t length_in)
1868 DECLARE_BITMAP(new_page_tables, I915_PDES);
1869 struct drm_device *dev = vm->dev;
1870 struct drm_i915_private *dev_priv = dev->dev_private;
1871 struct i915_hw_ppgtt *ppgtt =
1872 container_of(vm, struct i915_hw_ppgtt, base);
1873 struct i915_page_table *pt;
1874 uint32_t start, length, start_save, length_save;
1878 if (WARN_ON(start_in + length_in > ppgtt->base.total))
1881 start = start_save = start_in;
1882 length = length_save = length_in;
1884 bitmap_zero(new_page_tables, I915_PDES);
1886 /* The allocation is done in two stages so that we can bail out with
1887 * minimal amount of pain. The first stage finds new page tables that
1888 * need allocation. The second stage marks use ptes within the page
1891 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1892 if (pt != vm->scratch_pt) {
1893 WARN_ON(bitmap_empty(pt->used_ptes, GEN6_PTES));
1897 /* We've already allocated a page table */
1898 WARN_ON(!bitmap_empty(pt->used_ptes, GEN6_PTES));
1906 gen6_initialize_pt(vm, pt);
1908 ppgtt->pd.page_table[pde] = pt;
1909 __set_bit(pde, new_page_tables);
1910 trace_i915_page_table_entry_alloc(vm, pde, start, GEN6_PDE_SHIFT);
1914 length = length_save;
1916 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1917 DECLARE_BITMAP(tmp_bitmap, GEN6_PTES);
1919 bitmap_zero(tmp_bitmap, GEN6_PTES);
1920 bitmap_set(tmp_bitmap, gen6_pte_index(start),
1921 gen6_pte_count(start, length));
1923 if (__test_and_clear_bit(pde, new_page_tables))
1924 gen6_write_pde(&ppgtt->pd, pde, pt);
1926 trace_i915_page_table_entry_map(vm, pde, pt,
1927 gen6_pte_index(start),
1928 gen6_pte_count(start, length),
1930 bitmap_or(pt->used_ptes, tmp_bitmap, pt->used_ptes,
1934 WARN_ON(!bitmap_empty(new_page_tables, I915_PDES));
1936 /* Make sure write is complete before other code can use this page
1937 * table. Also require for WC mapped PTEs */
1938 readl(dev_priv->gtt.gsm);
1940 mark_tlbs_dirty(ppgtt);
1944 for_each_set_bit(pde, new_page_tables, I915_PDES) {
1945 struct i915_page_table *pt = ppgtt->pd.page_table[pde];
1947 ppgtt->pd.page_table[pde] = vm->scratch_pt;
1948 free_pt(vm->dev, pt);
1951 mark_tlbs_dirty(ppgtt);
1955 static int gen6_init_scratch(struct i915_address_space *vm)
1957 struct drm_device *dev = vm->dev;
1959 vm->scratch_page = alloc_scratch_page(dev);
1960 if (IS_ERR(vm->scratch_page))
1961 return PTR_ERR(vm->scratch_page);
1963 vm->scratch_pt = alloc_pt(dev);
1964 if (IS_ERR(vm->scratch_pt)) {
1965 free_scratch_page(dev, vm->scratch_page);
1966 return PTR_ERR(vm->scratch_pt);
1969 gen6_initialize_pt(vm, vm->scratch_pt);
1974 static void gen6_free_scratch(struct i915_address_space *vm)
1976 struct drm_device *dev = vm->dev;
1978 free_pt(dev, vm->scratch_pt);
1979 free_scratch_page(dev, vm->scratch_page);
1982 static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
1984 struct i915_hw_ppgtt *ppgtt =
1985 container_of(vm, struct i915_hw_ppgtt, base);
1986 struct i915_page_table *pt;
1989 drm_mm_remove_node(&ppgtt->node);
1991 gen6_for_all_pdes(pt, ppgtt, pde) {
1992 if (pt != vm->scratch_pt)
1993 free_pt(ppgtt->base.dev, pt);
1996 gen6_free_scratch(vm);
1999 static int gen6_ppgtt_allocate_page_directories(struct i915_hw_ppgtt *ppgtt)
2001 struct i915_address_space *vm = &ppgtt->base;
2002 struct drm_device *dev = ppgtt->base.dev;
2003 struct drm_i915_private *dev_priv = dev->dev_private;
2004 bool retried = false;
2007 /* PPGTT PDEs reside in the GGTT and consists of 512 entries. The
2008 * allocator works in address space sizes, so it's multiplied by page
2009 * size. We allocate at the top of the GTT to avoid fragmentation.
2011 BUG_ON(!drm_mm_initialized(&dev_priv->gtt.base.mm));
2013 ret = gen6_init_scratch(vm);
2018 ret = drm_mm_insert_node_in_range_generic(&dev_priv->gtt.base.mm,
2019 &ppgtt->node, GEN6_PD_SIZE,
2021 0, dev_priv->gtt.base.total,
2023 if (ret == -ENOSPC && !retried) {
2024 ret = i915_gem_evict_something(dev, &dev_priv->gtt.base,
2025 GEN6_PD_SIZE, GEN6_PD_ALIGN,
2027 0, dev_priv->gtt.base.total,
2040 if (ppgtt->node.start < dev_priv->gtt.mappable_end)
2041 DRM_DEBUG("Forced to use aperture for PDEs\n");
2046 gen6_free_scratch(vm);
2050 static int gen6_ppgtt_alloc(struct i915_hw_ppgtt *ppgtt)
2052 return gen6_ppgtt_allocate_page_directories(ppgtt);
2055 static void gen6_scratch_va_range(struct i915_hw_ppgtt *ppgtt,
2056 uint64_t start, uint64_t length)
2058 struct i915_page_table *unused;
2061 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde)
2062 ppgtt->pd.page_table[pde] = ppgtt->base.scratch_pt;
2065 static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
2067 struct drm_device *dev = ppgtt->base.dev;
2068 struct drm_i915_private *dev_priv = dev->dev_private;
2071 ppgtt->base.pte_encode = dev_priv->gtt.base.pte_encode;
2073 ppgtt->switch_mm = gen6_mm_switch;
2074 } else if (IS_HASWELL(dev)) {
2075 ppgtt->switch_mm = hsw_mm_switch;
2076 } else if (IS_GEN7(dev)) {
2077 ppgtt->switch_mm = gen7_mm_switch;
2081 if (intel_vgpu_active(dev))
2082 ppgtt->switch_mm = vgpu_mm_switch;
2084 ret = gen6_ppgtt_alloc(ppgtt);
2088 ppgtt->base.allocate_va_range = gen6_alloc_va_range;
2089 ppgtt->base.clear_range = gen6_ppgtt_clear_range;
2090 ppgtt->base.insert_entries = gen6_ppgtt_insert_entries;
2091 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
2092 ppgtt->base.bind_vma = ppgtt_bind_vma;
2093 ppgtt->base.cleanup = gen6_ppgtt_cleanup;
2094 ppgtt->base.start = 0;
2095 ppgtt->base.total = I915_PDES * GEN6_PTES * PAGE_SIZE;
2096 ppgtt->debug_dump = gen6_dump_ppgtt;
2098 ppgtt->pd.base.ggtt_offset =
2099 ppgtt->node.start / PAGE_SIZE * sizeof(gen6_pte_t);
2101 ppgtt->pd_addr = (gen6_pte_t __iomem *)dev_priv->gtt.gsm +
2102 ppgtt->pd.base.ggtt_offset / sizeof(gen6_pte_t);
2104 gen6_scratch_va_range(ppgtt, 0, ppgtt->base.total);
2106 gen6_write_page_range(dev_priv, &ppgtt->pd, 0, ppgtt->base.total);
2108 DRM_DEBUG_DRIVER("Allocated pde space (%lldM) at GTT entry: %llx\n",
2109 ppgtt->node.size >> 20,
2110 ppgtt->node.start / PAGE_SIZE);
2112 DRM_DEBUG("Adding PPGTT at offset %x\n",
2113 ppgtt->pd.base.ggtt_offset << 10);
2118 static int __hw_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
2120 ppgtt->base.dev = dev;
2122 if (INTEL_INFO(dev)->gen < 8)
2123 return gen6_ppgtt_init(ppgtt);
2125 return gen8_ppgtt_init(ppgtt);
2128 static void i915_address_space_init(struct i915_address_space *vm,
2129 struct drm_i915_private *dev_priv)
2131 drm_mm_init(&vm->mm, vm->start, vm->total);
2132 vm->dev = dev_priv->dev;
2133 INIT_LIST_HEAD(&vm->active_list);
2134 INIT_LIST_HEAD(&vm->inactive_list);
2135 list_add_tail(&vm->global_link, &dev_priv->vm_list);
2138 static void gtt_write_workarounds(struct drm_device *dev)
2140 struct drm_i915_private *dev_priv = dev->dev_private;
2142 /* This function is for gtt related workarounds. This function is
2143 * called on driver load and after a GPU reset, so you can place
2144 * workarounds here even if they get overwritten by GPU reset.
2146 /* WaIncreaseDefaultTLBEntries:chv,bdw,skl,bxt */
2147 if (IS_BROADWELL(dev))
2148 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN8_L3_LRA_1_GPGPU_DEFAULT_VALUE_BDW);
2149 else if (IS_CHERRYVIEW(dev))
2150 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN8_L3_LRA_1_GPGPU_DEFAULT_VALUE_CHV);
2151 else if (IS_SKYLAKE(dev))
2152 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN9_L3_LRA_1_GPGPU_DEFAULT_VALUE_SKL);
2153 else if (IS_BROXTON(dev))
2154 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN9_L3_LRA_1_GPGPU_DEFAULT_VALUE_BXT);
2157 int i915_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
2159 struct drm_i915_private *dev_priv = dev->dev_private;
2162 ret = __hw_ppgtt_init(dev, ppgtt);
2164 kref_init(&ppgtt->ref);
2165 i915_address_space_init(&ppgtt->base, dev_priv);
2171 int i915_ppgtt_init_hw(struct drm_device *dev)
2173 gtt_write_workarounds(dev);
2175 /* In the case of execlists, PPGTT is enabled by the context descriptor
2176 * and the PDPs are contained within the context itself. We don't
2177 * need to do anything here. */
2178 if (i915.enable_execlists)
2181 if (!USES_PPGTT(dev))
2185 gen6_ppgtt_enable(dev);
2186 else if (IS_GEN7(dev))
2187 gen7_ppgtt_enable(dev);
2188 else if (INTEL_INFO(dev)->gen >= 8)
2189 gen8_ppgtt_enable(dev);
2191 MISSING_CASE(INTEL_INFO(dev)->gen);
2196 int i915_ppgtt_init_ring(struct drm_i915_gem_request *req)
2198 struct drm_i915_private *dev_priv = req->ring->dev->dev_private;
2199 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
2201 if (i915.enable_execlists)
2207 return ppgtt->switch_mm(ppgtt, req);
2210 struct i915_hw_ppgtt *
2211 i915_ppgtt_create(struct drm_device *dev, struct drm_i915_file_private *fpriv)
2213 struct i915_hw_ppgtt *ppgtt;
2216 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2218 return ERR_PTR(-ENOMEM);
2220 ret = i915_ppgtt_init(dev, ppgtt);
2223 return ERR_PTR(ret);
2226 ppgtt->file_priv = fpriv;
2228 trace_i915_ppgtt_create(&ppgtt->base);
2233 void i915_ppgtt_release(struct kref *kref)
2235 struct i915_hw_ppgtt *ppgtt =
2236 container_of(kref, struct i915_hw_ppgtt, ref);
2238 trace_i915_ppgtt_release(&ppgtt->base);
2240 /* vmas should already be unbound */
2241 WARN_ON(!list_empty(&ppgtt->base.active_list));
2242 WARN_ON(!list_empty(&ppgtt->base.inactive_list));
2244 list_del(&ppgtt->base.global_link);
2245 drm_mm_takedown(&ppgtt->base.mm);
2247 ppgtt->base.cleanup(&ppgtt->base);
2251 extern int intel_iommu_gfx_mapped;
2252 /* Certain Gen5 chipsets require require idling the GPU before
2253 * unmapping anything from the GTT when VT-d is enabled.
2255 static bool needs_idle_maps(struct drm_device *dev)
2257 #ifdef CONFIG_INTEL_IOMMU
2258 /* Query intel_iommu to see if we need the workaround. Presumably that
2261 if (IS_GEN5(dev) && IS_MOBILE(dev) && intel_iommu_gfx_mapped)
2267 static bool do_idling(struct drm_i915_private *dev_priv)
2269 bool ret = dev_priv->mm.interruptible;
2271 if (unlikely(dev_priv->gtt.do_idle_maps)) {
2272 dev_priv->mm.interruptible = false;
2273 if (i915_gpu_idle(dev_priv->dev)) {
2274 DRM_ERROR("Couldn't idle GPU\n");
2275 /* Wait a bit, in hopes it avoids the hang */
2283 static void undo_idling(struct drm_i915_private *dev_priv, bool interruptible)
2285 if (unlikely(dev_priv->gtt.do_idle_maps))
2286 dev_priv->mm.interruptible = interruptible;
2289 void i915_check_and_clear_faults(struct drm_device *dev)
2291 struct drm_i915_private *dev_priv = dev->dev_private;
2292 struct intel_engine_cs *ring;
2295 if (INTEL_INFO(dev)->gen < 6)
2298 for_each_ring(ring, dev_priv, i) {
2300 fault_reg = I915_READ(RING_FAULT_REG(ring));
2301 if (fault_reg & RING_FAULT_VALID) {
2303 DRM_DEBUG_DRIVER("Unexpected fault\n"
2305 "\tAddress space: %s\n"
2308 fault_reg & PAGE_MASK,
2309 fault_reg & RING_FAULT_GTTSEL_MASK ? "GGTT" : "PPGTT",
2310 RING_FAULT_SRCID(fault_reg),
2311 RING_FAULT_FAULT_TYPE(fault_reg));
2313 I915_WRITE(RING_FAULT_REG(ring),
2314 fault_reg & ~RING_FAULT_VALID);
2317 POSTING_READ(RING_FAULT_REG(&dev_priv->ring[RCS]));
2320 static void i915_ggtt_flush(struct drm_i915_private *dev_priv)
2322 if (INTEL_INFO(dev_priv->dev)->gen < 6) {
2323 intel_gtt_chipset_flush();
2325 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2326 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2330 void i915_gem_suspend_gtt_mappings(struct drm_device *dev)
2332 struct drm_i915_private *dev_priv = dev->dev_private;
2334 /* Don't bother messing with faults pre GEN6 as we have little
2335 * documentation supporting that it's a good idea.
2337 if (INTEL_INFO(dev)->gen < 6)
2340 i915_check_and_clear_faults(dev);
2342 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
2343 dev_priv->gtt.base.start,
2344 dev_priv->gtt.base.total,
2347 i915_ggtt_flush(dev_priv);
2350 int i915_gem_gtt_prepare_object(struct drm_i915_gem_object *obj)
2352 if (!dma_map_sg(&obj->base.dev->pdev->dev,
2353 obj->pages->sgl, obj->pages->nents,
2354 PCI_DMA_BIDIRECTIONAL))
2360 static void gen8_set_pte(void __iomem *addr, gen8_pte_t pte)
2365 iowrite32((u32)pte, addr);
2366 iowrite32(pte >> 32, addr + 4);
2370 static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
2371 struct sg_table *st,
2373 enum i915_cache_level level, u32 unused)
2375 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2376 unsigned first_entry = start >> PAGE_SHIFT;
2377 gen8_pte_t __iomem *gtt_entries =
2378 (gen8_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
2380 struct sg_page_iter sg_iter;
2381 dma_addr_t addr = 0; /* shut up gcc */
2384 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2386 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
2387 addr = sg_dma_address(sg_iter.sg) +
2388 (sg_iter.sg_pgoffset << PAGE_SHIFT);
2389 gen8_set_pte(>t_entries[i],
2390 gen8_pte_encode(addr, level, true));
2395 * XXX: This serves as a posting read to make sure that the PTE has
2396 * actually been updated. There is some concern that even though
2397 * registers and PTEs are within the same BAR that they are potentially
2398 * of NUMA access patterns. Therefore, even with the way we assume
2399 * hardware should work, we must keep this posting read for paranoia.
2402 WARN_ON(readq(>t_entries[i-1])
2403 != gen8_pte_encode(addr, level, true));
2405 /* This next bit makes the above posting read even more important. We
2406 * want to flush the TLBs only after we're certain all the PTE updates
2409 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2410 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2412 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2415 struct insert_entries {
2416 struct i915_address_space *vm;
2417 struct sg_table *st;
2419 enum i915_cache_level level;
2423 static int gen8_ggtt_insert_entries__cb(void *_arg)
2425 struct insert_entries *arg = _arg;
2426 gen8_ggtt_insert_entries(arg->vm, arg->st,
2427 arg->start, arg->level, arg->flags);
2431 static void gen8_ggtt_insert_entries__BKL(struct i915_address_space *vm,
2432 struct sg_table *st,
2434 enum i915_cache_level level,
2437 struct insert_entries arg = { vm, st, start, level, flags };
2438 #ifndef __DragonFly__
2439 stop_machine(gen8_ggtt_insert_entries__cb, &arg, NULL);
2441 /* XXX: is this enough ?
2442 * See Linux commit 5bab6f60cb4d1417ad7c599166bcfec87529c1a2 */
2444 gen8_ggtt_insert_entries__cb(&arg);
2450 * Binds an object into the global gtt with the specified cache level. The object
2451 * will be accessible to the GPU via commands whose operands reference offsets
2452 * within the global GTT as well as accessible by the GPU through the GMADR
2453 * mapped BAR (dev_priv->mm.gtt->gtt).
2455 static void gen6_ggtt_insert_entries(struct i915_address_space *vm,
2456 struct sg_table *st,
2458 enum i915_cache_level level, u32 flags)
2460 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2461 unsigned first_entry = start >> PAGE_SHIFT;
2462 gen6_pte_t __iomem *gtt_entries =
2463 (gen6_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
2465 struct sg_page_iter sg_iter;
2466 dma_addr_t addr = 0;
2469 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2471 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
2472 addr = sg_page_iter_dma_address(&sg_iter);
2473 iowrite32(vm->pte_encode(addr, level, true, flags), >t_entries[i]);
2477 /* XXX: This serves as a posting read to make sure that the PTE has
2478 * actually been updated. There is some concern that even though
2479 * registers and PTEs are within the same BAR that they are potentially
2480 * of NUMA access patterns. Therefore, even with the way we assume
2481 * hardware should work, we must keep this posting read for paranoia.
2484 unsigned long gtt = readl(>t_entries[i-1]);
2485 WARN_ON(gtt != vm->pte_encode(addr, level, true, flags));
2488 /* This next bit makes the above posting read even more important. We
2489 * want to flush the TLBs only after we're certain all the PTE updates
2492 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2493 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2495 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2498 static void gen8_ggtt_clear_range(struct i915_address_space *vm,
2503 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2504 unsigned first_entry = start >> PAGE_SHIFT;
2505 unsigned num_entries = length >> PAGE_SHIFT;
2506 gen8_pte_t scratch_pte, __iomem *gtt_base =
2507 (gen8_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
2508 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
2512 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2514 if (WARN(num_entries > max_entries,
2515 "First entry = %d; Num entries = %d (max=%d)\n",
2516 first_entry, num_entries, max_entries))
2517 num_entries = max_entries;
2519 scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
2522 for (i = 0; i < num_entries; i++)
2523 gen8_set_pte(>t_base[i], scratch_pte);
2526 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2529 static void gen6_ggtt_clear_range(struct i915_address_space *vm,
2534 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2535 unsigned first_entry = start >> PAGE_SHIFT;
2536 unsigned num_entries = length >> PAGE_SHIFT;
2537 gen6_pte_t scratch_pte, __iomem *gtt_base =
2538 (gen6_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
2539 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
2543 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2545 if (WARN(num_entries > max_entries,
2546 "First entry = %d; Num entries = %d (max=%d)\n",
2547 first_entry, num_entries, max_entries))
2548 num_entries = max_entries;
2550 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
2551 I915_CACHE_LLC, use_scratch, 0);
2553 for (i = 0; i < num_entries; i++)
2554 iowrite32(scratch_pte, >t_base[i]);
2557 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2560 static void i915_ggtt_insert_entries(struct i915_address_space *vm,
2561 struct sg_table *pages,
2563 enum i915_cache_level cache_level, u32 unused)
2565 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2566 unsigned int flags = (cache_level == I915_CACHE_NONE) ?
2567 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
2570 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2572 intel_gtt_insert_sg_entries(pages, start >> PAGE_SHIFT, flags);
2574 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2578 static void i915_ggtt_clear_range(struct i915_address_space *vm,
2583 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2584 unsigned first_entry = start >> PAGE_SHIFT;
2585 unsigned num_entries = length >> PAGE_SHIFT;
2588 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2590 intel_gtt_clear_range(first_entry, num_entries);
2592 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2595 static int ggtt_bind_vma(struct i915_vma *vma,
2596 enum i915_cache_level cache_level,
2599 struct drm_i915_gem_object *obj = vma->obj;
2603 ret = i915_get_ggtt_vma_pages(vma);
2607 /* Currently applicable only to VLV */
2609 pte_flags |= PTE_READ_ONLY;
2611 vma->vm->insert_entries(vma->vm, vma->ggtt_view.pages,
2613 cache_level, pte_flags);
2616 * Without aliasing PPGTT there's no difference between
2617 * GLOBAL/LOCAL_BIND, it's all the same ptes. Hence unconditionally
2618 * upgrade to both bound if we bind either to avoid double-binding.
2620 vma->bound |= GLOBAL_BIND | LOCAL_BIND;
2625 static int aliasing_gtt_bind_vma(struct i915_vma *vma,
2626 enum i915_cache_level cache_level,
2629 struct drm_device *dev = vma->vm->dev;
2630 struct drm_i915_private *dev_priv = dev->dev_private;
2631 struct drm_i915_gem_object *obj = vma->obj;
2632 struct sg_table *pages = obj->pages;
2636 ret = i915_get_ggtt_vma_pages(vma);
2639 pages = vma->ggtt_view.pages;
2641 /* Currently applicable only to VLV */
2643 pte_flags |= PTE_READ_ONLY;
2646 if (flags & GLOBAL_BIND) {
2647 vma->vm->insert_entries(vma->vm, pages,
2649 cache_level, pte_flags);
2652 if (flags & LOCAL_BIND) {
2653 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
2654 appgtt->base.insert_entries(&appgtt->base, pages,
2656 cache_level, pte_flags);
2662 static void ggtt_unbind_vma(struct i915_vma *vma)
2664 struct drm_device *dev = vma->vm->dev;
2665 struct drm_i915_private *dev_priv = dev->dev_private;
2666 struct drm_i915_gem_object *obj = vma->obj;
2667 const uint64_t size = min_t(uint64_t,
2671 if (vma->bound & GLOBAL_BIND) {
2672 vma->vm->clear_range(vma->vm,
2678 if (dev_priv->mm.aliasing_ppgtt && vma->bound & LOCAL_BIND) {
2679 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
2681 appgtt->base.clear_range(&appgtt->base,
2688 void i915_gem_gtt_finish_object(struct drm_i915_gem_object *obj)
2690 struct drm_device *dev = obj->base.dev;
2691 struct drm_i915_private *dev_priv = dev->dev_private;
2694 interruptible = do_idling(dev_priv);
2696 dma_unmap_sg(&dev->pdev->dev, obj->pages->sgl, obj->pages->nents,
2697 PCI_DMA_BIDIRECTIONAL);
2699 undo_idling(dev_priv, interruptible);
2702 static void i915_gtt_color_adjust(struct drm_mm_node *node,
2703 unsigned long color,
2707 if (node->color != color)
2710 if (!list_empty(&node->node_list)) {
2711 node = list_entry(node->node_list.next,
2714 if (node->allocated && node->color != color)
2719 static int i915_gem_setup_global_gtt(struct drm_device *dev,
2724 /* Let GEM Manage all of the aperture.
2726 * However, leave one page at the end still bound to the scratch page.
2727 * There are a number of places where the hardware apparently prefetches
2728 * past the end of the object, and we've seen multiple hangs with the
2729 * GPU head pointer stuck in a batchbuffer bound at the last page of the
2730 * aperture. One page should be enough to keep any prefetching inside
2733 struct drm_i915_private *dev_priv = dev->dev_private;
2734 struct i915_address_space *ggtt_vm = &dev_priv->gtt.base;
2735 struct drm_mm_node *entry;
2736 struct drm_i915_gem_object *obj;
2737 unsigned long hole_start, hole_end;
2739 unsigned long mappable;
2742 mappable = min(end, mappable_end) - start;
2743 BUG_ON(mappable_end > end);
2745 ggtt_vm->start = start;
2747 /* Subtract the guard page before address space initialization to
2748 * shrink the range used by drm_mm */
2749 ggtt_vm->total = end - start - PAGE_SIZE;
2750 i915_address_space_init(ggtt_vm, dev_priv);
2751 ggtt_vm->total += PAGE_SIZE;
2753 if (intel_vgpu_active(dev)) {
2754 ret = intel_vgt_balloon(dev);
2760 ggtt_vm->mm.color_adjust = i915_gtt_color_adjust;
2762 /* Mark any preallocated objects as occupied */
2763 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
2764 struct i915_vma *vma = i915_gem_obj_to_vma(obj, ggtt_vm);
2766 DRM_DEBUG_KMS("reserving preallocated space: %llx + %zx\n",
2767 i915_gem_obj_ggtt_offset(obj), obj->base.size);
2769 WARN_ON(i915_gem_obj_ggtt_bound(obj));
2770 ret = drm_mm_reserve_node(&ggtt_vm->mm, &vma->node);
2772 DRM_DEBUG_KMS("Reservation failed: %i\n", ret);
2775 vma->bound |= GLOBAL_BIND;
2776 __i915_vma_set_map_and_fenceable(vma);
2777 list_add_tail(&vma->vm_link, &ggtt_vm->inactive_list);
2780 /* Clear any non-preallocated blocks */
2781 drm_mm_for_each_hole(entry, &ggtt_vm->mm, hole_start, hole_end) {
2782 DRM_DEBUG_KMS("clearing unused GTT space: [%lx, %lx]\n",
2783 hole_start, hole_end);
2784 ggtt_vm->clear_range(ggtt_vm, hole_start,
2785 hole_end - hole_start, true);
2788 #ifdef __DragonFly__
2789 device_printf(dev->dev->bsddev,
2790 "taking over the fictitious range 0x%llx-0x%llx\n",
2791 dev_priv->gtt.mappable_base + start, dev_priv->gtt.mappable_base + start + mappable);
2792 error = -vm_phys_fictitious_reg_range(dev_priv->gtt.mappable_base + start,
2793 dev_priv->gtt.mappable_base + start + mappable, VM_MEMATTR_WRITE_COMBINING);
2796 /* And finally clear the reserved guard page */
2797 ggtt_vm->clear_range(ggtt_vm, end - PAGE_SIZE, PAGE_SIZE, true);
2799 if (USES_PPGTT(dev) && !USES_FULL_PPGTT(dev)) {
2800 struct i915_hw_ppgtt *ppgtt;
2802 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2806 ret = __hw_ppgtt_init(dev, ppgtt);
2808 ppgtt->base.cleanup(&ppgtt->base);
2813 if (ppgtt->base.allocate_va_range)
2814 ret = ppgtt->base.allocate_va_range(&ppgtt->base, 0,
2817 ppgtt->base.cleanup(&ppgtt->base);
2822 ppgtt->base.clear_range(&ppgtt->base,
2827 dev_priv->mm.aliasing_ppgtt = ppgtt;
2828 WARN_ON(dev_priv->gtt.base.bind_vma != ggtt_bind_vma);
2829 dev_priv->gtt.base.bind_vma = aliasing_gtt_bind_vma;
2835 void i915_gem_init_global_gtt(struct drm_device *dev)
2837 struct drm_i915_private *dev_priv = dev->dev_private;
2838 u64 gtt_size, mappable_size;
2840 gtt_size = dev_priv->gtt.base.total;
2841 mappable_size = dev_priv->gtt.mappable_end;
2843 i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
2846 void i915_global_gtt_cleanup(struct drm_device *dev)
2848 struct drm_i915_private *dev_priv = dev->dev_private;
2849 struct i915_address_space *vm = &dev_priv->gtt.base;
2851 if (dev_priv->mm.aliasing_ppgtt) {
2852 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
2854 ppgtt->base.cleanup(&ppgtt->base);
2857 i915_gem_cleanup_stolen(dev);
2859 if (drm_mm_initialized(&vm->mm)) {
2860 if (intel_vgpu_active(dev))
2861 intel_vgt_deballoon();
2863 drm_mm_takedown(&vm->mm);
2864 list_del(&vm->global_link);
2870 static unsigned int gen6_get_total_gtt_size(u16 snb_gmch_ctl)
2872 snb_gmch_ctl >>= SNB_GMCH_GGMS_SHIFT;
2873 snb_gmch_ctl &= SNB_GMCH_GGMS_MASK;
2874 return snb_gmch_ctl << 20;
2877 static unsigned int gen8_get_total_gtt_size(u16 bdw_gmch_ctl)
2879 bdw_gmch_ctl >>= BDW_GMCH_GGMS_SHIFT;
2880 bdw_gmch_ctl &= BDW_GMCH_GGMS_MASK;
2882 bdw_gmch_ctl = 1 << bdw_gmch_ctl;
2884 #ifdef CONFIG_X86_32
2885 /* Limit 32b platforms to a 2GB GGTT: 4 << 20 / pte size * PAGE_SIZE */
2886 if (bdw_gmch_ctl > 4)
2890 return bdw_gmch_ctl << 20;
2893 static unsigned int chv_get_total_gtt_size(u16 gmch_ctrl)
2895 gmch_ctrl >>= SNB_GMCH_GGMS_SHIFT;
2896 gmch_ctrl &= SNB_GMCH_GGMS_MASK;
2899 return 1 << (20 + gmch_ctrl);
2904 static size_t gen6_get_stolen_size(u16 snb_gmch_ctl)
2906 snb_gmch_ctl >>= SNB_GMCH_GMS_SHIFT;
2907 snb_gmch_ctl &= SNB_GMCH_GMS_MASK;
2908 return snb_gmch_ctl << 25; /* 32 MB units */
2911 static size_t gen8_get_stolen_size(u16 bdw_gmch_ctl)
2913 bdw_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2914 bdw_gmch_ctl &= BDW_GMCH_GMS_MASK;
2915 return bdw_gmch_ctl << 25; /* 32 MB units */
2918 static size_t chv_get_stolen_size(u16 gmch_ctrl)
2920 gmch_ctrl >>= SNB_GMCH_GMS_SHIFT;
2921 gmch_ctrl &= SNB_GMCH_GMS_MASK;
2924 * 0x0 to 0x10: 32MB increments starting at 0MB
2925 * 0x11 to 0x16: 4MB increments starting at 8MB
2926 * 0x17 to 0x1d: 4MB increments start at 36MB
2928 if (gmch_ctrl < 0x11)
2929 return gmch_ctrl << 25;
2930 else if (gmch_ctrl < 0x17)
2931 return (gmch_ctrl - 0x11 + 2) << 22;
2933 return (gmch_ctrl - 0x17 + 9) << 22;
2936 static size_t gen9_get_stolen_size(u16 gen9_gmch_ctl)
2938 gen9_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2939 gen9_gmch_ctl &= BDW_GMCH_GMS_MASK;
2941 if (gen9_gmch_ctl < 0xf0)
2942 return gen9_gmch_ctl << 25; /* 32 MB units */
2944 /* 4MB increments starting at 0xf0 for 4MB */
2945 return (gen9_gmch_ctl - 0xf0 + 1) << 22;
2948 static int ggtt_probe_common(struct drm_device *dev,
2951 struct drm_i915_private *dev_priv = dev->dev_private;
2952 struct i915_page_scratch *scratch_page;
2953 phys_addr_t gtt_phys_addr;
2955 /* For Modern GENs the PTEs and register space are split in the BAR */
2956 gtt_phys_addr = pci_resource_start(dev->pdev, 0) +
2957 (pci_resource_len(dev->pdev, 0) / 2);
2960 * On BXT writes larger than 64 bit to the GTT pagetable range will be
2961 * dropped. For WC mappings in general we have 64 byte burst writes
2962 * when the WC buffer is flushed, so we can't use it, but have to
2963 * resort to an uncached mapping. The WC issue is easily caught by the
2964 * readback check when writing GTT PTE entries.
2966 if (IS_BROXTON(dev))
2967 dev_priv->gtt.gsm = ioremap_nocache(gtt_phys_addr, gtt_size);
2969 dev_priv->gtt.gsm = ioremap_wc(gtt_phys_addr, gtt_size);
2970 if (!dev_priv->gtt.gsm) {
2971 DRM_ERROR("Failed to map the gtt page table\n");
2975 scratch_page = alloc_scratch_page(dev);
2976 if (IS_ERR(scratch_page)) {
2977 DRM_ERROR("Scratch setup failed\n");
2978 /* iounmap will also get called at remove, but meh */
2979 iounmap(dev_priv->gtt.gsm);
2980 return PTR_ERR(scratch_page);
2983 dev_priv->gtt.base.scratch_page = scratch_page;
2988 /* The GGTT and PPGTT need a private PPAT setup in order to handle cacheability
2989 * bits. When using advanced contexts each context stores its own PAT, but
2990 * writing this data shouldn't be harmful even in those cases. */
2991 static void bdw_setup_private_ppat(struct drm_i915_private *dev_priv)
2995 pat = GEN8_PPAT(0, GEN8_PPAT_WB | GEN8_PPAT_LLC) | /* for normal objects, no eLLC */
2996 GEN8_PPAT(1, GEN8_PPAT_WC | GEN8_PPAT_LLCELLC) | /* for something pointing to ptes? */
2997 GEN8_PPAT(2, GEN8_PPAT_WT | GEN8_PPAT_LLCELLC) | /* for scanout with eLLC */
2998 GEN8_PPAT(3, GEN8_PPAT_UC) | /* Uncached objects, mostly for scanout */
2999 GEN8_PPAT(4, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(0)) |
3000 GEN8_PPAT(5, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(1)) |
3001 GEN8_PPAT(6, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(2)) |
3002 GEN8_PPAT(7, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3));
3004 if (!USES_PPGTT(dev_priv->dev))
3005 /* Spec: "For GGTT, there is NO pat_sel[2:0] from the entry,
3006 * so RTL will always use the value corresponding to
3008 * So let's disable cache for GGTT to avoid screen corruptions.
3009 * MOCS still can be used though.
3010 * - System agent ggtt writes (i.e. cpu gtt mmaps) already work
3011 * before this patch, i.e. the same uncached + snooping access
3012 * like on gen6/7 seems to be in effect.
3013 * - So this just fixes blitter/render access. Again it looks
3014 * like it's not just uncached access, but uncached + snooping.
3015 * So we can still hold onto all our assumptions wrt cpu
3016 * clflushing on LLC machines.
3018 pat = GEN8_PPAT(0, GEN8_PPAT_UC);
3020 /* XXX: spec defines this as 2 distinct registers. It's unclear if a 64b
3021 * write would work. */
3022 I915_WRITE(GEN8_PRIVATE_PAT_LO, pat);
3023 I915_WRITE(GEN8_PRIVATE_PAT_HI, pat >> 32);
3026 static void chv_setup_private_ppat(struct drm_i915_private *dev_priv)
3031 * Map WB on BDW to snooped on CHV.
3033 * Only the snoop bit has meaning for CHV, the rest is
3036 * The hardware will never snoop for certain types of accesses:
3037 * - CPU GTT (GMADR->GGTT->no snoop->memory)
3038 * - PPGTT page tables
3039 * - some other special cycles
3041 * As with BDW, we also need to consider the following for GT accesses:
3042 * "For GGTT, there is NO pat_sel[2:0] from the entry,
3043 * so RTL will always use the value corresponding to
3045 * Which means we must set the snoop bit in PAT entry 0
3046 * in order to keep the global status page working.
3048 pat = GEN8_PPAT(0, CHV_PPAT_SNOOP) |
3052 GEN8_PPAT(4, CHV_PPAT_SNOOP) |
3053 GEN8_PPAT(5, CHV_PPAT_SNOOP) |
3054 GEN8_PPAT(6, CHV_PPAT_SNOOP) |
3055 GEN8_PPAT(7, CHV_PPAT_SNOOP);
3057 I915_WRITE(GEN8_PRIVATE_PAT_LO, pat);
3058 I915_WRITE(GEN8_PRIVATE_PAT_HI, pat >> 32);
3061 static int gen8_gmch_probe(struct drm_device *dev,
3064 phys_addr_t *mappable_base,
3067 struct drm_i915_private *dev_priv = dev->dev_private;
3072 /* TODO: We're not aware of mappable constraints on gen8 yet */
3073 *mappable_base = pci_resource_start(dev->pdev, 2);
3074 *mappable_end = pci_resource_len(dev->pdev, 2);
3077 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(39)))
3078 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(39));
3081 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
3083 if (INTEL_INFO(dev)->gen >= 9) {
3084 *stolen = gen9_get_stolen_size(snb_gmch_ctl);
3085 gtt_size = gen8_get_total_gtt_size(snb_gmch_ctl);
3086 } else if (IS_CHERRYVIEW(dev)) {
3087 *stolen = chv_get_stolen_size(snb_gmch_ctl);
3088 gtt_size = chv_get_total_gtt_size(snb_gmch_ctl);
3090 *stolen = gen8_get_stolen_size(snb_gmch_ctl);
3091 gtt_size = gen8_get_total_gtt_size(snb_gmch_ctl);
3094 *gtt_total = (gtt_size / sizeof(gen8_pte_t)) << PAGE_SHIFT;
3096 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
3097 chv_setup_private_ppat(dev_priv);
3099 bdw_setup_private_ppat(dev_priv);
3101 ret = ggtt_probe_common(dev, gtt_size);
3103 dev_priv->gtt.base.clear_range = gen8_ggtt_clear_range;
3104 dev_priv->gtt.base.insert_entries = gen8_ggtt_insert_entries;
3105 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
3106 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
3108 if (IS_CHERRYVIEW(dev_priv))
3109 dev_priv->gtt.base.insert_entries = gen8_ggtt_insert_entries__BKL;
3114 static int gen6_gmch_probe(struct drm_device *dev,
3117 phys_addr_t *mappable_base,
3120 struct drm_i915_private *dev_priv = dev->dev_private;
3121 unsigned int gtt_size;
3125 *mappable_base = pci_resource_start(dev->pdev, 2);
3126 *mappable_end = pci_resource_len(dev->pdev, 2);
3128 /* 64/512MB is the current min/max we actually know of, but this is just
3129 * a coarse sanity check.
3131 if ((*mappable_end < (64<<20) || (*mappable_end > (512<<20)))) {
3132 DRM_ERROR("Unknown GMADR size (%llx)\n", dev_priv->gtt.mappable_end);
3137 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(40)))
3138 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(40));
3140 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
3142 *stolen = gen6_get_stolen_size(snb_gmch_ctl);
3144 gtt_size = gen6_get_total_gtt_size(snb_gmch_ctl);
3145 *gtt_total = (gtt_size / sizeof(gen6_pte_t)) << PAGE_SHIFT;
3147 ret = ggtt_probe_common(dev, gtt_size);
3149 dev_priv->gtt.base.clear_range = gen6_ggtt_clear_range;
3150 dev_priv->gtt.base.insert_entries = gen6_ggtt_insert_entries;
3151 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
3152 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
3157 static void gen6_gmch_remove(struct i915_address_space *vm)
3160 struct i915_gtt *gtt = container_of(vm, struct i915_gtt, base);
3163 free_scratch_page(vm->dev, vm->scratch_page);
3166 static int i915_gmch_probe(struct drm_device *dev,
3169 phys_addr_t *mappable_base,
3172 struct drm_i915_private *dev_priv = dev->dev_private;
3176 ret = intel_gmch_probe(dev_priv->bridge_dev, dev_priv->dev->pdev, NULL);
3178 DRM_ERROR("failed to set up gmch\n");
3183 intel_gtt_get(gtt_total, stolen, mappable_base, mappable_end);
3185 dev_priv->gtt.do_idle_maps = needs_idle_maps(dev_priv->dev);
3186 dev_priv->gtt.base.insert_entries = i915_ggtt_insert_entries;
3187 dev_priv->gtt.base.clear_range = i915_ggtt_clear_range;
3188 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
3189 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
3191 if (unlikely(dev_priv->gtt.do_idle_maps))
3192 DRM_INFO("applying Ironlake quirks for intel_iommu\n");
3197 static void i915_gmch_remove(struct i915_address_space *vm)
3199 intel_gmch_remove();
3202 int i915_gem_gtt_init(struct drm_device *dev)
3204 struct drm_i915_private *dev_priv = dev->dev_private;
3205 struct i915_gtt *gtt = &dev_priv->gtt;
3208 if (INTEL_INFO(dev)->gen <= 5) {
3209 gtt->gtt_probe = i915_gmch_probe;
3210 gtt->base.cleanup = i915_gmch_remove;
3211 } else if (INTEL_INFO(dev)->gen < 8) {
3212 gtt->gtt_probe = gen6_gmch_probe;
3213 gtt->base.cleanup = gen6_gmch_remove;
3214 if (IS_HASWELL(dev) && dev_priv->ellc_size)
3215 gtt->base.pte_encode = iris_pte_encode;
3216 else if (IS_HASWELL(dev))
3217 gtt->base.pte_encode = hsw_pte_encode;
3218 else if (IS_VALLEYVIEW(dev))
3219 gtt->base.pte_encode = byt_pte_encode;
3220 else if (INTEL_INFO(dev)->gen >= 7)
3221 gtt->base.pte_encode = ivb_pte_encode;
3223 gtt->base.pte_encode = snb_pte_encode;
3225 dev_priv->gtt.gtt_probe = gen8_gmch_probe;
3226 dev_priv->gtt.base.cleanup = gen6_gmch_remove;
3229 gtt->base.dev = dev;
3230 gtt->base.is_ggtt = true;
3232 ret = gtt->gtt_probe(dev, >t->base.total, >t->stolen_size,
3233 >t->mappable_base, >t->mappable_end);
3238 * Initialise stolen early so that we may reserve preallocated
3239 * objects for the BIOS to KMS transition.
3241 ret = i915_gem_init_stolen(dev);
3243 goto out_gtt_cleanup;
3245 /* GMADR is the PCI mmio aperture into the global GTT. */
3246 DRM_INFO("Memory usable by graphics device = %lluM\n",
3247 gtt->base.total >> 20);
3248 DRM_DEBUG_DRIVER("GMADR size = %lldM\n", gtt->mappable_end >> 20);
3249 DRM_DEBUG_DRIVER("GTT stolen size = %zdM\n", gtt->stolen_size >> 20);
3250 #ifdef CONFIG_INTEL_IOMMU
3251 if (intel_iommu_gfx_mapped)
3252 DRM_INFO("VT-d active for gfx access\n");
3255 * i915.enable_ppgtt is read-only, so do an early pass to validate the
3256 * user's requested state against the hardware/driver capabilities. We
3257 * do this now so that we can print out any log messages once rather
3258 * than every time we check intel_enable_ppgtt().
3260 i915.enable_ppgtt = sanitize_enable_ppgtt(dev, i915.enable_ppgtt);
3261 DRM_DEBUG_DRIVER("ppgtt mode: %i\n", i915.enable_ppgtt);
3266 gtt->base.cleanup(&dev_priv->gtt.base);
3271 void i915_gem_restore_gtt_mappings(struct drm_device *dev)
3273 struct drm_i915_private *dev_priv = dev->dev_private;
3274 struct drm_i915_gem_object *obj;
3275 struct i915_address_space *vm;
3276 struct i915_vma *vma;
3279 i915_check_and_clear_faults(dev);
3281 /* First fill our portion of the GTT with scratch pages */
3282 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
3283 dev_priv->gtt.base.start,
3284 dev_priv->gtt.base.total,
3287 /* Cache flush objects bound into GGTT and rebind them. */
3288 vm = &dev_priv->gtt.base;
3289 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
3291 list_for_each_entry(vma, &obj->vma_list, obj_link) {
3295 WARN_ON(i915_vma_bind(vma, obj->cache_level,
3302 i915_gem_clflush_object(obj, obj->pin_display);
3305 if (INTEL_INFO(dev)->gen >= 8) {
3306 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
3307 chv_setup_private_ppat(dev_priv);
3309 bdw_setup_private_ppat(dev_priv);
3314 if (USES_PPGTT(dev)) {
3315 list_for_each_entry(vm, &dev_priv->vm_list, global_link) {
3316 /* TODO: Perhaps it shouldn't be gen6 specific */
3318 struct i915_hw_ppgtt *ppgtt =
3319 container_of(vm, struct i915_hw_ppgtt,
3322 if (i915_is_ggtt(vm))
3323 ppgtt = dev_priv->mm.aliasing_ppgtt;
3325 gen6_write_page_range(dev_priv, &ppgtt->pd,
3326 0, ppgtt->base.total);
3330 i915_ggtt_flush(dev_priv);
3333 static struct i915_vma *
3334 __i915_gem_vma_create(struct drm_i915_gem_object *obj,
3335 struct i915_address_space *vm,
3336 const struct i915_ggtt_view *ggtt_view)
3338 struct i915_vma *vma;
3340 if (WARN_ON(i915_is_ggtt(vm) != !!ggtt_view))
3341 return ERR_PTR(-EINVAL);
3343 vma = kzalloc(sizeof(*vma), GFP_KERNEL);
3345 return ERR_PTR(-ENOMEM);
3347 INIT_LIST_HEAD(&vma->vm_link);
3348 INIT_LIST_HEAD(&vma->obj_link);
3349 INIT_LIST_HEAD(&vma->exec_list);
3352 vma->is_ggtt = i915_is_ggtt(vm);
3354 if (i915_is_ggtt(vm))
3355 vma->ggtt_view = *ggtt_view;
3357 i915_ppgtt_get(i915_vm_to_ppgtt(vm));
3359 list_add_tail(&vma->obj_link, &obj->vma_list);
3365 i915_gem_obj_lookup_or_create_vma(struct drm_i915_gem_object *obj,
3366 struct i915_address_space *vm)
3368 struct i915_vma *vma;
3370 vma = i915_gem_obj_to_vma(obj, vm);
3372 vma = __i915_gem_vma_create(obj, vm,
3373 i915_is_ggtt(vm) ? &i915_ggtt_view_normal : NULL);
3379 i915_gem_obj_lookup_or_create_ggtt_vma(struct drm_i915_gem_object *obj,
3380 const struct i915_ggtt_view *view)
3382 struct i915_address_space *ggtt = i915_obj_to_ggtt(obj);
3383 struct i915_vma *vma;
3386 return ERR_PTR(-EINVAL);
3388 vma = i915_gem_obj_to_ggtt_view(obj, view);
3394 vma = __i915_gem_vma_create(obj, ggtt, view);
3400 static struct scatterlist *
3401 rotate_pages(const dma_addr_t *in, unsigned int offset,
3402 unsigned int width, unsigned int height,
3403 unsigned int stride,
3404 struct sg_table *st, struct scatterlist *sg)
3406 unsigned int column, row;
3407 unsigned int src_idx;
3414 for (column = 0; column < width; column++) {
3415 src_idx = stride * (height - 1) + column;
3416 for (row = 0; row < height; row++) {
3418 /* We don't need the pages, but need to initialize
3419 * the entries so the sg list can be happily traversed.
3420 * The only thing we need are DMA addresses.
3422 sg_set_page(sg, NULL, PAGE_SIZE, 0);
3423 sg_dma_address(sg) = in[offset + src_idx];
3424 sg_dma_len(sg) = PAGE_SIZE;
3433 static struct sg_table *
3434 intel_rotate_fb_obj_pages(struct intel_rotation_info *rot_info,
3435 struct drm_i915_gem_object *obj)
3437 unsigned int size_pages = rot_info->size >> PAGE_SHIFT;
3438 unsigned int size_pages_uv;
3439 struct sg_page_iter sg_iter;
3441 dma_addr_t *page_addr_list;
3442 struct sg_table *st;
3443 unsigned int uv_start_page;
3444 struct scatterlist *sg;
3447 /* Allocate a temporary list of source pages for random access. */
3448 page_addr_list = drm_malloc_ab(obj->base.size / PAGE_SIZE,
3449 sizeof(dma_addr_t));
3450 if (!page_addr_list)
3451 return ERR_PTR(ret);
3453 /* Account for UV plane with NV12. */
3454 if (rot_info->pixel_format == DRM_FORMAT_NV12)
3455 size_pages_uv = rot_info->size_uv >> PAGE_SHIFT;
3459 /* Allocate target SG list. */
3460 st = kmalloc(sizeof(*st), M_DRM, M_WAITOK);
3464 ret = sg_alloc_table(st, size_pages + size_pages_uv, GFP_KERNEL);
3468 /* Populate source page list from the object. */
3470 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) {
3471 page_addr_list[i] = sg_page_iter_dma_address(&sg_iter);
3475 /* Rotate the pages. */
3476 sg = rotate_pages(page_addr_list, 0,
3477 rot_info->width_pages, rot_info->height_pages,
3478 rot_info->width_pages,
3481 /* Append the UV plane if NV12. */
3482 if (rot_info->pixel_format == DRM_FORMAT_NV12) {
3483 uv_start_page = size_pages;
3485 /* Check for tile-row un-alignment. */
3486 if (offset_in_page(rot_info->uv_offset))
3489 rot_info->uv_start_page = uv_start_page;
3491 rotate_pages(page_addr_list, uv_start_page,
3492 rot_info->width_pages_uv,
3493 rot_info->height_pages_uv,
3494 rot_info->width_pages_uv,
3499 "Created rotated page mapping for object size %zu (pitch=%u, height=%u, pixel_format=0x%x, %ux%u tiles, %u pages (%u plane 0)).\n",
3500 obj->base.size, rot_info->pitch, rot_info->height,
3501 rot_info->pixel_format, rot_info->width_pages,
3502 rot_info->height_pages, size_pages + size_pages_uv,
3505 drm_free_large(page_addr_list);
3512 drm_free_large(page_addr_list);
3515 "Failed to create rotated mapping for object size %zu! (%d) (pitch=%u, height=%u, pixel_format=0x%x, %ux%u tiles, %u pages (%u plane 0))\n",
3516 obj->base.size, ret, rot_info->pitch, rot_info->height,
3517 rot_info->pixel_format, rot_info->width_pages,
3518 rot_info->height_pages, size_pages + size_pages_uv,
3520 return ERR_PTR(ret);
3523 static struct sg_table *
3524 intel_partial_pages(const struct i915_ggtt_view *view,
3525 struct drm_i915_gem_object *obj)
3527 struct sg_table *st;
3528 struct scatterlist *sg;
3529 struct sg_page_iter obj_sg_iter;
3532 st = kmalloc(sizeof(*st), M_DRM, M_WAITOK);
3536 ret = sg_alloc_table(st, view->params.partial.size, GFP_KERNEL);
3542 for_each_sg_page(obj->pages->sgl, &obj_sg_iter, obj->pages->nents,
3543 view->params.partial.offset)
3545 if (st->nents >= view->params.partial.size)
3548 sg_set_page(sg, NULL, PAGE_SIZE, 0);
3549 sg_dma_address(sg) = sg_page_iter_dma_address(&obj_sg_iter);
3550 sg_dma_len(sg) = PAGE_SIZE;
3561 return ERR_PTR(ret);
3565 i915_get_ggtt_vma_pages(struct i915_vma *vma)
3569 if (vma->ggtt_view.pages)
3572 if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL)
3573 vma->ggtt_view.pages = vma->obj->pages;
3574 else if (vma->ggtt_view.type == I915_GGTT_VIEW_ROTATED)
3575 vma->ggtt_view.pages =
3576 intel_rotate_fb_obj_pages(&vma->ggtt_view.params.rotated, vma->obj);
3577 else if (vma->ggtt_view.type == I915_GGTT_VIEW_PARTIAL)
3578 vma->ggtt_view.pages =
3579 intel_partial_pages(&vma->ggtt_view, vma->obj);
3581 WARN_ONCE(1, "GGTT view %u not implemented!\n",
3582 vma->ggtt_view.type);
3584 if (!vma->ggtt_view.pages) {
3585 DRM_ERROR("Failed to get pages for GGTT view type %u!\n",
3586 vma->ggtt_view.type);
3588 } else if (IS_ERR(vma->ggtt_view.pages)) {
3589 ret = PTR_ERR(vma->ggtt_view.pages);
3590 vma->ggtt_view.pages = NULL;
3591 DRM_ERROR("Failed to get pages for VMA view type %u (%d)!\n",
3592 vma->ggtt_view.type, ret);
3599 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
3601 * @cache_level: mapping cache level
3602 * @flags: flags like global or local mapping
3604 * DMA addresses are taken from the scatter-gather table of this object (or of
3605 * this VMA in case of non-default GGTT views) and PTE entries set up.
3606 * Note that DMA addresses are also the only part of the SG table we care about.
3608 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
3614 if (WARN_ON(flags == 0))
3618 if (flags & PIN_GLOBAL)
3619 bind_flags |= GLOBAL_BIND;
3620 if (flags & PIN_USER)
3621 bind_flags |= LOCAL_BIND;
3623 if (flags & PIN_UPDATE)
3624 bind_flags |= vma->bound;
3626 bind_flags &= ~vma->bound;
3628 if (bind_flags == 0)
3631 if (vma->bound == 0 && vma->vm->allocate_va_range) {
3632 /* XXX: i915_vma_pin() will fix this +- hack */
3634 trace_i915_va_alloc(vma);
3635 ret = vma->vm->allocate_va_range(vma->vm,
3643 ret = vma->vm->bind_vma(vma, cache_level, bind_flags);
3647 vma->bound |= bind_flags;
3653 * i915_ggtt_view_size - Get the size of a GGTT view.
3654 * @obj: Object the view is of.
3655 * @view: The view in question.
3657 * @return The size of the GGTT view in bytes.
3660 i915_ggtt_view_size(struct drm_i915_gem_object *obj,
3661 const struct i915_ggtt_view *view)
3663 if (view->type == I915_GGTT_VIEW_NORMAL) {
3664 return obj->base.size;
3665 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
3666 return view->params.rotated.size;
3667 } else if (view->type == I915_GGTT_VIEW_PARTIAL) {
3668 return view->params.partial.size << PAGE_SHIFT;
3670 WARN_ONCE(1, "GGTT view %u not implemented!\n", view->type);
3671 return obj->base.size;