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>
35 #include <linux/highmem.h>
38 * DOC: Global GTT views
40 * Background and previous state
42 * Historically objects could exists (be bound) in global GTT space only as
43 * singular instances with a view representing all of the object's backing pages
44 * in a linear fashion. This view will be called a normal view.
46 * To support multiple views of the same object, where the number of mapped
47 * pages is not equal to the backing store, or where the layout of the pages
48 * is not linear, concept of a GGTT view was added.
50 * One example of an alternative view is a stereo display driven by a single
51 * image. In this case we would have a framebuffer looking like this
57 * Above would represent a normal GGTT view as normally mapped for GPU or CPU
58 * rendering. In contrast, fed to the display engine would be an alternative
59 * view which could look something like this:
64 * In this example both the size and layout of pages in the alternative view is
65 * different from the normal view.
67 * Implementation and usage
69 * GGTT views are implemented using VMAs and are distinguished via enum
70 * i915_ggtt_view_type and struct i915_ggtt_view.
72 * A new flavour of core GEM functions which work with GGTT bound objects were
73 * added with the _ggtt_ infix, and sometimes with _view postfix to avoid
74 * renaming in large amounts of code. They take the struct i915_ggtt_view
75 * parameter encapsulating all metadata required to implement a view.
77 * As a helper for callers which are only interested in the normal view,
78 * globally const i915_ggtt_view_normal singleton instance exists. All old core
79 * GEM API functions, the ones not taking the view parameter, are operating on,
80 * or with the normal GGTT view.
82 * Code wanting to add or use a new GGTT view needs to:
84 * 1. Add a new enum with a suitable name.
85 * 2. Extend the metadata in the i915_ggtt_view structure if required.
86 * 3. Add support to i915_get_vma_pages().
88 * New views are required to build a scatter-gather table from within the
89 * i915_get_vma_pages function. This table is stored in the vma.ggtt_view and
90 * exists for the lifetime of an VMA.
92 * Core API is designed to have copy semantics which means that passed in
93 * struct i915_ggtt_view does not need to be persistent (left around after
94 * calling the core API functions).
99 i915_get_ggtt_vma_pages(struct i915_vma *vma);
101 const struct i915_ggtt_view i915_ggtt_view_normal;
102 const struct i915_ggtt_view i915_ggtt_view_rotated = {
103 .type = I915_GGTT_VIEW_ROTATED
106 static int sanitize_enable_ppgtt(struct drm_device *dev, int enable_ppgtt)
108 bool has_aliasing_ppgtt;
111 has_aliasing_ppgtt = INTEL_INFO(dev)->gen >= 6;
112 has_full_ppgtt = INTEL_INFO(dev)->gen >= 7;
114 if (intel_vgpu_active(dev))
115 has_full_ppgtt = false; /* emulation is too hard */
118 * We don't allow disabling PPGTT for gen9+ as it's a requirement for
119 * execlists, the sole mechanism available to submit work.
121 if (INTEL_INFO(dev)->gen < 9 &&
122 (enable_ppgtt == 0 || !has_aliasing_ppgtt))
125 if (enable_ppgtt == 1)
128 if (enable_ppgtt == 2 && has_full_ppgtt)
131 #ifdef CONFIG_INTEL_IOMMU
132 /* Disable ppgtt on SNB if VT-d is on. */
133 if (INTEL_INFO(dev)->gen == 6 && intel_iommu_gfx_mapped) {
134 DRM_INFO("Disabling PPGTT because VT-d is on\n");
139 /* Early VLV doesn't have this */
140 if (IS_VALLEYVIEW(dev) && !IS_CHERRYVIEW(dev) &&
141 dev->pdev->revision < 0xb) {
142 DRM_DEBUG_DRIVER("disabling PPGTT on pre-B3 step VLV\n");
146 if (INTEL_INFO(dev)->gen >= 8 && i915.enable_execlists)
149 return has_aliasing_ppgtt ? 1 : 0;
152 static int ppgtt_bind_vma(struct i915_vma *vma,
153 enum i915_cache_level cache_level,
158 /* Currently applicable only to VLV */
160 pte_flags |= PTE_READ_ONLY;
162 vma->vm->insert_entries(vma->vm, vma->obj->pages, vma->node.start,
163 cache_level, pte_flags);
168 static void ppgtt_unbind_vma(struct i915_vma *vma)
170 vma->vm->clear_range(vma->vm,
176 static gen8_pte_t gen8_pte_encode(dma_addr_t addr,
177 enum i915_cache_level level,
180 gen8_pte_t pte = valid ? _PAGE_PRESENT | _PAGE_RW : 0;
184 case I915_CACHE_NONE:
185 pte |= PPAT_UNCACHED_INDEX;
188 pte |= PPAT_DISPLAY_ELLC_INDEX;
191 pte |= PPAT_CACHED_INDEX;
198 static gen8_pde_t gen8_pde_encode(const dma_addr_t addr,
199 const enum i915_cache_level level)
201 gen8_pde_t pde = _PAGE_PRESENT | _PAGE_RW;
203 if (level != I915_CACHE_NONE)
204 pde |= PPAT_CACHED_PDE_INDEX;
206 pde |= PPAT_UNCACHED_INDEX;
210 #define gen8_pdpe_encode gen8_pde_encode
211 #define gen8_pml4e_encode gen8_pde_encode
213 static gen6_pte_t snb_pte_encode(dma_addr_t addr,
214 enum i915_cache_level level,
215 bool valid, u32 unused)
217 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
218 pte |= GEN6_PTE_ADDR_ENCODE(addr);
221 case I915_CACHE_L3_LLC:
223 pte |= GEN6_PTE_CACHE_LLC;
225 case I915_CACHE_NONE:
226 pte |= GEN6_PTE_UNCACHED;
235 static gen6_pte_t ivb_pte_encode(dma_addr_t addr,
236 enum i915_cache_level level,
237 bool valid, u32 unused)
239 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
240 pte |= GEN6_PTE_ADDR_ENCODE(addr);
243 case I915_CACHE_L3_LLC:
244 pte |= GEN7_PTE_CACHE_L3_LLC;
247 pte |= GEN6_PTE_CACHE_LLC;
249 case I915_CACHE_NONE:
250 pte |= GEN6_PTE_UNCACHED;
259 static gen6_pte_t byt_pte_encode(dma_addr_t addr,
260 enum i915_cache_level level,
261 bool valid, u32 flags)
263 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
264 pte |= GEN6_PTE_ADDR_ENCODE(addr);
266 if (!(flags & PTE_READ_ONLY))
267 pte |= BYT_PTE_WRITEABLE;
269 if (level != I915_CACHE_NONE)
270 pte |= BYT_PTE_SNOOPED_BY_CPU_CACHES;
275 static gen6_pte_t hsw_pte_encode(dma_addr_t addr,
276 enum i915_cache_level level,
277 bool valid, u32 unused)
279 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
280 pte |= HSW_PTE_ADDR_ENCODE(addr);
282 if (level != I915_CACHE_NONE)
283 pte |= HSW_WB_LLC_AGE3;
288 static gen6_pte_t iris_pte_encode(dma_addr_t addr,
289 enum i915_cache_level level,
290 bool valid, u32 unused)
292 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
293 pte |= HSW_PTE_ADDR_ENCODE(addr);
296 case I915_CACHE_NONE:
299 pte |= HSW_WT_ELLC_LLC_AGE3;
302 pte |= HSW_WB_ELLC_LLC_AGE3;
309 static int __setup_page_dma(struct drm_device *dev,
310 struct i915_page_dma *p, gfp_t flags)
312 struct device *device = dev->pdev->dev;
314 p->page = alloc_page(flags);
318 p->daddr = dma_map_page(device,
319 p->page, 0, 4096, PCI_DMA_BIDIRECTIONAL);
321 if (dma_mapping_error(device, p->daddr)) {
322 __free_page(p->page);
329 static int setup_page_dma(struct drm_device *dev, struct i915_page_dma *p)
331 return __setup_page_dma(dev, p, GFP_KERNEL);
334 static void cleanup_page_dma(struct drm_device *dev, struct i915_page_dma *p)
336 if (WARN_ON(!p->page))
339 dma_unmap_page(dev->pdev->dev, p->daddr, 4096, PCI_DMA_BIDIRECTIONAL);
340 __free_page(p->page);
341 memset(p, 0, sizeof(*p));
344 static void *kmap_page_dma(struct i915_page_dma *p)
346 return kmap_atomic(p->page);
349 /* We use the flushing unmap only with ppgtt structures:
350 * page directories, page tables and scratch pages.
352 static void kunmap_page_dma(struct drm_device *dev, void *vaddr)
354 /* There are only few exceptions for gen >=6. chv and bxt.
355 * And we are not sure about the latter so play safe for now.
357 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
358 drm_clflush_virt_range(vaddr, PAGE_SIZE);
360 kunmap_atomic(vaddr);
363 #define kmap_px(px) kmap_page_dma(px_base(px))
364 #define kunmap_px(ppgtt, vaddr) kunmap_page_dma((ppgtt)->base.dev, (vaddr))
366 #define setup_px(dev, px) setup_page_dma((dev), px_base(px))
367 #define cleanup_px(dev, px) cleanup_page_dma((dev), px_base(px))
368 #define fill_px(dev, px, v) fill_page_dma((dev), px_base(px), (v))
369 #define fill32_px(dev, px, v) fill_page_dma_32((dev), px_base(px), (v))
371 static void fill_page_dma(struct drm_device *dev, struct i915_page_dma *p,
375 uint64_t * const vaddr = kmap_page_dma(p);
377 for (i = 0; i < 512; i++)
380 kunmap_page_dma(dev, vaddr);
383 static void fill_page_dma_32(struct drm_device *dev, struct i915_page_dma *p,
384 const uint32_t val32)
390 fill_page_dma(dev, p, v);
393 static struct i915_page_scratch *alloc_scratch_page(struct drm_device *dev)
395 struct i915_page_scratch *sp;
398 sp = kzalloc(sizeof(*sp), GFP_KERNEL);
400 return ERR_PTR(-ENOMEM);
402 ret = __setup_page_dma(dev, px_base(sp), GFP_DMA32 | __GFP_ZERO);
408 set_pages_uc(px_page(sp), 1);
413 static void free_scratch_page(struct drm_device *dev,
414 struct i915_page_scratch *sp)
416 set_pages_wb(px_page(sp), 1);
422 static struct i915_page_table *alloc_pt(struct drm_device *dev)
424 struct i915_page_table *pt;
425 const size_t count = INTEL_INFO(dev)->gen >= 8 ?
426 GEN8_PTES : GEN6_PTES;
429 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
431 return ERR_PTR(-ENOMEM);
433 pt->used_ptes = kcalloc(BITS_TO_LONGS(count), sizeof(*pt->used_ptes),
439 ret = setup_px(dev, pt);
446 kfree(pt->used_ptes);
453 static void free_pt(struct drm_device *dev, struct i915_page_table *pt)
456 kfree(pt->used_ptes);
461 * alloc_pt_range() - Allocate a multiple page tables
462 * @pd: The page directory which will have at least @count entries
463 * available to point to the allocated page tables.
464 * @pde: First page directory entry for which we are allocating.
465 * @count: Number of pages to allocate.
468 * Allocates multiple page table pages and sets the appropriate entries in the
469 * page table structure within the page directory. Function cleans up after
470 * itself on any failures.
472 * Return: 0 if allocation succeeded.
474 static int alloc_pt_range(struct i915_page_directory *pd, uint16_t pde, size_t count,
475 struct drm_device *dev)
479 /* 512 is the max page tables per page_directory on any platform. */
480 if (WARN_ON(pde + count > I915_PDES))
483 for (i = pde; i < pde + count; i++) {
484 struct i915_page_table *pt = alloc_pt(dev);
490 WARN(pd->page_table[i],
491 "Leaking page directory entry %d (%p)\n",
492 i, pd->page_table[i]);
493 pd->page_table[i] = pt;
500 free_pt(dev, pd->page_table[i]);
504 static void gen8_initialize_pt(struct i915_address_space *vm,
505 struct i915_page_table *pt)
507 gen8_pte_t scratch_pte;
509 scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
510 I915_CACHE_LLC, true);
512 fill_px(vm->dev, pt, scratch_pte);
515 static void gen6_initialize_pt(struct i915_address_space *vm,
516 struct i915_page_table *pt)
518 gen6_pte_t scratch_pte;
520 WARN_ON(px_dma(vm->scratch_page) == 0);
522 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
523 I915_CACHE_LLC, true, 0);
525 fill32_px(vm->dev, pt, scratch_pte);
528 static struct i915_page_directory *alloc_pd(struct drm_device *dev)
530 struct i915_page_directory *pd;
533 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
535 return ERR_PTR(-ENOMEM);
537 pd->used_pdes = kcalloc(BITS_TO_LONGS(I915_PDES),
538 sizeof(*pd->used_pdes), GFP_KERNEL);
542 ret = setup_px(dev, pd);
549 kfree(pd->used_pdes);
556 static void free_pd(struct drm_device *dev, struct i915_page_directory *pd)
560 kfree(pd->used_pdes);
565 static void gen8_initialize_pd(struct i915_address_space *vm,
566 struct i915_page_directory *pd)
568 gen8_pde_t scratch_pde;
570 scratch_pde = gen8_pde_encode(px_dma(vm->scratch_pt), I915_CACHE_LLC);
572 fill_px(vm->dev, pd, scratch_pde);
575 static int __pdp_init(struct drm_device *dev,
576 struct i915_page_directory_pointer *pdp)
578 size_t pdpes = I915_PDPES_PER_PDP(dev);
580 pdp->used_pdpes = kcalloc(BITS_TO_LONGS(pdpes),
581 sizeof(unsigned long),
583 if (!pdp->used_pdpes)
586 pdp->page_directory = kcalloc(pdpes, sizeof(*pdp->page_directory),
588 if (!pdp->page_directory) {
589 kfree(pdp->used_pdpes);
590 /* the PDP might be the statically allocated top level. Keep it
591 * as clean as possible */
592 pdp->used_pdpes = NULL;
599 static void __pdp_fini(struct i915_page_directory_pointer *pdp)
601 kfree(pdp->used_pdpes);
602 kfree(pdp->page_directory);
603 pdp->page_directory = NULL;
607 i915_page_directory_pointer *alloc_pdp(struct drm_device *dev)
609 struct i915_page_directory_pointer *pdp;
612 WARN_ON(!USES_FULL_48BIT_PPGTT(dev));
614 pdp = kzalloc(sizeof(*pdp), GFP_KERNEL);
616 return ERR_PTR(-ENOMEM);
618 ret = __pdp_init(dev, pdp);
622 ret = setup_px(dev, pdp);
636 static void free_pdp(struct drm_device *dev,
637 struct i915_page_directory_pointer *pdp)
640 if (USES_FULL_48BIT_PPGTT(dev)) {
641 cleanup_px(dev, pdp);
646 static void gen8_initialize_pdp(struct i915_address_space *vm,
647 struct i915_page_directory_pointer *pdp)
649 gen8_ppgtt_pdpe_t scratch_pdpe;
651 scratch_pdpe = gen8_pdpe_encode(px_dma(vm->scratch_pd), I915_CACHE_LLC);
653 fill_px(vm->dev, pdp, scratch_pdpe);
656 static void gen8_initialize_pml4(struct i915_address_space *vm,
657 struct i915_pml4 *pml4)
659 gen8_ppgtt_pml4e_t scratch_pml4e;
661 scratch_pml4e = gen8_pml4e_encode(px_dma(vm->scratch_pdp),
664 fill_px(vm->dev, pml4, scratch_pml4e);
668 gen8_setup_page_directory(struct i915_hw_ppgtt *ppgtt,
669 struct i915_page_directory_pointer *pdp,
670 struct i915_page_directory *pd,
673 gen8_ppgtt_pdpe_t *page_directorypo;
675 if (!USES_FULL_48BIT_PPGTT(ppgtt->base.dev))
678 page_directorypo = kmap_px(pdp);
679 page_directorypo[index] = gen8_pdpe_encode(px_dma(pd), I915_CACHE_LLC);
680 kunmap_px(ppgtt, page_directorypo);
684 gen8_setup_page_directory_pointer(struct i915_hw_ppgtt *ppgtt,
685 struct i915_pml4 *pml4,
686 struct i915_page_directory_pointer *pdp,
689 gen8_ppgtt_pml4e_t *pagemap = kmap_px(pml4);
691 WARN_ON(!USES_FULL_48BIT_PPGTT(ppgtt->base.dev));
692 pagemap[index] = gen8_pml4e_encode(px_dma(pdp), I915_CACHE_LLC);
693 kunmap_px(ppgtt, pagemap);
696 /* Broadwell Page Directory Pointer Descriptors */
697 static int gen8_write_pdp(struct drm_i915_gem_request *req,
701 struct intel_engine_cs *ring = req->ring;
706 ret = intel_ring_begin(req, 6);
710 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
711 intel_ring_emit(ring, GEN8_RING_PDP_UDW(ring, entry));
712 intel_ring_emit(ring, upper_32_bits(addr));
713 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
714 intel_ring_emit(ring, GEN8_RING_PDP_LDW(ring, entry));
715 intel_ring_emit(ring, lower_32_bits(addr));
716 intel_ring_advance(ring);
721 static int gen8_legacy_mm_switch(struct i915_hw_ppgtt *ppgtt,
722 struct drm_i915_gem_request *req)
726 for (i = GEN8_LEGACY_PDPES - 1; i >= 0; i--) {
727 const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i);
729 ret = gen8_write_pdp(req, i, pd_daddr);
737 static int gen8_48b_mm_switch(struct i915_hw_ppgtt *ppgtt,
738 struct drm_i915_gem_request *req)
740 return gen8_write_pdp(req, 0, px_dma(&ppgtt->pml4));
743 static void gen8_ppgtt_clear_pte_range(struct i915_address_space *vm,
744 struct i915_page_directory_pointer *pdp,
747 gen8_pte_t scratch_pte)
749 struct i915_hw_ppgtt *ppgtt =
750 container_of(vm, struct i915_hw_ppgtt, base);
751 gen8_pte_t *pt_vaddr;
752 unsigned pdpe = gen8_pdpe_index(start);
753 unsigned pde = gen8_pde_index(start);
754 unsigned pte = gen8_pte_index(start);
755 unsigned num_entries = length >> PAGE_SHIFT;
756 unsigned last_pte, i;
761 while (num_entries) {
762 struct i915_page_directory *pd;
763 struct i915_page_table *pt;
765 if (WARN_ON(!pdp->page_directory[pdpe]))
768 pd = pdp->page_directory[pdpe];
770 if (WARN_ON(!pd->page_table[pde]))
773 pt = pd->page_table[pde];
775 if (WARN_ON(!px_page(pt)))
778 last_pte = pte + num_entries;
779 if (last_pte > GEN8_PTES)
780 last_pte = GEN8_PTES;
782 pt_vaddr = kmap_px(pt);
784 for (i = pte; i < last_pte; i++) {
785 pt_vaddr[i] = scratch_pte;
789 kunmap_px(ppgtt, pt);
792 if (++pde == I915_PDES) {
793 if (++pdpe == I915_PDPES_PER_PDP(vm->dev))
800 static void gen8_ppgtt_clear_range(struct i915_address_space *vm,
805 struct i915_hw_ppgtt *ppgtt =
806 container_of(vm, struct i915_hw_ppgtt, base);
807 gen8_pte_t scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
808 I915_CACHE_LLC, use_scratch);
810 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
811 gen8_ppgtt_clear_pte_range(vm, &ppgtt->pdp, start, length,
814 uint64_t templ4, pml4e;
815 struct i915_page_directory_pointer *pdp;
817 gen8_for_each_pml4e(pdp, &ppgtt->pml4, start, length, templ4, pml4e) {
818 gen8_ppgtt_clear_pte_range(vm, pdp, start, length,
825 gen8_ppgtt_insert_pte_entries(struct i915_address_space *vm,
826 struct i915_page_directory_pointer *pdp,
827 struct sg_page_iter *sg_iter,
829 enum i915_cache_level cache_level)
831 struct i915_hw_ppgtt *ppgtt =
832 container_of(vm, struct i915_hw_ppgtt, base);
833 gen8_pte_t *pt_vaddr;
834 unsigned pdpe = gen8_pdpe_index(start);
835 unsigned pde = gen8_pde_index(start);
836 unsigned pte = gen8_pte_index(start);
840 while (__sg_page_iter_next(sg_iter)) {
841 if (pt_vaddr == NULL) {
842 struct i915_page_directory *pd = pdp->page_directory[pdpe];
843 struct i915_page_table *pt = pd->page_table[pde];
845 pt_vaddr = kmap_px(pt);
849 gen8_pte_encode(sg_page_iter_dma_address(sg_iter),
851 if (++pte == GEN8_PTES) {
852 kunmap_px(ppgtt, pt_vaddr);
854 if (++pde == I915_PDES) {
855 if (++pdpe == I915_PDPES_PER_PDP(vm->dev))
864 kunmap_px(ppgtt, pt_vaddr);
867 static void gen8_ppgtt_insert_entries(struct i915_address_space *vm,
868 struct sg_table *pages,
870 enum i915_cache_level cache_level,
873 struct i915_hw_ppgtt *ppgtt =
874 container_of(vm, struct i915_hw_ppgtt, base);
875 struct sg_page_iter sg_iter;
877 __sg_page_iter_start(&sg_iter, pages->sgl, sg_nents(pages->sgl), 0);
879 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
880 gen8_ppgtt_insert_pte_entries(vm, &ppgtt->pdp, &sg_iter, start,
883 struct i915_page_directory_pointer *pdp;
884 uint64_t templ4, pml4e;
885 uint64_t length = (uint64_t)pages->orig_nents << PAGE_SHIFT;
887 gen8_for_each_pml4e(pdp, &ppgtt->pml4, start, length, templ4, pml4e) {
888 gen8_ppgtt_insert_pte_entries(vm, pdp, &sg_iter,
894 static void gen8_free_page_tables(struct drm_device *dev,
895 struct i915_page_directory *pd)
902 for_each_set_bit(i, pd->used_pdes, I915_PDES) {
903 if (WARN_ON(!pd->page_table[i]))
906 free_pt(dev, pd->page_table[i]);
907 pd->page_table[i] = NULL;
911 static int gen8_init_scratch(struct i915_address_space *vm)
913 struct drm_device *dev = vm->dev;
915 vm->scratch_page = alloc_scratch_page(dev);
916 if (IS_ERR(vm->scratch_page))
917 return PTR_ERR(vm->scratch_page);
919 vm->scratch_pt = alloc_pt(dev);
920 if (IS_ERR(vm->scratch_pt)) {
921 free_scratch_page(dev, vm->scratch_page);
922 return PTR_ERR(vm->scratch_pt);
925 vm->scratch_pd = alloc_pd(dev);
926 if (IS_ERR(vm->scratch_pd)) {
927 free_pt(dev, vm->scratch_pt);
928 free_scratch_page(dev, vm->scratch_page);
929 return PTR_ERR(vm->scratch_pd);
932 if (USES_FULL_48BIT_PPGTT(dev)) {
933 vm->scratch_pdp = alloc_pdp(dev);
934 if (IS_ERR(vm->scratch_pdp)) {
935 free_pd(dev, vm->scratch_pd);
936 free_pt(dev, vm->scratch_pt);
937 free_scratch_page(dev, vm->scratch_page);
938 return PTR_ERR(vm->scratch_pdp);
942 gen8_initialize_pt(vm, vm->scratch_pt);
943 gen8_initialize_pd(vm, vm->scratch_pd);
944 if (USES_FULL_48BIT_PPGTT(dev))
945 gen8_initialize_pdp(vm, vm->scratch_pdp);
950 static int gen8_ppgtt_notify_vgt(struct i915_hw_ppgtt *ppgtt, bool create)
952 enum vgt_g2v_type msg;
953 struct drm_device *dev = ppgtt->base.dev;
954 struct drm_i915_private *dev_priv = dev->dev_private;
955 unsigned int offset = vgtif_reg(pdp0_lo);
958 if (USES_FULL_48BIT_PPGTT(dev)) {
959 u64 daddr = px_dma(&ppgtt->pml4);
961 I915_WRITE(offset, lower_32_bits(daddr));
962 I915_WRITE(offset + 4, upper_32_bits(daddr));
964 msg = (create ? VGT_G2V_PPGTT_L4_PAGE_TABLE_CREATE :
965 VGT_G2V_PPGTT_L4_PAGE_TABLE_DESTROY);
967 for (i = 0; i < GEN8_LEGACY_PDPES; i++) {
968 u64 daddr = i915_page_dir_dma_addr(ppgtt, i);
970 I915_WRITE(offset, lower_32_bits(daddr));
971 I915_WRITE(offset + 4, upper_32_bits(daddr));
976 msg = (create ? VGT_G2V_PPGTT_L3_PAGE_TABLE_CREATE :
977 VGT_G2V_PPGTT_L3_PAGE_TABLE_DESTROY);
980 I915_WRITE(vgtif_reg(g2v_notify), msg);
985 static void gen8_free_scratch(struct i915_address_space *vm)
987 struct drm_device *dev = vm->dev;
989 if (USES_FULL_48BIT_PPGTT(dev))
990 free_pdp(dev, vm->scratch_pdp);
991 free_pd(dev, vm->scratch_pd);
992 free_pt(dev, vm->scratch_pt);
993 free_scratch_page(dev, vm->scratch_page);
996 static void gen8_ppgtt_cleanup_3lvl(struct drm_device *dev,
997 struct i915_page_directory_pointer *pdp)
1001 for_each_set_bit(i, pdp->used_pdpes, I915_PDPES_PER_PDP(dev)) {
1002 if (WARN_ON(!pdp->page_directory[i]))
1005 gen8_free_page_tables(dev, pdp->page_directory[i]);
1006 free_pd(dev, pdp->page_directory[i]);
1012 static void gen8_ppgtt_cleanup_4lvl(struct i915_hw_ppgtt *ppgtt)
1016 for_each_set_bit(i, ppgtt->pml4.used_pml4es, GEN8_PML4ES_PER_PML4) {
1017 if (WARN_ON(!ppgtt->pml4.pdps[i]))
1020 gen8_ppgtt_cleanup_3lvl(ppgtt->base.dev, ppgtt->pml4.pdps[i]);
1023 cleanup_px(ppgtt->base.dev, &ppgtt->pml4);
1026 static void gen8_ppgtt_cleanup(struct i915_address_space *vm)
1028 struct i915_hw_ppgtt *ppgtt =
1029 container_of(vm, struct i915_hw_ppgtt, base);
1031 if (intel_vgpu_active(vm->dev))
1032 gen8_ppgtt_notify_vgt(ppgtt, false);
1034 if (!USES_FULL_48BIT_PPGTT(ppgtt->base.dev))
1035 gen8_ppgtt_cleanup_3lvl(ppgtt->base.dev, &ppgtt->pdp);
1037 gen8_ppgtt_cleanup_4lvl(ppgtt);
1039 gen8_free_scratch(vm);
1043 * gen8_ppgtt_alloc_pagetabs() - Allocate page tables for VA range.
1044 * @vm: Master vm structure.
1045 * @pd: Page directory for this address range.
1046 * @start: Starting virtual address to begin allocations.
1047 * @length: Size of the allocations.
1048 * @new_pts: Bitmap set by function with new allocations. Likely used by the
1049 * caller to free on error.
1051 * Allocate the required number of page tables. Extremely similar to
1052 * gen8_ppgtt_alloc_page_directories(). The main difference is here we are limited by
1053 * the page directory boundary (instead of the page directory pointer). That
1054 * boundary is 1GB virtual. Therefore, unlike gen8_ppgtt_alloc_page_directories(), it is
1055 * possible, and likely that the caller will need to use multiple calls of this
1056 * function to achieve the appropriate allocation.
1058 * Return: 0 if success; negative error code otherwise.
1060 static int gen8_ppgtt_alloc_pagetabs(struct i915_address_space *vm,
1061 struct i915_page_directory *pd,
1064 unsigned long *new_pts)
1066 struct drm_device *dev = vm->dev;
1067 struct i915_page_table *pt;
1071 gen8_for_each_pde(pt, pd, start, length, temp, pde) {
1072 /* Don't reallocate page tables */
1073 if (test_bit(pde, pd->used_pdes)) {
1074 /* Scratch is never allocated this way */
1075 WARN_ON(pt == vm->scratch_pt);
1083 gen8_initialize_pt(vm, pt);
1084 pd->page_table[pde] = pt;
1085 __set_bit(pde, new_pts);
1086 trace_i915_page_table_entry_alloc(vm, pde, start, GEN8_PDE_SHIFT);
1092 for_each_set_bit(pde, new_pts, I915_PDES)
1093 free_pt(dev, pd->page_table[pde]);
1099 * gen8_ppgtt_alloc_page_directories() - Allocate page directories for VA range.
1100 * @vm: Master vm structure.
1101 * @pdp: Page directory pointer for this address range.
1102 * @start: Starting virtual address to begin allocations.
1103 * @length: Size of the allocations.
1104 * @new_pds: Bitmap set by function with new allocations. Likely used by the
1105 * caller to free on error.
1107 * Allocate the required number of page directories starting at the pde index of
1108 * @start, and ending at the pde index @start + @length. This function will skip
1109 * over already allocated page directories within the range, and only allocate
1110 * new ones, setting the appropriate pointer within the pdp as well as the
1111 * correct position in the bitmap @new_pds.
1113 * The function will only allocate the pages within the range for a give page
1114 * directory pointer. In other words, if @start + @length straddles a virtually
1115 * addressed PDP boundary (512GB for 4k pages), there will be more allocations
1116 * required by the caller, This is not currently possible, and the BUG in the
1117 * code will prevent it.
1119 * Return: 0 if success; negative error code otherwise.
1122 gen8_ppgtt_alloc_page_directories(struct i915_address_space *vm,
1123 struct i915_page_directory_pointer *pdp,
1126 unsigned long *new_pds)
1128 struct drm_device *dev = vm->dev;
1129 struct i915_page_directory *pd;
1132 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1134 WARN_ON(!bitmap_empty(new_pds, pdpes));
1136 gen8_for_each_pdpe(pd, pdp, start, length, temp, pdpe) {
1137 if (test_bit(pdpe, pdp->used_pdpes))
1144 gen8_initialize_pd(vm, pd);
1145 pdp->page_directory[pdpe] = pd;
1146 __set_bit(pdpe, new_pds);
1147 trace_i915_page_directory_entry_alloc(vm, pdpe, start, GEN8_PDPE_SHIFT);
1153 for_each_set_bit(pdpe, new_pds, pdpes)
1154 free_pd(dev, pdp->page_directory[pdpe]);
1160 * gen8_ppgtt_alloc_page_dirpointers() - Allocate pdps for VA range.
1161 * @vm: Master vm structure.
1162 * @pml4: Page map level 4 for this address range.
1163 * @start: Starting virtual address to begin allocations.
1164 * @length: Size of the allocations.
1165 * @new_pdps: Bitmap set by function with new allocations. Likely used by the
1166 * caller to free on error.
1168 * Allocate the required number of page directory pointers. Extremely similar to
1169 * gen8_ppgtt_alloc_page_directories() and gen8_ppgtt_alloc_pagetabs().
1170 * The main difference is here we are limited by the pml4 boundary (instead of
1171 * the page directory pointer).
1173 * Return: 0 if success; negative error code otherwise.
1176 gen8_ppgtt_alloc_page_dirpointers(struct i915_address_space *vm,
1177 struct i915_pml4 *pml4,
1180 unsigned long *new_pdps)
1182 struct drm_device *dev = vm->dev;
1183 struct i915_page_directory_pointer *pdp;
1187 WARN_ON(!bitmap_empty(new_pdps, GEN8_PML4ES_PER_PML4));
1189 gen8_for_each_pml4e(pdp, pml4, start, length, temp, pml4e) {
1190 if (!test_bit(pml4e, pml4->used_pml4es)) {
1191 pdp = alloc_pdp(dev);
1195 gen8_initialize_pdp(vm, pdp);
1196 pml4->pdps[pml4e] = pdp;
1197 __set_bit(pml4e, new_pdps);
1198 trace_i915_page_directory_pointer_entry_alloc(vm,
1208 for_each_set_bit(pml4e, new_pdps, GEN8_PML4ES_PER_PML4)
1209 free_pdp(dev, pml4->pdps[pml4e]);
1215 free_gen8_temp_bitmaps(unsigned long *new_pds, unsigned long *new_pts)
1221 /* Fills in the page directory bitmap, and the array of page tables bitmap. Both
1222 * of these are based on the number of PDPEs in the system.
1225 int __must_check alloc_gen8_temp_bitmaps(unsigned long **new_pds,
1226 unsigned long **new_pts,
1232 pds = kcalloc(BITS_TO_LONGS(pdpes), sizeof(unsigned long), GFP_TEMPORARY);
1236 pts = kcalloc(pdpes, BITS_TO_LONGS(I915_PDES) * sizeof(unsigned long),
1247 free_gen8_temp_bitmaps(pds, pts);
1251 /* PDE TLBs are a pain to invalidate on GEN8+. When we modify
1252 * the page table structures, we mark them dirty so that
1253 * context switching/execlist queuing code takes extra steps
1254 * to ensure that tlbs are flushed.
1256 static void mark_tlbs_dirty(struct i915_hw_ppgtt *ppgtt)
1258 ppgtt->pd_dirty_rings = INTEL_INFO(ppgtt->base.dev)->ring_mask;
1261 static int gen8_alloc_va_range_3lvl(struct i915_address_space *vm,
1262 struct i915_page_directory_pointer *pdp,
1266 struct i915_hw_ppgtt *ppgtt =
1267 container_of(vm, struct i915_hw_ppgtt, base);
1268 unsigned long *new_page_dirs, *new_page_tables;
1269 struct drm_device *dev = vm->dev;
1270 struct i915_page_directory *pd;
1271 const uint64_t orig_start = start;
1272 const uint64_t orig_length = length;
1275 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1278 /* Wrap is never okay since we can only represent 48b, and we don't
1279 * actually use the other side of the canonical address space.
1281 if (WARN_ON(start + length < start))
1284 if (WARN_ON(start + length > vm->total))
1287 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables, pdpes);
1291 /* Do the allocations first so we can easily bail out */
1292 ret = gen8_ppgtt_alloc_page_directories(vm, pdp, start, length,
1295 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1299 /* For every page directory referenced, allocate page tables */
1300 gen8_for_each_pdpe(pd, pdp, start, length, temp, pdpe) {
1301 ret = gen8_ppgtt_alloc_pagetabs(vm, pd, start, length,
1302 new_page_tables + pdpe * BITS_TO_LONGS(I915_PDES));
1308 length = orig_length;
1310 /* Allocations have completed successfully, so set the bitmaps, and do
1312 gen8_for_each_pdpe(pd, pdp, start, length, temp, pdpe) {
1313 gen8_pde_t *const page_directory = kmap_px(pd);
1314 struct i915_page_table *pt;
1315 uint64_t pd_len = length;
1316 uint64_t pd_start = start;
1319 /* Every pd should be allocated, we just did that above. */
1322 gen8_for_each_pde(pt, pd, pd_start, pd_len, temp, pde) {
1323 /* Same reasoning as pd */
1326 WARN_ON(!gen8_pte_count(pd_start, pd_len));
1328 /* Set our used ptes within the page table */
1329 bitmap_set(pt->used_ptes,
1330 gen8_pte_index(pd_start),
1331 gen8_pte_count(pd_start, pd_len));
1333 /* Our pde is now pointing to the pagetable, pt */
1334 __set_bit(pde, pd->used_pdes);
1336 /* Map the PDE to the page table */
1337 page_directory[pde] = gen8_pde_encode(px_dma(pt),
1339 trace_i915_page_table_entry_map(&ppgtt->base, pde, pt,
1340 gen8_pte_index(start),
1341 gen8_pte_count(start, length),
1344 /* NB: We haven't yet mapped ptes to pages. At this
1345 * point we're still relying on insert_entries() */
1348 kunmap_px(ppgtt, page_directory);
1349 __set_bit(pdpe, pdp->used_pdpes);
1350 gen8_setup_page_directory(ppgtt, pdp, pd, pdpe);
1353 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1354 mark_tlbs_dirty(ppgtt);
1359 for_each_set_bit(temp, new_page_tables + pdpe *
1360 BITS_TO_LONGS(I915_PDES), I915_PDES)
1361 free_pt(dev, pdp->page_directory[pdpe]->page_table[temp]);
1364 for_each_set_bit(pdpe, new_page_dirs, pdpes)
1365 free_pd(dev, pdp->page_directory[pdpe]);
1367 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1368 mark_tlbs_dirty(ppgtt);
1372 static int gen8_alloc_va_range_4lvl(struct i915_address_space *vm,
1373 struct i915_pml4 *pml4,
1377 DECLARE_BITMAP(new_pdps, GEN8_PML4ES_PER_PML4);
1378 struct i915_hw_ppgtt *ppgtt =
1379 container_of(vm, struct i915_hw_ppgtt, base);
1380 struct i915_page_directory_pointer *pdp;
1381 uint64_t temp, pml4e;
1384 /* Do the pml4 allocations first, so we don't need to track the newly
1385 * allocated tables below the pdp */
1386 bitmap_zero(new_pdps, GEN8_PML4ES_PER_PML4);
1388 /* The pagedirectory and pagetable allocations are done in the shared 3
1389 * and 4 level code. Just allocate the pdps.
1391 ret = gen8_ppgtt_alloc_page_dirpointers(vm, pml4, start, length,
1396 WARN(bitmap_weight(new_pdps, GEN8_PML4ES_PER_PML4) > 2,
1397 "The allocation has spanned more than 512GB. "
1398 "It is highly likely this is incorrect.");
1400 gen8_for_each_pml4e(pdp, pml4, start, length, temp, pml4e) {
1403 ret = gen8_alloc_va_range_3lvl(vm, pdp, start, length);
1407 gen8_setup_page_directory_pointer(ppgtt, pml4, pdp, pml4e);
1410 bitmap_or(pml4->used_pml4es, new_pdps, pml4->used_pml4es,
1411 GEN8_PML4ES_PER_PML4);
1416 for_each_set_bit(pml4e, new_pdps, GEN8_PML4ES_PER_PML4)
1417 gen8_ppgtt_cleanup_3lvl(vm->dev, pml4->pdps[pml4e]);
1422 static int gen8_alloc_va_range(struct i915_address_space *vm,
1423 uint64_t start, uint64_t length)
1425 struct i915_hw_ppgtt *ppgtt =
1426 container_of(vm, struct i915_hw_ppgtt, base);
1428 if (USES_FULL_48BIT_PPGTT(vm->dev))
1429 return gen8_alloc_va_range_4lvl(vm, &ppgtt->pml4, start, length);
1431 return gen8_alloc_va_range_3lvl(vm, &ppgtt->pdp, start, length);
1434 static void gen8_dump_pdp(struct i915_page_directory_pointer *pdp,
1435 uint64_t start, uint64_t length,
1436 gen8_pte_t scratch_pte,
1439 struct i915_page_directory *pd;
1443 gen8_for_each_pdpe(pd, pdp, start, length, temp, pdpe) {
1444 struct i915_page_table *pt;
1445 uint64_t pd_len = length;
1446 uint64_t pd_start = start;
1449 if (!test_bit(pdpe, pdp->used_pdpes))
1452 seq_printf(m, "\tPDPE #%d\n", pdpe);
1453 gen8_for_each_pde(pt, pd, pd_start, pd_len, temp, pde) {
1455 gen8_pte_t *pt_vaddr;
1457 if (!test_bit(pde, pd->used_pdes))
1460 pt_vaddr = kmap_px(pt);
1461 for (pte = 0; pte < GEN8_PTES; pte += 4) {
1463 (pdpe << GEN8_PDPE_SHIFT) |
1464 (pde << GEN8_PDE_SHIFT) |
1465 (pte << GEN8_PTE_SHIFT);
1469 for (i = 0; i < 4; i++)
1470 if (pt_vaddr[pte + i] != scratch_pte)
1475 seq_printf(m, "\t\t0x%lx [%03d,%03d,%04d]: =", va, pdpe, pde, pte);
1476 for (i = 0; i < 4; i++) {
1477 if (pt_vaddr[pte + i] != scratch_pte)
1478 seq_printf(m, " %lx", pt_vaddr[pte + i]);
1480 seq_puts(m, " SCRATCH ");
1484 /* don't use kunmap_px, it could trigger
1485 * an unnecessary flush.
1487 kunmap_atomic(pt_vaddr);
1492 static void gen8_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
1494 struct i915_address_space *vm = &ppgtt->base;
1495 uint64_t start = ppgtt->base.start;
1496 uint64_t length = ppgtt->base.total;
1497 gen8_pte_t scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
1498 I915_CACHE_LLC, true);
1500 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
1501 gen8_dump_pdp(&ppgtt->pdp, start, length, scratch_pte, m);
1503 uint64_t templ4, pml4e;
1504 struct i915_pml4 *pml4 = &ppgtt->pml4;
1505 struct i915_page_directory_pointer *pdp;
1507 gen8_for_each_pml4e(pdp, pml4, start, length, templ4, pml4e) {
1508 if (!test_bit(pml4e, pml4->used_pml4es))
1511 seq_printf(m, " PML4E #%lu\n", pml4e);
1512 gen8_dump_pdp(pdp, start, length, scratch_pte, m);
1517 static int gen8_preallocate_top_level_pdps(struct i915_hw_ppgtt *ppgtt)
1519 unsigned long *new_page_dirs, *new_page_tables;
1520 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1523 /* We allocate temp bitmap for page tables for no gain
1524 * but as this is for init only, lets keep the things simple
1526 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables, pdpes);
1530 /* Allocate for all pdps regardless of how the ppgtt
1533 ret = gen8_ppgtt_alloc_page_directories(&ppgtt->base, &ppgtt->pdp,
1537 *ppgtt->pdp.used_pdpes = *new_page_dirs;
1539 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1545 * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers
1546 * with a net effect resembling a 2-level page table in normal x86 terms. Each
1547 * PDP represents 1GB of memory 4 * 512 * 512 * 4096 = 4GB legacy 32b address
1551 static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
1555 ret = gen8_init_scratch(&ppgtt->base);
1559 ppgtt->base.start = 0;
1560 ppgtt->base.cleanup = gen8_ppgtt_cleanup;
1561 ppgtt->base.allocate_va_range = gen8_alloc_va_range;
1562 ppgtt->base.insert_entries = gen8_ppgtt_insert_entries;
1563 ppgtt->base.clear_range = gen8_ppgtt_clear_range;
1564 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
1565 ppgtt->base.bind_vma = ppgtt_bind_vma;
1566 ppgtt->debug_dump = gen8_dump_ppgtt;
1568 if (USES_FULL_48BIT_PPGTT(ppgtt->base.dev)) {
1569 ret = setup_px(ppgtt->base.dev, &ppgtt->pml4);
1573 gen8_initialize_pml4(&ppgtt->base, &ppgtt->pml4);
1575 ppgtt->base.total = 1ULL << 48;
1576 ppgtt->switch_mm = gen8_48b_mm_switch;
1578 ret = __pdp_init(ppgtt->base.dev, &ppgtt->pdp);
1582 #define CONFIG_X86_32 0
1583 ppgtt->base.total = 1ULL << 32;
1584 ppgtt->switch_mm = gen8_legacy_mm_switch;
1585 trace_i915_page_directory_pointer_entry_alloc(&ppgtt->base,
1589 if (intel_vgpu_active(ppgtt->base.dev)) {
1590 ret = gen8_preallocate_top_level_pdps(ppgtt);
1596 if (intel_vgpu_active(ppgtt->base.dev))
1597 gen8_ppgtt_notify_vgt(ppgtt, true);
1602 gen8_free_scratch(&ppgtt->base);
1606 static int gen8_aliasing_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
1608 struct drm_device *dev = ppgtt->base.dev;
1609 struct drm_i915_private *dev_priv = dev->dev_private;
1610 uint64_t start = 0, size = dev_priv->gtt.base.total;
1613 ret = gen8_ppgtt_init(ppgtt);
1617 /* Aliasing PPGTT has to always work and be mapped because of the way we
1618 * use RESTORE_INHIBIT in the context switch. This will be fixed
1620 ret = gen8_alloc_va_range(&ppgtt->base, start, size);
1622 free_pd(ppgtt->base.dev, ppgtt->base.scratch_pd);
1623 free_pt(ppgtt->base.dev, ppgtt->base.scratch_pt);
1627 ppgtt->base.allocate_va_range = NULL;
1628 ppgtt->base.clear_range(&ppgtt->base, 0, ppgtt->base.total, true);
1633 static void gen6_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
1635 struct i915_address_space *vm = &ppgtt->base;
1636 struct i915_page_table *unused;
1637 gen6_pte_t scratch_pte;
1639 uint32_t pte, pde, temp;
1640 uint32_t start = ppgtt->base.start, length = ppgtt->base.total;
1642 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1643 I915_CACHE_LLC, true, 0);
1645 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde) {
1647 gen6_pte_t *pt_vaddr;
1648 const dma_addr_t pt_addr = px_dma(ppgtt->pd.page_table[pde]);
1649 pd_entry = readl(ppgtt->pd_addr + pde);
1650 expected = (GEN6_PDE_ADDR_ENCODE(pt_addr) | GEN6_PDE_VALID);
1652 if (pd_entry != expected)
1653 seq_printf(m, "\tPDE #%d mismatch: Actual PDE: %x Expected PDE: %x\n",
1657 seq_printf(m, "\tPDE: %x\n", pd_entry);
1659 pt_vaddr = kmap_px(ppgtt->pd.page_table[pde]);
1661 for (pte = 0; pte < GEN6_PTES; pte+=4) {
1663 (pde * PAGE_SIZE * GEN6_PTES) +
1667 for (i = 0; i < 4; i++)
1668 if (pt_vaddr[pte + i] != scratch_pte)
1673 seq_printf(m, "\t\t0x%lx [%03d,%04d]: =", va, pde, pte);
1674 for (i = 0; i < 4; i++) {
1675 if (pt_vaddr[pte + i] != scratch_pte)
1676 seq_printf(m, " %08x", pt_vaddr[pte + i]);
1678 seq_puts(m, " SCRATCH ");
1682 kunmap_px(ppgtt, pt_vaddr);
1686 /* Write pde (index) from the page directory @pd to the page table @pt */
1687 static void gen6_write_pde(struct i915_page_directory *pd,
1688 const int pde, struct i915_page_table *pt)
1690 /* Caller needs to make sure the write completes if necessary */
1691 struct i915_hw_ppgtt *ppgtt =
1692 container_of(pd, struct i915_hw_ppgtt, pd);
1695 pd_entry = GEN6_PDE_ADDR_ENCODE(px_dma(pt));
1696 pd_entry |= GEN6_PDE_VALID;
1698 writel(pd_entry, ppgtt->pd_addr + pde);
1701 /* Write all the page tables found in the ppgtt structure to incrementing page
1703 static void gen6_write_page_range(struct drm_i915_private *dev_priv,
1704 struct i915_page_directory *pd,
1705 uint32_t start, uint32_t length)
1707 struct i915_page_table *pt;
1710 gen6_for_each_pde(pt, pd, start, length, temp, pde)
1711 gen6_write_pde(pd, pde, pt);
1713 /* Make sure write is complete before other code can use this page
1714 * table. Also require for WC mapped PTEs */
1715 readl(dev_priv->gtt.gsm);
1718 static uint32_t get_pd_offset(struct i915_hw_ppgtt *ppgtt)
1720 BUG_ON(ppgtt->pd.base.ggtt_offset & 0x3f);
1722 return (ppgtt->pd.base.ggtt_offset / 64) << 16;
1725 static int hsw_mm_switch(struct i915_hw_ppgtt *ppgtt,
1726 struct drm_i915_gem_request *req)
1728 struct intel_engine_cs *ring = req->ring;
1731 /* NB: TLBs must be flushed and invalidated before a switch */
1732 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1736 ret = intel_ring_begin(req, 6);
1740 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1741 intel_ring_emit(ring, RING_PP_DIR_DCLV(ring));
1742 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1743 intel_ring_emit(ring, RING_PP_DIR_BASE(ring));
1744 intel_ring_emit(ring, get_pd_offset(ppgtt));
1745 intel_ring_emit(ring, MI_NOOP);
1746 intel_ring_advance(ring);
1751 static int vgpu_mm_switch(struct i915_hw_ppgtt *ppgtt,
1752 struct drm_i915_gem_request *req)
1754 struct intel_engine_cs *ring = req->ring;
1755 struct drm_i915_private *dev_priv = to_i915(ppgtt->base.dev);
1757 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
1758 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
1762 static int gen7_mm_switch(struct i915_hw_ppgtt *ppgtt,
1763 struct drm_i915_gem_request *req)
1765 struct intel_engine_cs *ring = req->ring;
1768 /* NB: TLBs must be flushed and invalidated before a switch */
1769 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1773 ret = intel_ring_begin(req, 6);
1777 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1778 intel_ring_emit(ring, RING_PP_DIR_DCLV(ring));
1779 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1780 intel_ring_emit(ring, RING_PP_DIR_BASE(ring));
1781 intel_ring_emit(ring, get_pd_offset(ppgtt));
1782 intel_ring_emit(ring, MI_NOOP);
1783 intel_ring_advance(ring);
1785 /* XXX: RCS is the only one to auto invalidate the TLBs? */
1786 if (ring->id != RCS) {
1787 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1795 static int gen6_mm_switch(struct i915_hw_ppgtt *ppgtt,
1796 struct drm_i915_gem_request *req)
1798 struct intel_engine_cs *ring = req->ring;
1799 struct drm_device *dev = ppgtt->base.dev;
1800 struct drm_i915_private *dev_priv = dev->dev_private;
1803 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
1804 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
1806 POSTING_READ(RING_PP_DIR_DCLV(ring));
1811 static void gen8_ppgtt_enable(struct drm_device *dev)
1813 struct drm_i915_private *dev_priv = dev->dev_private;
1814 struct intel_engine_cs *ring;
1817 for_each_ring(ring, dev_priv, j) {
1818 u32 four_level = USES_FULL_48BIT_PPGTT(dev) ? GEN8_GFX_PPGTT_48B : 0;
1819 I915_WRITE(RING_MODE_GEN7(ring),
1820 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE | four_level));
1824 static void gen7_ppgtt_enable(struct drm_device *dev)
1826 struct drm_i915_private *dev_priv = dev->dev_private;
1827 struct intel_engine_cs *ring;
1828 uint32_t ecochk, ecobits;
1831 ecobits = I915_READ(GAC_ECO_BITS);
1832 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B);
1834 ecochk = I915_READ(GAM_ECOCHK);
1835 if (IS_HASWELL(dev)) {
1836 ecochk |= ECOCHK_PPGTT_WB_HSW;
1838 ecochk |= ECOCHK_PPGTT_LLC_IVB;
1839 ecochk &= ~ECOCHK_PPGTT_GFDT_IVB;
1841 I915_WRITE(GAM_ECOCHK, ecochk);
1843 for_each_ring(ring, dev_priv, i) {
1844 /* GFX_MODE is per-ring on gen7+ */
1845 I915_WRITE(RING_MODE_GEN7(ring),
1846 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1850 static void gen6_ppgtt_enable(struct drm_device *dev)
1852 struct drm_i915_private *dev_priv = dev->dev_private;
1853 uint32_t ecochk, gab_ctl, ecobits;
1855 ecobits = I915_READ(GAC_ECO_BITS);
1856 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_SNB_BIT |
1857 ECOBITS_PPGTT_CACHE64B);
1859 gab_ctl = I915_READ(GAB_CTL);
1860 I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT);
1862 ecochk = I915_READ(GAM_ECOCHK);
1863 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT | ECOCHK_PPGTT_CACHE64B);
1865 I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1868 /* PPGTT support for Sandybdrige/Gen6 and later */
1869 static void gen6_ppgtt_clear_range(struct i915_address_space *vm,
1874 struct i915_hw_ppgtt *ppgtt =
1875 container_of(vm, struct i915_hw_ppgtt, base);
1876 gen6_pte_t *pt_vaddr, scratch_pte;
1877 unsigned first_entry = start >> PAGE_SHIFT;
1878 unsigned num_entries = length >> PAGE_SHIFT;
1879 unsigned act_pt = first_entry / GEN6_PTES;
1880 unsigned first_pte = first_entry % GEN6_PTES;
1881 unsigned last_pte, i;
1883 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1884 I915_CACHE_LLC, true, 0);
1886 while (num_entries) {
1887 last_pte = first_pte + num_entries;
1888 if (last_pte > GEN6_PTES)
1889 last_pte = GEN6_PTES;
1891 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1893 for (i = first_pte; i < last_pte; i++)
1894 pt_vaddr[i] = scratch_pte;
1896 kunmap_px(ppgtt, pt_vaddr);
1898 num_entries -= last_pte - first_pte;
1904 static void gen6_ppgtt_insert_entries(struct i915_address_space *vm,
1905 struct sg_table *pages,
1907 enum i915_cache_level cache_level, u32 flags)
1909 struct i915_hw_ppgtt *ppgtt =
1910 container_of(vm, struct i915_hw_ppgtt, base);
1911 gen6_pte_t *pt_vaddr;
1912 unsigned first_entry = start >> PAGE_SHIFT;
1913 unsigned act_pt = first_entry / GEN6_PTES;
1914 unsigned act_pte = first_entry % GEN6_PTES;
1915 struct sg_page_iter sg_iter;
1918 for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) {
1919 if (pt_vaddr == NULL)
1920 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1923 vm->pte_encode(sg_page_iter_dma_address(&sg_iter),
1924 cache_level, true, flags);
1926 if (++act_pte == GEN6_PTES) {
1927 kunmap_px(ppgtt, pt_vaddr);
1934 kunmap_px(ppgtt, pt_vaddr);
1937 static int gen6_alloc_va_range(struct i915_address_space *vm,
1938 uint64_t start_in, uint64_t length_in)
1940 DECLARE_BITMAP(new_page_tables, I915_PDES);
1941 struct drm_device *dev = vm->dev;
1942 struct drm_i915_private *dev_priv = dev->dev_private;
1943 struct i915_hw_ppgtt *ppgtt =
1944 container_of(vm, struct i915_hw_ppgtt, base);
1945 struct i915_page_table *pt;
1946 uint32_t start, length, start_save, length_save;
1950 if (WARN_ON(start_in + length_in > ppgtt->base.total))
1953 start = start_save = start_in;
1954 length = length_save = length_in;
1956 bitmap_zero(new_page_tables, I915_PDES);
1958 /* The allocation is done in two stages so that we can bail out with
1959 * minimal amount of pain. The first stage finds new page tables that
1960 * need allocation. The second stage marks use ptes within the page
1963 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1964 if (pt != vm->scratch_pt) {
1965 // WARN_ON(bitmap_empty(pt->used_ptes, GEN6_PTES));
1969 /* We've already allocated a page table */
1970 WARN_ON(!bitmap_empty(pt->used_ptes, GEN6_PTES));
1978 gen6_initialize_pt(vm, pt);
1980 ppgtt->pd.page_table[pde] = pt;
1981 __set_bit(pde, new_page_tables);
1982 trace_i915_page_table_entry_alloc(vm, pde, start, GEN6_PDE_SHIFT);
1986 length = length_save;
1988 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1989 DECLARE_BITMAP(tmp_bitmap, GEN6_PTES);
1991 bitmap_zero(tmp_bitmap, GEN6_PTES);
1992 bitmap_set(tmp_bitmap, gen6_pte_index(start),
1993 gen6_pte_count(start, length));
1995 if (__test_and_clear_bit(pde, new_page_tables))
1996 gen6_write_pde(&ppgtt->pd, pde, pt);
1998 trace_i915_page_table_entry_map(vm, pde, pt,
1999 gen6_pte_index(start),
2000 gen6_pte_count(start, length),
2002 bitmap_or(pt->used_ptes, tmp_bitmap, pt->used_ptes,
2006 WARN_ON(!bitmap_empty(new_page_tables, I915_PDES));
2008 /* Make sure write is complete before other code can use this page
2009 * table. Also require for WC mapped PTEs */
2010 readl(dev_priv->gtt.gsm);
2012 mark_tlbs_dirty(ppgtt);
2016 for_each_set_bit(pde, new_page_tables, I915_PDES) {
2017 struct i915_page_table *pt = ppgtt->pd.page_table[pde];
2019 ppgtt->pd.page_table[pde] = vm->scratch_pt;
2020 free_pt(vm->dev, pt);
2023 mark_tlbs_dirty(ppgtt);
2027 static int gen6_init_scratch(struct i915_address_space *vm)
2029 struct drm_device *dev = vm->dev;
2031 vm->scratch_page = alloc_scratch_page(dev);
2032 if (IS_ERR(vm->scratch_page))
2033 return PTR_ERR(vm->scratch_page);
2035 vm->scratch_pt = alloc_pt(dev);
2036 if (IS_ERR(vm->scratch_pt)) {
2037 free_scratch_page(dev, vm->scratch_page);
2038 return PTR_ERR(vm->scratch_pt);
2041 gen6_initialize_pt(vm, vm->scratch_pt);
2046 static void gen6_free_scratch(struct i915_address_space *vm)
2048 struct drm_device *dev = vm->dev;
2050 free_pt(dev, vm->scratch_pt);
2051 free_scratch_page(dev, vm->scratch_page);
2054 static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
2056 struct i915_hw_ppgtt *ppgtt =
2057 container_of(vm, struct i915_hw_ppgtt, base);
2058 struct i915_page_table *pt;
2061 drm_mm_remove_node(&ppgtt->node);
2063 gen6_for_all_pdes(pt, ppgtt, pde) {
2064 if (pt != vm->scratch_pt)
2065 free_pt(ppgtt->base.dev, pt);
2068 gen6_free_scratch(vm);
2071 static int gen6_ppgtt_allocate_page_directories(struct i915_hw_ppgtt *ppgtt)
2073 struct i915_address_space *vm = &ppgtt->base;
2074 struct drm_device *dev = ppgtt->base.dev;
2075 struct drm_i915_private *dev_priv = dev->dev_private;
2076 bool retried = false;
2079 /* PPGTT PDEs reside in the GGTT and consists of 512 entries. The
2080 * allocator works in address space sizes, so it's multiplied by page
2081 * size. We allocate at the top of the GTT to avoid fragmentation.
2083 BUG_ON(!drm_mm_initialized(&dev_priv->gtt.base.mm));
2085 ret = gen6_init_scratch(vm);
2090 ret = drm_mm_insert_node_in_range_generic(&dev_priv->gtt.base.mm,
2091 &ppgtt->node, GEN6_PD_SIZE,
2093 0, dev_priv->gtt.base.total,
2095 if (ret == -ENOSPC && !retried) {
2096 ret = i915_gem_evict_something(dev, &dev_priv->gtt.base,
2097 GEN6_PD_SIZE, GEN6_PD_ALIGN,
2099 0, dev_priv->gtt.base.total,
2112 if (ppgtt->node.start < dev_priv->gtt.mappable_end)
2113 DRM_DEBUG("Forced to use aperture for PDEs\n");
2118 gen6_free_scratch(vm);
2122 static int gen6_ppgtt_alloc(struct i915_hw_ppgtt *ppgtt)
2124 return gen6_ppgtt_allocate_page_directories(ppgtt);
2127 static void gen6_scratch_va_range(struct i915_hw_ppgtt *ppgtt,
2128 uint64_t start, uint64_t length)
2130 struct i915_page_table *unused;
2133 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde)
2134 ppgtt->pd.page_table[pde] = ppgtt->base.scratch_pt;
2137 static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt, bool aliasing)
2139 struct drm_device *dev = ppgtt->base.dev;
2140 struct drm_i915_private *dev_priv = dev->dev_private;
2143 ppgtt->base.pte_encode = dev_priv->gtt.base.pte_encode;
2145 ppgtt->switch_mm = gen6_mm_switch;
2146 } else if (IS_HASWELL(dev)) {
2147 ppgtt->switch_mm = hsw_mm_switch;
2148 } else if (IS_GEN7(dev)) {
2149 ppgtt->switch_mm = gen7_mm_switch;
2153 if (intel_vgpu_active(dev))
2154 ppgtt->switch_mm = vgpu_mm_switch;
2156 ret = gen6_ppgtt_alloc(ppgtt);
2161 /* preallocate all pts */
2162 ret = alloc_pt_range(&ppgtt->pd, 0, I915_PDES,
2166 gen6_ppgtt_cleanup(&ppgtt->base);
2171 ppgtt->base.allocate_va_range = gen6_alloc_va_range;
2172 ppgtt->base.clear_range = gen6_ppgtt_clear_range;
2173 ppgtt->base.insert_entries = gen6_ppgtt_insert_entries;
2174 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
2175 ppgtt->base.bind_vma = ppgtt_bind_vma;
2176 ppgtt->base.cleanup = gen6_ppgtt_cleanup;
2177 ppgtt->base.start = 0;
2178 ppgtt->base.total = I915_PDES * GEN6_PTES * PAGE_SIZE;
2179 ppgtt->debug_dump = gen6_dump_ppgtt;
2181 ppgtt->pd.base.ggtt_offset =
2182 ppgtt->node.start / PAGE_SIZE * sizeof(gen6_pte_t);
2184 ppgtt->pd_addr = (gen6_pte_t __iomem *)dev_priv->gtt.gsm +
2185 ppgtt->pd.base.ggtt_offset / sizeof(gen6_pte_t);
2188 ppgtt->base.clear_range(&ppgtt->base, 0, ppgtt->base.total, true);
2190 gen6_scratch_va_range(ppgtt, 0, ppgtt->base.total);
2192 gen6_write_page_range(dev_priv, &ppgtt->pd, 0, ppgtt->base.total);
2194 DRM_DEBUG_DRIVER("Allocated pde space (%ldM) at GTT entry: %lx\n",
2195 ppgtt->node.size >> 20,
2196 ppgtt->node.start / PAGE_SIZE);
2198 DRM_DEBUG("Adding PPGTT at offset %x\n",
2199 ppgtt->pd.base.ggtt_offset << 10);
2204 static int __hw_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt,
2207 ppgtt->base.dev = dev;
2209 if (INTEL_INFO(dev)->gen < 8)
2210 return gen6_ppgtt_init(ppgtt, aliasing);
2212 return gen8_aliasing_ppgtt_init(ppgtt);
2214 return gen8_ppgtt_init(ppgtt);
2217 static void i915_address_space_init(struct i915_address_space *vm,
2218 struct drm_i915_private *dev_priv)
2220 drm_mm_init(&vm->mm, vm->start, vm->total);
2221 vm->dev = dev_priv->dev;
2222 INIT_LIST_HEAD(&vm->active_list);
2223 INIT_LIST_HEAD(&vm->inactive_list);
2224 list_add_tail(&vm->global_link, &dev_priv->vm_list);
2227 int i915_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
2229 struct drm_i915_private *dev_priv = dev->dev_private;
2232 ret = __hw_ppgtt_init(dev, ppgtt, false);
2234 kref_init(&ppgtt->ref);
2235 i915_address_space_init(&ppgtt->base, dev_priv);
2241 int i915_ppgtt_init_hw(struct drm_device *dev)
2243 /* In the case of execlists, PPGTT is enabled by the context descriptor
2244 * and the PDPs are contained within the context itself. We don't
2245 * need to do anything here. */
2246 if (i915.enable_execlists)
2249 if (!USES_PPGTT(dev))
2253 gen6_ppgtt_enable(dev);
2254 else if (IS_GEN7(dev))
2255 gen7_ppgtt_enable(dev);
2256 else if (INTEL_INFO(dev)->gen >= 8)
2257 gen8_ppgtt_enable(dev);
2259 MISSING_CASE(INTEL_INFO(dev)->gen);
2264 int i915_ppgtt_init_ring(struct drm_i915_gem_request *req)
2266 struct drm_i915_private *dev_priv = req->ring->dev->dev_private;
2267 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
2269 if (i915.enable_execlists)
2275 return ppgtt->switch_mm(ppgtt, req);
2278 struct i915_hw_ppgtt *
2279 i915_ppgtt_create(struct drm_device *dev, struct drm_i915_file_private *fpriv)
2281 struct i915_hw_ppgtt *ppgtt;
2284 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2286 return ERR_PTR(-ENOMEM);
2288 ret = i915_ppgtt_init(dev, ppgtt);
2291 return ERR_PTR(ret);
2294 ppgtt->file_priv = fpriv;
2296 trace_i915_ppgtt_create(&ppgtt->base);
2301 void i915_ppgtt_release(struct kref *kref)
2303 struct i915_hw_ppgtt *ppgtt =
2304 container_of(kref, struct i915_hw_ppgtt, ref);
2306 trace_i915_ppgtt_release(&ppgtt->base);
2308 /* vmas should already be unbound */
2309 WARN_ON(!list_empty(&ppgtt->base.active_list));
2310 WARN_ON(!list_empty(&ppgtt->base.inactive_list));
2312 list_del(&ppgtt->base.global_link);
2313 drm_mm_takedown(&ppgtt->base.mm);
2315 ppgtt->base.cleanup(&ppgtt->base);
2319 extern int intel_iommu_gfx_mapped;
2320 /* Certain Gen5 chipsets require require idling the GPU before
2321 * unmapping anything from the GTT when VT-d is enabled.
2323 static bool needs_idle_maps(struct drm_device *dev)
2325 #ifdef CONFIG_INTEL_IOMMU
2326 /* Query intel_iommu to see if we need the workaround. Presumably that
2329 if (IS_GEN5(dev) && IS_MOBILE(dev) && intel_iommu_gfx_mapped)
2335 static bool do_idling(struct drm_i915_private *dev_priv)
2337 bool ret = dev_priv->mm.interruptible;
2339 if (unlikely(dev_priv->gtt.do_idle_maps)) {
2340 dev_priv->mm.interruptible = false;
2341 if (i915_gpu_idle(dev_priv->dev)) {
2342 DRM_ERROR("Couldn't idle GPU\n");
2343 /* Wait a bit, in hopes it avoids the hang */
2351 static void undo_idling(struct drm_i915_private *dev_priv, bool interruptible)
2353 if (unlikely(dev_priv->gtt.do_idle_maps))
2354 dev_priv->mm.interruptible = interruptible;
2357 void i915_check_and_clear_faults(struct drm_device *dev)
2359 struct drm_i915_private *dev_priv = dev->dev_private;
2360 struct intel_engine_cs *ring;
2363 if (INTEL_INFO(dev)->gen < 6)
2366 for_each_ring(ring, dev_priv, i) {
2368 fault_reg = I915_READ(RING_FAULT_REG(ring));
2369 if (fault_reg & RING_FAULT_VALID) {
2371 DRM_DEBUG_DRIVER("Unexpected fault\n"
2373 "\tAddress space: %s\n"
2376 fault_reg & PAGE_MASK,
2377 fault_reg & RING_FAULT_GTTSEL_MASK ? "GGTT" : "PPGTT",
2378 RING_FAULT_SRCID(fault_reg),
2379 RING_FAULT_FAULT_TYPE(fault_reg));
2381 I915_WRITE(RING_FAULT_REG(ring),
2382 fault_reg & ~RING_FAULT_VALID);
2385 POSTING_READ(RING_FAULT_REG(&dev_priv->ring[RCS]));
2388 static void i915_ggtt_flush(struct drm_i915_private *dev_priv)
2390 if (INTEL_INFO(dev_priv->dev)->gen < 6) {
2391 intel_gtt_chipset_flush();
2393 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2394 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2398 void i915_gem_suspend_gtt_mappings(struct drm_device *dev)
2400 struct drm_i915_private *dev_priv = dev->dev_private;
2402 /* Don't bother messing with faults pre GEN6 as we have little
2403 * documentation supporting that it's a good idea.
2405 if (INTEL_INFO(dev)->gen < 6)
2408 i915_check_and_clear_faults(dev);
2410 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
2411 dev_priv->gtt.base.start,
2412 dev_priv->gtt.base.total,
2415 i915_ggtt_flush(dev_priv);
2418 int i915_gem_gtt_prepare_object(struct drm_i915_gem_object *obj)
2420 if (!dma_map_sg(obj->base.dev->pdev->dev,
2421 obj->pages->sgl, obj->pages->nents,
2422 PCI_DMA_BIDIRECTIONAL))
2428 static void gen8_set_pte(void __iomem *addr, gen8_pte_t pte)
2433 iowrite32((u32)pte, addr);
2434 iowrite32(pte >> 32, addr + 4);
2438 static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
2439 struct sg_table *st,
2441 enum i915_cache_level level, u32 unused)
2443 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2444 unsigned first_entry = start >> PAGE_SHIFT;
2445 gen8_pte_t __iomem *gtt_entries =
2446 (gen8_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
2448 struct sg_page_iter sg_iter;
2449 dma_addr_t addr = 0; /* shut up gcc */
2451 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
2452 addr = sg_dma_address(sg_iter.sg) +
2453 (sg_iter.sg_pgoffset << PAGE_SHIFT);
2454 gen8_set_pte(>t_entries[i],
2455 gen8_pte_encode(addr, level, true));
2460 * XXX: This serves as a posting read to make sure that the PTE has
2461 * actually been updated. There is some concern that even though
2462 * registers and PTEs are within the same BAR that they are potentially
2463 * of NUMA access patterns. Therefore, even with the way we assume
2464 * hardware should work, we must keep this posting read for paranoia.
2467 WARN_ON(readq(>t_entries[i-1])
2468 != gen8_pte_encode(addr, level, true));
2470 /* This next bit makes the above posting read even more important. We
2471 * want to flush the TLBs only after we're certain all the PTE updates
2474 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2475 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2479 * Binds an object into the global gtt with the specified cache level. The object
2480 * will be accessible to the GPU via commands whose operands reference offsets
2481 * within the global GTT as well as accessible by the GPU through the GMADR
2482 * mapped BAR (dev_priv->mm.gtt->gtt).
2484 static void gen6_ggtt_insert_entries(struct i915_address_space *vm,
2485 struct sg_table *st,
2487 enum i915_cache_level level, u32 flags)
2489 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2490 unsigned first_entry = start >> PAGE_SHIFT;
2491 gen6_pte_t __iomem *gtt_entries =
2492 (gen6_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
2494 struct sg_page_iter sg_iter;
2495 dma_addr_t addr = 0;
2497 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
2498 addr = sg_page_iter_dma_address(&sg_iter);
2499 iowrite32(vm->pte_encode(addr, level, true, flags), >t_entries[i]);
2503 /* XXX: This serves as a posting read to make sure that the PTE has
2504 * actually been updated. There is some concern that even though
2505 * registers and PTEs are within the same BAR that they are potentially
2506 * of NUMA access patterns. Therefore, even with the way we assume
2507 * hardware should work, we must keep this posting read for paranoia.
2510 unsigned long gtt = readl(>t_entries[i-1]);
2511 WARN_ON(gtt != vm->pte_encode(addr, level, true, flags));
2514 /* This next bit makes the above posting read even more important. We
2515 * want to flush the TLBs only after we're certain all the PTE updates
2518 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2519 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2522 static void gen8_ggtt_clear_range(struct i915_address_space *vm,
2527 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2528 unsigned first_entry = start >> PAGE_SHIFT;
2529 unsigned num_entries = length >> PAGE_SHIFT;
2530 gen8_pte_t scratch_pte, __iomem *gtt_base =
2531 (gen8_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
2532 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
2535 if (WARN(num_entries > max_entries,
2536 "First entry = %d; Num entries = %d (max=%d)\n",
2537 first_entry, num_entries, max_entries))
2538 num_entries = max_entries;
2540 scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
2543 for (i = 0; i < num_entries; i++)
2544 gen8_set_pte(>t_base[i], scratch_pte);
2548 static void gen6_ggtt_clear_range(struct i915_address_space *vm,
2553 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2554 unsigned first_entry = start >> PAGE_SHIFT;
2555 unsigned num_entries = length >> PAGE_SHIFT;
2556 gen6_pte_t scratch_pte, __iomem *gtt_base =
2557 (gen6_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
2558 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
2561 if (WARN(num_entries > max_entries,
2562 "First entry = %d; Num entries = %d (max=%d)\n",
2563 first_entry, num_entries, max_entries))
2564 num_entries = max_entries;
2566 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
2567 I915_CACHE_LLC, use_scratch, 0);
2569 for (i = 0; i < num_entries; i++)
2570 iowrite32(scratch_pte, >t_base[i]);
2574 static void i915_ggtt_insert_entries(struct i915_address_space *vm,
2575 struct sg_table *pages,
2577 enum i915_cache_level cache_level, u32 unused)
2579 unsigned int flags = (cache_level == I915_CACHE_NONE) ?
2580 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
2582 intel_gtt_insert_sg_entries(pages, start >> PAGE_SHIFT, flags);
2585 static void i915_ggtt_clear_range(struct i915_address_space *vm,
2590 unsigned first_entry = start >> PAGE_SHIFT;
2591 unsigned num_entries = length >> PAGE_SHIFT;
2592 intel_gtt_clear_range(first_entry, num_entries);
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 unsigned long mappable;
2737 struct drm_mm_node *entry;
2738 struct drm_i915_gem_object *obj;
2739 unsigned long hole_start, hole_end;
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: %lx + %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->mm_list, &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,
2790 "taking over the fictitious range 0x%lx-0x%lx\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, true);
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 if (drm_mm_initialized(&vm->mm)) {
2858 if (intel_vgpu_active(dev))
2859 intel_vgt_deballoon();
2861 drm_mm_takedown(&vm->mm);
2862 list_del(&vm->global_link);
2868 static unsigned int gen6_get_total_gtt_size(u16 snb_gmch_ctl)
2870 snb_gmch_ctl >>= SNB_GMCH_GGMS_SHIFT;
2871 snb_gmch_ctl &= SNB_GMCH_GGMS_MASK;
2872 return snb_gmch_ctl << 20;
2875 static unsigned int gen8_get_total_gtt_size(u16 bdw_gmch_ctl)
2877 bdw_gmch_ctl >>= BDW_GMCH_GGMS_SHIFT;
2878 bdw_gmch_ctl &= BDW_GMCH_GGMS_MASK;
2880 bdw_gmch_ctl = 1 << bdw_gmch_ctl;
2882 #ifdef CONFIG_X86_32
2883 /* Limit 32b platforms to a 2GB GGTT: 4 << 20 / pte size * PAGE_SIZE */
2884 if (bdw_gmch_ctl > 4)
2888 return bdw_gmch_ctl << 20;
2891 static unsigned int chv_get_total_gtt_size(u16 gmch_ctrl)
2893 gmch_ctrl >>= SNB_GMCH_GGMS_SHIFT;
2894 gmch_ctrl &= SNB_GMCH_GGMS_MASK;
2897 return 1 << (20 + gmch_ctrl);
2902 static size_t gen6_get_stolen_size(u16 snb_gmch_ctl)
2904 snb_gmch_ctl >>= SNB_GMCH_GMS_SHIFT;
2905 snb_gmch_ctl &= SNB_GMCH_GMS_MASK;
2906 return snb_gmch_ctl << 25; /* 32 MB units */
2909 static size_t gen8_get_stolen_size(u16 bdw_gmch_ctl)
2911 bdw_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2912 bdw_gmch_ctl &= BDW_GMCH_GMS_MASK;
2913 return bdw_gmch_ctl << 25; /* 32 MB units */
2916 static size_t chv_get_stolen_size(u16 gmch_ctrl)
2918 gmch_ctrl >>= SNB_GMCH_GMS_SHIFT;
2919 gmch_ctrl &= SNB_GMCH_GMS_MASK;
2922 * 0x0 to 0x10: 32MB increments starting at 0MB
2923 * 0x11 to 0x16: 4MB increments starting at 8MB
2924 * 0x17 to 0x1d: 4MB increments start at 36MB
2926 if (gmch_ctrl < 0x11)
2927 return gmch_ctrl << 25;
2928 else if (gmch_ctrl < 0x17)
2929 return (gmch_ctrl - 0x11 + 2) << 22;
2931 return (gmch_ctrl - 0x17 + 9) << 22;
2934 static size_t gen9_get_stolen_size(u16 gen9_gmch_ctl)
2936 gen9_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2937 gen9_gmch_ctl &= BDW_GMCH_GMS_MASK;
2939 if (gen9_gmch_ctl < 0xf0)
2940 return gen9_gmch_ctl << 25; /* 32 MB units */
2942 /* 4MB increments starting at 0xf0 for 4MB */
2943 return (gen9_gmch_ctl - 0xf0 + 1) << 22;
2946 static int ggtt_probe_common(struct drm_device *dev,
2949 struct drm_i915_private *dev_priv = dev->dev_private;
2950 struct i915_page_scratch *scratch_page;
2951 phys_addr_t gtt_phys_addr;
2953 /* For Modern GENs the PTEs and register space are split in the BAR */
2954 gtt_phys_addr = pci_resource_start(dev->pdev, 0) +
2955 (pci_resource_len(dev->pdev, 0) / 2);
2958 * On BXT writes larger than 64 bit to the GTT pagetable range will be
2959 * dropped. For WC mappings in general we have 64 byte burst writes
2960 * when the WC buffer is flushed, so we can't use it, but have to
2961 * resort to an uncached mapping. The WC issue is easily caught by the
2962 * readback check when writing GTT PTE entries.
2964 if (IS_BROXTON(dev))
2965 dev_priv->gtt.gsm = ioremap_nocache(gtt_phys_addr, gtt_size);
2967 dev_priv->gtt.gsm = ioremap_wc(gtt_phys_addr, gtt_size);
2968 if (!dev_priv->gtt.gsm) {
2969 DRM_ERROR("Failed to map the gtt page table\n");
2973 scratch_page = alloc_scratch_page(dev);
2974 if (IS_ERR(scratch_page)) {
2975 DRM_ERROR("Scratch setup failed\n");
2976 /* iounmap will also get called at remove, but meh */
2977 iounmap(dev_priv->gtt.gsm);
2978 return PTR_ERR(scratch_page);
2981 dev_priv->gtt.base.scratch_page = scratch_page;
2986 /* The GGTT and PPGTT need a private PPAT setup in order to handle cacheability
2987 * bits. When using advanced contexts each context stores its own PAT, but
2988 * writing this data shouldn't be harmful even in those cases. */
2989 static void bdw_setup_private_ppat(struct drm_i915_private *dev_priv)
2993 pat = GEN8_PPAT(0, GEN8_PPAT_WB | GEN8_PPAT_LLC) | /* for normal objects, no eLLC */
2994 GEN8_PPAT(1, GEN8_PPAT_WC | GEN8_PPAT_LLCELLC) | /* for something pointing to ptes? */
2995 GEN8_PPAT(2, GEN8_PPAT_WT | GEN8_PPAT_LLCELLC) | /* for scanout with eLLC */
2996 GEN8_PPAT(3, GEN8_PPAT_UC) | /* Uncached objects, mostly for scanout */
2997 GEN8_PPAT(4, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(0)) |
2998 GEN8_PPAT(5, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(1)) |
2999 GEN8_PPAT(6, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(2)) |
3000 GEN8_PPAT(7, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3));
3002 if (!USES_PPGTT(dev_priv->dev))
3003 /* Spec: "For GGTT, there is NO pat_sel[2:0] from the entry,
3004 * so RTL will always use the value corresponding to
3006 * So let's disable cache for GGTT to avoid screen corruptions.
3007 * MOCS still can be used though.
3008 * - System agent ggtt writes (i.e. cpu gtt mmaps) already work
3009 * before this patch, i.e. the same uncached + snooping access
3010 * like on gen6/7 seems to be in effect.
3011 * - So this just fixes blitter/render access. Again it looks
3012 * like it's not just uncached access, but uncached + snooping.
3013 * So we can still hold onto all our assumptions wrt cpu
3014 * clflushing on LLC machines.
3016 pat = GEN8_PPAT(0, GEN8_PPAT_UC);
3018 /* XXX: spec defines this as 2 distinct registers. It's unclear if a 64b
3019 * write would work. */
3020 I915_WRITE(GEN8_PRIVATE_PAT_LO, pat);
3021 I915_WRITE(GEN8_PRIVATE_PAT_HI, pat >> 32);
3024 static void chv_setup_private_ppat(struct drm_i915_private *dev_priv)
3029 * Map WB on BDW to snooped on CHV.
3031 * Only the snoop bit has meaning for CHV, the rest is
3034 * The hardware will never snoop for certain types of accesses:
3035 * - CPU GTT (GMADR->GGTT->no snoop->memory)
3036 * - PPGTT page tables
3037 * - some other special cycles
3039 * As with BDW, we also need to consider the following for GT accesses:
3040 * "For GGTT, there is NO pat_sel[2:0] from the entry,
3041 * so RTL will always use the value corresponding to
3043 * Which means we must set the snoop bit in PAT entry 0
3044 * in order to keep the global status page working.
3046 pat = GEN8_PPAT(0, CHV_PPAT_SNOOP) |
3050 GEN8_PPAT(4, CHV_PPAT_SNOOP) |
3051 GEN8_PPAT(5, CHV_PPAT_SNOOP) |
3052 GEN8_PPAT(6, CHV_PPAT_SNOOP) |
3053 GEN8_PPAT(7, CHV_PPAT_SNOOP);
3055 I915_WRITE(GEN8_PRIVATE_PAT_LO, pat);
3056 I915_WRITE(GEN8_PRIVATE_PAT_HI, pat >> 32);
3059 static int gen8_gmch_probe(struct drm_device *dev,
3062 phys_addr_t *mappable_base,
3065 struct drm_i915_private *dev_priv = dev->dev_private;
3070 /* TODO: We're not aware of mappable constraints on gen8 yet */
3071 *mappable_base = pci_resource_start(dev->pdev, 2);
3072 *mappable_end = pci_resource_len(dev->pdev, 2);
3075 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(39)))
3076 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(39));
3079 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
3081 if (INTEL_INFO(dev)->gen >= 9) {
3082 *stolen = gen9_get_stolen_size(snb_gmch_ctl);
3083 gtt_size = gen8_get_total_gtt_size(snb_gmch_ctl);
3084 } else if (IS_CHERRYVIEW(dev)) {
3085 *stolen = chv_get_stolen_size(snb_gmch_ctl);
3086 gtt_size = chv_get_total_gtt_size(snb_gmch_ctl);
3088 *stolen = gen8_get_stolen_size(snb_gmch_ctl);
3089 gtt_size = gen8_get_total_gtt_size(snb_gmch_ctl);
3092 *gtt_total = (gtt_size / sizeof(gen8_pte_t)) << PAGE_SHIFT;
3094 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
3095 chv_setup_private_ppat(dev_priv);
3097 bdw_setup_private_ppat(dev_priv);
3099 ret = ggtt_probe_common(dev, gtt_size);
3101 dev_priv->gtt.base.clear_range = gen8_ggtt_clear_range;
3102 dev_priv->gtt.base.insert_entries = gen8_ggtt_insert_entries;
3103 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
3104 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
3109 static int gen6_gmch_probe(struct drm_device *dev,
3112 phys_addr_t *mappable_base,
3115 struct drm_i915_private *dev_priv = dev->dev_private;
3116 unsigned int gtt_size;
3120 *mappable_base = pci_resource_start(dev->pdev, 2);
3121 *mappable_end = pci_resource_len(dev->pdev, 2);
3123 /* 64/512MB is the current min/max we actually know of, but this is just
3124 * a coarse sanity check.
3126 if ((*mappable_end < (64<<20) || (*mappable_end > (512<<20)))) {
3127 DRM_ERROR("Unknown GMADR size (%lx)\n",
3128 dev_priv->gtt.mappable_end);
3133 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(40)))
3134 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(40));
3136 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
3138 *stolen = gen6_get_stolen_size(snb_gmch_ctl);
3140 gtt_size = gen6_get_total_gtt_size(snb_gmch_ctl);
3141 *gtt_total = (gtt_size / sizeof(gen6_pte_t)) << PAGE_SHIFT;
3143 ret = ggtt_probe_common(dev, gtt_size);
3145 dev_priv->gtt.base.clear_range = gen6_ggtt_clear_range;
3146 dev_priv->gtt.base.insert_entries = gen6_ggtt_insert_entries;
3147 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
3148 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
3153 static void gen6_gmch_remove(struct i915_address_space *vm)
3155 struct i915_gtt *gtt = container_of(vm, struct i915_gtt, base);
3158 free_scratch_page(vm->dev, vm->scratch_page);
3161 static int i915_gmch_probe(struct drm_device *dev,
3164 phys_addr_t *mappable_base,
3167 struct drm_i915_private *dev_priv = dev->dev_private;
3171 ret = intel_gmch_probe(dev_priv->bridge_dev, dev_priv->dev->pdev, NULL);
3173 DRM_ERROR("failed to set up gmch\n");
3178 intel_gtt_get(gtt_total, stolen, mappable_base, mappable_end);
3180 dev_priv->gtt.do_idle_maps = needs_idle_maps(dev_priv->dev);
3181 dev_priv->gtt.base.insert_entries = i915_ggtt_insert_entries;
3182 dev_priv->gtt.base.clear_range = i915_ggtt_clear_range;
3183 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
3184 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
3186 if (unlikely(dev_priv->gtt.do_idle_maps))
3187 DRM_INFO("applying Ironlake quirks for intel_iommu\n");
3192 static void i915_gmch_remove(struct i915_address_space *vm)
3194 intel_gmch_remove();
3197 int i915_gem_gtt_init(struct drm_device *dev)
3199 struct drm_i915_private *dev_priv = dev->dev_private;
3200 struct i915_gtt *gtt = &dev_priv->gtt;
3203 if (INTEL_INFO(dev)->gen <= 5) {
3204 gtt->gtt_probe = i915_gmch_probe;
3205 gtt->base.cleanup = i915_gmch_remove;
3206 } else if (INTEL_INFO(dev)->gen < 8) {
3207 gtt->gtt_probe = gen6_gmch_probe;
3208 gtt->base.cleanup = gen6_gmch_remove;
3209 if (IS_HASWELL(dev) && dev_priv->ellc_size)
3210 gtt->base.pte_encode = iris_pte_encode;
3211 else if (IS_HASWELL(dev))
3212 gtt->base.pte_encode = hsw_pte_encode;
3213 else if (IS_VALLEYVIEW(dev))
3214 gtt->base.pte_encode = byt_pte_encode;
3215 else if (INTEL_INFO(dev)->gen >= 7)
3216 gtt->base.pte_encode = ivb_pte_encode;
3218 gtt->base.pte_encode = snb_pte_encode;
3220 dev_priv->gtt.gtt_probe = gen8_gmch_probe;
3221 dev_priv->gtt.base.cleanup = gen6_gmch_remove;
3224 gtt->base.dev = dev;
3226 ret = gtt->gtt_probe(dev, >t->base.total, >t->stolen_size,
3227 >t->mappable_base, >t->mappable_end);
3231 /* GMADR is the PCI mmio aperture into the global GTT. */
3232 DRM_INFO("Memory usable by graphics device = %luM\n",
3233 gtt->base.total >> 20);
3234 DRM_DEBUG_DRIVER("GMADR size = %ldM\n", gtt->mappable_end >> 20);
3235 DRM_DEBUG_DRIVER("GTT stolen size = %zdM\n", gtt->stolen_size >> 20);
3236 #ifdef CONFIG_INTEL_IOMMU
3237 if (intel_iommu_gfx_mapped)
3238 DRM_INFO("VT-d active for gfx access\n");
3241 * i915.enable_ppgtt is read-only, so do an early pass to validate the
3242 * user's requested state against the hardware/driver capabilities. We
3243 * do this now so that we can print out any log messages once rather
3244 * than every time we check intel_enable_ppgtt().
3246 i915.enable_ppgtt = sanitize_enable_ppgtt(dev, i915.enable_ppgtt);
3247 DRM_DEBUG_DRIVER("ppgtt mode: %i\n", i915.enable_ppgtt);
3252 void i915_gem_restore_gtt_mappings(struct drm_device *dev)
3254 struct drm_i915_private *dev_priv = dev->dev_private;
3255 struct drm_i915_gem_object *obj;
3256 struct i915_address_space *vm;
3258 i915_check_and_clear_faults(dev);
3260 /* First fill our portion of the GTT with scratch pages */
3261 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
3262 dev_priv->gtt.base.start,
3263 dev_priv->gtt.base.total,
3266 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
3267 struct i915_vma *vma = i915_gem_obj_to_vma(obj,
3268 &dev_priv->gtt.base);
3272 i915_gem_clflush_object(obj, obj->pin_display);
3273 WARN_ON(i915_vma_bind(vma, obj->cache_level, PIN_UPDATE));
3277 if (INTEL_INFO(dev)->gen >= 8) {
3278 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
3279 chv_setup_private_ppat(dev_priv);
3281 bdw_setup_private_ppat(dev_priv);
3286 if (USES_PPGTT(dev)) {
3287 list_for_each_entry(vm, &dev_priv->vm_list, global_link) {
3288 /* TODO: Perhaps it shouldn't be gen6 specific */
3290 struct i915_hw_ppgtt *ppgtt =
3291 container_of(vm, struct i915_hw_ppgtt,
3294 if (i915_is_ggtt(vm))
3295 ppgtt = dev_priv->mm.aliasing_ppgtt;
3297 gen6_write_page_range(dev_priv, &ppgtt->pd,
3298 0, ppgtt->base.total);
3302 i915_ggtt_flush(dev_priv);
3305 static struct i915_vma *
3306 __i915_gem_vma_create(struct drm_i915_gem_object *obj,
3307 struct i915_address_space *vm,
3308 const struct i915_ggtt_view *ggtt_view)
3310 struct i915_vma *vma;
3312 if (WARN_ON(i915_is_ggtt(vm) != !!ggtt_view))
3313 return ERR_PTR(-EINVAL);
3315 vma = kzalloc(sizeof(*vma), GFP_KERNEL);
3317 return ERR_PTR(-ENOMEM);
3319 INIT_LIST_HEAD(&vma->vma_link);
3320 INIT_LIST_HEAD(&vma->mm_list);
3321 INIT_LIST_HEAD(&vma->exec_list);
3325 if (i915_is_ggtt(vm))
3326 vma->ggtt_view = *ggtt_view;
3328 list_add_tail(&vma->vma_link, &obj->vma_list);
3329 if (!i915_is_ggtt(vm))
3330 i915_ppgtt_get(i915_vm_to_ppgtt(vm));
3336 i915_gem_obj_lookup_or_create_vma(struct drm_i915_gem_object *obj,
3337 struct i915_address_space *vm)
3339 struct i915_vma *vma;
3341 vma = i915_gem_obj_to_vma(obj, vm);
3343 vma = __i915_gem_vma_create(obj, vm,
3344 i915_is_ggtt(vm) ? &i915_ggtt_view_normal : NULL);
3350 i915_gem_obj_lookup_or_create_ggtt_vma(struct drm_i915_gem_object *obj,
3351 const struct i915_ggtt_view *view)
3353 struct i915_address_space *ggtt = i915_obj_to_ggtt(obj);
3354 struct i915_vma *vma;
3357 return ERR_PTR(-EINVAL);
3359 vma = i915_gem_obj_to_ggtt_view(obj, view);
3365 vma = __i915_gem_vma_create(obj, ggtt, view);
3371 static struct scatterlist *
3372 rotate_pages(dma_addr_t *in, unsigned int offset,
3373 unsigned int width, unsigned int height,
3374 struct sg_table *st, struct scatterlist *sg)
3376 unsigned int column, row;
3377 unsigned int src_idx;
3384 for (column = 0; column < width; column++) {
3385 src_idx = width * (height - 1) + column;
3386 for (row = 0; row < height; row++) {
3388 /* We don't need the pages, but need to initialize
3389 * the entries so the sg list can be happily traversed.
3390 * The only thing we need are DMA addresses.
3392 sg_set_page(sg, NULL, PAGE_SIZE, 0);
3393 sg_dma_address(sg) = in[offset + src_idx];
3394 sg_dma_len(sg) = PAGE_SIZE;
3403 static struct sg_table *
3404 intel_rotate_fb_obj_pages(struct i915_ggtt_view *ggtt_view,
3405 struct drm_i915_gem_object *obj)
3407 struct intel_rotation_info *rot_info = &ggtt_view->rotation_info;
3408 unsigned int size_pages = rot_info->size >> PAGE_SHIFT;
3409 unsigned int size_pages_uv;
3410 struct sg_page_iter sg_iter;
3412 dma_addr_t *page_addr_list;
3413 struct sg_table *st;
3414 unsigned int uv_start_page;
3415 struct scatterlist *sg;
3418 /* Allocate a temporary list of source pages for random access. */
3419 page_addr_list = drm_malloc_ab(obj->base.size / PAGE_SIZE,
3420 sizeof(dma_addr_t));
3421 if (!page_addr_list)
3422 return ERR_PTR(ret);
3424 /* Account for UV plane with NV12. */
3425 if (rot_info->pixel_format == DRM_FORMAT_NV12)
3426 size_pages_uv = rot_info->size_uv >> PAGE_SHIFT;
3430 /* Allocate target SG list. */
3431 st = kmalloc(sizeof(*st), M_DRM, M_WAITOK);
3435 ret = sg_alloc_table(st, size_pages + size_pages_uv, GFP_KERNEL);
3439 /* Populate source page list from the object. */
3441 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) {
3442 page_addr_list[i] = sg_page_iter_dma_address(&sg_iter);
3446 /* Rotate the pages. */
3447 sg = rotate_pages(page_addr_list, 0,
3448 rot_info->width_pages, rot_info->height_pages,
3451 /* Append the UV plane if NV12. */
3452 if (rot_info->pixel_format == DRM_FORMAT_NV12) {
3453 uv_start_page = size_pages;
3455 /* Check for tile-row un-alignment. */
3456 if (offset_in_page(rot_info->uv_offset))
3459 rot_info->uv_start_page = uv_start_page;
3461 rotate_pages(page_addr_list, uv_start_page,
3462 rot_info->width_pages_uv,
3463 rot_info->height_pages_uv,
3468 "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",
3469 obj->base.size, rot_info->pitch, rot_info->height,
3470 rot_info->pixel_format, rot_info->width_pages,
3471 rot_info->height_pages, size_pages + size_pages_uv,
3474 drm_free_large(page_addr_list);
3481 drm_free_large(page_addr_list);
3484 "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",
3485 obj->base.size, ret, rot_info->pitch, rot_info->height,
3486 rot_info->pixel_format, rot_info->width_pages,
3487 rot_info->height_pages, size_pages + size_pages_uv,
3489 return ERR_PTR(ret);
3492 static struct sg_table *
3493 intel_partial_pages(const struct i915_ggtt_view *view,
3494 struct drm_i915_gem_object *obj)
3496 struct sg_table *st;
3497 struct scatterlist *sg;
3498 struct sg_page_iter obj_sg_iter;
3501 st = kmalloc(sizeof(*st), M_DRM, M_WAITOK);
3505 ret = sg_alloc_table(st, view->params.partial.size, GFP_KERNEL);
3511 for_each_sg_page(obj->pages->sgl, &obj_sg_iter, obj->pages->nents,
3512 view->params.partial.offset)
3514 if (st->nents >= view->params.partial.size)
3517 sg_set_page(sg, NULL, PAGE_SIZE, 0);
3518 sg_dma_address(sg) = sg_page_iter_dma_address(&obj_sg_iter);
3519 sg_dma_len(sg) = PAGE_SIZE;
3530 return ERR_PTR(ret);
3534 i915_get_ggtt_vma_pages(struct i915_vma *vma)
3538 if (vma->ggtt_view.pages)
3541 if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL)
3542 vma->ggtt_view.pages = vma->obj->pages;
3543 else if (vma->ggtt_view.type == I915_GGTT_VIEW_ROTATED)
3544 vma->ggtt_view.pages =
3545 intel_rotate_fb_obj_pages(&vma->ggtt_view, vma->obj);
3546 else if (vma->ggtt_view.type == I915_GGTT_VIEW_PARTIAL)
3547 vma->ggtt_view.pages =
3548 intel_partial_pages(&vma->ggtt_view, vma->obj);
3550 WARN_ONCE(1, "GGTT view %u not implemented!\n",
3551 vma->ggtt_view.type);
3553 if (!vma->ggtt_view.pages) {
3554 DRM_ERROR("Failed to get pages for GGTT view type %u!\n",
3555 vma->ggtt_view.type);
3557 } else if (IS_ERR(vma->ggtt_view.pages)) {
3558 ret = PTR_ERR(vma->ggtt_view.pages);
3559 vma->ggtt_view.pages = NULL;
3560 DRM_ERROR("Failed to get pages for VMA view type %u (%d)!\n",
3561 vma->ggtt_view.type, ret);
3568 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
3570 * @cache_level: mapping cache level
3571 * @flags: flags like global or local mapping
3573 * DMA addresses are taken from the scatter-gather table of this object (or of
3574 * this VMA in case of non-default GGTT views) and PTE entries set up.
3575 * Note that DMA addresses are also the only part of the SG table we care about.
3577 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
3583 if (WARN_ON(flags == 0))
3587 if (flags & PIN_GLOBAL)
3588 bind_flags |= GLOBAL_BIND;
3589 if (flags & PIN_USER)
3590 bind_flags |= LOCAL_BIND;
3592 if (flags & PIN_UPDATE)
3593 bind_flags |= vma->bound;
3595 bind_flags &= ~vma->bound;
3597 if (bind_flags == 0)
3600 if (vma->bound == 0 && vma->vm->allocate_va_range) {
3601 trace_i915_va_alloc(vma->vm,
3604 VM_TO_TRACE_NAME(vma->vm));
3606 /* XXX: i915_vma_pin() will fix this +- hack */
3608 ret = vma->vm->allocate_va_range(vma->vm,
3616 ret = vma->vm->bind_vma(vma, cache_level, bind_flags);
3620 vma->bound |= bind_flags;
3626 * i915_ggtt_view_size - Get the size of a GGTT view.
3627 * @obj: Object the view is of.
3628 * @view: The view in question.
3630 * @return The size of the GGTT view in bytes.
3633 i915_ggtt_view_size(struct drm_i915_gem_object *obj,
3634 const struct i915_ggtt_view *view)
3636 if (view->type == I915_GGTT_VIEW_NORMAL) {
3637 return obj->base.size;
3638 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
3639 return view->rotation_info.size;
3640 } else if (view->type == I915_GGTT_VIEW_PARTIAL) {
3641 return view->params.partial.size << PAGE_SHIFT;
3643 WARN_ONCE(1, "GGTT view %u not implemented!\n", view->type);
3644 return obj->base.size;