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>
27 #include <linux/stop_machine.h>
29 #include <drm/i915_drm.h>
31 #include "i915_vgpu.h"
32 #include "i915_trace.h"
33 #include "intel_drv.h"
36 * DOC: Global GTT views
38 * Background and previous state
40 * Historically objects could exists (be bound) in global GTT space only as
41 * singular instances with a view representing all of the object's backing pages
42 * in a linear fashion. This view will be called a normal view.
44 * To support multiple views of the same object, where the number of mapped
45 * pages is not equal to the backing store, or where the layout of the pages
46 * is not linear, concept of a GGTT view was added.
48 * One example of an alternative view is a stereo display driven by a single
49 * image. In this case we would have a framebuffer looking like this
55 * Above would represent a normal GGTT view as normally mapped for GPU or CPU
56 * rendering. In contrast, fed to the display engine would be an alternative
57 * view which could look something like this:
62 * In this example both the size and layout of pages in the alternative view is
63 * different from the normal view.
65 * Implementation and usage
67 * GGTT views are implemented using VMAs and are distinguished via enum
68 * i915_ggtt_view_type and struct i915_ggtt_view.
70 * A new flavour of core GEM functions which work with GGTT bound objects were
71 * added with the _ggtt_ infix, and sometimes with _view postfix to avoid
72 * renaming in large amounts of code. They take the struct i915_ggtt_view
73 * parameter encapsulating all metadata required to implement a view.
75 * As a helper for callers which are only interested in the normal view,
76 * globally const i915_ggtt_view_normal singleton instance exists. All old core
77 * GEM API functions, the ones not taking the view parameter, are operating on,
78 * or with the normal GGTT view.
80 * Code wanting to add or use a new GGTT view needs to:
82 * 1. Add a new enum with a suitable name.
83 * 2. Extend the metadata in the i915_ggtt_view structure if required.
84 * 3. Add support to i915_get_vma_pages().
86 * New views are required to build a scatter-gather table from within the
87 * i915_get_vma_pages function. This table is stored in the vma.ggtt_view and
88 * exists for the lifetime of an VMA.
90 * Core API is designed to have copy semantics which means that passed in
91 * struct i915_ggtt_view does not need to be persistent (left around after
92 * calling the core API functions).
97 i915_get_ggtt_vma_pages(struct i915_vma *vma);
99 const struct i915_ggtt_view i915_ggtt_view_normal;
100 const struct i915_ggtt_view i915_ggtt_view_rotated = {
101 .type = I915_GGTT_VIEW_ROTATED
104 static int sanitize_enable_ppgtt(struct drm_device *dev, int enable_ppgtt)
106 bool has_aliasing_ppgtt;
108 bool has_full_48bit_ppgtt;
110 has_aliasing_ppgtt = INTEL_INFO(dev)->gen >= 6;
111 has_full_ppgtt = INTEL_INFO(dev)->gen >= 7;
112 has_full_48bit_ppgtt = IS_BROADWELL(dev) || INTEL_INFO(dev)->gen >= 9;
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 if (enable_ppgtt == 3 && has_full_48bit_ppgtt)
134 #ifdef CONFIG_INTEL_IOMMU
135 /* Disable ppgtt on SNB if VT-d is on. */
136 if (INTEL_INFO(dev)->gen == 6 && intel_iommu_gfx_mapped) {
137 DRM_INFO("Disabling PPGTT because VT-d is on\n");
142 /* Early VLV doesn't have this */
143 if (IS_VALLEYVIEW(dev) && !IS_CHERRYVIEW(dev) &&
144 dev->pdev->revision < 0xb) {
145 DRM_DEBUG_DRIVER("disabling PPGTT on pre-B3 step VLV\n");
149 if (INTEL_INFO(dev)->gen >= 8 && i915.enable_execlists)
150 return has_full_48bit_ppgtt ? 3 : 2;
152 return has_aliasing_ppgtt ? 1 : 0;
155 static int ppgtt_bind_vma(struct i915_vma *vma,
156 enum i915_cache_level cache_level,
161 /* Currently applicable only to VLV */
163 pte_flags |= PTE_READ_ONLY;
165 vma->vm->insert_entries(vma->vm, vma->obj->pages, vma->node.start,
166 cache_level, pte_flags);
171 static void ppgtt_unbind_vma(struct i915_vma *vma)
173 vma->vm->clear_range(vma->vm,
179 static gen8_pte_t gen8_pte_encode(dma_addr_t addr,
180 enum i915_cache_level level,
183 gen8_pte_t pte = valid ? _PAGE_PRESENT | _PAGE_RW : 0;
187 case I915_CACHE_NONE:
188 pte |= PPAT_UNCACHED_INDEX;
191 pte |= PPAT_DISPLAY_ELLC_INDEX;
194 pte |= PPAT_CACHED_INDEX;
201 static gen8_pde_t gen8_pde_encode(const dma_addr_t addr,
202 const enum i915_cache_level level)
204 gen8_pde_t pde = _PAGE_PRESENT | _PAGE_RW;
206 if (level != I915_CACHE_NONE)
207 pde |= PPAT_CACHED_PDE_INDEX;
209 pde |= PPAT_UNCACHED_INDEX;
213 #define gen8_pdpe_encode gen8_pde_encode
214 #define gen8_pml4e_encode gen8_pde_encode
216 static gen6_pte_t snb_pte_encode(dma_addr_t addr,
217 enum i915_cache_level level,
218 bool valid, u32 unused)
220 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
221 pte |= GEN6_PTE_ADDR_ENCODE(addr);
224 case I915_CACHE_L3_LLC:
226 pte |= GEN6_PTE_CACHE_LLC;
228 case I915_CACHE_NONE:
229 pte |= GEN6_PTE_UNCACHED;
238 static gen6_pte_t ivb_pte_encode(dma_addr_t addr,
239 enum i915_cache_level level,
240 bool valid, u32 unused)
242 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
243 pte |= GEN6_PTE_ADDR_ENCODE(addr);
246 case I915_CACHE_L3_LLC:
247 pte |= GEN7_PTE_CACHE_L3_LLC;
250 pte |= GEN6_PTE_CACHE_LLC;
252 case I915_CACHE_NONE:
253 pte |= GEN6_PTE_UNCACHED;
262 static gen6_pte_t byt_pte_encode(dma_addr_t addr,
263 enum i915_cache_level level,
264 bool valid, u32 flags)
266 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
267 pte |= GEN6_PTE_ADDR_ENCODE(addr);
269 if (!(flags & PTE_READ_ONLY))
270 pte |= BYT_PTE_WRITEABLE;
272 if (level != I915_CACHE_NONE)
273 pte |= BYT_PTE_SNOOPED_BY_CPU_CACHES;
278 static gen6_pte_t hsw_pte_encode(dma_addr_t addr,
279 enum i915_cache_level level,
280 bool valid, u32 unused)
282 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
283 pte |= HSW_PTE_ADDR_ENCODE(addr);
285 if (level != I915_CACHE_NONE)
286 pte |= HSW_WB_LLC_AGE3;
291 static gen6_pte_t iris_pte_encode(dma_addr_t addr,
292 enum i915_cache_level level,
293 bool valid, u32 unused)
295 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
296 pte |= HSW_PTE_ADDR_ENCODE(addr);
299 case I915_CACHE_NONE:
302 pte |= HSW_WT_ELLC_LLC_AGE3;
305 pte |= HSW_WB_ELLC_LLC_AGE3;
312 static int __setup_page_dma(struct drm_device *dev,
313 struct i915_page_dma *p, gfp_t flags)
315 struct device *device = &dev->pdev->dev;
317 p->page = alloc_page(flags);
321 p->daddr = dma_map_page(device,
322 p->page, 0, 4096, PCI_DMA_BIDIRECTIONAL);
324 if (dma_mapping_error(device, p->daddr)) {
325 __free_page(p->page);
332 static int setup_page_dma(struct drm_device *dev, struct i915_page_dma *p)
334 return __setup_page_dma(dev, p, GFP_KERNEL);
337 static void cleanup_page_dma(struct drm_device *dev, struct i915_page_dma *p)
339 if (WARN_ON(!p->page))
342 dma_unmap_page(&dev->pdev->dev, p->daddr, 4096, PCI_DMA_BIDIRECTIONAL);
343 __free_page(p->page);
344 memset(p, 0, sizeof(*p));
347 static void *kmap_page_dma(struct i915_page_dma *p)
349 return kmap_atomic(p->page);
352 /* We use the flushing unmap only with ppgtt structures:
353 * page directories, page tables and scratch pages.
355 static void kunmap_page_dma(struct drm_device *dev, void *vaddr)
357 /* There are only few exceptions for gen >=6. chv and bxt.
358 * And we are not sure about the latter so play safe for now.
360 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
361 drm_clflush_virt_range(vaddr, PAGE_SIZE);
363 kunmap_atomic(vaddr);
366 #define kmap_px(px) kmap_page_dma(px_base(px))
367 #define kunmap_px(ppgtt, vaddr) kunmap_page_dma((ppgtt)->base.dev, (vaddr))
369 #define setup_px(dev, px) setup_page_dma((dev), px_base(px))
370 #define cleanup_px(dev, px) cleanup_page_dma((dev), px_base(px))
371 #define fill_px(dev, px, v) fill_page_dma((dev), px_base(px), (v))
372 #define fill32_px(dev, px, v) fill_page_dma_32((dev), px_base(px), (v))
374 static void fill_page_dma(struct drm_device *dev, struct i915_page_dma *p,
378 uint64_t * const vaddr = kmap_page_dma(p);
380 for (i = 0; i < 512; i++)
383 kunmap_page_dma(dev, vaddr);
386 static void fill_page_dma_32(struct drm_device *dev, struct i915_page_dma *p,
387 const uint32_t val32)
393 fill_page_dma(dev, p, v);
396 static struct i915_page_scratch *alloc_scratch_page(struct drm_device *dev)
398 struct i915_page_scratch *sp;
401 sp = kzalloc(sizeof(*sp), GFP_KERNEL);
403 return ERR_PTR(-ENOMEM);
405 ret = __setup_page_dma(dev, px_base(sp), GFP_DMA32 | __GFP_ZERO);
411 set_pages_uc(px_page(sp), 1);
416 static void free_scratch_page(struct drm_device *dev,
417 struct i915_page_scratch *sp)
419 set_pages_wb(px_page(sp), 1);
425 static struct i915_page_table *alloc_pt(struct drm_device *dev)
427 struct i915_page_table *pt;
428 const size_t count = INTEL_INFO(dev)->gen >= 8 ?
429 GEN8_PTES : GEN6_PTES;
432 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
434 return ERR_PTR(-ENOMEM);
436 pt->used_ptes = kcalloc(BITS_TO_LONGS(count), sizeof(*pt->used_ptes),
442 ret = setup_px(dev, pt);
449 kfree(pt->used_ptes);
456 static void free_pt(struct drm_device *dev, struct i915_page_table *pt)
459 kfree(pt->used_ptes);
463 static void gen8_initialize_pt(struct i915_address_space *vm,
464 struct i915_page_table *pt)
466 gen8_pte_t scratch_pte;
468 scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
469 I915_CACHE_LLC, true);
471 fill_px(vm->dev, pt, scratch_pte);
474 static void gen6_initialize_pt(struct i915_address_space *vm,
475 struct i915_page_table *pt)
477 gen6_pte_t scratch_pte;
479 WARN_ON(px_dma(vm->scratch_page) == 0);
481 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
482 I915_CACHE_LLC, true, 0);
484 fill32_px(vm->dev, pt, scratch_pte);
487 static struct i915_page_directory *alloc_pd(struct drm_device *dev)
489 struct i915_page_directory *pd;
492 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
494 return ERR_PTR(-ENOMEM);
496 pd->used_pdes = kcalloc(BITS_TO_LONGS(I915_PDES),
497 sizeof(*pd->used_pdes), GFP_KERNEL);
501 ret = setup_px(dev, pd);
508 kfree(pd->used_pdes);
515 static void free_pd(struct drm_device *dev, struct i915_page_directory *pd)
519 kfree(pd->used_pdes);
524 static void gen8_initialize_pd(struct i915_address_space *vm,
525 struct i915_page_directory *pd)
527 gen8_pde_t scratch_pde;
529 scratch_pde = gen8_pde_encode(px_dma(vm->scratch_pt), I915_CACHE_LLC);
531 fill_px(vm->dev, pd, scratch_pde);
534 static int __pdp_init(struct drm_device *dev,
535 struct i915_page_directory_pointer *pdp)
537 size_t pdpes = I915_PDPES_PER_PDP(dev);
539 pdp->used_pdpes = kcalloc(BITS_TO_LONGS(pdpes),
540 sizeof(unsigned long),
542 if (!pdp->used_pdpes)
545 pdp->page_directory = kcalloc(pdpes, sizeof(*pdp->page_directory),
547 if (!pdp->page_directory) {
548 kfree(pdp->used_pdpes);
549 /* the PDP might be the statically allocated top level. Keep it
550 * as clean as possible */
551 pdp->used_pdpes = NULL;
558 static void __pdp_fini(struct i915_page_directory_pointer *pdp)
560 kfree(pdp->used_pdpes);
561 kfree(pdp->page_directory);
562 pdp->page_directory = NULL;
566 i915_page_directory_pointer *alloc_pdp(struct drm_device *dev)
568 struct i915_page_directory_pointer *pdp;
571 WARN_ON(!USES_FULL_48BIT_PPGTT(dev));
573 pdp = kzalloc(sizeof(*pdp), GFP_KERNEL);
575 return ERR_PTR(-ENOMEM);
577 ret = __pdp_init(dev, pdp);
581 ret = setup_px(dev, pdp);
595 static void free_pdp(struct drm_device *dev,
596 struct i915_page_directory_pointer *pdp)
599 if (USES_FULL_48BIT_PPGTT(dev)) {
600 cleanup_px(dev, pdp);
605 static void gen8_initialize_pdp(struct i915_address_space *vm,
606 struct i915_page_directory_pointer *pdp)
608 gen8_ppgtt_pdpe_t scratch_pdpe;
610 scratch_pdpe = gen8_pdpe_encode(px_dma(vm->scratch_pd), I915_CACHE_LLC);
612 fill_px(vm->dev, pdp, scratch_pdpe);
615 static void gen8_initialize_pml4(struct i915_address_space *vm,
616 struct i915_pml4 *pml4)
618 gen8_ppgtt_pml4e_t scratch_pml4e;
620 scratch_pml4e = gen8_pml4e_encode(px_dma(vm->scratch_pdp),
623 fill_px(vm->dev, pml4, scratch_pml4e);
627 gen8_setup_page_directory(struct i915_hw_ppgtt *ppgtt,
628 struct i915_page_directory_pointer *pdp,
629 struct i915_page_directory *pd,
632 gen8_ppgtt_pdpe_t *page_directorypo;
634 if (!USES_FULL_48BIT_PPGTT(ppgtt->base.dev))
637 page_directorypo = kmap_px(pdp);
638 page_directorypo[index] = gen8_pdpe_encode(px_dma(pd), I915_CACHE_LLC);
639 kunmap_px(ppgtt, page_directorypo);
643 gen8_setup_page_directory_pointer(struct i915_hw_ppgtt *ppgtt,
644 struct i915_pml4 *pml4,
645 struct i915_page_directory_pointer *pdp,
648 gen8_ppgtt_pml4e_t *pagemap = kmap_px(pml4);
650 WARN_ON(!USES_FULL_48BIT_PPGTT(ppgtt->base.dev));
651 pagemap[index] = gen8_pml4e_encode(px_dma(pdp), I915_CACHE_LLC);
652 kunmap_px(ppgtt, pagemap);
655 /* Broadwell Page Directory Pointer Descriptors */
656 static int gen8_write_pdp(struct drm_i915_gem_request *req,
660 struct intel_engine_cs *ring = req->ring;
665 ret = intel_ring_begin(req, 6);
669 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
670 intel_ring_emit_reg(ring, GEN8_RING_PDP_UDW(ring, entry));
671 intel_ring_emit(ring, upper_32_bits(addr));
672 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
673 intel_ring_emit_reg(ring, GEN8_RING_PDP_LDW(ring, entry));
674 intel_ring_emit(ring, lower_32_bits(addr));
675 intel_ring_advance(ring);
680 static int gen8_legacy_mm_switch(struct i915_hw_ppgtt *ppgtt,
681 struct drm_i915_gem_request *req)
685 for (i = GEN8_LEGACY_PDPES - 1; i >= 0; i--) {
686 const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i);
688 ret = gen8_write_pdp(req, i, pd_daddr);
696 static int gen8_48b_mm_switch(struct i915_hw_ppgtt *ppgtt,
697 struct drm_i915_gem_request *req)
699 return gen8_write_pdp(req, 0, px_dma(&ppgtt->pml4));
702 static void gen8_ppgtt_clear_pte_range(struct i915_address_space *vm,
703 struct i915_page_directory_pointer *pdp,
706 gen8_pte_t scratch_pte)
708 struct i915_hw_ppgtt *ppgtt =
709 container_of(vm, struct i915_hw_ppgtt, base);
710 gen8_pte_t *pt_vaddr;
711 unsigned pdpe = gen8_pdpe_index(start);
712 unsigned pde = gen8_pde_index(start);
713 unsigned pte = gen8_pte_index(start);
714 unsigned num_entries = length >> PAGE_SHIFT;
715 unsigned last_pte, i;
720 while (num_entries) {
721 struct i915_page_directory *pd;
722 struct i915_page_table *pt;
724 if (WARN_ON(!pdp->page_directory[pdpe]))
727 pd = pdp->page_directory[pdpe];
729 if (WARN_ON(!pd->page_table[pde]))
732 pt = pd->page_table[pde];
734 if (WARN_ON(!px_page(pt)))
737 last_pte = pte + num_entries;
738 if (last_pte > GEN8_PTES)
739 last_pte = GEN8_PTES;
741 pt_vaddr = kmap_px(pt);
743 for (i = pte; i < last_pte; i++) {
744 pt_vaddr[i] = scratch_pte;
748 kunmap_px(ppgtt, pt);
751 if (++pde == I915_PDES) {
752 if (++pdpe == I915_PDPES_PER_PDP(vm->dev))
759 static void gen8_ppgtt_clear_range(struct i915_address_space *vm,
764 struct i915_hw_ppgtt *ppgtt =
765 container_of(vm, struct i915_hw_ppgtt, base);
766 gen8_pte_t scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
767 I915_CACHE_LLC, use_scratch);
769 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
770 gen8_ppgtt_clear_pte_range(vm, &ppgtt->pdp, start, length,
773 uint64_t templ4, pml4e;
774 struct i915_page_directory_pointer *pdp;
776 gen8_for_each_pml4e(pdp, &ppgtt->pml4, start, length, templ4, pml4e) {
777 gen8_ppgtt_clear_pte_range(vm, pdp, start, length,
784 gen8_ppgtt_insert_pte_entries(struct i915_address_space *vm,
785 struct i915_page_directory_pointer *pdp,
786 struct sg_page_iter *sg_iter,
788 enum i915_cache_level cache_level)
790 struct i915_hw_ppgtt *ppgtt =
791 container_of(vm, struct i915_hw_ppgtt, base);
792 gen8_pte_t *pt_vaddr;
793 unsigned pdpe = gen8_pdpe_index(start);
794 unsigned pde = gen8_pde_index(start);
795 unsigned pte = gen8_pte_index(start);
799 while (__sg_page_iter_next(sg_iter)) {
800 if (pt_vaddr == NULL) {
801 struct i915_page_directory *pd = pdp->page_directory[pdpe];
802 struct i915_page_table *pt = pd->page_table[pde];
803 pt_vaddr = kmap_px(pt);
807 gen8_pte_encode(sg_page_iter_dma_address(sg_iter),
809 if (++pte == GEN8_PTES) {
810 kunmap_px(ppgtt, pt_vaddr);
812 if (++pde == I915_PDES) {
813 if (++pdpe == I915_PDPES_PER_PDP(vm->dev))
822 kunmap_px(ppgtt, pt_vaddr);
825 static void gen8_ppgtt_insert_entries(struct i915_address_space *vm,
826 struct sg_table *pages,
828 enum i915_cache_level cache_level,
831 struct i915_hw_ppgtt *ppgtt =
832 container_of(vm, struct i915_hw_ppgtt, base);
833 struct sg_page_iter sg_iter;
835 __sg_page_iter_start(&sg_iter, pages->sgl, sg_nents(pages->sgl), 0);
837 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
838 gen8_ppgtt_insert_pte_entries(vm, &ppgtt->pdp, &sg_iter, start,
841 struct i915_page_directory_pointer *pdp;
842 uint64_t templ4, pml4e;
843 uint64_t length = (uint64_t)pages->orig_nents << PAGE_SHIFT;
845 gen8_for_each_pml4e(pdp, &ppgtt->pml4, start, length, templ4, pml4e) {
846 gen8_ppgtt_insert_pte_entries(vm, pdp, &sg_iter,
852 static void gen8_free_page_tables(struct drm_device *dev,
853 struct i915_page_directory *pd)
860 for_each_set_bit(i, pd->used_pdes, I915_PDES) {
861 if (WARN_ON(!pd->page_table[i]))
864 free_pt(dev, pd->page_table[i]);
865 pd->page_table[i] = NULL;
869 static int gen8_init_scratch(struct i915_address_space *vm)
871 struct drm_device *dev = vm->dev;
873 vm->scratch_page = alloc_scratch_page(dev);
874 if (IS_ERR(vm->scratch_page))
875 return PTR_ERR(vm->scratch_page);
877 vm->scratch_pt = alloc_pt(dev);
878 if (IS_ERR(vm->scratch_pt)) {
879 free_scratch_page(dev, vm->scratch_page);
880 return PTR_ERR(vm->scratch_pt);
883 vm->scratch_pd = alloc_pd(dev);
884 if (IS_ERR(vm->scratch_pd)) {
885 free_pt(dev, vm->scratch_pt);
886 free_scratch_page(dev, vm->scratch_page);
887 return PTR_ERR(vm->scratch_pd);
890 if (USES_FULL_48BIT_PPGTT(dev)) {
891 vm->scratch_pdp = alloc_pdp(dev);
892 if (IS_ERR(vm->scratch_pdp)) {
893 free_pd(dev, vm->scratch_pd);
894 free_pt(dev, vm->scratch_pt);
895 free_scratch_page(dev, vm->scratch_page);
896 return PTR_ERR(vm->scratch_pdp);
900 gen8_initialize_pt(vm, vm->scratch_pt);
901 gen8_initialize_pd(vm, vm->scratch_pd);
902 if (USES_FULL_48BIT_PPGTT(dev))
903 gen8_initialize_pdp(vm, vm->scratch_pdp);
908 static int gen8_ppgtt_notify_vgt(struct i915_hw_ppgtt *ppgtt, bool create)
910 enum vgt_g2v_type msg;
911 struct drm_device *dev = ppgtt->base.dev;
912 struct drm_i915_private *dev_priv = dev->dev_private;
915 if (USES_FULL_48BIT_PPGTT(dev)) {
916 u64 daddr = px_dma(&ppgtt->pml4);
918 I915_WRITE(vgtif_reg(pdp[0].lo), lower_32_bits(daddr));
919 I915_WRITE(vgtif_reg(pdp[0].hi), upper_32_bits(daddr));
921 msg = (create ? VGT_G2V_PPGTT_L4_PAGE_TABLE_CREATE :
922 VGT_G2V_PPGTT_L4_PAGE_TABLE_DESTROY);
924 for (i = 0; i < GEN8_LEGACY_PDPES; i++) {
925 u64 daddr = i915_page_dir_dma_addr(ppgtt, i);
927 I915_WRITE(vgtif_reg(pdp[i].lo), lower_32_bits(daddr));
928 I915_WRITE(vgtif_reg(pdp[i].hi), upper_32_bits(daddr));
931 msg = (create ? VGT_G2V_PPGTT_L3_PAGE_TABLE_CREATE :
932 VGT_G2V_PPGTT_L3_PAGE_TABLE_DESTROY);
935 I915_WRITE(vgtif_reg(g2v_notify), msg);
940 static void gen8_free_scratch(struct i915_address_space *vm)
942 struct drm_device *dev = vm->dev;
944 if (USES_FULL_48BIT_PPGTT(dev))
945 free_pdp(dev, vm->scratch_pdp);
946 free_pd(dev, vm->scratch_pd);
947 free_pt(dev, vm->scratch_pt);
948 free_scratch_page(dev, vm->scratch_page);
951 static void gen8_ppgtt_cleanup_3lvl(struct drm_device *dev,
952 struct i915_page_directory_pointer *pdp)
956 for_each_set_bit(i, pdp->used_pdpes, I915_PDPES_PER_PDP(dev)) {
957 if (WARN_ON(!pdp->page_directory[i]))
960 gen8_free_page_tables(dev, pdp->page_directory[i]);
961 free_pd(dev, pdp->page_directory[i]);
967 static void gen8_ppgtt_cleanup_4lvl(struct i915_hw_ppgtt *ppgtt)
971 for_each_set_bit(i, ppgtt->pml4.used_pml4es, GEN8_PML4ES_PER_PML4) {
972 if (WARN_ON(!ppgtt->pml4.pdps[i]))
975 gen8_ppgtt_cleanup_3lvl(ppgtt->base.dev, ppgtt->pml4.pdps[i]);
978 cleanup_px(ppgtt->base.dev, &ppgtt->pml4);
981 static void gen8_ppgtt_cleanup(struct i915_address_space *vm)
983 struct i915_hw_ppgtt *ppgtt =
984 container_of(vm, struct i915_hw_ppgtt, base);
986 if (intel_vgpu_active(vm->dev))
987 gen8_ppgtt_notify_vgt(ppgtt, false);
989 if (!USES_FULL_48BIT_PPGTT(ppgtt->base.dev))
990 gen8_ppgtt_cleanup_3lvl(ppgtt->base.dev, &ppgtt->pdp);
992 gen8_ppgtt_cleanup_4lvl(ppgtt);
994 gen8_free_scratch(vm);
998 * gen8_ppgtt_alloc_pagetabs() - Allocate page tables for VA range.
999 * @vm: Master vm structure.
1000 * @pd: Page directory for this address range.
1001 * @start: Starting virtual address to begin allocations.
1002 * @length: Size of the allocations.
1003 * @new_pts: Bitmap set by function with new allocations. Likely used by the
1004 * caller to free on error.
1006 * Allocate the required number of page tables. Extremely similar to
1007 * gen8_ppgtt_alloc_page_directories(). The main difference is here we are limited by
1008 * the page directory boundary (instead of the page directory pointer). That
1009 * boundary is 1GB virtual. Therefore, unlike gen8_ppgtt_alloc_page_directories(), it is
1010 * possible, and likely that the caller will need to use multiple calls of this
1011 * function to achieve the appropriate allocation.
1013 * Return: 0 if success; negative error code otherwise.
1015 static int gen8_ppgtt_alloc_pagetabs(struct i915_address_space *vm,
1016 struct i915_page_directory *pd,
1019 unsigned long *new_pts)
1021 struct drm_device *dev = vm->dev;
1022 struct i915_page_table *pt;
1026 gen8_for_each_pde(pt, pd, start, length, temp, pde) {
1027 /* Don't reallocate page tables */
1028 if (test_bit(pde, pd->used_pdes)) {
1029 /* Scratch is never allocated this way */
1030 WARN_ON(pt == vm->scratch_pt);
1038 gen8_initialize_pt(vm, pt);
1039 pd->page_table[pde] = pt;
1040 __set_bit(pde, new_pts);
1041 trace_i915_page_table_entry_alloc(vm, pde, start, GEN8_PDE_SHIFT);
1047 for_each_set_bit(pde, new_pts, I915_PDES)
1048 free_pt(dev, pd->page_table[pde]);
1054 * gen8_ppgtt_alloc_page_directories() - Allocate page directories for VA range.
1055 * @vm: Master vm structure.
1056 * @pdp: Page directory pointer for this address range.
1057 * @start: Starting virtual address to begin allocations.
1058 * @length: Size of the allocations.
1059 * @new_pds: Bitmap set by function with new allocations. Likely used by the
1060 * caller to free on error.
1062 * Allocate the required number of page directories starting at the pde index of
1063 * @start, and ending at the pde index @start + @length. This function will skip
1064 * over already allocated page directories within the range, and only allocate
1065 * new ones, setting the appropriate pointer within the pdp as well as the
1066 * correct position in the bitmap @new_pds.
1068 * The function will only allocate the pages within the range for a give page
1069 * directory pointer. In other words, if @start + @length straddles a virtually
1070 * addressed PDP boundary (512GB for 4k pages), there will be more allocations
1071 * required by the caller, This is not currently possible, and the BUG in the
1072 * code will prevent it.
1074 * Return: 0 if success; negative error code otherwise.
1077 gen8_ppgtt_alloc_page_directories(struct i915_address_space *vm,
1078 struct i915_page_directory_pointer *pdp,
1081 unsigned long *new_pds)
1083 struct drm_device *dev = vm->dev;
1084 struct i915_page_directory *pd;
1087 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1089 WARN_ON(!bitmap_empty(new_pds, pdpes));
1091 gen8_for_each_pdpe(pd, pdp, start, length, temp, pdpe) {
1092 if (test_bit(pdpe, pdp->used_pdpes))
1099 gen8_initialize_pd(vm, pd);
1100 pdp->page_directory[pdpe] = pd;
1101 __set_bit(pdpe, new_pds);
1102 trace_i915_page_directory_entry_alloc(vm, pdpe, start, GEN8_PDPE_SHIFT);
1108 for_each_set_bit(pdpe, new_pds, pdpes)
1109 free_pd(dev, pdp->page_directory[pdpe]);
1115 * gen8_ppgtt_alloc_page_dirpointers() - Allocate pdps for VA range.
1116 * @vm: Master vm structure.
1117 * @pml4: Page map level 4 for this address range.
1118 * @start: Starting virtual address to begin allocations.
1119 * @length: Size of the allocations.
1120 * @new_pdps: Bitmap set by function with new allocations. Likely used by the
1121 * caller to free on error.
1123 * Allocate the required number of page directory pointers. Extremely similar to
1124 * gen8_ppgtt_alloc_page_directories() and gen8_ppgtt_alloc_pagetabs().
1125 * The main difference is here we are limited by the pml4 boundary (instead of
1126 * the page directory pointer).
1128 * Return: 0 if success; negative error code otherwise.
1131 gen8_ppgtt_alloc_page_dirpointers(struct i915_address_space *vm,
1132 struct i915_pml4 *pml4,
1135 unsigned long *new_pdps)
1137 struct drm_device *dev = vm->dev;
1138 struct i915_page_directory_pointer *pdp;
1142 WARN_ON(!bitmap_empty(new_pdps, GEN8_PML4ES_PER_PML4));
1144 gen8_for_each_pml4e(pdp, pml4, start, length, temp, pml4e) {
1145 if (!test_bit(pml4e, pml4->used_pml4es)) {
1146 pdp = alloc_pdp(dev);
1150 gen8_initialize_pdp(vm, pdp);
1151 pml4->pdps[pml4e] = pdp;
1152 __set_bit(pml4e, new_pdps);
1153 trace_i915_page_directory_pointer_entry_alloc(vm,
1163 for_each_set_bit(pml4e, new_pdps, GEN8_PML4ES_PER_PML4)
1164 free_pdp(dev, pml4->pdps[pml4e]);
1170 free_gen8_temp_bitmaps(unsigned long *new_pds, unsigned long *new_pts)
1176 /* Fills in the page directory bitmap, and the array of page tables bitmap. Both
1177 * of these are based on the number of PDPEs in the system.
1180 int __must_check alloc_gen8_temp_bitmaps(unsigned long **new_pds,
1181 unsigned long **new_pts,
1187 pds = kcalloc(BITS_TO_LONGS(pdpes), sizeof(unsigned long), GFP_TEMPORARY);
1191 pts = kcalloc(pdpes, BITS_TO_LONGS(I915_PDES) * sizeof(unsigned long),
1202 free_gen8_temp_bitmaps(pds, pts);
1206 /* PDE TLBs are a pain to invalidate on GEN8+. When we modify
1207 * the page table structures, we mark them dirty so that
1208 * context switching/execlist queuing code takes extra steps
1209 * to ensure that tlbs are flushed.
1211 static void mark_tlbs_dirty(struct i915_hw_ppgtt *ppgtt)
1213 ppgtt->pd_dirty_rings = INTEL_INFO(ppgtt->base.dev)->ring_mask;
1216 static int gen8_alloc_va_range_3lvl(struct i915_address_space *vm,
1217 struct i915_page_directory_pointer *pdp,
1221 struct i915_hw_ppgtt *ppgtt =
1222 container_of(vm, struct i915_hw_ppgtt, base);
1223 unsigned long *new_page_dirs, *new_page_tables;
1224 struct drm_device *dev = vm->dev;
1225 struct i915_page_directory *pd;
1226 const uint64_t orig_start = start;
1227 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, temp, 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, temp, 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, temp, 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);
1314 for_each_set_bit(temp, new_page_tables + pdpe *
1315 BITS_TO_LONGS(I915_PDES), I915_PDES)
1316 free_pt(dev, pdp->page_directory[pdpe]->page_table[temp]);
1319 for_each_set_bit(pdpe, new_page_dirs, pdpes)
1320 free_pd(dev, pdp->page_directory[pdpe]);
1322 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1323 mark_tlbs_dirty(ppgtt);
1327 static int gen8_alloc_va_range_4lvl(struct i915_address_space *vm,
1328 struct i915_pml4 *pml4,
1332 DECLARE_BITMAP(new_pdps, GEN8_PML4ES_PER_PML4);
1333 struct i915_hw_ppgtt *ppgtt =
1334 container_of(vm, struct i915_hw_ppgtt, base);
1335 struct i915_page_directory_pointer *pdp;
1336 uint64_t temp, pml4e;
1339 /* Do the pml4 allocations first, so we don't need to track the newly
1340 * allocated tables below the pdp */
1341 bitmap_zero(new_pdps, GEN8_PML4ES_PER_PML4);
1343 /* The pagedirectory and pagetable allocations are done in the shared 3
1344 * and 4 level code. Just allocate the pdps.
1346 ret = gen8_ppgtt_alloc_page_dirpointers(vm, pml4, start, length,
1351 WARN(bitmap_weight(new_pdps, GEN8_PML4ES_PER_PML4) > 2,
1352 "The allocation has spanned more than 512GB. "
1353 "It is highly likely this is incorrect.");
1355 gen8_for_each_pml4e(pdp, pml4, start, length, temp, pml4e) {
1358 ret = gen8_alloc_va_range_3lvl(vm, pdp, start, length);
1362 gen8_setup_page_directory_pointer(ppgtt, pml4, pdp, pml4e);
1365 bitmap_or(pml4->used_pml4es, new_pdps, pml4->used_pml4es,
1366 GEN8_PML4ES_PER_PML4);
1371 for_each_set_bit(pml4e, new_pdps, GEN8_PML4ES_PER_PML4)
1372 gen8_ppgtt_cleanup_3lvl(vm->dev, pml4->pdps[pml4e]);
1377 static int gen8_alloc_va_range(struct i915_address_space *vm,
1378 uint64_t start, uint64_t length)
1380 struct i915_hw_ppgtt *ppgtt =
1381 container_of(vm, struct i915_hw_ppgtt, base);
1383 if (USES_FULL_48BIT_PPGTT(vm->dev))
1384 return gen8_alloc_va_range_4lvl(vm, &ppgtt->pml4, start, length);
1386 return gen8_alloc_va_range_3lvl(vm, &ppgtt->pdp, start, length);
1389 static void gen8_dump_pdp(struct i915_page_directory_pointer *pdp,
1390 uint64_t start, uint64_t length,
1391 gen8_pte_t scratch_pte,
1394 struct i915_page_directory *pd;
1398 gen8_for_each_pdpe(pd, pdp, start, length, temp, pdpe) {
1399 struct i915_page_table *pt;
1400 uint64_t pd_len = length;
1401 uint64_t pd_start = start;
1404 if (!test_bit(pdpe, pdp->used_pdpes))
1407 seq_printf(m, "\tPDPE #%d\n", pdpe);
1408 gen8_for_each_pde(pt, pd, pd_start, pd_len, temp, pde) {
1410 gen8_pte_t *pt_vaddr;
1412 if (!test_bit(pde, pd->used_pdes))
1415 pt_vaddr = kmap_px(pt);
1416 for (pte = 0; pte < GEN8_PTES; pte += 4) {
1418 (pdpe << GEN8_PDPE_SHIFT) |
1419 (pde << GEN8_PDE_SHIFT) |
1420 (pte << GEN8_PTE_SHIFT);
1424 for (i = 0; i < 4; i++)
1425 if (pt_vaddr[pte + i] != scratch_pte)
1430 seq_printf(m, "\t\t0x%llx [%03d,%03d,%04d]: =", va, pdpe, pde, pte);
1431 for (i = 0; i < 4; i++) {
1432 if (pt_vaddr[pte + i] != scratch_pte)
1433 seq_printf(m, " %llx", pt_vaddr[pte + i]);
1435 seq_puts(m, " SCRATCH ");
1439 /* don't use kunmap_px, it could trigger
1440 * an unnecessary flush.
1442 kunmap_atomic(pt_vaddr);
1447 static void gen8_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
1449 struct i915_address_space *vm = &ppgtt->base;
1450 uint64_t start = ppgtt->base.start;
1451 uint64_t length = ppgtt->base.total;
1452 gen8_pte_t scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
1453 I915_CACHE_LLC, true);
1455 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
1456 gen8_dump_pdp(&ppgtt->pdp, start, length, scratch_pte, m);
1458 uint64_t templ4, pml4e;
1459 struct i915_pml4 *pml4 = &ppgtt->pml4;
1460 struct i915_page_directory_pointer *pdp;
1462 gen8_for_each_pml4e(pdp, pml4, start, length, templ4, pml4e) {
1463 if (!test_bit(pml4e, pml4->used_pml4es))
1466 seq_printf(m, " PML4E #%llu\n", pml4e);
1467 gen8_dump_pdp(pdp, start, length, scratch_pte, m);
1472 static int gen8_preallocate_top_level_pdps(struct i915_hw_ppgtt *ppgtt)
1474 unsigned long *new_page_dirs, *new_page_tables;
1475 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1478 /* We allocate temp bitmap for page tables for no gain
1479 * but as this is for init only, lets keep the things simple
1481 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables, pdpes);
1485 /* Allocate for all pdps regardless of how the ppgtt
1488 ret = gen8_ppgtt_alloc_page_directories(&ppgtt->base, &ppgtt->pdp,
1492 *ppgtt->pdp.used_pdpes = *new_page_dirs;
1494 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1500 * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers
1501 * with a net effect resembling a 2-level page table in normal x86 terms. Each
1502 * PDP represents 1GB of memory 4 * 512 * 512 * 4096 = 4GB legacy 32b address
1506 static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
1510 ret = gen8_init_scratch(&ppgtt->base);
1514 ppgtt->base.start = 0;
1515 ppgtt->base.cleanup = gen8_ppgtt_cleanup;
1516 ppgtt->base.allocate_va_range = gen8_alloc_va_range;
1517 ppgtt->base.insert_entries = gen8_ppgtt_insert_entries;
1518 ppgtt->base.clear_range = gen8_ppgtt_clear_range;
1519 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
1520 ppgtt->base.bind_vma = ppgtt_bind_vma;
1521 ppgtt->debug_dump = gen8_dump_ppgtt;
1523 if (USES_FULL_48BIT_PPGTT(ppgtt->base.dev)) {
1524 ret = setup_px(ppgtt->base.dev, &ppgtt->pml4);
1528 gen8_initialize_pml4(&ppgtt->base, &ppgtt->pml4);
1530 ppgtt->base.total = 1ULL << 48;
1531 ppgtt->switch_mm = gen8_48b_mm_switch;
1533 ret = __pdp_init(ppgtt->base.dev, &ppgtt->pdp);
1537 ppgtt->base.total = 1ULL << 32;
1538 ppgtt->switch_mm = gen8_legacy_mm_switch;
1539 trace_i915_page_directory_pointer_entry_alloc(&ppgtt->base,
1543 if (intel_vgpu_active(ppgtt->base.dev)) {
1544 ret = gen8_preallocate_top_level_pdps(ppgtt);
1550 if (intel_vgpu_active(ppgtt->base.dev))
1551 gen8_ppgtt_notify_vgt(ppgtt, true);
1556 gen8_free_scratch(&ppgtt->base);
1560 static void gen6_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
1562 struct i915_address_space *vm = &ppgtt->base;
1563 struct i915_page_table *unused;
1564 gen6_pte_t scratch_pte;
1566 uint32_t pte, pde, temp;
1567 uint32_t start = ppgtt->base.start, length = ppgtt->base.total;
1569 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1570 I915_CACHE_LLC, true, 0);
1572 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde) {
1574 gen6_pte_t *pt_vaddr;
1575 const dma_addr_t pt_addr = px_dma(ppgtt->pd.page_table[pde]);
1576 pd_entry = readl(ppgtt->pd_addr + pde);
1577 expected = (GEN6_PDE_ADDR_ENCODE(pt_addr) | GEN6_PDE_VALID);
1579 if (pd_entry != expected)
1580 seq_printf(m, "\tPDE #%d mismatch: Actual PDE: %x Expected PDE: %x\n",
1584 seq_printf(m, "\tPDE: %x\n", pd_entry);
1586 pt_vaddr = kmap_px(ppgtt->pd.page_table[pde]);
1588 for (pte = 0; pte < GEN6_PTES; pte+=4) {
1590 (pde * PAGE_SIZE * GEN6_PTES) +
1594 for (i = 0; i < 4; i++)
1595 if (pt_vaddr[pte + i] != scratch_pte)
1600 seq_printf(m, "\t\t0x%lx [%03d,%04d]: =", va, pde, pte);
1601 for (i = 0; i < 4; i++) {
1602 if (pt_vaddr[pte + i] != scratch_pte)
1603 seq_printf(m, " %08x", pt_vaddr[pte + i]);
1605 seq_puts(m, " SCRATCH ");
1609 kunmap_px(ppgtt, pt_vaddr);
1613 /* Write pde (index) from the page directory @pd to the page table @pt */
1614 static void gen6_write_pde(struct i915_page_directory *pd,
1615 const int pde, struct i915_page_table *pt)
1617 /* Caller needs to make sure the write completes if necessary */
1618 struct i915_hw_ppgtt *ppgtt =
1619 container_of(pd, struct i915_hw_ppgtt, pd);
1622 pd_entry = GEN6_PDE_ADDR_ENCODE(px_dma(pt));
1623 pd_entry |= GEN6_PDE_VALID;
1625 writel(pd_entry, ppgtt->pd_addr + pde);
1628 /* Write all the page tables found in the ppgtt structure to incrementing page
1630 static void gen6_write_page_range(struct drm_i915_private *dev_priv,
1631 struct i915_page_directory *pd,
1632 uint32_t start, uint32_t length)
1634 struct i915_page_table *pt;
1637 gen6_for_each_pde(pt, pd, start, length, temp, pde)
1638 gen6_write_pde(pd, pde, pt);
1640 /* Make sure write is complete before other code can use this page
1641 * table. Also require for WC mapped PTEs */
1642 readl(dev_priv->gtt.gsm);
1645 static uint32_t get_pd_offset(struct i915_hw_ppgtt *ppgtt)
1647 BUG_ON(ppgtt->pd.base.ggtt_offset & 0x3f);
1649 return (ppgtt->pd.base.ggtt_offset / 64) << 16;
1652 static int hsw_mm_switch(struct i915_hw_ppgtt *ppgtt,
1653 struct drm_i915_gem_request *req)
1655 struct intel_engine_cs *ring = req->ring;
1658 /* NB: TLBs must be flushed and invalidated before a switch */
1659 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1663 ret = intel_ring_begin(req, 6);
1667 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1668 intel_ring_emit_reg(ring, RING_PP_DIR_DCLV(ring));
1669 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1670 intel_ring_emit_reg(ring, RING_PP_DIR_BASE(ring));
1671 intel_ring_emit(ring, get_pd_offset(ppgtt));
1672 intel_ring_emit(ring, MI_NOOP);
1673 intel_ring_advance(ring);
1678 static int vgpu_mm_switch(struct i915_hw_ppgtt *ppgtt,
1679 struct drm_i915_gem_request *req)
1681 struct intel_engine_cs *ring = req->ring;
1682 struct drm_i915_private *dev_priv = to_i915(ppgtt->base.dev);
1684 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
1685 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
1689 static int gen7_mm_switch(struct i915_hw_ppgtt *ppgtt,
1690 struct drm_i915_gem_request *req)
1692 struct intel_engine_cs *ring = req->ring;
1695 /* NB: TLBs must be flushed and invalidated before a switch */
1696 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1700 ret = intel_ring_begin(req, 6);
1704 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1705 intel_ring_emit_reg(ring, RING_PP_DIR_DCLV(ring));
1706 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1707 intel_ring_emit_reg(ring, RING_PP_DIR_BASE(ring));
1708 intel_ring_emit(ring, get_pd_offset(ppgtt));
1709 intel_ring_emit(ring, MI_NOOP);
1710 intel_ring_advance(ring);
1712 /* XXX: RCS is the only one to auto invalidate the TLBs? */
1713 if (ring->id != RCS) {
1714 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1722 static int gen6_mm_switch(struct i915_hw_ppgtt *ppgtt,
1723 struct drm_i915_gem_request *req)
1725 struct intel_engine_cs *ring = req->ring;
1726 struct drm_device *dev = ppgtt->base.dev;
1727 struct drm_i915_private *dev_priv = dev->dev_private;
1730 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
1731 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
1733 POSTING_READ(RING_PP_DIR_DCLV(ring));
1738 static void gen8_ppgtt_enable(struct drm_device *dev)
1740 struct drm_i915_private *dev_priv = dev->dev_private;
1741 struct intel_engine_cs *ring;
1744 for_each_ring(ring, dev_priv, j) {
1745 u32 four_level = USES_FULL_48BIT_PPGTT(dev) ? GEN8_GFX_PPGTT_48B : 0;
1746 I915_WRITE(RING_MODE_GEN7(ring),
1747 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE | four_level));
1751 static void gen7_ppgtt_enable(struct drm_device *dev)
1753 struct drm_i915_private *dev_priv = dev->dev_private;
1754 struct intel_engine_cs *ring;
1755 uint32_t ecochk, ecobits;
1758 ecobits = I915_READ(GAC_ECO_BITS);
1759 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B);
1761 ecochk = I915_READ(GAM_ECOCHK);
1762 if (IS_HASWELL(dev)) {
1763 ecochk |= ECOCHK_PPGTT_WB_HSW;
1765 ecochk |= ECOCHK_PPGTT_LLC_IVB;
1766 ecochk &= ~ECOCHK_PPGTT_GFDT_IVB;
1768 I915_WRITE(GAM_ECOCHK, ecochk);
1770 for_each_ring(ring, dev_priv, i) {
1771 /* GFX_MODE is per-ring on gen7+ */
1772 I915_WRITE(RING_MODE_GEN7(ring),
1773 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1777 static void gen6_ppgtt_enable(struct drm_device *dev)
1779 struct drm_i915_private *dev_priv = dev->dev_private;
1780 uint32_t ecochk, gab_ctl, ecobits;
1782 ecobits = I915_READ(GAC_ECO_BITS);
1783 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_SNB_BIT |
1784 ECOBITS_PPGTT_CACHE64B);
1786 gab_ctl = I915_READ(GAB_CTL);
1787 I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT);
1789 ecochk = I915_READ(GAM_ECOCHK);
1790 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT | ECOCHK_PPGTT_CACHE64B);
1792 I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1795 /* PPGTT support for Sandybdrige/Gen6 and later */
1796 static void gen6_ppgtt_clear_range(struct i915_address_space *vm,
1801 struct i915_hw_ppgtt *ppgtt =
1802 container_of(vm, struct i915_hw_ppgtt, base);
1803 gen6_pte_t *pt_vaddr, scratch_pte;
1804 unsigned first_entry = start >> PAGE_SHIFT;
1805 unsigned num_entries = length >> PAGE_SHIFT;
1806 unsigned act_pt = first_entry / GEN6_PTES;
1807 unsigned first_pte = first_entry % GEN6_PTES;
1808 unsigned last_pte, i;
1810 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1811 I915_CACHE_LLC, true, 0);
1813 while (num_entries) {
1814 last_pte = first_pte + num_entries;
1815 if (last_pte > GEN6_PTES)
1816 last_pte = GEN6_PTES;
1818 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1820 for (i = first_pte; i < last_pte; i++)
1821 pt_vaddr[i] = scratch_pte;
1823 kunmap_px(ppgtt, pt_vaddr);
1825 num_entries -= last_pte - first_pte;
1831 static void gen6_ppgtt_insert_entries(struct i915_address_space *vm,
1832 struct sg_table *pages,
1834 enum i915_cache_level cache_level, u32 flags)
1836 struct i915_hw_ppgtt *ppgtt =
1837 container_of(vm, struct i915_hw_ppgtt, base);
1838 gen6_pte_t *pt_vaddr;
1839 unsigned first_entry = start >> PAGE_SHIFT;
1840 unsigned act_pt = first_entry / GEN6_PTES;
1841 unsigned act_pte = first_entry % GEN6_PTES;
1842 struct sg_page_iter sg_iter;
1845 for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) {
1846 if (pt_vaddr == NULL)
1847 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1850 vm->pte_encode(sg_page_iter_dma_address(&sg_iter),
1851 cache_level, true, flags);
1853 if (++act_pte == GEN6_PTES) {
1854 kunmap_px(ppgtt, pt_vaddr);
1861 kunmap_px(ppgtt, pt_vaddr);
1864 static int gen6_alloc_va_range(struct i915_address_space *vm,
1865 uint64_t start_in, uint64_t length_in)
1867 DECLARE_BITMAP(new_page_tables, I915_PDES);
1868 struct drm_device *dev = vm->dev;
1869 struct drm_i915_private *dev_priv = dev->dev_private;
1870 struct i915_hw_ppgtt *ppgtt =
1871 container_of(vm, struct i915_hw_ppgtt, base);
1872 struct i915_page_table *pt;
1873 uint32_t start, length, start_save, length_save;
1877 if (WARN_ON(start_in + length_in > ppgtt->base.total))
1880 start = start_save = start_in;
1881 length = length_save = length_in;
1883 bitmap_zero(new_page_tables, I915_PDES);
1885 /* The allocation is done in two stages so that we can bail out with
1886 * minimal amount of pain. The first stage finds new page tables that
1887 * need allocation. The second stage marks use ptes within the page
1890 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1891 if (pt != vm->scratch_pt) {
1892 WARN_ON(bitmap_empty(pt->used_ptes, GEN6_PTES));
1896 /* We've already allocated a page table */
1897 WARN_ON(!bitmap_empty(pt->used_ptes, GEN6_PTES));
1905 gen6_initialize_pt(vm, pt);
1907 ppgtt->pd.page_table[pde] = pt;
1908 __set_bit(pde, new_page_tables);
1909 trace_i915_page_table_entry_alloc(vm, pde, start, GEN6_PDE_SHIFT);
1913 length = length_save;
1915 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1916 DECLARE_BITMAP(tmp_bitmap, GEN6_PTES);
1918 bitmap_zero(tmp_bitmap, GEN6_PTES);
1919 bitmap_set(tmp_bitmap, gen6_pte_index(start),
1920 gen6_pte_count(start, length));
1922 if (__test_and_clear_bit(pde, new_page_tables))
1923 gen6_write_pde(&ppgtt->pd, pde, pt);
1925 trace_i915_page_table_entry_map(vm, pde, pt,
1926 gen6_pte_index(start),
1927 gen6_pte_count(start, length),
1929 bitmap_or(pt->used_ptes, tmp_bitmap, pt->used_ptes,
1933 WARN_ON(!bitmap_empty(new_page_tables, I915_PDES));
1935 /* Make sure write is complete before other code can use this page
1936 * table. Also require for WC mapped PTEs */
1937 readl(dev_priv->gtt.gsm);
1939 mark_tlbs_dirty(ppgtt);
1943 for_each_set_bit(pde, new_page_tables, I915_PDES) {
1944 struct i915_page_table *pt = ppgtt->pd.page_table[pde];
1946 ppgtt->pd.page_table[pde] = vm->scratch_pt;
1947 free_pt(vm->dev, pt);
1950 mark_tlbs_dirty(ppgtt);
1954 static int gen6_init_scratch(struct i915_address_space *vm)
1956 struct drm_device *dev = vm->dev;
1958 vm->scratch_page = alloc_scratch_page(dev);
1959 if (IS_ERR(vm->scratch_page))
1960 return PTR_ERR(vm->scratch_page);
1962 vm->scratch_pt = alloc_pt(dev);
1963 if (IS_ERR(vm->scratch_pt)) {
1964 free_scratch_page(dev, vm->scratch_page);
1965 return PTR_ERR(vm->scratch_pt);
1968 gen6_initialize_pt(vm, vm->scratch_pt);
1973 static void gen6_free_scratch(struct i915_address_space *vm)
1975 struct drm_device *dev = vm->dev;
1977 free_pt(dev, vm->scratch_pt);
1978 free_scratch_page(dev, vm->scratch_page);
1981 static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
1983 struct i915_hw_ppgtt *ppgtt =
1984 container_of(vm, struct i915_hw_ppgtt, base);
1985 struct i915_page_table *pt;
1988 drm_mm_remove_node(&ppgtt->node);
1990 gen6_for_all_pdes(pt, ppgtt, pde) {
1991 if (pt != vm->scratch_pt)
1992 free_pt(ppgtt->base.dev, pt);
1995 gen6_free_scratch(vm);
1998 static int gen6_ppgtt_allocate_page_directories(struct i915_hw_ppgtt *ppgtt)
2000 struct i915_address_space *vm = &ppgtt->base;
2001 struct drm_device *dev = ppgtt->base.dev;
2002 struct drm_i915_private *dev_priv = dev->dev_private;
2003 bool retried = false;
2006 /* PPGTT PDEs reside in the GGTT and consists of 512 entries. The
2007 * allocator works in address space sizes, so it's multiplied by page
2008 * size. We allocate at the top of the GTT to avoid fragmentation.
2010 BUG_ON(!drm_mm_initialized(&dev_priv->gtt.base.mm));
2012 ret = gen6_init_scratch(vm);
2017 ret = drm_mm_insert_node_in_range_generic(&dev_priv->gtt.base.mm,
2018 &ppgtt->node, GEN6_PD_SIZE,
2020 0, dev_priv->gtt.base.total,
2022 if (ret == -ENOSPC && !retried) {
2023 ret = i915_gem_evict_something(dev, &dev_priv->gtt.base,
2024 GEN6_PD_SIZE, GEN6_PD_ALIGN,
2026 0, dev_priv->gtt.base.total,
2039 if (ppgtt->node.start < dev_priv->gtt.mappable_end)
2040 DRM_DEBUG("Forced to use aperture for PDEs\n");
2045 gen6_free_scratch(vm);
2049 static int gen6_ppgtt_alloc(struct i915_hw_ppgtt *ppgtt)
2051 return gen6_ppgtt_allocate_page_directories(ppgtt);
2054 static void gen6_scratch_va_range(struct i915_hw_ppgtt *ppgtt,
2055 uint64_t start, uint64_t length)
2057 struct i915_page_table *unused;
2060 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde)
2061 ppgtt->pd.page_table[pde] = ppgtt->base.scratch_pt;
2064 static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
2066 struct drm_device *dev = ppgtt->base.dev;
2067 struct drm_i915_private *dev_priv = dev->dev_private;
2070 ppgtt->base.pte_encode = dev_priv->gtt.base.pte_encode;
2072 ppgtt->switch_mm = gen6_mm_switch;
2073 } else if (IS_HASWELL(dev)) {
2074 ppgtt->switch_mm = hsw_mm_switch;
2075 } else if (IS_GEN7(dev)) {
2076 ppgtt->switch_mm = gen7_mm_switch;
2080 if (intel_vgpu_active(dev))
2081 ppgtt->switch_mm = vgpu_mm_switch;
2083 ret = gen6_ppgtt_alloc(ppgtt);
2087 ppgtt->base.allocate_va_range = gen6_alloc_va_range;
2088 ppgtt->base.clear_range = gen6_ppgtt_clear_range;
2089 ppgtt->base.insert_entries = gen6_ppgtt_insert_entries;
2090 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
2091 ppgtt->base.bind_vma = ppgtt_bind_vma;
2092 ppgtt->base.cleanup = gen6_ppgtt_cleanup;
2093 ppgtt->base.start = 0;
2094 ppgtt->base.total = I915_PDES * GEN6_PTES * PAGE_SIZE;
2095 ppgtt->debug_dump = gen6_dump_ppgtt;
2097 ppgtt->pd.base.ggtt_offset =
2098 ppgtt->node.start / PAGE_SIZE * sizeof(gen6_pte_t);
2100 ppgtt->pd_addr = (gen6_pte_t __iomem *)dev_priv->gtt.gsm +
2101 ppgtt->pd.base.ggtt_offset / sizeof(gen6_pte_t);
2103 gen6_scratch_va_range(ppgtt, 0, ppgtt->base.total);
2105 gen6_write_page_range(dev_priv, &ppgtt->pd, 0, ppgtt->base.total);
2107 DRM_DEBUG_DRIVER("Allocated pde space (%lldM) at GTT entry: %llx\n",
2108 ppgtt->node.size >> 20,
2109 ppgtt->node.start / PAGE_SIZE);
2111 DRM_DEBUG("Adding PPGTT at offset %x\n",
2112 ppgtt->pd.base.ggtt_offset << 10);
2117 static int __hw_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
2119 ppgtt->base.dev = dev;
2121 if (INTEL_INFO(dev)->gen < 8)
2122 return gen6_ppgtt_init(ppgtt);
2124 return gen8_ppgtt_init(ppgtt);
2127 static void i915_address_space_init(struct i915_address_space *vm,
2128 struct drm_i915_private *dev_priv)
2130 drm_mm_init(&vm->mm, vm->start, vm->total);
2131 vm->dev = dev_priv->dev;
2132 INIT_LIST_HEAD(&vm->active_list);
2133 INIT_LIST_HEAD(&vm->inactive_list);
2134 list_add_tail(&vm->global_link, &dev_priv->vm_list);
2137 int i915_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
2139 struct drm_i915_private *dev_priv = dev->dev_private;
2142 ret = __hw_ppgtt_init(dev, ppgtt);
2144 kref_init(&ppgtt->ref);
2145 i915_address_space_init(&ppgtt->base, dev_priv);
2151 int i915_ppgtt_init_hw(struct drm_device *dev)
2153 /* In the case of execlists, PPGTT is enabled by the context descriptor
2154 * and the PDPs are contained within the context itself. We don't
2155 * need to do anything here. */
2156 if (i915.enable_execlists)
2159 if (!USES_PPGTT(dev))
2163 gen6_ppgtt_enable(dev);
2164 else if (IS_GEN7(dev))
2165 gen7_ppgtt_enable(dev);
2166 else if (INTEL_INFO(dev)->gen >= 8)
2167 gen8_ppgtt_enable(dev);
2169 MISSING_CASE(INTEL_INFO(dev)->gen);
2174 int i915_ppgtt_init_ring(struct drm_i915_gem_request *req)
2176 struct drm_i915_private *dev_priv = req->ring->dev->dev_private;
2177 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
2179 if (i915.enable_execlists)
2185 return ppgtt->switch_mm(ppgtt, req);
2188 struct i915_hw_ppgtt *
2189 i915_ppgtt_create(struct drm_device *dev, struct drm_i915_file_private *fpriv)
2191 struct i915_hw_ppgtt *ppgtt;
2194 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2196 return ERR_PTR(-ENOMEM);
2198 ret = i915_ppgtt_init(dev, ppgtt);
2201 return ERR_PTR(ret);
2204 ppgtt->file_priv = fpriv;
2206 trace_i915_ppgtt_create(&ppgtt->base);
2211 void i915_ppgtt_release(struct kref *kref)
2213 struct i915_hw_ppgtt *ppgtt =
2214 container_of(kref, struct i915_hw_ppgtt, ref);
2216 trace_i915_ppgtt_release(&ppgtt->base);
2218 /* vmas should already be unbound */
2219 WARN_ON(!list_empty(&ppgtt->base.active_list));
2220 WARN_ON(!list_empty(&ppgtt->base.inactive_list));
2222 list_del(&ppgtt->base.global_link);
2223 drm_mm_takedown(&ppgtt->base.mm);
2225 ppgtt->base.cleanup(&ppgtt->base);
2229 extern int intel_iommu_gfx_mapped;
2230 /* Certain Gen5 chipsets require require idling the GPU before
2231 * unmapping anything from the GTT when VT-d is enabled.
2233 static bool needs_idle_maps(struct drm_device *dev)
2235 #ifdef CONFIG_INTEL_IOMMU
2236 /* Query intel_iommu to see if we need the workaround. Presumably that
2239 if (IS_GEN5(dev) && IS_MOBILE(dev) && intel_iommu_gfx_mapped)
2245 static bool do_idling(struct drm_i915_private *dev_priv)
2247 bool ret = dev_priv->mm.interruptible;
2249 if (unlikely(dev_priv->gtt.do_idle_maps)) {
2250 dev_priv->mm.interruptible = false;
2251 if (i915_gpu_idle(dev_priv->dev)) {
2252 DRM_ERROR("Couldn't idle GPU\n");
2253 /* Wait a bit, in hopes it avoids the hang */
2261 static void undo_idling(struct drm_i915_private *dev_priv, bool interruptible)
2263 if (unlikely(dev_priv->gtt.do_idle_maps))
2264 dev_priv->mm.interruptible = interruptible;
2267 void i915_check_and_clear_faults(struct drm_device *dev)
2269 struct drm_i915_private *dev_priv = dev->dev_private;
2270 struct intel_engine_cs *ring;
2273 if (INTEL_INFO(dev)->gen < 6)
2276 for_each_ring(ring, dev_priv, i) {
2278 fault_reg = I915_READ(RING_FAULT_REG(ring));
2279 if (fault_reg & RING_FAULT_VALID) {
2280 DRM_DEBUG_DRIVER("Unexpected fault\n"
2282 "\tAddress space: %s\n"
2285 fault_reg & PAGE_MASK,
2286 fault_reg & RING_FAULT_GTTSEL_MASK ? "GGTT" : "PPGTT",
2287 RING_FAULT_SRCID(fault_reg),
2288 RING_FAULT_FAULT_TYPE(fault_reg));
2289 I915_WRITE(RING_FAULT_REG(ring),
2290 fault_reg & ~RING_FAULT_VALID);
2293 POSTING_READ(RING_FAULT_REG(&dev_priv->ring[RCS]));
2296 static void i915_ggtt_flush(struct drm_i915_private *dev_priv)
2298 if (INTEL_INFO(dev_priv->dev)->gen < 6) {
2299 intel_gtt_chipset_flush();
2301 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2302 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2306 void i915_gem_suspend_gtt_mappings(struct drm_device *dev)
2308 struct drm_i915_private *dev_priv = dev->dev_private;
2310 /* Don't bother messing with faults pre GEN6 as we have little
2311 * documentation supporting that it's a good idea.
2313 if (INTEL_INFO(dev)->gen < 6)
2316 i915_check_and_clear_faults(dev);
2318 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
2319 dev_priv->gtt.base.start,
2320 dev_priv->gtt.base.total,
2323 i915_ggtt_flush(dev_priv);
2326 int i915_gem_gtt_prepare_object(struct drm_i915_gem_object *obj)
2328 if (!dma_map_sg(&obj->base.dev->pdev->dev,
2329 obj->pages->sgl, obj->pages->nents,
2330 PCI_DMA_BIDIRECTIONAL))
2336 static void gen8_set_pte(void __iomem *addr, gen8_pte_t pte)
2341 iowrite32((u32)pte, addr);
2342 iowrite32(pte >> 32, addr + 4);
2346 static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
2347 struct sg_table *st,
2349 enum i915_cache_level level, u32 unused)
2351 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2352 unsigned first_entry = start >> PAGE_SHIFT;
2353 gen8_pte_t __iomem *gtt_entries =
2354 (gen8_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
2356 struct sg_page_iter sg_iter;
2357 dma_addr_t addr = 0; /* shut up gcc */
2359 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
2360 addr = sg_dma_address(sg_iter.sg) +
2361 (sg_iter.sg_pgoffset << PAGE_SHIFT);
2362 gen8_set_pte(>t_entries[i],
2363 gen8_pte_encode(addr, level, true));
2368 * XXX: This serves as a posting read to make sure that the PTE has
2369 * actually been updated. There is some concern that even though
2370 * registers and PTEs are within the same BAR that they are potentially
2371 * of NUMA access patterns. Therefore, even with the way we assume
2372 * hardware should work, we must keep this posting read for paranoia.
2375 WARN_ON(readq(>t_entries[i-1])
2376 != gen8_pte_encode(addr, level, true));
2378 /* This next bit makes the above posting read even more important. We
2379 * want to flush the TLBs only after we're certain all the PTE updates
2382 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2383 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2387 * Binds an object into the global gtt with the specified cache level. The object
2388 * will be accessible to the GPU via commands whose operands reference offsets
2389 * within the global GTT as well as accessible by the GPU through the GMADR
2390 * mapped BAR (dev_priv->mm.gtt->gtt).
2392 static void gen6_ggtt_insert_entries(struct i915_address_space *vm,
2393 struct sg_table *st,
2395 enum i915_cache_level level, u32 flags)
2397 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2398 unsigned first_entry = start >> PAGE_SHIFT;
2399 gen6_pte_t __iomem *gtt_entries =
2400 (gen6_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
2402 struct sg_page_iter sg_iter;
2403 dma_addr_t addr = 0;
2405 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
2406 addr = sg_page_iter_dma_address(&sg_iter);
2407 iowrite32(vm->pte_encode(addr, level, true, flags), >t_entries[i]);
2411 /* XXX: This serves as a posting read to make sure that the PTE has
2412 * actually been updated. There is some concern that even though
2413 * registers and PTEs are within the same BAR that they are potentially
2414 * of NUMA access patterns. Therefore, even with the way we assume
2415 * hardware should work, we must keep this posting read for paranoia.
2418 unsigned long gtt = readl(>t_entries[i-1]);
2419 WARN_ON(gtt != vm->pte_encode(addr, level, true, flags));
2422 /* This next bit makes the above posting read even more important. We
2423 * want to flush the TLBs only after we're certain all the PTE updates
2426 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2427 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2430 static void gen8_ggtt_clear_range(struct i915_address_space *vm,
2435 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2436 unsigned first_entry = start >> PAGE_SHIFT;
2437 unsigned num_entries = length >> PAGE_SHIFT;
2438 gen8_pte_t scratch_pte, __iomem *gtt_base =
2439 (gen8_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
2440 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
2443 if (WARN(num_entries > max_entries,
2444 "First entry = %d; Num entries = %d (max=%d)\n",
2445 first_entry, num_entries, max_entries))
2446 num_entries = max_entries;
2448 scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
2451 for (i = 0; i < num_entries; i++)
2452 gen8_set_pte(>t_base[i], scratch_pte);
2456 static void gen6_ggtt_clear_range(struct i915_address_space *vm,
2461 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2462 unsigned first_entry = start >> PAGE_SHIFT;
2463 unsigned num_entries = length >> PAGE_SHIFT;
2464 gen6_pte_t scratch_pte, __iomem *gtt_base =
2465 (gen6_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
2466 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
2469 if (WARN(num_entries > max_entries,
2470 "First entry = %d; Num entries = %d (max=%d)\n",
2471 first_entry, num_entries, max_entries))
2472 num_entries = max_entries;
2474 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
2475 I915_CACHE_LLC, use_scratch, 0);
2477 for (i = 0; i < num_entries; i++)
2478 iowrite32(scratch_pte, >t_base[i]);
2482 static void i915_ggtt_insert_entries(struct i915_address_space *vm,
2483 struct sg_table *pages,
2485 enum i915_cache_level cache_level, u32 unused)
2487 unsigned int flags = (cache_level == I915_CACHE_NONE) ?
2488 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
2490 intel_gtt_insert_sg_entries(pages, start >> PAGE_SHIFT, flags);
2494 static void i915_ggtt_clear_range(struct i915_address_space *vm,
2499 unsigned first_entry = start >> PAGE_SHIFT;
2500 unsigned num_entries = length >> PAGE_SHIFT;
2501 intel_gtt_clear_range(first_entry, num_entries);
2504 static int ggtt_bind_vma(struct i915_vma *vma,
2505 enum i915_cache_level cache_level,
2508 struct drm_i915_gem_object *obj = vma->obj;
2512 ret = i915_get_ggtt_vma_pages(vma);
2516 /* Currently applicable only to VLV */
2518 pte_flags |= PTE_READ_ONLY;
2520 vma->vm->insert_entries(vma->vm, vma->ggtt_view.pages,
2522 cache_level, pte_flags);
2525 * Without aliasing PPGTT there's no difference between
2526 * GLOBAL/LOCAL_BIND, it's all the same ptes. Hence unconditionally
2527 * upgrade to both bound if we bind either to avoid double-binding.
2529 vma->bound |= GLOBAL_BIND | LOCAL_BIND;
2534 struct ggtt_bind_vma__cb {
2535 struct i915_vma *vma;
2536 enum i915_cache_level cache_level;
2540 static int ggtt_bind_vma__cb(void *_arg)
2542 struct ggtt_bind_vma__cb *arg = _arg;
2543 return ggtt_bind_vma(arg->vma, arg->cache_level, arg->flags);
2546 static int ggtt_bind_vma__BKL(struct i915_vma *vma,
2547 enum i915_cache_level cache_level,
2550 struct ggtt_bind_vma__cb arg = { vma, cache_level, flags };
2551 return stop_machine(ggtt_bind_vma__cb, &arg, NULL);
2554 static int aliasing_gtt_bind_vma(struct i915_vma *vma,
2555 enum i915_cache_level cache_level,
2558 struct drm_device *dev = vma->vm->dev;
2559 struct drm_i915_private *dev_priv = dev->dev_private;
2560 struct drm_i915_gem_object *obj = vma->obj;
2561 struct sg_table *pages = obj->pages;
2565 ret = i915_get_ggtt_vma_pages(vma);
2568 pages = vma->ggtt_view.pages;
2570 /* Currently applicable only to VLV */
2572 pte_flags |= PTE_READ_ONLY;
2575 if (flags & GLOBAL_BIND) {
2576 vma->vm->insert_entries(vma->vm, pages,
2578 cache_level, pte_flags);
2581 if (flags & LOCAL_BIND) {
2582 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
2583 appgtt->base.insert_entries(&appgtt->base, pages,
2585 cache_level, pte_flags);
2591 static void ggtt_unbind_vma(struct i915_vma *vma)
2593 struct drm_device *dev = vma->vm->dev;
2594 struct drm_i915_private *dev_priv = dev->dev_private;
2595 struct drm_i915_gem_object *obj = vma->obj;
2596 const uint64_t size = min_t(uint64_t,
2600 if (vma->bound & GLOBAL_BIND) {
2601 vma->vm->clear_range(vma->vm,
2607 if (dev_priv->mm.aliasing_ppgtt && vma->bound & LOCAL_BIND) {
2608 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
2610 appgtt->base.clear_range(&appgtt->base,
2617 void i915_gem_gtt_finish_object(struct drm_i915_gem_object *obj)
2619 struct drm_device *dev = obj->base.dev;
2620 struct drm_i915_private *dev_priv = dev->dev_private;
2623 interruptible = do_idling(dev_priv);
2625 dma_unmap_sg(&dev->pdev->dev, obj->pages->sgl, obj->pages->nents,
2626 PCI_DMA_BIDIRECTIONAL);
2628 undo_idling(dev_priv, interruptible);
2631 static void i915_gtt_color_adjust(struct drm_mm_node *node,
2632 unsigned long color,
2636 if (node->color != color)
2639 if (!list_empty(&node->node_list)) {
2640 node = list_entry(node->node_list.next,
2643 if (node->allocated && node->color != color)
2648 static int i915_gem_setup_global_gtt(struct drm_device *dev,
2653 /* Let GEM Manage all of the aperture.
2655 * However, leave one page at the end still bound to the scratch page.
2656 * There are a number of places where the hardware apparently prefetches
2657 * past the end of the object, and we've seen multiple hangs with the
2658 * GPU head pointer stuck in a batchbuffer bound at the last page of the
2659 * aperture. One page should be enough to keep any prefetching inside
2662 struct drm_i915_private *dev_priv = dev->dev_private;
2663 struct i915_address_space *ggtt_vm = &dev_priv->gtt.base;
2664 struct drm_mm_node *entry;
2665 struct drm_i915_gem_object *obj;
2666 unsigned long hole_start, hole_end;
2669 BUG_ON(mappable_end > end);
2671 ggtt_vm->start = start;
2673 /* Subtract the guard page before address space initialization to
2674 * shrink the range used by drm_mm */
2675 ggtt_vm->total = end - start - PAGE_SIZE;
2676 i915_address_space_init(ggtt_vm, dev_priv);
2677 ggtt_vm->total += PAGE_SIZE;
2679 if (intel_vgpu_active(dev)) {
2680 ret = intel_vgt_balloon(dev);
2686 ggtt_vm->mm.color_adjust = i915_gtt_color_adjust;
2688 /* Mark any preallocated objects as occupied */
2689 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
2690 struct i915_vma *vma = i915_gem_obj_to_vma(obj, ggtt_vm);
2692 DRM_DEBUG_KMS("reserving preallocated space: %llx + %zx\n",
2693 i915_gem_obj_ggtt_offset(obj), obj->base.size);
2695 WARN_ON(i915_gem_obj_ggtt_bound(obj));
2696 ret = drm_mm_reserve_node(&ggtt_vm->mm, &vma->node);
2698 DRM_DEBUG_KMS("Reservation failed: %i\n", ret);
2701 vma->bound |= GLOBAL_BIND;
2702 list_add_tail(&vma->mm_list, &ggtt_vm->inactive_list);
2705 /* Clear any non-preallocated blocks */
2706 drm_mm_for_each_hole(entry, &ggtt_vm->mm, hole_start, hole_end) {
2707 DRM_DEBUG_KMS("clearing unused GTT space: [%lx, %lx]\n",
2708 hole_start, hole_end);
2709 ggtt_vm->clear_range(ggtt_vm, hole_start,
2710 hole_end - hole_start, true);
2713 /* And finally clear the reserved guard page */
2714 ggtt_vm->clear_range(ggtt_vm, end - PAGE_SIZE, PAGE_SIZE, true);
2716 if (USES_PPGTT(dev) && !USES_FULL_PPGTT(dev)) {
2717 struct i915_hw_ppgtt *ppgtt;
2719 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2723 ret = __hw_ppgtt_init(dev, ppgtt);
2725 ppgtt->base.cleanup(&ppgtt->base);
2730 if (ppgtt->base.allocate_va_range)
2731 ret = ppgtt->base.allocate_va_range(&ppgtt->base, 0,
2734 ppgtt->base.cleanup(&ppgtt->base);
2739 ppgtt->base.clear_range(&ppgtt->base,
2744 dev_priv->mm.aliasing_ppgtt = ppgtt;
2745 WARN_ON(dev_priv->gtt.base.bind_vma != ggtt_bind_vma);
2746 dev_priv->gtt.base.bind_vma = aliasing_gtt_bind_vma;
2752 void i915_gem_init_global_gtt(struct drm_device *dev)
2754 struct drm_i915_private *dev_priv = dev->dev_private;
2755 u64 gtt_size, mappable_size;
2757 gtt_size = dev_priv->gtt.base.total;
2758 mappable_size = dev_priv->gtt.mappable_end;
2760 i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
2763 void i915_global_gtt_cleanup(struct drm_device *dev)
2765 struct drm_i915_private *dev_priv = dev->dev_private;
2766 struct i915_address_space *vm = &dev_priv->gtt.base;
2768 if (dev_priv->mm.aliasing_ppgtt) {
2769 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
2771 ppgtt->base.cleanup(&ppgtt->base);
2774 if (drm_mm_initialized(&vm->mm)) {
2775 if (intel_vgpu_active(dev))
2776 intel_vgt_deballoon();
2778 drm_mm_takedown(&vm->mm);
2779 list_del(&vm->global_link);
2785 static unsigned int gen6_get_total_gtt_size(u16 snb_gmch_ctl)
2787 snb_gmch_ctl >>= SNB_GMCH_GGMS_SHIFT;
2788 snb_gmch_ctl &= SNB_GMCH_GGMS_MASK;
2789 return snb_gmch_ctl << 20;
2792 static unsigned int gen8_get_total_gtt_size(u16 bdw_gmch_ctl)
2794 bdw_gmch_ctl >>= BDW_GMCH_GGMS_SHIFT;
2795 bdw_gmch_ctl &= BDW_GMCH_GGMS_MASK;
2797 bdw_gmch_ctl = 1 << bdw_gmch_ctl;
2799 #ifdef CONFIG_X86_32
2800 /* Limit 32b platforms to a 2GB GGTT: 4 << 20 / pte size * PAGE_SIZE */
2801 if (bdw_gmch_ctl > 4)
2805 return bdw_gmch_ctl << 20;
2808 static unsigned int chv_get_total_gtt_size(u16 gmch_ctrl)
2810 gmch_ctrl >>= SNB_GMCH_GGMS_SHIFT;
2811 gmch_ctrl &= SNB_GMCH_GGMS_MASK;
2814 return 1 << (20 + gmch_ctrl);
2819 static size_t gen6_get_stolen_size(u16 snb_gmch_ctl)
2821 snb_gmch_ctl >>= SNB_GMCH_GMS_SHIFT;
2822 snb_gmch_ctl &= SNB_GMCH_GMS_MASK;
2823 return snb_gmch_ctl << 25; /* 32 MB units */
2826 static size_t gen8_get_stolen_size(u16 bdw_gmch_ctl)
2828 bdw_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2829 bdw_gmch_ctl &= BDW_GMCH_GMS_MASK;
2830 return bdw_gmch_ctl << 25; /* 32 MB units */
2833 static size_t chv_get_stolen_size(u16 gmch_ctrl)
2835 gmch_ctrl >>= SNB_GMCH_GMS_SHIFT;
2836 gmch_ctrl &= SNB_GMCH_GMS_MASK;
2839 * 0x0 to 0x10: 32MB increments starting at 0MB
2840 * 0x11 to 0x16: 4MB increments starting at 8MB
2841 * 0x17 to 0x1d: 4MB increments start at 36MB
2843 if (gmch_ctrl < 0x11)
2844 return gmch_ctrl << 25;
2845 else if (gmch_ctrl < 0x17)
2846 return (gmch_ctrl - 0x11 + 2) << 22;
2848 return (gmch_ctrl - 0x17 + 9) << 22;
2851 static size_t gen9_get_stolen_size(u16 gen9_gmch_ctl)
2853 gen9_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2854 gen9_gmch_ctl &= BDW_GMCH_GMS_MASK;
2856 if (gen9_gmch_ctl < 0xf0)
2857 return gen9_gmch_ctl << 25; /* 32 MB units */
2859 /* 4MB increments starting at 0xf0 for 4MB */
2860 return (gen9_gmch_ctl - 0xf0 + 1) << 22;
2863 static int ggtt_probe_common(struct drm_device *dev,
2866 struct drm_i915_private *dev_priv = dev->dev_private;
2867 struct i915_page_scratch *scratch_page;
2868 phys_addr_t gtt_phys_addr;
2870 /* For Modern GENs the PTEs and register space are split in the BAR */
2871 gtt_phys_addr = pci_resource_start(dev->pdev, 0) +
2872 (pci_resource_len(dev->pdev, 0) / 2);
2875 * On BXT writes larger than 64 bit to the GTT pagetable range will be
2876 * dropped. For WC mappings in general we have 64 byte burst writes
2877 * when the WC buffer is flushed, so we can't use it, but have to
2878 * resort to an uncached mapping. The WC issue is easily caught by the
2879 * readback check when writing GTT PTE entries.
2881 if (IS_BROXTON(dev))
2882 dev_priv->gtt.gsm = ioremap_nocache(gtt_phys_addr, gtt_size);
2884 dev_priv->gtt.gsm = ioremap_wc(gtt_phys_addr, gtt_size);
2885 if (!dev_priv->gtt.gsm) {
2886 DRM_ERROR("Failed to map the gtt page table\n");
2890 scratch_page = alloc_scratch_page(dev);
2891 if (IS_ERR(scratch_page)) {
2892 DRM_ERROR("Scratch setup failed\n");
2893 /* iounmap will also get called at remove, but meh */
2894 iounmap(dev_priv->gtt.gsm);
2895 return PTR_ERR(scratch_page);
2898 dev_priv->gtt.base.scratch_page = scratch_page;
2903 /* The GGTT and PPGTT need a private PPAT setup in order to handle cacheability
2904 * bits. When using advanced contexts each context stores its own PAT, but
2905 * writing this data shouldn't be harmful even in those cases. */
2906 static void bdw_setup_private_ppat(struct drm_i915_private *dev_priv)
2910 pat = GEN8_PPAT(0, GEN8_PPAT_WB | GEN8_PPAT_LLC) | /* for normal objects, no eLLC */
2911 GEN8_PPAT(1, GEN8_PPAT_WC | GEN8_PPAT_LLCELLC) | /* for something pointing to ptes? */
2912 GEN8_PPAT(2, GEN8_PPAT_WT | GEN8_PPAT_LLCELLC) | /* for scanout with eLLC */
2913 GEN8_PPAT(3, GEN8_PPAT_UC) | /* Uncached objects, mostly for scanout */
2914 GEN8_PPAT(4, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(0)) |
2915 GEN8_PPAT(5, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(1)) |
2916 GEN8_PPAT(6, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(2)) |
2917 GEN8_PPAT(7, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3));
2919 if (!USES_PPGTT(dev_priv->dev))
2920 /* Spec: "For GGTT, there is NO pat_sel[2:0] from the entry,
2921 * so RTL will always use the value corresponding to
2923 * So let's disable cache for GGTT to avoid screen corruptions.
2924 * MOCS still can be used though.
2925 * - System agent ggtt writes (i.e. cpu gtt mmaps) already work
2926 * before this patch, i.e. the same uncached + snooping access
2927 * like on gen6/7 seems to be in effect.
2928 * - So this just fixes blitter/render access. Again it looks
2929 * like it's not just uncached access, but uncached + snooping.
2930 * So we can still hold onto all our assumptions wrt cpu
2931 * clflushing on LLC machines.
2933 pat = GEN8_PPAT(0, GEN8_PPAT_UC);
2935 /* XXX: spec defines this as 2 distinct registers. It's unclear if a 64b
2936 * write would work. */
2937 I915_WRITE(GEN8_PRIVATE_PAT_LO, pat);
2938 I915_WRITE(GEN8_PRIVATE_PAT_HI, pat >> 32);
2941 static void chv_setup_private_ppat(struct drm_i915_private *dev_priv)
2946 * Map WB on BDW to snooped on CHV.
2948 * Only the snoop bit has meaning for CHV, the rest is
2951 * The hardware will never snoop for certain types of accesses:
2952 * - CPU GTT (GMADR->GGTT->no snoop->memory)
2953 * - PPGTT page tables
2954 * - some other special cycles
2956 * As with BDW, we also need to consider the following for GT accesses:
2957 * "For GGTT, there is NO pat_sel[2:0] from the entry,
2958 * so RTL will always use the value corresponding to
2960 * Which means we must set the snoop bit in PAT entry 0
2961 * in order to keep the global status page working.
2963 pat = GEN8_PPAT(0, CHV_PPAT_SNOOP) |
2967 GEN8_PPAT(4, CHV_PPAT_SNOOP) |
2968 GEN8_PPAT(5, CHV_PPAT_SNOOP) |
2969 GEN8_PPAT(6, CHV_PPAT_SNOOP) |
2970 GEN8_PPAT(7, CHV_PPAT_SNOOP);
2972 I915_WRITE(GEN8_PRIVATE_PAT_LO, pat);
2973 I915_WRITE(GEN8_PRIVATE_PAT_HI, pat >> 32);
2976 static int gen8_gmch_probe(struct drm_device *dev,
2979 phys_addr_t *mappable_base,
2982 struct drm_i915_private *dev_priv = dev->dev_private;
2987 /* TODO: We're not aware of mappable constraints on gen8 yet */
2988 *mappable_base = pci_resource_start(dev->pdev, 2);
2989 *mappable_end = pci_resource_len(dev->pdev, 2);
2991 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(39)))
2992 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(39));
2994 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
2996 if (INTEL_INFO(dev)->gen >= 9) {
2997 *stolen = gen9_get_stolen_size(snb_gmch_ctl);
2998 gtt_size = gen8_get_total_gtt_size(snb_gmch_ctl);
2999 } else if (IS_CHERRYVIEW(dev)) {
3000 *stolen = chv_get_stolen_size(snb_gmch_ctl);
3001 gtt_size = chv_get_total_gtt_size(snb_gmch_ctl);
3003 *stolen = gen8_get_stolen_size(snb_gmch_ctl);
3004 gtt_size = gen8_get_total_gtt_size(snb_gmch_ctl);
3007 *gtt_total = (gtt_size / sizeof(gen8_pte_t)) << PAGE_SHIFT;
3009 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
3010 chv_setup_private_ppat(dev_priv);
3012 bdw_setup_private_ppat(dev_priv);
3014 ret = ggtt_probe_common(dev, gtt_size);
3016 dev_priv->gtt.base.clear_range = gen8_ggtt_clear_range;
3017 dev_priv->gtt.base.insert_entries = gen8_ggtt_insert_entries;
3018 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
3019 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
3021 if (IS_CHERRYVIEW(dev))
3022 dev_priv->gtt.base.bind_vma = ggtt_bind_vma__BKL;
3027 static int gen6_gmch_probe(struct drm_device *dev,
3030 phys_addr_t *mappable_base,
3033 struct drm_i915_private *dev_priv = dev->dev_private;
3034 unsigned int gtt_size;
3038 *mappable_base = pci_resource_start(dev->pdev, 2);
3039 *mappable_end = pci_resource_len(dev->pdev, 2);
3041 /* 64/512MB is the current min/max we actually know of, but this is just
3042 * a coarse sanity check.
3044 if ((*mappable_end < (64<<20) || (*mappable_end > (512<<20)))) {
3045 DRM_ERROR("Unknown GMADR size (%llx)\n",
3046 dev_priv->gtt.mappable_end);
3050 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(40)))
3051 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(40));
3052 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
3054 *stolen = gen6_get_stolen_size(snb_gmch_ctl);
3056 gtt_size = gen6_get_total_gtt_size(snb_gmch_ctl);
3057 *gtt_total = (gtt_size / sizeof(gen6_pte_t)) << PAGE_SHIFT;
3059 ret = ggtt_probe_common(dev, gtt_size);
3061 dev_priv->gtt.base.clear_range = gen6_ggtt_clear_range;
3062 dev_priv->gtt.base.insert_entries = gen6_ggtt_insert_entries;
3063 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
3064 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
3069 static void gen6_gmch_remove(struct i915_address_space *vm)
3072 struct i915_gtt *gtt = container_of(vm, struct i915_gtt, base);
3075 free_scratch_page(vm->dev, vm->scratch_page);
3078 static int i915_gmch_probe(struct drm_device *dev,
3081 phys_addr_t *mappable_base,
3084 struct drm_i915_private *dev_priv = dev->dev_private;
3087 ret = intel_gmch_probe(dev_priv->bridge_dev, dev_priv->dev->pdev, NULL);
3089 DRM_ERROR("failed to set up gmch\n");
3093 intel_gtt_get(gtt_total, stolen, mappable_base, mappable_end);
3095 dev_priv->gtt.do_idle_maps = needs_idle_maps(dev_priv->dev);
3096 dev_priv->gtt.base.insert_entries = i915_ggtt_insert_entries;
3097 dev_priv->gtt.base.clear_range = i915_ggtt_clear_range;
3098 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
3099 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
3101 if (unlikely(dev_priv->gtt.do_idle_maps))
3102 DRM_INFO("applying Ironlake quirks for intel_iommu\n");
3107 static void i915_gmch_remove(struct i915_address_space *vm)
3109 intel_gmch_remove();
3112 int i915_gem_gtt_init(struct drm_device *dev)
3114 struct drm_i915_private *dev_priv = dev->dev_private;
3115 struct i915_gtt *gtt = &dev_priv->gtt;
3118 if (INTEL_INFO(dev)->gen <= 5) {
3119 gtt->gtt_probe = i915_gmch_probe;
3120 gtt->base.cleanup = i915_gmch_remove;
3121 } else if (INTEL_INFO(dev)->gen < 8) {
3122 gtt->gtt_probe = gen6_gmch_probe;
3123 gtt->base.cleanup = gen6_gmch_remove;
3124 if (IS_HASWELL(dev) && dev_priv->ellc_size)
3125 gtt->base.pte_encode = iris_pte_encode;
3126 else if (IS_HASWELL(dev))
3127 gtt->base.pte_encode = hsw_pte_encode;
3128 else if (IS_VALLEYVIEW(dev))
3129 gtt->base.pte_encode = byt_pte_encode;
3130 else if (INTEL_INFO(dev)->gen >= 7)
3131 gtt->base.pte_encode = ivb_pte_encode;
3133 gtt->base.pte_encode = snb_pte_encode;
3135 dev_priv->gtt.gtt_probe = gen8_gmch_probe;
3136 dev_priv->gtt.base.cleanup = gen6_gmch_remove;
3139 gtt->base.dev = dev;
3141 ret = gtt->gtt_probe(dev, >t->base.total, >t->stolen_size,
3142 >t->mappable_base, >t->mappable_end);
3146 /* GMADR is the PCI mmio aperture into the global GTT. */
3147 DRM_INFO("Memory usable by graphics device = %lluM\n",
3148 gtt->base.total >> 20);
3149 DRM_DEBUG_DRIVER("GMADR size = %lldM\n", gtt->mappable_end >> 20);
3150 DRM_DEBUG_DRIVER("GTT stolen size = %zdM\n", gtt->stolen_size >> 20);
3151 #ifdef CONFIG_INTEL_IOMMU
3152 if (intel_iommu_gfx_mapped)
3153 DRM_INFO("VT-d active for gfx access\n");
3156 * i915.enable_ppgtt is read-only, so do an early pass to validate the
3157 * user's requested state against the hardware/driver capabilities. We
3158 * do this now so that we can print out any log messages once rather
3159 * than every time we check intel_enable_ppgtt().
3161 i915.enable_ppgtt = sanitize_enable_ppgtt(dev, i915.enable_ppgtt);
3162 DRM_DEBUG_DRIVER("ppgtt mode: %i\n", i915.enable_ppgtt);
3167 void i915_gem_restore_gtt_mappings(struct drm_device *dev)
3169 struct drm_i915_private *dev_priv = dev->dev_private;
3170 struct drm_i915_gem_object *obj;
3171 struct i915_address_space *vm;
3172 struct i915_vma *vma;
3175 i915_check_and_clear_faults(dev);
3177 /* First fill our portion of the GTT with scratch pages */
3178 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
3179 dev_priv->gtt.base.start,
3180 dev_priv->gtt.base.total,
3183 /* Cache flush objects bound into GGTT and rebind them. */
3184 vm = &dev_priv->gtt.base;
3185 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
3187 list_for_each_entry(vma, &obj->vma_list, vma_link) {
3191 WARN_ON(i915_vma_bind(vma, obj->cache_level,
3198 i915_gem_clflush_object(obj, obj->pin_display);
3201 if (INTEL_INFO(dev)->gen >= 8) {
3202 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
3203 chv_setup_private_ppat(dev_priv);
3205 bdw_setup_private_ppat(dev_priv);
3210 if (USES_PPGTT(dev)) {
3211 list_for_each_entry(vm, &dev_priv->vm_list, global_link) {
3212 /* TODO: Perhaps it shouldn't be gen6 specific */
3214 struct i915_hw_ppgtt *ppgtt =
3215 container_of(vm, struct i915_hw_ppgtt,
3218 if (i915_is_ggtt(vm))
3219 ppgtt = dev_priv->mm.aliasing_ppgtt;
3221 gen6_write_page_range(dev_priv, &ppgtt->pd,
3222 0, ppgtt->base.total);
3226 i915_ggtt_flush(dev_priv);
3229 static struct i915_vma *
3230 __i915_gem_vma_create(struct drm_i915_gem_object *obj,
3231 struct i915_address_space *vm,
3232 const struct i915_ggtt_view *ggtt_view)
3234 struct i915_vma *vma;
3236 if (WARN_ON(i915_is_ggtt(vm) != !!ggtt_view))
3237 return ERR_PTR(-EINVAL);
3239 vma = kmem_cache_zalloc(to_i915(obj->base.dev)->vmas, GFP_KERNEL);
3241 return ERR_PTR(-ENOMEM);
3243 INIT_LIST_HEAD(&vma->vma_link);
3244 INIT_LIST_HEAD(&vma->mm_list);
3245 INIT_LIST_HEAD(&vma->exec_list);
3249 if (i915_is_ggtt(vm))
3250 vma->ggtt_view = *ggtt_view;
3252 list_add_tail(&vma->vma_link, &obj->vma_list);
3253 if (!i915_is_ggtt(vm))
3254 i915_ppgtt_get(i915_vm_to_ppgtt(vm));
3260 i915_gem_obj_lookup_or_create_vma(struct drm_i915_gem_object *obj,
3261 struct i915_address_space *vm)
3263 struct i915_vma *vma;
3265 vma = i915_gem_obj_to_vma(obj, vm);
3267 vma = __i915_gem_vma_create(obj, vm,
3268 i915_is_ggtt(vm) ? &i915_ggtt_view_normal : NULL);
3274 i915_gem_obj_lookup_or_create_ggtt_vma(struct drm_i915_gem_object *obj,
3275 const struct i915_ggtt_view *view)
3277 struct i915_address_space *ggtt = i915_obj_to_ggtt(obj);
3278 struct i915_vma *vma;
3281 return ERR_PTR(-EINVAL);
3283 vma = i915_gem_obj_to_ggtt_view(obj, view);
3289 vma = __i915_gem_vma_create(obj, ggtt, view);
3295 static struct scatterlist *
3296 rotate_pages(dma_addr_t *in, unsigned int offset,
3297 unsigned int width, unsigned int height,
3298 struct sg_table *st, struct scatterlist *sg)
3300 unsigned int column, row;
3301 unsigned int src_idx;
3308 for (column = 0; column < width; column++) {
3309 src_idx = width * (height - 1) + column;
3310 for (row = 0; row < height; row++) {
3312 /* We don't need the pages, but need to initialize
3313 * the entries so the sg list can be happily traversed.
3314 * The only thing we need are DMA addresses.
3316 sg_set_page(sg, NULL, PAGE_SIZE, 0);
3317 sg_dma_address(sg) = in[offset + src_idx];
3318 sg_dma_len(sg) = PAGE_SIZE;
3327 static struct sg_table *
3328 intel_rotate_fb_obj_pages(struct i915_ggtt_view *ggtt_view,
3329 struct drm_i915_gem_object *obj)
3331 struct intel_rotation_info *rot_info = &ggtt_view->params.rotation_info;
3332 unsigned int size_pages = rot_info->size >> PAGE_SHIFT;
3333 unsigned int size_pages_uv;
3334 struct sg_page_iter sg_iter;
3336 dma_addr_t *page_addr_list;
3337 struct sg_table *st;
3338 unsigned int uv_start_page;
3339 struct scatterlist *sg;
3342 /* Allocate a temporary list of source pages for random access. */
3343 page_addr_list = drm_malloc_ab(obj->base.size / PAGE_SIZE,
3344 sizeof(dma_addr_t));
3345 if (!page_addr_list)
3346 return ERR_PTR(ret);
3348 /* Account for UV plane with NV12. */
3349 if (rot_info->pixel_format == DRM_FORMAT_NV12)
3350 size_pages_uv = rot_info->size_uv >> PAGE_SHIFT;
3354 /* Allocate target SG list. */
3355 st = kmalloc(sizeof(*st), GFP_KERNEL);
3359 ret = sg_alloc_table(st, size_pages + size_pages_uv, GFP_KERNEL);
3363 /* Populate source page list from the object. */
3365 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) {
3366 page_addr_list[i] = sg_page_iter_dma_address(&sg_iter);
3370 /* Rotate the pages. */
3371 sg = rotate_pages(page_addr_list, 0,
3372 rot_info->width_pages, rot_info->height_pages,
3375 /* Append the UV plane if NV12. */
3376 if (rot_info->pixel_format == DRM_FORMAT_NV12) {
3377 uv_start_page = size_pages;
3379 /* Check for tile-row un-alignment. */
3380 if (offset_in_page(rot_info->uv_offset))
3383 rot_info->uv_start_page = uv_start_page;
3385 rotate_pages(page_addr_list, uv_start_page,
3386 rot_info->width_pages_uv,
3387 rot_info->height_pages_uv,
3392 "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",
3393 obj->base.size, rot_info->pitch, rot_info->height,
3394 rot_info->pixel_format, rot_info->width_pages,
3395 rot_info->height_pages, size_pages + size_pages_uv,
3398 drm_free_large(page_addr_list);
3405 drm_free_large(page_addr_list);
3408 "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",
3409 obj->base.size, ret, rot_info->pitch, rot_info->height,
3410 rot_info->pixel_format, rot_info->width_pages,
3411 rot_info->height_pages, size_pages + size_pages_uv,
3413 return ERR_PTR(ret);
3416 static struct sg_table *
3417 intel_partial_pages(const struct i915_ggtt_view *view,
3418 struct drm_i915_gem_object *obj)
3420 struct sg_table *st;
3421 struct scatterlist *sg;
3422 struct sg_page_iter obj_sg_iter;
3425 st = kmalloc(sizeof(*st), GFP_KERNEL);
3429 ret = sg_alloc_table(st, view->params.partial.size, GFP_KERNEL);
3435 for_each_sg_page(obj->pages->sgl, &obj_sg_iter, obj->pages->nents,
3436 view->params.partial.offset)
3438 if (st->nents >= view->params.partial.size)
3441 sg_set_page(sg, NULL, PAGE_SIZE, 0);
3442 sg_dma_address(sg) = sg_page_iter_dma_address(&obj_sg_iter);
3443 sg_dma_len(sg) = PAGE_SIZE;
3454 return ERR_PTR(ret);
3458 i915_get_ggtt_vma_pages(struct i915_vma *vma)
3462 if (vma->ggtt_view.pages)
3465 if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL)
3466 vma->ggtt_view.pages = vma->obj->pages;
3467 else if (vma->ggtt_view.type == I915_GGTT_VIEW_ROTATED)
3468 vma->ggtt_view.pages =
3469 intel_rotate_fb_obj_pages(&vma->ggtt_view, vma->obj);
3470 else if (vma->ggtt_view.type == I915_GGTT_VIEW_PARTIAL)
3471 vma->ggtt_view.pages =
3472 intel_partial_pages(&vma->ggtt_view, vma->obj);
3474 WARN_ONCE(1, "GGTT view %u not implemented!\n",
3475 vma->ggtt_view.type);
3477 if (!vma->ggtt_view.pages) {
3478 DRM_ERROR("Failed to get pages for GGTT view type %u!\n",
3479 vma->ggtt_view.type);
3481 } else if (IS_ERR(vma->ggtt_view.pages)) {
3482 ret = PTR_ERR(vma->ggtt_view.pages);
3483 vma->ggtt_view.pages = NULL;
3484 DRM_ERROR("Failed to get pages for VMA view type %u (%d)!\n",
3485 vma->ggtt_view.type, ret);
3492 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
3494 * @cache_level: mapping cache level
3495 * @flags: flags like global or local mapping
3497 * DMA addresses are taken from the scatter-gather table of this object (or of
3498 * this VMA in case of non-default GGTT views) and PTE entries set up.
3499 * Note that DMA addresses are also the only part of the SG table we care about.
3501 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
3507 if (WARN_ON(flags == 0))
3511 if (flags & PIN_GLOBAL)
3512 bind_flags |= GLOBAL_BIND;
3513 if (flags & PIN_USER)
3514 bind_flags |= LOCAL_BIND;
3516 if (flags & PIN_UPDATE)
3517 bind_flags |= vma->bound;
3519 bind_flags &= ~vma->bound;
3521 if (bind_flags == 0)
3524 if (vma->bound == 0 && vma->vm->allocate_va_range) {
3525 trace_i915_va_alloc(vma->vm,
3528 VM_TO_TRACE_NAME(vma->vm));
3530 /* XXX: i915_vma_pin() will fix this +- hack */
3532 ret = vma->vm->allocate_va_range(vma->vm,
3540 ret = vma->vm->bind_vma(vma, cache_level, bind_flags);
3544 vma->bound |= bind_flags;
3550 * i915_ggtt_view_size - Get the size of a GGTT view.
3551 * @obj: Object the view is of.
3552 * @view: The view in question.
3554 * @return The size of the GGTT view in bytes.
3557 i915_ggtt_view_size(struct drm_i915_gem_object *obj,
3558 const struct i915_ggtt_view *view)
3560 if (view->type == I915_GGTT_VIEW_NORMAL) {
3561 return obj->base.size;
3562 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
3563 return view->params.rotation_info.size;
3564 } else if (view->type == I915_GGTT_VIEW_PARTIAL) {
3565 return view->params.partial.size << PAGE_SHIFT;
3567 WARN_ONCE(1, "GGTT view %u not implemented!\n", view->type);
3568 return obj->base.size;