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,
774 struct i915_page_directory_pointer *pdp;
776 gen8_for_each_pml4e(pdp, &ppgtt->pml4, start, length, 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;
843 uint64_t length = (uint64_t)pages->orig_nents << PAGE_SHIFT;
845 gen8_for_each_pml4e(pdp, &ppgtt->pml4, start, length, 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;
1025 gen8_for_each_pde(pt, pd, start, length, pde) {
1026 /* Don't reallocate page tables */
1027 if (test_bit(pde, pd->used_pdes)) {
1028 /* Scratch is never allocated this way */
1029 WARN_ON(pt == vm->scratch_pt);
1037 gen8_initialize_pt(vm, pt);
1038 pd->page_table[pde] = pt;
1039 __set_bit(pde, new_pts);
1040 trace_i915_page_table_entry_alloc(vm, pde, start, GEN8_PDE_SHIFT);
1046 for_each_set_bit(pde, new_pts, I915_PDES)
1047 free_pt(dev, pd->page_table[pde]);
1053 * gen8_ppgtt_alloc_page_directories() - Allocate page directories for VA range.
1054 * @vm: Master vm structure.
1055 * @pdp: Page directory pointer for this address range.
1056 * @start: Starting virtual address to begin allocations.
1057 * @length: Size of the allocations.
1058 * @new_pds: Bitmap set by function with new allocations. Likely used by the
1059 * caller to free on error.
1061 * Allocate the required number of page directories starting at the pde index of
1062 * @start, and ending at the pde index @start + @length. This function will skip
1063 * over already allocated page directories within the range, and only allocate
1064 * new ones, setting the appropriate pointer within the pdp as well as the
1065 * correct position in the bitmap @new_pds.
1067 * The function will only allocate the pages within the range for a give page
1068 * directory pointer. In other words, if @start + @length straddles a virtually
1069 * addressed PDP boundary (512GB for 4k pages), there will be more allocations
1070 * required by the caller, This is not currently possible, and the BUG in the
1071 * code will prevent it.
1073 * Return: 0 if success; negative error code otherwise.
1076 gen8_ppgtt_alloc_page_directories(struct i915_address_space *vm,
1077 struct i915_page_directory_pointer *pdp,
1080 unsigned long *new_pds)
1082 struct drm_device *dev = vm->dev;
1083 struct i915_page_directory *pd;
1085 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1087 WARN_ON(!bitmap_empty(new_pds, pdpes));
1089 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1090 if (test_bit(pdpe, pdp->used_pdpes))
1097 gen8_initialize_pd(vm, pd);
1098 pdp->page_directory[pdpe] = pd;
1099 __set_bit(pdpe, new_pds);
1100 trace_i915_page_directory_entry_alloc(vm, pdpe, start, GEN8_PDPE_SHIFT);
1106 for_each_set_bit(pdpe, new_pds, pdpes)
1107 free_pd(dev, pdp->page_directory[pdpe]);
1113 * gen8_ppgtt_alloc_page_dirpointers() - Allocate pdps for VA range.
1114 * @vm: Master vm structure.
1115 * @pml4: Page map level 4 for this address range.
1116 * @start: Starting virtual address to begin allocations.
1117 * @length: Size of the allocations.
1118 * @new_pdps: Bitmap set by function with new allocations. Likely used by the
1119 * caller to free on error.
1121 * Allocate the required number of page directory pointers. Extremely similar to
1122 * gen8_ppgtt_alloc_page_directories() and gen8_ppgtt_alloc_pagetabs().
1123 * The main difference is here we are limited by the pml4 boundary (instead of
1124 * the page directory pointer).
1126 * Return: 0 if success; negative error code otherwise.
1129 gen8_ppgtt_alloc_page_dirpointers(struct i915_address_space *vm,
1130 struct i915_pml4 *pml4,
1133 unsigned long *new_pdps)
1135 struct drm_device *dev = vm->dev;
1136 struct i915_page_directory_pointer *pdp;
1139 WARN_ON(!bitmap_empty(new_pdps, GEN8_PML4ES_PER_PML4));
1141 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1142 if (!test_bit(pml4e, pml4->used_pml4es)) {
1143 pdp = alloc_pdp(dev);
1147 gen8_initialize_pdp(vm, pdp);
1148 pml4->pdps[pml4e] = pdp;
1149 __set_bit(pml4e, new_pdps);
1150 trace_i915_page_directory_pointer_entry_alloc(vm,
1160 for_each_set_bit(pml4e, new_pdps, GEN8_PML4ES_PER_PML4)
1161 free_pdp(dev, pml4->pdps[pml4e]);
1167 free_gen8_temp_bitmaps(unsigned long *new_pds, unsigned long *new_pts)
1173 /* Fills in the page directory bitmap, and the array of page tables bitmap. Both
1174 * of these are based on the number of PDPEs in the system.
1177 int __must_check alloc_gen8_temp_bitmaps(unsigned long **new_pds,
1178 unsigned long **new_pts,
1184 pds = kcalloc(BITS_TO_LONGS(pdpes), sizeof(unsigned long), GFP_TEMPORARY);
1188 pts = kcalloc(pdpes, BITS_TO_LONGS(I915_PDES) * sizeof(unsigned long),
1199 free_gen8_temp_bitmaps(pds, pts);
1203 /* PDE TLBs are a pain to invalidate on GEN8+. When we modify
1204 * the page table structures, we mark them dirty so that
1205 * context switching/execlist queuing code takes extra steps
1206 * to ensure that tlbs are flushed.
1208 static void mark_tlbs_dirty(struct i915_hw_ppgtt *ppgtt)
1210 ppgtt->pd_dirty_rings = INTEL_INFO(ppgtt->base.dev)->ring_mask;
1213 static int gen8_alloc_va_range_3lvl(struct i915_address_space *vm,
1214 struct i915_page_directory_pointer *pdp,
1218 struct i915_hw_ppgtt *ppgtt =
1219 container_of(vm, struct i915_hw_ppgtt, base);
1220 unsigned long *new_page_dirs, *new_page_tables;
1221 struct drm_device *dev = vm->dev;
1222 struct i915_page_directory *pd;
1223 const uint64_t orig_start = start;
1224 const uint64_t orig_length = length;
1226 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1229 /* Wrap is never okay since we can only represent 48b, and we don't
1230 * actually use the other side of the canonical address space.
1232 if (WARN_ON(start + length < start))
1235 if (WARN_ON(start + length > vm->total))
1238 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables, pdpes);
1242 /* Do the allocations first so we can easily bail out */
1243 ret = gen8_ppgtt_alloc_page_directories(vm, pdp, start, length,
1246 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1250 /* For every page directory referenced, allocate page tables */
1251 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1252 ret = gen8_ppgtt_alloc_pagetabs(vm, pd, start, length,
1253 new_page_tables + pdpe * BITS_TO_LONGS(I915_PDES));
1259 length = orig_length;
1261 /* Allocations have completed successfully, so set the bitmaps, and do
1263 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1264 gen8_pde_t *const page_directory = kmap_px(pd);
1265 struct i915_page_table *pt;
1266 uint64_t pd_len = length;
1267 uint64_t pd_start = start;
1270 /* Every pd should be allocated, we just did that above. */
1273 gen8_for_each_pde(pt, pd, pd_start, pd_len, pde) {
1274 /* Same reasoning as pd */
1277 WARN_ON(!gen8_pte_count(pd_start, pd_len));
1279 /* Set our used ptes within the page table */
1280 bitmap_set(pt->used_ptes,
1281 gen8_pte_index(pd_start),
1282 gen8_pte_count(pd_start, pd_len));
1284 /* Our pde is now pointing to the pagetable, pt */
1285 __set_bit(pde, pd->used_pdes);
1287 /* Map the PDE to the page table */
1288 page_directory[pde] = gen8_pde_encode(px_dma(pt),
1290 trace_i915_page_table_entry_map(&ppgtt->base, pde, pt,
1291 gen8_pte_index(start),
1292 gen8_pte_count(start, length),
1295 /* NB: We haven't yet mapped ptes to pages. At this
1296 * point we're still relying on insert_entries() */
1299 kunmap_px(ppgtt, page_directory);
1300 __set_bit(pdpe, pdp->used_pdpes);
1301 gen8_setup_page_directory(ppgtt, pdp, pd, pdpe);
1304 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1305 mark_tlbs_dirty(ppgtt);
1312 for_each_set_bit(temp, new_page_tables + pdpe *
1313 BITS_TO_LONGS(I915_PDES), I915_PDES)
1314 free_pt(dev, pdp->page_directory[pdpe]->page_table[temp]);
1317 for_each_set_bit(pdpe, new_page_dirs, pdpes)
1318 free_pd(dev, pdp->page_directory[pdpe]);
1320 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1321 mark_tlbs_dirty(ppgtt);
1325 static int gen8_alloc_va_range_4lvl(struct i915_address_space *vm,
1326 struct i915_pml4 *pml4,
1330 DECLARE_BITMAP(new_pdps, GEN8_PML4ES_PER_PML4);
1331 struct i915_hw_ppgtt *ppgtt =
1332 container_of(vm, struct i915_hw_ppgtt, base);
1333 struct i915_page_directory_pointer *pdp;
1337 /* Do the pml4 allocations first, so we don't need to track the newly
1338 * allocated tables below the pdp */
1339 bitmap_zero(new_pdps, GEN8_PML4ES_PER_PML4);
1341 /* The pagedirectory and pagetable allocations are done in the shared 3
1342 * and 4 level code. Just allocate the pdps.
1344 ret = gen8_ppgtt_alloc_page_dirpointers(vm, pml4, start, length,
1349 WARN(bitmap_weight(new_pdps, GEN8_PML4ES_PER_PML4) > 2,
1350 "The allocation has spanned more than 512GB. "
1351 "It is highly likely this is incorrect.");
1353 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1356 ret = gen8_alloc_va_range_3lvl(vm, pdp, start, length);
1360 gen8_setup_page_directory_pointer(ppgtt, pml4, pdp, pml4e);
1363 bitmap_or(pml4->used_pml4es, new_pdps, pml4->used_pml4es,
1364 GEN8_PML4ES_PER_PML4);
1369 for_each_set_bit(pml4e, new_pdps, GEN8_PML4ES_PER_PML4)
1370 gen8_ppgtt_cleanup_3lvl(vm->dev, pml4->pdps[pml4e]);
1375 static int gen8_alloc_va_range(struct i915_address_space *vm,
1376 uint64_t start, uint64_t length)
1378 struct i915_hw_ppgtt *ppgtt =
1379 container_of(vm, struct i915_hw_ppgtt, base);
1381 if (USES_FULL_48BIT_PPGTT(vm->dev))
1382 return gen8_alloc_va_range_4lvl(vm, &ppgtt->pml4, start, length);
1384 return gen8_alloc_va_range_3lvl(vm, &ppgtt->pdp, start, length);
1387 static void gen8_dump_pdp(struct i915_page_directory_pointer *pdp,
1388 uint64_t start, uint64_t length,
1389 gen8_pte_t scratch_pte,
1392 struct i915_page_directory *pd;
1395 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1396 struct i915_page_table *pt;
1397 uint64_t pd_len = length;
1398 uint64_t pd_start = start;
1401 if (!test_bit(pdpe, pdp->used_pdpes))
1404 seq_printf(m, "\tPDPE #%d\n", pdpe);
1405 gen8_for_each_pde(pt, pd, pd_start, pd_len, pde) {
1407 gen8_pte_t *pt_vaddr;
1409 if (!test_bit(pde, pd->used_pdes))
1412 pt_vaddr = kmap_px(pt);
1413 for (pte = 0; pte < GEN8_PTES; pte += 4) {
1415 (pdpe << GEN8_PDPE_SHIFT) |
1416 (pde << GEN8_PDE_SHIFT) |
1417 (pte << GEN8_PTE_SHIFT);
1421 for (i = 0; i < 4; i++)
1422 if (pt_vaddr[pte + i] != scratch_pte)
1427 seq_printf(m, "\t\t0x%llx [%03d,%03d,%04d]: =", va, pdpe, pde, pte);
1428 for (i = 0; i < 4; i++) {
1429 if (pt_vaddr[pte + i] != scratch_pte)
1430 seq_printf(m, " %llx", pt_vaddr[pte + i]);
1432 seq_puts(m, " SCRATCH ");
1436 /* don't use kunmap_px, it could trigger
1437 * an unnecessary flush.
1439 kunmap_atomic(pt_vaddr);
1444 static void gen8_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
1446 struct i915_address_space *vm = &ppgtt->base;
1447 uint64_t start = ppgtt->base.start;
1448 uint64_t length = ppgtt->base.total;
1449 gen8_pte_t scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
1450 I915_CACHE_LLC, true);
1452 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
1453 gen8_dump_pdp(&ppgtt->pdp, start, length, scratch_pte, m);
1456 struct i915_pml4 *pml4 = &ppgtt->pml4;
1457 struct i915_page_directory_pointer *pdp;
1459 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1460 if (!test_bit(pml4e, pml4->used_pml4es))
1463 seq_printf(m, " PML4E #%llu\n", pml4e);
1464 gen8_dump_pdp(pdp, start, length, scratch_pte, m);
1469 static int gen8_preallocate_top_level_pdps(struct i915_hw_ppgtt *ppgtt)
1471 unsigned long *new_page_dirs, *new_page_tables;
1472 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1475 /* We allocate temp bitmap for page tables for no gain
1476 * but as this is for init only, lets keep the things simple
1478 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables, pdpes);
1482 /* Allocate for all pdps regardless of how the ppgtt
1485 ret = gen8_ppgtt_alloc_page_directories(&ppgtt->base, &ppgtt->pdp,
1489 *ppgtt->pdp.used_pdpes = *new_page_dirs;
1491 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1497 * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers
1498 * with a net effect resembling a 2-level page table in normal x86 terms. Each
1499 * PDP represents 1GB of memory 4 * 512 * 512 * 4096 = 4GB legacy 32b address
1503 static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
1507 ret = gen8_init_scratch(&ppgtt->base);
1511 ppgtt->base.start = 0;
1512 ppgtt->base.cleanup = gen8_ppgtt_cleanup;
1513 ppgtt->base.allocate_va_range = gen8_alloc_va_range;
1514 ppgtt->base.insert_entries = gen8_ppgtt_insert_entries;
1515 ppgtt->base.clear_range = gen8_ppgtt_clear_range;
1516 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
1517 ppgtt->base.bind_vma = ppgtt_bind_vma;
1518 ppgtt->debug_dump = gen8_dump_ppgtt;
1520 if (USES_FULL_48BIT_PPGTT(ppgtt->base.dev)) {
1521 ret = setup_px(ppgtt->base.dev, &ppgtt->pml4);
1525 gen8_initialize_pml4(&ppgtt->base, &ppgtt->pml4);
1527 ppgtt->base.total = 1ULL << 48;
1528 ppgtt->switch_mm = gen8_48b_mm_switch;
1530 ret = __pdp_init(ppgtt->base.dev, &ppgtt->pdp);
1534 ppgtt->base.total = 1ULL << 32;
1535 ppgtt->switch_mm = gen8_legacy_mm_switch;
1536 trace_i915_page_directory_pointer_entry_alloc(&ppgtt->base,
1540 if (intel_vgpu_active(ppgtt->base.dev)) {
1541 ret = gen8_preallocate_top_level_pdps(ppgtt);
1547 if (intel_vgpu_active(ppgtt->base.dev))
1548 gen8_ppgtt_notify_vgt(ppgtt, true);
1553 gen8_free_scratch(&ppgtt->base);
1557 static void gen6_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
1559 struct i915_address_space *vm = &ppgtt->base;
1560 struct i915_page_table *unused;
1561 gen6_pte_t scratch_pte;
1563 uint32_t pte, pde, temp;
1564 uint32_t start = ppgtt->base.start, length = ppgtt->base.total;
1566 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1567 I915_CACHE_LLC, true, 0);
1569 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde) {
1571 gen6_pte_t *pt_vaddr;
1572 const dma_addr_t pt_addr = px_dma(ppgtt->pd.page_table[pde]);
1573 pd_entry = readl(ppgtt->pd_addr + pde);
1574 expected = (GEN6_PDE_ADDR_ENCODE(pt_addr) | GEN6_PDE_VALID);
1576 if (pd_entry != expected)
1577 seq_printf(m, "\tPDE #%d mismatch: Actual PDE: %x Expected PDE: %x\n",
1581 seq_printf(m, "\tPDE: %x\n", pd_entry);
1583 pt_vaddr = kmap_px(ppgtt->pd.page_table[pde]);
1585 for (pte = 0; pte < GEN6_PTES; pte+=4) {
1587 (pde * PAGE_SIZE * GEN6_PTES) +
1591 for (i = 0; i < 4; i++)
1592 if (pt_vaddr[pte + i] != scratch_pte)
1597 seq_printf(m, "\t\t0x%lx [%03d,%04d]: =", va, pde, pte);
1598 for (i = 0; i < 4; i++) {
1599 if (pt_vaddr[pte + i] != scratch_pte)
1600 seq_printf(m, " %08x", pt_vaddr[pte + i]);
1602 seq_puts(m, " SCRATCH ");
1606 kunmap_px(ppgtt, pt_vaddr);
1610 /* Write pde (index) from the page directory @pd to the page table @pt */
1611 static void gen6_write_pde(struct i915_page_directory *pd,
1612 const int pde, struct i915_page_table *pt)
1614 /* Caller needs to make sure the write completes if necessary */
1615 struct i915_hw_ppgtt *ppgtt =
1616 container_of(pd, struct i915_hw_ppgtt, pd);
1619 pd_entry = GEN6_PDE_ADDR_ENCODE(px_dma(pt));
1620 pd_entry |= GEN6_PDE_VALID;
1622 writel(pd_entry, ppgtt->pd_addr + pde);
1625 /* Write all the page tables found in the ppgtt structure to incrementing page
1627 static void gen6_write_page_range(struct drm_i915_private *dev_priv,
1628 struct i915_page_directory *pd,
1629 uint32_t start, uint32_t length)
1631 struct i915_page_table *pt;
1634 gen6_for_each_pde(pt, pd, start, length, temp, pde)
1635 gen6_write_pde(pd, pde, pt);
1637 /* Make sure write is complete before other code can use this page
1638 * table. Also require for WC mapped PTEs */
1639 readl(dev_priv->gtt.gsm);
1642 static uint32_t get_pd_offset(struct i915_hw_ppgtt *ppgtt)
1644 BUG_ON(ppgtt->pd.base.ggtt_offset & 0x3f);
1646 return (ppgtt->pd.base.ggtt_offset / 64) << 16;
1649 static int hsw_mm_switch(struct i915_hw_ppgtt *ppgtt,
1650 struct drm_i915_gem_request *req)
1652 struct intel_engine_cs *ring = req->ring;
1655 /* NB: TLBs must be flushed and invalidated before a switch */
1656 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1660 ret = intel_ring_begin(req, 6);
1664 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1665 intel_ring_emit_reg(ring, RING_PP_DIR_DCLV(ring));
1666 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1667 intel_ring_emit_reg(ring, RING_PP_DIR_BASE(ring));
1668 intel_ring_emit(ring, get_pd_offset(ppgtt));
1669 intel_ring_emit(ring, MI_NOOP);
1670 intel_ring_advance(ring);
1675 static int vgpu_mm_switch(struct i915_hw_ppgtt *ppgtt,
1676 struct drm_i915_gem_request *req)
1678 struct intel_engine_cs *ring = req->ring;
1679 struct drm_i915_private *dev_priv = to_i915(ppgtt->base.dev);
1681 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
1682 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
1686 static int gen7_mm_switch(struct i915_hw_ppgtt *ppgtt,
1687 struct drm_i915_gem_request *req)
1689 struct intel_engine_cs *ring = req->ring;
1692 /* NB: TLBs must be flushed and invalidated before a switch */
1693 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1697 ret = intel_ring_begin(req, 6);
1701 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1702 intel_ring_emit_reg(ring, RING_PP_DIR_DCLV(ring));
1703 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1704 intel_ring_emit_reg(ring, RING_PP_DIR_BASE(ring));
1705 intel_ring_emit(ring, get_pd_offset(ppgtt));
1706 intel_ring_emit(ring, MI_NOOP);
1707 intel_ring_advance(ring);
1709 /* XXX: RCS is the only one to auto invalidate the TLBs? */
1710 if (ring->id != RCS) {
1711 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1719 static int gen6_mm_switch(struct i915_hw_ppgtt *ppgtt,
1720 struct drm_i915_gem_request *req)
1722 struct intel_engine_cs *ring = req->ring;
1723 struct drm_device *dev = ppgtt->base.dev;
1724 struct drm_i915_private *dev_priv = dev->dev_private;
1727 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
1728 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
1730 POSTING_READ(RING_PP_DIR_DCLV(ring));
1735 static void gen8_ppgtt_enable(struct drm_device *dev)
1737 struct drm_i915_private *dev_priv = dev->dev_private;
1738 struct intel_engine_cs *ring;
1741 for_each_ring(ring, dev_priv, j) {
1742 u32 four_level = USES_FULL_48BIT_PPGTT(dev) ? GEN8_GFX_PPGTT_48B : 0;
1743 I915_WRITE(RING_MODE_GEN7(ring),
1744 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE | four_level));
1748 static void gen7_ppgtt_enable(struct drm_device *dev)
1750 struct drm_i915_private *dev_priv = dev->dev_private;
1751 struct intel_engine_cs *ring;
1752 uint32_t ecochk, ecobits;
1755 ecobits = I915_READ(GAC_ECO_BITS);
1756 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B);
1758 ecochk = I915_READ(GAM_ECOCHK);
1759 if (IS_HASWELL(dev)) {
1760 ecochk |= ECOCHK_PPGTT_WB_HSW;
1762 ecochk |= ECOCHK_PPGTT_LLC_IVB;
1763 ecochk &= ~ECOCHK_PPGTT_GFDT_IVB;
1765 I915_WRITE(GAM_ECOCHK, ecochk);
1767 for_each_ring(ring, dev_priv, i) {
1768 /* GFX_MODE is per-ring on gen7+ */
1769 I915_WRITE(RING_MODE_GEN7(ring),
1770 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1774 static void gen6_ppgtt_enable(struct drm_device *dev)
1776 struct drm_i915_private *dev_priv = dev->dev_private;
1777 uint32_t ecochk, gab_ctl, ecobits;
1779 ecobits = I915_READ(GAC_ECO_BITS);
1780 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_SNB_BIT |
1781 ECOBITS_PPGTT_CACHE64B);
1783 gab_ctl = I915_READ(GAB_CTL);
1784 I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT);
1786 ecochk = I915_READ(GAM_ECOCHK);
1787 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT | ECOCHK_PPGTT_CACHE64B);
1789 I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1792 /* PPGTT support for Sandybdrige/Gen6 and later */
1793 static void gen6_ppgtt_clear_range(struct i915_address_space *vm,
1798 struct i915_hw_ppgtt *ppgtt =
1799 container_of(vm, struct i915_hw_ppgtt, base);
1800 gen6_pte_t *pt_vaddr, scratch_pte;
1801 unsigned first_entry = start >> PAGE_SHIFT;
1802 unsigned num_entries = length >> PAGE_SHIFT;
1803 unsigned act_pt = first_entry / GEN6_PTES;
1804 unsigned first_pte = first_entry % GEN6_PTES;
1805 unsigned last_pte, i;
1807 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1808 I915_CACHE_LLC, true, 0);
1810 while (num_entries) {
1811 last_pte = first_pte + num_entries;
1812 if (last_pte > GEN6_PTES)
1813 last_pte = GEN6_PTES;
1815 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1817 for (i = first_pte; i < last_pte; i++)
1818 pt_vaddr[i] = scratch_pte;
1820 kunmap_px(ppgtt, pt_vaddr);
1822 num_entries -= last_pte - first_pte;
1828 static void gen6_ppgtt_insert_entries(struct i915_address_space *vm,
1829 struct sg_table *pages,
1831 enum i915_cache_level cache_level, u32 flags)
1833 struct i915_hw_ppgtt *ppgtt =
1834 container_of(vm, struct i915_hw_ppgtt, base);
1835 gen6_pte_t *pt_vaddr;
1836 unsigned first_entry = start >> PAGE_SHIFT;
1837 unsigned act_pt = first_entry / GEN6_PTES;
1838 unsigned act_pte = first_entry % GEN6_PTES;
1839 struct sg_page_iter sg_iter;
1842 for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) {
1843 if (pt_vaddr == NULL)
1844 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1847 vm->pte_encode(sg_page_iter_dma_address(&sg_iter),
1848 cache_level, true, flags);
1850 if (++act_pte == GEN6_PTES) {
1851 kunmap_px(ppgtt, pt_vaddr);
1858 kunmap_px(ppgtt, pt_vaddr);
1861 static int gen6_alloc_va_range(struct i915_address_space *vm,
1862 uint64_t start_in, uint64_t length_in)
1864 DECLARE_BITMAP(new_page_tables, I915_PDES);
1865 struct drm_device *dev = vm->dev;
1866 struct drm_i915_private *dev_priv = dev->dev_private;
1867 struct i915_hw_ppgtt *ppgtt =
1868 container_of(vm, struct i915_hw_ppgtt, base);
1869 struct i915_page_table *pt;
1870 uint32_t start, length, start_save, length_save;
1874 if (WARN_ON(start_in + length_in > ppgtt->base.total))
1877 start = start_save = start_in;
1878 length = length_save = length_in;
1880 bitmap_zero(new_page_tables, I915_PDES);
1882 /* The allocation is done in two stages so that we can bail out with
1883 * minimal amount of pain. The first stage finds new page tables that
1884 * need allocation. The second stage marks use ptes within the page
1887 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1888 if (pt != vm->scratch_pt) {
1889 WARN_ON(bitmap_empty(pt->used_ptes, GEN6_PTES));
1893 /* We've already allocated a page table */
1894 WARN_ON(!bitmap_empty(pt->used_ptes, GEN6_PTES));
1902 gen6_initialize_pt(vm, pt);
1904 ppgtt->pd.page_table[pde] = pt;
1905 __set_bit(pde, new_page_tables);
1906 trace_i915_page_table_entry_alloc(vm, pde, start, GEN6_PDE_SHIFT);
1910 length = length_save;
1912 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1913 DECLARE_BITMAP(tmp_bitmap, GEN6_PTES);
1915 bitmap_zero(tmp_bitmap, GEN6_PTES);
1916 bitmap_set(tmp_bitmap, gen6_pte_index(start),
1917 gen6_pte_count(start, length));
1919 if (__test_and_clear_bit(pde, new_page_tables))
1920 gen6_write_pde(&ppgtt->pd, pde, pt);
1922 trace_i915_page_table_entry_map(vm, pde, pt,
1923 gen6_pte_index(start),
1924 gen6_pte_count(start, length),
1926 bitmap_or(pt->used_ptes, tmp_bitmap, pt->used_ptes,
1930 WARN_ON(!bitmap_empty(new_page_tables, I915_PDES));
1932 /* Make sure write is complete before other code can use this page
1933 * table. Also require for WC mapped PTEs */
1934 readl(dev_priv->gtt.gsm);
1936 mark_tlbs_dirty(ppgtt);
1940 for_each_set_bit(pde, new_page_tables, I915_PDES) {
1941 struct i915_page_table *pt = ppgtt->pd.page_table[pde];
1943 ppgtt->pd.page_table[pde] = vm->scratch_pt;
1944 free_pt(vm->dev, pt);
1947 mark_tlbs_dirty(ppgtt);
1951 static int gen6_init_scratch(struct i915_address_space *vm)
1953 struct drm_device *dev = vm->dev;
1955 vm->scratch_page = alloc_scratch_page(dev);
1956 if (IS_ERR(vm->scratch_page))
1957 return PTR_ERR(vm->scratch_page);
1959 vm->scratch_pt = alloc_pt(dev);
1960 if (IS_ERR(vm->scratch_pt)) {
1961 free_scratch_page(dev, vm->scratch_page);
1962 return PTR_ERR(vm->scratch_pt);
1965 gen6_initialize_pt(vm, vm->scratch_pt);
1970 static void gen6_free_scratch(struct i915_address_space *vm)
1972 struct drm_device *dev = vm->dev;
1974 free_pt(dev, vm->scratch_pt);
1975 free_scratch_page(dev, vm->scratch_page);
1978 static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
1980 struct i915_hw_ppgtt *ppgtt =
1981 container_of(vm, struct i915_hw_ppgtt, base);
1982 struct i915_page_table *pt;
1985 drm_mm_remove_node(&ppgtt->node);
1987 gen6_for_all_pdes(pt, ppgtt, pde) {
1988 if (pt != vm->scratch_pt)
1989 free_pt(ppgtt->base.dev, pt);
1992 gen6_free_scratch(vm);
1995 static int gen6_ppgtt_allocate_page_directories(struct i915_hw_ppgtt *ppgtt)
1997 struct i915_address_space *vm = &ppgtt->base;
1998 struct drm_device *dev = ppgtt->base.dev;
1999 struct drm_i915_private *dev_priv = dev->dev_private;
2000 bool retried = false;
2003 /* PPGTT PDEs reside in the GGTT and consists of 512 entries. The
2004 * allocator works in address space sizes, so it's multiplied by page
2005 * size. We allocate at the top of the GTT to avoid fragmentation.
2007 BUG_ON(!drm_mm_initialized(&dev_priv->gtt.base.mm));
2009 ret = gen6_init_scratch(vm);
2014 ret = drm_mm_insert_node_in_range_generic(&dev_priv->gtt.base.mm,
2015 &ppgtt->node, GEN6_PD_SIZE,
2017 0, dev_priv->gtt.base.total,
2019 if (ret == -ENOSPC && !retried) {
2020 ret = i915_gem_evict_something(dev, &dev_priv->gtt.base,
2021 GEN6_PD_SIZE, GEN6_PD_ALIGN,
2023 0, dev_priv->gtt.base.total,
2036 if (ppgtt->node.start < dev_priv->gtt.mappable_end)
2037 DRM_DEBUG("Forced to use aperture for PDEs\n");
2042 gen6_free_scratch(vm);
2046 static int gen6_ppgtt_alloc(struct i915_hw_ppgtt *ppgtt)
2048 return gen6_ppgtt_allocate_page_directories(ppgtt);
2051 static void gen6_scratch_va_range(struct i915_hw_ppgtt *ppgtt,
2052 uint64_t start, uint64_t length)
2054 struct i915_page_table *unused;
2057 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde)
2058 ppgtt->pd.page_table[pde] = ppgtt->base.scratch_pt;
2061 static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
2063 struct drm_device *dev = ppgtt->base.dev;
2064 struct drm_i915_private *dev_priv = dev->dev_private;
2067 ppgtt->base.pte_encode = dev_priv->gtt.base.pte_encode;
2069 ppgtt->switch_mm = gen6_mm_switch;
2070 } else if (IS_HASWELL(dev)) {
2071 ppgtt->switch_mm = hsw_mm_switch;
2072 } else if (IS_GEN7(dev)) {
2073 ppgtt->switch_mm = gen7_mm_switch;
2077 if (intel_vgpu_active(dev))
2078 ppgtt->switch_mm = vgpu_mm_switch;
2080 ret = gen6_ppgtt_alloc(ppgtt);
2084 ppgtt->base.allocate_va_range = gen6_alloc_va_range;
2085 ppgtt->base.clear_range = gen6_ppgtt_clear_range;
2086 ppgtt->base.insert_entries = gen6_ppgtt_insert_entries;
2087 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
2088 ppgtt->base.bind_vma = ppgtt_bind_vma;
2089 ppgtt->base.cleanup = gen6_ppgtt_cleanup;
2090 ppgtt->base.start = 0;
2091 ppgtt->base.total = I915_PDES * GEN6_PTES * PAGE_SIZE;
2092 ppgtt->debug_dump = gen6_dump_ppgtt;
2094 ppgtt->pd.base.ggtt_offset =
2095 ppgtt->node.start / PAGE_SIZE * sizeof(gen6_pte_t);
2097 ppgtt->pd_addr = (gen6_pte_t __iomem *)dev_priv->gtt.gsm +
2098 ppgtt->pd.base.ggtt_offset / sizeof(gen6_pte_t);
2100 gen6_scratch_va_range(ppgtt, 0, ppgtt->base.total);
2102 gen6_write_page_range(dev_priv, &ppgtt->pd, 0, ppgtt->base.total);
2104 DRM_DEBUG_DRIVER("Allocated pde space (%lldM) at GTT entry: %llx\n",
2105 ppgtt->node.size >> 20,
2106 ppgtt->node.start / PAGE_SIZE);
2108 DRM_DEBUG("Adding PPGTT at offset %x\n",
2109 ppgtt->pd.base.ggtt_offset << 10);
2114 static int __hw_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
2116 ppgtt->base.dev = dev;
2118 if (INTEL_INFO(dev)->gen < 8)
2119 return gen6_ppgtt_init(ppgtt);
2121 return gen8_ppgtt_init(ppgtt);
2124 static void i915_address_space_init(struct i915_address_space *vm,
2125 struct drm_i915_private *dev_priv)
2127 drm_mm_init(&vm->mm, vm->start, vm->total);
2128 vm->dev = dev_priv->dev;
2129 INIT_LIST_HEAD(&vm->active_list);
2130 INIT_LIST_HEAD(&vm->inactive_list);
2131 list_add_tail(&vm->global_link, &dev_priv->vm_list);
2134 int i915_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
2136 struct drm_i915_private *dev_priv = dev->dev_private;
2139 ret = __hw_ppgtt_init(dev, ppgtt);
2141 kref_init(&ppgtt->ref);
2142 i915_address_space_init(&ppgtt->base, dev_priv);
2148 int i915_ppgtt_init_hw(struct drm_device *dev)
2150 /* In the case of execlists, PPGTT is enabled by the context descriptor
2151 * and the PDPs are contained within the context itself. We don't
2152 * need to do anything here. */
2153 if (i915.enable_execlists)
2156 if (!USES_PPGTT(dev))
2160 gen6_ppgtt_enable(dev);
2161 else if (IS_GEN7(dev))
2162 gen7_ppgtt_enable(dev);
2163 else if (INTEL_INFO(dev)->gen >= 8)
2164 gen8_ppgtt_enable(dev);
2166 MISSING_CASE(INTEL_INFO(dev)->gen);
2171 int i915_ppgtt_init_ring(struct drm_i915_gem_request *req)
2173 struct drm_i915_private *dev_priv = req->ring->dev->dev_private;
2174 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
2176 if (i915.enable_execlists)
2182 return ppgtt->switch_mm(ppgtt, req);
2185 struct i915_hw_ppgtt *
2186 i915_ppgtt_create(struct drm_device *dev, struct drm_i915_file_private *fpriv)
2188 struct i915_hw_ppgtt *ppgtt;
2191 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2193 return ERR_PTR(-ENOMEM);
2195 ret = i915_ppgtt_init(dev, ppgtt);
2198 return ERR_PTR(ret);
2201 ppgtt->file_priv = fpriv;
2203 trace_i915_ppgtt_create(&ppgtt->base);
2208 void i915_ppgtt_release(struct kref *kref)
2210 struct i915_hw_ppgtt *ppgtt =
2211 container_of(kref, struct i915_hw_ppgtt, ref);
2213 trace_i915_ppgtt_release(&ppgtt->base);
2215 /* vmas should already be unbound */
2216 WARN_ON(!list_empty(&ppgtt->base.active_list));
2217 WARN_ON(!list_empty(&ppgtt->base.inactive_list));
2219 list_del(&ppgtt->base.global_link);
2220 drm_mm_takedown(&ppgtt->base.mm);
2222 ppgtt->base.cleanup(&ppgtt->base);
2226 extern int intel_iommu_gfx_mapped;
2227 /* Certain Gen5 chipsets require require idling the GPU before
2228 * unmapping anything from the GTT when VT-d is enabled.
2230 static bool needs_idle_maps(struct drm_device *dev)
2232 #ifdef CONFIG_INTEL_IOMMU
2233 /* Query intel_iommu to see if we need the workaround. Presumably that
2236 if (IS_GEN5(dev) && IS_MOBILE(dev) && intel_iommu_gfx_mapped)
2242 static bool do_idling(struct drm_i915_private *dev_priv)
2244 bool ret = dev_priv->mm.interruptible;
2246 if (unlikely(dev_priv->gtt.do_idle_maps)) {
2247 dev_priv->mm.interruptible = false;
2248 if (i915_gpu_idle(dev_priv->dev)) {
2249 DRM_ERROR("Couldn't idle GPU\n");
2250 /* Wait a bit, in hopes it avoids the hang */
2258 static void undo_idling(struct drm_i915_private *dev_priv, bool interruptible)
2260 if (unlikely(dev_priv->gtt.do_idle_maps))
2261 dev_priv->mm.interruptible = interruptible;
2264 void i915_check_and_clear_faults(struct drm_device *dev)
2266 struct drm_i915_private *dev_priv = dev->dev_private;
2267 struct intel_engine_cs *ring;
2270 if (INTEL_INFO(dev)->gen < 6)
2273 for_each_ring(ring, dev_priv, i) {
2275 fault_reg = I915_READ(RING_FAULT_REG(ring));
2276 if (fault_reg & RING_FAULT_VALID) {
2277 DRM_DEBUG_DRIVER("Unexpected fault\n"
2279 "\tAddress space: %s\n"
2282 fault_reg & PAGE_MASK,
2283 fault_reg & RING_FAULT_GTTSEL_MASK ? "GGTT" : "PPGTT",
2284 RING_FAULT_SRCID(fault_reg),
2285 RING_FAULT_FAULT_TYPE(fault_reg));
2286 I915_WRITE(RING_FAULT_REG(ring),
2287 fault_reg & ~RING_FAULT_VALID);
2290 POSTING_READ(RING_FAULT_REG(&dev_priv->ring[RCS]));
2293 static void i915_ggtt_flush(struct drm_i915_private *dev_priv)
2295 if (INTEL_INFO(dev_priv->dev)->gen < 6) {
2296 intel_gtt_chipset_flush();
2298 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2299 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2303 void i915_gem_suspend_gtt_mappings(struct drm_device *dev)
2305 struct drm_i915_private *dev_priv = dev->dev_private;
2307 /* Don't bother messing with faults pre GEN6 as we have little
2308 * documentation supporting that it's a good idea.
2310 if (INTEL_INFO(dev)->gen < 6)
2313 i915_check_and_clear_faults(dev);
2315 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
2316 dev_priv->gtt.base.start,
2317 dev_priv->gtt.base.total,
2320 i915_ggtt_flush(dev_priv);
2323 int i915_gem_gtt_prepare_object(struct drm_i915_gem_object *obj)
2325 if (!dma_map_sg(&obj->base.dev->pdev->dev,
2326 obj->pages->sgl, obj->pages->nents,
2327 PCI_DMA_BIDIRECTIONAL))
2333 static void gen8_set_pte(void __iomem *addr, gen8_pte_t pte)
2338 iowrite32((u32)pte, addr);
2339 iowrite32(pte >> 32, addr + 4);
2343 static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
2344 struct sg_table *st,
2346 enum i915_cache_level level, u32 unused)
2348 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2349 unsigned first_entry = start >> PAGE_SHIFT;
2350 gen8_pte_t __iomem *gtt_entries =
2351 (gen8_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
2353 struct sg_page_iter sg_iter;
2354 dma_addr_t addr = 0; /* shut up gcc */
2356 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
2357 addr = sg_dma_address(sg_iter.sg) +
2358 (sg_iter.sg_pgoffset << PAGE_SHIFT);
2359 gen8_set_pte(>t_entries[i],
2360 gen8_pte_encode(addr, level, true));
2365 * XXX: This serves as a posting read to make sure that the PTE has
2366 * actually been updated. There is some concern that even though
2367 * registers and PTEs are within the same BAR that they are potentially
2368 * of NUMA access patterns. Therefore, even with the way we assume
2369 * hardware should work, we must keep this posting read for paranoia.
2372 WARN_ON(readq(>t_entries[i-1])
2373 != gen8_pte_encode(addr, level, true));
2375 /* This next bit makes the above posting read even more important. We
2376 * want to flush the TLBs only after we're certain all the PTE updates
2379 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2380 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2384 * Binds an object into the global gtt with the specified cache level. The object
2385 * will be accessible to the GPU via commands whose operands reference offsets
2386 * within the global GTT as well as accessible by the GPU through the GMADR
2387 * mapped BAR (dev_priv->mm.gtt->gtt).
2389 static void gen6_ggtt_insert_entries(struct i915_address_space *vm,
2390 struct sg_table *st,
2392 enum i915_cache_level level, u32 flags)
2394 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2395 unsigned first_entry = start >> PAGE_SHIFT;
2396 gen6_pte_t __iomem *gtt_entries =
2397 (gen6_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
2399 struct sg_page_iter sg_iter;
2400 dma_addr_t addr = 0;
2402 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
2403 addr = sg_page_iter_dma_address(&sg_iter);
2404 iowrite32(vm->pte_encode(addr, level, true, flags), >t_entries[i]);
2408 /* XXX: This serves as a posting read to make sure that the PTE has
2409 * actually been updated. There is some concern that even though
2410 * registers and PTEs are within the same BAR that they are potentially
2411 * of NUMA access patterns. Therefore, even with the way we assume
2412 * hardware should work, we must keep this posting read for paranoia.
2415 unsigned long gtt = readl(>t_entries[i-1]);
2416 WARN_ON(gtt != vm->pte_encode(addr, level, true, flags));
2419 /* This next bit makes the above posting read even more important. We
2420 * want to flush the TLBs only after we're certain all the PTE updates
2423 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2424 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2427 static void gen8_ggtt_clear_range(struct i915_address_space *vm,
2432 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2433 unsigned first_entry = start >> PAGE_SHIFT;
2434 unsigned num_entries = length >> PAGE_SHIFT;
2435 gen8_pte_t scratch_pte, __iomem *gtt_base =
2436 (gen8_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
2437 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
2440 if (WARN(num_entries > max_entries,
2441 "First entry = %d; Num entries = %d (max=%d)\n",
2442 first_entry, num_entries, max_entries))
2443 num_entries = max_entries;
2445 scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
2448 for (i = 0; i < num_entries; i++)
2449 gen8_set_pte(>t_base[i], scratch_pte);
2453 static void gen6_ggtt_clear_range(struct i915_address_space *vm,
2458 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2459 unsigned first_entry = start >> PAGE_SHIFT;
2460 unsigned num_entries = length >> PAGE_SHIFT;
2461 gen6_pte_t scratch_pte, __iomem *gtt_base =
2462 (gen6_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
2463 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
2466 if (WARN(num_entries > max_entries,
2467 "First entry = %d; Num entries = %d (max=%d)\n",
2468 first_entry, num_entries, max_entries))
2469 num_entries = max_entries;
2471 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
2472 I915_CACHE_LLC, use_scratch, 0);
2474 for (i = 0; i < num_entries; i++)
2475 iowrite32(scratch_pte, >t_base[i]);
2479 static void i915_ggtt_insert_entries(struct i915_address_space *vm,
2480 struct sg_table *pages,
2482 enum i915_cache_level cache_level, u32 unused)
2484 unsigned int flags = (cache_level == I915_CACHE_NONE) ?
2485 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
2487 intel_gtt_insert_sg_entries(pages, start >> PAGE_SHIFT, flags);
2491 static void i915_ggtt_clear_range(struct i915_address_space *vm,
2496 unsigned first_entry = start >> PAGE_SHIFT;
2497 unsigned num_entries = length >> PAGE_SHIFT;
2498 intel_gtt_clear_range(first_entry, num_entries);
2501 static int ggtt_bind_vma(struct i915_vma *vma,
2502 enum i915_cache_level cache_level,
2505 struct drm_i915_gem_object *obj = vma->obj;
2509 ret = i915_get_ggtt_vma_pages(vma);
2513 /* Currently applicable only to VLV */
2515 pte_flags |= PTE_READ_ONLY;
2517 vma->vm->insert_entries(vma->vm, vma->ggtt_view.pages,
2519 cache_level, pte_flags);
2522 * Without aliasing PPGTT there's no difference between
2523 * GLOBAL/LOCAL_BIND, it's all the same ptes. Hence unconditionally
2524 * upgrade to both bound if we bind either to avoid double-binding.
2526 vma->bound |= GLOBAL_BIND | LOCAL_BIND;
2531 struct ggtt_bind_vma__cb {
2532 struct i915_vma *vma;
2533 enum i915_cache_level cache_level;
2537 static int ggtt_bind_vma__cb(void *_arg)
2539 struct ggtt_bind_vma__cb *arg = _arg;
2540 return ggtt_bind_vma(arg->vma, arg->cache_level, arg->flags);
2543 static int ggtt_bind_vma__BKL(struct i915_vma *vma,
2544 enum i915_cache_level cache_level,
2547 struct ggtt_bind_vma__cb arg = { vma, cache_level, flags };
2548 return stop_machine(ggtt_bind_vma__cb, &arg, NULL);
2551 static int aliasing_gtt_bind_vma(struct i915_vma *vma,
2552 enum i915_cache_level cache_level,
2555 struct drm_device *dev = vma->vm->dev;
2556 struct drm_i915_private *dev_priv = dev->dev_private;
2557 struct drm_i915_gem_object *obj = vma->obj;
2558 struct sg_table *pages = obj->pages;
2562 ret = i915_get_ggtt_vma_pages(vma);
2565 pages = vma->ggtt_view.pages;
2567 /* Currently applicable only to VLV */
2569 pte_flags |= PTE_READ_ONLY;
2572 if (flags & GLOBAL_BIND) {
2573 vma->vm->insert_entries(vma->vm, pages,
2575 cache_level, pte_flags);
2578 if (flags & LOCAL_BIND) {
2579 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
2580 appgtt->base.insert_entries(&appgtt->base, pages,
2582 cache_level, pte_flags);
2588 static void ggtt_unbind_vma(struct i915_vma *vma)
2590 struct drm_device *dev = vma->vm->dev;
2591 struct drm_i915_private *dev_priv = dev->dev_private;
2592 struct drm_i915_gem_object *obj = vma->obj;
2593 const uint64_t size = min_t(uint64_t,
2597 if (vma->bound & GLOBAL_BIND) {
2598 vma->vm->clear_range(vma->vm,
2604 if (dev_priv->mm.aliasing_ppgtt && vma->bound & LOCAL_BIND) {
2605 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
2607 appgtt->base.clear_range(&appgtt->base,
2614 void i915_gem_gtt_finish_object(struct drm_i915_gem_object *obj)
2616 struct drm_device *dev = obj->base.dev;
2617 struct drm_i915_private *dev_priv = dev->dev_private;
2620 interruptible = do_idling(dev_priv);
2622 dma_unmap_sg(&dev->pdev->dev, obj->pages->sgl, obj->pages->nents,
2623 PCI_DMA_BIDIRECTIONAL);
2625 undo_idling(dev_priv, interruptible);
2628 static void i915_gtt_color_adjust(struct drm_mm_node *node,
2629 unsigned long color,
2633 if (node->color != color)
2636 if (!list_empty(&node->node_list)) {
2637 node = list_entry(node->node_list.next,
2640 if (node->allocated && node->color != color)
2645 static int i915_gem_setup_global_gtt(struct drm_device *dev,
2650 /* Let GEM Manage all of the aperture.
2652 * However, leave one page at the end still bound to the scratch page.
2653 * There are a number of places where the hardware apparently prefetches
2654 * past the end of the object, and we've seen multiple hangs with the
2655 * GPU head pointer stuck in a batchbuffer bound at the last page of the
2656 * aperture. One page should be enough to keep any prefetching inside
2659 struct drm_i915_private *dev_priv = dev->dev_private;
2660 struct i915_address_space *ggtt_vm = &dev_priv->gtt.base;
2661 struct drm_mm_node *entry;
2662 struct drm_i915_gem_object *obj;
2663 unsigned long hole_start, hole_end;
2666 BUG_ON(mappable_end > end);
2668 ggtt_vm->start = start;
2670 /* Subtract the guard page before address space initialization to
2671 * shrink the range used by drm_mm */
2672 ggtt_vm->total = end - start - PAGE_SIZE;
2673 i915_address_space_init(ggtt_vm, dev_priv);
2674 ggtt_vm->total += PAGE_SIZE;
2676 if (intel_vgpu_active(dev)) {
2677 ret = intel_vgt_balloon(dev);
2683 ggtt_vm->mm.color_adjust = i915_gtt_color_adjust;
2685 /* Mark any preallocated objects as occupied */
2686 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
2687 struct i915_vma *vma = i915_gem_obj_to_vma(obj, ggtt_vm);
2689 DRM_DEBUG_KMS("reserving preallocated space: %llx + %zx\n",
2690 i915_gem_obj_ggtt_offset(obj), obj->base.size);
2692 WARN_ON(i915_gem_obj_ggtt_bound(obj));
2693 ret = drm_mm_reserve_node(&ggtt_vm->mm, &vma->node);
2695 DRM_DEBUG_KMS("Reservation failed: %i\n", ret);
2698 vma->bound |= GLOBAL_BIND;
2699 list_add_tail(&vma->mm_list, &ggtt_vm->inactive_list);
2702 /* Clear any non-preallocated blocks */
2703 drm_mm_for_each_hole(entry, &ggtt_vm->mm, hole_start, hole_end) {
2704 DRM_DEBUG_KMS("clearing unused GTT space: [%lx, %lx]\n",
2705 hole_start, hole_end);
2706 ggtt_vm->clear_range(ggtt_vm, hole_start,
2707 hole_end - hole_start, true);
2710 /* And finally clear the reserved guard page */
2711 ggtt_vm->clear_range(ggtt_vm, end - PAGE_SIZE, PAGE_SIZE, true);
2713 if (USES_PPGTT(dev) && !USES_FULL_PPGTT(dev)) {
2714 struct i915_hw_ppgtt *ppgtt;
2716 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2720 ret = __hw_ppgtt_init(dev, ppgtt);
2722 ppgtt->base.cleanup(&ppgtt->base);
2727 if (ppgtt->base.allocate_va_range)
2728 ret = ppgtt->base.allocate_va_range(&ppgtt->base, 0,
2731 ppgtt->base.cleanup(&ppgtt->base);
2736 ppgtt->base.clear_range(&ppgtt->base,
2741 dev_priv->mm.aliasing_ppgtt = ppgtt;
2742 WARN_ON(dev_priv->gtt.base.bind_vma != ggtt_bind_vma);
2743 dev_priv->gtt.base.bind_vma = aliasing_gtt_bind_vma;
2749 void i915_gem_init_global_gtt(struct drm_device *dev)
2751 struct drm_i915_private *dev_priv = dev->dev_private;
2752 u64 gtt_size, mappable_size;
2754 gtt_size = dev_priv->gtt.base.total;
2755 mappable_size = dev_priv->gtt.mappable_end;
2757 i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
2760 void i915_global_gtt_cleanup(struct drm_device *dev)
2762 struct drm_i915_private *dev_priv = dev->dev_private;
2763 struct i915_address_space *vm = &dev_priv->gtt.base;
2765 if (dev_priv->mm.aliasing_ppgtt) {
2766 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
2768 ppgtt->base.cleanup(&ppgtt->base);
2771 if (drm_mm_initialized(&vm->mm)) {
2772 if (intel_vgpu_active(dev))
2773 intel_vgt_deballoon();
2775 drm_mm_takedown(&vm->mm);
2776 list_del(&vm->global_link);
2782 static unsigned int gen6_get_total_gtt_size(u16 snb_gmch_ctl)
2784 snb_gmch_ctl >>= SNB_GMCH_GGMS_SHIFT;
2785 snb_gmch_ctl &= SNB_GMCH_GGMS_MASK;
2786 return snb_gmch_ctl << 20;
2789 static unsigned int gen8_get_total_gtt_size(u16 bdw_gmch_ctl)
2791 bdw_gmch_ctl >>= BDW_GMCH_GGMS_SHIFT;
2792 bdw_gmch_ctl &= BDW_GMCH_GGMS_MASK;
2794 bdw_gmch_ctl = 1 << bdw_gmch_ctl;
2796 #ifdef CONFIG_X86_32
2797 /* Limit 32b platforms to a 2GB GGTT: 4 << 20 / pte size * PAGE_SIZE */
2798 if (bdw_gmch_ctl > 4)
2802 return bdw_gmch_ctl << 20;
2805 static unsigned int chv_get_total_gtt_size(u16 gmch_ctrl)
2807 gmch_ctrl >>= SNB_GMCH_GGMS_SHIFT;
2808 gmch_ctrl &= SNB_GMCH_GGMS_MASK;
2811 return 1 << (20 + gmch_ctrl);
2816 static size_t gen6_get_stolen_size(u16 snb_gmch_ctl)
2818 snb_gmch_ctl >>= SNB_GMCH_GMS_SHIFT;
2819 snb_gmch_ctl &= SNB_GMCH_GMS_MASK;
2820 return snb_gmch_ctl << 25; /* 32 MB units */
2823 static size_t gen8_get_stolen_size(u16 bdw_gmch_ctl)
2825 bdw_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2826 bdw_gmch_ctl &= BDW_GMCH_GMS_MASK;
2827 return bdw_gmch_ctl << 25; /* 32 MB units */
2830 static size_t chv_get_stolen_size(u16 gmch_ctrl)
2832 gmch_ctrl >>= SNB_GMCH_GMS_SHIFT;
2833 gmch_ctrl &= SNB_GMCH_GMS_MASK;
2836 * 0x0 to 0x10: 32MB increments starting at 0MB
2837 * 0x11 to 0x16: 4MB increments starting at 8MB
2838 * 0x17 to 0x1d: 4MB increments start at 36MB
2840 if (gmch_ctrl < 0x11)
2841 return gmch_ctrl << 25;
2842 else if (gmch_ctrl < 0x17)
2843 return (gmch_ctrl - 0x11 + 2) << 22;
2845 return (gmch_ctrl - 0x17 + 9) << 22;
2848 static size_t gen9_get_stolen_size(u16 gen9_gmch_ctl)
2850 gen9_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2851 gen9_gmch_ctl &= BDW_GMCH_GMS_MASK;
2853 if (gen9_gmch_ctl < 0xf0)
2854 return gen9_gmch_ctl << 25; /* 32 MB units */
2856 /* 4MB increments starting at 0xf0 for 4MB */
2857 return (gen9_gmch_ctl - 0xf0 + 1) << 22;
2860 static int ggtt_probe_common(struct drm_device *dev,
2863 struct drm_i915_private *dev_priv = dev->dev_private;
2864 struct i915_page_scratch *scratch_page;
2865 phys_addr_t gtt_phys_addr;
2867 /* For Modern GENs the PTEs and register space are split in the BAR */
2868 gtt_phys_addr = pci_resource_start(dev->pdev, 0) +
2869 (pci_resource_len(dev->pdev, 0) / 2);
2872 * On BXT writes larger than 64 bit to the GTT pagetable range will be
2873 * dropped. For WC mappings in general we have 64 byte burst writes
2874 * when the WC buffer is flushed, so we can't use it, but have to
2875 * resort to an uncached mapping. The WC issue is easily caught by the
2876 * readback check when writing GTT PTE entries.
2878 if (IS_BROXTON(dev))
2879 dev_priv->gtt.gsm = ioremap_nocache(gtt_phys_addr, gtt_size);
2881 dev_priv->gtt.gsm = ioremap_wc(gtt_phys_addr, gtt_size);
2882 if (!dev_priv->gtt.gsm) {
2883 DRM_ERROR("Failed to map the gtt page table\n");
2887 scratch_page = alloc_scratch_page(dev);
2888 if (IS_ERR(scratch_page)) {
2889 DRM_ERROR("Scratch setup failed\n");
2890 /* iounmap will also get called at remove, but meh */
2891 iounmap(dev_priv->gtt.gsm);
2892 return PTR_ERR(scratch_page);
2895 dev_priv->gtt.base.scratch_page = scratch_page;
2900 /* The GGTT and PPGTT need a private PPAT setup in order to handle cacheability
2901 * bits. When using advanced contexts each context stores its own PAT, but
2902 * writing this data shouldn't be harmful even in those cases. */
2903 static void bdw_setup_private_ppat(struct drm_i915_private *dev_priv)
2907 pat = GEN8_PPAT(0, GEN8_PPAT_WB | GEN8_PPAT_LLC) | /* for normal objects, no eLLC */
2908 GEN8_PPAT(1, GEN8_PPAT_WC | GEN8_PPAT_LLCELLC) | /* for something pointing to ptes? */
2909 GEN8_PPAT(2, GEN8_PPAT_WT | GEN8_PPAT_LLCELLC) | /* for scanout with eLLC */
2910 GEN8_PPAT(3, GEN8_PPAT_UC) | /* Uncached objects, mostly for scanout */
2911 GEN8_PPAT(4, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(0)) |
2912 GEN8_PPAT(5, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(1)) |
2913 GEN8_PPAT(6, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(2)) |
2914 GEN8_PPAT(7, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3));
2916 if (!USES_PPGTT(dev_priv->dev))
2917 /* Spec: "For GGTT, there is NO pat_sel[2:0] from the entry,
2918 * so RTL will always use the value corresponding to
2920 * So let's disable cache for GGTT to avoid screen corruptions.
2921 * MOCS still can be used though.
2922 * - System agent ggtt writes (i.e. cpu gtt mmaps) already work
2923 * before this patch, i.e. the same uncached + snooping access
2924 * like on gen6/7 seems to be in effect.
2925 * - So this just fixes blitter/render access. Again it looks
2926 * like it's not just uncached access, but uncached + snooping.
2927 * So we can still hold onto all our assumptions wrt cpu
2928 * clflushing on LLC machines.
2930 pat = GEN8_PPAT(0, GEN8_PPAT_UC);
2932 /* XXX: spec defines this as 2 distinct registers. It's unclear if a 64b
2933 * write would work. */
2934 I915_WRITE(GEN8_PRIVATE_PAT_LO, pat);
2935 I915_WRITE(GEN8_PRIVATE_PAT_HI, pat >> 32);
2938 static void chv_setup_private_ppat(struct drm_i915_private *dev_priv)
2943 * Map WB on BDW to snooped on CHV.
2945 * Only the snoop bit has meaning for CHV, the rest is
2948 * The hardware will never snoop for certain types of accesses:
2949 * - CPU GTT (GMADR->GGTT->no snoop->memory)
2950 * - PPGTT page tables
2951 * - some other special cycles
2953 * As with BDW, we also need to consider the following for GT accesses:
2954 * "For GGTT, there is NO pat_sel[2:0] from the entry,
2955 * so RTL will always use the value corresponding to
2957 * Which means we must set the snoop bit in PAT entry 0
2958 * in order to keep the global status page working.
2960 pat = GEN8_PPAT(0, CHV_PPAT_SNOOP) |
2964 GEN8_PPAT(4, CHV_PPAT_SNOOP) |
2965 GEN8_PPAT(5, CHV_PPAT_SNOOP) |
2966 GEN8_PPAT(6, CHV_PPAT_SNOOP) |
2967 GEN8_PPAT(7, CHV_PPAT_SNOOP);
2969 I915_WRITE(GEN8_PRIVATE_PAT_LO, pat);
2970 I915_WRITE(GEN8_PRIVATE_PAT_HI, pat >> 32);
2973 static int gen8_gmch_probe(struct drm_device *dev,
2976 phys_addr_t *mappable_base,
2979 struct drm_i915_private *dev_priv = dev->dev_private;
2984 /* TODO: We're not aware of mappable constraints on gen8 yet */
2985 *mappable_base = pci_resource_start(dev->pdev, 2);
2986 *mappable_end = pci_resource_len(dev->pdev, 2);
2988 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(39)))
2989 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(39));
2991 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
2993 if (INTEL_INFO(dev)->gen >= 9) {
2994 *stolen = gen9_get_stolen_size(snb_gmch_ctl);
2995 gtt_size = gen8_get_total_gtt_size(snb_gmch_ctl);
2996 } else if (IS_CHERRYVIEW(dev)) {
2997 *stolen = chv_get_stolen_size(snb_gmch_ctl);
2998 gtt_size = chv_get_total_gtt_size(snb_gmch_ctl);
3000 *stolen = gen8_get_stolen_size(snb_gmch_ctl);
3001 gtt_size = gen8_get_total_gtt_size(snb_gmch_ctl);
3004 *gtt_total = (gtt_size / sizeof(gen8_pte_t)) << PAGE_SHIFT;
3006 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
3007 chv_setup_private_ppat(dev_priv);
3009 bdw_setup_private_ppat(dev_priv);
3011 ret = ggtt_probe_common(dev, gtt_size);
3013 dev_priv->gtt.base.clear_range = gen8_ggtt_clear_range;
3014 dev_priv->gtt.base.insert_entries = gen8_ggtt_insert_entries;
3015 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
3016 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
3018 if (IS_CHERRYVIEW(dev))
3019 dev_priv->gtt.base.bind_vma = ggtt_bind_vma__BKL;
3024 static int gen6_gmch_probe(struct drm_device *dev,
3027 phys_addr_t *mappable_base,
3030 struct drm_i915_private *dev_priv = dev->dev_private;
3031 unsigned int gtt_size;
3035 *mappable_base = pci_resource_start(dev->pdev, 2);
3036 *mappable_end = pci_resource_len(dev->pdev, 2);
3038 /* 64/512MB is the current min/max we actually know of, but this is just
3039 * a coarse sanity check.
3041 if ((*mappable_end < (64<<20) || (*mappable_end > (512<<20)))) {
3042 DRM_ERROR("Unknown GMADR size (%llx)\n",
3043 dev_priv->gtt.mappable_end);
3047 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(40)))
3048 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(40));
3049 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
3051 *stolen = gen6_get_stolen_size(snb_gmch_ctl);
3053 gtt_size = gen6_get_total_gtt_size(snb_gmch_ctl);
3054 *gtt_total = (gtt_size / sizeof(gen6_pte_t)) << PAGE_SHIFT;
3056 ret = ggtt_probe_common(dev, gtt_size);
3058 dev_priv->gtt.base.clear_range = gen6_ggtt_clear_range;
3059 dev_priv->gtt.base.insert_entries = gen6_ggtt_insert_entries;
3060 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
3061 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
3066 static void gen6_gmch_remove(struct i915_address_space *vm)
3069 struct i915_gtt *gtt = container_of(vm, struct i915_gtt, base);
3072 free_scratch_page(vm->dev, vm->scratch_page);
3075 static int i915_gmch_probe(struct drm_device *dev,
3078 phys_addr_t *mappable_base,
3081 struct drm_i915_private *dev_priv = dev->dev_private;
3084 ret = intel_gmch_probe(dev_priv->bridge_dev, dev_priv->dev->pdev, NULL);
3086 DRM_ERROR("failed to set up gmch\n");
3090 intel_gtt_get(gtt_total, stolen, mappable_base, mappable_end);
3092 dev_priv->gtt.do_idle_maps = needs_idle_maps(dev_priv->dev);
3093 dev_priv->gtt.base.insert_entries = i915_ggtt_insert_entries;
3094 dev_priv->gtt.base.clear_range = i915_ggtt_clear_range;
3095 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
3096 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
3098 if (unlikely(dev_priv->gtt.do_idle_maps))
3099 DRM_INFO("applying Ironlake quirks for intel_iommu\n");
3104 static void i915_gmch_remove(struct i915_address_space *vm)
3106 intel_gmch_remove();
3109 int i915_gem_gtt_init(struct drm_device *dev)
3111 struct drm_i915_private *dev_priv = dev->dev_private;
3112 struct i915_gtt *gtt = &dev_priv->gtt;
3115 if (INTEL_INFO(dev)->gen <= 5) {
3116 gtt->gtt_probe = i915_gmch_probe;
3117 gtt->base.cleanup = i915_gmch_remove;
3118 } else if (INTEL_INFO(dev)->gen < 8) {
3119 gtt->gtt_probe = gen6_gmch_probe;
3120 gtt->base.cleanup = gen6_gmch_remove;
3121 if (IS_HASWELL(dev) && dev_priv->ellc_size)
3122 gtt->base.pte_encode = iris_pte_encode;
3123 else if (IS_HASWELL(dev))
3124 gtt->base.pte_encode = hsw_pte_encode;
3125 else if (IS_VALLEYVIEW(dev))
3126 gtt->base.pte_encode = byt_pte_encode;
3127 else if (INTEL_INFO(dev)->gen >= 7)
3128 gtt->base.pte_encode = ivb_pte_encode;
3130 gtt->base.pte_encode = snb_pte_encode;
3132 dev_priv->gtt.gtt_probe = gen8_gmch_probe;
3133 dev_priv->gtt.base.cleanup = gen6_gmch_remove;
3136 gtt->base.dev = dev;
3138 ret = gtt->gtt_probe(dev, >t->base.total, >t->stolen_size,
3139 >t->mappable_base, >t->mappable_end);
3143 /* GMADR is the PCI mmio aperture into the global GTT. */
3144 DRM_INFO("Memory usable by graphics device = %lluM\n",
3145 gtt->base.total >> 20);
3146 DRM_DEBUG_DRIVER("GMADR size = %lldM\n", gtt->mappable_end >> 20);
3147 DRM_DEBUG_DRIVER("GTT stolen size = %zdM\n", gtt->stolen_size >> 20);
3148 #ifdef CONFIG_INTEL_IOMMU
3149 if (intel_iommu_gfx_mapped)
3150 DRM_INFO("VT-d active for gfx access\n");
3153 * i915.enable_ppgtt is read-only, so do an early pass to validate the
3154 * user's requested state against the hardware/driver capabilities. We
3155 * do this now so that we can print out any log messages once rather
3156 * than every time we check intel_enable_ppgtt().
3158 i915.enable_ppgtt = sanitize_enable_ppgtt(dev, i915.enable_ppgtt);
3159 DRM_DEBUG_DRIVER("ppgtt mode: %i\n", i915.enable_ppgtt);
3164 void i915_gem_restore_gtt_mappings(struct drm_device *dev)
3166 struct drm_i915_private *dev_priv = dev->dev_private;
3167 struct drm_i915_gem_object *obj;
3168 struct i915_address_space *vm;
3169 struct i915_vma *vma;
3172 i915_check_and_clear_faults(dev);
3174 /* First fill our portion of the GTT with scratch pages */
3175 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
3176 dev_priv->gtt.base.start,
3177 dev_priv->gtt.base.total,
3180 /* Cache flush objects bound into GGTT and rebind them. */
3181 vm = &dev_priv->gtt.base;
3182 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
3184 list_for_each_entry(vma, &obj->vma_list, vma_link) {
3188 WARN_ON(i915_vma_bind(vma, obj->cache_level,
3195 i915_gem_clflush_object(obj, obj->pin_display);
3198 if (INTEL_INFO(dev)->gen >= 8) {
3199 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
3200 chv_setup_private_ppat(dev_priv);
3202 bdw_setup_private_ppat(dev_priv);
3207 if (USES_PPGTT(dev)) {
3208 list_for_each_entry(vm, &dev_priv->vm_list, global_link) {
3209 /* TODO: Perhaps it shouldn't be gen6 specific */
3211 struct i915_hw_ppgtt *ppgtt =
3212 container_of(vm, struct i915_hw_ppgtt,
3215 if (i915_is_ggtt(vm))
3216 ppgtt = dev_priv->mm.aliasing_ppgtt;
3218 gen6_write_page_range(dev_priv, &ppgtt->pd,
3219 0, ppgtt->base.total);
3223 i915_ggtt_flush(dev_priv);
3226 static struct i915_vma *
3227 __i915_gem_vma_create(struct drm_i915_gem_object *obj,
3228 struct i915_address_space *vm,
3229 const struct i915_ggtt_view *ggtt_view)
3231 struct i915_vma *vma;
3233 if (WARN_ON(i915_is_ggtt(vm) != !!ggtt_view))
3234 return ERR_PTR(-EINVAL);
3236 vma = kmem_cache_zalloc(to_i915(obj->base.dev)->vmas, GFP_KERNEL);
3238 return ERR_PTR(-ENOMEM);
3240 INIT_LIST_HEAD(&vma->vma_link);
3241 INIT_LIST_HEAD(&vma->mm_list);
3242 INIT_LIST_HEAD(&vma->exec_list);
3246 if (i915_is_ggtt(vm))
3247 vma->ggtt_view = *ggtt_view;
3249 list_add_tail(&vma->vma_link, &obj->vma_list);
3250 if (!i915_is_ggtt(vm))
3251 i915_ppgtt_get(i915_vm_to_ppgtt(vm));
3257 i915_gem_obj_lookup_or_create_vma(struct drm_i915_gem_object *obj,
3258 struct i915_address_space *vm)
3260 struct i915_vma *vma;
3262 vma = i915_gem_obj_to_vma(obj, vm);
3264 vma = __i915_gem_vma_create(obj, vm,
3265 i915_is_ggtt(vm) ? &i915_ggtt_view_normal : NULL);
3271 i915_gem_obj_lookup_or_create_ggtt_vma(struct drm_i915_gem_object *obj,
3272 const struct i915_ggtt_view *view)
3274 struct i915_address_space *ggtt = i915_obj_to_ggtt(obj);
3275 struct i915_vma *vma;
3278 return ERR_PTR(-EINVAL);
3280 vma = i915_gem_obj_to_ggtt_view(obj, view);
3286 vma = __i915_gem_vma_create(obj, ggtt, view);
3292 static struct scatterlist *
3293 rotate_pages(dma_addr_t *in, unsigned int offset,
3294 unsigned int width, unsigned int height,
3295 struct sg_table *st, struct scatterlist *sg)
3297 unsigned int column, row;
3298 unsigned int src_idx;
3305 for (column = 0; column < width; column++) {
3306 src_idx = width * (height - 1) + column;
3307 for (row = 0; row < height; row++) {
3309 /* We don't need the pages, but need to initialize
3310 * the entries so the sg list can be happily traversed.
3311 * The only thing we need are DMA addresses.
3313 sg_set_page(sg, NULL, PAGE_SIZE, 0);
3314 sg_dma_address(sg) = in[offset + src_idx];
3315 sg_dma_len(sg) = PAGE_SIZE;
3324 static struct sg_table *
3325 intel_rotate_fb_obj_pages(struct i915_ggtt_view *ggtt_view,
3326 struct drm_i915_gem_object *obj)
3328 struct intel_rotation_info *rot_info = &ggtt_view->params.rotation_info;
3329 unsigned int size_pages = rot_info->size >> PAGE_SHIFT;
3330 unsigned int size_pages_uv;
3331 struct sg_page_iter sg_iter;
3333 dma_addr_t *page_addr_list;
3334 struct sg_table *st;
3335 unsigned int uv_start_page;
3336 struct scatterlist *sg;
3339 /* Allocate a temporary list of source pages for random access. */
3340 page_addr_list = drm_malloc_ab(obj->base.size / PAGE_SIZE,
3341 sizeof(dma_addr_t));
3342 if (!page_addr_list)
3343 return ERR_PTR(ret);
3345 /* Account for UV plane with NV12. */
3346 if (rot_info->pixel_format == DRM_FORMAT_NV12)
3347 size_pages_uv = rot_info->size_uv >> PAGE_SHIFT;
3351 /* Allocate target SG list. */
3352 st = kmalloc(sizeof(*st), GFP_KERNEL);
3356 ret = sg_alloc_table(st, size_pages + size_pages_uv, GFP_KERNEL);
3360 /* Populate source page list from the object. */
3362 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) {
3363 page_addr_list[i] = sg_page_iter_dma_address(&sg_iter);
3367 /* Rotate the pages. */
3368 sg = rotate_pages(page_addr_list, 0,
3369 rot_info->width_pages, rot_info->height_pages,
3372 /* Append the UV plane if NV12. */
3373 if (rot_info->pixel_format == DRM_FORMAT_NV12) {
3374 uv_start_page = size_pages;
3376 /* Check for tile-row un-alignment. */
3377 if (offset_in_page(rot_info->uv_offset))
3380 rot_info->uv_start_page = uv_start_page;
3382 rotate_pages(page_addr_list, uv_start_page,
3383 rot_info->width_pages_uv,
3384 rot_info->height_pages_uv,
3389 "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",
3390 obj->base.size, rot_info->pitch, rot_info->height,
3391 rot_info->pixel_format, rot_info->width_pages,
3392 rot_info->height_pages, size_pages + size_pages_uv,
3395 drm_free_large(page_addr_list);
3402 drm_free_large(page_addr_list);
3405 "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",
3406 obj->base.size, ret, rot_info->pitch, rot_info->height,
3407 rot_info->pixel_format, rot_info->width_pages,
3408 rot_info->height_pages, size_pages + size_pages_uv,
3410 return ERR_PTR(ret);
3413 static struct sg_table *
3414 intel_partial_pages(const struct i915_ggtt_view *view,
3415 struct drm_i915_gem_object *obj)
3417 struct sg_table *st;
3418 struct scatterlist *sg;
3419 struct sg_page_iter obj_sg_iter;
3422 st = kmalloc(sizeof(*st), GFP_KERNEL);
3426 ret = sg_alloc_table(st, view->params.partial.size, GFP_KERNEL);
3432 for_each_sg_page(obj->pages->sgl, &obj_sg_iter, obj->pages->nents,
3433 view->params.partial.offset)
3435 if (st->nents >= view->params.partial.size)
3438 sg_set_page(sg, NULL, PAGE_SIZE, 0);
3439 sg_dma_address(sg) = sg_page_iter_dma_address(&obj_sg_iter);
3440 sg_dma_len(sg) = PAGE_SIZE;
3451 return ERR_PTR(ret);
3455 i915_get_ggtt_vma_pages(struct i915_vma *vma)
3459 if (vma->ggtt_view.pages)
3462 if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL)
3463 vma->ggtt_view.pages = vma->obj->pages;
3464 else if (vma->ggtt_view.type == I915_GGTT_VIEW_ROTATED)
3465 vma->ggtt_view.pages =
3466 intel_rotate_fb_obj_pages(&vma->ggtt_view, vma->obj);
3467 else if (vma->ggtt_view.type == I915_GGTT_VIEW_PARTIAL)
3468 vma->ggtt_view.pages =
3469 intel_partial_pages(&vma->ggtt_view, vma->obj);
3471 WARN_ONCE(1, "GGTT view %u not implemented!\n",
3472 vma->ggtt_view.type);
3474 if (!vma->ggtt_view.pages) {
3475 DRM_ERROR("Failed to get pages for GGTT view type %u!\n",
3476 vma->ggtt_view.type);
3478 } else if (IS_ERR(vma->ggtt_view.pages)) {
3479 ret = PTR_ERR(vma->ggtt_view.pages);
3480 vma->ggtt_view.pages = NULL;
3481 DRM_ERROR("Failed to get pages for VMA view type %u (%d)!\n",
3482 vma->ggtt_view.type, ret);
3489 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
3491 * @cache_level: mapping cache level
3492 * @flags: flags like global or local mapping
3494 * DMA addresses are taken from the scatter-gather table of this object (or of
3495 * this VMA in case of non-default GGTT views) and PTE entries set up.
3496 * Note that DMA addresses are also the only part of the SG table we care about.
3498 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
3504 if (WARN_ON(flags == 0))
3508 if (flags & PIN_GLOBAL)
3509 bind_flags |= GLOBAL_BIND;
3510 if (flags & PIN_USER)
3511 bind_flags |= LOCAL_BIND;
3513 if (flags & PIN_UPDATE)
3514 bind_flags |= vma->bound;
3516 bind_flags &= ~vma->bound;
3518 if (bind_flags == 0)
3521 if (vma->bound == 0 && vma->vm->allocate_va_range) {
3522 trace_i915_va_alloc(vma->vm,
3525 VM_TO_TRACE_NAME(vma->vm));
3527 /* XXX: i915_vma_pin() will fix this +- hack */
3529 ret = vma->vm->allocate_va_range(vma->vm,
3537 ret = vma->vm->bind_vma(vma, cache_level, bind_flags);
3541 vma->bound |= bind_flags;
3547 * i915_ggtt_view_size - Get the size of a GGTT view.
3548 * @obj: Object the view is of.
3549 * @view: The view in question.
3551 * @return The size of the GGTT view in bytes.
3554 i915_ggtt_view_size(struct drm_i915_gem_object *obj,
3555 const struct i915_ggtt_view *view)
3557 if (view->type == I915_GGTT_VIEW_NORMAL) {
3558 return obj->base.size;
3559 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
3560 return view->params.rotation_info.size;
3561 } else if (view->type == I915_GGTT_VIEW_PARTIAL) {
3562 return view->params.partial.size << PAGE_SHIFT;
3564 WARN_ONCE(1, "GGTT view %u not implemented!\n", view->type);
3565 return obj->base.size;