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).
96 static inline struct i915_ggtt *
97 i915_vm_to_ggtt(struct i915_address_space *vm)
99 GEM_BUG_ON(!i915_is_ggtt(vm));
100 return container_of(vm, struct i915_ggtt, base);
104 i915_get_ggtt_vma_pages(struct i915_vma *vma);
106 const struct i915_ggtt_view i915_ggtt_view_normal = {
107 .type = I915_GGTT_VIEW_NORMAL,
109 const struct i915_ggtt_view i915_ggtt_view_rotated = {
110 .type = I915_GGTT_VIEW_ROTATED,
113 int intel_sanitize_enable_ppgtt(struct drm_i915_private *dev_priv,
116 bool has_aliasing_ppgtt;
118 bool has_full_48bit_ppgtt;
120 has_aliasing_ppgtt = INTEL_GEN(dev_priv) >= 6;
121 has_full_ppgtt = INTEL_GEN(dev_priv) >= 7;
122 has_full_48bit_ppgtt =
123 IS_BROADWELL(dev_priv) || INTEL_GEN(dev_priv) >= 9;
125 if (intel_vgpu_active(dev_priv))
126 has_full_ppgtt = false; /* emulation is too hard */
128 if (!has_aliasing_ppgtt)
132 * We don't allow disabling PPGTT for gen9+ as it's a requirement for
133 * execlists, the sole mechanism available to submit work.
135 if (enable_ppgtt == 0 && INTEL_GEN(dev_priv) < 9)
138 if (enable_ppgtt == 1)
141 if (enable_ppgtt == 2 && has_full_ppgtt)
144 if (enable_ppgtt == 3 && has_full_48bit_ppgtt)
147 #ifdef CONFIG_INTEL_IOMMU
148 /* Disable ppgtt on SNB if VT-d is on. */
149 if (IS_GEN6(dev_priv) && intel_iommu_gfx_mapped) {
150 DRM_INFO("Disabling PPGTT because VT-d is on\n");
155 /* Early VLV doesn't have this */
156 if (IS_VALLEYVIEW(dev_priv) && dev_priv->dev->pdev->revision < 0xb) {
157 DRM_DEBUG_DRIVER("disabling PPGTT on pre-B3 step VLV\n");
161 if (INTEL_GEN(dev_priv) >= 8 && i915.enable_execlists)
162 return has_full_48bit_ppgtt ? 3 : 2;
164 return has_aliasing_ppgtt ? 1 : 0;
167 static int ppgtt_bind_vma(struct i915_vma *vma,
168 enum i915_cache_level cache_level,
173 /* Currently applicable only to VLV */
175 pte_flags |= PTE_READ_ONLY;
177 vma->vm->insert_entries(vma->vm, vma->obj->pages, vma->node.start,
178 cache_level, pte_flags);
183 static void ppgtt_unbind_vma(struct i915_vma *vma)
185 vma->vm->clear_range(vma->vm,
191 static gen8_pte_t gen8_pte_encode(dma_addr_t addr,
192 enum i915_cache_level level,
195 gen8_pte_t pte = valid ? _PAGE_PRESENT | _PAGE_RW : 0;
199 case I915_CACHE_NONE:
200 pte |= PPAT_UNCACHED_INDEX;
203 pte |= PPAT_DISPLAY_ELLC_INDEX;
206 pte |= PPAT_CACHED_INDEX;
213 static gen8_pde_t gen8_pde_encode(const dma_addr_t addr,
214 const enum i915_cache_level level)
216 gen8_pde_t pde = _PAGE_PRESENT | _PAGE_RW;
218 if (level != I915_CACHE_NONE)
219 pde |= PPAT_CACHED_PDE_INDEX;
221 pde |= PPAT_UNCACHED_INDEX;
225 #define gen8_pdpe_encode gen8_pde_encode
226 #define gen8_pml4e_encode gen8_pde_encode
228 static gen6_pte_t snb_pte_encode(dma_addr_t addr,
229 enum i915_cache_level level,
230 bool valid, u32 unused)
232 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
233 pte |= GEN6_PTE_ADDR_ENCODE(addr);
236 case I915_CACHE_L3_LLC:
238 pte |= GEN6_PTE_CACHE_LLC;
240 case I915_CACHE_NONE:
241 pte |= GEN6_PTE_UNCACHED;
250 static gen6_pte_t ivb_pte_encode(dma_addr_t addr,
251 enum i915_cache_level level,
252 bool valid, u32 unused)
254 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
255 pte |= GEN6_PTE_ADDR_ENCODE(addr);
258 case I915_CACHE_L3_LLC:
259 pte |= GEN7_PTE_CACHE_L3_LLC;
262 pte |= GEN6_PTE_CACHE_LLC;
264 case I915_CACHE_NONE:
265 pte |= GEN6_PTE_UNCACHED;
274 static gen6_pte_t byt_pte_encode(dma_addr_t addr,
275 enum i915_cache_level level,
276 bool valid, u32 flags)
278 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
279 pte |= GEN6_PTE_ADDR_ENCODE(addr);
281 if (!(flags & PTE_READ_ONLY))
282 pte |= BYT_PTE_WRITEABLE;
284 if (level != I915_CACHE_NONE)
285 pte |= BYT_PTE_SNOOPED_BY_CPU_CACHES;
290 static gen6_pte_t hsw_pte_encode(dma_addr_t addr,
291 enum i915_cache_level level,
292 bool valid, u32 unused)
294 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
295 pte |= HSW_PTE_ADDR_ENCODE(addr);
297 if (level != I915_CACHE_NONE)
298 pte |= HSW_WB_LLC_AGE3;
303 static gen6_pte_t iris_pte_encode(dma_addr_t addr,
304 enum i915_cache_level level,
305 bool valid, u32 unused)
307 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
308 pte |= HSW_PTE_ADDR_ENCODE(addr);
311 case I915_CACHE_NONE:
314 pte |= HSW_WT_ELLC_LLC_AGE3;
317 pte |= HSW_WB_ELLC_LLC_AGE3;
324 static int __setup_page_dma(struct drm_device *dev,
325 struct i915_page_dma *p, gfp_t flags)
327 struct device *device = &dev->pdev->dev;
329 p->page = alloc_page(flags);
333 p->daddr = dma_map_page(device,
334 p->page, 0, 4096, PCI_DMA_BIDIRECTIONAL);
336 if (dma_mapping_error(device, p->daddr)) {
337 __free_page(p->page);
344 static int setup_page_dma(struct drm_device *dev, struct i915_page_dma *p)
346 return __setup_page_dma(dev, p, GFP_KERNEL);
349 static void cleanup_page_dma(struct drm_device *dev, struct i915_page_dma *p)
351 if (WARN_ON(!p->page))
354 dma_unmap_page(&dev->pdev->dev, p->daddr, 4096, PCI_DMA_BIDIRECTIONAL);
355 __free_page(p->page);
356 memset(p, 0, sizeof(*p));
359 static void *kmap_page_dma(struct i915_page_dma *p)
361 return kmap_atomic(p->page);
364 /* We use the flushing unmap only with ppgtt structures:
365 * page directories, page tables and scratch pages.
367 static void kunmap_page_dma(struct drm_device *dev, void *vaddr)
369 /* There are only few exceptions for gen >=6. chv and bxt.
370 * And we are not sure about the latter so play safe for now.
372 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
373 drm_clflush_virt_range(vaddr, PAGE_SIZE);
375 kunmap_atomic(vaddr);
378 #define kmap_px(px) kmap_page_dma(px_base(px))
379 #define kunmap_px(ppgtt, vaddr) kunmap_page_dma((ppgtt)->base.dev, (vaddr))
381 #define setup_px(dev, px) setup_page_dma((dev), px_base(px))
382 #define cleanup_px(dev, px) cleanup_page_dma((dev), px_base(px))
383 #define fill_px(dev, px, v) fill_page_dma((dev), px_base(px), (v))
384 #define fill32_px(dev, px, v) fill_page_dma_32((dev), px_base(px), (v))
386 static void fill_page_dma(struct drm_device *dev, struct i915_page_dma *p,
390 uint64_t * const vaddr = kmap_page_dma(p);
392 for (i = 0; i < 512; i++)
395 kunmap_page_dma(dev, vaddr);
398 static void fill_page_dma_32(struct drm_device *dev, struct i915_page_dma *p,
399 const uint32_t val32)
405 fill_page_dma(dev, p, v);
408 static struct i915_page_scratch *alloc_scratch_page(struct drm_device *dev)
410 struct i915_page_scratch *sp;
413 sp = kzalloc(sizeof(*sp), GFP_KERNEL);
415 return ERR_PTR(-ENOMEM);
417 ret = __setup_page_dma(dev, px_base(sp), GFP_DMA32 | __GFP_ZERO);
423 set_pages_uc(px_page(sp), 1);
428 static void free_scratch_page(struct drm_device *dev,
429 struct i915_page_scratch *sp)
431 set_pages_wb(px_page(sp), 1);
437 static struct i915_page_table *alloc_pt(struct drm_device *dev)
439 struct i915_page_table *pt;
440 const size_t count = INTEL_INFO(dev)->gen >= 8 ?
441 GEN8_PTES : GEN6_PTES;
444 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
446 return ERR_PTR(-ENOMEM);
448 pt->used_ptes = kcalloc(BITS_TO_LONGS(count), sizeof(*pt->used_ptes),
454 ret = setup_px(dev, pt);
461 kfree(pt->used_ptes);
468 static void free_pt(struct drm_device *dev, struct i915_page_table *pt)
471 kfree(pt->used_ptes);
475 static void gen8_initialize_pt(struct i915_address_space *vm,
476 struct i915_page_table *pt)
478 gen8_pte_t scratch_pte;
480 scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
481 I915_CACHE_LLC, true);
483 fill_px(vm->dev, pt, scratch_pte);
486 static void gen6_initialize_pt(struct i915_address_space *vm,
487 struct i915_page_table *pt)
489 gen6_pte_t scratch_pte;
491 WARN_ON(px_dma(vm->scratch_page) == 0);
493 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
494 I915_CACHE_LLC, true, 0);
496 fill32_px(vm->dev, pt, scratch_pte);
499 static struct i915_page_directory *alloc_pd(struct drm_device *dev)
501 struct i915_page_directory *pd;
504 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
506 return ERR_PTR(-ENOMEM);
508 pd->used_pdes = kcalloc(BITS_TO_LONGS(I915_PDES),
509 sizeof(*pd->used_pdes), GFP_KERNEL);
513 ret = setup_px(dev, pd);
520 kfree(pd->used_pdes);
527 static void free_pd(struct drm_device *dev, struct i915_page_directory *pd)
531 kfree(pd->used_pdes);
536 static void gen8_initialize_pd(struct i915_address_space *vm,
537 struct i915_page_directory *pd)
539 gen8_pde_t scratch_pde;
541 scratch_pde = gen8_pde_encode(px_dma(vm->scratch_pt), I915_CACHE_LLC);
543 fill_px(vm->dev, pd, scratch_pde);
546 static int __pdp_init(struct drm_device *dev,
547 struct i915_page_directory_pointer *pdp)
549 size_t pdpes = I915_PDPES_PER_PDP(dev);
551 pdp->used_pdpes = kcalloc(BITS_TO_LONGS(pdpes),
552 sizeof(unsigned long),
554 if (!pdp->used_pdpes)
557 pdp->page_directory = kcalloc(pdpes, sizeof(*pdp->page_directory),
559 if (!pdp->page_directory) {
560 kfree(pdp->used_pdpes);
561 /* the PDP might be the statically allocated top level. Keep it
562 * as clean as possible */
563 pdp->used_pdpes = NULL;
570 static void __pdp_fini(struct i915_page_directory_pointer *pdp)
572 kfree(pdp->used_pdpes);
573 kfree(pdp->page_directory);
574 pdp->page_directory = NULL;
578 i915_page_directory_pointer *alloc_pdp(struct drm_device *dev)
580 struct i915_page_directory_pointer *pdp;
583 WARN_ON(!USES_FULL_48BIT_PPGTT(dev));
585 pdp = kzalloc(sizeof(*pdp), GFP_KERNEL);
587 return ERR_PTR(-ENOMEM);
589 ret = __pdp_init(dev, pdp);
593 ret = setup_px(dev, pdp);
607 static void free_pdp(struct drm_device *dev,
608 struct i915_page_directory_pointer *pdp)
611 if (USES_FULL_48BIT_PPGTT(dev)) {
612 cleanup_px(dev, pdp);
617 static void gen8_initialize_pdp(struct i915_address_space *vm,
618 struct i915_page_directory_pointer *pdp)
620 gen8_ppgtt_pdpe_t scratch_pdpe;
622 scratch_pdpe = gen8_pdpe_encode(px_dma(vm->scratch_pd), I915_CACHE_LLC);
624 fill_px(vm->dev, pdp, scratch_pdpe);
627 static void gen8_initialize_pml4(struct i915_address_space *vm,
628 struct i915_pml4 *pml4)
630 gen8_ppgtt_pml4e_t scratch_pml4e;
632 scratch_pml4e = gen8_pml4e_encode(px_dma(vm->scratch_pdp),
635 fill_px(vm->dev, pml4, scratch_pml4e);
639 gen8_setup_page_directory(struct i915_hw_ppgtt *ppgtt,
640 struct i915_page_directory_pointer *pdp,
641 struct i915_page_directory *pd,
644 gen8_ppgtt_pdpe_t *page_directorypo;
646 if (!USES_FULL_48BIT_PPGTT(ppgtt->base.dev))
649 page_directorypo = kmap_px(pdp);
650 page_directorypo[index] = gen8_pdpe_encode(px_dma(pd), I915_CACHE_LLC);
651 kunmap_px(ppgtt, page_directorypo);
655 gen8_setup_page_directory_pointer(struct i915_hw_ppgtt *ppgtt,
656 struct i915_pml4 *pml4,
657 struct i915_page_directory_pointer *pdp,
660 gen8_ppgtt_pml4e_t *pagemap = kmap_px(pml4);
662 WARN_ON(!USES_FULL_48BIT_PPGTT(ppgtt->base.dev));
663 pagemap[index] = gen8_pml4e_encode(px_dma(pdp), I915_CACHE_LLC);
664 kunmap_px(ppgtt, pagemap);
667 /* Broadwell Page Directory Pointer Descriptors */
668 static int gen8_write_pdp(struct drm_i915_gem_request *req,
672 struct intel_engine_cs *engine = req->engine;
677 ret = intel_ring_begin(req, 6);
681 intel_ring_emit(engine, MI_LOAD_REGISTER_IMM(1));
682 intel_ring_emit_reg(engine, GEN8_RING_PDP_UDW(engine, entry));
683 intel_ring_emit(engine, upper_32_bits(addr));
684 intel_ring_emit(engine, MI_LOAD_REGISTER_IMM(1));
685 intel_ring_emit_reg(engine, GEN8_RING_PDP_LDW(engine, entry));
686 intel_ring_emit(engine, lower_32_bits(addr));
687 intel_ring_advance(engine);
692 static int gen8_legacy_mm_switch(struct i915_hw_ppgtt *ppgtt,
693 struct drm_i915_gem_request *req)
697 for (i = GEN8_LEGACY_PDPES - 1; i >= 0; i--) {
698 const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i);
700 ret = gen8_write_pdp(req, i, pd_daddr);
708 static int gen8_48b_mm_switch(struct i915_hw_ppgtt *ppgtt,
709 struct drm_i915_gem_request *req)
711 return gen8_write_pdp(req, 0, px_dma(&ppgtt->pml4));
714 static void gen8_ppgtt_clear_pte_range(struct i915_address_space *vm,
715 struct i915_page_directory_pointer *pdp,
718 gen8_pte_t scratch_pte)
720 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
721 gen8_pte_t *pt_vaddr;
722 unsigned pdpe = gen8_pdpe_index(start);
723 unsigned pde = gen8_pde_index(start);
724 unsigned pte = gen8_pte_index(start);
725 unsigned num_entries = length >> PAGE_SHIFT;
726 unsigned last_pte, i;
731 while (num_entries) {
732 struct i915_page_directory *pd;
733 struct i915_page_table *pt;
735 if (WARN_ON(!pdp->page_directory[pdpe]))
738 pd = pdp->page_directory[pdpe];
740 if (WARN_ON(!pd->page_table[pde]))
743 pt = pd->page_table[pde];
745 if (WARN_ON(!px_page(pt)))
748 last_pte = pte + num_entries;
749 if (last_pte > GEN8_PTES)
750 last_pte = GEN8_PTES;
752 pt_vaddr = kmap_px(pt);
754 for (i = pte; i < last_pte; i++) {
755 pt_vaddr[i] = scratch_pte;
759 kunmap_px(ppgtt, pt_vaddr);
762 if (++pde == I915_PDES) {
763 if (++pdpe == I915_PDPES_PER_PDP(vm->dev))
770 static void gen8_ppgtt_clear_range(struct i915_address_space *vm,
775 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
776 gen8_pte_t scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
777 I915_CACHE_LLC, use_scratch);
779 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
780 gen8_ppgtt_clear_pte_range(vm, &ppgtt->pdp, start, length,
784 struct i915_page_directory_pointer *pdp;
786 gen8_for_each_pml4e(pdp, &ppgtt->pml4, start, length, pml4e) {
787 gen8_ppgtt_clear_pte_range(vm, pdp, start, length,
794 gen8_ppgtt_insert_pte_entries(struct i915_address_space *vm,
795 struct i915_page_directory_pointer *pdp,
796 struct sg_page_iter *sg_iter,
798 enum i915_cache_level cache_level)
800 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
801 gen8_pte_t *pt_vaddr;
802 unsigned pdpe = gen8_pdpe_index(start);
803 unsigned pde = gen8_pde_index(start);
804 unsigned pte = gen8_pte_index(start);
808 while (__sg_page_iter_next(sg_iter)) {
809 if (pt_vaddr == NULL) {
810 struct i915_page_directory *pd = pdp->page_directory[pdpe];
811 struct i915_page_table *pt = pd->page_table[pde];
812 pt_vaddr = kmap_px(pt);
816 gen8_pte_encode(sg_page_iter_dma_address(sg_iter),
818 if (++pte == GEN8_PTES) {
819 kunmap_px(ppgtt, pt_vaddr);
821 if (++pde == I915_PDES) {
822 if (++pdpe == I915_PDPES_PER_PDP(vm->dev))
831 kunmap_px(ppgtt, pt_vaddr);
834 static void gen8_ppgtt_insert_entries(struct i915_address_space *vm,
835 struct sg_table *pages,
837 enum i915_cache_level cache_level,
840 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
841 struct sg_page_iter sg_iter;
843 __sg_page_iter_start(&sg_iter, pages->sgl, sg_nents(pages->sgl), 0);
845 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
846 gen8_ppgtt_insert_pte_entries(vm, &ppgtt->pdp, &sg_iter, start,
849 struct i915_page_directory_pointer *pdp;
851 uint64_t length = (uint64_t)pages->orig_nents << PAGE_SHIFT;
853 gen8_for_each_pml4e(pdp, &ppgtt->pml4, start, length, pml4e) {
854 gen8_ppgtt_insert_pte_entries(vm, pdp, &sg_iter,
860 static void gen8_free_page_tables(struct drm_device *dev,
861 struct i915_page_directory *pd)
868 for_each_set_bit(i, pd->used_pdes, I915_PDES) {
869 if (WARN_ON(!pd->page_table[i]))
872 free_pt(dev, pd->page_table[i]);
873 pd->page_table[i] = NULL;
877 static int gen8_init_scratch(struct i915_address_space *vm)
879 struct drm_device *dev = vm->dev;
882 vm->scratch_page = alloc_scratch_page(dev);
883 if (IS_ERR(vm->scratch_page))
884 return PTR_ERR(vm->scratch_page);
886 vm->scratch_pt = alloc_pt(dev);
887 if (IS_ERR(vm->scratch_pt)) {
888 ret = PTR_ERR(vm->scratch_pt);
889 goto free_scratch_page;
892 vm->scratch_pd = alloc_pd(dev);
893 if (IS_ERR(vm->scratch_pd)) {
894 ret = PTR_ERR(vm->scratch_pd);
898 if (USES_FULL_48BIT_PPGTT(dev)) {
899 vm->scratch_pdp = alloc_pdp(dev);
900 if (IS_ERR(vm->scratch_pdp)) {
901 ret = PTR_ERR(vm->scratch_pdp);
906 gen8_initialize_pt(vm, vm->scratch_pt);
907 gen8_initialize_pd(vm, vm->scratch_pd);
908 if (USES_FULL_48BIT_PPGTT(dev))
909 gen8_initialize_pdp(vm, vm->scratch_pdp);
914 free_pd(dev, vm->scratch_pd);
916 free_pt(dev, vm->scratch_pt);
918 free_scratch_page(dev, vm->scratch_page);
923 static int gen8_ppgtt_notify_vgt(struct i915_hw_ppgtt *ppgtt, bool create)
925 enum vgt_g2v_type msg;
926 struct drm_i915_private *dev_priv = to_i915(ppgtt->base.dev);
929 if (USES_FULL_48BIT_PPGTT(dev_priv)) {
930 u64 daddr = px_dma(&ppgtt->pml4);
932 I915_WRITE(vgtif_reg(pdp[0].lo), lower_32_bits(daddr));
933 I915_WRITE(vgtif_reg(pdp[0].hi), upper_32_bits(daddr));
935 msg = (create ? VGT_G2V_PPGTT_L4_PAGE_TABLE_CREATE :
936 VGT_G2V_PPGTT_L4_PAGE_TABLE_DESTROY);
938 for (i = 0; i < GEN8_LEGACY_PDPES; i++) {
939 u64 daddr = i915_page_dir_dma_addr(ppgtt, i);
941 I915_WRITE(vgtif_reg(pdp[i].lo), lower_32_bits(daddr));
942 I915_WRITE(vgtif_reg(pdp[i].hi), upper_32_bits(daddr));
945 msg = (create ? VGT_G2V_PPGTT_L3_PAGE_TABLE_CREATE :
946 VGT_G2V_PPGTT_L3_PAGE_TABLE_DESTROY);
949 I915_WRITE(vgtif_reg(g2v_notify), msg);
954 static void gen8_free_scratch(struct i915_address_space *vm)
956 struct drm_device *dev = vm->dev;
958 if (USES_FULL_48BIT_PPGTT(dev))
959 free_pdp(dev, vm->scratch_pdp);
960 free_pd(dev, vm->scratch_pd);
961 free_pt(dev, vm->scratch_pt);
962 free_scratch_page(dev, vm->scratch_page);
965 static void gen8_ppgtt_cleanup_3lvl(struct drm_device *dev,
966 struct i915_page_directory_pointer *pdp)
970 for_each_set_bit(i, pdp->used_pdpes, I915_PDPES_PER_PDP(dev)) {
971 if (WARN_ON(!pdp->page_directory[i]))
974 gen8_free_page_tables(dev, pdp->page_directory[i]);
975 free_pd(dev, pdp->page_directory[i]);
981 static void gen8_ppgtt_cleanup_4lvl(struct i915_hw_ppgtt *ppgtt)
985 for_each_set_bit(i, ppgtt->pml4.used_pml4es, GEN8_PML4ES_PER_PML4) {
986 if (WARN_ON(!ppgtt->pml4.pdps[i]))
989 gen8_ppgtt_cleanup_3lvl(ppgtt->base.dev, ppgtt->pml4.pdps[i]);
992 cleanup_px(ppgtt->base.dev, &ppgtt->pml4);
995 static void gen8_ppgtt_cleanup(struct i915_address_space *vm)
997 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
999 if (intel_vgpu_active(to_i915(vm->dev)))
1000 gen8_ppgtt_notify_vgt(ppgtt, false);
1002 if (!USES_FULL_48BIT_PPGTT(ppgtt->base.dev))
1003 gen8_ppgtt_cleanup_3lvl(ppgtt->base.dev, &ppgtt->pdp);
1005 gen8_ppgtt_cleanup_4lvl(ppgtt);
1007 gen8_free_scratch(vm);
1011 * gen8_ppgtt_alloc_pagetabs() - Allocate page tables for VA range.
1012 * @vm: Master vm structure.
1013 * @pd: Page directory for this address range.
1014 * @start: Starting virtual address to begin allocations.
1015 * @length: Size of the allocations.
1016 * @new_pts: Bitmap set by function with new allocations. Likely used by the
1017 * caller to free on error.
1019 * Allocate the required number of page tables. Extremely similar to
1020 * gen8_ppgtt_alloc_page_directories(). The main difference is here we are limited by
1021 * the page directory boundary (instead of the page directory pointer). That
1022 * boundary is 1GB virtual. Therefore, unlike gen8_ppgtt_alloc_page_directories(), it is
1023 * possible, and likely that the caller will need to use multiple calls of this
1024 * function to achieve the appropriate allocation.
1026 * Return: 0 if success; negative error code otherwise.
1028 static int gen8_ppgtt_alloc_pagetabs(struct i915_address_space *vm,
1029 struct i915_page_directory *pd,
1032 unsigned long *new_pts)
1034 struct drm_device *dev = vm->dev;
1035 struct i915_page_table *pt;
1038 gen8_for_each_pde(pt, pd, start, length, pde) {
1039 /* Don't reallocate page tables */
1040 if (test_bit(pde, pd->used_pdes)) {
1041 /* Scratch is never allocated this way */
1042 WARN_ON(pt == vm->scratch_pt);
1050 gen8_initialize_pt(vm, pt);
1051 pd->page_table[pde] = pt;
1052 __set_bit(pde, new_pts);
1053 trace_i915_page_table_entry_alloc(vm, pde, start, GEN8_PDE_SHIFT);
1059 for_each_set_bit(pde, new_pts, I915_PDES)
1060 free_pt(dev, pd->page_table[pde]);
1066 * gen8_ppgtt_alloc_page_directories() - Allocate page directories for VA range.
1067 * @vm: Master vm structure.
1068 * @pdp: Page directory pointer for this address range.
1069 * @start: Starting virtual address to begin allocations.
1070 * @length: Size of the allocations.
1071 * @new_pds: Bitmap set by function with new allocations. Likely used by the
1072 * caller to free on error.
1074 * Allocate the required number of page directories starting at the pde index of
1075 * @start, and ending at the pde index @start + @length. This function will skip
1076 * over already allocated page directories within the range, and only allocate
1077 * new ones, setting the appropriate pointer within the pdp as well as the
1078 * correct position in the bitmap @new_pds.
1080 * The function will only allocate the pages within the range for a give page
1081 * directory pointer. In other words, if @start + @length straddles a virtually
1082 * addressed PDP boundary (512GB for 4k pages), there will be more allocations
1083 * required by the caller, This is not currently possible, and the BUG in the
1084 * code will prevent it.
1086 * Return: 0 if success; negative error code otherwise.
1089 gen8_ppgtt_alloc_page_directories(struct i915_address_space *vm,
1090 struct i915_page_directory_pointer *pdp,
1093 unsigned long *new_pds)
1095 struct drm_device *dev = vm->dev;
1096 struct i915_page_directory *pd;
1098 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1100 WARN_ON(!bitmap_empty(new_pds, pdpes));
1102 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1103 if (test_bit(pdpe, pdp->used_pdpes))
1110 gen8_initialize_pd(vm, pd);
1111 pdp->page_directory[pdpe] = pd;
1112 __set_bit(pdpe, new_pds);
1113 trace_i915_page_directory_entry_alloc(vm, pdpe, start, GEN8_PDPE_SHIFT);
1119 for_each_set_bit(pdpe, new_pds, pdpes)
1120 free_pd(dev, pdp->page_directory[pdpe]);
1126 * gen8_ppgtt_alloc_page_dirpointers() - Allocate pdps for VA range.
1127 * @vm: Master vm structure.
1128 * @pml4: Page map level 4 for this address range.
1129 * @start: Starting virtual address to begin allocations.
1130 * @length: Size of the allocations.
1131 * @new_pdps: Bitmap set by function with new allocations. Likely used by the
1132 * caller to free on error.
1134 * Allocate the required number of page directory pointers. Extremely similar to
1135 * gen8_ppgtt_alloc_page_directories() and gen8_ppgtt_alloc_pagetabs().
1136 * The main difference is here we are limited by the pml4 boundary (instead of
1137 * the page directory pointer).
1139 * Return: 0 if success; negative error code otherwise.
1142 gen8_ppgtt_alloc_page_dirpointers(struct i915_address_space *vm,
1143 struct i915_pml4 *pml4,
1146 unsigned long *new_pdps)
1148 struct drm_device *dev = vm->dev;
1149 struct i915_page_directory_pointer *pdp;
1152 WARN_ON(!bitmap_empty(new_pdps, GEN8_PML4ES_PER_PML4));
1154 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1155 if (!test_bit(pml4e, pml4->used_pml4es)) {
1156 pdp = alloc_pdp(dev);
1160 gen8_initialize_pdp(vm, pdp);
1161 pml4->pdps[pml4e] = pdp;
1162 __set_bit(pml4e, new_pdps);
1163 trace_i915_page_directory_pointer_entry_alloc(vm,
1173 for_each_set_bit(pml4e, new_pdps, GEN8_PML4ES_PER_PML4)
1174 free_pdp(dev, pml4->pdps[pml4e]);
1180 free_gen8_temp_bitmaps(unsigned long *new_pds, unsigned long *new_pts)
1186 /* Fills in the page directory bitmap, and the array of page tables bitmap. Both
1187 * of these are based on the number of PDPEs in the system.
1190 int __must_check alloc_gen8_temp_bitmaps(unsigned long **new_pds,
1191 unsigned long **new_pts,
1197 pds = kcalloc(BITS_TO_LONGS(pdpes), sizeof(unsigned long), GFP_TEMPORARY);
1201 pts = kcalloc(pdpes, BITS_TO_LONGS(I915_PDES) * sizeof(unsigned long),
1212 free_gen8_temp_bitmaps(pds, pts);
1216 /* PDE TLBs are a pain to invalidate on GEN8+. When we modify
1217 * the page table structures, we mark them dirty so that
1218 * context switching/execlist queuing code takes extra steps
1219 * to ensure that tlbs are flushed.
1221 static void mark_tlbs_dirty(struct i915_hw_ppgtt *ppgtt)
1223 ppgtt->pd_dirty_rings = INTEL_INFO(ppgtt->base.dev)->ring_mask;
1226 static int gen8_alloc_va_range_3lvl(struct i915_address_space *vm,
1227 struct i915_page_directory_pointer *pdp,
1231 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1232 unsigned long *new_page_dirs, *new_page_tables;
1233 struct drm_device *dev = vm->dev;
1234 struct i915_page_directory *pd;
1235 const uint64_t orig_start = start;
1236 const uint64_t orig_length = length;
1238 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1241 /* Wrap is never okay since we can only represent 48b, and we don't
1242 * actually use the other side of the canonical address space.
1244 if (WARN_ON(start + length < start))
1247 if (WARN_ON(start + length > vm->total))
1250 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables, pdpes);
1254 /* Do the allocations first so we can easily bail out */
1255 ret = gen8_ppgtt_alloc_page_directories(vm, pdp, start, length,
1258 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1262 /* For every page directory referenced, allocate page tables */
1263 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1264 ret = gen8_ppgtt_alloc_pagetabs(vm, pd, start, length,
1265 new_page_tables + pdpe * BITS_TO_LONGS(I915_PDES));
1271 length = orig_length;
1273 /* Allocations have completed successfully, so set the bitmaps, and do
1275 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1276 gen8_pde_t *const page_directory = kmap_px(pd);
1277 struct i915_page_table *pt;
1278 uint64_t pd_len = length;
1279 uint64_t pd_start = start;
1282 /* Every pd should be allocated, we just did that above. */
1285 gen8_for_each_pde(pt, pd, pd_start, pd_len, pde) {
1286 /* Same reasoning as pd */
1289 WARN_ON(!gen8_pte_count(pd_start, pd_len));
1291 /* Set our used ptes within the page table */
1292 bitmap_set(pt->used_ptes,
1293 gen8_pte_index(pd_start),
1294 gen8_pte_count(pd_start, pd_len));
1296 /* Our pde is now pointing to the pagetable, pt */
1297 __set_bit(pde, pd->used_pdes);
1299 /* Map the PDE to the page table */
1300 page_directory[pde] = gen8_pde_encode(px_dma(pt),
1302 trace_i915_page_table_entry_map(&ppgtt->base, pde, pt,
1303 gen8_pte_index(start),
1304 gen8_pte_count(start, length),
1307 /* NB: We haven't yet mapped ptes to pages. At this
1308 * point we're still relying on insert_entries() */
1311 kunmap_px(ppgtt, page_directory);
1312 __set_bit(pdpe, pdp->used_pdpes);
1313 gen8_setup_page_directory(ppgtt, pdp, pd, pdpe);
1316 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1317 mark_tlbs_dirty(ppgtt);
1324 for_each_set_bit(temp, new_page_tables + pdpe *
1325 BITS_TO_LONGS(I915_PDES), I915_PDES)
1326 free_pt(dev, pdp->page_directory[pdpe]->page_table[temp]);
1329 for_each_set_bit(pdpe, new_page_dirs, pdpes)
1330 free_pd(dev, pdp->page_directory[pdpe]);
1332 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1333 mark_tlbs_dirty(ppgtt);
1337 static int gen8_alloc_va_range_4lvl(struct i915_address_space *vm,
1338 struct i915_pml4 *pml4,
1342 DECLARE_BITMAP(new_pdps, GEN8_PML4ES_PER_PML4);
1343 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1344 struct i915_page_directory_pointer *pdp;
1348 /* Do the pml4 allocations first, so we don't need to track the newly
1349 * allocated tables below the pdp */
1350 bitmap_zero(new_pdps, GEN8_PML4ES_PER_PML4);
1352 /* The pagedirectory and pagetable allocations are done in the shared 3
1353 * and 4 level code. Just allocate the pdps.
1355 ret = gen8_ppgtt_alloc_page_dirpointers(vm, pml4, start, length,
1360 WARN(bitmap_weight(new_pdps, GEN8_PML4ES_PER_PML4) > 2,
1361 "The allocation has spanned more than 512GB. "
1362 "It is highly likely this is incorrect.");
1364 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1367 ret = gen8_alloc_va_range_3lvl(vm, pdp, start, length);
1371 gen8_setup_page_directory_pointer(ppgtt, pml4, pdp, pml4e);
1374 bitmap_or(pml4->used_pml4es, new_pdps, pml4->used_pml4es,
1375 GEN8_PML4ES_PER_PML4);
1380 for_each_set_bit(pml4e, new_pdps, GEN8_PML4ES_PER_PML4)
1381 gen8_ppgtt_cleanup_3lvl(vm->dev, pml4->pdps[pml4e]);
1386 static int gen8_alloc_va_range(struct i915_address_space *vm,
1387 uint64_t start, uint64_t length)
1389 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1391 if (USES_FULL_48BIT_PPGTT(vm->dev))
1392 return gen8_alloc_va_range_4lvl(vm, &ppgtt->pml4, start, length);
1394 return gen8_alloc_va_range_3lvl(vm, &ppgtt->pdp, start, length);
1397 static void gen8_dump_pdp(struct i915_page_directory_pointer *pdp,
1398 uint64_t start, uint64_t length,
1399 gen8_pte_t scratch_pte,
1402 struct i915_page_directory *pd;
1405 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1406 struct i915_page_table *pt;
1407 uint64_t pd_len = length;
1408 uint64_t pd_start = start;
1411 if (!test_bit(pdpe, pdp->used_pdpes))
1414 seq_printf(m, "\tPDPE #%d\n", pdpe);
1415 gen8_for_each_pde(pt, pd, pd_start, pd_len, pde) {
1417 gen8_pte_t *pt_vaddr;
1419 if (!test_bit(pde, pd->used_pdes))
1422 pt_vaddr = kmap_px(pt);
1423 for (pte = 0; pte < GEN8_PTES; pte += 4) {
1425 (pdpe << GEN8_PDPE_SHIFT) |
1426 (pde << GEN8_PDE_SHIFT) |
1427 (pte << GEN8_PTE_SHIFT);
1431 for (i = 0; i < 4; i++)
1432 if (pt_vaddr[pte + i] != scratch_pte)
1437 seq_printf(m, "\t\t0x%llx [%03d,%03d,%04d]: =", va, pdpe, pde, pte);
1438 for (i = 0; i < 4; i++) {
1439 if (pt_vaddr[pte + i] != scratch_pte)
1440 seq_printf(m, " %llx", pt_vaddr[pte + i]);
1442 seq_puts(m, " SCRATCH ");
1446 /* don't use kunmap_px, it could trigger
1447 * an unnecessary flush.
1449 kunmap_atomic(pt_vaddr);
1454 static void gen8_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
1456 struct i915_address_space *vm = &ppgtt->base;
1457 uint64_t start = ppgtt->base.start;
1458 uint64_t length = ppgtt->base.total;
1459 gen8_pte_t scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
1460 I915_CACHE_LLC, true);
1462 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
1463 gen8_dump_pdp(&ppgtt->pdp, start, length, scratch_pte, m);
1466 struct i915_pml4 *pml4 = &ppgtt->pml4;
1467 struct i915_page_directory_pointer *pdp;
1469 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1470 if (!test_bit(pml4e, pml4->used_pml4es))
1473 seq_printf(m, " PML4E #%llu\n", pml4e);
1474 gen8_dump_pdp(pdp, start, length, scratch_pte, m);
1479 static int gen8_preallocate_top_level_pdps(struct i915_hw_ppgtt *ppgtt)
1481 unsigned long *new_page_dirs, *new_page_tables;
1482 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1485 /* We allocate temp bitmap for page tables for no gain
1486 * but as this is for init only, lets keep the things simple
1488 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables, pdpes);
1492 /* Allocate for all pdps regardless of how the ppgtt
1495 ret = gen8_ppgtt_alloc_page_directories(&ppgtt->base, &ppgtt->pdp,
1499 *ppgtt->pdp.used_pdpes = *new_page_dirs;
1501 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1507 * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers
1508 * with a net effect resembling a 2-level page table in normal x86 terms. Each
1509 * PDP represents 1GB of memory 4 * 512 * 512 * 4096 = 4GB legacy 32b address
1513 static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
1517 ret = gen8_init_scratch(&ppgtt->base);
1521 ppgtt->base.start = 0;
1522 ppgtt->base.cleanup = gen8_ppgtt_cleanup;
1523 ppgtt->base.allocate_va_range = gen8_alloc_va_range;
1524 ppgtt->base.insert_entries = gen8_ppgtt_insert_entries;
1525 ppgtt->base.clear_range = gen8_ppgtt_clear_range;
1526 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
1527 ppgtt->base.bind_vma = ppgtt_bind_vma;
1528 ppgtt->debug_dump = gen8_dump_ppgtt;
1530 if (USES_FULL_48BIT_PPGTT(ppgtt->base.dev)) {
1531 ret = setup_px(ppgtt->base.dev, &ppgtt->pml4);
1535 gen8_initialize_pml4(&ppgtt->base, &ppgtt->pml4);
1537 ppgtt->base.total = 1ULL << 48;
1538 ppgtt->switch_mm = gen8_48b_mm_switch;
1540 ret = __pdp_init(ppgtt->base.dev, &ppgtt->pdp);
1544 ppgtt->base.total = 1ULL << 32;
1545 ppgtt->switch_mm = gen8_legacy_mm_switch;
1546 trace_i915_page_directory_pointer_entry_alloc(&ppgtt->base,
1550 if (intel_vgpu_active(to_i915(ppgtt->base.dev))) {
1551 ret = gen8_preallocate_top_level_pdps(ppgtt);
1557 if (intel_vgpu_active(to_i915(ppgtt->base.dev)))
1558 gen8_ppgtt_notify_vgt(ppgtt, true);
1563 gen8_free_scratch(&ppgtt->base);
1567 static void gen6_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
1569 struct i915_address_space *vm = &ppgtt->base;
1570 struct i915_page_table *unused;
1571 gen6_pte_t scratch_pte;
1574 uint32_t start = ppgtt->base.start, length = ppgtt->base.total;
1576 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1577 I915_CACHE_LLC, true, 0);
1579 gen6_for_each_pde(unused, &ppgtt->pd, start, length, pde) {
1581 gen6_pte_t *pt_vaddr;
1582 const dma_addr_t pt_addr = px_dma(ppgtt->pd.page_table[pde]);
1583 pd_entry = readl(ppgtt->pd_addr + pde);
1584 expected = (GEN6_PDE_ADDR_ENCODE(pt_addr) | GEN6_PDE_VALID);
1586 if (pd_entry != expected)
1587 seq_printf(m, "\tPDE #%d mismatch: Actual PDE: %x Expected PDE: %x\n",
1591 seq_printf(m, "\tPDE: %x\n", pd_entry);
1593 pt_vaddr = kmap_px(ppgtt->pd.page_table[pde]);
1595 for (pte = 0; pte < GEN6_PTES; pte+=4) {
1597 (pde * PAGE_SIZE * GEN6_PTES) +
1601 for (i = 0; i < 4; i++)
1602 if (pt_vaddr[pte + i] != scratch_pte)
1607 seq_printf(m, "\t\t0x%lx [%03d,%04d]: =", va, pde, pte);
1608 for (i = 0; i < 4; i++) {
1609 if (pt_vaddr[pte + i] != scratch_pte)
1610 seq_printf(m, " %08x", pt_vaddr[pte + i]);
1612 seq_puts(m, " SCRATCH ");
1616 kunmap_px(ppgtt, pt_vaddr);
1620 /* Write pde (index) from the page directory @pd to the page table @pt */
1621 static void gen6_write_pde(struct i915_page_directory *pd,
1622 const int pde, struct i915_page_table *pt)
1624 /* Caller needs to make sure the write completes if necessary */
1625 struct i915_hw_ppgtt *ppgtt =
1626 container_of(pd, struct i915_hw_ppgtt, pd);
1629 pd_entry = GEN6_PDE_ADDR_ENCODE(px_dma(pt));
1630 pd_entry |= GEN6_PDE_VALID;
1632 writel(pd_entry, ppgtt->pd_addr + pde);
1635 /* Write all the page tables found in the ppgtt structure to incrementing page
1637 static void gen6_write_page_range(struct drm_i915_private *dev_priv,
1638 struct i915_page_directory *pd,
1639 uint32_t start, uint32_t length)
1641 struct i915_ggtt *ggtt = &dev_priv->ggtt;
1642 struct i915_page_table *pt;
1645 gen6_for_each_pde(pt, pd, start, length, pde)
1646 gen6_write_pde(pd, pde, pt);
1648 /* Make sure write is complete before other code can use this page
1649 * table. Also require for WC mapped PTEs */
1653 static uint32_t get_pd_offset(struct i915_hw_ppgtt *ppgtt)
1655 BUG_ON(ppgtt->pd.base.ggtt_offset & 0x3f);
1657 return (ppgtt->pd.base.ggtt_offset / 64) << 16;
1660 static int hsw_mm_switch(struct i915_hw_ppgtt *ppgtt,
1661 struct drm_i915_gem_request *req)
1663 struct intel_engine_cs *engine = req->engine;
1666 /* NB: TLBs must be flushed and invalidated before a switch */
1667 ret = engine->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1671 ret = intel_ring_begin(req, 6);
1675 intel_ring_emit(engine, MI_LOAD_REGISTER_IMM(2));
1676 intel_ring_emit_reg(engine, RING_PP_DIR_DCLV(engine));
1677 intel_ring_emit(engine, PP_DIR_DCLV_2G);
1678 intel_ring_emit_reg(engine, RING_PP_DIR_BASE(engine));
1679 intel_ring_emit(engine, get_pd_offset(ppgtt));
1680 intel_ring_emit(engine, MI_NOOP);
1681 intel_ring_advance(engine);
1686 static int vgpu_mm_switch(struct i915_hw_ppgtt *ppgtt,
1687 struct drm_i915_gem_request *req)
1689 struct intel_engine_cs *engine = req->engine;
1690 struct drm_i915_private *dev_priv = to_i915(ppgtt->base.dev);
1692 I915_WRITE(RING_PP_DIR_DCLV(engine), PP_DIR_DCLV_2G);
1693 I915_WRITE(RING_PP_DIR_BASE(engine), get_pd_offset(ppgtt));
1697 static int gen7_mm_switch(struct i915_hw_ppgtt *ppgtt,
1698 struct drm_i915_gem_request *req)
1700 struct intel_engine_cs *engine = req->engine;
1703 /* NB: TLBs must be flushed and invalidated before a switch */
1704 ret = engine->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1708 ret = intel_ring_begin(req, 6);
1712 intel_ring_emit(engine, MI_LOAD_REGISTER_IMM(2));
1713 intel_ring_emit_reg(engine, RING_PP_DIR_DCLV(engine));
1714 intel_ring_emit(engine, PP_DIR_DCLV_2G);
1715 intel_ring_emit_reg(engine, RING_PP_DIR_BASE(engine));
1716 intel_ring_emit(engine, get_pd_offset(ppgtt));
1717 intel_ring_emit(engine, MI_NOOP);
1718 intel_ring_advance(engine);
1720 /* XXX: RCS is the only one to auto invalidate the TLBs? */
1721 if (engine->id != RCS) {
1722 ret = engine->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1730 static int gen6_mm_switch(struct i915_hw_ppgtt *ppgtt,
1731 struct drm_i915_gem_request *req)
1733 struct intel_engine_cs *engine = req->engine;
1734 struct drm_device *dev = ppgtt->base.dev;
1735 struct drm_i915_private *dev_priv = to_i915(dev);
1738 I915_WRITE(RING_PP_DIR_DCLV(engine), PP_DIR_DCLV_2G);
1739 I915_WRITE(RING_PP_DIR_BASE(engine), get_pd_offset(ppgtt));
1741 POSTING_READ(RING_PP_DIR_DCLV(engine));
1746 static void gen8_ppgtt_enable(struct drm_device *dev)
1748 struct drm_i915_private *dev_priv = to_i915(dev);
1749 struct intel_engine_cs *engine;
1751 for_each_engine(engine, dev_priv) {
1752 u32 four_level = USES_FULL_48BIT_PPGTT(dev) ? GEN8_GFX_PPGTT_48B : 0;
1753 I915_WRITE(RING_MODE_GEN7(engine),
1754 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE | four_level));
1758 static void gen7_ppgtt_enable(struct drm_device *dev)
1760 struct drm_i915_private *dev_priv = to_i915(dev);
1761 struct intel_engine_cs *engine;
1762 uint32_t ecochk, ecobits;
1764 ecobits = I915_READ(GAC_ECO_BITS);
1765 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B);
1767 ecochk = I915_READ(GAM_ECOCHK);
1768 if (IS_HASWELL(dev)) {
1769 ecochk |= ECOCHK_PPGTT_WB_HSW;
1771 ecochk |= ECOCHK_PPGTT_LLC_IVB;
1772 ecochk &= ~ECOCHK_PPGTT_GFDT_IVB;
1774 I915_WRITE(GAM_ECOCHK, ecochk);
1776 for_each_engine(engine, dev_priv) {
1777 /* GFX_MODE is per-ring on gen7+ */
1778 I915_WRITE(RING_MODE_GEN7(engine),
1779 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1783 static void gen6_ppgtt_enable(struct drm_device *dev)
1785 struct drm_i915_private *dev_priv = to_i915(dev);
1786 uint32_t ecochk, gab_ctl, ecobits;
1788 ecobits = I915_READ(GAC_ECO_BITS);
1789 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_SNB_BIT |
1790 ECOBITS_PPGTT_CACHE64B);
1792 gab_ctl = I915_READ(GAB_CTL);
1793 I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT);
1795 ecochk = I915_READ(GAM_ECOCHK);
1796 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT | ECOCHK_PPGTT_CACHE64B);
1798 I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1801 /* PPGTT support for Sandybdrige/Gen6 and later */
1802 static void gen6_ppgtt_clear_range(struct i915_address_space *vm,
1807 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1808 gen6_pte_t *pt_vaddr, scratch_pte;
1809 unsigned first_entry = start >> PAGE_SHIFT;
1810 unsigned num_entries = length >> PAGE_SHIFT;
1811 unsigned act_pt = first_entry / GEN6_PTES;
1812 unsigned first_pte = first_entry % GEN6_PTES;
1813 unsigned last_pte, i;
1815 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1816 I915_CACHE_LLC, true, 0);
1818 while (num_entries) {
1819 last_pte = first_pte + num_entries;
1820 if (last_pte > GEN6_PTES)
1821 last_pte = GEN6_PTES;
1823 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1825 for (i = first_pte; i < last_pte; i++)
1826 pt_vaddr[i] = scratch_pte;
1828 kunmap_px(ppgtt, pt_vaddr);
1830 num_entries -= last_pte - first_pte;
1836 static void gen6_ppgtt_insert_entries(struct i915_address_space *vm,
1837 struct sg_table *pages,
1839 enum i915_cache_level cache_level, u32 flags)
1841 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1842 unsigned first_entry = start >> PAGE_SHIFT;
1843 unsigned act_pt = first_entry / GEN6_PTES;
1844 unsigned act_pte = first_entry % GEN6_PTES;
1845 gen6_pte_t *pt_vaddr = NULL;
1846 struct sgt_iter sgt_iter;
1849 for_each_sgt_dma(addr, sgt_iter, pages) {
1850 if (pt_vaddr == NULL)
1851 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1854 vm->pte_encode(addr, cache_level, true, flags);
1856 if (++act_pte == GEN6_PTES) {
1857 kunmap_px(ppgtt, pt_vaddr);
1865 kunmap_px(ppgtt, pt_vaddr);
1868 static int gen6_alloc_va_range(struct i915_address_space *vm,
1869 uint64_t start_in, uint64_t length_in)
1871 DECLARE_BITMAP(new_page_tables, I915_PDES);
1872 struct drm_device *dev = vm->dev;
1873 struct drm_i915_private *dev_priv = to_i915(dev);
1874 struct i915_ggtt *ggtt = &dev_priv->ggtt;
1875 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1876 struct i915_page_table *pt;
1877 uint32_t start, length, start_save, length_save;
1881 if (WARN_ON(start_in + length_in > ppgtt->base.total))
1884 start = start_save = start_in;
1885 length = length_save = length_in;
1887 bitmap_zero(new_page_tables, I915_PDES);
1889 /* The allocation is done in two stages so that we can bail out with
1890 * minimal amount of pain. The first stage finds new page tables that
1891 * need allocation. The second stage marks use ptes within the page
1894 gen6_for_each_pde(pt, &ppgtt->pd, start, length, pde) {
1895 if (pt != vm->scratch_pt) {
1896 WARN_ON(bitmap_empty(pt->used_ptes, GEN6_PTES));
1900 /* We've already allocated a page table */
1901 WARN_ON(!bitmap_empty(pt->used_ptes, GEN6_PTES));
1909 gen6_initialize_pt(vm, pt);
1911 ppgtt->pd.page_table[pde] = pt;
1912 __set_bit(pde, new_page_tables);
1913 trace_i915_page_table_entry_alloc(vm, pde, start, GEN6_PDE_SHIFT);
1917 length = length_save;
1919 gen6_for_each_pde(pt, &ppgtt->pd, start, length, pde) {
1920 DECLARE_BITMAP(tmp_bitmap, GEN6_PTES);
1922 bitmap_zero(tmp_bitmap, GEN6_PTES);
1923 bitmap_set(tmp_bitmap, gen6_pte_index(start),
1924 gen6_pte_count(start, length));
1926 if (__test_and_clear_bit(pde, new_page_tables))
1927 gen6_write_pde(&ppgtt->pd, pde, pt);
1929 trace_i915_page_table_entry_map(vm, pde, pt,
1930 gen6_pte_index(start),
1931 gen6_pte_count(start, length),
1933 bitmap_or(pt->used_ptes, tmp_bitmap, pt->used_ptes,
1937 WARN_ON(!bitmap_empty(new_page_tables, I915_PDES));
1939 /* Make sure write is complete before other code can use this page
1940 * table. Also require for WC mapped PTEs */
1943 mark_tlbs_dirty(ppgtt);
1947 for_each_set_bit(pde, new_page_tables, I915_PDES) {
1948 struct i915_page_table *pt = ppgtt->pd.page_table[pde];
1950 ppgtt->pd.page_table[pde] = vm->scratch_pt;
1951 free_pt(vm->dev, pt);
1954 mark_tlbs_dirty(ppgtt);
1958 static int gen6_init_scratch(struct i915_address_space *vm)
1960 struct drm_device *dev = vm->dev;
1962 vm->scratch_page = alloc_scratch_page(dev);
1963 if (IS_ERR(vm->scratch_page))
1964 return PTR_ERR(vm->scratch_page);
1966 vm->scratch_pt = alloc_pt(dev);
1967 if (IS_ERR(vm->scratch_pt)) {
1968 free_scratch_page(dev, vm->scratch_page);
1969 return PTR_ERR(vm->scratch_pt);
1972 gen6_initialize_pt(vm, vm->scratch_pt);
1977 static void gen6_free_scratch(struct i915_address_space *vm)
1979 struct drm_device *dev = vm->dev;
1981 free_pt(dev, vm->scratch_pt);
1982 free_scratch_page(dev, vm->scratch_page);
1985 static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
1987 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1988 struct i915_page_directory *pd = &ppgtt->pd;
1989 struct drm_device *dev = vm->dev;
1990 struct i915_page_table *pt;
1993 drm_mm_remove_node(&ppgtt->node);
1995 gen6_for_all_pdes(pt, pd, pde)
1996 if (pt != vm->scratch_pt)
1999 gen6_free_scratch(vm);
2002 static int gen6_ppgtt_allocate_page_directories(struct i915_hw_ppgtt *ppgtt)
2004 struct i915_address_space *vm = &ppgtt->base;
2005 struct drm_device *dev = ppgtt->base.dev;
2006 struct drm_i915_private *dev_priv = to_i915(dev);
2007 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2008 bool retried = false;
2011 /* PPGTT PDEs reside in the GGTT and consists of 512 entries. The
2012 * allocator works in address space sizes, so it's multiplied by page
2013 * size. We allocate at the top of the GTT to avoid fragmentation.
2015 BUG_ON(!drm_mm_initialized(&ggtt->base.mm));
2017 ret = gen6_init_scratch(vm);
2022 ret = drm_mm_insert_node_in_range_generic(&ggtt->base.mm,
2023 &ppgtt->node, GEN6_PD_SIZE,
2025 0, ggtt->base.total,
2027 if (ret == -ENOSPC && !retried) {
2028 ret = i915_gem_evict_something(dev, &ggtt->base,
2029 GEN6_PD_SIZE, GEN6_PD_ALIGN,
2031 0, ggtt->base.total,
2044 if (ppgtt->node.start < ggtt->mappable_end)
2045 DRM_DEBUG("Forced to use aperture for PDEs\n");
2050 gen6_free_scratch(vm);
2054 static int gen6_ppgtt_alloc(struct i915_hw_ppgtt *ppgtt)
2056 return gen6_ppgtt_allocate_page_directories(ppgtt);
2059 static void gen6_scratch_va_range(struct i915_hw_ppgtt *ppgtt,
2060 uint64_t start, uint64_t length)
2062 struct i915_page_table *unused;
2065 gen6_for_each_pde(unused, &ppgtt->pd, start, length, pde)
2066 ppgtt->pd.page_table[pde] = ppgtt->base.scratch_pt;
2069 static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
2071 struct drm_device *dev = ppgtt->base.dev;
2072 struct drm_i915_private *dev_priv = to_i915(dev);
2073 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2076 ppgtt->base.pte_encode = ggtt->base.pte_encode;
2078 ppgtt->switch_mm = gen6_mm_switch;
2079 } else if (IS_HASWELL(dev)) {
2080 ppgtt->switch_mm = hsw_mm_switch;
2081 } else if (IS_GEN7(dev)) {
2082 ppgtt->switch_mm = gen7_mm_switch;
2086 if (intel_vgpu_active(dev_priv))
2087 ppgtt->switch_mm = vgpu_mm_switch;
2089 ret = gen6_ppgtt_alloc(ppgtt);
2093 ppgtt->base.allocate_va_range = gen6_alloc_va_range;
2094 ppgtt->base.clear_range = gen6_ppgtt_clear_range;
2095 ppgtt->base.insert_entries = gen6_ppgtt_insert_entries;
2096 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
2097 ppgtt->base.bind_vma = ppgtt_bind_vma;
2098 ppgtt->base.cleanup = gen6_ppgtt_cleanup;
2099 ppgtt->base.start = 0;
2100 ppgtt->base.total = I915_PDES * GEN6_PTES * PAGE_SIZE;
2101 ppgtt->debug_dump = gen6_dump_ppgtt;
2103 ppgtt->pd.base.ggtt_offset =
2104 ppgtt->node.start / PAGE_SIZE * sizeof(gen6_pte_t);
2106 ppgtt->pd_addr = (gen6_pte_t __iomem *)ggtt->gsm +
2107 ppgtt->pd.base.ggtt_offset / sizeof(gen6_pte_t);
2109 gen6_scratch_va_range(ppgtt, 0, ppgtt->base.total);
2111 gen6_write_page_range(dev_priv, &ppgtt->pd, 0, ppgtt->base.total);
2113 DRM_DEBUG_DRIVER("Allocated pde space (%lldM) at GTT entry: %llx\n",
2114 ppgtt->node.size >> 20,
2115 ppgtt->node.start / PAGE_SIZE);
2117 DRM_DEBUG("Adding PPGTT at offset %x\n",
2118 ppgtt->pd.base.ggtt_offset << 10);
2123 static int __hw_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
2125 ppgtt->base.dev = dev;
2127 if (INTEL_INFO(dev)->gen < 8)
2128 return gen6_ppgtt_init(ppgtt);
2130 return gen8_ppgtt_init(ppgtt);
2133 static void i915_address_space_init(struct i915_address_space *vm,
2134 struct drm_i915_private *dev_priv)
2136 drm_mm_init(&vm->mm, vm->start, vm->total);
2137 vm->dev = dev_priv->dev;
2138 INIT_LIST_HEAD(&vm->active_list);
2139 INIT_LIST_HEAD(&vm->inactive_list);
2140 list_add_tail(&vm->global_link, &dev_priv->vm_list);
2143 static void gtt_write_workarounds(struct drm_device *dev)
2145 struct drm_i915_private *dev_priv = to_i915(dev);
2147 /* This function is for gtt related workarounds. This function is
2148 * called on driver load and after a GPU reset, so you can place
2149 * workarounds here even if they get overwritten by GPU reset.
2151 /* WaIncreaseDefaultTLBEntries:chv,bdw,skl,bxt */
2152 if (IS_BROADWELL(dev))
2153 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN8_L3_LRA_1_GPGPU_DEFAULT_VALUE_BDW);
2154 else if (IS_CHERRYVIEW(dev))
2155 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN8_L3_LRA_1_GPGPU_DEFAULT_VALUE_CHV);
2156 else if (IS_SKYLAKE(dev))
2157 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN9_L3_LRA_1_GPGPU_DEFAULT_VALUE_SKL);
2158 else if (IS_BROXTON(dev))
2159 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN9_L3_LRA_1_GPGPU_DEFAULT_VALUE_BXT);
2162 static int i915_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
2164 struct drm_i915_private *dev_priv = to_i915(dev);
2167 ret = __hw_ppgtt_init(dev, ppgtt);
2169 kref_init(&ppgtt->ref);
2170 i915_address_space_init(&ppgtt->base, dev_priv);
2176 int i915_ppgtt_init_hw(struct drm_device *dev)
2178 gtt_write_workarounds(dev);
2180 /* In the case of execlists, PPGTT is enabled by the context descriptor
2181 * and the PDPs are contained within the context itself. We don't
2182 * need to do anything here. */
2183 if (i915.enable_execlists)
2186 if (!USES_PPGTT(dev))
2190 gen6_ppgtt_enable(dev);
2191 else if (IS_GEN7(dev))
2192 gen7_ppgtt_enable(dev);
2193 else if (INTEL_INFO(dev)->gen >= 8)
2194 gen8_ppgtt_enable(dev);
2196 MISSING_CASE(INTEL_INFO(dev)->gen);
2201 struct i915_hw_ppgtt *
2202 i915_ppgtt_create(struct drm_device *dev, struct drm_i915_file_private *fpriv)
2204 struct i915_hw_ppgtt *ppgtt;
2207 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2209 return ERR_PTR(-ENOMEM);
2211 ret = i915_ppgtt_init(dev, ppgtt);
2214 return ERR_PTR(ret);
2217 ppgtt->file_priv = fpriv;
2219 trace_i915_ppgtt_create(&ppgtt->base);
2224 void i915_ppgtt_release(struct kref *kref)
2226 struct i915_hw_ppgtt *ppgtt =
2227 container_of(kref, struct i915_hw_ppgtt, ref);
2229 trace_i915_ppgtt_release(&ppgtt->base);
2231 /* vmas should already be unbound */
2232 WARN_ON(!list_empty(&ppgtt->base.active_list));
2233 WARN_ON(!list_empty(&ppgtt->base.inactive_list));
2235 list_del(&ppgtt->base.global_link);
2236 drm_mm_takedown(&ppgtt->base.mm);
2238 ppgtt->base.cleanup(&ppgtt->base);
2242 extern int intel_iommu_gfx_mapped;
2243 /* Certain Gen5 chipsets require require idling the GPU before
2244 * unmapping anything from the GTT when VT-d is enabled.
2246 static bool needs_idle_maps(struct drm_device *dev)
2248 #ifdef CONFIG_INTEL_IOMMU
2249 /* Query intel_iommu to see if we need the workaround. Presumably that
2252 if (IS_GEN5(dev) && IS_MOBILE(dev) && intel_iommu_gfx_mapped)
2258 static bool do_idling(struct drm_i915_private *dev_priv)
2260 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2261 bool ret = dev_priv->mm.interruptible;
2263 if (unlikely(ggtt->do_idle_maps)) {
2264 dev_priv->mm.interruptible = false;
2265 if (i915_gem_wait_for_idle(dev_priv)) {
2266 DRM_ERROR("Failed to wait for idle; VT'd may hang.\n");
2267 /* Wait a bit, in hopes it avoids the hang */
2275 static void undo_idling(struct drm_i915_private *dev_priv, bool interruptible)
2277 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2279 if (unlikely(ggtt->do_idle_maps))
2280 dev_priv->mm.interruptible = interruptible;
2283 void i915_check_and_clear_faults(struct drm_i915_private *dev_priv)
2285 struct intel_engine_cs *engine;
2287 if (INTEL_INFO(dev_priv)->gen < 6)
2290 for_each_engine(engine, dev_priv) {
2292 fault_reg = I915_READ(RING_FAULT_REG(engine));
2293 if (fault_reg & RING_FAULT_VALID) {
2294 DRM_DEBUG_DRIVER("Unexpected fault\n"
2296 "\tAddress space: %s\n"
2299 fault_reg & PAGE_MASK,
2300 fault_reg & RING_FAULT_GTTSEL_MASK ? "GGTT" : "PPGTT",
2301 RING_FAULT_SRCID(fault_reg),
2302 RING_FAULT_FAULT_TYPE(fault_reg));
2303 I915_WRITE(RING_FAULT_REG(engine),
2304 fault_reg & ~RING_FAULT_VALID);
2307 POSTING_READ(RING_FAULT_REG(&dev_priv->engine[RCS]));
2310 static void i915_ggtt_flush(struct drm_i915_private *dev_priv)
2312 if (INTEL_INFO(dev_priv)->gen < 6) {
2313 intel_gtt_chipset_flush();
2315 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2316 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2320 void i915_gem_suspend_gtt_mappings(struct drm_device *dev)
2322 struct drm_i915_private *dev_priv = to_i915(dev);
2323 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2325 /* Don't bother messing with faults pre GEN6 as we have little
2326 * documentation supporting that it's a good idea.
2328 if (INTEL_INFO(dev)->gen < 6)
2331 i915_check_and_clear_faults(dev_priv);
2333 ggtt->base.clear_range(&ggtt->base, ggtt->base.start, ggtt->base.total,
2336 i915_ggtt_flush(dev_priv);
2339 int i915_gem_gtt_prepare_object(struct drm_i915_gem_object *obj)
2341 if (!dma_map_sg(&obj->base.dev->pdev->dev,
2342 obj->pages->sgl, obj->pages->nents,
2343 PCI_DMA_BIDIRECTIONAL))
2349 static void gen8_set_pte(void __iomem *addr, gen8_pte_t pte)
2354 iowrite32((u32)pte, addr);
2355 iowrite32(pte >> 32, addr + 4);
2359 static void gen8_ggtt_insert_page(struct i915_address_space *vm,
2362 enum i915_cache_level level,
2365 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2366 gen8_pte_t __iomem *pte =
2367 (gen8_pte_t __iomem *)dev_priv->ggtt.gsm +
2368 (offset >> PAGE_SHIFT);
2371 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2373 gen8_set_pte(pte, gen8_pte_encode(addr, level, true));
2375 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2376 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2378 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2381 static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
2382 struct sg_table *st,
2384 enum i915_cache_level level, u32 unused)
2386 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2387 struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
2388 struct sgt_iter sgt_iter;
2389 gen8_pte_t __iomem *gtt_entries;
2390 gen8_pte_t gtt_entry;
2395 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2397 gtt_entries = (gen8_pte_t __iomem *)ggtt->gsm + (start >> PAGE_SHIFT);
2399 for_each_sgt_dma(addr, sgt_iter, st) {
2400 gtt_entry = gen8_pte_encode(addr, level, true);
2401 gen8_set_pte(>t_entries[i++], gtt_entry);
2405 * XXX: This serves as a posting read to make sure that the PTE has
2406 * actually been updated. There is some concern that even though
2407 * registers and PTEs are within the same BAR that they are potentially
2408 * of NUMA access patterns. Therefore, even with the way we assume
2409 * hardware should work, we must keep this posting read for paranoia.
2412 WARN_ON(readq(>t_entries[i-1]) != gtt_entry);
2414 /* This next bit makes the above posting read even more important. We
2415 * want to flush the TLBs only after we're certain all the PTE updates
2418 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2419 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2421 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2424 struct insert_entries {
2425 struct i915_address_space *vm;
2426 struct sg_table *st;
2428 enum i915_cache_level level;
2432 static int gen8_ggtt_insert_entries__cb(void *_arg)
2434 struct insert_entries *arg = _arg;
2435 gen8_ggtt_insert_entries(arg->vm, arg->st,
2436 arg->start, arg->level, arg->flags);
2440 static void gen8_ggtt_insert_entries__BKL(struct i915_address_space *vm,
2441 struct sg_table *st,
2443 enum i915_cache_level level,
2446 struct insert_entries arg = { vm, st, start, level, flags };
2447 stop_machine(gen8_ggtt_insert_entries__cb, &arg, NULL);
2450 static void gen6_ggtt_insert_page(struct i915_address_space *vm,
2453 enum i915_cache_level level,
2456 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2457 gen6_pte_t __iomem *pte =
2458 (gen6_pte_t __iomem *)dev_priv->ggtt.gsm +
2459 (offset >> PAGE_SHIFT);
2462 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2464 iowrite32(vm->pte_encode(addr, level, true, flags), pte);
2466 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2467 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2469 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2473 * Binds an object into the global gtt with the specified cache level. The object
2474 * will be accessible to the GPU via commands whose operands reference offsets
2475 * within the global GTT as well as accessible by the GPU through the GMADR
2476 * mapped BAR (dev_priv->mm.gtt->gtt).
2478 static void gen6_ggtt_insert_entries(struct i915_address_space *vm,
2479 struct sg_table *st,
2481 enum i915_cache_level level, u32 flags)
2483 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2484 struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
2485 struct sgt_iter sgt_iter;
2486 gen6_pte_t __iomem *gtt_entries;
2487 gen6_pte_t gtt_entry;
2492 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2494 gtt_entries = (gen6_pte_t __iomem *)ggtt->gsm + (start >> PAGE_SHIFT);
2496 for_each_sgt_dma(addr, sgt_iter, st) {
2497 gtt_entry = vm->pte_encode(addr, level, true, flags);
2498 iowrite32(gtt_entry, >t_entries[i++]);
2501 /* XXX: This serves as a posting read to make sure that the PTE has
2502 * actually been updated. There is some concern that even though
2503 * registers and PTEs are within the same BAR that they are potentially
2504 * of NUMA access patterns. Therefore, even with the way we assume
2505 * hardware should work, we must keep this posting read for paranoia.
2508 WARN_ON(readl(>t_entries[i-1]) != gtt_entry);
2510 /* This next bit makes the above posting read even more important. We
2511 * want to flush the TLBs only after we're certain all the PTE updates
2514 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2515 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2517 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2520 static void nop_clear_range(struct i915_address_space *vm,
2527 static void gen8_ggtt_clear_range(struct i915_address_space *vm,
2532 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2533 struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
2534 unsigned first_entry = start >> PAGE_SHIFT;
2535 unsigned num_entries = length >> PAGE_SHIFT;
2536 gen8_pte_t scratch_pte, __iomem *gtt_base =
2537 (gen8_pte_t __iomem *)ggtt->gsm + first_entry;
2538 const int max_entries = ggtt_total_entries(ggtt) - first_entry;
2542 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2544 if (WARN(num_entries > max_entries,
2545 "First entry = %d; Num entries = %d (max=%d)\n",
2546 first_entry, num_entries, max_entries))
2547 num_entries = max_entries;
2549 scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
2552 for (i = 0; i < num_entries; i++)
2553 gen8_set_pte(>t_base[i], scratch_pte);
2556 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2559 static void gen6_ggtt_clear_range(struct i915_address_space *vm,
2564 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2565 struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
2566 unsigned first_entry = start >> PAGE_SHIFT;
2567 unsigned num_entries = length >> PAGE_SHIFT;
2568 gen6_pte_t scratch_pte, __iomem *gtt_base =
2569 (gen6_pte_t __iomem *)ggtt->gsm + first_entry;
2570 const int max_entries = ggtt_total_entries(ggtt) - first_entry;
2574 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2576 if (WARN(num_entries > max_entries,
2577 "First entry = %d; Num entries = %d (max=%d)\n",
2578 first_entry, num_entries, max_entries))
2579 num_entries = max_entries;
2581 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
2582 I915_CACHE_LLC, use_scratch, 0);
2584 for (i = 0; i < num_entries; i++)
2585 iowrite32(scratch_pte, >t_base[i]);
2588 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2591 static void i915_ggtt_insert_page(struct i915_address_space *vm,
2594 enum i915_cache_level cache_level,
2597 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2598 unsigned int flags = (cache_level == I915_CACHE_NONE) ?
2599 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
2602 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2604 intel_gtt_insert_page(addr, offset >> PAGE_SHIFT, flags);
2606 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2609 static void i915_ggtt_insert_entries(struct i915_address_space *vm,
2610 struct sg_table *pages,
2612 enum i915_cache_level cache_level, u32 unused)
2614 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2615 unsigned int flags = (cache_level == I915_CACHE_NONE) ?
2616 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
2619 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2621 intel_gtt_insert_sg_entries(pages, start >> PAGE_SHIFT, flags);
2623 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2627 static void i915_ggtt_clear_range(struct i915_address_space *vm,
2632 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2633 unsigned first_entry = start >> PAGE_SHIFT;
2634 unsigned num_entries = length >> PAGE_SHIFT;
2637 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2639 intel_gtt_clear_range(first_entry, num_entries);
2641 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2644 static int ggtt_bind_vma(struct i915_vma *vma,
2645 enum i915_cache_level cache_level,
2648 struct drm_i915_gem_object *obj = vma->obj;
2652 ret = i915_get_ggtt_vma_pages(vma);
2656 /* Currently applicable only to VLV */
2658 pte_flags |= PTE_READ_ONLY;
2660 vma->vm->insert_entries(vma->vm, vma->ggtt_view.pages,
2662 cache_level, pte_flags);
2665 * Without aliasing PPGTT there's no difference between
2666 * GLOBAL/LOCAL_BIND, it's all the same ptes. Hence unconditionally
2667 * upgrade to both bound if we bind either to avoid double-binding.
2669 vma->bound |= GLOBAL_BIND | LOCAL_BIND;
2674 static int aliasing_gtt_bind_vma(struct i915_vma *vma,
2675 enum i915_cache_level cache_level,
2681 ret = i915_get_ggtt_vma_pages(vma);
2685 /* Currently applicable only to VLV */
2687 if (vma->obj->gt_ro)
2688 pte_flags |= PTE_READ_ONLY;
2691 if (flags & GLOBAL_BIND) {
2692 vma->vm->insert_entries(vma->vm,
2693 vma->ggtt_view.pages,
2695 cache_level, pte_flags);
2698 if (flags & LOCAL_BIND) {
2699 struct i915_hw_ppgtt *appgtt =
2700 to_i915(vma->vm->dev)->mm.aliasing_ppgtt;
2701 appgtt->base.insert_entries(&appgtt->base,
2702 vma->ggtt_view.pages,
2704 cache_level, pte_flags);
2710 static void ggtt_unbind_vma(struct i915_vma *vma)
2712 struct drm_device *dev = vma->vm->dev;
2713 struct drm_i915_private *dev_priv = to_i915(dev);
2714 struct drm_i915_gem_object *obj = vma->obj;
2715 const uint64_t size = min_t(uint64_t,
2719 if (vma->bound & GLOBAL_BIND) {
2720 vma->vm->clear_range(vma->vm,
2726 if (dev_priv->mm.aliasing_ppgtt && vma->bound & LOCAL_BIND) {
2727 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
2729 appgtt->base.clear_range(&appgtt->base,
2736 void i915_gem_gtt_finish_object(struct drm_i915_gem_object *obj)
2738 struct drm_device *dev = obj->base.dev;
2739 struct drm_i915_private *dev_priv = to_i915(dev);
2742 interruptible = do_idling(dev_priv);
2744 dma_unmap_sg(&dev->pdev->dev, obj->pages->sgl, obj->pages->nents,
2745 PCI_DMA_BIDIRECTIONAL);
2747 undo_idling(dev_priv, interruptible);
2750 static void i915_gtt_color_adjust(struct drm_mm_node *node,
2751 unsigned long color,
2755 if (node->color != color)
2758 if (!list_empty(&node->node_list)) {
2759 node = list_entry(node->node_list.next,
2762 if (node->allocated && node->color != color)
2767 static int i915_gem_setup_global_gtt(struct drm_device *dev,
2772 /* Let GEM Manage all of the aperture.
2774 * However, leave one page at the end still bound to the scratch page.
2775 * There are a number of places where the hardware apparently prefetches
2776 * past the end of the object, and we've seen multiple hangs with the
2777 * GPU head pointer stuck in a batchbuffer bound at the last page of the
2778 * aperture. One page should be enough to keep any prefetching inside
2781 struct drm_i915_private *dev_priv = to_i915(dev);
2782 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2783 struct drm_mm_node *entry;
2784 struct drm_i915_gem_object *obj;
2785 unsigned long hole_start, hole_end;
2788 BUG_ON(mappable_end > end);
2790 ggtt->base.start = start;
2792 /* Subtract the guard page before address space initialization to
2793 * shrink the range used by drm_mm */
2794 ggtt->base.total = end - start - PAGE_SIZE;
2795 i915_address_space_init(&ggtt->base, dev_priv);
2796 ggtt->base.total += PAGE_SIZE;
2798 ret = intel_vgt_balloon(dev_priv);
2803 ggtt->base.mm.color_adjust = i915_gtt_color_adjust;
2805 /* Mark any preallocated objects as occupied */
2806 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
2807 struct i915_vma *vma = i915_gem_obj_to_vma(obj, &ggtt->base);
2809 DRM_DEBUG_KMS("reserving preallocated space: %llx + %zx\n",
2810 i915_gem_obj_ggtt_offset(obj), obj->base.size);
2812 WARN_ON(i915_gem_obj_ggtt_bound(obj));
2813 ret = drm_mm_reserve_node(&ggtt->base.mm, &vma->node);
2815 DRM_DEBUG_KMS("Reservation failed: %i\n", ret);
2818 vma->bound |= GLOBAL_BIND;
2819 __i915_vma_set_map_and_fenceable(vma);
2820 list_add_tail(&vma->vm_link, &ggtt->base.inactive_list);
2823 /* Clear any non-preallocated blocks */
2824 drm_mm_for_each_hole(entry, &ggtt->base.mm, hole_start, hole_end) {
2825 DRM_DEBUG_KMS("clearing unused GTT space: [%lx, %lx]\n",
2826 hole_start, hole_end);
2827 ggtt->base.clear_range(&ggtt->base, hole_start,
2828 hole_end - hole_start, true);
2831 /* And finally clear the reserved guard page */
2832 ggtt->base.clear_range(&ggtt->base, end - PAGE_SIZE, PAGE_SIZE, true);
2834 if (USES_PPGTT(dev) && !USES_FULL_PPGTT(dev)) {
2835 struct i915_hw_ppgtt *ppgtt;
2837 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2841 ret = __hw_ppgtt_init(dev, ppgtt);
2843 ppgtt->base.cleanup(&ppgtt->base);
2848 if (ppgtt->base.allocate_va_range)
2849 ret = ppgtt->base.allocate_va_range(&ppgtt->base, 0,
2852 ppgtt->base.cleanup(&ppgtt->base);
2857 ppgtt->base.clear_range(&ppgtt->base,
2862 dev_priv->mm.aliasing_ppgtt = ppgtt;
2863 WARN_ON(ggtt->base.bind_vma != ggtt_bind_vma);
2864 ggtt->base.bind_vma = aliasing_gtt_bind_vma;
2871 * i915_gem_init_ggtt - Initialize GEM for Global GTT
2874 void i915_gem_init_ggtt(struct drm_device *dev)
2876 struct drm_i915_private *dev_priv = to_i915(dev);
2877 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2879 i915_gem_setup_global_gtt(dev, 0, ggtt->mappable_end, ggtt->base.total);
2883 * i915_ggtt_cleanup_hw - Clean up GGTT hardware initialization
2886 void i915_ggtt_cleanup_hw(struct drm_device *dev)
2888 struct drm_i915_private *dev_priv = to_i915(dev);
2889 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2891 if (dev_priv->mm.aliasing_ppgtt) {
2892 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
2894 ppgtt->base.cleanup(&ppgtt->base);
2897 i915_gem_cleanup_stolen(dev);
2899 if (drm_mm_initialized(&ggtt->base.mm)) {
2900 intel_vgt_deballoon(dev_priv);
2902 drm_mm_takedown(&ggtt->base.mm);
2903 list_del(&ggtt->base.global_link);
2906 ggtt->base.cleanup(&ggtt->base);
2909 static unsigned int gen6_get_total_gtt_size(u16 snb_gmch_ctl)
2911 snb_gmch_ctl >>= SNB_GMCH_GGMS_SHIFT;
2912 snb_gmch_ctl &= SNB_GMCH_GGMS_MASK;
2913 return snb_gmch_ctl << 20;
2916 static unsigned int gen8_get_total_gtt_size(u16 bdw_gmch_ctl)
2918 bdw_gmch_ctl >>= BDW_GMCH_GGMS_SHIFT;
2919 bdw_gmch_ctl &= BDW_GMCH_GGMS_MASK;
2921 bdw_gmch_ctl = 1 << bdw_gmch_ctl;
2923 #ifdef CONFIG_X86_32
2924 /* Limit 32b platforms to a 2GB GGTT: 4 << 20 / pte size * PAGE_SIZE */
2925 if (bdw_gmch_ctl > 4)
2929 return bdw_gmch_ctl << 20;
2932 static unsigned int chv_get_total_gtt_size(u16 gmch_ctrl)
2934 gmch_ctrl >>= SNB_GMCH_GGMS_SHIFT;
2935 gmch_ctrl &= SNB_GMCH_GGMS_MASK;
2938 return 1 << (20 + gmch_ctrl);
2943 static size_t gen6_get_stolen_size(u16 snb_gmch_ctl)
2945 snb_gmch_ctl >>= SNB_GMCH_GMS_SHIFT;
2946 snb_gmch_ctl &= SNB_GMCH_GMS_MASK;
2947 return snb_gmch_ctl << 25; /* 32 MB units */
2950 static size_t gen8_get_stolen_size(u16 bdw_gmch_ctl)
2952 bdw_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2953 bdw_gmch_ctl &= BDW_GMCH_GMS_MASK;
2954 return bdw_gmch_ctl << 25; /* 32 MB units */
2957 static size_t chv_get_stolen_size(u16 gmch_ctrl)
2959 gmch_ctrl >>= SNB_GMCH_GMS_SHIFT;
2960 gmch_ctrl &= SNB_GMCH_GMS_MASK;
2963 * 0x0 to 0x10: 32MB increments starting at 0MB
2964 * 0x11 to 0x16: 4MB increments starting at 8MB
2965 * 0x17 to 0x1d: 4MB increments start at 36MB
2967 if (gmch_ctrl < 0x11)
2968 return gmch_ctrl << 25;
2969 else if (gmch_ctrl < 0x17)
2970 return (gmch_ctrl - 0x11 + 2) << 22;
2972 return (gmch_ctrl - 0x17 + 9) << 22;
2975 static size_t gen9_get_stolen_size(u16 gen9_gmch_ctl)
2977 gen9_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2978 gen9_gmch_ctl &= BDW_GMCH_GMS_MASK;
2980 if (gen9_gmch_ctl < 0xf0)
2981 return gen9_gmch_ctl << 25; /* 32 MB units */
2983 /* 4MB increments starting at 0xf0 for 4MB */
2984 return (gen9_gmch_ctl - 0xf0 + 1) << 22;
2987 static int ggtt_probe_common(struct drm_device *dev,
2990 struct drm_i915_private *dev_priv = to_i915(dev);
2991 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2992 struct i915_page_scratch *scratch_page;
2993 phys_addr_t ggtt_phys_addr;
2995 /* For Modern GENs the PTEs and register space are split in the BAR */
2996 ggtt_phys_addr = pci_resource_start(dev->pdev, 0) +
2997 (pci_resource_len(dev->pdev, 0) / 2);
3000 * On BXT writes larger than 64 bit to the GTT pagetable range will be
3001 * dropped. For WC mappings in general we have 64 byte burst writes
3002 * when the WC buffer is flushed, so we can't use it, but have to
3003 * resort to an uncached mapping. The WC issue is easily caught by the
3004 * readback check when writing GTT PTE entries.
3006 if (IS_BROXTON(dev))
3007 ggtt->gsm = ioremap_nocache(ggtt_phys_addr, gtt_size);
3009 ggtt->gsm = ioremap_wc(ggtt_phys_addr, gtt_size);
3011 DRM_ERROR("Failed to map the gtt page table\n");
3015 scratch_page = alloc_scratch_page(dev);
3016 if (IS_ERR(scratch_page)) {
3017 DRM_ERROR("Scratch setup failed\n");
3018 /* iounmap will also get called at remove, but meh */
3020 return PTR_ERR(scratch_page);
3023 ggtt->base.scratch_page = scratch_page;
3028 /* The GGTT and PPGTT need a private PPAT setup in order to handle cacheability
3029 * bits. When using advanced contexts each context stores its own PAT, but
3030 * writing this data shouldn't be harmful even in those cases. */
3031 static void bdw_setup_private_ppat(struct drm_i915_private *dev_priv)
3035 pat = GEN8_PPAT(0, GEN8_PPAT_WB | GEN8_PPAT_LLC) | /* for normal objects, no eLLC */
3036 GEN8_PPAT(1, GEN8_PPAT_WC | GEN8_PPAT_LLCELLC) | /* for something pointing to ptes? */
3037 GEN8_PPAT(2, GEN8_PPAT_WT | GEN8_PPAT_LLCELLC) | /* for scanout with eLLC */
3038 GEN8_PPAT(3, GEN8_PPAT_UC) | /* Uncached objects, mostly for scanout */
3039 GEN8_PPAT(4, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(0)) |
3040 GEN8_PPAT(5, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(1)) |
3041 GEN8_PPAT(6, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(2)) |
3042 GEN8_PPAT(7, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3));
3044 if (!USES_PPGTT(dev_priv))
3045 /* Spec: "For GGTT, there is NO pat_sel[2:0] from the entry,
3046 * so RTL will always use the value corresponding to
3048 * So let's disable cache for GGTT to avoid screen corruptions.
3049 * MOCS still can be used though.
3050 * - System agent ggtt writes (i.e. cpu gtt mmaps) already work
3051 * before this patch, i.e. the same uncached + snooping access
3052 * like on gen6/7 seems to be in effect.
3053 * - So this just fixes blitter/render access. Again it looks
3054 * like it's not just uncached access, but uncached + snooping.
3055 * So we can still hold onto all our assumptions wrt cpu
3056 * clflushing on LLC machines.
3058 pat = GEN8_PPAT(0, GEN8_PPAT_UC);
3060 /* XXX: spec defines this as 2 distinct registers. It's unclear if a 64b
3061 * write would work. */
3062 I915_WRITE(GEN8_PRIVATE_PAT_LO, pat);
3063 I915_WRITE(GEN8_PRIVATE_PAT_HI, pat >> 32);
3066 static void chv_setup_private_ppat(struct drm_i915_private *dev_priv)
3071 * Map WB on BDW to snooped on CHV.
3073 * Only the snoop bit has meaning for CHV, the rest is
3076 * The hardware will never snoop for certain types of accesses:
3077 * - CPU GTT (GMADR->GGTT->no snoop->memory)
3078 * - PPGTT page tables
3079 * - some other special cycles
3081 * As with BDW, we also need to consider the following for GT accesses:
3082 * "For GGTT, there is NO pat_sel[2:0] from the entry,
3083 * so RTL will always use the value corresponding to
3085 * Which means we must set the snoop bit in PAT entry 0
3086 * in order to keep the global status page working.
3088 pat = GEN8_PPAT(0, CHV_PPAT_SNOOP) |
3092 GEN8_PPAT(4, CHV_PPAT_SNOOP) |
3093 GEN8_PPAT(5, CHV_PPAT_SNOOP) |
3094 GEN8_PPAT(6, CHV_PPAT_SNOOP) |
3095 GEN8_PPAT(7, CHV_PPAT_SNOOP);
3097 I915_WRITE(GEN8_PRIVATE_PAT_LO, pat);
3098 I915_WRITE(GEN8_PRIVATE_PAT_HI, pat >> 32);
3101 static int gen8_gmch_probe(struct i915_ggtt *ggtt)
3103 struct drm_device *dev = ggtt->base.dev;
3104 struct drm_i915_private *dev_priv = to_i915(dev);
3108 /* TODO: We're not aware of mappable constraints on gen8 yet */
3109 ggtt->mappable_base = pci_resource_start(dev->pdev, 2);
3110 ggtt->mappable_end = pci_resource_len(dev->pdev, 2);
3112 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(39)))
3113 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(39));
3115 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
3117 if (INTEL_INFO(dev)->gen >= 9) {
3118 ggtt->stolen_size = gen9_get_stolen_size(snb_gmch_ctl);
3119 ggtt->size = gen8_get_total_gtt_size(snb_gmch_ctl);
3120 } else if (IS_CHERRYVIEW(dev)) {
3121 ggtt->stolen_size = chv_get_stolen_size(snb_gmch_ctl);
3122 ggtt->size = chv_get_total_gtt_size(snb_gmch_ctl);
3124 ggtt->stolen_size = gen8_get_stolen_size(snb_gmch_ctl);
3125 ggtt->size = gen8_get_total_gtt_size(snb_gmch_ctl);
3128 ggtt->base.total = (ggtt->size / sizeof(gen8_pte_t)) << PAGE_SHIFT;
3130 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
3131 chv_setup_private_ppat(dev_priv);
3133 bdw_setup_private_ppat(dev_priv);
3135 ret = ggtt_probe_common(dev, ggtt->size);
3137 ggtt->base.bind_vma = ggtt_bind_vma;
3138 ggtt->base.unbind_vma = ggtt_unbind_vma;
3139 ggtt->base.insert_page = gen8_ggtt_insert_page;
3140 ggtt->base.clear_range = nop_clear_range;
3141 if (!USES_FULL_PPGTT(dev_priv))
3142 ggtt->base.clear_range = gen8_ggtt_clear_range;
3144 ggtt->base.insert_entries = gen8_ggtt_insert_entries;
3145 if (IS_CHERRYVIEW(dev_priv))
3146 ggtt->base.insert_entries = gen8_ggtt_insert_entries__BKL;
3151 static int gen6_gmch_probe(struct i915_ggtt *ggtt)
3153 struct drm_device *dev = ggtt->base.dev;
3157 ggtt->mappable_base = pci_resource_start(dev->pdev, 2);
3158 ggtt->mappable_end = pci_resource_len(dev->pdev, 2);
3160 /* 64/512MB is the current min/max we actually know of, but this is just
3161 * a coarse sanity check.
3163 if ((ggtt->mappable_end < (64<<20) || (ggtt->mappable_end > (512<<20)))) {
3164 DRM_ERROR("Unknown GMADR size (%llx)\n", ggtt->mappable_end);
3168 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(40)))
3169 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(40));
3170 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
3172 ggtt->stolen_size = gen6_get_stolen_size(snb_gmch_ctl);
3173 ggtt->size = gen6_get_total_gtt_size(snb_gmch_ctl);
3174 ggtt->base.total = (ggtt->size / sizeof(gen6_pte_t)) << PAGE_SHIFT;
3176 ret = ggtt_probe_common(dev, ggtt->size);
3178 ggtt->base.clear_range = gen6_ggtt_clear_range;
3179 ggtt->base.insert_page = gen6_ggtt_insert_page;
3180 ggtt->base.insert_entries = gen6_ggtt_insert_entries;
3181 ggtt->base.bind_vma = ggtt_bind_vma;
3182 ggtt->base.unbind_vma = ggtt_unbind_vma;
3187 static void gen6_gmch_remove(struct i915_address_space *vm)
3189 struct i915_ggtt *ggtt = container_of(vm, struct i915_ggtt, base);
3192 free_scratch_page(vm->dev, vm->scratch_page);
3195 static int i915_gmch_probe(struct i915_ggtt *ggtt)
3197 struct drm_device *dev = ggtt->base.dev;
3198 struct drm_i915_private *dev_priv = to_i915(dev);
3201 ret = intel_gmch_probe(dev_priv->bridge_dev, dev_priv->dev->pdev, NULL);
3203 DRM_ERROR("failed to set up gmch\n");
3207 intel_gtt_get(&ggtt->base.total, &ggtt->stolen_size,
3208 &ggtt->mappable_base, &ggtt->mappable_end);
3210 ggtt->do_idle_maps = needs_idle_maps(dev_priv->dev);
3211 ggtt->base.insert_page = i915_ggtt_insert_page;
3212 ggtt->base.insert_entries = i915_ggtt_insert_entries;
3213 ggtt->base.clear_range = i915_ggtt_clear_range;
3214 ggtt->base.bind_vma = ggtt_bind_vma;
3215 ggtt->base.unbind_vma = ggtt_unbind_vma;
3217 if (unlikely(ggtt->do_idle_maps))
3218 DRM_INFO("applying Ironlake quirks for intel_iommu\n");
3223 static void i915_gmch_remove(struct i915_address_space *vm)
3225 intel_gmch_remove();
3229 * i915_ggtt_init_hw - Initialize GGTT hardware
3232 int i915_ggtt_init_hw(struct drm_device *dev)
3234 struct drm_i915_private *dev_priv = to_i915(dev);
3235 struct i915_ggtt *ggtt = &dev_priv->ggtt;
3238 if (INTEL_INFO(dev)->gen <= 5) {
3239 ggtt->probe = i915_gmch_probe;
3240 ggtt->base.cleanup = i915_gmch_remove;
3241 } else if (INTEL_INFO(dev)->gen < 8) {
3242 ggtt->probe = gen6_gmch_probe;
3243 ggtt->base.cleanup = gen6_gmch_remove;
3246 ggtt->base.pte_encode = iris_pte_encode;
3247 else if (IS_HASWELL(dev))
3248 ggtt->base.pte_encode = hsw_pte_encode;
3249 else if (IS_VALLEYVIEW(dev))
3250 ggtt->base.pte_encode = byt_pte_encode;
3251 else if (INTEL_INFO(dev)->gen >= 7)
3252 ggtt->base.pte_encode = ivb_pte_encode;
3254 ggtt->base.pte_encode = snb_pte_encode;
3256 ggtt->probe = gen8_gmch_probe;
3257 ggtt->base.cleanup = gen6_gmch_remove;
3260 ggtt->base.dev = dev;
3261 ggtt->base.is_ggtt = true;
3263 ret = ggtt->probe(ggtt);
3267 if ((ggtt->base.total - 1) >> 32) {
3268 DRM_ERROR("We never expected a Global GTT with more than 32bits"
3269 "of address space! Found %lldM!\n",
3270 ggtt->base.total >> 20);
3271 ggtt->base.total = 1ULL << 32;
3272 ggtt->mappable_end = min(ggtt->mappable_end, ggtt->base.total);
3276 * Initialise stolen early so that we may reserve preallocated
3277 * objects for the BIOS to KMS transition.
3279 ret = i915_gem_init_stolen(dev);
3281 goto out_gtt_cleanup;
3283 /* GMADR is the PCI mmio aperture into the global GTT. */
3284 DRM_INFO("Memory usable by graphics device = %lluM\n",
3285 ggtt->base.total >> 20);
3286 DRM_DEBUG_DRIVER("GMADR size = %lldM\n", ggtt->mappable_end >> 20);
3287 DRM_DEBUG_DRIVER("GTT stolen size = %zdM\n", ggtt->stolen_size >> 20);
3288 #ifdef CONFIG_INTEL_IOMMU
3289 if (intel_iommu_gfx_mapped)
3290 DRM_INFO("VT-d active for gfx access\n");
3296 ggtt->base.cleanup(&ggtt->base);
3301 int i915_ggtt_enable_hw(struct drm_device *dev)
3303 if (INTEL_INFO(dev)->gen < 6 && !intel_enable_gtt())
3309 void i915_gem_restore_gtt_mappings(struct drm_device *dev)
3311 struct drm_i915_private *dev_priv = to_i915(dev);
3312 struct i915_ggtt *ggtt = &dev_priv->ggtt;
3313 struct drm_i915_gem_object *obj;
3314 struct i915_vma *vma;
3316 i915_check_and_clear_faults(dev_priv);
3318 /* First fill our portion of the GTT with scratch pages */
3319 ggtt->base.clear_range(&ggtt->base, ggtt->base.start, ggtt->base.total,
3322 /* Cache flush objects bound into GGTT and rebind them. */
3323 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
3324 list_for_each_entry(vma, &obj->vma_list, obj_link) {
3325 if (vma->vm != &ggtt->base)
3328 WARN_ON(i915_vma_bind(vma, obj->cache_level,
3332 if (obj->pin_display)
3333 WARN_ON(i915_gem_object_set_to_gtt_domain(obj, false));
3336 if (INTEL_INFO(dev)->gen >= 8) {
3337 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
3338 chv_setup_private_ppat(dev_priv);
3340 bdw_setup_private_ppat(dev_priv);
3345 if (USES_PPGTT(dev)) {
3346 struct i915_address_space *vm;
3348 list_for_each_entry(vm, &dev_priv->vm_list, global_link) {
3349 /* TODO: Perhaps it shouldn't be gen6 specific */
3351 struct i915_hw_ppgtt *ppgtt;
3354 ppgtt = dev_priv->mm.aliasing_ppgtt;
3356 ppgtt = i915_vm_to_ppgtt(vm);
3358 gen6_write_page_range(dev_priv, &ppgtt->pd,
3359 0, ppgtt->base.total);
3363 i915_ggtt_flush(dev_priv);
3366 static struct i915_vma *
3367 __i915_gem_vma_create(struct drm_i915_gem_object *obj,
3368 struct i915_address_space *vm,
3369 const struct i915_ggtt_view *ggtt_view)
3371 struct i915_vma *vma;
3373 if (WARN_ON(i915_is_ggtt(vm) != !!ggtt_view))
3374 return ERR_PTR(-EINVAL);
3376 vma = kmem_cache_zalloc(to_i915(obj->base.dev)->vmas, GFP_KERNEL);
3378 return ERR_PTR(-ENOMEM);
3380 INIT_LIST_HEAD(&vma->vm_link);
3381 INIT_LIST_HEAD(&vma->obj_link);
3382 INIT_LIST_HEAD(&vma->exec_list);
3385 vma->is_ggtt = i915_is_ggtt(vm);
3387 if (i915_is_ggtt(vm))
3388 vma->ggtt_view = *ggtt_view;
3390 i915_ppgtt_get(i915_vm_to_ppgtt(vm));
3392 list_add_tail(&vma->obj_link, &obj->vma_list);
3398 i915_gem_obj_lookup_or_create_vma(struct drm_i915_gem_object *obj,
3399 struct i915_address_space *vm)
3401 struct i915_vma *vma;
3403 vma = i915_gem_obj_to_vma(obj, vm);
3405 vma = __i915_gem_vma_create(obj, vm,
3406 i915_is_ggtt(vm) ? &i915_ggtt_view_normal : NULL);
3412 i915_gem_obj_lookup_or_create_ggtt_vma(struct drm_i915_gem_object *obj,
3413 const struct i915_ggtt_view *view)
3415 struct drm_device *dev = obj->base.dev;
3416 struct drm_i915_private *dev_priv = to_i915(dev);
3417 struct i915_ggtt *ggtt = &dev_priv->ggtt;
3418 struct i915_vma *vma = i915_gem_obj_to_ggtt_view(obj, view);
3421 vma = __i915_gem_vma_create(obj, &ggtt->base, view);
3427 static struct scatterlist *
3428 rotate_pages(const dma_addr_t *in, unsigned int offset,
3429 unsigned int width, unsigned int height,
3430 unsigned int stride,
3431 struct sg_table *st, struct scatterlist *sg)
3433 unsigned int column, row;
3434 unsigned int src_idx;
3436 for (column = 0; column < width; column++) {
3437 src_idx = stride * (height - 1) + column;
3438 for (row = 0; row < height; row++) {
3440 /* We don't need the pages, but need to initialize
3441 * the entries so the sg list can be happily traversed.
3442 * The only thing we need are DMA addresses.
3444 sg_set_page(sg, NULL, PAGE_SIZE, 0);
3445 sg_dma_address(sg) = in[offset + src_idx];
3446 sg_dma_len(sg) = PAGE_SIZE;
3455 static struct sg_table *
3456 intel_rotate_fb_obj_pages(struct intel_rotation_info *rot_info,
3457 struct drm_i915_gem_object *obj)
3459 const size_t n_pages = obj->base.size / PAGE_SIZE;
3460 unsigned int size_pages = rot_info->plane[0].width * rot_info->plane[0].height;
3461 unsigned int size_pages_uv;
3462 struct sgt_iter sgt_iter;
3463 dma_addr_t dma_addr;
3465 dma_addr_t *page_addr_list;
3466 struct sg_table *st;
3467 unsigned int uv_start_page;
3468 struct scatterlist *sg;
3471 /* Allocate a temporary list of source pages for random access. */
3472 page_addr_list = drm_malloc_gfp(n_pages,
3475 if (!page_addr_list)
3476 return ERR_PTR(ret);
3478 /* Account for UV plane with NV12. */
3479 if (rot_info->pixel_format == DRM_FORMAT_NV12)
3480 size_pages_uv = rot_info->plane[1].width * rot_info->plane[1].height;
3484 /* Allocate target SG list. */
3485 st = kmalloc(sizeof(*st), GFP_KERNEL);
3489 ret = sg_alloc_table(st, size_pages + size_pages_uv, GFP_KERNEL);
3493 /* Populate source page list from the object. */
3495 for_each_sgt_dma(dma_addr, sgt_iter, obj->pages)
3496 page_addr_list[i++] = dma_addr;
3498 GEM_BUG_ON(i != n_pages);
3502 /* Rotate the pages. */
3503 sg = rotate_pages(page_addr_list, 0,
3504 rot_info->plane[0].width, rot_info->plane[0].height,
3505 rot_info->plane[0].width,
3508 /* Append the UV plane if NV12. */
3509 if (rot_info->pixel_format == DRM_FORMAT_NV12) {
3510 uv_start_page = size_pages;
3512 /* Check for tile-row un-alignment. */
3513 if (offset_in_page(rot_info->uv_offset))
3516 rot_info->uv_start_page = uv_start_page;
3518 sg = rotate_pages(page_addr_list, rot_info->uv_start_page,
3519 rot_info->plane[1].width, rot_info->plane[1].height,
3520 rot_info->plane[1].width,
3524 DRM_DEBUG_KMS("Created rotated page mapping for object size %zu (%ux%u tiles, %u pages (%u plane 0)).\n",
3525 obj->base.size, rot_info->plane[0].width,
3526 rot_info->plane[0].height, size_pages + size_pages_uv,
3529 drm_free_large(page_addr_list);
3536 drm_free_large(page_addr_list);
3538 DRM_DEBUG_KMS("Failed to create rotated mapping for object size %zu! (%d) (%ux%u tiles, %u pages (%u plane 0))\n",
3539 obj->base.size, ret, rot_info->plane[0].width,
3540 rot_info->plane[0].height, size_pages + size_pages_uv,
3542 return ERR_PTR(ret);
3545 static struct sg_table *
3546 intel_partial_pages(const struct i915_ggtt_view *view,
3547 struct drm_i915_gem_object *obj)
3549 struct sg_table *st;
3550 struct scatterlist *sg;
3551 struct sg_page_iter obj_sg_iter;
3554 st = kmalloc(sizeof(*st), GFP_KERNEL);
3558 ret = sg_alloc_table(st, view->params.partial.size, GFP_KERNEL);
3564 for_each_sg_page(obj->pages->sgl, &obj_sg_iter, obj->pages->nents,
3565 view->params.partial.offset)
3567 if (st->nents >= view->params.partial.size)
3570 sg_set_page(sg, NULL, PAGE_SIZE, 0);
3571 sg_dma_address(sg) = sg_page_iter_dma_address(&obj_sg_iter);
3572 sg_dma_len(sg) = PAGE_SIZE;
3583 return ERR_PTR(ret);
3587 i915_get_ggtt_vma_pages(struct i915_vma *vma)
3591 if (vma->ggtt_view.pages)
3594 if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL)
3595 vma->ggtt_view.pages = vma->obj->pages;
3596 else if (vma->ggtt_view.type == I915_GGTT_VIEW_ROTATED)
3597 vma->ggtt_view.pages =
3598 intel_rotate_fb_obj_pages(&vma->ggtt_view.params.rotated, vma->obj);
3599 else if (vma->ggtt_view.type == I915_GGTT_VIEW_PARTIAL)
3600 vma->ggtt_view.pages =
3601 intel_partial_pages(&vma->ggtt_view, vma->obj);
3603 WARN_ONCE(1, "GGTT view %u not implemented!\n",
3604 vma->ggtt_view.type);
3606 if (!vma->ggtt_view.pages) {
3607 DRM_ERROR("Failed to get pages for GGTT view type %u!\n",
3608 vma->ggtt_view.type);
3610 } else if (IS_ERR(vma->ggtt_view.pages)) {
3611 ret = PTR_ERR(vma->ggtt_view.pages);
3612 vma->ggtt_view.pages = NULL;
3613 DRM_ERROR("Failed to get pages for VMA view type %u (%d)!\n",
3614 vma->ggtt_view.type, ret);
3621 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
3623 * @cache_level: mapping cache level
3624 * @flags: flags like global or local mapping
3626 * DMA addresses are taken from the scatter-gather table of this object (or of
3627 * this VMA in case of non-default GGTT views) and PTE entries set up.
3628 * Note that DMA addresses are also the only part of the SG table we care about.
3630 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
3636 if (WARN_ON(flags == 0))
3640 if (flags & PIN_GLOBAL)
3641 bind_flags |= GLOBAL_BIND;
3642 if (flags & PIN_USER)
3643 bind_flags |= LOCAL_BIND;
3645 if (flags & PIN_UPDATE)
3646 bind_flags |= vma->bound;
3648 bind_flags &= ~vma->bound;
3650 if (bind_flags == 0)
3653 if (vma->bound == 0 && vma->vm->allocate_va_range) {
3654 /* XXX: i915_vma_pin() will fix this +- hack */
3656 trace_i915_va_alloc(vma);
3657 ret = vma->vm->allocate_va_range(vma->vm,
3665 ret = vma->vm->bind_vma(vma, cache_level, bind_flags);
3669 vma->bound |= bind_flags;
3675 * i915_ggtt_view_size - Get the size of a GGTT view.
3676 * @obj: Object the view is of.
3677 * @view: The view in question.
3679 * @return The size of the GGTT view in bytes.
3682 i915_ggtt_view_size(struct drm_i915_gem_object *obj,
3683 const struct i915_ggtt_view *view)
3685 if (view->type == I915_GGTT_VIEW_NORMAL) {
3686 return obj->base.size;
3687 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
3688 return intel_rotation_info_size(&view->params.rotated) << PAGE_SHIFT;
3689 } else if (view->type == I915_GGTT_VIEW_PARTIAL) {
3690 return view->params.partial.size << PAGE_SHIFT;
3692 WARN_ONCE(1, "GGTT view %u not implemented!\n", view->type);
3693 return obj->base.size;
3697 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
3701 lockdep_assert_held(&vma->vm->dev->struct_mutex);
3702 if (WARN_ON(!vma->obj->map_and_fenceable))
3703 return ERR_PTR(-ENODEV);
3705 GEM_BUG_ON(!vma->is_ggtt);
3706 GEM_BUG_ON((vma->bound & GLOBAL_BIND) == 0);
3710 ptr = io_mapping_map_wc(i915_vm_to_ggtt(vma->vm)->mappable,
3714 return ERR_PTR(-ENOMEM);