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) && dev->pdev->revision < 0xb) {
144 DRM_DEBUG_DRIVER("disabling PPGTT on pre-B3 step VLV\n");
148 if (INTEL_INFO(dev)->gen >= 8 && i915.enable_execlists)
149 return has_full_48bit_ppgtt ? 3 : 2;
151 return has_aliasing_ppgtt ? 1 : 0;
154 static int ppgtt_bind_vma(struct i915_vma *vma,
155 enum i915_cache_level cache_level,
160 /* Currently applicable only to VLV */
162 pte_flags |= PTE_READ_ONLY;
164 vma->vm->insert_entries(vma->vm, vma->obj->pages, vma->node.start,
165 cache_level, pte_flags);
170 static void ppgtt_unbind_vma(struct i915_vma *vma)
172 vma->vm->clear_range(vma->vm,
178 static gen8_pte_t gen8_pte_encode(dma_addr_t addr,
179 enum i915_cache_level level,
182 gen8_pte_t pte = valid ? _PAGE_PRESENT | _PAGE_RW : 0;
186 case I915_CACHE_NONE:
187 pte |= PPAT_UNCACHED_INDEX;
190 pte |= PPAT_DISPLAY_ELLC_INDEX;
193 pte |= PPAT_CACHED_INDEX;
200 static gen8_pde_t gen8_pde_encode(const dma_addr_t addr,
201 const enum i915_cache_level level)
203 gen8_pde_t pde = _PAGE_PRESENT | _PAGE_RW;
205 if (level != I915_CACHE_NONE)
206 pde |= PPAT_CACHED_PDE_INDEX;
208 pde |= PPAT_UNCACHED_INDEX;
212 #define gen8_pdpe_encode gen8_pde_encode
213 #define gen8_pml4e_encode gen8_pde_encode
215 static gen6_pte_t snb_pte_encode(dma_addr_t addr,
216 enum i915_cache_level level,
217 bool valid, u32 unused)
219 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
220 pte |= GEN6_PTE_ADDR_ENCODE(addr);
223 case I915_CACHE_L3_LLC:
225 pte |= GEN6_PTE_CACHE_LLC;
227 case I915_CACHE_NONE:
228 pte |= GEN6_PTE_UNCACHED;
237 static gen6_pte_t ivb_pte_encode(dma_addr_t addr,
238 enum i915_cache_level level,
239 bool valid, u32 unused)
241 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
242 pte |= GEN6_PTE_ADDR_ENCODE(addr);
245 case I915_CACHE_L3_LLC:
246 pte |= GEN7_PTE_CACHE_L3_LLC;
249 pte |= GEN6_PTE_CACHE_LLC;
251 case I915_CACHE_NONE:
252 pte |= GEN6_PTE_UNCACHED;
261 static gen6_pte_t byt_pte_encode(dma_addr_t addr,
262 enum i915_cache_level level,
263 bool valid, u32 flags)
265 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
266 pte |= GEN6_PTE_ADDR_ENCODE(addr);
268 if (!(flags & PTE_READ_ONLY))
269 pte |= BYT_PTE_WRITEABLE;
271 if (level != I915_CACHE_NONE)
272 pte |= BYT_PTE_SNOOPED_BY_CPU_CACHES;
277 static gen6_pte_t hsw_pte_encode(dma_addr_t addr,
278 enum i915_cache_level level,
279 bool valid, u32 unused)
281 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
282 pte |= HSW_PTE_ADDR_ENCODE(addr);
284 if (level != I915_CACHE_NONE)
285 pte |= HSW_WB_LLC_AGE3;
290 static gen6_pte_t iris_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);
298 case I915_CACHE_NONE:
301 pte |= HSW_WT_ELLC_LLC_AGE3;
304 pte |= HSW_WB_ELLC_LLC_AGE3;
311 static int __setup_page_dma(struct drm_device *dev,
312 struct i915_page_dma *p, gfp_t flags)
314 struct device *device = &dev->pdev->dev;
316 p->page = alloc_page(flags);
320 p->daddr = dma_map_page(device,
321 p->page, 0, 4096, PCI_DMA_BIDIRECTIONAL);
323 if (dma_mapping_error(device, p->daddr)) {
324 __free_page(p->page);
331 static int setup_page_dma(struct drm_device *dev, struct i915_page_dma *p)
333 return __setup_page_dma(dev, p, GFP_KERNEL);
336 static void cleanup_page_dma(struct drm_device *dev, struct i915_page_dma *p)
338 if (WARN_ON(!p->page))
341 dma_unmap_page(&dev->pdev->dev, p->daddr, 4096, PCI_DMA_BIDIRECTIONAL);
342 __free_page(p->page);
343 memset(p, 0, sizeof(*p));
346 static void *kmap_page_dma(struct i915_page_dma *p)
348 return kmap_atomic(p->page);
351 /* We use the flushing unmap only with ppgtt structures:
352 * page directories, page tables and scratch pages.
354 static void kunmap_page_dma(struct drm_device *dev, void *vaddr)
356 /* There are only few exceptions for gen >=6. chv and bxt.
357 * And we are not sure about the latter so play safe for now.
359 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
360 drm_clflush_virt_range(vaddr, PAGE_SIZE);
362 kunmap_atomic(vaddr);
365 #define kmap_px(px) kmap_page_dma(px_base(px))
366 #define kunmap_px(ppgtt, vaddr) kunmap_page_dma((ppgtt)->base.dev, (vaddr))
368 #define setup_px(dev, px) setup_page_dma((dev), px_base(px))
369 #define cleanup_px(dev, px) cleanup_page_dma((dev), px_base(px))
370 #define fill_px(dev, px, v) fill_page_dma((dev), px_base(px), (v))
371 #define fill32_px(dev, px, v) fill_page_dma_32((dev), px_base(px), (v))
373 static void fill_page_dma(struct drm_device *dev, struct i915_page_dma *p,
377 uint64_t * const vaddr = kmap_page_dma(p);
379 for (i = 0; i < 512; i++)
382 kunmap_page_dma(dev, vaddr);
385 static void fill_page_dma_32(struct drm_device *dev, struct i915_page_dma *p,
386 const uint32_t val32)
392 fill_page_dma(dev, p, v);
395 static struct i915_page_scratch *alloc_scratch_page(struct drm_device *dev)
397 struct i915_page_scratch *sp;
400 sp = kzalloc(sizeof(*sp), GFP_KERNEL);
402 return ERR_PTR(-ENOMEM);
404 ret = __setup_page_dma(dev, px_base(sp), GFP_DMA32 | __GFP_ZERO);
410 set_pages_uc(px_page(sp), 1);
415 static void free_scratch_page(struct drm_device *dev,
416 struct i915_page_scratch *sp)
418 set_pages_wb(px_page(sp), 1);
424 static struct i915_page_table *alloc_pt(struct drm_device *dev)
426 struct i915_page_table *pt;
427 const size_t count = INTEL_INFO(dev)->gen >= 8 ?
428 GEN8_PTES : GEN6_PTES;
431 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
433 return ERR_PTR(-ENOMEM);
435 pt->used_ptes = kcalloc(BITS_TO_LONGS(count), sizeof(*pt->used_ptes),
441 ret = setup_px(dev, pt);
448 kfree(pt->used_ptes);
455 static void free_pt(struct drm_device *dev, struct i915_page_table *pt)
458 kfree(pt->used_ptes);
462 static void gen8_initialize_pt(struct i915_address_space *vm,
463 struct i915_page_table *pt)
465 gen8_pte_t scratch_pte;
467 scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
468 I915_CACHE_LLC, true);
470 fill_px(vm->dev, pt, scratch_pte);
473 static void gen6_initialize_pt(struct i915_address_space *vm,
474 struct i915_page_table *pt)
476 gen6_pte_t scratch_pte;
478 WARN_ON(px_dma(vm->scratch_page) == 0);
480 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
481 I915_CACHE_LLC, true, 0);
483 fill32_px(vm->dev, pt, scratch_pte);
486 static struct i915_page_directory *alloc_pd(struct drm_device *dev)
488 struct i915_page_directory *pd;
491 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
493 return ERR_PTR(-ENOMEM);
495 pd->used_pdes = kcalloc(BITS_TO_LONGS(I915_PDES),
496 sizeof(*pd->used_pdes), GFP_KERNEL);
500 ret = setup_px(dev, pd);
507 kfree(pd->used_pdes);
514 static void free_pd(struct drm_device *dev, struct i915_page_directory *pd)
518 kfree(pd->used_pdes);
523 static void gen8_initialize_pd(struct i915_address_space *vm,
524 struct i915_page_directory *pd)
526 gen8_pde_t scratch_pde;
528 scratch_pde = gen8_pde_encode(px_dma(vm->scratch_pt), I915_CACHE_LLC);
530 fill_px(vm->dev, pd, scratch_pde);
533 static int __pdp_init(struct drm_device *dev,
534 struct i915_page_directory_pointer *pdp)
536 size_t pdpes = I915_PDPES_PER_PDP(dev);
538 pdp->used_pdpes = kcalloc(BITS_TO_LONGS(pdpes),
539 sizeof(unsigned long),
541 if (!pdp->used_pdpes)
544 pdp->page_directory = kcalloc(pdpes, sizeof(*pdp->page_directory),
546 if (!pdp->page_directory) {
547 kfree(pdp->used_pdpes);
548 /* the PDP might be the statically allocated top level. Keep it
549 * as clean as possible */
550 pdp->used_pdpes = NULL;
557 static void __pdp_fini(struct i915_page_directory_pointer *pdp)
559 kfree(pdp->used_pdpes);
560 kfree(pdp->page_directory);
561 pdp->page_directory = NULL;
565 i915_page_directory_pointer *alloc_pdp(struct drm_device *dev)
567 struct i915_page_directory_pointer *pdp;
570 WARN_ON(!USES_FULL_48BIT_PPGTT(dev));
572 pdp = kzalloc(sizeof(*pdp), GFP_KERNEL);
574 return ERR_PTR(-ENOMEM);
576 ret = __pdp_init(dev, pdp);
580 ret = setup_px(dev, pdp);
594 static void free_pdp(struct drm_device *dev,
595 struct i915_page_directory_pointer *pdp)
598 if (USES_FULL_48BIT_PPGTT(dev)) {
599 cleanup_px(dev, pdp);
604 static void gen8_initialize_pdp(struct i915_address_space *vm,
605 struct i915_page_directory_pointer *pdp)
607 gen8_ppgtt_pdpe_t scratch_pdpe;
609 scratch_pdpe = gen8_pdpe_encode(px_dma(vm->scratch_pd), I915_CACHE_LLC);
611 fill_px(vm->dev, pdp, scratch_pdpe);
614 static void gen8_initialize_pml4(struct i915_address_space *vm,
615 struct i915_pml4 *pml4)
617 gen8_ppgtt_pml4e_t scratch_pml4e;
619 scratch_pml4e = gen8_pml4e_encode(px_dma(vm->scratch_pdp),
622 fill_px(vm->dev, pml4, scratch_pml4e);
626 gen8_setup_page_directory(struct i915_hw_ppgtt *ppgtt,
627 struct i915_page_directory_pointer *pdp,
628 struct i915_page_directory *pd,
631 gen8_ppgtt_pdpe_t *page_directorypo;
633 if (!USES_FULL_48BIT_PPGTT(ppgtt->base.dev))
636 page_directorypo = kmap_px(pdp);
637 page_directorypo[index] = gen8_pdpe_encode(px_dma(pd), I915_CACHE_LLC);
638 kunmap_px(ppgtt, page_directorypo);
642 gen8_setup_page_directory_pointer(struct i915_hw_ppgtt *ppgtt,
643 struct i915_pml4 *pml4,
644 struct i915_page_directory_pointer *pdp,
647 gen8_ppgtt_pml4e_t *pagemap = kmap_px(pml4);
649 WARN_ON(!USES_FULL_48BIT_PPGTT(ppgtt->base.dev));
650 pagemap[index] = gen8_pml4e_encode(px_dma(pdp), I915_CACHE_LLC);
651 kunmap_px(ppgtt, pagemap);
654 /* Broadwell Page Directory Pointer Descriptors */
655 static int gen8_write_pdp(struct drm_i915_gem_request *req,
659 struct intel_engine_cs *ring = req->ring;
664 ret = intel_ring_begin(req, 6);
668 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
669 intel_ring_emit_reg(ring, GEN8_RING_PDP_UDW(ring, entry));
670 intel_ring_emit(ring, upper_32_bits(addr));
671 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
672 intel_ring_emit_reg(ring, GEN8_RING_PDP_LDW(ring, entry));
673 intel_ring_emit(ring, lower_32_bits(addr));
674 intel_ring_advance(ring);
679 static int gen8_legacy_mm_switch(struct i915_hw_ppgtt *ppgtt,
680 struct drm_i915_gem_request *req)
684 for (i = GEN8_LEGACY_PDPES - 1; i >= 0; i--) {
685 const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i);
687 ret = gen8_write_pdp(req, i, pd_daddr);
695 static int gen8_48b_mm_switch(struct i915_hw_ppgtt *ppgtt,
696 struct drm_i915_gem_request *req)
698 return gen8_write_pdp(req, 0, px_dma(&ppgtt->pml4));
701 static void gen8_ppgtt_clear_pte_range(struct i915_address_space *vm,
702 struct i915_page_directory_pointer *pdp,
705 gen8_pte_t scratch_pte)
707 struct i915_hw_ppgtt *ppgtt =
708 container_of(vm, struct i915_hw_ppgtt, base);
709 gen8_pte_t *pt_vaddr;
710 unsigned pdpe = gen8_pdpe_index(start);
711 unsigned pde = gen8_pde_index(start);
712 unsigned pte = gen8_pte_index(start);
713 unsigned num_entries = length >> PAGE_SHIFT;
714 unsigned last_pte, i;
719 while (num_entries) {
720 struct i915_page_directory *pd;
721 struct i915_page_table *pt;
723 if (WARN_ON(!pdp->page_directory[pdpe]))
726 pd = pdp->page_directory[pdpe];
728 if (WARN_ON(!pd->page_table[pde]))
731 pt = pd->page_table[pde];
733 if (WARN_ON(!px_page(pt)))
736 last_pte = pte + num_entries;
737 if (last_pte > GEN8_PTES)
738 last_pte = GEN8_PTES;
740 pt_vaddr = kmap_px(pt);
742 for (i = pte; i < last_pte; i++) {
743 pt_vaddr[i] = scratch_pte;
747 kunmap_px(ppgtt, pt);
750 if (++pde == I915_PDES) {
751 if (++pdpe == I915_PDPES_PER_PDP(vm->dev))
758 static void gen8_ppgtt_clear_range(struct i915_address_space *vm,
763 struct i915_hw_ppgtt *ppgtt =
764 container_of(vm, struct i915_hw_ppgtt, base);
765 gen8_pte_t scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
766 I915_CACHE_LLC, use_scratch);
768 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
769 gen8_ppgtt_clear_pte_range(vm, &ppgtt->pdp, start, length,
773 struct i915_page_directory_pointer *pdp;
775 gen8_for_each_pml4e(pdp, &ppgtt->pml4, start, length, pml4e) {
776 gen8_ppgtt_clear_pte_range(vm, pdp, start, length,
783 gen8_ppgtt_insert_pte_entries(struct i915_address_space *vm,
784 struct i915_page_directory_pointer *pdp,
785 struct sg_page_iter *sg_iter,
787 enum i915_cache_level cache_level)
789 struct i915_hw_ppgtt *ppgtt =
790 container_of(vm, struct i915_hw_ppgtt, base);
791 gen8_pte_t *pt_vaddr;
792 unsigned pdpe = gen8_pdpe_index(start);
793 unsigned pde = gen8_pde_index(start);
794 unsigned pte = gen8_pte_index(start);
798 while (__sg_page_iter_next(sg_iter)) {
799 if (pt_vaddr == NULL) {
800 struct i915_page_directory *pd = pdp->page_directory[pdpe];
801 struct i915_page_table *pt = pd->page_table[pde];
802 pt_vaddr = kmap_px(pt);
806 gen8_pte_encode(sg_page_iter_dma_address(sg_iter),
808 if (++pte == GEN8_PTES) {
809 kunmap_px(ppgtt, pt_vaddr);
811 if (++pde == I915_PDES) {
812 if (++pdpe == I915_PDPES_PER_PDP(vm->dev))
821 kunmap_px(ppgtt, pt_vaddr);
824 static void gen8_ppgtt_insert_entries(struct i915_address_space *vm,
825 struct sg_table *pages,
827 enum i915_cache_level cache_level,
830 struct i915_hw_ppgtt *ppgtt =
831 container_of(vm, struct i915_hw_ppgtt, base);
832 struct sg_page_iter sg_iter;
834 __sg_page_iter_start(&sg_iter, pages->sgl, sg_nents(pages->sgl), 0);
836 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
837 gen8_ppgtt_insert_pte_entries(vm, &ppgtt->pdp, &sg_iter, start,
840 struct i915_page_directory_pointer *pdp;
842 uint64_t length = (uint64_t)pages->orig_nents << PAGE_SHIFT;
844 gen8_for_each_pml4e(pdp, &ppgtt->pml4, start, length, pml4e) {
845 gen8_ppgtt_insert_pte_entries(vm, pdp, &sg_iter,
851 static void gen8_free_page_tables(struct drm_device *dev,
852 struct i915_page_directory *pd)
859 for_each_set_bit(i, pd->used_pdes, I915_PDES) {
860 if (WARN_ON(!pd->page_table[i]))
863 free_pt(dev, pd->page_table[i]);
864 pd->page_table[i] = NULL;
868 static int gen8_init_scratch(struct i915_address_space *vm)
870 struct drm_device *dev = vm->dev;
872 vm->scratch_page = alloc_scratch_page(dev);
873 if (IS_ERR(vm->scratch_page))
874 return PTR_ERR(vm->scratch_page);
876 vm->scratch_pt = alloc_pt(dev);
877 if (IS_ERR(vm->scratch_pt)) {
878 free_scratch_page(dev, vm->scratch_page);
879 return PTR_ERR(vm->scratch_pt);
882 vm->scratch_pd = alloc_pd(dev);
883 if (IS_ERR(vm->scratch_pd)) {
884 free_pt(dev, vm->scratch_pt);
885 free_scratch_page(dev, vm->scratch_page);
886 return PTR_ERR(vm->scratch_pd);
889 if (USES_FULL_48BIT_PPGTT(dev)) {
890 vm->scratch_pdp = alloc_pdp(dev);
891 if (IS_ERR(vm->scratch_pdp)) {
892 free_pd(dev, vm->scratch_pd);
893 free_pt(dev, vm->scratch_pt);
894 free_scratch_page(dev, vm->scratch_page);
895 return PTR_ERR(vm->scratch_pdp);
899 gen8_initialize_pt(vm, vm->scratch_pt);
900 gen8_initialize_pd(vm, vm->scratch_pd);
901 if (USES_FULL_48BIT_PPGTT(dev))
902 gen8_initialize_pdp(vm, vm->scratch_pdp);
907 static int gen8_ppgtt_notify_vgt(struct i915_hw_ppgtt *ppgtt, bool create)
909 enum vgt_g2v_type msg;
910 struct drm_device *dev = ppgtt->base.dev;
911 struct drm_i915_private *dev_priv = dev->dev_private;
914 if (USES_FULL_48BIT_PPGTT(dev)) {
915 u64 daddr = px_dma(&ppgtt->pml4);
917 I915_WRITE(vgtif_reg(pdp[0].lo), lower_32_bits(daddr));
918 I915_WRITE(vgtif_reg(pdp[0].hi), upper_32_bits(daddr));
920 msg = (create ? VGT_G2V_PPGTT_L4_PAGE_TABLE_CREATE :
921 VGT_G2V_PPGTT_L4_PAGE_TABLE_DESTROY);
923 for (i = 0; i < GEN8_LEGACY_PDPES; i++) {
924 u64 daddr = i915_page_dir_dma_addr(ppgtt, i);
926 I915_WRITE(vgtif_reg(pdp[i].lo), lower_32_bits(daddr));
927 I915_WRITE(vgtif_reg(pdp[i].hi), upper_32_bits(daddr));
930 msg = (create ? VGT_G2V_PPGTT_L3_PAGE_TABLE_CREATE :
931 VGT_G2V_PPGTT_L3_PAGE_TABLE_DESTROY);
934 I915_WRITE(vgtif_reg(g2v_notify), msg);
939 static void gen8_free_scratch(struct i915_address_space *vm)
941 struct drm_device *dev = vm->dev;
943 if (USES_FULL_48BIT_PPGTT(dev))
944 free_pdp(dev, vm->scratch_pdp);
945 free_pd(dev, vm->scratch_pd);
946 free_pt(dev, vm->scratch_pt);
947 free_scratch_page(dev, vm->scratch_page);
950 static void gen8_ppgtt_cleanup_3lvl(struct drm_device *dev,
951 struct i915_page_directory_pointer *pdp)
955 for_each_set_bit(i, pdp->used_pdpes, I915_PDPES_PER_PDP(dev)) {
956 if (WARN_ON(!pdp->page_directory[i]))
959 gen8_free_page_tables(dev, pdp->page_directory[i]);
960 free_pd(dev, pdp->page_directory[i]);
966 static void gen8_ppgtt_cleanup_4lvl(struct i915_hw_ppgtt *ppgtt)
970 for_each_set_bit(i, ppgtt->pml4.used_pml4es, GEN8_PML4ES_PER_PML4) {
971 if (WARN_ON(!ppgtt->pml4.pdps[i]))
974 gen8_ppgtt_cleanup_3lvl(ppgtt->base.dev, ppgtt->pml4.pdps[i]);
977 cleanup_px(ppgtt->base.dev, &ppgtt->pml4);
980 static void gen8_ppgtt_cleanup(struct i915_address_space *vm)
982 struct i915_hw_ppgtt *ppgtt =
983 container_of(vm, struct i915_hw_ppgtt, base);
985 if (intel_vgpu_active(vm->dev))
986 gen8_ppgtt_notify_vgt(ppgtt, false);
988 if (!USES_FULL_48BIT_PPGTT(ppgtt->base.dev))
989 gen8_ppgtt_cleanup_3lvl(ppgtt->base.dev, &ppgtt->pdp);
991 gen8_ppgtt_cleanup_4lvl(ppgtt);
993 gen8_free_scratch(vm);
997 * gen8_ppgtt_alloc_pagetabs() - Allocate page tables for VA range.
998 * @vm: Master vm structure.
999 * @pd: Page directory for this address range.
1000 * @start: Starting virtual address to begin allocations.
1001 * @length: Size of the allocations.
1002 * @new_pts: Bitmap set by function with new allocations. Likely used by the
1003 * caller to free on error.
1005 * Allocate the required number of page tables. Extremely similar to
1006 * gen8_ppgtt_alloc_page_directories(). The main difference is here we are limited by
1007 * the page directory boundary (instead of the page directory pointer). That
1008 * boundary is 1GB virtual. Therefore, unlike gen8_ppgtt_alloc_page_directories(), it is
1009 * possible, and likely that the caller will need to use multiple calls of this
1010 * function to achieve the appropriate allocation.
1012 * Return: 0 if success; negative error code otherwise.
1014 static int gen8_ppgtt_alloc_pagetabs(struct i915_address_space *vm,
1015 struct i915_page_directory *pd,
1018 unsigned long *new_pts)
1020 struct drm_device *dev = vm->dev;
1021 struct i915_page_table *pt;
1024 gen8_for_each_pde(pt, pd, start, length, pde) {
1025 /* Don't reallocate page tables */
1026 if (test_bit(pde, pd->used_pdes)) {
1027 /* Scratch is never allocated this way */
1028 WARN_ON(pt == vm->scratch_pt);
1036 gen8_initialize_pt(vm, pt);
1037 pd->page_table[pde] = pt;
1038 __set_bit(pde, new_pts);
1039 trace_i915_page_table_entry_alloc(vm, pde, start, GEN8_PDE_SHIFT);
1045 for_each_set_bit(pde, new_pts, I915_PDES)
1046 free_pt(dev, pd->page_table[pde]);
1052 * gen8_ppgtt_alloc_page_directories() - Allocate page directories for VA range.
1053 * @vm: Master vm structure.
1054 * @pdp: Page directory pointer for this address range.
1055 * @start: Starting virtual address to begin allocations.
1056 * @length: Size of the allocations.
1057 * @new_pds: Bitmap set by function with new allocations. Likely used by the
1058 * caller to free on error.
1060 * Allocate the required number of page directories starting at the pde index of
1061 * @start, and ending at the pde index @start + @length. This function will skip
1062 * over already allocated page directories within the range, and only allocate
1063 * new ones, setting the appropriate pointer within the pdp as well as the
1064 * correct position in the bitmap @new_pds.
1066 * The function will only allocate the pages within the range for a give page
1067 * directory pointer. In other words, if @start + @length straddles a virtually
1068 * addressed PDP boundary (512GB for 4k pages), there will be more allocations
1069 * required by the caller, This is not currently possible, and the BUG in the
1070 * code will prevent it.
1072 * Return: 0 if success; negative error code otherwise.
1075 gen8_ppgtt_alloc_page_directories(struct i915_address_space *vm,
1076 struct i915_page_directory_pointer *pdp,
1079 unsigned long *new_pds)
1081 struct drm_device *dev = vm->dev;
1082 struct i915_page_directory *pd;
1084 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1086 WARN_ON(!bitmap_empty(new_pds, pdpes));
1088 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1089 if (test_bit(pdpe, pdp->used_pdpes))
1096 gen8_initialize_pd(vm, pd);
1097 pdp->page_directory[pdpe] = pd;
1098 __set_bit(pdpe, new_pds);
1099 trace_i915_page_directory_entry_alloc(vm, pdpe, start, GEN8_PDPE_SHIFT);
1105 for_each_set_bit(pdpe, new_pds, pdpes)
1106 free_pd(dev, pdp->page_directory[pdpe]);
1112 * gen8_ppgtt_alloc_page_dirpointers() - Allocate pdps for VA range.
1113 * @vm: Master vm structure.
1114 * @pml4: Page map level 4 for this address range.
1115 * @start: Starting virtual address to begin allocations.
1116 * @length: Size of the allocations.
1117 * @new_pdps: Bitmap set by function with new allocations. Likely used by the
1118 * caller to free on error.
1120 * Allocate the required number of page directory pointers. Extremely similar to
1121 * gen8_ppgtt_alloc_page_directories() and gen8_ppgtt_alloc_pagetabs().
1122 * The main difference is here we are limited by the pml4 boundary (instead of
1123 * the page directory pointer).
1125 * Return: 0 if success; negative error code otherwise.
1128 gen8_ppgtt_alloc_page_dirpointers(struct i915_address_space *vm,
1129 struct i915_pml4 *pml4,
1132 unsigned long *new_pdps)
1134 struct drm_device *dev = vm->dev;
1135 struct i915_page_directory_pointer *pdp;
1138 WARN_ON(!bitmap_empty(new_pdps, GEN8_PML4ES_PER_PML4));
1140 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1141 if (!test_bit(pml4e, pml4->used_pml4es)) {
1142 pdp = alloc_pdp(dev);
1146 gen8_initialize_pdp(vm, pdp);
1147 pml4->pdps[pml4e] = pdp;
1148 __set_bit(pml4e, new_pdps);
1149 trace_i915_page_directory_pointer_entry_alloc(vm,
1159 for_each_set_bit(pml4e, new_pdps, GEN8_PML4ES_PER_PML4)
1160 free_pdp(dev, pml4->pdps[pml4e]);
1166 free_gen8_temp_bitmaps(unsigned long *new_pds, unsigned long *new_pts)
1172 /* Fills in the page directory bitmap, and the array of page tables bitmap. Both
1173 * of these are based on the number of PDPEs in the system.
1176 int __must_check alloc_gen8_temp_bitmaps(unsigned long **new_pds,
1177 unsigned long **new_pts,
1183 pds = kcalloc(BITS_TO_LONGS(pdpes), sizeof(unsigned long), GFP_TEMPORARY);
1187 pts = kcalloc(pdpes, BITS_TO_LONGS(I915_PDES) * sizeof(unsigned long),
1198 free_gen8_temp_bitmaps(pds, pts);
1202 /* PDE TLBs are a pain to invalidate on GEN8+. When we modify
1203 * the page table structures, we mark them dirty so that
1204 * context switching/execlist queuing code takes extra steps
1205 * to ensure that tlbs are flushed.
1207 static void mark_tlbs_dirty(struct i915_hw_ppgtt *ppgtt)
1209 ppgtt->pd_dirty_rings = INTEL_INFO(ppgtt->base.dev)->ring_mask;
1212 static int gen8_alloc_va_range_3lvl(struct i915_address_space *vm,
1213 struct i915_page_directory_pointer *pdp,
1217 struct i915_hw_ppgtt *ppgtt =
1218 container_of(vm, struct i915_hw_ppgtt, base);
1219 unsigned long *new_page_dirs, *new_page_tables;
1220 struct drm_device *dev = vm->dev;
1221 struct i915_page_directory *pd;
1222 const uint64_t orig_start = start;
1223 const uint64_t orig_length = length;
1225 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1228 /* Wrap is never okay since we can only represent 48b, and we don't
1229 * actually use the other side of the canonical address space.
1231 if (WARN_ON(start + length < start))
1234 if (WARN_ON(start + length > vm->total))
1237 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables, pdpes);
1241 /* Do the allocations first so we can easily bail out */
1242 ret = gen8_ppgtt_alloc_page_directories(vm, pdp, start, length,
1245 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1249 /* For every page directory referenced, allocate page tables */
1250 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1251 ret = gen8_ppgtt_alloc_pagetabs(vm, pd, start, length,
1252 new_page_tables + pdpe * BITS_TO_LONGS(I915_PDES));
1258 length = orig_length;
1260 /* Allocations have completed successfully, so set the bitmaps, and do
1262 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1263 gen8_pde_t *const page_directory = kmap_px(pd);
1264 struct i915_page_table *pt;
1265 uint64_t pd_len = length;
1266 uint64_t pd_start = start;
1269 /* Every pd should be allocated, we just did that above. */
1272 gen8_for_each_pde(pt, pd, pd_start, pd_len, pde) {
1273 /* Same reasoning as pd */
1276 WARN_ON(!gen8_pte_count(pd_start, pd_len));
1278 /* Set our used ptes within the page table */
1279 bitmap_set(pt->used_ptes,
1280 gen8_pte_index(pd_start),
1281 gen8_pte_count(pd_start, pd_len));
1283 /* Our pde is now pointing to the pagetable, pt */
1284 __set_bit(pde, pd->used_pdes);
1286 /* Map the PDE to the page table */
1287 page_directory[pde] = gen8_pde_encode(px_dma(pt),
1289 trace_i915_page_table_entry_map(&ppgtt->base, pde, pt,
1290 gen8_pte_index(start),
1291 gen8_pte_count(start, length),
1294 /* NB: We haven't yet mapped ptes to pages. At this
1295 * point we're still relying on insert_entries() */
1298 kunmap_px(ppgtt, page_directory);
1299 __set_bit(pdpe, pdp->used_pdpes);
1300 gen8_setup_page_directory(ppgtt, pdp, pd, pdpe);
1303 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1304 mark_tlbs_dirty(ppgtt);
1311 for_each_set_bit(temp, new_page_tables + pdpe *
1312 BITS_TO_LONGS(I915_PDES), I915_PDES)
1313 free_pt(dev, pdp->page_directory[pdpe]->page_table[temp]);
1316 for_each_set_bit(pdpe, new_page_dirs, pdpes)
1317 free_pd(dev, pdp->page_directory[pdpe]);
1319 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1320 mark_tlbs_dirty(ppgtt);
1324 static int gen8_alloc_va_range_4lvl(struct i915_address_space *vm,
1325 struct i915_pml4 *pml4,
1329 DECLARE_BITMAP(new_pdps, GEN8_PML4ES_PER_PML4);
1330 struct i915_hw_ppgtt *ppgtt =
1331 container_of(vm, struct i915_hw_ppgtt, base);
1332 struct i915_page_directory_pointer *pdp;
1336 /* Do the pml4 allocations first, so we don't need to track the newly
1337 * allocated tables below the pdp */
1338 bitmap_zero(new_pdps, GEN8_PML4ES_PER_PML4);
1340 /* The pagedirectory and pagetable allocations are done in the shared 3
1341 * and 4 level code. Just allocate the pdps.
1343 ret = gen8_ppgtt_alloc_page_dirpointers(vm, pml4, start, length,
1348 WARN(bitmap_weight(new_pdps, GEN8_PML4ES_PER_PML4) > 2,
1349 "The allocation has spanned more than 512GB. "
1350 "It is highly likely this is incorrect.");
1352 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1355 ret = gen8_alloc_va_range_3lvl(vm, pdp, start, length);
1359 gen8_setup_page_directory_pointer(ppgtt, pml4, pdp, pml4e);
1362 bitmap_or(pml4->used_pml4es, new_pdps, pml4->used_pml4es,
1363 GEN8_PML4ES_PER_PML4);
1368 for_each_set_bit(pml4e, new_pdps, GEN8_PML4ES_PER_PML4)
1369 gen8_ppgtt_cleanup_3lvl(vm->dev, pml4->pdps[pml4e]);
1374 static int gen8_alloc_va_range(struct i915_address_space *vm,
1375 uint64_t start, uint64_t length)
1377 struct i915_hw_ppgtt *ppgtt =
1378 container_of(vm, struct i915_hw_ppgtt, base);
1380 if (USES_FULL_48BIT_PPGTT(vm->dev))
1381 return gen8_alloc_va_range_4lvl(vm, &ppgtt->pml4, start, length);
1383 return gen8_alloc_va_range_3lvl(vm, &ppgtt->pdp, start, length);
1386 static void gen8_dump_pdp(struct i915_page_directory_pointer *pdp,
1387 uint64_t start, uint64_t length,
1388 gen8_pte_t scratch_pte,
1391 struct i915_page_directory *pd;
1394 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1395 struct i915_page_table *pt;
1396 uint64_t pd_len = length;
1397 uint64_t pd_start = start;
1400 if (!test_bit(pdpe, pdp->used_pdpes))
1403 seq_printf(m, "\tPDPE #%d\n", pdpe);
1404 gen8_for_each_pde(pt, pd, pd_start, pd_len, pde) {
1406 gen8_pte_t *pt_vaddr;
1408 if (!test_bit(pde, pd->used_pdes))
1411 pt_vaddr = kmap_px(pt);
1412 for (pte = 0; pte < GEN8_PTES; pte += 4) {
1414 (pdpe << GEN8_PDPE_SHIFT) |
1415 (pde << GEN8_PDE_SHIFT) |
1416 (pte << GEN8_PTE_SHIFT);
1420 for (i = 0; i < 4; i++)
1421 if (pt_vaddr[pte + i] != scratch_pte)
1426 seq_printf(m, "\t\t0x%llx [%03d,%03d,%04d]: =", va, pdpe, pde, pte);
1427 for (i = 0; i < 4; i++) {
1428 if (pt_vaddr[pte + i] != scratch_pte)
1429 seq_printf(m, " %llx", pt_vaddr[pte + i]);
1431 seq_puts(m, " SCRATCH ");
1435 /* don't use kunmap_px, it could trigger
1436 * an unnecessary flush.
1438 kunmap_atomic(pt_vaddr);
1443 static void gen8_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
1445 struct i915_address_space *vm = &ppgtt->base;
1446 uint64_t start = ppgtt->base.start;
1447 uint64_t length = ppgtt->base.total;
1448 gen8_pte_t scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
1449 I915_CACHE_LLC, true);
1451 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
1452 gen8_dump_pdp(&ppgtt->pdp, start, length, scratch_pte, m);
1455 struct i915_pml4 *pml4 = &ppgtt->pml4;
1456 struct i915_page_directory_pointer *pdp;
1458 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1459 if (!test_bit(pml4e, pml4->used_pml4es))
1462 seq_printf(m, " PML4E #%llu\n", pml4e);
1463 gen8_dump_pdp(pdp, start, length, scratch_pte, m);
1468 static int gen8_preallocate_top_level_pdps(struct i915_hw_ppgtt *ppgtt)
1470 unsigned long *new_page_dirs, *new_page_tables;
1471 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1474 /* We allocate temp bitmap for page tables for no gain
1475 * but as this is for init only, lets keep the things simple
1477 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables, pdpes);
1481 /* Allocate for all pdps regardless of how the ppgtt
1484 ret = gen8_ppgtt_alloc_page_directories(&ppgtt->base, &ppgtt->pdp,
1488 *ppgtt->pdp.used_pdpes = *new_page_dirs;
1490 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1496 * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers
1497 * with a net effect resembling a 2-level page table in normal x86 terms. Each
1498 * PDP represents 1GB of memory 4 * 512 * 512 * 4096 = 4GB legacy 32b address
1502 static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
1506 ret = gen8_init_scratch(&ppgtt->base);
1510 ppgtt->base.start = 0;
1511 ppgtt->base.cleanup = gen8_ppgtt_cleanup;
1512 ppgtt->base.allocate_va_range = gen8_alloc_va_range;
1513 ppgtt->base.insert_entries = gen8_ppgtt_insert_entries;
1514 ppgtt->base.clear_range = gen8_ppgtt_clear_range;
1515 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
1516 ppgtt->base.bind_vma = ppgtt_bind_vma;
1517 ppgtt->debug_dump = gen8_dump_ppgtt;
1519 if (USES_FULL_48BIT_PPGTT(ppgtt->base.dev)) {
1520 ret = setup_px(ppgtt->base.dev, &ppgtt->pml4);
1524 gen8_initialize_pml4(&ppgtt->base, &ppgtt->pml4);
1526 ppgtt->base.total = 1ULL << 48;
1527 ppgtt->switch_mm = gen8_48b_mm_switch;
1529 ret = __pdp_init(ppgtt->base.dev, &ppgtt->pdp);
1533 ppgtt->base.total = 1ULL << 32;
1534 ppgtt->switch_mm = gen8_legacy_mm_switch;
1535 trace_i915_page_directory_pointer_entry_alloc(&ppgtt->base,
1539 if (intel_vgpu_active(ppgtt->base.dev)) {
1540 ret = gen8_preallocate_top_level_pdps(ppgtt);
1546 if (intel_vgpu_active(ppgtt->base.dev))
1547 gen8_ppgtt_notify_vgt(ppgtt, true);
1552 gen8_free_scratch(&ppgtt->base);
1556 static void gen6_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
1558 struct i915_address_space *vm = &ppgtt->base;
1559 struct i915_page_table *unused;
1560 gen6_pte_t scratch_pte;
1562 uint32_t pte, pde, temp;
1563 uint32_t start = ppgtt->base.start, length = ppgtt->base.total;
1565 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1566 I915_CACHE_LLC, true, 0);
1568 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde) {
1570 gen6_pte_t *pt_vaddr;
1571 const dma_addr_t pt_addr = px_dma(ppgtt->pd.page_table[pde]);
1572 pd_entry = readl(ppgtt->pd_addr + pde);
1573 expected = (GEN6_PDE_ADDR_ENCODE(pt_addr) | GEN6_PDE_VALID);
1575 if (pd_entry != expected)
1576 seq_printf(m, "\tPDE #%d mismatch: Actual PDE: %x Expected PDE: %x\n",
1580 seq_printf(m, "\tPDE: %x\n", pd_entry);
1582 pt_vaddr = kmap_px(ppgtt->pd.page_table[pde]);
1584 for (pte = 0; pte < GEN6_PTES; pte+=4) {
1586 (pde * PAGE_SIZE * GEN6_PTES) +
1590 for (i = 0; i < 4; i++)
1591 if (pt_vaddr[pte + i] != scratch_pte)
1596 seq_printf(m, "\t\t0x%lx [%03d,%04d]: =", va, pde, pte);
1597 for (i = 0; i < 4; i++) {
1598 if (pt_vaddr[pte + i] != scratch_pte)
1599 seq_printf(m, " %08x", pt_vaddr[pte + i]);
1601 seq_puts(m, " SCRATCH ");
1605 kunmap_px(ppgtt, pt_vaddr);
1609 /* Write pde (index) from the page directory @pd to the page table @pt */
1610 static void gen6_write_pde(struct i915_page_directory *pd,
1611 const int pde, struct i915_page_table *pt)
1613 /* Caller needs to make sure the write completes if necessary */
1614 struct i915_hw_ppgtt *ppgtt =
1615 container_of(pd, struct i915_hw_ppgtt, pd);
1618 pd_entry = GEN6_PDE_ADDR_ENCODE(px_dma(pt));
1619 pd_entry |= GEN6_PDE_VALID;
1621 writel(pd_entry, ppgtt->pd_addr + pde);
1624 /* Write all the page tables found in the ppgtt structure to incrementing page
1626 static void gen6_write_page_range(struct drm_i915_private *dev_priv,
1627 struct i915_page_directory *pd,
1628 uint32_t start, uint32_t length)
1630 struct i915_page_table *pt;
1633 gen6_for_each_pde(pt, pd, start, length, temp, pde)
1634 gen6_write_pde(pd, pde, pt);
1636 /* Make sure write is complete before other code can use this page
1637 * table. Also require for WC mapped PTEs */
1638 readl(dev_priv->gtt.gsm);
1641 static uint32_t get_pd_offset(struct i915_hw_ppgtt *ppgtt)
1643 BUG_ON(ppgtt->pd.base.ggtt_offset & 0x3f);
1645 return (ppgtt->pd.base.ggtt_offset / 64) << 16;
1648 static int hsw_mm_switch(struct i915_hw_ppgtt *ppgtt,
1649 struct drm_i915_gem_request *req)
1651 struct intel_engine_cs *ring = req->ring;
1654 /* NB: TLBs must be flushed and invalidated before a switch */
1655 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1659 ret = intel_ring_begin(req, 6);
1663 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1664 intel_ring_emit_reg(ring, RING_PP_DIR_DCLV(ring));
1665 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1666 intel_ring_emit_reg(ring, RING_PP_DIR_BASE(ring));
1667 intel_ring_emit(ring, get_pd_offset(ppgtt));
1668 intel_ring_emit(ring, MI_NOOP);
1669 intel_ring_advance(ring);
1674 static int vgpu_mm_switch(struct i915_hw_ppgtt *ppgtt,
1675 struct drm_i915_gem_request *req)
1677 struct intel_engine_cs *ring = req->ring;
1678 struct drm_i915_private *dev_priv = to_i915(ppgtt->base.dev);
1680 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
1681 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
1685 static int gen7_mm_switch(struct i915_hw_ppgtt *ppgtt,
1686 struct drm_i915_gem_request *req)
1688 struct intel_engine_cs *ring = req->ring;
1691 /* NB: TLBs must be flushed and invalidated before a switch */
1692 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1696 ret = intel_ring_begin(req, 6);
1700 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1701 intel_ring_emit_reg(ring, RING_PP_DIR_DCLV(ring));
1702 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1703 intel_ring_emit_reg(ring, RING_PP_DIR_BASE(ring));
1704 intel_ring_emit(ring, get_pd_offset(ppgtt));
1705 intel_ring_emit(ring, MI_NOOP);
1706 intel_ring_advance(ring);
1708 /* XXX: RCS is the only one to auto invalidate the TLBs? */
1709 if (ring->id != RCS) {
1710 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1718 static int gen6_mm_switch(struct i915_hw_ppgtt *ppgtt,
1719 struct drm_i915_gem_request *req)
1721 struct intel_engine_cs *ring = req->ring;
1722 struct drm_device *dev = ppgtt->base.dev;
1723 struct drm_i915_private *dev_priv = dev->dev_private;
1726 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
1727 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
1729 POSTING_READ(RING_PP_DIR_DCLV(ring));
1734 static void gen8_ppgtt_enable(struct drm_device *dev)
1736 struct drm_i915_private *dev_priv = dev->dev_private;
1737 struct intel_engine_cs *ring;
1740 for_each_ring(ring, dev_priv, j) {
1741 u32 four_level = USES_FULL_48BIT_PPGTT(dev) ? GEN8_GFX_PPGTT_48B : 0;
1742 I915_WRITE(RING_MODE_GEN7(ring),
1743 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE | four_level));
1747 static void gen7_ppgtt_enable(struct drm_device *dev)
1749 struct drm_i915_private *dev_priv = dev->dev_private;
1750 struct intel_engine_cs *ring;
1751 uint32_t ecochk, ecobits;
1754 ecobits = I915_READ(GAC_ECO_BITS);
1755 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B);
1757 ecochk = I915_READ(GAM_ECOCHK);
1758 if (IS_HASWELL(dev)) {
1759 ecochk |= ECOCHK_PPGTT_WB_HSW;
1761 ecochk |= ECOCHK_PPGTT_LLC_IVB;
1762 ecochk &= ~ECOCHK_PPGTT_GFDT_IVB;
1764 I915_WRITE(GAM_ECOCHK, ecochk);
1766 for_each_ring(ring, dev_priv, i) {
1767 /* GFX_MODE is per-ring on gen7+ */
1768 I915_WRITE(RING_MODE_GEN7(ring),
1769 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1773 static void gen6_ppgtt_enable(struct drm_device *dev)
1775 struct drm_i915_private *dev_priv = dev->dev_private;
1776 uint32_t ecochk, gab_ctl, ecobits;
1778 ecobits = I915_READ(GAC_ECO_BITS);
1779 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_SNB_BIT |
1780 ECOBITS_PPGTT_CACHE64B);
1782 gab_ctl = I915_READ(GAB_CTL);
1783 I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT);
1785 ecochk = I915_READ(GAM_ECOCHK);
1786 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT | ECOCHK_PPGTT_CACHE64B);
1788 I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1791 /* PPGTT support for Sandybdrige/Gen6 and later */
1792 static void gen6_ppgtt_clear_range(struct i915_address_space *vm,
1797 struct i915_hw_ppgtt *ppgtt =
1798 container_of(vm, struct i915_hw_ppgtt, base);
1799 gen6_pte_t *pt_vaddr, scratch_pte;
1800 unsigned first_entry = start >> PAGE_SHIFT;
1801 unsigned num_entries = length >> PAGE_SHIFT;
1802 unsigned act_pt = first_entry / GEN6_PTES;
1803 unsigned first_pte = first_entry % GEN6_PTES;
1804 unsigned last_pte, i;
1806 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1807 I915_CACHE_LLC, true, 0);
1809 while (num_entries) {
1810 last_pte = first_pte + num_entries;
1811 if (last_pte > GEN6_PTES)
1812 last_pte = GEN6_PTES;
1814 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1816 for (i = first_pte; i < last_pte; i++)
1817 pt_vaddr[i] = scratch_pte;
1819 kunmap_px(ppgtt, pt_vaddr);
1821 num_entries -= last_pte - first_pte;
1827 static void gen6_ppgtt_insert_entries(struct i915_address_space *vm,
1828 struct sg_table *pages,
1830 enum i915_cache_level cache_level, u32 flags)
1832 struct i915_hw_ppgtt *ppgtt =
1833 container_of(vm, struct i915_hw_ppgtt, base);
1834 gen6_pte_t *pt_vaddr;
1835 unsigned first_entry = start >> PAGE_SHIFT;
1836 unsigned act_pt = first_entry / GEN6_PTES;
1837 unsigned act_pte = first_entry % GEN6_PTES;
1838 struct sg_page_iter sg_iter;
1841 for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) {
1842 if (pt_vaddr == NULL)
1843 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1846 vm->pte_encode(sg_page_iter_dma_address(&sg_iter),
1847 cache_level, true, flags);
1849 if (++act_pte == GEN6_PTES) {
1850 kunmap_px(ppgtt, pt_vaddr);
1857 kunmap_px(ppgtt, pt_vaddr);
1860 static int gen6_alloc_va_range(struct i915_address_space *vm,
1861 uint64_t start_in, uint64_t length_in)
1863 DECLARE_BITMAP(new_page_tables, I915_PDES);
1864 struct drm_device *dev = vm->dev;
1865 struct drm_i915_private *dev_priv = dev->dev_private;
1866 struct i915_hw_ppgtt *ppgtt =
1867 container_of(vm, struct i915_hw_ppgtt, base);
1868 struct i915_page_table *pt;
1869 uint32_t start, length, start_save, length_save;
1873 if (WARN_ON(start_in + length_in > ppgtt->base.total))
1876 start = start_save = start_in;
1877 length = length_save = length_in;
1879 bitmap_zero(new_page_tables, I915_PDES);
1881 /* The allocation is done in two stages so that we can bail out with
1882 * minimal amount of pain. The first stage finds new page tables that
1883 * need allocation. The second stage marks use ptes within the page
1886 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1887 if (pt != vm->scratch_pt) {
1888 WARN_ON(bitmap_empty(pt->used_ptes, GEN6_PTES));
1892 /* We've already allocated a page table */
1893 WARN_ON(!bitmap_empty(pt->used_ptes, GEN6_PTES));
1901 gen6_initialize_pt(vm, pt);
1903 ppgtt->pd.page_table[pde] = pt;
1904 __set_bit(pde, new_page_tables);
1905 trace_i915_page_table_entry_alloc(vm, pde, start, GEN6_PDE_SHIFT);
1909 length = length_save;
1911 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1912 DECLARE_BITMAP(tmp_bitmap, GEN6_PTES);
1914 bitmap_zero(tmp_bitmap, GEN6_PTES);
1915 bitmap_set(tmp_bitmap, gen6_pte_index(start),
1916 gen6_pte_count(start, length));
1918 if (__test_and_clear_bit(pde, new_page_tables))
1919 gen6_write_pde(&ppgtt->pd, pde, pt);
1921 trace_i915_page_table_entry_map(vm, pde, pt,
1922 gen6_pte_index(start),
1923 gen6_pte_count(start, length),
1925 bitmap_or(pt->used_ptes, tmp_bitmap, pt->used_ptes,
1929 WARN_ON(!bitmap_empty(new_page_tables, I915_PDES));
1931 /* Make sure write is complete before other code can use this page
1932 * table. Also require for WC mapped PTEs */
1933 readl(dev_priv->gtt.gsm);
1935 mark_tlbs_dirty(ppgtt);
1939 for_each_set_bit(pde, new_page_tables, I915_PDES) {
1940 struct i915_page_table *pt = ppgtt->pd.page_table[pde];
1942 ppgtt->pd.page_table[pde] = vm->scratch_pt;
1943 free_pt(vm->dev, pt);
1946 mark_tlbs_dirty(ppgtt);
1950 static int gen6_init_scratch(struct i915_address_space *vm)
1952 struct drm_device *dev = vm->dev;
1954 vm->scratch_page = alloc_scratch_page(dev);
1955 if (IS_ERR(vm->scratch_page))
1956 return PTR_ERR(vm->scratch_page);
1958 vm->scratch_pt = alloc_pt(dev);
1959 if (IS_ERR(vm->scratch_pt)) {
1960 free_scratch_page(dev, vm->scratch_page);
1961 return PTR_ERR(vm->scratch_pt);
1964 gen6_initialize_pt(vm, vm->scratch_pt);
1969 static void gen6_free_scratch(struct i915_address_space *vm)
1971 struct drm_device *dev = vm->dev;
1973 free_pt(dev, vm->scratch_pt);
1974 free_scratch_page(dev, vm->scratch_page);
1977 static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
1979 struct i915_hw_ppgtt *ppgtt =
1980 container_of(vm, struct i915_hw_ppgtt, base);
1981 struct i915_page_table *pt;
1984 drm_mm_remove_node(&ppgtt->node);
1986 gen6_for_all_pdes(pt, ppgtt, pde) {
1987 if (pt != vm->scratch_pt)
1988 free_pt(ppgtt->base.dev, pt);
1991 gen6_free_scratch(vm);
1994 static int gen6_ppgtt_allocate_page_directories(struct i915_hw_ppgtt *ppgtt)
1996 struct i915_address_space *vm = &ppgtt->base;
1997 struct drm_device *dev = ppgtt->base.dev;
1998 struct drm_i915_private *dev_priv = dev->dev_private;
1999 bool retried = false;
2002 /* PPGTT PDEs reside in the GGTT and consists of 512 entries. The
2003 * allocator works in address space sizes, so it's multiplied by page
2004 * size. We allocate at the top of the GTT to avoid fragmentation.
2006 BUG_ON(!drm_mm_initialized(&dev_priv->gtt.base.mm));
2008 ret = gen6_init_scratch(vm);
2013 ret = drm_mm_insert_node_in_range_generic(&dev_priv->gtt.base.mm,
2014 &ppgtt->node, GEN6_PD_SIZE,
2016 0, dev_priv->gtt.base.total,
2018 if (ret == -ENOSPC && !retried) {
2019 ret = i915_gem_evict_something(dev, &dev_priv->gtt.base,
2020 GEN6_PD_SIZE, GEN6_PD_ALIGN,
2022 0, dev_priv->gtt.base.total,
2035 if (ppgtt->node.start < dev_priv->gtt.mappable_end)
2036 DRM_DEBUG("Forced to use aperture for PDEs\n");
2041 gen6_free_scratch(vm);
2045 static int gen6_ppgtt_alloc(struct i915_hw_ppgtt *ppgtt)
2047 return gen6_ppgtt_allocate_page_directories(ppgtt);
2050 static void gen6_scratch_va_range(struct i915_hw_ppgtt *ppgtt,
2051 uint64_t start, uint64_t length)
2053 struct i915_page_table *unused;
2056 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde)
2057 ppgtt->pd.page_table[pde] = ppgtt->base.scratch_pt;
2060 static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
2062 struct drm_device *dev = ppgtt->base.dev;
2063 struct drm_i915_private *dev_priv = dev->dev_private;
2066 ppgtt->base.pte_encode = dev_priv->gtt.base.pte_encode;
2068 ppgtt->switch_mm = gen6_mm_switch;
2069 } else if (IS_HASWELL(dev)) {
2070 ppgtt->switch_mm = hsw_mm_switch;
2071 } else if (IS_GEN7(dev)) {
2072 ppgtt->switch_mm = gen7_mm_switch;
2076 if (intel_vgpu_active(dev))
2077 ppgtt->switch_mm = vgpu_mm_switch;
2079 ret = gen6_ppgtt_alloc(ppgtt);
2083 ppgtt->base.allocate_va_range = gen6_alloc_va_range;
2084 ppgtt->base.clear_range = gen6_ppgtt_clear_range;
2085 ppgtt->base.insert_entries = gen6_ppgtt_insert_entries;
2086 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
2087 ppgtt->base.bind_vma = ppgtt_bind_vma;
2088 ppgtt->base.cleanup = gen6_ppgtt_cleanup;
2089 ppgtt->base.start = 0;
2090 ppgtt->base.total = I915_PDES * GEN6_PTES * PAGE_SIZE;
2091 ppgtt->debug_dump = gen6_dump_ppgtt;
2093 ppgtt->pd.base.ggtt_offset =
2094 ppgtt->node.start / PAGE_SIZE * sizeof(gen6_pte_t);
2096 ppgtt->pd_addr = (gen6_pte_t __iomem *)dev_priv->gtt.gsm +
2097 ppgtt->pd.base.ggtt_offset / sizeof(gen6_pte_t);
2099 gen6_scratch_va_range(ppgtt, 0, ppgtt->base.total);
2101 gen6_write_page_range(dev_priv, &ppgtt->pd, 0, ppgtt->base.total);
2103 DRM_DEBUG_DRIVER("Allocated pde space (%lldM) at GTT entry: %llx\n",
2104 ppgtt->node.size >> 20,
2105 ppgtt->node.start / PAGE_SIZE);
2107 DRM_DEBUG("Adding PPGTT at offset %x\n",
2108 ppgtt->pd.base.ggtt_offset << 10);
2113 static int __hw_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
2115 ppgtt->base.dev = dev;
2117 if (INTEL_INFO(dev)->gen < 8)
2118 return gen6_ppgtt_init(ppgtt);
2120 return gen8_ppgtt_init(ppgtt);
2123 static void i915_address_space_init(struct i915_address_space *vm,
2124 struct drm_i915_private *dev_priv)
2126 drm_mm_init(&vm->mm, vm->start, vm->total);
2127 vm->dev = dev_priv->dev;
2128 INIT_LIST_HEAD(&vm->active_list);
2129 INIT_LIST_HEAD(&vm->inactive_list);
2130 list_add_tail(&vm->global_link, &dev_priv->vm_list);
2133 int i915_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
2135 struct drm_i915_private *dev_priv = dev->dev_private;
2138 ret = __hw_ppgtt_init(dev, ppgtt);
2140 kref_init(&ppgtt->ref);
2141 i915_address_space_init(&ppgtt->base, dev_priv);
2147 int i915_ppgtt_init_hw(struct drm_device *dev)
2149 /* In the case of execlists, PPGTT is enabled by the context descriptor
2150 * and the PDPs are contained within the context itself. We don't
2151 * need to do anything here. */
2152 if (i915.enable_execlists)
2155 if (!USES_PPGTT(dev))
2159 gen6_ppgtt_enable(dev);
2160 else if (IS_GEN7(dev))
2161 gen7_ppgtt_enable(dev);
2162 else if (INTEL_INFO(dev)->gen >= 8)
2163 gen8_ppgtt_enable(dev);
2165 MISSING_CASE(INTEL_INFO(dev)->gen);
2170 int i915_ppgtt_init_ring(struct drm_i915_gem_request *req)
2172 struct drm_i915_private *dev_priv = req->ring->dev->dev_private;
2173 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
2175 if (i915.enable_execlists)
2181 return ppgtt->switch_mm(ppgtt, req);
2184 struct i915_hw_ppgtt *
2185 i915_ppgtt_create(struct drm_device *dev, struct drm_i915_file_private *fpriv)
2187 struct i915_hw_ppgtt *ppgtt;
2190 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2192 return ERR_PTR(-ENOMEM);
2194 ret = i915_ppgtt_init(dev, ppgtt);
2197 return ERR_PTR(ret);
2200 ppgtt->file_priv = fpriv;
2202 trace_i915_ppgtt_create(&ppgtt->base);
2207 void i915_ppgtt_release(struct kref *kref)
2209 struct i915_hw_ppgtt *ppgtt =
2210 container_of(kref, struct i915_hw_ppgtt, ref);
2212 trace_i915_ppgtt_release(&ppgtt->base);
2214 /* vmas should already be unbound */
2215 WARN_ON(!list_empty(&ppgtt->base.active_list));
2216 WARN_ON(!list_empty(&ppgtt->base.inactive_list));
2218 list_del(&ppgtt->base.global_link);
2219 drm_mm_takedown(&ppgtt->base.mm);
2221 ppgtt->base.cleanup(&ppgtt->base);
2225 extern int intel_iommu_gfx_mapped;
2226 /* Certain Gen5 chipsets require require idling the GPU before
2227 * unmapping anything from the GTT when VT-d is enabled.
2229 static bool needs_idle_maps(struct drm_device *dev)
2231 #ifdef CONFIG_INTEL_IOMMU
2232 /* Query intel_iommu to see if we need the workaround. Presumably that
2235 if (IS_GEN5(dev) && IS_MOBILE(dev) && intel_iommu_gfx_mapped)
2241 static bool do_idling(struct drm_i915_private *dev_priv)
2243 bool ret = dev_priv->mm.interruptible;
2245 if (unlikely(dev_priv->gtt.do_idle_maps)) {
2246 dev_priv->mm.interruptible = false;
2247 if (i915_gpu_idle(dev_priv->dev)) {
2248 DRM_ERROR("Couldn't idle GPU\n");
2249 /* Wait a bit, in hopes it avoids the hang */
2257 static void undo_idling(struct drm_i915_private *dev_priv, bool interruptible)
2259 if (unlikely(dev_priv->gtt.do_idle_maps))
2260 dev_priv->mm.interruptible = interruptible;
2263 void i915_check_and_clear_faults(struct drm_device *dev)
2265 struct drm_i915_private *dev_priv = dev->dev_private;
2266 struct intel_engine_cs *ring;
2269 if (INTEL_INFO(dev)->gen < 6)
2272 for_each_ring(ring, dev_priv, i) {
2274 fault_reg = I915_READ(RING_FAULT_REG(ring));
2275 if (fault_reg & RING_FAULT_VALID) {
2276 DRM_DEBUG_DRIVER("Unexpected fault\n"
2278 "\tAddress space: %s\n"
2281 fault_reg & PAGE_MASK,
2282 fault_reg & RING_FAULT_GTTSEL_MASK ? "GGTT" : "PPGTT",
2283 RING_FAULT_SRCID(fault_reg),
2284 RING_FAULT_FAULT_TYPE(fault_reg));
2285 I915_WRITE(RING_FAULT_REG(ring),
2286 fault_reg & ~RING_FAULT_VALID);
2289 POSTING_READ(RING_FAULT_REG(&dev_priv->ring[RCS]));
2292 static void i915_ggtt_flush(struct drm_i915_private *dev_priv)
2294 if (INTEL_INFO(dev_priv->dev)->gen < 6) {
2295 intel_gtt_chipset_flush();
2297 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2298 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2302 void i915_gem_suspend_gtt_mappings(struct drm_device *dev)
2304 struct drm_i915_private *dev_priv = dev->dev_private;
2306 /* Don't bother messing with faults pre GEN6 as we have little
2307 * documentation supporting that it's a good idea.
2309 if (INTEL_INFO(dev)->gen < 6)
2312 i915_check_and_clear_faults(dev);
2314 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
2315 dev_priv->gtt.base.start,
2316 dev_priv->gtt.base.total,
2319 i915_ggtt_flush(dev_priv);
2322 int i915_gem_gtt_prepare_object(struct drm_i915_gem_object *obj)
2324 if (!dma_map_sg(&obj->base.dev->pdev->dev,
2325 obj->pages->sgl, obj->pages->nents,
2326 PCI_DMA_BIDIRECTIONAL))
2332 static void gen8_set_pte(void __iomem *addr, gen8_pte_t pte)
2337 iowrite32((u32)pte, addr);
2338 iowrite32(pte >> 32, addr + 4);
2342 static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
2343 struct sg_table *st,
2345 enum i915_cache_level level, u32 unused)
2347 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2348 unsigned first_entry = start >> PAGE_SHIFT;
2349 gen8_pte_t __iomem *gtt_entries =
2350 (gen8_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
2352 struct sg_page_iter sg_iter;
2353 dma_addr_t addr = 0; /* shut up gcc */
2356 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2358 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
2359 addr = sg_dma_address(sg_iter.sg) +
2360 (sg_iter.sg_pgoffset << PAGE_SHIFT);
2361 gen8_set_pte(>t_entries[i],
2362 gen8_pte_encode(addr, level, true));
2367 * XXX: This serves as a posting read to make sure that the PTE has
2368 * actually been updated. There is some concern that even though
2369 * registers and PTEs are within the same BAR that they are potentially
2370 * of NUMA access patterns. Therefore, even with the way we assume
2371 * hardware should work, we must keep this posting read for paranoia.
2374 WARN_ON(readq(>t_entries[i-1])
2375 != gen8_pte_encode(addr, level, true));
2377 /* This next bit makes the above posting read even more important. We
2378 * want to flush the TLBs only after we're certain all the PTE updates
2381 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2382 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2384 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2387 struct insert_entries {
2388 struct i915_address_space *vm;
2389 struct sg_table *st;
2391 enum i915_cache_level level;
2395 static int gen8_ggtt_insert_entries__cb(void *_arg)
2397 struct insert_entries *arg = _arg;
2398 gen8_ggtt_insert_entries(arg->vm, arg->st,
2399 arg->start, arg->level, arg->flags);
2403 static void gen8_ggtt_insert_entries__BKL(struct i915_address_space *vm,
2404 struct sg_table *st,
2406 enum i915_cache_level level,
2409 struct insert_entries arg = { vm, st, start, level, flags };
2410 stop_machine(gen8_ggtt_insert_entries__cb, &arg, NULL);
2414 * Binds an object into the global gtt with the specified cache level. The object
2415 * will be accessible to the GPU via commands whose operands reference offsets
2416 * within the global GTT as well as accessible by the GPU through the GMADR
2417 * mapped BAR (dev_priv->mm.gtt->gtt).
2419 static void gen6_ggtt_insert_entries(struct i915_address_space *vm,
2420 struct sg_table *st,
2422 enum i915_cache_level level, u32 flags)
2424 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2425 unsigned first_entry = start >> PAGE_SHIFT;
2426 gen6_pte_t __iomem *gtt_entries =
2427 (gen6_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
2429 struct sg_page_iter sg_iter;
2430 dma_addr_t addr = 0;
2433 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2435 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
2436 addr = sg_page_iter_dma_address(&sg_iter);
2437 iowrite32(vm->pte_encode(addr, level, true, flags), >t_entries[i]);
2441 /* XXX: This serves as a posting read to make sure that the PTE has
2442 * actually been updated. There is some concern that even though
2443 * registers and PTEs are within the same BAR that they are potentially
2444 * of NUMA access patterns. Therefore, even with the way we assume
2445 * hardware should work, we must keep this posting read for paranoia.
2448 unsigned long gtt = readl(>t_entries[i-1]);
2449 WARN_ON(gtt != vm->pte_encode(addr, level, true, flags));
2452 /* This next bit makes the above posting read even more important. We
2453 * want to flush the TLBs only after we're certain all the PTE updates
2456 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2457 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2459 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2462 static void gen8_ggtt_clear_range(struct i915_address_space *vm,
2467 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2468 unsigned first_entry = start >> PAGE_SHIFT;
2469 unsigned num_entries = length >> PAGE_SHIFT;
2470 gen8_pte_t scratch_pte, __iomem *gtt_base =
2471 (gen8_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
2472 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
2476 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2478 if (WARN(num_entries > max_entries,
2479 "First entry = %d; Num entries = %d (max=%d)\n",
2480 first_entry, num_entries, max_entries))
2481 num_entries = max_entries;
2483 scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
2486 for (i = 0; i < num_entries; i++)
2487 gen8_set_pte(>t_base[i], scratch_pte);
2490 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2493 static void gen6_ggtt_clear_range(struct i915_address_space *vm,
2498 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2499 unsigned first_entry = start >> PAGE_SHIFT;
2500 unsigned num_entries = length >> PAGE_SHIFT;
2501 gen6_pte_t scratch_pte, __iomem *gtt_base =
2502 (gen6_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
2503 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
2507 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2509 if (WARN(num_entries > max_entries,
2510 "First entry = %d; Num entries = %d (max=%d)\n",
2511 first_entry, num_entries, max_entries))
2512 num_entries = max_entries;
2514 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
2515 I915_CACHE_LLC, use_scratch, 0);
2517 for (i = 0; i < num_entries; i++)
2518 iowrite32(scratch_pte, >t_base[i]);
2521 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2524 static void i915_ggtt_insert_entries(struct i915_address_space *vm,
2525 struct sg_table *pages,
2527 enum i915_cache_level cache_level, u32 unused)
2529 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2530 unsigned int flags = (cache_level == I915_CACHE_NONE) ?
2531 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
2534 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2536 intel_gtt_insert_sg_entries(pages, start >> PAGE_SHIFT, flags);
2538 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2542 static void i915_ggtt_clear_range(struct i915_address_space *vm,
2547 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2548 unsigned first_entry = start >> PAGE_SHIFT;
2549 unsigned num_entries = length >> PAGE_SHIFT;
2552 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2554 intel_gtt_clear_range(first_entry, num_entries);
2556 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2559 static int ggtt_bind_vma(struct i915_vma *vma,
2560 enum i915_cache_level cache_level,
2563 struct drm_i915_gem_object *obj = vma->obj;
2567 ret = i915_get_ggtt_vma_pages(vma);
2571 /* Currently applicable only to VLV */
2573 pte_flags |= PTE_READ_ONLY;
2575 vma->vm->insert_entries(vma->vm, vma->ggtt_view.pages,
2577 cache_level, pte_flags);
2580 * Without aliasing PPGTT there's no difference between
2581 * GLOBAL/LOCAL_BIND, it's all the same ptes. Hence unconditionally
2582 * upgrade to both bound if we bind either to avoid double-binding.
2584 vma->bound |= GLOBAL_BIND | LOCAL_BIND;
2589 static int aliasing_gtt_bind_vma(struct i915_vma *vma,
2590 enum i915_cache_level cache_level,
2593 struct drm_device *dev = vma->vm->dev;
2594 struct drm_i915_private *dev_priv = dev->dev_private;
2595 struct drm_i915_gem_object *obj = vma->obj;
2596 struct sg_table *pages = obj->pages;
2600 ret = i915_get_ggtt_vma_pages(vma);
2603 pages = vma->ggtt_view.pages;
2605 /* Currently applicable only to VLV */
2607 pte_flags |= PTE_READ_ONLY;
2610 if (flags & GLOBAL_BIND) {
2611 vma->vm->insert_entries(vma->vm, pages,
2613 cache_level, pte_flags);
2616 if (flags & LOCAL_BIND) {
2617 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
2618 appgtt->base.insert_entries(&appgtt->base, pages,
2620 cache_level, pte_flags);
2626 static void ggtt_unbind_vma(struct i915_vma *vma)
2628 struct drm_device *dev = vma->vm->dev;
2629 struct drm_i915_private *dev_priv = dev->dev_private;
2630 struct drm_i915_gem_object *obj = vma->obj;
2631 const uint64_t size = min_t(uint64_t,
2635 if (vma->bound & GLOBAL_BIND) {
2636 vma->vm->clear_range(vma->vm,
2642 if (dev_priv->mm.aliasing_ppgtt && vma->bound & LOCAL_BIND) {
2643 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
2645 appgtt->base.clear_range(&appgtt->base,
2652 void i915_gem_gtt_finish_object(struct drm_i915_gem_object *obj)
2654 struct drm_device *dev = obj->base.dev;
2655 struct drm_i915_private *dev_priv = dev->dev_private;
2658 interruptible = do_idling(dev_priv);
2660 dma_unmap_sg(&dev->pdev->dev, obj->pages->sgl, obj->pages->nents,
2661 PCI_DMA_BIDIRECTIONAL);
2663 undo_idling(dev_priv, interruptible);
2666 static void i915_gtt_color_adjust(struct drm_mm_node *node,
2667 unsigned long color,
2671 if (node->color != color)
2674 if (!list_empty(&node->node_list)) {
2675 node = list_entry(node->node_list.next,
2678 if (node->allocated && node->color != color)
2683 static int i915_gem_setup_global_gtt(struct drm_device *dev,
2688 /* Let GEM Manage all of the aperture.
2690 * However, leave one page at the end still bound to the scratch page.
2691 * There are a number of places where the hardware apparently prefetches
2692 * past the end of the object, and we've seen multiple hangs with the
2693 * GPU head pointer stuck in a batchbuffer bound at the last page of the
2694 * aperture. One page should be enough to keep any prefetching inside
2697 struct drm_i915_private *dev_priv = dev->dev_private;
2698 struct i915_address_space *ggtt_vm = &dev_priv->gtt.base;
2699 struct drm_mm_node *entry;
2700 struct drm_i915_gem_object *obj;
2701 unsigned long hole_start, hole_end;
2704 BUG_ON(mappable_end > end);
2706 ggtt_vm->start = start;
2708 /* Subtract the guard page before address space initialization to
2709 * shrink the range used by drm_mm */
2710 ggtt_vm->total = end - start - PAGE_SIZE;
2711 i915_address_space_init(ggtt_vm, dev_priv);
2712 ggtt_vm->total += PAGE_SIZE;
2714 if (intel_vgpu_active(dev)) {
2715 ret = intel_vgt_balloon(dev);
2721 ggtt_vm->mm.color_adjust = i915_gtt_color_adjust;
2723 /* Mark any preallocated objects as occupied */
2724 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
2725 struct i915_vma *vma = i915_gem_obj_to_vma(obj, ggtt_vm);
2727 DRM_DEBUG_KMS("reserving preallocated space: %llx + %zx\n",
2728 i915_gem_obj_ggtt_offset(obj), obj->base.size);
2730 WARN_ON(i915_gem_obj_ggtt_bound(obj));
2731 ret = drm_mm_reserve_node(&ggtt_vm->mm, &vma->node);
2733 DRM_DEBUG_KMS("Reservation failed: %i\n", ret);
2736 vma->bound |= GLOBAL_BIND;
2737 __i915_vma_set_map_and_fenceable(vma);
2738 list_add_tail(&vma->mm_list, &ggtt_vm->inactive_list);
2741 /* Clear any non-preallocated blocks */
2742 drm_mm_for_each_hole(entry, &ggtt_vm->mm, hole_start, hole_end) {
2743 DRM_DEBUG_KMS("clearing unused GTT space: [%lx, %lx]\n",
2744 hole_start, hole_end);
2745 ggtt_vm->clear_range(ggtt_vm, hole_start,
2746 hole_end - hole_start, true);
2749 /* And finally clear the reserved guard page */
2750 ggtt_vm->clear_range(ggtt_vm, end - PAGE_SIZE, PAGE_SIZE, true);
2752 if (USES_PPGTT(dev) && !USES_FULL_PPGTT(dev)) {
2753 struct i915_hw_ppgtt *ppgtt;
2755 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2759 ret = __hw_ppgtt_init(dev, ppgtt);
2761 ppgtt->base.cleanup(&ppgtt->base);
2766 if (ppgtt->base.allocate_va_range)
2767 ret = ppgtt->base.allocate_va_range(&ppgtt->base, 0,
2770 ppgtt->base.cleanup(&ppgtt->base);
2775 ppgtt->base.clear_range(&ppgtt->base,
2780 dev_priv->mm.aliasing_ppgtt = ppgtt;
2781 WARN_ON(dev_priv->gtt.base.bind_vma != ggtt_bind_vma);
2782 dev_priv->gtt.base.bind_vma = aliasing_gtt_bind_vma;
2788 void i915_gem_init_global_gtt(struct drm_device *dev)
2790 struct drm_i915_private *dev_priv = dev->dev_private;
2791 u64 gtt_size, mappable_size;
2793 gtt_size = dev_priv->gtt.base.total;
2794 mappable_size = dev_priv->gtt.mappable_end;
2796 i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
2799 void i915_global_gtt_cleanup(struct drm_device *dev)
2801 struct drm_i915_private *dev_priv = dev->dev_private;
2802 struct i915_address_space *vm = &dev_priv->gtt.base;
2804 if (dev_priv->mm.aliasing_ppgtt) {
2805 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
2807 ppgtt->base.cleanup(&ppgtt->base);
2810 if (drm_mm_initialized(&vm->mm)) {
2811 if (intel_vgpu_active(dev))
2812 intel_vgt_deballoon();
2814 drm_mm_takedown(&vm->mm);
2815 list_del(&vm->global_link);
2821 static unsigned int gen6_get_total_gtt_size(u16 snb_gmch_ctl)
2823 snb_gmch_ctl >>= SNB_GMCH_GGMS_SHIFT;
2824 snb_gmch_ctl &= SNB_GMCH_GGMS_MASK;
2825 return snb_gmch_ctl << 20;
2828 static unsigned int gen8_get_total_gtt_size(u16 bdw_gmch_ctl)
2830 bdw_gmch_ctl >>= BDW_GMCH_GGMS_SHIFT;
2831 bdw_gmch_ctl &= BDW_GMCH_GGMS_MASK;
2833 bdw_gmch_ctl = 1 << bdw_gmch_ctl;
2835 #ifdef CONFIG_X86_32
2836 /* Limit 32b platforms to a 2GB GGTT: 4 << 20 / pte size * PAGE_SIZE */
2837 if (bdw_gmch_ctl > 4)
2841 return bdw_gmch_ctl << 20;
2844 static unsigned int chv_get_total_gtt_size(u16 gmch_ctrl)
2846 gmch_ctrl >>= SNB_GMCH_GGMS_SHIFT;
2847 gmch_ctrl &= SNB_GMCH_GGMS_MASK;
2850 return 1 << (20 + gmch_ctrl);
2855 static size_t gen6_get_stolen_size(u16 snb_gmch_ctl)
2857 snb_gmch_ctl >>= SNB_GMCH_GMS_SHIFT;
2858 snb_gmch_ctl &= SNB_GMCH_GMS_MASK;
2859 return snb_gmch_ctl << 25; /* 32 MB units */
2862 static size_t gen8_get_stolen_size(u16 bdw_gmch_ctl)
2864 bdw_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2865 bdw_gmch_ctl &= BDW_GMCH_GMS_MASK;
2866 return bdw_gmch_ctl << 25; /* 32 MB units */
2869 static size_t chv_get_stolen_size(u16 gmch_ctrl)
2871 gmch_ctrl >>= SNB_GMCH_GMS_SHIFT;
2872 gmch_ctrl &= SNB_GMCH_GMS_MASK;
2875 * 0x0 to 0x10: 32MB increments starting at 0MB
2876 * 0x11 to 0x16: 4MB increments starting at 8MB
2877 * 0x17 to 0x1d: 4MB increments start at 36MB
2879 if (gmch_ctrl < 0x11)
2880 return gmch_ctrl << 25;
2881 else if (gmch_ctrl < 0x17)
2882 return (gmch_ctrl - 0x11 + 2) << 22;
2884 return (gmch_ctrl - 0x17 + 9) << 22;
2887 static size_t gen9_get_stolen_size(u16 gen9_gmch_ctl)
2889 gen9_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2890 gen9_gmch_ctl &= BDW_GMCH_GMS_MASK;
2892 if (gen9_gmch_ctl < 0xf0)
2893 return gen9_gmch_ctl << 25; /* 32 MB units */
2895 /* 4MB increments starting at 0xf0 for 4MB */
2896 return (gen9_gmch_ctl - 0xf0 + 1) << 22;
2899 static int ggtt_probe_common(struct drm_device *dev,
2902 struct drm_i915_private *dev_priv = dev->dev_private;
2903 struct i915_page_scratch *scratch_page;
2904 phys_addr_t gtt_phys_addr;
2906 /* For Modern GENs the PTEs and register space are split in the BAR */
2907 gtt_phys_addr = pci_resource_start(dev->pdev, 0) +
2908 (pci_resource_len(dev->pdev, 0) / 2);
2911 * On BXT writes larger than 64 bit to the GTT pagetable range will be
2912 * dropped. For WC mappings in general we have 64 byte burst writes
2913 * when the WC buffer is flushed, so we can't use it, but have to
2914 * resort to an uncached mapping. The WC issue is easily caught by the
2915 * readback check when writing GTT PTE entries.
2917 if (IS_BROXTON(dev))
2918 dev_priv->gtt.gsm = ioremap_nocache(gtt_phys_addr, gtt_size);
2920 dev_priv->gtt.gsm = ioremap_wc(gtt_phys_addr, gtt_size);
2921 if (!dev_priv->gtt.gsm) {
2922 DRM_ERROR("Failed to map the gtt page table\n");
2926 scratch_page = alloc_scratch_page(dev);
2927 if (IS_ERR(scratch_page)) {
2928 DRM_ERROR("Scratch setup failed\n");
2929 /* iounmap will also get called at remove, but meh */
2930 iounmap(dev_priv->gtt.gsm);
2931 return PTR_ERR(scratch_page);
2934 dev_priv->gtt.base.scratch_page = scratch_page;
2939 /* The GGTT and PPGTT need a private PPAT setup in order to handle cacheability
2940 * bits. When using advanced contexts each context stores its own PAT, but
2941 * writing this data shouldn't be harmful even in those cases. */
2942 static void bdw_setup_private_ppat(struct drm_i915_private *dev_priv)
2946 pat = GEN8_PPAT(0, GEN8_PPAT_WB | GEN8_PPAT_LLC) | /* for normal objects, no eLLC */
2947 GEN8_PPAT(1, GEN8_PPAT_WC | GEN8_PPAT_LLCELLC) | /* for something pointing to ptes? */
2948 GEN8_PPAT(2, GEN8_PPAT_WT | GEN8_PPAT_LLCELLC) | /* for scanout with eLLC */
2949 GEN8_PPAT(3, GEN8_PPAT_UC) | /* Uncached objects, mostly for scanout */
2950 GEN8_PPAT(4, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(0)) |
2951 GEN8_PPAT(5, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(1)) |
2952 GEN8_PPAT(6, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(2)) |
2953 GEN8_PPAT(7, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3));
2955 if (!USES_PPGTT(dev_priv->dev))
2956 /* Spec: "For GGTT, there is NO pat_sel[2:0] from the entry,
2957 * so RTL will always use the value corresponding to
2959 * So let's disable cache for GGTT to avoid screen corruptions.
2960 * MOCS still can be used though.
2961 * - System agent ggtt writes (i.e. cpu gtt mmaps) already work
2962 * before this patch, i.e. the same uncached + snooping access
2963 * like on gen6/7 seems to be in effect.
2964 * - So this just fixes blitter/render access. Again it looks
2965 * like it's not just uncached access, but uncached + snooping.
2966 * So we can still hold onto all our assumptions wrt cpu
2967 * clflushing on LLC machines.
2969 pat = GEN8_PPAT(0, GEN8_PPAT_UC);
2971 /* XXX: spec defines this as 2 distinct registers. It's unclear if a 64b
2972 * write would work. */
2973 I915_WRITE(GEN8_PRIVATE_PAT_LO, pat);
2974 I915_WRITE(GEN8_PRIVATE_PAT_HI, pat >> 32);
2977 static void chv_setup_private_ppat(struct drm_i915_private *dev_priv)
2982 * Map WB on BDW to snooped on CHV.
2984 * Only the snoop bit has meaning for CHV, the rest is
2987 * The hardware will never snoop for certain types of accesses:
2988 * - CPU GTT (GMADR->GGTT->no snoop->memory)
2989 * - PPGTT page tables
2990 * - some other special cycles
2992 * As with BDW, we also need to consider the following for GT accesses:
2993 * "For GGTT, there is NO pat_sel[2:0] from the entry,
2994 * so RTL will always use the value corresponding to
2996 * Which means we must set the snoop bit in PAT entry 0
2997 * in order to keep the global status page working.
2999 pat = GEN8_PPAT(0, CHV_PPAT_SNOOP) |
3003 GEN8_PPAT(4, CHV_PPAT_SNOOP) |
3004 GEN8_PPAT(5, CHV_PPAT_SNOOP) |
3005 GEN8_PPAT(6, CHV_PPAT_SNOOP) |
3006 GEN8_PPAT(7, CHV_PPAT_SNOOP);
3008 I915_WRITE(GEN8_PRIVATE_PAT_LO, pat);
3009 I915_WRITE(GEN8_PRIVATE_PAT_HI, pat >> 32);
3012 static int gen8_gmch_probe(struct drm_device *dev,
3015 phys_addr_t *mappable_base,
3018 struct drm_i915_private *dev_priv = dev->dev_private;
3023 /* TODO: We're not aware of mappable constraints on gen8 yet */
3024 *mappable_base = pci_resource_start(dev->pdev, 2);
3025 *mappable_end = pci_resource_len(dev->pdev, 2);
3027 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(39)))
3028 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(39));
3030 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
3032 if (INTEL_INFO(dev)->gen >= 9) {
3033 *stolen = gen9_get_stolen_size(snb_gmch_ctl);
3034 gtt_size = gen8_get_total_gtt_size(snb_gmch_ctl);
3035 } else if (IS_CHERRYVIEW(dev)) {
3036 *stolen = chv_get_stolen_size(snb_gmch_ctl);
3037 gtt_size = chv_get_total_gtt_size(snb_gmch_ctl);
3039 *stolen = gen8_get_stolen_size(snb_gmch_ctl);
3040 gtt_size = gen8_get_total_gtt_size(snb_gmch_ctl);
3043 *gtt_total = (gtt_size / sizeof(gen8_pte_t)) << PAGE_SHIFT;
3045 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
3046 chv_setup_private_ppat(dev_priv);
3048 bdw_setup_private_ppat(dev_priv);
3050 ret = ggtt_probe_common(dev, gtt_size);
3052 dev_priv->gtt.base.clear_range = gen8_ggtt_clear_range;
3053 dev_priv->gtt.base.insert_entries = gen8_ggtt_insert_entries;
3054 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
3055 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
3057 if (IS_CHERRYVIEW(dev_priv))
3058 dev_priv->gtt.base.insert_entries = gen8_ggtt_insert_entries__BKL;
3063 static int gen6_gmch_probe(struct drm_device *dev,
3066 phys_addr_t *mappable_base,
3069 struct drm_i915_private *dev_priv = dev->dev_private;
3070 unsigned int gtt_size;
3074 *mappable_base = pci_resource_start(dev->pdev, 2);
3075 *mappable_end = pci_resource_len(dev->pdev, 2);
3077 /* 64/512MB is the current min/max we actually know of, but this is just
3078 * a coarse sanity check.
3080 if ((*mappable_end < (64<<20) || (*mappable_end > (512<<20)))) {
3081 DRM_ERROR("Unknown GMADR size (%llx)\n",
3082 dev_priv->gtt.mappable_end);
3086 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(40)))
3087 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(40));
3088 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
3090 *stolen = gen6_get_stolen_size(snb_gmch_ctl);
3092 gtt_size = gen6_get_total_gtt_size(snb_gmch_ctl);
3093 *gtt_total = (gtt_size / sizeof(gen6_pte_t)) << PAGE_SHIFT;
3095 ret = ggtt_probe_common(dev, gtt_size);
3097 dev_priv->gtt.base.clear_range = gen6_ggtt_clear_range;
3098 dev_priv->gtt.base.insert_entries = gen6_ggtt_insert_entries;
3099 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
3100 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
3105 static void gen6_gmch_remove(struct i915_address_space *vm)
3108 struct i915_gtt *gtt = container_of(vm, struct i915_gtt, base);
3111 free_scratch_page(vm->dev, vm->scratch_page);
3114 static int i915_gmch_probe(struct drm_device *dev,
3117 phys_addr_t *mappable_base,
3120 struct drm_i915_private *dev_priv = dev->dev_private;
3123 ret = intel_gmch_probe(dev_priv->bridge_dev, dev_priv->dev->pdev, NULL);
3125 DRM_ERROR("failed to set up gmch\n");
3129 intel_gtt_get(gtt_total, stolen, mappable_base, mappable_end);
3131 dev_priv->gtt.do_idle_maps = needs_idle_maps(dev_priv->dev);
3132 dev_priv->gtt.base.insert_entries = i915_ggtt_insert_entries;
3133 dev_priv->gtt.base.clear_range = i915_ggtt_clear_range;
3134 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
3135 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
3137 if (unlikely(dev_priv->gtt.do_idle_maps))
3138 DRM_INFO("applying Ironlake quirks for intel_iommu\n");
3143 static void i915_gmch_remove(struct i915_address_space *vm)
3145 intel_gmch_remove();
3148 int i915_gem_gtt_init(struct drm_device *dev)
3150 struct drm_i915_private *dev_priv = dev->dev_private;
3151 struct i915_gtt *gtt = &dev_priv->gtt;
3154 if (INTEL_INFO(dev)->gen <= 5) {
3155 gtt->gtt_probe = i915_gmch_probe;
3156 gtt->base.cleanup = i915_gmch_remove;
3157 } else if (INTEL_INFO(dev)->gen < 8) {
3158 gtt->gtt_probe = gen6_gmch_probe;
3159 gtt->base.cleanup = gen6_gmch_remove;
3160 if (IS_HASWELL(dev) && dev_priv->ellc_size)
3161 gtt->base.pte_encode = iris_pte_encode;
3162 else if (IS_HASWELL(dev))
3163 gtt->base.pte_encode = hsw_pte_encode;
3164 else if (IS_VALLEYVIEW(dev))
3165 gtt->base.pte_encode = byt_pte_encode;
3166 else if (INTEL_INFO(dev)->gen >= 7)
3167 gtt->base.pte_encode = ivb_pte_encode;
3169 gtt->base.pte_encode = snb_pte_encode;
3171 dev_priv->gtt.gtt_probe = gen8_gmch_probe;
3172 dev_priv->gtt.base.cleanup = gen6_gmch_remove;
3175 gtt->base.dev = dev;
3177 ret = gtt->gtt_probe(dev, >t->base.total, >t->stolen_size,
3178 >t->mappable_base, >t->mappable_end);
3182 /* GMADR is the PCI mmio aperture into the global GTT. */
3183 DRM_INFO("Memory usable by graphics device = %lluM\n",
3184 gtt->base.total >> 20);
3185 DRM_DEBUG_DRIVER("GMADR size = %lldM\n", gtt->mappable_end >> 20);
3186 DRM_DEBUG_DRIVER("GTT stolen size = %zdM\n", gtt->stolen_size >> 20);
3187 #ifdef CONFIG_INTEL_IOMMU
3188 if (intel_iommu_gfx_mapped)
3189 DRM_INFO("VT-d active for gfx access\n");
3192 * i915.enable_ppgtt is read-only, so do an early pass to validate the
3193 * user's requested state against the hardware/driver capabilities. We
3194 * do this now so that we can print out any log messages once rather
3195 * than every time we check intel_enable_ppgtt().
3197 i915.enable_ppgtt = sanitize_enable_ppgtt(dev, i915.enable_ppgtt);
3198 DRM_DEBUG_DRIVER("ppgtt mode: %i\n", i915.enable_ppgtt);
3203 void i915_gem_restore_gtt_mappings(struct drm_device *dev)
3205 struct drm_i915_private *dev_priv = dev->dev_private;
3206 struct drm_i915_gem_object *obj;
3207 struct i915_address_space *vm;
3208 struct i915_vma *vma;
3211 i915_check_and_clear_faults(dev);
3213 /* First fill our portion of the GTT with scratch pages */
3214 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
3215 dev_priv->gtt.base.start,
3216 dev_priv->gtt.base.total,
3219 /* Cache flush objects bound into GGTT and rebind them. */
3220 vm = &dev_priv->gtt.base;
3221 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
3223 list_for_each_entry(vma, &obj->vma_list, vma_link) {
3227 WARN_ON(i915_vma_bind(vma, obj->cache_level,
3234 i915_gem_clflush_object(obj, obj->pin_display);
3237 if (INTEL_INFO(dev)->gen >= 8) {
3238 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
3239 chv_setup_private_ppat(dev_priv);
3241 bdw_setup_private_ppat(dev_priv);
3246 if (USES_PPGTT(dev)) {
3247 list_for_each_entry(vm, &dev_priv->vm_list, global_link) {
3248 /* TODO: Perhaps it shouldn't be gen6 specific */
3250 struct i915_hw_ppgtt *ppgtt =
3251 container_of(vm, struct i915_hw_ppgtt,
3254 if (i915_is_ggtt(vm))
3255 ppgtt = dev_priv->mm.aliasing_ppgtt;
3257 gen6_write_page_range(dev_priv, &ppgtt->pd,
3258 0, ppgtt->base.total);
3262 i915_ggtt_flush(dev_priv);
3265 static struct i915_vma *
3266 __i915_gem_vma_create(struct drm_i915_gem_object *obj,
3267 struct i915_address_space *vm,
3268 const struct i915_ggtt_view *ggtt_view)
3270 struct i915_vma *vma;
3272 if (WARN_ON(i915_is_ggtt(vm) != !!ggtt_view))
3273 return ERR_PTR(-EINVAL);
3275 vma = kmem_cache_zalloc(to_i915(obj->base.dev)->vmas, GFP_KERNEL);
3277 return ERR_PTR(-ENOMEM);
3279 INIT_LIST_HEAD(&vma->vma_link);
3280 INIT_LIST_HEAD(&vma->mm_list);
3281 INIT_LIST_HEAD(&vma->exec_list);
3285 if (i915_is_ggtt(vm))
3286 vma->ggtt_view = *ggtt_view;
3288 list_add_tail(&vma->vma_link, &obj->vma_list);
3289 if (!i915_is_ggtt(vm))
3290 i915_ppgtt_get(i915_vm_to_ppgtt(vm));
3296 i915_gem_obj_lookup_or_create_vma(struct drm_i915_gem_object *obj,
3297 struct i915_address_space *vm)
3299 struct i915_vma *vma;
3301 vma = i915_gem_obj_to_vma(obj, vm);
3303 vma = __i915_gem_vma_create(obj, vm,
3304 i915_is_ggtt(vm) ? &i915_ggtt_view_normal : NULL);
3310 i915_gem_obj_lookup_or_create_ggtt_vma(struct drm_i915_gem_object *obj,
3311 const struct i915_ggtt_view *view)
3313 struct i915_address_space *ggtt = i915_obj_to_ggtt(obj);
3314 struct i915_vma *vma;
3317 return ERR_PTR(-EINVAL);
3319 vma = i915_gem_obj_to_ggtt_view(obj, view);
3325 vma = __i915_gem_vma_create(obj, ggtt, view);
3331 static struct scatterlist *
3332 rotate_pages(dma_addr_t *in, unsigned int offset,
3333 unsigned int width, unsigned int height,
3334 struct sg_table *st, struct scatterlist *sg)
3336 unsigned int column, row;
3337 unsigned int src_idx;
3344 for (column = 0; column < width; column++) {
3345 src_idx = width * (height - 1) + column;
3346 for (row = 0; row < height; row++) {
3348 /* We don't need the pages, but need to initialize
3349 * the entries so the sg list can be happily traversed.
3350 * The only thing we need are DMA addresses.
3352 sg_set_page(sg, NULL, PAGE_SIZE, 0);
3353 sg_dma_address(sg) = in[offset + src_idx];
3354 sg_dma_len(sg) = PAGE_SIZE;
3363 static struct sg_table *
3364 intel_rotate_fb_obj_pages(struct i915_ggtt_view *ggtt_view,
3365 struct drm_i915_gem_object *obj)
3367 struct intel_rotation_info *rot_info = &ggtt_view->params.rotation_info;
3368 unsigned int size_pages = rot_info->size >> PAGE_SHIFT;
3369 unsigned int size_pages_uv;
3370 struct sg_page_iter sg_iter;
3372 dma_addr_t *page_addr_list;
3373 struct sg_table *st;
3374 unsigned int uv_start_page;
3375 struct scatterlist *sg;
3378 /* Allocate a temporary list of source pages for random access. */
3379 page_addr_list = drm_malloc_ab(obj->base.size / PAGE_SIZE,
3380 sizeof(dma_addr_t));
3381 if (!page_addr_list)
3382 return ERR_PTR(ret);
3384 /* Account for UV plane with NV12. */
3385 if (rot_info->pixel_format == DRM_FORMAT_NV12)
3386 size_pages_uv = rot_info->size_uv >> PAGE_SHIFT;
3390 /* Allocate target SG list. */
3391 st = kmalloc(sizeof(*st), GFP_KERNEL);
3395 ret = sg_alloc_table(st, size_pages + size_pages_uv, GFP_KERNEL);
3399 /* Populate source page list from the object. */
3401 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) {
3402 page_addr_list[i] = sg_page_iter_dma_address(&sg_iter);
3406 /* Rotate the pages. */
3407 sg = rotate_pages(page_addr_list, 0,
3408 rot_info->width_pages, rot_info->height_pages,
3411 /* Append the UV plane if NV12. */
3412 if (rot_info->pixel_format == DRM_FORMAT_NV12) {
3413 uv_start_page = size_pages;
3415 /* Check for tile-row un-alignment. */
3416 if (offset_in_page(rot_info->uv_offset))
3419 rot_info->uv_start_page = uv_start_page;
3421 rotate_pages(page_addr_list, uv_start_page,
3422 rot_info->width_pages_uv,
3423 rot_info->height_pages_uv,
3428 "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",
3429 obj->base.size, rot_info->pitch, rot_info->height,
3430 rot_info->pixel_format, rot_info->width_pages,
3431 rot_info->height_pages, size_pages + size_pages_uv,
3434 drm_free_large(page_addr_list);
3441 drm_free_large(page_addr_list);
3444 "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",
3445 obj->base.size, ret, rot_info->pitch, rot_info->height,
3446 rot_info->pixel_format, rot_info->width_pages,
3447 rot_info->height_pages, size_pages + size_pages_uv,
3449 return ERR_PTR(ret);
3452 static struct sg_table *
3453 intel_partial_pages(const struct i915_ggtt_view *view,
3454 struct drm_i915_gem_object *obj)
3456 struct sg_table *st;
3457 struct scatterlist *sg;
3458 struct sg_page_iter obj_sg_iter;
3461 st = kmalloc(sizeof(*st), GFP_KERNEL);
3465 ret = sg_alloc_table(st, view->params.partial.size, GFP_KERNEL);
3471 for_each_sg_page(obj->pages->sgl, &obj_sg_iter, obj->pages->nents,
3472 view->params.partial.offset)
3474 if (st->nents >= view->params.partial.size)
3477 sg_set_page(sg, NULL, PAGE_SIZE, 0);
3478 sg_dma_address(sg) = sg_page_iter_dma_address(&obj_sg_iter);
3479 sg_dma_len(sg) = PAGE_SIZE;
3490 return ERR_PTR(ret);
3494 i915_get_ggtt_vma_pages(struct i915_vma *vma)
3498 if (vma->ggtt_view.pages)
3501 if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL)
3502 vma->ggtt_view.pages = vma->obj->pages;
3503 else if (vma->ggtt_view.type == I915_GGTT_VIEW_ROTATED)
3504 vma->ggtt_view.pages =
3505 intel_rotate_fb_obj_pages(&vma->ggtt_view, vma->obj);
3506 else if (vma->ggtt_view.type == I915_GGTT_VIEW_PARTIAL)
3507 vma->ggtt_view.pages =
3508 intel_partial_pages(&vma->ggtt_view, vma->obj);
3510 WARN_ONCE(1, "GGTT view %u not implemented!\n",
3511 vma->ggtt_view.type);
3513 if (!vma->ggtt_view.pages) {
3514 DRM_ERROR("Failed to get pages for GGTT view type %u!\n",
3515 vma->ggtt_view.type);
3517 } else if (IS_ERR(vma->ggtt_view.pages)) {
3518 ret = PTR_ERR(vma->ggtt_view.pages);
3519 vma->ggtt_view.pages = NULL;
3520 DRM_ERROR("Failed to get pages for VMA view type %u (%d)!\n",
3521 vma->ggtt_view.type, ret);
3528 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
3530 * @cache_level: mapping cache level
3531 * @flags: flags like global or local mapping
3533 * DMA addresses are taken from the scatter-gather table of this object (or of
3534 * this VMA in case of non-default GGTT views) and PTE entries set up.
3535 * Note that DMA addresses are also the only part of the SG table we care about.
3537 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
3543 if (WARN_ON(flags == 0))
3547 if (flags & PIN_GLOBAL)
3548 bind_flags |= GLOBAL_BIND;
3549 if (flags & PIN_USER)
3550 bind_flags |= LOCAL_BIND;
3552 if (flags & PIN_UPDATE)
3553 bind_flags |= vma->bound;
3555 bind_flags &= ~vma->bound;
3557 if (bind_flags == 0)
3560 if (vma->bound == 0 && vma->vm->allocate_va_range) {
3561 trace_i915_va_alloc(vma->vm,
3564 VM_TO_TRACE_NAME(vma->vm));
3566 /* XXX: i915_vma_pin() will fix this +- hack */
3568 ret = vma->vm->allocate_va_range(vma->vm,
3576 ret = vma->vm->bind_vma(vma, cache_level, bind_flags);
3580 vma->bound |= bind_flags;
3586 * i915_ggtt_view_size - Get the size of a GGTT view.
3587 * @obj: Object the view is of.
3588 * @view: The view in question.
3590 * @return The size of the GGTT view in bytes.
3593 i915_ggtt_view_size(struct drm_i915_gem_object *obj,
3594 const struct i915_ggtt_view *view)
3596 if (view->type == I915_GGTT_VIEW_NORMAL) {
3597 return obj->base.size;
3598 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
3599 return view->params.rotation_info.size;
3600 } else if (view->type == I915_GGTT_VIEW_PARTIAL) {
3601 return view->params.partial.size << PAGE_SHIFT;
3603 WARN_ONCE(1, "GGTT view %u not implemented!\n", view->type);
3604 return obj->base.size;