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 .type = I915_GGTT_VIEW_NORMAL,
102 const struct i915_ggtt_view i915_ggtt_view_rotated = {
103 .type = I915_GGTT_VIEW_ROTATED,
106 static int sanitize_enable_ppgtt(struct drm_device *dev, int enable_ppgtt)
108 bool has_aliasing_ppgtt;
110 bool has_full_48bit_ppgtt;
112 has_aliasing_ppgtt = INTEL_INFO(dev)->gen >= 6;
113 has_full_ppgtt = INTEL_INFO(dev)->gen >= 7;
114 has_full_48bit_ppgtt = IS_BROADWELL(dev) || INTEL_INFO(dev)->gen >= 9;
116 if (intel_vgpu_active(dev))
117 has_full_ppgtt = false; /* emulation is too hard */
120 * We don't allow disabling PPGTT for gen9+ as it's a requirement for
121 * execlists, the sole mechanism available to submit work.
123 if (INTEL_INFO(dev)->gen < 9 &&
124 (enable_ppgtt == 0 || !has_aliasing_ppgtt))
127 if (enable_ppgtt == 1)
130 if (enable_ppgtt == 2 && has_full_ppgtt)
133 if (enable_ppgtt == 3 && has_full_48bit_ppgtt)
136 #ifdef CONFIG_INTEL_IOMMU
137 /* Disable ppgtt on SNB if VT-d is on. */
138 if (INTEL_INFO(dev)->gen == 6 && intel_iommu_gfx_mapped) {
139 DRM_INFO("Disabling PPGTT because VT-d is on\n");
144 /* Early VLV doesn't have this */
145 if (IS_VALLEYVIEW(dev) && dev->pdev->revision < 0xb) {
146 DRM_DEBUG_DRIVER("disabling PPGTT on pre-B3 step VLV\n");
150 if (INTEL_INFO(dev)->gen >= 8 && i915.enable_execlists)
151 return has_full_48bit_ppgtt ? 3 : 2;
153 return has_aliasing_ppgtt ? 1 : 0;
156 static int ppgtt_bind_vma(struct i915_vma *vma,
157 enum i915_cache_level cache_level,
162 /* Currently applicable only to VLV */
164 pte_flags |= PTE_READ_ONLY;
166 vma->vm->insert_entries(vma->vm, vma->obj->pages, vma->node.start,
167 cache_level, pte_flags);
172 static void ppgtt_unbind_vma(struct i915_vma *vma)
174 vma->vm->clear_range(vma->vm,
180 static gen8_pte_t gen8_pte_encode(dma_addr_t addr,
181 enum i915_cache_level level,
184 gen8_pte_t pte = valid ? _PAGE_PRESENT | _PAGE_RW : 0;
188 case I915_CACHE_NONE:
189 pte |= PPAT_UNCACHED_INDEX;
192 pte |= PPAT_DISPLAY_ELLC_INDEX;
195 pte |= PPAT_CACHED_INDEX;
202 static gen8_pde_t gen8_pde_encode(const dma_addr_t addr,
203 const enum i915_cache_level level)
205 gen8_pde_t pde = _PAGE_PRESENT | _PAGE_RW;
207 if (level != I915_CACHE_NONE)
208 pde |= PPAT_CACHED_PDE_INDEX;
210 pde |= PPAT_UNCACHED_INDEX;
214 #define gen8_pdpe_encode gen8_pde_encode
215 #define gen8_pml4e_encode gen8_pde_encode
217 static gen6_pte_t snb_pte_encode(dma_addr_t addr,
218 enum i915_cache_level level,
219 bool valid, u32 unused)
221 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
222 pte |= GEN6_PTE_ADDR_ENCODE(addr);
225 case I915_CACHE_L3_LLC:
227 pte |= GEN6_PTE_CACHE_LLC;
229 case I915_CACHE_NONE:
230 pte |= GEN6_PTE_UNCACHED;
239 static gen6_pte_t ivb_pte_encode(dma_addr_t addr,
240 enum i915_cache_level level,
241 bool valid, u32 unused)
243 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
244 pte |= GEN6_PTE_ADDR_ENCODE(addr);
247 case I915_CACHE_L3_LLC:
248 pte |= GEN7_PTE_CACHE_L3_LLC;
251 pte |= GEN6_PTE_CACHE_LLC;
253 case I915_CACHE_NONE:
254 pte |= GEN6_PTE_UNCACHED;
263 static gen6_pte_t byt_pte_encode(dma_addr_t addr,
264 enum i915_cache_level level,
265 bool valid, u32 flags)
267 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
268 pte |= GEN6_PTE_ADDR_ENCODE(addr);
270 if (!(flags & PTE_READ_ONLY))
271 pte |= BYT_PTE_WRITEABLE;
273 if (level != I915_CACHE_NONE)
274 pte |= BYT_PTE_SNOOPED_BY_CPU_CACHES;
279 static gen6_pte_t hsw_pte_encode(dma_addr_t addr,
280 enum i915_cache_level level,
281 bool valid, u32 unused)
283 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
284 pte |= HSW_PTE_ADDR_ENCODE(addr);
286 if (level != I915_CACHE_NONE)
287 pte |= HSW_WB_LLC_AGE3;
292 static gen6_pte_t iris_pte_encode(dma_addr_t addr,
293 enum i915_cache_level level,
294 bool valid, u32 unused)
296 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
297 pte |= HSW_PTE_ADDR_ENCODE(addr);
300 case I915_CACHE_NONE:
303 pte |= HSW_WT_ELLC_LLC_AGE3;
306 pte |= HSW_WB_ELLC_LLC_AGE3;
313 static int __setup_page_dma(struct drm_device *dev,
314 struct i915_page_dma *p, gfp_t flags)
316 struct device *device = &dev->pdev->dev;
318 p->page = alloc_page(flags);
322 p->daddr = dma_map_page(device,
323 p->page, 0, 4096, PCI_DMA_BIDIRECTIONAL);
325 if (dma_mapping_error(device, p->daddr)) {
326 __free_page(p->page);
333 static int setup_page_dma(struct drm_device *dev, struct i915_page_dma *p)
335 return __setup_page_dma(dev, p, GFP_KERNEL);
338 static void cleanup_page_dma(struct drm_device *dev, struct i915_page_dma *p)
340 if (WARN_ON(!p->page))
343 dma_unmap_page(&dev->pdev->dev, p->daddr, 4096, PCI_DMA_BIDIRECTIONAL);
344 __free_page(p->page);
345 memset(p, 0, sizeof(*p));
348 static void *kmap_page_dma(struct i915_page_dma *p)
350 return kmap_atomic(p->page);
353 /* We use the flushing unmap only with ppgtt structures:
354 * page directories, page tables and scratch pages.
356 static void kunmap_page_dma(struct drm_device *dev, void *vaddr)
358 /* There are only few exceptions for gen >=6. chv and bxt.
359 * And we are not sure about the latter so play safe for now.
361 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
362 drm_clflush_virt_range(vaddr, PAGE_SIZE);
364 kunmap_atomic(vaddr);
367 #define kmap_px(px) kmap_page_dma(px_base(px))
368 #define kunmap_px(ppgtt, vaddr) kunmap_page_dma((ppgtt)->base.dev, (vaddr))
370 #define setup_px(dev, px) setup_page_dma((dev), px_base(px))
371 #define cleanup_px(dev, px) cleanup_page_dma((dev), px_base(px))
372 #define fill_px(dev, px, v) fill_page_dma((dev), px_base(px), (v))
373 #define fill32_px(dev, px, v) fill_page_dma_32((dev), px_base(px), (v))
375 static void fill_page_dma(struct drm_device *dev, struct i915_page_dma *p,
379 uint64_t * const vaddr = kmap_page_dma(p);
381 for (i = 0; i < 512; i++)
384 kunmap_page_dma(dev, vaddr);
387 static void fill_page_dma_32(struct drm_device *dev, struct i915_page_dma *p,
388 const uint32_t val32)
394 fill_page_dma(dev, p, v);
397 static struct i915_page_scratch *alloc_scratch_page(struct drm_device *dev)
399 struct i915_page_scratch *sp;
402 sp = kzalloc(sizeof(*sp), GFP_KERNEL);
404 return ERR_PTR(-ENOMEM);
406 ret = __setup_page_dma(dev, px_base(sp), GFP_DMA32 | __GFP_ZERO);
412 set_pages_uc(px_page(sp), 1);
417 static void free_scratch_page(struct drm_device *dev,
418 struct i915_page_scratch *sp)
420 set_pages_wb(px_page(sp), 1);
426 static struct i915_page_table *alloc_pt(struct drm_device *dev)
428 struct i915_page_table *pt;
429 const size_t count = INTEL_INFO(dev)->gen >= 8 ?
430 GEN8_PTES : GEN6_PTES;
433 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
435 return ERR_PTR(-ENOMEM);
437 pt->used_ptes = kcalloc(BITS_TO_LONGS(count), sizeof(*pt->used_ptes),
443 ret = setup_px(dev, pt);
450 kfree(pt->used_ptes);
457 static void free_pt(struct drm_device *dev, struct i915_page_table *pt)
460 kfree(pt->used_ptes);
464 static void gen8_initialize_pt(struct i915_address_space *vm,
465 struct i915_page_table *pt)
467 gen8_pte_t scratch_pte;
469 scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
470 I915_CACHE_LLC, true);
472 fill_px(vm->dev, pt, scratch_pte);
475 static void gen6_initialize_pt(struct i915_address_space *vm,
476 struct i915_page_table *pt)
478 gen6_pte_t scratch_pte;
480 WARN_ON(px_dma(vm->scratch_page) == 0);
482 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
483 I915_CACHE_LLC, true, 0);
485 fill32_px(vm->dev, pt, scratch_pte);
488 static struct i915_page_directory *alloc_pd(struct drm_device *dev)
490 struct i915_page_directory *pd;
493 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
495 return ERR_PTR(-ENOMEM);
497 pd->used_pdes = kcalloc(BITS_TO_LONGS(I915_PDES),
498 sizeof(*pd->used_pdes), GFP_KERNEL);
502 ret = setup_px(dev, pd);
509 kfree(pd->used_pdes);
516 static void free_pd(struct drm_device *dev, struct i915_page_directory *pd)
520 kfree(pd->used_pdes);
525 static void gen8_initialize_pd(struct i915_address_space *vm,
526 struct i915_page_directory *pd)
528 gen8_pde_t scratch_pde;
530 scratch_pde = gen8_pde_encode(px_dma(vm->scratch_pt), I915_CACHE_LLC);
532 fill_px(vm->dev, pd, scratch_pde);
535 static int __pdp_init(struct drm_device *dev,
536 struct i915_page_directory_pointer *pdp)
538 size_t pdpes = I915_PDPES_PER_PDP(dev);
540 pdp->used_pdpes = kcalloc(BITS_TO_LONGS(pdpes),
541 sizeof(unsigned long),
543 if (!pdp->used_pdpes)
546 pdp->page_directory = kcalloc(pdpes, sizeof(*pdp->page_directory),
548 if (!pdp->page_directory) {
549 kfree(pdp->used_pdpes);
550 /* the PDP might be the statically allocated top level. Keep it
551 * as clean as possible */
552 pdp->used_pdpes = NULL;
559 static void __pdp_fini(struct i915_page_directory_pointer *pdp)
561 kfree(pdp->used_pdpes);
562 kfree(pdp->page_directory);
563 pdp->page_directory = NULL;
567 i915_page_directory_pointer *alloc_pdp(struct drm_device *dev)
569 struct i915_page_directory_pointer *pdp;
572 WARN_ON(!USES_FULL_48BIT_PPGTT(dev));
574 pdp = kzalloc(sizeof(*pdp), GFP_KERNEL);
576 return ERR_PTR(-ENOMEM);
578 ret = __pdp_init(dev, pdp);
582 ret = setup_px(dev, pdp);
596 static void free_pdp(struct drm_device *dev,
597 struct i915_page_directory_pointer *pdp)
600 if (USES_FULL_48BIT_PPGTT(dev)) {
601 cleanup_px(dev, pdp);
606 static void gen8_initialize_pdp(struct i915_address_space *vm,
607 struct i915_page_directory_pointer *pdp)
609 gen8_ppgtt_pdpe_t scratch_pdpe;
611 scratch_pdpe = gen8_pdpe_encode(px_dma(vm->scratch_pd), I915_CACHE_LLC);
613 fill_px(vm->dev, pdp, scratch_pdpe);
616 static void gen8_initialize_pml4(struct i915_address_space *vm,
617 struct i915_pml4 *pml4)
619 gen8_ppgtt_pml4e_t scratch_pml4e;
621 scratch_pml4e = gen8_pml4e_encode(px_dma(vm->scratch_pdp),
624 fill_px(vm->dev, pml4, scratch_pml4e);
628 gen8_setup_page_directory(struct i915_hw_ppgtt *ppgtt,
629 struct i915_page_directory_pointer *pdp,
630 struct i915_page_directory *pd,
633 gen8_ppgtt_pdpe_t *page_directorypo;
635 if (!USES_FULL_48BIT_PPGTT(ppgtt->base.dev))
638 page_directorypo = kmap_px(pdp);
639 page_directorypo[index] = gen8_pdpe_encode(px_dma(pd), I915_CACHE_LLC);
640 kunmap_px(ppgtt, page_directorypo);
644 gen8_setup_page_directory_pointer(struct i915_hw_ppgtt *ppgtt,
645 struct i915_pml4 *pml4,
646 struct i915_page_directory_pointer *pdp,
649 gen8_ppgtt_pml4e_t *pagemap = kmap_px(pml4);
651 WARN_ON(!USES_FULL_48BIT_PPGTT(ppgtt->base.dev));
652 pagemap[index] = gen8_pml4e_encode(px_dma(pdp), I915_CACHE_LLC);
653 kunmap_px(ppgtt, pagemap);
656 /* Broadwell Page Directory Pointer Descriptors */
657 static int gen8_write_pdp(struct drm_i915_gem_request *req,
661 struct intel_engine_cs *ring = req->ring;
666 ret = intel_ring_begin(req, 6);
670 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
671 intel_ring_emit_reg(ring, GEN8_RING_PDP_UDW(ring, entry));
672 intel_ring_emit(ring, upper_32_bits(addr));
673 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
674 intel_ring_emit_reg(ring, GEN8_RING_PDP_LDW(ring, entry));
675 intel_ring_emit(ring, lower_32_bits(addr));
676 intel_ring_advance(ring);
681 static int gen8_legacy_mm_switch(struct i915_hw_ppgtt *ppgtt,
682 struct drm_i915_gem_request *req)
686 for (i = GEN8_LEGACY_PDPES - 1; i >= 0; i--) {
687 const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i);
689 ret = gen8_write_pdp(req, i, pd_daddr);
697 static int gen8_48b_mm_switch(struct i915_hw_ppgtt *ppgtt,
698 struct drm_i915_gem_request *req)
700 return gen8_write_pdp(req, 0, px_dma(&ppgtt->pml4));
703 static void gen8_ppgtt_clear_pte_range(struct i915_address_space *vm,
704 struct i915_page_directory_pointer *pdp,
707 gen8_pte_t scratch_pte)
709 struct i915_hw_ppgtt *ppgtt =
710 container_of(vm, struct i915_hw_ppgtt, base);
711 gen8_pte_t *pt_vaddr;
712 unsigned pdpe = gen8_pdpe_index(start);
713 unsigned pde = gen8_pde_index(start);
714 unsigned pte = gen8_pte_index(start);
715 unsigned num_entries = length >> PAGE_SHIFT;
716 unsigned last_pte, i;
721 while (num_entries) {
722 struct i915_page_directory *pd;
723 struct i915_page_table *pt;
725 if (WARN_ON(!pdp->page_directory[pdpe]))
728 pd = pdp->page_directory[pdpe];
730 if (WARN_ON(!pd->page_table[pde]))
733 pt = pd->page_table[pde];
735 if (WARN_ON(!px_page(pt)))
738 last_pte = pte + num_entries;
739 if (last_pte > GEN8_PTES)
740 last_pte = GEN8_PTES;
742 pt_vaddr = kmap_px(pt);
744 for (i = pte; i < last_pte; i++) {
745 pt_vaddr[i] = scratch_pte;
749 kunmap_px(ppgtt, pt);
752 if (++pde == I915_PDES) {
753 if (++pdpe == I915_PDPES_PER_PDP(vm->dev))
760 static void gen8_ppgtt_clear_range(struct i915_address_space *vm,
765 struct i915_hw_ppgtt *ppgtt =
766 container_of(vm, struct i915_hw_ppgtt, base);
767 gen8_pte_t scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
768 I915_CACHE_LLC, use_scratch);
770 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
771 gen8_ppgtt_clear_pte_range(vm, &ppgtt->pdp, start, length,
775 struct i915_page_directory_pointer *pdp;
777 gen8_for_each_pml4e(pdp, &ppgtt->pml4, start, length, pml4e) {
778 gen8_ppgtt_clear_pte_range(vm, pdp, start, length,
785 gen8_ppgtt_insert_pte_entries(struct i915_address_space *vm,
786 struct i915_page_directory_pointer *pdp,
787 struct sg_page_iter *sg_iter,
789 enum i915_cache_level cache_level)
791 struct i915_hw_ppgtt *ppgtt =
792 container_of(vm, struct i915_hw_ppgtt, base);
793 gen8_pte_t *pt_vaddr;
794 unsigned pdpe = gen8_pdpe_index(start);
795 unsigned pde = gen8_pde_index(start);
796 unsigned pte = gen8_pte_index(start);
800 while (__sg_page_iter_next(sg_iter)) {
801 if (pt_vaddr == NULL) {
802 struct i915_page_directory *pd = pdp->page_directory[pdpe];
803 struct i915_page_table *pt = pd->page_table[pde];
804 pt_vaddr = kmap_px(pt);
808 gen8_pte_encode(sg_page_iter_dma_address(sg_iter),
810 if (++pte == GEN8_PTES) {
811 kunmap_px(ppgtt, pt_vaddr);
813 if (++pde == I915_PDES) {
814 if (++pdpe == I915_PDPES_PER_PDP(vm->dev))
823 kunmap_px(ppgtt, pt_vaddr);
826 static void gen8_ppgtt_insert_entries(struct i915_address_space *vm,
827 struct sg_table *pages,
829 enum i915_cache_level cache_level,
832 struct i915_hw_ppgtt *ppgtt =
833 container_of(vm, struct i915_hw_ppgtt, base);
834 struct sg_page_iter sg_iter;
836 __sg_page_iter_start(&sg_iter, pages->sgl, sg_nents(pages->sgl), 0);
838 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
839 gen8_ppgtt_insert_pte_entries(vm, &ppgtt->pdp, &sg_iter, start,
842 struct i915_page_directory_pointer *pdp;
844 uint64_t length = (uint64_t)pages->orig_nents << PAGE_SHIFT;
846 gen8_for_each_pml4e(pdp, &ppgtt->pml4, start, length, pml4e) {
847 gen8_ppgtt_insert_pte_entries(vm, pdp, &sg_iter,
853 static void gen8_free_page_tables(struct drm_device *dev,
854 struct i915_page_directory *pd)
861 for_each_set_bit(i, pd->used_pdes, I915_PDES) {
862 if (WARN_ON(!pd->page_table[i]))
865 free_pt(dev, pd->page_table[i]);
866 pd->page_table[i] = NULL;
870 static int gen8_init_scratch(struct i915_address_space *vm)
872 struct drm_device *dev = vm->dev;
874 vm->scratch_page = alloc_scratch_page(dev);
875 if (IS_ERR(vm->scratch_page))
876 return PTR_ERR(vm->scratch_page);
878 vm->scratch_pt = alloc_pt(dev);
879 if (IS_ERR(vm->scratch_pt)) {
880 free_scratch_page(dev, vm->scratch_page);
881 return PTR_ERR(vm->scratch_pt);
884 vm->scratch_pd = alloc_pd(dev);
885 if (IS_ERR(vm->scratch_pd)) {
886 free_pt(dev, vm->scratch_pt);
887 free_scratch_page(dev, vm->scratch_page);
888 return PTR_ERR(vm->scratch_pd);
891 if (USES_FULL_48BIT_PPGTT(dev)) {
892 vm->scratch_pdp = alloc_pdp(dev);
893 if (IS_ERR(vm->scratch_pdp)) {
894 free_pd(dev, vm->scratch_pd);
895 free_pt(dev, vm->scratch_pt);
896 free_scratch_page(dev, vm->scratch_page);
897 return PTR_ERR(vm->scratch_pdp);
901 gen8_initialize_pt(vm, vm->scratch_pt);
902 gen8_initialize_pd(vm, vm->scratch_pd);
903 if (USES_FULL_48BIT_PPGTT(dev))
904 gen8_initialize_pdp(vm, vm->scratch_pdp);
909 static int gen8_ppgtt_notify_vgt(struct i915_hw_ppgtt *ppgtt, bool create)
911 enum vgt_g2v_type msg;
912 struct drm_device *dev = ppgtt->base.dev;
913 struct drm_i915_private *dev_priv = dev->dev_private;
916 if (USES_FULL_48BIT_PPGTT(dev)) {
917 u64 daddr = px_dma(&ppgtt->pml4);
919 I915_WRITE(vgtif_reg(pdp[0].lo), lower_32_bits(daddr));
920 I915_WRITE(vgtif_reg(pdp[0].hi), upper_32_bits(daddr));
922 msg = (create ? VGT_G2V_PPGTT_L4_PAGE_TABLE_CREATE :
923 VGT_G2V_PPGTT_L4_PAGE_TABLE_DESTROY);
925 for (i = 0; i < GEN8_LEGACY_PDPES; i++) {
926 u64 daddr = i915_page_dir_dma_addr(ppgtt, i);
928 I915_WRITE(vgtif_reg(pdp[i].lo), lower_32_bits(daddr));
929 I915_WRITE(vgtif_reg(pdp[i].hi), upper_32_bits(daddr));
932 msg = (create ? VGT_G2V_PPGTT_L3_PAGE_TABLE_CREATE :
933 VGT_G2V_PPGTT_L3_PAGE_TABLE_DESTROY);
936 I915_WRITE(vgtif_reg(g2v_notify), msg);
941 static void gen8_free_scratch(struct i915_address_space *vm)
943 struct drm_device *dev = vm->dev;
945 if (USES_FULL_48BIT_PPGTT(dev))
946 free_pdp(dev, vm->scratch_pdp);
947 free_pd(dev, vm->scratch_pd);
948 free_pt(dev, vm->scratch_pt);
949 free_scratch_page(dev, vm->scratch_page);
952 static void gen8_ppgtt_cleanup_3lvl(struct drm_device *dev,
953 struct i915_page_directory_pointer *pdp)
957 for_each_set_bit(i, pdp->used_pdpes, I915_PDPES_PER_PDP(dev)) {
958 if (WARN_ON(!pdp->page_directory[i]))
961 gen8_free_page_tables(dev, pdp->page_directory[i]);
962 free_pd(dev, pdp->page_directory[i]);
968 static void gen8_ppgtt_cleanup_4lvl(struct i915_hw_ppgtt *ppgtt)
972 for_each_set_bit(i, ppgtt->pml4.used_pml4es, GEN8_PML4ES_PER_PML4) {
973 if (WARN_ON(!ppgtt->pml4.pdps[i]))
976 gen8_ppgtt_cleanup_3lvl(ppgtt->base.dev, ppgtt->pml4.pdps[i]);
979 cleanup_px(ppgtt->base.dev, &ppgtt->pml4);
982 static void gen8_ppgtt_cleanup(struct i915_address_space *vm)
984 struct i915_hw_ppgtt *ppgtt =
985 container_of(vm, struct i915_hw_ppgtt, base);
987 if (intel_vgpu_active(vm->dev))
988 gen8_ppgtt_notify_vgt(ppgtt, false);
990 if (!USES_FULL_48BIT_PPGTT(ppgtt->base.dev))
991 gen8_ppgtt_cleanup_3lvl(ppgtt->base.dev, &ppgtt->pdp);
993 gen8_ppgtt_cleanup_4lvl(ppgtt);
995 gen8_free_scratch(vm);
999 * gen8_ppgtt_alloc_pagetabs() - Allocate page tables for VA range.
1000 * @vm: Master vm structure.
1001 * @pd: Page directory for this address range.
1002 * @start: Starting virtual address to begin allocations.
1003 * @length: Size of the allocations.
1004 * @new_pts: Bitmap set by function with new allocations. Likely used by the
1005 * caller to free on error.
1007 * Allocate the required number of page tables. Extremely similar to
1008 * gen8_ppgtt_alloc_page_directories(). The main difference is here we are limited by
1009 * the page directory boundary (instead of the page directory pointer). That
1010 * boundary is 1GB virtual. Therefore, unlike gen8_ppgtt_alloc_page_directories(), it is
1011 * possible, and likely that the caller will need to use multiple calls of this
1012 * function to achieve the appropriate allocation.
1014 * Return: 0 if success; negative error code otherwise.
1016 static int gen8_ppgtt_alloc_pagetabs(struct i915_address_space *vm,
1017 struct i915_page_directory *pd,
1020 unsigned long *new_pts)
1022 struct drm_device *dev = vm->dev;
1023 struct i915_page_table *pt;
1026 gen8_for_each_pde(pt, pd, start, length, pde) {
1027 /* Don't reallocate page tables */
1028 if (test_bit(pde, pd->used_pdes)) {
1029 /* Scratch is never allocated this way */
1030 WARN_ON(pt == vm->scratch_pt);
1038 gen8_initialize_pt(vm, pt);
1039 pd->page_table[pde] = pt;
1040 __set_bit(pde, new_pts);
1041 trace_i915_page_table_entry_alloc(vm, pde, start, GEN8_PDE_SHIFT);
1047 for_each_set_bit(pde, new_pts, I915_PDES)
1048 free_pt(dev, pd->page_table[pde]);
1054 * gen8_ppgtt_alloc_page_directories() - Allocate page directories for VA range.
1055 * @vm: Master vm structure.
1056 * @pdp: Page directory pointer for this address range.
1057 * @start: Starting virtual address to begin allocations.
1058 * @length: Size of the allocations.
1059 * @new_pds: Bitmap set by function with new allocations. Likely used by the
1060 * caller to free on error.
1062 * Allocate the required number of page directories starting at the pde index of
1063 * @start, and ending at the pde index @start + @length. This function will skip
1064 * over already allocated page directories within the range, and only allocate
1065 * new ones, setting the appropriate pointer within the pdp as well as the
1066 * correct position in the bitmap @new_pds.
1068 * The function will only allocate the pages within the range for a give page
1069 * directory pointer. In other words, if @start + @length straddles a virtually
1070 * addressed PDP boundary (512GB for 4k pages), there will be more allocations
1071 * required by the caller, This is not currently possible, and the BUG in the
1072 * code will prevent it.
1074 * Return: 0 if success; negative error code otherwise.
1077 gen8_ppgtt_alloc_page_directories(struct i915_address_space *vm,
1078 struct i915_page_directory_pointer *pdp,
1081 unsigned long *new_pds)
1083 struct drm_device *dev = vm->dev;
1084 struct i915_page_directory *pd;
1086 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1088 WARN_ON(!bitmap_empty(new_pds, pdpes));
1090 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1091 if (test_bit(pdpe, pdp->used_pdpes))
1098 gen8_initialize_pd(vm, pd);
1099 pdp->page_directory[pdpe] = pd;
1100 __set_bit(pdpe, new_pds);
1101 trace_i915_page_directory_entry_alloc(vm, pdpe, start, GEN8_PDPE_SHIFT);
1107 for_each_set_bit(pdpe, new_pds, pdpes)
1108 free_pd(dev, pdp->page_directory[pdpe]);
1114 * gen8_ppgtt_alloc_page_dirpointers() - Allocate pdps for VA range.
1115 * @vm: Master vm structure.
1116 * @pml4: Page map level 4 for this address range.
1117 * @start: Starting virtual address to begin allocations.
1118 * @length: Size of the allocations.
1119 * @new_pdps: Bitmap set by function with new allocations. Likely used by the
1120 * caller to free on error.
1122 * Allocate the required number of page directory pointers. Extremely similar to
1123 * gen8_ppgtt_alloc_page_directories() and gen8_ppgtt_alloc_pagetabs().
1124 * The main difference is here we are limited by the pml4 boundary (instead of
1125 * the page directory pointer).
1127 * Return: 0 if success; negative error code otherwise.
1130 gen8_ppgtt_alloc_page_dirpointers(struct i915_address_space *vm,
1131 struct i915_pml4 *pml4,
1134 unsigned long *new_pdps)
1136 struct drm_device *dev = vm->dev;
1137 struct i915_page_directory_pointer *pdp;
1140 WARN_ON(!bitmap_empty(new_pdps, GEN8_PML4ES_PER_PML4));
1142 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1143 if (!test_bit(pml4e, pml4->used_pml4es)) {
1144 pdp = alloc_pdp(dev);
1148 gen8_initialize_pdp(vm, pdp);
1149 pml4->pdps[pml4e] = pdp;
1150 __set_bit(pml4e, new_pdps);
1151 trace_i915_page_directory_pointer_entry_alloc(vm,
1161 for_each_set_bit(pml4e, new_pdps, GEN8_PML4ES_PER_PML4)
1162 free_pdp(dev, pml4->pdps[pml4e]);
1168 free_gen8_temp_bitmaps(unsigned long *new_pds, unsigned long *new_pts)
1174 /* Fills in the page directory bitmap, and the array of page tables bitmap. Both
1175 * of these are based on the number of PDPEs in the system.
1178 int __must_check alloc_gen8_temp_bitmaps(unsigned long **new_pds,
1179 unsigned long **new_pts,
1185 pds = kcalloc(BITS_TO_LONGS(pdpes), sizeof(unsigned long), GFP_TEMPORARY);
1189 pts = kcalloc(pdpes, BITS_TO_LONGS(I915_PDES) * sizeof(unsigned long),
1200 free_gen8_temp_bitmaps(pds, pts);
1204 /* PDE TLBs are a pain to invalidate on GEN8+. When we modify
1205 * the page table structures, we mark them dirty so that
1206 * context switching/execlist queuing code takes extra steps
1207 * to ensure that tlbs are flushed.
1209 static void mark_tlbs_dirty(struct i915_hw_ppgtt *ppgtt)
1211 ppgtt->pd_dirty_rings = INTEL_INFO(ppgtt->base.dev)->ring_mask;
1214 static int gen8_alloc_va_range_3lvl(struct i915_address_space *vm,
1215 struct i915_page_directory_pointer *pdp,
1219 struct i915_hw_ppgtt *ppgtt =
1220 container_of(vm, struct i915_hw_ppgtt, base);
1221 unsigned long *new_page_dirs, *new_page_tables;
1222 struct drm_device *dev = vm->dev;
1223 struct i915_page_directory *pd;
1224 const uint64_t orig_start = start;
1225 const uint64_t orig_length = length;
1227 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1230 /* Wrap is never okay since we can only represent 48b, and we don't
1231 * actually use the other side of the canonical address space.
1233 if (WARN_ON(start + length < start))
1236 if (WARN_ON(start + length > vm->total))
1239 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables, pdpes);
1243 /* Do the allocations first so we can easily bail out */
1244 ret = gen8_ppgtt_alloc_page_directories(vm, pdp, start, length,
1247 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1251 /* For every page directory referenced, allocate page tables */
1252 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1253 ret = gen8_ppgtt_alloc_pagetabs(vm, pd, start, length,
1254 new_page_tables + pdpe * BITS_TO_LONGS(I915_PDES));
1260 length = orig_length;
1262 /* Allocations have completed successfully, so set the bitmaps, and do
1264 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1265 gen8_pde_t *const page_directory = kmap_px(pd);
1266 struct i915_page_table *pt;
1267 uint64_t pd_len = length;
1268 uint64_t pd_start = start;
1271 /* Every pd should be allocated, we just did that above. */
1274 gen8_for_each_pde(pt, pd, pd_start, pd_len, pde) {
1275 /* Same reasoning as pd */
1278 WARN_ON(!gen8_pte_count(pd_start, pd_len));
1280 /* Set our used ptes within the page table */
1281 bitmap_set(pt->used_ptes,
1282 gen8_pte_index(pd_start),
1283 gen8_pte_count(pd_start, pd_len));
1285 /* Our pde is now pointing to the pagetable, pt */
1286 __set_bit(pde, pd->used_pdes);
1288 /* Map the PDE to the page table */
1289 page_directory[pde] = gen8_pde_encode(px_dma(pt),
1291 trace_i915_page_table_entry_map(&ppgtt->base, pde, pt,
1292 gen8_pte_index(start),
1293 gen8_pte_count(start, length),
1296 /* NB: We haven't yet mapped ptes to pages. At this
1297 * point we're still relying on insert_entries() */
1300 kunmap_px(ppgtt, page_directory);
1301 __set_bit(pdpe, pdp->used_pdpes);
1302 gen8_setup_page_directory(ppgtt, pdp, pd, pdpe);
1305 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1306 mark_tlbs_dirty(ppgtt);
1313 for_each_set_bit(temp, new_page_tables + pdpe *
1314 BITS_TO_LONGS(I915_PDES), I915_PDES)
1315 free_pt(dev, pdp->page_directory[pdpe]->page_table[temp]);
1318 for_each_set_bit(pdpe, new_page_dirs, pdpes)
1319 free_pd(dev, pdp->page_directory[pdpe]);
1321 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1322 mark_tlbs_dirty(ppgtt);
1326 static int gen8_alloc_va_range_4lvl(struct i915_address_space *vm,
1327 struct i915_pml4 *pml4,
1331 DECLARE_BITMAP(new_pdps, GEN8_PML4ES_PER_PML4);
1332 struct i915_hw_ppgtt *ppgtt =
1333 container_of(vm, struct i915_hw_ppgtt, base);
1334 struct i915_page_directory_pointer *pdp;
1338 /* Do the pml4 allocations first, so we don't need to track the newly
1339 * allocated tables below the pdp */
1340 bitmap_zero(new_pdps, GEN8_PML4ES_PER_PML4);
1342 /* The pagedirectory and pagetable allocations are done in the shared 3
1343 * and 4 level code. Just allocate the pdps.
1345 ret = gen8_ppgtt_alloc_page_dirpointers(vm, pml4, start, length,
1350 WARN(bitmap_weight(new_pdps, GEN8_PML4ES_PER_PML4) > 2,
1351 "The allocation has spanned more than 512GB. "
1352 "It is highly likely this is incorrect.");
1354 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1357 ret = gen8_alloc_va_range_3lvl(vm, pdp, start, length);
1361 gen8_setup_page_directory_pointer(ppgtt, pml4, pdp, pml4e);
1364 bitmap_or(pml4->used_pml4es, new_pdps, pml4->used_pml4es,
1365 GEN8_PML4ES_PER_PML4);
1370 for_each_set_bit(pml4e, new_pdps, GEN8_PML4ES_PER_PML4)
1371 gen8_ppgtt_cleanup_3lvl(vm->dev, pml4->pdps[pml4e]);
1376 static int gen8_alloc_va_range(struct i915_address_space *vm,
1377 uint64_t start, uint64_t length)
1379 struct i915_hw_ppgtt *ppgtt =
1380 container_of(vm, struct i915_hw_ppgtt, base);
1382 if (USES_FULL_48BIT_PPGTT(vm->dev))
1383 return gen8_alloc_va_range_4lvl(vm, &ppgtt->pml4, start, length);
1385 return gen8_alloc_va_range_3lvl(vm, &ppgtt->pdp, start, length);
1388 static void gen8_dump_pdp(struct i915_page_directory_pointer *pdp,
1389 uint64_t start, uint64_t length,
1390 gen8_pte_t scratch_pte,
1393 struct i915_page_directory *pd;
1396 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1397 struct i915_page_table *pt;
1398 uint64_t pd_len = length;
1399 uint64_t pd_start = start;
1402 if (!test_bit(pdpe, pdp->used_pdpes))
1405 seq_printf(m, "\tPDPE #%d\n", pdpe);
1406 gen8_for_each_pde(pt, pd, pd_start, pd_len, pde) {
1408 gen8_pte_t *pt_vaddr;
1410 if (!test_bit(pde, pd->used_pdes))
1413 pt_vaddr = kmap_px(pt);
1414 for (pte = 0; pte < GEN8_PTES; pte += 4) {
1416 (pdpe << GEN8_PDPE_SHIFT) |
1417 (pde << GEN8_PDE_SHIFT) |
1418 (pte << GEN8_PTE_SHIFT);
1422 for (i = 0; i < 4; i++)
1423 if (pt_vaddr[pte + i] != scratch_pte)
1428 seq_printf(m, "\t\t0x%llx [%03d,%03d,%04d]: =", va, pdpe, pde, pte);
1429 for (i = 0; i < 4; i++) {
1430 if (pt_vaddr[pte + i] != scratch_pte)
1431 seq_printf(m, " %llx", pt_vaddr[pte + i]);
1433 seq_puts(m, " SCRATCH ");
1437 /* don't use kunmap_px, it could trigger
1438 * an unnecessary flush.
1440 kunmap_atomic(pt_vaddr);
1445 static void gen8_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
1447 struct i915_address_space *vm = &ppgtt->base;
1448 uint64_t start = ppgtt->base.start;
1449 uint64_t length = ppgtt->base.total;
1450 gen8_pte_t scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
1451 I915_CACHE_LLC, true);
1453 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
1454 gen8_dump_pdp(&ppgtt->pdp, start, length, scratch_pte, m);
1457 struct i915_pml4 *pml4 = &ppgtt->pml4;
1458 struct i915_page_directory_pointer *pdp;
1460 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1461 if (!test_bit(pml4e, pml4->used_pml4es))
1464 seq_printf(m, " PML4E #%llu\n", pml4e);
1465 gen8_dump_pdp(pdp, start, length, scratch_pte, m);
1470 static int gen8_preallocate_top_level_pdps(struct i915_hw_ppgtt *ppgtt)
1472 unsigned long *new_page_dirs, *new_page_tables;
1473 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1476 /* We allocate temp bitmap for page tables for no gain
1477 * but as this is for init only, lets keep the things simple
1479 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables, pdpes);
1483 /* Allocate for all pdps regardless of how the ppgtt
1486 ret = gen8_ppgtt_alloc_page_directories(&ppgtt->base, &ppgtt->pdp,
1490 *ppgtt->pdp.used_pdpes = *new_page_dirs;
1492 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1498 * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers
1499 * with a net effect resembling a 2-level page table in normal x86 terms. Each
1500 * PDP represents 1GB of memory 4 * 512 * 512 * 4096 = 4GB legacy 32b address
1504 static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
1508 ret = gen8_init_scratch(&ppgtt->base);
1512 ppgtt->base.start = 0;
1513 ppgtt->base.cleanup = gen8_ppgtt_cleanup;
1514 ppgtt->base.allocate_va_range = gen8_alloc_va_range;
1515 ppgtt->base.insert_entries = gen8_ppgtt_insert_entries;
1516 ppgtt->base.clear_range = gen8_ppgtt_clear_range;
1517 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
1518 ppgtt->base.bind_vma = ppgtt_bind_vma;
1519 ppgtt->debug_dump = gen8_dump_ppgtt;
1521 if (USES_FULL_48BIT_PPGTT(ppgtt->base.dev)) {
1522 ret = setup_px(ppgtt->base.dev, &ppgtt->pml4);
1526 gen8_initialize_pml4(&ppgtt->base, &ppgtt->pml4);
1528 ppgtt->base.total = 1ULL << 48;
1529 ppgtt->switch_mm = gen8_48b_mm_switch;
1531 ret = __pdp_init(ppgtt->base.dev, &ppgtt->pdp);
1535 ppgtt->base.total = 1ULL << 32;
1536 ppgtt->switch_mm = gen8_legacy_mm_switch;
1537 trace_i915_page_directory_pointer_entry_alloc(&ppgtt->base,
1541 if (intel_vgpu_active(ppgtt->base.dev)) {
1542 ret = gen8_preallocate_top_level_pdps(ppgtt);
1548 if (intel_vgpu_active(ppgtt->base.dev))
1549 gen8_ppgtt_notify_vgt(ppgtt, true);
1554 gen8_free_scratch(&ppgtt->base);
1558 static void gen6_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
1560 struct i915_address_space *vm = &ppgtt->base;
1561 struct i915_page_table *unused;
1562 gen6_pte_t scratch_pte;
1564 uint32_t pte, pde, temp;
1565 uint32_t start = ppgtt->base.start, length = ppgtt->base.total;
1567 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1568 I915_CACHE_LLC, true, 0);
1570 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde) {
1572 gen6_pte_t *pt_vaddr;
1573 const dma_addr_t pt_addr = px_dma(ppgtt->pd.page_table[pde]);
1574 pd_entry = readl(ppgtt->pd_addr + pde);
1575 expected = (GEN6_PDE_ADDR_ENCODE(pt_addr) | GEN6_PDE_VALID);
1577 if (pd_entry != expected)
1578 seq_printf(m, "\tPDE #%d mismatch: Actual PDE: %x Expected PDE: %x\n",
1582 seq_printf(m, "\tPDE: %x\n", pd_entry);
1584 pt_vaddr = kmap_px(ppgtt->pd.page_table[pde]);
1586 for (pte = 0; pte < GEN6_PTES; pte+=4) {
1588 (pde * PAGE_SIZE * GEN6_PTES) +
1592 for (i = 0; i < 4; i++)
1593 if (pt_vaddr[pte + i] != scratch_pte)
1598 seq_printf(m, "\t\t0x%lx [%03d,%04d]: =", va, pde, pte);
1599 for (i = 0; i < 4; i++) {
1600 if (pt_vaddr[pte + i] != scratch_pte)
1601 seq_printf(m, " %08x", pt_vaddr[pte + i]);
1603 seq_puts(m, " SCRATCH ");
1607 kunmap_px(ppgtt, pt_vaddr);
1611 /* Write pde (index) from the page directory @pd to the page table @pt */
1612 static void gen6_write_pde(struct i915_page_directory *pd,
1613 const int pde, struct i915_page_table *pt)
1615 /* Caller needs to make sure the write completes if necessary */
1616 struct i915_hw_ppgtt *ppgtt =
1617 container_of(pd, struct i915_hw_ppgtt, pd);
1620 pd_entry = GEN6_PDE_ADDR_ENCODE(px_dma(pt));
1621 pd_entry |= GEN6_PDE_VALID;
1623 writel(pd_entry, ppgtt->pd_addr + pde);
1626 /* Write all the page tables found in the ppgtt structure to incrementing page
1628 static void gen6_write_page_range(struct drm_i915_private *dev_priv,
1629 struct i915_page_directory *pd,
1630 uint32_t start, uint32_t length)
1632 struct i915_page_table *pt;
1635 gen6_for_each_pde(pt, pd, start, length, temp, pde)
1636 gen6_write_pde(pd, pde, pt);
1638 /* Make sure write is complete before other code can use this page
1639 * table. Also require for WC mapped PTEs */
1640 readl(dev_priv->gtt.gsm);
1643 static uint32_t get_pd_offset(struct i915_hw_ppgtt *ppgtt)
1645 BUG_ON(ppgtt->pd.base.ggtt_offset & 0x3f);
1647 return (ppgtt->pd.base.ggtt_offset / 64) << 16;
1650 static int hsw_mm_switch(struct i915_hw_ppgtt *ppgtt,
1651 struct drm_i915_gem_request *req)
1653 struct intel_engine_cs *ring = req->ring;
1656 /* NB: TLBs must be flushed and invalidated before a switch */
1657 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1661 ret = intel_ring_begin(req, 6);
1665 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1666 intel_ring_emit_reg(ring, RING_PP_DIR_DCLV(ring));
1667 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1668 intel_ring_emit_reg(ring, RING_PP_DIR_BASE(ring));
1669 intel_ring_emit(ring, get_pd_offset(ppgtt));
1670 intel_ring_emit(ring, MI_NOOP);
1671 intel_ring_advance(ring);
1676 static int vgpu_mm_switch(struct i915_hw_ppgtt *ppgtt,
1677 struct drm_i915_gem_request *req)
1679 struct intel_engine_cs *ring = req->ring;
1680 struct drm_i915_private *dev_priv = to_i915(ppgtt->base.dev);
1682 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
1683 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
1687 static int gen7_mm_switch(struct i915_hw_ppgtt *ppgtt,
1688 struct drm_i915_gem_request *req)
1690 struct intel_engine_cs *ring = req->ring;
1693 /* NB: TLBs must be flushed and invalidated before a switch */
1694 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1698 ret = intel_ring_begin(req, 6);
1702 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1703 intel_ring_emit_reg(ring, RING_PP_DIR_DCLV(ring));
1704 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1705 intel_ring_emit_reg(ring, RING_PP_DIR_BASE(ring));
1706 intel_ring_emit(ring, get_pd_offset(ppgtt));
1707 intel_ring_emit(ring, MI_NOOP);
1708 intel_ring_advance(ring);
1710 /* XXX: RCS is the only one to auto invalidate the TLBs? */
1711 if (ring->id != RCS) {
1712 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1720 static int gen6_mm_switch(struct i915_hw_ppgtt *ppgtt,
1721 struct drm_i915_gem_request *req)
1723 struct intel_engine_cs *ring = req->ring;
1724 struct drm_device *dev = ppgtt->base.dev;
1725 struct drm_i915_private *dev_priv = dev->dev_private;
1728 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
1729 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
1731 POSTING_READ(RING_PP_DIR_DCLV(ring));
1736 static void gen8_ppgtt_enable(struct drm_device *dev)
1738 struct drm_i915_private *dev_priv = dev->dev_private;
1739 struct intel_engine_cs *ring;
1742 for_each_ring(ring, dev_priv, j) {
1743 u32 four_level = USES_FULL_48BIT_PPGTT(dev) ? GEN8_GFX_PPGTT_48B : 0;
1744 I915_WRITE(RING_MODE_GEN7(ring),
1745 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE | four_level));
1749 static void gen7_ppgtt_enable(struct drm_device *dev)
1751 struct drm_i915_private *dev_priv = dev->dev_private;
1752 struct intel_engine_cs *ring;
1753 uint32_t ecochk, ecobits;
1756 ecobits = I915_READ(GAC_ECO_BITS);
1757 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B);
1759 ecochk = I915_READ(GAM_ECOCHK);
1760 if (IS_HASWELL(dev)) {
1761 ecochk |= ECOCHK_PPGTT_WB_HSW;
1763 ecochk |= ECOCHK_PPGTT_LLC_IVB;
1764 ecochk &= ~ECOCHK_PPGTT_GFDT_IVB;
1766 I915_WRITE(GAM_ECOCHK, ecochk);
1768 for_each_ring(ring, dev_priv, i) {
1769 /* GFX_MODE is per-ring on gen7+ */
1770 I915_WRITE(RING_MODE_GEN7(ring),
1771 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1775 static void gen6_ppgtt_enable(struct drm_device *dev)
1777 struct drm_i915_private *dev_priv = dev->dev_private;
1778 uint32_t ecochk, gab_ctl, ecobits;
1780 ecobits = I915_READ(GAC_ECO_BITS);
1781 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_SNB_BIT |
1782 ECOBITS_PPGTT_CACHE64B);
1784 gab_ctl = I915_READ(GAB_CTL);
1785 I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT);
1787 ecochk = I915_READ(GAM_ECOCHK);
1788 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT | ECOCHK_PPGTT_CACHE64B);
1790 I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1793 /* PPGTT support for Sandybdrige/Gen6 and later */
1794 static void gen6_ppgtt_clear_range(struct i915_address_space *vm,
1799 struct i915_hw_ppgtt *ppgtt =
1800 container_of(vm, struct i915_hw_ppgtt, base);
1801 gen6_pte_t *pt_vaddr, scratch_pte;
1802 unsigned first_entry = start >> PAGE_SHIFT;
1803 unsigned num_entries = length >> PAGE_SHIFT;
1804 unsigned act_pt = first_entry / GEN6_PTES;
1805 unsigned first_pte = first_entry % GEN6_PTES;
1806 unsigned last_pte, i;
1808 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1809 I915_CACHE_LLC, true, 0);
1811 while (num_entries) {
1812 last_pte = first_pte + num_entries;
1813 if (last_pte > GEN6_PTES)
1814 last_pte = GEN6_PTES;
1816 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1818 for (i = first_pte; i < last_pte; i++)
1819 pt_vaddr[i] = scratch_pte;
1821 kunmap_px(ppgtt, pt_vaddr);
1823 num_entries -= last_pte - first_pte;
1829 static void gen6_ppgtt_insert_entries(struct i915_address_space *vm,
1830 struct sg_table *pages,
1832 enum i915_cache_level cache_level, u32 flags)
1834 struct i915_hw_ppgtt *ppgtt =
1835 container_of(vm, struct i915_hw_ppgtt, base);
1836 gen6_pte_t *pt_vaddr;
1837 unsigned first_entry = start >> PAGE_SHIFT;
1838 unsigned act_pt = first_entry / GEN6_PTES;
1839 unsigned act_pte = first_entry % GEN6_PTES;
1840 struct sg_page_iter sg_iter;
1843 for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) {
1844 if (pt_vaddr == NULL)
1845 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1848 vm->pte_encode(sg_page_iter_dma_address(&sg_iter),
1849 cache_level, true, flags);
1851 if (++act_pte == GEN6_PTES) {
1852 kunmap_px(ppgtt, pt_vaddr);
1859 kunmap_px(ppgtt, pt_vaddr);
1862 static int gen6_alloc_va_range(struct i915_address_space *vm,
1863 uint64_t start_in, uint64_t length_in)
1865 DECLARE_BITMAP(new_page_tables, I915_PDES);
1866 struct drm_device *dev = vm->dev;
1867 struct drm_i915_private *dev_priv = dev->dev_private;
1868 struct i915_hw_ppgtt *ppgtt =
1869 container_of(vm, struct i915_hw_ppgtt, base);
1870 struct i915_page_table *pt;
1871 uint32_t start, length, start_save, length_save;
1875 if (WARN_ON(start_in + length_in > ppgtt->base.total))
1878 start = start_save = start_in;
1879 length = length_save = length_in;
1881 bitmap_zero(new_page_tables, I915_PDES);
1883 /* The allocation is done in two stages so that we can bail out with
1884 * minimal amount of pain. The first stage finds new page tables that
1885 * need allocation. The second stage marks use ptes within the page
1888 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1889 if (pt != vm->scratch_pt) {
1890 WARN_ON(bitmap_empty(pt->used_ptes, GEN6_PTES));
1894 /* We've already allocated a page table */
1895 WARN_ON(!bitmap_empty(pt->used_ptes, GEN6_PTES));
1903 gen6_initialize_pt(vm, pt);
1905 ppgtt->pd.page_table[pde] = pt;
1906 __set_bit(pde, new_page_tables);
1907 trace_i915_page_table_entry_alloc(vm, pde, start, GEN6_PDE_SHIFT);
1911 length = length_save;
1913 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1914 DECLARE_BITMAP(tmp_bitmap, GEN6_PTES);
1916 bitmap_zero(tmp_bitmap, GEN6_PTES);
1917 bitmap_set(tmp_bitmap, gen6_pte_index(start),
1918 gen6_pte_count(start, length));
1920 if (__test_and_clear_bit(pde, new_page_tables))
1921 gen6_write_pde(&ppgtt->pd, pde, pt);
1923 trace_i915_page_table_entry_map(vm, pde, pt,
1924 gen6_pte_index(start),
1925 gen6_pte_count(start, length),
1927 bitmap_or(pt->used_ptes, tmp_bitmap, pt->used_ptes,
1931 WARN_ON(!bitmap_empty(new_page_tables, I915_PDES));
1933 /* Make sure write is complete before other code can use this page
1934 * table. Also require for WC mapped PTEs */
1935 readl(dev_priv->gtt.gsm);
1937 mark_tlbs_dirty(ppgtt);
1941 for_each_set_bit(pde, new_page_tables, I915_PDES) {
1942 struct i915_page_table *pt = ppgtt->pd.page_table[pde];
1944 ppgtt->pd.page_table[pde] = vm->scratch_pt;
1945 free_pt(vm->dev, pt);
1948 mark_tlbs_dirty(ppgtt);
1952 static int gen6_init_scratch(struct i915_address_space *vm)
1954 struct drm_device *dev = vm->dev;
1956 vm->scratch_page = alloc_scratch_page(dev);
1957 if (IS_ERR(vm->scratch_page))
1958 return PTR_ERR(vm->scratch_page);
1960 vm->scratch_pt = alloc_pt(dev);
1961 if (IS_ERR(vm->scratch_pt)) {
1962 free_scratch_page(dev, vm->scratch_page);
1963 return PTR_ERR(vm->scratch_pt);
1966 gen6_initialize_pt(vm, vm->scratch_pt);
1971 static void gen6_free_scratch(struct i915_address_space *vm)
1973 struct drm_device *dev = vm->dev;
1975 free_pt(dev, vm->scratch_pt);
1976 free_scratch_page(dev, vm->scratch_page);
1979 static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
1981 struct i915_hw_ppgtt *ppgtt =
1982 container_of(vm, struct i915_hw_ppgtt, base);
1983 struct i915_page_table *pt;
1986 drm_mm_remove_node(&ppgtt->node);
1988 gen6_for_all_pdes(pt, ppgtt, pde) {
1989 if (pt != vm->scratch_pt)
1990 free_pt(ppgtt->base.dev, pt);
1993 gen6_free_scratch(vm);
1996 static int gen6_ppgtt_allocate_page_directories(struct i915_hw_ppgtt *ppgtt)
1998 struct i915_address_space *vm = &ppgtt->base;
1999 struct drm_device *dev = ppgtt->base.dev;
2000 struct drm_i915_private *dev_priv = dev->dev_private;
2001 bool retried = false;
2004 /* PPGTT PDEs reside in the GGTT and consists of 512 entries. The
2005 * allocator works in address space sizes, so it's multiplied by page
2006 * size. We allocate at the top of the GTT to avoid fragmentation.
2008 BUG_ON(!drm_mm_initialized(&dev_priv->gtt.base.mm));
2010 ret = gen6_init_scratch(vm);
2015 ret = drm_mm_insert_node_in_range_generic(&dev_priv->gtt.base.mm,
2016 &ppgtt->node, GEN6_PD_SIZE,
2018 0, dev_priv->gtt.base.total,
2020 if (ret == -ENOSPC && !retried) {
2021 ret = i915_gem_evict_something(dev, &dev_priv->gtt.base,
2022 GEN6_PD_SIZE, GEN6_PD_ALIGN,
2024 0, dev_priv->gtt.base.total,
2037 if (ppgtt->node.start < dev_priv->gtt.mappable_end)
2038 DRM_DEBUG("Forced to use aperture for PDEs\n");
2043 gen6_free_scratch(vm);
2047 static int gen6_ppgtt_alloc(struct i915_hw_ppgtt *ppgtt)
2049 return gen6_ppgtt_allocate_page_directories(ppgtt);
2052 static void gen6_scratch_va_range(struct i915_hw_ppgtt *ppgtt,
2053 uint64_t start, uint64_t length)
2055 struct i915_page_table *unused;
2058 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde)
2059 ppgtt->pd.page_table[pde] = ppgtt->base.scratch_pt;
2062 static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
2064 struct drm_device *dev = ppgtt->base.dev;
2065 struct drm_i915_private *dev_priv = dev->dev_private;
2068 ppgtt->base.pte_encode = dev_priv->gtt.base.pte_encode;
2070 ppgtt->switch_mm = gen6_mm_switch;
2071 } else if (IS_HASWELL(dev)) {
2072 ppgtt->switch_mm = hsw_mm_switch;
2073 } else if (IS_GEN7(dev)) {
2074 ppgtt->switch_mm = gen7_mm_switch;
2078 if (intel_vgpu_active(dev))
2079 ppgtt->switch_mm = vgpu_mm_switch;
2081 ret = gen6_ppgtt_alloc(ppgtt);
2085 ppgtt->base.allocate_va_range = gen6_alloc_va_range;
2086 ppgtt->base.clear_range = gen6_ppgtt_clear_range;
2087 ppgtt->base.insert_entries = gen6_ppgtt_insert_entries;
2088 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
2089 ppgtt->base.bind_vma = ppgtt_bind_vma;
2090 ppgtt->base.cleanup = gen6_ppgtt_cleanup;
2091 ppgtt->base.start = 0;
2092 ppgtt->base.total = I915_PDES * GEN6_PTES * PAGE_SIZE;
2093 ppgtt->debug_dump = gen6_dump_ppgtt;
2095 ppgtt->pd.base.ggtt_offset =
2096 ppgtt->node.start / PAGE_SIZE * sizeof(gen6_pte_t);
2098 ppgtt->pd_addr = (gen6_pte_t __iomem *)dev_priv->gtt.gsm +
2099 ppgtt->pd.base.ggtt_offset / sizeof(gen6_pte_t);
2101 gen6_scratch_va_range(ppgtt, 0, ppgtt->base.total);
2103 gen6_write_page_range(dev_priv, &ppgtt->pd, 0, ppgtt->base.total);
2105 DRM_DEBUG_DRIVER("Allocated pde space (%lldM) at GTT entry: %llx\n",
2106 ppgtt->node.size >> 20,
2107 ppgtt->node.start / PAGE_SIZE);
2109 DRM_DEBUG("Adding PPGTT at offset %x\n",
2110 ppgtt->pd.base.ggtt_offset << 10);
2115 static int __hw_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
2117 ppgtt->base.dev = dev;
2119 if (INTEL_INFO(dev)->gen < 8)
2120 return gen6_ppgtt_init(ppgtt);
2122 return gen8_ppgtt_init(ppgtt);
2125 static void i915_address_space_init(struct i915_address_space *vm,
2126 struct drm_i915_private *dev_priv)
2128 drm_mm_init(&vm->mm, vm->start, vm->total);
2129 vm->dev = dev_priv->dev;
2130 INIT_LIST_HEAD(&vm->active_list);
2131 INIT_LIST_HEAD(&vm->inactive_list);
2132 list_add_tail(&vm->global_link, &dev_priv->vm_list);
2135 static void gtt_write_workarounds(struct drm_device *dev)
2137 struct drm_i915_private *dev_priv = dev->dev_private;
2139 /* This function is for gtt related workarounds. This function is
2140 * called on driver load and after a GPU reset, so you can place
2141 * workarounds here even if they get overwritten by GPU reset.
2143 /* WaIncreaseDefaultTLBEntries:chv,bdw,skl,bxt */
2144 if (IS_BROADWELL(dev))
2145 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN8_L3_LRA_1_GPGPU_DEFAULT_VALUE_BDW);
2146 else if (IS_CHERRYVIEW(dev))
2147 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN8_L3_LRA_1_GPGPU_DEFAULT_VALUE_CHV);
2148 else if (IS_SKYLAKE(dev))
2149 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN9_L3_LRA_1_GPGPU_DEFAULT_VALUE_SKL);
2150 else if (IS_BROXTON(dev))
2151 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN9_L3_LRA_1_GPGPU_DEFAULT_VALUE_BXT);
2154 int i915_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
2156 struct drm_i915_private *dev_priv = dev->dev_private;
2159 ret = __hw_ppgtt_init(dev, ppgtt);
2161 kref_init(&ppgtt->ref);
2162 i915_address_space_init(&ppgtt->base, dev_priv);
2168 int i915_ppgtt_init_hw(struct drm_device *dev)
2170 gtt_write_workarounds(dev);
2172 /* In the case of execlists, PPGTT is enabled by the context descriptor
2173 * and the PDPs are contained within the context itself. We don't
2174 * need to do anything here. */
2175 if (i915.enable_execlists)
2178 if (!USES_PPGTT(dev))
2182 gen6_ppgtt_enable(dev);
2183 else if (IS_GEN7(dev))
2184 gen7_ppgtt_enable(dev);
2185 else if (INTEL_INFO(dev)->gen >= 8)
2186 gen8_ppgtt_enable(dev);
2188 MISSING_CASE(INTEL_INFO(dev)->gen);
2193 int i915_ppgtt_init_ring(struct drm_i915_gem_request *req)
2195 struct drm_i915_private *dev_priv = req->ring->dev->dev_private;
2196 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
2198 if (i915.enable_execlists)
2204 return ppgtt->switch_mm(ppgtt, req);
2207 struct i915_hw_ppgtt *
2208 i915_ppgtt_create(struct drm_device *dev, struct drm_i915_file_private *fpriv)
2210 struct i915_hw_ppgtt *ppgtt;
2213 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2215 return ERR_PTR(-ENOMEM);
2217 ret = i915_ppgtt_init(dev, ppgtt);
2220 return ERR_PTR(ret);
2223 ppgtt->file_priv = fpriv;
2225 trace_i915_ppgtt_create(&ppgtt->base);
2230 void i915_ppgtt_release(struct kref *kref)
2232 struct i915_hw_ppgtt *ppgtt =
2233 container_of(kref, struct i915_hw_ppgtt, ref);
2235 trace_i915_ppgtt_release(&ppgtt->base);
2237 /* vmas should already be unbound */
2238 WARN_ON(!list_empty(&ppgtt->base.active_list));
2239 WARN_ON(!list_empty(&ppgtt->base.inactive_list));
2241 list_del(&ppgtt->base.global_link);
2242 drm_mm_takedown(&ppgtt->base.mm);
2244 ppgtt->base.cleanup(&ppgtt->base);
2248 extern int intel_iommu_gfx_mapped;
2249 /* Certain Gen5 chipsets require require idling the GPU before
2250 * unmapping anything from the GTT when VT-d is enabled.
2252 static bool needs_idle_maps(struct drm_device *dev)
2254 #ifdef CONFIG_INTEL_IOMMU
2255 /* Query intel_iommu to see if we need the workaround. Presumably that
2258 if (IS_GEN5(dev) && IS_MOBILE(dev) && intel_iommu_gfx_mapped)
2264 static bool do_idling(struct drm_i915_private *dev_priv)
2266 bool ret = dev_priv->mm.interruptible;
2268 if (unlikely(dev_priv->gtt.do_idle_maps)) {
2269 dev_priv->mm.interruptible = false;
2270 if (i915_gpu_idle(dev_priv->dev)) {
2271 DRM_ERROR("Couldn't idle GPU\n");
2272 /* Wait a bit, in hopes it avoids the hang */
2280 static void undo_idling(struct drm_i915_private *dev_priv, bool interruptible)
2282 if (unlikely(dev_priv->gtt.do_idle_maps))
2283 dev_priv->mm.interruptible = interruptible;
2286 void i915_check_and_clear_faults(struct drm_device *dev)
2288 struct drm_i915_private *dev_priv = dev->dev_private;
2289 struct intel_engine_cs *ring;
2292 if (INTEL_INFO(dev)->gen < 6)
2295 for_each_ring(ring, dev_priv, i) {
2297 fault_reg = I915_READ(RING_FAULT_REG(ring));
2298 if (fault_reg & RING_FAULT_VALID) {
2299 DRM_DEBUG_DRIVER("Unexpected fault\n"
2301 "\tAddress space: %s\n"
2304 fault_reg & PAGE_MASK,
2305 fault_reg & RING_FAULT_GTTSEL_MASK ? "GGTT" : "PPGTT",
2306 RING_FAULT_SRCID(fault_reg),
2307 RING_FAULT_FAULT_TYPE(fault_reg));
2308 I915_WRITE(RING_FAULT_REG(ring),
2309 fault_reg & ~RING_FAULT_VALID);
2312 POSTING_READ(RING_FAULT_REG(&dev_priv->ring[RCS]));
2315 static void i915_ggtt_flush(struct drm_i915_private *dev_priv)
2317 if (INTEL_INFO(dev_priv->dev)->gen < 6) {
2318 intel_gtt_chipset_flush();
2320 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2321 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2325 void i915_gem_suspend_gtt_mappings(struct drm_device *dev)
2327 struct drm_i915_private *dev_priv = dev->dev_private;
2329 /* Don't bother messing with faults pre GEN6 as we have little
2330 * documentation supporting that it's a good idea.
2332 if (INTEL_INFO(dev)->gen < 6)
2335 i915_check_and_clear_faults(dev);
2337 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
2338 dev_priv->gtt.base.start,
2339 dev_priv->gtt.base.total,
2342 i915_ggtt_flush(dev_priv);
2345 int i915_gem_gtt_prepare_object(struct drm_i915_gem_object *obj)
2347 if (!dma_map_sg(&obj->base.dev->pdev->dev,
2348 obj->pages->sgl, obj->pages->nents,
2349 PCI_DMA_BIDIRECTIONAL))
2355 static void gen8_set_pte(void __iomem *addr, gen8_pte_t pte)
2360 iowrite32((u32)pte, addr);
2361 iowrite32(pte >> 32, addr + 4);
2365 static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
2366 struct sg_table *st,
2368 enum i915_cache_level level, u32 unused)
2370 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2371 unsigned first_entry = start >> PAGE_SHIFT;
2372 gen8_pte_t __iomem *gtt_entries =
2373 (gen8_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
2375 struct sg_page_iter sg_iter;
2376 dma_addr_t addr = 0; /* shut up gcc */
2379 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2381 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
2382 addr = sg_dma_address(sg_iter.sg) +
2383 (sg_iter.sg_pgoffset << PAGE_SHIFT);
2384 gen8_set_pte(>t_entries[i],
2385 gen8_pte_encode(addr, level, true));
2390 * XXX: This serves as a posting read to make sure that the PTE has
2391 * actually been updated. There is some concern that even though
2392 * registers and PTEs are within the same BAR that they are potentially
2393 * of NUMA access patterns. Therefore, even with the way we assume
2394 * hardware should work, we must keep this posting read for paranoia.
2397 WARN_ON(readq(>t_entries[i-1])
2398 != gen8_pte_encode(addr, level, true));
2400 /* This next bit makes the above posting read even more important. We
2401 * want to flush the TLBs only after we're certain all the PTE updates
2404 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2405 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2407 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2410 struct insert_entries {
2411 struct i915_address_space *vm;
2412 struct sg_table *st;
2414 enum i915_cache_level level;
2418 static int gen8_ggtt_insert_entries__cb(void *_arg)
2420 struct insert_entries *arg = _arg;
2421 gen8_ggtt_insert_entries(arg->vm, arg->st,
2422 arg->start, arg->level, arg->flags);
2426 static void gen8_ggtt_insert_entries__BKL(struct i915_address_space *vm,
2427 struct sg_table *st,
2429 enum i915_cache_level level,
2432 struct insert_entries arg = { vm, st, start, level, flags };
2433 stop_machine(gen8_ggtt_insert_entries__cb, &arg, NULL);
2437 * Binds an object into the global gtt with the specified cache level. The object
2438 * will be accessible to the GPU via commands whose operands reference offsets
2439 * within the global GTT as well as accessible by the GPU through the GMADR
2440 * mapped BAR (dev_priv->mm.gtt->gtt).
2442 static void gen6_ggtt_insert_entries(struct i915_address_space *vm,
2443 struct sg_table *st,
2445 enum i915_cache_level level, u32 flags)
2447 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2448 unsigned first_entry = start >> PAGE_SHIFT;
2449 gen6_pte_t __iomem *gtt_entries =
2450 (gen6_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
2452 struct sg_page_iter sg_iter;
2453 dma_addr_t addr = 0;
2456 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2458 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
2459 addr = sg_page_iter_dma_address(&sg_iter);
2460 iowrite32(vm->pte_encode(addr, level, true, flags), >t_entries[i]);
2464 /* XXX: This serves as a posting read to make sure that the PTE has
2465 * actually been updated. There is some concern that even though
2466 * registers and PTEs are within the same BAR that they are potentially
2467 * of NUMA access patterns. Therefore, even with the way we assume
2468 * hardware should work, we must keep this posting read for paranoia.
2471 unsigned long gtt = readl(>t_entries[i-1]);
2472 WARN_ON(gtt != vm->pte_encode(addr, level, true, flags));
2475 /* This next bit makes the above posting read even more important. We
2476 * want to flush the TLBs only after we're certain all the PTE updates
2479 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2480 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2482 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2485 static void gen8_ggtt_clear_range(struct i915_address_space *vm,
2490 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2491 unsigned first_entry = start >> PAGE_SHIFT;
2492 unsigned num_entries = length >> PAGE_SHIFT;
2493 gen8_pte_t scratch_pte, __iomem *gtt_base =
2494 (gen8_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
2495 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
2499 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2501 if (WARN(num_entries > max_entries,
2502 "First entry = %d; Num entries = %d (max=%d)\n",
2503 first_entry, num_entries, max_entries))
2504 num_entries = max_entries;
2506 scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
2509 for (i = 0; i < num_entries; i++)
2510 gen8_set_pte(>t_base[i], scratch_pte);
2513 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2516 static void gen6_ggtt_clear_range(struct i915_address_space *vm,
2521 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2522 unsigned first_entry = start >> PAGE_SHIFT;
2523 unsigned num_entries = length >> PAGE_SHIFT;
2524 gen6_pte_t scratch_pte, __iomem *gtt_base =
2525 (gen6_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
2526 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
2530 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2532 if (WARN(num_entries > max_entries,
2533 "First entry = %d; Num entries = %d (max=%d)\n",
2534 first_entry, num_entries, max_entries))
2535 num_entries = max_entries;
2537 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
2538 I915_CACHE_LLC, use_scratch, 0);
2540 for (i = 0; i < num_entries; i++)
2541 iowrite32(scratch_pte, >t_base[i]);
2544 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2547 static void i915_ggtt_insert_entries(struct i915_address_space *vm,
2548 struct sg_table *pages,
2550 enum i915_cache_level cache_level, u32 unused)
2552 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2553 unsigned int flags = (cache_level == I915_CACHE_NONE) ?
2554 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
2557 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2559 intel_gtt_insert_sg_entries(pages, start >> PAGE_SHIFT, flags);
2561 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2565 static void i915_ggtt_clear_range(struct i915_address_space *vm,
2570 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2571 unsigned first_entry = start >> PAGE_SHIFT;
2572 unsigned num_entries = length >> PAGE_SHIFT;
2575 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2577 intel_gtt_clear_range(first_entry, num_entries);
2579 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2582 static int ggtt_bind_vma(struct i915_vma *vma,
2583 enum i915_cache_level cache_level,
2586 struct drm_i915_gem_object *obj = vma->obj;
2590 ret = i915_get_ggtt_vma_pages(vma);
2594 /* Currently applicable only to VLV */
2596 pte_flags |= PTE_READ_ONLY;
2598 vma->vm->insert_entries(vma->vm, vma->ggtt_view.pages,
2600 cache_level, pte_flags);
2603 * Without aliasing PPGTT there's no difference between
2604 * GLOBAL/LOCAL_BIND, it's all the same ptes. Hence unconditionally
2605 * upgrade to both bound if we bind either to avoid double-binding.
2607 vma->bound |= GLOBAL_BIND | LOCAL_BIND;
2612 static int aliasing_gtt_bind_vma(struct i915_vma *vma,
2613 enum i915_cache_level cache_level,
2616 struct drm_device *dev = vma->vm->dev;
2617 struct drm_i915_private *dev_priv = dev->dev_private;
2618 struct drm_i915_gem_object *obj = vma->obj;
2619 struct sg_table *pages = obj->pages;
2623 ret = i915_get_ggtt_vma_pages(vma);
2626 pages = vma->ggtt_view.pages;
2628 /* Currently applicable only to VLV */
2630 pte_flags |= PTE_READ_ONLY;
2633 if (flags & GLOBAL_BIND) {
2634 vma->vm->insert_entries(vma->vm, pages,
2636 cache_level, pte_flags);
2639 if (flags & LOCAL_BIND) {
2640 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
2641 appgtt->base.insert_entries(&appgtt->base, pages,
2643 cache_level, pte_flags);
2649 static void ggtt_unbind_vma(struct i915_vma *vma)
2651 struct drm_device *dev = vma->vm->dev;
2652 struct drm_i915_private *dev_priv = dev->dev_private;
2653 struct drm_i915_gem_object *obj = vma->obj;
2654 const uint64_t size = min_t(uint64_t,
2658 if (vma->bound & GLOBAL_BIND) {
2659 vma->vm->clear_range(vma->vm,
2665 if (dev_priv->mm.aliasing_ppgtt && vma->bound & LOCAL_BIND) {
2666 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
2668 appgtt->base.clear_range(&appgtt->base,
2675 void i915_gem_gtt_finish_object(struct drm_i915_gem_object *obj)
2677 struct drm_device *dev = obj->base.dev;
2678 struct drm_i915_private *dev_priv = dev->dev_private;
2681 interruptible = do_idling(dev_priv);
2683 dma_unmap_sg(&dev->pdev->dev, obj->pages->sgl, obj->pages->nents,
2684 PCI_DMA_BIDIRECTIONAL);
2686 undo_idling(dev_priv, interruptible);
2689 static void i915_gtt_color_adjust(struct drm_mm_node *node,
2690 unsigned long color,
2694 if (node->color != color)
2697 if (!list_empty(&node->node_list)) {
2698 node = list_entry(node->node_list.next,
2701 if (node->allocated && node->color != color)
2706 static int i915_gem_setup_global_gtt(struct drm_device *dev,
2711 /* Let GEM Manage all of the aperture.
2713 * However, leave one page at the end still bound to the scratch page.
2714 * There are a number of places where the hardware apparently prefetches
2715 * past the end of the object, and we've seen multiple hangs with the
2716 * GPU head pointer stuck in a batchbuffer bound at the last page of the
2717 * aperture. One page should be enough to keep any prefetching inside
2720 struct drm_i915_private *dev_priv = dev->dev_private;
2721 struct i915_address_space *ggtt_vm = &dev_priv->gtt.base;
2722 struct drm_mm_node *entry;
2723 struct drm_i915_gem_object *obj;
2724 unsigned long hole_start, hole_end;
2727 BUG_ON(mappable_end > end);
2729 ggtt_vm->start = start;
2731 /* Subtract the guard page before address space initialization to
2732 * shrink the range used by drm_mm */
2733 ggtt_vm->total = end - start - PAGE_SIZE;
2734 i915_address_space_init(ggtt_vm, dev_priv);
2735 ggtt_vm->total += PAGE_SIZE;
2737 if (intel_vgpu_active(dev)) {
2738 ret = intel_vgt_balloon(dev);
2744 ggtt_vm->mm.color_adjust = i915_gtt_color_adjust;
2746 /* Mark any preallocated objects as occupied */
2747 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
2748 struct i915_vma *vma = i915_gem_obj_to_vma(obj, ggtt_vm);
2750 DRM_DEBUG_KMS("reserving preallocated space: %llx + %zx\n",
2751 i915_gem_obj_ggtt_offset(obj), obj->base.size);
2753 WARN_ON(i915_gem_obj_ggtt_bound(obj));
2754 ret = drm_mm_reserve_node(&ggtt_vm->mm, &vma->node);
2756 DRM_DEBUG_KMS("Reservation failed: %i\n", ret);
2759 vma->bound |= GLOBAL_BIND;
2760 __i915_vma_set_map_and_fenceable(vma);
2761 list_add_tail(&vma->vm_link, &ggtt_vm->inactive_list);
2764 /* Clear any non-preallocated blocks */
2765 drm_mm_for_each_hole(entry, &ggtt_vm->mm, hole_start, hole_end) {
2766 DRM_DEBUG_KMS("clearing unused GTT space: [%lx, %lx]\n",
2767 hole_start, hole_end);
2768 ggtt_vm->clear_range(ggtt_vm, hole_start,
2769 hole_end - hole_start, true);
2772 /* And finally clear the reserved guard page */
2773 ggtt_vm->clear_range(ggtt_vm, end - PAGE_SIZE, PAGE_SIZE, true);
2775 if (USES_PPGTT(dev) && !USES_FULL_PPGTT(dev)) {
2776 struct i915_hw_ppgtt *ppgtt;
2778 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2782 ret = __hw_ppgtt_init(dev, ppgtt);
2784 ppgtt->base.cleanup(&ppgtt->base);
2789 if (ppgtt->base.allocate_va_range)
2790 ret = ppgtt->base.allocate_va_range(&ppgtt->base, 0,
2793 ppgtt->base.cleanup(&ppgtt->base);
2798 ppgtt->base.clear_range(&ppgtt->base,
2803 dev_priv->mm.aliasing_ppgtt = ppgtt;
2804 WARN_ON(dev_priv->gtt.base.bind_vma != ggtt_bind_vma);
2805 dev_priv->gtt.base.bind_vma = aliasing_gtt_bind_vma;
2811 void i915_gem_init_global_gtt(struct drm_device *dev)
2813 struct drm_i915_private *dev_priv = dev->dev_private;
2814 u64 gtt_size, mappable_size;
2816 gtt_size = dev_priv->gtt.base.total;
2817 mappable_size = dev_priv->gtt.mappable_end;
2819 i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
2822 void i915_global_gtt_cleanup(struct drm_device *dev)
2824 struct drm_i915_private *dev_priv = dev->dev_private;
2825 struct i915_address_space *vm = &dev_priv->gtt.base;
2827 if (dev_priv->mm.aliasing_ppgtt) {
2828 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
2830 ppgtt->base.cleanup(&ppgtt->base);
2833 i915_gem_cleanup_stolen(dev);
2835 if (drm_mm_initialized(&vm->mm)) {
2836 if (intel_vgpu_active(dev))
2837 intel_vgt_deballoon();
2839 drm_mm_takedown(&vm->mm);
2840 list_del(&vm->global_link);
2846 static unsigned int gen6_get_total_gtt_size(u16 snb_gmch_ctl)
2848 snb_gmch_ctl >>= SNB_GMCH_GGMS_SHIFT;
2849 snb_gmch_ctl &= SNB_GMCH_GGMS_MASK;
2850 return snb_gmch_ctl << 20;
2853 static unsigned int gen8_get_total_gtt_size(u16 bdw_gmch_ctl)
2855 bdw_gmch_ctl >>= BDW_GMCH_GGMS_SHIFT;
2856 bdw_gmch_ctl &= BDW_GMCH_GGMS_MASK;
2858 bdw_gmch_ctl = 1 << bdw_gmch_ctl;
2860 #ifdef CONFIG_X86_32
2861 /* Limit 32b platforms to a 2GB GGTT: 4 << 20 / pte size * PAGE_SIZE */
2862 if (bdw_gmch_ctl > 4)
2866 return bdw_gmch_ctl << 20;
2869 static unsigned int chv_get_total_gtt_size(u16 gmch_ctrl)
2871 gmch_ctrl >>= SNB_GMCH_GGMS_SHIFT;
2872 gmch_ctrl &= SNB_GMCH_GGMS_MASK;
2875 return 1 << (20 + gmch_ctrl);
2880 static size_t gen6_get_stolen_size(u16 snb_gmch_ctl)
2882 snb_gmch_ctl >>= SNB_GMCH_GMS_SHIFT;
2883 snb_gmch_ctl &= SNB_GMCH_GMS_MASK;
2884 return snb_gmch_ctl << 25; /* 32 MB units */
2887 static size_t gen8_get_stolen_size(u16 bdw_gmch_ctl)
2889 bdw_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2890 bdw_gmch_ctl &= BDW_GMCH_GMS_MASK;
2891 return bdw_gmch_ctl << 25; /* 32 MB units */
2894 static size_t chv_get_stolen_size(u16 gmch_ctrl)
2896 gmch_ctrl >>= SNB_GMCH_GMS_SHIFT;
2897 gmch_ctrl &= SNB_GMCH_GMS_MASK;
2900 * 0x0 to 0x10: 32MB increments starting at 0MB
2901 * 0x11 to 0x16: 4MB increments starting at 8MB
2902 * 0x17 to 0x1d: 4MB increments start at 36MB
2904 if (gmch_ctrl < 0x11)
2905 return gmch_ctrl << 25;
2906 else if (gmch_ctrl < 0x17)
2907 return (gmch_ctrl - 0x11 + 2) << 22;
2909 return (gmch_ctrl - 0x17 + 9) << 22;
2912 static size_t gen9_get_stolen_size(u16 gen9_gmch_ctl)
2914 gen9_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2915 gen9_gmch_ctl &= BDW_GMCH_GMS_MASK;
2917 if (gen9_gmch_ctl < 0xf0)
2918 return gen9_gmch_ctl << 25; /* 32 MB units */
2920 /* 4MB increments starting at 0xf0 for 4MB */
2921 return (gen9_gmch_ctl - 0xf0 + 1) << 22;
2924 static int ggtt_probe_common(struct drm_device *dev,
2927 struct drm_i915_private *dev_priv = dev->dev_private;
2928 struct i915_page_scratch *scratch_page;
2929 phys_addr_t gtt_phys_addr;
2931 /* For Modern GENs the PTEs and register space are split in the BAR */
2932 gtt_phys_addr = pci_resource_start(dev->pdev, 0) +
2933 (pci_resource_len(dev->pdev, 0) / 2);
2936 * On BXT writes larger than 64 bit to the GTT pagetable range will be
2937 * dropped. For WC mappings in general we have 64 byte burst writes
2938 * when the WC buffer is flushed, so we can't use it, but have to
2939 * resort to an uncached mapping. The WC issue is easily caught by the
2940 * readback check when writing GTT PTE entries.
2942 if (IS_BROXTON(dev))
2943 dev_priv->gtt.gsm = ioremap_nocache(gtt_phys_addr, gtt_size);
2945 dev_priv->gtt.gsm = ioremap_wc(gtt_phys_addr, gtt_size);
2946 if (!dev_priv->gtt.gsm) {
2947 DRM_ERROR("Failed to map the gtt page table\n");
2951 scratch_page = alloc_scratch_page(dev);
2952 if (IS_ERR(scratch_page)) {
2953 DRM_ERROR("Scratch setup failed\n");
2954 /* iounmap will also get called at remove, but meh */
2955 iounmap(dev_priv->gtt.gsm);
2956 return PTR_ERR(scratch_page);
2959 dev_priv->gtt.base.scratch_page = scratch_page;
2964 /* The GGTT and PPGTT need a private PPAT setup in order to handle cacheability
2965 * bits. When using advanced contexts each context stores its own PAT, but
2966 * writing this data shouldn't be harmful even in those cases. */
2967 static void bdw_setup_private_ppat(struct drm_i915_private *dev_priv)
2971 pat = GEN8_PPAT(0, GEN8_PPAT_WB | GEN8_PPAT_LLC) | /* for normal objects, no eLLC */
2972 GEN8_PPAT(1, GEN8_PPAT_WC | GEN8_PPAT_LLCELLC) | /* for something pointing to ptes? */
2973 GEN8_PPAT(2, GEN8_PPAT_WT | GEN8_PPAT_LLCELLC) | /* for scanout with eLLC */
2974 GEN8_PPAT(3, GEN8_PPAT_UC) | /* Uncached objects, mostly for scanout */
2975 GEN8_PPAT(4, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(0)) |
2976 GEN8_PPAT(5, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(1)) |
2977 GEN8_PPAT(6, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(2)) |
2978 GEN8_PPAT(7, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3));
2980 if (!USES_PPGTT(dev_priv->dev))
2981 /* Spec: "For GGTT, there is NO pat_sel[2:0] from the entry,
2982 * so RTL will always use the value corresponding to
2984 * So let's disable cache for GGTT to avoid screen corruptions.
2985 * MOCS still can be used though.
2986 * - System agent ggtt writes (i.e. cpu gtt mmaps) already work
2987 * before this patch, i.e. the same uncached + snooping access
2988 * like on gen6/7 seems to be in effect.
2989 * - So this just fixes blitter/render access. Again it looks
2990 * like it's not just uncached access, but uncached + snooping.
2991 * So we can still hold onto all our assumptions wrt cpu
2992 * clflushing on LLC machines.
2994 pat = GEN8_PPAT(0, GEN8_PPAT_UC);
2996 /* XXX: spec defines this as 2 distinct registers. It's unclear if a 64b
2997 * write would work. */
2998 I915_WRITE(GEN8_PRIVATE_PAT_LO, pat);
2999 I915_WRITE(GEN8_PRIVATE_PAT_HI, pat >> 32);
3002 static void chv_setup_private_ppat(struct drm_i915_private *dev_priv)
3007 * Map WB on BDW to snooped on CHV.
3009 * Only the snoop bit has meaning for CHV, the rest is
3012 * The hardware will never snoop for certain types of accesses:
3013 * - CPU GTT (GMADR->GGTT->no snoop->memory)
3014 * - PPGTT page tables
3015 * - some other special cycles
3017 * As with BDW, we also need to consider the following for GT accesses:
3018 * "For GGTT, there is NO pat_sel[2:0] from the entry,
3019 * so RTL will always use the value corresponding to
3021 * Which means we must set the snoop bit in PAT entry 0
3022 * in order to keep the global status page working.
3024 pat = GEN8_PPAT(0, CHV_PPAT_SNOOP) |
3028 GEN8_PPAT(4, CHV_PPAT_SNOOP) |
3029 GEN8_PPAT(5, CHV_PPAT_SNOOP) |
3030 GEN8_PPAT(6, CHV_PPAT_SNOOP) |
3031 GEN8_PPAT(7, CHV_PPAT_SNOOP);
3033 I915_WRITE(GEN8_PRIVATE_PAT_LO, pat);
3034 I915_WRITE(GEN8_PRIVATE_PAT_HI, pat >> 32);
3037 static int gen8_gmch_probe(struct drm_device *dev,
3040 phys_addr_t *mappable_base,
3043 struct drm_i915_private *dev_priv = dev->dev_private;
3048 /* TODO: We're not aware of mappable constraints on gen8 yet */
3049 *mappable_base = pci_resource_start(dev->pdev, 2);
3050 *mappable_end = pci_resource_len(dev->pdev, 2);
3052 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(39)))
3053 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(39));
3055 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
3057 if (INTEL_INFO(dev)->gen >= 9) {
3058 *stolen = gen9_get_stolen_size(snb_gmch_ctl);
3059 gtt_size = gen8_get_total_gtt_size(snb_gmch_ctl);
3060 } else if (IS_CHERRYVIEW(dev)) {
3061 *stolen = chv_get_stolen_size(snb_gmch_ctl);
3062 gtt_size = chv_get_total_gtt_size(snb_gmch_ctl);
3064 *stolen = gen8_get_stolen_size(snb_gmch_ctl);
3065 gtt_size = gen8_get_total_gtt_size(snb_gmch_ctl);
3068 *gtt_total = (gtt_size / sizeof(gen8_pte_t)) << PAGE_SHIFT;
3070 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
3071 chv_setup_private_ppat(dev_priv);
3073 bdw_setup_private_ppat(dev_priv);
3075 ret = ggtt_probe_common(dev, gtt_size);
3077 dev_priv->gtt.base.clear_range = gen8_ggtt_clear_range;
3078 dev_priv->gtt.base.insert_entries = gen8_ggtt_insert_entries;
3079 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
3080 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
3082 if (IS_CHERRYVIEW(dev_priv))
3083 dev_priv->gtt.base.insert_entries = gen8_ggtt_insert_entries__BKL;
3088 static int gen6_gmch_probe(struct drm_device *dev,
3091 phys_addr_t *mappable_base,
3094 struct drm_i915_private *dev_priv = dev->dev_private;
3095 unsigned int gtt_size;
3099 *mappable_base = pci_resource_start(dev->pdev, 2);
3100 *mappable_end = pci_resource_len(dev->pdev, 2);
3102 /* 64/512MB is the current min/max we actually know of, but this is just
3103 * a coarse sanity check.
3105 if ((*mappable_end < (64<<20) || (*mappable_end > (512<<20)))) {
3106 DRM_ERROR("Unknown GMADR size (%llx)\n",
3107 dev_priv->gtt.mappable_end);
3111 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(40)))
3112 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(40));
3113 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
3115 *stolen = gen6_get_stolen_size(snb_gmch_ctl);
3117 gtt_size = gen6_get_total_gtt_size(snb_gmch_ctl);
3118 *gtt_total = (gtt_size / sizeof(gen6_pte_t)) << PAGE_SHIFT;
3120 ret = ggtt_probe_common(dev, gtt_size);
3122 dev_priv->gtt.base.clear_range = gen6_ggtt_clear_range;
3123 dev_priv->gtt.base.insert_entries = gen6_ggtt_insert_entries;
3124 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
3125 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
3130 static void gen6_gmch_remove(struct i915_address_space *vm)
3133 struct i915_gtt *gtt = container_of(vm, struct i915_gtt, base);
3136 free_scratch_page(vm->dev, vm->scratch_page);
3139 static int i915_gmch_probe(struct drm_device *dev,
3142 phys_addr_t *mappable_base,
3145 struct drm_i915_private *dev_priv = dev->dev_private;
3148 ret = intel_gmch_probe(dev_priv->bridge_dev, dev_priv->dev->pdev, NULL);
3150 DRM_ERROR("failed to set up gmch\n");
3154 intel_gtt_get(gtt_total, stolen, mappable_base, mappable_end);
3156 dev_priv->gtt.do_idle_maps = needs_idle_maps(dev_priv->dev);
3157 dev_priv->gtt.base.insert_entries = i915_ggtt_insert_entries;
3158 dev_priv->gtt.base.clear_range = i915_ggtt_clear_range;
3159 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
3160 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
3162 if (unlikely(dev_priv->gtt.do_idle_maps))
3163 DRM_INFO("applying Ironlake quirks for intel_iommu\n");
3168 static void i915_gmch_remove(struct i915_address_space *vm)
3170 intel_gmch_remove();
3173 int i915_gem_gtt_init(struct drm_device *dev)
3175 struct drm_i915_private *dev_priv = dev->dev_private;
3176 struct i915_gtt *gtt = &dev_priv->gtt;
3179 if (INTEL_INFO(dev)->gen <= 5) {
3180 gtt->gtt_probe = i915_gmch_probe;
3181 gtt->base.cleanup = i915_gmch_remove;
3182 } else if (INTEL_INFO(dev)->gen < 8) {
3183 gtt->gtt_probe = gen6_gmch_probe;
3184 gtt->base.cleanup = gen6_gmch_remove;
3185 if (IS_HASWELL(dev) && dev_priv->ellc_size)
3186 gtt->base.pte_encode = iris_pte_encode;
3187 else if (IS_HASWELL(dev))
3188 gtt->base.pte_encode = hsw_pte_encode;
3189 else if (IS_VALLEYVIEW(dev))
3190 gtt->base.pte_encode = byt_pte_encode;
3191 else if (INTEL_INFO(dev)->gen >= 7)
3192 gtt->base.pte_encode = ivb_pte_encode;
3194 gtt->base.pte_encode = snb_pte_encode;
3196 dev_priv->gtt.gtt_probe = gen8_gmch_probe;
3197 dev_priv->gtt.base.cleanup = gen6_gmch_remove;
3200 gtt->base.dev = dev;
3201 gtt->base.is_ggtt = true;
3203 ret = gtt->gtt_probe(dev, >t->base.total, >t->stolen_size,
3204 >t->mappable_base, >t->mappable_end);
3209 * Initialise stolen early so that we may reserve preallocated
3210 * objects for the BIOS to KMS transition.
3212 ret = i915_gem_init_stolen(dev);
3214 goto out_gtt_cleanup;
3216 /* GMADR is the PCI mmio aperture into the global GTT. */
3217 DRM_INFO("Memory usable by graphics device = %lluM\n",
3218 gtt->base.total >> 20);
3219 DRM_DEBUG_DRIVER("GMADR size = %lldM\n", gtt->mappable_end >> 20);
3220 DRM_DEBUG_DRIVER("GTT stolen size = %zdM\n", gtt->stolen_size >> 20);
3221 #ifdef CONFIG_INTEL_IOMMU
3222 if (intel_iommu_gfx_mapped)
3223 DRM_INFO("VT-d active for gfx access\n");
3226 * i915.enable_ppgtt is read-only, so do an early pass to validate the
3227 * user's requested state against the hardware/driver capabilities. We
3228 * do this now so that we can print out any log messages once rather
3229 * than every time we check intel_enable_ppgtt().
3231 i915.enable_ppgtt = sanitize_enable_ppgtt(dev, i915.enable_ppgtt);
3232 DRM_DEBUG_DRIVER("ppgtt mode: %i\n", i915.enable_ppgtt);
3237 gtt->base.cleanup(&dev_priv->gtt.base);
3242 void i915_gem_restore_gtt_mappings(struct drm_device *dev)
3244 struct drm_i915_private *dev_priv = dev->dev_private;
3245 struct drm_i915_gem_object *obj;
3246 struct i915_address_space *vm;
3247 struct i915_vma *vma;
3250 i915_check_and_clear_faults(dev);
3252 /* First fill our portion of the GTT with scratch pages */
3253 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
3254 dev_priv->gtt.base.start,
3255 dev_priv->gtt.base.total,
3258 /* Cache flush objects bound into GGTT and rebind them. */
3259 vm = &dev_priv->gtt.base;
3260 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
3262 list_for_each_entry(vma, &obj->vma_list, obj_link) {
3266 WARN_ON(i915_vma_bind(vma, obj->cache_level,
3273 i915_gem_clflush_object(obj, obj->pin_display);
3276 if (INTEL_INFO(dev)->gen >= 8) {
3277 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
3278 chv_setup_private_ppat(dev_priv);
3280 bdw_setup_private_ppat(dev_priv);
3285 if (USES_PPGTT(dev)) {
3286 list_for_each_entry(vm, &dev_priv->vm_list, global_link) {
3287 /* TODO: Perhaps it shouldn't be gen6 specific */
3289 struct i915_hw_ppgtt *ppgtt =
3290 container_of(vm, struct i915_hw_ppgtt,
3293 if (i915_is_ggtt(vm))
3294 ppgtt = dev_priv->mm.aliasing_ppgtt;
3296 gen6_write_page_range(dev_priv, &ppgtt->pd,
3297 0, ppgtt->base.total);
3301 i915_ggtt_flush(dev_priv);
3304 static struct i915_vma *
3305 __i915_gem_vma_create(struct drm_i915_gem_object *obj,
3306 struct i915_address_space *vm,
3307 const struct i915_ggtt_view *ggtt_view)
3309 struct i915_vma *vma;
3311 if (WARN_ON(i915_is_ggtt(vm) != !!ggtt_view))
3312 return ERR_PTR(-EINVAL);
3314 vma = kmem_cache_zalloc(to_i915(obj->base.dev)->vmas, GFP_KERNEL);
3316 return ERR_PTR(-ENOMEM);
3318 INIT_LIST_HEAD(&vma->vm_link);
3319 INIT_LIST_HEAD(&vma->obj_link);
3320 INIT_LIST_HEAD(&vma->exec_list);
3323 vma->is_ggtt = i915_is_ggtt(vm);
3325 if (i915_is_ggtt(vm))
3326 vma->ggtt_view = *ggtt_view;
3328 i915_ppgtt_get(i915_vm_to_ppgtt(vm));
3330 list_add_tail(&vma->obj_link, &obj->vma_list);
3336 i915_gem_obj_lookup_or_create_vma(struct drm_i915_gem_object *obj,
3337 struct i915_address_space *vm)
3339 struct i915_vma *vma;
3341 vma = i915_gem_obj_to_vma(obj, vm);
3343 vma = __i915_gem_vma_create(obj, vm,
3344 i915_is_ggtt(vm) ? &i915_ggtt_view_normal : NULL);
3350 i915_gem_obj_lookup_or_create_ggtt_vma(struct drm_i915_gem_object *obj,
3351 const struct i915_ggtt_view *view)
3353 struct i915_address_space *ggtt = i915_obj_to_ggtt(obj);
3354 struct i915_vma *vma;
3357 return ERR_PTR(-EINVAL);
3359 vma = i915_gem_obj_to_ggtt_view(obj, view);
3365 vma = __i915_gem_vma_create(obj, ggtt, view);
3371 static struct scatterlist *
3372 rotate_pages(const dma_addr_t *in, unsigned int offset,
3373 unsigned int width, unsigned int height,
3374 unsigned int stride,
3375 struct sg_table *st, struct scatterlist *sg)
3377 unsigned int column, row;
3378 unsigned int src_idx;
3380 for (column = 0; column < width; column++) {
3381 src_idx = stride * (height - 1) + column;
3382 for (row = 0; row < height; row++) {
3384 /* We don't need the pages, but need to initialize
3385 * the entries so the sg list can be happily traversed.
3386 * The only thing we need are DMA addresses.
3388 sg_set_page(sg, NULL, PAGE_SIZE, 0);
3389 sg_dma_address(sg) = in[offset + src_idx];
3390 sg_dma_len(sg) = PAGE_SIZE;
3399 static struct sg_table *
3400 intel_rotate_fb_obj_pages(struct intel_rotation_info *rot_info,
3401 struct drm_i915_gem_object *obj)
3403 unsigned int size_pages = rot_info->plane[0].width * rot_info->plane[0].height;
3404 unsigned int size_pages_uv;
3405 struct sg_page_iter sg_iter;
3407 dma_addr_t *page_addr_list;
3408 struct sg_table *st;
3409 unsigned int uv_start_page;
3410 struct scatterlist *sg;
3413 /* Allocate a temporary list of source pages for random access. */
3414 page_addr_list = drm_malloc_ab(obj->base.size / PAGE_SIZE,
3415 sizeof(dma_addr_t));
3416 if (!page_addr_list)
3417 return ERR_PTR(ret);
3419 /* Account for UV plane with NV12. */
3420 if (rot_info->pixel_format == DRM_FORMAT_NV12)
3421 size_pages_uv = rot_info->plane[1].width * rot_info->plane[1].height;
3425 /* Allocate target SG list. */
3426 st = kmalloc(sizeof(*st), GFP_KERNEL);
3430 ret = sg_alloc_table(st, size_pages + size_pages_uv, GFP_KERNEL);
3434 /* Populate source page list from the object. */
3436 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) {
3437 page_addr_list[i] = sg_page_iter_dma_address(&sg_iter);
3444 /* Rotate the pages. */
3445 sg = rotate_pages(page_addr_list, 0,
3446 rot_info->plane[0].width, rot_info->plane[0].height,
3447 rot_info->plane[0].width,
3450 /* Append the UV plane if NV12. */
3451 if (rot_info->pixel_format == DRM_FORMAT_NV12) {
3452 uv_start_page = size_pages;
3454 /* Check for tile-row un-alignment. */
3455 if (offset_in_page(rot_info->uv_offset))
3458 rot_info->uv_start_page = uv_start_page;
3460 sg = rotate_pages(page_addr_list, rot_info->uv_start_page,
3461 rot_info->plane[1].width, rot_info->plane[1].height,
3462 rot_info->plane[1].width,
3466 DRM_DEBUG_KMS("Created rotated page mapping for object size %zu (%ux%u tiles, %u pages (%u plane 0)).\n",
3467 obj->base.size, rot_info->plane[0].width,
3468 rot_info->plane[0].height, size_pages + size_pages_uv,
3471 drm_free_large(page_addr_list);
3478 drm_free_large(page_addr_list);
3480 DRM_DEBUG_KMS("Failed to create rotated mapping for object size %zu! (%d) (%ux%u tiles, %u pages (%u plane 0))\n",
3481 obj->base.size, ret, rot_info->plane[0].width,
3482 rot_info->plane[0].height, size_pages + size_pages_uv,
3484 return ERR_PTR(ret);
3487 static struct sg_table *
3488 intel_partial_pages(const struct i915_ggtt_view *view,
3489 struct drm_i915_gem_object *obj)
3491 struct sg_table *st;
3492 struct scatterlist *sg;
3493 struct sg_page_iter obj_sg_iter;
3496 st = kmalloc(sizeof(*st), GFP_KERNEL);
3500 ret = sg_alloc_table(st, view->params.partial.size, GFP_KERNEL);
3506 for_each_sg_page(obj->pages->sgl, &obj_sg_iter, obj->pages->nents,
3507 view->params.partial.offset)
3509 if (st->nents >= view->params.partial.size)
3512 sg_set_page(sg, NULL, PAGE_SIZE, 0);
3513 sg_dma_address(sg) = sg_page_iter_dma_address(&obj_sg_iter);
3514 sg_dma_len(sg) = PAGE_SIZE;
3525 return ERR_PTR(ret);
3529 i915_get_ggtt_vma_pages(struct i915_vma *vma)
3533 if (vma->ggtt_view.pages)
3536 if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL)
3537 vma->ggtt_view.pages = vma->obj->pages;
3538 else if (vma->ggtt_view.type == I915_GGTT_VIEW_ROTATED)
3539 vma->ggtt_view.pages =
3540 intel_rotate_fb_obj_pages(&vma->ggtt_view.params.rotated, vma->obj);
3541 else if (vma->ggtt_view.type == I915_GGTT_VIEW_PARTIAL)
3542 vma->ggtt_view.pages =
3543 intel_partial_pages(&vma->ggtt_view, vma->obj);
3545 WARN_ONCE(1, "GGTT view %u not implemented!\n",
3546 vma->ggtt_view.type);
3548 if (!vma->ggtt_view.pages) {
3549 DRM_ERROR("Failed to get pages for GGTT view type %u!\n",
3550 vma->ggtt_view.type);
3552 } else if (IS_ERR(vma->ggtt_view.pages)) {
3553 ret = PTR_ERR(vma->ggtt_view.pages);
3554 vma->ggtt_view.pages = NULL;
3555 DRM_ERROR("Failed to get pages for VMA view type %u (%d)!\n",
3556 vma->ggtt_view.type, ret);
3563 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
3565 * @cache_level: mapping cache level
3566 * @flags: flags like global or local mapping
3568 * DMA addresses are taken from the scatter-gather table of this object (or of
3569 * this VMA in case of non-default GGTT views) and PTE entries set up.
3570 * Note that DMA addresses are also the only part of the SG table we care about.
3572 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
3578 if (WARN_ON(flags == 0))
3582 if (flags & PIN_GLOBAL)
3583 bind_flags |= GLOBAL_BIND;
3584 if (flags & PIN_USER)
3585 bind_flags |= LOCAL_BIND;
3587 if (flags & PIN_UPDATE)
3588 bind_flags |= vma->bound;
3590 bind_flags &= ~vma->bound;
3592 if (bind_flags == 0)
3595 if (vma->bound == 0 && vma->vm->allocate_va_range) {
3596 /* XXX: i915_vma_pin() will fix this +- hack */
3598 trace_i915_va_alloc(vma);
3599 ret = vma->vm->allocate_va_range(vma->vm,
3607 ret = vma->vm->bind_vma(vma, cache_level, bind_flags);
3611 vma->bound |= bind_flags;
3617 * i915_ggtt_view_size - Get the size of a GGTT view.
3618 * @obj: Object the view is of.
3619 * @view: The view in question.
3621 * @return The size of the GGTT view in bytes.
3624 i915_ggtt_view_size(struct drm_i915_gem_object *obj,
3625 const struct i915_ggtt_view *view)
3627 if (view->type == I915_GGTT_VIEW_NORMAL) {
3628 return obj->base.size;
3629 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
3630 return intel_rotation_info_size(&view->params.rotated) << PAGE_SHIFT;
3631 } else if (view->type == I915_GGTT_VIEW_PARTIAL) {
3632 return view->params.partial.size << PAGE_SHIFT;
3634 WARN_ONCE(1, "GGTT view %u not implemented!\n", view->type);
3635 return obj->base.size;