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 *engine = req->engine;
666 ret = intel_ring_begin(req, 6);
670 intel_ring_emit(engine, MI_LOAD_REGISTER_IMM(1));
671 intel_ring_emit_reg(engine, GEN8_RING_PDP_UDW(engine, entry));
672 intel_ring_emit(engine, upper_32_bits(addr));
673 intel_ring_emit(engine, MI_LOAD_REGISTER_IMM(1));
674 intel_ring_emit_reg(engine, GEN8_RING_PDP_LDW(engine, entry));
675 intel_ring_emit(engine, lower_32_bits(addr));
676 intel_ring_advance(engine);
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 = i915_vm_to_ppgtt(vm);
710 gen8_pte_t *pt_vaddr;
711 unsigned pdpe = gen8_pdpe_index(start);
712 unsigned pde = gen8_pde_index(start);
713 unsigned pte = gen8_pte_index(start);
714 unsigned num_entries = length >> PAGE_SHIFT;
715 unsigned last_pte, i;
720 while (num_entries) {
721 struct i915_page_directory *pd;
722 struct i915_page_table *pt;
724 if (WARN_ON(!pdp->page_directory[pdpe]))
727 pd = pdp->page_directory[pdpe];
729 if (WARN_ON(!pd->page_table[pde]))
732 pt = pd->page_table[pde];
734 if (WARN_ON(!px_page(pt)))
737 last_pte = pte + num_entries;
738 if (last_pte > GEN8_PTES)
739 last_pte = GEN8_PTES;
741 pt_vaddr = kmap_px(pt);
743 for (i = pte; i < last_pte; i++) {
744 pt_vaddr[i] = scratch_pte;
748 kunmap_px(ppgtt, pt_vaddr);
751 if (++pde == I915_PDES) {
752 if (++pdpe == I915_PDPES_PER_PDP(vm->dev))
759 static void gen8_ppgtt_clear_range(struct i915_address_space *vm,
764 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
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 = i915_vm_to_ppgtt(vm);
790 gen8_pte_t *pt_vaddr;
791 unsigned pdpe = gen8_pdpe_index(start);
792 unsigned pde = gen8_pde_index(start);
793 unsigned pte = gen8_pte_index(start);
797 while (__sg_page_iter_next(sg_iter)) {
798 if (pt_vaddr == NULL) {
799 struct i915_page_directory *pd = pdp->page_directory[pdpe];
800 struct i915_page_table *pt = pd->page_table[pde];
801 pt_vaddr = kmap_px(pt);
805 gen8_pte_encode(sg_page_iter_dma_address(sg_iter),
807 if (++pte == GEN8_PTES) {
808 kunmap_px(ppgtt, pt_vaddr);
810 if (++pde == I915_PDES) {
811 if (++pdpe == I915_PDPES_PER_PDP(vm->dev))
820 kunmap_px(ppgtt, pt_vaddr);
823 static void gen8_ppgtt_insert_entries(struct i915_address_space *vm,
824 struct sg_table *pages,
826 enum i915_cache_level cache_level,
829 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
830 struct sg_page_iter sg_iter;
832 __sg_page_iter_start(&sg_iter, pages->sgl, sg_nents(pages->sgl), 0);
834 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
835 gen8_ppgtt_insert_pte_entries(vm, &ppgtt->pdp, &sg_iter, start,
838 struct i915_page_directory_pointer *pdp;
840 uint64_t length = (uint64_t)pages->orig_nents << PAGE_SHIFT;
842 gen8_for_each_pml4e(pdp, &ppgtt->pml4, start, length, pml4e) {
843 gen8_ppgtt_insert_pte_entries(vm, pdp, &sg_iter,
849 static void gen8_free_page_tables(struct drm_device *dev,
850 struct i915_page_directory *pd)
857 for_each_set_bit(i, pd->used_pdes, I915_PDES) {
858 if (WARN_ON(!pd->page_table[i]))
861 free_pt(dev, pd->page_table[i]);
862 pd->page_table[i] = NULL;
866 static int gen8_init_scratch(struct i915_address_space *vm)
868 struct drm_device *dev = vm->dev;
870 vm->scratch_page = alloc_scratch_page(dev);
871 if (IS_ERR(vm->scratch_page))
872 return PTR_ERR(vm->scratch_page);
874 vm->scratch_pt = alloc_pt(dev);
875 if (IS_ERR(vm->scratch_pt)) {
876 free_scratch_page(dev, vm->scratch_page);
877 return PTR_ERR(vm->scratch_pt);
880 vm->scratch_pd = alloc_pd(dev);
881 if (IS_ERR(vm->scratch_pd)) {
882 free_pt(dev, vm->scratch_pt);
883 free_scratch_page(dev, vm->scratch_page);
884 return PTR_ERR(vm->scratch_pd);
887 if (USES_FULL_48BIT_PPGTT(dev)) {
888 vm->scratch_pdp = alloc_pdp(dev);
889 if (IS_ERR(vm->scratch_pdp)) {
890 free_pd(dev, vm->scratch_pd);
891 free_pt(dev, vm->scratch_pt);
892 free_scratch_page(dev, vm->scratch_page);
893 return PTR_ERR(vm->scratch_pdp);
897 gen8_initialize_pt(vm, vm->scratch_pt);
898 gen8_initialize_pd(vm, vm->scratch_pd);
899 if (USES_FULL_48BIT_PPGTT(dev))
900 gen8_initialize_pdp(vm, vm->scratch_pdp);
905 static int gen8_ppgtt_notify_vgt(struct i915_hw_ppgtt *ppgtt, bool create)
907 enum vgt_g2v_type msg;
908 struct drm_i915_private *dev_priv = to_i915(ppgtt->base.dev);
911 if (USES_FULL_48BIT_PPGTT(dev_priv)) {
912 u64 daddr = px_dma(&ppgtt->pml4);
914 I915_WRITE(vgtif_reg(pdp[0].lo), lower_32_bits(daddr));
915 I915_WRITE(vgtif_reg(pdp[0].hi), upper_32_bits(daddr));
917 msg = (create ? VGT_G2V_PPGTT_L4_PAGE_TABLE_CREATE :
918 VGT_G2V_PPGTT_L4_PAGE_TABLE_DESTROY);
920 for (i = 0; i < GEN8_LEGACY_PDPES; i++) {
921 u64 daddr = i915_page_dir_dma_addr(ppgtt, i);
923 I915_WRITE(vgtif_reg(pdp[i].lo), lower_32_bits(daddr));
924 I915_WRITE(vgtif_reg(pdp[i].hi), upper_32_bits(daddr));
927 msg = (create ? VGT_G2V_PPGTT_L3_PAGE_TABLE_CREATE :
928 VGT_G2V_PPGTT_L3_PAGE_TABLE_DESTROY);
931 I915_WRITE(vgtif_reg(g2v_notify), msg);
936 static void gen8_free_scratch(struct i915_address_space *vm)
938 struct drm_device *dev = vm->dev;
940 if (USES_FULL_48BIT_PPGTT(dev))
941 free_pdp(dev, vm->scratch_pdp);
942 free_pd(dev, vm->scratch_pd);
943 free_pt(dev, vm->scratch_pt);
944 free_scratch_page(dev, vm->scratch_page);
947 static void gen8_ppgtt_cleanup_3lvl(struct drm_device *dev,
948 struct i915_page_directory_pointer *pdp)
952 for_each_set_bit(i, pdp->used_pdpes, I915_PDPES_PER_PDP(dev)) {
953 if (WARN_ON(!pdp->page_directory[i]))
956 gen8_free_page_tables(dev, pdp->page_directory[i]);
957 free_pd(dev, pdp->page_directory[i]);
963 static void gen8_ppgtt_cleanup_4lvl(struct i915_hw_ppgtt *ppgtt)
967 for_each_set_bit(i, ppgtt->pml4.used_pml4es, GEN8_PML4ES_PER_PML4) {
968 if (WARN_ON(!ppgtt->pml4.pdps[i]))
971 gen8_ppgtt_cleanup_3lvl(ppgtt->base.dev, ppgtt->pml4.pdps[i]);
974 cleanup_px(ppgtt->base.dev, &ppgtt->pml4);
977 static void gen8_ppgtt_cleanup(struct i915_address_space *vm)
979 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
981 if (intel_vgpu_active(vm->dev))
982 gen8_ppgtt_notify_vgt(ppgtt, false);
984 if (!USES_FULL_48BIT_PPGTT(ppgtt->base.dev))
985 gen8_ppgtt_cleanup_3lvl(ppgtt->base.dev, &ppgtt->pdp);
987 gen8_ppgtt_cleanup_4lvl(ppgtt);
989 gen8_free_scratch(vm);
993 * gen8_ppgtt_alloc_pagetabs() - Allocate page tables for VA range.
994 * @vm: Master vm structure.
995 * @pd: Page directory for this address range.
996 * @start: Starting virtual address to begin allocations.
997 * @length: Size of the allocations.
998 * @new_pts: Bitmap set by function with new allocations. Likely used by the
999 * caller to free on error.
1001 * Allocate the required number of page tables. Extremely similar to
1002 * gen8_ppgtt_alloc_page_directories(). The main difference is here we are limited by
1003 * the page directory boundary (instead of the page directory pointer). That
1004 * boundary is 1GB virtual. Therefore, unlike gen8_ppgtt_alloc_page_directories(), it is
1005 * possible, and likely that the caller will need to use multiple calls of this
1006 * function to achieve the appropriate allocation.
1008 * Return: 0 if success; negative error code otherwise.
1010 static int gen8_ppgtt_alloc_pagetabs(struct i915_address_space *vm,
1011 struct i915_page_directory *pd,
1014 unsigned long *new_pts)
1016 struct drm_device *dev = vm->dev;
1017 struct i915_page_table *pt;
1020 gen8_for_each_pde(pt, pd, start, length, pde) {
1021 /* Don't reallocate page tables */
1022 if (test_bit(pde, pd->used_pdes)) {
1023 /* Scratch is never allocated this way */
1024 WARN_ON(pt == vm->scratch_pt);
1032 gen8_initialize_pt(vm, pt);
1033 pd->page_table[pde] = pt;
1034 __set_bit(pde, new_pts);
1035 trace_i915_page_table_entry_alloc(vm, pde, start, GEN8_PDE_SHIFT);
1041 for_each_set_bit(pde, new_pts, I915_PDES)
1042 free_pt(dev, pd->page_table[pde]);
1048 * gen8_ppgtt_alloc_page_directories() - Allocate page directories for VA range.
1049 * @vm: Master vm structure.
1050 * @pdp: Page directory pointer for this address range.
1051 * @start: Starting virtual address to begin allocations.
1052 * @length: Size of the allocations.
1053 * @new_pds: Bitmap set by function with new allocations. Likely used by the
1054 * caller to free on error.
1056 * Allocate the required number of page directories starting at the pde index of
1057 * @start, and ending at the pde index @start + @length. This function will skip
1058 * over already allocated page directories within the range, and only allocate
1059 * new ones, setting the appropriate pointer within the pdp as well as the
1060 * correct position in the bitmap @new_pds.
1062 * The function will only allocate the pages within the range for a give page
1063 * directory pointer. In other words, if @start + @length straddles a virtually
1064 * addressed PDP boundary (512GB for 4k pages), there will be more allocations
1065 * required by the caller, This is not currently possible, and the BUG in the
1066 * code will prevent it.
1068 * Return: 0 if success; negative error code otherwise.
1071 gen8_ppgtt_alloc_page_directories(struct i915_address_space *vm,
1072 struct i915_page_directory_pointer *pdp,
1075 unsigned long *new_pds)
1077 struct drm_device *dev = vm->dev;
1078 struct i915_page_directory *pd;
1080 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1082 WARN_ON(!bitmap_empty(new_pds, pdpes));
1084 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1085 if (test_bit(pdpe, pdp->used_pdpes))
1092 gen8_initialize_pd(vm, pd);
1093 pdp->page_directory[pdpe] = pd;
1094 __set_bit(pdpe, new_pds);
1095 trace_i915_page_directory_entry_alloc(vm, pdpe, start, GEN8_PDPE_SHIFT);
1101 for_each_set_bit(pdpe, new_pds, pdpes)
1102 free_pd(dev, pdp->page_directory[pdpe]);
1108 * gen8_ppgtt_alloc_page_dirpointers() - Allocate pdps for VA range.
1109 * @vm: Master vm structure.
1110 * @pml4: Page map level 4 for this address range.
1111 * @start: Starting virtual address to begin allocations.
1112 * @length: Size of the allocations.
1113 * @new_pdps: Bitmap set by function with new allocations. Likely used by the
1114 * caller to free on error.
1116 * Allocate the required number of page directory pointers. Extremely similar to
1117 * gen8_ppgtt_alloc_page_directories() and gen8_ppgtt_alloc_pagetabs().
1118 * The main difference is here we are limited by the pml4 boundary (instead of
1119 * the page directory pointer).
1121 * Return: 0 if success; negative error code otherwise.
1124 gen8_ppgtt_alloc_page_dirpointers(struct i915_address_space *vm,
1125 struct i915_pml4 *pml4,
1128 unsigned long *new_pdps)
1130 struct drm_device *dev = vm->dev;
1131 struct i915_page_directory_pointer *pdp;
1134 WARN_ON(!bitmap_empty(new_pdps, GEN8_PML4ES_PER_PML4));
1136 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1137 if (!test_bit(pml4e, pml4->used_pml4es)) {
1138 pdp = alloc_pdp(dev);
1142 gen8_initialize_pdp(vm, pdp);
1143 pml4->pdps[pml4e] = pdp;
1144 __set_bit(pml4e, new_pdps);
1145 trace_i915_page_directory_pointer_entry_alloc(vm,
1155 for_each_set_bit(pml4e, new_pdps, GEN8_PML4ES_PER_PML4)
1156 free_pdp(dev, pml4->pdps[pml4e]);
1162 free_gen8_temp_bitmaps(unsigned long *new_pds, unsigned long *new_pts)
1168 /* Fills in the page directory bitmap, and the array of page tables bitmap. Both
1169 * of these are based on the number of PDPEs in the system.
1172 int __must_check alloc_gen8_temp_bitmaps(unsigned long **new_pds,
1173 unsigned long **new_pts,
1179 pds = kcalloc(BITS_TO_LONGS(pdpes), sizeof(unsigned long), GFP_TEMPORARY);
1183 pts = kcalloc(pdpes, BITS_TO_LONGS(I915_PDES) * sizeof(unsigned long),
1194 free_gen8_temp_bitmaps(pds, pts);
1198 /* PDE TLBs are a pain to invalidate on GEN8+. When we modify
1199 * the page table structures, we mark them dirty so that
1200 * context switching/execlist queuing code takes extra steps
1201 * to ensure that tlbs are flushed.
1203 static void mark_tlbs_dirty(struct i915_hw_ppgtt *ppgtt)
1205 ppgtt->pd_dirty_rings = INTEL_INFO(ppgtt->base.dev)->ring_mask;
1208 static int gen8_alloc_va_range_3lvl(struct i915_address_space *vm,
1209 struct i915_page_directory_pointer *pdp,
1213 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1214 unsigned long *new_page_dirs, *new_page_tables;
1215 struct drm_device *dev = vm->dev;
1216 struct i915_page_directory *pd;
1217 const uint64_t orig_start = start;
1218 const uint64_t orig_length = length;
1220 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1223 /* Wrap is never okay since we can only represent 48b, and we don't
1224 * actually use the other side of the canonical address space.
1226 if (WARN_ON(start + length < start))
1229 if (WARN_ON(start + length > vm->total))
1232 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables, pdpes);
1236 /* Do the allocations first so we can easily bail out */
1237 ret = gen8_ppgtt_alloc_page_directories(vm, pdp, start, length,
1240 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1244 /* For every page directory referenced, allocate page tables */
1245 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1246 ret = gen8_ppgtt_alloc_pagetabs(vm, pd, start, length,
1247 new_page_tables + pdpe * BITS_TO_LONGS(I915_PDES));
1253 length = orig_length;
1255 /* Allocations have completed successfully, so set the bitmaps, and do
1257 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1258 gen8_pde_t *const page_directory = kmap_px(pd);
1259 struct i915_page_table *pt;
1260 uint64_t pd_len = length;
1261 uint64_t pd_start = start;
1264 /* Every pd should be allocated, we just did that above. */
1267 gen8_for_each_pde(pt, pd, pd_start, pd_len, pde) {
1268 /* Same reasoning as pd */
1271 WARN_ON(!gen8_pte_count(pd_start, pd_len));
1273 /* Set our used ptes within the page table */
1274 bitmap_set(pt->used_ptes,
1275 gen8_pte_index(pd_start),
1276 gen8_pte_count(pd_start, pd_len));
1278 /* Our pde is now pointing to the pagetable, pt */
1279 __set_bit(pde, pd->used_pdes);
1281 /* Map the PDE to the page table */
1282 page_directory[pde] = gen8_pde_encode(px_dma(pt),
1284 trace_i915_page_table_entry_map(&ppgtt->base, pde, pt,
1285 gen8_pte_index(start),
1286 gen8_pte_count(start, length),
1289 /* NB: We haven't yet mapped ptes to pages. At this
1290 * point we're still relying on insert_entries() */
1293 kunmap_px(ppgtt, page_directory);
1294 __set_bit(pdpe, pdp->used_pdpes);
1295 gen8_setup_page_directory(ppgtt, pdp, pd, pdpe);
1298 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1299 mark_tlbs_dirty(ppgtt);
1306 for_each_set_bit(temp, new_page_tables + pdpe *
1307 BITS_TO_LONGS(I915_PDES), I915_PDES)
1308 free_pt(dev, pdp->page_directory[pdpe]->page_table[temp]);
1311 for_each_set_bit(pdpe, new_page_dirs, pdpes)
1312 free_pd(dev, pdp->page_directory[pdpe]);
1314 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1315 mark_tlbs_dirty(ppgtt);
1319 static int gen8_alloc_va_range_4lvl(struct i915_address_space *vm,
1320 struct i915_pml4 *pml4,
1324 DECLARE_BITMAP(new_pdps, GEN8_PML4ES_PER_PML4);
1325 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1326 struct i915_page_directory_pointer *pdp;
1330 /* Do the pml4 allocations first, so we don't need to track the newly
1331 * allocated tables below the pdp */
1332 bitmap_zero(new_pdps, GEN8_PML4ES_PER_PML4);
1334 /* The pagedirectory and pagetable allocations are done in the shared 3
1335 * and 4 level code. Just allocate the pdps.
1337 ret = gen8_ppgtt_alloc_page_dirpointers(vm, pml4, start, length,
1342 WARN(bitmap_weight(new_pdps, GEN8_PML4ES_PER_PML4) > 2,
1343 "The allocation has spanned more than 512GB. "
1344 "It is highly likely this is incorrect.");
1346 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1349 ret = gen8_alloc_va_range_3lvl(vm, pdp, start, length);
1353 gen8_setup_page_directory_pointer(ppgtt, pml4, pdp, pml4e);
1356 bitmap_or(pml4->used_pml4es, new_pdps, pml4->used_pml4es,
1357 GEN8_PML4ES_PER_PML4);
1362 for_each_set_bit(pml4e, new_pdps, GEN8_PML4ES_PER_PML4)
1363 gen8_ppgtt_cleanup_3lvl(vm->dev, pml4->pdps[pml4e]);
1368 static int gen8_alloc_va_range(struct i915_address_space *vm,
1369 uint64_t start, uint64_t length)
1371 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1373 if (USES_FULL_48BIT_PPGTT(vm->dev))
1374 return gen8_alloc_va_range_4lvl(vm, &ppgtt->pml4, start, length);
1376 return gen8_alloc_va_range_3lvl(vm, &ppgtt->pdp, start, length);
1379 static void gen8_dump_pdp(struct i915_page_directory_pointer *pdp,
1380 uint64_t start, uint64_t length,
1381 gen8_pte_t scratch_pte,
1384 struct i915_page_directory *pd;
1387 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1388 struct i915_page_table *pt;
1389 uint64_t pd_len = length;
1390 uint64_t pd_start = start;
1393 if (!test_bit(pdpe, pdp->used_pdpes))
1396 seq_printf(m, "\tPDPE #%d\n", pdpe);
1397 gen8_for_each_pde(pt, pd, pd_start, pd_len, pde) {
1399 gen8_pte_t *pt_vaddr;
1401 if (!test_bit(pde, pd->used_pdes))
1404 pt_vaddr = kmap_px(pt);
1405 for (pte = 0; pte < GEN8_PTES; pte += 4) {
1407 (pdpe << GEN8_PDPE_SHIFT) |
1408 (pde << GEN8_PDE_SHIFT) |
1409 (pte << GEN8_PTE_SHIFT);
1413 for (i = 0; i < 4; i++)
1414 if (pt_vaddr[pte + i] != scratch_pte)
1419 seq_printf(m, "\t\t0x%llx [%03d,%03d,%04d]: =", va, pdpe, pde, pte);
1420 for (i = 0; i < 4; i++) {
1421 if (pt_vaddr[pte + i] != scratch_pte)
1422 seq_printf(m, " %llx", pt_vaddr[pte + i]);
1424 seq_puts(m, " SCRATCH ");
1428 /* don't use kunmap_px, it could trigger
1429 * an unnecessary flush.
1431 kunmap_atomic(pt_vaddr);
1436 static void gen8_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
1438 struct i915_address_space *vm = &ppgtt->base;
1439 uint64_t start = ppgtt->base.start;
1440 uint64_t length = ppgtt->base.total;
1441 gen8_pte_t scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
1442 I915_CACHE_LLC, true);
1444 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
1445 gen8_dump_pdp(&ppgtt->pdp, start, length, scratch_pte, m);
1448 struct i915_pml4 *pml4 = &ppgtt->pml4;
1449 struct i915_page_directory_pointer *pdp;
1451 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1452 if (!test_bit(pml4e, pml4->used_pml4es))
1455 seq_printf(m, " PML4E #%llu\n", pml4e);
1456 gen8_dump_pdp(pdp, start, length, scratch_pte, m);
1461 static int gen8_preallocate_top_level_pdps(struct i915_hw_ppgtt *ppgtt)
1463 unsigned long *new_page_dirs, *new_page_tables;
1464 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1467 /* We allocate temp bitmap for page tables for no gain
1468 * but as this is for init only, lets keep the things simple
1470 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables, pdpes);
1474 /* Allocate for all pdps regardless of how the ppgtt
1477 ret = gen8_ppgtt_alloc_page_directories(&ppgtt->base, &ppgtt->pdp,
1481 *ppgtt->pdp.used_pdpes = *new_page_dirs;
1483 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1489 * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers
1490 * with a net effect resembling a 2-level page table in normal x86 terms. Each
1491 * PDP represents 1GB of memory 4 * 512 * 512 * 4096 = 4GB legacy 32b address
1495 static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
1499 ret = gen8_init_scratch(&ppgtt->base);
1503 ppgtt->base.start = 0;
1504 ppgtt->base.cleanup = gen8_ppgtt_cleanup;
1505 ppgtt->base.allocate_va_range = gen8_alloc_va_range;
1506 ppgtt->base.insert_entries = gen8_ppgtt_insert_entries;
1507 ppgtt->base.clear_range = gen8_ppgtt_clear_range;
1508 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
1509 ppgtt->base.bind_vma = ppgtt_bind_vma;
1510 ppgtt->debug_dump = gen8_dump_ppgtt;
1512 if (USES_FULL_48BIT_PPGTT(ppgtt->base.dev)) {
1513 ret = setup_px(ppgtt->base.dev, &ppgtt->pml4);
1517 gen8_initialize_pml4(&ppgtt->base, &ppgtt->pml4);
1519 ppgtt->base.total = 1ULL << 48;
1520 ppgtt->switch_mm = gen8_48b_mm_switch;
1522 ret = __pdp_init(ppgtt->base.dev, &ppgtt->pdp);
1526 ppgtt->base.total = 1ULL << 32;
1527 ppgtt->switch_mm = gen8_legacy_mm_switch;
1528 trace_i915_page_directory_pointer_entry_alloc(&ppgtt->base,
1532 if (intel_vgpu_active(ppgtt->base.dev)) {
1533 ret = gen8_preallocate_top_level_pdps(ppgtt);
1539 if (intel_vgpu_active(ppgtt->base.dev))
1540 gen8_ppgtt_notify_vgt(ppgtt, true);
1545 gen8_free_scratch(&ppgtt->base);
1549 static void gen6_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
1551 struct i915_address_space *vm = &ppgtt->base;
1552 struct i915_page_table *unused;
1553 gen6_pte_t scratch_pte;
1555 uint32_t pte, pde, temp;
1556 uint32_t start = ppgtt->base.start, length = ppgtt->base.total;
1558 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1559 I915_CACHE_LLC, true, 0);
1561 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde) {
1563 gen6_pte_t *pt_vaddr;
1564 const dma_addr_t pt_addr = px_dma(ppgtt->pd.page_table[pde]);
1565 pd_entry = readl(ppgtt->pd_addr + pde);
1566 expected = (GEN6_PDE_ADDR_ENCODE(pt_addr) | GEN6_PDE_VALID);
1568 if (pd_entry != expected)
1569 seq_printf(m, "\tPDE #%d mismatch: Actual PDE: %x Expected PDE: %x\n",
1573 seq_printf(m, "\tPDE: %x\n", pd_entry);
1575 pt_vaddr = kmap_px(ppgtt->pd.page_table[pde]);
1577 for (pte = 0; pte < GEN6_PTES; pte+=4) {
1579 (pde * PAGE_SIZE * GEN6_PTES) +
1583 for (i = 0; i < 4; i++)
1584 if (pt_vaddr[pte + i] != scratch_pte)
1589 seq_printf(m, "\t\t0x%lx [%03d,%04d]: =", va, pde, pte);
1590 for (i = 0; i < 4; i++) {
1591 if (pt_vaddr[pte + i] != scratch_pte)
1592 seq_printf(m, " %08x", pt_vaddr[pte + i]);
1594 seq_puts(m, " SCRATCH ");
1598 kunmap_px(ppgtt, pt_vaddr);
1602 /* Write pde (index) from the page directory @pd to the page table @pt */
1603 static void gen6_write_pde(struct i915_page_directory *pd,
1604 const int pde, struct i915_page_table *pt)
1606 /* Caller needs to make sure the write completes if necessary */
1607 struct i915_hw_ppgtt *ppgtt =
1608 container_of(pd, struct i915_hw_ppgtt, pd);
1611 pd_entry = GEN6_PDE_ADDR_ENCODE(px_dma(pt));
1612 pd_entry |= GEN6_PDE_VALID;
1614 writel(pd_entry, ppgtt->pd_addr + pde);
1617 /* Write all the page tables found in the ppgtt structure to incrementing page
1619 static void gen6_write_page_range(struct drm_i915_private *dev_priv,
1620 struct i915_page_directory *pd,
1621 uint32_t start, uint32_t length)
1623 struct i915_ggtt *ggtt = &dev_priv->ggtt;
1624 struct i915_page_table *pt;
1627 gen6_for_each_pde(pt, pd, start, length, temp, pde)
1628 gen6_write_pde(pd, pde, pt);
1630 /* Make sure write is complete before other code can use this page
1631 * table. Also require for WC mapped PTEs */
1635 static uint32_t get_pd_offset(struct i915_hw_ppgtt *ppgtt)
1637 BUG_ON(ppgtt->pd.base.ggtt_offset & 0x3f);
1639 return (ppgtt->pd.base.ggtt_offset / 64) << 16;
1642 static int hsw_mm_switch(struct i915_hw_ppgtt *ppgtt,
1643 struct drm_i915_gem_request *req)
1645 struct intel_engine_cs *engine = req->engine;
1648 /* NB: TLBs must be flushed and invalidated before a switch */
1649 ret = engine->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1653 ret = intel_ring_begin(req, 6);
1657 intel_ring_emit(engine, MI_LOAD_REGISTER_IMM(2));
1658 intel_ring_emit_reg(engine, RING_PP_DIR_DCLV(engine));
1659 intel_ring_emit(engine, PP_DIR_DCLV_2G);
1660 intel_ring_emit_reg(engine, RING_PP_DIR_BASE(engine));
1661 intel_ring_emit(engine, get_pd_offset(ppgtt));
1662 intel_ring_emit(engine, MI_NOOP);
1663 intel_ring_advance(engine);
1668 static int vgpu_mm_switch(struct i915_hw_ppgtt *ppgtt,
1669 struct drm_i915_gem_request *req)
1671 struct intel_engine_cs *engine = req->engine;
1672 struct drm_i915_private *dev_priv = to_i915(ppgtt->base.dev);
1674 I915_WRITE(RING_PP_DIR_DCLV(engine), PP_DIR_DCLV_2G);
1675 I915_WRITE(RING_PP_DIR_BASE(engine), get_pd_offset(ppgtt));
1679 static int gen7_mm_switch(struct i915_hw_ppgtt *ppgtt,
1680 struct drm_i915_gem_request *req)
1682 struct intel_engine_cs *engine = req->engine;
1685 /* NB: TLBs must be flushed and invalidated before a switch */
1686 ret = engine->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1690 ret = intel_ring_begin(req, 6);
1694 intel_ring_emit(engine, MI_LOAD_REGISTER_IMM(2));
1695 intel_ring_emit_reg(engine, RING_PP_DIR_DCLV(engine));
1696 intel_ring_emit(engine, PP_DIR_DCLV_2G);
1697 intel_ring_emit_reg(engine, RING_PP_DIR_BASE(engine));
1698 intel_ring_emit(engine, get_pd_offset(ppgtt));
1699 intel_ring_emit(engine, MI_NOOP);
1700 intel_ring_advance(engine);
1702 /* XXX: RCS is the only one to auto invalidate the TLBs? */
1703 if (engine->id != RCS) {
1704 ret = engine->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1712 static int gen6_mm_switch(struct i915_hw_ppgtt *ppgtt,
1713 struct drm_i915_gem_request *req)
1715 struct intel_engine_cs *engine = req->engine;
1716 struct drm_device *dev = ppgtt->base.dev;
1717 struct drm_i915_private *dev_priv = dev->dev_private;
1720 I915_WRITE(RING_PP_DIR_DCLV(engine), PP_DIR_DCLV_2G);
1721 I915_WRITE(RING_PP_DIR_BASE(engine), get_pd_offset(ppgtt));
1723 POSTING_READ(RING_PP_DIR_DCLV(engine));
1728 static void gen8_ppgtt_enable(struct drm_device *dev)
1730 struct drm_i915_private *dev_priv = dev->dev_private;
1731 struct intel_engine_cs *engine;
1733 for_each_engine(engine, dev_priv) {
1734 u32 four_level = USES_FULL_48BIT_PPGTT(dev) ? GEN8_GFX_PPGTT_48B : 0;
1735 I915_WRITE(RING_MODE_GEN7(engine),
1736 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE | four_level));
1740 static void gen7_ppgtt_enable(struct drm_device *dev)
1742 struct drm_i915_private *dev_priv = dev->dev_private;
1743 struct intel_engine_cs *engine;
1744 uint32_t ecochk, ecobits;
1746 ecobits = I915_READ(GAC_ECO_BITS);
1747 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B);
1749 ecochk = I915_READ(GAM_ECOCHK);
1750 if (IS_HASWELL(dev)) {
1751 ecochk |= ECOCHK_PPGTT_WB_HSW;
1753 ecochk |= ECOCHK_PPGTT_LLC_IVB;
1754 ecochk &= ~ECOCHK_PPGTT_GFDT_IVB;
1756 I915_WRITE(GAM_ECOCHK, ecochk);
1758 for_each_engine(engine, dev_priv) {
1759 /* GFX_MODE is per-ring on gen7+ */
1760 I915_WRITE(RING_MODE_GEN7(engine),
1761 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1765 static void gen6_ppgtt_enable(struct drm_device *dev)
1767 struct drm_i915_private *dev_priv = dev->dev_private;
1768 uint32_t ecochk, gab_ctl, ecobits;
1770 ecobits = I915_READ(GAC_ECO_BITS);
1771 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_SNB_BIT |
1772 ECOBITS_PPGTT_CACHE64B);
1774 gab_ctl = I915_READ(GAB_CTL);
1775 I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT);
1777 ecochk = I915_READ(GAM_ECOCHK);
1778 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT | ECOCHK_PPGTT_CACHE64B);
1780 I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1783 /* PPGTT support for Sandybdrige/Gen6 and later */
1784 static void gen6_ppgtt_clear_range(struct i915_address_space *vm,
1789 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1790 gen6_pte_t *pt_vaddr, scratch_pte;
1791 unsigned first_entry = start >> PAGE_SHIFT;
1792 unsigned num_entries = length >> PAGE_SHIFT;
1793 unsigned act_pt = first_entry / GEN6_PTES;
1794 unsigned first_pte = first_entry % GEN6_PTES;
1795 unsigned last_pte, i;
1797 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1798 I915_CACHE_LLC, true, 0);
1800 while (num_entries) {
1801 last_pte = first_pte + num_entries;
1802 if (last_pte > GEN6_PTES)
1803 last_pte = GEN6_PTES;
1805 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1807 for (i = first_pte; i < last_pte; i++)
1808 pt_vaddr[i] = scratch_pte;
1810 kunmap_px(ppgtt, pt_vaddr);
1812 num_entries -= last_pte - first_pte;
1818 static void gen6_ppgtt_insert_entries(struct i915_address_space *vm,
1819 struct sg_table *pages,
1821 enum i915_cache_level cache_level, u32 flags)
1823 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1824 gen6_pte_t *pt_vaddr;
1825 unsigned first_entry = start >> PAGE_SHIFT;
1826 unsigned act_pt = first_entry / GEN6_PTES;
1827 unsigned act_pte = first_entry % GEN6_PTES;
1828 struct sg_page_iter sg_iter;
1831 for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) {
1832 if (pt_vaddr == NULL)
1833 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1836 vm->pte_encode(sg_page_iter_dma_address(&sg_iter),
1837 cache_level, true, flags);
1839 if (++act_pte == GEN6_PTES) {
1840 kunmap_px(ppgtt, pt_vaddr);
1847 kunmap_px(ppgtt, pt_vaddr);
1850 static int gen6_alloc_va_range(struct i915_address_space *vm,
1851 uint64_t start_in, uint64_t length_in)
1853 DECLARE_BITMAP(new_page_tables, I915_PDES);
1854 struct drm_device *dev = vm->dev;
1855 struct drm_i915_private *dev_priv = to_i915(dev);
1856 struct i915_ggtt *ggtt = &dev_priv->ggtt;
1857 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1858 struct i915_page_table *pt;
1859 uint32_t start, length, start_save, length_save;
1863 if (WARN_ON(start_in + length_in > ppgtt->base.total))
1866 start = start_save = start_in;
1867 length = length_save = length_in;
1869 bitmap_zero(new_page_tables, I915_PDES);
1871 /* The allocation is done in two stages so that we can bail out with
1872 * minimal amount of pain. The first stage finds new page tables that
1873 * need allocation. The second stage marks use ptes within the page
1876 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1877 if (pt != vm->scratch_pt) {
1878 WARN_ON(bitmap_empty(pt->used_ptes, GEN6_PTES));
1882 /* We've already allocated a page table */
1883 WARN_ON(!bitmap_empty(pt->used_ptes, GEN6_PTES));
1891 gen6_initialize_pt(vm, pt);
1893 ppgtt->pd.page_table[pde] = pt;
1894 __set_bit(pde, new_page_tables);
1895 trace_i915_page_table_entry_alloc(vm, pde, start, GEN6_PDE_SHIFT);
1899 length = length_save;
1901 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1902 DECLARE_BITMAP(tmp_bitmap, GEN6_PTES);
1904 bitmap_zero(tmp_bitmap, GEN6_PTES);
1905 bitmap_set(tmp_bitmap, gen6_pte_index(start),
1906 gen6_pte_count(start, length));
1908 if (__test_and_clear_bit(pde, new_page_tables))
1909 gen6_write_pde(&ppgtt->pd, pde, pt);
1911 trace_i915_page_table_entry_map(vm, pde, pt,
1912 gen6_pte_index(start),
1913 gen6_pte_count(start, length),
1915 bitmap_or(pt->used_ptes, tmp_bitmap, pt->used_ptes,
1919 WARN_ON(!bitmap_empty(new_page_tables, I915_PDES));
1921 /* Make sure write is complete before other code can use this page
1922 * table. Also require for WC mapped PTEs */
1925 mark_tlbs_dirty(ppgtt);
1929 for_each_set_bit(pde, new_page_tables, I915_PDES) {
1930 struct i915_page_table *pt = ppgtt->pd.page_table[pde];
1932 ppgtt->pd.page_table[pde] = vm->scratch_pt;
1933 free_pt(vm->dev, pt);
1936 mark_tlbs_dirty(ppgtt);
1940 static int gen6_init_scratch(struct i915_address_space *vm)
1942 struct drm_device *dev = vm->dev;
1944 vm->scratch_page = alloc_scratch_page(dev);
1945 if (IS_ERR(vm->scratch_page))
1946 return PTR_ERR(vm->scratch_page);
1948 vm->scratch_pt = alloc_pt(dev);
1949 if (IS_ERR(vm->scratch_pt)) {
1950 free_scratch_page(dev, vm->scratch_page);
1951 return PTR_ERR(vm->scratch_pt);
1954 gen6_initialize_pt(vm, vm->scratch_pt);
1959 static void gen6_free_scratch(struct i915_address_space *vm)
1961 struct drm_device *dev = vm->dev;
1963 free_pt(dev, vm->scratch_pt);
1964 free_scratch_page(dev, vm->scratch_page);
1967 static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
1969 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1970 struct i915_page_table *pt;
1973 drm_mm_remove_node(&ppgtt->node);
1975 gen6_for_all_pdes(pt, ppgtt, pde) {
1976 if (pt != vm->scratch_pt)
1977 free_pt(ppgtt->base.dev, pt);
1980 gen6_free_scratch(vm);
1983 static int gen6_ppgtt_allocate_page_directories(struct i915_hw_ppgtt *ppgtt)
1985 struct i915_address_space *vm = &ppgtt->base;
1986 struct drm_device *dev = ppgtt->base.dev;
1987 struct drm_i915_private *dev_priv = to_i915(dev);
1988 struct i915_ggtt *ggtt = &dev_priv->ggtt;
1989 bool retried = false;
1992 /* PPGTT PDEs reside in the GGTT and consists of 512 entries. The
1993 * allocator works in address space sizes, so it's multiplied by page
1994 * size. We allocate at the top of the GTT to avoid fragmentation.
1996 BUG_ON(!drm_mm_initialized(&ggtt->base.mm));
1998 ret = gen6_init_scratch(vm);
2003 ret = drm_mm_insert_node_in_range_generic(&ggtt->base.mm,
2004 &ppgtt->node, GEN6_PD_SIZE,
2006 0, ggtt->base.total,
2008 if (ret == -ENOSPC && !retried) {
2009 ret = i915_gem_evict_something(dev, &ggtt->base,
2010 GEN6_PD_SIZE, GEN6_PD_ALIGN,
2012 0, ggtt->base.total,
2025 if (ppgtt->node.start < ggtt->mappable_end)
2026 DRM_DEBUG("Forced to use aperture for PDEs\n");
2031 gen6_free_scratch(vm);
2035 static int gen6_ppgtt_alloc(struct i915_hw_ppgtt *ppgtt)
2037 return gen6_ppgtt_allocate_page_directories(ppgtt);
2040 static void gen6_scratch_va_range(struct i915_hw_ppgtt *ppgtt,
2041 uint64_t start, uint64_t length)
2043 struct i915_page_table *unused;
2046 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde)
2047 ppgtt->pd.page_table[pde] = ppgtt->base.scratch_pt;
2050 static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
2052 struct drm_device *dev = ppgtt->base.dev;
2053 struct drm_i915_private *dev_priv = to_i915(dev);
2054 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2057 ppgtt->base.pte_encode = ggtt->base.pte_encode;
2059 ppgtt->switch_mm = gen6_mm_switch;
2060 } else if (IS_HASWELL(dev)) {
2061 ppgtt->switch_mm = hsw_mm_switch;
2062 } else if (IS_GEN7(dev)) {
2063 ppgtt->switch_mm = gen7_mm_switch;
2067 if (intel_vgpu_active(dev))
2068 ppgtt->switch_mm = vgpu_mm_switch;
2070 ret = gen6_ppgtt_alloc(ppgtt);
2074 ppgtt->base.allocate_va_range = gen6_alloc_va_range;
2075 ppgtt->base.clear_range = gen6_ppgtt_clear_range;
2076 ppgtt->base.insert_entries = gen6_ppgtt_insert_entries;
2077 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
2078 ppgtt->base.bind_vma = ppgtt_bind_vma;
2079 ppgtt->base.cleanup = gen6_ppgtt_cleanup;
2080 ppgtt->base.start = 0;
2081 ppgtt->base.total = I915_PDES * GEN6_PTES * PAGE_SIZE;
2082 ppgtt->debug_dump = gen6_dump_ppgtt;
2084 ppgtt->pd.base.ggtt_offset =
2085 ppgtt->node.start / PAGE_SIZE * sizeof(gen6_pte_t);
2087 ppgtt->pd_addr = (gen6_pte_t __iomem *)ggtt->gsm +
2088 ppgtt->pd.base.ggtt_offset / sizeof(gen6_pte_t);
2090 gen6_scratch_va_range(ppgtt, 0, ppgtt->base.total);
2092 gen6_write_page_range(dev_priv, &ppgtt->pd, 0, ppgtt->base.total);
2094 DRM_DEBUG_DRIVER("Allocated pde space (%lldM) at GTT entry: %llx\n",
2095 ppgtt->node.size >> 20,
2096 ppgtt->node.start / PAGE_SIZE);
2098 DRM_DEBUG("Adding PPGTT at offset %x\n",
2099 ppgtt->pd.base.ggtt_offset << 10);
2104 static int __hw_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
2106 ppgtt->base.dev = dev;
2108 if (INTEL_INFO(dev)->gen < 8)
2109 return gen6_ppgtt_init(ppgtt);
2111 return gen8_ppgtt_init(ppgtt);
2114 static void i915_address_space_init(struct i915_address_space *vm,
2115 struct drm_i915_private *dev_priv)
2117 drm_mm_init(&vm->mm, vm->start, vm->total);
2118 vm->dev = dev_priv->dev;
2119 INIT_LIST_HEAD(&vm->active_list);
2120 INIT_LIST_HEAD(&vm->inactive_list);
2121 list_add_tail(&vm->global_link, &dev_priv->vm_list);
2124 static void gtt_write_workarounds(struct drm_device *dev)
2126 struct drm_i915_private *dev_priv = dev->dev_private;
2128 /* This function is for gtt related workarounds. This function is
2129 * called on driver load and after a GPU reset, so you can place
2130 * workarounds here even if they get overwritten by GPU reset.
2132 /* WaIncreaseDefaultTLBEntries:chv,bdw,skl,bxt */
2133 if (IS_BROADWELL(dev))
2134 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN8_L3_LRA_1_GPGPU_DEFAULT_VALUE_BDW);
2135 else if (IS_CHERRYVIEW(dev))
2136 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN8_L3_LRA_1_GPGPU_DEFAULT_VALUE_CHV);
2137 else if (IS_SKYLAKE(dev))
2138 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN9_L3_LRA_1_GPGPU_DEFAULT_VALUE_SKL);
2139 else if (IS_BROXTON(dev))
2140 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN9_L3_LRA_1_GPGPU_DEFAULT_VALUE_BXT);
2143 int i915_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
2145 struct drm_i915_private *dev_priv = dev->dev_private;
2148 ret = __hw_ppgtt_init(dev, ppgtt);
2150 kref_init(&ppgtt->ref);
2151 i915_address_space_init(&ppgtt->base, dev_priv);
2157 int i915_ppgtt_init_hw(struct drm_device *dev)
2159 gtt_write_workarounds(dev);
2161 /* In the case of execlists, PPGTT is enabled by the context descriptor
2162 * and the PDPs are contained within the context itself. We don't
2163 * need to do anything here. */
2164 if (i915.enable_execlists)
2167 if (!USES_PPGTT(dev))
2171 gen6_ppgtt_enable(dev);
2172 else if (IS_GEN7(dev))
2173 gen7_ppgtt_enable(dev);
2174 else if (INTEL_INFO(dev)->gen >= 8)
2175 gen8_ppgtt_enable(dev);
2177 MISSING_CASE(INTEL_INFO(dev)->gen);
2182 int i915_ppgtt_init_ring(struct drm_i915_gem_request *req)
2184 struct drm_i915_private *dev_priv = req->i915;
2185 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
2187 if (i915.enable_execlists)
2193 return ppgtt->switch_mm(ppgtt, req);
2196 struct i915_hw_ppgtt *
2197 i915_ppgtt_create(struct drm_device *dev, struct drm_i915_file_private *fpriv)
2199 struct i915_hw_ppgtt *ppgtt;
2202 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2204 return ERR_PTR(-ENOMEM);
2206 ret = i915_ppgtt_init(dev, ppgtt);
2209 return ERR_PTR(ret);
2212 ppgtt->file_priv = fpriv;
2214 trace_i915_ppgtt_create(&ppgtt->base);
2219 void i915_ppgtt_release(struct kref *kref)
2221 struct i915_hw_ppgtt *ppgtt =
2222 container_of(kref, struct i915_hw_ppgtt, ref);
2224 trace_i915_ppgtt_release(&ppgtt->base);
2226 /* vmas should already be unbound */
2227 WARN_ON(!list_empty(&ppgtt->base.active_list));
2228 WARN_ON(!list_empty(&ppgtt->base.inactive_list));
2230 list_del(&ppgtt->base.global_link);
2231 drm_mm_takedown(&ppgtt->base.mm);
2233 ppgtt->base.cleanup(&ppgtt->base);
2237 extern int intel_iommu_gfx_mapped;
2238 /* Certain Gen5 chipsets require require idling the GPU before
2239 * unmapping anything from the GTT when VT-d is enabled.
2241 static bool needs_idle_maps(struct drm_device *dev)
2243 #ifdef CONFIG_INTEL_IOMMU
2244 /* Query intel_iommu to see if we need the workaround. Presumably that
2247 if (IS_GEN5(dev) && IS_MOBILE(dev) && intel_iommu_gfx_mapped)
2253 static bool do_idling(struct drm_i915_private *dev_priv)
2255 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2256 bool ret = dev_priv->mm.interruptible;
2258 if (unlikely(ggtt->do_idle_maps)) {
2259 dev_priv->mm.interruptible = false;
2260 if (i915_gpu_idle(dev_priv->dev)) {
2261 DRM_ERROR("Couldn't idle GPU\n");
2262 /* Wait a bit, in hopes it avoids the hang */
2270 static void undo_idling(struct drm_i915_private *dev_priv, bool interruptible)
2272 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2274 if (unlikely(ggtt->do_idle_maps))
2275 dev_priv->mm.interruptible = interruptible;
2278 void i915_check_and_clear_faults(struct drm_device *dev)
2280 struct drm_i915_private *dev_priv = dev->dev_private;
2281 struct intel_engine_cs *engine;
2283 if (INTEL_INFO(dev)->gen < 6)
2286 for_each_engine(engine, dev_priv) {
2288 fault_reg = I915_READ(RING_FAULT_REG(engine));
2289 if (fault_reg & RING_FAULT_VALID) {
2290 DRM_DEBUG_DRIVER("Unexpected fault\n"
2292 "\tAddress space: %s\n"
2295 fault_reg & PAGE_MASK,
2296 fault_reg & RING_FAULT_GTTSEL_MASK ? "GGTT" : "PPGTT",
2297 RING_FAULT_SRCID(fault_reg),
2298 RING_FAULT_FAULT_TYPE(fault_reg));
2299 I915_WRITE(RING_FAULT_REG(engine),
2300 fault_reg & ~RING_FAULT_VALID);
2303 POSTING_READ(RING_FAULT_REG(&dev_priv->engine[RCS]));
2306 static void i915_ggtt_flush(struct drm_i915_private *dev_priv)
2308 if (INTEL_INFO(dev_priv)->gen < 6) {
2309 intel_gtt_chipset_flush();
2311 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2312 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2316 void i915_gem_suspend_gtt_mappings(struct drm_device *dev)
2318 struct drm_i915_private *dev_priv = to_i915(dev);
2319 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2321 /* Don't bother messing with faults pre GEN6 as we have little
2322 * documentation supporting that it's a good idea.
2324 if (INTEL_INFO(dev)->gen < 6)
2327 i915_check_and_clear_faults(dev);
2329 ggtt->base.clear_range(&ggtt->base, ggtt->base.start, ggtt->base.total,
2332 i915_ggtt_flush(dev_priv);
2335 int i915_gem_gtt_prepare_object(struct drm_i915_gem_object *obj)
2337 if (!dma_map_sg(&obj->base.dev->pdev->dev,
2338 obj->pages->sgl, obj->pages->nents,
2339 PCI_DMA_BIDIRECTIONAL))
2345 static void gen8_set_pte(void __iomem *addr, gen8_pte_t pte)
2350 iowrite32((u32)pte, addr);
2351 iowrite32(pte >> 32, addr + 4);
2355 static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
2356 struct sg_table *st,
2358 enum i915_cache_level level, u32 unused)
2360 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2361 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2362 unsigned first_entry = start >> PAGE_SHIFT;
2363 gen8_pte_t __iomem *gtt_entries =
2364 (gen8_pte_t __iomem *)ggtt->gsm + first_entry;
2366 struct sg_page_iter sg_iter;
2367 dma_addr_t addr = 0; /* shut up gcc */
2370 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2372 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
2373 addr = sg_dma_address(sg_iter.sg) +
2374 (sg_iter.sg_pgoffset << PAGE_SHIFT);
2375 gen8_set_pte(>t_entries[i],
2376 gen8_pte_encode(addr, level, true));
2381 * XXX: This serves as a posting read to make sure that the PTE has
2382 * actually been updated. There is some concern that even though
2383 * registers and PTEs are within the same BAR that they are potentially
2384 * of NUMA access patterns. Therefore, even with the way we assume
2385 * hardware should work, we must keep this posting read for paranoia.
2388 WARN_ON(readq(>t_entries[i-1])
2389 != gen8_pte_encode(addr, level, true));
2391 /* This next bit makes the above posting read even more important. We
2392 * want to flush the TLBs only after we're certain all the PTE updates
2395 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2396 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2398 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2401 struct insert_entries {
2402 struct i915_address_space *vm;
2403 struct sg_table *st;
2405 enum i915_cache_level level;
2409 static int gen8_ggtt_insert_entries__cb(void *_arg)
2411 struct insert_entries *arg = _arg;
2412 gen8_ggtt_insert_entries(arg->vm, arg->st,
2413 arg->start, arg->level, arg->flags);
2417 static void gen8_ggtt_insert_entries__BKL(struct i915_address_space *vm,
2418 struct sg_table *st,
2420 enum i915_cache_level level,
2423 struct insert_entries arg = { vm, st, start, level, flags };
2424 stop_machine(gen8_ggtt_insert_entries__cb, &arg, NULL);
2428 * Binds an object into the global gtt with the specified cache level. The object
2429 * will be accessible to the GPU via commands whose operands reference offsets
2430 * within the global GTT as well as accessible by the GPU through the GMADR
2431 * mapped BAR (dev_priv->mm.gtt->gtt).
2433 static void gen6_ggtt_insert_entries(struct i915_address_space *vm,
2434 struct sg_table *st,
2436 enum i915_cache_level level, u32 flags)
2438 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2439 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2440 unsigned first_entry = start >> PAGE_SHIFT;
2441 gen6_pte_t __iomem *gtt_entries =
2442 (gen6_pte_t __iomem *)ggtt->gsm + first_entry;
2444 struct sg_page_iter sg_iter;
2445 dma_addr_t addr = 0;
2448 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2450 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
2451 addr = sg_page_iter_dma_address(&sg_iter);
2452 iowrite32(vm->pte_encode(addr, level, true, flags), >t_entries[i]);
2456 /* XXX: This serves as a posting read to make sure that the PTE has
2457 * actually been updated. There is some concern that even though
2458 * registers and PTEs are within the same BAR that they are potentially
2459 * of NUMA access patterns. Therefore, even with the way we assume
2460 * hardware should work, we must keep this posting read for paranoia.
2463 unsigned long gtt = readl(>t_entries[i-1]);
2464 WARN_ON(gtt != vm->pte_encode(addr, level, true, flags));
2467 /* This next bit makes the above posting read even more important. We
2468 * want to flush the TLBs only after we're certain all the PTE updates
2471 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2472 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2474 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2477 static void gen8_ggtt_clear_range(struct i915_address_space *vm,
2482 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2483 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2484 unsigned first_entry = start >> PAGE_SHIFT;
2485 unsigned num_entries = length >> PAGE_SHIFT;
2486 gen8_pte_t scratch_pte, __iomem *gtt_base =
2487 (gen8_pte_t __iomem *)ggtt->gsm + first_entry;
2488 const int max_entries = ggtt_total_entries(ggtt) - first_entry;
2492 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2494 if (WARN(num_entries > max_entries,
2495 "First entry = %d; Num entries = %d (max=%d)\n",
2496 first_entry, num_entries, max_entries))
2497 num_entries = max_entries;
2499 scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
2502 for (i = 0; i < num_entries; i++)
2503 gen8_set_pte(>t_base[i], scratch_pte);
2506 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2509 static void gen6_ggtt_clear_range(struct i915_address_space *vm,
2514 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2515 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2516 unsigned first_entry = start >> PAGE_SHIFT;
2517 unsigned num_entries = length >> PAGE_SHIFT;
2518 gen6_pte_t scratch_pte, __iomem *gtt_base =
2519 (gen6_pte_t __iomem *)ggtt->gsm + first_entry;
2520 const int max_entries = ggtt_total_entries(ggtt) - first_entry;
2524 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2526 if (WARN(num_entries > max_entries,
2527 "First entry = %d; Num entries = %d (max=%d)\n",
2528 first_entry, num_entries, max_entries))
2529 num_entries = max_entries;
2531 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
2532 I915_CACHE_LLC, use_scratch, 0);
2534 for (i = 0; i < num_entries; i++)
2535 iowrite32(scratch_pte, >t_base[i]);
2538 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2541 static void i915_ggtt_insert_entries(struct i915_address_space *vm,
2542 struct sg_table *pages,
2544 enum i915_cache_level cache_level, u32 unused)
2546 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2547 unsigned int flags = (cache_level == I915_CACHE_NONE) ?
2548 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
2551 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2553 intel_gtt_insert_sg_entries(pages, start >> PAGE_SHIFT, flags);
2555 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2559 static void i915_ggtt_clear_range(struct i915_address_space *vm,
2564 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2565 unsigned first_entry = start >> PAGE_SHIFT;
2566 unsigned num_entries = length >> PAGE_SHIFT;
2569 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2571 intel_gtt_clear_range(first_entry, num_entries);
2573 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2576 static int ggtt_bind_vma(struct i915_vma *vma,
2577 enum i915_cache_level cache_level,
2580 struct drm_i915_gem_object *obj = vma->obj;
2584 ret = i915_get_ggtt_vma_pages(vma);
2588 /* Currently applicable only to VLV */
2590 pte_flags |= PTE_READ_ONLY;
2592 vma->vm->insert_entries(vma->vm, vma->ggtt_view.pages,
2594 cache_level, pte_flags);
2597 * Without aliasing PPGTT there's no difference between
2598 * GLOBAL/LOCAL_BIND, it's all the same ptes. Hence unconditionally
2599 * upgrade to both bound if we bind either to avoid double-binding.
2601 vma->bound |= GLOBAL_BIND | LOCAL_BIND;
2606 static int aliasing_gtt_bind_vma(struct i915_vma *vma,
2607 enum i915_cache_level cache_level,
2613 ret = i915_get_ggtt_vma_pages(vma);
2617 /* Currently applicable only to VLV */
2619 if (vma->obj->gt_ro)
2620 pte_flags |= PTE_READ_ONLY;
2623 if (flags & GLOBAL_BIND) {
2624 vma->vm->insert_entries(vma->vm,
2625 vma->ggtt_view.pages,
2627 cache_level, pte_flags);
2630 if (flags & LOCAL_BIND) {
2631 struct i915_hw_ppgtt *appgtt =
2632 to_i915(vma->vm->dev)->mm.aliasing_ppgtt;
2633 appgtt->base.insert_entries(&appgtt->base,
2634 vma->ggtt_view.pages,
2636 cache_level, pte_flags);
2642 static void ggtt_unbind_vma(struct i915_vma *vma)
2644 struct drm_device *dev = vma->vm->dev;
2645 struct drm_i915_private *dev_priv = dev->dev_private;
2646 struct drm_i915_gem_object *obj = vma->obj;
2647 const uint64_t size = min_t(uint64_t,
2651 if (vma->bound & GLOBAL_BIND) {
2652 vma->vm->clear_range(vma->vm,
2658 if (dev_priv->mm.aliasing_ppgtt && vma->bound & LOCAL_BIND) {
2659 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
2661 appgtt->base.clear_range(&appgtt->base,
2668 void i915_gem_gtt_finish_object(struct drm_i915_gem_object *obj)
2670 struct drm_device *dev = obj->base.dev;
2671 struct drm_i915_private *dev_priv = dev->dev_private;
2674 interruptible = do_idling(dev_priv);
2676 dma_unmap_sg(&dev->pdev->dev, obj->pages->sgl, obj->pages->nents,
2677 PCI_DMA_BIDIRECTIONAL);
2679 undo_idling(dev_priv, interruptible);
2682 static void i915_gtt_color_adjust(struct drm_mm_node *node,
2683 unsigned long color,
2687 if (node->color != color)
2690 if (!list_empty(&node->node_list)) {
2691 node = list_entry(node->node_list.next,
2694 if (node->allocated && node->color != color)
2699 static int i915_gem_setup_global_gtt(struct drm_device *dev,
2704 /* Let GEM Manage all of the aperture.
2706 * However, leave one page at the end still bound to the scratch page.
2707 * There are a number of places where the hardware apparently prefetches
2708 * past the end of the object, and we've seen multiple hangs with the
2709 * GPU head pointer stuck in a batchbuffer bound at the last page of the
2710 * aperture. One page should be enough to keep any prefetching inside
2713 struct drm_i915_private *dev_priv = to_i915(dev);
2714 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2715 struct drm_mm_node *entry;
2716 struct drm_i915_gem_object *obj;
2717 unsigned long hole_start, hole_end;
2720 BUG_ON(mappable_end > end);
2722 ggtt->base.start = start;
2724 /* Subtract the guard page before address space initialization to
2725 * shrink the range used by drm_mm */
2726 ggtt->base.total = end - start - PAGE_SIZE;
2727 i915_address_space_init(&ggtt->base, dev_priv);
2728 ggtt->base.total += PAGE_SIZE;
2730 if (intel_vgpu_active(dev)) {
2731 ret = intel_vgt_balloon(dev);
2737 ggtt->base.mm.color_adjust = i915_gtt_color_adjust;
2739 /* Mark any preallocated objects as occupied */
2740 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
2741 struct i915_vma *vma = i915_gem_obj_to_vma(obj, &ggtt->base);
2743 DRM_DEBUG_KMS("reserving preallocated space: %llx + %zx\n",
2744 i915_gem_obj_ggtt_offset(obj), obj->base.size);
2746 WARN_ON(i915_gem_obj_ggtt_bound(obj));
2747 ret = drm_mm_reserve_node(&ggtt->base.mm, &vma->node);
2749 DRM_DEBUG_KMS("Reservation failed: %i\n", ret);
2752 vma->bound |= GLOBAL_BIND;
2753 __i915_vma_set_map_and_fenceable(vma);
2754 list_add_tail(&vma->vm_link, &ggtt->base.inactive_list);
2757 /* Clear any non-preallocated blocks */
2758 drm_mm_for_each_hole(entry, &ggtt->base.mm, hole_start, hole_end) {
2759 DRM_DEBUG_KMS("clearing unused GTT space: [%lx, %lx]\n",
2760 hole_start, hole_end);
2761 ggtt->base.clear_range(&ggtt->base, hole_start,
2762 hole_end - hole_start, true);
2765 /* And finally clear the reserved guard page */
2766 ggtt->base.clear_range(&ggtt->base, end - PAGE_SIZE, PAGE_SIZE, true);
2768 if (USES_PPGTT(dev) && !USES_FULL_PPGTT(dev)) {
2769 struct i915_hw_ppgtt *ppgtt;
2771 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2775 ret = __hw_ppgtt_init(dev, ppgtt);
2777 ppgtt->base.cleanup(&ppgtt->base);
2782 if (ppgtt->base.allocate_va_range)
2783 ret = ppgtt->base.allocate_va_range(&ppgtt->base, 0,
2786 ppgtt->base.cleanup(&ppgtt->base);
2791 ppgtt->base.clear_range(&ppgtt->base,
2796 dev_priv->mm.aliasing_ppgtt = ppgtt;
2797 WARN_ON(ggtt->base.bind_vma != ggtt_bind_vma);
2798 ggtt->base.bind_vma = aliasing_gtt_bind_vma;
2805 * i915_gem_init_ggtt - Initialize GEM for Global GTT
2808 void i915_gem_init_ggtt(struct drm_device *dev)
2810 struct drm_i915_private *dev_priv = to_i915(dev);
2811 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2813 i915_gem_setup_global_gtt(dev, 0, ggtt->mappable_end, ggtt->base.total);
2817 * i915_ggtt_cleanup_hw - Clean up GGTT hardware initialization
2820 void i915_ggtt_cleanup_hw(struct drm_device *dev)
2822 struct drm_i915_private *dev_priv = to_i915(dev);
2823 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2825 if (dev_priv->mm.aliasing_ppgtt) {
2826 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
2828 ppgtt->base.cleanup(&ppgtt->base);
2831 i915_gem_cleanup_stolen(dev);
2833 if (drm_mm_initialized(&ggtt->base.mm)) {
2834 if (intel_vgpu_active(dev))
2835 intel_vgt_deballoon();
2837 drm_mm_takedown(&ggtt->base.mm);
2838 list_del(&ggtt->base.global_link);
2841 ggtt->base.cleanup(&ggtt->base);
2844 static unsigned int gen6_get_total_gtt_size(u16 snb_gmch_ctl)
2846 snb_gmch_ctl >>= SNB_GMCH_GGMS_SHIFT;
2847 snb_gmch_ctl &= SNB_GMCH_GGMS_MASK;
2848 return snb_gmch_ctl << 20;
2851 static unsigned int gen8_get_total_gtt_size(u16 bdw_gmch_ctl)
2853 bdw_gmch_ctl >>= BDW_GMCH_GGMS_SHIFT;
2854 bdw_gmch_ctl &= BDW_GMCH_GGMS_MASK;
2856 bdw_gmch_ctl = 1 << bdw_gmch_ctl;
2858 #ifdef CONFIG_X86_32
2859 /* Limit 32b platforms to a 2GB GGTT: 4 << 20 / pte size * PAGE_SIZE */
2860 if (bdw_gmch_ctl > 4)
2864 return bdw_gmch_ctl << 20;
2867 static unsigned int chv_get_total_gtt_size(u16 gmch_ctrl)
2869 gmch_ctrl >>= SNB_GMCH_GGMS_SHIFT;
2870 gmch_ctrl &= SNB_GMCH_GGMS_MASK;
2873 return 1 << (20 + gmch_ctrl);
2878 static size_t gen6_get_stolen_size(u16 snb_gmch_ctl)
2880 snb_gmch_ctl >>= SNB_GMCH_GMS_SHIFT;
2881 snb_gmch_ctl &= SNB_GMCH_GMS_MASK;
2882 return snb_gmch_ctl << 25; /* 32 MB units */
2885 static size_t gen8_get_stolen_size(u16 bdw_gmch_ctl)
2887 bdw_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2888 bdw_gmch_ctl &= BDW_GMCH_GMS_MASK;
2889 return bdw_gmch_ctl << 25; /* 32 MB units */
2892 static size_t chv_get_stolen_size(u16 gmch_ctrl)
2894 gmch_ctrl >>= SNB_GMCH_GMS_SHIFT;
2895 gmch_ctrl &= SNB_GMCH_GMS_MASK;
2898 * 0x0 to 0x10: 32MB increments starting at 0MB
2899 * 0x11 to 0x16: 4MB increments starting at 8MB
2900 * 0x17 to 0x1d: 4MB increments start at 36MB
2902 if (gmch_ctrl < 0x11)
2903 return gmch_ctrl << 25;
2904 else if (gmch_ctrl < 0x17)
2905 return (gmch_ctrl - 0x11 + 2) << 22;
2907 return (gmch_ctrl - 0x17 + 9) << 22;
2910 static size_t gen9_get_stolen_size(u16 gen9_gmch_ctl)
2912 gen9_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2913 gen9_gmch_ctl &= BDW_GMCH_GMS_MASK;
2915 if (gen9_gmch_ctl < 0xf0)
2916 return gen9_gmch_ctl << 25; /* 32 MB units */
2918 /* 4MB increments starting at 0xf0 for 4MB */
2919 return (gen9_gmch_ctl - 0xf0 + 1) << 22;
2922 static int ggtt_probe_common(struct drm_device *dev,
2925 struct drm_i915_private *dev_priv = to_i915(dev);
2926 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2927 struct i915_page_scratch *scratch_page;
2928 phys_addr_t ggtt_phys_addr;
2930 /* For Modern GENs the PTEs and register space are split in the BAR */
2931 ggtt_phys_addr = pci_resource_start(dev->pdev, 0) +
2932 (pci_resource_len(dev->pdev, 0) / 2);
2935 * On BXT writes larger than 64 bit to the GTT pagetable range will be
2936 * dropped. For WC mappings in general we have 64 byte burst writes
2937 * when the WC buffer is flushed, so we can't use it, but have to
2938 * resort to an uncached mapping. The WC issue is easily caught by the
2939 * readback check when writing GTT PTE entries.
2941 if (IS_BROXTON(dev))
2942 ggtt->gsm = ioremap_nocache(ggtt_phys_addr, gtt_size);
2944 ggtt->gsm = ioremap_wc(ggtt_phys_addr, gtt_size);
2946 DRM_ERROR("Failed to map the gtt page table\n");
2950 scratch_page = alloc_scratch_page(dev);
2951 if (IS_ERR(scratch_page)) {
2952 DRM_ERROR("Scratch setup failed\n");
2953 /* iounmap will also get called at remove, but meh */
2955 return PTR_ERR(scratch_page);
2958 ggtt->base.scratch_page = scratch_page;
2963 /* The GGTT and PPGTT need a private PPAT setup in order to handle cacheability
2964 * bits. When using advanced contexts each context stores its own PAT, but
2965 * writing this data shouldn't be harmful even in those cases. */
2966 static void bdw_setup_private_ppat(struct drm_i915_private *dev_priv)
2970 pat = GEN8_PPAT(0, GEN8_PPAT_WB | GEN8_PPAT_LLC) | /* for normal objects, no eLLC */
2971 GEN8_PPAT(1, GEN8_PPAT_WC | GEN8_PPAT_LLCELLC) | /* for something pointing to ptes? */
2972 GEN8_PPAT(2, GEN8_PPAT_WT | GEN8_PPAT_LLCELLC) | /* for scanout with eLLC */
2973 GEN8_PPAT(3, GEN8_PPAT_UC) | /* Uncached objects, mostly for scanout */
2974 GEN8_PPAT(4, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(0)) |
2975 GEN8_PPAT(5, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(1)) |
2976 GEN8_PPAT(6, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(2)) |
2977 GEN8_PPAT(7, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3));
2979 if (!USES_PPGTT(dev_priv))
2980 /* Spec: "For GGTT, there is NO pat_sel[2:0] from the entry,
2981 * so RTL will always use the value corresponding to
2983 * So let's disable cache for GGTT to avoid screen corruptions.
2984 * MOCS still can be used though.
2985 * - System agent ggtt writes (i.e. cpu gtt mmaps) already work
2986 * before this patch, i.e. the same uncached + snooping access
2987 * like on gen6/7 seems to be in effect.
2988 * - So this just fixes blitter/render access. Again it looks
2989 * like it's not just uncached access, but uncached + snooping.
2990 * So we can still hold onto all our assumptions wrt cpu
2991 * clflushing on LLC machines.
2993 pat = GEN8_PPAT(0, GEN8_PPAT_UC);
2995 /* XXX: spec defines this as 2 distinct registers. It's unclear if a 64b
2996 * write would work. */
2997 I915_WRITE(GEN8_PRIVATE_PAT_LO, pat);
2998 I915_WRITE(GEN8_PRIVATE_PAT_HI, pat >> 32);
3001 static void chv_setup_private_ppat(struct drm_i915_private *dev_priv)
3006 * Map WB on BDW to snooped on CHV.
3008 * Only the snoop bit has meaning for CHV, the rest is
3011 * The hardware will never snoop for certain types of accesses:
3012 * - CPU GTT (GMADR->GGTT->no snoop->memory)
3013 * - PPGTT page tables
3014 * - some other special cycles
3016 * As with BDW, we also need to consider the following for GT accesses:
3017 * "For GGTT, there is NO pat_sel[2:0] from the entry,
3018 * so RTL will always use the value corresponding to
3020 * Which means we must set the snoop bit in PAT entry 0
3021 * in order to keep the global status page working.
3023 pat = GEN8_PPAT(0, CHV_PPAT_SNOOP) |
3027 GEN8_PPAT(4, CHV_PPAT_SNOOP) |
3028 GEN8_PPAT(5, CHV_PPAT_SNOOP) |
3029 GEN8_PPAT(6, CHV_PPAT_SNOOP) |
3030 GEN8_PPAT(7, CHV_PPAT_SNOOP);
3032 I915_WRITE(GEN8_PRIVATE_PAT_LO, pat);
3033 I915_WRITE(GEN8_PRIVATE_PAT_HI, pat >> 32);
3036 static int gen8_gmch_probe(struct i915_ggtt *ggtt)
3038 struct drm_device *dev = ggtt->base.dev;
3039 struct drm_i915_private *dev_priv = to_i915(dev);
3043 /* TODO: We're not aware of mappable constraints on gen8 yet */
3044 ggtt->mappable_base = pci_resource_start(dev->pdev, 2);
3045 ggtt->mappable_end = pci_resource_len(dev->pdev, 2);
3047 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(39)))
3048 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(39));
3050 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
3052 if (INTEL_INFO(dev)->gen >= 9) {
3053 ggtt->stolen_size = gen9_get_stolen_size(snb_gmch_ctl);
3054 ggtt->size = gen8_get_total_gtt_size(snb_gmch_ctl);
3055 } else if (IS_CHERRYVIEW(dev)) {
3056 ggtt->stolen_size = chv_get_stolen_size(snb_gmch_ctl);
3057 ggtt->size = chv_get_total_gtt_size(snb_gmch_ctl);
3059 ggtt->stolen_size = gen8_get_stolen_size(snb_gmch_ctl);
3060 ggtt->size = gen8_get_total_gtt_size(snb_gmch_ctl);
3063 ggtt->base.total = (ggtt->size / sizeof(gen8_pte_t)) << PAGE_SHIFT;
3065 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
3066 chv_setup_private_ppat(dev_priv);
3068 bdw_setup_private_ppat(dev_priv);
3070 ret = ggtt_probe_common(dev, ggtt->size);
3072 ggtt->base.clear_range = gen8_ggtt_clear_range;
3073 if (IS_CHERRYVIEW(dev_priv))
3074 ggtt->base.insert_entries = gen8_ggtt_insert_entries__BKL;
3076 ggtt->base.insert_entries = gen8_ggtt_insert_entries;
3077 ggtt->base.bind_vma = ggtt_bind_vma;
3078 ggtt->base.unbind_vma = ggtt_unbind_vma;
3083 static int gen6_gmch_probe(struct i915_ggtt *ggtt)
3085 struct drm_device *dev = ggtt->base.dev;
3089 ggtt->mappable_base = pci_resource_start(dev->pdev, 2);
3090 ggtt->mappable_end = pci_resource_len(dev->pdev, 2);
3092 /* 64/512MB is the current min/max we actually know of, but this is just
3093 * a coarse sanity check.
3095 if ((ggtt->mappable_end < (64<<20) || (ggtt->mappable_end > (512<<20)))) {
3096 DRM_ERROR("Unknown GMADR size (%llx)\n", ggtt->mappable_end);
3100 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(40)))
3101 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(40));
3102 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
3104 ggtt->stolen_size = gen6_get_stolen_size(snb_gmch_ctl);
3105 ggtt->size = gen6_get_total_gtt_size(snb_gmch_ctl);
3106 ggtt->base.total = (ggtt->size / sizeof(gen6_pte_t)) << PAGE_SHIFT;
3108 ret = ggtt_probe_common(dev, ggtt->size);
3110 ggtt->base.clear_range = gen6_ggtt_clear_range;
3111 ggtt->base.insert_entries = gen6_ggtt_insert_entries;
3112 ggtt->base.bind_vma = ggtt_bind_vma;
3113 ggtt->base.unbind_vma = ggtt_unbind_vma;
3118 static void gen6_gmch_remove(struct i915_address_space *vm)
3120 struct i915_ggtt *ggtt = container_of(vm, struct i915_ggtt, base);
3123 free_scratch_page(vm->dev, vm->scratch_page);
3126 static int i915_gmch_probe(struct i915_ggtt *ggtt)
3128 struct drm_device *dev = ggtt->base.dev;
3129 struct drm_i915_private *dev_priv = to_i915(dev);
3132 ret = intel_gmch_probe(dev_priv->bridge_dev, dev_priv->dev->pdev, NULL);
3134 DRM_ERROR("failed to set up gmch\n");
3138 intel_gtt_get(&ggtt->base.total, &ggtt->stolen_size,
3139 &ggtt->mappable_base, &ggtt->mappable_end);
3141 ggtt->do_idle_maps = needs_idle_maps(dev_priv->dev);
3142 ggtt->base.insert_entries = i915_ggtt_insert_entries;
3143 ggtt->base.clear_range = i915_ggtt_clear_range;
3144 ggtt->base.bind_vma = ggtt_bind_vma;
3145 ggtt->base.unbind_vma = ggtt_unbind_vma;
3147 if (unlikely(ggtt->do_idle_maps))
3148 DRM_INFO("applying Ironlake quirks for intel_iommu\n");
3153 static void i915_gmch_remove(struct i915_address_space *vm)
3155 intel_gmch_remove();
3159 * i915_ggtt_init_hw - Initialize GGTT hardware
3162 int i915_ggtt_init_hw(struct drm_device *dev)
3164 struct drm_i915_private *dev_priv = to_i915(dev);
3165 struct i915_ggtt *ggtt = &dev_priv->ggtt;
3168 if (INTEL_INFO(dev)->gen <= 5) {
3169 ggtt->probe = i915_gmch_probe;
3170 ggtt->base.cleanup = i915_gmch_remove;
3171 } else if (INTEL_INFO(dev)->gen < 8) {
3172 ggtt->probe = gen6_gmch_probe;
3173 ggtt->base.cleanup = gen6_gmch_remove;
3176 ggtt->base.pte_encode = iris_pte_encode;
3177 else if (IS_HASWELL(dev))
3178 ggtt->base.pte_encode = hsw_pte_encode;
3179 else if (IS_VALLEYVIEW(dev))
3180 ggtt->base.pte_encode = byt_pte_encode;
3181 else if (INTEL_INFO(dev)->gen >= 7)
3182 ggtt->base.pte_encode = ivb_pte_encode;
3184 ggtt->base.pte_encode = snb_pte_encode;
3186 ggtt->probe = gen8_gmch_probe;
3187 ggtt->base.cleanup = gen6_gmch_remove;
3190 ggtt->base.dev = dev;
3191 ggtt->base.is_ggtt = true;
3193 ret = ggtt->probe(ggtt);
3197 if ((ggtt->base.total - 1) >> 32) {
3198 DRM_ERROR("We never expected a Global GTT with more than 32bits"
3199 "of address space! Found %lldM!\n",
3200 ggtt->base.total >> 20);
3201 ggtt->base.total = 1ULL << 32;
3202 ggtt->mappable_end = min(ggtt->mappable_end, ggtt->base.total);
3206 * Initialise stolen early so that we may reserve preallocated
3207 * objects for the BIOS to KMS transition.
3209 ret = i915_gem_init_stolen(dev);
3211 goto out_gtt_cleanup;
3213 /* GMADR is the PCI mmio aperture into the global GTT. */
3214 DRM_INFO("Memory usable by graphics device = %lluM\n",
3215 ggtt->base.total >> 20);
3216 DRM_DEBUG_DRIVER("GMADR size = %lldM\n", ggtt->mappable_end >> 20);
3217 DRM_DEBUG_DRIVER("GTT stolen size = %zdM\n", ggtt->stolen_size >> 20);
3218 #ifdef CONFIG_INTEL_IOMMU
3219 if (intel_iommu_gfx_mapped)
3220 DRM_INFO("VT-d active for gfx access\n");
3223 * i915.enable_ppgtt is read-only, so do an early pass to validate the
3224 * user's requested state against the hardware/driver capabilities. We
3225 * do this now so that we can print out any log messages once rather
3226 * than every time we check intel_enable_ppgtt().
3228 i915.enable_ppgtt = sanitize_enable_ppgtt(dev, i915.enable_ppgtt);
3229 DRM_DEBUG_DRIVER("ppgtt mode: %i\n", i915.enable_ppgtt);
3234 ggtt->base.cleanup(&ggtt->base);
3239 int i915_ggtt_enable_hw(struct drm_device *dev)
3241 if (INTEL_INFO(dev)->gen < 6 && !intel_enable_gtt())
3247 void i915_gem_restore_gtt_mappings(struct drm_device *dev)
3249 struct drm_i915_private *dev_priv = to_i915(dev);
3250 struct i915_ggtt *ggtt = &dev_priv->ggtt;
3251 struct drm_i915_gem_object *obj;
3252 struct i915_vma *vma;
3255 i915_check_and_clear_faults(dev);
3257 /* First fill our portion of the GTT with scratch pages */
3258 ggtt->base.clear_range(&ggtt->base, ggtt->base.start, ggtt->base.total,
3261 /* Cache flush objects bound into GGTT and rebind them. */
3262 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
3264 list_for_each_entry(vma, &obj->vma_list, obj_link) {
3265 if (vma->vm != &ggtt->base)
3268 WARN_ON(i915_vma_bind(vma, obj->cache_level,
3275 i915_gem_clflush_object(obj, obj->pin_display);
3278 if (INTEL_INFO(dev)->gen >= 8) {
3279 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
3280 chv_setup_private_ppat(dev_priv);
3282 bdw_setup_private_ppat(dev_priv);
3287 if (USES_PPGTT(dev)) {
3288 struct i915_address_space *vm;
3290 list_for_each_entry(vm, &dev_priv->vm_list, global_link) {
3291 /* TODO: Perhaps it shouldn't be gen6 specific */
3293 struct i915_hw_ppgtt *ppgtt;
3296 ppgtt = dev_priv->mm.aliasing_ppgtt;
3298 ppgtt = i915_vm_to_ppgtt(vm);
3300 gen6_write_page_range(dev_priv, &ppgtt->pd,
3301 0, ppgtt->base.total);
3305 i915_ggtt_flush(dev_priv);
3308 static struct i915_vma *
3309 __i915_gem_vma_create(struct drm_i915_gem_object *obj,
3310 struct i915_address_space *vm,
3311 const struct i915_ggtt_view *ggtt_view)
3313 struct i915_vma *vma;
3315 if (WARN_ON(i915_is_ggtt(vm) != !!ggtt_view))
3316 return ERR_PTR(-EINVAL);
3318 vma = kmem_cache_zalloc(to_i915(obj->base.dev)->vmas, GFP_KERNEL);
3320 return ERR_PTR(-ENOMEM);
3322 INIT_LIST_HEAD(&vma->vm_link);
3323 INIT_LIST_HEAD(&vma->obj_link);
3324 INIT_LIST_HEAD(&vma->exec_list);
3327 vma->is_ggtt = i915_is_ggtt(vm);
3329 if (i915_is_ggtt(vm))
3330 vma->ggtt_view = *ggtt_view;
3332 i915_ppgtt_get(i915_vm_to_ppgtt(vm));
3334 list_add_tail(&vma->obj_link, &obj->vma_list);
3340 i915_gem_obj_lookup_or_create_vma(struct drm_i915_gem_object *obj,
3341 struct i915_address_space *vm)
3343 struct i915_vma *vma;
3345 vma = i915_gem_obj_to_vma(obj, vm);
3347 vma = __i915_gem_vma_create(obj, vm,
3348 i915_is_ggtt(vm) ? &i915_ggtt_view_normal : NULL);
3354 i915_gem_obj_lookup_or_create_ggtt_vma(struct drm_i915_gem_object *obj,
3355 const struct i915_ggtt_view *view)
3357 struct drm_device *dev = obj->base.dev;
3358 struct drm_i915_private *dev_priv = to_i915(dev);
3359 struct i915_ggtt *ggtt = &dev_priv->ggtt;
3360 struct i915_vma *vma = i915_gem_obj_to_ggtt_view(obj, view);
3363 vma = __i915_gem_vma_create(obj, &ggtt->base, view);
3369 static struct scatterlist *
3370 rotate_pages(const dma_addr_t *in, unsigned int offset,
3371 unsigned int width, unsigned int height,
3372 unsigned int stride,
3373 struct sg_table *st, struct scatterlist *sg)
3375 unsigned int column, row;
3376 unsigned int src_idx;
3378 for (column = 0; column < width; column++) {
3379 src_idx = stride * (height - 1) + column;
3380 for (row = 0; row < height; row++) {
3382 /* We don't need the pages, but need to initialize
3383 * the entries so the sg list can be happily traversed.
3384 * The only thing we need are DMA addresses.
3386 sg_set_page(sg, NULL, PAGE_SIZE, 0);
3387 sg_dma_address(sg) = in[offset + src_idx];
3388 sg_dma_len(sg) = PAGE_SIZE;
3397 static struct sg_table *
3398 intel_rotate_fb_obj_pages(struct intel_rotation_info *rot_info,
3399 struct drm_i915_gem_object *obj)
3401 unsigned int size_pages = rot_info->plane[0].width * rot_info->plane[0].height;
3402 unsigned int size_pages_uv;
3403 struct sg_page_iter sg_iter;
3405 dma_addr_t *page_addr_list;
3406 struct sg_table *st;
3407 unsigned int uv_start_page;
3408 struct scatterlist *sg;
3411 /* Allocate a temporary list of source pages for random access. */
3412 page_addr_list = drm_malloc_gfp(obj->base.size / PAGE_SIZE,
3415 if (!page_addr_list)
3416 return ERR_PTR(ret);
3418 /* Account for UV plane with NV12. */
3419 if (rot_info->pixel_format == DRM_FORMAT_NV12)
3420 size_pages_uv = rot_info->plane[1].width * rot_info->plane[1].height;
3424 /* Allocate target SG list. */
3425 st = kmalloc(sizeof(*st), GFP_KERNEL);
3429 ret = sg_alloc_table(st, size_pages + size_pages_uv, GFP_KERNEL);
3433 /* Populate source page list from the object. */
3435 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) {
3436 page_addr_list[i] = sg_page_iter_dma_address(&sg_iter);
3443 /* Rotate the pages. */
3444 sg = rotate_pages(page_addr_list, 0,
3445 rot_info->plane[0].width, rot_info->plane[0].height,
3446 rot_info->plane[0].width,
3449 /* Append the UV plane if NV12. */
3450 if (rot_info->pixel_format == DRM_FORMAT_NV12) {
3451 uv_start_page = size_pages;
3453 /* Check for tile-row un-alignment. */
3454 if (offset_in_page(rot_info->uv_offset))
3457 rot_info->uv_start_page = uv_start_page;
3459 sg = rotate_pages(page_addr_list, rot_info->uv_start_page,
3460 rot_info->plane[1].width, rot_info->plane[1].height,
3461 rot_info->plane[1].width,
3465 DRM_DEBUG_KMS("Created rotated page mapping for object size %zu (%ux%u tiles, %u pages (%u plane 0)).\n",
3466 obj->base.size, rot_info->plane[0].width,
3467 rot_info->plane[0].height, size_pages + size_pages_uv,
3470 drm_free_large(page_addr_list);
3477 drm_free_large(page_addr_list);
3479 DRM_DEBUG_KMS("Failed to create rotated mapping for object size %zu! (%d) (%ux%u tiles, %u pages (%u plane 0))\n",
3480 obj->base.size, ret, rot_info->plane[0].width,
3481 rot_info->plane[0].height, size_pages + size_pages_uv,
3483 return ERR_PTR(ret);
3486 static struct sg_table *
3487 intel_partial_pages(const struct i915_ggtt_view *view,
3488 struct drm_i915_gem_object *obj)
3490 struct sg_table *st;
3491 struct scatterlist *sg;
3492 struct sg_page_iter obj_sg_iter;
3495 st = kmalloc(sizeof(*st), GFP_KERNEL);
3499 ret = sg_alloc_table(st, view->params.partial.size, GFP_KERNEL);
3505 for_each_sg_page(obj->pages->sgl, &obj_sg_iter, obj->pages->nents,
3506 view->params.partial.offset)
3508 if (st->nents >= view->params.partial.size)
3511 sg_set_page(sg, NULL, PAGE_SIZE, 0);
3512 sg_dma_address(sg) = sg_page_iter_dma_address(&obj_sg_iter);
3513 sg_dma_len(sg) = PAGE_SIZE;
3524 return ERR_PTR(ret);
3528 i915_get_ggtt_vma_pages(struct i915_vma *vma)
3532 if (vma->ggtt_view.pages)
3535 if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL)
3536 vma->ggtt_view.pages = vma->obj->pages;
3537 else if (vma->ggtt_view.type == I915_GGTT_VIEW_ROTATED)
3538 vma->ggtt_view.pages =
3539 intel_rotate_fb_obj_pages(&vma->ggtt_view.params.rotated, vma->obj);
3540 else if (vma->ggtt_view.type == I915_GGTT_VIEW_PARTIAL)
3541 vma->ggtt_view.pages =
3542 intel_partial_pages(&vma->ggtt_view, vma->obj);
3544 WARN_ONCE(1, "GGTT view %u not implemented!\n",
3545 vma->ggtt_view.type);
3547 if (!vma->ggtt_view.pages) {
3548 DRM_ERROR("Failed to get pages for GGTT view type %u!\n",
3549 vma->ggtt_view.type);
3551 } else if (IS_ERR(vma->ggtt_view.pages)) {
3552 ret = PTR_ERR(vma->ggtt_view.pages);
3553 vma->ggtt_view.pages = NULL;
3554 DRM_ERROR("Failed to get pages for VMA view type %u (%d)!\n",
3555 vma->ggtt_view.type, ret);
3562 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
3564 * @cache_level: mapping cache level
3565 * @flags: flags like global or local mapping
3567 * DMA addresses are taken from the scatter-gather table of this object (or of
3568 * this VMA in case of non-default GGTT views) and PTE entries set up.
3569 * Note that DMA addresses are also the only part of the SG table we care about.
3571 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
3577 if (WARN_ON(flags == 0))
3581 if (flags & PIN_GLOBAL)
3582 bind_flags |= GLOBAL_BIND;
3583 if (flags & PIN_USER)
3584 bind_flags |= LOCAL_BIND;
3586 if (flags & PIN_UPDATE)
3587 bind_flags |= vma->bound;
3589 bind_flags &= ~vma->bound;
3591 if (bind_flags == 0)
3594 if (vma->bound == 0 && vma->vm->allocate_va_range) {
3595 /* XXX: i915_vma_pin() will fix this +- hack */
3597 trace_i915_va_alloc(vma);
3598 ret = vma->vm->allocate_va_range(vma->vm,
3606 ret = vma->vm->bind_vma(vma, cache_level, bind_flags);
3610 vma->bound |= bind_flags;
3616 * i915_ggtt_view_size - Get the size of a GGTT view.
3617 * @obj: Object the view is of.
3618 * @view: The view in question.
3620 * @return The size of the GGTT view in bytes.
3623 i915_ggtt_view_size(struct drm_i915_gem_object *obj,
3624 const struct i915_ggtt_view *view)
3626 if (view->type == I915_GGTT_VIEW_NORMAL) {
3627 return obj->base.size;
3628 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
3629 return intel_rotation_info_size(&view->params.rotated) << PAGE_SHIFT;
3630 } else if (view->type == I915_GGTT_VIEW_PARTIAL) {
3631 return view->params.partial.size << PAGE_SHIFT;
3633 WARN_ONCE(1, "GGTT view %u not implemented!\n", view->type);
3634 return obj->base.size;