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_device *dev = ppgtt->base.dev;
909 struct drm_i915_private *dev_priv = dev->dev_private;
912 if (USES_FULL_48BIT_PPGTT(dev)) {
913 u64 daddr = px_dma(&ppgtt->pml4);
915 I915_WRITE(vgtif_reg(pdp[0].lo), lower_32_bits(daddr));
916 I915_WRITE(vgtif_reg(pdp[0].hi), upper_32_bits(daddr));
918 msg = (create ? VGT_G2V_PPGTT_L4_PAGE_TABLE_CREATE :
919 VGT_G2V_PPGTT_L4_PAGE_TABLE_DESTROY);
921 for (i = 0; i < GEN8_LEGACY_PDPES; i++) {
922 u64 daddr = i915_page_dir_dma_addr(ppgtt, i);
924 I915_WRITE(vgtif_reg(pdp[i].lo), lower_32_bits(daddr));
925 I915_WRITE(vgtif_reg(pdp[i].hi), upper_32_bits(daddr));
928 msg = (create ? VGT_G2V_PPGTT_L3_PAGE_TABLE_CREATE :
929 VGT_G2V_PPGTT_L3_PAGE_TABLE_DESTROY);
932 I915_WRITE(vgtif_reg(g2v_notify), msg);
937 static void gen8_free_scratch(struct i915_address_space *vm)
939 struct drm_device *dev = vm->dev;
941 if (USES_FULL_48BIT_PPGTT(dev))
942 free_pdp(dev, vm->scratch_pdp);
943 free_pd(dev, vm->scratch_pd);
944 free_pt(dev, vm->scratch_pt);
945 free_scratch_page(dev, vm->scratch_page);
948 static void gen8_ppgtt_cleanup_3lvl(struct drm_device *dev,
949 struct i915_page_directory_pointer *pdp)
953 for_each_set_bit(i, pdp->used_pdpes, I915_PDPES_PER_PDP(dev)) {
954 if (WARN_ON(!pdp->page_directory[i]))
957 gen8_free_page_tables(dev, pdp->page_directory[i]);
958 free_pd(dev, pdp->page_directory[i]);
964 static void gen8_ppgtt_cleanup_4lvl(struct i915_hw_ppgtt *ppgtt)
968 for_each_set_bit(i, ppgtt->pml4.used_pml4es, GEN8_PML4ES_PER_PML4) {
969 if (WARN_ON(!ppgtt->pml4.pdps[i]))
972 gen8_ppgtt_cleanup_3lvl(ppgtt->base.dev, ppgtt->pml4.pdps[i]);
975 cleanup_px(ppgtt->base.dev, &ppgtt->pml4);
978 static void gen8_ppgtt_cleanup(struct i915_address_space *vm)
980 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
982 if (intel_vgpu_active(vm->dev))
983 gen8_ppgtt_notify_vgt(ppgtt, false);
985 if (!USES_FULL_48BIT_PPGTT(ppgtt->base.dev))
986 gen8_ppgtt_cleanup_3lvl(ppgtt->base.dev, &ppgtt->pdp);
988 gen8_ppgtt_cleanup_4lvl(ppgtt);
990 gen8_free_scratch(vm);
994 * gen8_ppgtt_alloc_pagetabs() - Allocate page tables for VA range.
995 * @vm: Master vm structure.
996 * @pd: Page directory for this address range.
997 * @start: Starting virtual address to begin allocations.
998 * @length: Size of the allocations.
999 * @new_pts: Bitmap set by function with new allocations. Likely used by the
1000 * caller to free on error.
1002 * Allocate the required number of page tables. Extremely similar to
1003 * gen8_ppgtt_alloc_page_directories(). The main difference is here we are limited by
1004 * the page directory boundary (instead of the page directory pointer). That
1005 * boundary is 1GB virtual. Therefore, unlike gen8_ppgtt_alloc_page_directories(), it is
1006 * possible, and likely that the caller will need to use multiple calls of this
1007 * function to achieve the appropriate allocation.
1009 * Return: 0 if success; negative error code otherwise.
1011 static int gen8_ppgtt_alloc_pagetabs(struct i915_address_space *vm,
1012 struct i915_page_directory *pd,
1015 unsigned long *new_pts)
1017 struct drm_device *dev = vm->dev;
1018 struct i915_page_table *pt;
1021 gen8_for_each_pde(pt, pd, start, length, pde) {
1022 /* Don't reallocate page tables */
1023 if (test_bit(pde, pd->used_pdes)) {
1024 /* Scratch is never allocated this way */
1025 WARN_ON(pt == vm->scratch_pt);
1033 gen8_initialize_pt(vm, pt);
1034 pd->page_table[pde] = pt;
1035 __set_bit(pde, new_pts);
1036 trace_i915_page_table_entry_alloc(vm, pde, start, GEN8_PDE_SHIFT);
1042 for_each_set_bit(pde, new_pts, I915_PDES)
1043 free_pt(dev, pd->page_table[pde]);
1049 * gen8_ppgtt_alloc_page_directories() - Allocate page directories for VA range.
1050 * @vm: Master vm structure.
1051 * @pdp: Page directory pointer for this address range.
1052 * @start: Starting virtual address to begin allocations.
1053 * @length: Size of the allocations.
1054 * @new_pds: Bitmap set by function with new allocations. Likely used by the
1055 * caller to free on error.
1057 * Allocate the required number of page directories starting at the pde index of
1058 * @start, and ending at the pde index @start + @length. This function will skip
1059 * over already allocated page directories within the range, and only allocate
1060 * new ones, setting the appropriate pointer within the pdp as well as the
1061 * correct position in the bitmap @new_pds.
1063 * The function will only allocate the pages within the range for a give page
1064 * directory pointer. In other words, if @start + @length straddles a virtually
1065 * addressed PDP boundary (512GB for 4k pages), there will be more allocations
1066 * required by the caller, This is not currently possible, and the BUG in the
1067 * code will prevent it.
1069 * Return: 0 if success; negative error code otherwise.
1072 gen8_ppgtt_alloc_page_directories(struct i915_address_space *vm,
1073 struct i915_page_directory_pointer *pdp,
1076 unsigned long *new_pds)
1078 struct drm_device *dev = vm->dev;
1079 struct i915_page_directory *pd;
1081 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1083 WARN_ON(!bitmap_empty(new_pds, pdpes));
1085 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1086 if (test_bit(pdpe, pdp->used_pdpes))
1093 gen8_initialize_pd(vm, pd);
1094 pdp->page_directory[pdpe] = pd;
1095 __set_bit(pdpe, new_pds);
1096 trace_i915_page_directory_entry_alloc(vm, pdpe, start, GEN8_PDPE_SHIFT);
1102 for_each_set_bit(pdpe, new_pds, pdpes)
1103 free_pd(dev, pdp->page_directory[pdpe]);
1109 * gen8_ppgtt_alloc_page_dirpointers() - Allocate pdps for VA range.
1110 * @vm: Master vm structure.
1111 * @pml4: Page map level 4 for this address range.
1112 * @start: Starting virtual address to begin allocations.
1113 * @length: Size of the allocations.
1114 * @new_pdps: Bitmap set by function with new allocations. Likely used by the
1115 * caller to free on error.
1117 * Allocate the required number of page directory pointers. Extremely similar to
1118 * gen8_ppgtt_alloc_page_directories() and gen8_ppgtt_alloc_pagetabs().
1119 * The main difference is here we are limited by the pml4 boundary (instead of
1120 * the page directory pointer).
1122 * Return: 0 if success; negative error code otherwise.
1125 gen8_ppgtt_alloc_page_dirpointers(struct i915_address_space *vm,
1126 struct i915_pml4 *pml4,
1129 unsigned long *new_pdps)
1131 struct drm_device *dev = vm->dev;
1132 struct i915_page_directory_pointer *pdp;
1135 WARN_ON(!bitmap_empty(new_pdps, GEN8_PML4ES_PER_PML4));
1137 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1138 if (!test_bit(pml4e, pml4->used_pml4es)) {
1139 pdp = alloc_pdp(dev);
1143 gen8_initialize_pdp(vm, pdp);
1144 pml4->pdps[pml4e] = pdp;
1145 __set_bit(pml4e, new_pdps);
1146 trace_i915_page_directory_pointer_entry_alloc(vm,
1156 for_each_set_bit(pml4e, new_pdps, GEN8_PML4ES_PER_PML4)
1157 free_pdp(dev, pml4->pdps[pml4e]);
1163 free_gen8_temp_bitmaps(unsigned long *new_pds, unsigned long *new_pts)
1169 /* Fills in the page directory bitmap, and the array of page tables bitmap. Both
1170 * of these are based on the number of PDPEs in the system.
1173 int __must_check alloc_gen8_temp_bitmaps(unsigned long **new_pds,
1174 unsigned long **new_pts,
1180 pds = kcalloc(BITS_TO_LONGS(pdpes), sizeof(unsigned long), GFP_TEMPORARY);
1184 pts = kcalloc(pdpes, BITS_TO_LONGS(I915_PDES) * sizeof(unsigned long),
1195 free_gen8_temp_bitmaps(pds, pts);
1199 /* PDE TLBs are a pain to invalidate on GEN8+. When we modify
1200 * the page table structures, we mark them dirty so that
1201 * context switching/execlist queuing code takes extra steps
1202 * to ensure that tlbs are flushed.
1204 static void mark_tlbs_dirty(struct i915_hw_ppgtt *ppgtt)
1206 ppgtt->pd_dirty_rings = INTEL_INFO(ppgtt->base.dev)->ring_mask;
1209 static int gen8_alloc_va_range_3lvl(struct i915_address_space *vm,
1210 struct i915_page_directory_pointer *pdp,
1214 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1215 unsigned long *new_page_dirs, *new_page_tables;
1216 struct drm_device *dev = vm->dev;
1217 struct i915_page_directory *pd;
1218 const uint64_t orig_start = start;
1219 const uint64_t orig_length = length;
1221 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1224 /* Wrap is never okay since we can only represent 48b, and we don't
1225 * actually use the other side of the canonical address space.
1227 if (WARN_ON(start + length < start))
1230 if (WARN_ON(start + length > vm->total))
1233 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables, pdpes);
1237 /* Do the allocations first so we can easily bail out */
1238 ret = gen8_ppgtt_alloc_page_directories(vm, pdp, start, length,
1241 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1245 /* For every page directory referenced, allocate page tables */
1246 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1247 ret = gen8_ppgtt_alloc_pagetabs(vm, pd, start, length,
1248 new_page_tables + pdpe * BITS_TO_LONGS(I915_PDES));
1254 length = orig_length;
1256 /* Allocations have completed successfully, so set the bitmaps, and do
1258 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1259 gen8_pde_t *const page_directory = kmap_px(pd);
1260 struct i915_page_table *pt;
1261 uint64_t pd_len = length;
1262 uint64_t pd_start = start;
1265 /* Every pd should be allocated, we just did that above. */
1268 gen8_for_each_pde(pt, pd, pd_start, pd_len, pde) {
1269 /* Same reasoning as pd */
1272 WARN_ON(!gen8_pte_count(pd_start, pd_len));
1274 /* Set our used ptes within the page table */
1275 bitmap_set(pt->used_ptes,
1276 gen8_pte_index(pd_start),
1277 gen8_pte_count(pd_start, pd_len));
1279 /* Our pde is now pointing to the pagetable, pt */
1280 __set_bit(pde, pd->used_pdes);
1282 /* Map the PDE to the page table */
1283 page_directory[pde] = gen8_pde_encode(px_dma(pt),
1285 trace_i915_page_table_entry_map(&ppgtt->base, pde, pt,
1286 gen8_pte_index(start),
1287 gen8_pte_count(start, length),
1290 /* NB: We haven't yet mapped ptes to pages. At this
1291 * point we're still relying on insert_entries() */
1294 kunmap_px(ppgtt, page_directory);
1295 __set_bit(pdpe, pdp->used_pdpes);
1296 gen8_setup_page_directory(ppgtt, pdp, pd, pdpe);
1299 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1300 mark_tlbs_dirty(ppgtt);
1307 for_each_set_bit(temp, new_page_tables + pdpe *
1308 BITS_TO_LONGS(I915_PDES), I915_PDES)
1309 free_pt(dev, pdp->page_directory[pdpe]->page_table[temp]);
1312 for_each_set_bit(pdpe, new_page_dirs, pdpes)
1313 free_pd(dev, pdp->page_directory[pdpe]);
1315 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1316 mark_tlbs_dirty(ppgtt);
1320 static int gen8_alloc_va_range_4lvl(struct i915_address_space *vm,
1321 struct i915_pml4 *pml4,
1325 DECLARE_BITMAP(new_pdps, GEN8_PML4ES_PER_PML4);
1326 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1327 struct i915_page_directory_pointer *pdp;
1331 /* Do the pml4 allocations first, so we don't need to track the newly
1332 * allocated tables below the pdp */
1333 bitmap_zero(new_pdps, GEN8_PML4ES_PER_PML4);
1335 /* The pagedirectory and pagetable allocations are done in the shared 3
1336 * and 4 level code. Just allocate the pdps.
1338 ret = gen8_ppgtt_alloc_page_dirpointers(vm, pml4, start, length,
1343 WARN(bitmap_weight(new_pdps, GEN8_PML4ES_PER_PML4) > 2,
1344 "The allocation has spanned more than 512GB. "
1345 "It is highly likely this is incorrect.");
1347 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1350 ret = gen8_alloc_va_range_3lvl(vm, pdp, start, length);
1354 gen8_setup_page_directory_pointer(ppgtt, pml4, pdp, pml4e);
1357 bitmap_or(pml4->used_pml4es, new_pdps, pml4->used_pml4es,
1358 GEN8_PML4ES_PER_PML4);
1363 for_each_set_bit(pml4e, new_pdps, GEN8_PML4ES_PER_PML4)
1364 gen8_ppgtt_cleanup_3lvl(vm->dev, pml4->pdps[pml4e]);
1369 static int gen8_alloc_va_range(struct i915_address_space *vm,
1370 uint64_t start, uint64_t length)
1372 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1374 if (USES_FULL_48BIT_PPGTT(vm->dev))
1375 return gen8_alloc_va_range_4lvl(vm, &ppgtt->pml4, start, length);
1377 return gen8_alloc_va_range_3lvl(vm, &ppgtt->pdp, start, length);
1380 static void gen8_dump_pdp(struct i915_page_directory_pointer *pdp,
1381 uint64_t start, uint64_t length,
1382 gen8_pte_t scratch_pte,
1385 struct i915_page_directory *pd;
1388 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1389 struct i915_page_table *pt;
1390 uint64_t pd_len = length;
1391 uint64_t pd_start = start;
1394 if (!test_bit(pdpe, pdp->used_pdpes))
1397 seq_printf(m, "\tPDPE #%d\n", pdpe);
1398 gen8_for_each_pde(pt, pd, pd_start, pd_len, pde) {
1400 gen8_pte_t *pt_vaddr;
1402 if (!test_bit(pde, pd->used_pdes))
1405 pt_vaddr = kmap_px(pt);
1406 for (pte = 0; pte < GEN8_PTES; pte += 4) {
1408 (pdpe << GEN8_PDPE_SHIFT) |
1409 (pde << GEN8_PDE_SHIFT) |
1410 (pte << GEN8_PTE_SHIFT);
1414 for (i = 0; i < 4; i++)
1415 if (pt_vaddr[pte + i] != scratch_pte)
1420 seq_printf(m, "\t\t0x%llx [%03d,%03d,%04d]: =", va, pdpe, pde, pte);
1421 for (i = 0; i < 4; i++) {
1422 if (pt_vaddr[pte + i] != scratch_pte)
1423 seq_printf(m, " %llx", pt_vaddr[pte + i]);
1425 seq_puts(m, " SCRATCH ");
1429 /* don't use kunmap_px, it could trigger
1430 * an unnecessary flush.
1432 kunmap_atomic(pt_vaddr);
1437 static void gen8_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
1439 struct i915_address_space *vm = &ppgtt->base;
1440 uint64_t start = ppgtt->base.start;
1441 uint64_t length = ppgtt->base.total;
1442 gen8_pte_t scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
1443 I915_CACHE_LLC, true);
1445 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
1446 gen8_dump_pdp(&ppgtt->pdp, start, length, scratch_pte, m);
1449 struct i915_pml4 *pml4 = &ppgtt->pml4;
1450 struct i915_page_directory_pointer *pdp;
1452 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1453 if (!test_bit(pml4e, pml4->used_pml4es))
1456 seq_printf(m, " PML4E #%llu\n", pml4e);
1457 gen8_dump_pdp(pdp, start, length, scratch_pte, m);
1462 static int gen8_preallocate_top_level_pdps(struct i915_hw_ppgtt *ppgtt)
1464 unsigned long *new_page_dirs, *new_page_tables;
1465 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1468 /* We allocate temp bitmap for page tables for no gain
1469 * but as this is for init only, lets keep the things simple
1471 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables, pdpes);
1475 /* Allocate for all pdps regardless of how the ppgtt
1478 ret = gen8_ppgtt_alloc_page_directories(&ppgtt->base, &ppgtt->pdp,
1482 *ppgtt->pdp.used_pdpes = *new_page_dirs;
1484 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1490 * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers
1491 * with a net effect resembling a 2-level page table in normal x86 terms. Each
1492 * PDP represents 1GB of memory 4 * 512 * 512 * 4096 = 4GB legacy 32b address
1496 static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
1500 ret = gen8_init_scratch(&ppgtt->base);
1504 ppgtt->base.start = 0;
1505 ppgtt->base.cleanup = gen8_ppgtt_cleanup;
1506 ppgtt->base.allocate_va_range = gen8_alloc_va_range;
1507 ppgtt->base.insert_entries = gen8_ppgtt_insert_entries;
1508 ppgtt->base.clear_range = gen8_ppgtt_clear_range;
1509 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
1510 ppgtt->base.bind_vma = ppgtt_bind_vma;
1511 ppgtt->debug_dump = gen8_dump_ppgtt;
1513 if (USES_FULL_48BIT_PPGTT(ppgtt->base.dev)) {
1514 ret = setup_px(ppgtt->base.dev, &ppgtt->pml4);
1518 gen8_initialize_pml4(&ppgtt->base, &ppgtt->pml4);
1520 ppgtt->base.total = 1ULL << 48;
1521 ppgtt->switch_mm = gen8_48b_mm_switch;
1523 ret = __pdp_init(ppgtt->base.dev, &ppgtt->pdp);
1527 ppgtt->base.total = 1ULL << 32;
1528 ppgtt->switch_mm = gen8_legacy_mm_switch;
1529 trace_i915_page_directory_pointer_entry_alloc(&ppgtt->base,
1533 if (intel_vgpu_active(ppgtt->base.dev)) {
1534 ret = gen8_preallocate_top_level_pdps(ppgtt);
1540 if (intel_vgpu_active(ppgtt->base.dev))
1541 gen8_ppgtt_notify_vgt(ppgtt, true);
1546 gen8_free_scratch(&ppgtt->base);
1550 static void gen6_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
1552 struct i915_address_space *vm = &ppgtt->base;
1553 struct i915_page_table *unused;
1554 gen6_pte_t scratch_pte;
1556 uint32_t pte, pde, temp;
1557 uint32_t start = ppgtt->base.start, length = ppgtt->base.total;
1559 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1560 I915_CACHE_LLC, true, 0);
1562 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde) {
1564 gen6_pte_t *pt_vaddr;
1565 const dma_addr_t pt_addr = px_dma(ppgtt->pd.page_table[pde]);
1566 pd_entry = readl(ppgtt->pd_addr + pde);
1567 expected = (GEN6_PDE_ADDR_ENCODE(pt_addr) | GEN6_PDE_VALID);
1569 if (pd_entry != expected)
1570 seq_printf(m, "\tPDE #%d mismatch: Actual PDE: %x Expected PDE: %x\n",
1574 seq_printf(m, "\tPDE: %x\n", pd_entry);
1576 pt_vaddr = kmap_px(ppgtt->pd.page_table[pde]);
1578 for (pte = 0; pte < GEN6_PTES; pte+=4) {
1580 (pde * PAGE_SIZE * GEN6_PTES) +
1584 for (i = 0; i < 4; i++)
1585 if (pt_vaddr[pte + i] != scratch_pte)
1590 seq_printf(m, "\t\t0x%lx [%03d,%04d]: =", va, pde, pte);
1591 for (i = 0; i < 4; i++) {
1592 if (pt_vaddr[pte + i] != scratch_pte)
1593 seq_printf(m, " %08x", pt_vaddr[pte + i]);
1595 seq_puts(m, " SCRATCH ");
1599 kunmap_px(ppgtt, pt_vaddr);
1603 /* Write pde (index) from the page directory @pd to the page table @pt */
1604 static void gen6_write_pde(struct i915_page_directory *pd,
1605 const int pde, struct i915_page_table *pt)
1607 /* Caller needs to make sure the write completes if necessary */
1608 struct i915_hw_ppgtt *ppgtt =
1609 container_of(pd, struct i915_hw_ppgtt, pd);
1612 pd_entry = GEN6_PDE_ADDR_ENCODE(px_dma(pt));
1613 pd_entry |= GEN6_PDE_VALID;
1615 writel(pd_entry, ppgtt->pd_addr + pde);
1618 /* Write all the page tables found in the ppgtt structure to incrementing page
1620 static void gen6_write_page_range(struct drm_i915_private *dev_priv,
1621 struct i915_page_directory *pd,
1622 uint32_t start, uint32_t length)
1624 struct i915_ggtt *ggtt = &dev_priv->ggtt;
1625 struct i915_page_table *pt;
1628 gen6_for_each_pde(pt, pd, start, length, temp, pde)
1629 gen6_write_pde(pd, pde, pt);
1631 /* Make sure write is complete before other code can use this page
1632 * table. Also require for WC mapped PTEs */
1636 static uint32_t get_pd_offset(struct i915_hw_ppgtt *ppgtt)
1638 BUG_ON(ppgtt->pd.base.ggtt_offset & 0x3f);
1640 return (ppgtt->pd.base.ggtt_offset / 64) << 16;
1643 static int hsw_mm_switch(struct i915_hw_ppgtt *ppgtt,
1644 struct drm_i915_gem_request *req)
1646 struct intel_engine_cs *engine = req->engine;
1649 /* NB: TLBs must be flushed and invalidated before a switch */
1650 ret = engine->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1654 ret = intel_ring_begin(req, 6);
1658 intel_ring_emit(engine, MI_LOAD_REGISTER_IMM(2));
1659 intel_ring_emit_reg(engine, RING_PP_DIR_DCLV(engine));
1660 intel_ring_emit(engine, PP_DIR_DCLV_2G);
1661 intel_ring_emit_reg(engine, RING_PP_DIR_BASE(engine));
1662 intel_ring_emit(engine, get_pd_offset(ppgtt));
1663 intel_ring_emit(engine, MI_NOOP);
1664 intel_ring_advance(engine);
1669 static int vgpu_mm_switch(struct i915_hw_ppgtt *ppgtt,
1670 struct drm_i915_gem_request *req)
1672 struct intel_engine_cs *engine = req->engine;
1673 struct drm_i915_private *dev_priv = to_i915(ppgtt->base.dev);
1675 I915_WRITE(RING_PP_DIR_DCLV(engine), PP_DIR_DCLV_2G);
1676 I915_WRITE(RING_PP_DIR_BASE(engine), get_pd_offset(ppgtt));
1680 static int gen7_mm_switch(struct i915_hw_ppgtt *ppgtt,
1681 struct drm_i915_gem_request *req)
1683 struct intel_engine_cs *engine = req->engine;
1686 /* NB: TLBs must be flushed and invalidated before a switch */
1687 ret = engine->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1691 ret = intel_ring_begin(req, 6);
1695 intel_ring_emit(engine, MI_LOAD_REGISTER_IMM(2));
1696 intel_ring_emit_reg(engine, RING_PP_DIR_DCLV(engine));
1697 intel_ring_emit(engine, PP_DIR_DCLV_2G);
1698 intel_ring_emit_reg(engine, RING_PP_DIR_BASE(engine));
1699 intel_ring_emit(engine, get_pd_offset(ppgtt));
1700 intel_ring_emit(engine, MI_NOOP);
1701 intel_ring_advance(engine);
1703 /* XXX: RCS is the only one to auto invalidate the TLBs? */
1704 if (engine->id != RCS) {
1705 ret = engine->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1713 static int gen6_mm_switch(struct i915_hw_ppgtt *ppgtt,
1714 struct drm_i915_gem_request *req)
1716 struct intel_engine_cs *engine = req->engine;
1717 struct drm_device *dev = ppgtt->base.dev;
1718 struct drm_i915_private *dev_priv = dev->dev_private;
1721 I915_WRITE(RING_PP_DIR_DCLV(engine), PP_DIR_DCLV_2G);
1722 I915_WRITE(RING_PP_DIR_BASE(engine), get_pd_offset(ppgtt));
1724 POSTING_READ(RING_PP_DIR_DCLV(engine));
1729 static void gen8_ppgtt_enable(struct drm_device *dev)
1731 struct drm_i915_private *dev_priv = dev->dev_private;
1732 struct intel_engine_cs *engine;
1734 for_each_engine(engine, dev_priv) {
1735 u32 four_level = USES_FULL_48BIT_PPGTT(dev) ? GEN8_GFX_PPGTT_48B : 0;
1736 I915_WRITE(RING_MODE_GEN7(engine),
1737 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE | four_level));
1741 static void gen7_ppgtt_enable(struct drm_device *dev)
1743 struct drm_i915_private *dev_priv = dev->dev_private;
1744 struct intel_engine_cs *engine;
1745 uint32_t ecochk, ecobits;
1747 ecobits = I915_READ(GAC_ECO_BITS);
1748 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B);
1750 ecochk = I915_READ(GAM_ECOCHK);
1751 if (IS_HASWELL(dev)) {
1752 ecochk |= ECOCHK_PPGTT_WB_HSW;
1754 ecochk |= ECOCHK_PPGTT_LLC_IVB;
1755 ecochk &= ~ECOCHK_PPGTT_GFDT_IVB;
1757 I915_WRITE(GAM_ECOCHK, ecochk);
1759 for_each_engine(engine, dev_priv) {
1760 /* GFX_MODE is per-ring on gen7+ */
1761 I915_WRITE(RING_MODE_GEN7(engine),
1762 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1766 static void gen6_ppgtt_enable(struct drm_device *dev)
1768 struct drm_i915_private *dev_priv = dev->dev_private;
1769 uint32_t ecochk, gab_ctl, ecobits;
1771 ecobits = I915_READ(GAC_ECO_BITS);
1772 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_SNB_BIT |
1773 ECOBITS_PPGTT_CACHE64B);
1775 gab_ctl = I915_READ(GAB_CTL);
1776 I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT);
1778 ecochk = I915_READ(GAM_ECOCHK);
1779 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT | ECOCHK_PPGTT_CACHE64B);
1781 I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1784 /* PPGTT support for Sandybdrige/Gen6 and later */
1785 static void gen6_ppgtt_clear_range(struct i915_address_space *vm,
1790 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1791 gen6_pte_t *pt_vaddr, scratch_pte;
1792 unsigned first_entry = start >> PAGE_SHIFT;
1793 unsigned num_entries = length >> PAGE_SHIFT;
1794 unsigned act_pt = first_entry / GEN6_PTES;
1795 unsigned first_pte = first_entry % GEN6_PTES;
1796 unsigned last_pte, i;
1798 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1799 I915_CACHE_LLC, true, 0);
1801 while (num_entries) {
1802 last_pte = first_pte + num_entries;
1803 if (last_pte > GEN6_PTES)
1804 last_pte = GEN6_PTES;
1806 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1808 for (i = first_pte; i < last_pte; i++)
1809 pt_vaddr[i] = scratch_pte;
1811 kunmap_px(ppgtt, pt_vaddr);
1813 num_entries -= last_pte - first_pte;
1819 static void gen6_ppgtt_insert_entries(struct i915_address_space *vm,
1820 struct sg_table *pages,
1822 enum i915_cache_level cache_level, u32 flags)
1824 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1825 gen6_pte_t *pt_vaddr;
1826 unsigned first_entry = start >> PAGE_SHIFT;
1827 unsigned act_pt = first_entry / GEN6_PTES;
1828 unsigned act_pte = first_entry % GEN6_PTES;
1829 struct sg_page_iter sg_iter;
1832 for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) {
1833 if (pt_vaddr == NULL)
1834 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1837 vm->pte_encode(sg_page_iter_dma_address(&sg_iter),
1838 cache_level, true, flags);
1840 if (++act_pte == GEN6_PTES) {
1841 kunmap_px(ppgtt, pt_vaddr);
1848 kunmap_px(ppgtt, pt_vaddr);
1851 static int gen6_alloc_va_range(struct i915_address_space *vm,
1852 uint64_t start_in, uint64_t length_in)
1854 DECLARE_BITMAP(new_page_tables, I915_PDES);
1855 struct drm_device *dev = vm->dev;
1856 struct drm_i915_private *dev_priv = to_i915(dev);
1857 struct i915_ggtt *ggtt = &dev_priv->ggtt;
1858 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1859 struct i915_page_table *pt;
1860 uint32_t start, length, start_save, length_save;
1864 if (WARN_ON(start_in + length_in > ppgtt->base.total))
1867 start = start_save = start_in;
1868 length = length_save = length_in;
1870 bitmap_zero(new_page_tables, I915_PDES);
1872 /* The allocation is done in two stages so that we can bail out with
1873 * minimal amount of pain. The first stage finds new page tables that
1874 * need allocation. The second stage marks use ptes within the page
1877 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1878 if (pt != vm->scratch_pt) {
1879 WARN_ON(bitmap_empty(pt->used_ptes, GEN6_PTES));
1883 /* We've already allocated a page table */
1884 WARN_ON(!bitmap_empty(pt->used_ptes, GEN6_PTES));
1892 gen6_initialize_pt(vm, pt);
1894 ppgtt->pd.page_table[pde] = pt;
1895 __set_bit(pde, new_page_tables);
1896 trace_i915_page_table_entry_alloc(vm, pde, start, GEN6_PDE_SHIFT);
1900 length = length_save;
1902 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1903 DECLARE_BITMAP(tmp_bitmap, GEN6_PTES);
1905 bitmap_zero(tmp_bitmap, GEN6_PTES);
1906 bitmap_set(tmp_bitmap, gen6_pte_index(start),
1907 gen6_pte_count(start, length));
1909 if (__test_and_clear_bit(pde, new_page_tables))
1910 gen6_write_pde(&ppgtt->pd, pde, pt);
1912 trace_i915_page_table_entry_map(vm, pde, pt,
1913 gen6_pte_index(start),
1914 gen6_pte_count(start, length),
1916 bitmap_or(pt->used_ptes, tmp_bitmap, pt->used_ptes,
1920 WARN_ON(!bitmap_empty(new_page_tables, I915_PDES));
1922 /* Make sure write is complete before other code can use this page
1923 * table. Also require for WC mapped PTEs */
1926 mark_tlbs_dirty(ppgtt);
1930 for_each_set_bit(pde, new_page_tables, I915_PDES) {
1931 struct i915_page_table *pt = ppgtt->pd.page_table[pde];
1933 ppgtt->pd.page_table[pde] = vm->scratch_pt;
1934 free_pt(vm->dev, pt);
1937 mark_tlbs_dirty(ppgtt);
1941 static int gen6_init_scratch(struct i915_address_space *vm)
1943 struct drm_device *dev = vm->dev;
1945 vm->scratch_page = alloc_scratch_page(dev);
1946 if (IS_ERR(vm->scratch_page))
1947 return PTR_ERR(vm->scratch_page);
1949 vm->scratch_pt = alloc_pt(dev);
1950 if (IS_ERR(vm->scratch_pt)) {
1951 free_scratch_page(dev, vm->scratch_page);
1952 return PTR_ERR(vm->scratch_pt);
1955 gen6_initialize_pt(vm, vm->scratch_pt);
1960 static void gen6_free_scratch(struct i915_address_space *vm)
1962 struct drm_device *dev = vm->dev;
1964 free_pt(dev, vm->scratch_pt);
1965 free_scratch_page(dev, vm->scratch_page);
1968 static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
1970 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1971 struct i915_page_table *pt;
1974 drm_mm_remove_node(&ppgtt->node);
1976 gen6_for_all_pdes(pt, ppgtt, pde) {
1977 if (pt != vm->scratch_pt)
1978 free_pt(ppgtt->base.dev, pt);
1981 gen6_free_scratch(vm);
1984 static int gen6_ppgtt_allocate_page_directories(struct i915_hw_ppgtt *ppgtt)
1986 struct i915_address_space *vm = &ppgtt->base;
1987 struct drm_device *dev = ppgtt->base.dev;
1988 struct drm_i915_private *dev_priv = to_i915(dev);
1989 struct i915_ggtt *ggtt = &dev_priv->ggtt;
1990 bool retried = false;
1993 /* PPGTT PDEs reside in the GGTT and consists of 512 entries. The
1994 * allocator works in address space sizes, so it's multiplied by page
1995 * size. We allocate at the top of the GTT to avoid fragmentation.
1997 BUG_ON(!drm_mm_initialized(&ggtt->base.mm));
1999 ret = gen6_init_scratch(vm);
2004 ret = drm_mm_insert_node_in_range_generic(&ggtt->base.mm,
2005 &ppgtt->node, GEN6_PD_SIZE,
2007 0, ggtt->base.total,
2009 if (ret == -ENOSPC && !retried) {
2010 ret = i915_gem_evict_something(dev, &ggtt->base,
2011 GEN6_PD_SIZE, GEN6_PD_ALIGN,
2013 0, ggtt->base.total,
2026 if (ppgtt->node.start < ggtt->mappable_end)
2027 DRM_DEBUG("Forced to use aperture for PDEs\n");
2032 gen6_free_scratch(vm);
2036 static int gen6_ppgtt_alloc(struct i915_hw_ppgtt *ppgtt)
2038 return gen6_ppgtt_allocate_page_directories(ppgtt);
2041 static void gen6_scratch_va_range(struct i915_hw_ppgtt *ppgtt,
2042 uint64_t start, uint64_t length)
2044 struct i915_page_table *unused;
2047 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde)
2048 ppgtt->pd.page_table[pde] = ppgtt->base.scratch_pt;
2051 static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
2053 struct drm_device *dev = ppgtt->base.dev;
2054 struct drm_i915_private *dev_priv = to_i915(dev);
2055 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2058 ppgtt->base.pte_encode = ggtt->base.pte_encode;
2060 ppgtt->switch_mm = gen6_mm_switch;
2061 } else if (IS_HASWELL(dev)) {
2062 ppgtt->switch_mm = hsw_mm_switch;
2063 } else if (IS_GEN7(dev)) {
2064 ppgtt->switch_mm = gen7_mm_switch;
2068 if (intel_vgpu_active(dev))
2069 ppgtt->switch_mm = vgpu_mm_switch;
2071 ret = gen6_ppgtt_alloc(ppgtt);
2075 ppgtt->base.allocate_va_range = gen6_alloc_va_range;
2076 ppgtt->base.clear_range = gen6_ppgtt_clear_range;
2077 ppgtt->base.insert_entries = gen6_ppgtt_insert_entries;
2078 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
2079 ppgtt->base.bind_vma = ppgtt_bind_vma;
2080 ppgtt->base.cleanup = gen6_ppgtt_cleanup;
2081 ppgtt->base.start = 0;
2082 ppgtt->base.total = I915_PDES * GEN6_PTES * PAGE_SIZE;
2083 ppgtt->debug_dump = gen6_dump_ppgtt;
2085 ppgtt->pd.base.ggtt_offset =
2086 ppgtt->node.start / PAGE_SIZE * sizeof(gen6_pte_t);
2088 ppgtt->pd_addr = (gen6_pte_t __iomem *)ggtt->gsm +
2089 ppgtt->pd.base.ggtt_offset / sizeof(gen6_pte_t);
2091 gen6_scratch_va_range(ppgtt, 0, ppgtt->base.total);
2093 gen6_write_page_range(dev_priv, &ppgtt->pd, 0, ppgtt->base.total);
2095 DRM_DEBUG_DRIVER("Allocated pde space (%lldM) at GTT entry: %llx\n",
2096 ppgtt->node.size >> 20,
2097 ppgtt->node.start / PAGE_SIZE);
2099 DRM_DEBUG("Adding PPGTT at offset %x\n",
2100 ppgtt->pd.base.ggtt_offset << 10);
2105 static int __hw_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
2107 ppgtt->base.dev = dev;
2109 if (INTEL_INFO(dev)->gen < 8)
2110 return gen6_ppgtt_init(ppgtt);
2112 return gen8_ppgtt_init(ppgtt);
2115 static void i915_address_space_init(struct i915_address_space *vm,
2116 struct drm_i915_private *dev_priv)
2118 drm_mm_init(&vm->mm, vm->start, vm->total);
2119 vm->dev = dev_priv->dev;
2120 INIT_LIST_HEAD(&vm->active_list);
2121 INIT_LIST_HEAD(&vm->inactive_list);
2122 list_add_tail(&vm->global_link, &dev_priv->vm_list);
2125 static void gtt_write_workarounds(struct drm_device *dev)
2127 struct drm_i915_private *dev_priv = dev->dev_private;
2129 /* This function is for gtt related workarounds. This function is
2130 * called on driver load and after a GPU reset, so you can place
2131 * workarounds here even if they get overwritten by GPU reset.
2133 /* WaIncreaseDefaultTLBEntries:chv,bdw,skl,bxt */
2134 if (IS_BROADWELL(dev))
2135 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN8_L3_LRA_1_GPGPU_DEFAULT_VALUE_BDW);
2136 else if (IS_CHERRYVIEW(dev))
2137 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN8_L3_LRA_1_GPGPU_DEFAULT_VALUE_CHV);
2138 else if (IS_SKYLAKE(dev))
2139 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN9_L3_LRA_1_GPGPU_DEFAULT_VALUE_SKL);
2140 else if (IS_BROXTON(dev))
2141 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN9_L3_LRA_1_GPGPU_DEFAULT_VALUE_BXT);
2144 int i915_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
2146 struct drm_i915_private *dev_priv = dev->dev_private;
2149 ret = __hw_ppgtt_init(dev, ppgtt);
2151 kref_init(&ppgtt->ref);
2152 i915_address_space_init(&ppgtt->base, dev_priv);
2158 int i915_ppgtt_init_hw(struct drm_device *dev)
2160 gtt_write_workarounds(dev);
2162 /* In the case of execlists, PPGTT is enabled by the context descriptor
2163 * and the PDPs are contained within the context itself. We don't
2164 * need to do anything here. */
2165 if (i915.enable_execlists)
2168 if (!USES_PPGTT(dev))
2172 gen6_ppgtt_enable(dev);
2173 else if (IS_GEN7(dev))
2174 gen7_ppgtt_enable(dev);
2175 else if (INTEL_INFO(dev)->gen >= 8)
2176 gen8_ppgtt_enable(dev);
2178 MISSING_CASE(INTEL_INFO(dev)->gen);
2183 int i915_ppgtt_init_ring(struct drm_i915_gem_request *req)
2185 struct drm_i915_private *dev_priv = req->i915;
2186 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
2188 if (i915.enable_execlists)
2194 return ppgtt->switch_mm(ppgtt, req);
2197 struct i915_hw_ppgtt *
2198 i915_ppgtt_create(struct drm_device *dev, struct drm_i915_file_private *fpriv)
2200 struct i915_hw_ppgtt *ppgtt;
2203 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2205 return ERR_PTR(-ENOMEM);
2207 ret = i915_ppgtt_init(dev, ppgtt);
2210 return ERR_PTR(ret);
2213 ppgtt->file_priv = fpriv;
2215 trace_i915_ppgtt_create(&ppgtt->base);
2220 void i915_ppgtt_release(struct kref *kref)
2222 struct i915_hw_ppgtt *ppgtt =
2223 container_of(kref, struct i915_hw_ppgtt, ref);
2225 trace_i915_ppgtt_release(&ppgtt->base);
2227 /* vmas should already be unbound */
2228 WARN_ON(!list_empty(&ppgtt->base.active_list));
2229 WARN_ON(!list_empty(&ppgtt->base.inactive_list));
2231 list_del(&ppgtt->base.global_link);
2232 drm_mm_takedown(&ppgtt->base.mm);
2234 ppgtt->base.cleanup(&ppgtt->base);
2238 extern int intel_iommu_gfx_mapped;
2239 /* Certain Gen5 chipsets require require idling the GPU before
2240 * unmapping anything from the GTT when VT-d is enabled.
2242 static bool needs_idle_maps(struct drm_device *dev)
2244 #ifdef CONFIG_INTEL_IOMMU
2245 /* Query intel_iommu to see if we need the workaround. Presumably that
2248 if (IS_GEN5(dev) && IS_MOBILE(dev) && intel_iommu_gfx_mapped)
2254 static bool do_idling(struct drm_i915_private *dev_priv)
2256 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2257 bool ret = dev_priv->mm.interruptible;
2259 if (unlikely(ggtt->do_idle_maps)) {
2260 dev_priv->mm.interruptible = false;
2261 if (i915_gpu_idle(dev_priv->dev)) {
2262 DRM_ERROR("Couldn't idle GPU\n");
2263 /* Wait a bit, in hopes it avoids the hang */
2271 static void undo_idling(struct drm_i915_private *dev_priv, bool interruptible)
2273 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2275 if (unlikely(ggtt->do_idle_maps))
2276 dev_priv->mm.interruptible = interruptible;
2279 void i915_check_and_clear_faults(struct drm_device *dev)
2281 struct drm_i915_private *dev_priv = dev->dev_private;
2282 struct intel_engine_cs *engine;
2284 if (INTEL_INFO(dev)->gen < 6)
2287 for_each_engine(engine, dev_priv) {
2289 fault_reg = I915_READ(RING_FAULT_REG(engine));
2290 if (fault_reg & RING_FAULT_VALID) {
2291 DRM_DEBUG_DRIVER("Unexpected fault\n"
2293 "\tAddress space: %s\n"
2296 fault_reg & PAGE_MASK,
2297 fault_reg & RING_FAULT_GTTSEL_MASK ? "GGTT" : "PPGTT",
2298 RING_FAULT_SRCID(fault_reg),
2299 RING_FAULT_FAULT_TYPE(fault_reg));
2300 I915_WRITE(RING_FAULT_REG(engine),
2301 fault_reg & ~RING_FAULT_VALID);
2304 POSTING_READ(RING_FAULT_REG(&dev_priv->engine[RCS]));
2307 static void i915_ggtt_flush(struct drm_i915_private *dev_priv)
2309 if (INTEL_INFO(dev_priv)->gen < 6) {
2310 intel_gtt_chipset_flush();
2312 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2313 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2317 void i915_gem_suspend_gtt_mappings(struct drm_device *dev)
2319 struct drm_i915_private *dev_priv = to_i915(dev);
2320 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2322 /* Don't bother messing with faults pre GEN6 as we have little
2323 * documentation supporting that it's a good idea.
2325 if (INTEL_INFO(dev)->gen < 6)
2328 i915_check_and_clear_faults(dev);
2330 ggtt->base.clear_range(&ggtt->base, ggtt->base.start, ggtt->base.total,
2333 i915_ggtt_flush(dev_priv);
2336 int i915_gem_gtt_prepare_object(struct drm_i915_gem_object *obj)
2338 if (!dma_map_sg(&obj->base.dev->pdev->dev,
2339 obj->pages->sgl, obj->pages->nents,
2340 PCI_DMA_BIDIRECTIONAL))
2346 static void gen8_set_pte(void __iomem *addr, gen8_pte_t pte)
2351 iowrite32((u32)pte, addr);
2352 iowrite32(pte >> 32, addr + 4);
2356 static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
2357 struct sg_table *st,
2359 enum i915_cache_level level, u32 unused)
2361 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2362 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2363 unsigned first_entry = start >> PAGE_SHIFT;
2364 gen8_pte_t __iomem *gtt_entries =
2365 (gen8_pte_t __iomem *)ggtt->gsm + first_entry;
2367 struct sg_page_iter sg_iter;
2368 dma_addr_t addr = 0; /* shut up gcc */
2371 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2373 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
2374 addr = sg_dma_address(sg_iter.sg) +
2375 (sg_iter.sg_pgoffset << PAGE_SHIFT);
2376 gen8_set_pte(>t_entries[i],
2377 gen8_pte_encode(addr, level, true));
2382 * XXX: This serves as a posting read to make sure that the PTE has
2383 * actually been updated. There is some concern that even though
2384 * registers and PTEs are within the same BAR that they are potentially
2385 * of NUMA access patterns. Therefore, even with the way we assume
2386 * hardware should work, we must keep this posting read for paranoia.
2389 WARN_ON(readq(>t_entries[i-1])
2390 != gen8_pte_encode(addr, level, true));
2392 /* This next bit makes the above posting read even more important. We
2393 * want to flush the TLBs only after we're certain all the PTE updates
2396 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2397 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2399 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2402 struct insert_entries {
2403 struct i915_address_space *vm;
2404 struct sg_table *st;
2406 enum i915_cache_level level;
2410 static int gen8_ggtt_insert_entries__cb(void *_arg)
2412 struct insert_entries *arg = _arg;
2413 gen8_ggtt_insert_entries(arg->vm, arg->st,
2414 arg->start, arg->level, arg->flags);
2418 static void gen8_ggtt_insert_entries__BKL(struct i915_address_space *vm,
2419 struct sg_table *st,
2421 enum i915_cache_level level,
2424 struct insert_entries arg = { vm, st, start, level, flags };
2425 stop_machine(gen8_ggtt_insert_entries__cb, &arg, NULL);
2429 * Binds an object into the global gtt with the specified cache level. The object
2430 * will be accessible to the GPU via commands whose operands reference offsets
2431 * within the global GTT as well as accessible by the GPU through the GMADR
2432 * mapped BAR (dev_priv->mm.gtt->gtt).
2434 static void gen6_ggtt_insert_entries(struct i915_address_space *vm,
2435 struct sg_table *st,
2437 enum i915_cache_level level, u32 flags)
2439 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2440 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2441 unsigned first_entry = start >> PAGE_SHIFT;
2442 gen6_pte_t __iomem *gtt_entries =
2443 (gen6_pte_t __iomem *)ggtt->gsm + first_entry;
2445 struct sg_page_iter sg_iter;
2446 dma_addr_t addr = 0;
2449 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2451 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
2452 addr = sg_page_iter_dma_address(&sg_iter);
2453 iowrite32(vm->pte_encode(addr, level, true, flags), >t_entries[i]);
2457 /* XXX: This serves as a posting read to make sure that the PTE has
2458 * actually been updated. There is some concern that even though
2459 * registers and PTEs are within the same BAR that they are potentially
2460 * of NUMA access patterns. Therefore, even with the way we assume
2461 * hardware should work, we must keep this posting read for paranoia.
2464 unsigned long gtt = readl(>t_entries[i-1]);
2465 WARN_ON(gtt != vm->pte_encode(addr, level, true, flags));
2468 /* This next bit makes the above posting read even more important. We
2469 * want to flush the TLBs only after we're certain all the PTE updates
2472 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2473 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2475 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2478 static void gen8_ggtt_clear_range(struct i915_address_space *vm,
2483 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2484 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2485 unsigned first_entry = start >> PAGE_SHIFT;
2486 unsigned num_entries = length >> PAGE_SHIFT;
2487 gen8_pte_t scratch_pte, __iomem *gtt_base =
2488 (gen8_pte_t __iomem *)ggtt->gsm + first_entry;
2489 const int max_entries = ggtt_total_entries(ggtt) - first_entry;
2493 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2495 if (WARN(num_entries > max_entries,
2496 "First entry = %d; Num entries = %d (max=%d)\n",
2497 first_entry, num_entries, max_entries))
2498 num_entries = max_entries;
2500 scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
2503 for (i = 0; i < num_entries; i++)
2504 gen8_set_pte(>t_base[i], scratch_pte);
2507 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2510 static void gen6_ggtt_clear_range(struct i915_address_space *vm,
2515 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2516 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2517 unsigned first_entry = start >> PAGE_SHIFT;
2518 unsigned num_entries = length >> PAGE_SHIFT;
2519 gen6_pte_t scratch_pte, __iomem *gtt_base =
2520 (gen6_pte_t __iomem *)ggtt->gsm + first_entry;
2521 const int max_entries = ggtt_total_entries(ggtt) - first_entry;
2525 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2527 if (WARN(num_entries > max_entries,
2528 "First entry = %d; Num entries = %d (max=%d)\n",
2529 first_entry, num_entries, max_entries))
2530 num_entries = max_entries;
2532 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
2533 I915_CACHE_LLC, use_scratch, 0);
2535 for (i = 0; i < num_entries; i++)
2536 iowrite32(scratch_pte, >t_base[i]);
2539 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2542 static void i915_ggtt_insert_entries(struct i915_address_space *vm,
2543 struct sg_table *pages,
2545 enum i915_cache_level cache_level, u32 unused)
2547 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2548 unsigned int flags = (cache_level == I915_CACHE_NONE) ?
2549 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
2552 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2554 intel_gtt_insert_sg_entries(pages, start >> PAGE_SHIFT, flags);
2556 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2560 static void i915_ggtt_clear_range(struct i915_address_space *vm,
2565 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2566 unsigned first_entry = start >> PAGE_SHIFT;
2567 unsigned num_entries = length >> PAGE_SHIFT;
2570 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2572 intel_gtt_clear_range(first_entry, num_entries);
2574 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2577 static int ggtt_bind_vma(struct i915_vma *vma,
2578 enum i915_cache_level cache_level,
2581 struct drm_i915_gem_object *obj = vma->obj;
2585 ret = i915_get_ggtt_vma_pages(vma);
2589 /* Currently applicable only to VLV */
2591 pte_flags |= PTE_READ_ONLY;
2593 vma->vm->insert_entries(vma->vm, vma->ggtt_view.pages,
2595 cache_level, pte_flags);
2598 * Without aliasing PPGTT there's no difference between
2599 * GLOBAL/LOCAL_BIND, it's all the same ptes. Hence unconditionally
2600 * upgrade to both bound if we bind either to avoid double-binding.
2602 vma->bound |= GLOBAL_BIND | LOCAL_BIND;
2607 static int aliasing_gtt_bind_vma(struct i915_vma *vma,
2608 enum i915_cache_level cache_level,
2614 ret = i915_get_ggtt_vma_pages(vma);
2618 /* Currently applicable only to VLV */
2620 if (vma->obj->gt_ro)
2621 pte_flags |= PTE_READ_ONLY;
2624 if (flags & GLOBAL_BIND) {
2625 vma->vm->insert_entries(vma->vm,
2626 vma->ggtt_view.pages,
2628 cache_level, pte_flags);
2631 if (flags & LOCAL_BIND) {
2632 struct i915_hw_ppgtt *appgtt =
2633 to_i915(vma->vm->dev)->mm.aliasing_ppgtt;
2634 appgtt->base.insert_entries(&appgtt->base,
2635 vma->ggtt_view.pages,
2637 cache_level, pte_flags);
2643 static void ggtt_unbind_vma(struct i915_vma *vma)
2645 struct drm_device *dev = vma->vm->dev;
2646 struct drm_i915_private *dev_priv = dev->dev_private;
2647 struct drm_i915_gem_object *obj = vma->obj;
2648 const uint64_t size = min_t(uint64_t,
2652 if (vma->bound & GLOBAL_BIND) {
2653 vma->vm->clear_range(vma->vm,
2659 if (dev_priv->mm.aliasing_ppgtt && vma->bound & LOCAL_BIND) {
2660 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
2662 appgtt->base.clear_range(&appgtt->base,
2669 void i915_gem_gtt_finish_object(struct drm_i915_gem_object *obj)
2671 struct drm_device *dev = obj->base.dev;
2672 struct drm_i915_private *dev_priv = dev->dev_private;
2675 interruptible = do_idling(dev_priv);
2677 dma_unmap_sg(&dev->pdev->dev, obj->pages->sgl, obj->pages->nents,
2678 PCI_DMA_BIDIRECTIONAL);
2680 undo_idling(dev_priv, interruptible);
2683 static void i915_gtt_color_adjust(struct drm_mm_node *node,
2684 unsigned long color,
2688 if (node->color != color)
2691 if (!list_empty(&node->node_list)) {
2692 node = list_entry(node->node_list.next,
2695 if (node->allocated && node->color != color)
2700 static int i915_gem_setup_global_gtt(struct drm_device *dev,
2705 /* Let GEM Manage all of the aperture.
2707 * However, leave one page at the end still bound to the scratch page.
2708 * There are a number of places where the hardware apparently prefetches
2709 * past the end of the object, and we've seen multiple hangs with the
2710 * GPU head pointer stuck in a batchbuffer bound at the last page of the
2711 * aperture. One page should be enough to keep any prefetching inside
2714 struct drm_i915_private *dev_priv = to_i915(dev);
2715 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2716 struct drm_mm_node *entry;
2717 struct drm_i915_gem_object *obj;
2718 unsigned long hole_start, hole_end;
2721 BUG_ON(mappable_end > end);
2723 ggtt->base.start = start;
2725 /* Subtract the guard page before address space initialization to
2726 * shrink the range used by drm_mm */
2727 ggtt->base.total = end - start - PAGE_SIZE;
2728 i915_address_space_init(&ggtt->base, dev_priv);
2729 ggtt->base.total += PAGE_SIZE;
2731 if (intel_vgpu_active(dev)) {
2732 ret = intel_vgt_balloon(dev);
2738 ggtt->base.mm.color_adjust = i915_gtt_color_adjust;
2740 /* Mark any preallocated objects as occupied */
2741 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
2742 struct i915_vma *vma = i915_gem_obj_to_vma(obj, &ggtt->base);
2744 DRM_DEBUG_KMS("reserving preallocated space: %llx + %zx\n",
2745 i915_gem_obj_ggtt_offset(obj), obj->base.size);
2747 WARN_ON(i915_gem_obj_ggtt_bound(obj));
2748 ret = drm_mm_reserve_node(&ggtt->base.mm, &vma->node);
2750 DRM_DEBUG_KMS("Reservation failed: %i\n", ret);
2753 vma->bound |= GLOBAL_BIND;
2754 __i915_vma_set_map_and_fenceable(vma);
2755 list_add_tail(&vma->vm_link, &ggtt->base.inactive_list);
2758 /* Clear any non-preallocated blocks */
2759 drm_mm_for_each_hole(entry, &ggtt->base.mm, hole_start, hole_end) {
2760 DRM_DEBUG_KMS("clearing unused GTT space: [%lx, %lx]\n",
2761 hole_start, hole_end);
2762 ggtt->base.clear_range(&ggtt->base, hole_start,
2763 hole_end - hole_start, true);
2766 /* And finally clear the reserved guard page */
2767 ggtt->base.clear_range(&ggtt->base, end - PAGE_SIZE, PAGE_SIZE, true);
2769 if (USES_PPGTT(dev) && !USES_FULL_PPGTT(dev)) {
2770 struct i915_hw_ppgtt *ppgtt;
2772 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2776 ret = __hw_ppgtt_init(dev, ppgtt);
2778 ppgtt->base.cleanup(&ppgtt->base);
2783 if (ppgtt->base.allocate_va_range)
2784 ret = ppgtt->base.allocate_va_range(&ppgtt->base, 0,
2787 ppgtt->base.cleanup(&ppgtt->base);
2792 ppgtt->base.clear_range(&ppgtt->base,
2797 dev_priv->mm.aliasing_ppgtt = ppgtt;
2798 WARN_ON(ggtt->base.bind_vma != ggtt_bind_vma);
2799 ggtt->base.bind_vma = aliasing_gtt_bind_vma;
2806 * i915_gem_init_ggtt - Initialize GEM for Global GTT
2809 void i915_gem_init_ggtt(struct drm_device *dev)
2811 struct drm_i915_private *dev_priv = to_i915(dev);
2812 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2814 i915_gem_setup_global_gtt(dev, 0, ggtt->mappable_end, ggtt->base.total);
2818 * i915_ggtt_cleanup_hw - Clean up GGTT hardware initialization
2821 void i915_ggtt_cleanup_hw(struct drm_device *dev)
2823 struct drm_i915_private *dev_priv = to_i915(dev);
2824 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2826 if (dev_priv->mm.aliasing_ppgtt) {
2827 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
2829 ppgtt->base.cleanup(&ppgtt->base);
2832 i915_gem_cleanup_stolen(dev);
2834 if (drm_mm_initialized(&ggtt->base.mm)) {
2835 if (intel_vgpu_active(dev))
2836 intel_vgt_deballoon();
2838 drm_mm_takedown(&ggtt->base.mm);
2839 list_del(&ggtt->base.global_link);
2842 ggtt->base.cleanup(&ggtt->base);
2845 static unsigned int gen6_get_total_gtt_size(u16 snb_gmch_ctl)
2847 snb_gmch_ctl >>= SNB_GMCH_GGMS_SHIFT;
2848 snb_gmch_ctl &= SNB_GMCH_GGMS_MASK;
2849 return snb_gmch_ctl << 20;
2852 static unsigned int gen8_get_total_gtt_size(u16 bdw_gmch_ctl)
2854 bdw_gmch_ctl >>= BDW_GMCH_GGMS_SHIFT;
2855 bdw_gmch_ctl &= BDW_GMCH_GGMS_MASK;
2857 bdw_gmch_ctl = 1 << bdw_gmch_ctl;
2859 #ifdef CONFIG_X86_32
2860 /* Limit 32b platforms to a 2GB GGTT: 4 << 20 / pte size * PAGE_SIZE */
2861 if (bdw_gmch_ctl > 4)
2865 return bdw_gmch_ctl << 20;
2868 static unsigned int chv_get_total_gtt_size(u16 gmch_ctrl)
2870 gmch_ctrl >>= SNB_GMCH_GGMS_SHIFT;
2871 gmch_ctrl &= SNB_GMCH_GGMS_MASK;
2874 return 1 << (20 + gmch_ctrl);
2879 static size_t gen6_get_stolen_size(u16 snb_gmch_ctl)
2881 snb_gmch_ctl >>= SNB_GMCH_GMS_SHIFT;
2882 snb_gmch_ctl &= SNB_GMCH_GMS_MASK;
2883 return snb_gmch_ctl << 25; /* 32 MB units */
2886 static size_t gen8_get_stolen_size(u16 bdw_gmch_ctl)
2888 bdw_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2889 bdw_gmch_ctl &= BDW_GMCH_GMS_MASK;
2890 return bdw_gmch_ctl << 25; /* 32 MB units */
2893 static size_t chv_get_stolen_size(u16 gmch_ctrl)
2895 gmch_ctrl >>= SNB_GMCH_GMS_SHIFT;
2896 gmch_ctrl &= SNB_GMCH_GMS_MASK;
2899 * 0x0 to 0x10: 32MB increments starting at 0MB
2900 * 0x11 to 0x16: 4MB increments starting at 8MB
2901 * 0x17 to 0x1d: 4MB increments start at 36MB
2903 if (gmch_ctrl < 0x11)
2904 return gmch_ctrl << 25;
2905 else if (gmch_ctrl < 0x17)
2906 return (gmch_ctrl - 0x11 + 2) << 22;
2908 return (gmch_ctrl - 0x17 + 9) << 22;
2911 static size_t gen9_get_stolen_size(u16 gen9_gmch_ctl)
2913 gen9_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2914 gen9_gmch_ctl &= BDW_GMCH_GMS_MASK;
2916 if (gen9_gmch_ctl < 0xf0)
2917 return gen9_gmch_ctl << 25; /* 32 MB units */
2919 /* 4MB increments starting at 0xf0 for 4MB */
2920 return (gen9_gmch_ctl - 0xf0 + 1) << 22;
2923 static int ggtt_probe_common(struct drm_device *dev,
2926 struct drm_i915_private *dev_priv = to_i915(dev);
2927 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2928 struct i915_page_scratch *scratch_page;
2929 phys_addr_t ggtt_phys_addr;
2931 /* For Modern GENs the PTEs and register space are split in the BAR */
2932 ggtt_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 ggtt->gsm = ioremap_nocache(ggtt_phys_addr, gtt_size);
2945 ggtt->gsm = ioremap_wc(ggtt_phys_addr, gtt_size);
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 */
2956 return PTR_ERR(scratch_page);
2959 ggtt->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))
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 i915_ggtt *ggtt)
3039 struct drm_device *dev = ggtt->base.dev;
3040 struct drm_i915_private *dev_priv = to_i915(dev);
3044 /* TODO: We're not aware of mappable constraints on gen8 yet */
3045 ggtt->mappable_base = pci_resource_start(dev->pdev, 2);
3046 ggtt->mappable_end = pci_resource_len(dev->pdev, 2);
3048 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(39)))
3049 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(39));
3051 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
3053 if (INTEL_INFO(dev)->gen >= 9) {
3054 ggtt->stolen_size = gen9_get_stolen_size(snb_gmch_ctl);
3055 ggtt->size = gen8_get_total_gtt_size(snb_gmch_ctl);
3056 } else if (IS_CHERRYVIEW(dev)) {
3057 ggtt->stolen_size = chv_get_stolen_size(snb_gmch_ctl);
3058 ggtt->size = chv_get_total_gtt_size(snb_gmch_ctl);
3060 ggtt->stolen_size = gen8_get_stolen_size(snb_gmch_ctl);
3061 ggtt->size = gen8_get_total_gtt_size(snb_gmch_ctl);
3064 ggtt->base.total = (ggtt->size / sizeof(gen8_pte_t)) << PAGE_SHIFT;
3066 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
3067 chv_setup_private_ppat(dev_priv);
3069 bdw_setup_private_ppat(dev_priv);
3071 ret = ggtt_probe_common(dev, ggtt->size);
3073 ggtt->base.clear_range = gen8_ggtt_clear_range;
3074 if (IS_CHERRYVIEW(dev_priv))
3075 ggtt->base.insert_entries = gen8_ggtt_insert_entries__BKL;
3077 ggtt->base.insert_entries = gen8_ggtt_insert_entries;
3078 ggtt->base.bind_vma = ggtt_bind_vma;
3079 ggtt->base.unbind_vma = ggtt_unbind_vma;
3084 static int gen6_gmch_probe(struct i915_ggtt *ggtt)
3086 struct drm_device *dev = ggtt->base.dev;
3090 ggtt->mappable_base = pci_resource_start(dev->pdev, 2);
3091 ggtt->mappable_end = pci_resource_len(dev->pdev, 2);
3093 /* 64/512MB is the current min/max we actually know of, but this is just
3094 * a coarse sanity check.
3096 if ((ggtt->mappable_end < (64<<20) || (ggtt->mappable_end > (512<<20)))) {
3097 DRM_ERROR("Unknown GMADR size (%llx)\n", ggtt->mappable_end);
3101 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(40)))
3102 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(40));
3103 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
3105 ggtt->stolen_size = gen6_get_stolen_size(snb_gmch_ctl);
3106 ggtt->size = gen6_get_total_gtt_size(snb_gmch_ctl);
3107 ggtt->base.total = (ggtt->size / sizeof(gen6_pte_t)) << PAGE_SHIFT;
3109 ret = ggtt_probe_common(dev, ggtt->size);
3111 ggtt->base.clear_range = gen6_ggtt_clear_range;
3112 ggtt->base.insert_entries = gen6_ggtt_insert_entries;
3113 ggtt->base.bind_vma = ggtt_bind_vma;
3114 ggtt->base.unbind_vma = ggtt_unbind_vma;
3119 static void gen6_gmch_remove(struct i915_address_space *vm)
3121 struct i915_ggtt *ggtt = container_of(vm, struct i915_ggtt, base);
3124 free_scratch_page(vm->dev, vm->scratch_page);
3127 static int i915_gmch_probe(struct i915_ggtt *ggtt)
3129 struct drm_device *dev = ggtt->base.dev;
3130 struct drm_i915_private *dev_priv = to_i915(dev);
3133 ret = intel_gmch_probe(dev_priv->bridge_dev, dev_priv->dev->pdev, NULL);
3135 DRM_ERROR("failed to set up gmch\n");
3139 intel_gtt_get(&ggtt->base.total, &ggtt->stolen_size,
3140 &ggtt->mappable_base, &ggtt->mappable_end);
3142 ggtt->do_idle_maps = needs_idle_maps(dev_priv->dev);
3143 ggtt->base.insert_entries = i915_ggtt_insert_entries;
3144 ggtt->base.clear_range = i915_ggtt_clear_range;
3145 ggtt->base.bind_vma = ggtt_bind_vma;
3146 ggtt->base.unbind_vma = ggtt_unbind_vma;
3148 if (unlikely(ggtt->do_idle_maps))
3149 DRM_INFO("applying Ironlake quirks for intel_iommu\n");
3154 static void i915_gmch_remove(struct i915_address_space *vm)
3156 intel_gmch_remove();
3160 * i915_ggtt_init_hw - Initialize GGTT hardware
3163 int i915_ggtt_init_hw(struct drm_device *dev)
3165 struct drm_i915_private *dev_priv = to_i915(dev);
3166 struct i915_ggtt *ggtt = &dev_priv->ggtt;
3169 if (INTEL_INFO(dev)->gen <= 5) {
3170 ggtt->probe = i915_gmch_probe;
3171 ggtt->base.cleanup = i915_gmch_remove;
3172 } else if (INTEL_INFO(dev)->gen < 8) {
3173 ggtt->probe = gen6_gmch_probe;
3174 ggtt->base.cleanup = gen6_gmch_remove;
3177 ggtt->base.pte_encode = iris_pte_encode;
3178 else if (IS_HASWELL(dev))
3179 ggtt->base.pte_encode = hsw_pte_encode;
3180 else if (IS_VALLEYVIEW(dev))
3181 ggtt->base.pte_encode = byt_pte_encode;
3182 else if (INTEL_INFO(dev)->gen >= 7)
3183 ggtt->base.pte_encode = ivb_pte_encode;
3185 ggtt->base.pte_encode = snb_pte_encode;
3187 ggtt->probe = gen8_gmch_probe;
3188 ggtt->base.cleanup = gen6_gmch_remove;
3191 ggtt->base.dev = dev;
3192 ggtt->base.is_ggtt = true;
3194 ret = ggtt->probe(ggtt);
3198 if ((ggtt->base.total - 1) >> 32) {
3199 DRM_ERROR("We never expected a Global GTT with more than 32bits"
3200 "of address space! Found %lldM!\n",
3201 ggtt->base.total >> 20);
3202 ggtt->base.total = 1ULL << 32;
3203 ggtt->mappable_end = min(ggtt->mappable_end, ggtt->base.total);
3207 * Initialise stolen early so that we may reserve preallocated
3208 * objects for the BIOS to KMS transition.
3210 ret = i915_gem_init_stolen(dev);
3212 goto out_gtt_cleanup;
3214 /* GMADR is the PCI mmio aperture into the global GTT. */
3215 DRM_INFO("Memory usable by graphics device = %lluM\n",
3216 ggtt->base.total >> 20);
3217 DRM_DEBUG_DRIVER("GMADR size = %lldM\n", ggtt->mappable_end >> 20);
3218 DRM_DEBUG_DRIVER("GTT stolen size = %zdM\n", ggtt->stolen_size >> 20);
3219 #ifdef CONFIG_INTEL_IOMMU
3220 if (intel_iommu_gfx_mapped)
3221 DRM_INFO("VT-d active for gfx access\n");
3224 * i915.enable_ppgtt is read-only, so do an early pass to validate the
3225 * user's requested state against the hardware/driver capabilities. We
3226 * do this now so that we can print out any log messages once rather
3227 * than every time we check intel_enable_ppgtt().
3229 i915.enable_ppgtt = sanitize_enable_ppgtt(dev, i915.enable_ppgtt);
3230 DRM_DEBUG_DRIVER("ppgtt mode: %i\n", i915.enable_ppgtt);
3235 ggtt->base.cleanup(&ggtt->base);
3240 void i915_gem_restore_gtt_mappings(struct drm_device *dev)
3242 struct drm_i915_private *dev_priv = to_i915(dev);
3243 struct i915_ggtt *ggtt = &dev_priv->ggtt;
3244 struct drm_i915_gem_object *obj;
3245 struct i915_vma *vma;
3248 i915_check_and_clear_faults(dev);
3250 /* First fill our portion of the GTT with scratch pages */
3251 ggtt->base.clear_range(&ggtt->base, ggtt->base.start, ggtt->base.total,
3254 /* Cache flush objects bound into GGTT and rebind them. */
3255 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
3257 list_for_each_entry(vma, &obj->vma_list, obj_link) {
3258 if (vma->vm != &ggtt->base)
3261 WARN_ON(i915_vma_bind(vma, obj->cache_level,
3268 i915_gem_clflush_object(obj, obj->pin_display);
3271 if (INTEL_INFO(dev)->gen >= 8) {
3272 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
3273 chv_setup_private_ppat(dev_priv);
3275 bdw_setup_private_ppat(dev_priv);
3280 if (USES_PPGTT(dev)) {
3281 struct i915_address_space *vm;
3283 list_for_each_entry(vm, &dev_priv->vm_list, global_link) {
3284 /* TODO: Perhaps it shouldn't be gen6 specific */
3286 struct i915_hw_ppgtt *ppgtt;
3289 ppgtt = dev_priv->mm.aliasing_ppgtt;
3291 ppgtt = i915_vm_to_ppgtt(vm);
3293 gen6_write_page_range(dev_priv, &ppgtt->pd,
3294 0, ppgtt->base.total);
3298 i915_ggtt_flush(dev_priv);
3301 static struct i915_vma *
3302 __i915_gem_vma_create(struct drm_i915_gem_object *obj,
3303 struct i915_address_space *vm,
3304 const struct i915_ggtt_view *ggtt_view)
3306 struct i915_vma *vma;
3308 if (WARN_ON(i915_is_ggtt(vm) != !!ggtt_view))
3309 return ERR_PTR(-EINVAL);
3311 vma = kmem_cache_zalloc(to_i915(obj->base.dev)->vmas, GFP_KERNEL);
3313 return ERR_PTR(-ENOMEM);
3315 INIT_LIST_HEAD(&vma->vm_link);
3316 INIT_LIST_HEAD(&vma->obj_link);
3317 INIT_LIST_HEAD(&vma->exec_list);
3320 vma->is_ggtt = i915_is_ggtt(vm);
3322 if (i915_is_ggtt(vm))
3323 vma->ggtt_view = *ggtt_view;
3325 i915_ppgtt_get(i915_vm_to_ppgtt(vm));
3327 list_add_tail(&vma->obj_link, &obj->vma_list);
3333 i915_gem_obj_lookup_or_create_vma(struct drm_i915_gem_object *obj,
3334 struct i915_address_space *vm)
3336 struct i915_vma *vma;
3338 vma = i915_gem_obj_to_vma(obj, vm);
3340 vma = __i915_gem_vma_create(obj, vm,
3341 i915_is_ggtt(vm) ? &i915_ggtt_view_normal : NULL);
3347 i915_gem_obj_lookup_or_create_ggtt_vma(struct drm_i915_gem_object *obj,
3348 const struct i915_ggtt_view *view)
3350 struct drm_device *dev = obj->base.dev;
3351 struct drm_i915_private *dev_priv = to_i915(dev);
3352 struct i915_ggtt *ggtt = &dev_priv->ggtt;
3353 struct i915_vma *vma = i915_gem_obj_to_ggtt_view(obj, view);
3356 vma = __i915_gem_vma_create(obj, &ggtt->base, view);
3362 static struct scatterlist *
3363 rotate_pages(const dma_addr_t *in, unsigned int offset,
3364 unsigned int width, unsigned int height,
3365 unsigned int stride,
3366 struct sg_table *st, struct scatterlist *sg)
3368 unsigned int column, row;
3369 unsigned int src_idx;
3371 for (column = 0; column < width; column++) {
3372 src_idx = stride * (height - 1) + column;
3373 for (row = 0; row < height; row++) {
3375 /* We don't need the pages, but need to initialize
3376 * the entries so the sg list can be happily traversed.
3377 * The only thing we need are DMA addresses.
3379 sg_set_page(sg, NULL, PAGE_SIZE, 0);
3380 sg_dma_address(sg) = in[offset + src_idx];
3381 sg_dma_len(sg) = PAGE_SIZE;
3390 static struct sg_table *
3391 intel_rotate_fb_obj_pages(struct intel_rotation_info *rot_info,
3392 struct drm_i915_gem_object *obj)
3394 unsigned int size_pages = rot_info->plane[0].width * rot_info->plane[0].height;
3395 unsigned int size_pages_uv;
3396 struct sg_page_iter sg_iter;
3398 dma_addr_t *page_addr_list;
3399 struct sg_table *st;
3400 unsigned int uv_start_page;
3401 struct scatterlist *sg;
3404 /* Allocate a temporary list of source pages for random access. */
3405 page_addr_list = drm_malloc_gfp(obj->base.size / PAGE_SIZE,
3408 if (!page_addr_list)
3409 return ERR_PTR(ret);
3411 /* Account for UV plane with NV12. */
3412 if (rot_info->pixel_format == DRM_FORMAT_NV12)
3413 size_pages_uv = rot_info->plane[1].width * rot_info->plane[1].height;
3417 /* Allocate target SG list. */
3418 st = kmalloc(sizeof(*st), GFP_KERNEL);
3422 ret = sg_alloc_table(st, size_pages + size_pages_uv, GFP_KERNEL);
3426 /* Populate source page list from the object. */
3428 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) {
3429 page_addr_list[i] = sg_page_iter_dma_address(&sg_iter);
3436 /* Rotate the pages. */
3437 sg = rotate_pages(page_addr_list, 0,
3438 rot_info->plane[0].width, rot_info->plane[0].height,
3439 rot_info->plane[0].width,
3442 /* Append the UV plane if NV12. */
3443 if (rot_info->pixel_format == DRM_FORMAT_NV12) {
3444 uv_start_page = size_pages;
3446 /* Check for tile-row un-alignment. */
3447 if (offset_in_page(rot_info->uv_offset))
3450 rot_info->uv_start_page = uv_start_page;
3452 sg = rotate_pages(page_addr_list, rot_info->uv_start_page,
3453 rot_info->plane[1].width, rot_info->plane[1].height,
3454 rot_info->plane[1].width,
3458 DRM_DEBUG_KMS("Created rotated page mapping for object size %zu (%ux%u tiles, %u pages (%u plane 0)).\n",
3459 obj->base.size, rot_info->plane[0].width,
3460 rot_info->plane[0].height, size_pages + size_pages_uv,
3463 drm_free_large(page_addr_list);
3470 drm_free_large(page_addr_list);
3472 DRM_DEBUG_KMS("Failed to create rotated mapping for object size %zu! (%d) (%ux%u tiles, %u pages (%u plane 0))\n",
3473 obj->base.size, ret, rot_info->plane[0].width,
3474 rot_info->plane[0].height, size_pages + size_pages_uv,
3476 return ERR_PTR(ret);
3479 static struct sg_table *
3480 intel_partial_pages(const struct i915_ggtt_view *view,
3481 struct drm_i915_gem_object *obj)
3483 struct sg_table *st;
3484 struct scatterlist *sg;
3485 struct sg_page_iter obj_sg_iter;
3488 st = kmalloc(sizeof(*st), GFP_KERNEL);
3492 ret = sg_alloc_table(st, view->params.partial.size, GFP_KERNEL);
3498 for_each_sg_page(obj->pages->sgl, &obj_sg_iter, obj->pages->nents,
3499 view->params.partial.offset)
3501 if (st->nents >= view->params.partial.size)
3504 sg_set_page(sg, NULL, PAGE_SIZE, 0);
3505 sg_dma_address(sg) = sg_page_iter_dma_address(&obj_sg_iter);
3506 sg_dma_len(sg) = PAGE_SIZE;
3517 return ERR_PTR(ret);
3521 i915_get_ggtt_vma_pages(struct i915_vma *vma)
3525 if (vma->ggtt_view.pages)
3528 if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL)
3529 vma->ggtt_view.pages = vma->obj->pages;
3530 else if (vma->ggtt_view.type == I915_GGTT_VIEW_ROTATED)
3531 vma->ggtt_view.pages =
3532 intel_rotate_fb_obj_pages(&vma->ggtt_view.params.rotated, vma->obj);
3533 else if (vma->ggtt_view.type == I915_GGTT_VIEW_PARTIAL)
3534 vma->ggtt_view.pages =
3535 intel_partial_pages(&vma->ggtt_view, vma->obj);
3537 WARN_ONCE(1, "GGTT view %u not implemented!\n",
3538 vma->ggtt_view.type);
3540 if (!vma->ggtt_view.pages) {
3541 DRM_ERROR("Failed to get pages for GGTT view type %u!\n",
3542 vma->ggtt_view.type);
3544 } else if (IS_ERR(vma->ggtt_view.pages)) {
3545 ret = PTR_ERR(vma->ggtt_view.pages);
3546 vma->ggtt_view.pages = NULL;
3547 DRM_ERROR("Failed to get pages for VMA view type %u (%d)!\n",
3548 vma->ggtt_view.type, ret);
3555 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
3557 * @cache_level: mapping cache level
3558 * @flags: flags like global or local mapping
3560 * DMA addresses are taken from the scatter-gather table of this object (or of
3561 * this VMA in case of non-default GGTT views) and PTE entries set up.
3562 * Note that DMA addresses are also the only part of the SG table we care about.
3564 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
3570 if (WARN_ON(flags == 0))
3574 if (flags & PIN_GLOBAL)
3575 bind_flags |= GLOBAL_BIND;
3576 if (flags & PIN_USER)
3577 bind_flags |= LOCAL_BIND;
3579 if (flags & PIN_UPDATE)
3580 bind_flags |= vma->bound;
3582 bind_flags &= ~vma->bound;
3584 if (bind_flags == 0)
3587 if (vma->bound == 0 && vma->vm->allocate_va_range) {
3588 /* XXX: i915_vma_pin() will fix this +- hack */
3590 trace_i915_va_alloc(vma);
3591 ret = vma->vm->allocate_va_range(vma->vm,
3599 ret = vma->vm->bind_vma(vma, cache_level, bind_flags);
3603 vma->bound |= bind_flags;
3609 * i915_ggtt_view_size - Get the size of a GGTT view.
3610 * @obj: Object the view is of.
3611 * @view: The view in question.
3613 * @return The size of the GGTT view in bytes.
3616 i915_ggtt_view_size(struct drm_i915_gem_object *obj,
3617 const struct i915_ggtt_view *view)
3619 if (view->type == I915_GGTT_VIEW_NORMAL) {
3620 return obj->base.size;
3621 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
3622 return intel_rotation_info_size(&view->params.rotated) << PAGE_SHIFT;
3623 } else if (view->type == I915_GGTT_VIEW_PARTIAL) {
3624 return view->params.partial.size << PAGE_SHIFT;
3626 WARN_ONCE(1, "GGTT view %u not implemented!\n", view->type);
3627 return obj->base.size;