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"
35 #include <sys/mplock2.h>
38 * DOC: Global GTT views
40 * Background and previous state
42 * Historically objects could exists (be bound) in global GTT space only as
43 * singular instances with a view representing all of the object's backing pages
44 * in a linear fashion. This view will be called a normal view.
46 * To support multiple views of the same object, where the number of mapped
47 * pages is not equal to the backing store, or where the layout of the pages
48 * is not linear, concept of a GGTT view was added.
50 * One example of an alternative view is a stereo display driven by a single
51 * image. In this case we would have a framebuffer looking like this
57 * Above would represent a normal GGTT view as normally mapped for GPU or CPU
58 * rendering. In contrast, fed to the display engine would be an alternative
59 * view which could look something like this:
64 * In this example both the size and layout of pages in the alternative view is
65 * different from the normal view.
67 * Implementation and usage
69 * GGTT views are implemented using VMAs and are distinguished via enum
70 * i915_ggtt_view_type and struct i915_ggtt_view.
72 * A new flavour of core GEM functions which work with GGTT bound objects were
73 * added with the _ggtt_ infix, and sometimes with _view postfix to avoid
74 * renaming in large amounts of code. They take the struct i915_ggtt_view
75 * parameter encapsulating all metadata required to implement a view.
77 * As a helper for callers which are only interested in the normal view,
78 * globally const i915_ggtt_view_normal singleton instance exists. All old core
79 * GEM API functions, the ones not taking the view parameter, are operating on,
80 * or with the normal GGTT view.
82 * Code wanting to add or use a new GGTT view needs to:
84 * 1. Add a new enum with a suitable name.
85 * 2. Extend the metadata in the i915_ggtt_view structure if required.
86 * 3. Add support to i915_get_vma_pages().
88 * New views are required to build a scatter-gather table from within the
89 * i915_get_vma_pages function. This table is stored in the vma.ggtt_view and
90 * exists for the lifetime of an VMA.
92 * Core API is designed to have copy semantics which means that passed in
93 * struct i915_ggtt_view does not need to be persistent (left around after
94 * calling the core API functions).
98 static inline struct i915_ggtt *
99 i915_vm_to_ggtt(struct i915_address_space *vm)
101 GEM_BUG_ON(!i915_is_ggtt(vm));
102 return container_of(vm, struct i915_ggtt, base);
106 i915_get_ggtt_vma_pages(struct i915_vma *vma);
108 const struct i915_ggtt_view i915_ggtt_view_normal = {
109 .type = I915_GGTT_VIEW_NORMAL,
111 const struct i915_ggtt_view i915_ggtt_view_rotated = {
112 .type = I915_GGTT_VIEW_ROTATED,
115 int intel_sanitize_enable_ppgtt(struct drm_i915_private *dev_priv,
118 bool has_aliasing_ppgtt;
120 bool has_full_48bit_ppgtt;
122 has_aliasing_ppgtt = INTEL_GEN(dev_priv) >= 6;
123 has_full_ppgtt = INTEL_GEN(dev_priv) >= 7;
124 has_full_48bit_ppgtt =
125 IS_BROADWELL(dev_priv) || INTEL_GEN(dev_priv) >= 9;
127 if (intel_vgpu_active(dev_priv)) {
128 /* emulation is too hard */
129 has_full_ppgtt = false;
130 has_full_48bit_ppgtt = false;
133 if (!has_aliasing_ppgtt)
137 * We don't allow disabling PPGTT for gen9+ as it's a requirement for
138 * execlists, the sole mechanism available to submit work.
140 if (enable_ppgtt == 0 && INTEL_GEN(dev_priv) < 9)
143 if (enable_ppgtt == 1)
146 if (enable_ppgtt == 2 && has_full_ppgtt)
149 if (enable_ppgtt == 3 && has_full_48bit_ppgtt)
152 #ifdef CONFIG_INTEL_IOMMU
153 /* Disable ppgtt on SNB if VT-d is on. */
154 if (IS_GEN6(dev_priv) && intel_iommu_gfx_mapped) {
155 DRM_INFO("Disabling PPGTT because VT-d is on\n");
160 /* Early VLV doesn't have this */
161 if (IS_VALLEYVIEW(dev_priv) && dev_priv->drm.pdev->revision < 0xb) {
162 DRM_DEBUG_DRIVER("disabling PPGTT on pre-B3 step VLV\n");
166 if (INTEL_GEN(dev_priv) >= 8 && i915.enable_execlists && has_full_ppgtt)
167 return has_full_48bit_ppgtt ? 3 : 2;
169 return has_aliasing_ppgtt ? 1 : 0;
172 static int ppgtt_bind_vma(struct i915_vma *vma,
173 enum i915_cache_level cache_level,
178 /* Currently applicable only to VLV */
180 pte_flags |= PTE_READ_ONLY;
182 vma->vm->insert_entries(vma->vm, vma->obj->pages, vma->node.start,
183 cache_level, pte_flags);
188 static void ppgtt_unbind_vma(struct i915_vma *vma)
190 vma->vm->clear_range(vma->vm,
196 static gen8_pte_t gen8_pte_encode(dma_addr_t addr,
197 enum i915_cache_level level,
200 gen8_pte_t pte = valid ? _PAGE_PRESENT | _PAGE_RW : 0;
204 case I915_CACHE_NONE:
205 pte |= PPAT_UNCACHED_INDEX;
208 pte |= PPAT_DISPLAY_ELLC_INDEX;
211 pte |= PPAT_CACHED_INDEX;
218 static gen8_pde_t gen8_pde_encode(const dma_addr_t addr,
219 const enum i915_cache_level level)
221 gen8_pde_t pde = _PAGE_PRESENT | _PAGE_RW;
223 if (level != I915_CACHE_NONE)
224 pde |= PPAT_CACHED_PDE_INDEX;
226 pde |= PPAT_UNCACHED_INDEX;
230 #define gen8_pdpe_encode gen8_pde_encode
231 #define gen8_pml4e_encode gen8_pde_encode
233 static gen6_pte_t snb_pte_encode(dma_addr_t addr,
234 enum i915_cache_level level,
235 bool valid, u32 unused)
237 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
238 pte |= GEN6_PTE_ADDR_ENCODE(addr);
241 case I915_CACHE_L3_LLC:
243 pte |= GEN6_PTE_CACHE_LLC;
245 case I915_CACHE_NONE:
246 pte |= GEN6_PTE_UNCACHED;
255 static gen6_pte_t ivb_pte_encode(dma_addr_t addr,
256 enum i915_cache_level level,
257 bool valid, u32 unused)
259 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
260 pte |= GEN6_PTE_ADDR_ENCODE(addr);
263 case I915_CACHE_L3_LLC:
264 pte |= GEN7_PTE_CACHE_L3_LLC;
267 pte |= GEN6_PTE_CACHE_LLC;
269 case I915_CACHE_NONE:
270 pte |= GEN6_PTE_UNCACHED;
279 static gen6_pte_t byt_pte_encode(dma_addr_t addr,
280 enum i915_cache_level level,
281 bool valid, u32 flags)
283 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
284 pte |= GEN6_PTE_ADDR_ENCODE(addr);
286 if (!(flags & PTE_READ_ONLY))
287 pte |= BYT_PTE_WRITEABLE;
289 if (level != I915_CACHE_NONE)
290 pte |= BYT_PTE_SNOOPED_BY_CPU_CACHES;
295 static gen6_pte_t hsw_pte_encode(dma_addr_t addr,
296 enum i915_cache_level level,
297 bool valid, u32 unused)
299 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
300 pte |= HSW_PTE_ADDR_ENCODE(addr);
302 if (level != I915_CACHE_NONE)
303 pte |= HSW_WB_LLC_AGE3;
308 static gen6_pte_t iris_pte_encode(dma_addr_t addr,
309 enum i915_cache_level level,
310 bool valid, u32 unused)
312 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
313 pte |= HSW_PTE_ADDR_ENCODE(addr);
316 case I915_CACHE_NONE:
319 pte |= HSW_WT_ELLC_LLC_AGE3;
322 pte |= HSW_WB_ELLC_LLC_AGE3;
329 static int __setup_page_dma(struct drm_device *dev,
330 struct i915_page_dma *p, gfp_t flags)
332 struct device *device = &dev->pdev->dev;
334 p->page = alloc_page(flags);
338 p->daddr = dma_map_page(device,
339 p->page, 0, 4096, PCI_DMA_BIDIRECTIONAL);
341 if (dma_mapping_error(device, p->daddr)) {
342 __free_page(p->page);
349 static int setup_page_dma(struct drm_device *dev, struct i915_page_dma *p)
351 return __setup_page_dma(dev, p, GFP_KERNEL);
354 static void cleanup_page_dma(struct drm_device *dev, struct i915_page_dma *p)
356 if (WARN_ON(!p->page))
359 dma_unmap_page(&dev->pdev->dev, p->daddr, 4096, PCI_DMA_BIDIRECTIONAL);
360 __free_page(p->page);
361 memset(p, 0, sizeof(*p));
364 static void *kmap_page_dma(struct i915_page_dma *p)
366 return kmap_atomic(p->page);
369 /* We use the flushing unmap only with ppgtt structures:
370 * page directories, page tables and scratch pages.
372 static void kunmap_page_dma(struct drm_device *dev, void *vaddr)
374 /* There are only few exceptions for gen >=6. chv and bxt.
375 * And we are not sure about the latter so play safe for now.
377 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
378 drm_clflush_virt_range(vaddr, PAGE_SIZE);
380 kunmap_atomic(vaddr);
383 #define kmap_px(px) kmap_page_dma(px_base(px))
384 #define kunmap_px(ppgtt, vaddr) kunmap_page_dma((ppgtt)->base.dev, (vaddr))
386 #define setup_px(dev, px) setup_page_dma((dev), px_base(px))
387 #define cleanup_px(dev, px) cleanup_page_dma((dev), px_base(px))
388 #define fill_px(dev, px, v) fill_page_dma((dev), px_base(px), (v))
389 #define fill32_px(dev, px, v) fill_page_dma_32((dev), px_base(px), (v))
391 static void fill_page_dma(struct drm_device *dev, struct i915_page_dma *p,
395 uint64_t * const vaddr = kmap_page_dma(p);
397 for (i = 0; i < 512; i++)
400 kunmap_page_dma(dev, vaddr);
403 static void fill_page_dma_32(struct drm_device *dev, struct i915_page_dma *p,
404 const uint32_t val32)
410 fill_page_dma(dev, p, v);
413 static struct i915_page_scratch *alloc_scratch_page(struct drm_device *dev)
415 struct i915_page_scratch *sp;
418 sp = kzalloc(sizeof(*sp), GFP_KERNEL);
420 return ERR_PTR(-ENOMEM);
422 ret = __setup_page_dma(dev, px_base(sp), GFP_DMA32 | __GFP_ZERO);
428 set_pages_uc(px_page(sp), 1);
433 static void free_scratch_page(struct drm_device *dev,
434 struct i915_page_scratch *sp)
436 set_pages_wb(px_page(sp), 1);
442 static struct i915_page_table *alloc_pt(struct drm_device *dev)
444 struct i915_page_table *pt;
445 const size_t count = INTEL_INFO(dev)->gen >= 8 ?
446 GEN8_PTES : GEN6_PTES;
449 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
451 return ERR_PTR(-ENOMEM);
453 pt->used_ptes = kcalloc(BITS_TO_LONGS(count), sizeof(*pt->used_ptes),
459 ret = setup_px(dev, pt);
466 kfree(pt->used_ptes);
473 static void free_pt(struct drm_device *dev, struct i915_page_table *pt)
476 kfree(pt->used_ptes);
480 static void gen8_initialize_pt(struct i915_address_space *vm,
481 struct i915_page_table *pt)
483 gen8_pte_t scratch_pte;
485 scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
486 I915_CACHE_LLC, true);
488 fill_px(vm->dev, pt, scratch_pte);
491 static void gen6_initialize_pt(struct i915_address_space *vm,
492 struct i915_page_table *pt)
494 gen6_pte_t scratch_pte;
496 WARN_ON(px_dma(vm->scratch_page) == 0);
498 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
499 I915_CACHE_LLC, true, 0);
501 fill32_px(vm->dev, pt, scratch_pte);
504 static struct i915_page_directory *alloc_pd(struct drm_device *dev)
506 struct i915_page_directory *pd;
509 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
511 return ERR_PTR(-ENOMEM);
513 pd->used_pdes = kcalloc(BITS_TO_LONGS(I915_PDES),
514 sizeof(*pd->used_pdes), GFP_KERNEL);
518 ret = setup_px(dev, pd);
525 kfree(pd->used_pdes);
532 static void free_pd(struct drm_device *dev, struct i915_page_directory *pd)
536 kfree(pd->used_pdes);
541 static void gen8_initialize_pd(struct i915_address_space *vm,
542 struct i915_page_directory *pd)
544 gen8_pde_t scratch_pde;
546 scratch_pde = gen8_pde_encode(px_dma(vm->scratch_pt), I915_CACHE_LLC);
548 fill_px(vm->dev, pd, scratch_pde);
551 static int __pdp_init(struct drm_device *dev,
552 struct i915_page_directory_pointer *pdp)
554 size_t pdpes = I915_PDPES_PER_PDP(dev);
556 pdp->used_pdpes = kcalloc(BITS_TO_LONGS(pdpes),
557 sizeof(unsigned long),
559 if (!pdp->used_pdpes)
562 pdp->page_directory = kcalloc(pdpes, sizeof(*pdp->page_directory),
564 if (!pdp->page_directory) {
565 kfree(pdp->used_pdpes);
566 /* the PDP might be the statically allocated top level. Keep it
567 * as clean as possible */
568 pdp->used_pdpes = NULL;
575 static void __pdp_fini(struct i915_page_directory_pointer *pdp)
577 kfree(pdp->used_pdpes);
578 kfree(pdp->page_directory);
579 pdp->page_directory = NULL;
583 i915_page_directory_pointer *alloc_pdp(struct drm_device *dev)
585 struct i915_page_directory_pointer *pdp;
588 WARN_ON(!USES_FULL_48BIT_PPGTT(dev));
590 pdp = kzalloc(sizeof(*pdp), GFP_KERNEL);
592 return ERR_PTR(-ENOMEM);
594 ret = __pdp_init(dev, pdp);
598 ret = setup_px(dev, pdp);
612 static void free_pdp(struct drm_device *dev,
613 struct i915_page_directory_pointer *pdp)
616 if (USES_FULL_48BIT_PPGTT(dev)) {
617 cleanup_px(dev, pdp);
622 static void gen8_initialize_pdp(struct i915_address_space *vm,
623 struct i915_page_directory_pointer *pdp)
625 gen8_ppgtt_pdpe_t scratch_pdpe;
627 scratch_pdpe = gen8_pdpe_encode(px_dma(vm->scratch_pd), I915_CACHE_LLC);
629 fill_px(vm->dev, pdp, scratch_pdpe);
632 static void gen8_initialize_pml4(struct i915_address_space *vm,
633 struct i915_pml4 *pml4)
635 gen8_ppgtt_pml4e_t scratch_pml4e;
637 scratch_pml4e = gen8_pml4e_encode(px_dma(vm->scratch_pdp),
640 fill_px(vm->dev, pml4, scratch_pml4e);
644 gen8_setup_page_directory(struct i915_hw_ppgtt *ppgtt,
645 struct i915_page_directory_pointer *pdp,
646 struct i915_page_directory *pd,
649 gen8_ppgtt_pdpe_t *page_directorypo;
651 if (!USES_FULL_48BIT_PPGTT(ppgtt->base.dev))
654 page_directorypo = kmap_px(pdp);
655 page_directorypo[index] = gen8_pdpe_encode(px_dma(pd), I915_CACHE_LLC);
656 kunmap_px(ppgtt, page_directorypo);
660 gen8_setup_page_directory_pointer(struct i915_hw_ppgtt *ppgtt,
661 struct i915_pml4 *pml4,
662 struct i915_page_directory_pointer *pdp,
665 gen8_ppgtt_pml4e_t *pagemap = kmap_px(pml4);
667 WARN_ON(!USES_FULL_48BIT_PPGTT(ppgtt->base.dev));
668 pagemap[index] = gen8_pml4e_encode(px_dma(pdp), I915_CACHE_LLC);
669 kunmap_px(ppgtt, pagemap);
672 /* Broadwell Page Directory Pointer Descriptors */
673 static int gen8_write_pdp(struct drm_i915_gem_request *req,
677 struct intel_engine_cs *engine = req->engine;
682 ret = intel_ring_begin(req, 6);
686 intel_ring_emit(engine, MI_LOAD_REGISTER_IMM(1));
687 intel_ring_emit_reg(engine, GEN8_RING_PDP_UDW(engine, entry));
688 intel_ring_emit(engine, upper_32_bits(addr));
689 intel_ring_emit(engine, MI_LOAD_REGISTER_IMM(1));
690 intel_ring_emit_reg(engine, GEN8_RING_PDP_LDW(engine, entry));
691 intel_ring_emit(engine, lower_32_bits(addr));
692 intel_ring_advance(engine);
697 static int gen8_legacy_mm_switch(struct i915_hw_ppgtt *ppgtt,
698 struct drm_i915_gem_request *req)
702 for (i = GEN8_LEGACY_PDPES - 1; i >= 0; i--) {
703 const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i);
705 ret = gen8_write_pdp(req, i, pd_daddr);
713 static int gen8_48b_mm_switch(struct i915_hw_ppgtt *ppgtt,
714 struct drm_i915_gem_request *req)
716 return gen8_write_pdp(req, 0, px_dma(&ppgtt->pml4));
719 static void gen8_ppgtt_clear_pte_range(struct i915_address_space *vm,
720 struct i915_page_directory_pointer *pdp,
723 gen8_pte_t scratch_pte)
725 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
726 gen8_pte_t *pt_vaddr;
727 unsigned pdpe = gen8_pdpe_index(start);
728 unsigned pde = gen8_pde_index(start);
729 unsigned pte = gen8_pte_index(start);
730 unsigned num_entries = length >> PAGE_SHIFT;
731 unsigned last_pte, i;
736 while (num_entries) {
737 struct i915_page_directory *pd;
738 struct i915_page_table *pt;
740 if (WARN_ON(!pdp->page_directory[pdpe]))
743 pd = pdp->page_directory[pdpe];
745 if (WARN_ON(!pd->page_table[pde]))
748 pt = pd->page_table[pde];
750 if (WARN_ON(!px_page(pt)))
753 last_pte = pte + num_entries;
754 if (last_pte > GEN8_PTES)
755 last_pte = GEN8_PTES;
757 pt_vaddr = kmap_px(pt);
759 for (i = pte; i < last_pte; i++) {
760 pt_vaddr[i] = scratch_pte;
764 kunmap_px(ppgtt, pt_vaddr);
767 if (++pde == I915_PDES) {
768 if (++pdpe == I915_PDPES_PER_PDP(vm->dev))
775 static void gen8_ppgtt_clear_range(struct i915_address_space *vm,
780 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
781 gen8_pte_t scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
782 I915_CACHE_LLC, use_scratch);
784 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
785 gen8_ppgtt_clear_pte_range(vm, &ppgtt->pdp, start, length,
789 struct i915_page_directory_pointer *pdp;
791 gen8_for_each_pml4e(pdp, &ppgtt->pml4, start, length, pml4e) {
792 gen8_ppgtt_clear_pte_range(vm, pdp, start, length,
799 gen8_ppgtt_insert_pte_entries(struct i915_address_space *vm,
800 struct i915_page_directory_pointer *pdp,
801 struct sg_page_iter *sg_iter,
803 enum i915_cache_level cache_level)
805 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
806 gen8_pte_t *pt_vaddr;
807 unsigned pdpe = gen8_pdpe_index(start);
808 unsigned pde = gen8_pde_index(start);
809 unsigned pte = gen8_pte_index(start);
813 while (__sg_page_iter_next(sg_iter)) {
814 if (pt_vaddr == NULL) {
815 struct i915_page_directory *pd = pdp->page_directory[pdpe];
816 struct i915_page_table *pt = pd->page_table[pde];
817 pt_vaddr = kmap_px(pt);
821 gen8_pte_encode(sg_page_iter_dma_address(sg_iter),
823 if (++pte == GEN8_PTES) {
824 kunmap_px(ppgtt, pt_vaddr);
826 if (++pde == I915_PDES) {
827 if (++pdpe == I915_PDPES_PER_PDP(vm->dev))
836 kunmap_px(ppgtt, pt_vaddr);
839 static void gen8_ppgtt_insert_entries(struct i915_address_space *vm,
840 struct sg_table *pages,
842 enum i915_cache_level cache_level,
845 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
846 struct sg_page_iter sg_iter;
848 __sg_page_iter_start(&sg_iter, pages->sgl, sg_nents(pages->sgl), 0);
850 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
851 gen8_ppgtt_insert_pte_entries(vm, &ppgtt->pdp, &sg_iter, start,
854 struct i915_page_directory_pointer *pdp;
856 uint64_t length = (uint64_t)pages->orig_nents << PAGE_SHIFT;
858 gen8_for_each_pml4e(pdp, &ppgtt->pml4, start, length, pml4e) {
859 gen8_ppgtt_insert_pte_entries(vm, pdp, &sg_iter,
865 static void gen8_free_page_tables(struct drm_device *dev,
866 struct i915_page_directory *pd)
873 for_each_set_bit(i, pd->used_pdes, I915_PDES) {
874 if (WARN_ON(!pd->page_table[i]))
877 free_pt(dev, pd->page_table[i]);
878 pd->page_table[i] = NULL;
882 static int gen8_init_scratch(struct i915_address_space *vm)
884 struct drm_device *dev = vm->dev;
887 vm->scratch_page = alloc_scratch_page(dev);
888 if (IS_ERR(vm->scratch_page))
889 return PTR_ERR(vm->scratch_page);
891 vm->scratch_pt = alloc_pt(dev);
892 if (IS_ERR(vm->scratch_pt)) {
893 ret = PTR_ERR(vm->scratch_pt);
894 goto free_scratch_page;
897 vm->scratch_pd = alloc_pd(dev);
898 if (IS_ERR(vm->scratch_pd)) {
899 ret = PTR_ERR(vm->scratch_pd);
903 if (USES_FULL_48BIT_PPGTT(dev)) {
904 vm->scratch_pdp = alloc_pdp(dev);
905 if (IS_ERR(vm->scratch_pdp)) {
906 ret = PTR_ERR(vm->scratch_pdp);
911 gen8_initialize_pt(vm, vm->scratch_pt);
912 gen8_initialize_pd(vm, vm->scratch_pd);
913 if (USES_FULL_48BIT_PPGTT(dev))
914 gen8_initialize_pdp(vm, vm->scratch_pdp);
919 free_pd(dev, vm->scratch_pd);
921 free_pt(dev, vm->scratch_pt);
923 free_scratch_page(dev, vm->scratch_page);
928 static int gen8_ppgtt_notify_vgt(struct i915_hw_ppgtt *ppgtt, bool create)
930 enum vgt_g2v_type msg;
931 struct drm_i915_private *dev_priv = to_i915(ppgtt->base.dev);
934 if (USES_FULL_48BIT_PPGTT(dev_priv)) {
935 u64 daddr = px_dma(&ppgtt->pml4);
937 I915_WRITE(vgtif_reg(pdp[0].lo), lower_32_bits(daddr));
938 I915_WRITE(vgtif_reg(pdp[0].hi), upper_32_bits(daddr));
940 msg = (create ? VGT_G2V_PPGTT_L4_PAGE_TABLE_CREATE :
941 VGT_G2V_PPGTT_L4_PAGE_TABLE_DESTROY);
943 for (i = 0; i < GEN8_LEGACY_PDPES; i++) {
944 u64 daddr = i915_page_dir_dma_addr(ppgtt, i);
946 I915_WRITE(vgtif_reg(pdp[i].lo), lower_32_bits(daddr));
947 I915_WRITE(vgtif_reg(pdp[i].hi), upper_32_bits(daddr));
950 msg = (create ? VGT_G2V_PPGTT_L3_PAGE_TABLE_CREATE :
951 VGT_G2V_PPGTT_L3_PAGE_TABLE_DESTROY);
954 I915_WRITE(vgtif_reg(g2v_notify), msg);
959 static void gen8_free_scratch(struct i915_address_space *vm)
961 struct drm_device *dev = vm->dev;
963 if (USES_FULL_48BIT_PPGTT(dev))
964 free_pdp(dev, vm->scratch_pdp);
965 free_pd(dev, vm->scratch_pd);
966 free_pt(dev, vm->scratch_pt);
967 free_scratch_page(dev, vm->scratch_page);
970 static void gen8_ppgtt_cleanup_3lvl(struct drm_device *dev,
971 struct i915_page_directory_pointer *pdp)
975 for_each_set_bit(i, pdp->used_pdpes, I915_PDPES_PER_PDP(dev)) {
976 if (WARN_ON(!pdp->page_directory[i]))
979 gen8_free_page_tables(dev, pdp->page_directory[i]);
980 free_pd(dev, pdp->page_directory[i]);
986 static void gen8_ppgtt_cleanup_4lvl(struct i915_hw_ppgtt *ppgtt)
990 for_each_set_bit(i, ppgtt->pml4.used_pml4es, GEN8_PML4ES_PER_PML4) {
991 if (WARN_ON(!ppgtt->pml4.pdps[i]))
994 gen8_ppgtt_cleanup_3lvl(ppgtt->base.dev, ppgtt->pml4.pdps[i]);
997 cleanup_px(ppgtt->base.dev, &ppgtt->pml4);
1000 static void gen8_ppgtt_cleanup(struct i915_address_space *vm)
1002 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1004 if (intel_vgpu_active(to_i915(vm->dev)))
1005 gen8_ppgtt_notify_vgt(ppgtt, false);
1007 if (!USES_FULL_48BIT_PPGTT(ppgtt->base.dev))
1008 gen8_ppgtt_cleanup_3lvl(ppgtt->base.dev, &ppgtt->pdp);
1010 gen8_ppgtt_cleanup_4lvl(ppgtt);
1012 gen8_free_scratch(vm);
1016 * gen8_ppgtt_alloc_pagetabs() - Allocate page tables for VA range.
1017 * @vm: Master vm structure.
1018 * @pd: Page directory for this address range.
1019 * @start: Starting virtual address to begin allocations.
1020 * @length: Size of the allocations.
1021 * @new_pts: Bitmap set by function with new allocations. Likely used by the
1022 * caller to free on error.
1024 * Allocate the required number of page tables. Extremely similar to
1025 * gen8_ppgtt_alloc_page_directories(). The main difference is here we are limited by
1026 * the page directory boundary (instead of the page directory pointer). That
1027 * boundary is 1GB virtual. Therefore, unlike gen8_ppgtt_alloc_page_directories(), it is
1028 * possible, and likely that the caller will need to use multiple calls of this
1029 * function to achieve the appropriate allocation.
1031 * Return: 0 if success; negative error code otherwise.
1033 static int gen8_ppgtt_alloc_pagetabs(struct i915_address_space *vm,
1034 struct i915_page_directory *pd,
1037 unsigned long *new_pts)
1039 struct drm_device *dev = vm->dev;
1040 struct i915_page_table *pt;
1043 gen8_for_each_pde(pt, pd, start, length, pde) {
1044 /* Don't reallocate page tables */
1045 if (test_bit(pde, pd->used_pdes)) {
1046 /* Scratch is never allocated this way */
1047 WARN_ON(pt == vm->scratch_pt);
1055 gen8_initialize_pt(vm, pt);
1056 pd->page_table[pde] = pt;
1057 __set_bit(pde, new_pts);
1058 trace_i915_page_table_entry_alloc(vm, pde, start, GEN8_PDE_SHIFT);
1064 for_each_set_bit(pde, new_pts, I915_PDES)
1065 free_pt(dev, pd->page_table[pde]);
1071 * gen8_ppgtt_alloc_page_directories() - Allocate page directories for VA range.
1072 * @vm: Master vm structure.
1073 * @pdp: Page directory pointer for this address range.
1074 * @start: Starting virtual address to begin allocations.
1075 * @length: Size of the allocations.
1076 * @new_pds: Bitmap set by function with new allocations. Likely used by the
1077 * caller to free on error.
1079 * Allocate the required number of page directories starting at the pde index of
1080 * @start, and ending at the pde index @start + @length. This function will skip
1081 * over already allocated page directories within the range, and only allocate
1082 * new ones, setting the appropriate pointer within the pdp as well as the
1083 * correct position in the bitmap @new_pds.
1085 * The function will only allocate the pages within the range for a give page
1086 * directory pointer. In other words, if @start + @length straddles a virtually
1087 * addressed PDP boundary (512GB for 4k pages), there will be more allocations
1088 * required by the caller, This is not currently possible, and the BUG in the
1089 * code will prevent it.
1091 * Return: 0 if success; negative error code otherwise.
1094 gen8_ppgtt_alloc_page_directories(struct i915_address_space *vm,
1095 struct i915_page_directory_pointer *pdp,
1098 unsigned long *new_pds)
1100 struct drm_device *dev = vm->dev;
1101 struct i915_page_directory *pd;
1103 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1105 WARN_ON(!bitmap_empty(new_pds, pdpes));
1107 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1108 if (test_bit(pdpe, pdp->used_pdpes))
1115 gen8_initialize_pd(vm, pd);
1116 pdp->page_directory[pdpe] = pd;
1117 __set_bit(pdpe, new_pds);
1118 trace_i915_page_directory_entry_alloc(vm, pdpe, start, GEN8_PDPE_SHIFT);
1124 for_each_set_bit(pdpe, new_pds, pdpes)
1125 free_pd(dev, pdp->page_directory[pdpe]);
1131 * gen8_ppgtt_alloc_page_dirpointers() - Allocate pdps for VA range.
1132 * @vm: Master vm structure.
1133 * @pml4: Page map level 4 for this address range.
1134 * @start: Starting virtual address to begin allocations.
1135 * @length: Size of the allocations.
1136 * @new_pdps: Bitmap set by function with new allocations. Likely used by the
1137 * caller to free on error.
1139 * Allocate the required number of page directory pointers. Extremely similar to
1140 * gen8_ppgtt_alloc_page_directories() and gen8_ppgtt_alloc_pagetabs().
1141 * The main difference is here we are limited by the pml4 boundary (instead of
1142 * the page directory pointer).
1144 * Return: 0 if success; negative error code otherwise.
1147 gen8_ppgtt_alloc_page_dirpointers(struct i915_address_space *vm,
1148 struct i915_pml4 *pml4,
1151 unsigned long *new_pdps)
1153 struct drm_device *dev = vm->dev;
1154 struct i915_page_directory_pointer *pdp;
1157 WARN_ON(!bitmap_empty(new_pdps, GEN8_PML4ES_PER_PML4));
1159 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1160 if (!test_bit(pml4e, pml4->used_pml4es)) {
1161 pdp = alloc_pdp(dev);
1165 gen8_initialize_pdp(vm, pdp);
1166 pml4->pdps[pml4e] = pdp;
1167 __set_bit(pml4e, new_pdps);
1168 trace_i915_page_directory_pointer_entry_alloc(vm,
1178 for_each_set_bit(pml4e, new_pdps, GEN8_PML4ES_PER_PML4)
1179 free_pdp(dev, pml4->pdps[pml4e]);
1185 free_gen8_temp_bitmaps(unsigned long *new_pds, unsigned long *new_pts)
1191 /* Fills in the page directory bitmap, and the array of page tables bitmap. Both
1192 * of these are based on the number of PDPEs in the system.
1195 int __must_check alloc_gen8_temp_bitmaps(unsigned long **new_pds,
1196 unsigned long **new_pts,
1202 pds = kcalloc(BITS_TO_LONGS(pdpes), sizeof(unsigned long), GFP_TEMPORARY);
1206 pts = kcalloc(pdpes, BITS_TO_LONGS(I915_PDES) * sizeof(unsigned long),
1217 free_gen8_temp_bitmaps(pds, pts);
1221 /* PDE TLBs are a pain to invalidate on GEN8+. When we modify
1222 * the page table structures, we mark them dirty so that
1223 * context switching/execlist queuing code takes extra steps
1224 * to ensure that tlbs are flushed.
1226 static void mark_tlbs_dirty(struct i915_hw_ppgtt *ppgtt)
1228 ppgtt->pd_dirty_rings = INTEL_INFO(ppgtt->base.dev)->ring_mask;
1231 static int gen8_alloc_va_range_3lvl(struct i915_address_space *vm,
1232 struct i915_page_directory_pointer *pdp,
1236 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1237 unsigned long *new_page_dirs, *new_page_tables;
1238 struct drm_device *dev = vm->dev;
1239 struct i915_page_directory *pd;
1240 const uint64_t orig_start = start;
1241 const uint64_t orig_length = length;
1243 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1246 /* Wrap is never okay since we can only represent 48b, and we don't
1247 * actually use the other side of the canonical address space.
1249 if (WARN_ON(start + length < start))
1252 if (WARN_ON(start + length > vm->total))
1255 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables, pdpes);
1259 /* Do the allocations first so we can easily bail out */
1260 ret = gen8_ppgtt_alloc_page_directories(vm, pdp, start, length,
1263 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1267 /* For every page directory referenced, allocate page tables */
1268 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1269 ret = gen8_ppgtt_alloc_pagetabs(vm, pd, start, length,
1270 new_page_tables + pdpe * BITS_TO_LONGS(I915_PDES));
1276 length = orig_length;
1278 /* Allocations have completed successfully, so set the bitmaps, and do
1280 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1281 gen8_pde_t *const page_directory = kmap_px(pd);
1282 struct i915_page_table *pt;
1283 uint64_t pd_len = length;
1284 uint64_t pd_start = start;
1287 /* Every pd should be allocated, we just did that above. */
1290 gen8_for_each_pde(pt, pd, pd_start, pd_len, pde) {
1291 /* Same reasoning as pd */
1294 WARN_ON(!gen8_pte_count(pd_start, pd_len));
1296 /* Set our used ptes within the page table */
1297 bitmap_set(pt->used_ptes,
1298 gen8_pte_index(pd_start),
1299 gen8_pte_count(pd_start, pd_len));
1301 /* Our pde is now pointing to the pagetable, pt */
1302 __set_bit(pde, pd->used_pdes);
1304 /* Map the PDE to the page table */
1305 page_directory[pde] = gen8_pde_encode(px_dma(pt),
1307 trace_i915_page_table_entry_map(&ppgtt->base, pde, pt,
1308 gen8_pte_index(start),
1309 gen8_pte_count(start, length),
1312 /* NB: We haven't yet mapped ptes to pages. At this
1313 * point we're still relying on insert_entries() */
1316 kunmap_px(ppgtt, page_directory);
1317 __set_bit(pdpe, pdp->used_pdpes);
1318 gen8_setup_page_directory(ppgtt, pdp, pd, pdpe);
1321 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1322 mark_tlbs_dirty(ppgtt);
1329 for_each_set_bit(temp, new_page_tables + pdpe *
1330 BITS_TO_LONGS(I915_PDES), I915_PDES)
1331 free_pt(dev, pdp->page_directory[pdpe]->page_table[temp]);
1334 for_each_set_bit(pdpe, new_page_dirs, pdpes)
1335 free_pd(dev, pdp->page_directory[pdpe]);
1337 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1338 mark_tlbs_dirty(ppgtt);
1342 static int gen8_alloc_va_range_4lvl(struct i915_address_space *vm,
1343 struct i915_pml4 *pml4,
1347 DECLARE_BITMAP(new_pdps, GEN8_PML4ES_PER_PML4);
1348 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1349 struct i915_page_directory_pointer *pdp;
1353 /* Do the pml4 allocations first, so we don't need to track the newly
1354 * allocated tables below the pdp */
1355 bitmap_zero(new_pdps, GEN8_PML4ES_PER_PML4);
1357 /* The pagedirectory and pagetable allocations are done in the shared 3
1358 * and 4 level code. Just allocate the pdps.
1360 ret = gen8_ppgtt_alloc_page_dirpointers(vm, pml4, start, length,
1365 WARN(bitmap_weight(new_pdps, GEN8_PML4ES_PER_PML4) > 2,
1366 "The allocation has spanned more than 512GB. "
1367 "It is highly likely this is incorrect.");
1369 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1372 ret = gen8_alloc_va_range_3lvl(vm, pdp, start, length);
1376 gen8_setup_page_directory_pointer(ppgtt, pml4, pdp, pml4e);
1379 bitmap_or(pml4->used_pml4es, new_pdps, pml4->used_pml4es,
1380 GEN8_PML4ES_PER_PML4);
1385 for_each_set_bit(pml4e, new_pdps, GEN8_PML4ES_PER_PML4)
1386 gen8_ppgtt_cleanup_3lvl(vm->dev, pml4->pdps[pml4e]);
1391 static int gen8_alloc_va_range(struct i915_address_space *vm,
1392 uint64_t start, uint64_t length)
1394 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1396 if (USES_FULL_48BIT_PPGTT(vm->dev))
1397 return gen8_alloc_va_range_4lvl(vm, &ppgtt->pml4, start, length);
1399 return gen8_alloc_va_range_3lvl(vm, &ppgtt->pdp, start, length);
1402 static void gen8_dump_pdp(struct i915_page_directory_pointer *pdp,
1403 uint64_t start, uint64_t length,
1404 gen8_pte_t scratch_pte,
1407 struct i915_page_directory *pd;
1410 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1411 struct i915_page_table *pt;
1412 uint64_t pd_len = length;
1413 uint64_t pd_start = start;
1416 if (!test_bit(pdpe, pdp->used_pdpes))
1419 seq_printf(m, "\tPDPE #%d\n", pdpe);
1420 gen8_for_each_pde(pt, pd, pd_start, pd_len, pde) {
1422 gen8_pte_t *pt_vaddr;
1424 if (!test_bit(pde, pd->used_pdes))
1427 pt_vaddr = kmap_px(pt);
1428 for (pte = 0; pte < GEN8_PTES; pte += 4) {
1430 (pdpe << GEN8_PDPE_SHIFT) |
1431 (pde << GEN8_PDE_SHIFT) |
1432 (pte << GEN8_PTE_SHIFT);
1436 for (i = 0; i < 4; i++)
1437 if (pt_vaddr[pte + i] != scratch_pte)
1442 seq_printf(m, "\t\t0x%lx [%03d,%03d,%04d]: =", va, pdpe, pde, pte);
1443 for (i = 0; i < 4; i++) {
1444 if (pt_vaddr[pte + i] != scratch_pte)
1445 seq_printf(m, " %lx", pt_vaddr[pte + i]);
1447 seq_puts(m, " SCRATCH ");
1451 /* don't use kunmap_px, it could trigger
1452 * an unnecessary flush.
1454 kunmap_atomic(pt_vaddr);
1459 static void gen8_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
1461 struct i915_address_space *vm = &ppgtt->base;
1462 uint64_t start = ppgtt->base.start;
1463 uint64_t length = ppgtt->base.total;
1464 gen8_pte_t scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
1465 I915_CACHE_LLC, true);
1467 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
1468 gen8_dump_pdp(&ppgtt->pdp, start, length, scratch_pte, m);
1471 struct i915_pml4 *pml4 = &ppgtt->pml4;
1472 struct i915_page_directory_pointer *pdp;
1474 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1475 if (!test_bit(pml4e, pml4->used_pml4es))
1478 seq_printf(m, " PML4E #%lu\n", pml4e);
1479 gen8_dump_pdp(pdp, start, length, scratch_pte, m);
1484 static int gen8_preallocate_top_level_pdps(struct i915_hw_ppgtt *ppgtt)
1486 unsigned long *new_page_dirs, *new_page_tables;
1487 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1490 /* We allocate temp bitmap for page tables for no gain
1491 * but as this is for init only, lets keep the things simple
1493 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables, pdpes);
1497 /* Allocate for all pdps regardless of how the ppgtt
1500 ret = gen8_ppgtt_alloc_page_directories(&ppgtt->base, &ppgtt->pdp,
1504 *ppgtt->pdp.used_pdpes = *new_page_dirs;
1506 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1512 * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers
1513 * with a net effect resembling a 2-level page table in normal x86 terms. Each
1514 * PDP represents 1GB of memory 4 * 512 * 512 * 4096 = 4GB legacy 32b address
1518 static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
1522 ret = gen8_init_scratch(&ppgtt->base);
1526 ppgtt->base.start = 0;
1527 ppgtt->base.cleanup = gen8_ppgtt_cleanup;
1528 ppgtt->base.allocate_va_range = gen8_alloc_va_range;
1529 ppgtt->base.insert_entries = gen8_ppgtt_insert_entries;
1530 ppgtt->base.clear_range = gen8_ppgtt_clear_range;
1531 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
1532 ppgtt->base.bind_vma = ppgtt_bind_vma;
1533 ppgtt->debug_dump = gen8_dump_ppgtt;
1535 if (USES_FULL_48BIT_PPGTT(ppgtt->base.dev)) {
1536 ret = setup_px(ppgtt->base.dev, &ppgtt->pml4);
1540 gen8_initialize_pml4(&ppgtt->base, &ppgtt->pml4);
1542 ppgtt->base.total = 1ULL << 48;
1543 ppgtt->switch_mm = gen8_48b_mm_switch;
1545 ret = __pdp_init(ppgtt->base.dev, &ppgtt->pdp);
1549 ppgtt->base.total = 1ULL << 32;
1550 ppgtt->switch_mm = gen8_legacy_mm_switch;
1551 trace_i915_page_directory_pointer_entry_alloc(&ppgtt->base,
1555 if (intel_vgpu_active(to_i915(ppgtt->base.dev))) {
1556 ret = gen8_preallocate_top_level_pdps(ppgtt);
1562 if (intel_vgpu_active(to_i915(ppgtt->base.dev)))
1563 gen8_ppgtt_notify_vgt(ppgtt, true);
1568 gen8_free_scratch(&ppgtt->base);
1572 static void gen6_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
1574 struct i915_address_space *vm = &ppgtt->base;
1575 struct i915_page_table *unused;
1576 gen6_pte_t scratch_pte;
1579 uint32_t start = ppgtt->base.start, length = ppgtt->base.total;
1581 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1582 I915_CACHE_LLC, true, 0);
1584 gen6_for_each_pde(unused, &ppgtt->pd, start, length, pde) {
1586 gen6_pte_t *pt_vaddr;
1587 const dma_addr_t pt_addr = px_dma(ppgtt->pd.page_table[pde]);
1588 pd_entry = readl(ppgtt->pd_addr + pde);
1589 expected = (GEN6_PDE_ADDR_ENCODE(pt_addr) | GEN6_PDE_VALID);
1591 if (pd_entry != expected)
1592 seq_printf(m, "\tPDE #%d mismatch: Actual PDE: %x Expected PDE: %x\n",
1596 seq_printf(m, "\tPDE: %x\n", pd_entry);
1598 pt_vaddr = kmap_px(ppgtt->pd.page_table[pde]);
1600 for (pte = 0; pte < GEN6_PTES; pte+=4) {
1602 (pde * PAGE_SIZE * GEN6_PTES) +
1606 for (i = 0; i < 4; i++)
1607 if (pt_vaddr[pte + i] != scratch_pte)
1612 seq_printf(m, "\t\t0x%lx [%03d,%04d]: =", va, pde, pte);
1613 for (i = 0; i < 4; i++) {
1614 if (pt_vaddr[pte + i] != scratch_pte)
1615 seq_printf(m, " %08x", pt_vaddr[pte + i]);
1617 seq_puts(m, " SCRATCH ");
1621 kunmap_px(ppgtt, pt_vaddr);
1625 /* Write pde (index) from the page directory @pd to the page table @pt */
1626 static void gen6_write_pde(struct i915_page_directory *pd,
1627 const int pde, struct i915_page_table *pt)
1629 /* Caller needs to make sure the write completes if necessary */
1630 struct i915_hw_ppgtt *ppgtt =
1631 container_of(pd, struct i915_hw_ppgtt, pd);
1634 pd_entry = GEN6_PDE_ADDR_ENCODE(px_dma(pt));
1635 pd_entry |= GEN6_PDE_VALID;
1637 writel(pd_entry, ppgtt->pd_addr + pde);
1640 /* Write all the page tables found in the ppgtt structure to incrementing page
1642 static void gen6_write_page_range(struct drm_i915_private *dev_priv,
1643 struct i915_page_directory *pd,
1644 uint32_t start, uint32_t length)
1646 struct i915_ggtt *ggtt = &dev_priv->ggtt;
1647 struct i915_page_table *pt;
1650 gen6_for_each_pde(pt, pd, start, length, pde)
1651 gen6_write_pde(pd, pde, pt);
1653 /* Make sure write is complete before other code can use this page
1654 * table. Also require for WC mapped PTEs */
1658 static uint32_t get_pd_offset(struct i915_hw_ppgtt *ppgtt)
1660 BUG_ON(ppgtt->pd.base.ggtt_offset & 0x3f);
1662 return (ppgtt->pd.base.ggtt_offset / 64) << 16;
1665 static int hsw_mm_switch(struct i915_hw_ppgtt *ppgtt,
1666 struct drm_i915_gem_request *req)
1668 struct intel_engine_cs *engine = req->engine;
1671 /* NB: TLBs must be flushed and invalidated before a switch */
1672 ret = engine->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1676 ret = intel_ring_begin(req, 6);
1680 intel_ring_emit(engine, MI_LOAD_REGISTER_IMM(2));
1681 intel_ring_emit_reg(engine, RING_PP_DIR_DCLV(engine));
1682 intel_ring_emit(engine, PP_DIR_DCLV_2G);
1683 intel_ring_emit_reg(engine, RING_PP_DIR_BASE(engine));
1684 intel_ring_emit(engine, get_pd_offset(ppgtt));
1685 intel_ring_emit(engine, MI_NOOP);
1686 intel_ring_advance(engine);
1691 static int gen7_mm_switch(struct i915_hw_ppgtt *ppgtt,
1692 struct drm_i915_gem_request *req)
1694 struct intel_engine_cs *engine = req->engine;
1697 /* NB: TLBs must be flushed and invalidated before a switch */
1698 ret = engine->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1702 ret = intel_ring_begin(req, 6);
1706 intel_ring_emit(engine, MI_LOAD_REGISTER_IMM(2));
1707 intel_ring_emit_reg(engine, RING_PP_DIR_DCLV(engine));
1708 intel_ring_emit(engine, PP_DIR_DCLV_2G);
1709 intel_ring_emit_reg(engine, RING_PP_DIR_BASE(engine));
1710 intel_ring_emit(engine, get_pd_offset(ppgtt));
1711 intel_ring_emit(engine, MI_NOOP);
1712 intel_ring_advance(engine);
1714 /* XXX: RCS is the only one to auto invalidate the TLBs? */
1715 if (engine->id != RCS) {
1716 ret = engine->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1724 static int gen6_mm_switch(struct i915_hw_ppgtt *ppgtt,
1725 struct drm_i915_gem_request *req)
1727 struct intel_engine_cs *engine = req->engine;
1728 struct drm_i915_private *dev_priv = req->i915;
1730 I915_WRITE(RING_PP_DIR_DCLV(engine), PP_DIR_DCLV_2G);
1731 I915_WRITE(RING_PP_DIR_BASE(engine), get_pd_offset(ppgtt));
1735 static void gen8_ppgtt_enable(struct drm_device *dev)
1737 struct drm_i915_private *dev_priv = to_i915(dev);
1738 struct intel_engine_cs *engine;
1740 for_each_engine(engine, dev_priv) {
1741 u32 four_level = USES_FULL_48BIT_PPGTT(dev) ? GEN8_GFX_PPGTT_48B : 0;
1742 I915_WRITE(RING_MODE_GEN7(engine),
1743 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE | four_level));
1747 static void gen7_ppgtt_enable(struct drm_device *dev)
1749 struct drm_i915_private *dev_priv = to_i915(dev);
1750 struct intel_engine_cs *engine;
1751 uint32_t ecochk, ecobits;
1753 ecobits = I915_READ(GAC_ECO_BITS);
1754 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B);
1756 ecochk = I915_READ(GAM_ECOCHK);
1757 if (IS_HASWELL(dev)) {
1758 ecochk |= ECOCHK_PPGTT_WB_HSW;
1760 ecochk |= ECOCHK_PPGTT_LLC_IVB;
1761 ecochk &= ~ECOCHK_PPGTT_GFDT_IVB;
1763 I915_WRITE(GAM_ECOCHK, ecochk);
1765 for_each_engine(engine, dev_priv) {
1766 /* GFX_MODE is per-ring on gen7+ */
1767 I915_WRITE(RING_MODE_GEN7(engine),
1768 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1772 static void gen6_ppgtt_enable(struct drm_device *dev)
1774 struct drm_i915_private *dev_priv = to_i915(dev);
1775 uint32_t ecochk, gab_ctl, ecobits;
1777 ecobits = I915_READ(GAC_ECO_BITS);
1778 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_SNB_BIT |
1779 ECOBITS_PPGTT_CACHE64B);
1781 gab_ctl = I915_READ(GAB_CTL);
1782 I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT);
1784 ecochk = I915_READ(GAM_ECOCHK);
1785 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT | ECOCHK_PPGTT_CACHE64B);
1787 I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1790 /* PPGTT support for Sandybdrige/Gen6 and later */
1791 static void gen6_ppgtt_clear_range(struct i915_address_space *vm,
1796 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1797 gen6_pte_t *pt_vaddr, scratch_pte;
1798 unsigned first_entry = start >> PAGE_SHIFT;
1799 unsigned num_entries = length >> PAGE_SHIFT;
1800 unsigned act_pt = first_entry / GEN6_PTES;
1801 unsigned first_pte = first_entry % GEN6_PTES;
1802 unsigned last_pte, i;
1804 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1805 I915_CACHE_LLC, true, 0);
1807 while (num_entries) {
1808 last_pte = first_pte + num_entries;
1809 if (last_pte > GEN6_PTES)
1810 last_pte = GEN6_PTES;
1812 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1814 for (i = first_pte; i < last_pte; i++)
1815 pt_vaddr[i] = scratch_pte;
1817 kunmap_px(ppgtt, pt_vaddr);
1819 num_entries -= last_pte - first_pte;
1825 static void gen6_ppgtt_insert_entries(struct i915_address_space *vm,
1826 struct sg_table *pages,
1828 enum i915_cache_level cache_level, u32 flags)
1830 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1831 unsigned first_entry = start >> PAGE_SHIFT;
1832 unsigned act_pt = first_entry / GEN6_PTES;
1833 unsigned act_pte = first_entry % GEN6_PTES;
1834 gen6_pte_t *pt_vaddr = NULL;
1835 struct sgt_iter sgt_iter;
1838 for_each_sgt_dma(addr, sgt_iter, pages) {
1839 if (pt_vaddr == NULL)
1840 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1843 vm->pte_encode(addr, cache_level, true, flags);
1845 if (++act_pte == GEN6_PTES) {
1846 kunmap_px(ppgtt, pt_vaddr);
1854 kunmap_px(ppgtt, pt_vaddr);
1857 static int gen6_alloc_va_range(struct i915_address_space *vm,
1858 uint64_t start_in, uint64_t length_in)
1860 DECLARE_BITMAP(new_page_tables, I915_PDES);
1861 struct drm_device *dev = vm->dev;
1862 struct drm_i915_private *dev_priv = to_i915(dev);
1863 struct i915_ggtt *ggtt = &dev_priv->ggtt;
1864 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1865 struct i915_page_table *pt;
1866 uint32_t start, length, start_save, length_save;
1870 if (WARN_ON(start_in + length_in > ppgtt->base.total))
1873 start = start_save = start_in;
1874 length = length_save = length_in;
1876 bitmap_zero(new_page_tables, I915_PDES);
1878 /* The allocation is done in two stages so that we can bail out with
1879 * minimal amount of pain. The first stage finds new page tables that
1880 * need allocation. The second stage marks use ptes within the page
1883 gen6_for_each_pde(pt, &ppgtt->pd, start, length, pde) {
1884 if (pt != vm->scratch_pt) {
1885 WARN_ON(bitmap_empty(pt->used_ptes, GEN6_PTES));
1889 /* We've already allocated a page table */
1890 WARN_ON(!bitmap_empty(pt->used_ptes, GEN6_PTES));
1898 gen6_initialize_pt(vm, pt);
1900 ppgtt->pd.page_table[pde] = pt;
1901 __set_bit(pde, new_page_tables);
1902 trace_i915_page_table_entry_alloc(vm, pde, start, GEN6_PDE_SHIFT);
1906 length = length_save;
1908 gen6_for_each_pde(pt, &ppgtt->pd, start, length, pde) {
1909 DECLARE_BITMAP(tmp_bitmap, GEN6_PTES);
1911 bitmap_zero(tmp_bitmap, GEN6_PTES);
1912 bitmap_set(tmp_bitmap, gen6_pte_index(start),
1913 gen6_pte_count(start, length));
1915 if (__test_and_clear_bit(pde, new_page_tables))
1916 gen6_write_pde(&ppgtt->pd, pde, pt);
1918 trace_i915_page_table_entry_map(vm, pde, pt,
1919 gen6_pte_index(start),
1920 gen6_pte_count(start, length),
1922 bitmap_or(pt->used_ptes, tmp_bitmap, pt->used_ptes,
1926 WARN_ON(!bitmap_empty(new_page_tables, I915_PDES));
1928 /* Make sure write is complete before other code can use this page
1929 * table. Also require for WC mapped PTEs */
1932 mark_tlbs_dirty(ppgtt);
1936 for_each_set_bit(pde, new_page_tables, I915_PDES) {
1937 struct i915_page_table *pt = ppgtt->pd.page_table[pde];
1939 ppgtt->pd.page_table[pde] = vm->scratch_pt;
1940 free_pt(vm->dev, pt);
1943 mark_tlbs_dirty(ppgtt);
1947 static int gen6_init_scratch(struct i915_address_space *vm)
1949 struct drm_device *dev = vm->dev;
1951 vm->scratch_page = alloc_scratch_page(dev);
1952 if (IS_ERR(vm->scratch_page))
1953 return PTR_ERR(vm->scratch_page);
1955 vm->scratch_pt = alloc_pt(dev);
1956 if (IS_ERR(vm->scratch_pt)) {
1957 free_scratch_page(dev, vm->scratch_page);
1958 return PTR_ERR(vm->scratch_pt);
1961 gen6_initialize_pt(vm, vm->scratch_pt);
1966 static void gen6_free_scratch(struct i915_address_space *vm)
1968 struct drm_device *dev = vm->dev;
1970 free_pt(dev, vm->scratch_pt);
1971 free_scratch_page(dev, vm->scratch_page);
1974 static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
1976 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1977 struct i915_page_directory *pd = &ppgtt->pd;
1978 struct drm_device *dev = vm->dev;
1979 struct i915_page_table *pt;
1982 drm_mm_remove_node(&ppgtt->node);
1984 gen6_for_all_pdes(pt, pd, pde)
1985 if (pt != vm->scratch_pt)
1988 gen6_free_scratch(vm);
1991 static int gen6_ppgtt_allocate_page_directories(struct i915_hw_ppgtt *ppgtt)
1993 struct i915_address_space *vm = &ppgtt->base;
1994 struct drm_device *dev = ppgtt->base.dev;
1995 struct drm_i915_private *dev_priv = to_i915(dev);
1996 struct i915_ggtt *ggtt = &dev_priv->ggtt;
1997 bool retried = false;
2000 /* PPGTT PDEs reside in the GGTT and consists of 512 entries. The
2001 * allocator works in address space sizes, so it's multiplied by page
2002 * size. We allocate at the top of the GTT to avoid fragmentation.
2004 BUG_ON(!drm_mm_initialized(&ggtt->base.mm));
2006 ret = gen6_init_scratch(vm);
2011 ret = drm_mm_insert_node_in_range_generic(&ggtt->base.mm,
2012 &ppgtt->node, GEN6_PD_SIZE,
2014 0, ggtt->base.total,
2016 if (ret == -ENOSPC && !retried) {
2017 ret = i915_gem_evict_something(dev, &ggtt->base,
2018 GEN6_PD_SIZE, GEN6_PD_ALIGN,
2020 0, ggtt->base.total,
2033 if (ppgtt->node.start < ggtt->mappable_end)
2034 DRM_DEBUG("Forced to use aperture for PDEs\n");
2039 gen6_free_scratch(vm);
2043 static int gen6_ppgtt_alloc(struct i915_hw_ppgtt *ppgtt)
2045 return gen6_ppgtt_allocate_page_directories(ppgtt);
2048 static void gen6_scratch_va_range(struct i915_hw_ppgtt *ppgtt,
2049 uint64_t start, uint64_t length)
2051 struct i915_page_table *unused;
2054 gen6_for_each_pde(unused, &ppgtt->pd, start, length, pde)
2055 ppgtt->pd.page_table[pde] = ppgtt->base.scratch_pt;
2058 static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
2060 struct drm_device *dev = ppgtt->base.dev;
2061 struct drm_i915_private *dev_priv = to_i915(dev);
2062 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2065 ppgtt->base.pte_encode = ggtt->base.pte_encode;
2066 if (intel_vgpu_active(dev_priv) || IS_GEN6(dev))
2067 ppgtt->switch_mm = gen6_mm_switch;
2068 else if (IS_HASWELL(dev))
2069 ppgtt->switch_mm = hsw_mm_switch;
2070 else if (IS_GEN7(dev))
2071 ppgtt->switch_mm = gen7_mm_switch;
2075 ret = gen6_ppgtt_alloc(ppgtt);
2079 ppgtt->base.allocate_va_range = gen6_alloc_va_range;
2080 ppgtt->base.clear_range = gen6_ppgtt_clear_range;
2081 ppgtt->base.insert_entries = gen6_ppgtt_insert_entries;
2082 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
2083 ppgtt->base.bind_vma = ppgtt_bind_vma;
2084 ppgtt->base.cleanup = gen6_ppgtt_cleanup;
2085 ppgtt->base.start = 0;
2086 ppgtt->base.total = I915_PDES * GEN6_PTES * PAGE_SIZE;
2087 ppgtt->debug_dump = gen6_dump_ppgtt;
2089 ppgtt->pd.base.ggtt_offset =
2090 ppgtt->node.start / PAGE_SIZE * sizeof(gen6_pte_t);
2092 ppgtt->pd_addr = (gen6_pte_t __iomem *)ggtt->gsm +
2093 ppgtt->pd.base.ggtt_offset / sizeof(gen6_pte_t);
2095 gen6_scratch_va_range(ppgtt, 0, ppgtt->base.total);
2097 gen6_write_page_range(dev_priv, &ppgtt->pd, 0, ppgtt->base.total);
2099 DRM_DEBUG_DRIVER("Allocated pde space (%lldM) at GTT entry: %llx\n",
2100 ppgtt->node.size >> 20,
2101 ppgtt->node.start / PAGE_SIZE);
2103 DRM_DEBUG("Adding PPGTT at offset %x\n",
2104 ppgtt->pd.base.ggtt_offset << 10);
2109 static int __hw_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
2111 ppgtt->base.dev = dev;
2113 if (INTEL_INFO(dev)->gen < 8)
2114 return gen6_ppgtt_init(ppgtt);
2116 return gen8_ppgtt_init(ppgtt);
2119 static void i915_address_space_init(struct i915_address_space *vm,
2120 struct drm_i915_private *dev_priv)
2122 drm_mm_init(&vm->mm, vm->start, vm->total);
2123 vm->dev = &dev_priv->drm;
2124 INIT_LIST_HEAD(&vm->active_list);
2125 INIT_LIST_HEAD(&vm->inactive_list);
2126 list_add_tail(&vm->global_link, &dev_priv->vm_list);
2129 static void gtt_write_workarounds(struct drm_device *dev)
2131 struct drm_i915_private *dev_priv = to_i915(dev);
2133 /* This function is for gtt related workarounds. This function is
2134 * called on driver load and after a GPU reset, so you can place
2135 * workarounds here even if they get overwritten by GPU reset.
2137 /* WaIncreaseDefaultTLBEntries:chv,bdw,skl,bxt */
2138 if (IS_BROADWELL(dev))
2139 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN8_L3_LRA_1_GPGPU_DEFAULT_VALUE_BDW);
2140 else if (IS_CHERRYVIEW(dev))
2141 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN8_L3_LRA_1_GPGPU_DEFAULT_VALUE_CHV);
2142 else if (IS_SKYLAKE(dev))
2143 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN9_L3_LRA_1_GPGPU_DEFAULT_VALUE_SKL);
2144 else if (IS_BROXTON(dev))
2145 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN9_L3_LRA_1_GPGPU_DEFAULT_VALUE_BXT);
2148 static int i915_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
2150 struct drm_i915_private *dev_priv = to_i915(dev);
2153 ret = __hw_ppgtt_init(dev, ppgtt);
2155 kref_init(&ppgtt->ref);
2156 i915_address_space_init(&ppgtt->base, dev_priv);
2162 int i915_ppgtt_init_hw(struct drm_device *dev)
2164 gtt_write_workarounds(dev);
2166 /* In the case of execlists, PPGTT is enabled by the context descriptor
2167 * and the PDPs are contained within the context itself. We don't
2168 * need to do anything here. */
2169 if (i915.enable_execlists)
2172 if (!USES_PPGTT(dev))
2176 gen6_ppgtt_enable(dev);
2177 else if (IS_GEN7(dev))
2178 gen7_ppgtt_enable(dev);
2179 else if (INTEL_INFO(dev)->gen >= 8)
2180 gen8_ppgtt_enable(dev);
2182 MISSING_CASE(INTEL_INFO(dev)->gen);
2187 struct i915_hw_ppgtt *
2188 i915_ppgtt_create(struct drm_device *dev, struct drm_i915_file_private *fpriv)
2190 struct i915_hw_ppgtt *ppgtt;
2193 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2195 return ERR_PTR(-ENOMEM);
2197 ret = i915_ppgtt_init(dev, ppgtt);
2200 return ERR_PTR(ret);
2203 ppgtt->file_priv = fpriv;
2205 trace_i915_ppgtt_create(&ppgtt->base);
2210 void i915_ppgtt_release(struct kref *kref)
2212 struct i915_hw_ppgtt *ppgtt =
2213 container_of(kref, struct i915_hw_ppgtt, ref);
2215 trace_i915_ppgtt_release(&ppgtt->base);
2217 /* vmas should already be unbound */
2218 WARN_ON(!list_empty(&ppgtt->base.active_list));
2219 WARN_ON(!list_empty(&ppgtt->base.inactive_list));
2221 list_del(&ppgtt->base.global_link);
2222 drm_mm_takedown(&ppgtt->base.mm);
2224 ppgtt->base.cleanup(&ppgtt->base);
2228 extern int intel_iommu_gfx_mapped;
2229 /* Certain Gen5 chipsets require require idling the GPU before
2230 * unmapping anything from the GTT when VT-d is enabled.
2232 static bool needs_idle_maps(struct drm_device *dev)
2234 #ifdef CONFIG_INTEL_IOMMU
2235 /* Query intel_iommu to see if we need the workaround. Presumably that
2238 if (IS_GEN5(dev) && IS_MOBILE(dev) && intel_iommu_gfx_mapped)
2244 static bool do_idling(struct drm_i915_private *dev_priv)
2246 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2247 bool ret = dev_priv->mm.interruptible;
2249 if (unlikely(ggtt->do_idle_maps)) {
2250 dev_priv->mm.interruptible = false;
2251 if (i915_gem_wait_for_idle(dev_priv)) {
2252 DRM_ERROR("Failed to wait for idle; VT'd may hang.\n");
2253 /* Wait a bit, in hopes it avoids the hang */
2261 static void undo_idling(struct drm_i915_private *dev_priv, bool interruptible)
2263 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2265 if (unlikely(ggtt->do_idle_maps))
2266 dev_priv->mm.interruptible = interruptible;
2269 void i915_check_and_clear_faults(struct drm_i915_private *dev_priv)
2271 struct intel_engine_cs *engine;
2273 if (INTEL_INFO(dev_priv)->gen < 6)
2276 for_each_engine(engine, dev_priv) {
2278 fault_reg = I915_READ(RING_FAULT_REG(engine));
2279 if (fault_reg & RING_FAULT_VALID) {
2280 DRM_DEBUG_DRIVER("Unexpected fault\n"
2282 "\tAddress space: %s\n"
2285 fault_reg & LINUX_PAGE_MASK,
2286 fault_reg & RING_FAULT_GTTSEL_MASK ? "GGTT" : "PPGTT",
2287 RING_FAULT_SRCID(fault_reg),
2288 RING_FAULT_FAULT_TYPE(fault_reg));
2289 I915_WRITE(RING_FAULT_REG(engine),
2290 fault_reg & ~RING_FAULT_VALID);
2293 POSTING_READ(RING_FAULT_REG(&dev_priv->engine[RCS]));
2296 static void i915_ggtt_flush(struct drm_i915_private *dev_priv)
2298 if (INTEL_INFO(dev_priv)->gen < 6) {
2299 intel_gtt_chipset_flush();
2301 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2302 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2306 void i915_gem_suspend_gtt_mappings(struct drm_device *dev)
2308 struct drm_i915_private *dev_priv = to_i915(dev);
2309 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2311 /* Don't bother messing with faults pre GEN6 as we have little
2312 * documentation supporting that it's a good idea.
2314 if (INTEL_INFO(dev)->gen < 6)
2317 i915_check_and_clear_faults(dev_priv);
2319 ggtt->base.clear_range(&ggtt->base, ggtt->base.start, ggtt->base.total,
2322 i915_ggtt_flush(dev_priv);
2325 int i915_gem_gtt_prepare_object(struct drm_i915_gem_object *obj)
2327 if (!dma_map_sg(&obj->base.dev->pdev->dev,
2328 obj->pages->sgl, obj->pages->nents,
2329 PCI_DMA_BIDIRECTIONAL))
2335 static void gen8_set_pte(void __iomem *addr, gen8_pte_t pte)
2340 iowrite32((u32)pte, addr);
2341 iowrite32(pte >> 32, addr + 4);
2345 static void gen8_ggtt_insert_page(struct i915_address_space *vm,
2348 enum i915_cache_level level,
2351 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2352 gen8_pte_t __iomem *pte =
2353 (gen8_pte_t __iomem *)dev_priv->ggtt.gsm +
2354 (offset >> PAGE_SHIFT);
2357 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2359 gen8_set_pte(pte, gen8_pte_encode(addr, level, true));
2361 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2362 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2364 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2367 static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
2368 struct sg_table *st,
2370 enum i915_cache_level level, u32 unused)
2372 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2373 struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
2374 struct sgt_iter sgt_iter;
2375 gen8_pte_t __iomem *gtt_entries;
2376 gen8_pte_t gtt_entry;
2381 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2383 gtt_entries = (gen8_pte_t __iomem *)ggtt->gsm + (start >> PAGE_SHIFT);
2385 for_each_sgt_dma(addr, sgt_iter, st) {
2386 gtt_entry = gen8_pte_encode(addr, level, true);
2387 gen8_set_pte(>t_entries[i++], gtt_entry);
2391 * XXX: This serves as a posting read to make sure that the PTE has
2392 * actually been updated. There is some concern that even though
2393 * registers and PTEs are within the same BAR that they are potentially
2394 * of NUMA access patterns. Therefore, even with the way we assume
2395 * hardware should work, we must keep this posting read for paranoia.
2398 WARN_ON(readq(>t_entries[i-1]) != gtt_entry);
2400 /* This next bit makes the above posting read even more important. We
2401 * want to flush the TLBs only after we're certain all the PTE updates
2404 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2405 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2407 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2410 struct insert_entries {
2411 struct i915_address_space *vm;
2412 struct sg_table *st;
2414 enum i915_cache_level level;
2418 static int gen8_ggtt_insert_entries__cb(void *_arg)
2420 struct insert_entries *arg = _arg;
2421 gen8_ggtt_insert_entries(arg->vm, arg->st,
2422 arg->start, arg->level, arg->flags);
2426 static void gen8_ggtt_insert_entries__BKL(struct i915_address_space *vm,
2427 struct sg_table *st,
2429 enum i915_cache_level level,
2432 struct insert_entries arg = { vm, st, start, level, flags };
2433 #ifndef __DragonFly__
2434 stop_machine(gen8_ggtt_insert_entries__cb, &arg, NULL);
2436 /* XXX: is this enough ?
2437 * See Linux commit 5bab6f60cb4d1417ad7c599166bcfec87529c1a2 */
2439 gen8_ggtt_insert_entries__cb(&arg);
2444 static void gen6_ggtt_insert_page(struct i915_address_space *vm,
2447 enum i915_cache_level level,
2450 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2451 gen6_pte_t __iomem *pte =
2452 (gen6_pte_t __iomem *)dev_priv->ggtt.gsm +
2453 (offset >> PAGE_SHIFT);
2456 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2458 iowrite32(vm->pte_encode(addr, level, true, flags), pte);
2460 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2461 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2463 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2467 * Binds an object into the global gtt with the specified cache level. The object
2468 * will be accessible to the GPU via commands whose operands reference offsets
2469 * within the global GTT as well as accessible by the GPU through the GMADR
2470 * mapped BAR (dev_priv->mm.gtt->gtt).
2472 static void gen6_ggtt_insert_entries(struct i915_address_space *vm,
2473 struct sg_table *st,
2475 enum i915_cache_level level, u32 flags)
2477 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2478 struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
2479 struct sgt_iter sgt_iter;
2480 gen6_pte_t __iomem *gtt_entries;
2481 gen6_pte_t gtt_entry;
2486 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2488 gtt_entries = (gen6_pte_t __iomem *)ggtt->gsm + (start >> PAGE_SHIFT);
2490 for_each_sgt_dma(addr, sgt_iter, st) {
2491 gtt_entry = vm->pte_encode(addr, level, true, flags);
2492 iowrite32(gtt_entry, >t_entries[i++]);
2495 /* XXX: This serves as a posting read to make sure that the PTE has
2496 * actually been updated. There is some concern that even though
2497 * registers and PTEs are within the same BAR that they are potentially
2498 * of NUMA access patterns. Therefore, even with the way we assume
2499 * hardware should work, we must keep this posting read for paranoia.
2502 WARN_ON(readl(>t_entries[i-1]) != gtt_entry);
2504 /* This next bit makes the above posting read even more important. We
2505 * want to flush the TLBs only after we're certain all the PTE updates
2508 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2509 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2511 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2514 static void nop_clear_range(struct i915_address_space *vm,
2521 static void gen8_ggtt_clear_range(struct i915_address_space *vm,
2526 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2527 struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
2528 unsigned first_entry = start >> PAGE_SHIFT;
2529 unsigned num_entries = length >> PAGE_SHIFT;
2530 gen8_pte_t scratch_pte, __iomem *gtt_base =
2531 (gen8_pte_t __iomem *)ggtt->gsm + first_entry;
2532 const int max_entries = ggtt_total_entries(ggtt) - first_entry;
2536 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2538 if (WARN(num_entries > max_entries,
2539 "First entry = %d; Num entries = %d (max=%d)\n",
2540 first_entry, num_entries, max_entries))
2541 num_entries = max_entries;
2543 scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
2546 for (i = 0; i < num_entries; i++)
2547 gen8_set_pte(>t_base[i], scratch_pte);
2550 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2553 static void gen6_ggtt_clear_range(struct i915_address_space *vm,
2558 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2559 struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
2560 unsigned first_entry = start >> PAGE_SHIFT;
2561 unsigned num_entries = length >> PAGE_SHIFT;
2562 gen6_pte_t scratch_pte, __iomem *gtt_base =
2563 (gen6_pte_t __iomem *)ggtt->gsm + first_entry;
2564 const int max_entries = ggtt_total_entries(ggtt) - first_entry;
2568 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2570 if (WARN(num_entries > max_entries,
2571 "First entry = %d; Num entries = %d (max=%d)\n",
2572 first_entry, num_entries, max_entries))
2573 num_entries = max_entries;
2575 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
2576 I915_CACHE_LLC, use_scratch, 0);
2578 for (i = 0; i < num_entries; i++)
2579 iowrite32(scratch_pte, >t_base[i]);
2582 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2585 static void i915_ggtt_insert_page(struct i915_address_space *vm,
2588 enum i915_cache_level cache_level,
2591 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2592 unsigned int flags = (cache_level == I915_CACHE_NONE) ?
2593 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
2596 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2598 intel_gtt_insert_page(addr, offset >> PAGE_SHIFT, flags);
2600 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2603 static void i915_ggtt_insert_entries(struct i915_address_space *vm,
2604 struct sg_table *pages,
2606 enum i915_cache_level cache_level, u32 unused)
2608 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2609 unsigned int flags = (cache_level == I915_CACHE_NONE) ?
2610 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
2613 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2615 intel_gtt_insert_sg_entries(pages, start >> PAGE_SHIFT, flags);
2617 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2621 static void i915_ggtt_clear_range(struct i915_address_space *vm,
2626 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2627 unsigned first_entry = start >> PAGE_SHIFT;
2628 unsigned num_entries = length >> PAGE_SHIFT;
2631 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2633 intel_gtt_clear_range(first_entry, num_entries);
2635 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2638 static int ggtt_bind_vma(struct i915_vma *vma,
2639 enum i915_cache_level cache_level,
2642 struct drm_i915_gem_object *obj = vma->obj;
2646 ret = i915_get_ggtt_vma_pages(vma);
2650 /* Currently applicable only to VLV */
2652 pte_flags |= PTE_READ_ONLY;
2654 vma->vm->insert_entries(vma->vm, vma->ggtt_view.pages,
2656 cache_level, pte_flags);
2659 * Without aliasing PPGTT there's no difference between
2660 * GLOBAL/LOCAL_BIND, it's all the same ptes. Hence unconditionally
2661 * upgrade to both bound if we bind either to avoid double-binding.
2663 vma->bound |= GLOBAL_BIND | LOCAL_BIND;
2668 static int aliasing_gtt_bind_vma(struct i915_vma *vma,
2669 enum i915_cache_level cache_level,
2675 ret = i915_get_ggtt_vma_pages(vma);
2679 /* Currently applicable only to VLV */
2681 if (vma->obj->gt_ro)
2682 pte_flags |= PTE_READ_ONLY;
2685 if (flags & GLOBAL_BIND) {
2686 vma->vm->insert_entries(vma->vm,
2687 vma->ggtt_view.pages,
2689 cache_level, pte_flags);
2692 if (flags & LOCAL_BIND) {
2693 struct i915_hw_ppgtt *appgtt =
2694 to_i915(vma->vm->dev)->mm.aliasing_ppgtt;
2695 appgtt->base.insert_entries(&appgtt->base,
2696 vma->ggtt_view.pages,
2698 cache_level, pte_flags);
2704 static void ggtt_unbind_vma(struct i915_vma *vma)
2706 struct drm_device *dev = vma->vm->dev;
2707 struct drm_i915_private *dev_priv = to_i915(dev);
2708 struct drm_i915_gem_object *obj = vma->obj;
2709 const uint64_t size = min_t(uint64_t,
2713 if (vma->bound & GLOBAL_BIND) {
2714 vma->vm->clear_range(vma->vm,
2720 if (dev_priv->mm.aliasing_ppgtt && vma->bound & LOCAL_BIND) {
2721 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
2723 appgtt->base.clear_range(&appgtt->base,
2730 void i915_gem_gtt_finish_object(struct drm_i915_gem_object *obj)
2732 struct drm_device *dev = obj->base.dev;
2733 struct drm_i915_private *dev_priv = to_i915(dev);
2736 interruptible = do_idling(dev_priv);
2738 dma_unmap_sg(&dev->pdev->dev, obj->pages->sgl, obj->pages->nents,
2739 PCI_DMA_BIDIRECTIONAL);
2741 undo_idling(dev_priv, interruptible);
2744 static void i915_gtt_color_adjust(struct drm_mm_node *node,
2745 unsigned long color,
2749 if (node->color != color)
2752 if (!list_empty(&node->node_list)) {
2753 node = list_entry(node->node_list.next,
2756 if (node->allocated && node->color != color)
2761 static int i915_gem_setup_global_gtt(struct drm_device *dev,
2766 /* Let GEM Manage all of the aperture.
2768 * However, leave one page at the end still bound to the scratch page.
2769 * There are a number of places where the hardware apparently prefetches
2770 * past the end of the object, and we've seen multiple hangs with the
2771 * GPU head pointer stuck in a batchbuffer bound at the last page of the
2772 * aperture. One page should be enough to keep any prefetching inside
2775 struct drm_i915_private *dev_priv = to_i915(dev);
2776 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2777 struct drm_mm_node *entry;
2778 struct drm_i915_gem_object *obj;
2779 unsigned long hole_start, hole_end;
2781 unsigned long mappable = min(end, mappable_end) - start;
2784 BUG_ON(mappable_end > end);
2786 ggtt->base.start = start;
2788 /* Subtract the guard page before address space initialization to
2789 * shrink the range used by drm_mm */
2790 ggtt->base.total = end - start - PAGE_SIZE;
2791 i915_address_space_init(&ggtt->base, dev_priv);
2792 ggtt->base.total += PAGE_SIZE;
2794 ret = intel_vgt_balloon(dev_priv);
2799 ggtt->base.mm.color_adjust = i915_gtt_color_adjust;
2801 /* Mark any preallocated objects as occupied */
2802 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
2803 struct i915_vma *vma = i915_gem_obj_to_vma(obj, &ggtt->base);
2805 DRM_DEBUG_KMS("reserving preallocated space: %llx + %zx\n",
2806 i915_gem_obj_ggtt_offset(obj), obj->base.size);
2808 WARN_ON(i915_gem_obj_ggtt_bound(obj));
2809 ret = drm_mm_reserve_node(&ggtt->base.mm, &vma->node);
2811 DRM_DEBUG_KMS("Reservation failed: %i\n", ret);
2814 vma->bound |= GLOBAL_BIND;
2815 __i915_vma_set_map_and_fenceable(vma);
2816 list_add_tail(&vma->vm_link, &ggtt->base.inactive_list);
2819 /* Clear any non-preallocated blocks */
2820 drm_mm_for_each_hole(entry, &ggtt->base.mm, hole_start, hole_end) {
2821 DRM_DEBUG_KMS("clearing unused GTT space: [%lx, %lx]\n",
2822 hole_start, hole_end);
2823 ggtt->base.clear_range(&ggtt->base, hole_start,
2824 hole_end - hole_start, true);
2827 #ifdef __DragonFly__
2828 device_printf(dev->dev->bsddev,
2829 "taking over the fictitious range 0x%llx-0x%llx\n",
2830 dev_priv->ggtt.mappable_base + start, dev_priv->ggtt.mappable_base + start + mappable);
2831 error = -vm_phys_fictitious_reg_range(dev_priv->ggtt.mappable_base + start,
2832 dev_priv->ggtt.mappable_base + start + mappable, VM_MEMATTR_WRITE_COMBINING);
2835 /* And finally clear the reserved guard page */
2836 ggtt->base.clear_range(&ggtt->base, end - PAGE_SIZE, PAGE_SIZE, true);
2838 if (USES_PPGTT(dev) && !USES_FULL_PPGTT(dev)) {
2839 struct i915_hw_ppgtt *ppgtt;
2841 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2845 ret = __hw_ppgtt_init(dev, ppgtt);
2847 ppgtt->base.cleanup(&ppgtt->base);
2852 if (ppgtt->base.allocate_va_range)
2853 ret = ppgtt->base.allocate_va_range(&ppgtt->base, 0,
2856 ppgtt->base.cleanup(&ppgtt->base);
2861 ppgtt->base.clear_range(&ppgtt->base,
2866 dev_priv->mm.aliasing_ppgtt = ppgtt;
2867 WARN_ON(ggtt->base.bind_vma != ggtt_bind_vma);
2868 ggtt->base.bind_vma = aliasing_gtt_bind_vma;
2875 * i915_gem_init_ggtt - Initialize GEM for Global GTT
2878 void i915_gem_init_ggtt(struct drm_device *dev)
2880 struct drm_i915_private *dev_priv = to_i915(dev);
2881 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2883 i915_gem_setup_global_gtt(dev, 0, ggtt->mappable_end, ggtt->base.total);
2887 * i915_ggtt_cleanup_hw - Clean up GGTT hardware initialization
2890 void i915_ggtt_cleanup_hw(struct drm_device *dev)
2892 struct drm_i915_private *dev_priv = to_i915(dev);
2893 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2895 if (dev_priv->mm.aliasing_ppgtt) {
2896 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
2898 ppgtt->base.cleanup(&ppgtt->base);
2902 i915_gem_cleanup_stolen(dev);
2904 if (drm_mm_initialized(&ggtt->base.mm)) {
2905 intel_vgt_deballoon(dev_priv);
2907 drm_mm_takedown(&ggtt->base.mm);
2908 list_del(&ggtt->base.global_link);
2911 ggtt->base.cleanup(&ggtt->base);
2914 static unsigned int gen6_get_total_gtt_size(u16 snb_gmch_ctl)
2916 snb_gmch_ctl >>= SNB_GMCH_GGMS_SHIFT;
2917 snb_gmch_ctl &= SNB_GMCH_GGMS_MASK;
2918 return snb_gmch_ctl << 20;
2921 static unsigned int gen8_get_total_gtt_size(u16 bdw_gmch_ctl)
2923 bdw_gmch_ctl >>= BDW_GMCH_GGMS_SHIFT;
2924 bdw_gmch_ctl &= BDW_GMCH_GGMS_MASK;
2926 bdw_gmch_ctl = 1 << bdw_gmch_ctl;
2928 #ifdef CONFIG_X86_32
2929 /* Limit 32b platforms to a 2GB GGTT: 4 << 20 / pte size * PAGE_SIZE */
2930 if (bdw_gmch_ctl > 4)
2934 return bdw_gmch_ctl << 20;
2937 static unsigned int chv_get_total_gtt_size(u16 gmch_ctrl)
2939 gmch_ctrl >>= SNB_GMCH_GGMS_SHIFT;
2940 gmch_ctrl &= SNB_GMCH_GGMS_MASK;
2943 return 1 << (20 + gmch_ctrl);
2948 static size_t gen6_get_stolen_size(u16 snb_gmch_ctl)
2950 snb_gmch_ctl >>= SNB_GMCH_GMS_SHIFT;
2951 snb_gmch_ctl &= SNB_GMCH_GMS_MASK;
2952 return snb_gmch_ctl << 25; /* 32 MB units */
2955 static size_t gen8_get_stolen_size(u16 bdw_gmch_ctl)
2957 bdw_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2958 bdw_gmch_ctl &= BDW_GMCH_GMS_MASK;
2959 return bdw_gmch_ctl << 25; /* 32 MB units */
2962 static size_t chv_get_stolen_size(u16 gmch_ctrl)
2964 gmch_ctrl >>= SNB_GMCH_GMS_SHIFT;
2965 gmch_ctrl &= SNB_GMCH_GMS_MASK;
2968 * 0x0 to 0x10: 32MB increments starting at 0MB
2969 * 0x11 to 0x16: 4MB increments starting at 8MB
2970 * 0x17 to 0x1d: 4MB increments start at 36MB
2972 if (gmch_ctrl < 0x11)
2973 return gmch_ctrl << 25;
2974 else if (gmch_ctrl < 0x17)
2975 return (gmch_ctrl - 0x11 + 2) << 22;
2977 return (gmch_ctrl - 0x17 + 9) << 22;
2980 static size_t gen9_get_stolen_size(u16 gen9_gmch_ctl)
2982 gen9_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2983 gen9_gmch_ctl &= BDW_GMCH_GMS_MASK;
2985 if (gen9_gmch_ctl < 0xf0)
2986 return gen9_gmch_ctl << 25; /* 32 MB units */
2988 /* 4MB increments starting at 0xf0 for 4MB */
2989 return (gen9_gmch_ctl - 0xf0 + 1) << 22;
2992 static int ggtt_probe_common(struct drm_device *dev,
2995 struct drm_i915_private *dev_priv = to_i915(dev);
2996 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2997 struct i915_page_scratch *scratch_page;
2998 phys_addr_t ggtt_phys_addr;
3000 /* For Modern GENs the PTEs and register space are split in the BAR */
3001 ggtt_phys_addr = pci_resource_start(dev->pdev, 0) +
3002 (pci_resource_len(dev->pdev, 0) / 2);
3005 * On BXT writes larger than 64 bit to the GTT pagetable range will be
3006 * dropped. For WC mappings in general we have 64 byte burst writes
3007 * when the WC buffer is flushed, so we can't use it, but have to
3008 * resort to an uncached mapping. The WC issue is easily caught by the
3009 * readback check when writing GTT PTE entries.
3011 if (IS_BROXTON(dev))
3012 ggtt->gsm = ioremap_nocache(ggtt_phys_addr, gtt_size);
3014 ggtt->gsm = ioremap_wc(ggtt_phys_addr, gtt_size);
3016 DRM_ERROR("Failed to map the gtt page table\n");
3020 scratch_page = alloc_scratch_page(dev);
3021 if (IS_ERR(scratch_page)) {
3022 DRM_ERROR("Scratch setup failed\n");
3023 /* iounmap will also get called at remove, but meh */
3025 return PTR_ERR(scratch_page);
3028 ggtt->base.scratch_page = scratch_page;
3033 /* The GGTT and PPGTT need a private PPAT setup in order to handle cacheability
3034 * bits. When using advanced contexts each context stores its own PAT, but
3035 * writing this data shouldn't be harmful even in those cases. */
3036 static void bdw_setup_private_ppat(struct drm_i915_private *dev_priv)
3040 pat = GEN8_PPAT(0, GEN8_PPAT_WB | GEN8_PPAT_LLC) | /* for normal objects, no eLLC */
3041 GEN8_PPAT(1, GEN8_PPAT_WC | GEN8_PPAT_LLCELLC) | /* for something pointing to ptes? */
3042 GEN8_PPAT(2, GEN8_PPAT_WT | GEN8_PPAT_LLCELLC) | /* for scanout with eLLC */
3043 GEN8_PPAT(3, GEN8_PPAT_UC) | /* Uncached objects, mostly for scanout */
3044 GEN8_PPAT(4, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(0)) |
3045 GEN8_PPAT(5, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(1)) |
3046 GEN8_PPAT(6, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(2)) |
3047 GEN8_PPAT(7, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3));
3049 if (!USES_PPGTT(dev_priv))
3050 /* Spec: "For GGTT, there is NO pat_sel[2:0] from the entry,
3051 * so RTL will always use the value corresponding to
3053 * So let's disable cache for GGTT to avoid screen corruptions.
3054 * MOCS still can be used though.
3055 * - System agent ggtt writes (i.e. cpu gtt mmaps) already work
3056 * before this patch, i.e. the same uncached + snooping access
3057 * like on gen6/7 seems to be in effect.
3058 * - So this just fixes blitter/render access. Again it looks
3059 * like it's not just uncached access, but uncached + snooping.
3060 * So we can still hold onto all our assumptions wrt cpu
3061 * clflushing on LLC machines.
3063 pat = GEN8_PPAT(0, GEN8_PPAT_UC);
3065 /* XXX: spec defines this as 2 distinct registers. It's unclear if a 64b
3066 * write would work. */
3067 I915_WRITE(GEN8_PRIVATE_PAT_LO, pat);
3068 I915_WRITE(GEN8_PRIVATE_PAT_HI, pat >> 32);
3071 static void chv_setup_private_ppat(struct drm_i915_private *dev_priv)
3076 * Map WB on BDW to snooped on CHV.
3078 * Only the snoop bit has meaning for CHV, the rest is
3081 * The hardware will never snoop for certain types of accesses:
3082 * - CPU GTT (GMADR->GGTT->no snoop->memory)
3083 * - PPGTT page tables
3084 * - some other special cycles
3086 * As with BDW, we also need to consider the following for GT accesses:
3087 * "For GGTT, there is NO pat_sel[2:0] from the entry,
3088 * so RTL will always use the value corresponding to
3090 * Which means we must set the snoop bit in PAT entry 0
3091 * in order to keep the global status page working.
3093 pat = GEN8_PPAT(0, CHV_PPAT_SNOOP) |
3097 GEN8_PPAT(4, CHV_PPAT_SNOOP) |
3098 GEN8_PPAT(5, CHV_PPAT_SNOOP) |
3099 GEN8_PPAT(6, CHV_PPAT_SNOOP) |
3100 GEN8_PPAT(7, CHV_PPAT_SNOOP);
3102 I915_WRITE(GEN8_PRIVATE_PAT_LO, pat);
3103 I915_WRITE(GEN8_PRIVATE_PAT_HI, pat >> 32);
3106 static int gen8_gmch_probe(struct i915_ggtt *ggtt)
3108 struct drm_device *dev = ggtt->base.dev;
3109 struct drm_i915_private *dev_priv = to_i915(dev);
3113 /* TODO: We're not aware of mappable constraints on gen8 yet */
3114 ggtt->mappable_base = pci_resource_start(dev->pdev, 2);
3115 ggtt->mappable_end = pci_resource_len(dev->pdev, 2);
3118 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(39)))
3119 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(39));
3122 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
3124 if (INTEL_INFO(dev)->gen >= 9) {
3125 ggtt->stolen_size = gen9_get_stolen_size(snb_gmch_ctl);
3126 ggtt->size = gen8_get_total_gtt_size(snb_gmch_ctl);
3127 } else if (IS_CHERRYVIEW(dev)) {
3128 ggtt->stolen_size = chv_get_stolen_size(snb_gmch_ctl);
3129 ggtt->size = chv_get_total_gtt_size(snb_gmch_ctl);
3131 ggtt->stolen_size = gen8_get_stolen_size(snb_gmch_ctl);
3132 ggtt->size = gen8_get_total_gtt_size(snb_gmch_ctl);
3135 ggtt->base.total = (ggtt->size / sizeof(gen8_pte_t)) << PAGE_SHIFT;
3137 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
3138 chv_setup_private_ppat(dev_priv);
3140 bdw_setup_private_ppat(dev_priv);
3142 ret = ggtt_probe_common(dev, ggtt->size);
3144 ggtt->base.bind_vma = ggtt_bind_vma;
3145 ggtt->base.unbind_vma = ggtt_unbind_vma;
3146 ggtt->base.insert_page = gen8_ggtt_insert_page;
3147 ggtt->base.clear_range = nop_clear_range;
3148 if (!USES_FULL_PPGTT(dev_priv) || intel_scanout_needs_vtd_wa(dev_priv))
3149 ggtt->base.clear_range = gen8_ggtt_clear_range;
3151 ggtt->base.insert_entries = gen8_ggtt_insert_entries;
3152 if (IS_CHERRYVIEW(dev_priv))
3153 ggtt->base.insert_entries = gen8_ggtt_insert_entries__BKL;
3158 static int gen6_gmch_probe(struct i915_ggtt *ggtt)
3160 struct drm_device *dev = ggtt->base.dev;
3164 ggtt->mappable_base = pci_resource_start(dev->pdev, 2);
3165 ggtt->mappable_end = pci_resource_len(dev->pdev, 2);
3167 /* 64/512MB is the current min/max we actually know of, but this is just
3168 * a coarse sanity check.
3170 if ((ggtt->mappable_end < (64<<20) || (ggtt->mappable_end > (512<<20)))) {
3171 DRM_ERROR("Unknown GMADR size (%llx)\n", ggtt->mappable_end);
3176 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(40)))
3177 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(40));
3179 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
3181 ggtt->stolen_size = gen6_get_stolen_size(snb_gmch_ctl);
3182 ggtt->size = gen6_get_total_gtt_size(snb_gmch_ctl);
3183 ggtt->base.total = (ggtt->size / sizeof(gen6_pte_t)) << PAGE_SHIFT;
3185 ret = ggtt_probe_common(dev, ggtt->size);
3187 ggtt->base.clear_range = gen6_ggtt_clear_range;
3188 ggtt->base.insert_page = gen6_ggtt_insert_page;
3189 ggtt->base.insert_entries = gen6_ggtt_insert_entries;
3190 ggtt->base.bind_vma = ggtt_bind_vma;
3191 ggtt->base.unbind_vma = ggtt_unbind_vma;
3196 static void gen6_gmch_remove(struct i915_address_space *vm)
3198 struct i915_ggtt *ggtt = container_of(vm, struct i915_ggtt, base);
3201 free_scratch_page(vm->dev, vm->scratch_page);
3204 static int i915_gmch_probe(struct i915_ggtt *ggtt)
3206 struct drm_device *dev = ggtt->base.dev;
3207 struct drm_i915_private *dev_priv = to_i915(dev);
3211 ret = intel_gmch_probe(dev_priv->bridge_dev, dev_priv->drm.pdev, NULL);
3213 DRM_ERROR("failed to set up gmch\n");
3218 intel_gtt_get(&ggtt->base.total, &ggtt->stolen_size,
3219 &ggtt->mappable_base, &ggtt->mappable_end);
3221 ggtt->do_idle_maps = needs_idle_maps(&dev_priv->drm);
3222 ggtt->base.insert_page = i915_ggtt_insert_page;
3223 ggtt->base.insert_entries = i915_ggtt_insert_entries;
3224 ggtt->base.clear_range = i915_ggtt_clear_range;
3225 ggtt->base.bind_vma = ggtt_bind_vma;
3226 ggtt->base.unbind_vma = ggtt_unbind_vma;
3228 if (unlikely(ggtt->do_idle_maps))
3229 DRM_INFO("applying Ironlake quirks for intel_iommu\n");
3234 static void i915_gmch_remove(struct i915_address_space *vm)
3236 intel_gmch_remove();
3240 * i915_ggtt_init_hw - Initialize GGTT hardware
3243 int i915_ggtt_init_hw(struct drm_device *dev)
3245 struct drm_i915_private *dev_priv = to_i915(dev);
3246 struct i915_ggtt *ggtt = &dev_priv->ggtt;
3249 if (INTEL_INFO(dev)->gen <= 5) {
3250 ggtt->probe = i915_gmch_probe;
3251 ggtt->base.cleanup = i915_gmch_remove;
3252 } else if (INTEL_INFO(dev)->gen < 8) {
3253 ggtt->probe = gen6_gmch_probe;
3254 ggtt->base.cleanup = gen6_gmch_remove;
3257 ggtt->base.pte_encode = iris_pte_encode;
3258 else if (IS_HASWELL(dev))
3259 ggtt->base.pte_encode = hsw_pte_encode;
3260 else if (IS_VALLEYVIEW(dev))
3261 ggtt->base.pte_encode = byt_pte_encode;
3262 else if (INTEL_INFO(dev)->gen >= 7)
3263 ggtt->base.pte_encode = ivb_pte_encode;
3265 ggtt->base.pte_encode = snb_pte_encode;
3267 ggtt->probe = gen8_gmch_probe;
3268 ggtt->base.cleanup = gen6_gmch_remove;
3271 ggtt->base.dev = dev;
3272 ggtt->base.is_ggtt = true;
3274 ret = ggtt->probe(ggtt);
3278 if ((ggtt->base.total - 1) >> 32) {
3279 DRM_ERROR("We never expected a Global GTT with more than 32bits"
3280 "of address space! Found %lldM!\n",
3281 ggtt->base.total >> 20);
3282 ggtt->base.total = 1ULL << 32;
3283 ggtt->mappable_end = min(ggtt->mappable_end, ggtt->base.total);
3287 * Initialise stolen early so that we may reserve preallocated
3288 * objects for the BIOS to KMS transition.
3290 ret = i915_gem_init_stolen(dev);
3292 goto out_gtt_cleanup;
3294 /* GMADR is the PCI mmio aperture into the global GTT. */
3295 DRM_INFO("Memory usable by graphics device = %lluM\n",
3296 ggtt->base.total >> 20);
3297 DRM_DEBUG_DRIVER("GMADR size = %lldM\n", ggtt->mappable_end >> 20);
3298 DRM_DEBUG_DRIVER("GTT stolen size = %zdM\n", ggtt->stolen_size >> 20);
3299 #ifdef CONFIG_INTEL_IOMMU
3300 if (intel_iommu_gfx_mapped)
3301 DRM_INFO("VT-d active for gfx access\n");
3307 ggtt->base.cleanup(&ggtt->base);
3312 int i915_ggtt_enable_hw(struct drm_device *dev)
3314 if (INTEL_INFO(dev)->gen < 6 && !intel_enable_gtt())
3320 void i915_gem_restore_gtt_mappings(struct drm_device *dev)
3322 struct drm_i915_private *dev_priv = to_i915(dev);
3323 struct i915_ggtt *ggtt = &dev_priv->ggtt;
3324 struct drm_i915_gem_object *obj;
3325 struct i915_vma *vma;
3327 i915_check_and_clear_faults(dev_priv);
3329 /* First fill our portion of the GTT with scratch pages */
3330 ggtt->base.clear_range(&ggtt->base, ggtt->base.start, ggtt->base.total,
3333 /* Cache flush objects bound into GGTT and rebind them. */
3334 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
3335 list_for_each_entry(vma, &obj->vma_list, obj_link) {
3336 if (vma->vm != &ggtt->base)
3339 WARN_ON(i915_vma_bind(vma, obj->cache_level,
3343 if (obj->pin_display)
3344 WARN_ON(i915_gem_object_set_to_gtt_domain(obj, false));
3347 if (INTEL_INFO(dev)->gen >= 8) {
3348 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
3349 chv_setup_private_ppat(dev_priv);
3351 bdw_setup_private_ppat(dev_priv);
3356 if (USES_PPGTT(dev)) {
3357 struct i915_address_space *vm;
3359 list_for_each_entry(vm, &dev_priv->vm_list, global_link) {
3360 /* TODO: Perhaps it shouldn't be gen6 specific */
3362 struct i915_hw_ppgtt *ppgtt;
3365 ppgtt = dev_priv->mm.aliasing_ppgtt;
3367 ppgtt = i915_vm_to_ppgtt(vm);
3369 gen6_write_page_range(dev_priv, &ppgtt->pd,
3370 0, ppgtt->base.total);
3374 i915_ggtt_flush(dev_priv);
3377 static struct i915_vma *
3378 __i915_gem_vma_create(struct drm_i915_gem_object *obj,
3379 struct i915_address_space *vm,
3380 const struct i915_ggtt_view *ggtt_view)
3382 struct i915_vma *vma;
3384 if (WARN_ON(i915_is_ggtt(vm) != !!ggtt_view))
3385 return ERR_PTR(-EINVAL);
3387 vma = kzalloc(sizeof(*vma), GFP_KERNEL);
3389 return ERR_PTR(-ENOMEM);
3391 INIT_LIST_HEAD(&vma->vm_link);
3392 INIT_LIST_HEAD(&vma->obj_link);
3393 INIT_LIST_HEAD(&vma->exec_list);
3396 vma->is_ggtt = i915_is_ggtt(vm);
3398 if (i915_is_ggtt(vm))
3399 vma->ggtt_view = *ggtt_view;
3401 i915_ppgtt_get(i915_vm_to_ppgtt(vm));
3403 list_add_tail(&vma->obj_link, &obj->vma_list);
3409 i915_gem_obj_lookup_or_create_vma(struct drm_i915_gem_object *obj,
3410 struct i915_address_space *vm)
3412 struct i915_vma *vma;
3414 vma = i915_gem_obj_to_vma(obj, vm);
3416 vma = __i915_gem_vma_create(obj, vm,
3417 i915_is_ggtt(vm) ? &i915_ggtt_view_normal : NULL);
3423 i915_gem_obj_lookup_or_create_ggtt_vma(struct drm_i915_gem_object *obj,
3424 const struct i915_ggtt_view *view)
3426 struct drm_device *dev = obj->base.dev;
3427 struct drm_i915_private *dev_priv = to_i915(dev);
3428 struct i915_ggtt *ggtt = &dev_priv->ggtt;
3429 struct i915_vma *vma = i915_gem_obj_to_ggtt_view(obj, view);
3432 vma = __i915_gem_vma_create(obj, &ggtt->base, view);
3438 static struct scatterlist *
3439 rotate_pages(const dma_addr_t *in, unsigned int offset,
3440 unsigned int width, unsigned int height,
3441 unsigned int stride,
3442 struct sg_table *st, struct scatterlist *sg)
3444 unsigned int column, row;
3445 unsigned int src_idx;
3447 for (column = 0; column < width; column++) {
3448 src_idx = stride * (height - 1) + column;
3449 for (row = 0; row < height; row++) {
3451 /* We don't need the pages, but need to initialize
3452 * the entries so the sg list can be happily traversed.
3453 * The only thing we need are DMA addresses.
3455 sg_set_page(sg, NULL, PAGE_SIZE, 0);
3456 sg_dma_address(sg) = in[offset + src_idx];
3457 sg_dma_len(sg) = PAGE_SIZE;
3466 static struct sg_table *
3467 intel_rotate_fb_obj_pages(struct intel_rotation_info *rot_info,
3468 struct drm_i915_gem_object *obj)
3470 const size_t n_pages = obj->base.size / PAGE_SIZE;
3471 unsigned int size_pages = rot_info->plane[0].width * rot_info->plane[0].height;
3472 unsigned int size_pages_uv;
3473 struct sgt_iter sgt_iter;
3474 dma_addr_t dma_addr;
3476 dma_addr_t *page_addr_list;
3477 struct sg_table *st;
3478 unsigned int uv_start_page;
3479 struct scatterlist *sg;
3482 /* Allocate a temporary list of source pages for random access. */
3483 page_addr_list = drm_malloc_gfp(n_pages,
3486 if (!page_addr_list)
3487 return ERR_PTR(ret);
3489 /* Account for UV plane with NV12. */
3490 if (rot_info->pixel_format == DRM_FORMAT_NV12)
3491 size_pages_uv = rot_info->plane[1].width * rot_info->plane[1].height;
3495 /* Allocate target SG list. */
3496 st = kmalloc(sizeof(*st), M_DRM, GFP_KERNEL);
3500 ret = sg_alloc_table(st, size_pages + size_pages_uv, GFP_KERNEL);
3504 /* Populate source page list from the object. */
3506 for_each_sgt_dma(dma_addr, sgt_iter, obj->pages)
3507 page_addr_list[i++] = dma_addr;
3509 GEM_BUG_ON(i != n_pages);
3513 /* Rotate the pages. */
3514 sg = rotate_pages(page_addr_list, 0,
3515 rot_info->plane[0].width, rot_info->plane[0].height,
3516 rot_info->plane[0].width,
3519 /* Append the UV plane if NV12. */
3520 if (rot_info->pixel_format == DRM_FORMAT_NV12) {
3521 uv_start_page = size_pages;
3523 /* Check for tile-row un-alignment. */
3524 if (offset_in_page(rot_info->uv_offset))
3527 rot_info->uv_start_page = uv_start_page;
3529 sg = rotate_pages(page_addr_list, rot_info->uv_start_page,
3530 rot_info->plane[1].width, rot_info->plane[1].height,
3531 rot_info->plane[1].width,
3535 DRM_DEBUG_KMS("Created rotated page mapping for object size %zu (%ux%u tiles, %u pages (%u plane 0)).\n",
3536 obj->base.size, rot_info->plane[0].width,
3537 rot_info->plane[0].height, size_pages + size_pages_uv,
3540 drm_free_large(page_addr_list);
3547 drm_free_large(page_addr_list);
3549 DRM_DEBUG_KMS("Failed to create rotated mapping for object size %zu! (%d) (%ux%u tiles, %u pages (%u plane 0))\n",
3550 obj->base.size, ret, rot_info->plane[0].width,
3551 rot_info->plane[0].height, size_pages + size_pages_uv,
3553 return ERR_PTR(ret);
3556 static struct sg_table *
3557 intel_partial_pages(const struct i915_ggtt_view *view,
3558 struct drm_i915_gem_object *obj)
3560 struct sg_table *st;
3561 struct scatterlist *sg;
3562 struct sg_page_iter obj_sg_iter;
3565 st = kmalloc(sizeof(*st), M_DRM, GFP_KERNEL);
3569 ret = sg_alloc_table(st, view->params.partial.size, GFP_KERNEL);
3575 for_each_sg_page(obj->pages->sgl, &obj_sg_iter, obj->pages->nents,
3576 view->params.partial.offset)
3578 if (st->nents >= view->params.partial.size)
3581 sg_set_page(sg, NULL, PAGE_SIZE, 0);
3582 sg_dma_address(sg) = sg_page_iter_dma_address(&obj_sg_iter);
3583 sg_dma_len(sg) = PAGE_SIZE;
3594 return ERR_PTR(ret);
3598 i915_get_ggtt_vma_pages(struct i915_vma *vma)
3602 if (vma->ggtt_view.pages)
3605 if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL)
3606 vma->ggtt_view.pages = vma->obj->pages;
3607 else if (vma->ggtt_view.type == I915_GGTT_VIEW_ROTATED)
3608 vma->ggtt_view.pages =
3609 intel_rotate_fb_obj_pages(&vma->ggtt_view.params.rotated, vma->obj);
3610 else if (vma->ggtt_view.type == I915_GGTT_VIEW_PARTIAL)
3611 vma->ggtt_view.pages =
3612 intel_partial_pages(&vma->ggtt_view, vma->obj);
3614 WARN_ONCE(1, "GGTT view %u not implemented!\n",
3615 vma->ggtt_view.type);
3617 if (!vma->ggtt_view.pages) {
3618 DRM_ERROR("Failed to get pages for GGTT view type %u!\n",
3619 vma->ggtt_view.type);
3621 } else if (IS_ERR(vma->ggtt_view.pages)) {
3622 ret = PTR_ERR(vma->ggtt_view.pages);
3623 vma->ggtt_view.pages = NULL;
3624 DRM_ERROR("Failed to get pages for VMA view type %u (%d)!\n",
3625 vma->ggtt_view.type, ret);
3632 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
3634 * @cache_level: mapping cache level
3635 * @flags: flags like global or local mapping
3637 * DMA addresses are taken from the scatter-gather table of this object (or of
3638 * this VMA in case of non-default GGTT views) and PTE entries set up.
3639 * Note that DMA addresses are also the only part of the SG table we care about.
3641 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
3647 if (WARN_ON(flags == 0))
3651 if (flags & PIN_GLOBAL)
3652 bind_flags |= GLOBAL_BIND;
3653 if (flags & PIN_USER)
3654 bind_flags |= LOCAL_BIND;
3656 if (flags & PIN_UPDATE)
3657 bind_flags |= vma->bound;
3659 bind_flags &= ~vma->bound;
3661 if (bind_flags == 0)
3664 if (vma->bound == 0 && vma->vm->allocate_va_range) {
3665 /* XXX: i915_vma_pin() will fix this +- hack */
3667 trace_i915_va_alloc(vma);
3668 ret = vma->vm->allocate_va_range(vma->vm,
3676 ret = vma->vm->bind_vma(vma, cache_level, bind_flags);
3680 vma->bound |= bind_flags;
3686 * i915_ggtt_view_size - Get the size of a GGTT view.
3687 * @obj: Object the view is of.
3688 * @view: The view in question.
3690 * @return The size of the GGTT view in bytes.
3693 i915_ggtt_view_size(struct drm_i915_gem_object *obj,
3694 const struct i915_ggtt_view *view)
3696 if (view->type == I915_GGTT_VIEW_NORMAL) {
3697 return obj->base.size;
3698 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
3699 return intel_rotation_info_size(&view->params.rotated) << PAGE_SHIFT;
3700 } else if (view->type == I915_GGTT_VIEW_PARTIAL) {
3701 return view->params.partial.size << PAGE_SHIFT;
3703 WARN_ONCE(1, "GGTT view %u not implemented!\n", view->type);
3704 return obj->base.size;
3708 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
3712 lockdep_assert_held(&vma->vm->dev->struct_mutex);
3713 if (WARN_ON(!vma->obj->map_and_fenceable))
3714 return IO_ERR_PTR(-ENODEV);
3716 GEM_BUG_ON(!vma->is_ggtt);
3717 GEM_BUG_ON((vma->bound & GLOBAL_BIND) == 0);
3721 ptr = io_mapping_map_wc(i915_vm_to_ggtt(vma->vm)->mappable,
3725 return IO_ERR_PTR(-ENOMEM);