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).
96 static inline struct i915_ggtt *
97 i915_vm_to_ggtt(struct i915_address_space *vm)
99 GEM_BUG_ON(!i915_is_ggtt(vm));
100 return container_of(vm, struct i915_ggtt, base);
104 i915_get_ggtt_vma_pages(struct i915_vma *vma);
106 const struct i915_ggtt_view i915_ggtt_view_normal = {
107 .type = I915_GGTT_VIEW_NORMAL,
109 const struct i915_ggtt_view i915_ggtt_view_rotated = {
110 .type = I915_GGTT_VIEW_ROTATED,
113 int intel_sanitize_enable_ppgtt(struct drm_i915_private *dev_priv,
116 bool has_aliasing_ppgtt;
118 bool has_full_48bit_ppgtt;
120 has_aliasing_ppgtt = INTEL_GEN(dev_priv) >= 6;
121 has_full_ppgtt = INTEL_GEN(dev_priv) >= 7;
122 has_full_48bit_ppgtt =
123 IS_BROADWELL(dev_priv) || INTEL_GEN(dev_priv) >= 9;
125 if (intel_vgpu_active(dev_priv))
126 has_full_ppgtt = false; /* emulation is too hard */
128 if (!has_aliasing_ppgtt)
132 * We don't allow disabling PPGTT for gen9+ as it's a requirement for
133 * execlists, the sole mechanism available to submit work.
135 if (enable_ppgtt == 0 && INTEL_GEN(dev_priv) < 9)
138 if (enable_ppgtt == 1)
141 if (enable_ppgtt == 2 && has_full_ppgtt)
144 if (enable_ppgtt == 3 && has_full_48bit_ppgtt)
147 #ifdef CONFIG_INTEL_IOMMU
148 /* Disable ppgtt on SNB if VT-d is on. */
149 if (IS_GEN6(dev_priv) && intel_iommu_gfx_mapped) {
150 DRM_INFO("Disabling PPGTT because VT-d is on\n");
155 /* Early VLV doesn't have this */
156 if (IS_VALLEYVIEW(dev_priv) && dev_priv->drm.pdev->revision < 0xb) {
157 DRM_DEBUG_DRIVER("disabling PPGTT on pre-B3 step VLV\n");
161 if (INTEL_GEN(dev_priv) >= 8 && i915.enable_execlists)
162 return has_full_48bit_ppgtt ? 3 : 2;
164 return has_aliasing_ppgtt ? 1 : 0;
167 static int ppgtt_bind_vma(struct i915_vma *vma,
168 enum i915_cache_level cache_level,
173 /* Currently applicable only to VLV */
175 pte_flags |= PTE_READ_ONLY;
177 vma->vm->insert_entries(vma->vm, vma->obj->pages, vma->node.start,
178 cache_level, pte_flags);
183 static void ppgtt_unbind_vma(struct i915_vma *vma)
185 vma->vm->clear_range(vma->vm,
191 static gen8_pte_t gen8_pte_encode(dma_addr_t addr,
192 enum i915_cache_level level,
195 gen8_pte_t pte = valid ? _PAGE_PRESENT | _PAGE_RW : 0;
199 case I915_CACHE_NONE:
200 pte |= PPAT_UNCACHED_INDEX;
203 pte |= PPAT_DISPLAY_ELLC_INDEX;
206 pte |= PPAT_CACHED_INDEX;
213 static gen8_pde_t gen8_pde_encode(const dma_addr_t addr,
214 const enum i915_cache_level level)
216 gen8_pde_t pde = _PAGE_PRESENT | _PAGE_RW;
218 if (level != I915_CACHE_NONE)
219 pde |= PPAT_CACHED_PDE_INDEX;
221 pde |= PPAT_UNCACHED_INDEX;
225 #define gen8_pdpe_encode gen8_pde_encode
226 #define gen8_pml4e_encode gen8_pde_encode
228 static gen6_pte_t snb_pte_encode(dma_addr_t addr,
229 enum i915_cache_level level,
230 bool valid, u32 unused)
232 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
233 pte |= GEN6_PTE_ADDR_ENCODE(addr);
236 case I915_CACHE_L3_LLC:
238 pte |= GEN6_PTE_CACHE_LLC;
240 case I915_CACHE_NONE:
241 pte |= GEN6_PTE_UNCACHED;
250 static gen6_pte_t ivb_pte_encode(dma_addr_t addr,
251 enum i915_cache_level level,
252 bool valid, u32 unused)
254 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
255 pte |= GEN6_PTE_ADDR_ENCODE(addr);
258 case I915_CACHE_L3_LLC:
259 pte |= GEN7_PTE_CACHE_L3_LLC;
262 pte |= GEN6_PTE_CACHE_LLC;
264 case I915_CACHE_NONE:
265 pte |= GEN6_PTE_UNCACHED;
274 static gen6_pte_t byt_pte_encode(dma_addr_t addr,
275 enum i915_cache_level level,
276 bool valid, u32 flags)
278 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
279 pte |= GEN6_PTE_ADDR_ENCODE(addr);
281 if (!(flags & PTE_READ_ONLY))
282 pte |= BYT_PTE_WRITEABLE;
284 if (level != I915_CACHE_NONE)
285 pte |= BYT_PTE_SNOOPED_BY_CPU_CACHES;
290 static gen6_pte_t hsw_pte_encode(dma_addr_t addr,
291 enum i915_cache_level level,
292 bool valid, u32 unused)
294 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
295 pte |= HSW_PTE_ADDR_ENCODE(addr);
297 if (level != I915_CACHE_NONE)
298 pte |= HSW_WB_LLC_AGE3;
303 static gen6_pte_t iris_pte_encode(dma_addr_t addr,
304 enum i915_cache_level level,
305 bool valid, u32 unused)
307 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
308 pte |= HSW_PTE_ADDR_ENCODE(addr);
311 case I915_CACHE_NONE:
314 pte |= HSW_WT_ELLC_LLC_AGE3;
317 pte |= HSW_WB_ELLC_LLC_AGE3;
324 static int __setup_page_dma(struct drm_device *dev,
325 struct i915_page_dma *p, gfp_t flags)
327 struct device *device = &dev->pdev->dev;
329 p->page = alloc_page(flags);
333 p->daddr = dma_map_page(device,
334 p->page, 0, 4096, PCI_DMA_BIDIRECTIONAL);
336 if (dma_mapping_error(device, p->daddr)) {
337 __free_page(p->page);
344 static int setup_page_dma(struct drm_device *dev, struct i915_page_dma *p)
346 return __setup_page_dma(dev, p, GFP_KERNEL);
349 static void cleanup_page_dma(struct drm_device *dev, struct i915_page_dma *p)
351 if (WARN_ON(!p->page))
354 dma_unmap_page(&dev->pdev->dev, p->daddr, 4096, PCI_DMA_BIDIRECTIONAL);
355 __free_page(p->page);
356 memset(p, 0, sizeof(*p));
359 static void *kmap_page_dma(struct i915_page_dma *p)
361 return kmap_atomic(p->page);
364 /* We use the flushing unmap only with ppgtt structures:
365 * page directories, page tables and scratch pages.
367 static void kunmap_page_dma(struct drm_device *dev, void *vaddr)
369 /* There are only few exceptions for gen >=6. chv and bxt.
370 * And we are not sure about the latter so play safe for now.
372 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
373 drm_clflush_virt_range(vaddr, PAGE_SIZE);
375 kunmap_atomic(vaddr);
378 #define kmap_px(px) kmap_page_dma(px_base(px))
379 #define kunmap_px(ppgtt, vaddr) kunmap_page_dma((ppgtt)->base.dev, (vaddr))
381 #define setup_px(dev, px) setup_page_dma((dev), px_base(px))
382 #define cleanup_px(dev, px) cleanup_page_dma((dev), px_base(px))
383 #define fill_px(dev, px, v) fill_page_dma((dev), px_base(px), (v))
384 #define fill32_px(dev, px, v) fill_page_dma_32((dev), px_base(px), (v))
386 static void fill_page_dma(struct drm_device *dev, struct i915_page_dma *p,
390 uint64_t * const vaddr = kmap_page_dma(p);
392 for (i = 0; i < 512; i++)
395 kunmap_page_dma(dev, vaddr);
398 static void fill_page_dma_32(struct drm_device *dev, struct i915_page_dma *p,
399 const uint32_t val32)
405 fill_page_dma(dev, p, v);
408 static struct i915_page_scratch *alloc_scratch_page(struct drm_device *dev)
410 struct i915_page_scratch *sp;
413 sp = kzalloc(sizeof(*sp), GFP_KERNEL);
415 return ERR_PTR(-ENOMEM);
417 ret = __setup_page_dma(dev, px_base(sp), GFP_DMA32 | __GFP_ZERO);
423 set_pages_uc(px_page(sp), 1);
428 static void free_scratch_page(struct drm_device *dev,
429 struct i915_page_scratch *sp)
431 set_pages_wb(px_page(sp), 1);
437 static struct i915_page_table *alloc_pt(struct drm_device *dev)
439 struct i915_page_table *pt;
440 const size_t count = INTEL_INFO(dev)->gen >= 8 ?
441 GEN8_PTES : GEN6_PTES;
444 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
446 return ERR_PTR(-ENOMEM);
448 pt->used_ptes = kcalloc(BITS_TO_LONGS(count), sizeof(*pt->used_ptes),
454 ret = setup_px(dev, pt);
461 kfree(pt->used_ptes);
468 static void free_pt(struct drm_device *dev, struct i915_page_table *pt)
471 kfree(pt->used_ptes);
475 static void gen8_initialize_pt(struct i915_address_space *vm,
476 struct i915_page_table *pt)
478 gen8_pte_t scratch_pte;
480 scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
481 I915_CACHE_LLC, true);
483 fill_px(vm->dev, pt, scratch_pte);
486 static void gen6_initialize_pt(struct i915_address_space *vm,
487 struct i915_page_table *pt)
489 gen6_pte_t scratch_pte;
491 WARN_ON(px_dma(vm->scratch_page) == 0);
493 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
494 I915_CACHE_LLC, true, 0);
496 fill32_px(vm->dev, pt, scratch_pte);
499 static struct i915_page_directory *alloc_pd(struct drm_device *dev)
501 struct i915_page_directory *pd;
504 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
506 return ERR_PTR(-ENOMEM);
508 pd->used_pdes = kcalloc(BITS_TO_LONGS(I915_PDES),
509 sizeof(*pd->used_pdes), GFP_KERNEL);
513 ret = setup_px(dev, pd);
520 kfree(pd->used_pdes);
527 static void free_pd(struct drm_device *dev, struct i915_page_directory *pd)
531 kfree(pd->used_pdes);
536 static void gen8_initialize_pd(struct i915_address_space *vm,
537 struct i915_page_directory *pd)
539 gen8_pde_t scratch_pde;
541 scratch_pde = gen8_pde_encode(px_dma(vm->scratch_pt), I915_CACHE_LLC);
543 fill_px(vm->dev, pd, scratch_pde);
546 static int __pdp_init(struct drm_device *dev,
547 struct i915_page_directory_pointer *pdp)
549 size_t pdpes = I915_PDPES_PER_PDP(dev);
551 pdp->used_pdpes = kcalloc(BITS_TO_LONGS(pdpes),
552 sizeof(unsigned long),
554 if (!pdp->used_pdpes)
557 pdp->page_directory = kcalloc(pdpes, sizeof(*pdp->page_directory),
559 if (!pdp->page_directory) {
560 kfree(pdp->used_pdpes);
561 /* the PDP might be the statically allocated top level. Keep it
562 * as clean as possible */
563 pdp->used_pdpes = NULL;
570 static void __pdp_fini(struct i915_page_directory_pointer *pdp)
572 kfree(pdp->used_pdpes);
573 kfree(pdp->page_directory);
574 pdp->page_directory = NULL;
578 i915_page_directory_pointer *alloc_pdp(struct drm_device *dev)
580 struct i915_page_directory_pointer *pdp;
583 WARN_ON(!USES_FULL_48BIT_PPGTT(dev));
585 pdp = kzalloc(sizeof(*pdp), GFP_KERNEL);
587 return ERR_PTR(-ENOMEM);
589 ret = __pdp_init(dev, pdp);
593 ret = setup_px(dev, pdp);
607 static void free_pdp(struct drm_device *dev,
608 struct i915_page_directory_pointer *pdp)
611 if (USES_FULL_48BIT_PPGTT(dev)) {
612 cleanup_px(dev, pdp);
617 static void gen8_initialize_pdp(struct i915_address_space *vm,
618 struct i915_page_directory_pointer *pdp)
620 gen8_ppgtt_pdpe_t scratch_pdpe;
622 scratch_pdpe = gen8_pdpe_encode(px_dma(vm->scratch_pd), I915_CACHE_LLC);
624 fill_px(vm->dev, pdp, scratch_pdpe);
627 static void gen8_initialize_pml4(struct i915_address_space *vm,
628 struct i915_pml4 *pml4)
630 gen8_ppgtt_pml4e_t scratch_pml4e;
632 scratch_pml4e = gen8_pml4e_encode(px_dma(vm->scratch_pdp),
635 fill_px(vm->dev, pml4, scratch_pml4e);
639 gen8_setup_page_directory(struct i915_hw_ppgtt *ppgtt,
640 struct i915_page_directory_pointer *pdp,
641 struct i915_page_directory *pd,
644 gen8_ppgtt_pdpe_t *page_directorypo;
646 if (!USES_FULL_48BIT_PPGTT(ppgtt->base.dev))
649 page_directorypo = kmap_px(pdp);
650 page_directorypo[index] = gen8_pdpe_encode(px_dma(pd), I915_CACHE_LLC);
651 kunmap_px(ppgtt, page_directorypo);
655 gen8_setup_page_directory_pointer(struct i915_hw_ppgtt *ppgtt,
656 struct i915_pml4 *pml4,
657 struct i915_page_directory_pointer *pdp,
660 gen8_ppgtt_pml4e_t *pagemap = kmap_px(pml4);
662 WARN_ON(!USES_FULL_48BIT_PPGTT(ppgtt->base.dev));
663 pagemap[index] = gen8_pml4e_encode(px_dma(pdp), I915_CACHE_LLC);
664 kunmap_px(ppgtt, pagemap);
667 /* Broadwell Page Directory Pointer Descriptors */
668 static int gen8_write_pdp(struct drm_i915_gem_request *req,
672 struct intel_ring *ring = req->ring;
673 struct intel_engine_cs *engine = req->engine;
678 ret = intel_ring_begin(req, 6);
682 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
683 intel_ring_emit_reg(ring, GEN8_RING_PDP_UDW(engine, entry));
684 intel_ring_emit(ring, upper_32_bits(addr));
685 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
686 intel_ring_emit_reg(ring, GEN8_RING_PDP_LDW(engine, entry));
687 intel_ring_emit(ring, lower_32_bits(addr));
688 intel_ring_advance(ring);
693 static int gen8_legacy_mm_switch(struct i915_hw_ppgtt *ppgtt,
694 struct drm_i915_gem_request *req)
698 for (i = GEN8_LEGACY_PDPES - 1; i >= 0; i--) {
699 const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i);
701 ret = gen8_write_pdp(req, i, pd_daddr);
709 static int gen8_48b_mm_switch(struct i915_hw_ppgtt *ppgtt,
710 struct drm_i915_gem_request *req)
712 return gen8_write_pdp(req, 0, px_dma(&ppgtt->pml4));
715 static void gen8_ppgtt_clear_pte_range(struct i915_address_space *vm,
716 struct i915_page_directory_pointer *pdp,
719 gen8_pte_t scratch_pte)
721 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
722 gen8_pte_t *pt_vaddr;
723 unsigned pdpe = gen8_pdpe_index(start);
724 unsigned pde = gen8_pde_index(start);
725 unsigned pte = gen8_pte_index(start);
726 unsigned num_entries = length >> PAGE_SHIFT;
727 unsigned last_pte, i;
732 while (num_entries) {
733 struct i915_page_directory *pd;
734 struct i915_page_table *pt;
736 if (WARN_ON(!pdp->page_directory[pdpe]))
739 pd = pdp->page_directory[pdpe];
741 if (WARN_ON(!pd->page_table[pde]))
744 pt = pd->page_table[pde];
746 if (WARN_ON(!px_page(pt)))
749 last_pte = pte + num_entries;
750 if (last_pte > GEN8_PTES)
751 last_pte = GEN8_PTES;
753 pt_vaddr = kmap_px(pt);
755 for (i = pte; i < last_pte; i++) {
756 pt_vaddr[i] = scratch_pte;
760 kunmap_px(ppgtt, pt_vaddr);
763 if (++pde == I915_PDES) {
764 if (++pdpe == I915_PDPES_PER_PDP(vm->dev))
771 static void gen8_ppgtt_clear_range(struct i915_address_space *vm,
776 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
777 gen8_pte_t scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
778 I915_CACHE_LLC, use_scratch);
780 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
781 gen8_ppgtt_clear_pte_range(vm, &ppgtt->pdp, start, length,
785 struct i915_page_directory_pointer *pdp;
787 gen8_for_each_pml4e(pdp, &ppgtt->pml4, start, length, pml4e) {
788 gen8_ppgtt_clear_pte_range(vm, pdp, start, length,
795 gen8_ppgtt_insert_pte_entries(struct i915_address_space *vm,
796 struct i915_page_directory_pointer *pdp,
797 struct sg_page_iter *sg_iter,
799 enum i915_cache_level cache_level)
801 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
802 gen8_pte_t *pt_vaddr;
803 unsigned pdpe = gen8_pdpe_index(start);
804 unsigned pde = gen8_pde_index(start);
805 unsigned pte = gen8_pte_index(start);
809 while (__sg_page_iter_next(sg_iter)) {
810 if (pt_vaddr == NULL) {
811 struct i915_page_directory *pd = pdp->page_directory[pdpe];
812 struct i915_page_table *pt = pd->page_table[pde];
813 pt_vaddr = kmap_px(pt);
817 gen8_pte_encode(sg_page_iter_dma_address(sg_iter),
819 if (++pte == GEN8_PTES) {
820 kunmap_px(ppgtt, pt_vaddr);
822 if (++pde == I915_PDES) {
823 if (++pdpe == I915_PDPES_PER_PDP(vm->dev))
832 kunmap_px(ppgtt, pt_vaddr);
835 static void gen8_ppgtt_insert_entries(struct i915_address_space *vm,
836 struct sg_table *pages,
838 enum i915_cache_level cache_level,
841 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
842 struct sg_page_iter sg_iter;
844 __sg_page_iter_start(&sg_iter, pages->sgl, sg_nents(pages->sgl), 0);
846 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
847 gen8_ppgtt_insert_pte_entries(vm, &ppgtt->pdp, &sg_iter, start,
850 struct i915_page_directory_pointer *pdp;
852 uint64_t length = (uint64_t)pages->orig_nents << PAGE_SHIFT;
854 gen8_for_each_pml4e(pdp, &ppgtt->pml4, start, length, pml4e) {
855 gen8_ppgtt_insert_pte_entries(vm, pdp, &sg_iter,
861 static void gen8_free_page_tables(struct drm_device *dev,
862 struct i915_page_directory *pd)
869 for_each_set_bit(i, pd->used_pdes, I915_PDES) {
870 if (WARN_ON(!pd->page_table[i]))
873 free_pt(dev, pd->page_table[i]);
874 pd->page_table[i] = NULL;
878 static int gen8_init_scratch(struct i915_address_space *vm)
880 struct drm_device *dev = vm->dev;
883 vm->scratch_page = alloc_scratch_page(dev);
884 if (IS_ERR(vm->scratch_page))
885 return PTR_ERR(vm->scratch_page);
887 vm->scratch_pt = alloc_pt(dev);
888 if (IS_ERR(vm->scratch_pt)) {
889 ret = PTR_ERR(vm->scratch_pt);
890 goto free_scratch_page;
893 vm->scratch_pd = alloc_pd(dev);
894 if (IS_ERR(vm->scratch_pd)) {
895 ret = PTR_ERR(vm->scratch_pd);
899 if (USES_FULL_48BIT_PPGTT(dev)) {
900 vm->scratch_pdp = alloc_pdp(dev);
901 if (IS_ERR(vm->scratch_pdp)) {
902 ret = PTR_ERR(vm->scratch_pdp);
907 gen8_initialize_pt(vm, vm->scratch_pt);
908 gen8_initialize_pd(vm, vm->scratch_pd);
909 if (USES_FULL_48BIT_PPGTT(dev))
910 gen8_initialize_pdp(vm, vm->scratch_pdp);
915 free_pd(dev, vm->scratch_pd);
917 free_pt(dev, vm->scratch_pt);
919 free_scratch_page(dev, vm->scratch_page);
924 static int gen8_ppgtt_notify_vgt(struct i915_hw_ppgtt *ppgtt, bool create)
926 enum vgt_g2v_type msg;
927 struct drm_i915_private *dev_priv = to_i915(ppgtt->base.dev);
930 if (USES_FULL_48BIT_PPGTT(dev_priv)) {
931 u64 daddr = px_dma(&ppgtt->pml4);
933 I915_WRITE(vgtif_reg(pdp[0].lo), lower_32_bits(daddr));
934 I915_WRITE(vgtif_reg(pdp[0].hi), upper_32_bits(daddr));
936 msg = (create ? VGT_G2V_PPGTT_L4_PAGE_TABLE_CREATE :
937 VGT_G2V_PPGTT_L4_PAGE_TABLE_DESTROY);
939 for (i = 0; i < GEN8_LEGACY_PDPES; i++) {
940 u64 daddr = i915_page_dir_dma_addr(ppgtt, i);
942 I915_WRITE(vgtif_reg(pdp[i].lo), lower_32_bits(daddr));
943 I915_WRITE(vgtif_reg(pdp[i].hi), upper_32_bits(daddr));
946 msg = (create ? VGT_G2V_PPGTT_L3_PAGE_TABLE_CREATE :
947 VGT_G2V_PPGTT_L3_PAGE_TABLE_DESTROY);
950 I915_WRITE(vgtif_reg(g2v_notify), msg);
955 static void gen8_free_scratch(struct i915_address_space *vm)
957 struct drm_device *dev = vm->dev;
959 if (USES_FULL_48BIT_PPGTT(dev))
960 free_pdp(dev, vm->scratch_pdp);
961 free_pd(dev, vm->scratch_pd);
962 free_pt(dev, vm->scratch_pt);
963 free_scratch_page(dev, vm->scratch_page);
966 static void gen8_ppgtt_cleanup_3lvl(struct drm_device *dev,
967 struct i915_page_directory_pointer *pdp)
971 for_each_set_bit(i, pdp->used_pdpes, I915_PDPES_PER_PDP(dev)) {
972 if (WARN_ON(!pdp->page_directory[i]))
975 gen8_free_page_tables(dev, pdp->page_directory[i]);
976 free_pd(dev, pdp->page_directory[i]);
982 static void gen8_ppgtt_cleanup_4lvl(struct i915_hw_ppgtt *ppgtt)
986 for_each_set_bit(i, ppgtt->pml4.used_pml4es, GEN8_PML4ES_PER_PML4) {
987 if (WARN_ON(!ppgtt->pml4.pdps[i]))
990 gen8_ppgtt_cleanup_3lvl(ppgtt->base.dev, ppgtt->pml4.pdps[i]);
993 cleanup_px(ppgtt->base.dev, &ppgtt->pml4);
996 static void gen8_ppgtt_cleanup(struct i915_address_space *vm)
998 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1000 if (intel_vgpu_active(to_i915(vm->dev)))
1001 gen8_ppgtt_notify_vgt(ppgtt, false);
1003 if (!USES_FULL_48BIT_PPGTT(ppgtt->base.dev))
1004 gen8_ppgtt_cleanup_3lvl(ppgtt->base.dev, &ppgtt->pdp);
1006 gen8_ppgtt_cleanup_4lvl(ppgtt);
1008 gen8_free_scratch(vm);
1012 * gen8_ppgtt_alloc_pagetabs() - Allocate page tables for VA range.
1013 * @vm: Master vm structure.
1014 * @pd: Page directory for this address range.
1015 * @start: Starting virtual address to begin allocations.
1016 * @length: Size of the allocations.
1017 * @new_pts: Bitmap set by function with new allocations. Likely used by the
1018 * caller to free on error.
1020 * Allocate the required number of page tables. Extremely similar to
1021 * gen8_ppgtt_alloc_page_directories(). The main difference is here we are limited by
1022 * the page directory boundary (instead of the page directory pointer). That
1023 * boundary is 1GB virtual. Therefore, unlike gen8_ppgtt_alloc_page_directories(), it is
1024 * possible, and likely that the caller will need to use multiple calls of this
1025 * function to achieve the appropriate allocation.
1027 * Return: 0 if success; negative error code otherwise.
1029 static int gen8_ppgtt_alloc_pagetabs(struct i915_address_space *vm,
1030 struct i915_page_directory *pd,
1033 unsigned long *new_pts)
1035 struct drm_device *dev = vm->dev;
1036 struct i915_page_table *pt;
1039 gen8_for_each_pde(pt, pd, start, length, pde) {
1040 /* Don't reallocate page tables */
1041 if (test_bit(pde, pd->used_pdes)) {
1042 /* Scratch is never allocated this way */
1043 WARN_ON(pt == vm->scratch_pt);
1051 gen8_initialize_pt(vm, pt);
1052 pd->page_table[pde] = pt;
1053 __set_bit(pde, new_pts);
1054 trace_i915_page_table_entry_alloc(vm, pde, start, GEN8_PDE_SHIFT);
1060 for_each_set_bit(pde, new_pts, I915_PDES)
1061 free_pt(dev, pd->page_table[pde]);
1067 * gen8_ppgtt_alloc_page_directories() - Allocate page directories for VA range.
1068 * @vm: Master vm structure.
1069 * @pdp: Page directory pointer for this address range.
1070 * @start: Starting virtual address to begin allocations.
1071 * @length: Size of the allocations.
1072 * @new_pds: Bitmap set by function with new allocations. Likely used by the
1073 * caller to free on error.
1075 * Allocate the required number of page directories starting at the pde index of
1076 * @start, and ending at the pde index @start + @length. This function will skip
1077 * over already allocated page directories within the range, and only allocate
1078 * new ones, setting the appropriate pointer within the pdp as well as the
1079 * correct position in the bitmap @new_pds.
1081 * The function will only allocate the pages within the range for a give page
1082 * directory pointer. In other words, if @start + @length straddles a virtually
1083 * addressed PDP boundary (512GB for 4k pages), there will be more allocations
1084 * required by the caller, This is not currently possible, and the BUG in the
1085 * code will prevent it.
1087 * Return: 0 if success; negative error code otherwise.
1090 gen8_ppgtt_alloc_page_directories(struct i915_address_space *vm,
1091 struct i915_page_directory_pointer *pdp,
1094 unsigned long *new_pds)
1096 struct drm_device *dev = vm->dev;
1097 struct i915_page_directory *pd;
1099 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1101 WARN_ON(!bitmap_empty(new_pds, pdpes));
1103 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1104 if (test_bit(pdpe, pdp->used_pdpes))
1111 gen8_initialize_pd(vm, pd);
1112 pdp->page_directory[pdpe] = pd;
1113 __set_bit(pdpe, new_pds);
1114 trace_i915_page_directory_entry_alloc(vm, pdpe, start, GEN8_PDPE_SHIFT);
1120 for_each_set_bit(pdpe, new_pds, pdpes)
1121 free_pd(dev, pdp->page_directory[pdpe]);
1127 * gen8_ppgtt_alloc_page_dirpointers() - Allocate pdps for VA range.
1128 * @vm: Master vm structure.
1129 * @pml4: Page map level 4 for this address range.
1130 * @start: Starting virtual address to begin allocations.
1131 * @length: Size of the allocations.
1132 * @new_pdps: Bitmap set by function with new allocations. Likely used by the
1133 * caller to free on error.
1135 * Allocate the required number of page directory pointers. Extremely similar to
1136 * gen8_ppgtt_alloc_page_directories() and gen8_ppgtt_alloc_pagetabs().
1137 * The main difference is here we are limited by the pml4 boundary (instead of
1138 * the page directory pointer).
1140 * Return: 0 if success; negative error code otherwise.
1143 gen8_ppgtt_alloc_page_dirpointers(struct i915_address_space *vm,
1144 struct i915_pml4 *pml4,
1147 unsigned long *new_pdps)
1149 struct drm_device *dev = vm->dev;
1150 struct i915_page_directory_pointer *pdp;
1153 WARN_ON(!bitmap_empty(new_pdps, GEN8_PML4ES_PER_PML4));
1155 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1156 if (!test_bit(pml4e, pml4->used_pml4es)) {
1157 pdp = alloc_pdp(dev);
1161 gen8_initialize_pdp(vm, pdp);
1162 pml4->pdps[pml4e] = pdp;
1163 __set_bit(pml4e, new_pdps);
1164 trace_i915_page_directory_pointer_entry_alloc(vm,
1174 for_each_set_bit(pml4e, new_pdps, GEN8_PML4ES_PER_PML4)
1175 free_pdp(dev, pml4->pdps[pml4e]);
1181 free_gen8_temp_bitmaps(unsigned long *new_pds, unsigned long *new_pts)
1187 /* Fills in the page directory bitmap, and the array of page tables bitmap. Both
1188 * of these are based on the number of PDPEs in the system.
1191 int __must_check alloc_gen8_temp_bitmaps(unsigned long **new_pds,
1192 unsigned long **new_pts,
1198 pds = kcalloc(BITS_TO_LONGS(pdpes), sizeof(unsigned long), GFP_TEMPORARY);
1202 pts = kcalloc(pdpes, BITS_TO_LONGS(I915_PDES) * sizeof(unsigned long),
1213 free_gen8_temp_bitmaps(pds, pts);
1217 /* PDE TLBs are a pain to invalidate on GEN8+. When we modify
1218 * the page table structures, we mark them dirty so that
1219 * context switching/execlist queuing code takes extra steps
1220 * to ensure that tlbs are flushed.
1222 static void mark_tlbs_dirty(struct i915_hw_ppgtt *ppgtt)
1224 ppgtt->pd_dirty_rings = INTEL_INFO(ppgtt->base.dev)->ring_mask;
1227 static int gen8_alloc_va_range_3lvl(struct i915_address_space *vm,
1228 struct i915_page_directory_pointer *pdp,
1232 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1233 unsigned long *new_page_dirs, *new_page_tables;
1234 struct drm_device *dev = vm->dev;
1235 struct i915_page_directory *pd;
1236 const uint64_t orig_start = start;
1237 const uint64_t orig_length = length;
1239 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1242 /* Wrap is never okay since we can only represent 48b, and we don't
1243 * actually use the other side of the canonical address space.
1245 if (WARN_ON(start + length < start))
1248 if (WARN_ON(start + length > vm->total))
1251 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables, pdpes);
1255 /* Do the allocations first so we can easily bail out */
1256 ret = gen8_ppgtt_alloc_page_directories(vm, pdp, start, length,
1259 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1263 /* For every page directory referenced, allocate page tables */
1264 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1265 ret = gen8_ppgtt_alloc_pagetabs(vm, pd, start, length,
1266 new_page_tables + pdpe * BITS_TO_LONGS(I915_PDES));
1272 length = orig_length;
1274 /* Allocations have completed successfully, so set the bitmaps, and do
1276 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1277 gen8_pde_t *const page_directory = kmap_px(pd);
1278 struct i915_page_table *pt;
1279 uint64_t pd_len = length;
1280 uint64_t pd_start = start;
1283 /* Every pd should be allocated, we just did that above. */
1286 gen8_for_each_pde(pt, pd, pd_start, pd_len, pde) {
1287 /* Same reasoning as pd */
1290 WARN_ON(!gen8_pte_count(pd_start, pd_len));
1292 /* Set our used ptes within the page table */
1293 bitmap_set(pt->used_ptes,
1294 gen8_pte_index(pd_start),
1295 gen8_pte_count(pd_start, pd_len));
1297 /* Our pde is now pointing to the pagetable, pt */
1298 __set_bit(pde, pd->used_pdes);
1300 /* Map the PDE to the page table */
1301 page_directory[pde] = gen8_pde_encode(px_dma(pt),
1303 trace_i915_page_table_entry_map(&ppgtt->base, pde, pt,
1304 gen8_pte_index(start),
1305 gen8_pte_count(start, length),
1308 /* NB: We haven't yet mapped ptes to pages. At this
1309 * point we're still relying on insert_entries() */
1312 kunmap_px(ppgtt, page_directory);
1313 __set_bit(pdpe, pdp->used_pdpes);
1314 gen8_setup_page_directory(ppgtt, pdp, pd, pdpe);
1317 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1318 mark_tlbs_dirty(ppgtt);
1325 for_each_set_bit(temp, new_page_tables + pdpe *
1326 BITS_TO_LONGS(I915_PDES), I915_PDES)
1327 free_pt(dev, pdp->page_directory[pdpe]->page_table[temp]);
1330 for_each_set_bit(pdpe, new_page_dirs, pdpes)
1331 free_pd(dev, pdp->page_directory[pdpe]);
1333 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1334 mark_tlbs_dirty(ppgtt);
1338 static int gen8_alloc_va_range_4lvl(struct i915_address_space *vm,
1339 struct i915_pml4 *pml4,
1343 DECLARE_BITMAP(new_pdps, GEN8_PML4ES_PER_PML4);
1344 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1345 struct i915_page_directory_pointer *pdp;
1349 /* Do the pml4 allocations first, so we don't need to track the newly
1350 * allocated tables below the pdp */
1351 bitmap_zero(new_pdps, GEN8_PML4ES_PER_PML4);
1353 /* The pagedirectory and pagetable allocations are done in the shared 3
1354 * and 4 level code. Just allocate the pdps.
1356 ret = gen8_ppgtt_alloc_page_dirpointers(vm, pml4, start, length,
1361 WARN(bitmap_weight(new_pdps, GEN8_PML4ES_PER_PML4) > 2,
1362 "The allocation has spanned more than 512GB. "
1363 "It is highly likely this is incorrect.");
1365 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1368 ret = gen8_alloc_va_range_3lvl(vm, pdp, start, length);
1372 gen8_setup_page_directory_pointer(ppgtt, pml4, pdp, pml4e);
1375 bitmap_or(pml4->used_pml4es, new_pdps, pml4->used_pml4es,
1376 GEN8_PML4ES_PER_PML4);
1381 for_each_set_bit(pml4e, new_pdps, GEN8_PML4ES_PER_PML4)
1382 gen8_ppgtt_cleanup_3lvl(vm->dev, pml4->pdps[pml4e]);
1387 static int gen8_alloc_va_range(struct i915_address_space *vm,
1388 uint64_t start, uint64_t length)
1390 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1392 if (USES_FULL_48BIT_PPGTT(vm->dev))
1393 return gen8_alloc_va_range_4lvl(vm, &ppgtt->pml4, start, length);
1395 return gen8_alloc_va_range_3lvl(vm, &ppgtt->pdp, start, length);
1398 static void gen8_dump_pdp(struct i915_page_directory_pointer *pdp,
1399 uint64_t start, uint64_t length,
1400 gen8_pte_t scratch_pte,
1403 struct i915_page_directory *pd;
1406 gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1407 struct i915_page_table *pt;
1408 uint64_t pd_len = length;
1409 uint64_t pd_start = start;
1412 if (!test_bit(pdpe, pdp->used_pdpes))
1415 seq_printf(m, "\tPDPE #%d\n", pdpe);
1416 gen8_for_each_pde(pt, pd, pd_start, pd_len, pde) {
1418 gen8_pte_t *pt_vaddr;
1420 if (!test_bit(pde, pd->used_pdes))
1423 pt_vaddr = kmap_px(pt);
1424 for (pte = 0; pte < GEN8_PTES; pte += 4) {
1426 (pdpe << GEN8_PDPE_SHIFT) |
1427 (pde << GEN8_PDE_SHIFT) |
1428 (pte << GEN8_PTE_SHIFT);
1432 for (i = 0; i < 4; i++)
1433 if (pt_vaddr[pte + i] != scratch_pte)
1438 seq_printf(m, "\t\t0x%lx [%03d,%03d,%04d]: =", va, pdpe, pde, pte);
1439 for (i = 0; i < 4; i++) {
1440 if (pt_vaddr[pte + i] != scratch_pte)
1441 seq_printf(m, " %lx", pt_vaddr[pte + i]);
1443 seq_puts(m, " SCRATCH ");
1447 /* don't use kunmap_px, it could trigger
1448 * an unnecessary flush.
1450 kunmap_atomic(pt_vaddr);
1455 static void gen8_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
1457 struct i915_address_space *vm = &ppgtt->base;
1458 uint64_t start = ppgtt->base.start;
1459 uint64_t length = ppgtt->base.total;
1460 gen8_pte_t scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
1461 I915_CACHE_LLC, true);
1463 if (!USES_FULL_48BIT_PPGTT(vm->dev)) {
1464 gen8_dump_pdp(&ppgtt->pdp, start, length, scratch_pte, m);
1467 struct i915_pml4 *pml4 = &ppgtt->pml4;
1468 struct i915_page_directory_pointer *pdp;
1470 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1471 if (!test_bit(pml4e, pml4->used_pml4es))
1474 seq_printf(m, " PML4E #%lu\n", pml4e);
1475 gen8_dump_pdp(pdp, start, length, scratch_pte, m);
1480 static int gen8_preallocate_top_level_pdps(struct i915_hw_ppgtt *ppgtt)
1482 unsigned long *new_page_dirs, *new_page_tables;
1483 uint32_t pdpes = I915_PDPES_PER_PDP(dev);
1486 /* We allocate temp bitmap for page tables for no gain
1487 * but as this is for init only, lets keep the things simple
1489 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables, pdpes);
1493 /* Allocate for all pdps regardless of how the ppgtt
1496 ret = gen8_ppgtt_alloc_page_directories(&ppgtt->base, &ppgtt->pdp,
1500 *ppgtt->pdp.used_pdpes = *new_page_dirs;
1502 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1508 * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers
1509 * with a net effect resembling a 2-level page table in normal x86 terms. Each
1510 * PDP represents 1GB of memory 4 * 512 * 512 * 4096 = 4GB legacy 32b address
1514 static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
1518 ret = gen8_init_scratch(&ppgtt->base);
1522 ppgtt->base.start = 0;
1523 ppgtt->base.cleanup = gen8_ppgtt_cleanup;
1524 ppgtt->base.allocate_va_range = gen8_alloc_va_range;
1525 ppgtt->base.insert_entries = gen8_ppgtt_insert_entries;
1526 ppgtt->base.clear_range = gen8_ppgtt_clear_range;
1527 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
1528 ppgtt->base.bind_vma = ppgtt_bind_vma;
1529 ppgtt->debug_dump = gen8_dump_ppgtt;
1531 if (USES_FULL_48BIT_PPGTT(ppgtt->base.dev)) {
1532 ret = setup_px(ppgtt->base.dev, &ppgtt->pml4);
1536 gen8_initialize_pml4(&ppgtt->base, &ppgtt->pml4);
1538 ppgtt->base.total = 1ULL << 48;
1539 ppgtt->switch_mm = gen8_48b_mm_switch;
1541 ret = __pdp_init(ppgtt->base.dev, &ppgtt->pdp);
1545 ppgtt->base.total = 1ULL << 32;
1546 ppgtt->switch_mm = gen8_legacy_mm_switch;
1547 trace_i915_page_directory_pointer_entry_alloc(&ppgtt->base,
1551 if (intel_vgpu_active(to_i915(ppgtt->base.dev))) {
1552 ret = gen8_preallocate_top_level_pdps(ppgtt);
1558 if (intel_vgpu_active(to_i915(ppgtt->base.dev)))
1559 gen8_ppgtt_notify_vgt(ppgtt, true);
1564 gen8_free_scratch(&ppgtt->base);
1568 static void gen6_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
1570 struct i915_address_space *vm = &ppgtt->base;
1571 struct i915_page_table *unused;
1572 gen6_pte_t scratch_pte;
1575 uint32_t start = ppgtt->base.start, length = ppgtt->base.total;
1577 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1578 I915_CACHE_LLC, true, 0);
1580 gen6_for_each_pde(unused, &ppgtt->pd, start, length, pde) {
1582 gen6_pte_t *pt_vaddr;
1583 const dma_addr_t pt_addr = px_dma(ppgtt->pd.page_table[pde]);
1584 pd_entry = readl(ppgtt->pd_addr + pde);
1585 expected = (GEN6_PDE_ADDR_ENCODE(pt_addr) | GEN6_PDE_VALID);
1587 if (pd_entry != expected)
1588 seq_printf(m, "\tPDE #%d mismatch: Actual PDE: %x Expected PDE: %x\n",
1592 seq_printf(m, "\tPDE: %x\n", pd_entry);
1594 pt_vaddr = kmap_px(ppgtt->pd.page_table[pde]);
1596 for (pte = 0; pte < GEN6_PTES; pte+=4) {
1598 (pde * PAGE_SIZE * GEN6_PTES) +
1602 for (i = 0; i < 4; i++)
1603 if (pt_vaddr[pte + i] != scratch_pte)
1608 seq_printf(m, "\t\t0x%lx [%03d,%04d]: =", va, pde, pte);
1609 for (i = 0; i < 4; i++) {
1610 if (pt_vaddr[pte + i] != scratch_pte)
1611 seq_printf(m, " %08x", pt_vaddr[pte + i]);
1613 seq_puts(m, " SCRATCH ");
1617 kunmap_px(ppgtt, pt_vaddr);
1621 /* Write pde (index) from the page directory @pd to the page table @pt */
1622 static void gen6_write_pde(struct i915_page_directory *pd,
1623 const int pde, struct i915_page_table *pt)
1625 /* Caller needs to make sure the write completes if necessary */
1626 struct i915_hw_ppgtt *ppgtt =
1627 container_of(pd, struct i915_hw_ppgtt, pd);
1630 pd_entry = GEN6_PDE_ADDR_ENCODE(px_dma(pt));
1631 pd_entry |= GEN6_PDE_VALID;
1633 writel(pd_entry, ppgtt->pd_addr + pde);
1636 /* Write all the page tables found in the ppgtt structure to incrementing page
1638 static void gen6_write_page_range(struct drm_i915_private *dev_priv,
1639 struct i915_page_directory *pd,
1640 uint32_t start, uint32_t length)
1642 struct i915_ggtt *ggtt = &dev_priv->ggtt;
1643 struct i915_page_table *pt;
1646 gen6_for_each_pde(pt, pd, start, length, pde)
1647 gen6_write_pde(pd, pde, pt);
1649 /* Make sure write is complete before other code can use this page
1650 * table. Also require for WC mapped PTEs */
1654 static uint32_t get_pd_offset(struct i915_hw_ppgtt *ppgtt)
1656 BUG_ON(ppgtt->pd.base.ggtt_offset & 0x3f);
1658 return (ppgtt->pd.base.ggtt_offset / 64) << 16;
1661 static int hsw_mm_switch(struct i915_hw_ppgtt *ppgtt,
1662 struct drm_i915_gem_request *req)
1664 struct intel_ring *ring = req->ring;
1665 struct intel_engine_cs *engine = req->engine;
1668 /* NB: TLBs must be flushed and invalidated before a switch */
1669 ret = engine->emit_flush(req, EMIT_INVALIDATE | EMIT_FLUSH);
1673 ret = intel_ring_begin(req, 6);
1677 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1678 intel_ring_emit_reg(ring, RING_PP_DIR_DCLV(engine));
1679 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1680 intel_ring_emit_reg(ring, RING_PP_DIR_BASE(engine));
1681 intel_ring_emit(ring, get_pd_offset(ppgtt));
1682 intel_ring_emit(ring, MI_NOOP);
1683 intel_ring_advance(ring);
1688 static int gen7_mm_switch(struct i915_hw_ppgtt *ppgtt,
1689 struct drm_i915_gem_request *req)
1691 struct intel_ring *ring = req->ring;
1692 struct intel_engine_cs *engine = req->engine;
1695 /* NB: TLBs must be flushed and invalidated before a switch */
1696 ret = engine->emit_flush(req, EMIT_INVALIDATE | EMIT_FLUSH);
1700 ret = intel_ring_begin(req, 6);
1704 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1705 intel_ring_emit_reg(ring, RING_PP_DIR_DCLV(engine));
1706 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1707 intel_ring_emit_reg(ring, RING_PP_DIR_BASE(engine));
1708 intel_ring_emit(ring, get_pd_offset(ppgtt));
1709 intel_ring_emit(ring, MI_NOOP);
1710 intel_ring_advance(ring);
1712 /* XXX: RCS is the only one to auto invalidate the TLBs? */
1713 if (engine->id != RCS) {
1714 ret = engine->emit_flush(req, EMIT_INVALIDATE | EMIT_FLUSH);
1722 static int gen6_mm_switch(struct i915_hw_ppgtt *ppgtt,
1723 struct drm_i915_gem_request *req)
1725 struct intel_engine_cs *engine = req->engine;
1726 struct drm_i915_private *dev_priv = req->i915;
1728 I915_WRITE(RING_PP_DIR_DCLV(engine), PP_DIR_DCLV_2G);
1729 I915_WRITE(RING_PP_DIR_BASE(engine), get_pd_offset(ppgtt));
1733 static void gen8_ppgtt_enable(struct drm_device *dev)
1735 struct drm_i915_private *dev_priv = to_i915(dev);
1736 struct intel_engine_cs *engine;
1738 for_each_engine(engine, dev_priv) {
1739 u32 four_level = USES_FULL_48BIT_PPGTT(dev) ? GEN8_GFX_PPGTT_48B : 0;
1740 I915_WRITE(RING_MODE_GEN7(engine),
1741 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE | four_level));
1745 static void gen7_ppgtt_enable(struct drm_device *dev)
1747 struct drm_i915_private *dev_priv = to_i915(dev);
1748 struct intel_engine_cs *engine;
1749 uint32_t ecochk, ecobits;
1751 ecobits = I915_READ(GAC_ECO_BITS);
1752 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B);
1754 ecochk = I915_READ(GAM_ECOCHK);
1755 if (IS_HASWELL(dev)) {
1756 ecochk |= ECOCHK_PPGTT_WB_HSW;
1758 ecochk |= ECOCHK_PPGTT_LLC_IVB;
1759 ecochk &= ~ECOCHK_PPGTT_GFDT_IVB;
1761 I915_WRITE(GAM_ECOCHK, ecochk);
1763 for_each_engine(engine, dev_priv) {
1764 /* GFX_MODE is per-ring on gen7+ */
1765 I915_WRITE(RING_MODE_GEN7(engine),
1766 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1770 static void gen6_ppgtt_enable(struct drm_device *dev)
1772 struct drm_i915_private *dev_priv = to_i915(dev);
1773 uint32_t ecochk, gab_ctl, ecobits;
1775 ecobits = I915_READ(GAC_ECO_BITS);
1776 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_SNB_BIT |
1777 ECOBITS_PPGTT_CACHE64B);
1779 gab_ctl = I915_READ(GAB_CTL);
1780 I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT);
1782 ecochk = I915_READ(GAM_ECOCHK);
1783 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT | ECOCHK_PPGTT_CACHE64B);
1785 I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1788 /* PPGTT support for Sandybdrige/Gen6 and later */
1789 static void gen6_ppgtt_clear_range(struct i915_address_space *vm,
1794 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1795 gen6_pte_t *pt_vaddr, scratch_pte;
1796 unsigned first_entry = start >> PAGE_SHIFT;
1797 unsigned num_entries = length >> PAGE_SHIFT;
1798 unsigned act_pt = first_entry / GEN6_PTES;
1799 unsigned first_pte = first_entry % GEN6_PTES;
1800 unsigned last_pte, i;
1802 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1803 I915_CACHE_LLC, true, 0);
1805 while (num_entries) {
1806 last_pte = first_pte + num_entries;
1807 if (last_pte > GEN6_PTES)
1808 last_pte = GEN6_PTES;
1810 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1812 for (i = first_pte; i < last_pte; i++)
1813 pt_vaddr[i] = scratch_pte;
1815 kunmap_px(ppgtt, pt_vaddr);
1817 num_entries -= last_pte - first_pte;
1823 static void gen6_ppgtt_insert_entries(struct i915_address_space *vm,
1824 struct sg_table *pages,
1826 enum i915_cache_level cache_level, u32 flags)
1828 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1829 unsigned first_entry = start >> PAGE_SHIFT;
1830 unsigned act_pt = first_entry / GEN6_PTES;
1831 unsigned act_pte = first_entry % GEN6_PTES;
1832 gen6_pte_t *pt_vaddr = NULL;
1833 struct sgt_iter sgt_iter;
1836 for_each_sgt_dma(addr, sgt_iter, pages) {
1837 if (pt_vaddr == NULL)
1838 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1841 vm->pte_encode(addr, cache_level, true, flags);
1843 if (++act_pte == GEN6_PTES) {
1844 kunmap_px(ppgtt, pt_vaddr);
1852 kunmap_px(ppgtt, pt_vaddr);
1855 static int gen6_alloc_va_range(struct i915_address_space *vm,
1856 uint64_t start_in, uint64_t length_in)
1858 DECLARE_BITMAP(new_page_tables, I915_PDES);
1859 struct drm_device *dev = vm->dev;
1860 struct drm_i915_private *dev_priv = to_i915(dev);
1861 struct i915_ggtt *ggtt = &dev_priv->ggtt;
1862 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1863 struct i915_page_table *pt;
1864 uint32_t start, length, start_save, length_save;
1868 if (WARN_ON(start_in + length_in > ppgtt->base.total))
1871 start = start_save = start_in;
1872 length = length_save = length_in;
1874 bitmap_zero(new_page_tables, I915_PDES);
1876 /* The allocation is done in two stages so that we can bail out with
1877 * minimal amount of pain. The first stage finds new page tables that
1878 * need allocation. The second stage marks use ptes within the page
1881 gen6_for_each_pde(pt, &ppgtt->pd, start, length, pde) {
1882 if (pt != vm->scratch_pt) {
1883 WARN_ON(bitmap_empty(pt->used_ptes, GEN6_PTES));
1887 /* We've already allocated a page table */
1888 WARN_ON(!bitmap_empty(pt->used_ptes, GEN6_PTES));
1896 gen6_initialize_pt(vm, pt);
1898 ppgtt->pd.page_table[pde] = pt;
1899 __set_bit(pde, new_page_tables);
1900 trace_i915_page_table_entry_alloc(vm, pde, start, GEN6_PDE_SHIFT);
1904 length = length_save;
1906 gen6_for_each_pde(pt, &ppgtt->pd, start, length, pde) {
1907 DECLARE_BITMAP(tmp_bitmap, GEN6_PTES);
1909 bitmap_zero(tmp_bitmap, GEN6_PTES);
1910 bitmap_set(tmp_bitmap, gen6_pte_index(start),
1911 gen6_pte_count(start, length));
1913 if (__test_and_clear_bit(pde, new_page_tables))
1914 gen6_write_pde(&ppgtt->pd, pde, pt);
1916 trace_i915_page_table_entry_map(vm, pde, pt,
1917 gen6_pte_index(start),
1918 gen6_pte_count(start, length),
1920 bitmap_or(pt->used_ptes, tmp_bitmap, pt->used_ptes,
1924 WARN_ON(!bitmap_empty(new_page_tables, I915_PDES));
1926 /* Make sure write is complete before other code can use this page
1927 * table. Also require for WC mapped PTEs */
1930 mark_tlbs_dirty(ppgtt);
1934 for_each_set_bit(pde, new_page_tables, I915_PDES) {
1935 struct i915_page_table *pt = ppgtt->pd.page_table[pde];
1937 ppgtt->pd.page_table[pde] = vm->scratch_pt;
1938 free_pt(vm->dev, pt);
1941 mark_tlbs_dirty(ppgtt);
1945 static int gen6_init_scratch(struct i915_address_space *vm)
1947 struct drm_device *dev = vm->dev;
1949 vm->scratch_page = alloc_scratch_page(dev);
1950 if (IS_ERR(vm->scratch_page))
1951 return PTR_ERR(vm->scratch_page);
1953 vm->scratch_pt = alloc_pt(dev);
1954 if (IS_ERR(vm->scratch_pt)) {
1955 free_scratch_page(dev, vm->scratch_page);
1956 return PTR_ERR(vm->scratch_pt);
1959 gen6_initialize_pt(vm, vm->scratch_pt);
1964 static void gen6_free_scratch(struct i915_address_space *vm)
1966 struct drm_device *dev = vm->dev;
1968 free_pt(dev, vm->scratch_pt);
1969 free_scratch_page(dev, vm->scratch_page);
1972 static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
1974 struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1975 struct i915_page_directory *pd = &ppgtt->pd;
1976 struct drm_device *dev = vm->dev;
1977 struct i915_page_table *pt;
1980 drm_mm_remove_node(&ppgtt->node);
1982 gen6_for_all_pdes(pt, pd, pde)
1983 if (pt != vm->scratch_pt)
1986 gen6_free_scratch(vm);
1989 static int gen6_ppgtt_allocate_page_directories(struct i915_hw_ppgtt *ppgtt)
1991 struct i915_address_space *vm = &ppgtt->base;
1992 struct drm_device *dev = ppgtt->base.dev;
1993 struct drm_i915_private *dev_priv = to_i915(dev);
1994 struct i915_ggtt *ggtt = &dev_priv->ggtt;
1995 bool retried = false;
1998 /* PPGTT PDEs reside in the GGTT and consists of 512 entries. The
1999 * allocator works in address space sizes, so it's multiplied by page
2000 * size. We allocate at the top of the GTT to avoid fragmentation.
2002 BUG_ON(!drm_mm_initialized(&ggtt->base.mm));
2004 ret = gen6_init_scratch(vm);
2009 ret = drm_mm_insert_node_in_range_generic(&ggtt->base.mm,
2010 &ppgtt->node, GEN6_PD_SIZE,
2012 0, ggtt->base.total,
2014 if (ret == -ENOSPC && !retried) {
2015 ret = i915_gem_evict_something(&ggtt->base,
2016 GEN6_PD_SIZE, GEN6_PD_ALIGN,
2018 0, ggtt->base.total,
2031 if (ppgtt->node.start < ggtt->mappable_end)
2032 DRM_DEBUG("Forced to use aperture for PDEs\n");
2037 gen6_free_scratch(vm);
2041 static int gen6_ppgtt_alloc(struct i915_hw_ppgtt *ppgtt)
2043 return gen6_ppgtt_allocate_page_directories(ppgtt);
2046 static void gen6_scratch_va_range(struct i915_hw_ppgtt *ppgtt,
2047 uint64_t start, uint64_t length)
2049 struct i915_page_table *unused;
2052 gen6_for_each_pde(unused, &ppgtt->pd, start, length, pde)
2053 ppgtt->pd.page_table[pde] = ppgtt->base.scratch_pt;
2056 static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
2058 struct drm_device *dev = ppgtt->base.dev;
2059 struct drm_i915_private *dev_priv = to_i915(dev);
2060 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2063 ppgtt->base.pte_encode = ggtt->base.pte_encode;
2064 if (intel_vgpu_active(dev_priv) || IS_GEN6(dev))
2065 ppgtt->switch_mm = gen6_mm_switch;
2066 else if (IS_HASWELL(dev))
2067 ppgtt->switch_mm = hsw_mm_switch;
2068 else if (IS_GEN7(dev))
2069 ppgtt->switch_mm = gen7_mm_switch;
2073 ret = gen6_ppgtt_alloc(ppgtt);
2077 ppgtt->base.allocate_va_range = gen6_alloc_va_range;
2078 ppgtt->base.clear_range = gen6_ppgtt_clear_range;
2079 ppgtt->base.insert_entries = gen6_ppgtt_insert_entries;
2080 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
2081 ppgtt->base.bind_vma = ppgtt_bind_vma;
2082 ppgtt->base.cleanup = gen6_ppgtt_cleanup;
2083 ppgtt->base.start = 0;
2084 ppgtt->base.total = I915_PDES * GEN6_PTES * PAGE_SIZE;
2085 ppgtt->debug_dump = gen6_dump_ppgtt;
2087 ppgtt->pd.base.ggtt_offset =
2088 ppgtt->node.start / PAGE_SIZE * sizeof(gen6_pte_t);
2090 ppgtt->pd_addr = (gen6_pte_t __iomem *)ggtt->gsm +
2091 ppgtt->pd.base.ggtt_offset / sizeof(gen6_pte_t);
2093 gen6_scratch_va_range(ppgtt, 0, ppgtt->base.total);
2095 gen6_write_page_range(dev_priv, &ppgtt->pd, 0, ppgtt->base.total);
2097 DRM_DEBUG_DRIVER("Allocated pde space (%lldM) at GTT entry: %llx\n",
2098 ppgtt->node.size >> 20,
2099 ppgtt->node.start / PAGE_SIZE);
2101 DRM_DEBUG("Adding PPGTT at offset %x\n",
2102 ppgtt->pd.base.ggtt_offset << 10);
2107 static int __hw_ppgtt_init(struct i915_hw_ppgtt *ppgtt,
2108 struct drm_i915_private *dev_priv)
2110 ppgtt->base.dev = &dev_priv->drm;
2112 if (INTEL_INFO(dev_priv)->gen < 8)
2113 return gen6_ppgtt_init(ppgtt);
2115 return gen8_ppgtt_init(ppgtt);
2118 static void i915_address_space_init(struct i915_address_space *vm,
2119 struct drm_i915_private *dev_priv)
2121 drm_mm_init(&vm->mm, vm->start, vm->total);
2122 INIT_LIST_HEAD(&vm->active_list);
2123 INIT_LIST_HEAD(&vm->inactive_list);
2124 INIT_LIST_HEAD(&vm->unbound_list);
2125 list_add_tail(&vm->global_link, &dev_priv->vm_list);
2128 static void gtt_write_workarounds(struct drm_device *dev)
2130 struct drm_i915_private *dev_priv = to_i915(dev);
2132 /* This function is for gtt related workarounds. This function is
2133 * called on driver load and after a GPU reset, so you can place
2134 * workarounds here even if they get overwritten by GPU reset.
2136 /* WaIncreaseDefaultTLBEntries:chv,bdw,skl,bxt */
2137 if (IS_BROADWELL(dev))
2138 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN8_L3_LRA_1_GPGPU_DEFAULT_VALUE_BDW);
2139 else if (IS_CHERRYVIEW(dev))
2140 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN8_L3_LRA_1_GPGPU_DEFAULT_VALUE_CHV);
2141 else if (IS_SKYLAKE(dev))
2142 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN9_L3_LRA_1_GPGPU_DEFAULT_VALUE_SKL);
2143 else if (IS_BROXTON(dev))
2144 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN9_L3_LRA_1_GPGPU_DEFAULT_VALUE_BXT);
2147 static int i915_ppgtt_init(struct i915_hw_ppgtt *ppgtt,
2148 struct drm_i915_private *dev_priv,
2149 struct drm_i915_file_private *file_priv)
2153 ret = __hw_ppgtt_init(ppgtt, dev_priv);
2155 kref_init(&ppgtt->ref);
2156 i915_address_space_init(&ppgtt->base, dev_priv);
2157 ppgtt->base.file = file_priv;
2163 int i915_ppgtt_init_hw(struct drm_device *dev)
2165 gtt_write_workarounds(dev);
2167 /* In the case of execlists, PPGTT is enabled by the context descriptor
2168 * and the PDPs are contained within the context itself. We don't
2169 * need to do anything here. */
2170 if (i915.enable_execlists)
2173 if (!USES_PPGTT(dev))
2177 gen6_ppgtt_enable(dev);
2178 else if (IS_GEN7(dev))
2179 gen7_ppgtt_enable(dev);
2180 else if (INTEL_INFO(dev)->gen >= 8)
2181 gen8_ppgtt_enable(dev);
2183 MISSING_CASE(INTEL_INFO(dev)->gen);
2188 struct i915_hw_ppgtt *
2189 i915_ppgtt_create(struct drm_i915_private *dev_priv,
2190 struct drm_i915_file_private *fpriv)
2192 struct i915_hw_ppgtt *ppgtt;
2195 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2197 return ERR_PTR(-ENOMEM);
2199 ret = i915_ppgtt_init(ppgtt, dev_priv, fpriv);
2202 return ERR_PTR(ret);
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 and destroyed */
2218 WARN_ON(!list_empty(&ppgtt->base.active_list));
2219 WARN_ON(!list_empty(&ppgtt->base.inactive_list));
2220 WARN_ON(!list_empty(&ppgtt->base.unbound_list));
2222 list_del(&ppgtt->base.global_link);
2223 drm_mm_takedown(&ppgtt->base.mm);
2225 ppgtt->base.cleanup(&ppgtt->base);
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_i915_private *dev_priv)
2234 #ifdef CONFIG_INTEL_IOMMU
2235 /* Query intel_iommu to see if we need the workaround. Presumably that
2238 if (IS_GEN5(dev_priv) && IS_MOBILE(dev_priv) && intel_iommu_gfx_mapped)
2244 void i915_check_and_clear_faults(struct drm_i915_private *dev_priv)
2246 struct intel_engine_cs *engine;
2248 if (INTEL_INFO(dev_priv)->gen < 6)
2251 for_each_engine(engine, dev_priv) {
2253 fault_reg = I915_READ(RING_FAULT_REG(engine));
2254 if (fault_reg & RING_FAULT_VALID) {
2255 DRM_DEBUG_DRIVER("Unexpected fault\n"
2257 "\tAddress space: %s\n"
2260 fault_reg & LINUX_PAGE_MASK,
2261 fault_reg & RING_FAULT_GTTSEL_MASK ? "GGTT" : "PPGTT",
2262 RING_FAULT_SRCID(fault_reg),
2263 RING_FAULT_FAULT_TYPE(fault_reg));
2264 I915_WRITE(RING_FAULT_REG(engine),
2265 fault_reg & ~RING_FAULT_VALID);
2268 POSTING_READ(RING_FAULT_REG(&dev_priv->engine[RCS]));
2271 static void i915_ggtt_flush(struct drm_i915_private *dev_priv)
2273 if (INTEL_INFO(dev_priv)->gen < 6) {
2274 intel_gtt_chipset_flush();
2276 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2277 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2281 void i915_gem_suspend_gtt_mappings(struct drm_device *dev)
2283 struct drm_i915_private *dev_priv = to_i915(dev);
2284 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2286 /* Don't bother messing with faults pre GEN6 as we have little
2287 * documentation supporting that it's a good idea.
2289 if (INTEL_INFO(dev)->gen < 6)
2292 i915_check_and_clear_faults(dev_priv);
2294 ggtt->base.clear_range(&ggtt->base, ggtt->base.start, ggtt->base.total,
2297 i915_ggtt_flush(dev_priv);
2300 int i915_gem_gtt_prepare_object(struct drm_i915_gem_object *obj)
2302 if (!dma_map_sg(&obj->base.dev->pdev->dev,
2303 obj->pages->sgl, obj->pages->nents,
2304 PCI_DMA_BIDIRECTIONAL))
2310 static void gen8_set_pte(void __iomem *addr, gen8_pte_t pte)
2315 iowrite32((u32)pte, addr);
2316 iowrite32(pte >> 32, addr + 4);
2320 static void gen8_ggtt_insert_page(struct i915_address_space *vm,
2323 enum i915_cache_level level,
2326 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2327 gen8_pte_t __iomem *pte =
2328 (gen8_pte_t __iomem *)dev_priv->ggtt.gsm +
2329 (offset >> PAGE_SHIFT);
2332 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2334 gen8_set_pte(pte, gen8_pte_encode(addr, level, true));
2336 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2337 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2339 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2342 static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
2343 struct sg_table *st,
2345 enum i915_cache_level level, u32 unused)
2347 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2348 struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
2349 struct sgt_iter sgt_iter;
2350 gen8_pte_t __iomem *gtt_entries;
2351 gen8_pte_t gtt_entry;
2356 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2358 gtt_entries = (gen8_pte_t __iomem *)ggtt->gsm + (start >> PAGE_SHIFT);
2360 for_each_sgt_dma(addr, sgt_iter, st) {
2361 gtt_entry = gen8_pte_encode(addr, level, true);
2362 gen8_set_pte(>t_entries[i++], gtt_entry);
2366 * XXX: This serves as a posting read to make sure that the PTE has
2367 * actually been updated. There is some concern that even though
2368 * registers and PTEs are within the same BAR that they are potentially
2369 * of NUMA access patterns. Therefore, even with the way we assume
2370 * hardware should work, we must keep this posting read for paranoia.
2373 WARN_ON(readq(>t_entries[i-1]) != gtt_entry);
2375 /* This next bit makes the above posting read even more important. We
2376 * want to flush the TLBs only after we're certain all the PTE updates
2379 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2380 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2382 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2385 struct insert_entries {
2386 struct i915_address_space *vm;
2387 struct sg_table *st;
2389 enum i915_cache_level level;
2393 static int gen8_ggtt_insert_entries__cb(void *_arg)
2395 struct insert_entries *arg = _arg;
2396 gen8_ggtt_insert_entries(arg->vm, arg->st,
2397 arg->start, arg->level, arg->flags);
2401 static void gen8_ggtt_insert_entries__BKL(struct i915_address_space *vm,
2402 struct sg_table *st,
2404 enum i915_cache_level level,
2407 struct insert_entries arg = { vm, st, start, level, flags };
2408 stop_machine(gen8_ggtt_insert_entries__cb, &arg, NULL);
2411 static void gen6_ggtt_insert_page(struct i915_address_space *vm,
2414 enum i915_cache_level level,
2417 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2418 gen6_pte_t __iomem *pte =
2419 (gen6_pte_t __iomem *)dev_priv->ggtt.gsm +
2420 (offset >> PAGE_SHIFT);
2423 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2425 iowrite32(vm->pte_encode(addr, level, true, flags), pte);
2427 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2428 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2430 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2434 * Binds an object into the global gtt with the specified cache level. The object
2435 * will be accessible to the GPU via commands whose operands reference offsets
2436 * within the global GTT as well as accessible by the GPU through the GMADR
2437 * mapped BAR (dev_priv->mm.gtt->gtt).
2439 static void gen6_ggtt_insert_entries(struct i915_address_space *vm,
2440 struct sg_table *st,
2442 enum i915_cache_level level, u32 flags)
2444 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2445 struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
2446 struct sgt_iter sgt_iter;
2447 gen6_pte_t __iomem *gtt_entries;
2448 gen6_pte_t gtt_entry;
2453 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2455 gtt_entries = (gen6_pte_t __iomem *)ggtt->gsm + (start >> PAGE_SHIFT);
2457 for_each_sgt_dma(addr, sgt_iter, st) {
2458 gtt_entry = vm->pte_encode(addr, level, true, flags);
2459 iowrite32(gtt_entry, >t_entries[i++]);
2462 /* XXX: This serves as a posting read to make sure that the PTE has
2463 * actually been updated. There is some concern that even though
2464 * registers and PTEs are within the same BAR that they are potentially
2465 * of NUMA access patterns. Therefore, even with the way we assume
2466 * hardware should work, we must keep this posting read for paranoia.
2469 WARN_ON(readl(>t_entries[i-1]) != gtt_entry);
2471 /* This next bit makes the above posting read even more important. We
2472 * want to flush the TLBs only after we're certain all the PTE updates
2475 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2476 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2478 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2481 static void nop_clear_range(struct i915_address_space *vm,
2488 static void gen8_ggtt_clear_range(struct i915_address_space *vm,
2493 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2494 struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
2495 unsigned first_entry = start >> PAGE_SHIFT;
2496 unsigned num_entries = length >> PAGE_SHIFT;
2497 gen8_pte_t scratch_pte, __iomem *gtt_base =
2498 (gen8_pte_t __iomem *)ggtt->gsm + first_entry;
2499 const int max_entries = ggtt_total_entries(ggtt) - first_entry;
2503 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2505 if (WARN(num_entries > max_entries,
2506 "First entry = %d; Num entries = %d (max=%d)\n",
2507 first_entry, num_entries, max_entries))
2508 num_entries = max_entries;
2510 scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
2513 for (i = 0; i < num_entries; i++)
2514 gen8_set_pte(>t_base[i], scratch_pte);
2517 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2520 static void gen6_ggtt_clear_range(struct i915_address_space *vm,
2525 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2526 struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
2527 unsigned first_entry = start >> PAGE_SHIFT;
2528 unsigned num_entries = length >> PAGE_SHIFT;
2529 gen6_pte_t scratch_pte, __iomem *gtt_base =
2530 (gen6_pte_t __iomem *)ggtt->gsm + first_entry;
2531 const int max_entries = ggtt_total_entries(ggtt) - first_entry;
2535 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2537 if (WARN(num_entries > max_entries,
2538 "First entry = %d; Num entries = %d (max=%d)\n",
2539 first_entry, num_entries, max_entries))
2540 num_entries = max_entries;
2542 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
2543 I915_CACHE_LLC, use_scratch, 0);
2545 for (i = 0; i < num_entries; i++)
2546 iowrite32(scratch_pte, >t_base[i]);
2549 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2552 static void i915_ggtt_insert_page(struct i915_address_space *vm,
2555 enum i915_cache_level cache_level,
2558 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2559 unsigned int flags = (cache_level == I915_CACHE_NONE) ?
2560 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
2563 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2565 intel_gtt_insert_page(addr, offset >> PAGE_SHIFT, flags);
2567 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2570 static void i915_ggtt_insert_entries(struct i915_address_space *vm,
2571 struct sg_table *pages,
2573 enum i915_cache_level cache_level, u32 unused)
2575 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2576 unsigned int flags = (cache_level == I915_CACHE_NONE) ?
2577 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
2580 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2582 intel_gtt_insert_sg_entries(pages, start >> PAGE_SHIFT, flags);
2584 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2588 static void i915_ggtt_clear_range(struct i915_address_space *vm,
2593 struct drm_i915_private *dev_priv = to_i915(vm->dev);
2594 unsigned first_entry = start >> PAGE_SHIFT;
2595 unsigned num_entries = length >> PAGE_SHIFT;
2598 rpm_atomic_seq = assert_rpm_atomic_begin(dev_priv);
2600 intel_gtt_clear_range(first_entry, num_entries);
2602 assert_rpm_atomic_end(dev_priv, rpm_atomic_seq);
2605 static int ggtt_bind_vma(struct i915_vma *vma,
2606 enum i915_cache_level cache_level,
2609 struct drm_i915_gem_object *obj = vma->obj;
2613 ret = i915_get_ggtt_vma_pages(vma);
2617 /* Currently applicable only to VLV */
2619 pte_flags |= PTE_READ_ONLY;
2621 vma->vm->insert_entries(vma->vm, vma->ggtt_view.pages,
2623 cache_level, pte_flags);
2626 * Without aliasing PPGTT there's no difference between
2627 * GLOBAL/LOCAL_BIND, it's all the same ptes. Hence unconditionally
2628 * upgrade to both bound if we bind either to avoid double-binding.
2630 vma->flags |= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
2635 static int aliasing_gtt_bind_vma(struct i915_vma *vma,
2636 enum i915_cache_level cache_level,
2642 ret = i915_get_ggtt_vma_pages(vma);
2646 /* Currently applicable only to VLV */
2648 if (vma->obj->gt_ro)
2649 pte_flags |= PTE_READ_ONLY;
2652 if (flags & I915_VMA_GLOBAL_BIND) {
2653 vma->vm->insert_entries(vma->vm,
2654 vma->ggtt_view.pages,
2656 cache_level, pte_flags);
2659 if (flags & I915_VMA_LOCAL_BIND) {
2660 struct i915_hw_ppgtt *appgtt =
2661 to_i915(vma->vm->dev)->mm.aliasing_ppgtt;
2662 appgtt->base.insert_entries(&appgtt->base,
2663 vma->ggtt_view.pages,
2665 cache_level, pte_flags);
2671 static void ggtt_unbind_vma(struct i915_vma *vma)
2673 struct i915_hw_ppgtt *appgtt = to_i915(vma->vm->dev)->mm.aliasing_ppgtt;
2674 const u64 size = min(vma->size, vma->node.size);
2676 if (vma->flags & I915_VMA_GLOBAL_BIND)
2677 vma->vm->clear_range(vma->vm,
2678 vma->node.start, size,
2681 if (vma->flags & I915_VMA_LOCAL_BIND && appgtt)
2682 appgtt->base.clear_range(&appgtt->base,
2683 vma->node.start, size,
2687 void i915_gem_gtt_finish_object(struct drm_i915_gem_object *obj)
2689 struct drm_device *dev = obj->base.dev;
2690 struct drm_i915_private *dev_priv = to_i915(dev);
2691 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2693 if (unlikely(ggtt->do_idle_maps)) {
2694 if (i915_gem_wait_for_idle(dev_priv, false)) {
2695 DRM_ERROR("Failed to wait for idle; VT'd may hang.\n");
2696 /* Wait a bit, in hopes it avoids the hang */
2701 dma_unmap_sg(&dev->pdev->dev, obj->pages->sgl, obj->pages->nents,
2702 PCI_DMA_BIDIRECTIONAL);
2705 static void i915_gtt_color_adjust(struct drm_mm_node *node,
2706 unsigned long color,
2710 if (node->color != color)
2713 node = list_first_entry_or_null(&node->node_list,
2716 if (node && node->allocated && node->color != color)
2720 int i915_gem_init_ggtt(struct drm_i915_private *dev_priv)
2722 /* Let GEM Manage all of the aperture.
2724 * However, leave one page at the end still bound to the scratch page.
2725 * There are a number of places where the hardware apparently prefetches
2726 * past the end of the object, and we've seen multiple hangs with the
2727 * GPU head pointer stuck in a batchbuffer bound at the last page of the
2728 * aperture. One page should be enough to keep any prefetching inside
2731 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2732 unsigned long hole_start, hole_end;
2733 struct drm_mm_node *entry;
2735 unsigned long mappable = min(ggtt->base.total, ggtt->mappable_end);
2737 ret = intel_vgt_balloon(dev_priv);
2741 /* Clear any non-preallocated blocks */
2742 drm_mm_for_each_hole(entry, &ggtt->base.mm, hole_start, hole_end) {
2743 DRM_DEBUG_KMS("clearing unused GTT space: [%lx, %lx]\n",
2744 hole_start, hole_end);
2745 ggtt->base.clear_range(&ggtt->base, hole_start,
2746 hole_end - hole_start, true);
2749 #ifdef __DragonFly__
2750 DRM_INFO("taking over the fictitious range 0x%llx-0x%llx\n",
2751 dev_priv->ggtt.mappable_base, dev_priv->ggtt.mappable_end);
2752 vm_phys_fictitious_reg_range(dev_priv->ggtt.mappable_base,
2753 dev_priv->ggtt.mappable_base + mappable, VM_MEMATTR_WRITE_COMBINING);
2756 /* And finally clear the reserved guard page */
2757 ggtt->base.clear_range(&ggtt->base,
2758 ggtt->base.total - PAGE_SIZE, PAGE_SIZE,
2761 if (USES_PPGTT(dev_priv) && !USES_FULL_PPGTT(dev_priv)) {
2762 struct i915_hw_ppgtt *ppgtt;
2764 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2768 ret = __hw_ppgtt_init(ppgtt, dev_priv);
2774 if (ppgtt->base.allocate_va_range)
2775 ret = ppgtt->base.allocate_va_range(&ppgtt->base, 0,
2778 ppgtt->base.cleanup(&ppgtt->base);
2783 ppgtt->base.clear_range(&ppgtt->base,
2788 dev_priv->mm.aliasing_ppgtt = ppgtt;
2789 WARN_ON(ggtt->base.bind_vma != ggtt_bind_vma);
2790 ggtt->base.bind_vma = aliasing_gtt_bind_vma;
2797 * i915_ggtt_cleanup_hw - Clean up GGTT hardware initialization
2798 * @dev_priv: i915 device
2800 void i915_ggtt_cleanup_hw(struct drm_i915_private *dev_priv)
2802 struct i915_ggtt *ggtt = &dev_priv->ggtt;
2804 if (dev_priv->mm.aliasing_ppgtt) {
2805 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
2807 ppgtt->base.cleanup(&ppgtt->base);
2811 i915_gem_cleanup_stolen(&dev_priv->drm);
2813 if (drm_mm_initialized(&ggtt->base.mm)) {
2814 intel_vgt_deballoon(dev_priv);
2816 drm_mm_takedown(&ggtt->base.mm);
2817 list_del(&ggtt->base.global_link);
2820 ggtt->base.cleanup(&ggtt->base);
2822 arch_phys_wc_del(ggtt->mtrr);
2823 io_mapping_free(ggtt->mappable);
2826 static unsigned int gen6_get_total_gtt_size(u16 snb_gmch_ctl)
2828 snb_gmch_ctl >>= SNB_GMCH_GGMS_SHIFT;
2829 snb_gmch_ctl &= SNB_GMCH_GGMS_MASK;
2830 return snb_gmch_ctl << 20;
2833 static unsigned int gen8_get_total_gtt_size(u16 bdw_gmch_ctl)
2835 bdw_gmch_ctl >>= BDW_GMCH_GGMS_SHIFT;
2836 bdw_gmch_ctl &= BDW_GMCH_GGMS_MASK;
2838 bdw_gmch_ctl = 1 << bdw_gmch_ctl;
2840 #ifdef CONFIG_X86_32
2841 /* Limit 32b platforms to a 2GB GGTT: 4 << 20 / pte size * PAGE_SIZE */
2842 if (bdw_gmch_ctl > 4)
2846 return bdw_gmch_ctl << 20;
2849 static unsigned int chv_get_total_gtt_size(u16 gmch_ctrl)
2851 gmch_ctrl >>= SNB_GMCH_GGMS_SHIFT;
2852 gmch_ctrl &= SNB_GMCH_GGMS_MASK;
2855 return 1 << (20 + gmch_ctrl);
2860 static size_t gen6_get_stolen_size(u16 snb_gmch_ctl)
2862 snb_gmch_ctl >>= SNB_GMCH_GMS_SHIFT;
2863 snb_gmch_ctl &= SNB_GMCH_GMS_MASK;
2864 return snb_gmch_ctl << 25; /* 32 MB units */
2867 static size_t gen8_get_stolen_size(u16 bdw_gmch_ctl)
2869 bdw_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2870 bdw_gmch_ctl &= BDW_GMCH_GMS_MASK;
2871 return bdw_gmch_ctl << 25; /* 32 MB units */
2874 static size_t chv_get_stolen_size(u16 gmch_ctrl)
2876 gmch_ctrl >>= SNB_GMCH_GMS_SHIFT;
2877 gmch_ctrl &= SNB_GMCH_GMS_MASK;
2880 * 0x0 to 0x10: 32MB increments starting at 0MB
2881 * 0x11 to 0x16: 4MB increments starting at 8MB
2882 * 0x17 to 0x1d: 4MB increments start at 36MB
2884 if (gmch_ctrl < 0x11)
2885 return gmch_ctrl << 25;
2886 else if (gmch_ctrl < 0x17)
2887 return (gmch_ctrl - 0x11 + 2) << 22;
2889 return (gmch_ctrl - 0x17 + 9) << 22;
2892 static size_t gen9_get_stolen_size(u16 gen9_gmch_ctl)
2894 gen9_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2895 gen9_gmch_ctl &= BDW_GMCH_GMS_MASK;
2897 if (gen9_gmch_ctl < 0xf0)
2898 return gen9_gmch_ctl << 25; /* 32 MB units */
2900 /* 4MB increments starting at 0xf0 for 4MB */
2901 return (gen9_gmch_ctl - 0xf0 + 1) << 22;
2904 static int ggtt_probe_common(struct i915_ggtt *ggtt, u64 size)
2906 struct pci_dev *pdev = ggtt->base.dev->pdev;
2907 struct i915_page_scratch *scratch_page;
2908 phys_addr_t phys_addr;
2910 /* For Modern GENs the PTEs and register space are split in the BAR */
2911 phys_addr = pci_resource_start(pdev, 0) + pci_resource_len(pdev, 0) / 2;
2914 * On BXT writes larger than 64 bit to the GTT pagetable range will be
2915 * dropped. For WC mappings in general we have 64 byte burst writes
2916 * when the WC buffer is flushed, so we can't use it, but have to
2917 * resort to an uncached mapping. The WC issue is easily caught by the
2918 * readback check when writing GTT PTE entries.
2920 if (IS_BROXTON(ggtt->base.dev))
2921 ggtt->gsm = ioremap_nocache(phys_addr, size);
2923 ggtt->gsm = ioremap_wc(phys_addr, size);
2925 DRM_ERROR("Failed to map the ggtt page table\n");
2929 scratch_page = alloc_scratch_page(ggtt->base.dev);
2930 if (IS_ERR(scratch_page)) {
2931 DRM_ERROR("Scratch setup failed\n");
2932 /* iounmap will also get called at remove, but meh */
2934 return PTR_ERR(scratch_page);
2937 ggtt->base.scratch_page = scratch_page;
2942 /* The GGTT and PPGTT need a private PPAT setup in order to handle cacheability
2943 * bits. When using advanced contexts each context stores its own PAT, but
2944 * writing this data shouldn't be harmful even in those cases. */
2945 static void bdw_setup_private_ppat(struct drm_i915_private *dev_priv)
2949 pat = GEN8_PPAT(0, GEN8_PPAT_WB | GEN8_PPAT_LLC) | /* for normal objects, no eLLC */
2950 GEN8_PPAT(1, GEN8_PPAT_WC | GEN8_PPAT_LLCELLC) | /* for something pointing to ptes? */
2951 GEN8_PPAT(2, GEN8_PPAT_WT | GEN8_PPAT_LLCELLC) | /* for scanout with eLLC */
2952 GEN8_PPAT(3, GEN8_PPAT_UC) | /* Uncached objects, mostly for scanout */
2953 GEN8_PPAT(4, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(0)) |
2954 GEN8_PPAT(5, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(1)) |
2955 GEN8_PPAT(6, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(2)) |
2956 GEN8_PPAT(7, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3));
2958 if (!USES_PPGTT(dev_priv))
2959 /* Spec: "For GGTT, there is NO pat_sel[2:0] from the entry,
2960 * so RTL will always use the value corresponding to
2962 * So let's disable cache for GGTT to avoid screen corruptions.
2963 * MOCS still can be used though.
2964 * - System agent ggtt writes (i.e. cpu gtt mmaps) already work
2965 * before this patch, i.e. the same uncached + snooping access
2966 * like on gen6/7 seems to be in effect.
2967 * - So this just fixes blitter/render access. Again it looks
2968 * like it's not just uncached access, but uncached + snooping.
2969 * So we can still hold onto all our assumptions wrt cpu
2970 * clflushing on LLC machines.
2972 pat = GEN8_PPAT(0, GEN8_PPAT_UC);
2974 /* XXX: spec defines this as 2 distinct registers. It's unclear if a 64b
2975 * write would work. */
2976 I915_WRITE(GEN8_PRIVATE_PAT_LO, pat);
2977 I915_WRITE(GEN8_PRIVATE_PAT_HI, pat >> 32);
2980 static void chv_setup_private_ppat(struct drm_i915_private *dev_priv)
2985 * Map WB on BDW to snooped on CHV.
2987 * Only the snoop bit has meaning for CHV, the rest is
2990 * The hardware will never snoop for certain types of accesses:
2991 * - CPU GTT (GMADR->GGTT->no snoop->memory)
2992 * - PPGTT page tables
2993 * - some other special cycles
2995 * As with BDW, we also need to consider the following for GT accesses:
2996 * "For GGTT, there is NO pat_sel[2:0] from the entry,
2997 * so RTL will always use the value corresponding to
2999 * Which means we must set the snoop bit in PAT entry 0
3000 * in order to keep the global status page working.
3002 pat = GEN8_PPAT(0, CHV_PPAT_SNOOP) |
3006 GEN8_PPAT(4, CHV_PPAT_SNOOP) |
3007 GEN8_PPAT(5, CHV_PPAT_SNOOP) |
3008 GEN8_PPAT(6, CHV_PPAT_SNOOP) |
3009 GEN8_PPAT(7, CHV_PPAT_SNOOP);
3011 I915_WRITE(GEN8_PRIVATE_PAT_LO, pat);
3012 I915_WRITE(GEN8_PRIVATE_PAT_HI, pat >> 32);
3015 static void gen6_gmch_remove(struct i915_address_space *vm)
3017 struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
3020 free_scratch_page(vm->dev, vm->scratch_page);
3023 static int gen8_gmch_probe(struct i915_ggtt *ggtt)
3025 struct drm_i915_private *dev_priv = to_i915(ggtt->base.dev);
3026 struct pci_dev *pdev = dev_priv->drm.pdev;
3030 /* TODO: We're not aware of mappable constraints on gen8 yet */
3031 ggtt->mappable_base = pci_resource_start(pdev, 2);
3032 ggtt->mappable_end = pci_resource_len(pdev, 2);
3035 if (!pci_set_dma_mask(pdev, DMA_BIT_MASK(39)))
3036 pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(39));
3039 pci_read_config_word(pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
3041 if (INTEL_GEN(dev_priv) >= 9) {
3042 ggtt->stolen_size = gen9_get_stolen_size(snb_gmch_ctl);
3043 size = gen8_get_total_gtt_size(snb_gmch_ctl);
3044 } else if (IS_CHERRYVIEW(dev_priv)) {
3045 ggtt->stolen_size = chv_get_stolen_size(snb_gmch_ctl);
3046 size = chv_get_total_gtt_size(snb_gmch_ctl);
3048 ggtt->stolen_size = gen8_get_stolen_size(snb_gmch_ctl);
3049 size = gen8_get_total_gtt_size(snb_gmch_ctl);
3052 ggtt->base.total = (size / sizeof(gen8_pte_t)) << PAGE_SHIFT;
3054 if (IS_CHERRYVIEW(dev_priv) || IS_BROXTON(dev_priv))
3055 chv_setup_private_ppat(dev_priv);
3057 bdw_setup_private_ppat(dev_priv);
3059 ggtt->base.cleanup = gen6_gmch_remove;
3060 ggtt->base.bind_vma = ggtt_bind_vma;
3061 ggtt->base.unbind_vma = ggtt_unbind_vma;
3062 ggtt->base.insert_page = gen8_ggtt_insert_page;
3063 ggtt->base.clear_range = nop_clear_range;
3064 if (!USES_FULL_PPGTT(dev_priv) || intel_scanout_needs_vtd_wa(dev_priv))
3065 ggtt->base.clear_range = gen8_ggtt_clear_range;
3067 ggtt->base.insert_entries = gen8_ggtt_insert_entries;
3068 if (IS_CHERRYVIEW(dev_priv))
3069 ggtt->base.insert_entries = gen8_ggtt_insert_entries__BKL;
3071 return ggtt_probe_common(ggtt, size);
3074 static int gen6_gmch_probe(struct i915_ggtt *ggtt)
3076 struct drm_i915_private *dev_priv = to_i915(ggtt->base.dev);
3077 struct pci_dev *pdev = dev_priv->drm.pdev;
3081 ggtt->mappable_base = pci_resource_start(pdev, 2);
3082 ggtt->mappable_end = pci_resource_len(pdev, 2);
3084 /* 64/512MB is the current min/max we actually know of, but this is just
3085 * a coarse sanity check.
3087 if (ggtt->mappable_end < (64<<20) || ggtt->mappable_end > (512<<20)) {
3088 DRM_ERROR("Unknown GMADR size (%llx)\n", ggtt->mappable_end);
3093 if (!pci_set_dma_mask(pdev, DMA_BIT_MASK(40)))
3094 pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(40));
3096 pci_read_config_word(pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
3098 ggtt->stolen_size = gen6_get_stolen_size(snb_gmch_ctl);
3100 size = gen6_get_total_gtt_size(snb_gmch_ctl);
3101 ggtt->base.total = (size / sizeof(gen6_pte_t)) << PAGE_SHIFT;
3103 ggtt->base.clear_range = gen6_ggtt_clear_range;
3104 ggtt->base.insert_page = gen6_ggtt_insert_page;
3105 ggtt->base.insert_entries = gen6_ggtt_insert_entries;
3106 ggtt->base.bind_vma = ggtt_bind_vma;
3107 ggtt->base.unbind_vma = ggtt_unbind_vma;
3108 ggtt->base.cleanup = gen6_gmch_remove;
3110 if (HAS_EDRAM(dev_priv))
3111 ggtt->base.pte_encode = iris_pte_encode;
3112 else if (IS_HASWELL(dev_priv))
3113 ggtt->base.pte_encode = hsw_pte_encode;
3114 else if (IS_VALLEYVIEW(dev_priv))
3115 ggtt->base.pte_encode = byt_pte_encode;
3116 else if (INTEL_GEN(dev_priv) >= 7)
3117 ggtt->base.pte_encode = ivb_pte_encode;
3119 ggtt->base.pte_encode = snb_pte_encode;
3121 return ggtt_probe_common(ggtt, size);
3124 static void i915_gmch_remove(struct i915_address_space *vm)
3126 intel_gmch_remove();
3129 static int i915_gmch_probe(struct i915_ggtt *ggtt)
3131 struct drm_i915_private *dev_priv = to_i915(ggtt->base.dev);
3135 ret = intel_gmch_probe(dev_priv->bridge_dev, dev_priv->drm.pdev, NULL);
3137 DRM_ERROR("failed to set up gmch\n");
3142 intel_gtt_get(&ggtt->base.total, &ggtt->stolen_size,
3143 &ggtt->mappable_base, &ggtt->mappable_end);
3145 ggtt->do_idle_maps = needs_idle_maps(dev_priv);
3146 ggtt->base.insert_page = i915_ggtt_insert_page;
3147 ggtt->base.insert_entries = i915_ggtt_insert_entries;
3148 ggtt->base.clear_range = i915_ggtt_clear_range;
3149 ggtt->base.bind_vma = ggtt_bind_vma;
3150 ggtt->base.unbind_vma = ggtt_unbind_vma;
3151 ggtt->base.cleanup = i915_gmch_remove;
3153 if (unlikely(ggtt->do_idle_maps))
3154 DRM_INFO("applying Ironlake quirks for intel_iommu\n");
3160 * i915_ggtt_probe_hw - Probe GGTT hardware location
3161 * @dev_priv: i915 device
3163 int i915_ggtt_probe_hw(struct drm_i915_private *dev_priv)
3165 struct i915_ggtt *ggtt = &dev_priv->ggtt;
3168 ggtt->base.dev = &dev_priv->drm;
3170 if (INTEL_GEN(dev_priv) <= 5)
3171 ret = i915_gmch_probe(ggtt);
3172 else if (INTEL_GEN(dev_priv) < 8)
3173 ret = gen6_gmch_probe(ggtt);
3175 ret = gen8_gmch_probe(ggtt);
3179 if ((ggtt->base.total - 1) >> 32) {
3180 DRM_ERROR("We never expected a Global GTT with more than 32bits"
3181 " of address space! Found %lldM!\n",
3182 ggtt->base.total >> 20);
3183 ggtt->base.total = 1ULL << 32;
3184 ggtt->mappable_end = min(ggtt->mappable_end, ggtt->base.total);
3187 if (ggtt->mappable_end > ggtt->base.total) {
3188 DRM_ERROR("mappable aperture extends past end of GGTT,"
3189 " aperture=%llx, total=%llx\n",
3190 ggtt->mappable_end, ggtt->base.total);
3191 ggtt->mappable_end = ggtt->base.total;
3194 /* GMADR is the PCI mmio aperture into the global GTT. */
3195 DRM_INFO("Memory usable by graphics device = %lluM\n",
3196 ggtt->base.total >> 20);
3197 DRM_DEBUG_DRIVER("GMADR size = %lldM\n", ggtt->mappable_end >> 20);
3198 DRM_DEBUG_DRIVER("GTT stolen size = %zdM\n", ggtt->stolen_size >> 20);
3199 #ifdef CONFIG_INTEL_IOMMU
3200 if (intel_iommu_gfx_mapped)
3201 DRM_INFO("VT-d active for gfx access\n");
3208 * i915_ggtt_init_hw - Initialize GGTT hardware
3209 * @dev_priv: i915 device
3211 int i915_ggtt_init_hw(struct drm_i915_private *dev_priv)
3213 struct i915_ggtt *ggtt = &dev_priv->ggtt;
3216 INIT_LIST_HEAD(&dev_priv->vm_list);
3218 /* Subtract the guard page before address space initialization to
3219 * shrink the range used by drm_mm.
3221 ggtt->base.total -= PAGE_SIZE;
3222 i915_address_space_init(&ggtt->base, dev_priv);
3223 ggtt->base.total += PAGE_SIZE;
3224 if (!HAS_LLC(dev_priv))
3225 ggtt->base.mm.color_adjust = i915_gtt_color_adjust;
3228 io_mapping_create_wc(ggtt->mappable_base, ggtt->mappable_end);
3229 if (!ggtt->mappable) {
3231 goto out_gtt_cleanup;
3234 ggtt->mtrr = arch_phys_wc_add(ggtt->mappable_base, ggtt->mappable_end);
3237 * Initialise stolen early so that we may reserve preallocated
3238 * objects for the BIOS to KMS transition.
3240 ret = i915_gem_init_stolen(&dev_priv->drm);
3242 goto out_gtt_cleanup;
3247 ggtt->base.cleanup(&ggtt->base);
3251 int i915_ggtt_enable_hw(struct drm_i915_private *dev_priv)
3253 if (INTEL_GEN(dev_priv) < 6 && !intel_enable_gtt())
3259 void i915_gem_restore_gtt_mappings(struct drm_device *dev)
3261 struct drm_i915_private *dev_priv = to_i915(dev);
3262 struct i915_ggtt *ggtt = &dev_priv->ggtt;
3263 struct drm_i915_gem_object *obj;
3264 struct i915_vma *vma;
3266 i915_check_and_clear_faults(dev_priv);
3268 /* First fill our portion of the GTT with scratch pages */
3269 ggtt->base.clear_range(&ggtt->base, ggtt->base.start, ggtt->base.total,
3272 /* Cache flush objects bound into GGTT and rebind them. */
3273 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
3274 list_for_each_entry(vma, &obj->vma_list, obj_link) {
3275 if (vma->vm != &ggtt->base)
3278 WARN_ON(i915_vma_bind(vma, obj->cache_level,
3282 if (obj->pin_display)
3283 WARN_ON(i915_gem_object_set_to_gtt_domain(obj, false));
3286 if (INTEL_INFO(dev)->gen >= 8) {
3287 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
3288 chv_setup_private_ppat(dev_priv);
3290 bdw_setup_private_ppat(dev_priv);
3295 if (USES_PPGTT(dev)) {
3296 struct i915_address_space *vm;
3298 list_for_each_entry(vm, &dev_priv->vm_list, global_link) {
3299 /* TODO: Perhaps it shouldn't be gen6 specific */
3301 struct i915_hw_ppgtt *ppgtt;
3303 if (i915_is_ggtt(vm))
3304 ppgtt = dev_priv->mm.aliasing_ppgtt;
3306 ppgtt = i915_vm_to_ppgtt(vm);
3308 gen6_write_page_range(dev_priv, &ppgtt->pd,
3309 0, ppgtt->base.total);
3313 i915_ggtt_flush(dev_priv);
3317 i915_vma_retire(struct i915_gem_active *active,
3318 struct drm_i915_gem_request *rq)
3320 const unsigned int idx = rq->engine->id;
3321 struct i915_vma *vma =
3322 container_of(active, struct i915_vma, last_read[idx]);
3324 GEM_BUG_ON(!i915_vma_has_active_engine(vma, idx));
3326 i915_vma_clear_active(vma, idx);
3327 if (i915_vma_is_active(vma))
3330 list_move_tail(&vma->vm_link, &vma->vm->inactive_list);
3331 if (unlikely(i915_vma_is_closed(vma) && !i915_vma_is_pinned(vma)))
3332 WARN_ON(i915_vma_unbind(vma));
3335 void i915_vma_destroy(struct i915_vma *vma)
3337 GEM_BUG_ON(vma->node.allocated);
3338 GEM_BUG_ON(i915_vma_is_active(vma));
3339 GEM_BUG_ON(!i915_vma_is_closed(vma));
3341 list_del(&vma->vm_link);
3342 if (!i915_vma_is_ggtt(vma))
3343 i915_ppgtt_put(i915_vm_to_ppgtt(vma->vm));
3345 kmem_cache_free(to_i915(vma->obj->base.dev)->vmas, vma);
3348 void i915_vma_close(struct i915_vma *vma)
3350 GEM_BUG_ON(i915_vma_is_closed(vma));
3351 vma->flags |= I915_VMA_CLOSED;
3353 list_del_init(&vma->obj_link);
3354 if (!i915_vma_is_active(vma) && !i915_vma_is_pinned(vma))
3355 WARN_ON(i915_vma_unbind(vma));
3358 static struct i915_vma *
3359 __i915_gem_vma_create(struct drm_i915_gem_object *obj,
3360 struct i915_address_space *vm,
3361 const struct i915_ggtt_view *view)
3363 struct i915_vma *vma;
3366 GEM_BUG_ON(vm->closed);
3368 if (WARN_ON(i915_is_ggtt(vm) != !!view))
3369 return ERR_PTR(-EINVAL);
3371 vma = kzalloc(sizeof(*vma), GFP_KERNEL);
3373 return ERR_PTR(-ENOMEM);
3375 INIT_LIST_HEAD(&vma->obj_link);
3376 INIT_LIST_HEAD(&vma->exec_list);
3377 for (i = 0; i < ARRAY_SIZE(vma->last_read); i++)
3378 init_request_active(&vma->last_read[i], i915_vma_retire);
3379 list_add(&vma->vm_link, &vm->unbound_list);
3382 vma->size = obj->base.size;
3384 if (i915_is_ggtt(vm)) {
3385 vma->flags |= I915_VMA_GGTT;
3386 vma->ggtt_view = *view;
3387 if (view->type == I915_GGTT_VIEW_PARTIAL) {
3388 vma->size = view->params.partial.size;
3389 vma->size <<= PAGE_SHIFT;
3390 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
3392 intel_rotation_info_size(&view->params.rotated);
3393 vma->size <<= PAGE_SHIFT;
3396 i915_ppgtt_get(i915_vm_to_ppgtt(vm));
3399 list_add_tail(&vma->obj_link, &obj->vma_list);
3405 i915_gem_obj_lookup_or_create_vma(struct drm_i915_gem_object *obj,
3406 struct i915_address_space *vm)
3408 struct i915_vma *vma;
3410 vma = i915_gem_obj_to_vma(obj, vm);
3412 vma = __i915_gem_vma_create(obj, vm,
3413 i915_is_ggtt(vm) ? &i915_ggtt_view_normal : NULL);
3419 i915_gem_obj_lookup_or_create_ggtt_vma(struct drm_i915_gem_object *obj,
3420 const struct i915_ggtt_view *view)
3422 struct drm_device *dev = obj->base.dev;
3423 struct drm_i915_private *dev_priv = to_i915(dev);
3424 struct i915_ggtt *ggtt = &dev_priv->ggtt;
3425 struct i915_vma *vma = i915_gem_obj_to_ggtt_view(obj, view);
3430 vma = __i915_gem_vma_create(obj, &ggtt->base, view);
3432 GEM_BUG_ON(i915_vma_is_closed(vma));
3437 static struct scatterlist *
3438 rotate_pages(const dma_addr_t *in, unsigned int offset,
3439 unsigned int width, unsigned int height,
3440 unsigned int stride,
3441 struct sg_table *st, struct scatterlist *sg)
3443 unsigned int column, row;
3444 unsigned int src_idx;
3446 for (column = 0; column < width; column++) {
3447 src_idx = stride * (height - 1) + column;
3448 for (row = 0; row < height; row++) {
3450 /* We don't need the pages, but need to initialize
3451 * the entries so the sg list can be happily traversed.
3452 * The only thing we need are DMA addresses.
3454 sg_set_page(sg, NULL, PAGE_SIZE, 0);
3455 sg_dma_address(sg) = in[offset + src_idx];
3456 sg_dma_len(sg) = PAGE_SIZE;
3465 static struct sg_table *
3466 intel_rotate_fb_obj_pages(struct intel_rotation_info *rot_info,
3467 struct drm_i915_gem_object *obj)
3469 const size_t n_pages = obj->base.size / PAGE_SIZE;
3470 unsigned int size_pages = rot_info->plane[0].width * rot_info->plane[0].height;
3471 unsigned int size_pages_uv;
3472 struct sgt_iter sgt_iter;
3473 dma_addr_t dma_addr;
3475 dma_addr_t *page_addr_list;
3476 struct sg_table *st;
3477 unsigned int uv_start_page;
3478 struct scatterlist *sg;
3481 /* Allocate a temporary list of source pages for random access. */
3482 page_addr_list = drm_malloc_gfp(n_pages,
3485 if (!page_addr_list)
3486 return ERR_PTR(ret);
3488 /* Account for UV plane with NV12. */
3489 if (rot_info->pixel_format == DRM_FORMAT_NV12)
3490 size_pages_uv = rot_info->plane[1].width * rot_info->plane[1].height;
3494 /* Allocate target SG list. */
3495 st = kmalloc(sizeof(*st), M_DRM, GFP_KERNEL);
3499 ret = sg_alloc_table(st, size_pages + size_pages_uv, GFP_KERNEL);
3503 /* Populate source page list from the object. */
3505 for_each_sgt_dma(dma_addr, sgt_iter, obj->pages)
3506 page_addr_list[i++] = dma_addr;
3508 GEM_BUG_ON(i != n_pages);
3512 /* Rotate the pages. */
3513 sg = rotate_pages(page_addr_list, 0,
3514 rot_info->plane[0].width, rot_info->plane[0].height,
3515 rot_info->plane[0].width,
3518 /* Append the UV plane if NV12. */
3519 if (rot_info->pixel_format == DRM_FORMAT_NV12) {
3520 uv_start_page = size_pages;
3522 /* Check for tile-row un-alignment. */
3523 if (offset_in_page(rot_info->uv_offset))
3526 rot_info->uv_start_page = uv_start_page;
3528 sg = rotate_pages(page_addr_list, rot_info->uv_start_page,
3529 rot_info->plane[1].width, rot_info->plane[1].height,
3530 rot_info->plane[1].width,
3534 DRM_DEBUG_KMS("Created rotated page mapping for object size %zu (%ux%u tiles, %u pages (%u plane 0)).\n",
3535 obj->base.size, rot_info->plane[0].width,
3536 rot_info->plane[0].height, size_pages + size_pages_uv,
3539 drm_free_large(page_addr_list);
3546 drm_free_large(page_addr_list);
3548 DRM_DEBUG_KMS("Failed to create rotated mapping for object size %zu! (%d) (%ux%u tiles, %u pages (%u plane 0))\n",
3549 obj->base.size, ret, rot_info->plane[0].width,
3550 rot_info->plane[0].height, size_pages + size_pages_uv,
3552 return ERR_PTR(ret);
3555 static struct sg_table *
3556 intel_partial_pages(const struct i915_ggtt_view *view,
3557 struct drm_i915_gem_object *obj)
3559 struct sg_table *st;
3560 struct scatterlist *sg;
3561 struct sg_page_iter obj_sg_iter;
3564 st = kmalloc(sizeof(*st), M_DRM, GFP_KERNEL);
3568 ret = sg_alloc_table(st, view->params.partial.size, GFP_KERNEL);
3574 for_each_sg_page(obj->pages->sgl, &obj_sg_iter, obj->pages->nents,
3575 view->params.partial.offset)
3577 if (st->nents >= view->params.partial.size)
3580 sg_set_page(sg, NULL, PAGE_SIZE, 0);
3581 sg_dma_address(sg) = sg_page_iter_dma_address(&obj_sg_iter);
3582 sg_dma_len(sg) = PAGE_SIZE;
3593 return ERR_PTR(ret);
3597 i915_get_ggtt_vma_pages(struct i915_vma *vma)
3601 if (vma->ggtt_view.pages)
3604 if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL)
3605 vma->ggtt_view.pages = vma->obj->pages;
3606 else if (vma->ggtt_view.type == I915_GGTT_VIEW_ROTATED)
3607 vma->ggtt_view.pages =
3608 intel_rotate_fb_obj_pages(&vma->ggtt_view.params.rotated, vma->obj);
3609 else if (vma->ggtt_view.type == I915_GGTT_VIEW_PARTIAL)
3610 vma->ggtt_view.pages =
3611 intel_partial_pages(&vma->ggtt_view, vma->obj);
3613 WARN_ONCE(1, "GGTT view %u not implemented!\n",
3614 vma->ggtt_view.type);
3616 if (!vma->ggtt_view.pages) {
3617 DRM_ERROR("Failed to get pages for GGTT view type %u!\n",
3618 vma->ggtt_view.type);
3620 } else if (IS_ERR(vma->ggtt_view.pages)) {
3621 ret = PTR_ERR(vma->ggtt_view.pages);
3622 vma->ggtt_view.pages = NULL;
3623 DRM_ERROR("Failed to get pages for VMA view type %u (%d)!\n",
3624 vma->ggtt_view.type, ret);
3631 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
3633 * @cache_level: mapping cache level
3634 * @flags: flags like global or local mapping
3636 * DMA addresses are taken from the scatter-gather table of this object (or of
3637 * this VMA in case of non-default GGTT views) and PTE entries set up.
3638 * Note that DMA addresses are also the only part of the SG table we care about.
3640 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 |= I915_VMA_GLOBAL_BIND;
3653 if (flags & PIN_USER)
3654 bind_flags |= I915_VMA_LOCAL_BIND;
3656 vma_flags = vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND);
3657 if (flags & PIN_UPDATE)
3658 bind_flags |= vma_flags;
3660 bind_flags &= ~vma_flags;
3661 if (bind_flags == 0)
3664 if (vma_flags == 0 && vma->vm->allocate_va_range) {
3665 trace_i915_va_alloc(vma);
3666 ret = vma->vm->allocate_va_range(vma->vm,
3673 ret = vma->vm->bind_vma(vma, cache_level, bind_flags);
3677 vma->flags |= bind_flags;
3681 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
3685 lockdep_assert_held(&vma->vm->dev->struct_mutex);
3686 if (WARN_ON(!vma->obj->map_and_fenceable))
3687 return IO_ERR_PTR(-ENODEV);
3689 GEM_BUG_ON(!i915_vma_is_ggtt(vma));
3690 GEM_BUG_ON((vma->flags & I915_VMA_GLOBAL_BIND) == 0);
3694 ptr = io_mapping_map_wc(i915_vm_to_ggtt(vma->vm)->mappable,
3698 return IO_ERR_PTR(-ENOMEM);
3703 __i915_vma_pin(vma);