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>
28 #include <drm/i915_drm.h>
30 #include "i915_vgpu.h"
31 #include "i915_trace.h"
32 #include "intel_drv.h"
35 * DOC: Global GTT views
37 * Background and previous state
39 * Historically objects could exists (be bound) in global GTT space only as
40 * singular instances with a view representing all of the object's backing pages
41 * in a linear fashion. This view will be called a normal view.
43 * To support multiple views of the same object, where the number of mapped
44 * pages is not equal to the backing store, or where the layout of the pages
45 * is not linear, concept of a GGTT view was added.
47 * One example of an alternative view is a stereo display driven by a single
48 * image. In this case we would have a framebuffer looking like this
54 * Above would represent a normal GGTT view as normally mapped for GPU or CPU
55 * rendering. In contrast, fed to the display engine would be an alternative
56 * view which could look something like this:
61 * In this example both the size and layout of pages in the alternative view is
62 * different from the normal view.
64 * Implementation and usage
66 * GGTT views are implemented using VMAs and are distinguished via enum
67 * i915_ggtt_view_type and struct i915_ggtt_view.
69 * A new flavour of core GEM functions which work with GGTT bound objects were
70 * added with the _ggtt_ infix, and sometimes with _view postfix to avoid
71 * renaming in large amounts of code. They take the struct i915_ggtt_view
72 * parameter encapsulating all metadata required to implement a view.
74 * As a helper for callers which are only interested in the normal view,
75 * globally const i915_ggtt_view_normal singleton instance exists. All old core
76 * GEM API functions, the ones not taking the view parameter, are operating on,
77 * or with the normal GGTT view.
79 * Code wanting to add or use a new GGTT view needs to:
81 * 1. Add a new enum with a suitable name.
82 * 2. Extend the metadata in the i915_ggtt_view structure if required.
83 * 3. Add support to i915_get_vma_pages().
85 * New views are required to build a scatter-gather table from within the
86 * i915_get_vma_pages function. This table is stored in the vma.ggtt_view and
87 * exists for the lifetime of an VMA.
89 * Core API is designed to have copy semantics which means that passed in
90 * struct i915_ggtt_view does not need to be persistent (left around after
91 * calling the core API functions).
96 i915_get_ggtt_vma_pages(struct i915_vma *vma);
98 const struct i915_ggtt_view i915_ggtt_view_normal;
99 const struct i915_ggtt_view i915_ggtt_view_rotated = {
100 .type = I915_GGTT_VIEW_ROTATED
103 static int sanitize_enable_ppgtt(struct drm_device *dev, int enable_ppgtt)
105 bool has_aliasing_ppgtt;
108 has_aliasing_ppgtt = INTEL_INFO(dev)->gen >= 6;
109 has_full_ppgtt = INTEL_INFO(dev)->gen >= 7;
111 if (intel_vgpu_active(dev))
112 has_full_ppgtt = false; /* emulation is too hard */
115 * We don't allow disabling PPGTT for gen9+ as it's a requirement for
116 * execlists, the sole mechanism available to submit work.
118 if (INTEL_INFO(dev)->gen < 9 &&
119 (enable_ppgtt == 0 || !has_aliasing_ppgtt))
122 if (enable_ppgtt == 1)
125 if (enable_ppgtt == 2 && has_full_ppgtt)
128 #ifdef CONFIG_INTEL_IOMMU
129 /* Disable ppgtt on SNB if VT-d is on. */
130 if (INTEL_INFO(dev)->gen == 6 && intel_iommu_gfx_mapped) {
131 DRM_INFO("Disabling PPGTT because VT-d is on\n");
136 /* Early VLV doesn't have this */
137 if (IS_VALLEYVIEW(dev) && !IS_CHERRYVIEW(dev) &&
138 dev->pdev->revision < 0xb) {
139 DRM_DEBUG_DRIVER("disabling PPGTT on pre-B3 step VLV\n");
143 if (INTEL_INFO(dev)->gen >= 8 && i915.enable_execlists)
146 return has_aliasing_ppgtt ? 1 : 0;
149 static int ppgtt_bind_vma(struct i915_vma *vma,
150 enum i915_cache_level cache_level,
155 /* Currently applicable only to VLV */
157 pte_flags |= PTE_READ_ONLY;
159 vma->vm->insert_entries(vma->vm, vma->obj->pages, vma->node.start,
160 cache_level, pte_flags);
165 static void ppgtt_unbind_vma(struct i915_vma *vma)
167 vma->vm->clear_range(vma->vm,
173 static gen8_pte_t gen8_pte_encode(dma_addr_t addr,
174 enum i915_cache_level level,
177 gen8_pte_t pte = valid ? _PAGE_PRESENT | _PAGE_RW : 0;
181 case I915_CACHE_NONE:
182 pte |= PPAT_UNCACHED_INDEX;
185 pte |= PPAT_DISPLAY_ELLC_INDEX;
188 pte |= PPAT_CACHED_INDEX;
195 static gen8_pde_t gen8_pde_encode(struct drm_device *dev,
197 enum i915_cache_level level)
199 gen8_pde_t pde = _PAGE_PRESENT | _PAGE_RW;
201 if (level != I915_CACHE_NONE)
202 pde |= PPAT_CACHED_PDE_INDEX;
204 pde |= PPAT_UNCACHED_INDEX;
208 static gen6_pte_t snb_pte_encode(dma_addr_t addr,
209 enum i915_cache_level level,
210 bool valid, u32 unused)
212 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
213 pte |= GEN6_PTE_ADDR_ENCODE(addr);
216 case I915_CACHE_L3_LLC:
218 pte |= GEN6_PTE_CACHE_LLC;
220 case I915_CACHE_NONE:
221 pte |= GEN6_PTE_UNCACHED;
230 static gen6_pte_t ivb_pte_encode(dma_addr_t addr,
231 enum i915_cache_level level,
232 bool valid, u32 unused)
234 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
235 pte |= GEN6_PTE_ADDR_ENCODE(addr);
238 case I915_CACHE_L3_LLC:
239 pte |= GEN7_PTE_CACHE_L3_LLC;
242 pte |= GEN6_PTE_CACHE_LLC;
244 case I915_CACHE_NONE:
245 pte |= GEN6_PTE_UNCACHED;
254 static gen6_pte_t byt_pte_encode(dma_addr_t addr,
255 enum i915_cache_level level,
256 bool valid, u32 flags)
258 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
259 pte |= GEN6_PTE_ADDR_ENCODE(addr);
261 if (!(flags & PTE_READ_ONLY))
262 pte |= BYT_PTE_WRITEABLE;
264 if (level != I915_CACHE_NONE)
265 pte |= BYT_PTE_SNOOPED_BY_CPU_CACHES;
270 static gen6_pte_t hsw_pte_encode(dma_addr_t addr,
271 enum i915_cache_level level,
272 bool valid, u32 unused)
274 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
275 pte |= HSW_PTE_ADDR_ENCODE(addr);
277 if (level != I915_CACHE_NONE)
278 pte |= HSW_WB_LLC_AGE3;
283 static gen6_pte_t iris_pte_encode(dma_addr_t addr,
284 enum i915_cache_level level,
285 bool valid, u32 unused)
287 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
288 pte |= HSW_PTE_ADDR_ENCODE(addr);
291 case I915_CACHE_NONE:
294 pte |= HSW_WT_ELLC_LLC_AGE3;
297 pte |= HSW_WB_ELLC_LLC_AGE3;
304 static int setup_page_dma(struct drm_device *dev, struct i915_page_dma *p)
306 struct device *device = &dev->pdev->dev;
308 p->page = alloc_page(GFP_KERNEL);
312 p->daddr = dma_map_page(device,
313 p->page, 0, 4096, PCI_DMA_BIDIRECTIONAL);
315 if (dma_mapping_error(device, p->daddr)) {
316 __free_page(p->page);
323 static void cleanup_page_dma(struct drm_device *dev, struct i915_page_dma *p)
325 if (WARN_ON(!p->page))
328 dma_unmap_page(&dev->pdev->dev, p->daddr, 4096, PCI_DMA_BIDIRECTIONAL);
329 __free_page(p->page);
330 memset(p, 0, sizeof(*p));
333 static void *kmap_page_dma(struct i915_page_dma *p)
335 return kmap_atomic(p->page);
338 /* We use the flushing unmap only with ppgtt structures:
339 * page directories, page tables and scratch pages.
341 static void kunmap_page_dma(struct drm_device *dev, void *vaddr)
343 /* There are only few exceptions for gen >=6. chv and bxt.
344 * And we are not sure about the latter so play safe for now.
346 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
347 drm_clflush_virt_range(vaddr, PAGE_SIZE);
349 kunmap_atomic(vaddr);
352 #define kmap_px(px) kmap_page_dma(&(px)->base)
353 #define kunmap_px(ppgtt, vaddr) kunmap_page_dma((ppgtt)->base.dev, (vaddr))
355 static void fill_page_dma(struct drm_device *dev, struct i915_page_dma *p,
359 uint64_t * const vaddr = kmap_page_dma(p);
361 for (i = 0; i < 512; i++)
364 kunmap_page_dma(dev, vaddr);
367 static void fill_page_dma_32(struct drm_device *dev, struct i915_page_dma *p,
368 const uint32_t val32)
374 fill_page_dma(dev, p, v);
377 static void free_pt(struct drm_device *dev, struct i915_page_table *pt)
379 cleanup_page_dma(dev, &pt->base);
380 kfree(pt->used_ptes);
384 static void gen8_initialize_pt(struct i915_address_space *vm,
385 struct i915_page_table *pt)
387 gen8_pte_t scratch_pte;
389 scratch_pte = gen8_pte_encode(vm->scratch.addr, I915_CACHE_LLC, true);
391 fill_page_dma(vm->dev, &pt->base, scratch_pte);
394 static struct i915_page_table *alloc_pt(struct drm_device *dev)
396 struct i915_page_table *pt;
397 const size_t count = INTEL_INFO(dev)->gen >= 8 ?
398 GEN8_PTES : GEN6_PTES;
401 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
403 return ERR_PTR(-ENOMEM);
405 pt->used_ptes = kcalloc(BITS_TO_LONGS(count), sizeof(*pt->used_ptes),
411 ret = setup_page_dma(dev, &pt->base);
418 kfree(pt->used_ptes);
425 static void free_pd(struct drm_device *dev, struct i915_page_directory *pd)
428 cleanup_page_dma(dev, &pd->base);
429 kfree(pd->used_pdes);
434 static struct i915_page_directory *alloc_pd(struct drm_device *dev)
436 struct i915_page_directory *pd;
439 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
441 return ERR_PTR(-ENOMEM);
443 pd->used_pdes = kcalloc(BITS_TO_LONGS(I915_PDES),
444 sizeof(*pd->used_pdes), GFP_KERNEL);
448 ret = setup_page_dma(dev, &pd->base);
455 kfree(pd->used_pdes);
462 /* Broadwell Page Directory Pointer Descriptors */
463 static int gen8_write_pdp(struct drm_i915_gem_request *req,
467 struct intel_engine_cs *ring = req->ring;
472 ret = intel_ring_begin(req, 6);
476 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
477 intel_ring_emit(ring, GEN8_RING_PDP_UDW(ring, entry));
478 intel_ring_emit(ring, upper_32_bits(addr));
479 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
480 intel_ring_emit(ring, GEN8_RING_PDP_LDW(ring, entry));
481 intel_ring_emit(ring, lower_32_bits(addr));
482 intel_ring_advance(ring);
487 static int gen8_mm_switch(struct i915_hw_ppgtt *ppgtt,
488 struct drm_i915_gem_request *req)
492 for (i = GEN8_LEGACY_PDPES - 1; i >= 0; i--) {
493 const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i);
495 ret = gen8_write_pdp(req, i, pd_daddr);
503 static void gen8_ppgtt_clear_range(struct i915_address_space *vm,
508 struct i915_hw_ppgtt *ppgtt =
509 container_of(vm, struct i915_hw_ppgtt, base);
510 gen8_pte_t *pt_vaddr, scratch_pte;
511 unsigned pdpe = start >> GEN8_PDPE_SHIFT & GEN8_PDPE_MASK;
512 unsigned pde = start >> GEN8_PDE_SHIFT & GEN8_PDE_MASK;
513 unsigned pte = start >> GEN8_PTE_SHIFT & GEN8_PTE_MASK;
514 unsigned num_entries = length >> PAGE_SHIFT;
515 unsigned last_pte, i;
517 scratch_pte = gen8_pte_encode(ppgtt->base.scratch.addr,
518 I915_CACHE_LLC, use_scratch);
520 while (num_entries) {
521 struct i915_page_directory *pd;
522 struct i915_page_table *pt;
524 if (WARN_ON(!ppgtt->pdp.page_directory[pdpe]))
527 pd = ppgtt->pdp.page_directory[pdpe];
529 if (WARN_ON(!pd->page_table[pde]))
532 pt = pd->page_table[pde];
534 if (WARN_ON(!pt->base.page))
537 last_pte = pte + num_entries;
538 if (last_pte > GEN8_PTES)
539 last_pte = GEN8_PTES;
541 pt_vaddr = kmap_px(pt);
543 for (i = pte; i < last_pte; i++) {
544 pt_vaddr[i] = scratch_pte;
548 kunmap_px(ppgtt, pt);
551 if (++pde == I915_PDES) {
558 static void gen8_ppgtt_insert_entries(struct i915_address_space *vm,
559 struct sg_table *pages,
561 enum i915_cache_level cache_level, u32 unused)
563 struct i915_hw_ppgtt *ppgtt =
564 container_of(vm, struct i915_hw_ppgtt, base);
565 gen8_pte_t *pt_vaddr;
566 unsigned pdpe = start >> GEN8_PDPE_SHIFT & GEN8_PDPE_MASK;
567 unsigned pde = start >> GEN8_PDE_SHIFT & GEN8_PDE_MASK;
568 unsigned pte = start >> GEN8_PTE_SHIFT & GEN8_PTE_MASK;
569 struct sg_page_iter sg_iter;
573 for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) {
574 if (WARN_ON(pdpe >= GEN8_LEGACY_PDPES))
577 if (pt_vaddr == NULL) {
578 struct i915_page_directory *pd = ppgtt->pdp.page_directory[pdpe];
579 struct i915_page_table *pt = pd->page_table[pde];
580 pt_vaddr = kmap_px(pt);
584 gen8_pte_encode(sg_page_iter_dma_address(&sg_iter),
586 if (++pte == GEN8_PTES) {
587 kunmap_px(ppgtt, pt_vaddr);
589 if (++pde == I915_PDES) {
598 kunmap_px(ppgtt, pt_vaddr);
601 static void __gen8_do_map_pt(gen8_pde_t * const pde,
602 struct i915_page_table *pt,
603 struct drm_device *dev)
606 gen8_pde_encode(dev, pt->base.daddr, I915_CACHE_LLC);
610 static void gen8_initialize_pd(struct i915_address_space *vm,
611 struct i915_page_directory *pd)
613 struct i915_hw_ppgtt *ppgtt =
614 container_of(vm, struct i915_hw_ppgtt, base);
615 gen8_pde_t scratch_pde;
617 scratch_pde = gen8_pde_encode(vm->dev, ppgtt->scratch_pt->base.daddr,
620 fill_page_dma(vm->dev, &pd->base, scratch_pde);
623 static void gen8_free_page_tables(struct i915_page_directory *pd, struct drm_device *dev)
630 for_each_set_bit(i, pd->used_pdes, I915_PDES) {
631 if (WARN_ON(!pd->page_table[i]))
634 free_pt(dev, pd->page_table[i]);
635 pd->page_table[i] = NULL;
639 static void gen8_ppgtt_cleanup(struct i915_address_space *vm)
641 struct i915_hw_ppgtt *ppgtt =
642 container_of(vm, struct i915_hw_ppgtt, base);
645 for_each_set_bit(i, ppgtt->pdp.used_pdpes, GEN8_LEGACY_PDPES) {
646 if (WARN_ON(!ppgtt->pdp.page_directory[i]))
649 gen8_free_page_tables(ppgtt->pdp.page_directory[i], ppgtt->base.dev);
650 free_pd(ppgtt->base.dev, ppgtt->pdp.page_directory[i]);
653 free_pd(ppgtt->base.dev, ppgtt->scratch_pd);
654 free_pt(ppgtt->base.dev, ppgtt->scratch_pt);
658 * gen8_ppgtt_alloc_pagetabs() - Allocate page tables for VA range.
659 * @ppgtt: Master ppgtt structure.
660 * @pd: Page directory for this address range.
661 * @start: Starting virtual address to begin allocations.
662 * @length Size of the allocations.
663 * @new_pts: Bitmap set by function with new allocations. Likely used by the
664 * caller to free on error.
666 * Allocate the required number of page tables. Extremely similar to
667 * gen8_ppgtt_alloc_page_directories(). The main difference is here we are limited by
668 * the page directory boundary (instead of the page directory pointer). That
669 * boundary is 1GB virtual. Therefore, unlike gen8_ppgtt_alloc_page_directories(), it is
670 * possible, and likely that the caller will need to use multiple calls of this
671 * function to achieve the appropriate allocation.
673 * Return: 0 if success; negative error code otherwise.
675 static int gen8_ppgtt_alloc_pagetabs(struct i915_hw_ppgtt *ppgtt,
676 struct i915_page_directory *pd,
679 unsigned long *new_pts)
681 struct drm_device *dev = ppgtt->base.dev;
682 struct i915_page_table *pt;
686 gen8_for_each_pde(pt, pd, start, length, temp, pde) {
687 /* Don't reallocate page tables */
689 /* Scratch is never allocated this way */
690 WARN_ON(pt == ppgtt->scratch_pt);
698 gen8_initialize_pt(&ppgtt->base, pt);
699 pd->page_table[pde] = pt;
700 set_bit(pde, new_pts);
706 for_each_set_bit(pde, new_pts, I915_PDES)
707 free_pt(dev, pd->page_table[pde]);
713 * gen8_ppgtt_alloc_page_directories() - Allocate page directories for VA range.
714 * @ppgtt: Master ppgtt structure.
715 * @pdp: Page directory pointer for this address range.
716 * @start: Starting virtual address to begin allocations.
717 * @length Size of the allocations.
718 * @new_pds Bitmap set by function with new allocations. Likely used by the
719 * caller to free on error.
721 * Allocate the required number of page directories starting at the pde index of
722 * @start, and ending at the pde index @start + @length. This function will skip
723 * over already allocated page directories within the range, and only allocate
724 * new ones, setting the appropriate pointer within the pdp as well as the
725 * correct position in the bitmap @new_pds.
727 * The function will only allocate the pages within the range for a give page
728 * directory pointer. In other words, if @start + @length straddles a virtually
729 * addressed PDP boundary (512GB for 4k pages), there will be more allocations
730 * required by the caller, This is not currently possible, and the BUG in the
731 * code will prevent it.
733 * Return: 0 if success; negative error code otherwise.
735 static int gen8_ppgtt_alloc_page_directories(struct i915_hw_ppgtt *ppgtt,
736 struct i915_page_directory_pointer *pdp,
739 unsigned long *new_pds)
741 struct drm_device *dev = ppgtt->base.dev;
742 struct i915_page_directory *pd;
746 WARN_ON(!bitmap_empty(new_pds, GEN8_LEGACY_PDPES));
748 gen8_for_each_pdpe(pd, pdp, start, length, temp, pdpe) {
756 gen8_initialize_pd(&ppgtt->base, pd);
757 pdp->page_directory[pdpe] = pd;
758 set_bit(pdpe, new_pds);
764 for_each_set_bit(pdpe, new_pds, GEN8_LEGACY_PDPES)
765 free_pd(dev, pdp->page_directory[pdpe]);
771 free_gen8_temp_bitmaps(unsigned long *new_pds, unsigned long **new_pts)
775 for (i = 0; i < GEN8_LEGACY_PDPES; i++)
781 /* Fills in the page directory bitmap, and the array of page tables bitmap. Both
782 * of these are based on the number of PDPEs in the system.
785 int __must_check alloc_gen8_temp_bitmaps(unsigned long **new_pds,
786 unsigned long ***new_pts)
792 pds = kcalloc(BITS_TO_LONGS(GEN8_LEGACY_PDPES), sizeof(unsigned long), GFP_KERNEL);
796 pts = kcalloc(GEN8_LEGACY_PDPES, sizeof(unsigned long *), GFP_KERNEL);
802 for (i = 0; i < GEN8_LEGACY_PDPES; i++) {
803 pts[i] = kcalloc(BITS_TO_LONGS(I915_PDES),
804 sizeof(unsigned long), GFP_KERNEL);
815 free_gen8_temp_bitmaps(pds, pts);
819 /* PDE TLBs are a pain to invalidate on GEN8+. When we modify
820 * the page table structures, we mark them dirty so that
821 * context switching/execlist queuing code takes extra steps
822 * to ensure that tlbs are flushed.
824 static void mark_tlbs_dirty(struct i915_hw_ppgtt *ppgtt)
826 ppgtt->pd_dirty_rings = INTEL_INFO(ppgtt->base.dev)->ring_mask;
829 static int gen8_alloc_va_range(struct i915_address_space *vm,
833 struct i915_hw_ppgtt *ppgtt =
834 container_of(vm, struct i915_hw_ppgtt, base);
835 unsigned long *new_page_dirs, **new_page_tables;
836 struct i915_page_directory *pd;
837 const uint64_t orig_start = start;
838 const uint64_t orig_length = length;
843 /* Wrap is never okay since we can only represent 48b, and we don't
844 * actually use the other side of the canonical address space.
846 if (WARN_ON(start + length < start))
849 if (WARN_ON(start + length > ppgtt->base.total))
852 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables);
856 /* Do the allocations first so we can easily bail out */
857 ret = gen8_ppgtt_alloc_page_directories(ppgtt, &ppgtt->pdp, start, length,
860 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
864 /* For every page directory referenced, allocate page tables */
865 gen8_for_each_pdpe(pd, &ppgtt->pdp, start, length, temp, pdpe) {
866 ret = gen8_ppgtt_alloc_pagetabs(ppgtt, pd, start, length,
867 new_page_tables[pdpe]);
873 length = orig_length;
875 /* Allocations have completed successfully, so set the bitmaps, and do
877 gen8_for_each_pdpe(pd, &ppgtt->pdp, start, length, temp, pdpe) {
878 gen8_pde_t *const page_directory = kmap_px(pd);
879 struct i915_page_table *pt;
880 uint64_t pd_len = gen8_clamp_pd(start, length);
881 uint64_t pd_start = start;
884 /* Every pd should be allocated, we just did that above. */
887 gen8_for_each_pde(pt, pd, pd_start, pd_len, temp, pde) {
888 /* Same reasoning as pd */
891 WARN_ON(!gen8_pte_count(pd_start, pd_len));
893 /* Set our used ptes within the page table */
894 bitmap_set(pt->used_ptes,
895 gen8_pte_index(pd_start),
896 gen8_pte_count(pd_start, pd_len));
898 /* Our pde is now pointing to the pagetable, pt */
899 set_bit(pde, pd->used_pdes);
901 /* Map the PDE to the page table */
902 __gen8_do_map_pt(page_directory + pde, pt, vm->dev);
904 /* NB: We haven't yet mapped ptes to pages. At this
905 * point we're still relying on insert_entries() */
908 kunmap_px(ppgtt, page_directory);
910 set_bit(pdpe, ppgtt->pdp.used_pdpes);
913 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
914 mark_tlbs_dirty(ppgtt);
919 for_each_set_bit(temp, new_page_tables[pdpe], I915_PDES)
920 free_pt(vm->dev, ppgtt->pdp.page_directory[pdpe]->page_table[temp]);
923 for_each_set_bit(pdpe, new_page_dirs, GEN8_LEGACY_PDPES)
924 free_pd(vm->dev, ppgtt->pdp.page_directory[pdpe]);
926 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
927 mark_tlbs_dirty(ppgtt);
932 * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers
933 * with a net effect resembling a 2-level page table in normal x86 terms. Each
934 * PDP represents 1GB of memory 4 * 512 * 512 * 4096 = 4GB legacy 32b address
938 static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
940 ppgtt->scratch_pt = alloc_pt(ppgtt->base.dev);
941 if (IS_ERR(ppgtt->scratch_pt))
942 return PTR_ERR(ppgtt->scratch_pt);
944 ppgtt->scratch_pd = alloc_pd(ppgtt->base.dev);
945 if (IS_ERR(ppgtt->scratch_pd))
946 return PTR_ERR(ppgtt->scratch_pd);
948 gen8_initialize_pt(&ppgtt->base, ppgtt->scratch_pt);
949 gen8_initialize_pd(&ppgtt->base, ppgtt->scratch_pd);
951 ppgtt->base.start = 0;
952 ppgtt->base.total = 1ULL << 32;
953 if (IS_ENABLED(CONFIG_X86_32))
954 /* While we have a proliferation of size_t variables
955 * we cannot represent the full ppgtt size on 32bit,
956 * so limit it to the same size as the GGTT (currently
959 ppgtt->base.total = to_i915(ppgtt->base.dev)->gtt.base.total;
960 ppgtt->base.cleanup = gen8_ppgtt_cleanup;
961 ppgtt->base.allocate_va_range = gen8_alloc_va_range;
962 ppgtt->base.insert_entries = gen8_ppgtt_insert_entries;
963 ppgtt->base.clear_range = gen8_ppgtt_clear_range;
964 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
965 ppgtt->base.bind_vma = ppgtt_bind_vma;
967 ppgtt->switch_mm = gen8_mm_switch;
972 static void gen6_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
974 struct i915_address_space *vm = &ppgtt->base;
975 struct i915_page_table *unused;
976 gen6_pte_t scratch_pte;
978 uint32_t pte, pde, temp;
979 uint32_t start = ppgtt->base.start, length = ppgtt->base.total;
981 scratch_pte = vm->pte_encode(vm->scratch.addr, I915_CACHE_LLC, true, 0);
983 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde) {
985 gen6_pte_t *pt_vaddr;
986 dma_addr_t pt_addr = ppgtt->pd.page_table[pde]->base.daddr;
987 pd_entry = readl(ppgtt->pd_addr + pde);
988 expected = (GEN6_PDE_ADDR_ENCODE(pt_addr) | GEN6_PDE_VALID);
990 if (pd_entry != expected)
991 seq_printf(m, "\tPDE #%d mismatch: Actual PDE: %x Expected PDE: %x\n",
995 seq_printf(m, "\tPDE: %x\n", pd_entry);
997 pt_vaddr = kmap_px(ppgtt->pd.page_table[pde]);
999 for (pte = 0; pte < GEN6_PTES; pte+=4) {
1001 (pde * PAGE_SIZE * GEN6_PTES) +
1005 for (i = 0; i < 4; i++)
1006 if (pt_vaddr[pte + i] != scratch_pte)
1011 seq_printf(m, "\t\t0x%lx [%03d,%04d]: =", va, pde, pte);
1012 for (i = 0; i < 4; i++) {
1013 if (pt_vaddr[pte + i] != scratch_pte)
1014 seq_printf(m, " %08x", pt_vaddr[pte + i]);
1016 seq_puts(m, " SCRATCH ");
1020 kunmap_px(ppgtt, pt_vaddr);
1024 /* Write pde (index) from the page directory @pd to the page table @pt */
1025 static void gen6_write_pde(struct i915_page_directory *pd,
1026 const int pde, struct i915_page_table *pt)
1028 /* Caller needs to make sure the write completes if necessary */
1029 struct i915_hw_ppgtt *ppgtt =
1030 container_of(pd, struct i915_hw_ppgtt, pd);
1033 pd_entry = GEN6_PDE_ADDR_ENCODE(pt->base.daddr);
1034 pd_entry |= GEN6_PDE_VALID;
1036 writel(pd_entry, ppgtt->pd_addr + pde);
1039 /* Write all the page tables found in the ppgtt structure to incrementing page
1041 static void gen6_write_page_range(struct drm_i915_private *dev_priv,
1042 struct i915_page_directory *pd,
1043 uint32_t start, uint32_t length)
1045 struct i915_page_table *pt;
1048 gen6_for_each_pde(pt, pd, start, length, temp, pde)
1049 gen6_write_pde(pd, pde, pt);
1051 /* Make sure write is complete before other code can use this page
1052 * table. Also require for WC mapped PTEs */
1053 readl(dev_priv->gtt.gsm);
1056 static uint32_t get_pd_offset(struct i915_hw_ppgtt *ppgtt)
1058 BUG_ON(ppgtt->pd.base.ggtt_offset & 0x3f);
1060 return (ppgtt->pd.base.ggtt_offset / 64) << 16;
1063 static int hsw_mm_switch(struct i915_hw_ppgtt *ppgtt,
1064 struct drm_i915_gem_request *req)
1066 struct intel_engine_cs *ring = req->ring;
1069 /* NB: TLBs must be flushed and invalidated before a switch */
1070 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1074 ret = intel_ring_begin(req, 6);
1078 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1079 intel_ring_emit(ring, RING_PP_DIR_DCLV(ring));
1080 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1081 intel_ring_emit(ring, RING_PP_DIR_BASE(ring));
1082 intel_ring_emit(ring, get_pd_offset(ppgtt));
1083 intel_ring_emit(ring, MI_NOOP);
1084 intel_ring_advance(ring);
1089 static int vgpu_mm_switch(struct i915_hw_ppgtt *ppgtt,
1090 struct drm_i915_gem_request *req)
1092 struct intel_engine_cs *ring = req->ring;
1093 struct drm_i915_private *dev_priv = to_i915(ppgtt->base.dev);
1095 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
1096 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
1100 static int gen7_mm_switch(struct i915_hw_ppgtt *ppgtt,
1101 struct drm_i915_gem_request *req)
1103 struct intel_engine_cs *ring = req->ring;
1106 /* NB: TLBs must be flushed and invalidated before a switch */
1107 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1111 ret = intel_ring_begin(req, 6);
1115 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1116 intel_ring_emit(ring, RING_PP_DIR_DCLV(ring));
1117 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1118 intel_ring_emit(ring, RING_PP_DIR_BASE(ring));
1119 intel_ring_emit(ring, get_pd_offset(ppgtt));
1120 intel_ring_emit(ring, MI_NOOP);
1121 intel_ring_advance(ring);
1123 /* XXX: RCS is the only one to auto invalidate the TLBs? */
1124 if (ring->id != RCS) {
1125 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1133 static int gen6_mm_switch(struct i915_hw_ppgtt *ppgtt,
1134 struct drm_i915_gem_request *req)
1136 struct intel_engine_cs *ring = req->ring;
1137 struct drm_device *dev = ppgtt->base.dev;
1138 struct drm_i915_private *dev_priv = dev->dev_private;
1141 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
1142 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
1144 POSTING_READ(RING_PP_DIR_DCLV(ring));
1149 static void gen8_ppgtt_enable(struct drm_device *dev)
1151 struct drm_i915_private *dev_priv = dev->dev_private;
1152 struct intel_engine_cs *ring;
1155 for_each_ring(ring, dev_priv, j) {
1156 I915_WRITE(RING_MODE_GEN7(ring),
1157 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1161 static void gen7_ppgtt_enable(struct drm_device *dev)
1163 struct drm_i915_private *dev_priv = dev->dev_private;
1164 struct intel_engine_cs *ring;
1165 uint32_t ecochk, ecobits;
1168 ecobits = I915_READ(GAC_ECO_BITS);
1169 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B);
1171 ecochk = I915_READ(GAM_ECOCHK);
1172 if (IS_HASWELL(dev)) {
1173 ecochk |= ECOCHK_PPGTT_WB_HSW;
1175 ecochk |= ECOCHK_PPGTT_LLC_IVB;
1176 ecochk &= ~ECOCHK_PPGTT_GFDT_IVB;
1178 I915_WRITE(GAM_ECOCHK, ecochk);
1180 for_each_ring(ring, dev_priv, i) {
1181 /* GFX_MODE is per-ring on gen7+ */
1182 I915_WRITE(RING_MODE_GEN7(ring),
1183 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1187 static void gen6_ppgtt_enable(struct drm_device *dev)
1189 struct drm_i915_private *dev_priv = dev->dev_private;
1190 uint32_t ecochk, gab_ctl, ecobits;
1192 ecobits = I915_READ(GAC_ECO_BITS);
1193 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_SNB_BIT |
1194 ECOBITS_PPGTT_CACHE64B);
1196 gab_ctl = I915_READ(GAB_CTL);
1197 I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT);
1199 ecochk = I915_READ(GAM_ECOCHK);
1200 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT | ECOCHK_PPGTT_CACHE64B);
1202 I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1205 /* PPGTT support for Sandybdrige/Gen6 and later */
1206 static void gen6_ppgtt_clear_range(struct i915_address_space *vm,
1211 struct i915_hw_ppgtt *ppgtt =
1212 container_of(vm, struct i915_hw_ppgtt, base);
1213 gen6_pte_t *pt_vaddr, scratch_pte;
1214 unsigned first_entry = start >> PAGE_SHIFT;
1215 unsigned num_entries = length >> PAGE_SHIFT;
1216 unsigned act_pt = first_entry / GEN6_PTES;
1217 unsigned first_pte = first_entry % GEN6_PTES;
1218 unsigned last_pte, i;
1220 scratch_pte = vm->pte_encode(vm->scratch.addr, I915_CACHE_LLC, true, 0);
1222 while (num_entries) {
1223 last_pte = first_pte + num_entries;
1224 if (last_pte > GEN6_PTES)
1225 last_pte = GEN6_PTES;
1227 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1229 for (i = first_pte; i < last_pte; i++)
1230 pt_vaddr[i] = scratch_pte;
1232 kunmap_px(ppgtt, pt_vaddr);
1234 num_entries -= last_pte - first_pte;
1240 static void gen6_ppgtt_insert_entries(struct i915_address_space *vm,
1241 struct sg_table *pages,
1243 enum i915_cache_level cache_level, u32 flags)
1245 struct i915_hw_ppgtt *ppgtt =
1246 container_of(vm, struct i915_hw_ppgtt, base);
1247 gen6_pte_t *pt_vaddr;
1248 unsigned first_entry = start >> PAGE_SHIFT;
1249 unsigned act_pt = first_entry / GEN6_PTES;
1250 unsigned act_pte = first_entry % GEN6_PTES;
1251 struct sg_page_iter sg_iter;
1254 for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) {
1255 if (pt_vaddr == NULL)
1256 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1259 vm->pte_encode(sg_page_iter_dma_address(&sg_iter),
1260 cache_level, true, flags);
1262 if (++act_pte == GEN6_PTES) {
1263 kunmap_px(ppgtt, pt_vaddr);
1270 kunmap_px(ppgtt, pt_vaddr);
1273 static void gen6_initialize_pt(struct i915_address_space *vm,
1274 struct i915_page_table *pt)
1276 gen6_pte_t scratch_pte;
1278 WARN_ON(vm->scratch.addr == 0);
1280 scratch_pte = vm->pte_encode(vm->scratch.addr, I915_CACHE_LLC, true, 0);
1282 fill_page_dma_32(vm->dev, &pt->base, scratch_pte);
1285 static int gen6_alloc_va_range(struct i915_address_space *vm,
1286 uint64_t start_in, uint64_t length_in)
1288 DECLARE_BITMAP(new_page_tables, I915_PDES);
1289 struct drm_device *dev = vm->dev;
1290 struct drm_i915_private *dev_priv = dev->dev_private;
1291 struct i915_hw_ppgtt *ppgtt =
1292 container_of(vm, struct i915_hw_ppgtt, base);
1293 struct i915_page_table *pt;
1294 uint32_t start, length, start_save, length_save;
1298 if (WARN_ON(start_in + length_in > ppgtt->base.total))
1301 start = start_save = start_in;
1302 length = length_save = length_in;
1304 bitmap_zero(new_page_tables, I915_PDES);
1306 /* The allocation is done in two stages so that we can bail out with
1307 * minimal amount of pain. The first stage finds new page tables that
1308 * need allocation. The second stage marks use ptes within the page
1311 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1312 if (pt != ppgtt->scratch_pt) {
1313 WARN_ON(bitmap_empty(pt->used_ptes, GEN6_PTES));
1317 /* We've already allocated a page table */
1318 WARN_ON(!bitmap_empty(pt->used_ptes, GEN6_PTES));
1326 gen6_initialize_pt(vm, pt);
1328 ppgtt->pd.page_table[pde] = pt;
1329 set_bit(pde, new_page_tables);
1330 trace_i915_page_table_entry_alloc(vm, pde, start, GEN6_PDE_SHIFT);
1334 length = length_save;
1336 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1337 DECLARE_BITMAP(tmp_bitmap, GEN6_PTES);
1339 bitmap_zero(tmp_bitmap, GEN6_PTES);
1340 bitmap_set(tmp_bitmap, gen6_pte_index(start),
1341 gen6_pte_count(start, length));
1343 if (test_and_clear_bit(pde, new_page_tables))
1344 gen6_write_pde(&ppgtt->pd, pde, pt);
1346 trace_i915_page_table_entry_map(vm, pde, pt,
1347 gen6_pte_index(start),
1348 gen6_pte_count(start, length),
1350 bitmap_or(pt->used_ptes, tmp_bitmap, pt->used_ptes,
1354 WARN_ON(!bitmap_empty(new_page_tables, I915_PDES));
1356 /* Make sure write is complete before other code can use this page
1357 * table. Also require for WC mapped PTEs */
1358 readl(dev_priv->gtt.gsm);
1360 mark_tlbs_dirty(ppgtt);
1364 for_each_set_bit(pde, new_page_tables, I915_PDES) {
1365 struct i915_page_table *pt = ppgtt->pd.page_table[pde];
1367 ppgtt->pd.page_table[pde] = ppgtt->scratch_pt;
1368 free_pt(vm->dev, pt);
1371 mark_tlbs_dirty(ppgtt);
1375 static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
1377 struct i915_hw_ppgtt *ppgtt =
1378 container_of(vm, struct i915_hw_ppgtt, base);
1379 struct i915_page_table *pt;
1383 drm_mm_remove_node(&ppgtt->node);
1385 gen6_for_all_pdes(pt, ppgtt, pde) {
1386 if (pt != ppgtt->scratch_pt)
1387 free_pt(ppgtt->base.dev, pt);
1390 free_pt(ppgtt->base.dev, ppgtt->scratch_pt);
1393 static int gen6_ppgtt_allocate_page_directories(struct i915_hw_ppgtt *ppgtt)
1395 struct drm_device *dev = ppgtt->base.dev;
1396 struct drm_i915_private *dev_priv = dev->dev_private;
1397 bool retried = false;
1400 /* PPGTT PDEs reside in the GGTT and consists of 512 entries. The
1401 * allocator works in address space sizes, so it's multiplied by page
1402 * size. We allocate at the top of the GTT to avoid fragmentation.
1404 BUG_ON(!drm_mm_initialized(&dev_priv->gtt.base.mm));
1405 ppgtt->scratch_pt = alloc_pt(ppgtt->base.dev);
1406 if (IS_ERR(ppgtt->scratch_pt))
1407 return PTR_ERR(ppgtt->scratch_pt);
1409 gen6_initialize_pt(&ppgtt->base, ppgtt->scratch_pt);
1412 ret = drm_mm_insert_node_in_range_generic(&dev_priv->gtt.base.mm,
1413 &ppgtt->node, GEN6_PD_SIZE,
1415 0, dev_priv->gtt.base.total,
1417 if (ret == -ENOSPC && !retried) {
1418 ret = i915_gem_evict_something(dev, &dev_priv->gtt.base,
1419 GEN6_PD_SIZE, GEN6_PD_ALIGN,
1421 0, dev_priv->gtt.base.total,
1434 if (ppgtt->node.start < dev_priv->gtt.mappable_end)
1435 DRM_DEBUG("Forced to use aperture for PDEs\n");
1440 free_pt(ppgtt->base.dev, ppgtt->scratch_pt);
1444 static int gen6_ppgtt_alloc(struct i915_hw_ppgtt *ppgtt)
1446 return gen6_ppgtt_allocate_page_directories(ppgtt);
1449 static void gen6_scratch_va_range(struct i915_hw_ppgtt *ppgtt,
1450 uint64_t start, uint64_t length)
1452 struct i915_page_table *unused;
1455 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde)
1456 ppgtt->pd.page_table[pde] = ppgtt->scratch_pt;
1459 static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
1461 struct drm_device *dev = ppgtt->base.dev;
1462 struct drm_i915_private *dev_priv = dev->dev_private;
1465 ppgtt->base.pte_encode = dev_priv->gtt.base.pte_encode;
1467 ppgtt->switch_mm = gen6_mm_switch;
1468 } else if (IS_HASWELL(dev)) {
1469 ppgtt->switch_mm = hsw_mm_switch;
1470 } else if (IS_GEN7(dev)) {
1471 ppgtt->switch_mm = gen7_mm_switch;
1475 if (intel_vgpu_active(dev))
1476 ppgtt->switch_mm = vgpu_mm_switch;
1478 ret = gen6_ppgtt_alloc(ppgtt);
1482 ppgtt->base.allocate_va_range = gen6_alloc_va_range;
1483 ppgtt->base.clear_range = gen6_ppgtt_clear_range;
1484 ppgtt->base.insert_entries = gen6_ppgtt_insert_entries;
1485 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
1486 ppgtt->base.bind_vma = ppgtt_bind_vma;
1487 ppgtt->base.cleanup = gen6_ppgtt_cleanup;
1488 ppgtt->base.start = 0;
1489 ppgtt->base.total = I915_PDES * GEN6_PTES * PAGE_SIZE;
1490 ppgtt->debug_dump = gen6_dump_ppgtt;
1492 ppgtt->pd.base.ggtt_offset =
1493 ppgtt->node.start / PAGE_SIZE * sizeof(gen6_pte_t);
1495 ppgtt->pd_addr = (gen6_pte_t __iomem *)dev_priv->gtt.gsm +
1496 ppgtt->pd.base.ggtt_offset / sizeof(gen6_pte_t);
1498 gen6_scratch_va_range(ppgtt, 0, ppgtt->base.total);
1500 gen6_write_page_range(dev_priv, &ppgtt->pd, 0, ppgtt->base.total);
1502 DRM_DEBUG_DRIVER("Allocated pde space (%lldM) at GTT entry: %llx\n",
1503 ppgtt->node.size >> 20,
1504 ppgtt->node.start / PAGE_SIZE);
1506 DRM_DEBUG("Adding PPGTT at offset %x\n",
1507 ppgtt->pd.base.ggtt_offset << 10);
1512 static int __hw_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
1514 struct drm_i915_private *dev_priv = dev->dev_private;
1516 ppgtt->base.dev = dev;
1517 ppgtt->base.scratch = dev_priv->gtt.base.scratch;
1519 if (INTEL_INFO(dev)->gen < 8)
1520 return gen6_ppgtt_init(ppgtt);
1522 return gen8_ppgtt_init(ppgtt);
1524 int i915_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
1526 struct drm_i915_private *dev_priv = dev->dev_private;
1529 ret = __hw_ppgtt_init(dev, ppgtt);
1531 kref_init(&ppgtt->ref);
1532 drm_mm_init(&ppgtt->base.mm, ppgtt->base.start,
1534 i915_init_vm(dev_priv, &ppgtt->base);
1540 int i915_ppgtt_init_hw(struct drm_device *dev)
1542 /* In the case of execlists, PPGTT is enabled by the context descriptor
1543 * and the PDPs are contained within the context itself. We don't
1544 * need to do anything here. */
1545 if (i915.enable_execlists)
1548 if (!USES_PPGTT(dev))
1552 gen6_ppgtt_enable(dev);
1553 else if (IS_GEN7(dev))
1554 gen7_ppgtt_enable(dev);
1555 else if (INTEL_INFO(dev)->gen >= 8)
1556 gen8_ppgtt_enable(dev);
1558 MISSING_CASE(INTEL_INFO(dev)->gen);
1563 int i915_ppgtt_init_ring(struct drm_i915_gem_request *req)
1565 struct drm_i915_private *dev_priv = req->ring->dev->dev_private;
1566 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
1568 if (i915.enable_execlists)
1574 return ppgtt->switch_mm(ppgtt, req);
1577 struct i915_hw_ppgtt *
1578 i915_ppgtt_create(struct drm_device *dev, struct drm_i915_file_private *fpriv)
1580 struct i915_hw_ppgtt *ppgtt;
1583 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
1585 return ERR_PTR(-ENOMEM);
1587 ret = i915_ppgtt_init(dev, ppgtt);
1590 return ERR_PTR(ret);
1593 ppgtt->file_priv = fpriv;
1595 trace_i915_ppgtt_create(&ppgtt->base);
1600 void i915_ppgtt_release(struct kref *kref)
1602 struct i915_hw_ppgtt *ppgtt =
1603 container_of(kref, struct i915_hw_ppgtt, ref);
1605 trace_i915_ppgtt_release(&ppgtt->base);
1607 /* vmas should already be unbound */
1608 WARN_ON(!list_empty(&ppgtt->base.active_list));
1609 WARN_ON(!list_empty(&ppgtt->base.inactive_list));
1611 list_del(&ppgtt->base.global_link);
1612 drm_mm_takedown(&ppgtt->base.mm);
1614 ppgtt->base.cleanup(&ppgtt->base);
1618 extern int intel_iommu_gfx_mapped;
1619 /* Certain Gen5 chipsets require require idling the GPU before
1620 * unmapping anything from the GTT when VT-d is enabled.
1622 static bool needs_idle_maps(struct drm_device *dev)
1624 #ifdef CONFIG_INTEL_IOMMU
1625 /* Query intel_iommu to see if we need the workaround. Presumably that
1628 if (IS_GEN5(dev) && IS_MOBILE(dev) && intel_iommu_gfx_mapped)
1634 static bool do_idling(struct drm_i915_private *dev_priv)
1636 bool ret = dev_priv->mm.interruptible;
1638 if (unlikely(dev_priv->gtt.do_idle_maps)) {
1639 dev_priv->mm.interruptible = false;
1640 if (i915_gpu_idle(dev_priv->dev)) {
1641 DRM_ERROR("Couldn't idle GPU\n");
1642 /* Wait a bit, in hopes it avoids the hang */
1650 static void undo_idling(struct drm_i915_private *dev_priv, bool interruptible)
1652 if (unlikely(dev_priv->gtt.do_idle_maps))
1653 dev_priv->mm.interruptible = interruptible;
1656 void i915_check_and_clear_faults(struct drm_device *dev)
1658 struct drm_i915_private *dev_priv = dev->dev_private;
1659 struct intel_engine_cs *ring;
1662 if (INTEL_INFO(dev)->gen < 6)
1665 for_each_ring(ring, dev_priv, i) {
1667 fault_reg = I915_READ(RING_FAULT_REG(ring));
1668 if (fault_reg & RING_FAULT_VALID) {
1669 DRM_DEBUG_DRIVER("Unexpected fault\n"
1671 "\tAddress space: %s\n"
1674 fault_reg & PAGE_MASK,
1675 fault_reg & RING_FAULT_GTTSEL_MASK ? "GGTT" : "PPGTT",
1676 RING_FAULT_SRCID(fault_reg),
1677 RING_FAULT_FAULT_TYPE(fault_reg));
1678 I915_WRITE(RING_FAULT_REG(ring),
1679 fault_reg & ~RING_FAULT_VALID);
1682 POSTING_READ(RING_FAULT_REG(&dev_priv->ring[RCS]));
1685 static void i915_ggtt_flush(struct drm_i915_private *dev_priv)
1687 if (INTEL_INFO(dev_priv->dev)->gen < 6) {
1688 intel_gtt_chipset_flush();
1690 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
1691 POSTING_READ(GFX_FLSH_CNTL_GEN6);
1695 void i915_gem_suspend_gtt_mappings(struct drm_device *dev)
1697 struct drm_i915_private *dev_priv = dev->dev_private;
1699 /* Don't bother messing with faults pre GEN6 as we have little
1700 * documentation supporting that it's a good idea.
1702 if (INTEL_INFO(dev)->gen < 6)
1705 i915_check_and_clear_faults(dev);
1707 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
1708 dev_priv->gtt.base.start,
1709 dev_priv->gtt.base.total,
1712 i915_ggtt_flush(dev_priv);
1715 int i915_gem_gtt_prepare_object(struct drm_i915_gem_object *obj)
1717 if (obj->has_dma_mapping)
1720 if (!dma_map_sg(&obj->base.dev->pdev->dev,
1721 obj->pages->sgl, obj->pages->nents,
1722 PCI_DMA_BIDIRECTIONAL))
1728 static void gen8_set_pte(void __iomem *addr, gen8_pte_t pte)
1733 iowrite32((u32)pte, addr);
1734 iowrite32(pte >> 32, addr + 4);
1738 static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
1739 struct sg_table *st,
1741 enum i915_cache_level level, u32 unused)
1743 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1744 unsigned first_entry = start >> PAGE_SHIFT;
1745 gen8_pte_t __iomem *gtt_entries =
1746 (gen8_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
1748 struct sg_page_iter sg_iter;
1749 dma_addr_t addr = 0; /* shut up gcc */
1751 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
1752 addr = sg_dma_address(sg_iter.sg) +
1753 (sg_iter.sg_pgoffset << PAGE_SHIFT);
1754 gen8_set_pte(>t_entries[i],
1755 gen8_pte_encode(addr, level, true));
1760 * XXX: This serves as a posting read to make sure that the PTE has
1761 * actually been updated. There is some concern that even though
1762 * registers and PTEs are within the same BAR that they are potentially
1763 * of NUMA access patterns. Therefore, even with the way we assume
1764 * hardware should work, we must keep this posting read for paranoia.
1767 WARN_ON(readq(>t_entries[i-1])
1768 != gen8_pte_encode(addr, level, true));
1770 /* This next bit makes the above posting read even more important. We
1771 * want to flush the TLBs only after we're certain all the PTE updates
1774 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
1775 POSTING_READ(GFX_FLSH_CNTL_GEN6);
1779 * Binds an object into the global gtt with the specified cache level. The object
1780 * will be accessible to the GPU via commands whose operands reference offsets
1781 * within the global GTT as well as accessible by the GPU through the GMADR
1782 * mapped BAR (dev_priv->mm.gtt->gtt).
1784 static void gen6_ggtt_insert_entries(struct i915_address_space *vm,
1785 struct sg_table *st,
1787 enum i915_cache_level level, u32 flags)
1789 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1790 unsigned first_entry = start >> PAGE_SHIFT;
1791 gen6_pte_t __iomem *gtt_entries =
1792 (gen6_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
1794 struct sg_page_iter sg_iter;
1795 dma_addr_t addr = 0;
1797 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
1798 addr = sg_page_iter_dma_address(&sg_iter);
1799 iowrite32(vm->pte_encode(addr, level, true, flags), >t_entries[i]);
1803 /* XXX: This serves as a posting read to make sure that the PTE has
1804 * actually been updated. There is some concern that even though
1805 * registers and PTEs are within the same BAR that they are potentially
1806 * of NUMA access patterns. Therefore, even with the way we assume
1807 * hardware should work, we must keep this posting read for paranoia.
1810 unsigned long gtt = readl(>t_entries[i-1]);
1811 WARN_ON(gtt != vm->pte_encode(addr, level, true, flags));
1814 /* This next bit makes the above posting read even more important. We
1815 * want to flush the TLBs only after we're certain all the PTE updates
1818 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
1819 POSTING_READ(GFX_FLSH_CNTL_GEN6);
1822 static void gen8_ggtt_clear_range(struct i915_address_space *vm,
1827 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1828 unsigned first_entry = start >> PAGE_SHIFT;
1829 unsigned num_entries = length >> PAGE_SHIFT;
1830 gen8_pte_t scratch_pte, __iomem *gtt_base =
1831 (gen8_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
1832 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
1835 if (WARN(num_entries > max_entries,
1836 "First entry = %d; Num entries = %d (max=%d)\n",
1837 first_entry, num_entries, max_entries))
1838 num_entries = max_entries;
1840 scratch_pte = gen8_pte_encode(vm->scratch.addr,
1843 for (i = 0; i < num_entries; i++)
1844 gen8_set_pte(>t_base[i], scratch_pte);
1848 static void gen6_ggtt_clear_range(struct i915_address_space *vm,
1853 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1854 unsigned first_entry = start >> PAGE_SHIFT;
1855 unsigned num_entries = length >> PAGE_SHIFT;
1856 gen6_pte_t scratch_pte, __iomem *gtt_base =
1857 (gen6_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
1858 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
1861 if (WARN(num_entries > max_entries,
1862 "First entry = %d; Num entries = %d (max=%d)\n",
1863 first_entry, num_entries, max_entries))
1864 num_entries = max_entries;
1866 scratch_pte = vm->pte_encode(vm->scratch.addr, I915_CACHE_LLC, use_scratch, 0);
1868 for (i = 0; i < num_entries; i++)
1869 iowrite32(scratch_pte, >t_base[i]);
1873 static void i915_ggtt_insert_entries(struct i915_address_space *vm,
1874 struct sg_table *pages,
1876 enum i915_cache_level cache_level, u32 unused)
1878 unsigned int flags = (cache_level == I915_CACHE_NONE) ?
1879 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
1881 intel_gtt_insert_sg_entries(pages, start >> PAGE_SHIFT, flags);
1885 static void i915_ggtt_clear_range(struct i915_address_space *vm,
1890 unsigned first_entry = start >> PAGE_SHIFT;
1891 unsigned num_entries = length >> PAGE_SHIFT;
1892 intel_gtt_clear_range(first_entry, num_entries);
1895 static int ggtt_bind_vma(struct i915_vma *vma,
1896 enum i915_cache_level cache_level,
1899 struct drm_device *dev = vma->vm->dev;
1900 struct drm_i915_private *dev_priv = dev->dev_private;
1901 struct drm_i915_gem_object *obj = vma->obj;
1902 struct sg_table *pages = obj->pages;
1906 ret = i915_get_ggtt_vma_pages(vma);
1909 pages = vma->ggtt_view.pages;
1911 /* Currently applicable only to VLV */
1913 pte_flags |= PTE_READ_ONLY;
1916 if (!dev_priv->mm.aliasing_ppgtt || flags & GLOBAL_BIND) {
1917 vma->vm->insert_entries(vma->vm, pages,
1919 cache_level, pte_flags);
1922 if (dev_priv->mm.aliasing_ppgtt && flags & LOCAL_BIND) {
1923 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
1924 appgtt->base.insert_entries(&appgtt->base, pages,
1926 cache_level, pte_flags);
1932 static void ggtt_unbind_vma(struct i915_vma *vma)
1934 struct drm_device *dev = vma->vm->dev;
1935 struct drm_i915_private *dev_priv = dev->dev_private;
1936 struct drm_i915_gem_object *obj = vma->obj;
1937 const uint64_t size = min_t(uint64_t,
1941 if (vma->bound & GLOBAL_BIND) {
1942 vma->vm->clear_range(vma->vm,
1948 if (dev_priv->mm.aliasing_ppgtt && vma->bound & LOCAL_BIND) {
1949 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
1951 appgtt->base.clear_range(&appgtt->base,
1958 void i915_gem_gtt_finish_object(struct drm_i915_gem_object *obj)
1960 struct drm_device *dev = obj->base.dev;
1961 struct drm_i915_private *dev_priv = dev->dev_private;
1964 interruptible = do_idling(dev_priv);
1966 if (!obj->has_dma_mapping)
1967 dma_unmap_sg(&dev->pdev->dev,
1968 obj->pages->sgl, obj->pages->nents,
1969 PCI_DMA_BIDIRECTIONAL);
1971 undo_idling(dev_priv, interruptible);
1974 static void i915_gtt_color_adjust(struct drm_mm_node *node,
1975 unsigned long color,
1979 if (node->color != color)
1982 if (!list_empty(&node->node_list)) {
1983 node = list_entry(node->node_list.next,
1986 if (node->allocated && node->color != color)
1991 static int i915_gem_setup_global_gtt(struct drm_device *dev,
1992 unsigned long start,
1993 unsigned long mappable_end,
1996 /* Let GEM Manage all of the aperture.
1998 * However, leave one page at the end still bound to the scratch page.
1999 * There are a number of places where the hardware apparently prefetches
2000 * past the end of the object, and we've seen multiple hangs with the
2001 * GPU head pointer stuck in a batchbuffer bound at the last page of the
2002 * aperture. One page should be enough to keep any prefetching inside
2005 struct drm_i915_private *dev_priv = dev->dev_private;
2006 struct i915_address_space *ggtt_vm = &dev_priv->gtt.base;
2007 struct drm_mm_node *entry;
2008 struct drm_i915_gem_object *obj;
2009 unsigned long hole_start, hole_end;
2012 BUG_ON(mappable_end > end);
2014 /* Subtract the guard page ... */
2015 drm_mm_init(&ggtt_vm->mm, start, end - start - PAGE_SIZE);
2017 dev_priv->gtt.base.start = start;
2018 dev_priv->gtt.base.total = end - start;
2020 if (intel_vgpu_active(dev)) {
2021 ret = intel_vgt_balloon(dev);
2027 dev_priv->gtt.base.mm.color_adjust = i915_gtt_color_adjust;
2029 /* Mark any preallocated objects as occupied */
2030 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
2031 struct i915_vma *vma = i915_gem_obj_to_vma(obj, ggtt_vm);
2033 DRM_DEBUG_KMS("reserving preallocated space: %lx + %zx\n",
2034 i915_gem_obj_ggtt_offset(obj), obj->base.size);
2036 WARN_ON(i915_gem_obj_ggtt_bound(obj));
2037 ret = drm_mm_reserve_node(&ggtt_vm->mm, &vma->node);
2039 DRM_DEBUG_KMS("Reservation failed: %i\n", ret);
2042 vma->bound |= GLOBAL_BIND;
2045 /* Clear any non-preallocated blocks */
2046 drm_mm_for_each_hole(entry, &ggtt_vm->mm, hole_start, hole_end) {
2047 DRM_DEBUG_KMS("clearing unused GTT space: [%lx, %lx]\n",
2048 hole_start, hole_end);
2049 ggtt_vm->clear_range(ggtt_vm, hole_start,
2050 hole_end - hole_start, true);
2053 /* And finally clear the reserved guard page */
2054 ggtt_vm->clear_range(ggtt_vm, end - PAGE_SIZE, PAGE_SIZE, true);
2056 if (USES_PPGTT(dev) && !USES_FULL_PPGTT(dev)) {
2057 struct i915_hw_ppgtt *ppgtt;
2059 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2063 ret = __hw_ppgtt_init(dev, ppgtt);
2065 ppgtt->base.cleanup(&ppgtt->base);
2070 if (ppgtt->base.allocate_va_range)
2071 ret = ppgtt->base.allocate_va_range(&ppgtt->base, 0,
2074 ppgtt->base.cleanup(&ppgtt->base);
2079 ppgtt->base.clear_range(&ppgtt->base,
2084 dev_priv->mm.aliasing_ppgtt = ppgtt;
2090 void i915_gem_init_global_gtt(struct drm_device *dev)
2092 struct drm_i915_private *dev_priv = dev->dev_private;
2093 u64 gtt_size, mappable_size;
2095 gtt_size = dev_priv->gtt.base.total;
2096 mappable_size = dev_priv->gtt.mappable_end;
2098 i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
2101 void i915_global_gtt_cleanup(struct drm_device *dev)
2103 struct drm_i915_private *dev_priv = dev->dev_private;
2104 struct i915_address_space *vm = &dev_priv->gtt.base;
2106 if (dev_priv->mm.aliasing_ppgtt) {
2107 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
2109 ppgtt->base.cleanup(&ppgtt->base);
2112 if (drm_mm_initialized(&vm->mm)) {
2113 if (intel_vgpu_active(dev))
2114 intel_vgt_deballoon();
2116 drm_mm_takedown(&vm->mm);
2117 list_del(&vm->global_link);
2123 static int setup_scratch_page(struct drm_device *dev)
2125 struct drm_i915_private *dev_priv = dev->dev_private;
2127 dma_addr_t dma_addr;
2129 page = alloc_page(GFP_KERNEL | GFP_DMA32 | __GFP_ZERO);
2132 set_pages_uc(page, 1);
2134 #ifdef CONFIG_INTEL_IOMMU
2135 dma_addr = pci_map_page(dev->pdev, page, 0, PAGE_SIZE,
2136 PCI_DMA_BIDIRECTIONAL);
2137 if (pci_dma_mapping_error(dev->pdev, dma_addr)) {
2142 dma_addr = page_to_phys(page);
2144 dev_priv->gtt.base.scratch.page = page;
2145 dev_priv->gtt.base.scratch.addr = dma_addr;
2150 static void teardown_scratch_page(struct drm_device *dev)
2152 struct drm_i915_private *dev_priv = dev->dev_private;
2153 struct page *page = dev_priv->gtt.base.scratch.page;
2155 set_pages_wb(page, 1);
2156 pci_unmap_page(dev->pdev, dev_priv->gtt.base.scratch.addr,
2157 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
2161 static unsigned int gen6_get_total_gtt_size(u16 snb_gmch_ctl)
2163 snb_gmch_ctl >>= SNB_GMCH_GGMS_SHIFT;
2164 snb_gmch_ctl &= SNB_GMCH_GGMS_MASK;
2165 return snb_gmch_ctl << 20;
2168 static unsigned int gen8_get_total_gtt_size(u16 bdw_gmch_ctl)
2170 bdw_gmch_ctl >>= BDW_GMCH_GGMS_SHIFT;
2171 bdw_gmch_ctl &= BDW_GMCH_GGMS_MASK;
2173 bdw_gmch_ctl = 1 << bdw_gmch_ctl;
2175 #ifdef CONFIG_X86_32
2176 /* Limit 32b platforms to a 2GB GGTT: 4 << 20 / pte size * PAGE_SIZE */
2177 if (bdw_gmch_ctl > 4)
2181 return bdw_gmch_ctl << 20;
2184 static unsigned int chv_get_total_gtt_size(u16 gmch_ctrl)
2186 gmch_ctrl >>= SNB_GMCH_GGMS_SHIFT;
2187 gmch_ctrl &= SNB_GMCH_GGMS_MASK;
2190 return 1 << (20 + gmch_ctrl);
2195 static size_t gen6_get_stolen_size(u16 snb_gmch_ctl)
2197 snb_gmch_ctl >>= SNB_GMCH_GMS_SHIFT;
2198 snb_gmch_ctl &= SNB_GMCH_GMS_MASK;
2199 return snb_gmch_ctl << 25; /* 32 MB units */
2202 static size_t gen8_get_stolen_size(u16 bdw_gmch_ctl)
2204 bdw_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2205 bdw_gmch_ctl &= BDW_GMCH_GMS_MASK;
2206 return bdw_gmch_ctl << 25; /* 32 MB units */
2209 static size_t chv_get_stolen_size(u16 gmch_ctrl)
2211 gmch_ctrl >>= SNB_GMCH_GMS_SHIFT;
2212 gmch_ctrl &= SNB_GMCH_GMS_MASK;
2215 * 0x0 to 0x10: 32MB increments starting at 0MB
2216 * 0x11 to 0x16: 4MB increments starting at 8MB
2217 * 0x17 to 0x1d: 4MB increments start at 36MB
2219 if (gmch_ctrl < 0x11)
2220 return gmch_ctrl << 25;
2221 else if (gmch_ctrl < 0x17)
2222 return (gmch_ctrl - 0x11 + 2) << 22;
2224 return (gmch_ctrl - 0x17 + 9) << 22;
2227 static size_t gen9_get_stolen_size(u16 gen9_gmch_ctl)
2229 gen9_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2230 gen9_gmch_ctl &= BDW_GMCH_GMS_MASK;
2232 if (gen9_gmch_ctl < 0xf0)
2233 return gen9_gmch_ctl << 25; /* 32 MB units */
2235 /* 4MB increments starting at 0xf0 for 4MB */
2236 return (gen9_gmch_ctl - 0xf0 + 1) << 22;
2239 static int ggtt_probe_common(struct drm_device *dev,
2242 struct drm_i915_private *dev_priv = dev->dev_private;
2243 phys_addr_t gtt_phys_addr;
2246 /* For Modern GENs the PTEs and register space are split in the BAR */
2247 gtt_phys_addr = pci_resource_start(dev->pdev, 0) +
2248 (pci_resource_len(dev->pdev, 0) / 2);
2251 * On BXT writes larger than 64 bit to the GTT pagetable range will be
2252 * dropped. For WC mappings in general we have 64 byte burst writes
2253 * when the WC buffer is flushed, so we can't use it, but have to
2254 * resort to an uncached mapping. The WC issue is easily caught by the
2255 * readback check when writing GTT PTE entries.
2257 if (IS_BROXTON(dev))
2258 dev_priv->gtt.gsm = ioremap_nocache(gtt_phys_addr, gtt_size);
2260 dev_priv->gtt.gsm = ioremap_wc(gtt_phys_addr, gtt_size);
2261 if (!dev_priv->gtt.gsm) {
2262 DRM_ERROR("Failed to map the gtt page table\n");
2266 ret = setup_scratch_page(dev);
2268 DRM_ERROR("Scratch setup failed\n");
2269 /* iounmap will also get called at remove, but meh */
2270 iounmap(dev_priv->gtt.gsm);
2276 /* The GGTT and PPGTT need a private PPAT setup in order to handle cacheability
2277 * bits. When using advanced contexts each context stores its own PAT, but
2278 * writing this data shouldn't be harmful even in those cases. */
2279 static void bdw_setup_private_ppat(struct drm_i915_private *dev_priv)
2283 pat = GEN8_PPAT(0, GEN8_PPAT_WB | GEN8_PPAT_LLC) | /* for normal objects, no eLLC */
2284 GEN8_PPAT(1, GEN8_PPAT_WC | GEN8_PPAT_LLCELLC) | /* for something pointing to ptes? */
2285 GEN8_PPAT(2, GEN8_PPAT_WT | GEN8_PPAT_LLCELLC) | /* for scanout with eLLC */
2286 GEN8_PPAT(3, GEN8_PPAT_UC) | /* Uncached objects, mostly for scanout */
2287 GEN8_PPAT(4, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(0)) |
2288 GEN8_PPAT(5, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(1)) |
2289 GEN8_PPAT(6, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(2)) |
2290 GEN8_PPAT(7, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3));
2292 if (!USES_PPGTT(dev_priv->dev))
2293 /* Spec: "For GGTT, there is NO pat_sel[2:0] from the entry,
2294 * so RTL will always use the value corresponding to
2296 * So let's disable cache for GGTT to avoid screen corruptions.
2297 * MOCS still can be used though.
2298 * - System agent ggtt writes (i.e. cpu gtt mmaps) already work
2299 * before this patch, i.e. the same uncached + snooping access
2300 * like on gen6/7 seems to be in effect.
2301 * - So this just fixes blitter/render access. Again it looks
2302 * like it's not just uncached access, but uncached + snooping.
2303 * So we can still hold onto all our assumptions wrt cpu
2304 * clflushing on LLC machines.
2306 pat = GEN8_PPAT(0, GEN8_PPAT_UC);
2308 /* XXX: spec defines this as 2 distinct registers. It's unclear if a 64b
2309 * write would work. */
2310 I915_WRITE(GEN8_PRIVATE_PAT, pat);
2311 I915_WRITE(GEN8_PRIVATE_PAT + 4, pat >> 32);
2314 static void chv_setup_private_ppat(struct drm_i915_private *dev_priv)
2319 * Map WB on BDW to snooped on CHV.
2321 * Only the snoop bit has meaning for CHV, the rest is
2324 * The hardware will never snoop for certain types of accesses:
2325 * - CPU GTT (GMADR->GGTT->no snoop->memory)
2326 * - PPGTT page tables
2327 * - some other special cycles
2329 * As with BDW, we also need to consider the following for GT accesses:
2330 * "For GGTT, there is NO pat_sel[2:0] from the entry,
2331 * so RTL will always use the value corresponding to
2333 * Which means we must set the snoop bit in PAT entry 0
2334 * in order to keep the global status page working.
2336 pat = GEN8_PPAT(0, CHV_PPAT_SNOOP) |
2340 GEN8_PPAT(4, CHV_PPAT_SNOOP) |
2341 GEN8_PPAT(5, CHV_PPAT_SNOOP) |
2342 GEN8_PPAT(6, CHV_PPAT_SNOOP) |
2343 GEN8_PPAT(7, CHV_PPAT_SNOOP);
2345 I915_WRITE(GEN8_PRIVATE_PAT, pat);
2346 I915_WRITE(GEN8_PRIVATE_PAT + 4, pat >> 32);
2349 static int gen8_gmch_probe(struct drm_device *dev,
2352 phys_addr_t *mappable_base,
2355 struct drm_i915_private *dev_priv = dev->dev_private;
2360 /* TODO: We're not aware of mappable constraints on gen8 yet */
2361 *mappable_base = pci_resource_start(dev->pdev, 2);
2362 *mappable_end = pci_resource_len(dev->pdev, 2);
2364 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(39)))
2365 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(39));
2367 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
2369 if (INTEL_INFO(dev)->gen >= 9) {
2370 *stolen = gen9_get_stolen_size(snb_gmch_ctl);
2371 gtt_size = gen8_get_total_gtt_size(snb_gmch_ctl);
2372 } else if (IS_CHERRYVIEW(dev)) {
2373 *stolen = chv_get_stolen_size(snb_gmch_ctl);
2374 gtt_size = chv_get_total_gtt_size(snb_gmch_ctl);
2376 *stolen = gen8_get_stolen_size(snb_gmch_ctl);
2377 gtt_size = gen8_get_total_gtt_size(snb_gmch_ctl);
2380 *gtt_total = (gtt_size / sizeof(gen8_pte_t)) << PAGE_SHIFT;
2382 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
2383 chv_setup_private_ppat(dev_priv);
2385 bdw_setup_private_ppat(dev_priv);
2387 ret = ggtt_probe_common(dev, gtt_size);
2389 dev_priv->gtt.base.clear_range = gen8_ggtt_clear_range;
2390 dev_priv->gtt.base.insert_entries = gen8_ggtt_insert_entries;
2391 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
2392 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
2397 static int gen6_gmch_probe(struct drm_device *dev,
2400 phys_addr_t *mappable_base,
2403 struct drm_i915_private *dev_priv = dev->dev_private;
2404 unsigned int gtt_size;
2408 *mappable_base = pci_resource_start(dev->pdev, 2);
2409 *mappable_end = pci_resource_len(dev->pdev, 2);
2411 /* 64/512MB is the current min/max we actually know of, but this is just
2412 * a coarse sanity check.
2414 if ((*mappable_end < (64<<20) || (*mappable_end > (512<<20)))) {
2415 DRM_ERROR("Unknown GMADR size (%llx)\n",
2416 dev_priv->gtt.mappable_end);
2420 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(40)))
2421 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(40));
2422 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
2424 *stolen = gen6_get_stolen_size(snb_gmch_ctl);
2426 gtt_size = gen6_get_total_gtt_size(snb_gmch_ctl);
2427 *gtt_total = (gtt_size / sizeof(gen6_pte_t)) << PAGE_SHIFT;
2429 ret = ggtt_probe_common(dev, gtt_size);
2431 dev_priv->gtt.base.clear_range = gen6_ggtt_clear_range;
2432 dev_priv->gtt.base.insert_entries = gen6_ggtt_insert_entries;
2433 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
2434 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
2439 static void gen6_gmch_remove(struct i915_address_space *vm)
2442 struct i915_gtt *gtt = container_of(vm, struct i915_gtt, base);
2445 teardown_scratch_page(vm->dev);
2448 static int i915_gmch_probe(struct drm_device *dev,
2451 phys_addr_t *mappable_base,
2454 struct drm_i915_private *dev_priv = dev->dev_private;
2457 ret = intel_gmch_probe(dev_priv->bridge_dev, dev_priv->dev->pdev, NULL);
2459 DRM_ERROR("failed to set up gmch\n");
2463 intel_gtt_get(gtt_total, stolen, mappable_base, mappable_end);
2465 dev_priv->gtt.do_idle_maps = needs_idle_maps(dev_priv->dev);
2466 dev_priv->gtt.base.insert_entries = i915_ggtt_insert_entries;
2467 dev_priv->gtt.base.clear_range = i915_ggtt_clear_range;
2468 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
2469 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
2471 if (unlikely(dev_priv->gtt.do_idle_maps))
2472 DRM_INFO("applying Ironlake quirks for intel_iommu\n");
2477 static void i915_gmch_remove(struct i915_address_space *vm)
2479 intel_gmch_remove();
2482 int i915_gem_gtt_init(struct drm_device *dev)
2484 struct drm_i915_private *dev_priv = dev->dev_private;
2485 struct i915_gtt *gtt = &dev_priv->gtt;
2488 if (INTEL_INFO(dev)->gen <= 5) {
2489 gtt->gtt_probe = i915_gmch_probe;
2490 gtt->base.cleanup = i915_gmch_remove;
2491 } else if (INTEL_INFO(dev)->gen < 8) {
2492 gtt->gtt_probe = gen6_gmch_probe;
2493 gtt->base.cleanup = gen6_gmch_remove;
2494 if (IS_HASWELL(dev) && dev_priv->ellc_size)
2495 gtt->base.pte_encode = iris_pte_encode;
2496 else if (IS_HASWELL(dev))
2497 gtt->base.pte_encode = hsw_pte_encode;
2498 else if (IS_VALLEYVIEW(dev))
2499 gtt->base.pte_encode = byt_pte_encode;
2500 else if (INTEL_INFO(dev)->gen >= 7)
2501 gtt->base.pte_encode = ivb_pte_encode;
2503 gtt->base.pte_encode = snb_pte_encode;
2505 dev_priv->gtt.gtt_probe = gen8_gmch_probe;
2506 dev_priv->gtt.base.cleanup = gen6_gmch_remove;
2509 ret = gtt->gtt_probe(dev, >t->base.total, >t->stolen_size,
2510 >t->mappable_base, >t->mappable_end);
2514 gtt->base.dev = dev;
2516 /* GMADR is the PCI mmio aperture into the global GTT. */
2517 DRM_INFO("Memory usable by graphics device = %lluM\n",
2518 gtt->base.total >> 20);
2519 DRM_DEBUG_DRIVER("GMADR size = %lldM\n", gtt->mappable_end >> 20);
2520 DRM_DEBUG_DRIVER("GTT stolen size = %zdM\n", gtt->stolen_size >> 20);
2521 #ifdef CONFIG_INTEL_IOMMU
2522 if (intel_iommu_gfx_mapped)
2523 DRM_INFO("VT-d active for gfx access\n");
2526 * i915.enable_ppgtt is read-only, so do an early pass to validate the
2527 * user's requested state against the hardware/driver capabilities. We
2528 * do this now so that we can print out any log messages once rather
2529 * than every time we check intel_enable_ppgtt().
2531 i915.enable_ppgtt = sanitize_enable_ppgtt(dev, i915.enable_ppgtt);
2532 DRM_DEBUG_DRIVER("ppgtt mode: %i\n", i915.enable_ppgtt);
2537 void i915_gem_restore_gtt_mappings(struct drm_device *dev)
2539 struct drm_i915_private *dev_priv = dev->dev_private;
2540 struct drm_i915_gem_object *obj;
2541 struct i915_address_space *vm;
2543 i915_check_and_clear_faults(dev);
2545 /* First fill our portion of the GTT with scratch pages */
2546 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
2547 dev_priv->gtt.base.start,
2548 dev_priv->gtt.base.total,
2551 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
2552 struct i915_vma *vma = i915_gem_obj_to_vma(obj,
2553 &dev_priv->gtt.base);
2557 i915_gem_clflush_object(obj, obj->pin_display);
2558 WARN_ON(i915_vma_bind(vma, obj->cache_level, PIN_UPDATE));
2562 if (INTEL_INFO(dev)->gen >= 8) {
2563 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
2564 chv_setup_private_ppat(dev_priv);
2566 bdw_setup_private_ppat(dev_priv);
2571 if (USES_PPGTT(dev)) {
2572 list_for_each_entry(vm, &dev_priv->vm_list, global_link) {
2573 /* TODO: Perhaps it shouldn't be gen6 specific */
2575 struct i915_hw_ppgtt *ppgtt =
2576 container_of(vm, struct i915_hw_ppgtt,
2579 if (i915_is_ggtt(vm))
2580 ppgtt = dev_priv->mm.aliasing_ppgtt;
2582 gen6_write_page_range(dev_priv, &ppgtt->pd,
2583 0, ppgtt->base.total);
2587 i915_ggtt_flush(dev_priv);
2590 static struct i915_vma *
2591 __i915_gem_vma_create(struct drm_i915_gem_object *obj,
2592 struct i915_address_space *vm,
2593 const struct i915_ggtt_view *ggtt_view)
2595 struct i915_vma *vma;
2597 if (WARN_ON(i915_is_ggtt(vm) != !!ggtt_view))
2598 return ERR_PTR(-EINVAL);
2600 vma = kmem_cache_zalloc(to_i915(obj->base.dev)->vmas, GFP_KERNEL);
2602 return ERR_PTR(-ENOMEM);
2604 INIT_LIST_HEAD(&vma->vma_link);
2605 INIT_LIST_HEAD(&vma->mm_list);
2606 INIT_LIST_HEAD(&vma->exec_list);
2610 if (i915_is_ggtt(vm))
2611 vma->ggtt_view = *ggtt_view;
2613 list_add_tail(&vma->vma_link, &obj->vma_list);
2614 if (!i915_is_ggtt(vm))
2615 i915_ppgtt_get(i915_vm_to_ppgtt(vm));
2621 i915_gem_obj_lookup_or_create_vma(struct drm_i915_gem_object *obj,
2622 struct i915_address_space *vm)
2624 struct i915_vma *vma;
2626 vma = i915_gem_obj_to_vma(obj, vm);
2628 vma = __i915_gem_vma_create(obj, vm,
2629 i915_is_ggtt(vm) ? &i915_ggtt_view_normal : NULL);
2635 i915_gem_obj_lookup_or_create_ggtt_vma(struct drm_i915_gem_object *obj,
2636 const struct i915_ggtt_view *view)
2638 struct i915_address_space *ggtt = i915_obj_to_ggtt(obj);
2639 struct i915_vma *vma;
2642 return ERR_PTR(-EINVAL);
2644 vma = i915_gem_obj_to_ggtt_view(obj, view);
2650 vma = __i915_gem_vma_create(obj, ggtt, view);
2657 rotate_pages(dma_addr_t *in, unsigned int width, unsigned int height,
2658 struct sg_table *st)
2660 unsigned int column, row;
2661 unsigned int src_idx;
2662 struct scatterlist *sg = st->sgl;
2666 for (column = 0; column < width; column++) {
2667 src_idx = width * (height - 1) + column;
2668 for (row = 0; row < height; row++) {
2670 /* We don't need the pages, but need to initialize
2671 * the entries so the sg list can be happily traversed.
2672 * The only thing we need are DMA addresses.
2674 sg_set_page(sg, NULL, PAGE_SIZE, 0);
2675 sg_dma_address(sg) = in[src_idx];
2676 sg_dma_len(sg) = PAGE_SIZE;
2683 static struct sg_table *
2684 intel_rotate_fb_obj_pages(struct i915_ggtt_view *ggtt_view,
2685 struct drm_i915_gem_object *obj)
2687 struct intel_rotation_info *rot_info = &ggtt_view->rotation_info;
2688 unsigned int size_pages = rot_info->size >> PAGE_SHIFT;
2689 struct sg_page_iter sg_iter;
2691 dma_addr_t *page_addr_list;
2692 struct sg_table *st;
2695 /* Allocate a temporary list of source pages for random access. */
2696 page_addr_list = drm_malloc_ab(obj->base.size / PAGE_SIZE,
2697 sizeof(dma_addr_t));
2698 if (!page_addr_list)
2699 return ERR_PTR(ret);
2701 /* Allocate target SG list. */
2702 st = kmalloc(sizeof(*st), GFP_KERNEL);
2706 ret = sg_alloc_table(st, size_pages, GFP_KERNEL);
2710 /* Populate source page list from the object. */
2712 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) {
2713 page_addr_list[i] = sg_page_iter_dma_address(&sg_iter);
2717 /* Rotate the pages. */
2718 rotate_pages(page_addr_list,
2719 rot_info->width_pages, rot_info->height_pages,
2723 "Created rotated page mapping for object size %zu (pitch=%u, height=%u, pixel_format=0x%x, %ux%u tiles, %u pages).\n",
2724 obj->base.size, rot_info->pitch, rot_info->height,
2725 rot_info->pixel_format, rot_info->width_pages,
2726 rot_info->height_pages, size_pages);
2728 drm_free_large(page_addr_list);
2735 drm_free_large(page_addr_list);
2738 "Failed to create rotated mapping for object size %zu! (%d) (pitch=%u, height=%u, pixel_format=0x%x, %ux%u tiles, %u pages)\n",
2739 obj->base.size, ret, rot_info->pitch, rot_info->height,
2740 rot_info->pixel_format, rot_info->width_pages,
2741 rot_info->height_pages, size_pages);
2742 return ERR_PTR(ret);
2745 static struct sg_table *
2746 intel_partial_pages(const struct i915_ggtt_view *view,
2747 struct drm_i915_gem_object *obj)
2749 struct sg_table *st;
2750 struct scatterlist *sg;
2751 struct sg_page_iter obj_sg_iter;
2754 st = kmalloc(sizeof(*st), GFP_KERNEL);
2758 ret = sg_alloc_table(st, view->params.partial.size, GFP_KERNEL);
2764 for_each_sg_page(obj->pages->sgl, &obj_sg_iter, obj->pages->nents,
2765 view->params.partial.offset)
2767 if (st->nents >= view->params.partial.size)
2770 sg_set_page(sg, NULL, PAGE_SIZE, 0);
2771 sg_dma_address(sg) = sg_page_iter_dma_address(&obj_sg_iter);
2772 sg_dma_len(sg) = PAGE_SIZE;
2783 return ERR_PTR(ret);
2787 i915_get_ggtt_vma_pages(struct i915_vma *vma)
2791 if (vma->ggtt_view.pages)
2794 if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL)
2795 vma->ggtt_view.pages = vma->obj->pages;
2796 else if (vma->ggtt_view.type == I915_GGTT_VIEW_ROTATED)
2797 vma->ggtt_view.pages =
2798 intel_rotate_fb_obj_pages(&vma->ggtt_view, vma->obj);
2799 else if (vma->ggtt_view.type == I915_GGTT_VIEW_PARTIAL)
2800 vma->ggtt_view.pages =
2801 intel_partial_pages(&vma->ggtt_view, vma->obj);
2803 WARN_ONCE(1, "GGTT view %u not implemented!\n",
2804 vma->ggtt_view.type);
2806 if (!vma->ggtt_view.pages) {
2807 DRM_ERROR("Failed to get pages for GGTT view type %u!\n",
2808 vma->ggtt_view.type);
2810 } else if (IS_ERR(vma->ggtt_view.pages)) {
2811 ret = PTR_ERR(vma->ggtt_view.pages);
2812 vma->ggtt_view.pages = NULL;
2813 DRM_ERROR("Failed to get pages for VMA view type %u (%d)!\n",
2814 vma->ggtt_view.type, ret);
2821 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
2823 * @cache_level: mapping cache level
2824 * @flags: flags like global or local mapping
2826 * DMA addresses are taken from the scatter-gather table of this object (or of
2827 * this VMA in case of non-default GGTT views) and PTE entries set up.
2828 * Note that DMA addresses are also the only part of the SG table we care about.
2830 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
2836 if (WARN_ON(flags == 0))
2840 if (flags & PIN_GLOBAL)
2841 bind_flags |= GLOBAL_BIND;
2842 if (flags & PIN_USER)
2843 bind_flags |= LOCAL_BIND;
2845 if (flags & PIN_UPDATE)
2846 bind_flags |= vma->bound;
2848 bind_flags &= ~vma->bound;
2850 if (bind_flags == 0)
2853 if (vma->bound == 0 && vma->vm->allocate_va_range) {
2854 trace_i915_va_alloc(vma->vm,
2857 VM_TO_TRACE_NAME(vma->vm));
2859 ret = vma->vm->allocate_va_range(vma->vm,
2866 ret = vma->vm->bind_vma(vma, cache_level, bind_flags);
2870 vma->bound |= bind_flags;
2876 * i915_ggtt_view_size - Get the size of a GGTT view.
2877 * @obj: Object the view is of.
2878 * @view: The view in question.
2880 * @return The size of the GGTT view in bytes.
2883 i915_ggtt_view_size(struct drm_i915_gem_object *obj,
2884 const struct i915_ggtt_view *view)
2886 if (view->type == I915_GGTT_VIEW_NORMAL) {
2887 return obj->base.size;
2888 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
2889 return view->rotation_info.size;
2890 } else if (view->type == I915_GGTT_VIEW_PARTIAL) {
2891 return view->params.partial.size << PAGE_SHIFT;
2893 WARN_ONCE(1, "GGTT view %u not implemented!\n", view->type);
2894 return obj->base.size;