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 #define i915_dma_unmap_single(px, dev) \
305 __i915_dma_unmap_single((px)->daddr, dev)
307 static void __i915_dma_unmap_single(dma_addr_t daddr,
308 struct drm_device *dev)
310 struct device *device = &dev->pdev->dev;
312 dma_unmap_page(device, daddr, 4096, PCI_DMA_BIDIRECTIONAL);
316 * i915_dma_map_single() - Create a dma mapping for a page table/dir/etc.
317 * @px: Page table/dir/etc to get a DMA map for
320 * Page table allocations are unified across all gens. They always require a
321 * single 4k allocation, as well as a DMA mapping. If we keep the structs
322 * symmetric here, the simple macro covers us for every page table type.
324 * Return: 0 if success.
326 #define i915_dma_map_single(px, dev) \
327 i915_dma_map_page_single((px)->page, (dev), &(px)->daddr)
329 static int i915_dma_map_page_single(struct page *page,
330 struct drm_device *dev,
333 struct device *device = &dev->pdev->dev;
335 *daddr = dma_map_page(device, page, 0, 4096, PCI_DMA_BIDIRECTIONAL);
336 if (dma_mapping_error(device, *daddr))
342 static void unmap_and_free_pt(struct i915_page_table *pt,
343 struct drm_device *dev)
345 if (WARN_ON(!pt->page))
348 i915_dma_unmap_single(pt, dev);
349 __free_page(pt->page);
350 kfree(pt->used_ptes);
354 static void gen8_initialize_pt(struct i915_address_space *vm,
355 struct i915_page_table *pt)
357 gen8_pte_t *pt_vaddr, scratch_pte;
360 pt_vaddr = kmap_atomic(pt->page);
361 scratch_pte = gen8_pte_encode(vm->scratch.addr,
362 I915_CACHE_LLC, true);
364 for (i = 0; i < GEN8_PTES; i++)
365 pt_vaddr[i] = scratch_pte;
367 if (!HAS_LLC(vm->dev))
368 drm_clflush_virt_range(pt_vaddr, PAGE_SIZE);
369 kunmap_atomic(pt_vaddr);
372 static struct i915_page_table *alloc_pt_single(struct drm_device *dev)
374 struct i915_page_table *pt;
375 const size_t count = INTEL_INFO(dev)->gen >= 8 ?
376 GEN8_PTES : GEN6_PTES;
379 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
381 return ERR_PTR(-ENOMEM);
383 pt->used_ptes = kcalloc(BITS_TO_LONGS(count), sizeof(*pt->used_ptes),
389 pt->page = alloc_page(GFP_KERNEL);
393 ret = i915_dma_map_single(pt, dev);
400 __free_page(pt->page);
402 kfree(pt->used_ptes);
409 static void unmap_and_free_pd(struct i915_page_directory *pd,
410 struct drm_device *dev)
413 i915_dma_unmap_single(pd, dev);
414 __free_page(pd->page);
415 kfree(pd->used_pdes);
420 static struct i915_page_directory *alloc_pd_single(struct drm_device *dev)
422 struct i915_page_directory *pd;
425 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
427 return ERR_PTR(-ENOMEM);
429 pd->used_pdes = kcalloc(BITS_TO_LONGS(I915_PDES),
430 sizeof(*pd->used_pdes), GFP_KERNEL);
434 pd->page = alloc_page(GFP_KERNEL);
438 ret = i915_dma_map_single(pd, dev);
445 __free_page(pd->page);
447 kfree(pd->used_pdes);
454 /* Broadwell Page Directory Pointer Descriptors */
455 static int gen8_write_pdp(struct intel_engine_cs *ring,
463 ret = intel_ring_begin(ring, 6);
467 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
468 intel_ring_emit(ring, GEN8_RING_PDP_UDW(ring, entry));
469 intel_ring_emit(ring, upper_32_bits(addr));
470 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
471 intel_ring_emit(ring, GEN8_RING_PDP_LDW(ring, entry));
472 intel_ring_emit(ring, lower_32_bits(addr));
473 intel_ring_advance(ring);
478 static int gen8_mm_switch(struct i915_hw_ppgtt *ppgtt,
479 struct intel_engine_cs *ring)
483 for (i = GEN8_LEGACY_PDPES - 1; i >= 0; i--) {
484 struct i915_page_directory *pd = ppgtt->pdp.page_directory[i];
485 dma_addr_t pd_daddr = pd ? pd->daddr : ppgtt->scratch_pd->daddr;
486 /* The page directory might be NULL, but we need to clear out
487 * whatever the previous context might have used. */
488 ret = gen8_write_pdp(ring, i, pd_daddr);
496 static void gen8_ppgtt_clear_range(struct i915_address_space *vm,
501 struct i915_hw_ppgtt *ppgtt =
502 container_of(vm, struct i915_hw_ppgtt, base);
503 gen8_pte_t *pt_vaddr, scratch_pte;
504 unsigned pdpe = start >> GEN8_PDPE_SHIFT & GEN8_PDPE_MASK;
505 unsigned pde = start >> GEN8_PDE_SHIFT & GEN8_PDE_MASK;
506 unsigned pte = start >> GEN8_PTE_SHIFT & GEN8_PTE_MASK;
507 unsigned num_entries = length >> PAGE_SHIFT;
508 unsigned last_pte, i;
510 scratch_pte = gen8_pte_encode(ppgtt->base.scratch.addr,
511 I915_CACHE_LLC, use_scratch);
513 while (num_entries) {
514 struct i915_page_directory *pd;
515 struct i915_page_table *pt;
516 struct page *page_table;
518 if (WARN_ON(!ppgtt->pdp.page_directory[pdpe]))
521 pd = ppgtt->pdp.page_directory[pdpe];
523 if (WARN_ON(!pd->page_table[pde]))
526 pt = pd->page_table[pde];
528 if (WARN_ON(!pt->page))
531 page_table = pt->page;
533 last_pte = pte + num_entries;
534 if (last_pte > GEN8_PTES)
535 last_pte = GEN8_PTES;
537 pt_vaddr = kmap_atomic(page_table);
539 for (i = pte; i < last_pte; i++) {
540 pt_vaddr[i] = scratch_pte;
544 if (!HAS_LLC(ppgtt->base.dev))
545 drm_clflush_virt_range(pt_vaddr, PAGE_SIZE);
546 kunmap_atomic(pt_vaddr);
549 if (++pde == I915_PDES) {
556 static void gen8_ppgtt_insert_entries(struct i915_address_space *vm,
557 struct sg_table *pages,
559 enum i915_cache_level cache_level, u32 unused)
561 struct i915_hw_ppgtt *ppgtt =
562 container_of(vm, struct i915_hw_ppgtt, base);
563 gen8_pte_t *pt_vaddr;
564 unsigned pdpe = start >> GEN8_PDPE_SHIFT & GEN8_PDPE_MASK;
565 unsigned pde = start >> GEN8_PDE_SHIFT & GEN8_PDE_MASK;
566 unsigned pte = start >> GEN8_PTE_SHIFT & GEN8_PTE_MASK;
567 struct sg_page_iter sg_iter;
571 for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) {
572 if (WARN_ON(pdpe >= GEN8_LEGACY_PDPES))
575 if (pt_vaddr == NULL) {
576 struct i915_page_directory *pd = ppgtt->pdp.page_directory[pdpe];
577 struct i915_page_table *pt = pd->page_table[pde];
578 struct page *page_table = pt->page;
580 pt_vaddr = kmap_atomic(page_table);
584 gen8_pte_encode(sg_page_iter_dma_address(&sg_iter),
586 if (++pte == GEN8_PTES) {
587 if (!HAS_LLC(ppgtt->base.dev))
588 drm_clflush_virt_range(pt_vaddr, PAGE_SIZE);
589 kunmap_atomic(pt_vaddr);
591 if (++pde == I915_PDES) {
599 if (!HAS_LLC(ppgtt->base.dev))
600 drm_clflush_virt_range(pt_vaddr, PAGE_SIZE);
601 kunmap_atomic(pt_vaddr);
605 static void __gen8_do_map_pt(gen8_pde_t * const pde,
606 struct i915_page_table *pt,
607 struct drm_device *dev)
610 gen8_pde_encode(dev, pt->daddr, I915_CACHE_LLC);
614 static void gen8_initialize_pd(struct i915_address_space *vm,
615 struct i915_page_directory *pd)
617 struct i915_hw_ppgtt *ppgtt =
618 container_of(vm, struct i915_hw_ppgtt, base);
619 gen8_pde_t *page_directory;
620 struct i915_page_table *pt;
623 page_directory = kmap_atomic(pd->page);
624 pt = ppgtt->scratch_pt;
625 for (i = 0; i < I915_PDES; i++)
626 /* Map the PDE to the page table */
627 __gen8_do_map_pt(page_directory + i, pt, vm->dev);
629 if (!HAS_LLC(vm->dev))
630 drm_clflush_virt_range(page_directory, PAGE_SIZE);
631 kunmap_atomic(page_directory);
634 static void gen8_free_page_tables(struct i915_page_directory *pd, struct drm_device *dev)
641 for_each_set_bit(i, pd->used_pdes, I915_PDES) {
642 if (WARN_ON(!pd->page_table[i]))
645 unmap_and_free_pt(pd->page_table[i], dev);
646 pd->page_table[i] = NULL;
650 static void gen8_ppgtt_cleanup(struct i915_address_space *vm)
652 struct i915_hw_ppgtt *ppgtt =
653 container_of(vm, struct i915_hw_ppgtt, base);
656 for_each_set_bit(i, ppgtt->pdp.used_pdpes, GEN8_LEGACY_PDPES) {
657 if (WARN_ON(!ppgtt->pdp.page_directory[i]))
660 gen8_free_page_tables(ppgtt->pdp.page_directory[i], ppgtt->base.dev);
661 unmap_and_free_pd(ppgtt->pdp.page_directory[i], ppgtt->base.dev);
664 unmap_and_free_pd(ppgtt->scratch_pd, ppgtt->base.dev);
665 unmap_and_free_pt(ppgtt->scratch_pt, ppgtt->base.dev);
669 * gen8_ppgtt_alloc_pagetabs() - Allocate page tables for VA range.
670 * @ppgtt: Master ppgtt structure.
671 * @pd: Page directory for this address range.
672 * @start: Starting virtual address to begin allocations.
673 * @length Size of the allocations.
674 * @new_pts: Bitmap set by function with new allocations. Likely used by the
675 * caller to free on error.
677 * Allocate the required number of page tables. Extremely similar to
678 * gen8_ppgtt_alloc_page_directories(). The main difference is here we are limited by
679 * the page directory boundary (instead of the page directory pointer). That
680 * boundary is 1GB virtual. Therefore, unlike gen8_ppgtt_alloc_page_directories(), it is
681 * possible, and likely that the caller will need to use multiple calls of this
682 * function to achieve the appropriate allocation.
684 * Return: 0 if success; negative error code otherwise.
686 static int gen8_ppgtt_alloc_pagetabs(struct i915_hw_ppgtt *ppgtt,
687 struct i915_page_directory *pd,
690 unsigned long *new_pts)
692 struct drm_device *dev = ppgtt->base.dev;
693 struct i915_page_table *pt;
697 gen8_for_each_pde(pt, pd, start, length, temp, pde) {
698 /* Don't reallocate page tables */
700 /* Scratch is never allocated this way */
701 WARN_ON(pt == ppgtt->scratch_pt);
705 pt = alloc_pt_single(dev);
709 gen8_initialize_pt(&ppgtt->base, pt);
710 pd->page_table[pde] = pt;
711 set_bit(pde, new_pts);
717 for_each_set_bit(pde, new_pts, I915_PDES)
718 unmap_and_free_pt(pd->page_table[pde], dev);
724 * gen8_ppgtt_alloc_page_directories() - Allocate page directories for VA range.
725 * @ppgtt: Master ppgtt structure.
726 * @pdp: Page directory pointer for this address range.
727 * @start: Starting virtual address to begin allocations.
728 * @length Size of the allocations.
729 * @new_pds Bitmap set by function with new allocations. Likely used by the
730 * caller to free on error.
732 * Allocate the required number of page directories starting at the pde index of
733 * @start, and ending at the pde index @start + @length. This function will skip
734 * over already allocated page directories within the range, and only allocate
735 * new ones, setting the appropriate pointer within the pdp as well as the
736 * correct position in the bitmap @new_pds.
738 * The function will only allocate the pages within the range for a give page
739 * directory pointer. In other words, if @start + @length straddles a virtually
740 * addressed PDP boundary (512GB for 4k pages), there will be more allocations
741 * required by the caller, This is not currently possible, and the BUG in the
742 * code will prevent it.
744 * Return: 0 if success; negative error code otherwise.
746 static int gen8_ppgtt_alloc_page_directories(struct i915_hw_ppgtt *ppgtt,
747 struct i915_page_directory_pointer *pdp,
750 unsigned long *new_pds)
752 struct drm_device *dev = ppgtt->base.dev;
753 struct i915_page_directory *pd;
757 WARN_ON(!bitmap_empty(new_pds, GEN8_LEGACY_PDPES));
759 /* FIXME: upper bound must not overflow 32 bits */
760 WARN_ON((start + length) >= (1ULL << 32));
762 gen8_for_each_pdpe(pd, pdp, start, length, temp, pdpe) {
766 pd = alloc_pd_single(dev);
770 gen8_initialize_pd(&ppgtt->base, pd);
771 pdp->page_directory[pdpe] = pd;
772 set_bit(pdpe, new_pds);
778 for_each_set_bit(pdpe, new_pds, GEN8_LEGACY_PDPES)
779 unmap_and_free_pd(pdp->page_directory[pdpe], dev);
785 free_gen8_temp_bitmaps(unsigned long *new_pds, unsigned long **new_pts)
789 for (i = 0; i < GEN8_LEGACY_PDPES; i++)
795 /* Fills in the page directory bitmap, and the array of page tables bitmap. Both
796 * of these are based on the number of PDPEs in the system.
799 int __must_check alloc_gen8_temp_bitmaps(unsigned long **new_pds,
800 unsigned long ***new_pts)
806 pds = kcalloc(BITS_TO_LONGS(GEN8_LEGACY_PDPES), sizeof(unsigned long), GFP_KERNEL);
810 pts = kcalloc(GEN8_LEGACY_PDPES, sizeof(unsigned long *), GFP_KERNEL);
816 for (i = 0; i < GEN8_LEGACY_PDPES; i++) {
817 pts[i] = kcalloc(BITS_TO_LONGS(I915_PDES),
818 sizeof(unsigned long), GFP_KERNEL);
829 free_gen8_temp_bitmaps(pds, pts);
833 static int gen8_alloc_va_range(struct i915_address_space *vm,
837 struct i915_hw_ppgtt *ppgtt =
838 container_of(vm, struct i915_hw_ppgtt, base);
839 unsigned long *new_page_dirs, **new_page_tables;
840 struct i915_page_directory *pd;
841 const uint64_t orig_start = start;
842 const uint64_t orig_length = length;
847 /* Wrap is never okay since we can only represent 48b, and we don't
848 * actually use the other side of the canonical address space.
850 if (WARN_ON(start + length < start))
853 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables);
857 /* Do the allocations first so we can easily bail out */
858 ret = gen8_ppgtt_alloc_page_directories(ppgtt, &ppgtt->pdp, start, length,
861 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
865 /* For every page directory referenced, allocate page tables */
866 gen8_for_each_pdpe(pd, &ppgtt->pdp, start, length, temp, pdpe) {
867 ret = gen8_ppgtt_alloc_pagetabs(ppgtt, pd, start, length,
868 new_page_tables[pdpe]);
874 length = orig_length;
876 /* Allocations have completed successfully, so set the bitmaps, and do
878 gen8_for_each_pdpe(pd, &ppgtt->pdp, start, length, temp, pdpe) {
879 gen8_pde_t *const page_directory = kmap_atomic(pd->page);
880 struct i915_page_table *pt;
881 uint64_t pd_len = gen8_clamp_pd(start, length);
882 uint64_t pd_start = start;
885 /* Every pd should be allocated, we just did that above. */
888 gen8_for_each_pde(pt, pd, pd_start, pd_len, temp, pde) {
889 /* Same reasoning as pd */
892 WARN_ON(!gen8_pte_count(pd_start, pd_len));
894 /* Set our used ptes within the page table */
895 bitmap_set(pt->used_ptes,
896 gen8_pte_index(pd_start),
897 gen8_pte_count(pd_start, pd_len));
899 /* Our pde is now pointing to the pagetable, pt */
900 set_bit(pde, pd->used_pdes);
902 /* Map the PDE to the page table */
903 __gen8_do_map_pt(page_directory + pde, pt, vm->dev);
905 /* NB: We haven't yet mapped ptes to pages. At this
906 * point we're still relying on insert_entries() */
909 if (!HAS_LLC(vm->dev))
910 drm_clflush_virt_range(page_directory, PAGE_SIZE);
912 kunmap_atomic(page_directory);
914 set_bit(pdpe, ppgtt->pdp.used_pdpes);
917 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
922 for_each_set_bit(temp, new_page_tables[pdpe], I915_PDES)
923 unmap_and_free_pt(ppgtt->pdp.page_directory[pdpe]->page_table[temp], vm->dev);
926 for_each_set_bit(pdpe, new_page_dirs, GEN8_LEGACY_PDPES)
927 unmap_and_free_pd(ppgtt->pdp.page_directory[pdpe], vm->dev);
929 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
934 * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers
935 * with a net effect resembling a 2-level page table in normal x86 terms. Each
936 * PDP represents 1GB of memory 4 * 512 * 512 * 4096 = 4GB legacy 32b address
940 static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
942 ppgtt->scratch_pt = alloc_pt_single(ppgtt->base.dev);
943 if (IS_ERR(ppgtt->scratch_pt))
944 return PTR_ERR(ppgtt->scratch_pt);
946 ppgtt->scratch_pd = alloc_pd_single(ppgtt->base.dev);
947 if (IS_ERR(ppgtt->scratch_pd))
948 return PTR_ERR(ppgtt->scratch_pd);
950 gen8_initialize_pt(&ppgtt->base, ppgtt->scratch_pt);
951 gen8_initialize_pd(&ppgtt->base, ppgtt->scratch_pd);
953 ppgtt->base.start = 0;
954 ppgtt->base.total = 1ULL << 32;
955 ppgtt->base.cleanup = gen8_ppgtt_cleanup;
956 ppgtt->base.allocate_va_range = gen8_alloc_va_range;
957 ppgtt->base.insert_entries = gen8_ppgtt_insert_entries;
958 ppgtt->base.clear_range = gen8_ppgtt_clear_range;
959 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
960 ppgtt->base.bind_vma = ppgtt_bind_vma;
962 ppgtt->switch_mm = gen8_mm_switch;
967 static void gen6_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
969 struct i915_address_space *vm = &ppgtt->base;
970 struct i915_page_table *unused;
971 gen6_pte_t scratch_pte;
973 uint32_t pte, pde, temp;
974 uint32_t start = ppgtt->base.start, length = ppgtt->base.total;
976 scratch_pte = vm->pte_encode(vm->scratch.addr, I915_CACHE_LLC, true, 0);
978 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde) {
980 gen6_pte_t *pt_vaddr;
981 dma_addr_t pt_addr = ppgtt->pd.page_table[pde]->daddr;
982 pd_entry = readl(ppgtt->pd_addr + pde);
983 expected = (GEN6_PDE_ADDR_ENCODE(pt_addr) | GEN6_PDE_VALID);
985 if (pd_entry != expected)
986 seq_printf(m, "\tPDE #%d mismatch: Actual PDE: %x Expected PDE: %x\n",
990 seq_printf(m, "\tPDE: %x\n", pd_entry);
992 pt_vaddr = kmap_atomic(ppgtt->pd.page_table[pde]->page);
993 for (pte = 0; pte < GEN6_PTES; pte+=4) {
995 (pde * PAGE_SIZE * GEN6_PTES) +
999 for (i = 0; i < 4; i++)
1000 if (pt_vaddr[pte + i] != scratch_pte)
1005 seq_printf(m, "\t\t0x%lx [%03d,%04d]: =", va, pde, pte);
1006 for (i = 0; i < 4; i++) {
1007 if (pt_vaddr[pte + i] != scratch_pte)
1008 seq_printf(m, " %08x", pt_vaddr[pte + i]);
1010 seq_puts(m, " SCRATCH ");
1014 kunmap_atomic(pt_vaddr);
1018 /* Write pde (index) from the page directory @pd to the page table @pt */
1019 static void gen6_write_pde(struct i915_page_directory *pd,
1020 const int pde, struct i915_page_table *pt)
1022 /* Caller needs to make sure the write completes if necessary */
1023 struct i915_hw_ppgtt *ppgtt =
1024 container_of(pd, struct i915_hw_ppgtt, pd);
1027 pd_entry = GEN6_PDE_ADDR_ENCODE(pt->daddr);
1028 pd_entry |= GEN6_PDE_VALID;
1030 writel(pd_entry, ppgtt->pd_addr + pde);
1033 /* Write all the page tables found in the ppgtt structure to incrementing page
1035 static void gen6_write_page_range(struct drm_i915_private *dev_priv,
1036 struct i915_page_directory *pd,
1037 uint32_t start, uint32_t length)
1039 struct i915_page_table *pt;
1042 gen6_for_each_pde(pt, pd, start, length, temp, pde)
1043 gen6_write_pde(pd, pde, pt);
1045 /* Make sure write is complete before other code can use this page
1046 * table. Also require for WC mapped PTEs */
1047 readl(dev_priv->gtt.gsm);
1050 static uint32_t get_pd_offset(struct i915_hw_ppgtt *ppgtt)
1052 BUG_ON(ppgtt->pd.pd_offset & 0x3f);
1054 return (ppgtt->pd.pd_offset / 64) << 16;
1057 static int hsw_mm_switch(struct i915_hw_ppgtt *ppgtt,
1058 struct intel_engine_cs *ring)
1062 /* NB: TLBs must be flushed and invalidated before a switch */
1063 ret = ring->flush(ring, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1067 ret = intel_ring_begin(ring, 6);
1071 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1072 intel_ring_emit(ring, RING_PP_DIR_DCLV(ring));
1073 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1074 intel_ring_emit(ring, RING_PP_DIR_BASE(ring));
1075 intel_ring_emit(ring, get_pd_offset(ppgtt));
1076 intel_ring_emit(ring, MI_NOOP);
1077 intel_ring_advance(ring);
1082 static int vgpu_mm_switch(struct i915_hw_ppgtt *ppgtt,
1083 struct intel_engine_cs *ring)
1085 struct drm_i915_private *dev_priv = to_i915(ppgtt->base.dev);
1087 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
1088 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
1092 static int gen7_mm_switch(struct i915_hw_ppgtt *ppgtt,
1093 struct intel_engine_cs *ring)
1097 /* NB: TLBs must be flushed and invalidated before a switch */
1098 ret = ring->flush(ring, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1102 ret = intel_ring_begin(ring, 6);
1106 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1107 intel_ring_emit(ring, RING_PP_DIR_DCLV(ring));
1108 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1109 intel_ring_emit(ring, RING_PP_DIR_BASE(ring));
1110 intel_ring_emit(ring, get_pd_offset(ppgtt));
1111 intel_ring_emit(ring, MI_NOOP);
1112 intel_ring_advance(ring);
1114 /* XXX: RCS is the only one to auto invalidate the TLBs? */
1115 if (ring->id != RCS) {
1116 ret = ring->flush(ring, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1124 static int gen6_mm_switch(struct i915_hw_ppgtt *ppgtt,
1125 struct intel_engine_cs *ring)
1127 struct drm_device *dev = ppgtt->base.dev;
1128 struct drm_i915_private *dev_priv = dev->dev_private;
1131 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
1132 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
1134 POSTING_READ(RING_PP_DIR_DCLV(ring));
1139 static void gen8_ppgtt_enable(struct drm_device *dev)
1141 struct drm_i915_private *dev_priv = dev->dev_private;
1142 struct intel_engine_cs *ring;
1145 for_each_ring(ring, dev_priv, j) {
1146 I915_WRITE(RING_MODE_GEN7(ring),
1147 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1151 static void gen7_ppgtt_enable(struct drm_device *dev)
1153 struct drm_i915_private *dev_priv = dev->dev_private;
1154 struct intel_engine_cs *ring;
1155 uint32_t ecochk, ecobits;
1158 ecobits = I915_READ(GAC_ECO_BITS);
1159 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B);
1161 ecochk = I915_READ(GAM_ECOCHK);
1162 if (IS_HASWELL(dev)) {
1163 ecochk |= ECOCHK_PPGTT_WB_HSW;
1165 ecochk |= ECOCHK_PPGTT_LLC_IVB;
1166 ecochk &= ~ECOCHK_PPGTT_GFDT_IVB;
1168 I915_WRITE(GAM_ECOCHK, ecochk);
1170 for_each_ring(ring, dev_priv, i) {
1171 /* GFX_MODE is per-ring on gen7+ */
1172 I915_WRITE(RING_MODE_GEN7(ring),
1173 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1177 static void gen6_ppgtt_enable(struct drm_device *dev)
1179 struct drm_i915_private *dev_priv = dev->dev_private;
1180 uint32_t ecochk, gab_ctl, ecobits;
1182 ecobits = I915_READ(GAC_ECO_BITS);
1183 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_SNB_BIT |
1184 ECOBITS_PPGTT_CACHE64B);
1186 gab_ctl = I915_READ(GAB_CTL);
1187 I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT);
1189 ecochk = I915_READ(GAM_ECOCHK);
1190 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT | ECOCHK_PPGTT_CACHE64B);
1192 I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1195 /* PPGTT support for Sandybdrige/Gen6 and later */
1196 static void gen6_ppgtt_clear_range(struct i915_address_space *vm,
1201 struct i915_hw_ppgtt *ppgtt =
1202 container_of(vm, struct i915_hw_ppgtt, base);
1203 gen6_pte_t *pt_vaddr, scratch_pte;
1204 unsigned first_entry = start >> PAGE_SHIFT;
1205 unsigned num_entries = length >> PAGE_SHIFT;
1206 unsigned act_pt = first_entry / GEN6_PTES;
1207 unsigned first_pte = first_entry % GEN6_PTES;
1208 unsigned last_pte, i;
1210 scratch_pte = vm->pte_encode(vm->scratch.addr, I915_CACHE_LLC, true, 0);
1212 while (num_entries) {
1213 last_pte = first_pte + num_entries;
1214 if (last_pte > GEN6_PTES)
1215 last_pte = GEN6_PTES;
1217 pt_vaddr = kmap_atomic(ppgtt->pd.page_table[act_pt]->page);
1219 for (i = first_pte; i < last_pte; i++)
1220 pt_vaddr[i] = scratch_pte;
1222 kunmap_atomic(pt_vaddr);
1224 num_entries -= last_pte - first_pte;
1230 static void gen6_ppgtt_insert_entries(struct i915_address_space *vm,
1231 struct sg_table *pages,
1233 enum i915_cache_level cache_level, u32 flags)
1235 struct i915_hw_ppgtt *ppgtt =
1236 container_of(vm, struct i915_hw_ppgtt, base);
1237 gen6_pte_t *pt_vaddr;
1238 unsigned first_entry = start >> PAGE_SHIFT;
1239 unsigned act_pt = first_entry / GEN6_PTES;
1240 unsigned act_pte = first_entry % GEN6_PTES;
1241 struct sg_page_iter sg_iter;
1244 for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) {
1245 if (pt_vaddr == NULL)
1246 pt_vaddr = kmap_atomic(ppgtt->pd.page_table[act_pt]->page);
1249 vm->pte_encode(sg_page_iter_dma_address(&sg_iter),
1250 cache_level, true, flags);
1252 if (++act_pte == GEN6_PTES) {
1253 kunmap_atomic(pt_vaddr);
1260 kunmap_atomic(pt_vaddr);
1263 /* PDE TLBs are a pain invalidate pre GEN8. It requires a context reload. If we
1264 * are switching between contexts with the same LRCA, we also must do a force
1267 static void mark_tlbs_dirty(struct i915_hw_ppgtt *ppgtt)
1269 /* If current vm != vm, */
1270 ppgtt->pd_dirty_rings = INTEL_INFO(ppgtt->base.dev)->ring_mask;
1273 static void gen6_initialize_pt(struct i915_address_space *vm,
1274 struct i915_page_table *pt)
1276 gen6_pte_t *pt_vaddr, scratch_pte;
1279 WARN_ON(vm->scratch.addr == 0);
1281 scratch_pte = vm->pte_encode(vm->scratch.addr,
1282 I915_CACHE_LLC, true, 0);
1284 pt_vaddr = kmap_atomic(pt->page);
1286 for (i = 0; i < GEN6_PTES; i++)
1287 pt_vaddr[i] = scratch_pte;
1289 kunmap_atomic(pt_vaddr);
1292 static int gen6_alloc_va_range(struct i915_address_space *vm,
1293 uint64_t start, uint64_t length)
1295 DECLARE_BITMAP(new_page_tables, I915_PDES);
1296 struct drm_device *dev = vm->dev;
1297 struct drm_i915_private *dev_priv = dev->dev_private;
1298 struct i915_hw_ppgtt *ppgtt =
1299 container_of(vm, struct i915_hw_ppgtt, base);
1300 struct i915_page_table *pt;
1301 const uint32_t start_save = start, length_save = length;
1305 WARN_ON(upper_32_bits(start));
1307 bitmap_zero(new_page_tables, I915_PDES);
1309 /* The allocation is done in two stages so that we can bail out with
1310 * minimal amount of pain. The first stage finds new page tables that
1311 * need allocation. The second stage marks use ptes within the page
1314 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1315 if (pt != ppgtt->scratch_pt) {
1316 WARN_ON(bitmap_empty(pt->used_ptes, GEN6_PTES));
1320 /* We've already allocated a page table */
1321 WARN_ON(!bitmap_empty(pt->used_ptes, GEN6_PTES));
1323 pt = alloc_pt_single(dev);
1329 gen6_initialize_pt(vm, pt);
1331 ppgtt->pd.page_table[pde] = pt;
1332 set_bit(pde, new_page_tables);
1333 trace_i915_page_table_entry_alloc(vm, pde, start, GEN6_PDE_SHIFT);
1337 length = length_save;
1339 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1340 DECLARE_BITMAP(tmp_bitmap, GEN6_PTES);
1342 bitmap_zero(tmp_bitmap, GEN6_PTES);
1343 bitmap_set(tmp_bitmap, gen6_pte_index(start),
1344 gen6_pte_count(start, length));
1346 if (test_and_clear_bit(pde, new_page_tables))
1347 gen6_write_pde(&ppgtt->pd, pde, pt);
1349 trace_i915_page_table_entry_map(vm, pde, pt,
1350 gen6_pte_index(start),
1351 gen6_pte_count(start, length),
1353 bitmap_or(pt->used_ptes, tmp_bitmap, pt->used_ptes,
1357 WARN_ON(!bitmap_empty(new_page_tables, I915_PDES));
1359 /* Make sure write is complete before other code can use this page
1360 * table. Also require for WC mapped PTEs */
1361 readl(dev_priv->gtt.gsm);
1363 mark_tlbs_dirty(ppgtt);
1367 for_each_set_bit(pde, new_page_tables, I915_PDES) {
1368 struct i915_page_table *pt = ppgtt->pd.page_table[pde];
1370 ppgtt->pd.page_table[pde] = ppgtt->scratch_pt;
1371 unmap_and_free_pt(pt, vm->dev);
1374 mark_tlbs_dirty(ppgtt);
1378 static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
1380 struct i915_hw_ppgtt *ppgtt =
1381 container_of(vm, struct i915_hw_ppgtt, base);
1382 struct i915_page_table *pt;
1386 drm_mm_remove_node(&ppgtt->node);
1388 gen6_for_all_pdes(pt, ppgtt, pde) {
1389 if (pt != ppgtt->scratch_pt)
1390 unmap_and_free_pt(pt, ppgtt->base.dev);
1393 unmap_and_free_pt(ppgtt->scratch_pt, ppgtt->base.dev);
1394 unmap_and_free_pd(&ppgtt->pd, ppgtt->base.dev);
1397 static int gen6_ppgtt_allocate_page_directories(struct i915_hw_ppgtt *ppgtt)
1399 struct drm_device *dev = ppgtt->base.dev;
1400 struct drm_i915_private *dev_priv = dev->dev_private;
1401 bool retried = false;
1404 /* PPGTT PDEs reside in the GGTT and consists of 512 entries. The
1405 * allocator works in address space sizes, so it's multiplied by page
1406 * size. We allocate at the top of the GTT to avoid fragmentation.
1408 BUG_ON(!drm_mm_initialized(&dev_priv->gtt.base.mm));
1409 ppgtt->scratch_pt = alloc_pt_single(ppgtt->base.dev);
1410 if (IS_ERR(ppgtt->scratch_pt))
1411 return PTR_ERR(ppgtt->scratch_pt);
1413 gen6_initialize_pt(&ppgtt->base, ppgtt->scratch_pt);
1416 ret = drm_mm_insert_node_in_range_generic(&dev_priv->gtt.base.mm,
1417 &ppgtt->node, GEN6_PD_SIZE,
1419 0, dev_priv->gtt.base.total,
1421 if (ret == -ENOSPC && !retried) {
1422 ret = i915_gem_evict_something(dev, &dev_priv->gtt.base,
1423 GEN6_PD_SIZE, GEN6_PD_ALIGN,
1425 0, dev_priv->gtt.base.total,
1438 if (ppgtt->node.start < dev_priv->gtt.mappable_end)
1439 DRM_DEBUG("Forced to use aperture for PDEs\n");
1444 unmap_and_free_pt(ppgtt->scratch_pt, ppgtt->base.dev);
1448 static int gen6_ppgtt_alloc(struct i915_hw_ppgtt *ppgtt)
1450 return gen6_ppgtt_allocate_page_directories(ppgtt);
1453 static void gen6_scratch_va_range(struct i915_hw_ppgtt *ppgtt,
1454 uint64_t start, uint64_t length)
1456 struct i915_page_table *unused;
1459 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde)
1460 ppgtt->pd.page_table[pde] = ppgtt->scratch_pt;
1463 static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
1465 struct drm_device *dev = ppgtt->base.dev;
1466 struct drm_i915_private *dev_priv = dev->dev_private;
1469 ppgtt->base.pte_encode = dev_priv->gtt.base.pte_encode;
1471 ppgtt->switch_mm = gen6_mm_switch;
1472 } else if (IS_HASWELL(dev)) {
1473 ppgtt->switch_mm = hsw_mm_switch;
1474 } else if (IS_GEN7(dev)) {
1475 ppgtt->switch_mm = gen7_mm_switch;
1479 if (intel_vgpu_active(dev))
1480 ppgtt->switch_mm = vgpu_mm_switch;
1482 ret = gen6_ppgtt_alloc(ppgtt);
1486 ppgtt->base.allocate_va_range = gen6_alloc_va_range;
1487 ppgtt->base.clear_range = gen6_ppgtt_clear_range;
1488 ppgtt->base.insert_entries = gen6_ppgtt_insert_entries;
1489 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
1490 ppgtt->base.bind_vma = ppgtt_bind_vma;
1491 ppgtt->base.cleanup = gen6_ppgtt_cleanup;
1492 ppgtt->base.start = 0;
1493 ppgtt->base.total = I915_PDES * GEN6_PTES * PAGE_SIZE;
1494 ppgtt->debug_dump = gen6_dump_ppgtt;
1496 ppgtt->pd.pd_offset =
1497 ppgtt->node.start / PAGE_SIZE * sizeof(gen6_pte_t);
1499 ppgtt->pd_addr = (gen6_pte_t __iomem *)dev_priv->gtt.gsm +
1500 ppgtt->pd.pd_offset / sizeof(gen6_pte_t);
1502 gen6_scratch_va_range(ppgtt, 0, ppgtt->base.total);
1504 gen6_write_page_range(dev_priv, &ppgtt->pd, 0, ppgtt->base.total);
1506 DRM_DEBUG_DRIVER("Allocated pde space (%lldM) at GTT entry: %llx\n",
1507 ppgtt->node.size >> 20,
1508 ppgtt->node.start / PAGE_SIZE);
1510 DRM_DEBUG("Adding PPGTT at offset %x\n",
1511 ppgtt->pd.pd_offset << 10);
1516 static int __hw_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
1518 struct drm_i915_private *dev_priv = dev->dev_private;
1520 ppgtt->base.dev = dev;
1521 ppgtt->base.scratch = dev_priv->gtt.base.scratch;
1523 if (INTEL_INFO(dev)->gen < 8)
1524 return gen6_ppgtt_init(ppgtt);
1526 return gen8_ppgtt_init(ppgtt);
1528 int i915_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
1530 struct drm_i915_private *dev_priv = dev->dev_private;
1533 ret = __hw_ppgtt_init(dev, ppgtt);
1535 kref_init(&ppgtt->ref);
1536 drm_mm_init(&ppgtt->base.mm, ppgtt->base.start,
1538 i915_init_vm(dev_priv, &ppgtt->base);
1544 int i915_ppgtt_init_hw(struct drm_device *dev)
1546 struct drm_i915_private *dev_priv = dev->dev_private;
1547 struct intel_engine_cs *ring;
1548 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
1551 /* In the case of execlists, PPGTT is enabled by the context descriptor
1552 * and the PDPs are contained within the context itself. We don't
1553 * need to do anything here. */
1554 if (i915.enable_execlists)
1557 if (!USES_PPGTT(dev))
1561 gen6_ppgtt_enable(dev);
1562 else if (IS_GEN7(dev))
1563 gen7_ppgtt_enable(dev);
1564 else if (INTEL_INFO(dev)->gen >= 8)
1565 gen8_ppgtt_enable(dev);
1567 MISSING_CASE(INTEL_INFO(dev)->gen);
1570 for_each_ring(ring, dev_priv, i) {
1571 ret = ppgtt->switch_mm(ppgtt, ring);
1579 struct i915_hw_ppgtt *
1580 i915_ppgtt_create(struct drm_device *dev, struct drm_i915_file_private *fpriv)
1582 struct i915_hw_ppgtt *ppgtt;
1585 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
1587 return ERR_PTR(-ENOMEM);
1589 ret = i915_ppgtt_init(dev, ppgtt);
1592 return ERR_PTR(ret);
1595 ppgtt->file_priv = fpriv;
1597 trace_i915_ppgtt_create(&ppgtt->base);
1602 void i915_ppgtt_release(struct kref *kref)
1604 struct i915_hw_ppgtt *ppgtt =
1605 container_of(kref, struct i915_hw_ppgtt, ref);
1607 trace_i915_ppgtt_release(&ppgtt->base);
1609 /* vmas should already be unbound */
1610 WARN_ON(!list_empty(&ppgtt->base.active_list));
1611 WARN_ON(!list_empty(&ppgtt->base.inactive_list));
1613 list_del(&ppgtt->base.global_link);
1614 drm_mm_takedown(&ppgtt->base.mm);
1616 ppgtt->base.cleanup(&ppgtt->base);
1620 extern int intel_iommu_gfx_mapped;
1621 /* Certain Gen5 chipsets require require idling the GPU before
1622 * unmapping anything from the GTT when VT-d is enabled.
1624 static bool needs_idle_maps(struct drm_device *dev)
1626 #ifdef CONFIG_INTEL_IOMMU
1627 /* Query intel_iommu to see if we need the workaround. Presumably that
1630 if (IS_GEN5(dev) && IS_MOBILE(dev) && intel_iommu_gfx_mapped)
1636 static bool do_idling(struct drm_i915_private *dev_priv)
1638 bool ret = dev_priv->mm.interruptible;
1640 if (unlikely(dev_priv->gtt.do_idle_maps)) {
1641 dev_priv->mm.interruptible = false;
1642 if (i915_gpu_idle(dev_priv->dev)) {
1643 DRM_ERROR("Couldn't idle GPU\n");
1644 /* Wait a bit, in hopes it avoids the hang */
1652 static void undo_idling(struct drm_i915_private *dev_priv, bool interruptible)
1654 if (unlikely(dev_priv->gtt.do_idle_maps))
1655 dev_priv->mm.interruptible = interruptible;
1658 void i915_check_and_clear_faults(struct drm_device *dev)
1660 struct drm_i915_private *dev_priv = dev->dev_private;
1661 struct intel_engine_cs *ring;
1664 if (INTEL_INFO(dev)->gen < 6)
1667 for_each_ring(ring, dev_priv, i) {
1669 fault_reg = I915_READ(RING_FAULT_REG(ring));
1670 if (fault_reg & RING_FAULT_VALID) {
1671 DRM_DEBUG_DRIVER("Unexpected fault\n"
1673 "\tAddress space: %s\n"
1676 fault_reg & PAGE_MASK,
1677 fault_reg & RING_FAULT_GTTSEL_MASK ? "GGTT" : "PPGTT",
1678 RING_FAULT_SRCID(fault_reg),
1679 RING_FAULT_FAULT_TYPE(fault_reg));
1680 I915_WRITE(RING_FAULT_REG(ring),
1681 fault_reg & ~RING_FAULT_VALID);
1684 POSTING_READ(RING_FAULT_REG(&dev_priv->ring[RCS]));
1687 static void i915_ggtt_flush(struct drm_i915_private *dev_priv)
1689 if (INTEL_INFO(dev_priv->dev)->gen < 6) {
1690 intel_gtt_chipset_flush();
1692 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
1693 POSTING_READ(GFX_FLSH_CNTL_GEN6);
1697 void i915_gem_suspend_gtt_mappings(struct drm_device *dev)
1699 struct drm_i915_private *dev_priv = dev->dev_private;
1701 /* Don't bother messing with faults pre GEN6 as we have little
1702 * documentation supporting that it's a good idea.
1704 if (INTEL_INFO(dev)->gen < 6)
1707 i915_check_and_clear_faults(dev);
1709 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
1710 dev_priv->gtt.base.start,
1711 dev_priv->gtt.base.total,
1714 i915_ggtt_flush(dev_priv);
1717 int i915_gem_gtt_prepare_object(struct drm_i915_gem_object *obj)
1719 if (obj->has_dma_mapping)
1722 if (!dma_map_sg(&obj->base.dev->pdev->dev,
1723 obj->pages->sgl, obj->pages->nents,
1724 PCI_DMA_BIDIRECTIONAL))
1730 static void gen8_set_pte(void __iomem *addr, gen8_pte_t pte)
1735 iowrite32((u32)pte, addr);
1736 iowrite32(pte >> 32, addr + 4);
1740 static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
1741 struct sg_table *st,
1743 enum i915_cache_level level, u32 unused)
1745 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1746 unsigned first_entry = start >> PAGE_SHIFT;
1747 gen8_pte_t __iomem *gtt_entries =
1748 (gen8_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
1750 struct sg_page_iter sg_iter;
1751 dma_addr_t addr = 0; /* shut up gcc */
1753 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
1754 addr = sg_dma_address(sg_iter.sg) +
1755 (sg_iter.sg_pgoffset << PAGE_SHIFT);
1756 gen8_set_pte(>t_entries[i],
1757 gen8_pte_encode(addr, level, true));
1762 * XXX: This serves as a posting read to make sure that the PTE has
1763 * actually been updated. There is some concern that even though
1764 * registers and PTEs are within the same BAR that they are potentially
1765 * of NUMA access patterns. Therefore, even with the way we assume
1766 * hardware should work, we must keep this posting read for paranoia.
1769 WARN_ON(readq(>t_entries[i-1])
1770 != gen8_pte_encode(addr, level, true));
1772 /* This next bit makes the above posting read even more important. We
1773 * want to flush the TLBs only after we're certain all the PTE updates
1776 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
1777 POSTING_READ(GFX_FLSH_CNTL_GEN6);
1781 * Binds an object into the global gtt with the specified cache level. The object
1782 * will be accessible to the GPU via commands whose operands reference offsets
1783 * within the global GTT as well as accessible by the GPU through the GMADR
1784 * mapped BAR (dev_priv->mm.gtt->gtt).
1786 static void gen6_ggtt_insert_entries(struct i915_address_space *vm,
1787 struct sg_table *st,
1789 enum i915_cache_level level, u32 flags)
1791 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1792 unsigned first_entry = start >> PAGE_SHIFT;
1793 gen6_pte_t __iomem *gtt_entries =
1794 (gen6_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
1796 struct sg_page_iter sg_iter;
1797 dma_addr_t addr = 0;
1799 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
1800 addr = sg_page_iter_dma_address(&sg_iter);
1801 iowrite32(vm->pte_encode(addr, level, true, flags), >t_entries[i]);
1805 /* XXX: This serves as a posting read to make sure that the PTE has
1806 * actually been updated. There is some concern that even though
1807 * registers and PTEs are within the same BAR that they are potentially
1808 * of NUMA access patterns. Therefore, even with the way we assume
1809 * hardware should work, we must keep this posting read for paranoia.
1812 unsigned long gtt = readl(>t_entries[i-1]);
1813 WARN_ON(gtt != vm->pte_encode(addr, level, true, flags));
1816 /* This next bit makes the above posting read even more important. We
1817 * want to flush the TLBs only after we're certain all the PTE updates
1820 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
1821 POSTING_READ(GFX_FLSH_CNTL_GEN6);
1824 static void gen8_ggtt_clear_range(struct i915_address_space *vm,
1829 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1830 unsigned first_entry = start >> PAGE_SHIFT;
1831 unsigned num_entries = length >> PAGE_SHIFT;
1832 gen8_pte_t scratch_pte, __iomem *gtt_base =
1833 (gen8_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
1834 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
1837 if (WARN(num_entries > max_entries,
1838 "First entry = %d; Num entries = %d (max=%d)\n",
1839 first_entry, num_entries, max_entries))
1840 num_entries = max_entries;
1842 scratch_pte = gen8_pte_encode(vm->scratch.addr,
1845 for (i = 0; i < num_entries; i++)
1846 gen8_set_pte(>t_base[i], scratch_pte);
1850 static void gen6_ggtt_clear_range(struct i915_address_space *vm,
1855 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1856 unsigned first_entry = start >> PAGE_SHIFT;
1857 unsigned num_entries = length >> PAGE_SHIFT;
1858 gen6_pte_t scratch_pte, __iomem *gtt_base =
1859 (gen6_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
1860 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
1863 if (WARN(num_entries > max_entries,
1864 "First entry = %d; Num entries = %d (max=%d)\n",
1865 first_entry, num_entries, max_entries))
1866 num_entries = max_entries;
1868 scratch_pte = vm->pte_encode(vm->scratch.addr, I915_CACHE_LLC, use_scratch, 0);
1870 for (i = 0; i < num_entries; i++)
1871 iowrite32(scratch_pte, >t_base[i]);
1875 static void i915_ggtt_insert_entries(struct i915_address_space *vm,
1876 struct sg_table *pages,
1878 enum i915_cache_level cache_level, u32 unused)
1880 unsigned int flags = (cache_level == I915_CACHE_NONE) ?
1881 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
1883 intel_gtt_insert_sg_entries(pages, start >> PAGE_SHIFT, flags);
1887 static void i915_ggtt_clear_range(struct i915_address_space *vm,
1892 unsigned first_entry = start >> PAGE_SHIFT;
1893 unsigned num_entries = length >> PAGE_SHIFT;
1894 intel_gtt_clear_range(first_entry, num_entries);
1897 static int ggtt_bind_vma(struct i915_vma *vma,
1898 enum i915_cache_level cache_level,
1901 struct drm_device *dev = vma->vm->dev;
1902 struct drm_i915_private *dev_priv = dev->dev_private;
1903 struct drm_i915_gem_object *obj = vma->obj;
1904 struct sg_table *pages = obj->pages;
1908 ret = i915_get_ggtt_vma_pages(vma);
1911 pages = vma->ggtt_view.pages;
1913 /* Currently applicable only to VLV */
1915 pte_flags |= PTE_READ_ONLY;
1918 if (!dev_priv->mm.aliasing_ppgtt || flags & GLOBAL_BIND) {
1919 vma->vm->insert_entries(vma->vm, pages,
1921 cache_level, pte_flags);
1924 if (dev_priv->mm.aliasing_ppgtt && flags & LOCAL_BIND) {
1925 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
1926 appgtt->base.insert_entries(&appgtt->base, pages,
1928 cache_level, pte_flags);
1934 static void ggtt_unbind_vma(struct i915_vma *vma)
1936 struct drm_device *dev = vma->vm->dev;
1937 struct drm_i915_private *dev_priv = dev->dev_private;
1938 struct drm_i915_gem_object *obj = vma->obj;
1939 const uint64_t size = min_t(uint64_t,
1943 if (vma->bound & GLOBAL_BIND) {
1944 vma->vm->clear_range(vma->vm,
1950 if (dev_priv->mm.aliasing_ppgtt && vma->bound & LOCAL_BIND) {
1951 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
1953 appgtt->base.clear_range(&appgtt->base,
1960 void i915_gem_gtt_finish_object(struct drm_i915_gem_object *obj)
1962 struct drm_device *dev = obj->base.dev;
1963 struct drm_i915_private *dev_priv = dev->dev_private;
1966 interruptible = do_idling(dev_priv);
1968 if (!obj->has_dma_mapping)
1969 dma_unmap_sg(&dev->pdev->dev,
1970 obj->pages->sgl, obj->pages->nents,
1971 PCI_DMA_BIDIRECTIONAL);
1973 undo_idling(dev_priv, interruptible);
1976 static void i915_gtt_color_adjust(struct drm_mm_node *node,
1977 unsigned long color,
1981 if (node->color != color)
1984 if (!list_empty(&node->node_list)) {
1985 node = list_entry(node->node_list.next,
1988 if (node->allocated && node->color != color)
1993 static int i915_gem_setup_global_gtt(struct drm_device *dev,
1994 unsigned long start,
1995 unsigned long mappable_end,
1998 /* Let GEM Manage all of the aperture.
2000 * However, leave one page at the end still bound to the scratch page.
2001 * There are a number of places where the hardware apparently prefetches
2002 * past the end of the object, and we've seen multiple hangs with the
2003 * GPU head pointer stuck in a batchbuffer bound at the last page of the
2004 * aperture. One page should be enough to keep any prefetching inside
2007 struct drm_i915_private *dev_priv = dev->dev_private;
2008 struct i915_address_space *ggtt_vm = &dev_priv->gtt.base;
2009 struct drm_mm_node *entry;
2010 struct drm_i915_gem_object *obj;
2011 unsigned long hole_start, hole_end;
2014 BUG_ON(mappable_end > end);
2016 /* Subtract the guard page ... */
2017 drm_mm_init(&ggtt_vm->mm, start, end - start - PAGE_SIZE);
2019 dev_priv->gtt.base.start = start;
2020 dev_priv->gtt.base.total = end - start;
2022 if (intel_vgpu_active(dev)) {
2023 ret = intel_vgt_balloon(dev);
2029 dev_priv->gtt.base.mm.color_adjust = i915_gtt_color_adjust;
2031 /* Mark any preallocated objects as occupied */
2032 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
2033 struct i915_vma *vma = i915_gem_obj_to_vma(obj, ggtt_vm);
2035 DRM_DEBUG_KMS("reserving preallocated space: %lx + %zx\n",
2036 i915_gem_obj_ggtt_offset(obj), obj->base.size);
2038 WARN_ON(i915_gem_obj_ggtt_bound(obj));
2039 ret = drm_mm_reserve_node(&ggtt_vm->mm, &vma->node);
2041 DRM_DEBUG_KMS("Reservation failed: %i\n", ret);
2044 vma->bound |= GLOBAL_BIND;
2047 /* Clear any non-preallocated blocks */
2048 drm_mm_for_each_hole(entry, &ggtt_vm->mm, hole_start, hole_end) {
2049 DRM_DEBUG_KMS("clearing unused GTT space: [%lx, %lx]\n",
2050 hole_start, hole_end);
2051 ggtt_vm->clear_range(ggtt_vm, hole_start,
2052 hole_end - hole_start, true);
2055 /* And finally clear the reserved guard page */
2056 ggtt_vm->clear_range(ggtt_vm, end - PAGE_SIZE, PAGE_SIZE, true);
2058 if (USES_PPGTT(dev) && !USES_FULL_PPGTT(dev)) {
2059 struct i915_hw_ppgtt *ppgtt;
2061 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2065 ret = __hw_ppgtt_init(dev, ppgtt);
2067 ppgtt->base.cleanup(&ppgtt->base);
2072 if (ppgtt->base.allocate_va_range)
2073 ret = ppgtt->base.allocate_va_range(&ppgtt->base, 0,
2076 ppgtt->base.cleanup(&ppgtt->base);
2081 ppgtt->base.clear_range(&ppgtt->base,
2086 dev_priv->mm.aliasing_ppgtt = ppgtt;
2092 void i915_gem_init_global_gtt(struct drm_device *dev)
2094 struct drm_i915_private *dev_priv = dev->dev_private;
2095 unsigned long gtt_size, mappable_size;
2097 gtt_size = dev_priv->gtt.base.total;
2098 mappable_size = dev_priv->gtt.mappable_end;
2100 i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
2103 void i915_global_gtt_cleanup(struct drm_device *dev)
2105 struct drm_i915_private *dev_priv = dev->dev_private;
2106 struct i915_address_space *vm = &dev_priv->gtt.base;
2108 if (dev_priv->mm.aliasing_ppgtt) {
2109 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
2111 ppgtt->base.cleanup(&ppgtt->base);
2114 if (drm_mm_initialized(&vm->mm)) {
2115 if (intel_vgpu_active(dev))
2116 intel_vgt_deballoon();
2118 drm_mm_takedown(&vm->mm);
2119 list_del(&vm->global_link);
2125 static int setup_scratch_page(struct drm_device *dev)
2127 struct drm_i915_private *dev_priv = dev->dev_private;
2129 dma_addr_t dma_addr;
2131 page = alloc_page(GFP_KERNEL | GFP_DMA32 | __GFP_ZERO);
2134 set_pages_uc(page, 1);
2136 #ifdef CONFIG_INTEL_IOMMU
2137 dma_addr = pci_map_page(dev->pdev, page, 0, PAGE_SIZE,
2138 PCI_DMA_BIDIRECTIONAL);
2139 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,
2353 unsigned long *mappable_end)
2355 struct drm_i915_private *dev_priv = dev->dev_private;
2356 unsigned int gtt_size;
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,
2401 unsigned long *mappable_end)
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 (%lx)\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,
2452 unsigned long *mappable_end)
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 = %zdM\n",
2518 gtt->base.total >> 20);
2519 DRM_DEBUG_DRIVER("GMADR size = %ldM\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 drm_device *dev = obj->base.dev;
2688 struct intel_rotation_info *rot_info = &ggtt_view->rotation_info;
2689 unsigned long size, pages, rot_pages;
2690 struct sg_page_iter sg_iter;
2692 dma_addr_t *page_addr_list;
2693 struct sg_table *st;
2694 unsigned int tile_pitch, tile_height;
2695 unsigned int width_pages, height_pages;
2698 pages = obj->base.size / PAGE_SIZE;
2700 /* Calculate tiling geometry. */
2701 tile_height = intel_tile_height(dev, rot_info->pixel_format,
2702 rot_info->fb_modifier);
2703 tile_pitch = PAGE_SIZE / tile_height;
2704 width_pages = DIV_ROUND_UP(rot_info->pitch, tile_pitch);
2705 height_pages = DIV_ROUND_UP(rot_info->height, tile_height);
2706 rot_pages = width_pages * height_pages;
2707 size = rot_pages * PAGE_SIZE;
2709 /* Allocate a temporary list of source pages for random access. */
2710 page_addr_list = drm_malloc_ab(pages, sizeof(dma_addr_t));
2711 if (!page_addr_list)
2712 return ERR_PTR(ret);
2714 /* Allocate target SG list. */
2715 st = kmalloc(sizeof(*st), GFP_KERNEL);
2719 ret = sg_alloc_table(st, rot_pages, GFP_KERNEL);
2723 /* Populate source page list from the object. */
2725 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) {
2726 page_addr_list[i] = sg_page_iter_dma_address(&sg_iter);
2730 /* Rotate the pages. */
2731 rotate_pages(page_addr_list, width_pages, height_pages, st);
2734 "Created rotated page mapping for object size %lu (pitch=%u, height=%u, pixel_format=0x%x, %ux%u tiles, %lu pages).\n",
2735 size, rot_info->pitch, rot_info->height,
2736 rot_info->pixel_format, width_pages, height_pages,
2739 drm_free_large(page_addr_list);
2746 drm_free_large(page_addr_list);
2749 "Failed to create rotated mapping for object size %lu! (%d) (pitch=%u, height=%u, pixel_format=0x%x, %ux%u tiles, %lu pages)\n",
2750 size, ret, rot_info->pitch, rot_info->height,
2751 rot_info->pixel_format, width_pages, height_pages,
2753 return ERR_PTR(ret);
2756 static struct sg_table *
2757 intel_partial_pages(const struct i915_ggtt_view *view,
2758 struct drm_i915_gem_object *obj)
2760 struct sg_table *st;
2761 struct scatterlist *sg;
2762 struct sg_page_iter obj_sg_iter;
2765 st = kmalloc(sizeof(*st), GFP_KERNEL);
2769 ret = sg_alloc_table(st, view->params.partial.size, GFP_KERNEL);
2775 for_each_sg_page(obj->pages->sgl, &obj_sg_iter, obj->pages->nents,
2776 view->params.partial.offset)
2778 if (st->nents >= view->params.partial.size)
2781 sg_set_page(sg, NULL, PAGE_SIZE, 0);
2782 sg_dma_address(sg) = sg_page_iter_dma_address(&obj_sg_iter);
2783 sg_dma_len(sg) = PAGE_SIZE;
2794 return ERR_PTR(ret);
2798 i915_get_ggtt_vma_pages(struct i915_vma *vma)
2802 if (vma->ggtt_view.pages)
2805 if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL)
2806 vma->ggtt_view.pages = vma->obj->pages;
2807 else if (vma->ggtt_view.type == I915_GGTT_VIEW_ROTATED)
2808 vma->ggtt_view.pages =
2809 intel_rotate_fb_obj_pages(&vma->ggtt_view, vma->obj);
2810 else if (vma->ggtt_view.type == I915_GGTT_VIEW_PARTIAL)
2811 vma->ggtt_view.pages =
2812 intel_partial_pages(&vma->ggtt_view, vma->obj);
2814 WARN_ONCE(1, "GGTT view %u not implemented!\n",
2815 vma->ggtt_view.type);
2817 if (!vma->ggtt_view.pages) {
2818 DRM_ERROR("Failed to get pages for GGTT view type %u!\n",
2819 vma->ggtt_view.type);
2821 } else if (IS_ERR(vma->ggtt_view.pages)) {
2822 ret = PTR_ERR(vma->ggtt_view.pages);
2823 vma->ggtt_view.pages = NULL;
2824 DRM_ERROR("Failed to get pages for VMA view type %u (%d)!\n",
2825 vma->ggtt_view.type, ret);
2832 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
2834 * @cache_level: mapping cache level
2835 * @flags: flags like global or local mapping
2837 * DMA addresses are taken from the scatter-gather table of this object (or of
2838 * this VMA in case of non-default GGTT views) and PTE entries set up.
2839 * Note that DMA addresses are also the only part of the SG table we care about.
2841 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
2847 if (WARN_ON(flags == 0))
2851 if (flags & PIN_GLOBAL)
2852 bind_flags |= GLOBAL_BIND;
2853 if (flags & PIN_USER)
2854 bind_flags |= LOCAL_BIND;
2856 if (flags & PIN_UPDATE)
2857 bind_flags |= vma->bound;
2859 bind_flags &= ~vma->bound;
2861 if (bind_flags == 0)
2864 if (vma->bound == 0 && vma->vm->allocate_va_range) {
2865 trace_i915_va_alloc(vma->vm,
2868 VM_TO_TRACE_NAME(vma->vm));
2870 ret = vma->vm->allocate_va_range(vma->vm,
2877 ret = vma->vm->bind_vma(vma, cache_level, bind_flags);
2881 vma->bound |= bind_flags;
2887 * i915_ggtt_view_size - Get the size of a GGTT view.
2888 * @obj: Object the view is of.
2889 * @view: The view in question.
2891 * @return The size of the GGTT view in bytes.
2894 i915_ggtt_view_size(struct drm_i915_gem_object *obj,
2895 const struct i915_ggtt_view *view)
2897 if (view->type == I915_GGTT_VIEW_NORMAL ||
2898 view->type == I915_GGTT_VIEW_ROTATED) {
2899 return obj->base.size;
2900 } else if (view->type == I915_GGTT_VIEW_PARTIAL) {
2901 return view->params.partial.size << PAGE_SHIFT;
2903 WARN_ONCE(1, "GGTT view %u not implemented!\n", view->type);
2904 return obj->base.size;