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(const dma_addr_t addr,
196 const enum i915_cache_level level)
198 gen8_pde_t pde = _PAGE_PRESENT | _PAGE_RW;
200 if (level != I915_CACHE_NONE)
201 pde |= PPAT_CACHED_PDE_INDEX;
203 pde |= PPAT_UNCACHED_INDEX;
207 static gen6_pte_t snb_pte_encode(dma_addr_t addr,
208 enum i915_cache_level level,
209 bool valid, u32 unused)
211 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
212 pte |= GEN6_PTE_ADDR_ENCODE(addr);
215 case I915_CACHE_L3_LLC:
217 pte |= GEN6_PTE_CACHE_LLC;
219 case I915_CACHE_NONE:
220 pte |= GEN6_PTE_UNCACHED;
229 static gen6_pte_t ivb_pte_encode(dma_addr_t addr,
230 enum i915_cache_level level,
231 bool valid, u32 unused)
233 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
234 pte |= GEN6_PTE_ADDR_ENCODE(addr);
237 case I915_CACHE_L3_LLC:
238 pte |= GEN7_PTE_CACHE_L3_LLC;
241 pte |= GEN6_PTE_CACHE_LLC;
243 case I915_CACHE_NONE:
244 pte |= GEN6_PTE_UNCACHED;
253 static gen6_pte_t byt_pte_encode(dma_addr_t addr,
254 enum i915_cache_level level,
255 bool valid, u32 flags)
257 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
258 pte |= GEN6_PTE_ADDR_ENCODE(addr);
260 if (!(flags & PTE_READ_ONLY))
261 pte |= BYT_PTE_WRITEABLE;
263 if (level != I915_CACHE_NONE)
264 pte |= BYT_PTE_SNOOPED_BY_CPU_CACHES;
269 static gen6_pte_t hsw_pte_encode(dma_addr_t addr,
270 enum i915_cache_level level,
271 bool valid, u32 unused)
273 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
274 pte |= HSW_PTE_ADDR_ENCODE(addr);
276 if (level != I915_CACHE_NONE)
277 pte |= HSW_WB_LLC_AGE3;
282 static gen6_pte_t iris_pte_encode(dma_addr_t addr,
283 enum i915_cache_level level,
284 bool valid, u32 unused)
286 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
287 pte |= HSW_PTE_ADDR_ENCODE(addr);
290 case I915_CACHE_NONE:
293 pte |= HSW_WT_ELLC_LLC_AGE3;
296 pte |= HSW_WB_ELLC_LLC_AGE3;
303 static int __setup_page_dma(struct drm_device *dev,
304 struct i915_page_dma *p, gfp_t flags)
306 struct device *device = &dev->pdev->dev;
308 p->page = alloc_page(flags);
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 int setup_page_dma(struct drm_device *dev, struct i915_page_dma *p)
325 return __setup_page_dma(dev, p, GFP_KERNEL);
328 static void cleanup_page_dma(struct drm_device *dev, struct i915_page_dma *p)
330 if (WARN_ON(!p->page))
333 dma_unmap_page(&dev->pdev->dev, p->daddr, 4096, PCI_DMA_BIDIRECTIONAL);
334 __free_page(p->page);
335 memset(p, 0, sizeof(*p));
338 static void *kmap_page_dma(struct i915_page_dma *p)
340 return kmap_atomic(p->page);
343 /* We use the flushing unmap only with ppgtt structures:
344 * page directories, page tables and scratch pages.
346 static void kunmap_page_dma(struct drm_device *dev, void *vaddr)
348 /* There are only few exceptions for gen >=6. chv and bxt.
349 * And we are not sure about the latter so play safe for now.
351 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
352 drm_clflush_virt_range(vaddr, PAGE_SIZE);
354 kunmap_atomic(vaddr);
357 #define kmap_px(px) kmap_page_dma(px_base(px))
358 #define kunmap_px(ppgtt, vaddr) kunmap_page_dma((ppgtt)->base.dev, (vaddr))
360 #define setup_px(dev, px) setup_page_dma((dev), px_base(px))
361 #define cleanup_px(dev, px) cleanup_page_dma((dev), px_base(px))
362 #define fill_px(dev, px, v) fill_page_dma((dev), px_base(px), (v))
363 #define fill32_px(dev, px, v) fill_page_dma_32((dev), px_base(px), (v))
365 static void fill_page_dma(struct drm_device *dev, struct i915_page_dma *p,
369 uint64_t * const vaddr = kmap_page_dma(p);
371 for (i = 0; i < 512; i++)
374 kunmap_page_dma(dev, vaddr);
377 static void fill_page_dma_32(struct drm_device *dev, struct i915_page_dma *p,
378 const uint32_t val32)
384 fill_page_dma(dev, p, v);
387 static void free_pt(struct drm_device *dev, struct i915_page_table *pt)
390 kfree(pt->used_ptes);
394 static void gen8_initialize_pt(struct i915_address_space *vm,
395 struct i915_page_table *pt)
397 gen8_pte_t scratch_pte;
399 scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
400 I915_CACHE_LLC, true);
402 fill_px(vm->dev, pt, scratch_pte);
405 static struct i915_page_table *alloc_pt(struct drm_device *dev)
407 struct i915_page_table *pt;
408 const size_t count = INTEL_INFO(dev)->gen >= 8 ?
409 GEN8_PTES : GEN6_PTES;
412 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
414 return ERR_PTR(-ENOMEM);
416 pt->used_ptes = kcalloc(BITS_TO_LONGS(count), sizeof(*pt->used_ptes),
422 ret = setup_px(dev, pt);
429 kfree(pt->used_ptes);
436 static void free_pd(struct drm_device *dev, struct i915_page_directory *pd)
440 kfree(pd->used_pdes);
445 static struct i915_page_directory *alloc_pd(struct drm_device *dev)
447 struct i915_page_directory *pd;
450 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
452 return ERR_PTR(-ENOMEM);
454 pd->used_pdes = kcalloc(BITS_TO_LONGS(I915_PDES),
455 sizeof(*pd->used_pdes), GFP_KERNEL);
459 ret = setup_px(dev, pd);
466 kfree(pd->used_pdes);
473 /* Broadwell Page Directory Pointer Descriptors */
474 static int gen8_write_pdp(struct drm_i915_gem_request *req,
478 struct intel_engine_cs *ring = req->ring;
483 ret = intel_ring_begin(req, 6);
487 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
488 intel_ring_emit(ring, GEN8_RING_PDP_UDW(ring, entry));
489 intel_ring_emit(ring, upper_32_bits(addr));
490 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
491 intel_ring_emit(ring, GEN8_RING_PDP_LDW(ring, entry));
492 intel_ring_emit(ring, lower_32_bits(addr));
493 intel_ring_advance(ring);
498 static int gen8_mm_switch(struct i915_hw_ppgtt *ppgtt,
499 struct drm_i915_gem_request *req)
503 for (i = GEN8_LEGACY_PDPES - 1; i >= 0; i--) {
504 const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i);
506 ret = gen8_write_pdp(req, i, pd_daddr);
514 static void gen8_ppgtt_clear_range(struct i915_address_space *vm,
519 struct i915_hw_ppgtt *ppgtt =
520 container_of(vm, struct i915_hw_ppgtt, base);
521 gen8_pte_t *pt_vaddr, scratch_pte;
522 unsigned pdpe = start >> GEN8_PDPE_SHIFT & GEN8_PDPE_MASK;
523 unsigned pde = start >> GEN8_PDE_SHIFT & GEN8_PDE_MASK;
524 unsigned pte = start >> GEN8_PTE_SHIFT & GEN8_PTE_MASK;
525 unsigned num_entries = length >> PAGE_SHIFT;
526 unsigned last_pte, i;
528 scratch_pte = gen8_pte_encode(px_dma(ppgtt->base.scratch_page),
529 I915_CACHE_LLC, use_scratch);
531 while (num_entries) {
532 struct i915_page_directory *pd;
533 struct i915_page_table *pt;
535 if (WARN_ON(!ppgtt->pdp.page_directory[pdpe]))
538 pd = ppgtt->pdp.page_directory[pdpe];
540 if (WARN_ON(!pd->page_table[pde]))
543 pt = pd->page_table[pde];
545 if (WARN_ON(!px_page(pt)))
548 last_pte = pte + num_entries;
549 if (last_pte > GEN8_PTES)
550 last_pte = GEN8_PTES;
552 pt_vaddr = kmap_px(pt);
554 for (i = pte; i < last_pte; i++) {
555 pt_vaddr[i] = scratch_pte;
559 kunmap_px(ppgtt, pt);
562 if (++pde == I915_PDES) {
569 static void gen8_ppgtt_insert_entries(struct i915_address_space *vm,
570 struct sg_table *pages,
572 enum i915_cache_level cache_level, u32 unused)
574 struct i915_hw_ppgtt *ppgtt =
575 container_of(vm, struct i915_hw_ppgtt, base);
576 gen8_pte_t *pt_vaddr;
577 unsigned pdpe = start >> GEN8_PDPE_SHIFT & GEN8_PDPE_MASK;
578 unsigned pde = start >> GEN8_PDE_SHIFT & GEN8_PDE_MASK;
579 unsigned pte = start >> GEN8_PTE_SHIFT & GEN8_PTE_MASK;
580 struct sg_page_iter sg_iter;
584 for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) {
585 if (WARN_ON(pdpe >= GEN8_LEGACY_PDPES))
588 if (pt_vaddr == NULL) {
589 struct i915_page_directory *pd = ppgtt->pdp.page_directory[pdpe];
590 struct i915_page_table *pt = pd->page_table[pde];
591 pt_vaddr = kmap_px(pt);
595 gen8_pte_encode(sg_page_iter_dma_address(&sg_iter),
597 if (++pte == GEN8_PTES) {
598 kunmap_px(ppgtt, pt_vaddr);
600 if (++pde == I915_PDES) {
609 kunmap_px(ppgtt, pt_vaddr);
612 static void gen8_initialize_pd(struct i915_address_space *vm,
613 struct i915_page_directory *pd)
615 struct i915_hw_ppgtt *ppgtt =
616 container_of(vm, struct i915_hw_ppgtt, base);
617 gen8_pde_t scratch_pde;
619 scratch_pde = gen8_pde_encode(px_dma(ppgtt->scratch_pt),
622 fill_px(vm->dev, pd, scratch_pde);
625 static void gen8_free_page_tables(struct i915_page_directory *pd, struct drm_device *dev)
632 for_each_set_bit(i, pd->used_pdes, I915_PDES) {
633 if (WARN_ON(!pd->page_table[i]))
636 free_pt(dev, pd->page_table[i]);
637 pd->page_table[i] = NULL;
641 static void gen8_ppgtt_cleanup(struct i915_address_space *vm)
643 struct i915_hw_ppgtt *ppgtt =
644 container_of(vm, struct i915_hw_ppgtt, base);
647 for_each_set_bit(i, ppgtt->pdp.used_pdpes, GEN8_LEGACY_PDPES) {
648 if (WARN_ON(!ppgtt->pdp.page_directory[i]))
651 gen8_free_page_tables(ppgtt->pdp.page_directory[i], ppgtt->base.dev);
652 free_pd(ppgtt->base.dev, ppgtt->pdp.page_directory[i]);
655 free_pd(ppgtt->base.dev, ppgtt->scratch_pd);
656 free_pt(ppgtt->base.dev, ppgtt->scratch_pt);
660 * gen8_ppgtt_alloc_pagetabs() - Allocate page tables for VA range.
661 * @ppgtt: Master ppgtt structure.
662 * @pd: Page directory for this address range.
663 * @start: Starting virtual address to begin allocations.
664 * @length Size of the allocations.
665 * @new_pts: Bitmap set by function with new allocations. Likely used by the
666 * caller to free on error.
668 * Allocate the required number of page tables. Extremely similar to
669 * gen8_ppgtt_alloc_page_directories(). The main difference is here we are limited by
670 * the page directory boundary (instead of the page directory pointer). That
671 * boundary is 1GB virtual. Therefore, unlike gen8_ppgtt_alloc_page_directories(), it is
672 * possible, and likely that the caller will need to use multiple calls of this
673 * function to achieve the appropriate allocation.
675 * Return: 0 if success; negative error code otherwise.
677 static int gen8_ppgtt_alloc_pagetabs(struct i915_hw_ppgtt *ppgtt,
678 struct i915_page_directory *pd,
681 unsigned long *new_pts)
683 struct drm_device *dev = ppgtt->base.dev;
684 struct i915_page_table *pt;
688 gen8_for_each_pde(pt, pd, start, length, temp, pde) {
689 /* Don't reallocate page tables */
691 /* Scratch is never allocated this way */
692 WARN_ON(pt == ppgtt->scratch_pt);
700 gen8_initialize_pt(&ppgtt->base, pt);
701 pd->page_table[pde] = pt;
702 set_bit(pde, new_pts);
708 for_each_set_bit(pde, new_pts, I915_PDES)
709 free_pt(dev, pd->page_table[pde]);
715 * gen8_ppgtt_alloc_page_directories() - Allocate page directories for VA range.
716 * @ppgtt: Master ppgtt structure.
717 * @pdp: Page directory pointer for this address range.
718 * @start: Starting virtual address to begin allocations.
719 * @length Size of the allocations.
720 * @new_pds Bitmap set by function with new allocations. Likely used by the
721 * caller to free on error.
723 * Allocate the required number of page directories starting at the pde index of
724 * @start, and ending at the pde index @start + @length. This function will skip
725 * over already allocated page directories within the range, and only allocate
726 * new ones, setting the appropriate pointer within the pdp as well as the
727 * correct position in the bitmap @new_pds.
729 * The function will only allocate the pages within the range for a give page
730 * directory pointer. In other words, if @start + @length straddles a virtually
731 * addressed PDP boundary (512GB for 4k pages), there will be more allocations
732 * required by the caller, This is not currently possible, and the BUG in the
733 * code will prevent it.
735 * Return: 0 if success; negative error code otherwise.
737 static int gen8_ppgtt_alloc_page_directories(struct i915_hw_ppgtt *ppgtt,
738 struct i915_page_directory_pointer *pdp,
741 unsigned long *new_pds)
743 struct drm_device *dev = ppgtt->base.dev;
744 struct i915_page_directory *pd;
748 WARN_ON(!bitmap_empty(new_pds, GEN8_LEGACY_PDPES));
750 gen8_for_each_pdpe(pd, pdp, start, length, temp, pdpe) {
758 gen8_initialize_pd(&ppgtt->base, pd);
759 pdp->page_directory[pdpe] = pd;
760 set_bit(pdpe, new_pds);
766 for_each_set_bit(pdpe, new_pds, GEN8_LEGACY_PDPES)
767 free_pd(dev, pdp->page_directory[pdpe]);
773 free_gen8_temp_bitmaps(unsigned long *new_pds, unsigned long **new_pts)
777 for (i = 0; i < GEN8_LEGACY_PDPES; i++)
783 /* Fills in the page directory bitmap, and the array of page tables bitmap. Both
784 * of these are based on the number of PDPEs in the system.
787 int __must_check alloc_gen8_temp_bitmaps(unsigned long **new_pds,
788 unsigned long ***new_pts)
794 pds = kcalloc(BITS_TO_LONGS(GEN8_LEGACY_PDPES), sizeof(unsigned long), GFP_KERNEL);
798 pts = kcalloc(GEN8_LEGACY_PDPES, sizeof(unsigned long *), GFP_KERNEL);
804 for (i = 0; i < GEN8_LEGACY_PDPES; i++) {
805 pts[i] = kcalloc(BITS_TO_LONGS(I915_PDES),
806 sizeof(unsigned long), GFP_KERNEL);
817 free_gen8_temp_bitmaps(pds, pts);
821 /* PDE TLBs are a pain to invalidate on GEN8+. When we modify
822 * the page table structures, we mark them dirty so that
823 * context switching/execlist queuing code takes extra steps
824 * to ensure that tlbs are flushed.
826 static void mark_tlbs_dirty(struct i915_hw_ppgtt *ppgtt)
828 ppgtt->pd_dirty_rings = INTEL_INFO(ppgtt->base.dev)->ring_mask;
831 static int gen8_alloc_va_range(struct i915_address_space *vm,
835 struct i915_hw_ppgtt *ppgtt =
836 container_of(vm, struct i915_hw_ppgtt, base);
837 unsigned long *new_page_dirs, **new_page_tables;
838 struct i915_page_directory *pd;
839 const uint64_t orig_start = start;
840 const uint64_t orig_length = length;
845 /* Wrap is never okay since we can only represent 48b, and we don't
846 * actually use the other side of the canonical address space.
848 if (WARN_ON(start + length < start))
851 if (WARN_ON(start + length > ppgtt->base.total))
854 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables);
858 /* Do the allocations first so we can easily bail out */
859 ret = gen8_ppgtt_alloc_page_directories(ppgtt, &ppgtt->pdp, start, length,
862 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
866 /* For every page directory referenced, allocate page tables */
867 gen8_for_each_pdpe(pd, &ppgtt->pdp, start, length, temp, pdpe) {
868 ret = gen8_ppgtt_alloc_pagetabs(ppgtt, pd, start, length,
869 new_page_tables[pdpe]);
875 length = orig_length;
877 /* Allocations have completed successfully, so set the bitmaps, and do
879 gen8_for_each_pdpe(pd, &ppgtt->pdp, start, length, temp, pdpe) {
880 gen8_pde_t *const page_directory = kmap_px(pd);
881 struct i915_page_table *pt;
882 uint64_t pd_len = gen8_clamp_pd(start, length);
883 uint64_t pd_start = start;
886 /* Every pd should be allocated, we just did that above. */
889 gen8_for_each_pde(pt, pd, pd_start, pd_len, temp, pde) {
890 /* Same reasoning as pd */
893 WARN_ON(!gen8_pte_count(pd_start, pd_len));
895 /* Set our used ptes within the page table */
896 bitmap_set(pt->used_ptes,
897 gen8_pte_index(pd_start),
898 gen8_pte_count(pd_start, pd_len));
900 /* Our pde is now pointing to the pagetable, pt */
901 set_bit(pde, pd->used_pdes);
903 /* Map the PDE to the page table */
904 page_directory[pde] = gen8_pde_encode(px_dma(pt),
907 /* NB: We haven't yet mapped ptes to pages. At this
908 * point we're still relying on insert_entries() */
911 kunmap_px(ppgtt, page_directory);
913 set_bit(pdpe, ppgtt->pdp.used_pdpes);
916 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
917 mark_tlbs_dirty(ppgtt);
922 for_each_set_bit(temp, new_page_tables[pdpe], I915_PDES)
923 free_pt(vm->dev, ppgtt->pdp.page_directory[pdpe]->page_table[temp]);
926 for_each_set_bit(pdpe, new_page_dirs, GEN8_LEGACY_PDPES)
927 free_pd(vm->dev, ppgtt->pdp.page_directory[pdpe]);
929 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
930 mark_tlbs_dirty(ppgtt);
935 * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers
936 * with a net effect resembling a 2-level page table in normal x86 terms. Each
937 * PDP represents 1GB of memory 4 * 512 * 512 * 4096 = 4GB legacy 32b address
941 static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
943 ppgtt->scratch_pt = alloc_pt(ppgtt->base.dev);
944 if (IS_ERR(ppgtt->scratch_pt))
945 return PTR_ERR(ppgtt->scratch_pt);
947 ppgtt->scratch_pd = alloc_pd(ppgtt->base.dev);
948 if (IS_ERR(ppgtt->scratch_pd))
949 return PTR_ERR(ppgtt->scratch_pd);
951 gen8_initialize_pt(&ppgtt->base, ppgtt->scratch_pt);
952 gen8_initialize_pd(&ppgtt->base, ppgtt->scratch_pd);
954 ppgtt->base.start = 0;
955 ppgtt->base.total = 1ULL << 32;
956 if (IS_ENABLED(CONFIG_X86_32))
957 /* While we have a proliferation of size_t variables
958 * we cannot represent the full ppgtt size on 32bit,
959 * so limit it to the same size as the GGTT (currently
962 ppgtt->base.total = to_i915(ppgtt->base.dev)->gtt.base.total;
963 ppgtt->base.cleanup = gen8_ppgtt_cleanup;
964 ppgtt->base.allocate_va_range = gen8_alloc_va_range;
965 ppgtt->base.insert_entries = gen8_ppgtt_insert_entries;
966 ppgtt->base.clear_range = gen8_ppgtt_clear_range;
967 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
968 ppgtt->base.bind_vma = ppgtt_bind_vma;
970 ppgtt->switch_mm = gen8_mm_switch;
975 static void gen6_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
977 struct i915_address_space *vm = &ppgtt->base;
978 struct i915_page_table *unused;
979 gen6_pte_t scratch_pte;
981 uint32_t pte, pde, temp;
982 uint32_t start = ppgtt->base.start, length = ppgtt->base.total;
984 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page), I915_CACHE_LLC, true, 0);
986 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde) {
988 gen6_pte_t *pt_vaddr;
989 const dma_addr_t pt_addr = px_dma(ppgtt->pd.page_table[pde]);
990 pd_entry = readl(ppgtt->pd_addr + pde);
991 expected = (GEN6_PDE_ADDR_ENCODE(pt_addr) | GEN6_PDE_VALID);
993 if (pd_entry != expected)
994 seq_printf(m, "\tPDE #%d mismatch: Actual PDE: %x Expected PDE: %x\n",
998 seq_printf(m, "\tPDE: %x\n", pd_entry);
1000 pt_vaddr = kmap_px(ppgtt->pd.page_table[pde]);
1002 for (pte = 0; pte < GEN6_PTES; pte+=4) {
1004 (pde * PAGE_SIZE * GEN6_PTES) +
1008 for (i = 0; i < 4; i++)
1009 if (pt_vaddr[pte + i] != scratch_pte)
1014 seq_printf(m, "\t\t0x%lx [%03d,%04d]: =", va, pde, pte);
1015 for (i = 0; i < 4; i++) {
1016 if (pt_vaddr[pte + i] != scratch_pte)
1017 seq_printf(m, " %08x", pt_vaddr[pte + i]);
1019 seq_puts(m, " SCRATCH ");
1023 kunmap_px(ppgtt, pt_vaddr);
1027 /* Write pde (index) from the page directory @pd to the page table @pt */
1028 static void gen6_write_pde(struct i915_page_directory *pd,
1029 const int pde, struct i915_page_table *pt)
1031 /* Caller needs to make sure the write completes if necessary */
1032 struct i915_hw_ppgtt *ppgtt =
1033 container_of(pd, struct i915_hw_ppgtt, pd);
1036 pd_entry = GEN6_PDE_ADDR_ENCODE(px_dma(pt));
1037 pd_entry |= GEN6_PDE_VALID;
1039 writel(pd_entry, ppgtt->pd_addr + pde);
1042 /* Write all the page tables found in the ppgtt structure to incrementing page
1044 static void gen6_write_page_range(struct drm_i915_private *dev_priv,
1045 struct i915_page_directory *pd,
1046 uint32_t start, uint32_t length)
1048 struct i915_page_table *pt;
1051 gen6_for_each_pde(pt, pd, start, length, temp, pde)
1052 gen6_write_pde(pd, pde, pt);
1054 /* Make sure write is complete before other code can use this page
1055 * table. Also require for WC mapped PTEs */
1056 readl(dev_priv->gtt.gsm);
1059 static uint32_t get_pd_offset(struct i915_hw_ppgtt *ppgtt)
1061 BUG_ON(ppgtt->pd.base.ggtt_offset & 0x3f);
1063 return (ppgtt->pd.base.ggtt_offset / 64) << 16;
1066 static int hsw_mm_switch(struct i915_hw_ppgtt *ppgtt,
1067 struct drm_i915_gem_request *req)
1069 struct intel_engine_cs *ring = req->ring;
1072 /* NB: TLBs must be flushed and invalidated before a switch */
1073 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1077 ret = intel_ring_begin(req, 6);
1081 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1082 intel_ring_emit(ring, RING_PP_DIR_DCLV(ring));
1083 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1084 intel_ring_emit(ring, RING_PP_DIR_BASE(ring));
1085 intel_ring_emit(ring, get_pd_offset(ppgtt));
1086 intel_ring_emit(ring, MI_NOOP);
1087 intel_ring_advance(ring);
1092 static int vgpu_mm_switch(struct i915_hw_ppgtt *ppgtt,
1093 struct drm_i915_gem_request *req)
1095 struct intel_engine_cs *ring = req->ring;
1096 struct drm_i915_private *dev_priv = to_i915(ppgtt->base.dev);
1098 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
1099 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
1103 static int gen7_mm_switch(struct i915_hw_ppgtt *ppgtt,
1104 struct drm_i915_gem_request *req)
1106 struct intel_engine_cs *ring = req->ring;
1109 /* NB: TLBs must be flushed and invalidated before a switch */
1110 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1114 ret = intel_ring_begin(req, 6);
1118 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1119 intel_ring_emit(ring, RING_PP_DIR_DCLV(ring));
1120 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1121 intel_ring_emit(ring, RING_PP_DIR_BASE(ring));
1122 intel_ring_emit(ring, get_pd_offset(ppgtt));
1123 intel_ring_emit(ring, MI_NOOP);
1124 intel_ring_advance(ring);
1126 /* XXX: RCS is the only one to auto invalidate the TLBs? */
1127 if (ring->id != RCS) {
1128 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1136 static int gen6_mm_switch(struct i915_hw_ppgtt *ppgtt,
1137 struct drm_i915_gem_request *req)
1139 struct intel_engine_cs *ring = req->ring;
1140 struct drm_device *dev = ppgtt->base.dev;
1141 struct drm_i915_private *dev_priv = dev->dev_private;
1144 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
1145 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
1147 POSTING_READ(RING_PP_DIR_DCLV(ring));
1152 static void gen8_ppgtt_enable(struct drm_device *dev)
1154 struct drm_i915_private *dev_priv = dev->dev_private;
1155 struct intel_engine_cs *ring;
1158 for_each_ring(ring, dev_priv, j) {
1159 I915_WRITE(RING_MODE_GEN7(ring),
1160 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1164 static void gen7_ppgtt_enable(struct drm_device *dev)
1166 struct drm_i915_private *dev_priv = dev->dev_private;
1167 struct intel_engine_cs *ring;
1168 uint32_t ecochk, ecobits;
1171 ecobits = I915_READ(GAC_ECO_BITS);
1172 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B);
1174 ecochk = I915_READ(GAM_ECOCHK);
1175 if (IS_HASWELL(dev)) {
1176 ecochk |= ECOCHK_PPGTT_WB_HSW;
1178 ecochk |= ECOCHK_PPGTT_LLC_IVB;
1179 ecochk &= ~ECOCHK_PPGTT_GFDT_IVB;
1181 I915_WRITE(GAM_ECOCHK, ecochk);
1183 for_each_ring(ring, dev_priv, i) {
1184 /* GFX_MODE is per-ring on gen7+ */
1185 I915_WRITE(RING_MODE_GEN7(ring),
1186 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1190 static void gen6_ppgtt_enable(struct drm_device *dev)
1192 struct drm_i915_private *dev_priv = dev->dev_private;
1193 uint32_t ecochk, gab_ctl, ecobits;
1195 ecobits = I915_READ(GAC_ECO_BITS);
1196 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_SNB_BIT |
1197 ECOBITS_PPGTT_CACHE64B);
1199 gab_ctl = I915_READ(GAB_CTL);
1200 I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT);
1202 ecochk = I915_READ(GAM_ECOCHK);
1203 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT | ECOCHK_PPGTT_CACHE64B);
1205 I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1208 /* PPGTT support for Sandybdrige/Gen6 and later */
1209 static void gen6_ppgtt_clear_range(struct i915_address_space *vm,
1214 struct i915_hw_ppgtt *ppgtt =
1215 container_of(vm, struct i915_hw_ppgtt, base);
1216 gen6_pte_t *pt_vaddr, scratch_pte;
1217 unsigned first_entry = start >> PAGE_SHIFT;
1218 unsigned num_entries = length >> PAGE_SHIFT;
1219 unsigned act_pt = first_entry / GEN6_PTES;
1220 unsigned first_pte = first_entry % GEN6_PTES;
1221 unsigned last_pte, i;
1223 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1224 I915_CACHE_LLC, true, 0);
1226 while (num_entries) {
1227 last_pte = first_pte + num_entries;
1228 if (last_pte > GEN6_PTES)
1229 last_pte = GEN6_PTES;
1231 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1233 for (i = first_pte; i < last_pte; i++)
1234 pt_vaddr[i] = scratch_pte;
1236 kunmap_px(ppgtt, pt_vaddr);
1238 num_entries -= last_pte - first_pte;
1244 static void gen6_ppgtt_insert_entries(struct i915_address_space *vm,
1245 struct sg_table *pages,
1247 enum i915_cache_level cache_level, u32 flags)
1249 struct i915_hw_ppgtt *ppgtt =
1250 container_of(vm, struct i915_hw_ppgtt, base);
1251 gen6_pte_t *pt_vaddr;
1252 unsigned first_entry = start >> PAGE_SHIFT;
1253 unsigned act_pt = first_entry / GEN6_PTES;
1254 unsigned act_pte = first_entry % GEN6_PTES;
1255 struct sg_page_iter sg_iter;
1258 for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) {
1259 if (pt_vaddr == NULL)
1260 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1263 vm->pte_encode(sg_page_iter_dma_address(&sg_iter),
1264 cache_level, true, flags);
1266 if (++act_pte == GEN6_PTES) {
1267 kunmap_px(ppgtt, pt_vaddr);
1274 kunmap_px(ppgtt, pt_vaddr);
1277 static void gen6_initialize_pt(struct i915_address_space *vm,
1278 struct i915_page_table *pt)
1280 gen6_pte_t scratch_pte;
1282 WARN_ON(px_dma(vm->scratch_page) == 0);
1284 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1285 I915_CACHE_LLC, true, 0);
1287 fill32_px(vm->dev, pt, scratch_pte);
1290 static int gen6_alloc_va_range(struct i915_address_space *vm,
1291 uint64_t start_in, uint64_t length_in)
1293 DECLARE_BITMAP(new_page_tables, I915_PDES);
1294 struct drm_device *dev = vm->dev;
1295 struct drm_i915_private *dev_priv = dev->dev_private;
1296 struct i915_hw_ppgtt *ppgtt =
1297 container_of(vm, struct i915_hw_ppgtt, base);
1298 struct i915_page_table *pt;
1299 uint32_t start, length, start_save, length_save;
1303 if (WARN_ON(start_in + length_in > ppgtt->base.total))
1306 start = start_save = start_in;
1307 length = length_save = length_in;
1309 bitmap_zero(new_page_tables, I915_PDES);
1311 /* The allocation is done in two stages so that we can bail out with
1312 * minimal amount of pain. The first stage finds new page tables that
1313 * need allocation. The second stage marks use ptes within the page
1316 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1317 if (pt != ppgtt->scratch_pt) {
1318 WARN_ON(bitmap_empty(pt->used_ptes, GEN6_PTES));
1322 /* We've already allocated a page table */
1323 WARN_ON(!bitmap_empty(pt->used_ptes, GEN6_PTES));
1331 gen6_initialize_pt(vm, pt);
1333 ppgtt->pd.page_table[pde] = pt;
1334 set_bit(pde, new_page_tables);
1335 trace_i915_page_table_entry_alloc(vm, pde, start, GEN6_PDE_SHIFT);
1339 length = length_save;
1341 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1342 DECLARE_BITMAP(tmp_bitmap, GEN6_PTES);
1344 bitmap_zero(tmp_bitmap, GEN6_PTES);
1345 bitmap_set(tmp_bitmap, gen6_pte_index(start),
1346 gen6_pte_count(start, length));
1348 if (test_and_clear_bit(pde, new_page_tables))
1349 gen6_write_pde(&ppgtt->pd, pde, pt);
1351 trace_i915_page_table_entry_map(vm, pde, pt,
1352 gen6_pte_index(start),
1353 gen6_pte_count(start, length),
1355 bitmap_or(pt->used_ptes, tmp_bitmap, pt->used_ptes,
1359 WARN_ON(!bitmap_empty(new_page_tables, I915_PDES));
1361 /* Make sure write is complete before other code can use this page
1362 * table. Also require for WC mapped PTEs */
1363 readl(dev_priv->gtt.gsm);
1365 mark_tlbs_dirty(ppgtt);
1369 for_each_set_bit(pde, new_page_tables, I915_PDES) {
1370 struct i915_page_table *pt = ppgtt->pd.page_table[pde];
1372 ppgtt->pd.page_table[pde] = ppgtt->scratch_pt;
1373 free_pt(vm->dev, pt);
1376 mark_tlbs_dirty(ppgtt);
1380 static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
1382 struct i915_hw_ppgtt *ppgtt =
1383 container_of(vm, struct i915_hw_ppgtt, base);
1384 struct i915_page_table *pt;
1388 drm_mm_remove_node(&ppgtt->node);
1390 gen6_for_all_pdes(pt, ppgtt, pde) {
1391 if (pt != ppgtt->scratch_pt)
1392 free_pt(ppgtt->base.dev, pt);
1395 free_pt(ppgtt->base.dev, ppgtt->scratch_pt);
1398 static int gen6_ppgtt_allocate_page_directories(struct i915_hw_ppgtt *ppgtt)
1400 struct drm_device *dev = ppgtt->base.dev;
1401 struct drm_i915_private *dev_priv = dev->dev_private;
1402 bool retried = false;
1405 /* PPGTT PDEs reside in the GGTT and consists of 512 entries. The
1406 * allocator works in address space sizes, so it's multiplied by page
1407 * size. We allocate at the top of the GTT to avoid fragmentation.
1409 BUG_ON(!drm_mm_initialized(&dev_priv->gtt.base.mm));
1410 ppgtt->scratch_pt = alloc_pt(ppgtt->base.dev);
1411 if (IS_ERR(ppgtt->scratch_pt))
1412 return PTR_ERR(ppgtt->scratch_pt);
1414 gen6_initialize_pt(&ppgtt->base, ppgtt->scratch_pt);
1417 ret = drm_mm_insert_node_in_range_generic(&dev_priv->gtt.base.mm,
1418 &ppgtt->node, GEN6_PD_SIZE,
1420 0, dev_priv->gtt.base.total,
1422 if (ret == -ENOSPC && !retried) {
1423 ret = i915_gem_evict_something(dev, &dev_priv->gtt.base,
1424 GEN6_PD_SIZE, GEN6_PD_ALIGN,
1426 0, dev_priv->gtt.base.total,
1439 if (ppgtt->node.start < dev_priv->gtt.mappable_end)
1440 DRM_DEBUG("Forced to use aperture for PDEs\n");
1445 free_pt(ppgtt->base.dev, ppgtt->scratch_pt);
1449 static int gen6_ppgtt_alloc(struct i915_hw_ppgtt *ppgtt)
1451 return gen6_ppgtt_allocate_page_directories(ppgtt);
1454 static void gen6_scratch_va_range(struct i915_hw_ppgtt *ppgtt,
1455 uint64_t start, uint64_t length)
1457 struct i915_page_table *unused;
1460 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde)
1461 ppgtt->pd.page_table[pde] = ppgtt->scratch_pt;
1464 static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
1466 struct drm_device *dev = ppgtt->base.dev;
1467 struct drm_i915_private *dev_priv = dev->dev_private;
1470 ppgtt->base.pte_encode = dev_priv->gtt.base.pte_encode;
1472 ppgtt->switch_mm = gen6_mm_switch;
1473 } else if (IS_HASWELL(dev)) {
1474 ppgtt->switch_mm = hsw_mm_switch;
1475 } else if (IS_GEN7(dev)) {
1476 ppgtt->switch_mm = gen7_mm_switch;
1480 if (intel_vgpu_active(dev))
1481 ppgtt->switch_mm = vgpu_mm_switch;
1483 ret = gen6_ppgtt_alloc(ppgtt);
1487 ppgtt->base.allocate_va_range = gen6_alloc_va_range;
1488 ppgtt->base.clear_range = gen6_ppgtt_clear_range;
1489 ppgtt->base.insert_entries = gen6_ppgtt_insert_entries;
1490 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
1491 ppgtt->base.bind_vma = ppgtt_bind_vma;
1492 ppgtt->base.cleanup = gen6_ppgtt_cleanup;
1493 ppgtt->base.start = 0;
1494 ppgtt->base.total = I915_PDES * GEN6_PTES * PAGE_SIZE;
1495 ppgtt->debug_dump = gen6_dump_ppgtt;
1497 ppgtt->pd.base.ggtt_offset =
1498 ppgtt->node.start / PAGE_SIZE * sizeof(gen6_pte_t);
1500 ppgtt->pd_addr = (gen6_pte_t __iomem *)dev_priv->gtt.gsm +
1501 ppgtt->pd.base.ggtt_offset / sizeof(gen6_pte_t);
1503 gen6_scratch_va_range(ppgtt, 0, ppgtt->base.total);
1505 gen6_write_page_range(dev_priv, &ppgtt->pd, 0, ppgtt->base.total);
1507 DRM_DEBUG_DRIVER("Allocated pde space (%lldM) at GTT entry: %llx\n",
1508 ppgtt->node.size >> 20,
1509 ppgtt->node.start / PAGE_SIZE);
1511 DRM_DEBUG("Adding PPGTT at offset %x\n",
1512 ppgtt->pd.base.ggtt_offset << 10);
1517 static int __hw_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
1519 struct drm_i915_private *dev_priv = dev->dev_private;
1521 ppgtt->base.dev = dev;
1522 ppgtt->base.scratch_page = dev_priv->gtt.base.scratch_page;
1524 if (INTEL_INFO(dev)->gen < 8)
1525 return gen6_ppgtt_init(ppgtt);
1527 return gen8_ppgtt_init(ppgtt);
1530 int i915_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
1532 struct drm_i915_private *dev_priv = dev->dev_private;
1535 ret = __hw_ppgtt_init(dev, ppgtt);
1537 kref_init(&ppgtt->ref);
1538 drm_mm_init(&ppgtt->base.mm, ppgtt->base.start,
1540 i915_init_vm(dev_priv, &ppgtt->base);
1546 int i915_ppgtt_init_hw(struct drm_device *dev)
1548 /* In the case of execlists, PPGTT is enabled by the context descriptor
1549 * and the PDPs are contained within the context itself. We don't
1550 * need to do anything here. */
1551 if (i915.enable_execlists)
1554 if (!USES_PPGTT(dev))
1558 gen6_ppgtt_enable(dev);
1559 else if (IS_GEN7(dev))
1560 gen7_ppgtt_enable(dev);
1561 else if (INTEL_INFO(dev)->gen >= 8)
1562 gen8_ppgtt_enable(dev);
1564 MISSING_CASE(INTEL_INFO(dev)->gen);
1569 int i915_ppgtt_init_ring(struct drm_i915_gem_request *req)
1571 struct drm_i915_private *dev_priv = req->ring->dev->dev_private;
1572 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
1574 if (i915.enable_execlists)
1580 return ppgtt->switch_mm(ppgtt, req);
1583 struct i915_hw_ppgtt *
1584 i915_ppgtt_create(struct drm_device *dev, struct drm_i915_file_private *fpriv)
1586 struct i915_hw_ppgtt *ppgtt;
1589 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
1591 return ERR_PTR(-ENOMEM);
1593 ret = i915_ppgtt_init(dev, ppgtt);
1596 return ERR_PTR(ret);
1599 ppgtt->file_priv = fpriv;
1601 trace_i915_ppgtt_create(&ppgtt->base);
1606 void i915_ppgtt_release(struct kref *kref)
1608 struct i915_hw_ppgtt *ppgtt =
1609 container_of(kref, struct i915_hw_ppgtt, ref);
1611 trace_i915_ppgtt_release(&ppgtt->base);
1613 /* vmas should already be unbound */
1614 WARN_ON(!list_empty(&ppgtt->base.active_list));
1615 WARN_ON(!list_empty(&ppgtt->base.inactive_list));
1617 list_del(&ppgtt->base.global_link);
1618 drm_mm_takedown(&ppgtt->base.mm);
1620 ppgtt->base.cleanup(&ppgtt->base);
1624 extern int intel_iommu_gfx_mapped;
1625 /* Certain Gen5 chipsets require require idling the GPU before
1626 * unmapping anything from the GTT when VT-d is enabled.
1628 static bool needs_idle_maps(struct drm_device *dev)
1630 #ifdef CONFIG_INTEL_IOMMU
1631 /* Query intel_iommu to see if we need the workaround. Presumably that
1634 if (IS_GEN5(dev) && IS_MOBILE(dev) && intel_iommu_gfx_mapped)
1640 static bool do_idling(struct drm_i915_private *dev_priv)
1642 bool ret = dev_priv->mm.interruptible;
1644 if (unlikely(dev_priv->gtt.do_idle_maps)) {
1645 dev_priv->mm.interruptible = false;
1646 if (i915_gpu_idle(dev_priv->dev)) {
1647 DRM_ERROR("Couldn't idle GPU\n");
1648 /* Wait a bit, in hopes it avoids the hang */
1656 static void undo_idling(struct drm_i915_private *dev_priv, bool interruptible)
1658 if (unlikely(dev_priv->gtt.do_idle_maps))
1659 dev_priv->mm.interruptible = interruptible;
1662 void i915_check_and_clear_faults(struct drm_device *dev)
1664 struct drm_i915_private *dev_priv = dev->dev_private;
1665 struct intel_engine_cs *ring;
1668 if (INTEL_INFO(dev)->gen < 6)
1671 for_each_ring(ring, dev_priv, i) {
1673 fault_reg = I915_READ(RING_FAULT_REG(ring));
1674 if (fault_reg & RING_FAULT_VALID) {
1675 DRM_DEBUG_DRIVER("Unexpected fault\n"
1677 "\tAddress space: %s\n"
1680 fault_reg & PAGE_MASK,
1681 fault_reg & RING_FAULT_GTTSEL_MASK ? "GGTT" : "PPGTT",
1682 RING_FAULT_SRCID(fault_reg),
1683 RING_FAULT_FAULT_TYPE(fault_reg));
1684 I915_WRITE(RING_FAULT_REG(ring),
1685 fault_reg & ~RING_FAULT_VALID);
1688 POSTING_READ(RING_FAULT_REG(&dev_priv->ring[RCS]));
1691 static void i915_ggtt_flush(struct drm_i915_private *dev_priv)
1693 if (INTEL_INFO(dev_priv->dev)->gen < 6) {
1694 intel_gtt_chipset_flush();
1696 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
1697 POSTING_READ(GFX_FLSH_CNTL_GEN6);
1701 void i915_gem_suspend_gtt_mappings(struct drm_device *dev)
1703 struct drm_i915_private *dev_priv = dev->dev_private;
1705 /* Don't bother messing with faults pre GEN6 as we have little
1706 * documentation supporting that it's a good idea.
1708 if (INTEL_INFO(dev)->gen < 6)
1711 i915_check_and_clear_faults(dev);
1713 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
1714 dev_priv->gtt.base.start,
1715 dev_priv->gtt.base.total,
1718 i915_ggtt_flush(dev_priv);
1721 int i915_gem_gtt_prepare_object(struct drm_i915_gem_object *obj)
1723 if (obj->has_dma_mapping)
1726 if (!dma_map_sg(&obj->base.dev->pdev->dev,
1727 obj->pages->sgl, obj->pages->nents,
1728 PCI_DMA_BIDIRECTIONAL))
1734 static void gen8_set_pte(void __iomem *addr, gen8_pte_t pte)
1739 iowrite32((u32)pte, addr);
1740 iowrite32(pte >> 32, addr + 4);
1744 static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
1745 struct sg_table *st,
1747 enum i915_cache_level level, u32 unused)
1749 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1750 unsigned first_entry = start >> PAGE_SHIFT;
1751 gen8_pte_t __iomem *gtt_entries =
1752 (gen8_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
1754 struct sg_page_iter sg_iter;
1755 dma_addr_t addr = 0; /* shut up gcc */
1757 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
1758 addr = sg_dma_address(sg_iter.sg) +
1759 (sg_iter.sg_pgoffset << PAGE_SHIFT);
1760 gen8_set_pte(>t_entries[i],
1761 gen8_pte_encode(addr, level, true));
1766 * XXX: This serves as a posting read to make sure that the PTE has
1767 * actually been updated. There is some concern that even though
1768 * registers and PTEs are within the same BAR that they are potentially
1769 * of NUMA access patterns. Therefore, even with the way we assume
1770 * hardware should work, we must keep this posting read for paranoia.
1773 WARN_ON(readq(>t_entries[i-1])
1774 != gen8_pte_encode(addr, level, true));
1776 /* This next bit makes the above posting read even more important. We
1777 * want to flush the TLBs only after we're certain all the PTE updates
1780 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
1781 POSTING_READ(GFX_FLSH_CNTL_GEN6);
1785 * Binds an object into the global gtt with the specified cache level. The object
1786 * will be accessible to the GPU via commands whose operands reference offsets
1787 * within the global GTT as well as accessible by the GPU through the GMADR
1788 * mapped BAR (dev_priv->mm.gtt->gtt).
1790 static void gen6_ggtt_insert_entries(struct i915_address_space *vm,
1791 struct sg_table *st,
1793 enum i915_cache_level level, u32 flags)
1795 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1796 unsigned first_entry = start >> PAGE_SHIFT;
1797 gen6_pte_t __iomem *gtt_entries =
1798 (gen6_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
1800 struct sg_page_iter sg_iter;
1801 dma_addr_t addr = 0;
1803 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
1804 addr = sg_page_iter_dma_address(&sg_iter);
1805 iowrite32(vm->pte_encode(addr, level, true, flags), >t_entries[i]);
1809 /* XXX: This serves as a posting read to make sure that the PTE has
1810 * actually been updated. There is some concern that even though
1811 * registers and PTEs are within the same BAR that they are potentially
1812 * of NUMA access patterns. Therefore, even with the way we assume
1813 * hardware should work, we must keep this posting read for paranoia.
1816 unsigned long gtt = readl(>t_entries[i-1]);
1817 WARN_ON(gtt != vm->pte_encode(addr, level, true, flags));
1820 /* This next bit makes the above posting read even more important. We
1821 * want to flush the TLBs only after we're certain all the PTE updates
1824 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
1825 POSTING_READ(GFX_FLSH_CNTL_GEN6);
1828 static void gen8_ggtt_clear_range(struct i915_address_space *vm,
1833 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1834 unsigned first_entry = start >> PAGE_SHIFT;
1835 unsigned num_entries = length >> PAGE_SHIFT;
1836 gen8_pte_t scratch_pte, __iomem *gtt_base =
1837 (gen8_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
1838 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
1841 if (WARN(num_entries > max_entries,
1842 "First entry = %d; Num entries = %d (max=%d)\n",
1843 first_entry, num_entries, max_entries))
1844 num_entries = max_entries;
1846 scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
1849 for (i = 0; i < num_entries; i++)
1850 gen8_set_pte(>t_base[i], scratch_pte);
1854 static void gen6_ggtt_clear_range(struct i915_address_space *vm,
1859 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1860 unsigned first_entry = start >> PAGE_SHIFT;
1861 unsigned num_entries = length >> PAGE_SHIFT;
1862 gen6_pte_t scratch_pte, __iomem *gtt_base =
1863 (gen6_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
1864 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
1867 if (WARN(num_entries > max_entries,
1868 "First entry = %d; Num entries = %d (max=%d)\n",
1869 first_entry, num_entries, max_entries))
1870 num_entries = max_entries;
1872 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1873 I915_CACHE_LLC, use_scratch, 0);
1875 for (i = 0; i < num_entries; i++)
1876 iowrite32(scratch_pte, >t_base[i]);
1880 static void i915_ggtt_insert_entries(struct i915_address_space *vm,
1881 struct sg_table *pages,
1883 enum i915_cache_level cache_level, u32 unused)
1885 unsigned int flags = (cache_level == I915_CACHE_NONE) ?
1886 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
1888 intel_gtt_insert_sg_entries(pages, start >> PAGE_SHIFT, flags);
1892 static void i915_ggtt_clear_range(struct i915_address_space *vm,
1897 unsigned first_entry = start >> PAGE_SHIFT;
1898 unsigned num_entries = length >> PAGE_SHIFT;
1899 intel_gtt_clear_range(first_entry, num_entries);
1902 static int ggtt_bind_vma(struct i915_vma *vma,
1903 enum i915_cache_level cache_level,
1906 struct drm_device *dev = vma->vm->dev;
1907 struct drm_i915_private *dev_priv = dev->dev_private;
1908 struct drm_i915_gem_object *obj = vma->obj;
1909 struct sg_table *pages = obj->pages;
1913 ret = i915_get_ggtt_vma_pages(vma);
1916 pages = vma->ggtt_view.pages;
1918 /* Currently applicable only to VLV */
1920 pte_flags |= PTE_READ_ONLY;
1923 if (!dev_priv->mm.aliasing_ppgtt || flags & GLOBAL_BIND) {
1924 vma->vm->insert_entries(vma->vm, pages,
1926 cache_level, pte_flags);
1929 if (dev_priv->mm.aliasing_ppgtt && flags & LOCAL_BIND) {
1930 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
1931 appgtt->base.insert_entries(&appgtt->base, pages,
1933 cache_level, pte_flags);
1939 static void ggtt_unbind_vma(struct i915_vma *vma)
1941 struct drm_device *dev = vma->vm->dev;
1942 struct drm_i915_private *dev_priv = dev->dev_private;
1943 struct drm_i915_gem_object *obj = vma->obj;
1944 const uint64_t size = min_t(uint64_t,
1948 if (vma->bound & GLOBAL_BIND) {
1949 vma->vm->clear_range(vma->vm,
1955 if (dev_priv->mm.aliasing_ppgtt && vma->bound & LOCAL_BIND) {
1956 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
1958 appgtt->base.clear_range(&appgtt->base,
1965 void i915_gem_gtt_finish_object(struct drm_i915_gem_object *obj)
1967 struct drm_device *dev = obj->base.dev;
1968 struct drm_i915_private *dev_priv = dev->dev_private;
1971 interruptible = do_idling(dev_priv);
1973 if (!obj->has_dma_mapping)
1974 dma_unmap_sg(&dev->pdev->dev,
1975 obj->pages->sgl, obj->pages->nents,
1976 PCI_DMA_BIDIRECTIONAL);
1978 undo_idling(dev_priv, interruptible);
1981 static void i915_gtt_color_adjust(struct drm_mm_node *node,
1982 unsigned long color,
1986 if (node->color != color)
1989 if (!list_empty(&node->node_list)) {
1990 node = list_entry(node->node_list.next,
1993 if (node->allocated && node->color != color)
1998 static int i915_gem_setup_global_gtt(struct drm_device *dev,
1999 unsigned long start,
2000 unsigned long mappable_end,
2003 /* Let GEM Manage all of the aperture.
2005 * However, leave one page at the end still bound to the scratch page.
2006 * There are a number of places where the hardware apparently prefetches
2007 * past the end of the object, and we've seen multiple hangs with the
2008 * GPU head pointer stuck in a batchbuffer bound at the last page of the
2009 * aperture. One page should be enough to keep any prefetching inside
2012 struct drm_i915_private *dev_priv = dev->dev_private;
2013 struct i915_address_space *ggtt_vm = &dev_priv->gtt.base;
2014 struct drm_mm_node *entry;
2015 struct drm_i915_gem_object *obj;
2016 unsigned long hole_start, hole_end;
2019 BUG_ON(mappable_end > end);
2021 /* Subtract the guard page ... */
2022 drm_mm_init(&ggtt_vm->mm, start, end - start - PAGE_SIZE);
2024 dev_priv->gtt.base.start = start;
2025 dev_priv->gtt.base.total = end - start;
2027 if (intel_vgpu_active(dev)) {
2028 ret = intel_vgt_balloon(dev);
2034 dev_priv->gtt.base.mm.color_adjust = i915_gtt_color_adjust;
2036 /* Mark any preallocated objects as occupied */
2037 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
2038 struct i915_vma *vma = i915_gem_obj_to_vma(obj, ggtt_vm);
2040 DRM_DEBUG_KMS("reserving preallocated space: %lx + %zx\n",
2041 i915_gem_obj_ggtt_offset(obj), obj->base.size);
2043 WARN_ON(i915_gem_obj_ggtt_bound(obj));
2044 ret = drm_mm_reserve_node(&ggtt_vm->mm, &vma->node);
2046 DRM_DEBUG_KMS("Reservation failed: %i\n", ret);
2049 vma->bound |= GLOBAL_BIND;
2052 /* Clear any non-preallocated blocks */
2053 drm_mm_for_each_hole(entry, &ggtt_vm->mm, hole_start, hole_end) {
2054 DRM_DEBUG_KMS("clearing unused GTT space: [%lx, %lx]\n",
2055 hole_start, hole_end);
2056 ggtt_vm->clear_range(ggtt_vm, hole_start,
2057 hole_end - hole_start, true);
2060 /* And finally clear the reserved guard page */
2061 ggtt_vm->clear_range(ggtt_vm, end - PAGE_SIZE, PAGE_SIZE, true);
2063 if (USES_PPGTT(dev) && !USES_FULL_PPGTT(dev)) {
2064 struct i915_hw_ppgtt *ppgtt;
2066 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2070 ret = __hw_ppgtt_init(dev, ppgtt);
2072 ppgtt->base.cleanup(&ppgtt->base);
2077 if (ppgtt->base.allocate_va_range)
2078 ret = ppgtt->base.allocate_va_range(&ppgtt->base, 0,
2081 ppgtt->base.cleanup(&ppgtt->base);
2086 ppgtt->base.clear_range(&ppgtt->base,
2091 dev_priv->mm.aliasing_ppgtt = ppgtt;
2097 void i915_gem_init_global_gtt(struct drm_device *dev)
2099 struct drm_i915_private *dev_priv = dev->dev_private;
2100 u64 gtt_size, mappable_size;
2102 gtt_size = dev_priv->gtt.base.total;
2103 mappable_size = dev_priv->gtt.mappable_end;
2105 i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
2108 void i915_global_gtt_cleanup(struct drm_device *dev)
2110 struct drm_i915_private *dev_priv = dev->dev_private;
2111 struct i915_address_space *vm = &dev_priv->gtt.base;
2113 if (dev_priv->mm.aliasing_ppgtt) {
2114 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
2116 ppgtt->base.cleanup(&ppgtt->base);
2119 if (drm_mm_initialized(&vm->mm)) {
2120 if (intel_vgpu_active(dev))
2121 intel_vgt_deballoon();
2123 drm_mm_takedown(&vm->mm);
2124 list_del(&vm->global_link);
2130 static int alloc_scratch_page(struct i915_address_space *vm)
2132 struct i915_page_scratch *sp;
2135 WARN_ON(vm->scratch_page);
2137 sp = kzalloc(sizeof(*sp), GFP_KERNEL);
2141 ret = __setup_page_dma(vm->dev, px_base(sp), GFP_DMA32 | __GFP_ZERO);
2147 set_pages_uc(px_page(sp), 1);
2149 vm->scratch_page = sp;
2154 static void free_scratch_page(struct i915_address_space *vm)
2156 struct i915_page_scratch *sp = vm->scratch_page;
2158 set_pages_wb(px_page(sp), 1);
2160 cleanup_px(vm->dev, sp);
2163 vm->scratch_page = NULL;
2166 static unsigned int gen6_get_total_gtt_size(u16 snb_gmch_ctl)
2168 snb_gmch_ctl >>= SNB_GMCH_GGMS_SHIFT;
2169 snb_gmch_ctl &= SNB_GMCH_GGMS_MASK;
2170 return snb_gmch_ctl << 20;
2173 static unsigned int gen8_get_total_gtt_size(u16 bdw_gmch_ctl)
2175 bdw_gmch_ctl >>= BDW_GMCH_GGMS_SHIFT;
2176 bdw_gmch_ctl &= BDW_GMCH_GGMS_MASK;
2178 bdw_gmch_ctl = 1 << bdw_gmch_ctl;
2180 #ifdef CONFIG_X86_32
2181 /* Limit 32b platforms to a 2GB GGTT: 4 << 20 / pte size * PAGE_SIZE */
2182 if (bdw_gmch_ctl > 4)
2186 return bdw_gmch_ctl << 20;
2189 static unsigned int chv_get_total_gtt_size(u16 gmch_ctrl)
2191 gmch_ctrl >>= SNB_GMCH_GGMS_SHIFT;
2192 gmch_ctrl &= SNB_GMCH_GGMS_MASK;
2195 return 1 << (20 + gmch_ctrl);
2200 static size_t gen6_get_stolen_size(u16 snb_gmch_ctl)
2202 snb_gmch_ctl >>= SNB_GMCH_GMS_SHIFT;
2203 snb_gmch_ctl &= SNB_GMCH_GMS_MASK;
2204 return snb_gmch_ctl << 25; /* 32 MB units */
2207 static size_t gen8_get_stolen_size(u16 bdw_gmch_ctl)
2209 bdw_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2210 bdw_gmch_ctl &= BDW_GMCH_GMS_MASK;
2211 return bdw_gmch_ctl << 25; /* 32 MB units */
2214 static size_t chv_get_stolen_size(u16 gmch_ctrl)
2216 gmch_ctrl >>= SNB_GMCH_GMS_SHIFT;
2217 gmch_ctrl &= SNB_GMCH_GMS_MASK;
2220 * 0x0 to 0x10: 32MB increments starting at 0MB
2221 * 0x11 to 0x16: 4MB increments starting at 8MB
2222 * 0x17 to 0x1d: 4MB increments start at 36MB
2224 if (gmch_ctrl < 0x11)
2225 return gmch_ctrl << 25;
2226 else if (gmch_ctrl < 0x17)
2227 return (gmch_ctrl - 0x11 + 2) << 22;
2229 return (gmch_ctrl - 0x17 + 9) << 22;
2232 static size_t gen9_get_stolen_size(u16 gen9_gmch_ctl)
2234 gen9_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2235 gen9_gmch_ctl &= BDW_GMCH_GMS_MASK;
2237 if (gen9_gmch_ctl < 0xf0)
2238 return gen9_gmch_ctl << 25; /* 32 MB units */
2240 /* 4MB increments starting at 0xf0 for 4MB */
2241 return (gen9_gmch_ctl - 0xf0 + 1) << 22;
2244 static int ggtt_probe_common(struct drm_device *dev,
2247 struct drm_i915_private *dev_priv = dev->dev_private;
2248 phys_addr_t gtt_phys_addr;
2251 /* For Modern GENs the PTEs and register space are split in the BAR */
2252 gtt_phys_addr = pci_resource_start(dev->pdev, 0) +
2253 (pci_resource_len(dev->pdev, 0) / 2);
2256 * On BXT writes larger than 64 bit to the GTT pagetable range will be
2257 * dropped. For WC mappings in general we have 64 byte burst writes
2258 * when the WC buffer is flushed, so we can't use it, but have to
2259 * resort to an uncached mapping. The WC issue is easily caught by the
2260 * readback check when writing GTT PTE entries.
2262 if (IS_BROXTON(dev))
2263 dev_priv->gtt.gsm = ioremap_nocache(gtt_phys_addr, gtt_size);
2265 dev_priv->gtt.gsm = ioremap_wc(gtt_phys_addr, gtt_size);
2266 if (!dev_priv->gtt.gsm) {
2267 DRM_ERROR("Failed to map the gtt page table\n");
2271 ret = alloc_scratch_page(&dev_priv->gtt.base);
2273 DRM_ERROR("Scratch setup failed\n");
2274 /* iounmap will also get called at remove, but meh */
2275 iounmap(dev_priv->gtt.gsm);
2281 /* The GGTT and PPGTT need a private PPAT setup in order to handle cacheability
2282 * bits. When using advanced contexts each context stores its own PAT, but
2283 * writing this data shouldn't be harmful even in those cases. */
2284 static void bdw_setup_private_ppat(struct drm_i915_private *dev_priv)
2288 pat = GEN8_PPAT(0, GEN8_PPAT_WB | GEN8_PPAT_LLC) | /* for normal objects, no eLLC */
2289 GEN8_PPAT(1, GEN8_PPAT_WC | GEN8_PPAT_LLCELLC) | /* for something pointing to ptes? */
2290 GEN8_PPAT(2, GEN8_PPAT_WT | GEN8_PPAT_LLCELLC) | /* for scanout with eLLC */
2291 GEN8_PPAT(3, GEN8_PPAT_UC) | /* Uncached objects, mostly for scanout */
2292 GEN8_PPAT(4, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(0)) |
2293 GEN8_PPAT(5, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(1)) |
2294 GEN8_PPAT(6, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(2)) |
2295 GEN8_PPAT(7, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3));
2297 if (!USES_PPGTT(dev_priv->dev))
2298 /* Spec: "For GGTT, there is NO pat_sel[2:0] from the entry,
2299 * so RTL will always use the value corresponding to
2301 * So let's disable cache for GGTT to avoid screen corruptions.
2302 * MOCS still can be used though.
2303 * - System agent ggtt writes (i.e. cpu gtt mmaps) already work
2304 * before this patch, i.e. the same uncached + snooping access
2305 * like on gen6/7 seems to be in effect.
2306 * - So this just fixes blitter/render access. Again it looks
2307 * like it's not just uncached access, but uncached + snooping.
2308 * So we can still hold onto all our assumptions wrt cpu
2309 * clflushing on LLC machines.
2311 pat = GEN8_PPAT(0, GEN8_PPAT_UC);
2313 /* XXX: spec defines this as 2 distinct registers. It's unclear if a 64b
2314 * write would work. */
2315 I915_WRITE(GEN8_PRIVATE_PAT, pat);
2316 I915_WRITE(GEN8_PRIVATE_PAT + 4, pat >> 32);
2319 static void chv_setup_private_ppat(struct drm_i915_private *dev_priv)
2324 * Map WB on BDW to snooped on CHV.
2326 * Only the snoop bit has meaning for CHV, the rest is
2329 * The hardware will never snoop for certain types of accesses:
2330 * - CPU GTT (GMADR->GGTT->no snoop->memory)
2331 * - PPGTT page tables
2332 * - some other special cycles
2334 * As with BDW, we also need to consider the following for GT accesses:
2335 * "For GGTT, there is NO pat_sel[2:0] from the entry,
2336 * so RTL will always use the value corresponding to
2338 * Which means we must set the snoop bit in PAT entry 0
2339 * in order to keep the global status page working.
2341 pat = GEN8_PPAT(0, CHV_PPAT_SNOOP) |
2345 GEN8_PPAT(4, CHV_PPAT_SNOOP) |
2346 GEN8_PPAT(5, CHV_PPAT_SNOOP) |
2347 GEN8_PPAT(6, CHV_PPAT_SNOOP) |
2348 GEN8_PPAT(7, CHV_PPAT_SNOOP);
2350 I915_WRITE(GEN8_PRIVATE_PAT, pat);
2351 I915_WRITE(GEN8_PRIVATE_PAT + 4, pat >> 32);
2354 static int gen8_gmch_probe(struct drm_device *dev,
2357 phys_addr_t *mappable_base,
2360 struct drm_i915_private *dev_priv = dev->dev_private;
2365 /* TODO: We're not aware of mappable constraints on gen8 yet */
2366 *mappable_base = pci_resource_start(dev->pdev, 2);
2367 *mappable_end = pci_resource_len(dev->pdev, 2);
2369 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(39)))
2370 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(39));
2372 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
2374 if (INTEL_INFO(dev)->gen >= 9) {
2375 *stolen = gen9_get_stolen_size(snb_gmch_ctl);
2376 gtt_size = gen8_get_total_gtt_size(snb_gmch_ctl);
2377 } else if (IS_CHERRYVIEW(dev)) {
2378 *stolen = chv_get_stolen_size(snb_gmch_ctl);
2379 gtt_size = chv_get_total_gtt_size(snb_gmch_ctl);
2381 *stolen = gen8_get_stolen_size(snb_gmch_ctl);
2382 gtt_size = gen8_get_total_gtt_size(snb_gmch_ctl);
2385 *gtt_total = (gtt_size / sizeof(gen8_pte_t)) << PAGE_SHIFT;
2387 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
2388 chv_setup_private_ppat(dev_priv);
2390 bdw_setup_private_ppat(dev_priv);
2392 ret = ggtt_probe_common(dev, gtt_size);
2394 dev_priv->gtt.base.clear_range = gen8_ggtt_clear_range;
2395 dev_priv->gtt.base.insert_entries = gen8_ggtt_insert_entries;
2396 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
2397 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
2402 static int gen6_gmch_probe(struct drm_device *dev,
2405 phys_addr_t *mappable_base,
2408 struct drm_i915_private *dev_priv = dev->dev_private;
2409 unsigned int gtt_size;
2413 *mappable_base = pci_resource_start(dev->pdev, 2);
2414 *mappable_end = pci_resource_len(dev->pdev, 2);
2416 /* 64/512MB is the current min/max we actually know of, but this is just
2417 * a coarse sanity check.
2419 if ((*mappable_end < (64<<20) || (*mappable_end > (512<<20)))) {
2420 DRM_ERROR("Unknown GMADR size (%llx)\n",
2421 dev_priv->gtt.mappable_end);
2425 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(40)))
2426 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(40));
2427 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
2429 *stolen = gen6_get_stolen_size(snb_gmch_ctl);
2431 gtt_size = gen6_get_total_gtt_size(snb_gmch_ctl);
2432 *gtt_total = (gtt_size / sizeof(gen6_pte_t)) << PAGE_SHIFT;
2434 ret = ggtt_probe_common(dev, gtt_size);
2436 dev_priv->gtt.base.clear_range = gen6_ggtt_clear_range;
2437 dev_priv->gtt.base.insert_entries = gen6_ggtt_insert_entries;
2438 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
2439 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
2444 static void gen6_gmch_remove(struct i915_address_space *vm)
2447 struct i915_gtt *gtt = container_of(vm, struct i915_gtt, base);
2450 free_scratch_page(vm);
2453 static int i915_gmch_probe(struct drm_device *dev,
2456 phys_addr_t *mappable_base,
2459 struct drm_i915_private *dev_priv = dev->dev_private;
2462 ret = intel_gmch_probe(dev_priv->bridge_dev, dev_priv->dev->pdev, NULL);
2464 DRM_ERROR("failed to set up gmch\n");
2468 intel_gtt_get(gtt_total, stolen, mappable_base, mappable_end);
2470 dev_priv->gtt.do_idle_maps = needs_idle_maps(dev_priv->dev);
2471 dev_priv->gtt.base.insert_entries = i915_ggtt_insert_entries;
2472 dev_priv->gtt.base.clear_range = i915_ggtt_clear_range;
2473 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
2474 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
2476 if (unlikely(dev_priv->gtt.do_idle_maps))
2477 DRM_INFO("applying Ironlake quirks for intel_iommu\n");
2482 static void i915_gmch_remove(struct i915_address_space *vm)
2484 intel_gmch_remove();
2487 int i915_gem_gtt_init(struct drm_device *dev)
2489 struct drm_i915_private *dev_priv = dev->dev_private;
2490 struct i915_gtt *gtt = &dev_priv->gtt;
2493 if (INTEL_INFO(dev)->gen <= 5) {
2494 gtt->gtt_probe = i915_gmch_probe;
2495 gtt->base.cleanup = i915_gmch_remove;
2496 } else if (INTEL_INFO(dev)->gen < 8) {
2497 gtt->gtt_probe = gen6_gmch_probe;
2498 gtt->base.cleanup = gen6_gmch_remove;
2499 if (IS_HASWELL(dev) && dev_priv->ellc_size)
2500 gtt->base.pte_encode = iris_pte_encode;
2501 else if (IS_HASWELL(dev))
2502 gtt->base.pte_encode = hsw_pte_encode;
2503 else if (IS_VALLEYVIEW(dev))
2504 gtt->base.pte_encode = byt_pte_encode;
2505 else if (INTEL_INFO(dev)->gen >= 7)
2506 gtt->base.pte_encode = ivb_pte_encode;
2508 gtt->base.pte_encode = snb_pte_encode;
2510 dev_priv->gtt.gtt_probe = gen8_gmch_probe;
2511 dev_priv->gtt.base.cleanup = gen6_gmch_remove;
2514 gtt->base.dev = dev;
2516 ret = gtt->gtt_probe(dev, >t->base.total, >t->stolen_size,
2517 >t->mappable_base, >t->mappable_end);
2521 /* GMADR is the PCI mmio aperture into the global GTT. */
2522 DRM_INFO("Memory usable by graphics device = %lluM\n",
2523 gtt->base.total >> 20);
2524 DRM_DEBUG_DRIVER("GMADR size = %lldM\n", gtt->mappable_end >> 20);
2525 DRM_DEBUG_DRIVER("GTT stolen size = %zdM\n", gtt->stolen_size >> 20);
2526 #ifdef CONFIG_INTEL_IOMMU
2527 if (intel_iommu_gfx_mapped)
2528 DRM_INFO("VT-d active for gfx access\n");
2531 * i915.enable_ppgtt is read-only, so do an early pass to validate the
2532 * user's requested state against the hardware/driver capabilities. We
2533 * do this now so that we can print out any log messages once rather
2534 * than every time we check intel_enable_ppgtt().
2536 i915.enable_ppgtt = sanitize_enable_ppgtt(dev, i915.enable_ppgtt);
2537 DRM_DEBUG_DRIVER("ppgtt mode: %i\n", i915.enable_ppgtt);
2542 void i915_gem_restore_gtt_mappings(struct drm_device *dev)
2544 struct drm_i915_private *dev_priv = dev->dev_private;
2545 struct drm_i915_gem_object *obj;
2546 struct i915_address_space *vm;
2548 i915_check_and_clear_faults(dev);
2550 /* First fill our portion of the GTT with scratch pages */
2551 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
2552 dev_priv->gtt.base.start,
2553 dev_priv->gtt.base.total,
2556 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
2557 struct i915_vma *vma = i915_gem_obj_to_vma(obj,
2558 &dev_priv->gtt.base);
2562 i915_gem_clflush_object(obj, obj->pin_display);
2563 WARN_ON(i915_vma_bind(vma, obj->cache_level, PIN_UPDATE));
2567 if (INTEL_INFO(dev)->gen >= 8) {
2568 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
2569 chv_setup_private_ppat(dev_priv);
2571 bdw_setup_private_ppat(dev_priv);
2576 if (USES_PPGTT(dev)) {
2577 list_for_each_entry(vm, &dev_priv->vm_list, global_link) {
2578 /* TODO: Perhaps it shouldn't be gen6 specific */
2580 struct i915_hw_ppgtt *ppgtt =
2581 container_of(vm, struct i915_hw_ppgtt,
2584 if (i915_is_ggtt(vm))
2585 ppgtt = dev_priv->mm.aliasing_ppgtt;
2587 gen6_write_page_range(dev_priv, &ppgtt->pd,
2588 0, ppgtt->base.total);
2592 i915_ggtt_flush(dev_priv);
2595 static struct i915_vma *
2596 __i915_gem_vma_create(struct drm_i915_gem_object *obj,
2597 struct i915_address_space *vm,
2598 const struct i915_ggtt_view *ggtt_view)
2600 struct i915_vma *vma;
2602 if (WARN_ON(i915_is_ggtt(vm) != !!ggtt_view))
2603 return ERR_PTR(-EINVAL);
2605 vma = kmem_cache_zalloc(to_i915(obj->base.dev)->vmas, GFP_KERNEL);
2607 return ERR_PTR(-ENOMEM);
2609 INIT_LIST_HEAD(&vma->vma_link);
2610 INIT_LIST_HEAD(&vma->mm_list);
2611 INIT_LIST_HEAD(&vma->exec_list);
2615 if (i915_is_ggtt(vm))
2616 vma->ggtt_view = *ggtt_view;
2618 list_add_tail(&vma->vma_link, &obj->vma_list);
2619 if (!i915_is_ggtt(vm))
2620 i915_ppgtt_get(i915_vm_to_ppgtt(vm));
2626 i915_gem_obj_lookup_or_create_vma(struct drm_i915_gem_object *obj,
2627 struct i915_address_space *vm)
2629 struct i915_vma *vma;
2631 vma = i915_gem_obj_to_vma(obj, vm);
2633 vma = __i915_gem_vma_create(obj, vm,
2634 i915_is_ggtt(vm) ? &i915_ggtt_view_normal : NULL);
2640 i915_gem_obj_lookup_or_create_ggtt_vma(struct drm_i915_gem_object *obj,
2641 const struct i915_ggtt_view *view)
2643 struct i915_address_space *ggtt = i915_obj_to_ggtt(obj);
2644 struct i915_vma *vma;
2647 return ERR_PTR(-EINVAL);
2649 vma = i915_gem_obj_to_ggtt_view(obj, view);
2655 vma = __i915_gem_vma_create(obj, ggtt, view);
2662 rotate_pages(dma_addr_t *in, unsigned int width, unsigned int height,
2663 struct sg_table *st)
2665 unsigned int column, row;
2666 unsigned int src_idx;
2667 struct scatterlist *sg = st->sgl;
2671 for (column = 0; column < width; column++) {
2672 src_idx = width * (height - 1) + column;
2673 for (row = 0; row < height; row++) {
2675 /* We don't need the pages, but need to initialize
2676 * the entries so the sg list can be happily traversed.
2677 * The only thing we need are DMA addresses.
2679 sg_set_page(sg, NULL, PAGE_SIZE, 0);
2680 sg_dma_address(sg) = in[src_idx];
2681 sg_dma_len(sg) = PAGE_SIZE;
2688 static struct sg_table *
2689 intel_rotate_fb_obj_pages(struct i915_ggtt_view *ggtt_view,
2690 struct drm_i915_gem_object *obj)
2692 struct intel_rotation_info *rot_info = &ggtt_view->rotation_info;
2693 unsigned int size_pages = rot_info->size >> PAGE_SHIFT;
2694 struct sg_page_iter sg_iter;
2696 dma_addr_t *page_addr_list;
2697 struct sg_table *st;
2700 /* Allocate a temporary list of source pages for random access. */
2701 page_addr_list = drm_malloc_ab(obj->base.size / PAGE_SIZE,
2702 sizeof(dma_addr_t));
2703 if (!page_addr_list)
2704 return ERR_PTR(ret);
2706 /* Allocate target SG list. */
2707 st = kmalloc(sizeof(*st), GFP_KERNEL);
2711 ret = sg_alloc_table(st, size_pages, GFP_KERNEL);
2715 /* Populate source page list from the object. */
2717 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) {
2718 page_addr_list[i] = sg_page_iter_dma_address(&sg_iter);
2722 /* Rotate the pages. */
2723 rotate_pages(page_addr_list,
2724 rot_info->width_pages, rot_info->height_pages,
2728 "Created rotated page mapping for object size %zu (pitch=%u, height=%u, pixel_format=0x%x, %ux%u tiles, %u pages).\n",
2729 obj->base.size, rot_info->pitch, rot_info->height,
2730 rot_info->pixel_format, rot_info->width_pages,
2731 rot_info->height_pages, size_pages);
2733 drm_free_large(page_addr_list);
2740 drm_free_large(page_addr_list);
2743 "Failed to create rotated mapping for object size %zu! (%d) (pitch=%u, height=%u, pixel_format=0x%x, %ux%u tiles, %u pages)\n",
2744 obj->base.size, ret, rot_info->pitch, rot_info->height,
2745 rot_info->pixel_format, rot_info->width_pages,
2746 rot_info->height_pages, size_pages);
2747 return ERR_PTR(ret);
2750 static struct sg_table *
2751 intel_partial_pages(const struct i915_ggtt_view *view,
2752 struct drm_i915_gem_object *obj)
2754 struct sg_table *st;
2755 struct scatterlist *sg;
2756 struct sg_page_iter obj_sg_iter;
2759 st = kmalloc(sizeof(*st), GFP_KERNEL);
2763 ret = sg_alloc_table(st, view->params.partial.size, GFP_KERNEL);
2769 for_each_sg_page(obj->pages->sgl, &obj_sg_iter, obj->pages->nents,
2770 view->params.partial.offset)
2772 if (st->nents >= view->params.partial.size)
2775 sg_set_page(sg, NULL, PAGE_SIZE, 0);
2776 sg_dma_address(sg) = sg_page_iter_dma_address(&obj_sg_iter);
2777 sg_dma_len(sg) = PAGE_SIZE;
2788 return ERR_PTR(ret);
2792 i915_get_ggtt_vma_pages(struct i915_vma *vma)
2796 if (vma->ggtt_view.pages)
2799 if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL)
2800 vma->ggtt_view.pages = vma->obj->pages;
2801 else if (vma->ggtt_view.type == I915_GGTT_VIEW_ROTATED)
2802 vma->ggtt_view.pages =
2803 intel_rotate_fb_obj_pages(&vma->ggtt_view, vma->obj);
2804 else if (vma->ggtt_view.type == I915_GGTT_VIEW_PARTIAL)
2805 vma->ggtt_view.pages =
2806 intel_partial_pages(&vma->ggtt_view, vma->obj);
2808 WARN_ONCE(1, "GGTT view %u not implemented!\n",
2809 vma->ggtt_view.type);
2811 if (!vma->ggtt_view.pages) {
2812 DRM_ERROR("Failed to get pages for GGTT view type %u!\n",
2813 vma->ggtt_view.type);
2815 } else if (IS_ERR(vma->ggtt_view.pages)) {
2816 ret = PTR_ERR(vma->ggtt_view.pages);
2817 vma->ggtt_view.pages = NULL;
2818 DRM_ERROR("Failed to get pages for VMA view type %u (%d)!\n",
2819 vma->ggtt_view.type, ret);
2826 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
2828 * @cache_level: mapping cache level
2829 * @flags: flags like global or local mapping
2831 * DMA addresses are taken from the scatter-gather table of this object (or of
2832 * this VMA in case of non-default GGTT views) and PTE entries set up.
2833 * Note that DMA addresses are also the only part of the SG table we care about.
2835 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
2841 if (WARN_ON(flags == 0))
2845 if (flags & PIN_GLOBAL)
2846 bind_flags |= GLOBAL_BIND;
2847 if (flags & PIN_USER)
2848 bind_flags |= LOCAL_BIND;
2850 if (flags & PIN_UPDATE)
2851 bind_flags |= vma->bound;
2853 bind_flags &= ~vma->bound;
2855 if (bind_flags == 0)
2858 if (vma->bound == 0 && vma->vm->allocate_va_range) {
2859 trace_i915_va_alloc(vma->vm,
2862 VM_TO_TRACE_NAME(vma->vm));
2864 /* XXX: i915_vma_pin() will fix this +- hack */
2866 ret = vma->vm->allocate_va_range(vma->vm,
2874 ret = vma->vm->bind_vma(vma, cache_level, bind_flags);
2878 vma->bound |= bind_flags;
2884 * i915_ggtt_view_size - Get the size of a GGTT view.
2885 * @obj: Object the view is of.
2886 * @view: The view in question.
2888 * @return The size of the GGTT view in bytes.
2891 i915_ggtt_view_size(struct drm_i915_gem_object *obj,
2892 const struct i915_ggtt_view *view)
2894 if (view->type == I915_GGTT_VIEW_NORMAL) {
2895 return obj->base.size;
2896 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
2897 return view->rotation_info.size;
2898 } else if (view->type == I915_GGTT_VIEW_PARTIAL) {
2899 return view->params.partial.size << PAGE_SHIFT;
2901 WARN_ONCE(1, "GGTT view %u not implemented!\n", view->type);
2902 return obj->base.size;