2 * Copyright © 2010 Daniel Vetter
3 * Copyright © 2011-2014 Intel Corporation
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
26 #include <linux/seq_file.h>
28 #include <drm/i915_drm.h>
30 #include "i915_vgpu.h"
31 #include "i915_trace.h"
32 #include "intel_drv.h"
35 * DOC: Global GTT views
37 * Background and previous state
39 * Historically objects could exists (be bound) in global GTT space only as
40 * singular instances with a view representing all of the object's backing pages
41 * in a linear fashion. This view will be called a normal view.
43 * To support multiple views of the same object, where the number of mapped
44 * pages is not equal to the backing store, or where the layout of the pages
45 * is not linear, concept of a GGTT view was added.
47 * One example of an alternative view is a stereo display driven by a single
48 * image. In this case we would have a framebuffer looking like this
54 * Above would represent a normal GGTT view as normally mapped for GPU or CPU
55 * rendering. In contrast, fed to the display engine would be an alternative
56 * view which could look something like this:
61 * In this example both the size and layout of pages in the alternative view is
62 * different from the normal view.
64 * Implementation and usage
66 * GGTT views are implemented using VMAs and are distinguished via enum
67 * i915_ggtt_view_type and struct i915_ggtt_view.
69 * A new flavour of core GEM functions which work with GGTT bound objects were
70 * added with the _ggtt_ infix, and sometimes with _view postfix to avoid
71 * renaming in large amounts of code. They take the struct i915_ggtt_view
72 * parameter encapsulating all metadata required to implement a view.
74 * As a helper for callers which are only interested in the normal view,
75 * globally const i915_ggtt_view_normal singleton instance exists. All old core
76 * GEM API functions, the ones not taking the view parameter, are operating on,
77 * or with the normal GGTT view.
79 * Code wanting to add or use a new GGTT view needs to:
81 * 1. Add a new enum with a suitable name.
82 * 2. Extend the metadata in the i915_ggtt_view structure if required.
83 * 3. Add support to i915_get_vma_pages().
85 * New views are required to build a scatter-gather table from within the
86 * i915_get_vma_pages function. This table is stored in the vma.ggtt_view and
87 * exists for the lifetime of an VMA.
89 * Core API is designed to have copy semantics which means that passed in
90 * struct i915_ggtt_view does not need to be persistent (left around after
91 * calling the core API functions).
96 i915_get_ggtt_vma_pages(struct i915_vma *vma);
98 const struct i915_ggtt_view i915_ggtt_view_normal;
99 const struct i915_ggtt_view i915_ggtt_view_rotated = {
100 .type = I915_GGTT_VIEW_ROTATED
103 static int sanitize_enable_ppgtt(struct drm_device *dev, int enable_ppgtt)
105 bool has_aliasing_ppgtt;
108 has_aliasing_ppgtt = INTEL_INFO(dev)->gen >= 6;
109 has_full_ppgtt = INTEL_INFO(dev)->gen >= 7;
111 if (intel_vgpu_active(dev))
112 has_full_ppgtt = false; /* emulation is too hard */
115 * We don't allow disabling PPGTT for gen9+ as it's a requirement for
116 * execlists, the sole mechanism available to submit work.
118 if (INTEL_INFO(dev)->gen < 9 &&
119 (enable_ppgtt == 0 || !has_aliasing_ppgtt))
122 if (enable_ppgtt == 1)
125 if (enable_ppgtt == 2 && has_full_ppgtt)
128 #ifdef CONFIG_INTEL_IOMMU
129 /* Disable ppgtt on SNB if VT-d is on. */
130 if (INTEL_INFO(dev)->gen == 6 && intel_iommu_gfx_mapped) {
131 DRM_INFO("Disabling PPGTT because VT-d is on\n");
136 /* Early VLV doesn't have this */
137 if (IS_VALLEYVIEW(dev) && !IS_CHERRYVIEW(dev) &&
138 dev->pdev->revision < 0xb) {
139 DRM_DEBUG_DRIVER("disabling PPGTT on pre-B3 step VLV\n");
143 if (INTEL_INFO(dev)->gen >= 8 && i915.enable_execlists)
146 return has_aliasing_ppgtt ? 1 : 0;
149 static int ppgtt_bind_vma(struct i915_vma *vma,
150 enum i915_cache_level cache_level,
155 /* Currently applicable only to VLV */
157 pte_flags |= PTE_READ_ONLY;
159 vma->vm->insert_entries(vma->vm, vma->obj->pages, vma->node.start,
160 cache_level, pte_flags);
165 static void ppgtt_unbind_vma(struct i915_vma *vma)
167 vma->vm->clear_range(vma->vm,
173 static gen8_pte_t gen8_pte_encode(dma_addr_t addr,
174 enum i915_cache_level level,
177 gen8_pte_t pte = valid ? _PAGE_PRESENT | _PAGE_RW : 0;
181 case I915_CACHE_NONE:
182 pte |= PPAT_UNCACHED_INDEX;
185 pte |= PPAT_DISPLAY_ELLC_INDEX;
188 pte |= PPAT_CACHED_INDEX;
195 static gen8_pde_t gen8_pde_encode(struct drm_device *dev,
197 enum i915_cache_level level)
199 gen8_pde_t pde = _PAGE_PRESENT | _PAGE_RW;
201 if (level != I915_CACHE_NONE)
202 pde |= PPAT_CACHED_PDE_INDEX;
204 pde |= PPAT_UNCACHED_INDEX;
208 static gen6_pte_t snb_pte_encode(dma_addr_t addr,
209 enum i915_cache_level level,
210 bool valid, u32 unused)
212 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
213 pte |= GEN6_PTE_ADDR_ENCODE(addr);
216 case I915_CACHE_L3_LLC:
218 pte |= GEN6_PTE_CACHE_LLC;
220 case I915_CACHE_NONE:
221 pte |= GEN6_PTE_UNCACHED;
230 static gen6_pte_t ivb_pte_encode(dma_addr_t addr,
231 enum i915_cache_level level,
232 bool valid, u32 unused)
234 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
235 pte |= GEN6_PTE_ADDR_ENCODE(addr);
238 case I915_CACHE_L3_LLC:
239 pte |= GEN7_PTE_CACHE_L3_LLC;
242 pte |= GEN6_PTE_CACHE_LLC;
244 case I915_CACHE_NONE:
245 pte |= GEN6_PTE_UNCACHED;
254 static gen6_pte_t byt_pte_encode(dma_addr_t addr,
255 enum i915_cache_level level,
256 bool valid, u32 flags)
258 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
259 pte |= GEN6_PTE_ADDR_ENCODE(addr);
261 if (!(flags & PTE_READ_ONLY))
262 pte |= BYT_PTE_WRITEABLE;
264 if (level != I915_CACHE_NONE)
265 pte |= BYT_PTE_SNOOPED_BY_CPU_CACHES;
270 static gen6_pte_t hsw_pte_encode(dma_addr_t addr,
271 enum i915_cache_level level,
272 bool valid, u32 unused)
274 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
275 pte |= HSW_PTE_ADDR_ENCODE(addr);
277 if (level != I915_CACHE_NONE)
278 pte |= HSW_WB_LLC_AGE3;
283 static gen6_pte_t iris_pte_encode(dma_addr_t addr,
284 enum i915_cache_level level,
285 bool valid, u32 unused)
287 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
288 pte |= HSW_PTE_ADDR_ENCODE(addr);
291 case I915_CACHE_NONE:
294 pte |= HSW_WT_ELLC_LLC_AGE3;
297 pte |= HSW_WB_ELLC_LLC_AGE3;
304 static int setup_page_dma(struct drm_device *dev, struct i915_page_dma *p)
306 struct device *device = &dev->pdev->dev;
308 p->page = alloc_page(GFP_KERNEL);
312 p->daddr = dma_map_page(device,
313 p->page, 0, 4096, PCI_DMA_BIDIRECTIONAL);
315 if (dma_mapping_error(device, p->daddr)) {
316 __free_page(p->page);
323 static void cleanup_page_dma(struct drm_device *dev, struct i915_page_dma *p)
325 if (WARN_ON(!p->page))
328 dma_unmap_page(&dev->pdev->dev, p->daddr, 4096, PCI_DMA_BIDIRECTIONAL);
329 __free_page(p->page);
330 memset(p, 0, sizeof(*p));
333 static void *kmap_page_dma(struct i915_page_dma *p)
335 return kmap_atomic(p->page);
338 /* We use the flushing unmap only with ppgtt structures:
339 * page directories, page tables and scratch pages.
341 static void kunmap_page_dma(struct drm_device *dev, void *vaddr)
343 /* There are only few exceptions for gen >=6. chv and bxt.
344 * And we are not sure about the latter so play safe for now.
346 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
347 drm_clflush_virt_range(vaddr, PAGE_SIZE);
349 kunmap_atomic(vaddr);
352 #define kmap_px(px) kmap_page_dma(px_base(px))
353 #define kunmap_px(ppgtt, vaddr) kunmap_page_dma((ppgtt)->base.dev, (vaddr))
355 #define setup_px(dev, px) setup_page_dma((dev), px_base(px))
356 #define cleanup_px(dev, px) cleanup_page_dma((dev), px_base(px))
357 #define fill_px(dev, px, v) fill_page_dma((dev), px_base(px), (v))
358 #define fill32_px(dev, px, v) fill_page_dma_32((dev), px_base(px), (v))
360 static void fill_page_dma(struct drm_device *dev, struct i915_page_dma *p,
364 uint64_t * const vaddr = kmap_page_dma(p);
366 for (i = 0; i < 512; i++)
369 kunmap_page_dma(dev, vaddr);
372 static void fill_page_dma_32(struct drm_device *dev, struct i915_page_dma *p,
373 const uint32_t val32)
379 fill_page_dma(dev, p, v);
382 static void free_pt(struct drm_device *dev, struct i915_page_table *pt)
385 kfree(pt->used_ptes);
389 static void gen8_initialize_pt(struct i915_address_space *vm,
390 struct i915_page_table *pt)
392 gen8_pte_t scratch_pte;
394 scratch_pte = gen8_pte_encode(vm->scratch.addr, I915_CACHE_LLC, true);
396 fill_px(vm->dev, pt, scratch_pte);
399 static struct i915_page_table *alloc_pt(struct drm_device *dev)
401 struct i915_page_table *pt;
402 const size_t count = INTEL_INFO(dev)->gen >= 8 ?
403 GEN8_PTES : GEN6_PTES;
406 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
408 return ERR_PTR(-ENOMEM);
410 pt->used_ptes = kcalloc(BITS_TO_LONGS(count), sizeof(*pt->used_ptes),
416 ret = setup_px(dev, pt);
423 kfree(pt->used_ptes);
430 static void free_pd(struct drm_device *dev, struct i915_page_directory *pd)
434 kfree(pd->used_pdes);
439 static struct i915_page_directory *alloc_pd(struct drm_device *dev)
441 struct i915_page_directory *pd;
444 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
446 return ERR_PTR(-ENOMEM);
448 pd->used_pdes = kcalloc(BITS_TO_LONGS(I915_PDES),
449 sizeof(*pd->used_pdes), GFP_KERNEL);
453 ret = setup_px(dev, pd);
460 kfree(pd->used_pdes);
467 /* Broadwell Page Directory Pointer Descriptors */
468 static int gen8_write_pdp(struct drm_i915_gem_request *req,
472 struct intel_engine_cs *ring = req->ring;
477 ret = intel_ring_begin(req, 6);
481 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
482 intel_ring_emit(ring, GEN8_RING_PDP_UDW(ring, entry));
483 intel_ring_emit(ring, upper_32_bits(addr));
484 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
485 intel_ring_emit(ring, GEN8_RING_PDP_LDW(ring, entry));
486 intel_ring_emit(ring, lower_32_bits(addr));
487 intel_ring_advance(ring);
492 static int gen8_mm_switch(struct i915_hw_ppgtt *ppgtt,
493 struct drm_i915_gem_request *req)
497 for (i = GEN8_LEGACY_PDPES - 1; i >= 0; i--) {
498 const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i);
500 ret = gen8_write_pdp(req, i, pd_daddr);
508 static void gen8_ppgtt_clear_range(struct i915_address_space *vm,
513 struct i915_hw_ppgtt *ppgtt =
514 container_of(vm, struct i915_hw_ppgtt, base);
515 gen8_pte_t *pt_vaddr, scratch_pte;
516 unsigned pdpe = start >> GEN8_PDPE_SHIFT & GEN8_PDPE_MASK;
517 unsigned pde = start >> GEN8_PDE_SHIFT & GEN8_PDE_MASK;
518 unsigned pte = start >> GEN8_PTE_SHIFT & GEN8_PTE_MASK;
519 unsigned num_entries = length >> PAGE_SHIFT;
520 unsigned last_pte, i;
522 scratch_pte = gen8_pte_encode(ppgtt->base.scratch.addr,
523 I915_CACHE_LLC, use_scratch);
525 while (num_entries) {
526 struct i915_page_directory *pd;
527 struct i915_page_table *pt;
529 if (WARN_ON(!ppgtt->pdp.page_directory[pdpe]))
532 pd = ppgtt->pdp.page_directory[pdpe];
534 if (WARN_ON(!pd->page_table[pde]))
537 pt = pd->page_table[pde];
539 if (WARN_ON(!px_page(pt)))
542 last_pte = pte + num_entries;
543 if (last_pte > GEN8_PTES)
544 last_pte = GEN8_PTES;
546 pt_vaddr = kmap_px(pt);
548 for (i = pte; i < last_pte; i++) {
549 pt_vaddr[i] = scratch_pte;
553 kunmap_px(ppgtt, pt);
556 if (++pde == I915_PDES) {
563 static void gen8_ppgtt_insert_entries(struct i915_address_space *vm,
564 struct sg_table *pages,
566 enum i915_cache_level cache_level, u32 unused)
568 struct i915_hw_ppgtt *ppgtt =
569 container_of(vm, struct i915_hw_ppgtt, base);
570 gen8_pte_t *pt_vaddr;
571 unsigned pdpe = start >> GEN8_PDPE_SHIFT & GEN8_PDPE_MASK;
572 unsigned pde = start >> GEN8_PDE_SHIFT & GEN8_PDE_MASK;
573 unsigned pte = start >> GEN8_PTE_SHIFT & GEN8_PTE_MASK;
574 struct sg_page_iter sg_iter;
578 for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) {
579 if (WARN_ON(pdpe >= GEN8_LEGACY_PDPES))
582 if (pt_vaddr == NULL) {
583 struct i915_page_directory *pd = ppgtt->pdp.page_directory[pdpe];
584 struct i915_page_table *pt = pd->page_table[pde];
585 pt_vaddr = kmap_px(pt);
589 gen8_pte_encode(sg_page_iter_dma_address(&sg_iter),
591 if (++pte == GEN8_PTES) {
592 kunmap_px(ppgtt, pt_vaddr);
594 if (++pde == I915_PDES) {
603 kunmap_px(ppgtt, pt_vaddr);
606 static void __gen8_do_map_pt(gen8_pde_t * const pde,
607 struct i915_page_table *pt,
608 struct drm_device *dev)
611 gen8_pde_encode(dev, px_dma(pt), I915_CACHE_LLC);
615 static void gen8_initialize_pd(struct i915_address_space *vm,
616 struct i915_page_directory *pd)
618 struct i915_hw_ppgtt *ppgtt =
619 container_of(vm, struct i915_hw_ppgtt, base);
620 gen8_pde_t scratch_pde;
622 scratch_pde = gen8_pde_encode(vm->dev, px_dma(ppgtt->scratch_pt),
625 fill_px(vm->dev, pd, scratch_pde);
628 static void gen8_free_page_tables(struct i915_page_directory *pd, struct drm_device *dev)
635 for_each_set_bit(i, pd->used_pdes, I915_PDES) {
636 if (WARN_ON(!pd->page_table[i]))
639 free_pt(dev, pd->page_table[i]);
640 pd->page_table[i] = NULL;
644 static void gen8_ppgtt_cleanup(struct i915_address_space *vm)
646 struct i915_hw_ppgtt *ppgtt =
647 container_of(vm, struct i915_hw_ppgtt, base);
650 for_each_set_bit(i, ppgtt->pdp.used_pdpes, GEN8_LEGACY_PDPES) {
651 if (WARN_ON(!ppgtt->pdp.page_directory[i]))
654 gen8_free_page_tables(ppgtt->pdp.page_directory[i], ppgtt->base.dev);
655 free_pd(ppgtt->base.dev, ppgtt->pdp.page_directory[i]);
658 free_pd(ppgtt->base.dev, ppgtt->scratch_pd);
659 free_pt(ppgtt->base.dev, ppgtt->scratch_pt);
663 * gen8_ppgtt_alloc_pagetabs() - Allocate page tables for VA range.
664 * @ppgtt: Master ppgtt structure.
665 * @pd: Page directory for this address range.
666 * @start: Starting virtual address to begin allocations.
667 * @length Size of the allocations.
668 * @new_pts: Bitmap set by function with new allocations. Likely used by the
669 * caller to free on error.
671 * Allocate the required number of page tables. Extremely similar to
672 * gen8_ppgtt_alloc_page_directories(). The main difference is here we are limited by
673 * the page directory boundary (instead of the page directory pointer). That
674 * boundary is 1GB virtual. Therefore, unlike gen8_ppgtt_alloc_page_directories(), it is
675 * possible, and likely that the caller will need to use multiple calls of this
676 * function to achieve the appropriate allocation.
678 * Return: 0 if success; negative error code otherwise.
680 static int gen8_ppgtt_alloc_pagetabs(struct i915_hw_ppgtt *ppgtt,
681 struct i915_page_directory *pd,
684 unsigned long *new_pts)
686 struct drm_device *dev = ppgtt->base.dev;
687 struct i915_page_table *pt;
691 gen8_for_each_pde(pt, pd, start, length, temp, pde) {
692 /* Don't reallocate page tables */
694 /* Scratch is never allocated this way */
695 WARN_ON(pt == ppgtt->scratch_pt);
703 gen8_initialize_pt(&ppgtt->base, pt);
704 pd->page_table[pde] = pt;
705 set_bit(pde, new_pts);
711 for_each_set_bit(pde, new_pts, I915_PDES)
712 free_pt(dev, pd->page_table[pde]);
718 * gen8_ppgtt_alloc_page_directories() - Allocate page directories for VA range.
719 * @ppgtt: Master ppgtt structure.
720 * @pdp: Page directory pointer for this address range.
721 * @start: Starting virtual address to begin allocations.
722 * @length Size of the allocations.
723 * @new_pds Bitmap set by function with new allocations. Likely used by the
724 * caller to free on error.
726 * Allocate the required number of page directories starting at the pde index of
727 * @start, and ending at the pde index @start + @length. This function will skip
728 * over already allocated page directories within the range, and only allocate
729 * new ones, setting the appropriate pointer within the pdp as well as the
730 * correct position in the bitmap @new_pds.
732 * The function will only allocate the pages within the range for a give page
733 * directory pointer. In other words, if @start + @length straddles a virtually
734 * addressed PDP boundary (512GB for 4k pages), there will be more allocations
735 * required by the caller, This is not currently possible, and the BUG in the
736 * code will prevent it.
738 * Return: 0 if success; negative error code otherwise.
740 static int gen8_ppgtt_alloc_page_directories(struct i915_hw_ppgtt *ppgtt,
741 struct i915_page_directory_pointer *pdp,
744 unsigned long *new_pds)
746 struct drm_device *dev = ppgtt->base.dev;
747 struct i915_page_directory *pd;
751 WARN_ON(!bitmap_empty(new_pds, GEN8_LEGACY_PDPES));
753 gen8_for_each_pdpe(pd, pdp, start, length, temp, pdpe) {
761 gen8_initialize_pd(&ppgtt->base, pd);
762 pdp->page_directory[pdpe] = pd;
763 set_bit(pdpe, new_pds);
769 for_each_set_bit(pdpe, new_pds, GEN8_LEGACY_PDPES)
770 free_pd(dev, pdp->page_directory[pdpe]);
776 free_gen8_temp_bitmaps(unsigned long *new_pds, unsigned long **new_pts)
780 for (i = 0; i < GEN8_LEGACY_PDPES; i++)
786 /* Fills in the page directory bitmap, and the array of page tables bitmap. Both
787 * of these are based on the number of PDPEs in the system.
790 int __must_check alloc_gen8_temp_bitmaps(unsigned long **new_pds,
791 unsigned long ***new_pts)
797 pds = kcalloc(BITS_TO_LONGS(GEN8_LEGACY_PDPES), sizeof(unsigned long), GFP_KERNEL);
801 pts = kcalloc(GEN8_LEGACY_PDPES, sizeof(unsigned long *), GFP_KERNEL);
807 for (i = 0; i < GEN8_LEGACY_PDPES; i++) {
808 pts[i] = kcalloc(BITS_TO_LONGS(I915_PDES),
809 sizeof(unsigned long), GFP_KERNEL);
820 free_gen8_temp_bitmaps(pds, pts);
824 /* PDE TLBs are a pain to invalidate on GEN8+. When we modify
825 * the page table structures, we mark them dirty so that
826 * context switching/execlist queuing code takes extra steps
827 * to ensure that tlbs are flushed.
829 static void mark_tlbs_dirty(struct i915_hw_ppgtt *ppgtt)
831 ppgtt->pd_dirty_rings = INTEL_INFO(ppgtt->base.dev)->ring_mask;
834 static int gen8_alloc_va_range(struct i915_address_space *vm,
838 struct i915_hw_ppgtt *ppgtt =
839 container_of(vm, struct i915_hw_ppgtt, base);
840 unsigned long *new_page_dirs, **new_page_tables;
841 struct i915_page_directory *pd;
842 const uint64_t orig_start = start;
843 const uint64_t orig_length = length;
848 /* Wrap is never okay since we can only represent 48b, and we don't
849 * actually use the other side of the canonical address space.
851 if (WARN_ON(start + length < start))
854 if (WARN_ON(start + length > ppgtt->base.total))
857 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables);
861 /* Do the allocations first so we can easily bail out */
862 ret = gen8_ppgtt_alloc_page_directories(ppgtt, &ppgtt->pdp, start, length,
865 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
869 /* For every page directory referenced, allocate page tables */
870 gen8_for_each_pdpe(pd, &ppgtt->pdp, start, length, temp, pdpe) {
871 ret = gen8_ppgtt_alloc_pagetabs(ppgtt, pd, start, length,
872 new_page_tables[pdpe]);
878 length = orig_length;
880 /* Allocations have completed successfully, so set the bitmaps, and do
882 gen8_for_each_pdpe(pd, &ppgtt->pdp, start, length, temp, pdpe) {
883 gen8_pde_t *const page_directory = kmap_px(pd);
884 struct i915_page_table *pt;
885 uint64_t pd_len = gen8_clamp_pd(start, length);
886 uint64_t pd_start = start;
889 /* Every pd should be allocated, we just did that above. */
892 gen8_for_each_pde(pt, pd, pd_start, pd_len, temp, pde) {
893 /* Same reasoning as pd */
896 WARN_ON(!gen8_pte_count(pd_start, pd_len));
898 /* Set our used ptes within the page table */
899 bitmap_set(pt->used_ptes,
900 gen8_pte_index(pd_start),
901 gen8_pte_count(pd_start, pd_len));
903 /* Our pde is now pointing to the pagetable, pt */
904 set_bit(pde, pd->used_pdes);
906 /* Map the PDE to the page table */
907 __gen8_do_map_pt(page_directory + pde, pt, vm->dev);
909 /* NB: We haven't yet mapped ptes to pages. At this
910 * point we're still relying on insert_entries() */
913 kunmap_px(ppgtt, page_directory);
915 set_bit(pdpe, ppgtt->pdp.used_pdpes);
918 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
919 mark_tlbs_dirty(ppgtt);
924 for_each_set_bit(temp, new_page_tables[pdpe], I915_PDES)
925 free_pt(vm->dev, ppgtt->pdp.page_directory[pdpe]->page_table[temp]);
928 for_each_set_bit(pdpe, new_page_dirs, GEN8_LEGACY_PDPES)
929 free_pd(vm->dev, ppgtt->pdp.page_directory[pdpe]);
931 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
932 mark_tlbs_dirty(ppgtt);
937 * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers
938 * with a net effect resembling a 2-level page table in normal x86 terms. Each
939 * PDP represents 1GB of memory 4 * 512 * 512 * 4096 = 4GB legacy 32b address
943 static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
945 ppgtt->scratch_pt = alloc_pt(ppgtt->base.dev);
946 if (IS_ERR(ppgtt->scratch_pt))
947 return PTR_ERR(ppgtt->scratch_pt);
949 ppgtt->scratch_pd = alloc_pd(ppgtt->base.dev);
950 if (IS_ERR(ppgtt->scratch_pd))
951 return PTR_ERR(ppgtt->scratch_pd);
953 gen8_initialize_pt(&ppgtt->base, ppgtt->scratch_pt);
954 gen8_initialize_pd(&ppgtt->base, ppgtt->scratch_pd);
956 ppgtt->base.start = 0;
957 ppgtt->base.total = 1ULL << 32;
958 if (IS_ENABLED(CONFIG_X86_32))
959 /* While we have a proliferation of size_t variables
960 * we cannot represent the full ppgtt size on 32bit,
961 * so limit it to the same size as the GGTT (currently
964 ppgtt->base.total = to_i915(ppgtt->base.dev)->gtt.base.total;
965 ppgtt->base.cleanup = gen8_ppgtt_cleanup;
966 ppgtt->base.allocate_va_range = gen8_alloc_va_range;
967 ppgtt->base.insert_entries = gen8_ppgtt_insert_entries;
968 ppgtt->base.clear_range = gen8_ppgtt_clear_range;
969 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
970 ppgtt->base.bind_vma = ppgtt_bind_vma;
972 ppgtt->switch_mm = gen8_mm_switch;
977 static void gen6_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
979 struct i915_address_space *vm = &ppgtt->base;
980 struct i915_page_table *unused;
981 gen6_pte_t scratch_pte;
983 uint32_t pte, pde, temp;
984 uint32_t start = ppgtt->base.start, length = ppgtt->base.total;
986 scratch_pte = vm->pte_encode(vm->scratch.addr, I915_CACHE_LLC, true, 0);
988 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde) {
990 gen6_pte_t *pt_vaddr;
991 const dma_addr_t pt_addr = px_dma(ppgtt->pd.page_table[pde]);
992 pd_entry = readl(ppgtt->pd_addr + pde);
993 expected = (GEN6_PDE_ADDR_ENCODE(pt_addr) | GEN6_PDE_VALID);
995 if (pd_entry != expected)
996 seq_printf(m, "\tPDE #%d mismatch: Actual PDE: %x Expected PDE: %x\n",
1000 seq_printf(m, "\tPDE: %x\n", pd_entry);
1002 pt_vaddr = kmap_px(ppgtt->pd.page_table[pde]);
1004 for (pte = 0; pte < GEN6_PTES; pte+=4) {
1006 (pde * PAGE_SIZE * GEN6_PTES) +
1010 for (i = 0; i < 4; i++)
1011 if (pt_vaddr[pte + i] != scratch_pte)
1016 seq_printf(m, "\t\t0x%lx [%03d,%04d]: =", va, pde, pte);
1017 for (i = 0; i < 4; i++) {
1018 if (pt_vaddr[pte + i] != scratch_pte)
1019 seq_printf(m, " %08x", pt_vaddr[pte + i]);
1021 seq_puts(m, " SCRATCH ");
1025 kunmap_px(ppgtt, pt_vaddr);
1029 /* Write pde (index) from the page directory @pd to the page table @pt */
1030 static void gen6_write_pde(struct i915_page_directory *pd,
1031 const int pde, struct i915_page_table *pt)
1033 /* Caller needs to make sure the write completes if necessary */
1034 struct i915_hw_ppgtt *ppgtt =
1035 container_of(pd, struct i915_hw_ppgtt, pd);
1038 pd_entry = GEN6_PDE_ADDR_ENCODE(px_dma(pt));
1039 pd_entry |= GEN6_PDE_VALID;
1041 writel(pd_entry, ppgtt->pd_addr + pde);
1044 /* Write all the page tables found in the ppgtt structure to incrementing page
1046 static void gen6_write_page_range(struct drm_i915_private *dev_priv,
1047 struct i915_page_directory *pd,
1048 uint32_t start, uint32_t length)
1050 struct i915_page_table *pt;
1053 gen6_for_each_pde(pt, pd, start, length, temp, pde)
1054 gen6_write_pde(pd, pde, pt);
1056 /* Make sure write is complete before other code can use this page
1057 * table. Also require for WC mapped PTEs */
1058 readl(dev_priv->gtt.gsm);
1061 static uint32_t get_pd_offset(struct i915_hw_ppgtt *ppgtt)
1063 BUG_ON(ppgtt->pd.base.ggtt_offset & 0x3f);
1065 return (ppgtt->pd.base.ggtt_offset / 64) << 16;
1068 static int hsw_mm_switch(struct i915_hw_ppgtt *ppgtt,
1069 struct drm_i915_gem_request *req)
1071 struct intel_engine_cs *ring = req->ring;
1074 /* NB: TLBs must be flushed and invalidated before a switch */
1075 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1079 ret = intel_ring_begin(req, 6);
1083 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1084 intel_ring_emit(ring, RING_PP_DIR_DCLV(ring));
1085 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1086 intel_ring_emit(ring, RING_PP_DIR_BASE(ring));
1087 intel_ring_emit(ring, get_pd_offset(ppgtt));
1088 intel_ring_emit(ring, MI_NOOP);
1089 intel_ring_advance(ring);
1094 static int vgpu_mm_switch(struct i915_hw_ppgtt *ppgtt,
1095 struct drm_i915_gem_request *req)
1097 struct intel_engine_cs *ring = req->ring;
1098 struct drm_i915_private *dev_priv = to_i915(ppgtt->base.dev);
1100 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
1101 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
1105 static int gen7_mm_switch(struct i915_hw_ppgtt *ppgtt,
1106 struct drm_i915_gem_request *req)
1108 struct intel_engine_cs *ring = req->ring;
1111 /* NB: TLBs must be flushed and invalidated before a switch */
1112 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1116 ret = intel_ring_begin(req, 6);
1120 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1121 intel_ring_emit(ring, RING_PP_DIR_DCLV(ring));
1122 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1123 intel_ring_emit(ring, RING_PP_DIR_BASE(ring));
1124 intel_ring_emit(ring, get_pd_offset(ppgtt));
1125 intel_ring_emit(ring, MI_NOOP);
1126 intel_ring_advance(ring);
1128 /* XXX: RCS is the only one to auto invalidate the TLBs? */
1129 if (ring->id != RCS) {
1130 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1138 static int gen6_mm_switch(struct i915_hw_ppgtt *ppgtt,
1139 struct drm_i915_gem_request *req)
1141 struct intel_engine_cs *ring = req->ring;
1142 struct drm_device *dev = ppgtt->base.dev;
1143 struct drm_i915_private *dev_priv = dev->dev_private;
1146 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
1147 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
1149 POSTING_READ(RING_PP_DIR_DCLV(ring));
1154 static void gen8_ppgtt_enable(struct drm_device *dev)
1156 struct drm_i915_private *dev_priv = dev->dev_private;
1157 struct intel_engine_cs *ring;
1160 for_each_ring(ring, dev_priv, j) {
1161 I915_WRITE(RING_MODE_GEN7(ring),
1162 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1166 static void gen7_ppgtt_enable(struct drm_device *dev)
1168 struct drm_i915_private *dev_priv = dev->dev_private;
1169 struct intel_engine_cs *ring;
1170 uint32_t ecochk, ecobits;
1173 ecobits = I915_READ(GAC_ECO_BITS);
1174 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B);
1176 ecochk = I915_READ(GAM_ECOCHK);
1177 if (IS_HASWELL(dev)) {
1178 ecochk |= ECOCHK_PPGTT_WB_HSW;
1180 ecochk |= ECOCHK_PPGTT_LLC_IVB;
1181 ecochk &= ~ECOCHK_PPGTT_GFDT_IVB;
1183 I915_WRITE(GAM_ECOCHK, ecochk);
1185 for_each_ring(ring, dev_priv, i) {
1186 /* GFX_MODE is per-ring on gen7+ */
1187 I915_WRITE(RING_MODE_GEN7(ring),
1188 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1192 static void gen6_ppgtt_enable(struct drm_device *dev)
1194 struct drm_i915_private *dev_priv = dev->dev_private;
1195 uint32_t ecochk, gab_ctl, ecobits;
1197 ecobits = I915_READ(GAC_ECO_BITS);
1198 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_SNB_BIT |
1199 ECOBITS_PPGTT_CACHE64B);
1201 gab_ctl = I915_READ(GAB_CTL);
1202 I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT);
1204 ecochk = I915_READ(GAM_ECOCHK);
1205 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT | ECOCHK_PPGTT_CACHE64B);
1207 I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1210 /* PPGTT support for Sandybdrige/Gen6 and later */
1211 static void gen6_ppgtt_clear_range(struct i915_address_space *vm,
1216 struct i915_hw_ppgtt *ppgtt =
1217 container_of(vm, struct i915_hw_ppgtt, base);
1218 gen6_pte_t *pt_vaddr, scratch_pte;
1219 unsigned first_entry = start >> PAGE_SHIFT;
1220 unsigned num_entries = length >> PAGE_SHIFT;
1221 unsigned act_pt = first_entry / GEN6_PTES;
1222 unsigned first_pte = first_entry % GEN6_PTES;
1223 unsigned last_pte, i;
1225 scratch_pte = vm->pte_encode(vm->scratch.addr, I915_CACHE_LLC, true, 0);
1227 while (num_entries) {
1228 last_pte = first_pte + num_entries;
1229 if (last_pte > GEN6_PTES)
1230 last_pte = GEN6_PTES;
1232 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1234 for (i = first_pte; i < last_pte; i++)
1235 pt_vaddr[i] = scratch_pte;
1237 kunmap_px(ppgtt, pt_vaddr);
1239 num_entries -= last_pte - first_pte;
1245 static void gen6_ppgtt_insert_entries(struct i915_address_space *vm,
1246 struct sg_table *pages,
1248 enum i915_cache_level cache_level, u32 flags)
1250 struct i915_hw_ppgtt *ppgtt =
1251 container_of(vm, struct i915_hw_ppgtt, base);
1252 gen6_pte_t *pt_vaddr;
1253 unsigned first_entry = start >> PAGE_SHIFT;
1254 unsigned act_pt = first_entry / GEN6_PTES;
1255 unsigned act_pte = first_entry % GEN6_PTES;
1256 struct sg_page_iter sg_iter;
1259 for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) {
1260 if (pt_vaddr == NULL)
1261 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1264 vm->pte_encode(sg_page_iter_dma_address(&sg_iter),
1265 cache_level, true, flags);
1267 if (++act_pte == GEN6_PTES) {
1268 kunmap_px(ppgtt, pt_vaddr);
1275 kunmap_px(ppgtt, pt_vaddr);
1278 static void gen6_initialize_pt(struct i915_address_space *vm,
1279 struct i915_page_table *pt)
1281 gen6_pte_t scratch_pte;
1283 WARN_ON(vm->scratch.addr == 0);
1285 scratch_pte = vm->pte_encode(vm->scratch.addr, 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 = dev_priv->gtt.base.scratch;
1524 if (INTEL_INFO(dev)->gen < 8)
1525 return gen6_ppgtt_init(ppgtt);
1527 return gen8_ppgtt_init(ppgtt);
1529 int i915_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
1531 struct drm_i915_private *dev_priv = dev->dev_private;
1534 ret = __hw_ppgtt_init(dev, ppgtt);
1536 kref_init(&ppgtt->ref);
1537 drm_mm_init(&ppgtt->base.mm, ppgtt->base.start,
1539 i915_init_vm(dev_priv, &ppgtt->base);
1545 int i915_ppgtt_init_hw(struct drm_device *dev)
1547 /* In the case of execlists, PPGTT is enabled by the context descriptor
1548 * and the PDPs are contained within the context itself. We don't
1549 * need to do anything here. */
1550 if (i915.enable_execlists)
1553 if (!USES_PPGTT(dev))
1557 gen6_ppgtt_enable(dev);
1558 else if (IS_GEN7(dev))
1559 gen7_ppgtt_enable(dev);
1560 else if (INTEL_INFO(dev)->gen >= 8)
1561 gen8_ppgtt_enable(dev);
1563 MISSING_CASE(INTEL_INFO(dev)->gen);
1568 int i915_ppgtt_init_ring(struct drm_i915_gem_request *req)
1570 struct drm_i915_private *dev_priv = req->ring->dev->dev_private;
1571 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
1573 if (i915.enable_execlists)
1579 return ppgtt->switch_mm(ppgtt, req);
1582 struct i915_hw_ppgtt *
1583 i915_ppgtt_create(struct drm_device *dev, struct drm_i915_file_private *fpriv)
1585 struct i915_hw_ppgtt *ppgtt;
1588 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
1590 return ERR_PTR(-ENOMEM);
1592 ret = i915_ppgtt_init(dev, ppgtt);
1595 return ERR_PTR(ret);
1598 ppgtt->file_priv = fpriv;
1600 trace_i915_ppgtt_create(&ppgtt->base);
1605 void i915_ppgtt_release(struct kref *kref)
1607 struct i915_hw_ppgtt *ppgtt =
1608 container_of(kref, struct i915_hw_ppgtt, ref);
1610 trace_i915_ppgtt_release(&ppgtt->base);
1612 /* vmas should already be unbound */
1613 WARN_ON(!list_empty(&ppgtt->base.active_list));
1614 WARN_ON(!list_empty(&ppgtt->base.inactive_list));
1616 list_del(&ppgtt->base.global_link);
1617 drm_mm_takedown(&ppgtt->base.mm);
1619 ppgtt->base.cleanup(&ppgtt->base);
1623 extern int intel_iommu_gfx_mapped;
1624 /* Certain Gen5 chipsets require require idling the GPU before
1625 * unmapping anything from the GTT when VT-d is enabled.
1627 static bool needs_idle_maps(struct drm_device *dev)
1629 #ifdef CONFIG_INTEL_IOMMU
1630 /* Query intel_iommu to see if we need the workaround. Presumably that
1633 if (IS_GEN5(dev) && IS_MOBILE(dev) && intel_iommu_gfx_mapped)
1639 static bool do_idling(struct drm_i915_private *dev_priv)
1641 bool ret = dev_priv->mm.interruptible;
1643 if (unlikely(dev_priv->gtt.do_idle_maps)) {
1644 dev_priv->mm.interruptible = false;
1645 if (i915_gpu_idle(dev_priv->dev)) {
1646 DRM_ERROR("Couldn't idle GPU\n");
1647 /* Wait a bit, in hopes it avoids the hang */
1655 static void undo_idling(struct drm_i915_private *dev_priv, bool interruptible)
1657 if (unlikely(dev_priv->gtt.do_idle_maps))
1658 dev_priv->mm.interruptible = interruptible;
1661 void i915_check_and_clear_faults(struct drm_device *dev)
1663 struct drm_i915_private *dev_priv = dev->dev_private;
1664 struct intel_engine_cs *ring;
1667 if (INTEL_INFO(dev)->gen < 6)
1670 for_each_ring(ring, dev_priv, i) {
1672 fault_reg = I915_READ(RING_FAULT_REG(ring));
1673 if (fault_reg & RING_FAULT_VALID) {
1674 DRM_DEBUG_DRIVER("Unexpected fault\n"
1676 "\tAddress space: %s\n"
1679 fault_reg & PAGE_MASK,
1680 fault_reg & RING_FAULT_GTTSEL_MASK ? "GGTT" : "PPGTT",
1681 RING_FAULT_SRCID(fault_reg),
1682 RING_FAULT_FAULT_TYPE(fault_reg));
1683 I915_WRITE(RING_FAULT_REG(ring),
1684 fault_reg & ~RING_FAULT_VALID);
1687 POSTING_READ(RING_FAULT_REG(&dev_priv->ring[RCS]));
1690 static void i915_ggtt_flush(struct drm_i915_private *dev_priv)
1692 if (INTEL_INFO(dev_priv->dev)->gen < 6) {
1693 intel_gtt_chipset_flush();
1695 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
1696 POSTING_READ(GFX_FLSH_CNTL_GEN6);
1700 void i915_gem_suspend_gtt_mappings(struct drm_device *dev)
1702 struct drm_i915_private *dev_priv = dev->dev_private;
1704 /* Don't bother messing with faults pre GEN6 as we have little
1705 * documentation supporting that it's a good idea.
1707 if (INTEL_INFO(dev)->gen < 6)
1710 i915_check_and_clear_faults(dev);
1712 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
1713 dev_priv->gtt.base.start,
1714 dev_priv->gtt.base.total,
1717 i915_ggtt_flush(dev_priv);
1720 int i915_gem_gtt_prepare_object(struct drm_i915_gem_object *obj)
1722 if (obj->has_dma_mapping)
1725 if (!dma_map_sg(&obj->base.dev->pdev->dev,
1726 obj->pages->sgl, obj->pages->nents,
1727 PCI_DMA_BIDIRECTIONAL))
1733 static void gen8_set_pte(void __iomem *addr, gen8_pte_t pte)
1738 iowrite32((u32)pte, addr);
1739 iowrite32(pte >> 32, addr + 4);
1743 static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
1744 struct sg_table *st,
1746 enum i915_cache_level level, u32 unused)
1748 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1749 unsigned first_entry = start >> PAGE_SHIFT;
1750 gen8_pte_t __iomem *gtt_entries =
1751 (gen8_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
1753 struct sg_page_iter sg_iter;
1754 dma_addr_t addr = 0; /* shut up gcc */
1756 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
1757 addr = sg_dma_address(sg_iter.sg) +
1758 (sg_iter.sg_pgoffset << PAGE_SHIFT);
1759 gen8_set_pte(>t_entries[i],
1760 gen8_pte_encode(addr, level, true));
1765 * XXX: This serves as a posting read to make sure that the PTE has
1766 * actually been updated. There is some concern that even though
1767 * registers and PTEs are within the same BAR that they are potentially
1768 * of NUMA access patterns. Therefore, even with the way we assume
1769 * hardware should work, we must keep this posting read for paranoia.
1772 WARN_ON(readq(>t_entries[i-1])
1773 != gen8_pte_encode(addr, level, true));
1775 /* This next bit makes the above posting read even more important. We
1776 * want to flush the TLBs only after we're certain all the PTE updates
1779 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
1780 POSTING_READ(GFX_FLSH_CNTL_GEN6);
1784 * Binds an object into the global gtt with the specified cache level. The object
1785 * will be accessible to the GPU via commands whose operands reference offsets
1786 * within the global GTT as well as accessible by the GPU through the GMADR
1787 * mapped BAR (dev_priv->mm.gtt->gtt).
1789 static void gen6_ggtt_insert_entries(struct i915_address_space *vm,
1790 struct sg_table *st,
1792 enum i915_cache_level level, u32 flags)
1794 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1795 unsigned first_entry = start >> PAGE_SHIFT;
1796 gen6_pte_t __iomem *gtt_entries =
1797 (gen6_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
1799 struct sg_page_iter sg_iter;
1800 dma_addr_t addr = 0;
1802 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
1803 addr = sg_page_iter_dma_address(&sg_iter);
1804 iowrite32(vm->pte_encode(addr, level, true, flags), >t_entries[i]);
1808 /* XXX: This serves as a posting read to make sure that the PTE has
1809 * actually been updated. There is some concern that even though
1810 * registers and PTEs are within the same BAR that they are potentially
1811 * of NUMA access patterns. Therefore, even with the way we assume
1812 * hardware should work, we must keep this posting read for paranoia.
1815 unsigned long gtt = readl(>t_entries[i-1]);
1816 WARN_ON(gtt != vm->pte_encode(addr, level, true, flags));
1819 /* This next bit makes the above posting read even more important. We
1820 * want to flush the TLBs only after we're certain all the PTE updates
1823 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
1824 POSTING_READ(GFX_FLSH_CNTL_GEN6);
1827 static void gen8_ggtt_clear_range(struct i915_address_space *vm,
1832 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1833 unsigned first_entry = start >> PAGE_SHIFT;
1834 unsigned num_entries = length >> PAGE_SHIFT;
1835 gen8_pte_t scratch_pte, __iomem *gtt_base =
1836 (gen8_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
1837 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
1840 if (WARN(num_entries > max_entries,
1841 "First entry = %d; Num entries = %d (max=%d)\n",
1842 first_entry, num_entries, max_entries))
1843 num_entries = max_entries;
1845 scratch_pte = gen8_pte_encode(vm->scratch.addr,
1848 for (i = 0; i < num_entries; i++)
1849 gen8_set_pte(>t_base[i], scratch_pte);
1853 static void gen6_ggtt_clear_range(struct i915_address_space *vm,
1858 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1859 unsigned first_entry = start >> PAGE_SHIFT;
1860 unsigned num_entries = length >> PAGE_SHIFT;
1861 gen6_pte_t scratch_pte, __iomem *gtt_base =
1862 (gen6_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
1863 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
1866 if (WARN(num_entries > max_entries,
1867 "First entry = %d; Num entries = %d (max=%d)\n",
1868 first_entry, num_entries, max_entries))
1869 num_entries = max_entries;
1871 scratch_pte = vm->pte_encode(vm->scratch.addr, I915_CACHE_LLC, use_scratch, 0);
1873 for (i = 0; i < num_entries; i++)
1874 iowrite32(scratch_pte, >t_base[i]);
1878 static void i915_ggtt_insert_entries(struct i915_address_space *vm,
1879 struct sg_table *pages,
1881 enum i915_cache_level cache_level, u32 unused)
1883 unsigned int flags = (cache_level == I915_CACHE_NONE) ?
1884 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
1886 intel_gtt_insert_sg_entries(pages, start >> PAGE_SHIFT, flags);
1890 static void i915_ggtt_clear_range(struct i915_address_space *vm,
1895 unsigned first_entry = start >> PAGE_SHIFT;
1896 unsigned num_entries = length >> PAGE_SHIFT;
1897 intel_gtt_clear_range(first_entry, num_entries);
1900 static int ggtt_bind_vma(struct i915_vma *vma,
1901 enum i915_cache_level cache_level,
1904 struct drm_device *dev = vma->vm->dev;
1905 struct drm_i915_private *dev_priv = dev->dev_private;
1906 struct drm_i915_gem_object *obj = vma->obj;
1907 struct sg_table *pages = obj->pages;
1911 ret = i915_get_ggtt_vma_pages(vma);
1914 pages = vma->ggtt_view.pages;
1916 /* Currently applicable only to VLV */
1918 pte_flags |= PTE_READ_ONLY;
1921 if (!dev_priv->mm.aliasing_ppgtt || flags & GLOBAL_BIND) {
1922 vma->vm->insert_entries(vma->vm, pages,
1924 cache_level, pte_flags);
1927 if (dev_priv->mm.aliasing_ppgtt && flags & LOCAL_BIND) {
1928 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
1929 appgtt->base.insert_entries(&appgtt->base, pages,
1931 cache_level, pte_flags);
1937 static void ggtt_unbind_vma(struct i915_vma *vma)
1939 struct drm_device *dev = vma->vm->dev;
1940 struct drm_i915_private *dev_priv = dev->dev_private;
1941 struct drm_i915_gem_object *obj = vma->obj;
1942 const uint64_t size = min_t(uint64_t,
1946 if (vma->bound & GLOBAL_BIND) {
1947 vma->vm->clear_range(vma->vm,
1953 if (dev_priv->mm.aliasing_ppgtt && vma->bound & LOCAL_BIND) {
1954 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
1956 appgtt->base.clear_range(&appgtt->base,
1963 void i915_gem_gtt_finish_object(struct drm_i915_gem_object *obj)
1965 struct drm_device *dev = obj->base.dev;
1966 struct drm_i915_private *dev_priv = dev->dev_private;
1969 interruptible = do_idling(dev_priv);
1971 if (!obj->has_dma_mapping)
1972 dma_unmap_sg(&dev->pdev->dev,
1973 obj->pages->sgl, obj->pages->nents,
1974 PCI_DMA_BIDIRECTIONAL);
1976 undo_idling(dev_priv, interruptible);
1979 static void i915_gtt_color_adjust(struct drm_mm_node *node,
1980 unsigned long color,
1984 if (node->color != color)
1987 if (!list_empty(&node->node_list)) {
1988 node = list_entry(node->node_list.next,
1991 if (node->allocated && node->color != color)
1996 static int i915_gem_setup_global_gtt(struct drm_device *dev,
1997 unsigned long start,
1998 unsigned long mappable_end,
2001 /* Let GEM Manage all of the aperture.
2003 * However, leave one page at the end still bound to the scratch page.
2004 * There are a number of places where the hardware apparently prefetches
2005 * past the end of the object, and we've seen multiple hangs with the
2006 * GPU head pointer stuck in a batchbuffer bound at the last page of the
2007 * aperture. One page should be enough to keep any prefetching inside
2010 struct drm_i915_private *dev_priv = dev->dev_private;
2011 struct i915_address_space *ggtt_vm = &dev_priv->gtt.base;
2012 struct drm_mm_node *entry;
2013 struct drm_i915_gem_object *obj;
2014 unsigned long hole_start, hole_end;
2017 BUG_ON(mappable_end > end);
2019 /* Subtract the guard page ... */
2020 drm_mm_init(&ggtt_vm->mm, start, end - start - PAGE_SIZE);
2022 dev_priv->gtt.base.start = start;
2023 dev_priv->gtt.base.total = end - start;
2025 if (intel_vgpu_active(dev)) {
2026 ret = intel_vgt_balloon(dev);
2032 dev_priv->gtt.base.mm.color_adjust = i915_gtt_color_adjust;
2034 /* Mark any preallocated objects as occupied */
2035 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
2036 struct i915_vma *vma = i915_gem_obj_to_vma(obj, ggtt_vm);
2038 DRM_DEBUG_KMS("reserving preallocated space: %lx + %zx\n",
2039 i915_gem_obj_ggtt_offset(obj), obj->base.size);
2041 WARN_ON(i915_gem_obj_ggtt_bound(obj));
2042 ret = drm_mm_reserve_node(&ggtt_vm->mm, &vma->node);
2044 DRM_DEBUG_KMS("Reservation failed: %i\n", ret);
2047 vma->bound |= GLOBAL_BIND;
2050 /* Clear any non-preallocated blocks */
2051 drm_mm_for_each_hole(entry, &ggtt_vm->mm, hole_start, hole_end) {
2052 DRM_DEBUG_KMS("clearing unused GTT space: [%lx, %lx]\n",
2053 hole_start, hole_end);
2054 ggtt_vm->clear_range(ggtt_vm, hole_start,
2055 hole_end - hole_start, true);
2058 /* And finally clear the reserved guard page */
2059 ggtt_vm->clear_range(ggtt_vm, end - PAGE_SIZE, PAGE_SIZE, true);
2061 if (USES_PPGTT(dev) && !USES_FULL_PPGTT(dev)) {
2062 struct i915_hw_ppgtt *ppgtt;
2064 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2068 ret = __hw_ppgtt_init(dev, ppgtt);
2070 ppgtt->base.cleanup(&ppgtt->base);
2075 if (ppgtt->base.allocate_va_range)
2076 ret = ppgtt->base.allocate_va_range(&ppgtt->base, 0,
2079 ppgtt->base.cleanup(&ppgtt->base);
2084 ppgtt->base.clear_range(&ppgtt->base,
2089 dev_priv->mm.aliasing_ppgtt = ppgtt;
2095 void i915_gem_init_global_gtt(struct drm_device *dev)
2097 struct drm_i915_private *dev_priv = dev->dev_private;
2098 u64 gtt_size, mappable_size;
2100 gtt_size = dev_priv->gtt.base.total;
2101 mappable_size = dev_priv->gtt.mappable_end;
2103 i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
2106 void i915_global_gtt_cleanup(struct drm_device *dev)
2108 struct drm_i915_private *dev_priv = dev->dev_private;
2109 struct i915_address_space *vm = &dev_priv->gtt.base;
2111 if (dev_priv->mm.aliasing_ppgtt) {
2112 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
2114 ppgtt->base.cleanup(&ppgtt->base);
2117 if (drm_mm_initialized(&vm->mm)) {
2118 if (intel_vgpu_active(dev))
2119 intel_vgt_deballoon();
2121 drm_mm_takedown(&vm->mm);
2122 list_del(&vm->global_link);
2128 static int setup_scratch_page(struct drm_device *dev)
2130 struct drm_i915_private *dev_priv = dev->dev_private;
2132 dma_addr_t dma_addr;
2134 page = alloc_page(GFP_KERNEL | GFP_DMA32 | __GFP_ZERO);
2137 set_pages_uc(page, 1);
2139 #ifdef CONFIG_INTEL_IOMMU
2140 dma_addr = pci_map_page(dev->pdev, page, 0, PAGE_SIZE,
2141 PCI_DMA_BIDIRECTIONAL);
2142 if (pci_dma_mapping_error(dev->pdev, dma_addr)) {
2147 dma_addr = page_to_phys(page);
2149 dev_priv->gtt.base.scratch.page = page;
2150 dev_priv->gtt.base.scratch.addr = dma_addr;
2155 static void teardown_scratch_page(struct drm_device *dev)
2157 struct drm_i915_private *dev_priv = dev->dev_private;
2158 struct page *page = dev_priv->gtt.base.scratch.page;
2160 set_pages_wb(page, 1);
2161 pci_unmap_page(dev->pdev, dev_priv->gtt.base.scratch.addr,
2162 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
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 = setup_scratch_page(dev);
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 teardown_scratch_page(vm->dev);
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 ret = gtt->gtt_probe(dev, >t->base.total, >t->stolen_size,
2515 >t->mappable_base, >t->mappable_end);
2519 gtt->base.dev = dev;
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 ret = vma->vm->allocate_va_range(vma->vm,
2871 ret = vma->vm->bind_vma(vma, cache_level, bind_flags);
2875 vma->bound |= bind_flags;
2881 * i915_ggtt_view_size - Get the size of a GGTT view.
2882 * @obj: Object the view is of.
2883 * @view: The view in question.
2885 * @return The size of the GGTT view in bytes.
2888 i915_ggtt_view_size(struct drm_i915_gem_object *obj,
2889 const struct i915_ggtt_view *view)
2891 if (view->type == I915_GGTT_VIEW_NORMAL) {
2892 return obj->base.size;
2893 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
2894 return view->rotation_info.size;
2895 } else if (view->type == I915_GGTT_VIEW_PARTIAL) {
2896 return view->params.partial.size << PAGE_SHIFT;
2898 WARN_ONCE(1, "GGTT view %u not implemented!\n", view->type);
2899 return obj->base.size;