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 struct i915_page_scratch *alloc_scratch_page(struct drm_device *dev)
389 struct i915_page_scratch *sp;
392 sp = kzalloc(sizeof(*sp), GFP_KERNEL);
394 return ERR_PTR(-ENOMEM);
396 ret = __setup_page_dma(dev, px_base(sp), GFP_DMA32 | __GFP_ZERO);
402 set_pages_uc(px_page(sp), 1);
407 static void free_scratch_page(struct drm_device *dev,
408 struct i915_page_scratch *sp)
410 set_pages_wb(px_page(sp), 1);
416 static struct i915_page_table *alloc_pt(struct drm_device *dev)
418 struct i915_page_table *pt;
419 const size_t count = INTEL_INFO(dev)->gen >= 8 ?
420 GEN8_PTES : GEN6_PTES;
423 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
425 return ERR_PTR(-ENOMEM);
427 pt->used_ptes = kcalloc(BITS_TO_LONGS(count), sizeof(*pt->used_ptes),
433 ret = setup_px(dev, pt);
440 kfree(pt->used_ptes);
447 static void free_pt(struct drm_device *dev, struct i915_page_table *pt)
450 kfree(pt->used_ptes);
454 static void gen8_initialize_pt(struct i915_address_space *vm,
455 struct i915_page_table *pt)
457 gen8_pte_t scratch_pte;
459 scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
460 I915_CACHE_LLC, true);
462 fill_px(vm->dev, pt, scratch_pte);
465 static void gen6_initialize_pt(struct i915_address_space *vm,
466 struct i915_page_table *pt)
468 gen6_pte_t scratch_pte;
470 WARN_ON(px_dma(vm->scratch_page) == 0);
472 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
473 I915_CACHE_LLC, true, 0);
475 fill32_px(vm->dev, pt, scratch_pte);
478 static struct i915_page_directory *alloc_pd(struct drm_device *dev)
480 struct i915_page_directory *pd;
483 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
485 return ERR_PTR(-ENOMEM);
487 pd->used_pdes = kcalloc(BITS_TO_LONGS(I915_PDES),
488 sizeof(*pd->used_pdes), GFP_KERNEL);
492 ret = setup_px(dev, pd);
499 kfree(pd->used_pdes);
506 static void free_pd(struct drm_device *dev, struct i915_page_directory *pd)
510 kfree(pd->used_pdes);
515 static void gen8_initialize_pd(struct i915_address_space *vm,
516 struct i915_page_directory *pd)
518 gen8_pde_t scratch_pde;
520 scratch_pde = gen8_pde_encode(px_dma(vm->scratch_pt), I915_CACHE_LLC);
522 fill_px(vm->dev, pd, scratch_pde);
525 /* Broadwell Page Directory Pointer Descriptors */
526 static int gen8_write_pdp(struct drm_i915_gem_request *req,
530 struct intel_engine_cs *ring = req->ring;
535 ret = intel_ring_begin(req, 6);
539 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
540 intel_ring_emit(ring, GEN8_RING_PDP_UDW(ring, entry));
541 intel_ring_emit(ring, upper_32_bits(addr));
542 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
543 intel_ring_emit(ring, GEN8_RING_PDP_LDW(ring, entry));
544 intel_ring_emit(ring, lower_32_bits(addr));
545 intel_ring_advance(ring);
550 static int gen8_mm_switch(struct i915_hw_ppgtt *ppgtt,
551 struct drm_i915_gem_request *req)
555 for (i = GEN8_LEGACY_PDPES - 1; i >= 0; i--) {
556 const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i);
558 ret = gen8_write_pdp(req, i, pd_daddr);
566 static void gen8_ppgtt_clear_range(struct i915_address_space *vm,
571 struct i915_hw_ppgtt *ppgtt =
572 container_of(vm, struct i915_hw_ppgtt, base);
573 gen8_pte_t *pt_vaddr, scratch_pte;
574 unsigned pdpe = start >> GEN8_PDPE_SHIFT & GEN8_PDPE_MASK;
575 unsigned pde = start >> GEN8_PDE_SHIFT & GEN8_PDE_MASK;
576 unsigned pte = start >> GEN8_PTE_SHIFT & GEN8_PTE_MASK;
577 unsigned num_entries = length >> PAGE_SHIFT;
578 unsigned last_pte, i;
580 scratch_pte = gen8_pte_encode(px_dma(ppgtt->base.scratch_page),
581 I915_CACHE_LLC, use_scratch);
583 while (num_entries) {
584 struct i915_page_directory *pd;
585 struct i915_page_table *pt;
587 if (WARN_ON(!ppgtt->pdp.page_directory[pdpe]))
590 pd = ppgtt->pdp.page_directory[pdpe];
592 if (WARN_ON(!pd->page_table[pde]))
595 pt = pd->page_table[pde];
597 if (WARN_ON(!px_page(pt)))
600 last_pte = pte + num_entries;
601 if (last_pte > GEN8_PTES)
602 last_pte = GEN8_PTES;
604 pt_vaddr = kmap_px(pt);
606 for (i = pte; i < last_pte; i++) {
607 pt_vaddr[i] = scratch_pte;
611 kunmap_px(ppgtt, pt);
614 if (++pde == I915_PDES) {
621 static void gen8_ppgtt_insert_entries(struct i915_address_space *vm,
622 struct sg_table *pages,
624 enum i915_cache_level cache_level, u32 unused)
626 struct i915_hw_ppgtt *ppgtt =
627 container_of(vm, struct i915_hw_ppgtt, base);
628 gen8_pte_t *pt_vaddr;
629 unsigned pdpe = start >> GEN8_PDPE_SHIFT & GEN8_PDPE_MASK;
630 unsigned pde = start >> GEN8_PDE_SHIFT & GEN8_PDE_MASK;
631 unsigned pte = start >> GEN8_PTE_SHIFT & GEN8_PTE_MASK;
632 struct sg_page_iter sg_iter;
636 for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) {
637 if (WARN_ON(pdpe >= GEN8_LEGACY_PDPES))
640 if (pt_vaddr == NULL) {
641 struct i915_page_directory *pd = ppgtt->pdp.page_directory[pdpe];
642 struct i915_page_table *pt = pd->page_table[pde];
643 pt_vaddr = kmap_px(pt);
647 gen8_pte_encode(sg_page_iter_dma_address(&sg_iter),
649 if (++pte == GEN8_PTES) {
650 kunmap_px(ppgtt, pt_vaddr);
652 if (++pde == I915_PDES) {
661 kunmap_px(ppgtt, pt_vaddr);
664 static void gen8_free_page_tables(struct drm_device *dev,
665 struct i915_page_directory *pd)
672 for_each_set_bit(i, pd->used_pdes, I915_PDES) {
673 if (WARN_ON(!pd->page_table[i]))
676 free_pt(dev, pd->page_table[i]);
677 pd->page_table[i] = NULL;
681 static void gen8_ppgtt_cleanup(struct i915_address_space *vm)
683 struct i915_hw_ppgtt *ppgtt =
684 container_of(vm, struct i915_hw_ppgtt, base);
687 for_each_set_bit(i, ppgtt->pdp.used_pdpes, GEN8_LEGACY_PDPES) {
688 if (WARN_ON(!ppgtt->pdp.page_directory[i]))
691 gen8_free_page_tables(ppgtt->base.dev,
692 ppgtt->pdp.page_directory[i]);
693 free_pd(ppgtt->base.dev, ppgtt->pdp.page_directory[i]);
696 free_pd(vm->dev, vm->scratch_pd);
697 free_pt(vm->dev, vm->scratch_pt);
701 * gen8_ppgtt_alloc_pagetabs() - Allocate page tables for VA range.
702 * @ppgtt: Master ppgtt structure.
703 * @pd: Page directory for this address range.
704 * @start: Starting virtual address to begin allocations.
705 * @length Size of the allocations.
706 * @new_pts: Bitmap set by function with new allocations. Likely used by the
707 * caller to free on error.
709 * Allocate the required number of page tables. Extremely similar to
710 * gen8_ppgtt_alloc_page_directories(). The main difference is here we are limited by
711 * the page directory boundary (instead of the page directory pointer). That
712 * boundary is 1GB virtual. Therefore, unlike gen8_ppgtt_alloc_page_directories(), it is
713 * possible, and likely that the caller will need to use multiple calls of this
714 * function to achieve the appropriate allocation.
716 * Return: 0 if success; negative error code otherwise.
718 static int gen8_ppgtt_alloc_pagetabs(struct i915_hw_ppgtt *ppgtt,
719 struct i915_page_directory *pd,
722 unsigned long *new_pts)
724 struct drm_device *dev = ppgtt->base.dev;
725 struct i915_page_table *pt;
729 gen8_for_each_pde(pt, pd, start, length, temp, pde) {
730 /* Don't reallocate page tables */
732 /* Scratch is never allocated this way */
733 WARN_ON(pt == ppgtt->base.scratch_pt);
741 gen8_initialize_pt(&ppgtt->base, pt);
742 pd->page_table[pde] = pt;
743 __set_bit(pde, new_pts);
749 for_each_set_bit(pde, new_pts, I915_PDES)
750 free_pt(dev, pd->page_table[pde]);
756 * gen8_ppgtt_alloc_page_directories() - Allocate page directories for VA range.
757 * @ppgtt: Master ppgtt structure.
758 * @pdp: Page directory pointer for this address range.
759 * @start: Starting virtual address to begin allocations.
760 * @length Size of the allocations.
761 * @new_pds Bitmap set by function with new allocations. Likely used by the
762 * caller to free on error.
764 * Allocate the required number of page directories starting at the pde index of
765 * @start, and ending at the pde index @start + @length. This function will skip
766 * over already allocated page directories within the range, and only allocate
767 * new ones, setting the appropriate pointer within the pdp as well as the
768 * correct position in the bitmap @new_pds.
770 * The function will only allocate the pages within the range for a give page
771 * directory pointer. In other words, if @start + @length straddles a virtually
772 * addressed PDP boundary (512GB for 4k pages), there will be more allocations
773 * required by the caller, This is not currently possible, and the BUG in the
774 * code will prevent it.
776 * Return: 0 if success; negative error code otherwise.
778 static int gen8_ppgtt_alloc_page_directories(struct i915_hw_ppgtt *ppgtt,
779 struct i915_page_directory_pointer *pdp,
782 unsigned long *new_pds)
784 struct drm_device *dev = ppgtt->base.dev;
785 struct i915_page_directory *pd;
789 WARN_ON(!bitmap_empty(new_pds, GEN8_LEGACY_PDPES));
791 gen8_for_each_pdpe(pd, pdp, start, length, temp, pdpe) {
799 gen8_initialize_pd(&ppgtt->base, pd);
800 pdp->page_directory[pdpe] = pd;
801 __set_bit(pdpe, new_pds);
807 for_each_set_bit(pdpe, new_pds, GEN8_LEGACY_PDPES)
808 free_pd(dev, pdp->page_directory[pdpe]);
814 free_gen8_temp_bitmaps(unsigned long *new_pds, unsigned long **new_pts)
818 for (i = 0; i < GEN8_LEGACY_PDPES; i++)
824 /* Fills in the page directory bitmap, and the array of page tables bitmap. Both
825 * of these are based on the number of PDPEs in the system.
828 int __must_check alloc_gen8_temp_bitmaps(unsigned long **new_pds,
829 unsigned long ***new_pts)
835 pds = kcalloc(BITS_TO_LONGS(GEN8_LEGACY_PDPES), sizeof(unsigned long), GFP_KERNEL);
839 pts = kcalloc(GEN8_LEGACY_PDPES, sizeof(unsigned long *), GFP_KERNEL);
845 for (i = 0; i < GEN8_LEGACY_PDPES; i++) {
846 pts[i] = kcalloc(BITS_TO_LONGS(I915_PDES),
847 sizeof(unsigned long), GFP_KERNEL);
858 free_gen8_temp_bitmaps(pds, pts);
862 /* PDE TLBs are a pain to invalidate on GEN8+. When we modify
863 * the page table structures, we mark them dirty so that
864 * context switching/execlist queuing code takes extra steps
865 * to ensure that tlbs are flushed.
867 static void mark_tlbs_dirty(struct i915_hw_ppgtt *ppgtt)
869 ppgtt->pd_dirty_rings = INTEL_INFO(ppgtt->base.dev)->ring_mask;
872 static int gen8_alloc_va_range(struct i915_address_space *vm,
876 struct i915_hw_ppgtt *ppgtt =
877 container_of(vm, struct i915_hw_ppgtt, base);
878 unsigned long *new_page_dirs, **new_page_tables;
879 struct i915_page_directory *pd;
880 const uint64_t orig_start = start;
881 const uint64_t orig_length = length;
886 /* Wrap is never okay since we can only represent 48b, and we don't
887 * actually use the other side of the canonical address space.
889 if (WARN_ON(start + length < start))
892 if (WARN_ON(start + length > ppgtt->base.total))
895 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables);
899 /* Do the allocations first so we can easily bail out */
900 ret = gen8_ppgtt_alloc_page_directories(ppgtt, &ppgtt->pdp, start, length,
903 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
907 /* For every page directory referenced, allocate page tables */
908 gen8_for_each_pdpe(pd, &ppgtt->pdp, start, length, temp, pdpe) {
909 ret = gen8_ppgtt_alloc_pagetabs(ppgtt, pd, start, length,
910 new_page_tables[pdpe]);
916 length = orig_length;
918 /* Allocations have completed successfully, so set the bitmaps, and do
920 gen8_for_each_pdpe(pd, &ppgtt->pdp, start, length, temp, pdpe) {
921 gen8_pde_t *const page_directory = kmap_px(pd);
922 struct i915_page_table *pt;
923 uint64_t pd_len = gen8_clamp_pd(start, length);
924 uint64_t pd_start = start;
927 /* Every pd should be allocated, we just did that above. */
930 gen8_for_each_pde(pt, pd, pd_start, pd_len, temp, pde) {
931 /* Same reasoning as pd */
934 WARN_ON(!gen8_pte_count(pd_start, pd_len));
936 /* Set our used ptes within the page table */
937 bitmap_set(pt->used_ptes,
938 gen8_pte_index(pd_start),
939 gen8_pte_count(pd_start, pd_len));
941 /* Our pde is now pointing to the pagetable, pt */
942 __set_bit(pde, pd->used_pdes);
944 /* Map the PDE to the page table */
945 page_directory[pde] = gen8_pde_encode(px_dma(pt),
948 /* NB: We haven't yet mapped ptes to pages. At this
949 * point we're still relying on insert_entries() */
952 kunmap_px(ppgtt, page_directory);
954 __set_bit(pdpe, ppgtt->pdp.used_pdpes);
957 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
958 mark_tlbs_dirty(ppgtt);
963 for_each_set_bit(temp, new_page_tables[pdpe], I915_PDES)
964 free_pt(vm->dev, ppgtt->pdp.page_directory[pdpe]->page_table[temp]);
967 for_each_set_bit(pdpe, new_page_dirs, GEN8_LEGACY_PDPES)
968 free_pd(vm->dev, ppgtt->pdp.page_directory[pdpe]);
970 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
971 mark_tlbs_dirty(ppgtt);
976 * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers
977 * with a net effect resembling a 2-level page table in normal x86 terms. Each
978 * PDP represents 1GB of memory 4 * 512 * 512 * 4096 = 4GB legacy 32b address
982 static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
984 ppgtt->base.scratch_pt = alloc_pt(ppgtt->base.dev);
985 if (IS_ERR(ppgtt->base.scratch_pt))
986 return PTR_ERR(ppgtt->base.scratch_pt);
988 ppgtt->base.scratch_pd = alloc_pd(ppgtt->base.dev);
989 if (IS_ERR(ppgtt->base.scratch_pd))
990 return PTR_ERR(ppgtt->base.scratch_pd);
992 gen8_initialize_pt(&ppgtt->base, ppgtt->base.scratch_pt);
993 gen8_initialize_pd(&ppgtt->base, ppgtt->base.scratch_pd);
995 ppgtt->base.start = 0;
996 ppgtt->base.total = 1ULL << 32;
997 if (IS_ENABLED(CONFIG_X86_32))
998 /* While we have a proliferation of size_t variables
999 * we cannot represent the full ppgtt size on 32bit,
1000 * so limit it to the same size as the GGTT (currently
1003 ppgtt->base.total = to_i915(ppgtt->base.dev)->gtt.base.total;
1004 ppgtt->base.cleanup = gen8_ppgtt_cleanup;
1005 ppgtt->base.allocate_va_range = gen8_alloc_va_range;
1006 ppgtt->base.insert_entries = gen8_ppgtt_insert_entries;
1007 ppgtt->base.clear_range = gen8_ppgtt_clear_range;
1008 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
1009 ppgtt->base.bind_vma = ppgtt_bind_vma;
1011 ppgtt->switch_mm = gen8_mm_switch;
1016 static void gen6_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
1018 struct i915_address_space *vm = &ppgtt->base;
1019 struct i915_page_table *unused;
1020 gen6_pte_t scratch_pte;
1022 uint32_t pte, pde, temp;
1023 uint32_t start = ppgtt->base.start, length = ppgtt->base.total;
1025 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1026 I915_CACHE_LLC, true, 0);
1028 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde) {
1030 gen6_pte_t *pt_vaddr;
1031 const dma_addr_t pt_addr = px_dma(ppgtt->pd.page_table[pde]);
1032 pd_entry = readl(ppgtt->pd_addr + pde);
1033 expected = (GEN6_PDE_ADDR_ENCODE(pt_addr) | GEN6_PDE_VALID);
1035 if (pd_entry != expected)
1036 seq_printf(m, "\tPDE #%d mismatch: Actual PDE: %x Expected PDE: %x\n",
1040 seq_printf(m, "\tPDE: %x\n", pd_entry);
1042 pt_vaddr = kmap_px(ppgtt->pd.page_table[pde]);
1044 for (pte = 0; pte < GEN6_PTES; pte+=4) {
1046 (pde * PAGE_SIZE * GEN6_PTES) +
1050 for (i = 0; i < 4; i++)
1051 if (pt_vaddr[pte + i] != scratch_pte)
1056 seq_printf(m, "\t\t0x%lx [%03d,%04d]: =", va, pde, pte);
1057 for (i = 0; i < 4; i++) {
1058 if (pt_vaddr[pte + i] != scratch_pte)
1059 seq_printf(m, " %08x", pt_vaddr[pte + i]);
1061 seq_puts(m, " SCRATCH ");
1065 kunmap_px(ppgtt, pt_vaddr);
1069 /* Write pde (index) from the page directory @pd to the page table @pt */
1070 static void gen6_write_pde(struct i915_page_directory *pd,
1071 const int pde, struct i915_page_table *pt)
1073 /* Caller needs to make sure the write completes if necessary */
1074 struct i915_hw_ppgtt *ppgtt =
1075 container_of(pd, struct i915_hw_ppgtt, pd);
1078 pd_entry = GEN6_PDE_ADDR_ENCODE(px_dma(pt));
1079 pd_entry |= GEN6_PDE_VALID;
1081 writel(pd_entry, ppgtt->pd_addr + pde);
1084 /* Write all the page tables found in the ppgtt structure to incrementing page
1086 static void gen6_write_page_range(struct drm_i915_private *dev_priv,
1087 struct i915_page_directory *pd,
1088 uint32_t start, uint32_t length)
1090 struct i915_page_table *pt;
1093 gen6_for_each_pde(pt, pd, start, length, temp, pde)
1094 gen6_write_pde(pd, pde, pt);
1096 /* Make sure write is complete before other code can use this page
1097 * table. Also require for WC mapped PTEs */
1098 readl(dev_priv->gtt.gsm);
1101 static uint32_t get_pd_offset(struct i915_hw_ppgtt *ppgtt)
1103 BUG_ON(ppgtt->pd.base.ggtt_offset & 0x3f);
1105 return (ppgtt->pd.base.ggtt_offset / 64) << 16;
1108 static int hsw_mm_switch(struct i915_hw_ppgtt *ppgtt,
1109 struct drm_i915_gem_request *req)
1111 struct intel_engine_cs *ring = req->ring;
1114 /* NB: TLBs must be flushed and invalidated before a switch */
1115 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1119 ret = intel_ring_begin(req, 6);
1123 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1124 intel_ring_emit(ring, RING_PP_DIR_DCLV(ring));
1125 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1126 intel_ring_emit(ring, RING_PP_DIR_BASE(ring));
1127 intel_ring_emit(ring, get_pd_offset(ppgtt));
1128 intel_ring_emit(ring, MI_NOOP);
1129 intel_ring_advance(ring);
1134 static int vgpu_mm_switch(struct i915_hw_ppgtt *ppgtt,
1135 struct drm_i915_gem_request *req)
1137 struct intel_engine_cs *ring = req->ring;
1138 struct drm_i915_private *dev_priv = to_i915(ppgtt->base.dev);
1140 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
1141 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
1145 static int gen7_mm_switch(struct i915_hw_ppgtt *ppgtt,
1146 struct drm_i915_gem_request *req)
1148 struct intel_engine_cs *ring = req->ring;
1151 /* NB: TLBs must be flushed and invalidated before a switch */
1152 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1156 ret = intel_ring_begin(req, 6);
1160 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1161 intel_ring_emit(ring, RING_PP_DIR_DCLV(ring));
1162 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1163 intel_ring_emit(ring, RING_PP_DIR_BASE(ring));
1164 intel_ring_emit(ring, get_pd_offset(ppgtt));
1165 intel_ring_emit(ring, MI_NOOP);
1166 intel_ring_advance(ring);
1168 /* XXX: RCS is the only one to auto invalidate the TLBs? */
1169 if (ring->id != RCS) {
1170 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1178 static int gen6_mm_switch(struct i915_hw_ppgtt *ppgtt,
1179 struct drm_i915_gem_request *req)
1181 struct intel_engine_cs *ring = req->ring;
1182 struct drm_device *dev = ppgtt->base.dev;
1183 struct drm_i915_private *dev_priv = dev->dev_private;
1186 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
1187 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
1189 POSTING_READ(RING_PP_DIR_DCLV(ring));
1194 static void gen8_ppgtt_enable(struct drm_device *dev)
1196 struct drm_i915_private *dev_priv = dev->dev_private;
1197 struct intel_engine_cs *ring;
1200 for_each_ring(ring, dev_priv, j) {
1201 I915_WRITE(RING_MODE_GEN7(ring),
1202 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1206 static void gen7_ppgtt_enable(struct drm_device *dev)
1208 struct drm_i915_private *dev_priv = dev->dev_private;
1209 struct intel_engine_cs *ring;
1210 uint32_t ecochk, ecobits;
1213 ecobits = I915_READ(GAC_ECO_BITS);
1214 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B);
1216 ecochk = I915_READ(GAM_ECOCHK);
1217 if (IS_HASWELL(dev)) {
1218 ecochk |= ECOCHK_PPGTT_WB_HSW;
1220 ecochk |= ECOCHK_PPGTT_LLC_IVB;
1221 ecochk &= ~ECOCHK_PPGTT_GFDT_IVB;
1223 I915_WRITE(GAM_ECOCHK, ecochk);
1225 for_each_ring(ring, dev_priv, i) {
1226 /* GFX_MODE is per-ring on gen7+ */
1227 I915_WRITE(RING_MODE_GEN7(ring),
1228 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1232 static void gen6_ppgtt_enable(struct drm_device *dev)
1234 struct drm_i915_private *dev_priv = dev->dev_private;
1235 uint32_t ecochk, gab_ctl, ecobits;
1237 ecobits = I915_READ(GAC_ECO_BITS);
1238 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_SNB_BIT |
1239 ECOBITS_PPGTT_CACHE64B);
1241 gab_ctl = I915_READ(GAB_CTL);
1242 I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT);
1244 ecochk = I915_READ(GAM_ECOCHK);
1245 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT | ECOCHK_PPGTT_CACHE64B);
1247 I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1250 /* PPGTT support for Sandybdrige/Gen6 and later */
1251 static void gen6_ppgtt_clear_range(struct i915_address_space *vm,
1256 struct i915_hw_ppgtt *ppgtt =
1257 container_of(vm, struct i915_hw_ppgtt, base);
1258 gen6_pte_t *pt_vaddr, scratch_pte;
1259 unsigned first_entry = start >> PAGE_SHIFT;
1260 unsigned num_entries = length >> PAGE_SHIFT;
1261 unsigned act_pt = first_entry / GEN6_PTES;
1262 unsigned first_pte = first_entry % GEN6_PTES;
1263 unsigned last_pte, i;
1265 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1266 I915_CACHE_LLC, true, 0);
1268 while (num_entries) {
1269 last_pte = first_pte + num_entries;
1270 if (last_pte > GEN6_PTES)
1271 last_pte = GEN6_PTES;
1273 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1275 for (i = first_pte; i < last_pte; i++)
1276 pt_vaddr[i] = scratch_pte;
1278 kunmap_px(ppgtt, pt_vaddr);
1280 num_entries -= last_pte - first_pte;
1286 static void gen6_ppgtt_insert_entries(struct i915_address_space *vm,
1287 struct sg_table *pages,
1289 enum i915_cache_level cache_level, u32 flags)
1291 struct i915_hw_ppgtt *ppgtt =
1292 container_of(vm, struct i915_hw_ppgtt, base);
1293 gen6_pte_t *pt_vaddr;
1294 unsigned first_entry = start >> PAGE_SHIFT;
1295 unsigned act_pt = first_entry / GEN6_PTES;
1296 unsigned act_pte = first_entry % GEN6_PTES;
1297 struct sg_page_iter sg_iter;
1300 for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) {
1301 if (pt_vaddr == NULL)
1302 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1305 vm->pte_encode(sg_page_iter_dma_address(&sg_iter),
1306 cache_level, true, flags);
1308 if (++act_pte == GEN6_PTES) {
1309 kunmap_px(ppgtt, pt_vaddr);
1316 kunmap_px(ppgtt, pt_vaddr);
1319 static int gen6_alloc_va_range(struct i915_address_space *vm,
1320 uint64_t start_in, uint64_t length_in)
1322 DECLARE_BITMAP(new_page_tables, I915_PDES);
1323 struct drm_device *dev = vm->dev;
1324 struct drm_i915_private *dev_priv = dev->dev_private;
1325 struct i915_hw_ppgtt *ppgtt =
1326 container_of(vm, struct i915_hw_ppgtt, base);
1327 struct i915_page_table *pt;
1328 uint32_t start, length, start_save, length_save;
1332 if (WARN_ON(start_in + length_in > ppgtt->base.total))
1335 start = start_save = start_in;
1336 length = length_save = length_in;
1338 bitmap_zero(new_page_tables, I915_PDES);
1340 /* The allocation is done in two stages so that we can bail out with
1341 * minimal amount of pain. The first stage finds new page tables that
1342 * need allocation. The second stage marks use ptes within the page
1345 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1346 if (pt != vm->scratch_pt) {
1347 WARN_ON(bitmap_empty(pt->used_ptes, GEN6_PTES));
1351 /* We've already allocated a page table */
1352 WARN_ON(!bitmap_empty(pt->used_ptes, GEN6_PTES));
1360 gen6_initialize_pt(vm, pt);
1362 ppgtt->pd.page_table[pde] = pt;
1363 __set_bit(pde, new_page_tables);
1364 trace_i915_page_table_entry_alloc(vm, pde, start, GEN6_PDE_SHIFT);
1368 length = length_save;
1370 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1371 DECLARE_BITMAP(tmp_bitmap, GEN6_PTES);
1373 bitmap_zero(tmp_bitmap, GEN6_PTES);
1374 bitmap_set(tmp_bitmap, gen6_pte_index(start),
1375 gen6_pte_count(start, length));
1377 if (__test_and_clear_bit(pde, new_page_tables))
1378 gen6_write_pde(&ppgtt->pd, pde, pt);
1380 trace_i915_page_table_entry_map(vm, pde, pt,
1381 gen6_pte_index(start),
1382 gen6_pte_count(start, length),
1384 bitmap_or(pt->used_ptes, tmp_bitmap, pt->used_ptes,
1388 WARN_ON(!bitmap_empty(new_page_tables, I915_PDES));
1390 /* Make sure write is complete before other code can use this page
1391 * table. Also require for WC mapped PTEs */
1392 readl(dev_priv->gtt.gsm);
1394 mark_tlbs_dirty(ppgtt);
1398 for_each_set_bit(pde, new_page_tables, I915_PDES) {
1399 struct i915_page_table *pt = ppgtt->pd.page_table[pde];
1401 ppgtt->pd.page_table[pde] = vm->scratch_pt;
1402 free_pt(vm->dev, pt);
1405 mark_tlbs_dirty(ppgtt);
1409 static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
1411 struct i915_hw_ppgtt *ppgtt =
1412 container_of(vm, struct i915_hw_ppgtt, base);
1413 struct i915_page_table *pt;
1416 drm_mm_remove_node(&ppgtt->node);
1418 gen6_for_all_pdes(pt, ppgtt, pde) {
1419 if (pt != vm->scratch_pt)
1420 free_pt(ppgtt->base.dev, pt);
1423 free_pt(vm->dev, vm->scratch_pt);
1426 static int gen6_ppgtt_allocate_page_directories(struct i915_hw_ppgtt *ppgtt)
1428 struct drm_device *dev = ppgtt->base.dev;
1429 struct drm_i915_private *dev_priv = dev->dev_private;
1430 bool retried = false;
1433 /* PPGTT PDEs reside in the GGTT and consists of 512 entries. The
1434 * allocator works in address space sizes, so it's multiplied by page
1435 * size. We allocate at the top of the GTT to avoid fragmentation.
1437 BUG_ON(!drm_mm_initialized(&dev_priv->gtt.base.mm));
1438 ppgtt->base.scratch_pt = alloc_pt(ppgtt->base.dev);
1439 if (IS_ERR(ppgtt->base.scratch_pt))
1440 return PTR_ERR(ppgtt->base.scratch_pt);
1442 gen6_initialize_pt(&ppgtt->base, ppgtt->base.scratch_pt);
1445 ret = drm_mm_insert_node_in_range_generic(&dev_priv->gtt.base.mm,
1446 &ppgtt->node, GEN6_PD_SIZE,
1448 0, dev_priv->gtt.base.total,
1450 if (ret == -ENOSPC && !retried) {
1451 ret = i915_gem_evict_something(dev, &dev_priv->gtt.base,
1452 GEN6_PD_SIZE, GEN6_PD_ALIGN,
1454 0, dev_priv->gtt.base.total,
1467 if (ppgtt->node.start < dev_priv->gtt.mappable_end)
1468 DRM_DEBUG("Forced to use aperture for PDEs\n");
1473 free_pt(ppgtt->base.dev, ppgtt->base.scratch_pt);
1477 static int gen6_ppgtt_alloc(struct i915_hw_ppgtt *ppgtt)
1479 return gen6_ppgtt_allocate_page_directories(ppgtt);
1482 static void gen6_scratch_va_range(struct i915_hw_ppgtt *ppgtt,
1483 uint64_t start, uint64_t length)
1485 struct i915_page_table *unused;
1488 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde)
1489 ppgtt->pd.page_table[pde] = ppgtt->base.scratch_pt;
1492 static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
1494 struct drm_device *dev = ppgtt->base.dev;
1495 struct drm_i915_private *dev_priv = dev->dev_private;
1498 ppgtt->base.pte_encode = dev_priv->gtt.base.pte_encode;
1500 ppgtt->switch_mm = gen6_mm_switch;
1501 } else if (IS_HASWELL(dev)) {
1502 ppgtt->switch_mm = hsw_mm_switch;
1503 } else if (IS_GEN7(dev)) {
1504 ppgtt->switch_mm = gen7_mm_switch;
1508 if (intel_vgpu_active(dev))
1509 ppgtt->switch_mm = vgpu_mm_switch;
1511 ret = gen6_ppgtt_alloc(ppgtt);
1515 ppgtt->base.allocate_va_range = gen6_alloc_va_range;
1516 ppgtt->base.clear_range = gen6_ppgtt_clear_range;
1517 ppgtt->base.insert_entries = gen6_ppgtt_insert_entries;
1518 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
1519 ppgtt->base.bind_vma = ppgtt_bind_vma;
1520 ppgtt->base.cleanup = gen6_ppgtt_cleanup;
1521 ppgtt->base.start = 0;
1522 ppgtt->base.total = I915_PDES * GEN6_PTES * PAGE_SIZE;
1523 ppgtt->debug_dump = gen6_dump_ppgtt;
1525 ppgtt->pd.base.ggtt_offset =
1526 ppgtt->node.start / PAGE_SIZE * sizeof(gen6_pte_t);
1528 ppgtt->pd_addr = (gen6_pte_t __iomem *)dev_priv->gtt.gsm +
1529 ppgtt->pd.base.ggtt_offset / sizeof(gen6_pte_t);
1531 gen6_scratch_va_range(ppgtt, 0, ppgtt->base.total);
1533 gen6_write_page_range(dev_priv, &ppgtt->pd, 0, ppgtt->base.total);
1535 DRM_DEBUG_DRIVER("Allocated pde space (%lldM) at GTT entry: %llx\n",
1536 ppgtt->node.size >> 20,
1537 ppgtt->node.start / PAGE_SIZE);
1539 DRM_DEBUG("Adding PPGTT at offset %x\n",
1540 ppgtt->pd.base.ggtt_offset << 10);
1545 static int __hw_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
1547 struct drm_i915_private *dev_priv = dev->dev_private;
1549 ppgtt->base.dev = dev;
1550 ppgtt->base.scratch_page = dev_priv->gtt.base.scratch_page;
1552 if (INTEL_INFO(dev)->gen < 8)
1553 return gen6_ppgtt_init(ppgtt);
1555 return gen8_ppgtt_init(ppgtt);
1558 int i915_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
1560 struct drm_i915_private *dev_priv = dev->dev_private;
1563 ret = __hw_ppgtt_init(dev, ppgtt);
1565 kref_init(&ppgtt->ref);
1566 drm_mm_init(&ppgtt->base.mm, ppgtt->base.start,
1568 i915_init_vm(dev_priv, &ppgtt->base);
1574 int i915_ppgtt_init_hw(struct drm_device *dev)
1576 /* In the case of execlists, PPGTT is enabled by the context descriptor
1577 * and the PDPs are contained within the context itself. We don't
1578 * need to do anything here. */
1579 if (i915.enable_execlists)
1582 if (!USES_PPGTT(dev))
1586 gen6_ppgtt_enable(dev);
1587 else if (IS_GEN7(dev))
1588 gen7_ppgtt_enable(dev);
1589 else if (INTEL_INFO(dev)->gen >= 8)
1590 gen8_ppgtt_enable(dev);
1592 MISSING_CASE(INTEL_INFO(dev)->gen);
1597 int i915_ppgtt_init_ring(struct drm_i915_gem_request *req)
1599 struct drm_i915_private *dev_priv = req->ring->dev->dev_private;
1600 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
1602 if (i915.enable_execlists)
1608 return ppgtt->switch_mm(ppgtt, req);
1611 struct i915_hw_ppgtt *
1612 i915_ppgtt_create(struct drm_device *dev, struct drm_i915_file_private *fpriv)
1614 struct i915_hw_ppgtt *ppgtt;
1617 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
1619 return ERR_PTR(-ENOMEM);
1621 ret = i915_ppgtt_init(dev, ppgtt);
1624 return ERR_PTR(ret);
1627 ppgtt->file_priv = fpriv;
1629 trace_i915_ppgtt_create(&ppgtt->base);
1634 void i915_ppgtt_release(struct kref *kref)
1636 struct i915_hw_ppgtt *ppgtt =
1637 container_of(kref, struct i915_hw_ppgtt, ref);
1639 trace_i915_ppgtt_release(&ppgtt->base);
1641 /* vmas should already be unbound */
1642 WARN_ON(!list_empty(&ppgtt->base.active_list));
1643 WARN_ON(!list_empty(&ppgtt->base.inactive_list));
1645 list_del(&ppgtt->base.global_link);
1646 drm_mm_takedown(&ppgtt->base.mm);
1648 ppgtt->base.cleanup(&ppgtt->base);
1652 extern int intel_iommu_gfx_mapped;
1653 /* Certain Gen5 chipsets require require idling the GPU before
1654 * unmapping anything from the GTT when VT-d is enabled.
1656 static bool needs_idle_maps(struct drm_device *dev)
1658 #ifdef CONFIG_INTEL_IOMMU
1659 /* Query intel_iommu to see if we need the workaround. Presumably that
1662 if (IS_GEN5(dev) && IS_MOBILE(dev) && intel_iommu_gfx_mapped)
1668 static bool do_idling(struct drm_i915_private *dev_priv)
1670 bool ret = dev_priv->mm.interruptible;
1672 if (unlikely(dev_priv->gtt.do_idle_maps)) {
1673 dev_priv->mm.interruptible = false;
1674 if (i915_gpu_idle(dev_priv->dev)) {
1675 DRM_ERROR("Couldn't idle GPU\n");
1676 /* Wait a bit, in hopes it avoids the hang */
1684 static void undo_idling(struct drm_i915_private *dev_priv, bool interruptible)
1686 if (unlikely(dev_priv->gtt.do_idle_maps))
1687 dev_priv->mm.interruptible = interruptible;
1690 void i915_check_and_clear_faults(struct drm_device *dev)
1692 struct drm_i915_private *dev_priv = dev->dev_private;
1693 struct intel_engine_cs *ring;
1696 if (INTEL_INFO(dev)->gen < 6)
1699 for_each_ring(ring, dev_priv, i) {
1701 fault_reg = I915_READ(RING_FAULT_REG(ring));
1702 if (fault_reg & RING_FAULT_VALID) {
1703 DRM_DEBUG_DRIVER("Unexpected fault\n"
1705 "\tAddress space: %s\n"
1708 fault_reg & PAGE_MASK,
1709 fault_reg & RING_FAULT_GTTSEL_MASK ? "GGTT" : "PPGTT",
1710 RING_FAULT_SRCID(fault_reg),
1711 RING_FAULT_FAULT_TYPE(fault_reg));
1712 I915_WRITE(RING_FAULT_REG(ring),
1713 fault_reg & ~RING_FAULT_VALID);
1716 POSTING_READ(RING_FAULT_REG(&dev_priv->ring[RCS]));
1719 static void i915_ggtt_flush(struct drm_i915_private *dev_priv)
1721 if (INTEL_INFO(dev_priv->dev)->gen < 6) {
1722 intel_gtt_chipset_flush();
1724 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
1725 POSTING_READ(GFX_FLSH_CNTL_GEN6);
1729 void i915_gem_suspend_gtt_mappings(struct drm_device *dev)
1731 struct drm_i915_private *dev_priv = dev->dev_private;
1733 /* Don't bother messing with faults pre GEN6 as we have little
1734 * documentation supporting that it's a good idea.
1736 if (INTEL_INFO(dev)->gen < 6)
1739 i915_check_and_clear_faults(dev);
1741 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
1742 dev_priv->gtt.base.start,
1743 dev_priv->gtt.base.total,
1746 i915_ggtt_flush(dev_priv);
1749 int i915_gem_gtt_prepare_object(struct drm_i915_gem_object *obj)
1751 if (obj->has_dma_mapping)
1754 if (!dma_map_sg(&obj->base.dev->pdev->dev,
1755 obj->pages->sgl, obj->pages->nents,
1756 PCI_DMA_BIDIRECTIONAL))
1762 static void gen8_set_pte(void __iomem *addr, gen8_pte_t pte)
1767 iowrite32((u32)pte, addr);
1768 iowrite32(pte >> 32, addr + 4);
1772 static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
1773 struct sg_table *st,
1775 enum i915_cache_level level, u32 unused)
1777 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1778 unsigned first_entry = start >> PAGE_SHIFT;
1779 gen8_pte_t __iomem *gtt_entries =
1780 (gen8_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
1782 struct sg_page_iter sg_iter;
1783 dma_addr_t addr = 0; /* shut up gcc */
1785 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
1786 addr = sg_dma_address(sg_iter.sg) +
1787 (sg_iter.sg_pgoffset << PAGE_SHIFT);
1788 gen8_set_pte(>t_entries[i],
1789 gen8_pte_encode(addr, level, true));
1794 * XXX: This serves as a posting read to make sure that the PTE has
1795 * actually been updated. There is some concern that even though
1796 * registers and PTEs are within the same BAR that they are potentially
1797 * of NUMA access patterns. Therefore, even with the way we assume
1798 * hardware should work, we must keep this posting read for paranoia.
1801 WARN_ON(readq(>t_entries[i-1])
1802 != gen8_pte_encode(addr, level, true));
1804 /* This next bit makes the above posting read even more important. We
1805 * want to flush the TLBs only after we're certain all the PTE updates
1808 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
1809 POSTING_READ(GFX_FLSH_CNTL_GEN6);
1813 * Binds an object into the global gtt with the specified cache level. The object
1814 * will be accessible to the GPU via commands whose operands reference offsets
1815 * within the global GTT as well as accessible by the GPU through the GMADR
1816 * mapped BAR (dev_priv->mm.gtt->gtt).
1818 static void gen6_ggtt_insert_entries(struct i915_address_space *vm,
1819 struct sg_table *st,
1821 enum i915_cache_level level, u32 flags)
1823 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1824 unsigned first_entry = start >> PAGE_SHIFT;
1825 gen6_pte_t __iomem *gtt_entries =
1826 (gen6_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
1828 struct sg_page_iter sg_iter;
1829 dma_addr_t addr = 0;
1831 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
1832 addr = sg_page_iter_dma_address(&sg_iter);
1833 iowrite32(vm->pte_encode(addr, level, true, flags), >t_entries[i]);
1837 /* XXX: This serves as a posting read to make sure that the PTE has
1838 * actually been updated. There is some concern that even though
1839 * registers and PTEs are within the same BAR that they are potentially
1840 * of NUMA access patterns. Therefore, even with the way we assume
1841 * hardware should work, we must keep this posting read for paranoia.
1844 unsigned long gtt = readl(>t_entries[i-1]);
1845 WARN_ON(gtt != vm->pte_encode(addr, level, true, flags));
1848 /* This next bit makes the above posting read even more important. We
1849 * want to flush the TLBs only after we're certain all the PTE updates
1852 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
1853 POSTING_READ(GFX_FLSH_CNTL_GEN6);
1856 static void gen8_ggtt_clear_range(struct i915_address_space *vm,
1861 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1862 unsigned first_entry = start >> PAGE_SHIFT;
1863 unsigned num_entries = length >> PAGE_SHIFT;
1864 gen8_pte_t scratch_pte, __iomem *gtt_base =
1865 (gen8_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
1866 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
1869 if (WARN(num_entries > max_entries,
1870 "First entry = %d; Num entries = %d (max=%d)\n",
1871 first_entry, num_entries, max_entries))
1872 num_entries = max_entries;
1874 scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
1877 for (i = 0; i < num_entries; i++)
1878 gen8_set_pte(>t_base[i], scratch_pte);
1882 static void gen6_ggtt_clear_range(struct i915_address_space *vm,
1887 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1888 unsigned first_entry = start >> PAGE_SHIFT;
1889 unsigned num_entries = length >> PAGE_SHIFT;
1890 gen6_pte_t scratch_pte, __iomem *gtt_base =
1891 (gen6_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
1892 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
1895 if (WARN(num_entries > max_entries,
1896 "First entry = %d; Num entries = %d (max=%d)\n",
1897 first_entry, num_entries, max_entries))
1898 num_entries = max_entries;
1900 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1901 I915_CACHE_LLC, use_scratch, 0);
1903 for (i = 0; i < num_entries; i++)
1904 iowrite32(scratch_pte, >t_base[i]);
1908 static void i915_ggtt_insert_entries(struct i915_address_space *vm,
1909 struct sg_table *pages,
1911 enum i915_cache_level cache_level, u32 unused)
1913 unsigned int flags = (cache_level == I915_CACHE_NONE) ?
1914 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
1916 intel_gtt_insert_sg_entries(pages, start >> PAGE_SHIFT, flags);
1920 static void i915_ggtt_clear_range(struct i915_address_space *vm,
1925 unsigned first_entry = start >> PAGE_SHIFT;
1926 unsigned num_entries = length >> PAGE_SHIFT;
1927 intel_gtt_clear_range(first_entry, num_entries);
1930 static int ggtt_bind_vma(struct i915_vma *vma,
1931 enum i915_cache_level cache_level,
1934 struct drm_device *dev = vma->vm->dev;
1935 struct drm_i915_private *dev_priv = dev->dev_private;
1936 struct drm_i915_gem_object *obj = vma->obj;
1937 struct sg_table *pages = obj->pages;
1941 ret = i915_get_ggtt_vma_pages(vma);
1944 pages = vma->ggtt_view.pages;
1946 /* Currently applicable only to VLV */
1948 pte_flags |= PTE_READ_ONLY;
1951 if (!dev_priv->mm.aliasing_ppgtt || flags & GLOBAL_BIND) {
1952 vma->vm->insert_entries(vma->vm, pages,
1954 cache_level, pte_flags);
1957 if (dev_priv->mm.aliasing_ppgtt && flags & LOCAL_BIND) {
1958 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
1959 appgtt->base.insert_entries(&appgtt->base, pages,
1961 cache_level, pte_flags);
1967 static void ggtt_unbind_vma(struct i915_vma *vma)
1969 struct drm_device *dev = vma->vm->dev;
1970 struct drm_i915_private *dev_priv = dev->dev_private;
1971 struct drm_i915_gem_object *obj = vma->obj;
1972 const uint64_t size = min_t(uint64_t,
1976 if (vma->bound & GLOBAL_BIND) {
1977 vma->vm->clear_range(vma->vm,
1983 if (dev_priv->mm.aliasing_ppgtt && vma->bound & LOCAL_BIND) {
1984 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
1986 appgtt->base.clear_range(&appgtt->base,
1993 void i915_gem_gtt_finish_object(struct drm_i915_gem_object *obj)
1995 struct drm_device *dev = obj->base.dev;
1996 struct drm_i915_private *dev_priv = dev->dev_private;
1999 interruptible = do_idling(dev_priv);
2001 if (!obj->has_dma_mapping)
2002 dma_unmap_sg(&dev->pdev->dev,
2003 obj->pages->sgl, obj->pages->nents,
2004 PCI_DMA_BIDIRECTIONAL);
2006 undo_idling(dev_priv, interruptible);
2009 static void i915_gtt_color_adjust(struct drm_mm_node *node,
2010 unsigned long color,
2014 if (node->color != color)
2017 if (!list_empty(&node->node_list)) {
2018 node = list_entry(node->node_list.next,
2021 if (node->allocated && node->color != color)
2026 static int i915_gem_setup_global_gtt(struct drm_device *dev,
2027 unsigned long start,
2028 unsigned long mappable_end,
2031 /* Let GEM Manage all of the aperture.
2033 * However, leave one page at the end still bound to the scratch page.
2034 * There are a number of places where the hardware apparently prefetches
2035 * past the end of the object, and we've seen multiple hangs with the
2036 * GPU head pointer stuck in a batchbuffer bound at the last page of the
2037 * aperture. One page should be enough to keep any prefetching inside
2040 struct drm_i915_private *dev_priv = dev->dev_private;
2041 struct i915_address_space *ggtt_vm = &dev_priv->gtt.base;
2042 struct drm_mm_node *entry;
2043 struct drm_i915_gem_object *obj;
2044 unsigned long hole_start, hole_end;
2047 BUG_ON(mappable_end > end);
2049 /* Subtract the guard page ... */
2050 drm_mm_init(&ggtt_vm->mm, start, end - start - PAGE_SIZE);
2052 dev_priv->gtt.base.start = start;
2053 dev_priv->gtt.base.total = end - start;
2055 if (intel_vgpu_active(dev)) {
2056 ret = intel_vgt_balloon(dev);
2062 dev_priv->gtt.base.mm.color_adjust = i915_gtt_color_adjust;
2064 /* Mark any preallocated objects as occupied */
2065 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
2066 struct i915_vma *vma = i915_gem_obj_to_vma(obj, ggtt_vm);
2068 DRM_DEBUG_KMS("reserving preallocated space: %lx + %zx\n",
2069 i915_gem_obj_ggtt_offset(obj), obj->base.size);
2071 WARN_ON(i915_gem_obj_ggtt_bound(obj));
2072 ret = drm_mm_reserve_node(&ggtt_vm->mm, &vma->node);
2074 DRM_DEBUG_KMS("Reservation failed: %i\n", ret);
2077 vma->bound |= GLOBAL_BIND;
2080 /* Clear any non-preallocated blocks */
2081 drm_mm_for_each_hole(entry, &ggtt_vm->mm, hole_start, hole_end) {
2082 DRM_DEBUG_KMS("clearing unused GTT space: [%lx, %lx]\n",
2083 hole_start, hole_end);
2084 ggtt_vm->clear_range(ggtt_vm, hole_start,
2085 hole_end - hole_start, true);
2088 /* And finally clear the reserved guard page */
2089 ggtt_vm->clear_range(ggtt_vm, end - PAGE_SIZE, PAGE_SIZE, true);
2091 if (USES_PPGTT(dev) && !USES_FULL_PPGTT(dev)) {
2092 struct i915_hw_ppgtt *ppgtt;
2094 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2098 ret = __hw_ppgtt_init(dev, ppgtt);
2100 ppgtt->base.cleanup(&ppgtt->base);
2105 if (ppgtt->base.allocate_va_range)
2106 ret = ppgtt->base.allocate_va_range(&ppgtt->base, 0,
2109 ppgtt->base.cleanup(&ppgtt->base);
2114 ppgtt->base.clear_range(&ppgtt->base,
2119 dev_priv->mm.aliasing_ppgtt = ppgtt;
2125 void i915_gem_init_global_gtt(struct drm_device *dev)
2127 struct drm_i915_private *dev_priv = dev->dev_private;
2128 u64 gtt_size, mappable_size;
2130 gtt_size = dev_priv->gtt.base.total;
2131 mappable_size = dev_priv->gtt.mappable_end;
2133 i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
2136 void i915_global_gtt_cleanup(struct drm_device *dev)
2138 struct drm_i915_private *dev_priv = dev->dev_private;
2139 struct i915_address_space *vm = &dev_priv->gtt.base;
2141 if (dev_priv->mm.aliasing_ppgtt) {
2142 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
2144 ppgtt->base.cleanup(&ppgtt->base);
2147 if (drm_mm_initialized(&vm->mm)) {
2148 if (intel_vgpu_active(dev))
2149 intel_vgt_deballoon();
2151 drm_mm_takedown(&vm->mm);
2152 list_del(&vm->global_link);
2158 static unsigned int gen6_get_total_gtt_size(u16 snb_gmch_ctl)
2160 snb_gmch_ctl >>= SNB_GMCH_GGMS_SHIFT;
2161 snb_gmch_ctl &= SNB_GMCH_GGMS_MASK;
2162 return snb_gmch_ctl << 20;
2165 static unsigned int gen8_get_total_gtt_size(u16 bdw_gmch_ctl)
2167 bdw_gmch_ctl >>= BDW_GMCH_GGMS_SHIFT;
2168 bdw_gmch_ctl &= BDW_GMCH_GGMS_MASK;
2170 bdw_gmch_ctl = 1 << bdw_gmch_ctl;
2172 #ifdef CONFIG_X86_32
2173 /* Limit 32b platforms to a 2GB GGTT: 4 << 20 / pte size * PAGE_SIZE */
2174 if (bdw_gmch_ctl > 4)
2178 return bdw_gmch_ctl << 20;
2181 static unsigned int chv_get_total_gtt_size(u16 gmch_ctrl)
2183 gmch_ctrl >>= SNB_GMCH_GGMS_SHIFT;
2184 gmch_ctrl &= SNB_GMCH_GGMS_MASK;
2187 return 1 << (20 + gmch_ctrl);
2192 static size_t gen6_get_stolen_size(u16 snb_gmch_ctl)
2194 snb_gmch_ctl >>= SNB_GMCH_GMS_SHIFT;
2195 snb_gmch_ctl &= SNB_GMCH_GMS_MASK;
2196 return snb_gmch_ctl << 25; /* 32 MB units */
2199 static size_t gen8_get_stolen_size(u16 bdw_gmch_ctl)
2201 bdw_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2202 bdw_gmch_ctl &= BDW_GMCH_GMS_MASK;
2203 return bdw_gmch_ctl << 25; /* 32 MB units */
2206 static size_t chv_get_stolen_size(u16 gmch_ctrl)
2208 gmch_ctrl >>= SNB_GMCH_GMS_SHIFT;
2209 gmch_ctrl &= SNB_GMCH_GMS_MASK;
2212 * 0x0 to 0x10: 32MB increments starting at 0MB
2213 * 0x11 to 0x16: 4MB increments starting at 8MB
2214 * 0x17 to 0x1d: 4MB increments start at 36MB
2216 if (gmch_ctrl < 0x11)
2217 return gmch_ctrl << 25;
2218 else if (gmch_ctrl < 0x17)
2219 return (gmch_ctrl - 0x11 + 2) << 22;
2221 return (gmch_ctrl - 0x17 + 9) << 22;
2224 static size_t gen9_get_stolen_size(u16 gen9_gmch_ctl)
2226 gen9_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2227 gen9_gmch_ctl &= BDW_GMCH_GMS_MASK;
2229 if (gen9_gmch_ctl < 0xf0)
2230 return gen9_gmch_ctl << 25; /* 32 MB units */
2232 /* 4MB increments starting at 0xf0 for 4MB */
2233 return (gen9_gmch_ctl - 0xf0 + 1) << 22;
2236 static int ggtt_probe_common(struct drm_device *dev,
2239 struct drm_i915_private *dev_priv = dev->dev_private;
2240 struct i915_page_scratch *scratch_page;
2241 phys_addr_t gtt_phys_addr;
2243 /* For Modern GENs the PTEs and register space are split in the BAR */
2244 gtt_phys_addr = pci_resource_start(dev->pdev, 0) +
2245 (pci_resource_len(dev->pdev, 0) / 2);
2248 * On BXT writes larger than 64 bit to the GTT pagetable range will be
2249 * dropped. For WC mappings in general we have 64 byte burst writes
2250 * when the WC buffer is flushed, so we can't use it, but have to
2251 * resort to an uncached mapping. The WC issue is easily caught by the
2252 * readback check when writing GTT PTE entries.
2254 if (IS_BROXTON(dev))
2255 dev_priv->gtt.gsm = ioremap_nocache(gtt_phys_addr, gtt_size);
2257 dev_priv->gtt.gsm = ioremap_wc(gtt_phys_addr, gtt_size);
2258 if (!dev_priv->gtt.gsm) {
2259 DRM_ERROR("Failed to map the gtt page table\n");
2263 scratch_page = alloc_scratch_page(dev);
2264 if (IS_ERR(scratch_page)) {
2265 DRM_ERROR("Scratch setup failed\n");
2266 /* iounmap will also get called at remove, but meh */
2267 iounmap(dev_priv->gtt.gsm);
2268 return PTR_ERR(scratch_page);
2271 dev_priv->gtt.base.scratch_page = scratch_page;
2276 /* The GGTT and PPGTT need a private PPAT setup in order to handle cacheability
2277 * bits. When using advanced contexts each context stores its own PAT, but
2278 * writing this data shouldn't be harmful even in those cases. */
2279 static void bdw_setup_private_ppat(struct drm_i915_private *dev_priv)
2283 pat = GEN8_PPAT(0, GEN8_PPAT_WB | GEN8_PPAT_LLC) | /* for normal objects, no eLLC */
2284 GEN8_PPAT(1, GEN8_PPAT_WC | GEN8_PPAT_LLCELLC) | /* for something pointing to ptes? */
2285 GEN8_PPAT(2, GEN8_PPAT_WT | GEN8_PPAT_LLCELLC) | /* for scanout with eLLC */
2286 GEN8_PPAT(3, GEN8_PPAT_UC) | /* Uncached objects, mostly for scanout */
2287 GEN8_PPAT(4, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(0)) |
2288 GEN8_PPAT(5, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(1)) |
2289 GEN8_PPAT(6, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(2)) |
2290 GEN8_PPAT(7, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3));
2292 if (!USES_PPGTT(dev_priv->dev))
2293 /* Spec: "For GGTT, there is NO pat_sel[2:0] from the entry,
2294 * so RTL will always use the value corresponding to
2296 * So let's disable cache for GGTT to avoid screen corruptions.
2297 * MOCS still can be used though.
2298 * - System agent ggtt writes (i.e. cpu gtt mmaps) already work
2299 * before this patch, i.e. the same uncached + snooping access
2300 * like on gen6/7 seems to be in effect.
2301 * - So this just fixes blitter/render access. Again it looks
2302 * like it's not just uncached access, but uncached + snooping.
2303 * So we can still hold onto all our assumptions wrt cpu
2304 * clflushing on LLC machines.
2306 pat = GEN8_PPAT(0, GEN8_PPAT_UC);
2308 /* XXX: spec defines this as 2 distinct registers. It's unclear if a 64b
2309 * write would work. */
2310 I915_WRITE(GEN8_PRIVATE_PAT, pat);
2311 I915_WRITE(GEN8_PRIVATE_PAT + 4, pat >> 32);
2314 static void chv_setup_private_ppat(struct drm_i915_private *dev_priv)
2319 * Map WB on BDW to snooped on CHV.
2321 * Only the snoop bit has meaning for CHV, the rest is
2324 * The hardware will never snoop for certain types of accesses:
2325 * - CPU GTT (GMADR->GGTT->no snoop->memory)
2326 * - PPGTT page tables
2327 * - some other special cycles
2329 * As with BDW, we also need to consider the following for GT accesses:
2330 * "For GGTT, there is NO pat_sel[2:0] from the entry,
2331 * so RTL will always use the value corresponding to
2333 * Which means we must set the snoop bit in PAT entry 0
2334 * in order to keep the global status page working.
2336 pat = GEN8_PPAT(0, CHV_PPAT_SNOOP) |
2340 GEN8_PPAT(4, CHV_PPAT_SNOOP) |
2341 GEN8_PPAT(5, CHV_PPAT_SNOOP) |
2342 GEN8_PPAT(6, CHV_PPAT_SNOOP) |
2343 GEN8_PPAT(7, CHV_PPAT_SNOOP);
2345 I915_WRITE(GEN8_PRIVATE_PAT, pat);
2346 I915_WRITE(GEN8_PRIVATE_PAT + 4, pat >> 32);
2349 static int gen8_gmch_probe(struct drm_device *dev,
2352 phys_addr_t *mappable_base,
2355 struct drm_i915_private *dev_priv = dev->dev_private;
2360 /* TODO: We're not aware of mappable constraints on gen8 yet */
2361 *mappable_base = pci_resource_start(dev->pdev, 2);
2362 *mappable_end = pci_resource_len(dev->pdev, 2);
2364 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(39)))
2365 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(39));
2367 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
2369 if (INTEL_INFO(dev)->gen >= 9) {
2370 *stolen = gen9_get_stolen_size(snb_gmch_ctl);
2371 gtt_size = gen8_get_total_gtt_size(snb_gmch_ctl);
2372 } else if (IS_CHERRYVIEW(dev)) {
2373 *stolen = chv_get_stolen_size(snb_gmch_ctl);
2374 gtt_size = chv_get_total_gtt_size(snb_gmch_ctl);
2376 *stolen = gen8_get_stolen_size(snb_gmch_ctl);
2377 gtt_size = gen8_get_total_gtt_size(snb_gmch_ctl);
2380 *gtt_total = (gtt_size / sizeof(gen8_pte_t)) << PAGE_SHIFT;
2382 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
2383 chv_setup_private_ppat(dev_priv);
2385 bdw_setup_private_ppat(dev_priv);
2387 ret = ggtt_probe_common(dev, gtt_size);
2389 dev_priv->gtt.base.clear_range = gen8_ggtt_clear_range;
2390 dev_priv->gtt.base.insert_entries = gen8_ggtt_insert_entries;
2391 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
2392 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
2397 static int gen6_gmch_probe(struct drm_device *dev,
2400 phys_addr_t *mappable_base,
2403 struct drm_i915_private *dev_priv = dev->dev_private;
2404 unsigned int gtt_size;
2408 *mappable_base = pci_resource_start(dev->pdev, 2);
2409 *mappable_end = pci_resource_len(dev->pdev, 2);
2411 /* 64/512MB is the current min/max we actually know of, but this is just
2412 * a coarse sanity check.
2414 if ((*mappable_end < (64<<20) || (*mappable_end > (512<<20)))) {
2415 DRM_ERROR("Unknown GMADR size (%llx)\n",
2416 dev_priv->gtt.mappable_end);
2420 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(40)))
2421 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(40));
2422 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
2424 *stolen = gen6_get_stolen_size(snb_gmch_ctl);
2426 gtt_size = gen6_get_total_gtt_size(snb_gmch_ctl);
2427 *gtt_total = (gtt_size / sizeof(gen6_pte_t)) << PAGE_SHIFT;
2429 ret = ggtt_probe_common(dev, gtt_size);
2431 dev_priv->gtt.base.clear_range = gen6_ggtt_clear_range;
2432 dev_priv->gtt.base.insert_entries = gen6_ggtt_insert_entries;
2433 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
2434 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
2439 static void gen6_gmch_remove(struct i915_address_space *vm)
2442 struct i915_gtt *gtt = container_of(vm, struct i915_gtt, base);
2445 free_scratch_page(vm->dev, vm->scratch_page);
2448 static int i915_gmch_probe(struct drm_device *dev,
2451 phys_addr_t *mappable_base,
2454 struct drm_i915_private *dev_priv = dev->dev_private;
2457 ret = intel_gmch_probe(dev_priv->bridge_dev, dev_priv->dev->pdev, NULL);
2459 DRM_ERROR("failed to set up gmch\n");
2463 intel_gtt_get(gtt_total, stolen, mappable_base, mappable_end);
2465 dev_priv->gtt.do_idle_maps = needs_idle_maps(dev_priv->dev);
2466 dev_priv->gtt.base.insert_entries = i915_ggtt_insert_entries;
2467 dev_priv->gtt.base.clear_range = i915_ggtt_clear_range;
2468 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
2469 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
2471 if (unlikely(dev_priv->gtt.do_idle_maps))
2472 DRM_INFO("applying Ironlake quirks for intel_iommu\n");
2477 static void i915_gmch_remove(struct i915_address_space *vm)
2479 intel_gmch_remove();
2482 int i915_gem_gtt_init(struct drm_device *dev)
2484 struct drm_i915_private *dev_priv = dev->dev_private;
2485 struct i915_gtt *gtt = &dev_priv->gtt;
2488 if (INTEL_INFO(dev)->gen <= 5) {
2489 gtt->gtt_probe = i915_gmch_probe;
2490 gtt->base.cleanup = i915_gmch_remove;
2491 } else if (INTEL_INFO(dev)->gen < 8) {
2492 gtt->gtt_probe = gen6_gmch_probe;
2493 gtt->base.cleanup = gen6_gmch_remove;
2494 if (IS_HASWELL(dev) && dev_priv->ellc_size)
2495 gtt->base.pte_encode = iris_pte_encode;
2496 else if (IS_HASWELL(dev))
2497 gtt->base.pte_encode = hsw_pte_encode;
2498 else if (IS_VALLEYVIEW(dev))
2499 gtt->base.pte_encode = byt_pte_encode;
2500 else if (INTEL_INFO(dev)->gen >= 7)
2501 gtt->base.pte_encode = ivb_pte_encode;
2503 gtt->base.pte_encode = snb_pte_encode;
2505 dev_priv->gtt.gtt_probe = gen8_gmch_probe;
2506 dev_priv->gtt.base.cleanup = gen6_gmch_remove;
2509 gtt->base.dev = dev;
2511 ret = gtt->gtt_probe(dev, >t->base.total, >t->stolen_size,
2512 >t->mappable_base, >t->mappable_end);
2516 /* GMADR is the PCI mmio aperture into the global GTT. */
2517 DRM_INFO("Memory usable by graphics device = %lluM\n",
2518 gtt->base.total >> 20);
2519 DRM_DEBUG_DRIVER("GMADR size = %lldM\n", gtt->mappable_end >> 20);
2520 DRM_DEBUG_DRIVER("GTT stolen size = %zdM\n", gtt->stolen_size >> 20);
2521 #ifdef CONFIG_INTEL_IOMMU
2522 if (intel_iommu_gfx_mapped)
2523 DRM_INFO("VT-d active for gfx access\n");
2526 * i915.enable_ppgtt is read-only, so do an early pass to validate the
2527 * user's requested state against the hardware/driver capabilities. We
2528 * do this now so that we can print out any log messages once rather
2529 * than every time we check intel_enable_ppgtt().
2531 i915.enable_ppgtt = sanitize_enable_ppgtt(dev, i915.enable_ppgtt);
2532 DRM_DEBUG_DRIVER("ppgtt mode: %i\n", i915.enable_ppgtt);
2537 void i915_gem_restore_gtt_mappings(struct drm_device *dev)
2539 struct drm_i915_private *dev_priv = dev->dev_private;
2540 struct drm_i915_gem_object *obj;
2541 struct i915_address_space *vm;
2543 i915_check_and_clear_faults(dev);
2545 /* First fill our portion of the GTT with scratch pages */
2546 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
2547 dev_priv->gtt.base.start,
2548 dev_priv->gtt.base.total,
2551 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
2552 struct i915_vma *vma = i915_gem_obj_to_vma(obj,
2553 &dev_priv->gtt.base);
2557 i915_gem_clflush_object(obj, obj->pin_display);
2558 WARN_ON(i915_vma_bind(vma, obj->cache_level, PIN_UPDATE));
2562 if (INTEL_INFO(dev)->gen >= 8) {
2563 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
2564 chv_setup_private_ppat(dev_priv);
2566 bdw_setup_private_ppat(dev_priv);
2571 if (USES_PPGTT(dev)) {
2572 list_for_each_entry(vm, &dev_priv->vm_list, global_link) {
2573 /* TODO: Perhaps it shouldn't be gen6 specific */
2575 struct i915_hw_ppgtt *ppgtt =
2576 container_of(vm, struct i915_hw_ppgtt,
2579 if (i915_is_ggtt(vm))
2580 ppgtt = dev_priv->mm.aliasing_ppgtt;
2582 gen6_write_page_range(dev_priv, &ppgtt->pd,
2583 0, ppgtt->base.total);
2587 i915_ggtt_flush(dev_priv);
2590 static struct i915_vma *
2591 __i915_gem_vma_create(struct drm_i915_gem_object *obj,
2592 struct i915_address_space *vm,
2593 const struct i915_ggtt_view *ggtt_view)
2595 struct i915_vma *vma;
2597 if (WARN_ON(i915_is_ggtt(vm) != !!ggtt_view))
2598 return ERR_PTR(-EINVAL);
2600 vma = kmem_cache_zalloc(to_i915(obj->base.dev)->vmas, GFP_KERNEL);
2602 return ERR_PTR(-ENOMEM);
2604 INIT_LIST_HEAD(&vma->vma_link);
2605 INIT_LIST_HEAD(&vma->mm_list);
2606 INIT_LIST_HEAD(&vma->exec_list);
2610 if (i915_is_ggtt(vm))
2611 vma->ggtt_view = *ggtt_view;
2613 list_add_tail(&vma->vma_link, &obj->vma_list);
2614 if (!i915_is_ggtt(vm))
2615 i915_ppgtt_get(i915_vm_to_ppgtt(vm));
2621 i915_gem_obj_lookup_or_create_vma(struct drm_i915_gem_object *obj,
2622 struct i915_address_space *vm)
2624 struct i915_vma *vma;
2626 vma = i915_gem_obj_to_vma(obj, vm);
2628 vma = __i915_gem_vma_create(obj, vm,
2629 i915_is_ggtt(vm) ? &i915_ggtt_view_normal : NULL);
2635 i915_gem_obj_lookup_or_create_ggtt_vma(struct drm_i915_gem_object *obj,
2636 const struct i915_ggtt_view *view)
2638 struct i915_address_space *ggtt = i915_obj_to_ggtt(obj);
2639 struct i915_vma *vma;
2642 return ERR_PTR(-EINVAL);
2644 vma = i915_gem_obj_to_ggtt_view(obj, view);
2650 vma = __i915_gem_vma_create(obj, ggtt, view);
2657 rotate_pages(dma_addr_t *in, unsigned int width, unsigned int height,
2658 struct sg_table *st)
2660 unsigned int column, row;
2661 unsigned int src_idx;
2662 struct scatterlist *sg = st->sgl;
2666 for (column = 0; column < width; column++) {
2667 src_idx = width * (height - 1) + column;
2668 for (row = 0; row < height; row++) {
2670 /* We don't need the pages, but need to initialize
2671 * the entries so the sg list can be happily traversed.
2672 * The only thing we need are DMA addresses.
2674 sg_set_page(sg, NULL, PAGE_SIZE, 0);
2675 sg_dma_address(sg) = in[src_idx];
2676 sg_dma_len(sg) = PAGE_SIZE;
2683 static struct sg_table *
2684 intel_rotate_fb_obj_pages(struct i915_ggtt_view *ggtt_view,
2685 struct drm_i915_gem_object *obj)
2687 struct intel_rotation_info *rot_info = &ggtt_view->rotation_info;
2688 unsigned int size_pages = rot_info->size >> PAGE_SHIFT;
2689 struct sg_page_iter sg_iter;
2691 dma_addr_t *page_addr_list;
2692 struct sg_table *st;
2695 /* Allocate a temporary list of source pages for random access. */
2696 page_addr_list = drm_malloc_ab(obj->base.size / PAGE_SIZE,
2697 sizeof(dma_addr_t));
2698 if (!page_addr_list)
2699 return ERR_PTR(ret);
2701 /* Allocate target SG list. */
2702 st = kmalloc(sizeof(*st), GFP_KERNEL);
2706 ret = sg_alloc_table(st, size_pages, GFP_KERNEL);
2710 /* Populate source page list from the object. */
2712 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) {
2713 page_addr_list[i] = sg_page_iter_dma_address(&sg_iter);
2717 /* Rotate the pages. */
2718 rotate_pages(page_addr_list,
2719 rot_info->width_pages, rot_info->height_pages,
2723 "Created rotated page mapping for object size %zu (pitch=%u, height=%u, pixel_format=0x%x, %ux%u tiles, %u pages).\n",
2724 obj->base.size, rot_info->pitch, rot_info->height,
2725 rot_info->pixel_format, rot_info->width_pages,
2726 rot_info->height_pages, size_pages);
2728 drm_free_large(page_addr_list);
2735 drm_free_large(page_addr_list);
2738 "Failed to create rotated mapping for object size %zu! (%d) (pitch=%u, height=%u, pixel_format=0x%x, %ux%u tiles, %u pages)\n",
2739 obj->base.size, ret, rot_info->pitch, rot_info->height,
2740 rot_info->pixel_format, rot_info->width_pages,
2741 rot_info->height_pages, size_pages);
2742 return ERR_PTR(ret);
2745 static struct sg_table *
2746 intel_partial_pages(const struct i915_ggtt_view *view,
2747 struct drm_i915_gem_object *obj)
2749 struct sg_table *st;
2750 struct scatterlist *sg;
2751 struct sg_page_iter obj_sg_iter;
2754 st = kmalloc(sizeof(*st), GFP_KERNEL);
2758 ret = sg_alloc_table(st, view->params.partial.size, GFP_KERNEL);
2764 for_each_sg_page(obj->pages->sgl, &obj_sg_iter, obj->pages->nents,
2765 view->params.partial.offset)
2767 if (st->nents >= view->params.partial.size)
2770 sg_set_page(sg, NULL, PAGE_SIZE, 0);
2771 sg_dma_address(sg) = sg_page_iter_dma_address(&obj_sg_iter);
2772 sg_dma_len(sg) = PAGE_SIZE;
2783 return ERR_PTR(ret);
2787 i915_get_ggtt_vma_pages(struct i915_vma *vma)
2791 if (vma->ggtt_view.pages)
2794 if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL)
2795 vma->ggtt_view.pages = vma->obj->pages;
2796 else if (vma->ggtt_view.type == I915_GGTT_VIEW_ROTATED)
2797 vma->ggtt_view.pages =
2798 intel_rotate_fb_obj_pages(&vma->ggtt_view, vma->obj);
2799 else if (vma->ggtt_view.type == I915_GGTT_VIEW_PARTIAL)
2800 vma->ggtt_view.pages =
2801 intel_partial_pages(&vma->ggtt_view, vma->obj);
2803 WARN_ONCE(1, "GGTT view %u not implemented!\n",
2804 vma->ggtt_view.type);
2806 if (!vma->ggtt_view.pages) {
2807 DRM_ERROR("Failed to get pages for GGTT view type %u!\n",
2808 vma->ggtt_view.type);
2810 } else if (IS_ERR(vma->ggtt_view.pages)) {
2811 ret = PTR_ERR(vma->ggtt_view.pages);
2812 vma->ggtt_view.pages = NULL;
2813 DRM_ERROR("Failed to get pages for VMA view type %u (%d)!\n",
2814 vma->ggtt_view.type, ret);
2821 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
2823 * @cache_level: mapping cache level
2824 * @flags: flags like global or local mapping
2826 * DMA addresses are taken from the scatter-gather table of this object (or of
2827 * this VMA in case of non-default GGTT views) and PTE entries set up.
2828 * Note that DMA addresses are also the only part of the SG table we care about.
2830 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
2836 if (WARN_ON(flags == 0))
2840 if (flags & PIN_GLOBAL)
2841 bind_flags |= GLOBAL_BIND;
2842 if (flags & PIN_USER)
2843 bind_flags |= LOCAL_BIND;
2845 if (flags & PIN_UPDATE)
2846 bind_flags |= vma->bound;
2848 bind_flags &= ~vma->bound;
2850 if (bind_flags == 0)
2853 if (vma->bound == 0 && vma->vm->allocate_va_range) {
2854 trace_i915_va_alloc(vma->vm,
2857 VM_TO_TRACE_NAME(vma->vm));
2859 /* XXX: i915_vma_pin() will fix this +- hack */
2861 ret = vma->vm->allocate_va_range(vma->vm,
2869 ret = vma->vm->bind_vma(vma, cache_level, bind_flags);
2873 vma->bound |= bind_flags;
2879 * i915_ggtt_view_size - Get the size of a GGTT view.
2880 * @obj: Object the view is of.
2881 * @view: The view in question.
2883 * @return The size of the GGTT view in bytes.
2886 i915_ggtt_view_size(struct drm_i915_gem_object *obj,
2887 const struct i915_ggtt_view *view)
2889 if (view->type == I915_GGTT_VIEW_NORMAL) {
2890 return obj->base.size;
2891 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
2892 return view->rotation_info.size;
2893 } else if (view->type == I915_GGTT_VIEW_PARTIAL) {
2894 return view->params.partial.size << PAGE_SHIFT;
2896 WARN_ONCE(1, "GGTT view %u not implemented!\n", view->type);
2897 return obj->base.size;