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"
34 #include <linux/bitmap.h>
35 #include <linux/highmem.h>
38 * DOC: Global GTT views
40 * Background and previous state
42 * Historically objects could exists (be bound) in global GTT space only as
43 * singular instances with a view representing all of the object's backing pages
44 * in a linear fashion. This view will be called a normal view.
46 * To support multiple views of the same object, where the number of mapped
47 * pages is not equal to the backing store, or where the layout of the pages
48 * is not linear, concept of a GGTT view was added.
50 * One example of an alternative view is a stereo display driven by a single
51 * image. In this case we would have a framebuffer looking like this
57 * Above would represent a normal GGTT view as normally mapped for GPU or CPU
58 * rendering. In contrast, fed to the display engine would be an alternative
59 * view which could look something like this:
64 * In this example both the size and layout of pages in the alternative view is
65 * different from the normal view.
67 * Implementation and usage
69 * GGTT views are implemented using VMAs and are distinguished via enum
70 * i915_ggtt_view_type and struct i915_ggtt_view.
72 * A new flavour of core GEM functions which work with GGTT bound objects were
73 * added with the _ggtt_ infix, and sometimes with _view postfix to avoid
74 * renaming in large amounts of code. They take the struct i915_ggtt_view
75 * parameter encapsulating all metadata required to implement a view.
77 * As a helper for callers which are only interested in the normal view,
78 * globally const i915_ggtt_view_normal singleton instance exists. All old core
79 * GEM API functions, the ones not taking the view parameter, are operating on,
80 * or with the normal GGTT view.
82 * Code wanting to add or use a new GGTT view needs to:
84 * 1. Add a new enum with a suitable name.
85 * 2. Extend the metadata in the i915_ggtt_view structure if required.
86 * 3. Add support to i915_get_vma_pages().
88 * New views are required to build a scatter-gather table from within the
89 * i915_get_vma_pages function. This table is stored in the vma.ggtt_view and
90 * exists for the lifetime of an VMA.
92 * Core API is designed to have copy semantics which means that passed in
93 * struct i915_ggtt_view does not need to be persistent (left around after
94 * calling the core API functions).
99 i915_get_ggtt_vma_pages(struct i915_vma *vma);
101 const struct i915_ggtt_view i915_ggtt_view_normal;
102 const struct i915_ggtt_view i915_ggtt_view_rotated = {
103 .type = I915_GGTT_VIEW_ROTATED
106 static int sanitize_enable_ppgtt(struct drm_device *dev, int enable_ppgtt)
108 bool has_aliasing_ppgtt;
111 has_aliasing_ppgtt = INTEL_INFO(dev)->gen >= 6;
112 has_full_ppgtt = INTEL_INFO(dev)->gen >= 7;
114 if (intel_vgpu_active(dev))
115 has_full_ppgtt = false; /* emulation is too hard */
118 * We don't allow disabling PPGTT for gen9+ as it's a requirement for
119 * execlists, the sole mechanism available to submit work.
121 if (INTEL_INFO(dev)->gen < 9 &&
122 (enable_ppgtt == 0 || !has_aliasing_ppgtt))
125 if (enable_ppgtt == 1)
128 if (enable_ppgtt == 2 && has_full_ppgtt)
131 #ifdef CONFIG_INTEL_IOMMU
132 /* Disable ppgtt on SNB if VT-d is on. */
133 if (INTEL_INFO(dev)->gen == 6 && intel_iommu_gfx_mapped) {
134 DRM_INFO("Disabling PPGTT because VT-d is on\n");
139 /* Early VLV doesn't have this */
140 if (IS_VALLEYVIEW(dev) && !IS_CHERRYVIEW(dev) &&
141 dev->pdev->revision < 0xb) {
142 DRM_DEBUG_DRIVER("disabling PPGTT on pre-B3 step VLV\n");
146 if (INTEL_INFO(dev)->gen >= 8 && i915.enable_execlists)
149 return has_aliasing_ppgtt ? 1 : 0;
152 static int ppgtt_bind_vma(struct i915_vma *vma,
153 enum i915_cache_level cache_level,
158 /* Currently applicable only to VLV */
160 pte_flags |= PTE_READ_ONLY;
162 vma->vm->insert_entries(vma->vm, vma->obj->pages, vma->node.start,
163 cache_level, pte_flags);
168 static void ppgtt_unbind_vma(struct i915_vma *vma)
170 vma->vm->clear_range(vma->vm,
176 static gen8_pte_t gen8_pte_encode(dma_addr_t addr,
177 enum i915_cache_level level,
180 gen8_pte_t pte = valid ? _PAGE_PRESENT | _PAGE_RW : 0;
184 case I915_CACHE_NONE:
185 pte |= PPAT_UNCACHED_INDEX;
188 pte |= PPAT_DISPLAY_ELLC_INDEX;
191 pte |= PPAT_CACHED_INDEX;
198 static gen8_pde_t gen8_pde_encode(const dma_addr_t addr,
199 const enum i915_cache_level level)
201 gen8_pde_t pde = _PAGE_PRESENT | _PAGE_RW;
203 if (level != I915_CACHE_NONE)
204 pde |= PPAT_CACHED_PDE_INDEX;
206 pde |= PPAT_UNCACHED_INDEX;
210 static gen6_pte_t snb_pte_encode(dma_addr_t addr,
211 enum i915_cache_level level,
212 bool valid, u32 unused)
214 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
215 pte |= GEN6_PTE_ADDR_ENCODE(addr);
218 case I915_CACHE_L3_LLC:
220 pte |= GEN6_PTE_CACHE_LLC;
222 case I915_CACHE_NONE:
223 pte |= GEN6_PTE_UNCACHED;
232 static gen6_pte_t ivb_pte_encode(dma_addr_t addr,
233 enum i915_cache_level level,
234 bool valid, u32 unused)
236 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
237 pte |= GEN6_PTE_ADDR_ENCODE(addr);
240 case I915_CACHE_L3_LLC:
241 pte |= GEN7_PTE_CACHE_L3_LLC;
244 pte |= GEN6_PTE_CACHE_LLC;
246 case I915_CACHE_NONE:
247 pte |= GEN6_PTE_UNCACHED;
256 static gen6_pte_t byt_pte_encode(dma_addr_t addr,
257 enum i915_cache_level level,
258 bool valid, u32 flags)
260 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
261 pte |= GEN6_PTE_ADDR_ENCODE(addr);
263 if (!(flags & PTE_READ_ONLY))
264 pte |= BYT_PTE_WRITEABLE;
266 if (level != I915_CACHE_NONE)
267 pte |= BYT_PTE_SNOOPED_BY_CPU_CACHES;
272 static gen6_pte_t hsw_pte_encode(dma_addr_t addr,
273 enum i915_cache_level level,
274 bool valid, u32 unused)
276 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
277 pte |= HSW_PTE_ADDR_ENCODE(addr);
279 if (level != I915_CACHE_NONE)
280 pte |= HSW_WB_LLC_AGE3;
285 static gen6_pte_t iris_pte_encode(dma_addr_t addr,
286 enum i915_cache_level level,
287 bool valid, u32 unused)
289 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
290 pte |= HSW_PTE_ADDR_ENCODE(addr);
293 case I915_CACHE_NONE:
296 pte |= HSW_WT_ELLC_LLC_AGE3;
299 pte |= HSW_WB_ELLC_LLC_AGE3;
306 static int __setup_page_dma(struct drm_device *dev,
307 struct i915_page_dma *p, gfp_t flags)
309 struct device *device = dev->pdev->dev;
311 p->page = alloc_page(flags);
315 p->daddr = dma_map_page(device,
316 p->page, 0, 4096, PCI_DMA_BIDIRECTIONAL);
318 if (dma_mapping_error(device, p->daddr)) {
319 __free_page(p->page);
326 static int setup_page_dma(struct drm_device *dev, struct i915_page_dma *p)
328 return __setup_page_dma(dev, p, GFP_KERNEL);
331 static void cleanup_page_dma(struct drm_device *dev, struct i915_page_dma *p)
333 if (WARN_ON(!p->page))
336 dma_unmap_page(dev->pdev->dev, p->daddr, 4096, PCI_DMA_BIDIRECTIONAL);
337 __free_page(p->page);
338 memset(p, 0, sizeof(*p));
341 static void *kmap_page_dma(struct i915_page_dma *p)
343 return kmap_atomic(p->page);
346 /* We use the flushing unmap only with ppgtt structures:
347 * page directories, page tables and scratch pages.
349 static void kunmap_page_dma(struct drm_device *dev, void *vaddr)
351 /* There are only few exceptions for gen >=6. chv and bxt.
352 * And we are not sure about the latter so play safe for now.
354 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
355 drm_clflush_virt_range(vaddr, PAGE_SIZE);
357 kunmap_atomic(vaddr);
360 #define kmap_px(px) kmap_page_dma(px_base(px))
361 #define kunmap_px(ppgtt, vaddr) kunmap_page_dma((ppgtt)->base.dev, (vaddr))
363 #define setup_px(dev, px) setup_page_dma((dev), px_base(px))
364 #define cleanup_px(dev, px) cleanup_page_dma((dev), px_base(px))
365 #define fill_px(dev, px, v) fill_page_dma((dev), px_base(px), (v))
366 #define fill32_px(dev, px, v) fill_page_dma_32((dev), px_base(px), (v))
368 static void fill_page_dma(struct drm_device *dev, struct i915_page_dma *p,
372 uint64_t * const vaddr = kmap_page_dma(p);
374 for (i = 0; i < 512; i++)
377 kunmap_page_dma(dev, vaddr);
380 static void fill_page_dma_32(struct drm_device *dev, struct i915_page_dma *p,
381 const uint32_t val32)
387 fill_page_dma(dev, p, v);
390 static struct i915_page_scratch *alloc_scratch_page(struct drm_device *dev)
392 struct i915_page_scratch *sp;
395 sp = kzalloc(sizeof(*sp), GFP_KERNEL);
397 return ERR_PTR(-ENOMEM);
399 ret = __setup_page_dma(dev, px_base(sp), GFP_DMA32 | __GFP_ZERO);
405 set_pages_uc(px_page(sp), 1);
410 static void free_scratch_page(struct drm_device *dev,
411 struct i915_page_scratch *sp)
413 set_pages_wb(px_page(sp), 1);
419 static struct i915_page_table *alloc_pt(struct drm_device *dev)
421 struct i915_page_table *pt;
422 const size_t count = INTEL_INFO(dev)->gen >= 8 ?
423 GEN8_PTES : GEN6_PTES;
426 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
428 return ERR_PTR(-ENOMEM);
430 pt->used_ptes = kcalloc(BITS_TO_LONGS(count), sizeof(*pt->used_ptes),
436 ret = setup_px(dev, pt);
443 kfree(pt->used_ptes);
450 static void free_pt(struct drm_device *dev, struct i915_page_table *pt)
453 kfree(pt->used_ptes);
458 * alloc_pt_range() - Allocate a multiple page tables
459 * @pd: The page directory which will have at least @count entries
460 * available to point to the allocated page tables.
461 * @pde: First page directory entry for which we are allocating.
462 * @count: Number of pages to allocate.
465 * Allocates multiple page table pages and sets the appropriate entries in the
466 * page table structure within the page directory. Function cleans up after
467 * itself on any failures.
469 * Return: 0 if allocation succeeded.
471 static int alloc_pt_range(struct i915_page_directory *pd, uint16_t pde, size_t count,
472 struct drm_device *dev)
476 /* 512 is the max page tables per page_directory on any platform. */
477 if (WARN_ON(pde + count > I915_PDES))
480 for (i = pde; i < pde + count; i++) {
481 struct i915_page_table *pt = alloc_pt(dev);
487 WARN(pd->page_table[i],
488 "Leaking page directory entry %d (%p)\n",
489 i, pd->page_table[i]);
490 pd->page_table[i] = pt;
497 free_pt(dev, pd->page_table[i]);
501 static void gen8_initialize_pt(struct i915_address_space *vm,
502 struct i915_page_table *pt)
504 gen8_pte_t scratch_pte;
506 scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
507 I915_CACHE_LLC, true);
509 fill_px(vm->dev, pt, scratch_pte);
512 static void gen6_initialize_pt(struct i915_address_space *vm,
513 struct i915_page_table *pt)
515 gen6_pte_t scratch_pte;
517 WARN_ON(px_dma(vm->scratch_page) == 0);
519 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
520 I915_CACHE_LLC, true, 0);
522 fill32_px(vm->dev, pt, scratch_pte);
525 static struct i915_page_directory *alloc_pd(struct drm_device *dev)
527 struct i915_page_directory *pd;
530 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
532 return ERR_PTR(-ENOMEM);
534 pd->used_pdes = kcalloc(BITS_TO_LONGS(I915_PDES),
535 sizeof(*pd->used_pdes), GFP_KERNEL);
539 ret = setup_px(dev, pd);
546 kfree(pd->used_pdes);
553 static void free_pd(struct drm_device *dev, struct i915_page_directory *pd)
557 kfree(pd->used_pdes);
562 static void gen8_initialize_pd(struct i915_address_space *vm,
563 struct i915_page_directory *pd)
565 gen8_pde_t scratch_pde;
567 scratch_pde = gen8_pde_encode(px_dma(vm->scratch_pt), I915_CACHE_LLC);
569 fill_px(vm->dev, pd, scratch_pde);
572 /* Broadwell Page Directory Pointer Descriptors */
573 static int gen8_write_pdp(struct drm_i915_gem_request *req,
577 struct intel_engine_cs *ring = req->ring;
582 ret = intel_ring_begin(req, 6);
586 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
587 intel_ring_emit(ring, GEN8_RING_PDP_UDW(ring, entry));
588 intel_ring_emit(ring, upper_32_bits(addr));
589 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
590 intel_ring_emit(ring, GEN8_RING_PDP_LDW(ring, entry));
591 intel_ring_emit(ring, lower_32_bits(addr));
592 intel_ring_advance(ring);
597 static int gen8_mm_switch(struct i915_hw_ppgtt *ppgtt,
598 struct drm_i915_gem_request *req)
602 for (i = GEN8_LEGACY_PDPES - 1; i >= 0; i--) {
603 const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i);
605 ret = gen8_write_pdp(req, i, pd_daddr);
613 static void gen8_ppgtt_clear_range(struct i915_address_space *vm,
618 struct i915_hw_ppgtt *ppgtt =
619 container_of(vm, struct i915_hw_ppgtt, base);
620 gen8_pte_t *pt_vaddr, scratch_pte;
621 unsigned pdpe = start >> GEN8_PDPE_SHIFT & GEN8_PDPE_MASK;
622 unsigned pde = start >> GEN8_PDE_SHIFT & GEN8_PDE_MASK;
623 unsigned pte = start >> GEN8_PTE_SHIFT & GEN8_PTE_MASK;
624 unsigned num_entries = length >> PAGE_SHIFT;
625 unsigned last_pte, i;
627 scratch_pte = gen8_pte_encode(px_dma(ppgtt->base.scratch_page),
628 I915_CACHE_LLC, use_scratch);
630 while (num_entries) {
631 struct i915_page_directory *pd;
632 struct i915_page_table *pt;
634 if (WARN_ON(!ppgtt->pdp.page_directory[pdpe]))
637 pd = ppgtt->pdp.page_directory[pdpe];
639 if (WARN_ON(!pd->page_table[pde]))
642 pt = pd->page_table[pde];
644 if (WARN_ON(!px_page(pt)))
647 last_pte = pte + num_entries;
648 if (last_pte > GEN8_PTES)
649 last_pte = GEN8_PTES;
651 pt_vaddr = kmap_px(pt);
653 for (i = pte; i < last_pte; i++) {
654 pt_vaddr[i] = scratch_pte;
658 kunmap_px(ppgtt, pt);
661 if (++pde == I915_PDES) {
668 static void gen8_ppgtt_insert_entries(struct i915_address_space *vm,
669 struct sg_table *pages,
671 enum i915_cache_level cache_level, u32 unused)
673 struct i915_hw_ppgtt *ppgtt =
674 container_of(vm, struct i915_hw_ppgtt, base);
675 gen8_pte_t *pt_vaddr;
676 unsigned pdpe = start >> GEN8_PDPE_SHIFT & GEN8_PDPE_MASK;
677 unsigned pde = start >> GEN8_PDE_SHIFT & GEN8_PDE_MASK;
678 unsigned pte = start >> GEN8_PTE_SHIFT & GEN8_PTE_MASK;
679 struct sg_page_iter sg_iter;
683 for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) {
684 if (WARN_ON(pdpe >= GEN8_LEGACY_PDPES))
687 if (pt_vaddr == NULL) {
688 struct i915_page_directory *pd = ppgtt->pdp.page_directory[pdpe];
689 struct i915_page_table *pt = pd->page_table[pde];
691 pt_vaddr = kmap_px(pt);
695 gen8_pte_encode(sg_page_iter_dma_address(&sg_iter),
697 if (++pte == GEN8_PTES) {
698 kunmap_px(ppgtt, pt_vaddr);
700 if (++pde == I915_PDES) {
709 kunmap_px(ppgtt, pt_vaddr);
712 static void gen8_free_page_tables(struct drm_device *dev,
713 struct i915_page_directory *pd)
720 for_each_set_bit(i, pd->used_pdes, I915_PDES) {
721 if (WARN_ON(!pd->page_table[i]))
724 free_pt(dev, pd->page_table[i]);
725 pd->page_table[i] = NULL;
729 static int gen8_init_scratch(struct i915_address_space *vm)
731 struct drm_device *dev = vm->dev;
733 vm->scratch_page = alloc_scratch_page(dev);
734 if (IS_ERR(vm->scratch_page))
735 return PTR_ERR(vm->scratch_page);
737 vm->scratch_pt = alloc_pt(dev);
738 if (IS_ERR(vm->scratch_pt)) {
739 free_scratch_page(dev, vm->scratch_page);
740 return PTR_ERR(vm->scratch_pt);
743 vm->scratch_pd = alloc_pd(dev);
744 if (IS_ERR(vm->scratch_pd)) {
745 free_pt(dev, vm->scratch_pt);
746 free_scratch_page(dev, vm->scratch_page);
747 return PTR_ERR(vm->scratch_pd);
750 gen8_initialize_pt(vm, vm->scratch_pt);
751 gen8_initialize_pd(vm, vm->scratch_pd);
756 static void gen8_free_scratch(struct i915_address_space *vm)
758 struct drm_device *dev = vm->dev;
760 free_pd(dev, vm->scratch_pd);
761 free_pt(dev, vm->scratch_pt);
762 free_scratch_page(dev, vm->scratch_page);
765 static void gen8_ppgtt_cleanup(struct i915_address_space *vm)
767 struct i915_hw_ppgtt *ppgtt =
768 container_of(vm, struct i915_hw_ppgtt, base);
771 for_each_set_bit(i, ppgtt->pdp.used_pdpes, GEN8_LEGACY_PDPES) {
772 if (WARN_ON(!ppgtt->pdp.page_directory[i]))
775 gen8_free_page_tables(ppgtt->base.dev,
776 ppgtt->pdp.page_directory[i]);
777 free_pd(ppgtt->base.dev, ppgtt->pdp.page_directory[i]);
780 gen8_free_scratch(vm);
784 * gen8_ppgtt_alloc_pagetabs() - Allocate page tables for VA range.
785 * @ppgtt: Master ppgtt structure.
786 * @pd: Page directory for this address range.
787 * @start: Starting virtual address to begin allocations.
788 * @length Size of the allocations.
789 * @new_pts: Bitmap set by function with new allocations. Likely used by the
790 * caller to free on error.
792 * Allocate the required number of page tables. Extremely similar to
793 * gen8_ppgtt_alloc_page_directories(). The main difference is here we are limited by
794 * the page directory boundary (instead of the page directory pointer). That
795 * boundary is 1GB virtual. Therefore, unlike gen8_ppgtt_alloc_page_directories(), it is
796 * possible, and likely that the caller will need to use multiple calls of this
797 * function to achieve the appropriate allocation.
799 * Return: 0 if success; negative error code otherwise.
801 static int gen8_ppgtt_alloc_pagetabs(struct i915_hw_ppgtt *ppgtt,
802 struct i915_page_directory *pd,
805 unsigned long *new_pts)
807 struct drm_device *dev = ppgtt->base.dev;
808 struct i915_page_table *pt;
812 gen8_for_each_pde(pt, pd, start, length, temp, pde) {
813 /* Don't reallocate page tables */
815 /* Scratch is never allocated this way */
816 WARN_ON(pt == ppgtt->base.scratch_pt);
824 gen8_initialize_pt(&ppgtt->base, pt);
825 pd->page_table[pde] = pt;
826 __set_bit(pde, new_pts);
832 for_each_set_bit(pde, new_pts, I915_PDES)
833 free_pt(dev, pd->page_table[pde]);
839 * gen8_ppgtt_alloc_page_directories() - Allocate page directories for VA range.
840 * @ppgtt: Master ppgtt structure.
841 * @pdp: Page directory pointer for this address range.
842 * @start: Starting virtual address to begin allocations.
843 * @length Size of the allocations.
844 * @new_pds Bitmap set by function with new allocations. Likely used by the
845 * caller to free on error.
847 * Allocate the required number of page directories starting at the pde index of
848 * @start, and ending at the pde index @start + @length. This function will skip
849 * over already allocated page directories within the range, and only allocate
850 * new ones, setting the appropriate pointer within the pdp as well as the
851 * correct position in the bitmap @new_pds.
853 * The function will only allocate the pages within the range for a give page
854 * directory pointer. In other words, if @start + @length straddles a virtually
855 * addressed PDP boundary (512GB for 4k pages), there will be more allocations
856 * required by the caller, This is not currently possible, and the BUG in the
857 * code will prevent it.
859 * Return: 0 if success; negative error code otherwise.
861 static int gen8_ppgtt_alloc_page_directories(struct i915_hw_ppgtt *ppgtt,
862 struct i915_page_directory_pointer *pdp,
865 unsigned long *new_pds)
867 struct drm_device *dev = ppgtt->base.dev;
868 struct i915_page_directory *pd;
872 WARN_ON(!bitmap_empty(new_pds, GEN8_LEGACY_PDPES));
874 gen8_for_each_pdpe(pd, pdp, start, length, temp, pdpe) {
882 gen8_initialize_pd(&ppgtt->base, pd);
883 pdp->page_directory[pdpe] = pd;
884 __set_bit(pdpe, new_pds);
890 for_each_set_bit(pdpe, new_pds, GEN8_LEGACY_PDPES)
891 free_pd(dev, pdp->page_directory[pdpe]);
897 free_gen8_temp_bitmaps(unsigned long *new_pds, unsigned long **new_pts)
901 for (i = 0; i < GEN8_LEGACY_PDPES; i++)
907 /* Fills in the page directory bitmap, and the array of page tables bitmap. Both
908 * of these are based on the number of PDPEs in the system.
911 int __must_check alloc_gen8_temp_bitmaps(unsigned long **new_pds,
912 unsigned long ***new_pts)
918 pds = kcalloc(BITS_TO_LONGS(GEN8_LEGACY_PDPES), sizeof(unsigned long), GFP_KERNEL);
922 pts = kcalloc(GEN8_LEGACY_PDPES, sizeof(unsigned long *), GFP_KERNEL);
928 for (i = 0; i < GEN8_LEGACY_PDPES; i++) {
929 pts[i] = kcalloc(BITS_TO_LONGS(I915_PDES),
930 sizeof(unsigned long), GFP_KERNEL);
941 free_gen8_temp_bitmaps(pds, pts);
945 /* PDE TLBs are a pain to invalidate on GEN8+. When we modify
946 * the page table structures, we mark them dirty so that
947 * context switching/execlist queuing code takes extra steps
948 * to ensure that tlbs are flushed.
950 static void mark_tlbs_dirty(struct i915_hw_ppgtt *ppgtt)
952 ppgtt->pd_dirty_rings = INTEL_INFO(ppgtt->base.dev)->ring_mask;
955 static int gen8_alloc_va_range(struct i915_address_space *vm,
959 struct i915_hw_ppgtt *ppgtt =
960 container_of(vm, struct i915_hw_ppgtt, base);
961 unsigned long *new_page_dirs, **new_page_tables;
962 struct i915_page_directory *pd;
963 const uint64_t orig_start = start;
964 const uint64_t orig_length = length;
969 /* Wrap is never okay since we can only represent 48b, and we don't
970 * actually use the other side of the canonical address space.
972 if (WARN_ON(start + length < start))
975 if (WARN_ON(start + length > ppgtt->base.total))
978 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables);
982 /* Do the allocations first so we can easily bail out */
983 ret = gen8_ppgtt_alloc_page_directories(ppgtt, &ppgtt->pdp, start, length,
986 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
990 /* For every page directory referenced, allocate page tables */
991 gen8_for_each_pdpe(pd, &ppgtt->pdp, start, length, temp, pdpe) {
992 ret = gen8_ppgtt_alloc_pagetabs(ppgtt, pd, start, length,
993 new_page_tables[pdpe]);
999 length = orig_length;
1001 /* Allocations have completed successfully, so set the bitmaps, and do
1003 gen8_for_each_pdpe(pd, &ppgtt->pdp, start, length, temp, pdpe) {
1004 gen8_pde_t *const page_directory = kmap_px(pd);
1005 struct i915_page_table *pt;
1006 uint64_t pd_len = gen8_clamp_pd(start, length);
1007 uint64_t pd_start = start;
1010 /* Every pd should be allocated, we just did that above. */
1013 gen8_for_each_pde(pt, pd, pd_start, pd_len, temp, pde) {
1014 /* Same reasoning as pd */
1017 WARN_ON(!gen8_pte_count(pd_start, pd_len));
1019 /* Set our used ptes within the page table */
1020 bitmap_set(pt->used_ptes,
1021 gen8_pte_index(pd_start),
1022 gen8_pte_count(pd_start, pd_len));
1024 /* Our pde is now pointing to the pagetable, pt */
1025 __set_bit(pde, pd->used_pdes);
1027 /* Map the PDE to the page table */
1028 page_directory[pde] = gen8_pde_encode(px_dma(pt),
1031 /* NB: We haven't yet mapped ptes to pages. At this
1032 * point we're still relying on insert_entries() */
1035 kunmap_px(ppgtt, page_directory);
1037 __set_bit(pdpe, ppgtt->pdp.used_pdpes);
1040 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1041 mark_tlbs_dirty(ppgtt);
1046 for_each_set_bit(temp, new_page_tables[pdpe], I915_PDES)
1047 free_pt(vm->dev, ppgtt->pdp.page_directory[pdpe]->page_table[temp]);
1050 for_each_set_bit(pdpe, new_page_dirs, GEN8_LEGACY_PDPES)
1051 free_pd(vm->dev, ppgtt->pdp.page_directory[pdpe]);
1053 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables);
1054 mark_tlbs_dirty(ppgtt);
1059 * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers
1060 * with a net effect resembling a 2-level page table in normal x86 terms. Each
1061 * PDP represents 1GB of memory 4 * 512 * 512 * 4096 = 4GB legacy 32b address
1065 static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
1069 ret = gen8_init_scratch(&ppgtt->base);
1073 ppgtt->base.start = 0;
1074 #ifndef __DragonFly__
1075 ppgtt->base.total = 1ULL << 32;
1076 if (IS_ENABLED(CONFIG_X86_32))
1077 /* While we have a proliferation of size_t variables
1078 * we cannot represent the full ppgtt size on 32bit,
1079 * so limit it to the same size as the GGTT (currently
1083 ppgtt->base.total = to_i915(ppgtt->base.dev)->gtt.base.total;
1084 ppgtt->base.cleanup = gen8_ppgtt_cleanup;
1085 ppgtt->base.allocate_va_range = gen8_alloc_va_range;
1086 ppgtt->base.insert_entries = gen8_ppgtt_insert_entries;
1087 ppgtt->base.clear_range = gen8_ppgtt_clear_range;
1088 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
1089 ppgtt->base.bind_vma = ppgtt_bind_vma;
1091 ppgtt->switch_mm = gen8_mm_switch;
1096 static int gen8_aliasing_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
1098 struct drm_device *dev = ppgtt->base.dev;
1099 struct drm_i915_private *dev_priv = dev->dev_private;
1100 uint64_t start = 0, size = dev_priv->gtt.base.total;
1103 ret = gen8_ppgtt_init(ppgtt);
1107 /* Aliasing PPGTT has to always work and be mapped because of the way we
1108 * use RESTORE_INHIBIT in the context switch. This will be fixed
1110 ret = gen8_alloc_va_range(&ppgtt->base, start, size);
1112 free_pd(ppgtt->base.dev, ppgtt->base.scratch_pd);
1113 free_pt(ppgtt->base.dev, ppgtt->base.scratch_pt);
1117 ppgtt->base.allocate_va_range = NULL;
1118 ppgtt->base.clear_range(&ppgtt->base, 0, ppgtt->base.total, true);
1123 static void gen6_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
1125 struct i915_address_space *vm = &ppgtt->base;
1126 struct i915_page_table *unused;
1127 gen6_pte_t scratch_pte;
1129 uint32_t pte, pde, temp;
1130 uint32_t start = ppgtt->base.start, length = ppgtt->base.total;
1132 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1133 I915_CACHE_LLC, true, 0);
1135 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde) {
1137 gen6_pte_t *pt_vaddr;
1138 const dma_addr_t pt_addr = px_dma(ppgtt->pd.page_table[pde]);
1139 pd_entry = readl(ppgtt->pd_addr + pde);
1140 expected = (GEN6_PDE_ADDR_ENCODE(pt_addr) | GEN6_PDE_VALID);
1142 if (pd_entry != expected)
1143 seq_printf(m, "\tPDE #%d mismatch: Actual PDE: %x Expected PDE: %x\n",
1147 seq_printf(m, "\tPDE: %x\n", pd_entry);
1149 pt_vaddr = kmap_px(ppgtt->pd.page_table[pde]);
1151 for (pte = 0; pte < GEN6_PTES; pte+=4) {
1153 (pde * PAGE_SIZE * GEN6_PTES) +
1157 for (i = 0; i < 4; i++)
1158 if (pt_vaddr[pte + i] != scratch_pte)
1163 seq_printf(m, "\t\t0x%lx [%03d,%04d]: =", va, pde, pte);
1164 for (i = 0; i < 4; i++) {
1165 if (pt_vaddr[pte + i] != scratch_pte)
1166 seq_printf(m, " %08x", pt_vaddr[pte + i]);
1168 seq_puts(m, " SCRATCH ");
1172 kunmap_px(ppgtt, pt_vaddr);
1176 /* Write pde (index) from the page directory @pd to the page table @pt */
1177 static void gen6_write_pde(struct i915_page_directory *pd,
1178 const int pde, struct i915_page_table *pt)
1180 /* Caller needs to make sure the write completes if necessary */
1181 struct i915_hw_ppgtt *ppgtt =
1182 container_of(pd, struct i915_hw_ppgtt, pd);
1185 pd_entry = GEN6_PDE_ADDR_ENCODE(px_dma(pt));
1186 pd_entry |= GEN6_PDE_VALID;
1188 writel(pd_entry, ppgtt->pd_addr + pde);
1191 /* Write all the page tables found in the ppgtt structure to incrementing page
1193 static void gen6_write_page_range(struct drm_i915_private *dev_priv,
1194 struct i915_page_directory *pd,
1195 uint32_t start, uint32_t length)
1197 struct i915_page_table *pt;
1200 gen6_for_each_pde(pt, pd, start, length, temp, pde)
1201 gen6_write_pde(pd, pde, pt);
1203 /* Make sure write is complete before other code can use this page
1204 * table. Also require for WC mapped PTEs */
1205 readl(dev_priv->gtt.gsm);
1208 static uint32_t get_pd_offset(struct i915_hw_ppgtt *ppgtt)
1210 BUG_ON(ppgtt->pd.base.ggtt_offset & 0x3f);
1212 return (ppgtt->pd.base.ggtt_offset / 64) << 16;
1215 static int hsw_mm_switch(struct i915_hw_ppgtt *ppgtt,
1216 struct drm_i915_gem_request *req)
1218 struct intel_engine_cs *ring = req->ring;
1221 /* NB: TLBs must be flushed and invalidated before a switch */
1222 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1226 ret = intel_ring_begin(req, 6);
1230 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1231 intel_ring_emit(ring, RING_PP_DIR_DCLV(ring));
1232 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1233 intel_ring_emit(ring, RING_PP_DIR_BASE(ring));
1234 intel_ring_emit(ring, get_pd_offset(ppgtt));
1235 intel_ring_emit(ring, MI_NOOP);
1236 intel_ring_advance(ring);
1241 static int vgpu_mm_switch(struct i915_hw_ppgtt *ppgtt,
1242 struct drm_i915_gem_request *req)
1244 struct intel_engine_cs *ring = req->ring;
1245 struct drm_i915_private *dev_priv = to_i915(ppgtt->base.dev);
1247 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
1248 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
1252 static int gen7_mm_switch(struct i915_hw_ppgtt *ppgtt,
1253 struct drm_i915_gem_request *req)
1255 struct intel_engine_cs *ring = req->ring;
1258 /* NB: TLBs must be flushed and invalidated before a switch */
1259 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1263 ret = intel_ring_begin(req, 6);
1267 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1268 intel_ring_emit(ring, RING_PP_DIR_DCLV(ring));
1269 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1270 intel_ring_emit(ring, RING_PP_DIR_BASE(ring));
1271 intel_ring_emit(ring, get_pd_offset(ppgtt));
1272 intel_ring_emit(ring, MI_NOOP);
1273 intel_ring_advance(ring);
1275 /* XXX: RCS is the only one to auto invalidate the TLBs? */
1276 if (ring->id != RCS) {
1277 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1285 static int gen6_mm_switch(struct i915_hw_ppgtt *ppgtt,
1286 struct drm_i915_gem_request *req)
1288 struct intel_engine_cs *ring = req->ring;
1289 struct drm_device *dev = ppgtt->base.dev;
1290 struct drm_i915_private *dev_priv = dev->dev_private;
1293 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
1294 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
1296 POSTING_READ(RING_PP_DIR_DCLV(ring));
1301 static void gen8_ppgtt_enable(struct drm_device *dev)
1303 struct drm_i915_private *dev_priv = dev->dev_private;
1304 struct intel_engine_cs *ring;
1307 for_each_ring(ring, dev_priv, j) {
1308 I915_WRITE(RING_MODE_GEN7(ring),
1309 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1313 static void gen7_ppgtt_enable(struct drm_device *dev)
1315 struct drm_i915_private *dev_priv = dev->dev_private;
1316 struct intel_engine_cs *ring;
1317 uint32_t ecochk, ecobits;
1320 ecobits = I915_READ(GAC_ECO_BITS);
1321 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B);
1323 ecochk = I915_READ(GAM_ECOCHK);
1324 if (IS_HASWELL(dev)) {
1325 ecochk |= ECOCHK_PPGTT_WB_HSW;
1327 ecochk |= ECOCHK_PPGTT_LLC_IVB;
1328 ecochk &= ~ECOCHK_PPGTT_GFDT_IVB;
1330 I915_WRITE(GAM_ECOCHK, ecochk);
1332 for_each_ring(ring, dev_priv, i) {
1333 /* GFX_MODE is per-ring on gen7+ */
1334 I915_WRITE(RING_MODE_GEN7(ring),
1335 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1339 static void gen6_ppgtt_enable(struct drm_device *dev)
1341 struct drm_i915_private *dev_priv = dev->dev_private;
1342 uint32_t ecochk, gab_ctl, ecobits;
1344 ecobits = I915_READ(GAC_ECO_BITS);
1345 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_SNB_BIT |
1346 ECOBITS_PPGTT_CACHE64B);
1348 gab_ctl = I915_READ(GAB_CTL);
1349 I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT);
1351 ecochk = I915_READ(GAM_ECOCHK);
1352 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT | ECOCHK_PPGTT_CACHE64B);
1354 I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1357 /* PPGTT support for Sandybdrige/Gen6 and later */
1358 static void gen6_ppgtt_clear_range(struct i915_address_space *vm,
1363 struct i915_hw_ppgtt *ppgtt =
1364 container_of(vm, struct i915_hw_ppgtt, base);
1365 gen6_pte_t *pt_vaddr, scratch_pte;
1366 unsigned first_entry = start >> PAGE_SHIFT;
1367 unsigned num_entries = length >> PAGE_SHIFT;
1368 unsigned act_pt = first_entry / GEN6_PTES;
1369 unsigned first_pte = first_entry % GEN6_PTES;
1370 unsigned last_pte, i;
1372 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
1373 I915_CACHE_LLC, true, 0);
1375 while (num_entries) {
1376 last_pte = first_pte + num_entries;
1377 if (last_pte > GEN6_PTES)
1378 last_pte = GEN6_PTES;
1380 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1382 for (i = first_pte; i < last_pte; i++)
1383 pt_vaddr[i] = scratch_pte;
1385 kunmap_px(ppgtt, pt_vaddr);
1387 num_entries -= last_pte - first_pte;
1393 static void gen6_ppgtt_insert_entries(struct i915_address_space *vm,
1394 struct sg_table *pages,
1396 enum i915_cache_level cache_level, u32 flags)
1398 struct i915_hw_ppgtt *ppgtt =
1399 container_of(vm, struct i915_hw_ppgtt, base);
1400 gen6_pte_t *pt_vaddr;
1401 unsigned first_entry = start >> PAGE_SHIFT;
1402 unsigned act_pt = first_entry / GEN6_PTES;
1403 unsigned act_pte = first_entry % GEN6_PTES;
1404 struct sg_page_iter sg_iter;
1407 for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) {
1408 if (pt_vaddr == NULL)
1409 pt_vaddr = kmap_px(ppgtt->pd.page_table[act_pt]);
1412 vm->pte_encode(sg_page_iter_dma_address(&sg_iter),
1413 cache_level, true, flags);
1415 if (++act_pte == GEN6_PTES) {
1416 kunmap_px(ppgtt, pt_vaddr);
1423 kunmap_px(ppgtt, pt_vaddr);
1426 static int gen6_alloc_va_range(struct i915_address_space *vm,
1427 uint64_t start_in, uint64_t length_in)
1429 DECLARE_BITMAP(new_page_tables, I915_PDES);
1430 struct drm_device *dev = vm->dev;
1431 struct drm_i915_private *dev_priv = dev->dev_private;
1432 struct i915_hw_ppgtt *ppgtt =
1433 container_of(vm, struct i915_hw_ppgtt, base);
1434 struct i915_page_table *pt;
1435 uint32_t start, length, start_save, length_save;
1439 if (WARN_ON(start_in + length_in > ppgtt->base.total))
1442 start = start_save = start_in;
1443 length = length_save = length_in;
1445 bitmap_zero(new_page_tables, I915_PDES);
1447 /* The allocation is done in two stages so that we can bail out with
1448 * minimal amount of pain. The first stage finds new page tables that
1449 * need allocation. The second stage marks use ptes within the page
1452 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1453 if (pt != vm->scratch_pt) {
1454 // WARN_ON(bitmap_empty(pt->used_ptes, GEN6_PTES));
1458 /* We've already allocated a page table */
1459 WARN_ON(!bitmap_empty(pt->used_ptes, GEN6_PTES));
1467 gen6_initialize_pt(vm, pt);
1469 ppgtt->pd.page_table[pde] = pt;
1470 __set_bit(pde, new_page_tables);
1471 trace_i915_page_table_entry_alloc(vm, pde, start, GEN6_PDE_SHIFT);
1475 length = length_save;
1477 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1478 DECLARE_BITMAP(tmp_bitmap, GEN6_PTES);
1480 bitmap_zero(tmp_bitmap, GEN6_PTES);
1481 bitmap_set(tmp_bitmap, gen6_pte_index(start),
1482 gen6_pte_count(start, length));
1484 if (__test_and_clear_bit(pde, new_page_tables))
1485 gen6_write_pde(&ppgtt->pd, pde, pt);
1487 trace_i915_page_table_entry_map(vm, pde, pt,
1488 gen6_pte_index(start),
1489 gen6_pte_count(start, length),
1491 bitmap_or(pt->used_ptes, tmp_bitmap, pt->used_ptes,
1495 WARN_ON(!bitmap_empty(new_page_tables, I915_PDES));
1497 /* Make sure write is complete before other code can use this page
1498 * table. Also require for WC mapped PTEs */
1499 readl(dev_priv->gtt.gsm);
1501 mark_tlbs_dirty(ppgtt);
1505 for_each_set_bit(pde, new_page_tables, I915_PDES) {
1506 struct i915_page_table *pt = ppgtt->pd.page_table[pde];
1508 ppgtt->pd.page_table[pde] = vm->scratch_pt;
1509 free_pt(vm->dev, pt);
1512 mark_tlbs_dirty(ppgtt);
1516 static int gen6_init_scratch(struct i915_address_space *vm)
1518 struct drm_device *dev = vm->dev;
1520 vm->scratch_page = alloc_scratch_page(dev);
1521 if (IS_ERR(vm->scratch_page))
1522 return PTR_ERR(vm->scratch_page);
1524 vm->scratch_pt = alloc_pt(dev);
1525 if (IS_ERR(vm->scratch_pt)) {
1526 free_scratch_page(dev, vm->scratch_page);
1527 return PTR_ERR(vm->scratch_pt);
1530 gen6_initialize_pt(vm, vm->scratch_pt);
1535 static void gen6_free_scratch(struct i915_address_space *vm)
1537 struct drm_device *dev = vm->dev;
1539 free_pt(dev, vm->scratch_pt);
1540 free_scratch_page(dev, vm->scratch_page);
1543 static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
1545 struct i915_hw_ppgtt *ppgtt =
1546 container_of(vm, struct i915_hw_ppgtt, base);
1547 struct i915_page_table *pt;
1550 drm_mm_remove_node(&ppgtt->node);
1552 gen6_for_all_pdes(pt, ppgtt, pde) {
1553 if (pt != vm->scratch_pt)
1554 free_pt(ppgtt->base.dev, pt);
1557 gen6_free_scratch(vm);
1560 static int gen6_ppgtt_allocate_page_directories(struct i915_hw_ppgtt *ppgtt)
1562 struct i915_address_space *vm = &ppgtt->base;
1563 struct drm_device *dev = ppgtt->base.dev;
1564 struct drm_i915_private *dev_priv = dev->dev_private;
1565 bool retried = false;
1568 /* PPGTT PDEs reside in the GGTT and consists of 512 entries. The
1569 * allocator works in address space sizes, so it's multiplied by page
1570 * size. We allocate at the top of the GTT to avoid fragmentation.
1572 BUG_ON(!drm_mm_initialized(&dev_priv->gtt.base.mm));
1574 ret = gen6_init_scratch(vm);
1579 ret = drm_mm_insert_node_in_range_generic(&dev_priv->gtt.base.mm,
1580 &ppgtt->node, GEN6_PD_SIZE,
1582 0, dev_priv->gtt.base.total,
1584 if (ret == -ENOSPC && !retried) {
1585 ret = i915_gem_evict_something(dev, &dev_priv->gtt.base,
1586 GEN6_PD_SIZE, GEN6_PD_ALIGN,
1588 0, dev_priv->gtt.base.total,
1601 if (ppgtt->node.start < dev_priv->gtt.mappable_end)
1602 DRM_DEBUG("Forced to use aperture for PDEs\n");
1607 gen6_free_scratch(vm);
1611 static int gen6_ppgtt_alloc(struct i915_hw_ppgtt *ppgtt)
1613 return gen6_ppgtt_allocate_page_directories(ppgtt);
1616 static void gen6_scratch_va_range(struct i915_hw_ppgtt *ppgtt,
1617 uint64_t start, uint64_t length)
1619 struct i915_page_table *unused;
1622 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde)
1623 ppgtt->pd.page_table[pde] = ppgtt->base.scratch_pt;
1626 static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt, bool aliasing)
1628 struct drm_device *dev = ppgtt->base.dev;
1629 struct drm_i915_private *dev_priv = dev->dev_private;
1632 ppgtt->base.pte_encode = dev_priv->gtt.base.pte_encode;
1634 ppgtt->switch_mm = gen6_mm_switch;
1635 } else if (IS_HASWELL(dev)) {
1636 ppgtt->switch_mm = hsw_mm_switch;
1637 } else if (IS_GEN7(dev)) {
1638 ppgtt->switch_mm = gen7_mm_switch;
1642 if (intel_vgpu_active(dev))
1643 ppgtt->switch_mm = vgpu_mm_switch;
1645 ret = gen6_ppgtt_alloc(ppgtt);
1650 /* preallocate all pts */
1651 ret = alloc_pt_range(&ppgtt->pd, 0, I915_PDES,
1655 gen6_ppgtt_cleanup(&ppgtt->base);
1660 ppgtt->base.allocate_va_range = gen6_alloc_va_range;
1661 ppgtt->base.clear_range = gen6_ppgtt_clear_range;
1662 ppgtt->base.insert_entries = gen6_ppgtt_insert_entries;
1663 ppgtt->base.unbind_vma = ppgtt_unbind_vma;
1664 ppgtt->base.bind_vma = ppgtt_bind_vma;
1665 ppgtt->base.cleanup = gen6_ppgtt_cleanup;
1666 ppgtt->base.start = 0;
1667 ppgtt->base.total = I915_PDES * GEN6_PTES * PAGE_SIZE;
1668 ppgtt->debug_dump = gen6_dump_ppgtt;
1670 ppgtt->pd.base.ggtt_offset =
1671 ppgtt->node.start / PAGE_SIZE * sizeof(gen6_pte_t);
1673 ppgtt->pd_addr = (gen6_pte_t __iomem *)dev_priv->gtt.gsm +
1674 ppgtt->pd.base.ggtt_offset / sizeof(gen6_pte_t);
1677 ppgtt->base.clear_range(&ppgtt->base, 0, ppgtt->base.total, true);
1679 gen6_scratch_va_range(ppgtt, 0, ppgtt->base.total);
1681 gen6_write_page_range(dev_priv, &ppgtt->pd, 0, ppgtt->base.total);
1683 DRM_DEBUG_DRIVER("Allocated pde space (%ldM) at GTT entry: %lx\n",
1684 ppgtt->node.size >> 20,
1685 ppgtt->node.start / PAGE_SIZE);
1687 DRM_DEBUG("Adding PPGTT at offset %x\n",
1688 ppgtt->pd.base.ggtt_offset << 10);
1693 static int __hw_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt,
1696 ppgtt->base.dev = dev;
1698 if (INTEL_INFO(dev)->gen < 8)
1699 return gen6_ppgtt_init(ppgtt, aliasing);
1701 return gen8_aliasing_ppgtt_init(ppgtt);
1703 return gen8_ppgtt_init(ppgtt);
1706 int i915_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
1708 struct drm_i915_private *dev_priv = dev->dev_private;
1711 ret = __hw_ppgtt_init(dev, ppgtt, false);
1713 kref_init(&ppgtt->ref);
1714 drm_mm_init(&ppgtt->base.mm, ppgtt->base.start,
1716 i915_init_vm(dev_priv, &ppgtt->base);
1722 int i915_ppgtt_init_hw(struct drm_device *dev)
1724 /* In the case of execlists, PPGTT is enabled by the context descriptor
1725 * and the PDPs are contained within the context itself. We don't
1726 * need to do anything here. */
1727 if (i915.enable_execlists)
1730 if (!USES_PPGTT(dev))
1734 gen6_ppgtt_enable(dev);
1735 else if (IS_GEN7(dev))
1736 gen7_ppgtt_enable(dev);
1737 else if (INTEL_INFO(dev)->gen >= 8)
1738 gen8_ppgtt_enable(dev);
1740 MISSING_CASE(INTEL_INFO(dev)->gen);
1745 int i915_ppgtt_init_ring(struct drm_i915_gem_request *req)
1747 struct drm_i915_private *dev_priv = req->ring->dev->dev_private;
1748 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
1750 if (i915.enable_execlists)
1756 return ppgtt->switch_mm(ppgtt, req);
1759 struct i915_hw_ppgtt *
1760 i915_ppgtt_create(struct drm_device *dev, struct drm_i915_file_private *fpriv)
1762 struct i915_hw_ppgtt *ppgtt;
1765 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
1767 return ERR_PTR(-ENOMEM);
1769 ret = i915_ppgtt_init(dev, ppgtt);
1772 return ERR_PTR(ret);
1775 ppgtt->file_priv = fpriv;
1777 trace_i915_ppgtt_create(&ppgtt->base);
1782 void i915_ppgtt_release(struct kref *kref)
1784 struct i915_hw_ppgtt *ppgtt =
1785 container_of(kref, struct i915_hw_ppgtt, ref);
1787 trace_i915_ppgtt_release(&ppgtt->base);
1789 /* vmas should already be unbound */
1790 WARN_ON(!list_empty(&ppgtt->base.active_list));
1791 WARN_ON(!list_empty(&ppgtt->base.inactive_list));
1793 list_del(&ppgtt->base.global_link);
1794 drm_mm_takedown(&ppgtt->base.mm);
1796 ppgtt->base.cleanup(&ppgtt->base);
1800 extern int intel_iommu_gfx_mapped;
1801 /* Certain Gen5 chipsets require require idling the GPU before
1802 * unmapping anything from the GTT when VT-d is enabled.
1804 static bool needs_idle_maps(struct drm_device *dev)
1806 #ifdef CONFIG_INTEL_IOMMU
1807 /* Query intel_iommu to see if we need the workaround. Presumably that
1810 if (IS_GEN5(dev) && IS_MOBILE(dev) && intel_iommu_gfx_mapped)
1816 static bool do_idling(struct drm_i915_private *dev_priv)
1818 bool ret = dev_priv->mm.interruptible;
1820 if (unlikely(dev_priv->gtt.do_idle_maps)) {
1821 dev_priv->mm.interruptible = false;
1822 if (i915_gpu_idle(dev_priv->dev)) {
1823 DRM_ERROR("Couldn't idle GPU\n");
1824 /* Wait a bit, in hopes it avoids the hang */
1832 static void undo_idling(struct drm_i915_private *dev_priv, bool interruptible)
1834 if (unlikely(dev_priv->gtt.do_idle_maps))
1835 dev_priv->mm.interruptible = interruptible;
1838 void i915_check_and_clear_faults(struct drm_device *dev)
1840 struct drm_i915_private *dev_priv = dev->dev_private;
1841 struct intel_engine_cs *ring;
1844 if (INTEL_INFO(dev)->gen < 6)
1847 for_each_ring(ring, dev_priv, i) {
1849 fault_reg = I915_READ(RING_FAULT_REG(ring));
1850 if (fault_reg & RING_FAULT_VALID) {
1852 DRM_DEBUG_DRIVER("Unexpected fault\n"
1854 "\tAddress space: %s\n"
1857 fault_reg & PAGE_MASK,
1858 fault_reg & RING_FAULT_GTTSEL_MASK ? "GGTT" : "PPGTT",
1859 RING_FAULT_SRCID(fault_reg),
1860 RING_FAULT_FAULT_TYPE(fault_reg));
1862 I915_WRITE(RING_FAULT_REG(ring),
1863 fault_reg & ~RING_FAULT_VALID);
1866 POSTING_READ(RING_FAULT_REG(&dev_priv->ring[RCS]));
1869 static void i915_ggtt_flush(struct drm_i915_private *dev_priv)
1871 if (INTEL_INFO(dev_priv->dev)->gen < 6) {
1872 intel_gtt_chipset_flush();
1874 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
1875 POSTING_READ(GFX_FLSH_CNTL_GEN6);
1879 void i915_gem_suspend_gtt_mappings(struct drm_device *dev)
1881 struct drm_i915_private *dev_priv = dev->dev_private;
1883 /* Don't bother messing with faults pre GEN6 as we have little
1884 * documentation supporting that it's a good idea.
1886 if (INTEL_INFO(dev)->gen < 6)
1889 i915_check_and_clear_faults(dev);
1891 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
1892 dev_priv->gtt.base.start,
1893 dev_priv->gtt.base.total,
1896 i915_ggtt_flush(dev_priv);
1899 int i915_gem_gtt_prepare_object(struct drm_i915_gem_object *obj)
1901 if (!dma_map_sg(obj->base.dev->pdev->dev,
1902 obj->pages->sgl, obj->pages->nents,
1903 PCI_DMA_BIDIRECTIONAL))
1909 static void gen8_set_pte(void __iomem *addr, gen8_pte_t pte)
1914 iowrite32((u32)pte, addr);
1915 iowrite32(pte >> 32, addr + 4);
1919 static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
1920 struct sg_table *st,
1922 enum i915_cache_level level, u32 unused)
1924 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1925 unsigned first_entry = start >> PAGE_SHIFT;
1926 gen8_pte_t __iomem *gtt_entries =
1927 (gen8_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
1929 struct sg_page_iter sg_iter;
1930 dma_addr_t addr = 0; /* shut up gcc */
1932 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
1933 addr = sg_dma_address(sg_iter.sg) +
1934 (sg_iter.sg_pgoffset << PAGE_SHIFT);
1935 gen8_set_pte(>t_entries[i],
1936 gen8_pte_encode(addr, level, true));
1941 * XXX: This serves as a posting read to make sure that the PTE has
1942 * actually been updated. There is some concern that even though
1943 * registers and PTEs are within the same BAR that they are potentially
1944 * of NUMA access patterns. Therefore, even with the way we assume
1945 * hardware should work, we must keep this posting read for paranoia.
1948 WARN_ON(readq(>t_entries[i-1])
1949 != gen8_pte_encode(addr, level, true));
1951 /* This next bit makes the above posting read even more important. We
1952 * want to flush the TLBs only after we're certain all the PTE updates
1955 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
1956 POSTING_READ(GFX_FLSH_CNTL_GEN6);
1960 * Binds an object into the global gtt with the specified cache level. The object
1961 * will be accessible to the GPU via commands whose operands reference offsets
1962 * within the global GTT as well as accessible by the GPU through the GMADR
1963 * mapped BAR (dev_priv->mm.gtt->gtt).
1965 static void gen6_ggtt_insert_entries(struct i915_address_space *vm,
1966 struct sg_table *st,
1968 enum i915_cache_level level, u32 flags)
1970 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1971 unsigned first_entry = start >> PAGE_SHIFT;
1972 gen6_pte_t __iomem *gtt_entries =
1973 (gen6_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
1975 struct sg_page_iter sg_iter;
1976 dma_addr_t addr = 0;
1978 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
1979 addr = sg_page_iter_dma_address(&sg_iter);
1980 iowrite32(vm->pte_encode(addr, level, true, flags), >t_entries[i]);
1984 /* XXX: This serves as a posting read to make sure that the PTE has
1985 * actually been updated. There is some concern that even though
1986 * registers and PTEs are within the same BAR that they are potentially
1987 * of NUMA access patterns. Therefore, even with the way we assume
1988 * hardware should work, we must keep this posting read for paranoia.
1991 unsigned long gtt = readl(>t_entries[i-1]);
1992 WARN_ON(gtt != vm->pte_encode(addr, level, true, flags));
1995 /* This next bit makes the above posting read even more important. We
1996 * want to flush the TLBs only after we're certain all the PTE updates
1999 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
2000 POSTING_READ(GFX_FLSH_CNTL_GEN6);
2003 static void gen8_ggtt_clear_range(struct i915_address_space *vm,
2008 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2009 unsigned first_entry = start >> PAGE_SHIFT;
2010 unsigned num_entries = length >> PAGE_SHIFT;
2011 gen8_pte_t scratch_pte, __iomem *gtt_base =
2012 (gen8_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
2013 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
2016 if (WARN(num_entries > max_entries,
2017 "First entry = %d; Num entries = %d (max=%d)\n",
2018 first_entry, num_entries, max_entries))
2019 num_entries = max_entries;
2021 scratch_pte = gen8_pte_encode(px_dma(vm->scratch_page),
2024 for (i = 0; i < num_entries; i++)
2025 gen8_set_pte(>t_base[i], scratch_pte);
2029 static void gen6_ggtt_clear_range(struct i915_address_space *vm,
2034 struct drm_i915_private *dev_priv = vm->dev->dev_private;
2035 unsigned first_entry = start >> PAGE_SHIFT;
2036 unsigned num_entries = length >> PAGE_SHIFT;
2037 gen6_pte_t scratch_pte, __iomem *gtt_base =
2038 (gen6_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
2039 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
2042 if (WARN(num_entries > max_entries,
2043 "First entry = %d; Num entries = %d (max=%d)\n",
2044 first_entry, num_entries, max_entries))
2045 num_entries = max_entries;
2047 scratch_pte = vm->pte_encode(px_dma(vm->scratch_page),
2048 I915_CACHE_LLC, use_scratch, 0);
2050 for (i = 0; i < num_entries; i++)
2051 iowrite32(scratch_pte, >t_base[i]);
2055 static void i915_ggtt_insert_entries(struct i915_address_space *vm,
2056 struct sg_table *pages,
2058 enum i915_cache_level cache_level, u32 unused)
2060 unsigned int flags = (cache_level == I915_CACHE_NONE) ?
2061 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
2063 intel_gtt_insert_sg_entries(pages, start >> PAGE_SHIFT, flags);
2066 static void i915_ggtt_clear_range(struct i915_address_space *vm,
2071 unsigned first_entry = start >> PAGE_SHIFT;
2072 unsigned num_entries = length >> PAGE_SHIFT;
2073 intel_gtt_clear_range(first_entry, num_entries);
2076 static int ggtt_bind_vma(struct i915_vma *vma,
2077 enum i915_cache_level cache_level,
2080 struct drm_device *dev = vma->vm->dev;
2081 struct drm_i915_private *dev_priv = dev->dev_private;
2082 struct drm_i915_gem_object *obj = vma->obj;
2083 struct sg_table *pages = obj->pages;
2087 ret = i915_get_ggtt_vma_pages(vma);
2090 pages = vma->ggtt_view.pages;
2092 /* Currently applicable only to VLV */
2094 pte_flags |= PTE_READ_ONLY;
2097 if (!dev_priv->mm.aliasing_ppgtt || flags & GLOBAL_BIND) {
2098 vma->vm->insert_entries(vma->vm, pages,
2100 cache_level, pte_flags);
2103 if (dev_priv->mm.aliasing_ppgtt && flags & LOCAL_BIND) {
2104 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
2105 appgtt->base.insert_entries(&appgtt->base, pages,
2107 cache_level, pte_flags);
2113 static void ggtt_unbind_vma(struct i915_vma *vma)
2115 struct drm_device *dev = vma->vm->dev;
2116 struct drm_i915_private *dev_priv = dev->dev_private;
2117 struct drm_i915_gem_object *obj = vma->obj;
2118 const uint64_t size = min_t(uint64_t,
2122 if (vma->bound & GLOBAL_BIND) {
2123 vma->vm->clear_range(vma->vm,
2129 if (dev_priv->mm.aliasing_ppgtt && vma->bound & LOCAL_BIND) {
2130 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
2132 appgtt->base.clear_range(&appgtt->base,
2139 void i915_gem_gtt_finish_object(struct drm_i915_gem_object *obj)
2141 struct drm_device *dev = obj->base.dev;
2142 struct drm_i915_private *dev_priv = dev->dev_private;
2145 interruptible = do_idling(dev_priv);
2147 dma_unmap_sg(dev->pdev->dev, obj->pages->sgl, obj->pages->nents,
2148 PCI_DMA_BIDIRECTIONAL);
2150 undo_idling(dev_priv, interruptible);
2153 static void i915_gtt_color_adjust(struct drm_mm_node *node,
2154 unsigned long color,
2158 if (node->color != color)
2161 if (!list_empty(&node->node_list)) {
2162 node = list_entry(node->node_list.next,
2165 if (node->allocated && node->color != color)
2170 static int i915_gem_setup_global_gtt(struct drm_device *dev,
2171 unsigned long start,
2172 unsigned long mappable_end,
2175 /* Let GEM Manage all of the aperture.
2177 * However, leave one page at the end still bound to the scratch page.
2178 * There are a number of places where the hardware apparently prefetches
2179 * past the end of the object, and we've seen multiple hangs with the
2180 * GPU head pointer stuck in a batchbuffer bound at the last page of the
2181 * aperture. One page should be enough to keep any prefetching inside
2184 struct drm_i915_private *dev_priv = dev->dev_private;
2185 struct i915_address_space *ggtt_vm = &dev_priv->gtt.base;
2186 unsigned long mappable;
2188 struct drm_mm_node *entry;
2189 struct drm_i915_gem_object *obj;
2190 unsigned long hole_start, hole_end;
2193 mappable = min(end, mappable_end) - start;
2194 BUG_ON(mappable_end > end);
2196 /* Subtract the guard page ... */
2197 drm_mm_init(&ggtt_vm->mm, start, end - start - PAGE_SIZE);
2199 dev_priv->gtt.base.start = start;
2200 dev_priv->gtt.base.total = end - start;
2202 if (intel_vgpu_active(dev)) {
2203 ret = intel_vgt_balloon(dev);
2209 dev_priv->gtt.base.mm.color_adjust = i915_gtt_color_adjust;
2211 /* Mark any preallocated objects as occupied */
2212 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
2213 struct i915_vma *vma = i915_gem_obj_to_vma(obj, ggtt_vm);
2215 DRM_DEBUG_KMS("reserving preallocated space: %lx + %zx\n",
2216 i915_gem_obj_ggtt_offset(obj), obj->base.size);
2218 WARN_ON(i915_gem_obj_ggtt_bound(obj));
2219 ret = drm_mm_reserve_node(&ggtt_vm->mm, &vma->node);
2221 DRM_DEBUG_KMS("Reservation failed: %i\n", ret);
2224 vma->bound |= GLOBAL_BIND;
2227 /* Clear any non-preallocated blocks */
2228 drm_mm_for_each_hole(entry, &ggtt_vm->mm, hole_start, hole_end) {
2229 DRM_DEBUG_KMS("clearing unused GTT space: [%lx, %lx]\n",
2230 hole_start, hole_end);
2231 ggtt_vm->clear_range(ggtt_vm, hole_start,
2232 hole_end - hole_start, true);
2235 #ifdef __DragonFly__
2236 device_printf(dev->dev,
2237 "taking over the fictitious range 0x%lx-0x%lx\n",
2238 dev_priv->gtt.mappable_base + start, dev_priv->gtt.mappable_base + start + mappable);
2239 error = -vm_phys_fictitious_reg_range(dev_priv->gtt.mappable_base + start,
2240 dev_priv->gtt.mappable_base + start + mappable, VM_MEMATTR_WRITE_COMBINING);
2243 /* And finally clear the reserved guard page */
2244 ggtt_vm->clear_range(ggtt_vm, end - PAGE_SIZE, PAGE_SIZE, true);
2246 if (USES_PPGTT(dev) && !USES_FULL_PPGTT(dev)) {
2247 struct i915_hw_ppgtt *ppgtt;
2249 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2253 ret = __hw_ppgtt_init(dev, ppgtt, true);
2255 ppgtt->base.cleanup(&ppgtt->base);
2260 if (ppgtt->base.allocate_va_range)
2261 ret = ppgtt->base.allocate_va_range(&ppgtt->base, 0,
2264 ppgtt->base.cleanup(&ppgtt->base);
2269 ppgtt->base.clear_range(&ppgtt->base,
2274 dev_priv->mm.aliasing_ppgtt = ppgtt;
2280 void i915_gem_init_global_gtt(struct drm_device *dev)
2282 struct drm_i915_private *dev_priv = dev->dev_private;
2283 u64 gtt_size, mappable_size;
2285 gtt_size = dev_priv->gtt.base.total;
2286 mappable_size = dev_priv->gtt.mappable_end;
2288 i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
2291 void i915_global_gtt_cleanup(struct drm_device *dev)
2293 struct drm_i915_private *dev_priv = dev->dev_private;
2294 struct i915_address_space *vm = &dev_priv->gtt.base;
2296 if (dev_priv->mm.aliasing_ppgtt) {
2297 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
2299 ppgtt->base.cleanup(&ppgtt->base);
2302 if (drm_mm_initialized(&vm->mm)) {
2303 if (intel_vgpu_active(dev))
2304 intel_vgt_deballoon();
2306 drm_mm_takedown(&vm->mm);
2307 list_del(&vm->global_link);
2313 static unsigned int gen6_get_total_gtt_size(u16 snb_gmch_ctl)
2315 snb_gmch_ctl >>= SNB_GMCH_GGMS_SHIFT;
2316 snb_gmch_ctl &= SNB_GMCH_GGMS_MASK;
2317 return snb_gmch_ctl << 20;
2320 static unsigned int gen8_get_total_gtt_size(u16 bdw_gmch_ctl)
2322 bdw_gmch_ctl >>= BDW_GMCH_GGMS_SHIFT;
2323 bdw_gmch_ctl &= BDW_GMCH_GGMS_MASK;
2325 bdw_gmch_ctl = 1 << bdw_gmch_ctl;
2327 #ifdef CONFIG_X86_32
2328 /* Limit 32b platforms to a 2GB GGTT: 4 << 20 / pte size * PAGE_SIZE */
2329 if (bdw_gmch_ctl > 4)
2333 return bdw_gmch_ctl << 20;
2336 static unsigned int chv_get_total_gtt_size(u16 gmch_ctrl)
2338 gmch_ctrl >>= SNB_GMCH_GGMS_SHIFT;
2339 gmch_ctrl &= SNB_GMCH_GGMS_MASK;
2342 return 1 << (20 + gmch_ctrl);
2347 static size_t gen6_get_stolen_size(u16 snb_gmch_ctl)
2349 snb_gmch_ctl >>= SNB_GMCH_GMS_SHIFT;
2350 snb_gmch_ctl &= SNB_GMCH_GMS_MASK;
2351 return snb_gmch_ctl << 25; /* 32 MB units */
2354 static size_t gen8_get_stolen_size(u16 bdw_gmch_ctl)
2356 bdw_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2357 bdw_gmch_ctl &= BDW_GMCH_GMS_MASK;
2358 return bdw_gmch_ctl << 25; /* 32 MB units */
2361 static size_t chv_get_stolen_size(u16 gmch_ctrl)
2363 gmch_ctrl >>= SNB_GMCH_GMS_SHIFT;
2364 gmch_ctrl &= SNB_GMCH_GMS_MASK;
2367 * 0x0 to 0x10: 32MB increments starting at 0MB
2368 * 0x11 to 0x16: 4MB increments starting at 8MB
2369 * 0x17 to 0x1d: 4MB increments start at 36MB
2371 if (gmch_ctrl < 0x11)
2372 return gmch_ctrl << 25;
2373 else if (gmch_ctrl < 0x17)
2374 return (gmch_ctrl - 0x11 + 2) << 22;
2376 return (gmch_ctrl - 0x17 + 9) << 22;
2379 static size_t gen9_get_stolen_size(u16 gen9_gmch_ctl)
2381 gen9_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2382 gen9_gmch_ctl &= BDW_GMCH_GMS_MASK;
2384 if (gen9_gmch_ctl < 0xf0)
2385 return gen9_gmch_ctl << 25; /* 32 MB units */
2387 /* 4MB increments starting at 0xf0 for 4MB */
2388 return (gen9_gmch_ctl - 0xf0 + 1) << 22;
2391 static int ggtt_probe_common(struct drm_device *dev,
2394 struct drm_i915_private *dev_priv = dev->dev_private;
2395 struct i915_page_scratch *scratch_page;
2396 phys_addr_t gtt_phys_addr;
2398 /* For Modern GENs the PTEs and register space are split in the BAR */
2399 gtt_phys_addr = pci_resource_start(dev->pdev, 0) +
2400 (pci_resource_len(dev->pdev, 0) / 2);
2403 * On BXT writes larger than 64 bit to the GTT pagetable range will be
2404 * dropped. For WC mappings in general we have 64 byte burst writes
2405 * when the WC buffer is flushed, so we can't use it, but have to
2406 * resort to an uncached mapping. The WC issue is easily caught by the
2407 * readback check when writing GTT PTE entries.
2409 if (IS_BROXTON(dev))
2410 dev_priv->gtt.gsm = ioremap_nocache(gtt_phys_addr, gtt_size);
2412 dev_priv->gtt.gsm = ioremap_wc(gtt_phys_addr, gtt_size);
2413 if (!dev_priv->gtt.gsm) {
2414 DRM_ERROR("Failed to map the gtt page table\n");
2418 scratch_page = alloc_scratch_page(dev);
2419 if (IS_ERR(scratch_page)) {
2420 DRM_ERROR("Scratch setup failed\n");
2421 /* iounmap will also get called at remove, but meh */
2422 iounmap(dev_priv->gtt.gsm);
2423 return PTR_ERR(scratch_page);
2426 dev_priv->gtt.base.scratch_page = scratch_page;
2431 /* The GGTT and PPGTT need a private PPAT setup in order to handle cacheability
2432 * bits. When using advanced contexts each context stores its own PAT, but
2433 * writing this data shouldn't be harmful even in those cases. */
2434 static void bdw_setup_private_ppat(struct drm_i915_private *dev_priv)
2438 pat = GEN8_PPAT(0, GEN8_PPAT_WB | GEN8_PPAT_LLC) | /* for normal objects, no eLLC */
2439 GEN8_PPAT(1, GEN8_PPAT_WC | GEN8_PPAT_LLCELLC) | /* for something pointing to ptes? */
2440 GEN8_PPAT(2, GEN8_PPAT_WT | GEN8_PPAT_LLCELLC) | /* for scanout with eLLC */
2441 GEN8_PPAT(3, GEN8_PPAT_UC) | /* Uncached objects, mostly for scanout */
2442 GEN8_PPAT(4, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(0)) |
2443 GEN8_PPAT(5, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(1)) |
2444 GEN8_PPAT(6, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(2)) |
2445 GEN8_PPAT(7, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3));
2447 if (!USES_PPGTT(dev_priv->dev))
2448 /* Spec: "For GGTT, there is NO pat_sel[2:0] from the entry,
2449 * so RTL will always use the value corresponding to
2451 * So let's disable cache for GGTT to avoid screen corruptions.
2452 * MOCS still can be used though.
2453 * - System agent ggtt writes (i.e. cpu gtt mmaps) already work
2454 * before this patch, i.e. the same uncached + snooping access
2455 * like on gen6/7 seems to be in effect.
2456 * - So this just fixes blitter/render access. Again it looks
2457 * like it's not just uncached access, but uncached + snooping.
2458 * So we can still hold onto all our assumptions wrt cpu
2459 * clflushing on LLC machines.
2461 pat = GEN8_PPAT(0, GEN8_PPAT_UC);
2463 /* XXX: spec defines this as 2 distinct registers. It's unclear if a 64b
2464 * write would work. */
2465 I915_WRITE(GEN8_PRIVATE_PAT, pat);
2466 I915_WRITE(GEN8_PRIVATE_PAT + 4, pat >> 32);
2469 static void chv_setup_private_ppat(struct drm_i915_private *dev_priv)
2474 * Map WB on BDW to snooped on CHV.
2476 * Only the snoop bit has meaning for CHV, the rest is
2479 * The hardware will never snoop for certain types of accesses:
2480 * - CPU GTT (GMADR->GGTT->no snoop->memory)
2481 * - PPGTT page tables
2482 * - some other special cycles
2484 * As with BDW, we also need to consider the following for GT accesses:
2485 * "For GGTT, there is NO pat_sel[2:0] from the entry,
2486 * so RTL will always use the value corresponding to
2488 * Which means we must set the snoop bit in PAT entry 0
2489 * in order to keep the global status page working.
2491 pat = GEN8_PPAT(0, CHV_PPAT_SNOOP) |
2495 GEN8_PPAT(4, CHV_PPAT_SNOOP) |
2496 GEN8_PPAT(5, CHV_PPAT_SNOOP) |
2497 GEN8_PPAT(6, CHV_PPAT_SNOOP) |
2498 GEN8_PPAT(7, CHV_PPAT_SNOOP);
2500 I915_WRITE(GEN8_PRIVATE_PAT, pat);
2501 I915_WRITE(GEN8_PRIVATE_PAT + 4, pat >> 32);
2504 static int gen8_gmch_probe(struct drm_device *dev,
2507 phys_addr_t *mappable_base,
2510 struct drm_i915_private *dev_priv = dev->dev_private;
2515 /* TODO: We're not aware of mappable constraints on gen8 yet */
2516 *mappable_base = pci_resource_start(dev->pdev, 2);
2517 *mappable_end = pci_resource_len(dev->pdev, 2);
2520 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(39)))
2521 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(39));
2524 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
2526 if (INTEL_INFO(dev)->gen >= 9) {
2527 *stolen = gen9_get_stolen_size(snb_gmch_ctl);
2528 gtt_size = gen8_get_total_gtt_size(snb_gmch_ctl);
2529 } else if (IS_CHERRYVIEW(dev)) {
2530 *stolen = chv_get_stolen_size(snb_gmch_ctl);
2531 gtt_size = chv_get_total_gtt_size(snb_gmch_ctl);
2533 *stolen = gen8_get_stolen_size(snb_gmch_ctl);
2534 gtt_size = gen8_get_total_gtt_size(snb_gmch_ctl);
2537 *gtt_total = (gtt_size / sizeof(gen8_pte_t)) << PAGE_SHIFT;
2539 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
2540 chv_setup_private_ppat(dev_priv);
2542 bdw_setup_private_ppat(dev_priv);
2544 ret = ggtt_probe_common(dev, gtt_size);
2546 dev_priv->gtt.base.clear_range = gen8_ggtt_clear_range;
2547 dev_priv->gtt.base.insert_entries = gen8_ggtt_insert_entries;
2548 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
2549 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
2554 static int gen6_gmch_probe(struct drm_device *dev,
2557 phys_addr_t *mappable_base,
2560 struct drm_i915_private *dev_priv = dev->dev_private;
2561 unsigned int gtt_size;
2565 *mappable_base = pci_resource_start(dev->pdev, 2);
2566 *mappable_end = pci_resource_len(dev->pdev, 2);
2568 /* 64/512MB is the current min/max we actually know of, but this is just
2569 * a coarse sanity check.
2571 if ((*mappable_end < (64<<20) || (*mappable_end > (512<<20)))) {
2572 DRM_ERROR("Unknown GMADR size (%lx)\n",
2573 dev_priv->gtt.mappable_end);
2578 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(40)))
2579 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(40));
2581 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
2583 *stolen = gen6_get_stolen_size(snb_gmch_ctl);
2585 gtt_size = gen6_get_total_gtt_size(snb_gmch_ctl);
2586 *gtt_total = (gtt_size / sizeof(gen6_pte_t)) << PAGE_SHIFT;
2588 ret = ggtt_probe_common(dev, gtt_size);
2590 dev_priv->gtt.base.clear_range = gen6_ggtt_clear_range;
2591 dev_priv->gtt.base.insert_entries = gen6_ggtt_insert_entries;
2592 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
2593 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
2598 static void gen6_gmch_remove(struct i915_address_space *vm)
2600 struct i915_gtt *gtt = container_of(vm, struct i915_gtt, base);
2603 free_scratch_page(vm->dev, vm->scratch_page);
2606 static int i915_gmch_probe(struct drm_device *dev,
2609 phys_addr_t *mappable_base,
2612 struct drm_i915_private *dev_priv = dev->dev_private;
2616 ret = intel_gmch_probe(dev_priv->bridge_dev, dev_priv->dev->pdev, NULL);
2618 DRM_ERROR("failed to set up gmch\n");
2623 intel_gtt_get(gtt_total, stolen, mappable_base, mappable_end);
2625 dev_priv->gtt.do_idle_maps = needs_idle_maps(dev_priv->dev);
2626 dev_priv->gtt.base.insert_entries = i915_ggtt_insert_entries;
2627 dev_priv->gtt.base.clear_range = i915_ggtt_clear_range;
2628 dev_priv->gtt.base.bind_vma = ggtt_bind_vma;
2629 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma;
2631 if (unlikely(dev_priv->gtt.do_idle_maps))
2632 DRM_INFO("applying Ironlake quirks for intel_iommu\n");
2637 static void i915_gmch_remove(struct i915_address_space *vm)
2639 intel_gmch_remove();
2642 int i915_gem_gtt_init(struct drm_device *dev)
2644 struct drm_i915_private *dev_priv = dev->dev_private;
2645 struct i915_gtt *gtt = &dev_priv->gtt;
2648 if (INTEL_INFO(dev)->gen <= 5) {
2649 gtt->gtt_probe = i915_gmch_probe;
2650 gtt->base.cleanup = i915_gmch_remove;
2651 } else if (INTEL_INFO(dev)->gen < 8) {
2652 gtt->gtt_probe = gen6_gmch_probe;
2653 gtt->base.cleanup = gen6_gmch_remove;
2654 if (IS_HASWELL(dev) && dev_priv->ellc_size)
2655 gtt->base.pte_encode = iris_pte_encode;
2656 else if (IS_HASWELL(dev))
2657 gtt->base.pte_encode = hsw_pte_encode;
2658 else if (IS_VALLEYVIEW(dev))
2659 gtt->base.pte_encode = byt_pte_encode;
2660 else if (INTEL_INFO(dev)->gen >= 7)
2661 gtt->base.pte_encode = ivb_pte_encode;
2663 gtt->base.pte_encode = snb_pte_encode;
2665 dev_priv->gtt.gtt_probe = gen8_gmch_probe;
2666 dev_priv->gtt.base.cleanup = gen6_gmch_remove;
2669 gtt->base.dev = dev;
2671 ret = gtt->gtt_probe(dev, >t->base.total, >t->stolen_size,
2672 >t->mappable_base, >t->mappable_end);
2676 /* GMADR is the PCI mmio aperture into the global GTT. */
2677 DRM_INFO("Memory usable by graphics device = %luM\n",
2678 gtt->base.total >> 20);
2679 DRM_DEBUG_DRIVER("GMADR size = %ldM\n", gtt->mappable_end >> 20);
2680 DRM_DEBUG_DRIVER("GTT stolen size = %zdM\n", gtt->stolen_size >> 20);
2681 #ifdef CONFIG_INTEL_IOMMU
2682 if (intel_iommu_gfx_mapped)
2683 DRM_INFO("VT-d active for gfx access\n");
2686 * i915.enable_ppgtt is read-only, so do an early pass to validate the
2687 * user's requested state against the hardware/driver capabilities. We
2688 * do this now so that we can print out any log messages once rather
2689 * than every time we check intel_enable_ppgtt().
2691 i915.enable_ppgtt = sanitize_enable_ppgtt(dev, i915.enable_ppgtt);
2692 DRM_DEBUG_DRIVER("ppgtt mode: %i\n", i915.enable_ppgtt);
2697 void i915_gem_restore_gtt_mappings(struct drm_device *dev)
2699 struct drm_i915_private *dev_priv = dev->dev_private;
2700 struct drm_i915_gem_object *obj;
2701 struct i915_address_space *vm;
2703 i915_check_and_clear_faults(dev);
2705 /* First fill our portion of the GTT with scratch pages */
2706 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
2707 dev_priv->gtt.base.start,
2708 dev_priv->gtt.base.total,
2711 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
2712 struct i915_vma *vma = i915_gem_obj_to_vma(obj,
2713 &dev_priv->gtt.base);
2717 i915_gem_clflush_object(obj, obj->pin_display);
2718 WARN_ON(i915_vma_bind(vma, obj->cache_level, PIN_UPDATE));
2722 if (INTEL_INFO(dev)->gen >= 8) {
2723 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
2724 chv_setup_private_ppat(dev_priv);
2726 bdw_setup_private_ppat(dev_priv);
2731 if (USES_PPGTT(dev)) {
2732 list_for_each_entry(vm, &dev_priv->vm_list, global_link) {
2733 /* TODO: Perhaps it shouldn't be gen6 specific */
2735 struct i915_hw_ppgtt *ppgtt =
2736 container_of(vm, struct i915_hw_ppgtt,
2739 if (i915_is_ggtt(vm))
2740 ppgtt = dev_priv->mm.aliasing_ppgtt;
2742 gen6_write_page_range(dev_priv, &ppgtt->pd,
2743 0, ppgtt->base.total);
2747 i915_ggtt_flush(dev_priv);
2750 static struct i915_vma *
2751 __i915_gem_vma_create(struct drm_i915_gem_object *obj,
2752 struct i915_address_space *vm,
2753 const struct i915_ggtt_view *ggtt_view)
2755 struct i915_vma *vma;
2757 if (WARN_ON(i915_is_ggtt(vm) != !!ggtt_view))
2758 return ERR_PTR(-EINVAL);
2760 vma = kzalloc(sizeof(*vma), GFP_KERNEL);
2762 return ERR_PTR(-ENOMEM);
2764 INIT_LIST_HEAD(&vma->vma_link);
2765 INIT_LIST_HEAD(&vma->mm_list);
2766 INIT_LIST_HEAD(&vma->exec_list);
2770 if (i915_is_ggtt(vm))
2771 vma->ggtt_view = *ggtt_view;
2773 list_add_tail(&vma->vma_link, &obj->vma_list);
2774 if (!i915_is_ggtt(vm))
2775 i915_ppgtt_get(i915_vm_to_ppgtt(vm));
2781 i915_gem_obj_lookup_or_create_vma(struct drm_i915_gem_object *obj,
2782 struct i915_address_space *vm)
2784 struct i915_vma *vma;
2786 vma = i915_gem_obj_to_vma(obj, vm);
2788 vma = __i915_gem_vma_create(obj, vm,
2789 i915_is_ggtt(vm) ? &i915_ggtt_view_normal : NULL);
2795 i915_gem_obj_lookup_or_create_ggtt_vma(struct drm_i915_gem_object *obj,
2796 const struct i915_ggtt_view *view)
2798 struct i915_address_space *ggtt = i915_obj_to_ggtt(obj);
2799 struct i915_vma *vma;
2802 return ERR_PTR(-EINVAL);
2804 vma = i915_gem_obj_to_ggtt_view(obj, view);
2810 vma = __i915_gem_vma_create(obj, ggtt, view);
2817 rotate_pages(dma_addr_t *in, unsigned int width, unsigned int height,
2818 struct sg_table *st)
2820 unsigned int column, row;
2821 unsigned int src_idx;
2822 struct scatterlist *sg = st->sgl;
2826 for (column = 0; column < width; column++) {
2827 src_idx = width * (height - 1) + column;
2828 for (row = 0; row < height; row++) {
2830 /* We don't need the pages, but need to initialize
2831 * the entries so the sg list can be happily traversed.
2832 * The only thing we need are DMA addresses.
2834 sg_set_page(sg, NULL, PAGE_SIZE, 0);
2835 sg_dma_address(sg) = in[src_idx];
2836 sg_dma_len(sg) = PAGE_SIZE;
2843 static struct sg_table *
2844 intel_rotate_fb_obj_pages(struct i915_ggtt_view *ggtt_view,
2845 struct drm_i915_gem_object *obj)
2847 struct intel_rotation_info *rot_info = &ggtt_view->rotation_info;
2848 unsigned int size_pages = rot_info->size >> PAGE_SHIFT;
2849 struct sg_page_iter sg_iter;
2851 dma_addr_t *page_addr_list;
2852 struct sg_table *st;
2855 /* Allocate a temporary list of source pages for random access. */
2856 page_addr_list = drm_malloc_ab(obj->base.size / PAGE_SIZE,
2857 sizeof(dma_addr_t));
2858 if (!page_addr_list)
2859 return ERR_PTR(ret);
2861 /* Allocate target SG list. */
2862 st = kmalloc(sizeof(*st), M_DRM, M_WAITOK);
2866 ret = sg_alloc_table(st, size_pages, GFP_KERNEL);
2870 /* Populate source page list from the object. */
2872 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) {
2873 page_addr_list[i] = sg_page_iter_dma_address(&sg_iter);
2877 /* Rotate the pages. */
2878 rotate_pages(page_addr_list,
2879 rot_info->width_pages, rot_info->height_pages,
2883 "Created rotated page mapping for object size %zu (pitch=%u, height=%u, pixel_format=0x%x, %ux%u tiles, %u pages).\n",
2884 obj->base.size, rot_info->pitch, rot_info->height,
2885 rot_info->pixel_format, rot_info->width_pages,
2886 rot_info->height_pages, size_pages);
2888 drm_free_large(page_addr_list);
2895 drm_free_large(page_addr_list);
2898 "Failed to create rotated mapping for object size %zu! (%d) (pitch=%u, height=%u, pixel_format=0x%x, %ux%u tiles, %u pages)\n",
2899 obj->base.size, ret, rot_info->pitch, rot_info->height,
2900 rot_info->pixel_format, rot_info->width_pages,
2901 rot_info->height_pages, size_pages);
2902 return ERR_PTR(ret);
2905 static struct sg_table *
2906 intel_partial_pages(const struct i915_ggtt_view *view,
2907 struct drm_i915_gem_object *obj)
2909 struct sg_table *st;
2910 struct scatterlist *sg;
2911 struct sg_page_iter obj_sg_iter;
2914 st = kmalloc(sizeof(*st), M_DRM, M_WAITOK);
2918 ret = sg_alloc_table(st, view->params.partial.size, GFP_KERNEL);
2924 for_each_sg_page(obj->pages->sgl, &obj_sg_iter, obj->pages->nents,
2925 view->params.partial.offset)
2927 if (st->nents >= view->params.partial.size)
2930 sg_set_page(sg, NULL, PAGE_SIZE, 0);
2931 sg_dma_address(sg) = sg_page_iter_dma_address(&obj_sg_iter);
2932 sg_dma_len(sg) = PAGE_SIZE;
2943 return ERR_PTR(ret);
2947 i915_get_ggtt_vma_pages(struct i915_vma *vma)
2951 if (vma->ggtt_view.pages)
2954 if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL)
2955 vma->ggtt_view.pages = vma->obj->pages;
2956 else if (vma->ggtt_view.type == I915_GGTT_VIEW_ROTATED)
2957 vma->ggtt_view.pages =
2958 intel_rotate_fb_obj_pages(&vma->ggtt_view, vma->obj);
2959 else if (vma->ggtt_view.type == I915_GGTT_VIEW_PARTIAL)
2960 vma->ggtt_view.pages =
2961 intel_partial_pages(&vma->ggtt_view, vma->obj);
2963 WARN_ONCE(1, "GGTT view %u not implemented!\n",
2964 vma->ggtt_view.type);
2966 if (!vma->ggtt_view.pages) {
2967 DRM_ERROR("Failed to get pages for GGTT view type %u!\n",
2968 vma->ggtt_view.type);
2970 } else if (IS_ERR(vma->ggtt_view.pages)) {
2971 ret = PTR_ERR(vma->ggtt_view.pages);
2972 vma->ggtt_view.pages = NULL;
2973 DRM_ERROR("Failed to get pages for VMA view type %u (%d)!\n",
2974 vma->ggtt_view.type, ret);
2981 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
2983 * @cache_level: mapping cache level
2984 * @flags: flags like global or local mapping
2986 * DMA addresses are taken from the scatter-gather table of this object (or of
2987 * this VMA in case of non-default GGTT views) and PTE entries set up.
2988 * Note that DMA addresses are also the only part of the SG table we care about.
2990 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
2996 if (WARN_ON(flags == 0))
3000 if (flags & PIN_GLOBAL)
3001 bind_flags |= GLOBAL_BIND;
3002 if (flags & PIN_USER)
3003 bind_flags |= LOCAL_BIND;
3005 if (flags & PIN_UPDATE)
3006 bind_flags |= vma->bound;
3008 bind_flags &= ~vma->bound;
3010 if (bind_flags == 0)
3013 if (vma->bound == 0 && vma->vm->allocate_va_range) {
3014 trace_i915_va_alloc(vma->vm,
3017 VM_TO_TRACE_NAME(vma->vm));
3019 /* XXX: i915_vma_pin() will fix this +- hack */
3021 ret = vma->vm->allocate_va_range(vma->vm,
3029 ret = vma->vm->bind_vma(vma, cache_level, bind_flags);
3033 vma->bound |= bind_flags;
3039 * i915_ggtt_view_size - Get the size of a GGTT view.
3040 * @obj: Object the view is of.
3041 * @view: The view in question.
3043 * @return The size of the GGTT view in bytes.
3046 i915_ggtt_view_size(struct drm_i915_gem_object *obj,
3047 const struct i915_ggtt_view *view)
3049 if (view->type == I915_GGTT_VIEW_NORMAL) {
3050 return obj->base.size;
3051 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
3052 return view->rotation_info.size;
3053 } else if (view->type == I915_GGTT_VIEW_PARTIAL) {
3054 return view->params.partial.size << PAGE_SHIFT;
3056 WARN_ONCE(1, "GGTT view %u not implemented!\n", view->type);
3057 return obj->base.size;