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).
95 const struct i915_ggtt_view i915_ggtt_view_normal;
97 static void bdw_setup_private_ppat(struct drm_i915_private *dev_priv);
98 static void chv_setup_private_ppat(struct drm_i915_private *dev_priv);
100 static int sanitize_enable_ppgtt(struct drm_device *dev, int enable_ppgtt)
102 bool has_aliasing_ppgtt;
105 has_aliasing_ppgtt = INTEL_INFO(dev)->gen >= 6;
106 has_full_ppgtt = INTEL_INFO(dev)->gen >= 7;
108 if (intel_vgpu_active(dev))
109 has_full_ppgtt = false; /* emulation is too hard */
112 * We don't allow disabling PPGTT for gen9+ as it's a requirement for
113 * execlists, the sole mechanism available to submit work.
115 if (INTEL_INFO(dev)->gen < 9 &&
116 (enable_ppgtt == 0 || !has_aliasing_ppgtt))
119 if (enable_ppgtt == 1)
122 if (enable_ppgtt == 2 && has_full_ppgtt)
125 #ifdef CONFIG_INTEL_IOMMU
126 /* Disable ppgtt on SNB if VT-d is on. */
127 if (INTEL_INFO(dev)->gen == 6 && intel_iommu_gfx_mapped) {
128 DRM_INFO("Disabling PPGTT because VT-d is on\n");
133 /* Early VLV doesn't have this */
134 if (IS_VALLEYVIEW(dev) && !IS_CHERRYVIEW(dev) &&
135 dev->pdev->revision < 0xb) {
136 DRM_DEBUG_DRIVER("disabling PPGTT on pre-B3 step VLV\n");
140 if (INTEL_INFO(dev)->gen >= 8 && i915.enable_execlists)
143 return has_aliasing_ppgtt ? 1 : 0;
146 static void ppgtt_bind_vma(struct i915_vma *vma,
147 enum i915_cache_level cache_level,
149 static void ppgtt_unbind_vma(struct i915_vma *vma);
151 static inline gen8_pte_t gen8_pte_encode(dma_addr_t addr,
152 enum i915_cache_level level,
155 gen8_pte_t pte = valid ? _PAGE_PRESENT | _PAGE_RW : 0;
159 case I915_CACHE_NONE:
160 pte |= PPAT_UNCACHED_INDEX;
163 pte |= PPAT_DISPLAY_ELLC_INDEX;
166 pte |= PPAT_CACHED_INDEX;
173 static inline gen8_pde_t gen8_pde_encode(struct drm_device *dev,
175 enum i915_cache_level level)
177 gen8_pde_t pde = _PAGE_PRESENT | _PAGE_RW;
179 if (level != I915_CACHE_NONE)
180 pde |= PPAT_CACHED_PDE_INDEX;
182 pde |= PPAT_UNCACHED_INDEX;
186 static gen6_pte_t snb_pte_encode(dma_addr_t addr,
187 enum i915_cache_level level,
188 bool valid, u32 unused)
190 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
191 pte |= GEN6_PTE_ADDR_ENCODE(addr);
194 case I915_CACHE_L3_LLC:
196 pte |= GEN6_PTE_CACHE_LLC;
198 case I915_CACHE_NONE:
199 pte |= GEN6_PTE_UNCACHED;
208 static gen6_pte_t ivb_pte_encode(dma_addr_t addr,
209 enum i915_cache_level level,
210 bool valid, u32 unused)
212 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
213 pte |= GEN6_PTE_ADDR_ENCODE(addr);
216 case I915_CACHE_L3_LLC:
217 pte |= GEN7_PTE_CACHE_L3_LLC;
220 pte |= GEN6_PTE_CACHE_LLC;
222 case I915_CACHE_NONE:
223 pte |= GEN6_PTE_UNCACHED;
232 static gen6_pte_t byt_pte_encode(dma_addr_t addr,
233 enum i915_cache_level level,
234 bool valid, u32 flags)
236 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
237 pte |= GEN6_PTE_ADDR_ENCODE(addr);
239 if (!(flags & PTE_READ_ONLY))
240 pte |= BYT_PTE_WRITEABLE;
242 if (level != I915_CACHE_NONE)
243 pte |= BYT_PTE_SNOOPED_BY_CPU_CACHES;
248 static gen6_pte_t hsw_pte_encode(dma_addr_t addr,
249 enum i915_cache_level level,
250 bool valid, u32 unused)
252 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
253 pte |= HSW_PTE_ADDR_ENCODE(addr);
255 if (level != I915_CACHE_NONE)
256 pte |= HSW_WB_LLC_AGE3;
261 static gen6_pte_t iris_pte_encode(dma_addr_t addr,
262 enum i915_cache_level level,
263 bool valid, u32 unused)
265 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
266 pte |= HSW_PTE_ADDR_ENCODE(addr);
269 case I915_CACHE_NONE:
272 pte |= HSW_WT_ELLC_LLC_AGE3;
275 pte |= HSW_WB_ELLC_LLC_AGE3;
282 #define i915_dma_unmap_single(px, dev) \
283 __i915_dma_unmap_single((px)->daddr, dev)
285 static inline void __i915_dma_unmap_single(dma_addr_t daddr,
286 struct drm_device *dev)
288 struct device *device = &dev->pdev->dev;
290 dma_unmap_page(device, daddr, 4096, PCI_DMA_BIDIRECTIONAL);
294 * i915_dma_map_single() - Create a dma mapping for a page table/dir/etc.
295 * @px: Page table/dir/etc to get a DMA map for
298 * Page table allocations are unified across all gens. They always require a
299 * single 4k allocation, as well as a DMA mapping. If we keep the structs
300 * symmetric here, the simple macro covers us for every page table type.
302 * Return: 0 if success.
304 #define i915_dma_map_single(px, dev) \
305 i915_dma_map_page_single((px)->page, (dev), &(px)->daddr)
307 static inline int i915_dma_map_page_single(struct page *page,
308 struct drm_device *dev,
311 struct device *device = &dev->pdev->dev;
313 *daddr = dma_map_page(device, page, 0, 4096, PCI_DMA_BIDIRECTIONAL);
314 return dma_mapping_error(device, *daddr);
317 static void unmap_and_free_pt(struct i915_page_table_entry *pt,
318 struct drm_device *dev)
320 if (WARN_ON(!pt->page))
323 i915_dma_unmap_single(pt, dev);
324 __free_page(pt->page);
325 kfree(pt->used_ptes);
329 static struct i915_page_table_entry *alloc_pt_single(struct drm_device *dev)
331 struct i915_page_table_entry *pt;
332 const size_t count = INTEL_INFO(dev)->gen >= 8 ?
333 GEN8_PTES : GEN6_PTES;
336 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
338 return ERR_PTR(-ENOMEM);
340 pt->used_ptes = kcalloc(BITS_TO_LONGS(count), sizeof(*pt->used_ptes),
346 pt->page = alloc_page(GFP_KERNEL | __GFP_ZERO);
350 ret = i915_dma_map_single(pt, dev);
357 __free_page(pt->page);
359 kfree(pt->used_ptes);
367 * alloc_pt_range() - Allocate a multiple page tables
368 * @pd: The page directory which will have at least @count entries
369 * available to point to the allocated page tables.
370 * @pde: First page directory entry for which we are allocating.
371 * @count: Number of pages to allocate.
374 * Allocates multiple page table pages and sets the appropriate entries in the
375 * page table structure within the page directory. Function cleans up after
376 * itself on any failures.
378 * Return: 0 if allocation succeeded.
380 static int alloc_pt_range(struct i915_page_directory_entry *pd, uint16_t pde, size_t count,
381 struct drm_device *dev)
385 /* 512 is the max page tables per page_directory on any platform. */
386 if (WARN_ON(pde + count > I915_PDES))
389 for (i = pde; i < pde + count; i++) {
390 struct i915_page_table_entry *pt = alloc_pt_single(dev);
396 WARN(pd->page_table[i],
397 "Leaking page directory entry %d (%p)\n",
398 i, pd->page_table[i]);
399 pd->page_table[i] = pt;
406 unmap_and_free_pt(pd->page_table[i], dev);
410 static void unmap_and_free_pd(struct i915_page_directory_entry *pd)
413 __free_page(pd->page);
418 static struct i915_page_directory_entry *alloc_pd_single(void)
420 struct i915_page_directory_entry *pd;
422 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
424 return ERR_PTR(-ENOMEM);
426 pd->page = alloc_page(GFP_KERNEL | __GFP_ZERO);
429 return ERR_PTR(-ENOMEM);
435 /* Broadwell Page Directory Pointer Descriptors */
436 static int gen8_write_pdp(struct intel_engine_cs *ring, unsigned entry,
443 ret = intel_ring_begin(ring, 6);
447 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
448 intel_ring_emit(ring, GEN8_RING_PDP_UDW(ring, entry));
449 intel_ring_emit(ring, (u32)(val >> 32));
450 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
451 intel_ring_emit(ring, GEN8_RING_PDP_LDW(ring, entry));
452 intel_ring_emit(ring, (u32)(val));
453 intel_ring_advance(ring);
458 static int gen8_mm_switch(struct i915_hw_ppgtt *ppgtt,
459 struct intel_engine_cs *ring)
463 /* bit of a hack to find the actual last used pd */
464 int used_pd = ppgtt->num_pd_entries / I915_PDES;
466 for (i = used_pd - 1; i >= 0; i--) {
467 dma_addr_t addr = ppgtt->pdp.page_directory[i]->daddr;
468 ret = gen8_write_pdp(ring, i, addr);
476 static void gen8_ppgtt_clear_range(struct i915_address_space *vm,
481 struct i915_hw_ppgtt *ppgtt =
482 container_of(vm, struct i915_hw_ppgtt, base);
483 gen8_pte_t *pt_vaddr, scratch_pte;
484 unsigned pdpe = start >> GEN8_PDPE_SHIFT & GEN8_PDPE_MASK;
485 unsigned pde = start >> GEN8_PDE_SHIFT & GEN8_PDE_MASK;
486 unsigned pte = start >> GEN8_PTE_SHIFT & GEN8_PTE_MASK;
487 unsigned num_entries = length >> PAGE_SHIFT;
488 unsigned last_pte, i;
490 scratch_pte = gen8_pte_encode(ppgtt->base.scratch.addr,
491 I915_CACHE_LLC, use_scratch);
493 while (num_entries) {
494 struct i915_page_directory_entry *pd;
495 struct i915_page_table_entry *pt;
496 struct page *page_table;
498 if (WARN_ON(!ppgtt->pdp.page_directory[pdpe]))
501 pd = ppgtt->pdp.page_directory[pdpe];
503 if (WARN_ON(!pd->page_table[pde]))
506 pt = pd->page_table[pde];
508 if (WARN_ON(!pt->page))
511 page_table = pt->page;
513 last_pte = pte + num_entries;
514 if (last_pte > GEN8_PTES)
515 last_pte = GEN8_PTES;
517 pt_vaddr = kmap_atomic(page_table);
519 for (i = pte; i < last_pte; i++) {
520 pt_vaddr[i] = scratch_pte;
524 if (!HAS_LLC(ppgtt->base.dev))
525 drm_clflush_virt_range(pt_vaddr, PAGE_SIZE);
526 kunmap_atomic(pt_vaddr);
529 if (++pde == I915_PDES) {
536 static void gen8_ppgtt_insert_entries(struct i915_address_space *vm,
537 struct sg_table *pages,
539 enum i915_cache_level cache_level, u32 unused)
541 struct i915_hw_ppgtt *ppgtt =
542 container_of(vm, struct i915_hw_ppgtt, base);
543 gen8_pte_t *pt_vaddr;
544 unsigned pdpe = start >> GEN8_PDPE_SHIFT & GEN8_PDPE_MASK;
545 unsigned pde = start >> GEN8_PDE_SHIFT & GEN8_PDE_MASK;
546 unsigned pte = start >> GEN8_PTE_SHIFT & GEN8_PTE_MASK;
547 struct sg_page_iter sg_iter;
551 for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) {
552 if (WARN_ON(pdpe >= GEN8_LEGACY_PDPES))
555 if (pt_vaddr == NULL) {
556 struct i915_page_directory_entry *pd = ppgtt->pdp.page_directory[pdpe];
557 struct i915_page_table_entry *pt = pd->page_table[pde];
558 struct page *page_table = pt->page;
560 pt_vaddr = kmap_atomic(page_table);
564 gen8_pte_encode(sg_page_iter_dma_address(&sg_iter),
566 if (++pte == GEN8_PTES) {
567 if (!HAS_LLC(ppgtt->base.dev))
568 drm_clflush_virt_range(pt_vaddr, PAGE_SIZE);
569 kunmap_atomic(pt_vaddr);
571 if (++pde == I915_PDES) {
579 if (!HAS_LLC(ppgtt->base.dev))
580 drm_clflush_virt_range(pt_vaddr, PAGE_SIZE);
581 kunmap_atomic(pt_vaddr);
585 static void gen8_free_page_tables(struct i915_page_directory_entry *pd, struct drm_device *dev)
592 for (i = 0; i < I915_PDES; i++) {
593 if (WARN_ON(!pd->page_table[i]))
596 unmap_and_free_pt(pd->page_table[i], dev);
597 pd->page_table[i] = NULL;
601 static void gen8_ppgtt_free(struct i915_hw_ppgtt *ppgtt)
605 for (i = 0; i < ppgtt->num_pd_pages; i++) {
606 if (WARN_ON(!ppgtt->pdp.page_directory[i]))
609 gen8_free_page_tables(ppgtt->pdp.page_directory[i], ppgtt->base.dev);
610 unmap_and_free_pd(ppgtt->pdp.page_directory[i]);
614 static void gen8_ppgtt_unmap_pages(struct i915_hw_ppgtt *ppgtt)
616 struct pci_dev *hwdev = ppgtt->base.dev->pdev;
619 for (i = 0; i < ppgtt->num_pd_pages; i++) {
620 /* TODO: In the future we'll support sparse mappings, so this
621 * will have to change. */
622 if (!ppgtt->pdp.page_directory[i]->daddr)
625 pci_unmap_page(hwdev, ppgtt->pdp.page_directory[i]->daddr, PAGE_SIZE,
626 PCI_DMA_BIDIRECTIONAL);
628 for (j = 0; j < I915_PDES; j++) {
629 struct i915_page_directory_entry *pd = ppgtt->pdp.page_directory[i];
630 struct i915_page_table_entry *pt;
633 if (WARN_ON(!pd->page_table[j]))
636 pt = pd->page_table[j];
640 pci_unmap_page(hwdev, addr, PAGE_SIZE,
641 PCI_DMA_BIDIRECTIONAL);
646 static void gen8_ppgtt_cleanup(struct i915_address_space *vm)
648 struct i915_hw_ppgtt *ppgtt =
649 container_of(vm, struct i915_hw_ppgtt, base);
651 gen8_ppgtt_unmap_pages(ppgtt);
652 gen8_ppgtt_free(ppgtt);
655 static int gen8_ppgtt_allocate_page_tables(struct i915_hw_ppgtt *ppgtt)
659 for (i = 0; i < ppgtt->num_pd_pages; i++) {
660 ret = alloc_pt_range(ppgtt->pdp.page_directory[i],
661 0, I915_PDES, ppgtt->base.dev);
670 gen8_free_page_tables(ppgtt->pdp.page_directory[i], ppgtt->base.dev);
675 static int gen8_ppgtt_allocate_page_directories(struct i915_hw_ppgtt *ppgtt,
680 for (i = 0; i < max_pdp; i++) {
681 ppgtt->pdp.page_directory[i] = alloc_pd_single();
682 if (IS_ERR(ppgtt->pdp.page_directory[i]))
686 ppgtt->num_pd_pages = max_pdp;
687 BUG_ON(ppgtt->num_pd_pages > GEN8_LEGACY_PDPES);
693 unmap_and_free_pd(ppgtt->pdp.page_directory[i]);
698 static int gen8_ppgtt_alloc(struct i915_hw_ppgtt *ppgtt,
703 ret = gen8_ppgtt_allocate_page_directories(ppgtt, max_pdp);
707 ret = gen8_ppgtt_allocate_page_tables(ppgtt);
711 ppgtt->num_pd_entries = max_pdp * I915_PDES;
716 gen8_ppgtt_free(ppgtt);
720 static int gen8_ppgtt_setup_page_directories(struct i915_hw_ppgtt *ppgtt,
726 pd_addr = pci_map_page(ppgtt->base.dev->pdev,
727 ppgtt->pdp.page_directory[pd]->page, 0,
728 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
730 ret = pci_dma_mapping_error(ppgtt->base.dev->pdev, pd_addr);
734 ppgtt->pdp.page_directory[pd]->daddr = pd_addr;
739 static int gen8_ppgtt_setup_page_tables(struct i915_hw_ppgtt *ppgtt,
744 struct i915_page_directory_entry *pdir = ppgtt->pdp.page_directory[pd];
745 struct i915_page_table_entry *ptab = pdir->page_table[pt];
746 struct page *p = ptab->page;
749 pt_addr = pci_map_page(ppgtt->base.dev->pdev,
750 p, 0, PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
751 ret = pci_dma_mapping_error(ppgtt->base.dev->pdev, pt_addr);
755 ptab->daddr = pt_addr;
761 * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers
762 * with a net effect resembling a 2-level page table in normal x86 terms. Each
763 * PDP represents 1GB of memory 4 * 512 * 512 * 4096 = 4GB legacy 32b address
766 * FIXME: split allocation into smaller pieces. For now we only ever do this
767 * once, but with full PPGTT, the multiple contiguous allocations will be bad.
768 * TODO: Do something with the size parameter
770 static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt, uint64_t size)
772 const int max_pdp = DIV_ROUND_UP(size, 1 << 30);
773 const int min_pt_pages = I915_PDES * max_pdp;
777 DRM_INFO("Pages will be wasted unless GTT size (%llu) is divisible by 1GB\n", size);
779 /* 1. Do all our allocations for page directories and page tables.
780 * We allocate more than was asked so that we can point the unused parts
781 * to valid entries that point to scratch page. Dynamic page tables
782 * will fix this eventually.
784 ret = gen8_ppgtt_alloc(ppgtt, GEN8_LEGACY_PDPES);
789 * 2. Create DMA mappings for the page directories and page tables.
791 for (i = 0; i < GEN8_LEGACY_PDPES; i++) {
792 ret = gen8_ppgtt_setup_page_directories(ppgtt, i);
796 for (j = 0; j < I915_PDES; j++) {
797 ret = gen8_ppgtt_setup_page_tables(ppgtt, i, j);
804 * 3. Map all the page directory entires to point to the page tables
807 * For now, the PPGTT helper functions all require that the PDEs are
808 * plugged in correctly. So we do that now/here. For aliasing PPGTT, we
809 * will never need to touch the PDEs again.
811 for (i = 0; i < GEN8_LEGACY_PDPES; i++) {
812 struct i915_page_directory_entry *pd = ppgtt->pdp.page_directory[i];
813 gen8_pde_t *pd_vaddr;
814 pd_vaddr = kmap_atomic(ppgtt->pdp.page_directory[i]->page);
815 for (j = 0; j < I915_PDES; j++) {
816 struct i915_page_table_entry *pt = pd->page_table[j];
817 dma_addr_t addr = pt->daddr;
818 pd_vaddr[j] = gen8_pde_encode(ppgtt->base.dev, addr,
821 if (!HAS_LLC(ppgtt->base.dev))
822 drm_clflush_virt_range(pd_vaddr, PAGE_SIZE);
823 kunmap_atomic(pd_vaddr);
826 ppgtt->switch_mm = gen8_mm_switch;
827 ppgtt->base.clear_range = gen8_ppgtt_clear_range;
828 ppgtt->base.insert_entries = gen8_ppgtt_insert_entries;
829 ppgtt->base.cleanup = gen8_ppgtt_cleanup;
830 ppgtt->base.start = 0;
832 /* This is the area that we advertise as usable for the caller */
833 ppgtt->base.total = max_pdp * I915_PDES * GEN8_PTES * PAGE_SIZE;
835 /* Set all ptes to a valid scratch page. Also above requested space */
836 ppgtt->base.clear_range(&ppgtt->base, 0,
837 ppgtt->num_pd_pages * GEN8_PTES * PAGE_SIZE,
840 DRM_DEBUG_DRIVER("Allocated %d pages for page directories (%d wasted)\n",
841 ppgtt->num_pd_pages, ppgtt->num_pd_pages - max_pdp);
842 DRM_DEBUG_DRIVER("Allocated %d pages for page tables (%lld wasted)\n",
843 ppgtt->num_pd_entries,
844 (ppgtt->num_pd_entries - min_pt_pages) + size % (1<<30));
848 gen8_ppgtt_unmap_pages(ppgtt);
849 gen8_ppgtt_free(ppgtt);
853 static void gen6_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
855 struct drm_i915_private *dev_priv = ppgtt->base.dev->dev_private;
856 struct i915_address_space *vm = &ppgtt->base;
857 gen6_pte_t __iomem *pd_addr;
858 gen6_pte_t scratch_pte;
862 scratch_pte = vm->pte_encode(vm->scratch.addr, I915_CACHE_LLC, true, 0);
864 pd_addr = (gen6_pte_t __iomem *)dev_priv->gtt.gsm +
865 ppgtt->pd.pd_offset / sizeof(gen6_pte_t);
867 seq_printf(m, " VM %p (pd_offset %x-%x):\n", vm,
869 ppgtt->pd.pd_offset + ppgtt->num_pd_entries);
870 for (pde = 0; pde < ppgtt->num_pd_entries; pde++) {
872 gen6_pte_t *pt_vaddr;
873 dma_addr_t pt_addr = ppgtt->pd.page_table[pde]->daddr;
874 pd_entry = readl(pd_addr + pde);
875 expected = (GEN6_PDE_ADDR_ENCODE(pt_addr) | GEN6_PDE_VALID);
877 if (pd_entry != expected)
878 seq_printf(m, "\tPDE #%d mismatch: Actual PDE: %x Expected PDE: %x\n",
882 seq_printf(m, "\tPDE: %x\n", pd_entry);
884 pt_vaddr = kmap_atomic(ppgtt->pd.page_table[pde]->page);
885 for (pte = 0; pte < GEN6_PTES; pte+=4) {
887 (pde * PAGE_SIZE * GEN6_PTES) +
891 for (i = 0; i < 4; i++)
892 if (pt_vaddr[pte + i] != scratch_pte)
897 seq_printf(m, "\t\t0x%lx [%03d,%04d]: =", va, pde, pte);
898 for (i = 0; i < 4; i++) {
899 if (pt_vaddr[pte + i] != scratch_pte)
900 seq_printf(m, " %08x", pt_vaddr[pte + i]);
902 seq_puts(m, " SCRATCH ");
906 kunmap_atomic(pt_vaddr);
910 /* Write pde (index) from the page directory @pd to the page table @pt */
911 static void gen6_write_pde(struct i915_page_directory_entry *pd,
912 const int pde, struct i915_page_table_entry *pt)
914 /* Caller needs to make sure the write completes if necessary */
915 struct i915_hw_ppgtt *ppgtt =
916 container_of(pd, struct i915_hw_ppgtt, pd);
919 pd_entry = GEN6_PDE_ADDR_ENCODE(pt->daddr);
920 pd_entry |= GEN6_PDE_VALID;
922 writel(pd_entry, ppgtt->pd_addr + pde);
925 /* Write all the page tables found in the ppgtt structure to incrementing page
927 static void gen6_write_page_range(struct drm_i915_private *dev_priv,
928 struct i915_page_directory_entry *pd,
929 uint32_t start, uint32_t length)
931 struct i915_page_table_entry *pt;
934 gen6_for_each_pde(pt, pd, start, length, temp, pde)
935 gen6_write_pde(pd, pde, pt);
937 /* Make sure write is complete before other code can use this page
938 * table. Also require for WC mapped PTEs */
939 readl(dev_priv->gtt.gsm);
942 static uint32_t get_pd_offset(struct i915_hw_ppgtt *ppgtt)
944 BUG_ON(ppgtt->pd.pd_offset & 0x3f);
946 return (ppgtt->pd.pd_offset / 64) << 16;
949 static int hsw_mm_switch(struct i915_hw_ppgtt *ppgtt,
950 struct intel_engine_cs *ring)
954 /* NB: TLBs must be flushed and invalidated before a switch */
955 ret = ring->flush(ring, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
959 ret = intel_ring_begin(ring, 6);
963 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
964 intel_ring_emit(ring, RING_PP_DIR_DCLV(ring));
965 intel_ring_emit(ring, PP_DIR_DCLV_2G);
966 intel_ring_emit(ring, RING_PP_DIR_BASE(ring));
967 intel_ring_emit(ring, get_pd_offset(ppgtt));
968 intel_ring_emit(ring, MI_NOOP);
969 intel_ring_advance(ring);
974 static int vgpu_mm_switch(struct i915_hw_ppgtt *ppgtt,
975 struct intel_engine_cs *ring)
977 struct drm_i915_private *dev_priv = to_i915(ppgtt->base.dev);
979 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
980 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
984 static int gen7_mm_switch(struct i915_hw_ppgtt *ppgtt,
985 struct intel_engine_cs *ring)
989 /* NB: TLBs must be flushed and invalidated before a switch */
990 ret = ring->flush(ring, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
994 ret = intel_ring_begin(ring, 6);
998 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
999 intel_ring_emit(ring, RING_PP_DIR_DCLV(ring));
1000 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1001 intel_ring_emit(ring, RING_PP_DIR_BASE(ring));
1002 intel_ring_emit(ring, get_pd_offset(ppgtt));
1003 intel_ring_emit(ring, MI_NOOP);
1004 intel_ring_advance(ring);
1006 /* XXX: RCS is the only one to auto invalidate the TLBs? */
1007 if (ring->id != RCS) {
1008 ret = ring->flush(ring, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1016 static int gen6_mm_switch(struct i915_hw_ppgtt *ppgtt,
1017 struct intel_engine_cs *ring)
1019 struct drm_device *dev = ppgtt->base.dev;
1020 struct drm_i915_private *dev_priv = dev->dev_private;
1023 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
1024 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
1026 POSTING_READ(RING_PP_DIR_DCLV(ring));
1031 static void gen8_ppgtt_enable(struct drm_device *dev)
1033 struct drm_i915_private *dev_priv = dev->dev_private;
1034 struct intel_engine_cs *ring;
1037 for_each_ring(ring, dev_priv, j) {
1038 I915_WRITE(RING_MODE_GEN7(ring),
1039 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1043 static void gen7_ppgtt_enable(struct drm_device *dev)
1045 struct drm_i915_private *dev_priv = dev->dev_private;
1046 struct intel_engine_cs *ring;
1047 uint32_t ecochk, ecobits;
1050 ecobits = I915_READ(GAC_ECO_BITS);
1051 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B);
1053 ecochk = I915_READ(GAM_ECOCHK);
1054 if (IS_HASWELL(dev)) {
1055 ecochk |= ECOCHK_PPGTT_WB_HSW;
1057 ecochk |= ECOCHK_PPGTT_LLC_IVB;
1058 ecochk &= ~ECOCHK_PPGTT_GFDT_IVB;
1060 I915_WRITE(GAM_ECOCHK, ecochk);
1062 for_each_ring(ring, dev_priv, i) {
1063 /* GFX_MODE is per-ring on gen7+ */
1064 I915_WRITE(RING_MODE_GEN7(ring),
1065 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1069 static void gen6_ppgtt_enable(struct drm_device *dev)
1071 struct drm_i915_private *dev_priv = dev->dev_private;
1072 uint32_t ecochk, gab_ctl, ecobits;
1074 ecobits = I915_READ(GAC_ECO_BITS);
1075 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_SNB_BIT |
1076 ECOBITS_PPGTT_CACHE64B);
1078 gab_ctl = I915_READ(GAB_CTL);
1079 I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT);
1081 ecochk = I915_READ(GAM_ECOCHK);
1082 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT | ECOCHK_PPGTT_CACHE64B);
1084 I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1087 /* PPGTT support for Sandybdrige/Gen6 and later */
1088 static void gen6_ppgtt_clear_range(struct i915_address_space *vm,
1093 struct i915_hw_ppgtt *ppgtt =
1094 container_of(vm, struct i915_hw_ppgtt, base);
1095 gen6_pte_t *pt_vaddr, scratch_pte;
1096 unsigned first_entry = start >> PAGE_SHIFT;
1097 unsigned num_entries = length >> PAGE_SHIFT;
1098 unsigned act_pt = first_entry / GEN6_PTES;
1099 unsigned first_pte = first_entry % GEN6_PTES;
1100 unsigned last_pte, i;
1102 scratch_pte = vm->pte_encode(vm->scratch.addr, I915_CACHE_LLC, true, 0);
1104 while (num_entries) {
1105 last_pte = first_pte + num_entries;
1106 if (last_pte > GEN6_PTES)
1107 last_pte = GEN6_PTES;
1109 pt_vaddr = kmap_atomic(ppgtt->pd.page_table[act_pt]->page);
1111 for (i = first_pte; i < last_pte; i++)
1112 pt_vaddr[i] = scratch_pte;
1114 kunmap_atomic(pt_vaddr);
1116 num_entries -= last_pte - first_pte;
1122 static void gen6_ppgtt_insert_entries(struct i915_address_space *vm,
1123 struct sg_table *pages,
1125 enum i915_cache_level cache_level, u32 flags)
1127 struct i915_hw_ppgtt *ppgtt =
1128 container_of(vm, struct i915_hw_ppgtt, base);
1129 gen6_pte_t *pt_vaddr;
1130 unsigned first_entry = start >> PAGE_SHIFT;
1131 unsigned act_pt = first_entry / GEN6_PTES;
1132 unsigned act_pte = first_entry % GEN6_PTES;
1133 struct sg_page_iter sg_iter;
1136 for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) {
1137 if (pt_vaddr == NULL)
1138 pt_vaddr = kmap_atomic(ppgtt->pd.page_table[act_pt]->page);
1141 vm->pte_encode(sg_page_iter_dma_address(&sg_iter),
1142 cache_level, true, flags);
1144 if (++act_pte == GEN6_PTES) {
1145 kunmap_atomic(pt_vaddr);
1152 kunmap_atomic(pt_vaddr);
1155 static void gen6_ppgtt_unmap_pages(struct i915_hw_ppgtt *ppgtt)
1159 for (i = 0; i < ppgtt->num_pd_entries; i++)
1160 pci_unmap_page(ppgtt->base.dev->pdev,
1161 ppgtt->pd.page_table[i]->daddr,
1162 4096, PCI_DMA_BIDIRECTIONAL);
1165 static int gen6_alloc_va_range(struct i915_address_space *vm,
1166 uint64_t start, uint64_t length)
1168 struct i915_hw_ppgtt *ppgtt =
1169 container_of(vm, struct i915_hw_ppgtt, base);
1170 struct i915_page_table_entry *pt;
1173 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1174 DECLARE_BITMAP(tmp_bitmap, GEN6_PTES);
1176 bitmap_zero(tmp_bitmap, GEN6_PTES);
1177 bitmap_set(tmp_bitmap, gen6_pte_index(start),
1178 gen6_pte_count(start, length));
1180 bitmap_or(pt->used_ptes, pt->used_ptes, tmp_bitmap,
1187 static void gen6_ppgtt_free(struct i915_hw_ppgtt *ppgtt)
1191 for (i = 0; i < ppgtt->num_pd_entries; i++)
1192 unmap_and_free_pt(ppgtt->pd.page_table[i], ppgtt->base.dev);
1194 unmap_and_free_pd(&ppgtt->pd);
1197 static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
1199 struct i915_hw_ppgtt *ppgtt =
1200 container_of(vm, struct i915_hw_ppgtt, base);
1202 drm_mm_remove_node(&ppgtt->node);
1204 gen6_ppgtt_unmap_pages(ppgtt);
1205 gen6_ppgtt_free(ppgtt);
1208 static int gen6_ppgtt_allocate_page_directories(struct i915_hw_ppgtt *ppgtt)
1210 struct drm_device *dev = ppgtt->base.dev;
1211 struct drm_i915_private *dev_priv = dev->dev_private;
1212 bool retried = false;
1215 /* PPGTT PDEs reside in the GGTT and consists of 512 entries. The
1216 * allocator works in address space sizes, so it's multiplied by page
1217 * size. We allocate at the top of the GTT to avoid fragmentation.
1219 BUG_ON(!drm_mm_initialized(&dev_priv->gtt.base.mm));
1221 ret = drm_mm_insert_node_in_range_generic(&dev_priv->gtt.base.mm,
1222 &ppgtt->node, GEN6_PD_SIZE,
1224 0, dev_priv->gtt.base.total,
1226 if (ret == -ENOSPC && !retried) {
1227 ret = i915_gem_evict_something(dev, &dev_priv->gtt.base,
1228 GEN6_PD_SIZE, GEN6_PD_ALIGN,
1230 0, dev_priv->gtt.base.total,
1243 if (ppgtt->node.start < dev_priv->gtt.mappable_end)
1244 DRM_DEBUG("Forced to use aperture for PDEs\n");
1246 ppgtt->num_pd_entries = I915_PDES;
1253 static int gen6_ppgtt_alloc(struct i915_hw_ppgtt *ppgtt)
1257 ret = gen6_ppgtt_allocate_page_directories(ppgtt);
1261 ret = alloc_pt_range(&ppgtt->pd, 0, ppgtt->num_pd_entries,
1265 drm_mm_remove_node(&ppgtt->node);
1272 static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
1274 struct drm_device *dev = ppgtt->base.dev;
1275 struct drm_i915_private *dev_priv = dev->dev_private;
1278 ppgtt->base.pte_encode = dev_priv->gtt.base.pte_encode;
1280 ppgtt->switch_mm = gen6_mm_switch;
1281 } else if (IS_HASWELL(dev)) {
1282 ppgtt->switch_mm = hsw_mm_switch;
1283 } else if (IS_GEN7(dev)) {
1284 ppgtt->switch_mm = gen7_mm_switch;
1288 if (intel_vgpu_active(dev))
1289 ppgtt->switch_mm = vgpu_mm_switch;
1291 ret = gen6_ppgtt_alloc(ppgtt);
1295 ppgtt->base.allocate_va_range = gen6_alloc_va_range;
1296 ppgtt->base.clear_range = gen6_ppgtt_clear_range;
1297 ppgtt->base.insert_entries = gen6_ppgtt_insert_entries;
1298 ppgtt->base.cleanup = gen6_ppgtt_cleanup;
1299 ppgtt->base.start = 0;
1300 ppgtt->base.total = ppgtt->num_pd_entries * GEN6_PTES * PAGE_SIZE;
1301 ppgtt->debug_dump = gen6_dump_ppgtt;
1303 ppgtt->pd.pd_offset =
1304 ppgtt->node.start / PAGE_SIZE * sizeof(gen6_pte_t);
1306 ppgtt->pd_addr = (gen6_pte_t __iomem *)dev_priv->gtt.gsm +
1307 ppgtt->pd.pd_offset / sizeof(gen6_pte_t);
1309 ppgtt->base.clear_range(&ppgtt->base, 0, ppgtt->base.total, true);
1311 gen6_write_page_range(dev_priv, &ppgtt->pd, 0, ppgtt->base.total);
1313 DRM_DEBUG_DRIVER("Allocated pde space (%lldM) at GTT entry: %llx\n",
1314 ppgtt->node.size >> 20,
1315 ppgtt->node.start / PAGE_SIZE);
1317 DRM_DEBUG("Adding PPGTT at offset %x\n",
1318 ppgtt->pd.pd_offset << 10);
1323 static int __hw_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
1325 struct drm_i915_private *dev_priv = dev->dev_private;
1327 ppgtt->base.dev = dev;
1328 ppgtt->base.scratch = dev_priv->gtt.base.scratch;
1330 if (INTEL_INFO(dev)->gen < 8)
1331 return gen6_ppgtt_init(ppgtt);
1333 return gen8_ppgtt_init(ppgtt, dev_priv->gtt.base.total);
1335 int i915_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
1337 struct drm_i915_private *dev_priv = dev->dev_private;
1340 ret = __hw_ppgtt_init(dev, ppgtt);
1342 kref_init(&ppgtt->ref);
1343 drm_mm_init(&ppgtt->base.mm, ppgtt->base.start,
1345 i915_init_vm(dev_priv, &ppgtt->base);
1351 int i915_ppgtt_init_hw(struct drm_device *dev)
1353 struct drm_i915_private *dev_priv = dev->dev_private;
1354 struct intel_engine_cs *ring;
1355 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
1358 /* In the case of execlists, PPGTT is enabled by the context descriptor
1359 * and the PDPs are contained within the context itself. We don't
1360 * need to do anything here. */
1361 if (i915.enable_execlists)
1364 if (!USES_PPGTT(dev))
1368 gen6_ppgtt_enable(dev);
1369 else if (IS_GEN7(dev))
1370 gen7_ppgtt_enable(dev);
1371 else if (INTEL_INFO(dev)->gen >= 8)
1372 gen8_ppgtt_enable(dev);
1374 MISSING_CASE(INTEL_INFO(dev)->gen);
1377 for_each_ring(ring, dev_priv, i) {
1378 ret = ppgtt->switch_mm(ppgtt, ring);
1386 struct i915_hw_ppgtt *
1387 i915_ppgtt_create(struct drm_device *dev, struct drm_i915_file_private *fpriv)
1389 struct i915_hw_ppgtt *ppgtt;
1392 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
1394 return ERR_PTR(-ENOMEM);
1396 ret = i915_ppgtt_init(dev, ppgtt);
1399 return ERR_PTR(ret);
1402 ppgtt->file_priv = fpriv;
1404 trace_i915_ppgtt_create(&ppgtt->base);
1409 void i915_ppgtt_release(struct kref *kref)
1411 struct i915_hw_ppgtt *ppgtt =
1412 container_of(kref, struct i915_hw_ppgtt, ref);
1414 trace_i915_ppgtt_release(&ppgtt->base);
1416 /* vmas should already be unbound */
1417 WARN_ON(!list_empty(&ppgtt->base.active_list));
1418 WARN_ON(!list_empty(&ppgtt->base.inactive_list));
1420 list_del(&ppgtt->base.global_link);
1421 drm_mm_takedown(&ppgtt->base.mm);
1423 ppgtt->base.cleanup(&ppgtt->base);
1428 ppgtt_bind_vma(struct i915_vma *vma,
1429 enum i915_cache_level cache_level,
1432 /* Currently applicable only to VLV */
1433 if (vma->obj->gt_ro)
1434 flags |= PTE_READ_ONLY;
1436 vma->vm->insert_entries(vma->vm, vma->obj->pages, vma->node.start,
1437 cache_level, flags);
1440 static void ppgtt_unbind_vma(struct i915_vma *vma)
1442 vma->vm->clear_range(vma->vm,
1444 vma->obj->base.size,
1448 extern int intel_iommu_gfx_mapped;
1449 /* Certain Gen5 chipsets require require idling the GPU before
1450 * unmapping anything from the GTT when VT-d is enabled.
1452 static inline bool needs_idle_maps(struct drm_device *dev)
1454 #ifdef CONFIG_INTEL_IOMMU
1455 /* Query intel_iommu to see if we need the workaround. Presumably that
1458 if (IS_GEN5(dev) && IS_MOBILE(dev) && intel_iommu_gfx_mapped)
1464 static bool do_idling(struct drm_i915_private *dev_priv)
1466 bool ret = dev_priv->mm.interruptible;
1468 if (unlikely(dev_priv->gtt.do_idle_maps)) {
1469 dev_priv->mm.interruptible = false;
1470 if (i915_gpu_idle(dev_priv->dev)) {
1471 DRM_ERROR("Couldn't idle GPU\n");
1472 /* Wait a bit, in hopes it avoids the hang */
1480 static void undo_idling(struct drm_i915_private *dev_priv, bool interruptible)
1482 if (unlikely(dev_priv->gtt.do_idle_maps))
1483 dev_priv->mm.interruptible = interruptible;
1486 void i915_check_and_clear_faults(struct drm_device *dev)
1488 struct drm_i915_private *dev_priv = dev->dev_private;
1489 struct intel_engine_cs *ring;
1492 if (INTEL_INFO(dev)->gen < 6)
1495 for_each_ring(ring, dev_priv, i) {
1497 fault_reg = I915_READ(RING_FAULT_REG(ring));
1498 if (fault_reg & RING_FAULT_VALID) {
1499 DRM_DEBUG_DRIVER("Unexpected fault\n"
1501 "\tAddress space: %s\n"
1504 fault_reg & PAGE_MASK,
1505 fault_reg & RING_FAULT_GTTSEL_MASK ? "GGTT" : "PPGTT",
1506 RING_FAULT_SRCID(fault_reg),
1507 RING_FAULT_FAULT_TYPE(fault_reg));
1508 I915_WRITE(RING_FAULT_REG(ring),
1509 fault_reg & ~RING_FAULT_VALID);
1512 POSTING_READ(RING_FAULT_REG(&dev_priv->ring[RCS]));
1515 static void i915_ggtt_flush(struct drm_i915_private *dev_priv)
1517 if (INTEL_INFO(dev_priv->dev)->gen < 6) {
1518 intel_gtt_chipset_flush();
1520 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
1521 POSTING_READ(GFX_FLSH_CNTL_GEN6);
1525 void i915_gem_suspend_gtt_mappings(struct drm_device *dev)
1527 struct drm_i915_private *dev_priv = dev->dev_private;
1529 /* Don't bother messing with faults pre GEN6 as we have little
1530 * documentation supporting that it's a good idea.
1532 if (INTEL_INFO(dev)->gen < 6)
1535 i915_check_and_clear_faults(dev);
1537 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
1538 dev_priv->gtt.base.start,
1539 dev_priv->gtt.base.total,
1542 i915_ggtt_flush(dev_priv);
1545 void i915_gem_restore_gtt_mappings(struct drm_device *dev)
1547 struct drm_i915_private *dev_priv = dev->dev_private;
1548 struct drm_i915_gem_object *obj;
1549 struct i915_address_space *vm;
1551 i915_check_and_clear_faults(dev);
1553 /* First fill our portion of the GTT with scratch pages */
1554 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
1555 dev_priv->gtt.base.start,
1556 dev_priv->gtt.base.total,
1559 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
1560 struct i915_vma *vma = i915_gem_obj_to_vma(obj,
1561 &dev_priv->gtt.base);
1565 i915_gem_clflush_object(obj, obj->pin_display);
1566 /* The bind_vma code tries to be smart about tracking mappings.
1567 * Unfortunately above, we've just wiped out the mappings
1568 * without telling our object about it. So we need to fake it.
1570 * Bind is not expected to fail since this is only called on
1571 * resume and assumption is all requirements exist already.
1573 vma->bound &= ~GLOBAL_BIND;
1574 WARN_ON(i915_vma_bind(vma, obj->cache_level, GLOBAL_BIND));
1578 if (INTEL_INFO(dev)->gen >= 8) {
1579 if (IS_CHERRYVIEW(dev))
1580 chv_setup_private_ppat(dev_priv);
1582 bdw_setup_private_ppat(dev_priv);
1587 if (USES_PPGTT(dev)) {
1588 list_for_each_entry(vm, &dev_priv->vm_list, global_link) {
1589 /* TODO: Perhaps it shouldn't be gen6 specific */
1591 struct i915_hw_ppgtt *ppgtt =
1592 container_of(vm, struct i915_hw_ppgtt,
1595 if (i915_is_ggtt(vm))
1596 ppgtt = dev_priv->mm.aliasing_ppgtt;
1598 gen6_write_page_range(dev_priv, &ppgtt->pd,
1599 0, ppgtt->base.total);
1603 i915_ggtt_flush(dev_priv);
1606 int i915_gem_gtt_prepare_object(struct drm_i915_gem_object *obj)
1608 if (obj->has_dma_mapping)
1611 if (!dma_map_sg(&obj->base.dev->pdev->dev,
1612 obj->pages->sgl, obj->pages->nents,
1613 PCI_DMA_BIDIRECTIONAL))
1619 static inline void gen8_set_pte(void __iomem *addr, gen8_pte_t pte)
1624 iowrite32((u32)pte, addr);
1625 iowrite32(pte >> 32, addr + 4);
1629 static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
1630 struct sg_table *st,
1632 enum i915_cache_level level, u32 unused)
1634 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1635 unsigned first_entry = start >> PAGE_SHIFT;
1636 gen8_pte_t __iomem *gtt_entries =
1637 (gen8_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
1639 struct sg_page_iter sg_iter;
1640 dma_addr_t addr = 0; /* shut up gcc */
1642 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
1643 addr = sg_dma_address(sg_iter.sg) +
1644 (sg_iter.sg_pgoffset << PAGE_SHIFT);
1645 gen8_set_pte(>t_entries[i],
1646 gen8_pte_encode(addr, level, true));
1651 * XXX: This serves as a posting read to make sure that the PTE has
1652 * actually been updated. There is some concern that even though
1653 * registers and PTEs are within the same BAR that they are potentially
1654 * of NUMA access patterns. Therefore, even with the way we assume
1655 * hardware should work, we must keep this posting read for paranoia.
1658 WARN_ON(readq(>t_entries[i-1])
1659 != gen8_pte_encode(addr, level, true));
1661 /* This next bit makes the above posting read even more important. We
1662 * want to flush the TLBs only after we're certain all the PTE updates
1665 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
1666 POSTING_READ(GFX_FLSH_CNTL_GEN6);
1670 * Binds an object into the global gtt with the specified cache level. The object
1671 * will be accessible to the GPU via commands whose operands reference offsets
1672 * within the global GTT as well as accessible by the GPU through the GMADR
1673 * mapped BAR (dev_priv->mm.gtt->gtt).
1675 static void gen6_ggtt_insert_entries(struct i915_address_space *vm,
1676 struct sg_table *st,
1678 enum i915_cache_level level, u32 flags)
1680 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1681 unsigned first_entry = start >> PAGE_SHIFT;
1682 gen6_pte_t __iomem *gtt_entries =
1683 (gen6_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
1685 struct sg_page_iter sg_iter;
1686 dma_addr_t addr = 0;
1688 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
1689 addr = sg_page_iter_dma_address(&sg_iter);
1690 iowrite32(vm->pte_encode(addr, level, true, flags), >t_entries[i]);
1694 /* XXX: This serves as a posting read to make sure that the PTE has
1695 * actually been updated. There is some concern that even though
1696 * registers and PTEs are within the same BAR that they are potentially
1697 * of NUMA access patterns. Therefore, even with the way we assume
1698 * hardware should work, we must keep this posting read for paranoia.
1701 unsigned long gtt = readl(>t_entries[i-1]);
1702 WARN_ON(gtt != vm->pte_encode(addr, level, true, flags));
1705 /* This next bit makes the above posting read even more important. We
1706 * want to flush the TLBs only after we're certain all the PTE updates
1709 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
1710 POSTING_READ(GFX_FLSH_CNTL_GEN6);
1713 static void gen8_ggtt_clear_range(struct i915_address_space *vm,
1718 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1719 unsigned first_entry = start >> PAGE_SHIFT;
1720 unsigned num_entries = length >> PAGE_SHIFT;
1721 gen8_pte_t scratch_pte, __iomem *gtt_base =
1722 (gen8_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
1723 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
1726 if (WARN(num_entries > max_entries,
1727 "First entry = %d; Num entries = %d (max=%d)\n",
1728 first_entry, num_entries, max_entries))
1729 num_entries = max_entries;
1731 scratch_pte = gen8_pte_encode(vm->scratch.addr,
1734 for (i = 0; i < num_entries; i++)
1735 gen8_set_pte(>t_base[i], scratch_pte);
1739 static void gen6_ggtt_clear_range(struct i915_address_space *vm,
1744 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1745 unsigned first_entry = start >> PAGE_SHIFT;
1746 unsigned num_entries = length >> PAGE_SHIFT;
1747 gen6_pte_t scratch_pte, __iomem *gtt_base =
1748 (gen6_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
1749 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
1752 if (WARN(num_entries > max_entries,
1753 "First entry = %d; Num entries = %d (max=%d)\n",
1754 first_entry, num_entries, max_entries))
1755 num_entries = max_entries;
1757 scratch_pte = vm->pte_encode(vm->scratch.addr, I915_CACHE_LLC, use_scratch, 0);
1759 for (i = 0; i < num_entries; i++)
1760 iowrite32(scratch_pte, >t_base[i]);
1765 static void i915_ggtt_bind_vma(struct i915_vma *vma,
1766 enum i915_cache_level cache_level,
1769 const unsigned long entry = vma->node.start >> PAGE_SHIFT;
1770 unsigned int flags = (cache_level == I915_CACHE_NONE) ?
1771 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
1773 BUG_ON(!i915_is_ggtt(vma->vm));
1774 intel_gtt_insert_sg_entries(vma->ggtt_view.pages, entry, flags);
1775 vma->bound = GLOBAL_BIND;
1778 static void i915_ggtt_clear_range(struct i915_address_space *vm,
1783 unsigned first_entry = start >> PAGE_SHIFT;
1784 unsigned num_entries = length >> PAGE_SHIFT;
1785 intel_gtt_clear_range(first_entry, num_entries);
1788 static void i915_ggtt_unbind_vma(struct i915_vma *vma)
1790 const unsigned int first = vma->node.start >> PAGE_SHIFT;
1791 const unsigned int size = vma->obj->base.size >> PAGE_SHIFT;
1793 BUG_ON(!i915_is_ggtt(vma->vm));
1795 intel_gtt_clear_range(first, size);
1798 static void ggtt_bind_vma(struct i915_vma *vma,
1799 enum i915_cache_level cache_level,
1802 struct drm_device *dev = vma->vm->dev;
1803 struct drm_i915_private *dev_priv = dev->dev_private;
1804 struct drm_i915_gem_object *obj = vma->obj;
1805 struct sg_table *pages = obj->pages;
1807 /* Currently applicable only to VLV */
1809 flags |= PTE_READ_ONLY;
1811 if (i915_is_ggtt(vma->vm))
1812 pages = vma->ggtt_view.pages;
1814 /* If there is no aliasing PPGTT, or the caller needs a global mapping,
1815 * or we have a global mapping already but the cacheability flags have
1816 * changed, set the global PTEs.
1818 * If there is an aliasing PPGTT it is anecdotally faster, so use that
1819 * instead if none of the above hold true.
1821 * NB: A global mapping should only be needed for special regions like
1822 * "gtt mappable", SNB errata, or if specified via special execbuf
1823 * flags. At all other times, the GPU will use the aliasing PPGTT.
1825 if (!dev_priv->mm.aliasing_ppgtt || flags & GLOBAL_BIND) {
1826 if (!(vma->bound & GLOBAL_BIND) ||
1827 (cache_level != obj->cache_level)) {
1828 vma->vm->insert_entries(vma->vm, pages,
1830 cache_level, flags);
1831 vma->bound |= GLOBAL_BIND;
1835 if (dev_priv->mm.aliasing_ppgtt &&
1836 (!(vma->bound & LOCAL_BIND) ||
1837 (cache_level != obj->cache_level))) {
1838 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
1839 appgtt->base.insert_entries(&appgtt->base, pages,
1841 cache_level, flags);
1842 vma->bound |= LOCAL_BIND;
1846 static void ggtt_unbind_vma(struct i915_vma *vma)
1848 struct drm_device *dev = vma->vm->dev;
1849 struct drm_i915_private *dev_priv = dev->dev_private;
1850 struct drm_i915_gem_object *obj = vma->obj;
1852 if (vma->bound & GLOBAL_BIND) {
1853 vma->vm->clear_range(vma->vm,
1857 vma->bound &= ~GLOBAL_BIND;
1860 if (vma->bound & LOCAL_BIND) {
1861 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
1862 appgtt->base.clear_range(&appgtt->base,
1866 vma->bound &= ~LOCAL_BIND;
1870 void i915_gem_gtt_finish_object(struct drm_i915_gem_object *obj)
1872 struct drm_device *dev = obj->base.dev;
1873 struct drm_i915_private *dev_priv = dev->dev_private;
1876 interruptible = do_idling(dev_priv);
1878 if (!obj->has_dma_mapping)
1879 dma_unmap_sg(&dev->pdev->dev,
1880 obj->pages->sgl, obj->pages->nents,
1881 PCI_DMA_BIDIRECTIONAL);
1883 undo_idling(dev_priv, interruptible);
1886 static void i915_gtt_color_adjust(struct drm_mm_node *node,
1887 unsigned long color,
1891 if (node->color != color)
1894 if (!list_empty(&node->node_list)) {
1895 node = list_entry(node->node_list.next,
1898 if (node->allocated && node->color != color)
1903 static int i915_gem_setup_global_gtt(struct drm_device *dev,
1904 unsigned long start,
1905 unsigned long mappable_end,
1908 /* Let GEM Manage all of the aperture.
1910 * However, leave one page at the end still bound to the scratch page.
1911 * There are a number of places where the hardware apparently prefetches
1912 * past the end of the object, and we've seen multiple hangs with the
1913 * GPU head pointer stuck in a batchbuffer bound at the last page of the
1914 * aperture. One page should be enough to keep any prefetching inside
1917 struct drm_i915_private *dev_priv = dev->dev_private;
1918 struct i915_address_space *ggtt_vm = &dev_priv->gtt.base;
1919 struct drm_mm_node *entry;
1920 struct drm_i915_gem_object *obj;
1921 unsigned long hole_start, hole_end;
1924 BUG_ON(mappable_end > end);
1926 /* Subtract the guard page ... */
1927 drm_mm_init(&ggtt_vm->mm, start, end - start - PAGE_SIZE);
1929 dev_priv->gtt.base.start = start;
1930 dev_priv->gtt.base.total = end - start;
1932 if (intel_vgpu_active(dev)) {
1933 ret = intel_vgt_balloon(dev);
1939 dev_priv->gtt.base.mm.color_adjust = i915_gtt_color_adjust;
1941 /* Mark any preallocated objects as occupied */
1942 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
1943 struct i915_vma *vma = i915_gem_obj_to_vma(obj, ggtt_vm);
1945 DRM_DEBUG_KMS("reserving preallocated space: %lx + %zx\n",
1946 i915_gem_obj_ggtt_offset(obj), obj->base.size);
1948 WARN_ON(i915_gem_obj_ggtt_bound(obj));
1949 ret = drm_mm_reserve_node(&ggtt_vm->mm, &vma->node);
1951 DRM_DEBUG_KMS("Reservation failed: %i\n", ret);
1954 vma->bound |= GLOBAL_BIND;
1957 /* Clear any non-preallocated blocks */
1958 drm_mm_for_each_hole(entry, &ggtt_vm->mm, hole_start, hole_end) {
1959 DRM_DEBUG_KMS("clearing unused GTT space: [%lx, %lx]\n",
1960 hole_start, hole_end);
1961 ggtt_vm->clear_range(ggtt_vm, hole_start,
1962 hole_end - hole_start, true);
1965 /* And finally clear the reserved guard page */
1966 ggtt_vm->clear_range(ggtt_vm, end - PAGE_SIZE, PAGE_SIZE, true);
1968 if (USES_PPGTT(dev) && !USES_FULL_PPGTT(dev)) {
1969 struct i915_hw_ppgtt *ppgtt;
1971 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
1975 ret = __hw_ppgtt_init(dev, ppgtt);
1979 dev_priv->mm.aliasing_ppgtt = ppgtt;
1985 void i915_gem_init_global_gtt(struct drm_device *dev)
1987 struct drm_i915_private *dev_priv = dev->dev_private;
1988 unsigned long gtt_size, mappable_size;
1990 gtt_size = dev_priv->gtt.base.total;
1991 mappable_size = dev_priv->gtt.mappable_end;
1993 i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
1996 void i915_global_gtt_cleanup(struct drm_device *dev)
1998 struct drm_i915_private *dev_priv = dev->dev_private;
1999 struct i915_address_space *vm = &dev_priv->gtt.base;
2001 if (dev_priv->mm.aliasing_ppgtt) {
2002 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
2004 ppgtt->base.cleanup(&ppgtt->base);
2007 if (drm_mm_initialized(&vm->mm)) {
2008 if (intel_vgpu_active(dev))
2009 intel_vgt_deballoon();
2011 drm_mm_takedown(&vm->mm);
2012 list_del(&vm->global_link);
2018 static int setup_scratch_page(struct drm_device *dev)
2020 struct drm_i915_private *dev_priv = dev->dev_private;
2022 dma_addr_t dma_addr;
2024 page = alloc_page(GFP_KERNEL | GFP_DMA32 | __GFP_ZERO);
2027 set_pages_uc(page, 1);
2029 #ifdef CONFIG_INTEL_IOMMU
2030 dma_addr = pci_map_page(dev->pdev, page, 0, PAGE_SIZE,
2031 PCI_DMA_BIDIRECTIONAL);
2032 if (pci_dma_mapping_error(dev->pdev, dma_addr))
2035 dma_addr = page_to_phys(page);
2037 dev_priv->gtt.base.scratch.page = page;
2038 dev_priv->gtt.base.scratch.addr = dma_addr;
2043 static void teardown_scratch_page(struct drm_device *dev)
2045 struct drm_i915_private *dev_priv = dev->dev_private;
2046 struct page *page = dev_priv->gtt.base.scratch.page;
2048 set_pages_wb(page, 1);
2049 pci_unmap_page(dev->pdev, dev_priv->gtt.base.scratch.addr,
2050 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
2054 static inline unsigned int gen6_get_total_gtt_size(u16 snb_gmch_ctl)
2056 snb_gmch_ctl >>= SNB_GMCH_GGMS_SHIFT;
2057 snb_gmch_ctl &= SNB_GMCH_GGMS_MASK;
2058 return snb_gmch_ctl << 20;
2061 static inline unsigned int gen8_get_total_gtt_size(u16 bdw_gmch_ctl)
2063 bdw_gmch_ctl >>= BDW_GMCH_GGMS_SHIFT;
2064 bdw_gmch_ctl &= BDW_GMCH_GGMS_MASK;
2066 bdw_gmch_ctl = 1 << bdw_gmch_ctl;
2068 #ifdef CONFIG_X86_32
2069 /* Limit 32b platforms to a 2GB GGTT: 4 << 20 / pte size * PAGE_SIZE */
2070 if (bdw_gmch_ctl > 4)
2074 return bdw_gmch_ctl << 20;
2077 static inline unsigned int chv_get_total_gtt_size(u16 gmch_ctrl)
2079 gmch_ctrl >>= SNB_GMCH_GGMS_SHIFT;
2080 gmch_ctrl &= SNB_GMCH_GGMS_MASK;
2083 return 1 << (20 + gmch_ctrl);
2088 static inline size_t gen6_get_stolen_size(u16 snb_gmch_ctl)
2090 snb_gmch_ctl >>= SNB_GMCH_GMS_SHIFT;
2091 snb_gmch_ctl &= SNB_GMCH_GMS_MASK;
2092 return snb_gmch_ctl << 25; /* 32 MB units */
2095 static inline size_t gen8_get_stolen_size(u16 bdw_gmch_ctl)
2097 bdw_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2098 bdw_gmch_ctl &= BDW_GMCH_GMS_MASK;
2099 return bdw_gmch_ctl << 25; /* 32 MB units */
2102 static size_t chv_get_stolen_size(u16 gmch_ctrl)
2104 gmch_ctrl >>= SNB_GMCH_GMS_SHIFT;
2105 gmch_ctrl &= SNB_GMCH_GMS_MASK;
2108 * 0x0 to 0x10: 32MB increments starting at 0MB
2109 * 0x11 to 0x16: 4MB increments starting at 8MB
2110 * 0x17 to 0x1d: 4MB increments start at 36MB
2112 if (gmch_ctrl < 0x11)
2113 return gmch_ctrl << 25;
2114 else if (gmch_ctrl < 0x17)
2115 return (gmch_ctrl - 0x11 + 2) << 22;
2117 return (gmch_ctrl - 0x17 + 9) << 22;
2120 static size_t gen9_get_stolen_size(u16 gen9_gmch_ctl)
2122 gen9_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2123 gen9_gmch_ctl &= BDW_GMCH_GMS_MASK;
2125 if (gen9_gmch_ctl < 0xf0)
2126 return gen9_gmch_ctl << 25; /* 32 MB units */
2128 /* 4MB increments starting at 0xf0 for 4MB */
2129 return (gen9_gmch_ctl - 0xf0 + 1) << 22;
2132 static int ggtt_probe_common(struct drm_device *dev,
2135 struct drm_i915_private *dev_priv = dev->dev_private;
2136 phys_addr_t gtt_phys_addr;
2139 /* For Modern GENs the PTEs and register space are split in the BAR */
2140 gtt_phys_addr = pci_resource_start(dev->pdev, 0) +
2141 (pci_resource_len(dev->pdev, 0) / 2);
2143 dev_priv->gtt.gsm = ioremap_wc(gtt_phys_addr, gtt_size);
2144 if (!dev_priv->gtt.gsm) {
2145 DRM_ERROR("Failed to map the gtt page table\n");
2149 ret = setup_scratch_page(dev);
2151 DRM_ERROR("Scratch setup failed\n");
2152 /* iounmap will also get called at remove, but meh */
2153 iounmap(dev_priv->gtt.gsm);
2159 /* The GGTT and PPGTT need a private PPAT setup in order to handle cacheability
2160 * bits. When using advanced contexts each context stores its own PAT, but
2161 * writing this data shouldn't be harmful even in those cases. */
2162 static void bdw_setup_private_ppat(struct drm_i915_private *dev_priv)
2166 pat = GEN8_PPAT(0, GEN8_PPAT_WB | GEN8_PPAT_LLC) | /* for normal objects, no eLLC */
2167 GEN8_PPAT(1, GEN8_PPAT_WC | GEN8_PPAT_LLCELLC) | /* for something pointing to ptes? */
2168 GEN8_PPAT(2, GEN8_PPAT_WT | GEN8_PPAT_LLCELLC) | /* for scanout with eLLC */
2169 GEN8_PPAT(3, GEN8_PPAT_UC) | /* Uncached objects, mostly for scanout */
2170 GEN8_PPAT(4, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(0)) |
2171 GEN8_PPAT(5, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(1)) |
2172 GEN8_PPAT(6, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(2)) |
2173 GEN8_PPAT(7, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3));
2175 if (!USES_PPGTT(dev_priv->dev))
2176 /* Spec: "For GGTT, there is NO pat_sel[2:0] from the entry,
2177 * so RTL will always use the value corresponding to
2179 * So let's disable cache for GGTT to avoid screen corruptions.
2180 * MOCS still can be used though.
2181 * - System agent ggtt writes (i.e. cpu gtt mmaps) already work
2182 * before this patch, i.e. the same uncached + snooping access
2183 * like on gen6/7 seems to be in effect.
2184 * - So this just fixes blitter/render access. Again it looks
2185 * like it's not just uncached access, but uncached + snooping.
2186 * So we can still hold onto all our assumptions wrt cpu
2187 * clflushing on LLC machines.
2189 pat = GEN8_PPAT(0, GEN8_PPAT_UC);
2191 /* XXX: spec defines this as 2 distinct registers. It's unclear if a 64b
2192 * write would work. */
2193 I915_WRITE(GEN8_PRIVATE_PAT, pat);
2194 I915_WRITE(GEN8_PRIVATE_PAT + 4, pat >> 32);
2197 static void chv_setup_private_ppat(struct drm_i915_private *dev_priv)
2202 * Map WB on BDW to snooped on CHV.
2204 * Only the snoop bit has meaning for CHV, the rest is
2207 * The hardware will never snoop for certain types of accesses:
2208 * - CPU GTT (GMADR->GGTT->no snoop->memory)
2209 * - PPGTT page tables
2210 * - some other special cycles
2212 * As with BDW, we also need to consider the following for GT accesses:
2213 * "For GGTT, there is NO pat_sel[2:0] from the entry,
2214 * so RTL will always use the value corresponding to
2216 * Which means we must set the snoop bit in PAT entry 0
2217 * in order to keep the global status page working.
2219 pat = GEN8_PPAT(0, CHV_PPAT_SNOOP) |
2223 GEN8_PPAT(4, CHV_PPAT_SNOOP) |
2224 GEN8_PPAT(5, CHV_PPAT_SNOOP) |
2225 GEN8_PPAT(6, CHV_PPAT_SNOOP) |
2226 GEN8_PPAT(7, CHV_PPAT_SNOOP);
2228 I915_WRITE(GEN8_PRIVATE_PAT, pat);
2229 I915_WRITE(GEN8_PRIVATE_PAT + 4, pat >> 32);
2232 static int gen8_gmch_probe(struct drm_device *dev,
2235 phys_addr_t *mappable_base,
2236 unsigned long *mappable_end)
2238 struct drm_i915_private *dev_priv = dev->dev_private;
2239 unsigned int gtt_size;
2243 /* TODO: We're not aware of mappable constraints on gen8 yet */
2244 *mappable_base = pci_resource_start(dev->pdev, 2);
2245 *mappable_end = pci_resource_len(dev->pdev, 2);
2247 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(39)))
2248 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(39));
2250 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
2252 if (INTEL_INFO(dev)->gen >= 9) {
2253 *stolen = gen9_get_stolen_size(snb_gmch_ctl);
2254 gtt_size = gen8_get_total_gtt_size(snb_gmch_ctl);
2255 } else if (IS_CHERRYVIEW(dev)) {
2256 *stolen = chv_get_stolen_size(snb_gmch_ctl);
2257 gtt_size = chv_get_total_gtt_size(snb_gmch_ctl);
2259 *stolen = gen8_get_stolen_size(snb_gmch_ctl);
2260 gtt_size = gen8_get_total_gtt_size(snb_gmch_ctl);
2263 *gtt_total = (gtt_size / sizeof(gen8_pte_t)) << PAGE_SHIFT;
2265 if (IS_CHERRYVIEW(dev))
2266 chv_setup_private_ppat(dev_priv);
2268 bdw_setup_private_ppat(dev_priv);
2270 ret = ggtt_probe_common(dev, gtt_size);
2272 dev_priv->gtt.base.clear_range = gen8_ggtt_clear_range;
2273 dev_priv->gtt.base.insert_entries = gen8_ggtt_insert_entries;
2278 static int gen6_gmch_probe(struct drm_device *dev,
2281 phys_addr_t *mappable_base,
2282 unsigned long *mappable_end)
2284 struct drm_i915_private *dev_priv = dev->dev_private;
2285 unsigned int gtt_size;
2289 *mappable_base = pci_resource_start(dev->pdev, 2);
2290 *mappable_end = pci_resource_len(dev->pdev, 2);
2292 /* 64/512MB is the current min/max we actually know of, but this is just
2293 * a coarse sanity check.
2295 if ((*mappable_end < (64<<20) || (*mappable_end > (512<<20)))) {
2296 DRM_ERROR("Unknown GMADR size (%lx)\n",
2297 dev_priv->gtt.mappable_end);
2301 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(40)))
2302 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(40));
2303 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
2305 *stolen = gen6_get_stolen_size(snb_gmch_ctl);
2307 gtt_size = gen6_get_total_gtt_size(snb_gmch_ctl);
2308 *gtt_total = (gtt_size / sizeof(gen6_pte_t)) << PAGE_SHIFT;
2310 ret = ggtt_probe_common(dev, gtt_size);
2312 dev_priv->gtt.base.clear_range = gen6_ggtt_clear_range;
2313 dev_priv->gtt.base.insert_entries = gen6_ggtt_insert_entries;
2318 static void gen6_gmch_remove(struct i915_address_space *vm)
2321 struct i915_gtt *gtt = container_of(vm, struct i915_gtt, base);
2324 teardown_scratch_page(vm->dev);
2327 static int i915_gmch_probe(struct drm_device *dev,
2330 phys_addr_t *mappable_base,
2331 unsigned long *mappable_end)
2333 struct drm_i915_private *dev_priv = dev->dev_private;
2336 ret = intel_gmch_probe(dev_priv->bridge_dev, dev_priv->dev->pdev, NULL);
2338 DRM_ERROR("failed to set up gmch\n");
2342 intel_gtt_get(gtt_total, stolen, mappable_base, mappable_end);
2344 dev_priv->gtt.do_idle_maps = needs_idle_maps(dev_priv->dev);
2345 dev_priv->gtt.base.clear_range = i915_ggtt_clear_range;
2347 if (unlikely(dev_priv->gtt.do_idle_maps))
2348 DRM_INFO("applying Ironlake quirks for intel_iommu\n");
2353 static void i915_gmch_remove(struct i915_address_space *vm)
2355 intel_gmch_remove();
2358 int i915_gem_gtt_init(struct drm_device *dev)
2360 struct drm_i915_private *dev_priv = dev->dev_private;
2361 struct i915_gtt *gtt = &dev_priv->gtt;
2364 if (INTEL_INFO(dev)->gen <= 5) {
2365 gtt->gtt_probe = i915_gmch_probe;
2366 gtt->base.cleanup = i915_gmch_remove;
2367 } else if (INTEL_INFO(dev)->gen < 8) {
2368 gtt->gtt_probe = gen6_gmch_probe;
2369 gtt->base.cleanup = gen6_gmch_remove;
2370 if (IS_HASWELL(dev) && dev_priv->ellc_size)
2371 gtt->base.pte_encode = iris_pte_encode;
2372 else if (IS_HASWELL(dev))
2373 gtt->base.pte_encode = hsw_pte_encode;
2374 else if (IS_VALLEYVIEW(dev))
2375 gtt->base.pte_encode = byt_pte_encode;
2376 else if (INTEL_INFO(dev)->gen >= 7)
2377 gtt->base.pte_encode = ivb_pte_encode;
2379 gtt->base.pte_encode = snb_pte_encode;
2381 dev_priv->gtt.gtt_probe = gen8_gmch_probe;
2382 dev_priv->gtt.base.cleanup = gen6_gmch_remove;
2385 ret = gtt->gtt_probe(dev, >t->base.total, >t->stolen_size,
2386 >t->mappable_base, >t->mappable_end);
2390 gtt->base.dev = dev;
2392 /* GMADR is the PCI mmio aperture into the global GTT. */
2393 DRM_INFO("Memory usable by graphics device = %zdM\n",
2394 gtt->base.total >> 20);
2395 DRM_DEBUG_DRIVER("GMADR size = %ldM\n", gtt->mappable_end >> 20);
2396 DRM_DEBUG_DRIVER("GTT stolen size = %zdM\n", gtt->stolen_size >> 20);
2397 #ifdef CONFIG_INTEL_IOMMU
2398 if (intel_iommu_gfx_mapped)
2399 DRM_INFO("VT-d active for gfx access\n");
2402 * i915.enable_ppgtt is read-only, so do an early pass to validate the
2403 * user's requested state against the hardware/driver capabilities. We
2404 * do this now so that we can print out any log messages once rather
2405 * than every time we check intel_enable_ppgtt().
2407 i915.enable_ppgtt = sanitize_enable_ppgtt(dev, i915.enable_ppgtt);
2408 DRM_DEBUG_DRIVER("ppgtt mode: %i\n", i915.enable_ppgtt);
2413 static struct i915_vma *
2414 __i915_gem_vma_create(struct drm_i915_gem_object *obj,
2415 struct i915_address_space *vm,
2416 const struct i915_ggtt_view *ggtt_view)
2418 struct i915_vma *vma;
2420 if (WARN_ON(i915_is_ggtt(vm) != !!ggtt_view))
2421 return ERR_PTR(-EINVAL);
2422 vma = kzalloc(sizeof(*vma), GFP_KERNEL);
2424 return ERR_PTR(-ENOMEM);
2426 INIT_LIST_HEAD(&vma->vma_link);
2427 INIT_LIST_HEAD(&vma->mm_list);
2428 INIT_LIST_HEAD(&vma->exec_list);
2432 if (INTEL_INFO(vm->dev)->gen >= 6) {
2433 if (i915_is_ggtt(vm)) {
2434 vma->ggtt_view = *ggtt_view;
2436 vma->unbind_vma = ggtt_unbind_vma;
2437 vma->bind_vma = ggtt_bind_vma;
2439 vma->unbind_vma = ppgtt_unbind_vma;
2440 vma->bind_vma = ppgtt_bind_vma;
2443 BUG_ON(!i915_is_ggtt(vm));
2444 vma->ggtt_view = *ggtt_view;
2445 vma->unbind_vma = i915_ggtt_unbind_vma;
2446 vma->bind_vma = i915_ggtt_bind_vma;
2449 list_add_tail(&vma->vma_link, &obj->vma_list);
2450 if (!i915_is_ggtt(vm))
2451 i915_ppgtt_get(i915_vm_to_ppgtt(vm));
2457 i915_gem_obj_lookup_or_create_vma(struct drm_i915_gem_object *obj,
2458 struct i915_address_space *vm)
2460 struct i915_vma *vma;
2462 vma = i915_gem_obj_to_vma(obj, vm);
2464 vma = __i915_gem_vma_create(obj, vm,
2465 i915_is_ggtt(vm) ? &i915_ggtt_view_normal : NULL);
2471 i915_gem_obj_lookup_or_create_ggtt_vma(struct drm_i915_gem_object *obj,
2472 const struct i915_ggtt_view *view)
2474 struct i915_address_space *ggtt = i915_obj_to_ggtt(obj);
2475 struct i915_vma *vma;
2478 return ERR_PTR(-EINVAL);
2480 vma = i915_gem_obj_to_ggtt_view(obj, view);
2486 vma = __i915_gem_vma_create(obj, ggtt, view);
2494 int i915_get_ggtt_vma_pages(struct i915_vma *vma)
2496 if (vma->ggtt_view.pages)
2499 if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL)
2500 vma->ggtt_view.pages = vma->obj->pages;
2502 WARN_ONCE(1, "GGTT view %u not implemented!\n",
2503 vma->ggtt_view.type);
2505 if (!vma->ggtt_view.pages) {
2506 DRM_ERROR("Failed to get pages for GGTT view type %u!\n",
2507 vma->ggtt_view.type);
2515 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
2517 * @cache_level: mapping cache level
2518 * @flags: flags like global or local mapping
2520 * DMA addresses are taken from the scatter-gather table of this object (or of
2521 * this VMA in case of non-default GGTT views) and PTE entries set up.
2522 * Note that DMA addresses are also the only part of the SG table we care about.
2524 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
2527 if (i915_is_ggtt(vma->vm)) {
2528 int ret = i915_get_ggtt_vma_pages(vma);
2534 vma->bind_vma(vma, cache_level, flags);