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;
96 const struct i915_ggtt_view i915_ggtt_view_rotated = {
97 .type = I915_GGTT_VIEW_ROTATED
100 static void bdw_setup_private_ppat(struct drm_i915_private *dev_priv);
101 static void chv_setup_private_ppat(struct drm_i915_private *dev_priv);
103 static int sanitize_enable_ppgtt(struct drm_device *dev, int enable_ppgtt)
105 bool has_aliasing_ppgtt;
108 has_aliasing_ppgtt = INTEL_INFO(dev)->gen >= 6;
109 has_full_ppgtt = INTEL_INFO(dev)->gen >= 7;
111 if (intel_vgpu_active(dev))
112 has_full_ppgtt = false; /* emulation is too hard */
115 * We don't allow disabling PPGTT for gen9+ as it's a requirement for
116 * execlists, the sole mechanism available to submit work.
118 if (INTEL_INFO(dev)->gen < 9 &&
119 (enable_ppgtt == 0 || !has_aliasing_ppgtt))
122 if (enable_ppgtt == 1)
125 if (enable_ppgtt == 2 && has_full_ppgtt)
128 #ifdef CONFIG_INTEL_IOMMU
129 /* Disable ppgtt on SNB if VT-d is on. */
130 if (INTEL_INFO(dev)->gen == 6 && intel_iommu_gfx_mapped) {
131 DRM_INFO("Disabling PPGTT because VT-d is on\n");
136 /* Early VLV doesn't have this */
137 if (IS_VALLEYVIEW(dev) && !IS_CHERRYVIEW(dev) &&
138 dev->pdev->revision < 0xb) {
139 DRM_DEBUG_DRIVER("disabling PPGTT on pre-B3 step VLV\n");
143 if (INTEL_INFO(dev)->gen >= 8 && i915.enable_execlists)
146 return has_aliasing_ppgtt ? 1 : 0;
149 static void ppgtt_bind_vma(struct i915_vma *vma,
150 enum i915_cache_level cache_level,
152 static void ppgtt_unbind_vma(struct i915_vma *vma);
154 static inline gen8_pte_t gen8_pte_encode(dma_addr_t addr,
155 enum i915_cache_level level,
158 gen8_pte_t pte = valid ? _PAGE_PRESENT | _PAGE_RW : 0;
162 case I915_CACHE_NONE:
163 pte |= PPAT_UNCACHED_INDEX;
166 pte |= PPAT_DISPLAY_ELLC_INDEX;
169 pte |= PPAT_CACHED_INDEX;
176 static inline gen8_pde_t gen8_pde_encode(struct drm_device *dev,
178 enum i915_cache_level level)
180 gen8_pde_t pde = _PAGE_PRESENT | _PAGE_RW;
182 if (level != I915_CACHE_NONE)
183 pde |= PPAT_CACHED_PDE_INDEX;
185 pde |= PPAT_UNCACHED_INDEX;
189 static gen6_pte_t snb_pte_encode(dma_addr_t addr,
190 enum i915_cache_level level,
191 bool valid, u32 unused)
193 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
194 pte |= GEN6_PTE_ADDR_ENCODE(addr);
197 case I915_CACHE_L3_LLC:
199 pte |= GEN6_PTE_CACHE_LLC;
201 case I915_CACHE_NONE:
202 pte |= GEN6_PTE_UNCACHED;
211 static gen6_pte_t ivb_pte_encode(dma_addr_t addr,
212 enum i915_cache_level level,
213 bool valid, u32 unused)
215 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
216 pte |= GEN6_PTE_ADDR_ENCODE(addr);
219 case I915_CACHE_L3_LLC:
220 pte |= GEN7_PTE_CACHE_L3_LLC;
223 pte |= GEN6_PTE_CACHE_LLC;
225 case I915_CACHE_NONE:
226 pte |= GEN6_PTE_UNCACHED;
235 static gen6_pte_t byt_pte_encode(dma_addr_t addr,
236 enum i915_cache_level level,
237 bool valid, u32 flags)
239 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
240 pte |= GEN6_PTE_ADDR_ENCODE(addr);
242 if (!(flags & PTE_READ_ONLY))
243 pte |= BYT_PTE_WRITEABLE;
245 if (level != I915_CACHE_NONE)
246 pte |= BYT_PTE_SNOOPED_BY_CPU_CACHES;
251 static gen6_pte_t hsw_pte_encode(dma_addr_t addr,
252 enum i915_cache_level level,
253 bool valid, u32 unused)
255 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
256 pte |= HSW_PTE_ADDR_ENCODE(addr);
258 if (level != I915_CACHE_NONE)
259 pte |= HSW_WB_LLC_AGE3;
264 static gen6_pte_t iris_pte_encode(dma_addr_t addr,
265 enum i915_cache_level level,
266 bool valid, u32 unused)
268 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0;
269 pte |= HSW_PTE_ADDR_ENCODE(addr);
272 case I915_CACHE_NONE:
275 pte |= HSW_WT_ELLC_LLC_AGE3;
278 pte |= HSW_WB_ELLC_LLC_AGE3;
285 #define i915_dma_unmap_single(px, dev) \
286 __i915_dma_unmap_single((px)->daddr, dev)
288 static inline void __i915_dma_unmap_single(dma_addr_t daddr,
289 struct drm_device *dev)
291 struct device *device = &dev->pdev->dev;
293 dma_unmap_page(device, daddr, 4096, PCI_DMA_BIDIRECTIONAL);
297 * i915_dma_map_single() - Create a dma mapping for a page table/dir/etc.
298 * @px: Page table/dir/etc to get a DMA map for
301 * Page table allocations are unified across all gens. They always require a
302 * single 4k allocation, as well as a DMA mapping. If we keep the structs
303 * symmetric here, the simple macro covers us for every page table type.
305 * Return: 0 if success.
307 #define i915_dma_map_single(px, dev) \
308 i915_dma_map_page_single((px)->page, (dev), &(px)->daddr)
310 static inline int i915_dma_map_page_single(struct page *page,
311 struct drm_device *dev,
314 struct device *device = &dev->pdev->dev;
316 *daddr = dma_map_page(device, page, 0, 4096, PCI_DMA_BIDIRECTIONAL);
317 if (dma_mapping_error(device, *daddr))
323 static void unmap_and_free_pt(struct i915_page_table *pt,
324 struct drm_device *dev)
326 if (WARN_ON(!pt->page))
329 i915_dma_unmap_single(pt, dev);
330 __free_page(pt->page);
331 kfree(pt->used_ptes);
335 static void gen8_initialize_pt(struct i915_address_space *vm,
336 struct i915_page_table *pt)
338 gen8_pte_t *pt_vaddr, scratch_pte;
341 pt_vaddr = kmap_atomic(pt->page);
342 scratch_pte = gen8_pte_encode(vm->scratch.addr,
343 I915_CACHE_LLC, true);
345 for (i = 0; i < GEN8_PTES; i++)
346 pt_vaddr[i] = scratch_pte;
348 if (!HAS_LLC(vm->dev))
349 drm_clflush_virt_range(pt_vaddr, PAGE_SIZE);
350 kunmap_atomic(pt_vaddr);
353 static struct i915_page_table *alloc_pt_single(struct drm_device *dev)
355 struct i915_page_table *pt;
356 const size_t count = INTEL_INFO(dev)->gen >= 8 ?
357 GEN8_PTES : GEN6_PTES;
360 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
362 return ERR_PTR(-ENOMEM);
364 pt->used_ptes = kcalloc(BITS_TO_LONGS(count), sizeof(*pt->used_ptes),
370 pt->page = alloc_page(GFP_KERNEL);
374 ret = i915_dma_map_single(pt, dev);
381 __free_page(pt->page);
383 kfree(pt->used_ptes);
391 * alloc_pt_range() - Allocate a multiple page tables
392 * @pd: The page directory which will have at least @count entries
393 * available to point to the allocated page tables.
394 * @pde: First page directory entry for which we are allocating.
395 * @count: Number of pages to allocate.
398 * Allocates multiple page table pages and sets the appropriate entries in the
399 * page table structure within the page directory. Function cleans up after
400 * itself on any failures.
402 * Return: 0 if allocation succeeded.
404 static int alloc_pt_range(struct i915_page_directory *pd, uint16_t pde, size_t count,
405 struct drm_device *dev)
409 /* 512 is the max page tables per page_directory on any platform. */
410 if (WARN_ON(pde + count > I915_PDES))
413 for (i = pde; i < pde + count; i++) {
414 struct i915_page_table *pt = alloc_pt_single(dev);
420 WARN(pd->page_table[i],
421 "Leaking page directory entry %d (%p)\n",
422 i, pd->page_table[i]);
423 pd->page_table[i] = pt;
430 unmap_and_free_pt(pd->page_table[i], dev);
434 static void unmap_and_free_pd(struct i915_page_directory *pd)
437 __free_page(pd->page);
442 static struct i915_page_directory *alloc_pd_single(void)
444 struct i915_page_directory *pd;
446 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
448 return ERR_PTR(-ENOMEM);
450 pd->page = alloc_page(GFP_KERNEL);
453 return ERR_PTR(-ENOMEM);
459 /* Broadwell Page Directory Pointer Descriptors */
460 static int gen8_write_pdp(struct intel_engine_cs *ring, unsigned entry,
467 ret = intel_ring_begin(ring, 6);
471 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
472 intel_ring_emit(ring, GEN8_RING_PDP_UDW(ring, entry));
473 intel_ring_emit(ring, (u32)(val >> 32));
474 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
475 intel_ring_emit(ring, GEN8_RING_PDP_LDW(ring, entry));
476 intel_ring_emit(ring, (u32)(val));
477 intel_ring_advance(ring);
482 static int gen8_mm_switch(struct i915_hw_ppgtt *ppgtt,
483 struct intel_engine_cs *ring)
487 /* bit of a hack to find the actual last used pd */
488 int used_pd = ppgtt->num_pd_entries / I915_PDES;
490 for (i = used_pd - 1; i >= 0; i--) {
491 dma_addr_t addr = ppgtt->pdp.page_directory[i]->daddr;
492 ret = gen8_write_pdp(ring, i, addr);
500 static void gen8_ppgtt_clear_range(struct i915_address_space *vm,
505 struct i915_hw_ppgtt *ppgtt =
506 container_of(vm, struct i915_hw_ppgtt, base);
507 gen8_pte_t *pt_vaddr, scratch_pte;
508 unsigned pdpe = start >> GEN8_PDPE_SHIFT & GEN8_PDPE_MASK;
509 unsigned pde = start >> GEN8_PDE_SHIFT & GEN8_PDE_MASK;
510 unsigned pte = start >> GEN8_PTE_SHIFT & GEN8_PTE_MASK;
511 unsigned num_entries = length >> PAGE_SHIFT;
512 unsigned last_pte, i;
514 scratch_pte = gen8_pte_encode(ppgtt->base.scratch.addr,
515 I915_CACHE_LLC, use_scratch);
517 while (num_entries) {
518 struct i915_page_directory *pd;
519 struct i915_page_table *pt;
520 struct page *page_table;
522 if (WARN_ON(!ppgtt->pdp.page_directory[pdpe]))
525 pd = ppgtt->pdp.page_directory[pdpe];
527 if (WARN_ON(!pd->page_table[pde]))
530 pt = pd->page_table[pde];
532 if (WARN_ON(!pt->page))
535 page_table = pt->page;
537 last_pte = pte + num_entries;
538 if (last_pte > GEN8_PTES)
539 last_pte = GEN8_PTES;
541 pt_vaddr = kmap_atomic(page_table);
543 for (i = pte; i < last_pte; i++) {
544 pt_vaddr[i] = scratch_pte;
548 if (!HAS_LLC(ppgtt->base.dev))
549 drm_clflush_virt_range(pt_vaddr, PAGE_SIZE);
550 kunmap_atomic(pt_vaddr);
553 if (++pde == I915_PDES) {
560 static void gen8_ppgtt_insert_entries(struct i915_address_space *vm,
561 struct sg_table *pages,
563 enum i915_cache_level cache_level, u32 unused)
565 struct i915_hw_ppgtt *ppgtt =
566 container_of(vm, struct i915_hw_ppgtt, base);
567 gen8_pte_t *pt_vaddr;
568 unsigned pdpe = start >> GEN8_PDPE_SHIFT & GEN8_PDPE_MASK;
569 unsigned pde = start >> GEN8_PDE_SHIFT & GEN8_PDE_MASK;
570 unsigned pte = start >> GEN8_PTE_SHIFT & GEN8_PTE_MASK;
571 struct sg_page_iter sg_iter;
575 for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) {
576 if (WARN_ON(pdpe >= GEN8_LEGACY_PDPES))
579 if (pt_vaddr == NULL) {
580 struct i915_page_directory *pd = ppgtt->pdp.page_directory[pdpe];
581 struct i915_page_table *pt = pd->page_table[pde];
582 struct page *page_table = pt->page;
584 pt_vaddr = kmap_atomic(page_table);
588 gen8_pte_encode(sg_page_iter_dma_address(&sg_iter),
590 if (++pte == GEN8_PTES) {
591 if (!HAS_LLC(ppgtt->base.dev))
592 drm_clflush_virt_range(pt_vaddr, PAGE_SIZE);
593 kunmap_atomic(pt_vaddr);
595 if (++pde == I915_PDES) {
603 if (!HAS_LLC(ppgtt->base.dev))
604 drm_clflush_virt_range(pt_vaddr, PAGE_SIZE);
605 kunmap_atomic(pt_vaddr);
609 static void __gen8_do_map_pt(gen8_pde_t * const pde,
610 struct i915_page_table *pt,
611 struct drm_device *dev)
614 gen8_pde_encode(dev, pt->daddr, I915_CACHE_LLC);
618 static void gen8_initialize_pd(struct i915_address_space *vm,
619 struct i915_page_directory *pd)
621 struct i915_hw_ppgtt *ppgtt =
622 container_of(vm, struct i915_hw_ppgtt, base);
623 gen8_pde_t *page_directory;
624 struct i915_page_table *pt;
627 page_directory = kmap_atomic(pd->page);
628 pt = ppgtt->scratch_pt;
629 for (i = 0; i < I915_PDES; i++)
630 /* Map the PDE to the page table */
631 __gen8_do_map_pt(page_directory + i, pt, vm->dev);
633 if (!HAS_LLC(vm->dev))
634 drm_clflush_virt_range(page_directory, PAGE_SIZE);
636 kunmap_atomic(page_directory);
639 static void gen8_free_page_tables(struct i915_page_directory *pd, struct drm_device *dev)
646 for (i = 0; i < I915_PDES; i++) {
647 if (WARN_ON(!pd->page_table[i]))
650 unmap_and_free_pt(pd->page_table[i], dev);
651 pd->page_table[i] = NULL;
655 static void gen8_ppgtt_free(struct i915_hw_ppgtt *ppgtt)
659 for (i = 0; i < ppgtt->num_pd_pages; i++) {
660 if (WARN_ON(!ppgtt->pdp.page_directory[i]))
663 gen8_free_page_tables(ppgtt->pdp.page_directory[i], ppgtt->base.dev);
664 unmap_and_free_pd(ppgtt->pdp.page_directory[i]);
667 unmap_and_free_pt(ppgtt->scratch_pt, ppgtt->base.dev);
670 static void gen8_ppgtt_cleanup(struct i915_address_space *vm)
672 struct i915_hw_ppgtt *ppgtt =
673 container_of(vm, struct i915_hw_ppgtt, base);
675 gen8_ppgtt_free(ppgtt);
678 static int gen8_ppgtt_allocate_page_tables(struct i915_hw_ppgtt *ppgtt)
682 for (i = 0; i < ppgtt->num_pd_pages; i++) {
683 ret = alloc_pt_range(ppgtt->pdp.page_directory[i],
684 0, I915_PDES, ppgtt->base.dev);
693 gen8_free_page_tables(ppgtt->pdp.page_directory[i], ppgtt->base.dev);
698 static int gen8_ppgtt_alloc_page_directories(struct i915_page_directory_pointer *pdp,
702 struct i915_hw_ppgtt *ppgtt =
703 container_of(pdp, struct i915_hw_ppgtt, pdp);
704 struct i915_page_directory *unused;
708 /* FIXME: PPGTT container_of won't work for 64b */
709 WARN_ON((start + length) > 0x800000000ULL);
711 gen8_for_each_pdpe(unused, pdp, start, length, temp, pdpe) {
713 pdp->page_directory[pdpe] = alloc_pd_single();
714 if (IS_ERR(ppgtt->pdp.page_directory[pdpe]))
717 gen8_initialize_pd(&ppgtt->base,
718 ppgtt->pdp.page_directory[pdpe]);
719 ppgtt->num_pd_pages++;
722 /* XXX: Still alloc all page directories in systems with less than
723 * 4GB of memory. This won't be needed after a subsequent patch.
725 while (ppgtt->num_pd_pages < GEN8_LEGACY_PDPES) {
726 ppgtt->pdp.page_directory[ppgtt->num_pd_pages] = alloc_pd_single();
727 if (IS_ERR(ppgtt->pdp.page_directory[ppgtt->num_pd_pages]))
730 gen8_initialize_pd(&ppgtt->base,
731 ppgtt->pdp.page_directory[ppgtt->num_pd_pages]);
733 ppgtt->num_pd_pages++;
736 BUG_ON(ppgtt->num_pd_pages > GEN8_LEGACY_PDPES);
742 unmap_and_free_pd(ppgtt->pdp.page_directory[pdpe]);
743 ppgtt->num_pd_pages--;
746 WARN_ON(ppgtt->num_pd_pages);
751 static int gen8_ppgtt_alloc(struct i915_hw_ppgtt *ppgtt,
756 ret = gen8_ppgtt_alloc_page_directories(&ppgtt->pdp, ppgtt->base.start,
761 ret = gen8_ppgtt_allocate_page_tables(ppgtt);
765 ppgtt->num_pd_entries = max_pdp * I915_PDES;
770 gen8_ppgtt_free(ppgtt);
774 static int gen8_ppgtt_setup_page_directories(struct i915_hw_ppgtt *ppgtt,
780 pd_addr = pci_map_page(ppgtt->base.dev->pdev,
781 ppgtt->pdp.page_directory[pd]->page, 0,
782 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
784 ret = pci_dma_mapping_error(ppgtt->base.dev->pdev, pd_addr);
788 ppgtt->pdp.page_directory[pd]->daddr = pd_addr;
793 static int gen8_ppgtt_setup_page_tables(struct i915_hw_ppgtt *ppgtt,
798 struct i915_page_directory *pdir = ppgtt->pdp.page_directory[pd];
799 struct i915_page_table *ptab = pdir->page_table[pt];
800 struct page *p = ptab->page;
803 gen8_initialize_pt(&ppgtt->base, ptab);
805 pt_addr = pci_map_page(ppgtt->base.dev->pdev,
806 p, 0, PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
807 ret = pci_dma_mapping_error(ppgtt->base.dev->pdev, pt_addr);
811 ptab->daddr = pt_addr;
817 * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers
818 * with a net effect resembling a 2-level page table in normal x86 terms. Each
819 * PDP represents 1GB of memory 4 * 512 * 512 * 4096 = 4GB legacy 32b address
822 * FIXME: split allocation into smaller pieces. For now we only ever do this
823 * once, but with full PPGTT, the multiple contiguous allocations will be bad.
824 * TODO: Do something with the size parameter
826 static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt, uint64_t size)
828 const int max_pdp = DIV_ROUND_UP(size, 1 << 30);
829 const int min_pt_pages = I915_PDES * max_pdp;
833 DRM_INFO("Pages will be wasted unless GTT size (%llu) is divisible by 1GB\n", size);
835 ppgtt->base.start = 0;
836 /* This is the area that we advertise as usable for the caller */
837 ppgtt->base.total = max_pdp * I915_PDES * GEN8_PTES * PAGE_SIZE;
838 WARN_ON(ppgtt->base.total == 0);
840 ppgtt->scratch_pt = alloc_pt_single(ppgtt->base.dev);
841 if (IS_ERR(ppgtt->scratch_pt))
842 return PTR_ERR(ppgtt->scratch_pt);
844 gen8_initialize_pt(&ppgtt->base, ppgtt->scratch_pt);
846 /* 1. Do all our allocations for page directories and page tables.
847 * We allocate more than was asked so that we can point the unused parts
848 * to valid entries that point to scratch page. Dynamic page tables
849 * will fix this eventually.
851 ret = gen8_ppgtt_alloc(ppgtt, GEN8_LEGACY_PDPES);
856 * 2. Create DMA mappings for the page directories and page tables.
858 for (i = 0; i < GEN8_LEGACY_PDPES; i++) {
859 ret = gen8_ppgtt_setup_page_directories(ppgtt, i);
863 for (j = 0; j < I915_PDES; j++) {
864 ret = gen8_ppgtt_setup_page_tables(ppgtt, i, j);
871 * 3. Map all the page directory entries to point to the page tables
874 * For now, the PPGTT helper functions all require that the PDEs are
875 * plugged in correctly. So we do that now/here. For aliasing PPGTT, we
876 * will never need to touch the PDEs again.
878 for (i = 0; i < GEN8_LEGACY_PDPES; i++) {
879 struct i915_page_directory *pd = ppgtt->pdp.page_directory[i];
880 gen8_pde_t *pd_vaddr;
881 pd_vaddr = kmap_atomic(ppgtt->pdp.page_directory[i]->page);
882 for (j = 0; j < I915_PDES; j++) {
883 struct i915_page_table *pt = pd->page_table[j];
884 dma_addr_t addr = pt->daddr;
885 pd_vaddr[j] = gen8_pde_encode(ppgtt->base.dev, addr,
888 if (!HAS_LLC(ppgtt->base.dev))
889 drm_clflush_virt_range(pd_vaddr, PAGE_SIZE);
890 kunmap_atomic(pd_vaddr);
893 ppgtt->switch_mm = gen8_mm_switch;
894 ppgtt->base.clear_range = gen8_ppgtt_clear_range;
895 ppgtt->base.insert_entries = gen8_ppgtt_insert_entries;
896 ppgtt->base.cleanup = gen8_ppgtt_cleanup;
898 /* Set all ptes to a valid scratch page. Also above requested space */
899 ppgtt->base.clear_range(&ppgtt->base, 0,
900 ppgtt->num_pd_pages * GEN8_PTES * PAGE_SIZE,
903 DRM_DEBUG_DRIVER("Allocated %d pages for page directories (%d wasted)\n",
904 ppgtt->num_pd_pages, ppgtt->num_pd_pages - max_pdp);
905 DRM_DEBUG_DRIVER("Allocated %d pages for page tables (%lld wasted)\n",
906 ppgtt->num_pd_entries,
907 (ppgtt->num_pd_entries - min_pt_pages) + size % (1<<30));
911 gen8_ppgtt_free(ppgtt);
915 static void gen6_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
917 struct drm_i915_private *dev_priv = ppgtt->base.dev->dev_private;
918 struct i915_address_space *vm = &ppgtt->base;
919 gen6_pte_t __iomem *pd_addr;
920 gen6_pte_t scratch_pte;
924 scratch_pte = vm->pte_encode(vm->scratch.addr, I915_CACHE_LLC, true, 0);
926 pd_addr = (gen6_pte_t __iomem *)dev_priv->gtt.gsm +
927 ppgtt->pd.pd_offset / sizeof(gen6_pte_t);
929 seq_printf(m, " VM %p (pd_offset %x-%x):\n", vm,
931 ppgtt->pd.pd_offset + ppgtt->num_pd_entries);
932 for (pde = 0; pde < ppgtt->num_pd_entries; pde++) {
934 gen6_pte_t *pt_vaddr;
935 dma_addr_t pt_addr = ppgtt->pd.page_table[pde]->daddr;
936 pd_entry = readl(pd_addr + pde);
937 expected = (GEN6_PDE_ADDR_ENCODE(pt_addr) | GEN6_PDE_VALID);
939 if (pd_entry != expected)
940 seq_printf(m, "\tPDE #%d mismatch: Actual PDE: %x Expected PDE: %x\n",
944 seq_printf(m, "\tPDE: %x\n", pd_entry);
946 pt_vaddr = kmap_atomic(ppgtt->pd.page_table[pde]->page);
947 for (pte = 0; pte < GEN6_PTES; pte+=4) {
949 (pde * PAGE_SIZE * GEN6_PTES) +
953 for (i = 0; i < 4; i++)
954 if (pt_vaddr[pte + i] != scratch_pte)
959 seq_printf(m, "\t\t0x%lx [%03d,%04d]: =", va, pde, pte);
960 for (i = 0; i < 4; i++) {
961 if (pt_vaddr[pte + i] != scratch_pte)
962 seq_printf(m, " %08x", pt_vaddr[pte + i]);
964 seq_puts(m, " SCRATCH ");
968 kunmap_atomic(pt_vaddr);
972 /* Write pde (index) from the page directory @pd to the page table @pt */
973 static void gen6_write_pde(struct i915_page_directory *pd,
974 const int pde, struct i915_page_table *pt)
976 /* Caller needs to make sure the write completes if necessary */
977 struct i915_hw_ppgtt *ppgtt =
978 container_of(pd, struct i915_hw_ppgtt, pd);
981 pd_entry = GEN6_PDE_ADDR_ENCODE(pt->daddr);
982 pd_entry |= GEN6_PDE_VALID;
984 writel(pd_entry, ppgtt->pd_addr + pde);
987 /* Write all the page tables found in the ppgtt structure to incrementing page
989 static void gen6_write_page_range(struct drm_i915_private *dev_priv,
990 struct i915_page_directory *pd,
991 uint32_t start, uint32_t length)
993 struct i915_page_table *pt;
996 gen6_for_each_pde(pt, pd, start, length, temp, pde)
997 gen6_write_pde(pd, pde, pt);
999 /* Make sure write is complete before other code can use this page
1000 * table. Also require for WC mapped PTEs */
1001 readl(dev_priv->gtt.gsm);
1004 static uint32_t get_pd_offset(struct i915_hw_ppgtt *ppgtt)
1006 BUG_ON(ppgtt->pd.pd_offset & 0x3f);
1008 return (ppgtt->pd.pd_offset / 64) << 16;
1011 static int hsw_mm_switch(struct i915_hw_ppgtt *ppgtt,
1012 struct intel_engine_cs *ring)
1016 /* NB: TLBs must be flushed and invalidated before a switch */
1017 ret = ring->flush(ring, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1021 ret = intel_ring_begin(ring, 6);
1025 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1026 intel_ring_emit(ring, RING_PP_DIR_DCLV(ring));
1027 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1028 intel_ring_emit(ring, RING_PP_DIR_BASE(ring));
1029 intel_ring_emit(ring, get_pd_offset(ppgtt));
1030 intel_ring_emit(ring, MI_NOOP);
1031 intel_ring_advance(ring);
1036 static int vgpu_mm_switch(struct i915_hw_ppgtt *ppgtt,
1037 struct intel_engine_cs *ring)
1039 struct drm_i915_private *dev_priv = to_i915(ppgtt->base.dev);
1041 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
1042 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
1046 static int gen7_mm_switch(struct i915_hw_ppgtt *ppgtt,
1047 struct intel_engine_cs *ring)
1051 /* NB: TLBs must be flushed and invalidated before a switch */
1052 ret = ring->flush(ring, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1056 ret = intel_ring_begin(ring, 6);
1060 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2));
1061 intel_ring_emit(ring, RING_PP_DIR_DCLV(ring));
1062 intel_ring_emit(ring, PP_DIR_DCLV_2G);
1063 intel_ring_emit(ring, RING_PP_DIR_BASE(ring));
1064 intel_ring_emit(ring, get_pd_offset(ppgtt));
1065 intel_ring_emit(ring, MI_NOOP);
1066 intel_ring_advance(ring);
1068 /* XXX: RCS is the only one to auto invalidate the TLBs? */
1069 if (ring->id != RCS) {
1070 ret = ring->flush(ring, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
1078 static int gen6_mm_switch(struct i915_hw_ppgtt *ppgtt,
1079 struct intel_engine_cs *ring)
1081 struct drm_device *dev = ppgtt->base.dev;
1082 struct drm_i915_private *dev_priv = dev->dev_private;
1085 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
1086 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt));
1088 POSTING_READ(RING_PP_DIR_DCLV(ring));
1093 static void gen8_ppgtt_enable(struct drm_device *dev)
1095 struct drm_i915_private *dev_priv = dev->dev_private;
1096 struct intel_engine_cs *ring;
1099 for_each_ring(ring, dev_priv, j) {
1100 I915_WRITE(RING_MODE_GEN7(ring),
1101 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1105 static void gen7_ppgtt_enable(struct drm_device *dev)
1107 struct drm_i915_private *dev_priv = dev->dev_private;
1108 struct intel_engine_cs *ring;
1109 uint32_t ecochk, ecobits;
1112 ecobits = I915_READ(GAC_ECO_BITS);
1113 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B);
1115 ecochk = I915_READ(GAM_ECOCHK);
1116 if (IS_HASWELL(dev)) {
1117 ecochk |= ECOCHK_PPGTT_WB_HSW;
1119 ecochk |= ECOCHK_PPGTT_LLC_IVB;
1120 ecochk &= ~ECOCHK_PPGTT_GFDT_IVB;
1122 I915_WRITE(GAM_ECOCHK, ecochk);
1124 for_each_ring(ring, dev_priv, i) {
1125 /* GFX_MODE is per-ring on gen7+ */
1126 I915_WRITE(RING_MODE_GEN7(ring),
1127 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1131 static void gen6_ppgtt_enable(struct drm_device *dev)
1133 struct drm_i915_private *dev_priv = dev->dev_private;
1134 uint32_t ecochk, gab_ctl, ecobits;
1136 ecobits = I915_READ(GAC_ECO_BITS);
1137 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_SNB_BIT |
1138 ECOBITS_PPGTT_CACHE64B);
1140 gab_ctl = I915_READ(GAB_CTL);
1141 I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT);
1143 ecochk = I915_READ(GAM_ECOCHK);
1144 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT | ECOCHK_PPGTT_CACHE64B);
1146 I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1149 /* PPGTT support for Sandybdrige/Gen6 and later */
1150 static void gen6_ppgtt_clear_range(struct i915_address_space *vm,
1155 struct i915_hw_ppgtt *ppgtt =
1156 container_of(vm, struct i915_hw_ppgtt, base);
1157 gen6_pte_t *pt_vaddr, scratch_pte;
1158 unsigned first_entry = start >> PAGE_SHIFT;
1159 unsigned num_entries = length >> PAGE_SHIFT;
1160 unsigned act_pt = first_entry / GEN6_PTES;
1161 unsigned first_pte = first_entry % GEN6_PTES;
1162 unsigned last_pte, i;
1164 scratch_pte = vm->pte_encode(vm->scratch.addr, I915_CACHE_LLC, true, 0);
1166 while (num_entries) {
1167 last_pte = first_pte + num_entries;
1168 if (last_pte > GEN6_PTES)
1169 last_pte = GEN6_PTES;
1171 pt_vaddr = kmap_atomic(ppgtt->pd.page_table[act_pt]->page);
1173 for (i = first_pte; i < last_pte; i++)
1174 pt_vaddr[i] = scratch_pte;
1176 kunmap_atomic(pt_vaddr);
1178 num_entries -= last_pte - first_pte;
1184 static void gen6_ppgtt_insert_entries(struct i915_address_space *vm,
1185 struct sg_table *pages,
1187 enum i915_cache_level cache_level, u32 flags)
1189 struct i915_hw_ppgtt *ppgtt =
1190 container_of(vm, struct i915_hw_ppgtt, base);
1191 gen6_pte_t *pt_vaddr;
1192 unsigned first_entry = start >> PAGE_SHIFT;
1193 unsigned act_pt = first_entry / GEN6_PTES;
1194 unsigned act_pte = first_entry % GEN6_PTES;
1195 struct sg_page_iter sg_iter;
1198 for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) {
1199 if (pt_vaddr == NULL)
1200 pt_vaddr = kmap_atomic(ppgtt->pd.page_table[act_pt]->page);
1203 vm->pte_encode(sg_page_iter_dma_address(&sg_iter),
1204 cache_level, true, flags);
1206 if (++act_pte == GEN6_PTES) {
1207 kunmap_atomic(pt_vaddr);
1214 kunmap_atomic(pt_vaddr);
1217 /* PDE TLBs are a pain invalidate pre GEN8. It requires a context reload. If we
1218 * are switching between contexts with the same LRCA, we also must do a force
1221 static inline void mark_tlbs_dirty(struct i915_hw_ppgtt *ppgtt)
1223 /* If current vm != vm, */
1224 ppgtt->pd_dirty_rings = INTEL_INFO(ppgtt->base.dev)->ring_mask;
1227 static void gen6_initialize_pt(struct i915_address_space *vm,
1228 struct i915_page_table *pt)
1230 gen6_pte_t *pt_vaddr, scratch_pte;
1233 WARN_ON(vm->scratch.addr == 0);
1235 scratch_pte = vm->pte_encode(vm->scratch.addr,
1236 I915_CACHE_LLC, true, 0);
1238 pt_vaddr = kmap_atomic(pt->page);
1240 for (i = 0; i < GEN6_PTES; i++)
1241 pt_vaddr[i] = scratch_pte;
1243 kunmap_atomic(pt_vaddr);
1246 static int gen6_alloc_va_range(struct i915_address_space *vm,
1247 uint64_t start, uint64_t length)
1249 DECLARE_BITMAP(new_page_tables, I915_PDES);
1250 struct drm_device *dev = vm->dev;
1251 struct drm_i915_private *dev_priv = dev->dev_private;
1252 struct i915_hw_ppgtt *ppgtt =
1253 container_of(vm, struct i915_hw_ppgtt, base);
1254 struct i915_page_table *pt;
1255 const uint32_t start_save = start, length_save = length;
1259 WARN_ON(upper_32_bits(start));
1261 bitmap_zero(new_page_tables, I915_PDES);
1263 /* The allocation is done in two stages so that we can bail out with
1264 * minimal amount of pain. The first stage finds new page tables that
1265 * need allocation. The second stage marks use ptes within the page
1268 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1269 if (pt != ppgtt->scratch_pt) {
1270 WARN_ON(bitmap_empty(pt->used_ptes, GEN6_PTES));
1274 /* We've already allocated a page table */
1275 WARN_ON(!bitmap_empty(pt->used_ptes, GEN6_PTES));
1277 pt = alloc_pt_single(dev);
1283 gen6_initialize_pt(vm, pt);
1285 ppgtt->pd.page_table[pde] = pt;
1286 set_bit(pde, new_page_tables);
1287 trace_i915_page_table_entry_alloc(vm, pde, start, GEN6_PDE_SHIFT);
1291 length = length_save;
1293 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
1294 DECLARE_BITMAP(tmp_bitmap, GEN6_PTES);
1296 bitmap_zero(tmp_bitmap, GEN6_PTES);
1297 bitmap_set(tmp_bitmap, gen6_pte_index(start),
1298 gen6_pte_count(start, length));
1300 if (test_and_clear_bit(pde, new_page_tables))
1301 gen6_write_pde(&ppgtt->pd, pde, pt);
1303 trace_i915_page_table_entry_map(vm, pde, pt,
1304 gen6_pte_index(start),
1305 gen6_pte_count(start, length),
1307 bitmap_or(pt->used_ptes, tmp_bitmap, pt->used_ptes,
1311 WARN_ON(!bitmap_empty(new_page_tables, I915_PDES));
1313 /* Make sure write is complete before other code can use this page
1314 * table. Also require for WC mapped PTEs */
1315 readl(dev_priv->gtt.gsm);
1317 mark_tlbs_dirty(ppgtt);
1321 for_each_set_bit(pde, new_page_tables, I915_PDES) {
1322 struct i915_page_table *pt = ppgtt->pd.page_table[pde];
1324 ppgtt->pd.page_table[pde] = ppgtt->scratch_pt;
1325 unmap_and_free_pt(pt, vm->dev);
1328 mark_tlbs_dirty(ppgtt);
1332 static void gen6_ppgtt_free(struct i915_hw_ppgtt *ppgtt)
1336 for (i = 0; i < ppgtt->num_pd_entries; i++) {
1337 struct i915_page_table *pt = ppgtt->pd.page_table[i];
1339 if (pt != ppgtt->scratch_pt)
1340 unmap_and_free_pt(ppgtt->pd.page_table[i], ppgtt->base.dev);
1343 unmap_and_free_pt(ppgtt->scratch_pt, ppgtt->base.dev);
1344 unmap_and_free_pd(&ppgtt->pd);
1347 static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
1349 struct i915_hw_ppgtt *ppgtt =
1350 container_of(vm, struct i915_hw_ppgtt, base);
1352 drm_mm_remove_node(&ppgtt->node);
1354 gen6_ppgtt_free(ppgtt);
1357 static int gen6_ppgtt_allocate_page_directories(struct i915_hw_ppgtt *ppgtt)
1359 struct drm_device *dev = ppgtt->base.dev;
1360 struct drm_i915_private *dev_priv = dev->dev_private;
1361 bool retried = false;
1364 /* PPGTT PDEs reside in the GGTT and consists of 512 entries. The
1365 * allocator works in address space sizes, so it's multiplied by page
1366 * size. We allocate at the top of the GTT to avoid fragmentation.
1368 BUG_ON(!drm_mm_initialized(&dev_priv->gtt.base.mm));
1369 ppgtt->scratch_pt = alloc_pt_single(ppgtt->base.dev);
1370 if (IS_ERR(ppgtt->scratch_pt))
1371 return PTR_ERR(ppgtt->scratch_pt);
1373 gen6_initialize_pt(&ppgtt->base, ppgtt->scratch_pt);
1376 ret = drm_mm_insert_node_in_range_generic(&dev_priv->gtt.base.mm,
1377 &ppgtt->node, GEN6_PD_SIZE,
1379 0, dev_priv->gtt.base.total,
1381 if (ret == -ENOSPC && !retried) {
1382 ret = i915_gem_evict_something(dev, &dev_priv->gtt.base,
1383 GEN6_PD_SIZE, GEN6_PD_ALIGN,
1385 0, dev_priv->gtt.base.total,
1398 if (ppgtt->node.start < dev_priv->gtt.mappable_end)
1399 DRM_DEBUG("Forced to use aperture for PDEs\n");
1401 ppgtt->num_pd_entries = I915_PDES;
1405 unmap_and_free_pt(ppgtt->scratch_pt, ppgtt->base.dev);
1409 static int gen6_ppgtt_alloc(struct i915_hw_ppgtt *ppgtt)
1411 return gen6_ppgtt_allocate_page_directories(ppgtt);
1414 static void gen6_scratch_va_range(struct i915_hw_ppgtt *ppgtt,
1415 uint64_t start, uint64_t length)
1417 struct i915_page_table *unused;
1420 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde)
1421 ppgtt->pd.page_table[pde] = ppgtt->scratch_pt;
1424 static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt, bool aliasing)
1426 struct drm_device *dev = ppgtt->base.dev;
1427 struct drm_i915_private *dev_priv = dev->dev_private;
1430 ppgtt->base.pte_encode = dev_priv->gtt.base.pte_encode;
1432 ppgtt->switch_mm = gen6_mm_switch;
1433 } else if (IS_HASWELL(dev)) {
1434 ppgtt->switch_mm = hsw_mm_switch;
1435 } else if (IS_GEN7(dev)) {
1436 ppgtt->switch_mm = gen7_mm_switch;
1440 if (intel_vgpu_active(dev))
1441 ppgtt->switch_mm = vgpu_mm_switch;
1443 ret = gen6_ppgtt_alloc(ppgtt);
1448 /* preallocate all pts */
1449 ret = alloc_pt_range(&ppgtt->pd, 0, ppgtt->num_pd_entries,
1453 gen6_ppgtt_cleanup(&ppgtt->base);
1458 ppgtt->base.allocate_va_range = gen6_alloc_va_range;
1459 ppgtt->base.clear_range = gen6_ppgtt_clear_range;
1460 ppgtt->base.insert_entries = gen6_ppgtt_insert_entries;
1461 ppgtt->base.cleanup = gen6_ppgtt_cleanup;
1462 ppgtt->base.start = 0;
1463 ppgtt->base.total = ppgtt->num_pd_entries * GEN6_PTES * PAGE_SIZE;
1464 ppgtt->debug_dump = gen6_dump_ppgtt;
1466 ppgtt->pd.pd_offset =
1467 ppgtt->node.start / PAGE_SIZE * sizeof(gen6_pte_t);
1469 ppgtt->pd_addr = (gen6_pte_t __iomem *)dev_priv->gtt.gsm +
1470 ppgtt->pd.pd_offset / sizeof(gen6_pte_t);
1473 ppgtt->base.clear_range(&ppgtt->base, 0, ppgtt->base.total, true);
1475 gen6_scratch_va_range(ppgtt, 0, ppgtt->base.total);
1477 gen6_write_page_range(dev_priv, &ppgtt->pd, 0, ppgtt->base.total);
1479 DRM_DEBUG_DRIVER("Allocated pde space (%lldM) at GTT entry: %llx\n",
1480 ppgtt->node.size >> 20,
1481 ppgtt->node.start / PAGE_SIZE);
1483 DRM_DEBUG("Adding PPGTT at offset %x\n",
1484 ppgtt->pd.pd_offset << 10);
1489 static int __hw_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt,
1492 struct drm_i915_private *dev_priv = dev->dev_private;
1494 ppgtt->base.dev = dev;
1495 ppgtt->base.scratch = dev_priv->gtt.base.scratch;
1497 if (INTEL_INFO(dev)->gen < 8)
1498 return gen6_ppgtt_init(ppgtt, aliasing);
1500 return gen8_ppgtt_init(ppgtt, dev_priv->gtt.base.total);
1502 int i915_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
1504 struct drm_i915_private *dev_priv = dev->dev_private;
1507 ret = __hw_ppgtt_init(dev, ppgtt, false);
1509 kref_init(&ppgtt->ref);
1510 drm_mm_init(&ppgtt->base.mm, ppgtt->base.start,
1512 i915_init_vm(dev_priv, &ppgtt->base);
1518 int i915_ppgtt_init_hw(struct drm_device *dev)
1520 struct drm_i915_private *dev_priv = dev->dev_private;
1521 struct intel_engine_cs *ring;
1522 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
1525 /* In the case of execlists, PPGTT is enabled by the context descriptor
1526 * and the PDPs are contained within the context itself. We don't
1527 * need to do anything here. */
1528 if (i915.enable_execlists)
1531 if (!USES_PPGTT(dev))
1535 gen6_ppgtt_enable(dev);
1536 else if (IS_GEN7(dev))
1537 gen7_ppgtt_enable(dev);
1538 else if (INTEL_INFO(dev)->gen >= 8)
1539 gen8_ppgtt_enable(dev);
1541 MISSING_CASE(INTEL_INFO(dev)->gen);
1544 for_each_ring(ring, dev_priv, i) {
1545 ret = ppgtt->switch_mm(ppgtt, ring);
1553 struct i915_hw_ppgtt *
1554 i915_ppgtt_create(struct drm_device *dev, struct drm_i915_file_private *fpriv)
1556 struct i915_hw_ppgtt *ppgtt;
1559 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
1561 return ERR_PTR(-ENOMEM);
1563 ret = i915_ppgtt_init(dev, ppgtt);
1566 return ERR_PTR(ret);
1569 ppgtt->file_priv = fpriv;
1571 trace_i915_ppgtt_create(&ppgtt->base);
1576 void i915_ppgtt_release(struct kref *kref)
1578 struct i915_hw_ppgtt *ppgtt =
1579 container_of(kref, struct i915_hw_ppgtt, ref);
1581 trace_i915_ppgtt_release(&ppgtt->base);
1583 /* vmas should already be unbound */
1584 WARN_ON(!list_empty(&ppgtt->base.active_list));
1585 WARN_ON(!list_empty(&ppgtt->base.inactive_list));
1587 list_del(&ppgtt->base.global_link);
1588 drm_mm_takedown(&ppgtt->base.mm);
1590 ppgtt->base.cleanup(&ppgtt->base);
1595 ppgtt_bind_vma(struct i915_vma *vma,
1596 enum i915_cache_level cache_level,
1599 /* Currently applicable only to VLV */
1600 if (vma->obj->gt_ro)
1601 flags |= PTE_READ_ONLY;
1603 vma->vm->insert_entries(vma->vm, vma->obj->pages, vma->node.start,
1604 cache_level, flags);
1607 static void ppgtt_unbind_vma(struct i915_vma *vma)
1609 vma->vm->clear_range(vma->vm,
1611 vma->obj->base.size,
1615 extern int intel_iommu_gfx_mapped;
1616 /* Certain Gen5 chipsets require require idling the GPU before
1617 * unmapping anything from the GTT when VT-d is enabled.
1619 static inline bool needs_idle_maps(struct drm_device *dev)
1621 #ifdef CONFIG_INTEL_IOMMU
1622 /* Query intel_iommu to see if we need the workaround. Presumably that
1625 if (IS_GEN5(dev) && IS_MOBILE(dev) && intel_iommu_gfx_mapped)
1631 static bool do_idling(struct drm_i915_private *dev_priv)
1633 bool ret = dev_priv->mm.interruptible;
1635 if (unlikely(dev_priv->gtt.do_idle_maps)) {
1636 dev_priv->mm.interruptible = false;
1637 if (i915_gpu_idle(dev_priv->dev)) {
1638 DRM_ERROR("Couldn't idle GPU\n");
1639 /* Wait a bit, in hopes it avoids the hang */
1647 static void undo_idling(struct drm_i915_private *dev_priv, bool interruptible)
1649 if (unlikely(dev_priv->gtt.do_idle_maps))
1650 dev_priv->mm.interruptible = interruptible;
1653 void i915_check_and_clear_faults(struct drm_device *dev)
1655 struct drm_i915_private *dev_priv = dev->dev_private;
1656 struct intel_engine_cs *ring;
1659 if (INTEL_INFO(dev)->gen < 6)
1662 for_each_ring(ring, dev_priv, i) {
1664 fault_reg = I915_READ(RING_FAULT_REG(ring));
1665 if (fault_reg & RING_FAULT_VALID) {
1666 DRM_DEBUG_DRIVER("Unexpected fault\n"
1668 "\tAddress space: %s\n"
1671 fault_reg & PAGE_MASK,
1672 fault_reg & RING_FAULT_GTTSEL_MASK ? "GGTT" : "PPGTT",
1673 RING_FAULT_SRCID(fault_reg),
1674 RING_FAULT_FAULT_TYPE(fault_reg));
1675 I915_WRITE(RING_FAULT_REG(ring),
1676 fault_reg & ~RING_FAULT_VALID);
1679 POSTING_READ(RING_FAULT_REG(&dev_priv->ring[RCS]));
1682 static void i915_ggtt_flush(struct drm_i915_private *dev_priv)
1684 if (INTEL_INFO(dev_priv->dev)->gen < 6) {
1685 intel_gtt_chipset_flush();
1687 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
1688 POSTING_READ(GFX_FLSH_CNTL_GEN6);
1692 void i915_gem_suspend_gtt_mappings(struct drm_device *dev)
1694 struct drm_i915_private *dev_priv = dev->dev_private;
1696 /* Don't bother messing with faults pre GEN6 as we have little
1697 * documentation supporting that it's a good idea.
1699 if (INTEL_INFO(dev)->gen < 6)
1702 i915_check_and_clear_faults(dev);
1704 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
1705 dev_priv->gtt.base.start,
1706 dev_priv->gtt.base.total,
1709 i915_ggtt_flush(dev_priv);
1712 void i915_gem_restore_gtt_mappings(struct drm_device *dev)
1714 struct drm_i915_private *dev_priv = dev->dev_private;
1715 struct drm_i915_gem_object *obj;
1716 struct i915_address_space *vm;
1718 i915_check_and_clear_faults(dev);
1720 /* First fill our portion of the GTT with scratch pages */
1721 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
1722 dev_priv->gtt.base.start,
1723 dev_priv->gtt.base.total,
1726 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
1727 struct i915_vma *vma = i915_gem_obj_to_vma(obj,
1728 &dev_priv->gtt.base);
1732 i915_gem_clflush_object(obj, obj->pin_display);
1733 /* The bind_vma code tries to be smart about tracking mappings.
1734 * Unfortunately above, we've just wiped out the mappings
1735 * without telling our object about it. So we need to fake it.
1737 * Bind is not expected to fail since this is only called on
1738 * resume and assumption is all requirements exist already.
1740 vma->bound &= ~GLOBAL_BIND;
1741 WARN_ON(i915_vma_bind(vma, obj->cache_level, GLOBAL_BIND));
1745 if (INTEL_INFO(dev)->gen >= 8) {
1746 if (IS_CHERRYVIEW(dev))
1747 chv_setup_private_ppat(dev_priv);
1749 bdw_setup_private_ppat(dev_priv);
1754 if (USES_PPGTT(dev)) {
1755 list_for_each_entry(vm, &dev_priv->vm_list, global_link) {
1756 /* TODO: Perhaps it shouldn't be gen6 specific */
1758 struct i915_hw_ppgtt *ppgtt =
1759 container_of(vm, struct i915_hw_ppgtt,
1762 if (i915_is_ggtt(vm))
1763 ppgtt = dev_priv->mm.aliasing_ppgtt;
1765 gen6_write_page_range(dev_priv, &ppgtt->pd,
1766 0, ppgtt->base.total);
1770 i915_ggtt_flush(dev_priv);
1773 int i915_gem_gtt_prepare_object(struct drm_i915_gem_object *obj)
1775 if (obj->has_dma_mapping)
1778 if (!dma_map_sg(&obj->base.dev->pdev->dev,
1779 obj->pages->sgl, obj->pages->nents,
1780 PCI_DMA_BIDIRECTIONAL))
1786 static inline void gen8_set_pte(void __iomem *addr, gen8_pte_t pte)
1791 iowrite32((u32)pte, addr);
1792 iowrite32(pte >> 32, addr + 4);
1796 static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
1797 struct sg_table *st,
1799 enum i915_cache_level level, u32 unused)
1801 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1802 unsigned first_entry = start >> PAGE_SHIFT;
1803 gen8_pte_t __iomem *gtt_entries =
1804 (gen8_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
1806 struct sg_page_iter sg_iter;
1807 dma_addr_t addr = 0; /* shut up gcc */
1809 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
1810 addr = sg_dma_address(sg_iter.sg) +
1811 (sg_iter.sg_pgoffset << PAGE_SHIFT);
1812 gen8_set_pte(>t_entries[i],
1813 gen8_pte_encode(addr, level, true));
1818 * XXX: This serves as a posting read to make sure that the PTE has
1819 * actually been updated. There is some concern that even though
1820 * registers and PTEs are within the same BAR that they are potentially
1821 * of NUMA access patterns. Therefore, even with the way we assume
1822 * hardware should work, we must keep this posting read for paranoia.
1825 WARN_ON(readq(>t_entries[i-1])
1826 != gen8_pte_encode(addr, level, true));
1828 /* This next bit makes the above posting read even more important. We
1829 * want to flush the TLBs only after we're certain all the PTE updates
1832 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
1833 POSTING_READ(GFX_FLSH_CNTL_GEN6);
1837 * Binds an object into the global gtt with the specified cache level. The object
1838 * will be accessible to the GPU via commands whose operands reference offsets
1839 * within the global GTT as well as accessible by the GPU through the GMADR
1840 * mapped BAR (dev_priv->mm.gtt->gtt).
1842 static void gen6_ggtt_insert_entries(struct i915_address_space *vm,
1843 struct sg_table *st,
1845 enum i915_cache_level level, u32 flags)
1847 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1848 unsigned first_entry = start >> PAGE_SHIFT;
1849 gen6_pte_t __iomem *gtt_entries =
1850 (gen6_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
1852 struct sg_page_iter sg_iter;
1853 dma_addr_t addr = 0;
1855 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
1856 addr = sg_page_iter_dma_address(&sg_iter);
1857 iowrite32(vm->pte_encode(addr, level, true, flags), >t_entries[i]);
1861 /* XXX: This serves as a posting read to make sure that the PTE has
1862 * actually been updated. There is some concern that even though
1863 * registers and PTEs are within the same BAR that they are potentially
1864 * of NUMA access patterns. Therefore, even with the way we assume
1865 * hardware should work, we must keep this posting read for paranoia.
1868 unsigned long gtt = readl(>t_entries[i-1]);
1869 WARN_ON(gtt != vm->pte_encode(addr, level, true, flags));
1872 /* This next bit makes the above posting read even more important. We
1873 * want to flush the TLBs only after we're certain all the PTE updates
1876 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
1877 POSTING_READ(GFX_FLSH_CNTL_GEN6);
1880 static void gen8_ggtt_clear_range(struct i915_address_space *vm,
1885 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1886 unsigned first_entry = start >> PAGE_SHIFT;
1887 unsigned num_entries = length >> PAGE_SHIFT;
1888 gen8_pte_t scratch_pte, __iomem *gtt_base =
1889 (gen8_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
1890 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
1893 if (WARN(num_entries > max_entries,
1894 "First entry = %d; Num entries = %d (max=%d)\n",
1895 first_entry, num_entries, max_entries))
1896 num_entries = max_entries;
1898 scratch_pte = gen8_pte_encode(vm->scratch.addr,
1901 for (i = 0; i < num_entries; i++)
1902 gen8_set_pte(>t_base[i], scratch_pte);
1906 static void gen6_ggtt_clear_range(struct i915_address_space *vm,
1911 struct drm_i915_private *dev_priv = vm->dev->dev_private;
1912 unsigned first_entry = start >> PAGE_SHIFT;
1913 unsigned num_entries = length >> PAGE_SHIFT;
1914 gen6_pte_t scratch_pte, __iomem *gtt_base =
1915 (gen6_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
1916 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
1919 if (WARN(num_entries > max_entries,
1920 "First entry = %d; Num entries = %d (max=%d)\n",
1921 first_entry, num_entries, max_entries))
1922 num_entries = max_entries;
1924 scratch_pte = vm->pte_encode(vm->scratch.addr, I915_CACHE_LLC, use_scratch, 0);
1926 for (i = 0; i < num_entries; i++)
1927 iowrite32(scratch_pte, >t_base[i]);
1932 static void i915_ggtt_bind_vma(struct i915_vma *vma,
1933 enum i915_cache_level cache_level,
1936 const unsigned long entry = vma->node.start >> PAGE_SHIFT;
1937 unsigned int flags = (cache_level == I915_CACHE_NONE) ?
1938 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
1940 BUG_ON(!i915_is_ggtt(vma->vm));
1941 intel_gtt_insert_sg_entries(vma->ggtt_view.pages, entry, flags);
1942 vma->bound = GLOBAL_BIND;
1945 static void i915_ggtt_clear_range(struct i915_address_space *vm,
1950 unsigned first_entry = start >> PAGE_SHIFT;
1951 unsigned num_entries = length >> PAGE_SHIFT;
1952 intel_gtt_clear_range(first_entry, num_entries);
1955 static void i915_ggtt_unbind_vma(struct i915_vma *vma)
1957 const unsigned int first = vma->node.start >> PAGE_SHIFT;
1958 const unsigned int size = vma->obj->base.size >> PAGE_SHIFT;
1960 BUG_ON(!i915_is_ggtt(vma->vm));
1962 intel_gtt_clear_range(first, size);
1965 static void ggtt_bind_vma(struct i915_vma *vma,
1966 enum i915_cache_level cache_level,
1969 struct drm_device *dev = vma->vm->dev;
1970 struct drm_i915_private *dev_priv = dev->dev_private;
1971 struct drm_i915_gem_object *obj = vma->obj;
1972 struct sg_table *pages = obj->pages;
1974 /* Currently applicable only to VLV */
1976 flags |= PTE_READ_ONLY;
1978 if (i915_is_ggtt(vma->vm))
1979 pages = vma->ggtt_view.pages;
1981 /* If there is no aliasing PPGTT, or the caller needs a global mapping,
1982 * or we have a global mapping already but the cacheability flags have
1983 * changed, set the global PTEs.
1985 * If there is an aliasing PPGTT it is anecdotally faster, so use that
1986 * instead if none of the above hold true.
1988 * NB: A global mapping should only be needed for special regions like
1989 * "gtt mappable", SNB errata, or if specified via special execbuf
1990 * flags. At all other times, the GPU will use the aliasing PPGTT.
1992 if (!dev_priv->mm.aliasing_ppgtt || flags & GLOBAL_BIND) {
1993 if (!(vma->bound & GLOBAL_BIND) ||
1994 (cache_level != obj->cache_level)) {
1995 vma->vm->insert_entries(vma->vm, pages,
1997 cache_level, flags);
1998 vma->bound |= GLOBAL_BIND;
2002 if (dev_priv->mm.aliasing_ppgtt &&
2003 (!(vma->bound & LOCAL_BIND) ||
2004 (cache_level != obj->cache_level))) {
2005 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
2006 appgtt->base.insert_entries(&appgtt->base, pages,
2008 cache_level, flags);
2009 vma->bound |= LOCAL_BIND;
2013 static void ggtt_unbind_vma(struct i915_vma *vma)
2015 struct drm_device *dev = vma->vm->dev;
2016 struct drm_i915_private *dev_priv = dev->dev_private;
2017 struct drm_i915_gem_object *obj = vma->obj;
2019 if (vma->bound & GLOBAL_BIND) {
2020 vma->vm->clear_range(vma->vm,
2024 vma->bound &= ~GLOBAL_BIND;
2027 if (vma->bound & LOCAL_BIND) {
2028 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt;
2029 appgtt->base.clear_range(&appgtt->base,
2033 vma->bound &= ~LOCAL_BIND;
2037 void i915_gem_gtt_finish_object(struct drm_i915_gem_object *obj)
2039 struct drm_device *dev = obj->base.dev;
2040 struct drm_i915_private *dev_priv = dev->dev_private;
2043 interruptible = do_idling(dev_priv);
2045 if (!obj->has_dma_mapping)
2046 dma_unmap_sg(&dev->pdev->dev,
2047 obj->pages->sgl, obj->pages->nents,
2048 PCI_DMA_BIDIRECTIONAL);
2050 undo_idling(dev_priv, interruptible);
2053 static void i915_gtt_color_adjust(struct drm_mm_node *node,
2054 unsigned long color,
2058 if (node->color != color)
2061 if (!list_empty(&node->node_list)) {
2062 node = list_entry(node->node_list.next,
2065 if (node->allocated && node->color != color)
2070 static int i915_gem_setup_global_gtt(struct drm_device *dev,
2071 unsigned long start,
2072 unsigned long mappable_end,
2075 /* Let GEM Manage all of the aperture.
2077 * However, leave one page at the end still bound to the scratch page.
2078 * There are a number of places where the hardware apparently prefetches
2079 * past the end of the object, and we've seen multiple hangs with the
2080 * GPU head pointer stuck in a batchbuffer bound at the last page of the
2081 * aperture. One page should be enough to keep any prefetching inside
2084 struct drm_i915_private *dev_priv = dev->dev_private;
2085 struct i915_address_space *ggtt_vm = &dev_priv->gtt.base;
2086 struct drm_mm_node *entry;
2087 struct drm_i915_gem_object *obj;
2088 unsigned long hole_start, hole_end;
2091 BUG_ON(mappable_end > end);
2093 /* Subtract the guard page ... */
2094 drm_mm_init(&ggtt_vm->mm, start, end - start - PAGE_SIZE);
2096 dev_priv->gtt.base.start = start;
2097 dev_priv->gtt.base.total = end - start;
2099 if (intel_vgpu_active(dev)) {
2100 ret = intel_vgt_balloon(dev);
2106 dev_priv->gtt.base.mm.color_adjust = i915_gtt_color_adjust;
2108 /* Mark any preallocated objects as occupied */
2109 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
2110 struct i915_vma *vma = i915_gem_obj_to_vma(obj, ggtt_vm);
2112 DRM_DEBUG_KMS("reserving preallocated space: %lx + %zx\n",
2113 i915_gem_obj_ggtt_offset(obj), obj->base.size);
2115 WARN_ON(i915_gem_obj_ggtt_bound(obj));
2116 ret = drm_mm_reserve_node(&ggtt_vm->mm, &vma->node);
2118 DRM_DEBUG_KMS("Reservation failed: %i\n", ret);
2121 vma->bound |= GLOBAL_BIND;
2124 /* Clear any non-preallocated blocks */
2125 drm_mm_for_each_hole(entry, &ggtt_vm->mm, hole_start, hole_end) {
2126 DRM_DEBUG_KMS("clearing unused GTT space: [%lx, %lx]\n",
2127 hole_start, hole_end);
2128 ggtt_vm->clear_range(ggtt_vm, hole_start,
2129 hole_end - hole_start, true);
2132 /* And finally clear the reserved guard page */
2133 ggtt_vm->clear_range(ggtt_vm, end - PAGE_SIZE, PAGE_SIZE, true);
2135 if (USES_PPGTT(dev) && !USES_FULL_PPGTT(dev)) {
2136 struct i915_hw_ppgtt *ppgtt;
2138 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2142 ret = __hw_ppgtt_init(dev, ppgtt, true);
2148 dev_priv->mm.aliasing_ppgtt = ppgtt;
2154 void i915_gem_init_global_gtt(struct drm_device *dev)
2156 struct drm_i915_private *dev_priv = dev->dev_private;
2157 unsigned long gtt_size, mappable_size;
2159 gtt_size = dev_priv->gtt.base.total;
2160 mappable_size = dev_priv->gtt.mappable_end;
2162 i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
2165 void i915_global_gtt_cleanup(struct drm_device *dev)
2167 struct drm_i915_private *dev_priv = dev->dev_private;
2168 struct i915_address_space *vm = &dev_priv->gtt.base;
2170 if (dev_priv->mm.aliasing_ppgtt) {
2171 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
2173 ppgtt->base.cleanup(&ppgtt->base);
2176 if (drm_mm_initialized(&vm->mm)) {
2177 if (intel_vgpu_active(dev))
2178 intel_vgt_deballoon();
2180 drm_mm_takedown(&vm->mm);
2181 list_del(&vm->global_link);
2187 static int setup_scratch_page(struct drm_device *dev)
2189 struct drm_i915_private *dev_priv = dev->dev_private;
2191 dma_addr_t dma_addr;
2193 page = alloc_page(GFP_KERNEL | GFP_DMA32 | __GFP_ZERO);
2196 set_pages_uc(page, 1);
2198 #ifdef CONFIG_INTEL_IOMMU
2199 dma_addr = pci_map_page(dev->pdev, page, 0, PAGE_SIZE,
2200 PCI_DMA_BIDIRECTIONAL);
2201 if (pci_dma_mapping_error(dev->pdev, dma_addr))
2204 dma_addr = page_to_phys(page);
2206 dev_priv->gtt.base.scratch.page = page;
2207 dev_priv->gtt.base.scratch.addr = dma_addr;
2212 static void teardown_scratch_page(struct drm_device *dev)
2214 struct drm_i915_private *dev_priv = dev->dev_private;
2215 struct page *page = dev_priv->gtt.base.scratch.page;
2217 set_pages_wb(page, 1);
2218 pci_unmap_page(dev->pdev, dev_priv->gtt.base.scratch.addr,
2219 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
2223 static inline unsigned int gen6_get_total_gtt_size(u16 snb_gmch_ctl)
2225 snb_gmch_ctl >>= SNB_GMCH_GGMS_SHIFT;
2226 snb_gmch_ctl &= SNB_GMCH_GGMS_MASK;
2227 return snb_gmch_ctl << 20;
2230 static inline unsigned int gen8_get_total_gtt_size(u16 bdw_gmch_ctl)
2232 bdw_gmch_ctl >>= BDW_GMCH_GGMS_SHIFT;
2233 bdw_gmch_ctl &= BDW_GMCH_GGMS_MASK;
2235 bdw_gmch_ctl = 1 << bdw_gmch_ctl;
2237 #ifdef CONFIG_X86_32
2238 /* Limit 32b platforms to a 2GB GGTT: 4 << 20 / pte size * PAGE_SIZE */
2239 if (bdw_gmch_ctl > 4)
2243 return bdw_gmch_ctl << 20;
2246 static inline unsigned int chv_get_total_gtt_size(u16 gmch_ctrl)
2248 gmch_ctrl >>= SNB_GMCH_GGMS_SHIFT;
2249 gmch_ctrl &= SNB_GMCH_GGMS_MASK;
2252 return 1 << (20 + gmch_ctrl);
2257 static inline size_t gen6_get_stolen_size(u16 snb_gmch_ctl)
2259 snb_gmch_ctl >>= SNB_GMCH_GMS_SHIFT;
2260 snb_gmch_ctl &= SNB_GMCH_GMS_MASK;
2261 return snb_gmch_ctl << 25; /* 32 MB units */
2264 static inline size_t gen8_get_stolen_size(u16 bdw_gmch_ctl)
2266 bdw_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2267 bdw_gmch_ctl &= BDW_GMCH_GMS_MASK;
2268 return bdw_gmch_ctl << 25; /* 32 MB units */
2271 static size_t chv_get_stolen_size(u16 gmch_ctrl)
2273 gmch_ctrl >>= SNB_GMCH_GMS_SHIFT;
2274 gmch_ctrl &= SNB_GMCH_GMS_MASK;
2277 * 0x0 to 0x10: 32MB increments starting at 0MB
2278 * 0x11 to 0x16: 4MB increments starting at 8MB
2279 * 0x17 to 0x1d: 4MB increments start at 36MB
2281 if (gmch_ctrl < 0x11)
2282 return gmch_ctrl << 25;
2283 else if (gmch_ctrl < 0x17)
2284 return (gmch_ctrl - 0x11 + 2) << 22;
2286 return (gmch_ctrl - 0x17 + 9) << 22;
2289 static size_t gen9_get_stolen_size(u16 gen9_gmch_ctl)
2291 gen9_gmch_ctl >>= BDW_GMCH_GMS_SHIFT;
2292 gen9_gmch_ctl &= BDW_GMCH_GMS_MASK;
2294 if (gen9_gmch_ctl < 0xf0)
2295 return gen9_gmch_ctl << 25; /* 32 MB units */
2297 /* 4MB increments starting at 0xf0 for 4MB */
2298 return (gen9_gmch_ctl - 0xf0 + 1) << 22;
2301 static int ggtt_probe_common(struct drm_device *dev,
2304 struct drm_i915_private *dev_priv = dev->dev_private;
2305 phys_addr_t gtt_phys_addr;
2308 /* For Modern GENs the PTEs and register space are split in the BAR */
2309 gtt_phys_addr = pci_resource_start(dev->pdev, 0) +
2310 (pci_resource_len(dev->pdev, 0) / 2);
2312 dev_priv->gtt.gsm = ioremap_wc(gtt_phys_addr, gtt_size);
2313 if (!dev_priv->gtt.gsm) {
2314 DRM_ERROR("Failed to map the gtt page table\n");
2318 ret = setup_scratch_page(dev);
2320 DRM_ERROR("Scratch setup failed\n");
2321 /* iounmap will also get called at remove, but meh */
2322 iounmap(dev_priv->gtt.gsm);
2328 /* The GGTT and PPGTT need a private PPAT setup in order to handle cacheability
2329 * bits. When using advanced contexts each context stores its own PAT, but
2330 * writing this data shouldn't be harmful even in those cases. */
2331 static void bdw_setup_private_ppat(struct drm_i915_private *dev_priv)
2335 pat = GEN8_PPAT(0, GEN8_PPAT_WB | GEN8_PPAT_LLC) | /* for normal objects, no eLLC */
2336 GEN8_PPAT(1, GEN8_PPAT_WC | GEN8_PPAT_LLCELLC) | /* for something pointing to ptes? */
2337 GEN8_PPAT(2, GEN8_PPAT_WT | GEN8_PPAT_LLCELLC) | /* for scanout with eLLC */
2338 GEN8_PPAT(3, GEN8_PPAT_UC) | /* Uncached objects, mostly for scanout */
2339 GEN8_PPAT(4, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(0)) |
2340 GEN8_PPAT(5, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(1)) |
2341 GEN8_PPAT(6, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(2)) |
2342 GEN8_PPAT(7, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3));
2344 if (!USES_PPGTT(dev_priv->dev))
2345 /* Spec: "For GGTT, there is NO pat_sel[2:0] from the entry,
2346 * so RTL will always use the value corresponding to
2348 * So let's disable cache for GGTT to avoid screen corruptions.
2349 * MOCS still can be used though.
2350 * - System agent ggtt writes (i.e. cpu gtt mmaps) already work
2351 * before this patch, i.e. the same uncached + snooping access
2352 * like on gen6/7 seems to be in effect.
2353 * - So this just fixes blitter/render access. Again it looks
2354 * like it's not just uncached access, but uncached + snooping.
2355 * So we can still hold onto all our assumptions wrt cpu
2356 * clflushing on LLC machines.
2358 pat = GEN8_PPAT(0, GEN8_PPAT_UC);
2360 /* XXX: spec defines this as 2 distinct registers. It's unclear if a 64b
2361 * write would work. */
2362 I915_WRITE(GEN8_PRIVATE_PAT, pat);
2363 I915_WRITE(GEN8_PRIVATE_PAT + 4, pat >> 32);
2366 static void chv_setup_private_ppat(struct drm_i915_private *dev_priv)
2371 * Map WB on BDW to snooped on CHV.
2373 * Only the snoop bit has meaning for CHV, the rest is
2376 * The hardware will never snoop for certain types of accesses:
2377 * - CPU GTT (GMADR->GGTT->no snoop->memory)
2378 * - PPGTT page tables
2379 * - some other special cycles
2381 * As with BDW, we also need to consider the following for GT accesses:
2382 * "For GGTT, there is NO pat_sel[2:0] from the entry,
2383 * so RTL will always use the value corresponding to
2385 * Which means we must set the snoop bit in PAT entry 0
2386 * in order to keep the global status page working.
2388 pat = GEN8_PPAT(0, CHV_PPAT_SNOOP) |
2392 GEN8_PPAT(4, CHV_PPAT_SNOOP) |
2393 GEN8_PPAT(5, CHV_PPAT_SNOOP) |
2394 GEN8_PPAT(6, CHV_PPAT_SNOOP) |
2395 GEN8_PPAT(7, CHV_PPAT_SNOOP);
2397 I915_WRITE(GEN8_PRIVATE_PAT, pat);
2398 I915_WRITE(GEN8_PRIVATE_PAT + 4, pat >> 32);
2401 static int gen8_gmch_probe(struct drm_device *dev,
2404 phys_addr_t *mappable_base,
2405 unsigned long *mappable_end)
2407 struct drm_i915_private *dev_priv = dev->dev_private;
2408 unsigned int gtt_size;
2412 /* TODO: We're not aware of mappable constraints on gen8 yet */
2413 *mappable_base = pci_resource_start(dev->pdev, 2);
2414 *mappable_end = pci_resource_len(dev->pdev, 2);
2416 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(39)))
2417 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(39));
2419 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
2421 if (INTEL_INFO(dev)->gen >= 9) {
2422 *stolen = gen9_get_stolen_size(snb_gmch_ctl);
2423 gtt_size = gen8_get_total_gtt_size(snb_gmch_ctl);
2424 } else if (IS_CHERRYVIEW(dev)) {
2425 *stolen = chv_get_stolen_size(snb_gmch_ctl);
2426 gtt_size = chv_get_total_gtt_size(snb_gmch_ctl);
2428 *stolen = gen8_get_stolen_size(snb_gmch_ctl);
2429 gtt_size = gen8_get_total_gtt_size(snb_gmch_ctl);
2432 *gtt_total = (gtt_size / sizeof(gen8_pte_t)) << PAGE_SHIFT;
2434 if (IS_CHERRYVIEW(dev))
2435 chv_setup_private_ppat(dev_priv);
2437 bdw_setup_private_ppat(dev_priv);
2439 ret = ggtt_probe_common(dev, gtt_size);
2441 dev_priv->gtt.base.clear_range = gen8_ggtt_clear_range;
2442 dev_priv->gtt.base.insert_entries = gen8_ggtt_insert_entries;
2447 static int gen6_gmch_probe(struct drm_device *dev,
2450 phys_addr_t *mappable_base,
2451 unsigned long *mappable_end)
2453 struct drm_i915_private *dev_priv = dev->dev_private;
2454 unsigned int gtt_size;
2458 *mappable_base = pci_resource_start(dev->pdev, 2);
2459 *mappable_end = pci_resource_len(dev->pdev, 2);
2461 /* 64/512MB is the current min/max we actually know of, but this is just
2462 * a coarse sanity check.
2464 if ((*mappable_end < (64<<20) || (*mappable_end > (512<<20)))) {
2465 DRM_ERROR("Unknown GMADR size (%lx)\n",
2466 dev_priv->gtt.mappable_end);
2470 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(40)))
2471 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(40));
2472 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
2474 *stolen = gen6_get_stolen_size(snb_gmch_ctl);
2476 gtt_size = gen6_get_total_gtt_size(snb_gmch_ctl);
2477 *gtt_total = (gtt_size / sizeof(gen6_pte_t)) << PAGE_SHIFT;
2479 ret = ggtt_probe_common(dev, gtt_size);
2481 dev_priv->gtt.base.clear_range = gen6_ggtt_clear_range;
2482 dev_priv->gtt.base.insert_entries = gen6_ggtt_insert_entries;
2487 static void gen6_gmch_remove(struct i915_address_space *vm)
2490 struct i915_gtt *gtt = container_of(vm, struct i915_gtt, base);
2493 teardown_scratch_page(vm->dev);
2496 static int i915_gmch_probe(struct drm_device *dev,
2499 phys_addr_t *mappable_base,
2500 unsigned long *mappable_end)
2502 struct drm_i915_private *dev_priv = dev->dev_private;
2505 ret = intel_gmch_probe(dev_priv->bridge_dev, dev_priv->dev->pdev, NULL);
2507 DRM_ERROR("failed to set up gmch\n");
2511 intel_gtt_get(gtt_total, stolen, mappable_base, mappable_end);
2513 dev_priv->gtt.do_idle_maps = needs_idle_maps(dev_priv->dev);
2514 dev_priv->gtt.base.clear_range = i915_ggtt_clear_range;
2516 if (unlikely(dev_priv->gtt.do_idle_maps))
2517 DRM_INFO("applying Ironlake quirks for intel_iommu\n");
2522 static void i915_gmch_remove(struct i915_address_space *vm)
2524 intel_gmch_remove();
2527 int i915_gem_gtt_init(struct drm_device *dev)
2529 struct drm_i915_private *dev_priv = dev->dev_private;
2530 struct i915_gtt *gtt = &dev_priv->gtt;
2533 if (INTEL_INFO(dev)->gen <= 5) {
2534 gtt->gtt_probe = i915_gmch_probe;
2535 gtt->base.cleanup = i915_gmch_remove;
2536 } else if (INTEL_INFO(dev)->gen < 8) {
2537 gtt->gtt_probe = gen6_gmch_probe;
2538 gtt->base.cleanup = gen6_gmch_remove;
2539 if (IS_HASWELL(dev) && dev_priv->ellc_size)
2540 gtt->base.pte_encode = iris_pte_encode;
2541 else if (IS_HASWELL(dev))
2542 gtt->base.pte_encode = hsw_pte_encode;
2543 else if (IS_VALLEYVIEW(dev))
2544 gtt->base.pte_encode = byt_pte_encode;
2545 else if (INTEL_INFO(dev)->gen >= 7)
2546 gtt->base.pte_encode = ivb_pte_encode;
2548 gtt->base.pte_encode = snb_pte_encode;
2550 dev_priv->gtt.gtt_probe = gen8_gmch_probe;
2551 dev_priv->gtt.base.cleanup = gen6_gmch_remove;
2554 ret = gtt->gtt_probe(dev, >t->base.total, >t->stolen_size,
2555 >t->mappable_base, >t->mappable_end);
2559 gtt->base.dev = dev;
2561 /* GMADR is the PCI mmio aperture into the global GTT. */
2562 DRM_INFO("Memory usable by graphics device = %zdM\n",
2563 gtt->base.total >> 20);
2564 DRM_DEBUG_DRIVER("GMADR size = %ldM\n", gtt->mappable_end >> 20);
2565 DRM_DEBUG_DRIVER("GTT stolen size = %zdM\n", gtt->stolen_size >> 20);
2566 #ifdef CONFIG_INTEL_IOMMU
2567 if (intel_iommu_gfx_mapped)
2568 DRM_INFO("VT-d active for gfx access\n");
2571 * i915.enable_ppgtt is read-only, so do an early pass to validate the
2572 * user's requested state against the hardware/driver capabilities. We
2573 * do this now so that we can print out any log messages once rather
2574 * than every time we check intel_enable_ppgtt().
2576 i915.enable_ppgtt = sanitize_enable_ppgtt(dev, i915.enable_ppgtt);
2577 DRM_DEBUG_DRIVER("ppgtt mode: %i\n", i915.enable_ppgtt);
2582 static struct i915_vma *
2583 __i915_gem_vma_create(struct drm_i915_gem_object *obj,
2584 struct i915_address_space *vm,
2585 const struct i915_ggtt_view *ggtt_view)
2587 struct i915_vma *vma;
2589 if (WARN_ON(i915_is_ggtt(vm) != !!ggtt_view))
2590 return ERR_PTR(-EINVAL);
2591 vma = kzalloc(sizeof(*vma), GFP_KERNEL);
2593 return ERR_PTR(-ENOMEM);
2595 INIT_LIST_HEAD(&vma->vma_link);
2596 INIT_LIST_HEAD(&vma->mm_list);
2597 INIT_LIST_HEAD(&vma->exec_list);
2601 if (INTEL_INFO(vm->dev)->gen >= 6) {
2602 if (i915_is_ggtt(vm)) {
2603 vma->ggtt_view = *ggtt_view;
2605 vma->unbind_vma = ggtt_unbind_vma;
2606 vma->bind_vma = ggtt_bind_vma;
2608 vma->unbind_vma = ppgtt_unbind_vma;
2609 vma->bind_vma = ppgtt_bind_vma;
2612 BUG_ON(!i915_is_ggtt(vm));
2613 vma->ggtt_view = *ggtt_view;
2614 vma->unbind_vma = i915_ggtt_unbind_vma;
2615 vma->bind_vma = i915_ggtt_bind_vma;
2618 list_add_tail(&vma->vma_link, &obj->vma_list);
2619 if (!i915_is_ggtt(vm))
2620 i915_ppgtt_get(i915_vm_to_ppgtt(vm));
2626 i915_gem_obj_lookup_or_create_vma(struct drm_i915_gem_object *obj,
2627 struct i915_address_space *vm)
2629 struct i915_vma *vma;
2631 vma = i915_gem_obj_to_vma(obj, vm);
2633 vma = __i915_gem_vma_create(obj, vm,
2634 i915_is_ggtt(vm) ? &i915_ggtt_view_normal : NULL);
2640 i915_gem_obj_lookup_or_create_ggtt_vma(struct drm_i915_gem_object *obj,
2641 const struct i915_ggtt_view *view)
2643 struct i915_address_space *ggtt = i915_obj_to_ggtt(obj);
2644 struct i915_vma *vma;
2647 return ERR_PTR(-EINVAL);
2649 vma = i915_gem_obj_to_ggtt_view(obj, view);
2655 vma = __i915_gem_vma_create(obj, ggtt, view);
2662 rotate_pages(dma_addr_t *in, unsigned int width, unsigned int height,
2663 struct sg_table *st)
2665 unsigned int column, row;
2666 unsigned int src_idx;
2667 struct scatterlist *sg = st->sgl;
2671 for (column = 0; column < width; column++) {
2672 src_idx = width * (height - 1) + column;
2673 for (row = 0; row < height; row++) {
2675 /* We don't need the pages, but need to initialize
2676 * the entries so the sg list can be happily traversed.
2677 * The only thing we need are DMA addresses.
2679 sg_set_page(sg, NULL, PAGE_SIZE, 0);
2680 sg_dma_address(sg) = in[src_idx];
2681 sg_dma_len(sg) = PAGE_SIZE;
2688 static struct sg_table *
2689 intel_rotate_fb_obj_pages(struct i915_ggtt_view *ggtt_view,
2690 struct drm_i915_gem_object *obj)
2692 struct drm_device *dev = obj->base.dev;
2693 struct intel_rotation_info *rot_info = &ggtt_view->rotation_info;
2694 unsigned long size, pages, rot_pages;
2695 struct sg_page_iter sg_iter;
2697 dma_addr_t *page_addr_list;
2698 struct sg_table *st;
2699 unsigned int tile_pitch, tile_height;
2700 unsigned int width_pages, height_pages;
2703 pages = obj->base.size / PAGE_SIZE;
2705 /* Calculate tiling geometry. */
2706 tile_height = intel_tile_height(dev, rot_info->pixel_format,
2707 rot_info->fb_modifier);
2708 tile_pitch = PAGE_SIZE / tile_height;
2709 width_pages = DIV_ROUND_UP(rot_info->pitch, tile_pitch);
2710 height_pages = DIV_ROUND_UP(rot_info->height, tile_height);
2711 rot_pages = width_pages * height_pages;
2712 size = rot_pages * PAGE_SIZE;
2714 /* Allocate a temporary list of source pages for random access. */
2715 page_addr_list = drm_malloc_ab(pages, sizeof(dma_addr_t));
2716 if (!page_addr_list)
2717 return ERR_PTR(ret);
2719 /* Allocate target SG list. */
2720 st = kmalloc(sizeof(*st), GFP_KERNEL);
2724 ret = sg_alloc_table(st, rot_pages, GFP_KERNEL);
2728 /* Populate source page list from the object. */
2730 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) {
2731 page_addr_list[i] = sg_page_iter_dma_address(&sg_iter);
2735 /* Rotate the pages. */
2736 rotate_pages(page_addr_list, width_pages, height_pages, st);
2739 "Created rotated page mapping for object size %lu (pitch=%u, height=%u, pixel_format=0x%x, %ux%u tiles, %lu pages).\n",
2740 size, rot_info->pitch, rot_info->height,
2741 rot_info->pixel_format, width_pages, height_pages,
2744 drm_free_large(page_addr_list);
2751 drm_free_large(page_addr_list);
2754 "Failed to create rotated mapping for object size %lu! (%d) (pitch=%u, height=%u, pixel_format=0x%x, %ux%u tiles, %lu pages)\n",
2755 size, ret, rot_info->pitch, rot_info->height,
2756 rot_info->pixel_format, width_pages, height_pages,
2758 return ERR_PTR(ret);
2762 i915_get_ggtt_vma_pages(struct i915_vma *vma)
2766 if (vma->ggtt_view.pages)
2769 if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL)
2770 vma->ggtt_view.pages = vma->obj->pages;
2771 else if (vma->ggtt_view.type == I915_GGTT_VIEW_ROTATED)
2772 vma->ggtt_view.pages =
2773 intel_rotate_fb_obj_pages(&vma->ggtt_view, vma->obj);
2775 WARN_ONCE(1, "GGTT view %u not implemented!\n",
2776 vma->ggtt_view.type);
2778 if (!vma->ggtt_view.pages) {
2779 DRM_ERROR("Failed to get pages for GGTT view type %u!\n",
2780 vma->ggtt_view.type);
2782 } else if (IS_ERR(vma->ggtt_view.pages)) {
2783 ret = PTR_ERR(vma->ggtt_view.pages);
2784 vma->ggtt_view.pages = NULL;
2785 DRM_ERROR("Failed to get pages for VMA view type %u (%d)!\n",
2786 vma->ggtt_view.type, ret);
2793 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
2795 * @cache_level: mapping cache level
2796 * @flags: flags like global or local mapping
2798 * DMA addresses are taken from the scatter-gather table of this object (or of
2799 * this VMA in case of non-default GGTT views) and PTE entries set up.
2800 * Note that DMA addresses are also the only part of the SG table we care about.
2802 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
2805 if (i915_is_ggtt(vma->vm)) {
2806 int ret = i915_get_ggtt_vma_pages(vma);
2812 vma->bind_vma(vma, cache_level, flags);