2 * Copyright © 2008-2012 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Eric Anholt <eric@anholt.net>
25 * Chris Wilson <chris@chris-wilson.co.uk>
30 #include <drm/i915_drm.h>
33 #define KB(x) ((x) * 1024)
34 #define MB(x) (KB(x) * 1024)
37 * The BIOS typically reserves some of the system's memory for the exclusive
38 * use of the integrated graphics. This memory is no longer available for
39 * use by the OS and so the user finds that his system has less memory
40 * available than he put in. We refer to this memory as stolen.
42 * The BIOS will allocate its framebuffer from the stolen memory. Our
43 * goal is try to reuse that object for our own fbcon which must always
44 * be available for panics. Anything else we can reuse the stolen memory
48 int i915_gem_stolen_insert_node_in_range(struct drm_i915_private *dev_priv,
49 struct drm_mm_node *node, u64 size,
50 unsigned alignment, u64 start, u64 end)
54 if (!drm_mm_initialized(&dev_priv->mm.stolen))
57 mutex_lock(&dev_priv->mm.stolen_lock);
58 ret = drm_mm_insert_node_in_range(&dev_priv->mm.stolen, node,
60 start, end, DRM_MM_INSERT_BEST);
61 mutex_unlock(&dev_priv->mm.stolen_lock);
66 int i915_gem_stolen_insert_node(struct drm_i915_private *dev_priv,
67 struct drm_mm_node *node, u64 size,
70 return i915_gem_stolen_insert_node_in_range(dev_priv, node, size,
71 alignment, 0, U64_MAX);
74 void i915_gem_stolen_remove_node(struct drm_i915_private *dev_priv,
75 struct drm_mm_node *node)
77 mutex_lock(&dev_priv->mm.stolen_lock);
78 drm_mm_remove_node(node);
79 mutex_unlock(&dev_priv->mm.stolen_lock);
82 static dma_addr_t i915_stolen_to_dma(struct drm_i915_private *dev_priv)
84 struct pci_dev *pdev = dev_priv->drm.pdev;
85 struct i915_ggtt *ggtt = &dev_priv->ggtt;
89 /* Almost universally we can find the Graphics Base of Stolen Memory
90 * at register BSM (0x5c) in the igfx configuration space. On a few
91 * (desktop) machines this is also mirrored in the bridge device at
92 * different locations, or in the MCHBAR.
94 * On 865 we just check the TOUD register.
96 * On 830/845/85x the stolen memory base isn't available in any
97 * register. We need to calculate it as TOM-TSEG_SIZE-stolen_size.
101 if (INTEL_GEN(dev_priv) >= 3) {
104 pci_read_config_dword(pdev, INTEL_BSM, &bsm);
106 base = bsm & INTEL_BSM_MASK;
107 } else if (IS_I865G(dev_priv)) {
112 pci_bus_read_config_byte(pdev->bus, PCI_DEVFN(0, 0),
115 if (tmp & TSEG_ENABLE) {
116 switch (tmp & I845_TSEG_SIZE_MASK) {
117 case I845_TSEG_SIZE_512K:
120 case I845_TSEG_SIZE_1M:
126 pci_bus_read_config_word(pdev->bus, PCI_DEVFN(0, 0),
129 base = (toud << 16) + tseg_size;
130 } else if (IS_I85X(dev_priv)) {
135 pci_bus_read_config_byte(pdev->bus, PCI_DEVFN(0, 0),
138 if (tmp & TSEG_ENABLE)
141 pci_bus_read_config_byte(pdev->bus, PCI_DEVFN(0, 1),
145 base = tom - tseg_size - ggtt->stolen_size;
146 } else if (IS_I845G(dev_priv)) {
151 pci_bus_read_config_byte(pdev->bus, PCI_DEVFN(0, 0),
154 if (tmp & TSEG_ENABLE) {
155 switch (tmp & I845_TSEG_SIZE_MASK) {
156 case I845_TSEG_SIZE_512K:
159 case I845_TSEG_SIZE_1M:
165 pci_bus_read_config_byte(pdev->bus, PCI_DEVFN(0, 0),
169 base = tom - tseg_size - ggtt->stolen_size;
170 } else if (IS_I830(dev_priv)) {
175 pci_bus_read_config_byte(pdev->bus, PCI_DEVFN(0, 0),
178 if (tmp & TSEG_ENABLE) {
179 if (tmp & I830_TSEG_SIZE_1M)
185 pci_bus_read_config_byte(pdev->bus, PCI_DEVFN(0, 0),
189 base = tom - tseg_size - ggtt->stolen_size;
192 if (base == 0 || add_overflows(base, ggtt->stolen_size))
195 /* make sure we don't clobber the GTT if it's within stolen memory */
196 if (INTEL_GEN(dev_priv) <= 4 &&
197 !IS_G33(dev_priv) && !IS_PINEVIEW(dev_priv) && !IS_G4X(dev_priv)) {
199 dma_addr_t start, end;
201 { .start = base, .end = base + ggtt->stolen_size, },
202 { .start = base, .end = base + ggtt->stolen_size, },
204 u64 ggtt_start, ggtt_end;
206 ggtt_start = I915_READ(PGTBL_CTL);
207 if (IS_GEN4(dev_priv))
208 ggtt_start = (ggtt_start & PGTBL_ADDRESS_LO_MASK) |
209 (ggtt_start & PGTBL_ADDRESS_HI_MASK) << 28;
211 ggtt_start &= PGTBL_ADDRESS_LO_MASK;
212 ggtt_end = ggtt_start + ggtt_total_entries(ggtt) * 4;
214 if (ggtt_start >= stolen[0].start && ggtt_start < stolen[0].end)
215 stolen[0].end = ggtt_start;
216 if (ggtt_end > stolen[1].start && ggtt_end <= stolen[1].end)
217 stolen[1].start = ggtt_end;
219 /* pick the larger of the two chunks */
220 if (stolen[0].end - stolen[0].start >
221 stolen[1].end - stolen[1].start) {
222 base = stolen[0].start;
223 ggtt->stolen_size = stolen[0].end - stolen[0].start;
225 base = stolen[1].start;
226 ggtt->stolen_size = stolen[1].end - stolen[1].start;
229 if (stolen[0].start != stolen[1].start ||
230 stolen[0].end != stolen[1].end) {
231 dma_addr_t end = base + ggtt->stolen_size - 1;
233 DRM_DEBUG_KMS("GTT within stolen memory at 0x%llx-0x%llx\n",
234 (unsigned long long)ggtt_start,
235 (unsigned long long)ggtt_end - 1);
236 DRM_DEBUG_KMS("Stolen memory adjusted to %pad - %pad\n",
242 /* Verify that nothing else uses this physical address. Stolen
243 * memory should be reserved by the BIOS and hidden from the
244 * kernel. So if the region is already marked as busy, something
245 * is seriously wrong.
247 r = devm_request_mem_region(dev_priv->drm.dev, base, ggtt->stolen_size,
248 "Graphics Stolen Memory");
251 * One more attempt but this time requesting region from
252 * base + 1, as we have seen that this resolves the region
253 * conflict with the PCI Bus.
254 * This is a BIOS w/a: Some BIOS wrap stolen in the root
255 * PCI bus, but have an off-by-one error. Hence retry the
256 * reservation starting from 1 instead of 0.
257 * There's also BIOS with off-by-one on the other end.
259 r = devm_request_mem_region(dev_priv->drm.dev, base + 1,
260 ggtt->stolen_size - 2,
261 "Graphics Stolen Memory");
263 * GEN3 firmware likes to smash pci bridges into the stolen
264 * range. Apparently this works.
266 if (r == NULL && !IS_GEN3(dev_priv)) {
267 dma_addr_t end = base + ggtt->stolen_size;
269 DRM_ERROR("conflict detected with stolen region: [%pad - %pad]\n",
278 void i915_gem_cleanup_stolen(struct drm_device *dev)
280 struct drm_i915_private *dev_priv = to_i915(dev);
282 if (!drm_mm_initialized(&dev_priv->mm.stolen))
285 drm_mm_takedown(&dev_priv->mm.stolen);
288 static void g4x_get_stolen_reserved(struct drm_i915_private *dev_priv,
289 dma_addr_t *base, u32 *size)
291 struct i915_ggtt *ggtt = &dev_priv->ggtt;
292 uint32_t reg_val = I915_READ(IS_GM45(dev_priv) ?
293 CTG_STOLEN_RESERVED :
294 ELK_STOLEN_RESERVED);
295 dma_addr_t stolen_top = dev_priv->mm.stolen_base + ggtt->stolen_size;
297 *base = (reg_val & G4X_STOLEN_RESERVED_ADDR2_MASK) << 16;
299 WARN_ON((reg_val & G4X_STOLEN_RESERVED_ADDR1_MASK) < *base);
301 /* On these platforms, the register doesn't have a size field, so the
302 * size is the distance between the base and the top of the stolen
303 * memory. We also have the genuine case where base is zero and there's
304 * nothing reserved. */
308 *size = stolen_top - *base;
311 static void gen6_get_stolen_reserved(struct drm_i915_private *dev_priv,
312 dma_addr_t *base, u32 *size)
314 uint32_t reg_val = I915_READ(GEN6_STOLEN_RESERVED);
316 *base = reg_val & GEN6_STOLEN_RESERVED_ADDR_MASK;
318 switch (reg_val & GEN6_STOLEN_RESERVED_SIZE_MASK) {
319 case GEN6_STOLEN_RESERVED_1M:
322 case GEN6_STOLEN_RESERVED_512K:
325 case GEN6_STOLEN_RESERVED_256K:
328 case GEN6_STOLEN_RESERVED_128K:
333 MISSING_CASE(reg_val & GEN6_STOLEN_RESERVED_SIZE_MASK);
337 static void gen7_get_stolen_reserved(struct drm_i915_private *dev_priv,
338 dma_addr_t *base, u32 *size)
340 uint32_t reg_val = I915_READ(GEN6_STOLEN_RESERVED);
342 *base = reg_val & GEN7_STOLEN_RESERVED_ADDR_MASK;
344 switch (reg_val & GEN7_STOLEN_RESERVED_SIZE_MASK) {
345 case GEN7_STOLEN_RESERVED_1M:
348 case GEN7_STOLEN_RESERVED_256K:
353 MISSING_CASE(reg_val & GEN7_STOLEN_RESERVED_SIZE_MASK);
357 static void chv_get_stolen_reserved(struct drm_i915_private *dev_priv,
358 dma_addr_t *base, u32 *size)
360 uint32_t reg_val = I915_READ(GEN6_STOLEN_RESERVED);
362 *base = reg_val & GEN6_STOLEN_RESERVED_ADDR_MASK;
364 switch (reg_val & GEN8_STOLEN_RESERVED_SIZE_MASK) {
365 case GEN8_STOLEN_RESERVED_1M:
368 case GEN8_STOLEN_RESERVED_2M:
369 *size = 2 * 1024 * 1024;
371 case GEN8_STOLEN_RESERVED_4M:
372 *size = 4 * 1024 * 1024;
374 case GEN8_STOLEN_RESERVED_8M:
375 *size = 8 * 1024 * 1024;
378 *size = 8 * 1024 * 1024;
379 MISSING_CASE(reg_val & GEN8_STOLEN_RESERVED_SIZE_MASK);
383 static void bdw_get_stolen_reserved(struct drm_i915_private *dev_priv,
384 dma_addr_t *base, u32 *size)
386 struct i915_ggtt *ggtt = &dev_priv->ggtt;
387 uint32_t reg_val = I915_READ(GEN6_STOLEN_RESERVED);
388 dma_addr_t stolen_top;
390 stolen_top = dev_priv->mm.stolen_base + ggtt->stolen_size;
392 *base = reg_val & GEN6_STOLEN_RESERVED_ADDR_MASK;
394 /* On these platforms, the register doesn't have a size field, so the
395 * size is the distance between the base and the top of the stolen
396 * memory. We also have the genuine case where base is zero and there's
397 * nothing reserved. */
401 *size = stolen_top - *base;
404 int i915_gem_init_stolen(struct drm_i915_private *dev_priv)
406 struct i915_ggtt *ggtt = &dev_priv->ggtt;
407 dma_addr_t reserved_base, stolen_top;
408 u32 reserved_total, reserved_size;
409 u32 stolen_usable_start;
411 mutex_init(&dev_priv->mm.stolen_lock);
413 if (intel_vgpu_active(dev_priv)) {
414 DRM_INFO("iGVT-g active, disabling use of stolen memory\n");
418 if (intel_vtd_active() && INTEL_GEN(dev_priv) < 8) {
419 DRM_INFO("DMAR active, disabling use of stolen memory\n");
423 if (ggtt->stolen_size == 0)
426 dev_priv->mm.stolen_base = i915_stolen_to_dma(dev_priv);
427 if (dev_priv->mm.stolen_base == 0)
430 stolen_top = dev_priv->mm.stolen_base + ggtt->stolen_size;
434 switch (INTEL_INFO(dev_priv)->gen) {
439 if (IS_G4X(dev_priv))
440 g4x_get_stolen_reserved(dev_priv,
441 &reserved_base, &reserved_size);
444 /* Assume the gen6 maximum for the older platforms. */
445 reserved_size = 1024 * 1024;
446 reserved_base = stolen_top - reserved_size;
449 gen6_get_stolen_reserved(dev_priv,
450 &reserved_base, &reserved_size);
453 gen7_get_stolen_reserved(dev_priv,
454 &reserved_base, &reserved_size);
458 chv_get_stolen_reserved(dev_priv,
459 &reserved_base, &reserved_size);
461 bdw_get_stolen_reserved(dev_priv,
462 &reserved_base, &reserved_size);
466 /* It is possible for the reserved base to be zero, but the register
467 * field for size doesn't have a zero option. */
468 if (reserved_base == 0) {
470 reserved_base = stolen_top;
473 if (reserved_base < dev_priv->mm.stolen_base ||
474 reserved_base + reserved_size > stolen_top) {
475 dma_addr_t reserved_top = reserved_base + reserved_size;
476 DRM_DEBUG_KMS("Stolen reserved area [%pad - %pad] outside stolen memory [%pad - %pad]\n",
477 &reserved_base, &reserved_top,
478 &dev_priv->mm.stolen_base, &stolen_top);
482 ggtt->stolen_reserved_base = reserved_base;
483 ggtt->stolen_reserved_size = reserved_size;
485 /* It is possible for the reserved area to end before the end of stolen
486 * memory, so just consider the start. */
487 reserved_total = stolen_top - reserved_base;
489 DRM_DEBUG_KMS("Memory reserved for graphics device: %uK, usable: %uK\n",
490 ggtt->stolen_size >> 10,
491 (ggtt->stolen_size - reserved_total) >> 10);
493 stolen_usable_start = 0;
494 /* WaSkipStolenMemoryFirstPage:bdw+ */
495 if (INTEL_GEN(dev_priv) >= 8)
496 stolen_usable_start = 4096;
498 ggtt->stolen_usable_size =
499 ggtt->stolen_size - reserved_total - stolen_usable_start;
501 /* Basic memrange allocator for stolen space. */
502 drm_mm_init(&dev_priv->mm.stolen, stolen_usable_start,
503 ggtt->stolen_usable_size);
508 static struct sg_table *
509 i915_pages_create_for_stolen(struct drm_device *dev,
510 u32 offset, u32 size)
512 struct drm_i915_private *dev_priv = to_i915(dev);
514 struct scatterlist *sg;
516 GEM_BUG_ON(range_overflows(offset, size, dev_priv->ggtt.stolen_size));
518 /* We hide that we have no struct page backing our stolen object
519 * by wrapping the contiguous physical allocation with a fake
520 * dma mapping in a single scatterlist.
523 st = kmalloc(sizeof(*st), GFP_KERNEL);
525 return ERR_PTR(-ENOMEM);
527 if (sg_alloc_table(st, 1, GFP_KERNEL)) {
529 return ERR_PTR(-ENOMEM);
536 sg_dma_address(sg) = (dma_addr_t)dev_priv->mm.stolen_base + offset;
537 sg_dma_len(sg) = size;
542 static struct sg_table *
543 i915_gem_object_get_pages_stolen(struct drm_i915_gem_object *obj)
545 return i915_pages_create_for_stolen(obj->base.dev,
550 static void i915_gem_object_put_pages_stolen(struct drm_i915_gem_object *obj,
551 struct sg_table *pages)
553 /* Should only be called from i915_gem_object_release_stolen() */
554 sg_free_table(pages);
559 i915_gem_object_release_stolen(struct drm_i915_gem_object *obj)
561 struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
562 struct drm_mm_node *stolen = fetch_and_zero(&obj->stolen);
566 __i915_gem_object_unpin_pages(obj);
568 i915_gem_stolen_remove_node(dev_priv, stolen);
572 static const struct drm_i915_gem_object_ops i915_gem_object_stolen_ops = {
573 .get_pages = i915_gem_object_get_pages_stolen,
574 .put_pages = i915_gem_object_put_pages_stolen,
575 .release = i915_gem_object_release_stolen,
578 static struct drm_i915_gem_object *
579 _i915_gem_object_create_stolen(struct drm_i915_private *dev_priv,
580 struct drm_mm_node *stolen)
582 struct drm_i915_gem_object *obj;
583 unsigned int cache_level;
585 obj = i915_gem_object_alloc(dev_priv);
589 drm_gem_private_object_init(&dev_priv->drm, &obj->base, stolen->size);
590 i915_gem_object_init(obj, &i915_gem_object_stolen_ops);
592 obj->stolen = stolen;
593 obj->base.read_domains = I915_GEM_DOMAIN_CPU | I915_GEM_DOMAIN_GTT;
594 cache_level = HAS_LLC(dev_priv) ? I915_CACHE_LLC : I915_CACHE_NONE;
595 i915_gem_object_set_cache_coherency(obj, cache_level);
597 if (i915_gem_object_pin_pages(obj))
603 i915_gem_object_free(obj);
607 struct drm_i915_gem_object *
608 i915_gem_object_create_stolen(struct drm_i915_private *dev_priv, u32 size)
610 struct drm_i915_gem_object *obj;
611 struct drm_mm_node *stolen;
614 if (!drm_mm_initialized(&dev_priv->mm.stolen))
620 stolen = kzalloc(sizeof(*stolen), GFP_KERNEL);
624 ret = i915_gem_stolen_insert_node(dev_priv, stolen, size, 4096);
630 obj = _i915_gem_object_create_stolen(dev_priv, stolen);
634 i915_gem_stolen_remove_node(dev_priv, stolen);
639 struct drm_i915_gem_object *
640 i915_gem_object_create_stolen_for_preallocated(struct drm_i915_private *dev_priv,
645 struct i915_ggtt *ggtt = &dev_priv->ggtt;
646 struct drm_i915_gem_object *obj;
647 struct drm_mm_node *stolen;
648 struct i915_vma *vma;
651 if (!drm_mm_initialized(&dev_priv->mm.stolen))
654 lockdep_assert_held(&dev_priv->drm.struct_mutex);
656 DRM_DEBUG_KMS("creating preallocated stolen object: stolen_offset=%x, gtt_offset=%x, size=%x\n",
657 stolen_offset, gtt_offset, size);
659 /* KISS and expect everything to be page-aligned */
660 if (WARN_ON(size == 0) ||
661 WARN_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE)) ||
662 WARN_ON(!IS_ALIGNED(stolen_offset, I915_GTT_MIN_ALIGNMENT)))
665 stolen = kzalloc(sizeof(*stolen), GFP_KERNEL);
669 stolen->start = stolen_offset;
671 mutex_lock(&dev_priv->mm.stolen_lock);
672 ret = drm_mm_reserve_node(&dev_priv->mm.stolen, stolen);
673 mutex_unlock(&dev_priv->mm.stolen_lock);
675 DRM_DEBUG_KMS("failed to allocate stolen space\n");
680 obj = _i915_gem_object_create_stolen(dev_priv, stolen);
682 DRM_DEBUG_KMS("failed to allocate stolen object\n");
683 i915_gem_stolen_remove_node(dev_priv, stolen);
688 /* Some objects just need physical mem from stolen space */
689 if (gtt_offset == I915_GTT_OFFSET_NONE)
692 ret = i915_gem_object_pin_pages(obj);
696 vma = i915_vma_instance(obj, &ggtt->base, NULL);
702 /* To simplify the initialisation sequence between KMS and GTT,
703 * we allow construction of the stolen object prior to
704 * setting up the GTT space. The actual reservation will occur
707 ret = i915_gem_gtt_reserve(&ggtt->base, &vma->node,
708 size, gtt_offset, obj->cache_level,
711 DRM_DEBUG_KMS("failed to allocate stolen GTT space\n");
715 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
717 vma->pages = obj->mm.pages;
718 vma->flags |= I915_VMA_GLOBAL_BIND;
719 __i915_vma_set_map_and_fenceable(vma);
720 list_move_tail(&vma->vm_link, &ggtt->base.inactive_list);
721 list_move_tail(&obj->global_link, &dev_priv->mm.bound_list);
727 i915_gem_object_unpin_pages(obj);
729 i915_gem_object_put(obj);