2 * Copyright 2016 Advanced Micro Devices, Inc.
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 shall be included in
12 * all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
22 * Authors: Christian König
25 #include <linux/dma-mapping.h>
26 #include <drm/ttm/ttm_range_manager.h>
29 #include "amdgpu_vm.h"
30 #include "amdgpu_res_cursor.h"
31 #include "amdgpu_atomfirmware.h"
34 struct amdgpu_vram_reservation {
35 struct list_head node;
36 struct drm_mm_node mm_node;
39 static inline struct amdgpu_vram_mgr *
40 to_vram_mgr(struct ttm_resource_manager *man)
42 return container_of(man, struct amdgpu_vram_mgr, manager);
45 static inline struct amdgpu_device *
46 to_amdgpu_device(struct amdgpu_vram_mgr *mgr)
48 return container_of(mgr, struct amdgpu_device, mman.vram_mgr);
52 * DOC: mem_info_vram_total
54 * The amdgpu driver provides a sysfs API for reporting current total VRAM
55 * available on the device
56 * The file mem_info_vram_total is used for this and returns the total
57 * amount of VRAM in bytes
59 static ssize_t amdgpu_mem_info_vram_total_show(struct device *dev,
60 struct device_attribute *attr, char *buf)
62 struct drm_device *ddev = dev_get_drvdata(dev);
63 struct amdgpu_device *adev = drm_to_adev(ddev);
65 return sysfs_emit(buf, "%llu\n", adev->gmc.real_vram_size);
69 * DOC: mem_info_vis_vram_total
71 * The amdgpu driver provides a sysfs API for reporting current total
72 * visible VRAM available on the device
73 * The file mem_info_vis_vram_total is used for this and returns the total
74 * amount of visible VRAM in bytes
76 static ssize_t amdgpu_mem_info_vis_vram_total_show(struct device *dev,
77 struct device_attribute *attr, char *buf)
79 struct drm_device *ddev = dev_get_drvdata(dev);
80 struct amdgpu_device *adev = drm_to_adev(ddev);
82 return sysfs_emit(buf, "%llu\n", adev->gmc.visible_vram_size);
86 * DOC: mem_info_vram_used
88 * The amdgpu driver provides a sysfs API for reporting current total VRAM
89 * available on the device
90 * The file mem_info_vram_used is used for this and returns the total
91 * amount of currently used VRAM in bytes
93 static ssize_t amdgpu_mem_info_vram_used_show(struct device *dev,
94 struct device_attribute *attr,
97 struct drm_device *ddev = dev_get_drvdata(dev);
98 struct amdgpu_device *adev = drm_to_adev(ddev);
99 struct ttm_resource_manager *man;
101 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
102 return sysfs_emit(buf, "%llu\n", amdgpu_vram_mgr_usage(man));
106 * DOC: mem_info_vis_vram_used
108 * The amdgpu driver provides a sysfs API for reporting current total of
110 * The file mem_info_vis_vram_used is used for this and returns the total
111 * amount of currently used visible VRAM in bytes
113 static ssize_t amdgpu_mem_info_vis_vram_used_show(struct device *dev,
114 struct device_attribute *attr,
117 struct drm_device *ddev = dev_get_drvdata(dev);
118 struct amdgpu_device *adev = drm_to_adev(ddev);
119 struct ttm_resource_manager *man;
121 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
122 return sysfs_emit(buf, "%llu\n", amdgpu_vram_mgr_vis_usage(man));
126 * DOC: mem_info_vram_vendor
128 * The amdgpu driver provides a sysfs API for reporting the vendor of the
130 * The file mem_info_vram_vendor is used for this and returns the name of the
133 static ssize_t amdgpu_mem_info_vram_vendor(struct device *dev,
134 struct device_attribute *attr,
137 struct drm_device *ddev = dev_get_drvdata(dev);
138 struct amdgpu_device *adev = drm_to_adev(ddev);
140 switch (adev->gmc.vram_vendor) {
142 return sysfs_emit(buf, "samsung\n");
144 return sysfs_emit(buf, "infineon\n");
146 return sysfs_emit(buf, "elpida\n");
148 return sysfs_emit(buf, "etron\n");
150 return sysfs_emit(buf, "nanya\n");
152 return sysfs_emit(buf, "hynix\n");
154 return sysfs_emit(buf, "mosel\n");
156 return sysfs_emit(buf, "winbond\n");
158 return sysfs_emit(buf, "esmt\n");
160 return sysfs_emit(buf, "micron\n");
162 return sysfs_emit(buf, "unknown\n");
166 static DEVICE_ATTR(mem_info_vram_total, S_IRUGO,
167 amdgpu_mem_info_vram_total_show, NULL);
168 static DEVICE_ATTR(mem_info_vis_vram_total, S_IRUGO,
169 amdgpu_mem_info_vis_vram_total_show,NULL);
170 static DEVICE_ATTR(mem_info_vram_used, S_IRUGO,
171 amdgpu_mem_info_vram_used_show, NULL);
172 static DEVICE_ATTR(mem_info_vis_vram_used, S_IRUGO,
173 amdgpu_mem_info_vis_vram_used_show, NULL);
174 static DEVICE_ATTR(mem_info_vram_vendor, S_IRUGO,
175 amdgpu_mem_info_vram_vendor, NULL);
177 static struct attribute *amdgpu_vram_mgr_attributes[] = {
178 &dev_attr_mem_info_vram_total.attr,
179 &dev_attr_mem_info_vis_vram_total.attr,
180 &dev_attr_mem_info_vram_used.attr,
181 &dev_attr_mem_info_vis_vram_used.attr,
182 &dev_attr_mem_info_vram_vendor.attr,
186 const struct attribute_group amdgpu_vram_mgr_attr_group = {
187 .attrs = amdgpu_vram_mgr_attributes
191 * amdgpu_vram_mgr_vis_size - Calculate visible node size
193 * @adev: amdgpu_device pointer
194 * @node: MM node structure
196 * Calculate how many bytes of the MM node are inside visible VRAM
198 static u64 amdgpu_vram_mgr_vis_size(struct amdgpu_device *adev,
199 struct drm_mm_node *node)
201 uint64_t start = node->start << PAGE_SHIFT;
202 uint64_t end = (node->size + node->start) << PAGE_SHIFT;
204 if (start >= adev->gmc.visible_vram_size)
207 return (end > adev->gmc.visible_vram_size ?
208 adev->gmc.visible_vram_size : end) - start;
212 * amdgpu_vram_mgr_bo_visible_size - CPU visible BO size
214 * @bo: &amdgpu_bo buffer object (must be in VRAM)
217 * How much of the given &amdgpu_bo buffer object lies in CPU visible VRAM.
219 u64 amdgpu_vram_mgr_bo_visible_size(struct amdgpu_bo *bo)
221 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
222 struct ttm_resource *res = bo->tbo.resource;
223 unsigned pages = res->num_pages;
224 struct drm_mm_node *mm;
227 if (amdgpu_gmc_vram_full_visible(&adev->gmc))
228 return amdgpu_bo_size(bo);
230 if (res->start >= adev->gmc.visible_vram_size >> PAGE_SHIFT)
233 mm = &container_of(res, struct ttm_range_mgr_node, base)->mm_nodes[0];
234 for (usage = 0; pages; pages -= mm->size, mm++)
235 usage += amdgpu_vram_mgr_vis_size(adev, mm);
240 /* Commit the reservation of VRAM pages */
241 static void amdgpu_vram_mgr_do_reserve(struct ttm_resource_manager *man)
243 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
244 struct amdgpu_device *adev = to_amdgpu_device(mgr);
245 struct drm_mm *mm = &mgr->mm;
246 struct amdgpu_vram_reservation *rsv, *temp;
249 list_for_each_entry_safe(rsv, temp, &mgr->reservations_pending, node) {
250 if (drm_mm_reserve_node(mm, &rsv->mm_node))
253 dev_dbg(adev->dev, "Reservation 0x%llx - %lld, Succeeded\n",
254 rsv->mm_node.start, rsv->mm_node.size);
256 vis_usage = amdgpu_vram_mgr_vis_size(adev, &rsv->mm_node);
257 atomic64_add(vis_usage, &mgr->vis_usage);
258 atomic64_add(rsv->mm_node.size << PAGE_SHIFT, &mgr->usage);
259 list_move(&rsv->node, &mgr->reserved_pages);
264 * amdgpu_vram_mgr_reserve_range - Reserve a range from VRAM
266 * @man: TTM memory type manager
267 * @start: start address of the range in VRAM
268 * @size: size of the range
270 * Reserve memory from start addess with the specified size in VRAM
272 int amdgpu_vram_mgr_reserve_range(struct ttm_resource_manager *man,
273 uint64_t start, uint64_t size)
275 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
276 struct amdgpu_vram_reservation *rsv;
278 rsv = kzalloc(sizeof(*rsv), GFP_KERNEL);
282 INIT_LIST_HEAD(&rsv->node);
283 rsv->mm_node.start = start >> PAGE_SHIFT;
284 rsv->mm_node.size = size >> PAGE_SHIFT;
286 spin_lock(&mgr->lock);
287 list_add_tail(&mgr->reservations_pending, &rsv->node);
288 amdgpu_vram_mgr_do_reserve(man);
289 spin_unlock(&mgr->lock);
295 * amdgpu_vram_mgr_query_page_status - query the reservation status
297 * @man: TTM memory type manager
298 * @start: start address of a page in VRAM
301 * -EBUSY: the page is still hold and in pending list
302 * 0: the page has been reserved
303 * -ENOENT: the input page is not a reservation
305 int amdgpu_vram_mgr_query_page_status(struct ttm_resource_manager *man,
308 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
309 struct amdgpu_vram_reservation *rsv;
312 spin_lock(&mgr->lock);
314 list_for_each_entry(rsv, &mgr->reservations_pending, node) {
315 if ((rsv->mm_node.start <= start) &&
316 (start < (rsv->mm_node.start + rsv->mm_node.size))) {
322 list_for_each_entry(rsv, &mgr->reserved_pages, node) {
323 if ((rsv->mm_node.start <= start) &&
324 (start < (rsv->mm_node.start + rsv->mm_node.size))) {
332 spin_unlock(&mgr->lock);
337 * amdgpu_vram_mgr_virt_start - update virtual start address
339 * @mem: ttm_resource to update
340 * @node: just allocated node
342 * Calculate a virtual BO start address to easily check if everything is CPU
345 static void amdgpu_vram_mgr_virt_start(struct ttm_resource *mem,
346 struct drm_mm_node *node)
350 start = node->start + node->size;
351 if (start > mem->num_pages)
352 start -= mem->num_pages;
355 mem->start = max(mem->start, start);
359 * amdgpu_vram_mgr_new - allocate new ranges
361 * @man: TTM memory type manager
362 * @tbo: TTM BO we need this range for
363 * @place: placement flags and restrictions
364 * @mem: the resulting mem object
366 * Allocate VRAM for the given BO.
368 static int amdgpu_vram_mgr_new(struct ttm_resource_manager *man,
369 struct ttm_buffer_object *tbo,
370 const struct ttm_place *place,
371 struct ttm_resource **res)
373 unsigned long lpfn, num_nodes, pages_per_node, pages_left, pages;
374 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
375 struct amdgpu_device *adev = to_amdgpu_device(mgr);
376 uint64_t vis_usage = 0, mem_bytes, max_bytes;
377 struct ttm_range_mgr_node *node;
378 struct drm_mm *mm = &mgr->mm;
379 enum drm_mm_insert_mode mode;
387 max_bytes = adev->gmc.mc_vram_size;
388 if (tbo->type != ttm_bo_type_kernel)
389 max_bytes -= AMDGPU_VM_RESERVED_VRAM;
391 /* bail out quickly if there's likely not enough VRAM for this BO */
392 mem_bytes = tbo->base.size;
393 if (atomic64_add_return(mem_bytes, &mgr->usage) > max_bytes) {
398 if (place->flags & TTM_PL_FLAG_CONTIGUOUS) {
399 pages_per_node = ~0ul;
402 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
403 pages_per_node = HPAGE_PMD_NR;
406 pages_per_node = 2UL << (20UL - PAGE_SHIFT);
408 pages_per_node = max_t(uint32_t, pages_per_node,
409 tbo->page_alignment);
410 num_nodes = DIV_ROUND_UP_ULL(PFN_UP(mem_bytes), pages_per_node);
413 node = kvmalloc(struct_size(node, mm_nodes, num_nodes),
414 GFP_KERNEL | __GFP_ZERO);
420 ttm_resource_init(tbo, place, &node->base);
422 mode = DRM_MM_INSERT_BEST;
423 if (place->flags & TTM_PL_FLAG_TOPDOWN)
424 mode = DRM_MM_INSERT_HIGH;
426 pages_left = node->base.num_pages;
428 /* Limit maximum size to 2GB due to SG table limitations */
429 pages = min(pages_left, 2UL << (30 - PAGE_SHIFT));
432 spin_lock(&mgr->lock);
434 uint32_t alignment = tbo->page_alignment;
436 if (pages >= pages_per_node)
437 alignment = pages_per_node;
439 r = drm_mm_insert_node_in_range(mm, &node->mm_nodes[i], pages,
440 alignment, 0, place->fpfn,
443 if (pages > pages_per_node) {
444 if (is_power_of_2(pages))
447 pages = rounddown_pow_of_two(pages);
453 vis_usage += amdgpu_vram_mgr_vis_size(adev, &node->mm_nodes[i]);
454 amdgpu_vram_mgr_virt_start(&node->base, &node->mm_nodes[i]);
458 if (pages > pages_left)
461 spin_unlock(&mgr->lock);
464 node->base.placement |= TTM_PL_FLAG_CONTIGUOUS;
466 atomic64_add(vis_usage, &mgr->vis_usage);
472 drm_mm_remove_node(&node->mm_nodes[i]);
473 spin_unlock(&mgr->lock);
477 atomic64_sub(mem_bytes, &mgr->usage);
482 * amdgpu_vram_mgr_del - free ranges
484 * @man: TTM memory type manager
485 * @mem: TTM memory object
487 * Free the allocated VRAM again.
489 static void amdgpu_vram_mgr_del(struct ttm_resource_manager *man,
490 struct ttm_resource *res)
492 struct ttm_range_mgr_node *node = to_ttm_range_mgr_node(res);
493 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
494 struct amdgpu_device *adev = to_amdgpu_device(mgr);
495 uint64_t usage = 0, vis_usage = 0;
498 spin_lock(&mgr->lock);
499 for (i = 0, pages = res->num_pages; pages;
500 pages -= node->mm_nodes[i].size, ++i) {
501 struct drm_mm_node *mm = &node->mm_nodes[i];
503 drm_mm_remove_node(mm);
504 usage += mm->size << PAGE_SHIFT;
505 vis_usage += amdgpu_vram_mgr_vis_size(adev, mm);
507 amdgpu_vram_mgr_do_reserve(man);
508 spin_unlock(&mgr->lock);
510 atomic64_sub(usage, &mgr->usage);
511 atomic64_sub(vis_usage, &mgr->vis_usage);
517 * amdgpu_vram_mgr_alloc_sgt - allocate and fill a sg table
519 * @adev: amdgpu device pointer
520 * @mem: TTM memory object
521 * @offset: byte offset from the base of VRAM BO
522 * @length: number of bytes to export in sg_table
523 * @dev: the other device
524 * @dir: dma direction
525 * @sgt: resulting sg table
527 * Allocate and fill a sg table from a VRAM allocation.
529 int amdgpu_vram_mgr_alloc_sgt(struct amdgpu_device *adev,
530 struct ttm_resource *res,
531 u64 offset, u64 length,
533 enum dma_data_direction dir,
534 struct sg_table **sgt)
536 struct amdgpu_res_cursor cursor;
537 struct scatterlist *sg;
541 *sgt = kmalloc(sizeof(**sgt), GFP_KERNEL);
545 /* Determine the number of DRM_MM nodes to export */
546 amdgpu_res_first(res, offset, length, &cursor);
547 while (cursor.remaining) {
549 amdgpu_res_next(&cursor, cursor.size);
552 r = sg_alloc_table(*sgt, num_entries, GFP_KERNEL);
556 /* Initialize scatterlist nodes of sg_table */
557 for_each_sgtable_sg((*sgt), sg, i)
561 * Walk down DRM_MM nodes to populate scatterlist nodes
562 * @note: Use iterator api to get first the DRM_MM node
563 * and the number of bytes from it. Access the following
564 * DRM_MM node(s) if more buffer needs to exported
566 amdgpu_res_first(res, offset, length, &cursor);
567 for_each_sgtable_sg((*sgt), sg, i) {
568 phys_addr_t phys = cursor.start + adev->gmc.aper_base;
569 size_t size = cursor.size;
572 addr = dma_map_resource(dev, phys, size, dir,
573 DMA_ATTR_SKIP_CPU_SYNC);
574 r = dma_mapping_error(dev, addr);
578 sg_set_page(sg, NULL, size, 0);
579 sg_dma_address(sg) = addr;
580 sg_dma_len(sg) = size;
582 amdgpu_res_next(&cursor, cursor.size);
588 for_each_sgtable_sg((*sgt), sg, i) {
592 dma_unmap_resource(dev, sg->dma_address,
594 DMA_ATTR_SKIP_CPU_SYNC);
604 * amdgpu_vram_mgr_free_sgt - allocate and fill a sg table
606 * @dev: device pointer
607 * @dir: data direction of resource to unmap
608 * @sgt: sg table to free
610 * Free a previously allocate sg table.
612 void amdgpu_vram_mgr_free_sgt(struct device *dev,
613 enum dma_data_direction dir,
614 struct sg_table *sgt)
616 struct scatterlist *sg;
619 for_each_sgtable_sg(sgt, sg, i)
620 dma_unmap_resource(dev, sg->dma_address,
622 DMA_ATTR_SKIP_CPU_SYNC);
628 * amdgpu_vram_mgr_usage - how many bytes are used in this domain
630 * @man: TTM memory type manager
632 * Returns how many bytes are used in this domain.
634 uint64_t amdgpu_vram_mgr_usage(struct ttm_resource_manager *man)
636 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
638 return atomic64_read(&mgr->usage);
642 * amdgpu_vram_mgr_vis_usage - how many bytes are used in the visible part
644 * @man: TTM memory type manager
646 * Returns how many bytes are used in the visible part of VRAM
648 uint64_t amdgpu_vram_mgr_vis_usage(struct ttm_resource_manager *man)
650 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
652 return atomic64_read(&mgr->vis_usage);
656 * amdgpu_vram_mgr_debug - dump VRAM table
658 * @man: TTM memory type manager
659 * @printer: DRM printer to use
661 * Dump the table content using printk.
663 static void amdgpu_vram_mgr_debug(struct ttm_resource_manager *man,
664 struct drm_printer *printer)
666 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
668 spin_lock(&mgr->lock);
669 drm_mm_print(&mgr->mm, printer);
670 spin_unlock(&mgr->lock);
672 drm_printf(printer, "man size:%llu pages, ram usage:%lluMB, vis usage:%lluMB\n",
673 man->size, amdgpu_vram_mgr_usage(man) >> 20,
674 amdgpu_vram_mgr_vis_usage(man) >> 20);
677 static const struct ttm_resource_manager_func amdgpu_vram_mgr_func = {
678 .alloc = amdgpu_vram_mgr_new,
679 .free = amdgpu_vram_mgr_del,
680 .debug = amdgpu_vram_mgr_debug
684 * amdgpu_vram_mgr_init - init VRAM manager and DRM MM
686 * @adev: amdgpu_device pointer
688 * Allocate and initialize the VRAM manager.
690 int amdgpu_vram_mgr_init(struct amdgpu_device *adev)
692 struct amdgpu_vram_mgr *mgr = &adev->mman.vram_mgr;
693 struct ttm_resource_manager *man = &mgr->manager;
695 ttm_resource_manager_init(man, adev->gmc.real_vram_size >> PAGE_SHIFT);
697 man->func = &amdgpu_vram_mgr_func;
699 drm_mm_init(&mgr->mm, 0, man->size);
700 spin_lock_init(&mgr->lock);
701 INIT_LIST_HEAD(&mgr->reservations_pending);
702 INIT_LIST_HEAD(&mgr->reserved_pages);
704 ttm_set_driver_manager(&adev->mman.bdev, TTM_PL_VRAM, &mgr->manager);
705 ttm_resource_manager_set_used(man, true);
710 * amdgpu_vram_mgr_fini - free and destroy VRAM manager
712 * @adev: amdgpu_device pointer
714 * Destroy and free the VRAM manager, returns -EBUSY if ranges are still
715 * allocated inside it.
717 void amdgpu_vram_mgr_fini(struct amdgpu_device *adev)
719 struct amdgpu_vram_mgr *mgr = &adev->mman.vram_mgr;
720 struct ttm_resource_manager *man = &mgr->manager;
722 struct amdgpu_vram_reservation *rsv, *temp;
724 ttm_resource_manager_set_used(man, false);
726 ret = ttm_resource_manager_evict_all(&adev->mman.bdev, man);
730 spin_lock(&mgr->lock);
731 list_for_each_entry_safe(rsv, temp, &mgr->reservations_pending, node)
734 list_for_each_entry_safe(rsv, temp, &mgr->reserved_pages, node) {
735 drm_mm_remove_node(&rsv->mm_node);
738 drm_mm_takedown(&mgr->mm);
739 spin_unlock(&mgr->lock);
741 ttm_resource_manager_cleanup(man);
742 ttm_set_driver_manager(&adev->mman.bdev, TTM_PL_VRAM, NULL);