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>
27 #include "amdgpu_vm.h"
28 #include "amdgpu_res_cursor.h"
29 #include "amdgpu_atomfirmware.h"
32 struct amdgpu_vram_reservation {
33 struct list_head node;
34 struct drm_mm_node mm_node;
37 static inline struct amdgpu_vram_mgr *
38 to_vram_mgr(struct ttm_resource_manager *man)
40 return container_of(man, struct amdgpu_vram_mgr, manager);
43 static inline struct amdgpu_device *
44 to_amdgpu_device(struct amdgpu_vram_mgr *mgr)
46 return container_of(mgr, struct amdgpu_device, mman.vram_mgr);
50 * DOC: mem_info_vram_total
52 * The amdgpu driver provides a sysfs API for reporting current total VRAM
53 * available on the device
54 * The file mem_info_vram_total is used for this and returns the total
55 * amount of VRAM in bytes
57 static ssize_t amdgpu_mem_info_vram_total_show(struct device *dev,
58 struct device_attribute *attr, char *buf)
60 struct drm_device *ddev = dev_get_drvdata(dev);
61 struct amdgpu_device *adev = drm_to_adev(ddev);
63 return sysfs_emit(buf, "%llu\n", adev->gmc.real_vram_size);
67 * DOC: mem_info_vis_vram_total
69 * The amdgpu driver provides a sysfs API for reporting current total
70 * visible VRAM available on the device
71 * The file mem_info_vis_vram_total is used for this and returns the total
72 * amount of visible VRAM in bytes
74 static ssize_t amdgpu_mem_info_vis_vram_total_show(struct device *dev,
75 struct device_attribute *attr, char *buf)
77 struct drm_device *ddev = dev_get_drvdata(dev);
78 struct amdgpu_device *adev = drm_to_adev(ddev);
80 return sysfs_emit(buf, "%llu\n", adev->gmc.visible_vram_size);
84 * DOC: mem_info_vram_used
86 * The amdgpu driver provides a sysfs API for reporting current total VRAM
87 * available on the device
88 * The file mem_info_vram_used is used for this and returns the total
89 * amount of currently used VRAM in bytes
91 static ssize_t amdgpu_mem_info_vram_used_show(struct device *dev,
92 struct device_attribute *attr,
95 struct drm_device *ddev = dev_get_drvdata(dev);
96 struct amdgpu_device *adev = drm_to_adev(ddev);
97 struct ttm_resource_manager *man;
99 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
100 return sysfs_emit(buf, "%llu\n", amdgpu_vram_mgr_usage(man));
104 * DOC: mem_info_vis_vram_used
106 * The amdgpu driver provides a sysfs API for reporting current total of
108 * The file mem_info_vis_vram_used is used for this and returns the total
109 * amount of currently used visible VRAM in bytes
111 static ssize_t amdgpu_mem_info_vis_vram_used_show(struct device *dev,
112 struct device_attribute *attr,
115 struct drm_device *ddev = dev_get_drvdata(dev);
116 struct amdgpu_device *adev = drm_to_adev(ddev);
117 struct ttm_resource_manager *man;
119 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
120 return sysfs_emit(buf, "%llu\n", amdgpu_vram_mgr_vis_usage(man));
124 * DOC: mem_info_vram_vendor
126 * The amdgpu driver provides a sysfs API for reporting the vendor of the
128 * The file mem_info_vram_vendor is used for this and returns the name of the
131 static ssize_t amdgpu_mem_info_vram_vendor(struct device *dev,
132 struct device_attribute *attr,
135 struct drm_device *ddev = dev_get_drvdata(dev);
136 struct amdgpu_device *adev = drm_to_adev(ddev);
138 switch (adev->gmc.vram_vendor) {
140 return sysfs_emit(buf, "samsung\n");
142 return sysfs_emit(buf, "infineon\n");
144 return sysfs_emit(buf, "elpida\n");
146 return sysfs_emit(buf, "etron\n");
148 return sysfs_emit(buf, "nanya\n");
150 return sysfs_emit(buf, "hynix\n");
152 return sysfs_emit(buf, "mosel\n");
154 return sysfs_emit(buf, "winbond\n");
156 return sysfs_emit(buf, "esmt\n");
158 return sysfs_emit(buf, "micron\n");
160 return sysfs_emit(buf, "unknown\n");
164 static DEVICE_ATTR(mem_info_vram_total, S_IRUGO,
165 amdgpu_mem_info_vram_total_show, NULL);
166 static DEVICE_ATTR(mem_info_vis_vram_total, S_IRUGO,
167 amdgpu_mem_info_vis_vram_total_show,NULL);
168 static DEVICE_ATTR(mem_info_vram_used, S_IRUGO,
169 amdgpu_mem_info_vram_used_show, NULL);
170 static DEVICE_ATTR(mem_info_vis_vram_used, S_IRUGO,
171 amdgpu_mem_info_vis_vram_used_show, NULL);
172 static DEVICE_ATTR(mem_info_vram_vendor, S_IRUGO,
173 amdgpu_mem_info_vram_vendor, NULL);
175 static struct attribute *amdgpu_vram_mgr_attributes[] = {
176 &dev_attr_mem_info_vram_total.attr,
177 &dev_attr_mem_info_vis_vram_total.attr,
178 &dev_attr_mem_info_vram_used.attr,
179 &dev_attr_mem_info_vis_vram_used.attr,
180 &dev_attr_mem_info_vram_vendor.attr,
184 const struct attribute_group amdgpu_vram_mgr_attr_group = {
185 .attrs = amdgpu_vram_mgr_attributes
189 * amdgpu_vram_mgr_vis_size - Calculate visible node size
191 * @adev: amdgpu_device pointer
192 * @node: MM node structure
194 * Calculate how many bytes of the MM node are inside visible VRAM
196 static u64 amdgpu_vram_mgr_vis_size(struct amdgpu_device *adev,
197 struct drm_mm_node *node)
199 uint64_t start = node->start << PAGE_SHIFT;
200 uint64_t end = (node->size + node->start) << PAGE_SHIFT;
202 if (start >= adev->gmc.visible_vram_size)
205 return (end > adev->gmc.visible_vram_size ?
206 adev->gmc.visible_vram_size : end) - start;
210 * amdgpu_vram_mgr_bo_visible_size - CPU visible BO size
212 * @bo: &amdgpu_bo buffer object (must be in VRAM)
215 * How much of the given &amdgpu_bo buffer object lies in CPU visible VRAM.
217 u64 amdgpu_vram_mgr_bo_visible_size(struct amdgpu_bo *bo)
219 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
220 struct ttm_resource *mem = &bo->tbo.mem;
221 struct drm_mm_node *nodes = mem->mm_node;
222 unsigned pages = mem->num_pages;
225 if (amdgpu_gmc_vram_full_visible(&adev->gmc))
226 return amdgpu_bo_size(bo);
228 if (mem->start >= adev->gmc.visible_vram_size >> PAGE_SHIFT)
231 for (usage = 0; nodes && pages; pages -= nodes->size, nodes++)
232 usage += amdgpu_vram_mgr_vis_size(adev, nodes);
237 /* Commit the reservation of VRAM pages */
238 static void amdgpu_vram_mgr_do_reserve(struct ttm_resource_manager *man)
240 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
241 struct amdgpu_device *adev = to_amdgpu_device(mgr);
242 struct drm_mm *mm = &mgr->mm;
243 struct amdgpu_vram_reservation *rsv, *temp;
246 list_for_each_entry_safe(rsv, temp, &mgr->reservations_pending, node) {
247 if (drm_mm_reserve_node(mm, &rsv->mm_node))
250 dev_dbg(adev->dev, "Reservation 0x%llx - %lld, Succeeded\n",
251 rsv->mm_node.start, rsv->mm_node.size);
253 vis_usage = amdgpu_vram_mgr_vis_size(adev, &rsv->mm_node);
254 atomic64_add(vis_usage, &mgr->vis_usage);
255 atomic64_add(rsv->mm_node.size << PAGE_SHIFT, &mgr->usage);
256 list_move(&rsv->node, &mgr->reserved_pages);
261 * amdgpu_vram_mgr_reserve_range - Reserve a range from VRAM
263 * @man: TTM memory type manager
264 * @start: start address of the range in VRAM
265 * @size: size of the range
267 * Reserve memory from start addess with the specified size in VRAM
269 int amdgpu_vram_mgr_reserve_range(struct ttm_resource_manager *man,
270 uint64_t start, uint64_t size)
272 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
273 struct amdgpu_vram_reservation *rsv;
275 rsv = kzalloc(sizeof(*rsv), GFP_KERNEL);
279 INIT_LIST_HEAD(&rsv->node);
280 rsv->mm_node.start = start >> PAGE_SHIFT;
281 rsv->mm_node.size = size >> PAGE_SHIFT;
283 spin_lock(&mgr->lock);
284 list_add_tail(&mgr->reservations_pending, &rsv->node);
285 amdgpu_vram_mgr_do_reserve(man);
286 spin_unlock(&mgr->lock);
292 * amdgpu_vram_mgr_query_page_status - query the reservation status
294 * @man: TTM memory type manager
295 * @start: start address of a page in VRAM
298 * -EBUSY: the page is still hold and in pending list
299 * 0: the page has been reserved
300 * -ENOENT: the input page is not a reservation
302 int amdgpu_vram_mgr_query_page_status(struct ttm_resource_manager *man,
305 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
306 struct amdgpu_vram_reservation *rsv;
309 spin_lock(&mgr->lock);
311 list_for_each_entry(rsv, &mgr->reservations_pending, node) {
312 if ((rsv->mm_node.start <= start) &&
313 (start < (rsv->mm_node.start + rsv->mm_node.size))) {
319 list_for_each_entry(rsv, &mgr->reserved_pages, node) {
320 if ((rsv->mm_node.start <= start) &&
321 (start < (rsv->mm_node.start + rsv->mm_node.size))) {
329 spin_unlock(&mgr->lock);
334 * amdgpu_vram_mgr_virt_start - update virtual start address
336 * @mem: ttm_resource to update
337 * @node: just allocated node
339 * Calculate a virtual BO start address to easily check if everything is CPU
342 static void amdgpu_vram_mgr_virt_start(struct ttm_resource *mem,
343 struct drm_mm_node *node)
347 start = node->start + node->size;
348 if (start > mem->num_pages)
349 start -= mem->num_pages;
352 mem->start = max(mem->start, start);
356 * amdgpu_vram_mgr_new - allocate new ranges
358 * @man: TTM memory type manager
359 * @tbo: TTM BO we need this range for
360 * @place: placement flags and restrictions
361 * @mem: the resulting mem object
363 * Allocate VRAM for the given BO.
365 static int amdgpu_vram_mgr_new(struct ttm_resource_manager *man,
366 struct ttm_buffer_object *tbo,
367 const struct ttm_place *place,
368 struct ttm_resource *mem)
370 unsigned long lpfn, num_nodes, pages_per_node, pages_left, pages;
371 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
372 struct amdgpu_device *adev = to_amdgpu_device(mgr);
373 uint64_t vis_usage = 0, mem_bytes, max_bytes;
374 struct drm_mm *mm = &mgr->mm;
375 enum drm_mm_insert_mode mode;
376 struct drm_mm_node *nodes;
384 max_bytes = adev->gmc.mc_vram_size;
385 if (tbo->type != ttm_bo_type_kernel)
386 max_bytes -= AMDGPU_VM_RESERVED_VRAM;
388 /* bail out quickly if there's likely not enough VRAM for this BO */
389 mem_bytes = (u64)mem->num_pages << PAGE_SHIFT;
390 if (atomic64_add_return(mem_bytes, &mgr->usage) > max_bytes) {
391 atomic64_sub(mem_bytes, &mgr->usage);
395 if (place->flags & TTM_PL_FLAG_CONTIGUOUS) {
396 pages_per_node = ~0ul;
399 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
400 pages_per_node = HPAGE_PMD_NR;
403 pages_per_node = 2UL << (20UL - PAGE_SHIFT);
405 pages_per_node = max_t(uint32_t, pages_per_node,
406 tbo->page_alignment);
407 num_nodes = DIV_ROUND_UP(mem->num_pages, pages_per_node);
410 nodes = kvmalloc_array((uint32_t)num_nodes, sizeof(*nodes),
411 GFP_KERNEL | __GFP_ZERO);
413 atomic64_sub(mem_bytes, &mgr->usage);
417 mode = DRM_MM_INSERT_BEST;
418 if (place->flags & TTM_PL_FLAG_TOPDOWN)
419 mode = DRM_MM_INSERT_HIGH;
422 pages_left = mem->num_pages;
424 /* Limit maximum size to 2GB due to SG table limitations */
425 pages = min(pages_left, 2UL << (30 - PAGE_SHIFT));
428 spin_lock(&mgr->lock);
430 uint32_t alignment = tbo->page_alignment;
432 if (pages >= pages_per_node)
433 alignment = pages_per_node;
435 r = drm_mm_insert_node_in_range(mm, &nodes[i], pages, alignment,
436 0, place->fpfn, lpfn, mode);
438 if (pages > pages_per_node) {
439 if (is_power_of_2(pages))
442 pages = rounddown_pow_of_two(pages);
448 vis_usage += amdgpu_vram_mgr_vis_size(adev, &nodes[i]);
449 amdgpu_vram_mgr_virt_start(mem, &nodes[i]);
453 if (pages > pages_left)
456 spin_unlock(&mgr->lock);
459 mem->placement |= TTM_PL_FLAG_CONTIGUOUS;
461 atomic64_add(vis_usage, &mgr->vis_usage);
462 mem->mm_node = nodes;
467 drm_mm_remove_node(&nodes[i]);
468 spin_unlock(&mgr->lock);
469 atomic64_sub(mem->num_pages << PAGE_SHIFT, &mgr->usage);
476 * amdgpu_vram_mgr_del - free ranges
478 * @man: TTM memory type manager
479 * @mem: TTM memory object
481 * Free the allocated VRAM again.
483 static void amdgpu_vram_mgr_del(struct ttm_resource_manager *man,
484 struct ttm_resource *mem)
486 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
487 struct amdgpu_device *adev = to_amdgpu_device(mgr);
488 struct drm_mm_node *nodes = mem->mm_node;
489 uint64_t usage = 0, vis_usage = 0;
490 unsigned pages = mem->num_pages;
495 spin_lock(&mgr->lock);
497 pages -= nodes->size;
498 drm_mm_remove_node(nodes);
499 usage += nodes->size << PAGE_SHIFT;
500 vis_usage += amdgpu_vram_mgr_vis_size(adev, nodes);
503 amdgpu_vram_mgr_do_reserve(man);
504 spin_unlock(&mgr->lock);
506 atomic64_sub(usage, &mgr->usage);
507 atomic64_sub(vis_usage, &mgr->vis_usage);
509 kvfree(mem->mm_node);
514 * amdgpu_vram_mgr_alloc_sgt - allocate and fill a sg table
516 * @adev: amdgpu device pointer
517 * @mem: TTM memory object
518 * @offset: byte offset from the base of VRAM BO
519 * @length: number of bytes to export in sg_table
520 * @dev: the other device
521 * @dir: dma direction
522 * @sgt: resulting sg table
524 * Allocate and fill a sg table from a VRAM allocation.
526 int amdgpu_vram_mgr_alloc_sgt(struct amdgpu_device *adev,
527 struct ttm_resource *mem,
528 u64 offset, u64 length,
530 enum dma_data_direction dir,
531 struct sg_table **sgt)
533 struct amdgpu_res_cursor cursor;
534 struct scatterlist *sg;
538 *sgt = kmalloc(sizeof(**sgt), GFP_KERNEL);
542 /* Determine the number of DRM_MM nodes to export */
543 amdgpu_res_first(mem, offset, length, &cursor);
544 while (cursor.remaining) {
546 amdgpu_res_next(&cursor, cursor.size);
549 r = sg_alloc_table(*sgt, num_entries, GFP_KERNEL);
553 /* Initialize scatterlist nodes of sg_table */
554 for_each_sgtable_sg((*sgt), sg, i)
558 * Walk down DRM_MM nodes to populate scatterlist nodes
559 * @note: Use iterator api to get first the DRM_MM node
560 * and the number of bytes from it. Access the following
561 * DRM_MM node(s) if more buffer needs to exported
563 amdgpu_res_first(mem, offset, length, &cursor);
564 for_each_sgtable_sg((*sgt), sg, i) {
565 phys_addr_t phys = cursor.start + adev->gmc.aper_base;
566 size_t size = cursor.size;
569 addr = dma_map_resource(dev, phys, size, dir,
570 DMA_ATTR_SKIP_CPU_SYNC);
571 r = dma_mapping_error(dev, addr);
575 sg_set_page(sg, NULL, size, 0);
576 sg_dma_address(sg) = addr;
577 sg_dma_len(sg) = size;
579 amdgpu_res_next(&cursor, cursor.size);
585 for_each_sgtable_sg((*sgt), sg, i) {
589 dma_unmap_resource(dev, sg->dma_address,
591 DMA_ATTR_SKIP_CPU_SYNC);
601 * amdgpu_vram_mgr_free_sgt - allocate and fill a sg table
603 * @dev: device pointer
604 * @dir: data direction of resource to unmap
605 * @sgt: sg table to free
607 * Free a previously allocate sg table.
609 void amdgpu_vram_mgr_free_sgt(struct device *dev,
610 enum dma_data_direction dir,
611 struct sg_table *sgt)
613 struct scatterlist *sg;
616 for_each_sgtable_sg(sgt, sg, i)
617 dma_unmap_resource(dev, sg->dma_address,
619 DMA_ATTR_SKIP_CPU_SYNC);
625 * amdgpu_vram_mgr_usage - how many bytes are used in this domain
627 * @man: TTM memory type manager
629 * Returns how many bytes are used in this domain.
631 uint64_t amdgpu_vram_mgr_usage(struct ttm_resource_manager *man)
633 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
635 return atomic64_read(&mgr->usage);
639 * amdgpu_vram_mgr_vis_usage - how many bytes are used in the visible part
641 * @man: TTM memory type manager
643 * Returns how many bytes are used in the visible part of VRAM
645 uint64_t amdgpu_vram_mgr_vis_usage(struct ttm_resource_manager *man)
647 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
649 return atomic64_read(&mgr->vis_usage);
653 * amdgpu_vram_mgr_debug - dump VRAM table
655 * @man: TTM memory type manager
656 * @printer: DRM printer to use
658 * Dump the table content using printk.
660 static void amdgpu_vram_mgr_debug(struct ttm_resource_manager *man,
661 struct drm_printer *printer)
663 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
665 spin_lock(&mgr->lock);
666 drm_mm_print(&mgr->mm, printer);
667 spin_unlock(&mgr->lock);
669 drm_printf(printer, "man size:%llu pages, ram usage:%lluMB, vis usage:%lluMB\n",
670 man->size, amdgpu_vram_mgr_usage(man) >> 20,
671 amdgpu_vram_mgr_vis_usage(man) >> 20);
674 static const struct ttm_resource_manager_func amdgpu_vram_mgr_func = {
675 .alloc = amdgpu_vram_mgr_new,
676 .free = amdgpu_vram_mgr_del,
677 .debug = amdgpu_vram_mgr_debug
681 * amdgpu_vram_mgr_init - init VRAM manager and DRM MM
683 * @adev: amdgpu_device pointer
685 * Allocate and initialize the VRAM manager.
687 int amdgpu_vram_mgr_init(struct amdgpu_device *adev)
689 struct amdgpu_vram_mgr *mgr = &adev->mman.vram_mgr;
690 struct ttm_resource_manager *man = &mgr->manager;
692 ttm_resource_manager_init(man, adev->gmc.real_vram_size >> PAGE_SHIFT);
694 man->func = &amdgpu_vram_mgr_func;
696 drm_mm_init(&mgr->mm, 0, man->size);
697 spin_lock_init(&mgr->lock);
698 INIT_LIST_HEAD(&mgr->reservations_pending);
699 INIT_LIST_HEAD(&mgr->reserved_pages);
701 ttm_set_driver_manager(&adev->mman.bdev, TTM_PL_VRAM, &mgr->manager);
702 ttm_resource_manager_set_used(man, true);
707 * amdgpu_vram_mgr_fini - free and destroy VRAM manager
709 * @adev: amdgpu_device pointer
711 * Destroy and free the VRAM manager, returns -EBUSY if ranges are still
712 * allocated inside it.
714 void amdgpu_vram_mgr_fini(struct amdgpu_device *adev)
716 struct amdgpu_vram_mgr *mgr = &adev->mman.vram_mgr;
717 struct ttm_resource_manager *man = &mgr->manager;
719 struct amdgpu_vram_reservation *rsv, *temp;
721 ttm_resource_manager_set_used(man, false);
723 ret = ttm_resource_manager_evict_all(&adev->mman.bdev, man);
727 spin_lock(&mgr->lock);
728 list_for_each_entry_safe(rsv, temp, &mgr->reservations_pending, node)
731 list_for_each_entry_safe(rsv, temp, &mgr->reserved_pages, node) {
732 drm_mm_remove_node(&rsv->mm_node);
735 drm_mm_takedown(&mgr->mm);
736 spin_unlock(&mgr->lock);
738 ttm_resource_manager_cleanup(man);
739 ttm_set_driver_manager(&adev->mman.bdev, TTM_PL_VRAM, NULL);