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_atomfirmware.h"
31 static inline struct amdgpu_vram_mgr *to_vram_mgr(struct ttm_resource_manager *man)
33 return container_of(man, struct amdgpu_vram_mgr, manager);
36 static inline struct amdgpu_device *to_amdgpu_device(struct amdgpu_vram_mgr *mgr)
38 return container_of(mgr, struct amdgpu_device, mman.vram_mgr);
42 * DOC: mem_info_vram_total
44 * The amdgpu driver provides a sysfs API for reporting current total VRAM
45 * available on the device
46 * The file mem_info_vram_total is used for this and returns the total
47 * amount of VRAM in bytes
49 static ssize_t amdgpu_mem_info_vram_total_show(struct device *dev,
50 struct device_attribute *attr, char *buf)
52 struct drm_device *ddev = dev_get_drvdata(dev);
53 struct amdgpu_device *adev = drm_to_adev(ddev);
55 return snprintf(buf, PAGE_SIZE, "%llu\n", adev->gmc.real_vram_size);
59 * DOC: mem_info_vis_vram_total
61 * The amdgpu driver provides a sysfs API for reporting current total
62 * visible VRAM available on the device
63 * The file mem_info_vis_vram_total is used for this and returns the total
64 * amount of visible VRAM in bytes
66 static ssize_t amdgpu_mem_info_vis_vram_total_show(struct device *dev,
67 struct device_attribute *attr, char *buf)
69 struct drm_device *ddev = dev_get_drvdata(dev);
70 struct amdgpu_device *adev = drm_to_adev(ddev);
72 return snprintf(buf, PAGE_SIZE, "%llu\n", adev->gmc.visible_vram_size);
76 * DOC: mem_info_vram_used
78 * The amdgpu driver provides a sysfs API for reporting current total VRAM
79 * available on the device
80 * The file mem_info_vram_used is used for this and returns the total
81 * amount of currently used VRAM in bytes
83 static ssize_t amdgpu_mem_info_vram_used_show(struct device *dev,
84 struct device_attribute *attr, char *buf)
86 struct drm_device *ddev = dev_get_drvdata(dev);
87 struct amdgpu_device *adev = drm_to_adev(ddev);
88 struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
90 return snprintf(buf, PAGE_SIZE, "%llu\n",
91 amdgpu_vram_mgr_usage(man));
95 * DOC: mem_info_vis_vram_used
97 * The amdgpu driver provides a sysfs API for reporting current total of
99 * The file mem_info_vis_vram_used is used for this and returns the total
100 * amount of currently used visible VRAM in bytes
102 static ssize_t amdgpu_mem_info_vis_vram_used_show(struct device *dev,
103 struct device_attribute *attr, char *buf)
105 struct drm_device *ddev = dev_get_drvdata(dev);
106 struct amdgpu_device *adev = drm_to_adev(ddev);
107 struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
109 return snprintf(buf, PAGE_SIZE, "%llu\n",
110 amdgpu_vram_mgr_vis_usage(man));
113 static ssize_t amdgpu_mem_info_vram_vendor(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);
120 switch (adev->gmc.vram_vendor) {
122 return snprintf(buf, PAGE_SIZE, "samsung\n");
124 return snprintf(buf, PAGE_SIZE, "infineon\n");
126 return snprintf(buf, PAGE_SIZE, "elpida\n");
128 return snprintf(buf, PAGE_SIZE, "etron\n");
130 return snprintf(buf, PAGE_SIZE, "nanya\n");
132 return snprintf(buf, PAGE_SIZE, "hynix\n");
134 return snprintf(buf, PAGE_SIZE, "mosel\n");
136 return snprintf(buf, PAGE_SIZE, "winbond\n");
138 return snprintf(buf, PAGE_SIZE, "esmt\n");
140 return snprintf(buf, PAGE_SIZE, "micron\n");
142 return snprintf(buf, PAGE_SIZE, "unknown\n");
146 static DEVICE_ATTR(mem_info_vram_total, S_IRUGO,
147 amdgpu_mem_info_vram_total_show, NULL);
148 static DEVICE_ATTR(mem_info_vis_vram_total, S_IRUGO,
149 amdgpu_mem_info_vis_vram_total_show,NULL);
150 static DEVICE_ATTR(mem_info_vram_used, S_IRUGO,
151 amdgpu_mem_info_vram_used_show, NULL);
152 static DEVICE_ATTR(mem_info_vis_vram_used, S_IRUGO,
153 amdgpu_mem_info_vis_vram_used_show, NULL);
154 static DEVICE_ATTR(mem_info_vram_vendor, S_IRUGO,
155 amdgpu_mem_info_vram_vendor, NULL);
157 static const struct attribute *amdgpu_vram_mgr_attributes[] = {
158 &dev_attr_mem_info_vram_total.attr,
159 &dev_attr_mem_info_vis_vram_total.attr,
160 &dev_attr_mem_info_vram_used.attr,
161 &dev_attr_mem_info_vis_vram_used.attr,
162 &dev_attr_mem_info_vram_vendor.attr,
166 static const struct ttm_resource_manager_func amdgpu_vram_mgr_func;
169 * amdgpu_vram_mgr_init - init VRAM manager and DRM MM
171 * @adev: amdgpu_device pointer
173 * Allocate and initialize the VRAM manager.
175 int amdgpu_vram_mgr_init(struct amdgpu_device *adev)
177 struct amdgpu_vram_mgr *mgr = &adev->mman.vram_mgr;
178 struct ttm_resource_manager *man = &mgr->manager;
181 ttm_resource_manager_init(man, adev->gmc.real_vram_size >> PAGE_SHIFT);
183 man->func = &amdgpu_vram_mgr_func;
185 drm_mm_init(&mgr->mm, 0, man->size);
186 spin_lock_init(&mgr->lock);
187 INIT_LIST_HEAD(&mgr->reservations_pending);
188 INIT_LIST_HEAD(&mgr->reserved_pages);
190 /* Add the two VRAM-related sysfs files */
191 ret = sysfs_create_files(&adev->dev->kobj, amdgpu_vram_mgr_attributes);
193 DRM_ERROR("Failed to register sysfs\n");
195 ttm_set_driver_manager(&adev->mman.bdev, TTM_PL_VRAM, &mgr->manager);
196 ttm_resource_manager_set_used(man, true);
201 * amdgpu_vram_mgr_fini - free and destroy VRAM manager
203 * @adev: amdgpu_device pointer
205 * Destroy and free the VRAM manager, returns -EBUSY if ranges are still
206 * allocated inside it.
208 void amdgpu_vram_mgr_fini(struct amdgpu_device *adev)
210 struct amdgpu_vram_mgr *mgr = &adev->mman.vram_mgr;
211 struct ttm_resource_manager *man = &mgr->manager;
213 struct amdgpu_vram_reservation *rsv, *temp;
215 ttm_resource_manager_set_used(man, false);
217 ret = ttm_resource_manager_evict_all(&adev->mman.bdev, man);
221 spin_lock(&mgr->lock);
222 list_for_each_entry_safe(rsv, temp, &mgr->reservations_pending, node)
225 list_for_each_entry_safe(rsv, temp, &mgr->reserved_pages, node) {
226 drm_mm_remove_node(&rsv->mm_node);
229 drm_mm_takedown(&mgr->mm);
230 spin_unlock(&mgr->lock);
232 sysfs_remove_files(&adev->dev->kobj, amdgpu_vram_mgr_attributes);
234 ttm_resource_manager_cleanup(man);
235 ttm_set_driver_manager(&adev->mman.bdev, TTM_PL_VRAM, NULL);
239 * amdgpu_vram_mgr_vis_size - Calculate visible node size
241 * @adev: amdgpu_device pointer
242 * @node: MM node structure
244 * Calculate how many bytes of the MM node are inside visible VRAM
246 static u64 amdgpu_vram_mgr_vis_size(struct amdgpu_device *adev,
247 struct drm_mm_node *node)
249 uint64_t start = node->start << PAGE_SHIFT;
250 uint64_t end = (node->size + node->start) << PAGE_SHIFT;
252 if (start >= adev->gmc.visible_vram_size)
255 return (end > adev->gmc.visible_vram_size ?
256 adev->gmc.visible_vram_size : end) - start;
260 * amdgpu_vram_mgr_bo_visible_size - CPU visible BO size
262 * @bo: &amdgpu_bo buffer object (must be in VRAM)
265 * How much of the given &amdgpu_bo buffer object lies in CPU visible VRAM.
267 u64 amdgpu_vram_mgr_bo_visible_size(struct amdgpu_bo *bo)
269 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
270 struct ttm_resource *mem = &bo->tbo.mem;
271 struct drm_mm_node *nodes = mem->mm_node;
272 unsigned pages = mem->num_pages;
275 if (amdgpu_gmc_vram_full_visible(&adev->gmc))
276 return amdgpu_bo_size(bo);
278 if (mem->start >= adev->gmc.visible_vram_size >> PAGE_SHIFT)
281 for (usage = 0; nodes && pages; pages -= nodes->size, nodes++)
282 usage += amdgpu_vram_mgr_vis_size(adev, nodes);
287 static void amdgpu_vram_mgr_do_reserve(struct ttm_resource_manager *man)
289 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
290 struct amdgpu_device *adev = to_amdgpu_device(mgr);
291 struct drm_mm *mm = &mgr->mm;
292 struct amdgpu_vram_reservation *rsv, *temp;
295 list_for_each_entry_safe(rsv, temp, &mgr->reservations_pending, node) {
296 if (drm_mm_reserve_node(mm, &rsv->mm_node))
299 dev_dbg(adev->dev, "Reservation 0x%llx - %lld, Succeeded\n",
300 rsv->mm_node.start, rsv->mm_node.size);
302 vis_usage = amdgpu_vram_mgr_vis_size(adev, &rsv->mm_node);
303 atomic64_add(vis_usage, &mgr->vis_usage);
304 atomic64_add(rsv->mm_node.size << PAGE_SHIFT, &mgr->usage);
305 list_move(&rsv->node, &mgr->reserved_pages);
310 * amdgpu_vram_mgr_reserve_range - Reserve a range from VRAM
312 * @man: TTM memory type manager
313 * @start: start address of the range in VRAM
314 * @size: size of the range
316 * Reserve memory from start addess with the specified size in VRAM
318 int amdgpu_vram_mgr_reserve_range(struct ttm_resource_manager *man,
319 uint64_t start, uint64_t size)
321 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
322 struct amdgpu_vram_reservation *rsv;
324 rsv = kzalloc(sizeof(*rsv), GFP_KERNEL);
328 INIT_LIST_HEAD(&rsv->node);
329 rsv->mm_node.start = start >> PAGE_SHIFT;
330 rsv->mm_node.size = size >> PAGE_SHIFT;
332 spin_lock(&mgr->lock);
333 list_add_tail(&mgr->reservations_pending, &rsv->node);
334 amdgpu_vram_mgr_do_reserve(man);
335 spin_unlock(&mgr->lock);
341 * amdgpu_vram_mgr_query_page_status - query the reservation status
343 * @man: TTM memory type manager
344 * @start: start address of a page in VRAM
347 * -EBUSY: the page is still hold and in pending list
348 * 0: the page has been reserved
349 * -ENOENT: the input page is not a reservation
351 int amdgpu_vram_mgr_query_page_status(struct ttm_resource_manager *man,
354 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
355 struct amdgpu_vram_reservation *rsv;
358 spin_lock(&mgr->lock);
360 list_for_each_entry(rsv, &mgr->reservations_pending, node) {
361 if ((rsv->mm_node.start <= start) &&
362 (start < (rsv->mm_node.start + rsv->mm_node.size))) {
368 list_for_each_entry(rsv, &mgr->reserved_pages, node) {
369 if ((rsv->mm_node.start <= start) &&
370 (start < (rsv->mm_node.start + rsv->mm_node.size))) {
378 spin_unlock(&mgr->lock);
383 * amdgpu_vram_mgr_virt_start - update virtual start address
385 * @mem: ttm_resource to update
386 * @node: just allocated node
388 * Calculate a virtual BO start address to easily check if everything is CPU
391 static void amdgpu_vram_mgr_virt_start(struct ttm_resource *mem,
392 struct drm_mm_node *node)
396 start = node->start + node->size;
397 if (start > mem->num_pages)
398 start -= mem->num_pages;
401 mem->start = max(mem->start, start);
405 * amdgpu_vram_mgr_new - allocate new ranges
407 * @man: TTM memory type manager
408 * @tbo: TTM BO we need this range for
409 * @place: placement flags and restrictions
410 * @mem: the resulting mem object
412 * Allocate VRAM for the given BO.
414 static int amdgpu_vram_mgr_new(struct ttm_resource_manager *man,
415 struct ttm_buffer_object *tbo,
416 const struct ttm_place *place,
417 struct ttm_resource *mem)
419 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
420 struct amdgpu_device *adev = to_amdgpu_device(mgr);
421 struct drm_mm *mm = &mgr->mm;
422 struct drm_mm_node *nodes;
423 enum drm_mm_insert_mode mode;
424 unsigned long lpfn, num_nodes, pages_per_node, pages_left;
425 uint64_t vis_usage = 0, mem_bytes, max_bytes;
433 max_bytes = adev->gmc.mc_vram_size;
434 if (tbo->type != ttm_bo_type_kernel)
435 max_bytes -= AMDGPU_VM_RESERVED_VRAM;
437 /* bail out quickly if there's likely not enough VRAM for this BO */
438 mem_bytes = (u64)mem->num_pages << PAGE_SHIFT;
439 if (atomic64_add_return(mem_bytes, &mgr->usage) > max_bytes) {
440 atomic64_sub(mem_bytes, &mgr->usage);
444 if (place->flags & TTM_PL_FLAG_CONTIGUOUS) {
445 pages_per_node = ~0ul;
448 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
449 pages_per_node = HPAGE_PMD_NR;
452 pages_per_node = (2UL << (20UL - PAGE_SHIFT));
454 pages_per_node = max((uint32_t)pages_per_node, mem->page_alignment);
455 num_nodes = DIV_ROUND_UP(mem->num_pages, pages_per_node);
458 nodes = kvmalloc_array((uint32_t)num_nodes, sizeof(*nodes),
459 GFP_KERNEL | __GFP_ZERO);
461 atomic64_sub(mem_bytes, &mgr->usage);
465 mode = DRM_MM_INSERT_BEST;
466 if (place->flags & TTM_PL_FLAG_TOPDOWN)
467 mode = DRM_MM_INSERT_HIGH;
470 pages_left = mem->num_pages;
472 spin_lock(&mgr->lock);
473 for (i = 0; pages_left >= pages_per_node; ++i) {
474 unsigned long pages = rounddown_pow_of_two(pages_left);
476 /* Limit maximum size to 2GB due to SG table limitations */
477 pages = min(pages, (2UL << (30 - PAGE_SHIFT)));
479 r = drm_mm_insert_node_in_range(mm, &nodes[i], pages,
486 vis_usage += amdgpu_vram_mgr_vis_size(adev, &nodes[i]);
487 amdgpu_vram_mgr_virt_start(mem, &nodes[i]);
491 for (; pages_left; ++i) {
492 unsigned long pages = min(pages_left, pages_per_node);
493 uint32_t alignment = mem->page_alignment;
495 if (pages == pages_per_node)
496 alignment = pages_per_node;
498 r = drm_mm_insert_node_in_range(mm, &nodes[i],
505 vis_usage += amdgpu_vram_mgr_vis_size(adev, &nodes[i]);
506 amdgpu_vram_mgr_virt_start(mem, &nodes[i]);
509 spin_unlock(&mgr->lock);
511 atomic64_add(vis_usage, &mgr->vis_usage);
513 mem->mm_node = nodes;
519 drm_mm_remove_node(&nodes[i]);
520 spin_unlock(&mgr->lock);
521 atomic64_sub(mem->num_pages << PAGE_SHIFT, &mgr->usage);
528 * amdgpu_vram_mgr_del - free ranges
530 * @man: TTM memory type manager
531 * @mem: TTM memory object
533 * Free the allocated VRAM again.
535 static void amdgpu_vram_mgr_del(struct ttm_resource_manager *man,
536 struct ttm_resource *mem)
538 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
539 struct amdgpu_device *adev = to_amdgpu_device(mgr);
540 struct drm_mm_node *nodes = mem->mm_node;
541 uint64_t usage = 0, vis_usage = 0;
542 unsigned pages = mem->num_pages;
547 spin_lock(&mgr->lock);
549 pages -= nodes->size;
550 drm_mm_remove_node(nodes);
551 usage += nodes->size << PAGE_SHIFT;
552 vis_usage += amdgpu_vram_mgr_vis_size(adev, nodes);
555 amdgpu_vram_mgr_do_reserve(man);
556 spin_unlock(&mgr->lock);
558 atomic64_sub(usage, &mgr->usage);
559 atomic64_sub(vis_usage, &mgr->vis_usage);
561 kvfree(mem->mm_node);
566 * amdgpu_vram_mgr_alloc_sgt - allocate and fill a sg table
568 * @adev: amdgpu device pointer
569 * @mem: TTM memory object
570 * @dev: the other device
571 * @dir: dma direction
572 * @sgt: resulting sg table
574 * Allocate and fill a sg table from a VRAM allocation.
576 int amdgpu_vram_mgr_alloc_sgt(struct amdgpu_device *adev,
577 struct ttm_resource *mem,
579 enum dma_data_direction dir,
580 struct sg_table **sgt)
582 struct drm_mm_node *node;
583 struct scatterlist *sg;
588 *sgt = kmalloc(sizeof(**sgt), GFP_KERNEL);
592 for (pages = mem->num_pages, node = mem->mm_node;
593 pages; pages -= node->size, ++node)
596 r = sg_alloc_table(*sgt, num_entries, GFP_KERNEL);
600 for_each_sgtable_sg((*sgt), sg, i)
604 for_each_sgtable_sg((*sgt), sg, i) {
605 phys_addr_t phys = (node->start << PAGE_SHIFT) +
607 size_t size = node->size << PAGE_SHIFT;
611 addr = dma_map_resource(dev, phys, size, dir,
612 DMA_ATTR_SKIP_CPU_SYNC);
613 r = dma_mapping_error(dev, addr);
617 sg_set_page(sg, NULL, size, 0);
618 sg_dma_address(sg) = addr;
619 sg_dma_len(sg) = size;
624 for_each_sgtable_sg((*sgt), sg, i) {
628 dma_unmap_resource(dev, sg->dma_address,
630 DMA_ATTR_SKIP_CPU_SYNC);
640 * amdgpu_vram_mgr_free_sgt - allocate and fill a sg table
642 * @adev: amdgpu device pointer
643 * @dev: device pointer
644 * @dir: data direction of resource to unmap
645 * @sgt: sg table to free
647 * Free a previously allocate sg table.
649 void amdgpu_vram_mgr_free_sgt(struct amdgpu_device *adev,
651 enum dma_data_direction dir,
652 struct sg_table *sgt)
654 struct scatterlist *sg;
657 for_each_sgtable_sg(sgt, sg, i)
658 dma_unmap_resource(dev, sg->dma_address,
660 DMA_ATTR_SKIP_CPU_SYNC);
666 * amdgpu_vram_mgr_usage - how many bytes are used in this domain
668 * @man: TTM memory type manager
670 * Returns how many bytes are used in this domain.
672 uint64_t amdgpu_vram_mgr_usage(struct ttm_resource_manager *man)
674 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
676 return atomic64_read(&mgr->usage);
680 * amdgpu_vram_mgr_vis_usage - how many bytes are used in the visible part
682 * @man: TTM memory type manager
684 * Returns how many bytes are used in the visible part of VRAM
686 uint64_t amdgpu_vram_mgr_vis_usage(struct ttm_resource_manager *man)
688 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
690 return atomic64_read(&mgr->vis_usage);
694 * amdgpu_vram_mgr_debug - dump VRAM table
696 * @man: TTM memory type manager
697 * @printer: DRM printer to use
699 * Dump the table content using printk.
701 static void amdgpu_vram_mgr_debug(struct ttm_resource_manager *man,
702 struct drm_printer *printer)
704 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
706 spin_lock(&mgr->lock);
707 drm_mm_print(&mgr->mm, printer);
708 spin_unlock(&mgr->lock);
710 drm_printf(printer, "man size:%llu pages, ram usage:%lluMB, vis usage:%lluMB\n",
711 man->size, amdgpu_vram_mgr_usage(man) >> 20,
712 amdgpu_vram_mgr_vis_usage(man) >> 20);
715 static const struct ttm_resource_manager_func amdgpu_vram_mgr_func = {
716 .alloc = amdgpu_vram_mgr_new,
717 .free = amdgpu_vram_mgr_del,
718 .debug = amdgpu_vram_mgr_debug