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 man->available_caching = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC;
182 man->default_caching = TTM_PL_FLAG_WC;
184 ttm_resource_manager_init(man, adev->gmc.real_vram_size >> PAGE_SHIFT);
186 man->func = &amdgpu_vram_mgr_func;
188 drm_mm_init(&mgr->mm, 0, man->size);
189 spin_lock_init(&mgr->lock);
190 INIT_LIST_HEAD(&mgr->reservations_pending);
191 INIT_LIST_HEAD(&mgr->reserved_pages);
193 /* Add the two VRAM-related sysfs files */
194 ret = sysfs_create_files(&adev->dev->kobj, amdgpu_vram_mgr_attributes);
196 DRM_ERROR("Failed to register sysfs\n");
198 ttm_set_driver_manager(&adev->mman.bdev, TTM_PL_VRAM, &mgr->manager);
199 ttm_resource_manager_set_used(man, true);
204 * amdgpu_vram_mgr_fini - free and destroy VRAM manager
206 * @adev: amdgpu_device pointer
208 * Destroy and free the VRAM manager, returns -EBUSY if ranges are still
209 * allocated inside it.
211 void amdgpu_vram_mgr_fini(struct amdgpu_device *adev)
213 struct amdgpu_vram_mgr *mgr = &adev->mman.vram_mgr;
214 struct ttm_resource_manager *man = &mgr->manager;
216 struct amdgpu_vram_reservation *rsv, *temp;
218 ttm_resource_manager_set_used(man, false);
220 ret = ttm_resource_manager_force_list_clean(&adev->mman.bdev, man);
224 spin_lock(&mgr->lock);
225 list_for_each_entry_safe(rsv, temp, &mgr->reservations_pending, node)
228 list_for_each_entry_safe(rsv, temp, &mgr->reserved_pages, node) {
229 drm_mm_remove_node(&rsv->mm_node);
232 drm_mm_takedown(&mgr->mm);
233 spin_unlock(&mgr->lock);
235 sysfs_remove_files(&adev->dev->kobj, amdgpu_vram_mgr_attributes);
237 ttm_resource_manager_cleanup(man);
238 ttm_set_driver_manager(&adev->mman.bdev, TTM_PL_VRAM, NULL);
242 * amdgpu_vram_mgr_vis_size - Calculate visible node size
244 * @adev: amdgpu_device pointer
245 * @node: MM node structure
247 * Calculate how many bytes of the MM node are inside visible VRAM
249 static u64 amdgpu_vram_mgr_vis_size(struct amdgpu_device *adev,
250 struct drm_mm_node *node)
252 uint64_t start = node->start << PAGE_SHIFT;
253 uint64_t end = (node->size + node->start) << PAGE_SHIFT;
255 if (start >= adev->gmc.visible_vram_size)
258 return (end > adev->gmc.visible_vram_size ?
259 adev->gmc.visible_vram_size : end) - start;
263 * amdgpu_vram_mgr_bo_visible_size - CPU visible BO size
265 * @bo: &amdgpu_bo buffer object (must be in VRAM)
268 * How much of the given &amdgpu_bo buffer object lies in CPU visible VRAM.
270 u64 amdgpu_vram_mgr_bo_visible_size(struct amdgpu_bo *bo)
272 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
273 struct ttm_resource *mem = &bo->tbo.mem;
274 struct drm_mm_node *nodes = mem->mm_node;
275 unsigned pages = mem->num_pages;
278 if (amdgpu_gmc_vram_full_visible(&adev->gmc))
279 return amdgpu_bo_size(bo);
281 if (mem->start >= adev->gmc.visible_vram_size >> PAGE_SHIFT)
284 for (usage = 0; nodes && pages; pages -= nodes->size, nodes++)
285 usage += amdgpu_vram_mgr_vis_size(adev, nodes);
290 static void amdgpu_vram_mgr_do_reserve(struct ttm_resource_manager *man)
292 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
293 struct amdgpu_device *adev = to_amdgpu_device(mgr);
294 struct drm_mm *mm = &mgr->mm;
295 struct amdgpu_vram_reservation *rsv, *temp;
298 list_for_each_entry_safe(rsv, temp, &mgr->reservations_pending, node) {
299 if (drm_mm_reserve_node(mm, &rsv->mm_node))
302 dev_dbg(adev->dev, "Reservation 0x%llx - %lld, Successed\n",
303 rsv->mm_node.start, rsv->mm_node.size);
305 vis_usage = amdgpu_vram_mgr_vis_size(adev, &rsv->mm_node);
306 atomic64_add(vis_usage, &mgr->vis_usage);
307 atomic64_add(rsv->mm_node.size << PAGE_SHIFT, &mgr->usage);
308 list_move(&rsv->node, &mgr->reserved_pages);
313 * amdgpu_vram_mgr_reserve_range - Reserve a range from VRAM
315 * @man: TTM memory type manager
316 * @start: start address of the range in VRAM
317 * @size: size of the range
319 * Reserve memory from start addess with the specified size in VRAM
321 int amdgpu_vram_mgr_reserve_range(struct ttm_resource_manager *man,
322 uint64_t start, uint64_t size)
324 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
325 struct amdgpu_vram_reservation *rsv;
327 rsv = kzalloc(sizeof(*rsv), GFP_KERNEL);
331 INIT_LIST_HEAD(&rsv->node);
332 rsv->mm_node.start = start >> PAGE_SHIFT;
333 rsv->mm_node.size = size >> PAGE_SHIFT;
335 spin_lock(&mgr->lock);
336 list_add_tail(&mgr->reservations_pending, &rsv->node);
337 amdgpu_vram_mgr_do_reserve(man);
338 spin_unlock(&mgr->lock);
344 * amdgpu_vram_mgr_query_page_status - query the reservation status
346 * @man: TTM memory type manager
347 * @start: start address of a page in VRAM
350 * -EBUSY: the page is still hold and in pending list
351 * 0: the page has been reserved
352 * -ENOENT: the input page is not a reservation
354 int amdgpu_vram_mgr_query_page_status(struct ttm_resource_manager *man,
357 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
358 struct amdgpu_vram_reservation *rsv;
361 spin_lock(&mgr->lock);
363 list_for_each_entry(rsv, &mgr->reservations_pending, node) {
364 if ((rsv->mm_node.start <= start) &&
365 (start < (rsv->mm_node.start + rsv->mm_node.size))) {
371 list_for_each_entry(rsv, &mgr->reserved_pages, node) {
372 if ((rsv->mm_node.start <= start) &&
373 (start < (rsv->mm_node.start + rsv->mm_node.size))) {
381 spin_unlock(&mgr->lock);
386 * amdgpu_vram_mgr_virt_start - update virtual start address
388 * @mem: ttm_resource to update
389 * @node: just allocated node
391 * Calculate a virtual BO start address to easily check if everything is CPU
394 static void amdgpu_vram_mgr_virt_start(struct ttm_resource *mem,
395 struct drm_mm_node *node)
399 start = node->start + node->size;
400 if (start > mem->num_pages)
401 start -= mem->num_pages;
404 mem->start = max(mem->start, start);
408 * amdgpu_vram_mgr_new - allocate new ranges
410 * @man: TTM memory type manager
411 * @tbo: TTM BO we need this range for
412 * @place: placement flags and restrictions
413 * @mem: the resulting mem object
415 * Allocate VRAM for the given BO.
417 static int amdgpu_vram_mgr_new(struct ttm_resource_manager *man,
418 struct ttm_buffer_object *tbo,
419 const struct ttm_place *place,
420 struct ttm_resource *mem)
422 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
423 struct amdgpu_device *adev = to_amdgpu_device(mgr);
424 struct drm_mm *mm = &mgr->mm;
425 struct drm_mm_node *nodes;
426 enum drm_mm_insert_mode mode;
427 unsigned long lpfn, num_nodes, pages_per_node, pages_left;
428 uint64_t vis_usage = 0, mem_bytes, max_bytes;
436 max_bytes = adev->gmc.mc_vram_size;
437 if (tbo->type != ttm_bo_type_kernel)
438 max_bytes -= AMDGPU_VM_RESERVED_VRAM;
440 /* bail out quickly if there's likely not enough VRAM for this BO */
441 mem_bytes = (u64)mem->num_pages << PAGE_SHIFT;
442 if (atomic64_add_return(mem_bytes, &mgr->usage) > max_bytes) {
443 atomic64_sub(mem_bytes, &mgr->usage);
447 if (place->flags & TTM_PL_FLAG_CONTIGUOUS) {
448 pages_per_node = ~0ul;
451 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
452 pages_per_node = HPAGE_PMD_NR;
455 pages_per_node = (2UL << (20UL - PAGE_SHIFT));
457 pages_per_node = max((uint32_t)pages_per_node, mem->page_alignment);
458 num_nodes = DIV_ROUND_UP(mem->num_pages, pages_per_node);
461 nodes = kvmalloc_array((uint32_t)num_nodes, sizeof(*nodes),
462 GFP_KERNEL | __GFP_ZERO);
464 atomic64_sub(mem_bytes, &mgr->usage);
468 mode = DRM_MM_INSERT_BEST;
469 if (place->flags & TTM_PL_FLAG_TOPDOWN)
470 mode = DRM_MM_INSERT_HIGH;
473 pages_left = mem->num_pages;
475 spin_lock(&mgr->lock);
476 for (i = 0; pages_left >= pages_per_node; ++i) {
477 unsigned long pages = rounddown_pow_of_two(pages_left);
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 * @sgt: sg table to free
645 * Free a previously allocate sg table.
647 void amdgpu_vram_mgr_free_sgt(struct amdgpu_device *adev,
649 enum dma_data_direction dir,
650 struct sg_table *sgt)
652 struct scatterlist *sg;
655 for_each_sgtable_sg(sgt, sg, i)
656 dma_unmap_resource(dev, sg->dma_address,
658 DMA_ATTR_SKIP_CPU_SYNC);
664 * amdgpu_vram_mgr_usage - how many bytes are used in this domain
666 * @man: TTM memory type manager
668 * Returns how many bytes are used in this domain.
670 uint64_t amdgpu_vram_mgr_usage(struct ttm_resource_manager *man)
672 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
674 return atomic64_read(&mgr->usage);
678 * amdgpu_vram_mgr_vis_usage - how many bytes are used in the visible part
680 * @man: TTM memory type manager
682 * Returns how many bytes are used in the visible part of VRAM
684 uint64_t amdgpu_vram_mgr_vis_usage(struct ttm_resource_manager *man)
686 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
688 return atomic64_read(&mgr->vis_usage);
692 * amdgpu_vram_mgr_debug - dump VRAM table
694 * @man: TTM memory type manager
695 * @printer: DRM printer to use
697 * Dump the table content using printk.
699 static void amdgpu_vram_mgr_debug(struct ttm_resource_manager *man,
700 struct drm_printer *printer)
702 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
704 spin_lock(&mgr->lock);
705 drm_mm_print(&mgr->mm, printer);
706 spin_unlock(&mgr->lock);
708 drm_printf(printer, "man size:%llu pages, ram usage:%lluMB, vis usage:%lluMB\n",
709 man->size, amdgpu_vram_mgr_usage(man) >> 20,
710 amdgpu_vram_mgr_vis_usage(man) >> 20);
713 static const struct ttm_resource_manager_func amdgpu_vram_mgr_func = {
714 .alloc = amdgpu_vram_mgr_new,
715 .free = amdgpu_vram_mgr_del,
716 .debug = amdgpu_vram_mgr_debug