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 = ddev->dev_private;
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 = ddev->dev_private;
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 = ddev->dev_private;
88 struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
89 return snprintf(buf, PAGE_SIZE, "%llu\n",
90 amdgpu_vram_mgr_usage(man));
94 * DOC: mem_info_vis_vram_used
96 * The amdgpu driver provides a sysfs API for reporting current total of
98 * The file mem_info_vis_vram_used is used for this and returns the total
99 * amount of currently used visible VRAM in bytes
101 static ssize_t amdgpu_mem_info_vis_vram_used_show(struct device *dev,
102 struct device_attribute *attr, char *buf)
104 struct drm_device *ddev = dev_get_drvdata(dev);
105 struct amdgpu_device *adev = ddev->dev_private;
106 struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
107 return snprintf(buf, PAGE_SIZE, "%llu\n",
108 amdgpu_vram_mgr_vis_usage(man));
111 static ssize_t amdgpu_mem_info_vram_vendor(struct device *dev,
112 struct device_attribute *attr,
115 struct drm_device *ddev = dev_get_drvdata(dev);
116 struct amdgpu_device *adev = ddev->dev_private;
118 switch (adev->gmc.vram_vendor) {
120 return snprintf(buf, PAGE_SIZE, "samsung\n");
122 return snprintf(buf, PAGE_SIZE, "infineon\n");
124 return snprintf(buf, PAGE_SIZE, "elpida\n");
126 return snprintf(buf, PAGE_SIZE, "etron\n");
128 return snprintf(buf, PAGE_SIZE, "nanya\n");
130 return snprintf(buf, PAGE_SIZE, "hynix\n");
132 return snprintf(buf, PAGE_SIZE, "mosel\n");
134 return snprintf(buf, PAGE_SIZE, "winbond\n");
136 return snprintf(buf, PAGE_SIZE, "esmt\n");
138 return snprintf(buf, PAGE_SIZE, "micron\n");
140 return snprintf(buf, PAGE_SIZE, "unknown\n");
144 static DEVICE_ATTR(mem_info_vram_total, S_IRUGO,
145 amdgpu_mem_info_vram_total_show, NULL);
146 static DEVICE_ATTR(mem_info_vis_vram_total, S_IRUGO,
147 amdgpu_mem_info_vis_vram_total_show,NULL);
148 static DEVICE_ATTR(mem_info_vram_used, S_IRUGO,
149 amdgpu_mem_info_vram_used_show, NULL);
150 static DEVICE_ATTR(mem_info_vis_vram_used, S_IRUGO,
151 amdgpu_mem_info_vis_vram_used_show, NULL);
152 static DEVICE_ATTR(mem_info_vram_vendor, S_IRUGO,
153 amdgpu_mem_info_vram_vendor, NULL);
155 static const struct attribute *amdgpu_vram_mgr_attributes[] = {
156 &dev_attr_mem_info_vram_total.attr,
157 &dev_attr_mem_info_vis_vram_total.attr,
158 &dev_attr_mem_info_vram_used.attr,
159 &dev_attr_mem_info_vis_vram_used.attr,
160 &dev_attr_mem_info_vram_vendor.attr,
164 static const struct ttm_resource_manager_func amdgpu_vram_mgr_func;
167 * amdgpu_vram_mgr_init - init VRAM manager and DRM MM
169 * @man: TTM memory type manager
170 * @p_size: maximum size of VRAM
172 * Allocate and initialize the VRAM manager.
174 int amdgpu_vram_mgr_init(struct amdgpu_device *adev)
176 struct amdgpu_vram_mgr *mgr = &adev->mman.vram_mgr;
177 struct ttm_resource_manager *man = &mgr->manager;
180 man->available_caching = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC;
181 man->default_caching = TTM_PL_FLAG_WC;
183 ttm_resource_manager_init(man, adev->gmc.real_vram_size >> PAGE_SHIFT);
185 man->func = &amdgpu_vram_mgr_func;
187 drm_mm_init(&mgr->mm, 0, man->size);
188 spin_lock_init(&mgr->lock);
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 * @man: TTM memory type manager
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;
214 ttm_resource_manager_set_used(man, false);
216 ret = ttm_resource_manager_force_list_clean(&adev->mman.bdev, man);
220 spin_lock(&mgr->lock);
221 drm_mm_takedown(&mgr->mm);
222 spin_unlock(&mgr->lock);
224 sysfs_remove_files(&adev->dev->kobj, amdgpu_vram_mgr_attributes);
226 ttm_resource_manager_cleanup(man);
227 ttm_set_driver_manager(&adev->mman.bdev, TTM_PL_VRAM, NULL);
231 * amdgpu_vram_mgr_vis_size - Calculate visible node size
233 * @adev: amdgpu device structure
234 * @node: MM node structure
236 * Calculate how many bytes of the MM node are inside visible VRAM
238 static u64 amdgpu_vram_mgr_vis_size(struct amdgpu_device *adev,
239 struct drm_mm_node *node)
241 uint64_t start = node->start << PAGE_SHIFT;
242 uint64_t end = (node->size + node->start) << PAGE_SHIFT;
244 if (start >= adev->gmc.visible_vram_size)
247 return (end > adev->gmc.visible_vram_size ?
248 adev->gmc.visible_vram_size : end) - start;
252 * amdgpu_vram_mgr_bo_visible_size - CPU visible BO size
254 * @bo: &amdgpu_bo buffer object (must be in VRAM)
257 * How much of the given &amdgpu_bo buffer object lies in CPU visible VRAM.
259 u64 amdgpu_vram_mgr_bo_visible_size(struct amdgpu_bo *bo)
261 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
262 struct ttm_resource *mem = &bo->tbo.mem;
263 struct drm_mm_node *nodes = mem->mm_node;
264 unsigned pages = mem->num_pages;
267 if (amdgpu_gmc_vram_full_visible(&adev->gmc))
268 return amdgpu_bo_size(bo);
270 if (mem->start >= adev->gmc.visible_vram_size >> PAGE_SHIFT)
273 for (usage = 0; nodes && pages; pages -= nodes->size, nodes++)
274 usage += amdgpu_vram_mgr_vis_size(adev, nodes);
280 * amdgpu_vram_mgr_virt_start - update virtual start address
282 * @mem: ttm_resource to update
283 * @node: just allocated node
285 * Calculate a virtual BO start address to easily check if everything is CPU
288 static void amdgpu_vram_mgr_virt_start(struct ttm_resource *mem,
289 struct drm_mm_node *node)
293 start = node->start + node->size;
294 if (start > mem->num_pages)
295 start -= mem->num_pages;
298 mem->start = max(mem->start, start);
302 * amdgpu_vram_mgr_new - allocate new ranges
304 * @man: TTM memory type manager
305 * @tbo: TTM BO we need this range for
306 * @place: placement flags and restrictions
307 * @mem: the resulting mem object
309 * Allocate VRAM for the given BO.
311 static int amdgpu_vram_mgr_new(struct ttm_resource_manager *man,
312 struct ttm_buffer_object *tbo,
313 const struct ttm_place *place,
314 struct ttm_resource *mem)
316 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
317 struct amdgpu_device *adev = to_amdgpu_device(mgr);
318 struct drm_mm *mm = &mgr->mm;
319 struct drm_mm_node *nodes;
320 enum drm_mm_insert_mode mode;
321 unsigned long lpfn, num_nodes, pages_per_node, pages_left;
322 uint64_t vis_usage = 0, mem_bytes, max_bytes;
330 max_bytes = adev->gmc.mc_vram_size;
331 if (tbo->type != ttm_bo_type_kernel)
332 max_bytes -= AMDGPU_VM_RESERVED_VRAM;
334 /* bail out quickly if there's likely not enough VRAM for this BO */
335 mem_bytes = (u64)mem->num_pages << PAGE_SHIFT;
336 if (atomic64_add_return(mem_bytes, &mgr->usage) > max_bytes) {
337 atomic64_sub(mem_bytes, &mgr->usage);
341 if (place->flags & TTM_PL_FLAG_CONTIGUOUS) {
342 pages_per_node = ~0ul;
345 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
346 pages_per_node = HPAGE_PMD_NR;
349 pages_per_node = (2UL << (20UL - PAGE_SHIFT));
351 pages_per_node = max((uint32_t)pages_per_node, mem->page_alignment);
352 num_nodes = DIV_ROUND_UP(mem->num_pages, pages_per_node);
355 nodes = kvmalloc_array((uint32_t)num_nodes, sizeof(*nodes),
356 GFP_KERNEL | __GFP_ZERO);
358 atomic64_sub(mem_bytes, &mgr->usage);
362 mode = DRM_MM_INSERT_BEST;
363 if (place->flags & TTM_PL_FLAG_TOPDOWN)
364 mode = DRM_MM_INSERT_HIGH;
367 pages_left = mem->num_pages;
369 spin_lock(&mgr->lock);
370 for (i = 0; pages_left >= pages_per_node; ++i) {
371 unsigned long pages = rounddown_pow_of_two(pages_left);
373 r = drm_mm_insert_node_in_range(mm, &nodes[i], pages,
380 vis_usage += amdgpu_vram_mgr_vis_size(adev, &nodes[i]);
381 amdgpu_vram_mgr_virt_start(mem, &nodes[i]);
385 for (; pages_left; ++i) {
386 unsigned long pages = min(pages_left, pages_per_node);
387 uint32_t alignment = mem->page_alignment;
389 if (pages == pages_per_node)
390 alignment = pages_per_node;
392 r = drm_mm_insert_node_in_range(mm, &nodes[i],
399 vis_usage += amdgpu_vram_mgr_vis_size(adev, &nodes[i]);
400 amdgpu_vram_mgr_virt_start(mem, &nodes[i]);
403 spin_unlock(&mgr->lock);
405 atomic64_add(vis_usage, &mgr->vis_usage);
407 mem->mm_node = nodes;
413 drm_mm_remove_node(&nodes[i]);
414 spin_unlock(&mgr->lock);
415 atomic64_sub(mem->num_pages << PAGE_SHIFT, &mgr->usage);
422 * amdgpu_vram_mgr_del - free ranges
424 * @man: TTM memory type manager
425 * @mem: TTM memory object
427 * Free the allocated VRAM again.
429 static void amdgpu_vram_mgr_del(struct ttm_resource_manager *man,
430 struct ttm_resource *mem)
432 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
433 struct amdgpu_device *adev = to_amdgpu_device(mgr);
434 struct drm_mm_node *nodes = mem->mm_node;
435 uint64_t usage = 0, vis_usage = 0;
436 unsigned pages = mem->num_pages;
441 spin_lock(&mgr->lock);
443 pages -= nodes->size;
444 drm_mm_remove_node(nodes);
445 usage += nodes->size << PAGE_SHIFT;
446 vis_usage += amdgpu_vram_mgr_vis_size(adev, nodes);
449 spin_unlock(&mgr->lock);
451 atomic64_sub(usage, &mgr->usage);
452 atomic64_sub(vis_usage, &mgr->vis_usage);
454 kvfree(mem->mm_node);
459 * amdgpu_vram_mgr_alloc_sgt - allocate and fill a sg table
461 * @adev: amdgpu device pointer
462 * @mem: TTM memory object
463 * @dev: the other device
464 * @dir: dma direction
465 * @sgt: resulting sg table
467 * Allocate and fill a sg table from a VRAM allocation.
469 int amdgpu_vram_mgr_alloc_sgt(struct amdgpu_device *adev,
470 struct ttm_resource *mem,
472 enum dma_data_direction dir,
473 struct sg_table **sgt)
475 struct drm_mm_node *node;
476 struct scatterlist *sg;
481 *sgt = kmalloc(sizeof(*sg), GFP_KERNEL);
485 for (pages = mem->num_pages, node = mem->mm_node;
486 pages; pages -= node->size, ++node)
489 r = sg_alloc_table(*sgt, num_entries, GFP_KERNEL);
493 for_each_sgtable_sg((*sgt), sg, i)
497 for_each_sgtable_sg((*sgt), sg, i) {
498 phys_addr_t phys = (node->start << PAGE_SHIFT) +
500 size_t size = node->size << PAGE_SHIFT;
504 addr = dma_map_resource(dev, phys, size, dir,
505 DMA_ATTR_SKIP_CPU_SYNC);
506 r = dma_mapping_error(dev, addr);
510 sg_set_page(sg, NULL, size, 0);
511 sg_dma_address(sg) = addr;
512 sg_dma_len(sg) = size;
517 for_each_sgtable_sg((*sgt), sg, i) {
521 dma_unmap_resource(dev, sg->dma_address,
523 DMA_ATTR_SKIP_CPU_SYNC);
533 * amdgpu_vram_mgr_alloc_sgt - allocate and fill a sg table
535 * @adev: amdgpu device pointer
536 * @sgt: sg table to free
538 * Free a previously allocate sg table.
540 void amdgpu_vram_mgr_free_sgt(struct amdgpu_device *adev,
542 enum dma_data_direction dir,
543 struct sg_table *sgt)
545 struct scatterlist *sg;
548 for_each_sgtable_sg(sgt, sg, i)
549 dma_unmap_resource(dev, sg->dma_address,
551 DMA_ATTR_SKIP_CPU_SYNC);
557 * amdgpu_vram_mgr_usage - how many bytes are used in this domain
559 * @man: TTM memory type manager
561 * Returns how many bytes are used in this domain.
563 uint64_t amdgpu_vram_mgr_usage(struct ttm_resource_manager *man)
565 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
567 return atomic64_read(&mgr->usage);
571 * amdgpu_vram_mgr_vis_usage - how many bytes are used in the visible part
573 * @man: TTM memory type manager
575 * Returns how many bytes are used in the visible part of VRAM
577 uint64_t amdgpu_vram_mgr_vis_usage(struct ttm_resource_manager *man)
579 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
581 return atomic64_read(&mgr->vis_usage);
585 * amdgpu_vram_mgr_debug - dump VRAM table
587 * @man: TTM memory type manager
588 * @printer: DRM printer to use
590 * Dump the table content using printk.
592 static void amdgpu_vram_mgr_debug(struct ttm_resource_manager *man,
593 struct drm_printer *printer)
595 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
597 spin_lock(&mgr->lock);
598 drm_mm_print(&mgr->mm, printer);
599 spin_unlock(&mgr->lock);
601 drm_printf(printer, "man size:%llu pages, ram usage:%lluMB, vis usage:%lluMB\n",
602 man->size, amdgpu_vram_mgr_usage(man) >> 20,
603 amdgpu_vram_mgr_vis_usage(man) >> 20);
606 static const struct ttm_resource_manager_func amdgpu_vram_mgr_func = {
607 .alloc = amdgpu_vram_mgr_new,
608 .free = amdgpu_vram_mgr_del,
609 .debug = amdgpu_vram_mgr_debug