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 * @man: TTM memory type manager
172 * @p_size: maximum size of VRAM
174 * Allocate and initialize the VRAM manager.
176 int amdgpu_vram_mgr_init(struct amdgpu_device *adev)
178 struct amdgpu_vram_mgr *mgr = &adev->mman.vram_mgr;
179 struct ttm_resource_manager *man = &mgr->manager;
182 man->available_caching = TTM_PL_FLAG_UNCACHED | 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);
191 /* Add the two VRAM-related sysfs files */
192 ret = sysfs_create_files(&adev->dev->kobj, amdgpu_vram_mgr_attributes);
194 DRM_ERROR("Failed to register sysfs\n");
196 ttm_set_driver_manager(&adev->mman.bdev, TTM_PL_VRAM, &mgr->manager);
197 ttm_resource_manager_set_used(man, true);
202 * amdgpu_vram_mgr_fini - free and destroy VRAM manager
204 * @man: TTM memory type manager
206 * Destroy and free the VRAM manager, returns -EBUSY if ranges are still
207 * allocated inside it.
209 void amdgpu_vram_mgr_fini(struct amdgpu_device *adev)
211 struct amdgpu_vram_mgr *mgr = &adev->mman.vram_mgr;
212 struct ttm_resource_manager *man = &mgr->manager;
215 ttm_resource_manager_set_used(man, false);
217 ret = ttm_resource_manager_force_list_clean(&adev->mman.bdev, man);
221 spin_lock(&mgr->lock);
222 drm_mm_takedown(&mgr->mm);
223 spin_unlock(&mgr->lock);
225 sysfs_remove_files(&adev->dev->kobj, amdgpu_vram_mgr_attributes);
227 ttm_resource_manager_cleanup(man);
228 ttm_set_driver_manager(&adev->mman.bdev, TTM_PL_VRAM, NULL);
232 * amdgpu_vram_mgr_vis_size - Calculate visible node size
234 * @adev: amdgpu device structure
235 * @node: MM node structure
237 * Calculate how many bytes of the MM node are inside visible VRAM
239 static u64 amdgpu_vram_mgr_vis_size(struct amdgpu_device *adev,
240 struct drm_mm_node *node)
242 uint64_t start = node->start << PAGE_SHIFT;
243 uint64_t end = (node->size + node->start) << PAGE_SHIFT;
245 if (start >= adev->gmc.visible_vram_size)
248 return (end > adev->gmc.visible_vram_size ?
249 adev->gmc.visible_vram_size : end) - start;
253 * amdgpu_vram_mgr_bo_visible_size - CPU visible BO size
255 * @bo: &amdgpu_bo buffer object (must be in VRAM)
258 * How much of the given &amdgpu_bo buffer object lies in CPU visible VRAM.
260 u64 amdgpu_vram_mgr_bo_visible_size(struct amdgpu_bo *bo)
262 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
263 struct ttm_resource *mem = &bo->tbo.mem;
264 struct drm_mm_node *nodes = mem->mm_node;
265 unsigned pages = mem->num_pages;
268 if (amdgpu_gmc_vram_full_visible(&adev->gmc))
269 return amdgpu_bo_size(bo);
271 if (mem->start >= adev->gmc.visible_vram_size >> PAGE_SHIFT)
274 for (usage = 0; nodes && pages; pages -= nodes->size, nodes++)
275 usage += amdgpu_vram_mgr_vis_size(adev, nodes);
281 * amdgpu_vram_mgr_virt_start - update virtual start address
283 * @mem: ttm_resource to update
284 * @node: just allocated node
286 * Calculate a virtual BO start address to easily check if everything is CPU
289 static void amdgpu_vram_mgr_virt_start(struct ttm_resource *mem,
290 struct drm_mm_node *node)
294 start = node->start + node->size;
295 if (start > mem->num_pages)
296 start -= mem->num_pages;
299 mem->start = max(mem->start, start);
303 * amdgpu_vram_mgr_new - allocate new ranges
305 * @man: TTM memory type manager
306 * @tbo: TTM BO we need this range for
307 * @place: placement flags and restrictions
308 * @mem: the resulting mem object
310 * Allocate VRAM for the given BO.
312 static int amdgpu_vram_mgr_new(struct ttm_resource_manager *man,
313 struct ttm_buffer_object *tbo,
314 const struct ttm_place *place,
315 struct ttm_resource *mem)
317 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
318 struct amdgpu_device *adev = to_amdgpu_device(mgr);
319 struct drm_mm *mm = &mgr->mm;
320 struct drm_mm_node *nodes;
321 enum drm_mm_insert_mode mode;
322 unsigned long lpfn, num_nodes, pages_per_node, pages_left;
323 uint64_t vis_usage = 0, mem_bytes, max_bytes;
331 max_bytes = adev->gmc.mc_vram_size;
332 if (tbo->type != ttm_bo_type_kernel)
333 max_bytes -= AMDGPU_VM_RESERVED_VRAM;
335 /* bail out quickly if there's likely not enough VRAM for this BO */
336 mem_bytes = (u64)mem->num_pages << PAGE_SHIFT;
337 if (atomic64_add_return(mem_bytes, &mgr->usage) > max_bytes) {
338 atomic64_sub(mem_bytes, &mgr->usage);
342 if (place->flags & TTM_PL_FLAG_CONTIGUOUS) {
343 pages_per_node = ~0ul;
346 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
347 pages_per_node = HPAGE_PMD_NR;
350 pages_per_node = (2UL << (20UL - PAGE_SHIFT));
352 pages_per_node = max((uint32_t)pages_per_node, mem->page_alignment);
353 num_nodes = DIV_ROUND_UP(mem->num_pages, pages_per_node);
356 nodes = kvmalloc_array((uint32_t)num_nodes, sizeof(*nodes),
357 GFP_KERNEL | __GFP_ZERO);
359 atomic64_sub(mem_bytes, &mgr->usage);
363 mode = DRM_MM_INSERT_BEST;
364 if (place->flags & TTM_PL_FLAG_TOPDOWN)
365 mode = DRM_MM_INSERT_HIGH;
368 pages_left = mem->num_pages;
370 spin_lock(&mgr->lock);
371 for (i = 0; pages_left >= pages_per_node; ++i) {
372 unsigned long pages = rounddown_pow_of_two(pages_left);
374 r = drm_mm_insert_node_in_range(mm, &nodes[i], pages,
381 vis_usage += amdgpu_vram_mgr_vis_size(adev, &nodes[i]);
382 amdgpu_vram_mgr_virt_start(mem, &nodes[i]);
386 for (; pages_left; ++i) {
387 unsigned long pages = min(pages_left, pages_per_node);
388 uint32_t alignment = mem->page_alignment;
390 if (pages == pages_per_node)
391 alignment = pages_per_node;
393 r = drm_mm_insert_node_in_range(mm, &nodes[i],
400 vis_usage += amdgpu_vram_mgr_vis_size(adev, &nodes[i]);
401 amdgpu_vram_mgr_virt_start(mem, &nodes[i]);
404 spin_unlock(&mgr->lock);
406 atomic64_add(vis_usage, &mgr->vis_usage);
408 mem->mm_node = nodes;
414 drm_mm_remove_node(&nodes[i]);
415 spin_unlock(&mgr->lock);
416 atomic64_sub(mem->num_pages << PAGE_SHIFT, &mgr->usage);
423 * amdgpu_vram_mgr_del - free ranges
425 * @man: TTM memory type manager
426 * @mem: TTM memory object
428 * Free the allocated VRAM again.
430 static void amdgpu_vram_mgr_del(struct ttm_resource_manager *man,
431 struct ttm_resource *mem)
433 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
434 struct amdgpu_device *adev = to_amdgpu_device(mgr);
435 struct drm_mm_node *nodes = mem->mm_node;
436 uint64_t usage = 0, vis_usage = 0;
437 unsigned pages = mem->num_pages;
442 spin_lock(&mgr->lock);
444 pages -= nodes->size;
445 drm_mm_remove_node(nodes);
446 usage += nodes->size << PAGE_SHIFT;
447 vis_usage += amdgpu_vram_mgr_vis_size(adev, nodes);
450 spin_unlock(&mgr->lock);
452 atomic64_sub(usage, &mgr->usage);
453 atomic64_sub(vis_usage, &mgr->vis_usage);
455 kvfree(mem->mm_node);
460 * amdgpu_vram_mgr_alloc_sgt - allocate and fill a sg table
462 * @adev: amdgpu device pointer
463 * @mem: TTM memory object
464 * @dev: the other device
465 * @dir: dma direction
466 * @sgt: resulting sg table
468 * Allocate and fill a sg table from a VRAM allocation.
470 int amdgpu_vram_mgr_alloc_sgt(struct amdgpu_device *adev,
471 struct ttm_resource *mem,
473 enum dma_data_direction dir,
474 struct sg_table **sgt)
476 struct drm_mm_node *node;
477 struct scatterlist *sg;
482 *sgt = kmalloc(sizeof(**sgt), GFP_KERNEL);
486 for (pages = mem->num_pages, node = mem->mm_node;
487 pages; pages -= node->size, ++node)
490 r = sg_alloc_table(*sgt, num_entries, GFP_KERNEL);
494 for_each_sgtable_sg((*sgt), sg, i)
498 for_each_sgtable_sg((*sgt), sg, i) {
499 phys_addr_t phys = (node->start << PAGE_SHIFT) +
501 size_t size = node->size << PAGE_SHIFT;
505 addr = dma_map_resource(dev, phys, size, dir,
506 DMA_ATTR_SKIP_CPU_SYNC);
507 r = dma_mapping_error(dev, addr);
511 sg_set_page(sg, NULL, size, 0);
512 sg_dma_address(sg) = addr;
513 sg_dma_len(sg) = size;
518 for_each_sgtable_sg((*sgt), sg, i) {
522 dma_unmap_resource(dev, sg->dma_address,
524 DMA_ATTR_SKIP_CPU_SYNC);
534 * amdgpu_vram_mgr_alloc_sgt - allocate and fill a sg table
536 * @adev: amdgpu device pointer
537 * @sgt: sg table to free
539 * Free a previously allocate sg table.
541 void amdgpu_vram_mgr_free_sgt(struct amdgpu_device *adev,
543 enum dma_data_direction dir,
544 struct sg_table *sgt)
546 struct scatterlist *sg;
549 for_each_sgtable_sg(sgt, sg, i)
550 dma_unmap_resource(dev, sg->dma_address,
552 DMA_ATTR_SKIP_CPU_SYNC);
558 * amdgpu_vram_mgr_usage - how many bytes are used in this domain
560 * @man: TTM memory type manager
562 * Returns how many bytes are used in this domain.
564 uint64_t amdgpu_vram_mgr_usage(struct ttm_resource_manager *man)
566 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
568 return atomic64_read(&mgr->usage);
572 * amdgpu_vram_mgr_vis_usage - how many bytes are used in the visible part
574 * @man: TTM memory type manager
576 * Returns how many bytes are used in the visible part of VRAM
578 uint64_t amdgpu_vram_mgr_vis_usage(struct ttm_resource_manager *man)
580 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
582 return atomic64_read(&mgr->vis_usage);
586 * amdgpu_vram_mgr_debug - dump VRAM table
588 * @man: TTM memory type manager
589 * @printer: DRM printer to use
591 * Dump the table content using printk.
593 static void amdgpu_vram_mgr_debug(struct ttm_resource_manager *man,
594 struct drm_printer *printer)
596 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
598 spin_lock(&mgr->lock);
599 drm_mm_print(&mgr->mm, printer);
600 spin_unlock(&mgr->lock);
602 drm_printf(printer, "man size:%llu pages, ram usage:%lluMB, vis usage:%lluMB\n",
603 man->size, amdgpu_vram_mgr_usage(man) >> 20,
604 amdgpu_vram_mgr_vis_usage(man) >> 20);
607 static const struct ttm_resource_manager_func amdgpu_vram_mgr_func = {
608 .alloc = amdgpu_vram_mgr_new,
609 .free = amdgpu_vram_mgr_del,
610 .debug = amdgpu_vram_mgr_debug