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);
188 /* Add the two VRAM-related sysfs files */
189 ret = sysfs_create_files(&adev->dev->kobj, amdgpu_vram_mgr_attributes);
191 DRM_ERROR("Failed to register sysfs\n");
193 ttm_set_driver_manager(&adev->mman.bdev, TTM_PL_VRAM, &mgr->manager);
194 ttm_resource_manager_set_used(man, true);
199 * amdgpu_vram_mgr_fini - free and destroy VRAM manager
201 * @adev: amdgpu_device pointer
203 * Destroy and free the VRAM manager, returns -EBUSY if ranges are still
204 * allocated inside it.
206 void amdgpu_vram_mgr_fini(struct amdgpu_device *adev)
208 struct amdgpu_vram_mgr *mgr = &adev->mman.vram_mgr;
209 struct ttm_resource_manager *man = &mgr->manager;
212 ttm_resource_manager_set_used(man, false);
214 ret = ttm_resource_manager_force_list_clean(&adev->mman.bdev, man);
218 spin_lock(&mgr->lock);
219 drm_mm_takedown(&mgr->mm);
220 spin_unlock(&mgr->lock);
222 sysfs_remove_files(&adev->dev->kobj, amdgpu_vram_mgr_attributes);
224 ttm_resource_manager_cleanup(man);
225 ttm_set_driver_manager(&adev->mman.bdev, TTM_PL_VRAM, NULL);
229 * amdgpu_vram_mgr_vis_size - Calculate visible node size
231 * @adev: amdgpu_device pointer
232 * @node: MM node structure
234 * Calculate how many bytes of the MM node are inside visible VRAM
236 static u64 amdgpu_vram_mgr_vis_size(struct amdgpu_device *adev,
237 struct drm_mm_node *node)
239 uint64_t start = node->start << PAGE_SHIFT;
240 uint64_t end = (node->size + node->start) << PAGE_SHIFT;
242 if (start >= adev->gmc.visible_vram_size)
245 return (end > adev->gmc.visible_vram_size ?
246 adev->gmc.visible_vram_size : end) - start;
250 * amdgpu_vram_mgr_bo_visible_size - CPU visible BO size
252 * @bo: &amdgpu_bo buffer object (must be in VRAM)
255 * How much of the given &amdgpu_bo buffer object lies in CPU visible VRAM.
257 u64 amdgpu_vram_mgr_bo_visible_size(struct amdgpu_bo *bo)
259 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
260 struct ttm_resource *mem = &bo->tbo.mem;
261 struct drm_mm_node *nodes = mem->mm_node;
262 unsigned pages = mem->num_pages;
265 if (amdgpu_gmc_vram_full_visible(&adev->gmc))
266 return amdgpu_bo_size(bo);
268 if (mem->start >= adev->gmc.visible_vram_size >> PAGE_SHIFT)
271 for (usage = 0; nodes && pages; pages -= nodes->size, nodes++)
272 usage += amdgpu_vram_mgr_vis_size(adev, nodes);
278 * amdgpu_vram_mgr_virt_start - update virtual start address
280 * @mem: ttm_resource to update
281 * @node: just allocated node
283 * Calculate a virtual BO start address to easily check if everything is CPU
286 static void amdgpu_vram_mgr_virt_start(struct ttm_resource *mem,
287 struct drm_mm_node *node)
291 start = node->start + node->size;
292 if (start > mem->num_pages)
293 start -= mem->num_pages;
296 mem->start = max(mem->start, start);
300 * amdgpu_vram_mgr_new - allocate new ranges
302 * @man: TTM memory type manager
303 * @tbo: TTM BO we need this range for
304 * @place: placement flags and restrictions
305 * @mem: the resulting mem object
307 * Allocate VRAM for the given BO.
309 static int amdgpu_vram_mgr_new(struct ttm_resource_manager *man,
310 struct ttm_buffer_object *tbo,
311 const struct ttm_place *place,
312 struct ttm_resource *mem)
314 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
315 struct amdgpu_device *adev = to_amdgpu_device(mgr);
316 struct drm_mm *mm = &mgr->mm;
317 struct drm_mm_node *nodes;
318 enum drm_mm_insert_mode mode;
319 unsigned long lpfn, num_nodes, pages_per_node, pages_left;
320 uint64_t vis_usage = 0, mem_bytes, max_bytes;
328 max_bytes = adev->gmc.mc_vram_size;
329 if (tbo->type != ttm_bo_type_kernel)
330 max_bytes -= AMDGPU_VM_RESERVED_VRAM;
332 /* bail out quickly if there's likely not enough VRAM for this BO */
333 mem_bytes = (u64)mem->num_pages << PAGE_SHIFT;
334 if (atomic64_add_return(mem_bytes, &mgr->usage) > max_bytes) {
335 atomic64_sub(mem_bytes, &mgr->usage);
339 if (place->flags & TTM_PL_FLAG_CONTIGUOUS) {
340 pages_per_node = ~0ul;
343 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
344 pages_per_node = HPAGE_PMD_NR;
347 pages_per_node = (2UL << (20UL - PAGE_SHIFT));
349 pages_per_node = max((uint32_t)pages_per_node, mem->page_alignment);
350 num_nodes = DIV_ROUND_UP(mem->num_pages, pages_per_node);
353 nodes = kvmalloc_array((uint32_t)num_nodes, sizeof(*nodes),
354 GFP_KERNEL | __GFP_ZERO);
356 atomic64_sub(mem_bytes, &mgr->usage);
360 mode = DRM_MM_INSERT_BEST;
361 if (place->flags & TTM_PL_FLAG_TOPDOWN)
362 mode = DRM_MM_INSERT_HIGH;
365 pages_left = mem->num_pages;
367 spin_lock(&mgr->lock);
368 for (i = 0; pages_left >= pages_per_node; ++i) {
369 unsigned long pages = rounddown_pow_of_two(pages_left);
371 r = drm_mm_insert_node_in_range(mm, &nodes[i], pages,
378 vis_usage += amdgpu_vram_mgr_vis_size(adev, &nodes[i]);
379 amdgpu_vram_mgr_virt_start(mem, &nodes[i]);
383 for (; pages_left; ++i) {
384 unsigned long pages = min(pages_left, pages_per_node);
385 uint32_t alignment = mem->page_alignment;
387 if (pages == pages_per_node)
388 alignment = pages_per_node;
390 r = drm_mm_insert_node_in_range(mm, &nodes[i],
397 vis_usage += amdgpu_vram_mgr_vis_size(adev, &nodes[i]);
398 amdgpu_vram_mgr_virt_start(mem, &nodes[i]);
401 spin_unlock(&mgr->lock);
403 atomic64_add(vis_usage, &mgr->vis_usage);
405 mem->mm_node = nodes;
411 drm_mm_remove_node(&nodes[i]);
412 spin_unlock(&mgr->lock);
413 atomic64_sub(mem->num_pages << PAGE_SHIFT, &mgr->usage);
420 * amdgpu_vram_mgr_del - free ranges
422 * @man: TTM memory type manager
423 * @mem: TTM memory object
425 * Free the allocated VRAM again.
427 static void amdgpu_vram_mgr_del(struct ttm_resource_manager *man,
428 struct ttm_resource *mem)
430 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
431 struct amdgpu_device *adev = to_amdgpu_device(mgr);
432 struct drm_mm_node *nodes = mem->mm_node;
433 uint64_t usage = 0, vis_usage = 0;
434 unsigned pages = mem->num_pages;
439 spin_lock(&mgr->lock);
441 pages -= nodes->size;
442 drm_mm_remove_node(nodes);
443 usage += nodes->size << PAGE_SHIFT;
444 vis_usage += amdgpu_vram_mgr_vis_size(adev, nodes);
447 spin_unlock(&mgr->lock);
449 atomic64_sub(usage, &mgr->usage);
450 atomic64_sub(vis_usage, &mgr->vis_usage);
452 kvfree(mem->mm_node);
457 * amdgpu_vram_mgr_alloc_sgt - allocate and fill a sg table
459 * @adev: amdgpu device pointer
460 * @mem: TTM memory object
461 * @dev: the other device
462 * @dir: dma direction
463 * @sgt: resulting sg table
465 * Allocate and fill a sg table from a VRAM allocation.
467 int amdgpu_vram_mgr_alloc_sgt(struct amdgpu_device *adev,
468 struct ttm_resource *mem,
470 enum dma_data_direction dir,
471 struct sg_table **sgt)
473 struct drm_mm_node *node;
474 struct scatterlist *sg;
479 *sgt = kmalloc(sizeof(**sgt), GFP_KERNEL);
483 for (pages = mem->num_pages, node = mem->mm_node;
484 pages; pages -= node->size, ++node)
487 r = sg_alloc_table(*sgt, num_entries, GFP_KERNEL);
491 for_each_sgtable_sg((*sgt), sg, i)
495 for_each_sgtable_sg((*sgt), sg, i) {
496 phys_addr_t phys = (node->start << PAGE_SHIFT) +
498 size_t size = node->size << PAGE_SHIFT;
502 addr = dma_map_resource(dev, phys, size, dir,
503 DMA_ATTR_SKIP_CPU_SYNC);
504 r = dma_mapping_error(dev, addr);
508 sg_set_page(sg, NULL, size, 0);
509 sg_dma_address(sg) = addr;
510 sg_dma_len(sg) = size;
515 for_each_sgtable_sg((*sgt), sg, i) {
519 dma_unmap_resource(dev, sg->dma_address,
521 DMA_ATTR_SKIP_CPU_SYNC);
531 * amdgpu_vram_mgr_alloc_sgt - allocate and fill a sg table
533 * @adev: amdgpu device pointer
534 * @sgt: sg table to free
536 * Free a previously allocate sg table.
538 void amdgpu_vram_mgr_free_sgt(struct amdgpu_device *adev,
540 enum dma_data_direction dir,
541 struct sg_table *sgt)
543 struct scatterlist *sg;
546 for_each_sgtable_sg(sgt, sg, i)
547 dma_unmap_resource(dev, sg->dma_address,
549 DMA_ATTR_SKIP_CPU_SYNC);
555 * amdgpu_vram_mgr_usage - how many bytes are used in this domain
557 * @man: TTM memory type manager
559 * Returns how many bytes are used in this domain.
561 uint64_t amdgpu_vram_mgr_usage(struct ttm_resource_manager *man)
563 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
565 return atomic64_read(&mgr->usage);
569 * amdgpu_vram_mgr_vis_usage - how many bytes are used in the visible part
571 * @man: TTM memory type manager
573 * Returns how many bytes are used in the visible part of VRAM
575 uint64_t amdgpu_vram_mgr_vis_usage(struct ttm_resource_manager *man)
577 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
579 return atomic64_read(&mgr->vis_usage);
583 * amdgpu_vram_mgr_debug - dump VRAM table
585 * @man: TTM memory type manager
586 * @printer: DRM printer to use
588 * Dump the table content using printk.
590 static void amdgpu_vram_mgr_debug(struct ttm_resource_manager *man,
591 struct drm_printer *printer)
593 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
595 spin_lock(&mgr->lock);
596 drm_mm_print(&mgr->mm, printer);
597 spin_unlock(&mgr->lock);
599 drm_printf(printer, "man size:%llu pages, ram usage:%lluMB, vis usage:%lluMB\n",
600 man->size, amdgpu_vram_mgr_usage(man) >> 20,
601 amdgpu_vram_mgr_vis_usage(man) >> 20);
604 static const struct ttm_resource_manager_func amdgpu_vram_mgr_func = {
605 .alloc = amdgpu_vram_mgr_new,
606 .free = amdgpu_vram_mgr_del,
607 .debug = amdgpu_vram_mgr_debug