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 struct amdgpu_vram_mgr {
32 struct ttm_resource_manager manager;
37 struct amdgpu_device *adev;
40 static inline struct amdgpu_vram_mgr *to_vram_mgr(struct ttm_resource_manager *man)
42 return container_of(man, struct amdgpu_vram_mgr, manager);
46 * DOC: mem_info_vram_total
48 * The amdgpu driver provides a sysfs API for reporting current total VRAM
49 * available on the device
50 * The file mem_info_vram_total is used for this and returns the total
51 * amount of VRAM in bytes
53 static ssize_t amdgpu_mem_info_vram_total_show(struct device *dev,
54 struct device_attribute *attr, char *buf)
56 struct drm_device *ddev = dev_get_drvdata(dev);
57 struct amdgpu_device *adev = ddev->dev_private;
59 return snprintf(buf, PAGE_SIZE, "%llu\n", adev->gmc.real_vram_size);
63 * DOC: mem_info_vis_vram_total
65 * The amdgpu driver provides a sysfs API for reporting current total
66 * visible VRAM available on the device
67 * The file mem_info_vis_vram_total is used for this and returns the total
68 * amount of visible VRAM in bytes
70 static ssize_t amdgpu_mem_info_vis_vram_total_show(struct device *dev,
71 struct device_attribute *attr, char *buf)
73 struct drm_device *ddev = dev_get_drvdata(dev);
74 struct amdgpu_device *adev = ddev->dev_private;
76 return snprintf(buf, PAGE_SIZE, "%llu\n", adev->gmc.visible_vram_size);
80 * DOC: mem_info_vram_used
82 * The amdgpu driver provides a sysfs API for reporting current total VRAM
83 * available on the device
84 * The file mem_info_vram_used is used for this and returns the total
85 * amount of currently used VRAM in bytes
87 static ssize_t amdgpu_mem_info_vram_used_show(struct device *dev,
88 struct device_attribute *attr, char *buf)
90 struct drm_device *ddev = dev_get_drvdata(dev);
91 struct amdgpu_device *adev = ddev->dev_private;
92 struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
93 return snprintf(buf, PAGE_SIZE, "%llu\n",
94 amdgpu_vram_mgr_usage(man));
98 * DOC: mem_info_vis_vram_used
100 * The amdgpu driver provides a sysfs API for reporting current total of
102 * The file mem_info_vis_vram_used is used for this and returns the total
103 * amount of currently used visible VRAM in bytes
105 static ssize_t amdgpu_mem_info_vis_vram_used_show(struct device *dev,
106 struct device_attribute *attr, char *buf)
108 struct drm_device *ddev = dev_get_drvdata(dev);
109 struct amdgpu_device *adev = ddev->dev_private;
110 struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
111 return snprintf(buf, PAGE_SIZE, "%llu\n",
112 amdgpu_vram_mgr_vis_usage(man));
115 static ssize_t amdgpu_mem_info_vram_vendor(struct device *dev,
116 struct device_attribute *attr,
119 struct drm_device *ddev = dev_get_drvdata(dev);
120 struct amdgpu_device *adev = ddev->dev_private;
122 switch (adev->gmc.vram_vendor) {
124 return snprintf(buf, PAGE_SIZE, "samsung\n");
126 return snprintf(buf, PAGE_SIZE, "infineon\n");
128 return snprintf(buf, PAGE_SIZE, "elpida\n");
130 return snprintf(buf, PAGE_SIZE, "etron\n");
132 return snprintf(buf, PAGE_SIZE, "nanya\n");
134 return snprintf(buf, PAGE_SIZE, "hynix\n");
136 return snprintf(buf, PAGE_SIZE, "mosel\n");
138 return snprintf(buf, PAGE_SIZE, "winbond\n");
140 return snprintf(buf, PAGE_SIZE, "esmt\n");
142 return snprintf(buf, PAGE_SIZE, "micron\n");
144 return snprintf(buf, PAGE_SIZE, "unknown\n");
148 static DEVICE_ATTR(mem_info_vram_total, S_IRUGO,
149 amdgpu_mem_info_vram_total_show, NULL);
150 static DEVICE_ATTR(mem_info_vis_vram_total, S_IRUGO,
151 amdgpu_mem_info_vis_vram_total_show,NULL);
152 static DEVICE_ATTR(mem_info_vram_used, S_IRUGO,
153 amdgpu_mem_info_vram_used_show, NULL);
154 static DEVICE_ATTR(mem_info_vis_vram_used, S_IRUGO,
155 amdgpu_mem_info_vis_vram_used_show, NULL);
156 static DEVICE_ATTR(mem_info_vram_vendor, S_IRUGO,
157 amdgpu_mem_info_vram_vendor, NULL);
159 static const struct attribute *amdgpu_vram_mgr_attributes[] = {
160 &dev_attr_mem_info_vram_total.attr,
161 &dev_attr_mem_info_vis_vram_total.attr,
162 &dev_attr_mem_info_vram_used.attr,
163 &dev_attr_mem_info_vis_vram_used.attr,
164 &dev_attr_mem_info_vram_vendor.attr,
168 static const struct ttm_resource_manager_func amdgpu_vram_mgr_func;
171 * amdgpu_vram_mgr_init - init VRAM manager and DRM MM
173 * @man: TTM memory type manager
174 * @p_size: maximum size of VRAM
176 * Allocate and initialize the VRAM manager.
178 int amdgpu_vram_mgr_init(struct amdgpu_device *adev)
180 struct ttm_resource_manager *man;
181 struct amdgpu_vram_mgr *mgr;
184 mgr = kzalloc(sizeof(*mgr), GFP_KERNEL);
190 man->available_caching = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC;
191 man->default_caching = TTM_PL_FLAG_WC;
193 ttm_resource_manager_init(man, adev->gmc.real_vram_size >> PAGE_SHIFT);
195 man->func = &amdgpu_vram_mgr_func;
197 drm_mm_init(&mgr->mm, 0, man->size);
198 spin_lock_init(&mgr->lock);
202 /* Add the two VRAM-related sysfs files */
203 ret = sysfs_create_files(&adev->dev->kobj, amdgpu_vram_mgr_attributes);
205 DRM_ERROR("Failed to register sysfs\n");
207 ttm_set_driver_manager(&adev->mman.bdev, TTM_PL_VRAM, &mgr->manager);
208 ttm_resource_manager_set_used(man, true);
213 * amdgpu_vram_mgr_fini - free and destroy VRAM manager
215 * @man: TTM memory type manager
217 * Destroy and free the VRAM manager, returns -EBUSY if ranges are still
218 * allocated inside it.
220 void amdgpu_vram_mgr_fini(struct amdgpu_device *adev)
222 struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
223 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
226 ttm_resource_manager_set_used(man, false);
228 ret = ttm_resource_manager_force_list_clean(&adev->mman.bdev, man);
232 spin_lock(&mgr->lock);
233 drm_mm_takedown(&mgr->mm);
234 spin_unlock(&mgr->lock);
236 sysfs_remove_files(&adev->dev->kobj, amdgpu_vram_mgr_attributes);
238 ttm_resource_manager_cleanup(man);
239 ttm_set_driver_manager(&adev->mman.bdev, TTM_PL_VRAM, NULL);
244 * amdgpu_vram_mgr_vis_size - Calculate visible node size
246 * @adev: amdgpu device structure
247 * @node: MM node structure
249 * Calculate how many bytes of the MM node are inside visible VRAM
251 static u64 amdgpu_vram_mgr_vis_size(struct amdgpu_device *adev,
252 struct drm_mm_node *node)
254 uint64_t start = node->start << PAGE_SHIFT;
255 uint64_t end = (node->size + node->start) << PAGE_SHIFT;
257 if (start >= adev->gmc.visible_vram_size)
260 return (end > adev->gmc.visible_vram_size ?
261 adev->gmc.visible_vram_size : end) - start;
265 * amdgpu_vram_mgr_bo_visible_size - CPU visible BO size
267 * @bo: &amdgpu_bo buffer object (must be in VRAM)
270 * How much of the given &amdgpu_bo buffer object lies in CPU visible VRAM.
272 u64 amdgpu_vram_mgr_bo_visible_size(struct amdgpu_bo *bo)
274 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
275 struct ttm_mem_reg *mem = &bo->tbo.mem;
276 struct drm_mm_node *nodes = mem->mm_node;
277 unsigned pages = mem->num_pages;
280 if (amdgpu_gmc_vram_full_visible(&adev->gmc))
281 return amdgpu_bo_size(bo);
283 if (mem->start >= adev->gmc.visible_vram_size >> PAGE_SHIFT)
286 for (usage = 0; nodes && pages; pages -= nodes->size, nodes++)
287 usage += amdgpu_vram_mgr_vis_size(adev, nodes);
293 * amdgpu_vram_mgr_virt_start - update virtual start address
295 * @mem: ttm_mem_reg to update
296 * @node: just allocated node
298 * Calculate a virtual BO start address to easily check if everything is CPU
301 static void amdgpu_vram_mgr_virt_start(struct ttm_mem_reg *mem,
302 struct drm_mm_node *node)
306 start = node->start + node->size;
307 if (start > mem->num_pages)
308 start -= mem->num_pages;
311 mem->start = max(mem->start, start);
315 * amdgpu_vram_mgr_new - allocate new ranges
317 * @man: TTM memory type manager
318 * @tbo: TTM BO we need this range for
319 * @place: placement flags and restrictions
320 * @mem: the resulting mem object
322 * Allocate VRAM for the given BO.
324 static int amdgpu_vram_mgr_new(struct ttm_resource_manager *man,
325 struct ttm_buffer_object *tbo,
326 const struct ttm_place *place,
327 struct ttm_mem_reg *mem)
329 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
330 struct amdgpu_device *adev = mgr->adev;
331 struct drm_mm *mm = &mgr->mm;
332 struct drm_mm_node *nodes;
333 enum drm_mm_insert_mode mode;
334 unsigned long lpfn, num_nodes, pages_per_node, pages_left;
335 uint64_t vis_usage = 0, mem_bytes, max_bytes;
343 max_bytes = adev->gmc.mc_vram_size;
344 if (tbo->type != ttm_bo_type_kernel)
345 max_bytes -= AMDGPU_VM_RESERVED_VRAM;
347 /* bail out quickly if there's likely not enough VRAM for this BO */
348 mem_bytes = (u64)mem->num_pages << PAGE_SHIFT;
349 if (atomic64_add_return(mem_bytes, &mgr->usage) > max_bytes) {
350 atomic64_sub(mem_bytes, &mgr->usage);
354 if (place->flags & TTM_PL_FLAG_CONTIGUOUS) {
355 pages_per_node = ~0ul;
358 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
359 pages_per_node = HPAGE_PMD_NR;
362 pages_per_node = (2UL << (20UL - PAGE_SHIFT));
364 pages_per_node = max((uint32_t)pages_per_node, mem->page_alignment);
365 num_nodes = DIV_ROUND_UP(mem->num_pages, pages_per_node);
368 nodes = kvmalloc_array((uint32_t)num_nodes, sizeof(*nodes),
369 GFP_KERNEL | __GFP_ZERO);
371 atomic64_sub(mem_bytes, &mgr->usage);
375 mode = DRM_MM_INSERT_BEST;
376 if (place->flags & TTM_PL_FLAG_TOPDOWN)
377 mode = DRM_MM_INSERT_HIGH;
380 pages_left = mem->num_pages;
382 spin_lock(&mgr->lock);
383 for (i = 0; pages_left >= pages_per_node; ++i) {
384 unsigned long pages = rounddown_pow_of_two(pages_left);
386 r = drm_mm_insert_node_in_range(mm, &nodes[i], pages,
393 vis_usage += amdgpu_vram_mgr_vis_size(adev, &nodes[i]);
394 amdgpu_vram_mgr_virt_start(mem, &nodes[i]);
398 for (; pages_left; ++i) {
399 unsigned long pages = min(pages_left, pages_per_node);
400 uint32_t alignment = mem->page_alignment;
402 if (pages == pages_per_node)
403 alignment = pages_per_node;
405 r = drm_mm_insert_node_in_range(mm, &nodes[i],
412 vis_usage += amdgpu_vram_mgr_vis_size(adev, &nodes[i]);
413 amdgpu_vram_mgr_virt_start(mem, &nodes[i]);
416 spin_unlock(&mgr->lock);
418 atomic64_add(vis_usage, &mgr->vis_usage);
420 mem->mm_node = nodes;
426 drm_mm_remove_node(&nodes[i]);
427 spin_unlock(&mgr->lock);
428 atomic64_sub(mem->num_pages << PAGE_SHIFT, &mgr->usage);
435 * amdgpu_vram_mgr_del - free ranges
437 * @man: TTM memory type manager
438 * @tbo: TTM BO we need this range for
439 * @place: placement flags and restrictions
440 * @mem: TTM memory object
442 * Free the allocated VRAM again.
444 static void amdgpu_vram_mgr_del(struct ttm_resource_manager *man,
445 struct ttm_mem_reg *mem)
447 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
448 struct amdgpu_device *adev = mgr->adev;
449 struct drm_mm_node *nodes = mem->mm_node;
450 uint64_t usage = 0, vis_usage = 0;
451 unsigned pages = mem->num_pages;
456 spin_lock(&mgr->lock);
458 pages -= nodes->size;
459 drm_mm_remove_node(nodes);
460 usage += nodes->size << PAGE_SHIFT;
461 vis_usage += amdgpu_vram_mgr_vis_size(adev, nodes);
464 spin_unlock(&mgr->lock);
466 atomic64_sub(usage, &mgr->usage);
467 atomic64_sub(vis_usage, &mgr->vis_usage);
469 kvfree(mem->mm_node);
474 * amdgpu_vram_mgr_alloc_sgt - allocate and fill a sg table
476 * @adev: amdgpu device pointer
477 * @mem: TTM memory object
478 * @dev: the other device
479 * @dir: dma direction
480 * @sgt: resulting sg table
482 * Allocate and fill a sg table from a VRAM allocation.
484 int amdgpu_vram_mgr_alloc_sgt(struct amdgpu_device *adev,
485 struct ttm_mem_reg *mem,
487 enum dma_data_direction dir,
488 struct sg_table **sgt)
490 struct drm_mm_node *node;
491 struct scatterlist *sg;
496 *sgt = kmalloc(sizeof(*sg), GFP_KERNEL);
500 for (pages = mem->num_pages, node = mem->mm_node;
501 pages; pages -= node->size, ++node)
504 r = sg_alloc_table(*sgt, num_entries, GFP_KERNEL);
508 for_each_sgtable_sg((*sgt), sg, i)
512 for_each_sgtable_sg((*sgt), sg, i) {
513 phys_addr_t phys = (node->start << PAGE_SHIFT) +
515 size_t size = node->size << PAGE_SHIFT;
519 addr = dma_map_resource(dev, phys, size, dir,
520 DMA_ATTR_SKIP_CPU_SYNC);
521 r = dma_mapping_error(dev, addr);
525 sg_set_page(sg, NULL, size, 0);
526 sg_dma_address(sg) = addr;
527 sg_dma_len(sg) = size;
532 for_each_sgtable_sg((*sgt), sg, i) {
536 dma_unmap_resource(dev, sg->dma_address,
538 DMA_ATTR_SKIP_CPU_SYNC);
548 * amdgpu_vram_mgr_alloc_sgt - allocate and fill a sg table
550 * @adev: amdgpu device pointer
551 * @sgt: sg table to free
553 * Free a previously allocate sg table.
555 void amdgpu_vram_mgr_free_sgt(struct amdgpu_device *adev,
557 enum dma_data_direction dir,
558 struct sg_table *sgt)
560 struct scatterlist *sg;
563 for_each_sgtable_sg(sgt, sg, i)
564 dma_unmap_resource(dev, sg->dma_address,
566 DMA_ATTR_SKIP_CPU_SYNC);
572 * amdgpu_vram_mgr_usage - how many bytes are used in this domain
574 * @man: TTM memory type manager
576 * Returns how many bytes are used in this domain.
578 uint64_t amdgpu_vram_mgr_usage(struct ttm_resource_manager *man)
580 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
582 return atomic64_read(&mgr->usage);
586 * amdgpu_vram_mgr_vis_usage - how many bytes are used in the visible part
588 * @man: TTM memory type manager
590 * Returns how many bytes are used in the visible part of VRAM
592 uint64_t amdgpu_vram_mgr_vis_usage(struct ttm_resource_manager *man)
594 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
596 return atomic64_read(&mgr->vis_usage);
600 * amdgpu_vram_mgr_debug - dump VRAM table
602 * @man: TTM memory type manager
603 * @printer: DRM printer to use
605 * Dump the table content using printk.
607 static void amdgpu_vram_mgr_debug(struct ttm_resource_manager *man,
608 struct drm_printer *printer)
610 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
612 spin_lock(&mgr->lock);
613 drm_mm_print(&mgr->mm, printer);
614 spin_unlock(&mgr->lock);
616 drm_printf(printer, "man size:%llu pages, ram usage:%lluMB, vis usage:%lluMB\n",
617 man->size, amdgpu_vram_mgr_usage(man) >> 20,
618 amdgpu_vram_mgr_vis_usage(man) >> 20);
621 static const struct ttm_resource_manager_func amdgpu_vram_mgr_func = {
622 .get_node = amdgpu_vram_mgr_new,
623 .put_node = amdgpu_vram_mgr_del,
624 .debug = amdgpu_vram_mgr_debug