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_res_cursor.h"
29 #include "amdgpu_atomfirmware.h"
32 static inline struct amdgpu_vram_mgr *to_vram_mgr(struct ttm_resource_manager *man)
34 return container_of(man, struct amdgpu_vram_mgr, manager);
37 static inline struct amdgpu_device *to_amdgpu_device(struct amdgpu_vram_mgr *mgr)
39 return container_of(mgr, struct amdgpu_device, mman.vram_mgr);
43 * DOC: mem_info_vram_total
45 * The amdgpu driver provides a sysfs API for reporting current total VRAM
46 * available on the device
47 * The file mem_info_vram_total is used for this and returns the total
48 * amount of VRAM in bytes
50 static ssize_t amdgpu_mem_info_vram_total_show(struct device *dev,
51 struct device_attribute *attr, char *buf)
53 struct drm_device *ddev = dev_get_drvdata(dev);
54 struct amdgpu_device *adev = drm_to_adev(ddev);
56 return sysfs_emit(buf, "%llu\n", adev->gmc.real_vram_size);
60 * DOC: mem_info_vis_vram_total
62 * The amdgpu driver provides a sysfs API for reporting current total
63 * visible VRAM available on the device
64 * The file mem_info_vis_vram_total is used for this and returns the total
65 * amount of visible VRAM in bytes
67 static ssize_t amdgpu_mem_info_vis_vram_total_show(struct device *dev,
68 struct device_attribute *attr, char *buf)
70 struct drm_device *ddev = dev_get_drvdata(dev);
71 struct amdgpu_device *adev = drm_to_adev(ddev);
73 return sysfs_emit(buf, "%llu\n", adev->gmc.visible_vram_size);
77 * DOC: mem_info_vram_used
79 * The amdgpu driver provides a sysfs API for reporting current total VRAM
80 * available on the device
81 * The file mem_info_vram_used is used for this and returns the total
82 * amount of currently used VRAM in bytes
84 static ssize_t amdgpu_mem_info_vram_used_show(struct device *dev,
85 struct device_attribute *attr, char *buf)
87 struct drm_device *ddev = dev_get_drvdata(dev);
88 struct amdgpu_device *adev = drm_to_adev(ddev);
89 struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
91 return sysfs_emit(buf, "%llu\n", 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 sysfs_emit(buf, "%llu\n", amdgpu_vram_mgr_vis_usage(man));
112 static ssize_t amdgpu_mem_info_vram_vendor(struct device *dev,
113 struct device_attribute *attr,
116 struct drm_device *ddev = dev_get_drvdata(dev);
117 struct amdgpu_device *adev = drm_to_adev(ddev);
119 switch (adev->gmc.vram_vendor) {
121 return sysfs_emit(buf, "samsung\n");
123 return sysfs_emit(buf, "infineon\n");
125 return sysfs_emit(buf, "elpida\n");
127 return sysfs_emit(buf, "etron\n");
129 return sysfs_emit(buf, "nanya\n");
131 return sysfs_emit(buf, "hynix\n");
133 return sysfs_emit(buf, "mosel\n");
135 return sysfs_emit(buf, "winbond\n");
137 return sysfs_emit(buf, "esmt\n");
139 return sysfs_emit(buf, "micron\n");
141 return sysfs_emit(buf, "unknown\n");
145 static DEVICE_ATTR(mem_info_vram_total, S_IRUGO,
146 amdgpu_mem_info_vram_total_show, NULL);
147 static DEVICE_ATTR(mem_info_vis_vram_total, S_IRUGO,
148 amdgpu_mem_info_vis_vram_total_show,NULL);
149 static DEVICE_ATTR(mem_info_vram_used, S_IRUGO,
150 amdgpu_mem_info_vram_used_show, NULL);
151 static DEVICE_ATTR(mem_info_vis_vram_used, S_IRUGO,
152 amdgpu_mem_info_vis_vram_used_show, NULL);
153 static DEVICE_ATTR(mem_info_vram_vendor, S_IRUGO,
154 amdgpu_mem_info_vram_vendor, NULL);
156 static const struct attribute *amdgpu_vram_mgr_attributes[] = {
157 &dev_attr_mem_info_vram_total.attr,
158 &dev_attr_mem_info_vis_vram_total.attr,
159 &dev_attr_mem_info_vram_used.attr,
160 &dev_attr_mem_info_vis_vram_used.attr,
161 &dev_attr_mem_info_vram_vendor.attr,
165 static const struct ttm_resource_manager_func amdgpu_vram_mgr_func;
168 * amdgpu_vram_mgr_init - init VRAM manager and DRM MM
170 * @adev: amdgpu_device pointer
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 ttm_resource_manager_init(man, adev->gmc.real_vram_size >> PAGE_SHIFT);
182 man->func = &amdgpu_vram_mgr_func;
184 drm_mm_init(&mgr->mm, 0, man->size);
185 spin_lock_init(&mgr->lock);
186 INIT_LIST_HEAD(&mgr->reservations_pending);
187 INIT_LIST_HEAD(&mgr->reserved_pages);
189 /* Add the two VRAM-related sysfs files */
190 ret = sysfs_create_files(&adev->dev->kobj, amdgpu_vram_mgr_attributes);
192 DRM_ERROR("Failed to register sysfs\n");
194 ttm_set_driver_manager(&adev->mman.bdev, TTM_PL_VRAM, &mgr->manager);
195 ttm_resource_manager_set_used(man, true);
200 * amdgpu_vram_mgr_fini - free and destroy VRAM manager
202 * @adev: amdgpu_device pointer
204 * Destroy and free the VRAM manager, returns -EBUSY if ranges are still
205 * allocated inside it.
207 void amdgpu_vram_mgr_fini(struct amdgpu_device *adev)
209 struct amdgpu_vram_mgr *mgr = &adev->mman.vram_mgr;
210 struct ttm_resource_manager *man = &mgr->manager;
212 struct amdgpu_vram_reservation *rsv, *temp;
214 ttm_resource_manager_set_used(man, false);
216 ret = ttm_resource_manager_evict_all(&adev->mman.bdev, man);
220 spin_lock(&mgr->lock);
221 list_for_each_entry_safe(rsv, temp, &mgr->reservations_pending, node)
224 list_for_each_entry_safe(rsv, temp, &mgr->reserved_pages, node) {
225 drm_mm_remove_node(&rsv->mm_node);
228 drm_mm_takedown(&mgr->mm);
229 spin_unlock(&mgr->lock);
231 sysfs_remove_files(&adev->dev->kobj, amdgpu_vram_mgr_attributes);
233 ttm_resource_manager_cleanup(man);
234 ttm_set_driver_manager(&adev->mman.bdev, TTM_PL_VRAM, NULL);
238 * amdgpu_vram_mgr_vis_size - Calculate visible node size
240 * @adev: amdgpu_device pointer
241 * @node: MM node structure
243 * Calculate how many bytes of the MM node are inside visible VRAM
245 static u64 amdgpu_vram_mgr_vis_size(struct amdgpu_device *adev,
246 struct drm_mm_node *node)
248 uint64_t start = node->start << PAGE_SHIFT;
249 uint64_t end = (node->size + node->start) << PAGE_SHIFT;
251 if (start >= adev->gmc.visible_vram_size)
254 return (end > adev->gmc.visible_vram_size ?
255 adev->gmc.visible_vram_size : end) - start;
259 * amdgpu_vram_mgr_bo_visible_size - CPU visible BO size
261 * @bo: &amdgpu_bo buffer object (must be in VRAM)
264 * How much of the given &amdgpu_bo buffer object lies in CPU visible VRAM.
266 u64 amdgpu_vram_mgr_bo_visible_size(struct amdgpu_bo *bo)
268 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
269 struct ttm_resource *mem = &bo->tbo.mem;
270 struct drm_mm_node *nodes = mem->mm_node;
271 unsigned pages = mem->num_pages;
274 if (amdgpu_gmc_vram_full_visible(&adev->gmc))
275 return amdgpu_bo_size(bo);
277 if (mem->start >= adev->gmc.visible_vram_size >> PAGE_SHIFT)
280 for (usage = 0; nodes && pages; pages -= nodes->size, nodes++)
281 usage += amdgpu_vram_mgr_vis_size(adev, nodes);
286 static void amdgpu_vram_mgr_do_reserve(struct ttm_resource_manager *man)
288 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
289 struct amdgpu_device *adev = to_amdgpu_device(mgr);
290 struct drm_mm *mm = &mgr->mm;
291 struct amdgpu_vram_reservation *rsv, *temp;
294 list_for_each_entry_safe(rsv, temp, &mgr->reservations_pending, node) {
295 if (drm_mm_reserve_node(mm, &rsv->mm_node))
298 dev_dbg(adev->dev, "Reservation 0x%llx - %lld, Succeeded\n",
299 rsv->mm_node.start, rsv->mm_node.size);
301 vis_usage = amdgpu_vram_mgr_vis_size(adev, &rsv->mm_node);
302 atomic64_add(vis_usage, &mgr->vis_usage);
303 atomic64_add(rsv->mm_node.size << PAGE_SHIFT, &mgr->usage);
304 list_move(&rsv->node, &mgr->reserved_pages);
309 * amdgpu_vram_mgr_reserve_range - Reserve a range from VRAM
311 * @man: TTM memory type manager
312 * @start: start address of the range in VRAM
313 * @size: size of the range
315 * Reserve memory from start addess with the specified size in VRAM
317 int amdgpu_vram_mgr_reserve_range(struct ttm_resource_manager *man,
318 uint64_t start, uint64_t size)
320 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
321 struct amdgpu_vram_reservation *rsv;
323 rsv = kzalloc(sizeof(*rsv), GFP_KERNEL);
327 INIT_LIST_HEAD(&rsv->node);
328 rsv->mm_node.start = start >> PAGE_SHIFT;
329 rsv->mm_node.size = size >> PAGE_SHIFT;
331 spin_lock(&mgr->lock);
332 list_add_tail(&mgr->reservations_pending, &rsv->node);
333 amdgpu_vram_mgr_do_reserve(man);
334 spin_unlock(&mgr->lock);
340 * amdgpu_vram_mgr_query_page_status - query the reservation status
342 * @man: TTM memory type manager
343 * @start: start address of a page in VRAM
346 * -EBUSY: the page is still hold and in pending list
347 * 0: the page has been reserved
348 * -ENOENT: the input page is not a reservation
350 int amdgpu_vram_mgr_query_page_status(struct ttm_resource_manager *man,
353 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
354 struct amdgpu_vram_reservation *rsv;
357 spin_lock(&mgr->lock);
359 list_for_each_entry(rsv, &mgr->reservations_pending, node) {
360 if ((rsv->mm_node.start <= start) &&
361 (start < (rsv->mm_node.start + rsv->mm_node.size))) {
367 list_for_each_entry(rsv, &mgr->reserved_pages, node) {
368 if ((rsv->mm_node.start <= start) &&
369 (start < (rsv->mm_node.start + rsv->mm_node.size))) {
377 spin_unlock(&mgr->lock);
382 * amdgpu_vram_mgr_virt_start - update virtual start address
384 * @mem: ttm_resource to update
385 * @node: just allocated node
387 * Calculate a virtual BO start address to easily check if everything is CPU
390 static void amdgpu_vram_mgr_virt_start(struct ttm_resource *mem,
391 struct drm_mm_node *node)
395 start = node->start + node->size;
396 if (start > mem->num_pages)
397 start -= mem->num_pages;
400 mem->start = max(mem->start, start);
404 * amdgpu_vram_mgr_new - allocate new ranges
406 * @man: TTM memory type manager
407 * @tbo: TTM BO we need this range for
408 * @place: placement flags and restrictions
409 * @mem: the resulting mem object
411 * Allocate VRAM for the given BO.
413 static int amdgpu_vram_mgr_new(struct ttm_resource_manager *man,
414 struct ttm_buffer_object *tbo,
415 const struct ttm_place *place,
416 struct ttm_resource *mem)
418 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
419 struct amdgpu_device *adev = to_amdgpu_device(mgr);
420 struct drm_mm *mm = &mgr->mm;
421 struct drm_mm_node *nodes;
422 enum drm_mm_insert_mode mode;
423 unsigned long lpfn, num_nodes, pages_per_node, pages_left;
424 uint64_t vis_usage = 0, mem_bytes, max_bytes;
432 max_bytes = adev->gmc.mc_vram_size;
433 if (tbo->type != ttm_bo_type_kernel)
434 max_bytes -= AMDGPU_VM_RESERVED_VRAM;
436 /* bail out quickly if there's likely not enough VRAM for this BO */
437 mem_bytes = (u64)mem->num_pages << PAGE_SHIFT;
438 if (atomic64_add_return(mem_bytes, &mgr->usage) > max_bytes) {
439 atomic64_sub(mem_bytes, &mgr->usage);
443 if (place->flags & TTM_PL_FLAG_CONTIGUOUS) {
444 pages_per_node = ~0ul;
447 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
448 pages_per_node = HPAGE_PMD_NR;
451 pages_per_node = (2UL << (20UL - PAGE_SHIFT));
453 pages_per_node = max((uint32_t)pages_per_node, mem->page_alignment);
454 num_nodes = DIV_ROUND_UP(mem->num_pages, pages_per_node);
457 nodes = kvmalloc_array((uint32_t)num_nodes, sizeof(*nodes),
458 GFP_KERNEL | __GFP_ZERO);
460 atomic64_sub(mem_bytes, &mgr->usage);
464 mode = DRM_MM_INSERT_BEST;
465 if (place->flags & TTM_PL_FLAG_TOPDOWN)
466 mode = DRM_MM_INSERT_HIGH;
469 pages_left = mem->num_pages;
471 spin_lock(&mgr->lock);
472 for (i = 0; pages_left >= pages_per_node; ++i) {
473 unsigned long pages = rounddown_pow_of_two(pages_left);
475 /* Limit maximum size to 2GB due to SG table limitations */
476 pages = min(pages, (2UL << (30 - PAGE_SHIFT)));
478 r = drm_mm_insert_node_in_range(mm, &nodes[i], pages,
485 vis_usage += amdgpu_vram_mgr_vis_size(adev, &nodes[i]);
486 amdgpu_vram_mgr_virt_start(mem, &nodes[i]);
490 for (; pages_left; ++i) {
491 unsigned long pages = min(pages_left, pages_per_node);
492 uint32_t alignment = mem->page_alignment;
494 if (pages == pages_per_node)
495 alignment = pages_per_node;
497 r = drm_mm_insert_node_in_range(mm, &nodes[i],
504 vis_usage += amdgpu_vram_mgr_vis_size(adev, &nodes[i]);
505 amdgpu_vram_mgr_virt_start(mem, &nodes[i]);
508 spin_unlock(&mgr->lock);
510 atomic64_add(vis_usage, &mgr->vis_usage);
512 mem->mm_node = nodes;
518 drm_mm_remove_node(&nodes[i]);
519 spin_unlock(&mgr->lock);
520 atomic64_sub(mem->num_pages << PAGE_SHIFT, &mgr->usage);
527 * amdgpu_vram_mgr_del - free ranges
529 * @man: TTM memory type manager
530 * @mem: TTM memory object
532 * Free the allocated VRAM again.
534 static void amdgpu_vram_mgr_del(struct ttm_resource_manager *man,
535 struct ttm_resource *mem)
537 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
538 struct amdgpu_device *adev = to_amdgpu_device(mgr);
539 struct drm_mm_node *nodes = mem->mm_node;
540 uint64_t usage = 0, vis_usage = 0;
541 unsigned pages = mem->num_pages;
546 spin_lock(&mgr->lock);
548 pages -= nodes->size;
549 drm_mm_remove_node(nodes);
550 usage += nodes->size << PAGE_SHIFT;
551 vis_usage += amdgpu_vram_mgr_vis_size(adev, nodes);
554 amdgpu_vram_mgr_do_reserve(man);
555 spin_unlock(&mgr->lock);
557 atomic64_sub(usage, &mgr->usage);
558 atomic64_sub(vis_usage, &mgr->vis_usage);
560 kvfree(mem->mm_node);
565 * amdgpu_vram_mgr_alloc_sgt - allocate and fill a sg table
567 * @adev: amdgpu device pointer
568 * @mem: TTM memory object
569 * @offset: byte offset from the base of VRAM BO
570 * @length: number of bytes to export in sg_table
571 * @dev: the other device
572 * @dir: dma direction
573 * @sgt: resulting sg table
575 * Allocate and fill a sg table from a VRAM allocation.
577 int amdgpu_vram_mgr_alloc_sgt(struct amdgpu_device *adev,
578 struct ttm_resource *mem,
579 u64 offset, u64 length,
581 enum dma_data_direction dir,
582 struct sg_table **sgt)
584 struct amdgpu_res_cursor cursor;
585 struct scatterlist *sg;
589 *sgt = kmalloc(sizeof(**sgt), GFP_KERNEL);
593 /* Determine the number of DRM_MM nodes to export */
594 amdgpu_res_first(mem, offset, length, &cursor);
595 while (cursor.remaining) {
597 amdgpu_res_next(&cursor, cursor.size);
600 r = sg_alloc_table(*sgt, num_entries, GFP_KERNEL);
604 /* Initialize scatterlist nodes of sg_table */
605 for_each_sgtable_sg((*sgt), sg, i)
609 * Walk down DRM_MM nodes to populate scatterlist nodes
610 * @note: Use iterator api to get first the DRM_MM node
611 * and the number of bytes from it. Access the following
612 * DRM_MM node(s) if more buffer needs to exported
614 amdgpu_res_first(mem, offset, length, &cursor);
615 for_each_sgtable_sg((*sgt), sg, i) {
616 phys_addr_t phys = cursor.start + adev->gmc.aper_base;
617 size_t size = cursor.size;
620 addr = dma_map_resource(dev, phys, size, dir,
621 DMA_ATTR_SKIP_CPU_SYNC);
622 r = dma_mapping_error(dev, addr);
626 sg_set_page(sg, NULL, size, 0);
627 sg_dma_address(sg) = addr;
628 sg_dma_len(sg) = size;
630 amdgpu_res_next(&cursor, cursor.size);
636 for_each_sgtable_sg((*sgt), sg, i) {
640 dma_unmap_resource(dev, sg->dma_address,
642 DMA_ATTR_SKIP_CPU_SYNC);
652 * amdgpu_vram_mgr_free_sgt - allocate and fill a sg table
654 * @dev: device pointer
655 * @dir: data direction of resource to unmap
656 * @sgt: sg table to free
658 * Free a previously allocate sg table.
660 void amdgpu_vram_mgr_free_sgt(struct device *dev,
661 enum dma_data_direction dir,
662 struct sg_table *sgt)
664 struct scatterlist *sg;
667 for_each_sgtable_sg(sgt, sg, i)
668 dma_unmap_resource(dev, sg->dma_address,
670 DMA_ATTR_SKIP_CPU_SYNC);
676 * amdgpu_vram_mgr_usage - how many bytes are used in this domain
678 * @man: TTM memory type manager
680 * Returns how many bytes are used in this domain.
682 uint64_t amdgpu_vram_mgr_usage(struct ttm_resource_manager *man)
684 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
686 return atomic64_read(&mgr->usage);
690 * amdgpu_vram_mgr_vis_usage - how many bytes are used in the visible part
692 * @man: TTM memory type manager
694 * Returns how many bytes are used in the visible part of VRAM
696 uint64_t amdgpu_vram_mgr_vis_usage(struct ttm_resource_manager *man)
698 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
700 return atomic64_read(&mgr->vis_usage);
704 * amdgpu_vram_mgr_debug - dump VRAM table
706 * @man: TTM memory type manager
707 * @printer: DRM printer to use
709 * Dump the table content using printk.
711 static void amdgpu_vram_mgr_debug(struct ttm_resource_manager *man,
712 struct drm_printer *printer)
714 struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
716 spin_lock(&mgr->lock);
717 drm_mm_print(&mgr->mm, printer);
718 spin_unlock(&mgr->lock);
720 drm_printf(printer, "man size:%llu pages, ram usage:%lluMB, vis usage:%lluMB\n",
721 man->size, amdgpu_vram_mgr_usage(man) >> 20,
722 amdgpu_vram_mgr_vis_usage(man) >> 20);
725 static const struct ttm_resource_manager_func amdgpu_vram_mgr_func = {
726 .alloc = amdgpu_vram_mgr_new,
727 .free = amdgpu_vram_mgr_del,
728 .debug = amdgpu_vram_mgr_debug