1/*
2 * Copyright 2016 Advanced Micro Devices, Inc.
3 *
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:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
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.
21 *
22 * Authors: Christian König
23 */
24
25#include <drm/drmP.h>
26#include "amdgpu.h"
27
28struct amdgpu_gtt_mgr {
29 struct drm_mm mm;
30 spinlock_t lock;
31 atomic64_t available;
32};
33
34struct amdgpu_gtt_node {
35 struct drm_mm_node node;
36 struct ttm_buffer_object *tbo;
37};
38
39/**
40 * amdgpu_gtt_mgr_init - init GTT manager and DRM MM
41 *
42 * @man: TTM memory type manager
43 * @p_size: maximum size of GTT
44 *
45 * Allocate and initialize the GTT manager.
46 */
47static int amdgpu_gtt_mgr_init(struct ttm_mem_type_manager *man,
48 unsigned long p_size)
49{
50 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev);
51 struct amdgpu_gtt_mgr *mgr;
52 uint64_t start, size;
53
54 mgr = kzalloc(sizeof(*mgr), GFP_KERNEL);
55 if (!mgr)
56 return -ENOMEM;
57
58 start = AMDGPU_GTT_MAX_TRANSFER_SIZE * AMDGPU_GTT_NUM_TRANSFER_WINDOWS;
59 size = (adev->gmc.gart_size >> PAGE_SHIFT) - start;
60 drm_mm_init(&mgr->mm, start, size);
61 spin_lock_init(&mgr->lock);
62 atomic64_set(&mgr->available, p_size);
63 man->priv = mgr;
64 return 0;
65}
66
67/**
68 * amdgpu_gtt_mgr_fini - free and destroy GTT manager
69 *
70 * @man: TTM memory type manager
71 *
72 * Destroy and free the GTT manager, returns -EBUSY if ranges are still
73 * allocated inside it.
74 */
75static int amdgpu_gtt_mgr_fini(struct ttm_mem_type_manager *man)
76{
77 struct amdgpu_gtt_mgr *mgr = man->priv;
78 spin_lock(&mgr->lock);
79 drm_mm_takedown(&mgr->mm);
80 spin_unlock(&mgr->lock);
81 kfree(mgr);
82 man->priv = NULL;
83 return 0;
84}
85
86/**
87 * amdgpu_gtt_mgr_has_gart_addr - Check if mem has address space
88 *
89 * @mem: the mem object to check
90 *
91 * Check if a mem object has already address space allocated.
92 */
93bool amdgpu_gtt_mgr_has_gart_addr(struct ttm_mem_reg *mem)
94{
95 struct amdgpu_gtt_node *node = mem->mm_node;
96
97 return (node->node.start != AMDGPU_BO_INVALID_OFFSET);
98}
99
100/**
101 * amdgpu_gtt_mgr_alloc - allocate new ranges
102 *
103 * @man: TTM memory type manager
104 * @tbo: TTM BO we need this range for
105 * @place: placement flags and restrictions
106 * @mem: the resulting mem object
107 *
108 * Allocate the address space for a node.
109 */
110static int amdgpu_gtt_mgr_alloc(struct ttm_mem_type_manager *man,
111 struct ttm_buffer_object *tbo,
112 const struct ttm_place *place,
113 struct ttm_mem_reg *mem)
114{
115 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev);
116 struct amdgpu_gtt_mgr *mgr = man->priv;
117 struct amdgpu_gtt_node *node = mem->mm_node;
118 enum drm_mm_insert_mode mode;
119 unsigned long fpfn, lpfn;
120 int r;
121
122 if (amdgpu_gtt_mgr_has_gart_addr(mem))
123 return 0;
124
125 if (place)
126 fpfn = place->fpfn;
127 else
128 fpfn = 0;
129
130 if (place && place->lpfn)
131 lpfn = place->lpfn;
132 else
133 lpfn = adev->gart.num_cpu_pages;
134
135 mode = DRM_MM_INSERT_BEST;
136 if (place && place->flags & TTM_PL_FLAG_TOPDOWN)
137 mode = DRM_MM_INSERT_HIGH;
138
139 spin_lock(&mgr->lock);
140 r = drm_mm_insert_node_in_range(&mgr->mm, &node->node, mem->num_pages,
141 mem->page_alignment, 0, fpfn, lpfn,
142 mode);
143 spin_unlock(&mgr->lock);
144
145 if (!r)
146 mem->start = node->node.start;
147
148 return r;
149}
150
151/**
152 * amdgpu_gtt_mgr_new - allocate a new node
153 *
154 * @man: TTM memory type manager
155 * @tbo: TTM BO we need this range for
156 * @place: placement flags and restrictions
157 * @mem: the resulting mem object
158 *
159 * Dummy, allocate the node but no space for it yet.
160 */
161static int amdgpu_gtt_mgr_new(struct ttm_mem_type_manager *man,
162 struct ttm_buffer_object *tbo,
163 const struct ttm_place *place,
164 struct ttm_mem_reg *mem)
165{
166 struct amdgpu_gtt_mgr *mgr = man->priv;
167 struct amdgpu_gtt_node *node;
168 int r;
169
170 spin_lock(&mgr->lock);
171 if ((&tbo->mem == mem || tbo->mem.mem_type != TTM_PL_TT) &&
172 atomic64_read(&mgr->available) < mem->num_pages) {
173 spin_unlock(&mgr->lock);
174 return 0;
175 }
176 atomic64_sub(mem->num_pages, &mgr->available);
177 spin_unlock(&mgr->lock);
178
179 node = kzalloc(sizeof(*node), GFP_KERNEL);
180 if (!node) {
181 r = -ENOMEM;
182 goto err_out;
183 }
184
185 node->node.start = AMDGPU_BO_INVALID_OFFSET;
186 node->node.size = mem->num_pages;
187 node->tbo = tbo;
188 mem->mm_node = node;
189
190 if (place->fpfn || place->lpfn || place->flags & TTM_PL_FLAG_TOPDOWN) {
191 r = amdgpu_gtt_mgr_alloc(man, tbo, place, mem);
192 if (unlikely(r)) {
193 kfree(node);
194 mem->mm_node = NULL;
195 r = 0;
196 goto err_out;
197 }
198 } else {
199 mem->start = node->node.start;
200 }
201
202 return 0;
203err_out:
204 atomic64_add(mem->num_pages, &mgr->available);
205
206 return r;
207}
208
209/**
210 * amdgpu_gtt_mgr_del - free ranges
211 *
212 * @man: TTM memory type manager
213 * @tbo: TTM BO we need this range for
214 * @place: placement flags and restrictions
215 * @mem: TTM memory object
216 *
217 * Free the allocated GTT again.
218 */
219static void amdgpu_gtt_mgr_del(struct ttm_mem_type_manager *man,
220 struct ttm_mem_reg *mem)
221{
222 struct amdgpu_gtt_mgr *mgr = man->priv;
223 struct amdgpu_gtt_node *node = mem->mm_node;
224
225 if (!node)
226 return;
227
228 spin_lock(&mgr->lock);
229 if (node->node.start != AMDGPU_BO_INVALID_OFFSET)
230 drm_mm_remove_node(&node->node);
231 spin_unlock(&mgr->lock);
232 atomic64_add(mem->num_pages, &mgr->available);
233
234 kfree(node);
235 mem->mm_node = NULL;
236}
237
238/**
239 * amdgpu_gtt_mgr_usage - return usage of GTT domain
240 *
241 * @man: TTM memory type manager
242 *
243 * Return how many bytes are used in the GTT domain
244 */
245uint64_t amdgpu_gtt_mgr_usage(struct ttm_mem_type_manager *man)
246{
247 struct amdgpu_gtt_mgr *mgr = man->priv;
248 s64 result = man->size - atomic64_read(&mgr->available);
249
250 return (result > 0 ? result : 0) * PAGE_SIZE;
251}
252
253int amdgpu_gtt_mgr_recover(struct ttm_mem_type_manager *man)
254{
255 struct amdgpu_gtt_mgr *mgr = man->priv;
256 struct amdgpu_gtt_node *node;
257 struct drm_mm_node *mm_node;
258 int r = 0;
259
260 spin_lock(&mgr->lock);
261 drm_mm_for_each_node(mm_node, &mgr->mm) {
262 node = container_of(mm_node, struct amdgpu_gtt_node, node);
263 r = amdgpu_ttm_recover_gart(node->tbo);
264 if (r)
265 break;
266 }
267 spin_unlock(&mgr->lock);
268
269 return r;
270}
271
272/**
273 * amdgpu_gtt_mgr_debug - dump VRAM table
274 *
275 * @man: TTM memory type manager
276 * @printer: DRM printer to use
277 *
278 * Dump the table content using printk.
279 */
280static void amdgpu_gtt_mgr_debug(struct ttm_mem_type_manager *man,
281 struct drm_printer *printer)
282{
283 struct amdgpu_gtt_mgr *mgr = man->priv;
284
285 spin_lock(&mgr->lock);
286 drm_mm_print(&mgr->mm, printer);
287 spin_unlock(&mgr->lock);
288
289 drm_printf(printer, "man size:%llu pages, gtt available:%lld pages, usage:%lluMB\n",
290 man->size, (u64)atomic64_read(&mgr->available),
291 amdgpu_gtt_mgr_usage(man) >> 20);
292}
293
294const struct ttm_mem_type_manager_func amdgpu_gtt_mgr_func = {
295 .init = amdgpu_gtt_mgr_init,
296 .takedown = amdgpu_gtt_mgr_fini,
297 .get_node = amdgpu_gtt_mgr_new,
298 .put_node = amdgpu_gtt_mgr_del,
299 .debug = amdgpu_gtt_mgr_debug
300};
301