diff options
Diffstat (limited to 'drivers/gpu/nvgpu/gk20a/gk20a_allocator.h')
-rw-r--r-- | drivers/gpu/nvgpu/gk20a/gk20a_allocator.h | 200 |
1 files changed, 30 insertions, 170 deletions
diff --git a/drivers/gpu/nvgpu/gk20a/gk20a_allocator.h b/drivers/gpu/nvgpu/gk20a/gk20a_allocator.h index 74e23e6c..06819700 100644 --- a/drivers/gpu/nvgpu/gk20a/gk20a_allocator.h +++ b/drivers/gpu/nvgpu/gk20a/gk20a_allocator.h | |||
@@ -17,8 +17,6 @@ | |||
17 | #ifndef GK20A_ALLOCATOR_H | 17 | #ifndef GK20A_ALLOCATOR_H |
18 | #define GK20A_ALLOCATOR_H | 18 | #define GK20A_ALLOCATOR_H |
19 | 19 | ||
20 | #include <linux/list.h> | ||
21 | #include <linux/rbtree.h> | ||
22 | #include <linux/debugfs.h> | 20 | #include <linux/debugfs.h> |
23 | #include <linux/seq_file.h> | 21 | #include <linux/seq_file.h> |
24 | #include <linux/platform_device.h> | 22 | #include <linux/platform_device.h> |
@@ -63,138 +61,6 @@ struct gk20a_allocator_ops { | |||
63 | struct seq_file *s, int lock); | 61 | struct seq_file *s, int lock); |
64 | }; | 62 | }; |
65 | 63 | ||
66 | /* | ||
67 | * Each buddy is an element in a binary tree. | ||
68 | */ | ||
69 | struct gk20a_buddy { | ||
70 | struct gk20a_buddy *parent; /* Parent node. */ | ||
71 | struct gk20a_buddy *buddy; /* This node's buddy. */ | ||
72 | struct gk20a_buddy *left; /* Lower address sub-node. */ | ||
73 | struct gk20a_buddy *right; /* Higher address sub-node. */ | ||
74 | |||
75 | struct list_head buddy_entry; /* List entry for various lists. */ | ||
76 | struct rb_node alloced_entry; /* RB tree of allocations. */ | ||
77 | |||
78 | u64 start; /* Start address of this buddy. */ | ||
79 | u64 end; /* End address of this buddy. */ | ||
80 | u64 order; /* Buddy order. */ | ||
81 | |||
82 | #define BALLOC_BUDDY_ALLOCED 0x1 | ||
83 | #define BALLOC_BUDDY_SPLIT 0x2 | ||
84 | #define BALLOC_BUDDY_IN_LIST 0x4 | ||
85 | int flags; /* List of associated flags. */ | ||
86 | |||
87 | /* | ||
88 | * Size of the PDE this buddy is using. This allows for grouping like | ||
89 | * sized allocations into the same PDE. | ||
90 | */ | ||
91 | #define BALLOC_PTE_SIZE_ANY 0x0 | ||
92 | #define BALLOC_PTE_SIZE_SMALL 0x1 | ||
93 | #define BALLOC_PTE_SIZE_BIG 0x2 | ||
94 | int pte_size; | ||
95 | }; | ||
96 | |||
97 | #define __buddy_flag_ops(flag, flag_up) \ | ||
98 | static inline int buddy_is_ ## flag(struct gk20a_buddy *b) \ | ||
99 | { \ | ||
100 | return b->flags & BALLOC_BUDDY_ ## flag_up; \ | ||
101 | } \ | ||
102 | static inline void buddy_set_ ## flag(struct gk20a_buddy *b) \ | ||
103 | { \ | ||
104 | b->flags |= BALLOC_BUDDY_ ## flag_up; \ | ||
105 | } \ | ||
106 | static inline void buddy_clr_ ## flag(struct gk20a_buddy *b) \ | ||
107 | { \ | ||
108 | b->flags &= ~BALLOC_BUDDY_ ## flag_up; \ | ||
109 | } | ||
110 | |||
111 | /* | ||
112 | * int buddy_is_alloced(struct gk20a_buddy *b); | ||
113 | * void buddy_set_alloced(struct gk20a_buddy *b); | ||
114 | * void buddy_clr_alloced(struct gk20a_buddy *b); | ||
115 | * | ||
116 | * int buddy_is_split(struct gk20a_buddy *b); | ||
117 | * void buddy_set_split(struct gk20a_buddy *b); | ||
118 | * void buddy_clr_split(struct gk20a_buddy *b); | ||
119 | * | ||
120 | * int buddy_is_in_list(struct gk20a_buddy *b); | ||
121 | * void buddy_set_in_list(struct gk20a_buddy *b); | ||
122 | * void buddy_clr_in_list(struct gk20a_buddy *b); | ||
123 | */ | ||
124 | __buddy_flag_ops(alloced, ALLOCED); | ||
125 | __buddy_flag_ops(split, SPLIT); | ||
126 | __buddy_flag_ops(in_list, IN_LIST); | ||
127 | |||
128 | /* | ||
129 | * Keeps info for a fixed allocation. | ||
130 | */ | ||
131 | struct gk20a_fixed_alloc { | ||
132 | struct list_head buddies; /* List of buddies. */ | ||
133 | struct rb_node alloced_entry; /* RB tree of fixed allocations. */ | ||
134 | |||
135 | u64 start; /* Start of fixed block. */ | ||
136 | u64 end; /* End address. */ | ||
137 | }; | ||
138 | |||
139 | /* | ||
140 | * GPU buddy allocator for the various GPU address spaces. Each addressable unit | ||
141 | * doesn't have to correspond to a byte. In some cases each unit is a more | ||
142 | * complex object such as a comp_tag line or the like. | ||
143 | * | ||
144 | * The max order is computed based on the size of the minimum order and the size | ||
145 | * of the address space. | ||
146 | * | ||
147 | * order_size is the size of an order 0 buddy. | ||
148 | */ | ||
149 | struct gk20a_buddy_allocator { | ||
150 | struct gk20a_allocator *owner; /* Owner of this buddy allocator. */ | ||
151 | struct vm_gk20a *vm; /* Parent VM - can be NULL. */ | ||
152 | |||
153 | u64 base; /* Base address of the space. */ | ||
154 | u64 length; /* Length of the space. */ | ||
155 | u64 blk_size; /* Size of order 0 allocation. */ | ||
156 | u64 blk_shift; /* Shift to divide by blk_size. */ | ||
157 | |||
158 | int init; /* Non-zero if initialized. */ | ||
159 | |||
160 | /* Internal stuff. */ | ||
161 | u64 start; /* Real start (aligned to blk_size). */ | ||
162 | u64 end; /* Real end, trimmed if needed. */ | ||
163 | u64 count; /* Count of objects in space. */ | ||
164 | u64 blks; /* Count of blks in the space. */ | ||
165 | u64 max_order; /* Specific maximum order. */ | ||
166 | |||
167 | struct rb_root alloced_buddies; /* Outstanding allocations. */ | ||
168 | struct rb_root fixed_allocs; /* Outstanding fixed allocations. */ | ||
169 | |||
170 | /* | ||
171 | * Impose an upper bound on the maximum order. | ||
172 | */ | ||
173 | #define GPU_BALLOC_MAX_ORDER 31 | ||
174 | #define GPU_BALLOC_ORDER_LIST_LEN (GPU_BALLOC_MAX_ORDER + 1) | ||
175 | |||
176 | struct list_head buddy_list[GPU_BALLOC_ORDER_LIST_LEN]; | ||
177 | u64 buddy_list_len[GPU_BALLOC_ORDER_LIST_LEN]; | ||
178 | u64 buddy_list_split[GPU_BALLOC_ORDER_LIST_LEN]; | ||
179 | u64 buddy_list_alloced[GPU_BALLOC_ORDER_LIST_LEN]; | ||
180 | |||
181 | /* | ||
182 | * This is for when the allocator is managing a GVA space (the | ||
183 | * GPU_BALLOC_GVA_SPACE bit is set in @flags). This requires | ||
184 | * that we group like sized allocations into PDE blocks. | ||
185 | */ | ||
186 | u64 pte_blk_order; | ||
187 | |||
188 | int inited; | ||
189 | |||
190 | #define GPU_BALLOC_GVA_SPACE 0x1 | ||
191 | u64 flags; | ||
192 | |||
193 | u64 bytes_alloced; | ||
194 | u64 bytes_alloced_real; | ||
195 | u64 bytes_freed; | ||
196 | }; | ||
197 | |||
198 | struct gk20a_allocator { | 64 | struct gk20a_allocator { |
199 | char name[32]; | 65 | char name[32]; |
200 | struct mutex lock; | 66 | struct mutex lock; |
@@ -205,6 +71,11 @@ struct gk20a_allocator { | |||
205 | struct dentry *debugfs_entry; | 71 | struct dentry *debugfs_entry; |
206 | }; | 72 | }; |
207 | 73 | ||
74 | /* | ||
75 | * Allocator flags. | ||
76 | */ | ||
77 | #define GPU_BALLOC_GVA_SPACE 0x1 | ||
78 | |||
208 | static inline void alloc_lock(struct gk20a_allocator *a) | 79 | static inline void alloc_lock(struct gk20a_allocator *a) |
209 | { | 80 | { |
210 | mutex_lock(&a->lock); | 81 | mutex_lock(&a->lock); |
@@ -215,42 +86,6 @@ static inline void alloc_unlock(struct gk20a_allocator *a) | |||
215 | mutex_unlock(&a->lock); | 86 | mutex_unlock(&a->lock); |
216 | } | 87 | } |
217 | 88 | ||
218 | static inline struct gk20a_buddy_allocator *buddy_allocator( | ||
219 | struct gk20a_allocator *a) | ||
220 | { | ||
221 | return (struct gk20a_buddy_allocator *)a->priv; | ||
222 | } | ||
223 | |||
224 | static inline struct list_head *balloc_get_order_list( | ||
225 | struct gk20a_buddy_allocator *a, int order) | ||
226 | { | ||
227 | return &a->buddy_list[order]; | ||
228 | } | ||
229 | |||
230 | static inline u64 balloc_order_to_len(struct gk20a_buddy_allocator *a, | ||
231 | int order) | ||
232 | { | ||
233 | return (1 << order) * a->blk_size; | ||
234 | } | ||
235 | |||
236 | static inline u64 balloc_base_shift(struct gk20a_buddy_allocator *a, | ||
237 | u64 base) | ||
238 | { | ||
239 | return base - a->start; | ||
240 | } | ||
241 | |||
242 | static inline u64 balloc_base_unshift(struct gk20a_buddy_allocator *a, | ||
243 | u64 base) | ||
244 | { | ||
245 | return base + a->start; | ||
246 | } | ||
247 | |||
248 | static inline struct gk20a_allocator *balloc_owner( | ||
249 | struct gk20a_buddy_allocator *a) | ||
250 | { | ||
251 | return a->owner; | ||
252 | } | ||
253 | |||
254 | /* | 89 | /* |
255 | * Buddy allocator specific initializers. | 90 | * Buddy allocator specific initializers. |
256 | */ | 91 | */ |
@@ -262,6 +97,8 @@ int gk20a_buddy_allocator_init(struct gk20a_allocator *allocator, | |||
262 | const char *name, u64 base, u64 size, | 97 | const char *name, u64 base, u64 size, |
263 | u64 blk_size, u64 flags); | 98 | u64 blk_size, u64 flags); |
264 | 99 | ||
100 | #define GPU_BALLOC_MAX_ORDER 31 | ||
101 | |||
265 | /* | 102 | /* |
266 | * Allocator APIs. | 103 | * Allocator APIs. |
267 | */ | 104 | */ |
@@ -282,10 +119,33 @@ void gk20a_alloc_print_stats(struct gk20a_allocator *a, | |||
282 | struct seq_file *s, int lock); | 119 | struct seq_file *s, int lock); |
283 | 120 | ||
284 | /* | 121 | /* |
122 | * Common functionality for the internals of the allocators. | ||
123 | */ | ||
124 | void gk20a_init_alloc_debug(struct gk20a_allocator *a); | ||
125 | void gk20a_fini_alloc_debug(struct gk20a_allocator *a); | ||
126 | int __gk20a_alloc_common_init(struct gk20a_allocator *a, | ||
127 | const char *name, void *priv, | ||
128 | const struct gk20a_allocator_ops *ops); | ||
129 | |||
130 | /* | ||
285 | * Debug stuff. | 131 | * Debug stuff. |
286 | */ | 132 | */ |
133 | extern u32 gk20a_alloc_tracing_on; | ||
134 | |||
287 | void gk20a_alloc_debugfs_init(struct platform_device *pdev); | 135 | void gk20a_alloc_debugfs_init(struct platform_device *pdev); |
288 | 136 | ||
137 | #define gk20a_alloc_trace_func() \ | ||
138 | do { \ | ||
139 | if (gk20a_alloc_tracing_on) \ | ||
140 | trace_printk("%s\n", __func__); \ | ||
141 | } while (0) | ||
142 | |||
143 | #define gk20a_alloc_trace_func_done() \ | ||
144 | do { \ | ||
145 | if (gk20a_alloc_tracing_on) \ | ||
146 | trace_printk("%s_done\n", __func__); \ | ||
147 | } while (0) | ||
148 | |||
289 | #define __alloc_pstat(seq, allocator, fmt, arg...) \ | 149 | #define __alloc_pstat(seq, allocator, fmt, arg...) \ |
290 | do { \ | 150 | do { \ |
291 | if (s) \ | 151 | if (s) \ |