]>
Commit | Line | Data |
---|---|---|
30a5b536 DZ |
1 | /* |
2 | * mm/percpu-debug.c | |
3 | * | |
4 | * Copyright (C) 2017 Facebook Inc. | |
5 | * Copyright (C) 2017 Dennis Zhou <dennisz@fb.com> | |
6 | * | |
7 | * This file is released under the GPLv2. | |
8 | * | |
9 | * Prints statistics about the percpu allocator and backing chunks. | |
10 | */ | |
11 | #include <linux/debugfs.h> | |
12 | #include <linux/list.h> | |
13 | #include <linux/percpu.h> | |
14 | #include <linux/seq_file.h> | |
15 | #include <linux/sort.h> | |
16 | #include <linux/vmalloc.h> | |
17 | ||
18 | #include "percpu-internal.h" | |
19 | ||
20 | #define P(X, Y) \ | |
02459164 | 21 | seq_printf(m, " %-20s: %12lld\n", X, (long long int)Y) |
30a5b536 DZ |
22 | |
23 | struct percpu_stats pcpu_stats; | |
24 | struct pcpu_alloc_info pcpu_stats_ai; | |
25 | ||
26 | static int cmpint(const void *a, const void *b) | |
27 | { | |
28 | return *(int *)a - *(int *)b; | |
29 | } | |
30 | ||
31 | /* | |
40064aec | 32 | * Iterates over all chunks to find the max nr_alloc entries. |
30a5b536 | 33 | */ |
40064aec | 34 | static int find_max_nr_alloc(void) |
30a5b536 DZ |
35 | { |
36 | struct pcpu_chunk *chunk; | |
40064aec | 37 | int slot, max_nr_alloc; |
30a5b536 | 38 | |
40064aec | 39 | max_nr_alloc = 0; |
30a5b536 DZ |
40 | for (slot = 0; slot < pcpu_nr_slots; slot++) |
41 | list_for_each_entry(chunk, &pcpu_slot[slot], list) | |
40064aec | 42 | max_nr_alloc = max(max_nr_alloc, chunk->nr_alloc); |
30a5b536 | 43 | |
40064aec | 44 | return max_nr_alloc; |
30a5b536 DZ |
45 | } |
46 | ||
47 | /* | |
48 | * Prints out chunk state. Fragmentation is considered between | |
49 | * the beginning of the chunk to the last allocation. | |
40064aec DZF |
50 | * |
51 | * All statistics are in bytes unless stated otherwise. | |
30a5b536 DZ |
52 | */ |
53 | static void chunk_map_stats(struct seq_file *m, struct pcpu_chunk *chunk, | |
cd6a884d | 54 | int *buffer) |
30a5b536 | 55 | { |
40064aec | 56 | int i, last_alloc, as_len, start, end; |
30a5b536 DZ |
57 | int *alloc_sizes, *p; |
58 | /* statistics */ | |
59 | int sum_frag = 0, max_frag = 0; | |
60 | int cur_min_alloc = 0, cur_med_alloc = 0, cur_max_alloc = 0; | |
61 | ||
62 | alloc_sizes = buffer; | |
30a5b536 | 63 | |
40064aec DZF |
64 | /* |
65 | * find_last_bit returns the start value if nothing found. | |
66 | * Therefore, we must determine if it is a failure of find_last_bit | |
67 | * and set the appropriate value. | |
68 | */ | |
69 | last_alloc = find_last_bit(chunk->alloc_map, | |
70 | pcpu_chunk_map_bits(chunk) - | |
71 | chunk->end_offset / PCPU_MIN_ALLOC_SIZE - 1); | |
72 | last_alloc = test_bit(last_alloc, chunk->alloc_map) ? | |
73 | last_alloc + 1 : 0; | |
74 | ||
75 | as_len = 0; | |
76 | start = chunk->start_offset; | |
77 | ||
78 | /* | |
79 | * If a bit is set in the allocation map, the bound_map identifies | |
80 | * where the allocation ends. If the allocation is not set, the | |
81 | * bound_map does not identify free areas as it is only kept accurate | |
82 | * on allocation, not free. | |
83 | * | |
84 | * Positive values are allocations and negative values are free | |
85 | * fragments. | |
86 | */ | |
87 | while (start < last_alloc) { | |
88 | if (test_bit(start, chunk->alloc_map)) { | |
89 | end = find_next_bit(chunk->bound_map, last_alloc, | |
90 | start + 1); | |
91 | alloc_sizes[as_len] = 1; | |
92 | } else { | |
93 | end = find_next_bit(chunk->alloc_map, last_alloc, | |
94 | start + 1); | |
95 | alloc_sizes[as_len] = -1; | |
30a5b536 DZ |
96 | } |
97 | ||
40064aec DZF |
98 | alloc_sizes[as_len++] *= (end - start) * PCPU_MIN_ALLOC_SIZE; |
99 | ||
100 | start = end; | |
101 | } | |
102 | ||
103 | /* | |
104 | * The negative values are free fragments and thus sorting gives the | |
105 | * free fragments at the beginning in largest first order. | |
106 | */ | |
107 | if (as_len > 0) { | |
108 | sort(alloc_sizes, as_len, sizeof(int), cmpint, NULL); | |
30a5b536 | 109 | |
40064aec | 110 | /* iterate through the unallocated fragments */ |
30a5b536 DZ |
111 | for (i = 0, p = alloc_sizes; *p < 0 && i < as_len; i++, p++) { |
112 | sum_frag -= *p; | |
113 | max_frag = max(max_frag, -1 * (*p)); | |
114 | } | |
115 | ||
116 | cur_min_alloc = alloc_sizes[i]; | |
117 | cur_med_alloc = alloc_sizes[(i + as_len - 1) / 2]; | |
118 | cur_max_alloc = alloc_sizes[as_len - 1]; | |
119 | } | |
120 | ||
121 | P("nr_alloc", chunk->nr_alloc); | |
122 | P("max_alloc_size", chunk->max_alloc_size); | |
0cecf50c | 123 | P("empty_pop_pages", chunk->nr_empty_pop_pages); |
40064aec DZF |
124 | P("free_bytes", chunk->free_bytes); |
125 | P("contig_bytes", chunk->contig_bits * PCPU_MIN_ALLOC_SIZE); | |
30a5b536 DZ |
126 | P("sum_frag", sum_frag); |
127 | P("max_frag", max_frag); | |
128 | P("cur_min_alloc", cur_min_alloc); | |
129 | P("cur_med_alloc", cur_med_alloc); | |
130 | P("cur_max_alloc", cur_max_alloc); | |
131 | seq_putc(m, '\n'); | |
132 | } | |
133 | ||
134 | static int percpu_stats_show(struct seq_file *m, void *v) | |
135 | { | |
136 | struct pcpu_chunk *chunk; | |
40064aec | 137 | int slot, max_nr_alloc; |
cd6a884d | 138 | int *buffer; |
30a5b536 DZ |
139 | |
140 | alloc_buffer: | |
141 | spin_lock_irq(&pcpu_lock); | |
40064aec | 142 | max_nr_alloc = find_max_nr_alloc(); |
30a5b536 DZ |
143 | spin_unlock_irq(&pcpu_lock); |
144 | ||
40064aec DZF |
145 | /* there can be at most this many free and allocated fragments */ |
146 | buffer = vmalloc((2 * max_nr_alloc + 1) * sizeof(int)); | |
30a5b536 DZ |
147 | if (!buffer) |
148 | return -ENOMEM; | |
149 | ||
150 | spin_lock_irq(&pcpu_lock); | |
151 | ||
152 | /* if the buffer allocated earlier is too small */ | |
40064aec | 153 | if (max_nr_alloc < find_max_nr_alloc()) { |
30a5b536 DZ |
154 | spin_unlock_irq(&pcpu_lock); |
155 | vfree(buffer); | |
156 | goto alloc_buffer; | |
157 | } | |
158 | ||
159 | #define PL(X) \ | |
02459164 | 160 | seq_printf(m, " %-20s: %12lld\n", #X, (long long int)pcpu_stats_ai.X) |
30a5b536 DZ |
161 | |
162 | seq_printf(m, | |
163 | "Percpu Memory Statistics\n" | |
164 | "Allocation Info:\n" | |
165 | "----------------------------------------\n"); | |
166 | PL(unit_size); | |
167 | PL(static_size); | |
168 | PL(reserved_size); | |
169 | PL(dyn_size); | |
170 | PL(atom_size); | |
171 | PL(alloc_size); | |
172 | seq_putc(m, '\n'); | |
173 | ||
174 | #undef PL | |
175 | ||
176 | #define PU(X) \ | |
02459164 | 177 | seq_printf(m, " %-20s: %12llu\n", #X, (unsigned long long)pcpu_stats.X) |
30a5b536 DZ |
178 | |
179 | seq_printf(m, | |
180 | "Global Stats:\n" | |
181 | "----------------------------------------\n"); | |
182 | PU(nr_alloc); | |
183 | PU(nr_dealloc); | |
184 | PU(nr_cur_alloc); | |
185 | PU(nr_max_alloc); | |
186 | PU(nr_chunks); | |
187 | PU(nr_max_chunks); | |
188 | PU(min_alloc_size); | |
189 | PU(max_alloc_size); | |
6b9b6f39 | 190 | P("empty_pop_pages", pcpu_nr_empty_pop_pages); |
30a5b536 DZ |
191 | seq_putc(m, '\n'); |
192 | ||
193 | #undef PU | |
194 | ||
195 | seq_printf(m, | |
196 | "Per Chunk Stats:\n" | |
197 | "----------------------------------------\n"); | |
198 | ||
199 | if (pcpu_reserved_chunk) { | |
200 | seq_puts(m, "Chunk: <- Reserved Chunk\n"); | |
201 | chunk_map_stats(m, pcpu_reserved_chunk, buffer); | |
202 | } | |
203 | ||
204 | for (slot = 0; slot < pcpu_nr_slots; slot++) { | |
205 | list_for_each_entry(chunk, &pcpu_slot[slot], list) { | |
206 | if (chunk == pcpu_first_chunk) { | |
207 | seq_puts(m, "Chunk: <- First Chunk\n"); | |
208 | chunk_map_stats(m, chunk, buffer); | |
209 | ||
210 | ||
211 | } else { | |
212 | seq_puts(m, "Chunk:\n"); | |
213 | chunk_map_stats(m, chunk, buffer); | |
214 | } | |
215 | ||
216 | } | |
217 | } | |
218 | ||
219 | spin_unlock_irq(&pcpu_lock); | |
220 | ||
221 | vfree(buffer); | |
222 | ||
223 | return 0; | |
224 | } | |
225 | ||
226 | static int percpu_stats_open(struct inode *inode, struct file *filp) | |
227 | { | |
228 | return single_open(filp, percpu_stats_show, NULL); | |
229 | } | |
230 | ||
231 | static const struct file_operations percpu_stats_fops = { | |
232 | .open = percpu_stats_open, | |
233 | .read = seq_read, | |
234 | .llseek = seq_lseek, | |
235 | .release = single_release, | |
236 | }; | |
237 | ||
238 | static int __init init_percpu_stats_debugfs(void) | |
239 | { | |
240 | debugfs_create_file("percpu_stats", 0444, NULL, NULL, | |
241 | &percpu_stats_fops); | |
242 | ||
243 | return 0; | |
244 | } | |
245 | ||
246 | late_initcall(init_percpu_stats_debugfs); |