Commit | Line | Data |
---|---|---|
55716d26 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
30a5b536 DZ |
2 | /* |
3 | * mm/percpu-debug.c | |
4 | * | |
5 | * Copyright (C) 2017 Facebook Inc. | |
bfacd38f | 6 | * Copyright (C) 2017 Dennis Zhou <dennis@kernel.org> |
30a5b536 | 7 | * |
30a5b536 DZ |
8 | * Prints statistics about the percpu allocator and backing chunks. |
9 | */ | |
10 | #include <linux/debugfs.h> | |
11 | #include <linux/list.h> | |
12 | #include <linux/percpu.h> | |
13 | #include <linux/seq_file.h> | |
14 | #include <linux/sort.h> | |
15 | #include <linux/vmalloc.h> | |
16 | ||
17 | #include "percpu-internal.h" | |
18 | ||
19 | #define P(X, Y) \ | |
02459164 | 20 | seq_printf(m, " %-20s: %12lld\n", X, (long long int)Y) |
30a5b536 DZ |
21 | |
22 | struct percpu_stats pcpu_stats; | |
23 | struct pcpu_alloc_info pcpu_stats_ai; | |
24 | ||
25 | static int cmpint(const void *a, const void *b) | |
26 | { | |
27 | return *(int *)a - *(int *)b; | |
28 | } | |
29 | ||
30 | /* | |
40064aec | 31 | * Iterates over all chunks to find the max nr_alloc entries. |
30a5b536 | 32 | */ |
40064aec | 33 | static int find_max_nr_alloc(void) |
30a5b536 DZ |
34 | { |
35 | struct pcpu_chunk *chunk; | |
40064aec | 36 | int slot, max_nr_alloc; |
3c7be18a | 37 | enum pcpu_chunk_type type; |
30a5b536 | 38 | |
40064aec | 39 | max_nr_alloc = 0; |
3c7be18a RG |
40 | for (type = 0; type < PCPU_NR_CHUNK_TYPES; type++) |
41 | for (slot = 0; slot < pcpu_nr_slots; slot++) | |
42 | list_for_each_entry(chunk, &pcpu_chunk_list(type)[slot], | |
43 | list) | |
44 | max_nr_alloc = max(max_nr_alloc, | |
45 | chunk->nr_alloc); | |
30a5b536 | 46 | |
40064aec | 47 | return max_nr_alloc; |
30a5b536 DZ |
48 | } |
49 | ||
50 | /* | |
51 | * Prints out chunk state. Fragmentation is considered between | |
52 | * the beginning of the chunk to the last allocation. | |
40064aec DZF |
53 | * |
54 | * All statistics are in bytes unless stated otherwise. | |
30a5b536 DZ |
55 | */ |
56 | static void chunk_map_stats(struct seq_file *m, struct pcpu_chunk *chunk, | |
cd6a884d | 57 | int *buffer) |
30a5b536 | 58 | { |
92c14cab | 59 | struct pcpu_block_md *chunk_md = &chunk->chunk_md; |
40064aec | 60 | int i, last_alloc, as_len, start, end; |
30a5b536 DZ |
61 | int *alloc_sizes, *p; |
62 | /* statistics */ | |
63 | int sum_frag = 0, max_frag = 0; | |
64 | int cur_min_alloc = 0, cur_med_alloc = 0, cur_max_alloc = 0; | |
65 | ||
66 | alloc_sizes = buffer; | |
30a5b536 | 67 | |
40064aec DZF |
68 | /* |
69 | * find_last_bit returns the start value if nothing found. | |
70 | * Therefore, we must determine if it is a failure of find_last_bit | |
71 | * and set the appropriate value. | |
72 | */ | |
73 | last_alloc = find_last_bit(chunk->alloc_map, | |
74 | pcpu_chunk_map_bits(chunk) - | |
75 | chunk->end_offset / PCPU_MIN_ALLOC_SIZE - 1); | |
76 | last_alloc = test_bit(last_alloc, chunk->alloc_map) ? | |
77 | last_alloc + 1 : 0; | |
78 | ||
79 | as_len = 0; | |
2e08d20d | 80 | start = chunk->start_offset / PCPU_MIN_ALLOC_SIZE; |
40064aec DZF |
81 | |
82 | /* | |
83 | * If a bit is set in the allocation map, the bound_map identifies | |
84 | * where the allocation ends. If the allocation is not set, the | |
85 | * bound_map does not identify free areas as it is only kept accurate | |
86 | * on allocation, not free. | |
87 | * | |
88 | * Positive values are allocations and negative values are free | |
89 | * fragments. | |
90 | */ | |
91 | while (start < last_alloc) { | |
92 | if (test_bit(start, chunk->alloc_map)) { | |
93 | end = find_next_bit(chunk->bound_map, last_alloc, | |
94 | start + 1); | |
95 | alloc_sizes[as_len] = 1; | |
96 | } else { | |
97 | end = find_next_bit(chunk->alloc_map, last_alloc, | |
98 | start + 1); | |
99 | alloc_sizes[as_len] = -1; | |
30a5b536 DZ |
100 | } |
101 | ||
40064aec DZF |
102 | alloc_sizes[as_len++] *= (end - start) * PCPU_MIN_ALLOC_SIZE; |
103 | ||
104 | start = end; | |
105 | } | |
106 | ||
107 | /* | |
108 | * The negative values are free fragments and thus sorting gives the | |
109 | * free fragments at the beginning in largest first order. | |
110 | */ | |
111 | if (as_len > 0) { | |
112 | sort(alloc_sizes, as_len, sizeof(int), cmpint, NULL); | |
30a5b536 | 113 | |
40064aec | 114 | /* iterate through the unallocated fragments */ |
30a5b536 DZ |
115 | for (i = 0, p = alloc_sizes; *p < 0 && i < as_len; i++, p++) { |
116 | sum_frag -= *p; | |
117 | max_frag = max(max_frag, -1 * (*p)); | |
118 | } | |
119 | ||
120 | cur_min_alloc = alloc_sizes[i]; | |
121 | cur_med_alloc = alloc_sizes[(i + as_len - 1) / 2]; | |
122 | cur_max_alloc = alloc_sizes[as_len - 1]; | |
123 | } | |
124 | ||
125 | P("nr_alloc", chunk->nr_alloc); | |
126 | P("max_alloc_size", chunk->max_alloc_size); | |
0cecf50c | 127 | P("empty_pop_pages", chunk->nr_empty_pop_pages); |
92c14cab | 128 | P("first_bit", chunk_md->first_free); |
40064aec | 129 | P("free_bytes", chunk->free_bytes); |
92c14cab | 130 | P("contig_bytes", chunk_md->contig_hint * PCPU_MIN_ALLOC_SIZE); |
30a5b536 DZ |
131 | P("sum_frag", sum_frag); |
132 | P("max_frag", max_frag); | |
133 | P("cur_min_alloc", cur_min_alloc); | |
134 | P("cur_med_alloc", cur_med_alloc); | |
135 | P("cur_max_alloc", cur_max_alloc); | |
3c7be18a RG |
136 | #ifdef CONFIG_MEMCG_KMEM |
137 | P("memcg_aware", pcpu_is_memcg_chunk(pcpu_chunk_type(chunk))); | |
138 | #endif | |
30a5b536 DZ |
139 | seq_putc(m, '\n'); |
140 | } | |
141 | ||
142 | static int percpu_stats_show(struct seq_file *m, void *v) | |
143 | { | |
144 | struct pcpu_chunk *chunk; | |
40064aec | 145 | int slot, max_nr_alloc; |
cd6a884d | 146 | int *buffer; |
3c7be18a | 147 | enum pcpu_chunk_type type; |
0760fa3d | 148 | int nr_empty_pop_pages; |
30a5b536 DZ |
149 | |
150 | alloc_buffer: | |
151 | spin_lock_irq(&pcpu_lock); | |
40064aec | 152 | max_nr_alloc = find_max_nr_alloc(); |
30a5b536 DZ |
153 | spin_unlock_irq(&pcpu_lock); |
154 | ||
40064aec | 155 | /* there can be at most this many free and allocated fragments */ |
42bc47b3 | 156 | buffer = vmalloc(array_size(sizeof(int), (2 * max_nr_alloc + 1))); |
30a5b536 DZ |
157 | if (!buffer) |
158 | return -ENOMEM; | |
159 | ||
160 | spin_lock_irq(&pcpu_lock); | |
161 | ||
162 | /* if the buffer allocated earlier is too small */ | |
40064aec | 163 | if (max_nr_alloc < find_max_nr_alloc()) { |
30a5b536 DZ |
164 | spin_unlock_irq(&pcpu_lock); |
165 | vfree(buffer); | |
166 | goto alloc_buffer; | |
167 | } | |
168 | ||
0760fa3d RG |
169 | nr_empty_pop_pages = 0; |
170 | for (type = 0; type < PCPU_NR_CHUNK_TYPES; type++) | |
171 | nr_empty_pop_pages += pcpu_nr_empty_pop_pages[type]; | |
172 | ||
173 | #define PL(X) \ | |
02459164 | 174 | seq_printf(m, " %-20s: %12lld\n", #X, (long long int)pcpu_stats_ai.X) |
30a5b536 DZ |
175 | |
176 | seq_printf(m, | |
177 | "Percpu Memory Statistics\n" | |
178 | "Allocation Info:\n" | |
179 | "----------------------------------------\n"); | |
180 | PL(unit_size); | |
181 | PL(static_size); | |
182 | PL(reserved_size); | |
183 | PL(dyn_size); | |
184 | PL(atom_size); | |
185 | PL(alloc_size); | |
186 | seq_putc(m, '\n'); | |
187 | ||
188 | #undef PL | |
189 | ||
190 | #define PU(X) \ | |
02459164 | 191 | seq_printf(m, " %-20s: %12llu\n", #X, (unsigned long long)pcpu_stats.X) |
30a5b536 DZ |
192 | |
193 | seq_printf(m, | |
194 | "Global Stats:\n" | |
195 | "----------------------------------------\n"); | |
196 | PU(nr_alloc); | |
197 | PU(nr_dealloc); | |
198 | PU(nr_cur_alloc); | |
199 | PU(nr_max_alloc); | |
200 | PU(nr_chunks); | |
201 | PU(nr_max_chunks); | |
202 | PU(min_alloc_size); | |
203 | PU(max_alloc_size); | |
0760fa3d | 204 | P("empty_pop_pages", nr_empty_pop_pages); |
30a5b536 DZ |
205 | seq_putc(m, '\n'); |
206 | ||
207 | #undef PU | |
208 | ||
209 | seq_printf(m, | |
210 | "Per Chunk Stats:\n" | |
211 | "----------------------------------------\n"); | |
212 | ||
213 | if (pcpu_reserved_chunk) { | |
214 | seq_puts(m, "Chunk: <- Reserved Chunk\n"); | |
215 | chunk_map_stats(m, pcpu_reserved_chunk, buffer); | |
216 | } | |
217 | ||
3c7be18a RG |
218 | for (type = 0; type < PCPU_NR_CHUNK_TYPES; type++) { |
219 | for (slot = 0; slot < pcpu_nr_slots; slot++) { | |
220 | list_for_each_entry(chunk, &pcpu_chunk_list(type)[slot], | |
221 | list) { | |
222 | if (chunk == pcpu_first_chunk) { | |
223 | seq_puts(m, "Chunk: <- First Chunk\n"); | |
224 | chunk_map_stats(m, chunk, buffer); | |
225 | } else { | |
226 | seq_puts(m, "Chunk:\n"); | |
227 | chunk_map_stats(m, chunk, buffer); | |
228 | } | |
30a5b536 | 229 | } |
30a5b536 DZ |
230 | } |
231 | } | |
232 | ||
233 | spin_unlock_irq(&pcpu_lock); | |
234 | ||
235 | vfree(buffer); | |
236 | ||
237 | return 0; | |
238 | } | |
5ad35093 | 239 | DEFINE_SHOW_ATTRIBUTE(percpu_stats); |
30a5b536 DZ |
240 | |
241 | static int __init init_percpu_stats_debugfs(void) | |
242 | { | |
243 | debugfs_create_file("percpu_stats", 0444, NULL, NULL, | |
244 | &percpu_stats_fops); | |
245 | ||
246 | return 0; | |
247 | } | |
248 | ||
249 | late_initcall(init_percpu_stats_debugfs); |