1 | // SPDX-License-Identifier: GPL-2.0-only |
2 | /* |
3 | * mm/percpu-debug.c |
4 | * |
5 | * Copyright (C) 2017 Facebook Inc. |
6 | * Copyright (C) 2017 Dennis Zhou <dennis@kernel.org> |
7 | * |
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) \ |
20 | seq_printf(m, " %-20s: %12lld\n", X, (long long int)Y) |
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 | /* |
31 | * Iterates over all chunks to find the max nr_alloc entries. |
32 | */ |
33 | static int find_max_nr_alloc(void) |
34 | { |
35 | struct pcpu_chunk *chunk; |
36 | int slot, max_nr_alloc; |
37 | |
38 | max_nr_alloc = 0; |
39 | for (slot = 0; slot < pcpu_nr_slots; slot++) |
40 | list_for_each_entry(chunk, &pcpu_chunk_lists[slot], list) |
41 | max_nr_alloc = max(max_nr_alloc, chunk->nr_alloc); |
42 | |
43 | return max_nr_alloc; |
44 | } |
45 | |
46 | /* |
47 | * Prints out chunk state. Fragmentation is considered between |
48 | * the beginning of the chunk to the last allocation. |
49 | * |
50 | * All statistics are in bytes unless stated otherwise. |
51 | */ |
52 | static void chunk_map_stats(struct seq_file *m, struct pcpu_chunk *chunk, |
53 | int *buffer) |
54 | { |
55 | struct pcpu_block_md *chunk_md = &chunk->chunk_md; |
56 | int i, last_alloc, as_len, start, end; |
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; |
63 | |
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(addr: chunk->alloc_map, |
70 | size: 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 / PCPU_MIN_ALLOC_SIZE; |
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(addr: chunk->bound_map, size: last_alloc, |
90 | offset: start + 1); |
91 | alloc_sizes[as_len] = 1; |
92 | } else { |
93 | end = find_next_bit(addr: chunk->alloc_map, size: last_alloc, |
94 | offset: start + 1); |
95 | alloc_sizes[as_len] = -1; |
96 | } |
97 | |
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(base: alloc_sizes, num: as_len, size: sizeof(int), cmp_func: cmpint, NULL); |
109 | |
110 | /* iterate through the unallocated fragments */ |
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); |
123 | P("empty_pop_pages" , chunk->nr_empty_pop_pages); |
124 | P("first_bit" , chunk_md->first_free); |
125 | P("free_bytes" , chunk->free_bytes); |
126 | P("contig_bytes" , chunk_md->contig_hint * PCPU_MIN_ALLOC_SIZE); |
127 | P("sum_frag" , sum_frag); |
128 | P("max_frag" , max_frag); |
129 | P("cur_min_alloc" , cur_min_alloc); |
130 | P("cur_med_alloc" , cur_med_alloc); |
131 | P("cur_max_alloc" , cur_max_alloc); |
132 | seq_putc(m, c: '\n'); |
133 | } |
134 | |
135 | static int percpu_stats_show(struct seq_file *m, void *v) |
136 | { |
137 | struct pcpu_chunk *chunk; |
138 | int slot, max_nr_alloc; |
139 | int *buffer; |
140 | |
141 | alloc_buffer: |
142 | spin_lock_irq(lock: &pcpu_lock); |
143 | max_nr_alloc = find_max_nr_alloc(); |
144 | spin_unlock_irq(lock: &pcpu_lock); |
145 | |
146 | /* there can be at most this many free and allocated fragments */ |
147 | buffer = vmalloc_array(n: 2 * max_nr_alloc + 1, size: sizeof(int)); |
148 | if (!buffer) |
149 | return -ENOMEM; |
150 | |
151 | spin_lock_irq(lock: &pcpu_lock); |
152 | |
153 | /* if the buffer allocated earlier is too small */ |
154 | if (max_nr_alloc < find_max_nr_alloc()) { |
155 | spin_unlock_irq(lock: &pcpu_lock); |
156 | vfree(addr: buffer); |
157 | goto alloc_buffer; |
158 | } |
159 | |
160 | #define PL(X) \ |
161 | seq_printf(m, " %-20s: %12lld\n", #X, (long long int)pcpu_stats_ai.X) |
162 | |
163 | seq_printf(m, |
164 | fmt: "Percpu Memory Statistics\n" |
165 | "Allocation Info:\n" |
166 | "----------------------------------------\n" ); |
167 | PL(unit_size); |
168 | PL(static_size); |
169 | PL(reserved_size); |
170 | PL(dyn_size); |
171 | PL(atom_size); |
172 | PL(alloc_size); |
173 | seq_putc(m, c: '\n'); |
174 | |
175 | #undef PL |
176 | |
177 | #define PU(X) \ |
178 | seq_printf(m, " %-20s: %12llu\n", #X, (unsigned long long)pcpu_stats.X) |
179 | |
180 | seq_printf(m, |
181 | fmt: "Global Stats:\n" |
182 | "----------------------------------------\n" ); |
183 | PU(nr_alloc); |
184 | PU(nr_dealloc); |
185 | PU(nr_cur_alloc); |
186 | PU(nr_max_alloc); |
187 | PU(nr_chunks); |
188 | PU(nr_max_chunks); |
189 | PU(min_alloc_size); |
190 | PU(max_alloc_size); |
191 | P("empty_pop_pages" , pcpu_nr_empty_pop_pages); |
192 | seq_putc(m, c: '\n'); |
193 | |
194 | #undef PU |
195 | |
196 | seq_printf(m, |
197 | fmt: "Per Chunk Stats:\n" |
198 | "----------------------------------------\n" ); |
199 | |
200 | if (pcpu_reserved_chunk) { |
201 | seq_puts(m, s: "Chunk: <- Reserved Chunk\n" ); |
202 | chunk_map_stats(m, chunk: pcpu_reserved_chunk, buffer); |
203 | } |
204 | |
205 | for (slot = 0; slot < pcpu_nr_slots; slot++) { |
206 | list_for_each_entry(chunk, &pcpu_chunk_lists[slot], list) { |
207 | if (chunk == pcpu_first_chunk) |
208 | seq_puts(m, s: "Chunk: <- First Chunk\n" ); |
209 | else if (slot == pcpu_to_depopulate_slot) |
210 | seq_puts(m, s: "Chunk (to_depopulate)\n" ); |
211 | else if (slot == pcpu_sidelined_slot) |
212 | seq_puts(m, s: "Chunk (sidelined):\n" ); |
213 | else |
214 | seq_puts(m, s: "Chunk:\n" ); |
215 | chunk_map_stats(m, chunk, buffer); |
216 | } |
217 | } |
218 | |
219 | spin_unlock_irq(lock: &pcpu_lock); |
220 | |
221 | vfree(addr: buffer); |
222 | |
223 | return 0; |
224 | } |
225 | DEFINE_SHOW_ATTRIBUTE(percpu_stats); |
226 | |
227 | static int __init init_percpu_stats_debugfs(void) |
228 | { |
229 | debugfs_create_file(name: "percpu_stats" , mode: 0444, NULL, NULL, |
230 | fops: &percpu_stats_fops); |
231 | |
232 | return 0; |
233 | } |
234 | |
235 | late_initcall(init_percpu_stats_debugfs); |
236 | |