mm/percpu-stats.c

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