git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/movinggc.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Moving/copying garbage collector
   4  *
   5  * Copyright 2012 Google, Inc.
   6  */
   7
   8 #include "bcachefs.h"
   9 #include "alloc_background.h"
  10 #include "alloc_foreground.h"
  11 #include "btree_iter.h"
  12 #include "btree_update.h"
  13 #include "btree_write_buffer.h"
  14 #include "buckets.h"
  15 #include "clock.h"
  16 #include "disk_groups.h"
  17 #include "errcode.h"
  18 #include "error.h"
  19 #include "extents.h"
  20 #include "eytzinger.h"
  21 #include "io.h"
  22 #include "keylist.h"
  23 #include "lru.h"
  24 #include "move.h"
  25 #include "movinggc.h"
  26 #include "super-io.h"
  27
  28 #include <trace/events/bcachefs.h>
  29 #include <linux/freezer.h>
  30 #include <linux/kthread.h>
  31 #include <linux/math64.h>
  32 #include <linux/sched/task.h>
  33 #include <linux/sort.h>
  34 #include <linux/wait.h>
  35
  36 static int bch2_bucket_is_movable(struct btree_trans *trans,
  37                                   struct bpos bucket, u64 time, u8 *gen)
  38 {
  39         struct btree_iter iter;
  40         struct bkey_s_c k;
  41         struct bch_alloc_v4 _a;
  42         const struct bch_alloc_v4 *a;
  43         int ret;
  44
  45         if (bch2_bucket_is_open(trans->c, bucket.inode, bucket.offset))
  46                 return 0;
  47
  48         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, bucket, 0);
  49         k = bch2_btree_iter_peek_slot(&iter);
  50         ret = bkey_err(k);
  51         bch2_trans_iter_exit(trans, &iter);
  52
  53         if (ret)
  54                 return ret;
  55
  56         a = bch2_alloc_to_v4(k, &_a);
  57         *gen = a->gen;
  58         ret = (a->data_type == BCH_DATA_btree ||
  59                a->data_type == BCH_DATA_user) &&
  60                 a->fragmentation_lru &&
  61                 a->fragmentation_lru <= time;
  62
  63         if (ret) {
  64                 struct printbuf buf = PRINTBUF;
  65
  66                 bch2_bkey_val_to_text(&buf, trans->c, k);
  67                 pr_debug("%s", buf.buf);
  68                 printbuf_exit(&buf);
  69         }
  70
  71         return ret;
  72 }
  73
  74 static int bch2_copygc_next_bucket(struct btree_trans *trans,
  75                                    struct bpos *bucket, u8 *gen, struct bpos *pos)
  76 {
  77         struct btree_iter iter;
  78         struct bkey_s_c k;
  79         int ret;
  80
  81         ret = for_each_btree_key2_upto(trans, iter, BTREE_ID_lru,
  82                                   bpos_max(*pos, lru_pos(BCH_LRU_FRAGMENTATION_START, 0, 0)),
  83                                   lru_pos(BCH_LRU_FRAGMENTATION_START, U64_MAX, LRU_TIME_MAX),
  84                                   0, k, ({
  85                 *bucket = u64_to_bucket(k.k->p.offset);
  86
  87                 bch2_bucket_is_movable(trans, *bucket, lru_pos_time(k.k->p), gen);
  88         }));
  89
  90         *pos = iter.pos;
  91         if (ret < 0)
  92                 return ret;
  93         return ret ? 0 : -ENOENT;
  94 }
  95
  96 static int bch2_copygc(struct bch_fs *c)
  97 {
  98         struct bch_move_stats move_stats;
  99         struct btree_trans trans;
 100         struct moving_context ctxt;
 101         struct data_update_opts data_opts = {
 102                 .btree_insert_flags = BTREE_INSERT_USE_RESERVE|JOURNAL_WATERMARK_copygc,
 103         };
 104         struct bpos bucket;
 105         struct bpos pos;
 106         u8 gen = 0;
 107         unsigned nr_evacuated;
 108         int ret = 0;
 109
 110         bch2_move_stats_init(&move_stats, "copygc");
 111         bch2_moving_ctxt_init(&ctxt, c, NULL, &move_stats,
 112                               writepoint_ptr(&c->copygc_write_point),
 113                               false);
 114         bch2_trans_init(&trans, c, 0, 0);
 115
 116         ret = bch2_btree_write_buffer_flush(&trans);
 117         BUG_ON(ret);
 118
 119         for (nr_evacuated = 0, pos = POS_MIN;
 120              nr_evacuated < 32 && !ret;
 121              nr_evacuated++, pos = bpos_nosnap_successor(pos)) {
 122                 ret = bch2_copygc_next_bucket(&trans, &bucket, &gen, &pos) ?:
 123                         __bch2_evacuate_bucket(&trans, &ctxt, bucket, gen, data_opts);
 124                 if (bkey_eq(pos, POS_MAX))
 125                         break;
 126         }
 127
 128         bch2_trans_exit(&trans);
 129         bch2_moving_ctxt_exit(&ctxt);
 130
 131         /* no entries in LRU btree found, or got to end: */
 132         if (ret == -ENOENT)
 133                 ret = 0;
 134
 135         if (ret < 0 && !bch2_err_matches(ret, EROFS))
 136                 bch_err(c, "error from bch2_move_data() in copygc: %s", bch2_err_str(ret));
 137
 138         trace_and_count(c, copygc, c, atomic64_read(&move_stats.sectors_moved), 0, 0, 0);
 139         return ret;
 140 }
 141
 142 /*
 143  * Copygc runs when the amount of fragmented data is above some arbitrary
 144  * threshold:
 145  *
 146  * The threshold at the limit - when the device is full - is the amount of space
 147  * we reserved in bch2_recalc_capacity; we can't have more than that amount of
 148  * disk space stranded due to fragmentation and store everything we have
 149  * promised to store.
 150  *
 151  * But we don't want to be running copygc unnecessarily when the device still
 152  * has plenty of free space - rather, we want copygc to smoothly run every so
 153  * often and continually reduce the amount of fragmented space as the device
 154  * fills up. So, we increase the threshold by half the current free space.
 155  */
 156 unsigned long bch2_copygc_wait_amount(struct bch_fs *c)
 157 {
 158         struct bch_dev *ca;
 159         unsigned dev_idx;
 160         s64 wait = S64_MAX, fragmented_allowed, fragmented;
 161
 162         for_each_rw_member(ca, c, dev_idx) {
 163                 struct bch_dev_usage usage = bch2_dev_usage_read(ca);
 164
 165                 fragmented_allowed = ((__dev_buckets_available(ca, usage, RESERVE_none) *
 166                                        ca->mi.bucket_size) >> 1);
 167                 fragmented = usage.d[BCH_DATA_user].fragmented;
 168
 169                 wait = min(wait, max(0LL, fragmented_allowed - fragmented));
 170         }
 171
 172         return wait;
 173 }
 174
 175 void bch2_copygc_wait_to_text(struct printbuf *out, struct bch_fs *c)
 176 {
 177         prt_printf(out, "Currently waiting for:     ");
 178         prt_human_readable_u64(out, max(0LL, c->copygc_wait -
 179                                         atomic64_read(&c->io_clock[WRITE].now)) << 9);
 180         prt_newline(out);
 181
 182         prt_printf(out, "Currently calculated wait: ");
 183         prt_human_readable_u64(out, bch2_copygc_wait_amount(c));
 184         prt_newline(out);
 185 }
 186
 187 static int bch2_copygc_thread(void *arg)
 188 {
 189         struct bch_fs *c = arg;
 190         struct io_clock *clock = &c->io_clock[WRITE];
 191         u64 last, wait;
 192         int ret = 0;
 193
 194         set_freezable();
 195
 196         while (!ret && !kthread_should_stop()) {
 197                 cond_resched();
 198
 199                 if (kthread_wait_freezable(c->copy_gc_enabled))
 200                         break;
 201
 202                 last = atomic64_read(&clock->now);
 203                 wait = bch2_copygc_wait_amount(c);
 204
 205                 if (wait > clock->max_slop) {
 206                         trace_and_count(c, copygc_wait, c, wait, last + wait);
 207                         c->copygc_wait = last + wait;
 208                         bch2_kthread_io_clock_wait(clock, last + wait,
 209                                         MAX_SCHEDULE_TIMEOUT);
 210                         continue;
 211                 }
 212
 213                 c->copygc_wait = 0;
 214
 215                 c->copygc_running = true;
 216                 ret = bch2_copygc(c);
 217                 c->copygc_running = false;
 218
 219                 wake_up(&c->copygc_running_wq);
 220         }
 221
 222         return 0;
 223 }
 224
 225 void bch2_copygc_stop(struct bch_fs *c)
 226 {
 227         if (c->copygc_thread) {
 228                 kthread_stop(c->copygc_thread);
 229                 put_task_struct(c->copygc_thread);
 230         }
 231         c->copygc_thread = NULL;
 232 }
 233
 234 int bch2_copygc_start(struct bch_fs *c)
 235 {
 236         struct task_struct *t;
 237         int ret;
 238
 239         if (c->copygc_thread)
 240                 return 0;
 241
 242         if (c->opts.nochanges)
 243                 return 0;
 244
 245         if (bch2_fs_init_fault("copygc_start"))
 246                 return -ENOMEM;
 247
 248         t = kthread_create(bch2_copygc_thread, c, "bch-copygc/%s", c->name);
 249         ret = PTR_ERR_OR_ZERO(t);
 250         if (ret) {
 251                 bch_err(c, "error creating copygc thread: %s", bch2_err_str(ret));
 252                 return ret;
 253         }
 254
 255         get_task_struct(t);
 256
 257         c->copygc_thread = t;
 258         wake_up_process(c->copygc_thread);
 259
 260         return 0;
 261 }
 262
 263 void bch2_fs_copygc_init(struct bch_fs *c)
 264 {
 265         init_waitqueue_head(&c->copygc_running_wq);
 266         c->copygc_running = false;
 267 }