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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "alloc_foreground.h"
   5 #include "btree_iter.h"
   6 #include "buckets.h"
   7 #include "clock.h"
   8 #include "disk_groups.h"
   9 #include "extents.h"
  10 #include "io.h"
  11 #include "move.h"
  12 #include "rebalance.h"
  13 #include "super-io.h"
  14
  15 #include <linux/freezer.h>
  16 #include <linux/kthread.h>
  17 #include <linux/sched/cputime.h>
  18 #include <trace/events/bcachefs.h>
  19
  20 static inline bool rebalance_ptr_pred(struct bch_fs *c,
  21                                       struct extent_ptr_decoded p,
  22                                       struct bch_io_opts *io_opts)
  23 {
  24         if (io_opts->background_target &&
  25             !bch2_dev_in_target(c, p.ptr.dev, io_opts->background_target) &&
  26             !p.ptr.cached)
  27                 return true;
  28
  29         if (io_opts->background_compression &&
  30             p.crc.compression_type !=
  31             bch2_compression_opt_to_type[io_opts->background_compression])
  32                 return true;
  33
  34         return false;
  35 }
  36
  37 void bch2_rebalance_add_key(struct bch_fs *c,
  38                             struct bkey_s_c k,
  39                             struct bch_io_opts *io_opts)
  40 {
  41         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
  42         const union bch_extent_entry *entry;
  43         struct extent_ptr_decoded p;
  44
  45         if (!bkey_extent_is_data(k.k))
  46                 return;
  47
  48         if (!io_opts->background_target &&
  49             !io_opts->background_compression)
  50                 return;
  51
  52         bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
  53                 if (rebalance_ptr_pred(c, p, io_opts)) {
  54                         struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
  55
  56                         if (atomic64_add_return(p.crc.compressed_size,
  57                                                 &ca->rebalance_work) ==
  58                             p.crc.compressed_size)
  59                                 rebalance_wakeup(c);
  60                 }
  61 }
  62
  63 void bch2_rebalance_add_work(struct bch_fs *c, u64 sectors)
  64 {
  65         if (atomic64_add_return(sectors, &c->rebalance.work_unknown_dev) ==
  66             sectors)
  67                 rebalance_wakeup(c);
  68 }
  69
  70 static enum data_cmd rebalance_pred(struct bch_fs *c, void *arg,
  71                                     struct bkey_s_c k,
  72                                     struct bch_io_opts *io_opts,
  73                                     struct data_opts *data_opts)
  74 {
  75         switch (k.k->type) {
  76         case KEY_TYPE_extent: {
  77                 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
  78                 const union bch_extent_entry *entry;
  79                 struct extent_ptr_decoded p;
  80
  81                 /* Make sure we have room to add a new pointer: */
  82                 if (bkey_val_u64s(e.k) + BKEY_EXTENT_PTR_U64s_MAX >
  83                     BKEY_EXTENT_VAL_U64s_MAX)
  84                         return DATA_SKIP;
  85
  86                 extent_for_each_ptr_decode(e, p, entry)
  87                         if (rebalance_ptr_pred(c, p, io_opts))
  88                                 goto found;
  89
  90                 return DATA_SKIP;
  91 found:
  92                 data_opts->target               = io_opts->background_target;
  93                 data_opts->btree_insert_flags   = 0;
  94                 return DATA_ADD_REPLICAS;
  95         }
  96         default:
  97                 return DATA_SKIP;
  98         }
  99 }
 100
 101 struct rebalance_work {
 102         int             dev_most_full_idx;
 103         unsigned        dev_most_full_percent;
 104         u64             dev_most_full_work;
 105         u64             dev_most_full_capacity;
 106         u64             total_work;
 107 };
 108
 109 static void rebalance_work_accumulate(struct rebalance_work *w,
 110                 u64 dev_work, u64 unknown_dev, u64 capacity, int idx)
 111 {
 112         unsigned percent_full;
 113         u64 work = dev_work + unknown_dev;
 114
 115         if (work < dev_work || work < unknown_dev)
 116                 work = U64_MAX;
 117         work = min(work, capacity);
 118
 119         percent_full = div64_u64(work * 100, capacity);
 120
 121         if (percent_full >= w->dev_most_full_percent) {
 122                 w->dev_most_full_idx            = idx;
 123                 w->dev_most_full_percent        = percent_full;
 124                 w->dev_most_full_work           = work;
 125                 w->dev_most_full_capacity       = capacity;
 126         }
 127
 128         if (w->total_work + dev_work >= w->total_work &&
 129             w->total_work + dev_work >= dev_work)
 130                 w->total_work += dev_work;
 131 }
 132
 133 static struct rebalance_work rebalance_work(struct bch_fs *c)
 134 {
 135         struct bch_dev *ca;
 136         struct rebalance_work ret = { .dev_most_full_idx = -1 };
 137         u64 unknown_dev = atomic64_read(&c->rebalance.work_unknown_dev);
 138         unsigned i;
 139
 140         for_each_online_member(ca, c, i)
 141                 rebalance_work_accumulate(&ret,
 142                         atomic64_read(&ca->rebalance_work),
 143                         unknown_dev,
 144                         bucket_to_sector(ca, ca->mi.nbuckets -
 145                                          ca->mi.first_bucket),
 146                         i);
 147
 148         rebalance_work_accumulate(&ret,
 149                 unknown_dev, 0, c->capacity, -1);
 150
 151         return ret;
 152 }
 153
 154 static void rebalance_work_reset(struct bch_fs *c)
 155 {
 156         struct bch_dev *ca;
 157         unsigned i;
 158
 159         for_each_online_member(ca, c, i)
 160                 atomic64_set(&ca->rebalance_work, 0);
 161
 162         atomic64_set(&c->rebalance.work_unknown_dev, 0);
 163 }
 164
 165 static unsigned long curr_cputime(void)
 166 {
 167         u64 utime, stime;
 168
 169         task_cputime_adjusted(current, &utime, &stime);
 170         return nsecs_to_jiffies(utime + stime);
 171 }
 172
 173 static int bch2_rebalance_thread(void *arg)
 174 {
 175         struct bch_fs *c = arg;
 176         struct bch_fs_rebalance *r = &c->rebalance;
 177         struct io_clock *clock = &c->io_clock[WRITE];
 178         struct rebalance_work w, p;
 179         unsigned long start, prev_start;
 180         unsigned long prev_run_time, prev_run_cputime;
 181         unsigned long cputime, prev_cputime;
 182         unsigned long io_start;
 183         long throttle;
 184
 185         set_freezable();
 186
 187         io_start        = atomic_long_read(&clock->now);
 188         p               = rebalance_work(c);
 189         prev_start      = jiffies;
 190         prev_cputime    = curr_cputime();
 191
 192         while (!kthread_wait_freezable(r->enabled)) {
 193                 start                   = jiffies;
 194                 cputime                 = curr_cputime();
 195
 196                 prev_run_time           = start - prev_start;
 197                 prev_run_cputime        = cputime - prev_cputime;
 198
 199                 w                       = rebalance_work(c);
 200                 BUG_ON(!w.dev_most_full_capacity);
 201
 202                 if (!w.total_work) {
 203                         r->state = REBALANCE_WAITING;
 204                         kthread_wait_freezable(rebalance_work(c).total_work);
 205                         continue;
 206                 }
 207
 208                 /*
 209                  * If there isn't much work to do, throttle cpu usage:
 210                  */
 211                 throttle = prev_run_cputime * 100 /
 212                         max(1U, w.dev_most_full_percent) -
 213                         prev_run_time;
 214
 215                 if (w.dev_most_full_percent < 20 && throttle > 0) {
 216                         r->state = REBALANCE_THROTTLED;
 217                         r->throttled_until_iotime = io_start +
 218                                 div_u64(w.dev_most_full_capacity *
 219                                         (20 - w.dev_most_full_percent),
 220                                         50);
 221                         r->throttled_until_cputime = start + throttle;
 222
 223                         bch2_kthread_io_clock_wait(clock,
 224                                 r->throttled_until_iotime,
 225                                 throttle);
 226                         continue;
 227                 }
 228
 229                 /* minimum 1 mb/sec: */
 230                 r->pd.rate.rate =
 231                         max_t(u64, 1 << 11,
 232                               r->pd.rate.rate *
 233                               max(p.dev_most_full_percent, 1U) /
 234                               max(w.dev_most_full_percent, 1U));
 235
 236                 io_start        = atomic_long_read(&clock->now);
 237                 p               = w;
 238                 prev_start      = start;
 239                 prev_cputime    = cputime;
 240
 241                 r->state = REBALANCE_RUNNING;
 242                 memset(&r->move_stats, 0, sizeof(r->move_stats));
 243                 rebalance_work_reset(c);
 244
 245                 bch2_move_data(c,
 246                                /* ratelimiting disabled for now */
 247                                NULL, /*  &r->pd.rate, */
 248                                writepoint_ptr(&c->rebalance_write_point),
 249                                POS_MIN, POS_MAX,
 250                                rebalance_pred, NULL,
 251                                &r->move_stats);
 252         }
 253
 254         return 0;
 255 }
 256
 257 ssize_t bch2_rebalance_work_show(struct bch_fs *c, char *buf)
 258 {
 259         struct printbuf out = _PBUF(buf, PAGE_SIZE);
 260         struct bch_fs_rebalance *r = &c->rebalance;
 261         struct rebalance_work w = rebalance_work(c);
 262         char h1[21], h2[21];
 263
 264         bch2_hprint(&PBUF(h1), w.dev_most_full_work << 9);
 265         bch2_hprint(&PBUF(h2), w.dev_most_full_capacity << 9);
 266         pr_buf(&out, "fullest_dev (%i):\t%s/%s\n",
 267                w.dev_most_full_idx, h1, h2);
 268
 269         bch2_hprint(&PBUF(h1), w.total_work << 9);
 270         bch2_hprint(&PBUF(h2), c->capacity << 9);
 271         pr_buf(&out, "total work:\t\t%s/%s\n", h1, h2);
 272
 273         pr_buf(&out, "rate:\t\t\t%u\n", r->pd.rate.rate);
 274
 275         switch (r->state) {
 276         case REBALANCE_WAITING:
 277                 pr_buf(&out, "waiting\n");
 278                 break;
 279         case REBALANCE_THROTTLED:
 280                 bch2_hprint(&PBUF(h1),
 281                             (r->throttled_until_iotime -
 282                              atomic_long_read(&c->io_clock[WRITE].now)) << 9);
 283                 pr_buf(&out, "throttled for %lu sec or %s io\n",
 284                        (r->throttled_until_cputime - jiffies) / HZ,
 285                        h1);
 286                 break;
 287         case REBALANCE_RUNNING:
 288                 pr_buf(&out, "running\n");
 289                 pr_buf(&out, "pos %llu:%llu\n",
 290                        r->move_stats.pos.inode,
 291                        r->move_stats.pos.offset);
 292                 break;
 293         }
 294
 295         return out.pos - buf;
 296 }
 297
 298 void bch2_rebalance_stop(struct bch_fs *c)
 299 {
 300         struct task_struct *p;
 301
 302         c->rebalance.pd.rate.rate = UINT_MAX;
 303         bch2_ratelimit_reset(&c->rebalance.pd.rate);
 304
 305         p = rcu_dereference_protected(c->rebalance.thread, 1);
 306         c->rebalance.thread = NULL;
 307
 308         if (p) {
 309                 /* for sychronizing with rebalance_wakeup() */
 310                 synchronize_rcu();
 311
 312                 kthread_stop(p);
 313                 put_task_struct(p);
 314         }
 315 }
 316
 317 int bch2_rebalance_start(struct bch_fs *c)
 318 {
 319         struct task_struct *p;
 320
 321         if (c->opts.nochanges)
 322                 return 0;
 323
 324         p = kthread_create(bch2_rebalance_thread, c, "bch_rebalance");
 325         if (IS_ERR(p))
 326                 return PTR_ERR(p);
 327
 328         get_task_struct(p);
 329         rcu_assign_pointer(c->rebalance.thread, p);
 330         wake_up_process(p);
 331         return 0;
 332 }
 333
 334 void bch2_fs_rebalance_init(struct bch_fs *c)
 335 {
 336         bch2_pd_controller_init(&c->rebalance.pd);
 337
 338         atomic64_set(&c->rebalance.work_unknown_dev, S64_MAX);
 339 }