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 "errcode.h"
  10 #include "extents.h"
  11 #include "io.h"
  12 #include "move.h"
  13 #include "rebalance.h"
  14 #include "super-io.h"
  15 #include "trace.h"
  16
  17 #include <linux/freezer.h>
  18 #include <linux/kthread.h>
  19 #include <linux/sched/cputime.h>
  20
  21 /*
  22  * Check if an extent should be moved:
  23  * returns -1 if it should not be moved, or
  24  * device of pointer that should be moved, if known, or INT_MAX if unknown
  25  */
  26 static bool rebalance_pred(struct bch_fs *c, void *arg,
  27                            struct bkey_s_c k,
  28                            struct bch_io_opts *io_opts,
  29                            struct data_update_opts *data_opts)
  30 {
  31         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
  32         unsigned i;
  33
  34         data_opts->rewrite_ptrs         = 0;
  35         data_opts->target               = io_opts->background_target;
  36         data_opts->extra_replicas       = 0;
  37         data_opts->btree_insert_flags   = 0;
  38
  39         if (io_opts->background_compression &&
  40             !bch2_bkey_is_incompressible(k)) {
  41                 const union bch_extent_entry *entry;
  42                 struct extent_ptr_decoded p;
  43
  44                 i = 0;
  45                 bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
  46                         if (!p.ptr.cached &&
  47                             p.crc.compression_type !=
  48                             bch2_compression_opt_to_type[io_opts->background_compression])
  49                                 data_opts->rewrite_ptrs |= 1U << i;
  50                         i++;
  51                 }
  52         }
  53
  54         if (io_opts->background_target) {
  55                 const struct bch_extent_ptr *ptr;
  56
  57                 i = 0;
  58                 bkey_for_each_ptr(ptrs, ptr) {
  59                         if (!ptr->cached &&
  60                             !bch2_dev_in_target(c, ptr->dev, io_opts->background_target) &&
  61                             bch2_target_accepts_data(c, BCH_DATA_user, io_opts->background_target))
  62                                 data_opts->rewrite_ptrs |= 1U << i;
  63                         i++;
  64                 }
  65         }
  66
  67         return data_opts->rewrite_ptrs != 0;
  68 }
  69
  70 void bch2_rebalance_add_key(struct bch_fs *c,
  71                             struct bkey_s_c k,
  72                             struct bch_io_opts *io_opts)
  73 {
  74         struct data_update_opts update_opts = { 0 };
  75         struct bkey_ptrs_c ptrs;
  76         const struct bch_extent_ptr *ptr;
  77         unsigned i;
  78
  79         if (!rebalance_pred(c, NULL, k, io_opts, &update_opts))
  80                 return;
  81
  82         i = 0;
  83         ptrs = bch2_bkey_ptrs_c(k);
  84         bkey_for_each_ptr(ptrs, ptr) {
  85                 if ((1U << i) && update_opts.rewrite_ptrs)
  86                         if (atomic64_add_return(k.k->size,
  87                                         &bch_dev_bkey_exists(c, ptr->dev)->rebalance_work) ==
  88                             k.k->size)
  89                                 rebalance_wakeup(c);
  90                 i++;
  91         }
  92 }
  93
  94 void bch2_rebalance_add_work(struct bch_fs *c, u64 sectors)
  95 {
  96         if (atomic64_add_return(sectors, &c->rebalance.work_unknown_dev) ==
  97             sectors)
  98                 rebalance_wakeup(c);
  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         struct bch_move_stats move_stats;
 180         unsigned long start, prev_start;
 181         unsigned long prev_run_time, prev_run_cputime;
 182         unsigned long cputime, prev_cputime;
 183         u64 io_start;
 184         long throttle;
 185
 186         set_freezable();
 187
 188         io_start        = atomic64_read(&clock->now);
 189         p               = rebalance_work(c);
 190         prev_start      = jiffies;
 191         prev_cputime    = curr_cputime();
 192
 193         bch2_move_stats_init(&move_stats, "rebalance");
 194         while (!kthread_wait_freezable(r->enabled)) {
 195                 cond_resched();
 196
 197                 start                   = jiffies;
 198                 cputime                 = curr_cputime();
 199
 200                 prev_run_time           = start - prev_start;
 201                 prev_run_cputime        = cputime - prev_cputime;
 202
 203                 w                       = rebalance_work(c);
 204                 BUG_ON(!w.dev_most_full_capacity);
 205
 206                 if (!w.total_work) {
 207                         r->state = REBALANCE_WAITING;
 208                         kthread_wait_freezable(rebalance_work(c).total_work);
 209                         continue;
 210                 }
 211
 212                 /*
 213                  * If there isn't much work to do, throttle cpu usage:
 214                  */
 215                 throttle = prev_run_cputime * 100 /
 216                         max(1U, w.dev_most_full_percent) -
 217                         prev_run_time;
 218
 219                 if (w.dev_most_full_percent < 20 && throttle > 0) {
 220                         r->throttled_until_iotime = io_start +
 221                                 div_u64(w.dev_most_full_capacity *
 222                                         (20 - w.dev_most_full_percent),
 223                                         50);
 224
 225                         if (atomic64_read(&clock->now) + clock->max_slop <
 226                             r->throttled_until_iotime) {
 227                                 r->throttled_until_cputime = start + throttle;
 228                                 r->state = REBALANCE_THROTTLED;
 229
 230                                 bch2_kthread_io_clock_wait(clock,
 231                                         r->throttled_until_iotime,
 232                                         throttle);
 233                                 continue;
 234                         }
 235                 }
 236
 237                 /* minimum 1 mb/sec: */
 238                 r->pd.rate.rate =
 239                         max_t(u64, 1 << 11,
 240                               r->pd.rate.rate *
 241                               max(p.dev_most_full_percent, 1U) /
 242                               max(w.dev_most_full_percent, 1U));
 243
 244                 io_start        = atomic64_read(&clock->now);
 245                 p               = w;
 246                 prev_start      = start;
 247                 prev_cputime    = cputime;
 248
 249                 r->state = REBALANCE_RUNNING;
 250                 memset(&move_stats, 0, sizeof(move_stats));
 251                 rebalance_work_reset(c);
 252
 253                 bch2_move_data(c,
 254                                0,               POS_MIN,
 255                                BTREE_ID_NR,     POS_MAX,
 256                                /* ratelimiting disabled for now */
 257                                NULL, /*  &r->pd.rate, */
 258                                &move_stats,
 259                                writepoint_ptr(&c->rebalance_write_point),
 260                                true,
 261                                rebalance_pred, NULL);
 262         }
 263
 264         return 0;
 265 }
 266
 267 void bch2_rebalance_work_to_text(struct printbuf *out, struct bch_fs *c)
 268 {
 269         struct bch_fs_rebalance *r = &c->rebalance;
 270         struct rebalance_work w = rebalance_work(c);
 271
 272         if (!out->nr_tabstops)
 273                 printbuf_tabstop_push(out, 20);
 274
 275         prt_printf(out, "fullest_dev (%i):", w.dev_most_full_idx);
 276         prt_tab(out);
 277
 278         prt_human_readable_u64(out, w.dev_most_full_work << 9);
 279         prt_printf(out, "/");
 280         prt_human_readable_u64(out, w.dev_most_full_capacity << 9);
 281         prt_newline(out);
 282
 283         prt_printf(out, "total work:");
 284         prt_tab(out);
 285
 286         prt_human_readable_u64(out, w.total_work << 9);
 287         prt_printf(out, "/");
 288         prt_human_readable_u64(out, c->capacity << 9);
 289         prt_newline(out);
 290
 291         prt_printf(out, "rate:");
 292         prt_tab(out);
 293         prt_printf(out, "%u", r->pd.rate.rate);
 294         prt_newline(out);
 295
 296         switch (r->state) {
 297         case REBALANCE_WAITING:
 298                 prt_printf(out, "waiting");
 299                 break;
 300         case REBALANCE_THROTTLED:
 301                 prt_printf(out, "throttled for %lu sec or ",
 302                        (r->throttled_until_cputime - jiffies) / HZ);
 303                 prt_human_readable_u64(out,
 304                             (r->throttled_until_iotime -
 305                              atomic64_read(&c->io_clock[WRITE].now)) << 9);
 306                 prt_printf(out, " io");
 307                 break;
 308         case REBALANCE_RUNNING:
 309                 prt_printf(out, "running");
 310                 break;
 311         }
 312         prt_newline(out);
 313 }
 314
 315 void bch2_rebalance_stop(struct bch_fs *c)
 316 {
 317         struct task_struct *p;
 318
 319         c->rebalance.pd.rate.rate = UINT_MAX;
 320         bch2_ratelimit_reset(&c->rebalance.pd.rate);
 321
 322         p = rcu_dereference_protected(c->rebalance.thread, 1);
 323         c->rebalance.thread = NULL;
 324
 325         if (p) {
 326                 /* for sychronizing with rebalance_wakeup() */
 327                 synchronize_rcu();
 328
 329                 kthread_stop(p);
 330                 put_task_struct(p);
 331         }
 332 }
 333
 334 int bch2_rebalance_start(struct bch_fs *c)
 335 {
 336         struct task_struct *p;
 337         int ret;
 338
 339         if (c->rebalance.thread)
 340                 return 0;
 341
 342         if (c->opts.nochanges)
 343                 return 0;
 344
 345         p = kthread_create(bch2_rebalance_thread, c, "bch-rebalance/%s", c->name);
 346         ret = PTR_ERR_OR_ZERO(p);
 347         if (ret) {
 348                 bch_err(c, "error creating rebalance thread: %s", bch2_err_str(ret));
 349                 return ret;
 350         }
 351
 352         get_task_struct(p);
 353         rcu_assign_pointer(c->rebalance.thread, p);
 354         wake_up_process(p);
 355         return 0;
 356 }
 357
 358 void bch2_fs_rebalance_init(struct bch_fs *c)
 359 {
 360         bch2_pd_controller_init(&c->rebalance.pd);
 361
 362         atomic64_set(&c->rebalance.work_unknown_dev, S64_MAX);
 363 }