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 /*
  21  * Check if an extent should be moved:
  22  * returns -1 if it should not be moved, or
  23  * device of pointer that should be moved, if known, or INT_MAX if unknown
  24  */
  25 static bool rebalance_pred(struct bch_fs *c, void *arg,
  26                            struct bkey_s_c k,
  27                            struct bch_io_opts *io_opts,
  28                            struct data_update_opts *data_opts)
  29 {
  30         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
  31         unsigned i;
  32
  33         data_opts->rewrite_ptrs         = 0;
  34         data_opts->target               = io_opts->background_target;
  35         data_opts->extra_replicas       = 0;
  36         data_opts->btree_insert_flags   = 0;
  37
  38         if (io_opts->background_compression &&
  39             !bch2_bkey_is_incompressible(k)) {
  40                 const union bch_extent_entry *entry;
  41                 struct extent_ptr_decoded p;
  42
  43                 i = 0;
  44                 bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
  45                         if (!p.ptr.cached &&
  46                             p.crc.compression_type !=
  47                             bch2_compression_opt_to_type[io_opts->background_compression])
  48                                 data_opts->rewrite_ptrs |= 1U << i;
  49                         i++;
  50                 }
  51         }
  52
  53         if (io_opts->background_target) {
  54                 const struct bch_extent_ptr *ptr;
  55
  56                 i = 0;
  57                 bkey_for_each_ptr(ptrs, ptr) {
  58                         if (!ptr->cached &&
  59                             !bch2_dev_in_target(c, ptr->dev, io_opts->background_target))
  60                                 data_opts->rewrite_ptrs |= 1U << i;
  61                         i++;
  62                 }
  63         }
  64
  65         return data_opts->rewrite_ptrs != 0;
  66 }
  67
  68 void bch2_rebalance_add_key(struct bch_fs *c,
  69                             struct bkey_s_c k,
  70                             struct bch_io_opts *io_opts)
  71 {
  72         struct data_update_opts update_opts = { 0 };
  73         struct bkey_ptrs_c ptrs;
  74         const struct bch_extent_ptr *ptr;
  75         unsigned i;
  76
  77         if (!rebalance_pred(c, NULL, k, io_opts, &update_opts))
  78                 return;
  79
  80         i = 0;
  81         ptrs = bch2_bkey_ptrs_c(k);
  82         bkey_for_each_ptr(ptrs, ptr) {
  83                 if ((1U << i) && update_opts.rewrite_ptrs)
  84                         if (atomic64_add_return(k.k->size,
  85                                         &bch_dev_bkey_exists(c, ptr->dev)->rebalance_work) ==
  86                             k.k->size)
  87                                 rebalance_wakeup(c);
  88                 i++;
  89         }
  90 }
  91
  92 void bch2_rebalance_add_work(struct bch_fs *c, u64 sectors)
  93 {
  94         if (atomic64_add_return(sectors, &c->rebalance.work_unknown_dev) ==
  95             sectors)
  96                 rebalance_wakeup(c);
  97 }
  98
  99 struct rebalance_work {
 100         int             dev_most_full_idx;
 101         unsigned        dev_most_full_percent;
 102         u64             dev_most_full_work;
 103         u64             dev_most_full_capacity;
 104         u64             total_work;
 105 };
 106
 107 static void rebalance_work_accumulate(struct rebalance_work *w,
 108                 u64 dev_work, u64 unknown_dev, u64 capacity, int idx)
 109 {
 110         unsigned percent_full;
 111         u64 work = dev_work + unknown_dev;
 112
 113         if (work < dev_work || work < unknown_dev)
 114                 work = U64_MAX;
 115         work = min(work, capacity);
 116
 117         percent_full = div64_u64(work * 100, capacity);
 118
 119         if (percent_full >= w->dev_most_full_percent) {
 120                 w->dev_most_full_idx            = idx;
 121                 w->dev_most_full_percent        = percent_full;
 122                 w->dev_most_full_work           = work;
 123                 w->dev_most_full_capacity       = capacity;
 124         }
 125
 126         if (w->total_work + dev_work >= w->total_work &&
 127             w->total_work + dev_work >= dev_work)
 128                 w->total_work += dev_work;
 129 }
 130
 131 static struct rebalance_work rebalance_work(struct bch_fs *c)
 132 {
 133         struct bch_dev *ca;
 134         struct rebalance_work ret = { .dev_most_full_idx = -1 };
 135         u64 unknown_dev = atomic64_read(&c->rebalance.work_unknown_dev);
 136         unsigned i;
 137
 138         for_each_online_member(ca, c, i)
 139                 rebalance_work_accumulate(&ret,
 140                         atomic64_read(&ca->rebalance_work),
 141                         unknown_dev,
 142                         bucket_to_sector(ca, ca->mi.nbuckets -
 143                                          ca->mi.first_bucket),
 144                         i);
 145
 146         rebalance_work_accumulate(&ret,
 147                 unknown_dev, 0, c->capacity, -1);
 148
 149         return ret;
 150 }
 151
 152 static void rebalance_work_reset(struct bch_fs *c)
 153 {
 154         struct bch_dev *ca;
 155         unsigned i;
 156
 157         for_each_online_member(ca, c, i)
 158                 atomic64_set(&ca->rebalance_work, 0);
 159
 160         atomic64_set(&c->rebalance.work_unknown_dev, 0);
 161 }
 162
 163 static unsigned long curr_cputime(void)
 164 {
 165         u64 utime, stime;
 166
 167         task_cputime_adjusted(current, &utime, &stime);
 168         return nsecs_to_jiffies(utime + stime);
 169 }
 170
 171 static int bch2_rebalance_thread(void *arg)
 172 {
 173         struct bch_fs *c = arg;
 174         struct bch_fs_rebalance *r = &c->rebalance;
 175         struct io_clock *clock = &c->io_clock[WRITE];
 176         struct rebalance_work w, p;
 177         struct bch_move_stats move_stats;
 178         unsigned long start, prev_start;
 179         unsigned long prev_run_time, prev_run_cputime;
 180         unsigned long cputime, prev_cputime;
 181         u64 io_start;
 182         long throttle;
 183
 184         set_freezable();
 185
 186         io_start        = atomic64_read(&clock->now);
 187         p               = rebalance_work(c);
 188         prev_start      = jiffies;
 189         prev_cputime    = curr_cputime();
 190
 191         bch_move_stats_init(&move_stats, "rebalance");
 192         while (!kthread_wait_freezable(r->enabled)) {
 193                 cond_resched();
 194
 195                 start                   = jiffies;
 196                 cputime                 = curr_cputime();
 197
 198                 prev_run_time           = start - prev_start;
 199                 prev_run_cputime        = cputime - prev_cputime;
 200
 201                 w                       = rebalance_work(c);
 202                 BUG_ON(!w.dev_most_full_capacity);
 203
 204                 if (!w.total_work) {
 205                         r->state = REBALANCE_WAITING;
 206                         kthread_wait_freezable(rebalance_work(c).total_work);
 207                         continue;
 208                 }
 209
 210                 /*
 211                  * If there isn't much work to do, throttle cpu usage:
 212                  */
 213                 throttle = prev_run_cputime * 100 /
 214                         max(1U, w.dev_most_full_percent) -
 215                         prev_run_time;
 216
 217                 if (w.dev_most_full_percent < 20 && throttle > 0) {
 218                         r->throttled_until_iotime = io_start +
 219                                 div_u64(w.dev_most_full_capacity *
 220                                         (20 - w.dev_most_full_percent),
 221                                         50);
 222
 223                         if (atomic64_read(&clock->now) + clock->max_slop <
 224                             r->throttled_until_iotime) {
 225                                 r->throttled_until_cputime = start + throttle;
 226                                 r->state = REBALANCE_THROTTLED;
 227
 228                                 bch2_kthread_io_clock_wait(clock,
 229                                         r->throttled_until_iotime,
 230                                         throttle);
 231                                 continue;
 232                         }
 233                 }
 234
 235                 /* minimum 1 mb/sec: */
 236                 r->pd.rate.rate =
 237                         max_t(u64, 1 << 11,
 238                               r->pd.rate.rate *
 239                               max(p.dev_most_full_percent, 1U) /
 240                               max(w.dev_most_full_percent, 1U));
 241
 242                 io_start        = atomic64_read(&clock->now);
 243                 p               = w;
 244                 prev_start      = start;
 245                 prev_cputime    = cputime;
 246
 247                 r->state = REBALANCE_RUNNING;
 248                 memset(&move_stats, 0, sizeof(move_stats));
 249                 rebalance_work_reset(c);
 250
 251                 bch2_move_data(c,
 252                                0,               POS_MIN,
 253                                BTREE_ID_NR,     POS_MAX,
 254                                /* ratelimiting disabled for now */
 255                                NULL, /*  &r->pd.rate, */
 256                                &move_stats,
 257                                writepoint_ptr(&c->rebalance_write_point),
 258                                true,
 259                                rebalance_pred, NULL);
 260         }
 261
 262         return 0;
 263 }
 264
 265 void bch2_rebalance_work_to_text(struct printbuf *out, struct bch_fs *c)
 266 {
 267         struct bch_fs_rebalance *r = &c->rebalance;
 268         struct rebalance_work w = rebalance_work(c);
 269
 270         out->tabstops[0] = 20;
 271
 272         prt_printf(out, "fullest_dev (%i):", w.dev_most_full_idx);
 273         prt_tab(out);
 274
 275         prt_human_readable_u64(out, w.dev_most_full_work << 9);
 276         prt_printf(out, "/");
 277         prt_human_readable_u64(out, w.dev_most_full_capacity << 9);
 278         prt_newline(out);
 279
 280         prt_printf(out, "total work:");
 281         prt_tab(out);
 282
 283         prt_human_readable_u64(out, w.total_work << 9);
 284         prt_printf(out, "/");
 285         prt_human_readable_u64(out, c->capacity << 9);
 286         prt_newline(out);
 287
 288         prt_printf(out, "rate:");
 289         prt_tab(out);
 290         prt_printf(out, "%u", r->pd.rate.rate);
 291         prt_newline(out);
 292
 293         switch (r->state) {
 294         case REBALANCE_WAITING:
 295                 prt_printf(out, "waiting");
 296                 break;
 297         case REBALANCE_THROTTLED:
 298                 prt_printf(out, "throttled for %lu sec or ",
 299                        (r->throttled_until_cputime - jiffies) / HZ);
 300                 prt_human_readable_u64(out,
 301                             (r->throttled_until_iotime -
 302                              atomic64_read(&c->io_clock[WRITE].now)) << 9);
 303                 prt_printf(out, " io");
 304                 break;
 305         case REBALANCE_RUNNING:
 306                 prt_printf(out, "running");
 307                 break;
 308         }
 309         prt_newline(out);
 310 }
 311
 312 void bch2_rebalance_stop(struct bch_fs *c)
 313 {
 314         struct task_struct *p;
 315
 316         c->rebalance.pd.rate.rate = UINT_MAX;
 317         bch2_ratelimit_reset(&c->rebalance.pd.rate);
 318
 319         p = rcu_dereference_protected(c->rebalance.thread, 1);
 320         c->rebalance.thread = NULL;
 321
 322         if (p) {
 323                 /* for sychronizing with rebalance_wakeup() */
 324                 synchronize_rcu();
 325
 326                 kthread_stop(p);
 327                 put_task_struct(p);
 328         }
 329 }
 330
 331 int bch2_rebalance_start(struct bch_fs *c)
 332 {
 333         struct task_struct *p;
 334
 335         if (c->rebalance.thread)
 336                 return 0;
 337
 338         if (c->opts.nochanges)
 339                 return 0;
 340
 341         p = kthread_create(bch2_rebalance_thread, c, "bch-rebalance/%s", c->name);
 342         if (IS_ERR(p)) {
 343                 bch_err(c, "error creating rebalance thread: %li", PTR_ERR(p));
 344                 return PTR_ERR(p);
 345         }
 346
 347         get_task_struct(p);
 348         rcu_assign_pointer(c->rebalance.thread, p);
 349         wake_up_process(p);
 350         return 0;
 351 }
 352
 353 void bch2_fs_rebalance_init(struct bch_fs *c)
 354 {
 355         bch2_pd_controller_init(&c->rebalance.pd);
 356
 357         atomic64_set(&c->rebalance.work_unknown_dev, S64_MAX);
 358 }