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
  16 #include <linux/freezer.h>
  17 #include <linux/kthread.h>
  18 #include <linux/sched/cputime.h>
  19 #include <trace/events/bcachefs.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                                 data_opts->rewrite_ptrs |= 1U << i;
  62                         i++;
  63                 }
  64         }
  65
  66         return data_opts->rewrite_ptrs != 0;
  67 }
  68
  69 void bch2_rebalance_add_key(struct bch_fs *c,
  70                             struct bkey_s_c k,
  71                             struct bch_io_opts *io_opts)
  72 {
  73         struct data_update_opts update_opts = { 0 };
  74         struct bkey_ptrs_c ptrs;
  75         const struct bch_extent_ptr *ptr;
  76         unsigned i;
  77
  78         if (!rebalance_pred(c, NULL, k, io_opts, &update_opts))
  79                 return;
  80
  81         i = 0;
  82         ptrs = bch2_bkey_ptrs_c(k);
  83         bkey_for_each_ptr(ptrs, ptr) {
  84                 if ((1U << i) && update_opts.rewrite_ptrs)
  85                         if (atomic64_add_return(k.k->size,
  86                                         &bch_dev_bkey_exists(c, ptr->dev)->rebalance_work) ==
  87                             k.k->size)
  88                                 rebalance_wakeup(c);
  89                 i++;
  90         }
  91 }
  92
  93 void bch2_rebalance_add_work(struct bch_fs *c, u64 sectors)
  94 {
  95         if (atomic64_add_return(sectors, &c->rebalance.work_unknown_dev) ==
  96             sectors)
  97                 rebalance_wakeup(c);
  98 }
  99
 100 struct rebalance_work {
 101         int             dev_most_full_idx;
 102         unsigned        dev_most_full_percent;
 103         u64             dev_most_full_work;
 104         u64             dev_most_full_capacity;
 105         u64             total_work;
 106 };
 107
 108 static void rebalance_work_accumulate(struct rebalance_work *w,
 109                 u64 dev_work, u64 unknown_dev, u64 capacity, int idx)
 110 {
 111         unsigned percent_full;
 112         u64 work = dev_work + unknown_dev;
 113
 114         if (work < dev_work || work < unknown_dev)
 115                 work = U64_MAX;
 116         work = min(work, capacity);
 117
 118         percent_full = div64_u64(work * 100, capacity);
 119
 120         if (percent_full >= w->dev_most_full_percent) {
 121                 w->dev_most_full_idx            = idx;
 122                 w->dev_most_full_percent        = percent_full;
 123                 w->dev_most_full_work           = work;
 124                 w->dev_most_full_capacity       = capacity;
 125         }
 126
 127         if (w->total_work + dev_work >= w->total_work &&
 128             w->total_work + dev_work >= dev_work)
 129                 w->total_work += dev_work;
 130 }
 131
 132 static struct rebalance_work rebalance_work(struct bch_fs *c)
 133 {
 134         struct bch_dev *ca;
 135         struct rebalance_work ret = { .dev_most_full_idx = -1 };
 136         u64 unknown_dev = atomic64_read(&c->rebalance.work_unknown_dev);
 137         unsigned i;
 138
 139         for_each_online_member(ca, c, i)
 140                 rebalance_work_accumulate(&ret,
 141                         atomic64_read(&ca->rebalance_work),
 142                         unknown_dev,
 143                         bucket_to_sector(ca, ca->mi.nbuckets -
 144                                          ca->mi.first_bucket),
 145                         i);
 146
 147         rebalance_work_accumulate(&ret,
 148                 unknown_dev, 0, c->capacity, -1);
 149
 150         return ret;
 151 }
 152
 153 static void rebalance_work_reset(struct bch_fs *c)
 154 {
 155         struct bch_dev *ca;
 156         unsigned i;
 157
 158         for_each_online_member(ca, c, i)
 159                 atomic64_set(&ca->rebalance_work, 0);
 160
 161         atomic64_set(&c->rebalance.work_unknown_dev, 0);
 162 }
 163
 164 static unsigned long curr_cputime(void)
 165 {
 166         u64 utime, stime;
 167
 168         task_cputime_adjusted(current, &utime, &stime);
 169         return nsecs_to_jiffies(utime + stime);
 170 }
 171
 172 static int bch2_rebalance_thread(void *arg)
 173 {
 174         struct bch_fs *c = arg;
 175         struct bch_fs_rebalance *r = &c->rebalance;
 176         struct io_clock *clock = &c->io_clock[WRITE];
 177         struct rebalance_work w, p;
 178         struct bch_move_stats move_stats;
 179         unsigned long start, prev_start;
 180         unsigned long prev_run_time, prev_run_cputime;
 181         unsigned long cputime, prev_cputime;
 182         u64 io_start;
 183         long throttle;
 184
 185         set_freezable();
 186
 187         io_start        = atomic64_read(&clock->now);
 188         p               = rebalance_work(c);
 189         prev_start      = jiffies;
 190         prev_cputime    = curr_cputime();
 191
 192         bch2_move_stats_init(&move_stats, "rebalance");
 193         while (!kthread_wait_freezable(r->enabled)) {
 194                 cond_resched();
 195
 196                 start                   = jiffies;
 197                 cputime                 = curr_cputime();
 198
 199                 prev_run_time           = start - prev_start;
 200                 prev_run_cputime        = cputime - prev_cputime;
 201
 202                 w                       = rebalance_work(c);
 203                 BUG_ON(!w.dev_most_full_capacity);
 204
 205                 if (!w.total_work) {
 206                         r->state = REBALANCE_WAITING;
 207                         kthread_wait_freezable(rebalance_work(c).total_work);
 208                         continue;
 209                 }
 210
 211                 /*
 212                  * If there isn't much work to do, throttle cpu usage:
 213                  */
 214                 throttle = prev_run_cputime * 100 /
 215                         max(1U, w.dev_most_full_percent) -
 216                         prev_run_time;
 217
 218                 if (w.dev_most_full_percent < 20 && throttle > 0) {
 219                         r->throttled_until_iotime = io_start +
 220                                 div_u64(w.dev_most_full_capacity *
 221                                         (20 - w.dev_most_full_percent),
 222                                         50);
 223
 224                         if (atomic64_read(&clock->now) + clock->max_slop <
 225                             r->throttled_until_iotime) {
 226                                 r->throttled_until_cputime = start + throttle;
 227                                 r->state = REBALANCE_THROTTLED;
 228
 229                                 bch2_kthread_io_clock_wait(clock,
 230                                         r->throttled_until_iotime,
 231                                         throttle);
 232                                 continue;
 233                         }
 234                 }
 235
 236                 /* minimum 1 mb/sec: */
 237                 r->pd.rate.rate =
 238                         max_t(u64, 1 << 11,
 239                               r->pd.rate.rate *
 240                               max(p.dev_most_full_percent, 1U) /
 241                               max(w.dev_most_full_percent, 1U));
 242
 243                 io_start        = atomic64_read(&clock->now);
 244                 p               = w;
 245                 prev_start      = start;
 246                 prev_cputime    = cputime;
 247
 248                 r->state = REBALANCE_RUNNING;
 249                 memset(&move_stats, 0, sizeof(move_stats));
 250                 rebalance_work_reset(c);
 251
 252                 bch2_move_data(c,
 253                                0,               POS_MIN,
 254                                BTREE_ID_NR,     POS_MAX,
 255                                /* ratelimiting disabled for now */
 256                                NULL, /*  &r->pd.rate, */
 257                                &move_stats,
 258                                writepoint_ptr(&c->rebalance_write_point),
 259                                true,
 260                                rebalance_pred, NULL);
 261         }
 262
 263         return 0;
 264 }
 265
 266 void bch2_rebalance_work_to_text(struct printbuf *out, struct bch_fs *c)
 267 {
 268         struct bch_fs_rebalance *r = &c->rebalance;
 269         struct rebalance_work w = rebalance_work(c);
 270
 271         if (!out->nr_tabstops)
 272                 printbuf_tabstop_push(out, 20);
 273
 274         prt_printf(out, "fullest_dev (%i):", w.dev_most_full_idx);
 275         prt_tab(out);
 276
 277         prt_human_readable_u64(out, w.dev_most_full_work << 9);
 278         prt_printf(out, "/");
 279         prt_human_readable_u64(out, w.dev_most_full_capacity << 9);
 280         prt_newline(out);
 281
 282         prt_printf(out, "total work:");
 283         prt_tab(out);
 284
 285         prt_human_readable_u64(out, w.total_work << 9);
 286         prt_printf(out, "/");
 287         prt_human_readable_u64(out, c->capacity << 9);
 288         prt_newline(out);
 289
 290         prt_printf(out, "rate:");
 291         prt_tab(out);
 292         prt_printf(out, "%u", r->pd.rate.rate);
 293         prt_newline(out);
 294
 295         switch (r->state) {
 296         case REBALANCE_WAITING:
 297                 prt_printf(out, "waiting");
 298                 break;
 299         case REBALANCE_THROTTLED:
 300                 prt_printf(out, "throttled for %lu sec or ",
 301                        (r->throttled_until_cputime - jiffies) / HZ);
 302                 prt_human_readable_u64(out,
 303                             (r->throttled_until_iotime -
 304                              atomic64_read(&c->io_clock[WRITE].now)) << 9);
 305                 prt_printf(out, " io");
 306                 break;
 307         case REBALANCE_RUNNING:
 308                 prt_printf(out, "running");
 309                 break;
 310         }
 311         prt_newline(out);
 312 }
 313
 314 void bch2_rebalance_stop(struct bch_fs *c)
 315 {
 316         struct task_struct *p;
 317
 318         c->rebalance.pd.rate.rate = UINT_MAX;
 319         bch2_ratelimit_reset(&c->rebalance.pd.rate);
 320
 321         p = rcu_dereference_protected(c->rebalance.thread, 1);
 322         c->rebalance.thread = NULL;
 323
 324         if (p) {
 325                 /* for sychronizing with rebalance_wakeup() */
 326                 synchronize_rcu();
 327
 328                 kthread_stop(p);
 329                 put_task_struct(p);
 330         }
 331 }
 332
 333 int bch2_rebalance_start(struct bch_fs *c)
 334 {
 335         struct task_struct *p;
 336         int ret;
 337
 338         if (c->rebalance.thread)
 339                 return 0;
 340
 341         if (c->opts.nochanges)
 342                 return 0;
 343
 344         p = kthread_create(bch2_rebalance_thread, c, "bch-rebalance/%s", c->name);
 345         ret = PTR_ERR_OR_ZERO(p);
 346         if (ret) {
 347                 bch_err(c, "error creating rebalance thread: %s", bch2_err_str(ret));
 348                 return ret;
 349         }
 350
 351         get_task_struct(p);
 352         rcu_assign_pointer(c->rebalance.thread, p);
 353         wake_up_process(p);
 354         return 0;
 355 }
 356
 357 void bch2_fs_rebalance_init(struct bch_fs *c)
 358 {
 359         bch2_pd_controller_init(&c->rebalance.pd);
 360
 361         atomic64_set(&c->rebalance.work_unknown_dev, S64_MAX);
 362 }