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

   1
   2 #include "bcachefs.h"
   3 #include "alloc_foreground.h"
   4 #include "btree_iter.h"
   5 #include "buckets.h"
   6 #include "clock.h"
   7 #include "disk_groups.h"
   8 #include "extents.h"
   9 #include "io.h"
  10 #include "move.h"
  11 #include "rebalance.h"
  12 #include "super-io.h"
  13
  14 #include <linux/freezer.h>
  15 #include <linux/kthread.h>
  16 #include <linux/sched/cputime.h>
  17 #include <trace/events/bcachefs.h>
  18
  19 static inline bool rebalance_ptr_pred(struct bch_fs *c,
  20                                       struct extent_ptr_decoded p,
  21                                       struct bch_io_opts *io_opts)
  22 {
  23         if (io_opts->background_target &&
  24             !bch2_dev_in_target(c, p.ptr.dev, io_opts->background_target) &&
  25             !p.ptr.cached)
  26                 return true;
  27
  28         if (io_opts->background_compression &&
  29             p.crc.compression_type !=
  30             bch2_compression_opt_to_type[io_opts->background_compression])
  31                 return true;
  32
  33         return false;
  34 }
  35
  36 void bch2_rebalance_add_key(struct bch_fs *c,
  37                             struct bkey_s_c k,
  38                             struct bch_io_opts *io_opts)
  39 {
  40         const union bch_extent_entry *entry;
  41         struct extent_ptr_decoded p;
  42         struct bkey_s_c_extent e;
  43
  44         if (!bkey_extent_is_data(k.k))
  45                 return;
  46
  47         if (!io_opts->background_target &&
  48             !io_opts->background_compression)
  49                 return;
  50
  51         e = bkey_s_c_to_extent(k);
  52
  53         extent_for_each_ptr_decode(e, p, entry)
  54                 if (rebalance_ptr_pred(c, p, io_opts)) {
  55                         struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
  56
  57                         if (atomic64_add_return(p.crc.compressed_size,
  58                                                 &ca->rebalance_work) ==
  59                             p.crc.compressed_size)
  60                                 rebalance_wakeup(c);
  61                 }
  62 }
  63
  64 void bch2_rebalance_add_work(struct bch_fs *c, u64 sectors)
  65 {
  66         if (atomic64_add_return(sectors, &c->rebalance.work_unknown_dev) ==
  67             sectors)
  68                 rebalance_wakeup(c);
  69 }
  70
  71 static enum data_cmd rebalance_pred(struct bch_fs *c, void *arg,
  72                                     enum bkey_type type,
  73                                     struct bkey_s_c_extent e,
  74                                     struct bch_io_opts *io_opts,
  75                                     struct data_opts *data_opts)
  76 {
  77         const union bch_extent_entry *entry;
  78         struct extent_ptr_decoded p;
  79
  80         /* Make sure we have room to add a new pointer: */
  81         if (bkey_val_u64s(e.k) + BKEY_EXTENT_PTR_U64s_MAX >
  82             BKEY_EXTENT_VAL_U64s_MAX)
  83                 return DATA_SKIP;
  84
  85         extent_for_each_ptr_decode(e, p, entry)
  86                 if (rebalance_ptr_pred(c, p, io_opts))
  87                         goto found;
  88
  89         return DATA_SKIP;
  90 found:
  91         data_opts->target               = io_opts->background_target;
  92         data_opts->btree_insert_flags   = 0;
  93         return DATA_ADD_REPLICAS;
  94 }
  95
  96 struct rebalance_work {
  97         int             dev_most_full_idx;
  98         unsigned        dev_most_full_percent;
  99         u64             dev_most_full_work;
 100         u64             dev_most_full_capacity;
 101         u64             total_work;
 102 };
 103
 104 static void rebalance_work_accumulate(struct rebalance_work *w,
 105                 u64 dev_work, u64 unknown_dev, u64 capacity, int idx)
 106 {
 107         unsigned percent_full;
 108         u64 work = dev_work + unknown_dev;
 109
 110         if (work < dev_work || work < unknown_dev)
 111                 work = U64_MAX;
 112         work = min(work, capacity);
 113
 114         percent_full = div64_u64(work * 100, capacity);
 115
 116         if (percent_full >= w->dev_most_full_percent) {
 117                 w->dev_most_full_idx            = idx;
 118                 w->dev_most_full_percent        = percent_full;
 119                 w->dev_most_full_work           = work;
 120                 w->dev_most_full_capacity       = capacity;
 121         }
 122
 123         if (w->total_work + dev_work >= w->total_work &&
 124             w->total_work + dev_work >= dev_work)
 125                 w->total_work += dev_work;
 126 }
 127
 128 static struct rebalance_work rebalance_work(struct bch_fs *c)
 129 {
 130         struct bch_dev *ca;
 131         struct rebalance_work ret = { .dev_most_full_idx = -1 };
 132         u64 unknown_dev = atomic64_read(&c->rebalance.work_unknown_dev);
 133         unsigned i;
 134
 135         for_each_online_member(ca, c, i)
 136                 rebalance_work_accumulate(&ret,
 137                         atomic64_read(&ca->rebalance_work),
 138                         unknown_dev,
 139                         bucket_to_sector(ca, ca->mi.nbuckets -
 140                                          ca->mi.first_bucket),
 141                         i);
 142
 143         rebalance_work_accumulate(&ret,
 144                 unknown_dev, 0, c->capacity, -1);
 145
 146         return ret;
 147 }
 148
 149 static void rebalance_work_reset(struct bch_fs *c)
 150 {
 151         struct bch_dev *ca;
 152         unsigned i;
 153
 154         for_each_online_member(ca, c, i)
 155                 atomic64_set(&ca->rebalance_work, 0);
 156
 157         atomic64_set(&c->rebalance.work_unknown_dev, 0);
 158 }
 159
 160 static unsigned long curr_cputime(void)
 161 {
 162         u64 utime, stime;
 163
 164         task_cputime_adjusted(current, &utime, &stime);
 165         return nsecs_to_jiffies(utime + stime);
 166 }
 167
 168 static int bch2_rebalance_thread(void *arg)
 169 {
 170         struct bch_fs *c = arg;
 171         struct bch_fs_rebalance *r = &c->rebalance;
 172         struct io_clock *clock = &c->io_clock[WRITE];
 173         struct rebalance_work w, p;
 174         unsigned long start, prev_start;
 175         unsigned long prev_run_time, prev_run_cputime;
 176         unsigned long cputime, prev_cputime;
 177         unsigned long io_start;
 178         long throttle;
 179
 180         set_freezable();
 181
 182         io_start        = atomic_long_read(&clock->now);
 183         p               = rebalance_work(c);
 184         prev_start      = jiffies;
 185         prev_cputime    = curr_cputime();
 186
 187         while (!kthread_wait_freezable(r->enabled)) {
 188                 start                   = jiffies;
 189                 cputime                 = curr_cputime();
 190
 191                 prev_run_time           = start - prev_start;
 192                 prev_run_cputime        = cputime - prev_cputime;
 193
 194                 w                       = rebalance_work(c);
 195                 BUG_ON(!w.dev_most_full_capacity);
 196
 197                 if (!w.total_work) {
 198                         r->state = REBALANCE_WAITING;
 199                         kthread_wait_freezable(rebalance_work(c).total_work);
 200                         continue;
 201                 }
 202
 203                 /*
 204                  * If there isn't much work to do, throttle cpu usage:
 205                  */
 206                 throttle = prev_run_cputime * 100 /
 207                         max(1U, w.dev_most_full_percent) -
 208                         prev_run_time;
 209
 210                 if (w.dev_most_full_percent < 20 && throttle > 0) {
 211                         r->state = REBALANCE_THROTTLED;
 212                         r->throttled_until_iotime = io_start +
 213                                 div_u64(w.dev_most_full_capacity *
 214                                         (20 - w.dev_most_full_percent),
 215                                         50);
 216                         r->throttled_until_cputime = start + throttle;
 217
 218                         bch2_kthread_io_clock_wait(clock,
 219                                 r->throttled_until_iotime,
 220                                 throttle);
 221                         continue;
 222                 }
 223
 224                 /* minimum 1 mb/sec: */
 225                 r->pd.rate.rate =
 226                         max_t(u64, 1 << 11,
 227                               r->pd.rate.rate *
 228                               max(p.dev_most_full_percent, 1U) /
 229                               max(w.dev_most_full_percent, 1U));
 230
 231                 io_start        = atomic_long_read(&clock->now);
 232                 p               = w;
 233                 prev_start      = start;
 234                 prev_cputime    = cputime;
 235
 236                 r->state = REBALANCE_RUNNING;
 237                 memset(&r->move_stats, 0, sizeof(r->move_stats));
 238                 rebalance_work_reset(c);
 239
 240                 bch2_move_data(c,
 241                                /* ratelimiting disabled for now */
 242                                NULL, /*  &r->pd.rate, */
 243                                writepoint_ptr(&c->rebalance_write_point),
 244                                POS_MIN, POS_MAX,
 245                                rebalance_pred, NULL,
 246                                &r->move_stats);
 247         }
 248
 249         return 0;
 250 }
 251
 252 ssize_t bch2_rebalance_work_show(struct bch_fs *c, char *buf)
 253 {
 254         char *out = buf, *end = out + PAGE_SIZE;
 255         struct bch_fs_rebalance *r = &c->rebalance;
 256         struct rebalance_work w = rebalance_work(c);
 257         char h1[21], h2[21];
 258
 259         bch2_hprint(h1, w.dev_most_full_work << 9);
 260         bch2_hprint(h2, w.dev_most_full_capacity << 9);
 261         out += scnprintf(out, end - out,
 262                          "fullest_dev (%i):\t%s/%s\n",
 263                          w.dev_most_full_idx, h1, h2);
 264
 265         bch2_hprint(h1, w.total_work << 9);
 266         bch2_hprint(h2, c->capacity << 9);
 267         out += scnprintf(out, end - out,
 268                          "total work:\t\t%s/%s\n",
 269                          h1, h2);
 270
 271         out += scnprintf(out, end - out,
 272                          "rate:\t\t\t%u\n",
 273                          r->pd.rate.rate);
 274
 275         switch (r->state) {
 276         case REBALANCE_WAITING:
 277                 out += scnprintf(out, end - out, "waiting\n");
 278                 break;
 279         case REBALANCE_THROTTLED:
 280                 bch2_hprint(h1,
 281                             (r->throttled_until_iotime -
 282                              atomic_long_read(&c->io_clock[WRITE].now)) << 9);
 283                 out += scnprintf(out, end - out,
 284                                  "throttled for %lu sec or %s io\n",
 285                                  (r->throttled_until_cputime - jiffies) / HZ,
 286                                  h1);
 287                 break;
 288         case REBALANCE_RUNNING:
 289                 out += scnprintf(out, end - out, "running\n");
 290                 out += scnprintf(out, end - out, "pos %llu:%llu\n",
 291                                  r->move_stats.iter.pos.inode,
 292                                  r->move_stats.iter.pos.offset);
 293                 break;
 294         }
 295
 296         return out - buf;
 297 }
 298
 299 void bch2_rebalance_stop(struct bch_fs *c)
 300 {
 301         struct task_struct *p;
 302
 303         c->rebalance.pd.rate.rate = UINT_MAX;
 304         bch2_ratelimit_reset(&c->rebalance.pd.rate);
 305
 306         p = rcu_dereference_protected(c->rebalance.thread, 1);
 307         c->rebalance.thread = NULL;
 308
 309         if (p) {
 310                 /* for sychronizing with rebalance_wakeup() */
 311                 synchronize_rcu();
 312
 313                 kthread_stop(p);
 314                 put_task_struct(p);
 315         }
 316 }
 317
 318 int bch2_rebalance_start(struct bch_fs *c)
 319 {
 320         struct task_struct *p;
 321
 322         if (c->opts.nochanges)
 323                 return 0;
 324
 325         p = kthread_create(bch2_rebalance_thread, c, "bch_rebalance");
 326         if (IS_ERR(p))
 327                 return PTR_ERR(p);
 328
 329         get_task_struct(p);
 330         rcu_assign_pointer(c->rebalance.thread, p);
 331         wake_up_process(p);
 332         return 0;
 333 }
 334
 335 void bch2_fs_rebalance_init(struct bch_fs *c)
 336 {
 337         bch2_pd_controller_init(&c->rebalance.pd);
 338
 339         atomic64_set(&c->rebalance.work_unknown_dev, S64_MAX);
 340 }