git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/journal_types.h

   1 #ifndef _BCACHE_JOURNAL_TYPES_H
   2 #define _BCACHE_JOURNAL_TYPES_H
   3
   4 #include <linux/cache.h>
   5 #include <linux/workqueue.h>
   6
   7 #include "alloc_types.h"
   8 #include "fifo.h"
   9
  10 struct journal_res;
  11
  12 /*
  13  * We put two of these in struct journal; we used them for writes to the
  14  * journal that are being staged or in flight.
  15  */
  16 struct journal_buf {
  17         struct jset             *data;
  18
  19         struct closure_waitlist wait;
  20
  21         unsigned                size;
  22         unsigned                disk_sectors;
  23
  24         /*
  25          * ugh, prio_buckets are stupid - need to convert them to new
  26          * transaction machinery when it arrives
  27          */
  28         unsigned                nr_prio_buckets;
  29
  30         /* bloom filter: */
  31         unsigned long           has_inode[1024 / sizeof(unsigned long)];
  32 };
  33
  34 /*
  35  * Something that makes a journal entry dirty - i.e. a btree node that has to be
  36  * flushed:
  37  */
  38
  39 struct journal_entry_pin_list {
  40         struct list_head                list;
  41         atomic_t                        count;
  42 };
  43
  44 struct journal;
  45 struct journal_entry_pin;
  46 typedef void (*journal_pin_flush_fn)(struct journal *j,
  47                                 struct journal_entry_pin *, u64);
  48
  49 struct journal_entry_pin {
  50         struct list_head                list;
  51         journal_pin_flush_fn            flush;
  52         struct journal_entry_pin_list   *pin_list;
  53 };
  54
  55 /* corresponds to a btree node with a blacklisted bset: */
  56 struct blacklisted_node {
  57         __le64                  seq;
  58         enum btree_id           btree_id;
  59         struct bpos             pos;
  60 };
  61
  62 struct journal_seq_blacklist {
  63         struct list_head        list;
  64         u64                     seq;
  65         bool                    written;
  66         struct journal_entry_pin pin;
  67
  68         struct blacklisted_node *entries;
  69         size_t                  nr_entries;
  70 };
  71
  72 struct journal_res {
  73         bool                    ref;
  74         u8                      idx;
  75         u16                     u64s;
  76         u32                     offset;
  77         u64                     seq;
  78 };
  79
  80 union journal_res_state {
  81         struct {
  82                 atomic64_t      counter;
  83         };
  84
  85         struct {
  86                 u64             v;
  87         };
  88
  89         struct {
  90                 u64             cur_entry_offset:20,
  91                                 idx:1,
  92                                 prev_buf_unwritten:1,
  93                                 buf0_count:21,
  94                                 buf1_count:21;
  95         };
  96 };
  97
  98 /* bytes: */
  99 #define JOURNAL_ENTRY_SIZE_MIN          (64U << 10) /* 64k */
 100 #define JOURNAL_ENTRY_SIZE_MAX          (4U  << 20) /* 4M */
 101
 102 /*
 103  * We stash some journal state as sentinal values in cur_entry_offset:
 104  * note - cur_entry_offset is in units of u64s
 105  */
 106 #define JOURNAL_ENTRY_OFFSET_MAX        ((1U << 20) - 1)
 107
 108 #define JOURNAL_ENTRY_CLOSED_VAL        (JOURNAL_ENTRY_OFFSET_MAX - 1)
 109 #define JOURNAL_ENTRY_ERROR_VAL         (JOURNAL_ENTRY_OFFSET_MAX)
 110
 111 /*
 112  * JOURNAL_NEED_WRITE - current (pending) journal entry should be written ASAP,
 113  * either because something's waiting on the write to complete or because it's
 114  * been dirty too long and the timer's expired.
 115  */
 116
 117 enum {
 118         JOURNAL_REPLAY_DONE,
 119         JOURNAL_STARTED,
 120         JOURNAL_NEED_WRITE,
 121 };
 122
 123 /* Embedded in struct bch_fs */
 124 struct journal {
 125         /* Fastpath stuff up front: */
 126
 127         unsigned long           flags;
 128
 129         union journal_res_state reservations;
 130         unsigned                cur_entry_u64s;
 131         unsigned                prev_buf_sectors;
 132         unsigned                cur_buf_sectors;
 133         unsigned                buf_size_want;
 134
 135         /*
 136          * Two journal entries -- one is currently open for new entries, the
 137          * other is possibly being written out.
 138          */
 139         struct journal_buf      buf[2];
 140
 141         spinlock_t              lock;
 142
 143         /* Used when waiting because the journal was full */
 144         wait_queue_head_t       wait;
 145
 146         struct closure          io;
 147         struct delayed_work     write_work;
 148
 149         /* Sequence number of most recent journal entry (last entry in @pin) */
 150         atomic64_t              seq;
 151
 152         /* last_seq from the most recent journal entry written */
 153         u64                     last_seq_ondisk;
 154
 155         /*
 156          * FIFO of journal entries whose btree updates have not yet been
 157          * written out.
 158          *
 159          * Each entry is a reference count. The position in the FIFO is the
 160          * entry's sequence number relative to @seq.
 161          *
 162          * The journal entry itself holds a reference count, put when the
 163          * journal entry is written out. Each btree node modified by the journal
 164          * entry also holds a reference count, put when the btree node is
 165          * written.
 166          *
 167          * When a reference count reaches zero, the journal entry is no longer
 168          * needed. When all journal entries in the oldest journal bucket are no
 169          * longer needed, the bucket can be discarded and reused.
 170          */
 171         DECLARE_FIFO(struct journal_entry_pin_list, pin);
 172         struct journal_entry_pin_list *replay_pin_list;
 173
 174         /*
 175          * Protects the pin lists - the fifo itself is still protected by
 176          * j->lock though:
 177          */
 178         spinlock_t              pin_lock;
 179
 180         struct mutex            blacklist_lock;
 181         struct list_head        seq_blacklist;
 182
 183         BKEY_PADDED(key);
 184         struct dev_group        devs;
 185
 186         struct delayed_work     reclaim_work;
 187         unsigned long           last_flushed;
 188
 189         /* protects advancing ja->last_idx: */
 190         struct mutex            reclaim_lock;
 191
 192         /*
 193          * ugh: need to get prio_buckets converted over to the eventual new
 194          * transaction machinery
 195          */
 196         __le64                  prio_buckets[BCH_SB_MEMBERS_MAX];
 197         unsigned                nr_prio_buckets;
 198
 199         unsigned                write_delay_ms;
 200         unsigned                reclaim_delay_ms;
 201
 202         u64                     res_get_blocked_start;
 203         u64                     need_write_time;
 204         u64                     write_start_time;
 205
 206         struct time_stats       *write_time;
 207         struct time_stats       *delay_time;
 208         struct time_stats       *blocked_time;
 209         struct time_stats       *flush_seq_time;
 210
 211 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 212         struct lockdep_map      res_map;
 213 #endif
 214 };
 215
 216 /*
 217  * Embedded in struct bch_dev. First three fields refer to the array of journal
 218  * buckets, in bch_sb.
 219  */
 220 struct journal_device {
 221         /*
 222          * For each journal bucket, contains the max sequence number of the
 223          * journal writes it contains - so we know when a bucket can be reused.
 224          */
 225         u64                     *bucket_seq;
 226
 227         unsigned                sectors_free;
 228
 229         /* Journal bucket we're currently writing to */
 230         unsigned                cur_idx;
 231
 232         /* Last journal bucket that still contains an open journal entry */
 233
 234         /*
 235          * j->lock and j->reclaim_lock must both be held to modify, j->lock
 236          * sufficient to read:
 237          */
 238         unsigned                last_idx;
 239         unsigned                nr;
 240         u64                     *buckets;
 241
 242         /* Bio for journal reads/writes to this device */
 243         struct bio              *bio;
 244
 245         /* for bch_journal_read_device */
 246         struct closure          read;
 247 };
 248
 249 #endif /* _BCACHE_JOURNAL_TYPES_H */