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