linux/fs/bcachefs/journal.h
Kent Overstreet 916abefd43 bcachefs: better journal pipelining
Recently a severe performance regression was discovered, which bisected
to

  a6548c8b5e bcachefs: Avoid flushing the journal in the discard path

It turns out the old behaviour, which issued excessive journal flushes,
worked around a performance issue where queueing delays would cause the
journal to not be able to write quickly enough and stall.

The journal flushes masked the issue because they periodically flushed
the device write cache, reducing write latency for non flushes.

This patch reworks the journalling code to allow more than one
(non-flush) write to be in flight at a time. With this patch, doing 4k
random writes and an iodepth of 128, we are now able to hit 560k iops to
a Samsung 970 EVO Plus - previously, we were stuck in the ~200k range.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2024-03-10 15:34:08 -04:00

450 lines
14 KiB
C

/* SPDX-License-Identifier: GPL-2.0 */
#ifndef _BCACHEFS_JOURNAL_H
#define _BCACHEFS_JOURNAL_H
/*
* THE JOURNAL:
*
* The primary purpose of the journal is to log updates (insertions) to the
* b-tree, to avoid having to do synchronous updates to the b-tree on disk.
*
* Without the journal, the b-tree is always internally consistent on
* disk - and in fact, in the earliest incarnations bcache didn't have a journal
* but did handle unclean shutdowns by doing all index updates synchronously
* (with coalescing).
*
* Updates to interior nodes still happen synchronously and without the journal
* (for simplicity) - this may change eventually but updates to interior nodes
* are rare enough it's not a huge priority.
*
* This means the journal is relatively separate from the b-tree; it consists of
* just a list of keys and journal replay consists of just redoing those
* insertions in same order that they appear in the journal.
*
* PERSISTENCE:
*
* For synchronous updates (where we're waiting on the index update to hit
* disk), the journal entry will be written out immediately (or as soon as
* possible, if the write for the previous journal entry was still in flight).
*
* Synchronous updates are specified by passing a closure (@flush_cl) to
* bch2_btree_insert() or bch_btree_insert_node(), which then pass that parameter
* down to the journalling code. That closure will wait on the journal write to
* complete (via closure_wait()).
*
* If the index update wasn't synchronous, the journal entry will be
* written out after 10 ms have elapsed, by default (the delay_ms field
* in struct journal).
*
* JOURNAL ENTRIES:
*
* A journal entry is variable size (struct jset), it's got a fixed length
* header and then a variable number of struct jset_entry entries.
*
* Journal entries are identified by monotonically increasing 64 bit sequence
* numbers - jset->seq; other places in the code refer to this sequence number.
*
* A jset_entry entry contains one or more bkeys (which is what gets inserted
* into the b-tree). We need a container to indicate which b-tree the key is
* for; also, the roots of the various b-trees are stored in jset_entry entries
* (one for each b-tree) - this lets us add new b-tree types without changing
* the on disk format.
*
* We also keep some things in the journal header that are logically part of the
* superblock - all the things that are frequently updated. This is for future
* bcache on raw flash support; the superblock (which will become another
* journal) can't be moved or wear leveled, so it contains just enough
* information to find the main journal, and the superblock only has to be
* rewritten when we want to move/wear level the main journal.
*
* JOURNAL LAYOUT ON DISK:
*
* The journal is written to a ringbuffer of buckets (which is kept in the
* superblock); the individual buckets are not necessarily contiguous on disk
* which means that journal entries are not allowed to span buckets, but also
* that we can resize the journal at runtime if desired (unimplemented).
*
* The journal buckets exist in the same pool as all the other buckets that are
* managed by the allocator and garbage collection - garbage collection marks
* the journal buckets as metadata buckets.
*
* OPEN/DIRTY JOURNAL ENTRIES:
*
* Open/dirty journal entries are journal entries that contain b-tree updates
* that have not yet been written out to the b-tree on disk. We have to track
* which journal entries are dirty, and we also have to avoid wrapping around
* the journal and overwriting old but still dirty journal entries with new
* journal entries.
*
* On disk, this is represented with the "last_seq" field of struct jset;
* last_seq is the first sequence number that journal replay has to replay.
*
* To avoid overwriting dirty journal entries on disk, we keep a mapping (in
* journal_device->seq) of for each journal bucket, the highest sequence number
* any journal entry it contains. Then, by comparing that against last_seq we
* can determine whether that journal bucket contains dirty journal entries or
* not.
*
* To track which journal entries are dirty, we maintain a fifo of refcounts
* (where each entry corresponds to a specific sequence number) - when a ref
* goes to 0, that journal entry is no longer dirty.
*
* Journalling of index updates is done at the same time as the b-tree itself is
* being modified (see btree_insert_key()); when we add the key to the journal
* the pending b-tree write takes a ref on the journal entry the key was added
* to. If a pending b-tree write would need to take refs on multiple dirty
* journal entries, it only keeps the ref on the oldest one (since a newer
* journal entry will still be replayed if an older entry was dirty).
*
* JOURNAL FILLING UP:
*
* There are two ways the journal could fill up; either we could run out of
* space to write to, or we could have too many open journal entries and run out
* of room in the fifo of refcounts. Since those refcounts are decremented
* without any locking we can't safely resize that fifo, so we handle it the
* same way.
*
* If the journal fills up, we start flushing dirty btree nodes until we can
* allocate space for a journal write again - preferentially flushing btree
* nodes that are pinning the oldest journal entries first.
*/
#include <linux/hash.h>
#include "journal_types.h"
struct bch_fs;
static inline void journal_wake(struct journal *j)
{
wake_up(&j->wait);
closure_wake_up(&j->async_wait);
}
static inline struct journal_buf *journal_cur_buf(struct journal *j)
{
return j->buf + j->reservations.idx;
}
/* Sequence number of oldest dirty journal entry */
static inline u64 journal_last_seq(struct journal *j)
{
return j->pin.front;
}
static inline u64 journal_cur_seq(struct journal *j)
{
return atomic64_read(&j->seq);
}
static inline u64 journal_last_unwritten_seq(struct journal *j)
{
return j->seq_ondisk + 1;
}
static inline int journal_state_count(union journal_res_state s, int idx)
{
switch (idx) {
case 0: return s.buf0_count;
case 1: return s.buf1_count;
case 2: return s.buf2_count;
case 3: return s.buf3_count;
}
BUG();
}
static inline void journal_state_inc(union journal_res_state *s)
{
s->buf0_count += s->idx == 0;
s->buf1_count += s->idx == 1;
s->buf2_count += s->idx == 2;
s->buf3_count += s->idx == 3;
}
/*
* Amount of space that will be taken up by some keys in the journal (i.e.
* including the jset header)
*/
static inline unsigned jset_u64s(unsigned u64s)
{
return u64s + sizeof(struct jset_entry) / sizeof(u64);
}
static inline int journal_entry_overhead(struct journal *j)
{
return sizeof(struct jset) / sizeof(u64) + j->entry_u64s_reserved;
}
static inline struct jset_entry *
bch2_journal_add_entry_noreservation(struct journal_buf *buf, size_t u64s)
{
struct jset *jset = buf->data;
struct jset_entry *entry = vstruct_idx(jset, le32_to_cpu(jset->u64s));
memset(entry, 0, sizeof(*entry));
entry->u64s = cpu_to_le16(u64s);
le32_add_cpu(&jset->u64s, jset_u64s(u64s));
return entry;
}
static inline struct jset_entry *
journal_res_entry(struct journal *j, struct journal_res *res)
{
return vstruct_idx(j->buf[res->idx].data, res->offset);
}
static inline unsigned journal_entry_init(struct jset_entry *entry, unsigned type,
enum btree_id id, unsigned level,
unsigned u64s)
{
entry->u64s = cpu_to_le16(u64s);
entry->btree_id = id;
entry->level = level;
entry->type = type;
entry->pad[0] = 0;
entry->pad[1] = 0;
entry->pad[2] = 0;
return jset_u64s(u64s);
}
static inline unsigned journal_entry_set(struct jset_entry *entry, unsigned type,
enum btree_id id, unsigned level,
const void *data, unsigned u64s)
{
unsigned ret = journal_entry_init(entry, type, id, level, u64s);
memcpy_u64s_small(entry->_data, data, u64s);
return ret;
}
static inline struct jset_entry *
bch2_journal_add_entry(struct journal *j, struct journal_res *res,
unsigned type, enum btree_id id,
unsigned level, unsigned u64s)
{
struct jset_entry *entry = journal_res_entry(j, res);
unsigned actual = journal_entry_init(entry, type, id, level, u64s);
EBUG_ON(!res->ref);
EBUG_ON(actual > res->u64s);
res->offset += actual;
res->u64s -= actual;
return entry;
}
static inline bool journal_entry_empty(struct jset *j)
{
if (j->seq != j->last_seq)
return false;
vstruct_for_each(j, i)
if (i->type == BCH_JSET_ENTRY_btree_keys && i->u64s)
return false;
return true;
}
/*
* Drop reference on a buffer index and return true if the count has hit zero.
*/
static inline union journal_res_state journal_state_buf_put(struct journal *j, unsigned idx)
{
union journal_res_state s;
s.v = atomic64_sub_return(((union journal_res_state) {
.buf0_count = idx == 0,
.buf1_count = idx == 1,
.buf2_count = idx == 2,
.buf3_count = idx == 3,
}).v, &j->reservations.counter);
return s;
}
bool bch2_journal_entry_close(struct journal *);
void bch2_journal_do_writes(struct journal *);
void bch2_journal_buf_put_final(struct journal *, u64);
static inline void __bch2_journal_buf_put(struct journal *j, unsigned idx, u64 seq)
{
union journal_res_state s;
s = journal_state_buf_put(j, idx);
if (!journal_state_count(s, idx))
bch2_journal_buf_put_final(j, seq);
}
static inline void bch2_journal_buf_put(struct journal *j, unsigned idx, u64 seq)
{
union journal_res_state s;
s = journal_state_buf_put(j, idx);
if (!journal_state_count(s, idx)) {
spin_lock(&j->lock);
bch2_journal_buf_put_final(j, seq);
spin_unlock(&j->lock);
}
}
/*
* This function releases the journal write structure so other threads can
* then proceed to add their keys as well.
*/
static inline void bch2_journal_res_put(struct journal *j,
struct journal_res *res)
{
if (!res->ref)
return;
lock_release(&j->res_map, _THIS_IP_);
while (res->u64s)
bch2_journal_add_entry(j, res,
BCH_JSET_ENTRY_btree_keys,
0, 0, 0);
bch2_journal_buf_put(j, res->idx, res->seq);
res->ref = 0;
}
int bch2_journal_res_get_slowpath(struct journal *, struct journal_res *,
unsigned);
/* First bits for BCH_WATERMARK: */
enum journal_res_flags {
__JOURNAL_RES_GET_NONBLOCK = BCH_WATERMARK_BITS,
__JOURNAL_RES_GET_CHECK,
};
#define JOURNAL_RES_GET_NONBLOCK (1 << __JOURNAL_RES_GET_NONBLOCK)
#define JOURNAL_RES_GET_CHECK (1 << __JOURNAL_RES_GET_CHECK)
static inline int journal_res_get_fast(struct journal *j,
struct journal_res *res,
unsigned flags)
{
union journal_res_state old, new;
u64 v = atomic64_read(&j->reservations.counter);
do {
old.v = new.v = v;
/*
* Check if there is still room in the current journal
* entry:
*/
if (new.cur_entry_offset + res->u64s > j->cur_entry_u64s)
return 0;
EBUG_ON(!journal_state_count(new, new.idx));
if ((flags & BCH_WATERMARK_MASK) < j->watermark)
return 0;
new.cur_entry_offset += res->u64s;
journal_state_inc(&new);
/*
* If the refcount would overflow, we have to wait:
* XXX - tracepoint this:
*/
if (!journal_state_count(new, new.idx))
return 0;
if (flags & JOURNAL_RES_GET_CHECK)
return 1;
} while ((v = atomic64_cmpxchg(&j->reservations.counter,
old.v, new.v)) != old.v);
res->ref = true;
res->idx = old.idx;
res->offset = old.cur_entry_offset;
res->seq = le64_to_cpu(j->buf[old.idx].data->seq);
return 1;
}
static inline int bch2_journal_res_get(struct journal *j, struct journal_res *res,
unsigned u64s, unsigned flags)
{
int ret;
EBUG_ON(res->ref);
EBUG_ON(!test_bit(JOURNAL_STARTED, &j->flags));
res->u64s = u64s;
if (journal_res_get_fast(j, res, flags))
goto out;
ret = bch2_journal_res_get_slowpath(j, res, flags);
if (ret)
return ret;
out:
if (!(flags & JOURNAL_RES_GET_CHECK)) {
lock_acquire_shared(&j->res_map, 0,
(flags & JOURNAL_RES_GET_NONBLOCK) != 0,
NULL, _THIS_IP_);
EBUG_ON(!res->ref);
}
return 0;
}
/* journal_entry_res: */
void bch2_journal_entry_res_resize(struct journal *,
struct journal_entry_res *,
unsigned);
int bch2_journal_flush_seq_async(struct journal *, u64, struct closure *);
void bch2_journal_flush_async(struct journal *, struct closure *);
int bch2_journal_flush_seq(struct journal *, u64);
int bch2_journal_flush(struct journal *);
bool bch2_journal_noflush_seq(struct journal *, u64);
int bch2_journal_meta(struct journal *);
void bch2_journal_halt(struct journal *);
static inline int bch2_journal_error(struct journal *j)
{
return j->reservations.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL
? -EIO : 0;
}
struct bch_dev;
static inline void bch2_journal_set_replay_done(struct journal *j)
{
BUG_ON(!test_bit(JOURNAL_STARTED, &j->flags));
set_bit(JOURNAL_REPLAY_DONE, &j->flags);
}
void bch2_journal_unblock(struct journal *);
void bch2_journal_block(struct journal *);
struct journal_buf *bch2_next_write_buffer_flush_journal_buf(struct journal *j, u64 max_seq);
void __bch2_journal_debug_to_text(struct printbuf *, struct journal *);
void bch2_journal_debug_to_text(struct printbuf *, struct journal *);
void bch2_journal_pins_to_text(struct printbuf *, struct journal *);
bool bch2_journal_seq_pins_to_text(struct printbuf *, struct journal *, u64 *);
int bch2_set_nr_journal_buckets(struct bch_fs *, struct bch_dev *,
unsigned nr);
int bch2_dev_journal_alloc(struct bch_dev *);
int bch2_fs_journal_alloc(struct bch_fs *);
void bch2_dev_journal_stop(struct journal *, struct bch_dev *);
void bch2_fs_journal_stop(struct journal *);
int bch2_fs_journal_start(struct journal *, u64);
void bch2_dev_journal_exit(struct bch_dev *);
int bch2_dev_journal_init(struct bch_dev *, struct bch_sb *);
void bch2_fs_journal_exit(struct journal *);
int bch2_fs_journal_init(struct journal *);
#endif /* _BCACHEFS_JOURNAL_H */