+// SPDX-License-Identifier: GPL-2.0
/*
* bcachefs journalling code, for btree insertions
*
#include "alloc_foreground.h"
#include "bkey_methods.h"
#include "btree_gc.h"
+#include "btree_update.h"
#include "buckets.h"
+#include "error.h"
#include "journal.h"
#include "journal_io.h"
#include "journal_reclaim.h"
#include <trace/events/bcachefs.h>
+static u64 last_unwritten_seq(struct journal *j)
+{
+ union journal_res_state s = READ_ONCE(j->reservations);
+
+ lockdep_assert_held(&j->lock);
+
+ return journal_cur_seq(j) - ((s.idx - s.unwritten_idx) & JOURNAL_BUF_MASK);
+}
+
+static inline bool journal_seq_unwritten(struct journal *j, u64 seq)
+{
+ return seq >= last_unwritten_seq(j);
+}
+
static bool __journal_entry_is_open(union journal_res_state state)
{
return state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL;
return __journal_entry_is_open(j->reservations);
}
-static void journal_pin_new_entry(struct journal *j, int count)
+static inline struct journal_buf *
+journal_seq_to_buf(struct journal *j, u64 seq)
{
- struct journal_entry_pin_list *p;
+ struct journal_buf *buf = NULL;
- /*
- * The fifo_push() needs to happen at the same time as j->seq is
- * incremented for journal_last_seq() to be calculated correctly
- */
- atomic64_inc(&j->seq);
- p = fifo_push_ref(&j->pin);
+ EBUG_ON(seq > journal_cur_seq(j));
+ EBUG_ON(seq == journal_cur_seq(j) &&
+ j->reservations.cur_entry_offset == JOURNAL_ENTRY_CLOSED_VAL);
+
+ if (journal_seq_unwritten(j, seq)) {
+ buf = j->buf + (seq & JOURNAL_BUF_MASK);
+ EBUG_ON(le64_to_cpu(buf->data->seq) != seq);
+ }
+ return buf;
+}
+static void journal_pin_list_init(struct journal_entry_pin_list *p, int count)
+{
INIT_LIST_HEAD(&p->list);
+ INIT_LIST_HEAD(&p->key_cache_list);
INIT_LIST_HEAD(&p->flushed);
atomic_set(&p->count, count);
p->devs.nr = 0;
}
+static void journal_pin_new_entry(struct journal *j)
+{
+ /*
+ * The fifo_push() needs to happen at the same time as j->seq is
+ * incremented for journal_last_seq() to be calculated correctly
+ */
+ atomic64_inc(&j->seq);
+ journal_pin_list_init(fifo_push_ref(&j->pin), 1);
+}
+
static void bch2_journal_buf_init(struct journal *j)
{
struct journal_buf *buf = journal_cur_buf(j);
- memset(buf->has_inode, 0, sizeof(buf->has_inode));
+ bkey_extent_init(&buf->key);
+ buf->noflush = false;
+ buf->must_flush = false;
+ buf->separate_flush = false;
memset(buf->data, 0, sizeof(*buf->data));
buf->data->seq = cpu_to_le64(journal_cur_seq(j));
} while ((v = atomic64_cmpxchg(&j->reservations.counter,
old.v, new.v)) != old.v);
+ /*
+ * XXX: we're not using j->lock here because this can be called from
+ * interrupt context, this can race with journal_write_done()
+ */
+ if (!j->err_seq)
+ j->err_seq = journal_cur_seq(j);
journal_wake(j);
closure_wake_up(&journal_cur_buf(j)->wait);
}
/* journal entry close/open: */
-void __bch2_journal_buf_put(struct journal *j, bool need_write_just_set)
+void __bch2_journal_buf_put(struct journal *j)
{
- if (!need_write_just_set &&
- test_bit(JOURNAL_NEED_WRITE, &j->flags))
- bch2_time_stats_update(j->delay_time,
- j->need_write_time);
-
- clear_bit(JOURNAL_NEED_WRITE, &j->flags);
+ struct bch_fs *c = container_of(j, struct bch_fs, journal);
- closure_call(&j->io, bch2_journal_write, system_highpri_wq, NULL);
+ closure_call(&j->io, bch2_journal_write, c->io_complete_wq, NULL);
}
/*
* Returns true if journal entry is now closed:
+ *
+ * We don't close a journal_buf until the next journal_buf is finished writing,
+ * and can be opened again - this also initializes the next journal_buf:
*/
static bool __journal_entry_close(struct journal *j)
{
struct journal_buf *buf = journal_cur_buf(j);
union journal_res_state old, new;
u64 v = atomic64_read(&j->reservations.counter);
- bool set_need_write = false;
unsigned sectors;
lockdep_assert_held(&j->lock);
if (!test_bit(JOURNAL_NEED_WRITE, &j->flags)) {
set_bit(JOURNAL_NEED_WRITE, &j->flags);
j->need_write_time = local_clock();
- set_need_write = true;
}
- if (new.prev_buf_unwritten)
- return false;
-
new.cur_entry_offset = JOURNAL_ENTRY_CLOSED_VAL;
new.idx++;
- new.prev_buf_unwritten = 1;
+
+ if (new.idx == new.unwritten_idx)
+ return false;
BUG_ON(journal_state_count(new, new.idx));
} while ((v = atomic64_cmpxchg(&j->reservations.counter,
old.v, new.v)) != old.v);
+ /* Close out old buffer: */
buf->data->u64s = cpu_to_le32(old.cur_entry_offset);
sectors = vstruct_blocks_plus(buf->data, c->block_bits,
BUG_ON(sectors > buf->sectors);
buf->sectors = sectors;
- bkey_extent_init(&buf->key);
-
/*
* We have to set last_seq here, _before_ opening a new journal entry:
*
* Hence, we want update/set last_seq on the current journal entry right
* before we open a new one:
*/
- buf->data->last_seq = cpu_to_le64(journal_last_seq(j));
+ buf->last_seq = journal_last_seq(j);
+ buf->data->last_seq = cpu_to_le64(buf->last_seq);
- if (journal_entry_empty(buf->data))
- clear_bit(JOURNAL_NOT_EMPTY, &j->flags);
- else
- set_bit(JOURNAL_NOT_EMPTY, &j->flags);
+ __bch2_journal_pin_put(j, le64_to_cpu(buf->data->seq));
- journal_pin_new_entry(j, 1);
+ /* Initialize new buffer: */
+ journal_pin_new_entry(j);
bch2_journal_buf_init(j);
cancel_delayed_work(&j->write_work);
+ clear_bit(JOURNAL_NEED_WRITE, &j->flags);
bch2_journal_space_available(j);
- bch2_journal_buf_put(j, old.idx, set_need_write);
+ bch2_journal_buf_put(j, old.idx);
return true;
}
+static bool journal_entry_want_write(struct journal *j)
+{
+ union journal_res_state s = READ_ONCE(j->reservations);
+ bool ret = false;
+
+ /*
+ * Don't close it yet if we already have a write in flight, but do set
+ * NEED_WRITE:
+ */
+ if (s.idx != s.unwritten_idx)
+ set_bit(JOURNAL_NEED_WRITE, &j->flags);
+ else
+ ret = __journal_entry_close(j);
+
+ return ret;
+}
+
static bool journal_entry_close(struct journal *j)
{
bool ret;
spin_lock(&j->lock);
- ret = __journal_entry_close(j);
+ ret = journal_entry_want_write(j);
spin_unlock(&j->lock);
return ret;
*/
static int journal_entry_open(struct journal *j)
{
+ struct bch_fs *c = container_of(j, struct bch_fs, journal);
struct journal_buf *buf = journal_cur_buf(j);
union journal_res_state old, new;
int u64s;
u64 v;
+ BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
+
lockdep_assert_held(&j->lock);
BUG_ON(journal_entry_is_open(j));
if (j->blocked)
- return -EAGAIN;
+ return cur_entry_blocked;
if (j->cur_entry_error)
return j->cur_entry_error;
u64s = clamp_t(int, u64s, 0, JOURNAL_ENTRY_CLOSED_VAL - 1);
if (u64s <= le32_to_cpu(buf->data->u64s))
- return -ENOSPC;
+ return cur_entry_journal_full;
/*
* Must be set before marking the journal entry as open:
old.v = new.v = v;
if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL)
- return -EROFS;
+ return cur_entry_insufficient_devices;
/* Handle any already added entries */
new.cur_entry_offset = le32_to_cpu(buf->data->u64s);
j->res_get_blocked_start);
j->res_get_blocked_start = 0;
- mod_delayed_work(system_freezable_wq,
+ mod_delayed_work(c->io_complete_wq,
&j->write_work,
msecs_to_jiffies(j->write_delay_ms));
journal_wake(j);
static bool journal_quiesced(struct journal *j)
{
- union journal_res_state state = READ_ONCE(j->reservations);
- bool ret = !state.prev_buf_unwritten && !__journal_entry_is_open(state);
+ union journal_res_state s = READ_ONCE(j->reservations);
+ bool ret = s.idx == s.unwritten_idx && !__journal_entry_is_open(s);
if (!ret)
journal_entry_close(j);
journal_entry_close(j);
}
-/*
- * Given an inode number, if that inode number has data in the journal that
- * hasn't yet been flushed, return the journal sequence number that needs to be
- * flushed:
- */
-u64 bch2_inode_journal_seq(struct journal *j, u64 inode)
-{
- size_t h = hash_64(inode, ilog2(sizeof(j->buf[0].has_inode) * 8));
- u64 seq = 0;
-
- if (!test_bit(h, j->buf[0].has_inode) &&
- !test_bit(h, j->buf[1].has_inode))
- return 0;
-
- spin_lock(&j->lock);
- if (test_bit(h, journal_cur_buf(j)->has_inode))
- seq = journal_cur_seq(j);
- else if (test_bit(h, journal_prev_buf(j)->has_inode))
- seq = journal_cur_seq(j) - 1;
- spin_unlock(&j->lock);
-
- return seq;
-}
-
static int __journal_res_get(struct journal *j, struct journal_res *res,
unsigned flags)
{
* Don't want to close current journal entry, just need to
* invoke reclaim:
*/
- ret = -ENOSPC;
+ ret = cur_entry_journal_full;
goto unlock;
}
* there's still a previous one in flight:
*/
trace_journal_entry_full(c);
- ret = -EAGAIN;
+ ret = cur_entry_blocked;
} else {
ret = journal_entry_open(j);
}
unlock:
- if ((ret == -EAGAIN || ret == -ENOSPC) &&
- !j->res_get_blocked_start)
+ if ((ret && ret != cur_entry_insufficient_devices) &&
+ !j->res_get_blocked_start) {
j->res_get_blocked_start = local_clock() ?: 1;
+ trace_journal_full(c);
+ }
can_discard = j->can_discard;
spin_unlock(&j->lock);
if (!ret)
goto retry;
- if (ret == -ENOSPC) {
- BUG_ON(!can_discard && (flags & JOURNAL_RES_GET_RESERVED));
-
- /*
- * Journal is full - can't rely on reclaim from work item due to
- * freezing:
- */
- trace_journal_full(c);
+ if ((ret == cur_entry_journal_full ||
+ ret == cur_entry_journal_pin_full) &&
+ !can_discard &&
+ j->reservations.idx == j->reservations.unwritten_idx &&
+ (flags & JOURNAL_RES_GET_RESERVED)) {
+ char *journal_debug_buf = kmalloc(4096, GFP_ATOMIC);
+
+ bch_err(c, "Journal stuck!");
+ if (journal_debug_buf) {
+ bch2_journal_debug_to_text(&_PBUF(journal_debug_buf, 4096), j);
+ bch_err(c, "%s", journal_debug_buf);
+
+ bch2_journal_pins_to_text(&_PBUF(journal_debug_buf, 4096), j);
+ bch_err(c, "Journal pins:\n%s", journal_debug_buf);
+ kfree(journal_debug_buf);
+ }
- if (!(flags & JOURNAL_RES_GET_NONBLOCK)) {
- if (can_discard) {
- bch2_journal_do_discards(j);
- goto retry;
- }
+ bch2_fatal_error(c);
+ dump_stack();
+ }
- if (mutex_trylock(&j->reclaim_lock)) {
- bch2_journal_reclaim(j);
- mutex_unlock(&j->reclaim_lock);
- }
+ /*
+ * Journal is full - can't rely on reclaim from work item due to
+ * freezing:
+ */
+ if ((ret == cur_entry_journal_full ||
+ ret == cur_entry_journal_pin_full) &&
+ !(flags & JOURNAL_RES_GET_NONBLOCK)) {
+ if (can_discard) {
+ bch2_journal_do_discards(j);
+ goto retry;
}
- ret = -EAGAIN;
+ if (mutex_trylock(&j->reclaim_lock)) {
+ bch2_journal_reclaim(j);
+ mutex_unlock(&j->reclaim_lock);
+ }
}
- return ret;
+ return ret == cur_entry_insufficient_devices ? -EROFS : -EAGAIN;
}
/*
static bool journal_preres_available(struct journal *j,
struct journal_preres *res,
- unsigned new_u64s)
+ unsigned new_u64s,
+ unsigned flags)
{
- bool ret = bch2_journal_preres_get_fast(j, res, new_u64s);
+ bool ret = bch2_journal_preres_get_fast(j, res, new_u64s, flags, true);
- if (!ret)
- bch2_journal_reclaim_work(&j->reclaim_work.work);
+ if (!ret && mutex_trylock(&j->reclaim_lock)) {
+ bch2_journal_reclaim(j);
+ mutex_unlock(&j->reclaim_lock);
+ }
return ret;
}
int __bch2_journal_preres_get(struct journal *j,
struct journal_preres *res,
- unsigned new_u64s)
+ unsigned new_u64s,
+ unsigned flags)
{
int ret;
closure_wait_event(&j->preres_wait,
(ret = bch2_journal_error(j)) ||
- journal_preres_available(j, res, new_u64s));
+ journal_preres_available(j, res, new_u64s, flags));
return ret;
}
/* journal flushing: */
-u64 bch2_journal_last_unwritten_seq(struct journal *j)
-{
- u64 seq;
-
- spin_lock(&j->lock);
- seq = journal_cur_seq(j);
- if (j->reservations.prev_buf_unwritten)
- seq--;
- spin_unlock(&j->lock);
-
- return seq;
-}
-
/**
- * bch2_journal_open_seq_async - try to open a new journal entry if @seq isn't
- * open yet, or wait if we cannot
+ * bch2_journal_flush_seq_async - wait for a journal entry to be written
*
- * used by the btree interior update machinery, when it needs to write a new
- * btree root - every journal entry contains the roots of all the btrees, so it
- * doesn't need to bother with getting a journal reservation
+ * like bch2_journal_wait_on_seq, except that it triggers a write immediately if
+ * necessary
*/
-int bch2_journal_open_seq_async(struct journal *j, u64 seq, struct closure *cl)
+int bch2_journal_flush_seq_async(struct journal *j, u64 seq,
+ struct closure *parent)
{
- struct bch_fs *c = container_of(j, struct bch_fs, journal);
- int ret;
+ struct journal_buf *buf;
+ int ret = 0;
+
+ if (seq <= j->flushed_seq_ondisk)
+ return 1;
spin_lock(&j->lock);
- /*
- * Can't try to open more than one sequence number ahead:
- */
- BUG_ON(journal_cur_seq(j) < seq && !journal_entry_is_open(j));
+ if (WARN_ONCE(seq > journal_cur_seq(j),
+ "requested to flush journal seq %llu, but currently at %llu",
+ seq, journal_cur_seq(j)))
+ goto out;
- if (journal_cur_seq(j) > seq ||
- journal_entry_is_open(j)) {
- spin_unlock(&j->lock);
- return 0;
+ /* Recheck under lock: */
+ if (j->err_seq && seq >= j->err_seq) {
+ ret = -EIO;
+ goto out;
}
- if (journal_cur_seq(j) < seq &&
- !__journal_entry_close(j)) {
- /* haven't finished writing out the previous one: */
- trace_journal_entry_full(c);
- ret = -EAGAIN;
- } else {
- BUG_ON(journal_cur_seq(j) != seq);
-
- ret = journal_entry_open(j);
+ if (seq <= j->flushed_seq_ondisk) {
+ ret = 1;
+ goto out;
}
- if ((ret == -EAGAIN || ret == -ENOSPC) &&
- !j->res_get_blocked_start)
- j->res_get_blocked_start = local_clock() ?: 1;
-
- if (ret == -EAGAIN || ret == -ENOSPC)
- closure_wait(&j->async_wait, cl);
+ /* if seq was written, but not flushed - flush a newer one instead */
+ seq = max(seq, last_unwritten_seq(j));
- spin_unlock(&j->lock);
-
- if (ret == -ENOSPC) {
- trace_journal_full(c);
- bch2_journal_reclaim_work(&j->reclaim_work.work);
- ret = -EAGAIN;
- }
-
- return ret;
-}
+recheck_need_open:
+ if (seq == journal_cur_seq(j) && !journal_entry_is_open(j)) {
+ struct journal_res res = { 0 };
-static int journal_seq_error(struct journal *j, u64 seq)
-{
- union journal_res_state state = READ_ONCE(j->reservations);
-
- if (seq == journal_cur_seq(j))
- return bch2_journal_error(j);
-
- if (seq + 1 == journal_cur_seq(j) &&
- !state.prev_buf_unwritten &&
- seq > j->seq_ondisk)
- return -EIO;
-
- return 0;
-}
-
-static inline struct journal_buf *
-journal_seq_to_buf(struct journal *j, u64 seq)
-{
- /* seq should be for a journal entry that has been opened: */
- BUG_ON(seq > journal_cur_seq(j));
- BUG_ON(seq == journal_cur_seq(j) &&
- j->reservations.cur_entry_offset == JOURNAL_ENTRY_CLOSED_VAL);
-
- if (seq == journal_cur_seq(j))
- return journal_cur_buf(j);
- if (seq + 1 == journal_cur_seq(j) &&
- j->reservations.prev_buf_unwritten)
- return journal_prev_buf(j);
- return NULL;
-}
+ spin_unlock(&j->lock);
-/**
- * bch2_journal_wait_on_seq - wait for a journal entry to be written
- *
- * does _not_ cause @seq to be written immediately - if there is no other
- * activity to cause the relevant journal entry to be filled up or flushed it
- * can wait for an arbitrary amount of time (up to @j->write_delay_ms, which is
- * configurable).
- */
-void bch2_journal_wait_on_seq(struct journal *j, u64 seq,
- struct closure *parent)
-{
- struct journal_buf *buf;
+ ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
+ if (ret)
+ return ret;
- spin_lock(&j->lock);
+ seq = res.seq;
+ buf = j->buf + (seq & JOURNAL_BUF_MASK);
+ buf->must_flush = true;
+ set_bit(JOURNAL_NEED_WRITE, &j->flags);
- if ((buf = journal_seq_to_buf(j, seq))) {
- if (!closure_wait(&buf->wait, parent))
+ if (parent && !closure_wait(&buf->wait, parent))
BUG();
- if (seq == journal_cur_seq(j)) {
- smp_mb();
- if (bch2_journal_error(j))
- closure_wake_up(&buf->wait);
- }
- }
+ bch2_journal_res_put(j, &res);
- spin_unlock(&j->lock);
-}
-
-/**
- * bch2_journal_flush_seq_async - wait for a journal entry to be written
- *
- * like bch2_journal_wait_on_seq, except that it triggers a write immediately if
- * necessary
- */
-void bch2_journal_flush_seq_async(struct journal *j, u64 seq,
- struct closure *parent)
-{
- struct journal_buf *buf;
-
- spin_lock(&j->lock);
-
- if (parent &&
- (buf = journal_seq_to_buf(j, seq)))
- if (!closure_wait(&buf->wait, parent))
- BUG();
-
- if (seq == journal_cur_seq(j))
- __journal_entry_close(j);
- spin_unlock(&j->lock);
-}
+ spin_lock(&j->lock);
+ goto want_write;
+ }
-static int journal_seq_flushed(struct journal *j, u64 seq)
-{
- int ret;
+ /*
+ * if write was kicked off without a flush, flush the next sequence
+ * number instead
+ */
+ buf = journal_seq_to_buf(j, seq);
+ if (buf->noflush) {
+ seq++;
+ goto recheck_need_open;
+ }
- spin_lock(&j->lock);
- ret = seq <= j->seq_ondisk ? 1 : journal_seq_error(j, seq);
+ buf->must_flush = true;
+ if (parent && !closure_wait(&buf->wait, parent))
+ BUG();
+want_write:
if (seq == journal_cur_seq(j))
- __journal_entry_close(j);
+ journal_entry_want_write(j);
+out:
spin_unlock(&j->lock);
-
return ret;
}
u64 start_time = local_clock();
int ret, ret2;
- ret = wait_event_killable(j->wait, (ret2 = journal_seq_flushed(j, seq)));
+ ret = wait_event_interruptible(j->wait, (ret2 = bch2_journal_flush_seq_async(j, seq, NULL)));
- bch2_time_stats_update(j->flush_seq_time, start_time);
+ if (!ret)
+ bch2_time_stats_update(j->flush_seq_time, start_time);
return ret ?: ret2 < 0 ? ret2 : 0;
}
-/**
- * bch2_journal_meta_async - force a journal entry to be written
- */
-void bch2_journal_meta_async(struct journal *j, struct closure *parent)
-{
- struct journal_res res;
-
- memset(&res, 0, sizeof(res));
-
- bch2_journal_res_get(j, &res, jset_u64s(0), 0);
- bch2_journal_res_put(j, &res);
-
- bch2_journal_flush_seq_async(j, res.seq, parent);
-}
-
int bch2_journal_meta(struct journal *j)
{
struct journal_res res;
if (nr <= ja->nr)
return 0;
- ret = -ENOMEM;
new_buckets = kzalloc(nr * sizeof(u64), GFP_KERNEL);
new_bucket_seq = kzalloc(nr * sizeof(u64), GFP_KERNEL);
- if (!new_buckets || !new_bucket_seq)
+ if (!new_buckets || !new_bucket_seq) {
+ ret = -ENOMEM;
goto err;
+ }
journal_buckets = bch2_sb_resize_journal(&ca->disk_sb,
- nr + sizeof(*journal_buckets) / sizeof(u64));
- if (!journal_buckets)
+ nr + sizeof(*journal_buckets) / sizeof(u64));
+ if (!journal_buckets) {
+ ret = -ENOSPC;
goto err;
+ }
/*
* We may be called from the device add path, before the new device has
* actually been added to the running filesystem:
*/
- if (c)
+ if (!new_fs)
spin_lock(&c->journal.lock);
memcpy(new_buckets, ja->buckets, ja->nr * sizeof(u64));
swap(new_buckets, ja->buckets);
swap(new_bucket_seq, ja->bucket_seq);
- if (c)
+ if (!new_fs)
spin_unlock(&c->journal.lock);
while (ja->nr < nr) {
struct open_bucket *ob = NULL;
unsigned pos;
- long bucket;
+ long b;
if (new_fs) {
- bucket = bch2_bucket_alloc_new_fs(ca);
- if (bucket < 0) {
+ if (c)
+ percpu_down_read(&c->mark_lock);
+ b = bch2_bucket_alloc_new_fs(ca);
+ if (b < 0) {
+ percpu_up_read(&c->mark_lock);
ret = -ENOSPC;
goto err;
}
} else {
- ob = bch2_bucket_alloc(c, ca, RESERVE_ALLOC,
+ rcu_read_lock();
+ ob = bch2_bucket_alloc(c, ca, RESERVE_NONE,
false, cl);
+ rcu_read_unlock();
if (IS_ERR(ob)) {
ret = cl ? -EAGAIN : -ENOSPC;
goto err;
}
- bucket = sector_to_bucket(ca, ob->ptr.offset);
+ b = sector_to_bucket(ca, ob->ptr.offset);
}
- if (c) {
- percpu_down_read_preempt_disable(&c->mark_lock);
+ if (c)
spin_lock(&c->journal.lock);
- } else {
- preempt_disable();
- }
+
+ /*
+ * XXX
+ * For resize at runtime, we should be writing the new
+ * superblock before inserting into the journal array
+ */
pos = ja->nr ? (ja->cur_idx + 1) % ja->nr : 0;
__array_insert_item(ja->buckets, ja->nr, pos);
__array_insert_item(journal_buckets->buckets, ja->nr, pos);
ja->nr++;
- ja->buckets[pos] = bucket;
+ ja->buckets[pos] = b;
ja->bucket_seq[pos] = 0;
- journal_buckets->buckets[pos] = cpu_to_le64(bucket);
+ journal_buckets->buckets[pos] = cpu_to_le64(b);
if (pos <= ja->discard_idx)
ja->discard_idx = (ja->discard_idx + 1) % ja->nr;
if (pos <= ja->cur_idx)
ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
- bch2_mark_metadata_bucket(c, ca, bucket, BCH_DATA_JOURNAL,
- ca->mi.bucket_size,
- gc_phase(GC_PHASE_SB),
- 0);
-
- if (c) {
+ if (c)
spin_unlock(&c->journal.lock);
- percpu_up_read_preempt_enable(&c->mark_lock);
+
+ if (new_fs) {
+ bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
+ ca->mi.bucket_size,
+ gc_phase(GC_PHASE_SB),
+ 0);
+ if (c)
+ percpu_up_read(&c->mark_lock);
} else {
- preempt_enable();
- }
+ ret = bch2_trans_do(c, NULL, NULL, BTREE_INSERT_NOFAIL,
+ bch2_trans_mark_metadata_bucket(&trans, ca,
+ b, BCH_DATA_journal,
+ ca->mi.bucket_size));
- if (!new_fs)
bch2_open_bucket_put(c, ob);
- }
- ret = 0;
+ if (ret)
+ goto err;
+ }
+ }
err:
+ bch2_sb_resize_journal(&ca->disk_sb,
+ ja->nr + sizeof(*journal_buckets) / sizeof(u64));
kfree(new_bucket_seq);
kfree(new_buckets);
if (dynamic_fault("bcachefs:add:journal_alloc"))
return -ENOMEM;
+ /* 1/128th of the device by default: */
+ nr = ca->mi.nbuckets >> 7;
+
/*
- * clamp journal size to 1024 buckets or 512MB (in sectors), whichever
+ * clamp journal size to 8192 buckets or 8GB (in sectors), whichever
* is smaller:
*/
- nr = clamp_t(unsigned, ca->mi.nbuckets >> 8,
+ nr = clamp_t(unsigned, nr,
BCH_JOURNAL_BUCKETS_MIN,
- min(1 << 10,
- (1 << 20) / ca->mi.bucket_size));
+ min(1 << 13,
+ (1 << 24) / ca->mi.bucket_size));
return __bch2_set_nr_journal_buckets(ca, nr, true, NULL);
}
static bool bch2_journal_writing_to_device(struct journal *j, unsigned dev_idx)
{
union journal_res_state state;
- struct journal_buf *w;
- bool ret;
+ bool ret = false;
+ unsigned i;
spin_lock(&j->lock);
state = READ_ONCE(j->reservations);
- w = j->buf + !state.idx;
+ i = state.idx;
- ret = state.prev_buf_unwritten &&
- bch2_extent_has_device(bkey_i_to_s_c_extent(&w->key), dev_idx);
+ while (i != state.unwritten_idx) {
+ i = (i - 1) & JOURNAL_BUF_MASK;
+ if (bch2_bkey_has_device(bkey_i_to_s_c(&j->buf[i].key), dev_idx))
+ ret = true;
+ }
spin_unlock(&j->lock);
return ret;
void bch2_fs_journal_stop(struct journal *j)
{
- struct bch_fs *c = container_of(j, struct bch_fs, journal);
+ bch2_journal_flush_all_pins(j);
wait_event(j->wait, journal_entry_close(j));
- /* do we need to write another journal entry? */
- if (test_bit(JOURNAL_NOT_EMPTY, &j->flags) ||
- c->btree_roots_dirty)
- bch2_journal_meta(j);
+ /*
+ * Always write a new journal entry, to make sure the clock hands are up
+ * to date (and match the superblock)
+ */
+ bch2_journal_meta(j);
journal_quiesce(j);
BUG_ON(!bch2_journal_error(j) &&
- test_bit(JOURNAL_NOT_EMPTY, &j->flags));
+ test_bit(JOURNAL_REPLAY_DONE, &j->flags) &&
+ (journal_entry_is_open(j) ||
+ j->last_empty_seq + 1 != journal_cur_seq(j)));
cancel_delayed_work_sync(&j->write_work);
- cancel_delayed_work_sync(&j->reclaim_work);
+ bch2_journal_reclaim_stop(j);
}
int bch2_fs_journal_start(struct journal *j, u64 cur_seq,
u64 last_seq = cur_seq, nr, seq;
if (!list_empty(journal_entries))
- last_seq = le64_to_cpu(list_first_entry(journal_entries,
- struct journal_replay,
- list)->j.seq);
+ last_seq = le64_to_cpu(list_last_entry(journal_entries,
+ struct journal_replay, list)->j.last_seq);
nr = cur_seq - last_seq;
j->pin.back = cur_seq;
atomic64_set(&j->seq, cur_seq - 1);
- fifo_for_each_entry_ptr(p, &j->pin, seq) {
- INIT_LIST_HEAD(&p->list);
- INIT_LIST_HEAD(&p->flushed);
- atomic_set(&p->count, 1);
- p->devs.nr = 0;
- }
+ fifo_for_each_entry_ptr(p, &j->pin, seq)
+ journal_pin_list_init(p, 1);
list_for_each_entry(i, journal_entries, list) {
+ unsigned ptr;
+
seq = le64_to_cpu(i->j.seq);
+ BUG_ON(seq >= cur_seq);
+
+ if (seq < last_seq)
+ continue;
- BUG_ON(seq < last_seq || seq >= cur_seq);
+ p = journal_seq_pin(j, seq);
- journal_seq_pin(j, seq)->devs = i->devs;
+ p->devs.nr = 0;
+ for (ptr = 0; ptr < i->nr_ptrs; ptr++)
+ bch2_dev_list_add_dev(&p->devs, i->ptrs[ptr].dev);
}
spin_lock(&j->lock);
set_bit(JOURNAL_STARTED, &j->flags);
+ j->last_flush_write = jiffies;
+
+ journal_pin_new_entry(j);
+
+ j->reservations.idx = j->reservations.unwritten_idx = journal_cur_seq(j);
- journal_pin_new_entry(j, 1);
bch2_journal_buf_init(j);
c->last_bucket_seq_cleanup = journal_cur_seq(j);
bch2_journal_space_available(j);
spin_unlock(&j->lock);
- return 0;
+ return bch2_journal_reclaim_start(j);
}
/* init/exit: */
void bch2_fs_journal_exit(struct journal *j)
{
- kvpfree(j->buf[1].data, j->buf[1].buf_size);
- kvpfree(j->buf[0].data, j->buf[0].buf_size);
+ unsigned i;
+
+ for (i = 0; i < ARRAY_SIZE(j->buf); i++)
+ kvpfree(j->buf[i].data, j->buf[i].buf_size);
free_fifo(&j->pin);
}
{
struct bch_fs *c = container_of(j, struct bch_fs, journal);
static struct lock_class_key res_key;
+ unsigned i;
int ret = 0;
pr_verbose_init(c->opts, "");
spin_lock_init(&j->err_lock);
init_waitqueue_head(&j->wait);
INIT_DELAYED_WORK(&j->write_work, journal_write_work);
- INIT_DELAYED_WORK(&j->reclaim_work, bch2_journal_reclaim_work);
+ init_waitqueue_head(&j->reclaim_wait);
init_waitqueue_head(&j->pin_flush_wait);
mutex_init(&j->reclaim_lock);
mutex_init(&j->discard_lock);
lockdep_init_map(&j->res_map, "journal res", &res_key, 0);
- j->buf[0].buf_size = JOURNAL_ENTRY_SIZE_MIN;
- j->buf[1].buf_size = JOURNAL_ENTRY_SIZE_MIN;
j->write_delay_ms = 1000;
j->reclaim_delay_ms = 100;
- /* Btree roots: */
- j->entry_u64s_reserved +=
- BTREE_ID_NR * (JSET_KEYS_U64s + BKEY_EXTENT_U64s_MAX);
-
atomic64_set(&j->reservations.counter,
((union journal_res_state)
{ .cur_entry_offset = JOURNAL_ENTRY_CLOSED_VAL }).v);
- if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)) ||
- !(j->buf[0].data = kvpmalloc(j->buf[0].buf_size, GFP_KERNEL)) ||
- !(j->buf[1].data = kvpmalloc(j->buf[1].buf_size, GFP_KERNEL))) {
+ if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL))) {
ret = -ENOMEM;
goto out;
}
+ for (i = 0; i < ARRAY_SIZE(j->buf); i++) {
+ j->buf[i].buf_size = JOURNAL_ENTRY_SIZE_MIN;
+ j->buf[i].data = kvpmalloc(j->buf[i].buf_size, GFP_KERNEL);
+ if (!j->buf[i].data) {
+ ret = -ENOMEM;
+ goto out;
+ }
+ }
+
j->pin.front = j->pin.back = 1;
out:
pr_verbose_init(c->opts, "ret %i", ret);
/* debug: */
-ssize_t bch2_journal_print_debug(struct journal *j, char *buf)
+void __bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
{
- struct printbuf out = _PBUF(buf, PAGE_SIZE);
struct bch_fs *c = container_of(j, struct bch_fs, journal);
union journal_res_state s;
struct bch_dev *ca;
- unsigned iter;
+ unsigned i;
rcu_read_lock();
- spin_lock(&j->lock);
s = READ_ONCE(j->reservations);
- pr_buf(&out,
+ pr_buf(out,
"active journal entries:\t%llu\n"
"seq:\t\t\t%llu\n"
"last_seq:\t\t%llu\n"
"last_seq_ondisk:\t%llu\n"
+ "flushed_seq_ondisk:\t%llu\n"
"prereserved:\t\t%u/%u\n"
+ "each entry reserved:\t%u\n"
+ "nr flush writes:\t%llu\n"
+ "nr noflush writes:\t%llu\n"
+ "nr direct reclaim:\t%llu\n"
+ "nr background reclaim:\t%llu\n"
+ "reclaim kicked:\t\t%u\n"
+ "reclaim runs in:\t%u ms\n"
"current entry sectors:\t%u\n"
+ "current entry error:\t%u\n"
"current entry:\t\t",
fifo_used(&j->pin),
journal_cur_seq(j),
journal_last_seq(j),
j->last_seq_ondisk,
+ j->flushed_seq_ondisk,
j->prereserved.reserved,
j->prereserved.remaining,
- j->cur_entry_sectors);
+ j->entry_u64s_reserved,
+ j->nr_flush_writes,
+ j->nr_noflush_writes,
+ j->nr_direct_reclaim,
+ j->nr_background_reclaim,
+ j->reclaim_kicked,
+ jiffies_to_msecs(j->next_reclaim - jiffies),
+ j->cur_entry_sectors,
+ j->cur_entry_error);
switch (s.cur_entry_offset) {
case JOURNAL_ENTRY_ERROR_VAL:
- pr_buf(&out, "error\n");
+ pr_buf(out, "error\n");
break;
case JOURNAL_ENTRY_CLOSED_VAL:
- pr_buf(&out, "closed\n");
+ pr_buf(out, "closed\n");
break;
default:
- pr_buf(&out, "%u/%u\n",
+ pr_buf(out, "%u/%u\n",
s.cur_entry_offset,
j->cur_entry_u64s);
break;
}
- pr_buf(&out,
- "current entry refs:\t%u\n"
- "prev entry unwritten:\t",
- journal_state_count(s, s.idx));
+ pr_buf(out,
+ "current entry:\t\tidx %u refcount %u\n",
+ s.idx, journal_state_count(s, s.idx));
- if (s.prev_buf_unwritten)
- pr_buf(&out, "yes, ref %u sectors %u\n",
- journal_state_count(s, !s.idx),
- journal_prev_buf(j)->sectors);
- else
- pr_buf(&out, "no\n");
+ i = s.idx;
+ while (i != s.unwritten_idx) {
+ i = (i - 1) & JOURNAL_BUF_MASK;
- pr_buf(&out,
+ pr_buf(out, "unwritten entry:\tidx %u refcount %u sectors %u\n",
+ i, journal_state_count(s, i), j->buf[i].sectors);
+ }
+
+ pr_buf(out,
"need write:\t\t%i\n"
"replay done:\t\t%i\n",
test_bit(JOURNAL_NEED_WRITE, &j->flags),
test_bit(JOURNAL_REPLAY_DONE, &j->flags));
- for_each_member_device_rcu(ca, c, iter,
- &c->rw_devs[BCH_DATA_JOURNAL]) {
+ pr_buf(out, "space:\n");
+ pr_buf(out, "\tdiscarded\t%u:%u\n",
+ j->space[journal_space_discarded].next_entry,
+ j->space[journal_space_discarded].total);
+ pr_buf(out, "\tclean ondisk\t%u:%u\n",
+ j->space[journal_space_clean_ondisk].next_entry,
+ j->space[journal_space_clean_ondisk].total);
+ pr_buf(out, "\tclean\t\t%u:%u\n",
+ j->space[journal_space_clean].next_entry,
+ j->space[journal_space_clean].total);
+ pr_buf(out, "\ttotal\t\t%u:%u\n",
+ j->space[journal_space_total].next_entry,
+ j->space[journal_space_total].total);
+
+ for_each_member_device_rcu(ca, c, i,
+ &c->rw_devs[BCH_DATA_journal]) {
struct journal_device *ja = &ca->journal;
+ if (!test_bit(ca->dev_idx, c->rw_devs[BCH_DATA_journal].d))
+ continue;
+
if (!ja->nr)
continue;
- pr_buf(&out,
+ pr_buf(out,
"dev %u:\n"
"\tnr\t\t%u\n"
+ "\tbucket size\t%u\n"
"\tavailable\t%u:%u\n"
- "\tdiscard_idx\t\t%u\n"
- "\tdirty_idx_ondisk\t%u (seq %llu)\n"
- "\tdirty_idx\t\t%u (seq %llu)\n"
+ "\tdiscard_idx\t%u\n"
+ "\tdirty_ondisk\t%u (seq %llu)\n"
+ "\tdirty_idx\t%u (seq %llu)\n"
"\tcur_idx\t\t%u (seq %llu)\n",
- iter, ja->nr,
+ i, ja->nr, ca->mi.bucket_size,
bch2_journal_dev_buckets_available(j, ja, journal_space_discarded),
ja->sectors_free,
ja->discard_idx,
ja->cur_idx, ja->bucket_seq[ja->cur_idx]);
}
- spin_unlock(&j->lock);
rcu_read_unlock();
+}
- return out.pos - buf;
+void bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
+{
+ spin_lock(&j->lock);
+ __bch2_journal_debug_to_text(out, j);
+ spin_unlock(&j->lock);
}
-ssize_t bch2_journal_print_pins(struct journal *j, char *buf)
+void bch2_journal_pins_to_text(struct printbuf *out, struct journal *j)
{
- struct printbuf out = _PBUF(buf, PAGE_SIZE);
struct journal_entry_pin_list *pin_list;
struct journal_entry_pin *pin;
u64 i;
spin_lock(&j->lock);
fifo_for_each_entry_ptr(pin_list, &j->pin, i) {
- pr_buf(&out, "%llu: count %u\n",
+ pr_buf(out, "%llu: count %u\n",
i, atomic_read(&pin_list->count));
list_for_each_entry(pin, &pin_list->list, list)
- pr_buf(&out, "\t%p %pf\n",
+ pr_buf(out, "\t%px %ps\n",
pin, pin->flush);
if (!list_empty(&pin_list->flushed))
- pr_buf(&out, "flushed:\n");
+ pr_buf(out, "flushed:\n");
list_for_each_entry(pin, &pin_list->flushed, list)
- pr_buf(&out, "\t%p %pf\n",
+ pr_buf(out, "\t%px %ps\n",
pin, pin->flush);
}
spin_unlock(&j->lock);
-
- return out.pos - buf;
}
projects / bcachefs-tools-debian / blobdiff