#include "alloc_foreground.h"
#include "btree_io.h"
#include "btree_update_interior.h"
+#include "btree_write_buffer.h"
#include "buckets.h"
#include "checksum.h"
#include "disk_groups.h"
#include "sb-clean.h"
#include "trace.h"
+void bch2_journal_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
+ struct journal_replay *j)
+{
+ darray_for_each(j->ptrs, i) {
+ struct bch_dev *ca = bch_dev_bkey_exists(c, i->dev);
+ u64 offset;
+
+ div64_u64_rem(i->sector, ca->mi.bucket_size, &offset);
+
+ if (i != j->ptrs.data)
+ prt_printf(out, " ");
+ prt_printf(out, "%u:%u:%u (sector %llu)",
+ i->dev, i->bucket, i->bucket_offset, i->sector);
+ }
+}
+
+static void bch2_journal_replay_to_text(struct printbuf *out, struct bch_fs *c,
+ struct journal_replay *j)
+{
+ prt_printf(out, "seq %llu ", le64_to_cpu(j->j.seq));
+
+ bch2_journal_ptrs_to_text(out, c, j);
+
+ struct jset_entry *entry;
+ for_each_jset_entry_type(entry, &j->j, BCH_JSET_ENTRY_datetime) {
+ struct jset_entry_datetime *datetime =
+ container_of(entry, struct jset_entry_datetime, entry);
+ bch2_prt_datetime(out, le64_to_cpu(datetime->seconds));
+ break;
+ }
+}
+
static struct nonce journal_nonce(const struct jset *jset)
{
return (struct nonce) {{
}};
}
-static bool jset_csum_good(struct bch_fs *c, struct jset *j)
+static bool jset_csum_good(struct bch_fs *c, struct jset *j, struct bch_csum *csum)
{
- return bch2_checksum_type_valid(c, JSET_CSUM_TYPE(j)) &&
- !bch2_crc_cmp(j->csum,
- csum_vstruct(c, JSET_CSUM_TYPE(j), journal_nonce(j), j));
+ if (!bch2_checksum_type_valid(c, JSET_CSUM_TYPE(j))) {
+ *csum = (struct bch_csum) {};
+ return false;
+ }
+
+ *csum = csum_vstruct(c, JSET_CSUM_TYPE(j), journal_nonce(j), j);
+ return !bch2_crc_cmp(j->csum, *csum);
}
static inline u32 journal_entry_radix_idx(struct bch_fs *c, u64 seq)
BUG_ON(*p != i);
*p = NULL;
- kvpfree(i, offsetof(struct journal_replay, j) +
- vstruct_bytes(&i->j));
+ kvfree(i);
}
static void journal_replay_free(struct bch_fs *c, struct journal_replay *i)
{
struct genradix_iter iter;
struct journal_replay **_i, *i, *dup;
- struct journal_ptr *ptr;
size_t bytes = vstruct_bytes(j);
u64 last_seq = !JSET_NO_FLUSH(j) ? le64_to_cpu(j->last_seq) : 0;
+ struct printbuf buf = PRINTBUF;
int ret = JOURNAL_ENTRY_ADD_OK;
/* Is this entry older than the range we need? */
*/
dup = *_i;
if (dup) {
- if (bytes == vstruct_bytes(&dup->j) &&
- !memcmp(j, &dup->j, bytes)) {
- i = dup;
- goto found;
- }
+ bool identical = bytes == vstruct_bytes(&dup->j) &&
+ !memcmp(j, &dup->j, bytes);
+ bool not_identical = !identical &&
+ entry_ptr.csum_good &&
+ dup->csum_good;
+
+ bool same_device = false;
+ darray_for_each(dup->ptrs, ptr)
+ if (ptr->dev == ca->dev_idx)
+ same_device = true;
+
+ ret = darray_push(&dup->ptrs, entry_ptr);
+ if (ret)
+ goto out;
- if (!entry_ptr.csum_good) {
- i = dup;
- goto found;
- }
+ bch2_journal_replay_to_text(&buf, c, dup);
+
+ fsck_err_on(same_device,
+ c, journal_entry_dup_same_device,
+ "duplicate journal entry on same device\n %s",
+ buf.buf);
- if (!dup->csum_good)
+ fsck_err_on(not_identical,
+ c, journal_entry_replicas_data_mismatch,
+ "found duplicate but non identical journal entries\n %s",
+ buf.buf);
+
+ if (entry_ptr.csum_good && !identical)
goto replace;
- fsck_err(c, journal_entry_replicas_data_mismatch,
- "found duplicate but non identical journal entries (seq %llu)",
- le64_to_cpu(j->seq));
- i = dup;
- goto found;
+ goto out;
}
replace:
- i = kvpmalloc(offsetof(struct journal_replay, j) + bytes, GFP_KERNEL);
+ i = kvmalloc(offsetof(struct journal_replay, j) + bytes, GFP_KERNEL);
if (!i)
return -BCH_ERR_ENOMEM_journal_entry_add;
- i->nr_ptrs = 0;
+ darray_init(&i->ptrs);
i->csum_good = entry_ptr.csum_good;
i->ignore = false;
unsafe_memcpy(&i->j, j, bytes, "embedded variable length struct");
- i->ptrs[i->nr_ptrs++] = entry_ptr;
if (dup) {
- if (dup->nr_ptrs >= ARRAY_SIZE(dup->ptrs)) {
- bch_err(c, "found too many copies of journal entry %llu",
- le64_to_cpu(i->j.seq));
- dup->nr_ptrs = ARRAY_SIZE(dup->ptrs) - 1;
- }
-
/* The first ptr should represent the jset we kept: */
- memcpy(i->ptrs + i->nr_ptrs,
- dup->ptrs,
- sizeof(dup->ptrs[0]) * dup->nr_ptrs);
- i->nr_ptrs += dup->nr_ptrs;
+ darray_for_each(dup->ptrs, ptr)
+ darray_push(&i->ptrs, *ptr);
__journal_replay_free(c, dup);
+ } else {
+ darray_push(&i->ptrs, entry_ptr);
}
*_i = i;
- return 0;
-found:
- for (ptr = i->ptrs; ptr < i->ptrs + i->nr_ptrs; ptr++) {
- if (ptr->dev == ca->dev_idx) {
- bch_err(c, "duplicate journal entry %llu on same device",
- le64_to_cpu(i->j.seq));
- goto out;
- }
- }
-
- if (i->nr_ptrs >= ARRAY_SIZE(i->ptrs)) {
- bch_err(c, "found too many copies of journal entry %llu",
- le64_to_cpu(i->j.seq));
- goto out;
- }
-
- i->ptrs[i->nr_ptrs++] = entry_ptr;
out:
fsck_err:
+ printbuf_exit(&buf);
return ret;
}
return 0;
}
- return journal_validate_key(c, jset, entry, 1, entry->btree_id, k,
- version, big_endian, flags);
+ ret = journal_validate_key(c, jset, entry, 1, entry->btree_id, k,
+ version, big_endian, flags);
+ if (ret == FSCK_DELETED_KEY)
+ ret = 0;
fsck_err:
return ret;
}
struct jset_entry_data_usage *u =
container_of(entry, struct jset_entry_data_usage, entry);
unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
+ struct printbuf err = PRINTBUF;
int ret = 0;
if (journal_entry_err_on(bytes < sizeof(*u) ||
journal_entry_data_usage_bad_size,
"invalid journal entry usage: bad size")) {
journal_entry_null_range(entry, vstruct_next(entry));
- return ret;
+ goto out;
}
+ if (journal_entry_err_on(bch2_replicas_entry_validate(&u->r, c->disk_sb.sb, &err),
+ c, version, jset, entry,
+ journal_entry_data_usage_bad_size,
+ "invalid journal entry usage: %s", err.buf)) {
+ journal_entry_null_range(entry, vstruct_next(entry));
+ goto out;
+ }
+out:
fsck_err:
+ printbuf_exit(&err);
return ret;
}
prt_printf(out, "dev=%u", le32_to_cpu(u->dev));
for (i = 0; i < nr_types; i++) {
- if (i < BCH_DATA_NR)
- prt_printf(out, " %s", bch2_data_types[i]);
- else
- prt_printf(out, " (unknown data type %u)", i);
+ bch2_prt_data_type(out, i);
prt_printf(out, ": buckets=%llu sectors=%llu fragmented=%llu",
le64_to_cpu(u->d[i].buckets),
le64_to_cpu(u->d[i].sectors),
le64_to_cpu(u->d[i].fragmented));
}
-
- prt_printf(out, " buckets_ec: %llu", le64_to_cpu(u->buckets_ec));
}
static int journal_entry_log_validate(struct bch_fs *c,
journal_entry_btree_keys_to_text(out, c, entry);
}
+static int journal_entry_write_buffer_keys_validate(struct bch_fs *c,
+ struct jset *jset,
+ struct jset_entry *entry,
+ unsigned version, int big_endian,
+ enum bkey_invalid_flags flags)
+{
+ return journal_entry_btree_keys_validate(c, jset, entry,
+ version, big_endian, READ);
+}
+
+static void journal_entry_write_buffer_keys_to_text(struct printbuf *out, struct bch_fs *c,
+ struct jset_entry *entry)
+{
+ journal_entry_btree_keys_to_text(out, c, entry);
+}
+
+static int journal_entry_datetime_validate(struct bch_fs *c,
+ struct jset *jset,
+ struct jset_entry *entry,
+ unsigned version, int big_endian,
+ enum bkey_invalid_flags flags)
+{
+ unsigned bytes = vstruct_bytes(entry);
+ unsigned expected = 16;
+ int ret = 0;
+
+ if (journal_entry_err_on(vstruct_bytes(entry) < expected,
+ c, version, jset, entry,
+ journal_entry_dev_usage_bad_size,
+ "bad size (%u < %u)",
+ bytes, expected)) {
+ journal_entry_null_range(entry, vstruct_next(entry));
+ return ret;
+ }
+fsck_err:
+ return ret;
+}
+
+static void journal_entry_datetime_to_text(struct printbuf *out, struct bch_fs *c,
+ struct jset_entry *entry)
+{
+ struct jset_entry_datetime *datetime =
+ container_of(entry, struct jset_entry_datetime, entry);
+
+ bch2_prt_datetime(out, le64_to_cpu(datetime->seconds));
+}
+
struct jset_entry_ops {
int (*validate)(struct bch_fs *, struct jset *,
struct jset_entry *, unsigned, int,
static int jset_validate_entries(struct bch_fs *c, struct jset *jset,
enum bkey_invalid_flags flags)
{
- struct jset_entry *entry;
unsigned version = le32_to_cpu(jset->version);
int ret = 0;
return -BCH_ERR_ENOMEM_journal_read_buf_realloc;
new_size = roundup_pow_of_two(new_size);
- n = kvpmalloc(new_size, GFP_KERNEL);
+ n = kvmalloc(new_size, GFP_KERNEL);
if (!n)
return -BCH_ERR_ENOMEM_journal_read_buf_realloc;
- kvpfree(b->data, b->size);
+ kvfree(b->data);
b->data = n;
b->size = new_size;
return 0;
u64 offset = bucket_to_sector(ca, ja->buckets[bucket]),
end = offset + ca->mi.bucket_size;
bool saw_bad = false, csum_good;
+ struct printbuf err = PRINTBUF;
int ret = 0;
pr_debug("reading %u", bucket);
* found on a different device, and missing or
* no journal entries will be handled later
*/
- return 0;
+ goto out;
}
j = buf->data;
ret = journal_read_buf_realloc(buf,
vstruct_bytes(j));
if (ret)
- return ret;
+ goto err;
}
goto reread;
case JOURNAL_ENTRY_NONE:
if (!saw_bad)
- return 0;
+ goto out;
/*
* On checksum error we don't really trust the size
* field of the journal entry we read, so try reading
sectors = block_sectors(c);
goto next_block;
default:
- return ret;
+ goto err;
}
/*
* bucket:
*/
if (le64_to_cpu(j->seq) < ja->bucket_seq[bucket])
- return 0;
+ goto out;
ja->bucket_seq[bucket] = le64_to_cpu(j->seq);
- csum_good = jset_csum_good(c, j);
+ enum bch_csum_type csum_type = JSET_CSUM_TYPE(j);
+ struct bch_csum csum;
+ csum_good = jset_csum_good(c, j, &csum);
+
if (bch2_dev_io_err_on(!csum_good, ca, BCH_MEMBER_ERROR_checksum,
- "journal checksum error"))
+ "%s",
+ (printbuf_reset(&err),
+ prt_str(&err, "journal "),
+ bch2_csum_err_msg(&err, csum_type, j->csum, csum),
+ err.buf)))
saw_bad = true;
ret = bch2_encrypt(c, JSET_CSUM_TYPE(j), journal_nonce(j),
j->encrypted_start,
vstruct_end(j) - (void *) j->encrypted_start);
bch2_fs_fatal_err_on(ret, c,
- "error decrypting journal entry: %i", ret);
+ "error decrypting journal entry: %s",
+ bch2_err_str(ret));
mutex_lock(&jlist->lock);
ret = journal_entry_add(c, ca, (struct journal_ptr) {
case JOURNAL_ENTRY_ADD_OUT_OF_RANGE:
break;
default:
- return ret;
+ goto err;
}
next_block:
pr_debug("next");
j = ((void *) j) + (sectors << 9);
}
- return 0;
+out:
+ ret = 0;
+err:
+ printbuf_exit(&err);
+ return ret;
}
-static void bch2_journal_read_device(struct closure *cl)
+static CLOSURE_CALLBACK(bch2_journal_read_device)
{
- struct journal_device *ja =
- container_of(cl, struct journal_device, read);
+ closure_type(ja, struct journal_device, read);
struct bch_dev *ca = container_of(ja, struct bch_dev, journal);
struct bch_fs *c = ca->fs;
struct journal_list *jlist =
if (!r)
continue;
- for (i = 0; i < r->nr_ptrs; i++) {
- if (r->ptrs[i].dev == ca->dev_idx) {
- unsigned wrote = bucket_remainder(ca, r->ptrs[i].sector) +
+ darray_for_each(r->ptrs, i)
+ if (i->dev == ca->dev_idx) {
+ unsigned wrote = bucket_remainder(ca, i->sector) +
vstruct_sectors(&r->j, c->block_bits);
- ja->cur_idx = r->ptrs[i].bucket;
+ ja->cur_idx = i->bucket;
ja->sectors_free = ca->mi.bucket_size - wrote;
goto found;
}
- }
}
found:
mutex_unlock(&jlist->lock);
if (ja->bucket_seq[ja->cur_idx] &&
ja->sectors_free == ca->mi.bucket_size) {
+#if 0
+ /*
+ * Debug code for ZNS support, where we (probably) want to be
+ * correlated where we stopped in the journal to the zone write
+ * points:
+ */
bch_err(c, "ja->sectors_free == ca->mi.bucket_size");
bch_err(c, "cur_idx %u/%u", ja->cur_idx, ja->nr);
for (i = 0; i < 3; i++) {
bch_err(c, "bucket_seq[%u] = %llu", idx, ja->bucket_seq[idx]);
}
+#endif
ja->sectors_free = 0;
}
ja->dirty_idx = (ja->cur_idx + 1) % ja->nr;
out:
bch_verbose(c, "journal read done on device %s, ret %i", ca->name, ret);
- kvpfree(buf.data, buf.size);
+ kvfree(buf.data);
percpu_ref_put(&ca->io_ref);
closure_return(cl);
return;
goto out;
}
-void bch2_journal_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
- struct journal_replay *j)
-{
- unsigned i;
-
- for (i = 0; i < j->nr_ptrs; i++) {
- struct bch_dev *ca = bch_dev_bkey_exists(c, j->ptrs[i].dev);
- u64 offset;
-
- div64_u64_rem(j->ptrs[i].sector, ca->mi.bucket_size, &offset);
-
- if (i)
- prt_printf(out, " ");
- prt_printf(out, "%u:%u:%u (sector %llu)",
- j->ptrs[i].dev,
- j->ptrs[i].bucket,
- j->ptrs[i].bucket_offset,
- j->ptrs[i].sector);
- }
-}
-
int bch2_journal_read(struct bch_fs *c,
u64 *last_seq,
u64 *blacklist_seq,
struct journal_list jlist;
struct journal_replay *i, **_i, *prev = NULL;
struct genradix_iter radix_iter;
- struct bch_dev *ca;
- unsigned iter;
struct printbuf buf = PRINTBUF;
bool degraded = false, last_write_torn = false;
u64 seq;
jlist.last_seq = 0;
jlist.ret = 0;
- for_each_member_device(ca, c, iter) {
+ for_each_member_device(c, ca) {
if (!c->opts.fsck &&
!(bch2_dev_has_data(c, ca) & (1 << BCH_DATA_journal)))
continue;
.e.data_type = BCH_DATA_journal,
.e.nr_required = 1,
};
- unsigned ptr;
i = *_i;
if (!i || i->ignore)
continue;
- for (ptr = 0; ptr < i->nr_ptrs; ptr++) {
- ca = bch_dev_bkey_exists(c, i->ptrs[ptr].dev);
+ darray_for_each(i->ptrs, ptr) {
+ struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
- if (!i->ptrs[ptr].csum_good)
- bch_err_dev_offset(ca, i->ptrs[ptr].sector,
+ if (!ptr->csum_good)
+ bch_err_dev_offset(ca, ptr->sector,
"invalid journal checksum, seq %llu%s",
le64_to_cpu(i->j.seq),
i->csum_good ? " (had good copy on another device)" : "");
}
ret = jset_validate(c,
- bch_dev_bkey_exists(c, i->ptrs[0].dev),
+ bch_dev_bkey_exists(c, i->ptrs.data[0].dev),
&i->j,
- i->ptrs[0].sector,
+ i->ptrs.data[0].sector,
READ);
if (ret)
goto err;
- for (ptr = 0; ptr < i->nr_ptrs; ptr++)
- replicas.e.devs[replicas.e.nr_devs++] = i->ptrs[ptr].dev;
+ darray_for_each(i->ptrs, ptr)
+ replicas.e.devs[replicas.e.nr_devs++] = ptr->dev;
bch2_replicas_entry_sort(&replicas.e);
c->opts.foreground_target;
unsigned i, replicas = 0, replicas_want =
READ_ONCE(c->opts.metadata_replicas);
+ unsigned replicas_need = min_t(unsigned, replicas_want,
+ READ_ONCE(c->opts.metadata_replicas_required));
rcu_read_lock();
retry:
BUG_ON(bkey_val_u64s(&w->key.k) > BCH_REPLICAS_MAX);
- return replicas >= c->opts.metadata_replicas_required ? 0 : -EROFS;
+ return replicas >= replicas_need ? 0 : -EROFS;
}
static void journal_buf_realloc(struct journal *j, struct journal_buf *buf)
{
+ struct bch_fs *c = container_of(j, struct bch_fs, journal);
+
/* we aren't holding j->lock: */
unsigned new_size = READ_ONCE(j->buf_size_want);
void *new_buf;
if (buf->buf_size >= new_size)
return;
- new_buf = kvpmalloc(new_size, GFP_NOFS|__GFP_NOWARN);
+ size_t btree_write_buffer_size = new_size / 64;
+
+ if (bch2_btree_write_buffer_resize(c, btree_write_buffer_size))
+ return;
+
+ new_buf = kvmalloc(new_size, GFP_NOFS|__GFP_NOWARN);
if (!new_buf)
return;
swap(buf->buf_size, new_size);
spin_unlock(&j->lock);
- kvpfree(new_buf, new_size);
+ kvfree(new_buf);
}
static inline struct journal_buf *journal_last_unwritten_buf(struct journal *j)
return j->buf + (journal_last_unwritten_seq(j) & JOURNAL_BUF_MASK);
}
-static void journal_write_done(struct closure *cl)
+static CLOSURE_CALLBACK(journal_write_done)
{
- struct journal *j = container_of(cl, struct journal, io);
+ closure_type(w, struct journal_buf, io);
+ struct journal *j = container_of(w, struct journal, buf[w->idx]);
struct bch_fs *c = container_of(j, struct bch_fs, journal);
- struct journal_buf *w = journal_last_unwritten_buf(j);
struct bch_replicas_padded replicas;
union journal_res_state old, new;
- u64 v, seq;
+ u64 v, seq = le64_to_cpu(w->data->seq);
int err = 0;
- bch2_time_stats_update(!JSET_NO_FLUSH(w->data)
- ? j->flush_write_time
- : j->noflush_write_time, j->write_start_time);
+ time_stats_update(!JSET_NO_FLUSH(w->data)
+ ? j->flush_write_time
+ : j->noflush_write_time, j->write_start_time);
if (!w->devs_written.nr) {
bch_err(c, "unable to write journal to sufficient devices");
if (err)
bch2_fatal_error(c);
- spin_lock(&j->lock);
- seq = le64_to_cpu(w->data->seq);
+ closure_debug_destroy(cl);
+ spin_lock(&j->lock);
if (seq >= j->pin.front)
journal_seq_pin(j, seq)->devs = w->devs_written;
+ if (err && (!j->err_seq || seq < j->err_seq))
+ j->err_seq = seq;
+ w->write_done = true;
+
+ bool completed = false;
- if (!err) {
- if (!JSET_NO_FLUSH(w->data)) {
+ for (seq = journal_last_unwritten_seq(j);
+ seq <= journal_cur_seq(j);
+ seq++) {
+ w = j->buf + (seq & JOURNAL_BUF_MASK);
+ if (!w->write_done)
+ break;
+
+ if (!j->err_seq && !JSET_NO_FLUSH(w->data)) {
j->flushed_seq_ondisk = seq;
j->last_seq_ondisk = w->last_seq;
bch2_do_discards(c);
closure_wake_up(&c->freelist_wait);
-
bch2_reset_alloc_cursors(c);
}
- } else if (!j->err_seq || seq < j->err_seq)
- j->err_seq = seq;
- j->seq_ondisk = seq;
+ j->seq_ondisk = seq;
- /*
- * Updating last_seq_ondisk may let bch2_journal_reclaim_work() discard
- * more buckets:
- *
- * Must come before signaling write completion, for
- * bch2_fs_journal_stop():
- */
- if (j->watermark != BCH_WATERMARK_stripe)
- journal_reclaim_kick(&c->journal);
+ /*
+ * Updating last_seq_ondisk may let bch2_journal_reclaim_work() discard
+ * more buckets:
+ *
+ * Must come before signaling write completion, for
+ * bch2_fs_journal_stop():
+ */
+ if (j->watermark != BCH_WATERMARK_stripe)
+ journal_reclaim_kick(&c->journal);
- /* also must come before signalling write completion: */
- closure_debug_destroy(cl);
+ v = atomic64_read(&j->reservations.counter);
+ do {
+ old.v = new.v = v;
+ BUG_ON(journal_state_count(new, new.unwritten_idx));
+ BUG_ON(new.unwritten_idx != (seq & JOURNAL_BUF_MASK));
- v = atomic64_read(&j->reservations.counter);
- do {
- old.v = new.v = v;
- BUG_ON(journal_state_count(new, new.unwritten_idx));
+ new.unwritten_idx++;
+ } while ((v = atomic64_cmpxchg(&j->reservations.counter, old.v, new.v)) != old.v);
- new.unwritten_idx++;
- } while ((v = atomic64_cmpxchg(&j->reservations.counter,
- old.v, new.v)) != old.v);
+ completed = true;
+ }
- bch2_journal_space_available(j);
+ if (completed) {
+ bch2_journal_reclaim_fast(j);
+ bch2_journal_space_available(j);
- closure_wake_up(&w->wait);
- journal_wake(j);
+ track_event_change(&c->times[BCH_TIME_blocked_journal_max_in_flight], false);
- if (!journal_state_count(new, new.unwritten_idx) &&
- journal_last_unwritten_seq(j) <= journal_cur_seq(j)) {
- spin_unlock(&j->lock);
- closure_call(&j->io, bch2_journal_write, c->io_complete_wq, NULL);
- } else if (journal_last_unwritten_seq(j) == journal_cur_seq(j) &&
+ closure_wake_up(&w->wait);
+ journal_wake(j);
+ }
+
+ if (journal_last_unwritten_seq(j) == journal_cur_seq(j) &&
new.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL) {
struct journal_buf *buf = journal_cur_buf(j);
long delta = buf->expires - jiffies;
* previous entries still in flight - the current journal entry
* might want to be written now:
*/
-
- spin_unlock(&j->lock);
- mod_delayed_work(c->io_complete_wq, &j->write_work, max(0L, delta));
- } else {
- spin_unlock(&j->lock);
+ mod_delayed_work(j->wq, &j->write_work, max(0L, delta));
}
+
+ spin_unlock(&j->lock);
}
static void journal_write_endio(struct bio *bio)
{
- struct bch_dev *ca = bio->bi_private;
+ struct journal_bio *jbio = container_of(bio, struct journal_bio, bio);
+ struct bch_dev *ca = jbio->ca;
struct journal *j = &ca->fs->journal;
- struct journal_buf *w = journal_last_unwritten_buf(j);
- unsigned long flags;
+ struct journal_buf *w = j->buf + jbio->buf_idx;
if (bch2_dev_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_write,
"error writing journal entry %llu: %s",
le64_to_cpu(w->data->seq),
bch2_blk_status_to_str(bio->bi_status)) ||
bch2_meta_write_fault("journal")) {
+ unsigned long flags;
+
spin_lock_irqsave(&j->err_lock, flags);
bch2_dev_list_drop_dev(&w->devs_written, ca->dev_idx);
spin_unlock_irqrestore(&j->err_lock, flags);
}
- closure_put(&j->io);
+ closure_put(&w->io);
percpu_ref_put(&ca->io_ref);
}
-static void do_journal_write(struct closure *cl)
+static CLOSURE_CALLBACK(do_journal_write)
{
- struct journal *j = container_of(cl, struct journal, io);
+ closure_type(w, struct journal_buf, io);
+ struct journal *j = container_of(w, struct journal, buf[w->idx]);
struct bch_fs *c = container_of(j, struct bch_fs, journal);
- struct bch_dev *ca;
- struct journal_buf *w = journal_last_unwritten_buf(j);
- struct bch_extent_ptr *ptr;
- struct bio *bio;
unsigned sectors = vstruct_sectors(w->data, c->block_bits);
extent_for_each_ptr(bkey_i_to_s_extent(&w->key), ptr) {
- ca = bch_dev_bkey_exists(c, ptr->dev);
+ struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
+ struct journal_device *ja = &ca->journal;
+
if (!percpu_ref_tryget(&ca->io_ref)) {
/* XXX: fix this */
bch_err(c, "missing device for journal write\n");
this_cpu_add(ca->io_done->sectors[WRITE][BCH_DATA_journal],
sectors);
- bio = ca->journal.bio;
+ struct bio *bio = &ja->bio[w->idx]->bio;
bio_reset(bio, ca->disk_sb.bdev, REQ_OP_WRITE|REQ_SYNC|REQ_META);
bio->bi_iter.bi_sector = ptr->offset;
bio->bi_end_io = journal_write_endio;
trace_and_count(c, journal_write, bio);
closure_bio_submit(bio, cl);
- ca->journal.bucket_seq[ca->journal.cur_idx] =
- le64_to_cpu(w->data->seq);
+ ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
}
- continue_at(cl, journal_write_done, c->io_complete_wq);
+ continue_at(cl, journal_write_done, j->wq);
}
-static void bch2_journal_entries_postprocess(struct bch_fs *c, struct jset *jset)
+static int bch2_journal_write_prep(struct journal *j, struct journal_buf *w)
{
- struct jset_entry *i, *next, *prev = NULL;
+ struct bch_fs *c = container_of(j, struct bch_fs, journal);
+ struct jset_entry *start, *end;
+ struct jset *jset = w->data;
+ struct journal_keys_to_wb wb = { NULL };
+ unsigned sectors, bytes, u64s;
+ unsigned long btree_roots_have = 0;
+ bool validate_before_checksum = false;
+ u64 seq = le64_to_cpu(jset->seq);
+ int ret;
/*
* Simple compaction, dropping empty jset_entries (from journal
* If we wanted to be really fancy here, we could sort all the keys in
* the jset and drop keys that were overwritten - probably not worth it:
*/
- vstruct_for_each_safe(jset, i, next) {
+ vstruct_for_each(jset, i) {
unsigned u64s = le16_to_cpu(i->u64s);
/* Empty entry: */
if (!u64s)
continue;
- if (i->type == BCH_JSET_ENTRY_btree_root)
+ /*
+ * New btree roots are set by journalling them; when the journal
+ * entry gets written we have to propagate them to
+ * c->btree_roots
+ *
+ * But, every journal entry we write has to contain all the
+ * btree roots (at least for now); so after we copy btree roots
+ * to c->btree_roots we have to get any missing btree roots and
+ * add them to this journal entry:
+ */
+ switch (i->type) {
+ case BCH_JSET_ENTRY_btree_root:
bch2_journal_entry_to_btree_root(c, i);
-
- /* Can we merge with previous entry? */
- if (prev &&
- i->btree_id == prev->btree_id &&
- i->level == prev->level &&
- i->type == prev->type &&
- i->type == BCH_JSET_ENTRY_btree_keys &&
- le16_to_cpu(prev->u64s) + u64s <= U16_MAX) {
- memmove_u64s_down(vstruct_next(prev),
- i->_data,
- u64s);
- le16_add_cpu(&prev->u64s, u64s);
- continue;
+ __set_bit(i->btree_id, &btree_roots_have);
+ break;
+ case BCH_JSET_ENTRY_write_buffer_keys:
+ EBUG_ON(!w->need_flush_to_write_buffer);
+
+ if (!wb.wb)
+ bch2_journal_keys_to_write_buffer_start(c, &wb, seq);
+
+ struct bkey_i *k;
+ jset_entry_for_each_key(i, k) {
+ ret = bch2_journal_key_to_wb(c, &wb, i->btree_id, k);
+ if (ret) {
+ bch2_fs_fatal_error(c, "-ENOMEM flushing journal keys to btree write buffer");
+ bch2_journal_keys_to_write_buffer_end(c, &wb);
+ return ret;
+ }
+ }
+ i->type = BCH_JSET_ENTRY_btree_keys;
+ break;
}
-
- /* Couldn't merge, move i into new position (after prev): */
- prev = prev ? vstruct_next(prev) : jset->start;
- if (i != prev)
- memmove_u64s_down(prev, i, jset_u64s(u64s));
- }
-
- prev = prev ? vstruct_next(prev) : jset->start;
- jset->u64s = cpu_to_le32((u64 *) prev - jset->_data);
-}
-
-void bch2_journal_write(struct closure *cl)
-{
- struct journal *j = container_of(cl, struct journal, io);
- struct bch_fs *c = container_of(j, struct bch_fs, journal);
- struct bch_dev *ca;
- struct journal_buf *w = journal_last_unwritten_buf(j);
- struct bch_replicas_padded replicas;
- struct jset_entry *start, *end;
- struct jset *jset;
- struct bio *bio;
- struct printbuf journal_debug_buf = PRINTBUF;
- bool validate_before_checksum = false;
- unsigned i, sectors, bytes, u64s, nr_rw_members = 0;
- int ret;
-
- BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
-
- journal_buf_realloc(j, w);
- jset = w->data;
-
- j->write_start_time = local_clock();
-
- spin_lock(&j->lock);
-
- /*
- * If the journal is in an error state - we did an emergency shutdown -
- * we prefer to continue doing journal writes. We just mark them as
- * noflush so they'll never be used, but they'll still be visible by the
- * list_journal tool - this helps in debugging.
- *
- * There's a caveat: the first journal write after marking the
- * superblock dirty must always be a flush write, because on startup
- * from a clean shutdown we didn't necessarily read the journal and the
- * new journal write might overwrite whatever was in the journal
- * previously - we can't leave the journal without any flush writes in
- * it.
- *
- * So if we're in an error state, and we're still starting up, we don't
- * write anything at all.
- */
- if (!test_bit(JOURNAL_NEED_FLUSH_WRITE, &j->flags) &&
- (bch2_journal_error(j) ||
- w->noflush ||
- (!w->must_flush &&
- (jiffies - j->last_flush_write) < msecs_to_jiffies(c->opts.journal_flush_delay) &&
- test_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags)))) {
- w->noflush = true;
- SET_JSET_NO_FLUSH(jset, true);
- jset->last_seq = 0;
- w->last_seq = 0;
-
- j->nr_noflush_writes++;
- } else if (!bch2_journal_error(j)) {
- j->last_flush_write = jiffies;
- j->nr_flush_writes++;
- clear_bit(JOURNAL_NEED_FLUSH_WRITE, &j->flags);
- } else {
- spin_unlock(&j->lock);
- goto err;
}
- spin_unlock(&j->lock);
- /*
- * New btree roots are set by journalling them; when the journal entry
- * gets written we have to propagate them to c->btree_roots
- *
- * But, every journal entry we write has to contain all the btree roots
- * (at least for now); so after we copy btree roots to c->btree_roots we
- * have to get any missing btree roots and add them to this journal
- * entry:
- */
+ if (wb.wb)
+ bch2_journal_keys_to_write_buffer_end(c, &wb);
- bch2_journal_entries_postprocess(c, jset);
+ spin_lock(&c->journal.lock);
+ w->need_flush_to_write_buffer = false;
+ spin_unlock(&c->journal.lock);
start = end = vstruct_last(jset);
- end = bch2_btree_roots_to_journal_entries(c, jset->start, end);
+ end = bch2_btree_roots_to_journal_entries(c, end, btree_roots_have);
- bch2_journal_super_entries_add_common(c, &end,
- le64_to_cpu(jset->seq));
+ struct jset_entry_datetime *d =
+ container_of(jset_entry_init(&end, sizeof(*d)), struct jset_entry_datetime, entry);
+ d->entry.type = BCH_JSET_ENTRY_datetime;
+ d->seconds = cpu_to_le64(ktime_get_real_seconds());
+
+ bch2_journal_super_entries_add_common(c, &end, seq);
u64s = (u64 *) end - (u64 *) start;
BUG_ON(u64s > j->entry_u64s_reserved);
bch2_fs_fatal_error(c, "aieeee! journal write overran available space, %zu > %u (extra %u reserved %u/%u)",
vstruct_bytes(jset), w->sectors << 9,
u64s, w->u64s_reserved, j->entry_u64s_reserved);
- goto err;
+ return -EINVAL;
}
jset->magic = cpu_to_le64(jset_magic(c));
SET_JSET_CSUM_TYPE(jset, bch2_meta_checksum_type(c));
if (!JSET_NO_FLUSH(jset) && journal_entry_empty(jset))
- j->last_empty_seq = le64_to_cpu(jset->seq);
+ j->last_empty_seq = seq;
if (bch2_csum_type_is_encryption(JSET_CSUM_TYPE(jset)))
validate_before_checksum = true;
validate_before_checksum = true;
if (validate_before_checksum &&
- jset_validate(c, NULL, jset, 0, WRITE))
- goto err;
+ (ret = jset_validate(c, NULL, jset, 0, WRITE)))
+ return ret;
ret = bch2_encrypt(c, JSET_CSUM_TYPE(jset), journal_nonce(jset),
jset->encrypted_start,
vstruct_end(jset) - (void *) jset->encrypted_start);
if (bch2_fs_fatal_err_on(ret, c,
"error decrypting journal entry: %i", ret))
- goto err;
+ return ret;
jset->csum = csum_vstruct(c, JSET_CSUM_TYPE(jset),
journal_nonce(jset), jset);
if (!validate_before_checksum &&
- jset_validate(c, NULL, jset, 0, WRITE))
- goto err;
+ (ret = jset_validate(c, NULL, jset, 0, WRITE)))
+ return ret;
memset((void *) jset + bytes, 0, (sectors << 9) - bytes);
+ return 0;
+}
+
+static int bch2_journal_write_pick_flush(struct journal *j, struct journal_buf *w)
+{
+ struct bch_fs *c = container_of(j, struct bch_fs, journal);
+ int error = bch2_journal_error(j);
+
+ /*
+ * If the journal is in an error state - we did an emergency shutdown -
+ * we prefer to continue doing journal writes. We just mark them as
+ * noflush so they'll never be used, but they'll still be visible by the
+ * list_journal tool - this helps in debugging.
+ *
+ * There's a caveat: the first journal write after marking the
+ * superblock dirty must always be a flush write, because on startup
+ * from a clean shutdown we didn't necessarily read the journal and the
+ * new journal write might overwrite whatever was in the journal
+ * previously - we can't leave the journal without any flush writes in
+ * it.
+ *
+ * So if we're in an error state, and we're still starting up, we don't
+ * write anything at all.
+ */
+ if (error && test_bit(JOURNAL_NEED_FLUSH_WRITE, &j->flags))
+ return -EIO;
+
+ if (error ||
+ w->noflush ||
+ (!w->must_flush &&
+ (jiffies - j->last_flush_write) < msecs_to_jiffies(c->opts.journal_flush_delay) &&
+ test_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags))) {
+ w->noflush = true;
+ SET_JSET_NO_FLUSH(w->data, true);
+ w->data->last_seq = 0;
+ w->last_seq = 0;
+
+ j->nr_noflush_writes++;
+ } else {
+ j->last_flush_write = jiffies;
+ j->nr_flush_writes++;
+ clear_bit(JOURNAL_NEED_FLUSH_WRITE, &j->flags);
+ }
+
+ return 0;
+}
+
+CLOSURE_CALLBACK(bch2_journal_write)
+{
+ closure_type(w, struct journal_buf, io);
+ struct journal *j = container_of(w, struct journal, buf[w->idx]);
+ struct bch_fs *c = container_of(j, struct bch_fs, journal);
+ struct bch_replicas_padded replicas;
+ struct printbuf journal_debug_buf = PRINTBUF;
+ unsigned nr_rw_members = 0;
+ int ret;
+
+ for_each_rw_member(c, ca)
+ nr_rw_members++;
+
+ BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
+ BUG_ON(!w->write_started);
+ BUG_ON(w->write_allocated);
+ BUG_ON(w->write_done);
+
+ j->write_start_time = local_clock();
-retry_alloc:
spin_lock(&j->lock);
- ret = journal_write_alloc(j, w);
+ if (nr_rw_members > 1)
+ w->separate_flush = true;
+
+ ret = bch2_journal_write_pick_flush(j, w);
+ spin_unlock(&j->lock);
+ if (ret)
+ goto err;
+
+ mutex_lock(&j->buf_lock);
+ journal_buf_realloc(j, w);
+
+ ret = bch2_journal_write_prep(j, w);
+ mutex_unlock(&j->buf_lock);
+ if (ret)
+ goto err;
+
+ j->entry_bytes_written += vstruct_bytes(w->data);
+
+ while (1) {
+ spin_lock(&j->lock);
+ ret = journal_write_alloc(j, w);
+ if (!ret || !j->can_discard)
+ break;
- if (ret && j->can_discard) {
spin_unlock(&j->lock);
bch2_journal_do_discards(j);
- goto retry_alloc;
}
- if (ret)
+ if (ret) {
__bch2_journal_debug_to_text(&journal_debug_buf, j);
+ spin_unlock(&j->lock);
+ bch_err(c, "Unable to allocate journal write:\n%s",
+ journal_debug_buf.buf);
+ printbuf_exit(&journal_debug_buf);
+ goto err;
+ }
/*
* write is allocated, no longer need to account for it in
* bch2_journal_space_available():
*/
w->sectors = 0;
+ w->write_allocated = true;
/*
* journal entry has been compacted and allocated, recalculate space
* available:
*/
bch2_journal_space_available(j);
+ bch2_journal_do_writes(j);
spin_unlock(&j->lock);
- if (ret) {
- bch_err(c, "Unable to allocate journal write:\n%s",
- journal_debug_buf.buf);
- printbuf_exit(&journal_debug_buf);
- goto err;
- }
-
w->devs_written = bch2_bkey_devs(bkey_i_to_s_c(&w->key));
if (c->opts.nochanges)
goto no_io;
- for_each_rw_member(ca, c, i)
- nr_rw_members++;
-
- if (nr_rw_members > 1)
- w->separate_flush = true;
-
/*
* Mark journal replicas before we submit the write to guarantee
* recovery will find the journal entries after a crash.
if (ret)
goto err;
- if (!JSET_NO_FLUSH(jset) && w->separate_flush) {
- for_each_rw_member(ca, c, i) {
+ if (!JSET_NO_FLUSH(w->data))
+ closure_wait_event(&j->async_wait, j->seq_ondisk + 1 == le64_to_cpu(w->data->seq));
+
+ if (!JSET_NO_FLUSH(w->data) && w->separate_flush) {
+ for_each_rw_member(c, ca) {
percpu_ref_get(&ca->io_ref);
- bio = ca->journal.bio;
- bio_reset(bio, ca->disk_sb.bdev, REQ_OP_FLUSH);
+ struct journal_device *ja = &ca->journal;
+ struct bio *bio = &ja->bio[w->idx]->bio;
+ bio_reset(bio, ca->disk_sb.bdev,
+ REQ_OP_WRITE|REQ_SYNC|REQ_META|REQ_PREFLUSH);
bio->bi_end_io = journal_write_endio;
bio->bi_private = ca;
closure_bio_submit(bio, cl);
}
}
- continue_at(cl, do_journal_write, c->io_complete_wq);
+ continue_at(cl, do_journal_write, j->wq);
return;
no_io:
- continue_at(cl, journal_write_done, c->io_complete_wq);
+ continue_at(cl, journal_write_done, j->wq);
return;
err:
bch2_fatal_error(c);
- continue_at(cl, journal_write_done, c->io_complete_wq);
+ continue_at(cl, journal_write_done, j->wq);
}