#include "journal_reclaim.h"
#include "journal_seq_blacklist.h"
#include "replicas.h"
-
-#include <trace/events/bcachefs.h>
+#include "trace.h"
static struct nonce journal_nonce(const struct jset *jset)
{
journal_entry_radix_idx(c, le64_to_cpu(j->seq)),
GFP_KERNEL);
if (!_i)
- return -ENOMEM;
+ return -BCH_ERR_ENOMEM_journal_entry_add;
/*
* Duplicate journal entries? If so we want the one that didn't have a
replace:
i = kvpmalloc(offsetof(struct journal_replay, j) + bytes, GFP_KERNEL);
if (!i)
- return -ENOMEM;
+ return -BCH_ERR_ENOMEM_journal_entry_add;
i->nr_ptrs = 0;
i->csum_good = entry_ptr.csum_good;
i->ignore = false;
- memcpy(&i->j, j, bytes);
+ unsafe_memcpy(&i->j, j, bytes, "embedded variable length struct");
i->ptrs[i->nr_ptrs++] = entry_ptr;
if (dup) {
int ret = journal_validate_key(c, jset, entry,
entry->level,
entry->btree_id,
- k, version, big_endian, write);
+ k, version, big_endian,
+ write|BKEY_INVALID_JOURNAL);
if (ret == FSCK_DELETED_KEY)
continue;
struct bkey_i *k;
bool first = true;
- vstruct_for_each(entry, k) {
+ jset_entry_for_each_key(entry, k) {
if (!first) {
prt_newline(out);
prt_printf(out, "%s: ", bch2_jset_entry_types[entry->type]);
struct jset_entry *entry,
unsigned version, int big_endian, int write)
{
- return journal_entry_btree_keys_validate(c, jset, entry, version, big_endian, write);
+ return journal_entry_btree_keys_validate(c, jset, entry,
+ version, big_endian, READ);
}
static void journal_entry_overwrite_to_text(struct printbuf *out, struct bch_fs *c,
return JOURNAL_ENTRY_NONE;
version = le32_to_cpu(jset->version);
- if (journal_entry_err_on((version != BCH_JSET_VERSION_OLD &&
- version < bcachefs_metadata_version_min) ||
- version >= bcachefs_metadata_version_max,
- c, jset, NULL,
- "%s sector %llu seq %llu: unknown journal entry version %u",
+ if (journal_entry_err_on(!bch2_version_compatible(version), c, jset, NULL,
+ "%s sector %llu seq %llu: incompatible journal entry version %u.%u",
ca ? ca->name : c->name,
sector, le64_to_cpu(jset->seq),
- version)) {
+ BCH_VERSION_MAJOR(version),
+ BCH_VERSION_MINOR(version))) {
/* don't try to continue: */
return -EINVAL;
}
return JOURNAL_ENTRY_NONE;
version = le32_to_cpu(jset->version);
- if (journal_entry_err_on((version != BCH_JSET_VERSION_OLD &&
- version < bcachefs_metadata_version_min) ||
- version >= bcachefs_metadata_version_max,
- c, jset, NULL,
- "%s sector %llu seq %llu: unknown journal entry version %u",
+ if (journal_entry_err_on(!bch2_version_compatible(version), c, jset, NULL,
+ "%s sector %llu seq %llu: unknown journal entry version %u.%u",
ca ? ca->name : c->name,
sector, le64_to_cpu(jset->seq),
- version)) {
+ BCH_VERSION_MAJOR(version),
+ BCH_VERSION_MINOR(version))) {
/* don't try to continue: */
return -EINVAL;
}
/* the bios are sized for this many pages, max: */
if (new_size > JOURNAL_ENTRY_SIZE_MAX)
- return -ENOMEM;
+ return -BCH_ERR_ENOMEM_journal_read_buf_realloc;
new_size = roundup_pow_of_two(new_size);
n = kvpmalloc(new_size, GFP_KERNEL);
if (!n)
- return -ENOMEM;
+ return -BCH_ERR_ENOMEM_journal_read_buf_realloc;
kvpfree(b->data, b->size);
b->data = n;
struct journal_replay *r, **_r;
struct genradix_iter iter;
struct journal_read_buf buf = { NULL, 0 };
- u64 min_seq = U64_MAX;
unsigned i;
int ret = 0;
goto err;
}
- /* Find the journal bucket with the highest sequence number: */
- for (i = 0; i < ja->nr; i++) {
- if (ja->bucket_seq[i] > ja->bucket_seq[ja->cur_idx])
- ja->cur_idx = i;
-
- min_seq = min(ja->bucket_seq[i], min_seq);
- }
-
- /*
- * If there's duplicate journal entries in multiple buckets (which
- * definitely isn't supposed to happen, but...) - make sure to start
- * cur_idx at the last of those buckets, so we don't deadlock trying to
- * allocate
- */
- while (ja->bucket_seq[ja->cur_idx] > min_seq &&
- ja->bucket_seq[ja->cur_idx] ==
- ja->bucket_seq[(ja->cur_idx + 1) % ja->nr])
- ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
-
ja->sectors_free = ca->mi.bucket_size;
mutex_lock(&jlist->lock);
- genradix_for_each(&c->journal_entries, iter, _r) {
+ genradix_for_each_reverse(&c->journal_entries, iter, _r) {
r = *_r;
if (!r)
continue;
for (i = 0; i < r->nr_ptrs; i++) {
- if (r->ptrs[i].dev == ca->dev_idx &&
- sector_to_bucket(ca, r->ptrs[i].sector) == ja->buckets[ja->cur_idx]) {
+ if (r->ptrs[i].dev == ca->dev_idx) {
unsigned wrote = bucket_remainder(ca, r->ptrs[i].sector) +
vstruct_sectors(&r->j, c->block_bits);
- ja->sectors_free = min(ja->sectors_free,
- ca->mi.bucket_size - wrote);
+ ja->cur_idx = r->ptrs[i].bucket;
+ ja->sectors_free = ca->mi.bucket_size - wrote;
+ goto found;
}
}
}
+found:
mutex_unlock(&jlist->lock);
if (ja->bucket_seq[ja->cur_idx] &&
}
}
-int bch2_journal_read(struct bch_fs *c, u64 *blacklist_seq, u64 *start_seq)
+int bch2_journal_read(struct bch_fs *c,
+ u64 *last_seq,
+ u64 *blacklist_seq,
+ u64 *start_seq)
{
struct journal_list jlist;
struct journal_replay *i, **_i, *prev = NULL;
struct bch_dev *ca;
unsigned iter;
struct printbuf buf = PRINTBUF;
- size_t keys = 0, entries = 0;
- bool degraded = false;
- u64 seq, last_seq = 0;
+ bool degraded = false, last_write_torn = false;
+ u64 seq;
int ret = 0;
closure_init_stack(&jlist.cl);
if (jlist.ret)
return jlist.ret;
- *start_seq = 0;
+ *last_seq = 0;
+ *start_seq = 0;
+ *blacklist_seq = 0;
/*
* Find most recent flush entry, and ignore newer non flush entries -
* those entries will be blacklisted:
*/
genradix_for_each_reverse(&c->journal_entries, radix_iter, _i) {
+ int write = READ;
+
i = *_i;
if (!i || i->ignore)
continue;
if (!*start_seq)
- *start_seq = le64_to_cpu(i->j.seq) + 1;
-
- if (!JSET_NO_FLUSH(&i->j)) {
- int write = READ;
- if (journal_entry_err_on(le64_to_cpu(i->j.last_seq) > le64_to_cpu(i->j.seq),
- c, &i->j, NULL,
- "invalid journal entry: last_seq > seq (%llu > %llu)",
- le64_to_cpu(i->j.last_seq),
- le64_to_cpu(i->j.seq)))
- i->j.last_seq = i->j.seq;
-
- last_seq = le64_to_cpu(i->j.last_seq);
- *blacklist_seq = le64_to_cpu(i->j.seq) + 1;
- break;
+ *blacklist_seq = *start_seq = le64_to_cpu(i->j.seq) + 1;
+
+ if (JSET_NO_FLUSH(&i->j)) {
+ i->ignore = true;
+ continue;
}
- journal_replay_free(c, i);
+ if (!last_write_torn && !i->csum_good) {
+ last_write_torn = true;
+ i->ignore = true;
+ continue;
+ }
+
+ if (journal_entry_err_on(le64_to_cpu(i->j.last_seq) > le64_to_cpu(i->j.seq),
+ c, &i->j, NULL,
+ "invalid journal entry: last_seq > seq (%llu > %llu)",
+ le64_to_cpu(i->j.last_seq),
+ le64_to_cpu(i->j.seq)))
+ i->j.last_seq = i->j.seq;
+
+ *last_seq = le64_to_cpu(i->j.last_seq);
+ *blacklist_seq = le64_to_cpu(i->j.seq) + 1;
+ break;
}
if (!*start_seq) {
return 0;
}
- if (!last_seq) {
+ if (!*last_seq) {
fsck_err(c, "journal read done, but no entries found after dropping non-flushes");
- ret = -1;
- goto err;
+ return 0;
}
+ bch_info(c, "journal read done, replaying entries %llu-%llu",
+ *last_seq, *blacklist_seq - 1);
+
+ if (*start_seq != *blacklist_seq)
+ bch_info(c, "dropped unflushed entries %llu-%llu",
+ *blacklist_seq, *start_seq - 1);
+
/* Drop blacklisted entries and entries older than last_seq: */
genradix_for_each(&c->journal_entries, radix_iter, _i) {
i = *_i;
continue;
seq = le64_to_cpu(i->j.seq);
- if (seq < last_seq) {
+ if (seq < *last_seq) {
journal_replay_free(c, i);
continue;
}
if (bch2_journal_seq_is_blacklisted(c, seq, true)) {
fsck_err_on(!JSET_NO_FLUSH(&i->j), c,
"found blacklisted journal entry %llu", seq);
-
- journal_replay_free(c, i);
+ i->ignore = true;
}
}
/* Check for missing entries: */
- seq = last_seq;
+ seq = *last_seq;
genradix_for_each(&c->journal_entries, radix_iter, _i) {
i = *_i;
" prev at %s\n"
" next at %s",
missing_start, missing_end,
- last_seq, *blacklist_seq - 1,
+ *last_seq, *blacklist_seq - 1,
buf1.buf, buf2.buf);
printbuf_exit(&buf1);
}
genradix_for_each(&c->journal_entries, radix_iter, _i) {
- struct jset_entry *entry;
- struct bkey_i *k, *_n;
struct bch_replicas_padded replicas = {
.e.data_type = BCH_DATA_journal,
.e.nr_required = 1,
if (ret)
goto err;
}
-
- for_each_jset_key(k, _n, entry, &i->j)
- keys++;
- entries++;
}
-
- bch_info(c, "journal read done, %zu keys in %zu entries, seq %llu",
- keys, entries, *start_seq);
-
- if (*start_seq != *blacklist_seq)
- bch_info(c, "dropped unflushed entries %llu-%llu",
- *blacklist_seq, *start_seq - 1);
err:
fsck_err:
printbuf_exit(&buf);
if (!ca->mi.durability ||
ca->mi.state != BCH_MEMBER_STATE_rw ||
!ja->nr ||
- bch2_bkey_has_device(bkey_i_to_s_c(&w->key),
- ca->dev_idx) ||
+ bch2_bkey_has_device_c(bkey_i_to_s_c(&w->key), ca->dev_idx) ||
sectors > ja->sectors_free)
continue;
if (buf->buf_size >= new_size)
return;
- new_buf = kvpmalloc(new_size, GFP_NOIO|__GFP_NOWARN);
+ new_buf = kvpmalloc(new_size, GFP_NOFS|__GFP_NOWARN);
if (!new_buf)
return;
struct journal *j = container_of(cl, struct journal, io);
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;
int err = 0;
if (!w->devs_written.nr) {
bch_err(c, "unable to write journal to sufficient devices");
err = -EIO;
- } else {
- bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
- w->devs_written);
- if (bch2_mark_replicas(c, &replicas.e))
- err = -EIO;
}
-
if (err)
bch2_fatal_error(c);
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;
* Must come before signaling write completion, for
* bch2_fs_journal_stop():
*/
- if (j->watermark)
+ if (j->watermark != BCH_WATERMARK_stripe)
journal_reclaim_kick(&c->journal);
/* also must come before signalling write completion: */
return;
}
+static void bch2_journal_entries_postprocess(struct bch_fs *c, struct jset *jset)
+{
+ struct jset_entry *i, *next, *prev = NULL;
+
+ /*
+ * Simple compaction, dropping empty jset_entries (from journal
+ * reservations that weren't fully used) and merging jset_entries that
+ * can be.
+ *
+ * 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) {
+ unsigned u64s = le16_to_cpu(i->u64s);
+
+ /* Empty entry: */
+ if (!u64s)
+ continue;
+
+ if (i->type == 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;
+ }
+
+ /* 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;
j->write_start_time = local_clock();
spin_lock(&j->lock);
- if (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))) {
+
+ /*
+ * 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 {
+ } 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);
* entry:
*/
- bch2_journal_entries_to_btree_roots(c, jset);
+ bch2_journal_entries_postprocess(c, jset);
start = end = vstruct_last(jset);
BUG_ON(u64s > j->entry_u64s_reserved);
le32_add_cpu(&jset->u64s, u64s);
- BUG_ON(vstruct_sectors(jset, c->block_bits) > w->sectors);
+
+ sectors = vstruct_sectors(jset, c->block_bits);
+ bytes = vstruct_bytes(jset);
+
+ if (sectors > w->sectors) {
+ 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;
+ }
jset->magic = cpu_to_le64(jset_magic(c));
- jset->version = c->sb.version < bcachefs_metadata_version_bkey_renumber
- ? cpu_to_le32(BCH_JSET_VERSION_OLD)
- : cpu_to_le32(c->sb.version);
+ jset->version = cpu_to_le32(c->sb.version);
SET_JSET_BIG_ENDIAN(jset, CPU_BIG_ENDIAN);
SET_JSET_CSUM_TYPE(jset, bch2_meta_checksum_type(c));
jset_validate(c, NULL, jset, 0, WRITE))
goto err;
- sectors = vstruct_sectors(jset, c->block_bits);
- BUG_ON(sectors > w->sectors);
-
- bytes = vstruct_bytes(jset);
memset((void *) jset + bytes, 0, (sectors << 9) - bytes);
retry_alloc:
bch_err(c, "Unable to allocate journal write:\n%s",
journal_debug_buf.buf);
printbuf_exit(&journal_debug_buf);
- bch2_fatal_error(c);
- continue_at(cl, journal_write_done, c->io_complete_wq);
- return;
+ goto err;
}
w->devs_written = bch2_bkey_devs(bkey_i_to_s_c(&w->key));
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.
+ */
+ bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
+ w->devs_written);
+ ret = bch2_mark_replicas(c, &replicas.e);
+ if (ret)
+ goto err;
+
if (!JSET_NO_FLUSH(jset) && w->separate_flush) {
for_each_rw_member(ca, c, i) {
percpu_ref_get(&ca->io_ref);