X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Fjournal_io.c;h=65e58edb975a0259de3c6918dde1c9501c2c2da4;hb=93241a1c9a7dbef987f524a9dea3fc61e87e845b;hp=68113a08f1b9c4a42cab61b57277c9fb4bafe537;hpb=e0a51ccce8533a91c7cc0cd0adc5662697c9bcfa;p=bcachefs-tools-debian

diff --git a/libbcachefs/journal_io.c b/libbcachefs/journal_io.c
index 68113a0..65e58ed 100644
--- a/libbcachefs/journal_io.c
+++ b/libbcachefs/journal_io.c
@@ -4,18 +4,18 @@
 #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 "error.h"
-#include "io.h"
 #include "journal.h"
 #include "journal_io.h"
 #include "journal_reclaim.h"
 #include "journal_seq_blacklist.h"
 #include "replicas.h"
-
-#include <trace/events/bcachefs.h>
+#include "sb-clean.h"
+#include "trace.h"
 
 static struct nonce journal_nonce(const struct jset *jset)
 {
@@ -119,7 +119,7 @@ static int journal_entry_add(struct bch_fs *c, struct bch_dev *ca,
 				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
@@ -141,7 +141,8 @@ static int journal_entry_add(struct bch_fs *c, struct bch_dev *ca,
 		if (!dup->csum_good)
 			goto replace;
 
-		fsck_err(c, "found duplicate but non identical journal entries (seq %llu)",
+		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;
@@ -149,12 +150,12 @@ static int journal_entry_add(struct bch_fs *c, struct bch_dev *ca,
 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) {
@@ -209,38 +210,47 @@ static void journal_entry_null_range(void *start, void *end)
 #define JOURNAL_ENTRY_BAD	7
 
 static void journal_entry_err_msg(struct printbuf *out,
+				  u32 version,
 				  struct jset *jset,
 				  struct jset_entry *entry)
 {
-	prt_str(out, "invalid journal entry ");
-	if (entry)
-		prt_printf(out, "%s ", bch2_jset_entry_types[entry->type]);
-
-	if (!jset)
-		prt_printf(out, "in superblock");
-	else if (!entry)
-		prt_printf(out, "at seq %llu", le64_to_cpu(jset->seq));
-	else
-		prt_printf(out, "at offset %zi/%u seq %llu",
-			   (u64 *) entry - jset->_data,
-			   le32_to_cpu(jset->u64s),
-			   le64_to_cpu(jset->seq));
+	prt_str(out, "invalid journal entry, version=");
+	bch2_version_to_text(out, version);
+
+	if (entry) {
+		prt_str(out, " type=");
+		prt_str(out, bch2_jset_entry_types[entry->type]);
+	}
+
+	if (!jset) {
+		prt_printf(out, " in superblock");
+	} else {
+
+		prt_printf(out, " seq=%llu", le64_to_cpu(jset->seq));
+
+		if (entry)
+			prt_printf(out, " offset=%zi/%u",
+				   (u64 *) entry - jset->_data,
+				   le32_to_cpu(jset->u64s));
+	}
+
 	prt_str(out, ": ");
 }
 
-#define journal_entry_err(c, jset, entry, msg, ...)			\
+#define journal_entry_err(c, version, jset, entry, _err, msg, ...)	\
 ({									\
-	struct printbuf buf = PRINTBUF;					\
+	struct printbuf _buf = PRINTBUF;				\
 									\
-	journal_entry_err_msg(&buf, jset, entry);			\
-	prt_printf(&buf, msg, ##__VA_ARGS__);				\
+	journal_entry_err_msg(&_buf, version, jset, entry);		\
+	prt_printf(&_buf, msg, ##__VA_ARGS__);				\
 									\
-	switch (write) {						\
+	switch (flags & BKEY_INVALID_WRITE) {				\
 	case READ:							\
-		mustfix_fsck_err(c, "%s", buf.buf);			\
+		mustfix_fsck_err(c, _err, "%s", _buf.buf);		\
 		break;							\
 	case WRITE:							\
-		bch_err(c, "corrupt metadata before write: %s\n", buf.buf);\
+		bch2_sb_error_count(c, BCH_FSCK_ERR_##_err);		\
+		bch_err(c, "corrupt metadata before write: %s\n", _buf.buf);\
 		if (bch2_fs_inconsistent(c)) {				\
 			ret = -BCH_ERR_fsck_errors_not_fixed;		\
 			goto fsck_err;					\
@@ -248,12 +258,12 @@ static void journal_entry_err_msg(struct printbuf *out,
 		break;							\
 	}								\
 									\
-	printbuf_exit(&buf);						\
+	printbuf_exit(&_buf);						\
 	true;								\
 })
 
-#define journal_entry_err_on(cond, c, jset, entry, msg, ...)		\
-	((cond) ? journal_entry_err(c, jset, entry, msg, ##__VA_ARGS__) : false)
+#define journal_entry_err_on(cond, ...)					\
+	((cond) ? journal_entry_err(__VA_ARGS__) : false)
 
 #define FSCK_DELETED_KEY	5
 
@@ -262,13 +272,18 @@ static int journal_validate_key(struct bch_fs *c,
 				struct jset_entry *entry,
 				unsigned level, enum btree_id btree_id,
 				struct bkey_i *k,
-				unsigned version, int big_endian, int write)
+				unsigned version, int big_endian,
+				enum bkey_invalid_flags flags)
 {
+	int write = flags & BKEY_INVALID_WRITE;
 	void *next = vstruct_next(entry);
 	struct printbuf buf = PRINTBUF;
 	int ret = 0;
 
-	if (journal_entry_err_on(!k->k.u64s, c, jset, entry, "k->u64s 0")) {
+	if (journal_entry_err_on(!k->k.u64s,
+				 c, version, jset, entry,
+				 journal_entry_bkey_u64s_0,
+				 "k->u64s 0")) {
 		entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 		journal_entry_null_range(vstruct_next(entry), next);
 		return FSCK_DELETED_KEY;
@@ -276,7 +291,8 @@ static int journal_validate_key(struct bch_fs *c,
 
 	if (journal_entry_err_on((void *) bkey_next(k) >
 				 (void *) vstruct_next(entry),
-				 c, jset, entry,
+				 c, version, jset, entry,
+				 journal_entry_bkey_past_end,
 				 "extends past end of journal entry")) {
 		entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 		journal_entry_null_range(vstruct_next(entry), next);
@@ -284,7 +300,8 @@ static int journal_validate_key(struct bch_fs *c,
 	}
 
 	if (journal_entry_err_on(k->k.format != KEY_FORMAT_CURRENT,
-				 c, jset, entry,
+				 c, version, jset, entry,
+				 journal_entry_bkey_bad_format,
 				 "bad format %u", k->k.format)) {
 		le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 		memmove(k, bkey_next(k), next - (void *) bkey_next(k));
@@ -299,11 +316,7 @@ static int journal_validate_key(struct bch_fs *c,
 	if (bch2_bkey_invalid(c, bkey_i_to_s_c(k),
 			      __btree_node_type(level, btree_id), write, &buf)) {
 		printbuf_reset(&buf);
-		prt_printf(&buf, "invalid journal entry %s at offset %zi/%u seq %llu:",
-			   bch2_jset_entry_types[entry->type],
-			   (u64 *) entry - jset->_data,
-			   le32_to_cpu(jset->u64s),
-			   le64_to_cpu(jset->seq));
+		journal_entry_err_msg(&buf, version, jset, entry);
 		prt_newline(&buf);
 		printbuf_indent_add(&buf, 2);
 
@@ -312,7 +325,8 @@ static int journal_validate_key(struct bch_fs *c,
 		bch2_bkey_invalid(c, bkey_i_to_s_c(k),
 				  __btree_node_type(level, btree_id), write, &buf);
 
-		mustfix_fsck_err(c, "%s", buf.buf);
+		mustfix_fsck_err(c, journal_entry_bkey_invalid,
+				 "%s", buf.buf);
 
 		le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 		memmove(k, bkey_next(k), next - (void *) bkey_next(k));
@@ -331,9 +345,10 @@ fsck_err:
 }
 
 static int journal_entry_btree_keys_validate(struct bch_fs *c,
-					     struct jset *jset,
-					     struct jset_entry *entry,
-					     unsigned version, int big_endian, int write)
+				struct jset *jset,
+				struct jset_entry *entry,
+				unsigned version, int big_endian,
+				enum bkey_invalid_flags flags)
 {
 	struct bkey_i *k = entry->start;
 
@@ -341,7 +356,8 @@ static int journal_entry_btree_keys_validate(struct bch_fs *c,
 		int ret = journal_validate_key(c, jset, entry,
 					       entry->level,
 					       entry->btree_id,
-					       k, version, big_endian, write);
+					       k, version, big_endian,
+					       flags|BKEY_INVALID_JOURNAL);
 		if (ret == FSCK_DELETED_KEY)
 			continue;
 
@@ -357,28 +373,30 @@ static void journal_entry_btree_keys_to_text(struct printbuf *out, struct bch_fs
 	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]);
 		}
-		prt_printf(out, "btree=%s l=%u ", bch2_btree_ids[entry->btree_id], entry->level);
+		prt_printf(out, "btree=%s l=%u ", bch2_btree_id_str(entry->btree_id), entry->level);
 		bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(k));
 		first = false;
 	}
 }
 
 static int journal_entry_btree_root_validate(struct bch_fs *c,
-					     struct jset *jset,
-					     struct jset_entry *entry,
-					     unsigned version, int big_endian, int write)
+				struct jset *jset,
+				struct jset_entry *entry,
+				unsigned version, int big_endian,
+				enum bkey_invalid_flags flags)
 {
 	struct bkey_i *k = entry->start;
 	int ret = 0;
 
 	if (journal_entry_err_on(!entry->u64s ||
 				 le16_to_cpu(entry->u64s) != k->k.u64s,
-				 c, jset, entry,
+				 c, version, jset, entry,
+				 journal_entry_btree_root_bad_size,
 				 "invalid btree root journal entry: wrong number of keys")) {
 		void *next = vstruct_next(entry);
 		/*
@@ -392,7 +410,7 @@ static int journal_entry_btree_root_validate(struct bch_fs *c,
 	}
 
 	return journal_validate_key(c, jset, entry, 1, entry->btree_id, k,
-				    version, big_endian, write);
+				    version, big_endian, flags);
 fsck_err:
 	return ret;
 }
@@ -404,9 +422,10 @@ static void journal_entry_btree_root_to_text(struct printbuf *out, struct bch_fs
 }
 
 static int journal_entry_prio_ptrs_validate(struct bch_fs *c,
-					    struct jset *jset,
-					    struct jset_entry *entry,
-					    unsigned version, int big_endian, int write)
+				struct jset *jset,
+				struct jset_entry *entry,
+				unsigned version, int big_endian,
+				enum bkey_invalid_flags flags)
 {
 	/* obsolete, don't care: */
 	return 0;
@@ -418,14 +437,16 @@ static void journal_entry_prio_ptrs_to_text(struct printbuf *out, struct bch_fs
 }
 
 static int journal_entry_blacklist_validate(struct bch_fs *c,
-					    struct jset *jset,
-					    struct jset_entry *entry,
-					    unsigned version, int big_endian, int write)
+				struct jset *jset,
+				struct jset_entry *entry,
+				unsigned version, int big_endian,
+				enum bkey_invalid_flags flags)
 {
 	int ret = 0;
 
 	if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 1,
-				 c, jset, entry,
+				 c, version, jset, entry,
+				 journal_entry_blacklist_bad_size,
 		"invalid journal seq blacklist entry: bad size")) {
 		journal_entry_null_range(entry, vstruct_next(entry));
 	}
@@ -443,15 +464,17 @@ static void journal_entry_blacklist_to_text(struct printbuf *out, struct bch_fs
 }
 
 static int journal_entry_blacklist_v2_validate(struct bch_fs *c,
-					       struct jset *jset,
-					       struct jset_entry *entry,
-					       unsigned version, int big_endian, int write)
+				struct jset *jset,
+				struct jset_entry *entry,
+				unsigned version, int big_endian,
+				enum bkey_invalid_flags flags)
 {
 	struct jset_entry_blacklist_v2 *bl_entry;
 	int ret = 0;
 
 	if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 2,
-				 c, jset, entry,
+				 c, version, jset, entry,
+				 journal_entry_blacklist_v2_bad_size,
 		"invalid journal seq blacklist entry: bad size")) {
 		journal_entry_null_range(entry, vstruct_next(entry));
 		goto out;
@@ -461,7 +484,8 @@ static int journal_entry_blacklist_v2_validate(struct bch_fs *c,
 
 	if (journal_entry_err_on(le64_to_cpu(bl_entry->start) >
 				 le64_to_cpu(bl_entry->end),
-				 c, jset, entry,
+				 c, version, jset, entry,
+				 journal_entry_blacklist_v2_start_past_end,
 		"invalid journal seq blacklist entry: start > end")) {
 		journal_entry_null_range(entry, vstruct_next(entry));
 	}
@@ -482,9 +506,10 @@ static void journal_entry_blacklist_v2_to_text(struct printbuf *out, struct bch_
 }
 
 static int journal_entry_usage_validate(struct bch_fs *c,
-					struct jset *jset,
-					struct jset_entry *entry,
-					unsigned version, int big_endian, int write)
+				struct jset *jset,
+				struct jset_entry *entry,
+				unsigned version, int big_endian,
+				enum bkey_invalid_flags flags)
 {
 	struct jset_entry_usage *u =
 		container_of(entry, struct jset_entry_usage, entry);
@@ -492,7 +517,8 @@ static int journal_entry_usage_validate(struct bch_fs *c,
 	int ret = 0;
 
 	if (journal_entry_err_on(bytes < sizeof(*u),
-				 c, jset, entry,
+				 c, version, jset, entry,
+				 journal_entry_usage_bad_size,
 				 "invalid journal entry usage: bad size")) {
 		journal_entry_null_range(entry, vstruct_next(entry));
 		return ret;
@@ -514,24 +540,36 @@ static void journal_entry_usage_to_text(struct printbuf *out, struct bch_fs *c,
 }
 
 static int journal_entry_data_usage_validate(struct bch_fs *c,
-					struct jset *jset,
-					struct jset_entry *entry,
-					unsigned version, int big_endian, int write)
+				struct jset *jset,
+				struct jset_entry *entry,
+				unsigned version, int big_endian,
+				enum bkey_invalid_flags flags)
 {
 	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) ||
 				 bytes < sizeof(*u) + u->r.nr_devs,
-				 c, jset, entry,
+				 c, version, jset, entry,
+				 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;
 }
 
@@ -546,9 +584,10 @@ static void journal_entry_data_usage_to_text(struct printbuf *out, struct bch_fs
 }
 
 static int journal_entry_clock_validate(struct bch_fs *c,
-					struct jset *jset,
-					struct jset_entry *entry,
-					unsigned version, int big_endian, int write)
+				struct jset *jset,
+				struct jset_entry *entry,
+				unsigned version, int big_endian,
+				enum bkey_invalid_flags flags)
 {
 	struct jset_entry_clock *clock =
 		container_of(entry, struct jset_entry_clock, entry);
@@ -556,13 +595,17 @@ static int journal_entry_clock_validate(struct bch_fs *c,
 	int ret = 0;
 
 	if (journal_entry_err_on(bytes != sizeof(*clock),
-				 c, jset, entry, "bad size")) {
+				 c, version, jset, entry,
+				 journal_entry_clock_bad_size,
+				 "bad size")) {
 		journal_entry_null_range(entry, vstruct_next(entry));
 		return ret;
 	}
 
 	if (journal_entry_err_on(clock->rw > 1,
-				 c, jset, entry, "bad rw")) {
+				 c, version, jset, entry,
+				 journal_entry_clock_bad_rw,
+				 "bad rw")) {
 		journal_entry_null_range(entry, vstruct_next(entry));
 		return ret;
 	}
@@ -581,9 +624,10 @@ static void journal_entry_clock_to_text(struct printbuf *out, struct bch_fs *c,
 }
 
 static int journal_entry_dev_usage_validate(struct bch_fs *c,
-					    struct jset *jset,
-					    struct jset_entry *entry,
-					    unsigned version, int big_endian, int write)
+				struct jset *jset,
+				struct jset_entry *entry,
+				unsigned version, int big_endian,
+				enum bkey_invalid_flags flags)
 {
 	struct jset_entry_dev_usage *u =
 		container_of(entry, struct jset_entry_dev_usage, entry);
@@ -593,7 +637,9 @@ static int journal_entry_dev_usage_validate(struct bch_fs *c,
 	int ret = 0;
 
 	if (journal_entry_err_on(bytes < expected,
-				 c, jset, entry, "bad size (%u < %u)",
+				 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;
@@ -602,13 +648,17 @@ static int journal_entry_dev_usage_validate(struct bch_fs *c,
 	dev = le32_to_cpu(u->dev);
 
 	if (journal_entry_err_on(!bch2_dev_exists2(c, dev),
-				 c, jset, entry, "bad dev")) {
+				 c, version, jset, entry,
+				 journal_entry_dev_usage_bad_dev,
+				 "bad dev")) {
 		journal_entry_null_range(entry, vstruct_next(entry));
 		return ret;
 	}
 
 	if (journal_entry_err_on(u->pad,
-				 c, jset, entry, "bad pad")) {
+				 c, version, jset, entry,
+				 journal_entry_dev_usage_bad_pad,
+				 "bad pad")) {
 		journal_entry_null_range(entry, vstruct_next(entry));
 		return ret;
 	}
@@ -636,14 +686,13 @@ static void journal_entry_dev_usage_to_text(struct printbuf *out, struct bch_fs
 		       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,
-				      struct jset *jset,
-				      struct jset_entry *entry,
-				      unsigned version, int big_endian, int write)
+				struct jset *jset,
+				struct jset_entry *entry,
+				unsigned version, int big_endian,
+				enum bkey_invalid_flags flags)
 {
 	return 0;
 }
@@ -658,11 +707,13 @@ static void journal_entry_log_to_text(struct printbuf *out, struct bch_fs *c,
 }
 
 static int journal_entry_overwrite_validate(struct bch_fs *c,
-				      struct jset *jset,
-				      struct jset_entry *entry,
-				      unsigned version, int big_endian, int write)
+				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, 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,
@@ -671,9 +722,26 @@ static void journal_entry_overwrite_to_text(struct printbuf *out, struct bch_fs
 	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);
+}
+
 struct jset_entry_ops {
 	int (*validate)(struct bch_fs *, struct jset *,
-			struct jset_entry *, unsigned, int, int);
+			struct jset_entry *, unsigned, int,
+			enum bkey_invalid_flags);
 	void (*to_text)(struct printbuf *, struct bch_fs *, struct jset_entry *);
 };
 
@@ -690,11 +758,12 @@ static const struct jset_entry_ops bch2_jset_entry_ops[] = {
 int bch2_journal_entry_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
-				unsigned version, int big_endian, int write)
+				unsigned version, int big_endian,
+				enum bkey_invalid_flags flags)
 {
 	return entry->type < BCH_JSET_ENTRY_NR
 		? bch2_jset_entry_ops[entry->type].validate(c, jset, entry,
-				version, big_endian, write)
+				version, big_endian, flags)
 		: 0;
 }
 
@@ -710,22 +779,22 @@ void bch2_journal_entry_to_text(struct printbuf *out, struct bch_fs *c,
 }
 
 static int jset_validate_entries(struct bch_fs *c, struct jset *jset,
-				 int write)
+				 enum bkey_invalid_flags flags)
 {
-	struct jset_entry *entry;
+	unsigned version = le32_to_cpu(jset->version);
 	int ret = 0;
 
 	vstruct_for_each(jset, entry) {
-		if (journal_entry_err_on(vstruct_next(entry) >
-					 vstruct_last(jset), c, jset, entry,
+		if (journal_entry_err_on(vstruct_next(entry) > vstruct_last(jset),
+				c, version, jset, entry,
+				journal_entry_past_jset_end,
 				"journal entry extends past end of jset")) {
 			jset->u64s = cpu_to_le32((u64 *) entry - jset->_data);
 			break;
 		}
 
 		ret = bch2_journal_entry_validate(c, jset, entry,
-					le32_to_cpu(jset->version),
-					JSET_BIG_ENDIAN(jset), write);
+					version, JSET_BIG_ENDIAN(jset), flags);
 		if (ret)
 			break;
 	}
@@ -736,7 +805,7 @@ fsck_err:
 static int jset_validate(struct bch_fs *c,
 			 struct bch_dev *ca,
 			 struct jset *jset, u64 sector,
-			 int write)
+			 enum bkey_invalid_flags flags)
 {
 	unsigned version;
 	int ret = 0;
@@ -745,20 +814,21 @@ static int jset_validate(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, version, jset, NULL,
+			jset_unsupported_version,
+			"%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;
 	}
 
 	if (journal_entry_err_on(!bch2_checksum_type_valid(c, JSET_CSUM_TYPE(jset)),
-				 c, jset, NULL,
+			c, version, jset, NULL,
+			jset_unknown_csum,
 			"%s sector %llu seq %llu: journal entry with unknown csum type %llu",
 			ca ? ca->name : c->name,
 			sector, le64_to_cpu(jset->seq),
@@ -768,7 +838,8 @@ static int jset_validate(struct bch_fs *c,
 	/* last_seq is ignored when JSET_NO_FLUSH is true */
 	if (journal_entry_err_on(!JSET_NO_FLUSH(jset) &&
 				 le64_to_cpu(jset->last_seq) > le64_to_cpu(jset->seq),
-				 c, jset, NULL,
+				 c, version, jset, NULL,
+				 jset_last_seq_newer_than_seq,
 				 "invalid journal entry: last_seq > seq (%llu > %llu)",
 				 le64_to_cpu(jset->last_seq),
 				 le64_to_cpu(jset->seq))) {
@@ -776,7 +847,7 @@ static int jset_validate(struct bch_fs *c,
 		return JOURNAL_ENTRY_BAD;
 	}
 
-	ret = jset_validate_entries(c, jset, write);
+	ret = jset_validate_entries(c, jset, flags);
 fsck_err:
 	return ret;
 }
@@ -789,21 +860,21 @@ static int jset_validate_early(struct bch_fs *c,
 {
 	size_t bytes = vstruct_bytes(jset);
 	unsigned version;
-	int write = READ;
+	enum bkey_invalid_flags flags = BKEY_INVALID_JOURNAL;
 	int ret = 0;
 
 	if (le64_to_cpu(jset->magic) != jset_magic(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, version, jset, NULL,
+			jset_unsupported_version,
+			"%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;
 	}
@@ -813,7 +884,8 @@ static int jset_validate_early(struct bch_fs *c,
 		return JOURNAL_ENTRY_REREAD;
 
 	if (journal_entry_err_on(bytes > bucket_sectors_left << 9,
-				 c, jset, NULL,
+			c, version, jset, NULL,
+			jset_past_bucket_end,
 			"%s sector %llu seq %llu: journal entry too big (%zu bytes)",
 			ca ? ca->name : c->name,
 			sector, le64_to_cpu(jset->seq), bytes))
@@ -835,12 +907,12 @@ static int journal_read_buf_realloc(struct journal_read_buf *b,
 
 	/* 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;
@@ -882,7 +954,7 @@ reread:
 			ret = submit_bio_wait(bio);
 			kfree(bio);
 
-			if (bch2_dev_io_err_on(ret, ca,
+			if (bch2_dev_io_err_on(ret, ca, BCH_MEMBER_ERROR_read,
 					       "journal read error: sector %llu",
 					       offset) ||
 			    bch2_meta_read_fault("journal")) {
@@ -938,7 +1010,8 @@ reread:
 		ja->bucket_seq[bucket] = le64_to_cpu(j->seq);
 
 		csum_good = jset_csum_good(c, j);
-		if (!csum_good)
+		if (bch2_dev_io_err_on(!csum_good, ca, BCH_MEMBER_ERROR_checksum,
+				       "journal checksum error"))
 			saw_bad = true;
 
 		ret = bch2_encrypt(c, JSET_CSUM_TYPE(j), journal_nonce(j),
@@ -976,10 +1049,9 @@ next_block:
 	return 0;
 }
 
-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 =
@@ -987,7 +1059,6 @@ static void bch2_journal_read_device(struct closure *cl)
 	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;
 
@@ -1006,55 +1077,45 @@ static void bch2_journal_read_device(struct closure *cl)
 			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] &&
 	    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++) {
 			unsigned idx = (ja->cur_idx + ja->nr - 1 + i) % ja->nr;
+
 			bch_err(c, "bucket_seq[%u] = %llu", idx, ja->bucket_seq[idx]);
 		}
+#endif
 		ja->sectors_free = 0;
 	}
 
@@ -1099,17 +1160,17 @@ void bch2_journal_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
 	}
 }
 
-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 genradix_iter radix_iter;
-	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);
@@ -1117,7 +1178,7 @@ int bch2_journal_read(struct bch_fs *c, u64 *blacklist_seq, u64 *start_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;
@@ -1138,41 +1199,47 @@ int bch2_journal_read(struct bch_fs *c, u64 *blacklist_seq, u64 *start_seq)
 	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) {
+		enum bkey_invalid_flags flags = BKEY_INVALID_JOURNAL;
+
 		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;
-
-			pr_info("last flush %llu-%llu csum good %u",
-				le64_to_cpu(i->j.last_seq),
-				le64_to_cpu(i->j.seq),
-				i->csum_good);
-
-			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, le32_to_cpu(i->j.version), &i->j, NULL,
+					 jset_last_seq_newer_than_seq,
+					 "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) {
@@ -1180,12 +1247,19 @@ int bch2_journal_read(struct bch_fs *c, u64 *blacklist_seq, u64 *start_seq)
 		return 0;
 	}
 
-	if (!last_seq) {
-		fsck_err(c, "journal read done, but no entries found after dropping non-flushes");
-		ret = -1;
-		goto err;
+	if (!*last_seq) {
+		fsck_err(c, dirty_but_no_journal_entries_post_drop_nonflushes,
+			 "journal read done, but no entries found after dropping non-flushes");
+		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;
@@ -1194,21 +1268,21 @@ int bch2_journal_read(struct bch_fs *c, u64 *blacklist_seq, u64 *start_seq)
 			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,
+				    jset_seq_blacklisted,
 				    "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;
 
@@ -1242,11 +1316,12 @@ int bch2_journal_read(struct bch_fs *c, u64 *blacklist_seq, u64 *start_seq)
 			bch2_journal_ptrs_to_text(&buf2, c, i);
 
 			missing_end = seq - 1;
-			fsck_err(c, "journal entries %llu-%llu missing! (replaying %llu-%llu)\n"
+			fsck_err(c, journal_entries_missing,
+				 "journal entries %llu-%llu missing! (replaying %llu-%llu)\n"
 				 "  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);
@@ -1258,8 +1333,6 @@ int bch2_journal_read(struct bch_fs *c, u64 *blacklist_seq, u64 *start_seq)
 	}
 
 	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,
@@ -1274,10 +1347,10 @@ int bch2_journal_read(struct bch_fs *c, u64 *blacklist_seq, u64 *start_seq)
 			struct bch_dev *ca = bch_dev_bkey_exists(c, i->ptrs[ptr].dev);
 
 			if (!i->ptrs[ptr].csum_good)
-				printk(KERN_ERR "bcachefs (%s) sector %llu: invalid journal checksum, seq %llu%s\n",
-				       ca->name, i->ptrs[ptr].sector,
-				       le64_to_cpu(i->j.seq),
-				       i->csum_good ? " (had good copy on another device)" : "");
+				bch_err_dev_offset(ca, i->ptrs[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,
@@ -1293,34 +1366,20 @@ int bch2_journal_read(struct bch_fs *c, u64 *blacklist_seq, u64 *start_seq)
 
 		bch2_replicas_entry_sort(&replicas.e);
 
-		/*
-		 * If we're mounting in degraded mode - if we didn't read all
-		 * the devices - this is wrong:
-		 */
-
 		printbuf_reset(&buf);
 		bch2_replicas_entry_to_text(&buf, &replicas.e);
 
 		if (!degraded &&
-		    fsck_err_on(!bch2_replicas_marked(c, &replicas.e), c,
-				"superblock not marked as containing replicas %s",
-				buf.buf)) {
+		    !bch2_replicas_marked(c, &replicas.e) &&
+		    (le64_to_cpu(i->j.seq) == *last_seq ||
+		     fsck_err(c, journal_entry_replicas_not_marked,
+			      "superblock not marked as containing replicas for journal entry %llu\n  %s",
+			      le64_to_cpu(i->j.seq), buf.buf))) {
 			ret = bch2_mark_replicas(c, &replicas.e);
 			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);
@@ -1358,8 +1417,7 @@ static void __journal_write_alloc(struct journal *j,
 		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;
 
@@ -1385,16 +1443,21 @@ static void __journal_write_alloc(struct journal *j,
 }
 
 /**
- * journal_next_bucket - move on to the next journal bucket if possible
+ * journal_write_alloc - decide where to write next journal entry
+ *
+ * @j:		journal object
+ * @w:		journal buf (entry to be written)
+ *
+ * Returns: 0 on success, or -EROFS on failure
  */
-static int journal_write_alloc(struct journal *j, struct journal_buf *w,
-			       unsigned sectors)
+static int journal_write_alloc(struct journal *j, struct journal_buf *w)
 {
 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
 	struct bch_devs_mask devs;
 	struct journal_device *ja;
 	struct bch_dev *ca;
 	struct dev_alloc_list devs_sorted;
+	unsigned sectors = vstruct_sectors(w->data, c->block_bits);
 	unsigned target = c->opts.metadata_target ?:
 		c->opts.foreground_target;
 	unsigned i, replicas = 0, replicas_want =
@@ -1452,6 +1515,8 @@ done:
 
 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;
@@ -1459,7 +1524,12 @@ static void journal_buf_realloc(struct journal *j, struct journal_buf *buf)
 	if (buf->buf_size >= new_size)
 		return;
 
-	new_buf = kvpmalloc(new_size, GFP_NOIO|__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 = kvpmalloc(new_size, GFP_NOFS|__GFP_NOWARN);
 	if (!new_buf)
 		return;
 
@@ -1478,9 +1548,9 @@ 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(j, 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;
@@ -1518,6 +1588,8 @@ static void journal_write_done(struct closure *cl)
 
 			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;
@@ -1531,7 +1603,7 @@ static void journal_write_done(struct closure *cl)
 	 * 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: */
@@ -1546,13 +1618,18 @@ static void journal_write_done(struct closure *cl)
 	} while ((v = atomic64_cmpxchg(&j->reservations.counter,
 				       old.v, new.v)) != old.v);
 
+	bch2_journal_reclaim_fast(j);
 	bch2_journal_space_available(j);
 
+	track_event_change(&c->times[BCH_TIME_blocked_journal_max_in_flight],
+			   &j->max_in_flight_start, false);
+
 	closure_wake_up(&w->wait);
 	journal_wake(j);
 
 	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) &&
 		   new.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL) {
@@ -1565,10 +1642,11 @@ static void journal_write_done(struct closure *cl)
 		 * 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);
 	}
-
-	spin_unlock(&j->lock);
 }
 
 static void journal_write_endio(struct bio *bio)
@@ -1578,7 +1656,8 @@ static void journal_write_endio(struct bio *bio)
 	struct journal_buf *w = journal_last_unwritten_buf(j);
 	unsigned long flags;
 
-	if (bch2_dev_io_err_on(bio->bi_status, ca, "error writing journal entry %llu: %s",
+	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")) {
@@ -1591,9 +1670,9 @@ static void journal_write_endio(struct bio *bio)
 	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(j, 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);
@@ -1636,82 +1715,102 @@ static void do_journal_write(struct closure *cl)
 	}
 
 	continue_at(cl, journal_write_done, c->io_complete_wq);
-	return;
 }
 
-void bch2_journal_write(struct closure *cl)
+static int bch2_journal_write_prep(struct journal *j, struct journal_buf *w)
 {
-	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 jset_entry *start, *end;
-	struct jset *jset;
-	struct bio *bio;
-	struct printbuf journal_debug_buf = PRINTBUF;
+	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;
-	unsigned i, sectors, bytes, u64s, nr_rw_members = 0;
+	u64 seq = le64_to_cpu(jset->seq);
 	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 (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 {
-		j->last_flush_write = jiffies;
-		j->nr_flush_writes++;
-	}
-	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
+	 * Simple compaction, dropping empty jset_entries (from journal
+	 * reservations that weren't fully used) and merging jset_entries that
+	 * can be.
 	 *
-	 * 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 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(jset, i) {
+		unsigned u64s = le16_to_cpu(i->u64s);
+
+		/* Empty entry: */
+		if (!u64s)
+			continue;
+
+		/*
+		 * 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);
+			__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;
+		}
+	}
 
-	bch2_journal_entries_to_btree_roots(c, jset);
+	if (wb.wb)
+		bch2_journal_keys_to_write_buffer_end(c, &wb);
+	w->need_flush_to_write_buffer = false;
 
 	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));
+	bch2_journal_super_entries_add_common(c, &end, seq);
 	u64s	= (u64 *) end - (u64 *) start;
 	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);
+		return -EINVAL;
+	}
 
 	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));
 
 	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;
@@ -1720,41 +1819,120 @@ void bch2_journal_write(struct closure *cl)
 		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;
-
-	sectors = vstruct_sectors(jset, c->block_bits);
-	BUG_ON(sectors > w->sectors);
+	    (ret = jset_validate(c, NULL, jset, 0, WRITE)))
+		return ret;
 
-	bytes = vstruct_bytes(jset);
 	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(j, 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;
+	struct bio *bio;
+	struct printbuf journal_debug_buf = PRINTBUF;
+	unsigned nr_rw_members = 0;
+	int ret;
+
+	BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
+
+	j->write_start_time = local_clock();
 
-retry_alloc:
 	spin_lock(&j->lock);
-	ret = journal_write_alloc(j, w, sectors);
+	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
@@ -1769,28 +1947,29 @@ retry_alloc:
 	bch2_journal_space_available(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);
-		bch2_fatal_error(c);
-		continue_at(cl, journal_write_done, c->io_complete_wq);
-		return;
-	}
-
 	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)
+	for_each_rw_member(c, ca)
 		nr_rw_members++;
 
 	if (nr_rw_members > 1)
 		w->separate_flush = true;
 
-	if (!JSET_NO_FLUSH(jset) && w->separate_flush) {
-		for_each_rw_member(ca, c, i) {
+	/*
+	 * 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(w->data) && w->separate_flush) {
+		for_each_rw_member(c, ca) {
 			percpu_ref_get(&ca->io_ref);
 
 			bio = ca->journal.bio;