bcachefs: bch2_btree_write_buffer_flush_going_ro()

The write buffer needs to be specifically flushed when going RO: keys in
the journal that haven't yet been moved to the write buffer don't have a
journal pin yet.

This fixes numerous syzbot bugs, all with symptoms of still doing writes
after we've got RO.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

diff --git a/fs/bcachefs/btree_write_buffer.c b/fs/bcachefs/btree_write_buffer.c
index 3f56b584f8ec2cd0627ace63166af71f96a942f2..1639c60dffa0e4166a155017608a565c9f5fd8ad 100644 (file)
--- a/fs/bcachefs/btree_write_buffer.c
+++ b/fs/bcachefs/btree_write_buffer.c
@@ -277,6 +277,10 @@ static int bch2_btree_write_buffer_flush_locked(struct btree_trans *trans)
        bool accounting_replay_done = test_bit(BCH_FS_accounting_replay_done, &c->flags);
        int ret = 0;
 
+       ret = bch2_journal_error(&c->journal);
+       if (ret)
+               return ret;
+
        bch2_trans_unlock(trans);
        bch2_trans_begin(trans);
 
@@ -491,7 +495,8 @@ static int fetch_wb_keys_from_journal(struct bch_fs *c, u64 seq)
        return ret;
 }
 
-static int btree_write_buffer_flush_seq(struct btree_trans *trans, u64 seq)
+static int btree_write_buffer_flush_seq(struct btree_trans *trans, u64 seq,
+                                       bool *did_work)
 {
        struct bch_fs *c = trans->c;
        struct btree_write_buffer *wb = &c->btree_write_buffer;
@@ -502,6 +507,8 @@ static int btree_write_buffer_flush_seq(struct btree_trans *trans, u64 seq)
 
                fetch_from_journal_err = fetch_wb_keys_from_journal(c, seq);
 
+               *did_work |= wb->inc.keys.nr || wb->flushing.keys.nr;
+
                /*
                 * On memory allocation failure, bch2_btree_write_buffer_flush_locked()
                 * is not guaranteed to empty wb->inc:
@@ -521,17 +528,34 @@ static int bch2_btree_write_buffer_journal_flush(struct journal *j,
                                struct journal_entry_pin *_pin, u64 seq)
 {
        struct bch_fs *c = container_of(j, struct bch_fs, journal);
+       bool did_work = false;
 
-       return bch2_trans_run(c, btree_write_buffer_flush_seq(trans, seq));
+       return bch2_trans_run(c, btree_write_buffer_flush_seq(trans, seq, &did_work));
 }
 
 int bch2_btree_write_buffer_flush_sync(struct btree_trans *trans)
 {
        struct bch_fs *c = trans->c;
+       bool did_work = false;
 
        trace_and_count(c, write_buffer_flush_sync, trans, _RET_IP_);
 
-       return btree_write_buffer_flush_seq(trans, journal_cur_seq(&c->journal));
+       return btree_write_buffer_flush_seq(trans, journal_cur_seq(&c->journal), &did_work);
+}
+
+/*
+ * The write buffer requires flushing when going RO: keys in the journal for the
+ * write buffer don't have a journal pin yet
+ */
+bool bch2_btree_write_buffer_flush_going_ro(struct bch_fs *c)
+{
+       if (bch2_journal_error(&c->journal))
+               return false;
+
+       bool did_work = false;
+       bch2_trans_run(c, btree_write_buffer_flush_seq(trans,
+                               journal_cur_seq(&c->journal), &did_work));
+       return did_work;
 }
 
 int bch2_btree_write_buffer_flush_nocheck_rw(struct btree_trans *trans)

diff --git a/fs/bcachefs/btree_write_buffer.h b/fs/bcachefs/btree_write_buffer.h
index 725e7965421688e9e07df2320cfd63d2c7453b04..d535cea28bde99bdbec5eb3b9bce48372b3bee9c 100644 (file)
--- a/fs/bcachefs/btree_write_buffer.h
+++ b/fs/bcachefs/btree_write_buffer.h
@@ -21,6 +21,7 @@ static inline bool bch2_btree_write_buffer_must_wait(struct bch_fs *c)
 
 struct btree_trans;
 int bch2_btree_write_buffer_flush_sync(struct btree_trans *);
+bool bch2_btree_write_buffer_flush_going_ro(struct bch_fs *);
 int bch2_btree_write_buffer_flush_nocheck_rw(struct btree_trans *);
 int bch2_btree_write_buffer_tryflush(struct btree_trans *);
 

diff --git a/fs/bcachefs/super.c b/fs/bcachefs/super.c
index 657fd3759e7b06bb352f25eb80036d59f43901f5..a6ed9a0bf1c707ae42aa5a061facf45dae108c9a 100644 (file)
--- a/fs/bcachefs/super.c
+++ b/fs/bcachefs/super.c
@@ -272,6 +272,7 @@ static void __bch2_fs_read_only(struct bch_fs *c)
                clean_passes++;
 
                if (bch2_btree_interior_updates_flush(c) ||
+                   bch2_btree_write_buffer_flush_going_ro(c) ||
                    bch2_journal_flush_all_pins(&c->journal) ||
                    bch2_btree_flush_all_writes(c) ||
                    seq != atomic64_read(&c->journal.seq)) {