When a sector mode namespace is initially created, the arena's err_lock
is not initialized. If, on the other hand, the namespace already
exists, the mutex is initialized. To fix the issue, I moved the mutex
initialization into the arena_alloc, which is called by both
discover_arenas and create_arenas.
This was discovered on an older kernel where mutex_trylock checks the
count to determine whether the lock is held. Because the data structure
is kzalloc-d, that count was 0 (held), and I/O to the device would hang
forever waiting for the lock to be released (see btt_write_pg, for
example). Current kernels have a different mutex implementation that
checks for a non-null owner, and so this doesn't show up as a problem.
If that lock were ever contended, it might cause issues, but you'd have
to be really unlucky, I think.
Signed-off-by: Jeff Moyer <jmoyer(a)redhat.com>
diff --git a/drivers/nvdimm/btt.c b/drivers/nvdimm/btt.c
index e949e33..5860f99 100644
@@ -630,6 +630,7 @@ static struct arena_info *alloc_arena(struct btt *btt, size_t size,
arena->nd_btt = btt->nd_btt;
arena->sector_size = btt->sector_size;
@@ -758,7 +759,6 @@ static int discover_arenas(struct btt *btt)
arena->external_lba_start = cur_nlba;
parse_arena_meta(arena, super, cur_off);
ret = btt_freelist_init(arena);