verify: fix a bug with verify_async
There's a race between marking an io_u as deferred, completing it,
and checking of that flag. It cannot be done reliably for async
verify threads. So change the mechanism to have the verify_async
part pull the io_u completely, so we don't have to check for a flag
in it after we have run ->end_io().
This fixes a bug with verify_async, where fio would crash with
this message:
fio: io_u.c:1315: __get_io_u: Assertion `io_u->flags & IO_U_F_FREE' failed
This race has always existed, but was made considerably worse with
this commit:
commit 2ae0b204743d6b4048c6fffd46c6280a70f2ecd1
Author: Jens Axboe <axboe@kernel.dk>
Date: Tue May 28 14:16:55 2013 +0200
Replace list based free/busy/requeue list with FIFO + ring
Cache friendliness of the list is pretty low. This has
provably lower overhead.
since we moved from a single list holding the io, to a separate
verify list and io_u queues. The above bug is happening because
the io_u ends up on both the freelist and the pending verify list,
causing mayhem.
Reported-by: scameron@beardog.cce.hp.com
Signed-off-by: Jens Axboe <axboe@fb.com>
diff --git a/verify.h b/verify.h
index dba7743..bb3fd49 100644
--- a/verify.h
+++ b/verify.h
@@ -76,8 +76,8 @@
*/
extern void populate_verify_io_u(struct thread_data *, struct io_u *);
extern int __must_check get_next_verify(struct thread_data *td, struct io_u *);
-extern int __must_check verify_io_u(struct thread_data *, struct io_u *);
-extern int verify_io_u_async(struct thread_data *, struct io_u *);
+extern int __must_check verify_io_u(struct thread_data *, struct io_u **);
+extern int verify_io_u_async(struct thread_data *, struct io_u **);
extern void fill_verify_pattern(struct thread_data *td, void *p, unsigned int len, struct io_u *io_u, unsigned long seed, int use_seed);
extern void fill_buffer_pattern(struct thread_data *td, void *p, unsigned int len);
extern void fio_verify_init(struct thread_data *td);