epoll can call out to vfs_poll() with a file pointer that may race with the last 'fput()'. That would make f_count go down to zero, and while the ep->mtx locking means that the resulting file pointer tear-down will be blocked until the poll returns, it means that f_count is already dead, and any use of it won't actually get a reference to the file any more: it's dead regardless. Make sure we have a valid ref on the file pointer before we call down to vfs_poll() from the epoll routines. Link: https://lore.kernel.org/lkml/0000000000002d631f0615918f1e@xxxxxxxxxx/ Reported-by: syzbot+045b454ab35fd82a35fb@xxxxxxxxxxxxxxxxxxxxxxxxx Reviewed-by: Jens Axboe <axboe@xxxxxxxxx> Signed-off-by: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> --- Changes since v1: - add Link, Reported-by, and Jens' reviewed-by. And sign off on it because it looks fine to me and we have some testing now. - move epi_fget() closer to the user - more comments about the background - remove the rcu_read_lock(), with the comment explaining why it's not needed - note about returning zero rather than something like EPOLLERR|POLLHUP for a file that is going away fs/eventpoll.c | 42 +++++++++++++++++++++++++++++++++++++++++- 1 file changed, 41 insertions(+), 1 deletion(-) diff --git a/fs/eventpoll.c b/fs/eventpoll.c index 882b89edc52a..a3f0f868adc4 100644 --- a/fs/eventpoll.c +++ b/fs/eventpoll.c @@ -979,6 +979,37 @@ static __poll_t __ep_eventpoll_poll(struct file *file, poll_table *wait, int dep return res; } +/* + * The ffd.file pointer may be in the process of + * being torn down due to being closed, but we + * may not have finished eventpoll_release() yet. + * + * Normally, even with the atomic_long_inc_not_zero, + * the file may have been free'd and then gotten + * re-allocated to something else (since files are + * not RCU-delayed, they are SLAB_TYPESAFE_BY_RCU). + * + * But for epoll, users hold the ep->mtx mutex, and + * as such any file in the process of being free'd + * will block in eventpoll_release_file() and thus + * the underlying file allocation will not be free'd, + * and the file re-use cannot happen. + * + * For the same reason we can avoid a rcu_read_lock() + * around the operation - 'ffd.file' cannot go away + * even if the refcount has reached zero (but we must + * still not call out to ->poll() functions etc). + */ +static struct file *epi_fget(const struct epitem *epi) +{ + struct file *file; + + file = epi->ffd.file; + if (!atomic_long_inc_not_zero(&file->f_count)) + file = NULL; + return file; +} + /* * Differs from ep_eventpoll_poll() in that internal callers already have * the ep->mtx so we need to start from depth=1, such that mutex_lock_nested() @@ -987,14 +1018,23 @@ static __poll_t __ep_eventpoll_poll(struct file *file, poll_table *wait, int dep static __poll_t ep_item_poll(const struct epitem *epi, poll_table *pt, int depth) { - struct file *file = epi->ffd.file; + struct file *file = epi_fget(epi); __poll_t res; + /* + * We could return EPOLLERR | EPOLLHUP or something, + * but let's treat this more as "file doesn't exist, + * poll didn't happen". + */ + if (!file) + return 0; + pt->_key = epi->event.events; if (!is_file_epoll(file)) res = vfs_poll(file, pt); else res = __ep_eventpoll_poll(file, pt, depth); + fput(file); return res & epi->event.events; } -- 2.44.0.330.g4d18c88175
