Omer Tripp's analysis of a Spectre V1 gadget in __close_fd():

"1. __close_fd() is reachable via the close() syscall with a user-controlled fd.

2. If said bounds check is mispredicted, then a user-controlled address fdt->fd[fd] is obtained then dereferenced, and the value of a user-controlled address is loaded into the local variable file.

3. file is then passed as an argument to filp_close, where the cache lines secret + offsetof(f_op) and secret + offsetof(f_mode) are hot and vulnerable to a timing channel attack."

Address this by using array_index_nospec() to prevent speculation past the end of current->fdt.

Reported-by: Omer Tripp trippo@google.com Cc: stable@vger.kernel.org Signed-off-by: Greg Hackmann ghackmann@android.com --- v2: include Omer Tripp's analysis in commit message, and update my email address

fs/file.c | 2 ++ 1 file changed, 2 insertions(+)

diff --git a/fs/file.c b/fs/file.c index 7ffd6e9d103d..a80cf82be96b 100644 --- a/fs/file.c +++ b/fs/file.c @@ -18,6 +18,7 @@ #include <linux/bitops.h> #include <linux/spinlock.h> #include <linux/rcupdate.h> +#include <linux/nospec.h>

unsigned int sysctl_nr_open __read_mostly = 1024*1024; unsigned int sysctl_nr_open_min = BITS_PER_LONG; @@ -626,6 +627,7 @@ int __close_fd(struct files_struct *files, unsigned fd) fdt = files_fdtable(files); if (fd >= fdt->max_fds) goto out_unlock; + fd = array_index_nospec(fd, fdt->max_fds); file = fdt->fd[fd]; if (!file) goto out_unlock;