The array_index_nospec() check in io_sq_offload_start() is performed
before any checks on p->sq_thread_cpu are done. This means cpu is
clamped and therefore no error occurs when out-of-range values are
passed in from userspace. This is in violation of the specification
for io_ring_setup() and causes the io_ring_setup unit test in liburing
to regress.
Add a new bounds check on sq_thread_cpu at the start of
io_sq_offload_start() so we can exit the function early when bad
values are passed in.
Fixes: 975554b03edd ("io_uring: fix SQPOLL cpu validation")
Signed-off-by: Stephen Bates <sbates@xxxxxxxxxxxx>
---
fs/io_uring.c | 3 +++
1 file changed, 3 insertions(+)
diff --git a/fs/io_uring.c b/fs/io_uring.c
index 30a5687..e458470 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -2316,6 +2316,9 @@ static int io_sq_offload_start(struct io_ring_ctx *ctx,
{
int ret;
+ if (p->sq_thread_cpu >= nr_cpu_ids)
+ return -EINVAL;
+
init_waitqueue_head(&ctx->sqo_wait);
mmgrab(current->mm);
ctx->sqo_mm = current->mm;
--
2.7.4
