io_uring: fix races with buffer table unregister
authorPavel Begunkov <asml.silence@gmail.com>
Mon, 13 Jun 2022 05:30:06 +0000 (06:30 +0100)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Wed, 22 Jun 2022 12:22:00 +0000 (14:22 +0200)
[ Upstream commit d11d31fc5d8a96f707facee0babdcffaafa38de2 ]

Fixed buffer table quiesce might unlock ->uring_lock, potentially
letting new requests to be submitted, don't allow those requests to
use the table as they will race with unregistration.

Reported-and-tested-by: van fantasy <g1042620637@gmail.com>
Fixes: bd54b6fe3316ec ("io_uring: implement fixed buffers registration similar to fixed files")
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>
fs/io_uring.c

index 5f111a6..be21765 100644 (file)
@@ -8905,12 +8905,19 @@ static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
 
 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
 {
+       unsigned nr = ctx->nr_user_bufs;
        int ret;
 
        if (!ctx->buf_data)
                return -ENXIO;
 
+       /*
+        * Quiesce may unlock ->uring_lock, and while it's not held
+        * prevent new requests using the table.
+        */
+       ctx->nr_user_bufs = 0;
        ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
+       ctx->nr_user_bufs = nr;
        if (!ret)
                __io_sqe_buffers_unregister(ctx);
        return ret;