ioprio_set() and the per-request I/O priority of aio and io_uring only hand out IOPRIO_CLASS_RT to tasks that have CAP_SYS_ADMIN or CAP_SYS_NICE. The io.prio.class cgroup attribute has no such check. Writing "promote-to-rt", or its alias "none-to-rt", to it changes the requests of the cgroup to the RT class. The attribute exists in every cgroup that has the io controller enabled and it is not hierarchical. A user that got a cgroup delegated owns the files of the child cgroups it creates there, so it can create a child cgroup, move into it and write "promote-to-rt" to its io.prio.class. Its requests are then issued with IOPRIO_CLASS_RT although ioprio_set() fails with -EPERM for the same task, and mq-deadline dispatches them before the best-effort requests of everybody else. The delegation model in Documentation/admin-guide/cgroup-v2.rst expects the settings of all resource controllers to be hierarchical, so that a delegatee can only distribute further what it received from its parent. This attribute lets it raise the priority class of its own requests instead. Require CAP_SYS_ADMIN or CAP_SYS_NICE for the two policies that raise the priority class, like ioprio_check_cap() does. The other policies are left alone. Fixes: 556910e39249 ("block: Introduce the ioprio rq-qos policy") Cc: stable@vger.kernel.org Signed-off-by: Hungyu Lin --- block/blk-ioprio.c | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/block/blk-ioprio.c b/block/blk-ioprio.c index 8fa8bca35062..ea32db727b16 100644 --- a/block/blk-ioprio.c +++ b/block/blk-ioprio.c @@ -14,6 +14,7 @@ #include #include +#include #include #include #include "blk-cgroup.h" @@ -91,6 +92,9 @@ static ssize_t ioprio_set_prio_policy(struct kernfs_open_file *of, char *buf, ret = sysfs_match_string(policy_name, buf); if (ret < 0) return ret; + if ((ret == POLICY_PROMOTE_TO_RT || ret == POLICY_NONE_TO_RT) && + !capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE)) + return -EPERM; blkcg->prio_policy = ret; return nbytes; } -- 2.43.0