Re: [PATCH] kernel/taskstats: fix wrong nla type for {cgroup,task}stats policy

[Johannes Berg]

On Thu, 2020-03-26 at 13:08 -0700, Andrew Morton wrote:
> (cc's added)
> 
> On Wed, 25 Mar 2020 22:50:42 -0400 Yafang Shao <laoar.shao@xxxxxxxxx> wrote:
> 
> > After our server is upgraded to a newer kernel, we found that it
> > continuesly print a warning in the kernel message. The warning is,
> > [832984.946322] netlink: 'irmas.lc': attribute type 1 has an invalid length.
> > 
> > irmas.lc is one of our container monitor daemons, and it will use
> > CGROUPSTATS_CMD_GET to get the cgroupstats, that is similar with
> > tools/accounting/getdelays.c. We can also produce this warning with
> > getdelays. For example, after running bellow command
> > 	$ ./getdelays -C /sys/fs/cgroup/memory
> > then you can find a warning in dmesg,
> > [61607.229318] netlink: 'getdelays': attribute type 1 has an invalid length.

And looking at this ... well, that code is completely wrong?

E.g.

                rc = send_cmd(nl_sd, id, mypid, TASKSTATS_CMD_GET,
                              cmd_type, &tid, sizeof(__u32));

(cmd_type is one of TASKSTATS_CMD_ATTR_TGID, TASKSTATS_CMD_ATTR_PID)

or it might do

                rc = send_cmd(nl_sd, id, mypid, CGROUPSTATS_CMD_GET,
                              CGROUPSTATS_CMD_ATTR_FD, &cfd, sizeof(__u32));

so clearly it wants to produce a u32 attribute.

But then

static int send_cmd(int sd, __u16 nlmsg_type, __u32 nlmsg_pid,
             __u8 genl_cmd, __u16 nla_type,
             void *nla_data, int nla_len)
{
...

        na = (struct nlattr *) GENLMSG_DATA(&msg);

// this is still fine

        na->nla_type = nla_type;

// this is also fine

        na->nla_len = nla_len + 1 + NLA_HDRLEN;

// but this??? the nla_len of a netlink attribute should just be
// the len ... what's NLA_HDRLEN doing here? this isn't nested
// here we end up just reserving 1+NLA_HDRLEN too much space

        memcpy(NLA_DATA(na), nla_data, nla_len);

// but then it anyway only fills the first nla_len bytes, which
// is just like a regular attribute.

        msg.n.nlmsg_len += NLMSG_ALIGN(na->nla_len);
// note that this is also wrong - it should be 
// += NLA_ALIGN(NLA_HDRLEN + nla_len)



So really I think what happened here is precisely what we wanted -
David's kernel patch caught the broken userspace tool.

johannes