bug: fix assign vfsmnt correctly

[arthur-zhang]

Fixes

Description

cwd_read do not assign data->vfsmnt correctly.

reproduce:

sudo mount -t nfs  x.x.x.x:/data/dev/nfs_dir /home/arthur/my_run

$ mount -t
x.x.x.x:/data/dev/nfs_dir on /home/arthur/my_run
/dev/nvme0n1p5 on /home type ext4 (rw,relatime)

and cat a file in my_run dir

cat /home/arthur/my_run/a.txt

in tetragon side, it will get a path : /home/home/arthur/my_run/a.txt, but it should be /home/arthur/my_run/a.txt

cat-2152134 [003] d...1 3725192.671779: bpf_trace_printk: copy path: /home/home/arthur/my_run/a.txt

and in another production env, i found it will get path more than expected when i cat /tmp/tetragon. the prefix /scon/containers/01J6HEV7R29R4WXXS1N2CS9ATP/rootfs/ should not be obtained.

/scon/containers/01J6HEV7R29R4WXXS1N2CS9ATP/rootfs/tmp/tetragon

in the old code, mnt is pointer to the old data->mnt, which is not correct.

struct mount *mnt = data->mnt;
...
if (data->mnt != parent) {
    probe_read(&data->dentry, sizeof(data->dentry),
			    _(&mnt->mnt_mountpoint));
    data->mnt = parent;
    probe_read(&data->vfsmnt, sizeof(data->vfsmnt),
			    _(&mnt->mnt));
    return 0;
}

Changelog

Fix vfsmnt assign bug in bpf/process/bpf_process_event.h
Fixes

Description

cwd_read do not assign data->vfsmnt correctly.

reproduce:

sudo mount -t nfs  x.x.x.x:/data/dev/nfs_dir /home/arthur/my_run

$ mount -t
x.x.x.x:/data/dev/nfs_dir on /home/arthur/my_run
/dev/nvme0n1p5 on /home type ext4 (rw,relatime)

and cat a file in my_run dir

cat /home/arthur/my_run/a.txt

in tetragon side, it will get a path : /home/home/arthur/my_run/a.txt, but it should be /home/arthur/my_run/a.txt

cat-2152134 [003] d...1 3725192.671779: bpf_trace_printk: copy path: /home/home/arthur/my_run/a.txt

and in another production env, i found it will get path more than expected when i cat /tmp/tetragon. the prefix /scon/containers/01J6HEV7R29R4WXXS1N2CS9ATP/rootfs/ should not be obtained.

/scon/containers/01J6HEV7R29R4WXXS1N2CS9ATP/rootfs/tmp/tetragon

in the old code, mnt is pointer to the old data->mnt, which is not correct.

struct mount *mnt = data->mnt;
...
if (data->mnt != parent) {
    probe_read(&data->dentry, sizeof(data->dentry),
			    _(&mnt->mnt_mountpoint));
    data->mnt = parent;
    probe_read(&data->vfsmnt, sizeof(data->vfsmnt),
			    _(&mnt->mnt));
    return 0;
}

[olsajiri]

Fixes

Description

cwd_read do not assign data->vfsmnt correctly.

reproduce:

sudo mount -t nfs  x.x.x.x:/data/dev/nfs_dir /home/arthur/my_run

$ mount -t
x.x.x.x:/data/dev/nfs_dir on /home/arthur/my_run
/dev/nvme0n1p5 on /home type ext4 (rw,relatime)

and cat a file in my_run dir

cat /home/arthur/my_run/a.txt

in tetragon side, it will get a path : /home/home/arthur/my_run/a.txt, but it should be /home/arthur/my_run/a.txt

cat-2152134 [003] d...1 3725192.671779: bpf_trace_printk: copy path: /home/home/arthur/my_run/a.txt

and in another production env, i found it will get path more than expected when i cat /tmp/tetragon. the prefix /scon/containers/01J6HEV7R29R4WXXS1N2CS9ATP/rootfs/ should not be obtained.

/scon/containers/01J6HEV7R29R4WXXS1N2CS9ATP/rootfs/tmp/tetragon

in the old code, mnt is pointer to the old data->mnt, which is not correct.

struct mount *mnt = data->mnt;
...
if (data->mnt != parent) {
    probe_read(&data->dentry, sizeof(data->dentry),
			    _(&mnt->mnt_mountpoint));
    data->mnt = parent;
    probe_read(&data->vfsmnt, sizeof(data->vfsmnt),
			    _(&mnt->mnt));
    return 0;
}

Changelog

Fix vfsmnt assign bug in bpf/process/bpf_process_event.h

@arthur-zhang please put all this in the commit changelog, also any chance you could add test the issue? thanks

[olsajiri]

ok, parent->mnt is not pointer but actual vfsmnt object embedded in mount struct,
so we can just take the address from parent.. but I think you still need to use _() to get CORE involved

[arthur-zhang]

OK， i will fix it