Re: [PATCH bpf-next V9 1/3] bpf: new helper to obtain namespace data from current task

Yonghong, 

Thanks for the pointer, I fixed this bug, but I found another one that's triggered
now the test program I included in  tools/testing/selftests/bpf/test_pidns.
It's seemed that fname was not correctly setup when passing it to filename_lookup.
This is fixed now and I'm doing some more testing.
I think I'll remove the tests on samples/bpf as they are mostly end on -EPERM as 
the fix intended.
Is ok to remove them and just focus to finish the self tests code?.

Bests

On Wed, Aug 14, 2019 at 01:25:06AM -0400, carlos antonio neira bustos wrote:

Thank you very much!

Bests

El mié., 14 de ago. de 2019 00:50, Yonghong Song [off-list ref] escribió:

quoted

On 8/13/19 5:56 PM, Carlos Antonio Neira Bustos wrote:

quoted

On Tue, Aug 13, 2019 at 11:11:14PM +0000, Yonghong Song wrote:

quoted

On 8/13/19 11:47 AM, Carlos Neira wrote:

quoted

From: Carlos <redacted>

New bpf helper bpf_get_current_pidns_info.
This helper obtains the active namespace from current and returns
pid, tgid, device and namespace id as seen from that namespace,
allowing to instrument a process inside a container.

Signed-off-by: Carlos Neira <redacted>
---
   fs/internal.h            |  2 --
   fs/namei.c               |  1 -
   include/linux/bpf.h      |  1 +
   include/linux/namei.h    |  4 +++
   include/uapi/linux/bpf.h | 31 ++++++++++++++++++++++-
   kernel/bpf/core.c        |  1 +
   kernel/bpf/helpers.c     | 64

++++++++++++++++++++++++++++++++++++++++++++++++

quoted

quoted

quoted

   kernel/trace/bpf_trace.c |  2 ++
   8 files changed, 102 insertions(+), 4 deletions(-)

[...]

quoted

quoted

quoted

+BPF_CALL_2(bpf_get_current_pidns_info, struct bpf_pidns_info *,

pidns_info, u32,

quoted

quoted

quoted

+    size)
+{
+   const char *pidns_path = "/proc/self/ns/pid";
+   struct pid_namespace *pidns = NULL;
+   struct filename *tmp = NULL;
+   struct inode *inode;
+   struct path kp;
+   pid_t tgid = 0;
+   pid_t pid = 0;
+   int ret;
+   int len;

Thank you very much for catching this!.
Could you share how to replicate this bug?.

The config is attached. just run trace_ns_info and you
can reproduce the issue.

quoted

quoted

I am running your sample program and get the following kernel bug:

...
[   26.414825] BUG: sleeping function called from invalid context at
/data/users/yhs/work/net-next/fs
/dcache.c:843
[   26.416314] in_atomic(): 1, irqs_disabled(): 0, pid: 1911, name: ping
[   26.417189] CPU: 0 PID: 1911 Comm: ping Tainted: G        W
5.3.0-rc1+ #280
[   26.418182] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996),
BIOS 1.9.3-1.el7.centos 04/01/2
014
[   26.419393] Call Trace:
[   26.419697]  <IRQ>
[   26.419960]  dump_stack+0x46/0x5b
[   26.420434]  ___might_sleep+0xe4/0x110
[   26.420894]  dput+0x2a/0x200
[   26.421265]  walk_component+0x10c/0x280
[   26.421773]  link_path_walk+0x327/0x560
[   26.422280]  ? proc_ns_dir_readdir+0x1a0/0x1a0
[   26.422848]  ? path_init+0x232/0x330
[   26.423364]  path_lookupat+0x88/0x200
[   26.423808]  ? selinux_parse_skb.constprop.69+0x124/0x430
[   26.424521]  filename_lookup+0xaf/0x190
[   26.425031]  ? simple_attr_release+0x20/0x20
[   26.425560]  bpf_get_current_pidns_info+0xfa/0x190
[   26.426168]  bpf_prog_83627154cefed596+0xe66/0x1000
[   26.426779]  trace_call_bpf+0xb5/0x160
[   26.427317]  ? __netif_receive_skb_core+0x1/0xbb0
[   26.427929]  ? __netif_receive_skb_core+0x1/0xbb0
[   26.428496]  kprobe_perf_func+0x4d/0x280
[   26.428986]  ? tracing_record_taskinfo_skip+0x1a/0x30
[   26.429584]  ? tracing_record_taskinfo+0xe/0x80
[   26.430152]  ? ttwu_do_wakeup.isra.114+0xcf/0xf0
[   26.430737]  ? __netif_receive_skb_core+0x1/0xbb0
[   26.431334]  ? __netif_receive_skb_core+0x5/0xbb0
[   26.431930]  kprobe_ftrace_handler+0x90/0xf0
[   26.432495]  ftrace_ops_assist_func+0x63/0x100
[   26.433060]  0xffffffffc03180bf
[   26.433471]  ? __netif_receive_skb_core+0x1/0xbb0
...

To prevent we are running in arbitrary task (e.g., idle task)
context which may introduce sleeping issues, the following
probably appropriate:

         if (in_nmi() || in_softirq())
                 return -EPERM;

Anyway, if in nmi or softirq, the namespace and pid/tgid
we get may be just accidentally associated with the bpf running
context, but it could be in a different context. So such info
is not reliable any way.

quoted

+
+   if (unlikely(size != sizeof(struct bpf_pidns_info)))
+           return -EINVAL;
+   pidns = task_active_pid_ns(current);

[...]