Re: [PATCH net v4] bpf, net: Check cgroup_bpf_enabled() only once in do_sock_getsockopt()

On Wed, 2024-08-21 at 14:01 -0700, Alexei Starovoitov wrote:

 	 
External email : Please do not click links or open attachments until
you have verified the sender or the content.
 On Wed, Aug 21, 2024 at 2:30 AM Tze-nan Wu [off-list ref]
wrote:

quoted

The return value from `cgroup_bpf_enabled(CGROUP_GETSOCKOPT)` can

change

quoted

between the invocations of `BPF_CGROUP_GETSOCKOPT_MAX_OPTLEN` and
`BPF_CGROUP_RUN_PROG_GETSOCKOPT`.

If `cgroup_bpf_enabled(CGROUP_GETSOCKOPT)` changes from "false" to
"true" between the invocations of

`BPF_CGROUP_GETSOCKOPT_MAX_OPTLEN` and

quoted

`BPF_CGROUP_RUN_PROG_GETSOCKOPT`, `BPF_CGROUP_RUN_PROG_GETSOCKOPT`

will

quoted

receive an -EFAULT from

`__cgroup_bpf_run_filter_getsockopt(max_optlen=0)`

quoted

due to `get_user()` was not reached in

`BPF_CGROUP_GETSOCKOPT_MAX_OPTLEN`.

quoted

Scenario shown as below:

           `process A`                      `process B`
           -----------                      ------------
  BPF_CGROUP_GETSOCKOPT_MAX_OPTLEN
                                            enable

CGROUP_GETSOCKOPT

quoted

  BPF_CGROUP_RUN_PROG_GETSOCKOPT (-EFAULT)

To prevent this, invoke `cgroup_bpf_enabled()` only once and cache

the

quoted

result in a newly added local variable `enabled`.
Both `BPF_CGROUP_*` macros in `do_sock_getsockopt` will then check

their

quoted

condition using the same `enabled` variable as the condition

variable,

quoted

instead of using the return values from `cgroup_bpf_enabled` called

by

quoted

themselves as the condition variable(which could yield different

results).

quoted

This ensures that either both `BPF_CGROUP_*` macros pass the

condition

quoted

or neither does.

Fixes: 0d01da6afc54 ("bpf: implement getsockopt and setsockopt

hooks")

quoted

Co-developed-by: Yanghui Li <redacted>
Signed-off-by: Yanghui Li <redacted>
Co-developed-by: Cheng-Jui Wang <redacted>
Signed-off-by: Cheng-Jui Wang <redacted>
Signed-off-by: Tze-nan Wu <redacted>
---

Chagnes from v1 to v2: 

https://lore.kernel.org/all/20240819082513.27176-1-Tze-nan.Wu@mediatek.com/ (local)

quoted

  Instead of using cgroup_lock in the fastpath, invoke

cgroup_bpf_enabled

quoted

  only once and cache the value in the newly added variable

`enabled`.

quoted

  `BPF_CGROUP_*` macros in do_sock_getsockopt can then both check

their

quoted

  condition with the new variable `enable`, ensuring that either

they both

quoted

  passing the condition or both do not.

Chagnes from v2 to v3: 

https://lore.kernel.org/all/20240819155627.1367-1-Tze-nan.Wu@mediatek.com/ (local)

quoted

  Hide cgroup_bpf_enabled in the macro, and some modifications to

adapt

quoted

  the coding style.

Chagnes from v3 to v4: 

https://lore.kernel.org/all/20240820092942.16654-1-Tze-nan.Wu@mediatek.com/ (local)

quoted

  Add bpf tag to subject, and Fixes tag in body.

---
 include/linux/bpf-cgroup.h | 15 ++++++++-------
 net/socket.c               |  5 +++--
 2 files changed, 11 insertions(+), 9 deletions(-)

diff --git a/include/linux/bpf-cgroup.h b/include/linux/bpf-

cgroup.h

quoted

index fb3c3e7181e6..5afa2ac76aae 100644

--- a/include/linux/bpf-cgroup.h
+++ b/include/linux/bpf-cgroup.h

@@ -390,20 +390,20 @@ static inline bool

cgroup_bpf_sock_enabled(struct sock *sk,

quoted

        __ret;                                                     

            \

quoted

 })

-#define

BPF_CGROUP_GETSOCKOPT_MAX_OPTLEN(optlen)                             
 \

quoted

+#define BPF_CGROUP_GETSOCKOPT_MAX_OPTLEN(optlen,

enabled)                     \

quoted

 ({                                                                

            \

quoted

        int __ret =

0;                                                         \

quoted

-       if

(cgroup_bpf_enabled(CGROUP_GETSOCKOPT))                             \

quoted

+       enabled =

cgroup_bpf_enabled(CGROUP_GETSOCKOPT);                       \

quoted

+       if (enabled)

I suspect the compiler generates slow code after such a patch.
pw-bot: cr

What is the problem with double cgroup_bpf_enabled() check?
yes it might return two different values, so?

Depending on where the -EFAULT occurs, the problem could be different.
In our case, the -EFAULT is returned from getsockopt() during a
"bootup-critical property setting" flow in Android. As a result, the
property setting fails due it get -EFAULT from getsockopt(), causing
the device to fail the boot process.

Should the userspace caller always anticipate an -EFAULT from
getsockopt() if there's another process enables CGROUP_GETSOCKOPT
(possibly through the bpf() syscall) at the same time?
If that's the case, then I will handle this in userspace.

Thanks,
--tze-nan