Re: [PATCH RFC v2 net-next 05/16] bpf: introduce syscall(BPF, ...) and BPF maps

On Wed, Jul 23, 2014 at 11:02 AM, Kees Cook [off-list ref] wrote:

quoted

--- a/Documentation/networking/filter.txt
+++ b/Documentation/networking/filter.txt

+eBPF maps
+---------
+'maps' is a generic storage of different types for sharing data between kernel
+and userspace.
+
+The maps are accessed from user space via BPF syscall, which has commands:
+- create a map with given id, type and attributes
+  map_id = bpf_map_create(int map_id, map_type, struct nlattr *attr, int len)
+  returns positive map id or negative error

Looks like these docs need updating for the fd-based approach instead
of the map_id approach?

ohh, yes. updated it in srcs and in commit log, but forgot in docs.

quoted

+SYSCALL_DEFINE5(bpf, int, cmd, unsigned long, arg2, unsigned long, arg3,
+               unsigned long, arg4, unsigned long, arg5)
+{
+       if (!capable(CAP_SYS_ADMIN))
+               return -EPERM;

It might be valuable to have a comment here describing why this is
currently limited to CAP_SYS_ADMIN.

makes sense.
There are several reasons it should be limited to root initially:
- to phase changes in gradually
- verifier is not detecting pointer leaks yet
- full security audit wasn't performed
- tracing and network analytics are root only anyway
Currently eBPF is safe (non-crashing), since safety is relatively easy
to enforce by static analysis. For somebody with compiler background
it's natural to think about bounds, alignments, uninitialized access, etc.
So I'm confident that I didn't miss anything big in 'safety' aspect.
'Non-root security' is harder. I'll add pointer leak detection first
and will ask for more suggestions.

quoted

+       switch (cmd) {
+       case BPF_MAP_CREATE:
+               return map_create((enum bpf_map_type) arg2,
+                                 (struct nlattr __user *) arg3, (int) arg4);

I'd recommend requiring arg5 == 0 here, just for future flexibility.

Though I expect all extensions to go through nlattr attributes,
it's indeed cleaner to enforce arg5==0 here and for all other cmds.