This work moves the bpf loader into the iproute2 library and reworks
the tc specific parts into generic code. It's useful as we can then
more easily support new program types by just having the same ELF
loader backend. Joint work with Thomas Graf. I hacked a rough start
of a test suite to make sure nothing breaks [1] and looks all good.
[1] https://github.com/borkmann/clsact/blob/master/test_bpf.sh
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Thomas Graf <tgraf@suug.ch>
This big patch was compiled by vimgrepping for memset calls and changing
to C99 initializer if applicable. One notable exception is the
initialization of union bpf_attr in tc/tc_bpf.c: changing it would break
for older gcc versions (at least <=3.4.6).
Calls to memset for struct rtattr pointer fields for parse_rtattr*()
were just dropped since they are not needed.
The changes here allowed the compiler to discover some unused variables,
so get rid of them, too.
Signed-off-by: Phil Sutter <phil@nwl.cc>
Acked-by: David Ahern <dsa@cumulusnetworks.com>
Since we have all infrastructure in place now, allow atomic live updates
on program arrays. This can be very useful e.g. in case programs that are
being tail-called need to be replaced, f.e. when classifier functionality
needs to be changed, new protocols added/removed during runtime, etc.
Thus, provide a way for in-place code updates, minimal example: Given is
an object file cls.o that contains the entry point in section 'classifier',
has a globally pinned program array 'jmp' with 2 slots and id of 0, and
two tail called programs under section '0/0' (prog array key 0) and '0/1'
(prog array key 1), the section encoding for the loader is <id/key>.
Adding the filter loads everything into cls_bpf:
tc filter add dev foo parent ffff: bpf da obj cls.o
Now, the program under section '0/1' needs to be replaced with an updated
version that resides in the same section (also full path to tc's subfolder
of the mount point can be passed, e.g. /sys/fs/bpf/tc/globals/jmp):
tc exec bpf graft m:globals/jmp obj cls.o sec 0/1
In case the program resides under a different section 'foo', it can also
be injected into the program array like:
tc exec bpf graft m:globals/jmp key 1 obj cls.o sec foo
If the new tail called classifier program is already available as a pinned
object somewhere (here: /sys/fs/bpf/tc/progs/parser), it can be injected
into the prog array like:
tc exec bpf graft m:globals/jmp key 1 fd m:progs/parser
In the kernel, the program on key 1 is being atomically replaced and the
old one's refcount dropped.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
This larger work addresses one of the bigger remaining issues on
tc's eBPF frontend, that is, to allow for persistent file descriptors.
Whenever tc parses the ELF object, extracts and loads maps into the
kernel, these file descriptors will be out of reach after the tc
instance exits.
Meaning, for simple (unnested) programs which contain one or
multiple maps, the kernel holds a reference, and they will live
on inside the kernel until the program holding them is unloaded,
but they will be out of reach for user space, even worse with
(also multiple nested) tail calls.
For this issue, we introduced the concept of an agent that can
receive the set of file descriptors from the tc instance creating
them, in order to be able to further inspect/update map data for
a specific use case. However, while that is more tied towards
specific applications, it still doesn't easily allow for sharing
maps accross multiple tc instances and would require a daemon to
be running in the background. F.e. when a map should be shared by
two eBPF programs, one attached to ingress, one to egress, this
currently doesn't work with the tc frontend.
This work solves exactly that, i.e. if requested, maps can now be
_arbitrarily_ shared between object files (PIN_GLOBAL_NS) or within
a single object (but various program sections, PIN_OBJECT_NS) without
"loosing" the file descriptor set. To make that happen, we use eBPF
object pinning introduced in kernel commit b2197755b263 ("bpf: add
support for persistent maps/progs") for exactly this purpose.
The shipped examples/bpf/bpf_shared.c code from this patch can be
easily applied, for instance, as:
- classifier-classifier shared:
tc filter add dev foo parent 1: bpf obj shared.o sec egress
tc filter add dev foo parent ffff: bpf obj shared.o sec ingress
- classifier-action shared (here: late binding to a dummy classifier):
tc actions add action bpf obj shared.o sec egress pass index 42
tc filter add dev foo parent ffff: bpf obj shared.o sec ingress
tc filter add dev foo parent 1: bpf bytecode '1,6 0 0 4294967295,' \
action bpf index 42
The toy example increments a shared counter on egress and dumps its
value on ingress (if no sharing (PIN_NONE) would have been chosen,
map value is 0, of course, due to the two map instances being created):
[...]
<idle>-0 [002] ..s. 38264.788234: : map val: 4
<idle>-0 [002] ..s. 38264.788919: : map val: 4
<idle>-0 [002] ..s. 38264.789599: : map val: 5
[...]
... thus if both sections reference the pinned map(s) in question,
tc will take care of fetching the appropriate file descriptor.
The patch has been tested extensively on both, classifier and
action sides.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
This work follows upon commit 6256f8c9e4 ("tc, bpf: finalize eBPF
support for cls and act front-end") and takes up the idea proposed by
Hannes Frederic Sowa to spawn a shell (or any other command) that holds
generated eBPF map file descriptors.
File descriptors, based on their id, are being fetched from the same
unix domain socket as demonstrated in the bpf_agent, the shell spawned
via execvpe(2) and the map fds passed over the environment, and thus
are made available to applications in the fashion of std{in,out,err}
for read/write access, for example in case of iproute2's examples/bpf/:
# env | grep BPF
BPF_NUM_MAPS=3
BPF_MAP1=6 <- BPF_MAP_ID_QUEUE (id 1)
BPF_MAP0=5 <- BPF_MAP_ID_PROTO (id 0)
BPF_MAP2=7 <- BPF_MAP_ID_DROPS (id 2)
# ls -la /proc/self/fd
[...]
lrwx------. 1 root root 64 Apr 14 16:46 0 -> /dev/pts/4
lrwx------. 1 root root 64 Apr 14 16:46 1 -> /dev/pts/4
lrwx------. 1 root root 64 Apr 14 16:46 2 -> /dev/pts/4
[...]
lrwx------. 1 root root 64 Apr 14 16:46 5 -> anon_inode:bpf-map
lrwx------. 1 root root 64 Apr 14 16:46 6 -> anon_inode:bpf-map
lrwx------. 1 root root 64 Apr 14 16:46 7 -> anon_inode:bpf-map
The advantage (as opposed to the direct/native usage) is that now the
shell is map fd owner and applications can terminate and easily reattach
to descriptors w/o any kernel changes. Moreover, multiple applications
can easily read/write eBPF maps simultaneously.
To further allow users for experimenting with that, next step is to add
a small helper that can get along with simple data types, so that also
shell scripts can make use of bpf syscall, f.e to read/write into maps.
Generally, this allows for prepopulating maps, or any runtime altering
which could influence eBPF program behaviour (f.e. different run-time
classifications, skb modifications, ...), dumping of statistics, etc.
Reference: http://thread.gmane.org/gmane.linux.network/357471/focus=357860
Suggested-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
Acked-by: Alexei Starovoitov <ast@plumgrid.com>
