forked from Minki/linux
mainlining shenanigans
0bc23a1d1c
Alexei Starovoitov says:
====================
This patch set is the first real user of user mode driver facility. The
general use case for user mode driver is to ship vmlinux with preloaded BPF
programs. In this particular case the user mode driver populates bpffs instance
with two BPF iterators. In several months BPF_LSM project would need to preload
the kernel with its own set of BPF programs and attach to LSM hooks instead of
bpffs. BPF iterators and BPF_LSM are unstable from uapi perspective. They are
tracing based and peek into arbitrary kernel data structures. One can question
why a kernel module cannot embed BPF programs inside. The reason is that libbpf
is necessary to load them. First libbpf loads BPF Type Format, then creates BPF
maps, populates them. Then it relocates code sections inside BPF programs,
loads BPF programs, and finally attaches them to events. Theoretically libbpf
can be rewritten to work in the kernel, but that is massive undertaking. The
maintenance of in-kernel libbpf and user space libbpf would be another
challenge. Another obstacle to embedding BPF programs into kernel module is
sys_bpf api. Loading of programs, BTF, maps goes through the verifier. It
validates and optimizes the code. It's possible to provide in-kernel api to all
of sys_bpf commands (load progs, create maps, update maps, load BTF, etc), but
that is huge amount of work and forever maintenance headache.
Hence the decision is to ship vmlinux with user mode drivers that load
BPF programs. Just like kernel modules extend vmlinux BPF programs
are safe extensions of the kernel and some of them need to ship with vmlinux.
This patch set adds a kernel module with user mode driver that populates bpffs
with two BPF iterators.
$ mount bpffs /my/bpffs/ -t bpf
$ ls -la /my/bpffs/
total 4
drwxrwxrwt 2 root root 0 Jul 2 00:27 .
drwxr-xr-x 19 root root 4096 Jul 2 00:09 ..
-rw------- 1 root root 0 Jul 2 00:27 maps.debug
-rw------- 1 root root 0 Jul 2 00:27 progs.debug
The user mode driver will load BPF Type Formats, create BPF maps, populate BPF
maps, load two BPF programs, attach them to BPF iterators, and finally send two
bpf_link IDs back to the kernel.
The kernel will pin two bpf_links into newly mounted bpffs instance under
names "progs.debug" and "maps.debug". These two files become human readable.
$ cat /my/bpffs/progs.debug
id name attached
11 dump_bpf_map bpf_iter_bpf_map
12 dump_bpf_prog bpf_iter_bpf_prog
27 test_pkt_access
32 test_main test_pkt_access test_pkt_access
33 test_subprog1 test_pkt_access_subprog1 test_pkt_access
34 test_subprog2 test_pkt_access_subprog2 test_pkt_access
35 test_subprog3 test_pkt_access_subprog3 test_pkt_access
36 new_get_skb_len get_skb_len test_pkt_access
37 new_get_skb_ifindex get_skb_ifindex test_pkt_access
38 new_get_constant get_constant test_pkt_access
The BPF program dump_bpf_prog() in iterators.bpf.c is printing this data about
all BPF programs currently loaded in the system. This information is unstable
and will change from kernel to kernel.
In some sence this output is similar to 'bpftool prog show' that is using
stable api to retreive information about BPF programs. The BPF subsytems grows
quickly and there is always demand to show as much info about BPF things as
possible. But we cannot expose all that info via stable uapi of bpf syscall,
since the details change so much. Right now a BPF program can be attached to
only one other BPF program. Folks are working on patches to enable
multi-attach, but for debugging it's necessary to see the current state. There
is no uapi for that, but above output shows it:
37 new_get_skb_ifindex get_skb_ifindex test_pkt_access
38 new_get_constant get_constant test_pkt_access
[1] [2] [3]
[1] is the full name of BPF prog from BTF.
[2] is the name of function inside target BPF prog.
[3] is the name of target BPF prog.
[2] and [3] are not exposed via uapi, since they will change from single to
multi soon. There are many other cases where bpf internals are useful for
debugging, but shouldn't be exposed via uapi due to high rate of changes.
systemd mounts /sys/fs/bpf at the start, so this kernel module with user mode
driver needs to be available early. BPF_LSM most likely would need to preload
BPF programs even earlier.
Few interesting observations:
- though bpffs comes with two human readble files "progs.debug" and
"maps.debug" they can be removed. 'rm -f /sys/fs/bpf/progs.debug' will remove
bpf_link and kernel will automatically unload corresponding BPF progs, maps,
BTFs. In the future '-o remount' will be able to restore them. This is not
implemented yet.
- 'ps aux|grep bpf_preload' shows nothing. User mode driver loaded BPF
iterators and exited. Nothing is lingering in user space at this point.
- We can consider giving 0644 permissions to "progs.debug" and "maps.debug"
to allow unprivileged users see BPF things loaded in the system.
We cannot do so with "bpftool prog show", since it's using cap_sys_admin
parts of bpf syscall.
- The functionality split between core kernel, bpf_preload kernel module and
user mode driver is very similar to bpfilter style of interaction.
- Similar BPF iterators can be used as unstable extensions to /proc.
Like mounting /proc can prepopolate some subdirectory in there with
a BPF iterator that will print QUIC sockets instead of tcp and udp.
Changelog:
v5->v6:
- refactored Makefiles with Andrii's help
- switched to explicit $(MAKE) style
- switched to userldlibs instead of userldflags
- fixed build issue with libbpf Makefile due to invocation from kbuild
- fixed menuconfig order as spotted by Daniel
- introduced CONFIG_USERMODE_DRIVER bool that is selected by bpfilter and bpf_preload
v4->v5:
- addressed Song and Andrii feedback. s/pages/max_entries/
v3->v4:
- took THIS_MODULE in patch 3 as suggested by Daniel to simplify the code.
- converted BPF iterator to use BTF (when available) to print full BPF program name
instead of 16-byte truncated version.
This is something I've been using drgn scripts for.
Take a look at get_name() in iterators.bpf.c to see how short it is comparing
to what user space bpftool would have to do to print the same full name:
. get prog info via obj_info_by_fd
. do get_fd_by_id from info->btf_id
. fetch potentially large BTF of the program from the kernel
. parse that BTF in user space to figure out all type boundaries and string section
. read info->func_info to get btf_id of func_proto from there
. find that btf_id in the parsed BTF
That's quite a bit work for bpftool comparing to few lines in get_name().
I guess would be good to make bpftool do this info extraction anyway.
While doing this BTF reading in the kernel realized that the verifier is not smart
enough to follow double pointers (added to my todo list), otherwise get_name()
would have been even shorter.
v2->v3:
- fixed module unload race (Daniel)
- added selftest (Daniel)
- fixed build bot warning
v1->v2:
- changed names to 'progs.debug' and 'maps.debug' to hopefully better indicate
instability of the text output. Having dot in the name also guarantees
that these special files will not conflict with normal bpf objects pinned
in bpffs, since dot is disallowed for normal pins.
- instead of hard coding link_name in the core bpf moved into UMD.
- cleanedup error handling.
- addressed review comments from Yonghong and Andrii.
====================
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
|
||
|---|---|---|
| arch | ||
| block | ||
| certs | ||
| crypto | ||
| Documentation | ||
| drivers | ||
| fs | ||
| include | ||
| init | ||
| ipc | ||
| kernel | ||
| lib | ||
| LICENSES | ||
| mm | ||
| net | ||
| samples | ||
| scripts | ||
| security | ||
| sound | ||
| tools | ||
| usr | ||
| virt | ||
| .clang-format | ||
| .cocciconfig | ||
| .get_maintainer.ignore | ||
| .gitattributes | ||
| .gitignore | ||
| .mailmap | ||
| COPYING | ||
| CREDITS | ||
| Kbuild | ||
| Kconfig | ||
| MAINTAINERS | ||
| Makefile | ||
| README | ||
Linux kernel
============
There are several guides for kernel developers and users. These guides can
be rendered in a number of formats, like HTML and PDF. Please read
Documentation/admin-guide/README.rst first.
In order to build the documentation, use ``make htmldocs`` or
``make pdfdocs``. The formatted documentation can also be read online at:
https://www.kernel.org/doc/html/latest/
There are various text files in the Documentation/ subdirectory,
several of them using the Restructured Text markup notation.
Please read the Documentation/process/changes.rst file, as it contains the
requirements for building and running the kernel, and information about
the problems which may result by upgrading your kernel.