bpfloader/bpfloader.rc

# zygote-start is what officially starts netd (see //system/core/rootdir/init.rc)
# However, on some hardware it's started from post-fs-data as well, which is just
# a tad earlier.  There's no benefit to that though, since on 4.9+ P+ devices netd
# will just block until bpfloader finishes and sets the bpf.progs_loaded property.
#
# It is important that we start bpfloader after:
#   - /sys/fs/bpf is already mounted,
#   - apex (incl. rollback) is initialized (so that in the future we can load bpf
#     programs shipped as part of apex mainline modules)
#   - system properties have been set, this is because isBpfSupported() calls
#     getUncachedBpfSupportLevel() which depends on
#     ro.kernel.ebpf.supported, ro.product.first_api_level & ro.build.version.sdk
#   - logd is ready for us to log stuff
#
# At the same time we want to be as early as possible to reduce races and thus
# failures (before memory is fragmented, and cpu is busy running tons of other
# stuff) and we absolutely want to be before netd and the system boot slot is
# considered to have booted successfully.
#
on load_bpf_programs
    # Enable the eBPF JIT -- but do note that on 64-bit kernels it is likely
    # already force enabled by the kernel config option BPF_JIT_ALWAYS_ON
    write /proc/sys/net/core/bpf_jit_enable 1
    # Enable JIT kallsyms export for privileged users only
    write /proc/sys/net/core/bpf_jit_kallsyms 1
    start bpfloader

service bpfloader /system/bin/bpfloader
    capabilities CHOWN SYS_ADMIN
    #
    # Set RLIMIT_MEMLOCK to 1GiB for bpfloader
    #
    # Actually only 8MiB would be needed if bpfloader ran as its own uid.
    #
    # However, while the rlimit is per-thread, the accounting is system wide.
    # So, for example, if the graphics stack has already allocated 10MiB of
    # memlock data before bpfloader even gets a chance to run, it would fail
    # if its memlock rlimit is only 8MiB - since there would be none left for it.
    #
    # bpfloader succeeding is critical to system health, since a failure will
    # cause netd crashloop and thus system server crashloop... and the only
    # recovery is a full kernel reboot.
    #
    # We've had issues where devices would sometimes (rarely) boot into
    # a crashloop because bpfloader would occasionally lose a boot time
    # race against the graphics stack's boot time locked memory allocation.
    #
    # Thus bpfloader's memlock has to be 8MB higher then the locked memory
    # consumption of the root uid anywhere else in the system...
    # But we don't know what that is for all possible devices...
    #
    # Ideally, we'd simply grant bpfloader the IPC_LOCK capability and it
    # would simply ignore it's memlock rlimit... but it turns that this
    # capability is not even checked by the kernel's bpf system call.
    #
    # As such we simply use 1GiB as a reasonable approximation of infinity.
    #
    rlimit memlock 1073741824 1073741824
    oneshot
    reboot_on_failure reboot,bpfloader-failed
    # we're not really updatable, but want to be able to load bpf programs shipped in apexes
    updatable