# 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)
#   - 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
    exec_start bpfloader

service bpfloader /system/bin/netbpfload
    # netbpfload will do network bpf loading, then execute /system/bin/bpfloader
    capabilities CHOWN SYS_ADMIN NET_ADMIN
    # The following group memberships are a workaround for lack of DAC_OVERRIDE
    # and allow us to open (among other things) files that we created and are
    # no longer root owned (due to CHOWN) but still have group read access to
    # one of the following groups.  This is not perfect, but a more correct
    # solution requires significantly more effort to implement.
    group root graphics network_stack net_admin net_bw_acct net_bw_stats net_raw system
    user root
    #
    # 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
    #
    # How to debug bootloops caused by 'bpfloader-failed'.
    #
    # 1. On some lower RAM devices (like wembley) you may need to first enable developer mode
    #    (from the Settings app UI), and change the developer option "Logger buffer sizes"
    #    from the default (wembley: 64kB) to the maximum (1M) per log buffer.
    #    Otherwise buffer will overflow before you manage to dump it and you'll get useless logs.
    #
    # 2. comment out 'reboot_on_failure reboot,bpfloader-failed' below
    # 3. rebuild/reflash/reboot
    # 4. as the device is booting up capture bpfloader logs via:
    #    adb logcat -s 'bpfloader:*' 'LibBpfLoader:*' 'NetBpfLoad:*' 'NetBpfLoader:*'
    #
    # something like:
    #   $ adb reboot; sleep 1; adb wait-for-device; adb root; sleep 1; adb wait-for-device; adb logcat -s 'bpfloader:*' 'LibBpfLoader:*' 'NetBpfLoad:*' 'NetBpfLoader:*'
    # will take care of capturing logs as early as possible
    #
    # 5. look through the logs from the kernel's bpf verifier that bpfloader dumps out,
    #    it usually makes sense to search back from the end and find the particular
    #    bpf verifier failure that caused bpfloader to terminate early with an error code.
    #    This will probably be something along the lines of 'too many jumps' or
    #    'cannot prove return value is 0 or 1' or 'unsupported / unknown operation / helper',
    #    'invalid bpf_context access', etc.
    #
    reboot_on_failure reboot,bpfloader-failed
    # we're not really updatable, but want to be able to load bpf programs shipped in apexes
    updatable