Hi,
On Debian GNU/Hurd 0.9 (pfinet as shipped in the hurd source package,
hurd 1:0.9.git20230520-7), there is no RPC, no ioctl, and no procfs
file that returns per-interface byte / packet counters. Every tool
that reads /proc/net/dev (ifconfig byte counts, bmon, vnstat,
prometheus node_exporter, anything else) reports zero traffic on
every interface regardless of actual load.
Seven independent probes (in increasing order of authority):
1. /usr/include/x86_64-gnu/bits/ioctls.h defines SIOC* constants
for ADDR / DSTADDR / FLAGS / BRDADDR / CONF / NETMASK / METRIC /
ARP / MTU / INDEX / NAME / HWADDR only. No SIOCGIFSTATS,
no SIOCGIFMIB, no SIOCGIFDATA, no SIOCGIFCOUNTERS macro
anywhere under /usr/include.
2. pfinet.defs (subsystem 37000) at
/usr/include/x86_64-gnu/hurd/pfinet.defs declares the complete
pfinet RPC surface as exactly two routines: pfinet_siocgifconf
and pfinet_getroutes. The MIG-generated pfinet.h matches.
3. iioctl.defs (subsystem 112000) enumerates every per-iface
ioctl pfinet implements. No stats routine is reserved, not
even as an unimplemented slot. Slots 0-11, 13, 15, 18,
20-21, 26-32, 36, 38, 40-50, 53-89 are all marked skip.
4. Runtime SIOCGIFSTATS probe (30-line C, socket(AF_INET,
SOCK_DGRAM, 0) then the standard ioctl menu against lo and
/dev/eth0) returns errno=1073741849 "Inappropriate ioctl for
device" (Hurd-encoded ENOTTY). The baseline ioctls
(SIOCGIFCONF, SIOCGIFFLAGS) round-trip cleanly, so the socket
is genuinely talking to pfinet.
5. Binary symbol scan of /hurd/pfinet finds exactly one
_get_stats match: tunnel_get_stats, a Linux-net tunnel-device
internal callback, not an RPC entry. No pfinet_getstats,
no inet_stats, no siocgifstats strings anywhere in the binary.
6. Procfs walk: /proc/net/ does not exist on this image. The
top-level /proc/ namespace contains cmdline cpuinfo
filesystems hostinfo loadavg meminfo mounts route self
slabinfo stat swaps uptime version vmstat and per-pid dirs.
No snmp, no netstat, no dev. /proc/route exists at the top
level (not under net/) and carries the routing table, but
its Use and RefCnt columns are always 0.
7. Translator identity: showtrans /servers/socket/2 returns
/hurd/pfinet -6 /servers/socket/26; fsysopts /servers/socket/2
returns the live pfinet command line with all configured v4
and v6 addresses. So the AF_INET socket the probe in
finding 4 used is definitely talking to pfinet, not a stub.
The request is to add a per-iface stats getter so the counters
pfinet already maintains internally (via its Linux-net glue layer)
become readable.
Scope sketch (for a hurd hacker):
- Add a new iioctl routine. Obvious shape:
iioctl_siocgifstats (port_t iface)
-> (struct iioctl_net_device_stats)
returning a fixed-width record:
struct iioctl_net_device_stats {
uint64_t rx_bytes;
uint64_t tx_bytes;
uint64_t rx_packets;
uint64_t tx_packets;
uint64_t rx_errors;
uint64_t tx_errors;
uint64_t rx_dropped;
uint64_t tx_dropped;
};
The Linux glue layer pfinet wraps already maintains these
counters (linux/net/core/dev.c equivalent increments on every
recv and xmit). The patch needs the RPC surface and a
getter, not a new accounting layer.
- Alternative shape: extend pfinet.defs (subsystem 37000) with
a third routine pfinet_get_iface_stats. I prefer the iioctl
form because it parallels the existing per-iface getters
(SIOCGIFFLAGS, SIOCGIFADDR, ...) and reuses the same per-iface
port the caller already holds.
- One of the reserved iioctl.defs slots (0-11, 13, 15, 18,
20-21, 26-32, 36, 38, 40-50, 53-89 are all skip today) is
the right home.
- libc-side: add a SIOCGIFSTATS ioctl macro and a thin glibc
shim that issues the new MIG routine.
Estimation: small patch by absolute size (one MIG entry, one
getter, one libc shim). The risk is in agreeing the record shape
upstream so it does not have to change later.
No patch attached. Happy to test once a draft lands.
Alternative filing channel if a tracker entry is preferred:
https://savannah.gnu.org/bugs/?group=hurd
Thanks,
Borja Tarraso
[email protected]
