netconsole hands every console message to every enabled target, with no
bound on the rate.
At Meta, a few hosts caused some DoS and netconsd was OOM killed in
some regions, fed by 3 billion RCU messages from thousand hosts and by
a swap fault that retries forever, which logged 6 million copies of one
line on a single host in a few hours.
Both floods are being fixed where they are printed, by Paul and me for RCU
Tasks [1][2] and by me for the swap fault fix [3], but that only ever
covers the flood already found, not new issues that might be happening.
Use the ratelimit in the kernel to ratelimit messages from netconsole.
Give each target a token bucket, consulted once per message so that a
message split into several ncfrag packets is sent whole or dropped
whole.
It starts with a zero interval, which struct ratelimit_state treats as
unlimited, so nothing changes until ratelimit_interval_ms and
ratelimit_burst are written.
What the bucket discards leaves nothing on the wire. An extended target
sees the loss as a gap in the sequence number the header carries:
7.2.0-rc7-01460-g75848acaf136,13,628,252392647,-;netconsole selftest:
netcons_AGujw 1
7.2.0-rc7-01460-g75848acaf136,13,629,252392782,-;netconsole selftest:
netcons_AGujw 2
7.2.0-rc7-01460-g75848acaf136,13,678,254742125,-;netconsole selftest:
netcons_AGujw 1
Crash output is exempt. The bucket is skipped once
netconsole_kernel_dying() is true, which the first patch adds:
oops_in_progress covers panic(), TAINT_DIE covers an oops or a BUG(),
whose records only reach netconsole from the printer thread, after
oops_end() cleared the flag.
A target that has seen the kernel die stays unlimited until reboot.
[1] https://lore.kernel.org/all/[email protected]/
[2]
https://lore.kernel.org/all/[email protected]/
[3] https://lore.kernel.org/all/[email protected]/
Signed-off-by: Breno Leitao <[email protected]>
---
Changes in v3:
- New first patch, a fix that stands on its own: oops_only=1 receives
nothing on an oops today, because the backtrace reaches netconsole
from the printer thread, after oops_end() cleared oops_in_progress.
It tests TAINT_DIE too, behind netconsole_kernel_dying(), which the
bucket then reuses so that a limit cannot truncate a crash either.
(Sashiko)
- Restart the interval when either knob is written. A target is
configured while it floods, with the bucket already empty, and the
new limit only applied one interval later. (Sashiko)
- Reject an interval above INT_MAX ms. msecs_to_jiffies() saturates
below INT_MAX on 32-bit, so checking the jiffies value alone took a
write there that a 64-bit kernel turns down. (Sashiko)
- List both files in the target parameter table, and qualify the rule
underneath it, which says that only a disabled target can be updated.
(Sashiko)
- Link to v2:
https://patch.msgid.link/[email protected]
Changes in v2:
- Drop the patch that sent a "N messages dropped by rate limit" notice
to the target. (Gustavo Luiz Duarte).
- Link to v1:
https://patch.msgid.link/[email protected]
---
Breno Leitao (5):
netconsole: send the oops when oops_only is set
netconsole: add a per-target message rate limit
netconsole: allow configuring the rate limit interval through configfs
netconsole: allow configuring the rate limit burst through configfs
docs: netconsole: document rate limit feature
Documentation/networking/netconsole.rst | 71 ++++++++++++++----
drivers/net/netconsole.c | 126 +++++++++++++++++++++++++++++++-
2 files changed, 181 insertions(+), 16 deletions(-)
---
base-commit: c68a982815dcce5464e3bf2a31ac94f5146c04ca
change-id: 20260817-netcons_ratelimit-629b04a73c57
Best regards,
--
Breno Leitao <[email protected]>
