Bug #13938
closed
Kernel panic accessing the GUI over IPsec in certain environments when using nginx ``sendfile`` with unmapped mbufs
Added by Jim Pingle over 1 year ago.
Updated over 1 year ago.
Plus Target Version:
23.05
Description
Under certain conditions which have not yet been identified, it is possible to encounter a kernel kernel panic on FreeBSD main/14.0-CURRENT builds (e.g. Plus 23.01) when attempting to access the GUI over an IPsec tunnel. Thus far we have only received a small number of reports (2) and we have not been able to reproduce the panic in lab conditions.
A community member tracked it down to the use of sendfile
in nginx when used in combination with unmapped mbufs (kern.ipc.mb_use_ext_pgs=1
) both of which are enabled by default.
Users encountering this crash can take either one of two actions:
1. Disable unmapped mbufs by adding a tunable to set kern.ipc.mb_use_ext_pgs=0
OR
2. Disable sendfile
in nginx as described in https://forum.netgate.com/post/1084590
Full details on the forum thread, including backtraces and textdump archives:
https://forum.netgate.com/topic/176974/web-gui-crashes-after-upgrade-from-22-05-to-23-01
To make searching easier the backtrace this generates is:
Tracing pid 3765 tid 100406 td 0xfffffe00c65a4900
kdb_enter() at kdb_enter+0x32/frame 0xfffffe00c3d6f320
vpanic() at vpanic+0x182/frame 0xfffffe00c3d6f370
panic() at panic+0x43/frame 0xfffffe00c3d6f3d0
trap_fatal() at trap_fatal+0x409/frame 0xfffffe00c3d6f430
trap_pfault() at trap_pfault+0x4f/frame 0xfffffe00c3d6f490
calltrap() at calltrap+0x8/frame 0xfffffe00c3d6f490
--- trap 0xc, rip = 0xffffffff813187ba, rsp = 0xfffffe00c3d6f560, rbp = 0xfffffe00c3d6f560 ---
memcpy_erms() at memcpy_erms+0x10a/frame 0xfffffe00c3d6f560
m_unshare() at m_unshare+0x3de/frame 0xfffffe00c3d6f5e0
esp_output() at esp_output+0x186/frame 0xfffffe00c3d6f6d0
ipsec4_perform_request() at ipsec4_perform_request+0x1d2/frame 0xfffffe00c3d6f760
ipsec4_common_output() at ipsec4_common_output+0xa2/frame 0xfffffe00c3d6f7a0
ip_output() at ip_output+0x99d/frame 0xfffffe00c3d6f8a0
tcp_default_output() at tcp_default_output+0x1d2b/frame 0xfffffe00c3d6fa70
tcp_usr_ready() at tcp_usr_ready+0x1a1/frame 0xfffffe00c3d6fad0
sendfile_iodone() at sendfile_iodone+0x11c/frame 0xfffffe00c3d6fb10
vn_sendfile() at vn_sendfile+0x1663/frame 0xfffffe00c3d6fd70
sys_sendfile() at sys_sendfile+0xf7/frame 0xfffffe00c3d6fe00
amd64_syscall() at amd64_syscall+0x10c/frame 0xfffffe00c3d6ff30
fast_syscall_common() at fast_syscall_common+0xf8/frame 0xfffffe00c3d6ff30
--- syscall (393, FreeBSD ELF64, sys_sendfile), rip = 0x8254b84ba, rsp = 0x8209bed68, rbp = 0x8209bf640 ---
The workarounds used here also seem to apply at least partially to connections over OpenVPN tunnels.
See: https://forum.netgate.com/post/1090118
In that case there is no kernel panic but it seems to crash nginx requiring a reboot.
- Status changed from New to Feedback
- % Done changed from 0 to 100
- Assignee changed from Mateusz Guzik to Christian McDonald
We will now disable sendfile mode. Sendfile has little to no benefit for us on pfSense.
This feature of nginx has been problematic upstream for a while, with it being broken and fixed several times.
Sendfile is really only useful when serving static files from UFS filesystem.
- Status changed from Feedback to Resolved
sendfile
is off
in all nginx
configurations now, for the GUI and Captive Portal.
Seeing as this is a bug in mbuf handling, I would argue the thing to do is to flip the unmapped buf support off -- there may be other programs out there using sendfile, no point of them crashing the system.
- Status changed from Resolved to Feedback
- Status changed from Feedback to Resolved
kern.ipc.mb_use_ext_pgs has been disabled for 2 weeks now.
Marking as resolved.
Also available in: Atom
PDF