Date: Mon, 2 May 2011 23:01:06 -0700 From: Jeremy Chadwick <freebsd@jdc.parodius.com> To: Vlad Galu <dudu@dudu.ro> Cc: freebsd-stable@freebsd.org, freebsd-pf@freebsd.org Subject: Re: RELENG_8 pf stack issue (state count spiraling out of control) Message-ID: <20110503060106.GA36331@icarus.home.lan> In-Reply-To: <BANLkTi=R9fQQZQmLG7HJH%2Bfm3xrJv=HPqA@mail.gmail.com> References: <20110503015854.GA31444@icarus.home.lan> <BANLkTi=R9fQQZQmLG7HJH%2Bfm3xrJv=HPqA@mail.gmail.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Tue, May 03, 2011 at 07:22:10AM +0200, Vlad Galu wrote: > On Tue, May 3, 2011 at 3:58 AM, Jeremy Chadwick <freebsd@jdc.parodius.com>wrote: > > > (Please keep me CC'd as I'm not subscribed to freebsd-pf. And apologies > > for cross-posting, but the issue is severe enough that I wanted to make > > it known on -stable) > > > > The below issue I'm describing is from a machine running 8.2-PRERELEASE > > (RELENG_8) using src dated Tue Feb 15 05:46:02 PST 2011. > > > > Please read the story in full, as I have taken the time to describe > > everything I did, plus log output, as well as induce a panic via "call > > doadump" from ddb so I have a capture of the system at the time. I also > > have a theory as to what caused the problem, but how to trigger it is > > unknown; it may be a rare race condition. > > > > > > This morning I woke up to find a report from one of our users that he > > could not connect to a specific TCP port (not SSH) on one of our > > servers. I also found that I couldn't SSH into the same box. Serial > > console was working fine, and the serial console log showed no sign of > > any problems. > > > > I started to debug the issue of me not being able to SSH into the > > machine and within a few minutes became immediately concerned: pfctl > > indicated we had reached the maximum number permitted state table > > entries (10,000). > > > > ============================================================ > > # pfctl -s info > > Status: Enabled for 76 days 06:49:10 Debug: Urgent > > > > Interface Stats for em0 IPv4 IPv6 > > Bytes In 8969748840 0 > > Bytes Out 8296135477 0 > > Packets In > > Passed 128211763 0 > > Blocked 621379 0 > > Packets Out > > Passed 138483868 0 > > Blocked 2579 0 > > > > State Table Total Rate > > current entries 10000 > > searches 267316807 40.6/s > > inserts 4440553 0.7/s > > removals 4430553 0.7/s > > Counters > > match 5067474 0.8/s > > bad-offset 0 0.0/s > > fragment 324 0.0/s > > short 0 0.0/s > > normalize 32 0.0/s > > memory 336946 0.1/s > > bad-timestamp 0 0.0/s > > congestion 0 0.0/s > > ip-option 0 0.0/s > > proto-cksum 1611 0.0/s > > state-mismatch 509 0.0/s > > state-insert 0 0.0/s > > state-limit 0 0.0/s > > src-limit 0 0.0/s > > synproxy 0 0.0/s > > > > # pfctl -s memory > > states hard limit 10000 > > src-nodes hard limit 10000 > > frags hard limit 5000 > > tables hard limit 1000 > > table-entries hard limit 100000 > > ============================================================ > > > > The above is mainly for em0 (our WAN interface); our LAN interface (em1) > > was not impacted because we use "set skip on em1". And it's a good > > thing too: we have lots of LAN-based services that this machine provides > > that would have been impacted. We also use "set skip on lo0". > > > > I immediately went to look at our monitoring graphs, which monitor pf > > state (specifically state table entries), polled via bsnmpd(8). This > > data is even more frightening: > > > > http://jdc.parodius.com/freebsd/pf-issue/pf_states-day.png > > http://jdc.parodius.com/freebsd/pf-issue/pf_states-week.png > > > > Literally something was spiraling out of control, starting at approx. > > 2011/05/01 (Sun) at 12:30 PDT. The situation became dire at approx. > > 19:45 PDT the same day, but I wasn't aware of it until said user brought > > an issue to my attention. > > > > You can see from the network I/O graphs (taken from SNMP polling our > > switch, NOT from the host/box itself) that there was no DoS attack or > > anything like that occurring -- this was something within FreeBSD > > itself. More evidence of that will become apparent. > > > > http://jdc.parodius.com/freebsd/pf-issue/port_03-day.png > > http://jdc.parodius.com/freebsd/pf-issue/port_03-week.png > > > > The first thing I did was "/etc/rc.d/pf reload". This command hung. > > Any attempt to send Ctrl-C/SIGINT did nothing. I was able to > > Ctrl-Z/SIGSTOP it, then use kill %1, but the actual reload process did > > not truly die (despite csh stating "Terminated"). The only way to kill > > it was to kill -9. > > > > Attempts to shut down any daemons which utilised the network -- > > including things that only used em1 -- would not shut down. This > > included things like postfix, mysqld, and some inet-based services. I > > was forced to kill -9 them. Things like bsnmpd, however, did shut down. > > > > Equally as uncomfortable, "shutdown -r now" did not reboot the system. > > That is to say, wall(1)'s announcement was shown, but the actual > > stopping of services did not begin. > > > > The next thing I tried was "/etc/rc.d/pf stop", which worked. Then I > > did "/etc/rc.d/pf start", which also worked. However, what I saw next > > surely indicated a bug in the pf layer somewhere -- "pfctl -s states" > > and "pfctl -s info" disagreed on the state count: > > > > ============================================================ > > # pfctl -s info > > Status: Enabled for 0 days 00:00:16 Debug: Urgent > > > > Interface Stats for em0 IPv4 IPv6 > > Bytes In 3459 0 > > Bytes Out 0 0 > > Packets In > > Passed 0 0 > > Blocked 29 0 > > Packets Out > > Passed 0 0 > > Blocked 0 0 > > > > State Table Total Rate > > current entries 10000 > > searches 29 1.8/s > > inserts 0 0.0/s > > removals 0 0.0/s > > Counters > > match 29 1.8/s > > bad-offset 0 0.0/s > > fragment 0 0.0/s > > short 0 0.0/s > > normalize 0 0.0/s > > memory 18 1.1/s > > bad-timestamp 0 0.0/s > > congestion 0 0.0/s > > ip-option 0 0.0/s > > proto-cksum 0 0.0/s > > state-mismatch 0 0.0/s > > state-insert 0 0.0/s > > state-limit 0 0.0/s > > src-limit 0 0.0/s > > synproxy 0 0.0/s > > > > # pfctl -s state | wc -l > > 0 > > ============================================================ > > > > The "pf uptime" shown above, by the way, matches system uptime. > > > > I then attempted "pfctl -F state", but nothing changed (looked the same > > as above). > > > > Since I could not reboot the box, I was forced to drop to ddb via serial > > console. I did some commands like "ps" and the like, and then "call > > doadump" to induce a kernel panic, and then "reboot" (which worked). > > > > Once the machine came back up, savecore(8) ran, wrote the data out, and > > everything has been fine since. /var/crash/core.txt.0 is ~68KBytes and > > I do not feel comfortable sharing its content publicly, but will be > > happy to hand it to developer(s) who are interested. Relevant tidbits I > > can discern: > > > > ------------------------------------------------------------------------ > > ps -axl > > > > UID PID PPID CPU PRI NI VSZ RSS MWCHAN STAT TT TIME COMMAND > > 0 422 0 0 -16 0 0 0 pftm DL ?? 1362773081:04.00 > > [pfpurge] > > ------------------------------------------------------------------------ > > > > ------------------------------------------------------------------------ > > vmstat -z > > > > ITEM SIZE LIMIT USED FREE REQUESTS > > FAILURES > > pfsrctrpl: 152, 10000, 0, 0, 0, > > 0 > > pfrulepl: 912, 0, 40, 88, 806, > > 0 > > pfstatepl: 392, 10000, 10000, 0, 4440553, > > 341638 > > pfaltqpl: 240, 0, 0, 0, 0, > > 0 > > pfpooladdrpl: 88, 0, 0, 0, 0, > > 0 > > pfrktable: 1296, 1002, 4, 20, 112, > > 0 > > pfrkentry: 216, 100008, 603, 891, 15384, > > 0 > > pfrkentry2: 216, 0, 0, 0, 0, > > 0 > > pffrent: 32, 5050, 0, 303, 1620, > > 0 > > pffrag: 80, 0, 0, 135, 807, > > 0 > > pffrcache: 80, 10035, 0, 0, 0, > > 0 > > pffrcent: 24, 50022, 0, 0, 0, > > 0 > > pfstatescrub: 40, 0, 0, 0, 0, > > 0 > > pfiaddrpl: 120, 0, 0, 0, 0, > > 0 > > pfospfen: 112, 0, 696, 30, 18096, > > 0 > > pfosfp: 40, 0, 407, 97, 10582, > > 0 > > ------------------------------------------------------------------------ > > > > You can see evidence of processes not exiting/doing what they should do > > here: > > > > ------------------------------------------------------------------------ > > fstat > > > > USER CMD PID FD MOUNT INUM MODE SZ|DV R/W > > root shutdown 91155 root / 2 drwxr-xr-x 512 r > > root shutdown 91155 wd / 2 drwxr-xr-x 512 r > > root shutdown 91155 text / 47195 -r-sr-x--- 15912 r > > root shutdown 91155 0 /dev 38 crw------- ttyu0 rw > > root shutdown 91155 1 /dev 38 crw------- ttyu0 rw > > root shutdown 91155 2 /dev 38 crw------- ttyu0 rw > > root sh 91129 root / 2 drwxr-xr-x 512 r > > root sh 91129 wd / 2 drwxr-xr-x 512 r > > root sh 91129 text / 44 -r-xr-xr-x 134848 r > > root sh 91129 0 /dev 38 crw------- ttyu0 rw > > root sh 91129 1* pipe ffffff01e78fc9e0 <-> ffffff01e78fc888 > > 0 rw > > root sh 91129 2 /dev 20 crw-rw-rw- null w > > root shutdown 91115 root / 2 drwxr-xr-x 512 r > > root shutdown 91115 wd /storage 5 drwx------ 37 r > > root shutdown 91115 text / 47195 -r-sr-x--- 15912 r > > root shutdown 91115 0 /dev 38 crw------- ttyu0 rw > > root shutdown 91115 1 /dev 38 crw------- ttyu0 rw > > root shutdown 91115 2 /dev 38 crw------- ttyu0 rw > > root shutdown 91115 3* local dgram ffffff008ff92960 > > root sh 90818 root / 2 drwxr-xr-x 512 r > > root sh 90818 wd / 70659 drwxr-xr-x 512 r > > root sh 90818 text / 44 -r-xr-xr-x 134848 r > > root sh 90818 0 /dev 38 crw------- ttyu0 rw > > root sh 90818 1* pipe ffffff0043f1ecb8 <-> ffffff0043f1eb60 > > 0 rw > > root sh 90818 2 /dev 20 crw-rw-rw- null w > > root csh 90802 root / 2 drwxr-xr-x 512 r > > root csh 90802 wd / 2 drwxr-xr-x 512 r > > root csh 90802 text / 51 -r-xr-xr-x 358752 r > > root csh 90802 15 /dev 38 crw------- ttyu0 rw > > root csh 90802 16 /dev 38 crw------- ttyu0 rw > > root csh 90802 17 /dev 38 crw------- ttyu0 rw > > root csh 90802 18 /dev 38 crw------- ttyu0 rw > > root csh 90802 19 /dev 38 crw------- ttyu0 rw > > ------------------------------------------------------------------------ > > > > No indication of mbuf exhaustion, putting further focus on the pf stack: > > > > ------------------------------------------------------------------------ > > netstat -m > > > > 2054/1786/3840 mbufs in use (current/cache/total) > > 2048/1414/3462/25600 mbuf clusters in use (current/cache/total/max) > > 2048/896 mbuf+clusters out of packet secondary zone in use (current/cache) > > 0/320/320/12800 4k (page size) jumbo clusters in use > > (current/cache/total/max) > > 0/0/0/19200 9k jumbo clusters in use (current/cache/total/max) > > 0/0/0/12800 16k jumbo clusters in use (current/cache/total/max) > > 4609K/4554K/9164K bytes allocated to network (current/cache/total) > > 0/0/0 requests for mbufs denied (mbufs/clusters/mbuf+clusters) > > 0/0/0 requests for jumbo clusters denied (4k/9k/16k) > > 0 requests for sfbufs denied > > 0 requests for sfbufs delayed > > 0 requests for I/O initiated by sendfile > > 0 calls to protocol drain routines > > ------------------------------------------------------------------------ > > > > Here's one piece of core.0.txt which makes no sense to me -- the "rate" > > column. I have a very hard time believing that was the interrupt rate > > of all the relevant devices at the time (way too high). Maybe this data > > becomes wrong only during a coredump? The total column I could believe. > > > > ------------------------------------------------------------------------ > > vmstat -i > > > > interrupt total rate > > irq4: uart0 54768 912 > > irq6: fdc0 1 0 > > irq17: uhci1+ 172 2 > > irq23: uhci3 ehci1+ 2367 39 > > cpu0: timer 13183882632 219731377 > > irq256: em0 260491055 4341517 > > irq257: em1 127555036 2125917 > > irq258: ahci0 225923164 3765386 > > cpu2: timer 13183881837 219731363 > > cpu1: timer 13002196469 216703274 > > cpu3: timer 13183881783 219731363 > > Total 53167869284 886131154 > > ------------------------------------------------------------------------ > > > > Here's what a normal "vmstat -i" shows from the command-line: > > > > # vmstat -i > > interrupt total rate > > irq4: uart0 518 0 > > irq6: fdc0 1 0 > > irq23: uhci3 ehci1+ 145 0 > > cpu0: timer 19041199 1999 > > irq256: em0 614280 64 > > irq257: em1 168529 17 > > irq258: ahci0 355536 37 > > cpu2: timer 19040462 1999 > > cpu1: timer 19040458 1999 > > cpu3: timer 19040454 1999 > > Total 77301582 8119 > > > > We graph many aspects of this box, including CPU load, memory/swap > > usage, etc. and none show any sign that the interrupt rate on all of > > those devices was even remotely out of control. (I would expect to see > > CPU through the roof given the above data) > > > > I have since rebuilt/reinstalled world/kernel on the machine with the > > latest RELENG_8 code (box is now 8.2-STABLE #0: Mon May 2 14:44:18 PDT > > 2011), hoping whatever this was may have been fixed. > > > > As for what I think may have triggered it, but I have no hard evidence > > of such: on April 29th, I changed our pf.conf and did "/etc/rc.d/pf > > reload". The pf.conf change was a single line: > > > > Old: scrub on em0 all > > New: scrub in on em0 all > > > > Why it took the problem approximately 3 days to start is unknown. It's > > the only change we've made to the system (truly/honestly), and it was a > > change to pf.conf. > > > > If anyone has advice (or has seen the above problem), or is interested > > in debugging it -- as I said, I have a vmcore -- I'm happy to assist in > > any way I can. I would hate for someone else to get bit by this, and > > really am hoping its something that has been fixed between February and > > now. > > > > > I'm seeing this as well. You could change your scrub rules so that you > specifically avoid TCP reassembly (that creates states). Thank you very much. This helps -- and I'm also glad someone else has seen this behaviour. That confirms it's not specific to my equipment, which is good. Regarding scrubbing and TCP reassembly (option "reassemble tcp" according to pf.conf): I wasn't under the impression this option was enabled by default. This got me wondering what the defaults actually are (pf.conf(5) is somewhat vague in this regard, but it does state that "fragment reassemble" is enabled by default). The rule I use: scrub in on em0 all Appears to get evaluated into this (per "pfctl -s rules -v"): scrub in on em0 all fragment reassemble Did you mean to tell me to disable the "fragment reassemble" option (which is different from "reassemble tcp")? If so, how do I do that? It looks like I could use a "no scrub" rule, except I can't find any examples of this on the web or in the docs. Or is it just better to remove use of scrub entirely until whatever this is gets fixed? -- | Jeremy Chadwick jdc@parodius.com | | Parodius Networking http://www.parodius.com/ | | UNIX Systems Administrator Mountain View, CA, USA | | Making life hard for others since 1977. PGP 4BD6C0CB |
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20110503060106.GA36331>