|Summary:||igb (I210) + net.inet.ipsec.async_crypto=1 + aesni kill receiving queues and traffic|
|Product:||Base System||Reporter:||Lev A. Serebryakov <lev>|
|Component:||kern||Assignee:||freebsd-net mailing list <net>|
|Status:||In Progress ---|
|Severity:||Affects Only Me||CC:||ae, erj, garga, gnn, sbruno, sigsys|
Description Lev A. Serebryakov 2018-12-03 21:58:49 UTC
I'm benchmarking system with FreeBSD 13-CURRENT (r340933) configured as "VPN border router". It receives traffic via igb0 (I210) and pass it with encryption via IPsec tunnel mode via igb1 (I210 too). CPU supports AES-NI, so it is enabled. I'm using aes-128-gcm cipher. Incoming traffic has characteristics which allow to employ all 4 queues of igb0, which run on 4 cores of 4-core CPU, it could be seen in "top -SH" output and under heavy load queues consume 400% of CPU, 100% each. Everything works as expected. But when I set "sysctl net.inet.ipsec.async_crypto=1" strange thing happens: some igb0 queues stop to work. Sometimes it is 1 queue, sometimes 2 queues, and sometimes 3 (I never seen 4 queues stop to work, though). Stopped queues don't consume CPU and don't pass traffic, so 1/4, 2/4 or 3/4 of incoming traffic is simply lost. Even if traffic is very light (say, 64 packet per second!), traffic which was classified to "stopped" queues is lost. "tcpdump -i igb0" doesn't show it. Only reboot reset this behavior, setting "sysctl net.inet.ipsec.async_crypto=0" doesn't help.
Comment 1 Lev A. Serebryakov 2018-12-04 09:56:08 UTC
I want to additionally clarify, that OUTBOUND traffic via igb1 is encrypted by IPsec, but INBOUND queues of igb0 are blocked. I have "net.inet.ip.redirect" disabled (to enable fast-forward). I have "net.isr.dispatch=direct", as it is set by default. I have "dev.igb.N.iflib.tx_abdicate" disabled, as it is set by default.
Comment 2 Eric Joyner 2018-12-04 21:19:12 UTC
Can you clarify more about what happens when the queues top working? Do the 1-3 queues immediately stop working when you set that sysctl, or do they just randomly stop while receiving traffic? And to confirm, the queues on igb1 aren't affected by this setting?
Comment 3 Lev A. Serebryakov 2018-12-04 21:39:40 UTC
(In reply to Eric Joyner from comment #2) They randomly stop while receiving traffic. It could be several minutes, like all traffic is passing well for 2 minutes, then 1/4 of traffic is lost forever, then after another 30 seconds other 1/4 of traffic is lost (1/2 in total), then 1/2 of traffic could pass for another 3-4-5 minutes, but at least only 3/4 is lost and 3 queues are effectively stopped. There is no panics or any messages in kernel output. I never seen recovery, but, to be honest, I didn't wait more than 5 minutes. I could say, that ifconfig igb0 down && ifconfig igb0 up doesn't help. I could not say about igb1 (outbound interface) for sure, as I've checked inbound one with "tcpdump" and it shows that 1/4-1/2-3/4 of traffic is not seen even by tcpdump. Traffic is generated (with pkt-gen on other end of the link, NOT ON SAME SYSTEM!) with distinct pattern (regular loop over many source IPs), so holes in the traffic is well seen by eye (when packet rate is low). And igb1 only send traffic, it is unidirectional UDP test (as pkt-gen implies). And one more: only sending works. I could turn on "net.inet.ipsec.async_crypto=1" and send traffic from very this system with, say, iperf3 (use it as endpoint with IPsec and not as router), and it make things faster (it is why I've tried to turn it on for routing), and everything works. BTW, I could try to build kernel with INVARIANTS and WITNESSes...
Comment 4 Lev A. Serebryakov 2018-12-04 22:41:21 UTC
(In reply to Eric Joyner from comment #2) Oh. One more: this system has igb2 too (in addition to igb0 and igb1), which is used as management interface (to access system under test for configuration and diagnostics), and it continue to work no problem in both directions for traffic for this host and back (mostly ssh, of course).
Comment 5 Sean Bruno 2018-12-06 19:12:15 UTC
Lev: Can you document your test case here? I'm curious what the two endpoints of the test are and what you're setting up on your machines. I'm fairly ignorant of how to setup ipsec and I'm not sure what that is doing to cause your problems.
Comment 6 Lev A. Serebryakov 2018-12-07 12:37:31 UTC
(In reply to Sean Bruno from comment #5) I have three systems (they are separate physical systems, not VMs). (1) Manager. (2) Device Under Test ("DUT") (3) Mirror. Each system has 3 interfaces. One interface of each system is management one to connect from outside work, and these interfaces is not in scope of this description. Manager system has two interfaces in question: "outbound" and "inbound". - outbound has IP 10.1.0.2/24 and it is connected with "inbound" interface of DUT (via dedicated switch). - inbound has IP 10.10.10.2/24 and it is connected with "outbound" interface of "Mirror". Manager system doesn't have any special routing record. DUT system has two interfaces: "outbound" (igb1 in this ticket) and "inbound" (igb0 in this ticket). - "outbound" (igb1) has IP 10.2.0.1/24 and it is connected with "inbound" interface of "Mirror". - "inbound" (igb0) has IP 10.1.0.1/24 and it is connected with "outbound" interface of "Manager" (via dedicated switch). DUT has routing enabled and has "route -net 10.10.10.0/24 10.2.0.1". DUT has such IPSec settings: ============ add 10.2.0.1 10.2.0.2 esp 0x10001 -m tunnel -E aes-gcm-16 "wxyz0123456789abcdef"; add 10.2.0.1 10.2.0.` esp 0x10002 -m tunnel -E aes-gcm-16 "wxyz0123456789abcdef"; spdadd 10.1.0.0/24 10.10.10.0/24 udp -P out ipsec esp/tunnel/10.2.0.1-10.2.0.2/require; spdadd 10.10.10.0/24 10.1.0.0/24 udp -P in ipsec esp/tunnel/10.2.0.2-10.2.0.1/require; ============ Mirror system has two interfaces in question: "outbound" and "inbound". - outbound has IP 10.10.10.1/24 and it is connected with "inbound" interface of Manager. - inbound has IP 10.2.0.2/24 and it is connected with "outbound" interface of DUT. Mirror has routing enabled and has "route -net 10.1.0.0/24 10.2.0.2". Mirror has static ARP for 10.10.10.2-10.10.10.254 points to "Manager" "Inbound" interface. Mirror has such IPSec settings: ============ add 10.2.0.1 10.2.0.2 esp 0x10001 -m tunnel -E aes-gcm-16 "wxyz0123456789abcdef"; add 10.2.0.1 10.2.0.` esp 0x10002 -m tunnel -E aes-gcm-16 "wxyz0123456789abcdef"; spdadd 10.10.10.0/24 10.1.0.0/24 udp -P out ipsec esp/tunnel/10.2.0.2-10.2.0.1/require; spdadd 10.1.0.0/24 10.10.10.0/24 udp -P in ipsec esp/tunnel/10.2.0.1-10.2.0.2/require; ============ Ok, it is config. Really, it is loop "Manager -> DUT -> Mirror -> Manager" where connection between DUT and Mirror has additional IPsec config. Manager and Mirror are much more powerful than DUT and could pass full-wire-speed traffic without any problems with and without encryption. Now to test. Manager generates (with netmap's pkt-gen) UDP traffic with such characteristics: Transmit interface: "outbound" Dst MAC: DUT "inbound" Src IPs: 10.1.0.2:2000-10.1.0.5:2004 Dst IPs: 10.10.10.2:2000-10.10.10.128:2006 Manager receives all traffic (with netmap's pkt-gen) at "inbound" interface and measure bandwidth. Now, if DUT has default setting for async IPsec (turned off) it could pass 690Mbit/s or 199Kp/s. Any traffic lower than that passes without any losses. For example, if I generate traffic and speed 64P/s (without any prefixes!) I see each and any packet returned to Manager from Mirror via DUT. No problems here. If I turn on async IPsec ("sysctl net.inet.ipsec.async_crypto=1" on DUT), no matter which traffic is generated (I've tested with 64 packets per second, not kilo-packets, simple packets!) receive queues of DUT inbound interface (igb0) stop to work one by one.
Comment 7 Sean Bruno 2018-12-10 14:05:18 UTC
Adding gnn@ and ae@. I'm unsure what igb(4) could be doing here to interfere.