Skip site navigation (1)Skip section navigation (2)
Date:      Mon, 5 Apr 2021 00:08:40 +0200
From:      tuexen@freebsd.org
To:        Rick Macklem <rmacklem@uoguelph.ca>
Cc:        "Scheffenegger, Richard" <Richard.Scheffenegger@netapp.com>, Youssef GHORBAL <youssef.ghorbal@pasteur.fr>, "freebsd-net@freebsd.org" <freebsd-net@freebsd.org>
Subject:   Re: NFS Mount Hangs
Message-ID:  <2B189169-C0C9-4DE6-A01A-BE916F10BABA@freebsd.org>
In-Reply-To: <YQXPR0101MB096876B44F33BAD8991B62C8DD789@YQXPR0101MB0968.CANPRD01.PROD.OUTLOOK.COM>
References:  <C643BB9C-6B61-4DAC-8CF9-CE04EA7292D0@tildenparkcapital.com> <3750001D-3F1C-4D9A-A9D9-98BCA6CA65A4@tildenparkcapital.com> <33693DE3-7FF8-4FAB-9A75-75576B88A566@tildenparkcapital.com> <D67AF317-D238-4EC0-8C7F-22D54AD5144C@pasteur.fr> <YQXPR0101MB09684AB7BEFA911213604467DD669@YQXPR0101MB0968.CANPRD01.PROD.OUTLOOK.COM> <C87066D3-BBF1-44E1-8398-E4EB6903B0F2@tildenparkcapital.com> <8E745920-1092-4312-B251-B49D11FE8028@pasteur.fr> <YQXPR0101MB0968C44C7C82A3EB64F384D0DD7B9@YQXPR0101MB0968.CANPRD01.PROD.OUTLOOK.COM> <DEF8564D-0FE9-4C2C-9F3B-9BCDD423377C@freebsd.org> <YQXPR0101MB0968E0A17D8BCACFAF132225DD7A9@YQXPR0101MB0968.CANPRD01.PROD.OUTLOOK.COM> <SN4PR0601MB3728E392BCA494EAD49605FE86789@SN4PR0601MB3728.namprd06.prod.outlook.com> <YQXPR0101MB09686B4F921B96DCAFEBF874DD789@YQXPR0101MB0968.CANPRD01.PROD.OUTLOOK.COM> <765CE1CD-6AAB-4BEF-97C6-C2A1F0FF4AC5@freebsd.org> <YQXPR0101MB096876B44F33BAD8991B62C8DD789@YQXPR0101MB0968.CANPRD01.PROD.OUTLOOK.COM>

next in thread | previous in thread | raw e-mail | index | archive | help
> On 4. Apr 2021, at 22:28, Rick Macklem <rmacklem@uoguelph.ca> wrote:
>=20
> Oops, yes the packet capture is on freefall (forgot to mention =
that;-).
> You should be able to:
> % fetch https://people.freebsd.org/~rmacklem/linuxtofreenfs.pcap
>=20
> Some useful packet #s are:
> 1949 - partitioning starts
> 2005 - partition healed
> 2060 - last RST
> 2067 - SYN -> gets going again
>=20
> This was taken at the Linux end. I have FreeBSD end too, although I
> don't think it tells you anything more.
Hi Rick,

I would like to look at the FreeBSD side, too. Do you also know, what
the state of the TCP connection was when the SYN / ACK / RST game was
going on?
I would like to understand why the reestablishment of the connection
did not work...

Best regards
Michael
>=20
> Have fun with it, rick
>=20
>=20
> ________________________________________
> From: tuexen@freebsd.org <tuexen@freebsd.org>
> Sent: Sunday, April 4, 2021 12:41 PM
> To: Rick Macklem
> Cc: Scheffenegger, Richard; Youssef GHORBAL; freebsd-net@freebsd.org
> Subject: Re: NFS Mount Hangs
>=20
> CAUTION: This email originated from outside of the University of =
Guelph. Do not click links or open attachments unless you recognize the =
sender and know the content is safe. If in doubt, forward suspicious =
emails to IThelp@uoguelph.ca
>=20
>=20
>> On 4. Apr 2021, at 17:27, Rick Macklem <rmacklem@uoguelph.ca> wrote:
>>=20
>> Well, I'm going to cheat and top post, since this is elated info. and
>> not really part of the discussion...
>>=20
>> I've been testing network partitioning between a Linux client (5.2 =
kernel)
>> and a FreeBSD-current NFS server. I have not gotten a solid hang, but
>> I have had the Linux client doing "battle" with the FreeBSD server =
for
>> several minutes after un-partitioning the connection.
>>=20
>> The battle basically consists of the Linux client sending an RST, =
followed
>> by a SYN.
>> The FreeBSD server ignores the RST and just replies with the same old =
ack.
>> --> This varies from "just a SYN" that succeeds to 100+ cycles of the =
above
>>      over several minutes.
>>=20
>> I had thought that an RST was a "pretty heavy hammer", but FreeBSD =
seems
>> pretty good at ignoring it.
>>=20
>> A full packet capture of one of these is in =
/home/rmacklem/linuxtofreenfs.pcap
>> in case anyone wants to look at it.
> On freefall? I would like to take a look at it...
>=20
> Best regards
> Michael
>>=20
>> Here's a tcpdump snippet of the interesting part (see the *** =
comments):
>> 19:10:09.305775 IP nfsv4-new3.home.rick.nfsd > =
nfsv4-linux.home.rick.apex-mesh: Flags [P.], seq 202585:202749, ack =
212293, win 29128, options [nop,nop,TS val 2073636037 ecr 2671204825], =
length 164: NFS reply xid 613153685 reply ok 160 getattr NON 4 ids =
0/33554432 sz 0
>> 19:10:09.305850 IP nfsv4-linux.home.rick.apex-mesh > =
nfsv4-new3.home.rick.nfsd: Flags [.], ack 202749, win 501, options =
[nop,nop,TS val 2671204825 ecr 2073636037], length 0
>> *** Network is now partitioned...
>>=20
>> 19:10:09.407840 IP nfsv4-linux.home.rick.apex-mesh > =
nfsv4-new3.home.rick.nfsd: Flags [P.], seq 212293:212525, ack 202749, =
win 501, options [nop,nop,TS val 2671204927 ecr 2073636037], length 232: =
NFS request xid 629930901 228 getattr fh 0,1/53
>> 19:10:09.615779 IP nfsv4-linux.home.rick.apex-mesh > =
nfsv4-new3.home.rick.nfsd: Flags [P.], seq 212293:212525, ack 202749, =
win 501, options [nop,nop,TS val 2671205135 ecr 2073636037], length 232: =
NFS request xid 629930901 228 getattr fh 0,1/53
>> 19:10:09.823780 IP nfsv4-linux.home.rick.apex-mesh > =
nfsv4-new3.home.rick.nfsd: Flags [P.], seq 212293:212525, ack 202749, =
win 501, options [nop,nop,TS val 2671205343 ecr 2073636037], length 232: =
NFS request xid 629930901 228 getattr fh 0,1/53
>> *** Lots of lines snipped.
>>=20
>>=20
>> 19:13:41.295783 ARP, Request who-has nfsv4-new3.home.rick tell =
nfsv4-linux.home.rick, length 28
>> 19:13:42.319767 ARP, Request who-has nfsv4-new3.home.rick tell =
nfsv4-linux.home.rick, length 28
>> 19:13:46.351966 ARP, Request who-has nfsv4-new3.home.rick tell =
nfsv4-linux.home.rick, length 28
>> 19:13:47.375790 ARP, Request who-has nfsv4-new3.home.rick tell =
nfsv4-linux.home.rick, length 28
>> 19:13:48.399786 ARP, Request who-has nfsv4-new3.home.rick tell =
nfsv4-linux.home.rick, length 28
>> *** Network is now unpartitioned...
>>=20
>> 19:13:48.399990 ARP, Reply nfsv4-new3.home.rick is-at =
d4:be:d9:07:81:72 (oui Unknown), length 46
>> 19:13:48.400002 IP nfsv4-linux.home.rick.apex-mesh > =
nfsv4-new3.home.rick.nfsd: Flags [S], seq 416692300, win 64240, options =
[mss 1460,sackOK,TS val 2671421871 ecr 0,nop,wscale 7], length 0
>> 19:13:48.400185 IP nfsv4-new3.home.rick.nfsd > =
nfsv4-linux.home.rick.apex-mesh: Flags [.], ack 212293, win 29127, =
options [nop,nop,TS val 2073855137 ecr 2671204825], length 0
>> 19:13:48.400273 IP nfsv4-linux.home.rick.apex-mesh > =
nfsv4-new3.home.rick.nfsd: Flags [R], seq 964161458, win 0, length 0
>> 19:13:49.423833 IP nfsv4-linux.home.rick.apex-mesh > =
nfsv4-new3.home.rick.nfsd: Flags [S], seq 416692300, win 64240, options =
[mss 1460,sackOK,TS val 2671424943 ecr 0,nop,wscale 7], length 0
>> 19:13:49.424056 IP nfsv4-new3.home.rick.nfsd > =
nfsv4-linux.home.rick.apex-mesh: Flags [.], ack 212293, win 29127, =
options [nop,nop,TS val 2073856161 ecr 2671204825], length 0
>> *** This "battle" goes on for 223sec...
>>   I snipped out 13 cycles of this "Linux sends an RST, followed by =
SYN"
>>   "FreeBSD replies with same old ACK". In another test run I saw this
>>   cycle continue non-stop for several minutes. This time, the Linux
>>   client paused for a while (see ARPs below).
>>=20
>> 19:13:49.424101 IP nfsv4-linux.home.rick.apex-mesh > =
nfsv4-new3.home.rick.nfsd: Flags [R], seq 964161458, win 0, length 0
>> 19:13:53.455867 IP nfsv4-linux.home.rick.apex-mesh > =
nfsv4-new3.home.rick.nfsd: Flags [S], seq 416692300, win 64240, options =
[mss 1460,sackOK,TS val 2671428975 ecr 0,nop,wscale 7], length 0
>> 19:13:53.455991 IP nfsv4-new3.home.rick.nfsd > =
nfsv4-linux.home.rick.apex-mesh: Flags [.], ack 212293, win 29127, =
options [nop,nop,TS val 2073860193 ecr 2671204825], length 0
>> *** Snipped a bunch of stuff out, mostly ARPs, plus one more RST.
>>=20
>> 19:16:57.775780 ARP, Request who-has nfsv4-new3.home.rick tell =
nfsv4-linux.home.rick, length 28
>> 19:16:57.775937 ARP, Reply nfsv4-new3.home.rick is-at =
d4:be:d9:07:81:72 (oui Unknown), length 46
>> 19:16:57.980240 ARP, Request who-has nfsv4-new3.home.rick tell =
192.168.1.254, length 46
>> 19:16:58.555663 ARP, Request who-has nfsv4-new3.home.rick tell =
192.168.1.254, length 46
>> 19:17:00.104701 IP nfsv4-new3.home.rick.nfsd > =
nfsv4-linux.home.rick.apex-mesh: Flags [F.], seq 202749, ack 212293, win =
29128, options [nop,nop,TS val 2074046846 ecr 2671204825], length 0
>> 19:17:15.664354 IP nfsv4-new3.home.rick.nfsd > =
nfsv4-linux.home.rick.apex-mesh: Flags [F.], seq 202749, ack 212293, win =
29128, options [nop,nop,TS val 2074062406 ecr 2671204825], length 0
>> 19:17:31.239246 IP nfsv4-new3.home.rick.nfsd > =
nfsv4-linux.home.rick.apex-mesh: Flags [R.], seq 202750, ack 212293, win =
0, options [nop,nop,TS val 2074077981 ecr 2671204825], length 0
>> *** FreeBSD finally acknowledges the RST 38sec after Linux sent the =
last
>>   of 13 (100+ for another test run).
>>=20
>> 19:17:51.535979 IP nfsv4-linux.home.rick.apex-mesh > =
nfsv4-new3.home.rick.nfsd: Flags [S], seq 4247692373, win 64240, options =
[mss 1460,sackOK,TS val 2671667055 ecr 0,nop,wscale 7], length 0
>> 19:17:51.536130 IP nfsv4-new3.home.rick.nfsd > =
nfsv4-linux.home.rick.apex-mesh: Flags [S.], seq 661237469, ack =
4247692374, win 65535, options [mss 1460,nop,wscale 6,sackOK,TS val =
2074098278 ecr 2671667055], length 0
>> *** Now back in business...
>>=20
>> 19:17:51.536218 IP nfsv4-linux.home.rick.apex-mesh > =
nfsv4-new3.home.rick.nfsd: Flags [.], ack 1, win 502, options =
[nop,nop,TS val 2671667055 ecr 2074098278], length 0
>> 19:17:51.536295 IP nfsv4-linux.home.rick.apex-mesh > =
nfsv4-new3.home.rick.nfsd: Flags [P.], seq 1:233, ack 1, win 502, =
options [nop,nop,TS val 2671667056 ecr 2074098278], length 232: NFS =
request xid 629930901 228 getattr fh 0,1/53
>> 19:17:51.536346 IP nfsv4-linux.home.rick.apex-mesh > =
nfsv4-new3.home.rick.nfsd: Flags [P.], seq 233:505, ack 1, win 502, =
options [nop,nop,TS val 2671667056 ecr 2074098278], length 272: NFS =
request xid 697039765 132 getattr fh 0,1/53
>> 19:17:51.536515 IP nfsv4-new3.home.rick.nfsd > =
nfsv4-linux.home.rick.apex-mesh: Flags [.], ack 505, win 29128, options =
[nop,nop,TS val 2074098279 ecr 2671667056], length 0
>> 19:17:51.536553 IP nfsv4-linux.home.rick.apex-mesh > =
nfsv4-new3.home.rick.nfsd: Flags [P.], seq 505:641, ack 1, win 502, =
options [nop,nop,TS val 2671667056 ecr 2074098279], length 136: NFS =
request xid 730594197 132 getattr fh 0,1/53
>> 19:17:51.536562 IP nfsv4-new3.home.rick.nfsd > =
nfsv4-linux.home.rick.apex-mesh: Flags [P.], seq 1:49, ack 505, win =
29128, options [nop,nop,TS val 2074098279 ecr 2671667056], length 48: =
NFS reply xid 697039765 reply ok 44 getattr ERROR: unk 10063
>>=20
>> This error 10063 after the partition heals is also "bad news". It =
indicates the Session
>> (which is supposed to maintain "exactly once" RPC semantics is =
broken). I'll admit I
>> suspect a Linux client bug, but will be investigating further.
>>=20
>> So, hopefully TCP conversant folk can confirm if the above is correct =
behaviour
>> or if the RST should be ack'd sooner?
>>=20
>> I could also see this becoming a "forever" TCP battle for other =
versions of Linux client.
>>=20
>> rick
>>=20
>>=20
>> ________________________________________
>> From: Scheffenegger, Richard <Richard.Scheffenegger@netapp.com>
>> Sent: Sunday, April 4, 2021 7:50 AM
>> To: Rick Macklem; tuexen@freebsd.org
>> Cc: Youssef GHORBAL; freebsd-net@freebsd.org
>> Subject: Re: NFS Mount Hangs
>>=20
>> CAUTION: This email originated from outside of the University of =
Guelph. Do not click links or open attachments unless you recognize the =
sender and know the content is safe. If in doubt, forward suspicious =
emails to IThelp@uoguelph.ca
>>=20
>>=20
>> For what it=E2=80=98s worth, suse found two bugs in the linux =
nfconntrack (stateful firewall), and pfifo-fast scheduler, which could =
conspire to make tcp sessions hang forever.
>>=20
>> One is a missed updaten when the c=C3=B6ient is not using the =
noresvport moint option, which makes tje firewall think rsts are illegal =
(and drop them);
>>=20
>> The fast scheduler can run into an issue if only a single packet =
should be forwarded (note that this is not the default scheduler, but =
often recommended for perf, as it runs lockless and lower cpu cost that =
pfq (default). If no other/additional packet pushes out that last packet =
of a flow, it can become stuck forever...
>>=20
>> I can try getting the relevant bug info next week...
>>=20
>> ________________________________
>> Von: owner-freebsd-net@freebsd.org <owner-freebsd-net@freebsd.org> im =
Auftrag von Rick Macklem <rmacklem@uoguelph.ca>
>> Gesendet: Friday, April 2, 2021 11:31:01 PM
>> An: tuexen@freebsd.org <tuexen@freebsd.org>
>> Cc: Youssef GHORBAL <youssef.ghorbal@pasteur.fr>; =
freebsd-net@freebsd.org <freebsd-net@freebsd.org>
>> Betreff: Re: NFS Mount Hangs
>>=20
>> NetApp Security WARNING: This is an external email. Do not click =
links or open attachments unless you recognize the sender and know the =
content is safe.
>>=20
>>=20
>>=20
>>=20
>> tuexen@freebsd.org wrote:
>>>> On 2. Apr 2021, at 02:07, Rick Macklem <rmacklem@uoguelph.ca> =
wrote:
>>>>=20
>>>> I hope you don't mind a top post...
>>>> I've been testing network partitioning between the only Linux =
client
>>>> I have (5.2 kernel) and a FreeBSD server with the xprtdied.patch
>>>> (does soshutdown(..SHUT_WR) when it knows the socket is broken)
>>>> applied to it.
>>>>=20
>>>> I'm not enough of a TCP guy to know if this is useful, but here's =
what
>>>> I see...
>>>>=20
>>>> While partitioned:
>>>> On the FreeBSD server end, the socket either goes to CLOSED during
>>>> the network partition or stays ESTABLISHED.
>>> If it goes to CLOSED you called shutdown(, SHUT_WR) and the peer =
also
>>> sent a FIN, but you never called close() on the socket.
>>> If the socket stays in ESTABLISHED, there is no communication =
ongoing,
>>> I guess, and therefore the server does not even detect that the peer
>>> is not reachable.
>>>> On the Linux end, the socket seems to remain ESTABLISHED for a
>>>> little while, and then disappears.
>>> So how does Linux detect the peer is not reachable?
>> Well, here's what I see in a packet capture in the Linux client once
>> I partition it (just unplug the net cable):
>> - lots of retransmits of the same segment (with ACK) for 54sec
>> - then only ARP queries
>>=20
>> Once I plug the net cable back in:
>> - ARP works
>> - one more retransmit of the same segement
>> - receives RST from FreeBSD
>> ** So, is this now a "new" TCP connection, despite
>>   using the same port#.
>>   --> It matters for NFS, since "new connection"
>>          implies "must retry all outstanding RPCs".
>> - sends SYN
>> - receives SYN, ACK from FreeBSD
>> --> connection starts working again
>> Always uses same port#.
>>=20
>> On the FreeBSD server end:
>> - receives the last retransmit of the segment (with ACK)
>> - sends RST
>> - receives SYN
>> - sends SYN, ACK
>>=20
>> I thought that there was no RST in the capture I looked at
>> yesterday, so I'm not sure if FreeBSD always sends an RST,
>> but the Linux client behaviour was the same. (Sent a SYN, etc).
>> The socket disappears from the Linux "netstat -a" and I
>> suspect that happens after about 54sec, but I am not sure
>> about the timing.
>>=20
>>>>=20
>>>> After unpartitioning:
>>>> On the FreeBSD server end, you get another socket showing up at
>>>> the same port#
>>>> Active Internet connections (including servers)
>>>> Proto Recv-Q Send-Q Local Address          Foreign Address        =
(state)
>>>> tcp4       0      0 nfsv4-new3.nfsd        nfsv4-linux.678        =
ESTABLISHED
>>>> tcp4       0      0 nfsv4-new3.nfsd        nfsv4-linux.678        =
CLOSED
>>>>=20
>>>> The Linux client shows the same connection ESTABLISHED.
>> But disappears from "netstat -a" for a while during the partitioning.
>>=20
>>>> (The mount sometimes reports an error. I haven't looked at packet
>>>> traces to see if it retries RPCs or why the errors occur.)
>> I have now done so, as above.
>>=20
>>>> --> However I never get hangs.
>>>> Sometimes it goes to SYN_SENT for a while and the FreeBSD server
>>>> shows FIN_WAIT_1, but then both ends go to ESTABLISHED and the
>>>> mount starts working again.
>>>>=20
>>>> The most obvious thing is that the Linux client always keeps using
>>>> the same port#. (The FreeBSD client will use a different port# when
>>>> it does a TCP reconnect after no response from the NFS server for
>>>> a little while.)
>>>>=20
>>>> What do those TCP conversant think?
>>> I guess you are you are never calling close() on the socket, for =
with
>>> the connection state is CLOSED.
>> Ok, that makes sense. For this case the Linux client has not done a
>> BindConnectionToSession to re-assign the back channel.
>> I'll have to bug them about this. However, I'll bet they'll answer
>> that I have to tell them the back channel needs re-assignment
>> or something like that.
>>=20
>> I am pretty certain they are broken, in that the client needs to
>> retry all outstanding RPCs.
>>=20
>> For others, here's the long winded version of this that I just
>> put on the phabricator review:
>> In the server side kernel RPC, the socket (struct socket *) is in a
>> structure called SVCXPRT (normally pointed to by "xprt").
>> These structures a ref counted and the soclose() is done
>> when the ref. cnt goes to zero. My understanding is that
>> "struct socket *" is free'd by soclose() so this cannot be done
>> before the xprt ref. cnt goes to zero.
>>=20
>> For NFSv4.1/4.2 there is something called a back channel
>> which means that a "xprt" is used for server->client RPCs,
>> although the TCP connection is established by the client
>> to the server.
>> --> This back channel holds a ref cnt on "xprt" until the
>>=20
>>    client re-assigns it to a different TCP connection
>>    via an operation called BindConnectionToSession
>>    and the Linux client is not doing this soon enough,
>>   it appears.
>>=20
>> So, the soclose() is delayed, which is why I think the
>> TCP connection gets stuck in CLOSE_WAIT and that is
>> why I've added the soshutdown(..SHUT_WR) calls,
>> which can happen before the client gets around to
>> re-assigning the back channel.
>>=20
>> Thanks for your help with this Michael, rick
>>=20
>> Best regards
>> Michael
>>>=20
>>> rick
>>> ps: I can capture packets while doing this, if anyone has a use
>>>    for them.
>>>=20
>>>=20
>>>=20
>>>=20
>>>=20
>>>=20
>>> ________________________________________
>>> From: owner-freebsd-net@freebsd.org <owner-freebsd-net@freebsd.org> =
on behalf of Youssef  GHORBAL <youssef.ghorbal@pasteur.fr>
>>> Sent: Saturday, March 27, 2021 6:57 PM
>>> To: Jason Breitman
>>> Cc: Rick Macklem; freebsd-net@freebsd.org
>>> Subject: Re: NFS Mount Hangs
>>>=20
>>> CAUTION: This email originated from outside of the University of =
Guelph. Do not click links or open attachments unless you recognize the =
sender and know the content is safe. If in doubt, forward suspicious =
emails to IThelp@uoguelph.ca
>>>=20
>>>=20
>>>=20
>>>=20
>>> On 27 Mar 2021, at 13:20, Jason Breitman =
<jbreitman@tildenparkcapital.com<mailto:jbreitman@tildenparkcapital.com>> =
wrote:
>>>=20
>>> The issue happened again so we can say that disabling TSO and LRO on =
the NIC did not resolve this issue.
>>> # ifconfig lagg0 -rxcsum -rxcsum6 -txcsum -txcsum6 -lro -tso =
-vlanhwtso
>>> # ifconfig lagg0
>>> lagg0: flags=3D8943<UP,BROADCAST,RUNNING,PROMISC,SIMPLEX,MULTICAST> =
metric 0 mtu 1500
>>>      =
options=3D8100b8<VLAN_MTU,VLAN_HWTAGGING,JUMBO_MTU,VLAN_HWCSUM,VLAN_HWFILT=
ER>
>>>=20
>>> We can also say that the sysctl settings did not resolve this issue.
>>>=20
>>> # sysctl net.inet.tcp.fast_finwait2_recycle=3D1
>>> net.inet.tcp.fast_finwait2_recycle: 0 -> 1
>>>=20
>>> # sysctl net.inet.tcp.finwait2_timeout=3D1000
>>> net.inet.tcp.finwait2_timeout: 60000 -> 1000
>>>=20
>>> I don=E2=80=99t think those will do anything in your case since the =
FIN_WAIT2 are on the client side and those sysctls are for BSD.
>>> By the way it seems that Linux recycles automatically TCP sessions =
in FIN_WAIT2 after 60 seconds (sysctl net.ipv4.tcp_fin_timeout)
>>>=20
>>> tcp_fin_timeout (integer; default: 60; since Linux 2.2)
>>>            This specifies how many seconds to wait for a final FIN
>>>            packet before the socket is forcibly closed.  This is
>>>            strictly a violation of the TCP specification, but
>>>            required to prevent denial-of-service attacks.  In Linux
>>>            2.2, the default value was 180.
>>>=20
>>> So I don=E2=80=99t get why it stucks in the FIN_WAIT2 state anyway.
>>>=20
>>> You really need to have a packet capture during the outage (client =
and server side) so you=E2=80=99ll get over the wire chat and start =
speculating from there.
>>> No need to capture the beginning of the outage for now. All you have =
to do, is run a tcpdump for 10 minutes or so when you notice a client =
stuck.
>>>=20
>>> * I have not rebooted the NFS Server nor have I restarted nfsd, but =
do not believe that is required as these settings are at the TCP level =
and I would expect new sessions to use the updated settings.
>>>=20
>>> The issue occurred after 5 days following a reboot of the client =
machines.
>>> I ran the capture information again to make use of the situation.
>>>=20
>>> #!/bin/sh
>>>=20
>>> while true
>>> do
>>> /bin/date >> /tmp/nfs-hang.log
>>> /bin/ps axHl | grep nfsd | grep -v grep >> /tmp/nfs-hang.log
>>> /usr/bin/procstat -kk 2947 >> /tmp/nfs-hang.log
>>> /usr/bin/procstat -kk 2944 >> /tmp/nfs-hang.log
>>> /bin/sleep 60
>>> done
>>>=20
>>>=20
>>> On the NFS Server
>>> Active Internet connections (including servers)
>>> Proto Recv-Q Send-Q Local Address          Foreign Address        =
(state)
>>> tcp4       0      0 NFS.Server.IP.X.2049      NFS.Client.IP.X.48286  =
   CLOSE_WAIT
>>>=20
>>> On the NFS Client
>>> tcp        0      0 NFS.Client.IP.X:48286      NFS.Server.IP.X:2049  =
     FIN_WAIT2
>>>=20
>>>=20
>>>=20
>>> You had also asked for the output below.
>>>=20
>>> # nfsstat -E -s
>>> BackChannelCtBindConnToSes
>>>          0            0
>>>=20
>>> # sysctl vfs.nfsd.request_space_throttle_count
>>> vfs.nfsd.request_space_throttle_count: 0
>>>=20
>>> I see that you are testing a patch and I look forward to seeing the =
results.
>>>=20
>>>=20
>>> Jason Breitman
>>>=20
>>>=20
>>> On Mar 21, 2021, at 6:21 PM, Rick Macklem =
<rmacklem@uoguelph.ca<mailto:rmacklem@uoguelph.ca>> wrote:
>>>=20
>>> Youssef GHORBAL =
<youssef.ghorbal@pasteur.fr<mailto:youssef.ghorbal@pasteur.fr>> wrote:
>>>> Hi Jason,
>>>>=20
>>>>> On 17 Mar 2021, at 18:17, Jason Breitman =
<jbreitman@tildenparkcapital.com<mailto:jbreitman@tildenparkcapital.com>> =
wrote:
>>>>>=20
>>>>> Please review the details below and let me know if there is a =
setting that I should apply to my FreeBSD NFS Server or if there is a =
bug fix that I can apply to resolve my issue.
>>>>> I shared this information with the linux-nfs mailing list and they =
believe the issue is on the server side.
>>>>>=20
>>>>> Issue
>>>>> NFSv4 mounts periodically hang on the NFS Client.
>>>>>=20
>>>>> During this time, it is possible to manually mount from another =
NFS Server on the NFS Client having issues.
>>>>> Also, other NFS Clients are successfully mounting from the NFS =
Server in question.
>>>>> Rebooting the NFS Client appears to be the only solution.
>>>>=20
>>>> I had experienced a similar weird situation with periodically stuck =
Linux NFS clients >mounting Isilon NFS servers (Isilon is FreeBSD based =
but they seem to have there >own nfsd)
>>> Yes, my understanding is that Isilon uses a proprietary user space =
nfsd and
>>> not the kernel based RPC and nfsd in FreeBSD.
>>>=20
>>>> We=E2=80=99ve had better luck and we did manage to have packet =
captures on both sides >during the issue. The gist of it goes like =
follows:
>>>>=20
>>>> - Data flows correctly between SERVER and the CLIENT
>>>> - At some point SERVER starts decreasing it's TCP Receive Window =
until it reachs 0
>>>> - The client (eager to send data) can only ack data sent by SERVER.
>>>> - When SERVER was done sending data, the client starts sending TCP =
Window >Probes hoping that the TCP Window opens again so he can flush =
its buffers.
>>>> - SERVER responds with a TCP Zero Window to those probes.
>>> Having the window size drop to zero is not necessarily incorrect.
>>> If the server is overloaded (has a backlog of NFS requests), it can =
stop doing
>>> soreceive() on the socket (so the socket rcv buffer can fill up and =
the TCP window
>>> closes). This results in "backpressure" to stop the NFS client from =
flooding the
>>> NFS server with requests.
>>> --> However, once the backlog is handled, the nfsd should start to =
soreceive()
>>> again and this shouls cause the window to open back up.
>>> --> Maybe this is broken in the socket/TCP code. I quickly got lost =
in
>>> tcp_output() when it decides what to do about the rcvwin.
>>>=20
>>>> - After 6 minutes (the NFS server default Idle timeout) SERVER =
racefully closes the >TCP connection sending a FIN Packet (and still a =
TCP Window 0)
>>> This probably does not happen for Jason's case, since the 6minute =
timeout
>>> is disabled when the TCP connection is assigned as a backchannel =
(most likely
>>> the case for NFSv4.1).
>>>=20
>>>> - CLIENT ACK that FIN.
>>>> - SERVER goes in FIN_WAIT_2 state
>>>> - CLIENT closes its half part part of the socket and goes in =
LAST_ACK state.
>>>> - FIN is never sent by the client since there still data in its =
SendQ and receiver TCP >Window is still 0. At this stage the client =
starts sending TCP Window Probes again >and again hoping that the server =
opens its TCP Window so it can flush it's buffers >and terminate its =
side of the socket.
>>>> - SERVER keeps responding with a TCP Zero Window to those probes.
>>>> =3D> The last two steps goes on and on for hours/days freezing the =
NFS mount bound >to that TCP session.
>>>>=20
>>>> If we had a situation where CLIENT was responsible for closing the =
TCP Window (and >initiating the TCP FIN first) and server wanting to =
send data we=E2=80=99ll end up in the same >state as you I think.
>>>>=20
>>>> We=E2=80=99ve never had the root cause of why the SERVER decided to =
close the TCP >Window and no more acccept data, the fix on the Isilon =
part was to recycle more >aggressively the FIN_WAIT_2 sockets =
(net.inet.tcp.fast_finwait2_recycle=3D1 & =
>net.inet.tcp.finwait2_timeout=3D5000). Once the socket recycled and at =
the next >occurence of CLIENT TCP Window probe, SERVER sends a RST, =
triggering the >teardown of the session on the client side, a new TCP =
handchake, etc and traffic >flows again (NFS starts responding)
>>>>=20
>>>> To avoid rebooting the client (and before the aggressive FIN_WAIT_2 =
was >implemented on the Isilon side) we=E2=80=99ve added a check script =
on the client that detects >LAST_ACK sockets on the client and through =
iptables rule enforces a TCP RST, >Something like: -A OUTPUT -p tcp -d =
$nfs_server_addr --sport $local_port -j REJECT >--reject-with tcp-reset =
(the script removes this iptables rule as soon as the LAST_ACK =
>disappears)
>>>>=20
>>>> The bottom line would be to have a packet capture during the outage =
(client and/or >server side), it will show you at least the shape of the =
TCP exchange when NFS is >stuck.
>>> Interesting story and good work w.r.t. sluething, Youssef, thanks.
>>>=20
>>> I looked at Jason's log and it shows everything is ok w.r.t the nfsd =
threads.
>>> (They're just waiting for RPC requests.)
>>> However, I do now think I know why the soclose() does not happen.
>>> When the TCP connection is assigned as a backchannel, that takes a =
reference
>>> cnt on the structure. This refcnt won't be released until the =
connection is
>>> replaced by a BindConnectiotoSession operation from the client. But =
that won't
>>> happen until the client creates a new TCP connection.
>>> --> No refcnt release-->no refcnt of 0-->no soclose().
>>>=20
>>> I've created the attached patch (completely different from the =
previous one)
>>> that adds soshutdown(SHUT_WR) calls in the three places where the =
TCP
>>> connection is going away. This seems to get it past CLOSE_WAIT =
without a
>>> soclose().
>>> --> I know you are not comfortable with patching your server, but I =
do think
>>> this change will get the socket shutdown to complete.
>>>=20
>>> There are a couple more things you can check on the server...
>>> # nfsstat -E -s
>>> --> Look for the count under "BindConnToSes".
>>> --> If non-zero, backchannels have been assigned
>>> # sysctl -a | fgrep request_space_throttle_count
>>> --> If non-zero, the server has been overloaded at some point.
>>>=20
>>> I think the attached patch might work around the problem.
>>> The code that should open up the receive window needs to be checked.
>>> I am also looking at enabling the 6minute timeout when a backchannel =
is
>>> assigned.
>>>=20
>>> rick
>>>=20
>>> Youssef
>>>=20
>>> _______________________________________________
>>> freebsd-net@freebsd.org<mailto:freebsd-net@freebsd.org> mailing list
>>> =
https://urldefense.com/v3/__https://lists.freebsd.org/mailman/listinfo/fre=
ebsd-net__;!!JFdNOqOXpB6UZW0!_c2MFNbir59GXudWPVdE5bNBm-qqjXeBuJ2UEmFv5OZci=
Lj4ObR_drJNv5yryaERfIbhKR2d$
>>> To unsubscribe, send any mail to =
"freebsd-net-unsubscribe@freebsd.org<mailto:freebsd-net-unsubscribe@freebs=
d.org>"
>>> <xprtdied.patch>
>>>=20
>>> <nfs-hang.log.gz>
>>>=20
>>> _______________________________________________
>>> freebsd-net@freebsd.org mailing list
>>> https://lists.freebsd.org/mailman/listinfo/freebsd-net
>>> To unsubscribe, send any mail to =
"freebsd-net-unsubscribe@freebsd.org"
>>> _______________________________________________
>>> freebsd-net@freebsd.org mailing list
>>> https://lists.freebsd.org/mailman/listinfo/freebsd-net
>>> To unsubscribe, send any mail to =
"freebsd-net-unsubscribe@freebsd.org"
>>=20
>> _______________________________________________
>> freebsd-net@freebsd.org mailing list
>> https://lists.freebsd.org/mailman/listinfo/freebsd-net
>> To unsubscribe, send any mail to =
"freebsd-net-unsubscribe@freebsd.org"
>> _______________________________________________
>> freebsd-net@freebsd.org mailing list
>> https://lists.freebsd.org/mailman/listinfo/freebsd-net
>> To unsubscribe, send any mail to =
"freebsd-net-unsubscribe@freebsd.org"
>=20




Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?2B189169-C0C9-4DE6-A01A-BE916F10BABA>