On Tue, Apr 13, 2021 at 02:59:27PM -0400, Olga Kornievskaia wrote: > On Tue, Apr 13, 2021 at 1:17 PM J. Bruce Fields <bfields@xxxxxxxxxxxx> wrote: > > > > On Tue, Apr 13, 2021 at 09:31:37AM -0400, Olga Kornievskaia wrote: > > > On Tue, Apr 13, 2021 at 3:08 AM Rick Macklem <rmacklem@xxxxxxxxxxx> wrote: > > > > > > > > Hi, > > > > > > > > During testing of a Fedora Core 30 (5.2.10 kernel) against a FreeBSD > > > > server (4.1 mount), I have been simulating a network partitioning > > > > for a few minutes (until the TCP connection goes to SYN_SENT on > > > > the Linux client). > > > > > > > > Sometimes, after the network partition heals, the FreeBSD server > > > > replies NFS4ERR_SEQ_MISORDERED. > > > > Looking at the packet trace, the seqid for the slot has advanced by > > > > 2 instead of 1. An RPC request for old-seqid + 1 never seems to get > > > > sent. > > > > --> Since sending an RPC with "seqid + 2" but never sending one > > > > that is "seqid + 1" for a slot seems harmless, I have added an optional > > > > hack (can be turned off), to allow this case instead of replying > > > > NFS4ERR_SEQ_MISORDERED for it. The code will still reply > > > > NFS4ERR_SEQ_MISORDERED if an RPC for the slot with > > > > "old seqid + 1" in it. > > > > --> Yes, doing this hack is a violation of RFC5661, but I've > > > > done it anyhow. > > > > > > > > If you are interested in a packet capture with this in it: > > > > fetch https://people.freebsd.org/~rmacklem/linuxtofreenfs.pcap > > > > - then look at packet #1945 and #2072 > > > > --> You'll see that slot #1 seqid goes from 4 to 6. There is no > > > > slot#1 seqid 5 RPC sent on the wire. > > > > (This packet capture was taken on the Linux client using > > > > tcpdump.) > > > > --> Btw, the "RST battle" you'll see in the above trace between > > > > #2005 and #2068 that goes on until the FreeBSD > > > > krpc/NFS times out the connection after 6min. seems to be a recent > > > > FreeBSD TCP bug. > > > > I have reproduced this seqid advances by 2 on an older system > > > > that does not "RST battle" and allows the reconnect right away, > > > > once the network partition is healed, so it does seem to be > > > > relevant to this bug. > > > > > > > > Someday, I will get around to upgrading to a more recent Linux > > > > kernel and will test to see if I can still reproduce this bug. > > > > On 5.2.10, it is intermittent and does not occur every time I > > > > do the network partitioning test. > > > > > > > > Mostly just fyi, rick > > > > > > Hi Rick, > > > > > > I think this is happening because slotid=1 had something queued up > > > using seqid=5 and that was interrupted because the connection was > > > RSTed. For the interrupted slot, the client would send solo SEQUENCE > > > with +1 seqid. > > > > Doesn't the client send the solo SEQUENCE with seqid 5 in that case? > > No it sends with seq+1 because NFS layer client doesn't know if seqid > actually was actually transmitted before the connection got caught > (and/or received by the server). But then the MISORDERED tells the client it wasn't received, so the client follows up with a call with seqid 5--is that what happens? Sorry, I seem to recall we went through this all a couple years ago, but now I've forgotten how it works. --b.
