Date: Thu, 20 May 2021 19:38:57 -0700 From: Mark Millard <marklmi@yahoo.com> To: Rick Macklem <rmacklem@uoguelph.ca> Cc: FreeBSD-STABLE Mailing List <freebsd-stable@freebsd.org> Subject: Re: releng/13 release/13.0.0 : odd/incorrect diff result over nfs (in a zfs file systems context) Message-ID: <04D7264A-206B-4281-B452-779B01EA3327@yahoo.com> In-Reply-To: <YTOPR0101MB0970A1257E4DD37335D5B52EDD299@YTOPR0101MB0970.CANPRD01.PROD.OUTLOOK.COM> References: <623369D9-5EE5-4FEF-B9AD-56499E8F1C09.ref@yahoo.com> <623369D9-5EE5-4FEF-B9AD-56499E8F1C09@yahoo.com> <YQXPR0101MB0968B29934D7BD73FCA73907DD299@YQXPR0101MB0968.CANPRD01.PROD.OUTLOOK.COM> <YTOPR0101MB0970A1257E4DD37335D5B52EDD299@YTOPR0101MB0970.CANPRD01.PROD.OUTLOOK.COM>
next in thread | previous in thread | raw e-mail | index | archive | help
> On 2021-May-20, at 18:09, Rick Macklem <rmacklem@uoguelph.ca> wrote: >=20 > Oh, one additional thing that I'll dare to top post... > r367492 broke the TCP upcalls that the NFS server uses, such > that intermittent hangs of NFS mounts to FreeBSD13 servers can occur. > This has not yet been resolved in "main" etc and could explain > why an RPC could time out for a soft mount. See later notes that I added: soft mount is not required to see the problem. > You can revert the patch in r367492 to avoid the problem. If I understand right, you are indicating that this would not apply to the non-soft mount case that I got. > Disabling TSO, LRO are also de-facto standard things to do when > you observe weird NFS behaviour, because they are often broken > in various network device drivers. I'll have to figure out how to experiment with such. Things are at defaults rather generally on the systems. I'm not literate in the subject areas. I'm the only user of the machines and network. It is not outward facing. It is a rather small EtherNet network. > rick >=20 > ________________________________________ > From: owner-freebsd-stable@freebsd.org = <owner-freebsd-stable@freebsd.org> on behalf of Rick Macklem = <rmacklem@uoguelph.ca> > Sent: Thursday, May 20, 2021 8:55 PM > To: FreeBSD-STABLE Mailing List; Mark Millard > Subject: Re: releng/13 release/13.0.0 : odd/incorrect diff result over = nfs (in a zfs file systems context) >=20 > Mark Millard wrote: >> [I warn that I'm a fairly minimal user of NFS >> mounts, not knowing all that much. I'm mostly >> reporting this in case it ends up as evidence >> via eventually matching up with others observing >> possibly related oddities.] >>=20 >> I got the following odd sequence (that I've >> mixed notes into). It involved a diff -r over NFS >> showing differences (files missing) and then a >> later diff finding matches for the same files, >> no file system changes made on either machine. >> I'm unable to reproduce the oddity on demand. >>=20 >> Note: A larger scope diff -r originally returned the >> below as well, but doing the narrower diff -r did >> repeat the result and that is what I show. (I >> make no use of devel/ice .) >>=20 >> # diff -r /usr/ports/devel/ice/files /mnt/devel/ice/files | more >> Only in /usr/ports/devel/ice/files: Make.rules.FreeBSD . . . >> Only in /usr/ports/devel/ice/files: patch-scripts-TestUtil.py >>=20 >> Note: The above was not expected. So I tried: >>=20 >> # ls -Tld /mnt/devel/ice/files/* >> -rw-r--r-- 1 root wheel 755 Apr 21 21:07:54 2021 = /mnt/devel/ice/files/Make.rules.FreeBSD . . . >> -rw-r--r-- 1 root wheel 2588 Apr 21 21:07:54 2021 = /mnt/devel/ice/files/patch-scripts-TestUtil.py >>=20 >> Note: So that indicated that the files were there on the >> machine that /mnt references. So attempting the original >> diff -r again: >>=20 >> # diff -r /usr/ports/devel/ice/files /mnt/devel/ice/files | more >> # >>=20 >> (Empty difference.) >>=20 >> Note: So after the explicit "ls -Tld /mnt/devel/ice/files/*" >> the odd result of the diff -r no longer happened: no >> differences reported. >>=20 >>=20 >>=20 >> For reference (both machines reported): >>=20 >> . . . >> The original mount command was on CA72_16Gp_ZFS: >>=20 >> # mount -onoatime,soft 192.168.1.170:/usr/ports/ /mnt/ > The likely explanation for this is your use of a "soft" mount. > - If the NFS server is slow to respond or there is a temporary network = issue, > the RPC request can time out and then the > syscall can fail with EINT/ETIMEDOUT. Since almost nothing, = including the > readdir(3) libc functions expect syscalls to fail this way... > Then the cached directory is messed up. > Doing the "ls" read the directory again and fixed the problem. >=20 > Try to reproduce it for a mount without the "soft" option. > (If a mount point is hung, due to an unresponsive server "umount -N = /mnt" > can usually get rid of it.) > Personally, I thought "soft" was a bad idea when Sun introduced it in = NFS in 1985 > and I still feel that way. > --> If you can reproduce it without "soft" then I can't explain it. > To be honest, the directory reading/caching code in the NFSv3 = client > hasn't changed significantly in literally decades, as far as I = can remember. Well . . . trying an even wider scope diff than the original . . . # umount /mnt/ # mount -onoatime 192.168.1.170:/usr/ports/ /mnt/ # diff -r /usr/ports/ /mnt/ | more Only in /mnt/databases/mongodb42/files/aarch64: = patch-src_third__party_mozjs-60_ Only in /usr/ports/databases/mongodb42/files/aarch64: = patch-src_third__party_mozjs-60_platform_aarch64_freebsd_build_Unified__cp= p__js__src25.cpp Only in /usr/ports/devel/ice/files: Make.rules.FreeBSD Only in /usr/ports/devel/ice/files: patch-config-Make.common.rules Only in /usr/ports/devel/ice/files: patch-cpp-Makefile . . . Only in /usr/ports/devel/ice/files: = patch-python-test-Slice-unicodePaths-run.py Only in /usr/ports/devel/ice/files: patch-scripts-Expect.py Only in /usr/ports/devel/ice/files: patch-scripts-IceGridAdmin.py Only in /usr/ports/devel/ice/files: patch-scripts-TestUtil.py So the devel/ice files showed up again. But 2 other lines show up, one finding a file supposedly only on /mnt/. . . QUOTE Only in /mnt/databases/mongodb42/files/aarch64: = patch-src_third__party_mozjs-60_ END QUOTE That seems to be a truncated file name. Looking directly on the machine = that /mnt/ references (hitting tab at the end of the partial name to show a list): # ls -Tld = /usr/ports/databases/mongodb42/files/aarch64/patch-src_third__party_mozjs-= 60_ = /usr/ports/databases/mongodb42/files/aarch64/patch-src_third__party_mozjs-= 60_gen-config.sh =20 = /usr/ports/databases/mongodb42/files/aarch64/patch-src_third__party_mozjs-= 60_platform_aarch64_freebsd_build_js-confdefs.h =20 = /usr/ports/databases/mongodb42/files/aarch64/patch-src_third__party_mozjs-= 60_platform_aarch64_freebsd_build_Unified__cpp__js__src0.cpp =20 = /usr/ports/databases/mongodb42/files/aarch64/patch-src_third__party_mozjs-= 60_platform_aarch64_freebsd_build_Unified__cpp__js__src1.cpp =20 . . . =20 = /usr/ports/databases/mongodb42/files/aarch64/patch-src_third__party_mozjs-= 60_platform_aarch64_freebsd_build_Unified__cpp__js__src9.cpp =20 = /usr/ports/databases/mongodb42/files/aarch64/patch-src_third__party_mozjs-= 60_platform_aarch64_freebsd_include_js-config.h =20 The other machine agrees (machine-local usage). The other of the 2 new names is one of the matches to the prefix: QUOTE Only in /usr/ports/databases/mongodb42/files/aarch64: = patch-src_third__party_mozjs-60_platform_aarch64_freebsd_build_Unified__cp= p__js__src25.cpp END QUOTE For reference: I've not gotten any console messages about anything during these. > One additional thing to note is that cached directory contents are = invalidated > when the directory's ctime changes. I'm not aware of anything that should have been touching the /usr/ports file systems on either machine any time near my diff activities. (I'm the only system user.) > I am not sure how/if/when ZFS changes a > directory's ctime. However, if it was badly broken, I'd hear about = this a lot. > (If the ZFS change to ZoL has changed its ctime handling, that might = also explain it > and I'll be hearing a lot more soon as FreeBSD13 becomes adopted. I = never use ZFS and, > as such, never test with it.) I recently decided to try using bectl, which lead to my recent ZFS-based system experiments. This means I can boot the stable/13 or main [so: 14] that I last built and try the same experiments with the same /usr/ports file sysystems. releng/13 's release/13.0.0 , stable/13 , and main are all non-debug builds as stands. I could add debug builds of any or all, but it would take a while. (aarch64 4-core Cortex-A72 contexts.) > --> For UFS, if you use mtime, directory caching does not work as = well, which is > why the client directory caching code uses ctime and not mtime = to detect that > a directory has changed and cached directory blocks need to be = invalidated. >=20 > Jason Bacon did report a directory reading issue some months ago that = never > quite got resolved, although I recall he said he couldn't reproduce it = after a > system update, so he thought it was related to some local change he = had made. > (I can't remember his email or I'd add him to the cc so he could = remind me what > his case was. I do recall it being somewhat reproducible and happened = for both > UFS and ZFS.) >> The network is just a local EtherNet. >=20 =3D=3D=3D Mark Millard marklmi at yahoo.com ( dsl-only.net went away in early 2018-Mar)
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?04D7264A-206B-4281-B452-779B01EA3327>