Date: Wed, 20 Apr 2022 09:11:03 -0700 From: Doug Ambrisko <ambrisko@ambrisko.com> To: Mateusz Guzik <mjguzik@gmail.com> Cc: freebsd-current@freebsd.org Subject: Re: nullfs and ZFS issues Message-ID: <YmAwl0FDI1/ugpTg@ambrisko.com> In-Reply-To: <CAGudoHGv9Es4Qov9YSHQ=8jAOdiAmyWA0fCE1bgAStv%2Bn=J2dw@mail.gmail.com> References: <Yl31Frx6HyLVl4tE@ambrisko.com> <CAGudoHEqjs4QoAqvkvW5JdSOMZ_QNjd3XU65kULxgabsOva5Xw@mail.gmail.com> <CAGudoHGBfVFcsCbhC=MCRFPzCtVRYCa1pCU7cGuuJq1fOv6ttg@mail.gmail.com> <CAGudoHGP5MTaF_LKanCh88ufHM6mBdzicQg-KLdfw0xGA-AxJQ@mail.gmail.com> <Yl7cP3RTWROiQLX5@ambrisko.com> <CAGudoHGv9Es4Qov9YSHQ=8jAOdiAmyWA0fCE1bgAStv%2Bn=J2dw@mail.gmail.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Wed, Apr 20, 2022 at 11:43:10AM +0200, Mateusz Guzik wrote: | On 4/19/22, Doug Ambrisko <ambrisko@ambrisko.com> wrote: | > On Tue, Apr 19, 2022 at 11:47:22AM +0200, Mateusz Guzik wrote: | > | Try this: https://people.freebsd.org/~mjg/vnlru_free_pick.diff | > | | > | this is not committable but should validate whether it works fine | > | > As a POC it's working. I see the vnode count for the nullfs and | > ZFS go up. The ARC cache also goes up until it exceeds the ARC max. | > size tten the vnodes for nullfs and ZFS goes down. The ARC cache goes | > down as well. This all repeats over and over. The systems seems | > healthy. No excessive running of arc_prune or arc_evict. | > | > My only comment is that the vnode freeing seems a bit agressive. | > Going from ~15,000 to ~200 vnode for nullfs and the same for ZFS. | > The ARC drops from 70M to 7M (max is set at 64M) for this unit | > test. | > | | Can you check what kind of shrinking is requested by arc to begin | with? I imagine encountering a nullfs vnode may end up recycling 2 | instead of 1, but even repeated a lot it does not explain the above. I dug it into a bit more and think there could be a bug in: module/zfs/arc.c arc_evict_meta_balanced(uint64_t meta_used) prune += zfs_arc_meta_prune; //arc_prune_async(prune); arc_prune_async(zfs_arc_meta_prune); Since arc_prune_async, is queuing up a run of arc_prune_task for each call it is actually already accumulating the zfs_arc_meta_prune amount. It makes the count to vnlru_free_impl get really big quickly since it is looping via restart. 1 HELLO arc_prune_task 164 ticks 2147465958 count 20480000 dmesg | grep arc_prune_task | uniq -c 14 HELLO arc_prune_task 164 ticks -2147343772 count 100 50 HELLO arc_prune_task 164 ticks -2147343771 count 100 46 HELLO arc_prune_task 164 ticks -2147343770 count 100 49 HELLO arc_prune_task 164 ticks -2147343769 count 100 44 HELLO arc_prune_task 164 ticks -2147343768 count 100 116 HELLO arc_prune_task 164 ticks -2147343767 count 100 1541 HELLO arc_prune_task 164 ticks -2147343766 count 100 53 HELLO arc_prune_task 164 ticks -2147343101 count 100 100 HELLO arc_prune_task 164 ticks -2147343100 count 100 75 HELLO arc_prune_task 164 ticks -2147343099 count 100 52 HELLO arc_prune_task 164 ticks -2147343098 count 100 50 HELLO arc_prune_task 164 ticks -2147343097 count 100 51 HELLO arc_prune_task 164 ticks -2147343096 count 100 783 HELLO arc_prune_task 164 ticks -2147343095 count 100 884 HELLO arc_prune_task 164 ticks -2147343094 count 100 Note I shrunk vfs.zfs.arc.meta_prune=100 to see how that might help. Changing it to 1, helps more! I see less agressive swings. I added printf("HELLO %s %d ticks %d count %ld\n",__FUNCTION__,__LINE__,ticks,nr_scan); to arc_prune_task. Adjusting both sysctl vfs.zfs.arc.meta_adjust_restarts=1 sysctl vfs.zfs.arc.meta_prune=100 without changing arc_prune_async(prune) helps avoid excessive swings. Thanks, Doug A. | > | On 4/19/22, Mateusz Guzik <mjguzik@gmail.com> wrote: | > | > On 4/19/22, Mateusz Guzik <mjguzik@gmail.com> wrote: | > | >> On 4/19/22, Doug Ambrisko <ambrisko@ambrisko.com> wrote: | > | >>> I've switched my laptop to use nullfs and ZFS. Previously, I used | > | >>> localhost NFS mounts instead of nullfs when nullfs would complain | > | >>> that it couldn't mount. Since that check has been removed, I've | > | >>> switched to nullfs only. However, every so often my laptop would | > | >>> get slow and the the ARC evict and prune thread would consume two | > | >>> cores 100% until I rebooted. I had a 1G max. ARC and have increased | > | >>> it to 2G now. Looking into this has uncovered some issues: | > | >>> - nullfs would prevent vnlru_free_vfsops from doing anything | > | >>> when called from ZFS arc_prune_task | > | >>> - nullfs would hang onto a bunch of vnodes unless mounted with | > | >>> nocache | > | >>> - nullfs and nocache would break untar. This has been fixed | > now. | > | >>> | > | >>> With nullfs, nocache and settings max vnodes to a low number I can | > | >>> keep the ARC around the max. without evict and prune consuming | > | >>> 100% of 2 cores. This doesn't seem like the best solution but it | > | >>> better then when the ARC starts spinning. | > | >>> | > | >>> Looking into this issue with bhyve and a md drive for testing I | > create | > | >>> a brand new zpool mounted as /test and then nullfs mount /test to | > /mnt. | > | >>> I loop through untaring the Linux kernel into the nullfs mount, rm | > -rf | > | >>> it | > | >>> and repeat. I set the ARC to the smallest value I can. Untarring | > the | > | >>> Linux kernel was enough to get the ARC evict and prune to spin since | > | >>> they couldn't evict/prune anything. | > | >>> | > | >>> Looking at vnlru_free_vfsops called from ZFS arc_prune_task I see it | > | >>> static int | > | >>> vnlru_free_impl(int count, struct vfsops *mnt_op, struct vnode | > *mvp) | > | >>> { | > | >>> ... | > | >>> | > | >>> for (;;) { | > | >>> ... | > | >>> vp = TAILQ_NEXT(vp, v_vnodelist); | > | >>> ... | > | >>> | > | >>> /* | > | >>> * Don't recycle if our vnode is from different type | > | >>> * of mount point. Note that mp is type-safe, the | > | >>> * check does not reach unmapped address even if | > | >>> * vnode is reclaimed. | > | >>> */ | > | >>> if (mnt_op != NULL && (mp = vp->v_mount) != NULL && | > | >>> mp->mnt_op != mnt_op) { | > | >>> continue; | > | >>> } | > | >>> ... | > | >>> | > | >>> The vp ends up being the nulfs mount and then hits the continue | > | >>> even though the passed in mvp is on ZFS. If I do a hack to | > | >>> comment out the continue then I see the ARC, nullfs vnodes and | > | >>> ZFS vnodes grow. When the ARC calls arc_prune_task that calls | > | >>> vnlru_free_vfsops and now the vnodes go down for nullfs and ZFS. | > | >>> The ARC cache usage also goes down. Then they increase again until | > | >>> the ARC gets full and then they go down again. So with this hack | > | >>> I don't need nocache passed to nullfs and I don't need to limit | > | >>> the max vnodes. Doing multiple untars in parallel over and over | > | >>> doesn't seem to cause any issues for this test. I'm not saying | > | >>> commenting out continue is the fix but a simple POC test. | > | >>> | > | >> | > | >> I don't see an easy way to say "this is a nullfs vnode holding onto a | > | >> zfs vnode". Perhaps the routine can be extrended with issuing a nullfs | > | >> callback, if the module is loaded. | > | >> | > | >> In the meantime I think a good enough(tm) fix would be to check that | > | >> nothing was freed and fallback to good old regular clean up without | > | >> filtering by vfsops. This would be very similar to what you are doing | > | >> with your hack. | > | >> | > | > | > | > Now that I wrote this perhaps an acceptable hack would be to extend | > | > struct mount with a pointer to "lower layer" mount (if any) and patch | > | > the vfsops check to also look there. | > | > | > | >> | > | >>> It appears that when ZFS is asking for cached vnodes to be | > | >>> free'd nullfs also needs to free some up as well so that | > | >>> they are free'd on the VFS level. It seems that vnlru_free_impl | > | >>> should allow some of the related nullfs vnodes to be free'd so | > | >>> the ZFS ones can be free'd and reduce the size of the ARC. | > | >>> | > | >>> BTW, I also hacked the kernel and mount to show the vnodes used | > | >>> per mount ie. mount -v: | > | >>> test on /test (zfs, NFS exported, local, nfsv4acls, fsid | > | >>> 2b23b2a1de21ed66, | > | >>> vnodes: count 13846 lazy 0) | > | >>> /test on /mnt (nullfs, NFS exported, local, nfsv4acls, fsid | > | >>> 11ff002929000000, vnodes: count 13846 lazy 0) | > | >>> | > | >>> Now I can easily see how the vnodes are used without going into ddb. | > | >>> On my laptop I have various vnet jails and nullfs mount my homedir | > into | > | >>> them so pretty much everything goes through nullfs to ZFS. I'm | > limping | > | >>> along with the nullfs nocache and small number of vnodes but it would | > be | > | >>> nice to not need that. | > | >>> | > | >>> Thanks, | > | >>> | > | >>> Doug A. | > | >>> | > | >>> | > | >> | > | >> | > | >> -- | > | >> Mateusz Guzik <mjguzik gmail.com> | > | >> | > | > | > | > | > | > -- | > | > Mateusz Guzik <mjguzik gmail.com> | > | > | > | | > | | > | -- | > | Mateusz Guzik <mjguzik gmail.com> | > | | | -- | Mateusz Guzik <mjguzik gmail.com>
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?YmAwl0FDI1/ugpTg>