Re: git: 6468cd8e0ef9 - main - mount: add vnode usage per file system with mount -v
Date: Tue, 14 Jun 2022 16:50:09 UTC
On Mon, Jun 13, 2022 at 09:56:38PM +0200, Mateusz Guzik wrote:
| On 6/13/22, Doug Ambrisko <ambrisko@ambrisko.com> wrote:
| > On Mon, Jun 13, 2022 at 06:43:31PM +0200, Mateusz Guzik wrote:
| > | On 6/13/22, Doug Ambrisko <ambrisko@freebsd.org> wrote:
| > | > The branch main has been updated by ambrisko:
| > | >
| > | > URL:
| > | >
| > https://cgit.FreeBSD.org/src/commit/?id=6468cd8e0ef9d1d3331e9de26cd2be59bc778494
| > | >
| > | > commit 6468cd8e0ef9d1d3331e9de26cd2be59bc778494
| > | > Author: Doug Ambrisko <ambrisko@FreeBSD.org>
| > | > AuthorDate: 2022-06-13 14:56:38 +0000
| > | > Commit: Doug Ambrisko <ambrisko@FreeBSD.org>
| > | > CommitDate: 2022-06-13 14:56:38 +0000
| > | >
| > | > mount: add vnode usage per file system with mount -v
| > | >
| > | > This avoids the need to drop into the ddb to figure out vnode
| > | > usage per file system. It helps to see if they are or are not
| > | > being freed. Suggestion to report active vnode count was from
| > | > kib@
| > | >
| > | > Reviewed by: kib
| > | > Differential Revision: https://reviews.freebsd.org/D35436
| > | > ---
| > | > sbin/mount/mount.c | 7 +++++++
| > | > sys/kern/vfs_mount.c | 12 ++++++++++++
| > | > sys/sys/mount.h | 4 +++-
| > | > 3 files changed, 22 insertions(+), 1 deletion(-)
| > | >
| > | > diff --git a/sbin/mount/mount.c b/sbin/mount/mount.c
| > | > index 79d9d6cb0caf..bd3d0073c474 100644
| > | > --- a/sbin/mount/mount.c
| > | > +++ b/sbin/mount/mount.c
| > | > @@ -692,6 +692,13 @@ prmount(struct statfs *sfp)
| > | > xo_emit("{D:, }{Lw:fsid}{:fsid}", fsidbuf);
| > | > free(fsidbuf);
| > | > }
| > | > + if (sfp->f_nvnodelistsize != 0 || sfp->f_avnodecount != 0) {
| > | > + xo_open_container("vnodes");
| > | > + xo_emit("{D:,
| > | > }{Lwc:vnodes}{Lw:count}{w:count/%ju}{Lw:active}{:active/%ju}",
| > | > + (uintmax_t)sfp->f_nvnodelistsize,
| > | > + (uintmax_t)sfp->f_avnodecount);
| > | > + xo_close_container("vnodes");
| > | > + }
| > | > }
| > | > xo_emit("{D:)}\n");
| > | > }
| > | > diff --git a/sys/kern/vfs_mount.c b/sys/kern/vfs_mount.c
| > | > index 71a40fd97a9c..e3818b67e841 100644
| > | > --- a/sys/kern/vfs_mount.c
| > | > +++ b/sys/kern/vfs_mount.c
| > | > @@ -2610,6 +2610,8 @@ vfs_copyopt(struct vfsoptlist *opts, const char
| > *name,
| > | > void *dest, int len)
| > | > int
| > | > __vfs_statfs(struct mount *mp, struct statfs *sbp)
| > | > {
| > | > + struct vnode *vp;
| > | > + uint32_t count;
| > | >
| > | > /*
| > | > * Filesystems only fill in part of the structure for updates, we
| > | > @@ -2624,6 +2626,16 @@ __vfs_statfs(struct mount *mp, struct statfs
| > *sbp)
| > | > sbp->f_version = STATFS_VERSION;
| > | > sbp->f_namemax = NAME_MAX;
| > | > sbp->f_flags = mp->mnt_flag & MNT_VISFLAGMASK;
| > | > + sbp->f_nvnodelistsize = mp->mnt_nvnodelistsize;
| > | > +
| > | > + count = 0;
| > | > + MNT_ILOCK(mp);
| > | > + TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
| > | > + if (vrefcnt(vp) > 0) /* racy but does not matter */
| > | > + count++;
| > | > + }
| > | > + MNT_IUNLOCK(mp);
| > | > + sbp->f_avnodecount = count;
| > | >
| > |
| > | libc uses statfs for dir walk (see gen/fts.c), most notably find
| > | immediately runs into it. As such the linear scan by default is a
| > | non-starter.
| > |
| > | I don't know if mount is the right place to dump this kind of info to
| > | begin with, but even so, it should only happen with a dedicated flag.
| > |
| > | As statfs does not take any flags on its own, there is no way to
| > | prevent it from doing the above walk. Perhaps a dedicated sysctl which
| > | takes mount point id could do the walk instead, when asked.
| > |
| > | Short of making the walk optional I'm afraid this will have to be
| > reverted.
| >
| > Just to be clear, this isn't breaking things but is not optimal for
| > things that don't need this extra info.
| >
|
| It's not "not optimal", it's a significant overhead which taxes
| frequent users which don't benefit from it.
|
| For more data I plugged dtrace -n 'fbt::__vfs_statfs:entry {
| @[execname] = count(); }' while package building, then i got tons of
| hits:
| [snip]
| expr 13992
| install 14090
| dirname 14921
| mv 17404
| ghc-stage1 17577
| grep 18998
| xgcc 23832
| cpp 29282
| cc1 36961
| sh 70575
| rm 73904
| ld.lld 87784
| sed 88803
| c++ 98175
| cat 115811
| cc 449725
|
I rather not revert it all but revert the active part that kib@ thought
would be useful versus lazy vnodes. This is what I propose to commit:
diff --git a/sbin/mount/mount.c b/sbin/mount/mount.c
index bd3d0073c47..6c986907bcd 100644
--- a/sbin/mount/mount.c
+++ b/sbin/mount/mount.c
@@ -692,11 +692,10 @@ prmount(struct statfs *sfp)
xo_emit("{D:, }{Lw:fsid}{:fsid}", fsidbuf);
free(fsidbuf);
}
- if (sfp->f_nvnodelistsize != 0 || sfp->f_avnodecount != 0) {
+ if (sfp->f_nvnodelistsize != 0) {
xo_open_container("vnodes");
- xo_emit("{D:, }{Lwc:vnodes}{Lw:count}{w:count/%ju}{Lw:active}{:active/%ju}",
- (uintmax_t)sfp->f_nvnodelistsize,
- (uintmax_t)sfp->f_avnodecount);
+ xo_emit("{D:, }{Lwc:vnodes}{Lw:count}{w:count/%ju}",
+ (uintmax_t)sfp->f_nvnodelistsize);
xo_close_container("vnodes");
}
}
diff --git a/sys/kern/vfs_mount.c b/sys/kern/vfs_mount.c
index e3818b67e84..3c55f83633a 100644
--- a/sys/kern/vfs_mount.c
+++ b/sys/kern/vfs_mount.c
@@ -2610,9 +2610,6 @@ vfs_copyopt(struct vfsoptlist *opts, const char *name, void *dest, int len)
int
__vfs_statfs(struct mount *mp, struct statfs *sbp)
{
- struct vnode *vp;
- uint32_t count;
-
/*
* Filesystems only fill in part of the structure for updates, we
* have to read the entirety first to get all content.
@@ -2628,15 +2625,6 @@ __vfs_statfs(struct mount *mp, struct statfs *sbp)
sbp->f_flags = mp->mnt_flag & MNT_VISFLAGMASK;
sbp->f_nvnodelistsize = mp->mnt_nvnodelistsize;
- count = 0;
- MNT_ILOCK(mp);
- TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
- if (vrefcnt(vp) > 0) /* racy but does not matter */
- count++;
- }
- MNT_IUNLOCK(mp);
- sbp->f_avnodecount = count;
-
return (mp->mnt_op->vfs_statfs(mp, sbp));
}
diff --git a/sys/sys/mount.h b/sys/sys/mount.h
index edac64171f9..ffb2676258f 100644
--- a/sys/sys/mount.h
+++ b/sys/sys/mount.h
@@ -92,7 +92,7 @@ struct statfs {
uint64_t f_syncreads; /* count of sync reads since mount */
uint64_t f_asyncreads; /* count of async reads since mount */
uint32_t f_nvnodelistsize; /* # of vnodes */
- uint32_t f_avnodecount; /* # of active vnodes */
+ uint32_t f_spare0; /* unused spare */
uint64_t f_spare[9]; /* unused spare */
uint32_t f_namemax; /* maximum filename length */
uid_t f_owner; /* user that mounted the filesystem */
the single assignment of:
sbp->f_nvnodelistsize = mp->mnt_nvnodelistsize;
shouldn't cause much of a performance issue. Output is now:
root@client:~ # mount -v
192.168.35.1:/data/home/ambrisko/netboot on / (nfs, fsid 01ff003a3a000000, vnodes: count 567 )
devfs on /dev (devfs, fsid 00ff007171000000, vnodes: count 27 )
procfs on /proc (procfs, local, fsid 02ff000202000000, vnodes: count 2 )
linprocfs on /compat/linux/proc (linprocfs, local, fsid 03ff00b5b5000000, vnodes: count 2 )
linsysfs on /compat/linux/sys (linsysfs, local, fsid 04ff008a8a000000, vnodes: count 2 )
fdescfs on /dev/fd (fdescfs, fsid 05ff005959000000, vnodes: count 3 )
tmpfs on /tmp (tmpfs, local, fsid 06ff008787000000, vnodes: count 6 )
tmpfs on /var/tmp (tmpfs, local, fsid 07ff008787000000, vnodes: count 2 )
tmpfs on /var/run (tmpfs, local, fsid 08ff008787000000, vnodes: count 8 )
devfs on /compat/linux/dev (devfs, fsid 09ff007171000000, vnodes: count 4 )
fdescfs on /compat/linux/dev/fd (fdescfs, fsid 0aff005959000000, vnodes: count 2 )
tmpfs on /compat/linux/dev/shm (tmpfs, local, fsid 0bff008787000000, vnodes: count 2 )
192.168.31.1:/data on /data (nfs, fsid 0cff003a3a000000, vnodes: count 2 )
root@client:~ #
Thanks,
Doug A.