Date: Thu, 16 Apr 1998 21:00:29 +0800 From: Peter Wemm <peter@netplex.com.au> To: Bruce Evans <bde@zeta.org.au> Cc: cvs-all@FreeBSD.ORG, cvs-committers@FreeBSD.ORG, cvs-sys@FreeBSD.ORG, dyson@FreeBSD.ORG Subject: Re: cvs commit: src/sys/kern vfs_subr.c Message-ID: <199804161300.VAA16055@spinner.netplex.com.au> In-Reply-To: Your message of "Thu, 16 Apr 1998 18:51:05 %2B1000." <199804160851.SAA04053@godzilla.zeta.org.au>
next in thread | previous in thread | raw e-mail | index | archive | help
Bruce Evans wrote:
> > Modified files:
> > sys/kern vfs_subr.c
> > Log:
> > When the softdep conversion took place, the periodic vfs_msync() from
> > update got lost. This is responsible for ensuring that dirty mmap() pages
> > get periodically written to disk. Without it, long time mmap's might not
> > have their dirty pages written out at all of the system crashes or isn't
> > cleanly shut down. This could be nasty if you've got a long-running
> > writing via mmap(), dirty pages used to get written to disk within 30
> > seconds or so.
>
> sync_fsync() seems to be called too often for this (approx. every 3
> seconds on an idle system).
Well, from what I could see, it's called however often it needs in order
to sync each filesystem one-at-a-time over a 30 second period. The old
update process called vfs_msync() n times (n = number of mountpoints)
every 30 seconds. The way I understand things, it's now being called
at exactly the same rate overall but at a more even spread.
> vfs_msync() seems to do a lot every time it is called.
Yes.. It walks through the vnodes attached to a given filesystem. From
the look of it, there could be some optimizations there.. We could
eliminate uninteresting vnodes before wasting time doing VOP's, getting
simple_locks, etc.
Index: vfs_subr.c
===================================================================
RCS file: /home/ncvs/src/sys/kern/vfs_subr.c,v
retrieving revision 1.150
diff -u -2 -r1.150 vfs_subr.c
--- vfs_subr.c 1998/04/16 03:31:26 1.150
+++ vfs_subr.c 1998/04/16 12:57:05
@@ -2405,4 +2405,5 @@
vfs_msync(struct mount *mp, int flags) {
struct vnode *vp, *nvp;
+ struct vm_object *obj;
int anyio, tries;
@@ -2417,4 +2418,9 @@
goto loop;
}
+
+ /* skip vnodes that are uninteresting before locking */
+ obj = vp->v_object;
+ if (obj == NULL || (obj->flags & OBJ_MIGHTBEDIRTY) == 0)
+ continue;
if ((vp->v_flag & VXLOCK) ||
With a few thousand vnodes, all those VOP_ISLOCKED() and
simple_lock()/unlock calls have got to add up. It should be harmless to
test these variables outside of any critical section protection.
> We also lost control of the update interval.
Yes, but it is compile-time tweakable, so those with a desperate need to
tweak it can still do so, although on a power-of-two basis.
> We also lost waking the sync daemon in vm_pageout_scan(). There is
> still a wakeup on vfs_update_wakeup, but nothing sleeps on it any
> more.
>
> A better quick fix for most of this is probably to resurrect the
> update daemon and only call vfs_msync() from it.
I don't think there's much gain to be had overall.. The old update called
sync() every 30 seconds. sync does:
simple_lock(&mountlist_slock);
for (mp = mountlist.cqh_first; mp != (void *)&mountlist; mp = nmp) {
[..]
vfs_msync(mp, MNT_NOWAIT);
[..]
}
simple_unlock(&mountlist_slock);
So, if you have 10 filesystems, you get 20 vfs_msync's per minute under
update, but you get a whole glut of disk writes in one major hit.
Under syncer, you get 1 vfs_msync() every 3 seconds, stepping through each
filesystem one at a time. You still have 20 vfs_msync's per minute.
> Bruce
>
Cheers,
-Peter
--
Peter Wemm <peter@netplex.com.au> Netplex Consulting
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe cvs-all" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199804161300.VAA16055>