On Sun, Jun 28, 2020 at 12:18:06PM -0400, sven falempin wrote:
> On Sun, Jun 28, 2020 at 2:40 AM Bryan Linton <b...@shoshoni.info> wrote:
>
> > On 2020-06-27 19:29:31, Bob Beck <b...@obtuse.com> wrote:
> > >
> > > No.
> > >
> > > I know *exactly* what needbuf is but to attempt to diagnose what your
> > > problem is we need exact details. especially:
> > >
> > > 1) The configuration of your system including all the details of the
> > filesystems
> > > you have mounted, all options used, etc.
> > >
> > > 2) The script you are using to generate the problem (Not a paraphrasing
> > of what
> > > you think the script does) What filesystems it is using.
> > >
> >
> > Not the OP, but this problem sounds almost exactly like the bug I
> > reported last year.
> >
> > There is a detailed list of steps I used to reproduce the bug in
> > the following bug report.
> >
> > https://marc.info/?l=openbsd-bugs&m=156412299418191
> >
> > I was even able to bisect and identify the commit which first
> > caused the breakage for me.
> >
> >
> > ---8<---
> >
> > CVSROOT: /cvs
> > Module name: src
> > Changes by: b...@cvs.openbsd.org 2019/05/08 06:40:57
> >
> > Modified files:
> > sys/kern : vfs_bio.c vfs_biomem.c
> >
> > Log message:
> > Modify the buffer cache to always flip recovered DMA buffers high.
> >
> > This also modifies the backoff logic to only back off what is requested
> > and not a "mimimum" amount. Tested by me, benno@, tedu@ anda ports build
> > by naddy@.
> >
> > ok tedu@
> >
> > ---8<---
> >
> > However, I have since migrated away from using vnd(4)s since I was
> > able to find other solutions that worked for my use cases. So I
> > may not be able to provide much additional information other than
> > what is contained in the above bug report.
> >
> > --
> > Bryan
> >
> > >
> > >
> >
> >
> Reproduction of BUG.
>
>
> # optional
> mkdir /tmpfs
> mount_mfs -o rw -s 2500M swap /tmpfs # i mounted through fstab so this line
> is not tested
> #the bug
> /bin/dd if=/dev/zero of=/tmpfs/img.dd count=0 bs=1 seek=2500000000
> vnconfig vnd3 /tmpfs/img.dd
> printf "a a\n\n\n\nw\nq\n" | disklabel -E vnd3
> newfs vnd3a
> mount /dev/vnd3a /mnt
> cd /tmp && ftp https://cdn.openbsd.org/pub/OpenBSD/6.7/amd64/base67.tgz
> cd /mnt
> #will occur here (the mkdir was ub beedbuf state for a while ...
> for v in 1 2 3 4 5 6 7 8 9; do mkdir /tmp/$v; tar xzvf /tmp/base67.tgz -C
> /mnt/$v; done
>
> Ready to test patches.
>
>
So, your problem is that you have your vnd created in an mfs
filesystem, when I run your test with the vnd backed by a regular
filesystem (withe softdep even) it works fine.
The trouble happens when your VND has buffers cached in it's
"filesystem" but then is not flushing them out to the underlyin file
(vnode) that you have your vnd backed by. On normal filesystems this
works fine, since vnd tells the lower layer to not cache the writes
and to do them syncrhonously, to avoid an explosion of delayed writes
and dependencies of buffers.
The problem happens when we convert syncryonous bwrites to
asynchronous bdwrites if the fileystem is mounted ASYNC, which,
curiously, MFS always is (I don't know why, it doesn't really make any
sense, and I might even look at changing that) All the writes you do
end up being delayed anc chewing up more buffer space. And they are
all tied to one vnode (your image). once you exhaust the buffer
space, the cleaner runs, but as you have noticed can't clean out your
vnode until the syncer runs (every 60 seconds). This is why your
thing "takes a long time", and things stall in need buffer. softdep
has deep dark voodoo in it to avoid this problem and therefore when I
use a softdep filesystem instead of an ASYNC filesystem it works.
Anyway, what's below fixes your issue on my machine. I'm not sure I'm
happy that it's the final fix but it does fix it. there are many other
dragons lurking in here.
Index: sys/kern/vfs_bio.c
===================================================================
RCS file: /cvs/src/sys/kern/vfs_bio.c,v
retrieving revision 1.200
diff -u -p -u -p -r1.200 vfs_bio.c
--- sys/kern/vfs_bio.c 29 Apr 2020 02:25:48 -0000 1.200
+++ sys/kern/vfs_bio.c 29 Jun 2020 15:18:21 -0000
@@ -706,8 +706,14 @@ bwrite(struct buf *bp)
*/
async = ISSET(bp->b_flags, B_ASYNC);
if (!async && mp && ISSET(mp->mnt_flag, MNT_ASYNC)) {
- bdwrite(bp);
- return (0);
+ /*
+ * Don't convert writes from VND on async filesystems
+ * that already have delayed writes in the upper layer.
+ */
+ if (!ISSET(bp->b_flags, B_NOCACHE)) {
+ bdwrite(bp);
+ return (0);
+ }
}
/*
