Date: Tue, 17 May 2016 07:13:55 +1000 (EST) From: Bruce Evans <brde@optusnet.com.au> To: fs@freebsd.org Cc: rmacklem@freebsd.org Subject: fixes for i/o counting in nfs Message-ID: <20160517063058.E2021@besplex.bde.org>
next in thread | raw e-mail | index | archive | help
nfs doesn't count block inputs in resource usage. It seems to count
block outputs well enough (not very well, since buffering and threading
causes some i/o's to be done by other threads where the counts are hard
to see and harder to associate with the actual user).
nfs doesn't support per-mount i/o counts for either input and output.
These patches are for an old version of oldnfs. They apply cleanly
to oldnfs in FreeBSD-10.
I'm not sure if I found all the i/o's and don't trust my thread and mount
pointer handling, but they seem to work reasonably and never find a null
pointer. The per-mount i/o counts were easier to do than for most
file systems since nfs isn't handicapped by using geom. The corresponding
code in g_vfs_strategy() has a harder time finding the mount point and
often fails, so must check for null pointers and not work when it can't
find the mount point. This code doesn't even exist in the version that
this patch is for (except I patch it in).
X Index: nfs_bio.c
X ===================================================================
X --- nfs_bio.c (revision 181737)
X +++ nfs_bio.c (working copy)
X @@ -1568,6 +1581,14 @@
X case VREG:
X uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE;
X nfsstats.read_bios++;
X + if (td == NULL)
X + curthread->td_ru.ru_inblock++; /* XXX */
X + else
X + td->td_ru.ru_inblock++; /* XXX? */
These are XXX'ed since I don't know if td is ever null or always right
when it is non-null. But this seems to work right -- some counts go
to normal threads and some to nfsiod's.
X + if (LK_HOLDER(bp->b_lock.lk_lock) == LK_KERNPROC)
X + vp->v_mount->mnt_stat.f_asyncreads++; /* XXX */
X + else
X + vp->v_mount->mnt_stat.f_syncreads++;
X error = (nmp->nm_rpcops->nr_readrpc)(vp, uiop, cr);
This is XXX'ed since I don't trust the LK_KERNPROC check at all. This
was blindly copied from g_vfs_strategy(). A separate count for async
_reads_ is not very useful anyway. It is mostly for read-ahead. Most
reads should be ahead, but complicated buffering in hardware and software
makes them hard to count and the counts not very useful.
X
X if (!error) {
X @@ -1674,10 +1695,16 @@
X io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
X uiop->uio_rw = UIO_WRITE;
X nfsstats.write_bios++;
X + if (td == NULL)
X + curthread->td_ru.ru_oublock++; /* XXX */
X + else
X + td->td_ru.ru_oublock++; /* XXX? */
As above.
X
X if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC)
X + vp->v_mount->mnt_stat.f_asyncwrites++,
X iomode = NFSV3WRITE_UNSTABLE;
X else
X + vp->v_mount->mnt_stat.f_syncwrites++,
X iomode = NFSV3WRITE_FILESYNC;
Here the sync/async decision is easy to make correctly.
The patch uses a comma splice hack to keep the patch small.
X
X error = (nmp->nm_rpcops->nr_writerpc)(vp, uiop, cr, &iomode, &must_commit);
X Index: nfs_vnops.c
X ===================================================================
X --- nfs_vnops.c (revision 181737)
X +++ nfs_vnops.c (working copy)
X @@ -3138,7 +3290,6 @@
X bp->b_iocmd = BIO_WRITE;
X
X bufobj_wref(bp->b_bufobj);
X - curthread->td_ru.ru_oublock++;
X splx(s);
X
X /*
This is now counted in nfs_bio.c, and the results are much the same.
Apparently it makes little difference to always use curthread. In file
systems generally, we pass around td's and they are usually useless, but
if they are good for anything at all then it is to record the (first)
originator of the i/o so as to charge the originator and not a daemon.
I don't know if they are used for that.
Bruce
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20160517063058.E2021>