On Wed, Apr 04, 2012 at 03:51:38PM +0200, Mike Belopuhov wrote:
> On Wed, Apr 04, 2012 at 14:42 +0400, Alexander Polakov wrote:
> > This is a diff from NetBSD pr.34583:
> > http://gnats.netbsd.org/cgi-bin/query-pr-single.pl?number=34583
> >
> > Quoting the author:
> >
> > "I noticed that when writing large file (hundreds of megabytes)
> > to an msdos disk, the writing speed to a file decreases with the
> > file length.
> > Since I have some experience with messydos filesystems (I wrote
> > MSH: for the Amiga) I took a look.
> > The obvious suspicion with operations that slow down with the
> > length of a file is an excessive traversal of the FAT cluster
> > chain. However, there is a cache that caches 2 positions: the
> > last cluster in the file, and the "most recently looked up" one.
> > Debugging info showed however that frequent full traversals were
> > still made. So, apparently when extending a file and after
> > updating the end cluster, the previous end is again needed.
> > Adding a 3rd entry in the cache, which keeps the end position
> > from just before extending a file.
> > This has the desired effect of keeping the write speed constant.
> > (What it is that needs that position I have not been able to
> > ascertain from the filesystem code; it doesn't seem to make
> > sense, actually, to read or write clusters before the original
> > EOF. I was hoping to find the place where the cache is trashed
> > and rewrite it to get the desired info from it beforehand, so
> > that the extra cache entry is again unneeded, but alas.)"
> >
> > While there, I changed 0 to NULL for two pointer arguments of
> > extendfile().
> >
>
> i agree that this is a great find. i don't really like the diff though.
> i see no point in introducing this macro. what do others think?
Agreed. I like this version better.
.... Ken
>
> Index: msdosfs/denode.h
> ===================================================================
> RCS file: /cvs/src/sys/msdosfs/denode.h,v
> retrieving revision 1.23
> diff -u -p -r1.23 denode.h
> --- msdosfs/denode.h 17 Jul 2010 19:27:07 -0000 1.23
> +++ msdosfs/denode.h 4 Apr 2012 12:20:23 -0000
> @@ -116,10 +116,11 @@ struct fatcache {
> * cache is probably pretty worthless if a file is opened by multiple
> * processes.
> */
> -#define FC_SIZE 2 /* number of entries in the cache */
> +#define FC_SIZE 3 /* number of entries in the cache */
> #define FC_LASTMAP 0 /* entry the last call to pcbmap()
> resolved
> * to */
> #define FC_LASTFC 1 /* entry for the last cluster in the
> file */
> +#define FC_OLASTFC 2 /* entry for the previous last cluster
> */
>
> #define FCE_EMPTY 0xffffffff /* doesn't represent an actual
> cluster # */
>
> Index: msdosfs/msdosfs_fat.c
> ===================================================================
> RCS file: /cvs/src/sys/msdosfs/msdosfs_fat.c,v
> retrieving revision 1.22
> diff -u -p -r1.22 msdosfs_fat.c
> --- msdosfs/msdosfs_fat.c 4 Jul 2011 04:30:41 -0000 1.22
> +++ msdosfs/msdosfs_fat.c 4 Apr 2012 12:20:26 -0000
> @@ -952,6 +952,13 @@ extendfile(struct denode *dep, uint32_t
> return (error);
> }
>
> + /*
> + * Preserve value for the last cluster before extending the file
> + * to speed up further lookups.
> + */
> + fc_setcache(dep, FC_OLASTFC, dep->de_fc[FC_LASTFC].fc_frcn,
> + dep->de_fc[FC_LASTFC].fc_fsrcn);
> +
> while (count > 0) {
> /*
> * Allocate a new cluster chain and cat onto the end of the
