On Sun, Feb 21, 2010 at 05:17:37PM -0800, Tim Kientzle wrote:
> Juergen,
>
> Could you try the current version of read_open_filename from:
>
> http://libarchive.googlecode.com/svn/trunk/libarchive/archive_read_open_filename.c
>
> You should be able to just copy it into your FreeBSD source
> tree and recompile.
>
Small but important bug:
Index: archive_read_open_filename.c
@@ -94,7 +94,7 @@ archive_read_open_filename(struct archiv
struct read_file_data *mine;
void *buffer;
int fd;
- int is_disk_like;
+ int is_disk_like = 0;
#if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
off_t mediasize = 0;
#elif defined(__NetBSD__) || defined(__OpenBSD__)
> Duane Hesser's comments clarified for me that disk and tape
> I/O do require fundamentally different strategies, so I
> refactored this code to clearly break out those
> strategies. This in turn allowed me to do some
> straightforward I/O optimization of the disk case,
> which sped up your example by about 100x. I have a 2G
> DVD here on which "tar tvf /dev/acd0" took 90s originally,
> takes 14s with your patch, and now takes only 0.6s.
> For comparison, after mounting the disk, "find /mnt"
> takes 1.6s (which drops to .1s after the cache is warmed).
> It would be interesting to also compare two ways of
> copying all of the files: "tar xvf /dev/acd0"
> on an unmounted disk and "cp -R" on a mounted disk.
>
> The non-disk cases are all still handled generically, but
> there are now clearly-labeled holes where someone could
> add optimized I/O strategies for tapes, sockets, etc.
> I've even included a number of TODO comments about
> what strategies I think are worth exploring, including
> tricks like using MTIO ioctls on tape, mmap() for disk,
> and nonblocking I/O for sockets and pipes. I expect
> any one of these is around a dozen lines of code, so
> should be within the reach of even fairly junior
> developers.
Btw... the lseek + read combinations could also be replaced with
pread(2), for the least amount of syscalls. :) (Especially since I
noticed this version at least when ran on an optical disc with iso9660
ends up doing lots of lseek()s towards the end without ever reading
anything from the disc anymore...)
And to test the above fix, I ran the Linux-built bsdtar in the
linuxolator, and after confirming it now does read all of the disc
again I went and patched the remaining compatibility problems, i.e.
I now have disks appear as block devices for Linux processes (there
already was commented out code for that in linux_stats.c, I hope my
version is now `correct enough' to be usable [1]), and I made a simple
patch to make lseek SEEK_END (L_XTND in the source) dtrt on disk
devices too by simply invoking the DIOCGMEDIASIZE ioctl there. [2]
Patches are here: (made on stable/8, if they don't apply on head
I'll have to make extra versions for that...)
http://people.freebsd.org/~nox/linuxdisk-blk.patch [1]
http://people.freebsd.org/~nox/lseek-seek_end.patch [2]
And yes, with these patches the Linux bsdtar now also runs fast
on FreeBSD. :)
Cheers,
Juergen
_______________________________________________
freebsd-hackers@freebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-hackers
To unsubscribe, send any mail to "freebsd-hackers-unsubscr...@freebsd.org"
