On 17/06/07 17:30 +0100, Thiemo Seufer wrote: > Lauri Leukkunen wrote: > > Attached patch fixes the linux-user path mangling code for use with > > real target root filesystems that have nasty symlinks and lots of > > files. The old code is terribly slow and can easily end up going > > through the entire host system /usr hierarchy in a recursive loop. > > > > Compared to the previous version of this patch, fixes an issue with > > attempting to free() a pointer returned by GNU basename(). > > I gave it a try, and found: > - It loses the path caching (which should be done incrementally to > avoid the slowness you observe)
Current path mangling is not just slow, it gets stuck in a loop for 15 minutes on my box and eats all my RAM until it exhausts all virtual memory: 3976 lleukkun 18 0 2641m 1.5g 188 R 68 77.4 1:26.36 qemu-arm So I don't know if losing a bit of caching is that bad in comparison. I've implemented a kinda similar path mangling system for scratchbox2, there I'm caching the translations in a cache dir as symlinks, cuts down the mangling overhead quite a bit, but on the other hand sb2 needs to do a lot more work to determine how to mangle the path. Not sure it's worth the trouble for qemu. It seems on a current debian/testing it no longer works at all, it used to hog all memory and then (maybe after getting a NULL from malloc()?) continued, now it just eats all swap and keeps going as long as the kernel permits. If you want to see the effect get a maemo rootfs from here: http://repository.maemo.org/stable/3.1/armel/maemo-sdk-rootstrap_3.1_armel.tgz and try giving the extracted dir to qemu-arm -L /path/to/the/rootfs It gets tripped up by the /usr/X11R6/bin/X11 -> . symlink, don't ask me why such a link is needed ;) > - It misses to do error handling on malloc() returns > - It is broken, an emulated "/bin/ls /.." shows it ascends above > the root. I'll try to provide an updated patch with these addressed tomorrow. /lauri
