Date: Wed, 18 Jun 2003 17:35:19 +0200 From: Poul-Henning Kamp <phk@phk.freebsd.dk> To: arch@freebsd.org Subject: userland access to devices is moving! Message-ID: <41012.1055950519@critter.freebsd.dk>
next in thread | raw e-mail | index | archive | help
I sat down and hacked up a simple prototype to test the concept I have been rambling about for some years: Going directly from filedescriptor to device driver thus bypassing the vnode, devfs and specfs layer. I implemented this for /dev/null and /dev/zero, and ran the simple benchmark "dd if=/dev/zero of=/dev/null count=1000000" Before: N 3 Average: 44.900752667 Stddev: 0.049906338 After: N 3 Average: 18.460190333 Stddev: 0.074019507 That is 26.4 microseconds saved for each read(2)+write(2) operation, or 41% improvement on my Athlon 700MHz machine. A bit more locking will probably be needed, so this will erode some of this number, but there will be something left I'm sure :-) The largest impact of this is that VOP_OPEN(), vn_open() and vn_open_cred() grows an argument (the fdesc index) which existing callers need to pass a -1, the rest is relatively local hacking in devfs and some adjustments in the descriptor code. I have overall found that the implementation of this is not as hard as I imagined, and if I doubted it before, I am now certain that this is the right way to go. I should have tried this long time ago... patch at: http://phk.freebsd.dk/patch/fdesc.patch -- Poul-Henning Kamp | UNIX since Zilog Zeus 3.20 phk@FreeBSD.ORG | TCP/IP since RFC 956 FreeBSD committer | BSD since 4.3-tahoe Never attribute to malice what can adequately be explained by incompetence.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?41012.1055950519>