From: Hanna Linder <hannal@us...> - 2002-02-08 20:03:54
LSE Con Call Minutes Feb 8, 2002
I. Hanna apologizes profusely. She will never cut and paste
the con call number again. The web page always has the right
I. William Lee Irwin III and Rik Van Riel - rmap stuff
The most recent and stable version is rmap 12d on 2.4.17
which Rik just sent out last night. Rik wants future work to look
very carefully at the locking in the vm to see what could be
simplified. He said two important things are 1) split the page
lock and 2) make a rule that if you are going to modify two
objects you need to lock them both.
When you add a page on an inode you take a lock on the
lru, inode, pagecache lists but no lock on the page. So other
people can still do stuff to the page. It would simplify the
vm if simple rule is followed that if you modify two objects
need to lock them both. That would get rid of a whole lot of
complexities. I would like to split the page lock to make it
easier to maintain and understand (ie diff between io and locking).
Bill said oh that might explain some of the problems
ive been seeing with truncate.
Gerrit asked about testing and said The LTP and OSDL should
be sent the newest patches. Rik said many people have tested it
(approx 600 people downloaded it from his site) mainly on desktops.
Gerrit said the most testing help to be offered would be on large
Bill Hartners group has started on 2.4.16/17 but low freemem
falls very low and causes problems. (might be a filesystem problem?).
On that 8-way system it looks like hte number of buffer heads
allocated grows to a very large number. Rik said early himem code
may have had this problem. He may need to add code to rmap to
strip buffer heads (?).
Gerrit asked if the rmap code would help with greater than
4gb bounce buffers problem? Rik said the rmap code might help
a little. They might try on an 8gb system with jens axboe's
bounce buffer patch and the newest rmap patch. But there is still
a problem with not utilizeing greater than 4bg memory.
Bill has merged the radix tree pagecache and rmap and offered
to send it to Rik. But Rik wants to wait until the radix tree
pagecache stabilizes until merging it with rmap.
Gerrit asked Rik if he has thought about using bitkeeper. Rik
said he has been using bitkeeper for quite a while. But he still has to
merge to linus' tree.
Rik really wants to simplify the locking in the vm.
Gerrit mentioned read copy update mechanism which originally
was written to help simplify locking as well as increase
performance. Rik said that might be an option.
(for more info read the logs on irc for #kernelnewbies or
go to http://www.surriel.com)
-minutes compiled by hannal@...
Any mistakes are my own.