* Re: [PATCH] Re: 2.4 VM & refill_inactive_scan()
[not found] ` <Pine.LNX.4.10.10010271147160.1850-100000@penguin.transmeta.com>
@ 2000-11-02 12:29 ` Stephen C. Tweedie
0 siblings, 0 replies; only message in thread
From: Stephen C. Tweedie @ 2000-11-02 12:29 UTC (permalink / raw)
To: Linus Torvalds; +Cc: Rik van Riel, Ingo Molnar, linux-mm, Stephen Tweedie
Hi,
On Fri, Oct 27, 2000 at 11:52:35AM -0700, Linus Torvalds wrote:
>
> On Fri, 27 Oct 2000, Rik van Riel wrote:
> >
> > There is one big point left though...
> >
> > Raw IO into a page which just gets unmapped by a process.
> > For NFS we'll somehow need to get the credentials for that
> > page so we can write back the data after it is detached from
> > the process...
>
> I really think that you should just add "struct file" to the kiobuf array,
> and expand "maplist".
OK, done. I've broken the diff into chunks, because there are a
couple of other kiobuf fixes to go in at the same time. The kiobuf
user-land mapping code can now map the struct file and propagate
dirty flags into it once a read operation into the memory completes.
> In fact, you should expand "maplist" anyway, because right now kiobuf's
> cannot handle the case of multiple partial pages: you can be partial only
> at the beginning or the end, which means that kiobuf's are worthless for
> stuff like "sendmsg()" that can do scatter-gather.
That's why the IO functions themselves take a kiovec, not a kiobuf, as
input, and why the iobuf.c code has utility functions for dealing with
entire kiovecs at once. The iovec is the scatter-gather unit for IO,
but the individual kiobuf is the unit for mapping of pages. That way,
you can take two different kiobufs, owned by different system
components, and do a single IO on them (eg. take an mmap()ed kiobuf
containing an http header that an application just wrote, and a kiobuf
mapped over the page cache containing a file, and send them over the
wire as a single network packet).
I've considered adding support for a real struct kiovec to allow
passing of kiovec structs as first-class data objects, but for now I
haven't come across anything in the implementation which can't be
handled cleanly enough by a counted vector of kiobufs.
Cheers,
Stephen
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org. For more info on Linux MM,
see: http://www.linux.eu.org/Linux-MM/
^ permalink raw reply [flat|nested] only message in thread