Messages in this thread | | | From | David Howells <> | Subject | Re: [PATCH 22/43] CacheFiles: Add a hook to write a single page of data to an inode [ver #46] | Date | Thu, 02 Apr 2009 18:22:32 +0100 |
| |
Christoph Hellwig <hch@infradead.org> wrote:
> > I don't think "write_one_page" sounds like a particularly good new > > API addition. > > I also thing it's not a nice one. I still haven't seen a really good > explanation of why it can't just use plain ->write
->write requires:
(1) A struct file. Refer to the ENFILE problem that my patch to do this raised. The problem is that doing a tar of, say, a kernel tree immediately opens ~30000 files internally, and then userspace falls over with ENFILE all over the place. My tests multiply that by 8.
You (at least I think it was you) refused to countenance allocating file structs internally to the kernel that weren't accounted.
(2) A pointer to a buffer. That means kmapping a page for the duration of the write.
->write is quite heavy. For something like ext3 it goes down through quite a few layers, in and out of the fs, VM and VFS. All this takes both time and stack space.
I want to write a whole page from the kernel at a page-boundary in the file. That means it is possible to use an extremely optimised aop to do that.
Really, what I want to do is have the filesystem issue its BIOs directly on the netfs page, but the Ext3 people I talked to at the time thought that would be too difficult to track.
David
| |