Re: Upcoming: fscache rewrite

[David Howells <dhowells-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>]

Matthew Wilcox <willy-wEGCiKHe2LqWVfeAwA7xHQ@public.gmane.org> wrote:

> I suspect you don't need to call find_get_pages_contig().  If you look
> at __readahead_batch() in pagemap.h, it does basically what you want
> (other than being wrapped up inside the readahead iterator).  You require
> the pages already be pinned in the xarray, so there's no need for the
> page_cache_get_speculative() dance that find_get_pages_contig) does,
> nor the check for xa_is_value().

I'll have a look at that.

> My main concern with your patchset is that it introduces a new page flag

Technically, the flag already exists - I'm just using it for something
different than the old fscache code used it for.

> to sleep on which basically means "I am writing this page to the fscache".
> I don't understand why you need it; you've elevated the refcount on
> the pages so they're not going to get reused for another purpose.
> All it does (as far as I can tell) is make a task calling truncate()
> wait for the page to finish being written to cache, which isn't actually
> necessary.

It's also used to prevent starting overlapping async DIO writes to the cache.

See fscache_read_done(), where it's set to cover writing what we've just read
from the server to the cache, and afs_write_back_from_locked_page(), where
it's set to cover writing the data to be written back to the cache.

David