page-based i/o

[Dean Gaudet <dg...@arctic.org>]

Ed asked me for more details on what I mean when I talk about "paged based
zero copy i/o". 

While writing mod_mmap_static I was thinking about the primitives that the
core requires of the filesystem.  What exactly is it that ties us into the
filesystem?  and how would we abstract it?  The metadata (last modified
time, file length) is actually pretty easy to abstract.  It's also easy to
define an "index" function so that MultiViews and such can be implemented. 
And with layered I/O we can hide the actual details of how you access
these "virtual" files. 

But therein lies an inefficiency.  If we had only bread() for reading
virtual files, then we would enforce at least one copy of the data. 
bread() supplies the place that the caller wants to see the data, and so
the bread() code has to copy it.  But there's very little reason that
bread() callers have to supply the buffer... bread() itself could supply
the buffer.  Call this new interface page_read().  It looks something like
this:

    typedef struct {
	const void *data;
	size_t data_len; /* amt of data on page which is valid */
	... other stuff necessary for managing the page pool ...
    } a_page_head;

    /* returns NULL if an error or EOF occurs, on EOF errno will be
     * set to 0
     */
    a_page_head *page_read(BUFF *fb);

    /* queues entire page for writing, returns 0 on success, -1 on
     * error
     */
    int page_write(BUFF *fb, a_page_head *);

It's very important that a_page_head structures point to the data page
rather than be part of the data page.  This way we can build a_page_head
structures which refer to parts of mmap()d memory.

This stuff is a little more tricky to do, but is a big win for performance.
With this integrated into our layered I/O it means that we can have
zero-copy performance while still getting the advantages of layering.

But note I'm glossing over a bunch of details... like the fact that we
have to decide if a_page_heads are shared data, and hence need reference
counting (i.e. I said "queues for writing" up there, which means some
bit of the a_page_head data has to be kept until its actually written).
Similarly for the page data.

There are other tricks in this area that we can take advantage of --
like interprocess communication on architectures that do page flipping.
On these boxes if you write() something that's page-aligned and page-sized
to a pipe or unix socket, and the other end read()s into a page-aligned
page-sized buffer then the kernel can get away without copying any data.
It just marks the two pages as shared copy-on-write, and only when
they're written to will the copy be made.  So to make this work, your
writer uses a ring of 2+ page-aligned/sized buffers so that it's not
writing on something the reader is still reading.

Dean