On Sun, 18 Sep 2016 06:33:52 +0100
Al Viro [off-list ref] wrote:

[finally Cc'd to fsdevel - should've done that several iterations upthread]

On Wed, Sep 14, 2016 at 01:39:25PM +1000, Nicholas Piggin wrote:

quoted

Should not be so bad, but I don't have hard numbers for you. PAGEVEC_SIZE
is 14, and that's conceptually rather similar operation (walk radix tree;
grab pages). OTOH many archs are heavier and do locking and vmas walking etc.

Documentation/features/vm/pte_special/arch-support.txt

But even for those, at 16 entries, the bulk of the cost *should* be hitting
struct page cachelines and refcounting. The rest should mostly stay in cache.  

OK...  That's actually important only for vmsplice_to_pipe() and 16-page
array seems to be doing fine there.

Another question, now that you've finally resurfaced: could you reconstruct
the story with page-stealing and breakage(s) thereof that had lead to
commit 485ddb4b9741bafb70b22e5c1f9b4f37dc3e85bd
Author: Nick Piggin [off-list ref]
Date:   Tue Mar 27 08:55:08 2007 +0200

    1/2 splice: dont steal

I realize that it had been 9 years ago, but anything resembling a braindump
would be very welcome.  Note that there is a couple of ->splice_write()
instances that _do_ use ->steal() (fuse_dev_splice_write() and virtio_console
port_fops_splice_write()) and I wonder if they suffer from the same problems;
your commit message is rather short on details, unfortunately.  FUSE one
is especially interesting...

Without looking through all the patches again, I believe the issue was
just that filesystems were not expecting (or at least, not audited to
expect) pages being added to their pagecache in that particular state
(they'd expect to go through ->readpage or see !uptodate in prepare_write).

If some wanted to attach metadata to uptodate pages for example, this
may have caused a problem. It wasn't some big fundamental problem, just a
mechanical one.

Thanks,
Nick'