Re: copy on write for splice() from file to pipe?

From: Stefan Metzmacher
Date: Fri Feb 10 2023 - 15:45:43 EST


Am 09.02.23 um 20:48 schrieb Linus Torvalds via samba-technical:
On Thu, Feb 9, 2023 at 11:36 AM Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:

I guarantee that you will only slow things down with some odd async_memcpy.

Extended note: even if the copies themselves would then be done
concurrently with other work (so "not faster, but more parallel"), the
synchronization required at the end would then end up being costly
enough to eat up any possible win. Plus you'd still end up with a
fundamental problem of "what if the data changes in the meantime".

And that's ignoring all the practical problems of actually starting
the async copy, which traditionally requires virtual to physical
translations (where "physical" is whatever the DMA address space is).

So I don't think there are any actual real cases of async memory copy
engines being even _remotely_ better than memcpy outside of
microcontrollers (and historical computers before caches - people may
still remember things like the Amiga blitter fondly).

Again, the exception ends up being if you can actually use real DMA to
not do a memory copy, but to transfer data directly to or from the
device. That's in some way what 'splice()' is designed to allow you to
do, but exactly by the pipe part ending up being the "conceptual
buffer" for the zero-copy pages.

So this is exactly *why* splicing from a file all the way to the
network will then show any file changes that have happened in between
that "splice started" and "network card got the data". You're supposed
to use splice only when you can guarantee the data stability (or,
alternatively, when you simply don't care about the data stability,
and getting the changed data is perfectly fine).

Ok, thanks for the explanation!

Looking at this patch from David Howells :
https://lore.kernel.org/linux-fsdevel/909202.1675959337@xxxxxxxxxxxxxxxxxxxxxx/
indicates that we don't have that problem with O_DIRECT as it operates
on dedicated pages (not part of the shared page cache). And these pages
might be filled via DMA (depending on the filesystem and block device).

Is my understanding correct?

Together with this patch:
https://lore.kernel.org/linux-fsdevel/20230209102954.528942-4-dhowells@xxxxxxxxxx/

I guess it would be easy to pass a flag (maybe SPLICE_F_FORCE_COPY)
down to generic_file_splice_read() and let it create dedicated pages
and use memcpy() from the page cache to the dedicated pages.
This would mean one memcpy(), but it would allow the pipe to be used
for the splice() to the socket, tee(). This might be easier than
using pread() followed by vmsplice(SPLICE_F_GIFT).

The usage of SPLICE_F_GIFT is very confusing to me...
I found this example in libkcapi using the kernel crypto sockets:
https://github.com/smuellerDD/libkcapi/blob/master/lib/kcapi-kernel-if.c#L324
where it just passes SPLICE_F_GIFT together with an iovec passed from the
caller of the library.

To me it's not clear if the caller can still use it's buffers referenced by
the passed iovec...

metze