On Tue, Jul 27, 2021 at 4:14 AM Andreas Gruenbacher <agruenba@xxxxxxxxxx> wrote:
>
> On Tue, Jul 27, 2021 at 11:30 AM David Laight <David.Laight@xxxxxxxxxx> wrote:
> >
> > Is it actually worth doing any more than ensuring the first byte
> > of the buffer is paged in before entering the block that has
> > to disable page faults?
>
> We definitely do want to process as many pages as we can, especially
> if allocations are involved during a write.
Yeah, from an efficiency standpoint, once you start walking page
tables, it's probably best to just handle as much as you can.
But once you get an error, I don't think it should be "everything is bad".
This is a bit annoying, because while *most* users really just want
that "everything is good", *some* users might just want to handle the
partial success case.
It's why "copy_to/from_user()" returns the number of bytes *not*
written, rather than -EFAULT like get/put_user(). 99% of all users
just want to know "did I write all bytes" (and then checking for a
zero return is a simple and cheap verification of "everything was
ok").
But then very occasionally, you hit a case where you actually want to
know how much of a copy worked. It's rare, but it happens, and the
read/write system calls tend to be the main user of it.
And yes, the fact that "copy_to/from_user()" doesn't return an error
(like get/put_user() does) has confused people many times over the
years. It's annoying, but it's required by those (few) users that
really do want to handle that partial case.
I think this iov_iter_fault_in_readable/writeable() case should do the same.
And no, it's not new to Andreas' patch. iov_iter_fault_in_readable()
is doing the "everything has to be good" thing already.
Which maybe implies that nobody cares about partial reads/writes. Or
it's very very rare - I've seen code that handles page faults in user
space, but it's admittedly been some very special CPU
simulator/emulator checkpointing stuff.
Linus
