Re: Splicing to/from a tty

[Greg Kroah-Hartman <gregkh@linuxfoundation.org>]

On Mon, Jan 18, 2021 at 05:38:55PM -0800, Linus Torvalds wrote:
> On Mon, Jan 18, 2021 at 2:20 PM Linus Torvalds
> <torvalds@linux-foundation.org> wrote:
> >
> > So it's not a "real" patch, but with improved buffer handling in
> > tty_read(), I think this is actually quite close.
> 
> Hmm.
> 
> I somehow ended up working on this all because it's a Monday, and I
> don't see a lot of pull requests early in the week.
> 
> And I think I have a solution for the HDLC "we may need to copy a
> packet that might be up to 64kB" issue, that isn't really all that
> ugly.
> 
> We can just iterate over a random "cookie" that the line discipline
> can use any way it wants to. In the case of n_hdlc, it can just put
> the 'rbuf' thing it has into that cookie, and then it can copy it all
> piece-meal until it is all used up. And if it runs out of space in the
> middle, it will return -EOVERFLOW, and we're all good.
> 
> The only other thing such a line discipline needs is the offset into
> the cookie, but the iterator has to maintain that anyway, so that's
> simple enough.
> 
> So here's a fourth patch for this thing today, this time with what I
> think is actually a working model for the buffer handling.
> 
> Other line disciplines *could* use the cookie if they want to. I
> didn't do any of that, though.
> 
> The normal n_tty line discipline, for example, could easily just loop
> over the data. It doesn't need an offset or that 'rbuf' pointer, but
> it still needs to know whether the call is the first one or not,
> because the first time the n_tty line discipline is called it may need
> to wait for a minimum number of characters or whatever the termios
> settings say - but obviously once it has waited for it once, it
> shouldn't wait for it again the next time around (only on the next
> actual full read()). IOW, it would be wrong if the termios said "wait
> for 5 characters", and then it saw 68 characters, copied the first 64,
> in the first iteration, and than saw "oh, now there are only 4
> characters left so now I have to wait for a fifth".
> 
> So n_tty could use the cookie purely to see whether it's the first
> iteration or not, and it could just set the cookie to a random value
> (it always starts out as NULL) to just show what state it is in.
> 
> I did *NOT* do that, because it's not technically necessary - unlike
> the hdlc packet case, n_tty returning a partial result is not wrong
> per se even if we might have reasons to improve on it later.
> 
> What do people think about this?
> 
> Also, does anybody have any test-code for the HDLC case? I did find an
> interesting comment when doing a Debian code search:
> 
>   /* Bloody hell... readv doesn't work with N_HDLC line discipline... GRR! */
> 
> and yes, this model would allow us to handle readv() properly for hdlc
> (and no, the old one did not, because it really wanted to see the
> whole packet in *one* user buffer).
> 
> But I have no idea if hdlc is even relevant any more, and if anybody
> really cares.
> 
> Anybody?


This looks sane, but I'm still missing what the goal of this is here.
It's nice from a "don't make the ldisc do the userspace copy", point of
view, but what is the next step in order to tie that into splice?

I ask as I also have reports that sysfs binary files are now failing for
this same reason, so I need to make the same change for them and it's
not excatly obvious what to do:
	https://lore.kernel.org/r/1adf9aa4-ed7e-8f05-a354-57419d61ec18@codeaurora.org

thanks,

greg k-h