Re: [PATCH v3 0/4] tty: TX helpers

[Ilpo Järvinen]

On Wed, 7 Sep 2022, Greg Kroah-Hartman wrote:

> On Wed, Sep 07, 2022 at 03:32:44PM +0300, Ilpo Järvinen wrote:
> > On Wed, 7 Sep 2022, Greg Kroah-Hartman wrote:
> > 
> > > On Wed, Sep 07, 2022 at 03:21:28PM +0300, Ilpo Järvinen wrote:
> > > > On Wed, 7 Sep 2022, Arnd Bergmann wrote:
> > > > 
> > > > > On Wed, Sep 7, 2022, at 12:16 PM, Ilpo Järvinen wrote:
> > > > > > On Wed, 7 Sep 2022, Jiri Slaby wrote:
> > > > > >> On 06. 09. 22, 13:30, Johan Hovold wrote:
> > > > > >> > On Tue, Sep 06, 2022 at 12:48:01PM +0200, Jiri Slaby wrote:
> > > > > >> > NAK
> > > > > >> 
> > > > > >> I'd love to come up with something nicer. That would be a function in
> > > > > >> serial-core calling hooks like I had [1] for example. But provided all those
> > > > > >> CPU workarounds/thunks, it'd be quite expensive to call two functions per
> > > > > >> character.
> > > > > >> 
> > > > > >> Or creating a static inline (having ± the macro content) and the hooks as
> > > > > >> parameters and hope for optimizations to eliminate thunks (also suggested in
> > > > > >> the past [1]).
> > > > > >> 
> > > > > >> [1] https://lore.kernel.org/all/20220411105405.9519-1-jslaby@xxxxxxx/
> > > > > >
> > > > > > I second Jiri here.
> > > > > >
> > > > > > Saving lines in drivers is not that important compared with all removing 
> > > > > > all the variants of the same thing that have crept there over the years.
> > > > > >
> > > > > > I suspect the main reason for the variants is that everybody just used 
> > > > > > other drivers as examples and therefore we've a few "main" variant 
> > > > > > branches depending on which of the drivers was used as an example for the 
> > > > > > other. That is hardly a good enough reason to keep them different and as 
> > > > > > long as each driver keeps its own function for this, it will eventually 
> > > > > > lead to similar differentiation so e.g. a one-time band-aid similarization 
> > > > > > would not help in the long run.
> > > > > >
> > > > > > Also, I don't understand why you see it unreadable when the actual code is 
> > > > > > out in the open in that macro. It's formatted much better than e.g. 
> > > > > > read_poll_timeout() if you want an example of something that is hardly 
> > > > > > readable ;-). I agree though there's a learning-curve, albeit small, that 
> > > > > > it actually creates a function but that doesn't seem to me as big of an 
> > > > > > obstacle you seem to think.
> > > > > 
> > > > > I think it would help to replace the macro that defines
> > > > > the function with a set of macros that can be used in
> > > > > function bodies. This would avoid the __VA_ARGS__ stuff
> > > > > and allow readers that are unfamiliar with tty drivers to
> > > > > treat it as a function call.
> > > > > 
> > > > > So e.g. instead of 
> > > > > 
> > > > > static DEFINE_UART_PORT_TX_HELPER_LIMITED(altera_jtaguart_do_tx_chars,
> > > > > 		true,
> > > > > 		writel(ch, port->membase + ALTERA_JTAGUART_DATA_REG),
> > > > > 		({}));
> > > > > 
> > > > > the altera_jtaguart driver would contain a function like
> > > > > 
> > > > > static int altera_jtaguart_do_tx_chars(struct uart_port *port,
> > > > >                                        unsigned int count)
> > > > > {
> > > > >        char ch;
> > > > > 
> > > > >        return uart_port_tx_helper_limited(port, ch, count, true,
> > > > >                 writel(ch, port->membase + ALTERA_JTAGUART_DATA_REG),
> > > > >                 ({}));
> > > > > }
> > > > > 
> > > > > or some variation of that. It's a few more lines, but those
> > > > > extra lines would help me understand what is actually going on
> > > > > while still avoiding the usual bugs and duplication.
> > > > > 
> > > > > If the caller of that function is itself trivial (like
> > > > > serial21285_tx_chars), then the intermediate function can
> > > > > be omitted in order to save some of the extra complexity.
> > > > 
> > > > I'd be ok with that. There's still a small startle factor associated to 
> > > > passing that writel(...) as an argument to a "function" but it's the same 
> > > > for other things such as read_poll_timeout() so not an end of the world.
> > > 
> > > That's going to incure the function-pointer-indirection-call for every
> > > character that Jiri's original submission had, so I don't think this is
> > > a very viable solution, sorry.
> > 
> > I don't think you got what Arnd meant. It must still be technically a 
> > #define because you cannot pass
> > 	writel(ch, port->membase + ALTERA_JTAGUART_DATA_REG)
> > as an argument to a real function like he did in the example above.
> > It's similar to how read_poll_timeout() and friends are #defines despite 
> > being lowercased.
> 
> Ok, no, I don't understand what Arnd meant here then :(

I think he did just 2 things (compared with 2/4 of this series):

- __DEFINE_UART_PORT_TX_HELPER() doesn't create a function but just wraps 
the macro body with ({ }). Therefore, a driver must use it inside 
a function rather than the macro creating a new function with name.
- Use lowercase instead of uppercase (this is a trivial change)

When a driver is using these "function" that are just made look functions 
but are macros for real, it will not add function-pointer-indirection-call 
per character but pulls the whole loop from the macro into the function 
in the driver and injects those putchar, etc. directly into that loop. 
...It is just like read_poll_timeout() constructs the loop.

-- 
 i.