Re: [Bug #14388] keyboard under X with 2.6.31

[Linus Torvalds]

On Tue, 13 Oct 2009, Alan Cox wrote:
> 
> There is a simple reason the locking is sufficient. If you can call the
> function from two places at once in your serial driver at the same you've
> scrambled the data order so you've already lost.

Umm. No, Alan.

You also can race with:

 - whoever is _reading_ the buffer, and due to memory ordering may see the 
   update to the buffer length _before_ it actually sees the data itself. 
   That spinlock does all the memory ordering too.

 - scrambling the data order with two writers is certainly less annoying 
   than potentially screwing up ->used entirely, and having the memcpy's 
   overflow the buffer. Both writers may have decided that there is enough 
   room for each one - but that does not mean that there is enough room 
   for _both_.

Now, I do agree that generally there should be locking at a higher level, 
and you should never see two concurrent writers. But even if the locking 
is only for reading, the old locking is simply _wrong_.

> >   pointless: they then call tty_insert_flip_string(), which means that the 
> >   tty_buffer_request_room() call was totally redundant ]
> 
> It's a performance tweak. With a 3G USB modem or similar device running
> at 20Mbits or more being able to generate one allocation per chunk
> received for DMA made a measurable performance difference on some
> platforms. 

Have you even _read_ the code, Alan?

It's not a f*cking performance tweak, and you're ludicrous to claim it is. 
It's pointless, and it's making the code _slower_ rather than faster.

Lookie here, Alan - the common sequence is crap like this:

	tty_buffer_request_room(tty, buf->size);
	tty_insert_flip_string(tty, buf->base, buf->size);

and anybody who claims that is a "performance tweak" doesn't know what the 
hell he is talking about.

Look again.

The first thing that tty_insert_flup_string() does is to re-do the same 
tty_buffer_request_room() call. 

Performance tweak? No. Most of them are stupid, pointless, and worthless. 
Many of them do it for a single character too.

Not all, no. One or two seem to do one tty_buffer_request_room() call, and 
then some one-byte-at-a-time thing, but quite frankly, those are sure as 
hell not going to push lots of data quickly that way either.

Maybe there is some driver where there's a point to it, but from a quick 
grep, I couldn't find any.

			Linus
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/