Re: recursive spinlocks. Shoot.

[Johannes Erdfelt]

On Mon, May 19, 2003, Pete Zaitcev <[e-mail]> wrote:
> >> Let's quote the example from rubini & corbet of the sbull block device
> >> driver. The request function ends like so:
> > 
> > defective locking in a driver is no excuse to pamper over it with
> > recusrive shite.
> 
> Arjan is a little too harsh here, but on the principle I happen
> to agree, because I worked with systems which allow recursive locks.
> They very often cover up for programmer's lack of basic understanding.
> Worse, sometimes even experienced programmers can do poorly.
> I ran into the latter cathegory of code when fixing so-called
> "presto" in Solaris (now replaced by Encore-originated code).
> 
> Normal spinlocks are not without problems, in particular people
> tend to write:
> 
>    void urb_rm_priv_locked(struct urb *) {
>      ......
>    }
>    void urb_rm_priv(struct urb *u) {
>      spin_lock_irqsave();
>      urb_rm_prin_locked(u);
>      spin_unlock_irqrestore();
>    }
> 
> Which eats a stack frame. We make this tradeoff on purpose,
> as a lesser evil.
> 
> BTW, I do not see Linus and his leutenants rebuking the onslaught
> of recursive ingenuity in this thread. Ignoring the hogwash,
> or waiting and watching?

If past experience is any example, I don't think Linus is completely
against recursive spinlocks.

The uhci driver used a simple implementation at one point in time
because of a tricky locking situation. We eventually discovered a non
recursive method of handling it and ditched the code.

Linus actually helped with the code a little bit.

That being said, I'm happy we found an alternative solution and ditched
the recursive spinlock code. I agree with much of your sentiments about
them as well.

JE