Re: crash in pipe_create()

[Andrew Doran <ad%netbsd.org@localhost>]

On Tue, Feb 10, 2009 at 07:23:26PM +0900, enami tsugutomo wrote:
> Andrew Doran <ad%netbsd.org@localhost> writes:
> 
> >                                                       I think it would be
> > good for pool_cache_get() to invalidate the pool_cache if it cannot fetch
> > an item from the pool. This would flush per-CPU caches back to the pool.
> > 
> >     item = pool_get(&pc->pc_pool, PR_NOWAIT);
> >     if (__predict_false(item == NULL) && (flags & PR_WAITOK) != 0) {
> >             mutex_enter(&pc->pc_lock);
> >             inv = ratecheck(&pc->pc_invaltime, &pool_cache_invaltime);
> >             mutex_exit(&pc->pc_lock);
> >             if (inv) {
> >                     pool_cache_invalidate(pc);
> >             }
> >             item = pool_get(&pc->pc_pool, PR_WAITOK);
> >     }
> 
> Sounds cool.  (But pool_cache_invalidate() doesn't invalidate per CPU
> cache does it?)

Ah right, I had forgotten. It was done that way to satisfy the requirements
of the existing callers. It could be made to invalidate per-CPU caches.

There are only two calls left in the kernel:

- The x86 pmap no longer needs the call.

- The alpha pmap's usage appears broken. At a glance it seems to assume a
  non-preemptive non-MP kernel, i.e. that pool_cache_invalidate is part
  of a greater atomic unit of work (pmap_growkernel).

Andrew