On Tue, 10 Jun 1997 gt1355b@prism.gatech.edu wrote:
> This is on my home computer, which is a Pentium with everything
> up-to-date, most things compiled as modules (ethernet, cdrom, sound
> card, printer). It's stock 2.1.42 except for the mmap patch (which it
> needs to stay up for more than a couple of hours).
OK, I've found the patch you mean at LinuxHQ.
> Here's the first crash:
>
> kmem_cache_reap() called within int!
> Scheduling in interrupt
Yeek! try_to_free_page() is being called during an interrupt!
> Using `/build/kernel/linux-2.1.42/System.map' to map addresses to symbols.
>
> Trace: c0145a8f <dev_alloc_skb+f/28>
...and here is the call.
The priority is GFP_ATOMIC, but the mmap patch doesn't honour this.
> Here's the second error:
>
> kmem_cache_reap() called within int!
> last message repeated 2 times
> Aiee, killing interrupt handler
> eth0: Tx request while isr active.
Same problem as above.
I do not have a clean 2.1.42 tree to apply the patch to, but from reading
the patch, mm/page_alloc.c/__get_free_page() will call try_to_free_page()
if the pri is GFP_ATOMIC and the free-pool cannot honour the request..
This is a big NO as the functions called by try_to_free_page() can block -
hence the panics!
The solution is to stop the call for GFP_ATOMIC (and for GFP_BUFFER when
the nr_free_pages is low).
Try changing the code for __get_free_page() to look like;
repeat:
spin_lock_irqsave(&page_alloc_lock, flags);
if ((priority==GFP_ATOMIC) || nr_free_pages > reserved_pages) {
RMQUEUE(order, dma);}
}
spin_unlock_irqrestore(&page_alloc_lock, flags);
if (priority==GFP_ATOMIC || priority==GFP_BUFFER)
return NULL;
if (!tries) {
Hopefully I'm reading the mmap diff correctly :(
Perhaps someone with a tree who can apply the mmap-diff, can make a
patch for this change (and check my suggested mod). Hopefully the orignal
author will pick this up....
Regards,
markhe
-----------------------------------------------------
Mark Hemment UNIX/C Software Engineer (contractor)
"Success has many fathers. Failure is a b**tard"
-----------------------------------------------------