Re: [RFC 0/2] new kfifo API

[Stefani Seibold <stefani@seibold.net>]

Am Montag, den 03.08.2009, 20:27 +0200 schrieb Andi Kleen:
> Stefani Seibold <stefani@seibold.net> writes:
> >   the buffer. No extra indirection is needed to access the fifo buffer.
> > - Lockless access: if only one reader and one writer is active on the fifo,
> >   which is the common use case, there is no additional locking necessary.
> 
> Would it also be NMI safe?  I've been looking at a new log buffer
> for MCE/NMI. One option was the fifo in ftrace, but it seems
> so big that it would blow up the machine check code considerably.
> 

Yes, it is. If you have only one reader and one writer there is no
locking necessary.

> There was also an own new fifo that Ying Huang implemented,
> but that one wasn't very popular.
> 
> This might be indeed an alternative. Requirement is NMI-safeness.
> 
> > The API:
> > --------
> >
> > struct kfifo *kfifo_alloc(unsigned long size, gfp_t gfp_mask)
> >  Dynamically allocates a new fifo and returns the address
> >  @size: the size of the internal buffer to be allocated.
> >  @gfp_mask: get_free_pages mask, passed to kmalloc()
> 
> For the MCE use case this would need to be able to optionally use
> bootmem because the first initialization happens too early.
> 

You can use a global variable for your fifo if you like. Or you can
greate a fifo inside a bootmem object. Which this API you have the
freedom of choice ;-)

> -Andi
> 

Stefani