Re: [RFC, 2.6] a simple FIFO implementation

[Andrea Arcangeli <andrea@novell.com>]

this is nice, I had to write a ring buffer myself last month for
bootcache (you can find the patch on l-k searching for "bootcache"). It
was fun so I don't mind but certainly it took me a few reboots to make
it work ;)

My main issue with this is that I don't like to use kmalloc, I expect
most people will use a page anyways, I'm using alloc_page myself (and I
may want to switch to vmalloc to get a larger buffer, that's fine for
bootcache since the allocation is in a slow path). I wonder if it worth
to generalize the allocator passing down a callback or something like
that. I can still use kmalloc but it'd be just a waste of some memory
and risk fragmentation for >PAGE_SIZE (OTOH the callback as well will
waste some byte).

The other issue with the locking is that I will not need locking since
I've my own external locking used for other stuff too that serializes
the fifo as well, so I wonder if the "spinlock_t *" could as well be
passed down to kfifo_get so I won't need to allocate the spinlock
structure as well inside the kfifo.  Otherwise I could start to use such
a spinlock inside the kfifo for the external locking too (and then I
could call only the __ functions), that means guys outside your
kfifo.[ch] would use the kfifo->lock which doesn't sound that clean,
kfifo using an external lock passed down by the caller as parameter
looks more robust.