[lkml]   [2018]   [May]   [5]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH] percpu_ida: Use _irqsave() instead of local_irq_save() + spin_lock
On Sat, May 05, 2018 at 08:10:20AM -0600, Jens Axboe wrote:
> On 5/4/18 9:51 PM, Matthew Wilcox wrote:
> > On Fri, May 04, 2018 at 04:22:16PM -0700, Andrew Morton wrote:
> >> I'm feeling a bit hostile toward lib/percpu_ida.c in general ;) It has
> >> very few users and seems rather complicated (what's with that
> >> schedule() in percpu_ida_alloc?). I'm suspecting and hoping that if
> >> someone can figure out what the requirements were, this could all be
> >> zapped and reimplemented using something else which we already have.
> >
> > Note that I have no code in percpu_ida ... it's quite different from
> > the regular IDA. But I have noticed the stunning similarity between the
> > percpu_ida and the code in lib/sbitmap.c. I have no idea which one is
> > better, but they're essentially doing the same thing.
> Not sure where you see that "stunning similarity"? The sbitmap code is
> basically the blk-mq tagging sparse bitmaps, abstracted into a generally
> usable form. The percpu_ida design works fine for lower utilization, but
> it fell apart for the tagging use case where we can easily run at full
> utilization. percpu_ida has percpu caches, sbitmap gets away with just
> percpu hints. These caches are why it doesn't work well for > 50%
> utilization. sbitmap also supports shallow operations, and online
> resizing. Outside of the sharing the same basic functionality of "give
> me some free ID", I really don't see a lot of similarities. In terms of
> functionality, yes, I don't think it would be hard to get rid of
> percpu_ida and just replace it with sbitmap. Probably a worthwhile
> pursuit.

Yes, I meant stunning similarity in terms of functionality, rather than
implementation. I didn't intend to imply that you'd filed off the serial
numbers, given it a fresh coat of paint and called it yours ;-)

I've been looking into what it'll take to replace percpu_ida with sbitmap.
The good news is that there's large chunks of the percpu_ida API that
aren't being used, and the better news is that there's actually only
one percpu_ida, although it gets used by a lot of target drivers.

Looking at the functions in the header file ...

percpu_ida_alloc - seven drivers, all sess_tag_pool
percpu_ida_free - seven drivers, all sess_tag_pool
percpu_ida_destroy - target_core_transport.c (sess_tag_pool)
percpu_ida_init - target_core_transport.c (sess_tag_pool)
percpu_ida_for_each_free - unused
percpu_ida_free_tags - unused

percpu_ida_alloc uses 'state' in a little bit of an unusual way. It seems
to me that TASK_RUNNING means "Do not sleep", and any other value means
"sleep in this TASK_ state". As I understand the sbitmap code, that
means we want an sbitmap_queue.

init and destroy seem to map to sbitmap_queue_init_node and
sbitmap_queue_free. percpu_ida_free maps to sbitmap_queue_clear.
percpu_ida_alloc(x, TASK_RUNNING) maps to sbitmap_queue_get, and any
other state is going to involve the kind of code we see in blk_mq_get_tag.

Does all of that make sense, or have I missed something?

And, Kent, do you see any reason to keep percpu_ida around? Is there
an important way in which it's superior to sbitmap?

 \ /
  Last update: 2018-05-05 17:52    [W:0.088 / U:0.776 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site