SMP problem with uma_zalloc

[Bosko Milekic]

On Mon, Jul 21, 2003 at 09:03:00AM +0200, Harti Brandt wrote:
> On Sat, 19 Jul 2003, Bosko Milekic wrote:
> 
> BM>
> BM>On Sat, Jul 19, 2003 at 08:31:26PM +0200, Lara & Harti Brandt wrote:
> BM>[...]
> BM>> Well the problem is, that nothing is starved. I have an idle machine and
> BM>> a zone that I have limited to 60 or so items. When allocating the 2nd
> BM>> item I get block on the zone limit. Usually I get unblocked whenever I
> BM>> free an item. This will however not happen, because I have neither
> BM>> reached the limit nor is there memory pressure in the system to which I
> BM>> could react. I simply may be blocked forever.
> BM>
> BM>  UMA_ZFLAG_FULL is set on the zone prior to the msleep().  This means
> BM>  that the next free will result in your wakeup, as the next free will
> BM>  be sent to the zone internally, and not the pcpu cache.
> 
> But there is no free to come. To explain where we have the problem:
> 
> the HARP ATM code uses a zone in the IP code to allocate control blocks
> for VCCs. The zone is limited to 100 items which evaluates to 1 page.
> When I start an interface, first the signalling vcc=5 is opened. This
> allocates one item from the zone, all the other items go into the CPU
> cache. Next I start ILMI. ILMI tries to open its vcc=16. While this works
> on UP machines (the zone allocator will find a free item in the CPU
> cache), on my 2-proc machine half of the time ILMI gets blocked on the
> zonelimit. And it blocks there forever, because, of course nobody is going
> to free the one and only allocated item. On a four processor machine the
> blocking probability will be 75%.
> 
> So in order to be able to get out N items from a zone (given that there is
> no shortage of memory) one has to set the limit to N + nproc *
> items_per_allocation, which one cannot do because he doesn't know
> items_per_allocation.

  It sounds to me like your example is really not the general-case one.
  Basically, you're using a zone capped off at 1 page.  Currently in
  UMA, this is the size of the slab.  So, basically, you have this whole
  zone (with all associated overhead) so as to serve a maximum of only
  one slab.  This defeats most of the assumptions made when the zone is
  created with PCPU caches.  The zone maximum exists to prevent more
  than the specified amount of resources to be allocated toward the
  given zone; I don't think that the intention was "to ensure that if
  the maximum items aren't allocated, there will always be one
  available," despite the fact that that is the effective behavior on
  UP.

  The solution to your really small zone problem is to either make the
  zone bigger, or to hack at UMA to export the UMA_ZONE_INTERNAL API
  properly so that you can skip the pcpu caches for all allocations and
  go straight to the zone.  I'd suggest that you make the zone bigger,
  unless there's a Really Good reason not to.

  In mb_alloc (for mbufs) I had implemented something that in this sort
  of scenario would dip into the other caches and transfer over what I
  called a "bucket" to the current cpu cache.  Although in this
  scenario, it seems like that sort of solution would do what you want,
  some more thought into its behavior reveals that in fact it pessimizes
  the situation.  To give you a better idea, let's consider what happens
  in this specific scenario, where a "bucket" would be all of a page.
  The allocator would make an attempt to allocate from its pcpu cache
  but would find it empty, so it would then attempt to steal a bucket
  from the second cpu's cache.  There, it would find the bucket, move it
  to its cpu's cache, and grab an item from it.  However, a thread on
  the second cpu may then attempt to grab an item, and the bucket will
  just ping-pong from pcpu cache to pcpu cache; the problem that the
  allocator was trying to solve for such really small zones was in fact
  still there - because of the general assumptions made in the design
  with respect to the size of most zones that it dealt with - only
  instead of failing the allocation, it was pessimizing it.

> harti

Regards,
-- 
Bosko Milekic  *  bmilekic at technokratis.com  *  bmilekic at FreeBSD.org
TECHNOkRATIS Consulting Services  *  http://www.technokratis.com/