SMP problem with uma_zalloc

[Harti Brandt]

On Mon, 21 Jul 2003, Bosko Milekic wrote:

BM>
BM>On Mon, Jul 21, 2003 at 09:03:00AM +0200, Harti Brandt wrote:
BM>> On Sat, 19 Jul 2003, Bosko Milekic wrote:
BM>>
BM>> BM>
BM>> BM>On Sat, Jul 19, 2003 at 08:31:26PM +0200, Lara & Harti Brandt wrote:
BM>> BM>[...]
BM>> BM>> Well the problem is, that nothing is starved. I have an idle machine and
BM>> BM>> a zone that I have limited to 60 or so items. When allocating the 2nd
BM>> BM>> item I get block on the zone limit. Usually I get unblocked whenever I
BM>> BM>> free an item. This will however not happen, because I have neither
BM>> BM>> reached the limit nor is there memory pressure in the system to which I
BM>> BM>> could react. I simply may be blocked forever.
BM>> BM>
BM>> BM>  UMA_ZFLAG_FULL is set on the zone prior to the msleep().  This means
BM>> BM>  that the next free will result in your wakeup, as the next free will
BM>> BM>  be sent to the zone internally, and not the pcpu cache.
BM>>
BM>> But there is no free to come. To explain where we have the problem:
BM>>
BM>> the HARP ATM code uses a zone in the IP code to allocate control blocks
BM>> for VCCs. The zone is limited to 100 items which evaluates to 1 page.
BM>> When I start an interface, first the signalling vcc=5 is opened. This
BM>> allocates one item from the zone, all the other items go into the CPU
BM>> cache. Next I start ILMI. ILMI tries to open its vcc=16. While this works
BM>> on UP machines (the zone allocator will find a free item in the CPU
BM>> cache), on my 2-proc machine half of the time ILMI gets blocked on the
BM>> zonelimit. And it blocks there forever, because, of course nobody is going
BM>> to free the one and only allocated item. On a four processor machine the
BM>> blocking probability will be 75%.
BM>>
BM>> So in order to be able to get out N items from a zone (given that there is
BM>> no shortage of memory) one has to set the limit to N + nproc *
BM>> items_per_allocation, which one cannot do because he doesn't know
BM>> items_per_allocation.
BM>
BM>  It sounds to me like your example is really not the general-case one.
BM>  Basically, you're using a zone capped off at 1 page.  Currently in
BM>  UMA, this is the size of the slab.  So, basically, you have this whole
BM>  zone (with all associated overhead) so as to serve a maximum of only
BM>  one slab.  This defeats most of the assumptions made when the zone is
BM>  created with PCPU caches.  The zone maximum exists to prevent more
BM>  than the specified amount of resources to be allocated toward the
BM>  given zone; I don't think that the intention was "to ensure that if
BM>  the maximum items aren't allocated, there will always be one
BM>  available," despite the fact that that is the effective behavior on
BM>  UP.
BM>
BM>  The solution to your really small zone problem is to either make the
BM>  zone bigger, or to hack at UMA to export the UMA_ZONE_INTERNAL API
BM>  properly so that you can skip the pcpu caches for all allocations and
BM>  go straight to the zone.  I'd suggest that you make the zone bigger,
BM>  unless there's a Really Good reason not to.

I think I take two paths: for stuffs like VCC where there may be a large
number I will just remove the limit. The limits were a leftover when the
ATM code had its own memory pool code. For stuff where there is a high
probability that only a handful (usually 1 or 2) of them will be allocated
(network interfaces) I will try to make it to use malloc().

How do you think about adding a paragraph for uma_zone_set_max to the man
page?:

An upper limit of items in the zone can be specified with a call to
uma_zone_set_max. This limits the total number of items which includes:
allocated items, free items and free items in the per-cpu caches. On
systems with more than one CPU it may not be possible to allocate the
specified number of items, because all of the remaining free items may
be in the caches of the other CPUs when the limit is hit.

Regards,
harti

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

-- 
harti brandt,
http://www.fokus.fraunhofer.de/research/cc/cats/employees/hartmut.brandt/private
brandt at fokus.fraunhofer.de, harti at freebsd.org