Date: Thu, 13 Mar 2014 18:50:21 -0700 From: John-Mark Gurney <jmg@funkthat.com> To: Rick Macklem <rmacklem@uoguelph.ca> Cc: Freebsd hackers list <freebsd-hackers@freebsd.org>, Garrett Wollman <wollman@freebsd.org> Subject: Re: kernel memory allocator: UMA or malloc? Message-ID: <20140314015021.GN32089@funkthat.com> In-Reply-To: <1783335610.22308389.1394749339304.JavaMail.root@uoguelph.ca> References: <20140313054659.GG32089@funkthat.com> <1783335610.22308389.1394749339304.JavaMail.root@uoguelph.ca>
next in thread | previous in thread | raw e-mail | index | archive | help
Rick Macklem wrote this message on Thu, Mar 13, 2014 at 18:22 -0400:
> John-Mark Gurney wrote:
> > Rick Macklem wrote this message on Wed, Mar 12, 2014 at 21:59 -0400:
> > > John-Mark Gurney wrote:
> > > > Rick Macklem wrote this message on Tue, Mar 11, 2014 at 21:32
> > > > -0400:
> > > > > I've been working on a patch provided by wollman@, where
> > > > > he uses UMA instead of malloc() to allocate an iovec array
> > > > > for use by the NFS server's read.
> > > > >
> > > > > So, my question is:
> > > > > When is it preferable to use UMA(9) vs malloc(9) if the
> > > > > allocation is going to be a fixed size?
> > > >
> > > > UMA has benefits if the structure size is uniform and a non-power
> > > > of
> > > > 2..
> > > > In this case, it can pack the items more densely, say, a 192 byte
> > > > allocation can fit 21 allocations in a 4k page size verse malloc
> > > > which
> > > > would round it up to 256 bytes leaving only 16 per page... These
> > > > counts per page are probably different as UMA may keep some
> > > > information
> > > > in the page...
> > > >
> > > Ok, this one might apply. I need to look at the size.
> > >
> > > > It also has the benefit of being able to keep allocations "half
> > > > alive"..
> > > > "freed" objects can be partly initalized with references to
> > > > buffers
> > > > and
> > > > other allocations still held by them... Then if the systems needs
> > > > to
> > > > fully free your allocation, it can, and will call your function
> > > > to
> > > > release these remaining resources... look at the ctor/dtor
> > > > uminit/fini
> > > > functions in uma(9) for more info...
> > > >
> > > > uma also allows you to set a hard limit on the number of
> > > > allocations
> > > > the zone provides...
> > > >
> > > Yep. None of the above applies to this case, but thanks for the
> > > good points
> > > for a future case. (I've seen where this gets used for the
> > > "secondary zone"
> > > for mbufs+cluster.)
> > >
> > > > Hope this helps...
> > > >
> > > Yes, it did. Thanks.
> > >
> > > Does anyone know if there is a significant performance difference
> > > if the allocation
> > > is a power of 2 and the "half alive" cases don't apply?
> >
> > From my understanding, the malloc case is "slightly" slower as it
> > needs to look up which bucket to use, but after the lookup, the
> > buckets
> > are UMA, so the performance will be the same...
> >
> > > Thanks all for your help, rick
> > > ps: Garrett's patch switched to using a fixed size allocation and
> > > using UMA(9).
> > > Since I have found that a uma allocation request with M_WAITOK
> > > can get the thread
> > > stuck sleeping in "btalloc", I am a bit shy of using it when
> > > I've never
> >
> > Hmm... I took a look at the code, and if you're stuck in btalloc,
> > either pause(9) isn't working, or you're looping, which probably
> > means
> > you're really low on memory...
> >
> Well, this was an i386 with the default of about 400Mbytes of kernel
> memory (address space if I understand it correctly). Since it seemed
> to persist in this state, I assumed that it was looping and, therefore,
> wasn't able to find a page sized and page aligned chunk of kernel
> address space to use. (The rest of the system was still running ok.)
It looks like vm.phys_free would have some useful information about the
availability of free memory... I'm not sure if this is where the
allocators get their memory or not.. I was about to say it seamed weird
we only have 16K as the largest allocation, but that's 16MEGs..
> I did email about this and since no one had a better explanation/fix,
> I avoided the problem by using M_NOWAIT on the m_getjcl() call.
>
> Although I couldn't reproduce this reliably, it seemed to happen more
> easily when my code was doing a mix of MCLBYTES and MJUMPAGESIZE cluster
> allocation. Again, just a hunch, but maybe the MCLBYTE cluster allocations
> were fragmenting the address space to the point where a page sized chunk
> aligned to a page boundary couldn't be found.
By definition, you would be out of memory if there is not a page free
(that is aligned to a page boundary, which all pages are)...
It'd be interesting to put a printf w/ the pause to see if it is
looping, and to get a sysctl -a from the machine when it is happening...
> Alternately, the code for M_WAITOK is broken in some way not obvious
> to me.
>
> Either way, I avoid it by using M_NOWAIT. I also fall back on:
> MGET(..M_WAITOK);
> MCLGET(..M_NOWAIT);
> which has a "side effect" of draining the mbuf cluster zone if the
> MCLGET(..M_NOWAIT) fails to get a cluster. (For some reason m_getcl()
> and m_getjcl() do not drain the cluster zone when they fail?)
Why aren't you using m_getcl(9) which does both of the above automaticly
for you? And is faster, since there is a special uma zone that has
both an mbuf and an mbuf cluster paired up already?
> One of the advantages of having very old/small hardware to test on;-)
:)
> > > had a problem with malloc(). Btw, this was for a pagesize
> > > cluster allocation,
> > > so it might be related to the alignment requirement (and
> > > running on a small
> > > i386, so the kernel address space is relatively small).
> >
> > Yeh, if you put additional alignment requirements, that's probably
> > it,
> > but if you needed these alignment requirements, how was malloc
> > satisfying your request?
> >
> This was for a m_getjcl(MJUMAGEIZE, M_WAITOK..), so for this case
> I've never done a malloc(). The code in head (which my patch uses as
> a fallback when m_getjcl(..M_NOWAIT..) fails does (as above):
> MGET(..M_WAITOK);
> MCLGET(..M_NOWAIT);
When that fails, an netstat -m would also be useful to see what the
stats think of the availability of page size clusters...
> > > I do see that switching to a fixed size allocation to cover the
> > > common
> > > case is a good idea, but I'm not sure if setting up a uma zone
> > > is worth
> > > the effort over malloc()?
> >
> > I'd say it depends upon how many and the number... If you're
> > allocating
> > many megabytes of memory, and the wastage is 50%+, then think about
> > it, but if it's just a few objects, then the coding time and
> > maintenance isn't worth it..
> >
> Btw, I think the allocation is a power of 2. (It is a power of 2 times
> sizeof(struct iovec) and it looks to me that sizeof(struct iovec) is
> a power of 2 as well. (I know i386 is 8 and I think most 64bits arches
> will make it 16, since it is a pointer and a size_t.)
yes, struct iovec is 16 on amd64...
(kgdb) print sizeof(struct iovec)
$1 = 16
> This was part of Garrett's patch, so I'll admit I would have been to
> lazy to do it.;-) Now it's in the current patch, so unless there seems
> to be a reason to take it out..??
>
> Garrett mentioned that UMA(9) has a per-CPU cache. I'll admit I don't
> know what that implies?
a per-CPU cache means that on an SMP system, you can lock the local
pool instead of grabing a global lock.. This will be MUCH faster as
the local lock won't have to bounce around CPUs like a global lock
does, plus it should never contend which really puts the breaks on
sync primities...
> - I might guess that a per-CPU cache would be useful for items that get
> re-allocated a lot with minimal change to the data in the slab.
> --> It seems to me that if most of the bytes in the slab have the
> same bits, then you might improve hit rate on the CPU's memory
> caches, but since I haven't looked at this, I could be way off??
caching will help some, but the lock is the main one...
> - For this case, the iovec array that is allocated is filled in with
> different mbuf data addresses each time, so minimal change doesn`t
> apply.
So, this is where a UMA half alive object could be helpful... Say that
you always need to allocate an iovec + 8 mbuf clusters to populate the
iovec... What you can do is have a uma uminit function that allocates
the memory for the iovec and 8 mbuf clusters, and populates the iovec
w/ the correct addresses... Then when you call uma_zalloc, the iovec
is already initalized, and you just go on your merry way instead of
doing all that work... when you uma_zfree, you don't have to worry
about loosing the clusters as the next uma_zalloc might return the
same object w/ the clusters already present... When the system gets
low on memory, it will call your fini function which will need to
free the clusters....
> - Does the per-CPU cache help w.r.t. UMA(9) internal code perf?
>
> So, lots of questions that I don't have an answer for. However, unless
> there is a downside to using UMA(9) for this, the code is written and
> I'm ok with it.
Nope, not really...
--
John-Mark Gurney Voice: +1 415 225 5579
"All that I will do, has been done, All that I have, has not."
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20140314015021.GN32089>