[SSSD] [PATCH] mmap_cache: Use two chains for hash collision.

Simo Sorce simo at redhat.com
Thu Sep 12 15:54:13 UTC 2013


On Thu, 2013-09-12 at 16:51 +0200, Jakub Hrozek wrote:
> On Thu, Sep 12, 2013 at 09:40:57AM -0400, Simo Sorce wrote:
> > On Thu, 2013-09-12 at 09:05 +0200, Lukas Slebodnik wrote:
> > > On (11/09/13 17:55), Simo Sorce wrote:
> > > >On Wed, 2013-09-11 at 21:21 +0200, Lukas Slebodnik wrote:
> > > >> On (11/09/13 12:50), Simo Sorce wrote:
> > > >> >On Wed, 2013-09-11 at 17:39 +0200, Lukas Slebodnik wrote:
> > > >> >> On (11/09/13 16:30), Michal Židek wrote:
> > > >> >> >On 09/11/2013 04:16 PM, Simo Sorce wrote:
> > > >> >> >>On Wed, 2013-09-11 at 15:58 +0200, Lukas Slebodnik wrote:
> > > >> >> >>>On (11/09/13 09:39), Simo Sorce wrote:
> > > >> >> >>>>On Wed, 2013-09-11 at 12:24 +0200, Lukas Slebodnik wrote:
> > > >> >> >>>>>ehlo,
> > > >> >> >>>>>
> > > >> >> >>>>>Two patches are attached.
> > > >> >> >>>>>
> > > >> >> >>>>>The 1st patch reverts commit 4662725ffef62b3b2502481438effa7c8fef9f80.
> > > >> >> >>>>>"mmap_cache: Skip records which doesn't have same hash"
> > > >> >> >>>>>  --next patch lokks nicer with recerted commit 4662725ffe
> > > >> >> >>>>>
> > > >> >> >>>>>
> > > >> >> >>>>>The 2nd patch changes behaviour of record chaining.
> > > >> >> >>>>>
> > > >> >> >>>>>LS
> > > >> >> >>>>
> > > >> >> >>>>NACK, you cannot just change the record structure w/o bumping the whole
> > > >> >> >>>>mmap MAJOR version number.
> > > >> >> >>>>
> > > >> >> >>>>Aside:
> > > >> >> >>>>
> > > >> >> >>>>If we are not I am thinking whether we should do some other
> > > >> >> >>>>optimizations to keep the size down.
> > > >> >> >>>>
> > > >> >> >>>>For example, hashes are generally relatively small quantities as we do
> > > >> >> >>>>not allow huge tables. If we can limit tables size to a max of 65k
> > > >> >> >>>>elements we could change the hash type from uint32_t to uint16_t, and
> > > >> >> >>>>end up using the same size we are using now.
> > > >> >> >>>>We already limit the total mmap size to 4G as rel_prt_t is a uint32_t,
> > > >> >> >>>>and in practice we'll never allow maps with a very large size as it eats
> > > >> >> >>>>into the memory available to 32 bit processes.
> > > >> >> >>>>So perhaps 65k for the hash tables is an acceptable limit. (I think we
> > > >> >> >>>>currently have a 100k limit which is probably a little bit overblown as
> > > >> >> >>>>a default anyway.
> > > >> >> >>>>
> > > >> >> >>>>Simo.
> > > >> >> >>>>
> > > >> >> >>>I would rather reduce "uint64_t expire" into "uint32_t expire_offset"
> > > >> >> >>>
> > > >> >> >>>expire_offset will be time since creation of mmap cache.
> > > >> >> >>>and uint64_t "creation_time" will be stored in "struct sss_mc_ctx"
> > > >> >> >>
> > > >> >> >>Uhmm, I do agree that 4billion seconds offset is probably ok, but then
> > > >> >> >>it means the information is split in 2 different places, and this will
> > > >> >> >>make clients more complicated, as clients will have to keep track of the
> > > >> >> >>creation time + expiration offset now ...
> > > >> >> >>
> > > >> >> >>>Do you agree?
> > > >> >> >>
> > > >> >> >>I am not 100% convinced yet, although I do agree that it would save
> > > >> >> >>quite some space.
> > > >> >> >>
> > > >> >> >>>Do I need to increment SSS_MC_MAJOR_VNO or only SSS_MC_MINOR_VNO?
> > > >> >> >>
> > > >> >> >>Major, you are completely changing the structure, an older client cannot
> > > >> >> >>be allowed to use the new map unless it explicitly know the layout
> > > >> >> >>changed.
> > > >> >> >
> > > >> >> >Well, if we already have to change the major version, I would go with
> > > >> >> >the two 32 bit next1 and nex2 members. If we really want to keep the
> > > >> >> >current size of the record (btw. why? We have to change the major ver.
> > > >> >> >anyway) then I like the two 16 bit pointers next1 and next2 more than
> > > >> >> >the 32 bit time offset (the memcache is just another cache layer, so it
> > > >> >> >does not have to be "gigantic"... 16 bit pointers are just fine IMO).
> > > >> >> >
> > > >> >> I don't like 16 bit next pointers.
> > > >> >> 
> > > >> >> Default value of hash table size is 100000 (for passwd)
> > > >> >> It is 53% greater value than UINT_16_MAX.
> > > >> >> If we decided to enlarge this default value in future, we would be still
> > > >> >> limited by UINT_16_MAX.
> > > >> >> 
> > > >> >> I think it is not good idea to reduce size of next pointers.
> > > >> >
> > > >> >I am thinking we might want to change the way we create the way we
> > > >> >handle record size to see if we can make a change that will be more
> > > >> >future proof.
> > > >> >
> > > >> >If we stored the size of the records in the header and always calculate
> > > >> >the position of b2 based on that information then we could introduce a
> > > >> >variable size record structure. This would allow in future to increase
> > > >> >the size of the structure for optional parameters w/o necessarily having
> > > >> >to bump up the major version.
> > > >> >
> > > >> >However I want to have at the very least 64bit alignemtn always
> > > >> >preserverd, so from 32 bytes we want to grow in 8 bytes steps, which
> > > >> >means adding a uint32_t padding before b2.
> > > >> totally agree.
> > > >> 
> > > >> SSS_MC_MAJOR_VNO was increased.
> > > >> padding was added. sizeof(struct sss_mc_ctx) will be 40
> > > >> 
> > > >> Updated patches are attached.
> > > >
> > > >It looks ok to me.
> > > >
> > > >However, just to be clear we are planning on pushing this to 1.12 only
> > > >right ?
> > > >
> > > >I am not hugely comfortable with pushing it to existing releases.
> > > >
> > > >Simo.
> > > >
> > > 
> > > The main problem is that existing releases could crash in some rare situations
> > > There is a report in bugzilla one week old.
> > > https://bugzilla.redhat.com/show_bug.cgi?id=997406#c12
> > > 
> > > User reported crash and he had latest sssd 1-10 with all
> > > mmap_cache patches; but it didn't help. sssd_nss crashed.
> > > 
> > > We had a long discussion about solution (revalidation of chains), but we did
> > > not find any simple "back-compatible" fix.
> > > 
> > > How do we want to fix it in older branches If we don't want to backport patch
> > > from previous mail.
> > 
> > The problem is that a database format change is not to be taken lightly
> > as it kills completely the use of the mmap cache until all long lived
> > processes are restarted. If we can detect the problem I would rather
> > nuke the caches.
> > Deletions are rare, they happen only when a user disappear IIRC, so if
> > we have a simple method of detecting when a deletion may cause issue we
> > can simply reset the mmap cache then it would have less dramatic effects
> > I think. But I am open to other ideas.
> > 
> > Simo.
> > 
> 
> Hm, perhaps for this problem we could hack up something like Michal's
> previous patch to detect the problem and cp the corrupt database for
> dissection?

As long as we can reliably detect the problem without returning wrong
answers :)

The best would be to detect on the server side that the operation of
removing the record will probably cause issues and just reset the whole
ccache, a bit drastic indeed but better than having a corrupt chain.

Simo.


-- 
Simo Sorce * Red Hat, Inc * New York




More information about the sssd-devel mailing list