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

Jakub Hrozek jhrozek at redhat.com
Thu Sep 12 09:26:13 UTC 2013


On Thu, Sep 12, 2013 at 09:05:07AM +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.

Can you describe why? Are you afraid of regressions? (Genuine
questions, sorry, I still don't know the memcache code well)

Would it help to have a memcache unit test?

> >
> >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.

Right, I'd like to backport a fix for the above problem to all existing
branches if possible.



More information about the sssd-devel mailing list