On Thu, Jan 5, 2017 at 5:07 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Robert Haas <robertmhaas@gmail.com> writes:
>> On Thu, Jan 5, 2017 at 4:33 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>>> Better documentation seems required, but really the whole design seems
>>> rather wacko. Backends must agree on numeric tranche IDs, but every
>>> backend has its own copy of the tranche name? How do we even know what
>>> agreement is? And every one has to "register" every tranche ID for
>>> itself? Why in the world isn't registration done *once* and the tranche
>>> name stored in shared memory?
>
>> Well, with the original design, that wasn't feasible, because each
>> backend had to store not only the name (which was constant across all
>> backends) but also the array_base (which might not be, if the locks
>> were stored in DSM) and array_stride. Since commit
>> 3761fe3c20bb040b15f0e8da58d824631da00caa, it would be much easier to
>> do what you're proposing. It's still not without difficulties,
>> though. There are 65,536 possible tranche IDs, and allocating an
>> array of 64k pointers in shared memory would consume half a megabyte
>> of shared memory the vast majority of which would normally be
>> completely unused. So I'm not very enthused about that solution, but
>> you aren't the first person to propose it.
>
> So, um, how do we know that backend A and backend B have the same idea
> about what tranche id 37 means?
Well, if they just call C exposed functions at random with arguments
picked out of a hat, then we don't. But if they use the APIs in the
manner documented in lwlock.h, then we do:
/** There is another, more flexible method of obtaining lwlocks. First, call* LWLockNewTrancheId just once to obtain a
trancheID; this allocates from* a shared counter. Next, each individual process using the tranche should* call
LWLockRegisterTranche()to associate that tranche ID with a name.* Finally, LWLockInitialize should be called just once
perlwlock, passing* the tranche ID as an argument.** It may seem strange that each process using the tranche must
registerit* separately, but dynamic shared memory segments aren't guaranteed to be* mapped at the same address in all
coordinatingbackends, so storing the* registration in the main shared memory segment wouldn't work for that case.*/
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company