Re: Shared memory and processes

On Thu, Apr 28, 2016 at 12:56 PM, <david(at)andl(dot)org> wrote:
> Does Postgres provide a convenient way for one process to pass data to
> another using shared memory?

1. Look at ShmemInitStruct, ShmemInitHash (as in hash table),
ShmemInitQueue in storage/shmem.h. These use memory that is mapped at
the same address in every backend (process) which is convenient for
sharing data structures with internal pointers. The amount of memory
available is fixed at cluster startup however.

2. Look at dsm_XXX in storage/dsm.h. This subsystem provides
segments of memory that is "dynamic" in the sense that it is limited
only by your system's available virtual memory, but you have to
explicitly attach these segment in any backend that wants to access
them by passing a handle around and the memory may be mapped at any
address in each backend, so you need to work harder to build data
structures that reference each other (using offsets rather than
pointers, that kind of thing). DSM segments won't work well if you
try to create large numbers of them, so you'll need to provide a way
to manage the space inside a smallish number of large chunks of DSM
memory yourself (this is a problem I'm working to fix by providing a
general purpose allocator backed by a bunch of DSM segments -- watch
this space). LWLocks (our usual lock primitive for cases where
spinlocks are inappropriate) currently don't work correctly inside DSM
segments (this too will be fixed).

--
Thomas Munro
http://www.enterprisedb.com