Re: Improvements and refactoring in shmem.c - Mailing list pgsql-hackers

On Thu, Jan 29, 2026 at 5:58 PM Heikki Linnakangas <hlinnaka@iki.fi> wrote:
>
> On 29/01/2026 11:56, Ashutosh Bapat wrote:
> > 0001: Adds assertions to InitShmemAccess() and InitShmemAllocation
> > which indicate the conditions, EXEC_BACKEND and IsUnderPostmaster,
> > these functions are expected to be called. I found these annotations
> > to be useful when modifying these functions to handle multiple Shmem
> > segments required by buffer pool resizing project [1].
> >
> > 0002: We use two different methods to pass ShmemIndex and ShmemLock
> > respectively to new backends even though both the structures are
> > allocated before creating named structures in the shared memory. This
> > patch consistently uses the same method - passing via PGShmemHeader.
> > Further the patch gets rid of InitShmemAllocation and moves that code
> > inside InitShmemAccess() itself. Maybe that's overkill but at least we
> > would be able to call InitShmemAllocation() from InitShmemAccess() and
> > declare first as static within shmem.c. That way we avoid a minor risk
> > of InitShmemAllocation() being called twice.
> >
> > We may achieve consistency by passing ShmemIndex through
> > BackendParameter, but I haven't tried that.
>
> Hmm, I agree it's inconsistent currently. And I'd like to reduce the use
> of BackendParameters, I don't like that mechanism.
>
> I don't much like moving the 'shmem_lock' pointer to PGShmemHeader
> either though. It feels like a modularity violation, as no one else than
> shmem.c should be accessing those fields. The same goes for the existing
> 'index' and 'freeoffset' fields really.
>
> Also, does it make sense to have those fields in the "shim" shmem
> segment that PGSharedMemoryCreate() creates? It's a little confusing, we
> don't keep the 'freeoffset' in the shim up-to-date, for example.
>
> One idea is to move all those fields to another struct, see attached.
> What do you think?

This looks good. Good that we got rid of ShmemAllocUnlocked() too.

A nitpick should content_offset be contentOffset (like totalSize) or
content_offset (like dsm_control)? I am ok either way.

A minor discomfort I have is ShmemBase, which is the starting address
that the allocator will use, remains outside of ShmemAllocatorData().
But it doesn't change once set so no need to "share" it through the
memory and also that will create a self-referencing pointer within the
shared memory. So it's fine.

>
> > @@ -35,6 +36,7 @@ typedef struct PGShmemHeader  /* standard header for all Postgres shmem */
> >         Size            freeoffset;             /* offset to first free space */
> >         dsm_handle      dsm_control;    /* ID of dynamic shared memory control seg */
> >         void       *index;                      /* pointer to ShmemIndex table */
> > +       slock_t    *shmem_lock;         /* spinlock for shmem allocation */
> >  #ifndef WIN32                                  /* Windows doesn't have useful inode#s */
> >         dev_t           device;                 /* device data directory is on */
> >         ino_t           inode;                  /* inode number of data directory */
>
> Could this contain the spinlock itself, instead of just a pointer to it?
> Would be a little simpler. (That's moot if we go with my approach though)

Yes, that would have been better just like dsm_control - I missed
that. Not needed with your approach.

--
Best Wishes,
Ashutosh Bapat