Thanks!
I've pushed the first two parts, improving the shmem accounting.
I've done a couple additional adjustments before commit, namely:
1) I've renamed hash_get_init_size() to just hash_get_size(). We've not
indicated that it's "initial" allocation before, I don't think we need
to indicate to do that now. So I've just removed the "shared" part, as
that's no longer relevant.
2) I've removed one of the compute_buckets_and_segs() in hash_create,
because AFAICS we don't need to recalculate the nsegs - we already have
the correct value in the hctl, so I just used that.
The fact that we call compute_buckets_and_segs() in multiple places,
relying on getting the exact same result seemed a bit fragile to me.
With this we now have just two callers (instead of 3), and I don't think
we can do better - the first call is in hash_get_size(), and I don't see
a convenient way to pass the info init_hash() where we do the 2nd call.
3) In 0002, I've reworked how ProcGlobalShmemSize() calculates the size
of the required memory - the earlier patch added PGProcShmemSize(), but
the ProcGlobalShmemSize() redid the calculation all over. I changed that
to call the new function, and also added a new FastPathLockShmemSize()
to do the same for the fast-path lock arrays (which I realized is now
allocated as a separate area, so it'll be shown separately).
4) I realized PGProcShmemSize() did the calculation slightly differently
from ProcGlobalShmemSize() - by directly multiplying the values, not
using mul_size(). I don't recall why we use mul_size(), but it seems
like the "usual" way. So I did it that way.
I notice we still do the plain multiplication later when setting the
ProcGlobal fields. Seems a bit weird, but we always did that - the patch
does not really change that.
I'll now mark this as committed. I haven't done about the alignment. My
conclusion from the discussion was we don't quite need to do that, but
if we do I think it's a matter for a separate patch - perhaps something
like the 0003.
Thanks for the patch, reviews, etc.
--
Tomas Vondra
