Re: [PROPOSAL] Shared Ispell dictionaries

Hello,

On Tue, May 15, 2018 at 05:02:43PM -0400, Robert Haas wrote:
> On Tue, Mar 27, 2018 at 8:19 AM, Arthur Zakirov
> <a(dot)zakirov(at)postgrespro(dot)ru> wrote:
> > Yes, there is dsm_pin_mapping() for this. But it is necessary to keep a
> > segment even if there are no attached processes. From 0003:
> >
> > + /* Remain attached until end of postmaster */
> > + dsm_pin_segment(seg);
> > + /* Remain attached until end of session */
> > + dsm_pin_mapping(seg);
>
> I don't quite understand the problem you're trying to solve here, but:
>
> 1. Unless dsm_pin_segment() is called, a DSM segment will
> automatically be removed when there are no remaining mappings.
>
> 2. Unless dsm_pin_mapping() is called, a DSM segment will be unmapped
> when the currently-in-scope resource owner is cleaned up, like at the
> end of the query. If it is called, then the mapping will stick around
> until the backend exits.

I tried to solve the case when DSM segment remains mapped even a
dictionary was dropped. It may happen in the following situation:

Backend 1:

=# select ts_lexize('english_shared', 'test');
-- The dictionary is loaded into DSM, the segment and the mapping is
pinned
...
-- Call ts_lexize() from backend 2 below
=# drop text search dictionary english_shared;
-- The segment and the mapping is unpinned, see ts_dict_shmem_release()

Backend 2:

=# select ts_lexize('english_shared', 'test');
-- The dictionary got from DSM, the mapping is pinned
...
-- The dictionary was dropped by backend 1, but the mapping still is
pinned

As you can see the DSM still is pinned by backend 2. Later I fixed it by
checking do we need to unping segments. In the current version of the
patch do_ts_dict_shmem_release() is called in
lookup_ts_dictionary_cache(). It unpins segments if text search cache
was invalidated. It unpins all segments, but I think it is ok since
text search changes should be infrequent.

> If you pin the mapping or the segment and later no longer want it
> pinned, there are dsm_unpin_mapping() and dsm_unpin_segment()
> functions available, too. So it seems like what you might want to do
> is pin the segment when it's created, and then unpin it if it's
> stale/obsolete. The latter won't remove it immediately, but will once
> all the mappings are gone.

Yes, dsm_unpin_mapping() and dsm_unpin_segment() will be called when the
dictionary is dropped or altered in the current version of the patch. I
descriped the approach above.

In sum, I think the problem is mostly solved. Backend 2 unpins the
segment in next ts_lexize() call. But if backend 2 doesn't call
ts_lexize() (or other TS function) anymore the segment will remain mapped.
It is the only problem I see for now.

I hope the description is clear. I attached the rebased patch.

--
Arthur Zakirov
Postgres Professional: http://www.postgrespro.com
Russian Postgres Company