Content-Type: text/plain;
	charset=utf-8
Mime-Version: 1.0 (Mac OS X Mail 16.0 \(3864.400.21\))
Subject: Re: GetCachedPlan() refactor: move execution lock acquisition out
From: Chao Li <li.evan.chao@gmail.com>
In-Reply-To: 
 <CA+HiwqEXyLiCdi7MQgmU8Y+aPVcOf8JpWfVA4szCmJtZAqj7Fg@mail.gmail.com>
Date: Thu, 16 Apr 2026 10:34:58 +0800
Cc: Kirill Reshke <reshkekirill@gmail.com>,
 PostgreSQL-development <pgsql-hackers@postgresql.org>
Content-Transfer-Encoding: quoted-printable
Message-Id: <6B155E97-EDE7-4CBE-8196-E709F6A4B33D@gmail.com>
References: 
 <CA+HiwqE1ntHy2h9zJ9v3MwAkoGAveSERcHWkDTTZnP0kxWqbKQ@mail.gmail.com>
 <CALdSSPh=8qLa1kPsBUnPXMmigaCABiNTwmwg1wH+TmQWH=f_Ww@mail.gmail.com>
 <CA+HiwqH7TcHb=P7qHF-9jnXD3APLUWbD4QS_OJ3d-5qaC+Puzg@mail.gmail.com>
 <CA+HiwqEXyLiCdi7MQgmU8Y+aPVcOf8JpWfVA4szCmJtZAqj7Fg@mail.gmail.com>
To: Amit Langote <amitlangote09@gmail.com>
Archived-At: 
 <https://www.postgresql.org/message-id/6B155E97-EDE7-4CBE-8196-E709F6A4B33D%40gmail.com>
Precedence: bulk


> On Apr 15, 2026, at 21:46, Amit Langote <amitlangote09@gmail.com> =
wrote:
>=20
> On Wed, Apr 15, 2026 at 11:14=E2=80=AFAM Amit Langote =
<amitlangote09@gmail.com> wrote:
>> On Wed, Apr 15, 2026 at 1:08=E2=80=AFAM Kirill Reshke =
<reshkekirill@gmail.com> wrote:
>>>> +static bool
>>>> +LockCachedPlan(CachedPlan *cplan)
>>>> +{
>>>> + AcquireExecutorLocks(cplan->stmt_list, true);
>>>> + if (!cplan->is_valid)
>>>> + {
>>>> + AcquireExecutorLocks(cplan->stmt_list, false);
>>>> + return false;
>>>> + }
>>>> + return true;
>>>> +}
>>>=20
>>> simply `return cplan->is_valid ` would be more Postgres-y here, isnt =
it?
>>=20
>> Agreed, will fix too.
>=20
> Thinking more about this bit, checking only cplan->is_valid after
> locking is not really equivalent to what CheckCachedPlan() was doing
> before.
>=20
> CheckCachedPlan() can also reset plan->is_valid based on other state
> (role, xmin), whereas in v1 the new check relied only on lock inval
> callbacks to have set is_valid to false. That means the new check is
> not enough. It may happen to be okay for the current callers in this
> patch, but it is not something I think is safe to rely on once there
> are future callers that do arbitrary work, such as ExecutorPrep(),
> even if carefully coded, between the original GetCachedPlan() /
> CheckCachedPlan() step and the later recheck after locking.
>=20
> Separately, I also realized that v1 was introducing redundant lock
> acquisition for freshly built plans, which is not really a no behavior
> change refactoring either.
>=20
> So in v2 I ended up reworking two parts more substantially:
>=20
> * The patch now factors the reused-generic-plan validity logic into
> RecheckCachedPlan().
>=20
> * It also adds an is_reused output argument to GetCachedPlan(), so
> callers can distinguish the reused-plan case from a freshly built plan
> and avoid the extra lock step in the latter case.
>=20
> --=20
> Thanks, Amit Langote
> <v2-0001-Move-execution-lock-acquisition-out-of-GetCachedP.patch>

V2 overall looks good, it preserves the current behavior.

I just have a question. RecheckCachedPlan is only called by =
LockCachedPlan and CheckCachedPlan, and both callers are static, why =
RecheckCachedPlan is exported?

Best regards,
--
Chao Li (Evan)
HighGo Software Co., Ltd.
https://www.highgo.com/