Message-ID: <e1ffb822-054e-4006-ac06-50532767f75b@iki.fi>
Date: Mon, 24 Jun 2024 13:29:38 +0300
MIME-Version: 1.0
User-Agent: Mozilla Thunderbird
Content-Language: en-US
From: Heikki Linnakangas <hlinnaka@iki.fi>
Subject: Injection point locking
To: pgsql-hackers <pgsql-hackers@postgresql.org>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit
Archived-At: <https://www.postgresql.org/message-id/e1ffb822-054e-4006-ac06-50532767f75b%40iki.fi>
Precedence: bulk

InjectionPointRun() acquires InjectionPointLock, looks up the hash 
entry, and releases the lock:

> 	LWLockAcquire(InjectionPointLock, LW_SHARED);
> 	entry_by_name = (InjectionPointEntry *)
> 		hash_search(InjectionPointHash, name,
> 					HASH_FIND, &found);
> 	LWLockRelease(InjectionPointLock);

Later, it reads fields from the entry it looked up:

> 		/* not found in local cache, so load and register */
> 		snprintf(path, MAXPGPATH, "%s/%s%s", pkglib_path,
> 				 entry_by_name->library, DLSUFFIX);

Isn't that a straightforward race condition, if the injection point is 
detached in between?


Another thing:

I wanted use injection points to inject an error early at backend 
startup, to write a test case for the scenarios that Jacob point out at 
https://www.postgresql.org/message-id/CAOYmi%2Bnwvu21mJ4DYKUa98HdfM_KZJi7B1MhyXtnsyOO-PB6Ww%40mail.gmail.com. 
But I can't do that, because InjectionPointRun() requires a PGPROC 
entry, because it uses an LWLock. That also makes it impossible to use 
injection points in the postmaster. Any chance we could allow injection 
points to be triggered without a PGPROC entry? Could we use a simple 
spinlock instead? With a fast path for the case that no injection points 
are attached or something?

-- 
Heikki Linnakangas
Neon (https://neon.tech)