Re: libpq: SCRAM client proofs serialize across threads on OpenSSL 3 (re-init of the HMAC per PBKDF2 iteration)

From: Matthew Mochalkin <matthew(dot)mochalkin(at)gmail(dot)com>
To: Jakub Wartak <jakub(dot)wartak(at)enterprisedb(dot)com>
Cc: pgsql-hackers(at)lists(dot)postgresql(dot)org
Subject: Re: libpq: SCRAM client proofs serialize across threads on OpenSSL 3 (re-init of the HMAC per PBKDF2 iteration)
Date: 2026-11-17 16:37:24
Message-ID: E839523C-1B40-4EFD-8839-8DDE0A845865@gmail.com
Views: Whole Thread | Raw Message | Download mbox | Resend email
Thread:
Lists: pgsql-hackers

Yes, same thing. The contention your perf report showed inside HMAC_Init_ex() is what the patch removes.

I ran pgbench --connect through pgbouncer as you suggested. Same pgbench binary (master), base and patched libpq swapped with LD_LIBRARY_PATH. pgbouncer 1.24.1, auth_type = scram-sha-256, the role's SCRAM verifier in auth_file, PostgreSQL 16 behind it, OpenSSL 3.5.7, 8 CPUs for the client.

pgbench --connect -n -c N -j N -T 20 -f select1.sql, connections per second:

-c = -j base patched
1 403 451
8 1,170 5,380
16 734 8,555

With the base libpq pgbench used 6-8 cores for those ~1,000 connections/s, the "1000% CPU" you saw. With the patch it uses about the same CPU for 5-8x the connections.

If you’d like to repeat it. Put the SCRAM verifier in pgbouncer's auth_file, not the plain password: with the plain password pgbouncer does the PBKDF2 itself and caps the test at ~380 logins/s no matter what libpq does. And at 16 threads the patched client runs out of ephemeral ports in 20 s unless tcp_tw_reuse is on.

Regards,
Matt Mochalkin

> On 17 Nov 2026, at 5:24 PM, Jakub Wartak <jakub(dot)wartak(at)enterprisedb(dot)com> wrote:
>
> On Mon, Nov 16, 2026 at 11:08 PM Matthew Mochalkin
> <matthew(dot)mochalkin(at)gmail(dot)com <mailto:matthew(dot)mochalkin(at)gmail(dot)com>> wrote:
>>
>> Hi,
>>
>> A multithreaded libpq client on OpenSSL 3 cannot open SCRAM-SHA-256 connections in parallel. The threads of one process get about the throughput of a single thread, and adding threads makes it worse. The same client as separate processes scales as expected. I ran into it with FrankenPHP (PHP's ZTS build served from the threads of one process, libpq 17.12, OpenSSL 3.5.7, Debian 13): 8 threads doing one connect per request reach ~600 connects/s against ~1,700 for 8 processes, and ~2,900 for the same 8 threads once the role uses md5 or trust.
>>
>> The cost is in scram_SaltedPassword(). It calls pg_hmac_init() for every one of the 4,096 iterations, and on OpenSSL 3 each HMAC_Init_ex(ctx, key, len, EVP_sha256(), NULL) does an implicit fetch of the digest, which takes process-wide locks inside libcrypto (openssl/openssl#19612 describes the same cost for EVP_DigestInit_ex). With one thread that is a per-iteration overhead; with several threads it is a lock convoy. The EVP_MAC conversion in 1f3b9bb109 keeps the per-iteration EVP_MAC_init() with the digest parameter and has the same behaviour.
>>
>> A 60-line C program with no libpq in it shows the effect. It runs the 4,096-iteration HMAC-SHA-256 loop in N threads or N forked processes, 8 CPUs, loops per second:
>>
>> pattern per iteration 1 thr 8 thr 16 thr 8 proc
>> HMAC_Init_ex(ctx, key, len, EVP_sha256(), NULL) 519 547 252 3,397 (libpq <= 17)
>> EVP_MAC_init(ctx, key, len, {digest=SHA256}) 600 1,191 547 4,582 (master)
>> HMAC_Init_ex(ctx, key, len, md_fetched_once, NULL) 1,150 5,684 9,832 6,366
>> HMAC_Init_ex(ctx, NULL, 0, NULL, NULL) 1,997 9,272 12,095 18,557 (reuse key and digest)
>> EVP_MAC_init(ctx, NULL, 0, NULL) 1,085 8,045 13,785 13,613
>>
>> And with libpq itself, PQconnectdb()+PQfinish() in a loop against PostgreSQL 16, scram role:
>>
>> role 1 thr 8 thr 16 thr 1 proc 8 proc 16 proc
>> scram-sha-256 237 592 268 222 1,641 2,386
>> trust 595 1,882 3,814 559 2,640 3,594
>>
>> This was discussed in February 2025 ("Optimize scram_SaltedPassword performance", https://www.postgresql.org/message-id/Z6B7QXlPwPXewaQZ%40paquier.xyz) That proposal was about single-thread latency and was dropped, partly because a faster PBKDF2 does not help the server much and partly because of the static state in the patch. The multithreaded case is different: it is not that the client is 4 times slower than it could be, it is that N threads are no faster than one, which for a threaded application server means the database connect rate of the whole process is capped around a few hundred per second. The fix does not need any static state: inside scram_SaltedPassword() the key is the same for every iteration, so the OpenSSL implementation can keep it in the context after the first pg_hmac_init() and re-init with a NULL key (HMAC_Init_ex(ctx, NULL, 0, NULL, NULL), or EVP_MAC_init(ctx, NULL, 0, NULL) on the EVP_MAC path). The fallback src/common/hmac.c can keep its current behaviour, or skip re-deriving the pads when the key pointer and length are unchanged.
>>
>> I can send a patch for the OpenSSL path if there is interest. The reproduction (Docker, PHP and C) is at https://github.com/mattleads/frankenphp-pgconnect-repro.
>
> To me it looks like another instance of [1], isn't it? If that would be true
> then pgbench --connect along with pgbouncer (quite to setup) could be enough
> to test/verify your patch idea too.
>
> [1] - https://www.postgresql.org/message-id/CAKZiRmw6jbShG%2BLJe0pY5G4w0ktoGwzPozsA%3DLhH_mN9wC%2Bbhg%40mail.gmail.com

In response to

Browse pgsql-hackers by date

  From Date Subject
Next Message Nathan Bossart 2026-11-17 16:42:48 create .abi-compliance-history file for v19
Previous Message Ashutosh Sharma 2026-11-17 16:24:52 Re: log XLogPrefetch stats at end of recovery