Newsgroups: php.internals
Path: news.php.net
Xref: news.php.net php.internals:128695
Feedback-ID: ifab94697:Fastmail
Precedence: list
MIME-Version: 1.0
Date: Sat, 13 Sep 2025 00:37:08 +0200
To: "Hans Krentel" <hanskrentel@yahoo.de>, internals@lists.php.net
Message-ID: <2a34b4a4-a1f3-40d1-a005-49a91a31b07c@app.fastmail.com>
In-Reply-To: <1757713032326.1442271469.4286265584@yahoo.de>
References: <f20b92f6-c2dc-4eae-bba3-faea45a6fe4a@app.fastmail.com>
 <1757713032326.1442271469.4286265584@yahoo.de>
Subject: Re: [PHP-DEV] merging r/w locks
Content-Type: multipart/alternative;
 boundary=d344e2f8712840f082949697bf2016e3
From: rob@bottled.codes ("Rob Landers")

--d344e2f8712840f082949697bf2016e3
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: quoted-printable



On Fri, Sep 12, 2025, at 23:39, Hans Krentel wrote:
>=20
>=20
>=20
> On Tuesday 09 September 2025 10:14:31 (+02:00), Rob Landers wrote:
>=20
> > Hi internals,
> >=20
> > Is there any hope to have https://github.com/php/php-src/pull/16565 =
merged before 8.5? It's too late to address all the locks, but having it=
 as part of the ABI would be nice so we can start fixing them for 8.6/9.=
0. In FrankenPHP, if something takes a lock, it can significantly affect=
 performance. By using r/w locks instead of exclusive locks, we can use =
shared locks when reading instead of exclusive locks.
> >=20
> > =E2=80=94 Rob
>=20
> My educated guess would be that environ is per process.
>=20
> -- hakre
>=20

Hi Hakre,

We already solved the environment locks in FrankenPHP (which uses thread=
s instead of processes in ZTS mode) by basically =E2=80=9Cdeleting=E2=80=
=9D the built-in environment functions and replacing them. Environment a=
ccess is not thread safe, and Go (the language of FrankenPHP) uses its o=
wn locks around environment access. Thus we needed to prevent concurrent=
 access to the environment in the two runtimes (TSRM vs. Go). That part =
of the PR will just help with other SAPI performance when using PHP comp=
iled in ZTS mode.

The main feature of the PR is not the environment locks, but the adding =
of non-exclusive locks to TSRM. For example, there are TSRM locks in OPc=
ache, but they=E2=80=99re only used *during writes* and *don=E2=80=99t p=
rotect reads*. This can lead to very subtle bugs and crashes that we see=
 sporadically in FrankenPHP. By enabling r/w locks (the new tsrm_rwlock_=
rlock in this PR), we can protect reads *and* writes, without degrading =
existing performance.

TSRM itself could use this as well, so that every time FrankenPHP spawns=
 a new PHP thread, it doesn=E2=80=99t need to contend for reading/writin=
g globals during startup (and for dealing with cases where TSRM cache ca=
n=E2=80=99t be used, which affects ALL global reads/writes if I understa=
nd correctly =E2=80=94 which is possible I don=E2=80=99t).

I believe there are several other instances that would benefit from r/w =
locks as well, such as mysqlnd, but it has been awhile since I looked at=
 this, and I=E2=80=99m just going off notes.

=E2=80=94 Rob
--d344e2f8712840f082949697bf2016e3
Content-Type: text/html; charset=utf-8
Content-Transfer-Encoding: quoted-printable

<!DOCTYPE html><html><head><title></title></head><body><div><br></div><d=
iv><br></div><div>On Fri, Sep 12, 2025, at 23:39, Hans Krentel wrote:</d=
iv><blockquote type=3D"cite" id=3D"qt" style=3D""><div><br></div><div><b=
r></div><div><br></div><div>On Tuesday 09 September 2025 10:14:31 (+02:0=
0), Rob Landers wrote:</div><div><br></div><div>&gt; Hi internals,</div>=
<div>&gt;&nbsp;</div><div>&gt; Is there any hope to have&nbsp;<a href=3D=
"https://github.com/php/php-src/pull/16565">https://github.com/php/php-s=
rc/pull/16565</a> merged before 8.5? It's too late to address all the lo=
cks, but having it as part of the ABI would be nice so we can start fixi=
ng them for 8.6/9.0. In FrankenPHP, if something takes a lock, it can si=
gnificantly affect performance. By using r/w locks instead of exclusive =
locks, we can use shared locks when reading instead of exclusive locks.<=
/div><div>&gt;&nbsp;</div><div>&gt; =E2=80=94 Rob</div><div><br></div><d=
iv>My educated guess would be that environ is per process.</div><div><br=
></div><div>-- hakre</div><div><br></div></blockquote><div><br></div><di=
v>Hi Hakre,</div><div><br></div><div>We already solved the environment l=
ocks in FrankenPHP (which uses threads instead of processes in ZTS mode)=
 by basically =E2=80=9Cdeleting=E2=80=9D the built-in environment functi=
ons and replacing them. Environment access is not thread safe, and Go (t=
he language of FrankenPHP) uses its own locks around environment access.=
 Thus we needed to prevent concurrent access to the environment in the t=
wo runtimes (TSRM vs. Go). That part of the PR will just help with other=
 SAPI performance when using PHP compiled in ZTS mode.</div><div><br></d=
iv><div>The main feature of the PR is not the environment locks, but the=
 adding of non-exclusive locks to TSRM. For example, there are TSRM lock=
s in OPcache, but they=E2=80=99re only used&nbsp;<i>during writes</i> an=
d <i>don=E2=80=99t protect reads</i>. This can lead to very subtle bugs =
and crashes that we see sporadically in FrankenPHP. By enabling r/w lock=
s (the new tsrm_rwlock_rlock in this PR), we can protect reads&nbsp;<b>a=
nd</b> writes, without degrading existing performance.</div><div><br></d=
iv><div>TSRM itself could use this as well, so that every time FrankenPH=
P spawns a new PHP thread, it doesn=E2=80=99t need to contend for readin=
g/writing globals during startup (and for dealing with cases where TSRM =
cache can=E2=80=99t be used, which affects ALL global reads/writes if I =
understand correctly =E2=80=94 which is possible I don=E2=80=99t).</div>=
<div><br></div><div>I believe there are several other instances that wou=
ld benefit from r/w locks as well, such as mysqlnd, but it has been awhi=
le since I looked at this, and I=E2=80=99m just going off notes.</div><d=
iv><br></div><div id=3D"sig121229152">=E2=80=94 Rob</div></body></html>
--d344e2f8712840f082949697bf2016e3--