plpython tests fail against python 3.14 under Valgrind

Alexander Lakhin <exclusion@gmail.com> writes:

As buildfarm animal skink shows, upgrade to Python 3.14 breaks plpython
tests under Valgrind [1]:
 61/400 plpython - postgresql:plpython/regress ERROR             49.88s   exit status 1

Yeah, I can reproduce this here, using Fedora 43 (with
python3-3.14.6-1.fc43.x86_64 and valgrind-3.27.1-1.fc43.x86_64).
As you say, it doesn't reproduce on master; I didn't take the trouble
to bisect.

It's also not reproduced on master, probably just because of a bit lesser
stack usage. I've bisected the anomaly and found what makes master pass
the tests: aeb07c55f.

Actually, I think it's the other way around: master uses more stack.
In v19, tzload() does this:

union local_storage *lsp = malloc(sizeof *lsp);
...
free(lsp);

while in master it does this:

union local_storage *lsp;
union local_storage ls;

lsp = &ls;

That's a sizeable amount of stack getting chewed:

(gdb) p sizeof(union local_storage)
$1 = 72240

However, we surely don't invoke tzload() while running any Python
code. What I think must be happening is that at postmaster start,
or possibly backend start, we invoke tzload() while setting the
timezone GUC, and this causes a bunch of stack pages to get allocated,
more than we have in the v19 code path. Somehow, python and valgrind
interact badly when there's not much pre-existing stack allocation.
It's pretty unclear which one is to blame, but it seems like it can't
be our fault. (Although ... surely the standard python executable
invokes libpython with little pre-existing stack? If that doesn't
trigger this problem, what are we doing differently?)

Anyway, I've confirmed that the attached patch makes the problem
go away here. Unless somebody wants to expend brain cells on
running the underlying issue to ground, I think we should just
apply this to the back branches and be happy. A variant plan
could be to do it like this #ifdef USE_VALGRIND and otherwise
keep the old code, but that seems unduly paranoid to me.

regards, tom lane

I wrote:

Anyway, I've confirmed that the attached patch makes the problem
go away here. Unless somebody wants to expend brain cells on
running the underlying issue to ground, I think we should just
apply this to the back branches and be happy. A variant plan
could be to do it like this #ifdef USE_VALGRIND and otherwise
keep the old code, but that seems unduly paranoid to me.

A bit of excavation in our git history reminded me that our old
behavior here was very intentional:

Author: Tom Lane <tgl@sss.pgh.pa.us>
Branch: master Release: REL9_6_BR [62c8421e8] 2016-07-07 11:28:17 -0400

Reduce stack space consumption in tzload().

While syncing our timezone code with IANA's updates in commit 1c1a7cbd6,
I'd chosen not to adopt the code they conditionally compile under #ifdef
ALL_STATE. The main thing that that drives is that the space for gmtime
and localtime timezone definitions isn't statically allocated, but is
malloc'd on first use. I reasoned we didn't need that logic: we don't have
localtime() at all, and we always initialize TimeZone to GMT so we always
need that one. But there is one other thing ALL_STATE does, which is to
make tzload() malloc its transient workspace instead of just declaring it
as a local variable. It turns out that that local variable occupies 78K.
Even worse is that, at least for common US timezone settings, there's a
recursive call to parse the "posixrules" zone name, making peak stack
consumption to select a time zone upwards of 150K. That's an uncomfortably
large fraction of our STACK_DEPTH_SLOP safety margin, and could result in
outright crashes if we try to reduce STACK_DEPTH_SLOP as has been discussed
recently. Furthermore, this means that the postmaster's peak stack
consumption is several times that of a backend running typical queries
(since, except on Windows, backends inherit the timezone GUC values and
don't ever run this code themselves unless you do SET TIMEZONE). That's
completely backwards from a safety perspective.

Hence, adopt the ALL_STATE rather than non-ALL_STATE variant of tzload(),
while not changing the other code aspects that symbol controls. The
risk of an ENOMEM error from malloc() seems less than that of a SIGSEGV
from stack overrun.

Now, we deleted the "posixrules" recursion logic in v13, so in all
active branches the amount of stack space required is 70-something K
not twice that. But re-reading this message, the postmaster
vs. backends point still seems like a compelling safety argument.
So I'd rather make master work like the old code instead of vice
versa.

Hence, what I'm now thinking is to malloc the space by default,
but use a stack allocation if USE_VALGRIND, with a comment explaining
that that masks an obscure valgrind-or-possibly-python bug; and to
do it the same way in all branches.

regards, tom lane