BUG #19476: Segmentation fault in contrib/spi

Hi,

On Tue, 12 May 2026 at 21:14, PG Bug reporting form <noreply@postgresql.org>
wrote:

CREATE EXTENSION refint;
CREATE TABLE c (id int4);
CREATE UNIQUE INDEX ci ON c(id);
CREATE TABLE b (refb int4);
CREATE INDEX bi ON b(refb);

CREATE TRIGGER at AFTER DELETE OR UPDATE ON c FOR EACH ROW
EXECUTE FUNCTION check_foreign_key (1, 'cascade', 'id', 'b', 'refb');

CREATE TRIGGER bt AFTER INSERT OR UPDATE ON b FOR EACH ROW
EXECUTE FUNCTION check_primary_key ('refb', 'c', 'id');

INSERT INTO c VALUES (10);
INSERT INTO b VALUES (10);
UPDATE c SET id = NULL WHERE id = 10;

Backtrace:
#0 __strlen_avx2 () at ../sysdeps/x86_64/multiarch/strlen-avx2.S:76
#1 0x0000563a92ca096e in quote_literal_cstr (rawstr=0x0) at quote.c:109
#2 0x00007fa4a7cce5f8 in check_foreign_key (fcinfo=<optimized out>) at
refint.c:489
#3 0x0000563a929c1d52 in ExecCallTriggerFunc
(trigdata=trigdata@entry=0x7fff8cae1100,
tgindx=tgindx@entry=0, finfo=finfo@entry=0x563aa5ae4438,
instr=instr@entry=0x0,
per_tuple_context=per_tuple_context@entry=0x563aa5adb9c0) at
trigger.c:2369
#4 0x0000563a929c4110 in AfterTriggerExecute (estate=<optimized out>,
event=0x563aa5aedbb0,
relInfo=0x563aa5ae4068, src_relInfo=<optimized out>,
dst_relInfo=<optimized out>,
trigdesc=0x563aa5ae4278, finfo=0x563aa5ae4438, instr=0x0,
per_tuple_context=<optimized out>, trig_tuple_slot1=0x0,
trig_tuple_slot2=0x0)
at trigger.c:4559
#5 afterTriggerInvokeEvents (events=events@entry=0x563aa5a75050,
firing_id=1,
estate=estate@entry=0x563aa5ae3b20, delete_ok=delete_ok@entry=false)
at
trigger.c:4800
#6 0x0000563a929c8e88 in AfterTriggerEndQuery
(estate=estate@entry=0x563aa5ae3b20)
at trigger.c:5182
#7 0x0000563a929ed504 in standard_ExecutorFinish
(queryDesc=0x563aa5996e20)
at execMain.c:443
#8 0x0000563a92bd96f8 in ProcessQuery (plan=0x563aa5af4168,
sourceText=0x563aa59df7e0 "UPDATE c SET id = NULL WHERE id = 10;",
params=0x0,
queryEnv=0x0, dest=0x563aa5aeae28, qc=0x7fff8cae1460) at pquery.c:194
#9 0x0000563a92bda41e in PortalRunMulti
(portal=portal@entry=0x563aa5a60510,
isTopLevel=isTopLevel@entry=true,
setHoldSnapshot=setHoldSnapshot@entry=false,
--Type <RET> for more, q to quit, c to continue without paging--c
dest=dest@entry=0x563aa5aeae28, altdest=altdest@entry=0x563aa5aeae28,
qc=qc@entry=0x7fff8cae1460) at pquery.c:1272
#10 0x0000563a92bda7f7 in PortalRun (portal=portal@entry=0x563aa5a60510,
count=count@entry=9223372036854775807, isTopLevel=isTopLevel@entry
=true,
dest=dest@entry=0x563aa5aeae28, altdest=altdest@entry=0x563aa5aeae28,
qc=qc@entry=0x7fff8cae1460) at pquery.c:788
#11 0x0000563a92bd63fa in exec_simple_query (
query_string=0x563aa59df7e0 "UPDATE c SET id = NULL WHERE id = 10;") at
postgres.c:1274
#12 0x0000563a92bd7fc5 in PostgresMain (dbname=<optimized out>,
username=<optimized out>)
at postgres.c:4770
#13 0x0000563a92bd23bd in BackendMain (startup_data=<optimized out>,
startup_data_len=<optimized out>) at backend_startup.c:124
#14 0x0000563a92b24932 in postmaster_child_launch (child_type=<optimized
out>, child_slot=1,
startup_data=startup_data@entry=0x7fff8cae1900,
startup_data_len=startup_data_len@entry=24,
client_sock=client_sock@entry=0x7fff8cae1920)
at launch_backend.c:290
#15 0x0000563a92b283d2 in BackendStartup (client_sock=0x7fff8cae1920) at
postmaster.c:3569
#16 ServerLoop () at postmaster.c:1703
#17 0x0000563a92b29ea6 in PostmasterMain (argc=argc@entry=3,
argv=argv@entry=0x563aa5985bf0)
at postmaster.c:1401
#18 0x0000563a92800a5a in main (argc=3, argv=0x563aa5985bf0) at main.c:227

Thanks for the report and the clear backtrace.

I had a look at this and I think the immediate crash comes from
check_foreign_key()'s cascade-update branch in contrib/spi/refint.c,
which calls SPI_getvalue() for the new key value and then passes the
result straight into snprintf("%s", ...):

nv = SPI_getvalue(newtuple, tupdesc, fn);
...
snprintf(sql + strlen(sql), sizeof(sql) - strlen(sql),
" %s = %s%s%s %s ",
args2[k], (is_char_type > 0) ? "'" : "",
nv, (is_char_type > 0) ? "'" : "", (k < nkeys) ? ", " : "");

For an UPDATE that sets the referenced key to NULL, SPI_getvalue()
returns NULL, so nv is NULL. On stock glibc/Postgres builds snprintf
substitutes "(null)" and SQL parses that as the NULL keyword, so the
cascade silently does the right thing. On builds with _FORTIFY_SOURCE
glibc's strlen check dereferences the NULL pointer instead, which
matches the backtrace you posted.

I am attaching a two-patch series that addresses this.

0001 is the minimal fix: emit the SQL NULL keyword when SPI_getvalue()
returns NULL, instead of passing a NULL pointer to snprintf. This is
small and targeted, and is sufficient on its own to fix the reported
crash. I did not add a regression test for 0001 because the bug only
manifests on hardened libc builds; on stock builds the bad snprintf
happens to produce a result that the SQL parser accepts.

0002 is a follow-up that I think makes sense but is not strictly
required for this bug. While looking at the cascade-update path I
noticed two pre-existing issues in the same code:

- The new key values were embedded into the prepared cascade UPDATE
query text as literals. Combined with plan caching (the cache key
only gained the operation type in 8cfbdf8f4df), the first set of
new values gets reused for every subsequent cascade UPDATE.

- Char-like new values were wrapped in single quotes without escaping,
so a value containing a single quote produced malformed SQL.

0002 switches that branch to use SPI parameters for the new key values
instead of embedding them into the SQL text. That naturally handles
NULLs via the SPI nulls array, lets a single cached plan handle later
UPDATEs with different keys, and avoids the quoting issue entirely.
It also adds a refint regression test that runs two cascade UPDATEs in
a row (one to NULL, one to a non-NULL value) which catches the cached-
plan reuse.

I'm happy with just 0001 if reviewers prefer to keep the fix minimal,
given that refint is intended mainly as a documentation example
(8cfbdf8f4df explicitly mentioned not back-patching changes here?).

Regards,
Ayush

Hello,

You have not used the very last version of refint.c which has been updated just yesterday:

commit 1ebda7da9a43d3ae3564d08612de9cb27fbaf482
Author: Nathan Bossart <nathan@postgresql.org>
Date: Mon May 11 05:13:48 2026 -0700

refint: Fix SQL injection and buffer overruns.

Maliciously crafted key value updates could achieve SQL injection
within check_foreign_key(). To fix, ensure new key values are
properly quoted and escaped in the internally generated SQL
statements. While at it, avoid potential buffer overruns by
replacing the stack buffers for internally generated SQL statements
with StringInfo.

Reported-by: Nikolay Samokhvalov <nik@postgres.ai>
Author: Nathan Bossart <nathandbossart@gmail.com>
Reviewed-by: Noah Misch <noah@leadboat.com>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: Fujii Masao <masao.fujii@gmail.com>
Security: CVE-2026-6637
Backpatch-through: 14

I think for patch v1-0001 you need to adapt the code with something like:

/* SPI_getvalue() returns NULL for SQL NULL values, so
* emit the NULL keyword rather than passing a NULL
* pointer to quote_litteral_cstr, which is undefined
* behavior.
*/
if (nv == NULL)
appendStringInfo(&sql, " %s = NULL %s",
args2[k], "");
else
appendStringInfo(&sql, " %s = %s ",
args2[k], quote_literal_cstr(nv));

Regards
Pierre

Hi,

On Wed, 13 May 2026 at 00:22, <pierre.forstmann@gmail.com> wrote:

Hello,

You have not used the very last version of refint.c which has been updated
just yesterday:

commit 1ebda7da9a43d3ae3564d08612de9cb27fbaf482
Author: Nathan Bossart <nathan@postgresql.org>
Date: Mon May 11 05:13:48 2026 -0700

refint: Fix SQL injection and buffer overruns.

Maliciously crafted key value updates could achieve SQL injection
within check_foreign_key(). To fix, ensure new key values are
properly quoted and escaped in the internally generated SQL
statements. While at it, avoid potential buffer overruns by
replacing the stack buffers for internally generated SQL statements
with StringInfo.

Reported-by: Nikolay Samokhvalov <nik@postgres.ai>
Author: Nathan Bossart <nathandbossart@gmail.com>
Reviewed-by: Noah Misch <noah@leadboat.com>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: Fujii Masao <masao.fujii@gmail.com>
Security: CVE-2026-6637
Backpatch-through: 14

You're right, thanks for catching this. I sent the v1 patches against
master from the day before; commit 260e97733bf (CVE-2026-6637) landed
in between and I had not noticed it. That commit rewrites the same
cascade-update path to use StringInfo and quote_literal_cstr(), so the
v1 patches do not apply on current master at all.

Importantly, after 260e97733bf the bug is also no longer dependent on
_FORTIFY_SOURCE: the new code calls quote_literal_cstr(nv) directly,
which dereferences nv via strlen() and segfaults on stock builds too.
I reproduced this on plain master built with --enable-cassert.

I have rebased the minimal fix on current master. It is essentially
the same shape as the snippet you suggested -- emit the NULL keyword
directly when SPI_getvalue() returns NULL, otherwise pass through
quote_literal_cstr() as today. Attached as v2-0001.

I dropped my earlier 0002 patch. The CVE fix already addressed the
quoting/escaping concerns that motivated half of it.

Regards,
Ayush

On Wed, May 13, 2026 at 12:57:47AM +0530, Ayush Tiwari wrote:

I have rebased the minimal fix on current master. It is essentially
the same shape as the snippet you suggested -- emit the NULL keyword
directly when SPI_getvalue() returns NULL, otherwise pass through
quote_literal_cstr() as today. Attached as v2-0001.

I dropped my earlier 0002 patch. The CVE fix already addressed the
quoting/escaping concerns that motivated half of it.

I'm confused why you dropped 0002. Reusing the new key values for
subsequent updates seems like a bug worth fixing. However, note that the
parameter symbol approach doesn't work well for cases like this:

CREATE EXTENSION refint;

CREATE TABLE p (a int);
CREATE TABLE f (a xid);
CREATE TRIGGER t
AFTER DELETE OR UPDATE ON p
FOR EACH ROW EXECUTE PROCEDURE
check_foreign_key(1, 'c', 'a', 'f', 'a');

INSERT INTO p VALUES (1);
UPDATE p SET a = 2;

With a rebased version of 0002 applied, the UPDATE statement fails like
this:

ERROR: column "a" is of type xid but expression is of type integer
LINE 1: update f set a = $2 where a = $1
^
HINT: You will need to rewrite or cast the expression.
QUERY: update f set a = $2 where a = $1

Presumably the problem is that we're using the wrong argument type for the
foreign key. I'm not sure this is trivial to fix; it seems like we'd need
to provide that information in CREATE TRIGGER or look up the foreign key
type within the trigger function itself. Perhaps the best we can do is to
avoid caching a plan in this case.

Regarding 0001, note that the refint docs state the following:

Note that the primary/unique key columns should be marked NOT NULL and
should have a unique index.

So maybe we could alternatively teach check_foreign_key() to either ERROR
or do nothing instead. On the other hand, given this case seemed to
accidentally work before the CVE fix, it's arguably worth fixing.

--
nathan

On Thu, May 14, 2026 at 11:01:55AM -0500, Nathan Bossart wrote:

Regarding 0001, note that the refint docs state the following:

Note that the primary/unique key columns should be marked NOT NULL and
should have a unique index.

So maybe we could alternatively teach check_foreign_key() to either ERROR
or do nothing instead. On the other hand, given this case seemed to
accidentally work before the CVE fix, it's arguably worth fixing.

Here is what I have staged for commit, which I intend to do shortly.

--
nathan

Hi,

On Thu, 14 May 2026 at 22:15, Nathan Bossart <nathandbossart@gmail.com>
wrote:

On Thu, May 14, 2026 at 11:01:55AM -0500, Nathan Bossart wrote:

Regarding 0001, note that the refint docs state the following:

Note that the primary/unique key columns should be marked NOT NULL

and

should have a unique index.

So maybe we could alternatively teach check_foreign_key() to either ERROR
or do nothing instead. On the other hand, given this case seemed to
accidentally work before the CVE fix, it's arguably worth fixing.

Here is what I have staged for commit, which I intend to do shortly.

Thanks, this version looks good to me. The compact form is fine, and I
agree with preserving the pre-CVE behavior for this misconfigured case
rather than turning it into an ERROR or no-op.

I also agree with your earlier point about the parameterized 0002: using
the triggered relation's key types for parameters is not generally right
for the referencing relation's SET targets.

I think the cached cascade UPDATE plan issue is worth pursuing,
I'll start a different hackers thread on that
probably using the "don't cache cascade UPDATE plans" approach you
suggested.

Regards,
Ayush

On Thu, May 14, 2026 at 10:42:07PM +0530, Ayush Tiwari wrote:

Thanks, this version looks good to me. The compact form is fine, and I
agree with preserving the pre-CVE behavior for this misconfigured case
rather than turning it into an ERROR or no-op.

Committed.

I think the cached cascade UPDATE plan issue is worth pursuing,
I'll start a different hackers thread on that
probably using the "don't cache cascade UPDATE plans" approach you
suggested.

Great, thanks.

--
nathan