On Mon, Oct 05, 2026 at 11:26:06PM +0000, maxml wrote:
> Subject: [PATCH] BUG/MINOR: resolvers: validate TXT RDATA before reading
> Hello,
> 
> Thank you for maintaining HAProxy.
> 
> While fuzzing HAProxy's DNS response validator, I found a one-byte read past
> the logical end of a DNS response when a TXT answer has RDLENGTH == 0.
> 
> The TXT answer parser loads the first character-string length byte before
> checking that the record data contains at least one byte. A TXT answer with
> RDLENGTH zero at the end of a DNS response leaves reader equal to bufend, so
> reader[0] crosses the parser's logical message boundary.
> 
> A minimized 31-byte response reproduces this as a one-byte
> heap-buffer-overflow in resolv_validate_dns_response() with Clang libFuzzer
> and AddressSanitizer. All 13 original AFL++ crash inputs reach the same source
> instruction.
> 
> The production caller currently uses a DNS_MAX_MSG_SIZE + 1 stack array while
> parsing no more than DNS_MAX_MSG_SIZE bytes, and the prematurely loaded value
> is discarded once the zero-length check fails. Thus I have not demonstrated a
> production crash, information disclosure, out-of-bounds write, or control-flow
> effect. The fix nevertheless restores the function's bufend contract and
> keeps the parser safe for exactly sized callers.
> 
> Reject zero-length TXT RDATA before loading its first byte. This preserves a
> valid empty TXT string, whose wire representation has RDLENGTH one followed by
> a zero character-string length byte.
> 
> The affected TXT parser was introduced by commit e418e828aa ("MINOR:
> resolvers: basic TXT record implementation") and is present in HAProxy 3.4.0.
> This fix must be backported to 3.4; older stable branches are not affected.
> 
> Testing performed:
> 
> - reproduced the exact resolvers.c:1375:31 ASan read on revision
> 5129223edcd4a011c3219434cda7501ad2f52405;
> - replayed and deduplicated all 13 original AFL++ findings;
> - applied this patch and verified that the minimized and original inputs no
> longer produce a sanitizer finding; and
> - checked that the patch applies cleanly to upstream master
> e4d1b143ba49b2364172395410f8b3da17caf9ae (2026-10-05).
> 
> I have attached poc.tar.gz. It is a source-only, machine-independent bundle
> containing the minimized and original inputs, root-cause analysis, recorded
> ASan report, proposed patch, integrity manifests, and scripts that reproduce
> the vulnerable result and validate the fix with ./run.sh.
> 
> The finding was discovered with AFL++ and analyzed and packaged with automated
> assistance. In accordance with HAProxy's security directions, the attachment
> includes a working reproducer and a tested patch.
> 
> Please let me know if you need any additional information or would like me to 
> test a
> proposed fix.
> 
> Regards,
> maxml

The patch seems to be good, the problem has no impact since the value won't be
used if the len is 0.

I pushed the patch mentioning you as the author. For your  next findings,
provide a patch in git format so it can be applied directly.

Thanks!


-- 
William Lallemand


Reply via email to