| Commit message (Collapse) | Author | Age | Files | Lines |
... | |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Currently, a read/write memory BIO pulls up the data via memmove() on each
read. This becomes very expensive when a lot of small reads are performed,
especially if there is a reasonable amount of data stored in the memory
BIO.
Instead, store a read offset into the buffer and only perform a memmove()
to pull up the data on a write, if we have read more than 4096 bytes. This
way we only perform memmove() when the space saving will potentially be of
benefit, while avoiding frequent memmove() in the case of small interleaved
reads and writes.
Should address oss-fuzz #19881.
ok inoguchi@ tb@
|
| |
|
|
|
|
|
|
|
|
|
| |
In order to fix and improve the memory BIO, we need to be able to track
more than just a single BUF_MEM *. Provide a struct bio_mem (which
currently only contains a BUF_MEM *) and rework the internals to use this
struct.
ok inoguchi@ tb@
|
|
|
|
|
|
|
|
| |
person. Rewrite or use singular they.
ok thfr@ sthen@ daniel@ ian@ job@ kmos@ jcs@ ratchov@ phessler@ and
others I'm likely missing on an earlier version.
feedback tj@, feedback and ok jmc@
|
|
|
|
|
|
|
|
| |
This is a first pass that uses sensible and consistent names for variables.
Call the BIO 'bio' (instead of 'a', 'b', 'bp', or 'h'), drop a bunch of
unnecessary casts, simplify some logic and add additional error checking.
With input from and ok tb@
|
|
|
|
| |
ok gnezdo@ miod@ jmc@
|
|
|
|
| |
ok jmc@
|
| |
|
| |
|
|
|
|
|
|
|
|
|
|
| |
A fix for this was previously commited in r1.32, however while this added
a bounds check the logic means we still fall through and perform the
overread. Fix the logic such that we only log the error if the bounds check
fails. While here, flip the test around such that we check for validity then
print (which is more readable and matches earlier code).
ok inoguchi@ tb@
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The current implementation uses an unsigned long, then switches to BN once
the arc exceeds its size. However, the complexity of BN_bn2dec() is
quadratic in the length of number being converted. This means that OIDs
with excessively large arcs take a lot of computation to convert to text.
While the X.660 specification states that arcs are unbounded, in reality
they are not overly large numbers - 640K^W64 bits ought to be enough for
any arc. Remove BN entirely, switch from unsigned long to uin64_t and fail
if an arc exceeds this size.
Identified via oss-fuzz timeouts - should fix #41028 and #44372.
ok tb@
|
|
|
|
|
|
|
|
|
| |
An IP address in a name constraint is actually an IP address concatenated
with a netmask, so it is twice as long as usual.
This fixes a third bug introduced in r1.3 and reported by Volker Schlecht
ok jsing
|
|
|
|
|
|
| |
Fixes a segfault reported by Volker Schlecht.
ok jsing
|
|
|
|
|
|
|
|
|
|
|
| |
a2i_GENERAL_NAME() modifies and returns the out argument that was
passed in unless out == NULL, in which case it returns something
freshly allocated. Thus, in v2i_GENERAL_NAME_ex() we must only free
ret if out == NULL so v2i_NAME_CONSTRAINTS() can free correctly.
Issue reported by Volker Schlecht
ok jsing
|
| |
|
|
|
|
|
|
|
|
|
|
| |
Currently OBJ_obj2nid() with NID_undef returns NID_ccitt - this is due to
doing a lookup on an empty value and having NID_undef conflict with an
uninitialised NID value.
Somewhat based on OpenSSL 0fb99904809.
ok tb@
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
no_sanitize_address attribute. ASAN doesn't seem to be able
to understand these lowlevel gymnastics with sigaltstack()
and segfaults in __intercept_memem().
This allows LibreSSL and other portable projects that use this
test run tests with ASAN enabled.
Issue reported and workaround suggested by Ilya Shipitsin
Paraphrasing millert: it's a little ugly but it's only a regress.
|
|
|
|
| |
int foo() to int foo(void)
|
|
|
|
|
|
|
|
|
| |
X509_STORE_add_crl() does not take ownership of the CRL, it bumps its
refcount. So nulling out the CRL from the stack will leak it.
Issue reported by KS Sreeram, thanks!
ok jsing
|
| |
|
| |
|
|
|
|
|
|
| |
Make sure kp is freed also on error.
ok jsing
|
| |
|
|
|
|
|
|
|
|
| |
a 'const uint8_t *a' to a 'const uint8_t a[32]' to match
the prototype in curve25519_internal.h and the other variant
inside OPENSSL_SMALL.
ok millert
|
|
|
|
|
|
|
|
|
|
| |
This is a very rarely used function and the crash is hard to reach in
practice. Instead of implementing BN_is_odd() badly by hand, just call
the real thing.
Reported by Guido Vranken
ok beck jsing
|
|
|
|
|
|
| |
From OpenSSL 6a009812, prompted by a report by Guido Vranken
ok beck jsing
|
|
|
|
|
|
|
|
|
|
|
| |
In order for SSL_get_error() to work with SSL_read_ex() and SSL_write_ex()
the error handling needs to be performed without checking i <= 0. This is
effectively part of OpenSSL 8051ab2b6f8 and should bring the behaviour of
SSL_get_error() largely inline with OpenSSL 1.1.
Issue reported by Johannes Nixdorf.
ok inoguchi@ tb@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
With the legaacy stack, it is possible to do a zero byte SSL_read() or
SSL_write() that triggers the handshake, but then returns zero without
SSL_ERROR_WANT_READ or SSL_ERROR_WANT_WRITE being flagged. This currently
works in the TLSv1.3 stack by returning TLS_IO_WANT_POLLIN or
TLS_IO_WANT_POLLOUT, which is then hidden by SSL_get_error().
However, due to upcoming changes to SSL_get_error() this will no longer be
the case. In order to maintain the existing legacy behaviour, explicitly
handle zero byte reads and writes in the TLSv1.3 stack, following
completion of a handshake.
ok inoguchi@ tb@
|
| |
|
|
|
|
| |
ok jmc@ sthen@ millert@
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
SSL_CTX_set_cipher_list() in OpenSSL 1.1 does not accept TLSv1.3 ciphers.
This wasn't a problem until now since the AEAD- ciphers were counted as
distinct from TLS_ ciphers by the regress test, so they were never used
in the {run,check}-cipher-${cipher}-client-${clib}-server-${slib} tests
With the renaming, the TLSv1.3 ciphers are now considered as common
ciphers, so they're tested. With openssl11 this results in
0:error:1410D0B9:SSL routines:SSL_CTX_set_cipher_list:no cipher match:ssl/ssl_lib.c:2573:
The design of these tests doesn't allow easily adding a call to
SSL_CTX_set_ciphersuites (since they also need to work with openssl 1.0.2)
so skip the TLS_* ciphers for the time being.
|
| |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
OpenSSL chose to break the previous naming convention for ciphers and
to adopt TLS_* "RFC" names instead. Unfortunately, these names are
exposed in several APIs and some language bindings test for these
non-standard names instead of cipher values, which is ... unfortunate
(others would say "plain crazy").
We currently have to maintain patches in regress and ports (p5-Net-SSLeay,
openssl-ruby-tests - which means that Ruby will pick this up at some point)
to work around this difference and that's just not worth the effort.
The old AEAD- names will become aliases and continue to work, but in
openssl ciphers and netcat output the TLS_* names will now be displayed.
"I would be very happy if this gets committed" bluhm
ok beck inoguchi, begrudgingly ok jsing
|
| |
|
|
|
|
|
|
|
|
| |
S3I has served us well, however now that libssl is fully opaque it is time
to say goodbye. Aside from removing the calloc/free/memset, the rest is
mechanical sed.
ok inoguchi@ tb@
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
LibreSSL's pc files effectively hardcode the version to 1.0.0 since
LibreSSL exists. That probably never made much sense. This causes
some pain for ports that "need 'openssl' ['>= +1.1.0'] found '1.0.0'"
or similar while they would build perfectly fine with LibreSSL.
This only affects OpenBSD. We do not put the actual LibreSSL version
in there since it may cause trouble
Discussed with sthen, millert, inoguchi, beck over the past year.
Diff from/ok sthen
|
|
|
|
|
|
|
|
| |
The trust anchor can't inherit, but the code says that it can inherit
just not if the leaf tries to inherit from that. This makes no sense
and doesn't match what is done on the asid side.
ok jsing
|
| |
|
| |
|
| |
|
|
|
|
| |
CRL's signature is invalid, not the certificate's.
|
|
|
|
|
|
| |
the code. Also add error checking where possible.
ok jsing
|
|
|
|
| |
ok jsing
|
|
|
|
|
|
|
|
| |
Remove the X509 argument as it is unused - this was passed so that
ssl_cert_type() can get the public key from the X509 object if the
EVP_PKEY argument is NULL, however this is never the case.
ok tb@
|
| |
|
|
|
|
|
|
|
|
| |
Rather than leaking libcrypto defines through the tls_sign_cb and
tls_signer_sign() interfaces, provide and use our own TLS_PADDING_*
defines.
ok inoguchi@ tb@
|