]> git.kaiwu.me - nginx.git/log
nginx.git
2 years agonginx-1.25.2-RELEASE release-1.25.2
Maxim Dounin [Tue, 15 Aug 2023 17:03:04 +0000 (20:03 +0300)]
nginx-1.25.2-RELEASE

2 years agoUpdated OpenSSL used for win32 builds.
Maxim Dounin [Tue, 15 Aug 2023 15:10:50 +0000 (18:10 +0300)]
Updated OpenSSL used for win32 builds.

2 years agoQUIC: path MTU discovery.
Roman Arutyunyan [Mon, 14 Aug 2023 05:21:27 +0000 (09:21 +0400)]
QUIC: path MTU discovery.

MTU selection starts by doubling the initial MTU until the first failure.
Then binary search is used to find the path MTU.

2 years agoQUIC: allowed ngx_quic_frame_sendto() to return NGX_AGAIN.
Roman Arutyunyan [Tue, 8 Aug 2023 06:43:17 +0000 (10:43 +0400)]
QUIC: allowed ngx_quic_frame_sendto() to return NGX_AGAIN.

Previously, NGX_AGAIN returned by ngx_quic_send() was treated by
ngx_quic_frame_sendto() as error, which triggered errors in its callers.
However, a blocked socket is not an error.  Now NGX_AGAIN is passed as is to
the ngx_quic_frame_sendto() callers, which can safely ignore it.

2 years agoQUIC: removed explicit packet padding for certain frames.
Roman Arutyunyan [Thu, 6 Jul 2023 07:30:47 +0000 (11:30 +0400)]
QUIC: removed explicit packet padding for certain frames.

The frames for which the padding is removed are PATH_CHALLENGE and
PATH_RESPONSE, which are sent separately by ngx_quic_frame_sendto().

2 years agoQUIC: removed path->limited flag.
Roman Arutyunyan [Thu, 6 Jul 2023 13:49:01 +0000 (17:49 +0400)]
QUIC: removed path->limited flag.

Its value is the opposite of path->validated.

2 years agoQUIC: fixed probe-congestion deadlock.
Roman Arutyunyan [Mon, 14 Aug 2023 04:28:30 +0000 (08:28 +0400)]
QUIC: fixed probe-congestion deadlock.

When probe timeout expired while congestion window was exhausted, probe PINGs
could not be sent.  As a result, lost packets could not be declared lost and
congestion window could not be freed for new packets.  This deadlock
continued until connection idle timeout expiration.

Now PINGs are sent separately from the frame queue without congestion control,
as specified by RFC 9002, Section 7:

  An endpoint MUST NOT send a packet if it would cause bytes_in_flight
  (see Appendix B.2) to be larger than the congestion window, unless the
  packet is sent on a PTO timer expiration (see Section 6.2) or when entering
  recovery (see Section 7.3.2).

2 years agoQUIC: fixed PTO expiration condition.
Roman Arutyunyan [Tue, 1 Aug 2023 07:21:59 +0000 (11:21 +0400)]
QUIC: fixed PTO expiration condition.

Previously, PTO handler analyzed the first packet in the sent queue for the
timeout expiration.  However, the last sent packet should be analyzed instead.
An example is timeout calculation in ngx_quic_set_lost_timer().

2 years agoQUIC: avoid accessing freed frame.
Roman Arutyunyan [Tue, 1 Aug 2023 07:20:04 +0000 (11:20 +0400)]
QUIC: avoid accessing freed frame.

Previously the field pnum of a potentially freed frame was accessed.  Now the
value is copied to a local variable.  The old behavior did not cause any
problems since the frame memory is not freed, but is moved to a free queue
instead.

2 years agoQUIC: fixed congesion control in GSO mode.
Roman Arutyunyan [Thu, 27 Jul 2023 09:35:42 +0000 (13:35 +0400)]
QUIC: fixed congesion control in GSO mode.

In non-GSO mode, a datagram is sent if congestion window is not exceeded by the
time of send.  The window could be exceeded by a small amount after the send.
In GSO mode, congestion window was checked in a similar way, but for all
concatenated datagrams as a whole.  This could result in exceeding congestion
window by a lot.  Now congestion window is checked for every datagram in GSO
mode as well.

2 years agoQUIC: always add ACK frame to the queue head.
Roman Arutyunyan [Thu, 10 Aug 2023 16:11:29 +0000 (20:11 +0400)]
QUIC: always add ACK frame to the queue head.

Previously it was added to the tail as all other frames.  However, if the
amount of queued data is large, it could delay the delivery of ACK, which
could trigger frames retransmissions and slow down the connection.

2 years agoQUIC: optimized ACK delay.
Roman Arutyunyan [Thu, 27 Jul 2023 12:37:17 +0000 (16:37 +0400)]
QUIC: optimized ACK delay.

Previously ACK was not generated if max_ack_delay was not yet expired and the
number of unacknowledged ack-eliciting packets was less than two, as allowed by
RFC 9000 13.2.1-13.2.2.  However this only makes sense to avoid sending ACK-only
packets, as explained by the RFC:

  On the other hand, reducing the frequency of packets that carry only
  acknowledgments reduces packet transmission and processing cost at both
  endpoints.

Now ACK is delayed only if output frame queue is empty.  Otherwise ACK is sent
immediately, which significantly improves QUIC performance with certain tests.

2 years agoSSL: avoid using OpenSSL config in build directory (ticket #2404).
Maxim Dounin [Tue, 20 Jun 2023 22:29:53 +0000 (01:29 +0300)]
SSL: avoid using OpenSSL config in build directory (ticket #2404).

With this change, the NGX_OPENSSL_NO_CONFIG macro is defined when nginx
is asked to build OpenSSL itself.  And with this macro automatic loading
of OpenSSL configuration (from the build directory) is prevented unless
the OPENSSL_CONF environment variable is explicitly set.

Note that not loading configuration is broken in OpenSSL 1.1.1 and 1.1.1a
(fixed in OpenSSL 1.1.1b, see https://github.com/openssl/openssl/issues/7350).
If nginx is used to compile these OpenSSL versions, configuring nginx with
NGX_OPENSSL_NO_CONFIG explicitly set to 0 might be used as a workaround.

2 years agoSSL: provided "nginx" appname when loading OpenSSL configs.
Maxim Dounin [Tue, 20 Jun 2023 22:29:55 +0000 (01:29 +0300)]
SSL: provided "nginx" appname when loading OpenSSL configs.

Following OpenSSL 0.9.8f, OpenSSL tries to load application-specific
configuration section first, and then falls back to the "openssl_conf"
default section if application-specific section is not found, by using
CONF_modules_load_file(CONF_MFLAGS_DEFAULT_SECTION).  Therefore this
change is not expected to introduce any compatibility issues with existing
configurations.  It does, however, make it easier to configure specific
OpenSSL settings for nginx in system-wide OpenSSL configuration
(ticket #2449).

Instead of checking OPENSSL_VERSION_NUMBER when using the OPENSSL_init_ssl()
interface, the code now tests for OPENSSL_INIT_LOAD_CONFIG to be defined and
true, and also explicitly excludes LibreSSL.  This ensures that this interface
is not used with BoringSSL and LibreSSL, which do not provide additional
library initialization settings, notably the OPENSSL_INIT_set_config_appname()
call.

2 years agoContrib: vim syntax, update core and 3rd party module directives.
Gena Makhomed [Mon, 24 Jul 2023 15:04:41 +0000 (18:04 +0300)]
Contrib: vim syntax, update core and 3rd party module directives.

List of 3rd party modules github repositories was obtained from
https://github.com/freebsd/freebsd-ports/blob/main/www/nginx-devel/Makefile.extmod

2 years agoCore: fixed environment variables on exit.
Maxim Dounin [Wed, 19 Jul 2023 02:09:23 +0000 (05:09 +0300)]
Core: fixed environment variables on exit.

Similarly to 6822:c045b4926b2c, environment variables introduced with
the "env" directive (and "NGINX_BPF_MAPS" added by QUIC) are now allocated
via ngx_alloc(), and explicitly freed by a cleanup handler if no longer used.

In collaboration with Sergey Kandaurov.

2 years agoHTTP/3: fixed $body_bytes_sent.
Sergey Kandaurov [Wed, 12 Jul 2023 11:27:35 +0000 (15:27 +0400)]
HTTP/3: fixed $body_bytes_sent.

2 years agoQUIC: use AEAD to encrypt address validation tokens.
Roman Arutyunyan [Thu, 8 Jun 2023 10:58:01 +0000 (14:58 +0400)]
QUIC: use AEAD to encrypt address validation tokens.

Previously used AES256-CBC is now substituted with AES256-GCM.  Although there
seem to be no tangible consequences of token integrity loss.

2 years agoQUIC: removed TLS1_3_CK_* macros wrap up.
Sergey Kandaurov [Fri, 16 Jun 2023 13:13:29 +0000 (17:13 +0400)]
QUIC: removed TLS1_3_CK_* macros wrap up.

They were preserved in 172705615d04 to ease transition from older BoringSSL.

2 years agoQUIC: style.
Sergey Kandaurov [Tue, 20 Jun 2023 13:59:02 +0000 (17:59 +0400)]
QUIC: style.

2 years agoQUIC: unified ngx_quic_tls_open() and ngx_quic_tls_seal().
Sergey Kandaurov [Tue, 20 Jun 2023 13:59:01 +0000 (17:59 +0400)]
QUIC: unified ngx_quic_tls_open() and ngx_quic_tls_seal().

2 years agoQUIC: TLS_AES_128_CCM_SHA256 cipher suite support.
Roman Arutyunyan [Tue, 20 Jun 2023 12:10:49 +0000 (16:10 +0400)]
QUIC: TLS_AES_128_CCM_SHA256 cipher suite support.

2 years agoQUIC: common cipher control constants instead of GCM-related.
Roman Arutyunyan [Fri, 9 Jun 2023 06:23:22 +0000 (10:23 +0400)]
QUIC: common cipher control constants instead of GCM-related.

The constants are used for both GCM and CHACHAPOLY.

2 years agoQUIC: a new constant for AEAD tag length.
Roman Arutyunyan [Fri, 9 Jun 2023 06:25:54 +0000 (10:25 +0400)]
QUIC: a new constant for AEAD tag length.

Previously used constant EVP_GCM_TLS_TAG_LEN had misleading name since it was
used not only with GCM, but also with CHACHAPOLY.  Now a new constant
NGX_QUIC_TAG_LEN introduced.  Luckily all AEAD algorithms used by QUIC have
the same tag length of 16.

2 years agoVersion bump.
Roman Arutyunyan [Tue, 20 Jun 2023 13:01:00 +0000 (17:01 +0400)]
Version bump.

2 years agorelease-1.25.1 tag
Maxim Dounin [Tue, 13 Jun 2023 15:08:10 +0000 (18:08 +0300)]
release-1.25.1 tag

2 years agonginx-1.25.1-RELEASE release-1.25.1
Maxim Dounin [Tue, 13 Jun 2023 15:08:09 +0000 (18:08 +0300)]
nginx-1.25.1-RELEASE

2 years agoQUIC: fixed rttvar on subsequent RTT samples (ticket #2505).
Sergey Kandaurov [Mon, 12 Jun 2023 19:38:56 +0000 (23:38 +0400)]
QUIC: fixed rttvar on subsequent RTT samples (ticket #2505).

Previously, computing rttvar used an updated smoothed_rtt value as per
RFC 9002, section 5.3, which appears to be specified in a wrong order.
A technical errata ID 7539 is reported.

2 years agoHTTP/2: removed server push (ticket #2432).
Sergey Kandaurov [Thu, 8 Jun 2023 12:56:46 +0000 (16:56 +0400)]
HTTP/2: removed server push (ticket #2432).

Although it has better implementation status than HTTP/3 server push,
it remains of limited use, with adoption numbers seen as negligible.
Per IETF 102 materials, server push was used only in 0.04% of sessions.
It was considered to be "difficult to use effectively" in RFC 9113.
Its use is further limited by badly matching to fetch/cache/connection
models in browsers, see related discussions linked from [1].

Server push was disabled in Chrome 106 [2].

The http2_push, http2_push_preload, and http2_max_concurrent_pushes
directives are made obsolete.  In particular, this essentially reverts
7201:641306096f5b and 7207:3d2b0b02bd3d.

[1] https://jakearchibald.com/2017/h2-push-tougher-than-i-thought/
[2] https://chromestatus.com/feature/6302414934114304

2 years agoSSL: removed the "ssl" directive.
Roman Arutyunyan [Thu, 8 Jun 2023 10:49:27 +0000 (14:49 +0400)]
SSL: removed the "ssl" directive.

It has been deprecated since 7270:46c0c7ef4913 (1.15.0) in favour of
the "ssl" parameter of the "listen" directive, which has been available
since 2224:109849282793 (0.7.14).

2 years agoHTTP/2: "http2" directive.
Roman Arutyunyan [Tue, 16 May 2023 12:30:08 +0000 (16:30 +0400)]
HTTP/2: "http2" directive.

The directive enables HTTP/2 in the current server.  The previous way to
enable HTTP/2 via "listen ... http2" is now deprecated.  The new approach
allows to share HTTP/2 and HTTP/0.9-1.1 on the same port.

For SSL connections, HTTP/2 is now selected by ALPN callback based on whether
the protocol is enabled in the virtual server chosen by SNI.  This however only
works since OpenSSL 1.0.2h, where ALPN callback is invoked after SNI callback.
For older versions of OpenSSL, HTTP/2 is enabled based on the default virtual
server configuration.

For plain TCP connections, HTTP/2 is now auto-detected by HTTP/2 preface, if
HTTP/2 is enabled in the default virtual server.  If preface is not matched,
HTTP/0.9-1.1 is assumed.

2 years agoQUIC: fixed compat with ciphers other than AES128 (ticket #2500).
Roman Arutyunyan [Sun, 28 May 2023 07:17:07 +0000 (11:17 +0400)]
QUIC: fixed compat with ciphers other than AES128 (ticket #2500).

Previously, rec.level field was not uninitialized in SSL_provide_quic_data().
As a result, its value was always ssl_encryption_initial.  Later in
ngx_quic_ciphers() such level resulted in resetting the cipher to
TLS1_3_CK_AES_128_GCM_SHA256 and using AES128 to encrypt the packet.

Now the level is initialized and the right cipher is used.

2 years agoVersion bump.
Roman Arutyunyan [Mon, 29 May 2023 11:03:31 +0000 (15:03 +0400)]
Version bump.

2 years agorelease-1.25.0 tag
Maxim Dounin [Tue, 23 May 2023 15:08:20 +0000 (18:08 +0300)]
release-1.25.0 tag

2 years agonginx-1.25.0-RELEASE release-1.25.0
Maxim Dounin [Tue, 23 May 2023 15:08:19 +0000 (18:08 +0300)]
nginx-1.25.0-RELEASE

2 years agoQUIC: fixed OpenSSL compat layer with OpenSSL master branch.
Sergey Kandaurov [Mon, 22 May 2023 20:45:18 +0000 (00:45 +0400)]
QUIC: fixed OpenSSL compat layer with OpenSSL master branch.

The layer is enabled as a fallback if the QUIC support is configured and the
BoringSSL API wasn't detected, or when using the --with-openssl option, also
compatible with QuicTLS and LibreSSL.  For the latter, the layer is assumed
to be present if QUIC was requested, so it needs to be undefined to prevent
QUIC API redefinition as appropriate.

A previously used approach to test the TLSEXT_TYPE_quic_transport_parameters
macro doesn't work with OpenSSL 3.2 master branch where this macro appeared
with incompatible QUIC API.  To fix the build there, the test is revised to
pass only for QuicTLS and LibreSSL.

2 years agoQUIC: fixed post-close use-after-free.
Roman Arutyunyan [Mon, 22 May 2023 11:59:42 +0000 (15:59 +0400)]
QUIC: fixed post-close use-after-free.

Previously, ngx_quic_close_connection() could be called in a way that QUIC
connection was accessed after the call.  In most cases the connection is not
closed right away, but close timeout is scheduled.  However, it's not always
the case.  Also, if the close process started earlier for a different reason,
calling ngx_quic_close_connection() may actually close the connection.  The
connection object should not be accessed after that.

Now, when possible, return statement is added to eliminate post-close connection
object access.  In other places ngx_quic_close_connection() is substituted with
posting close event.

Also, the new way of closing connection in ngx_quic_stream_cleanup_handler()
fixes another problem in this function.  Previously it passed stream connection
instead of QUIC connection to ngx_quic_close_connection().  This could result
in incomplete connection shutdown.  One consequence of that could be that QUIC
streams were freed without shutting down their application contexts.  This could
result in another use-after-free.

Found by Coverity (CID 1530402).

2 years agoQUIC: better sockaddr initialization.
Maxim Dounin [Sun, 21 May 2023 01:38:45 +0000 (04:38 +0300)]
QUIC: better sockaddr initialization.

The qsock->sockaddr field is a ngx_sockaddr_t union, and therefore can hold
any sockaddr (and union members, such qsock->sockaddr.sockaddr, can be used
to access appropriate variant of the sockaddr).  It is better to set it via
qsock->sockaddr itself though, and not qsock->sockaddr.sockaddr, so static
analyzers won't complain about out-of-bounds access.

Prodded by Coverity (CID 1530403).

2 years agoMerged with the quic branch.
Roman Arutyunyan [Fri, 19 May 2023 17:46:36 +0000 (21:46 +0400)]
Merged with the quic branch.

2 years agoRemoved README.
Roman Arutyunyan [Sun, 14 May 2023 06:31:24 +0000 (10:31 +0400)]
Removed README.

3 years agoHTTP/3: removed server push support.
Roman Arutyunyan [Fri, 12 May 2023 06:02:10 +0000 (10:02 +0400)]
HTTP/3: removed server push support.

2 years agoCommon tree insert function for QUIC and UDP connections.
Roman Arutyunyan [Sun, 14 May 2023 08:30:11 +0000 (12:30 +0400)]
Common tree insert function for QUIC and UDP connections.

Previously, ngx_udp_rbtree_insert_value() was used for plain UDP and
ngx_quic_rbtree_insert_value() was used for QUIC.  Because of this it was
impossible to initialize connection tree in ngx_create_listening() since
this function is not aware what kind of listening it creates.

Now ngx_udp_rbtree_insert_value() is used for both QUIC and UDP.  To make
is possible, a generic key field is added to ngx_udp_connection_t.  It keeps
client address for UDP and connection ID for QUIC.

2 years agoStream: removed QUIC support.
Roman Arutyunyan [Sun, 14 May 2023 08:05:35 +0000 (12:05 +0400)]
Stream: removed QUIC support.

3 years agoQUIC: style.
Maxim Dounin [Thu, 11 May 2023 15:48:01 +0000 (18:48 +0300)]
QUIC: style.

3 years agoHTTP/3: removed "http3" parameter of "listen" directive.
Roman Arutyunyan [Thu, 11 May 2023 09:22:10 +0000 (13:22 +0400)]
HTTP/3: removed "http3" parameter of "listen" directive.

The parameter has been deprecated since c851a2ed5ce8.

3 years agoQUIC: removed "quic_mtu" directive.
Roman Arutyunyan [Thu, 11 May 2023 06:37:51 +0000 (10:37 +0400)]
QUIC: removed "quic_mtu" directive.

The directive used to set the value of the "max_udp_payload_size" transport
parameter.  According to RFC 9000, Section 18.2, the value specifies the size
of buffer for reading incoming datagrams:

    This limit does act as an additional constraint on datagram size in
    the same way as the path MTU, but it is a property of the endpoint
    and not the path; see Section 14. It is expected that this is the
    space an endpoint dedicates to holding incoming packets.

Current QUIC implementation uses the maximum possible buffer size (65527) for
reading datagrams.

3 years agoQUIC: resized input datagram buffer from 65535 to 65527.
Roman Arutyunyan [Thu, 11 May 2023 05:49:34 +0000 (09:49 +0400)]
QUIC: resized input datagram buffer from 65535 to 65527.

The value of 65527 is the maximum permitted UDP payload size.

3 years agoQUIC: keep stream sockaddr and addr_text constant.
Roman Arutyunyan [Thu, 11 May 2023 15:40:11 +0000 (19:40 +0400)]
QUIC: keep stream sockaddr and addr_text constant.

HTTP and Stream variables $remote_addr and $binary_remote_addr rely on
constant client address, particularly because they are cacheable.
However, QUIC client may migrate to a new address.  While there's no perfect
way to handle this, the proposed solution is to copy client address to QUIC
stream at stream creation.

The change also fixes truncated $remote_addr if migration happened while the
stream was active.  The reason is addr_text string was copied to stream by
value.

3 years agoVariables: avoid possible buffer overrun with some "$sent_http_*".
Sergey Kandaurov [Mon, 1 May 2023 15:16:05 +0000 (19:16 +0400)]
Variables: avoid possible buffer overrun with some "$sent_http_*".

The existing logic to evaluate multi header "$sent_http_*" variables,
such as $sent_http_cache_control, as previously introduced in 1.23.0,
doesn't take into account that one or more elements can be cleared,
yet still present in a linked list, pointed to by the next field.
Such elements don't contribute to the resulting variable length, an
attempt to append a separator for them ends up in out of bounds write.

This is not possible with standard modules, though at least one third
party module is known to override multi header values this way, so it
makes sense to harden the logic.

The fix restores a generic boundary check.

3 years agoQUIC: set c->socklen for streams.
Roman Arutyunyan [Thu, 27 Apr 2023 15:49:05 +0000 (19:49 +0400)]
QUIC: set c->socklen for streams.

Previously, the value was not set and remained zero.  While in nginx code the
value of c->sockaddr is accessed without taking c->socklen into account,
invalid c->socklen could lead to unexpected results in third-party modules.

3 years agoQUIC: fixed addr_text after migration (ticket #2488).
Roman Arutyunyan [Thu, 27 Apr 2023 15:52:40 +0000 (19:52 +0400)]
QUIC: fixed addr_text after migration (ticket #2488).

Previously, the post-migration value of addr_text could be truncated, if
it was longer than the previous one.  Also, the new value always included
port, which should not be there.

3 years agoQUIC: reschedule path validation on path insertion/removal.
Sergey Kandaurov [Tue, 9 May 2023 15:42:40 +0000 (19:42 +0400)]
QUIC: reschedule path validation on path insertion/removal.

Two issues fixed:
- new path validation could be scheduled late
- a validated path could leave a spurious timer

3 years agoQUIC: lower bound path validation PTO.
Sergey Kandaurov [Tue, 9 May 2023 15:42:40 +0000 (19:42 +0400)]
QUIC: lower bound path validation PTO.

According to RFC 9000, 8.2.4. Failed Path Validation,
the following value is recommended as a validation timeout:

  A value of three times the larger of the current PTO
  or the PTO for the new path (using kInitialRtt, as
  defined in [QUIC-RECOVERY]) is RECOMMENDED.

The change adds PTO of the new path to the equation as the lower bound.

3 years agoQUIC: separated path validation retransmit backoff.
Sergey Kandaurov [Tue, 9 May 2023 15:42:39 +0000 (19:42 +0400)]
QUIC: separated path validation retransmit backoff.

Path validation packets containing PATH_CHALLENGE frames are sent separately
from regular frame queue, because of the need to use a decicated path and pad
the packets.  The packets are sent periodically, separately from the regular
probe/lost detection mechanism.  A path validation packet is resent up to 3
times, each time after PTO expiration, with increasing per-path PTO backoff.

3 years agoQUIC: removed check for in-flight packets in computing PTO.
Sergey Kandaurov [Tue, 9 May 2023 15:42:38 +0000 (19:42 +0400)]
QUIC: removed check for in-flight packets in computing PTO.

The check is needed for clients in order to unblock a server due to
anti-amplification limits, and it seems to make no sense for servers.
See RFC 9002, A.6 and A.8 for a further explanation.

This makes max_ack_delay to now always account, notably including
PATH_CHALLENGE timers as noted in the last paragraph of 9000, 9.4,
unlike when it was only used when there are packets in flight.

While here, fixed nearby style.

3 years agoQUIC: disabled datagram fragmentation.
Roman Arutyunyan [Sat, 6 May 2023 12:23:27 +0000 (16:23 +0400)]
QUIC: disabled datagram fragmentation.

As per RFC 9000, Section 14:

  UDP datagrams MUST NOT be fragmented at the IP layer.

3 years agoQUIC: fixed encryption level in ngx_quic_frame_sendto().
Roman Arutyunyan [Thu, 4 May 2023 15:29:34 +0000 (19:29 +0400)]
QUIC: fixed encryption level in ngx_quic_frame_sendto().

Previously, ssl_encryption_application was hardcoded.  Before 9553eea74f2a,
ngx_quic_frame_sendto() was used only for PATH_CHALLENGE/PATH_RESPONSE sent
at the application level only.  Since 9553eea74f2a, ngx_quic_frame_sendto()
is also used for CONNECTION_CLOSE, which can be sent at initial level after
SSL handshake error or rejection.  This resulted in packet encryption error.
Now level is copied from frame, which fixes the error.

3 years agoQUIC: optimized immediate close.
Roman Arutyunyan [Tue, 2 May 2023 13:54:53 +0000 (17:54 +0400)]
QUIC: optimized immediate close.

Previously, before sending CONNECTION_CLOSE to client, all pending frames
were sent.  This is redundant and could prevent CONNECTION_CLOSE from being
sent due to congestion control.  Now pending frames are freed and
CONNECTION_CLOSE is sent without congestion control, as advised by RFC 9002:

  Packets containing frames besides ACK or CONNECTION_CLOSE frames
  count toward congestion control limits and are considered to be in flight.

3 years agoQUIC: fixed split frames error handling.
Sergey Kandaurov [Thu, 4 May 2023 11:52:23 +0000 (15:52 +0400)]
QUIC: fixed split frames error handling.

Do not corrupt frame data chain pointer on ngx_quic_read_buffer() error.
The error leads to closing a QUIC connection where the frame may be used
as part of the QUIC connection tear down, which envolves writing pending
frames, including this one.

3 years agoHTTP/3: fixed ngx_http_v3_init_session() error handling.
Sergey Kandaurov [Thu, 4 May 2023 11:52:22 +0000 (15:52 +0400)]
HTTP/3: fixed ngx_http_v3_init_session() error handling.

A QUIC connection is not usable yet at this early stage of spin up.

3 years agoFixed segfault if regex studies list allocation fails.
Maxim Dounin [Tue, 18 Apr 2023 03:28:46 +0000 (06:28 +0300)]
Fixed segfault if regex studies list allocation fails.

The rcf->studies list is unconditionally accessed by ngx_regex_cleanup(),
and this used to cause NULL pointer dereference if allocation
failed.  Fix is to set cleanup handler only when allocation succeeds.

3 years agoAdded stream modules realip and ssl_preread to win32 builds.
Sergey Kandaurov [Mon, 17 Apr 2023 10:08:00 +0000 (14:08 +0400)]
Added stream modules realip and ssl_preread to win32 builds.

3 years agoYear 2023.
Sergey Kandaurov [Mon, 17 Apr 2023 10:07:59 +0000 (14:07 +0400)]
Year 2023.

3 years agoVersion bump.
Sergey Kandaurov [Mon, 17 Apr 2023 10:06:43 +0000 (14:06 +0400)]
Version bump.

3 years agoREADME: revised TLSv1.3 requirement for QUIC.
Roman Arutyunyan [Tue, 11 Apr 2023 14:29:20 +0000 (18:29 +0400)]
README: revised TLSv1.3 requirement for QUIC.

TLSv1.3 is enabled by default since d1cf09451ae8.

3 years agoStream: allow waiting on a blocked QUIC stream (ticket #2479).
Roman Arutyunyan [Thu, 6 Apr 2023 11:39:48 +0000 (15:39 +0400)]
Stream: allow waiting on a blocked QUIC stream (ticket #2479).

Previously, waiting on a shared connection was not allowed, because the only
type of such connection was plain UDP.  However, QUIC stream connections are
also shared since they share socket descriptor with the listen connection.
Meanwhile, it's perfectly normal to wait on such connections.

The issue manifested itself with stream write errors when the amount of data
exceeded stream buffer size or flow control.  Now no error is triggered
and Stream write module is allowed to wait for buffer space to become available.

3 years agoHTTP/3: fixed CANCEL_PUSH handling.
Sergey Kandaurov [Thu, 6 Apr 2023 14:18:41 +0000 (18:18 +0400)]
HTTP/3: fixed CANCEL_PUSH handling.

3 years agoQUIC: optimized sending stream response.
Roman Arutyunyan [Mon, 3 Apr 2023 12:17:12 +0000 (16:17 +0400)]
QUIC: optimized sending stream response.

When a stream is created by client, it's often the case that nginx will send
immediate response on that stream.  An example is HTTP/3 request stream, which
in most cases quickly replies with at least HTTP headers.

QUIC stream init handlers are called from a posted event.  Output QUIC
frames are also sent to client from a posted event, called the push event.
If the push event is posted before the stream init event, then output produced
by stream may trigger sending an extra UDP datagram.  To address this, push
event is now re-posted when a new stream init event is posted.

An example is handling 0-RTT packets.  Client typically sends an init packet
coalesced with a 0-RTT packet.  Previously, nginx replied with a padded CRYPTO
datagram, followed by a 1-RTT stream reply datagram.  Now CRYPTO and STREAM
packets are coalesced in one reply datagram, which saves bandwidth.

Other examples include coalescing 1-RTT first stream response, and
MAX_STREAMS/STREAM sent in response to ACK/STREAM.

3 years agoMerged with the default branch.
Sergey Kandaurov [Wed, 29 Mar 2023 07:14:25 +0000 (11:14 +0400)]
Merged with the default branch.

3 years agorelease-1.23.4 tag
Maxim Dounin [Tue, 28 Mar 2023 15:01:54 +0000 (18:01 +0300)]
release-1.23.4 tag

3 years agonginx-1.23.4-RELEASE release-1.23.4
Maxim Dounin [Tue, 28 Mar 2023 15:01:53 +0000 (18:01 +0300)]
nginx-1.23.4-RELEASE

3 years agoUpdated OpenSSL used for win32 builds.
Maxim Dounin [Mon, 27 Mar 2023 23:25:55 +0000 (02:25 +0300)]
Updated OpenSSL used for win32 builds.

3 years agoGzip: compatibility with recent zlib-ng versions.
Maxim Dounin [Mon, 27 Mar 2023 18:25:05 +0000 (21:25 +0300)]
Gzip: compatibility with recent zlib-ng versions.

It now uses custom alloc_aligned() wrapper for all allocations,
therefore all allocations are larger than expected by (64 + sizeof(void*)).
Further, they are seen as allocations of 1 element.  Relevant calculations
were adjusted to reflect this, and state allocation is now protected
with a flag to avoid misinterpreting other allocations as the zlib
deflate_state allocation.

Further, it no longer forces window bits to 13 on compression level 1,
so the comment was adjusted to reflect this.

3 years agoSSL: enabled TLSv1.3 by default.
Maxim Dounin [Thu, 23 Mar 2023 23:57:43 +0000 (02:57 +0300)]
SSL: enabled TLSv1.3 by default.

3 years agoMail: fixed handling of blocked client read events in proxy.
Maxim Dounin [Thu, 23 Mar 2023 23:53:21 +0000 (02:53 +0300)]
Mail: fixed handling of blocked client read events in proxy.

When establishing a connection to the backend, nginx blocks reading
from the client with ngx_mail_proxy_block_read().  Previously, such
events were lost, and in some cases this resulted in connection hangs.

Notably, this affected mail_imap_ssl.t on Windows, since the test
closes connections after requesting authentication, but without
waiting for any responses (so the connection close events might be
lost).

Fix is to post an event to read from the client after connecting to
the backend if there were blocked events.

3 years agoQUIC: style.
Roman Arutyunyan [Wed, 15 Mar 2023 15:57:15 +0000 (19:57 +0400)]
QUIC: style.

3 years agoHTTP/3: fixed OpenSSL compatibility layer initialization.
Sergey Kandaurov [Fri, 24 Mar 2023 15:49:50 +0000 (19:49 +0400)]
HTTP/3: fixed OpenSSL compatibility layer initialization.

SSL context is not present if the default server has neither certificates nor
ssl_reject_handshake enabled.  Previously, this led to null pointer dereference
before it would be caught with configuration checks.

Additionally, non-default servers with distinct SSL contexts need to initialize
compatibility layer in order to complete a QUIC handshake.

3 years agoSyslog: introduced error log handler.
Maxim Dounin [Fri, 10 Mar 2023 04:43:50 +0000 (07:43 +0300)]
Syslog: introduced error log handler.

This ensures that errors which happen during logging to syslog are logged
with proper context, such as "while logging to syslog" and the server name.

Prodded by Safar Safarly.

3 years agoSyslog: removed usage of ngx_cycle->log and ngx_cycle->hostname.
Maxim Dounin [Fri, 10 Mar 2023 04:43:40 +0000 (07:43 +0300)]
Syslog: removed usage of ngx_cycle->log and ngx_cycle->hostname.

During initial startup the ngx_cycle->hostname is not available, and
previously this resulted in incorrect logging.  Instead, hostname from the
configuration being parsed is now preserved in the syslog peer structure
and then used during logging.

Similarly, ngx_cycle->log might not match the configuration where the
syslog peer is defined if the configuration is not yet fully applied,
and previously this resulted in unexpected logging of syslog errors
and debug information.  Instead, cf->cycle->new_log is now referenced
in the syslog peer structure and used for logging, similarly to how it
is done in other modules.

3 years agoHTTP/2: finalize request as bad if header validation fails.
Maxim Dounin [Fri, 10 Mar 2023 03:47:53 +0000 (06:47 +0300)]
HTTP/2: finalize request as bad if header validation fails.

Similarly to 7192:d5a535774861, this avoids spurious zero statuses
in access.log, and in line with other header-related errors.

3 years agoHTTP/2: socket leak with "return 444" in error_page (ticket #2455).
Maxim Dounin [Fri, 10 Mar 2023 03:47:48 +0000 (06:47 +0300)]
HTTP/2: socket leak with "return 444" in error_page (ticket #2455).

Similarly to ticket #274 (7354:1812f1d79d84), early request finalization
without calling ngx_http_run_posted_requests() resulted in a connection
hang (a socket leak) if the 400 (Bad Request) error was generated in
ngx_http_v2_state_process_header() due to invalid request headers and
"return 444" was used in error_page 400.

3 years agoSSL: logging levels of errors observed with BoringSSL.
Maxim Dounin [Wed, 8 Mar 2023 19:22:47 +0000 (22:22 +0300)]
SSL: logging levels of errors observed with BoringSSL.

As tested with tlsfuzzer with BoringSSL, the following errors are
certainly client-related:

SSL_do_handshake() failed (SSL: error:10000066:SSL routines:OPENSSL_internal:BAD_ALERT)
SSL_do_handshake() failed (SSL: error:10000089:SSL routines:OPENSSL_internal:DECODE_ERROR)
SSL_do_handshake() failed (SSL: error:100000dc:SSL routines:OPENSSL_internal:TOO_MANY_WARNING_ALERTS)
SSL_do_handshake() failed (SSL: error:10000100:SSL routines:OPENSSL_internal:INVALID_COMPRESSION_LIST)
SSL_do_handshake() failed (SSL: error:10000102:SSL routines:OPENSSL_internal:MISSING_KEY_SHARE)
SSL_do_handshake() failed (SSL: error:1000010e:SSL routines:OPENSSL_internal:TOO_MUCH_SKIPPED_EARLY_DATA)
SSL_read() failed (SSL: error:100000b6:SSL routines:OPENSSL_internal:NO_RENEGOTIATION)

Accordingly, the SSL_R_BAD_ALERT, SSL_R_DECODE_ERROR,
SSL_R_TOO_MANY_WARNING_ALERTS, SSL_R_INVALID_COMPRESSION_LIST,
SSL_R_MISSING_KEY_SHARE, SSL_R_TOO_MUCH_SKIPPED_EARLY_DATA,
and SSL_R_NO_RENEGOTIATION errors are now logged at the "info" level.

3 years agoSSL: logging levels of errors observed with tlsfuzzer and LibreSSL.
Maxim Dounin [Wed, 8 Mar 2023 19:22:34 +0000 (22:22 +0300)]
SSL: logging levels of errors observed with tlsfuzzer and LibreSSL.

As tested with tlsfuzzer with LibreSSL 3.7.0, the following errors are
certainly client-related:

SSL_do_handshake() failed (SSL: error:14026073:SSL routines:ACCEPT_SR_CLNT_HELLO:bad packet length)
SSL_do_handshake() failed (SSL: error:1402612C:SSL routines:ACCEPT_SR_CLNT_HELLO:ssl3 session id too long)
SSL_do_handshake() failed (SSL: error:140380EA:SSL routines:ACCEPT_SR_KEY_EXCH:tls rsa encrypted value length is wrong)

Accordingly, the SSL_R_BAD_PACKET_LENGTH ("bad packet length"),
SSL_R_SSL3_SESSION_ID_TOO_LONG ("ssl3 session id too long"),
SSL_R_TLS_RSA_ENCRYPTED_VALUE_LENGTH_IS_WRONG ("tls rsa encrypted value
length is wrong") errors are now logged at the "info" level.

3 years agoSSL: logging levels of various errors reported with tlsfuzzer.
Maxim Dounin [Wed, 8 Mar 2023 19:21:59 +0000 (22:21 +0300)]
SSL: logging levels of various errors reported with tlsfuzzer.

To further differentiate client-related errors and adjust logging levels
of various SSL errors, nginx was tested with tlsfuzzer with multiple
OpenSSL versions (3.1.0-beta1, 3.0.8, 1.1.1t, 1.1.0l, 1.0.2u, 1.0.1u,
1.0.0s, 0.9.8zh).

The following errors were observed during tlsfuzzer runs with OpenSSL 3.0.8,
and are clearly client-related:

SSL_do_handshake() failed (SSL: error:0A000092:SSL routines::data length too long)
SSL_do_handshake() failed (SSL: error:0A0000A0:SSL routines::length too short)
SSL_do_handshake() failed (SSL: error:0A000124:SSL routines::bad legacy version)
SSL_do_handshake() failed (SSL: error:0A000178:SSL routines::no shared signature algorithms)

Accordingly, the SSL_R_DATA_LENGTH_TOO_LONG ("data length too long"),
SSL_R_LENGTH_TOO_SHORT ("length too short"), SSL_R_BAD_LEGACY_VERSION
("bad legacy version"), and SSL_R_NO_SHARED_SIGNATURE_ALGORITHMS
("no shared signature algorithms", misspelled as "sigature" in OpenSSL 1.0.2)
errors are now logged at the "info" level.

Additionally, the following errors were observed with OpenSSL 3.0.8 and
with TLSv1.3 enabled:

SSL_do_handshake() failed (SSL: error:0A00006F:SSL routines::bad digest length)
SSL_do_handshake() failed (SSL: error:0A000070:SSL routines::missing sigalgs extension)
SSL_do_handshake() failed (SSL: error:0A000096:SSL routines::encrypted length too long)
SSL_do_handshake() failed (SSL: error:0A00010F:SSL routines::bad length)
SSL_read() failed (SSL: error:0A00007A:SSL routines::bad key update)
SSL_read() failed (SSL: error:0A000125:SSL routines::mixed handshake and non handshake data)

Accordingly, the SSL_R_BAD_DIGEST_LENGTH ("bad digest length"),
SSL_R_MISSING_SIGALGS_EXTENSION ("missing sigalgs extension"),
SSL_R_ENCRYPTED_LENGTH_TOO_LONG ("encrypted length too long"),
SSL_R_BAD_LENGTH ("bad length"), SSL_R_BAD_KEY_UPDATE ("bad key update"),
and SSL_R_MIXED_HANDSHAKE_AND_NON_HANDSHAKE_DATA ("mixed handshake and non
handshake data") errors are now logged at the "info" level.

Additionally, the following errors were observed with OpenSSL 1.1.1t:

SSL_do_handshake() failed (SSL: error:14094091:SSL routines:ssl3_read_bytes:data between ccs and finished)
SSL_do_handshake() failed (SSL: error:14094199:SSL routines:ssl3_read_bytes:too many warn alerts)
SSL_read() failed (SSL: error:1408F0C6:SSL routines:ssl3_get_record:packet length too long)
SSL_read() failed (SSL: error:14094085:SSL routines:ssl3_read_bytes:ccs received early)

Accordingly, the SSL_R_CCS_RECEIVED_EARLY ("ccs received early"),
SSL_R_DATA_BETWEEN_CCS_AND_FINISHED ("data between ccs and finished"),
SSL_R_PACKET_LENGTH_TOO_LONG ("packet length too long"), and
SSL_R_TOO_MANY_WARN_ALERTS ("too many warn alerts") errors are now logged
at the "info" level.

Additionally, the following errors were observed with OpenSSL 1.0.2u:

SSL_do_handshake() failed (SSL: error:1407612A:SSL routines:SSL23_GET_CLIENT_HELLO:record too small)
SSL_do_handshake() failed (SSL: error:1408C09A:SSL routines:ssl3_get_finished:got a fin before a ccs)

Accordingly, the SSL_R_RECORD_TOO_SMALL ("record too small") and
SSL_R_GOT_A_FIN_BEFORE_A_CCS ("got a fin before a ccs") errors are now
logged at the "info" level.

No additional client-related errors were observed while testing with
OpenSSL 3.1.0-beta1, OpenSSL 1.1.0l, OpenSSL 1.0.1u, OpenSSL 1.0.0s,
and OpenSSL 0.9.8zh.

3 years agoSSL: switched to detect log level based on the last error.
Maxim Dounin [Wed, 8 Mar 2023 19:21:53 +0000 (22:21 +0300)]
SSL: switched to detect log level based on the last error.

In some cases there might be multiple errors in the OpenSSL error queue,
notably when a libcrypto call fails, and then the SSL layer generates
an error itself.  For example, the following errors were observed
with OpenSSL 3.0.8 with TLSv1.3 enabled:

SSL_do_handshake() failed (SSL: error:02800066:Diffie-Hellman routines::invalid public key error:0A000132:SSL routines::bad ecpoint)
SSL_do_handshake() failed (SSL: error:08000066:elliptic curve routines::invalid encoding error:0A000132:SSL routines::bad ecpoint)
SSL_do_handshake() failed (SSL: error:0800006B:elliptic curve routines::point is not on curve error:0A000132:SSL routines::bad ecpoint)

In such cases it seems to be better to determine logging level based on
the last error in the error queue (the one added by the SSL layer,
SSL_R_BAD_ECPOINT in all of the above example example errors).  To do so,
the ngx_ssl_connection_error() function was changed to use
ERR_peek_last_error().

3 years agoCore: stricter UTF-8 handling in ngx_utf8_decode().
Yugo Horie [Wed, 22 Feb 2023 23:09:50 +0000 (08:09 +0900)]
Core: stricter UTF-8 handling in ngx_utf8_decode().

An UTF-8 octet sequence cannot start with a 11111xxx byte (above 0xf8),
see https://datatracker.ietf.org/doc/html/rfc3629#section-3.  Previously,
such bytes were accepted by ngx_utf8_decode() and misinterpreted as 11110xxx
bytes (as in a 4-byte sequence).  While unlikely, this can potentially cause
issues.

Fix is to explicitly reject such bytes in ngx_utf8_decode().

3 years agoWin32: non-ASCII names in ngx_fs_bsize(), ngx_fs_available().
Maxim Dounin [Thu, 23 Feb 2023 17:50:03 +0000 (20:50 +0300)]
Win32: non-ASCII names in ngx_fs_bsize(), ngx_fs_available().

This fixes potentially incorrect cache size calculations and non-working
"min_free" when using cache in directories with non-ASCII names.

3 years agoWin32: removed attempt to use a drive letter in ngx_fs_bsize().
Maxim Dounin [Thu, 23 Feb 2023 17:50:00 +0000 (20:50 +0300)]
Win32: removed attempt to use a drive letter in ngx_fs_bsize().

Just a drive letter might not correctly represent file system being used,
notably when using symlinks (as created by "mklink /d").  As such, instead
of trying to call GetDiskFreeSpace() with just a drive letter, we now always
use GetDiskFreeSpace() with full path.

Further, it looks like the code to use just a drive letter never worked,
since it tried to test name[2] instead of name[1] to be ':'.

3 years agoWin32: non-ASCII names support in ngx_open_tempfile().
Maxim Dounin [Thu, 23 Feb 2023 17:49:57 +0000 (20:49 +0300)]
Win32: non-ASCII names support in ngx_open_tempfile().

This makes it possible to use temporary directories with non-ASCII characters,
either explicitly or via a prefix with non-ASCII characters in it.

3 years agoWin32: non-ASCII names support in ngx_rename_file().
Maxim Dounin [Thu, 23 Feb 2023 17:49:55 +0000 (20:49 +0300)]
Win32: non-ASCII names support in ngx_rename_file().

This makes it possible to upload files with non-ASCII characters
when using the dav module (ticket #1433).

3 years agoWin32: non-ASCII names support in ngx_delete_file().
Maxim Dounin [Thu, 23 Feb 2023 17:49:54 +0000 (20:49 +0300)]
Win32: non-ASCII names support in ngx_delete_file().

This makes it possible to delete files with non-ASCII characters
when using the dav module (ticket #1433).

3 years agoWin32: reworked ngx_win32_rename_file() to use nginx wrappers.
Maxim Dounin [Thu, 23 Feb 2023 17:49:52 +0000 (20:49 +0300)]
Win32: reworked ngx_win32_rename_file() to use nginx wrappers.

This ensures that ngx_win32_rename_file() will support non-ASCII names
when supported by the wrappers.

Notably, this is used by PUT requests in the dav module when overwriting
existing files with non-ASCII names (ticket #1433).

3 years agoWin32: reworked ngx_win32_rename_file() to check errors.
Maxim Dounin [Thu, 23 Feb 2023 17:49:50 +0000 (20:49 +0300)]
Win32: reworked ngx_win32_rename_file() to check errors.

Previously, ngx_win32_rename_file() retried on all errors returned by
MoveFile() to a temporary name.  It only make sense, however, to retry
when the destination file already exists, similarly to the condition
when ngx_win32_rename_file() is called.  Retrying on other errors is
meaningless and might result in an infinite loop.

3 years agoWin32: non-ASCII directory names support in ngx_delete_dir().
Maxim Dounin [Thu, 23 Feb 2023 17:49:47 +0000 (20:49 +0300)]
Win32: non-ASCII directory names support in ngx_delete_dir().

This makes it possible to delete directories with non-ASCII characters
when using the dav module (ticket #1433).

3 years agoWin32: non-ASCII directory names support in ngx_create_dir().
Maxim Dounin [Thu, 23 Feb 2023 17:49:45 +0000 (20:49 +0300)]
Win32: non-ASCII directory names support in ngx_create_dir().

This makes it possible to create directories under prefix with non-ASCII
characters, as well as makes it possible to create directories with non-ASCII
characters when using the dav module (ticket #1433).

To ensure that the dav module operations are restricted similarly to
other file operations (in particular, short names are not allowed), the
ngx_win32_check_filename() function is used.  It improved to support
checking of just dirname, and now can be used to check paths when creating
files or directories.

3 years agoWin32: non-ASCII directory names support in ngx_getcwd().
Maxim Dounin [Thu, 23 Feb 2023 17:49:44 +0000 (20:49 +0300)]
Win32: non-ASCII directory names support in ngx_getcwd().

This makes it possible to start nginx without a prefix explicitly set
in a directory with non-ASCII characters in it.

3 years agoWin32: non-ASCII names support in "include" with wildcards.
Maxim Dounin [Thu, 23 Feb 2023 17:49:41 +0000 (20:49 +0300)]
Win32: non-ASCII names support in "include" with wildcards.

Notably, ngx_open_glob() now supports opening directories with non-ASCII
characters, and pathnames returned by ngx_read_glob() are converted to UTF-8.

3 years agoWin32: non-ASCII names support in autoindex (ticket #458).
Maxim Dounin [Thu, 23 Feb 2023 17:49:39 +0000 (20:49 +0300)]
Win32: non-ASCII names support in autoindex (ticket #458).

Notably, ngx_open_dir() now supports opening directories with non-ASCII
characters, and directory entries returned by ngx_read_dir() are properly
converted to UTF-8.

3 years agoWin32: OpenSSL compilation for x64 targets with MSVC.
Maxim Dounin [Thu, 23 Feb 2023 15:16:08 +0000 (18:16 +0300)]
Win32: OpenSSL compilation for x64 targets with MSVC.

To ensure proper target selection the NGX_MACHINE variable is now set
based on the MSVC compiler output, and the OpenSSL target is set based
on it.

This is not important as long as "no-asm" is used (as in misc/GNUmakefile
and win32 build instructions), but might be beneficial if someone is trying
to build OpenSSL with assembler code.

3 years agoWin32: i386 now assumed when crossbuilding (ticket #2416).
Maxim Dounin [Thu, 23 Feb 2023 15:15:59 +0000 (18:15 +0300)]
Win32: i386 now assumed when crossbuilding (ticket #2416).

Previously, NGX_MACHINE was not set when crossbuilding, resulting in
NGX_ALIGNMENT=16 being used in 32-bit builds (if not explicitly set to a
correct value).  This in turn might result in memory corruption in
ngx_palloc() (as there are no usable aligned allocator on Windows, and
normal malloc() is used instead, which provides 8 byte alignment on
32-bit platforms).

To fix this, now i386 machine is set when crossbuilding, so nginx won't
assume strict alignment requirements.