]> git.kaiwu.me - nginx.git/log
nginx.git
20 months agoSwitched GNUmakefile from hg to git.
Roman Arutyunyan [Thu, 29 Aug 2024 12:03:58 +0000 (16:03 +0400)]
Switched GNUmakefile from hg to git.

20 months agoRemoved .hgtags file.
Roman Arutyunyan [Thu, 29 Aug 2024 12:03:27 +0000 (16:03 +0400)]
Removed .hgtags file.

20 months agoStream: OCSP stapling. default
Sergey Kandaurov [Thu, 22 Aug 2024 10:57:46 +0000 (14:57 +0400)]
Stream: OCSP stapling.

20 months agoStream: client certificate validation with OCSP.
Sergey Kandaurov [Thu, 22 Aug 2024 10:57:45 +0000 (14:57 +0400)]
Stream: client certificate validation with OCSP.

20 months agoVersion bump.
Sergey Kandaurov [Tue, 20 Aug 2024 17:18:30 +0000 (21:18 +0400)]
Version bump.

20 months agorelease-1.27.1 tag
Sergey Kandaurov [Mon, 12 Aug 2024 14:21:01 +0000 (18:21 +0400)]
release-1.27.1 tag

20 months agonginx-1.27.1-RELEASE release-1.27.1
Sergey Kandaurov [Mon, 12 Aug 2024 14:20:52 +0000 (18:20 +0400)]
nginx-1.27.1-RELEASE

20 months agoUpdated OpenSSL used for win32 builds.
Sergey Kandaurov [Mon, 12 Aug 2024 14:20:49 +0000 (18:20 +0400)]
Updated OpenSSL used for win32 builds.

20 months agoMp4: rejecting unordered chunks in stsc atom.
Roman Arutyunyan [Mon, 12 Aug 2024 14:20:45 +0000 (18:20 +0400)]
Mp4: rejecting unordered chunks in stsc atom.

Unordered chunks could result in trak->end_chunk smaller than trak->start_chunk
in ngx_http_mp4_crop_stsc_data().  Later in ngx_http_mp4_update_stco_atom()
this caused buffer overread while trying to calculate trak->end_offset.

20 months agoMp4: fixed buffer underread while updating stsz atom.
Roman Arutyunyan [Mon, 12 Aug 2024 14:20:43 +0000 (18:20 +0400)]
Mp4: fixed buffer underread while updating stsz atom.

While cropping an stsc atom in ngx_http_mp4_crop_stsc_data(), a 32-bit integer
overflow could happen, which could result in incorrect seeking and a very large
value stored in "samples".  This resulted in a large invalid value of
trak->end_chunk_samples.  This value is further used to calculate the value of
trak->end_chunk_samples_size in ngx_http_mp4_update_stsz_atom().  While doing
this, a large invalid value of trak->end_chunk_samples could result in reading
memory before stsz atom start.  This could potentially result in a segfault.

21 months agoStream ssl_preread: do not reallocate a parsed SNI host.
Sergey Kandaurov [Fri, 9 Aug 2024 15:12:26 +0000 (19:12 +0400)]
Stream ssl_preread: do not reallocate a parsed SNI host.

We own this memory from the session pool.

21 months agoQUIC: zero out existing keying material only.
Sergey Kandaurov [Fri, 9 Aug 2024 15:12:26 +0000 (19:12 +0400)]
QUIC: zero out existing keying material only.

Previously, this used to have extra ngx_explicit_memzero() calls
from within ngx_quic_keys_cleanup(), which might be suboptimal.

21 months agoQUIC: discarding 0-RTT keys.
Sergey Kandaurov [Fri, 9 Aug 2024 15:12:25 +0000 (19:12 +0400)]
QUIC: discarding 0-RTT keys.

For simplicity, this is done on successful decryption of a 1-RTT packet.

21 months agoTypo fixed.
Sergey Kandaurov [Fri, 9 Aug 2024 15:12:23 +0000 (19:12 +0400)]
Typo fixed.

21 months agoVersion bump.
Sergey Kandaurov [Fri, 9 Aug 2024 14:01:42 +0000 (18:01 +0400)]
Version bump.

21 months agoHTTP/2: close connections initialized during graceful shutdown.
Kasei Wang [Thu, 18 Jul 2024 13:43:25 +0000 (17:43 +0400)]
HTTP/2: close connections initialized during graceful shutdown.

In some rare cases, graceful shutdown may happen while initializing an HTTP/2
connection.  Previously, such a connection ignored the shutdown and remained
active.  Now it is gracefully closed prior to processing any streams to
eliminate the shutdown delay.

22 months agoStream: allow servers with no handler.
Roman Arutyunyan [Thu, 27 Jun 2024 13:29:56 +0000 (17:29 +0400)]
Stream: allow servers with no handler.

Previously handlers were mandatory.  However they are not always needed.
For example, a server configured with ssl_reject_handshake does not need a
handler.  Such servers required a fake handler to pass the check.  Now handler
absence check is moved to runtime.  If handler is missing, the connection is
closed with 500 code.

23 months agorelease-1.27.0 tag
Sergey Kandaurov [Tue, 28 May 2024 13:22:30 +0000 (17:22 +0400)]
release-1.27.0 tag

23 months agonginx-1.27.0-RELEASE release-1.27.0
Sergey Kandaurov [Tue, 28 May 2024 13:19:38 +0000 (17:19 +0400)]
nginx-1.27.0-RELEASE

23 months agoHTTP/3: fixed handling of zero-length literal field line.
Sergey Kandaurov [Tue, 28 May 2024 13:20:45 +0000 (17:20 +0400)]
HTTP/3: fixed handling of zero-length literal field line.

Previously, st->value was passed with NULL data pointer to header handlers.

23 months agoQUIC: ngx_quic_buffer_t use-after-free protection.
Roman Arutyunyan [Tue, 28 May 2024 13:19:21 +0000 (17:19 +0400)]
QUIC: ngx_quic_buffer_t use-after-free protection.

Previously the last chain field of ngx_quic_buffer_t could still reference freed
chains and buffers after calling ngx_quic_free_buffer().  While normally an
ngx_quic_buffer_t object should not be used after freeing, resetting last_chain
field would prevent a potential use-after-free.

23 months agoQUIC: ignore CRYPTO frames after handshake completion.
Roman Arutyunyan [Tue, 28 May 2024 13:19:08 +0000 (17:19 +0400)]
QUIC: ignore CRYPTO frames after handshake completion.

Sending handshake-level CRYPTO frames after the client's Finished message could
lead to memory disclosure and a potential segfault, if those frames are sent in
one packet with the Finished frame.

23 months agoHTTP/3: fixed dynamic table overflow.
Roman Arutyunyan [Tue, 28 May 2024 13:18:50 +0000 (17:18 +0400)]
HTTP/3: fixed dynamic table overflow.

While inserting a new entry into the dynamic table, first the entry is added,
and then older entries are evicted until table size is within capacity.  After
the first step, the number of entries may temporarily exceed the maximum
calculated from capacity by one entry, which previously caused table overflow.

The easiest way to trigger the issue is to keep adding entries with empty names
and values until first eviction.

The issue was introduced by 987bee4363d1.

23 months agoHTTP/3: decoder stream pre-creation.
Roman Arutyunyan [Tue, 28 May 2024 13:18:28 +0000 (17:18 +0400)]
HTTP/3: decoder stream pre-creation.

Previously a decoder stream was created on demand for sending Section
Acknowledgement, Stream Cancellation and Insert Count Increment.  If conditions
for sending any of these instructions never happen, a decoder stream is not
created at all.  These conditions include client not using the dynamic table and
no streams abandoned by server (RFC 9204, Section 2.2.2.2).  However RFC 9204,
Section 4.2 defines only one condition for not creating a decoder stream:

   An endpoint MAY avoid creating a decoder stream if its decoder sets
   the maximum capacity of the dynamic table to zero.

The change enables pre-creation of the decoder stream at HTTP/3 session
initialization if maximum dynamic table capacity is not zero.  Note that this
value is currently hardcoded to 4096 bytes and is not configurable, so the
stream is now always created.

Also, the change fixes a potential stack overflow when creating a decoder
stream in ngx_http_v3_send_cancel_stream() while draining a request stream by
ngx_drain_connections().  Creating a decoder stream involves calling
ngx_get_connection(), which calls ngx_drain_connections(), which will drain the
same request stream again.  If client's MAX_STREAMS for uni stream is high
enough, these recursive calls will continue until we run out of stack.
Otherwise, decoder stream creation will fail at some point and the request
stream connection will be drained.  This may result in use-after-free, since
this connection could still be referenced up the stack.

23 months agoQUIC: client transport parameter data length checking.
Sergey Kandaurov [Tue, 28 May 2024 13:17:19 +0000 (17:17 +0400)]
QUIC: client transport parameter data length checking.

2 years agoUpstream: variables support in proxy_limit_rate and friends.
J Carter [Sat, 25 Nov 2023 21:57:09 +0000 (21:57 +0000)]
Upstream: variables support in proxy_limit_rate and friends.

23 months agoOptimized chain link usage (ticket #2614).
Roman Arutyunyan [Thu, 23 May 2024 15:15:38 +0000 (19:15 +0400)]
Optimized chain link usage (ticket #2614).

Previously chain links could sometimes be dropped instead of being reused,
which could result in increased memory consumption during long requests.

A similar chain link issue in ngx_http_gzip_filter_module was fixed in
da46bfc484ef (1.11.10).

Based on a patch by Sangmin Lee.

23 months agoConfigure: fixed building libatomic test.
Edgar Bonet [Thu, 16 May 2024 09:15:10 +0000 (11:15 +0200)]
Configure: fixed building libatomic test.

Using "long *" instead of "AO_t *" leads either to -Wincompatible-pointer-types
or -Wpointer-sign warnings, depending on whether long and size_t are compatible
types (e.g., ILP32 versus LP64 data models).  Notably, -Wpointer-sign warnings
are enabled by default in Clang only, and -Wincompatible-pointer-types is an
error starting from GCC 14.

Signed-off-by: Edgar Bonet <bonet@grenoble.cnrs.fr>
2 years agoStream pass: disabled passing from or to udp.
Roman Arutyunyan [Fri, 3 May 2024 16:26:05 +0000 (20:26 +0400)]
Stream pass: disabled passing from or to udp.

Passing from udp was not possible for the most part due to preread buffer
restriction.  Passing to udp could occasionally work, but the connection would
still be bound to the original listen rbtree, which prevented it from being
deleted on connection closure.

2 years agoSSL: fixed possible configuration overwrite loading "engine:" keys.
Sergey Kandaurov [Fri, 3 May 2024 16:29:01 +0000 (20:29 +0400)]
SSL: fixed possible configuration overwrite loading "engine:" keys.

When loading certificate keys via ENGINE_load_private_key() in runtime,
it was possible to overwrite configuration on ENGINE_by_id() failure.
OpenSSL documention doesn't describe errors in details, the only reason
I found in the comment to example is when the engine is not available.

2 years agoHTTP/3: fixed handling of malformed request body length.
Sergey Kandaurov [Fri, 3 May 2024 16:28:32 +0000 (20:28 +0400)]
HTTP/3: fixed handling of malformed request body length.

Previously, a request body larger than declared in Content-Length resulted in
a 413 status code, because Content-Length was mistakenly used as the maximum
allowed request body, similar to client_max_body_size.  Following the HTTP/3
specification, such requests are now rejected with the 400 error as malformed.

2 years agoVersion bump.
Sergey Kandaurov [Fri, 3 May 2024 16:28:22 +0000 (20:28 +0400)]
Version bump.

2 years agorelease-1.25.5 tag
Roman Arutyunyan [Tue, 16 Apr 2024 14:29:59 +0000 (18:29 +0400)]
release-1.25.5 tag

2 years agonginx-1.25.5-RELEASE release-1.25.5
Roman Arutyunyan [Tue, 16 Apr 2024 14:27:50 +0000 (18:27 +0400)]
nginx-1.25.5-RELEASE

2 years agoStream pass: limited the number of passes per connection.
Roman Arutyunyan [Thu, 11 Apr 2024 07:37:30 +0000 (11:37 +0400)]
Stream pass: limited the number of passes per connection.

Previously a cycle in pass configuration resulted in stack overflow.

2 years agoQUIC: fixed close timer processing with early data.
Vladimir Khomutov [Wed, 10 Apr 2024 06:38:10 +0000 (09:38 +0300)]
QUIC: fixed close timer processing with early data.

The ngx_quic_run() function uses qc->close timer to limit the handshake
duration.  Normally it is removed by ngx_quic_do_init_streams() which is
called once when we are done with initial SSL processing.

The problem happens when the client sends early data and streams are
initialized in the ngx_quic_run() -> ngx_quic_handle_datagram() call.
The order of set/remove timer calls is now reversed; the close timer is
set up and the timer fires when assigned, starting the unexpected connection
close process.

The fix is to skip setting the timer if streams were initialized during
handling of the initial datagram.  The idle timer for quic is set anyway,
and stream-related timeouts are managed by application layer.

2 years agoConfigure: allow cross-compiling to Windows using Clang.
Piotr Sikora [Mon, 26 Feb 2024 20:00:48 +0000 (20:00 +0000)]
Configure: allow cross-compiling to Windows using Clang.

Signed-off-by: Piotr Sikora <piotr@aviatrix.com>
2 years agoConfigure: fixed "make install" when cross-compiling to Windows.
Piotr Sikora [Mon, 26 Feb 2024 20:00:46 +0000 (20:00 +0000)]
Configure: fixed "make install" when cross-compiling to Windows.

Signed-off-by: Piotr Sikora <piotr@aviatrix.com>
2 years agoConfigure: added support for Homebrew on Apple Silicon.
Piotr Sikora [Mon, 26 Feb 2024 20:00:43 +0000 (20:00 +0000)]
Configure: added support for Homebrew on Apple Silicon.

Signed-off-by: Piotr Sikora <piotr@aviatrix.com>
2 years agoConfigure: set cache line size for more architectures.
Sergey Kandaurov [Wed, 27 Mar 2024 15:36:51 +0000 (19:36 +0400)]
Configure: set cache line size for more architectures.

Based on a patch by Piotr Sikora.

2 years agoDetect cache line size at runtime on macOS.
Piotr Sikora [Mon, 26 Feb 2024 20:00:40 +0000 (20:00 +0000)]
Detect cache line size at runtime on macOS.

Notably, Apple Silicon CPUs have 128 byte cache line size,
which is twice the default configured for generic aarch64.

Signed-off-by: Piotr Sikora <piotr@aviatrix.com>
2 years agoConfigure: fixed Linux crypt_r() test to add libcrypt.
Sergey Kandaurov [Mon, 26 Feb 2024 20:00:38 +0000 (20:00 +0000)]
Configure: fixed Linux crypt_r() test to add libcrypt.

Previously, the resulting binary was successfully linked
because libcrypt was added in a separate test for crypt().

Patch by Piotr Sikora.

2 years agoWin32: fixed unique file index calculations.
Piotr Sikora [Mon, 26 Feb 2024 20:00:35 +0000 (20:00 +0000)]
Win32: fixed unique file index calculations.

The old code was breaking strict aliasing rules.

Signed-off-by: Piotr Sikora <piotr@aviatrix.com>
2 years agoRewrite: fixed "return" directive without response text.
Piotr Sikora [Mon, 26 Feb 2024 20:00:28 +0000 (20:00 +0000)]
Rewrite: fixed "return" directive without response text.

Previously, the response text wasn't initialized and the rewrite module
was sending response body set to NULL.

Found with UndefinedBehaviorSanitizer (pointer-overflow).

Signed-off-by: Piotr Sikora <piotr@aviatrix.com>
2 years agoFixed undefined behaviour with IPv4-mapped IPv6 addresses.
Sergey Kandaurov [Mon, 18 Mar 2024 13:14:30 +0000 (17:14 +0400)]
Fixed undefined behaviour with IPv4-mapped IPv6 addresses.

Previously, it could result when left-shifting signed integer due to implicit
integer promotion, such that the most significant bit appeared on the sign bit.

In practice, though, this results in the same left value as with an explicit
cast, at least on known compilers, such as GCC and Clang.  The reason is that
in_addr_t, which is equivalent to uint32_t and same as "unsigned int" in ILP32
and LP64 data type models, has the same type width as the intermediate after
integer promotion, so there's no side effects such as sign-extension.  This
explains why adding an explicit cast does not change object files in practice.

Found with UndefinedBehaviorSanitizer (shift).

Based on a patch by Piotr Sikora.

2 years agoGeo: fixed uninitialized memory access.
Piotr Sikora [Thu, 14 Mar 2024 14:37:20 +0000 (18:37 +0400)]
Geo: fixed uninitialized memory access.

While copying ngx_http_variable_value_t structures to geo binary base
in ngx_http_geo_copy_values(), and similarly in the stream module,
uninitialized parts of these structures are copied as well.  These
include the "escape" field and possible holes.  Calculating crc32 of
this data triggers uninitialized memory access.

Found with MemorySanitizer.

Signed-off-by: Piotr Sikora <piotr@aviatrix.com>
2 years agoStream: $server_name.
Sergey Kandaurov [Fri, 22 Mar 2024 14:39:50 +0000 (18:39 +0400)]
Stream: $server_name.

2 years agoStream: moved fastopen compatibility check.
Roman Arutyunyan [Thu, 18 Jan 2024 15:12:38 +0000 (19:12 +0400)]
Stream: moved fastopen compatibility check.

The move makes the code look similar to the corresponding code in http module.

2 years agoStream: the "setfib" parameter of the "listen" directive.
Sergey Kandaurov [Fri, 22 Mar 2024 10:53:19 +0000 (14:53 +0400)]
Stream: the "setfib" parameter of the "listen" directive.

The FreeBSD SO_SETFIB support.

2 years agoStream: the "accept_filter" parameter of the "listen" directive.
Sergey Kandaurov [Fri, 22 Mar 2024 10:53:19 +0000 (14:53 +0400)]
Stream: the "accept_filter" parameter of the "listen" directive.

The FreeBSD accept filters support.

2 years agoStream: the "deferred" parameter of the "listen" directive.
Sergey Kandaurov [Fri, 22 Mar 2024 10:53:19 +0000 (14:53 +0400)]
Stream: the "deferred" parameter of the "listen" directive.

The Linux TCP_DEFER_ACCEPT support.

2 years agoStream: reshuffled ngx_stream_listen_opt_t fields.
Sergey Kandaurov [Fri, 22 Mar 2024 10:53:19 +0000 (14:53 +0400)]
Stream: reshuffled ngx_stream_listen_opt_t fields.

In preparation for adding more parameters to the listen directive,
and to be in sync with the corresponding structure in the http module.
No functional changes.

2 years agoOverhauled some diagnostic messages akin to 1b05b9bbcebf.
Sergey Kandaurov [Fri, 22 Mar 2024 10:51:14 +0000 (14:51 +0400)]
Overhauled some diagnostic messages akin to 1b05b9bbcebf.

2 years agoStream: using ngx_stream_ssl_srv_conf_t *sscf naming convention.
Sergey Kandaurov [Fri, 22 Mar 2024 10:18:51 +0000 (14:18 +0400)]
Stream: using ngx_stream_ssl_srv_conf_t *sscf naming convention.

Originally, the stream module was developed based on the mail module,
following the existing style.  Then it was diverged to closely follow
the http module development.  This change updates style to use sscf
naming convention troughout the stream module, which matches the http
module code style.  No functional changes.

2 years agoStream: ngx_stream_pass_module.
Roman Arutyunyan [Wed, 21 Feb 2024 13:36:02 +0000 (17:36 +0400)]
Stream: ngx_stream_pass_module.

The module allows to pass connections from Stream to other modules such as HTTP
or Mail, as well as back to Stream.  Previously, this was only possible with
proxying.  Connections with preread buffer read out from socket cannot be
passed.

The module allows selective SSL termination based on SNI.

    stream {
        server {
            listen 8000 default_server;
            ssl_preread on;
            ...
        }

        server {
            listen 8000;
            server_name foo.example.com;
            pass 127.0.0.1:8001; # to HTTP
        }

        server {
            listen 8000;
            server_name bar.example.com;
            ...
        }
    }

    http {
        server {
            listen 8001 ssl;
            ...

            location / {
                root html;
            }
        }
    }

2 years agoStream: virtual servers.
Roman Arutyunyan [Thu, 14 Dec 2023 17:58:39 +0000 (21:58 +0400)]
Stream: virtual servers.

Server name is taken either from ngx_stream_ssl_module or
ngx_stream_ssl_preread_module.

The change adds "default_server" parameter to the "listen" directive,
as well as the following directives: "server_names_hash_max_size",
"server_names_hash_bucket_size", "server_name" and "ssl_reject_handshake".

2 years agoStream: socket peek in preread phase.
Roman Arutyunyan [Wed, 13 Dec 2023 14:04:55 +0000 (18:04 +0400)]
Stream: socket peek in preread phase.

Previously, preread buffer was always read out from socket, which made it
impossible to terminate SSL on the connection without introducing additional
SSL BIOs.  The following patches will rely on this.

Now, when possible, recv(MSG_PEEK) is used instead, which keeps data in socket.
It's called if SSL is not already terminated and if an egde-triggered event
method is used.  For epoll, EPOLLRDHUP support is also required.

2 years agoVersion bump.
Roman Arutyunyan [Thu, 21 Mar 2024 13:06:21 +0000 (17:06 +0400)]
Version bump.

2 years agorelease-1.25.4 tag
Sergey Kandaurov [Wed, 14 Feb 2024 16:03:00 +0000 (20:03 +0400)]
release-1.25.4 tag

2 years agonginx-1.25.4-RELEASE release-1.25.4
Sergey Kandaurov [Wed, 14 Feb 2024 11:55:46 +0000 (15:55 +0400)]
nginx-1.25.4-RELEASE

2 years agoUpdated OpenSSL and zlib used for win32 builds.
Sergey Kandaurov [Wed, 14 Feb 2024 11:55:42 +0000 (15:55 +0400)]
Updated OpenSSL and zlib used for win32 builds.

2 years agoQUIC: fixed stream cleanup (ticket #2586).
Roman Arutyunyan [Wed, 14 Feb 2024 11:55:37 +0000 (15:55 +0400)]
QUIC: fixed stream cleanup (ticket #2586).

Stream connection cleanup handler ngx_quic_stream_cleanup_handler() calls
ngx_quic_shutdown_stream() after which it resets the pointer from quic stream
to the connection (sc->connection = NULL).  Previously if this call failed,
sc->connection retained the old value, while the connection was freed by the
application code.  This resulted later in a second attempt to close the freed
connection, which lead to allocator double free error.

The fix is to reset the sc->connection pointer in case of error.

2 years agoQUIC: trial packet decryption in response to invalid key update.
Sergey Kandaurov [Wed, 14 Feb 2024 11:55:34 +0000 (15:55 +0400)]
QUIC: trial packet decryption in response to invalid key update.

Inspired by RFC 9001, Section 6.3, trial packet decryption with the current
keys is now used to avoid a timing side-channel signal.  Further, this fixes
segfault while accessing missing next keys (ticket #2585).

2 years agoQUIC: fixed unsent MTU probe acknowledgement.
Roman Arutyunyan [Wed, 14 Feb 2024 12:56:28 +0000 (16:56 +0400)]
QUIC: fixed unsent MTU probe acknowledgement.

Previously if an MTU probe send failed early in ngx_quic_frame_sendto()
due to allocation error or congestion control, the application level packet
number was not increased, but was still saved as MTU probe packet number.
Later when a packet with this number was acknowledged, the unsent MTU probe
was acknowledged as well.  This could result in discovering a bigger MTU than
supported by the path, which could lead to EMSGSIZE (Message too long) errors
while sending further packets.

The problem existed since PMTUD was introduced in 58afcd72446f (1.25.2).
Back then only the unlikely memory allocation error could trigger it.  However
in efcdaa66df2e congestion control was added to ngx_quic_frame_sendto() which
can now trigger the issue with a higher probability.

2 years agoHTTP/3: added more compatibility checks for "listen ... quic".
Sergey Kandaurov [Tue, 30 Jan 2024 15:19:26 +0000 (19:19 +0400)]
HTTP/3: added more compatibility checks for "listen ... quic".

Now "fastopen", "backlog", "accept_filter", "deferred", and "so_keepalive"
parameters are not allowed with "quic" in the "listen" directive.

Reported by Izorkin.

2 years agoSSL: fixed $ssl_curves allocation error handling.
Sergey Kandaurov [Tue, 30 Jan 2024 15:18:31 +0000 (19:18 +0400)]
SSL: fixed $ssl_curves allocation error handling.

2 years agoYear 2024.
Sergey Kandaurov [Tue, 30 Jan 2024 15:14:16 +0000 (19:14 +0400)]
Year 2024.

2 years agoUpstream: fixed usage of closed sockets with filter finalization.
Maxim Dounin [Tue, 30 Jan 2024 00:20:10 +0000 (03:20 +0300)]
Upstream: fixed usage of closed sockets with filter finalization.

When filter finalization is triggered when working with an upstream server,
and error_page redirects request processing to some simple handler,
ngx_http_request_finalize() triggers request termination when the response
is sent.  In particular, via the upstream cleanup handler, nginx will close
the upstream connection and the corresponding socket.

Still, this can happen to be with ngx_event_pipe() on stack.  While
the code will set p->downstream_error due to NGX_ERROR returned from the
output filter chain by filter finalization, otherwise the error will be
ignored till control returns to ngx_http_upstream_process_request().
And event pipe might try reading from the (already closed) socket, resulting
in "readv() failed (9: Bad file descriptor) while reading upstream" errors
(or even segfaults with SSL).

Such errors were seen with the following configuration:

    location /t2 {
        proxy_pass http://127.0.0.1:8080/big;

        image_filter_buffer 10m;
        image_filter   resize  150 100;
        error_page     415   = /empty;
    }

    location /empty {
        return 204;
    }

    location /big {
        # big enough static file
    }

Fix is to clear p->upstream in ngx_http_upstream_finalize_request(),
and ensure that p->upstream is checked in ngx_event_pipe_read_upstream()
and when handling events at ngx_event_pipe() exit.

2 years agoFixed request termination with AIO and subrequests (ticket #2555).
Maxim Dounin [Tue, 30 Jan 2024 00:20:05 +0000 (03:20 +0300)]
Fixed request termination with AIO and subrequests (ticket #2555).

When a request was terminated due to an error via ngx_http_terminate_request()
while an AIO operation was running in a subrequest, various issues were
observed.  This happened because ngx_http_request_finalizer() was only set
in the subrequest where ngx_http_terminate_request() was called, but not
in the subrequest where the AIO operation was running.  After completion
of the AIO operation normal processing of the subrequest was resumed, leading
to issues.

In particular, in case of the upstream module, termination of the request
called upstream cleanup, which closed the upstream connection.  Attempts to
further work with the upstream connection after AIO operation completion
resulted in segfaults in ngx_ssl_recv(), "readv() failed (9: Bad file
descriptor) while reading upstream" errors, or socket leaks.

In ticket #2555, issues were observed with the following configuration
with cache background update (with thread writing instrumented to
introduce a delay, when a client closes the connection during an update):

    location = /background-and-aio-write {
        proxy_pass ...
        proxy_cache one;
        proxy_cache_valid 200 1s;
        proxy_cache_background_update on;
        proxy_cache_use_stale updating;
        aio threads;
        aio_write on;
        limit_rate 1000;
    }

Similarly, the same issue can be seen with SSI, and can be caused by
errors in subrequests, such as in the following configuration
(where "/proxy" uses AIO, and "/sleep" returns 444 after some delay,
causing request termination):

    location = /ssi-active-boom {
        ssi on;
        ssi_types *;
        return 200 '
                   <!--#include virtual="/proxy" -->
                   <!--#include virtual="/sleep" -->
                   ';
        limit_rate 1000;
    }

Or the same with both AIO operation and the error in non-active subrequests
(which needs slightly different handling, see below):

    location = /ssi-non-active-boom {
        ssi on;
        ssi_types *;
        return 200 '
                   <!--#include virtual="/static" -->
                   <!--#include virtual="/proxy" -->
                   <!--#include virtual="/sleep" -->
                   ';
        limit_rate 1000;
    }

Similarly, issues can be observed with just static files.  However,
with static files potential impact is limited due to timeout safeguards
in ngx_http_writer(), and the fact that c->error is set during request
termination.

In a simple configuration with an AIO operation in the active subrequest,
such as in the following configuration, the connection is closed right
after completion of the AIO operation anyway, since ngx_http_writer()
tries to write to the connection and fails due to c->error set:

    location = /ssi-active-static-boom {
        ssi on;
        ssi_types *;
        return 200 '
                   <!--#include virtual="/static-aio" -->
                   <!--#include virtual="/sleep" -->
                   ';
        limit_rate 1000;
    }

In the following configuration, with an AIO operation in a non-active
subrequest, the connection is closed only after send_timeout expires:

    location = /ssi-non-active-static-boom {
        ssi on;
        ssi_types *;
        return 200 '
                   <!--#include virtual="/static" -->
                   <!--#include virtual="/static-aio" -->
                   <!--#include virtual="/sleep" -->
                   ';
        limit_rate 1000;
    }

Fix is to introduce r->main->terminated flag, which is to be checked
by AIO event handlers when the r->main->blocked counter is decremented.
When the flag is set, handlers are expected to wake up the connection
instead of the subrequest (which might be already cleaned up).

Additionally, now ngx_http_request_finalizer() is always set in the
active subrequest, so waking up the connection properly finalizes the
request even if termination happened in a non-active subrequest.

2 years agoAIO operations now add timers (ticket #2162).
Maxim Dounin [Mon, 29 Jan 2024 07:31:37 +0000 (10:31 +0300)]
AIO operations now add timers (ticket #2162).

Each AIO (thread IO) operation being run is now accompanied with 1-minute
timer.  This timer prevents unexpected shutdown of the worker process while
an AIO operation is running, and logs an alert if the operation is running
for too long.

This fixes "open socket left" alerts during worker processes shutdown
due to pending AIO (or thread IO) operations while corresponding requests
have no timers.  In particular, such errors were observed while reading
cache headers (ticket #2162), and with worker_shutdown_timeout.

2 years agoSilenced complaints about socket leaks on forced termination.
Maxim Dounin [Mon, 29 Jan 2024 07:29:39 +0000 (10:29 +0300)]
Silenced complaints about socket leaks on forced termination.

When graceful shutdown was requested, and then nginx was forced to
do fast shutdown, it used to (incorrectly) complain about open sockets
left in connections which weren't yet closed when fast shutdown
was requested.

Fix is to avoid complaining about open sockets when fast shutdown was
requested after graceful one.  Abnormal termination, if requested with
the WINCH signal, can still happen though.

2 years agoSSL: reasonable version for LibreSSL adjusted.
Sergey Kandaurov [Mon, 25 Dec 2023 17:15:48 +0000 (21:15 +0400)]
SSL: reasonable version for LibreSSL adjusted.

OPENSSL_VERSION_NUMBER is now redefined to 0x1010000fL for LibreSSL 3.5.0
and above.  Building with older LibreSSL versions, such as 2.8.0, may now
produce warnings (see cab37803ebb3) and may require appropriate compiler
options to suppress them.

Notably, this allows to start using SSL_get0_verified_chain() appeared
in OpenSSL 1.1.0 and LibreSSL 3.5.0, without additional macro tests.

Prodded by Ilya Shipitsin.

2 years agoSSL: disabled renegotiation checks with LibreSSL.
Sergey Kandaurov [Mon, 25 Dec 2023 17:15:47 +0000 (21:15 +0400)]
SSL: disabled renegotiation checks with LibreSSL.

Similar to 7356:e3ba4026c02d, as long as SSL_OP_NO_CLIENT_RENEGOTIATION
is defined, it is the library responsibility to prevent renegotiation.

Additionally, this allows to raise LibreSSL version used to redefine
OPENSSL_VERSION_NUMBER to 0x1010000fL, such that this won't result in
attempts to dereference SSL objects made opaque in LibreSSL 3.4.0.

Patch by Maxim Dounin.

2 years agoWin32: extended ngx_random() range to 0x7fffffff.
J Carter [Sat, 9 Dec 2023 08:38:14 +0000 (08:38 +0000)]
Win32: extended ngx_random() range to 0x7fffffff.

rand() is used on win32. RAND_MAX is implementation defined. win32's is
0x7fff.

Existing uses of ngx_random() rely upon 0x7fffffff range provided by
POSIX implementations of random().

2 years agoQUIC: fixed format specifier after a6f79f044de5.
Sergey Kandaurov [Fri, 15 Dec 2023 23:40:01 +0000 (03:40 +0400)]
QUIC: fixed format specifier after a6f79f044de5.

2 years agoQUIC: path aware in-flight bytes accounting.
Sergey Kandaurov [Tue, 12 Dec 2023 16:21:12 +0000 (20:21 +0400)]
QUIC: path aware in-flight bytes accounting.

On-packet acknowledgement is made path aware, as per RFC 9000, Section 9.4:
    Packets sent on the old path MUST NOT contribute to congestion control
    or RTT estimation for the new path.

To make this possible in a single congestion control context, the first packet
to be sent after the new path has been validated, which includes resetting the
congestion controller and RTT estimator, is now remembered in the connection.
Packets sent previously, such as on the old path, are not taken into account.

Note that although the packet number is saved per-connection, the added checks
affect application level packets only.  For non-application level packets,
which are only processed prior to the handshake is complete, the remembered
packet number remains set to zero.

2 years agoQUIC: reset RTT estimator for the new path.
Sergey Kandaurov [Tue, 12 Dec 2023 16:20:51 +0000 (20:20 +0400)]
QUIC: reset RTT estimator for the new path.

RTT is a property of the path, it must be reset on confirming a peer's
ownership of its new address.

2 years agoQUIC: path revalidation after expansion failure.
Roman Arutyunyan [Wed, 29 Nov 2023 06:58:21 +0000 (10:58 +0400)]
QUIC: path revalidation after expansion failure.

As per RFC 9000, Section 8.2.1:

    When an endpoint is unable to expand the datagram size to 1200 bytes due
    to the anti-amplification limit, the path MTU will not be validated.
    To ensure that the path MTU is large enough, the endpoint MUST perform a
    second path validation by sending a PATH_CHALLENGE frame in a datagram of
    at least 1200 bytes.

2 years agoQUIC: ngx_quic_frame_t time fields cleanup.
Roman Arutyunyan [Thu, 30 Nov 2023 11:03:06 +0000 (15:03 +0400)]
QUIC: ngx_quic_frame_t time fields cleanup.

The field "first" is removed.  It's unused since 909b989ec088.
The field "last" is renamed to "send_time".  It holds frame send time.

2 years agoQUIC: congestion control in ngx_quic_frame_sendto().
Roman Arutyunyan [Wed, 29 Nov 2023 17:41:29 +0000 (21:41 +0400)]
QUIC: congestion control in ngx_quic_frame_sendto().

Previously ngx_quic_frame_sendto() ignored congestion control and did not
contribute to in_flight counter.

Now congestion control window is checked unless ignore_congestion flag is set.
Also, in_flight counter is incremented and the frame is stored in ctx->sent
queue if it's ack-eliciting.  This behavior is now similar to
ngx_quic_output_packet().

2 years agoQUIC: ignore duplicate PATH_CHALLENGE frames.
Roman Arutyunyan [Wed, 22 Nov 2023 10:48:12 +0000 (14:48 +0400)]
QUIC: ignore duplicate PATH_CHALLENGE frames.

According to RFC 9000, an endpoint SHOULD NOT send multiple PATH_CHALLENGE
frames in a single packet.  The change adds a check to enforce this claim to
optimize server behavior.  Previously each PATH_CHALLENGE always resulted in a
single response datagram being sent to client.  The effect of this was however
limited by QUIC flood protection.

Also, PATH_CHALLENGE is explicitly disabled in Initial and Handshake levels,
see RFC 9000, Table 3.  However, technically it may be sent by client in 0-RTT
over a new path without actual migration, even though the migration itself is
prohibited during handshake.  This allows client to coalesce multiple 0-RTT
packets each carrying a PATH_CHALLENGE and end up with multiple PATH_CHALLENGEs
per datagram.  This again leads to suboptimal behavior, see above.  Since the
purpose of sending PATH_CHALLENGE frames in 0-RTT is unclear, these frames are
now only allowed in 1-RTT.  For 0-RTT they are silently ignored.

2 years agoQUIC: fixed anti-amplification with explicit send.
Roman Arutyunyan [Wed, 22 Nov 2023 10:52:21 +0000 (14:52 +0400)]
QUIC: fixed anti-amplification with explicit send.

Previously, when using ngx_quic_frame_sendto() to explicitly send a packet with
a single frame, anti-amplification limit was not properly enforced.  Even when
there was no quota left for the packet, it was sent anyway, but with no padding.
Now the packet is not sent at all.

This function is called to send PATH_CHALLENGE/PATH_RESPONSE, PMTUD and probe
packets.  For all these cases packet send is retried later in case the send was
not successful.

2 years agoQUIC: avoid partial expansion of PATH_CHALLENGE/PATH_RESPONSE.
Roman Arutyunyan [Wed, 29 Nov 2023 14:13:25 +0000 (18:13 +0400)]
QUIC: avoid partial expansion of PATH_CHALLENGE/PATH_RESPONSE.

By default packets with these frames are expanded to 1200 bytes.  Previously,
if anti-amplification limit did not allow this expansion, it was limited to
whatever size was allowed.  However RFC 9000 clearly states no partial
expansion should happen in both cases.

    Section 8.2.1.  Initiating Path Validation:

    An endpoint MUST expand datagrams that contain a PATH_CHALLENGE frame
    to at least the smallest allowed maximum datagram size of 1200 bytes,
    unless the anti-amplification limit for the path does not permit
    sending a datagram of this size.

    Section 8.2.2. Path Validation Responses:

    An endpoint MUST expand datagrams that contain a PATH_RESPONSE frame
    to at least the smallest allowed maximum datagram size of 1200 bytes.
    ...
    However, an endpoint MUST NOT expand the datagram containing the
    PATH_RESPONSE if the resulting data exceeds the anti-amplification limit.

2 years agoHTTP: uniform checks in ngx_http_alloc_large_header_buffer().
Vladimir Khomutov [Wed, 29 Nov 2023 08:13:05 +0000 (11:13 +0300)]
HTTP: uniform checks in ngx_http_alloc_large_header_buffer().

If URI is not fully parsed yet, some pointers are not set.  As a result,
the calculation of "new + (ptr - old)" expression is flawed.

According to C11, 6.5.6 Additive operators, p.9:

: When two pointers are subtracted, both shall point to elements
: of the same array object, or one past the last element of the
: array object

Since "ptr" is not set, subtraction leads to undefined behaviour, because
"ptr" and "old" are not in the same buffer (i.e. array objects).

Prodded by GCC undefined behaviour sanitizer.

2 years agoHTTP: removed unused r->port_start and r->port_end.
Vladimir Khomutov [Tue, 28 Nov 2023 09:57:14 +0000 (12:57 +0300)]
HTTP: removed unused r->port_start and r->port_end.

Neither r->port_start nor r->port_end were ever used.

The r->port_end is set by the parser, though it was never used by
the following code (and was never usable, since not copied by the
ngx_http_alloc_large_header_buffer() without r->port_start set).

2 years agoHTTP/3: added Huffman decoding error logging.
Sergey Kandaurov [Tue, 14 Nov 2023 11:26:02 +0000 (15:26 +0400)]
HTTP/3: added Huffman decoding error logging.

2 years agoAdjusted Huffman coding debug logging, missed in 7977:336084ff943b.
Sergey Kandaurov [Tue, 14 Nov 2023 10:50:03 +0000 (14:50 +0400)]
Adjusted Huffman coding debug logging, missed in 7977:336084ff943b.

Spotted by XingY Wang.

2 years agoQUIC: improved packet and frames debug tracing.
Vladimir Khomutov [Thu, 26 Oct 2023 20:35:09 +0000 (23:35 +0300)]
QUIC: improved packet and frames debug tracing.

Currently, packets generated by ngx_quic_frame_sendto() and
ngx_quic_send_early_cc() are not logged, thus making it hard
to read logs due to gaps appearing in packet numbers sequence.

At frames level, it is handy to see immediately packet number
in which they arrived or being sent.

2 years agoVersion bump.
Sergey Kandaurov [Thu, 26 Oct 2023 21:29:28 +0000 (01:29 +0400)]
Version bump.

2 years agorelease-1.25.3 tag
Maxim Dounin [Tue, 24 Oct 2023 13:46:47 +0000 (16:46 +0300)]
release-1.25.3 tag

2 years agonginx-1.25.3-RELEASE release-1.25.3
Maxim Dounin [Tue, 24 Oct 2023 13:46:46 +0000 (16:46 +0300)]
nginx-1.25.3-RELEASE

2 years agoUpdated OpenSSL and zlib used for win32 builds.
Maxim Dounin [Mon, 23 Oct 2023 18:50:26 +0000 (21:50 +0300)]
Updated OpenSSL and zlib used for win32 builds.

2 years agoHTTP/2: fixed buffer management with HTTP/2 auto-detection.
Sergey Kandaurov [Sat, 21 Oct 2023 14:48:24 +0000 (18:48 +0400)]
HTTP/2: fixed buffer management with HTTP/2 auto-detection.

As part of normal HTTP/2 processing, incomplete frames are saved in the
control state using a fixed size memcpy of NGX_HTTP_V2_STATE_BUFFER_SIZE.
For this matter, two state buffers are reserved in the HTTP/2 recv buffer.

As part of HTTP/2 auto-detection on plain TCP connections, initial data
is first read into a buffer specified by the client_header_buffer_size
directive that doesn't have state reservation.  Previously, this made it
possible to over-read the buffer as part of saving the state.

The fix is to read the available buffer size rather than a fixed size.
Although memcpy of a fixed size can produce a better optimized code,
handling of incomplete frames isn't a common execution path, so it was
sacrificed for the sake of simplicity of the fix.

2 years agoQUIC: explicitly zero out unused keying material.
Sergey Kandaurov [Fri, 20 Oct 2023 14:05:07 +0000 (18:05 +0400)]
QUIC: explicitly zero out unused keying material.

2 years agoQUIC: removed key field from ngx_quic_secret_t.
Sergey Kandaurov [Fri, 20 Oct 2023 14:05:07 +0000 (18:05 +0400)]
QUIC: removed key field from ngx_quic_secret_t.

It is made local as it is only needed now when creating crypto context.

BoringSSL lacks EVP interface for ChaCha20, providing instead
a function for one-shot encryption, thus hp is still preserved.

Based on a patch by Roman Arutyunyan.

2 years agoQUIC: simplified ngx_quic_ciphers() API.
Sergey Kandaurov [Fri, 20 Oct 2023 14:05:07 +0000 (18:05 +0400)]
QUIC: simplified ngx_quic_ciphers() API.

After conversion to reusable crypto ctx, now there's enough caller
context to remove the "level" argument from ngx_quic_ciphers().

2 years agoQUIC: cleaned up now unused ngx_quic_ciphers() calls.
Sergey Kandaurov [Fri, 20 Oct 2023 14:05:07 +0000 (18:05 +0400)]
QUIC: cleaned up now unused ngx_quic_ciphers() calls.

2 years agoQUIC: reusing crypto contexts for header protection.
Sergey Kandaurov [Fri, 20 Oct 2023 14:05:07 +0000 (18:05 +0400)]
QUIC: reusing crypto contexts for header protection.

2 years agoQUIC: common code for crypto open and seal operations.
Sergey Kandaurov [Fri, 20 Oct 2023 14:05:07 +0000 (18:05 +0400)]
QUIC: common code for crypto open and seal operations.

2 years agoQUIC: reusing crypto contexts for packet protection.
Sergey Kandaurov [Fri, 20 Oct 2023 14:05:07 +0000 (18:05 +0400)]
QUIC: reusing crypto contexts for packet protection.