RFC 9002, Section 6.1.1 defines packet reordering threshold as 3. Testing
shows that such low value leads to spurious packet losses followed by
congestion window collapse. The change implements dynamic packet threshold
detection based on in-flight packet range. Packet threshold is defined
as half the number of in-flight packets, with mininum value of 3.
Also, renamed ngx_quic_lost_threshold() to ngx_quic_time_threshold()
for better compliance with RFC 9002 terms.
Previosly the threshold was hardcoded at 10000. This value is too low for
high BDP networks. For example, if all frames are STREAM frames, and MTU
is 1500, the upper limit for congestion window would be roughly 15M
(10000 * 1500). With 100ms RTT it's just a 1.2Gbps network (15M * 10 * 8).
In reality, the limit is even lower because of other frame types. Also,
the number of frames that could be used simultaneously depends on the total
amount of data buffered in all server streams, and client flow control.
The change sets frame threshold based on max concurrent streams and stream
buffer size, the product of which is the maximum number of in-flight stream
data in all server streams at any moment. The value is divided by 2000 to
account for a typical MTU 1500 and the fact that not all frames are STREAM
frames.
If connection is network-limited, MTU probes have little chance of being
sent since congestion window is almost always full. As a result, PMTUD
may not be able to reach the real MTU and the connection may operate with
a reduced MTU. The solution is to ignore the congestion window. This may
lead to a temporary increase in in-flight count beyond congestion window.
As per RFC 9000, Section 14.4:
Loss of a QUIC packet that is carried in a PMTU probe is therefore
not a reliable indication of congestion and SHOULD NOT trigger a
congestion control reaction.
Previously, these functions operated on a per-level basis. This however
resulted in excessive logging of in_flight and will also led to extra
work detecting underutilized congestion window in the followup patches.
On some systems the value of ngx_current_msec is derived from monotonic
clock, for which the following is defined by POSIX:
For this clock, the value returned by clock_gettime() represents
the amount of time (in seconds and nanoseconds) since an unspecified
point in the past.
As as result, overflow protection is needed when comparing two ngx_msec_t.
The change adds such protection to the ngx_quic_detect_lost() function.
Since recovery_start field was initialized with ngx_current_msec, all
congestion events that happened within the same millisecond or cycle
iteration, were treated as in recovery mode.
Also, when handling persistent congestion, initializing recovery_start
with ngx_current_msec resulted in treating all sent packets as in recovery
mode, which violates RFC 9002, see example in Appendix B.8.
While here, also fixed recovery_start wrap protection. Previously it used
2 * max_idle_timeout time frame for all sent frames, which is not a
reliable protection since max_idle_timeout is unrelated to congestion
control. Now recovery_start <= now condition is enforced. Note that
recovery_start wrap is highly unlikely and can only occur on a
32-bit system if there are no congestion events for 24 days.
Previously, the expiration caused QUIC connection finalization even if
there are application-terminated streams finishing sending data. Such
finalization terminated these streams.
An easy way to trigger this is to request a large file from HTTP/3 over
a small MTU. In this case keepalive timeout expiration may abruptly
terminate the request stream.
Improved logging for simpler data extraction for plotting congestion
window graphs. In particular, added current milliseconds number from
ngx_current_msec.
While here, simplified logging text and removed irrelevant data.
Starting with OpenSSL 3.0, groups may be added externally with pluggable
KEM providers. Using SSL_get_negotiated_group(), which makes lookup in a
static table with known groups, doesn't allow to list such groups by names
leaving them in hex. Adding X25519MLKEM768 to the default group list in
OpenSSL 3.5 made this problem more visible. SSL_get0_group_name() and,
apparently, SSL_group_to_name() allow to resolve such provider-implemented
groups, which is also "generally preferred" over SSL_get_negotiated_group()
as documented in OpenSSL git commit 93d4f6133f.
This change makes external groups listing by name using SSL_group_to_name()
available since OpenSSL 3.0. To preserve "prime256v1" naming for the group
0x0017, and to avoid breaking BoringSSL and older OpenSSL versions support,
it is used supplementary for a group that appears to be unknown.
See https://github.com/openssl/openssl/issues/27137 for related discussion.
Passwords were not preserved in optimized SSL contexts, the bug had
appeared in d791b4aab (1.23.1), as in the following configuration:
server {
proxy_ssl_password_file password;
proxy_ssl_certificate $ssl_server_name.crt;
proxy_ssl_certificate_key $ssl_server_name.key;
location /original/ {
proxy_pass https://u1/;
}
location /optimized/ {
proxy_pass https://u2/;
}
}
The fix is to always preserve passwords, by copying to the configuration
pool, if dynamic certificates are used. This is done as part of merging
"ssl_passwords" configuration.
To minimize the number of copies, a preserved version is then used for
inheritance. A notable exception is inheritance of preserved empty
passwords to the context with statically configured certificates:
server {
proxy_ssl_certificate $ssl_server_name.crt;
proxy_ssl_certificate_key $ssl_server_name.key;
location / {
proxy_pass ...;
proxy_ssl_certificate example.com.crt;
proxy_ssl_certificate_key example.com.key;
}
}
In this case, an unmodified version (NULL) of empty passwords is set,
to allow reading them from the password prompt on nginx startup.
As an additional optimization, a preserved instance of inherited
configured passwords is set to the previous level, to inherit it
to other contexts:
server {
proxy_ssl_password_file password;
location /1/ {
proxy_pass https://u1/;
proxy_ssl_certificate $ssl_server_name.crt;
proxy_ssl_certificate_key $ssl_server_name.key;
}
location /2/ {
proxy_pass https://u2/;
proxy_ssl_certificate $ssl_server_name.crt;
proxy_ssl_certificate_key $ssl_server_name.key;
}
}
It was possible to write outside of the buffer used to keep UTF-8
decoded values when parsing conversion table configuration.
Since this happened before UTF-8 decoding, the fix is to check in
advance if character codes are of more than 3-byte sequence. Note
that this is already enforced by a later check for ngx_utf8_decode()
decoded values for 0xffff, which corresponds to the maximum value
encoded as a valid 3-byte sequence, so the fix does not affect the
valid values.
Found with AddressSanitizer.
Fixes GitHub issue #529.
As uncovered by recent addition in slice.t, a partially initialized
context, coupled with HTTP 206 response from stub backend, might be
accessed in the next slice subrequest.
Found by bad memory allocator simulation.
Upstream SSL sessions may be of a noticeably larger size with tickets
in TLSv1.2 and older versions, or with "stateless" tickets in TLSv1.3,
if a client certificate is saved into the session. Further, certain
stateless session resumption implemetations may store additional data.
Such one is JDK, known to also include server certificates in session
ticket data, which roughly doubles a decoded session size to slightly
beyond the previous limit. While it's believed to be an issue on the
JDK side, this change allows to save such sessions.
Another, innocent case is using RSA certificates with 8192 key size.
Previously, request might be left in inconsistent state in case of error,
which manifested in "http request count is zero" alerts when used by SSI
filter.
The fix is to reshuffle initialization order to postpone committing state
changes until after any potentially failing parts.
Found by bad memory allocator simulation.
In OpenSSL, session resumption always happens in the default SSL context,
prior to invoking the SNI callback. Further, unlike in TLSv1.2 and older
protocols, SSL_get_servername() returns values received in the resumption
handshake, which may be different from the value in the initial handshake.
Notably, this makes the restriction added in b720f650b insufficient for
sessions resumed with different SNI server name.
Considering the example from b720f650b, previously, a client was able to
request example.org by presenting a certificate for example.org, then to
resume and request example.com.
The fix is to reject handshakes resumed with a different server name, if
verification of client certificates is enabled in a corresponding server
configuration.
The directive sets a timeout during which a keepalive connection will
not be closed by nginx for connection reuse or graceful shutdown.
The change allows clients that send multiple requests over the same
connection without delay or with a small delay between them, to avoid
receiving a TCP RST in response to one of them. This excludes network
issues and non-graceful shutdown. As a side-effect, it also addresses
the TCP reset problem described in RFC 9112, Section 9.6, when the last
sent HTTP response could be damaged by a followup TCP RST. It is important
for non-idempotent requests, which cannot be retried by client.
It is not recommended to set keepalive_min_timeout to large values as
this can introduce an additional delay during graceful shutdown and may
restrict nginx from effective connection reuse.
The build location of the resulting libatomic_ops.a was changed in v7.4.0
after converting libatomic_ops to use libtool. The fix is to use library
from the install path, this allows building with both old and new versions.
Initially reported here:
https://mailman.nginx.org/pipermail/nginx/2018-April/056054.html
This can happen with certificates and certificate keys specified
with variables due to partial cache update in various scenarios:
- cache expiration with only one element of pair evicted
- on-disk update with non-cacheable encrypted keys
- non-atomic on-disk update
The fix is to retry with fresh data on X509_R_KEY_VALUES_MISMATCH.
A new directive "ssl_certificate_cache max=N [valid=time] [inactive=time]"
enables caching of SSL certificate chain and secret key objects specified
by "ssl_certificate" and "ssl_certificate_key" directives with variables.
Co-authored-by: Aleksei Bavshin <a.bavshin@nginx.com>
SSL object cache, as previously introduced in 1.27.2, did not take
into account encrypted certificate keys that might be unexpectedly
fetched from the cache regardless of the matching passphrase. To
avoid this, caching of encrypted certificate keys is now disabled
based on the passphrase callback invocation.
A notable exception is encrypted certificate keys configured without
ssl_password_file. They are loaded once resulting in the passphrase
prompt on startup and reused in other contexts as applicable.
Memory based objects are always inherited, engine based objects are
never inherited to adhere the volatile nature of engines, file based
objects are inherited subject to modification time and file index.
The previous behaviour to bypass cache from the old configuration cycle
is preserved with a new directive "ssl_object_cache_inheritable off;".
It now uses 5/4 times more memory for the pending buffer.
Further, a single allocation is now used, which takes additional 56 bytes
for deflate_allocs in 64-bit mode aligned to 16, to store sub-allocation
pointers, and the total allocation size now padded up to 128 bytes, which
takes theoretically 200 additional bytes in total. This fits though into
"4 * (64 + sizeof(void*))" additional space for ZALLOC used in zlib-ng
2.1.x versions. The comment was updated to reflect this.
While trying to close a stream in ngx_quic_close_streams() by calling its
read event handler, the next stream saved prior to that could be destroyed
recursively. This caused a segfault while trying to access the next stream.
The way the next stream could be destroyed in HTTP/3 is the following.
A request stream read event handler ngx_http_request_handler() could
end up calling ngx_http_v3_send_cancel_stream() to report a cancelled
request stream in the decoder stream. If sending stream cancellation
decoder instruction fails for any reason, and the decoder stream is the
next in order after the request stream, the issue is triggered.
The fix is to postpone calling read event handlers for all streams being
closed to avoid closing a released stream.
Previously, such packets were treated as long header packets with unknown
version 0, and a version negotiation packet was sent in response. This
could be used to set up an infinite traffic reflect loop with another nginx
instance.
Now version negotiation packets are ignored. As per RFC 9000, Section 6.1:
An endpoint MUST NOT send a Version Negotiation packet in response to
receiving a Version Negotiation packet.
Since 0-RTT and 1-RTT data exist in the same packet number space,
ngx_quic_discard_ctx incorrectly discards 1-RTT packets when
0-RTT keys are discarded.
The issue was introduced by 58b92177e7c3c50f77f807ab3846ad5c7bbf0ebe.
