It turns out no browsers implement HTTP/2 GOAWAY handling properly, and
large enough number of resources on a page results in failures to load
some resources. In particular, Chrome seems to experience errors if
loading of all resources requires more than 1 connection (while it
is usually able to retry requests at least once, even with 2 connections
there are occasional failures for some reason), Safari if loading requires
more than 3 connections, and Firefox if loading requires more than 10
connections (can be configured with network.http.request.max-attempts,
defaults to 10).
It does not seem to be possible to resolve this on nginx side, even strict
limiting of maximum concurrency does not help, and loading issues seems to
be triggered by merely queueing of a request for a particular connection.
The only available mitigation seems to use higher keepalive_requests value.
The new default is 1000 and matches previously used default for
http2_max_requests. It is expected to be enough for 99.98% of the pages
(https://httparchive.org/reports/state-of-the-web?start=latest#reqTotal)
even in Chrome.
Similar to lingering_time, it limits total connection lifetime before
keepalive is switched off. The default is 1 hour, which is close to
the total maximum connection lifetime possible with default
keepalive_requests and keepalive_timeout.
Firefox uses several idle streams for PRIORITY frames[1], and
"http2_max_concurrent_streams 1;" results in "client sent too many
PRIORITY frames" errors when a connection is established by Firefox.
Fix is to relax the PRIORITY frames limit to use at least 100 as
the initial value (which is the recommended by the HTTP/2 protocol
minimum limit on the number of concurrent streams, so it is not
unreasonable for clients to assume that similar number of idle streams
can be used for prioritization).
[1] https://hg.mozilla.org/mozilla-central/file/32a9e6e145d6e3071c3993a20bb603a2f388722b/netwerk/protocol/http/Http2Stream.cpp#l1270
In FreeBSD 13, eventfd(2) was added, and this breaks build
with --test-build-epoll and without --with-file-aio. Fix is
to move eventfd(2) detection to auto/os/linux, as it is used
only on Linux as a notification mechanism for epoll().
In current versions (all versions based on zlib 1.2.11, at least
since 2018) it no longer uses 64K hash and does not force window
bits to 13 if it is less than 13. That is, it needs just 16 bytes
more memory than normal zlib, so these bytes are simply added to
the normal size calculation.
In limit_req, auth_delay, and upstream code to check for broken
connections, tests for possible connection close by the client
did not work if the connection was already closed when relevant
event handler was set. This happened because there were no additional
events in case of edge-triggered event methods, and read events
were disabled in case of level-triggered ones.
Fix is to explicitly post a read event if the c->read->ready flag
is set.
For new data to be reported with eventport on Solaris,
ngx_handle_read_event() needs to be called after reading response
headers. To do so, ngx_http_upstream_process_non_buffered_upstream()
now called unconditionally if there are no prepread data. This
won't cause any read() syscalls as long as upstream connection
is not ready for reading (c->read->ready is not set), but will result
in proper handling of all events.
If we need to be notified about further events, ngx_handle_read_event()
needs to be called after a read event is processed. Without this,
an event can be removed from the kernel and won't be reported again,
notably when using oneshot event methods, such as eventport on Solaris.
While here, error handling is also added, similar to one present in
ngx_resolver_tcp_read(). This is not expected to make a difference
and mostly added for consistency.
If an attempt is made to delete an event which was already reported,
port_dissociate() returns an error. Fix is avoid doing anything if
ev->active is not set.
Possible alternative approach would be to avoid calling ngx_del_event()
at all if ev->active is not set. This approach, however, will require
something else to re-add the other event of the connection, since both
read and write events are dissociated if an event is reported on a file
descriptor. Currently ngx_eventport_del_event() re-associates write
event if called to delete read event, and vice versa.
If, at the start of an event loop iteration, there are any timers
in the past (including timers expiring now), the ngx_process_events()
function is called with zero timeout, and returns immediately even
if there are no events. But the following code only calls
ngx_event_expire_timers() if time actually changed, so this results
in nginx spinning in the event loop till current time changes.
While such timers are not expected to appear under normal conditions,
as all such timers should be removed on previous event loop iterations,
they still can appear due to bugs, zero timeouts set in the configuration
(if this is not explicitly handled by the code), or due to external
time changes on systems without clock_gettime(CLOCK_MONOTONIC).
Fix is to call ngx_event_expire_timers() unconditionally. Calling
it on each event loop iteration is not expected to be significant from
performance point of view, especially compared to a syscall in
ngx_process_events().
Without explicit handling, a zero timer was actually added, leading to
multiple unneeded syscalls. Further, sending GOAWAY frame early might
be beneficial for clients.
Reported by Sergey Kandaurov.
Unlike in 75e908236701, which added the logic to ngx_http_finalize_request(),
this change moves it to a more generic routine ngx_http_finalize_connection()
to cover cases when a request is finalized with NGX_DONE.
In particular, this fixes unwanted connection transition into the keepalive
state after receiving EOF while discarding request body. With edge-triggered
event methods that means the connection will last for extra seconds as set in
the keepalive_timeout directive.
The response size check introduced in 39501ce97e29 did not take into
account possible padding on DATA frames, resulting in incorrect
"upstream sent response body larger than indicated content length" errors
if upstream server used padding in responses with known length.
Fix is to check the actual size of response buffers produced by the code,
similarly to how it is done in other protocols, instead of checking
the size of DATA frames.
Reported at:
http://mailman.nginx.org/pipermail/nginx-devel/2021-March/013907.html
Activated with the "proxy_protocol" directive. Can be combined with
"listen ... proxy_protocol;" and "set_real_ip_from ...;" to pass
client address provided to nginx in the PROXY protocol header.
Activated with the "proxy_protocol" parameter of the "listen" directive.
Obtained information is passed to the auth_http script in Proxy-Protocol-Addr,
Proxy-Protocol-Port, Proxy-Protocol-Server-Addr, and Proxy-Protocol-Server-Port
headers.
Similarly to 40e8ce405859 in the stream module, this reduces the time
accept mutex is held. This also simplifies following changes to
introduce PROXY protocol support.
If we need to be notified about further events, ngx_handle_read_event()
needs to be called after a read event is processed. Without this,
an event can be removed from the kernel and won't be reported again,
notably when using oneshot event methods, such as eventport on Solaris.
For consistency, existing ngx_handle_read_event() call removed from
ngx_mail_read_command(), as this call only covers one of the code paths
where ngx_mail_read_command() returns NGX_AGAIN. Instead, appropriate
processing added to the callers, covering all code paths where NGX_AGAIN
is returned.
As long as a read event is blocked (ignored), ngx_handle_read_event()
needs to be called to make sure no further notifications will be
triggered when using level-triggered event methods, such as select() or
poll().
The "!rev->ready" test seems to be a typo, introduced in the original
commit (719:f30b1a75fd3b). The ngx_handle_write_event() code properly
tests for "rev->ready" instead.
Due to this typo, read events might be unexpectedly removed during
proxying after an event on the other part of the proxied connection.
Catched by mail proxying tests.
The strerrordesc_np() function, introduced in glibc 2.32, provides an
async-signal-safe way to obtain error messages. This makes it possible
to avoid copying error messages.
Previously, systems without sys_nerr (or _sys_nerr) were handled with an
assumption that errors start at 0 and continuous. This is, however, not
something POSIX requires, and not true on some platforms.
Notably, on Linux, where sys_nerr is no longer available for newly linked
binaries starting with glibc 2.32, there are gaps in error list, which
used to stop us from properly detecting maximum errno. Further, on
GNU/Hurd errors start at 0x40000001.
With this change, maximum errno detection is moved to the runtime code,
now able to ignore gaps, and also detects the first error if needed.
This fixes observed "Unknown error" messages as seen on Linux with
glibc 2.32 and on GNU/Hurd.
With this change, behaviour of HTTP/2 becomes even closer to HTTP/1.x,
and client_header_timeout instead of keepalive_timeout is used before
the first request is received.
This fixes HTTP/2 connections being closed even before the first request
if "keepalive_timeout 0;" was used in the configuration; the problem
appeared in f790816a0e87 (1.19.7).
Using default highlighting makes it possible to easily overrule
highlighting specified in the syntax file, see ":highlight-default"
in vim help for details.
Previously, PINGs and other frames extended possible keepalive time,
making it possible to keep an open HTTP/2 connection for a long time.
Now the connection is always closed as long as keepalive_timeout expires,
similarly to how it happens in HTTP/1.x.
Note that as a part of this change, incomplete frames are no longer
trigger a separate timeout, so http2_recv_timeout (replaced by
client_header_timeout in previous patches) is essentially cancelled.
The client_header_timeout is, however, used for SSL handshake and
while reading HEADERS frames.
