They refer to the same socket descriptor as our real connection, and
deleting them will stop processing of the connection.
Events of fake connections must not be activated, and if it happened there
is nothing we can do. The whole processing should be terminated as soon as
possible, but it is not obvious how to do this safely.
A quote from SPDY draft 2 specification: "The length of each name and
value must be greater than zero. A receiver of a zero-length name or
value must send a RST_STREAM with code PROTOCOL error."
But it appears that Chrome browser allows sending requests over SPDY/2
connection using JavaScript that contain headers with empty values.
For better compatibility across SPDY clients and to be compliant with
HTTP, such headers are no longer rejected.
Also, it is worth noting that in SPDY draft 3 the statement has been
changed so that it permits empty values for headers.
When matching a compiled regex against value in the "Referer" header field,
the length was calculated incorrectly for strings that start from "https://".
This might cause matching to fail for regexes with end-of-line anchors.
Patch by Liangbin Li.
If a relative path is set by variables, then the ngx_conf_full_name()
function was called while processing requests, which causes allocations
from the cycle pool.
A new function that takes pool as an argument was introduced.
Though there are several MIME types commonly used for JavaScript nowadays,
the most common being "text/javascript", "application/javascript", and
currently used by nginx "application/x-javascript", RFC 4329 prefers
"application/javascript".
The "charset_types" directive's default value was adjusted accordingly.
As per perlxs, C preprocessor directives should be at the first
non-whitespace of a line to avoid interpreting them as comments.
#if and #endif are moved so that there are no blank lines before them
to retain them as part of the function body.
Previously, after sending a header we always sent a last buffer and
finalized a request with code 0, even in case of errors. In some cases
this resulted in a loss of ability to detect the response wasn't complete
(e.g. if Content-Length was removed from a response by gzip filter).
This change tries to propogate to a client information that a response
isn't complete in such cases. In particular, with this change we no longer
pretend a returned response is complete if we wasn't able to create
a temporary file.
If an error code suggests the error wasn't fatal, we flush buffered data
and disable keepalive, then finalize request normally. This allows to to
propogate information about a problem to a client, while still sending all
the data we've got from an upstream.
No semantic changes expected, though some checks are done differently.
In particular, the r->cached flag is no longer explicitly checked. Instead,
we relay on u->header_sent not being set if a response is sent from
a cache.
The NGX_HTTP_CLIENT_CLOSED_REQUEST code is allowed to happen after we
started sending a response (much like NGX_HTTP_REQUEST_TIME_OUT), so there
is no need to reset response code to 0 in this case.
Checks were added to both buffered and unbuffered code paths to detect
and complain if a response is incomplete. Appropriate error codes are
now passed to ngx_http_upstream_finalize_request().
With this change in unbuffered mode we now use u->length set to -1 as an
indicator that EOF is allowed per protocol and used to indicate response
end (much like its with p->length in buffered mode). Proxy module was
changed to set u->length to 1 (instead of previously used -1) in case of
chunked transfer encoding used to comply with the above.
That is, by default we assume that response end is signalled by
a connection close. This seems to be better default, and in line
with u->pipe->length behaviour.
Memcached module was modified accordingly.
In case of upstream eof, only responses with u->pipe->length == -1
are now cached/stored. This ensures that unfinished chunked responses
are not cached.
Note well - previously used checks for u->headers_in.content_length_n are
preserved. This provides an additional level of protection if protol data
disagree with Content-Length header provided (e.g., a FastCGI response
is sent with wrong Content-Length, or an incomple SCGI or uwsgi response),
as well as protects from storing of responses to HEAD requests. This should
be reconsidered if we'll consider caching of responses to HEAD requests.
There is no real difference from previously used 0 as NGX_HTTP_* will
become 0 in ngx_http_upstream_finalize_request(), but the change
preserves information about a timeout a bit longer. Previous use of
ETIMEDOUT in one place was just wrong.
Note well that with cacheable responses there will be a difference
(code in ngx_http_upstream_finalize_request() will store the error
in cache), though this change doesn't touch cacheable case.
Previously, ngx_http_upstream_finalize_request(0) was used in most
cases after errors. While with current code there is no difference,
use of NGX_ERROR allows to pass a bit more information into
ngx_http_upstream_finalize_request().
In all cases ngx_http_upstream_finalize_request() with NGX_ERROR now used.
Previously used NGX_HTTP_INTERNAL_SERVER_ERROR in the subrequest in memory
case don't cause any harm, but inconsistent with other uses.
Previous code called ngx_http_finalize_request() with rc = 0. This is
ok if a response status was already set, but resulted in "000" being
logged if it wasn't. In particular this happened with limit_req
if a connection was prematurely closed during limit_req delay.
After a failed partial match we now check if there is another partial
match in previously matched substring to fix cases like "aab" in "aaab".
The ctx->saved string is now always sent if it's present on return
from the ngx_http_sub_parse() function (and reset accordingly). This
allows to release parts of previously matched data.
If a pattern was partially matched at a response end, partially matched
string wasn't send. E.g., a response "fo" was truncated to an empty response
if partially mathed by a pattern "foo".
It is possible to send FLAG_FIN in additional empty data frame, even if it is
known from the content-length header that request body is empty. And Firefox
actually behaves like this (see ticket #357).
To simplify code we sacrificed our microoptimization that did not work right
due to missing check in the ngx_http_spdy_state_data() function for rb->buf
set to NULL.
The exsltRegisterAll() needs to be called before XSLT stylesheets
are compiled, else stylesheet compilation hooks will not work. This
change fixes EXSLT Functions extension.
On Linux, sockaddr length is required to process unix socket addresses properly
due to unnamed sockets (which don't have sun_path set at all) and abstract
namespace sockets.
Minimal data length we expect for further calls was calculated incorrectly
if parsing stopped right after parsing chunk size. This might in theory
affect clients and/or backends using LF instead of CRLF.
Patch by Dmitry Popov.
When several "error_log" directives are specified in the same configuration
block, logs are written to all files with a matching log level.
All logs are stored in the singly-linked list that is sorted by log level in
the descending order.
Specific debug levels (NGX_LOG_DEBUG_HTTP,EVENT, etc.) are not supported
if several "error_log" directives are specified. In this case all logs
will use debug level that has largest absolute value.
Valgrind complains if we pass uninitialized memory to a syscall:
==36492== Syscall param sendmsg(msg.msg_iov[0]) points to uninitialised byte(s)
==36492== at 0x6B5E6A: sendmsg (in /usr/lib/system/libsystem_kernel.dylib)
==36492== by 0x10004288E: ngx_signal_worker_processes (ngx_process_cycle.c:527)
==36492== by 0x1000417A7: ngx_master_process_cycle (ngx_process_cycle.c:203)
==36492== by 0x100001F10: main (nginx.c:410)
==36492== Address 0x7fff5fbff71c is on thread 1's stack
Even initialization of all members of the structure passed isn't enough, as
there is padding which still remains uninitialized and results in Valgrind
complaint. Note there is no real problem here as data from uninitialized
memory isn't used.
If "stderr" was specified in one of the "error_log" directives,
stderr is not redirected to the first error_log on startup,
configuration reload, and reopening log files.
The parameter is mostly identical to http_404, and is expected to
be used in similar situations. The 403 code might be returned by
a backend instead of 404 on initial sync of new directories with rsync.
See here for feature request and additional details:
http://mailman.nginx.org/pipermail/nginx-ru/2013-April/050920.html
An invalid memcached reply that started with '\n' could cause
segmentation fault.
An invalid memcached reply "VALUE / 0 2\r?ok\r\nEND\r\n" was
considered as a valid response.
In addition, if memcached reports that the key was not found,
set u->headers_in.content_length_n to 0. This ensures that
ngx_http_memcached_filter() will not be called while previous
code relied on always intercepting 404. Initialization of
ctx->rest was moved to where it belongs.
Due to peer->checked always set since rev. c90801720a0c (1.3.0)
by round-robin and least_conn balancers (ip_hash not affected),
the code in ngx_http_upstream_free_round_robin_peer() function
incorrectly reset peer->fails too often.
Reported by Dmitry Popov,
http://mailman.nginx.org/pipermail/nginx-devel/2013-May/003720.html
There are two significant changes in this patch:
1) The <= 0 comparison is done with a signed type. This fixes the case
of ngx_time() being larger than r->lingering_time.
2) Calculation of r->lingering_time - ngx_time() is now always done
in the ngx_msec_t type. This ensures the calculation is correct
even if time_t is unsigned and differs in size from ngx_msec_t.
Thanks to Lanshun Zhou.
Previously, input pattern was kept only for regular expressions
with named captures, which resulted in error log entries without
input pattern for PCRE errors that occured while processing
regular expressions without them.
Signed-off-by: Piotr Sikora <piotr@cloudflare.com>
The $proxy_internal_body_length value might change during request lifetime,
notably if proxy_set_body used, and use of a cached value might result in
incorrect upstream requests.
Patch by Lanshun Zhou.
As of now, it allows to better control bandwidth limiting from additional
modules. It is also expected to be used to add variables support to
the limit_rate_after directive.
If nginx was compiled without --with-http_ssl_module, but with some
other module which uses OpenSSL (e.g. --with-mail_ssl_module), insufficient
preprocessor check resulted in build failure. The problem was introduced
by e0a3714a36f8 (1.3.14).
Reported by Roman Arutyunyan.
This is to avoid setting the TCP_NODELAY flag on SPDY socket in
ngx_http_upstream_send_response(). The latter works per request,
but in SPDY case it might affect other streams in connection.
As of 1.3.9, chunked request body may be available with
r->headers_in.content_length_n <= 0. Additionally, request body
may be in multiple buffers even if r->request_body_in_single_buf
was requested.
Dependancy tracking introduced in r5169 were not handled absolute path
names properly. Absolute names might appear in CORE_DEPS if --with-openssl
or --with-pcre configure arguments are used to build OpenSSL/PCRE
libraries.
Additionally, revert part of r5169 to set NGX_INCS from Makefile
variables. Makefile variables have $ngx_include_opt in them, which
might result in wrong include paths being used. As a side effect,
this also restores build with --with-http_perl_module and --without-http
at the same time.
Before 1.3.9 an attempt to read body in a subrequest only caused problems
if body wasn't already read or discarded in a main request. Starting with
1.3.9 it might also cause problems if body was discarded by a main request
before subrequest start.
Fix is to just ignore attempts to read request body in a subrequest, which
looks like right thing to do anyway.
To avoid further breaks it's now done properly, all the dependencies
are now passed to Makefile.PL. While here, fixed include list passed to
Makefile.PL to use Makefile variables rather than a list expanded during
configure.
Sorting of upstream servers by their weights is not required by
current balancing algorithms.
This will likely change mapping to backends served by ip_hash
weighted upstreams.
And corresponding variable $connections_waiting was added.
Previously, waiting connections were counted as the difference between
active connections and the sum of reading and writing connections.
That made it impossible to count more than one request in one connection
as reading or writing (as is the case for SPDY).
Also, we no longer count connections in handshake state as waiting.
This should improve behavior under deficiency of connections.
Since SSL handshake usually takes significant amount of time,
we exclude connections from reusable queue during this period
to avoid premature flush of them.
If proxy_pass to a host with dynamic resolution was used to handle
a subrequest, and host resolution failed, the main request wasn't run
till something else happened on the connection. E.g. request to "/zzz"
with the following configuration hanged:
addition_types *;
resolver 8.8.8.8;
location /test {
set $ihost xxx;
proxy_pass http://$ihost;
}
location /zzz {
add_after_body /test;
return 200 "test";
}
Report and original version of the patch by Lanshun Zhou,
http://mailman.nginx.org/pipermail/nginx-devel/2013-March/003476.html.
Code to reuse of r->request_body->buf in upstream module assumes it's
dedicated buffer, hence after 1.3.9 (r4931) it might reuse r->header_in
if client_body_in_file_only was set, resulting in original request
corruption. It is considered to be safer to always create a dedicated
buffer for rb->bufs to avoid such problems.
After introduction of chunked request body handling in 1.3.9 (r4931),
r->request_body->bufs buffers have b->start pointing to original buffer
start (and b->pos pointing to real data of this particular buffer).
While this is ok as per se, it caused bad things (usually original request
headers included before the request body) after reinit of the request
chain in ngx_http_upstream_reinit() while sending the request to a next
upstream server (which used to do b->pos = b->start for each buffer
in the request chain).
Patch by Piotr Sikora.
If c->recv() returns 0 there is no sense in using ngx_socket_errno for
logging, its value meaningless. (The code in question was copied from
ngx_http_keepalive_handler(), but ngx_socket_errno makes sense there as it's
used as a part of ECONNRESET handling, and the c->recv() call is preceeded
by the ngx_set_socket_errno(0) call.)
In r2411 setting of NGX_HTTP_GZIP_BUFFERED in c->buffered was moved from
ngx_http_gzip_filter_deflate_start() to ngx_http_gzip_filter_buffer() since
it was always called first. But in r2543 the "postpone_gzipping" directive
was introduced, and if postponed gzipping is disabled (the default setting),
ngx_http_gzip_filter_buffer() is not called at all.
We must always set NGX_HTTP_GZIP_BUFFERED after the start of compression
since there is always a trailer that is buffered.
There are no known cases when it leads to any problem with current code.
But we already had troubles in upcoming SPDY implementation.
Not only this is useful for the upcoming SPDY support, but it can
also help to improve HTTPS performance by enabling TLS False Start
in Chrome/Chromium browsers [1]. So, we always enable NPN for HTTPS
if it is supported by OpenSSL.
[1] http://www.imperialviolet.org/2012/04/11/falsestart.html
The c->single_connection was intended to be used as lock mechanism
to serialize modifications of request object from several threads
working with client and upstream connections. The flag is redundant
since threads in nginx have never been used that way.
In Linux 2.6.32, TCP_DEFER_ACCEPT was changed to accept connections
after the deferring period is finished without any data available.
(Reading from the socket returns EAGAIN in this case.)
Since in nginx TCP_DEFER_ACCEPT is set to "post_accept_timeout", we
do not need to wait longer if deferred accept returns with no data.
Previously, only the first request in a connection used timeout
value from the "client_header_timeout" directive while reading
header. All subsequent requests used "keepalive_timeout" for
that.
It happened because timeout of the read event was set to the
value of "keepalive_timeout" in ngx_http_set_keepalive(), but
was not removed when the next request arrived.
Previously, we always created an object and logged 400 (Bad Request)
in access log if a client closed connection without sending any data.
Such a connection was counted as "reading".
Since it's common for modern browsers to behave like this, it's no
longer considered an error if a client closes connection without
sending any data, and such a connection will be counted as "waiting".
Now, we do not log 400 (Bad Request) and keep memory footprint as
small as possible.
Previously, it was allocated from a connection pool and
was selectively freed for an idle keepalive connection.
The goal is to put coupled things in one chunk of memory,
and to simplify handling of request objects.
According to RFC 6066, client is not supposed to request a different server
name at the application layer. Server implementations that rely upon these
names being equal must validate that a client did not send a different name
in HTTP request. Current versions of Apache HTTP server always return 400
"Bad Request" in such cases.
There exist implementations however (e.g., SPDY) that rely on being able to
request different host names in one connection. Given this, we only reject
requests with differing host names if verification of client certificates
is enabled in a corresponding server configuration.
An example of configuration that might not work as expected:
server {
listen 433 ssl default;
return 404;
}
server {
listen 433 ssl;
server_name example.org;
ssl_client_certificate org.cert;
ssl_verify_client on;
}
server {
listen 433 ssl;
server_name example.com;
ssl_client_certificate com.cert;
ssl_verify_client on;
}
Previously, a client was able to request example.com by presenting
a certificate for example.org, and vice versa.
Not only this is consistent with a case without SNI, but this also
prevents abusing configurations that assume that the $host variable
is limited to one of the configured names for a server.
An example of potentially unsafe configuration:
server {
listen 443 ssl default_server;
...
}
server {
listen 443;
server_name example.com;
location / {
proxy_pass http://$host;
}
}
Note: it is possible to negotiate "example.com" by SNI, and to request
arbitrary host name that does not exist in the configuration above.
Previously, this was done only after the whole request header
was parsed, and if an error occurred earlier then the request
was processed in the default server (or server chosen by SNI),
while r->headers_in.server might be set to the value from the
Host: header or host from request line.
r->headers_in.server is in turn used for $host variable and
in HTTP redirects if "server_name_in_redirect" is disabled.
Without the change, configurations that rely on this during
error handling are potentially unsafe if SNI is used.
This change also allows to use server specific settings of
"underscores_in_headers", "ignore_invalid_headers", and
"large_client_header_buffers" directives for HTTP requests
and HTTPS requests without SNI.
The request object will not be created until SSL handshake is complete.
This simplifies adding another connection handler that does not need
request object right after handshake (e.g., SPDY).
There are also a few more intentional effects:
- the "client_header_buffer_size" directive will be taken from the
server configuration that was negotiated by SNI;
- SSL handshake errors and timeouts are not logged into access log
as bad requests;
- ngx_ssl_create_connection() is not called until the first byte of
ClientHello message was received. This also decreases memory
consumption if plain HTTP request is sent to SSL socket.
Previously, only the first request in a connection was assigned the
configuration selected by SNI. All subsequent requests initially
used the default server's configuration, ignoring SNI, which was
wrong.
Now all subsequent requests in a connection will initially use the
configuration selected by SNI. This is done by storing a pointer
to configuration in http connection object. It points to default
server's configuration initially, but changed upon receipt of SNI.
(The request's configuration can be further refined when parsing
the request line and Host: header.)
This change was not made specific to SNI as it also allows slightly
faster access to configuration without the request object.
This change helps to decouple ngx_http_ssl_servername() from the request
object.
Note: now we close connection in case of error during server name lookup
for request. Previously, we did so only for HTTP/0.9 requests.
In case multiple "Cache-Control" headers are sent to a client,
multiple values in $sent_http_cache_control were incorrectly
split by a semicolon. Now they are split by a comma.
In case of error in the read event handling we close a connection
by calling ngx_http_close_connection(), that also destroys connection
pool. Thereafter, an attempt to free a buffer (added in r4892) that
was allocated from the pool could cause SIGSEGV and is meaningless
as well (the buffer already freed with the pool).
It was added in r2717 and no longer needed since r2721,
where the termination was added to ngx_shm_alloc() and
ngx_init_zone_pool(). So then it only corrupts error
messages about ivalid zones.
This allows to proxy WebSockets by using configuration like this:
location /chat/ {
proxy_pass http://backend;
proxy_http_version 1.1;
proxy_set_header Upgrade $http_upgrade;
proxy_set_header Connection "upgrade";
}
Connection upgrade is allowed as long as it was requested by a client
via the Upgrade request header.
If fastcgi end request record was split between several network packets,
with fastcgi_keep_conn it was possible that connection was saved in incorrect
state (e.g. with padding bytes not yet read).
Checks for f->padding before state transitions make code hard to follow,
remove them and make sure we always do another loop iteration after
f->state is set to ngx_http_fastcgi_st_padding.
With fastcgi_keep_conn it was possible that connection was closed after
FCGI_STDERR record with zero padding and without any further data read yet.
This happended as f->state was set to ngx_http_fastcgi_st_padding and then
"break" happened, resulting in p->length being set to f->padding, i.e. 0
(which in turn resulted in connection close).
Fix is to make sure we continue the loop after f->state is set.
After introduction of chunked request body reading support in 1.3.9 (r4931),
the rb->bufs wasn't set if request body was fully preread while calling the
ngx_http_read_client_request_body() function.
Reported by Yichun Zhang (agentzh).
The "secure_link_secret" directive was always inherited from the outer
configuration level even when "secure_link" and "secure_link_md5" were
specified on the inner level.
Before the patch if proxy_method was specified at http{} level the code
to add trailing space wasn't executed, resulting in incorrect requests
to upstream.
The "proxy_bind", "fastcgi_bind", "uwsgi_bind", "scgi_bind" and
"memcached_bind" directives are now inherited; inherited value
can be reset by the "off" parameter. Duplicate directives are
now detected. Parameter value can now contain variables.
Upstreams created by "proxy_pass" with IP address and no port were
broken in 1.3.10, by not initializing port in u->sockaddr.
API change: ngx_parse_url() was modified to always initialize port
(in u->sockaddr and in u->port), even for the u->no_resolve case;
ngx_http_upstream() and ngx_http_upstream_add() were adopted.
