Previously, the source IP address of a response UDP datagram could differ from
the original datagram destination address. This could happen if the server UDP
socket is bound to a wildcard address and the network interface chosen to output
the response packet has a different default address than the destination address
of the original packet. For example, if two addresses from the same network are
configured on an interface.
Now source address is set explicitly if a response is sent for a server UDP
socket bound to a wildcard address.
This change adds "http_429" parameter to "proxy_next_upstream" for
retrying rate-limited requests, and to "proxy_cache_use_stale" for
serving stale cached responses after being rate-limited.
Signed-off-by: Piotr Sikora <piotrsikora@google.com>
This change adds reason phrase in status line and pretty response body
when "429" status code is used in "return", "limit_conn_status" and/or
"limit_req_status" directives.
Signed-off-by: Piotr Sikora <piotrsikora@google.com>
When a slice subrequest was redirected to a new location, its context was lost.
After its completion, a new slice subrequest for the same slice was created.
This could lead to infinite loop. Now the slice module makes sure each slice
subrequest starts output with the slice context available.
With post_action or subrequests, it is possible that the timer set for
wev->delayed will expire while the active subrequest write event handler
is not ready to handle this. This results in request hangs as observed
with limit_rate / sendfile_max_chunk and post_action (ticket #776) or
subrequests (ticket #1228).
Moving the handling to the connection event handler fixes the hangs observed,
and also slightly simplifies the code.
Since limit_req uses connection's write event to delay request processing,
it can conflict with timers in other subrequests. In particular, even
if applied to an active subrequest, it can break things if wev->delayed
is already set (due to limit_rate or sendfile_max_chunk), since after
limit_req finishes the wev->delayed flag will be set and no timer will be
active.
Fix is to use the wev->delayed flag in limit_req as well. This ensures that
wev->delayed won't be set after limit_req finishes, and also ensures that
limit_req's timers will be properly handled by other subrequests if the one
delayed by limit_req is not active.
All streams in connection must be finalized before the connection
itself can be finalized and all related memory is freed. That's
not always possible on the current event loop iteration.
Thus when the last stream is finalized, it sets the special read
event handler ngx_http_v2_handle_connection_handler() and posts
the event.
Previously, this handler didn't check the connection state and
could call the regular event handler on a connection that was
already in finalization stage. In the worst case that could
lead to a segmentation fault, since some data structures aren't
supposed to be used during connection finalization. Particularly,
the waiting queue can contain already freed streams, so the
WINDOW_UPDATE frame received by that moment could trigger
accessing to these freed streams.
Now, the connection error flag is explicitly checked in
ngx_http_v2_handle_connection_handler().
In order to finalize stream the error flag is set on fake connection and
either "write" or "read" event handler is called. The read events of fake
connections are always ready, but it's not the case with the write events.
When the ready flag isn't set, the error flag can be not checked in some
cases and as a result stream isn't finalized. Now the ready flag is
explicilty set on write events for proper finalization in all cases.
Previously, flow control didn't account for padding in DATA frames,
which meant that its view of the world could drift from peer's view
by up to 256 bytes per received padded DATA frame, which could lead
to a deadlock.
Signed-off-by: Piotr Sikora <piotrsikora@google.com>
Previously, its value accounted for payloads of HEADERS, CONTINUATION
and DATA frames, as well as frame headers of HEADERS and DATA frames,
but it didn't account for frame headers of CONTINUATION frames.
Signed-off-by: Piotr Sikora <piotrsikora@google.com>
Previously, connection write handler was called, resulting in wake up
of the active subrequest. This change makes it possible to read data
in non-active subrequests as well. For example, this allows SSI to
process instructions in non-active subrequests earlier and start
additional subrequests if needed, reducing overall response time.
If the subrequest is already finalized, the handler set with aio_write
may still be used by sendfile in threads when using range requests
(see also e4c1f5b32868, and the original note in 9fd738b85fad). Calling
already finalized subrequest's r->write_event_handler in practice
results in request hang in some cases.
Fix is to trigger connection event handler if the subrequest was already
finalized.
The ngx_linux_sendfile() function is now used for both normal sendfile()
and sendfile in threads. The ngx_linux_sendfile_thread() function was
modified to use the same interface as ngx_linux_sendfile(), and is simply
called from ngx_linux_sendfile() when threads are enabled.
Special return code NGX_DONE is used to indicate that a thread task was
posted and no further actions are needed.
If number of bytes sent is less that what we were sending, we now always
retry sending. This is needed for sendfile() in threads as the number
of bytes we are sending might have been changed since the thread task
was posted. And this is also needed for Linux 4.3+, as sendfile() might
be interrupted at any time and provides no indication if it was interrupted
or not (ticket #1174).
If of.err is 0, it means that there was a memory allocation error
and no further logging and/or processing is needed. The of.failed
string can be only accessed if of.err is not 0.
The directive configures a timeout to be used when gracefully shutting down
worker processes. When the timer expires, nginx will try to close all
the connections currently open to facilitate shutdown.
There is no need to cancel timers early if there are other timers blocking
shutdown anyway. Preserving such timers allows nginx to continue some
periodic work till the shutdown is actually possible.
With the new approach, timers with ev->cancelable are simply ignored when
checking if there are any timers left during shutdown.
The ev->timedout flag is set on first timer expiration, and never reset
after it. Due to this the code to stop the timer when the timer was
canceled never worked (except in a very specific time frame immediately
after start), and the timer was always armed again. This essentially
resulted in a buffer flush at the end of an event loop iteration.
This behaviour actually seems to be better than just stopping the flush
timer for the whole shutdown, so it is preserved as is instead of fixing
the code to actually remove the timer. It will be further improved by
upcoming changes to preserve cancelable timers if there are other timers
blocking shutdown.
Most notably, this fixes possible buffer overflows if number of large
client header buffers in a virtual server is different from the one in
the default server.
Reported by Daniil Bondarev.
Cloned subrequests should inherit r->content_handler. This way they will
be able to use the same location configuration as the original request
if there are "if" directives in the configuration.
Without r->content_handler inherited, the following configuration tries
to access a static file in the update request:
location / {
set $true 1;
if ($true) {
# nothing
}
proxy_pass http://backend;
proxy_cache one;
proxy_cache_use_stale updating;
proxy_cache_background_update on;
}
See http://mailman.nginx.org/pipermail/nginx/2017-February/053019.html for
initial report.
With "proxy_ignore_client_abort off" (the default), upstream module changes
r->read_event_handler to ngx_http_upstream_rd_check_broken_connection().
If the handler is not cleared during upstream finalization, it can be
triggered later, causing unexpected effects, if, for example, a request
was redirected to a different location using error_page or X-Accel-Redirect.
In particular, it makes "proxy_ignore_client_abort on" non-working after
a redirection in a configuration like this:
location = / {
error_page 502 = /error;
proxy_pass http://127.0.0.1:8082;
}
location /error {
proxy_pass http://127.0.0.1:8083;
proxy_ignore_client_abort on;
}
It is also known to cause segmentation faults with aio used, see
http://mailman.nginx.org/pipermail/nginx-ru/2015-August/056570.html.
Fix is to explicitly set r->read_event_handler to ngx_http_block_reading()
during upstream finalization, similar to how it is done in the request body
reading code and in the limit_req module.
This allows to store larger ETag values for proxy_cache_revalidate,
including ones generated as SHA256, and cache responses with longer
Vary (ticket #826).
In particular, this fixes caching of Amazon S3 responses with CORS
enabled, which now use "Vary: Origin, Access-Control-Request-Headers,
Access-Control-Request-Method".
Cache version bumped accordingly.
Previously, slice subrequest location was selected based on request URI.
If request is then redirected to a new location, its context array is cleared,
making the slice module loose current slice range information. This lead to
broken output. Now subrequests with the NGX_HTTP_SUBREQUEST_CLONE flag are
created for slices. Such subrequests stay in the same location as the parent
request and keep the right slice context.
Previously, there was no way to enable the proxy_cache_use_stale behavior by
reading the backend response. Now, stale-while-revalidate and stale-if-error
Cache-Control extensions (RFC 5861) are supported. They specify, how long a
stale response can be used when a cache entry is being updated, or in case of
an error.
The function may leave error in the error queue while returning success,
e.g., when taking a DSO reference to itself as of OpenSSL 1.1.0d:
https://git.openssl.org/?p=openssl.git;a=commit;h=4af9f7f
Notably, this fixes alert seen with statically linked OpenSSL on some platforms.
While here, check OPENSSL_init_ssl() return value.
Previously, buffer size was not changed from the one saved during
initial ngx_ssl_create_connection(), even if the buffer itself was not
yet created. Fix is to change c->ssl->buffer_size in the SNI callback.
Note that it should be also possible to update buffer size even in non-SNI
virtual hosts as long as the buffer is not yet allocated. This looks
like an overcomplication though.
The ngx_event_pipe() function wasn't called on write events with
wev->delayed set. As a result, threaded writing results weren't
properly collected in ngx_event_pipe_write_to_downstream() when a
write event was triggered for a completed write.
Further, this wasn't detected, as p->aio was reset by a thread completion
handler, and results were later collected in ngx_event_pipe_read_upstream()
instead of scheduling a new write of additional data. If this happened
on the last reading from an upstream, last part of the response was never
written to the cache file.
Similar problems might also happen in case of timeouts when writing to
client, as this also results in ngx_event_pipe() not being called on write
events. In this scenario socket leaks were observed.
Fix is to check if p->writing is set in ngx_event_pipe_read_upstream(), and
therefore collect results of previous write operations in case of read events
as well, similar to how we do so in ngx_event_pipe_write_downstream().
This is enough to fix the wev->delayed case. Additionally, we now call
ngx_event_pipe() from ngx_http_upstream_process_request() if there are
uncollected write operations (p->writing and !p->aio). This also fixes
the wev->timedout case.
The ngx_chain_coalesce_file() function may produce more bytes to send then
requested in the limit passed, as it aligns the last file position
to send to memory page boundary. As a result, (limit - send) may become
negative. This resulted in big positive number when converted to size_t
while calling ngx_output_chain_to_iovec().
Another part of the problem is in ngx_chain_coalesce_file(): it changes cl
to the next chain link even if the current buffer is only partially sent
due to limit.
Therefore, if a file buffer was not expected to be fully sent due to limit,
and was followed by a memory buffer, nginx called sendfile() with a part
of the file buffer, and the memory buffer in trailer. If there were enough
room in the socket buffer, this resulted in a part of the file buffer being
skipped, and corresponding part of the memory buffer sent instead.
The bug was introduced in 8e903522c17a (1.7.8). Configurations affected
are ones using limits, that is, limit_rate and/or sendfile_max_chunk, and
memory buffers after file ones (may happen when using subrequests or
with proxying with disk buffering).
Fix is to explicitly check if (send < limit) before constructing trailer
with ngx_output_chain_to_iovec(). Additionally, ngx_chain_coalesce_file()
was modified to preserve unfinished file buffers in cl.
Closing up to 32 connections might be too aggressive if worker_connections
is set to a comparable number (and/or there are only a small number of
reusable connections). If an occasional connection shorage happens in
such a configuration, it leads to closing all reusable connections instead
of gradually reducing keepalive timeout to a smaller value. To improve
granularity in such configurations we now close no more than 1/8 of all
reusable connections at once.
Suggested by Joel Cunningham.
A missing check could cause ngx_stream_ssl_handler() to be applied
to a non-ssl session, which resulted in a null pointer dereference
if ssl_verify_client is enabled.
The bug had appeared in 1.11.8 (41cb1b64561d).
Previously, an unavailable peer was considered recovered after a successful
proxy session to this peer. Until then, only a single client connection per
fail_timeout was allowed to be proxied to the peer.
Since stream sessions can be long, it may take indefinite time for a peer to
recover, limiting the ability of the peer to receive new connections.
Now, a peer is considered recovered after a successful TCP connection is
established to it. Balancers are notified of this event via the notify()
callback.
OpenSSL 1.1.0 now uses normal "nmake; nmake install" instead of using
custom "ms\do_ms.bat" script and "ms\nt.mak" makefile. And Configure
now requires --prefix to be absolute, and no longer derives --openssldir
from prefix (so it's specified explicitly). Generated libraries are now
called "libcrypto.lib" and "libssl.lib" instead of "libeay32.lib"
and "ssleay32.lib". Appropriate tests added to support both old and new
variants.
Additionally, openssl/lhash.h now triggers warning C4090 ('function' :
different 'const' qualifiers), so the warning was disabled.
There are lots of C4244 warnings (conversion from 'type1' to 'type2',
possible loss of data), so they were disabled.
The same applies to C4267 warnings (conversion from 'size_t' to 'type',
possible loss of data), most notably - conversion from ngx_str_t.len to
ngx_variable_value_t.len (which is unsigned:28). Additionally, there
is at least one case when it is not possible to fix the warning properly
without introducing win32-specific code: recv() on win32 uses "int len",
while POSIX defines "size_t len".
The ssize_t type now properly defined for 64-bit compilation with MSVC.
Caught by warning C4305 (truncation from '__int64' to 'ssize_t'), on
"cutoff = NGX_MAX_SIZE_T_VALUE / 10" in ngx_atosz()).
Several C4334 warnings (result of 32-bit shift implicitly converted to 64 bits)
were fixed by adding explicit conversions.
Several C4214 warnings (nonstandard extension used: bit field types other
than int) in ngx_http_script.h fixed by changing bit field types from
uintptr_t to unsigned.
Most notably, warning W8012 (comparing signed and unsigned values) reported
in multiple places where an unsigned value of small type (e.g., u_short) is
promoted to an int and compared to an unsigned value.
Warning W8072 (suspicious pointer arithmetic) disabled, it is reported
when we increment base pointer in ngx_shm_alloc().
These types are available with MSVC (at least since 2003, in stddef.h),
all variants of GCC (in stdint.h) and Watcom C. We need to define them
only for Borland C.
This implies ticket key size of 80 bytes instead of previously used 48,
as both HMAC and AES keys are 32 bytes now. When an old 48-byte ticket key
is provided, we fall back to using backward-compatible AES128 encryption.
OpenSSL switched to using AES256 in 1.1.0, and we are providing equivalent
security. While here, order of HMAC and AES keys was reverted to make
the implementation compatible with keys used by OpenSSL with
SSL_CTX_set_tlsext_ticket_keys().
Prodded by Christian Klinger.
The current version of HTTP/1.1 standard allows relative references in
redirects (https://tools.ietf.org/html/rfc7231#section-7.1.2).
Allow this form for redirects generated by nginx by introducing the new
directive absolute_redirect.
SSL version 3.0 can be specified by the client at the record level for
compatibility reasons. Previously, ssl_preread module rejected such
connections, presuming they don't have SNI. Now SSL 3.0 is allowed at
the record level.
The resolver handles SRV requests in two stages. In the first
stage it gets all SRV RRs, and in the second stage it resolves
the names from SRV RRs into addresses.
Previously, if a response to an SRV request was cached, the
queries to resolve names were not limited by a timeout. If a
response to any of these queries was not received, the SRV
request could never complete.
If a response to an SRV request was not cached, and some of the
queries to resolve names timed out, NGX_RESOLVE_TIMEDOUT was
returned instead of successfully resolved addresses.
To fix both issues, resolving of names is now always limited by
a timeout.
Changeset e7cb5deb951d breaks build on CentOS 5 with "dereferencing
type-punned pointer will break strict-aliasing rules" warning. It is
backed out.
Instead, to keep builds with BoringSSL happy, type of the "value"
variable changed to "char *", and an explicit cast added before calling
ngx_parse_http_time().
A bug was introduced by 82efcedb310b that could lead to timing out of
responses or segmentation fault, when accept_mutex was enabled.
The output queue in HTTP/2 can contain frames from different streams.
When the queue is sent, all related write handlers need to be called.
In order to do so, the streams were added to the h2c->posted queue
after handling sent frames. Then this queue was processed in
ngx_http_v2_write_handler().
If accept_mutex is enabled, the event's "ready" flag is set but its
handler is not called immediately. Instead, the event is added to
the ngx_posted_events queue. At the same time in this queue can be
events from upstream connections. Such events can result in sending
output queue before ngx_http_v2_write_handler() is triggered. And
at the time ngx_http_v2_write_handler() is called, the output queue
can be already empty with some streams added to h2c->posted.
But after 82efcedb310b, these streams weren't processed if all frames
have already been sent and the output queue was empty. This might lead
to a situation when a number of streams were get stuck in h2c->posted
queue for a long time. Eventually these streams might get closed by
the send timeout.
In the worst case this might also lead to a segmentation fault, if
already freed stream was left in the h2c->posted queue. This could
happen if one of the streams was terminated but wasn't closed, due to
the HEADERS frame or a partially sent DATA frame left in the output
queue. If this happened the ngx_http_v2_filter_cleanup() handler
removed the stream from the h2c->waiting or h2c->posted queue on
termination stage, before the frame has been sent, and the stream
was again added to the h2c->posted queue after the frame was sent.
In order to fix all these problems and simplify the code, write
events of fake stream connections are now added to ngx_posted_events
instead of using a custom h2c->posted queue.
By default, "map" creates cacheable variables [1]. With this
parameter it creates a non-cacheable variable.
An original idea was to deduce the cacheability of the "map"
variable by checking the cacheability of variables specified
in source and resulting values, but it turned to be too hard.
For example, a cacheable variable can be overridden with the
"set" directive or with the SSI "set" command. Also, keeping
"map" variables cacheable by default is good for performance
reasons. This required adding a new parameter.
[1] Before db699978a33f (1.11.0), the cacheability of the
"map" variable could vary depending on the cacheability of
variables specified in resulting values (ticket #1090).
This is believed to be a bug rather than a feature.
Removed code that would cause an endless loop, and removed condition
check that is always false. The first page in the slot list is
guaranteed to satisfy an allocation.
On exit environment allocated from a pool is no longer available, leading
to a segmentation fault if, for example, a library tries to use it from
an atexit() handler.
Fix is to allocate environment via ngx_alloc() instead, and explicitly
free it using a pool cleanup handler if it's no longer used (e.g., on
configuration reload).
In Perl 5.8.6 the default was switched to use putenv() when used as
embedded library unless "PL_use_safe_putenv = 0" is explicitly used
in the code. Therefore, for modern versions of Perl it is no longer
necessary to restore previous environment when calling perl_destruct().
For Perl compiled with threads, without PERL_SET_INTERP() the PL_curinterp
remains set to the first interpreter created (that is, one created at
original start). As a result after a reload Perl thinks that operations
are done withing a thread, and, most notably, denies to change environment.
For example, the following code properly works on original start,
but fails after a reload:
perl 'sub {
my $r = shift;
$r->send_http_header("text/plain");
$ENV{TZ} = "UTC";
$r->print("tz: " . $ENV{TZ} . " (localtime " . (localtime()) . ")\n");
$ENV{TZ} = "Europe/Moscow";
$r->print("tz: " . $ENV{TZ} . " (localtime " . (localtime()) . ")\n");
return OK;
}';
To fix this, PERL_SET_INTERP() added anywhere where PERL_SET_CONTEXT()
was previously used.
Note that PERL_SET_INTERP() doesn't seem to be documented anywhere.
Yet it is used in some other software, and also seems to be the only
solution possible.
Atom size is the sum of atom header size and atom data size. The
specification says that the first 4 bytes are set to one when
the atom size is greater than the maximum unsigned 32-bit value.
Which means atom header size should be considered when the
comparison takes place between atom data size and 0xffffffff.
The variable contains a list of curves as supported by the client.
Known curves are listed by their names, unknown ones are shown
in hex, e.g., "0x001d:prime256v1:secp521r1:secp384r1".
Note that OpenSSL uses session data for SSL_get1_curves(), and
it doesn't store full list of curves supported by the client when
serializing a session. As a result $ssl_curves is only available
for new sessions (and will be empty for reused ones).
The variable is only meaningful when using OpenSSL 1.0.2 and above.
With older versions the variable is empty.
The variable contains list of ciphers as supported by the client.
Known ciphers are listed by their names, unknown ones are shown
in hex, e.g., ""AES128-SHA:AES256-SHA:0x00ff".
The variable is fully supported only when using OpenSSL 1.0.2 and above.
With older version there is an attempt to provide some information
using SSL_get_shared_ciphers(). It only lists known ciphers though.
Moreover, as OpenSSL uses session data for SSL_get_shared_ciphers(),
and it doesn't store relevant data when serializing a session. As
a result $ssl_ciphers is only available for new sessions (and not
available for reused ones) when using OpenSSL older than 1.0.2.
Now in case of a verification failure $ssl_client_verify contains
"FAILED:<reason>", similar to Apache's SSL_CLIENT_VERIFY, e.g.,
"FAILED:certificate has expired".
Detailed description of possible errors can be found in the verify(1)
manual page as provided by OpenSSL.
Normally, the epoll module calls the read and write handlers depending
on whether EPOLLIN and EPOLLOUT are reported by epoll_wait(). No error
processing is done in the module, the handlers are expected to get an
error when doing I/O.
If an error event is reported without EPOLLIN and EPOLLOUT, the module
set both EPOLLIN and EPOLLOUT to ensure the error event is handled at
least in one active handler.
This works well unless the error is delivered along with only one of
EPOLLIN or EPOLLOUT, and the corresponding handler does not do any I/O.
For example, it happened when getting EPOLLERR|EPOLLOUT from
epoll_wait() upon receiving "ICMP port unreachable" while proxying UDP.
As the write handler had nothing to send it was not able to detect and
log an error, and did not switch to the next upstream.
The fix is to unconditionally set EPOLLIN and EPOLLOUT in case of an
error event. In the aforementioned case, this causes the read handler
to be called which does recv() and detects an error.
In addition to the epoll module, analogous changes were made in
devpoll/eventport/poll.
Previously, a request body bigger than "client_body_buffer_size" wasn't written
into a temporary file if it has been pre-read entirely. The preread buffer
is freed after processing, thus subsequent use of it might result in sending
corrupted body or cause a segfault.
On Linux, the rename syscall can be slow due to a global file system lock,
acquired for the entire rename operation, unless both old and new files are
in the same directory. To address this temporary files are now created
in the same directory as the expected resulting cache file when using the
"use_temp_path=off" parameter.
This change mostly reverts 99639bfdfa2a and 3281de8142f5, restoring the
behaviour as of a9138c35120d (with minor changes).
Holding a cache node lock doesn't make sense as we can't use caching
anyway, and results in "ignore long locked inactive cache entry" alerts
if a node is locked for a long time.
The same is done for unbuffered connections, as they can be alive for
a long time as well.
It configures a threshold in bytes, above which client range
requests are not cached. In such a case the client's Range
header is passed directly to a proxied server.
As the pointer to the first argument was tested instead of the argument
itself, array of arguments was always created, even if there were no
arguments. Fix is to test args[0] instead of args.
Found by Coverity (CID 1356862).
The only thing that default_port comparison did in the current
code is prevented implicit upstreams to the same address/port
from being aliased for http and https, e.g.:
proxy_pass http://10.0.0.1:12345;
proxy_pass https://10.0.0.1:12345;
This is inconsistent because it doesn't work for a similar case
with uswgi_pass:
uwsgi_pass uwsgi://10.0.0.1:12345;
uwsgi_pass suwsgi://10.0.0.1:12345;
or with an explicit upstream:
upstream u {
server 10.0.0.1:12345;
}
proxy_pass http://u;
proxy_pass https://u;
Before c9059bd5445b, default_port comparison was needed to
differentiate implicit upstreams in
proxy_pass http://example.com;
and
proxy_pass https://example.com;
as u->port was not set.
When an upstream{} block follows a proxy_pass reference to it,
such an upstream inherited port and default_port settings from
proxy_pass. This was different from when they came in another
order (see ticket #1059). Explicit upstreams should not have
port and default_port in any case.
This fixes the following case:
server { location / { proxy_pass http://u; } ... }
upstream u { server 127.0.0.1; }
server { location / { proxy_pass https://u; } ... }
but not the following:
server { location / { proxy_pass http://u; } ... }
server { location / { proxy_pass https://u; } ... }
upstream u { server 127.0.0.1; }
If proxy_pass (and friends) with variables evaluates an upstream
specified with literal address, nginx always created a per-request
upstream.
Now, if there's a matching upstream specified in the configuration
(either implicit or explicit), it will be used instead.
This fixes inconsistency in what is stored in the "host" field.
Normally it would contain the "host" part of the parsed URL
(e.g., proxy_pass with variables), but for the case of an
implicit upstream specified with literal address it contained
the text representation of the socket address (that is, host
including port for IP).
Now the "host" field always contains the "host" part of the URL,
while the text representation of the socket address is stored
in the newly added "name" field.
The ngx_http_upstream_create_round_robin_peer() function was
modified accordingly in a way to be compatible with the code
that does not know about the new "name" field.
The "stream" code was similarly modified except for not adding
compatibility in ngx_stream_upstream_create_round_robin_peer().
This change is also a prerequisite for the next change.
The new directive "http2_max_requests" is introduced. From users point of
view it works quite similar to "keepalive_requests" but has significantly
bigger default value that is more suitable for HTTP/2.
This allows to correctly parse "start" and "end" arguments without
null-termination (ticket #475), and also fixes rounding errors observed
with strtod() when using i387 instructions.
