Previously the stale-if-error extension of the Cache-Control upstream header
triggered the return of a stale response for all error conditions that can be
specified in the proxy_cache_use_stale directive. The list of these errors
includes both network/timeout/format errors, as well as some HTTP codes like
503, 504, 403, 429 etc. The latter prevented a cache entry from being updated
by a response with any of these HTTP codes during the stale-if-error period.
Now stale-if-error only works for network/timeout/format errors and ignores
the upstream HTTP code. The return of a stale response for certain HTTP codes
is still possible using the proxy_cache_use_stale directive.
This change also applies to the stale-while-revalidate extension of the
Cache-Control header, which triggers stale-if-error if it is missing.
Reported at
http://mailman.nginx.org/pipermail/nginx/2020-July/059723.html.
Reworked connections reuse, so closing connections is attempted in
advance, as long as number of free connections is less than 1/16 of
worker connections configured. This ensures that new connections can
be handled even if closing a reusable connection requires some time,
for example, for a lingering close (ticket #2017).
The 1/16 ratio is selected to be smaller than 1/8 used for disabling
accept when working with accept mutex, so nginx will try to balance
new connections to different workers first, and will start reusing
connections only if this won't help.
Previously, reusing connections happened silently and was only
visible in monitoring systems. This was shown to be not very user-friendly,
and administrators often didn't realize there were too few connections
available to withstand the load, and configured timeouts (keepalive_timeout
and http2_idle_timeout) were effectively reduced to keep things running.
To provide at least some information about this, a warning is now logged
(at most once per second, to avoid flooding the logs).
Sending shutdown when ngx_http_test_reading() detects the connection is
closed can result in "SSL_shutdown() failed (SSL: ... bad write retry)"
critical log messages if there are blocked writes.
Fix is to avoid sending shutdown via the c->ssl->no_send_shutdown flag,
similarly to how it is done in ngx_http_keepalive_handler() for kqueue
when pending EOF is detected.
Reported by Jan Prachař
(http://mailman.nginx.org/pipermail/nginx-devel/2018-December/011702.html).
Without the flag, SSL shutdown is attempted on such connections,
resulting in useless work and/or bogus "SSL_shutdown() failed
(SSL: ... bad write retry)" critical log messages if there are
blocked writes.
Previously, bidirectional shutdown never worked, due to two issues
in the code:
1. The code only tested SSL_ERROR_WANT_READ and SSL_ERROR_WANT_WRITE
when there was an error in the error queue, which cannot happen.
The bug was introduced in an attempt to fix unexpected error logging
as reported with OpenSSL 0.9.8g
(http://mailman.nginx.org/pipermail/nginx/2008-January/003084.html).
2. The code never called SSL_shutdown() for the second time to wait for
the peer's close_notify alert.
This change fixes both issues.
Note that after this change bidirectional shutdown is expected to work for
the first time, so c->ssl->no_wait_shutdown now makes a difference. This
is not a problem for HTTP code which always uses c->ssl->no_wait_shutdown,
but might be a problem for stream and mail code, as well as 3rd party
modules.
To minimize the effect of the change, the timeout, which was used to be 30
seconds and not configurable, though never actually used, is now set to
3 seconds. It is also expanded to apply to both SSL_ERROR_WANT_READ and
SSL_ERROR_WANT_WRITE, so timeout is properly set if writing to the socket
buffer is not possible.
If some additional data from a pipelined request happens to be
read into the body buffer, we copy it to r->header_in or allocate
an additional large client header buffer for it.
This ensures that copying won't write more than the buffer size
even if the buffer comes from hc->free and it is smaller than the large
client header buffer size in the virtual host configuration. This might
happen if size of large client header buffers is different in name-based
virtual hosts, similarly to the problem with number of buffers fixed
in 6926:e662cbf1b932.
After 05e42236e95b (1.19.1) responses with extra data might result in
zero size buffers being generated and "zero size buf" alerts in writer
(if f->rest happened to be 0 when processing additional stdout data).
Previously, the document generated by the xslt filter was always fully sent
to client even if a range was requested and response status was 206 with
appropriate Content-Range.
The xslt module is unable to serve a range because of suspending the header
filter chain. By the moment full response xml is buffered by the xslt filter,
range header filter is not called yet, but the range body filter has already
been called and did nothing.
The fix is to disable ranges by resetting the r->allow_ranges flag much like
the image filter that employs a similar technique.
The slice filter allows ranges for the response by setting the r->allow_ranges
flag, which enables the range filter. If the range was not requested, the
range filter adds an Accept-Ranges header to the response to signal the
support for ranges.
Previously, if an Accept-Ranges header was already present in the first slice
response, client received two copies of this header. Now, the slice filter
removes the Accept-Ranges header from the response prior to setting the
r->allow_ranges flag.
As long as the "Content-Length" header is given, we now make sure
it exactly matches the size of the response. If it doesn't,
the response is considered malformed and must not be forwarded
(https://tools.ietf.org/html/rfc7540#section-8.1.2.6). While it
is not really possible to "not forward" the response which is already
being forwarded, we generate an error instead, which is the closest
equivalent.
Previous behaviour was to pass everything to the client, but this
seems to be suboptimal and causes issues (ticket #1695). Also this
directly contradicts HTTP/2 specification requirements.
Note that the new behaviour for the gRPC proxy is more strict than that
applied in other variants of proxying. This is intentional, as HTTP/2
specification requires us to do so, while in other types of proxying
malformed responses from backends are well known and historically
tolerated.
Previous behaviour was to pass everything to the client, but this
seems to be suboptimal and causes issues (ticket #1695). Fix is to
drop extra data instead, as it naturally happens in most clients.
Additionally, we now also issue a warning if the response is too
short, and make sure the fact it is truncated is propagated to the
client. The u->error flag is introduced to make it possible to
propagate the error to the client in case of unbuffered proxying.
For responses to HEAD requests there is an exception: we do allow
both responses without body and responses with body matching the
Content-Length header.
Previous behaviour was to pass everything to the client, but this
seems to be suboptimal and causes issues (ticket #1695). Fix is to
drop extra data instead, as it naturally happens in most clients.
This change covers generic buffered and unbuffered filters as used
in the scgi and uwsgi modules. Appropriate input filter init
handlers are provided by the scgi and uwsgi modules to set corresponding
lengths.
Note that for responses to HEAD requests there is an exception:
we do allow any response length. This is because responses to HEAD
requests might be actual full responses, and it is up to nginx
to remove the response body. If caching is enabled, only full
responses matching the Content-Length header will be cached
(see b779728b180c).
Previously, additional data after final chunk was either ignored
(in the same buffer, or during unbuffered proxying) or sent to the
client (in the next buffer already if it was already read from the
socket). Now additional data are properly detected and ignored
in all cases. Additionally, a warning is now logged and keepalive
is disabled in the connection.
Previous behaviour was to pass everything to the client, but this
seems to be suboptimal and causes issues (ticket #1695). Fix is to
drop extra data instead, as it naturally happens in most clients.
If a memcached response was followed by a correct trailer, and then
the NUL character followed by some extra data - this was accepted by
the trailer checking code. This in turn resulted in ctx->rest underflow
and caused negative size buffer on the next reading from the upstream,
followed by the "negative size buf in writer" alert.
Fix is to always check for too long responses, so a correct trailer cannot
be followed by extra data.
After sending the GOAWAY frame, a connection is now closed using
the lingering close mechanism.
This allows for the reliable delivery of the GOAWAY frames, while
also fixing connection resets observed when http2_max_requests is
reached (ticket #1250), or with graceful shutdown (ticket #1544),
when some additional data from the client is received on a fully
closed connection.
For HTTP/2, the settings lingering_close, lingering_timeout, and
lingering_time are taken from the "server" level.
Using SSL_CTX_set_verify(SSL_VERIFY_PEER) implies that OpenSSL will
send a certificate request during an SSL handshake, leading to unexpected
certificate requests from browsers as long as there are any client
certificates installed. Given that ngx_ssl_trusted_certificate()
is called unconditionally by the ngx_http_ssl_module, this affected
all HTTPS servers. Broken by 699f6e55bbb4 (not released yet).
Fix is to set verify callback in the ngx_ssl_trusted_certificate() function
without changing the verify mode.
Clearing cache based on free space left on a file system is
expected to allow better disk utilization in some cases, notably
when disk space might be also used for something other than nginx
cache (including nginx own temporary files) and while loading
cache (when cache size might be inaccurate for a while, effectively
disabling max_size cache clearing).
Based on a patch by Adam Bambuch.
With XFS, using "allocsize=64m" mount option results in large preallocation
being reported in the st_blocks as returned by fstat() till the file is
closed. This in turn results in incorrect cache size calculations and
wrong clearing based on max_size.
To avoid too aggressive cache clearing on such volumes, st_blocks values
which result in sizes larger than st_size and eight blocks (an arbitrary
limit) are no longer trusted, and we use st_size instead.
The ngx_de_fs_size() counterpart is intentionally not modified, as
it is used on closed files and hence not affected by this problem.
NFS on Linux is known to report wsize as a block size (in both f_bsize
and f_frsize, both in statfs() and statvfs()). On the other hand,
typical file system block sizes on Linux (ext2/ext3/ext4, XFS) are limited
to pagesize. (With FAT, block sizes can be at least up to 512k in
extreme cases, but this doesn't really matter, see below.)
To avoid too aggressive cache clearing on NFS volumes on Linux, block
sizes larger than pagesize are now ignored.
Note that it is safe to ignore large block sizes. Since 3899:e7cd13b7f759
(1.0.1) cache size is calculated based on fstat() st_blocks, and rounding
to file system block size is preserved mostly for Windows.
Note well that on other OSes valid block sizes seen are at least up
to 65536. In particular, UFS on FreeBSD is known to work well with block
and fragment sizes set to 65536.
When validating second and further certificates, ssl callback could be called
twice to report the error. After the first call client connection is
terminated and its memory is released. Prior to the second call and in it
released connection memory is accessed.
Errors triggering this behavior:
- failure to create the request
- failure to start resolving OCSP responder name
- failure to start connecting to the OCSP responder
The fix is to rearrange the code to eliminate the second call.
The flush flag was not set when forwarding the request body to the uwsgi
server. When using uwsgi_pass suwsgi://..., this causes the uwsgi server
to wait indefinitely for the request body and eventually time out due to
SSL buffering.
This is essentially the same change as 4009:3183165283cc, which was made
to ngx_http_proxy_module.c.
This will fix the uwsgi bug https://github.com/unbit/uwsgi/issues/1490.
This ensures that certificate verification is properly logged to debug
log during upstream server certificate verification. This should help
with debugging various certificate issues.
Listening UNIX sockets were not removed on graceful shutdown, preventing
the next runs. The fix is to replace the custom socket closing code in
ngx_master_process_cycle() by the ngx_close_listening_sockets() call.
When changing binary, sending a SIGTERM to the new binary's master process
should not remove inherited UNIX sockets unless the old binary's master
process has exited.
Previously, invalid connection preface errors were only logged at debug
level, providing no visible feedback, in particular, when a plain text
HTTP/2 listening socket is erroneously used for HTTP/1.x connections.
Now these are explicitly logged at the info level, much like other
client-related errors.
When enabled, certificate status is stored in cache and is used to validate
the certificate in future requests.
New directive ssl_ocsp_cache is added to configure the cache.
OCSP validation for client certificates is enabled by the "ssl_ocsp" directive.
OCSP responder can be optionally specified by "ssl_ocsp_responder".
When session is reused, peer chain is not available for validation.
If the verified chain contains certificates from the peer chain not available
at the server, validation will fail.
Previously only the first responder address was used per each stapling update.
Now, in case of a network or parsing error, next address is used.
This also fixes the issue with unsupported responder address families
(ticket #1330).
As per https://tools.ietf.org/html/rfc7540#section-8.1,
: A server can send a complete response prior to the client
: sending an entire request if the response does not depend on
: any portion of the request that has not been sent and
: received. When this is true, a server MAY request that the
: client abort transmission of a request without error by
: sending a RST_STREAM with an error code of NO_ERROR after
: sending a complete response (i.e., a frame with the
: END_STREAM flag). Clients MUST NOT discard responses as a
: result of receiving such a RST_STREAM, though clients can
: always discard responses at their discretion for other
: reasons.
Previously, RST_STREAM(NO_ERROR) received from upstream after
a frame with the END_STREAM flag was incorrectly treated as an
error. Now, a single RST_STREAM(NO_ERROR) is properly handled.
This fixes problems observed with modern grpc-c [1], as well
as with the Go gRPC module.
[1] https://github.com/grpc/grpc/pull/1661
The request processing is delayed by a timer. Since nginx updates
internal time once at the start of each event loop iteration, this
normally ensures constant time delay, adding a mitigation from
time-based attacks.
A notable exception to this is the case when there are no additional
events before the timer expires. To ensure constant-time processing
in this case as well, we trigger an additional event loop iteration
by posting a dummy event for the next event loop iteration.
When "aio" or "aio threads" is used while processing the response body of an
in-memory background subrequest, the subrequest could be finalized with an aio
operation still in progress. Upon aio completion either parent request is
woken or the old r->write_event_handler is called again. The latter may result
in request errors. In either case post_subrequest handler is never called with
the full response body, which is typically expected when using in-memory
subrequests.
Currently in nginx background subrequests are created by the upstream module
and the mirror module. The issue does not manifest itself with these
subrequests because they are header-only. But it can manifest itself with
third-party modules which create in-memory background subrequests.
We used to have default error_page overwrite for 495, 496, and 497, so
a configuration like
error_page 495 /error;
will result in error 400, much like without any error_page configured.
The 494 status code was introduced later (in 3848:de59ad6bf557, nginx 0.9.4),
and relevant changes to ngx_http_core_error_page() were missed, resulting
in inconsistent behaviour of "error_page 494" - with error_page configured
it results in 494 being returned instead of 400.
Reported by Frank Liu,
http://mailman.nginx.org/pipermail/nginx/2020-February/058957.html.
In "co64" atom chunk start offset is a 64-bit unsigned integer. When trimming
the "mdat" atom, chunk offsets are casted to off_t values which are typically
64-bit signed integers. A specially crafted mp4 file with huge chunk offsets
may lead to off_t overflow and result in negative trim boundaries.
The consequences of the overflow are:
- Incorrect Content-Length header value in the response.
- Negative left boundary of the response file buffer holding the trimmed "mdat".
This leads to pread()/sendfile() errors followed by closing the client
connection.
On rare systems where off_t is a 32-bit integer, this scenario is also feasible
with the "stco" atom.
The fix is to add checks which make sure data chunks referenced by each track
are within the mp4 file boundaries. Additionally a few more checks are added to
ensure mp4 file consistency and log errors.
Duplicate "Host" headers were allowed in nginx 0.7.0 (revision b9de93d804ea)
as a workaround for some broken Motorola phones which used to generate
requests with two "Host" headers[1]. It is believed that this workaround
is no longer relevant.
[1] http://mailman.nginx.org/pipermail/nginx-ru/2008-May/017845.html
The "identity" transfer coding has been removed in RFC 7230. It is
believed that it is not used in real life, and at the same time it
provides a potential attack vector.
We anyway do not support more than one transfer encoding, so accepting
requests with multiple Transfer-Encoding headers doesn't make sense.
Further, we do not handle multiple headers, and ignore anything but
the first header.
Reported by Filippo Valsorda.
A connection could get stuck without timers if a client has partially sent
the HEADERS frame such that it was split on the individual header boundary.
In this case, it cannot be processed without the rest of the HEADERS frame.
The fix is to call ngx_http_v2_state_headers_save() in this case. Normally,
it would be called from the ngx_http_v2_state_header_block() handler on the
next iteration, when there is not enough data to continue processing. This
isn't the case if recv_buffer became empty and there's no more data to read.
With the recent change to prevent frames flood in d4448892a294,
nginx will finalize the connection with NGX_HTTP_V2_INTERNAL_ERROR
whenever flood is detected, causing nginx aborting or stopping if
the debug_points directive is used in nginx config.
Previous change 1ce3f01a4355 incorrectly introduced processing of the
ngx_posted_next_events queue at the end of operation, effectively making
posted next events a nop, since at the end of an event loop iteration
the queue is always empty. Correct approach is to move events to the
ngx_posted_events queue at an iteration start, as it was done previously.
Further, in some cases the c->read event might be already in the
ngx_posted_events queue, and calling ngx_post_event() with the
ngx_posted_next_events queue won't do anything. To make sure the event
will be correctly placed into the ngx_posted_next_events queue
we now check if it is already posted.
Introduced in 9d2ad2fb4423 available bytes handling in SSL relied
on connection read handler being overwritten to set the ready flag
and the amount of available bytes. This approach is, however, does
not work properly when connection read handler is changed, for example,
when switching to a next pipelined request, and can result in unexpected
connection timeouts, see here:
http://mailman.nginx.org/pipermail/nginx-devel/2019-December/012825.html
Fix is to introduce ngx_event_process_posted_next() instead, which
will set ready and available regardless of how event handler is set.
When ngx_http_v2_close_stream_handler() is used to retry stream close
after queued frames are sent, client timeouts on the stream can be
logged multiple times and/or in addition to already happened errors.
To resolve this, separate ngx_http_v2_retry_close_stream_handler()
was introduced, which does not try to log timeouts.
If a stream is closed with queued frames, it is possible that no further
write events will occur on the stream, leading to the socket leak.
To fix this, the stream's fake connection read handler is set to
ngx_http_v2_close_stream_handler(), to make sure that finalizing the
connection with ngx_http_v2_finalize_connection() will be able to
close the stream regardless of the current number of queued frames.
Additionally, the stream's fake connection fc->error flag is explicitly
set, so ngx_http_v2_handle_stream() will post a write event when queued
frames are finally sent even if stream flow control window is exhausted.
These checks were missed when chunked support was introduced. And also
added an explicit error message to ngx_http_dav_copy_move_handler()
(it was missed for some reason, in contrast to DELETE and MKCOL handlers).
While empty replacements were caught at run-time, parsing code
of the "rewrite" directive expects that a minimum length of the
"replacement" argument is 1.
If a rewritten URI has the null character, only a part of URI was
copied to a memory buffer allocated for path. In some setups this
could be exploited to expose uninitialized memory via the Location
header.
The "alias" directive cannot be used in the same location where URI
was rewritten. This has been detected in the "rewrite ... break"
case, but not when the standalone "break" directive was used.
This change also fixes proxy_pass with URI component in a similar
case:
location /aaa/ {
rewrite ^ /xxx/yyy;
break;
proxy_pass http://localhost:8080/bbb/;
}
Previously, the "/bbb/yyy" would be sent to a backend instead of
"/xxx/yyy". And if location's prefix was longer than the rewritten
URI, a segmentation fault might occur.
Previously, connections returned from keepalive cache had c->data
pointing to the keepalive cache item. While this shouldn't be a problem
for correct code, as c->data is not expected to be used before it is set,
explicitly clearing it might help to avoid confusion.
Previously only an rbtree was associated with a limit_conn. To make it
possible to associate more data with a limit_conn, shared context is introduced
similar to limit_req. Also, shared pool pointer is kept in a way similar to
limit_req.
Now a new structure ngx_proxy_protocol_t holds these fields. This allows
to add more PROXY protocol fields in the future without modifying the
connection structure.
With MinGW-w64, building 64-bit nginx binary with GCC 8 and above
results in warning due to cast of GetProcAddress() result to ngx_wsapoll_pt,
which GCC thinks is incorrect. Added intermediate cast to "void *" to
silence the warning.
FormatMessage() seems to return many errors which essentially indicate that
the language in question is not available. At least the following were
observed in the wild and during testing: ERROR_MUI_FILE_NOT_FOUND (15100)
(ticket #1868), ERROR_RESOURCE_TYPE_NOT_FOUND (1813). While documentation
says it should be ERROR_RESOURCE_LANG_NOT_FOUND (1815), this doesn't seem
to be the case.
As such, checking error code was removed, and as long as FormatMessage()
returns an error, we now always try the default language.
Added code to track number of bytes available in the socket.
This makes it possible to avoid looping for a long time while
working with fast enough peer when data are added to the socket buffer
faster than we are able to read and process data.
When kernel does not provide number of bytes available, it is
retrieved using ioctl(FIONREAD) as long as a buffer is filled by
SSL_read().
It is assumed that number of bytes returned by SSL_read() is close
to the number of bytes read from the socket, as we do not use
SSL compression. But even if it is not true for some reason, this
is not important, as we post an additional reading event anyway.
Note that data can be buffered at SSL layer, and it is not possible
to simply stop reading at some point and wait till the event will
be reported by the kernel again. This can be only done when there
are no data in SSL buffers, and there is no good way to find out if
it's the case.
Instead of trying to figure out if SSL buffers are empty, this patch
introduces events posted for the next event loop iteration - such
events will be processed only on the next event loop iteration,
after going into the kernel and retrieving additional events. This
seems to be simple and reliable approach.
This makes it possible to avoid looping for a long time while working
with a fast enough peer when data are added to the socket buffer faster
than we are able to read and process them (ticket #1431). This is
basically what we already do on FreeBSD with kqueue, where information
about the number of bytes in the socket buffer is returned by
the kevent() call.
With other event methods rev->available is now set to -1 when the socket
is ready for reading. Later in ngx_recv() and ngx_recv_chain(), if
full buffer is received, real number of bytes in the socket buffer is
retrieved using ioctl(FIONREAD). Reading more than this number of bytes
ensures that even with edge-triggered event methods the event will be
triggered again, so it is safe to stop processing of the socket and
switch to other connections.
Using ioctl(FIONREAD) only after reading a full buffer is an optimization.
With this approach we only call ioctl(FIONREAD) when there are at least
two recv()/readv() calls.
As long as there are data to read in the socket, yet the amount of data
is less than total size of the buffers in the chain, this saves one
unneeded read() syscall. Before this change, reading only stopped if
ngx_ssl_recv() returned no data, that is, two read() syscalls in a row
returned EAGAIN.
In SSL connections, data can be buffered by the SSL layer, and it is
wrong to avoid doing c->recv_chain() if c->read->available is 0 and
c->read->pending_eof is set. And tests show that the optimization in
question indeed can result in incorrect detection of premature connection
close if upstream closes the connection without sending a close notify
alert at the same time. Fix is to disable c->read->available optimization
for SSL connections.
