On some systems, openssl 1.0.x is still the default, but it has been
patched to contain all the recent security fixes. As a result of this
patching, it is possible for macro X509_V_FLAG_NO_ALT_CHAINS to be
defined, while the previous behavior of openssl to not look at trusted
chains first, remains.
Fix it: ensure X509_V_FLAG_TRUSTED_FIRST is always set, do not try to
probe for the behavior of openssl based on the existence ofmacros.
Closes#5530
Commit 4a4b63d forgot to set the expected SOCKS5 reply length when the
reply ATYP is X'01'. This resulted in erroneously expecting more bytes
when the request length is greater than the reply length (e.g., when
remotely resolving the hostname).
Closes#5527
When the method is updated inside libcurl we must still not change the
method as set by the user as then repeated transfers with that same
handle might not execute the same operation anymore!
This fixes the libcurl part of #5462
Test 1633 added to verify.
Closes#5499
`http_proxy` will not be available in `conndata` if `CURL_DISABLE_PROXY`
is enabled. Repair the build with that configuration.
Follow-up to f3d501dc67Closes#5498
"Null-checking k->str suggests that it may be null, but it has already
been dereferenced on all paths leading to the check" - and it can't
legally be NULL at this point. Remove check.
Detected by Coverity CID 1463884
Closes#5495
Since Win32 almost always will also have USE_WINSOCK,
we can reduce complexity and always use Sleep there.
Assisted-by: Jay Satiro
Reviewed-by: Daniel Stenberg
Follow up to #5343Closes#5489
... and free it as soon as the transfer is done. It removes the extra
alloc when a new size is set with setopt() and reduces memory for unused
easy handles.
In addition: the closure_handle now doesn't use an allocated buffer at
all but the smallest supported size as a stack based one.
Closes#5472
Using time_t and suseconds_t if suseconds_t is available,
long on Windows (maybe others in the future) and int elsewhere.
Also handle case of ULONG_MAX being greater or equal to INFINITE.
Assisted-by: Jay Satiro
Reviewed-by: Daniel Stenberg
Part of #5343
Make all functions in select.[ch] take timeout_ms as timediff_t
which should always be large enough and signed on all platforms
to take all possible timeout values and avoid type conversions.
Reviewed-by: Jay Satiro
Reviewed-by: Daniel Stenberg
Replaces #5107 and partially #5262
Related to #5240 and #5286Closes#5343
Tested with ngtcp2 built against the OpenSSL library. Additionally
tested with MultiSSL (NSS for TLS and ngtcp2+OpenSSL for QUIC).
The TLS backend (independent of QUIC) may or may not already have opened
the keylog file before. Therefore Curl_tls_keylog_open is always called
to ensure the file is open.
Tested following the same curl and tshark commands as in commit
"vtls: Extract and simplify key log file handling from OpenSSL" using
WolfSSL v4.4.0-stable-128-g5179503e8 from git master built with
`./configure --enable-all --enable-debug CFLAGS=-DHAVE_SECRET_CALLBACK`.
Full support for this feature requires certain wolfSSL build options,
see "Availability note" in lib/vtls/wolfssl.c for details.
Closes#5327
Create a set of routines for TLS key log file handling to enable reuse
with other TLS backends. Simplify the OpenSSL backend as follows:
- Drop the ENABLE_SSLKEYLOGFILE macro as it is unconditionally enabled.
- Do not perform dynamic memory allocation when preparing a log entry.
Unless the TLS specifications change we can suffice with a reasonable
fixed-size buffer.
- Simplify state tracking when SSL_CTX_set_keylog_callback is
unavailable. My original sslkeylog.c code included this tracking in
order to handle multiple calls to SSL_connect and detect new keys
after renegotiation (via SSL_read/SSL_write). For curl however we can
be sure that a single master secret eventually becomes available
after SSL_connect, so a simple flag is sufficient. An alternative to
the flag is examining SSL_state(), but this seems more complex and is
not pursued. Capturing keys after server renegotiation was already
unsupported in curl and remains unsupported.
Tested with curl built against OpenSSL 0.9.8zh, 1.0.2u, and 1.1.1f
(`SSLKEYLOGFILE=keys.txt curl -vkso /dev/null https://localhost:4433`)
against an OpenSSL 1.1.1f server configured with:
# Force non-TLSv1.3, use TLSv1.0 since 0.9.8 fails with 1.1 or 1.2
openssl s_server -www -tls1
# Likewise, but fail the server handshake.
openssl s_server -www -tls1 -Verify 2
# TLS 1.3 test. No need to test the failing server handshake.
openssl s_server -www -tls1_3
Verify that all secrets (1 for TLS 1.0, 4 for TLS 1.3) are correctly
written using Wireshark. For the first and third case, expect four
matches per connection (decrypted Server Finished, Client Finished, HTTP
Request, HTTP Response). For the second case where the handshake fails,
expect a decrypted Server Finished only.
tshark -i lo -pf tcp -otls.keylog_file:keys.txt -Tfields \
-eframe.number -eframe.time -etcp.stream -e_ws.col.Info \
-dtls.port==4433,http -ohttp.desegment_body:FALSE \
-Y 'tls.handshake.verify_data or http'
A single connection can easily be identified via the `tcp.stream` field.
For HTTP 1.x, it's a protocol error when the server sends more bytes
than announced. If this happens, don't reuse the connection, because the
start position of the next response is undefined.
Closes#5440
When USE_RESOLVE_ON_IPS is set (defined on macOS), it means that
numerical IP addresses still need to get "resolved" - but not with DoH.
Reported-by: Viktor Szakats
Fixes#5454Closes#5459
They're only limited to the maximum string input restrictions, not to
256 bytes.
Added test 1178 to verify
Reported-by: Will Roberts
Fixes#5448Closes#5449
Fixed the alt-svc parser to treat a newline as end of line.
The unit tests in test 1654 were done without CRLF and thus didn't quite
match the real world. Now they use CRLF as well.
Reported-by: Peter Wu
Assisted-by: Peter Wu
Assisted-by: Jay Satiro
Fixes#5445Closes#5446
This reverts commit 74623551f3.
Instead mark the function call with (void). Getting the return code and
using it instead triggered Coverity warning CID 1463596 because
snprintf() can return a negative value...
Closes#5441
... as returning a "" is not a good idea as the string is supposed to be
allocated and returning a const string will cause issues.
Reported-by: Brian Carpenter
Follow-up to ed35d6590eCloses#5405
This change introduces a generic way to provide binary data in setopt
options, called BLOBs.
This change introduces these new setopts:
CURLOPT_ISSUERCERT_BLOB, CURLOPT_PROXY_SSLCERT_BLOB,
CURLOPT_PROXY_SSLKEY_BLOB, CURLOPT_SSLCERT_BLOB and CURLOPT_SSLKEY_BLOB.
Reviewed-by: Daniel Stenberg
Closes#5357
- Stick to a single unified way to use structs
- Make checksrc complain on 'typedef struct {'
- Allow them in tests, public headers and examples
- Let MD4_CTX, MD5_CTX, and SHA256_CTX typedefs remain as they actually
typedef different types/structs depending on build conditions.
Closes#5338
Previously, after PASV and immediately after the data connection has
connected, the function would only return the control socket to wait for
which then made the data connection simply timeout and not get polled
correctly. This become obvious when running test 1631 and 1632 event-
based.
Use them only if `_UNICODE` is defined, in which case command-line
arguments have been converted to UTF-8.
Closes https://github.com/curl/curl/pull/3784
- use `wmain` instead of `main` when `_UNICODE` is defined [0]
- define `argv_item_t` as `wchar_t *` in this case
- use the curl_multibyte gear to convert the command-line arguments to
UTF-8
This makes it possible to pass parameters with characters outside of
the current locale on Windows, which is required for some tests, e.g.
the IDN tests. Out of the box, this currently only works with the
Visual Studio project files, which default to Unicode, and winbuild
with the `ENABLE_UNICODE` option.
[0] https://devblogs.microsoft.com/oldnewthing/?p=40643
Ref: https://github.com/curl/curl/issues/3747
Closes https://github.com/curl/curl/pull/3784
Fix theoretical integer overflow in Curl_auth_create_plain_message.
The security impact of the overflow was discussed on hackerone. We
agreed this is more of a theoretical vulnerability, as the integer
overflow would only be triggerable on systems using 32-bits size_t with
over 4GB of available memory space for the process.
Closes#5391
Since input passed to libcurl with CURLOPT_USERPWD and
CURLOPT_PROXYUSERPWD circumvents the regular string length check we have
in Curl_setstropt(), the input length limit is enforced in
Curl_parse_login_details too, separately.
Reported-by: Thomas Bouzerar
Closes#5383
When looking for a protocol match among supported schemes, check the
most "popular" schemes first. It has zero functionality difference and
for all practical purposes a speed difference will not be measureable
but it still think it makes sense to put the least likely matches last.
"Popularity" based on the 2019 user survey.
Closes#5377
Add three new CMake Find modules (using the curl license, but I grant
others the right to apply the CMake BSD license instead).
This CMake config is simpler than the autotools one because it assumes
ngtcp2 and nghttp3 to be used together. Another difference is that this
CMake config checks whether QUIC is actually supported by the TLS
library (patched OpenSSL or boringssl) since this can be a common
configuration mistake that could result in build errors later.
Unlike autotools, CMake does not warn you that the features are
experimental. The user is supposed to already know that and read the
documentation. It requires a very special build environment anyway.
Tested with ngtcp2+OpenSSL+nghttp3 and quiche+boringssl, both built from
current git master. Use `LD_DEBUG=files src/curl |& grep need` to figure
out which features (libldap-2.4, libssh2) to disable due to conflicts
with boringssl.
Closes#5359
If the QLOGDIR environment variable is set, enable qlogging.
... and create Curl_qlogdir() in the new generic vquic/vquic.c file for
QUIC functions that are backend independent.
Closes#5353
That return code is reserved for build-time conditional code not being
present while this was a regular run-time error from a Windows API.
Reported-by: wangp on github
Fixes#5349Closes#5350
Triggered by a crash detected by OSS-Fuzz after the dynbuf introduction in
ed35d6590e. This should make the trailer handling more straight forward and
hopefully less error-prone.
Deliver the trailer header to the callback already at receive-time. No
longer caches the trailers to get delivered at end of stream.
Bug: https://bugs.chromium.org/p/oss-fuzz/issues/detail?id=22030Closes#5348
In my very basic test that lists sftp://127.0.0.1/tmp/, this patched
code makes 161 allocations compared to 194 in git master. A 17%
reduction.
Closes#5336
quiche has the potential to log qlog files. To enable this, you must
build quiche with the qlog feature enabled `cargo build --features
qlog`. curl then passes a file descriptor to quiche, which takes
ownership of the file. The FD transfer only works on UNIX.
The convention is to enable logging when the QLOGDIR environment is
set. This should be a path to a folder where files are written with the
naming template <SCID>.qlog.
Co-authored-by: Lucas Pardue
Replaces #5337Closes#5341
A common set of functions instead of many separate implementations for
creating buffers that can grow when appending data to them. Existing
functionality has been ported over.
In my early basic testing, the total number of allocations seem at
roughly the same amount as before, possibly a few less.
See docs/DYNBUF.md for a description of the API.
Closes#5300
- Check for NULL entry parameter before attempting to deref entry in
Curl_resolver_is_resolved, like is already done in asyn-ares.
This is to silence cppcheck which does not seem to understand that
asyn-ares and asyn-thread have separate Curl_resolver_is_resolved
and those units are mutually exclusive. Prior to this change it warned
of a scenario where asyn-thread's Curl_resolver_is_resolved is called
with a NULL entry from asyn-ares, but that couldn't happen.
Reported-by: rl1987@users.noreply.github.com
Fixes https://github.com/curl/curl/issues/5326
More connection cache accesses are protected by locks.
CONNCACHE_* is a beter prefix for the connection cache lock macros.
Curl_attach_connnection: now called as soon as there's a connection
struct available and before the connection is added to the connection
cache.
Curl_disconnect: now assumes that the connection is already removed from
the connection cache.
Ref: #4915Closes#5009
Regression since 7.69.0 and 68fb25fa3f.
The code wrongly assigned 'from' instead of 'auth' which probably was a
copy and paste mistake from other code, leading to that auth could
remain NULL and later cause an error to be returned.
Assisted-by: Eric Sauvageau
Fixes#5294Closes#5295
Previously, options set explicitly through command line options could be
overridden by the configuration files parsed automatically when
ssh_connect() was called.
By calling ssh_options_parse_config() explicitly, the configuration
files are parsed before setting the options, avoiding the options
override. Once the configuration files are parsed, the automatic
configuration parsing is not executed.
Fixes#4972Closes#5283
Signed-off-by: Anderson Toshiyuki Sasaki <ansasaki@redhat.com>
Coverity found CID 1461718:
Integer handling issues (CONSTANT_EXPRESSION_RESULT) "timeout_ms >
9223372036854775807L" is always false regardless of the values of its
operands. This occurs as the logical second operand of "||".
Closes#5240
Prior to this change if there was a 303 reply to a PUT request then
the subsequent request to respond to that redirect would also be a PUT.
It was determined that was most likely incorrect based on the language
of the RFCs. Basically 303 means "see other" resource, which implies it
is most likely not the same resource, therefore we should not try to PUT
to that different resource.
Refer to the discussions in #5237 and #5248 for more information.
Fixes https://github.com/curl/curl/issues/5237
Closes https://github.com/curl/curl/pull/5248
GnuTLS 3.1.10 added new functions we want to use. That version was
released on Mar 22, 2013. Removing support for older versions also
greatly simplifies the code.
Ref: #5271Closes#5276
Detected by Coverity. CID 1462319.
"The same code is executed when the condition result is true or false,
because the code in the if-then branch and after the if statement is
identical."
Closes#5275
When cURL is compiled with support for multiple SSL backends, it is
possible to configure an SSL backend via `curl_global_sslset()`, but
only *before* `curl_global_init()` was called.
If another SSL backend should be used after that, a user might be
tempted to call `curl_global_cleanup()` to start over. However, we did
not foresee that use case and forgot to reset the SSL backend in that
cleanup.
Let's allow that use case.
Fixes#5255Closes#5257
Reported-by: davidedec on github
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
From libssh 0.9.0, ssh_key_type() returns different key types for ECDSA
keys depending on the curve.
Signed-off-by: Anderson Toshiyuki Sasaki <ansasaki@redhat.com>
Fixes#5252Closes#5253
- Fix schannel_send for the case when no timeout was set.
Prior to this change schannel would error if the socket was not ready
to send data and no timeout was set.
This commit is similar to parent commit 89dc6e0 which recently made the
same change for SOCKS, for the same reason. Basically it was not well
understood that when Curl_timeleft returns 0 it is not a timeout of 0 ms
but actually means no timeout.
Fixes https://github.com/curl/curl/issues/5177
Closes https://github.com/curl/curl/pull/5221
- Document in Curl_timeleft's comment block that returning 0 signals no
timeout (ie there's infinite time left).
- Fix SOCKS' Curl_blockread_all for the case when no timeout was set.
Prior to this change if the timeout had a value of 0 and that was passed
to SOCKET_READABLE it would return right away instead of blocking. That
was likely because it was not well understood that when Curl_timeleft
returns 0 it is not a timeout of 0 ms but actually means no timeout.
Ref: https://github.com/curl/curl/pull/5214#issuecomment-612512360
Closes https://github.com/curl/curl/pull/5220
Prior to this change gopher's blocking code would block forever,
ignoring any set timeout value.
Assisted-by: Jay Satiro
Reviewed-by: Daniel Stenberg
Similar to #5220 and #5221Closes#5214
When SRP is requested in the priority string, GnuTLS will disable
support for TLS 1.3. Before this change, curl would always add +SRP to
the priority list, effectively always disabling TLS 1.3 support.
With this change, +SRP is only added to the priority list when SRP
authentication is also requested. This also allows updating the error
handling here to not have to retry without SRP. This is because SRP is
only added when requested and in that case a retry is not needed.
Closes#5223
- If loss of data may occur converting a timediff_t to time_t and
the time value is > TIME_T_MAX then treat it as TIME_T_MAX.
This is a follow-up to 8843678 which removed the (time_t) typecast
from the macros so that conversion warnings could be identified.
Closes https://github.com/curl/curl/pull/5199
In a debug build, settting the environment variable "CURL_SMALLREQSEND"
will make the first HTTP request send not send more bytes than the set
amount, thus ending up verifying that the logic for handling a split
HTTP request send works correctly.
Restores the --head functionality to the curl utility which extracts
'protocol' that is stored that way.
Reported-by: James Fuller
Fixes#5196Closes#5198
In libcurl, CURLINFO_CONDITION_UNMET is used to avoid writing to the
output file if the server did not transfered a file based on time
condition. In the same manner, getting a 304 HTTP response back from the
server, for example after passing a custom If-Match-* header, also
fulfill this condition.
Fixes#5181Closes#5183
Currently, the TLS backend used by vquic/ngtcp2.c is selected at compile
time. Therefore OpenSSL support needs to be explicitly disabled.
Signed-off-by: Daiki Ueno <dueno@redhat.com>
Closes#5148
This updates the ngtcp2 OpenSSL backend to follow the API change in
commit 32e703164 of ngtcp2.
Notable changes are:
- ngtcp2_crypto_derive_and_install_{rx,tx}_key have been added to replace
ngtcp2_crypto_derive_and_install_key
- the 'side' argument of ngtcp2_crypto_derive_and_install_initial_key
has been removed
Fixes#5166Closes#5168
OpenSSL 3 deprecates SSL_CTX_load_verify_locations and the MD4, DES
functions we use.
Fix the MD4 and SSL_CTX_load_verify_locations warnings.
In configure, detect OpenSSL v3 and if so, inhibit the deprecation
warnings. OpenSSL v3 deprecates the DES functions we use for NTLM and
until we rewrite the code to use non-deprecated functions we better
ignore these warnings as they don't help us.
Closes#5139
Reported by the new script 'scripts/copyright.pl'. The script has a
regex whitelist for the files that don't need copyright headers.
Removed three (mostly usesless) README files from docs/
Closes#5141
.. because not all Windows build systems have those symbols, and even
those that do may be missing newer symbols (eg the Windows 7 SDK does
not define _WIN32_WINNT_WIN10).
Those symbols are used in build-time logic to decide which API to use
and prior to this change if the symbols were missing it would have
resulted in deprecated API being used when more recent functions were
available (eg GetVersionEx used instead of VerifyVersionInfo).
Reported-by: FuccDucc@users.noreply.github.com
Probably fixes https://github.com/curl/curl/issues/4995
Closes https://github.com/curl/curl/pull/5057
Prior to this change in libcurl debug builds http2 stream closure was
erroneously referred to as connection closure.
Before:
* nread <= 0, server closed connection, bailing
After:
* nread == 0, stream closed, bailing
Closes https://github.com/curl/curl/pull/5118
- Implement new option CURLSSLOPT_REVOKE_BEST_EFFORT and
--ssl-revoke-best-effort to allow a "best effort" revocation check.
A best effort revocation check ignores errors that the revocation check
was unable to take place. The reasoning is described in detail below and
discussed further in the PR.
---
When running e.g. with Fiddler, the schannel backend fails with an
unhelpful error message:
Unknown error (0x80092012) - The revocation function was unable
to check revocation for the certificate.
Sadly, many enterprise users who are stuck behind MITM proxies suffer
the very same problem.
This has been discussed in plenty of issues:
https://github.com/curl/curl/issues/3727,
https://github.com/curl/curl/issues/264, for example.
In the latter, a Microsoft Edge developer even made the case that the
common behavior is to ignore issues when a certificate has no recorded
distribution point for revocation lists, or when the server is offline.
This is also known as "best effort" strategy and addresses the Fiddler
issue.
Unfortunately, this strategy was not chosen as the default for schannel
(and is therefore a backend-specific behavior: OpenSSL seems to happily
ignore the offline servers and missing distribution points).
To maintain backward-compatibility, we therefore add a new flag
(`CURLSSLOPT_REVOKE_BEST_EFFORT`) and a new option
(`--ssl-revoke-best-effort`) to select the new behavior.
Due to the many related issues Git for Windows and GitHub Desktop, the
plan is to make this behavior the default in these software packages.
The test 2070 was added to verify this behavior, adapted from 310.
Based-on-work-by: georgeok <giorgos.n.oikonomou@gmail.com>
Co-authored-by: Markus Olsson <j.markus.olsson@gmail.com>
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Closes https://github.com/curl/curl/pull/4981
- If an easy handle is owned by a multi different from the one specified
then return CURLM_BAD_EASY_HANDLE.
Prior to this change I assume user error could cause corruption.
Closes https://github.com/curl/curl/pull/5116
Makes curl_easy_getinfo() of "variable" numerical content instead return
the number set in the env variable `CURL_TIME`.
Makes curl_version() of "variable" textual content. This guarantees a
stable version string which can be tested against. Environment variable
`CURL_VERSION` defines the content.
Assisted-by: Mathias Gumz
This commit adds support to generate JSON via the writeout feature:
-w "%{json}"
It leverages the existing infrastructure as much as possible. Thus,
generating the JSON on STDERR is possible by:
-w "%{stderr}%{json}"
This implements a variant of
https://github.com/curl/curl/wiki/JSON#--write-out-json.
Closes#4870
When libcurl retries a connection due to it being "seemingly dead" or by
REFUSED_STREAM, it will now only do it up five times before giving up,
to avoid never-ending loops.
Reported-by: Dima Tisnek
Bug: https://curl.haxx.se/mail/lib-2020-03/0044.htmlCloses#5074
Make sure each separate index in connn->tempaddr[] is used for a fixed
family (and only that family) during the connection process.
If family one takes a long time and family two fails immediately, the
previous logic could misbehave and retry the same family two address
repeatedly.
Reported-by: Paul Vixie
Reported-by: Jay Satiro
Fixes#5083Fixes#4954Closes#5089
- Ignore CURLE_NOT_BUILT_IN errors returned by c-ares functions in
curl_easy_duphandle.
Prior to this change if c-ares was used as the resolver backend and
either it was too old or libcurl was built without IPv6 support then
some of our resolver functions could return CURLE_NOT_BUILT_IN to
curl_easy_duphandle causing it to fail.
Caused by c8f086b which shipped in 7.69.1.
Reported-by: Karl Chen
Fixes https://github.com/curl/curl/issues/5097
Closes https://github.com/curl/curl/pull/5100
1. The socks4 state machine was broken in the host resolving phase
2. The code now insists on IPv4-only when using SOCKS4 as the protocol
only supports that.
Regression from #4907 and 4a4b63d, shipped in 7.69.0
Reported-by: amishmm on github
Bug: https://github.com/curl/curl/issues/5053#issuecomment-596191594Closes#5061
New test 666 checks this is effective.
As upload buffer size is significant in this kind of tests, shorten it
in similar test 652.
Fixes#4860Closes#4833
Reported-by: RuurdBeerstra on github
Input buffer filling may delay the data sending if data reads are slow.
To overcome this problem, file and callback data reads do not accumulate
in buffer anymore. All other data (memory data and mime framing) are
considered as fast and still concatenated in buffer.
As this may highly impact performance in terms of data overhead, an early
end of part data check is added to spare a read call.
When encoding a part's data, an encoder may require more bytes than made
available by a single read. In this case, the above rule does not apply
and reads are performed until the encoder is able to deliver some data.
Tests 643, 644, 645, 650 and 654 have been adapted to the output data
changes, with test data size reduced to avoid the boredom of long lists of
1-byte chunks in verification data.
New test 667 checks mimepost using single-byte read callback with encoder.
New test 668 checks the end of part data early detection.
Fixes#4826
Reported-by: MrdUkk on github
In case a read callback returns a status (pause, abort, eof,
error) instead of a byte count, drain the bytes read so far but
remember this status for further processing.
Takes care of not losing data when pausing, and properly resume a
paused mime structure when requested.
New tests 670-673 check unpausing cases, with easy or multi
interface and mime or form api.
Fixes#4813
Reported-by: MrdUkk on github
With c-ares the dns parameters lives in ares_channel. Store them in the
curl handle and set them again in easy_duphandle.
Regression introduced in #3228 (6765e6d), shipped in curl 7.63.0.
Fixes#4893Closes#5020
Signed-off-by: Ernst Sjöstrand <ernst.sjostrand@verisure.com>
- Don't check errno on wakeup socket if sread returned 0 since sread
doesn't set errno in that case.
This is a follow-up to cf7760a from several days ago which fixed
Curl_multi_wait to stop busy looping sread on the non-blocking wakeup
socket if it was closed (ie sread returns 0). Due to a logic error it
was still possible to busy loop in that case if errno == EINTR.
Closes https://github.com/curl/curl/pull/5047
As we have logic that checks if we get a >= 400 reponse code back before
the upload is done, which then got confused since it wasn't "done" but
yet there was no data to send!
Reported-by: IvanoG on github
Fixes#4996Closes#5002
