To better allow arguments like "1 to 9999" without flooding the terminal
with error messages, the given test cases range is now checked and only
test numbers with existing files are actually run.
The previous test certificate contained a MD5 hash which is not
supported using TLSv1.2 with Schannel on Windows 7 or newer.
See the update to this blog post on IEInternals / MSDN:
http://blogs.msdn.com/b/ieinternals/archive/2011/03/25/
misbehaving-https-servers-impair-tls-1.1-and-tls-1.2.aspx
"Update: If the server negotiates a TLS1.2 connection with a
Windows 7 or 8 schannel.dll-using client application, and it
provides a certificate chain which uses the (weak) MD5 hash
algorithm, the client will abort the connection (TCP/IP FIN)
upon receipt of the certificate."
When allowing NTLM, the re-use connection logic was too focused on
finding an existing NTLM connection to use and didn't properly allow
re-use of other ones. This made the logic not re-use perfectly re-usable
connections.
Added test case 1418 and 1419 to verify.
Regression brought in 8ae35102c (curl 7.35.0)
Reported-by: Jeff King
Bug: http://thread.gmane.org/gmane.comp.version-control.git/242213
This one is needed with the gcc options -fstack-protector-all -O2
That brings the number of suppressions for test 165 to four, and I
suspect I could find another two missing without trying very hard. I'm
beginning to think suppressions isn't the best way to handle these
kinds of cases.
Do not try to convert line-endings to CRLF on Windows by setting stdout
to binary mode, just like the curl tool does if --ascii is not specified.
This should prevent corrupted stdout line-ending output like CRCRLF.
In order to make the previously naive text-aware tests work with
binary mode on Windows, text-mode is disabled for them if it is not
actually part of the test case and line-endings are corrected.
According to RFC 2616 and RFC 2326 individual protocol elements, like
headers and except the actual content, are terminated by using CRLF.
Therefore the test data files for these protocols need to contain
mixed line-endings if the actual protocol elements use CRLF while
the file uses LF.
gcc 4.7.2 with -O2 will optimize Curl_connect by inlining some
functions two levels deep, which makes the valgrind suppression
fail to match. The underlying reason for these idna suppressions is
a gcc strlen optimization when compiling libidn; compiling it with
-fno-builtin-strlen makes this suppression unnecessary.
It seems the fips config option causes an error if FIPS mode was
not enabled at stunnel compile-time. FIPS support was disabled
by default in stunnel 5.00, so this is probably really only needed
on versions between 4.32 and 5.00.
This was already mostly being done, except that analysis after the
test still assumed that the valgrind log files would be available. An
alternative way to handle the valgrind + gdb combination could be to
enable one of the valgrind debugger hooks.
lib1515.c:38:26 warning: unused parameter 'curl'
lib1515.c:38:81 warning: unused parameter 'ptr'
lib1515.c:38:5 warning: no previous prototype for 'debug_callback'
lib1515.c:46:5 warning: no previous prototype for 'do_one_request'
lib1515.c:120:3 warning: ISO C90 forbids mixed declarations and code
As well as some code policing such as white space and braces.
Not comma, which is an inconsistency and a mistake probably inherited
from the examples section of RFC1867.
This bug has been present since the day curl started to support
multipart formposts, back in the 90s.
Reported-by: Rob Davies
Bug: http://curl.haxx.se/bug/view.cgi?id=1333
Fix for bug #1303 (030a2b8cb) was not complete.
libcurl still pruned DNS entries added manually
after detecting a dead connection. This test
checks such behavior.
Test-case 1515 reproduces bug #1303, where libcurl
would incorrectly prune DNS entries added via
CURLOPT_RESOLVE after the DNS_CACHE_TIMEOUT had
expired.
The test contains a cookie jar file where one of the cookies has an
expiry date of 1391252187 -- Sat, 1 Feb 2014 10:56:27 GMT which has
now expired. Updated to Wed, 14 Oct 2037 16:36:33 GMT as per test
179.
Reported-by: Adam Sampson
Bug: http://curl.haxx.se/bug/view.cgi?id=1330
Since the timer resolution is lower, there are actually cases that
the compared values are equal. Therefore we check for previous
timestamps being greater than the current one instead.
According to section 2.2 of RFC959 the End-of-Line is defined as:
The end-of-line sequence defines the separation of printing
lines. The sequence is Carriage Return, followed by Line Feed.
Verified by sniffing traffic between a Windows FTP client (FileZilla)
and Unix-hosted FTP server (ProFTPD).
It makes more sense to convert the expected output to [CR][LF] on
Windows than to force the actual, probably correct, output to [LF].
This way it is actually possible to see if curl outputs the correct
line-ending excepted by a text-aware test case.
Since the previous complex select function with initial support for
non-socket file descriptors, did not actually work correctly for
Console handles, this change simplifies the whole procedure by using
an internal waiting thread for the stdin console handle.
The previous implementation made it continuously trigger for the stdin
handle if it was being redirected to a parent process instead of
an actual Console input window.
This approach supports actual Console input handles as well as
anonymous Pipe handles which are used during input redirection.
It depends on the fact that ReadFile supports trying to read zero bytes
which makes it wait for the handle to become ready for reading.
Removed Unix-specific functionality in order to support Windows:
- select.epoll replaced with select.select
- SocketServer.ForkingMixIn replaced with SocketServer.ForkingMixIn
- socket.MSG_DONTWAIT replaced with socket.setblocking(False)
Even though epoll has a better performance and improved socket handling
than select, this change should not affect the actual test case.
Also, make the ftp server return a canned response that doesn't
cause XML verification problems. Although the test file format
isn't technically XML, it's still handy to be able to use XML
tools to verify and manipulate them.
Since /dev/stdout is not always emulated on Windows,
just skip the output option on Windows.
MinGW/msys support /dev/stdout only from a new login shell.
tstunnel on Windows does not support the pid option and is unable
to write to an output log that is already being used as a redirection
target for stdout. Therefore it does now output all log data to stdout
by default and secureserver.pl creates a fake pidfile on Windows.
The built-in memory debug system doesn't work with multi-threaded use so
instead of causing annoying false positives, disable the memory tracking
if the threaded resolver is used.
The Windows console version of stunnel is called "tstunnel", while
running "stunnel" on Windows spawns a new console window which
cannot be handled by the testsuite.
Previously LIST always returned a fixed hardcoded list that the ftp
server code knew about, mostly since the server didn't get any test case
number in the LIST scenario. Starting now, doing a CWD to a directory
named test-[number] will make the test server remember that number and
consider it a test case so that a subsequent LIST command will send the
<data> section of that test case back.
It allows LIST tests to be made more similar to how all other tests
work.
Test 100 was updated to provide its own directory listing.
Verify the change brought in commit 8e11731653061. It makes sure that
returning a failure from the progress callback even very early results
in the correct return code.
memdebug.h already contains all required definitions and including
curl_memory.h causes errors like the following:
tests/unit/unit1394.c:119: undefined reference to `Curl_cfree'
tests/unit/unit1394.c:120: undefined reference to `Curl_cfree'
Following commit 0aafd77fa4, replaced the internal usage of
FORMAT_OFF_T and FORMAT_OFF_TU with the external versions that we
expect API programmers to use.
This negates the need for separate definitions which were subtly
different under different platforms/compilers.
Following the addition of informational commands to the SMTP protocol,
the test server is no longer required to return the verified server
information in responses that curl only outputs in verbose mode.
Instead, a similar detection mechanism to that used by FTP, IMAP and
POP3 can now be used.
This commit replaces that of 9f260b5d66 because according to RFC-2449,
section 6, there is no APOP capability "...even though APOP is an
optional command in [POP3]. Clients discover server support of APOP by
the presence in the greeting banner of an initial challenge enclosed in
angle brackets."
SASL downgrade tests: 833, 835, 879, 881, 935 and 937 would fail as
they contained a minus sign in their authentication mechanism and this
would be missed by the custom reply parser.
Added support for downgrading the SASL authentication mechanism when the
decoding of CRAM-MD5, DIGEST-MD5 and NTLM messages fails. This enhances
the previously added support for graceful cancellation by allowing the
client to retry a lesser SASL mechanism such as LOGIN or PLAIN, or even
APOP / clear text (in the case of POP3 and IMAP) when supported by the
server.
To avoid the regression when users pass in passwords containing semi-
colons, we now drop the ability to set the login options with the same
options. Support for login options in CURLOPT_USERPWD was added in
7.31.0.
Test case 83 was modified to verify that colons and semi-colons can be
used as part of the password when using -u (CURLOPT_USERPWD).
Bug: http://curl.haxx.se/bug/view.cgi?id=1311
Reported-by: Petr Bahula
Assisted-by: Steve Holme
Signed-off-by: Daniel Stenberg <daniel@haxx.se>
A failure during authentication, which is performed as part of the
CONNECT phrase (for IMAP, POP3 and SMTP) is considered by the multi-
interface as being closed prematurely (aka a dead connection). As such
these protocols cannot issue the relevant QUIT or LOGOUT command.
Temporarily fixed the test cases until we can fix this properly.
The error code should not be sent as data as it isn't passed onto the
client as body data, so cannot be compared in the test suite against
expected data.
Although this option should have already been set, the SMTP module can
now download information from and send instructional commands to, an
SMTP server, requiring the option to be set in order to perform a mail
transfer.
As the IMAP regex could fail and $1 would not contain the command id
updated the unrecognised command response to be more generic and
realistic (like those used in the command handlers).
Additionally updated the POP3, SMTP and FTP responses.
A base64 string should be a multiple of 4 characters in length, not
contain any more than 2 padding characters and only contain padding
characters at the end of string. For example: Y3VybA==
Strings such as the following are considered invalid:
Y= - Invalid length
Y== - Invalid length
Y=== - More than two padding characters
Y=x= - Padding character contained within string
This is a regression since the switch to always-multi internally
c43127414d.
Test 1316 was modified since we now clearly call the Curl_client_write()
function when doing the LIST transfer part and then the
handler->protocol says FTP and ftpc.transfertype is 'A' which implies
text converting even though that the response is initially a HTTP
CONNECT response in this case.
As the URI, which is contained within the DIGEST-MD5 response, is
constructed from the service and realm, the encoded message differs
from that generated under POP3.
...as it is no longer required following capability and authentication
changes and is now causing problems following commit 49341628b5 as
the test number is obtained from the client address in the EHLO.
...to the client address as this frees the RCPT strings to contain
just an email address and by passing the test number into curl as the
client address remains consistent with POP3 and IMAP tests as they are
specified in the URL.
As someone reported on the mailing list a while back, the hard-coded
arbitrary timeout of 7s in test 1112 is not sufficient in some build
environments. At Arista Networks we build and test curl as part of our
automated build system, and we've run into this timeout 170 times so
far. Our build servers are typically quite busy building and testing a
lot of code in parallel, so despite being beefy machines with 32 cores
and 128GB of RAM we still hit this 7s timeout regularly.
URL: http://curl.haxx.se/mail/lib-2010-02/0200.html
As curl will send a HELO command after an negative EHLO response, added
the same detection from commit b07709f741 to the HELO handler to
ensure the test server is identified correctly and an upload isn't
performed.
Rather than detecting the TO address as missing in the DATA handler,
moved the detection to the RCPT command handler where an error response
can be generated.
Moved the special SMTP server detection code from the DATA command
handler, which happens further down the operation chain after EHLO,
MAIL and RCPT commands, to the EHLO command as it is the first command
to be generated by a SMTP operation as well as containing the special
"verifiedserver" string from the URL.
This not only makes it easier and quicker to detect but also means that
cURL doesn't need to specify "verifiedserver" as --mail-from and
--mail-rcpt arguments.
More importantly, this also makes the upcoming verification changes to
the RCPT handler easier to implement.
The libcurl date parser returns INT_MAX for all dates > 2037 so this
test is now made to use 2037 instead of 2038 to work the same for both
32bit and 64bit time_t systems.
Following changes to ftpserver.pl fixed the mail from address to be a
correctly formatted address otherwise the server response will be 501
Invalid address.
MAIl_smtp() will now check for a correctly formatted FROM address as
well as the optional SIZE parameter comparing it against the server
capability when specified.
Implement: Expired Cookies These following situation, curl removes
cookie(s) from struct CookieInfo if the cookie expired.
- Curl_cookie_add()
- Curl_cookie_getlist()
- cookie_output()
Renamed SUPPORTAUTH to AUTH and added support for specifying a list of
supported SASL mechanisms to return to the client.
Additionally added the directive to the FILEFORMAT document.
Renamed SUPPORTCAPA to CAPA and added support for specifying a list of
supported capabilities to return to the client.
Additionally added the directive to the FILEFORMAT document.
The message numbers given in the LIST response are an index into the
list, which are only valid for the current session, rather than being a
unique message identifier. An index would only be missing from the LIST
response if a DELE command had been issued within the same session and
had not been committed by the end of session QUIT command. Once
committed the POP3 server will regenerate the message numbers in the
next session to be contiguous again. As such our LIST response should
list message numbers contiguously until we support a DELE command in the
same session.
Should a POP3 user require the unique message ID for any or all
messages then they should use the extended UIDL command. This command
will be supported by the test ftpserver in an upcoming commit.
Corrected the call to logmsg() in the IMAP SEARCH handler from commit
4ae7b7ea69 as it should have been outputting the what argument and
not the test number.
The specified curl binary will then be used to verify the running
server(s) instead of the development version. This is very useful in
some cases when the development version fails to verify correctly as
then the test case may not run at all.
The actual test will still be run with the "normal" curl executable
(unless the test case specifies something differently).
... this also makes sure that the progess callback gets called more
often during TFTP transfers.
Added test 1238 to verify.
Bug: http://curl.haxx.se/bug/view.cgi?id=1269
Reported-by: Jo3
This function is meant to work *exactly* as curl_easy_perform() but will
use the event-based libcurl API internally instead of
curl_multi_perform(). To avoid relying on an actual event-based library
and to not use non-portable functions (like epoll or similar), there's a
rather inefficient emulation layer implemented on top of Curl_poll()
instead.
There's currently some convenience logging done in curl_easy_perform_ev
which helps when tracking down problems. They may be suitable to remove
or change once things seem to be fine enough.
curl has a new --test-event option when built with debug enabled that
then uses curl_easy_perform_ev() instead of curl_easy_perform(). If
built without debug, using --test-event will only output a warning
message.
NOTE: curl_easy_perform_ev() is not part if the public API on purpose.
It is only present in debug builds of libcurl and MUST NOT be considered
stable even then. Use it for libcurl-testing purposes only.
runtests.pl now features an -e command line option that makes it use
--test-event for all curl command line tests. The man page is updated.
libcurl truncates usernames and passwords it reads from .netrc to
LOGINSIZE and PASSWORDSIZE (64) characters without any indication to
the user, to ensure the values returned from Curl_parsenetrc fit in a
caller-provided buffer.
Fix the interface by passing back dynamically allocated buffers
allocated to fit the user's input. The parser still relies on a
256-character buffer to read each line, though.
So now you can include an ~246-character password in your .netrc,
instead of the previous limit of 63 characters.
Reported-by: Colby Ranger
The new multiply() function detects range value overflows. 32bit
machines will overflow on a 32bit boundary while 64bit hosts support
ranges up to the full 64 bit range.
Added test 1236 to verify.
Bug: http://curl.haxx.se/bug/view.cgi?id=1267
Reported-by: Will Dietz
A rather big overhaul and cleanup.
1 - curl wouldn't properly detect and reject globbing that ended with an
open brace if there were brackets or braces before it. Like "{}{" or
"[0-1]{"
2 - curl wouldn't properly reject empty lists so that "{}{}" would
result in curl getting (nil) strings in the output.
3 - By using strtoul() instead of sscanf() the code will now detected
over and underflows. It now also better parses the step argument to only
accept positive numbers and only step counters that is smaller than the
delta between the maximum and minimum numbers.
4 - By switching to unsigned longs instead of signed ints for the
counters, the max values for []-ranges are now very large (on 64bit
machines).
5 - Bumped the maximum number of globs in a single URL to 100 (from 10)
6 - Simplified the code somewhat and now it stores fixed strings as
single- entry lists. That's also one of the reasons why I did (5) as now
all strings between "globs" will take a slot in the array.
Added test 1234 and 1235 to verify. Updated test 87.
This commit fixes three separate bug reports.
Bug: http://curl.haxx.se/bug/view.cgi?id=1264
Bug: http://curl.haxx.se/bug/view.cgi?id=1265
Bug: http://curl.haxx.se/bug/view.cgi?id=1266
Reported-by: Will Dietz
CURLOPT_DNS_USE_GLOBAL_CACHE broke in commit c43127414d (been
broken since the libcurl 7.29.0 release). While this option has been
documented as deprecated for almost a decade and nobody even reported
this bug, it should remain functional.
Added test case 1512 to verify
This is a regression as this logic used to work. It isn't clear when it
broke, but I'm assuming in 7.28.0 when we went all-multi internally.
This likely never worked with the multi interface. As the failed
connection is detected once the multi state has reached DO_MORE, the
Curl_do_more() function was now expanded somewhat so that the
ftp_do_more() function can request to go "back" to the previous state
when it makes another attempt - using PASV.
Added test case 1233 to verify this fix. It has the little issue that it
assumes no service is listening/accepting connections on port 1...
Reported-by: byte_bucket in the #curl IRC channel
The internal function that's used to detect known file extensions for
the default Content-Type got the the wrong pointer passed in when
CURLFORM_BUFFER + CURLFORM_BUFFERPTR were used. This had the effect that
strlen() would be used which could lead to an out-of-bounds read (and
thus segfault). In most cases it would only lead to it not finding or
using the correct default content-type.
It also showed that test 554 and test 587 were testing for the
previous/wrong behavior and now they're updated as well.
Bug: http://curl.haxx.se/bug/view.cgi?id=1262
Reported-by: Konstantin Isakov
This reverts commit 7ed25cc, reinstating commit 8ec2cb5.
As of 18-jul-2013 we still do have code in libcurl that makes use of these
memory functions. Commit 8ec2cb5 comment still applies and is yet valid.
These memory functions are solely used in Windows builds, so all related
code is protected with '#ifdef WIN32' preprocessor conditional compilation
directives.
Specifically, wcsdup() _wcsdup() are used when building a Windows target with
UNICODE and USE_WINDOWS_SSPI preprocessor symbols defined. This is the case
when building a Windows UNICODE target with Windows native SSL/TLS support
enabled.
Realizing that wcsdup() _wcsdup() are used is a bit tricky given that usage
of these is hidden behind _tcsdup() which is MS way of dealing with code
that must tolerate UNICODE and non-UNICODE compilation. Additionally, MS
header files and those compatible from other compilers use this preprocessor
conditional compilation directive in order to select at compilation time
whether 'wide' or 'ansi' MS API functions are used.
Without this code, Windows build targets with Windows native SSL/TLS support
enabled and MemoryTracking support enabled misbehave in tracking memory usage,
regardless of being a UNICODE enabled build or not.
Fixed issue with static build for MSVC2010.
After some investigation I've discovered known issue
http://public.kitware.com/Bug/view.php?id=11240 When .rc file is linked
to static lib it fails with following linker error
LINK : warning LNK4068: /MACHINE not specified; defaulting to X86
file.obj : fatal error LNK1112: module machine type 'x64' conflicts with
target machine type 'X86'
Fix add target property /MACHINE: for MSVC generation.
Also removed old workarounds - it caused errors during msvc build.
Bug: http://curl.haxx.se/mail/lib-2013-07/0046.html
When doing multi-part formposts, libcurl used a pseudo-random value that
was seeded with time(). This turns out to be bad for users who formpost
data that is provided with users who then can guess how the boundary
string will look like and then they can forge a different formpost part
and trick the receiver.
My advice to such implementors is (still even after this change) to not
rely on the boundary strings being cryptographically strong. Fix your
code and logic to not depend on them that much!
I moved the Curl_rand() function into the sslgen.c source file now to be
able to take advantage of the SSL library's random function if it
provides one. If not, try to use the RANDOM_FILE for seeding and as a
last resort keep the old logic, just modified to also add microseconds
which makes it harder to properly guess the exact seed.
The formboundary() function in formdata.c is now using 64 bit entropy
for the boundary and therefore the string of dashes was reduced by 4
letters and there are 16 hex digits following it. The total length is
thus still the same.
Bug: http://curl.haxx.se/bug/view.cgi?id=1251
Reported-by: "Floris"
RFC3986 details how a path part passed in as part of a URI should be
"cleaned" from dot sequences before getting used. The described
algorithm is now implemented in lib/dotdot.c with the accompanied test
case in test 1395.
Bug: http://curl.haxx.se/bug/view.cgi?id=1200
Reported-by: Alex Vinnik
The initial fix to only compare full path names were done in commit
04f52e9b4d but found out to be incomplete. This takes should make the
change more complete and there's now two additional tests to verify
(test 31 and 62).
After curl_multi_wait() returns, this test checked that we got exactly
one file descriptor told to read from, but we cannot be sure that is
true. curl_multi_wait() will sometimes return earlier without any file
descriptor to handle, just just because it is a suitable time to call
*perform().
This problem showed up with commit 29bf0598.
Bug: http://curl.haxx.se/mail/lib-2013-06/0029.html
Reported-by: Fabian Keil
If an ipv6-address is provided to CONNECT, the last hexadecimal group in
the address will be used as the test number! For example the address
"[1234::ff]" would be treated as test case 255.
When sending the HTTP Authorization: header for digest, the user name
needs to be escaped if it contains a double-quote or backslash.
Test 1229 was added to verify
Reported and fixed by: Nach M. S
Bug: http://curl.haxx.se/bug/view.cgi?id=1230
I found a bug which cURL sends cookies to the path not to aim at.
For example:
- cURL sends a request to http://example.fake/hoge/
- server returns cookie which with path=/hoge;
the point is there is NOT the '/' end of path string.
- cURL sends a request to http://example.fake/hogege/ with the cookie.
The reason for this old "feature" is because that behavior is what is
described in the original netscape cookie spec:
http://curl.haxx.se/rfc/cookie_spec.html
The current cookie spec (RFC6265) clarifies the situation:
http://tools.ietf.org/html/rfc6265#section-5.2.4
This reverts commit 8ec2cb5544.
We don't have any code anywhere in libcurl (or the curl tool) that use
wcsdup so there's no such memory use to track. It seems to cause mild
problems with the Borland compiler though that we may avoid by reverting
this change again.
Bug: http://curl.haxx.se/mail/lib-2013-05/0070.html
If the mail sent during the transfer contains a terminating <CRLF> then
we should not send the first <CRLF> of the EOB as specified in RFC-5321.
Additionally don't send the <CRLF> if there is "no mail data" as the
DATA command already includes it.
The emails that are sent to the server during these tests were
incorrectly formatted as they contained one or more LF terminated lines
rather than being CRLF terminated as per Section 2.3.7 of RFC-2821.
This wasn't a problem for the test suite as the <stdin> data matched the
<upload> data but anyone using these tests as reference would be sending
incorrect data to a server.
Updated test903 and test904 following the addition of CURLOPT_SASL_IR
as the default behaviour of SMTP AUTH responses is now to not include
the initial response. New tests with --sasl-ir support to follow.
When connecting back to an FTP server after having sent PASV/EPSV,
libcurl sometimes didn't use the proxy properly even though the proxy
was used for the initial connect.
The function wrongly checked for the CURLOPT_PROXY variable to be set,
which made it act wrongly if the proxy information was set with an
environment variable.
Added test case 711 to verify (based on 707 which uses --socks5). Also
added test712 to verify another variation of setting the proxy: with
--proxy socks5://
Bug: http://curl.haxx.se/bug/view.cgi?id=1218
Reported-by: Zekun Ni
I couldn't figure out why the host key logic isn't working, but having
it set to yes prevents my SSH-based test cases to run. I also don't see
a strong need to use strict host key checking on this test server.
So I disabled it.
...instead of the 220 we otherwise expect.
Made the ftpserver.pl support sending a custom "welcome" and then
created test 1219 to verify this fix with such a 230 welcome.
Bug: http://curl.haxx.se/mail/lib-2013-02/0102.html
Reported by: Anders Havn
Accessing a file with an absolute path in the root dir but with no
directory specified was not handled correctly. This fix comes with four
new test cases that verify it.
Bug: http://curl.haxx.se/mail/lib-2013-04/0142.html
Reported by: Sam Deane
... and make 1216 also verify it with a file input
These tests verify commit 3604fde3d3c9b0d, the fix for the "cookie
domain tailmatch" vulnerability. See
http://curl.haxx.se/docs/adv_20130412.html
Previously it only compared credentials if the requested needle
connection wasn't using a proxy. This caused NTLM authentication
failures when using proxies as the authentication code wasn't send on
the connection where the challenge arrived.
Added test 1215 to verify: NTLM server authentication through a proxy
(This is a modified copy of test 67)
WINSOCK only:
Since FD_CLOSE is only signaled once, it may trigger at the same
time as FD_READ. Data actually being available makes it impossible
to detect that the connection was closed by checking that recv returns
zero. Another recv attempt could block the connection if it was
not closed. This workaround abuses exceptfds in conjunction with
readfds to signal that the connection has actually closed.
Since commit 57aeabcc1a, it handles errors on the control connection
while waiting for the data connection better.
Test 591 and 592 are updated accordingly.
