Igor Novoseltsev reported a problem with the multi socket API and
using timeouts and timers. It boiled down to a problem with
libcurl's use of GetTickCount() interally to figure out the
current time, while Igor's own application code used another
function call.
It made his app call the socket API timeout function a bit
_before_ libcurl would consider the timeout to trigger, and that
could easily lead to timeouts or stalls in the app. It seems
GetTickCount() in general often has no better resolution than
16ms and switching to the alternative function
QueryPerformanceCounter has its share of problems:
http://www.virtualdub.org/blog/pivot/entry.php?id=106
We address this problem by simply having libcurl treat timers
that already has occured or will occur within 40ms subject for
treatment. I'm confident that there are other implementations and
operating systems with similarly in accurate timer functions so
it makes sense to have applied generically and I don't believe we
sacrifice much by adding a 40ms inaccuracy on these timeouts.
Hauke Duden provided an example program that made the multi
interface crash. His example simply used the multi interface and
did first one FTP transfer and after completion it used a second
easy handle and did another FTP transfer on the same FTP server.
This triggered a bug in the "delayed easy handle kill" system
that curl uses: when an FTP connection is left alive it must keep
an easy handle around internally - only for the purpose of having
an easy handle when it later disconnects it. The code assumed
that when the easy handle was removed and an internal reference
was made, that version could be killed later on when a new easy
handle came using the same connection. This was wrong as Hauke's
example showed that the removed handle wasn't killed for real
until later. This caused a double close attempt => segfault.
The problem mentioned on Dec 10 2009
(http://curl.haxx.se/bug/view.cgi?id=2905220) was only partially fixed.
Partially because an easy handle can be associated with many connections in
the cache (e.g. if there is a redirect during the lifetime of the easy
handle). The previous patch only cleaned up the first one. The new fix now
removes the easy handle from all connections, not just the first one.
simply check for CURLM_CALL_MULTI_PERFORM internally. This has the added
benefit that this goes in line with my long-term wishes to get rid of the
CURLM_CALL_MULTI_PERFORM all together from the public API.
state, we return CURLM_CALL_MULTI_PERFORM unconditionally then so that we
can act faster like in the case the protocol-specific connect doesn't block
on anything and we can just persue on the next action immediately. It also
then avoids a case where curl_multi_fdset() would return -1.
accessing alredy freed memory and thus crash when using HTTPS (with
OpenSSL), multi interface and the CURLOPT_DEBUGFUNCTION and a certain order
of cleaning things up. I fixed it.
(http://curl.haxx.se/bug/view.cgi?id=2891591)
strdup() that could lead to segfault if it returned NULL. I extended his
suggest patch to now have Curl_retry_request() return a regular return code
and better check that.
Fix SIGSEGV on free'd easy_conn when pipe unexpectedly breaks
Fix data corruption issue with re-connected transfers
Fix use after free if we're completed but easy_conn not NULL
With the curl memory tracking feature decoupled from the debug build feature,
CURLDEBUG and DEBUGBUILD preprocessor symbol definitions are used as follows:
CURLDEBUG used for curl debug memory tracking specific code (--enable-curldebug)
DEBUGBUILD used for debug enabled specific code (--enable-debug)
KEEP_RECV to better match the general terminology: receive and send is what we
do from the (remote) servers. We read and write from and to the local fs.
(http://curl.haxx.se/bug/view.cgi?id=2784055) identifying a problem to
connect to SOCKS proxies when using the multi interface. It turned out to
almost not work at all previously. We need to wait for the TCP connect to
be properly verified before doing the SOCKS magic.
There's still a flaw in the FTP code for this.
FTP with the multi interface: when a transfer fails, like when aborted by a
write callback, the control connection was wrongly closed and thus not
re-used properly.
This change is also an attempt to cleanup the code somewhat in this area, as
now the FTP code attempts to keep (better) track on pending responses
necessary to get read in ftp_done().
When using the multi interface over HTTP and the server returns a Location
header, the running easy handle will get stuck in the CURLM_STATE_PERFORM
state, leaving the external event loop stuck waiting for data from the
ingoing socket (when using the curl_multi_socket_action stuff). While this
bug was pretty hard to find, it seems to require only a one-line fix. The
break statement on line 1374 in multi.c caused the function to skip the call
to multistate().
How to reproduce this bug? Well, that's another question. evhiperfifo.c in
the examples directory chokes on this bug only _sometimes_, probably
depending on how fast the URLs are added. One way of testing the bug out is
writing to hiper.fifo from more than one source at the same time.
pipelining, as libcurl could then easily get confused and A) work on the
handle that was not "first in queue" on a pipeline, or even B) tell the app
to REMOVE a socket while it was in use by a second handle in a pipeline. Both
errors caused hanging or stalling applications.
was actually ready to get done, as the internal time resolution is higher
than the returned millisecond timer. Therefore it could cause applications
running on fast processors to do short bursts of busy-loops.
curl_multi_timeout() will now only return 0 if the timeout is actually
alreay triggered.
now has an improved ability to do right when the multi interface (both
"regular" and multi_socket) is used for SCP and SFTP transfers. This should
result in (much) less busy-loop situations and thus less CPU usage with no
speed loss.
removing easy handles from multi handles when the easy handle is/was within
a HTTP pipeline. His bug report #2351653
(http://curl.haxx.se/bug/view.cgi?id=2351653) was also related and was
eventually fixed by a patch by Igor himself.
(http://curl.haxx.se/bug/view.cgi?id=2351645) that identified a problem with
the multi interface that occured if you removed an easy handle while in
progress and the handle was used in a HTTP pipeline.
eventually identified a flaw in how the multi_socket interface in some cases
missed to call the timeout callback when easy interfaces are removed and
added within the same millisecond.
libcurl to not tell the app properly when a socket was closed (when the name
resolve done by c-ares is done) and then immediately re-created and put to
use again (for the actual connection). Since the closure will make the
"watch status" get lost in several event-based systems libcurl will need to
tell the app about this close/re-create case.
the curl_multi_socket() API with HTTP pipelining enabled and could lead to
the pipeline basically stalling for a very long period of time until it took
off again.
go straight to DO
we had multiple states for which the internal function returned no socket at
all to wait for, with the effect that libcurl calls the socket callback (when
curl_multi_socket() is used) with REMOVE prematurely (as it would be added
again within very shortly)
redirections and thus cannot use CURLOPT_FOLLOWLOCATION easily, we now
introduce the new CURLINFO_REDIRECT_URL option that lets applications
extract the URL libcurl would've redirected to if it had been told to. This
then enables the application to continue to that URL as it thinks is
suitable, without having to re-implement the magic of creating the new URL
from the Location: header etc. Test 1029 verifies it.
such as the CURLOPT_SSL_CTX_FUNCTION one treat that as if it was a Location:
following. The patch that introduced this feature was done for 7.11.0, but
this code and functionality has been broken since about 7.15.4 (March 2006)
with the introduction of non-blocking OpenSSL "connects".
It was a hack to begin with and since it doesn't work and hasn't worked
correctly for a long time and nobody has even noticed, I consider it a very
suitable subject for plain removal. And so it was done.
CONNECT over a proxy. curl_multi_fdset() didn't report back the socket
properly during that state, due to a missing case in the switch in the
multi_getsock() function.
previously had a number of flaws, perhaps most notably when an application
fired up N transfers at once as then they wouldn't pipeline at all that
nicely as anyone would think... Test case 530 was also updated to take the
improved functionality into account.
is inited at the start of the DO action. I removed the Curl_transfer_keeper
struct completely, and I had to move out a few struct members (that had to
be set before DO or used after DONE) to the UrlState struct. The SingleRequest
struct is accessed with SessionHandle->req.
One of the biggest reasons for doing this was the bunch of duplicate struct
members in HandleData and Curl_transfer_keeper since it was really messy to
keep track of two variables with the same name and basically the same purpose!
do_init() and do_complete() which now are called first and last in the DO
function. It simplified the flow in multi.c and the functions got more
sensible names!
hash function for different hashes, and also expanded the default size for
the socket hash table used in multi handles to greatly enhance speed when
very many connections are added and the socket API is used.
when CURLM_CALL_MULTI_PERFORM is returned from curl_multi_socket*/perform,
to make applications using only curl_multi_socket() to properly function
when adding easy handles "on the fly". Bug report and test app provided by
Michael Wallner.
function that deprecates the curl_multi_socket() function. Using the new
function the application tell libcurl what action that was found in the
socket that it passes in. This gives a significant performance boost as it
allows libcurl to avoid a call to poll()/select() for every call to
curl_multi_socket*().
fixing some bugs:
o Don't mix GET and POST requests in a pipeline
o Fix the order in which requests are dispatched from the pipeline
o Fixed several curl bugs with pipelining when the server is returning
chunked encoding:
* Added states to chunked parsing for final CRLF
* Rewind buffer after parsing chunk with data remaining
* Moved chunked header initializing to a spot just before receiving
headers
the multi interface and connection re-use that could make a
curl_multi_remove_handle() ruin a pointer in another handle.
The second problem was less of an actual problem but more of minor quirk:
the re-using of connections wasn't properly checking if the connection was
marked for closure.
doing an FTP transfer is removed from a multi handle before completion. The
fix also fixed the "alive counter" to be correct on "premature removal" for
all protocols.
(http://curl.haxx.se/bug/view.cgi?id=1604956) which identified setting
CURLOPT_MAXCONNECTS to zero caused libcurl to SIGSEGV. Starting now, libcurl
will always internally use no less than 1 entry in the connection cache.
and while doing so it became apparent that the current timeout system for
the socket API really was a bit awkward since it become quite some work to
be sure we have the correct timeout set.
Jeff then provided the new CURLMOPT_TIMERFUNCTION that is yet another
callback the app can set to get to know when the general timeout time
changes and thus for an application like hiperfifo.c it makes everything a
lot easier and nicer. There's a CURLMOPT_TIMERDATA option too of course in
good old libcurl tradition.
would crash if a bad function sequence was used when shutting down after
using the multi interface (i.e using easy_cleanup after multi_cleanup) so
precautions have been added to make sure it doesn't any more - test case 529
was added to verify.
it basically was that we didn't remove the current connection from the pipe
list when following a redirect. Also in this commit: several cases of
additional debug code for debug builds helping to check and track down some
signs of run-time trouble.
handle that is part of a multi handle first removes the handle from the
stack.
- Added CURLOPT_SSL_SESSIONID_CACHE and --no-sessionid to disable SSL
session-ID re-use on demand since there obviously are broken servers out
there that misbehave with session-IDs used.
problem with it (SIGSEGV-style). It clearly showed that the existing
socket-state and state-difference function wasn't good enough so I rewrote
it and could then re-run Jeff's program without any crash. The previous
version clearly could miss to tell the application when a handle changed
from using one socket to using another.
While I was at it (as I could use this as a means to track this problem
down), I've now added a 'magic' number to the easy handle struct that is
inited at curl_easy_init() time and cleared at curl_easy_cleanup() time that
we can use internally to detect that an easy handle seems to be fine, or at
least not closed or freed (freeing in debug builds fill the area with 0x13
bytes but in normal builds we can of course not assume any particular data
in the freed areas).
we did wrong and patched it: When nodes were removed from the splay tree,
and we didn't properly remove it from the splay tree when an easy handle was
removed from a multi stack and thus we could wrongly leave a node in the
splay tree pointing to (bad) memory.
the crash was that libcurl internally was a bit confused about who owned the
DNS cache at all times so if you created an easy handle that uses a shared
DNS cache and added that to a multi handle it would crash. Now we keep more
careful internal track of exactly what kind of DNS cache each easy handle
uses: None, Private (allocated for and used only by this single handle),
Shared (points to a cache held by a shared object), Global (points to the
global cache) or Multi (points to the cache within the multi handle that is
automatically shared between all easy handles that are added with private
caches).
CURLOPT_MAX_RECV_SPEED_LARGE that limit tha maximum rate libcurl is allowed
to send or receive data. This kind of adds the the command line tool's
option --limit-rate to the library.
The rate limiting logic in the curl app is now removed and is instead
provided by libcurl itself. Transfer rate limiting will now also work for -d
and -F, which it didn't before.
multi stack and that easy handle had already been used to do one or more
easy interface transfers, as then the code threw away the previously used
DNS cache without properly freeing it.
(http://curl.haxx.se/bug/view.cgi?id=1431750) helped me identify and fix two
different but related bugs:
1) Removing an easy handle from a multi handle before the transfer is done
could leave a connection in the connection cache for that handle that is
in a state that isn't suitable for re-use. A subsequent re-use could then
read from a NULL pointer and segfault.
2) When an easy handle was removed from the multi handle, there could be an
outstanding c-ares DNS name resolve request. When the response arrived,
it caused havoc since the connection struct it "belonged" to could've
been freed already.
Now Curl_done() is called when an easy handle is removed from a multi handle
pre-maturely (that is, before the transfer was complteted). Curl_done() also
makes sure to cancel all (if any) outstanding c-ares requests.
an app can use to let libcurl only connect to a remote host and then extract
the socket from libcurl. libcurl will then not attempt to do any transfer at
all after the connect is done.
gets closed just after the request has been sent failed and did not re-issue
a request on a fresh reconnect like the easy interface did. Now it does!
(define CURL_MULTIEASY, run test case 160)
using a custom Host: header and curl fails to send a request on a re-used
persistent connection and thus creates a new connection and resends it. It
then sent two Host: headers. Cyrill's analysis was posted here:
http://curl.haxx.se/mail/archive-2005-01/0022.html
#1098843. In short, a shared DNS cache was setup for a multi handle and when
the shared cache was deleted before the individual easy handles, the latter
cleanups caused read/writes to already freed memory.
my_getaddrinfo() has been called (and wait has been set to TRUE), we check
if the name already is resolved and if so don't return wait status to the
parent. This can happen with IP-only names.
messages are stored, so that they don't have to be kept around for the multi
handle's entire life time. He also made it return failure codes properly
which it didn't do before.
I made the messages only get stored per easy-handle so that they can be
independently killed easier without ruining the "master list". It makes
the info_read() function slightly less beautiful as it has to scan for
messages to return, but it makes removing individual handles a lot easier
and less error prone.
- Use a global dns cache (via setting the tentatively named,
CURLOPT_DNS_USE_GLOBAL_CACHE option to true)
- Use a per-handle dns cache, by default
- Use a pooled dns cache when in the "multi" interface
