Dear Tobias,
Thank you for all the great information. This enabled me to
isolate the change which caused the issue.
So, for a bit of background, SWORD has no calls to mark critical
sections which might be problematic for re-entrant usage. This
has been due to the many implementations of threading across many
different platforms over the years, before C++11. But, as a
policy to support clients which desire to use SWORD in a
multithreaded manner, we do our best to make this safe by
advising clients to use separate SWMgr instances per thread.
There are still some shared objects in this scenario, but we do
our best to do all the writing to these shared objects upon
initialization. We broke this rule in commit 2760, which is what
caused your problem. SWORD have a facility to pool open file
handles, to help OSs which have small open file handle limits.
This work is done in FileMgr. Recently, to support Windows
Unicode path names (the commit you found which breaks your
multithreaded use), we rounded up all remaining native file IO
calls and replaced them to used FileMgr for the IO and then
extended FileMgr to handle Windows Unicode paths in a
Windows-specific manner. One of these changes was in
CURLFTPTransport, which is where you are having the issue. The
problem is that, where previously this class was directly opening
a FILE to do its writing, commit 2760 changed this to use FileMgr
to open the file, which involved the SWORD-wide file handle pool,
and since we are create a new file, we are always writing to this
shared pool container, which is not threadsafe. My guess is that
you have two threads trying to update the pool container at
exactly the same time. Using the file handle pool is usually
safe, because SWMgr "opens" all of its file handles on
initialization (these are not actually opening OS file handles,
but instead updating the file handle pool container with proxy
objects which delay actual OS open to on-demand, but the point is
this instance of shared file handle pool container writing is
done on creation of the SWMgr, afterward, the shared resource
file handle pool is only read and each object in the pool is
owned by only 1 thread if the "each thread must have its own
SWMgr" rule is followed.
Regardless of the details. I believe I have committed a fix for
you. In short, I have changed CURLFTPTransport to follow our
rule to avoid writing to shared objects when we might be
re-entrant. Here we now use FileMgr's methods which isolate OS
implementation, but not FileMgr's file handle pool (as it did not
previously use the pool before this commit). This should allow
this to still take advantage of the Windows OS-specific
implementation, and also avoid the critical section. Can you
please try SVN head and let me know if we are back to 20 out of
20 successes?
Thanks again for the very helpful debug log and exact revision
where failure began.
Troy
On 10/13/20 10:08 PM, Tobias Klein wrote:
I managed to get a backtrace to a segmentation fault using GDB.
It seems like the crash is happening in sword::FileMgr::open( ...
The starting point is sword::InstallMgr::refreshRemoteSource as
I was writing before.
Best regards,
Tobias
Program received signal SIGSEGV, Segmentation fault.
[Switching to Thread 0x7f1af3fff700 (LWP 220833)]
0x00007f1b027045a4 in sword::FileMgr::open(char const*, int,
int, bool) () from
/home/tobi/dev/ezra_project/node-sword-interface-git/build/Release/node_sword_interface.node
(gdb) backtrace
#0 0x00007f1b027045a4 in sword::FileMgr::open(char const*, int,
int, bool) () from
/home/tobi/dev/ezra_project/node-sword-interface-git/build/Release/node_sword_interface.node
#1 0x00007f1b0276ad7b in sword::(anonymous
namespace)::my_fwrite(void*, unsigned long, unsigned long, void*) ()
from
/home/tobi/dev/ezra_project/node-sword-interface-git/build/Release/node_sword_interface.node
#2 0x00007f1b180626bf in ?? () from
/usr/lib/x86_64-linux-gnu/libcurl-gnutls.so.4
#3 0x00007f1b18074a2b in ?? () from
/usr/lib/x86_64-linux-gnu/libcurl-gnutls.so.4
#4 0x00007f1b1807e2e4 in ?? () from
/usr/lib/x86_64-linux-gnu/libcurl-gnutls.so.4
#5 0x00007f1b1807f6f9 in curl_multi_perform () from
/usr/lib/x86_64-linux-gnu/libcurl-gnutls.so.4
#6 0x00007f1b18075d13 in curl_easy_perform () from
/usr/lib/x86_64-linux-gnu/libcurl-gnutls.so.4
#7 0x00007f1b0276b683 in sword::CURLFTPTransport::getURL(char
const*, char const*, sword::SWBuf*) () from
/home/tobi/dev/ezra_project/node-sword-interface-git/build/Release/node_sword_interface.node
#8 0x00007f1b0271d5d2 in
sword::InstallMgr::remoteCopy(sword::InstallSource*, char
const*, char const*, bool, char const*) ()
from
/home/tobi/dev/ezra_project/node-sword-interface-git/build/Release/node_sword_interface.node
#9 0x00007f1b0271edc7 in
sword::InstallMgr::refreshRemoteSource(sword::InstallSource*) ()
from
/home/tobi/dev/ezra_project/node-sword-interface-git/build/Release/node_sword_interface.node
#10 0x00007f1b026ad734 in
RepositoryInterface::refreshIndividualRemoteSource(std::__cxx11::basic_string<char,
std::char_traits<char>, std::allocator<char> >,
std::function<void (unsigned int)>*) ()
from
/home/tobi/dev/ezra_project/node-sword-interface-git/build/Release/node_sword_interface.node
#11 0x00007f1b026b17dd in
std::thread::_State_impl<std::thread::_Invoker<std::tuple<int
(RepositoryInterface::*)(std::__cxx11::basic_string<char,
std::char_traits<char>, std::allocator<char> >,
std::function<void (unsigned int)>*), RepositoryInterface*,
std::__cxx11::basic_string<char, std::char_traits<char>,
std::allocator<char> >, std::function<void (unsigned int)>*> >
>::_M_run() ()
from
/home/tobi/dev/ezra_project/node-sword-interface-git/build/Release/node_sword_interface.node
#12 0x00007f1b1d622cb4 in ?? () from
/usr/lib/x86_64-linux-gnu/libstdc++.so.6
#13 0x00007f1b1e20a609 in start_thread () from
/lib/x86_64-linux-gnu/libpthread.so.0
#14 0x00007f1b1e131103 in clone () from
/lib/x86_64-linux-gnu/libc.so.6
On 10/13/20 1:07 PM, Tobias Klein wrote:
Hi Troy,
I tested more SVN revisions of SWORD trunk (starting from my
stable version until I hit the bug) and I can now say that
SVN Rev. 3759 is the last SVN revision that works without
hanging for the below mentioned scenario. (20 out of 20 tests
successful)
SVN Rev. 3760 is the first SVN revision where the hanging
occurs. The commit message is "First cut at better isolation of
FileIO to FileMgr and providing a WIN32 impl with works with
wchar_t".
Modified files:
include/filemgr.h
include/swbuf.h
lib/bcppmake/libsword.bpr
src/mgr/curlftpt.cpp
src/mgr/curlhttpt.cpp
src/mgr/filemgr.cpp
src/mgr/installmgr.cpp
src/mgr/swmgr.cpp
src/utilfuns/utilstr.cpp
Maybe this helps to find the root-cause.
Best regards,
Tobias
On 10/12/20 9:20 PM, Tobias Klein wrote:
I'll see whether I can collect a stack trace. It may take some
time until I have it.
The multi-threaded "remote source refreshing" worked without
issues until recently.
Here is the code of the function that does the actual work in
a thread.
See
https://github.com/tobias-klein/node-sword-interface/blob/787160ccb4b3bab2a762d22f74031c7237edc803/src/sword_backend/repository_interface.cpp#L105.
intRepositoryInterface::refreshIndividualRemoteSource(stringremoteSourceName,
std::function<void(unsignedintprogress)>*progressCallback)
{
//cout << "Refreshing source " << remoteSourceName << endl <<
flush;
InstallSource* source= this->getRemoteSource(remoteSourceName);
intresult= this->_installMgr->refreshRemoteSource(source);
if(result!= 0) {
cerr<<"Failed to refresh source "<<remoteSourceName<<endl<<flush;
}
remoteSourceUpdateMutex.lock();
this->_remoteSourceUpdateCount++;
unsignedinttotalPercent=
(unsignedint)calculateIntPercentage<double>(this->_remoteSourceUpdateCount,
this->_remoteSourceCount);
if(progressCallback!= 0) {
(*progressCallback)(totalPercent);
}
remoteSourceUpdateMutex.unlock();
returnresult;
}
Best regards,
Tobias
On 10/12/20 9:01 PM, Troy A. Griffitts wrote:
Any luck getting a stack trace on crash?
Regarding the "multitheaded mode", I'd have to get a bit more
information as to exactly how you are sharing SWORD objects
across your threads. Generally, as a rule, you shouldn't. We
recommend a separate instance of SWMgr per thread and that
probably goes for InstallMgr, as well.
Troy
On October 12, 2020 8:29:31 PM GMT+02:00, Tobias Klein
<cont...@tklein.info> wrote:
Hi Troy,
I'm using curl on all three platforms.
Regarding the timeout configuration I have not changed
anything yet, to make this configurable in Ezra Project
is still on my todo list.
I just checked on Linux.
With the old version (May 18th 2020) no hanging or crash
in 10 out of 10 times.
WIth the new version (latest trunk / SWORD 1.9 RC3) I get
1 x crash, 2 x hanging, 7 x working.
I'm running the InstallMgr::refreshRemoteSource "in a
multi-threaded mode".
Best regards,
Tobias
On 10/12/20 6:59 PM, Troy A. Griffitts wrote:
Hi Tobias,
What transport library are you building with? ftplib or
curl?
Have you changed the value of our new timeout from the
default, I believe we decided on, 10 seconds?
Troy
On October 12, 2020 6:46:54 PM GMT+02:00, Tobias Klein
<cont...@tklein.info> wrote:
Hi Troy,
In my latest Ezra Project builds using SWORD trunk I’ve been noticing
random „hangs“ and crashes related to "updating remote sources“. I suppose it
must be around InstallMgr::refreshRemoteSource.
This was still rock solid when using SWORD trunk from May 18th
2020, but not so any more with the recent SWORD trunk.
Unfortunately I cannot pinpoint this more specifically. I just
wanted to first share this observation, because it’s worrying me.
I’ve been noticing this regression both on Windows and macOS. Need
to check later whether this also happens on Linux, cannot recall it right now.
Best regards,
Tobias
------------------------------------------------------------------------
sword-devel mailing list:sword-devel@crosswire.org
http://crosswire.org/mailman/listinfo/sword-devel
<http://crosswire.org/mailman/listinfo/sword-devel>
Instructions to unsubscribe/change your settings at above page
--
Sent from my Android device with K-9 Mail. Please excuse
my brevity.
--
Sent from my Android device with K-9 Mail. Please excuse my
brevity.
_______________________________________________
sword-devel mailing list:sword-devel@crosswire.org
http://crosswire.org/mailman/listinfo/sword-devel
Instructions to unsubscribe/change your settings at above page
_______________________________________________
sword-devel mailing list:sword-devel@crosswire.org
http://crosswire.org/mailman/listinfo/sword-devel
Instructions to unsubscribe/change your settings at above page
_______________________________________________
sword-devel mailing list:sword-devel@crosswire.org
http://crosswire.org/mailman/listinfo/sword-devel
Instructions to unsubscribe/change your settings at above page
_______________________________________________
sword-devel mailing list:sword-devel@crosswire.org
http://crosswire.org/mailman/listinfo/sword-devel
Instructions to unsubscribe/change your settings at above page
