Running concurrent requests segfaults

[simendsjo]

I tried running a couple of concurrent requests using ab, and it sometimes fails.
The line that breaks is this one: https://github.com/rejectedsoftware/vibe.d/blob/master/source/vibe/core/drivers/libevent2_tcp.d#L360
assert(client_ctx) fails after connectionCallback() is called in the line above.

[s-ludwig]

Interesting.. a closure bug? The client_ctx is modified nowhere...

[simendsjo]

Hmmm.. Seems there are some strange stuff going on. If I reduce the test-case to this:
void index(HttpServerRequest req, HttpServerResponse res) { auto cn = new Connection("localhost", "user", "password", "database"); scope(exit) cn.close(); }

Then I get other errors:

[7FFFF7EC8F00:7FFFF7F37380 WRN] A socket error occurred on fd 23: 32

Program received signal SIGSEGV, Segmentation fault.
0x0000000000586f0c in rt.deh2.terminate() ()
(gdb) l
1       dlopen.c: No such file or directory.
        in dlopen.c

Adding in a little more sql:
void index(HttpServerRequest req, HttpServerResponse res) { auto cn = new Connection("localhost", "user", "password", "database"); scope(exit) cn.close(); auto cmd = Command(cn, "SELECT * FROM user"); auto data = cmd.execSQLResult(); while(!data.empty) { data.popFront(); } }

Gives other problems:
Program received signal SIGSEGV, Segmentation fault. 0x00000000005851cd in gc.gcx.Gcx.mark() ()

But if I sprinkle with writeln's, I reproduce the error I first reported...

[simendsjo]

Oh.. And I get failures even without concurrent requests, but a little different. With all writeln's, I get the second reported issue rather than the first when running with 1 request at a time.

I really have no idea how I can be of any help finding the source of this bug.

[s-ludwig]

Unfortunately, I currently don't have any time left to debug the MySQL stuff. If nobody else does it, I will try to get to this in two or three weeks.

[simendsjo]

Hmm.. I'm running a branch with some changes now, and the problem is gone (at least I think the remaining problem is vibe or druntime).

[simendsjo]

It fails on a simple "Hello world" app too.

ab -n1000000 -c1000 http://localhost:8080/

Usually fails before reaching 100,000 for me.

Of course.. It runs fine for smaller number of concurrent users, so it's obviously not an issue for my use, but it might hide some other errors.

[s-ludwig]

Interesting. I do get some failed connections sometimes and I think it's a timing issue where the new connection is not accept'ed fast enough. It doesn't crash though. Hopefully when it's finished, the win32 back end will provide a better target for debugging this. Libevent hides a lot of things under it's abstractions...

[simendsjo]

I get strange segfaults on linux64 at least. Haven't tested on windows or x32.

[s-ludwig]

Closing for now due to non-reproducibility over a long period of time.