[SR-4677] Intermittent crash in TestFoundation/TestURLSession

[pushkarnk]

Previous ID	SR-4677
Radar	None
Original Reporter	@pushkarnk
Type	Bug
Status	Resolved
Resolution	Done

Additional Detail from JIRA

Votes	0
Component/s	libdispatch
Labels	Bug
Assignee	dgrove-oss (JIRA)
Priority	Medium

md5: 449ba526e5ae862c14b254de13d2ee31

Issue Description:

This time its a dispatch crash. Very intermittent. All I could retrieve from the crash is the backtrace:

* thread #​10, name = 'TestFoundation', stop reason = signal SIGSEGV: invalid address (fault address: 0x38)
  * frame #​0: 0x00007ffff7edee67 libdispatch.so`_dispatch_unote_register + 631
    frame #​1: 0x00007ffff7ed93fb libdispatch.so`_dispatch_source_refs_register + 123
    frame #​2: 0x00007ffff7eda1fb libdispatch.so`_dispatch_source_invoke + 2171
    frame #​3: 0x00007ffff7ed4bf6 libdispatch.so`_dispatch_queue_serial_drain + 326
    frame #​4: 0x00007ffff7ed5130 libdispatch.so`_dispatch_mgr_queue_drain + 64
    frame #​5: 0x00007ffff7edb2cd libdispatch.so`_dispatch_mgr_invoke + 141
    frame #​6: 0x00007ffff7edb08b libdispatch.so`_dispatch_mgr_thread + 11
    frame #​7: 0x00007ffff7ed73b4 libdispatch.so`_dispatch_root_queue_drain + 372
    frame #​8: 0x00007ffff7ed71b8 libdispatch.so`_dispatch_worker_thread + 440
    frame #​9: 0x00007ffff6c866ba libpthread.so.0`start_thread + 202
    frame #​10: 0x00007ffff5d1482d libc.so.6`__clone at clone.S:109

If I have other findings, I'll post them here.

[pushkarnk]

Saw this failure with "TestURLSession.test_dataTaskWithURL()" today

[swift-ci]

Comment by David Grove (JIRA)

Ran more than 10,000 times with no crashes. Unable to reproduce.

[swift-ci]

Comment by David Grove (JIRA)

I've been able to reproduce a crash with this backtrace using the FTP tests from https://github.com/saiHemak/swift-corelibs-foundation/tree/ftpimplementation.

There is definitely a local problem with this code; _dispatch_muxnote_create can return NULL under certain odd circumstances and dispatch_unote_register is blindly dereferencing the result of calling _dispatch_muxnote_create when it is NULL, causing a segfault. Putting in the obvious check for NULL and returning false from dispatch_unote_register appears to be sufficient to allow the FTP tests to pass. I'm preparing a dispatch PR with the fix.--

It is still possible there is a higher-level bug in the layers above here, but the lowlevel code is definitely missing a check for a NULL return that can (does) happen.

[swift-ci]

Comment by David Grove (JIRA)

submitted #300 with a proposed fix.

[swift-ci]

Comment by David Grove (JIRA)

PR merged.